content
stringlengths 1
15.9M
|
|---|
\section*{Methods}
\noindent \textbf{Nanofabrication.}
The devices are fabricated on a commercial LNOI wafer (supplied by NANOLN) with 610\,nm thick Z-cut LN thin film on 1.8\,$\mu$m silicon dioxide (SiO$_2$) on a silicon substrate. The bus waveguide is ultimately tapered to a width of 4\,$\mu$m at both facets to improve the fiber-to-chip coupling efficiency. The pattern is defined by a 100\,kV electron beam lithography system (Raith EBPG 5000+) with a negative FOx-16 resist, which is then developed in 25$\%$ TMAH solution to ensure a high contrast. The exposed pattern is transferred onto the LN thin film using an optimized inductively couple plasma (ICP) reactive ion etching (RIE) process with Ar$^+$ plasma. For the subsequent poling process, the radial nickel electrodes are initially patterned on top of the LN microring via the lift-off process. The periodic domain inversion is then enabled by keeping the silicon substrate as the electrical ground while applying several 600\,V, 250\,ms pulses on the electrodes at an elevated temperature of 250\,$^\circ$C. After removing the nickel electrodes, the chip is cleaved to expose the waveguide facets for fiber-to-chip coupling. The insertion losses are calibrated to be 8.4 and 11.1\,dB/facet for the infrared and near-visible lights, respectively.
\vspace{1 mm}
\noindent \textbf{Numerical simulation.}
The parametric signal and idler wavelengths as a function of the pump wavelength and temperature are numerically investigated using a commercial finite-difference-eigenmode solver (Lumerical MODE) based on the equations:
\begin{gather}
\frac{1}{\lambda_p} = \frac{1}{\lambda_s}+\frac{1}{\lambda_i}, \label{energy}\\
\frac{n_p(T)}{\lambda_p} = \frac{n_s(T)}{\lambda_s}+\frac{n_i(T)}{\lambda_i}+\frac{M}{2\pi R}, \label{momentum}
\end{gather}
where $\lambda_{p/s/i}$ and $n_{p/s/i}$ denote the wavelengths and refractive indices for the pump, signal, and idler lights, respectively. In the simulation, we used the Sellmeier equation for congruent LN in Ref.~\cite{Zelmon:97} and its wavelength-dependent thermo-optic coefficient is also considered~\cite{doi:10.1063/1.1988987}.
\def\section*{\textbf{references}}{\section*{\textbf{references}}}
\bibliographystyle{myaipnum4-1}
|
\section{Introduction}
Many domains of science, engineering, and economics make extensive use of models implemented as stochastic numerical simulators \citep{gourieroux1993indirect,ratmann2007using,alsing2018,brehmer2018constraining,karabatsos2018approximate,gonccalves2019training}. A key challenge when studying and validating such simulation-based models is the statistical identification of parameters which are consistent with observed data. In many cases, calculation of the likelihood is intractable or impractical, rendering conventional approaches inapplicable. The goal of simulation-based inference (SBI), also known as `likelihood-free inference', is to perform Bayesian inference without requiring numerical evaluation of the likelihood function \citep{sisson2018_chapter1,cranmer2019}. In SBI, it is generally not required that the simulator is differentiable, nor that one has access to its internal random variables.
In recent years, several new SBI algorithms have been developed \citep[e.g., ][]{gutmann2015,papamakarios2016,lueckmann2017,chan2018,greenberg2019,papamakarios2019a,prangle2019distilling,brehmer2020a,hermans2019,jarvenpaa2020,picchini2020,rodrigues2020,thomas2020}, energized, in part, by advances in probabilistic machine learning \citep{rezende2016,papamakarios2017,papamakarios2019c}. Despite---or possibly \emph{because}---of these rapid and exciting developments, it is currently difficult to assess how different approaches relate to each other theoretically and empirically: First, different studies often use different tasks and metrics for comparison, and comprehensive comparisons on multiple tasks and simulation budgets are rare.
Second, some commonly employed metrics might not be appropriate or might be biased through the choice of hyperparameters.
Third, the absence of a benchmark has made it necessary to reimplement tasks and algorithms for each new study. This practice is wasteful, and makes it hard to rapidly evaluate the potential of new algorithms.
Overall, it is difficult to discern the most promising approaches and decide on which algorithm to use when. These problems are exacerbated by the interdisciplinary nature of research on SBI, which has led to independent development and co-existence of closely-related algorithms in different disciplines.
There are many exciting challenges and opportunities ahead, such as the scaling of these algorithms to high-dimensional data, active selection of simulations, and gray-box settings, as outlined in \citet{cranmer2019}. To tackle such challenges, researchers will need an extensible framework to compare existing algorithms and test novel ideas. Carefully curated, a benchmark framework will make it easier for researchers to enter SBI research, and will fuel the development of new algorithms through community involvement, exchange of expertise and collaboration. Furthermore, benchmarking results could help practitioners to decide which algorithm to use on a given problem of interest, and thereby contribute to the dissemination of SBI.
The catalyzing effect of benchmarks has been evident, e.g.,~in computer vision
\citep{russakovsky2015imagenet},
speech recognition \citep{hirsch2000aurora,wang2018glue}, reinforcement learning \citep{bellemare2013arcade, duan2016benchmarking},
Bayesian deep learning \citep{filos2019systematic,wenzel2020}, and many other fields drawing on machine learning. Open benchmarks can be an important component of transparent and reproducible computational research.
Surprisingly, a benchmark framework for SBI has been lacking, possibly due to the challenging endeavor of designing benchmarking tasks and defining suitable performance metrics.
\input{figs/01_algorithms}
Here, we begin to address this challenge, and provide a benchmark framework for SBI to allow rapid and transparent comparisons of current and future SBI algorithms: First, we selected a set of initial algorithms representing distinct approaches to SBI \citep[\autoref{fig:algorithms}; ][]{cranmer2019}. Second, we analyzed multiple performance metrics which have been used in the SBI literature. Third, we implemented ten tasks including tasks popular in the field.
The shortcomings of commonly used metrics led us to focus on tasks for which a likelihood \emph{can} be evaluated, which allowed us to calculate reference (`ground-truth') posteriors.
These reference posteriors are made available to allow rapid evaluation of SBI algorithms. Code for the framework is available at \link{https://github.com/}{github.com/anonymized}{} and we maintain an interactive version of all results at \link{https://sbi-benchmark.github.io}{sbi-benchmark.github.io}{}.
The full potential of the benchmark will be realized when it is populated with additional community-contributed algorithms and tasks. However, our initial version already provides useful insights: 1) the choice of performance metric is critical; 2) the performance of the algorithms on some tasks leaves substantial room for improvement; 3) sequential estimation generally improves sample efficiency; 4) for small and moderate simulation budgets, neural-network based approaches outperform classical ABC algorithms, confirming recent progress in the field; and 5) there is no algorithm to rule them all. The performance ranking of algorithms is task-dependent, pointing to a need for better guidance or automated procedures for choosing which algorithm to use when. We highlight examples of how the benchmark can be used to diagnose shortcomings of algorithms and facilitate improvements. We end with a discussion of the limitations of the benchmark.
\section{Benchmark}
The benchmark consists of a set of algorithms, performance metrics and tasks. Given a prior $p(\mathbold{\theta})$ over parameters $\mathbold{\theta}$, a simulator to sample $\mathbold{x} \sim p(\mathbold{x}|\mathbold{\theta})$ and an observation $\mathbold{x}_o$, an algorithm returns an approximate posterior $q(\mathbold{\theta}|\mathbold{x}_o)$, or samples from it, $\mathbold{\theta} \sim q$. The approximate solution is tested, according to a performance metric, against a reference posterior $p(\mathbold{\theta}|\mathbold{x}_o)$.
\subsection{Algorithms}
Following the classification introduced in the review by \citet{cranmer2019}, we selected algorithms addressing SBI in four distinct ways, as schematically depicted in \autoref{fig:algorithms}. An important difference between algorithms is how new simulations are acquired: Sequential algorithms adaptively choose informative simulations to increase sample efficiency. While crucial for expensive simulators, it can require non-trivial algorithmic steps and hyperparameter choices. To evaluate whether the potential is realized empirically and justifies the algorithmic burden, we included sequential and non-sequential counterparts for algorithms of each category.
Keeping our initial selection focused allowed us to carefully consider implementation details and hyperparameters: We extensively explored performance and sensitivity to different choices in more than 10k runs, all results and details of which can be found in \autoref{appendix:hyperparams}. Our selection is briefly described below, full algorithm details are in \autoref{appendix:algorithms}.
\textbf{{\color{ABC}{REJ-ABC}}{} and {\color{SABC}{SMC-ABC}}{}.} Approximate Bayesian Computation \citep[ABC,][]{sisson2018_chapter1} is centered around the idea of Monte Carlo rejection sampling \citep{tavere1997,pritchard1999}. Parameters $\mathbold{\theta}$ are sampled from a proposal distribution, simulation outcomes $\mathbold{x}$ are compared with observed data $\mathbold{x}_o$, and are accepted or rejected depending on a (user-specified) distance function and rejection criterion. While rejection ABC ({\color{ABC}{REJ-ABC}}{}) uses the prior as a proposal distribution, the efficiency can be improved by using sequentially refined proposal distributions \citep[{\color{SABC}{SMC-ABC}}{},][]{beaumont2002, marjoram2006, sisson2007, toni2009, beaumont2009}. We implemented {\color{ABC}{REJ-ABC}}{} with quantile-based rejection and used the scheme of \citet{beaumont2009} for {\color{SABC}{SMC-ABC}}{}. We extensively varied hyperparameters and compared the implementation of an ABC-toolbox \citep{klinger2018} against our own (\autoref{appendix:hyperparams}). We investigated linear regression adjustment \citep{blum2010} and the summary statistics approach by \citet{prangle2014semi} (\suppfig{abc_additional}).
\textbf{{\color{NLE}{NLE}}{} and {\color{SNLE}{SNLE}}{}.} Likelihood estimation (or `synthetic likelihood') algorithms learn an approximation to the intractable likelihood, for an overview see \citet{sisson2018_chapter12}. While early incarnations focused on Gaussian approximations \citep[{\color{SL}{SL}}{}; ][]{wood2010}, recent versions utilize deep neural networks \citep{papamakarios2019a,lueckmann2019} to approximate a density over $\mathbold{x}$, followed by MCMC to obtain posterior samples. Since we primarily focused on these latter versions, we refer to them as neural likelihood estimation ({\color{NLE}{NLE}}{}) algorithms, and denote the sequential variant with proposals as {\color{SNLE}{SNLE}}{}. In particular, we used the scheme proposed by \citet{papamakarios2019a} which uses masked autoregressive flows \citep[MAFs, ][]{papamakarios2017} for density estimation. We improved MCMC sampling for ({\color{SNLE}{S}}){\color{NLE}{NLE}}{} and compared MAFs against Neural Spline Flows \citep[NSFs; ][]{durkan2019neural}, see \autoref{appendix:hyperparams}.
\textbf{{\color{NPE}{NPE}}{} and {\color{SNPE}{SNPE}}{}.} Instead of approximating the likelihood, these approaches directly target the posterior. Their origins date back to regression adjustment approaches \citep{blum2010}. Modern variants \citep{papamakarios2016,lueckmann2017,greenberg2019} use neural networks for density estimation (approximating a density over $\mathbold{\theta}$). Here, we used the recent algorithmic approach proposed by \citet{greenberg2019} for sequential acquisitions. We report performance using NSFs for density estimation, which outperformed MAFs (\autoref{appendix:hyperparams}).
\textbf{{\color{NRE}{NRE}}{} and {\color{SNRE}{SNRE}}{}.} Ratio Estimation approaches to SBI use classifiers to approximate density ratios \citep{izbicki2014high,pham2014note,cranmer2015,dutta2016likelihood,durkan2020,thomas2020}.
Here, we used the recent approach proposed by \citet{hermans2019} as implemented in \citet{durkan2020}: A neural network-based classifier approximates probability ratios and MCMC is used to obtain samples from the posterior. {\color{SNRE}{SNRE}}{} denotes the sequential variant of neural ratio estimation ({\color{NRE}{NRE}}{}). In \autoref{appendix:hyperparams} we compare different classifier architectures for ({\color{SNRE}{S}}){\color{NRE}{NRE}}{}.
In addition, we benchmarked Random Forest ABC \citep[{\color{RFABC}{RF-ABC}}{}; ][]{raynal2018}, a recent ABC variant, and Synthetic Likelihood \citep[SL; ][]{wood2010}, mentioned above. However, {\color{RFABC}{RF-ABC}}{} only targets individual parameters (i.e.~assumes posteriors to factorize), and {\color{SL}{SL}}{} requires new simulations for every MCMC step, thus requiring orders of magnitude more simulations than other algorithms. Therefore, we report results for these algorithms separately, in \suppfig{rf_abc} and \suppfig{sl}, respectively.
Algorithms can be grouped with respect to how their output is represented: 1) some return samples from the posterior, $\mathbold{\theta} \sim q(\mathbold{\theta}|\mathbold{x}_o)$ ({\color{ABC}{REJ-ABC}}{}, {\color{SABC}{SMC-ABC}}{}); 2) others return samples and allow evaluation of unnormalized posteriors $\tilde{q}(\mathbold{\theta}|\mathbold{x}_o)$ (({\color{SNLE}{S}}){\color{NLE}{NLE}}{}, ({\color{SNRE}{S}}){\color{NRE}{NRE}}{}); and 3) for some, the posterior density $q(\mathbold{\theta}|\mathbold{x}_o)$ can be evaluated and sampled directly, without MCMC (({\color{SNPE}{S}}){\color{NPE}{NPE}}{}). As discussed below, these properties constrain the metrics that can be used for comparison.
\subsection{Performance metrics}
Choice of a suitable performance metric is central to any benchmark. As the goal of SBI algorithms is to perform full inference, the `gold standard' would be to quantify the similarity between the true posterior and the inferred one with a suitable distance (or divergence) measure on probability distributions. This would require both access to the ground-truth posterior, and a reliable means of estimating similarity between (potentially) richly structured distributions.
Several performance metrics have been used in past research, depending on the constraints imposed by knowledge about ground-truth and the inference algorithm (see \autoref{table:metrics}). In real-world applications, typically only the observation $\mathbold{x}_o$ is known. However, in a benchmarking setting, it is reasonable to assume that one has at least access to the ground-truth parameters $\mathbold{\theta}_o$. There are two commonly used metrics which only require $\mathbold{\theta}_o$ and $\mathbold{x}_o$, but suffer severe drawbacks for our purposes:
\textbf{Probability $\mathbold{\theta}_o$}. The negative log probability of true parameters averaged over different $(\mathbold{\theta}_o, \mathbold{x}_o)$, $-\mathbb{E}[\log q(\mathbold{\theta}_o|\mathbold{x}_o)]$, has been used extensively in the literature \citep{papamakarios2016,durkan2018,greenberg2019,papamakarios2019a,durkan2020,hermans2019}. Its appeal lies in the fact that one does not need access to the ground-truth posterior. However, using it only for a small set of $(\mathbold{\theta}_o, \mathbold{x}_o)$ is highly problematic: It is only a valid performance measure if averaged over a large set of observations sampled from the prior \citep[][detailed discussion including connection to simulation-based calibration in \autoref{appendix:metrics}]{talts2018}. For reliable results, one would require inference for hundreds of $\mathbold{x}_o$ which is only feasible if inference is rapid (amortized) and the density $q$ can be evaluated directly (among the algorithms used here this applies only to {\color{NPE}{NPE}}{}).
\input{tables/01_metrics}
\textbf{Posterior-Predictive Checks (PPCs).} As the name implies, PPCs should be considered a mere check rather than a metric, although the \textit{median distance} between predictive samples and $\mathbold{x}_o$ has been reported in the SBI literature \citep{papamakarios2019a,greenberg2019,durkan2020}. A failure mode of such a metric is that an algorithm obtaining a good MAP point estimate, could perfectly pass this check even if the estimated posterior is poor. Empirically, we found median-distances (MEDDIST) to be in disagreement with other metrics (see \nameref{results}).
The shortcomings of these commonly-used metrics led us to focus on tasks for which it is possible to get samples from ground-truth posterior $\mathbold{\theta} \sim p$, thus allowing us to use metrics based on two-sample tests:
\textbf{Maximum Mean Discrepancy (MMD).} MMD \citep{gretton2012,sutherland2017} is a kernel-based 2-sample test. Recent papers \citep{papamakarios2019a,greenberg2019,hermans2019} reported MMD using translation-invariant Gaussian kernels with length scales determined by the median heuristic \citep{ramdas2015}. We empirically found that MMD can be sensitive to hyperparameter choices, in particular on posteriors with multiple modes and length scales \citep[see \nameref{results} and][]{liu2020}.
\textbf{Classifier 2-Sample Tests (C2ST).} C2STs \citep{friedman2004, lopez-paz2018} train a classifier to discriminate samples from the true and inferred posteriors, which makes them simple to apply and easy to interpret. Therefore, we prefer to report and compare algorithms in terms of accuracy in classification-based tests. In the context of SBI, C2ST has e.g.~been used in \citet{gutmann2018likelihood,dalmasso2019a}.
Other metrics that could be used include:
\textbf{Kernelized Stein Discrepancy (KSD).} KSD \citep{liu2016,chwialkowski2016} is a 1-sample test, which require access to $\nabla_{\mathbold{\theta}}\ \tilde{p}(\mathbold{\theta}|\mathbold{x}_o)$ rather than samples from $p$ ($\tilde{p}$ is the unnormalized posterior). Like MMD, current estimators use translation-invariant kernels.
\textbf{$f$-Divergences.} Divergences such as Total Variation (TV) divergence and KL divergences can only be computed when the densities of true and approximate posteriors can be evaluated (\autoref{table:metrics}). Thus, we did not use $f$-divergences for the benchmark.
Full discussion and details of metrics in \autoref{appendix:metrics}.
\subsection{Tasks}
The preceding considerations guided our selection of inference tasks: We focused on tasks for which reference posterior samples $\mathbold{\theta} \sim p$ can be obtained, to allow calculation of 2-sample tests. We focused on eight purely statistical problems and two problems relevant in applied domains, with diverse dimensionalities of parameters and data (details in \autoref{appendix:tasks}):
\textbf{Gaussian Linear/Gaussian Linear Uniform.} We included two versions of simple, linear, 10-d Gaussian models, in which the parameter $\mathbold{\theta}$ is the mean, and the covariance is fixed. The first version has a Gaussian (conjugate) prior, the second one a uniform prior. These tasks allow us to test how algorithms deal with trivial scaling of dimensionality, as well as truncated support.
\textbf{SLCP/SLCP Distractors.} A challenging inference task designed to have a simple likelihood and a complex posterior \citep{papamakarios2019a, greenberg2019}: The prior is uniform over five parameters $\mathbold{\theta}$ and the data are a set of four two-dimensional points sampled from a Gaussian likelihood whose mean and variance are nonlinear functions of $\mathbold{\theta}$. This induces a complex posterior with four symmetrical modes and vertical cut-offs. We included a second version with 92 additional, non-informative outputs (distractors) to test the ability to detect informative features.
\textbf{Bernoulli GLM/Bernoulli GLM Raw.} 10-parameter Generalized Linear Model (GLM) with Bernoulli observations. Inference was either performed on sufficient statistics (10-d) or raw data (100-d).
\textbf{Gaussian Mixture.} This inference task, introduced by \citet{sisson2007}, has become common in the ABC literature \citep{beaumont2009, toni2009, simola2020}. It consists of a mixture of two two-dimensional Gaussian distributions, one with much broader covariance than the other.
\textbf{Two Moons.} A two-dimensional task with a posterior that exhibits both global (bimodality) and local (crescent shape) structure \citep{greenberg2019} to illustrate how algorithms deal with multimodality.
\textbf{SIR.} Dynamical systems represent paradigmatic use cases for SBI. SIR is an influential epidemiological model describing the dynamics of the number of individuals in three possible states: susceptible $S$, infectious $I$, and recovered or deceased, $R$. We infer the contact rate $\beta$ and the mean recovery rate $\gamma$, given observed infection counts $I$ at 10 evenly-spaced time points.
\textbf{Lotka-Volterra.} An influential model in ecology describing the dynamics of two interacting species, widely used in SBI studies. We infer four parameters $\mathbold{\theta}$ related to species interaction, given the number of individuals in both populations at 10 evenly-spaced points in time.
\input{figs/02_two_moons_metrics}
\subsection{Experimental Setup}
For each task, we sampled 10 sets of true parameters from the prior and generated corresponding observations $(\mathbold{\theta}_o, \mathbold{x}_o)_{1 {:}10}$. For each observation, we generated 10k samples from the reference posterior. Some reference posteriors required a customised (likelihood-based) approach (\autoref{appendix:benchmark}).
In SBI, it is typically assumed that total computation cost is dominated by simulation time.
We therefore report performance at different simulation budgets. For each observation, each algorithm was run with a simulation budget ranging from 1k to 100k simulations.
For each run, we calculated metrics described above. To estimate C2ST accuracy, we trained a multilayer perceptron to tell apart approximate and reference posterior samples and performed five-fold cross-validation. We used two hidden layers, each with 10 times as many ReLu units as the dimensionality of the data.
We also measured and report runtimes (\autoref{appendix:runtimes}).
\subsection{Software}
\textbf{Code.} All code is released publicly at \link{https://github.com/}{github.com/anonymized}{}. Our framework includes tasks, reference posteriors, metrics, plotting, and infrastructure tooling and is designed to be 1) easily extensible, 2) used with external toolboxes implementing algorithms. All tasks are implemented as probabilistic programs in \texttt{Pyro} \citep{bingham2018}, so that likelihoods and gradients for reference posteriors can be extracted automatically. To make this possible for tasks that use ODEs, we developed a new interface between \texttt{DifferentialEquations.jl} \citep{rackauckas2017,bezanson2017julia} and \texttt{PyTorch} \citep{paszke2019}. In addition, specifying simulators in a probabilistic programming language has the advantage that `gray-box' algorithms \citep{brehmer2020a,cranmer2019} can be added in the future. We here evaluated algorithms implemented in \texttt{pyABC} \citep{klinger2018}, \texttt{pyabcranger} \citep{collin2020}, and \texttt{sbi} \citep{sbi}. See \autoref{appendix:benchmark} for details and existing SBI toolboxes.
\textbf{Reproducibility.} Instructions to reproduce experiments on cloud-based infrastructure are in \autoref{appendix:benchmark}.
\textbf{Website.} Along with the code, we provide a web interface which allows interactive exploration of all the results (\link{https://sbi-benchmark.github.io}{sbi-benchmark.github.io}{}; \autoref{appendix:website}).
\input{figs/03_results_metrics}
\section{Results}
\label{results}
We first consider empirical results on a single task, Two Moons, according to different metrics, which illustrate the following important insight:
\textbf{\#1: Choice of performance metric is key.} While C2ST results on Two Moons show that performance increases with higher simulation budgets and that sequential algorithms outperform non-sequential ones for low to medium budgets, these results were not reflected in MMD and MEDDIST (\autoref{fig:two_moons_metrics}): In our analyses, we found MMD to be sensitive to hyperparameter choices, in particular on tasks with complex posterior structure. When using the commonly employed median heuristic to set the kernel length scale on a task with multi-modal posteriors (like Two Moons), MMD had difficulty discerning markedly different posteriors. This can be `fixed' by using hyperparameters adapted to the task (\suppfig{mmd}). As discussed above, the median distance (though commonly used) can be `gamed' by a good point estimate even if the estimated posterior is poor and is thus not a suitable performance metric. Computation of KSD showed numerical problems on Two Moons, due to the gradient calculation.
We assessed relationships between metrics empirically via the correlations across tasks (\suppfig{metrics_correlations}). As discussed above, the log-probability of ground-truth parameters can be problematic when averaged over too few observations (e.g.,~10, as is common in the literature): indeed, this metric had a correlation of only 0.3 with C2ST on Two Moons and 0.6 on the SLCP task. Based on these considerations, we used C2ST for reporting performance (\autoref{fig:results_metrics}; results for MMD, KSD and median distance on the website).
Based on the comparison of the performance across all tasks, we highlight the following main points:
\textbf{\#2: These are not solved problems.} C2ST uses an interpretable scale (1 to 0.5), which makes it possible to conclude that, for several tasks, no algorithm could solve them with the specified budget (e.g.,~SLCP, Lotka-Volterra). This highlights that our problems---though conceptually simple---are challenging, and there is room for development of more powerful algorithms.
\textbf{\#3: Sequential estimation improves sample efficiency.} Our results show that sequential algorithms outperform non-sequential ones (\autoref{fig:results_metrics}). The difference was small on simple tasks (i.e.~linear Gaussian cases), yet pronounced on most others. However, we also found these methods to exhibit diminishing returns as the simulation budget grows, which points to an opportunity for future improvements.
\textbf{\#4: Density or ratio estimation-based algorithms generally outperform classical techniques.} {\color{ABC}{REJ-ABC}}{} and {\color{SABC}{SMC-ABC}}{} were generally outperformed by more recent techniques which use neural networks for density- or ratio-estimation, and which can therefore efficiently interpolate between different simulations (\autoref{fig:results_metrics}).
Without such model-based interpolation, even a simple 10-d Gaussian task can be challenging. However, classical rejection-based methods have a computational footprint that is orders of magnitudes smaller, as no network training is involved (\autoref{appendix:runtimes}). Thus, on low-dimensional problems and for cheap simulators, these methods can still be competitive. See \suppfig{abc_additional} for results with additional ABC variants \citep{blum2010,prangle2014semi} and \suppfig{rf_abc} for results on {\color{RFABC}{RF-ABC}}{}.
\pagebreak
\textbf{\#5: No one algorithm to rule them all.} Although sequential density or ratio estimation-based algorithms performed better than their non-sequential counterparts, there was no clear-cut answer as to which sequential method ({\color{SNLE}{SNLE}}{}, {\color{SNRE}{SNRE}}{}, and {\color{SNPE}{SNPE}}{}) should be preferred. To some degree, this is to be expected: these algorithms have distinct strengths that can play out differently depending on the problem structure \citep[see discussions e.g.,~in][]{greenberg2019, durkan2018, durkan2020}. However, this has not been shown systematically before.
We formulate some practical guidelines for choosing appropriate algorithms in \hyperref[box:advice]{Box~\ref*{box:advice}}.
\textbf{\#6: The benchmark can be used to diagnose implementation issues and improve algorithms.}
For example, ({\color{SNLE}{S}}){\color{NLE}{NLE}}{} and ({\color{SNRE}{S}}){\color{NRE}{NRE}}{} rely on MCMC sampling to compute posteriors, and this sampling step can limit the performance. Access to a reference posterior can help identify and improve such issues: We found that single chains initialized by sampling from the prior with axis-aligned slice sampling \citep[as introduced in][]{papamakarios2019a} frequently got stuck in single modes. Based on this observation, we changed the MCMC strategy (details in \autoref{appendix:algorithms}), which, though simple, yielded significant performance and speed improvements on the benchmark tasks. Similarly, ({\color{SNLE}{S}}){\color{NLE}{NLE}}{} and ({\color{SNRE}{S}}){\color{NRE}{NRE}}{} improved by transforming parameters to be unbounded: Without transformations, runs on some tasks can get stuck during MCMC sampling (e.g.,~Lotka-Volterra). While this is common advice for MCMC \citep{hogg2017}, it has been lacking in code and papers of SBI approaches.
We used the benchmark to systematically compare hyperparameters: For example, as density estimators
\input{04_practical_advice}
for ({\color{SNLE}{S}}){\color{NLE}{NLE}}{} and ({\color{SNPE}{S}}){\color{NPE}{NPE}}{}, we used NSFs \citep{durkan2020} which were developed after these algorithms were published. This revealed that higher capacity density estimators were beneficial for posterior but not likelihood estimation (detailed analysis in \autoref{appendix:hyperparams}).
These examples show how the benchmark makes it possible to diagnose problems and improve algorithms.
\section{Limitations}
\label{limitations}
Our benchmark, in its current form, has several limitations. First, the algorithms considered here do not cover the entire spectrum of SBI algorithms: We did not include sequential algorithms using active learning or Bayesian Optimization \citep{gutmann2015,jarvenpaa2019,lueckmann2019,aushev2020}, or `gray-box' algorithms, which use additional information about or from the simulator \citep[e.g.,][]{baydin2019etalumis,brehmer2020a}. We focused on approaches using neural networks for density estimation and did not compare to alternatives using Gaussian Processes \citep[e.g.,][]{meeds2014,wilkinson2014}. There are many other algorithms which the benchmark is currently lacking \citep[e.g.,][]{nott2014,ong2018,clarte2019, prangle2019distilling, priddle2019, picchini2020, radev2020bayesflow, rodrigues2020}. Keeping our initial selection small allowed us to carefully investigate hyperparameter choices. We focused on sequential algorithms with less sophisticated acquisition schemes and the black-box scenario, since we think these are important baselines for future comparisons.
Second, the tasks we considered do not cover the variety of possible challenges. Notably, while we have tasks with high dimensional data with and without structure, we have not included tasks with high-dimensional spatial structure, e.g.,~images. Such tasks would require algorithms that automatically learn summary statistics while exploring the structure of the data \citep[e.g., ][]{dinev2018,greenberg2019,hermans2019,chen2020neural}, an active research area.
Third, while we extensively investigated tuning choices and compared implementations, the results might nevertheless reflect our own areas of expertise.
Fourth, in line with common practice in SBI, results presented in the paper focused on performance as a function of the number of simulation calls. It is important to remember that differences in computation time can be substantial (see \autoref{appendix:runtimes}): For example, ({\color{SABC}{S}}){\color{ABC}{ABC}}{} was much faster than approaches requiring network training. Overall, sequential neural algorithms exhibited longest runtimes.
Fifth, for reasons described above, we focused on problems for which reference posteriors can be computed. This raises the question of how insights on these problems will generalize to `real-world' simulators. Notably, even these simple problems already identify clear differences between, and limitations of, different SBI approaches. Since it is not possible to rigorously compare the performance of different algorithms directly on `real-world' simulators due to the lack of appropriate metrics, we see the benchmark as a necessary stepping stone towards the development of (potentially automated) selection strategies for practical problems.
Sixth, in practice, the choice of algorithm can depend on aspects that are difficult to quantify: It will depend on the available information about a problem, the inference goal, and the speed of the simulator, among other considerations. We included some practical considerations and recommendations in \hyperref[box:advice]{Box~\ref*{box:advice}}.
Finally, benchmarking is an important tool, but not an end in itself---for example, conceptually new ideas might initially not yield competitive results but only reveal their true value later. Conversely, `overfitting' on benchmarks can lead to the illusion of progress, and result in an undue focus on small implementation details which might not generalize beyond it. It would certainly be possible to cheat on this benchmark: In particular, as the simulators are available, one could use samples (or even likelihoods) to excessively tune hyperparameters \textit{for each task}. This would hardly transfer to practice where such tuning is usually impossible (lack of metrics and expensive simulators). %
Therefore, we carefully compared choices and selected hyperparameters performing best \textit{across tasks} (\autoref{appendix:hyperparams}).
\section{Discussion}
\label{discussion}
Quantitatively evaluating, comparing and improving algorithms through benchmarking is at the core of progress in machine learning. We here provided an initial benchmark for simulation-based inference. If used sensibly, it will be an important tool for clarifying and expediting progress in SBI. We hope that the current results on multiple widely-used algorithms already provide insights into the state of the field, assist researchers with algorithm development, and that our recommendations for practitioners will help them in selecting appropriate algorithms.
We believe that the full potential of the benchmark will be revealed as more researchers participate and contribute. To facilitate this process, and allow users to quickly explore and compare algorithms, we are providing precomputed reference posteriors, a website (\link{https://sbi-benchmark.github.io}{sbi-benchmark.github.io}{}), and open-source code (\link{https://github.com/}{github.com/anonymized}{}).
\newpage
\section*{Acknowledgements}
\addcontentsline{toc}{section}{\protect\numberline{\thesection}{Acknowledgements}}
\label{acknowledgements}
We thank Álvaro Tejero-Cantero, Auguste Schulz, Conor Durkan, François Lanusse, Leandra White, Marcel Nonnenmacher, Michael Deistler, Pedro Rodrigues, Poornima Ramesh, Sören Becker and Theofanis Karaletsos for discussions and comments on the manuscript. In addition, J.-M.L. would like to thank the organisers and participants of the Likelihood-Free Inference Workshop hosted by the Simons Foundation for discussions, in particular, Danley Hsu, François Lanusse, George Papamakarios, Henri Pesonen, Joeri Hermans, Johann Brehmer, Kyle Cranmer, Owen Thomas and Umberto Simola. We also acknowledge and thank the Python \citep{van1995python} and Julia \citep{bezanson2017julia} communities for developing the tools enabling this work, including \texttt{Altair} \citep{altair}, \texttt{DifferentialEquations.jl} \citep{rackauckas2017}, \texttt{Hydra} \citep{hydra}, \texttt{kernel-gof} \citep{kgof}, \texttt{igms} \citep{igms}, \texttt{NumPy} \citep{harris2020}, \texttt{pandas} \citep{pandas}, \texttt{pyABC} \citep{klinger2018}, \texttt{pyabcranger} \citep{collin2020}, \texttt{Pyro} \citep{bingham2018}, \texttt{PyTorch} \citep{paszke2019}, \texttt{sbi} \citep{sbi}, \texttt{Scikit-learn} \citep{scikit-learn}, \texttt{torch-two-sample} \citep{tts}, and \texttt{vega-lite} \citep{vega-lite}.
This work was supported by the German Research Foundation (DFG; SFB 1233 PN 276693517, SFB 1089, SPP 2041, Germany’s Excellence Strategy – EXC number 2064/1 PN 390727645) and the German Federal Ministry of Education and Research (BMBF; project 'ADIMEM', FKZ 01IS18052 A-D).
\subsection{Maximum Mean Discrepancy (MMD)}
Maximum Mean Discrepancy (MMD) is an Integral Probability Metric (IPM). Linear and quadratic time estimates for using MMD as a two-sample test were derived in \cite{gretton2012}. MMD has been commonly used in the SBI literature with Gaussian kernels \citep{papamakarios2019a,greenberg2019,hermans2019}, setting a single length-scale hyperparameter by using a median heuristic \citep{ramdas2015}. We follow the same procedure, i.e., use Gaussian kernels with length-scale determined by the median heuristic on reference samples. MMDs are calculated using 10k samples from reference and approximate posteriors.
If simple kernels are used to compare distributions with complex, multimodal structure, distinct distributions can be mapped to nearby mean embeddings, resulting in low test power. On SLCP and Two Moons, for example, we found a translation-invariant kernel to be limiting, since it cannot adapt to the local structure (see \suppfig{mmd}). This is reflected in the low correlation of MMD and C2ST (\suppfig{metrics_correlations}). We emphasize that these issues are strictly related to simple kernels with hyperparameters commonly used in the literature. Posteriors of the Two Moons task have a structure similar to the blobs example of \citet{liu2020}, who argue for using learned kernels to overcome the aforementioned problem.
\subsection{Median distance (MEDDIST)}
\label{appendix:median_distance}
Posterior predictive checks (PPCs) use the posterior predictive distribution to predict new data, $\mathbold{x}^{\prime} \sim p(\mathbold{x}^{\prime}|\mathbold{x}_o) = \int p(\mathbold{x}^{\prime}|\mathbold{\theta}) q(\mathbold{\theta}|\mathbold{x}_o) \d\mathbold{\theta}$. The observed data $\mathbold{x}_o$ should look plausible under the posterior predictive distribution (\cite{gelman2004}, chapter 6). A particular PPC, used for example in \cite{papamakarios2019a, greenberg2019, durkan2020}, is to assess the median L2 distance between $N^{\prime}$ posterior predictive samples $\mathbold{x}^{\prime}$ and $\mathbold{x}_o$. The median is used since the mean would be more sensitive to outliers.
In the benchmark, we refer to this metric as median distance (MEDDIST) and drew $N^{\prime} = 10000$ samples from each posterior predictive distribution to compute it. In contrast with other metrics considered here, the median distance is computed in the space of data $\mathbold{x}$ and requires additional simulations (which could be expensive, depending on the simulator). The median distance should be considered a mere check rather than a metric and it does not necessarily test the structure of the estimated posterior.
\subsection{Kernelized Stein Discrepancy (KSD)}
Kernelized Stein Discrepancy (KSD) is a 1-sample goodness-of-fit test proposed independently by \citet{chwialkowski2016} and \citet{liu2016}. KSD tests samples from algorithms against the gradient of unnormalized true posterior density, $\nabla_{\mathbold{\theta}}\ \tilde{p}(\mathbold{\theta}|\mathbold{x}_o)$. We used KSD with Gaussian kernels, setting the length-scale through the median heuristic, and 10k samples from each algorithm.
\subsection{Classifier-based tests (C2ST)}
In classifier-based testing, a classifier is trained to distinguish samples of the true posterior $p(\mathbold{\theta}|\mathbold{x}_o)$ from samples of the estimated posterior $q(\mathbold{\theta}|\mathbold{x}_o)$. If the samples are indistinguishable, the classification performance should be at chance level, 0.5. Practical use and properties of classifier-based 2-sample testing (C2ST) are discussed in \citet{lopez-paz2018} \citep[see][for examples in the context of SBI]{gutmann2018likelihood,dalmasso2019a}.
To compute C2ST, we trained a two-layer neural network with 10 times as many ReLU units as the dimensionality of parameters, and optimize with Adam \citep{kingma2014adam}. Classifiers were trained on 10k z-scored samples from reference and approximate posterior each. Classification accuracy was reported using 5-fold cross-validation.
\subsection{Simulation-based calibration (SBC)}
In simulation-based calibration (SBC), samples $\mathbold{\theta}^\prime$ are drawn from the data-averaged posterior, i.e., the posterior obtained by running inference for many observations. When the posterior approximation is exact, $\mathbold{\theta}^\prime$ is distributed according to the prior \citep{talts2018}.
Let us briefly illustrate this: In SBC, we draw $\mathbold{\theta} \sim p(\mathbold{\theta}), \mathbold{x} \sim p(\mathbold{x}|\mathbold{\theta}), \mathbold{\theta}^\prime \sim q(\mathbold{\theta}^\prime|\mathbold{x})$, which implies a joint distribution $\pi(\mathbold{\theta}, \mathbold{x}, \mathbold{\theta}^\prime) = p(\mathbold{\theta}) p(\mathbold{x}|\mathbold{\theta}) q(\mathbold{\theta}^\prime|\mathbold{x})$. The marginal $\pi(\mathbold{\theta}^\prime)$ is then:
$$
\begin{aligned}
\pi(\mathbold{\theta}^\prime) = \int\int p(\mathbold{\theta}) p(\mathbold{x}|\mathbold{\theta}) q(\mathbold{\theta}^\prime|\mathbold{x}) \d\mathbold{x} \d\mathbold{\theta} = \int\int p(\mathbold{\theta}, \mathbold{x}) q(\mathbold{\theta}^\prime|\mathbold{x}) \d\mathbold{x} \d\mathbold{\theta} = \int p(\mathbold{x})\ q(\mathbold{\theta}^\prime|\mathbold{x}) \d\mathbold{x}.
\end{aligned}
$$
If the approximate posterior is the true posterior, the marginal on $\mathbold{\theta}^\prime$ is equal to the prior: If $q(\mathbold{\theta}^\prime|\mathbold{x}) = p(\mathbold{\theta}^\prime|\mathbold{x})$, then $\pi(\mathbold{\theta}^\prime)= \int p(\mathbold{x}, \mathbold{\theta}') \d\mathbold{x} = p(\mathbold{\theta}^\prime)$, i.e., one can set up a consistency test that is based on the distribution of $\mathbold{\theta}^\prime$ samples. \cite{talts2018} do this by using frequentist tests per dimension.
Note that SBC as described above is merely a consistency check. For example, if the approximate posterior were the prior, a calibration test as described above would not be able to detect this. This is a realistic failure mode in simulation-based inference. It could happen with rejection ABC in the limit $\epsilon \rightarrow \infty$, or when learned summary statistics have no information about $\mathbold{\theta}$. One way around this is issue is proposed in \cite{prangle2014diag}, who propose to restrict observations to a subset of all possible $\mathcal{X}$.
SBC is similar to the average negative log likelihood of true parameters described above, in that inference needs to be carried out for many observations generated by sampling from the prior. Running inference for hundreds of observations would become prohibitively expensive in terms of compute for most algorithms, which is why we do not rely on SBC in the benchmark.
\subsection[Negative log probability of true parameters (NLTP)]{Negative log probability of $\mathbold{\theta}_o$ (NLTP)}
\label{appendix:metrics:true_parameters}
In simulation-based inference, the average negative log likelihood of true parameters $-\mathbb{E}[\log q(\mathbold{\theta}_o|\mathbold{x}_o)]$ (NLTP) is commonly reported as a performance metric in the literature \citep{papamakarios2016,durkan2018,papamakarios2019a,greenberg2019,hermans2019,durkan2020}. An attractive property of this metric is that the access to the ground-truth posterior is not required.
It is important to point out, however, that calculating this metric on a single or small number of pairs $(\mathbold{\theta}_o, \mathbold{x}_o)$ is problematic. To illustrate the issue, consider the following example (as discussed in \cite{talts2018}): Consider $\theta \sim \mathcal{N}(0, 1^2), x | \theta \sim \mathcal{N}(\theta, {1}^2)$, and a single pair $(\mathbold{\theta}_o, \mathbold{x}_o)$ with $\theta_o=0$ and an implausible (but possible) $x_o=2.1$. In this case, the true posterior is $\mathcal{N}(\theta|1.05, 0.5^2)$ under which the $\theta_o$ has low probability since it is more than two standard deviations away from the posterior mean. If an algorithm fitted a wrong posterior, e.g., by overestimating the standard deviation as 1 instead of 0.5, the probability of $\theta_o$ under the estimated posterior would be higher than under the true posterior.
Therefore, a large number of pairs $(\mathbold{\theta}_o, \mathbold{x}_o)$ should be used. Indeed, in the limit of infinite number of pairs $(\mathbold{\theta}_o, \mathbold{x}_o)$, the metric converges to a $D_{\text{KL}}$:
$$
\begin{aligned}
&\mathbb{E}_{\mathbold{\theta}_o \sim p(\mathbold{\theta})} \mathbb{E}_{\mathbold{x}_o \sim p(\mathbold{x}|\mathbold{\theta}_o)}\big[ -\log q(\mathbold{\theta}_o|\mathbold{x}_o) \big] \\
=\ &\mathbb{E}_{\mathbold{x}_o \sim p(\mathbold{x}), \mathbold{\theta}_o \sim p(\mathbold{\theta}|\mathbold{x}_o)}\big[ -\log q(\mathbold{\theta}_o|\mathbold{x}_o) \big] \\
=\ &\mathbb{E}_{\mathbold{x}_o \sim p(\mathbold{x}), \mathbold{\theta}_o \sim p(\mathbold{\theta}|\mathbold{x}_o)}\big[ -\log q(\mathbold{\theta}_o|\mathbold{x}_o) + \log p(\mathbold{\theta}_o|\mathbold{x}_o) \big] - \mathbb{E}_{\mathbold{x}_o \sim p(\mathbold{x}), \mathbold{\theta}_o \sim p(\mathbold{\theta}|\mathbold{x}_o)}\big[\log p(\mathbold{\theta}_o|\mathbold{x}_o) \big] \\
=\ &\mathbb{E}_{\mathbold{x}_o \sim p(\mathbold{x})} D_{\text{KL}}(p(\mathbold{\theta} | \mathbold{x}_o) || q(\mathbold{\theta}|\mathbold{x}_o)) + \mathbb{E}_{\mathbold{x}_o \sim p(\mathbold{x})} \mathbb{H}(p(\mathbold{\theta} | \mathbold{x}_o))
\end{aligned}
$$
The first term in the final equation is the average $D_{\text{KL}}$ between true and approximate posteriors over all observations $\mathbold{x}_o$ that can be generated when sampling parameters $\mathbold{\theta}_o$ from the prior. The second term, the entropy term, would be the same for all algorithms compared.
In the context of this benchmark, we decided against using the probability of $\mathbold{\theta}_o$ as a metric: For all algorithms that are not amortized (all but one), evaluating posteriors at different $\mathbold{x}_o$ would require rerunning inference. As the computational requirements for running the benchmark at 10 observations per task are already high, running tasks for hundreds of observations would become prohibitively expensive.
\subsection{Gaussian Linear Uniform}
\label{appendix:task:gaussian_linear_uniform}
Inference of the mean of a 10-d Gaussian model, in which the covariance is fixed. The prior is uniform:
{\setlength{\extrarowheight}{5pt}
\begin{tabularx}{\textwidth}{@{}ll@{}}
\textbf{Prior} & %
$\mathcal{U}(-\mathbold{1}, \mathbold{1})$ \\
%
\textbf{Simulator} & %
\bigcell{l}{
$ \mathbold{x}|\mathbold{\theta} \sim \mathcal{N}(\mathbold{x}|\mathbold{m}_{\mathbold{\theta}}=\mathbold{\theta}, \mathbold{S} = 0.1 \odot \mathbold{I}) $
} \\
%
\textbf{Dimensionality} & %
$\mathbold{\theta} \in \mathbb{R}^{10}, \mathbold{x} \in \mathbb{R}^{10}$ \\
%
\end{tabularx}}
\subsection{Gaussian Linear}
\label{appendix:task:gaussian_linear}
Inference of the mean of a 10-d Gaussian model, in which the covariance is fixed. The (conjugate) prior is Gaussian:
{\setlength{\extrarowheight}{5pt}
\begin{tabularx}{\textwidth}{@{}ll@{}}
\textbf{Prior} & %
$\mathcal{N}(\mathbold{0}, 0.1 \odot \mathbold{I})$ \\
%
\textbf{Simulator} & %
\bigcell{l}{
$ \mathbold{x}|\mathbold{\theta} \sim \mathcal{N}(\mathbold{x}|\mathbold{m}_{\mathbold{\theta}}=\mathbold{\theta}, \mathbold{S} = 0.1 \odot \mathbold{I}) $
} \\
%
\textbf{Dimensionality} & %
$\mathbold{\theta} \in \mathbb{R}^{10}, \mathbold{x} \in \mathbb{R}^{10}$ \\
%
\end{tabularx}}
\subsection{SLCP}
\label{appendix:task:slcp}
A challenging inference task designed to have a simple likelihood and a complex posterior. The prior is uniform over five parameters $\mathbold{\theta}$ and the data are a set of four two-dimensional points sampled from a Gaussian likelihood whose mean and variance are nonlinear functions of $\mathbold{\theta}$:
{\setlength{\extrarowheight}{5pt}
\begin{tabularx}{\textwidth}{@{}ll@{}}
\textbf{Prior} & %
$\mathcal{U}(-\mathbold{3}, \mathbold{3})$ \\
%
\textbf{Simulator} & %
\bigcell{l}{
$ \mathbold{x}|\mathbold{\theta} = (\mathbold{x}_{1}, \ldots, \mathbold{x}_{4})$, $ \mathbold{x}_i \sim \mathcal{N}(\mathbold{m}_{\mathbold{\theta}}, \mathbold{S}_{\mathbold{\theta}}) $, \\
where $ \mathbold{m}_{\mathbold{\theta}} = \begin{bmatrix} {\theta_{1}} \\ {\theta_{2}} \end{bmatrix} $,
$ \mathbold{S}_{\mathbold{\theta}} = \begin{bmatrix} {s_{1}^{2}} & {\rho s_{1} s_{2}} \\ {\rho s_{1} s_{2}} & {s_{2}^{2}}\end{bmatrix} $,
$ s_{1}=\theta_{3}^{2}, s_{2}=\theta_{4}^{2}, \rho=\tanh \theta_{5}$
} \\
%
\textbf{Dimensionality} & %
$\mathbold{\theta} \in \mathbb{R}^5, \mathbold{x} \in \mathbb{R}^8$ \\
%
\textbf{References} & %
\bigcell{l}{\cite{papamakarios2019a,greenberg2019,hermans2019} \\
\cite{durkan2020}}
\end{tabularx}}
\subsection{Gaussian Mixture}
\label{appendix:task:gaussian_mixture}
This task is common in the ABC literature. It consists of inferring the common mean of a mixture of two two-dimensional Gaussian distributions, one with much broader covariance than the other:
{\setlength{\extrarowheight}{5pt}
\begin{tabularx}{\textwidth}{@{}ll@{}}
\textbf{Prior} & %
$\mathcal{U}(-\mathbold{10}, \mathbold{10})$ \\
%
\textbf{Simulator} & %
\bigcell{l}{
$ \mathbold{x}|\mathbold{\theta} \sim 0.5 \; \mathcal{N}(\mathbold{x}|\mathbold{m}_{\mathbold{\theta}}=\mathbold{\theta}, \mathbold{S} = \mathbold{I}) + $ $0.5 \; \mathcal{N}(\mathbold{x}|\mathbold{m}_{\mathbold{\theta}}=\mathbold{\theta}, \mathbold{S} = 0.01 \odot \mathbold{I}) $
} \\
%
\textbf{Dimensionality} & %
$\mathbold{\theta} \in \mathbb{R}^{2}, \mathbold{x} \in \mathbb{R}^{2}$ \\
%
\textbf{References} & %
\cite{sisson2007,beaumont2009,toni2009,simola2020}
\end{tabularx}}
\newpage
\subsection{Lotka-Volterra}
\label{appendix:task:lotka_volterra}
This is an influential model in ecology describing the dynamics of two interacting species, most commonly prey and predator interactions. Our task consists in the inference of four parameters $\mathbold{\theta}$ related to species interaction, given 20 summary statistics consisting of the number of individuals in both populations in 10 evenly-spaced points in time:
{\setlength{\extrarowheight}{5pt}
\begin{tabularx}{\textwidth}{@{}ll@{}}
\textbf{Prior} & %
\bigcell{l}{
$\alpha \sim \text{LogNormal}(-0.125, 0.5)$,
$\beta \sim \text{LogNormal}(-3, 0.5)$, \\
$\gamma \sim \text{LogNormal}(-0.125, 0.5)$,
$\delta \sim \text{LogNormal}(-3, 0.5)$
} \\
%
\textbf{Simulator} & %
\bigcell{l}{
$ \mathbold{x}|\mathbold{\theta} = (\mathbold{x}_{1}, \ldots, \mathbold{x}_{10})$,
$\mathbold{x}_{1,i} \sim \text{LogNormal}(\log(X),0.1)$,
$\mathbold{x}_{2,i} \sim \text{LogNormal}(\log(Y),0.1)$, \\
$X$ and $Y$ are simulated from \\
$\frac{dX}{dt} = \alpha X - \beta X Y$ \\
$\frac{dY}{dt} = -\gamma Y + \delta X Y$
} \\
%
\textbf{Dimensionality} & %
$\mathbold{\theta} \in \mathbb{R}^4, \mathbold{x} \in \mathbb{R}^{20}$ \\
%
\textbf{Fixed parameters} & %
\bigcell{l}{
Duration of task $T=20$. Initial conditions:
$(X(0),Y(0))= (30,1)$
}\\
%
\textbf{References} & %
\cite{lotka1920}
\end{tabularx}}
\subsection{SLCP with Distractors}
\label{appendix:task:slcp_distractors}
This task is similar to \ref{appendix:task:slcp}, with the difference that we add uninformative dimensions (distractors) to the observation:
{\setlength{\extrarowheight}{5pt}
\begin{tabularx}{\textwidth}{@{}ll@{}}
\textbf{Prior} & %
$\mathcal{U}(-\mathbold{3}, \mathbold{3})$ \\
%
\textbf{Simulator} & %
\bigcell{l}{
$ \mathbold{x}|\mathbold{\theta} = (\mathbold{x}_{1}, \ldots, \mathbold{x}_{100})$,
$\mathbold{x} = p(\mathbold{y})$, where $p$ re-orders the dimensions of $\mathbold{y}$ with a fixed
random \\ permutation, \\
$ \mathbold{y}_{[1:8]} \sim \mathcal{N}(\mathbold{m}_{\mathbold{\theta}}, \mathbold{S}_{\mathbold{\theta}}) $,
$ \mathbold{y}_{[9:100]} \sim \frac{1}{20}\sum_{i=1}^{20} t_2(\mathbold{\mu}^i,\mathbold{\Sigma}^i)$ \\
where $ \mathbold{m}_{\mathbold{\theta}} = \begin{bmatrix} {\theta_{1}} \\ {\theta_{2}} \end{bmatrix} $,
$ \mathbold{S}_{\mathbold{\theta}} = \begin{bmatrix} {s_{1}^{2}} & {\rho s_{1} s_{2}} \\ {\rho s_{1} s_{2}} & {s_{2}^{2}}\end{bmatrix} $,
$ s_{1}=\theta_{3}^{2}, s_{2}=\theta_{4}^{2}, \rho=\tanh \theta_{5}$, \\
$ \mathbold{\mu}^i \sim \mathcal{N}(0,15^2 \mathbold{I})$,
$ \mathbold{\Sigma}_{j,k}^i \sim \mathcal{N}(0,9)$, for $j>k$,
$ \mathbold{\Sigma}_{j,j}^i = 3 e^a$, where $a \sim \mathcal{N}(0,1)$,
$ \mathbold{\Sigma}_{j,k}^i = 0$ otherwise
} \\
%
\textbf{Dimensionality} & %
$\mathbold{\theta} \in \mathbb{R}^5, \mathbold{x} \in \mathbb{R}^{100}$ \\
%
\textbf{References} & %
\cite{greenberg2019}
\end{tabularx}}
\subsection{Two Moons}
\label{appendix:task:two_moons}
A two-dimensional task with a posterior that exhibits both global (bimodality) and local (crescent shape) structure to illustrate how algorithms deal with multimodality:
{\setlength{\extrarowheight}{5pt}
\begin{tabularx}{\textwidth}{@{}ll@{}}
\textbf{Prior} & %
$\mathcal{U}(-\mathbold{1}, \mathbold{1})$ \\
%
\textbf{Simulator} & %
\bigcell{l}{
$ \boldsymbol{x} | \mathbold{\theta} = \begin{bmatrix}
r \cos(\alpha)+0.25\\
r \sin(\alpha)
\end{bmatrix} + $
$\begin{bmatrix}
{-|\theta_1+\theta_2|}/{\sqrt{2}} \\
{(-\theta_{1}+\theta_{2}})/{\sqrt{2}} \\
\end{bmatrix} $,
where $ \alpha \sim \mathcal{U}(-\pi/{2}, \pi/2) $,
$ r \sim \mathcal{N}(0.1,0.01^{2}) $
} \\
%
\textbf{Dimensionality} & %
$\mathbold{\theta} \in \mathbb{R}^2, \mathbold{x} \in \mathbb{R}^2$ \\
%
\textbf{References} & %
\cite{greenberg2019}
\end{tabularx}}
\subsection{Bernoulli GLM}
\label{appendix:task:bernoulli_glm}
Inference of a 10-parameter Generalized linear model (GLM) with Bernoulli observations, and Gaussian prior with covariance matrix which encourages smoothness by penalizing the second-order differences in the vector of parameters \citep{DeNicolao1997}. The observations are the sufficient statistics for this GLM:
{\setlength{\extrarowheight}{5pt}
\begin{tabularx}{\textwidth}{@{}ll@{}}
\textbf{Prior} & %
\bigcell{l}{
$\beta \sim \mathcal{N}(0,2)$,
$\mathbold{f} \sim \mathcal{N}(\mathbold{0}, (\mathbold{F}^{\top} \mathbold{F})^{-1})$, \\
$\mathbold{F}_{i,i-2} = 1$,
$\mathbold{F}_{i,i-1} = -2$,
$\mathbold{F}_{i,i} = 1+\sqrt{\frac{i-1}{9}}$,
$\mathbold{F}_{i,j} = 0$ otherwise,
$1\leq i,j \leq9$
}\\
%
\textbf{Simulator} & %
\bigcell{l}{
$\mathbold{x}|\mathbold{\theta} = (\mathbold{x}_{1}, \ldots, \mathbold{x}_{10})$,
$\mathbold{x}_1 = \sum_i^T z_i$,
$\mathbold{x}_{2:10} = \frac{1}{\mathbold{x}_1} \mathbold{V} \mathbold{z}$, \\
$z_i \sim \mathrm{Bern}(\eta(\mathbf{v}_i^{\top} \mathbold{f} + \beta))$,
$\eta(\cdot)=\exp(\cdot)/(1 + \exp(\cdot))$, \\
frozen input between time bins $i-8$ and $i$:
$\mathbold{v}_i \sim \mathcal{N}(\mathbold{0},\mathbold{I})$,
$\mathbold{V} = [v_1, v_2, \ldots, v_T]$
} \\
%
\textbf{Dimensionality} & %
$\mathbold{\theta} \in \mathbb{R}^{10}, \mathbold{x} \in \mathbb{R}^{10}$ \\
%
\textbf{Fixed parameters} & %
\bigcell{l}{
Duration of task $T=100$.
} \\
%
\textbf{References} & %
\cite{lueckmann2017,gonccalves2019training}
\end{tabularx}}
\subsection{SIR}
\label{appendix:task:sir}
The SIR model is an epidemiological model describing the dynamics of the number of individuals in three possible states: susceptible $S$, infectious $I$, and recovered or deceased $R$.
The SIR task consists in inferring the contact rate $\beta$ and the mean recovery rate $\gamma$, given a sampled number of individuals in the infectious group $I$ in 10 evenly-spaced points in time:
{\setlength{\extrarowheight}{5pt}
\begin{tabularx}{\textwidth}{@{}ll@{}}
\textbf{Prior} & %
\bigcell{l}{
$\beta \sim \text{LogNormal}(\log(0.4), 0.5)$
$\gamma \sim \text{LogNormal}(\log(1/8), 0.2)$
} \\
%
\textbf{Simulator} & %
\bigcell{l}{
$ \mathbold{x}|\mathbold{\theta} = (x_{1}, \ldots, x_{10})$,
$x_i \sim \mathcal{B}(1000,\frac{I}{N}) $, where I is simulated from \\
$\frac{dS}{dt} = -\beta \frac{SI}{N}$ \\
$\frac{dI}{dt} = \beta \frac{SI}{N}-\gamma I$ \\
$\frac{dR}{dt} = \gamma I$
} \\
%
\textbf{Dimensionality} & %
$\mathbold{\theta} \in \mathbb{R}^2, \mathbold{x} \in \mathbb{R}^{10}$ \\
%
\textbf{Fixed parameters} & %
\bigcell{l}{
Population size $N=1000000$ and duration of task $T=160$. \\
Initial conditions:
$(S(0),I(0),R(0)) = (N-1,1,0)$
} \\
%
\textbf{References} & %
\cite{kermack1927}
\end{tabularx}}
\subsection{Bernoulli GLM Raw}
\label{appendix:task:bernoulli_glm_raw}
This task is similar to \ref{appendix:task:bernoulli_glm}, the sole difference being that the observations are not the sufficient statistics for the Bernoulli GLM process but the raw observations:
{\setlength{\extrarowheight}{5pt}
\begin{tabularx}{\textwidth}{@{}ll@{}}
\textbf{Prior} & %
\bigcell{l}{
$\beta \sim \mathcal{N}(0,2)$,
$\mathbold{f} \sim \mathcal{N}(\mathbold{0}, (\mathbold{F}^{\top} \mathbold{F})^{-1})$, \\
$\mathbold{F}_{i,i-2} = 1$, $\mathbold{F}_{i,i-1} = -2$,
$\mathbold{F}_{i,i} = 1+\sqrt{\frac{i-1}{9}}$,
$\mathbold{F}_{i,j} = 0$ otherwise
$1\leq i,j \leq9$
}\\
%
\textbf{Simulator} & %
\bigcell{l}{
$\mathbold{x}|\mathbold{\theta} = (\mathbold{x}_{1}, \ldots, \mathbold{x}_{100})$,
$x_i \sim \mathrm{Bern}(\eta(\mathbf{v}_i^{\top} \mathbold{f} + \beta))$,
$\eta(\cdot)=\exp(\cdot)/(1 + \exp(\cdot))$ \\
frozen input between time bins $i-8$ and $i$:
$\mathbold{v}_i \sim \mathcal{N}(\mathbold{0},\mathbold{I})$,
} \\
%
\textbf{Dimensionality} & %
$\mathbold{\theta} \in \mathbb{R}^{10}, \mathbold{x} \in \mathbb{R}^{100}$ \\
%
\textbf{Fixed parameters} & %
Duration of task $T=100$.\\
\end{tabularx}}
\section{Figures}
\label{appendix:figures}
\begin{figure*}[h!]
\centering
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 3,clip,width=\textwidth]{figs/abc_lra_sass/gaussian_linear_gaussian_linear_uniform_C2ST.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/abc_lra_sass/slcp_slcp_distractors_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/abc_lra_sass/bernoulli_glm_bernoulli_glm_raw_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/abc_lra_sass/gaussian_mixture_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/abc_lra_sass/two_moons_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/abc_lra_sass/sir_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 10 0 0,clip,width=\textwidth]{figs/abc_lra_sass/lotka_volterra_C2ST_no_labels.pdf}
\end{subfigure} %
%
\caption{
{\bf Additional ABC results with linear regression adjustment (LRA) and semi-automatic summary-statistics (SASS).} We ran ABC with post-hoc LRA \citep{beaumont2002, sisson2018_chapter3}. On some tasks, this led to an improvement relative to versions without post-hoc adjustment. On Two Moons (bimodal posterior), linear adjustment decreased performance. We implemented our own SASS \citep{prangle2014semi} with a third order polynomial feature expansion, and observed similar performance as with the implementation in \texttt{abcpy} toolbox \citep{abcpy-repo}.
%
}
%
\label{fig:abc_additional}
%
\end{figure*}
\newpage
\clearpage
\begin{figure*}[h!]
\centering
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 3,clip,width=\textwidth]{figs/rf_abc/gaussian_linear_gaussian_linear_uniform_C2ST.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/rf_abc/slcp_slcp_distractors_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/rf_abc/bernoulli_glm_bernoulli_glm_raw_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/rf_abc/gaussian_mixture_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/rf_abc/two_moons_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/rf_abc/sir_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 10 0 0,clip,width=\textwidth]{figs/rf_abc/lotka_volterra_C2ST_no_labels.pdf}
\end{subfigure}
%
\caption{
{\bf {\color{RFABC}{RF-ABC}}{} results.} Results for {\color{RFABC}{RF-ABC}}{} (as described in \ref{appendix:algorithms:rfabc}) compared to {\color{ABC}{REJ-ABC}}{} and {\color{SABC}{SMC-ABC}}{} on all benchmark tasks, using C2ST. Note that {\color{RFABC}{RF-ABC}}{} predicts each parameter individually, i.e. effectively assumes the posterior to be factorized-- this is only appropriate for the Gaussian Linear, Gaussian Linear Uniform, and Gaussian Mixture tasks. On other tasks, the posterior deviates markedly from being factorized, and therefore it is to be expected that {\color{RFABC}{RF-ABC}}{} performance is limited, even when using many samples. Each data point corresponds to the mean and 95\% confidence interval across 10 observations.
%
}
%
\label{fig:rf_abc}
%
\end{figure*}
\newpage
\clearpage
\begin{figure}[h!]
\centering
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 55 0 3,clip,width=\textwidth]{figs/sl/gaussian_linear_gaussian_linear_uniform_C2ST.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 55 0 0,clip,width=\textwidth]{figs/sl/slcp_slcp_distractors_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 55 0 0,clip,width=\textwidth]{figs/sl/bernoulli_glm_bernoulli_glm_raw_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 55 0 0,clip,width=\textwidth]{figs/sl/gaussian_mixture_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 55 0 0,clip,width=\textwidth]{figs/sl/two_moons_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 10 0 0,clip,width=\textwidth]{figs/sl/sir_C2ST_no_labels.pdf}
\end{subfigure}
%
%
\caption{
{\bf {\color{SL}{SL}}{} results.} Results for {\color{SL}{SL}}{} compared to {\color{NLE}{NLE}}{} and {\color{SNLE}{SNLE}}{} on benchmark tasks in terms of C2ST. Note that {\color{SL}{SL}}{} performs simulations at every MCMC step to approximate a Gaussian likelihood (see \ref{appendix:algorithms:sl} for details), and therefore it does not produce sensible results with the simulation budgets of other algorithms (between 1k and 100k), . In our experiments, {\color{SL}{SL}}{} required on the order of $10^8$ to $10^9$ simulations. For the SLCP Distractors and Lotka-Volterra stable estimation of covariances was not possible, which is why these tasks were omitted (details in \ref{appendix:algorithms:sl}). We do not report {\color{SL}{SL}}{} results in the main paper, given the huge difference in simulation budget. Each data point corresponds to the mean and 95\% confidence interval across 10 observations.
%
}
%
\label{fig:sl}
%
\end{figure}
\newpage
\clearpage
\begin{figure}[h!]
\centering
%
\begin{subfigure}[htbp]{0.45\textwidth}
\includegraphics[trim=0 0 0 0,clip,width=\textwidth]{figs/mmd/rej_abc.png}
\end{subfigure}
%
\begin{subfigure}[htbp]{0.45\textwidth}
\includegraphics[trim=0 0 0 0,clip,width=\textwidth]{figs/mmd/snle.png}
\end{subfigure}
%
\caption{
{\bf MMD on Two Moons.} When using MMD with the median heuristic (as commonly done, including in SBI papers), MMD is slightly lower for the posterior obtained by {\color{ABC}{REJ-ABC}}{} (left, blue samples), than for {\color{SNLE}{SNLE}}{} samples (right, green samples): 0.00729 ({\color{ABC}{REJ-ABC}}{}) versus 0.00772 ({\color{NLE}{NLE}}{}). This is at odds with the visual impression of the quality of the fit (reference samples in gray) as well as C2ST results: A classifier performed near chance level (.502) for {\color{SNLE}{SNLE}}{} samples while being able to tell apart {\color{ABC}{REJ-ABC}}{} samples from the reference with accuracy 0.794. When manually choosing a length scale on the median distance of a \textit{single crescent} (i.e., 0.09 instead of 1.78), MMD results were in agreement with C2ST results: 0.00738 ({\color{ABC}{REJ-ABC}}{}) versus 0.00035 ({\color{SNLE}{SNLE}}{}), i.e., they also suggested a better fit for {\color{SNLE}{SNLE}}{}. In the main paper, we prefer to report C2ST because we found it less sensitive to hyperparameters: reliance on the commonly used median heuristic can be problematic on tasks with complex posterior structure, e.g., multi-modality in Two Moons, as demonstrated here. We refer the interested reader to \citet{liu2020} for further illustrative examples of where MMD with Gaussian kernels can have limited power. We also want to point out that new kernel-based two sample tests are being actively developed which might make them easier to use on such problems in the future.
%
}
%
\label{fig:mmd}
%
\end{figure}
\newpage
\clearpage
\begin{figure*}[h!]
\centering
%
\begin{subfigure}[htbp]{0.28\textwidth}
\includegraphics[trim=0 0 0 0,clip,width=\textwidth]{figs/correlations/gaussian_linear.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{0.28\textwidth}
\includegraphics[trim=0 0 0 0,clip,width=\textwidth]{figs/correlations/gaussian_linear_uniform.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{0.28\textwidth}
\includegraphics[trim=0 0 0 0,clip,width=\textwidth]{figs/correlations/bernoulli_glm.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{0.28\textwidth}
\includegraphics[trim=0 0 0 0,clip,width=\textwidth]{figs/correlations/bernoulli_glm_raw.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{0.28\textwidth}
\includegraphics[trim=0 0 0 0,clip,width=\textwidth]{figs/correlations/slcp.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{0.28\textwidth}
\includegraphics[trim=0 0 0 0,clip,width=\textwidth]{figs/correlations/slcp_distractors.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{0.28\textwidth}
\includegraphics[trim=0 0 0 0,clip,width=\textwidth]{figs/correlations/gaussian_mixture.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{0.28\textwidth}
\includegraphics[trim=0 0 0 0,clip,width=\textwidth]{figs/correlations/two_moons.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{0.28\textwidth}
\includegraphics[trim=0 0 0 0,clip,width=\textwidth]{figs/correlations/sir.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{0.28\textwidth}
\includegraphics[trim=0 0 0 0,clip,width=\textwidth]{figs/correlations/lotka_volterra.pdf}
\end{subfigure}
%
\caption{
{\bf Correlations between metrics for all tasks.} NLTP is the negative log probability of true parameters. Note that calculation of KSD was numerically unstable when calculating gradients for SLCP Distractors and Two Moons, resulting in correlation of zero for these tasks.
%
}
%
\label{fig:metrics_correlations}
%
\end{figure*}
\newpage
\clearpage
\section{Website}
\label{appendix:website}
The companion website (\link{https://sbi-benchmark.github.io}{sbi-benchmark.github.io}{}) allows interactive comparisons in terms of all metrics. It also allows inspection of posterior samples of all runs, which we found insightful when choosing hyperparameters and diagnosing implementation issues. Two screenshots are provided in \autoref{fig:website}. \\[-0.5cm]
\begin{figure*}[h!]
\centering
%
\includegraphics[width=0.85\textwidth]{appendix/screenshots/metrics.png}
%
\includegraphics[width=0.85\textwidth]{appendix/screenshots/posterior.png}
%
\caption{
{\bf Screenshots from the companion website}. Top: Classification accuracy (C2ST) for a subset of sequential algorithms on the SLCP task. Bottom: {\color{SNLE}{SNLE}}{} posterior on SLCP for $\mathbold{x}_o^{(1)}$ at 100k simulations.
%
}
%
\label{fig:website}
%
\end{figure*}
\section{Tasks}
\label{appendix:tasks}
\input{appendix/tasks/gaussian_linear}
\input{appendix/tasks/gaussian_linear_uniform}
\input{appendix/tasks/slcp}
\input{appendix/tasks/slcp_distractors}
\input{appendix/tasks/bernoulli_glm}
\input{appendix/tasks/bernoulli_glm_raw}
\input{appendix/tasks/gaussian_mixture}
\input{appendix/tasks/two_moons}
\input{appendix/tasks/sir}
\input{appendix/tasks/lotka_volterra}
\subsection[Sequential Monte Carlo Approximate Bayesian Computation (SMC-ABC)]{Sequential Monte Carlo Approximate Bayesian Computation ({\color{SABC}{SMC-ABC}}{})}
\begin{algorithm}
\SetAlgoLined
\vspace{1mm}
Set schedule $\mathbold{\epsilon}$ (including initial $\epsilon_0$), population indicator $t=0$, and population size $N$ \\ \vspace{1mm}
Initialize weights $W_0 = 1/N$ uniformly \\
Sample initial population $\{\mathbold{\theta}_0^{(i)}\}$ using rejection sampling with $\epsilon_0$ \\ \vspace{2mm}
\While{in simulation budget}{
Increase population indicator $t=t+1$ \\
Set particle indicator $i=0$ \\ \vspace{1mm}
\While{$i < N$}{
Sample $\mathbold{\theta}^{\prime}$ from previous population $\{\mathbold{\theta}_{t-1}^{(i)}\}$ with weights $\{W_{t-1}^{(i)}\}$\;
Perturb $\mathbold{\theta}^{\prime}$: $\mathbold{\theta}^{\prime\prime} \sim K_t(\mathbold{\theta} | \mathbold{\theta}^{\prime})$ \\
Simulate data $x^{\prime\prime}$ from $p(\mathbold{x}| \mathbold{\theta}^{\prime\prime})$ \\ \vspace{1mm}
\eIf{$d(\mathbold{x}^{\prime\prime}, \mathbold{x}_o) \leq \epsilon_t$}{
Set $\mathbold{\theta}_t^{(i)} = \mathbold{\theta}^{\prime\prime}$ and $W_t^{i} = \frac{p(\mathbold{\theta}_t^{(i)})}{\sum_{j=1}^{N} W_{t-1}^j K_t(\mathbold{\theta}_t^{(i)} | \mathbold{\theta}_{t-1}^j)}$ \\
Increase particle indicator $i=i+1$ \\
}{
reject $\mathbold{\theta}^{\prime\prime}$ \\
}
}
Normalize weights so that $\sum_i W_t^{(i)} = 1$ \\
}
\textbf{return} {Weighted samples $\{\mathbold{\theta}_t^{(i)}\}$ from $\hat{p}(\mathbold{\theta} | d(\mathbold{x}, \mathbold{x}_o) \leq \epsilon)$} \vspace{1mm}
\caption{Population Monte Carlo ABC (ABC-PMC) as in \cite{beaumont2009}}
\label{algorithm:sabc}
\end{algorithm}
Sequential Monte Carlo Approximate Bayesian Computation ({\color{SABC}{SMC-ABC}}{}) algorithms \citep{beaumont2002, marjoram2006, sisson2007, toni2009} are an extension of the classical rejection ABC approach, inspired by importance sampling and sequential Monte Carlo sampling. Central to {\color{SABC}{SMC-ABC}}{} is the idea to approach the final set of samples from the approximate posterior by constructing a series of intermediate sets of samples slowly approaching the final set through perturbations.
Several variants have been developed \citep[e.g., ][]{sisson2007, beaumont2009, toni2009, simola2020}. Here, we used the scheme ABC-PMC scheme of \citet{beaumont2009} and refer to it as {\color{SABC}{SMC-ABC}}{} in the manuscript. More formally, the description of the ABC-PMC algorithm is as follows: Given observed data $\mathbold{x}_o$, a prior $p(\mathbold{\theta})$ over parameters of a simulation-based model $p(\mathbold{x}| \mathbold{\theta})$, a distance measure $d(\mathbold{x}, \mathbold{x}_o)$, a schedule of acceptance thresholds $\epsilon_i$, and a kernel $K(\mathbold{\theta} | \mathbold{\theta}^{\prime})$ to perturb intermediate samples, weighted samples of the approximate posterior are obtained as described in \algo{sabc}.
{\color{SABC}{SMC-ABC}}{} can improve the sampling efficiency compared to {\color{ABC}{REJ-ABC}}{} and avoids severe inefficiencies due to a mismatch between initial sampling and the target distribution. However, it comes with more hyperparameters that can require careful tuning to the problem at hand, e.g., the choice of distance measure, kernel, and $\epsilon$-schedule. Like, {\color{ABC}{REJ-ABC}}{}, {\color{SABC}{SMC-ABC}}{} suffers from the curse of dimensionality.
For the benchmark, we considered the popular toolbox \texttt{pyABC} \citep{klinger2018}. Additionally, to fully understand the details of the {\color{SABC}{SMC-ABC}}{} approach, we also implemented our own version. In the main paper we report results obtained with our implementation because it yielded slightly better results. A careful comparison of the two approaches, and the optimization of hyperparameters like $\epsilon$-schedule, population size and perturbation kernel variance across different tasks are shown in \autoref{appendix:hyperparams}. After optimization, the crucial parameters of {\color{SABC}{SMC-ABC}}{} were set to: $l_2$-norm as distance metric, quantile-based epsilon decay with 0.2 quantile, population size 100 for simulation budgets 1k and 10k, population size 1000 for simulation budget 100k, Gaussian perturbation kernel with empirical covariance from previous population scaled by 0.5. We obtained 10k samples required for calculation of metrics as follows: If a population is not complete within the simulation budget we completed it with accepted particles from the last population and recalculated all weights. We then fitted a KDE on all those particles and sampled 10k samples from the KDE.
\newpage
\subsection[Neural Ratio Estimation (NRE)]{Neural Ratio Estimation ({\color{NRE}{NRE}}{})}
\label{appendix:algorithms:nre}
\begin{algorithm}
\SetAlgoLined
Set optimization criterion $l$ (e.g., BCE)\\
\For{$j = 1 : N$}{
Sample $\mathbold{\theta}_j \sim p(\mathbold{\theta})$\\
Sample $\mathbold{\theta}_j' \sim p(\mathbold{\theta})$\\
Simulate $\mathbold{x}_j \sim p(\mathbold{x} | \mathbold{\theta}_j)$\\
}
$\mathbold{\phi} \leftarrow \argmin l(d_{\mathbold{\phi}}(\mathbold{x}_n, \mathbold{\theta}_n), 1) + l(d_{\mathbold{\phi}}(\mathbold{x}_n, \mathbold{\theta}_n'), 0)$\\
Parameterize $d_{\mathbold{\phi}}(\mathbold{x}, \mathbold{\theta})$\\
\textbf{return} {Samples from $\hat{p}(\mathbold{\theta} | \mathbold{x}_o)$ via MCMC; $d_{\mathbold{\phi}}(\mathbold{x}, \mathbold{\theta})$} \vspace{1mm}
\caption{Single round Neural Ratio Estimation as in \cite{hermans2019}}
\label{algorithm:nre}
\end{algorithm}
Neural ratio estimation ({\color{NRE}{NRE}}{}) uses neural-network based classifiers to approximate the posterior $p(\mathbold{\theta} | \mathbold{x}_o)$. While neural-network based approaches described in the previous sections use \textit{density estimation} to either estimate the likelihood (({\color{SNLE}{S}}){\color{NLE}{NLE}}) or the posterior (({\color{SNPE}{S}}){\color{NPE}{NPE}}), NRE algorithms (({\color{SNRE}{S}}){\color{NRE}{NRE}}) use \textit{classification} to estimate a ratio of likelihoods. The ratio can then be used for posterior evaluation or MCMC-based sampling.
Likelihood ratio estimation can be used for SBI because it allows to perform MCMC without evaluating the intractable likelihood. In MCMC, the transition probability from a current parameter $\mathbold{\theta}_t$ to a proposed parameter $\mathbold{\theta}'$ depends on the posterior ratio and in turn on the likelihood ratio between the two parameters:
$$
\begin{aligned}
\frac{p(\mathbold{\theta}' | \mathbold{x})}{p(\mathbold{\theta}_t | \mathbold{x})} &= \frac{p(\mathbold{\theta}') p(\mathbold{x} | \mathbold{\theta}') / p(\mathbold{x})}{p(\mathbold{\theta}_t) p(\mathbold{x} | \mathbold{\theta}_t) / p(\mathbold{x})} = \frac{p(\mathbold{\theta}') p(\mathbold{x} | \mathbold{\theta}')}{p(\mathbold{\theta}_t) p(\mathbold{x} | \mathbold{\theta}_t)}.
\end{aligned}
$$
Therefore, given a ratio estimator $r(\mathbold{x} |\mathbold{\theta}', \mathbold{\theta}_t) = \frac{p(\mathbold{x} | \mathbold{\theta}')}{p(\mathbold{x} | \mathbold{\theta}_t)}$ learned from simulations, one can perform MCMC to obtain samples from the posterior, even if evaluating $p(\mathbold{x} | \mathbold{\theta})$ is intractable.
\cite{hermans2019} proposed the following approach for MCMC with classifiers to approximate density ratios: A classifier is trained to distinguish samples from an arbitrary $(\mathbold{\theta}, \mathbold{x}) \sim p(\mathbold{x} | \mathbold{\theta})p(\mathbold{\theta})$ and samples from the marginal model $(\mathbold{\theta}, \mathbold{x}) \sim p(\mathbold{\theta})p(\mathbold{x})$. This results in a likelihood-to-evidence estimator that needs to be trained only once to be evaluated for any $\mathbold{\theta}$. The training of the classifier $d_{\mathbold{\phi}}(\mathbold{x}, \mathbold{\theta})$ proceeds by minimizing the binary cross-entropy loss (BCE), as outlined in \algo{nre}. Once the classifier $d_{\mathbold{\phi}}(\mathbold{x}, \mathbold{\theta})$ is parameterized, it can be used to perform MCMC to obtain samples from the posterior. The authors name their approach \textit{Amortized Approximate Likelihood Ratio MCMC} (AALR-MCMC): It is amortized because once the likelihood ratio estimator is trained, it is possible to run MCMC for any $\mathbold{x} \sim p(\mathbold{x})$.
Earlier ratio estimation algorithms for SBI \citep[e.g.,][]{izbicki2014high,pham2014note,cranmer2015,dutta2016likelihood} and their connections to recent methods are discussed in \cite{thomas2020}, as well as in \cite{durkan2020}. AALR-MCMC is closely related to LFIRE \citep{dutta2016likelihood} but trains an amortized classifier rather than a separate one per posterior evaluation. \cite{durkan2020} showed that the loss of AALR-MCMC is closely related to the atomic SNPE-C/APT approach of \citet{greenberg2019} ({\color{SNPE}{SNPE}}{}) and that both can be combined in a unified framework. \cite{durkan2020} changed the formulation of the loss function for training the classifier from binary to multi-class.
For the benchmark, we used neural ratio estimation ({\color{NRE}{NRE}}{}) as formulated by \citet{durkan2020} and implemented in the \texttt{sbi} toolbox \citep{sbi}. As a classifier, we used a residual network architecture (ResNet) with two hidden layers of 50 units and ReLU non-linearity, trained with Adam \citep{kingma2014adam}. Following the notation of \cite{durkan2020}, we used $K=10$ as the size of the contrasting set. For the MCMC step, we followed the same procedure as described for {\color{NLE}{NLE}}{}, i.e., using Slice Sampling with 100 chains, to obtain 10k samples from each approximate posterior. In \autoref{appendix:hyperparams}, we show results for all tasks obtained with a multi-layer perceptron (MLP) architecture with two hidden layers of 50 ReLu units, and batch normalization.
\newpage
\subsection[Random Forest Approximate Bayesian Computation (RF-ABC)]{Random Forest Approximate Bayesian Computation ({\color{RFABC}{RF-ABC}}{})}
\label{appendix:algorithms:rfabc}
\begin{algorithm}
\SetAlgoLined
Set $\mathcal{D}=\{\}$
Set simulation budget $N$\\
Set number of trees $B$\\
Set minimum node size $N_{min}$
\For{$n = 1 : N$}{
Sample $\mathbold{\theta}_n \sim p(\mathbold{\theta})$ \\
Simulate $\mathbold{x}_n \sim p(\mathbold{x} | \mathbold{\theta}_n)$ \\
Add $(\mathbold{\theta}_n, \mathbold{x}_n)$ to $\mathcal{D}$ \\
}
Run random forest regression of $\mathbold{x}$ on $\mathbold{\theta}$ using $\mathcal{D}$, $B$ and $N_{min}$\\
\textbf{return} {$N$ samples $\{\mathbold{\theta}^{(i)}\}$ and associated weights $\{w^{(i)}\}$ for drawing approximate posterior samples} \vspace{1mm}
\caption{Random Forest ABC (RF-ABC) as in \citet{raynal2018}}
\label{algorithm:rfabc}
\end{algorithm}
Random forest Approximate Bayesian Computation \citep[{\color{RFABC}{RF-ABC}}{},][]{pudlo2016, raynal2018} is a more recently developed ABC algorithm based on a regression approach. Similar to previous regression approaches to ABC \citep{beaumont2002, blum2010}, {\color{RFABC}{RF-ABC}}{} aims at improving classical ABC inference ({\color{ABC}{REJ-ABC}}{}, {\color{SABC}{SMC-ABC}}{}) in the setting of high-dimensional data.
The idea of the {\color{RFABC}{RF-ABC}}{} algorithm is to use random forests \citep[RF, ][]{breiman2001} to run a non-parametric regression of a set of potential summary statistics of the data on the corresponding parameters. That is, the RF regression is trained on data simulated from the model, such that the covariates are the summary statistics and the response variable is a parameter. For a detailed description of the algorithm, we refer to \citet{raynal2018}.
The only hyperparameters for the {\color{RFABC}{RF-ABC}}{} algorithm are the number of trees and the minimum node size for the RF regression. Following \citet{raynal2018}, we chose the default of 500 trees and a minimum of 5 nodes. The output of the algorithm is a RF weight for each of the simulated parameters. This set of weights can be used to calculate posterior quantiles or to obtain an approximate posterior density as described in \citet{raynal2018}. We obtained 10k posterior samples for the benchmark by using the random forest weights to sample from the simulated parameters. We used the implementation in the \texttt{abcranger} toolbox \citet{collin2020}.
One important property of {\color{RFABC}{RF-ABC}}{} is that it can only be applied in the unidimensional setting, i.e., for 1-D dimensional parameter spaces, or for multidimensional parameters spaces with the assumption that the posterior factorizes over parameters (thus ignoring potential posterior correlations). This assumptions holds only for a few tasks in our benchmark (Gaussian Linear, Gaussian Linear Uniform, Gaussian Mixture). Due to this inherent limitation, we report {\color{RFABC}{RF-ABC}}{} in the supplement (see \suppfig{rf_abc}).
\newpage
\subsection[Neural Likelihood Estimation (NLE)]{Neural Likelihood Estimation ({\color{NLE}{NLE}}{})}
\label{appendix:algorithms:nle}
\begin{algorithm}
\SetAlgoLined
Set $\mathcal{D}=\{\}$
\For{$n = 1 : N$}{
Sample $\mathbold{\theta}_n \sim p(\mathbold{\theta})$ \\
Simulate $\mathbold{x}_n \sim p(\mathbold{x} | \mathbold{\theta}_n)$ \\
Add $(\mathbold{\theta}_n, \mathbold{x}_n)$ to $\mathcal{D}$ \\
}
Train $q_{\mathbold{\psi}}(\mathbold{x} | \mathbold{\theta})$ on $\mathcal{D}$\\
\textbf{return} {Samples from $\hat{p}(\mathbold{\theta} | \mathbold{x}_o) \propto q_{\mathbold{\psi}}(\mathbold{x}_o | \mathbold{\theta}) p(\mathbold{\theta})$ via MCMC; $q_{\mathbold{\psi}}(\mathbold{x} | \mathbold{\theta})$} \vspace{1mm}
\caption{Single round Neural Likelihood as in \cite{papamakarios2019a}}
\label{algorithm:nle}
\end{algorithm}
Likelihood estimation approaches to SBI use density estimation to approximate the likelihood $p(\mathbold{x}_o | \mathbold{\theta})$. After learning a surrogate $q_{\mathbold{\psi}}$ ($\mathbold{\psi}$ denoting the parameters of the estimator) for the likelihood function, one can for example use Markov Chain Monte Carlo (MCMC) based sampling algorithms to obtain samples from the approximate posterior $\hat{p}(\mathbold{\theta} | \mathbold{x}_o)$. This idea dates back to using Gaussian approximations of the likelihood \citep{wood2010, sisson2018_chapter12}, and more recently, was extended to density estimation with neural networks \citep{papamakarios2019a, lueckmann2019}.
We refer to the single-round version of the (sequential) neural likelihood approach by \citet{papamakarios2019a} as {\color{NLE}{NLE}}{}, and outline it in \algo{nle}: Given a set of samples $\{\mathbold{\theta}_n, \mathbold{x}_n\}_{1:N}$ obtained by sampling $\mathbold{\theta}_n \sim p(\mathbold{\theta})$ from the prior and simulating $\mathbold{x}_n \sim p(\mathbold{x} | \mathbold{\theta}_n)$, we train a conditional neural density estimator $q_{\mathbold{\psi}}(\mathbold{x} | \mathbold{\theta})$ modelling the conditional of data given parameters on the set $\{\mathbold{\theta}_n, \mathbold{x}_n\}_{1:N}$. Training proceeds by maximizing the log likelihood $\sum_n \log q_{\mathbold{\psi}}(\mathbold{x} | \mathbold{\theta})$. Given enough simulations, a sufficiently flexible conditional neural density estimator approximates the likelihood in the support of the prior $p(\mathbold{\theta})$ \citep{papamakarios2019a}. Once $q_{\mathbold{\psi}}$ is trained, samples from the approximate posterior $\hat{p}(\mathbold{\theta} | \mathbold{x}_o)$ are obtained using MCMC sampling based on the approximate likelihood $\hat{p}(\mathbold{x}_o | \mathbold{\theta})$ and the prior $p(\mathbold{\theta})$.
\phantomsection
\label{appendix:mcmc-sampling}
For MCMC sampling, \citet{papamakarios2019a} suggest to use Slice Sampling \citep{neal2003} with a single chain. However, we observed that the accuracy of the obtained posterior samples can be substantially improved by changing the Slice Sampling scheme as follows: 1) Instead of a single chain, we used 100 parallel MCMC chains; 2) for initialization of the chains, we sampled 10k candidate parameters from the prior, evaluated them under the unnormalized approximate posterior, and used these values as weights to resample initial locations; 3) we transformed parameters to be unbounded as suggested e.g.~in \citet{bingham2018, carpenter2017,hogg2017}. In addition, we reimplemented the slice sampler to allow vectorized evaluations of the likelihood, which yielded significant computational speed-ups.
For the benchmark, we used as density estimator a Masked Autoregressive Flow \citep[MAF, ][]{papamakarios2017} with five flow transforms, each with two blocks and 50 hidden units, $\tanh$ non-linearity and batch normalization after each layer. For the MCMC step, we used the scheme as outlined above with 250 warm-up steps and ten-fold thinning, to obtain 10k samples from the approximate posterior (1k samples from each chain). In \autoref{appendix:hyperparams} we show results for all tasks obtained with a Neural Spline Flow \citep[NSF, ][]{durkan2019neural} for density estimation, using five flow transforms, two residual blocks of 50 hidden units each, ReLU non-linearity, and 10 bins.
\newpage
\subsection[Sequential Neural Posterior Estimation (SNPE)]{Sequential Neural Posterior Estimation ({\color{SNPE}{SNPE}}{})}
\begin{algorithm}
\label{appendix:algorithms:snpe-alg}
\SetAlgoLined
Set $\tilde{p}_1(\mathbold{\theta}) = p(\mathbold{\theta})$\\
$c \leftarrow 0$ \\
\For{$r = 1 : R$}{
\For{$j = 1 : N$}{
$c \leftarrow c + 1$\\
Sample $\mathbold{\theta}_{c} \sim \tilde{p}_{r}(\mathbold{\theta})$\\
Simulate $\mathbold{x}_{c} \sim p(\mathbold{x} | \mathbold{\theta}_{c})$\\
}
$V_r(\Theta) := \begin{cases} \binom{c}{M}^{-1} & \text{ if } \Theta = \{\mathbold{\theta}_{b_1}, \mathbold{\theta}_{b_1}, \ldots,\mathbold{\theta}_{b_M}\} \text{ and } 1 \leq b_1 < b_2 < \ldots < b_M \leq c \\ 0 & \text{ otherwise} \end{cases}$\\
$\mathbold{\phi} \leftarrow \argmin_{\mathbold{\phi}} \mathbb{E}_{\Theta \sim V_r(\Theta)} \left[ \sum_{\mathbold{\theta}_j \in \Theta} - \log \tilde{q}_{\mathbold{x}_{j}, \mathbold{\phi}}(\mathbold{\theta}_{j}) \right]$\\
Set $\tilde{p}_{r+1}(\mathbold{\theta}) := q_{F(\mathbold{x}_o, \mathbold{\phi})}(\mathbold{\theta})$\\
}
\textbf{return} {Samples from $\hat{p}_R(\mathbold{\theta} | \mathbold{x}_o)$; $q_{F(x, \mathbold{\phi})}(\mathbold{\theta})$} \vspace{1mm}
\caption{Sequential Neural Posterior Estimation with atomic proposals \citep{greenberg2019}}
\label{algorithm:snpe}
\end{algorithm}
Sequential Neural Posterior Estimation {\color{SNPE}{SNPE}}{} is the sequential analog of {\color{NPE}{NPE}}{}, and meant to increase sample efficiency (see also \autoref{appendix:algorithms:snle}). When the posterior is targeted directly, using a proposal distribution $\tilde{p}(\mathbold{\theta})$ different from
the prior requires a correction step---without it, the posterior under the proposal distribution would be inferred \citep{papamakarios2016}. This so-called proposal posterior is denoted by $\tilde{p}(\mathbold{\theta} | \mathbold{x})$:
$$
\begin{aligned}
\label{appendix:algorithms:snpe:proposaleq}
\tilde{p}(\mathbold{\theta} | \mathbold{x}) &= p(\mathbold{\theta} | \mathbold{x}) \frac{\tilde{p}(\mathbold{\theta}) p(\mathbold{x})}{p(\mathbold{\theta}) \tilde{p}(\mathbold{x})},
\end{aligned}
$$
\vspace{-0.1cm}
where $\tilde{p}(\mathbold{x}) = \int_{\mathbold{\theta}}\tilde{p}(\mathbold{\theta}) p(\mathbold{x} | \mathbold{\theta})$. Note that for $\tilde{p}(\mathbold{\theta})=p(\mathbold{\theta})$, it directly follows that $\tilde{p}(\mathbold{\theta} | \mathbold{x}) = p(\mathbold{\theta} | \mathbold{x})$.
There have been three different approaches to this correction step so far, leading to three versions of SNPE \citep{papamakarios2016,lueckmann2017,greenberg2019}. All three algorithms have in common
that they train a neural network $F(\mathbold{x}, \mathbold{\phi})$ to learn the parameters of a family of densities $q_{\mathbold{\psi}}$ to estimate the posterior. They differ in what is targeted by $q_{\mathbold{\psi}}$ and which loss is used for $F$.
SNPE-A \citep{papamakarios2016} trains $F$ to target the proposal posterior $\tilde{p}(\mathbold{\theta} | \mathbold{x})$ by minimizing the log likelihood loss $-\sum_n \log q_{\mathbold{\psi}}(\mathbold{\theta}_n | \mathbold{x}_n)$, and then post-hoc solves for $p(\mathbold{\theta} | \mathbold{x})$. The analytical post-hoc step places restrictions on $q_{\mathbold{\psi}}$, the proposal, and prior. \cite{papamakarios2016} used Gaussian mixture density networks, single Gaussians proposals, and Gaussian or uniform priors. SNPE-B \citep{lueckmann2017} trains $F$ with the importance weighted loss $-\sum_n \frac{p(\mathbold{\theta}_n}{\tilde{p}(\mathbold{\theta}_n)} \log q_{\mathbold{\psi}}(\mathbold{\theta}_n | \mathbold{x}_n)$ to directly recover $p(\mathbold{\theta} | \mathbold{x})$ without the need for post-hoc correction, removing restrictions with respect to $q_{\mathbold{\psi}}$, the proposal, and prior. However, the importance weights can have high variance during training, leading to inaccurate inference for some tasks \citep{greenberg2019}. SNPE-C (APT) \citep{greenberg2019} alleviates this issue by reparameterizing the problem such that it can infer the posterior by maximizing an estimated proposal posterior. It trains $F$ to approximate $p(\mathbold{\theta}|\mathbold{x})$ with $q_{F(\mathbold{x}, \mathbold{\phi})}(\mathbold{\theta})$, using a loss defined on the approximate proposal posterior $\tilde{q}_{\mathbold{x}, \mathbold{\phi}}(\mathbold{\theta})$. \cite{greenberg2019} introduce `atomic' proposals to allow for arbitrary choices of the density estimator, e.g., flows \citep{papamakarios2019c}: The loss on $\tilde{q}_{\mathbold{x}, \mathbold{\phi}}(\mathbold{\theta})$ is calculated as the expectation over proposal sets $\Theta$ sampled from a so-called `hyperproposal' $V(\Theta)$ as outlined in \algo{snpe} \citep[see][for full details]{greenberg2019}.
For the benchmark, we used the approach by \citet{greenberg2019} with `atomic' proposals and referred to it as {\color{SNPE}{SNPE}}{}. As density estimator, we used a Neural Spline Flow \citep{durkan2019neural} with the same settings as for {\color{NPE}{NPE}}{}. For the `atomic' proposals, we used $M=10$ atoms (larger $M$ was too demanding in terms of memory). The simulation budget was equally split across 10 rounds and for the final round, we obtained 10k samples from the approximate posterior $\hat{p}_R(\mathbold{\theta} | \mathbold{x}_o)$. In \autoref{appendix:hyperparams}, we compare NSFs to Masked Autoregressive Flows \citep[MAFs, ][]{papamakarios2017}, as used in \citet{greenberg2019,durkan2020}, with five flow transforms, each with two blocks and 50 hidden units, $\tanh$ non-linearity and batch normalization after each layer.
\newpage
\subsection[Sequential Neural Ratio Estimation (SNRE)]{Sequential Neural Ratio Estimation ({\color{SNRE}{SNRE}}{})}
\label{appendix:algorithms:snre}
\begin{algorithm}
\SetAlgoLined
Set optimization criterion $l$ (e.g., BCE)\\
Set $\tilde{p}(\mathbold{\theta}) = p(\mathbold{\theta})$\\
\For{$r = 1 : R$}{
\For{$j = 1 : N$}{
Sample $\mathbold{\theta}_j \sim \tilde{p}(\mathbold{\theta})$ (via $d_{\mathbold{\phi}}$ and MCMC)\\
Sample $\mathbold{\theta}_j' \sim \tilde{p}(\mathbold{\theta})$ (via $d_{\mathbold{\phi}}$ and MCMC)\\
Simulate $\mathbold{x}_j \sim p(\mathbold{x} | \mathbold{\theta}_j)$\\
}
$\mathbold{\phi} \leftarrow \argmin l(I'n, \mathbold{\theta}_n), 1) + l(d_{\mathbold{\phi}}(\mathbold{x}_n, \mathbold{\theta}_n'), 0)$\;
Parameterize $d_{\mathbold{\phi}}(\mathbold{x}, \mathbold{\theta})$\\
}
\textbf{return} {Samples from $\hat{p}(\mathbold{\theta} | \mathbold{x}_o)$ via MCMC; $d_{\mathbold{\phi}}(\mathbold{x}, \mathbold{\theta})$} \vspace{1mm}
\caption{Sequential Neural Ratio Estimation as in \citet{hermans2019}}
\label{algorithm:snre}
\end{algorithm}
Sequential Neural Ratio Estimation ({\color{SNRE}{SNRE}}{}) is the sequential version of {\color{NRE}{NRE}}{}, and meant to increase sample efficiency, at the cost of needing to train new classifiers for different $\mathbold{x}_o$.
A sequential version of neural ratio estimation was proposed by \citet{hermans2019}. As with other sequential algorithms, the idea is to replace the prior by a proposal distribution $\tilde{p}(\mathbold{\theta})$ that is focused on $\mathbold{x}_o$ in the sense that the sampled parameters $\mathbold{\theta}$ result in simulated data $\mathbold{x}$ that are informative about $\mathbold{x}_o$. The proposal for the next round is the posterior estimate from the previous round. The ratio estimator then becomes $\tilde{r}(\mathbold{x}, \mathbold{\theta})$ and is refined over rounds by training the underlying classifier with positive examples $(\mathbold{x}, \mathbold{\theta}) \sim p(\mathbold{x} | \mathbold{\theta}) \tilde{p}(\mathbold{\theta})$ and negative examples $(\mathbold{x}, \mathbold{\theta}) \sim p(\mathbold{x}) \tilde{p}(\mathbold{\theta})$. Exact posterior evaluation is not possible anymore, but samples can be obtained as before via MCMC. These steps are outlined in \algo{snre}.
For the benchmark, we used {\color{SNRE}{SNRE}}{} as formulated by \citet{durkan2020} and implemented in the \texttt{sbi} toolbox \citep{sbi}. The classifier had the same architecture as described for {\color{NRE}{NRE}}{}. For the MCMC step, we followed the same procedure as described for {\color{NLE}{NLE}}{}. The simulation budget was equally split across 10 rounds. In \autoref{appendix:hyperparams}, we show results for all tasks obtained with a multi-layer perceptron (MLP) architecture with two hidden layers of 50 ReLu units, and batch normalization.
\newpage
\subsection[Sequential Neural Likelihood Estimation (SNLE)]{Sequential Neural Likelihood Estimation ({\color{SNLE}{SNLE}}{})}
\label{appendix:algorithms:snle}
\begin{algorithm}
\SetAlgoLined
Set $\hat{p}_0(\mathbold{\theta} | \mathbold{x}_o) = p(\mathbold{\theta})$ and $\mathcal{D}=\{\}$
\For{r = 1 : R}{
\For{$n = 1 : N$}{
Sample $\mathbold{\theta}_n \sim \hat{p}_{r-1}(\mathbold{\theta} | \mathbold{x}_o)$ with MCMC \\
Simulate $\mathbold{x}_n \sim p(\mathbold{x} | \mathbold{\theta}_n)$\\
Add $(\mathbold{\theta}_n, \mathbold{x}_n)$ to $\mathcal{D}$\\
}
(Re-)train $q_{\mathbold{\psi}}(\mathbold{x} | \mathbold{\theta})$ on $\mathcal{D}$\\
Set $\hat{p}_r(\mathbold{\theta} | \mathbold{x}_o) \propto q_{\mathbold{\psi}}(\mathbold{x}_o | \mathbold{\theta}) p(\mathbold{\theta})$\\
}
\textbf{return} {Samples from $\hat{p}(\mathbold{\theta} | \mathbold{x}_o) \propto q_{\mathbold{\psi}}(\mathbold{x}_o | \mathbold{\theta}) p(\mathbold{\theta})$ via MCMC; $q_{\mathbold{\psi}}(\mathbold{x} | \mathbold{\theta})$} \vspace{1mm}
\caption{Sequential Neural Likelihood as in \cite{papamakarios2019a}}
\label{algorithm:snle}
\end{algorithm}
Sequential Neural Likelihood estimation \citep[SNLE or SNL, ][]{papamakarios2019a} extends the neural likelihood estimation approach described in the previous section to be sequential.
The idea behind sequential SBI algorithms is based on the following intuition: If for a particular inference problem, there is only a single $\mathbold{x}_o$ one is interested in, then simulating data using parameters from the entire prior space might be inefficient, leading to a training set $\mathcal{D}$ that contains training data $(\mathbold{\theta}, \mathbold{x})$ which carries little information about the posterior $p(\mathbold{\theta} | \mathbold{x}_o)$. Instead, to increase sample efficiency, one may draw training data points from a proposal distribution $\tilde{p}(\mathbold{\theta})$, ideally obtaining $\mathbold{\theta}$ for which $\mathbold{x}$ is close to $\mathbold{x}_o$. One candidate that has been commonly used in the literature for such a proposal is the approximate posterior distribution itself.
{\color{SNLE}{SNLE}}{} is a multi-round version of {\color{NLE}{NLE}}{}, where in each round new training samples are drawn from a proposal $\tilde{p}(\mathbold{\theta})$. The proposal is chosen to be the posterior estimate at $\mathbold{x}_o$ from the previous round $\hat{p}(\mathbold{\theta} | \mathbold{x}_o)$ and its samples are obtained using MCMC. The proposal controls where $q_{\mathbold{\psi}}(\mathbold{x} | \mathbold{\theta})$ is learned most accurately. Thus, by iterating over multiple rounds, a good approximation to the posterior can be learned more efficiently than by sampling all training data from the prior. {\color{SNLE}{SNLE}}{} is summarized in \algo{snle}.
For the benchmark, we used as density estimator a Masked Autoregressive Flow \citep{papamakarios2017}, and MCMC to obtain posterior samples after every round, both with the same settings as described for {\color{NLE}{NLE}}{}. The simulation budget was equally split across 10 rounds. In \autoref{appendix:hyperparams}, we show results for all tasks obtained with a Neural Spline Flow \citep[NSF, ][]{durkan2019neural} for density estimation, using five flow transforms, two residual blocks of 50 hidden units each, ReLU non-linearity, and 10 bins.
\newpage
\subsection[Synthetic Likelihood (SL)]{Synthetic Likelihood ({\color{SL}{SL}}{})}
\label{appendix:algorithms:sl}
\begin{algorithm}
\SetAlgoLined
Set number of simulations per step $M$\\
Set number of MCMC steps $T$\\
\For{$t=1:T$}{
Get new candidate $\mathbold{\theta}_t$ from MCMC scheme \\
Set $\mathcal{D}_t=\{\}$ \\
\For{$m=1 : M$}{
Simulate $\mathbold{x}_m \sim p(\mathbold{x} | \mathbold{\theta}_t)$ \\
Add $(\mathbold{\theta}_t, \mathbold{x}_m)$ to $\mathcal{D}_t$ \\
}
Use $\mathcal{D}_t$ to estimate mean and covariance of a Gaussian approximation of the likelihood $\hat{L}(\mathbold{x}_o | \theta_t)$ \\
Perform the next MCMC step using $\hat{L}(\mathbold{x}_o | \theta_t)$
}
\textbf{return} {$N$ samples $\{\mathbold{\theta}^{(i)}\}$ from MCMC chain} \vspace{1mm}
\caption{Synthetic Likelihood algorithm as in \citet{wood2010}}
\label{algorithm:sl}
\end{algorithm}
The Synthetic Likelihood ({\color{SL}{SL}}{}) approach circumvents the evaluation of the intractable likelihood by estimating a \textit{synthetic} one from simulated data or summary statistics. This approach was introduced by \citet{wood2010}. Its main motivation is that the classical ABC approach of comparing simulated and observed data with a distance metric can be problematic if parts of the differences are entirely noise-driven. \citet{wood2010} instead approximated the distribution of the summary statistics (the likelihood) of a nonlinear ecological dynamic system as a Gaussian distribution, thereby capturing the underlying noise as well. The approximation of the likelihood can then be used to obtain posterior sampling via Markov Chain Monte Carlo (MCMC) \citep{wood2010}.
The {\color{SL}{SL}}{} approach can be seen as the predecessor of the ({\color{SNLE}{S}}){\color{NLE}{NLE}}{} approaches: They replaced the Gaussian approximation of the likelihood with a much more flexible one that uses neural networks and normalizing flows (see \ref{appendix:algorithms:nle}). Moreover, there are modern approaches from the classical ABC field that further developed {\color{SL}{SL}}{} using a Gaussian approximation \citep[e.g.,][]{sisson2018_chapter12,priddle2019}.
For the benchmark, we implemented our own version of the algorithm proposed by \citet{wood2010}. We used Slice Sampling MCMC \citep{neal2003} and estimated the Gaussian likelihood from 100 samples at each sampling step. To ensure a positive definite covariance matrix, we added a small value $\epsilon$ to the diagonal of the estimated covariance matrix for some of the tasks. In particular, we used $\epsilon=0.01$ for SIR and Bernoulli GLM Raw tasks, and we tried without success $\epsilon=[0, 0.01, 0.1, 1.0]$ for Lotka-Volterra and SLCP with distractors. For all remaining tasks, we set $\epsilon=0$. For Slice Sampling, we used a single chain initialized with sequential importance sampling (SIR) as described for {\color{NLE}{NLE}}{}, 1k warm-up steps and no thinning, in order to keep the number of required simulations tractable. This resulted in an overall simulation budget on the order of $10^8$ to $10^9$ simulations per run in order to generate 10k posterior samples, as new simulations are required for every MCMC step.
The high simulation budget makes it problematic to directly compare {\color{SL}{SL}}{} and other other algorithms in the benchmark. Therefore, we report {\color{SL}{SL}}{} in the supplement (see \suppfig{sl}).
\newpage
\clearpage
\subsection[Neural Posterior Estimation (NPE)]{Neural Posterior Estimation ({\color{NPE}{NPE}}{})}
\label{apppendix:algorithms:npe}
\begin{algorithm}
\SetAlgoLined
\For{$j = 1 : N$}{
Sample $\mathbold{\theta}_j \sim p(\mathbold{\theta})$\\
Simulate $\mathbold{x}_j \sim p(\mathbold{x} | \mathbold{\theta}_j)$\\
}
$\mathbold{\phi} \leftarrow \argmin \sum_j^N - \log q_{F(\mathbold{x}_{j}, \mathbold{\phi})}(\mathbold{\theta}_{j})$\\
Set $\hat{p}(\mathbold{\theta} | \mathbold{x}_o) = q_{F(\mathbold{x}_o, \mathbold{\phi})}(\mathbold{\theta})$\\
\textbf{return} {Samples from $\hat{p}(\mathbold{\theta} | \mathbold{x}_o)$; $q_{F(\mathbold{x}, \mathbold{\phi})}(\mathbold{\theta})$} \vspace{1mm}
\caption{Single round Neural Posterior Estimation as in \cite{papamakarios2016}}
\label{algorithm:npe}
\end{algorithm}
{\color{NPE}{NPE}}{} uses conditional density estimation to directly estimate the posterior. This idea dates back to regression adjustment approaches \citep{blum2010} and was extended to density estimators using neural networks \citep{papamakarios2016} more recently.
As outlined in \algo{npe}, the approach is as follows: Given a prior over parameters $p(\mathbold{\theta})$ and a simulator, a set of training data points
$(\mathbold{\theta}, \mathbold{x})$ is generated. This training data is used to learn the parameters $\mathbold{\psi}$ of a conditional density estimator $q_{\mathbold{\psi}}(\mathbold{\theta} | x)$ using a neural network $F(\mathbold{x}, \mathbold{\phi})$, i.e., $\mathbold{\psi}=F(\mathbold{x}, \mathbold{\phi})$. The loss function is given by the negative log probability $-\log q_{\mathbold{\psi}}(\mathbold{\theta} | x)$. If the density
estimator $q$ is flexible enough and training data is infinite, this loss function
leads to perfect recovery of the ground-truth posterior \citep{papamakarios2016}.
For the benchmark, we used the approach by \citet{papamakarios2016} with a Neural Spline Flow \citep[NSF, ][]{durkan2019neural} as density estimator, using five flow transforms, two residual blocks of 50 hidden units each, ReLU non-linearity, and 10 bins. We sampled 10k samples from the approximate posterior $q_{F(\mathbold{x}_o, \mathbold{\phi})}(\mathbold{\theta})$. In \autoref{appendix:hyperparams}, we compare NSFs to Masked Autoregressive Flows \citep[MAFs, ][]{papamakarios2017}, as used in \cite{greenberg2019,durkan2020}, with five flow transforms, each with two blocks and 50 hidden units, $\tanh$ non-linearity and batch normalization after each layer.
\newpage
\subsection[Rejection Approximate Bayesian Computation (REJ-ABC)]{Rejection Approximate Bayesian Computation ({\color{ABC}{REJ-ABC}}{})}
\label{appendix:algorithms:mcabc}
\begin{algorithm}
\SetAlgoLined
\While{in simulation budget}{
Sample $\mathbold{\theta}^\prime$ from $p(\mathbold{\theta})$
Simulate data $\mathbold{x}^\prime$ from $p(\mathbold{x} | \mathbold{\theta}^\prime)$
\eIf{$d(\mathbold{x}^\prime, \mathbold{x}_o) \leq \epsilon$}{
Accept $\mathbold{\theta}^\prime$
}{
Reject $\mathbold{\theta}^\prime$
}
}
\textbf{return} {Accepted samples $\{\mathbold{\theta}^\prime\}$ from $\hat{p}(\mathbold{\theta} | d(\mathbold{x}, \mathbold{x}_o) \leq \epsilon$)}
\caption{Rejection ABC}
\label{algorithm:abc}
\end{algorithm}
Classical Approximate Bayesian Computation (ABC) is based on Monte Carlo rejection sampling \citep{tavere1997, pritchard1999}: In rejection ABC, the evaluation of the likelihood is replaced by a comparison between observed data $\mathbold{x}_o$ and simulated data $\mathbold{x}$, based on a distance measure $d(\mathbold{x}, \mathbold{x}_o)$. Samples $\mathbold{\theta}$ from the approximate posterior are obtained by collecting simulation parameters that result in simulated data that is close to the observed data.
More formally, given observed data $\mathbold{x}_o$, a prior $p(\mathbold{\theta})$ over parameters of simulation-based model $p(\mathbold{x} | \mathbold{\theta})$, a distance measure $d(\mathbold{x}, \mathbold{x}_o)$ and an acceptance threshold $\epsilon$, rejection ABC obtains parameter samples $\mathbold{\theta}$ from the approximate posterior as outlined in \algo{abc}.
In theory, rejection ABC obtains samples from the true posterior $p(\mathbold{\theta} | \mathbold{x}_o)$ in the limit $\epsilon \to 0$ and $N \to \infty$, where $N$ is the simulation budget. In practice, its accuracy depends on the trade-off between simulation budget and the rejection criterion $\epsilon$. Rejection ABC suffers from the curse of dimensionality, i.e., with linear increase in the dimensionality of $\mathbold{x}$, an exponential increase in simulation budget is required to maintain accurate results.
For the benchmark, we did not use a fixed $\epsilon$-threshold, but quantile-based rejection. Depending on the simulation budget (1k, 10k, 100k), we used a quantile of (0.1, 0.01, or, 0.001), so that {\color{ABC}{REJ-ABC}}{} returned 100 samples with smallest distance to $\mathbold{x}_o$ in each of these cases (see \autoref{appendix:hyperparams} for different hyperparameter choices). In order to compute metrics on 10k samples, we sampled from a KDE fitted on the accepted parameters (details about KDE resampling in \autoref{appendix:hyperparams}). {\color{ABC}{REJ-ABC}}{} requires the choice of the distance measure $d(\mathbold{x}, \mathbold{x}_o)$: here we used the $l_2$-norm.
\newpage
\section{Runtimes}
\label{appendix:runtimes}
In applications of SBI, simulations are commonly assumed to be the dominant cost. In order to make the benchmark feasible at this scale, we focused on simple simulators and optimized runtimes, e.g. we developed a new package bridging \texttt{DifferentialEquations.jl} \citep{rackauckas2017,bezanson2017julia} and \texttt{PyTorch} \citep{paszke2019} so that generating simulations for all implemented tasks is extremely fast. This differs from many cases in practice, where the runtime costs for an algorithm are often negligible compared to the cost of simulations. Having said that, algorithms show significant differences in runtime costs, which we measured and report here.
We recorded runtimes for all algorithms on all tasks. In principle, runtimes could be reduced by employing multi-CPU architectures, however, we decided for the single CPU setup to accurately compare runtimes across all algorithms and tasks. We did not employ GPUs for training neural-networks (NN). This is because the type of NNs used in the algorithms currently in the benchmark do not benefit much from GPU versus CPU training (e.g., no CNN architecture, rather shallow and narrow networks). In fact, running {\color{SNPE}{SNPE}}{} on SLCP using a GeForce GTX 1080 showed slightly longer runtimes than on CPU, due to the added overhead resulting from copying data back and forth to the device. Therefore, it was more economical and comparable to run the benchmark on CPUs.
All neural network-based algorithms were run on single 3.6 GHz CPU cores of AWS C5-instances. ABC algorithms were run on single CPU cores of an internal cluster with 2.4 GHz CPUs. We observed a difference in runtimes of less than 100ms when running ABC algorithms on the same hardware as used for neural network-based algorithms.
Figure \ref{fig:metrics_runtime} shows the recorded runtimes in minutes. We observed short runtimes for {\color{ABC}{REJ-ABC}}{} and {\color{SABC}{SMC-ABC}}{}, as these do not require NN training or MCMC. The sequential versions of all three NN-based algorithms yielded longer runtimes than the non-sequential versions because these involve 10 rounds of NN training. Among the sequential algorithms, {\color{SNPE}{SNPE}}{} showed the longest runtimes. Runtimes with MAFs instead of NSFs tend to be faster, e.g.~the difference between MAFs and NSFs using {\color{SNPE}{SNPE}}{} on SLCP at 100k simulations was about 50 minutes on average. We also emphasize that the speed of ({\color{SNLE}{S}}){\color{NLE}{NLE}}{} reported here was only obtained after vectorizing MCMC sampling. Without vectorization, runtime on the Gaussian Linear for {\color{SNLE}{SNLE}}{} was more than 36 hours instead of less than 2 hours (see \autoref{appendix:hyperparams}).
\begin{figure*}[h!]
\centering
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 3,clip,width=\textwidth]{figs/main_paper/gaussian_linear_gaussian_linear_uniform_RT.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/main_paper/slcp_slcp_distractors_RT_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/main_paper/bernoulli_glm_bernoulli_glm_raw_RT_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/main_paper/gaussian_mixture_RT_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/main_paper/two_moons_RT_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/main_paper/sir_RT_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 10 0 0,clip,width=\textwidth]{figs/main_paper/lotka_volterra_RT_no_labels.pdf}
\end{subfigure}
%
\caption{
{\bf Runtime on benchmark tasks.} Runtime of {\color{ABC}{REJ-ABC}}{}, {\color{SABC}{SMC-ABC}}{}, {\color{NLE}{NLE}}{}, {\color{SNLE}{SNLE}}{}, {\color{NPE}{NPE}}{}, {\color{SNPE}{SNPE}}{}, {\color{NRE}{NRE}}{}, {\color{SNRE}{SNRE}}{} in minutes, for 10 observations each, means and 95\% confidence intervals. Each run was allocated a single CPU core, see \autoref{appendix:runtimes} for details.
%
}
%
\label{fig:metrics_runtime}
%
\end{figure*}
\newpage
\clearpage
\section{Metrics}
\label{appendix:metrics}
\input{appendix/metrics/true_parameters}
\input{appendix/metrics/simulation_based_calibration}
\input{appendix/metrics/median_distance}
\input{appendix/metrics/mmd}
\input{appendix/metrics/c2st}
\input{appendix/metrics/ksd}
\section{Algorithms}
\label{appendix:algorithms}
\input{appendix/algorithms/rej_abc}
\input{appendix/algorithms/smc_abc}
\input{appendix/algorithms/nle}
\input{appendix/algorithms/snle}
\input{appendix/algorithms/npe}
\input{appendix/algorithms/snpe}
\input{appendix/algorithms/nre}
\input{appendix/algorithms/snre}
\input{appendix/algorithms/rf_abc}
\input{appendix/algorithms/sl}
\section{Benchmark}
\label{appendix:benchmark}
\subsection{Reference posteriors}
We generated 10k reference posterior samples for each observation. For the \nameref{appendix:task:gaussian_linear} task, reference samples were obtained by using the analytic solution for the true posterior. Similarly, for \nameref{appendix:task:gaussian_linear_uniform} and \nameref{appendix:task:gaussian_mixture}, the analytic solution was used, combined with an additional rejection step, in order to account for the bounded support of the posterior due to the use of a uniform prior. For the \nameref{appendix:task:two_moons} task, we devised a custom scheme based on the model equations, which samples both modes and rejects samples outside the prior bounds.
For \nameref{appendix:task:slcp}, \nameref{appendix:task:sir}, and \nameref{appendix:task:lotka_volterra}, we devised a likelihood-based procedure to ensure obtaining a valid set of reference posterior samples: First, we either used Sampling/Importance Resampling \citep{rubin1988using} (for \nameref{appendix:task:slcp}, \nameref{appendix:task:sir}) or Slice Sampling MCMC \citep{neal2003} (for \nameref{appendix:task:lotka_volterra}) to obtain a set of 10k proposal samples from the unnormalized posterior $f(\mathbold{\theta}) = \tilde{p}(\mathbold{\theta}|\mathbold{x}_o) = p(\mathbold{x}_o|\mathbold{\theta}) p(\mathbold{\theta})$. We used these proposal samples to train a density estimator, for which we used a neural spline flow (NSF) \citep{durkan2019neural}. Next, we created a mixture composed of the NSF and the prior with weights 0.9 and 0.1, respectively, as a proposal distribution $g(\mathbold{\theta})$ for rejection sampling \citep{martino2018accept}. Rejection sampling relies on finding a constant $M$ such that $f(\mathbold{\theta}) \leq M g(\mathbold{\theta})$ for all values of $\mathbold{\theta}$: To find this constant, we initialized $M=1$, sampled $\mathbold{\theta} \sim g(\mathbold{\theta})$, and updated $M=1.2 f(\mathbold{\theta}) / g(\mathbold{\theta})$ if $f(\mathbold{\theta}) / g(\mathbold{\theta}) > M$. This loop stopped only after at least 100k samples without updating $M$ were reached. We then used $M$, $f$, and $g$ to generate 10k reference posterior samples. We found that the NSF-based proposal distribution resulted in high acceptance rates. We used this custom scheme rather than relying on MCMC directly, since we found that standard MCMC approaches (Slice Sampling, HMC, and NUTS) all struggled with multi-modal posteriors and wanted to avoid bias in the reference samples, e.g.~due to correlations in MCMC chains.
As a sanity check, we ran this scheme twice on all tasks and observation and found that the resulting reference posterior samples were indistinguishable in terms of C2ST.
\subsection{Code}
We provide \texttt{sbibm}, a benchmarking framework that implements all tasks, reference posteriors, different metrics and tooling to run and analyse benchmark results at scale. The framework is available at:
\link{https://github.com/}{github.com/anonymized}{}
We make benchmarking new algorithms maximally easy by providing an open, modular framework for \textit{integration} with SBI toolboxes. We here evaluated algorithms implemented in \texttt{pyABC} \citep{klinger2018}, \texttt{pyabcranger} \citep{collin2020}, and \texttt{sbi} \citep{sbi}. We emphasize that the goal of \texttt{sbibm} is orthogonal to any toolbox: It could easily be used with other toolboxes, or even be used to compare results for the same algorithm implemented by different ones. There are currently several SBI toolboxes available or under active development. \texttt{elfi} \citep{elfi2018} is a general purpose toolbox, including ABC algorithms as well as BOLFI \citep{gutmann2015}. There are many toolboxes for ABC algorithms, e.g., \texttt{abcpy} \citep{abcpy-repo}, \texttt{astroABC} \citep{jennings2017astroabc}, \texttt{CosmoABC} \citep{ishida2015cosmoabc}, see also \citet{sisson2018_chapter13} for an overview. \texttt{carl} \citep{louppe2016} implements the algorithm by \citet{cranmer2015}. \texttt{hypothesis} \citep{hypothesis-repo}, and \texttt{pydelfi} \citep{pydelfi-repo} are SBI toolboxes under development.
\subsection{Reproducibility}
To ensure reproducibility of our results, we publicly released all code including instructions on how to run the benchmark on cloud-based infrastructure.
\section{Hyperparameter Choices}
\label{appendix:hyperparams}
In this section, we address two central questions for any benchmark: (1) how hyperparameters are chosen and (2) how sensitive results are to the respective choices.
Rather than tuning hyperparameters on a per-task basis, we changed hyperparameters on multiple or all tasks at once and selected configurations that worked best across tasks. We wanted to avoid overfitting on individual benchmark tasks and were instead interested in settings that can generalize across multiple tasks. In practice, tuning an algorithm on a given task would typically be impossible, due to the lack of suitable metrics that can be computed without reference posteriors as well as high computational demands that SBI tasks often have.
To find good general settings, we performed more than 10 000 individual runs. We explored hyperparameter choices that have not been previously reported, and revealed substantial improvements. The benchmark offers the possibility to systematically compare different choices and design better and more robust SBI algorithms.
\subsection[REJ-ABC]{{\color{ABC}{REJ-ABC}}{}}
Classical ABC algorithms have crucial hyperparameters, most importantly, the distance metric and acceptance tolerance $\epsilon$. We used our own implementation of {\color{ABC}{REJ-ABC}}{} as it is straightforward to implement (see \ref{appendix:algorithms:mcabc}). The distance metric was fixed to be the $l_2$-norm for all tasks and we varied different acceptance tolerances $\epsilon$ across tasks on which {\color{ABC}{REJ-ABC}}{} performed sufficiently well. Our implementation of {\color{ABC}{REJ-ABC}}{} is quantile based, i.e,. we select a quantile of the samples with the smallest distance to the observed data, which implicitly defines an $\epsilon$. The 10k samples needed for the comparison to the reference posterior samples are then resampled from the selected samples. In order to check whether this resampling significantly impaired performance, we alternatively fit a KDE in order to obtain 10k samples.
Below, we show results for different schedules of quantiles for each simulation budget, e.g., a schedule of 0.1, 0.01, 0.001 corresponds to the 10, 1 and 0.1 percent quantile, or the top 100 samples for each simulation budget. Across tasks and budgets the 0.1, 0.01, 0.001 quantile schedule performed best (\autoref{fig:rej_abc_sweep}). Performance showed improvement by the KDE fit, especially on the Gaussian tasks. We therefore report the version using the top 100 samples and KDE in the main paper.
\begin{figure*}[h!]
\centering
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 3,clip,width=\textwidth]{figs/hyperparameters_rej_abc/gaussian_linear_gaussian_linear_uniform_C2ST.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_rej_abc/slcp_slcp_distractors_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_rej_abc/bernoulli_glm_bernoulli_glm_raw_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_rej_abc/gaussian_mixture_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_rej_abc/two_moons_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_rej_abc/sir_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 10 0 0,clip,width=\textwidth]{figs/hyperparameters_rej_abc/lotka_volterra_C2ST_no_labels.pdf}
\end{subfigure}
%
\caption{
{\bf Hyperparameter selection for {\color{ABC}{REJ-ABC}}{}.} C2ST performance of different percentile schedules across simulation budgets (columns) for all tasks (rows). Top label for each plot column: number of samples retained, and optional KDE. Across tasks and budgets, the schedule of 0.1, 0.01, 0.001 percentiles, which corresponds to the top 100 samples closest to the observation, performed best. Each data point corresponds to the mean and 95\% confidence interval across 10 observations.
%
}
%
\label{fig:rej_abc_sweep}
%
\end{figure*}
\newpage
\clearpage
\subsection[SMC-ABC]{{\color{SABC}{SMC-ABC}}{}}
{\color{SABC}{SMC-ABC}}{} has several hyperparameters including the population size, the perturbation kernel, the epsilon schedule and the distance metric. In order to ensure that we report the best possible {\color{SABC}{SMC-ABC}}{} results for a fair comparison, we sweeped over three hyperparameters that are especially critical: the population size, the quantile used to select the epsilon from the distances of the particles of the previous population, and the scaling factor of the covariance of the Gaussian perturbation kernel. The remaining hyperparameters were fixed to values common in the literature: Gaussian perturbation kernel and $l2$-norm distance metric.
Additionally, we compared our implementation against one from the popular pyABC toolbox \citep{klinger2018} to which we refer as versions A and B respectively. We sweeped over these hyperparameters and optionally added a post-hoc KDE fit for drawing the samples needed for two-sample based performance metrics.
Overall, the parameter setting with a population size of 100, a kernel covariance scale of 0.5, and an epsilon quantile 0.2 performed best. Although the results of the two different implementations were qualitatively very similar (compare \autoref{fig:smc_abc_sweep_sbi} and \autoref{fig:smc_abc_sweep_pyabc}, respectively), version A was slightly better on the Gaussian tasks. Although we tried to match the implementations and the exact settings, there are small differences between the two, which might explain the difference in the results: Implementation B constructs the Gaussian perturbation kernel using kernel density estimation on the weighted samples of the previous population, whereas A constructs it using the mean and covariance estimated from samples from the previous population. The latter could be advantageous in case of a Gaussian-like (high-dimensional) posterior (Gaussian Mixture and Gaussian linear task) and disadvantageous in a non-Gaussian-like posteriors (e.g., Two Moons). We decided to report results for {\color{SABC}{SMC-ABC}}{} in the main paper using implementation A (ours) with population size 100 for simulation budgets 1k and 10k, and population size 1000 for simulation budget 100k, a kernel covariance scale of 0.5, and epsilon quantile 0.2. This choice of kernel covariance scale is different from recommendations in the literature \citep{sisson2007, beaumont2009}. We only found very small performance differences for different scales and note that our choice is in line with the recommendation of the \texttt{pyABC} toolbox \citep{pyabcKernel}, i.e., using a scale between 0 and 1. Performance showed improvement by the KDE fit, especially on the Gaussian tasks. We therefore report the version with KDE in the main paper.
\newpage
\clearpage
\begin{figure*}[h!]
\centering
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 130,clip,width=\textwidth]{figs/hyperparameters_smc_abc_ours/gaussian_linear_gaussian_linear_uniform_C2ST.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_smc_abc_ours/slcp_slcp_distractors_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_smc_abc_ours/bernoulli_glm_bernoulli_glm_raw_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_smc_abc_ours/gaussian_mixture_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_smc_abc_ours/two_moons_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_smc_abc_ours/sir_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 10 0 0,clip,width=\textwidth]{figs/hyperparameters_smc_abc_ours/lotka_volterra_C2ST_no_labels.pdf}
\end{subfigure}
%
\caption{
{\bf Hyperparameter selection for {\color{SABC}{SMC-ABC}}{} with our implementation}. Top label for each plot column: population size, kernel covariance scale, epsilon quantile/epsilon-decay parameter, and optional KDE. Each data point corresponds to the mean and 95\% confidence interval across 10 observations.
%
}
%
\label{fig:smc_abc_sweep_sbi}
%
\end{figure*}
\newpage
\clearpage
\begin{figure*}[h!]
\centering
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 30,clip,width=\textwidth]{figs/hyperparameters_smc_abc_pyabc/gaussian_linear_gaussian_linear_uniform_C2ST.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_smc_abc_pyabc/slcp_slcp_distractors_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_smc_abc_pyabc/bernoulli_glm_bernoulli_glm_raw_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_smc_abc_pyabc/gaussian_mixture_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_smc_abc_pyabc/two_moons_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_smc_abc_pyabc/sir_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 10 0 0,clip,width=\textwidth]{figs/hyperparameters_smc_abc_pyabc/lotka_volterra_C2ST_no_labels.pdf}
\end{subfigure}
%
\caption{
{\bf Hyperparameter selection for {\color{SABC}{SMC-ABC}}{}. with \texttt{pyABC} implementation}. Top label for each plot column: population size, kernel covariance scale, epsilon quantile/epsilon-decay parameter, and optional KDE. Each data point corresponds to the mean and 95\% confidence interval across 10 observations.
%
}
%
\label{fig:smc_abc_sweep_pyabc}
%
\end{figure*}
\newpage
\clearpage
\subsection[MCMC for (S)NLE and (S)NRE]{MCMC for ({\color{SNLE}{S}}){\color{NLE}{NLE}}{} and ({\color{SNRE}{S}}){\color{NRE}{NRE}}{}}
({\color{SNLE}{S}}){\color{NLE}{NLE}}{} and ({\color{SNRE}{S}}){\color{NRE}{NRE}}{} both rely on MCMC sampling, which has several hyperparameters. In line with \citet{papamakarios2019a} and \citet{durkan2020}, we used Slice Sampling \citep{neal2003}. However, we modified the MCMC schemes used in these papers and obtained significant improvements in performance and speed.
\textbf{Number of chains and initialization.} While \citet{papamakarios2019a,durkan2020} used a single chain with axis-aligned updates, we found that on tasks with multi-modal posteriors, it can be essential to run multiple MCMC chains in order to sample all modes. Performance on Two Moons, for example, was poor with a single chain, since usually only one of the crescent shapes was sampled. Rather than initialising chains by drawing initial locations from the prior, we found the resampling scheme as described in \ref{appendix:algorithms:nle} to work better for initialisation, and used 100 chains instead of a single one.
\textbf{Transformation of variables.} When implementing MCMC, it is common advice to transform problems to have unbounded support \citep{hogg2017}, although this has not been discussed in SBI papers or implemented in accompanying code. We found that without this transformation, MCMC sampling could get stuck in endless loops, e.g., on the Lotka-Volterra task. Apart from the transformation to unbounded space, we found z-scoring of parameters and data to be crucial for some tasks.
\textbf{Vectorization of MCMC sampling}. We reimplemented Slice Sampling so that all chains could perform likelihood evaluations in parallel. Evaluating likelihoods, e.g., in the case of ({\color{SNLE}{S}}){\color{NLE}{NLE}}{}, requires passes through a flow-based density estimator, which is significantly faster when batched. This allowed us to sample all chains in parallel rather than sequentially and yielded huge speed-ups: For example, {\color{SNLE}{SNLE}}{} on Gaussian Linear took more than 36 hours on average for 100k simulations without vectorization, and less than 2 hours with vectorization.
\subsection[Density estimator for (S)NLE]{Density estimator for ({\color{SNLE}{S}}){\color{NLE}{NLE}}{}}
Approaches based on neural networks (NN) tend to have many hyperparameters, including the concrete type of NN architecture and hyperparameters for training. We strove to keep our choices close to \cite{durkan2020}, which are the defaults in the toolbox we used \citep[\texttt{sbi},][]{sbi}.
While \cite{papamakarios2019a,durkan2020} used Masked Autoregressive Flows \citep[MAFs,][]{papamakarios2017} for density estimation, we explored how results change when using Neural Spline Flows \citep[NSFs,][]{durkan2019neural} for density estimation. These results are shown in \autoref{fig:snle_maf_nsf}.
\begin{figure*}[h!]
\centering
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 3,clip,width=\textwidth]{figs/hyperparameters_snle/gaussian_linear_gaussian_linear_uniform_C2ST.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_snle/slcp_slcp_distractors_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_snle/bernoulli_glm_bernoulli_glm_raw_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_snle/gaussian_mixture_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_snle/two_moons_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_snle/sir_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 10 0 0,clip,width=\textwidth]{figs/hyperparameters_snle/lotka_volterra_C2ST_no_labels.pdf}
\end{subfigure}
%
\caption{
{\bf Density estimator selection for ({\color{SNLE}{S}}){\color{NLE}{NLE}}{}.} Performance of ({\color{SNLE}{S}}){\color{NLE}{NLE}}{} in terms of C2ST across tasks using MAFs or NSFs for density estimation. Considering all tasks, NSFs generally performed worse, e.g., using NSFs significantly reduced performance on SIR and Lotka-Volterra, indicating that the added flexibility of NSFs was not needed for ({\color{SNLE}{S}}){\color{NLE}{NLE}}{}. We thus reported performance using MAFs in the main paper. Each data point corresponds to the mean and 95\% confidence interval across 10 observations.
}
%
\label{fig:snle_maf_nsf}
%
\end{figure*}
\newpage
\clearpage
\subsection[Density estimator for (S)NPE]{Density estimator for ({\color{SNPE}{S}}){\color{NPE}{NPE}}{}}
We performed the analogous experiments for ({\color{SNPE}{S}}){\color{NPE}{NPE}}{} as for ({\color{SNLE}{S}}){\color{NLE}{NLE}}{}: Here, we found NSFs to increase performance relative to MAFs (\autoref{fig:snpe_maf_nsf}). When directly estimating the posterior distribution, especially on tasks with complex multi-modal structure like Two Moons or SLCP, the additional flexibility offered by NSFs improved performance. With NSFs, artifacts from density transformation that were visible e.g. in Two Moons posteriors, vanished. To our knowledge, results on ({\color{SNPE}{S}}){\color{NPE}{NPE}}{} with NSFs have not been previously published.
\begin{figure*}[h!]
\centering
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 3,clip,width=\textwidth]{figs/hyperparameters_snpe/gaussian_linear_gaussian_linear_uniform_C2ST.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_snpe/slcp_slcp_distractors_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_snpe/bernoulli_glm_bernoulli_glm_raw_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_snpe/gaussian_mixture_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_snpe/two_moons_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_snpe/sir_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 10 0 0,clip,width=\textwidth]{figs/hyperparameters_snpe/lotka_volterra_C2ST_no_labels.pdf}
\end{subfigure}
%
\caption{
{\bf Density estimator selection for ({\color{SNPE}{S}}){\color{NPE}{NPE}}{}.} Performance of ({\color{SNPE}{S}}){\color{NPE}{NPE}}{} in terms of C2ST across tasks using MAFs or NSFs for density estimation. Considering all tasks, NSFs generally performed better, especially on Gaussian Mixture, Two Moons, and SIR. We thus reported performance using NSFs in the main paper. Each data point corresponds to the mean and 95\% confidence interval across 10 observations.
%
}
%
\label{fig:snpe_maf_nsf}
%
\end{figure*}
\newpage
\clearpage
\subsection[Density estimator for (S)NRE]{Classifier choice for ({\color{SNRE}{S}}){\color{NRE}{NRE}}{}}
For ({\color{SNRE}{S}}){\color{NRE}{NRE}}{}, we compared two different choices of classifier architectures: an MLP and a ResNet architecture, as described in \ref{appendix:algorithms:nre}. While results were similar for most tasks (\autoref{fig:snre_mlp_res}), we decided to use the ResNet architecture in the main paper due to the better performance on Two Moons and SIR for low to medium simulation budgets.
\begin{figure*}[h!]
\centering
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 3,clip,width=\textwidth]{figs/hyperparameters_snre/gaussian_linear_gaussian_linear_uniform_C2ST.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_snre/slcp_slcp_distractors_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_snre/bernoulli_glm_bernoulli_glm_raw_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_snre/gaussian_mixture_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_snre/two_moons_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 49 0 0,clip,width=\textwidth]{figs/hyperparameters_snre/sir_C2ST_no_labels.pdf}
\end{subfigure}
%
\begin{subfigure}[htbp]{\textwidth}
\includegraphics[trim=20 10 0 0,clip,width=\textwidth]{figs/hyperparameters_snre/lotka_volterra_C2ST_no_labels.pdf}
\end{subfigure}
%
\caption{
{\bf Classifier architecture for ({\color{SNRE}{S}}){\color{NRE}{NRE}}{}.} Performance of ({\color{SNRE}{S}}){\color{NRE}{NRE}}{} in terms of C2ST across tasks using MLPs or ResNets for classification. Considering all tasks, ResNets generally performed better, especially on Two Moons and SIR. We thus reported performance using ResNets in the main paper. Each data point corresponds to the mean and 95\% confidence interval across 10 observations.
%
}
%
\label{fig:snre_mlp_res}
%
\end{figure*}
\subsubsection*{\bibname}}
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{url}
\usepackage{amsfonts}
\usepackage{booktabs}
\usepackage{nicefrac}
\usepackage{microtype}
\usepackage{bibunits}
\bibliographystyle{humannat}
\defaultbibliography{references}
\defaultbibliographystyle{humannat}
\input{preamble/preamble.tex}
\input{preamble/math.tex}
\usepackage[page,header]{appendix}
\listfiles
\begin{document}
\runningauthor{Jan-Matthis Lueckmann, Jan Boelts, David S. Greenberg, Pedro J. Gon\c{c}alves, Jakob H. Macke}
\twocolumn[
\aistatstitle{Benchmarking Simulation-Based Inference}
\aistatsauthor{%
\textbf{Jan-Matthis Lueckmann}$^{1,2}$\hspace{0.4cm} \textbf{Jan Boelts}$^{2}$\hspace{0.4cm}
David S. \textbf{Greenberg}$^{2,3}$
\\
\textbf{Pedro J. Gon\c{c}alves}$^{4}$\hspace{0.4cm}
\textbf{Jakob H. Macke}$^{1,2,5}$\hspace{0.4cm}\vspace{0.2cm}
}
\aistatsaddress{$^{1}$University of Tübingen\hspace{0.4cm}$^{2}$Technical University of Munich\hspace{0.4cm}$^{3}$Helmholtz Centre Geesthacht\\$^{4}$Research Center caesar\hspace{0.4cm}$^{5}$Max Planck Institute for Intelligent Systems, Tübingen}
]
\special{papersize = 8.5in, 11in}
\setlength{\pdfpageheight}{11in}
\setlength{\pdfpagewidth}{8.5in}
\input{00_abstract}
\begin{bibunit}
\input{01_introduction}
\input{02_benchmark}
\input{03_results}
\input{05_discussion}
\addcontentsline{toc}{section}{\protect\numberline{\thesection}{References}}
\putbib
\end{bibunit}
\onecolumn
\newpage
\clearpage
\subsubsection*{\bibname}}
\begin{document}
\onecolumn
\aistatstitle{Instructions for Paper Submissions to AISTATS 2021: \\
Supplementary Materials}
\section{FORMATTING INSTRUCTIONS}
To prepare a supplementary pdf file, we ask the authors to use \texttt{aistats2021.sty} as a style file and to follow the same formatting instructions as in the main paper.
The only difference is that the supplementary material must be in a \emph{single-column} format.
You can use \texttt{supplement.tex} in our starter pack as a starting point, or append the supplementary content to the main paper and split the final PDF into two separate files.
Note that reviewers are under no obligation to examine your supplementary material.
\section{MISSING PROOFS}
The supplementary materials may contain detailed proofs of the results that are missing in the main paper.
\subsection{Proof of Lemma 3}
\textit{In this section, we present the detailed proof of Lemma 3 and then [ ... ]}
\section{ADDITIONAL EXPERIMENTS}
If you have additional experimental results, you may include them in the supplementary materials.
\subsection{The Effect of Regularization Parameter}
\textit{Our algorithm depends on the regularization parameter $\lambda$. Figure 1 below illustrates the effect of this parameter on the performance of our algorithm. As we can see, [ ... ]}
\vfill
\end{document}
\subsubsection*{\bibname}}
\begin{document}
\twocolumn[
\aistatstitle{Instructions for Paper Submissions to AISTATS 2021}
\aistatsauthor{ Author 1 \And Author 2 \And Author 3 }
\aistatsaddress{ Institution 1 \And Institution 2 \And Institution 3 } ]
\begin{abstract}
The Abstract paragraph should be indented 0.25 inch (1.5 picas) on
both left and right-hand margins. Use 10~point type, with a vertical
spacing of 11~points. The \textbf{Abstract} heading must be centered,
bold, and in point size 12. Two line spaces precede the
Abstract. The Abstract must be limited to one paragraph.
\end{abstract}
\section{GENERAL FORMATTING INSTRUCTIONS}
The camera-ready versions of the accepted papers are 8 pages,
plus any additional pages needed for references.
Papers are in 2 columns with the overall line width of 6.75~inches (41~picas).
Each column is 3.25~inches wide (19.5~picas). The space
between the columns is .25~inches wide (1.5~picas). The left margin is 0.88~inches (5.28~picas).
Use 10~point type with a vertical spacing of
11~points. Please use US Letter size paper instead of A4.
Paper title is 16~point, caps/lc, bold, centered between 2~horizontal rules.
Top rule is 4~points thick and bottom rule is 1~point thick.
Allow 1/4~inch space above and below title to rules.
Author descriptions are center-justified, initial caps. The lead
author is to be listed first (left-most), and the Co-authors are set
to follow. If up to three authors, use a single row of author
descriptions, each one center-justified, and all set side by side;
with more authors or unusually long names or institutions, use more
rows.
Use one-half line space between paragraphs, with no indent.
\section{FIRST LEVEL HEADINGS}
First level headings are all caps, flush left, bold, and in point size
12. Use one line space before the first level heading and one-half line space
after the first level heading.
\subsection{Second Level Heading}
Second level headings are initial caps, flush left, bold, and in point
size 10. Use one line space before the second level heading and one-half line
space after the second level heading.
\subsubsection{Third Level Heading}
Third level headings are flush left, initial caps, bold, and in point
size 10. Use one line space before the third level heading and one-half line
space after the third level heading.
\paragraph{Fourth Level Heading}
Fourth level headings must be flush left, initial caps, bold, and
Roman type. Use one line space before the fourth level heading, and
place the section text immediately after the heading with no line
break, but an 11 point horizontal space.
\subsection{Citations, Figure, References}
\subsubsection{Citations in Text}
Citations within the text should include the author's last name and
year, e.g., (Cheesman, 1985).
Be sure that the sentence reads
correctly if the citation is deleted: e.g., instead of ``As described
by (Cheesman, 1985), we first frobulate the widgets,'' write ``As
described by Cheesman (1985), we first frobulate the widgets.''
The references listed at the end of the paper can follow any style as long as it is used consistently.
\subsubsection{Footnotes}
Indicate footnotes with a number\footnote{Sample of the first
footnote.} in the text. Use 8 point type for footnotes. Place the
footnotes at the bottom of the column in which their markers appear,
continuing to the next column if required. Precede the footnote
section of a column with a 0.5 point horizontal rule 1~inch (6~picas)
long.\footnote{Sample of the second footnote.}
\subsubsection{Figures}
All artwork must be centered, neat, clean, and legible. All lines
should be very dark for purposes of reproduction, and art work should
not be hand-drawn. Figures may appear at the top of a column, at the
top of a page spanning multiple columns, inline within a column, or
with text wrapped around them, but the figure number and caption
always appear immediately below the figure. Leave 2 line spaces
between the figure and the caption. The figure caption is initial caps
and each figure should be numbered consecutively.
Make sure that the figure caption does not get separated from the
figure. Leave extra white space at the bottom of the page rather than
splitting the figure and figure caption.
\begin{figure}[h]
\vspace{.3in}
\centerline{\fbox{This figure intentionally left non-blank}}
\vspace{.3in}
\caption{Sample Figure Caption}
\end{figure}
\subsubsection{Tables}
All tables must be centered, neat, clean, and legible. Do not use hand-drawn tables.
Table number and title always appear above the table.
See Table~\ref{sample-table}.
Use one line space before the table title, one line space after the table title,
and one line space after the table. The table title must be
initial caps and each table numbered consecutively.
\begin{table}[h]
\caption{Sample Table Title} \label{sample-table}
\begin{center}
\begin{tabular}{ll}
\textbf{PART} &\textbf{DESCRIPTION} \\
\hline \\
Dendrite &Input terminal \\
Axon &Output terminal \\
Soma &Cell body (contains cell nucleus) \\
\end{tabular}
\end{center}
\end{table}
\section{SUPPLEMENTARY MATERIAL}
If you need to include additional appendices during submission, you can include them in the supplementary material file.
You can submit a single file of additional supplementary material which may be either a pdf file (such as proof details) or a zip file for other formats/more files (such as code or videos).
Note that reviewers are under no obligation to examine your supplementary material.
If you have only one supplementary pdf file, please upload it as is; otherwise gather everything to the single zip file.
You must use \texttt{aistats2021.sty} as a style file for your supplementary pdf file and follow the same formatting instructions as in the main paper.
The only difference is that it must be in a \emph{single-column} format.
You can use \texttt{supplement.tex} in our starter pack as a starting point.
Alternatively, you may append the supplementary content to the main paper and split the final PDF into two separate files.
\section{SUBMISSION INSTRUCTIONS}
To submit your paper to AISTATS 2021, please follow these instructions.
\begin{enumerate}
\item Download \texttt{aistats2021.sty}, \texttt{fancyhdr.sty}, and \texttt{sample\_paper.tex} provided in our starter pack.
Please, do not modify the style files as this might result in a formatting violation.
\item Use \texttt{sample\_paper.tex} as a starting point.
\item Begin your document with
\begin{flushleft}
\texttt{\textbackslash documentclass[twoside]\{article\}}\\
\texttt{\textbackslash usepackage\{aistats2021\}}
\end{flushleft}
The \texttt{twoside} option for the class article allows the
package \texttt{fancyhdr.sty} to include headings for even and odd
numbered pages.
\item When you are ready to submit the manuscript, compile the latex file to obtain the pdf file.
\item Check that the content of your submission, \emph{excluding} references, is limited to \textbf{8 pages}. The number of pages containing references alone is not limited.
\item Upload the PDF file along with other supplementary material files to the CMT website.
\end{enumerate}
\subsection{Camera-ready Papers}
If your papers are accepted, you will need to submit the camera-ready version. Please make sure that you follow these instructions:
\begin{enumerate}
\item Change the beginning of your document to
\begin{flushleft}
\texttt{\textbackslash documentclass[twoside]\{article\}}\\
\texttt{\textbackslash usepackage[accepted]\{aistats2021\}}
\end{flushleft}
The option \texttt{accepted} for the package
\texttt{aistats2021.sty} will write a copyright notice at the end of
the first column of the first page. This option will also print
headings for the paper. For the \emph{even} pages, the title of
the paper will be used as heading and for \emph{odd} pages the
author names will be used as heading. If the title of the paper
is too long or the number of authors is too large, the style will
print a warning message as heading. If this happens additional
commands can be used to place as headings shorter versions of the
title and the author names. This is explained in the next point.
\item If you get warning messages as described above, then
immediately after $\texttt{\textbackslash
begin\{document\}}$, write
\begin{flushleft}
\texttt{\textbackslash runningtitle\{Provide here an alternative
shorter version of the title of your paper\}}\\
\texttt{\textbackslash runningauthor\{Provide here the surnames of
the authors of your paper, all separated by commas\}}
\end{flushleft}
Note that the text that appears as argument in \texttt{\textbackslash
runningtitle} will be printed as a heading in the \emph{even}
pages. The text that appears as argument in \texttt{\textbackslash
runningauthor} will be printed as a heading in the \emph{odd}
pages. If even the author surnames do not fit, it is acceptable
to give a subset of author names followed by ``et al.''
\item The camera-ready versions of the accepted papers are 8
pages, plus any additional pages needed for references.
\item If you need to include additional appendices,
you can include them in the supplementary
material file.
\item Please, do not change the layout given by the above
instructions and by the style file.
\end{enumerate}
\subsubsection*{Acknowledgements}
All acknowledgments go at the end of the paper, including thanks to reviewers who gave useful comments, to colleagues who contributed to the ideas, and to funding agencies and corporate sponsors that provided financial support.
To preserve the anonymity, please include acknowledgments \emph{only} in the camera-ready papers.
\subsubsection*{References}
References follow the acknowledgements. Use an unnumbered third level
heading for the references section. Please use the same font
size for references as for the body of the paper---remember that
references do not count against your page length total.
|
\section{Introduction}
The COVID-19 pandemic is touching many aspects of human life. Information has been rapidly generated to inform the public about the pandemic, help people understand complicated mathematical forecasting models, and persuade them to make behavioral changes to mitigate the spread of the disease. In addition to all the textual information being disseminated about COVID-19, a significant number of visualizations have been created to help people make sense of the complexities of the pandemic and understand the evolving public health crisis. \textit{Crisis visualizations} (i.e., visual representations of crisis information such as disease prevalence, epidemiological simulations, and economic and social changes) are being circulated daily. COVID-19 visualizations have been produced by diverse content creators (e.g., scientists, government and healthcare officials, social media users, news media outlets) and disseminated to a large number of audiences. This pandemic may be the first time in history that such a large proportion of the public has been engaged with and responding to visualizations.
Research has shown that visualizations are useful to help the general public, who often have difficulties understanding complex crisis dynamics, making sense of crisis situations, assessing personal risk, and decision-making~\cite{stone2015effects}. Prior work has suggested that visualizing information helps reduce mental load and allows people to grasp content more quickly, compared with presenting that information in text~\cite{dur2014interactive}. Research has examined crisis visualization design from various perspectives, such as providing guidance on how to design risk maps effectively~\cite{dent2008thematic, fostermapping, roth2012visualizing} and visualizing uncertainty to communicate risks~\cite{dieckmann2015home, dieckmann2017seeing, frewer2002public, padilla2017effects, van2019communicating}. Moreover, there has been an ongoing discussion among researchers, practitioners, and the public regarding whether or not to create COVID-19 visualizations~\cite{Correll_medium} and how to design these visualizations responsibly~\cite{Makulec_medium}.
Given this proliferation of visualization production and consumption, there is an urgent and critical need to document and organize these efforts in a timely manner in order to provide guidance for emerging visualizations. Moreover, it is important to understand what visualizations have been designed and disseminated to the public because information consumption can impact people's behaviors, attitudes, and thus ultimately the path of the pandemic. Unfortunately, to date, there has not been a review of crisis visualizations that aim to communicate with the general public from a visualization perspective.
To address this research gap, we curated and analyzed 668 visualizations that communicate information about COVID-19. Through our analysis, we derived a conceptual framework that focuses on understanding \textbf{who}, uses \textbf{what data}, to communicate \textbf{what messages}, in \textbf{what form}, as well as emphasizing the \textbf{contexts (what circumstances)} in which crisis visualizations are created.
In this paper, we use this framework to discuss several characteristics of COVID-19 visualizations. First, we describe the variety of stakeholders who created the visualizations in our corpus. Second, we examine the characteristics of data within these visualization---including the data sources, data source citations, data quality and uncertainty. Third, we provide an explication of the six categories of messages inherent within these visualizations: informing of severity, forecasting trends and influences, explaining the nature of the crisis, guiding risk mitigation, communicating risk, vulnerability, and equity, and gauging the multifaceted impacts of the crisis. Fourth, we unpack the trends in the visualization techniques and encodings that have been used to communicate each type of message. Lastly, we explore the dynamic temporal context of crisis visualization design, specifically, how COVID-19 visualizations have changed over time.
Our work makes contributions to visualization, crisis informatics, and public health research and practice.
First, our findings will help to accelerate the design process of ongoing and future crisis visualizations by outlining currently used visualization methods and visual encoding techniques. By outlining these trends as well as open challenges, our findings can help inspire new visualization approaches and research areas and help designers avoid common pitfalls, thus accelerating innovation.
Second, our work contributes to crisis informatics research, a field of study that examines the role of information and communication technologies in times of crisis. Little work in this field has examined the nature of visualizations designed for the general public during times of crisis. Through our comprehensive analysis of crisis visualizations, we contribute new knowledge regarding the nature of such visualizations. Specifically, we provide a systematic synthesis of crisis visualizations that characterizes trends and patterns in who, communicated what messages, to whom, in the dynamic temporal context of the COVID-19 pandemic. In addition, by formalizing our findings around the COVID-19 crisis visualization landscape in a preliminary framework \jz{---by extending and blending existing models in communication research and visualization research---we provide a thinking tool that aims to descriptively capture dimensions of the crisis visualization design space and important areas that warrant further inquiry.}
Third, through our framework and accompanying findings that document the kinds of messaging inherent in COVID-19 visualizations, this work will help catalyze research at the intersection of visualization and public health that examines the impact of consuming COVID-19 visualizations on behaviors and attitudes. Specifically, we provide a discussion of ways in which future work can further examine questions around the creation and impact of crisis visualizations raised in each area of our framework. Our framework will provide researchers with a conceptual tool that can guide empirical investigation, by highlighting important areas of inquiry in the crisis visualization space (e.g., investigating how exposure to different categories of crisis visualizations deferentially impacts risk perceptions, prevention behaviors, and mental health).
\section{Related Work}
\label{sec:background}
\subsection{Crisis Informatics}
\label{sec:crisis_informatics}
Traditional risk and crisis communication has emphasized a one-way, top-down approach to information creation and dissemination (e.g., transmitting information from entities such as governmental authorities to the general public~\cite{hagar2010introduction, zhang2020ci}). Today, much risk and crisis communication has shifted to two-way information exchange between the institutions and governmental authorities and the general public, a shift which necessitates considerations of the needs and values of each stakeholder group. For example, social media platforms enable direct communication between the authorities and the public in various forms (e.g., following, mentioning, replying)~\cite{kim2018emergency, zhang2019social}, which enables two-way communication.
The involvement of the general public in crisis communication has transformed information and communications technology design to be more participatory~\cite{zhang2019social}, a paradigm shift that has helped fuel the development of the crisis informatics research area.
\textit{Crisis informatics} is an interdisciplinary area that examines the interconnectedness of people, organizations, information, and technology during crises~\cite{hagar2010introduction}. Crisis informatics research has studied how crisis and risk communication has occurred online~\cite{chauhan2017providing}, and the use of information and communications technology during crises~\cite{zhang2020ci}.
Despite the increasing interest in crisis informatics research, less work has examined the creation and dissemination of visualizations in times of crisis. Notable examples of work in this area include research examining ``public reporting'' and ``crisis mapping''---this research has examined participatory maps and crowdsourced crisis databases, and built computational models to provide risk warning and mitigation approaches~\cite{herranz2014multi, norheim2010crowdsourcing}. Another body of work has focused on the design and development of visualization tools to help people understand and explore trends arising during emergency situations and unfolding crises ~\cite{onorati2019social, romero2016towards}. However, overall, there has been little work that explores the intersection of visualization and crisis informatics. We address this gap, identifying challenges and opportunities in this field through a comprehensive analysis of crisis visualizations, such as how public engagement may shape the design of crisis visualization.
\subsection{Crisis Visualizations}
\label{sec:crisis_vis}
With increases in the amount of information about crises and associated risks has come an increase in the production and dissemination of visual representations of this information. \textit{Risk visualization} can be defined as ``the systematic effort of using images to augment the quality of risk communication along the entire risk management cycle''~\cite{roth2012visualizing}. Risk visualizations aim to depict the risks that people face and to help institutions and authorities communicate with the general public about risks. More recently, data collection, analysis, modeling, and visualizations of outbreak data has become increasingly complex, leading to an emerging field of \textit{outbreak analytics}, where visualization plays an important role in supporting sensemaking of complex outbreak data~\cite{polonsky2019outbreak}. The target audiences for visualizations of outbreak analytics mainly include health professionals and analysts. Visualizations in these fields may be defined distinctly depending on the contexts and approaches, the concepts are clearly interconnected---they are visual representations that aim to augment communication about undesirable situations that have inherent uncertainty, and that posit threats to human beings~\cite{zhang2020ci}. As such, we define \textit{crisis visualizations} as visual representations of information about undesirable situations that posit threats to human beings. These visualizations communicate information such as disease prevalence, epidemiological simulations, and economic and social changes. We will be using the term \textit{crisis visualizations} throughout the paper.
Prior work has explored data visualizations of historical pandemic events~\cite{lu2004web, preim2020survey, welhausen2015visualizing} \jz{and other crises, such as hurricane~\cite{Bica2019Com, padilla2017effects}}. This body of work has examined how visualizations influence risk perception~\cite{welhausen2015visualizing} and explored the design of visual analytic tools for modeling and simulation~\cite{polonsky2019outbreak}. A more recent survey on visual analytics for public health~\cite{preim2020survey} has presented a variety of requirements and visual analytics techniques to help guide epidemiologists and environmental health specialists conduct visual analyses of public health issues. There is thus a significant body of work that has focused on crisis visualizations that assume specialized audiences and require higher levels of numeracy and visualization literacy to interpret them. In contrast, our work examines crisis visualizations that aim to communicate with the general public.
To date, research focused on COVID-19 visualization has been limited. Several notable exceptions include recently published work by D’Ignazio and Klein~\cite{dignazio2020datafeminism}, as well as Bowe, Simmons, and Mattern~\cite{bowe2020learning}. Both works have provided critical consideration of broader issues, such as equity, power, and social justice associated with COVID-19 data. Bowe et al.~\cite{bowe2020learning} used the lens of ``embodiment'' to interpret COVID-19 visualizations and discussed the importance of creating more meaningful visualizations through the use of embodied and local data.
However, to our knowledge, existing work has not reported on syntheses of large crisis visualization datasets, and thus may not sufficiently depict a broad picture of COVID-19 visualizations. Considering that these crisis visualizations have received huge public engagement and the significant role of visualization in communicating crisis information, comprehensive and systematic analyses of COVID-19 visualizations are needed. Our work aims to identify emerging trends and patterns in currently used visualization methods and visual encoding techniques for COVID-19 crisis visualizations, as well as to identify challenges and common pitfalls.
\section{Methodology}
\label{sec:methodology}
This paper reports on our analysis of a COVID-19 visualization collection that we assembled. To compile this collection, we used opportunistic sampling, also called emergent sampling, a non-probability sampling method for data collection. Opportunistic sampling ``takes advantage of unforeseen opportunities after fieldwork has begun,'' and is particularly useful for synthesizing exploratory research areas~\cite{patton1990qualitative}. The time-sensitive and unpredictable evolution of the COVID-19 pandemic can make a priori sampling decisions challenging, thus motivating our use of opportunistic sampling, an approach that enabled us to flexibly collect visualizations as they were produced and disseminated. Several other works in crisis informatics have successfully used opportunistic sampling to study a variety of crisis events before (e.g., ~\cite{kassam2006mental, tang2020need}). However, we also need to point out the limitations of opportunistic sampling. Opportunistic sampling can suffer from selection bias and generalizability issues for findings, necessitating care when interpreting the results.
\subsection{Data Collection}
Our collection includes visual representations that communicate information about COVID-19, including both data visualizations (e.g., an interactive map, graph, chart, or diagram) and infographics (e.g., a static narrative or graphics). At the beginning of the data collection, early March 2020, we collected an initial set of visualizations. Later in March, to expand the collection, we began compiling a corpus of crisis visualizations focused on COVID-19 using image database searches (e.g., Google Images) and word-of-mouth contributions. We also searched on visualization blogs that contained special topics on the pandemic, such as \texttt{coronavirus} on FlowingData and \texttt{Science \& Health} on FiveThirtyEight. We then cleaned the collection by removing duplicated entries, so that each entry in the database had a unique URL. Whenever possible, we included the original visualization that appeared online in our database. Also, we removed entries with broken URLs (websites that were not maintained) on July 31, 2020. Note that some visualizations in our collection have been updated regularly; therefore our analysis reflects the state of these visualizations at the time we visited the website.
\subsection{Corpus}
\begin{figure*}[tb]
\centering
\includegraphics[width=\textwidth]{fig/framework.pdf}
\caption{A conceptual framework for understanding crisis visualizations, built upon existing models in communication research~\cite{braddock1958extension, lasswell1948structure} and visualization research~\cite{munzner2009nested}, focuses on \textit{who}, (uses) \textit{what data}, (to communicate) \textit{what message}, \textit{in what form}, highlighted in gray~\colorRect{frameworkColor}.
Our framework also emphasizes the role of circumstances and contexts in understanding crisis visualizations, indicated by the surrounded dashed lines \colorVLine{colorFramework} \colorHLine{colorFramework} \colorHLine{colorFramework} \colorHLine{colorFramework} \colorVLine{colorFramework}.
The back dashed arrow $\dashleftarrow$ indicates a non-linear flow of visualization design and communication.
}
\label{fig:framework}
\end{figure*}
To focus the scope of this work, we analyzed a subset of visualizations from our total collection of 1184 visualizations. The inclusion criteria for this work include visualizations in English that were published online and aimed to communicate with the general public. We used information outlets as a proxy for the target audience (i.e., general public vs. specialized audiences), for example, including visualizations shared on news media and governmental websites that provide information for the general public. We did not include academic publications, which typically assume specialized audiences. With these criteria in mind, this paper reports on our analysis of 668 visualizations (published between January 22 and July 31, 2020).
\subsection{Analysis}
To analyze our visualization corpus, we engaged in a two-step process that involved both inductive and deductive coding. This multi-phase process enabled a rigorous and systematic analysis of our data set. We first conducted an inductive analysis~\cite{thomas2006general} of the collected visualizations, to develop an initial set of codes that were grounded in the phenomena reflected within our corpus. At the beginning of our analysis in late April (when we collected our initial set of visualizations), two researchers first individually reviewed our entire collection to develop an initial understanding of the visualization themes arising in the data. The two researchers then inductively coded the first 100 visualizations in the collection, and then clustered related low-level codes to arrive at high-level themes that characterize the visualizations. The two researchers discussed the coding scheme regularly to review the evolving codes to achieve a mutual understanding and to refine the codebook. The codebook contains a table that has a list of the codes and their definitions that reflect the concept referred to by the code. The goal of the codebook was to maximize coherence among coders~\cite{creswell2017research}.
We also consulted with senior visualization researchers and had conversations with visualization practitioners and journalists who created COVID-19 visualizations to further refine the codebook, which allowed us to add new codes and reorganize the codes. This multiphase process enabled us to arrive at a detailed and multifaceted codebook with which to characterize the COVID-19 visualizations in our corpus.
Our final codebook included 61 codes within four categories: metadata (e.g., title, URL, publisher), type of visualizations (e.g., choropleth map), messages that the visualizations aim to communicate (e.g., informing of the severity), data handling such as whether the visualization used normalized data. Guided by the codebook, we then conducted a deductive analysis of the visualizations~\cite{thomas2006general}, which enabled us to exhaustively assess to what extent the various concepts in our codebook were reflected across our entire visualization corpus. We re-coded the previously-coded 100 visualizations, and continued coding the rest of the visualizations in the collection at that time. In each phase of coding, the researchers discussed any points of disagreement in the application of codes until we reached a consensus. The codebook and the corpus used for this paper are provided as supplemental material and also available at \href{https://osf.io/gf8zr/}{https://osf.io/gf8zr/}.
\begin{figure*}[tb]
\centering
\includegraphics[width=\textwidth]{fig/outlets_bar.pdf}
\caption{\jz{The distribution of COVID-19 visualizations in our corpus, by information outlets and type of messages communicated (the y-axis is in a logarithmic scale).}}
\label{fig:message_outlet}
\end{figure*}
\section{A Conceptual Framework for Crisis Visualizations}
\label{sec:unique_characteristics}
Through our analysis, we derived a conceptual framework for crisis visualizations (see ~\autoref{fig:framework}). Our framework focuses on \textbf{who}, uses \textbf{what data}, communicates \textbf{what messages} in \textbf{what form}, and emphasizes the \textbf{contexts (what circumstances)} in which visualizations are created, based on Lasswell’s model of communication~\cite{lasswell1948structure}, Braddock's extended formulation of Lasswell's model~\cite{braddock1958extension}\jz{, and Munzner's nested model~\cite{munzner2009nested}}. Lasswell's model of communication (1940)~\cite{lasswell1948structure} describes an act of communication by defining ``Who, said what, in which channel, to whom, with what effect?''. Critiques of the model include that it neglects the impact of environments on communication. Braddock (1958)~\cite{braddock1958extension} thus advocated the addition of ``under what circumstances’’ to the model, to emphasize the attributes of time and contexts, and their influences on the communication process. \jz{The nested model by Munzner~\cite{munzner2009nested} highlights domain situations, task and data abstractions, visual encoding, and algorithms. In line with this nested model, our proposed framework also emphasizes domain situations and data and visual encoding. The difference lies in the prescriptive versus the descriptive nature of the work. Munzner’s nested model ``provides prescriptive guidance for determining appropriate evaluation approaches by identifying threats to validity unique to each level''~\cite{munzner2009nested}. Our framework is descriptive in the sense of providing concepts and guiding further inquiry through organizing and analyzing surveyed visualizations to capture their commonalities. Moreover, the nested model does not explicitly focus on ``who'' (i.e., content producers), ``communicates what messages'', ``to whom'', ``with what effect''.} Building upon \jz{these existing models in communication and visualization research,} our conceptual framework aims to support the analysis of crisis visualization communication patterns, \jz{to capture dimensions of the crisis visualization design space, and highlight areas that are worth further inquiry.}
In our framework, the first component is \textit{who} is creating COVID-19 visualizations (briefly described in \autoref{sec:who}). The second component is \textit{what data} is used for creating visualizations (\autoref{sec:data}), which focuses on: a) understanding the diversity of data sources in crisis visualizations, b) data source citation practices, and c) practices of dealing with uncertainty caused by unavailable, missing, incomplete, and inconsistent crisis data. The third component is \textit{what message} is being communicated through existing visualizations (\autoref{sec:messages}). The fourth component is \textit{in what form} is data presented in visualizations (\autoref{sec:visual_forms}). The fifth component examines \textit{under what circumstances} visualizations are created (\autoref{sec:contexts}), with a particular focus on understanding the dynamic temporal nature crises.
As a preview, our findings suggest that crisis visualizations have been changed over time to fit the constantly changing crisis situations. Moreover, the impacts that visualizations make on people (e.g., shifting attitudes and public perceptions) then, in turn, impact how and what is being designed, and thus the process of design and communication is non-linear~(as reflected by the back dashed arrow $\dashleftarrow$).
While our findings do not focus on the final two components of the framework---(\textit{to whom}) and \textit{with what effect}---our framework and the insights gained from our analysis do suggest important areas for future inquiry into these components. We discuss such opportunities in the Discussion section (\autoref{sec:discussion}).
\section{Who created the visualizations?}
\label{sec:who}
In our collection, there were 174 distinct information outlets that shared COVID-19 visualizations. Most visualizations were shared in news media outlets (57\%, 380 out of 668, e.g., the New York Times), followed by non-governmental organizations (NGOs) and research institutes (14\%, 95 out of 668, e.g., Johns Hopkins University), governmental agencies and public health officials (9\%, 60 out of 668, e.g., United States Centers for Disease Control \& Prevention), independent media outlets (8\%, 52 out of 668, e.g., FlowingData), corporations (7\%, 51 out of 668, e.g., McKinsey \& Company), and social media posts (4\%, 27 out of 668, e.g., Twitter). \autoref{fig:message_outlet} summarizes the different messages focused on by each information outlet in our corpus. We will describe this messaging in more detail in~\autoref{sec:messages}.
\jz{
In terms of the visualization country of origin, the majority of the surveyed visualizations in our corpus are from the United States of America (66\%, 439 out of 668), followed by the United Kingdom (20\%, 131 out of 668), Canada (3\%, 18 out of 668), India (2\%, 16 out of 668). The rest 9\% belong to Switzerland (n=8), Japan (n=6), China (n=6), Germany (n=6), New Zealand (n=5), Belgium (n=4), Iraq (n=4), Singapore (n=3), South Africa (n=3), Qatar (n=3), Italy (n=2), Australia (n=2), Finland (n=2), Sweden (n=1), Portugal (n=1), Brazil (n=1), Iceland (n=1), Bangladesh (n=1), Nigeria (n=1), France (n=1), Denmark (n=1), and for two visualizations, we were unable to determine the country of origin. }
\section{What data has been used to create the visualizations?}
\label{sec:data}
In this section, we provide an overview of data sources described in our database, the current status of data source citations, and data source quality and uncertainty.
\subsection{Data Sources}
We identified 486 distinct data sources that were used to communicate various aspects of the pandemic; among which 200 data sources were used to create visualizations to convey the message of the severity of the pandemic (see ~\autoref{sec:inform_severity}). Recent work has discussed how available COVID-19 data has varied in format, which in turn has led to inconsistent and incomplete data reporting with varied quality~\cite{lau2020internationally}. We also see the sheer variety of data sources represented in our corpus. To help investigate data quality issues, groups of ``digital volunteers'' have self-organized---online communities of volunteers across fields and cultures who collaboratively collect, verify, and map crisis information across digital channels. For example, the team of 1Point3Acres~\cite{1Point3Acres}---a group of volunteers with different backgrounds such as computer science, public policy, public health, and biology---have engaged in fact-checking and cross-referenced different sources in an effort to produce more accurate COVID-19 data.
We also identified eight visualizations that used crowdsourced data sources. These visualizations were all from research institutes and NGOs. The crowdsourced data has been used to help understand activity in communities, such as examining additional characteristics of the pandemic, early detection, and local forecasting of the pandemic. In our corpus, for example, COVIDcast~\cite{cmu_covid} includes multiple crowdsourced indicators, such as doctor visits and hospital admissions (based on data from health system partners), self-reported symptoms through surveys (advertised by Facebook), search trends (data from Google), and COVID-19 antigen tests (data from Quidel). The resulting visualization was a choropleth map that integrated various crowdsourced indicators. Similarly, COVIDNearYou~\cite{covid_near_you} also applies crowdsourcing methodologies to bootstrap public health tracking (e.g., asking about website visitors' health status daily) as a way of identifying current and potential illness hotspots.
\subsection{Data Source Citations}
Including data source citations in visualizations is one of the most important ways of supporting data provenance. Data provenance refers to data source citations, references, methodological choices, relevant facts, and annotations of exceptions and correction~\cite{hullman2011visualization}. 93\% of visualizations in our collection cited the original data source, suggesting that the majority of the visualizations were rigorous in this respect. However, we still see that approximately 7\% did not include source citations, which undermines the credibility of those visualizations.
Among the uncited visualizations, 32\% were used to depict the severity of the pandemic, followed by communicating risks, vulnerability, and equity (24\%), gauging the impact of the pandemic (17\%), guiding risk mitigation (10\%), and explaining coronavirus (5\%). Some of these visualizations were screenshots shared on social media (e.g., Twitter, Reddit) without source citations. These visualizations went viral on the internet~\cite{viralgraph2020}, potentially spreading and bolstering fear and uncertainty, and driving harmful reactions such as taking medicine that has not been fully tested and not taking the pandemic seriously. Therefore, having data sources cited is important to improve the validity of the visualizations.
\subsection{Data Source Quality and Uncertainty}
Currently unavailable, missing, incomplete, and inconsistent information may cause uncertainty. Our analysis shows that COVID-19 visualization designers communicated uncertainty related to data quality in three ways: adding disclaimers, including explanations of the data collection (e.g., describing missing data or inconsistencies in reporting), and assessing the uncertainty caused by the data quality issues. For example, in our corpus, the COVID Tracking project~\cite{covid_tracking_project} offers data-quality grades, which assess how well each US state reports COVID-19 data. The grade is based on 16 separate factors within 5 categories, including reporting (how well states format and publish data), testing data completeness (whether a state is publishing complete basic testing data), patient outcomes (whether a state is reporting on COVID-19's effect on patients), demographics (whether a state reports the break-down information by demographic categories), and other (e.g., whether a state reports hospital capacity). Attempts such as these may help provide the public with guidelines for interpreting and formulating decisions based on data that has inherent uncertainty.
\section{What messages do the visualizations communicate?}
\label{sec:messages}
COVID-19 crisis visualizations communicate a wide range of information about the pandemic. Based on our analysis, we classify these messages as reflecting six purposes: (1) informing of the severity; (2) forecasting trends and influences; (3) explaining the nature of the crisis; (4) guiding risk mitigation; (5) communicating risk, vulnerability, and equity; and (6) gauging the multifaceted impacts of the crisis. Note that these message categories are not mutually exclusive. In our corpus, it is common for one visualization to convey more than one message.
\textbf{(1) Informing of the severity:}
The severity of a pandemic, as represented by the number of diagnoses, hospitalized patients, incubated patients, and deaths, is one essential aspect of public health communication~\cite{hastall2013severity} and crisis communication~\cite{cori2017key}. The importance of communicating the severity of a crisis is reflected by the large proportion of the visualizations in our database (231 out of 668, 35\% of the visualizations aim to communicate the severity of the pandemic). Indeed, this was the message category most represented in our corpus. By far, most visualizations were created to inform the public about the current state of the crisis, including the trajectory of the crisis over time and the geospatial spread of the crisis (218 out of 231, 94\%). A much smaller proportion (13 out of 231, 6\%) compared COVID-19 with historical events.
Most of the visualizations that compared with historical events (8 out of 13, 62\%) in our collection were published in the early stage of the outbreak (January to March 2020). One reason for this early focus on historical events may be that there was a greater level of uncertainty during the initial phase of the crisis, and as such comparing the pandemic with historical events may have been seen as a way to help the public make sense of what was happening.
\textbf{(2) Forecasting trends and influences:}
Visualizations have been created to estimate and forecast trends and influences in terms of estimation of current transmissibility, effective reproduction rate, and projections of the pandemic (57 out of 668, 9\%).
\textbf{(3) Explaining the nature of the crisis:}
A body of visualizations in our corpus used biomedical information to explain the nature of the COVID-19 crisis (95 out of 668, 14\%), including the structure of the virus (25 out of 95, 26\%), illness symptoms and timeline of diagnosis (26 out of 95, 27\%), how disease transmission happens (31 out of 95, 33\%), and approaches to contact tracing (13 out of 95, 14\%).
\textbf{(4) Guiding risk mitigation:}
Visualizations have been also created to guide risk mitigation (46 out of 668, 7\%), including providing practical instructional guidance to keep social distancing, how to prepare and use personal protective equipment (PPE), how to maintain proper hygiene, \jz{as well as justification for particular risk mitigation strategies}.
\begin{figure*}[tb]
\centering
\includegraphics[width=\textwidth]{fig/examples/temporal.pdf}
\caption{Examples of temporal visualizations:
(A) New reported cases by day in the United States, superimposed with a 7-day average line (by the New York Times~\cite{nyt_dashboard});
(B) A set of Pez Charts showing case number changes over time (by the Los Angeles Times~\cite{latimes_dashboard});
(C) A type of Growth Chart displaying the total number in the past week against the total number over time (by Aatish Bhatia~\cite{aatishb_growth_chart});
(D) Days since confirmed cases first reached 30 cases per day using event alignment (in a log scale) (by Our World In Data~\cite{owidcoronavirus}).}
\label{fig:temporal}
\end{figure*}
\textbf{(5) Communicating risk, vulnerability, and equity:}
Another set of visualizations aimed to communicate risk and unpack issues of vulnerability and equity (124 out of 668, 19\%). Previously identified potential risk factors for COVID-19 morbidity and mortality include age, race/ethnicity, gender, some medical conditions, the use of certain medications, poverty, crowding, certain occupations, and pregnancy~\cite{cdc_risk_factor}. These factors are also influenced by the social determinants of health, that is, social and economic conditions that influence the health of individuals and communities, such as income, education, employment, social support, and access to health care~\cite{lowcock2012social}.
\textbf{(6) Gauging the multifaceted impacts of the crisis:}
Several visualizations in our corpus depicted the initial impacts of the crisis (141 out of 668, 21\%), including \textit{government response and interventions}; \textit{economic disruptions} such as changes in unemployment, GDP, S\&P ratings, and business sales; \textit{social disruptions} due to school closures, quarantine, and lockdowns in response to government policies; and \textit{environmental impact} such as the change of urban pollution and vibration of the Earth's surface. The impact of the pandemic has been multifaceted and wide-ranging. Visualizations in this category attempt to convey the shifts in daily life due to the crisis, and also serve as evidence of the effectiveness of interventions designed to mitigate risks.
In summary, COVID-19 visualizations have been created to communicate a wide variety of messages. This message categorization provides an overall understanding of the focal areas of the visualizations in our corpus. In the next section, we use this message categorization to guide our in-depth examination of the visual approaches used to convey various messages about the pandemic.
\section{In what visual forms was the information presented?}
\label{sec:visual_forms}
We now turn to a discussion of how various visualization techniques and visual encodings have been used to create COVID-19 crisis visualizations that communicate each type of message identified above, along with some visualization examples. We also identify issues and challenges in existing COVID-19 visualizations, which will catalyze future research that tackles the challenges inherent within designing visualizations that address each type of messaging goal. Note that some screenshots used in this paper were cropped, and some visualization captions were not fully displayed here due to the space consideration. Care is needed when interpreting these visualization examples (e.g., to avoid potentially misleading messages or biases).
\subsection{Informing of the Severity}
\label{sec:inform_severity}
\begin{figure*}[tb]
\centering
\includegraphics[width=\textwidth]{fig/examples/geospatial.pdf}
\caption{Examples of geospatial visualizations:
(A) A choropleth map (not normalized) using a traditional projection plots the total number of cases (by the Centers for Disease Control and Prevention~\cite{cdc_choropleth_map});
(B) A choropleth map overlaying on a tile grid base map provides both the raw number version and a normalized view (by the USAFacts~\cite{usafacts_tilegrid_map});
(C) A proportional symbol map (bubble map) shows the total cases (by CNN~\cite{cnn_bubble_map});
(D) An alternative proportional symbol map shows the total deaths over time. Each triangle represents the number of deaths in a metropolitan area (by the New York Times~\cite{nyt_triangle_map}).
}
\label{fig:geospatial}
\end{figure*}
Visualizations that depicted the intensity of the pandemic mainly include traditional basic chart-based temporal visualization and map-based geospatial visualizations, and more advanced multivariate visualizations, as well as data-driven storytelling approaches.
\textbf{Temporal visualizations} focused on depicting the trajectory of the pandemic over time (see \autoref{fig:temporal}). The most basic solution was to visualize daily and cumulative numbers over a time period, mostly using bar charts, line charts, and area charts. Two major variations in visualizing the temporal change of the pandemic include using logarithmic (log) scales (e.g., \autoref{fig:temporal} (C-D)) and adding the moving average that minimizes random fluctuations in counts (e.g., \autoref{fig:temporal} (A)). By the time of our review, 14 visualizations provided both a log- and a linear-scale mode (10 enabled users to toggle between modes), 6 provided only log scales, and the rest used linear scales. The wide adoption of linear scale charts may be because they are easier to understand for the general public.
Novel visualization techniques were introduced, such as the Pez Charts~\includegraphics[height=\fontcharht\font`\B]{fig/pez_chart_new.pdf} as seen in ~\autoref{fig:temporal} (B), the Growth Charts~\includegraphics[height=\fontcharht\font`\B]{fig/trend_chart.pdf} as shown in ~\autoref{fig:temporal} (C), and the use of temporal event alignment.
The Pez Chart~\includegraphics[height=\fontcharht\font`\B]{fig/pez_chart_new.pdf}~\cite{pez_charts} aims to show case number changes in horizontally-placed blocks, with the x-axis representing time and the color of each block representing the values. When stacking multiple horizontal series together, it is easier to compare trends across regions. The first visualization that used the Pez Chart to show trends of the pandemic, in our corpus, was from the Los Angeles Times on April 1, 2020.
Moreover, \textit{temporal event alignment} is useful to reveal patterns that emerge by aligning the occurrence of events of interest over time~\cite{zhang2019eva}. The trajectory of the pandemic can be seen through an alignment of days because cases or deaths across regions (on the x-axis) are usually presented in a log scale. Comparing with the reference lines (i.e., the slope of a curve) helps demonstrate a doubling rate.
Another emerging solution is the \href{https://aatishb.com/covidtrends/}{Growth Chart}~\cite{aatishb_growth_chart} \includegraphics[height=\fontcharht\font`\B]{fig/trend_chart.pdf} (see ~\autoref{fig:temporal} (C)), which plots the total number of the confirmed cases as the x-axis and the weekly confirmed cases as the y-axis~\cite{aatishb_growth_chart}. Applying the log scale to both axes and using the total case number as the x-axis (rather than time) helps reveal trends and patterns (e.g., demonstrating if cases are exponentially growing in a region or if a region is on the path to containing the virus).
\begin{figure*}[tb]
\centering
\includegraphics[width=\textwidth]{fig/examples/multivariate.pdf}
\caption{Examples of multivariate visualizations:
(A) Multiple encodings to illustrate reported case change over time, including using the line thickness to show the current number relative to population and using line height and color to show the change in reported cases relative to a specific date (as a type of event alignment) (by the Washington Post~\cite{wp_sparkline_map});
(B) Concentric/ centered bubble maps showing both the positive and total tests (by POLITICO~\cite{politico_centricbubble}). }
\label{fig:multivariate}
\end{figure*}
Though these more advanced methods have been studied and used before in academic research, they were not initially used for communicating information about the pandemic with the general public. The patterns conveyed in these novel visualizations are particularly relevant.
\textbf{Geospatial visualizations} displayed variables of interest over geographical maps to demonstrate which regions were impacted and compare how the impact differs by regions. Among all the map-based visualizations we analyzed, choropleth maps were the most popular option (110 out of 218, 50\%), followed by proportional symbol maps (61 out of 218, 29\% bubble maps). ~\autoref{fig:geospatial} shows four types of representative maps.
Both choropleth and proportional symbol maps overlay information on a base map (i.e., maps containing reference information that may use different geospatial information depending on what the designer wants to communicate~\cite{dent2008thematic}). Most map-based visualizations in our collection used maps that represented the physical shapes of geographical regions. Traditional projections (e.g., Mercator or Robinson) are widely used, while a few others in our collection used the Cahill–Keyes projection ~\includegraphics[height=1.4\fontcharht\font`\B]{fig/cahill.pdf}~\cite{nyt_cahill_projection}, and the Armadillo projection \includegraphics[height=1.4\fontcharht\font`\B]{fig/armadillo.pdf} to provide a perspective with less distortions~\cite{scmp_armadillo_projection}. Different projections distort the globe in different ways, potentially impacting how the message is received by the audience. It was also common for visualizations to use tile grid maps, abstracting away the physical shape of geographical regions so that viewers can focus on the overlaying information.
While choropleth maps were the most frequently used geospatial visualization in our corpus, they do present a challenge in that color-coding raw data may mislead viewers. For example, if a highly-populated state has the same case number as a less-populated state, the color will be the same on the map. A reader may interpret the map as conveying that the intensity of cases is similar in the two states, while in actuality, the intensity in the less-populated state is higher. Prior work has shown that normalizing data may help remove this type of bias~\cite{dent2008thematic}. Typical normalization variations include normalizing by area, relevant population, a prior date, central tendency (e.g., mean, median, mode), and variability (e.g., standard deviation, above and below range)~\cite{foster2019map}. However, our findings suggest that only 37\% of the visualizations in our corpus used some form of normalization.
In addition, maps have applied various color schemes, including diverging (28\%), sequential (65\%), and a mix of diverging and sequential methods (7\%). Moreover, a majority of maps applied light background (e.g., white, light grey), but 17\% used dark background (e.g., black, dark blue). Prior work has examined color-related biases (e.g., dark-is-more, contrast-is-more bias) and the choice of background color~\cite{schloss2018mapping}. These design choices, while nuanced, may also impact how people perceive severity and risk in times of crisis.
\begin{figure*}[tb]
\centering
\includegraphics[width=\textwidth]{fig/examples/history_compare.pdf}
\caption{Example visualizations of comparing with historical events:
(A) A side-by-side table-like visualization that compares a variety of dimensions between the flu and COVID-19 (by Vox~\cite{vox_compare_history});
(B) A 3D view visualization that uses horizontal timelines along the Z-axis to present the history of pandemics (by Nicholas LePan~\cite{vc_hairball});
(C) A redesigned ``distorted'' 3D version of (B) shows the death toll of historical pandemics (by Ryu Sakai~\cite{medium_ryu}).}
\label{fig:history_compare}
\end{figure*}
\textbf{Multivariate visualizations} display two or more variables in a visualization, that aim to allow people to distill a complex multidimensional dataset and explore the patterns between variables.
In our corpus, 8 visualizations used multivariate approaches, such as superimposing small multiples on maps to show cumulative number of cases over time in different locations (e.g., sparkline maps), as shown in ~\autoref{fig:multivariate} (A), using animation to illustrate how things have changed over time, as shown in ~\autoref{fig:multivariate} (B), and using concentric bubble maps to show multiple metrics (e.g., positive and total tests) in absolute numbers. Though it is common to use multivariate visualizations in visualization research and public health data analytics among professionals, it was not commonly used in our corpus.
Visualizations that aim to provide \textbf{comparisons with historical events} mostly used table-like representations (\autoref{fig:history_compare} (A)) to compare characteristics between COVID-19 and other pandemics. Other visualizations embed 3D timelines~\cite{vc_hairball, medium_ryu} for easy comparisons, as shown in \autoref{fig:history_compare} (B-C).
\textbf{Crisis visualizations that \textit{turned to the felt experience}} were also designed to convey the severity of the pandemic, though unlike the previously-described data visualizations. For example, the visualizations created by the New York Times (\autoref{fig:griefloss}) had focused on personalization and emotions. These visualizations convey a sense of acknowledging grief and sadness resulting from the loss of lives due to the pandemic. Though we only found two visualizations in this category, they carried an important message that the death toll was not only a number---it reflects an indescribable loss of valued human life.
\textbf{Summary:} Our findings show various trends in the use of visual encodings, traditional chart-based techniques, and novel data visualization approaches within our corpus of COVID-19 visualizations. We identified some common issues and challenges within these visualizations, such as the lack of normalization. We also presented data-driven storytelling approaches to convey the severity of the pandemic. Future work should further examine the effect of different types of visualizations on people's risk perception, and identify the most and least effective approaches for visualizations.
\begin{figure}[tb]
\centering
\includegraphics[width=\linewidth]{fig/examples/grief.pdf}
\caption{Examples of acknowledging grief and sadness due to the loss of life: an interactive visualization using pictorial forms and scrollytelling, a visual storytelling approach that happens when visualization is revealed or changed as a user scrolls (by the New York Times~\cite{nyt_loss_interactive}). }
\label{fig:griefloss}
\end{figure}
\subsection{Forecasting Trends and Influences}
\label{subsec:forecasting}
\begin{figure*}[tb]
\centering
\includegraphics[width=\textwidth]{fig/examples/uncertainty_switch.pdf}
\caption{Examples of the uncertainty visualizations using different visual encodings at different time (by the University of Washington~\cite{uw_model}). (A) A visualization with uncertainty annotations with shades (April 7, 2020); (B) A visualization without uncertainty annotations (no shade) as the default view (September 11, 2020).}
\label{fig:uncertainty_switch}
\end{figure*}
\begin{figure*}[tb]
\centering
\includegraphics[width=\textwidth]{fig/examples/course_of_disease.pdf}
\caption{Example visualizations of explaining the nature of the crisis:
(A) A medical illustration created at the Centers for Disease Control and Prevention (CDC) reveals ultrastructural morphology exhibited by coronaviruses (by the CDC~\cite{cdc_coronavirus_illustration});
(B) A type of symptom-body map shows associated symptoms on the body shape (by The Globe and Mail~\cite{globemail_bodymap});
(C) A 3D visualization using scrollytelling vividly describes how possible transmission happens (by the New York Times~\cite{nyt_3d_simulation}).
}
\label{fig:course_of_diseaase}
\end{figure*}
Forecasting models intrinsically involve a certain level of uncertainty. Many visualizations in this category use graphical annotations to quantify uncertainty (48 out of 57, 84\%). Among the collected visualizations for forecasting and projection, major approaches to communicate uncertainty include: presenting prediction intervals (29 out of 48, 60\%), comparing projection results of multiple models or different scenarios (14 out of 48, 29\%), and communicating uncertainty in text (5 out of 48, 10\%).
Visualizing uncertainty intervals for projections (n=29) includes using solid line charts with shade \includegraphics[height=1.2\fontcharht\font`\B]{fig/solid_range.pdf} (16 out of 29, 55\%), a solid line indicating current and prior situations followed by a dashed line with shade indicating projection \includegraphics[height=1.2\fontcharht\font`\B]{fig/dashed_range.pdf} (6 out of 29, 21\% ), superimposing multiple levels of confidence intervals (4 out of 29, 14\%) shown with different opacity using interval funnels \includegraphics[height=1.2\fontcharht\font`\B]{fig/intervals_funnel.pdf} and interval range bars \includegraphics[height=1\fontcharht\font`\B]{fig/intervals_bar.pdf}, as well as showing confidence intervals with error bars (3 out of 29, 10\% for visualizing effective reproduction rate).
As more and more forecasting models are generated, visualizations are more commonly used to compare between different models or scenarios, e.g., using ensemble plots (n=9) and multiple hypothetical outcomes (n=5). The goal is to present differences in the assumptions and corresponding confidence intervals and to help the public understand the potential limitations of the forecasts. Multiple hypothetical outcomes can be useful for understanding various scenarios regarding what might happen in the future. For example, in our corpus, visualizations that juxtapose small multiple choropleth maps in the same view~\includegraphics[height=1.2\fontcharht\font`\B]{fig/map_small_multiples.png} show how different infection prevention and control measures might influence the outbreak's spread~\cite{nyt_sm_control}. This type of visualization may be able to intuitively help people understand the probability amidst this complex pandemic.
Among all the forecasting visualizations in our corpus, we noticed one case where the default view of the visualization switched from showing uncertainty annotations with shade (\autoref{fig:uncertainty_switch} (A)) to charts without uncertainty annotations (\autoref{fig:uncertainty_switch} (B)). Instead, users were able to toggle on and off the button \includegraphics[height=1\fontcharht\font`\B]{fig/toggle.pdf}.
\textbf{Summary:} Our findings summarize the design choices of whether to present uncertainties and the visual styles to communicate uncertainty among the COVID-19 visualizations. Future research needs to examine how to effectively communicate uncertainty to lay audiences to increase credibility and improve decision-making. Understanding the nuanced differences in visual encodings for uncertainty will help contribute to the increasing body of literature in uncertainty research.
\subsection{Explaining the Nature of the Crisis}
\label{sec:explain_course_of_disease}
\begin{figure*}[tb]
\centering
\includegraphics[width=0.95\textwidth]{fig/examples/risk_mitigation.pdf}
\caption{Example visualizations focused on guiding risk mitigation:
(A) An infographic designed by a group of volunteers in Japan using the cultural metaphor of tatami, to show how to engage in social distancing (by PANDAID~\cite{jp_tatami});
(B) A flatten-the-curve illustration (by Drew A. Harris~\cite{harris_flatten_twitter});
(C) An animated cartoon version of the flatten-the-curve illustration (by Dr Siouxsie Wiles et al.~\cite{siouxsie_flatten_animated}).}
\label{fig:risk_mitigation}
\vspace{-10pt}
\end{figure*}
Visualizations that \textbf{explained the causes, symptoms, and transmission} were mainly designed to facilitate communication of information generated by health professionals and scientists. For example, the medical illustrations team from the Centers for Disease Control and Prevention (CDC) aimed to make abstract medical concepts more approachable and comprehensible, as shown in~\autoref{fig:course_of_diseaase} (A):
\textit{``The colours were chosen for visual impact. The bold red of the S proteins contrasted by the gray of the viral wall, adds a feeling of alarm... Shadows add to the realism''}~\cite{cdc_coronavirus_illustration}. The choices of colors and shadows may be meant to spark the felt experience (i.e., how visualizations are felt by the users).
Besides, visualizations have also been created to help people \textbf{understand and compare symptoms} between COVID-19 and other conditions (e.g., allergies, cold, and, influenza). Side-by-side table-like visualizations have been widely used for symptom comparison. Some characteristics of this type of visualizations include using categorical color encoding to represent various frequencies of symptoms (like rare, sometimes, and common), and embedding pictorial aids next to the text. While side-by-side tables are easy for comparison, \textit{symptom-body maps} that map associated symptoms on the body shape (e.g.,~\autoref{fig:course_of_diseaase} (B)) help make the invisible become visible.
Messages that aim to \textbf{explain transmission} focus on explaining how the virus spreads or the various ways in which it can spread. For example, a 3D simulation visualization, as shown in ~\autoref{fig:course_of_diseaase} (C)~\cite{nyt_3d_simulation}, vividly explains the possible transmission routes and attempts to persuade people to keep social distance. To that end, these visualizations also aim to help people \textit{understand the reasonableness} of recommended actions (e.g., social distancing) and reinforce personal responsibility to reduce harm to other individuals and the community~\cite{reynolds2007cdc}.
\textbf{Visualizations of contact tracing} can be classified into explaining the process behind contact tracing and visualizing individual-level or group-level tracing. At an individual level, visualizations illustrate patient movement paths and timelines \includegraphics[height=1.2\fontcharht\font`\B]{fig/individual_contact_tracing.pdf} to explain how an individual might have come into contact with a person infected with the virus. At the group-level, node-link networks~\includegraphics[height=1.2\fontcharht\font`\B]{fig/node-link.pdf} explore clusters of core groups or super spreaders.
\textbf{Summary:} Communicating scientific research to the general public is important for scientists, yet it is also challenging to effectively communicate with diverse audiences~\cite{brownell2013science}. Visualization designers have made efforts to help bring science-based information closer to the public. Prior work suggests using pictorial aids in health communication is an effective way to facilitate understanding and improve recall of medical instructions~\cite{arcia2019helping}. Though it is promising to incorporate pictorial forms to help interpret textual explanations, it is important to keep in mind that these visual aids might be too complex to understand or they may fail in reflecting viewers' expectations~\cite{katz2006use}.
\subsection{Guiding Risk Mitigation}
\label{sec:guiding_risk_mitigation}
\textbf{Providing instructional guidance} was one of the common approaches to guiding the public in risk mitigation. Infographics, especially instructional graphics, have been produced and distributed widely to provide instructional guidance on mitigating strategies, such as how to maintain social distancing, use personal protective equipment~(PPE), and maintain hygiene.
Several visualizations (n=6) have incorporated some sort of cultural metaphors in the visualizations that aim to guide people on how to avoid risks.
For example, social distancing guidance designed by a Japanese organization adopted the metaphor of tatami (a type of mat in traditional Japanese-style rooms) to guide people to ``stay one tatami apart''~\cite{jp_tatami} (see ~\autoref{fig:risk_mitigation} (A)), while a Florida county reminded people to ``stay one alligator apart''~\cite{alligator_social_distancing}.
\begin{figure}[tb]
\centering
\includegraphics[width=\linewidth]{fig/examples/wp_health_disparity_map.pdf}
\caption{A set of choropleth maps show how COVID-19 further demonstrates existing health disparities in communities of color (by the Washington Post~\cite{wp_health_disparities}). }
\label{fig:wp_health_disparities}
\end{figure}
\begin{figure*}[tb]
\centering
\includegraphics[width=\textwidth]{fig/examples/equity.pdf}
\caption{Examples of visualizing risk, vulnerability, and equity:
(A) Scatter plots using scrollytelling show people who face the greatest risk of exposure to coronavirus according to the physical proximity to others (by Lazaro Gamio~\cite{nyt_vulnerable_occupation});
(B) 3D visualizations vividly depict the challenge of sheltering in limited spaces (by Reuters~\cite{reuters_small_shelters});
(C) A scrollytelling visualization explains approaches to rationing care, asking the question \textit{``when medical resources are limited, who should get care first?''} (by Feilding Cage~\cite{reuters_rationing_care}).
}
\label{fig:equity}
\end{figure*}
\textbf{Conceptual flattening-the-curve} charts (e.g.,~\autoref{fig:risk_mitigation} (B-C)) have become prevalent, particularly with the addition of a horizontal line marking ``healthcare system capacity''. We found 21 different versions of flattening-the-curve charts that aim to conceptually depict the management of the healthcare systems' capacity. They appear to emphasize personal responsibility in minimizing the unprecedented strain on the health system. However, these type of visuals have become controversial amidst the pandemic as they simplify complex pandemic situations (e.g., implying that it is good enough to keep the patient counts below the healthcare system's capacity) and they may also be misinterpreted by the public.
\textbf{Summary:} Though we did not systematically analyze the cultural elements in our corpus of visualizations, it is important for future work to examine cultural differences in the design of crisis visualizations, and the effect of culturally-tailored visualizations on people's risk perceptions and behaviors. Moreover, more work is needed to examine the effectiveness of data-driven empirical visualizations as compared to conceptual visuals like the flatten-the-curve illustrations.
\subsection{Communicating Risk, Vulnerability, and Equity}
\label{subsec:equity}
Our findings highlight a body of visualizations (7 out of 81, 9\%) that \textbf{used aggregated scores to communicate risks} (i.e., using a single value to reflect risks arising from multiple factors), such as \href{https://svi.cdc.gov/map.html}{the Social Vulnerability Index (SVI)} and Behavioral Risk Factor Surveillance System (BRFSS). The resulting visualizations were all choropleth maps (e.g., \autoref{fig:wp_health_disparities}), that revealed health disparities caused by multiple factors including majority-minority, highly vulnerable, overcrowded households, and most uninsured~\cite{wp_health_disparities}.
An increasing number of visualizations have added \textbf{disaggregated demographic data in their existing visualizations} as governmental authorities began to release demographic information of patients. Bar charts were used to display demographic distributions of affected populations. Age distribution was the most commonly visualized (n=39), followed by race (n=27) and gender (n=23). Some visualizations (n=5) unpacked the \textbf{influence of underlying health conditions} to understand the characteristics of affected population, and the likelihood of people seeking intensive care treatment.
Limited access to healthcare is another risk factor for severe disease~\cite{lowcock2012social}. People who have limited access to health care resources are likely to be more vulnerable during a public health crisis~\cite{reynolds2007cdc}. To help understand these factors, visualizations have been created to show \textbf{accessibility and allocation of resources} (n=4), such as visualizing the travel distance to nearest hospitals to get treatment and displaying the geospatial spread of tangible social support (e.g., food services).
In addition, visualizations have been created to \textbf{examine associations of socioeconomic status~(SES) and the pandemic} (15 out of 81, 19\%). Various types of visualizations have been adopted in this category, including scatter plots for displaying the relationship of physical proximity and the exposure to diseases, bar charts to show ranked order of jobs based on forced policies of physical distancing, and index charts that compare high and low income disparities in movement change.
Despite the effort to unpack and address the vulnerability and equity issues, we also found that visualizations in this category were not equally created across visualization outlets. In our corpus, visualizations created by the government agencies only show basic population distribution by age, race, and gender, with two visualizations displaying underlying condition. However, no government created visualizations examined other SES-related factors. Instead, work examining SES, accessibility to resources, and living conditions were created by the news outlets, independent media, companies, and NGOs, indicating the effort being put in revealing vulnerability and equity challenges, as shown in \autoref{fig:equity} (A-B). Scrollytelling visualizations have also been designed to explain who should get access to medical services first when resources are limited, e.g., sickest first, an equal chance, and maximizing treatment benefits~\cite{reuters_rationing_care}, as shown in \autoref{fig:equity} (C). Such visualizations highlight a socioeconomic challenge in terms of imperfectness of the healthcare system and may shed light on how to reform and improve the system.
\begin{figure*}[tb]
\centering
\includegraphics[width=\textwidth]{fig/examples/crisis_impact.pdf}
\caption{Example visualizations of gauging the multifaceted impacts of the crisis:
(A) A superimposed line chart shows how COVID-19 stalled climate change momentum, using single event alignment technique with the vertical line indicating the alignment event of Chinese Lunar New Year (by the Financial Times~\cite{ft_lunar_alignment});
(B) A line chart visualizes how the level of the earth's surface vibration has changed over time during the lockdown periods. Shaded areas indicate the interval-event duration (by the National Geographic Society~\cite{ng_earthquite});
(C) A ``misleading'' V-shape-like line chart that went viral on Twitter (by CNBC~\cite{CNBC_tweet}).
}
\label{fig:impact_crisis}
\end{figure*}
\textbf{Summary:} Our findings show that a number of visualizations created by a variety of information outlets help communicate risks, and unpack issues around vulnerability and equity amidst the COVID-19 crisis. It is important to communicate that some populations face greater risks than others because it highlights the inequities that exist within society and calls attention to areas in which concerted efforts are needed to close gaps in terms of how different communities are impacted by a crisis. However, when visualizing risk, vulnerability, and equity, one caveat is that deliberate efforts need to be made to prevent and counteract stigmatization that can arise by focusing on how the pandemic is negatively impacting a group.
\subsection{Gauging the Multifaceted Impacts of the Crisis}
\label{sec:gauge_impact}
We identified two challenges for visualizations that aim to demonstrate the impacts of the pandemic on human lives and society.
The first challenge is building the connection between the theme of the visualization and the major events amidst the crisis. Techniques that help building such connections include \textit{event alignment}, such as aligning by Lunar New Year to show the change of coal consumption in China, as shown in ~\autoref{fig:impact_crisis} (A), and event annotations that indicate when key events occurred, including \textit{point-event indicators without duration} (e.g., first death in Wuhan) and \textit{interval-event indicators} (e.g., duration shutdown order in effect), as shown in ~\autoref{fig:impact_crisis} (B).
The second challenge is how to vividly tell a story about the effect of certain interventions (e.g., public health and government interventions) without misleading audiences. One ``misleading'' visualization example that went viral on social media was the V-shape-like line chart presented by CNBC (see \autoref{fig:impact_crisis} (C)) to support the claim that the job market was recovered. The problem was the inappropriate choice of chart type. Many social media users pointed out this ``misleading'' chart issue in the thread of the discussion and proposed alternative design solutions like using various bar charts to improve the design.
\begin{figure*}[tb]
\centering
\includegraphics[width=\textwidth]{fig/examples/nyt_change_over_time.pdf}
\caption{Examples of how crisis visualizations have changed over time by the \textit{New York Times}~\cite{nyt_dashboard}. (A) A proportional symbol map (i.e., bubble map) superimposed on a qualitative map, published on March 3, 2020; (B) A bubble map published on April 8, 2020; (C) A choropleth map with categorical and sequential color schemes, published on May 28, 2020; (D) A choropleth map (with updated legend and menu), published on August 19, 2020.}
\label{fig:nyt_change_over_time}
\end{figure*}
\textbf{Summary:} Though the visualizations in this category seem to be general and can be designed at any time, we argue they are particularly important and relevant to crises. The design and dissemination of ``misleading'' visualizations in times of crisis pose extra challenges to the general public as these deceptive visualizations may impact people's trust of officials and organizations, and people's decision making. Considering the public engagement with existing crisis visualizations, we see how visualizations can be improved for more effective and responsible communication.
\jz{
\subsection{Narrative Visualization Approaches among COVID-19 Visualizations}
One growing area of work in the visualization space is that of narrative visualization~\cite{Kosara2013Story}, that is, ``a genre that combines interaction techniques for exploratory control over insights gained and communicative, rhetorical, and
persuasive techniques for conveying an intended story''~\cite{hullman2011visualization, Lee2015More}. As a means of characterizing the growing landscape of narrative visualizations, Segel and Heer conducted a design space analysis that categorized visualization genres, visual narrative tactics, and narrative structure tactics~\cite{Segel2010Narrative}. Hullman and Diakopoulos's~\cite{hullman2011visualization} work on visualization rhetoric provided an analytical framework to help people understand how design cues facilitate prioritization of particular interpretations in visualizations. Moreover, Moere et al. \cite{moere2011role} argued for making the information visualization research more inclusive by incorporating more reflection and critiques from visualization practice work (e.g., visualization activities that are conducted by commercial enterprises and freelance designers).
Though not all visualizations in our corpus are narrative visualizations ~\cite{Lee2015More}, some have applied narrative visualization approaches, such as annotation, animation, and scrollytelling. Building upon prior narrative visualization research, our work examined how narrative visualization approaches were applied in COVID-19 visualizations, with a focus on visualizations produced by practitioners (i.e., those available to the general public). }
\jz{
In the preceding sections, we have summarized how various visualization forms have been used to communicate a range of messages in the COVID-19 pandemic (\autoref{sec:inform_severity} to \autoref{sec:gauge_impact}). Beyond the visual encodings and techniques discussed thus far, we also found that narrative visualization techniques were variably used to communicate COVID-19 messages. For example, animation was most often used (39\%) to depict the severity of the pandemic by showing how fast coronavirus has spread across countries and how current cases have changed over the course of the pandemic. Animation was also used in other contexts, although less frequently. Specifically, animation was used in visualizations designed to: explain virus transmission (17\%), unpack health equity and vulnerability (17\%), guide people towards mitigating risks (13\%), gauge the impact of the crisis (10\%), and forecast trends and influences (4\%). Interestingly we found that scrollytelling techniques were only present in visualizations created by news outlets (e.g., the Washington Post, the New York Times, and Reuters).
}
\begin{figure*}[tb]
\centering
\includegraphics[width=\textwidth]{fig/examples/GA_misleading.pdf}
\caption{Examples of a ``misleading'' visualization and its corrected version by the Georgia Department of Public Health~\cite{GA_misleading}.
(A) A bar chart aims to show that daily cases were declining in five of the state's hardest-hit counties in Georgia. But note that the dates were arranged out of order;
(B) A corrected chart of (A) after being accused of data manipulation using misleading visualizations.}
\label{fig:GA_misleading}
\end{figure*}
\section{Under What Circumstances? Crisis Visualizations within Dynamic Temporal Contexts}
\label{sec:contexts}
So far, we have presented four components of our framework to examine COVID-19 visualizations. In this section, we draw attention to the complex and dynamic temporal contexts of the pandemic. Our findings show that 36\% of the visualizations in our corpus (243 out of 668) have been updated within our review time.
We categorize these changes as being \textbf{crisis-driven}, \textbf{public-driven}, and \textbf{content producer-driven}. These categories are not likely to be mutually exclusive nor do they possess strict boundaries, and they are inherently subjective given that we are not privy to the intentions designers held as they were creating the visualizations. Indeed, in our discussion below, we do not report on the \textit{intentions} of visualization creators as these visualizations were changed, but rather the ways in which the changing visualizations in our corpus \textit{reflect} the dynamic nature of: the pandemic (``crisis-driven changes''), public response to created visualizations (``public-driven changes''), and the strategies used to create visualizations (``content producer-driven changes''). The benefits of categorizing such changes over time, are that it helps us understand how temporal contexts play a role in visualization design, and enables us to highlight challenges and opportunities in visualization and crisis informatics research.
\subsection{Crisis-driven Changes}
Visualization changes can reflect an urgent need to adapt existing visualizations to better suit the constantly changing crisis situations; we refer to these changes as \textbf{``crisis-driven'' changes}. Existing visualizations may be insufficient to meet the challenges of visualizing crisis information as the crisis develops for a number of reasons, such as a change of data format released by health officials, or the inability of existing visualizations to reveal trends and patterns of the pandemic due to increasingly complex situations.
For example, \autoref{fig:nyt_change_over_time} shows four US maps that represent different stages of COVID-19. At the beginning of the pandemic, an early-stage map (see~\autoref{fig:nyt_change_over_time} (A)) applied dual encodings including proportional symbols (i.e., bubble map) and a qualitative thematic map (areas marked with yellow background) to show COVID-19 cases. Later on, in April, the default map was changed to a bubble map as COVID-19 cases had spread rapidly all over the country (see~\autoref{fig:nyt_change_over_time} (B)). In May, although the overall US case numbers were decreasing, some states witnessed an increasing number of cases again. To better reveal the trends and patterns, the default map changed to a choropleth map showing hot spots of how the number of new cases had changed in the last two weeks. The map classification (shown as the map legend at the top) fell into four categories: falling~\colorRect{colorLegendBlue}, about the same~\colorRect{colorLegendYellow}, rising (with three sequential colors~\colorRect{colorLegendLightOrange}\colorRect{colorLegendOrange}\colorRect{colorLegendRed}), and few or no cases~\colorRect{colorLegendLightGrey}, as shown in~\autoref{fig:nyt_change_over_time} (C). The granularity also changed from states in (A) to counties in (C) to contrast different trends in some rural areas and cites.
A more recent version of the map classification in~\autoref{fig:nyt_change_over_time} (D) has changed to an eight-level sequential color scheme~\colorRect{colorLegendYellow}\colorRect{colorLegendNYTLevel2}\colorRect{colorLegendNYTLevel3}\colorRect{colorLegendNYTLevel4}\colorRect{colorLegendNYTLevel5}\colorRect{colorLegendNYTLevel6}\colorRect{colorLegendNYTLevel7}\colorRect{colorLegendNYTLevel8}, along with the category of few or no cases~\colorRect{colorLegendLightGrey}, since the rising second wave of cases has became the theme in August 2020.
In addition, adding the moving average lines over a time-series chart was one of the most common changes in our corpus, as shown in ~\autoref{fig:temporal} (A). Prior work has suggested that the moving average provides a more stable view of the trend than daily change and helps to convey the effectiveness of COVID-19 surveillance and containment amidst the pandemic~\cite{ng2020evaluation}. Most visualizations in our corpus that were created early on in the pandemic did not use moving averages. To our knowledge, the earliest use of moving averages was from a Singaporean report on February 29, 2020~\cite{ng2020evaluation}. Later on, more visualizations started adding moving averages to better reflect the current state of the pandemic (n=30). For example, the New York Times added the 7-day average starting from April 8, 2020. This may be because the COVID-19 case and death data started to show apparent periodic fluctuation from late March to early April.
\begin{figure*}[tb]
\centering
\includegraphics[width=\textwidth]{fig/examples/GA_change_over_time.pdf}
\caption{Examples of how crisis visualizations have changed over time, by the Georgia Department of Public Health~\cite{GA_covid}:
(A) An early version of the COVID-19 dashboard (March 25, 2020) contains one (non-normalized) choropleth map and two pie charts showing the break-down COVID-19 data by age and sex;
(B) A version published in May includes choropleth maps (both showing the raw case numbers and normalized data) showing cases and deaths change over time (bottom left), and a set of bar charts showing the break-down COVID-19 information by sex, race, and ethnicity;
(C) A more recent version (August 29, 2020) provides more views to show daily status.
}
\label{fig:GA_change_over_time}
\end{figure*}
\subsection{Public-driven Changes}
Some changes in visualizations are prompted by the public participation and critique of visualization works; these changes related to the feedback from the public and can be termed \textbf{``public-driven'' changes}. The growing interests and practices of designing and disseminating COVID-19 visualizations have led to increasing public critiques and discussions on social media (e.g., Twitter and Reddit). This sort of collective intelligence from the public has called attention to necessary changes. For instance, ~\autoref{fig:GA_misleading} (A) shows a controversial example of potentially ``manipulative'' data using ``misleading'' visualizations to show a downward trend in cases~\cite{GA_misleading}. The Georgia Department of Public Health released a graph on May 17, 2020, showing COVID-19 cases had supposedly fallen among five hardest-hit Georgia counties from April 26 to May 9, with the dates on the chart being out of order, as shown in \autoref{fig:GA_misleading} (A). Many social media users critiqued this misleading visualization and pointed out the mistakes. The graph has been taken off the website since then. A spokeswoman for Gov. Brian Kemp tweeted on May 11, 2020: \textit{``The x axis was set up that way to show descending values to more easily demonstrate peak values and counties on those dates. Our mission failed. We apologize. It is fixed.''}~\cite{GA_tweet}. The team amended the error by replacing it with the chronological order, as shown in \autoref{fig:GA_misleading} (B). The corrected chart shows that cases were holding steady rather than declining.
It is difficult to assess the intentionality of the visualization design; one could argue that the graph was manipulated to intentionally deceive or was mistakenly created by the visualization tool. Yet the effect of such visualizations may have spawned misunderstanding and misinformation that will ultimately influence people's trust with authorities and their own decision-making processes when dealing with the crisis (e.g., they underestimate the severity of the crisis). As such, this example shows how public participation and critiques of misleading visualizations can be powerful ways of bringing immediate and effective changes to visualization designs.
\subsection{Content Producer-driven Changes}
Another type of change can be framed as \textbf{``content producer-driven changes''}---changes that are mainly made in terms of the visual encodings (e.g., color and shape). For example, despite the errors they made in previously published visualizations, the Georgia Department of Public Health has also made numerous improvements in COVID-19 visualizations. \autoref{fig:GA_change_over_time} provides a snapshot of how their daily status visualizations have changed from early March to late August 2020. There are several major changes in how they have visualized COVID-19 information, including changing color schemes, normalization methods (e.g., from non-normalized choropleth maps to normalized ones), and basic chart types (e.g., from line charts to bar charts to show how case numbers have changed over time).
\textbf{Summary:} In this section, we have presented several examples of changes being made to visualizations amidst the pandemic. Examining how visualizations have changed amidst the dynamic nature of the pandemic highlights a broader challenge in designing and evaluating crisis visualizations. That is, how can we incorporate and embrace visualization changes over time, while also being critical of and responsible in implementing such changes? We will further discuss the challenges and opportunities in the design and evaluation of crisis visualizations shortly in ~\autoref{sec:discussion}.
\section{General Discussion}
\label{sec:discussion}
\jz{A crisis is characterized by its intense and broad impact, high risk, urgency, fast-evolving nature, and high level of uncertainty~\cite{shaluf2003review}. As we have seen with the COVID-19 pandemic, these intersecting characteristics of a crisis can lead to a large number of publicly-available visualizations, all generated from different data sources and with the great potential to collectively impact public behavior. Furthermore, COVID-19 has demonstrated how the fast-evolving nature of a crises can result in constant changes in visualization techniques and messages. In this paper, we have described four components of our conceptual framework for crisis visualizations: \textit{who}, (uses) \textit{what data}, (to communicate) \textit{what message}, \textit{in what form}, and have presented how visualizations have changed over the course of the COVID-19 pandemic. Visualization creators have updated, adapted, and transformed the crisis visualizations regularly to suit various needs, different data, and to answer ever-changing public questions. From the consumer's side, there is also a tendency to view these crisis visualizations over and over, to monitor the situation rather than to understand it once.}
In this section, we further discuss our framework regarding its scope and utility, as well as a set of research directions drawing on our observations and analysis of existing COVID-19 crisis visualizations.
\subsection{Who}
Though we have briefly described the information outlets that have produced visualizations amidst COVID-19, we believe there is a need to further examine issues around their positioning and power (i.e., ``the current configuration of structural privilege and structural oppression in which some groups experience unearned advantages''~\cite{dignazio2020seven}). The increasing political and economic interests of news outlets and those individuals and entities who have the power to control the outlets pose challenges to neutrality in their reports~\cite{elejalde2018nature}.
Media bias (i.e., an internal bias exhibited in media coverage)~\cite{hamborg2019automated} impacts how people perceive the risks and the severity of the crisis~\cite{slovic1987perception}. As visualizations become more accessible and popular, there is a vital need to investigate to what extent visualization designs systematically vary between different media groups (e.g., those with different political biases).
Our findings suggest that different media outlets have shown different patterns in communicating various messages and message components (e.g., visualizations focused on socioeconomic status were all from news media). Future work should further examine these and other patterns in crisis visualization variation across information outlets. For example, how do left and right-leaning media outlets choose their color schemes for maps that seek to inform people of the severity of the crisis? How might media biases be related to visualization color choices? These questions are important to examine, as a color has different cultural meanings in different countries, and color can elicit an emotional response~\cite{welhausen2015visualizing}. Using different color schemes may influence how people perceive risks and the severity of the crisis. The importance of exploring this aspect is further supported by a recent study focused on COVID-19 visualizations~\cite{cay2020understanding}. The study shows that the choice of colors and design can greatly impact users' risk perception. Understanding the differences and similarities of visualization design across media outlets may help us examine media biases, positioning, and power issues inherent in widely distributed visualizations.
\subsection{What Data}
Issues of crisis data sources are prevalent, as we described in ~\autoref{sec:data}. Similar to the positioning and power issues reflected in analyses of ``who'' designs the visualizations, the same concerns also extend to ``what data'' has been used for creating these visualizations. Yet, the issues surrounding data sources do not only pertain to COVID-19. Research on historical crises (e.g., SARS, H1N1, and Ebola) have already documented similar issues with data sources, such as varied data formats for reporting and dissemination~\cite{cori2017key}. Each of these threats to data quality in turn threatens the integrity of resulting visualizations. As such, care must be taken when visualizing crisis data with varied quality. For example, we emphasize the importance of reporting on the source (and even the source of source) and recency of data. Crises such as the COVID-19 pandemic introduce particular threats to the production of trustworthy visual information. Having a trustworthy data source is crucial for creating visualizations, especially in times of crisis where misinformation and disinformation are rampant. As we described in~\autoref{sec:data}, 7\% of visualizations did not cite the original data sources, leaving the audiences uncertain about where the data came from. In addition to reporting data sources, we highlight the importance of reporting the recency of data, that is, the date of data collection and visualization production. And, our discussion of how COVID-19 visualizations have changed over time, highlight how visualizations produced at one point in time may not be sufficient for depicting the state of the crisis at a later time.
\subsection{What Message}
We have presented six high-level categories of messages that communicate information about COVID-19, including informing of severity, forecasting trends and influences, explaining the nature of the crisis, guiding risk mitigation, gauging the multifaceted impacts of the crisis, as well as communicating risk, vulnerability, and equity (\autoref{sec:messages}). Through our inductive and deductive analysis approach, we derived a categorization that paints a holistic picture of the messages within our corpus of visualizations. At the same time, given that these categories were derived from our specific COVID-19 visualization sample, future work should explore to what extent this categorization needs iteration or expansion to fit future crisis contexts.
\jz{By categorizing the messages conveyed in COVID-19 visualizations, we aim to showcase their diversity and catalyze research that evaluates them and their impact. Our analysis shows a large spectrum of crisis-related messages being conveyed and many visualizations used for each message. While crisis and non-crisis public health visualizations may convey some similar messages, COVID-19 provides a unique opportunity to develop the message categorization we present in this paper. Such categorizations are more challenging in other public health contexts where there are fewer visualizations produced.
In addition, some messages are especially salient in crisis visualizations. For example, these communication tools have a particular focus on helping people gauge the crises' impact, given the intense and broad influences that they can have on the world. As another example, when communicating severity, risk, and equity, crisis visualizations may especially highlight medical resources, since such resources are especially crucial in these contexts. Also, the amount of visualizations communicating each message fluctuates as crises evolve. For instance, as public interests shift, we see more COVID-19 visualizations have turned to conveying felt experience and acknowledging grief and uncertainty.}
Moreover, our categorization of messages (all starting with verbs) provides a high-level task analysis that can be helpful to guide task abstraction and visualization design. Conducting task abstraction is based on task analysis. The goal of task abstraction is to recast user tasks and goals from domain-specific languages to a generalized terminology to allow for better understanding and readability of domain tasks~\cite{brehmer2013multi}. For example, building upon our categorization of high-level messages and tasks for crisis visualizations, future work could consider using existing task abstraction frameworks, ranging from high-level~\cite{shneiderman1996eyes} to low-level abstraction approaches~\cite{amar2005low}, or works that aim to bridge between the high- and low-level abstraction (e.g.,~\cite{brehmer2013multi, zhang2019idmvis}). Our work represents an important first step towards categorizing a set of tasks, or goals, for crisis visualizations, which will help future work create task abstraction that can further guide the choice of crisis visualization forms. \jz{General guidelines may be largely applicable to visualization design amidst a crisis. However, additional considerations may be needed to suit specific crisis contexts; our framework and analysis can help support the building of such guidelines, by focusing attention on the unique features of crises.}
\subsection{In What Form}
We have characterized patterns and trends in our corpus as well as challenges inherent in the design of such crisis visualizations (see \autoref{sec:visual_forms}).
Our analysis provides both a horizontal overview that compares and contrasts across visualizations aiming to communicate the same type of messages (even with the same data source), as well as a vertical overview that examines how visualizations have changed over time. Our findings can help to bootstrap the creation of the next generation of crisis visualizations, by characterizing previously used visual techniques along with the potential pitfalls of such approaches. These findings can help future visualization designers to avoid reinventing the wheel, thus accelerating innovation and the more efficient production of visualizations that communicate crisis information.
As the COVID-19 pandemic evolves, there may be new opportunities and imperatives to develop novel visualization methods that help people make sense of trends and patterns within the pandemic. Yet, using new visualization methods may make it challenging for people to interpret the message accurately. For example, using a log scale may help better display a trend. However, one study found that people had a less accurate understanding of the trajectory of the pandemic when showed the number of deaths on a log scale~\cite{romano2020covid}. Therefore, care must be taken when using novel visualization techniques or approaches that the general public is less familiar with.
However, we should also think critically regarding the production of crisis visualizations. The great diversity and the number of visualizations in our corpus reflect an increasing public interest in visualization. Yet, this proliferation of visualizations has also led to issues such as the varied quality among the visualizations reflected in our corpus. Thinking back to the ``misleading'' visualization example from the Georgia Department of Public Health (\autoref{fig:GA_misleading}), the website has been visited 4.93 million times according to Similarweb (a website traffic tracker)~\cite{similarweb}. We cannot say for certain that the chart was created intentionally or unintentionally (e.g., due to lack of knowledge or technical mistake). But these charts may have caused negative consequences to people who had seen them (e.g., not following public health recommendations). However, such ``unintended consequences'' may be able to be anticipated in advance~\cite{parvin2020unintended}. Neglecting such unintended consequences may cause more serious issues and may become a way to marginalize the ethical questions at the root of design decisions.
Therefore, instead of advocating for a single yes or no response to the question of whether or not we should encourage more crisis visualizations, why not think about, for example, how we can design more effective crisis visualizations through collaborations between experts in design, public health, and crisis informatics? Though our work discusses a set of visualization techniques used, issues, and challenges in the context of crisis visualizations, more work is needed to build a more comprehensive glossary in this field. For example, building upon our work, future work can compile a list of visual components, encodings, and visualisation techniques that are specifically related to the crisis context. Such glossaries and visualization ``cheatsheets'' would be helpful for people who typically are under time pressure during crises, to create crisis visualizations more effectively and efficiently.
\subsection{Under What Circumstances}
Our findings show how visualizations have changed over time, including crisis-driven, public-driven, and content producer-driven changes. Our exploration of the dynamic temporal contexts in crisis visualizations offers a novel way to reflect on history, as visualization changes over time offer a means of reflecting on crisis events. Such a record of crisis visualizations and reflection upon them can be seen as a way of building \textit{cultural heritage} (i.e., artifacts and memories that help to characterize our society's experience of the COVID-19 pandemic), which in this context, can be seen as building up and examining our sense of cultural identity. Exploring cultural heritage involves ``renegotiating our identities and value systems by reworking the meanings of the past as the cultural, social and political needs of the present change''~\cite{liu2010grassroots}. As such, studies of visualization changes over time can be an important cultural reflection process, and we hope that our collection of visualizations, analysis of them, and the resulting framework can help to catalyze the preservation of important cultural records and future reflections on these records. On the one hand, visualizations need to be adapted to better present and communicate the constantly changing information as a crisis develops. On the other hand, after decades have passed, the documented changes of visualizations can help us better recall how the crisis developed ``back then''.
Moreover, as we mentioned before, the general public on social media participated in detecting, criticizing, and redesigning existing visualizations, including those ``misleading'' visualizations. Their collective critiques and redesign of these crisis visualizations---emerging from collective knowledge building---contributes a collective memory that may shape the form of cultural heritage. Scholars have pointed out that research examining how memory-focused technologies can operate at a collective and cultural level is an underexplored field ~\cite{liu2010grassroots, van2008technologies}. The ``grassroots heritage'' framework proposed by Liu~\cite{liu2010grassroots} offers inspiration for exploring the temporal aspects of visualizations (e.g., how these visualizations change, and why). For example, some concepts described in the grassroots heritage framework include an open collection of content on a societal scale, and fostering interpretation of these collections, which are in line with the scope of our work.
People are increasingly engaged with visualization design in times of crisis using social media. As such, \textit{people are changing culture as they participate}. We encourage future work that explores the public's engagement with crisis visualizations, the resulting changes to those visualizations, and how such participation in the building of cultural heritage amidst a pandemic might impact the public (e.g., by helping or hindering individuals' ability to cope with and make sense of the crisis).
\subsection{To Whom}
\textit{``To whom''} refers to the receiver of the message or an audience~\cite{lasswell1948structure}. People use visualization to communicate a message. A challenge of communication during a crisis is that not all populations can be reached effectively and equally~\cite{reynolds2007cdc}. It is more challenging for special populations to assess information due to various constraints, such as cognitive, physical impairment or \jz{sensory impairment~\cite{COVID_visual_impaired, Holloway20Nonvisual}}, language barriers, and lack of devices~\cite{reynolds2007cdc}. The effects of crisis communication through visualizations also depend on the demographics, literacy, numeracy, and personal traits of the audiences~\cite{borner2019data}. Moreover, the relationship to \textit{``what message''} and \textit{``who''} also plays a role in effecting visualization communication, due to pre-existing confirmation biases (e.g., strong distrust of organizations distributing messages or visualizations)~\cite{reynolds2005crisis, van2019communicating, zhang2020ci}. All these factors associated with the characteristics of the audiences and the relationship and trust between the targeted audiences and information outlets should be considered when tailoring visualization design and dissemination in crisis communication. Future work can make use of our categorization of messages and example visualizations to investigate how crisis visualizations are perceived by different populations.
\vspace{-5pt}
\subsection{With What Effect}
Though our work did not examine the effects of crisis visualizations generated in times of COVID-19 on the general public, there is a critical need to evaluate such effects. Indeed, a great body of prior work has explored the effects of visualization on various aspects of people's risk perception and decision-making process~\cite{lipkus2007numeric}. While our findings show that there have been many variations of emerging novel visualization encodings and techniques raised amidst the pandemic, there is still a need to further examine how visualization impacts people.
Our work provides a collection of COVID-19 visualizations and a basic framework to accelerate future work to examine how various visualization encodings and techniques within each type of message impacts people's cognition, emotion, trust, behaviors, mental health, and decision-making processes. Understanding these effects will help us design and communicate crisis information more effectively in the future.
In addition to the traditional usability and performance metrics for visualization evaluation (e.g., accuracy, completion time, and memorability), we suggest future work explore dimensions of the felt and lived experience. The examples of data-driven storytelling approaches of acknowledging grief and misery of the loss (described in \autoref{sec:inform_severity} may spark inspiration for future evaluation studies. Our notion is in line with some visualization scholars~\cite{boy2017showing, kennedy2018feeling, peck2019data, saket2016beyond} that also indicated the importance of paying attention to the impact of emotion in visualization. Yet, we also argue that there is more to explore regarding various aspects of assessing felt and lived experiences ``in the wild'' especially in times of crisis, where uncertainty, negative emotions, and distrust are rampant.
Moreover, the effect of dynamic temporal contexts in crisis visualization design brings up a new challenge in evaluation. Since visualizations have changed over time, what methodological approaches should we take into consideration when designing and evaluating visualizations? In addition to traditional considerations like whom do we evaluate? How many participants do we need? What do we evaluate? Other considerations include: what are the metrics we should consider to incorporate the waves of changes in visualizations over time? Shneiderman and Plaisant's Multi-dimensional In-depth Long-term Case studies (MILCs) might help to examine the long-term effects of visualizations~\cite{shneiderman2006strategies}. However, there is more work needed to further explore methodological and technical aspects of visualization evaluation ``in the wild'' during times of crisis.
It is also important to understand the similarities and differences of various aspects of effects among populations with different characteristics and beliefs in relationship and trust with authorities and organizations (as discussed in the \textit{``to whom''} component). Examining these aspects is crucial for creating an equitable society in which people of all backgrounds are able to utilize, benefit from, question, and challenge information that is disseminated about the humanitarian challenges that impact our world.
\section{Limitations}
Our work was limited by the sample of COVID-19 visualizations, not covering other historical crises. Therefore, more research is needed to test our framework and adapt it to better suit contexts of the crises. Moreover, our collection inevitably fails to capture all COVID-19 visualizations. Most of the examples (in English) in this paper were from publishers in the United States, and some were from other countries (e.g., Japan, Singapore). These examples in the paper may create bias as they may be seen as Western-centric, but they did not reflect our intention. Instead, we hope our work will spark further conversations around crisis visualizations.
\section{Conclusion}
Not a single visualization or message can tell the whole story about the pandemic. Instead, multiple views across different fields with a collective effort from multiple stakeholders may help better reveal the state of the crisis. Through our work in collecting, documenting, organizing, and analyzing hundreds of COVID-19 visualizations, we categorized the trends and patterns of these crisis visualizations, and challenges inherent in the design. We also contributed a conceptual framework of crisis visualization that helps guide future analysis of existing crisis visualization, design effective visualizations, and evaluate the effects of these visualizations. Our work aims to help future work further examine the space of crisis visualization.
\begin{acks}
We wish to thank Ben Shneiderman, Catherine Plaisant, John Stasko, and Alex Endert for discussions and advice, and our reviewers for their constructive feedback. We also thank Paul Kahn for co-leading the database and people who have facilitated with data collection, as well as Lin Shi, Jennifer Howell, Rumi Chunara, Racquel Fygenson, Helia Hossein-pour, Anamaria Crisan, Hugh Dubberly and Dubberly Design Office, Wellness Technology lab at Georgia Tech, Visualization group at Georgia Tech, and VIS lab at Northeastern University for feedback and various levels of support on this work. This work was supported by NSF award number \#2028374.
\end{acks}
\balance
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction} A topological space $X$ is called
profinite if there exists
an inverse system of finite discrete spaces for which its inverse limit is
homeomorphic to $X,$ equivalently by \cite[Proposition 2.4]{Magid} $X$ is profinite if it is compact Hausdorff and zero-dimensional. A topological group is profinite if it is profinite as a topological space. Important examples of profinite groups and spaces are the group of $p$-adic integers, $p$ prime, the Galois group on an arbitrary Galois extension, fundamental groups of connected schemes and the set of connected components of a compact Hausdorff space. For details in profinite spaces and profinite groups the interested reader may consult \cite{RZ} or \cite{WI}.
On the other hand, partial actions of groups appeared in the context of $C^*$-algebras
in \cite{Ruy1}, in which $C^*$-algebraic crossed products by partial automorphism were introduced and studied by analyzing view of their internal structure. Since then, partial group actions have appeared in many different context, such as the theory of operator algebras, Galois cohomology, Hopf algebras, Polish spaces, the theory of $\mathbb{R}$-trees and model theory (see \cite{DO} for a detail account in recent developments on partial actions). It is worth to mention that a deep application of partial actions is given in \cite{AraE} where it was provided a counter-example for a conjecture of M. R{\o}rdam and A. Sierakovski related the Banach-Tarski paradox.
A natural question is whether a partial action obtained as restriction of a corresponding collection of total maps on some superspace. In the topological context, this problem was studied in \cite{AB} and \cite{KL}. They showed that for any partial action $\eta=\{\eta_g\}_{g\in G}$ of a topological group $G$ on a topological space X, there is a topological space $Y$ and a continuous action $\mu$ of $G$ on $Y$ such that $X$ is embedded in $Y$ and $\eta$ is the restriction of $\mu$ to $X.$ Such a space $Y$ is called a globalization of $X.$ They also show that there is a minimal globalization $X_G$ called the enveloping space of $X$ (see Definition \ref{globa}). However, structural properties of $X$ are not in general inherited by $X_G,$ for instance in \cite{Ruy2} it is shown that the enveloping space of a partial action of a countable group on compact metric spaces is Hausdorff iff the domain of each $\eta_g$ is clopen for all $g \in G,$ while in \cite{PU1} there were established conditions for which $X_G$ is a Polish space provided that $X$ and $G$ are.
The present work is structured as follows. After the introduction, in Section \ref{pactions} we provide the necessary background and notations on (set theoretic) partial and topological partial actions, and their corresponding globalization. Also we give some preliminary results that will be needed in the work. In Section \ref{parprof} we work with partial actions on a profinite space $X$ and present in Theorem \ref{pro3.3} a sufficient condition for the space $X/\!\sim_G$ to be profinite, where $\sim_G$ is the orbit equivalence relation determined by the partial action (see equation \eqref{porbit}). Later we treat the problem of finding a continuous section to the quotient map $\pi_G: X\to X/\! \sim_G$ and show the existence of such a section when the group $G$ is finite (see Proposition \ref{pro3.6}) this result is extended in Theorem \ref{contgeral} to the case $G$ profinite, to do so one needs to work with partial actions of quotient groups, which are not possible to define in a natural way as classical (global) actions. Thus in Lemma \ref{lemaaux} and Lemma \ref{lem4.3} we use globalizations and induced partial actions to provide a partial action of a quotient group of $G$ on a certain quotient space. Some applications to quotient maps induced by enveloping actions and enveloping spaces are presented in Subsection \ref{apglob}, in particular having a continuous section of $\pi_G$ we show how to find a continuous section of $\Pi_G$ (the corresponding quotient map induced by globalization), at this point it is important to remark that the converse does not seem to be true, that is having a continuous section of $\Pi_G$ does not seem to imply that $\pi_G$ has a continuous section, items 4) and 5) of Proposition \ref{cor3.3} deal with this problem (see also Proposition \ref{borel}). We end this work with Section \ref{exelgroup} by considering the Exel semigroup, and the groupoid action in the context of profinite spaces.
\section{Preliminaries on partial actions}\label{pactions} In this section we state our conventions on partial actions, and prove some results that will be useful throughout the work. We start with the following.
\begin{defi} \cite[p. 87-88]{KL} Let $G$ be a group with identity element $1$ and $X$ be a set. A partially defined function $\eta: G\times X\dashrightarrow X$, $(g,x)\mapsto g\cdot x$ is called a (set theoretic) {\it partial action} of $G$ on $X$ if for each $g,h\in G$ and $x\in X$ the following assertions hold:
\begin{enumerate}
\item [(PA1)] If $\exists g\cdot x$, then $\exists g^{-1}\cdot (g\cdot x)$ and $g^{-1}\cdot (g\cdot x)=x$,
\item [(PA2)] If $\exists g\cdot(h\cdot x)$, then $\exists (gh)\cdot x$ and $g\cdot(h\cdot x)=(gh)\cdot x$,
\item [(PA3)] $\exists 1\cdot x$ and $1\cdot x=x,$
\end{enumerate}
where $\exists g\cdot x$ means that $g\cdot x$ is defined. We say that $\eta$ {\it acts} (globally) on $X$ if $\exists g\cdot x,$ for all $(g,x)\in G\times X.$
\end{defi}
\begin{rem} In \cite{KL} a partial action is only required to satisfy (PA1) and (PA2), and it is called unital if also (PA3) holds, so for us partial actions are always unital.
\end{rem}
Let $\eta$ be a partial action of $G$ on $X,$ $g\in G, x\in X$ and $ U$ a subset of $X.$ We fix the following notations:
\begin{itemize}
\item $G*U=\{(g,u)\in G\times U\mid \exists g\cdot u\}.$ In particular $G\ast X$ is the domain of $\eta.$
\item $G^x=\{g\in G\mid \exists g\cdot x\},$ $G^U=\bigcup\limits_{u\in U} G^u$ and $G^U\cdot U=\{g\cdot u\mid u\in U, g\in G^u\}.$
\item $X_g=\{x\in X\mid \exists g{}^{-1}\cdot x \}.$
\end{itemize}
Then $\eta$ induces a family of bijections $\{\eta_g\colon X_{g{}^{-1}}\ni x\mapsto g\cdot x\in X_g \}_{g\in G},$ such that $\eta_e$ is the identity of $X$ and $\eta_{g{}^{-1}}=\eta{}^{-1}_g.$ We also denoted this family by $\eta.$
The following result characterizes partial actions in terms of a family of bijections.
\begin{pro} {\rm \cite[Lemma 1.2]{QR} \label{fam} A partial action $\eta$ of $G$ on $X$ is a family $\eta=\{\eta_g\colon X_{g{}^{-1}}\to X_g\}_{g\in G},$ where $X_g\subseteq X,$ $\eta_g\colon X_{g{}^{-1}}\to X_g$ is bijective, for all $g\in G,$ and such that:
\begin{itemize}
\item[(i)]$X_1=X$ and $\eta_1=\rm{id}_X;$
\item[(ii)] $\eta_g( X_{g{}^{-1}}\cap X_h)=X_g\cap X_{gh};$
\item[(iii)] $\eta_g\eta_h\colon X_{h{}^{-1}}\cap X_{ h{}^{-1} g{}^{-1}}\to X_g\cap X_{gh},$ and $\eta_g\eta_h=\eta_{gh}$ in $ X_{h{}^{-1}}\cap X_{ h{}^{-1} g{}^{-1}};$
\end{itemize}
for all $g,h\in G.$}
\end{pro}
\begin{exe}\cite[Example 2.3]{PU1}\label{etahat}
Let $\eta$ be a partial action of $G$ on $X$. Consider the family of bijections $\hat{\eta}=\{\hat{\eta}_g: (G\times X)_{g^{-1}}\rightarrow (G\times X)_g\}_{g\in G},$ where $(G\times X)_g=G\times X_g$ and $\hat{\eta}_g(h,x)=(hg^{-1},\eta_g(x))$ for each $g\in G$ and $(h,x)\in G\times X_{g^{-1}}$. Then $\hat\eta$ is a partial action of $G$ on $G\times X.$
\end{exe}
The partial action $\hat\eta$ defined above, will be useful for our purposes.
\begin{defi}
Let $\eta: G\ast X \rightarrow X$ be a partial action of $G$ on $X$ and $U\subseteq X, $ then $U$ is $G$- invariant if $G^U\cdot U\subseteq U,$ or equivalently,
$G^U\cdot U= U.$\end{defi}
We have the next.
\begin{lem}\label{lem1.1.1}
Let $\eta$ be a partial action of $G$ on $X$ and $U$ a nonempty subset of $X$. Then the following statements are true:
\begin{enumerate}
\item [(i)] $\eta(G*U)=G^U\cdot U$.
\item [(ii)] $G^U\cdot U$ and its complement are $G$-invariant.
\end{enumerate}
\end{lem}
\begin{proof}
Statement $(i)$ is straightforward. To show $(ii)$ denote $A:=G^U\cdot U$ and $B=X\setminus A$. The assumption $G^A\cdot A \subseteq A$ follows from (PA2). Now we check $G^B\cdot B\subseteq B$. Take $y\in G^B\cdot B$, then there are $x\in B$ and $g\in G^x$ such that $\eta_g(x)=y$. Suppose $y\in A$ and let $m\in U, h\in G^m$ such that $\eta_h(m)=y$. Note that $\eta_h(m)\in X_h$ and $\eta_h(m)=y\in X_g$, then $m\in \eta_h^{-1}(X_h\cap X_g)=X_{h{}^{-1}}\cap X_{h{}^{-1} g}$ and by (iii) of Proposition \ref{fam} follows that $\eta_{g^{-1}h}(m)=x$, this implies that $x\in A$, which is a contradiction, therefore $y\in B$.
\end{proof}
From now on in this work $G$ will denote a topological group and $X$ a topological space, we endow $G\times X$ with the product topology and $G*X$ with the induced topology of subspace. Moreover $\eta: G*X\to X$ will denote a partial action. We say that $\eta$ is a {\it topological partial action} if every $X_g$ is open and $\eta_g$ is a homeomorphism, $g\in G,$ if moreover $\eta$ is continuous, we say that $\eta$
is a {\it continuous partial action}.
Respect the partial action $\hat\eta$ given in Example \ref{etahat} we have the next.
\begin{pro}\label{hatcon}
Let $\eta$ be a topological partial action of $G$ on $X$. Then $\eta$ is continuous if and only if $\hat{\eta}$ is continuous.
\end{pro}
\begin{proof}
Suppose that $\eta$ is continuous. It is clear that $\hat{\eta}$ is a topological partial action. To show that $\hat{\eta}$ is continuous set $\alpha:G*(G\times X)\ni (g, (h,x))\mapsto hg^{-1}\in G$ and
\begin{equation}\label{bt}\beta:G*(G\times X)\ni(g,(h,x))\mapsto \eta_g(x)\in X.\end{equation} We show that $\alpha$ and $\beta$ are continuous, for this consider $U\subseteq G$ an open set for which $\alpha(g,(h,x))\in U$ for some $(g,(h,x))\in G*(G\times X)$. Consider $V$ and $W$ open subsets of $G$ such that $(g,h)\in V\times W$ and $WV^{-1}\subseteq U$ and define $Z_0:= (V\times G\times W)\cap G*(G\times X)$. It is easy to see that $(g,(h,x))\in Z_0$ and $\alpha(Z_0)\subseteq WV^{-1}\subseteq U$. This shows that $\alpha$ is continuous. On the other hand, let $U\subseteq X$ be a open in $X$ such that $\beta(g,(h,x))\in U$ for some $(g,(h,x))$. Since $\eta$ is continuous there are open sets $T\subseteq G$, $H\subseteq X$ such that $(g,x)\in T\times H$ and $\eta((T\times H)\cap G*X)\subseteq U.$ Let $Z_1:=(T\times G\times H)\cap G*(G\times X)$ then $(g,(h,x))\in Z_1$ and $\beta(Z_1)\subseteq \eta((T\times X)\cap G*X)\subseteq U$. Thus we conclude that $\hat{\eta}$ is continuous.
Conversely, suppose that $\hat{\eta}$ is continuous. Let $U\subseteq X$ be an open set such that $\eta(g,x)\in U$ for some $(g,x)\in G*X$. Note that $(g,(1,x))\in G*(G\times X)$ and $\beta(g,(1,x))=\eta(g,x)\in U$. Since the map $\beta$ defined in \eqref{bt} satisfies $\beta=\pi\circ \hat\eta,$ where $\pi: G\times X\ni (g,x)\mapsto x\in X$ then it is continuous and there are open subsets $M,N\subseteq G$ and $V\subseteq X$ such that $(g,(1,x))\in M\times N\times V$ and $\beta(Z)\subseteq U$, where $Z:= (M\times N\times V)\cap G*(G\times X)$. Define $Z':= (M\times V)\cap G*X$. Note that $(g,x)\in Z'$, moreover $\eta(Z')\subseteq \beta(Z)\subseteq U,$ and $\eta$ is continuous as desired.
\end{proof}
\begin{lem}\label{lem1.1}
Let $\eta$ be a partial action of $G$ on $X.$ Then the following assertions hold:
\begin{itemize}
\item [(i)] If $U$ is an open subset of $X,$ $G^U\cdot U$ is open.
\item [(ii)] If $G*X$ is clopen then $X_g$ is clopen for all $g\in G.$
\end{itemize}
\end{lem}
\begin{proof}
The first assertion follows from the fact that for every $g\in G$ the set $X_g$ is open and $G^U\cdot U=\bigcup\limits_{g\in G^U} \eta_g(U\cap X_{g{}^{-1}}).$ We check the second one, take $g\in G$, first let us prove that $X_g$ is closed. Let $x\in X\setminus X_g,$ since $G*X$ is closed there open sets $U\subseteq G$ and $V\subseteq X$ such that $(g^{-1},x)\in U\times V\subseteq (G*X)^c$. Moreover, if $y\in V$ then $(g^{-1},y)\notin G*X$ and $y\notin X_g$. This shows that $X_g$ is closed.
To prove that $X_g$ is open, take $x\in X_g$ then $(g^{-1},x)\in G*X$ and there are open sets $U\subseteq G$ and $V\subseteq X$ such that $(g^{-1},x)\in U\times V\subseteq G*X$ from this we get $x\in V\subseteq X_g$ and $X_g$ is open.
\end{proof}
\subsection{Induced partial actions and globalization}
Let $u \colon G\times Y\to Y$ be a continuous action of $G$ on a topological space $Y$ and $X\subseteq Y$ an open set. For $g\in G,$ set
\begin{equation}\label{induced}X_g=X\cap u_g(X)\,\,\, \text{ and let}\,\,\, \eta_g=u_g\restriction X_{g{}^{-1}}.\end{equation} Then $\eta \colon G* X\ni (g,x)\mapsto \eta_g(x)\in X $ is a topological partial action of $G$ on $X.$ In this case we say that $\eta$ is {\it induced} by $u.$
An important question in the study of partial actions is whether they can be induced by global actions. In the topological sense, this turns out to be affirmative and a proof was given in \cite[Theorem 1.1]{AB} and independently in \cite[Section 3.1]{KL}. We recall their construction. Let $\eta$ be a topological partial action of $G$ on $X.$ Define an
equivalence relation on $ G\times X$ as follows:
\begin{equation}
\label{eqgl}
(g,x)R(h,y) \Longleftrightarrow x\in X_{g{}^{-1} h}\,\,\,\, \text{and}\, \,\,\, \eta_{h{}^{-1} g}(x)=y,
\end{equation}
and denote by $[g,x]$ the equivalence class of the pair $(g,x).$
Consider the set $X_G=(G\times X)/R$ endowed the quotient topology.
Then by \cite[Theorem 1.1]{AB} the action
\begin{equation}
\label{action}
\mu \colon G\times X_G\ni (g,[h,x])\to [gh,x]\in X_G,
\end{equation}
is continuous and the map
\begin{equation}
\label{iota}
\iota \colon X\ni x\mapsto [1,x]\in X_G
\end{equation} is a continuous injection such that $G\cdot \iota(X)=X_G$ and if $G\ast X$ is open then $\iota$ is open.
We finish this section with the next.
\begin{defi}\label{globa} Let $\eta$ be a topological partial action of $G$ on $X.$ The action $\mu$
provided by \eqref{action} is called the enveloping action of $\eta$ and the space $X_G$ is the
enveloping space or the globalization of $X.$ \end{defi}
\section{Partial actions on profinite spaces}\label{parprof}
Given a topological partial action of $G$ on $X,$ one can define the {\it orbit equivalence relation} $\sim_G$ on $X$ as follows:
\begin{equation}\label{porbit}x\sim_G y\Longleftrightarrow\,\, \exists g\in G^x \,\,{\rm such \,\,that}\,\, g\cdot x=y,\end{equation}
for each $x,y\in X$. The elements of $X/\!\sim_G$ are the {\it orbits} $G^x\cdot x$ with $x\in X$ and $X/\!\sim_G$ is endowed with the quotient topology, in particular by \cite[Lemma 3.2]{PU1} the {\it induced quotient map of $\eta$}
\begin{equation}\label{qmap}\pi_G: X\ni x\mapsto G^x\cdot x \in X/\!\sim_G\end{equation} is continuous and open.
Our next goal is to show that $X/\!\sim_G$ is profinite provided that $X$ is profinite and $G$ compact, but first for the reader's convenience we give the next.
\begin{defi} Two points $u$ and $v$ in $X$ can be separated if there are disjoint open subsets $U$ and $V$ of $X$ such that $u\in U ,$ $v\in V$ and $U\cup V=X.$
\end{defi}
It is known that any two different points in a compact space $X$ can be separated, if and only if, $X$ is Hausdorff and has a basis consisting of clopen sets (see for instance \cite[Proposition 2.3]{Magid}).
\begin{teo} \label{pro3.3}
Let $\eta: G* X\rightarrow X$ be a continuous partial action of a compact group $G$ on a profinite space $X$ such that $G* X$ is closed, then $X/\sim_G$ is a profinite space.
\end{teo}
\begin{proof}
First of all note that $X/\sim_G$ is compact. Now we show that different points $G^x\cdot x, G^y\cdot y\in X/\sim_G$ can be separated. Let $\mathcal{C}:=\{ U\subseteq X : U$ is clopen$, x\in U \}$, then $\mathcal{C}\neq \emptyset$. We claim that there exist $U\in \mathcal{C}$ such that $G^U\cdot U\cap G^y\cdot y=\emptyset$. On the contrary, for each $V\in \mathcal{C}$ the set $\tilde{F}_y(V)=\{ (g,v)\in G* V : \eta_g(v)=y\}$ is nonempty. Since $\tilde{F}_y(V)=\eta^{-1}(y)\cap G*V$, it is closed in $G*V$ and thus closed in $G\ast X$. Now, if $V_1, V_2\in \mathcal{C}$, then $\tilde{F}_y(V_1\cap V_2)\subseteq \tilde{F}_y(V_1)\cap \tilde{F}_y(V_2)$. In that sense, $\lbrace \tilde{F}_y(V)\rbrace_{V\in \mathcal{C}}$ is a family in $G\ast X$ with the finite intersection property. Thus, there exist $(g,v)\in \bigcap_{V\in \mathcal{C}}\tilde{F}_y(V)$ which implies $v=x$ and $\eta_g(x)=y$ and leads to a contradiction. Then there exist $U\in \mathcal{C}$ such that $G^U\cdot U\cap G^y\cdot y=\emptyset$. Now we check that $G^U\cdot U$ is clopen, indeed it is open thanks to Lemma \ref{lem1.1}, moreover since $G\ast U$ is compact, $X$ is Hausdorff and $\eta(G*U)=G^U\cdot U$, we have that $G^U\cdot U$ is closed. Let $A=G^U\cdot U$ and $B=X\setminus A$ by Lemma \ref{lem1.1.1} the sets $A$ and $B$ are $G$-invariant and clopen, then
$\pi_G^{-1}(\pi_G(A))=A$ y $\pi_G^{-1}(\pi_G(B))=B$ and $G^x\cdot x$ y $G^y\cdot y$ are separated by the sets $\pi_G(A)$ and $\pi_G(B),$ respectively.
\end{proof}
\subsection{On continuous sections of the quotient map}
In this section we are interested in providing conditions for which the quotient map $\pi_G$ defined in \eqref{qmap} has a continuous section, for this we adapt some of the ideas presented in \cite[Section 2.4]{Magid} to the context of partial actions. It is important to remark that the partial case is essentially more laborious than the classical one.
First we give the next.
\begin{defi}
Let $\eta: G\ast X\longrightarrow X$ be a topological partial action of a group $G$ on a set $X$. We say that $\eta$ is free if for each $(g,x)\in G\ast X$ such that $\eta(g,x)=x$, we have $g=1$.
\end{defi}
\begin{pro}\label{pro3.6}
Let $G$ be a finite and discrete group, $\eta$ a continuous and free partial action such that $G\ast X$ is clopen. Then $\pi_G:X\longrightarrow X/\!\sim_G$ has a continuous section.
\end{pro}
\begin{proof}
Let $x\in X$. We have that $\eta_g(x)\neq \eta_h(x)$ for each $g,h\in G^x$ such that $g\neq h$ because $\eta$ is free. Since $G$ is finite and $X$ profinite, for each $g\in G^x$ there is a clopen set $U_g$ such that $\eta_g(x)\in U_g$ and $U_g\cap U_h=\emptyset$, if $g\neq h$. Since $x\in \eta_g^{-1}(U_g\cap X_g)$ for each $g\in G^x$, $x\in V:= \bigcap_{g\in G^x}\eta_g^{-1}(U_g\cap X_g)$, then $V$ is clopen thanks to the second item of Lemma \ref{lem1.1}. We will show that $\pi_G\restriction_V:V\rightarrow \pi_G(V)$ is a homeomorphism. To show the injectivity, note that if $\eta_g(V)\cap V\neq\emptyset$ for some $g\in G^x$ then $g=1,$ indeed let $y\in \eta_g(V)\cap V$ since $V\subseteq \eta_{g^{-1}}(U_g\cap X_g)$, then $y\in U_g$. Thus $y\in U_g\cap U_1$ and $g=1$. Now, let $a,b\in V$ with $\pi_G(a)=\pi_G(b)$ then there exists $g\in G^a$ such that $\eta_g(a)=b$. Thus, $b\in \eta_g(X_{g^{-1}}\cap V)\cap V$ and for the previous reasoning $g=1$. This shows that $\pi_G\restriction_V$ is injective. Moreover, since $V$ is compact and $\pi_G(V)$ is Hausdorff, we have that $\pi_G\restriction_V$ is a homeomorphism. The inverse of $\pi_G\restriction_V$ is a continuous section of $\pi_G$ over $\pi_G(V)$. To finish the proof, note that $\pi_G(V)$ is clopen since $\pi_G^{-1}(\pi_G(V))=\bigcup_{g\in G}\eta_g(X_{g^{-1}}\cap V)$ is clopen. Thus, for each $x\in X$, there exist a clopen neighborhood $V_x=\pi_G(V)$ of $\pi_G(x)$ and a continuous section $q\restriction_{V_x}: V_x\rightarrow X$ of $\pi_G\restriction_V$. Since $X$ is compact there exist $m\in \mathbb{N}$ and $V_1, V_2, \cdots, V_m$ such that $X=\bigcup\limits_{i=1}^mV_i,$ moreover by \cite [Lemma 2.4]{Magid} there is a refinement $\{W_j\}_{j=1}^n$ of $\{V_i\}_{i=1}^m$ such that $W_i\cap W_j=\emptyset$, if $i\neq j$ and using the family $\{q\restriction_{W_j}\}_{j=1}^n$ we obtain a continuous section of $\pi_G$.
\end{proof}
\begin{exe} \cite[p. $22$]{Ruy} {\bf Partial Bernoulli action} Let $G$ be a discrete group and $X:= \lbrace 0,1\rbrace^G$. There is a continuous global action $\beta=\lbrace \beta_g\rbrace$, where for all $\omega\in X$, $\beta_g(\omega)=g\omega$. The topological partial Bernoulli action $\eta$ is obtained by restricting $\beta$ to the open set $\Omega_1=\lbrace \omega\in X : \omega(1)=1\rbrace$. Thus by \eqref{induced} $D_g:=\Omega_1\cap \beta_g(\Omega_1)= \lbrace \omega\in X : \omega(1)=1=\omega(g)\rbrace$ and $\eta_g= \beta_g\restriction_{D_g^{-1}}, g\in G$. Let us show that $G\ast \Omega_1$ is clopen. Let $\{(n_i, x_i)\}_{i\in I}$ be a convergent net to $(n,x)$, since $G$ is discrete then $(n_i)_{i\in I}$ is constant and for all $i\in I$, $n_i=n$. On the other hand, as $x_i\longrightarrow x$ then $1=x_i(1)\longrightarrow x(1)$ and from this $x(1)=1$. Similarly it is obtained that $x(n^{-1})=1$ and from the above we conclude $(n,x)\in G\ast \Omega_1$. On the other hand, take $(n,x)\in G*\Omega_1$. Then $x\in V=( \pi_{n^{-1}}\restriction_{\Omega_1})^{-1}(\{1\})$ and $(n,x)\in \{n\}\times V\subseteq G*\Omega_1$. This shows that $G*\Omega_1$ is clopen. Thus if $G$ is finite Theorem \ref{pro3.3} implies that $X/\!\sim_G$ is a profinite space
\end{exe}
We shall extend Proposition \ref{pro3.6} to the case when $G$ is profinite, but first we use globalizations to obtain partial actions of quotient groups on certain orbit spaces.
Let $\eta$ be a topological partial action of $G$ on $X$. Then for any subgroup $H$ of $G$ $\eta$ induces by restriction a topological partial action $\eta_H$ of $H$ on $X$. The corresponding orbit equivalence relation of $\eta_H$ is denoted by $\sim_H$. On the other hand it is necessary to clarify at this point that the orbits in the space $X_G/\sim_H$ will be denoted by $H[g,x]$ for an element $[g,x]\in X_G$.
\begin{lem}\label{lemaaux}
Let $\eta$ be a partial action of $G$ on $X$ with $G*X$ open and $X_G$ its enveloping space. Then for each subgroup $H$ the map
\begin{equation}\label{var}\varphi:X/\!\sim_H \ni H^x\cdot x\mapsto H[1,x]\in X_G/\sim_H \end{equation} is an embedding, that is continuous, open and injective. \end{lem}
\begin{proof}
First of all note that $\varphi$ is well defined. In fact, let $x,y\in X$ such that $x\sim_H y$ and take $h\in H^x$ with $\eta_h(x)=y$. Thus, $[1,y]\stackrel{\eqref{eqgl}}=[h,x]\stackrel{\eqref{action}}=\mu_h([1,x])$ and $[1,y]\sim_H[1,x]$. This shows that $\varphi$ is well defined, it is easy to check that $\varphi$ is injective. To prove that $\varphi$ is continuous, consider $\pi_H: X\rightarrow X/\!\sim_H$ and $\Pi_H:X_G\rightarrow X_G/\!\sim_H$ the corresponding projection functions. Since the map $\iota$ defined in \eqref{iota} is continuous and $\varphi\circ \pi_H=\Pi_H\circ \iota$, we conclude that $\varphi$ is continuous. It remains to check that $\varphi$ is open, let $U\subseteq X/\sim_H$ open, then $\varphi(U)=\Pi_H(\iota(\pi_H^{-1}(U)))$ is open because $\pi_H^{-1}(U)$ is open in $X$ and the functions $\iota$ and $\Pi_H$ are open.
\end{proof}
We proceed with the next.
\begin{lem}\label{lem4.3}
Let $\eta$ be a free partial action of a profinite group $G$ on a profinite space $X$, and $H$ a closed normal subgroup of $G$. Then there is a continuous free partial action $\eta_{G/H}$ of $G/H$ on $X/\!\sim_H$, such that the spaces $(X/\!\sim_H)/\!\sim_{G/H}$ and $X/\!\sim_G$ are homeomorphic.
\end{lem}
\begin{proof}
Let $\mu$ be the globalization of $\eta$. Then $\mu$ is continuous and induces a continuous action $\tau_{G/H}$ on $X_G/\!\sim_H$ as follows:
$$\tau_{gH}: X_G/\sim_H\ni \Pi_H([t,x]) \mapsto \Pi_H([gt,x])\in X_G/\sim_H,$$
for each $gH\in G/H$. Note that $\tau_{G/H}$ is free. In fact, let $g,t\in G$ and $x\in X$ such that $\Pi_H([gt,x])=\Pi_H([t,x]).$ Then $[hgt,x]=[t,x]$ for some $h\in H$ and by \eqref{eqgl} we have $x\in X_{(hgt)^{-1}t}$ and $\eta_{t^{-1}hgt}(x)=x$, moreover the fact that $\eta$ is free implies then $hg=1$ and $g\in H$, this shows that $gH=H$ and $\tau_{G/H}$ is free. Now let $\varphi$ be defined by \eqref{var} since $\varphi$ is open, by restricting $\tau_{G/H}$ to the open set $Im(\varphi)$ to obtain a partial action $\eta'_{G/H}=\{\eta'_{gH}:X_{g{}^{-1} H} \to X_{gH}\}_{gH\in G/H}$ of $G/H$ on $Im(\varphi),$ where $X_{gH}=\tau_{gH}(Im(\varphi))\cap Im(\varphi)$ and $\eta'_{gH}=\tau_{gH}\restriction{X_{g^{-1}H}}$ (see equation \ref{induced}), it is not difficult to see that $\eta'_{G/H}$ is continuous and free. On the other hand, using $\eta'_{G/H}$ and the embedding $\varphi$ we construct the partial action $\eta_{G/H}$ of $G/H$
on $T:=X/\sim_H$, where $T_{gH}=\varphi^{-1}(X_{gH})$ and
\begin{equation}\label{etaq}\eta_{gH}(x)=\varphi^{-1}(\eta'_{gH}(\varphi(x)))=\varphi^{-1}(\tau_{gH}(\varphi(x))),\end{equation} for each $x\in \varphi^{-1}(X_{g^{-1}H})$. It is easy to check that $\eta_{G/H}$ is free and continuous.
Let $\sim_{G/H}$ be the orbit relation in $T$ induced by $\eta_{G/H}$. We are going to show that $T/\!\sim_{G/H}$ and $X/\sim_G$ are homeomorphic. In that sense, consider the following diagram:
\begin{center}
$\xymatrix{X \ar[d]_-{\pi_H} \ar[r]^-{\pi_G}& X/\sim_G \\
T \ar[r]_-{\pi_{G/H}}& T/\sim_{G/H}\ar[u]^-{\psi}
}$
\end{center}
where $\psi$ is made such that the diagram commutes, that is
\begin{equation}\label{homeo}\psi(\pi_{G/H}(\pi_H(x)))=\pi_G(x),\end{equation} for each $x\in X$. Let us first prove that $\psi$ is well defined, take $z\in T/\!\sim_{G/H}$ and $x,y\in X$ such that $\pi_{G/H}(\pi_H(x))=\pi_{G/H}(\pi_H(y))=z$. Then there is $g\in G$ with $\eta_{gH}(\pi_H(x))=\pi_H(y)$, that is, $\varphi^{-1}(\tau_{gH}(\varphi(\pi_H(x))))=\pi_H(y)$ which implies $H[g,x]=H[1,y]$ and there is $h\in H$ such that $[hg,x]=[1,y]$. We deduce that $\eta_{hg}(x)=y$ and $\pi_G(x)=\pi_G(y)$ which shows that $\psi$ is well defined. Moreover the map $\psi$ is continuous and surjective.
Let us prove that $\psi$ is injective. Let $z_1,z_2\in T/\!\sim_{G/H}$ such that $\psi(z_1)=\psi(z_2)$, and let $x,y\in X$ such that $\pi_{G/H}(\pi_H(x))=z_1$ and $\pi_{G/H}(\pi_H(y))=z_2$. Since $\pi_G(x)=\pi_G(y)$, there is $g\in G^x$ such that $\eta_g(x)=y$. To prove that $z_1=z_2$, we need $t\in G$ such that $\eta_{tH}(\pi_H(x))=\pi_H(y)$. We claim that $\eta_{gH}(\pi_H(x))=\pi_H(y).$ In fact, note that \begin{center}
$\eta_{gH}(\pi_H(x))=\varphi^{-1}(\tau_{gH}(\varphi(\pi_H(x))))=\varphi^{-1}(H[g,x])$
\end{center}
and
\begin{center}
$\varphi(\pi_H(y))=H[1,y]=H[g,x],$
\end{center}
then $\eta_{gH}(\pi_H(x))=\pi_H(y)$. Thus $\psi$ is injective.
Finally let $U\subseteq T/\sim_{G/H}$ be an open set. Since $\pi_G$ is open, $\pi_G(\pi_H^{-1}(\pi_{G/H}^{-1}(U)))\subseteq X/\sim_G$ is open. Thus, $\psi(U)$ is open and $\psi$ is a homeomorphism. \end{proof}
Let $H_1, H_2$ be subgroups of $G$ such that $H_1\subseteq H_2$. We define $\pi_{H_1, H_2}: X/\!\sim_{H_1}\to X/\!\sim_{H_2}$ as the only map such that
\begin{equation}\label{pih12}\pi_{H_2}= \pi_{H_1, H_2}\circ \pi_{H_1},\end{equation} in particular for a subgroup $H$ of $G$ the map $\pi_{H,H}$ is the identity on $X/\!\sim_{H}.$
\begin{rem}\label{comen4.4}
According to the notations of Lemma \ref{lemaaux} and Lemma \ref{lem4.3} suppose that there exists a continuous section $\lambda$ of $\pi_{G/H}$. Note that $\lambda\circ \psi^{-1}$ is a continuous section of $\pi_{H,G}$. Indeed, take $x\in X$. We have
\begin{equation}\label{comp}(\lambda\circ\psi^{-1})(\pi_G(x))\stackrel{\eqref{homeo}}=(\lambda\circ\pi_{G/H})(\pi_H(x))=\pi_H(y),\end{equation}
where $y\in X$ is chosen such that $\lambda(\pi_{G/H}(\pi_H(x)))=\pi_H(y)$. From this follows that $\pi_{G/H}(\pi_H(x))=\pi_{G/H}(\pi_H(y))$ then there exists $g\in G$ such that $\eta_{gH}(\pi_H(x))=\pi_H(y)$, that is \begin{center}
$\pi_H(y)\stackrel{\eqref{etaq}}=\varphi^{-1}(\eta'_{gH}(\varphi(\pi_H(x)))\stackrel{\eqref{var}}=\varphi^{-1}(\tau_{gH}(\Pi_H([1,x])))=\varphi^{-1}(\Pi_H([g,x]))$,
\end{center}
then $\Pi_H([g,x])=\Pi_H([1,y])$ and there is $h\in H$ such that $\eta_{gh}(x)=y$. This shows that $\pi_G(x)=\pi_G(y). $ Moreover
$$[\pi_{H,G}\circ(\lambda \circ\psi^{-1})](\pi_G(x))\stackrel{\eqref{comp}}=\pi_{H,G}(\pi_H(y))\stackrel{\eqref{pih12}}=\pi_G(y)=\pi_G(x),$$ as desired.
\end{rem}
In order to show the main result of this section we recall the following:
\begin{lem} The following statements hold.
\begin{itemize}
\item \cite[Proposition 1.1.6]{WI}\label{lem4.5}
Let $\{X_i;f_i^j,I\}$ be an inverse system of compact Hausdorff topological spaces over a directed set $I$, and $(X;\pi_i)$ be its inverse limit. If $Y\subseteq X$ is such that $\pi_i(Y)=X_i$ for each $i\in I$, then $Y$ is dense in $X$.
\item \cite[Lemma 2.12]{Magid} \label{lem 4.6}
Let $G$ be a profinite group and $N$ a closed normal subgroup of $G$. If $N\neq\{1\}$, then there is a proper subgroup $M$ of $N$ such that $M$ is open in $N$ and normal in $G$.
\end{itemize}
\end{lem}
We write $N\trianglelefteq_{\rm cl} G$ to indicate that $N$ \ is a closed\ normal\ subgroup\ of $G.$ We give the next.
\begin{teo}\label{contgeral}
Let $\eta$ be a continuous and free partial action of a profinite group $G$ on a profinite space $X$ such that $G*X$ is clopen. Then $\pi_G:X\rightarrow X/\sim_G$ has a continuous section.
\end{teo}
\begin{proof}
Consider \begin{center}
$\mathcal{X}:=\{(F,r): F\trianglelefteq_{\rm cl} G\ and\ r\ is\ a\ continuous\ section\ of\ \pi_{F,G} \}$.
\end{center}
Note that $\mathcal{X}\neq\emptyset$ because $(G,{\rm id}_{\pi_G(X)})\in\mathcal{X}$. We define a partial order in $\mathcal{X}$ as follows: For each $(F,r)$ and $(F',r')\in \mathcal{X}$, we say that
\begin{equation*}(F,r)\leq (F',r')\,\,\text{ if and only if} \,\,F'\subseteq F\,\,\text{ and}\,\, \pi_{F',F}\circ r' =r.\end{equation*}
Let $\mathcal{C}=\{(F_i,r_i): i\in I\}$ be a chain in $\mathcal{X}$. Then $\{X/\sim_{F_i}; \pi_{F_i,F_j}; \mathcal{C}\}$ is an inverse system of topological spaces.
Consider $\displaystyle{ \lim_{\longleftarrow}}\ X/\sim_{F_i}$ and for each $i\in I$ let $\pi_i:\displaystyle{ \lim_{\longleftarrow}}\ X/\sim_{F_i}\rightarrow X/\sim_{F_i}$ the corresponding projection map, we are going to prove that $\displaystyle{ \lim_{\longleftarrow}}\ X/\sim_{F_i}$ and $X/\sim_F$ are homeomorphic, where $F=\bigcap_{i\in I}F_i$. In that sense, note that if $(F_i,r_i)\leq (F_j,r_j)$ then $\pi_{F_j,F_i}\circ \pi_{F_j}=\pi_{F_i}$ thanks to \eqref{pih12}, and by the universal property of $\displaystyle{ \lim_{\longleftarrow}}\ X/\sim_{F_i}$ there is a unique continuous map $e: X\rightarrow \displaystyle{ \lim_{\longleftarrow}}\ X/\sim_{F_i}$ such that $\pi_i\circ e=\pi_{F_i}$, for each $i\in I$. We claim that $e$ is surjective. In fact, since $\pi_{F_i}$ is surjective for each $i\in I$ we know that $e(X)$ is dense thanks to the first item of Lemma \ref{lem4.5}. Moreover $\displaystyle{ \lim_{\longleftarrow}}X/\sim_{F_i}$ is Hausdorff by Theorem \ref{pro3.3} for each $i$. Then $e(X)= \displaystyle{ \lim_{\longleftarrow}}\ X/\sim_{F_i}$.
The prove that induced function $\overline{e}: X/\sim_F\rightarrow \displaystyle{ \lim_{\longleftarrow}}\ X/\!\sim_{F_i}$ is homeomorphism is analogous to the prove given in \cite[Proposition 2.9]{Magid}.
Now, let $i,j\in I$ such that $(F_i,r_i)\leq(F_j,r_j)$. We have $r_i=\pi_{F_j,F_i}\circ r_j$, then by the universal property of $ \displaystyle{ \lim_{\longleftarrow}}\ X/\!\sim_{F_i}$ there is a unique continuous function $r: X/\sim_G\rightarrow \displaystyle{ \lim_{\longleftarrow}}\ X/\!\sim_{F_i}$ such that $\pi_i\circ r=r_i$ for each $i\in I$. Consider $q: X/\!\sim_G\rightarrow X/\!\sim_F$ defined by $q=\overline{e}^{-1}\circ r$. Let us prove that $(F,q)\in \mathcal{X},$ take $x\in X$ and $z:=G^x\cdot x$ we need to check that $\pi_{F,G}(q(z))=z$. Let $t\in X$ such that $q(z)=F^t\cdot t,$ then $r(z)=\overline{e}(F^t\cdot t)=e(t)$, that is $(r_i(G^x\cdot x))_{i\in I}=(\pi_{F_i}(t))_{i\in I}$. If $i\in I$ then
$$G^t\cdot t=\pi_{F_i,G}(\pi_{F_i}(t))=\pi_{F_i,G}(r_i(G^x\cdot x))=G^x\cdot x,$$
and $\pi_{F,G}(q(z))=G^t\cdot t=G^x\cdot x=z$. Finally we have to prove that $(F_i,r_i)\leq (F,q),$ for each $i\in I$. Let $i\in I$, $z=G^x\cdot x\in X/\sim_G$ and $t\in X$ such that $q(z)=F^t\cdot t$ it follows by the previous reasoning that $r_i(z)=\pi_{F_i}(t)$, which is $\pi_{F,F_i}(q(z))=\pi_{F,F_i}(F^t\cdot t)=\pi_{F_i}(t)=r_i(z)$. This shows that $(F_i,r_i)\leq (F,q)$ and by Zorn's Lemma we conclude that there is a maximal element $(N,t)\in \mathcal{X}.$
To finish the proof we show that $N=\{1\}.$ Suppose that there exist a open subgroup $M$ of $N$ such that $M$ is normal in $G$. Since $N$ acts continuously on $X$ we get by Lemma \ref{lem4.3} that $N/M$ acts partially and continuously on $T:=X/\sim_M$ and there is a homeomorphism $\psi: T/\!\sim_{N/M} \to X/\sim_N$. Moreover by \cite[Lemma 2.1.2]{RZ} the group $N/M$ is finite thus there is $\lambda: T/\sim_{N/M} \rightarrow T$ a continuous section of $\pi_{N/M}: T\rightarrow T/\sim_{N/M},$ thanks to Proposition \ref{pro3.6} and follows by \ref{comen4.4} that $\lambda\circ\psi^{-1}$ is a continuous section of $\pi_{M,N}$, then $\alpha:=\lambda\circ \psi^{-1}\circ t$ is a continuous section of $\pi_{M,G}$ and $(M,\alpha)\in \mathcal{X}$. Besides $$\pi_{M,N}\circ \alpha=\pi_{M,N}\circ \lambda\circ\psi^{-1}\circ t = Id_{X/\sim_N}\circ t=t,$$ then $(N,t)\leq (M,\alpha)$. Since $(N,t)$ is maximal, we conclude that $N=M$ and by the second item of Lemma \ref{lem4.5} we have that $N=\{1\}$, that is, $t$ is a continuous section of $\pi_G$.
\end{proof}
\begin{ejem}
Let $G$ be a profinite group with internal operation $\divideontimes$, then $G$ acts freely and continuously on itself with action $\divideontimes$. Let us take a open subgroup $U$ of $G$ and consider $\theta=\lbrace \theta_g:D_{g^{-1}}\longrightarrow D_g\rbrace_{g\in G}$ the restriction of $\divideontimes$ on $U$, where for each $g\in G$, $D_g=U\cap gU$. It's clear that $\theta$ is free and continuous. On the other hand note that $G\ast U=G\times U$ and from \cite[Lemma 2.1.2]{RZ} it follows that $G\ast U$ is clopen. By the Theorem \ref{contgeral} we conclude that $\pi_G: U\longrightarrow U/\sim_G$ admits a continuous section.
\end{ejem}
\begin{rem} Notice that in general the assumption that $\eta$ acts freely on $X,$ cannot be omitted even when $\eta$ acts globally, see for instance \cite[Example 5.6.8]{RZ}.
\end{rem}
\subsection{Applications to the globalization}\label{apglob}
Let $\mu$ be the globalization of $\eta$ and $X_G$ the enveloping space of $X$ according to Definition \ref{globa}, we apply our results to the study of $X_G$ and the relation between continuous sections of the maps $\pi_G$ and $\Pi_G,$ being $\Pi_G$ the corresponding quotient map of the enveloping action $\mu.$
\begin{pro}\label{cor3.3}
Let $\eta$ be a continuous partial action of a profinite group $G$ on a profinite space $X.$ Then the following statements hold.
\begin{enumerate}
\item If $G*X$ is closed, then $X_G$ is profinite.
\item If $\mu$ is free and $G*X$ is clopen, then $\Pi_G$ has a continuous section.
\item If $\pi_G$ has a continuous section so does $\Pi_G$.
\item If $G*X$ is open and $q$ is a continuous section of $\Pi_G$ such that ${\rm im}\, q\subseteq \iota(X),$ then $\pi_G$ has a continuous section.
\item If $\Pi_G$ and $\hat{\Pi}_G$ have continuous sections, then $\pi_G$ has continuous section, where $\hat{\Pi}_G$ is the quotient map $G\times X\to X_G.$
\end{enumerate}
\end{pro}
\begin{proof} 1) Let $\hat\eta$ be given by Example \ref{etahat}, by Proposition \ref{hatcon} the map $\hat\eta$ is continuous, denote $\sim_{\hat{G}}$ the orbit equivalence relation on $G\times X$ determined by $\hat{\eta}$ then by \cite[Theorem 3.3]{PU1} we get $X_G=(G\times X)/\!\sim_{\hat{G}}$. Thus using Theorem \ref{pro3.3} it is enough to check that $G*(G\times X)$ is closed in $G \times G\times X.$
In that sense, let $(g,(h,x))\notin G*(G\times X)$ then $(h,x)\notin G\times X_{g^{-1}}$ and $(g,x)\notin G*X,$ since $G*X$ is closed there are open sets $T\subseteq G$ and $U\subseteq X$ such that $(g,x)\subseteq T\times U \subseteq (G*X)^c$. Note that $(g,(h,x))\in T\times (G\times U)\subseteq (G*(G\times X))^c$. This shows that $G*(G\times X)$ is closed and we conclude that $X_G$ is profinite.
2)
We know that $X_G$ is profinite and that $\mu$ is continuous, we check that it is free.
Let $g\in G$ and $[h,x]\in X_G$ such that $[h,x]=\mu_g([h,x])=[gh,x]$, then by \eqref{eqgl} $x\in X_{h{}^{-1} gh}$ and $\eta_{h^{-1}g^{-1}h}(x)=x,$ thus $g=1$ because $\eta$ is free. This shows that $\mu$ is free and by Theorem \eqref{contgeral} we conclude the proof.
3) Suppose that $q:X/\!\sim_G\rightarrow X$ is a continuous section of $\pi_G$. Consider
$$s:\left(X_G/\sim_G\right)\ni G\cdot [1,x]\mapsto \iota(q(\pi_G(x))) \in X_G, $$ where $\iota$ is given by \eqref{iota}. We claim that $s$ is a continuous section of $\Pi_G$. First note that $s$ is well defined. In fact, let $x,y\in X$ such that $[1,x]\sim_G [1,y]$, we have that $\pi_G(x)=\pi_G(y)$. Let $z_x=q(\pi_G(x))$ and $z_y=q(\pi_G(y))$, then $\pi_G(z_x)=\pi_G(x)=\pi_G(y)=\pi_G(z_y)$. Thus \begin{center}
$s(G\cdot [1,x])=[1,z_x]=[1,z_y]=s(G\cdot [1,y])$,
\end{center}
and $s$ is well defined. Since $q$ and $\iota$ are continuous we get that $s$ is continuous. To finish the proof, take $x\in X$ and let $y_x\in X$ such that $q(\pi_G(x))=y_x$. Since $\pi_G(y_x)=\pi_G(x)$ there is $g\in G^x$ such that $\eta_g(x)=y_x$. Thus $\mu(g,[1,x])=[1,y_x]$ and we have
$$(\Pi_G\circ s)(G\cdot [1,x])=\Pi_G([1,y_x])=G\cdot [1,y_x]=G\cdot [1,x].$$ This shows that $s$ is continuous section of $\Pi_G$.
4) Let $r:\left(X/\!\sim_G\right)\ni G^x\cdot x\mapsto \iota^{-1}(q(G[1,x]))\in X.$ It is not difficult to check that $r$ is well defined, moreover the fact that $G*X$ is open implies that $\iota$ is open and thus $r$ is continuous. Finally take $x,z_x\in X$ such that $q(G[1,x])=[1,z_x],$ then $G[1,x]=G[1,z_x]$ which gives $\pi_G(x)=\pi_G(z_x),$ this implies $\pi_G(r(G^x\cdot x))=\pi_G(z_x)=\pi_G(x)$ and $r$ is a continuous section of $\pi_G$.
5) Let $q:X_G/\!\sim_G\rightarrow X_G$ and $t:X_G\rightarrow G\times X$ be continuous sections of $\Pi_G$ and $\hat{\Pi}_G,$ respectively. Define
\begin{center}
$p:X/\!\sim_G\ni G^x\cdot x\mapsto {\rm proj}_2(t(q(G\cdot [1,x]))) \in X.$
\end{center}
We check that $p$ is a continuous section of $\pi_G$. The map $p$ is well defined and continuous since $X/\sim_G\ni G^x\cdot x\mapsto G[1,x]\in X_G/\!\sim_G$ is continuous. On the other hand, take $x\in X$. Let $h,k\in G$ and $y,z\in X$ such that $q(G[1,x])=[h,y]$ and $t([h,y])=(k,z)$, then $G[1,x]=G[h,y]$ and $[k,z]=[h,y]$. Thus $\eta_{h^{-1}g}(x)=y$ and $\eta_{k^{-1}h}(y)=z$ for some $g\in G$, then $G^x\cdot x=G^z\cdot z$ and we have that $\pi_G(p(G^x\cdot x))=G^z\cdot z=G^x\cdot x$, as desired.
\end{proof}
It follows by \cite[Lemma 4.2]{PU2} that $\hat{\Pi}_G$ has a Borel section, provided that $G$ and $X$ are also a Polish group and a Polish space, respectively. Thus by 1) of Proposition \ref{cor3.3} and the proof 5) of the same Proposition we have the next.
\begin{pro}\label{borel} Let $\eta$ be a continuous partial action of a profinite polish group $G$ on a profinite Polish space $X$ such that $G*X$ is closed, then $\pi_G$ has a Borel section.
\end{pro}
\section{A remark on Exel's semigroup and the action groupoid}\label{exelgroup}
\subsection{The Exel's semigroup}
The Exel's semigroup $\mathcal{S}(G)$ of $G$ appeared for the first time in \cite{RUY3}, where it was shown that the actions of $\mathcal{S}(G)$ are in one-to-one correspondence with the partial actions of $G,$ both in the case of actions on a set. Later in \cite{KL} the authors proved that $\mathcal{S}(G)$ is isomorphic to the Birget-Rhodes expansion $\tilde{G}^R:=\{(A,g)\in \mathcal{P}(G)\times G: \{1,g\}\in A\}$ of $G,$ where $\mathcal{P}(G)$ denotes the finite subsets of $G,$ and thus $\mathcal{S}(G)$ is identified with $\tilde{G}^R.$
In the topological context this semigroup was considered in \cite{Choi} as follows: Let $G$ be a locally compact group and $\mathcal{K}_1(G)$ be the semilattice compact subsets of $G$ containing $1,$ we endow with the Vietoris topology, set
\begin{equation*}
\tilde{G}^R_c:=\{(A,g)\in \mathcal{K}_1(G)\times G: \{1,g\}\in A\},
\end{equation*} with product $(A, g)(B, h)=(A\cup gB, gh).$ It is shown in \cite{Choi} that $\tilde{G}^R_c$ is a topological inverse monoid.
Now suppose that $G$ is a profinite group, we shall check that $\tilde{G}^R_c$ is profinite. First of all it is not difficult to check that $\mathcal{K}_1(G)$ is a compact Hausdorff space, moreover by \cite[Proposition 8.6]{NW} the space $\mathcal{K}_1(G)$ is zero-dimensional, from this we get that $\mathcal{K}_1(G)\times G$ is profinite and thus it is enough to show that $\tilde{G}^R_c$ is closed in $\mathcal{K}_1(G)\times G$. In that sense, take $(A,g)\in ( \mathcal{K}_1(G)\times G) \setminus \tilde{G}^R_c$. Then $g\in A^c$ and there is a clopen $U\subseteq G$ such that $g\in U\subseteq A^c$. Note that $(A,g)\in \langle U^c\rangle\times U\subseteq ( \mathcal{K}_1(G)\times G)\setminus \tilde{G}^R_c$, and we conclude that $\tilde{G}^R_c$ is closed. Thus $\tilde{G}^R_c$ is profinite. The same reasoning is used to prove that $\tilde{G}^R$ is profinite and thus the Exel semigroup is profinite whenever $G$ is.
\subsection{The action groupoid}
The shortest definition of groupoid is the following: a groupoid $\mathcal G$ is a small category closed under inversions and in which every morphism is an isomorphism
According to \cite[p. 100]{KL} or \cite[p. 8]{Ruy} using a partial action $\eta$ of a group $G$ on a set $X$ we have the {\it action groupoid} $\mathcal G,$ where $\mathcal G= G*X$ as sets,
$\cdot^{-1}: \mathcal{G}\ni (g,x)\mapsto (g^{-1}, g\cdot x) \in \mathcal{G}$ is a unary and partial product $\ast: \mathcal{G}\times \mathcal{G}\longrightarrow \mathcal{G}$ is a partial binary operation defined as follows: if $(g,x), (h,y)\in \mathcal{G}$, then $(g,x)\ast (h,y)=(gh,y)$ provided that $x=h\cdot y$ and undefined otherwise. The groupoid ${\mathcal G}$ plays an important role in the connection of partial actions with star injective functor on ${\mathcal G}$ (see \cite[p. 602]{KL}) and as being isomorphic (as groupoids) to the graph groupoid it is also relevant in the study of $C^*$-algebras (see \cite{A2}).
If $G$ and $X$ are profinite and ${\mathcal G}=G*X$ is closed, then it is profinite as a topological space. It remains to check that the operations are continuous. We shall do this in the case that $G$ is finite (then discrete). We start with the inversion map. For this take $(g,x)\in \mathcal{G}$ and an open set $V\subseteq \mathcal{G}$ such that $(g^{-1},\eta_g(x))\in V$. Now there are open sets $T_1\subseteq G$ and $T_2\subseteq X$ such that $V=\mathcal{G}\cap T_1\times T_2$. Consider $W:=\lbrace g\rbrace\times \eta_{g^{-1}}(T_2\cap X_g)$, then $W\subseteq \mathcal{G}$ is open and $(g,x)\in W$. Let $(m,n)\in \cdot^{-1}(W)$, then there exists $(g,w)\in W$ such that $\cdot^{-1}(g,w)=(m,n)$ $(g^{-1},\theta_g(w))=(m,n)$ and we have $(m,n)\in V$ which shows that the map $\cdot^{-1}$ is continuous. Let us now show the continuity of $\ast,$ take $((g,\eta_h(x)); (h,x))\in \mathcal{G}\times \mathcal{G}$ and $U\subseteq \mathcal{G}$ is a open set such that $(gh,x)\in U$. As above there are open sets $U_1\subseteq G$ and $U_2\subseteq X$ such that $U=\mathcal{G}\cap U_1\times U_2$. Consider the open sets $W_1=\lbrace g\rbrace \times X_h,W_2=\lbrace h\rbrace \times (X_{h^{-1}}\cap U_2)$ and $W:=W_1\times W_2,$ then $((g,\theta_h(x)); (h,x))\in W.$ Finally, if $(s,t)\in \ast(W)$ there is $m\in X_{h^{-1}}\cap U_2$ such that $(gh,m)=\ast ((g,\theta_h(m));(h,m))=(s,t)$ and from this we get $(s,t)\in U$ and $\ast$ is continuous.
|
\section{Introduction} \label{sec:intro}
Molecular gas plays a prominent role in the evolution of galaxies by fuelling star-forming activities.
The bulk of the gas is the molecular hydrogen (H$_2$), but it does not have a permanent dipole moment.
A weak emission is only allowed at a high energy level, hampering direct observations of it in the typical ISM condition.
As alternative tracers, CO lines and dust have been used to weigh the total molecular gas in galaxies through the ``conversion factors" calibrated by the Milky Way and nearby galaxy observations.
With the advent of sensitive interferometric facilities such as the Atacama Large Millimeter/submillimeter Array (ALMA), the NOrthern Extended Millimeter Array (NOEMA), and the Karl G. Jansky Very Large Array (JVLA), the cosmic evolution of molecular gas up to $z\sim6$ is measured for many field galaxies (e.g., \citealt{Tacconi2020}).
However, it is still an open question whether the same calibration (or conversion factor) can be applied to galaxies in dense environments of clusters or their progenitors, protoclusters, because the calibration might be different from isolated field galaxies due to different excitation conditions, metallicity, and ambient radiation fields.
Such a test is necessary by using an independent method, whether the calibration of field galaxies can be applied in the same manner.
Fine structure lines of atomic carbon [\ion{C}{1}] is an alternative tracer of the molecular gas.
A traditional theoretical model of photodissociation regions (PDRs) viewed that the [\ion{C}{1}]\, line is originated from a narrow region at the interface between [\ion{C}{2}] and CO (e.g., \citealt{Tielens1985a, Tielens1985b, Hollenbach1999, Kaufman1999}),
but it was challenged by the observations that CO and [\ion{C}{1}]\, can coexist based on the constant column density ratio N([\ion{C}{1}]\,)/N(CO) over a wide range of ambient FUV field and physical conditions in the Orion clouds and Galactic center (\citealt{Ikeda1999,Ikeda2002,Ojha2001})
pointing to a potential use of [\ion{C}{1}]\, as a molecular gas tracer (\citealt{Gerin2000, Papadopoulos2004a}).
A tight, linear correlation between the CO~(1--0) and [\ion{C}{1}]\, line luminosities over five orders of magnitude for local galaxies strengthened this view (e.g., \citealt{Jiao2019}).
Not only is an independent molecular gas tracer, but [\ion{C}{1}]\, line can also be an efficient gas tracer than the lowest CO~(1--0) in terms of observing time with the current facilities.
The current JVLA can target CO (1--0) line for $z>1.3$ galaxies, observing the [\ion{C}{1}]\, can be more efficient than CO~(1--0) considering the positive K-correction at the same resolution and the robustness of [\ion{C}{1}]\,\,against CMB at very high redshifts ($z\gtrsim4$, \citealt{daCunha2013, Zhang2016}).
The frequency coverage of ALMA and NOEMA only allows $J>1$ transitions of CO for galaxies at $z\gtrsim0.5$ so that one needs an additional assumption of gas excitation to convert the line flux into the CO~(1--0) luminosity.
The lowest transition of [\ion{C}{1}]\, $^3P_1-$$^3P_0$ (hereafter written simply as the [\ion{C}{1}]\, (1--0) transition) is optically thin in most cases, less sensitive to excitation temperature compared to [\ion{C}{1}]\, $^3P_2-$$^3P_1$ (hereafter [\ion{C}{1}]\, (2--1)) when converting to a mass (\citealt{Weiss2005}).
Similar to CO~(1--0)-to-H$_2$ conversion factor, converting the [\ion{C}{1}]\, (1--0) line flux into a gas measurement needs an assumption of the [\ion{C}{1}]\,/[H$_2$] abundance ratio (hereafter $X_{\rm CI}$).
Recent calculations showed that $X_{\rm CI}$ is robust or well-behaved even in low metallicity and high cosmic ray environments (e.g., \citealt{Offner2013, Glover2016, Bisbas2017}).
So far, [\ion{C}{1}]\, line searches and detections at high-$z$ have happened in galaxies well above the star-forming main-sequence, or on the main-sequence in general fields (e.g., \citealt{Walter2011, Alaghband-Zadeh2013, Popping2017, Valentino2018, Valentino2020b} (hereafter V20); \citealt{Bourne2019}).
In this Paper, we perform a [CI] line search by targeting main-sequence galaxies associated with a protocluster at $z=2.49$, where CO and dust emissions are already detected (\citealt{minju2017a, minju2019a}, hereafter we refer to these two papers as \citetalias{minju2017a} and \citetalias{minju2019a}, respectively).
These main-sequence star-forming galaxies are H$\alpha$ emitters selected by a narrow-band filter technique using the Subaru telescope (\citealt{Tanaka2011}).
We report [\ion{C}{1}]\,(1--0) detections using ALMA and revisit the gas mass and compare gas masses derived from the three different tracers.
This paper is organized as follows.
In Section~\ref{sec:obs}, we present a summary of the ALMA Band 4 observations targeting the [\ion{C}{1}]\, line.
In Section~\ref{sec:results}, we show the [\ion{C}{1}]\, line detections and compare the line properties with the previous CO observations.
We discuss the derived molecular mass from different gas tracers in Section~\ref{sec:discussion}.
Finally, a summary is presented in Section~\ref{sec:summary}.
Throughout this paper, we assume a $\Lambda$ cold dark matter cosmology with $H_0$ =70 km s$^{-1}$ Mpc$^{-1}$, $\Omega _{\rm m}$ = 0.3, and $\Omega_{\Lambda}=0.7$.
The adopted initial mass function (IMF) is the Chabrier initial mass function (IMF; \citealt{Chabrier2003}) in the mass range of 0.1-100 $M_{\odot}$.
\section{Observations} \label{sec:obs}
\subsection{ALMA Band 4 observations} \label{sec:obsdetail}
We used ALMA Band~4 receivers to obtain the redshifted [\ion{C}{1}]\,~(1--0) emission (ID: ADS/JAO.ALMA \#2015.1.00152.S, PI: Minju Lee).
We designed our observations to use two-point positions, which is the same as the CO~(4--3) observations summarized in \citetalias{minju2019a}, covering a total of 16 out of 25 parent H$\alpha$ emitters (HAEs).
A total of either 38 or 39 antennas were used with the baseline length between 15 m and 3.1 km.
For each pointing, on-source time was $\simeq$95 mins to get the line emission.
We used four spectral windows (SPWs), two of each were placed in the upper and the lower sideband, respectively.
One SPW was set in the Frequency-Division Mode (FDM) to detect the redshifted [\ion{C}{1}]\, (1--0) with a channel width of 7.82 MHz ($\sim$ 16.6 km s$^{-1}$) covering 1.875 GHz bandwidth.
The remaining three SPWs were observed in the Time-Division Mode (TDM) with a 2.0-GHz bandwidth at the 15.6-MHz resolution to cover the dust continuum at 2 mm.
J2025+3343 and J2148+0657 were chosen for bandpass calibrators.
J2148+0657 and J1751+0939 were used for flux calibration.
A phase calibrator was J2114+2832.
We used $\mathtt{CASA}$ (\citealt{McMullin2007}) version 4.7.0 and 5.6.1 for the calibration of visibility data and imaging, respectively.
For visibility calibration, we used the pipeline script provided by the ALMA Regional Center staffs.
Images were produced by $\mathtt{CASA}$ task, $\mathtt{tclean}$, and deconvolved down to 2$\sigma$ noise level.
The CLEAN masks were created based on the positions of HAEs taking $1''$-radius circular regions.
The synthesized beam is $0^{\prime\prime}.42\times0^{\prime\prime}.29$ and the typical noise level is 0.10 mJy beam$^{-1}$ around the phase center at 80 km s$^{-1}$ for natural weighting.
We take the velocity resolution of 80 km s$^{-1}$ as a default, or otherwise specified.
Tapered images were also created by three different uvtaper parameters of $0''.3$, $0''.5$ and $0''.9$ in the $\mathtt{tclean}$ task to check the robustness of the line detections and the presence of extended emissions.
The corresponding beam sizes are $0''.53 \times 0''.43$, $0''.65 \times 0''.57$ and $0''.97 \times 0''.92$ for the uvtaper values of $0''.3$, $0''.5$ and $0''.9$, respectively.
We subtracted the continuum emission on the image domain using $\mathtt{imcontsub}$ (fitorder = 0) after we made smaller data cubes ($10''\times10''$) centered on the target HAEs.
This is to get improved results of continuum subtraction for galaxies not close to the phase center.
We then performed a primary beam correction.
Hereafter, we refer to the data cube that is corrected for the primary beam response and of which the continuum emission is subtracted.
\subsection{Detection criteria and flux measurements}\label{sec:criteria}
We impose the same detection criteria applied in the CO~(4--3) line detection (\citetalias{minju2019a}) that were found to be robust to identify line detections for known positions.
In summary, we regard a galaxy as detected by imposing double-step criteria.
Firstly, the [\ion{C}{1}]\, line detection candidates are selected if (a) a signal-to-noise (S/N) ratio at the peak position ($I_{\rm CI,peak}$) is equal to or greater than 4.5 ($S/N_{I_{\rm CI,peak}}\geq 4.5$).
We then regard these candidates as detected if the peak position spectrum satisfies either of the following criteria : (b) a peak flux density ($S_{\rm CI, peak}) \geq 3.5~\sigma$;
(c) at least two continuous channels, including a maximum-peak-flux channel have fluxes $> 2.5~\sigma$, where $\sigma$ is the average channel noise level, estimated from the line-free positions.
The noise estimate for the peak line intensity is based on the calculation after taking into account the channel noise, integrating channel range following \cite{Hainline2004} assuming the source is unresolved.
We also checked the noise level by bootstrapping from 30 random positions of the map, which matches the theoretical expectation.
The total line flux is measured by performing 2D gaussian fitting using $\mathtt{imfit}$ in CASA for a given aperture.
We check the growth curves in natural and tapered maps to find the optimal imaging parameters that give the highest S/Ns and total flux.
In summary, for HAE~3 and HAE~16, we use the $0''.3$-tapered map and a circular aperture size of $1''.0$.
For HAE~4, we use the natural-weighted map and a circular aperture size of $0''.7$ (but see Section~\ref{sec:det} for a different choice of aperture size for ``HAE~4b'').
For HAE~8 and HAE~9, we use the $0''.5$-tapered map and an aperture size of $1''.0$.
For HAE~8, we use 50 km s$^{-1}$ considering the narrow width known from the previous CO line detections.
For low S/N galaxies (HAE~4 and HAE~3), we also checked whether different velocity binnings (50 km s$^{-1}$ versus 80 km s$^{-1}$) change our results. We confirmed that the results did not change in terms of the line widths and fluxes within uncertainties between two different velocity binnings.
For [\ion{C}{1}]\,(1--0) non-detection, we calculate a 3$\sigma$ limit for galaxies by assuming the same line width and aperture sizes used in the CO~(4--3) if it is detected, and in other cases, we take a full-width half-maximum (FWHM) of 300 km $^{-1}$ for the integrated flux.
We note that except for HAE~5, which was not targeted for the [\ion{C}{1}]\,(1--0) and CO~(4--3) observations, galaxies with CO~(3--2) detection are a subset of the CO~(4--3) detected galaxy group, i.e., if CO~(3--2) line is detected, then CO~(4--3) is also detected. If a galaxy is not detected in any of the lines, we use the natural weight image and the aperture size of $0''.7$ for calculating the [\ion{C}{1}]\, flux limit.
The current detection criteria exclude a possible [\ion{C}{1}]\, detection with the {\it peak} S/N = 5.4 for HAE~7, which was previously detected in CO~(4--3), but not satisfying neither (b) nor (c) in Step 2 for the [\ion{C}{1}]\, analysis (the spectra are presented in Appendix~\ref{app:single}).
The 2D Gaussian best-fit flux ($0.28\pm0.09$) gives $\approx$10\% higher flux than the 3$\sigma$ upper limit placed by the noise level, but still within the uncertainty and it does not impact our later discussion.
The data cubes were also visually checked further for the candidate galaxies without any CO line detections because of the redshift uncertainty, which is $\Delta z\approx0.03$, coming from the original selection (i.e., narrow-band filter detected).
\subsection{Reanalysis of previous CO and dust measurements}\label{sec:reanalysis}
We reanalyzed the CO~(3--2), CO~(4--3), and dust data for unified analyses to reduce the systematic errors coming from different flux measurement methods.
The flux values in our previous studies were measured in different spatial and/or spectral resolutions and various methods.
Here we performed the same analyzing procedures applied in the measurement of the [\ion{C}{1}]\, line.
The beamsizes of the CO~(4--3), CO~(3--2) and dust are $0''.53 \times 0''.32$, $0''.89\times0''.66$, and $0''.78\times 0''.68$, respectively, for natural weighting.
As the beam size of the CO~(4--3) observations is very similar to that of [\ion{C}{1}]\, observations, we took the same aperture sizes and imaging parameters described in Section~\ref{sec:criteria} when doing $\mathtt{imfit}$.
We used the natural weighting map for CO~(3--2) and for dust continuum, taper parameter of $0''.5$, which gives the synthesized beam of $0''.89\times0''.82$.
For the CO~(3--2) data and dust, we used a fixed aperture size of $1''.0$ to measure the flux values.
The choice of the aperture sizes in flux measurement is based on our investigation to maximize the S/N.
However, we note that choosing $1''.0$ for all galaxies instead of different aperture sizes does not change our flux estimate.
Indeed, HAE~4 is the only galaxy using different aperture sizes between different lines, we note that [\ion{C}{1}]\, and CO~(4--3) fluxes using the $1''.0$-circular aperture for HAE~4 are consistent with the $0''.7$ measurement but with low S/N (but see Section~\ref{sec:det} for further discussion on HAE~4).
Since we probe close transitions between $J=3$ and $4$ for CO, the difference in radial distribution might be small.
We note, however, that for non-detection different aperture sizes can give $\approx40\%$ (or $\approx 0.14$ dex) change in the 3$\sigma$ upper limit.
The number of line detection does not change (i.e., seven galaxies for CO~(3--2), eleven galaxies for CO~(4--3), and four galaxies for dust), but the total flux values and signal-to-noise ratios are updated.
For CO~(4--3), HAE~3 and HAE~16 fluxes increased by $\approx40-50\%$ than \citetalias{minju2019a}. We also note that the S/Ns of HAE~1, 2, 7, 12 and 23, which are detected only in CO~(4--3), are lower than the previous measurements that we decided not to include in the line ratio analysis in Section~\ref{sec:lumiratio} (see their spectra in Appendix~\ref{app:single}).
The CO~(3--2) fluxes of HAE~8 and HAE~4 have increased roughly by a factor of two than the one in \citetalias{minju2017a}, where we took a peak flux from the smoothed maps, instead of fitting Gaussian for all galaxies (see the Appendix for the related discussions on the flux measurements in \citetalias{minju2017a}).
However, we note that the updated values are consistent with the values reported in \citet{Tadaki2019a} within uncertainties.
The updated fluxes do not change the main conclusion made in \citetalias{minju2017a} that (1) the mean gas fraction of galaxies with $\log{(M_{\rm star}/M_{\odot})} = [10.6, 11.3]$ is consistent with the field within the error bars and (2) the depletion time scale (or star-formation efficiency) decreases (increases) with increasing stellar mass but with an outlier of an AGN-dominated galaxy (HAE 5).
We summarize the flux values in Table~\ref{tab:lineflux}. Table~\ref{tab:lineflux} supersedes the previous measurements and we will regard the updated measurements for further analysis.
\section{Results} \label{sec:results}
\subsection{Detection of [\ion{C}{1}]\, and line properties compared to CO}\label{sec:det}
\begin{figure*}[tbh]
\centering
\includegraphics[width=0.85\textwidth, bb=0 0 1400 1800]{fig_spectrum.pdf
\caption{Left: [\ion{C}{1}]\, (1--0) line spectrum for detections at the [\ion{C}{1}]\, peak. Color-filled regions indicate the velocity range to integrate to derive the line flux for [\ion{C}{1}]\, (1--0), which gives the largest S/N. For comparison, CO~(4--3) and CO~(3--2) spectrum at the [\ion{C}{1}]\, peak position are shown with a base level shift of 2.0 and 1.0, respectively. The velocity resolution is 80 km s$^{-1}$ except for HAE~8. For HAE~8, the resolution is 50 km s$^{-1}$. Middle: [\ion{C}{1}]\, (1--0) line spectrum using a $1''.0$ aperture centered at the peak. CO~(3--2) spectrum using the same aperture size at the CO~(3--2) peak is plotted in black thinner line. Right: the velocity-integrated map of [\ion{C}{1}]\, (1--0). The size of the panel is $3''\times 3''$. Cross symbols indicate the peak positions of CO~(4--3) (black), CO~(3--2) (red), and H$\alpha$ (white) emissions. The contour levels are starting from $4\sigma$ in steps of $1\sigma$. Dashed contours are also plotted for $-4\sigma$ indicating no 4-$\sigma$ peak around the galaxies. The synthesized beam ($0''.53\times 0''.43$) is shown as an filled-ellipse on the bottom left of each panel. \label{fig:lines}}
\end{figure*}
We detect the [\ion{C}{1}]\, line from five galaxies previously detected in CO~(3--2) and CO~(4--3) lines, i.e., HAE~3, 4, 8, 9 and 16.
Three galaxies, HAE~3, 8, and 9, were detected in the 1.1 mm continuum data as well.
In the left panels of Figure~\ref{fig:lines}, we show the spectra of [\ion{C}{1}]\,, CO~(3--2), and CO~(4--3) emissions at the peak position of the [\ion{C}{1}]\,(1--0) emission for [CI] detections.
The spectra of the [\ion{C}{1}]\, and CO~(4--3) lines are taken from the $0''.3$-tapered maps and CO~(3--2) from the natural-weighted maps.
We find the line properties (widths and redshifts) of [\ion{C}{1}]\, centered at each peak position of the [\ion{C}{1}]\, and CO~(3--2) with a fixed aperture ($1''.0$) size are broadly consistent with those of CO~(3--2), except for HAE~3 and HAE~9 (Figure~\ref{fig:lines} middle panels).
These two galaxies show signatures of mergers in their stellar and/or CO maps (\citetalias{minju2017a,minju2019a}).
For HAE~3, the [\ion{C}{1}]\, line is narrower (FWHM = 172 $\pm$ 106 km s$^{-1}$) than two CO lines (FWHM = 500 km s$^{-1}$) which may indicate kinematically different origins of the emission lines. If we integrate the [C I] cube of HAE~3 over the same velocity range defined by two CO lines, the 1D Gaussian fitted flux ($0.15\pm0.08$) is consistent within uncertainties.
For HAE~9, we discussed its kinematic properties in \citetalias{minju2019a} that the galaxy is undergoing a gas-rich merger.
The red-shifted component of HAE~9 is not detected in [\ion{C}{1}]\, (or does not satisfy the detection criteria), which was detected in CO~(4--3).
Accordingly, in Table~\ref{tab:lineflux}, we list additional flux measurements for HAE~3 and HAE~9 (named HAE~3a and HAE~9a) to refer to the flux values for which integrating ranges are matched to those of the CO~(3--2) lines. In the later analysis of the gas measurements, we will use these values.
The peak positions of [\ion{C}{1}]\, are also consistent within the positional errors except for HAE~4 if the beam size and the S/Ns are taken into account (Figure~\ref{fig:lines} right panels).
HAE~4 shows a positional offset of $0.6''$ between H$\alpha$ and CO~(3--2) peak positions (\citetalias{minju2017a}) and extended H$\alpha$ emission.
The CO~(4--3) peak is closer to the H$\alpha$ peak in contrast to the offset from CO~(3--2) peak position.
The [\ion{C}{1}]\, line peak is also offset from the other line peak positions by $\approx 0''.5-0''.6$ ($\approx$ 4-5 kpc).
It may indicate different gas excitation conditions within the galaxy as discussed in \citetalias{minju2019a}.
For this reason, we take another flux measurement for HAE~4.
We list the line fluxes from different positions which give the highest S/N for individual emission lines (HAE~4), and from the CO~(4--3) peak position (HAE~4b, see Appendix~\ref{app:haes} for its spectrum).
For HAE~4b, we derive the fluxes using a $1''.0$ circular aperture to be consistent with other [\ion{C}{1}]\, detected galaxies.
Owing to the low S/N at the CO~(4--3) position, we do not derive the flux in the 2D map for HAE~4b using $\texttt{imfit}$ task in CASA, but by fitting the 1D spectrum with a single Gaussian.
We note that the aperture photometry in the 2D map, using the same circular aperture of $1''.0$, is lower than or comparable to the 1D Gaussian fit listed in Table~\ref{tab:lineflux}, i.e., $0.13\pm0.03$, $0.14\pm0.04$, $0.20\pm0.05$ for [\ion{C}{1}]\,, CO~(3--2) and CO~(4--3), respectively.
Other galaxies also show similar signatures in their CO and [\ion{C}{1}]\, line profiles; the line profiles taken from the the [\ion{C}{1}]\, peak position show a varying degree of relative line ratios at each velocity bin (Figure~\ref{fig:lines} left panels).
Discussing the resolved gas excitation conditions is beyond the scope of this paper, and this would need more sensitive and higher angular resolution observations to get more details.
In this paper, we assume that the line emissions are {\it globally} coming from the same galaxy.
\begin{deluxetable*}{ccccccccccc}
\tablecaption{{Summary of line observations for HAEs associated to 4C23.56 protocluster}\label{tab:lineflux}}
\tabletypesize{\footnotesize}
\tablewidth{0.99\textwidth}
\tablehead{
\colhead{ID} & \colhead{R.A.} & \colhead{Decl.} & \colhead{[\ion{C}{1}]\, beam} &\colhead{aperture([CI])} & \colhead{$\Delta v({\rm [C I]})_{\rm integ.}$} & \colhead{FWHM$_{\rm [CI](1-0)}$} & \colhead{$I_{\rm [CI]10}$} &\colhead{$I_{\rm CO32}$} & \colhead{$I_{\rm CO43}$} & \colhead{$S_{\rm 1.1mm}$} \\
\colhead{} & \colhead{J2000} & \colhead{J2000} & \colhead{$''\times''$} & \colhead{$''$} & \colhead{km s$^{-1}$} & \colhead{km s$^{-1}$} & \colhead{Jy km s$^{-1}$}& \colhead{Jy km s$^{-1}$} & \colhead{Jy km s$^{-1}$} & \colhead{mJy} \\
\colhead{(1)} & \colhead{(2)} & \colhead{(3)} &\colhead{(4)} & \colhead{(5)} & \colhead{(6)} & \colhead{(7)} & \colhead{(8)} & \colhead{(9)} & \colhead{(10)} & \colhead{(11)}}
\startdata
HAE1\tablenotemark{*} & 316.81166 & 23.52921 & 0.42$\times$0.29 & 0.7 & 560 & ... & $<0.17$ & $<0.18$ & 0.34$\pm$0.13 & $<0.25$\\
HAE2\tablenotemark{*} & 316.84074 & 23.53043 & 0.42$\times$0.29 & 0.7 & 400 & ... & $<0.14$ & $<0.12$ & 0.16$\pm$0.07 & $<0.36$\\
HAE3 & 316.83763 & 23.52055 & 0.53$\times$0.43 & 1.0 & 160 & 172$\pm$106 & 0.17$\pm$0.05 & 0.43$\pm$0.12 & 0.67$\pm$0.17 & 0.77$\pm$0.26\\
HAE3a\tablenotemark{$\dagger$} & 316.83763 & 23.52055 & 0.53$\times$0.43 & 1.0 & 560 & 172$\pm$106 & 0.15$\pm$0.08 & 0.43$\pm$0.12 & 0.63$\pm$0.18 & 0.77$\pm$0.26\\
HAE4 & 316.84013 & 23.52810 & 0.42$\times$0.29 & 0.7 & 480 & 228$\pm$130 & 0.20$\pm$0.06 & 0.42$\pm$0.10 & 0.19$\pm$0.06 & $<0.42$\\
HAE4b\tablenotemark{$\ddag$} & 316.84013 & 23.52810 & 0.53$\times$0.43 & 1.0 & 480 & 292$\pm$138 & 0.13$\pm$0.08 & 0.17$\pm$0.07 & 0.22$\pm$0.10 & $<0.42$\\
HAE5 & 316.82069 & 23.50846 & ... & ... & ... & ... & ... & 0.14$\pm$0.04 & ... & $<0.29$\\
HAE6 & 316.83955 & 23.52209 & 0.42$\times$0.29 & 0.7 & 300 & ... & $<0.09$ & $<0.14$ & $<0.11$ & $<0.31$\\
HAE7\tablenotemark{*} & 316.81468 & 23.52707 & 0.53$\times$0.43 & 1.0 & 1520 & ... & $<0.25$ & $<0.23$ & 0.43$\pm$0.12 & $<0.51$\\
HAE8 & 316.81638 & 23.52431 & 0.65$\times$0.57 & 1.0 & 150 & 154$\pm$62 & 0.24$\pm$0.04 & 0.55$\pm$0.09 & 0.49$\pm$0.06 & 0.90$\pm$0.19\\
HAE9 & 316.84409 & 23.52871 & 0.65$\times$0.57 & 1.0 & 560 & 410$\pm$171 & 0.38$\pm$0.09 & 0.64$\pm$0.10 & 0.61$\pm$0.12 & 1.52$\pm$0.29\\
HAE9a\tablenotemark{$\dagger$} & 316.84409 & 23.52871 & 0.65$\times$0.57 & 1.0 & 800 & 410$\pm$171 & 0.55$\pm$0.14 & 0.64$\pm$0.10 & 1.23$\pm$0.23 & 1.52$\pm$0.29\\
HAE10 & 316.81543 & 23.52003 & 0.53$\times$0.43 & 1.0 & 480 & ... & $<0.17$ & 0.53$\pm$0.12 & 0.26$\pm$0.10 & 0.50$\pm$0.16\\
HAE11 & 316.79338 & 23.51373 & ... & ... & ... & ... & ... & ..... & ...\\
HAE12\tablenotemark{*} & 316.81222 & 23.52988 & 0.42$\times$0.29 & 0.7 & 240 & ... & $<0.13$ & $<0.12$ & 0.19$\pm$0.07 & $<0.25$\\
HAE13 & 316.84092 & 23.52826 & 0.42$\times$0.29 & 0.7 & 300 & ... & $<0.09$ & $<0.10$ & $<0.11$ & $<0.32$\\
HAE14 & 316.83241 & 23.51417 & ... & 0.7 & 300 & ... & ... & $<0.11$ & ... & $<1.05$\\
HAE15 & 316.83315 & 23.51896 & ... & 0.7 & 300 & ... & ... & $<0.10$ & ... & ...\\
HAE16 & 316.81102 & 23.52107 & 0.53$\times$0.43 & 1.0 & 800 & 453$\pm$201 & 0.33$\pm$0.07 & 0.57$\pm$0.14 & 0.82$\pm$0.12 & $<0.94$\\
HAE17 & 316.82340 & 23.53068 & ... & ... & 300 & ... & ... & $<0.20$ & ... & ...\\
HAE18 & 316.84011 & 23.53366 & ... & ... & 300 & ... & ... & $<0.14$ & ... & ...\\
HAE19 & 316.84247 & 23.52944 & 0.42$\times$0.29 & 0.7 & 300 & ... & $<0.11$ & $<0.10$ & $<0.13$ & $<0.27$\\
HAE20 & 316.81228 & 23.52238 & 0.42$\times$0.29 & 0.7 & 300 & ... & $<0.08$ & $<0.13$ & $<0.10$ & ...\\
HAE21 & 316.81175 & 23.52857 & 0.42$\times$0.29 & 0.7 & 300 & ... & $<0.11$ & $<0.13$ & $<0.13$ & $<0.26$\\
HAE22 & 316.82441 & 23.52909 & ... & ... & 300 & ... & ... & $<0.20$ & ... & ...\\
HAE23\tablenotemark{*} & 316.81147 & 23.52184 & 0.42$\times$0.29 & 0.7 & 800 & ... & $<0.13$ & $<0.23$ & 0.27$\pm$0.09 & ...\\
HAE24 & 316.84344 & 23.54678 & ... & ... & ... & ... & ... & ..... & ...\\
HAE25 & 316.80971 & 23.53572 & ... & ... & ... & ... & ... & ..... & ...\\
\enddata
\tablecomments{\\
$\dagger$: flux values obtained by integrating the velocity range used in CO~(3--2) line (column 6) \\$\ddagger$: flux values at a fixed position to the CO~(4--3) peak, which is the closest ($\Delta r= 0''.15$) to the original H$\alpha$ emission. \\
$*$: HAE~1, 2, 7, 12, and 23 satisfy our detection criteria in CO~(4--3) but we note that they have lower significance levels of the line detection in comparison with other secure detections those have other line and/or dust detections (see Appendix~\ref{app:single} for their spectra). Therefore, we suggest readers use these values with caution.\\
Colums: (1) Source ID. For sources without a dagger mark($\dagger$) or a double dagger mark ($\ddag$) behind the ID, we list fluxes integrated over the velocity range with beams and aperture sizes that maximizes the S/N of the peak position flux and total flux, centered at the peak position of each line; (2) Right Ascension (J2000) of H$\alpha$ position from Subaru/MOIRCS observations (\citealt{Tanaka2011}; I. Tanaka in preparation); (3) Declination (J2000) of H$\alpha$ position; (4) Synthesized beam of the [\ion{C}{1}]\, images to measure the flux; (5) Aperture size to measure the [\ion{C}{1}]\, flux or the 3$\sigma$ upper limit; (6) Velocity integrating range to measure the [\ion{C}{1}]\, flux. If the [\ion{C}{1}]\, line is non-detection but other lines (i.e., CO~(4--3) or CO~(3--2)) are detected, we used the integrating range referring to their integrating ranges that is listed in this column. If no other lines are detected, we set the integrating range fixed as 300 km s$^{-1}$ for 3$\sigma$ upper limits; (7) Full-width half-maximum of the [\ion{C}{1}]\, line fitted with a single 1D Gaussian profile; (8) [\ion{C}{1}]\, (1--0) flux measured using CASA $\texttt{imfit}$ (but 1D Gaussian fit for HAE~4b) with the aperture size listed in Col. (5) and (6). Upper limits are 3$\sigma$; (9) CO~(3--2) flux using CASA $\texttt{imfit}$ (but 1D Gaussian fit for HAE~4b) with an aperture of $1''.0$. Upper limits are 3$\sigma$ using the value listed in Col (5) and (6); (10) CO~(4--3) flux using CASA $\texttt{imfit}$ (but 1D Gaussian fit for HAE~4b) with the same aperture size listed in Col. 5. Upper limits are 3$\sigma$ using the value listed in Col (5) and (6); (11) 1.1 mm dust continuum flux using an aperture of $1''.0$. All measurements presented here supersede the previous measurement reported in \citetalias{minju2017a} and \citetalias{minju2019a}.}
\end{deluxetable*}
\begin{figure}[tb]
\centering
\includegraphics[width=0.42\textwidth, bb=0 0 900 1800]{fig_lumi_ver5.pdf}
\caption{The relations between [\ion{C}{1}]\, luminosity and CO (4--3) luminosity (top) and between [\ion{C}{1}]\, luminosity and CO~(3--2) luminosity (bottom). Protocluster galaxies are shown as squares, and smaller squares (HAE~4 and ~9) are excluded in the gas mass calibration in Section~\ref{sec:calmass} (see the text and Table~\ref{tab:lineflux} caption for the description of the galaxy name tags). We also plot other data points from \citetalias{Valentino2020b} (circles), \citet{Spilker2014} (diamond), \citet{Boogaard2020} (crosses), and \citet{Popping2017} (X-cross). In each panel, the 3$\sigma$ upper limits for [CI] non-detection with CO line detections are also plotted; arrows for the protocluster members and left faced triangles for others. Red dashed and blue solid lines are the best-fit using the data points excluding the protocluster galaxies (see Table~\ref{tab:lumifit} and the main text). Using six secure detections (HAE~3a, 4b, 8, 9a, 10 and 16) with at least two line detections, dash-dotted and dotted lines are the best-fit depending on the model assumptions. Shaded regions indicate 1$\sigma$ uncertainties of the best-fit parameters for varying slopes.\label{fig:luminosity}}
\end{figure}
\subsection{The $L'_{\rm [C I](1-0)}$ -- $L'_{\rm CO(J-J-1)}$ relation}\label{sec:lumiratio}
\begin{deluxetable*}{ccccccc}
\tablecaption{$L'_{\rm CO}$ and $L'_{\rm CI}$ relation fit\label{tab:lumifit}}
\tabletypesize{\footnotesize}
\tablewidth{0.99\textwidth}
\tablehead{
\colhead{} & \multicolumn{3}{c}{$\log{L'_{\rm CO43}} = a \times (\log{L'_{\rm CI10}}-10) + b$}& \multicolumn{2}{c}{$\log{L'_{\rm CO43}} = a_{\rm fixed} \times (\log{L'_{\rm CI10}}-10) + b$}\\
\colhead{References} & \colhead{(1)+(2)+(3)} & \colhead{high-$z$ MS gals in (1), and (2)}& \colhead{protocluster (this work)} &\multicolumn{2}{c}{protocluster (this work)}}
\startdata
a & $1.04\pm0.02$ & $1.32\pm0.22$ & $1.16\pm0.23$ & fixed to $1.04$ & fixed to $1.32$\\
b & $10.32\pm0.02$ & $10.47\pm0.12$ & $10.46\pm0.08$ & $10.42\pm0.03$ & $10.51\pm0.03$\\
\hline
\\
\hline\hline
{} & \multicolumn{3}{c}{$\log{L'_{\rm CO32}} = a \times (\log{L'_{\rm CI10}}-10) + b$} & \multicolumn{2}{c}{$\log{L'_{\rm CO32}} = a_{\rm fixed} \times (\log{L'_{\rm CI10}}-10) + b$}\\
References& \multicolumn{2}{c}{(1)+(3)\tablenotemark{$*$}} & protocluster (this work) &\multicolumn{2}{c}{protocluster (this work)}\\
\hline
a& \multicolumn{2}{c}{$1.32\pm0.10$} & $0.43\pm0.20$ & \multicolumn{2}{c}{fixed to $1.32$}\\
b& \multicolumn{2}{c}{$10.36\pm0.05$} & $10.39\pm0.08$& \multicolumn{2}{c}{$10.70 \pm 0.08$}\\
\enddata
\tablecomments{
References (1) = \citetalias{Valentino2020b}, (2) = \citet{Boogaard2020}, (3) = \citet{Spilker2014}\\
$*$: We note that for this fitting, the sample is largely dominated by high-$z$ DSFGs and only one is a main-sequence galaxy at $z=2.2$ from \citet{Popping2017} that observed a compact star-forming galaxy.\\
The fit for the protocluster is using six secure galaxies at least two line detections (i.e., HAE~3a, 4b, 8, 9a, 10 and 16). If we include other [\ion{C}{1}]\, \,3$\sigma$ upper limits of the remaining galaxies for CO~(4--3) (see Section~\ref{sec:reanalysis} and Table~\ref{tab:lineflux} note regarding single line detections), the best-fit parameters change slightly but still consistent within the uncertainties; slope for CO~(4--3) becomes steeper with higher normalization, i.e., $a=1.40 \pm 0.26$, $b = 10.55 \pm 0.13$; if we fix the slope, $b= 10.39\pm0.04$ and $10.52 \pm 0.04$ for $a=1.04$ and $1.32$, respectively.}
\end{deluxetable*}
We compare the line luminosities between the [\ion{C}{1}]\,(1--0) and CO lines.
We first fit the literature values compiled in \citetalias{Valentino2020b} and $z\sim1$ galaxies from the ASPECS survey (\citealt{Boogaard2020}) as well as the mean values of the SPT sources from \citet{Spilker2014}.
The data from \citetalias{Valentino2020b} includes local IR-luminous galaxies (LIRGs), high-$z$ main-sequence galaxies, where most of the samples are at $z\sim1$ except for one at $z=2.2$, and high-$z$ dusty star-forming galaxies (DSFGs) and QSOs. We note that the local galaxies compiled in \citetalias{Valentino2020b} are representative of star-bursting population rather than typical spiral galaxies.
The mean deviation of these local galaxies from the main-sequence defined by \citet{Speagle2014} is $\langle\log\Delta {\rm MS}\rangle=1.3$. Hereafter, we define ``starburst'' galaxies to be galaxies well-above the main-sequence by 3$\sigma$ from the main-sequence (i.e., $\log{\Delta{\rm MS}}>0.6$).
We fit the data using the orthogonal distance regression to take into account uncertainties in both axes.
For fitting, we have included the non-detections using the 3-$\sigma$ upper limits if another line is detected, and the uncertainties of all data points are set to the 1-$\sigma$ noise level.
We use a linear model of $\log{L'_{\rm CO43\,,or\,CO32}} = a \times (\log{L'_{\rm CI10}}-10) + b$.
We fit for $L'_{\rm CO43}-L'_{\rm CI10}$ relation by excluding star-bursting population of local galaxies and high-$z$ DSFGs (i.e., $z\simeq1-2$ main-sequence galaxies only) as well.
For $L'_{\rm CO32}-L'_{\rm CI10}$ relation, the data sets become sparse and most of the galaxies available are high-$z$ DSFGs except for one from \citet{Popping2017}, which is a compact star-forming galaxy (cSFG) at $z=2.2$.
The best-fit slope values indicate super-linear relations between $L'_{\rm CO}$ and $L'_{\rm [CI](1-0)}$.
{A summary of the best-fit parameters is listed in Table~\ref{tab:lumifit}. Figure~\ref{fig:luminosity} shows the best-fit relations (red dashed -- all galaxies in the literature -- and blue solid lines -- high-$z$ main-sequence galaxies --).
We also note that the residual variance is higher for the CO~(3--2)--[C I] fit, partially owing to the limited data points. Simply by taking the best-fit models listed in the second column in Table~\ref{tab:lumifit}, i.e., (1)+(2)+(3) fit and (1)+(3) fit for $L'_{\rm CO43}-L'_{\rm CI10}$ and $L'_{\rm CO32}-L'_{\rm CI10}$, respectively), the brightness temperature (luminosity) ratio between CO~(4--3) and CO~(3--2) ($R_{43} = L'_{\rm CO43}/L'_{\rm CO32}$) becomes higher (smaller) for smaller (higher) $L'_{\rm [CI](1-0)}$ luminosity.
We perform the same exercise for the protocluster galaxies only while keeping in mind that the dynamic range is much narrower i.e., less than an order of magnitude in luminosities.
We fit using six galaxies with secure detections (i.e., at least two lines of detection; HAE~3a, 4b, 8, 9a, 10 and 16).
The best-fit parameters are summarized in Table~\ref{tab:lumifit} using the same fitting model but with two options of a variable slope and a fixed slope.
For $\log{L'_{\rm CO(4-3)}}-\log{L'_{\rm [CI](1-0)}}$ relation, the slope is super-linear for the protocluster members, consistent with other best-fit using the literature values.
The super-linear relation between $\log{L'_{\rm CO(4-3)}}$ and $\log{L'_{\rm [CI](1-0)}}$ may indicate that CO~(4--3) emission is not tracing the total gas mass, but either dense gas that directly forms stars and, hence star-formation, while the [\ion{C}{1}]\, emission is tracing the total gas mass, mimicking the global Kennicutt-Schmidt law (\citealt{Schmidt1959, Kennicutt1998a}).
Also, it may be worth noting that there is a hint of a steeper super-linear slope for high-$z$ galaxies than our protocluster galaxies.
Concerning this, there are a few studies in local galaxies (e.g., \citealt{Juneau2009, Bayet2009a}) that reported a super-linear correlation between far-infrared luminosity and low-$J$ CO transitions ($J<3$) spanning a wide IR luminosity range.
Some local galaxy surveys reported a linear relation instead (e.g., \citealt{Gao2004, Greve2014}) for $L_{FIR}>10^{11}\,L_{\odot}$ galaxies.
The difference is often explained as the different dense gas fractions (i.e, constantly higher dense gas fraction for galaxies with high IR luminosity; e.g., \citealt{Gao2004, Wu2005, Greve2014}).
\cite{Greve2014} discussed further for the $L_{\rm FIR}$-low-$J$ CO relation that a linear slope or higher normalization factor may be physically related to the existence of dense gas, which is limited by the Eddington limit (see also \citealt{Andrews2011}).
If we can apply a similar interpretation, a hint of a shallower slope (close to a linear relation) and a higher normalization factor in protocluster galaxies than other high-$z$ main-sequence galaxies might hint at the existence of more denser gas in protocluster galaxies than high-$z$ field main-sequence galaxies.
There is a hint of a systematically different relation in the protocluster for the $L'_{\rm CO(3-2)}-L'_{\rm [CI](1-0)}$ relation (at $1\sigma$).
Considering that the fit for the literature values is biased towards high-$z$ DSFGs and the fit for protocluster galaxies with a varying slope shows a sub-linear relation, our previous simple interpretation of $L'_{\rm CO}-L'_{\rm [CI](1-0)}$ as an analog of $L_{\rm FIR}-L'_{\rm CO}$ (or the K-S law) might not be always viable.
Or, the CO excitation condition might be different from high-$z$ DSFGs while CO~(3--2) line still traces dense gas.
Regarding this, it may be worth noting that the line ratio ($L'_{\rm CO(3-2)}/L'_{\rm [CI](1-0)}$) of cSFG from \citet{Popping2017} is comparable to the protocluster galaxies.
In any case, it is a surprising result to see such a large difference in these two close transitions but in a (similarly) narrow dynamic range.
The current data set is not sufficient to robustly conclude because of the the low number statistics and unknown systematic biases for currently available data sets in different environments.
Depending on our model selection for the fit (fixed slope or not), the relation between $L'_{CO43}/L'_{CO32}$ as a function of $L'_{\rm CI10}$ also varies (negative or positive), which further denotes the current limitation.
Therefore, we defer further discussions on the existence of systematically $L'_{\rm CO}-L'_{\rm [CI](1-0)}$ relations in different environments to a future work. We will have to wait for more data from future surveys for fairer comparison.
\subsection{Molecular gas mass from [\ion{C}{1}]\,, CO and dust}
\begin{deluxetable}{cccc}
\tablecaption{Derived gas mass based on [\ion{C}{1}]\,, CO and dust\label{tab:gasmass}}
\tabletypesize{\footnotesize}
\tablewidth{0.99\textwidth}
\tablehead{
\colhead{ID} & \colhead{$M_{\rm mol, CI}$} & \colhead{$M_{\rm mol, CO}$} & \colhead{$M_{\rm mol, RJ}$}\\
\colhead{} & {$\times 10^{10}\,M_{\odot}$} & {$\times 10^{10}\,M_{\odot}$} & {$\times 10^{10}\,M_{\odot}$}}
\startdata
HAE3a & 5.5$\pm$3.0 & 11.4$\pm$3.2 & 8.0$\pm$2.7\\
HAE4 & 7.4$\pm$2.2 & 11.2$\pm$2.7 & $<$ 4.4\\
HAE4b & 4.8$\pm$3.0 & 4.5$\pm$1.9 & $<$ 4.4\\
HAE8 & 8.9$\pm$1.5 & 14.6$\pm$2.4 & 9.4$\pm$2.0\\
HAE9a & 20.3$\pm$5.2 & 16.9$\pm$2.6 & 15.9$\pm$3.0\\
HAE10 & $<$ 6.3 & 14.1$\pm$3.2 & 5.2$\pm$1.7\\
HAE16 & 12.2$\pm$2.6 & 15.1$\pm$3.7 & $<$ 9.8\\
\hline
\enddata
\end{deluxetable}
\begin{figure}[t]
\centering
\includegraphics[width=0.48\textwidth, bb=0 0 1000 1000]{fig_mass_ver2.pdf}
\caption{Comparison between the derived molecular gas mass from [\ion{C}{1}]\,, CO and dust. Squares show the measurements from this work using 1.1 mm (brown) and CO~(3--2) (turquoise). Upper limits (3$\sigma$) are plotted as triangles. We compare our results with $z\sim1$ galaxies on the main-sequence and local galaxies compiled in \citetalias{Valentino2020b}, where both CO~(2--1) and [\ion{C}{1}]\, fluxes are available. For a comparison of dust-based measurements, we plot data points from \citet{Bourne2019} using the 1.2mm detections. The dashed line is to guide 1:1 relation. \label{fig:gasmass}}
\end{figure}
The relation between CO and [\ion{C}{1}]\, demonstrates that [\ion{C}{1}]\, traces the global gas mass well.
We derive gas mass from [\ion{C}{1}]\,, CO and dust to investigate whether the general assumptions are valid to protocluster galaxies.
Following \citet{Papadopoulos2004a}, we estimate the total molecular gas mass from [\ion{C}{1}]\, as
\begin{equation}
\frac{M^{\rm [CI]}_{\rm mol}}{M_{\odot}} = \frac{4 \pi \mu m_{\rm H_2}}{h\,c\,A_{10} \,X_{\rm CI} \, Q_{10}} \left( \frac{d^2_{\rm L}}{1+z}\right) S'_{\rm [CI]} \, \Delta V
\end{equation}
where $\mu = 1.36$ accounts for the mass of helium and $m_{\rm H_2}$ is the molecular mass of H$_2$ in $M_{\odot}$.
For the [\ion{C}{1}]\,(1--0) transition, the Einstein coefficient A$_{10}$ is $7.93\times10^{-8}$ s$^{-1}$ and the excitation factor $Q_{10}$ is determined by the gas density and temperature.
Without measurements of both [\ion{C}{1}]\,(1--0) and [\ion{C}{1}]\,(2--1), we cannot measure $Q_{10}$.
Here, we adopt a moderate value of $Q_{10} = 0.35$ for typical conditions of kinematic temperature $20<T_{\rm kin}<40$ K and density $300<n<10^4$ cm$^{-3}$ based on \citet{Papadopoulos2012b} and \citet{Jiao2017}.
$S'_{\rm [CI]} \, \Delta V$ is the measured [\ion{C}{1}]\, flux in Jy km s$^{-1}$. The remainders are luminosity distance $d_L$, speed of light $c$, and Planck constant in SI unit.
To convert the [\ion{C}{1}]\, mass into the H$_2$ mass, we need to assume $X_{\rm CI}$.
Observational constraints of the $\log{(X_{\rm CI})}$ for the Milky Way are in a range of [-5.7, -4.7] (\citealt{Frerking1989}) and the typically assumed value is -4.5 (\citealt{Weiss2003}).
For local galaxies without AGNs, based on the comparison with low $J$ CO-based gas measurements, they have higher abundance ratio of $\log{(X_{\rm CI})} = -4.2\pm0.2$ and comparable to high-$z$ DSFGs ([-3.9, -4.2], \citealt{Walter2011, Danielson2011, Alaghband-Zadeh2013, Valentino2018}).
Here, we adopt the mean value of $\log(X_{\rm CI}) = -4.8$ found in \citet{Valentino2018} that studied main-sequence galaxies at $z\sim1$.
For CO~(3--2) and dust, we use the same recipe adopted in \citetalias{minju2017a}.
In brief, we take a line luminosity ratio between CO~(3--2) and CO~(1--0) ($R_{13}$ = 1.9), a similar value measured or used in \citet{Dannerbauer2009}, and \citet{Tacconi2018} for $z=1-3$ main-sequence galaxies, but see e.g., \citet{Bolatto2015, Riechers2020, Boogaard2020} for a lower value, or a slightly higher average value reported in \citet{Daddi2015}.
We use the Milky Way (MW)-like CO-to-H$_2$ conversion factor ($\alpha_{\rm CO} = 4.36$~M$_{\odot}$ (K km s$^{-1}$ pc$^{2}$)$^{-1}$; e.g., \citealt{Bolatto2013, Genzel2015}).
Here, we don't correct for the metallicity-dependence of the $\alpha_{\rm CO}$ conversion factor.
For the underlying stellar mass range that we observe ($\log{(M_{\rm star}/M_{\odot})}>10.5$), the metallicity-dependent $\alpha_{\rm CO}$ factor only increases the gas estimate the up to 0.07 dex for the lowest mass galaxies (HAE~16 and HAE~10) and does not affect higher mass regime.
For dust-based measurements, we adopt the calibration described in \citet{Scoville2016}.
The details of the calculations are presented in \citetalias{minju2017a}.
Figure~\ref{fig:gasmass} compares the molecular gas masses derived from three different tracers, and Table~\ref{tab:gasmass} summarizes the derived values.
The gas mass derived from CO is different by -0.08 dex up to 0.35 dex compared to the [\ion{C}{1}]\,-based gas mass and 0.14 dex higher on average.
Meanwhile, the dust-based measurements are lower than the [\ion{C}{1}]\,-based mass by 0.02 dex on average, and the differences range between -0.11 dex and 0.16 dex.
The mean values of the difference compared to the [\ion{C}{1}]\,-based measurement are shown as the arrows along the vertical line of the top-left panel of Figure~\ref{fig:dev}.
In summary, three gas tracers are broadly in good agreement for the inferred molecular masses within $\approx$0.3 dex, but CO-based gas measurements give $\approx$40\% (0.14 dex) higher value than other gas tracers on average.
For comparison, we also plot the local and high-$z$ field [\ion{C}{1}]\, studies in Figure~\ref{fig:gasmass}.
The data points are taken from \citetalias{Valentino2020b} and \citet{Bourne2019}.
\citetalias{Valentino2020b} collected local galaxy and high-$z$ [\ion{C}{1}]\,(1--0) studies.
From their catalogs, we take the $z\sim1$ main-sequence measurements and local galaxies for which CO~(2--1) measurements are available.
For the $z\sim1$ galaxies, we assumed $R_{\rm 12} = 1.2$ and the same $\alpha_{\rm CO}$ and $X_{\rm CI}$ with protocluster galaxies given that they are on the main-sequence.
For local galaxies with CO~(2--1) we used the different $\alpha_{\rm CO} = 0.8$ and $\log{(X_{\rm CI})} = -4.2$ given that they are on average above the main-sequence ($\langle\log\Delta {\rm MS}\rangle$=1.3).
From \citet{Bourne2019}, we compare the dust-based measurements and [\ion{C}{1}]\,.
From the Table~2 in \citet{Bourne2019}, we used the flux values obtained from larger aperture sizes for Galaxy 1 and 2.
In general, they are well-matched with each other.
\section{Discussion} \label{sec:discussion}
\subsection{Self-consistent calibration from three tracers}\label{sec:calmass}
\begin{figure*}[t]
\centering
\includegraphics[width=0.85\textwidth, bb=0 0 1000 1000]{fig_dev_ver2.pdf
\caption{The gas mass comparison from CO~(3--2), dust and [\ion{C}{1}]\,(1--0) emissions. In the vertical axes, we plot the CO- and dust- based gas mass divided by [\ion{C}{1}]\,-based mass. In the horizontal axes, we plot stellar mass (top left), deviation from the main-sequence (top right), molecular gas-to-stellar mass ratio (bottom left) and depletion time scale (bottom right). The mean values of the galaxy properties are marked as dark blue arrows. The arrows plotted along the vertical axis of the top left panel is the average deviation of dust- based (turquiose, -0.02 dex) and CO-based (pink, +0.14 dex) from [\ion{C}{1}]\,- based including the upper limits.\label{fig:dev}}
\end{figure*}
The general key assumptions for deriving gas mass from different tracers are the following.
In the dust-based calibration of \citet{Scoville2016}, the mass-weighted dust temperature is fixed to $T_{\rm d}$ = 25 K and the dust-to-gas mass ratio is implicitly fixed.
In the CO-based calibration, we need to assume the $\alpha_{\rm CO}$ and CO gas excitation condition.
For the [\ion{C}{1}]\,-based calibration, we need to assume the $X_{\rm CI}$ abundance ratio (or [\ion{C}{1}]\,-conversion factor).
The excitation temperature may matter, especially for the higher transition [\ion{C}{1}]\,~(2--1) but has less impact for [\ion{C}{1}]\,~(1--0) (\citealt{Weiss2005}).
The value of $X_{\rm CI}$ is not constrained observationally in high-z galaxies.
In the following, we discuss the possible changes of the parameters ($T_{\rm d}$, gas-to-dust mass ratio, $\alpha_{\rm CO}\times R_{13}$ and $X_{\rm CI}$) to get self-consistent gas mass from different tracers focusing on non-detection case.
If the dust continuum is non-detection (i.e., HAE~4 and HAE~16), it may be attributed to a high dust temperature and/or intrinsically low dust mass (at fixed $T_{\rm d}$). For a case of high dust temperature, the dust SED would shift to a shorter wavelength hampering the detection at a given sensitivity.
It would imply the dust-based gas mass to an even lower value if the mass-weighted temperature is also high enforcing other gas measurements also to be lowered.
Alternatively, if the dust mass intrinsically is low, this would require higher gas-to-dust mass ratio to match other gas measurements.
The dust-to-gas mass ratio is closely connected to the metallicity (e.g., \citealt{Remy-Ruyer2014}).
Considering the mass-metallicity relation and the stellar mass of the galaxies, e.g., $\log{(M_{\rm star}/M_{\odot})} \approx 11.3$ (HAE4) and $\approx 10.6$ (HAE16), the scatter of the gas-to-dust ratio at fixed metallicity ($\sim0.37$ dex) could allow such variation.
Therefore, we can reasonably choose the 3$\sigma$ upper limit as face value for the following discussion.
Firstly, we consider whether the mild differences are depending on galaxy parameters, especially for the CO-based measurements, which show systematically higher value than the [\ion{C}{1}]\,-based gas mass.
Figure~\ref{fig:dev} compares the difference between the molecular gas masss derived from three different tracers as a function of stellar mass ($M_{\rm star}$), deviation from the main-sequence (sSFR/sSFR(MS)), gas-to-stellar mass ratio ($\mu_{\rm gas} = M_{\rm gas}/M_{\rm star}$) and gas depletion time scale ($\tau_{\rm depl.} = M_{\rm gas}/{\rm SFR}$) where the latter two are measured using the [\ion{C}{1}]\,-based gas measurements.
We do not find a statistically meaningful correlation with galactic parameters and the gas mass discrepancies.
For example, we cannot reject the null hypothesis that depletion time scale and the gas mass ratio of CO-based and [\ion{C}{1}]\,-based measurements, which is a proxy for the overestimated amount of $\alpha_{\rm CO}\times R_{13}$ at fixed $X_{\rm CI}$, are not correlated (i.e., Spearman's Rank coefficient of = -0.37 with a $p$-value of 0.47, but note that if HAE~4b is excluded, the strength of the correlation becomes stronger, -0.90 with a moderate $p$-value of 0.04.)
Given the small number statistics, we conclude that there is no clear dependency on the stellar mass (for $\log{M_{\rm star}/M_{\odot}} = [10.5,11.3]$), deviation from the main-sequence (for $\log{\Delta \rm MS} = [-0.1,0.2]$), gas-to-stellar mass ratio (for $\log{\mu_{\rm gas}} = [-0.6,0.5]$) and depletion time scale (for $\log({\tau_{\rm depl.}}$/Myr) $= [2.1, 3.3]$), and the differences are coming from galaxy-to-galaxy variations creating a scatter.
Detailed investigations for the gas mass cross-calibration for individual galaxies are presented in Appendix~\ref{app:crosscalibration}. We briefly summarize here the results as follows:
(1) the CO~(4--3)-to-[C I](1--0) luminosity ratio of two galaxies (HAE~3a and HAE~16) suggests the existence of dense gas that may resemble the ISM properties of DSFGs (see also Section~\ref{sec:lumiratio});
(2) the non-detection of [C I](1--0) in HAE~10, the current constraints on mass-weighted dust temperature and $\log{(X_{\rm CI})}$ suggest that it is more likely that the assumed $\alpha_{\rm CO}\times R_{13}$ would be lower than what is assumed.
We note all of these galaxies are all on the main-sequence defined by \citet{Speagle2014} (i.e., $\log{\Delta}$MS $\leq \pm0.2$).
Overall, there is a hint that the assumed $\alpha_{\rm CO}\times R_{13}$ needs to be lowered for HAE~3, 10, and 16 (thus reducing the CO-based gas mass), though it is not conclusive for the remaining galaxies. If the protocluster galaxies share similar properties, it is more reasonable to lower the CO-based gas mass (thus reducing the $\alpha_{\rm CO}\times R_{13}$) to match with dust and [\ion{C}{1}]\, based gas mass.
We finally note that the gas mass cross calibrations do not require more than a factor of $>0.6$ dex change that would characterize the protocluster galaxies well above the main-sequence where the assumptions of $\alpha_{\rm CO}\times R_{\rm 13}$ and $X_{\rm CI}$ are different.
\subsection{Gas depletion time scale compared to field galaxies}
\begin{figure*}[t]
\centering
\includegraphics[width=0.98\textwidth, bb=0 0 1500 500]{fig_depl_mstar_ver2.pdf
\caption{Gas depletion time scale for the protocluster galaxies as a function of stellar mass normalized by field scaling relation (depletion time scale) from \citet{Scoville2017b} (left), \citet{Tacconi2018} (middle) and \citet{Liu2019b} (right). In addition to five [\ion{C}{1}]\, detections (squares), non-detections for galaxies above $\log{(M_{\rm star}/M_{\odot})}>9.5$ are also shown as up-side-down triangles. Filled circles are other [\ion{C}{1}]\,-based data points for $z\sim1$ main-sequence for a reference of field galaxies from \citet{Bourne2019} and \citetalias{Valentino2020b}. \label{fig:depl}}
\end{figure*}
In the previous section, we showed that the [\ion{C}{1}]\, line is a reliable gas tracer of protocluster galaxies.
The same calibration of $\log(X_{\rm CI}) = -4.8$ applied for the field main-sequence galaxies can be used for these galaxies.
By investigating [\ion{C}{1}]\,-based gas mass, we gained one additional data point compared to 1.1 mm dust and more robust gas measurement compared to CO.
In this section, we revisit how gas scaling relation (i.e., gas depletion time scale determined by redshift, stellar mass, and deviation from the main sequence) differs from field galaxies based on the [\ion{C}{1}]\, observations.
Figure~\ref{fig:depl} shows the gas depletion time scale of the protocluster galaxies normalized by the (field) gas scaling relation determined in \citet{Scoville2017b}, \citet{Tacconi2018} and \citet{Liu2019b} at given stellar mass, redshift and deviation from the main sequence.
We plot for the [\ion{C}{1}]\, non-detections as well for those with CO~(4--3) line detections for galaxies with $\log{M_{\rm star}/M_{\odot}}>9.5$, where the stellar mass constraint is more reliable than the lower mass regime.
For non-detections, we assume the [\ion{C}{1}]\, line width the same as the CO~(4--3) width (the sixth column of Table~\ref{tab:lineflux}) to get the 3$\sigma$ upper limit of the depletion time scale.
Other [\ion{C}{1}]\,-based data points in circles are $z\sim1$ main sequence galaxies taken from \citet{Bourne2019} and \citetalias{Valentino2020b}.
Compared to high-$z$ ($z\sim1$) field main-sequence galaxies, we find a trend that the depletion time scale for main-sequence protocluster galaxies steeply decreases with stellar mass.
In comparison with field scaling relation, a stronger anti-correlation is found between depletion time scale normalized by three different scaling relations and stellar mass in protocluster galaxies.
We find negative Spearman coefficient values ($p$-value) of -0.81 (0.008, \citealt{Scoville2017b}), -0.68 (0.04, \citealt{Tacconi2018}), -0.53 (0.14, \citealt{Liu2019b}), when the 3$\sigma$ upper limits are all taken into account.
We also note that the correlation coefficient is -0.9 with p-value of 0.04 if only the secure detections are considered.
As control samples, field galaxies from \citet{Bourne2019}, and \citetalias{Valentino2020b} show less compelling evidence of such anti-correlation, i.e., Spearman coefficient ($p$-value) = -0.35 (0.06, \citealt{Scoville2017b}), -0.20 (0.30, \citealt{Tacconi2018}) and 0.23 (0.22, \citealt{Liu2019b}).
The gas scaling relation in \citet{Liu2019b} has the strongest stellar-mass dependency compared to the other two studies, which is why we find less significant signature of anti-correlation when normalized by \citet{Liu2019b} in the protocluster members and field galaxies, if any.
With the [\ion{C}{1}]\, line detections, we reconfirm the trend seen in \citetalias{minju2017a} and \citet{Tadaki2019a} that used CO/dust and CO only, respectively, though the absolute vertical scale would be slightly different due to the different flux values.
Overall, we find a signature of stronger anti-correlation between the stellar mass and the gas depletion time scale normalized by field scaling relation: the more the galaxy becomes massive, the faster it consumes or removes gas, though we caution that the number statistics and potential selection effects are remaining concerns.
What kind of physical mechanisms can derive this trend?
In \citetalias{minju2019a}, we discussed that the larger CO~(4--3) line widths for protocluster galaxies could arise from unresolved gas-rich mergers or smaller sizes.
These are not necessarily exclusive that mergers can also make galaxies smaller.
There are still other possible mechanisms to make the galaxy smaller, for example, compaction through disk-instability (e.g., \citealt{Bournaud2008, Genzel2008, Genel2012, Agertz2009, Ceverino2010, Zolotov2015, Tacchella2016, Barro2016}) and gas stripping (e.g., \citealt{Gunn1972, Haynes1984, Poggianti2017}).
Gas stripping in molecular phase remains inconclusive (e.g., \citealt{Stark1986, Kenney1989, Boselli1997}), but there is a growing body of evidence that more tightly bound molecular gas can even be stripped or disturbed (e.g., \citealt{Vollmer2008, Fumagalli2009, Boselli2014a, Zabel2019}).
We note also that the conventional picture of gas stripping is more likely to happen in low mass galaxies in massive clusters, but some group members also show such signature for neutral atomic gas (e.g., \citealt{VerdesMontenegro2001, Rasmussen2006, Hess2013}).
Considering these, the molecular gas stripping in massive galaxies in the group-like environments (i.e., protoclusters) may be less efficient, but it remains one of the possibilities.
At the same time, this protocluster also shows a hint that more massive galaxies tend to reside in denser regions (\citetalias{minju2017a}) that other environmentally driven mechanisms (gas strangulation; \citealt{Larson1980, Peng2015}) might play an additional role for the protocluster members, for earlier quenching.
All these mechanisms can be efficient in consuming gas or removing gas out of the galaxies, though the efficiency of these mechanisms in protoclusters is unknown observationally and
current data sets are not sufficient to reject or accept any of the possibilities.
In this regard, we may get more hints of possible mechanisms from either very massive ($M_{\rm star}/M_{\odot}>11$) or lower mass ($M_{\rm star}/M_{\odot}<10.5$) galaxies in protoclusters, which will also help to verify the observed trend.
\section{Summary and conclusion} \label{sec:summary}
We revisited the gas content of the protocluster by [\ion{C}{1}]\, line observations using ALMA.
We aimed to test if the same calibrations in measuring gas content using CO and dust can be applied to protocluster galaxies using an independent gas tracer of [\ion{C}{1}]\,.
Five galaxies are detected in [\ion{C}{1}]\, out of sixteen galaxies targeted.
We first compared the CO-to-[C I] luminosity relation with other studies.
It revealed the [\ion{C}{1}]\, is a good gas tracer for the protocluster galaxies.
The CO~(4--3)-to-[\ion{C}{1}]\, line relation further hinted at the presence of denser gas in protocluster galaxies compared to field galaxies though the dynamic range was still too narrow to robustly conclude on the existence of different relation in different environments.
We then compared the gas mass based on [\ion{C}{1}]\,, CO(3--2), and dust when at least one of the last two tracers was available.
Given that these five galaxies are all on the main-sequence, we adopted the general calibration methods applied to (field) main-sequence galaxies by assuming $\log{(X_{\rm CI})} = -4.8$ for the [\ion{C}{1}]\,-based gas measurement.
The [\ion{C}{1}]\,-based gas measurements are, in general, bridging the gap between CO~(3--2)-based and dust-based gas estimates.
They are lower or comparable to the CO-based gas measurements by -0.35 dex at the lowest.
The difference between [\ion{C}{1}]\,- and the dust- based measurements are mild by up to 0.16 with the mean difference of 0.02 dex.
Considering the CO~(4--3)-to-[\ion{C}{1}]\, line ratios and the parameter space allowed for typical main-sequence galaxies,
there are three galaxies that CO-based gas mass may be overestimated than the other tracers, suggesting different calibrations (or conversion factors).
Hence, the combination of $\alpha_{\rm CO}\times R_{13}$ needs to be lowered for these protocluster galaxies by up to $\approx$ 0.35 (0.43) dex to match the [\ion{C}{1}]\,-(dust-)based gas measurements.
Such adjustment is still within the uncertainty of the individual calibration methods, however.
We generally reconfirm our previous findings that the mean gas fraction is comparable to field galaxies for a stellar mass range of $\log{M_{\rm star}/M_{\odot}} = [10.6, 11.3]$ ($f_{\rm gas} = M_{\rm gas}/(M_{\rm gas}+M_{\rm star}) \approx 50\%$)).
At least for these five [\ion{C}{1}]\,-detected galaxies, the depletion time scale decreases more rapidly with stellar mass than field galaxies that might lead galaxies to quench earlier than field galaxies.
More detailed studies on more massive galaxies ($\log{M_{\rm star}/M_{\odot}}>11$) and less massive ones ($\log{M_{\rm star}/M_{\odot}}>10.5$) would help to disentangle potential physical mechanisms that are dominant than field galaxies making the observed trend.
Our findings support the idea that optically-thin gas tracer of [\ion{C}{1}]\, is indeed a good tracer for probing gas content even for protocluster galaxies.
Based on the similarity between dust-based and [\ion{C}{1}]\,-based gas measurement, dust observations would be a more efficient tool for a larger number of galaxies to improve statistics, especially for the galaxies on the main-sequence.
However, [\ion{C}{1}]\, can still be an efficient tracer for high-$z$ where getting CO~(1--0) is more expensive, and for such, the line can be used as an alternative anchor for dust-based calibration.
Finally, we caution against the generalization of our findings.
With limited number of galaxies detected in [\ion{C}{1}]\, (also CO and dust), it is difficult to verify whether there is any selection bias delivering such mass dependency.
Such bias could originate not only from the flux-limited galaxy selection (H$\alpha$) but also from the halo property (mass, evolutionary stage) of the protocluster.
More galaxies in the probed mass range ($\log{M_{\rm star}/M_{\odot}} = [10.5, 11]$) in differently selected protoclusters would also help to broaden our understanding of galaxy evolution in different environments.
\acknowledgments
We are immensely grateful to the anonymous referee for constructive comments. We thank Dr. Daizhong Liu for fruitful discussions and Dr. Frencesco Valentino for providing the observed data set before it became public. This paper makes use of the following ALMA data: ADS/JAO.ALMA \#2015.1.00152.S. ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan), together with NRC (Canada) and NSC and ASIAA (Taiwan) and KASI (Republic of Korea), in cooperation with the Republic of Chile.
Y.T is supported by NAOJ ALMA Scientific Research Grant No. 2018-09B.
\facilities{ALMA}
\software{astropy \citep{astropy}, Scikit-learn \citep{scikit-learn}, CASA \citep{McMullin2007}
}
|
\section{Appendix}
\begin{appendix}
\input{appendix/formulaedef}
\input{appendix/localnormalizeddeterminant}
\input{appendix/proofsjointdensityuscores}
\input{appendix/derivation}
\input{appendix/proofpropositionedge}
\input{appendix/proofpropositionallclose}
\input{appendix/proofgeoconditionaltounconditional}
\input{appendix/proofofLemmadiscussion}
\input{appendix/proofgeo3}
\input{appendix/auxiliarylemmas}
\input{appendix/simulations}
\end{appendix}
\begin{comment}
\subsection{Proof of Theorem \ref{cor:Poissonlimit}}
\label{sec:proofofcor:Poissonlimit}
\begin{proof}
The assumptions in this theorem implies all the assumptions in Theorem \ref{thm:Poissonultrahigh}, and hence by Theorem \ref{thm:Poissonultrahigh}
$$
d_{\operatorname{TV}}\left(\mathscr{L}\left(\bar{N}_\delta\right), \operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho}) \right) \to 0, \quad \text{ as } p \to \infty.
$$
By \eqref{eqn:jumpsizedislimit}, as $p\to \infty$ and $p^{1+\frac{1}{\delta}}(1-\rho)^{\frac{n-2}{2}}\to e_0$
$$
d_{\operatorname{TV}}\left(\operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho}), \operatorname{CP}(\lambda_{n,\delta}(e_0),\bm{\zeta}_{n,\delta})\right) \to 0,
$$
since for discrete random variable convergence in distribution is equivalent to convergence of their distribution in total variation.
Combining the last two equations immediately proves the theorem.
\end{proof}
\end{comment}
\begin{comment}
\section{recycle}
\begin{rem}
The condition $2p^{1+\frac{1}{\delta}}P_n(r_\rho)\leq \gamma$ specifies an implicit lower bound on the threshold $\rho$. To obtain an explicit lower bound, observe $2p^{1+\frac{1}{\delta}} \left(\sqrt{2(1-\rho)}\right)^{n-2} \leq \gamma $ is a sufficient condition of $2p^{1+\frac{1}{\delta}}P_n(r_\rho)\leq \gamma$, by Lemma \ref{pderivative} \ref{item:pderivativea}. Solving for $\rho$, then an explicit lower bound of $\rho$ sufficient for $2p^{1+\frac{1}{\delta}}P_n(r_\rho)\leq \gamma$ is
$$
\rho \geq 1 - \frac{1}{2} \left(\gamma 2^{-1}p^{-1-\frac{1}{\delta}}\right)^{\frac{2}{n-2}}.
$$
Even though the above result holds for any symmetric positive definite matrix $\bm{\Sigma}$, for the results to be effective, the upper bounds in the above results should be small. All the upper bounds contain the term $\mu_p \frac{\kappa}{p}$. Hence for the results to be useful, $\bm{\Sigma}$ has to have relatively concentrated eigenvalues and be row-$\kappa$ sparse with relative small sparsity level $\kappa/p$, such that $\mu_p \kappa/p$ is small. In the special case when $\bm{\Sigma}$ is diagonal, $\mu_p \kappa/p =1/p$ and $\mathbb{E} N_{\EE_\delta}^{(\bm{R})} = \xi_{p,n,\delta,\rho}$ since $\mu_p=\kappa=1$.
Moreover, suppose $\bm{\Sigma}$ is well-structured such that $\mu_p \kappa/p$ is small, say $\mu_p \kappa/p<1$, then in the upper bound in \ref{lem:mainprecorb} the term $\mu_p\left(\kappa/p\right)^2$ inside the parenthesis can be dropped since $\mu_p\left(\kappa/p\right)^2<1$. In other words, for both \ref{item:meanapprowspa} and \ref{lem:mainprecorb} the effective upper bound ,neglecting the coefficients depending on $\delta$ and $\gamma$, is $\mu_p\kappa/p$, where $\exp\left(\delta\gamma p^{-\frac{1}{\delta}} \right) \left( 1+2^{n-2}\gamma p^{-\frac{1}{\delta}} \right)$ is treated as a constant since it is between $1$ and the value by plugging 4 into $p$. \hfill $\Box$
\end{rem}
The proof of Proposition \ref{lem:mainprecor} is deferred to Section \ref{sec:proofofedgecor}. Proposition \ref{lem:mainprecor} states for given $n$, $p$, $\delta$ and $\gamma$, if the threshold $\rho$ is properly chosen, and $\bm{\Sigma}$ is row-$\kappa$ sparse and well-structured, then the distribution of $N_{E_\delta}^{(\bm{R})}/\psi(\delta)$ is approximately Poisson and its mean is close to that of the Poisson distribution. In the next subsection, we will built connections among $\{N_i^{(\bm{\Psi})}: \bm{\Psi} \in \{\bm{R},\bm{P}\}, i\in \{E_\delta, \breve{V}_\delta, V_\delta \} \}$ such that Proposition \ref{lem:mainprecor} can be extended to other $5$ quantities.
\end{comment}
\begin{comment}
\newpage
Most relevant typographical errors (just consider the correlation screening case): \\
In Large Correlation Screening Paper: \\
1) Assume each row is i.i.d. elliptically contoured is not enough. So that's why in this current paper a strong assumption that the whole data matrix is vector elliptically contoured (i.e. \eqref{eqn:ellden} in this current paper) is assumed. The following parts from your previous paper might be questionable:
(Page 8)"It is well known that when the rows of the data
matrix $X$ are i.i.d. and follow a diagonal elliptical distribution the $U$-scores are uniformly distributed on $S^{n-2}$, see for example [Sec. 2.7][1]"
(Page 14 last paragraph, and the following paragraphs on the next page) "In terms of the limiting value (3.5) of $E[N]$ the case where the columns of $X$ have
spherically contoured distribution is of special interest. In this case the $U$-scores are i.i.d.
uniformly distributed and $J =1$".
2) The proof of the edge screening on Page S-9 to S-10. Specifically, the equation on the second line on Page S-10 doesn't follow. Either the definition for the neighborhood or the definition of $A_k(i,j)$ need to adjusted.\\
In HUB DISCOVERY IN PARTIAL CORRELATION GRAPHICAL MODELS,
1) Same as the 1st point above
2) (Page 21) Same problem as 2nd point above. In the paragraph above inequality to bound $b_3$. "Next we bound the term b3 in (46). The set $A_k(\vec{i})=B_{\vec{i}}^c-\{\vec{i}\}$ indexes $\cdots$ " doesn't follow since the neighborhood defined in the paper doesn't guarantee $A_k(\vec{i})=B_{\vec{i}}^c-\{\vec{i}\}$ true for $A_k(\vec{i})$ defined in (25) in Hub screening paper.
3) (Page 21) the inequality for $b_2$ neglect some lower order and dominating terms. For those lower order terms, see Lemma \ref{item:eleine} \ref{item:eleineb} in this current paper.
4) (Page 19) Below equation (41) where "The line (41) follows from the identity $\binom{p-1-\delta}{l}\binom{p-1}{\delta}=\binom{p-1}{l+\delta}(\delta!)^{-1}$ and a change of index $\cdots$", $\binom{p-1-\delta}{l}\binom{p-1}{\delta}=\binom{p-1}{l+\delta}(\delta!)^{-1}$ is not correct.\\
Instead of proving the general case with some complicate quantities (i.e. $J$ and $\|\delta\|$) in the upper bound and then specify the results in the sparse case to simplify those complicate quantities, below (page 2-17) a new proof is presented to overcome the above-mentioned problems and to directly use the sparsity structure in the proof. The main result of the new proof for your previous result is collected in Lemma \ref{lem:mainprecor}. Lemma \ref{lem:mainprecor} is better that your previous result in the sense that for the error bound for the mean in \ref{item:meanapprowspa}, there is no $\sqrt{1-\rho}$. Moreover, the only assumption in this new proof is $\mu_p$ is finite, while in your previous assumption you have some additional assumptions on the derivative on conditional densities.
This version of new proof of your previous result (page 2-17) is almost the same as the previous version. The only difference is: in Lemma \ref{lem:totvardel1} and Lemma \ref{lem:tolvaredgedel2}, the bound of $b_1$ is improved to reduce the effect of $\mu_p^2$, while the upper bound for $b_2$ remains the same. So if you read the previous version, you may just read the new proof to bound for $b_1$. Also Lemma \ref{lem:mainprecor} is a new result added in this version.
Lemma \ref{lem:mainprecor} will be used or cited in this papers. So we need to consider how to address it: should we show (Page 2-17) in this paper and then criticize the previous papers? Or is there another better way?
\end{comment}
\subsection{Auxiliary lemmas}
\label{sec:auxlem}
\begin{lem}
\label{pderivative}
Let $P_n(r)$ be defined as in Section \ref{sec:closenessnumedge}. Suppose $n\geq 4$.
\begin{enumerate}[label=(\alph*)]
\item \label{item:pderivativea}
Recall $a_n =\frac{b_n}{2(n-2)}=\frac{\Gamma((n-1)/2)}{(n-2)\sqrt{\pi}\Gamma((n-2)/2)} \leq 1$. Then
$$
a_nr^{n-2}\left(1-\frac{\min\{r^2,4\}}{4}\right)^{\frac{n-4}{2}}\leq P_n(r) \leq a_nr^{n-2}
$$
and
\begin{equation}
-\frac{n-4}{8} a_nr^{n-2} \min\{r^2,4\}\leq P_n(r) - a_nr^{n-2}\leq 0. \label{eqn:Pcnrdif}
\end{equation}
\item \label{item:pderivativeb}
$\lim_{r\to 0} P_n(r) / \left(a_nr^{n-2} \right) =1$.
\item \label{item:pderivativec}
Let $0\leq \beta < 1< \alpha$ and $0<r\leq 2$. Then
$$
P_n(\alpha r)-P_n(\beta r)\leq (n-2) P_n(r) \alpha^{n-3} (\alpha - \beta).
$$
\item \label{item:pderivatived}
Consider $\alpha>1$ and $r>0$. Then
$$
P_n(\alpha r) \leq \alpha^{n-2}P_n(r).
$$
\item \label{item:pderivativee}
Consider $0<\beta<1$ and $0<r<2$. Then
$$
P_n(\beta r) \leq \beta^{n-2}\left( \frac{ 1-\frac{\beta^2 r^2}{4}} { 1-\frac{r^2}{4}}\right )^{\frac{n-4}{2}}P_n(r).
$$
\end{enumerate}
\end{lem}
\begin{proof}
(a)
It is easy to verify
\begin{equation}
P_n'(x)= \begin{cases} \frac{b_n}{2}x^{n-3}(1-\frac{x^2}{4})^{\frac{n-4}{2}} & x<{2}, \\
0 & x\geq 2. \end{cases} \label{eqn:pnrderivative}
\end{equation}
Consider $r>0$. Then
\begin{align}
\frac{P_n(r)}{a_nr^{n-2}} = \frac{P'_n(\xi)}{(n-2)a_n\xi^{n-3}}
= \left(1-\frac{\xi^2}{4}\right)^{\frac{n-4}{2}} \label{eqn:pderivativea1}
\end{align}
where in the first equality $\xi \in (0,\min\{r,2\}) $ due to the Cauchy Mean Value Theorem and $P_n(r)\not = 0$, and the second equality follows from \eqref{eqn:pnrderivative}. Equation \eqref{eqn:pderivativea1} directly implies
$$
\left(1-\frac{(\min\{r,2\})^2}{4}\right)^{\frac{n-4}{2}} \leq \frac{P_n(r)}{a_nr^{n-2}} \leq 1.
$$
\begin{comment}
Moreover by \eqref{eqn:pderivativea1},
\begin{align*}
\frac{P_n(r)}{a_nr^{n-2}}-1= \left(1-\frac{\xi^2}{4}\right)^{\frac{n-4}{2}} -1 \geq -\frac{n-4}{4} \xi^2 \geq -\frac{n-4}{4} \min\{r^2,4\} \numberthis \label{eqn:pderivativea2}.
\end{align*}
where the first inequality follows from mean value theorem when $n\geq 6$ and direct verification when $n=4,5$.
Then upper and lower bounds of \eqref{eqn:Pcnrdif} follows respectively from \eqref{eqn:pderivativea1} and \eqref{eqn:pderivativea2}.
\end{comment}
\noindent (b)
It follows directly by taking limit $r\to 0^+$ in \eqref{eqn:pderivativea1}.\\
\noindent (c)
Since $0\leq \beta < 1< \alpha$ and $0<r\leq 2$, $P_n(\alpha r)- P_n(\beta r)>0$ and $P_n(r)>0$. Then
\begin{align*}
\frac{P_n(\alpha r)- P_n(\beta r) }{ P_n(r) } &= \frac{\left(P_n(\alpha r)- P_n(\beta r)\right)-\left(P_n(\alpha\cdot 0)- P_n(\beta \cdot 0)\right) }{ P_n(r) - P_n(0)}\\
& = \frac{ \left. \frac{d}{dr}\left(P_n(\alpha r)- P_n(\beta r)\right) \right|_{r=\xi} }{\left. \frac{d}{dr}P_n( r) \right|_{r=\xi}} \\
&= \frac{\alpha^{n-2}\left( 1-\frac{\alpha^2\xi^2}{4}\right )^{\frac{n-4}{2}}-\beta^{n-2}\left( 1-\frac{\beta^2\xi^2}{4}\right )^{\frac{n-4}{2}}}{\left( 1-\frac{\xi^2}{4}\right )^{\frac{n-4}{2}}}\\
& \leq \alpha^{n-2}- \beta^{n-2} \numberthis \label{eqn:Pnderivativetemp} \\
& \leq (n-2)\alpha^{n-3}(\alpha-\beta),
\end{align*}
where the second equality follows from the Cauchy Mean Value Theorem with $\xi\in (0,r)$, the third equality follows from \eqref{eqn:pnrderivative} together with the fact that the numerator has to be positive, which imply $\alpha\xi<2$, the first inequality follows from $0\leq\beta<1<\alpha$, and the last inequality follows from mean value theorem.\\
\noindent (d)
When $r\geq 2$,
$
P_n(\alpha r) = P_n(r) =1
$
and the conclusion holds trivially. The case $0<r<2$ follows from \eqref{eqn:Pnderivativetemp} with $\beta=0$.\\
\noindent
(e)
Consider $0<\beta<1$ and $0<r<2$. Then
$$
\frac{P_n(\beta r)}{P_n(r)} = \frac{\beta^{n-2}\left( 1-\frac{\beta^2\xi^2}{4}\right )^{\frac{n-4}{2}}}{\left( 1-\frac{\xi^2}{4}\right )^{\frac{n-4}{2}}}\leq \frac{\beta^{n-2}\left( 1-\frac{\beta^2 r^2}{4}\right )^{\frac{n-4}{2}}}{\left( 1-\frac{r^2}{4}\right )^{\frac{n-4}{2}}},
$$
where the equality follows from Cauchy Mean Value Theorem with $\xi\in (0,r)$.
\end{proof}
\begin{lem} \label{prop:was1dis}
Consider $Z_1$ and $Z_2$ be two discrete random variable support on $[\delta]$. Then
$$
d_{\text{W}}\left(\mathscr{L}(Z_1), \mathscr{L}(Z_2)\right) \leq \frac{\delta-1}{2} \sum_{\ell=1}^{\delta}\left|P(Z_1=\ell)-P(Z_2=\ell)\right|.
$$
\end{lem}
\begin{proof}
By Remark 2.19 (iii) of Section 2.2 in \cite{villani2003topics},
\begin{align*}
d_{\text{W}}\left(\mathscr{L}(Z_1), \mathscr{L}(Z_2)\right)
= &
\sum_{i=1}^{\delta-1}|P(Z_1\leq i)-P(Z_2\leq i)|
\leq
\sum_{i=1}^{\delta-1} \sum_{j=1}^{i} |P(Z_1 = j)-P(Z_2= j)|.
\end{align*}
On the other hand, from the above equality,
\begin{align*}
d_{\text{W}}\left(\mathscr{L}(Z_1), \mathscr{L}(Z_2)\right)
= &
\sum_{i=1}^{\delta-1}|P(Z_1\geq i+1)-P(Z_2\geq i+1)| \\
\leq &
\sum_{i=1}^{\delta-1} \sum_{j=i+1}^{\delta} |P(Z_1 = j)-P(Z_2= j)|.
\end{align*}
Averaging the above two inequalities yields the desired conclusion.
\end{proof}
\begin{lem}
\label{prop:tvtomean}
For any integer-valued random variable $Z_1$ and $Z_2$,
$$
d_{\operatorname{TV}}\left(\mathscr{L}(Z_1), \mathscr{L}(Z_2)\right) \leq P( Z_1 \not = Z_2) \leq \mathbb{E} \left |Z_1 - Z_2 \right |.
$$
\end{lem}
\begin{proof}
\begin{align*}
d_{\operatorname{TV}}\left(\mathscr{L}(Z_1), \mathscr{L}(Z_2)\right) & = \max_{A \text{ Borel measurable }} \left| P(Z_1 \in A) - P(Z_2 \in A) \right| \\
& = \max_{A \text{ Borel measurable }} \left| P(Z_1 \in A, Z_1 \not = Z_2) - P(Z_2 \in A, Z_1\not = Z_2) \right| \\
& \leq P( Z_1 \not = Z_2) \\
& = P( |Z_1 - Z_2| \geq 1)\\
& \leq \mathbb{E} \left |Z_1 - Z_2 \right |.
\end{align*}
\end{proof}
\begin{lem} \label{eqn:cptv}
Consider two compound Poisson distributions $\operatorname{CP}(\lambda_1,\bm{\zeta}_1)$ and $\operatorname{CP}(\lambda_2,\bm{\zeta}_2)$. Then
\begin{align*}
& d_{\operatorname{TV}}\left(\operatorname{CP}(\lambda_1,\bm{\zeta}_1),\operatorname{CP}(\lambda_2,\bm{\zeta}_2)\right) \\
\leq &
\min\{\lambda_1,\lambda_2\} d_{\operatorname{TV}}(\bm{\zeta}_1,\bm{\zeta}_2) + d_{\operatorname{TV}}(\operatorname{Pois}(\lambda_1),\operatorname{Pois}(\lambda_2))\\
\leq & \min\{\lambda_1,\lambda_2\} d_{\operatorname{TV}}(\bm{\zeta}_1,\bm{\zeta}_2) + \min\left\{|\lambda_1-\lambda_2|,\sqrt{\frac{2}{e}}|\sqrt{\lambda_1}-\sqrt{\lambda_2}|\right\}.
\end{align*}
\end{lem}
\begin{proof}
By triangular inequality,
\begin{equation}
d_{\operatorname{TV}}\left(\operatorname{CP}(\lambda_1,\bm{\zeta}_1),\operatorname{CP}(\lambda_2,\bm{\zeta}_2)\right)\leq d_{\operatorname{TV}}\left(\operatorname{CP}(\lambda_1,\bm{\zeta}_1),\operatorname{CP}(\lambda_1,\bm{\zeta}_2)\right)+d_{\operatorname{TV}}\left(\operatorname{CP}(\lambda_1,\bm{\zeta}_2),\operatorname{CP}(\lambda_2,\bm{\zeta}_2)\right). \label{eqn:totalvariation}
\end{equation}
We will bound the two terms in the upper bound separately.
\noindent
\textbf{Step 1}: $d_{\operatorname{TV}}\left(\operatorname{CP}(\lambda_1,\bm{\zeta}_1),\operatorname{CP}(\lambda_1,\bm{\zeta}_2)\right)$\\
Consider $Z_1=\sum_{i=1}^N Y_i$ and $Z_2=\sum_{i=1}^N Y'_i$, where
$N\sim \operatorname{Pois}(\lambda_1)$,
$\{Y_i\}\overset{\text{i.i.d.}}{\sim} \bm{\zeta}_1$,
$\{Y'_i\}\overset{\text{i.i.d.}}{\sim} \bm{\zeta}_2$, and $N$ is independent of $\{Y_i\},\{Y'_i\}$. Then
\begin{align*}
d_{\operatorname{TV}}\left(\operatorname{CP}(\lambda_1,\bm{\zeta}_1),\operatorname{CP}(\lambda_1,\bm{\zeta}_2)\right) &= d_{\operatorname{TV}}\left(\mathscr{L}(Z_1), \mathscr{L}(Z_2) \right) \\
& \leq 1-P(Z_1 = Z_2) \\
& = 1- \mathbb{E} P(Z_1 = Z_2|N) \\
& \leq 1- \mathbb{E} \left(P(Y_1 = Y'_1)\right)^N,
\end{align*}
where the first inequality follows from Lemma \ref{prop:tvtomean}. Since the above inequality holds for any coupling $(Y_1,Y'_1)$ of $\zeta_1$ and $\zeta_2$, by Proposition 4.7 in \cite{levin2017markov} taking the infimum of all the coupling then yields
$$
d_{\operatorname{TV}}\left(\operatorname{CP}(\lambda_1,\bm{\zeta}_1), \operatorname{CP}(\lambda_1,\bm{\zeta}_2)\right)\leq 1- \mathbb{E} (1-d_{\operatorname{TV}}(\bm{\zeta}_1,\bm{\zeta}_2))^N \leq \lambda_1 d_{\operatorname{TV}}(\bm{\zeta}_1,\bm{\zeta}_2).
$$
\noindent
\textbf{Step 2}: $d_{\operatorname{TV}}\left(\operatorname{CP}(\lambda_1,\bm{\zeta}_2),\operatorname{CP}(\lambda_2,\bm{\zeta}_2)\right)$\\
Consider $Z_1=\sum_{i=1}^{N_1} Y'_i$ and $Z_2=\sum_{i=1}^N Y'_i$, where
$N_1\sim \operatorname{Pois}(\lambda_1)$, $N_2\sim \operatorname{Pois}(\lambda_2)$,
$\{Y'_i\}\overset{\text{i.i.d.}}{\sim} \bm{\zeta}_2$,
and $\{Y'_i\}$ is independent of $N_1,N_2$. Then
\begin{align*}
d_{\operatorname{TV}}\left(\operatorname{CP}(\lambda_1,\bm{\zeta}_2),\operatorname{CP}(\lambda_2,\bm{\zeta}_2)\right)
&= d_{\operatorname{TV}}\left(\mathscr{L}(Z_1), \mathscr{L}(Z_2) \right) \\
& \leq P(Z_1 \neq Z_2) \\
& = P(N_1 \neq N_2),
\end{align*}
where the first inequality follows from Lemma \ref{prop:tvtomean}. Since the above inequality holds for any coupling $(N_1,N_2)$ of $\operatorname{Pois}(\lambda_1)$ and $\operatorname{Pois}(\lambda_2)$, by Proposition 4.7 in \cite{levin2017markov} taking the infimum of all the coupling then yields $d_{\operatorname{TV}}\left(\operatorname{CP}(\lambda_1,\bm{\zeta}_2),\operatorname{CP}(\lambda_2,\bm{\zeta}_2)\right) \leq d_{\operatorname{TV}}\left(\operatorname{Pois}(\lambda_1), \operatorname{Pois}(\lambda_2) \right) $.
Plugging step 1 and step 2 into \eqref{eqn:totalvariation},
$$
d_{\operatorname{TV}}\left(\operatorname{CP}(\lambda_1,\bm{\zeta}_1),\operatorname{CP}(\lambda_2,\bm{\zeta}_2)\right) \leq \lambda_1 d_{\operatorname{TV}}(\bm{\zeta}_1,\bm{\zeta}_2) + d_{\operatorname{TV}}\left(\operatorname{Pois}(\lambda_1), \operatorname{Pois}(\lambda_2) \right).
$$
By symmetry property, the first term in the above upper bound can be replace by $\min\{\lambda_1,\lambda_2\} d_{\operatorname{TV}}(\bm{\zeta}_1,\bm{\zeta}_2)$. The proof is then completed by applying equation (2.2) of \cite{adell2006exact} to bound $d_{\operatorname{TV}}\left(\operatorname{Pois}(\lambda_1), \operatorname{Pois}(\lambda_2) \right)$.
\end{proof}
\begin{lem} \label{basine}
\begin{enumerate}[label=(\alph*)]
\item \label{item:basineh}
Let $p,p',m$ be positive integers such that $p\geq p'$. Then $$\prod\limits_{i=0}^m (p-i) - \prod\limits_{i=0}^m (p'-i) \leq (m+1)\left(\prod\limits_{i=0}^{m-1} (p-i)\right)(p-p').$$
\item \label{item:basinei}
Let $p,\delta,\kappa$ be positive integers such that $ \delta\leq p-1$. Then
$$
\prod\limits_{\ell=1}^\delta (p-\ell) -\prod\limits_{\ell=1}^\delta (p-\ell\kappa) \leq \frac{\delta(\delta+1)}{2} \left(\kappa-1\right) \prod_{\ell=1}^{\delta-1}(p-\ell).
$$
\item \label{item:basinej}
$\left( \frac{1+x}{1-x}\right)^2$ is increasing function on $[0,\frac{1}{2}]$ and $\left( \frac{1+x}{1-x}\right)^2\leq 1+16x$ for $0\leq x\leq \frac{1}{2}$. Then
\item \label{item:basinek}
$1+x-\frac{1}{1+x}\leq 2x$ for any $x\geq 0$.
\end{enumerate}
\end{lem}
\begin{proof}
\ref{item:basinej} and \ref{item:basinek} are simple quadratic inequalities and hence their proofs are omitted. \\
\noindent
(a)
Let $f(x)=\prod\limits_{i=0}^m (x-i)$. When $p'\geq m$, $f'(x)\leq (m+1) \prod_{i=0}^{m-1}(p-i) $ and the conclusion then follows by the mean value theorem. When $p'\leq m-1$,
$$
f(p)-f(p')\leq f(p) \leq (p-p')\prod_{i=0}^{m-1}(p-i).
$$
\noindent
(b)
Let $f(x)=\prod\limits_{\ell=1}^\delta (p-\ell x)$. When $p<\delta\kappa$,
$$
f(1)-f(\kappa)\leq f(1)\leq (\delta\kappa - \delta) \prod_{\ell=1}^{\delta-1}(p-\ell)\leq \frac{\delta(\delta+1)}{2} \left(\kappa-1\right) \prod_{\ell=1}^{\delta-1}(p-\ell).
$$
When $p\geq \delta\kappa$, $f'(x)\geq -\frac{\delta(\delta+1)}{2}\prod_{\ell=1}^{\delta-1}(p-\ell)$ for $x\in [1,\kappa]$. Then the conclusion follows by the mean value theorem.
\end{proof}
\begin{lem}[Perturbation Theory] \label{perthe}
Consider $\bm{D}\in \mathbb{S}^n$ and $\bm{E}\in \mathbb{S}^n$, where $\mathbb{S}^n$ is the set of all real symmetric matrices of dimension $n\times n$. Let $\{\lambda_i(\cdot)\}_{i=1}^n$ be the eigenvalues of corresponding matrix such that $\lambda_1(\cdot)\geq \lambda_2(\cdot)\geq \ldots\geq \lambda_n(\cdot)$.
\begin{enumerate}[label=(\alph*)]
\item \label{item:perthea}
$$
|\lambda_i(\bm{D}+\bm{E})-\lambda_i(\bm{D})|\leq \|\bm{E}\|_2 \quad (i=1,2,\ldots,n)
$$
\item Assume $\bm{E}=\omega x\xve^\top$, where $x\in S^{n-1}$. If $\omega\geq0$, then
$$
\lambda_i(\bm{D}+\bm{E})\in [\lambda_i(\bm{D}),\lambda_{i-1}(\bm{D})], \quad ( i=2,3,\ldots,n),
$$
while if $\omega\leq0$, then
$$
\lambda_i(\bm{D}+\bm{E})\in [\lambda_{i+1}(\bm{D}),\lambda_{i}(\bm{D})], \quad ( i=1,2,\ldots,n-1).
$$
In either case, there exist nonnegative $m_1,m_2,\ldots,m_n$ such that
$$
\lambda_i(\bm{D}+\bm{E})=\lambda_i(\bm{D})+m_i\omega, \quad (i=1,2,\ldots,n)
$$
with $m_1+m_2+\cdots+m_n=1$.
\item \label{item:perthec}
Assume $\bm{E}= \displaystyle\sum_{i=1}^m\omega_i x_ix_i^\top$, where $\{x_i\}_{i=1}^m \subset S^{n-1}$ and $\omega_i\geq 0$ for all $i$. Then
$$
\lambda_n(\bm{D}+\bm{E})\geq \lambda_n(\bm{D}).
$$
\end{enumerate}
\end{lem}
\begin{proof}
(a) and (b) is Corollary 8.1.6 and Theorem 8.1.8 in \cite{golub2012matrix}. (c) follows by induction on the smallest eigenvalue using part (b) for $\omega \geq 0$.
\end{proof}
\begin{lem}
\label{isoine}
Let $x_1,x_2$ be two vectors on $S^{n-1}$, and $\bm{D}\in \mathbb{R}^{n\times n}$ be an invertible matrix. Let $\text{S}_{\text{max}}(\bm{D})$ and $\text{S}_{\text{min}}(\bm{D})$ be respectively the largest and smallest singular value of $\bm{D}$. Define $\bar{\bm{w}}_i = \bm{D} x_i$ and $\bm{w}_i=\bar{\bm{w}}_i/\|\bar{\bm{w}}_i\|_2$, $(i=1,2)$. Then,
$$ \frac{\text{S}_{\text{min}}(\bm{D})}{\text{S}_{\text{max}}(\bm{D})} \|x_1-x_2\|_2 \leq \|\bm{w}_1-\bm{w}_2\|_2 \leq \frac{\text{S}_{\text{max}}(\bm{D})}{\text{S}_{\text{min}}(\bm{D})}\|x_1-x_2\|_2$$
\end{lem}
\begin{proof}
{\bf Part I (Upper Bound)}
\begin{figure}
\center
\includegraphics[scale=0.3]{coslaw.png}
\caption{$\bm{w}_1$ and $\bm{w}_2$ are the normalized vector of $\bar{\bm{w}}_1$ and $\bar{\bm{w}}_2$ respectively. }
\label{fig:geogra}
\end{figure}
Denote $\angle(\cdot,\cdot)$ the angle between two vectors. By the Law of Cosines,
$$
\cos(\angle(\bm{w}_1,\bm{w}_2))= \frac{\|\bm{w}_1\|_2^2+\|\bm{w}_2\|_2^2-\|\bm{w}_1-\bm{w}_2\|_2^2}{2\times \|\bm{w}_1\|_2 \times \|\bm{w}_2\|_2}= \frac{2-\|\bm{w}_1-\bm{w}_2\|_2^2}{2},
$$
and
$$
\cos(\angle(\bar{\bm{w}}_1,\bar{\bm{w}}_2))= \frac{\|\bar{\bm{w}}_1\|_2^2+\|\bar{\bm{w}}_2\|_2^2-\|\bar{\bm{w}}_1-\bar{\bm{w}}_2\|_2^2}{2\times \|\bar{z}_1\|_2 \times \|\bar{\bm{w}}_2\|_2}.
$$
Observing $\angle(\bm{w}_1,\bm{w}_2) = \angle(\bar{\bm{w}}_1,\bar{\bm{w}}_2)$, the right hand sides of the above two equations are equal. Solving for $\|\bm{w}_1-\bm{w}_2\|_2$, we get
$$
\|\bm{w}_1-\bm{w}_2\|_2^2 = \frac{\|\bar{\bm{w}}_1-\bar{\bm{w}}_2\|_2^2}{\|\bar{\bm{w}}_1\|_2\|\bar{\bm{w}}_2\|_2}+\left( 2 - \frac{\|\bar{\bm{w}}_2\|_2}{\|\bar{\bm{w}}_1\|_2} - \frac{\|\bar{\bm{w}}_1\|_2}{\|\bar{\bm{w}}_2\|_2}\right) \leq \frac{\|\bar{\bm{w}}_1-\bar{\bm{w}}_2\|_2^2}{\|\bar{\bm{w}}\|_2\|\bar{\bm{w}}_2\|_2}.
$$
Therefore,
\begin{align*}
\|\bm{w}_1-\bm{w}_2\|_2 &\leq \frac{\|\bar{\bm{w}}_1-\bar{\bm{w}}_2\|_2}{\sqrt{\|\bar{\bm{w}}_1\|_2\|\bar{\bm{w}}_2\|_2}} \\
& \leq \frac{\text{S}_{\text{max}}(D)\|x_1-x_2\|_2}{\sqrt{\text{S}_{\text{min}}(\bm{D})\|x_1\|_2\text{S}_{\text{min}}(\bm{D})\|x_2\|_2}} \\
& = \frac{\text{S}_{\text{max}}(\bm{D})}{\text{S}_{\text{min}}(\bm{D})}\|x_1-x_2\|_2.
\end{align*}
{\bf Part II(Lower Bound)}\\
Define $\bar{x}_i = \bm{D}^{-1}\bm{w}_i$, $(i=1,2)$. Note for $\forall i\in \{1,2\}$, $x_i$ and $\bar{x}_i$ are parallel to each other, since $x_i=\bm{D}^{-1}\bar{\bm{w}}_i$ and $\bar{\bm{w}}_i$ is parallel to $\bm{w}_i$. Thus, we conclude $x_i=\bar{x}_i/\|\bar{x}_i\|_2$, $(i=1,2)$. Reversing the role of $x_i$ and $\bm{w}_i$ in Part I, one has
$$
\|x_1-x_2\|_2 \leq \frac{\text{S}_{\text{max}}(\bm{D}^{-1})}{\text{S}_{\text{min}}(\bm{D}^{-1})}\|\bm{w}_1-\bm{w}_2\|_2.
$$
The lower bound follows from the relation $\frac{\text{S}_{\text{max}}(\bm{D}^{-1})}{\text{S}_{\text{min}}(\bm{D}^{-1})} = \frac{\text{S}_{\text{max}}(\bm{D})}{\text{S}_{\text{min}}(\bm{D})}$.
\end{proof}
\begin{lem}
\label{setrelation}
Let $\{\mathcal{D}_i\}_{i=1}^m$, $\{\mathcal{F}_i\}_{i=1}^m$, $\{\mathcal{G}_i\}_{i=1}^m$ and $\{\mathcal{H}_i\}_{i=1}^m$ be sets satisfying
$$
\mathcal{G}_i \subset \mathcal{D}_i\subset \mathcal{H}_i,\quad \mathcal{G}_i \subset \mathcal{F}_i\subset \mathcal{H}_i, \quad (i=1,2,\ldots,m).
$$
Then
\begin{enumerate}[label=(\alph*)]
\item \label{prop:setinclusiona}
$$
\left(\bigcap_{i=1}^m \mathcal{D}_i\right) \bigtriangleup \left(\bigcap_{i=1}^m \mathcal{F}_i\right) \subset \bigcup_{i=1}^m \left(\mathcal{H}_i \backslash \mathcal{G}_i \right)\bigcap \left(\bigcap_{j=1}^m\mathcal{H}_j\right)=\bigcup_{i=1}^m \left( \left(\mathcal{H}_i \backslash \mathcal{G}_i\right) \bigcap \left(\bigcap_{\substack{j=1\\j\not =i}}^m\mathcal{H}_j\right) \right) .
$$
\item \label{item:setinclusionb}
$$
\left(\bigcup_{i=1}^m \mathcal{D}_i\right) \bigtriangleup \left(\bigcup_{i=1}^m \mathcal{F}_i\right) \subset \bigcup_{i=1}^m \left(\mathcal{H}_i \backslash \mathcal{G}_i \right).
$$
\end{enumerate}
\end{lem}
\begin{proof}
(a)
Obviously,
\begin{equation}
\bigcap_{i=1}^m \mathcal{G}_i \subset \bigcap_{i=1}^m \mathcal{D}_i \subset \bigcap_{i=1}^m \mathcal{H}_i, \quad \bigcap_{i=1}^m \mathcal{G}_i \subset \bigcap_{i=1}^m \mathcal{F}_i \subset \bigcap_{i=1}^m \mathcal{H}_i. \label{eqn:setinc1}
\end{equation}
Thus,
\begin{align*}
\left(\bigcap_{i=1}^m \mathcal{D}_i\right) \bigtriangleup \left(\bigcap_{i=1}^m \mathcal{F}_i\right) &\subset \left(\bigcap_{i=1}^m \mathcal{H}_i\right) \backslash \left(\bigcap_{i=1}^m \mathcal{G}_i \right).\\
\end{align*}
Take $\forall \omega \in \left(\bigcap_{i=1}^m \mathcal{H}_i\right) \backslash \left(\bigcap_{i=1}^m \mathcal{G}_i \right)$, we know $\omega\in \bigcap_{i=1}^m \mathcal{H}_i$ and $\omega\not\in \bigcap_{i=1}^m \mathcal{G}_i$. The later fact shows $\exists j$ (which depends on $\omega$) such that $\omega \not\in \mathcal{G}_j$. Then,
\begin{equation}
\omega\in \left(\bigcap_{i=1}^m \mathcal{H}_i\right) \backslash \mathcal{G}_j\subset \mathcal{H}_j\backslash \mathcal{G}_j\subset \bigcup_{i=1}^m \left(\mathcal{H}_i \backslash \mathcal{G}_i \right). \label{eqn:setinc2}
\end{equation}
The proof is completed by combining \eqref{eqn:setinc1} and \eqref{eqn:setinc2}.\\
\noindent (b)
\begin{align*}
\left(\bigcup_{i=1}^m \mathcal{D}_i\right) \bigtriangleup \left(\bigcup_{i=1}^m \mathcal{F}_i\right) &= \left(\bigcup_{i=1}^m \mathcal{D}_i\right)^c \bigtriangleup \left(\bigcup_{i=1}^m \mathcal{F}_i\right)^c \\
& = \left(\bigcap_{i=1}^m \mathcal{D}_i^c\right) \bigtriangleup \left(\bigcap_{i=1}^m \mathcal{F}_i^c\right) \\
& \subset \bigcup_{i=1}^m \left(\mathcal{G}_i^c \backslash \mathcal{H}_i^c \right)\\
& = \bigcup_{i=1}^m \left( \mathcal{H}_i \backslash \mathcal{G}_i \right),
\end{align*}
where the inclusion step follows from \ref{prop:setinclusiona}.
\end{proof}
\begin{lem}
\label{proeig}
Let $\bm{Q}\in \mathbb{R}^{n \times m}(n\leq m)$, with each column $\bm{q}_i$ being i.i.d. $\operatorname{{\bf unif}}(\sqrt{n}S^{n-1})$. Let $\lambda_{\text{min}}$ and $\lambda_{\text{max}}$ be respectively the largest and smallest eigenvalue of $\frac{1}{m}\bm{Q}\Qma^\top$. Then with probability at least $ 1 - 2\exp(-ct^2)$,
\begin{equation}
\left[1-C\left(\sqrt{\frac{n}{m}}+\frac{t}{\sqrt{m}}\right)\right]^2 \leq \lambda_{\text{min}} \leq \lambda_{\text{max}} \leq \left[1+C\left(\sqrt{\frac{n}{m}}+\frac{t}{\sqrt{m}}\right)\right]^2, \label{eqn:eigupplowbou}
\end{equation}
where $c,C$ are absolute constants.
\end{lem}
\begin{proof}
Let $\text{S}_{\text{max}}$, $\text{S}_{\text{min}}$ be respectively the largest and smallest singular value of $\bm{Q}$. Since columns $\{\bm{q}_i\}_{i=1}^m$ are isotropic random vectors with subgaussian norm (or $\psi_2$ norm) being a constant, by applying Theorem 5.39 in \cite{vershynin2012introduction} to $Q^\top$,
\begin{equation}
\sqrt{m}-C(\sqrt{n}+t) \leq \text{S}_{\text{min}} \leq \text{S}_{\text{max}} \leq \sqrt{m}+C(\sqrt{n}+t),
\end{equation}
holds with probability at least $ 1 - 2\exp(-ct^2)$, where $c,C$ are absolute constants. The proof is completed by
$$
\lambda_{\text{max}} = \frac{1}{m}\text{S}_{\text{max}}^2, \quad \lambda_{\text{min}}=\frac{1}{m} \text{S}_{\text{min}}^2.
$$
\end{proof}
\begin{comment}
\begin{lem} \label{lem:bonferroni}
Let $\{a_i\}_{i=1}^m \subset [0,1]$.
$$
1+\sum_{j=1}^{m_1}(-1)^j\sum_{1\leq q_1<\cdots<q_j\leq m}a_{q_1}\cdots a_{q_j}
\leq \prod_{i=1}^m(1-a_i)\leq 1+\sum_{j=1}^{m_2}(-1)^j\sum_{1\leq q_1<\cdots<q_j\leq m}a_{q_1}\cdots a_{q_j},
$$
where $m_1\in [m]$,$m_2\in [m]\cup\{0\}$ are respectively any odd and even integer. In particular,
$$
1 - \sum_{q=1}^ma_q
\leq \prod_{i=1}^m(1-a_i)\leq 1
$$
\end{lem}
\begin{proof}
It can be proved by induction or follows directly from Bonferroni Inequalities.
\end{proof}
\end{comment}
\subsection{\ONE{An informal derivation of Proposition \ref{prop:edgecor}}}
\label{sec:derivationprop:edgecor}
In this subsection we provide motivations to the following two questions: 1) why is the distribution of the star subgraph counts $N_{E_\delta}^{(\bm{R})}$ approximately compound Poisson? 2) what are the associated parameters of the compound Poisson approximation to $N_{E_\delta}^{(\bm{R})}$?
We first introduce some notation to answer the first question.
Denote
\begin{equation}
C_\delta^{<}:=\{\vec{i}=(i_0,i_1,\cdots,i_\delta)\in [p]^{\delta+1}: i_1<i_2<\cdots<i_\delta, \text{ and } i_\ell\not= i_0,\ \forall\ \ell\in [\delta] \}. \label{eqn:Cdeltadef}
\end{equation}
Each $\vec{i}\in C_\delta^<$ corresponds to a group of $\delta$ vertices indexed by $\{i_j\}_{j=1}^{\delta}$ and a center indexed by $i_0$. For $\vec{i}\in C_\delta^<$, denote by
\begin{equation}
\Phi_{\vec{i}}^{(\bm{R})}= \prod_{j=1}^\delta \Phi_{i_0i_j}^{(\bm{R})} = 1\left(\bigcap_{j=1}^{\delta}\{\textbf{dist}(\bm{u}_{i_0},\bm{u}_{i_j})\leq r_\rho\}\right) \label{eqn:NEdeltaRdef}
\end{equation}
the indicator function of the event that vertex $i_0$ is connected to each vertex $i_j$ for $j\in [\delta]$ in the empirical correlation graph. Equivalently, when $\delta\geq 2$, $\Phi_{\vec{i}}^{(\bm{R})}$ is the indicator function of the event that there exists a star subgraph in $\mathcal{G}_{\rho}(\bm{R})$ with center at $i_0$ and with leaves $\{i_j\}_{j=1}^{\delta}$.
Then by definition
\begin{equation}
N_{\EE_\delta}^{(\bm{R})} =\sum_{\vec{i}\in C_\delta^<}\ \ \Phi_{\vec{i}}^{(\bm{R})}. \label{eqn:Nedgedef1}
\end{equation}
\index{$N_{\EE_\delta}^{(\bm{R})}$}
If $\Phi_{\vec{i}}^{(\bm{R})}$ for different $\vec{i}\in C^<_{\delta}$ are independent or weakly dependent, then the distribution of $N_{\EE_\delta}^{(\bm{R})}$, as a sum of independent or weakly dependent indicator random variables, might be expected to be approximately Poisson. This, however, is not the case due to high dependency among many terms in the summation. Specifically, for any $\vec{i}\in C_\delta^<$, $\Phi_{\vec{i}}^{(\bm{R})}$ is highly dependent on $\Phi_{\vec{j}}^{(\bm{R})}$ for any $\vec{j} \in S_{\vec{i}}$ where
\begin{align}
S_{\vec{i}} & :=\left\{\vec{j}\in C_\delta^< \backslash \{\vec{i}\}: \bigcup_{\ell=0}^\delta \{j_\ell\} = \bigcup_{\ell=0}^\delta \{i_\ell\} \right\}.
\label{eqn:Sidef}
\end{align}
$S_{\vec{i}}$ is the set of indexes sharing the same vertices with $\vec{i}$ but with center different from $i_0$ and thus $|S_{\vec{i}}| = \delta$.
Indeed, provided $\Phi_{\vec{i}}^{(\bm{R})}=1$, which is equivalent to $\textbf{dist}(\bm{u}_{i_0},\bm{u}_{i_j})\leq r_\rho$ for $\forall j\in [\delta]$,
$$
\textbf{dist}(\bm{u}_{i_1},\bm{u}_{i_j})\leq \textbf{dist}(\bm{u}_{i_1},\bm{u}_{i_0})+\textbf{dist}(\bm{u}_{i_0},\bm{u}_{i_j})\leq 2r_\rho, \text{ for } \forall 2\leq j \leq \delta.
$$
That is, $\{\bm{u}_{i_j}\}_{j=2}^{\delta+1}$ are all close to $\bm{u}_{i_1}$ and hence it is likely there are edges connecting them. In other words it is likely
for $\vec{i'}=(i_1,i_0,\cdots,i_\delta)$, $\Phi_{\vec{i'}}^{(\bm{R})}=1$, where we without loss of generality assume $i_0<i_j$ for $2\leq j \leq \delta$.
Let
\begin{equation}
U_{\vec{i}} = \sum_{\vec{j}\in S_{\vec{i}}} \Phi_{\vec{j}}^{(\bm{R})} \label{eqn:Uidef}
\end{equation}
be the sum of highly dependent terms of $\Phi_{\vec{i}}^{(\bm{R})}$. In summary, if there is an increment for $N_{\EE_\delta}^{(\bm{R})}$, say $\Phi_{\vec{i}}^{(\bm{R})}=1$, there is a certain probability that $U_{\vec{i}}$ is greater than $0$ due to the high dependence, causing each increment of $N_{\EE_\delta}^{(\bm{R})}$ to be greater than $1$ with a certain probability. This is typical behavior for a compound Poisson random variable.
To answer the second question, we next informally derive the parameters of the compound Poisson approximation to $N_{E_\delta}^{(\bm{R})}$ in the special case when $\bm{\Sigma}$ is diagonal. We require some additional definitions. Let $\left[\vec{i}\right]=\{i_0,i_1,\cdots,i_\delta\}$ be the unordered set of indexes, which is referred as \emph{index group}, of any $\vec{i}\in C_\delta^<$ and define $\left[C_\delta^{<}\right]:=\left\{\left[\vec{i}\right]: \vec{i}\in C_\delta^< \right\}$. It follows that $\left|\left[C_\delta^{<}\right]\right| =\binom{p}{\delta+1}$,
where $|\cdot|$ is the cardinality of a set. For a given group of $\delta+1$ indexes $\left[\vec{i}\right]$, $\Phi_{\vec{i}}^{(\bm{R})} + U_{\vec{i}}$ is the increment associated to this group and its value is between $0$ and $\delta+1$.
In the following informal argument we assume that the event that two different index groups both have non-zero increments has probability zero. This assumption will be verified in the proof of Proposition \ref{prop:edgecor}. Consequently the probability of increment size $\ell$ for $\ell\geq 1$ is proportional to the expectation of the fraction of the number of groups with increment $\ell$
\begin{align*}
&\mathbb{E} \frac{1}{\left|\left[C_\delta^{<}\right]\right|} \sum_{\left[\overset{}{\vec{i}}\right]\in \left[C_\delta^{<}\right] } 1\left( \Phi_{\vec{i}}^{(\bm{R})} + U_{\vec{i}} = \ell \right) \\
=& \frac{1}{\left|\left[C_\delta^{<}\right]\right|} \frac{1}{\ell} \mathbb{E} \sum_{\left[\overset{}{\vec{i}}\right]\in \left[C_\delta^{<}\right] } \left(\Phi_{\vec{i}}^{(\bm{R})} + U_{\vec{i}}\right) 1\left( \Phi_{\vec{i}}^{(\bm{R})} + U_{\vec{i}} = \ell \right) \\
=& \frac{1}{\left|\left[C_\delta^{<}\right]\right|} \frac{1}{\ell} \mathbb{E} \sum_{\vec{i}\in C_\delta^{<} } \Phi_{\vec{i}}^{(\bm{R})} 1\left( \Phi_{\vec{i}}^{(\bm{R})} + U_{\vec{i}} = \ell \right) \\
=& \frac{1}{\left|\left[C_\delta^{<}\right]\right|} \frac{1}{\ell} \sum_{\vec{i}\in C_\delta^{<} } P\left(\Phi_{\vec{i}}^{(\bm{R})}=1\right) P \left( \Phi_{\vec{i}}^{(\bm{R})} + U_{\vec{i}} = \ell | \Phi_{\vec{i}}^{(\bm{R})}=1 \right). \numberthis \label{eqn:jumprateint}
\end{align*}
Since $\bm{\Sigma}$ is assumed to be diagonal, $\{\bm{u}_i\}_{i=1}^p$ are i.i.d. $\operatorname{{\bf unif}}(S^{n-2})$ by Lemma \ref{lem:Uscoresdistribution} \ref{item:Uscoresdistributionb} and hence \eqref{eqn:jumprateint} becomes
\begin{align*}
\mathbb{E} \frac{1}{\left|\left[C_\delta^{<}\right]\right|} \sum_{\left[\vec{i}\right]\in \left[C_\delta^{<}\right] } 1\left( \Phi_{\vec{i}}^{(\bm{R})} + U_{\vec{i}} = \ell \right) = & \frac{1}{\left|\left[C_\delta^{<}\right]\right|}
\frac{1}{\ell} \left|C_\delta^<\right| (2P_n(r_\rho))^{\delta} \alpha(\ell,r_\rho) \\
= &
\frac{\delta+1}{\ell} (2P_n(r_\rho))^{\delta} \alpha(\ell,r_\rho),
\end{align*}
where $\alpha(\ell,r_\rho)$ is defined in \eqref{eqn:alphandeltarho}
and $P\left(\Phi_{\vec{i}}^{(\bm{R})}=1\right)=(2P_n(r_\rho))^{\delta}$ by conditioning on $\bm{u}_{i_0}$. As a consequence, the probability of increment size $\ell$ for $\ell\geq 1$ is:
\begin{align*}
\frac{\delta+1}{\ell} (2P_n(r_\rho))^{\delta} \alpha(\ell,r_\rho)/ \sum_{\ell =1}^{\delta+1}\left(\frac{\delta+1}{\ell} (2P_n(r_\rho))^{\delta} \alpha(\ell,r_\rho)\right)
= &\bm{\zeta}_{n,\delta,\rho}(\ell),
\end{align*}
where the equation follows from
\eqref{eqn:zetandelrhodef}. This argument implies that $\bm{\zeta}_{n,\delta,\rho}$ is the increment distribution of the compound Poisson approximation.
Since the mean of a compound Poisson distribution is the product of the arrival rate and the mean of the increment distribution, the arrival rate $\lambda$ of the compound Poisson approximation satisfies the following mean constraint:
$$
\lambda \mathbb{E} \bm{\zeta}_{n,\delta,\rho} = \mathbb{E} N_{\EE_\delta}^{(\bm{R})},
$$
where $\mathbb{E} \bm{\zeta}_{n,\delta,\rho}$ is the mean of $\bm{\zeta}_{n,\delta,\rho}$. One can easily verify $\mathbb{E} \bm{\zeta}_{n,\delta,\rho}= 1/\sum_{\ell=1}^{\delta+1} \left(\alpha(\ell,r_\rho)/\ell \right) $ and $\mathbb{E} N_{\EE_\delta}^{(\bm{R})} = \binom{p}{1}\binom{p-1}{\delta}\mathbb{E}\Phi_{\vec{i}}^{(\bm{R})} = \binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^\delta $ since $\bm{\Sigma}$ is diagonal. Hence the arrival rate for the compound Poisson is
\begin{align}
\lambda = \binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^\delta\sum_{\ell=1}^{\delta+1} \frac{\alpha(\ell,r_\rho)}{\ell}= \lambda_{p,n,\delta,\rho}.
\end{align}
This informal argument motivates the compound Poisson approximation $\operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho})$ to $N_{\EE_\delta}^{(\bm{R})}$ when $\bm{\Sigma}$ is diagonal.
\section{List of symbols in the main text}
\subsection{List of symbols in the main text}
\label{sec:symbols}
\begin{itemize}
\setlength\itemsep{0mm}
\item
$\|\cdot\|_2$ : Euclidean norm.
\item
$|\cdot|$ : the cardinality of a set.
\item
$C$ and $c$ : denote generic positive universal constants that often differ from line to line.
$C$ and $c$ with subscripts are positive constants depending on the parameter in their subscripts and may differ from line to line.
\item
$X$: $n\times p$ matrix of observations. It has density $f_{X}(X)$ given in \eqref{eqn:ellden}.
\item
$\bm{R}$ : $p\times p$ sample correlation matrix. Also denoted $\bm{\Psi}^{(\mathbf R)}$.
\item
$\bm{P}$ : $p\times p$ sample partial correlation matrix. Also denoted $\bm{\Psi}^{(\mathbf P)}$.
\item
$\rho \in [0, 1]$ : screening threshold applied to elements of matrices $\bm{R}$ or $\bm{P}$.
\item
$\bm{\Phi}^{(\bm{\Psi})}$ : the binary indicator matrix of the non-zero entries of the thresholded sample correlation or sample partial correlation matrix $\bm{\Psi}$ with entries $\Phi_{ij}^{(\bm{\Psi})}= \Phi_{ij}^{(\bm{\Psi})}(\rho)=1(|\Psi_{ij}|\geq \rho)$, where $\bm{\Psi} \in\{\bm{R},\bm{P}\}$.
\item
$\mathcal{G}_{\rho}(\bm{\Psi})$ : the graph associated with adjacency matrix $\bm{\Phi}^{(\bm{\Psi})}$. Specifically, it is the empirical correlation graph when
$\bm{\Psi} =\bm{R}$ or, equivalently, $\bm{\Psi}=\bm{R}$ and it is the empirical partial correlation graph when $\bm{\Psi} =\bm{P}$ or, equivalently, $\bm{\Psi}=\bm{P}$ (see Subsection \ref{Hubfra}).
\begin{comment}
\item
\YW{
$\mathcal{G}_{\rho}(\bm{R})=\mathcal{G}_{\rho}(\bm{\Psi})$ : empirical correlation graph for
$\bm{\Psi} =\bm{R}$ or, equivalently, $\bm{\Psi}=\bm{R}$ (see Subsection \ref{Hubfra}).
}
\item
$\mathcal{G}_{\rho}(\bm{\Psi})$ : empirical partial correlation graph for $\bm{\Psi} =\bm{P}$ or, equivalently, $\bm{\Psi}=\bm{P}$ (see Subsection \ref{Hubfra}).
\end{comment}
\item
$N_{V_\delta}^{(\bm{\Psi})}$ : number of vertices of degree at least $\delta$ in $\mathcal{G}_{\rho}(\bm{\Psi})$.
\item
$N_{\breve{V}_\delta}^{(\bm{\Psi})}$ : number of vertices of degree exactly $\delta$ in $\mathcal{G}_{\rho}(\bm{\Psi})$. The quantity $\{N_{\breve{V}_\delta}^{(\bm{\Psi})}\}_{\bm{\Psi} =0}^{n-1}$ is the empirical vertex degree distribution associated with the sample correlation graph ($\bm{\Psi} =\bm{R}$) or partial correlation graph ($\bm{\Psi} =\bm{P}$).
\item
$\Gamma_{\delta}$ : a star graph with $\delta$ edges.
\item
$N_{E_\delta}^{(\bm{\Psi})}$ : the number of subgraphs isomorphic to $\Gamma_{\delta}$ in $\mathcal{G}_{\rho}(\bm{\Psi})$ when $\delta \geq 2$. $N_{E_\delta}^{(\bm{\Psi})}$ is defined as twice the number of edges in $\mathcal{G}_{\rho}(\bm{\Psi})$ when $\delta =1$.
\item
$\bar{N}_{\delta}$ : a generic random variable denoting any of the quantities $\{N_{E_\delta}^{(\bm{\Psi})},N_{\breve{V}_\delta}^{(\bm{\Psi})},N_{V_\delta}^{(\bm{\Psi})}: \bm{\Psi} \in \{\bm{R},\bm{P}\} \}$.
\item
$\bm{U}=[\bm{u}_1,\ldots,\bm{u}_p]$ : $(n-1) \times p$ matrix of correlation U-scores such that $\bm{U}^\top \bm{U} = \bm{R} $.
\item
$Y=[y_1,\ldots,y_p]$ : $(n-1) \times p$ matrix of partial correlation Y-scores such that $Y^\top Y =\bm{P}$.
\item
$r_{\rho}=\sqrt{2(1-\rho)}$ : spherical cap radius parameter in terms of $\rho$.
\item
$S^{n-2}$ : unit sphere in $\mathbb{R}^{n-1}$, {i.e. $S^{n-2}=\{\bm{w}\in \mathbb{R}^{n-1}: \|\bm{w}\|_2=1$\}. Its area is denoted by $\textbf{Area}(S^{n-2})$ or $|S^{n-2}|$.
}
\item
$B^{n-2}$ : unit ball in $\mathbb{R}^{n-2}$, {i.e. $B^{n-2}=\{\bm{w}\in \mathbb{R}^{n-2}: \|\bm{w}\|_2\leq 1$\}. Its volume is denoted by $|B^{n-2}|$.}
\item
$\operatorname{{\bf unif}}(S^{n-2})$ : uniform distribution on $S^{n-2}$.
\item
$\operatorname{{\bf unif}}(B^{n-2})$ : uniform distribution in $B^{n-2}$.
\item
$P_n(r)$ : normalised area of a spherical cap on $S^{n-2}$ with radius $r$ (see \eqref{spharefor}).
\item
$\textbf{Ge}\left(\{\bm{v}_i \}_{i=1}^\delta, r \right)$ : the geometric graph generated by a set of points $\{\bm{v}_i \}_{i=1}^\delta\subset \mathbb{R}^{n-2}$ with ball radius parameter $r$ (see Subsection \ref{Hubfra}).
\item Universal vertex : a vertex of an undirected graph that is adjacent to all other vertices
\item
$\operatorname{{\bf NUV}}\left(\{\bm{v}_i \}_{i=1}^\delta, r \right)$ :
the number of universal vertices in $\textbf{Ge}\left(\{\bm{v}_i \}_{i=1}^\delta, r \right)$.
\item
$\textbf{PGe}\left(\{\bm{v}_i \}_{i=1}^m, r ; \mathscr{N} \right)$ : the pseudo geometric graph generated by $\{\bm{v}_i \}_{i=1}^m\subset \mathbb{R}^{\mathscr{N}}$ with radius paraneter $r$ (see Definition \ref{def:pge}).
\item
$\operatorname{ {\bf PNUV} }\left(\{\bm{v}_i \}_{i=1}^m, r ; \mathscr{N} \right)$ : the number of universal vertices in $\textbf{PGe}\left(\{\bm{v}_i \}_{i=1}^m, r ; \mathscr{N} \right)$.
\item
$\{\bm{u}'_i\}_i$ : a sequence of i.i.d. random vectors drawn from $\operatorname{{\bf unif}}(S^{n-2})$, the uniform distribution on $S^{n-2}$.
\item
$\alpha(\ell,r_\rho)=\alpha_{n,\delta}(\ell,r_\rho)$: the conditional probability that there are $\ell$ universal vertices in $\textbf{PGe} \left(\{\bm{u}'_i \}_{i=1}^{\delta+1}, r_\rho \right)$ (see \eqref{eqn:alphandeltarho}).
\item
$\bm{\zeta}_{n,\delta,\rho}$ : the discrete distribution supported on $[\delta+1]$ with $\bm{\zeta}_{n,\delta,\rho}(\ell)\propto \alpha(\ell,r_\rho)/\ell$ (see \eqref{eqn:zetandelrhodef}).
\item
$\operatorname{CP}(\lambda,\bm{\zeta})$ :
the compound Poisson distribution, where the arrival rate $\lambda$ is the rate for the underlying Poisson random variable and the increment distribution $\bm{\zeta}$ is the distribution of each increment.
\item
$\operatorname{CP}(\lambda_{p, n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho})$ :
the non-asymptotic compound Poisson distribution in Theorem \ref{thm:Poissonultrahigh}.\\
\item
$\overset{\mathcal{D}}{\to}$ : denotes convergence in distribution.
\item
$\lambda_{p, n,\delta,\rho}$ : a quantity defined as $\binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^\delta\sum_{\ell=1}^{\delta+1} \frac{\alpha(\ell,r_\rho)}{\ell}$. It is the arrival rate of $\operatorname{CP}(\lambda_{p, n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho})$.
\item
$\{\tilde{\bm{u}}_i\}_i$ : a sequence of i.i.d. random vectors drawn from $\operatorname{{\bf unif}}(B^{n-2})$, the uniform distribution in $B^{n-2}$.
\item
$\alpha_\ell$ : the probability that there are exactly $\ell-1$ universal vertices in $\textbf{Ge}\left(\{\tilde{\bm{u}_i}\}_{i=1}^\delta, 1 \right)$ (see \eqref{eqn:alphaelldef}). $\alpha_\ell=\lim_{\rho\to 1}\alpha(\ell,r_\rho)$ (see Lemma \ref{prop:rangeolimit} and the paragraph beneath it).
\item
$\bm{\zeta}_{n,\delta}$ : the discrete distribution supported on $[\delta+1]$ with $\bm{\zeta}_{n,\delta}(\ell)\propto \alpha_\ell/\ell$ (see \eqref{eqn:incrementsizelimitdef}). $\bm{\zeta}_{n,\delta}(\ell)=\lim_{\rho\to 1} \bm{\zeta}_{n,\delta,\rho}(\ell)$ (see Lemma \ref{lem:compoundpoissonlimit}).
\item
$\lambda_{n,\delta}(e_{n,\delta})$ : a quantity defined as $\frac{1}{\delta !} \left( e_{n,\delta}\right)^\delta \sum_{\ell=1}^{\delta+1}\frac{\alpha_\ell}{\ell}$. $\lambda_{n,\delta}(e_{n,\delta})=\lim_{p\rightarrow \infty} \lambda_{p,n,\delta,\rho}$ when $\rho$ satisfies $a_n2^{\frac{n}{2}}p^{1+\frac{1}{\delta}}(1-\rho)^{\frac{n-2}{2}}\to e_{n,\delta}$, where $a_n=\frac{\Gamma((n-1)/2)}{(n-2)\sqrt{\pi}\Gamma((n-2)/2)}$ and $e_{n,\delta}$ is some positive constant (see Lemma \ref{lem:compoundpoissonlimit}).
\item
$\operatorname{CP}(\lambda_{n,\delta},\bm{\zeta}_{n,\delta})$ :
the limiting compound Poisson distribution in Theorem \ref{cor:Poissonlimit}. Moreover, $\operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho}) \overset{\mathcal{D}}{\to} \operatorname{CP}(\lambda_{n,\delta}(e_{n,\delta}),\bm{\zeta}_{n,\delta})$, when $p\to \infty$ and $\rho$ satisfies $a_n2^{\frac{n}{2}}p^{1+\frac{1}{\delta}}(1-\rho)^{\frac{n-2}{2}}\to e_{n,\delta}$. (see Lemma \ref{lem:compoundpoissonlimit})
\item
$\mu(\bm{A})$ : normalized determinant of square matrix $\bm{A}$ (see Definition \ref{def:nordet}).
\item
$\mu_m(\bm{A})$ : local normalized determinant of degree $m$ (see the paragraph after Definition \ref{def:nordet}).
\item
$\mu_{n,m} (\bm{A})$ : inverse local normalized determinant. It is powers of $\mu_m(\bm{A})$ (see \eqref{eqn:munmA}).
\item
$C_\delta^<$ : the index set defined as $\{\vec{i}=(i_0,i_1,\cdots,i_\delta)\in [p]^{\delta+1}: i_1<i_2<\cdots<i_\delta, \text{ and } i_\ell\not= i_0,\ \forall\ \ell\in [\delta] \}$. Each $\vec{i}\in C_\delta^<$ corresponds to a group of $\delta$ vertices indexed by $\{i_j\}_{j=1}^{\delta}$ and a center indexed by $i_0$.
\item
$\Phi_{\vec{i}}^{(\bm{R})}$ : defined as $\prod_{j=1}^\delta \Phi_{i_0i_j}^{(\bm{R})}$.
It is the indicator random variable that vertex ${i_0}$ is adjacent to each vertex ${i_j}$ for $j\in [\delta]$ in $\mathcal{G}_\rho(\bm{R})$. Equivalently, it is the indicator function of the event that there exist a star subgraph in $\mathcal{G}_{\rho}(\bm{R})$ with center at $i_0$ and leaves $\{i_j\}_{j=1}^{\delta}$.
\item
$S_{\vec{i}}$ : defined as $\left\{\vec{j}\in C_\delta^< \backslash \{\vec{i}\}: \bigcup_{\ell=0}^\delta \{j_\ell\} = \bigcup_{\ell=0}^\delta \{i_\ell\} \right\}$ for $\vec{i}\in C_\delta^<$. It is the set of indices in $C_\delta^<$ that share the same vertices with $\vec{i}$ but with center different from $i_0$.
\item
$U_{\vec{i}}$ : defined as $\sum_{\vec{j}\in S_{\vec{i}}} \Phi_{\vec{j}}^{(\bm{R})}$. It is the sum of highly dependent terms of $\Phi_{\vec{i}}^{(\bm{R})}$. (see the discussions preceding and following \eqref{eqn:Uidef}).
\end{itemize}
\begin{comment}
\begin{tabular}{c|l}
$X$ & $n\times p$ matrix of observations\\
$\bm{R}$ & $p\times p$ sample correlation\\
$\bm{P}$ & $p\times p$ sample partial correlation\\
$\rho \in [0, 1]$ & screening threshold applied to matrix $\bm{R}$ or $\bm{P}$\\
$\mathcal{G}_{\rho}(\Psi^{(\bm{\Psi})})$ & empirical correlation graph $\Psi^{(\bm{R})}=\bm{R}$ or empirical partial correlation graph $\Psi^{(\bm{P})}=\bm{R}$ (see Subsection \ref{Hubfra})\\
$N_{V_\delta}^{(\bm{\Psi})}$ & number of vertices of degree at least $\delta$ in $\mathcal{G}_{\rho}(\Psi^{(\bm{\Psi})})$\\
$N_{\breve{V}_\delta}^{(\bm{\Psi})}$ & number of vertices of degree exactly $\delta$ in $\mathcal{G}_{\rho}(\Psi^{(\bm{\Psi})})$\\
$\Gamma_{\delta}$ & a star graph with $\delta$ edges\\
$N_{E_\delta}^{(\bm{\Psi})}$ & number of subgraphs isomorphic to $\Gamma_{\delta}$ in $\mathcal{G}_{\rho}(\Psi^{(\bm{\Psi})})$ when $\delta \geq 2$ and number of edges in $\mathcal{G}_{\rho}(\Psi^{(\bm{\Psi})})$ when $\delta =1$\\
$\bm{U}=[\bm{u}_1,\ldots,\bm{u}_p]$ & $n-1 \times p$ matrix of correlation U-scores such that $\bm{U}^\top \bm{U} $ is the sample correlation $\bm{R}$ \\
$Y=[y_1,\ldots,y_p]$ & $n-1 \times p$ matrix of correlation Y-scores such that $Y^\top Y $ is the sample partial correlation $\bm{P}$ \\
$r_{\rho}=\sqrt{2(1-\rho)}$ & spherical cap radius parameter in terms of $\rho$\\
$S^{n-2}$ & unit sphere in $\mathbb{R}^{n-1}$\\
$B^{n-2}$ & unit ball in $\mathbb{R}^{n-1}$\\
$P_n(r)$ & normalised area of the spherical cap with radius $r$\\
$\textbf{Ge}\left(\{\bm{v}_i \}_{i=1}^\delta, r \right)$ & the geometric graph generated by $\{\bm{v}_i \}_{i=1}^\delta\subset \mathbb{R}^{n-2}$ with parameter/radius $r$ (see Subsection \ref{Hubfra})\\
$\operatorname{{\bf NUV}}\left(\{\bm{v}_i \}_{i=1}^\delta, r \right)$ &
the number of vertices of maximum degree in $\textbf{Ge}\left(\{\bm{v}_i \}_{i=1}^\delta, r \right)$ \\
$\textbf{PGe}\left(\{\bm{v}_i \}_{i=1}^m, r ; \mathscr{N} \right)$ & the pseudo geometric graph generated by $\{\bm{v}_i \}_{i=1}^m\subset \mathbb{R}^{\mathscr{N}}$ with parameter/radius $r$ (see Definition \ref{def:pge})\\
$\operatorname{ {\bf PNUV} }\left(\{\bm{v}_i \}_{i=1}^m, r ; \mathscr{N} \right)$ & the number of vertices of maximum degree in $\textbf{PGe}\left(\{\bm{v}_i \}_{i=1}^m, r ; \mathscr{N} \right)$\\
$\operatorname{CP}(\lambda,\bm{\zeta})$ & a compound Poisson distribution, where $\lambda$ is the rate for the underlying Poisson random variable and $\bm{\zeta}$ is the distribution of each increment\\
$\bm{\zeta}_{n,\delta,\rho}$ & the distribution of each increment for the approximating compound Poisson when $p$ is finite (see \eqref{eqn:zetandelrhodef})\\
$\bm{\zeta}_{n,\delta}$ & the distribution of each increment for the limiting compound Poisson (see \eqref{eqn:incrementsizelimitdef})\\
$\mu(\bm{A})$ & normalized determinant (see Definition \ref{def:nordet})\\
$\mu_{n,m} (\bm{A})$ & see \eqref{eqn:munmA}\\
\end{tabular}
\end{comment}
\section{Controlling local normalized determinant by extreme eigenvalues}
\subsection{Controlling local normalized determinant by extreme eigenvalues}
\label{sec:localdeterminantvseigenvalues}
\begin{lem} \label{lem:sufass}
Let $\bm{A}$ be a symmetric positive definite matrix. Consider a sequence of symmetric positive definite matrices $\bm{\Sigma}\in \mathbb{R}^{p\times p}$ with increasing dimension $p$.
\begin{enumerate}[label=(\alph*)]
\item \label{item:sufassnewa}
$\mu_{n,m}(\bm{A})$ is bounded by powers of the largest local condition number:
$$
\mu_{n,m}(\bm{A}) \leq \begin{cases} \max\limits_{{\mathcal{I}}\subset [p]}\left(\frac{\lambda_{\text{max}}(\bm{A}_{\mathcal{I}})}{\lambda_{\text{min}}(\bm{A}_{\mathcal{I}})}\right)^{\frac{m(n-1)}{2}}, & \bm{A} \text{ not diagonal, } \\
1, & \bm{A} \text{ diagonal. }
\end{cases}
$$
\item \label{item:sufassa}
$\mu_{n,m}(\bm{A})$ is bounded by powers of the condition number:
$$
\mu_{n,m}(\bm{A}) \leq \begin{cases} \left(\frac{\lambda_{\text{max}}(\bm{A})}{\lambda_{\text{min}}(\bm{A})}\right)^{\frac{m(n-1)}{2}}, & \bm{A} \text{ not diagonal, } \\
1, & \bm{A} \text{ diagonal. }
\end{cases}
$$
\item If $\lambda_{\text{min}}\left(\bm{\Sigma}\right) \geq \underline{\lambda}$ and $\lambda_{\text{max}}\left(\bm{\Sigma}\right) \leq \overline{\lambda}$ for all $p$, then
$$
\mu_{n,m}(\bm{\Sigma}) \leq \begin{cases} \left(\frac{\overline{\lambda}}{\underline{\lambda}}\right)^{\frac{m(n-1)}{2}}, &\bm{\Sigma} \text{ not diagonal, } \\
1, & \bm{\Sigma} \text{ diagonal. }
\end{cases}
$$
\item Let $M>0$ be a constant. Suppose
for all $p$, $\sup_{1\leq i\leq p} \Sigma_{ii}\leq M $. Moreover suppose $\lambda_{\text{min}}\left(\bm{\Sigma}\right) \geq \underline{\lambda}$ for all $p$. Then
$$
\mu_{n,m}(\bm{\Sigma}) \leq \begin{cases} \left(\frac{Mm}{\underline{\lambda}}\right)^{\frac{m(n-1)}{2}}, &\bm{\Sigma} \text{ not diagonal, } \\
1, & \bm{\Sigma} \text{ diagonal. }
\end{cases}
$$
\end{enumerate}
\end{lem}
\begin{proof}
(a)
It follows directly by definition of $\mu_{n,m}(A)$ and $\mu(A_{\mathcal{I}})\geq \left(\frac{\lambda_{\text{min}}(A_{\mathcal{I}})}{\lambda_{\text{max}}(A_{\mathcal{I}})}\right)^{m}$.\\
\noindent (b)
Since $\mu_{n,m}(A)$ is increasing in $m$ as discussed after Example \ref{exa:normaliseddeterminant}, $\mu_{n,m}(A)\leq \mu_{n,p}(A)$. The proof is then complete by applying \ref{item:sufassnewa} to $\mu_{n,p}(A)$.\\
\noindent (c)
It follows directly from \ref{item:sufassa}.\\
\noindent (d)
Let $\mathcal{I}\subset [p]$ with $|\mathcal{I}|=m$. Since $\bm{\Sigma}_{\mathcal{I}}$ is symmetric positive definite, $|(\bm{\Sigma}_{\mathcal{I}})_{ij}|\leq \sqrt{(\bm{\Sigma}_{\mathcal{I}})_{ii}(\bm{\Sigma}_{\mathcal{I}})_{jj} }\leq M$ and thus $\lambda_{\text{max}}(\bm{\Sigma}_{\mathcal{I}}) = \|\bm{\Sigma}_{\mathcal{I}}\|_2 \leq \|\bm{\Sigma}_{\mathcal{I}}\|_F \leq Mm $. By the interlacing property (see Theorem 8.1.7 in \cite{golub2012matrix}), $\lambda_{\text{min}}(\bm{\Sigma}_{\mathcal{I}}))\geq \lambda_{\text{min}}(\bm{\Sigma}).$
The proof is then completed by applying part \ref{item:sufassa}.
\end{proof}
\subsection{Proofs in Section \ref{sec:limitingcompoundPoisson}}
\label{sec:proofoflem:randgeogra}
\subsubsection{Proof of Lemma \ref{lem:randgeogra} and Lemma \ref{prop:jumpsizedecay}}
\begin{figure}%
\labellist
\small \hair 2pt
\pinlabel $\bm{0}$ at 589 552
\pinlabel $\tilde{\bm{u}}_{1}$ at 952 547
\pinlabel $\sqrt{1-\left(\frac{\|\tilde{\bm{u}}_{1}\|_2}{2}\right)^2}$ [l] at 1013 1315
\pinlabel $\tilde{\bm{u}}_{1}$ at 952 547
\pinlabel 1 at 336 1005
\endlabellist
\centering
\includegraphics[width=0.45\linewidth]{graphs/circleintersection}%
\caption{The solid circle represents the unit Euclidean ball $B_2^n$ in $\mathbb{R}^n$ while the dash circle represents the unit ball centered at $\tilde{\bm{u}}_1$. Their intersection is the green region, which is contained in the ball with center at $\tilde{\bm{u}}_1/2$ and with radius $\sqrt{1-\left(\frac{\|\tilde{\bm{u}}_{1}\|_2}{2}\right)^2}$.
}.
\label{fig:ballintersection}
\end{figure}
\begin{proof}[Proof of Lemma \ref{lem:randgeogra}]\hfil
(a)
Denote by $\operatorname{deg}(\cdot)$ the degree of a vertex in $\textbf{Ge}\left(\{\tilde{\bm{u}}_i \}_{i=1}^m, 1 ; \mathscr{N} \right)$. Then by union bound,
\begin{align*}
P\left(\operatorname{{\bf NUV}}\left(\{\tilde{\bm{u}}_i \}_{i=1}^{\delta}, 1 ; n-2 \right) \geq 1 \right) \leq & \delta P\left(\operatorname{deg}(\tilde{\bm{u}}_1)= \delta-1 \right) \\
= & \delta \mathbb{E} \left(P(\operatorname{deg}(\tilde{\bm{u}}_1) = \delta-1 | \tilde{\bm{u}}_1 )\right)\\
= & \delta \mathbb{E} \left(P\left( \left.\|\tilde{\bm{u}}_1-\tilde{\bm{u}}_2\|_2\leq 1,\cdots, \|\tilde{\bm{u}}_1-\tilde{\bm{u}}_{\delta}\|_2\leq 1 \right| \tilde{\bm{u}}_1 \right) \right) \\
=& \delta \mathbb{E} \left( P\left( \left.\|\tilde{\bm{u}}_1-\tilde{\bm{u}}_2\|_2\leq 1 \right| \tilde{\bm{u}}_1 \right)^{\delta-1} \right), \numberthis \label{eqn:tep1}
\end{align*}
where the last equality follows by conditional independence.
As illustrated in Figure \ref{fig:ballintersection}, $P\left( \|\tilde{\bm{u}}_1-\tilde{\bm{u}}_2\|_2\leq 1 | \tilde{\bm{u}}_1 \right)$ is the ratio between Lebesgue measure of green region and $|B_2^{n-2}|$. Moreover, the Lebesgue measure of the green region is less than $$
\left(1-\left(\frac{\|\tilde{\bm{u}}_{1}\|_2}{2}\right)^2\right)^{\frac{n-2}{2}}|B_2^{n-2}|.
$$
Then
\begin{equation}
P\left( \|\tilde{\bm{u}}_1-\tilde{\bm{u}}_2\|_2\leq 1 | \tilde{\bm{u}}_1 \right) \leq \left (1-\left(\frac{\|\tilde{\bm{u}}_{1}\|_2}{2}\right)^2\right)^{\frac{n-2}{2}} \quad a.s. \label{eqn:u1u21cond}
\end{equation}
By combining \eqref{eqn:tep1} and \eqref{eqn:u1u21cond},
\begin{align*}
P\left(\operatorname{{\bf NUV}}\left(\{\tilde{\bm{u}}_i \}_{i=1}^{\delta}, 1 ; n-2 \right) \geq 1 \right)
\leq & \delta \mathbb{E} \left(1-\left(\frac{\|\tilde{\bm{u}}_{1}\|_2}{2}\right)^2\right)^{\frac{(n-2)(\delta-1)}{2}}\\
= & \delta (n-2)\int_{0}^1 \left(1-\frac{r^2}{4}\right)^{\frac{(n-2)(\delta-1)}{2}} r^{n-3}dr
\numberthis \label{eqn:inmcompletebetafun} \\
=& \delta (n-2) 2^{n-3}B\left(\frac{1}{4};\frac{n-2}{2},\frac{(n-2)(\delta-1)}{2}+1\right),
\end{align*}
where the first equality follows from expressing the integral in polar coordinates, and the last step follows from changing the variables $r=2\sqrt{y}$.\\
\noindent (b)
Denote $f\left(r;\alpha,\beta \right)=\left(1-\frac{r^2}{4}\right)^\alpha r^\beta$. Then it is easy to verify that for any $\alpha,\beta>0$,
\begin{equation}
\max_{r\in [0,1]} f\left(r;\alpha,\beta\right) = \begin{cases} f\left(1;\alpha,\beta\right) =\left(\frac{3}{4}\right)^\alpha & \text{ if } 3\beta \geq 2\alpha, \\ f\left(\sqrt{\frac{4\beta}{2\alpha+\beta}};\alpha,\beta\right) = \left(\frac{2\alpha}{2\alpha+\beta}\right)^\alpha \left(\frac{4\beta}{2\alpha+\beta}\right)^{\frac{\beta}{2}} & \text{ if } 3\beta \leq 2\alpha. \end{cases} \label{eqn:eleinetwocase}
\end{equation}
Moreover, $f(r;\alpha,\beta)$ is increasing on $[0,1]$ if $3\beta \geq 2\alpha$.
Let $\alpha = \frac{(n-2)(\delta-1)}{2} $ and $\beta=n-3$. If $\delta=2$, then for any $n\geq 4$, $3\beta \geq 2\alpha$ is satisfied. Then since $f(r;\alpha,\beta)$ is increasing on $[0,1]$,
\begin{equation}
\int_0^1 f\left(r;\alpha,\beta\right) dr \leq \sqrt{\frac{4}{5}}f\left(\sqrt{\frac{4}{5}};\alpha,\beta\right)+ \left(1-\sqrt{\frac{4}{5}}\right)f\left(1;\alpha,\beta\right). \label{eqn:m2N}
\end{equation}
If $\delta=3$, then for any $n\geq 5$, $3\beta \geq 2\alpha$ is satisfied and hence \eqref{eqn:m2N} holds. For $n = 4$, $3\beta \leq 2\alpha$ is satisfied and by \eqref{eqn:eleinetwocase},
\begin{equation}
\int_0^1 f\left(r;\alpha,\beta\right) dr \leq f\left(\sqrt{\frac{4\beta}{2\alpha+\beta}};\alpha,\beta\right) = f\left(\sqrt{\frac{4}{5}};\alpha,\beta\right). \label{eqn:m3N2}
\end{equation}
If $\delta \geq 4$, it is easy to see for any $n\geq 4$, $3\beta \leq 2\alpha$ holds. By \eqref{eqn:eleinetwocase}
\begin{align*}
&\int_0^1 f\left(r;\alpha,\beta\right) dr \\
\leq &
f\left(\sqrt{\frac{4\beta}{2\alpha+\beta}};\alpha,\beta\right) \\
= & \left(\frac{\delta-1}{\delta}\right)^{\frac{(n-2)(\delta-1)}{2}} \left(\frac{4}{\delta}\right)^{\frac{n-3}{2}} \left(\frac{n-2}{n-2-\frac{1}{\delta}}\right)^{\frac{\delta-1}{2}}\left( \left(\frac{n-2}{n-2-\frac{1}{\delta}}\right)^{\delta-1} \left(\frac{n-3}{n-2-\frac{1}{m}}\right) \right)^{\frac{n-3}{2}}\\
\leq & \exp\left(\frac{1}{4}\right) \left(\frac{\delta-1}{\delta}\right)^{\frac{(n-2)(\delta-1)}{2}} \left(\frac{4}{\delta}\right)^{\frac{n-3}{2}}, \numberthis \label{eqn:inedelta4}
\end{align*}
where the last step follows from $\left(\frac{n-2}{n-2-\frac{1}{\delta}}\right)^{\delta-1} \left(\frac{n-3}{n-2-\frac{1}{\delta}}\right)\leq 1$ and $\left(\frac{n-2}{n-2-\frac{1}{\delta}}\right)^{\frac{\delta-1}{2}}\leq \exp\left(\frac{1}{4}\right)$.
Then \eqref{eqn:m2N}, \eqref{eqn:m3N2}, \eqref{eqn:inedelta4} and the fact that $f\left(\sqrt{\frac{4}{5}};\alpha,\beta\right) = \left(\frac{4}{5}\right)^{\frac{(n-2)\delta-1}{2}} $ yield the conclusion.
\end{proof}
\begin{proof}[Proof of Lemma \ref{prop:jumpsizedecay}]
(a)
Notice that
\begin{align*}
d_{\operatorname{TV}}\left(\bm{\zeta}_{n,\delta}, \operatorname{Dirac}(1) \right) = & \frac{1}{2} \sum_{\ell=1}^{\delta+1}|\bm{\zeta}_{n,\delta}(\ell) - \operatorname{Dirac}(1) \left(\ell\right) | \\
= & \sum_{\ell=2}^{\delta+1}\bm{\zeta}_{n,\delta}(\ell) \numberthis \label{eqn:tvatleast2} \\
=& \frac{\sum_{\ell=2}^{\delta+1}\left(\alpha_\ell / \ell \right) }{\alpha_1+\sum_{\ell=2}^{\delta+1}\left(\alpha_\ell /\ell \right)} \\
\leq & \frac{\sum_{\ell=2}^{\delta+1}\alpha_\ell }{\alpha_1+\sum_{\ell=2}^{\delta+1}\alpha_\ell } \\
= & \sum_{\ell=2}^{\delta+1}\alpha_\ell. \numberthis \label{eqn:tvuppbou}
\end{align*}
\noindent (b)
It follows from that
\begin{align*}
\left| \lambda_{n,\delta}(e_{n,\delta}) - \frac{1}{\delta !} \left( e_{n,\delta}\right)^\delta \right| = & \frac{1}{\delta !} \left(e_{n,\delta}\right)^\delta\left| \sum_{\ell=1}^{\delta+1}\left(\alpha_\ell/\ell\right) -1\right|\leq \frac{1}{\delta !} \left( e_{n,\delta}\right)^\delta \frac{3}{2} \sum_{\ell=2}^{\delta+1}\alpha_\ell.
\end{align*}
(c) It follows directly by part \ref{item:jumpsizedecaya}, part \ref{item:jumpsizedecayb} and Lemma \ref{eqn:cptv}.
\end{proof}
\subsubsection{Proof of Lemma \ref{lem:randgeograconduppbou}}
\begin{proof}[Proof of Lemma \ref{lem:randgeograconduppbou}]
(a)
Denote
\begin{align*}
I:=P\left(\operatorname{{\bf NUV}}\left(\{\bm{u}'_i \}_{i=1}^{\delta+1}, r ; n-2 \right) \geq 2 |\operatorname{deg}(\bm{u}'_{{\delta+1}})=\delta\right).
\end{align*}
Then by the union bound
\begin{align*}
I = & P\left(\bigcup_{i=1}^{\delta}\{\operatorname{deg}(\bm{u}'_i)=\delta\} |\operatorname{deg}(\bm{u}'_{{\delta+1}})=\delta\right)\\
\leq & \delta P\left(\operatorname{deg}(\bm{u}'_1)=\delta |\operatorname{deg}(\bm{u}'_{{\delta+1}})=\delta\right)\\
=& \delta P\left(\operatorname{deg}(\bm{u}'_1)=\delta ,\operatorname{deg}(\bm{u}'_{{\delta+1}})=\delta\right)/ P\left(\operatorname{deg}(\bm{u}'_{{\delta+1}})=\delta \right). \numberthis \label{eqn:Iunionbou}
\end{align*}
Notice that
\begin{equation}
P\left(\operatorname{deg}(\bm{u}'_{{\delta+1}})=\delta \right) = \mathbb{E} \prod_{i=1}^{\delta}P(\bm{u}'_i\in \operatorname{SC}(r,\bm{u}'_{{\delta+1}})|\bm{u}'_{{\delta+1}})=(P_{n}(r))^{\delta}, \label{eqn:denominatordegm-1}
\end{equation}
where $\operatorname{SC}(r,\bm{u}'_{{\delta+1}})$ and $P_{n}(r)$ are defined in \eqref{eqn:SCdef} and the paragraph after \eqref{eqn:SCdef}. Moreover
\begin{align*}
&P\left(\operatorname{deg}(\bm{u}'_1)=\delta ,\operatorname{deg}(\bm{u}'_{{\delta+1}})=\delta\right) \\
= & \mathbb{E} P\left(\operatorname{deg}(\bm{u}'_1)=\delta ,\operatorname{deg}(\bm{u}'_{{\delta+1}})=\delta|\bm{u}'_{1} ,\bm{u}'_{{\delta+1}}\right)\\
=& \mathbb{E} 1(\|\bm{u}'_1-\bm{u}'_{{\delta+1}}\|_2\leq r) \prod_{i=2}^{\delta}P\left(\bm{u}'_i\in \operatorname{SC}(r,\bm{u}'_1)\cap \operatorname{SC}(r,\bm{u}'_{{\delta+1}})|\bm{u}'_{1} ,\bm{u}'_{{\delta+1}} \right)\\
\leq &\mathbb{E} 1(\|\bm{u}'_1-\bm{u}'_{{\delta+1}}\|_2\leq r) \left(P_{n}(h(r,\|\bm{u}'_1-\bm{u}'_{{\delta+1}}\|_2))\right)^{\delta-1} \numberthis \label{eqn:jointmaximaldeg}
\end{align*}
where the last inequality follows from Lemma \ref{lem:jointmaximaldgreeuppbou} with
$$
h(r,d) =\sqrt{2-\frac{2-r^2}{\sqrt{1-(\frac{d}{2})^2}}}.
$$
Observing the random quantity in the expectation of \eqref{eqn:jointmaximaldeg} only depends the distance between $\|\bm{u}'_1-\bm{u}'_{{\delta+1}}\|_2$, replace $\bm{u}'_{{\delta+1}}$ with $\bm{v}_0=(1,0,\ldots,0)$ will not change its value. Then
\begin{equation}
P\left(\operatorname{deg}(\bm{u}'_1)=\delta ,\operatorname{deg}(\bm{u}'_{{\delta+1}})=\delta\right) \leq \mathbb{E} 1(\|\bm{u}'_{1}-\bm{v}_0\|_2\leq r) \left(P_{n}(h(r,\|\bm{u}'_1-\bm{v}_0\|_2))\right)^{\delta-1}. \label{eqn:jointmaximaluppbou2}
\end{equation}
Use the following coordinate system for each $\bm{u}'_1 = \left(u_{j1}:1\leq j \leq n-1 \right)^\top$ in the region $\operatorname{SC}(r, \bm{v}_0)$:
\begin{equation*}
\begin{cases}
u_{11}=1-\frac{r^2 r_1^2}{2},\\
u_{21}=r_1r \sqrt{1-\frac{r^2 r_1^2}{4}} \cos(\theta_{2}),\\
\vdots \\
u_{j1}=r_1r \sqrt{1-\frac{r^2 r_1^2}{4}} \cos(\theta_{j})\prod\limits_{m=2}^{j-1}\sin(\theta_{m}),\\
\vdots \\
u_{(n-2)1}=r_1r \sqrt{1-\frac{r^2 r_1^2}{4}}\sin(\theta_{2})\cdots\sin(\theta_{n-3})\cos(\theta_{n-2}),\\
u_{(n-1)1}=r_1r \sqrt{1-\frac{r^2 r_1^2}{4}}\sin(\theta_{2})\cdots\sin(\theta_{n-3})\sin(\theta_{n-2}),
\end{cases}
\end{equation*}
where
\begin{equation}
r_1\in [0,1], \theta_{j}\in[0,\pi] \text{ for } 2\leq j\leq n-3 \text{ and } \theta_{n-2}\in[0,2\pi). \label{eqn:intregionnonprod}
\end{equation}
Then the right hand side of \eqref{eqn:jointmaximaluppbou2} become
\begin{align*}
& \mathbb{E} 1(\|\bm{u}'_{1}-\bm{v}_0\|_2\leq r) \left(P_{n}(h(r,\|\bm{u}'_1-\bm{v}_0\|_2))\right)^{\delta-1} \\
= & \frac{1}{\text{Area}(S^{n-2})} \int_0^1 \left(P_{n}(h(r,r_1r))\right)^{\delta-1}r^{n-2}r_1^{n-3}\left(1-\frac{r^2r_1^2}{4}\right)^{\frac{n-4}{2}}dr_1 \prod_{j=2}^{n-3}\int_0^{\pi}\sin^{n-2-j}(\theta_j)d\theta_j\\
= & \frac{1}{\int_0^{\pi}\sin^{n-3}(\theta)d\theta} \int_0^1 \left(P_{n}(h(r,r_1r))\right)^{\delta-1} r^{n-2}r_1^{n-3}\left(1-\frac{r^2r_1^2}{4}\right)^{\frac{n-4}{2}}dr_1 \\
= & \frac{r^{n-2}}{B(\frac{n-2}{2},\frac{1}{2})} \int_0^1 \left(P_{n}(h(r,r_1r))\right)^{\delta-1} r_1^{n-3}\left(1-\frac{r^2r_1^2}{4}\right)^{\frac{n-4}{2}}dr_1\\
\leq & \frac{r^{n-2}}{B(\frac{n-2}{2},\frac{1}{2})} \int_0^1 \left(P_{n}(h(r,r_1r))\right)^{\delta-1}r_1^{n-3}dr_1. \numberthis \label{eqn:jointmaximaluppbou3}
\end{align*}
Plug \eqref{eqn:denominatordegm-1}, \eqref{eqn:jointmaximaluppbou2} and \eqref{eqn:jointmaximaluppbou3} into \eqref{eqn:Iunionbou} and we obtain
\begin{align*}
I \leq & \delta \frac{r^{n-2}}{B(\frac{n-2}{2},\frac{1}{2})P_{n}(r)} \int_0^1 \left(\frac{P_{n}(h(r,r_1r))}{P_{n}(r)}\right)^{\delta-1}r_1^{n-3}dr_1 \\
= &\delta (n-2) \frac{a_nr^{n-2}}{P_{n}(r)} \int_0^1 \left(\frac{P_{n}(h(r,r_1r))}{P_{n}(r)}\right)^{\delta-1}r_1^{n-3}dr_1 \numberthis \label{eqn:Iuppbou3int}
\end{align*}
where the equality follows from $a_n =\frac{1}{ (n-2)B(\frac{n-2}{2},\frac{1}{2})}$. By Lemma \ref{pderivative} \ref{item:pderivativea},
\begin{equation}
\frac{a_nr^{n-2}}{P_{n}(r)}\leq \frac{1}{\left(1-\frac{r^2}{4}\right)^{\frac{n-4}{2}}}. \label{eqn:cnrPnratiouppbou}
\end{equation}
Since when $0<r_1<1$, $0<h(r,d)/r<1$, by Lemma \ref{pderivative} \ref{item:pderivativee},
\begin{align*}
\frac{P_n(h(r,r_1 r))}{P_n(r)} \leq & \left(\frac{h(r,r_1r)}{r}\right)^{n-2}\left( \frac{ 1-\frac{h^2(r,r_1r)}{4}} { 1-\frac{r^2}{4}}\right )^{\frac{n-4}{2}}\\
\leq & \left(1 -\left(\frac{r_1}{2}\right)^2 \right)^{\frac{n-2}{2}} \frac{1}{\sqrt{1-\left(\frac{r_1r}{2}\right)^2}}\left( \frac{ 1-\frac{h^2(r,r_1r)}{4}} { (1-\frac{r^2}{4})\sqrt{1-\left(\frac{r_1r}{2}\right)^2}}\right )^{\frac{n-4}{2}}\numberthis \label{eqn:Pnratiouppboutemp}
\end{align*}
where the second inequality follows from
$$
\left(\frac{h(r,r_1r)}{r}\right)^2 = \frac{1}{\sqrt{1-\left(\frac{r_1r}{2}\right)^2}}\left( \frac{-2(\frac{r_1}{2})^2}{1+\sqrt{1-(\frac{r_1r}{2})^2}}+1 \right) \leq \frac{1}{\sqrt{1-\left(\frac{r_1r}{2}\right)^2}}\left( -\left(\frac{r_1}{2}\right)^2+1 \right).
$$
Since $h^2(r,r_1r)$ is decreasing function of $r_1\in [0,1]$, \eqref{eqn:Pnratiouppboutemp} become
\begin{align*}
\frac{P_n(h(r,r_1 r))}{P_n(r)} \leq & \left(1 -\left(\frac{r_1}{2}\right)^2 \right)^{\frac{n-2}{2}} \frac{1}{\sqrt{1-\left(\frac{r}{2}\right)^2}}\left( \frac{ 1-\frac{h^2(r,r)}{4}} { (1-\frac{r^2}{4})\sqrt{1-\left(\frac{r}{2}\right)^2}}\right )^{\frac{n-4}{2}}\\
\leq & \left(1 -\left(\frac{r_1}{2}\right)^2 \right)^{\frac{n-2}{2}} \frac{1}{\sqrt{1-\left(\frac{r}{2}\right)^2}}\left( \frac{ 1} { (1-\frac{r^2}{4})}\right )^{\frac{n-4}{2}}, \numberthis \label{eqn:Pnratiouppbou}
\end{align*}
where the second inequality follows from
$$
1-\frac{h^2(r,r)}{4} \leq \sqrt{1-\frac{r^2}{4}}.
$$
Plugging \eqref{eqn:cnrPnratiouppbou} and \eqref{eqn:Pnratiouppbou} into \eqref{eqn:Iuppbou3int},
\begin{align*}
I\leq & \delta (n-2) \frac{1}{\left(1-\frac{r^2}{4}\right)^{\frac{n+\delta-5}{2}}}\left( \frac{ 1} { (1-\frac{r^2}{4})}\right )^{\frac{(n-4)(\delta-1)}{2}} \int_0^1 \left(1 -\left(\frac{r_1}{2}\right)^2 \right)^{\frac{(n-2){(\delta-1)}}{2}} r_1^{n-3}dr_1\\
=& \bar{h}\left(\frac{1}{\sqrt{1-r^2/4}},n,\delta\right) \delta (n-2) \int_0^1 \left(1 -\left(\frac{r_1}{2}\right)^2 \right)^{\frac{(n-2){(\delta-1)}}{2}} r_1^{n-3}dr_1.
\end{align*}
\noindent
(b)
Since $\frac{1}{\sqrt{1-r^2/4}}$ is decreasing and $\bar{h}(x,n,\delta)$ as a function of $x$ is increasing,
\begin{align*}
\bar{h}\left(\frac{1}{\sqrt{1-r^2/4}},n,\delta\right)
\leq &
\begin{cases}
\bar{h}\left(\left(\frac{5}{4}\right)^{\frac{1}{4}},n,\delta\right), & \delta = 2,3\\
\bar{h}\left(\left(\frac{\delta}{\delta-1}\right)^{\frac{1}{4}},n,\delta\right), & \delta \geq 4
\end{cases} \\
=&
\begin{cases}
\left(\sqrt{\frac{5}{4}}\right)^{\frac{n+\delta-5}{2}}\left(\sqrt{\frac{5}{4}}\right)^{\frac{(n-2)(\delta-1)}{2}}, & \delta = 2,3\\
\left(\sqrt{\frac{\delta}{\delta-1}}\right)^{\frac{n+\delta-5}{2}}\left(\sqrt{\frac{\delta}{\delta-1}}\right)^{\frac{(n-2)(\delta-1)}{2}}, & \delta \geq 4
\end{cases}. \numberthis \label{eqn:hbaruppbou}
\end{align*}
Then the proof is complete by combining part \ref{item:randgeograconduppboua}, Lemma \ref{lem:randgeogra} \ref{item:randgeograb} and \eqref{eqn:hbaruppbou}.\\
\noindent (c)
Similar to \eqref{eqn:tvuppbou}, we have
$$
d_{\operatorname{TV}}(\bm{\zeta}_{n,\delta,\rho},\operatorname{Dirac}(1))\leq \sum_{\ell=2}^{\delta+1} \alpha(\ell,r_{\rho})
= P\left(\operatorname{{\bf NUV}}\left(\{\bm{u}'_i \}_{i=1}^{\delta+1}, r_\rho ; n-2 \right) \geq 2 |\operatorname{deg}(\bm{u}'_{{\delta+1}})=\delta\right).
$$
where the equality follows from \eqref{eqn:alphalrhorgg}. Then the conclusion follows from part \ref{item:randgeograconduppboua} and part \ref{item:randgeograconduppboub} since $r_\rho$ satisfies the condition there.
\end{proof}
\begin{lem} \label{lem:jointmaximaldgreeuppbou}
Let $n\geq 3$ and $0<r<\sqrt{2}$. If $\bm{w}_1$ and $\bm{w}_2$ are two points in $S^{n-2}$ with $\|\bm{w}_1-\bm{w}_2\|_2 = d$ satisfying $2-2\sqrt{1-(d/2)^2}<r^2$, then
$$
P\left(\bm{u}'_1\in \operatorname{SC}(r,\bm{w}_1)\cap \operatorname{SC}(r,\bm{w}_{2}) \right) \leq P_{n}(h(r,d))
$$
where $\bm{u}'_1$ has distribution $\operatorname{{\bf unif}}(S^{n-2})$ and
$$
h(r,d) =\sqrt{2-\frac{2-r^2}{\sqrt{1-(\frac{d}{2})^2}}}.
$$
\end{lem}
\begin{figure}%
\labellist
\small \hair 2pt
\pinlabel $\bm{0}$ at 20 600
\pinlabel $\bm{w}_1$ at 727 12
\pinlabel $\bm{w}_2$ at 727 1252
\pinlabel $\bm{w}_3$ at 710 600
\pinlabel $\bm{w}_4$ at 780 680
\pinlabel $\bm{w}_5$ at 900 480
\pinlabel $r$ at 810 250
\pinlabel $r$ at 810 865
\pinlabel $\theta$ at 280 640
\pinlabel $1$ at 373 266
\pinlabel $1$ at 373 1070
\endlabellist
\centering
\includegraphics[width=7cm, height=4cm]{graphs/capintersectionuppbou1}%
\caption{$\bm{0}$ is the origin in $\mathbb{R}^{n-2}$ and $\bm{w}_1,\bm{w}_2,\bm{w}_4,\bm{w}_5$ are on $S^{n-2}$. $\bm{w}_3$ is the midpoint of $z_1$ and $z_2$, while $z_4$ is the midpoint of the shortest arc on $S^{n-2}$ connecting $\bm{w}_1$ and $\bm{w}_2$. $\bm{w}_5$ is one of the two intersection points of the boundary $\operatorname{SC}(r,\bm{w}_1)$ and the boundary of $\operatorname{SC}(r,\bm{w}_2)$. The angle between line segment $\bm{0}\bm{w}_4$ and $\bm{0}\bm{w}_5$ is $\theta$.
}
\label{fig:capintersection}
\end{figure}
\begin{proof} The proof is based on Figure \ref{fig:capintersection} and we use $|\cdot|$ to represent the length of a line segment in this proof. In the right triangle $\bm{0}\bm{w}_3\bm{w}_1$, the line segment $\bm{0}\bm{w}_3$ has length $|\bm{0}\bm{w}_3|=\sqrt{1-(d/2)^2}$. In the right triangle $\bm{w}_1\bm{w}_3\bm{w}_5$, $|\bm{w}_3\bm{w}_5| =\sqrt{r^2-(d/2)^2}$. In the triangle $\bm{0}\bm{w}_3\bm{w}_5$, by the law of Cosines,
$$
\cos(\theta) = \frac{2-r^2}{2\sqrt{1-(\frac{d}{2})^2}}.
$$
Then in the isosceles triangle, the line segment $\bm{w}_4\bm{w}_5$ has length
$$
|\bm{w}_4\bm{w}_5|=2\sin(\theta/2) = \sqrt{2(1-\cos(\theta))} = \sqrt{2-\frac{2-r^2}{\sqrt{1-(\frac{d}{2})^2}}} = h(r,d).
$$
It is easy to deduce that $|\bm{w}_1\bm{w}_4|=\sqrt{2-2\sqrt{1-(d/2)^2}}$. The condition $2-2\sqrt{1-(d/2)^2}<r^2$ entails that $\operatorname{SC}(r,\bm{w}_1)\cap \operatorname{SC}(r,\bm{w}_{2})\not= \emptyset$ and that $|\bm{w}_1\bm{w}_4|<|\bm{w}_4\bm{w}_5|=h(r,d)$. In this case $\operatorname{SC}(r,\bm{w}_1)\cap \operatorname{SC}(r,\bm{w}_{2}) \subset \operatorname{SC}(h(r,d),\bm{w}_4) $. Thus
$$
P\left(\bm{u}'_1\in \operatorname{SC}(r,\bm{w}_1)\cap \operatorname{SC}(r,\bm{w}_{2}) \right) \leq P\left(\bm{u}'_1\in \operatorname{SC}(h(r,d),\bm{w}_4)\right) = P_{n}(h(r,d)).
$$
\end{proof}
\subsubsection{Proof of Lemma \ref{lem:alphaldelta2}}
\begin{proof}[Proof of Lemma \ref{lem:alphaldelta2}]
When $\delta=2$, $\alpha_2=0$ since either both vertices have degree $1$ or none. Moreover,
\begin{align*}
\alpha_3
= &
P(\|\tilde{\bm{u}}_1-\tilde{\bm{u}}_2\|_2\leq 1)\\
= &
\mathbb{E} P(\|\tilde{\bm{u}}_1-\tilde{\bm{u}}_2\|_2\leq 1|\tilde{\bm{u}}_1)\\
\overset{(*)}{=} &
\mathbb{E} \frac{1}{\text{Vol}(B^{n-2})}\times2\times \frac{\pi^{(n-3)/2}}{\Gamma(\frac{n-3}{2}+1)}\int_{0}^{\arccos(\frac{\|\tilde{\bm{u}}_1\|_2}{2})} \sin^{n-2}(\theta)d\theta\\
\overset{(**)}{=} &
\frac{1}{\text{Vol}(B^{n-2})}\times2\times \frac{\pi^{(n-3)/2}}{\Gamma(\frac{n-3}{2}+1)}\frac{\text{Area}(S^{n-3})}{\text{Vol}(B^{n-2})} \int_0^1 r^{n-3} \int_{0}^{\arccos(\frac{r}{2})} \sin^{n-2}(\theta)d\theta dr\\
=& \frac{2(n-2)}{B(\frac{n-1}{2},\frac{1}{2})} \int_0^1 r^{n-3} \int_{0}^{\arccos(\frac{r}{2})} \sin^{n-2}(\theta)d\theta dr \numberthis \label{eqn:alpha3delta2formula1}
\end{align*}
where step $(*)$ follows from the Subsection ``Volume of a hyperspherical cap'' from \cite{li2011concise} and $\text{Vol}(B^{n-2})$ is the volume of $B^{n-2}$, step $(**)$ follows by observing the random quantity only depends on $\tilde{\bm{u}}_1$ through its Euclidean norm, and in the last step $B(\cdot,\cdot)$ is the Beta function. By Fubini's Theorem
\begin{align*}
&\int_0^1 r^{n-3} \int_{0}^{\arccos(\frac{r}{2})} \sin^{n-2}(\theta)d\theta dr \\
= & \int_{0}^{\frac{\pi}{3}} \int_0^1 r^{n-3} \sin^{n-2}(\theta)dr d\theta + \int_{\frac{\pi}{3}}^{\frac{\pi}{2}} \int_0^{2\cos(\theta)} r^{n-3} \sin^{n-2}(\theta)dr d\theta \\
=& \frac{3}{2(n-2)} \int_{0}^{\frac{\pi}{3}} \sin^{n-2}(\theta) d\theta
\end{align*}
Plugging the preceding formula into \eqref{eqn:alpha3delta2formula1}, $\alpha_3 = \frac{3}{2} I_{\frac{3}{4}}(\frac{n-1}{2},\frac{1}{2})$, where $I_x(a,b)$ is the regularized incomplete Beta function. $\alpha_1=1-\alpha_3$ follows from $\alpha_2=0$.
\end{proof}
\section{Proofs in Subsection \ref{sec:compoihighdim}}
\subsection{Proofs in Subsection \ref{sec:compoihighdim}}
\subsubsection{{Proof of Lemma \ref{prop:rangeo}}}
\label{sec:proofofprop:rangeo}
To utilize the notation we have defined in this paper, we make the following adjustments to the notation throughout this subsection. In this proof it suffices to prove the conclusion for any $\delta+1$ i.i.d. random points from $\operatorname{{\bf unif}}(S^{n-2})$. Without loss of generality assume in this subsection that the first $\delta+1$ U-scores $\{\bm{u}_i\}_{i=1}^{\delta+1}$ are independent. Another adjustment is to replace $r$ by $r_\rho$. With these adjustments Lemma \ref{prop:rangeo} is equivalent to the following: when $r_\rho <2/\sqrt{5}$, $\delta\geq 1$, for any $\ell\in [\delta+1]$,
\begin{align*}
&P\left(\operatorname{{\bf NUV}}\left(\{\bm{u}_i \}_{i=1}^{\delta+1}, r_\rho \right)=\ell|\operatorname{deg}(\bm{u}_{\delta+1})=\delta\right) \\
= & P\left(\operatorname{ {\bf PNUV} }\left(\{\bm{u}_i \}_{i=1}^{\delta+1}, r_\rho \right)=\ell|\operatorname{deg}(\bm{u}_{\delta+1})=\delta\right). \numberthis \label{eqn:proofequilem1}
\end{align*}
Take $\vec{i}=(\delta+1,1,\cdots,\delta)$. Recall the notation $\vec{i},\Phi_{\vec{i}}=\Phi^{(\bm{R})}_{\vec{i}},U_{\vec{i}}$ are defined in Subsection \ref{sec:closenessnumedge}, where the dependence of $\bm{R}$ in $\Phi_{\vec{i}}^{(\bm{R})}$ is suppressed throughout this subsection for the sake of clearer exposition. Then $\operatorname{ {\bf PNUV} }\left(\{\bm{u}_i \}_{i=1}^{\delta+1}, r_\rho \right)= U_{\vec{i}}+ \Phi_{\vec{i}}$. Moreover, the
event $\{\operatorname{deg}(\bm{u}_{\delta+1})=\delta\}$ in $\textbf{PGe}\left(\{\bm{u}_i \}_{i=1}^{\delta+1}, r_\rho \right)$ is the same as $\{\Phi_{\vec{i}} =1 \}$. Define $F_{ij}^{(q)}=\{\|\bm{u}_i-q\bm{u}_j\|_2 \leq r_\rho \}$. Then $1\left(\bigcap\limits_{j=1}^\delta F_{j(\delta+1)}^{(+1)}\right)$ is the indicator function of the event that the degree of vertex $\bm{u}_{\delta+1}$ in $\textbf{Ge}\left(\{\bm{u}_i \}_{i=1}^{\delta+1}, r_\rho \right)$ is $\delta$. Hence Lemma \ref{prop:rangeo} is also equivalent to the following: when $r_\rho <2/\sqrt{5}$, $\delta\geq 1$, for any $\ell\in [\delta+1]$,
\begin{equation}
P\left(U_{\vec{i}}+ \Phi_{\vec{i}}=\ell|\Phi_{\vec{i}}=1\right) = P\left(\operatorname{{\bf NUV}}\left(\{\bm{u}_i \}_{i=1}^{\delta+1}, r_\rho \right) =\ell \left| 1\left(\bigcap_{j=1}^\delta F_{j(\delta+1)}^{(+1)}\right) =1 \right. \right). \label{eqn:pseudogeogeoequiv}
\end{equation}
\begin{proof}[Proof of \eqref{eqn:pseudogeogeoequiv}]
For $\vec{q} =(q_1,q_2,\cdots,q_\delta) \in \{-1,+1\}^{\delta}$, denote $F_{\delta+1}^{(\vec{q})} = \bigcap_{j=1}^\delta F_{j(\delta+1)}^{(q_j)}$.
Observe that
$$
\{\Phi_{\vec{i}} =1 \} = \bigcap_{j=1}^{\delta} \ \ \bigcup_{q_j \in \{+1,-1\} } F_{j(\delta+1)}^{(q_j)} =
= \bigcup_{\vec{q}\in \{-1,+1\}^{\delta} } \ \ \ F_{\delta+1}^{(\vec{q})}.
$$
Since $r_\rho < 2/\sqrt{5}< \sqrt{2} $, $F_{j(\delta+1)}^{(-1)}$ and $F_{j(\delta+1)}^{(+1)}$ are disjoint for every $j\in [\delta]$, which implies $F_{\delta+1}^{(\vec{q})}$ for different $\vec{q} \in \{-1,+1\}^{\delta} $ are disjoint. Hence,
\begin{equation}
P (\Phi_{\vec{i}} =1, U_{\vec{i}}=\ell -1) = \sum_{\vec{q} \in \{-1,+1\}^{\delta}} P(F_{\delta+1}^{(\vec{q})}, U_{\vec{i}} =\ell -1). \label{eqn:temp1}
\end{equation}
Next observe that $1(F_{\delta+1}^{(\vec{q})})$ is a function of $\bm{u}_1,\cdots,\bm{u}_{\delta+1}$, and hence it has the same distribution when replacing $\bm{u}_i$ by $-\bm{u}_i$ for any $i\in [\delta]$. Moreover, replacing $\bm{u}_i$ by $-\bm{u}_i$ for any $i\in [\delta]$ wouldn't change $U_{\vec{i}}$. As a result, \eqref{eqn:temp1} implies
\begin{equation}
P (\Phi_{\vec{i}} =1, U_{\vec{i}}=\ell -1) = 2^{\delta} P\left(F_{\delta+1}^{(\vec{q}_0)}, U_{\vec{i}} =\ell -1\right), \label{eqn:temp2}
\end{equation}
where $\vec{q}_0 =(+1,+1,\cdots,+1)$ is the vector in $\mathbb{R} ^{\delta}$ with all its components $+1$.
Consider $\omega\in F_{\delta+1}^{(\vec{q}_0)}$. Then $\Phi_{\vec{i}}(\omega)=1$ or equivalently, $\Phi_{i(\delta+1)}^{(\bm{R})}(\omega)=1$ for any $i \in [\delta]$. Then
\begin{equation}
U_{\vec{i}}(\omega) = \sum_{\vec{j}\in S_{\vec{i}}} \Phi_{\vec{j}} (\omega) = \sum_{i=1}^\delta \prod_{\substack{j=1\\ j\not = i} }^{\delta+1} \Phi^{(\bm{R})}_{ij} (\omega)=\sum_{i=1}^\delta \prod_{\substack{j=1\\ j\not = i} }^{\delta} \Phi^{(\bm{R})}_{ij} (\omega). \label{eqn:Uitemptemp2}
\end{equation}
Since for any distinct $i,j\in [\delta]$, $\|\bm{u}_i(\omega) - \bm{u}_j (\omega) \|_2 \leq \| \bm{u}_i -\bm{u}_{\delta+1}(\omega)\|_2 + \| \bm{u}_i (\omega) -\bm{u}_{\delta+1}(\omega)\|_2 \leq 2r_\rho < 4/\sqrt{5} $, $\|\bm{u}_i(\omega) + \bm{u}_j (\omega) \|_2 =\sqrt{4- \|\bm{u}_i(\omega) - \bm{u}_j (\omega) \|_2^2 } > 2/\sqrt{5} > r_\rho $. Thus $\Phi^{(\bm{R})}_{ij}(\omega) = 1_{F_{ij}^{(+1)}}(\omega) $. That is, in the set $F_{\delta+1}^{(\vec{q}_0)}$, \eqref{eqn:Uitemptemp2} becomes
\begin{equation}
U_{\vec{i}} = \sum_{i=1}^\delta \prod_{\substack{j=1\\ j\not = i} }^\delta 1(F_{ij}^{(+1)}) = \operatorname{{\bf NUV}}\left(\{\bm{u}_i \}_{i=1}^{\delta}, r_\rho \right), \label{eqn:Uidef2}
\end{equation}
which implies
\begin{align*}
\left(\Phi_{\vec{i}} + U_{\vec{i}}\right) 1\left(F_{\delta+1}^{(\vec{q}_0)}\right) =& \left(1+ \operatorname{{\bf NUV}}\left(\{\bm{u}_i \}_{i=1}^{\delta}, r_\rho \right) \right) 1\left(F_{\delta+1}^{(\vec{q}_0)}\right) \\
= & \operatorname{{\bf NUV}}\left(\{\bm{u}_i \}_{i=1}^{\delta+1}, r_\rho \right) 1\left(F_{\delta+1}^{(\vec{q}_0)}\right) . \numberthis \label{eqn:temptemp1}
\end{align*}
Thus,
\begin{align*}
P\left(U_{\vec{i}}+ \Phi_{\vec{i}}=\ell|\Phi_{\vec{i}}=1\right) =&
\frac{2^{\delta} P\left(F_{\delta+1}^{(\vec{q}_0)}, U_{\vec{i}} =\ell -1\right)}{P \left(\Phi_{\vec{i}}=1\right)}\\
=&
\frac{2^{\delta} P\left(F_{\delta+1}^{(\vec{q}_0)}, U_{\vec{i}} =\ell -1\right)}{\left(2P_n(r_\rho)\right)^\delta} \\
=&
\frac{ P\left(F_{\delta+1}^{(\vec{q}_0)}, \operatorname{{\bf NUV}}\left(\{\bm{u}_i \}_{i=1}^{\delta+1}, r_\rho \right) =\ell \right)}{\left(P_n(r_\rho)\right)^\delta} \\
=&
P\left(\operatorname{{\bf NUV}}\left(\{\bm{u}_i \}_{i=1}^{\delta+1}, r_\rho \right) =\ell \left| 1\left(\bigcap_{j=1}^\delta F_{j(\delta+1)}^{(+1)}\right) =1 \right. \right),
\end{align*}
where the first equality follows from \eqref{eqn:temp2}, the second equality follows from Lemma \ref{item:prod}, and the third equality follows from \eqref{eqn:temptemp1}.
\end{proof}
\subsubsection{Proofs of Lemma \ref{prop:rangeolimit} and Lemma \ref{lem:compoundpoissonlimit}}
\label{sec:proofofprop:rangeolimit}
\begin{comment}
As in subsection \ref{sec:proofofprop:rangeo}, to utilize the notation we have defined in this paper, we make the following adjustments on the notation throughout this subsection. In this proof it suffices to prove the conclusion for any $\delta+1$ i.i.d. random points from $\operatorname{{\bf unif}}(S^{n-2})$. Without loss of generality assume in this subsection that the first $\delta+1$ U-scores $\{\bm{u}_i\}_{i=1}^{\delta+1}$ are independent. The last adjustment is to replace $r$ by $r_\rho$. With these adjustments Lemma \ref{prop:rangeolimit} is equivalent to prove: when $\delta\geq 1$ and $\{\tilde{\bm{u}}\}_{i=1}^{\delta} \overset{i.i.d.}{\sim} \operatorname{{\bf unif}}(B^{n-2})$, for any $\ell\in [\delta+1]$,
\begin{align*}
&\lim_{r_\rho\to 0^+}P\left(\operatorname{{\bf NUV}}\left(\{\bm{u}_i \}_{i=1}^{\delta+1}, r_\rho \right)=\ell|\operatorname{deg}(\bm{u}_{\delta+1})=\delta \right)\\
= & P\left(\operatorname{{\bf NUV}}\left(\{\tilde{\bm{u}}_i \}_{i=1}^{\delta}, 1 \right)=\ell-1\right), \numberthis \label{eqn:tobeproof2}
\end{align*}
\end{comment}
\begin{proof}[Proof of Lemma \ref{prop:rangeolimit}]
In the set $\{\operatorname{deg}(\bm{u}'_{\delta+1})=\delta\}$, it follows that
$$
\operatorname{{\bf NUV}}\left(\{\bm{u}'_i \}_{i=1}^{\delta+1}, r \right) = \operatorname{{\bf NUV}}\left( \{\bm{u}'_i \}_{i=1}^\delta, r \right)+1.
$$
Thus,
\begin{align*}
&P\left(\operatorname{deg}(\bm{u}'_{\delta+1})=\delta, \operatorname{{\bf NUV}}\left(\{\bm{u}'_i \}_{i=1}^{\delta+1}, r \right) =\ell\right) \\
=& \mathbb{E} 1\left(\{\operatorname{deg}(\bm{u}'_{\delta+1})=\delta\} \bigcap \{ \operatorname{{\bf NUV}}\left( \{\bm{u}'_i \}_{i=1}^\delta, r \right)= \ell -1 \} \right) \\
=&\mathbb{E}\left( \mathbb{E} \left( 1\left(\{\operatorname{deg}(\bm{u}'_{\delta+1})=\delta\} \bigcap \{ \operatorname{{\bf NUV}}\left( \{\bm{u}'_i \}_{i=1}^\delta, r \right)= \ell -1 \} \right) \big| \bm{u}'_{\delta+1} \right) \right) \\
=& \mathbb{E} \left( 1\left(\{\operatorname{deg}(\bm{u}'_{\delta+1})=\delta\} \bigcap \{ \operatorname{{\bf NUV}}\left( \{\bm{u}'_i \}_{i=1}^\delta, r \right)= \ell -1 \} \right) \big| \bm{u}'_{\delta+1} = \bm{v}_0 \right), \numberthis \label{eqn:tempcond}
\end{align*}
where the last equality follows from that
$$
\mathbb{E} \left( 1\left(\{\operatorname{deg}(\bm{u}'_{\delta+1})=\delta\} \bigcap \{ \operatorname{{\bf NUV}}\left( \{\bm{u}'_i \}_{i=1}^{\delta}, r \right)= \ell -1 \} \right) \big| \bm{u}'_{\delta+1} \right)
$$
as a function of the random variable $\bm{u}'_{\delta+1}$, due to the rotation invariance property of the distribution $\operatorname{{\bf unif}}(S^{n-2})$, equals to
$$
\mathbb{E} \left( 1\left(\{\operatorname{deg}(\bm{u}'_{\delta+1})=\delta\} \bigcap \{ \operatorname{{\bf NUV}}\left( \{\bm{u}'_i \}_{i=1}^\delta, r \right)= \ell -1 \} \right) \big| \bm{u}'_{\delta+1} = \bm{v}_0 \right)
$$
a.s. with $\bm{v}_0 = (1,0,0,\cdots,0)\in S^{n-2}$.
Under the condition $\bm{u}'_{\delta+1} = \bm{v}_0$, $1\left(\{\operatorname{deg}(\bm{u}'_{\delta+1})=\delta\}\right) = \prod_{i=1}^{\delta} 1\left( \bm{u}'_i \in \operatorname{SC}(r, \bm{v}_0) \right)$, where $\operatorname{SC}(r, \bm{v}_0)$ is defined in \eqref{eqn:SCdef}. Consider the coordinate system for $\bm{u}'_i = \left(u'_{ji}:1\leq j \leq n-1\right)^\top$ in the region $\operatorname{SC}(r, \bm{v}_0)$:
\begin{equation*}
\begin{cases}
u'_{1i}=1-\frac{r^2 r_i^2}{2},\\
u'_{2i}=r r_i\sqrt{1-\frac{r^2 r_i^2}{4}} \cos(\theta_{2i}),\\
\vdots & \text{\ \ for } 1\leq i \leq \delta,\\
u'_{ji}=r r_i\sqrt{1-\frac{r^2 r_i^2}{4}} \cos(\theta_{ji})\prod\limits_{m=2}^{j-1}\sin(\theta_{mi}),\\
\vdots \\
u'_{(n-2)i}=r r_i\sqrt{1-\frac{r^2 r_i^2}{4}}\sin(\theta_{2i})\cdots\sin(\theta_{(n-3)i})\cos(\theta_{(n-2)i}),\\
u'_{(n-1)i}=r r_i\sqrt{1-\frac{r^2 r_i^2}{4}}\sin(\theta_{2i})\cdots\sin(\theta_{(n-3)i})\sin(\theta_{(n-2)i}),
\end{cases}
\end{equation*}
where for each $i\in [\delta]$:
\begin{equation}
r_i\in [0,1], \theta_{ji}\in[0,\pi] \text{ for } 2\leq j\leq n-3 \text{ and } \theta_{(n-2)i}\in[0,2\pi). \label{eqn:intregion}
\end{equation}
Then,
\begin{align*}
&r^{-(n-2)\delta}\mathbb{E} \left( 1\left(\{\operatorname{deg}(\bm{u}'_{\delta+1})=\delta\} \bigcap \{ \operatorname{{\bf NUV}}\left( \{\bm{u}'_i \}_{i=1}^\delta, r \right)= \ell -1 \} \right) \big| \bm{u}'_{\delta+1} = \bm{v}_0 \right) \\
= & r^{-(n-2)\delta} \mathbb{E} \prod_{i=1}^{\delta} 1\left( \bm{u}'_i \in \operatorname{SC}(r, \bm{v}_0) \right) 1\left( \operatorname{{\bf NUV}}\left( \{\bm{u}'_i \}_{i=1}^\delta, r \right)= \ell -1\right) \\
\overset{(*)}{=}\ & \frac{r^{-(n-2)\delta}}{|S^{n-2}|^\delta} \int\cdots\int_{\Omega_0} 1\left( \operatorname{{\bf NUV}}\left( \{\bm{u}'_i \}_{i=1}^\delta, r \right)= \ell -1\right)\\
&\ \ \ \times \prod_{i=1}^{\delta}\left( r^{n-2}r_i^{n-3}\left(1-\frac{r^2 r^2_i}{4}\right)^{\frac{n-4}{2}} dr_i \prod_{j=2}^{n-2} \left(\sin^{n-2-j}(\theta_{ji}) d\theta_{ji} \right)\right) \\
=& \frac{1}{|S^{n-2}|^\delta} \int\cdots\int_{\Omega_0} 1\left( \operatorname{{\bf NUV}}\left( \{\bm{u}'_i \}_{i=1}^\delta, r \right)= \ell -1\right)\\
&\ \ \ \times\prod_{i=1}^{\delta}\left( r_i^{n-3}\left(1-\frac{r^2 r^2_i}{4}\right)^{\frac{n-4}{2}} dr_i \prod_{j=2}^{n-2} \left(\sin^{n-2-j}(\theta_{ji}) d\theta_{ji} \right)\right), \numberthis \label{eqn:temp4}
\end{align*}
where $\Omega_0$ in equality $(*)$ is the region described in \eqref{eqn:intregion}. Denote by $f(r)$ the integrand in \eqref{eqn:temp4}. $f(r)$ is a function of $r_i$, $\theta_{ji}$ for $2\leq j\leq n-2$ and $1\leq i \leq \delta$, of which the dependences are suppressed.
Note $\operatorname{{\bf NUV}}\left(\{\bm{v}_i \}_{i=1}^\delta, r \right)$ is a function of $(1\left(\|\bm{v}_i-\bm{v}_j\|_2\leq r\right): 1\leq i<j \leq \delta)$ and it does not depend on the specific location of each vertex. Thus in \eqref{eqn:temp4} $ \operatorname{{\bf NUV}}\left( \{\bm{u}'_i \}_{i=1}^\delta, r \right)$ is a function of $(1\left(\|\bm{u}'_i-\bm{u}'_j\|_2< r\right): 1\leq i<j \leq \delta)$ since $\|\bm{u}'_i-\bm{u}'_j\|_2 = r$, as a set of Lebesgue measure $0$, contributes nothing to the integral.
We then write
\begin{align*}
\operatorname{{\bf NUV}}\left( \{\bm{u}'_i \}_{i=1}^\delta, r \right) = & \chi(1\left(\|\bm{u}'_i-\bm{u}'_j\|_2< r\right): 1\leq i<j \leq \delta)\\
= & \chi\left(1\left(\frac{1}{r}\|\bm{u}'_i-\bm{u}'_j\|_2< 1\right): 1\leq i<j \leq \delta\right).
\end{align*}
Intrinsically, $1\left(\|\bm{u}'_i-\bm{u}'_j\|_2< r\right)$ is the indicator random variable about whether there is an edge between vertex $i$ and $j$, and the function $\chi$ is the function that takes all edge information among $\delta$ vertices as input and outputs the number of universal vertices.
Then as $r\to 0^+$,
\begin{align*}
\lim_{r\to 0^+} f(r)= & \prod_{i=1}^{\delta}\left( r_i^{n-3} \prod_{j=2}^{n-2} \left(\sin^{n-2-j}(\theta_{ji}) \right)\right)\\
& \times\lim_{r\to 0^+} 1\left(\chi\left(1\left(\frac{1}{r}\|\bm{u}'_i-\bm{u}'_j\|_2< 1\right): 1\leq i<j \leq \delta\right) = \ell-1 \right). \numberthis \label{eqn:temp3}
\end{align*}
Observe
\begin{align*}
&\lim_{r \to 0^+}\left(\frac{1}{r}\|\bm{u}'_i-\bm{u}'_j\|_2\right)^2 \\
= & \left( r_i \cos(\theta_{2i}) - r_j \cos(\theta_{2j}) \right)^2+\sum_{q=3}^{n-2} \left( r_i \prod_{m=2}^{q-1} \sin(\theta_{mi})\cos(\theta_{qi}) - r_j \prod_{m=2}^{q-1} \sin(\theta_{mj})\cos(\theta_{qj}) \right)^2 \\
& \quad + \left( r_i \prod_{m=2}^{n-2} \sin(\theta_{mi}) - r_j \prod_{m=2}^{n-2} \sin(\theta_{mj}) \right)^2. \numberthis \label{eqn:temp5}
\end{align*}
On $\Omega_0$, for $1\leq i \leq \delta$, define
\begin{equation}
\begin{cases}
\tilde{u}'_{1i} = r_i \cos(\theta_{2i}),\\
\tilde{u}'_{ji}=r_i \cos(\theta_{(j+1)i}) \prod\limits_{m=2}^{j}\sin(\theta_{mi}),\text{ for } 2\leq j \leq n-3 & \\
\tilde{u}'_{(n-2)i}= r_i\prod\limits_{m=2}^{n-2}\sin(\theta_{mi}),\\
\end{cases} \label{eqn:newballparametrization}
\end{equation}
and $\tilde{\bm{u}}'_i =(\tilde{u}'_{ji}:1\leq j \leq n-2)\in B^{n-2}$. Then by \eqref{eqn:temp5}
$$
\lim_{r\to 0^+}\frac{1}{r}\|\bm{u}'_i-\bm{u}'_j\|_2 = \| \tilde{\bm{u}}'_i -\tilde{\bm{u}}'_j\|_2,
$$
which, together with \eqref{eqn:temp3}, imply
\begin{align*}
&\lim_{r \to 0^+} f(r) \\
= & \prod_{i=1}^{\delta}\left( r_i^{n-3} \prod_{j=2}^{n-2} \left(\sin^{n-2-j}(\theta_{ji}) \right)\right) 1\left(\chi\left(1\left(\|\tilde{\bm{u}}'_i-\tilde{\bm{u}}'_j\|_2< 1\right): 1\leq i<j \leq \delta\right) = \ell-1 \right)\\
=&\prod_{i=1}^{\delta}\left( r_i^{n-3} \prod_{j=2}^{n-2} \left(\sin^{n-2-j}(\theta_{ji}) \right)\right) 1\left(\chi\left(1\left(\|\tilde{\bm{u}}'_i-\tilde{\bm{u}}'_j\|_2\leq 1\right): 1\leq i<j \leq \delta\right) = \ell-1 \right)\\
=& \prod_{i=1}^{\delta}\left( r_i^{n-3} \prod_{j=2}^{n-2} \left(\sin^{n-2-j}(\theta_{ji}) \right)\right) 1\left(\operatorname{{\bf NUV}}\left( \{\tilde{\bm{u}}'_i \}_{i=1}^\delta, 1 \right)= \ell -1\right),
\end{align*}
where the second equality holds $a.s.$ with respect to the Lebesgue measure on $\Omega_0$,
Moreover, $|f(r)|\leq 1$, which is integrable over the bounded set $\Omega_0$. Applying Dominated Convergence Theorem to \eqref{eqn:temp4},
\begin{align*}
&\lim_{r\to 0^+}r^{-(n-2)\delta}\mathbb{E} \left( 1\left(\{\operatorname{deg}(\bm{u}'_{\delta+1})=\delta\} \bigcap \{ \operatorname{{\bf NUV}}\left( \{\bm{u}'_i \}_{i=1}^\delta, r \right)= \ell -1 \} \right) \big| \bm{u}'_{\delta+1} = \bm{v}_0 \right) \\
=& \frac{1}{|S^{n-2}|^\delta} \int\cdots\int_{\Omega_0} 1\left( \operatorname{{\bf NUV}}\left( \{\tilde{\bm{u}}'_i \}_{i=1}^\delta, 1 \right)= \ell -1\right) \prod_{i=1}^{\delta} r_i^{n-3} dr_i \prod_{j=2}^{n-2} \sin^{n-2-j}(\theta_{ji}) d\theta_{ji} \\
= & \frac{|B^{n-2}|^\delta}{|S^{n-2}|^\delta} P\left( \operatorname{{\bf NUV}}\left( \{\tilde{\bm{u}}_i \}_{i=1}^\delta, 1 \right)= \ell -1\right), \numberthis \label{eqn:templimit}
\end{align*}
where the parametrization \eqref{eqn:newballparametrization} and the region $\Omega_0$ coincide with the spherical coordinates for $B^{n-2}$.
Thus
\begin{align*}
&\lim_{r \to 0^+} P\left(\operatorname{{\bf NUV}}\left(\{\bm{u}'_i \}_{i=1}^{\delta+1}, r \right) =\ell \left| \operatorname{deg}(\bm{u}'_{\delta+1})=\delta \right. \right) \\
= & \lim_{r \to 0^+} r^{-(n-2)\delta}P\left(\operatorname{deg}(\bm{u}'_{\delta+1})=\delta, \operatorname{{\bf NUV}}\left(\{\bm{u}'_i \}_{i=1}^{\delta+1}, r \right) =\ell\right) \frac{r^{(n-2)\delta}}{\left(P_n(r)\right)^\delta} \\
= & \frac{|B^{n-2}|^\delta}{|S^{n-2}|^\delta} P\left( \operatorname{{\bf NUV}}\left( \{\tilde{\bm{u}}_i \}_{i=1}^\delta, 1 \right)= \ell -1\right) \frac{1}{(a_n)^\delta} \\
= & \frac{1}{(a_n)^\delta} \frac{|B^{n-2}|^\delta}{|S^{n-2}|^\delta} P\left( \operatorname{{\bf NUV}}\left( \{\tilde{\bm{u}}_i \}_{i=1}^\delta, 1 \right)= \ell -1\right) \\
= & P\left( \operatorname{{\bf NUV}}\left( \{\tilde{\bm{u}}_i \}_{i=1}^\delta, 1 \right)= \ell -1\right), \end{align*}
where the second equality follows from \eqref{eqn:tempcond}, \eqref{eqn:templimit} and Lemma \ref{pderivative} \ref{item:pderivativeb}.
\end{proof}
\begin{proof}[Proof of Lemma \ref{lem:compoundpoissonlimit}]
By \eqref{eqn:alphalrhorgg}, Lemma \ref{prop:rangeolimit} and \eqref{eqn:alphaelldef},
\begin{equation}
\lim_{\rho\to 1^-}\alpha(\ell,r_\rho) = \lim_{r_\rho\to 0^+}\alpha(\ell,r_\rho) = \alpha_\ell, \quad \forall \ell \in [\delta+1], \label{eqn:alphalrholimit}
\end{equation}
and thus,
\begin{equation}
\lim_{\rho\to 1^-}\bm{\zeta}_{n,\delta,\rho}(\ell) = \bm{\zeta}_{n,\delta}(\ell), \quad \forall \ell \in [\delta+1]. \label{eqn:alphalimit}
\end{equation}
Note that
\begin{align*}
\left| 2p^{1+\frac{1}{\delta}}P_n(r_\rho) - e_{n,\delta}\right| \leq & \left|2p^{1+\frac{1}{\delta}}P_n(r_\rho)- 2a_n p^{1+\frac{1}{\delta}} r_\rho^{n-2} \right|+ \left| 2a_n p^{1+\frac{1}{\delta}} r_\rho^{n-2}- e_{n,\delta}\right| \\
\leq & 2a_n p^{1+\frac{1}{\delta}} r_\rho^{n-2} \left(\frac{n-4}{8}{r_\rho^2}\right) + \left| 2a_n p^{1+\frac{1}{\delta}} r_\rho^{n-2}- e_{n,\delta}\right|\\
= & 2a_n p^{1+\frac{1}{\delta}} r_\rho^{n-2} \left(\frac{n-4}{4}(1-\rho)\right) + \left| 2a_n p^{1+\frac{1}{\delta}} r_\rho^{n-2}- e_{n,\delta}\right|. \numberthis \label{eqn:slowupperbound}
\end{align*}
where the second inequality follows from Lemma \ref{pderivative} \ref{item:pderivativea}.
\begin{comment}
By Lemma \ref{pderivative} \ref{item:pderivativeb},
$$
\lim_{p\to\infty}2p^{1+\frac{1}{\delta}}P_n(r_\rho)=\lim_{p\to\infty} 2a_n p^{1+\frac{1}{\delta}} r_\rho^{n-2}=\lim_{p\to\infty}2^{\frac{n}{2}}a_n p^{1+\frac{1}{\delta}} (1-\rho)^{\frac{n-2}{2}}=e_{n,\delta}.
$$
\end{comment}
Then the preceding expression and \eqref{eqn:alphalrholimit} yield
\begin{equation}
\lim_{p\to\infty}\lambda_{p,n,\delta,\rho} =\lim_{p\to\infty} \frac{1}{\delta!}p^{\delta+1}(2P_n(r_\rho))^\delta\sum_{\ell=1}^{\delta+1} \frac{\alpha(\ell,r_\rho)}{\ell}=\lim_{p\to\infty} \frac{1}{\delta!}\left(e_{n,\delta}\right)^\delta \sum_{\ell=1}^{\delta+1} \frac{\alpha_\ell}{\ell}=\lambda_{n,\delta}(e_{n,\delta}). \label{eqn:lambdalimit}
\end{equation}
\eqref{eqn:alphalimit} and \eqref{eqn:lambdalimit} immediately yield the conclusion.
\end{proof}
\section{Proofs in Section \ref{sec:convergenceofmoments}}
\subsection{Proofs in Section \ref{sec:convergenceofmoments}}
\subsubsection{Proofs of Lemma \ref{lem:meanedge} and Proposition \ref{lem:2ndmoment}}
\begin{proof}[Proof of Lemma \ref{lem:meanedge}]
\begin{align*}
\mathbb{E} N_{E_{\delta}}^{(\bm{R})}-\binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta}
= & \sum_{ \substack{\vec{i}\in C_{\delta}^<\\ \bm{\Sigma}_{\vec{i}}\text{ not diagonal}} } \left(\mathbb{E} \prod_{j=1}^{\delta} \Phi_{i_0i_j}^{(\bm{R})}-(2P_n(r_\rho))^{\delta}\right) \numberthis \label{eqn:meantemp1} \\
\leq &
\left(\mu_{n,\delta+1}\left(\bm{\Sigma}\right)-1\right) \left( 2P_n(r_\rho)\right)^\delta \sum_{ \substack{\vec{i}\in C_{\delta}^<\\ \bm{\Sigma}_{\vec{i}}\text{ not diagonal}} }1,
\end{align*}
where the first inequality follows from Lemma \ref{lem:densitybou} \ref{item:densitybouc} and Lemma \ref{item:prod}. By \eqref{eqn:meantemp1},
$$
\mathbb{E} N_{E_{\delta}}^{(\bm{R})}-\binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta}\geq - (2P_n(r_\rho))^{\delta} \sum_{ \substack{\vec{i}\in C_{\delta}^<\\ \bm{\Sigma}_{\vec{i}}\text{ not diagonal}} } 1 .
$$
Combining the preceding two expressions,
\begin{align*}
\left| \mathbb{E} N_{E_{\delta}}^{(\bm{R})}-\binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta} \right|
\leq & \max\{1,\mu_{n,\delta+1}\left(\bm{\Sigma}\right)-1)\} \left( 2P_n(r_\rho)\right)^\delta \sum_{ \substack{\vec{i}\in C_{\delta}^<\\ \bm{\Sigma}_{\vec{i}}\text{ not diagonal}} }1\\
\leq &
\mu_{n,\delta+1}\left(\bm{\Sigma}\right) (2P_n(r_\rho))^{\delta} \frac{(\delta+1)}{2((\delta-1)!)} p^{\delta} (\kappa-1)\\
\leq &
\frac{(\delta+1)}{2((\delta-1)!)}\gamma^{\delta} \mu_{n,\delta+1}\left(\bm{\Sigma}\right) \frac{\kappa-1}{p},
\end{align*}
where the second inequality follows from Lemma \ref{lem:numofnondia}.
\end{proof}
\begin{proof}[Proof of Proposition \ref{lem:2ndmoment}]
Recall for $\vec{i}\in C_\delta^<$, $\Phi_{\vec{i}}^{(\bm{R})}$ is defined in \eqref{eqn:NEdeltaRdef}, $U_{\vec{i}}$ is defined in \eqref{eqn:Uidef}, and $Z_{\vec{i}}$ and $W_{\vec{i}}$ are defined in \eqref{eqn:WiZidef}. Then,
$$
N_{E_\delta}^{(\bm{R})} = \sum_{\vec{i}\in C^<_{\delta}} \Phi_{\vec{i}}^{(\bm{R})} = \Phi_{\vec{i}}^{(\bm{R})} + U_{\vec{i}} + Z_{\vec{i}} + W_{\vec{i}}.
$$
Then,
\begin{equation}
\left(N_{E_\delta}^{(\bm{R})}\right)^2 = \sum_{\vec{i}\in C^<_{\delta}} \Phi_{\vec{i}}^{(\bm{R})} \left( \Phi_{\vec{i}}^{(\bm{R})} + U_{\vec{i}} + Z_{\vec{i}} + W_{\vec{i}} \right). \label{eqn:squaredecom}
\end{equation}
\textbf{Step 1}:\\
Since $\Phi_{\vec{i}}^{(\bm{R})} + U_{\vec{i}}$ takes value in $[\delta+1]\bigcup \{0\}$,
\begin{align*}
\mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} \left( \Phi_{\vec{i}}^{(\bm{R})} + U_{\vec{i}}\right) = & \sum_{\ell=1}^{\delta+1} \ell\ \mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} {1} \left( \Phi_{\vec{i}}^{(\bm{R})} + U_{\vec{i}} = \ell \right) \\
=& \sum_{\ell=1}^{\delta+1} \ell\ P \left( \Phi_{\vec{i}}^{(\bm{R})} =1 \right) P \left( \Phi_{\vec{i}}^{(\bm{R})} + U_{\vec{i}} = \ell | \Phi_{\vec{i}}^{(\bm{R})} =1 \right).
\end{align*}
For $\vec{i}\in C^<_\delta$ such that $\bm{\Sigma}_{\vec{i}}$ diagonal, we have $P \left( \Phi_{\vec{i}}^{(\bm{R})} =1 \right) = (2P_n(r_\rho))^{\delta}$ by Lemma \ref{item:prod} and moreover $P \left( \Phi_{\vec{i}}^{(\bm{R})} + U_{\vec{i}} = \ell | \Phi_{\vec{i}}^{(\bm{R})} =1 \right) = \alpha(\ell,r_\rho)$ by \eqref{eqn:alphalrhorgg}. Thus in this case,
$$
\mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} \left( \Phi_{\vec{i}}^{(\bm{R})} + U_{\vec{i}}\right) = (2P_n(r_\rho))^{\delta} \sum_{\ell=1}^{\delta+1} \ell \alpha(\ell,r_\rho).
$$
Moreover, when $\vec{i}\in C^<_\delta$ such that $\bm{\Sigma}_{\vec{i}}$ is not diagonal, by Lemma \ref{lem:densitybou} \ref{item:densitybouc}
$$
\mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} \left( \Phi_{\vec{i}}^{(\bm{R})} + U_{\vec{i}}\right) \leq \mu_{n,\delta+1}(\bm{\Sigma}) (2P_n(r_\rho))^{\delta} \sum_{\ell=1}^{\delta+1} \ell \alpha(\ell,r_\rho).
$$
Then by the preceding two expressions,
\begin{align*}
&\sum_{\vec{i}\in C^<_{\delta}} \mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} \left( \Phi_{\vec{i}}^{(\bm{R})} + U_{\vec{i}}\right) - \binom{p}{1}\binom{p-1}{\delta} (2P_n(r_\rho))^{\delta} \sum_{\ell=1}^{\delta+1} \ell \alpha(\ell,r_\rho)\\
= &
\sum_{\substack{\vec{i}\in C^<_{\delta}\\ \bm{\Sigma}_{\vec{i}} \text{ not diagonal} }} \left( \mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} \left( \Phi_{\vec{i}}^{(\bm{R})} + U_{\vec{i}}\right) - (2P_n(r_\rho))^{\delta} \sum_{\ell=1}^{\delta+1} \ell \alpha(\ell,r_\rho) \right) \numberthis \label{eqn:firsttermtemp} \\
\leq
& \left(\mu_{n,\delta+1}(\bm{\Sigma}) -1 \right) (2P_n(r_\rho))^{\delta} \sum_{\ell=1}^{\delta+1} \ell \alpha(\ell,r_\rho) \sum_{\substack{\vec{i}\in C^<_{\delta}\\ \bm{\Sigma}_{\vec{i}} \text{ not diagonal} }} 1.
\end{align*}
By \eqref{eqn:firsttermtemp},
\begin{align*}
&\sum_{\vec{i}\in C^<_{\delta}} \mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} \left( \Phi_{\vec{i}}^{(\bm{R})} + U_{\vec{i}}\right) - \binom{p}{1}\binom{p-1}{\delta} (2P_n(r_\rho))^{\delta} \sum_{\ell=1}^{\delta+1} \ell \alpha(\ell,r_\rho) \\
\geq &
- (2P_n(r_\rho))^{\delta} \sum_{\ell=1}^{\delta+1} \ell \alpha(\ell,r_\rho) \sum_{\substack{\vec{i}\in C^<_{\delta}\\ \bm{\Sigma}_{\vec{i}} \text{ not diagonal} }} 1.
\end{align*}
By combining the preceding two expressions,
\begin{align*}
&\left|\sum_{\vec{i}\in C^<_{\delta}} \mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} \left( \Phi_{\vec{i}}^{(\bm{R})} + U_{\vec{i}}\right) - \binom{p}{1}\binom{p-1}{\delta} (2P_n(r_\rho))^{\delta}\sum_{\ell=1}^{\delta+1} \ell \alpha(\ell,r_\rho)\right| \\
\leq &
\mu_{n,\delta+1}(\bm{\Sigma}) (2P_n(r_\rho))^{\delta} \sum_{\ell=1}^{\delta+1} \ell \alpha(\ell,r_\rho) \sum_{\substack{\vec{i}\in C^<_{\delta}\\ \bm{\Sigma}_{\vec{i}} \text{ not diagonal} }} 1 \\
\leq & \mu_{n,\delta+1}(\bm{\Sigma}) (2P_n(r_\rho))^{\delta} \sum_{\ell=1}^{\delta+1} \ell \alpha(\ell,r_\rho) \frac{(\delta+1)}{2((\delta-1)!)} p^{\delta} (\kappa-1) \\
\leq & \mu_{n,\delta+1}(\bm{\Sigma}) (2p^{1+1/\delta}P_n(r_\rho))^{\delta} \frac{(\delta+1)^2}{2((\delta-1)!)} \frac{\kappa-1}{p}
\numberthis \label{eqn:2ndmomentstep1}
\end{align*}
where the second inequality follows from Lemma \ref{lem:numofnondia}.\\
\noindent
\textbf{Step 2}:\\
\begin{align*}
&\sum_{\vec{i}\in C^<_{\delta}} \mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} W_{\vec{i}} - \left(\binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta}\right)^2 \\
= &
\sum_{\vec{i}\in C^<_{\delta}} \sum_{\vec{j}\in T_{\vec{i}}} \mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} \Phi_{\vec{j}}^{(\bm{R})}-\sum_{\vec{i}\in C^<_{\delta}} \sum_{\vec{j}\in C^<_{\delta}} (2P_n(r_\rho))^{2\delta}\\
= &
\sum_{\vec{i}\in C^<_{\delta}} \sum_{\vec{j}\in T_{\vec{i}}} \mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} \mathbb{E}\Phi_{\vec{j}}^{(\bm{R})} - \sum_{\vec{i}\in C^<_{\delta}} \sum_{\vec{j}\in T_{\vec{i}}} (2P_n(r_\rho))^{2\delta} -\sum_{\vec{i}\in C^<_{\delta}} \sum_{\vec{j}\in C^<_{\delta}\backslash T_{\vec{i}}} (2P_n(r_\rho))^{2\delta}\\
= &
-\sum_{\vec{i}\in C^<_{\delta}} \sum_{\vec{j}\in C^<_{\delta}\backslash T_{\vec{i}}} (2P_n(r_\rho))^{2\delta}
\sum_{\substack{\vec{i}\in C^<_{\delta}\\ \bm{\Sigma}_{\vec{i}} \text{ not diagonal} }} \sum_{\vec{j}\in T_{\vec{i}}} \left(\mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} \Phi_{\vec{j}}^{(\bm{R})}
- (2P_n(r_\rho))^{2\delta}\right) \\
& +\sum_{\substack{\vec{i}\in C^<_{\delta}\\ \bm{\Sigma}_{\vec{i}} \text{ diagonal} }} \ \sum_{ \substack{\vec{j}\in T_{\vec{i}}\\ \bm{\Sigma}_{\vec{j}} \text{ not diagonal} }} \left(\mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} \Phi_{\vec{j}}^{(\bm{R})} - (2P_n(r_\rho))^{2\delta}\right),
\numberthis \label{eqn:3rdtermtemp}
\end{align*}
where the last equality follows from $\mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} =(2P_n(r_\rho))^{\delta}$ for $\vec{i}\in \left\{ \vec{j}\in C_\delta^<: \bm{\Sigma}_{\vec{j}} \text{ diagonal} \right\}$ by Lemma \ref{item:prod}. Then by \eqref{eqn:3rdtermtemp},
\begin{align*}
&\sum_{\vec{i}\in C^<_{\delta}} \mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} W_{\vec{i}} - \left(\binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta}\right)^2\\
\leq & \left(\mu_{n,2\delta+2}(\bm{\Sigma}) - 1\right)(2P_n(r_\rho))^{2\delta}
\left(\sum_{\substack{\vec{i}\in C^<_{\delta}\\ \bm{\Sigma}_{\vec{i}} \text{ not diagonal} }} \sum_{\vec{j}\in T_{\vec{i}}} 1 +\sum_{\substack{\vec{i}\in C^<_{\delta}\\ \bm{\Sigma}_{\vec{i}} \text{ diagonal} }} \ \sum_{ \substack{\vec{j}\in T_{\vec{i}}\\ \bm{\Sigma}_{\vec{j}} \text{ not diagonal} }} 1\right),
\end{align*}
where the inequality follows from Lemma \ref{lem:densitybou} \ref{item:densitybouc}. On the other hand, by \eqref{eqn:3rdtermtemp},
\begin{align*}
& \sum_{\vec{i}\in C^<_{\delta}} \mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} W_{\vec{i}} - \left(\binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta}\right)^2 \\
\geq & - (2P_n(r_\rho))^{2\delta} \left(\sum_{\substack{\vec{i}\in C^<_{\delta}\\ \bm{\Sigma}_{\vec{i}} \text{ not diagonal} }} \sum_{\vec{j}\in T_{\vec{i}}} 1 +\sum_{\substack{\vec{i}\in C^<_{\delta}\\ \bm{\Sigma}_{\vec{i}} \text{ diagonal} }} \ \sum_{ \substack{\vec{j}\in T_{\vec{i}}\\ \bm{\Sigma}_{\vec{j}} \text{ not diagonal} }} 1 \right) -\sum_{\vec{i}\in C^<_{\delta}} \sum_{\vec{j}\in C^<_{\delta}\backslash T_{\vec{i}}} (2P_n(r_\rho))^{2\delta}.
\end{align*}
Combining the preceding two displays,
\begin{align*}
& \left|\sum_{\vec{i}\in C^<_{\delta}} \mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} W_{\vec{i}} - \left(\binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta}\right)^2\right| \\
\leq &
\mu_{n,2\delta+2}(\bm{\Sigma}) (2P_n(r_\rho))^{2\delta}
\left(\sum_{\substack{\vec{i}\in C^<_{\delta}\\ \bm{\Sigma}_{\vec{i}} \text{ not diagonal} }} \sum_{\vec{j}\in T_{\vec{i}}} 1 +\sum_{\substack{\vec{i}\in C^<_{\delta}\\ \bm{\Sigma}_{\vec{i}} \text{ diagonal} }} \ \sum_{ \substack{\vec{j}\in T_{\vec{i}}\\ \bm{\Sigma}_{\vec{j}} \text{ not diagonal} }} 1\right) \\
&\quad \quad \quad + \sum_{\vec{i}\in C^<_{\delta}} \sum_{\vec{j}\in C^<_\delta\backslash T_{\vec{i}}} (2P_n(r_\rho))^{2\delta} \\
\leq &
\mu_{n,2\delta+2}(\bm{\Sigma}) (2P_n(r_\rho))^{2\delta} \left(2\sum_{\substack{\vec{i}\in C^<_{\delta}\\ \bm{\Sigma}_{\vec{i}} \text{ not diagonal} }} \sum_{\vec{j}\in C_\delta^<} 1\right) + \sum_{\vec{i}\in C^<_{\delta}} \sum_{\vec{j}\in C^<_\delta\backslash T_{\vec{i}}} (2P_n(r_\rho))^{2\delta}\\
\overset{(*)}{\leq} &
\mu_{n,2\delta+2}(\bm{\Sigma}) (2P_n(r_\rho))^{2\delta} 2\binom{p}{1}\binom{p-1}{\delta} \frac{(\delta+1)}{2((\delta-1)!)} p^{\delta} (\kappa-1) \\
&\quad \quad + \binom{p}{1}\binom{p-1}{\delta} \frac{(\delta+1)^2}{\delta !} p^{\delta}\kappa (2P_n(r_\rho))^{2\delta}\\
\leq&
\frac{2(\delta+1)^2}{(\delta!)^2} (2p^{1+1/\delta}P_n(r_\rho))^{2\delta} \mu_{n,2\delta+2}(\bm{\Sigma})\frac{\kappa}{p}, \numberthis \label{eqn:2ndmomentstep2}
\end{align*}
where step $(*)$ follows from Lemma \ref{lem:numofnondia} and \eqref{eqn:neiborsize}.
\noindent
\textbf{Step 3}:\\
Notice $\sum_{\vec{i}\in C^<_{\delta}} \mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} Z_{\vec{i}} = b_2$ as in \eqref{eqn:b2def} and thus satisfies the bound \eqref{eqn:b2upperbou2new}. Then by \eqref{eqn:squaredecom},
\begin{align*}
&\left| \mathbb{E} \left(N_{E_{\delta}}^{(\bm{R})}\right)^2-\binom{p}{1}\binom{p-1}{\delta} (2P_n(r_\rho))^{\delta}\sum_{\ell=1}^{\delta+1} \ell \alpha(\ell,r_\rho) -\left(\binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta}\right)^2 \right| \\
\leq & \left|\sum_{\vec{i}\in C^<_{\delta}} \mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} \left( \Phi_{\vec{i}}^{(\bm{R})} + U_{\vec{i}}\right) - \binom{p}{1}\binom{p-1}{\delta} (2P_n(r_\rho))^{\delta} \sum_{\ell=1}^{\delta+1} \ell \alpha(\ell,r_\rho) \right|\\
&+ \left|\sum_{\vec{i}\in C^<_{\delta}} \mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} W_{\vec{i}} - \left(\binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta}\right)^2\right| + b_2\\
\leq & C_{n,\delta} \left( (2p^{1+1/\delta}P_n(r_\rho))^{\delta} \left(1+2p^{1+1/\delta}P_n(r_\rho)\right)^{\delta} \mu_{n,2\delta+2}(\bm{\Sigma})\frac{\kappa}{p}\right.\\
&\left. \quad + p(2pP_n(r_\rho))^{\delta+1} \left(1+2pP_n(r_\rho)\right)^{\delta}\right), \numberthis \label{eqn:2ndmomentdelta}
\end{align*}
where the last inequality follows from \eqref{eqn:2ndmomentstep1}, \eqref{eqn:2ndmomentstep2} and \eqref{eqn:b2upperbou2new}. The proof is then completed by using the inequality $2p^{1+1/\delta}P_n(r_\rho)\leq \gamma$.
\end{proof}
\subsubsection{Proof of Proposition \ref{prop:allcloseL2}}
\begin{proof}[Proof of Proposition \ref{prop:allcloseL2} \ref{item:allcloseL2a}]\ \\
By taking the square of each of the terms in Lemma \ref{lem:6quantitiesine},
$$
\left(N_{E_\delta}^{(\bm{R})}\right)^2 - 2(\delta+1)N_{E_\delta}^{(\bm{R})} N_{E_{\delta+1}}^{(\bm{R})} \leq \left(N_{\breve{V}_\delta}^{(\bm{R})}\right)^2 \leq \left(N_{V_\delta}^{(\bm{R})}\right)^2 \leq \left(N_{E_\delta}^{(\bm{R})}\right)^2,
$$
which then implies for $\bar{N}_\delta \in \left\{ N_{\breve{V}_\delta}^{(\bm{R})}, N_{V_\delta}^{(\bm{R})} \right\} $
\begin{equation}
\left|\left(\bar{N}_{\delta}\right)^2-\left(N_{E_\delta}^{(\bm{R})}\right)^2\right| \leq 2(\delta+1)N_{E_\delta}^{(\bm{R})} N_{E_{\delta+1}}^{(\bm{R})} = 2(\delta+1)\sum_{\vec{i}\in C_{\delta+1}^<} \sum_{\vec{j}\in C_{\delta}^<} \Phi_{\vec{i}}^{(\bm{R})} \Phi_{\vec{j}}^{(\bm{R})}. \label{eqn:squarecorrelationuppbou}
\end{equation}
It suffices to establish an upper bound on $\mathbb{E} N_{E_{\delta+1}}^{(\bm{R})} N_{E_\delta}^{(\bm{R})} = \mathbb{E} \sum_{\vec{i}\in C_{\delta+1}^<} \sum_{\vec{j}\in C_{\delta}^<} \Phi_{\vec{i}}^{(\bm{R})} \Phi_{\vec{j}}^{(\bm{R})} $.
Observe for $\vec{j}\in J_{\vec{i}}$,
$$
\mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} \Phi_{\vec{j}}^{(\bm{R})} \leq \mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} \leq \mu_{n,\delta+2}(\bm{\Sigma}_{\vec{i}}) (2P_n(r_\rho))^{\delta+1}.
$$
For $\vec{j} \in T_{\vec{i}}$, $\left[\vec{j}\right]\cap \left[\vec{i}\right]=\emptyset$. Thus, if $\bm{\Sigma}_{\vec{i}\cup \vec{j}}$ is diagonal, $\mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} \Phi_{\vec{j}}^{(\bm{R})} = \mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} \mathbb{E} \Phi_{\vec{j}}^{(\bm{R})} =(2P_n(r_\rho))^{2\delta+1}$ by Lemma \ref{item:prod}. Then, for the general case when $\bm{\Sigma}_{\vec{i}\cup \vec{j}}$ is not necessarily diagonal, by Lemma \ref{lem:densitybou} \ref{item:densitybouc},
$$
\mathbb{E} \Phi_{\vec{i}}^{(\bm{R})} \Phi_{\vec{j}}^{(\bm{R})} \leq \mu_{n,2\delta+3}(\bm{\Sigma}_{\vec{i}\cup\vec{j}}) (2P_n(r_\rho))^{2\delta+1}
$$
By Lemma \ref{lem:densitybou} \ref{item:densitybouc}, Lemma \ref{item:prod} and Lemma \ref{lem:eleine}, it is straightforward that the conditions in Lemma \ref{lem:b2} with $q=\delta+1$ and $\theta_{\vec{i},\vec{j}}=\Phi_{\vec{i}}^{(\bm{R})}\Phi_{\vec{j}}^{(\bm{R})}$ are satisfied with $a=1$, $b=2P_n(2r_\rho)1(\delta\geq 2)+2P_n(r_\rho)1(\delta= 1)$ and $z=2P_n(r_\rho)$. Moreover, $b/z\leq 2^{n-2}1(\delta\geq 2)+1$ by Lemma \ref{pderivative} \ref{item:pderivatived}.
Thus, Lemma \ref{lem:b2wholesum} with $q=\delta+1$ and $\theta_{\vec{i},\vec{j}}=\Phi_{\vec{i}}^{(\bm{R})}\Phi_{\vec{j}}^{(\bm{R})}$, $a=1$, $b=2P_n(2r_\rho)1(\delta\geq 2)+2P_n(r_\rho)1(\delta= 1)$ and $z=2P_n(r_\rho)$, together with the fact that $pz\leq p^{1+\frac{1}{\delta}}z = 2p^{1+\frac{1}{\delta}}P_n(r_\rho) \leq \gamma$,
yield
\begin{align}
\mathbb{E} N_{E_{\delta+1}}^{(\bm{R})} N_{E_\delta}^{(\bm{R})} = \mathbb{E} \sum_{\vec{i}\in C_{\delta+1}^<} \sum_{\vec{j}\in C_{\delta}^<} \Phi_{\vec{i}}^{(\bm{R})} \Phi_{\vec{j}}^{(\bm{R})}
\leq C_{n,\delta,\gamma} \left(1+\mu_{n,2\delta+3}(\bm{\Sigma})\frac{\kappa-1}{p}\right)p^{-1/\delta}.
\label{eqn:NEdelta1NEdelta}
\end{align}
The proof is then complete by the preceding expression and \eqref{eqn:squarecorrelationuppbou}.
\end{proof}
We now present a few lemmas that are used in the proof of Proposition \ref{prop:allcloseL2} \ref{item:allcloseL2b} and \ref{item:allcloseL2c}. Recall that $F_{ij}(r_\rho),H_{ij}(r_\rho), G_{ij}(r_\rho), F_{\vec{i}}(r_\rho)$ are defined in Section \ref{sec:proofthmallclosec}.
\begin{lem}
\label{lem:inddifbou2ndmoment}
Suppose $p\geq n$. $1\leq \delta\leq q \leq p-1$. Then for any $\vec{i}\in C_q^<$, $\vec{j}\in C_{\delta}^<$, with probability $1$,
$$
\left|\Phi_{\vec{i}}^{(\bm{P})}\Phi_{\vec{j}}^{(\bm{P})}-\Phi_{\vec{i}}^{(\bm{R})}\Phi_{\vec{j}}^{(\bm{R})}\right|\leq \xi_{\vec{i},\vec{j}},
$$
where
$$
\xi_{\vec{i},\vec{j}} := 1\left( \bigcup\limits_{m=1}^q \left ( \left( H_{i_0i_m} \big \backslash G_{i_0i_m} \right) \bigcap H_{\vec{i},-m} \bigcap H_{\vec{j}} \right ) \bigcup \bigcup\limits_{\ell=1}^{\delta} \left ( \left( H_{j_0j_{\ell}} \big \backslash G_{j_0j_{\ell}} \right) \bigcap H_{\vec{j},-\ell} \bigcap H_{\vec{i}} \right ) \right).
$$
\end{lem}
\begin{proof}
\begin{align*}
\left|\Phi_{\vec{i}}^{(\bm{P})}\Phi_{\vec{j}}^{(\bm{P})}-\Phi_{\vec{i}}^{(\bm{R})}\Phi_{\vec{j}}^{(\bm{R})}\right| =& 1\left(\left(\bigcap\limits_{m=1}^q F_{i_0i_m} \bigcap \bigcap\limits_{\ell=1}^{\delta} F_{j_0j_\ell} \right)\bigtriangleup \left(\bigcap\limits_{m=1}^q S_{i_0i_m} \bigcap \bigcap\limits_{\ell=1}^{\delta} S_{j_0j_\ell} \right) \right)\leq \xi_{\vec{i},\vec{j}}
\end{align*}
where the inequality follows from \eqref{eqn:FGJSsetinclusion} and Lemma \ref{setrelation} \ref{prop:setinclusiona}.
\end{proof}
\begin{lem}
\label{lem:xietarelationship}
Let $p\geq n\geq 4$, $1\leq \delta\leq q \leq p-1$ and $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$. Suppose $\bm{\Sigma}$, after some row-column permutation, is $(\tau,\kappa)$ sparse with $\tau\leq \frac{p}{2}$. Let $t$ be any positive number, and suppose \eqref{eqn:postempass} holds. Then for any $\vec{i}\in C_q^<$, $\vec{j}\in C_{\delta}^<$, with probability $1$,
$$
\xi_{\vec{i},\vec{j}} 1\left(\mathcal{E}(t)\right) \leq \eta_{\vec{i},\vec{j}}(t),
$$
where
\begin{align*}
\eta_{\vec{i},\vec{j}}(t) := &
1 \left( \bigcup_{m=1}^q \left( \left(F_{i_0i_m}(\theta_1(t)r_\rho)\big \backslash F_{i_0i_m}\left(\frac{r_\rho}{\theta_1(t)}\right) \right) \bigcap F_{\vec{i},-m}(\theta_1(t)r_\rho) \bigcap F_{\vec{j}}(\theta_1(t)r_\rho) \right) \bigcup \right.\\
&\left. \quad \bigcup_{\ell=1}^\delta \left( \left(F_{j_0j_\ell}(\theta_1(t)r_\rho)\big \backslash F_{j_0j_\ell}\left(\frac{r_\rho}{\theta_1(t)}\right) \right) \bigcap F_{\vec{j},-\ell}(\theta_1(t)r_\rho) \bigcap F_{\vec{i}}(\theta_1(t)r_\rho) \right)
\right).
\end{align*}
\end{lem}
\begin{proof}
By \eqref{ctempdef}, $H_{ij}(r_\rho) \cap \mathcal{E}(t)\subset F_{ij}(\theta_1(t)r_\rho)$ and $G_{ij}(r_\rho) \cap \mathcal{E}(t)\supset F_{ij}\left(\frac{r_\rho}{\theta_1(t)}\right)$. Then
$$
\xi_{\vec{i},\vec{j}} 1\left(\mathcal{E}(t)\right)
\leq \eta_{\vec{i},\vec{j}}(t).
$$
\end{proof}
\begin{lem}
\label{lem:etadoublesumexpectation}
Let $p\geq n\geq 4$, $1\leq \delta\leq q \leq p-1$ and $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$. Suppose $\bm{\Sigma}$, after some row-column permutation, is $(\tau,\kappa)$ sparse with $\tau\leq \frac{p}{2}$. Let $t$ be any positive number, and suppose \eqref{eqn:postempass} holds. Suppose additionally $2p^{1+\frac{1}{\delta}}P_n(r_\rho)\leq \gamma$. Then
\begin{align*}
&\mathbb{E} \sum_{\vec{i}\in C_q^<} \sum_{\vec{j}\in C_\delta^<} \eta_{\vec{i},\vec{j}}(t) \\
\leq &
C_{n,q,\delta,\gamma} \left(\theta_1(t)\right)^{n(2\delta+q)} \left(1+\mu_{n,q+\delta+2}(\bm{\Sigma})\frac{\kappa-1}{p}\right) \left(\sqrt{\frac{1}{p}}+\frac{t}{\sqrt{p}} + \frac{\tau}{p}\right) p^{1-\frac{q}{\delta}}.
\end{align*}
\end{lem}
\begin{proof}
Note
\begin{align*}
\sum_{\vec{i}\in C_q^<} \sum_{\vec{j}\in C_\delta^<} \eta_{\vec{i},\vec{j}}(t) = \left(\sum_{\vec{i}\in C_q^<} \sum_{\vec{j}\in J_{\vec{i}}} + \sum_{\vec{i}\in C_q^<} \sum_{\vec{j}\in T_{\vec{i}}} +\sum_{\vec{i}\in C_q^<} \sum_{\vec{j}\in N_{\vec{i}}}\right)
\eta_{\vec{i},\vec{j}}(t).
\end{align*}
\noindent
\textbf{Step 1: $\vec{j}\in T_{\vec{i}}$}\\
By the union bound for indicator functions,
\begin{equation}
\eta_{\vec{i},\vec{j}}(t) \leq \eta_{\vec{i}}(t) 1\left(F_{\vec{j}}(\theta_1(t)r_\rho)\right) + \eta_{\vec{j}}(t) 1\left(F_{\vec{i}}(\theta_1(t)r_\rho)\right), \label{eqn:etaunionbound}
\end{equation}
where $\eta_{\vec{i}}(t)$ is defined in \eqref{eqn:etavecidef} with $\delta$ replaced by $q$. Then for $\vec{j}\in T_{\vec{i}}$
\begin{equation}
\mathbb{E}\eta_{\vec{i},\vec{j}}(t) \leq \mathbb{E}\eta_{\vec{i}}(t) P\left(F_{\vec{j}}(\theta_1(t)r_\rho)\right) + \mathbb{E}\eta_{\vec{j}}(t) P\left(F_{\vec{i}}(\theta_1(t)r_\rho)\right). \label{eqn:etadoubleindexuppbouTi}
\end{equation}
Moreover, for $\vec{j}\in T_{\vec{i}}$, $\bm{\Sigma}_{\vec{i}\cup \vec{j} }$ is diagonal if and only if $\bm{\Sigma}_{\vec{i}}$ and $\bm{\Sigma}_{\vec{j}}$ are both diagonal.
Now suppose $\bm{\Sigma}_{\vec{i}\cup \vec{j} }$ is diagonal. By conditioning on $\bm{u}_{j_0}$
$$
P\left(F_{\vec{j}}(\theta_1(t)r_\rho)\right) = \left(2P_n(\theta_1(t)r_\rho)\right)^\delta, \quad P\left(F_{\vec{i}}(\theta_1(t)r_\rho)\right) = \left(2P_n(\theta_1(t)r_\rho)\right)^q.
$$
The preceding expression, \eqref{eqn:etadoubleindexuppbouTi}, Lemma \ref{lem:inddifuppexpbou} applied to $\mathbb{E}\eta_{\vec{j}}(t)$, and Lemma \ref{lem:inddifuppexpbou} with $\delta=q$ applied to $\mathbb{E}\eta_{\vec{i}}(t)$ yield
$$
\mathbb{E} \eta_{\vec{i},\vec{j}}(t) \leq (\delta+q) 2\left(P_n(r_\rho \theta_1(t))-P_n\left(\frac{r_\rho}{\theta_1(t)} \right)\right) \left(2P_n\left(\theta_1(t)r_\rho\right)\right)^{q+\delta-1}.
$$
For the general case that $\bm{\Sigma}_{\vec{i}\cup \vec{j} }$ is not necessarily diagonal, by Lemma \ref{lem:densitybou} \ref{item:densitybouc}, for any $\vec{j}\in T_{\vec{i}}$
$$
\mathbb{E} \eta_{\vec{i},\vec{j}}(t) \leq \mu_{n,q+\delta+2}(\bm{\Sigma}_{\vec{i}\cup\vec{j}}) (\delta+q) 2\left(P_n(r_\rho \theta_1(t))-P_n\left(\frac{r_\rho}{\theta_1(t)} \right)\right) \left(2P_n\left(\theta_1(t)r_\rho\right)\right)^{q+\delta-1}.
$$
Then the condition in Lemma \ref{lem:quadraticsum} \ref{item:quadraticsumb} is satisfied with $\theta_{\vec{i},\vec{j}}=\eta_{\vec{i},\vec{j}}$, $z=2P_n(r_\rho \theta_1(t))$ and $a = (\delta+q) \frac{P_n(r_\rho \theta_1(t))-P_n\left(\frac{r_\rho}{\theta_1(t)} \right)}{ P_n(r_\rho \theta_1(t)) } $.\\
\noindent
\textbf{Step 2: $\vec{j}\in J_{\vec{i}}$}\\
\eqref{eqn:etaunionbound} implies
\begin{align}
\eta_{\vec{i},\vec{j}}(t)\leq & \eta_{\vec{i}}(t)+ \sum_{\ell=1}^{\delta} 1\left(F_{j_0j_\ell}(\theta_1(t)r_\rho)\big \backslash F_{j_0j_\ell}\left(\frac{r_\rho}{\theta_1(t)}\right) \right) 1(F_{\vec{i}}(\theta_1(t)r_\rho)). \label{eqn:step1etadoubleindexuppbou}
\end{align}
For $\vec{j}\in J_{\vec{i}}$, $j_0,j_\ell\in \left[ \vec{i} \right]$. If $i_0\in \{j_0,j_\ell\}$, without loss of generality, say $i_0=j_\ell$ and $j_0=i_\alpha$ for some $1\leq \alpha\leq q$. Then
\begin{align*}
&\mathbb{E} 1\left(F_{j_0j_\ell}(\theta_1(t)r_\rho)\big \backslash F_{j_0j_\ell}\left(\frac{r_\rho}{\theta_1(t)}\right) \right) 1(F_{\vec{i}}(\theta_1(t)r_\rho)) \\
= &
P \left( \left(F_{i_0i_\alpha}(\theta_1(t)r_\rho)\big \backslash F_{i_0i_\alpha}\left(\frac{r_\rho}{\theta_1(t)}\right) \right) \bigcap F_{\vec{i},-\alpha}(\theta_1(t)r_\rho) \right)\\
\leq &
\mu_{n,q+1}(\bm{\Sigma}_{\vec{i}}) 2\left(P_n(r_\rho \theta_1(t))-P_n\left(\frac{r_\rho}{\theta_1(t)} \right)\right) \left(2P_n\left(\theta_1(t)r_\rho\right)\right)^{q-1}, \numberthis \label{eqn:etadoubleindexcase1}
\end{align*}
where the last step follows from Lemma \ref{lem:inddifuppexpbou} with $\delta$ replace by $q$.
If $i_0\not \in \{j_0,j_\ell\}$, without loss of generality, let $j_0=i_\alpha$, $j_\ell=i_{\beta}$ for some $1\leq \alpha \neq \beta \leq q$. Suppose for now that $\bm{\Sigma}_{\vec{i}}$ is diagonal, and then
\begin{align*}
&\mathbb{E} 1\left(F_{j_0j_\ell}(\theta_1(t)r_\rho)\big \backslash F_{j_0j_\ell}\left(\frac{r_\rho}{\theta_1(t)}\right) \right) 1(F_{\vec{i}}(\theta_1(t)r_\rho)) \\
= &
\mathbb{E} 1\left(F_{i_\alpha i_\beta}(\theta_1(t)r_\rho)\big \backslash F_{i_\alpha i_\beta}\left(\frac{r_\rho}{\theta_1(t)}\right) \right) 1(F_{\vec{i}}(\theta_1(t)r_\rho)) \\
\overset{(*)}{=} &
(2P_n(\theta_1(t)r_\rho))^{q-2} \mathbb{E} 1\left(F_{i_\alpha i_\beta}(\theta_1(t)r_\rho)\setminus F_{i_\alpha i_\beta} \left(\frac{r_\rho}{\theta_1(t)}\right) \right) 1(F_{i_0i_\alpha}(\theta_1(t)r_\rho))1(F_{i_0i_\beta}(\theta_1(t)r_\rho))
\\
\overset{(**)}{\leq} &
2\left (P_n(\theta_1(t)r_\rho)-P_n\left(\frac{r_\rho}{\theta_1(t)}\right)\right) (2P_n(\theta_1(t)r_\rho))^{q-1} ,
\end{align*}
where the step $(*)$ follows from conditioning on $\bm{u}_{i_0},\bm{u}_{i_\alpha},\bm{u}_{i_\beta}$, and the step $(**)$ follows from dropping the term $1(F_{i_0i_\beta}(\theta_1(t)r_\rho))$ and then conditioning on $\bm{u}_{i_\alpha}$. For the general case that $\bm{\Sigma}_{\vec{i}}$ is not necessarily diagonal, by Lemma \ref{lem:densitybou} \ref{item:densitybouc},
\begin{align*}
&\mathbb{E} 1\left(F_{j_0j_\ell}(\theta_1(t)r_\rho)\big \backslash F_{j_0j_\ell}\left(\frac{r_\rho}{\theta_1(t)}\right) \right) 1(F_{\vec{i}}(\theta_1(t)r_\rho)) \\
\leq &
\mu_{n,q+1}(\bm{\Sigma}_{\vec{i}}) 2\left (P_n(\theta_1(t)r_\rho)-P_n\left(\frac{r_\rho}{\theta_1(t)}\right)\right) (2P_n(\theta_1(t)r_\rho))^{q-1}. \numberthis \label{eqn:etadoubleindextemp4}
\end{align*}
By combining \eqref{eqn:step1etadoubleindexuppbou}, \eqref{eqn:etadoubleindexcase1}, \eqref{eqn:etadoubleindextemp4} and Lemma \ref{lem:inddifuppexpbou} with $\delta$ replace by $q$,
$$
\mathbb{E} \eta_{\vec{i},\vec{j}}(t) \leq (q+\delta)\mu_{n,q+1}(\bm{\Sigma}_{\vec{i}}) 2\left (P_n(\theta_1(t)r_\rho)-P_n\left(\frac{r_\rho}{\theta_1(t)}\right)\right) (2P_n(\theta_1(t)r_\rho))^{q-1}.
$$
Then the conditions in Lemma \ref{lem:quadraticsum} \ref{item:quadraticsuma} with $\theta_{\vec{i},\vec{j}}=\eta_{\vec{i},\vec{j}}$, $z=2P_n(r_\rho\theta_1(t))$ and $a=(q+\delta) \frac{P_n(\theta_1(t)r_\rho)-P_n\left(\frac{r_\rho}{\theta_1(t)}\right)}{P_n(\theta_1(t)r_\rho)}$ is satisfied.\\
\noindent
\textbf{Step 3: $\vec{j}\in N_{\vec{i}}$}\\
It is straightforward by Lemma \ref{lem:densitybou} \ref{item:densitybouc}, Lemma \ref{item:prod} and Lemma \ref{lem:eleine} that the conditions in Lemma \ref{lem:b2} with $\theta_{\vec{i},\vec{j}}=\eta_{\vec{i},\vec{j}}$ are satisfied with $a=a_1=(q+\delta) \frac{P_n(\theta_1(t)r_\rho)-P_n\left(\frac{r_\rho}{\theta_1(t)}\right)}{P_n(\theta_1(t)r_\rho)}$, $b=2P_n(2r_\rho\theta_1(t))1(\delta\geq 2)+2P_n(r_\rho\theta_1(t))1(\delta= 1)$ and $z=2P_n(r_\rho\theta_1(t))$. Moreover, $b/z\leq 2^{n-2}1(\delta\geq 2)+1$ by Lemma \ref{pderivative} \ref{item:pderivatived}.
By Lemma \ref{lem:b2wholesum} with $\theta_{\vec{i},\vec{j}}=\eta_{\vec{i},\vec{j}}$, $a=(q+\delta) \frac{P_n(\theta_1(t)r_\rho)-P_n\left(\frac{r_\rho}{\theta_1(t)}\right)}{P_n(\theta_1(t)r_\rho)}$, $b=2P_n(2r_\rho\theta_1(t))1(\delta\geq 2)+2P_n(r_\rho\theta_1(t))1(\delta= 1)$ and $z=2P_n(r_\rho\theta_1(t))$
\begin{align*}
&\sum_{\vec{i}\in C_q^<} \sum_{\vec{j}\in C_\delta^<} \mathbb{E} \eta_{\vec{i},\vec{j}} \\
\leq &
C_{n,q,\delta} \left(p^{1+\frac{1}{\delta}}z\right)^q \left(1+(p^{1+\frac{1}{\delta}}z)^{\delta}\right)\left(1+pz\right)^{\delta-1} \left(1+\mu_{n,q+\delta+2}(\bm{\Sigma})\frac{\kappa-1}{p}\right) ap^{1-\frac{q}{p}}. \numberthis \label{eqn:etadoublesumtemp}
\end{align*}
\noindent
\textbf{Step 4}\\
Observe that
\begin{align}
pz \leq p^{1+\frac{1}{\delta}} z \leq \left(\theta_1(t)\right)^{n-2} 2p^{1+\frac{1}{\delta}}P_n(r_\rho)
\leq \left(\theta_1(t)\right)^{n-2} \gamma \label{eqn:pzbou}
\end{align}
where the second inequality follows from Lemma \ref{pderivative} \ref{item:pderivatived}. Moreover, by Lemma \ref{pderivative} \ref{item:pderivativec} and the fact that $\theta_1(t)\geq 1$,
\begin{equation}
a\leq (q+\delta) \frac{P_n(\theta_1(t)r_\rho)-P_n\left(\frac{r_\rho}{\theta_1(t)}\right)}{P_n(r_\rho)} \leq (q+\delta)(n-2)\left(\theta_1(t)\right)^{n-3}\left(\theta_1(t)-\frac{1}{\theta_1(t)}\right). \label{eqn:aupperbound}
\end{equation}
Plugging \eqref{eqn:pzbou} and \eqref{eqn:aupperbound} into \eqref{eqn:etadoublesumtemp} and by the fact that $\theta_1(t)\geq 1$,
\begin{align*}
\sum_{\vec{i}\in C_q^<} \sum_{\vec{j}\in C_\delta^<} \eta_{\vec{i},\vec{j}}
\leq &
C_{n,q,\delta,\gamma} \left(\theta_1(t)\right)^{n(2\delta+q)} \left(1+\mu_{n,q+\delta+2}(\bm{\Sigma})\frac{\kappa-1}{p}\right) \left(\theta_1(t)-\frac{1}{\theta_1(t)}\right)p^{1-\frac{q}{p}}\\
\leq & C_{n,q,\delta,\gamma} \left(\theta_1(t)\right)^{n(2\delta+q)} \left(1+\mu_{n,q+\delta+2}(\bm{\Sigma})\frac{\kappa-1}{p}\right) \left(\sqrt{\frac{n}{p}}+\frac{t}{\sqrt{p}} + n\frac{\tau}{p}\right) p^{1-\frac{q}{p}},
\end{align*}
where the last inequality follows from Lemma \ref{basine} \ref{item:basinek} and \eqref{ctempdef}.
\end{proof}
\begin{lem}
\label{lem:NEqNEdelta}
Let $p\geq n \geq 4$ and $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$. Suppose $\bm{\Sigma}$, after some row-column permutation, is $(\tau,\kappa)$ sparse with $\tau\leq \frac{p}{2}$. Consider $1\leq \delta \leq p-2$ and let $q\in \{\delta,\delta+1\}$. Suppose $2p^{1+\frac{1}{\delta}}P_n(r_\rho)\leq \gamma$ and $\left(\sqrt{\frac{n-1}{p}}+\sqrt{\frac{\delta\ln p}{p}}\right) \leq c$ hold for some positive and \ONE{sufficiently} small universal constant $c$. Then
\begin{align*}
\mathbb{E} \left|N_{E_q}^{(\bm{P})} N_{E_\delta}^{(\bm{P})}-N_{E_q}^{(\bm{R})} N_{E_\delta}^{(\bm{R})}\right|\leq C_{n,\delta,\gamma} \left(1+\mu_{n,q+\delta+2}(\bm{\Sigma})\frac{\kappa-1}{p}\right) \left(\frac{\sqrt{\ln p}}{\sqrt{p}} + \frac{\tau}{p}\right) p^{1-\frac{q}{\delta}}.
\end{align*}
\end{lem}
\begin{proof}
For $\bm{\Psi} \in \{\bm{R},\bm{P}\}$,
$$
N_{E_q}^{(\bm{\Psi})} N_{E_\delta}^{(\bm{\Psi})} = \sum_{\vec{i}\in C_q^<} \sum_{\vec{j}\in C_\delta^<} \Phi_{\vec{i}}^{(\bm{\Psi})} \Phi_{\vec{j}}^{(\bm{\Psi})}.
$$
Thus,
\begin{align*}
&\mathbb{E} \left|N_{E_q}^{(\bm{P})} N_{E_\delta}^{(\bm{P})}-N_{E_q}^{(\bm{R})} N_{E_\delta}^{(\bm{R})}\right| \\
\leq & \mathbb{E} \sum_{\vec{i}\in C_q^<} \sum_{\vec{j}\in C_\delta^<} \left|\Phi_{\vec{i}}^{(\bm{P})} \Phi_{\vec{j}}^{(\bm{P})} -\Phi_{\vec{i}}^{(\bm{R})} \Phi_{\vec{j}}^{(\bm{R})}\right|1\left(\mathcal{E}(t)\right) + \binom{p}{1}\binom{p-1}{q}\binom{p}{1}\binom{p-1}{\delta}P(\mathcal{E}^c(t))\\
\leq & \mathbb{E} \sum_{\vec{i}\in C_q^<} \sum_{\vec{j}\in C_\delta^<} \eta_{\vec{i},\vec{j}} + \frac{p^{q+\delta+2}}{\delta!q!} 2\exp(-c_1t^2) \numberthis \label{eqn:Nedgedelprnew}
\end{align*}
where the first inequality follows from $0\leq N_{\EE_\delta}^{(\bm{\Psi})} \leq \binom{p}{1}\binom{p-1}{\delta}$ for both $\bm{\Psi} =\bm{R}$ and $\bm{\Psi} = \bm{P}$, and the second inequality follows from Lemma \ref{lem:inddifbou2ndmoment}, Lemma \ref{lem:xietarelationship} and \eqref{eqn:proepst}.
Choose $t=s_0\sqrt{\ln p}$ with $s_0 = \sqrt{\left(\frac{9}{2}+2\delta\right)/c_1}$. Notice that $s_0 \geq \sqrt{\left(\frac{3}{2}+q+\delta + \frac{q}{\delta} \right)/c_1}$ since $q\in \{\delta,\delta+1\}$. Then
$$
2\exp(-c_1t^2)\leq 2\exp\left(-\left(\frac{3}{2}+q+\delta+\frac{q}{\delta}\right)\ln p\right)=\frac{2}{p^{\frac{3}{2}+q+\delta+\frac{q}{\delta}}}.
$$
Moreover, for any $c<\frac{1}{2\max\left \{\sqrt{\left(\frac{9}{2}+2\delta\right)/c_1},1\right\}\sqrt{2}C_1} $,
$$
\left(\sqrt{\frac{n-1}{p}}+\sqrt{\frac{\delta \ln p}{p}}\right) \leq c
$$
implies
\begin{equation}
\sqrt{2}C_1\left(\sqrt{\frac{n-1}{p}}+s_0 \sqrt{\frac{\ln p}{p}}\right) \leq \frac{1}{2}, \label{eqn:CEatleadelprenew}
\end{equation}
which is \eqref{eqn:postempass} with $t= s_0 \sqrt{\ln p}$. Then apply Lemma \ref{lem:etadoublesumexpectation} with $t=s_0 \sqrt{\ln p}$ to \eqref{eqn:Nedgedelprnew},
\begin{align*}
&\mathbb{E} \left|N_{E_q}^{(\bm{P})} N_{E_\delta}^{(\bm{P})}-N_{E_q}^{(\bm{R})} N_{E_\delta}^{(\bm{R})}\right|\\
\leq & C_{n,q,\delta,\gamma} \left(\theta_1(s_0\ln p)\right)^{n(2\delta+q)} \left(1+\mu_{n,q+\delta+2}(\bm{\Sigma})\frac{\kappa-1}{p}\right) \left(\sqrt{\frac{1}{p}}+\frac{s_0\sqrt{\ln p}}{\sqrt{p}} + \frac{\tau}{p}\right) p^{1-\frac{q}{\delta}}\\
&\ \ \ \ \ \ \ + \frac{2}{\delta!q!\sqrt{p}} p^{1-\frac{q}{\delta}}\\
\leq & C_{n,q,\delta,\gamma} \left(1+\mu_{n,q+\delta+2}(\bm{\Sigma})\frac{\kappa-1}{p}\right) \left(s_0\frac{\sqrt{\ln p}}{\sqrt{p}} + \frac{\tau}{p}\right) p^{1-\frac{q}{\delta}} + \frac{2}{\delta!q!\sqrt{p}}p^{1-\frac{q}{\delta}} \\
\leq & C_{n,q,\delta,\gamma} \left(1+\mu_{n,q+\delta+2}(\bm{\Sigma})\frac{\kappa-1}{p}\right) \left(\frac{\sqrt{\ln p}}{\sqrt{p}} + \frac{\tau}{p}\right) p^{1-\frac{q}{\delta}}\\
\leq & C_{n,\delta,\gamma} \left(1+\mu_{n,q+\delta+2}(\bm{\Sigma})\frac{\kappa-1}{p}\right) \left(\frac{\sqrt{\ln p}}{\sqrt{p}} + \frac{\tau}{p}\right) p^{1-\frac{q}{\delta}},
\end{align*}
where the second inequality follows from $\theta_1\left(s_0 \sqrt{\ln p}\right)\leq 9+ 4(n-1)=4n+5$ by \eqref{eqn:CEatleadelprenew} and $\tau\leq p/2$; and the last step follows from $q\in \{\delta,\delta+1\}$.
\end{proof}
\begin{proof}[Proof of Proposition \ref{prop:allcloseL2} \ref{item:allcloseL2b} and \ref{item:allcloseL2c}]
(b)
It follows directly from Lemma \ref{lem:NEqNEdelta} with $q=\delta$.\\
\noindent
(c)
By taking square of each terms in Lemma \ref{lem:6quantitiesine},
$$
\left(N_{E_\delta}^{(\bm{P})}\right)^2 - 2(\delta+1)N_{E_\delta}^{(\bm{P})} N_{E_{\delta+1}}^{(\bm{P})} \leq \left(N_{\breve{V}_\delta}^{(\bm{P})}\right)^2 \leq \left(N_{V_\delta}^{(\bm{P})}\right)^2 \leq \left(N_{E_\delta}^{(\bm{P})}\right)^2,
$$
which then implies for $\bar{N}_\delta \in \left\{ N_{\breve{V}_\delta}^{(\bm{P})}, N_{V_\delta}^{(\bm{P})} \right\} $
$$
\left|\bar{N}_\delta- \left(N_{E_\delta}^{(\bm{P})}\right)^2\right|\leq 2(\delta+1)\left(N_{E_{\delta+1}}^{(\bm{P})} N_{E_\delta}^{(\bm{P})} - N_{E_{\delta+1}}^{(\bm{R})} N_{E_\delta}^{(\bm{R})} \right) + 2(\delta+1) N_{E_{\delta+1}}^{(\bm{R})} N_{E_\delta}^{(\bm{R})} .
$$
By Lemma \ref{lem:NEqNEdelta} with $q=\delta+1$,
$$
\mathbb{E} \left|N_{E_{\delta+1}}^{(\bm{P})} N_{E_\delta}^{(\bm{P})}-N_{E_{\delta+1}}^{(\bm{R})} N_{E_\delta}^{(\bm{R})}\right|\leq C_{n,\delta,\gamma} \left(1+\mu_{n,2\delta+3}(\bm{\Sigma})\frac{\kappa-1}{p}\right) \left(\frac{\sqrt{\ln p}}{\sqrt{p}} + \frac{\tau}{p}\right) p^{-\frac{1}{\delta}}.
$$
The proof is then completed by combining the preceding two displays, \eqref{eqn:NEdelta1NEdelta} and the fact that
$$\left(\frac{\sqrt{\ln p}}{\sqrt{p}} + \frac{\tau}{p}\right)\leq 1.$$
\end{proof}
\section{Proofs in Subsection \ref{sec:allclose} } \label{sec:proofofallclose}
\subsection{Proofs in Subsection \ref{sec:allclose} } \label{sec:proofofallclose}
\subsubsection{Proof of Lemma \ref{lem:6quantitiesine}}
\begin{proof}[Proof of Lemma \ref{lem:6quantitiesine}]
$N_{\breve{V}_\delta}^{(\bm{R})} \leq N_{V_\delta}^{(\bm{R})} \leq N_{E_\delta}^{(\bm{R})}$ follows trivially from their definitions. It remains to show
\begin{equation}
N_{E_\delta}^{(\bm{R})}\leq (\delta+1)N_{E_{\delta+1}}^{(\bm{R})} + N_{\breve{V}_\delta}^{(\bm{R})}. \label{eqn:tobeproved}
\end{equation}
To see this, consider $\delta\geq 2$ and any vertex $i$ and denote its degree by $m$. If $m<\delta$, then it contributes zero to both sides of \eqref{eqn:tobeproved}. If $m=\delta$, then it contributes $1$ to both sides of \eqref{eqn:tobeproved}. If $m>\delta$, it contributes $\binom{m}{\delta}$ to left hand side of \eqref{eqn:tobeproved}, while contributes $(\delta+1)\binom{m}{\delta+1}=(m-\delta)\binom{m}{\delta}$. The above observation proves \eqref{eqn:tobeproved}. The case $\delta=1$ is similar and is omitted.
The above proof indeed applies to any graph and, in particular, the empirical partial correlation graph. So the second equation in the statement of the lemma holds.
\end{proof}
\subsubsection{Proof of Proposition \ref{thm:allclose} \ref{item:allclosea}}
By \eqref{eqn:corexaatupp}, it suffices to establish an upper bound on $\mathbb{E} N_{E_{\delta+1}}^{(\bm{R})}$.
\begin{lem} \label{lem:NEbou}
{Let $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$. Suppose $\bm{\Sigma}$ is row-$\kappa$ sparse. Let $\ell\in [p-1]$. Then
$$
\mathbb{E} N_{E_{\ell}}^{(\bm{R})} \leq \frac{1}{\ell!}\left(1+\ell^2\mu_{n,\ell+1}(\bm{\Sigma})\frac{\kappa-1}{p}\right)p\left( 2pP_n(r_\rho)\right)^\ell.
$$}
\end{lem}
\begin{proof}
\begin{align*}
\mathbb{E} N_{E_{\ell}}^{(\bm{R})} = & \sum_{\vec{i}\in C_{\ell}^< }\mathbb{E} \prod_{j=1}^{\ell} \Phi_{i_0i_j}^{(\bm{R})} \\
\leq & \sum_{\vec{i}\in C_\ell^<}\mu_{n,\ell+1}\left(\bm{\Sigma}_{\vec{i}}\right) \left( 2P_n(r_\rho)\right)^\ell \\
\leq & \frac{1}{\ell!}\left(1+\ell^2\mu_{n,\ell+1}(\bm{\Sigma})\frac{\kappa-1}{p}\right)p\left( 2pP_n(r_\rho)\right)^\ell,
\end{align*}
where the first inequality follows from Lemma \ref{item:prod}, and the second inequality follows from Lemma \ref{lem:summuupp}.
\end{proof}
\begin{proof}[Proof of Proposition \ref{thm:allclose} \ref{item:allclosea}] It follows from \eqref{eqn:corexaatupp}, Lemma \ref{lem:NEbou} and Lemma \ref{prop:tvtomean}.
\end{proof}
\subsubsection{Proof of Proposition \ref{thm:allclose} \ref{item:allclosec}}
\label{sec:proofthmallclosec}
Similar to \eqref{eqn:NEdeltaRdef} and \eqref{eqn:Nedgedef1}, denote
\begin{equation*}
\Phi_{\vec{i}}^{(\bm{R})}= \prod_{j=1}^\delta \Phi_{i_0i_j}^{(\bm{R})} = 1\left(\bigcap_{j=1}^{\delta}\{\textbf{dist}(\bm{u}_{i_0},\bm{u}_{i_j})\leq r_\rho\}\right). \label{eqn:NEdeltaPdef}
\end{equation*}
Then by definition
\begin{equation}
N_{\EE_\delta}^{(\bm{P})} =\sum_{\vec{i}\in C_\delta^<}\ \ \Phi_{\vec{i}}^{(\bm{P})}. \label{eqn:NEdeltaPdef1}
\end{equation}
\index{$N_{\EE_\delta}^{(\bm{R})}$}
By \eqref{eqn:Nedgedef1} and \eqref{eqn:NEdeltaPdef1},
$$
\left|N_{\EE_\delta}^{(\bm{P})}-N_{\EE_\delta}^{(\bm{R})}\right| \leq \sum_{\vec{i}\in C_\delta^<}|\Phi_{\vec{i}}^{(\bm{P})} - \Phi_{\vec{i}}^{(\bm{R})}|.
$$
The next three lemmas establish upper bound on $|\Phi_{\vec{i}}^{(\bm{P})} - \Phi_{\vec{i}}^{(\bm{R})}|$.
We may suppose $\bm{\Sigma}$ is $(\tau,\kappa)$ sparse throughout this proof and the proof of Proposition \ref{thm:allclose} \ref{item:allclosed} since the conclusion is invariant to permutation of the variables by Remark \ref{rem:firsttaurworemark}. As a result, the U-scores may be partitioned into $\hat{\bm{U}}\in \mathbb{R}^{(n-1)\times \tau}$ consisting of the first $\tau$ columns and $\check{\bm{U}}\in \mathbb{R}^{(n-1)\times (p-\tau)}$ consisting the remaining $p-\tau$ columns.
Denote $[\tau]=\{1,2,\cdots,\tau \}$.
Define a matrix $\check{\bm{B}}$ by
\begin{equation}
\check{\bm{B}} = \frac{n-1}{p-\tau}\check{\bm{U}}[\check{\bm{U}}]^\top
=
\frac{n-1}{p-\tau}\displaystyle\sum_{i\in [p]\backslash [\tau] }^{p}\bm{u}_i\bm{u}_i^\top. \label{Bp'def}
\end{equation}
Denote $\check{\bm{Q}}=\sqrt{n-1}\check{\bm{U}}$. Observe that $\check{\bm{Q}}$ has exactly $p-\tau$ independent columns and each column $\sqrt{n-1}\bm{u}_i\sim \operatorname{{\bf unif}}(\sqrt{n-1}S^{n-2})$. These observations immediately give us part \ref{item:bpproa} of the following.
\begin{lem}
\label{lem:bppro}
Let $\{\bm{u}_\alpha\}_{\alpha =1 }^ p$ be columns of $\bm{U}$ defined in Section \ref{sec:scorerep}. Let $\check{\bm{B}}$ be defined as in equation \eqref{Bp'def}.
\begin{enumerate}[label=(\alph*)]
\item \label{item:bpproa} Suppose $\bm{\Sigma}$ is $(\tau,\kappa)$ sparse. $\check{\bm{B}}=\frac{1}{p-\tau}\bm{Q}\Qma^\top$, where $\bm{Q}\in R^{(n-1)\times (p-\tau)}$ has independent columns with each column distributed as $\operatorname{{\bf unif}}(\sqrt{n-1}S^{n-2})$.
\item \label{item:bpprob} $|\lambda_{\text{max}}\left(\frac{p}{p-\tau}\bm{B}\right)-\lambda_{\text{max}}(\check{\bm{B}})|\leq \frac{n-1}{p-\tau}\tau$, and $\lambda_{\text{min}}\left(\frac{p}{p-\tau}\bm{B}\right)\geq \lambda_{\text{min}}(\check{\bm{B}})$.
\end{enumerate}
\end{lem}
\begin{proof}
(b) Recall $\bm{B}=\frac{n-1}{p}\sum_{i=1}^p \bm{u}_i\bm{u}_i^\top$. Then,
$$
\frac{p}{p-\tau}\bm{B}-\check{\bm{B}} = \frac{n-1}{p-\tau}\sum_{i\in [\tau] }\bm{u}_i\bm{u}_i^\top.
$$
By Lemma \ref{perthe} \ref{item:perthea}, we have:
\begin{align*}
|\lambda_{\text{max}}\left(\frac{p}{p-\tau}\bm{B}\right)-\lambda_{\text{max}}\left(\check{\bm{B}}\right)| &\leq \left\|\frac{n-1}{p-\tau}\sum_{i\in [\tau] }\bm{u}_i\bm{u}_i^\top\right\|_2\\
&\leq \frac{n-1}{p-\tau}\sum_{i\in [\tau] }\|\bm{u}_i\bm{u}_i^\top\|_2 \\
&\leq \frac{n-1}{p-\tau}\tau,
\end{align*}
where for the last inequality, we use the fact that $\bm{u}_i \in S^{n-2}$. Moreover, by Lemma \ref{perthe} \ref{item:perthec}, we get $\lambda_{\text{min}}\left(\frac{p}{p-\tau}\bm{B}\right)\geq \lambda_{\text{min}}(\check{\bm{B}})$.
\end{proof}
Denote $h_0\left(\check{\bm{B}}\right)= \frac{\lambda_{\text{max}}\left(\check{\bm{B}}\right)+\frac{n-1}{p-\tau}\tau}{\lambda_{\text{min}}\left(\check{\bm{B}}\right)}=\frac{S_\text{max}\left(\check{\bm{B}}\right)+\frac{n-1}{p-\tau}\tau}{S_\text{min}\left(\check{\bm{B}}\right)}$ to be the perturbational condition number of $\check{\bm{B}}$, where $\lambda_{\text{max}}\left(\check{\bm{B}}\right)$, $\lambda_{\text{min}}\left(\check{\bm{B}}\right)$, $S_{\text{max}}\left(\check{\bm{B}}\right)$, and $S_{\text{min}}\left(\check{\bm{B}}\right)$ are respectively the largest eigenvalue, smallest eigenvalue, largest singular value and smallest singular value of $\check{\bm{B}}$.
\begin{lem}
\label{uyrelation}
Suppose $p\geq n$. Let $\{\bm{u}_\alpha\}_{\alpha =1 }^ p$ and $\{y_\alpha\}_{\alpha =1 }^p$ be defined as in Section \ref{sec:scorerep}.
Consider distinct $i,j$ satisfying $1\leq i, j \leq p$. Then with probability $1$,
$$
\frac{1}{h_0\left(\check{\bm{B}}\right)}\|\bm{u}_i-\bm{u}_j\|_2 \leq \|y_i-y_j\|_2 \leq h_0\left(\check{\bm{B}}\right)\|\bm{u}_i-\bm{u}_j\|_2,
$$
and
$$
\frac{1}{h_0\left(\check{\bm{B}}\right)}\|\bm{u}_i+\bm{u}_j\|_2 \leq \|y_i+y_j\|_2 \leq h_0\left(\check{\bm{B}}\right)\|\bm{u}_i+\bm{u}_j\|_2.
$$
\end{lem}
\begin{proof}
Recall $y_\alpha=\bar{y}_\alpha/\|\bar{y}_\alpha\|_2$ and $\bar{y}_\alpha = \bm{B}^{-1}\bm{u}_\alpha$ $a.s.$, for $\alpha=i,j$. Apply the upper bound in Lemma \ref{isoine},
\begin{align*}
\|y_i-y_j\|_2
&\leq \frac{\lambda_{\text{max}}(\bm{B}^{-1})}{\lambda_{\text{min}}(\bm{B}^{-1})}\|\bm{u}_i-\bm{u}_j\|_2 \quad a.s. \\
& =
\frac{\lambda_{\text{max}}\left(\frac{p}{p-\tau}\bm{B}\right)}{\lambda_{\text{min}}\left(\frac{p}{p-\tau}\bm{B}\right)}\|\bm{u}_i-\bm{u}_j\|_2 \\
& \leq \frac{\lambda_{\text{max}}\left(\check{\bm{B}}\right)+\frac{n-1}{p-\tau}\tau}{\lambda_{\text{min}}\left(\check{\bm{B}}\right)}\|\bm{u}_i-\bm{u}_j\|_2 ,
\end{align*}
where the last inequality follows from Lemma \ref{lem:bppro} \ref{item:bpprob}. The lower bound of the first desired display follows similarly, by the lower bound in Lemma \ref{isoine}.
The second desired expression follows analogously.
\end{proof}
For $\{i,j\}\in [p]$ with $i\not =j$, $q\in\{-1,+1\}$, define
\begin{align*}
S_{ij}^{(q)}(r_\rho)& =\{\|y_i-qy_j\|_2 \leq r_\rho \}, \quad F_{ij}^{(q)}(r_\rho)=\{\|\bm{u}_i-q\bm{u}_j\|_2 \leq r_\rho \}, \numberthis \label{eqn:Fijdef} \\
G_{ij}^{(q)}(r_\rho)& =\left\{\|\bm{u}_i-q\bm{u}_j\|_2 \leq \frac{1}{h_0\left(\check{\bm{B}}\right)}r_\rho \right \}, \quad H_{ij}^{(q)}(r_\rho)=\left \{\|\bm{u}_i-q\bm{u}_j\|_2 \leq h_0\left(\check{\bm{B}}\right)r_\rho \right\}.
\end{align*}
Define $F_{ij}(r_\rho)=F_{ij}^{(-1)}(r_\rho)\cup F_{ij}^{(+1)}(r_\rho)$. $G_{ij}(r_\rho)$, $H_{ij}(r_\rho)$, $S_{ij}(r_\rho)$ are defined similarly. Using these notation, then $\Phi_{ij}^{(\bm{P})}(\rho)=1\left(S_{ij}(r_\rho)\right)$, and $\Phi_{ij}^{(\bm{R})}(\rho)=1\left(F_{ij}(r_\rho)\right)$. For $\vec{i}\in C_\delta^<$, denote
$$
H_{\vec{i}}(r_\rho)=\bigcap_{\ell=1}^{\delta} H_{i_0i_\ell}(r_\rho),\quad H_{\vec{i},-m}(r_\rho)=\bigcap_{\substack{\ell=1\\ \ell\neq m}}^{\delta} H_{i_0i_\ell}(r_\rho).
$$
When it is clear from the context, the dependence of $r_\rho$ for the above quantities will be suppressed. By Lemma \ref{uyrelation}, with probability $1$,
\begin{equation}
G_{ij}^{(q)} \subset S_{ij}^{(q)}\subset H_{ij}^{(q)}, \quad G_{ij}^{(q)} \subset F_{ij}^{(q)}\subset H_{ij}^{(q)}. \label{eqn:FGJSsetinclusion}
\end{equation}
\begin{lem}
\label{lem:inddifbou}
Suppose $p\geq n$. Consider $\delta\in [p-1]$. For any $\vec{i}\in C_\delta^<$, with probability $1$,
$$
\left| \Phi_{\vec{i}}^{(\bm{P})}-\Phi_{\vec{i}}^{(\bm{R})} \right|\leq \xi_{\vec{i}},
$$
where
$$
\xi_{\vec{i}} := 1\left( \bigcup\limits_{m=1}^\delta
\left ( \left( H_{i_0i_m} \big \backslash G_{i_0i_m} \right) \bigcap H_{\vec{i},-m} \right )\right).
$$
\end{lem}
\begin{proof}
Notice
$\Phi_{\vec{i}}^{(\bm{R})}=1\left(\bigcap\limits_{m=1}^\delta F_{i_0i_m}\right)$ and $\Phi_{\vec{i}}^{(\bm{P})}=1\left(\bigcap\limits_{m=1}^\delta S_{i_0i_m}\right)$. Let $\bigtriangleup$ denote the symmetrization difference of two sets. Then
\begin{align*}
\left| \Phi_{\vec{i}}^{(\bm{P})}-\Phi_{\vec{i}}^{(\bm{R})} \right|=
1\left(\left(\bigcap\limits_{m=1}^\delta F_{i_0i_m}\right)\bigtriangleup \left(\bigcap\limits_{m=1}^\delta S_{i_0i_m}\right) \right)
\leq \xi_{\vec{i}},
\end{align*}
where the inequality follows from \eqref{eqn:FGJSsetinclusion} and Lemma \ref{setrelation} \ref{prop:setinclusiona}.
\end{proof}
To obtain an upper bound on the expectation of $\xi_{\vec{i}}$, we first bound the expectation on a high-probability set. Define the set $\mathcal{E}(t)$, with $t$ being a parameter to be determined, by
\begin{align*}
\mathcal{E}(t) &= \left\{ \left[1-C_1\left(\sqrt{\frac{n-1}{p-\tau}}+\frac{t}{\sqrt{p-\tau}}\right)\right]^2 \leq \lambda_{\text{min}}(\check{\bm{B}})\right\}\bigcap \\
&\bigcap \left\{\lambda_{\text{max}}(\check{\bm{B}})\leq \left[1+C_1\left(\sqrt{\frac{n-1}{p-\tau}}+\frac{t}{\sqrt{p-\tau}}\right)\right]^2 \right\}, \numberthis \label{Epsdef}
\end{align*}
to be the set such that \eqref{eqn:eigupplowbou} in Lemma \ref{proeig} holds, i.e. the constant $C_1$ in $\mathcal{E}(t)$ is the same constant as $C$ in \eqref{eqn:eigupplowbou}. By Lemma \ref{lem:bppro} \ref{item:bpproa} and Lemma \ref{proeig},
\begin{equation}
P(\mathcal{E}^c(t))\leq 2\exp(-c_1t^2). \label{eqn:proepst}
\end{equation}
{ Since $\tau\leq\frac{p}{2}$},
\begin{equation}
\frac{n-1}{p-\tau}\tau \leq 2(n-1)\frac{\tau}{p}, \label{eqn:ntaupupp}
\end{equation}
and
\begin{equation}
C_1\left(\sqrt{\frac{n-1}{p-\tau}}+\frac{t}{\sqrt{p-\tau}}\right)\leq \sqrt{2}C_1\left(\sqrt{\frac{n-1}{p}}+\frac{t}{\sqrt{p}}\right). \label{eqn:c1nptauupp}
\end{equation}
Moreover, on $\mathcal{E}(t)$, and assuming
\begin{equation}
\sqrt{2}C_1\left(\sqrt{\frac{n-1}{p}}+\frac{t}{\sqrt{p}}\right) \leq \frac{1}{2}, \label{eqn:postempass}
\end{equation}
one has
\begin{multline*}
h_0(\bm{B}) \leq \frac{\left(1+C_1\left(\sqrt{\frac{n-1}{p-\tau}}+\frac{t}{\sqrt{p-\tau}}\right)\right)^2+\frac{n-1}{p-\tau}\tau}{\left(1-C_1\left(\sqrt{\frac{n-1}{p-\tau}}+\frac{t}{\sqrt{p-\tau}}\right)\right)^2}\\
\leq 1+ 16 \sqrt{2}C_1\left(\sqrt{\frac{n-1}{p}}+\frac{t}{\sqrt{p}}\right) + 8(n-1)\frac{\tau}{p}
\vcentcolon= \theta_1(t), \numberthis \label{ctempdef}
\end{multline*}
where the second inequality follows from \eqref{eqn:ntaupupp}, \eqref{eqn:c1nptauupp} and Lemma \ref{basine} \ref{item:basinej}.
For $\vec{i}\in C_\delta^<$, denote
$$
F_{\vec{i}}(r_\rho)=\bigcap_{\ell=1}^{\delta} F_{i_0i_\ell}(r_\rho),\quad F_{\vec{i},-m}(r_\rho)=\bigcap_{\substack{\ell=1\\ \ell\neq m}}^{\delta} F_{i_0i_\ell}(r_\rho).
$$
\begin{lem}
\label{lem:inddifuppexpbou}
Let $p\geq n\geq 4$, $\delta\in [p-1]$ and $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$. Suppose $\bm{\Sigma}$, after some row-column permutation, is $(\tau,\kappa)$ sparse with $\tau\leq \frac{p}{2}$. Let $t$ be any positive number, and suppose \eqref{eqn:postempass} holds. Then for any $\vec{i}\in C_\delta^<$, with probability $1$,
$$
\xi_{\vec{i}} 1\left(\mathcal{E}(t)\right) \leq \eta_{\vec{i}}(t),
$$
where
\begin{align}
\eta_{\vec{i}}(t) := &1 \left( \bigcup_{m=1}^\delta\left( \left(F_{i_0i_m}(\theta_1(t)r_\rho)\big \backslash F_{i_0i_m}\left(\frac{r_\rho}{\theta_1(t)}\right) \right) \bigcap F_{\vec{i},-m}(\theta_1(t)r_\rho) \right)\right). \label{eqn:etavecidef}
\end{align}
Moreover,
\begin{align*}
&\mathbb{E} 1 \left( \left(F_{i_0i_m}(\theta_1(t)r_\rho)\big \backslash F_{i_0i_m}\left(\frac{r_\rho}{\theta_1(t)}\right) \right) \bigcap F_{\vec{i},-m}(\theta_1(t)r_\rho) \right)\\
\leq & \mu_{n,\delta+1}(\bm{\Sigma}_{\vec{i}}) 2\left(P_n(r_\rho \theta_1(t))-P_n\left(\frac{r_\rho}{\theta_1(t)} \right)\right) \left(2P_n\left(\theta_1(t)r_\rho\right)\right)^{\delta-1}, \numberthis \label{eqn:etasingletermbound}
\end{align*}
and
\begin{align*}
\mathbb{E} \eta_{\vec{i}}(t)
\leq & \delta \mu_{n,\delta+1}(\bm{\Sigma}_{\vec{i}}) 2\left(P_n(r_\rho \theta_1(t))-P_n\left(\frac{r_\rho}{\theta_1(t)} \right)\right) \left(2P_n\left(\theta_1(t)r_\rho\right)\right)^{\delta-1} \\
\leq & \mu_{n,\delta+1}(\bm{\Sigma}_{\vec{i}}) \delta n (\theta_1(t))^{n\delta}\left(\theta_1(t)-\frac{1}{\theta_1(t)}\right) \left( 2P_n\left(r_\rho\right) \right)^{\delta}.
\end{align*}
\end{lem}
\begin{proof}
By \eqref{ctempdef}, $H_{ij}(r_\rho) \cap \mathcal{E}(t)\subset F_{ij}(\theta_1(t)r_\rho)$ and $G_{ij}(r_\rho) \cap \mathcal{E}(t)\supset F_{ij}\left(\frac{r_\rho}{\theta_1(t)}\right)$. Then
\begin{align*}
\xi_{\vec{i}} 1\left(\mathcal{E}(t)\right)
\leq \eta_{\vec{i}}(t). \numberthis \label{eqn:A1upp1}
\end{align*}
\begin{align*}
&\mathbb{E} 1 \left( \left(F_{i_0i_m}(\theta_1(t)r_\rho)\big \backslash F_{i_0i_m}\left(\frac{r_\rho}{\theta_1(t)}\right) \right) \bigcap F_{\vec{i},-m}(\theta_1(t)r_\rho) \right) \\
\leq & \mu_{n,\delta+1}(\bm{\Sigma}_{\vec{i}}) P \left( \left(\bigcup\limits_{q\in\{-1,+1\}}\left\{ \frac{r_\rho}{\theta_1(t)} <\|\bm{u}'_{i_0}-q\bm{u}'_{i_m}\|_2 \leq \theta_1(t)r_\rho \right\}\right) \bigcap \right.\\
& \ \ \ \ \ \ \ \left. \left(\bigcap_{\substack{\alpha=1 \\ \alpha\not=m}}^\delta \left(\bigcup\limits_{q\in\{-1,+1\}} \{ \|\bm{u}'_{i_0}-q\bm{u}'_{i_\alpha}\|_2\leq \theta_1(t)r_\rho \} \right)\right)\right) , \numberthis \label{eqn:expindupp}
\end{align*}
where the last inequality follows from Lemma \ref{lem:densitybou} \ref{item:densitybouc} with
$$
\bm{u}'_{i_0},\bm{u}'_{i_1},\cdots,\bm{u}'_{i_\delta}\overset{\text{i.i.d.}}{\sim} \operatorname{{\bf unif}}(S^{n-2}).
$$
For any $\bm{w}\in S^{n-2}$, define $\Omega_{\bm{w}}^{(q)} :=\{\bm{v} \in S^{n-2}:\frac{1}{\theta_1(t)} r_\rho < \|\bm{v}-q\bm{w}\|_2 \leq r_\rho \theta_1(t)\}$. Then
$$P\left(\bm{u}'_{im}\in \bigcup_{q\in\{-1,+1\}} \Omega_{\bm{w}}^{(q)} \right)=2\left(P_n(r_\rho \theta_1(t))-P_n\left(\frac{1}{\theta_1(t)}r_\rho \right)\right).$$
By conditioning on $\bm{u}'_{i_0}$, the term in right hand side of \eqref{eqn:expindupp} equals to
\begin{align*}
\mu_{n,\delta+1}(\bm{\Sigma}_{\vec{i}}) 2\left(P_n(r_\rho \theta_1(t))-P_n\left(\frac{1}{\theta_1(t)}r_\rho \right)\right) \left(2P_n\left(\theta_1(t)r_\rho\right)\right)^{\delta-1},
\end{align*}
which then proves \eqref{eqn:etasingletermbound}.
By the union bound,
\begin{align*}
&\mathbb{E} \eta_{\vec{i}}(t) \\
\leq &
\sum\limits_{m=1}^\delta \mathbb{E} 1 \left( \left(F_{i_0i_m}(\theta_1(t)r_\rho)\big \backslash F_{i_0i_m}\left(\frac{r_\rho}{\theta_1(t)}\right) \right) \bigcap F_{\vec{i},-m}(\theta_1(t)r_\rho) \right) \\
\overset{(*)}{\leq} &
\delta \mu_{n,\delta+1}(\bm{\Sigma}_{\vec{i}}) 2\left(P_n(r_\rho \theta_1(t))-P_n\left(\frac{1}{\theta_1(t)}r_\rho \right)\right) \left(2P_n\left(\theta_1(t)r_\rho\right)\right)^{\delta-1}\\
\overset{(**)}{\leq} & \delta\mu_{n,\delta+1}(\bm{\Sigma}_{\vec{i}})
2(n-2)P_n(r_\rho)\left(\theta_1(t)\right)^{n-3} \left(\theta_1(t)-\frac{1}{\theta_1(t)}\right)\left(2P_n(r_\rho \theta_1(t))\right)^{\delta-1}\\
\overset{(***)}{\leq} & \delta\mu_{n,\delta+1}(\bm{\Sigma}_{\vec{i}})
2(n-2)P_n(r_\rho) \left(\theta_1(t)\right)^{n-3}\left(\theta_1(t)-\frac{1}{\theta_1(t)}\right) \left(\left(\theta_1(t)\right)^{n-2} 2P_n\left(r_\rho\right) \right)^{(\delta-1)},
\end{align*}
where $(*)$ follows from \eqref{eqn:etasingletermbound}, $(**)$ follows from Lemma \ref{pderivative} \ref{item:pderivativec},
and $(***)$ follows from Lemma \ref{pderivative} \ref{item:pderivatived}.
\end{proof}
\begin{lem}
\label{lem:ideltaedgeindicatorbou}
Let $p\geq n\geq 4$, $\delta\in [p-1]$ and $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$.
Let $t$ be any positive number, and suppose \eqref{eqn:postempass} holds. Suppose $\bm{\Sigma}$, after some row-column permutation, is $(\tau,\kappa)$ sparse with $\tau\leq \frac{p}{2}$. Then
$$
\left|N_{E_{\delta}}^{(\bm{P})}-N_{E_{\delta}}^{(\bm{R})} \right| 1\left(\mathcal{E}(t) \right) \leq \sum_{\vec{i}\in C_\delta^<} \eta_{\vec{i}}(t)
$$
and
$$
\mathbb{E} \sum_{\vec{i}\in C_\delta^<} \eta_{\vec{i}}(t)
\leq \frac{Cn^2}{(\delta-1)!}\left(1+\delta^2\frac{\kappa-1}{p}\mu_{n,\delta+1}(\bm{\Sigma})\right) (\theta_1(t))^{n\delta}\left(\sqrt{\frac{1}{p}}+\frac{t}{\sqrt{p}} + \frac{\tau}{p}\right) p\left( 2pP_n\left(r_\rho\right) \right)^{\delta},
$$
where $C$ is an universal constant.
\end{lem}
\begin{proof}
\begin{align*}
\left|N_{E_{\delta}}^{(\bm{P})}-N_{E_{\delta}}^{(\bm{R})} \right| 1\left(\mathcal{E}(t) \right) \leq &
\sum_{\vec{i}\in C_\delta^<}\ \ \mathbb{E}\left| \Phi_{\vec{i}}^{(\bm{P})}
- \Phi_{\vec{i}}^{(\bm{R})}\right| 1\left(\mathcal{E}(t) \right) \\
\leq & \sum_{\vec{i}\in C_\delta^<} \eta_{\vec{i}}(t),
\end{align*}
where the last inequality follows from Lemma \ref{lem:inddifbou} and Lemma \ref{lem:inddifuppexpbou}.
By Lemma \ref{lem:inddifuppexpbou},
\begin{align*}
& \sum_{\vec{i}\in C_\delta^<} \mathbb{E} \eta_{\vec{i}} \\
\leq &
\sum_{\vec{i}\in C_\delta^<}\mu_{n,\delta+1}(\bm{\Sigma}_{\vec{i}}) \delta n (\theta_1(t))^{n\delta}\left(\theta_1(t)-\frac{1}{\theta_1(t)}\right) \left( 2P_n\left(r_\rho\right) \right)^{\delta}\\
\leq & \frac{p^{\delta+1}}{\delta!}\left(1+\delta^2\frac{\kappa-1}{p}\mu_{n,\delta+1}(\bm{\Sigma})\right)\delta n (\theta_1(t))^{n\delta}\left(\theta_1(t)-\frac{1}{\theta_1(t)}\right) \left( 2P_n\left(r_\rho\right) \right)^{\delta}\\
\leq & \frac{Cn}{(\delta-1)!}\left(1+\delta^2\frac{\kappa-1}{p}\mu_{n,\delta+1}(\bm{\Sigma})\right) (\theta_1(t))^{n\delta}\left(\sqrt{\frac{n}{p}}+\frac{t}{\sqrt{p}} + n\frac{\tau}{p}\right) \left( 2p^{1+\frac{1}{\delta}}P_n\left(r_\rho\right) \right)^{\delta} , \numberthis \label{eqn:temptemptemp1}
\end{align*}
where the third inequality follows from Lemma \ref{lem:summuupp} and the last inequality follows from Lemma \ref{basine} \ref{item:basinek} and \eqref{ctempdef}.
\end{proof}
\begin{lem} \label{lem:Nedgedelp}
Let $p\geq n \geq 4$, $\delta\in [p-1]$ and $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$. Suppose $\bm{\Sigma}$, after some row-column permutation, is $(\tau,\kappa)$ sparse with $\tau\leq \frac{p}{2}$. Suppose $2p^{1+\frac{1}{\delta}}P_n(r_\rho)\leq \gamma$ and $\left(\sqrt{\frac{n-1}{p}}+\sqrt{\frac{\delta\ln p}{p}}\right) \leq c$ hold for some positive and \ONE{sufficiently} small universal constant $c$. Then
\begin{align*}
\mathbb{E}\left|N_{\EE_\delta}^{(\bm{P})}-N_{\EE_\delta}^{(\bm{R})}\right|\leq C_{E_\delta}^{(\bm{P})} \left(1+\frac{\kappa-1}{p}\mu_{n,\delta+1}(\bm{\Sigma})\right)\left(\sqrt{\frac{\ln p}{p}}+\frac{\tau}{p}\right),
\end{align*}
where $C^{(\bm{P})}_{E_\delta} $ is defined in \eqref{eqn:CEatleadel}.
\end{lem}
\begin{proof}
\begin{align*}
\mathbb{E}\left|N_{\EE_\delta}^{(\bm{P})}-N_{\EE_\delta}^{(\bm{R})}\right|
&\leq \mathbb{E} \left|N_{\EE_\delta}^{(\bm{P})}-N_{\EE_\delta}^{(\bm{R})}\right|1\left(\mathcal{E}(t)\right) + \binom{p}{1}\binom{p-1}{\delta}P(\mathcal{E}^c(t)), \\
&\leq \mathbb{E} \left|N_{\EE_\delta}^{(\bm{P})}-N_{\EE_\delta}^{(\bm{R})}\right|1\left(\mathcal{E}(t)\right) + \frac{p^{\delta+1}}{\delta!}2\exp(-c_1t^2), \numberthis \label{eqn:Nedgedelpr}
\end{align*}
where the first inequality follows from $0\leq N_{\EE_\delta}^{(\bm{\Psi})} \leq \binom{p}{1}\binom{p-1}{\delta}$ for both $\bm{\Psi} =\bm{R}$ and $\bm{\Psi} = \bm{P}$, and the second inequality follows from \eqref{eqn:proepst}.
Choose $t=c_\delta\sqrt{\ln p}$ with $c_\delta=\sqrt{\frac{5\delta}{2c_1}} \geq \sqrt{\left(\frac{3}{2}+\delta\right)/c_1}$ such that
$$
2\exp(-c_1t^2)\leq 2\exp\left(-\left(\frac{3}{2}+\delta\right)\ln p\right)=\frac{2}{p^{\frac{3}{2}+\delta}}.
$$
Moreover, for any
\begin{equation}
c<\frac{1}{2\max\left \{\sqrt{\frac{5}{2c_1}},1\right\}\sqrt{2}C_1}, \label{eqn:cdef}
\end{equation}
the inequality
$$
\left(\sqrt{\frac{n-1}{p}}+\sqrt{\frac{\delta \ln p}{p}}\right) \leq c
$$
implies
\begin{equation}
\sqrt{2}C_1\left(\sqrt{\frac{n-1}{p}}+c_\delta\sqrt{\frac{\ln p}{p}}\right) \leq \frac{1}{2}, \label{eqn:CEatleadelpre}
\end{equation}
which is \eqref{eqn:postempass} with $t= c_\delta \sqrt{\ln p}$. Then apply Lemma \ref{lem:ideltaedgeindicatorbou} with $t=c_\delta \sqrt{\ln p}$ to \eqref{eqn:Nedgedelpr},
\begin{align*}
&\mathbb{E}\left|N_{\EE_\delta}^{(\bm{P})}-N_{\EE_\delta}^{(\bm{R})}\right|\\
\leq & \frac{Cn^2}{(\delta-1)!}\left(1+\delta^2\frac{\kappa-1}{p}\mu_{n,\delta+1}(\bm{\Sigma})\right) \left(\theta_1\left(c_\delta\sqrt{\ln p}\right)\right)^{n\delta}\left(\sqrt{\frac{1}{p}}+\frac{\sqrt{\delta \ln p}}{\sqrt{p}} + \frac{\tau}{p}\right) \gamma^{\delta} + \frac{2}{\delta!\sqrt{p}}\\
\leq & \frac{Cn^2\sqrt{\delta}}{(\delta-1)!}\left(1+\delta^2\frac{\kappa-1}{p}\mu_{n,\delta+1}(\bm{\Sigma})\right) (\theta_1(c_\delta \sqrt{\ln p}))^{n\delta}\left(\frac{\sqrt{\ln p}}{\sqrt{p}} + \frac{\tau}{p}\right) \gamma^{\delta} + \frac{2}{\delta!\sqrt{p}} \\
\leq & C_{E_\delta}^{(\bm{P})} \left(1+\frac{\kappa-1}{p}\mu_{n,\delta+1}(\bm{\Sigma})\right)\left(\sqrt{\frac{\ln p}{p}}+\frac{\tau}{p}\right),
\end{align*}
where
\begin{align*}
C_{E_\delta}^{(\bm{P})} = & \frac{Cn^2\delta^{\frac{5}{2}}}{(\delta-1)!} \left(\theta_1\left(c_\delta \sqrt{\ln p}\right)\right)^{n\delta} \gamma^{\delta}+\frac{2}{\delta!\sqrt{\ln p}}\\
\leq & \frac{Cn^2\delta^{\frac{5}{2}}}{(\delta-1)!} \left(4n+5\right)^{n\delta} \gamma^{\delta}+\frac{2}{\delta!}, \numberthis \label{eqn:CEatleadel}
\end{align*}
where the last step follows from $\theta_1\left(c_\delta \sqrt{\ln p}\right)\leq 9+ 4(n-1)=4n+5$ by \eqref{eqn:CEatleadelpre} and $\tau\leq p/2$.
\end{proof}
\begin{proof}[Proof of Proposition \ref{thm:allclose} \ref{item:allclosec}]
It follows directly from Lemma \ref{lem:Nedgedelp} and Lemma \ref{prop:tvtomean}.
\end{proof}
\subsubsection{Proof of Proposition \ref{thm:allclose} \ref{item:allclosed} }
By Lemma \ref{lem:6quantitiesine},
\[
N_{E_{\delta}}^{(\bm{P})}-(\delta+1)N_{E_{\delta+1}}^{(\bm{P})} - N_{E_{\delta}}^{(\bm{R})} \leq N_{\breve{V}_\delta}^{(\bm{P})}-N_{\breve{V}_\delta}^{(\bm{R})} \leq N_{E_{\delta}}^{(\bm{P})}-N_{E_{\delta}}^{(\bm{R})} +(\delta+1) N_{E_{\delta+1}}^{(\bm{R})},
\]
which implies
\begin{align*}
\left|N_{\breve{V}_\delta}^{(\bm{P})}-N_{\breve{V}_\delta}^{(\bm{R})}\right|\leq \left|N_{E_{\delta}}^{(\bm{P})} - N_{E_{\delta}}^{(\bm{R})}\right|+ (\delta+1)\left|N_{E_{\delta+1}}^{(\bm{P})}-N_{E_{\delta+1}}^{(\bm{R})}\right|+(\delta+1)N_{E_{\delta+1}}^{(\bm{R})}. \numberthis \label{eqn:Nvexacdifbou}
\end{align*}
\begin{lem} \label{lem:Nvexadelpr}
Let $p\geq n \geq 4$, $\delta\in [p-1]$ and $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$. Suppose $\bm{\Sigma}$, after some row-column permutation, is $(\tau,\kappa)$ sparse with $\tau\leq \frac{p}{2}$. Suppose $2p^{1+\frac{1}{\delta}}P_n(r_\rho)\leq \gamma$ and $\left(\sqrt{\frac{n-1}{p}}+\sqrt{\frac{\ln p}{p}}\right) \leq c$ hold for some positive and \ONE{sufficiently} small constant $c$. Then
$$\mathbb{E}\left|N_{\breve{V}_\delta}^{(\bm{P})}-N_{\breve{V}_\delta}^{(\bm{R})}\right|\leq C_{\breve{V}_\delta}^{(\bm{P})} \left(1+\frac{\kappa-1}{p}\mu_{n,\delta+2}(\bm{\Sigma})\right)\left(\sqrt{\frac{\ln p }{p}}+\frac{\tau}{p}+p^{-\frac{1}{\delta}} \right) $$
where $C_{\breve{V}_\delta}^{(\bm{P})}$ is defined in \eqref{eqn:CVexcdel}.
\end{lem}
\begin{proof}
Let $\mathcal{E}(t)$ be the same as in \eqref{Epsdef} with $t$ to be determined. Consider $\delta\in [p-2]$.
\begin{align*}
& \mathbb{E} \left|N_{\breve{V}_\delta}^{(\bm{P})}-N_{\breve{V}_\delta}^{(\bm{R})}\right| \\
\leq & \mathbb{E} \left|N_{\breve{V}_\delta}^{(\bm{P})}-N_{\breve{V}_\delta}^{(\bm{R})}\right| 1\left(\mathcal{E}(t)\right) + p P (\mathcal{E}^c(t)) \\
\leq & \mathbb{E} |N_{E_{\delta}}^{(\bm{P})} - N_{E_{\delta}}^{(\bm{R})}|1\left(\mathcal{E}(t)\right)+ (\delta+1)\mathbb{E}|N_{E_{\delta+1}}^{(\bm{P})}-N_{E_{\delta+1}}^{(\bm{R})}|1\left(\mathcal{E}(t)\right)+(\delta+1)\mathbb{E} N_{E_{\delta+1}}^{(\bm{R})} +2p e^{-c_1t^2}, \numberthis \label{eqn:Nvexadelpr}
\end{align*}
where the first inequality follows from $0\leq N_{\breve{V}_\delta}^{(\bm{\Psi})}\leq p$ for $\bm{\Psi} =\bm{R}$ and $\bm{P}$, the second inequality follows from \eqref{eqn:Nvexacdifbou} and \eqref{eqn:proepst}. If $\delta=p-1$, then
\begin{align*}
\mathbb{E} \left|N_{\breve{V}_{p-1}}^{(\bm{P})}-N_{\breve{V}_{p-1}}^{(\bm{R})}\right|
\leq & \mathbb{E} \left|N_{\breve{V}_{p-1}}^{(\bm{P})}-N_{\breve{V}_{p-1}}^{(\bm{R})}\right| 1\left(\mathcal{E}(t)\right) + p P (\mathcal{E}^c(t)) \\
\leq & \mathbb{E} \left|N_{E_{p-1}}^{(\bm{P})} - N_{E_{p-1}}^{(\bm{R})}\right|1\left(\mathcal{E}(t)\right) +2p \exp(-c_1t^2),
\end{align*}
which shows that \eqref{eqn:Nvexadelpr} also holds for $\delta=p-1$ with the convention $N_{E_{p}}^{(\bm{P})}=N_{E_{p}}^{(\bm{R})}=0$.
Choose $t=\sqrt{\frac{3}{c_1}\ln p}:=c_2 \sqrt{\ln p}$, such that $p\exp(-c_1t^2)=\frac{1}{p^2}$. Moreover, for any
\begin{equation}
c<\frac{1}{2\max\{c_2,1\}\sqrt{2}C_1}, \label{eqn:cuppbou2}
\end{equation}
the inequality
$$
\left(\sqrt{\frac{n-1}{p}}+\sqrt{\frac{\ln p}{p}}\right) \leq c
$$
implies
\begin{equation}
\sqrt{2}C_1\left(\sqrt{\frac{n-1}{p}}+c_2\sqrt{\frac{\ln p}{p}}\right) \leq \frac{1}{2}, \label{eqn:CEatleadelprenewnew}
\end{equation}
which is \eqref{eqn:postempass} with $t= c_2 \sqrt{\ln p}$. With $t= c_2 \sqrt{\ln p}$ Lemma \ref{lem:ideltaedgeindicatorbou} becomes:
\begin{align*}
& \mathbb{E} \left|N_{V_{\delta}}^{(\bm{P})}-N_{V_{\delta}}^{(\bm{R})} \right| 1\left(\mathcal{E}(t) \right)\\
\leq & \frac{Cn^2}{(\delta-1)!}\left(1+\delta^2\frac{\kappa-1}{p}\mu_{n,\delta+1}(\bm{\Sigma})\right) (\theta_1(c_2\sqrt{\ln p}))^{n\delta}\left(\frac{\sqrt{\ln p}}{\sqrt{p}} + \frac{\tau}{p}\right) p\left( 2pP_n\left(r_\rho\right) \right)^{\delta}, \numberthis \label{eqn:Etemp1}
\end{align*}
Then for $\delta\in [p-1]$ applying \eqref{eqn:Etemp1} with $\delta$, $\delta+1$ and Lemma \ref{lem:NEbou} to \eqref{eqn:Nvexadelpr}, together with $\mu_{n,\delta+1}(\bm{\Sigma})\leq \mu_{n,\delta+2}(\bm{\Sigma})$,
\begin{align*}
&\mathbb{E} \left|N_{\breve{V}_\delta}^{(\bm{P})}-N_{\breve{V}_\delta}^{(\bm{R})}\right|\\
\leq & \frac{Cn^2}{(\delta-1)!}\left(1+\delta^2\frac{\kappa-1}{p}\mu_{n,\delta+2}(\bm{\Sigma})\right) \left(\theta_1\left(c_2\sqrt{\ln p}\right)\right)^{n\delta}\left(\frac{\sqrt{\ln p}}{\sqrt{p}} + \frac{\tau}{p}\right) p\left( 2pP_n\left(r_\rho\right) \right)^{\delta}+ \\
&
{ (\delta+1)\frac{Cn^2}{\delta!}\left(1+(\delta+1)^2\frac{\kappa-1}{p}\mu_{n,\delta+2}(\bm{\Sigma})\right) \left(\theta_1\left(c_2\sqrt{\ln p}\right)\right)^{n(\delta+1)}\left(\frac{\sqrt{\ln p}}{\sqrt{p}} + \frac{\tau}{p}\right) p\left( 2pP_n\left(r_\rho\right) \right)^{\delta+1} } \\
& \ \ \ \ +(\delta+1)\frac{1}{(\delta+1)!}\left(1+(\delta+1)^2\mu_{n,\delta+2}(\bm{\Sigma})\frac{\kappa-1}{p}\right)p\left( 2pP_n(r_\rho)\right)^{\delta+1}
+ \frac{2}{p^2} \\
\leq & \frac{Cn^2}{(\delta-1)!}\left(1+(\delta+1)^2\frac{\kappa-1}{p}\mu_{n,\delta+2}(\bm{\Sigma})\right) \left(\theta_1\left(c_2\sqrt{\ln p}\right)\right)^{n(\delta+1)}\left(\frac{\sqrt{\ln p}}{\sqrt{p}} + \frac{\tau}{p}\right) \gamma^{\delta}\left(1+ \frac{\delta+1}{\delta}\gamma p^{-\frac{1}{\delta}}\right) \\
& \ \ \ \ +\frac{1}{\delta!}\left(1+(\delta+1)^2\mu_{n,\delta+2}(\bm{\Sigma})\frac{\kappa-1}{p}\right)\gamma^{\delta+1}p^{-\frac{1}{\delta}}
+ \frac{2}{p^2} \\
\leq & C_{\breve{V}_\delta}^{(\bm{P})} \left(1+\frac{\kappa-1}{p}\mu_{n,\delta+2}(\bm{\Sigma})\right)\left(\sqrt{\frac{\ln p }{p}}+\frac{\tau}{p}+p^{-\frac{1}{\delta}} \right),
\end{align*}
where in the last inequality
\begin{align*}
C_{\breve{V}_\delta}^{(\bm{P})}=&\frac{Cn^2(\delta+1)^2}{(\delta-1)!} \left(\theta_1\left(c_2\sqrt{\ln p}\right)\right)^{n(\delta+1)} \gamma^{\delta}(1+\gamma)\left(1+ \frac{\delta+1}{\delta}\gamma p^{-\frac{1}{\delta}}\right)
+ \frac{2}{p^{2-\frac{1}{\delta}}}\\
\leq & \frac{Cn^2(\delta+1)^2}{(\delta-1)!} \left(4n+5\right)^{n(\delta+1)} \gamma^{\delta}(1+\gamma)\left(1+ \frac{\delta+1}{\delta}\gamma p^{-\frac{1}{\delta}}\right)
+ 2. \numberthis \label{eqn:CVexcdel}
\end{align*}
where the last step follows from $\theta_1\left(c_2 \sqrt{\ln p}\right)\leq 9+ 4(n-1)=4n+5$ by \eqref{eqn:CEatleadelprenewnew} and $\tau\leq p/2$.
\end{proof}
\begin{proof}[Proof of Proposition \ref{thm:allclose} \ref{item:allclosed}]
Lemmas \ref{lem:Nvexadelpr} and \ref{prop:tvtomean} complete the proof of Proposition \ref{thm:allclose} \ref{item:allclosed}.
\end{proof}
\section{Proof of Proposition \ref{prop:edgecor}}
\subsection{Proof of Proposition \ref{prop:edgecor}}
\label{sec:proofprop:edgecor}
\subsubsection{Auxiliary lemmas for Proposition \ref{prop:edgecor}}
Recall for any $\delta\geq 1$, $C_\delta^<$ is defined in \eqref{eqn:Cdeltadef}. For $\vec{i}\in C_\ell^{<}$, define a symmetric positive definite matrix $\bm{\Sigma}_{\vec{i}}\in \mathbb{R}^{(\ell+1)\times (\ell+1)}$ to be the submatrix of $\bm{\Sigma}$, consisting of rows and columns $\bm{\Sigma}$ indexed by the ordered components $(i_0,i_1,\ldots,i_{\ell})$ of $\vec{i}$. Let $\left[\vec{i}\right]=\{i_0,i_1,\cdots,i_\ell\}$ be the unordered set of indices of any $\vec{i}\in C_\ell^<$. Then $\bm{\Sigma}_{\vec{i}}\in \bm{\Sigma}_{[\vec{i}]}$
and $\mu_{n,\ell+1}\left(\bm{\Sigma}_{\vec{i}}\right)=\mu_{n,\ell+1}\left(\bm{\Sigma}_{[\vec{i}]}\right)$, where $\bm{\Sigma}_{[\vec{i}]}$ and $\mu_{n,\ell+1}\left(\bm{\Sigma}_{[\vec{i}]}\right)$ are defined in the paragraph after Definition \ref{def:nordet}.
\begin{lem} \label{item:prod}
Suppose $X \sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$. Let $\ell\in[p-1]$. Consider $\vec{i}=(i_0,i_1,\cdots,i_\ell)\in C^<_\ell$.
\begin{equation*}
\mathbb{E} \prod_{q=1}^\ell\Phi_{i_0i_q}^{(\bm{R})} \leq \mu_{n,\ell+1}\left(\bm{\Sigma}_{\vec{i}}\right) \left( 2P_n(r_\rho)\right)^\ell. \end{equation*}
Moreover, when $\bm{\Sigma}_{\vec{i}}$ is diagonal, in the last expression the equality holds and $\mu_{n,\ell+1}\left(\bm{\Sigma}_{\vec{i}}\right)=1$.
\end{lem}
\begin{proof}
$ \prod\limits_{q=1}^\ell\Phi_{i_0i_q}^{(\bm{R})}$ is a nonnegative Borel Measurable function of $\bm{u}_j$ for $j\in \left[\vec{i}\right]$. By Lemma \ref{lem:densitybou} \ref{item:densitybouc}, it suffices to show
$$
\mathbb{E} \prod_{q=1}^\ell\Phi_{i_0i_q}^{(\bm{R})} \leq \left( 2P_n(r)\right)^\ell
$$
for the case $\bm{u}_j$ for $j\in \left[\vec{i}\right]$ are $\ell+1$ independent $\operatorname{{\bf unif}}(S^{n-2})$. The last inequality holds with equality, which follows from that the terms in the product on the left hand side are independent conditioned on $\bm{u}_i$.
\end{proof}
Lemma \ref{item:prod} suggests differentiating whether $\bm{\Sigma}_{\vec{i}}$ is diagonal or not since $\mu_{n,\ell+1}\left(\bm{\Sigma}_{\vec{i}}\right)=1$ when $\bm{\Sigma}_{\vec{i}}$ is diagonal. The next lemma establishes an upper bound on the number of $\vec{i}\in C_\delta^<$ such that $\bm{\Sigma}_{\vec{i}}$ is not diagonal.
\begin{lem}
\label{lem:numofnondia}
Let $\bm{\Sigma}$ be row-$\kappa$ sparse. Let $\delta\in [p-1]$. Then
$$
\sum_{\substack{\vec{i}\in C_\delta^{<}\\ \bm{\Sigma}_{\vec{i}} \text{ not diagonal}}}\ 1 \leq \frac{\delta(\delta+1)}{2}(\kappa-1)\binom{p}{\delta}\leq
\frac{(\delta+1)}{2((\delta-1)!)} p^{\delta} (\kappa-1).
$$
\end{lem}
\begin{proof}
Note that
$$
\sum_{\substack{\vec{i}\in C_\delta^{<}\\ \bm{\Sigma}_{\vec{i}} \text{ diagonal}}}\ 1 \geq \frac{1}{\delta!}p(p-\kappa)\ldots(p-\delta\kappa),
$$
where the $\frac{1}{\delta!}$ is due to in our definition $\vec{i}$ the index $i_1<\ldots <i_{\delta}$ are sorted. Then
\begin{align*}
\sum_{\substack{\vec{i}\in C_\delta^{<}\\ \bm{\Sigma}_{\vec{i}} \text{ not diagonal}}}\ 1 \leq \binom{p}{1}\binom{p-1}{\delta} -
\frac{1}{\delta!}p\prod_{\ell=1}^{\delta}(p-\ell\kappa)\leq \frac{\delta(\delta+1)}{2}(\kappa-1)\binom{p}{\delta},
\end{align*}
where the last inequality follows from Lemma \ref{basine} \ref{item:basinei}.
\begin{comment}
\begin{align*}
&\sum_{\substack{\vec{i}\in C_\delta^{<}\\ \bm{\Sigma}_{\vec{i}} \text{ not diagonal}}}\ 1 \\ =&\binom{p}{1}\binom{p-1}{\delta}-\frac{1}{\delta !}\sum_{i_0=1}^p\ \sum_{i_1\in [p]\backslash \mathcal{NZ}(i_0)}\ \sum_{i_2\in [p]\backslash \bigcup\limits_{\ell=0}^1\mathcal{NZ}(i_\ell)}\cdots \sum_{i_\delta\in [p]\backslash \bigcup\limits_{\ell=0}^{\delta-1}\mathcal{NZ}(i_\ell)}1\\
=&\frac{1}{\delta !}\sum_{i_0=1}^p\ \sum_{i_1\in [p]\backslash \{i_0\}}\ \sum_{i_2\in [p]\backslash \bigcup\limits_{\ell=0}^1\{i_\ell\}}\cdots \sum_{i_\delta\in [p]\backslash \bigcup\limits_{\ell=0}^{\delta-1}\{i_\ell\}}1\\
&-\frac{1}{\delta !}\sum_{i_0=1}^p\ \sum_{i_1\in [p]\backslash \mathcal{NZ}(i_0)}\ \sum_{i_2\in [p]\backslash \bigcup\limits_{\ell=0}^1\mathcal{NZ}(i_\ell)}\cdots \sum_{i_\delta\in [p]\backslash \bigcup\limits_{\ell=0}^{\delta-1}\mathcal{NZ}(i_\ell)}1\\
=&\frac{1}{\delta!}\sum_{m=1}^\delta\left( \sum_{i_0=1}^p\ \sum_{i_1\in [p]\backslash \{i_0\}}\cdots \sum_{i_{m-1}\in [p]\backslash \bigcup\limits_{\ell=0}^{m-2}\{i_\ell\}}\ \sum_{i_{m}\in \bigcup\limits_{\ell=0}^{m-1}\mathcal{NZ}(i_\ell)\backslash \bigcup\limits_{\ell=0}^{m-1}\{i_\ell\}}\ \sum_{i_{m+1}\in [p]\backslash \bigcup\limits_{\ell=0}^{m}\mathcal{NZ}(i_\ell)}\cdots \sum_{i_\delta\in [p]\backslash \bigcup\limits_{\ell=0}^{\delta-1}\mathcal{NZ}(i_\ell)}1 \right)\\
\leq & \frac{1}{\delta!}\sum_{m=1}^\delta\left( p\ p^{m-1} m(\kappa-1) p^{\delta-m} \right)\\
= & \frac{(\delta+1)}{2((\delta-1)!)} p^{\delta} (\kappa-1).
\end{align*}
\end{comment}
\end{proof}
Note $\kappa=1$, the Lemma \ref{lem:numofnondia} shows $\sum\limits_{\substack{\vec{i}\in C_\delta^{<}\\ \bm{\Sigma}_{\vec{i}} \text{ not diagonal}}}\ 1 = 0$, which means $\bm{\Sigma}$ is diagonal matrix. Next we present a lemma to bound $
\sum_{\vec{i}\in C_\ell^<}\mu_{n,\ell+1}\left(\bm{\Sigma}_{\vec{i}}\right)
$.
\begin{lem} \label{lem:summuupp}
$$
\sum_{\vec{i}\in C_\ell^<}\mu_{n,\ell+1}\left(\bm{\Sigma}_{\vec{i}}\right) \leq \frac{p^{\ell+1}}{\ell!}\left(1+\ell^2\mu_{n,\ell+1}(\bm{\Sigma})\frac{\kappa-1}{p}\right).
$$
\end{lem}
\begin{proof}
\begin{align*}
\sum_{\vec{i}\in C_\ell^<}\mu_{n,\ell+1}(\bm{\Sigma}_{\vec{i}}) =& \sum_{\substack{\vec{i}\in C_\ell^<\\ \bm{\Sigma}_{\vec{i} }\text{ diagonal }}}1+ \sum_{\substack{\vec{i}\in C_\ell^<\\ \bm{\Sigma}_{\vec{i} } \text{ not diagonal }}} \mu_{n,\ell+1}(\bm{\Sigma}_{\vec{i}})\\
\leq & \binom{p}{1}\binom{p-1}{\ell} + \mu_{n,\ell+1}(\bm{\Sigma}) \sum_{\substack{\vec{i}\in C_\ell^<\\ \bm{\Sigma}_{\vec{i} } \text{ not diagonal }}}1 \\
\leq & \frac{p^{\ell+1}}{\ell!}\left(1+\ell^2\mu_{n,\ell+1}(\bm{\Sigma})\frac{\kappa-1}{p}\right),
\end{align*}
where the first inequality follows from the Lemma \ref{lem:densitybou} \ref{item:densityboua}, and the second inequality follows from Lemma \ref{lem:numofnondia}.
\end{proof}
\begin{lem}\label{lem:eleine}
Let $X\sim \mathcal{VE}(\bm{\mu}, \bm{\Sigma}, g)$. Let $\{i_q\}_{q=0}^\alpha, \{j_q\}_{q=0}^\beta\subset [p]$ be respectively a sequence of $\alpha+1$ and $\beta+1$ distinct integers. Let $m\in [\min\{\alpha,\beta\}]$. Suppose $i_q=j_q$ for $q\in [m]$ and $i_q \not= j_{q'}$ for $q,q'\not\in [m]$. Denote $\mathcal{I} = \bigcup_{q=0}^\alpha \{i_q\} \bigcup \left(\bigcup_{q'=0}^{\beta} \{j_q'\}\right)$ and then $|\mathcal{I}|=\alpha+\beta-m+2$.
\begin{enumerate}[label=(\alph*)]
\item \label{item:eleineb}
Then
\begin{equation}
\mathbb{E} \left(\prod\limits_{q=1}^\alpha\Phi_{i_0i_q}^{(\bm{R})}\right) \left(\prod\limits_{q'=1}^\beta\Phi_{j_0j_{q'}}^{(\bm{R})}\right) \leq
\mu_{n,|\mathcal{I}|}\left(\bm{\Sigma}_{\mathcal{I}}\right) (2P_n(r_\rho))^{\alpha+\beta-m} \left(2P_n(2r_\rho)\right).
\label{eqn:prodmupp2}
\end{equation}
\item \label{item:eleinec}
Then
\begin{align*}
\mathbb{E}\Phi_{i_0j_0}^{(\bm{R})} \left(\prod\limits_{q=1}^\alpha\Phi_{i_0i_q}^{(\bm{R})}\right) \left(\prod\limits_{q'=1}^\beta\Phi_{j_0j_{q'}}^{(\bm{R})}\right)
\leq &
\mu_{n,|\mathcal{I}|}\left(\bm{\Sigma}_{\mathcal{I}}\right) \left(2P_n(r_\rho)\right)^{\alpha+\beta-m+1} \sum_{\ell=2}^{m+2} \alpha_{n,m+1}(\ell,r_\rho)\\
\leq &
\mu_{n,|\mathcal{I}|}\left(\bm{\Sigma}_{\mathcal{I}}\right) (2P_n(r_\rho))^{\alpha+\beta-m+1}. \numberthis \label{eqn:prodmupp3}
\end{align*}
The inequality \eqref{eqn:prodmupp3} also holds with $m=0$.
\end{enumerate}
\end{lem}
\begin{proof}
\noindent (a)
By Lemma \ref{lem:densitybou} \ref{item:densitybouc}, it suffices to prove
\eqref{eqn:prodmupp2} without $\mu_{n,|\mathcal{I}|}\left(\bm{\Sigma}_{\mathcal{I}}\right)$ for the case that $\{\bm{u}_j\}$ for $j\in \mathcal{I}$ are independent $\operatorname{{\bf unif}}(S^{n-2})$. Conditioned on $\bm{u}_{i_0}$ and $\bm{u}_{j_0}$, $\{\Phi_{i_0i_q}^{(\bm{R})}\Phi_{j_0i_q}^{(\bm{R})}\}_{q=1}^m$ are i.i.d., $\{\Phi_{i_0i_q}^{(\bm{R})}\}_{q=m+1}^\alpha\bigcup\{\Phi_{j_0j_{q'}}^{(\bm{R})}\}_{q'=m+1}^\beta$ are i.i.d. and moreover, every term in $\{\Phi_{i_0i_q}^{(\bm{R})}\Phi_{j_0i_q}^{(\bm{R})}\}_{q=1}^m$ is independent of every term in $\{\Phi_{i_0i_q}^{(\bm{R})}\}_{q=m+1}^\alpha\bigcup\{\Phi_{j_0j_{q'}}^{(\bm{R})}\}_{q'=m+1}^\beta$. Thus
\begin{align*}
&\mathbb{E}\left[ \left.\left(\prod\limits_{q=1}^\alpha\Phi_{i_0i_q}^{(\bm{R})}\right) \left(\prod\limits_{q'=1}^\beta\Phi_{j_0j_{q'}}^{(\bm{R})}\right) \right|\bm{u}_{i_0},\bm{u}_{j_0}\right]\\
=&
\mathbb{E}\left[\left.\prod_{q=1}^m \Phi_{i_0i_q}^{(\bm{R})}\Phi_{j_0i_q}^{(\bm{R})}\right|\bm{u}_{i_0},\bm{u}_{j_0}\right] \left(\mathbb{E}\left[ \left.\Phi_{i_0i_\alpha}^{(\bm{R})}{1}(\alpha>m)+\Phi_{j_0j_\beta}^{(\bm{R})}{1}(\alpha=m,\beta>m) \right|\bm{u}_{i_0},\bm{u}_{j_0}\right]\right)^{\alpha+\beta-2m}\\
=&
\mathbb{E}\left[\left. \prod_{q=1}^m \Phi_{i_0i_q}^{(\bm{R})}\Phi_{j_0i_q}^{(\bm{R})}\right|\bm{u}_{i_0},\bm{u}_{j_0}\right] \left(2P_n(r_\rho)\right)^{\alpha+\beta-2m} \numberthis \label{eqn:conine2} \\
=&
\left(\mathbb{E}\left[\left. \Phi_{i_0i_q}^{(\bm{R})}\Phi_{j_0i_q}^{(\bm{R})}\right|\bm{u}_{i_0},\bm{u}_{j_0}\right]\right)^m \left(2P_n(r_\rho)\right)^{\alpha+\beta-2m} \numberthis \label{eqn:conine}
\end{align*}
where for the first equality the convention $0^0=1$ is used if $\alpha=\beta=m$. Notice \eqref{eqn:conine} also holds for $m=0$.
Denote $ \overline{\operatorname{SC}}(r,\bm{w}) = \operatorname{SC}(r,\bm{w}) \cup \operatorname{SC}(r,-\bm{w})$. Then conditioned on $\bm{u}_{i_0}$ and $\bm{u}_{j_0}$,
\begin{align*}
&\Phi_{i_0i_1}^{(\bm{R})}\Phi_{j_0i_1}^{(\bm{R})}\\
=& 1\left(\bm{u}_{i_1}\in \overline{\operatorname{SC}}(r,\bm{u}_{i_0}) \cap \overline{\operatorname{SC}}(r,\bm{u}_{j_0})\right)\\ =&1(\|\bm{u}_{i_0}-\bm{u}_{j_0}\|_2\leq 2r_\rho\text{ or } \|\bm{u}_{i_0}+\bm{u}_{j_0}\|_2\leq 2r_\rho) 1\left(\bm{u}_{i_1}\in \overline{\operatorname{SC}}(r,\bm{u}_{i_0}) \cap \overline{\operatorname{SC}}(r,\bm{u}_{j_0})\right),
\end{align*}
where the last equality follows by noticing that $\overline{\operatorname{SC}}(r_\rho,\bm{u}_{i_0}) \cap \overline{\operatorname{SC}}(r_\rho,\bm{u}_{j_0})$ is non-empty only when $\|\bm{u}_{i_0}-\bm{u}_{j_0}\|_2\leq 2r_\rho$ or $\|\bm{u}_{i_0}+\bm{u}_{j_0}\|_2\leq 2r_\rho$. Plugging the above inequality into \eqref{eqn:conine}, we obtain
\begin{align*}
&\mathbb{E}\left[ \left.\left(\prod\limits_{q=1}^\alpha\Phi_{i_0i_q}^{(\bm{R})}\right) \left(\prod\limits_{q'=1}^\beta\Phi_{j_0j_{q'}}^{(\bm{R})}\right) \right|\bm{u}_{i_0},\bm{u}_{j_0}\right]\\
= & 1(\|\bm{u}_{i_0}-\bm{u}_{j_0}\|_2\leq 2r_\rho\text{ or } \|\bm{u}_{i_0}+\bm{u}_{j_0}\|_2\leq 2r_\rho)\left(\mathbb{E}\left[\left.\Phi_{i_0i_1}^{(\bm{R})}\Phi_{j_0i_1}^{(\bm{R})}\right|\bm{u}_{i_0},\bm{u}_{j_0}\right]\right)^m \left(2P_n(r_\rho)\right)^{\alpha+\beta-2m}\\
\leq & 1(\|\bm{u}_{i_0}-\bm{u}_{j_0}\|_2\leq 2r_\rho\text{ or } \|\bm{u}_{i_0}+\bm{u}_{j_0}\|_2\leq 2r_\rho)\left(2P_n(r_\rho)\right)^m \left(2P_n(r_\rho)\right)^{\alpha+\beta-2m}.
\end{align*}
The result then follows by taking expectation w.r.t. $\bm{u}_{i_0}$ and $\bm{u}_{j_0}$.\\
\noindent (b)
Similar to the proof of \ref{item:eleineb}, it suffices to prove
\eqref{eqn:prodmupp3} without $\mu_{n,|\mathcal{I}|}\left(\bm{\Sigma}_{\mathcal{I}}\right)$ for the case $\bm{u}_j$ for $j\in \mathcal{I}$ are independent $\operatorname{{\bf unif}}(S^{n-2})$. Conditioned on $\bm{u}_{i_0}$ and $\bm{u}_{j_0}$,
\begin{align*}
&\mathbb{E} \left[\Phi_{i_0j_0}^{(\bm{R})} \left.\left(\prod\limits_{q=1}^\alpha\Phi_{i_0i_q}^{(\bm{R})}\right) \left(\prod\limits_{q'=1}^\beta\Phi_{j_0j_{q'}}^{(\bm{R})}\right) \right|\bm{u}_{i_0},\bm{u}_{j_0}\right]\\
=&
\Phi_{i_0j_0}^{(\bm{R})} \mathbb{E}\left[ \prod_{q=1}^m \left.\Phi_{i_0i_q}^{(\bm{R})}\Phi_{j_0i_q}^{(\bm{R})}\right|\bm{u}_{i_0},\bm{u}_{j_0}\right] \left(2P_n(r_\rho)\right)^{\alpha+\beta-2m}.
\end{align*}
where the equality follows from \eqref{eqn:conine2}. Then
\begin{align*}
&\mathbb{E} \left[\Phi_{i_0j_0}^{(\bm{R})} \left(\prod\limits_{q=1}^\alpha\Phi_{i_0i_q}^{(\bm{R})}\right) \left(\prod\limits_{q'=1}^\beta\Phi_{j_0j_{q'}}^{(\bm{R})}\right)\right]\\
= &
\mathbb{E}\left[ \prod_{q=0}^m \Phi_{i_0i_q}^{(\bm{R})} \prod_{q'=0}^{m} \Phi_{j_0 i_{q'}}^{(\bm{R})} \right] \left(2P_n(r_\rho)\right)^{\alpha+\beta-2m}\\
= &
P\left( \prod_{q=0}^m \Phi_{i_0i_q}^{(\bm{R})} =1 \right) P\left( \left. \prod_{q'=0}^{m} \Phi_{j_0 i_{q'}}^{(\bm{R})} = 1\right|\prod_{q=0}^m \Phi_{i_0i_q}^{(\bm{R})} =1 \right) \left(2P_n(r_\rho)\right)^{\alpha+\beta-2m} \\
\overset{(*)}{\leq} &
\left(2P_n(r_\rho)\right)^{\alpha+\beta-m+1} \sum_{\ell=2}^{m+2} \alpha_{n,m+1}(\ell,r_\rho)\\
\leq &
\left(2P_n(r_\rho)\right)^{\alpha+\beta-m+1},
\end{align*}
where step $(*)$ follows from the definition of $\alpha_{n,\delta}(\ell,r_\rho)$ in \eqref{eqn:alphandeltarho}. Notice \eqref{eqn:conine} also holds for $m=0$.
\end{proof}
\subsubsection{Lemmas on double summations}
\label{sec:quadraticterms}
Denote $\vec{i}\cup \vec{j} = \left[\vec{i}\right]\bigcup\left[\vec{j}\right] $ for any $\vec{i}\in C_q^<$ and any $\vec{j}\in C_\delta^<$. Consider any $\theta_{\vec{i},\vec{j}}$ that is a non-negative function of $\bm{u}_{\ell}$ for $\ell\in \vec{i}\cup\vec{j}$ defined for $\vec{i}\in C_q^<$ and $\vec{j}\in C_\delta^<$ with $1\leq \delta\leq q\leq p-1$. In this section an upper bound on $\mathbb{E} \sum_{\vec{i}\in C_q^< }\sum_{\vec{j}\in C_\delta^<} \theta_{\vec{i},\vec{j}}$ is presented. The results in this subsection will be used in the proofs of Proposition \ref{prop:edgecor} and Proposition \ref{prop:allcloseL2}.
For $i\in [p]$, let
\begin{equation}
\mathcal{NZ}(i):=\{m\in [p]: \bm{\Sigma}_{im}\not =0\} \label{eqn:NZdef}
\end{equation}
denote the index of the variables that has non zero correlation with the $i$-th variable. For $\vec{i}\in C_q^<$, define $\mathcal{NZ}\left(\vec{i}\right) := \bigcup\limits_{\ell=0}^q\mathcal{NZ}(i_{\ell}) $. Since $\bm{\Sigma}$ is row-$\kappa$ sparse, for any $\vec{i}\in C_q^<$, $\left|\mathcal{NZ}\left(\vec{i}\right)\right|\leq (q+1)\kappa$, and
\begin{equation}
p_{\vec{i}}:=\left|[p]\backslash \mathcal{NZ}\left(\vec{i}\right)\right|\geq p-(q+1)\kappa. \label{eqn:p'upperbounew}
\end{equation}
Note that $p_{\vec{i}}$ denotes the number of variables that are independent of variables with index in $[\vec{i}]$.
\\
For $\vec{i}\in C_q^<$, define
\begin{align}
J_{\vec{i}} & := \left\{\vec{j}\in C_\delta^<: \bigcup_{\ell=0}^\delta \{j_\ell\} \subset \bigcup_{\ell=0}^q \{i_\ell\} \right\}\label{eqn:Sidefgeneral},\\
T_{\vec{i}} & :=\left\{\vec{j}\in C_\delta^<: \left(\bigcup_{\ell=0}^\delta \{j_\ell\} \right)\bigcap \left( \mathcal{NZ}\left(\vec{i}\right) \right) = \emptyset \right\}, \label{eqn:Tidef} \\
N_{\vec{i}} & := C_q^< \backslash J_{\vec{i}} \backslash T_{\vec{i}}. \label{eqn:Nidef}
\end{align}
Here $J_{\vec{i}}$ is the set of indices in $C_\delta^<$ consisting of coordinates as subsets of $\left[\vec{i}\right]$; $T_{\vec{i}}$ is the set of indices in $C_\delta^<$ consisting of coordinates outside the neighborhood of $\vec{i}$; $N_{\vec{i}}$ is the set of "correlated but not highly correlated" indices in $C_\delta^<$, i.e. the set of indices of which at least one coordinate is in the neighborhood of $\vec{i}$, but excluding those sets of indices of which the set of coordinates are subsets as that of $\vec{i}$.
The strategy is to decompose
$$
\mathbb{E} \sum_{\vec{i}\in C_q^< }\sum_{\vec{j}\in C_\delta^<} \theta_{\vec{i},\vec{j}} = \mathbb{E} \sum_{\vec{i}\in C_q^< } \sum_{\vec{j}\in J_{\vec{i}}}\theta_{\vec{i},\vec{j}} + \mathbb{E} \sum_{\vec{i}\in C_q^< } \sum_{\vec{j}\in T_{\vec{i}}} \theta_{\vec{i},\vec{j}} +\mathbb{E} \sum_{\vec{i}\in C_q^< } \sum_{\vec{j}\in N_{\vec{i}}}\theta_{\vec{i},\vec{j}}
$$
and bound each of the three terms.
The next result is an upper bound on the first two terms.
\begin{lem}
\label{lem:quadraticsum}
Let $p\geq n \geq 4$ and $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$. Suppose $\bm{\Sigma}$ is row-$\kappa$ sparse.
Consider any $\theta_{\vec{i},\vec{j}}$ that is a non-negative function of $\bm{u}_{\ell}$ for $\ell\in \vec{i}\cup\vec{j}$ defined for $\vec{i}\in C_q^<$ and $\vec{j}\in C_\delta^<$ with $1\leq \delta\leq q\leq p-1$.
\begin{enumerate}[label=(\alph*)]
\item \label{item:quadraticsuma}
Suppose there exist positive constants $a,z$ such that $\mathbb{E} \theta_{\vec{i},\vec{j}} \leq \mu_{n,q+1}(\bm{\Sigma}_{\vec{i}}) a z^q$ for any $\vec{j}\in J_{\vec{i}}$. Then
\begin{align*}
\sum_{\vec{i}\in C_q^<} \sum_{\vec{j}\in J_{\vec{i}}} \mathbb{E} \theta_{\vec{i},\vec{j}} \leq & ap (pz)^q \frac{q+1}{\delta!(q-\delta)!} \left(1+q^2\mu_{n,q+1}(\bm{\Sigma})\frac{\kappa-1}{p}\right)
\end{align*}
\item \label{item:quadraticsumb}
Suppose there exist positive constants $a,z$ such that $\mathbb{E} \theta_{\vec{i},\vec{j}} \leq \mu_{n,q+\delta+2}(\bm{\Sigma}_{\vec{i}\cup\vec{j}}) a z^{q+\delta}$ for any $\vec{j}\in T_{\vec{i}}$.
Then
\begin{align*}
\sum_{\vec{i}\in C_q^<} \sum_{\vec{j}\in T_{\vec{i}}} \mathbb{E} \theta_{\vec{i},\vec{j}} \leq& a p^2 (pz)^{q+\delta} \frac{3}{\delta! (q-1)!} \left(1+ \mu_{n,q+\delta+2}(\bm{\Sigma}) \frac{\kappa-1}{p} \right) .
\end{align*}
\end{enumerate}
\end{lem}
\begin{proof} (a)
Since $|J_{\vec{i}}|=\binom{q+1}{1}\binom{q}{\delta}$
\begin{align*}
\sum_{\vec{i}\in C_q^<} \sum_{\vec{j}\in J_{\vec{i}}} \mathbb{E} \theta_{\vec{i},\vec{j}}
\leq & a z^q \binom{q+1}{1}\binom{q}{\delta} \sum_{\vec{i}\in C_q^<}\mu_{n,q+1}(\bm{\Sigma}_{\vec{i}})\\
\leq & a p (pz)^q \frac{q+1}{\delta!(q-\delta)!} \left(1+q^2\mu_{n,q+1}(\bm{\Sigma})\frac{\kappa-1}{p}\right),
\end{align*}
where the last step follows from Lemma \ref{lem:summuupp}.\\
\noindent
(b)
\begin{align*}
&\sum_{\vec{i}\in C_q^<} \sum_{\vec{j}\in T_{\vec{i}}} \mathbb{E} \theta_{\vec{i},\vec{j}} \\
\leq &
az^{q+\delta} \sum_{\vec{i}\in C_q^<} \sum_{\vec{j}\in T_{\vec{i}}} \mu_{n,q+\delta+2}(\bm{\Sigma}_{\vec{i}\cup \vec{j}}) \\
\leq &
az^{q+\delta}
\left(\sum_{\substack{\vec{i}\in C_q^<\\ \bm{\Sigma}_{\vec{i}} \text{ diagonal }}} \sum_{\substack{\vec{j}\in T_{\vec{i}}\\ \bm{\Sigma}_{\vec{j}} \text{ diagonal } }} 1
+ \mu_{n,q+\delta+2}(\bm{\Sigma}) \sum_{\substack{\vec{i}\in C_q^<\\ \bm{\Sigma}_{\vec{i}} \text{ diagonal }}} \sum_{\substack{\vec{j}\in T_{\vec{i}}\\ \bm{\Sigma}_{\vec{j}} \text{ not diagonal } }}1+\right.\\
&\quad \left.\mu_{n,q+\delta+2}(\bm{\Sigma}) \sum_{\substack{\vec{i}\in C_q^<\\ \bm{\Sigma}_{\vec{i}} \text{ not diagonal }}} \sum_{\vec{j}\in T_{\vec{i}}} 1\right)
\\
\leq & az^{q+\delta}
\left(\binom{p}{1}\binom{p-1}{q} \binom{p}{1}\binom{p-1}{\delta}
+ \mu_{n,q+\delta+2}(\bm{\Sigma}) \binom{p}{1}\binom{p-1}{q} \sum_{\substack{\vec{j}\in C_\delta^<\\ \bm{\Sigma}_{\vec{j}} \text{ not diagonal } }}1+ \right. \\
&\quad \left.\mu_{n,q+\delta+2}(\bm{\Sigma}) \sum_{\substack{\vec{i}\in C_q^<\\ \bm{\Sigma}_{\vec{i}} \text{ not diagonal }}} \binom{p}{1}\binom{p-1}{\delta} \right)\\
\leq &
a p^2 (pz)^{q+\delta} \frac{3}{\delta! (q-1)!} \left(1+ \mu_{n,q+\delta+2}(\bm{\Sigma}) \frac{\kappa-1}{p} \right) ,
\end{align*}
where the second inequality follows from that for $\vec{j}\in T_{\vec{i}}$, $\bm{\Sigma}_{\vec{i}\cup \vec{j} }$ is diagonal if and only if $\bm{\Sigma}_{\vec{i}}$ and $\bm{\Sigma}_{\vec{j}}$ are both diagonal; and the last step follows from Lemma \ref{lem:numofnondia}.\\
\end{proof}
To control $\mathbb{E} \sum_{\vec{i}\in C_q^< } \sum_{\vec{j}\in N_{\vec{i}}}\theta_{\vec{i},\vec{j}}$, we further partition $N_{\vec{i}}$ into $6$ subsets as follows. For $\vec{i}\in C_q^<$ with $q\geq \delta$, define
\begin{align*}
\mathcal{K}_1\left(\vec{i}\right) & := \left\{\vec{j}\in N_{\vec{i}}:j_0=i_0 \right\},\\
\mathcal{K}_2\left(\vec{i}\right) & := \left\{\vec{j}\in N_{\vec{i}}:j_0\not=i_0, j_0\in \bigcup_{\ell=1}^q \{i_\ell\}, i_0\in \bigcup_{\ell=1}^\delta\{j_\ell\}\right\}, \\
\mathcal{K}_3\left(\vec{i}\right) & := \left\{\vec{j}\in N_{\vec{i}}:j_0\not=i_0, j_0\not\in \bigcup_{\ell=1}^q\{i_\ell\}, i_0\in \bigcup_{\ell=1}^\delta\{j_\ell\}\right\},\\
\mathcal{K}_4\left(\vec{i}\right) & := \left\{\vec{j}\in N_{\vec{i}}:j_0\not=i_0, j_0\in \bigcup_{\ell=1}^q \{i_\ell\}, i_0\not\in \bigcup_{\ell=1}^\delta\{j_\ell\} \right\},\\
\mathcal{K}_5\left(\vec{i}\right) & := \left\{\vec{j}\in N_{\vec{i}}:j_0\not=i_0, j_0\not\in \bigcup_{\ell=1}^q\{i_\ell\}, i_0\not\in \bigcup_{\ell=1}^\delta\{j_\ell\}, \left|\left(\bigcup_{\ell=1}^q\{i_\ell\}\right)\bigcap
\left(\bigcup_{\ell=1}^\delta \{j_\ell\}\right) \right| \geq 1 \right\},\\
\mathcal{K}_6\left(\vec{i}\right) & := \left\{\vec{j}\in N_{\vec{i}}:j_0\not=i_0, j_0\not\in \bigcup_{\ell=1}^q\{i_\ell\}, i_0\not\in \bigcup_{\ell=1}^\delta\{j_\ell\}, \left(\bigcup_{\ell=1}^q\{i_\ell\}\right) \bigcap \left(\bigcup_{\ell=1}^\delta \{j_\ell\}\right) = \emptyset \right\}.
\end{align*}
Then $N_{\vec{i}}= \cup_{w=1}^6\mathcal{K}_w\left(\vec{i}\right)$. Let $D_{\vec{i}}^m=\{\vec{j}\in N_{\vec{i}}: \left| \left(\cup_{\ell=1}^q\{i_\ell\}\right) \bigcap \left(\cup_{\ell=1}^\delta\{j_\ell\}\right) \right|=m \}.$
We are now in a good position to present
a lemma on $\mathbb{E} \sum_{\vec{i}\in C_q^< } \sum_{\vec{j}\in N_{\vec{i}}} \theta_{\vec{i},\vec{j}}$.
\begin{lem}
\label{lem:b2}
Let $p\geq n \geq 4$ and $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$. Suppose $\bm{\Sigma}$ is row-$\kappa$ sparse.
Consider any $\theta_{\vec{i},\vec{j}}$ that is a non-negative function of $\bm{u}_{\ell}$ for $\ell\in \vec{i}\cup\vec{j}$ defined for $\vec{i}\in C_q^<$ and $\vec{j}\in C_\delta^<$ with $1\leq \delta\leq q\leq p-1$. Suppose there exist positive constants $a,b,z$ such that $\theta_{\vec{i},\vec{j}}$ satisfies:
\begin{align*}
\mathbb{E} \theta_{\vec{i},\vec{j}} \leq \mu_{n,|\vec{i}\cup \vec{j}|}(\bm{\Sigma}_{\vec{i}\cup \vec{j}}) az^{q+\delta-m}, &\quad \forall\ \vec{j}\in \mathcal{K}_w\left(\vec{i}\right) \cap D_{\vec{i}}^m,\ \forall\ 0\leq m\leq \delta-1,\ \forall w\in\{1,3,4\}; \\
\mathbb{E} \theta_{\vec{i},\vec{j}} \leq \mu_{n,|\vec{i}\cup \vec{j}|}(\bm{\Sigma}_{\vec{i}\cup \vec{j}}) az^{q+\delta-m-1}, &\quad \forall\ \vec{j}\in \mathcal{K}_2\left(\vec{i}\right) \cap D_{\vec{i}}^m,\ \forall\ 0\leq m\leq \delta-2; \\
\mathbb{E} \theta_{\vec{i},\vec{j}} \leq \mu_{n,|\vec{i}\cup \vec{j}|}(\bm{\Sigma}_{\vec{i}\cup \vec{j}}) abz^{q+\delta-m}, &\quad \forall\ \vec{j}\in \mathcal{K}_5\left(\vec{i}\right) \cap D_{\vec{i}}^m,\ \forall\ 1\leq m\leq \delta;\\
\mathbb{E} \theta_{\vec{i},\vec{j}} \leq \mu_{n,|\vec{i}\cup \vec{j}|}(\bm{\Sigma}_{\vec{i}\cup \vec{j}}) az^{q+\delta}, &\quad \forall\ \vec{j}\in \mathcal{K}_6\left(\vec{i}\right).
\end{align*}
Then
\begin{align*}
\sum_{\vec{i}\in C_q^<} \sum_{\vec{j}\in N_{\vec{i}}} \mathbb{E} \theta_{\vec{i},\vec{j}} \leq& ap(pz)^{q+1}\left(1+ \mu_{n,q+\delta+1}\left(\bm{\Sigma}\right) (3q^2)\frac{\kappa-1}{p}\right)\left(1+pz\right)^{\delta-1} \delta\frac{4+b/z}{(\delta-1)!}+\\
&\quad\quad\quad ap^2 (pz)^{q+\delta} \frac{(\delta+1)(q+1)}{\delta! q!} \mu_{n,q+\delta+2}\left(\bm{\Sigma}\right)\frac{\kappa-1}{p}.
\end{align*}
\end{lem}
\begin{proof}
Since
\begin{align*}
N_{\vec{i}} = \bigcup_{w=1}^6\mathcal{K}_w\left(\vec{i}\right),
\end{align*}
\begin{equation}
\sum_{\vec{i}\in C_q^<} \sum_{\vec{j}\in N_{\vec{i}}} \mathbb{E} \theta_{\vec{i},\vec{j}} = \sum_{w=1}^6 \sum_{\vec{i}\in C_q^<}I_w\left(\vec{i}\right), \label{eqn:b2dec}
\end{equation}
with
$$
I_w\left(\vec{i}\right) := \sum_{\vec{j}\in \mathcal{K}_w\left(\vec{i}\right)} \mathbb{E}\theta_{\vec{i},\vec{j}}.
$$
\noindent \textbf{Case 1}: $p \geq q+\delta +2 $\\
Obviously $\mathcal{K}_1\left(\vec{i}\right)=\bigcup_{m=0}^{\delta-1} \left(\mathcal{K}_1\left(\vec{i}\right)\cap D_{\vec{i}}^m \right)$. Then for any $\vec{i}\in C_q^<$ satisfying $\bm{\Sigma}_{\vec{i}}$ diagonal,
\begin{align*}
&\left|\vec{j}\in \mathcal{K}_1\left(\vec{i}\right) \bigcap D_{\vec{i}}^m : \bm{\Sigma}_{\vec{i}\bigcup\vec{j}} \text{ diagonal } \right| \\
= &
\binom{q}{m} \frac{1}{(\delta-m)!} \sum_{ j_1\in [p]\backslash \mathcal{NZ}\left(\vec{i}\right) }\ \sum_{ \substack{ j_2\in [p]\backslash \mathcal{NZ}\left(\vec{i}\right) \\ j_2\not \in \mathcal{NZ}(j_1) } } \ \ \sum_{ \substack{ j_3\in [p]\backslash \mathcal{NZ}\left(\vec{i}\right) \\ j_3\not \in \cup_{\ell=1}^2 \mathcal{NZ}(j_\ell) } } \cdots \sum_{ \substack{ j_{\delta-m}\in [p]\backslash \mathcal{NZ}\left(\vec{i}\right) \\ j_{\delta-m}\not \in \cup_{\ell=1}^{\delta-m-1} \mathcal{NZ}(j_\ell) } } 1 \\
\geq &
\binom{q}{m} \frac{1}{(\delta-m)!}\prod_{\ell=0}^{\delta-m-1}\left(p_{\vec{i}}-\ell \kappa \right), \numberthis \label{eqn:K1mdia}
\end{align*}
where in the first inequality we assume without loss of generality that the components of $\vec{j}$ distinct from $\vec{i}$ are $j_1,j_2,\cdots,j_{\delta-m}$.
Then,
\begin{align*}
&\left|\vec{j}\in \mathcal{K}_1\left(\vec{i}\right) \bigcap D_{\vec{i}}^m : \bm{\Sigma}_{\vec{i}\bigcup\vec{j}} \text{ not diagonal } \right|\\
\leq &\binom{q}{m} \binom{p-1-q}{\delta-m} -\binom{q}{m} \frac{1}{(\delta-m)!}\prod_{\ell=0}^{\delta-m-1}\left(p_{\vec{i}}-\ell \kappa \right)\\
= &
\frac{\binom{q}{m}}{(\delta-m)!}\left(\prod_{\ell=0}^{\delta-m-1}\left(p-1-q-\ell \right) -\prod_{\ell=0}^{\delta-m-1}\left(p_{\vec{i}}-\ell \right)+\prod_{\ell=0}^{\delta-m-1}\left(p_{\vec{i}}-\ell \right) -\prod_{\ell=0}^{\delta-m-1}\left(p_{\vec{i}}-\ell \kappa \right) \right) \\
\leq &
\binom{q}{m} \frac{1}{(\delta-m)!}(\delta-m)p^{\delta-m-1}(q+\delta)(\kappa-1), \numberthis \label{eqn:K1mnotdia}
\end{align*}
where the first inequality follows from \eqref{eqn:K1mdia},
and the second inequality follows from Lemma \ref{basine} \ref{item:basineh}, \ref{item:basinei} and \eqref{eqn:p'upperbou}.
Then
\begin{align*}
&\sum_{\vec{i}\in C_q^<}I_1 \left(\vec{i}\right) \\
=&
\sum_{m=0}^{\delta-1}\ \left( \sum_{ \substack{\vec{i}\in C_q^<\\ \bm{\Sigma}_{\vec{i}} \text{ diagonal}}} \ \sum_{ \substack{\vec{j}\in \mathcal{K}_1\left(\vec{i}\right)\bigcap D_{\vec{i}}^m\\ \bm{\Sigma}_{\vec{i}\cup\vec{j} \text{ diagonal}}}}
+
\sum_{ \substack{\vec{i}\in C_q^<\\ \bm{\Sigma}_{\vec{i}} \text{ diagonal}}} \ \sum_{ \substack{\vec{j}\in \mathcal{K}_1\left(\vec{i}\right)\bigcap D_{\vec{i}}^m\\ \bm{\Sigma}_{\vec{i}\cup\vec{j} \text{ not diagonal}}}}
+
\sum_{ \substack{\vec{i}\in C_q^<\\ \bm{\Sigma}_{\vec{i}} \text{ not diagonal}}} \ \sum_{ \vec{j}\in \mathcal{K}_1\left(\vec{i}\right)\bigcap D_{\vec{i}}^m} \right) \mathbb{E}\theta_{\vec{i},\vec{j}}\\
\leq &
\sum_{m=0}^{\delta-1}\ \left( \sum_{ \substack{\vec{i}\in C_q^<\\ \bm{\Sigma}_{\vec{i}} \text{ diagonal}}} \ \sum_{ \substack{\vec{j}\in \mathcal{K}_1\left(\vec{i}\right)\bigcap D_{\vec{i}}^m\\ \bm{\Sigma}_{\vec{i}\cup\vec{j} \text{ diagonal}}}}
+\right.\\
&
\left.\mu_{n,q+\delta+1}\left(\bm{\Sigma}\right)\left(\sum_{ \substack{\vec{i}\in C_q^<\\ \bm{\Sigma}_{\vec{i}} \text{ diagonal}}} \ \sum_{ \substack{\vec{j}\in \mathcal{K}_1\left(\vec{i}\right)\bigcap D_{\vec{i}}^m\\ \bm{\Sigma}_{\vec{i}\cup\vec{j} \text{ not diagonal}}}} +\sum_{ \substack{\vec{i}\in C_q^<\\ \bm{\Sigma}_{\vec{i}} \text{ not diagonal}}} \ \sum_{ \vec{j}\in \mathcal{K}_1\left(\vec{i}\right)\bigcap D_{\vec{i}}^m}\right) \right) a z^{q+\delta-m}\\
\leq &
\sum_{m=0}^{\delta-1}\ \left( \binom{p}{1}\binom{p-1}{q} \binom{q}{m}\binom{p-1-q}{\delta-m} \right.\\
&\left. \ \ \ \ \ +\mu_{n,q+\delta+1}\left(\bm{\Sigma}\right)\binom{p}{1}\binom{p-1}{q} \binom{q}{m} \frac{\delta-m}{(\delta-m)!}p^{\delta-m-1} (q+\delta)(\kappa-1) \right. \\
&\quad \quad \quad \left. +\mu_{n,q+\delta+1}\left(\bm{\Sigma}\right)\frac{(q+1)}{2((q-1)!)} p^{q} (\kappa-1) \binom{q}{m}\binom{p-1-q}{\delta-m} \right) a z^{q+\delta-m}\\
\leq &
\sum_{m=0}^{\delta-1}\ \left( \frac{1}{m!(\delta-m)!} +\mu_{n,q+\delta+1}\left(\bm{\Sigma}\right)\frac{\kappa-1}{p} \left( \frac{(q+\delta)(\delta-m)}{m!(\delta-m)!} + \frac{q(q+1)}{m!(\delta-m)!2}\right) \right) ap\left( pz\right)^{q+\delta-m}\\
\leq &
ap\left(pz\right)^{q+1} \left( 1 +\mu_{n,q+\delta+1}\left(\bm{\Sigma}\right)(3q^2) \frac{\kappa-1}{p} \right) \frac{1}{(\delta-1)!} \sum_{m=0}^{\delta-1} \frac{(\delta-1)!}{m!(\delta-1-m)!} \left( pz\right)^{\delta-1-m} \\
= &
ap\left(pz\right)^{q+1} \left( 1 +\mu_{n,q+\delta+1}\left(\bm{\Sigma}\right)(3q^2) \frac{\kappa-1}{p} \right) \frac{1}{(\delta-1)!} \left(1+pz\right)^{\delta-1}, \numberthis \label{eqn:I1temp}
\end{align*}
where the first inequality follows from $\mu_{n,q+\delta-m+1}\left(\bm{\Sigma}\right)\leq \mu_{n,q+\delta+1}\left(\bm{\Sigma}\right)$, and the second inequality follows from Lemma \ref{lem:numofnondia} and \eqref{eqn:K1mnotdia}.
Obviously,
\begin{align*}
\mathcal{K}_2\left(\vec{i}\right)= \bigcup_{m=0}^{\delta-2}\left(\mathcal{K}_2\left(\vec{i}\right) \bigcap D_{\vec{i}}^m \right) , & \quad \mathcal{K}_3\left(\vec{i}\right)= \bigcup_{m=0}^{\delta-1}\left(\mathcal{K}_3\left(\vec{i}\right) \bigcap D_{\vec{i}}^m \right) ,\\ \mathcal{K}_4\left(\vec{i}\right)= \bigcup_{m=0}^{\delta-1}\left(\mathcal{K}_4\left(\vec{i}\right) \bigcap D_{\vec{i}}^m \right), & \quad \mathcal{K}_5\left(\vec{i}\right) = \bigcup_{m=1}^\delta \left(\mathcal{K}_5\left(\vec{i}\right) \bigcap D_{\vec{i}}^m \right) .
\end{align*}
Then, following a similar analysis to $\mathcal{K}_1(\vec{i})$, additionally with Lemma \ref{lem:eleine}, we obtain
\begin{align*}
\sum_{\vec{i}\in C_q^<}I_2 \left(\vec{i}\right) \leq & ap\left(pz\right)^{q+1} \left( 1 +\mu_{n,q+\delta}\left(\bm{\Sigma}\right)(3q^2) \frac{\kappa-1}{p} \right) \frac{1}{(\delta-2)!} \left(1+pz\right)^{\delta-2}1(\delta \geq 2), \numberthis \label{eqn:I2temp} \\
\sum_{\vec{i}\in C_q^<}I_3 \left(\vec{i}\right) \leq & ap(pz)^{q+1} \left( 1 +\mu_{n,q+\delta+1}\left(\bm{\Sigma}\right)(3q^2) \frac{\kappa-1}{p} \right) \frac{1}{(\delta-1)!} \left(1+pz\right)^{\delta-1}, \numberthis \label{eqn:I3temp} \\
\sum_{\vec{i}\in C_q^<}I_4 \left(\vec{i}\right) \leq & ap\left(pz\right)^{q+1} \left( 1 +\mu_{n,2\delta+1}\left(\bm{\Sigma}\right)(3q^2) \frac{\kappa-1}{p} \right) \frac{1}{(\delta-1)!} \left(1+pz\right)^{\delta-1}, \numberthis \label{eqn:I4temp} \\
\sum_{\vec{i}\in C_q^<}I_5 \left(\vec{i}\right)
\leq & a\left(\frac{b}{z}\right) p\left(pz\right)^{q+1} \left( 1 +\mu_{n,q+\delta+1}\left(\bm{\Sigma}\right)(3\delta^2) \frac{\kappa-1}{p} \right) \frac{1}{(\delta-1)!} \left(1+pz\right)^{\delta-1}. \numberthis \label{eqn:I5temp}
\end{align*}
The detailed derivation of the above inequalities are omitted for the sake of brevity.
Observe that
$$
\mathcal{K}_6\left(\vec{i}\right) =
\left\{\vec{j}\in C_q^<: \left(\bigcup_{\ell=0}^q\{i_\ell\}\right) \bigcap \left(\bigcup_{\ell=0}^\delta \{j_\ell\}\right) = \emptyset , \ \exists \ell\in [\delta]\cup\{0\} \text{ such that } j_\ell \in \mathcal{NZ}\left(\vec{i}\right) \right\} .
$$
Then,
\begin{align*}
\left|\mathcal{K}_6\left(\vec{i}\right)\right|=& \binom{p-1-q}{1}\binom{p-2-q}{\delta}-\binom{p_{\vec{i}}}{1}\binom{p_{\vec{i}}-1}{\delta}\\
\leq & \frac{1}{\delta!}(\delta+1)p^{\delta}(q+1)(\kappa-1), \numberthis \label{eqn:k6bou}
\end{align*}
where the last inequality follows from Lemma \ref{basine} \ref{item:basineh} and \eqref{eqn:p'upperbou}.
Thus,
\begin{align*}
\sum_{\vec{i}\in C_q^<}I_6\left(\vec{i}\right)
\leq & \binom{p}{1}\binom{p-1}{q} \frac{1}{\delta!}(\delta+1)p^{\delta}(q+1)(\kappa-1) \mu_{n,q+\delta+2}\left(\bm{\Sigma}\right) az^{q+\delta},\\
\leq & ap^2 (pz)^{q+\delta} \frac{(\delta+1)(q+1)}{\delta! q!} \mu_{n,q+\delta+2}\left(\bm{\Sigma}\right)\frac{\kappa-1}{p}, \numberthis \label{eqn:I6temp}
\end{align*}
where the first inequality follows from \eqref{eqn:k6bou} and Lemma \ref{lem:densitybou} \ref{item:densityboua}.
\noindent \textbf{Case 2}: $p < q+\delta +2 $ \\
We have imposed the condition $p \geq 2\delta +2 $ to derive \eqref{eqn:I1temp}, \eqref{eqn:I2temp}, \eqref{eqn:I3temp}, \eqref{eqn:I4temp}, \eqref{eqn:I5temp} and \eqref{eqn:I6temp}. However,
one can verify directly that these inequalities also hold when $p < q+\delta +2 $. We omit these tedious verifications here and take it for granted \eqref{eqn:I1temp}, \eqref{eqn:I2temp}, \eqref{eqn:I3temp}, \eqref{eqn:I4temp}, \eqref{eqn:I5temp} and \eqref{eqn:I6temp} holds for all $1\leq \delta\leq q\leq p-1$.
Thus combining \eqref{eqn:b2dec}, \eqref{eqn:I1temp}, \eqref{eqn:I2temp}, \eqref{eqn:I3temp}, \eqref{eqn:I4temp}, \eqref{eqn:I5temp} and \eqref{eqn:I6temp}, yield
\begin{align*}
\sum_{\vec{i}\in C_q^<} \sum_{\vec{j}\in N_{\vec{i}}} \mathbb{E} \theta_{\vec{i},\vec{j}} & \leq
ap(pz)^{q+1}\left(1+ \mu_{n,q+\delta+1}\left(\bm{\Sigma}\right) (3q^2)\frac{\kappa-1}{p}\right)\left(1+pz\right)^{\delta-1} \delta\frac{4+b/z}{(\delta-1)!}+\\
&\quad\quad\quad ap^2 (pz)^{q+\delta} \frac{(\delta+1)(q+1)}{\delta! q!} \mu_{n,q+\delta+2}\left(\bm{\Sigma}\right)\frac{\kappa-1}{p}.
\end{align*}
\end{proof}
Combining Lemma \ref{lem:quadraticsum} and Lemma \ref{lem:b2} immediately yields the following lemma.
\begin{lem}
\label{lem:b2wholesum}
Let $p\geq n \geq 4$ and $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$. Suppose $\bm{\Sigma}$ is row-$\kappa$ sparse. Consider any $\theta_{\vec{i},\vec{j}}$ that is a non-negative function of $\bm{u}_{\ell}$ for $\ell\in \vec{i}\cup\vec{j}$ defined for $\vec{i}\in C_q^<$ and $\vec{j}\in C_\delta^<$ with $1\leq \delta\leq q\leq p-1$. Suppose there exist $a,z,b$ such that all conditions in Lemma \ref{lem:quadraticsum} and in Lemma \ref{lem:b2} hold. Moreover suppose $b/z\leq c_{n,\delta,q}$ for some positive constant $c_{n,\delta,q}$ that only depends on $n$, $q$ and $\delta$. Then
\begin{align*}
& \sum_{\vec{i}\in C_q^<} \sum_{\vec{j}\in C_\delta^<} \mathbb{E} \theta_{\vec{i},\vec{j}} \\
\leq &
C_{n,q,\delta} \left(p^{1+\frac{1}{\delta}}z\right)^q \left(1+(p^{1+\frac{1}{\delta}}z)^{\delta}\right)\left(1+pz\right)^{\delta-1} \left(1+\mu_{n,q+\delta+2}(\bm{\Sigma})\frac{\kappa-1}{p}\right) ap^{1-\frac{q}{\delta}}
\end{align*}
\end{lem}
\subsubsection{Proof of Proposition \ref{prop:edgecor}}
In what follows in this subsection, for the sake of brevity, we write $C^<$ for $C_\delta^<$, and write $\Phi_{\vec{i}}$ for $\Phi_{\vec{i}}^{(\bm{R})}$, for any $\vec{i}\in C^<$.
\begin{proof}[Proof of Proposition \ref{prop:edgecor}]
Recall
$$
N_{E_{\delta}}^{(\bm{R})}=\sum_{\vec{i}\in C_\delta^<}\ \ \prod_{j=1}^\delta \Phi_{i_0i_j}^{(\bm{R})}=\sum_{\vec{i}\in C_\delta^<}\ \ \Phi_{\vec{i}}.
$$
To apply a Compound Poisson Approximation result, some additional notation need to be introduced. For $\vec{i}\in C^<$, let $S_{\vec{i}}$ be defined in \eqref{eqn:Sidef}, and let $T_{\vec{i}}$, $N_{\vec{i}}$ be defined respectively as in \eqref{eqn:Tidef}, \eqref{eqn:Nidef} with $q=\delta$. Here $T_{\vec{i}}$ is the set of indices consisting of coordinates outside the neighborhood of $\vec{i}$; $N_{\vec{i}}$ is the set of "correlated but not highly correlated" indices, i.e. the set of indices of which at least one component but not every component is in the neighborhood of $\vec{i}$.
Denote
\begin{equation}
W_{\vec{i}} = \sum_{\vec{j}\in T_{\vec{i}}} \Phi_{\vec{j}}, \quad Z_{\vec{i}} = \sum_{\vec{j}\in N_{\vec{i}}} \Phi_{\vec{j}}, \label{eqn:WiZidef}
\end{equation}
and recall $U_{\vec{i}} = \sum_{\vec{j}\in S_{\vec{i}}}\Phi_{\vec{j}}^{(\bm{R})}$ is defined in \eqref{eqn:Uidef}. Then $W_{\vec{i}}$ is independent of $U_{\vec{i}}$ and $\Phi_{\vec{i}}$. Further denote
\begin{align*}
\lambda_0 & = \sum_{\vec{i}\in C^<} \mathbb{E} \left( \frac{\Phi_{\vec{i}}}{\Phi_{\vec{i}}+U_{\vec{i}}} 1\left(\Phi_{\vec{i}}+U_{\vec{i}} \geq 1\right)\right),\\
\zeta_{0\ell} & = \frac{1}{\lambda_0 \ell}\sum_{\vec{i}\in C^<} \mathbb{E}\left( \Phi_{\vec{i}} 1\left(\Phi_{\vec{i}}+ U_{\vec{i}} = \ell \right) \right), \quad \forall \ell \geq 1 \numberthis \label{eqn:defofzeta0l}
\end{align*}
and a probability distribution $\bm{\zeta}_0$ on positive integers with $\bm{\zeta}_0(\ell) = \zeta_{0\ell}$. The mean of $\bm{\zeta}_0$ is $\mathbb{E} \bm{\zeta}_0 = \sum_{\ell \geq 1} \ell \zeta_{0\ell}$. Moreover, let $b_1 =\sum_{\vec{i}\in C^<} \mathbb{E} \Phi_{\vec{i}} \mathbb{E}(\Phi_{\vec{i}} \ + U_{\vec{i}}+Z_{\vec{i}}) $ and
\begin{equation}
b_2 = \sum_{\vec{i}\in C^<}\mathbb{E} \left( \Phi_{\vec{i}} Z_{\vec{i}} \right). \label{eqn:b2def}
\end{equation}
In this proof we write $\lambda$ and $\bm{\zeta}$ for $\lambda_{p,n,\delta,\rho}$ and $\bm{\zeta}_{n,\delta,\rho}\left( \ell \right)$ respectively when there is no confusion. By the compound Poisson Stein's approximation, i.e. (5.19) and (5.16) in \cite{barbour2001topics},
\begin{equation}
d_{\operatorname{TV}}\left(\mathscr{L}\left(N_{E_\delta}^{(\bm{R})}\right), \operatorname{CP}(\lambda,\bm{\zeta})\right) \leq e^{\lambda_0} \left(b_1+b_2+\lambda_0 d_{W}(\bm{\zeta}_0',\bm{\zeta}') \mathbb{E} \bm{\zeta}_0+ |\lambda_0\mathbb{E}\bm{\zeta_0} - \lambda\mathbb{E}\bm{\zeta} |\right), \label{eqn: compoibou}
\end{equation}
where $\bm{\zeta}_{0}'(\ell) = \ell \zeta_{0\ell}/\mathbb{E} \bm{\zeta}_{0} $ and $\bm{\zeta}'(\ell) = \ell \bm{\zeta}(\ell)/\mathbb{E} \bm{\zeta} $ for $\ell \in \mathbb{Z}_+$, the set of all positive integers. In \eqref{eqn: compoibou}, the distance $d_W$ is the Wasserstein $L_1$ metric on probability measures over the set of positive integers $\mathbb{Z}_+$:
$$
d_W(P,Q) = \sup_{f\in \text{Lip}_1} \left| \int f dP - \int f dQ \right|
$$
where $\text{Lip}_1=\{f:|f(r)-f(s)|\leq |r-s|, r,s \in \mathbb{Z}_+ \}$.
By Lemma \ref{prop:was1dis},
\begin{align*}
\lambda_0 d_{W}(\bm{\zeta}_0',\bm{\zeta}') \mathbb{E} \bm{\zeta}_0 \leq & \lambda_0 \mathbb{E} \bm{\zeta}_0 \frac{\delta}{2}\sum_{\ell=1}^{\delta+1}\ell\left|\frac{\lambda_0\zeta_{0\ell}}{\lambda_0\mathbb{E}\bm{\zeta}_0} - \frac{\lambda\bm{\zeta}(\ell)}{\lambda\mathbb{E}\bm{\zeta}} \right|\\
= & \frac{\delta}{2}\sum_{\ell=1}^{\delta+1}\ell\left|\left(\lambda_0\zeta_{0\ell} - \lambda \bm{\zeta}(\ell)\right) + \left(\lambda\mathbb{E}\bm{\zeta}-\lambda_0 \mathbb{E} \bm{\zeta}_0 \right)\frac{\lambda\bm{\zeta}(\ell)}{\lambda\mathbb{E}\bm{\zeta}} \right|\\
\leq & \frac{\delta}{2}\sum_{\ell=1}^{\delta+1}\ell\left|\lambda_0\zeta_{0\ell} - \lambda \bm{\zeta}(\ell)\right| + \frac{\delta}{2}\left| \lambda\mathbb{E}\bm{\zeta}-\lambda_0 \mathbb{E} \bm{\zeta}_0 \right|\sum_{\ell=1}^{\delta+1}\ell\frac{\lambda\bm{\zeta}(\ell)}{\lambda\mathbb{E}\bm{\zeta}} \\
\leq & \delta\sum_{\ell=1}^{\delta+1}\ell\left|\lambda_0\zeta_{0\ell} - \lambda \bm{\zeta}(\ell)\right|.
\end{align*}
Plugging the above inequalities into \eqref{eqn: compoibou},
\begin{equation}
d_{\operatorname{TV}}\left(\mathscr{L}\left(N_{E_\delta}^{(\bm{R})}\right), \operatorname{CP}(\lambda,\bm{\zeta})\right) \leq e^{\lambda_0} \left(b_1+b_2+ \left(\delta+1\right)\sum_{\ell=1}^{\delta+1}\ell\left|\lambda_0\zeta_{0\ell} - \lambda \bm{\zeta}(\ell)\right|\right). \label{eqn:compoibou1}
\end{equation}
It remains to estimate the quantities in the right hand side of \eqref{eqn:compoibou1}. \\
\noindent{\textbf{Part I. Upper bound for} $\lambda_0$ \textbf{and} $\sum_{\ell=1}^{\delta+1}\ell\left|\lambda_0\zeta_{0\ell} - \lambda \bm{\zeta}(\ell)\right|$ } \\
For $\ell \in [\delta+1]$,
\begin{align*}
&\left|\lambda_0\zeta_{0\ell} - \lambda\bm{\zeta}\left( \ell \right) \right|\\
= &\left|\lambda_0\zeta_{0\ell} - \frac{1}{\ell}\binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^\delta\alpha(\ell,r_\rho) \right| \\
\leq &\frac{1}{ \ell}\sum_{\vec{i}\in C^<} \left|\mathbb{E}\left( \Phi_{\vec{i}} 1\left(\Phi_{\vec{i}}+ U_{\vec{i}} = \ell \right) \right) - (2P_n(r_\rho))^\delta\alpha(\ell,r_\rho) \right|\\
= & \frac{1}{ \ell}\sum_{ \substack{\vec{i}\in C^< \\ \bm{\Sigma}_{\vec{i}} \text{ not diagonal} }} \ \ \left|\mathbb{E}\left( \Phi_{\vec{i}} 1\left(\Phi_{\vec{i}}+ U_{\vec{i}} = \ell \right) \right) - (2P_n(r_\rho))^\delta\alpha(\ell,r_\rho) \right| \\
\leq & \frac{1}{ \ell} (\mu_{n,\delta+1}\left(\bm{\Sigma}\right)+1) (2P_n(r_\rho))^\delta\alpha(\ell,r_\rho)\sum_{ \substack{\vec{i}\in C^< \\ \bm{\Sigma}_{\vec{i}} \text{ not diagonal} }} \ \ 1 \\
\leq & \frac{\alpha(\ell,r_\rho)}{ \ell} \mu_{n,\delta+1}\left(\bm{\Sigma}\right) \gamma^\delta \frac{(\delta+1)}{(\delta-1)!} \frac{\kappa-1}{p}, \numberthis \label{eqn:lambdazetabou}
\end{align*}
where the first inequality follows from the definition of $\zeta_{0\ell}$ in \eqref{eqn:defofzeta0l}, the second inequality follows from Lemma \ref{item:prod} and Lemma \ref{lem:densitybou} \ref{item:densityboua}, and the last inequality follows from Lemma \ref{lem:numofnondia} and $\mu_{n,\delta+1}\left(\bm{\Sigma}\right)\geq 1$.
Then,
\begin{equation}
|\lambda_0\mathbb{E}\bm{\zeta_0} - \lambda\mathbb{E}\bm{\zeta} | \leq \sum_{\ell=1}^{\delta+1} \ell\left|\lambda_0\zeta_{0\ell} - \lambda\bm{\zeta(\ell)} \right|\leq \mu_{n,\delta+1}\left(\bm{\Sigma}\right) \gamma^\delta \frac{(\delta+1)}{(\delta-1)!} \frac{\kappa-1}{p}, \label{eqn:lambdameandif}
\end{equation}
where the last inequality follows from \eqref{eqn:lambdazetabou}. As an immediate consequences,
\begin{align*}
\lambda_0
&\leq \lambda_0\mathbb{E}\bm{\zeta_0} \\
& \leq |\lambda_0\mathbb{E}\bm{\zeta_0} - \lambda\mathbb{E}\bm{\zeta} |+ \lambda\mathbb{E}\bm{\zeta} \\
& \leq \mu_{n,\delta+1}\left(\bm{\Sigma}\right) \gamma^\delta \frac{(\delta+1)}{(\delta-1)!} \frac{\kappa-1}{p} + \binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^\delta \\
& \leq \gamma^\delta \frac{(\delta+1)}{(\delta-1)!}\left( \mu_{n,\delta+1}\left(\bm{\Sigma}\right) \frac{\kappa-1}{p} +1 \right), \numberthis \label{eqn:lambda0uppbou}
\end{align*}
where the third inequality follows from \eqref{eqn:lambdameandif}.\\
\noindent {\textbf{Part II. Upper bound for } $b_1$}\\
Since $N_{\vec{i}}\cup S_{\vec{i}}\cup \{\vec{i}\}=C^<\backslash T_{\vec{i}}$,
\begin{equation}
b_1 = \sum_{\vec{i}\in C^<} \ \ \sum_{\vec{j}\in C^< \backslash T_{\vec{i}} } \mathbb{E}\Phi_{\vec{i}} \mathbb{E} \Phi_{\vec{j}}. \label{eqn:b1uppbounew}
\end{equation}
Given $\vec{i}\in C^<$, by \eqref{eqn:p'upperbounew} with $q=\delta$,
\begin{equation}
p_{\vec{i}}:=\left|[p]\backslash \mathcal{NZ}\left(\vec{i}\right)\right|\geq p-(\delta+1)\kappa. \label{eqn:p'upperbou}
\end{equation}
Since $\left|T_{\vec{i}}\right| = p_{\vec{i}}\binom{p_{\vec{i}}-1}{\delta}$,
\begin{equation}
\left| C^<\backslash T_{\vec{i}}\right|=p\binom{p-1}{\delta}-p_{\vec{i}}\binom{p_{\vec{i}}-1}{\delta
\leq \frac{1}{\delta !} (\delta+1)\left(\prod_{\alpha=0}^{\delta-1} (p-\alpha) \right) (p-p_{\vec{i}})\leq \frac{(\delta+1)^2}{\delta !} p^{\delta}\kappa, \label{eqn:neiborsize}
\end{equation}
where the first inequality follows from Lemma \ref{basine} \ref{item:basineh}.
\begin{comment}
provided $p_{\vec{i}}>\delta$, and the last inequality follows from \eqref{eqn:p'upperbou}. In the case $p_{\vec{i}}
\leq \delta $,
\begin{equation}
\left| C^<\backslash T_{\vec{i}}\right| \leq \frac{\prod_{\alpha=0}^{\delta-1}(p-\alpha)}{\delta!}(p-\delta)\leq \frac{\prod_{\alpha=0}^{\delta-1}(p-\alpha)}{\delta!}(p-p'_{\vec{i}})(\delta+1)\leq \frac{(\delta+1)^2}{\delta !} p^{\delta}\kappa. \label{eqn:neiborsize}
\end{equation}
\end{comment}
One straightforward upper bound is
\begin{align*}
b_1 & = \sum_{\vec{i}\in C^{<}} \ \sum_{\vec{j}\in C^<\backslash T_{\vec{i}}} \mathbb{E}\left( \prod_{l=1}^\delta \Phi_{i_0i_l}^{(\bm{R})} \right) \mathbb{E}\left( \prod_{l'=1}^\delta \Phi_{j_0j_{l'}}^{(\bm{R})} \right) \\
&\leq p\binom{p-1}{\delta}\frac{(\delta+1)^2}{\delta !} p^{\delta}\kappa\left(\mu_{n,\delta+1}\left(\bm{\Sigma}\right)\right)^2(2P_n(r_\rho))^{2\delta} \\
& \leq \left(\mu_{n,\delta+1}\left(\bm{\Sigma}\right)\right)^2\frac{(\delta+1)^2}{(\delta !)^2}\left(2p^{1+\frac{1}{\delta}}P_n(r_\rho)\right)^{2\delta}\frac{\kappa}{p},
\end{align*}
where the first inequality follows from Lemma \ref{item:prod}, Lemma \ref{lem:densitybou} \ref{item:densityboua} and \eqref{eqn:neiborsize}. The $\left(\mu_{n,\delta+1}\left(\bm{\Sigma}\right)\right)^2$ in the above upper bound is not very satisfactory, and can be improved by a more involved analysis.
Observe for given $\vec{i}\in C^<$,
\begin{align*}
&\left|\{\vec{j}\in C^<\backslash T_{\vec{i}}: \bm{\Sigma}_{\vec{j}} \text{ not diagonal} \} \right|\\
=& |C^<\backslash T_{\vec{i}}| - \left( \left|\{\vec{j}\in C^<: \bm{\Sigma}_{\vec{j}} \text{ diagonal} \} \right| -\left|\{\vec{j}\in T_{\vec{i}}: \bm{\Sigma}_{\vec{j}} \text{ diagonal} \} \right| \right)\\
\leq & \frac{1}{\delta !} (\delta+1)\left(\prod_{\alpha=0}^{\delta-1} (p-\alpha) \right) (p-p_{\vec{i}}) - \left( \left|\{\vec{j}\in C^<: \bm{\Sigma}_{\vec{j}} \text{ diagonal} \} \right| -\left|\{\vec{j}\in T_{\vec{i}}: \bm{\Sigma}_{\vec{j}} \text{ diagonal} \} \right| \right), \numberthis \label{eqn:neinotdiagonalsize}
\end{align*}
where the inequality follows from \eqref{eqn:neiborsize}.
Then
\begin{align*}
& \left|\{\vec{j}\in C^<: \bm{\Sigma}_{\vec{j}} \text{ diagonal} \} \right| -\left|\{\vec{j}\in T_{\vec{i}}: \bm{\Sigma}_{\vec{j}} \text{ diagonal} \} \right| \\
= &
\frac{1}{\delta !}\sum_{j_0=1}^p\ \sum_{j_1\in [p]\backslash \mathcal{NZ}(j_0)}\ \cdots \sum_{j_\delta\in [p]\backslash \bigcup\limits_{l=0}^{\delta-1}\mathcal{NZ}(j_l)}1- \frac{1}{\delta !}\sum_{\substack{j_0\in [p] \\ j_0 \not \in \mathcal{NZ}(\vec{i})} }\ \sum_{\substack{j_1\in [p]\backslash \mathcal{NZ}(j_0)\\ j_1 \not \in \mathcal{NZ}(\vec{i}) }}\ \cdots \sum_{\substack{j_\delta\in [p]\backslash \bigcup\limits_{l=0}^{\delta-1}\mathcal{NZ}(j_l)\\ j_\delta \not \in \mathcal{NZ}(\vec{i})}}1 \numberthis \label{eqn:substract} \\
=&
\frac{1}{\delta !} \sum_{m=0}^\delta \left(\sum_{\substack{j_0\in [p] \\ j_0 \not \in \mathcal{NZ}(\vec{i})} }\
\sum_{\substack{j_1\in [p]\backslash \mathcal{NZ}(j_0)\\ j_1 \not \in \mathcal{NZ}(\vec{i}) }}\ \cdots\right.\\ &\left.\quad \sum_{\substack{j_{m-1}\in [p]\backslash \bigcup\limits_{l=0}^{m-2}\mathcal{NZ}(j_l)\\ j_{m-1} \not \in \mathcal{NZ}(\vec{i})}}\ \sum_{\substack{j_m\in [p]\backslash \bigcup\limits_{l=0}^{m-1}\mathcal{NZ}(j_l)\\ j_m \in \mathcal{NZ}(\vec{i})}} \ \ \sum_{\substack{j_{m+1}\in [p]\backslash \bigcup\limits_{l=0}^{m}\mathcal{NZ}(j_l)}} \cdots \sum_{\substack{j_\delta\in [p]\backslash \bigcup\limits_{l=0}^{\delta-1}\mathcal{NZ}(j_l)}}1 \right)\\
= & \frac{1}{\delta !} \sum_{\substack{j_0\in [p] \\ j_0 \in \mathcal{NZ}(\vec{i})} }\ \sum_{\substack{j_1\in [p]\backslash \mathcal{NZ}(j_0) }}\ \cdots \sum_{\substack{j_\delta\in [p]\backslash \bigcup\limits_{l=0}^{\delta-1}\mathcal{NZ}(j_l)}}1+
\frac{1}{\delta !} \sum_{m=1}^\delta \left( \right.\\
&\left.\sum_{\substack{ j_m \in \mathcal{NZ}(\vec{i})}} \sum_{\substack{j_0\in [p] \\ j_0 \not \in \mathcal{NZ}(\vec{i})\\ j_0\not \in \mathcal{NZ}(j_m) } }\ \sum_{\substack{j_1\in [p]\backslash \mathcal{NZ}(j_0)\\ j_1 \not \in \mathcal{NZ}(\vec{i}) \\ j_1 \not \in \mathcal{NZ}(j_m) }}\ \cdots \sum_{\substack{j_{m-1}\in [p]\backslash \bigcup\limits_{l=0}^{m-2}\mathcal{NZ}(j_l)\\ j_{m-1} \not \in \mathcal{NZ}(\vec{i})\\ j_{m-1}\not \in \mathcal{NZ}(j_m)}}\ \ \ \sum_{\substack{j_{m+1}\in [p]\backslash \bigcup\limits_{l=0}^{m}\mathcal{NZ}(j_l)}} \cdots \sum_{\substack{j_\delta\in [p]\backslash \bigcup\limits_{l=0}^{\delta-1}\mathcal{NZ}(j_l)}}1 \right) \\
\geq & \frac{1}{\delta !} (p-p_{\vec{i}}) \prod_{\beta=1}^\delta (p-\beta \kappa)+
\frac{1}{\delta !} \sum_{m=1}^\delta (p-p_{\vec{i}}) \left(\prod_{\alpha=1}^m (p_{\vec{i}}-\alpha \kappa)\right) \left(\prod_{\beta=m+1}^\delta (p-\beta \kappa)\right) \\
\geq &
\frac{(\delta+1)}{\delta !} (p-p_{\vec{i}}) \prod_{\alpha=1}^\delta (p_{\vec{i}}-\alpha \kappa), \numberthis \label{eqn:neidiagonal}
\end{align*}
where the second equality follows by writing \eqref{eqn:substract} as a telescoping sum with the convention that the summation over $j_{-1}$ for $m=0$ and the summation over $j_{\delta+1}$ for $m=\delta$ vanish, and the third equality follows from changing the order of the summation for $m\geq 1$. Plugging \eqref{eqn:neidiagonal} into \eqref{eqn:neinotdiagonalsize},
\begin{align*}
&\left|\{\vec{j}\in C^<\backslash T_{\vec{i}}: \bm{\Sigma}_{\vec{j}} \text{ not diagonal} \} \right|\\
\leq & \frac{1}{\delta !} (\delta+1) (p-p_{\vec{i}}) \left(\prod_{\alpha=0}^{\delta-1} (p-\alpha) - \prod_{\alpha=0}^{\delta-1} (p_{\vec{i}}-1-\alpha) + \prod_{\alpha=1}^{\delta} (p_{\vec{i}}-\alpha)- \prod_{\alpha=1}^\delta (p_{\vec{i}}-\alpha \kappa) \right)\\
\leq & \frac{1}{\delta !} (\delta+1) (p-p_{\vec{i}}) \left(\delta p^{\delta-1} (p-p_{\vec{i}}+1) + \frac{\delta(\delta+1)}{2} p^{\delta-1} (\kappa-1) \right) \\
\leq & \frac{3\delta(\delta+1)^3}{\delta !} p^{\delta-1} \kappa^2, \numberthis \label{eqn:neinotdiasize2}
\end{align*}
where the second inequality follows from Lemma \ref{basine} \ref{item:basineh} and Lemma \ref{basine} \ref{item:basinei}, and the last inequality follows from \eqref{eqn:p'upperbou}.
\begin{comment}
In the case $p_{\vec{i}}\leq \delta\kappa$, $p\leq p_{\vec{i}}+(\delta+1)\kappa \leq (2\delta+1)\kappa$, and thus
\begin{align*}
\left|\{\vec{j}\in C^<\backslash T_{\vec{i}}: \bm{\Sigma}_{\vec{j}} \text{ not diagonal} \} \right|
&\leq \frac{1}{\delta !} (\delta+1)p^{\delta} (\delta+1) \kappa \\
& \leq \frac{1}{\delta !} (\delta+1)p^{\delta} (\delta+1) \kappa\\
&\leq \frac{(\delta+1)^2}{\delta !} p^{\delta-1} \kappa (2\delta+1)\kappa\\
& \leq \frac{3\delta(\delta+1)^3}{\delta !} p^{\delta-1} \kappa^2, \numberthis \label{eqn:neinotdiasize2}
\end{align*}
where the first inequality follows from \eqref{eqn:neinotdiagonalsize} and \eqref{eqn:p'upperbou}.
\end{comment}
Then for any $\vec{i}\in C^<$,
\begin{align*}
\sum_{\vec{j}\in C^<\backslash T_{\vec{i}}}\mu_{n,\delta+1}\left(\bm{\Sigma}_{\vec{j}}\right) \leq & \sum_{\substack{\vec{j}\in C^<\backslash T_{\vec{i}}\\ \bm{\Sigma}_{\vec{j}} \text{ diagonal}}}1 + \mu_{n,\delta+1}\left(\bm{\Sigma}\right) \sum_{\substack{\vec{j}\in C^<\backslash T_{\vec{i}}\\ \bm{\Sigma}_{\vec{j}} \text{ not diagonal}}}1\\
\leq & \frac{(\delta+1)^2}{\delta !} p^{\delta}\kappa + \mu_{n,\delta+1}\left(\bm{\Sigma}\right) \frac{3\delta(\delta+1)^3}{\delta !} p^{\delta-1} \kappa^2, \\
\leq & \frac{3\delta(\delta+1)^3}{\delta !} p^{\delta} \kappa\left(1+ \mu_{n,\delta+1}\left(\bm{\Sigma}\right) \frac{\kappa}{p} \right), \numberthis \label{eqn:sumti}
\end{align*}
where the first inequality follows from Lemma \ref{lem:densitybou} \ref{item:densityboua}, and the second inequality follows from \eqref{eqn:neiborsize}, \eqref{eqn:neinotdiasize2}.
Then following \eqref{eqn:b1uppbounew},
\begin{align*}
b_1 & \leq \sum_{\vec{i}\in C^{<}} \ \sum_{\vec{j}\in C^<\backslash T_{\vec{i}}} \mu_{n,\delta+1}\left(\bm{\Sigma}_{\vec{i}}\right)\mu_{n,\delta+1}\left(\bm{\Sigma}_{\vec{j}}\right) (2P_n(r_\rho))^{2\delta}\\
&\leq \frac{p^{\delta+1}}{\delta !}\left(1+\delta^2\mu_{n,\delta+1}(\bm{\Sigma})\frac{\kappa-1}{p}\right)
\frac{3\delta(\delta+1)^3}{\delta !} p^{\delta} \kappa\left(1+ \mu_{n,\delta+1}\left(\bm{\Sigma}\right) \frac{\kappa}{p} \right) (2P_n(r_\rho))^{2\delta} \\
& \leq \left(3\frac{\delta^3(\delta+1)^3}{(\delta!)^2}(2p^{1+\frac{1}{\delta}}P_n(r_\rho))^{2\delta}\right)\ \frac{\kappa}{p} \left(1+\mu_{n,\delta+1}(\bm{\Sigma})\frac{\kappa}{p}\right)^2, \numberthis \label{eqn:b1upperbou3}
\end{align*}
where the first inequality follows from Lemma \ref{item:prod}, the second inequality follows from Lemma \ref{lem:summuupp} and \eqref{eqn:sumti}.
\vspace{2mm}
\noindent{\textbf{Part III. Upper bound for } $b_2$}\\
Let $\mathcal{K}_w(\vec{i})$ and $D_{\vec{i}}^m$ be the same as in Subsection \ref{sec:quadraticterms} with $q=\delta$. By Lemma \ref{lem:densitybou} \ref{item:densitybouc}, Lemma \ref{item:prod} and Lemma \ref{lem:eleine},
it follows that the conditions in Lemma \ref{lem:b2} with $q=\delta$ and $\theta_{\vec{i},\vec{j}}=\Phi_{\vec{i}}\Phi_{\vec{j}}$ are satisfied with $a=1$, $b=2P_n(2r_\rho)1(\delta\geq 2)+2P_n(r_\rho)1(\delta= 1)$ and $z=2P_n(r_\rho)$. Moreover, $b/z\leq 2^{n-2}1(\delta\geq 2)+1$ by Lemma \ref{pderivative} \ref{item:pderivatived}. Thus by Lemma \ref{lem:b2} with $q=\delta$ and $\theta_{\vec{i},\vec{j}}=\Phi_{\vec{i}}\Phi_{\vec{j}}$,
\begin{align*}
b_2 & \leq p(2pP_n(r_\rho))^{\delta+1}\left(1+ \mu_{n,2\delta+2}\left(\bm{\Sigma}\right) (3\delta^2)\frac{\kappa-1}{p}\right)\left(1+2pP_n(r_\rho)\right)^{\delta-1} \delta\frac{5+2^{n-1}1(\delta \geq 2)}{(\delta-1)!}+\\
&\quad\quad\quad \frac{(\delta+1)^2}{(\delta!)^2}p^2(2pP_n(r_\rho))^{2\delta}\mu_{n,2\delta+2}\left(\bm{\Sigma}\right)\frac{\kappa-1}{p}. \numberthis \label{eqn:b2upperbou2new} \\
\leq & \gamma^{\delta+1}p^{-\frac{1}{\delta}}\left(1+ \mu_{n,2\delta+2}\left(\bm{\Sigma}\right) (3\delta^2)\frac{\kappa-1}{p}\right)\left(1+\gamma p^{-\frac{1}{\delta}}\right)^{\delta-1} \delta\frac{5+2^{n-1}1(\delta \geq 2)}{(\delta-1)!}+\\
&\quad\quad\quad \frac{(\delta+1)^2}{(\delta!)^2}\gamma^{2\delta}\mu_{n,2\delta+2}\left(\bm{\Sigma}\right)\frac{\kappa-1}{p}, \numberthis \label{eqn:b2upperbou2}
\end{align*}
where the last step follows from $2p^{1+\frac{1}{\delta}}P_n(r_\rho)\leq \gamma$.
By combining \eqref{eqn:compoibou1}, \eqref{eqn:lambdameandif}, \eqref{eqn:lambda0uppbou}, \eqref{eqn:b1upperbou3}, \eqref{eqn:b2upperbou2}, together with the assumption $2p^{1+\frac{1}{\delta}}P_n(r_\rho)\leq \gamma$,
\begin{align*}
&d_{\operatorname{TV}}\left(\mathscr{L}\left(N_{E_\delta}^{(\bm{R})}\right), \operatorname{CP}(\lambda,\bm{\zeta})\right) \\
\leq & C_{n,\delta,\gamma} \left(C'_{\delta,\gamma}\right)^{\mu_{n,\delta+1}\left(\bm{\Sigma}\right)(\kappa-1)/p}\left( \mu_{n,2\delta+2}\left(\bm{\Sigma}\right)\kappa/p \left(1 + \mu_{n,2\delta+2}\left(\bm{\Sigma}\right) \left(\kappa/p\right)^2 \right) + p^{-\frac{1}{\delta}} \right) ,
\end{align*}
where
\begin{equation}
C'_{\delta,\gamma} = \exp\left(\gamma^{\delta}\frac{\delta+1}{(\delta-1)!}\right), \label{eqn:C'delgamdef}
\end{equation}
and
\begin{equation}
C_{n,\delta,\gamma} = C\frac{\delta^6+\delta^22^{n-1}1(\delta \geq 2)}{\delta!} \gamma^{\delta+1}(1+\gamma)^\delta C'_{\delta,\gamma}. \label{eqn:Cndelgamdef}
\end{equation}
\end{proof}
\subsection{Proofs in Subsection \ref{sec:scorerep}}
\subsubsection{Proofs of Lemma \ref{lem:Binv} and Lemma \ref{lem:Uscoresdistribution}}
\label{sec:Uscoresproof}
\begin{proof}[Proof of Lemma \ref{lem:Binv}]
It suffices to prove that $\bm{U}$ has rank $n-1$ $a.s.$ By Lemma \ref{lem:Uscoresdistribution} \ref{item:Uscoresdistributiona}, it suffices to prove that $\left[\frac{\tilde{x}_1}{\|\tilde{x}_1\|_2},\ldots,\frac{\tilde{x}_p}{\|\tilde{x}_p\|_2}\right]$ has rank $n-1$ $a.s.$ Note that $\tilde{X}$ is of rank $n-1$ $a.s.$, since it has a density with respect to Lebesgue measure on $\mathbb{R}^{(n-1)p}$ and $p\geq n$. Then $\left[\frac{\tilde{x}_1}{\|\tilde{x}_1\|_2},\ldots,\frac{\tilde{x}_p}{\|\tilde{x}_p\|_2}\right]$ have rank $n-1$ $a.s.$ as well since it is obtained by normalizing the columns of $\tilde{X}$.
\end{proof}
\begin{proof}[Proof of Lemma \ref{lem:Uscoresdistribution}]
(a)
Consider $\hat{x}^{(i)}\overset{\text{i.i.d.}}{\sim}\mathcal{N}(\bm{0},\bm{\Sigma})$. Then $\hat{X} := [\hat{x}^{(1)},\ldots,\hat{x}^{(n)}]^\top = [\hat{x}_1,\ldots,\hat{x}_p] \in \mathbb{R}^{n\times p}$ has the distribution $\mathcal{VE}(\bm{0},\bm{\Sigma},g_0)$ where $g_0(w)=(2\pi)^{-\frac{n p}{2}}\exp(-\frac{1}{2}w)$.
The sample mean $\bar{\hat{x}}$ of $\hat{X}$ is given as a row vector
$$
\bar{\hat{x}}=\frac{1}{n}\displaystyle\sum_{i=1}^{n}\hat{x}^{(i)} = \frac{1}{n} \hat{X}^\top\bm{1}=(\bar{\hat{x}}_1,\ldots,\bar{\hat{x}}_p),
$$
where $\bar{\hat{x}}_i:=n^{-1}\sum_{j=1}^p {\hat{X}_{ij}}$. The sample covariance matrix $\hat{\bm{S}}$ of $\hat{X}$ is
\begin{equation*}
\hat{\bm{S}}=\frac{1}{n-1}\sum_{i=1}^{n} (\hat{x}^{(i)}-\bar{\hat{x}})(\hat{x}^{(i)}-\bar{\hat{x}})^\top
=\frac{1}{n-1}\hat{X}^\top \left(\bm{I}-\frac{1}{n}\bm{1}\bm{1}^\top\right) \hat{X}
\end{equation*}
The Z-scores of $\hat{X}$ is
$$
\hat{z}_i := \frac{\hat{x}_i-\bm{1}\bar{\hat{x}}_i}{\sqrt{\hat{S}_{ii}(n-1)}}= \frac{\left(\bm{I}-\frac{1}{n}\bm{1}\bm{1}^\top\right) \hat{x}_i}{\sqrt{\hat{S}_{ii}(n-1)}} .
$$
Recall $\bm{H}_{2:n}\in \mathbb{R}^{n\times (n-1)}$ is a matrix such that
$n \times (n-1)$ matrix $\bm{H} = [n^{-\frac{1}{2}}\bm{1}, \bm{H}_{2:n}]^\top$ is orthogonal, which satisfies $\bm{H}_{2:n} \bm{H}_{2:n}^\top = \bm{I} -\frac{1}{n}\bm{1}\bm{1}^T$. The U-scores of $\hat{X}$ is denoted $\hat{\bm{U}}=[\hat{\bm{u}}_1,\ldots,\hat{\bm{u}}_p]$ with $\hat{\bm{u}}_i=\bm{H}_{2:n}^\top z_i$ for $i\in [p]$.
By the theorem in Section 6 of \cite{anderson1992nonnormal}, the distribution of $\bm{U}$ is invariant to $g$ and $\bm{\mu}$. In particular, $\bm{U}$ has the same distribution as $\hat{\bm{U}}$. It suffices to study the distribution of $\hat{\bm{U}}$.
The proof of Theorem 3.3.2 in \cite{anderson2003introduction} establishes that
\begin{equation*}
\check{X} = \bm{H}_{2:n}^\top \hat{X} \in \mathbb{R} ^{(n-1) \times p}
\end{equation*}
has i.i.d. rows $\{\check{x}^{(i)}\}_{i=1}^{n-1}$ distributed as $\mathcal{N}(\bm{0},\bm{\Sigma})$, and
\begin{equation*}
\hat{\bm{S}} = \frac{1}{n-1}\hat{X}^\top \left(\bm{I}-\frac{1}{n}\bm{1}\bm{1}^\top\right) \hat{X} =\frac{1}{n-1} \left(\bm{H}_{2:n}^\top \hat{X}\right)^\top \bm{H}_{2:n}^\top \hat{X}= \frac{1}{n-1} \check{X}^\top \check{X}.
\end{equation*}
Then $\hat{S}_{ii} = \|\check{x}_i\|^2/(n-1) $ where $\check{x}_i$ is the $i$-th column of $\check{X}$. Thus
$$
\hat{\bm{u}}_i = \bm{H}_{2:n}^\top \frac{\left(\bm{I}-\frac{1}{n}\bm{1}\bm{1}^\top\right) \hat{x}_i}{\sqrt{\hat{S}_{ii}(n-1)}} = \frac{\bm{H}_{2:n}^\top \bm{H}_{2:n} \bm{H}_{2:n}^\top \hat{x}_i }{\sqrt{\|\check{x}_i\|^2}}=\frac{\check{x}_i}{\|\check{x}_i\|_2}.
$$
The proof is then complete by observing that $\check{X}$ and $\tilde{X}$ has the same distribution.\\
\noindent
(b)
Part \ref{item:Uscoresdistributionb} directly follows from part \ref{item:Uscoresdistributiona}.
\end{proof}
\subsubsection{Proof of Lemma \ref{lem:densitybou}} \label{sec:proofdensitybou}
Part \ref{item:densityboua} of Lemma \ref{lem:densitybou} follows from $\mu_m(\bm{\Sigma}_{\mathcal{J}})=\mu(\bm{\Sigma}_{\mathcal{J}})\geq \mu_m(\bm{\Sigma})$. Part \ref{item:densitybouc} immediately follows from part \ref{item:densityboub} and in the remainder of this subsection we give the proof of part \ref{item:densityboub}.
It suffices to prove the following statement:
$$
f_{\bm{u}_{j_1},\bm{u}_{j_2},\cdots,\bm{u}_{j_m}}(\bm{v}_1,\bm{v}_2,\cdots,\bm{v}_m) \leq \mu_{n,m}(\bm{\Sigma}_{\mathcal{J}}), \quad \forall \ \bm{v}_i\in S^{n-2}, \forall\ i\in [m].
$$
For notation convenience, we only present the proof for $m=p$ and $\mathcal{J}=[p]$ since the proof of the general $m$ and $\mathcal{J}$ follows the same proof procedure. When $m=p$ and $\mathcal{J}=[p]$, the statement of Lemma \ref{lem:densitybou} \ref{item:densityboub} becomes:
The joint density of columns of U-scores w.r.t. $\otimes^p \sigma^{n-1}$ is upper bounded by $\mu_{n,p}(\bm{\Sigma})$: \begin{equation}
f_{\bm{u}_1,\bm{u}_2,\cdots,\bm{u}_p}(\bm{v}_1,\bm{v}_2,\cdots,\bm{v}_p)\leq \mu_{n,p}(\bm{\Sigma}), \quad \forall \ \bm{v}_i\in S^{n-2}, \forall\ i\in [p]. \label{eqn:jointuuppbou}
\end{equation}
\begin{proof}[Proof of \eqref{eqn:jointuuppbou}]
Recall $\{\tilde{x}^{(i)}\}_{i=1}^{n-1}\subset \mathbb{R}^p$, the rows of $\tilde{X}$, are i.i.d. copy of $\mathcal{N}(\bm{0},\bm{\Sigma})$; $\{\tilde{x}_i\}_{i=1}^{p}$ are the columns of $\tilde{X}$; $\bm{u}_i\vcentcolon=\frac{\tilde{x}_i}{\|\tilde{x}_i\|_2}\in \mathbb{R}^{n-1}$ has distribution $\operatorname{{\bf unif}}(S^{n-2})$ for $i\in [p]$.
When $\bm{\Sigma}$ is symmetric positive definite and diagonal, $\{\tilde{x}_i\}_{i=1}^{p}$ are independent, which implies that $\{\bm{u}_i\}_{i=1}^{p}$ are independent. Thus in this case, the joint density of columns of U-scores w.r.t. $\otimes^p \sigma^{n-1}$ is $1$.
Consider general symmetric positive definite $\bm{\Sigma}$. The probability density of $\tilde{X}$ w.r.t. the Lebesgue measure on $\mathbb{R}^{(n-1)p}$ is
$$
f_{\tilde{X}}(\tilde{X}) = A\exp\left( -\frac{1}{2}\sum_{j=1}^{n-1}
\left(\tilde{x}^{(j)}\right)^\top\bm{\Sigma}^{-1}\tilde{x}^{(j)} \right),
$$
where the constant $A=\text{det}(\bm{\Sigma})^{-\frac{n-1}{2}}(2\pi)^{-\frac{(n-1)p}{2}}$.
Use the spherical transform for each column $\tilde{x}_i = \left(\tilde{X}_{ji}:1\leq j \leq n-1\right)^\top$:
\begin{equation*}
\begin{cases}
\tilde{X}_{1i}=R_i\cos(\theta_{1i}),\\
\tilde{X}_{2i}=R_i\sin(\theta_{1i})\cos(\theta_{2i}),\\ \vdots & \text{\ \ for } 1\leq i \leq p,\\
\tilde{X}_{(n-2)i}=R_i\sin(\theta_{1i})\sin(\theta_{2i})\cdots\sin(\theta_{(n-3)i})\cos(\theta_{(n-2)i}),\\ \tilde{X}_{(n-1)i}=R_i\sin(\theta_{1i})\sin(\theta_{2i})\cdots\sin(\theta_{(n-3)i})\sin(\theta_{(n-2)i}),
\end{cases}
\end{equation*}
where for each $i\in [p]$: $R_i\geq 0, \theta_{ji}\in[0,\pi]$ for $1\leq j\leq n-3$ and $\theta_{(n-2)i}\in[0,2\pi)$.
Denote $\bm{R}=(R_i:1\leq i\leq p)$ and $\bm{\Theta}=(\theta_{ji}: 1\leq i\leq p, 1\leq j\leq (n-2))$.
Then the joint density of $(\bm{R},\bm{\Theta})$ is:
\begin{align*}
&f_{\bm{R},\bm{\Theta}}(\bm{R},\bm{\Theta})\\
= &
A\exp\left( -\frac{1}{2}\sum_{j=1}^{n-1}
\left(\bm{h}^{(j)}\right)^\top\bm{\Sigma}^{-1}\bm{h}^{(j)} \right) \prod_{i=1}^{p} \left(R_i^{n-2}\prod_{j=1}^{n-2}\sin^{n-2-j}(\theta_{ji})\right),
\end{align*}
where
$$
\bm{h}^{(j)}=\left(R_i\cos(\theta_{ji})\prod\limits_{q=1}^{j-1}\sin(\theta_{qi}):1\leq i\leq p\right)^\top\in \mathbb{R}^{p} \text{ for } 1\leq j\leq n-2
$$
and
$$
\bm{h}^{(n-1)}=\left(R_i\prod\limits_{q=1}^{n-2}\sin(\theta_{qi}):1\leq i\leq p\right)^\top\in \mathbb{R}^p.
$$
Then the density of $\bm{\Theta}$ is:
\begin{align*}
&f_{\bm{\Theta}}(\bm{\Theta}) \\
= &
A \left(\prod_{i=1}^{p} \prod_{j=1}^{n-2}\sin^{n-2-j}(\theta_{ji}) \right) \int_{[0,\infty)^p}\exp\left(-\frac{1}{2}\sum_{j=1}^{n-1}
\left(\bm{h}^{(j)}\right)^\top\bm{\Sigma}^{-1}\bm{h}^{(j)} \right) \prod_{i=1}^{p} R_i^{n-2} dR_i\\
\leq &
A \left(\prod_{i=1}^{p} \prod_{j=1}^{n-2}\sin^{n-2-j}(\theta_{ji}) \right) \int_{[0,\infty)^p} \exp\left( -\frac{1}{2}\lambda_{\text{min}}(\bm{\Sigma}^{-1})\sum_{j=1}^{n-1}
\|\bm{h}^{(j)}\|_2^2 \right) \prod_{i=1}^{p} R_i^{n-2}dR_i\\
=&
A \left(\prod_{i=1}^{p} \prod_{j=1}^{n-2}\sin^{n-2-j}(\theta_{ji}) \right) \int_{[0,\infty)^p} \exp\left( -\frac{1}{2}[\lambda_{\text{max}}(\bm{\Sigma})]^{-1}\sum_{i=1}^{p}
R_i^2 \right) \prod_{i=1}^{p} R_i^{n-2} d R_i \\
=&
A \left(\prod_{i=1}^{p} \prod_{j=1}^{n-2}\sin^{n-2-j}(\theta_{ji}) \right) \left(\int_{[0,\infty)}\exp\left( -\frac{1}{2}[\lambda_{\text{max}}(\bm{\Sigma})]^{-1}
R_1^2 \right) R_1^{n-2} dR_1\right)^p \\
\overset{(m)}{=}&
A \left(\prod_{i=1}^{p} \prod_{j=1}^{n-2}\sin^{n-2-j}(\theta_{ji}) \right) \left( \left[\lambda_{\text{max}}(\bm{\Sigma})\right]^{\frac{n-1}{2}} \Gamma\left(\frac{n-1}{2}\right)2^{\frac{n-3}{2}}
\right)^p\\
\overset{(mm)}{=}&\left[\frac{\left(\lambda_{\text{max}}(\bm{\Sigma})\right)^p}{\text{det}(\bm{\Sigma})}\right]^\frac{n-1}{2} \frac{1}{(\text{Area}(S^{n-2}))^p} \prod_{i=1}^{p} \prod_{j=1}^{n-2}\sin^{n-2-j}(\theta_{ji})
\end{align*}
where equality $(m)$ follows from the integration of Chi distribution with degree $n-1$, and equality $(mm)$ follows from $|S^{(n-2)}|=2\pi^{\frac{n-1}{2}}/\Gamma((n-1)/2)$. The proof is complete by noticing $f_{\bm{\Theta}}(\bm{\Theta})$ is joint density of columns of U-scores expressed in spherical coordinate
and $\frac{1}{(\text{Area}(S^{n-2}))^p} \prod_{i=1}^{p} \prod_{j=1}^{n-2}\sin^{n-2-j}(\theta_{ji})$ is the joint distribution of $p$ independent $\operatorname{{\bf unif}}(S^{n-2})$ expressed as spherical coordinates.
\end{proof}
\section{Formulae for \texorpdfstring{$\{N_i^{(k)}: k\in \{\bm{R},\bm{P}\}, i\in \{E_\delta, \breve{V}_\delta, V_\delta \} \}$}{\text{the $6$ quantities}}}
\label{sec:formulae}
In this section, for $\bm{\Psi} =\bm{R}$ or $\bm{P}$, we write $N_{\breve{V}_\delta}^{(k)}$, $N_{V_\delta}^{(k)}$ and $N_{\EE_\delta}^{(k)}$ in terms of $\Phi_{ij}^{(k)}(\rho)$. The dependence of $\Phi_{ij}^{(k)}(\rho)$ on $\rho$ will be suppressed if it's clear from context.
Define
$$
\breve{C}_i(p-1,l) :=\{\vec{q}: q_1<\dots<q_l, q_{l+1}<\dots<q_{p-1},q_j\in[p]\backslash \{i\},\ q_j\not=q_{j'}, \forall j,j'\in [p-1] \text{ and } j\not = j'\},
$$
and
$$
C_\delta^{<}:=\{\vec{i}=(i_0,i_1,\cdots,i_\delta)\in [p]^{\delta+1}: i_1<i_2<\cdots<i_\delta, \text{ and } i_l\not= i_0, \forall\ l\in [\delta] \}.
$$
The indicator function when the $i$-the vertex has degree $\delta$ for $1\leq \delta\leq p-1$ is:
$$
\breve{\Phi}_{i,\delta}^{(k)} :=\sum_{\vec{q}\in \breve{C}_i(p-1,\delta)} \prod_{j=1}^\delta \Phi_{iq_j}^{(k)}\prod_{m=\delta+1}^{p-1}(1-\Phi^{(k)}_{iq_m})
$$
The number of vertices with degree $\delta$ for $1\leq \delta\leq p-1$ is:
$$
N_{\breve{V}_\delta}^{(k)}=\sum\limits_{i=1}^p{\breve{\Phi}_{i,\delta}^{(k)}}
$$
The indicator function when the $i$-the vertex has degree at least $\delta$ for $1\leq \delta\leq p-1$ is:
$$
\Phi_{i,\delta}^{(k)} :=\sum_{l=\delta}^{p-1}\breve{\Phi}_{i,l}^{(k)}
$$
The number of vertices with degree at least $\delta$ for $1\leq \delta\leq p-1$ is:
$$
N_{V_\delta}^{(k)}=\sum\limits_{i=1}^p{\Phi_{i,\delta}^{(k)}}
$$
For $1\leq \delta \leq p-1$, define
$$
\mathring{\Phi}_{i,\delta}^{(k)} := \sum_{\vec{q}\in \breve{C}_i(p-1,\delta)}\ \ \prod_{j=1}^\delta \Phi_{iq_j}^{(k)}
= \sum_{\substack{q_1<q_2<\cdots<q_\delta\\q_j\in [p]\backslash \{i\}, \forall j\in [\delta] }}\ \ \prod_{j=1}^\delta \Phi_{iq_j}^{(k)}. $$
$\mathring{\Phi}_{i,\delta}^{(k)}$ is the number of groups of vertices in $[p]\backslash\{i\}$ of size $\delta$ such that the vertex $i$ is connected to each vertex in that group. Thus $\mathring{\Phi}_{i,\delta}^{(k)}=\binom{\text{deg}^{(k)}(i)}{\delta}$, where $\text{deg}^{(k)}(i)$ is the degree of vertex $i$ in empirical correlation graph when $\bm{\Psi} =\bm{R}$ or in partial empirical correlation graph when $\bm{\Psi} =\bm{P}$. Then for $2\leq \delta \leq p-1$, the number of subgraphs isomorphic to $\Gamma_\delta$ with center at the $i$-th vertex is $\mathring{\Phi}_{i,\delta}^{(k)}$. Thus $N_{\EE_\delta}^{(k)}$ for $1\leq \delta\leq p-1$ is
$$
N_{\EE_\delta}^{(k)}=
\begin{cases}
\sum\limits_{i=1}^p\mathring{\Phi}_{i,\delta}^{(k)} & \text{ when } 2 \leq \delta \leq p-1, \\
2\sum\limits_{1\leq i<j \leq p} \Phi_{ij}^{(k)}= \sum\limits_{i=1}^p\mathring{\Phi}_{i,1}^{(k)} & \text{ when } \delta = 1.
\end{cases} = \sum\limits_{i=1}^p\mathring{\Phi}_{i,\delta}^{(k)} =\sum_{\vec{i}\in C_\delta^<}\ \ \prod_{j=1}^\delta \Phi_{i_0i_j}^{(k)}.
$$
\section{Numerical simulations and experiments}
\subsection{Numerical simulations and experiments}
\label{sec:simulation}
\begin{figure}[ht]
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[width=.8\linewidth]{graphs/cpvd_taukappadelta1limitp.png}
\caption{limiting compound Poisson}
\label{fig:finiteplimitcomparisona}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[width=.8\linewidth]{graphs/cpvd_taukappadelta1finitep.png}
\caption{finite $p$ compound Poisson approximation}
\label{fig:finiteplimitcomparisonb}
\end{subfigure}
\caption{The vertical axis of (a) is $d_{\operatorname{TV}}(\mathscr{L}\left({N_{V_1}^{(\bm{\Psi})}}\right),\operatorname{CP}(\lambda_{20,1}(1),\bm{\zeta}_{20,1}))$ and that of (b) is $d_{\operatorname{TV}}(\mathscr{L}\left(N_{V_1}^{(\bm{\Psi})}\right),\operatorname{CP}(\lambda_{p,20,1,\rho},\bm{\zeta}_{20,1,\rho}))$.
\ONE{The theoretical convergence results have respectively been established in Theorem \ref{cor:Poissonlimit} and Theorem \ref{thm:Poissonultrahigh}.} For both plots the samples are independently generated according to $\mathcal{N}(\bm{0},\bm{\Sigma})$ with $\bm{\Sigma}$ being a $(\tau=p^{0.6},\kappa=p^{0.8})$ sparse matrix for each $p$. The parameters are $n=20$, $\delta=1$ and the threshold $\rho$ is chosen according to \eqref{eqn:rhopformula} with $e_{n,\delta}=1$.
The blue curve is for the empirical correlation graph ($\bm{\Psi} =\bm{R}$) and the red curve is for the empirical partial correlation graph ($\bm{\Psi} =\bm{P}$). Note since $\delta=1$, $\bm{\zeta}_{20,1}=\operatorname{Dirac}(2)=\bm{\zeta}_{20,1,\rho}$, by Example \ref{exa:limitcompoidelta1}. As demonstrated by the plots, for both the empirical correlation and partial correlation graphs, the total variations in (a) decrease very slowly while the total variations in (b) converge to $0$ very fast, which has also been analytically discussed in Remark \ref{rem:comparisonsfinitelimit}.
\ONE{Our observation that the non-asymptotic compound Poisson distribution provides a better fit to the numerical simulations for small $p$ is a caveat to practitioners who may be tempted to use the Poisson approximation.}
}
\label{fig:finiteplimitcomparison}
\end{figure}
\begin{figure}[ht]
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[width=.8\linewidth]{graphs/cpp_taukappan35delta2trial2000p700limit.png}
\caption{Poisson limit when $p\to \infty$}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[width=.8\linewidth]{graphs/cpp_taukappan35delta2trial2000p700.png}
\caption{Poisson approximation for finite $p$}
\end{subfigure}
\caption{ The vertical axis of (a) is $d_{\operatorname{TV}}\left(\mathscr{L}\left(N_{V_\delta}^{(\bm{\Psi})}\right),\operatorname{Pois}\left(\frac{(e_{n,\delta})^\delta}{\delta!}\right) \right)$, where we replaced $\operatorname{CP}(\lambda_{n,\delta}(e_{n,\delta}), \bm{\zeta}_{n,\delta})$ in Theorem \ref{cor:Poissonlimit} by its approximation $\operatorname{Pois}(\frac{(e_{n,\delta})^\delta}{\delta!})$ as in Proposition \ref{prop:poiapplimit}. The vertical axis of (b) is $d_{\operatorname{TV}}(N_{V_\delta}^{(\bm{\Psi})},\operatorname{Pois}(\binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta}))$, where we replaced $\operatorname{CP}(\lambda_{p,n,\delta,\rho}, \bm{\zeta}_{n,\delta,\rho})$ in Theorem \ref{thm:Poissonultrahigh} by its approximation $\operatorname{Pois}(\binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta})$ as in Proposition \ref{prop:poiappfinitep}. For both plots the samples are independently generated according to $\mathcal{N}(\bm{0},\bm{\Sigma})$ with $\bm{\Sigma}$ being a $(\tau=p^{0.6},\kappa=p^{0.8})$ sparse matrix for each $p$. The parameters are $n=35$, $\delta=2$ and the threshold $\rho$ is chosen according to \eqref{eqn:rhopformula} with $e_{n,\delta}=1$.
As demonstrated by the plots, for both the empirical correlation and partial correlation graphs, the total variations in (a) decrease very slowly while the total variations in (b) converge to $0$ very fast. The fast convergence in Figure \ref{fig:Poissonfinitevslimit} (b) verifies the validity of using Poisson distribution $\operatorname{Pois}(\binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta})$ to approximate the distribution of random quantities in $\{N_{E_\delta}^{(\bm{\Psi})},N_{\breve{V}_\delta}^{(\bm{\Psi})},N_{V_\delta}^{(\bm{\Psi})}: \bm{\Psi} \in \{\bm{R},\bm{P}\} \}$ for large $n$. We now discuss the slow convergence behavior of Figure \ref{fig:Poissonfinitevslimit} (a).
Note that $n=35$ is large enough to guarantee that $\sum_{\ell=2}^{\delta+1}\alpha_\ell$ is small as indicated by Figure \ref{fig:incrementtodeltauppbou} (b), which implies that $\operatorname{CP}(\lambda_{n,\delta}(e_{n,\delta}), \bm{\zeta}_{n,\delta})$ is well approximated by $\operatorname{Pois}(\frac{(e_{n,\delta})^\delta}{\delta!})$ by Lemma \ref{prop:jumpsizedecay} \ref{item:jumpsizedecayc}. As a result, the extremely slow decrease in Figure \ref{fig:Poissonfinitevslimit} (a) is not because of using the Poisson approximation, but is
due to the slow convergence of Theorem \ref{cor:Poissonlimit}, which has been extensively discussed in Remark \ref{rem:comparisonsfinitelimit}.
This specific example additionally indicates that the slow convergence of Theorem \ref{cor:Poissonlimit} is due to slow convergence of $\lambda_{p,n,\delta,\rho}\to \lambda_{n,\delta}$ since the increment distribution in this large $n$ case are both close to $\operatorname{Dirac}(1)$. }
\label{fig:Poissonfinitevslimit}
\end{figure}
\begin{comment}
Note the increment distributions $\bm{\zeta}_{35,2}$ in (a) and $\bm{\zeta}_{35,2,\rho}$ in (b) are both replaced by $\operatorname{Dirac}(1)$ since $n=35$ is sufficiently large for $\delta=2$ as indicated by Figure \ref{fig:incrementtodeltauppbou} (b). That is, the number of samples $n=35$ is large enough for Corollary \ref{prop:jumpsizedecay} \ref{item:jumpsizedecaya} and
Lemma \ref{lem:randgeograconduppbou} \ref{item:randgeograconduppbouc} to be effective. The blue curve is for the empirical correlation graph (\bm{\Psi} =$\bm{R}$) and the red curve is for the empirical partial correlation graph (\bm{\Psi} =$\bm{P}$). As demonstrated by the plots, for both the empirical correlation and partial correlation graphs, the total variations in (a) decrease very slowly while the total variations in (b) converge to $0$ very fast. The fast convergence in Figure \ref{fig:Poissonfinitevslimit} (b) verifies the validity of using Poisson distribution $\operatorname{Pois}(\binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta})$ to approximate the distribution of random quantities in $\{N_{E_\delta}^{(\bm{\Psi})},N_{\breve{V}_\delta}^{(\bm{\Psi})},N_{V_\delta}^{(\bm{\Psi})}: \bm{\Psi} \in \{\bm{R},\bm{P}\} \}$ for large $n$. The extremely slow decrease in Figure \ref{fig:Poissonfinitevslimit} (a) is due to the slow convergence of Theorem \ref{cor:Poissonlimit}, which has been extensively discussed in Remark \ref{rem:comparisonsfinitelimit}.
This specific example indicates that the slow convergence of Theorem \ref{cor:Poissonlimit} is due to slow convergence of $\lambda_{p,n,\delta,\rho}\to \lambda_{n,\delta}$ since the increment distribution in this large $n$ case are both close to $\operatorname{Dirac}(1)$.
\end{comment}
\section{Responses to Associate Editor's comments (from email)}
\begin{enumerate}
\item \mycomment{Specifically, while there are a number of technical points raised by the reviewers, one of the more substantial concerns raised by multiple reviewers concerns the novelty/significance of your work compared to some of the most closely related prior work. (I will also note that two of the reviewers raised concerns about the suitability of Transactions on Information Theory as an appropriate venue -- while you should feel free to address this concern in your response, I will say that this is not an issue that I am worried about with this paper.)
}
\myresponse{
We thank the AE for giving us an opportunity to revise the paper and for his encouraging remarks, especially regarding the issue of the suitability of Transactions on Information Theory as an appropriate venue. Per your request, in this revision we have comprehensively addressed the technical points raised by the reviewers. This 20
page "Response to reviewer document" attests to the sheer amount of work that has gone into addressing the reviewers' feedback. We have also specifically addressed the significance or novelty of the results in the paper (or alleged lack thereof) by bringing forward multiple points which highlight the significant and original contributions of the paper. In particular, a new subsection has been introduced into Section \ref{sec:contribution}, which contains the following content:
\begin{quote}
\ONE{We summarize the principal contributions of the paper. As above $p$ denotes the number of variables and $n$ denotes the number of samples.
\begin{enumerate}
\item
The paper presents a unified and complete asymptotic analysis of the star subgraph counts, the counts of vertices of a given degree and the counts of vertices above a given degree in the random graphs obtained by thresholding the sample correlation and sample partial correlation. This unification of different types of random counts represents an important improvement over previous work \cite{hero2011large,hero2012hub} where only
the counts of vertices above a given degree are studied.
\item
We approximate the full distributions of the random counts for finite $p$ and as $p\to\infty$. In addition, we characterize the first and second moments of these random counts. This is a significant generalization of previous results \cite{hero2011large, hero2012hub} that only established approximations for the mean number of random counts and for the probabilities that these counts were positive.
\item
We obtains compound Poisson characterizations of the the distributions of the random counts.
The compound Poisson limit and approximation are well approximated by the standard Poisson limit when $n$ is moderately large (Section \ref{sec:limitingcompoundPoisson}). This result corrects and refines the claim in \cite{hero2012hub} that erroneously asserted a Poisson limit.
\item
The theory in this paper is developed under a novel sparsity condition on the population dispersion matrix. This sparsity condition, called $(\tau, \kappa)$ sparsity in Sec. \ref{sec:taukappaspa}, is significantly weaker than previously assumed conditions, which makes our theory more broadly applicable. Specifically, while the block sparsity condition in previous work \cite{hero2012hub,fan2008sure,firouzi2016two}
imposes that correlation can only occur locally in small blocks of variables, the $(\tau,\kappa)$ sparsity condition relaxes this condition to more general global correlation patterning.
\end{enumerate}
The implications of the above technical contributions are far reaching. We elaborate on these implications below.
Previous conditions on population partial correlation networks assume they are of lower dimension.
In particular, conditions such as block sparsity do not allow for completely connected partial correlation graphs which involve all $p$ variables. Such restrictive assumptions are difficult to validate and rule out many realistic population (partial) correlation structures. Overcoming this hurdle has been an open problem for several years. The newly introduced $(\tau,\kappa)$ sparsity condition on the population covariance matrix settles this longstanding problem by successfully allowing for completely connected partial correlation graphs over the entire set of features. It therefore marks a significant theoretical breakthrough with important implications for correlation screening approaches in application domains.
Historically, the literature on correlation estimation and graphical models has separated the treatments of covariance graph models and undirected graphical models (or inverse covariance graph models) \cite{cox2014multivariate}. Unifying the two classes of statistical models has been an open problem for the better part of almost 3 decades. While this separate treatment may be appropriate in low dimensional settings when there are few variables, it is not immediately obvious which of the two frameworks is appropriate for a given data set in modern ultra-high dimensional regimes. To our knowledge, the framework in this paper is the first in the graphical model or correlation graph estimation literature to propose methodology which brings both approaches under one umbrella.
The results in the paper also have significant relevance to applications. Recall that our Poisson and compound Poisson expressions effectively describe the number of false discoveries and hence allow us to obtain results for the familywise error rate (FWER) or k-FWER, that is the probability of obtaining k or more false discoveries. Note that we can also obtain the marginal distributions of correlation estimation which in turn allow us to obtain expressions for p-values for testing correlation estimates. These marginal p-values allow us to establish FDR control too using either the Benjamini-Hochberg \cite{benjamini1995controlling} or Benjamini-Yakutelli procedures \cite{benjamini2001control}. In summary, the correlations screening framework is sufficiently rich that it allows us to undertake statistical error control in terms of FWER, k-FWER and FDR. This is one of the main strengths of our results: a rigorous inferential framework in the ultra-high dimensional setting.
}
\end{quote}
Another change we made to emphasize the paper's contributions is the last part of the paragraph after Remark \ref{rem:remarkofthm1}, which we quoted below:
\begin{quote}
\ONE{Theorem \ref{cor:Poissonlimit} and Theorem \ref{thm:Poissonultrahigh} specify asymptotic compound Poisson limits and non-asymptotic
bounds on the full distribution of vertex counts. These limits correct and extend the Poisson limits that were falsely claimed in \cite{hero2012hub}, although the compound Poisson limit can be well approximated by the Poisson limit in the case of moderately large $n$ or large $\delta$ (see Sec. \ref{sec:limitingcompoundPoisson}).}
\end{quote}
}
\end{enumerate}
\begin{comment}
\begin{enumerate}
\item
\aoh{The paper presents a unified and complete asymptotic analysis of the edge and hub counts in the random graphs obtained by thresholding the sample correlation and sample partial correlation. In particular, convergence of the full distributions of the edge and hub counts is established as a consequence of finite $p$ bounds on variational norm differences from the respective limiting distributions (Thm XXXX).
In addition, we establish convergent limits of the second moments of the edge and hub counts as a consequence of finite $p$ bounds on deviations from these limits.
This is a significant generalization of previous results \cite{hero2011large, hero2012hub} that only established convergence of the mean number of edges and hub discoveries and the probabilities that these numbers were positive. }
\item
\aoh{The paper establishes that the number of high correlation hubs of fixed degree
converge to a compound Poisson limit as $p\rightarrow \infty$ for any finite $n>3$ (Theorem YYY).
The compound Poisson limit is well approximated by the standard Poisson limit when $n$ is moderately large (Sec. \ref{sec:limitingcompoundPoisson}). This result corrects and refines the claim in \cite{hero2012hub} that erroneously asserted a Poisson limit.}
\item
\aoh{ The theory in this paper is developed under a novel sparsity condition on the population dispersion matrix. This sparsity condition, called $(\tau, \kappa)$ sparsity in Sec. \ref{sec:taukappaspa}, is significantly weaker than previously assumed conditions, which makes our theory more broadly applicable. Specifically, while the block sparsity condition in previous work \cite{hero2012hub} ({\em cite other papers that impose block sparsity for the theory underlying SIS, SPARCS, graphical lasso, compressive sensing, etc}) imposes that correlation can only occur locally in small blocks of variables, the $(\tau,\kappa)$ sparsity condition relaxes this condition to more general global correlation patterning.
} \br{The implications of the results are far reaching as previous conditions effectively assume that population partial correlation network is effectively of lower dimension. Hence conditions such as block sparsity do not allow for a completely connected partial correlation graphs which involve all $p$ variables. Such restrictive assumptions are difficult to validate and rule out many realistic population (partial) correlation structures. Overcoming this hurdle has been an open problem for several years. The newly introduced $(\tau,\kappa)$ sparsity condition settles this longstanding problem by successfully allowing for completely connected partial correlation graphs over the entire set of features. It therefore marks a significant theoretical breakthrough with important implications for correlation screening approaches in application domains."} \yw{It is well written. I wonder would it be slightly better to have it somewhere in the introduction?}
\item
We demonstrate novel connection between correlation to random geometric graphs that have not been previously explored in previous papers. \aoh{If we do not have formal theorems on this in the paper it is best not to call this a principal contribution.}
Details are available at Section \ref{sec:rangeo}.
\item \br{Historically, the literature on correlation estimation and graphical models has separate treatments of covariance graph models and undirected graphical models (or inverse covariance graph models). Unifying the two classes of statistical models has been an open problem for the better part of almost 3 decades (please cite the Cox and Wermuth book). While this separate treatment may be appropriate in low dimensional settings when there are few variables, it is not immediately obvious which of the two frameworks is appropriate for a given data set in modern ultra-high dimensional regimes. To our knowledge, the framework in this paper is the first in the graphical model or correlation graph estimation literature to propose methodology which brings both approaches under one umbrella. } \yw{I still feel the credit of unification of the correlation and partial correlation should go to your previous hub screening paper. }
\end{enumerate}
}
\end{comment}
\newpage
\section{Responses to Referee 1's comments (filename: IT-21-0152.pdf)}
\mycomment{I want to start by acknowledging that, given the scope, the authors have provided a comprehensive
picture about different aspects of the statistical analysis (asymptotic distributions, their
non-asymptotic approximates, approximations for the distribution parameters, and the moments).
Clearly, there has been considerable effort into forming a complete picture. Having said that, I am
somewhat lukewarm about the scope, the exposition, the level of new insights (especially compared
to the existing literature, including those by two of the co-authors), and fitness to the IEEE
Trans. Info. Theory.}
\myresponse{
The reviewer is absolutely correct in mentioning that there has been a considerable effort into forming a complete picture of the problem at hand. We also agree however that there is scope for improvement in highlighting the scope and novelty of the results in the paper. To this end, we have made changes to our paper based on the Referee's specific comments and below we respond to your specific concerns about the scope, exposition, level of insights and fitness to TIT.
}
\begin{enumerate}
\item \mycomment{The exposition and precise statements of the problem solved are missing. The paper starts
with the promise of providing a unified framework for correlation mining. One expects that
irrespectively of the approach taken, there a clear formulation of the problem is provided.
Nevertheless, an objective is specified only the “correlation mining” problem is reduced
(simplified) to a random graph problem. Let me elaborate: given the data model, a heuristic
approach (threshold the pair-wise correlation values) is adopted to generate graphs. These
graphs are only very coarse and incomplete representations of the correlation structures
embedded in the data. The rest of the paper is entirely spent on analyzing these coarse
graphical representations. Perhaps under proper formulation, analyzing such graphs might
become interesting problems, but analyzing their statistical properties is not equivalent to
analyzing the correlation structure embedded in the data.
}
\myresponse{
We didn't fully understand the reviewer's statement that ``the exposition and precise statement of the problem solved are missing." In our original submission, in Sec \ref{sec:intro} we provided an expository description of our contribution, placing it into the context of previous work; in Sec \ref{Hubfra} we stated concisely the problem we are solving and throughout the rest of the paper we provided a precise theorem for our main result (Theorem \ref{cor:Poissonlimit} and Theorem \ref{thm:Poissonultrahigh}). In particular, in the 4th and 5th paragraph of Sec \ref{Hubfra} of the original submission we gave the following concise statement of the problem that we address:
\begin{quote} ``The focus of this paper is correlation and partial screening, which counts the number of
vertices of prescribed degree, the number of star subgraphs, or the number of edges in $\mathcal{G}_\rho(\bm{\Psi})$.
The objective is to characterize the
distributions of these counting statistics.
More specifically, \ldots This paper derives finite sample compound Poisson characterizations of the distributions of the $6$ random quantities $\{N_{E_\delta}^{(k)},N_{\breve{V}_\delta}^{(k)},N_{V_\delta}^{(k)}: k\in \{\bm{R},\bm{P}\} \}$, for $p$ finite and as $p\to\infty$, for suitably chosen $\rho$, under a sparsity assumption on the dispersion parameter $\bm{\Sigma}$."
\end{quote}
On the other hand, we completely agree with the reviewer that the exposition and precise statements of the problem solved are important and we have therefore included in the revision a new subsection (Sec. \ref{sec:contribution}, quoted above in our response to the AE) that describes the principal contributions of the paper.
The Referee is concerned that the our random graph formulation of the correlation mining problem is a ``heuristic approach" and a ``coarse and incomplete representation of the correlation structures embedded in the data."
We believe there is a misunderstanding here and do not believe the above are fully accurate characterizations of the work in the paper. Below we provide more details to clarify this and apologize for any confusion. First of all, any correlation matrix is fully specified by its associated family of random graphs obtained from thresholding the matrix over the full range of threshold levels so random graphs are not a "coarse or incomplete representation." Second of all, correlation mining based on a random graph representation of the population covariance is the basis for the class of Gaussian graphical models, which has a rich history in statistics and computer science (\cite{penrose2003random, koller2009probabilistic}) and have been widely used in applications.
We have also added the following sentence to the last paragraph of Section \ref{Hubfra} to emphasize the implication of our results to correlation mining:
\begin{quote}
\ONE{Such characterizations can be used to test the sparsity structure of the dispersion parameter $\bm{\Sigma}$ or to guide the choice of the threshold $\rho$ \cite{hero2011large, hero2012hub}.}
\end{quote}
We hope the these comments clarify the scope of our work and resolve any confusion caused.
}
\item \mycomment{Compared to the existing studies in [18] and [19], the key contribution in terms of the models
analyzed relates to the sparsity of the covariance matrix. Specifically, earlier studies focused
on the notions of block sparsity, and this paper aims to relax that and introduces a notion
of sparsity that falls in the gap between row sparsity and block sparsity. Specifically, the
covariance matrix, up to certain row/column permutations, has a covariance matrix that
consists of a fully diagonal part, and the other rows not a part of that diagonal part have row
sparsity. This new notion of sparsity is argued to represent a broader class of correlation
structures compared to matrices with block sparsity. This, in principle, is true. However,
in my opinion, such comparisons are somewhat an artifact of the coarse quantization of the
correlation structures by unweighted graphical models. That is, for instance, by changing the
quantization threshold ρ block and $(\eta, \kappa)$ sparse matrices can be transformed to one another.
Essentially, the very same problem considered, by using a different threshold ρ can also be
represented by a graph that has an adjacency that is block-sparse. The core question is now,
with two choices of thresholds and two associated adjacency matrices (one block sparse
and one $(\eta, \kappa)$ sparse) do we arrive at different conclusions about the underlying correlation
structure in the data?
}
\myresponse{The reviewer states well our contribution of extending the theory to a new sparse class of covariance matrices whose sparsity falls between row sparse and block sparse matrices. However, we emphasize that our paper makes several other contributions, including the novel compound Poisson limit/approximations and connections to random geometric graphs, as discussed in the aforementioned new subsection (Section \ref{sec:contribution}) added to the revision.
However, in the remaining reviewer comments there seems to be a misunderstanding of our analysis. Our assumption of $(\tau,\kappa)$ sparsity is on the {\it true (population) covariance matrix $\bm{\Sigma}$} and {\bf not} on {\it the sample covariance matrix}. Thus, in all of our analysis the true covariance sparsity structure is independent of the threshold applied to the sample correlation matrix or sample partial correlation matrix. Put another way, the sparsity of the sample covariance matrix plays no role in our analysis and the thresholding of the \emph{sample correlation} matrix has no bearing on the $(\tau,\kappa)$ or block sparsity of the true correlation.
}
\item \mycomment{A somewhat less significant (but still important) issue is the relevance of the scope to the
T-IT. Let me elaborate: the problem is certainly interesting, and in my opinion, its scope
falls within that of the T-IT. The approach taken, however, does not have an information-
theoretic flavor (even in spirit). Information-theoretically, one would expect that correlation
is viewed and treated at a more fundamental level. In this paper, correlation structure is essentially the pair-wise relationship between the coordinates. More fundamentally, however, one would be examining beyond pair-wise relationships (e.g., via joint distributions and
various marginals). Furthermore, the nature of results is essentially statistical (performance)
analysis for a heuristic “algorithm” that represents the correlation structure by a graphical
model. An information-theoretic aspect of the analysis would involve at least providing
algorithm-agnostic bounds on recovering the correlation structures (as an example). Overall, given the formulation, approach, and nature of results, in my opinion, a statistics venue
can be a better fit for this manuscript.
}
\myresponse{Information theory and information theorists have long been interested in the nature of correlation and the nature of coherence for high dimensional compression, compressive sensing, sensor networks, multi-antenna (MIMO) communications, and so forth. The IEEE Information Theory transactions has been a home for many such papers and, given the unifying theory we offer in our paper, we think that it will be an excellent place to publish our submission. In particular, the fact that we can obtain theoretical limits under a weaker sparsity assumption than block sparsity could inspire other areas of information theory to adopt the $(\tau, \kappa)$ sparsity framework, e.g., in compressive sensing the classical incoherence conditions equivalent to block diagonal sparsity structure on a certain correlation matrix. In addition, information theory has often been the home for unifying papers like ours at the nexus of of computer science, engineering, statistics and mathematics. In addition to the above-mentioned historical reasons, the Transactions on Information Theory is also increasingly publishing work where information theoretic tools are being brought to bear on important theoretical problems in data science involving high-dimensional or ``Big Data" applications - a context where the limits of signal recovery is still a nascent and largely underdeveloped field. For the above-mentioned reasons, we believe that our paper would be a good fit for the Transactions on Information Theory.
}
\item \mycomment{Finally, I would like to comment on the presentation quality: it is difficult to see the key
insights from the presentation and the statements of the results. Of course, I do not mean
that rigor should be compromised, but I believe there could be more effort put into making
the presentation (e.g., notations and flow) more reader-friendly (especially given the length
of the manuscript).
}
\myresponse{We agree with the Referee and acknowledge the challenges in balancing accessibility against mathematical rigor. In the original paper we had attempted to provide insights and remarks to improve the flow, e.g. by motivating the compound Poisson distribution by way of making connections to random geometric graphs in Section \ref{sec:rangeo}. But we agree with the Referee that our presentation can still be improved. We have made several changes to improve the presentation. Firstly we have added a subsection in the introduction to summarizes the contributions and comparisons to the previous papers to motivate readers, as discussed in detail in the response to your Comment 2. Secondly, we have moved the informal derivation to motivate Proposition \ref{prop:edgecor} in Section \ref{sec:closenessnumedge} to Section \ref{sec:derivationprop:edgecor} in the Appendix, which makes the flow in the main text cleaner.
Finally, we have provided additional clarifying comments throughout the paper to provide better linkages between the various theoretical results in the paper.
}
\end{enumerate}
\newpage
\section{Responses to Referee 2's comments (filename: IEEE\_Transactions\_screening\_ultra_high.pdf)}
\mycomment{In previous work, the authors
have established the limiting distributions of various vertex count statistics separately for
correlation and partial correlation screening in ultra-high dimensions. In this paper, they
extend these results by (a) relaxing the block sparsity assumption on the true sequence
of covariance matrices to a form of row-block sparsity which allows variables in a given
block to be correlated with variables outside the block, (b) developing a unified analysis
of the correlation and partial correlation settings, and (c) developing practically feasible
approximations to the parameters of the limiting compound Poisson distributions. These
results represent a significantly novel development, the paper is well-written and the technical
presentation is tight. I do think that more discussion regarding the assumptions and more
detailed simulations are needed to clarify the scope and practical feasibility of the results.
Here are my specific comments.} \\
\myresponse{Thanks for your kind comments on our paper, especially for mentioning that our paper represents a significantly novel development and that the paper is well-written. Furthermore, thank you also for stating that the technical presentation is tight; we have tried to be as precise as possible. We addresses your specific comments below.}
\begin{enumerate}
\item \mycomment{
While the approximation for finite $p$ (Theorem III.11) holds for any $\mu$ and $g$ because of
the invariance of the $U$-scores distribution, how do these quantities affect the constants
appearing in the bounds (and also the assumptions)?}
\myresponse{This is a good question and warrants further clarification. As discussed in Remark \ref{rem:Udisconsequence}, the distribution $\mathscr{L}(\tilde{N}_\delta )$ for $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$ is invariant to $\bm{\mu}$ and $g$. Thus none of the constants in our theorems depend on these quantities. This is because the left hand side $d_{\operatorname{TV}}\left(\mathscr{L}\left(N_{E_\delta}^{(\bm{R})}\right), \operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho})\right)$ in Proposition \ref{prop:edgecor} is also invariant to $\bm{\mu}$ and $g$. Furthermore, all of our assumptions are on $\bm{\Sigma}$, and $n,p,\rho$, which do not depend on $\bm{\mu}$ and $g$. We have added the following sentence in Remark \ref{rem:nonasythm} of Section \ref{sec:compoihighdim}:
\begin{quote}
\ONE{In particular, none of the constants in our theorems depend on $\bm{\mu}$ or $g$.}
\end{quote}
}
\item \mycomment{
Can the assumption of $(\tau, \kappa)$-sparsity be relaxed to “approximate” $(\tau, \kappa)$-sparsity, in the sense that the true sequence of covariance matrices can be approximated by $(\tau, \kappa)$-
sparse matrices so that the difference vanishes at an appropriate rate as $p\to\infty$?}
\myresponse{This is an interesting question. For empirical correlation graphs, the answer is yes. Indeed in Lemma \ref{item:Poissonultrahigha} we present such conclusions under the weaker sparsity condition: row-$\kappa$ sparsity.
For empirical partial correlation graphs the problem is open.
We suspect that our results continue to hold under an approximate $(\tau,\kappa)$ sparse assumption.
For the reviewer's benefit we show in Figure \ref{fig:covasscom} and Figure \ref{fig:ar1} below a preliminary result that suggests that our suspicion is correct. Below are two cases where the sparsity condition is violated. The first case is the rank $1$ spiked covariance model, where $\bm{\Sigma}$ is with diagonal entries $1$ and off diagonal entries $\xi$, and the corresponding numerical simulations are Figure \ref{fig:covasscom}. The second case is the AR(1) model, where $\bm{\Sigma}$ has diagonal $1$ and $k$th off diagonal entries $\xi^k$ with $\xi=0.9$ and $k=1,\ldots,p-1$, and the corresponding numerical simulation is Figure \ref{fig:ar1}.
While these numerical results are compelling evidence that our theory might extend to approximately sparse covariance, this question is our of scope of the current submission. It is a very interesting future direction, and this is
alluded to as a future direction in Section \ref{sec:conclusions} of our paper: ``A future line of work is to characterize the compound Poisson approximations for weaker sparsity conditions.'' We thank the reviewer for raising this point as we too concur that it warrants further and detailed study
\begin{figure}[h]
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[width=.8\linewidth]{graphs/cpvd_xigap07trial2000p1500.png}
\caption{gap=0.7}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[width=.8\linewidth]{graphs/cpvd_xigap09trial2000p1500.png}
\caption{gap=0.9}
\end{subfigure}
\caption{The vertical axis of both (a) and (b) is the approximation error $d_{\operatorname{TV}}(\mathscr{L}\left({N_{V_\delta}^{(\bm{\Psi})}}\right),\operatorname{CP}(\lambda_{n,\delta}(1),\bm{\zeta}_{n,\delta}))$ of the compound Poisson approximation $\operatorname{CP}(\lambda_{n,\delta}(1),\bm{\zeta}_{n,\delta})$ to the empirical distribution $\mathscr{L}\left({N_{V_\delta}^{(\bm{\Psi})}}\right)$.
For both plots the samples are independently generated according to $\mathcal{N}(\bm{0},\bm{\Sigma})$ with $\bm{\Sigma}$ being diagonal $1$ and off diagonal $\xi=\rho-\text{gap}$. The threshold $\rho$ is chosen according to \eqref{eqn:rhopformula} with $e_{n,\delta}=1$.
The blue curve corresponds to the empirical correlation graph (\bm{\Psi} =$\bm{R}$) and the red curve is for the empirical partial correlation graph (\bm{\Psi} =$\bm{P}$). \textbf{Implication of (a)}: As we can see, when the $\text{gap}=0.7$, the partial correlation vertex count $N_{V_\delta}^{(\bm{P})}$ approaches in distribution to the predicted compound Poisson while the distribution of the correlation vertex count $N_{V_\delta}^{(\bm{R})}$ does not. Note in this case the partial correlation matrix is also a rank $1$ spiked model with off diagonal $-\frac{\xi}{1+(p-2)\xi}$, which has smaller off-diagonal entries than the correlation matrix when $p\geq 2$, which might account for the fact that $N_{V_\delta}^{(\bm{P})}$ is closer to the predicted compound Poisson than $N_{V_\delta}^{(\bm{R})}$. \textbf{Implication of (b)}: For (b), we have larger gap, which results in smaller off-diagonal entries for both correlation and partial correlation matrix. In this case both $N_{V_\delta}^{(\bm{P})}$ and $N_{V_\delta}^{(\bm{R})}$ approach to the the non-asymptotic compound Poisson distributions.}
\label{fig:covasscom}
\end{figure}
\begin{figure}[h!]
\center
\includegraphics[scale=0.3]{cpvd_arxi09trial6000p1500.png}
\caption{The vertical axis of both (a) and (b) is the approximation error $d_{\operatorname{TV}}(\mathscr{L}\left({N_{V_\delta}^{(\bm{\Psi})}}\right),\operatorname{CP}(\lambda_{n,\delta}(1),\bm{\zeta}_{n,\delta}))$ of the compound Poisson approximation $\operatorname{CP}(\lambda_{n,\delta}(1),\bm{\zeta}_{n,\delta})$ to the empirical distribution $\mathscr{L}\left({N_{V_\delta}^{(\bm{\Psi})}}\right)$.
The samples are independently generated according to $\mathcal{N}(\bm{0},\bm{\Sigma})$ with $\bm{\Sigma}$ being AR(1) model where $\bm{\Sigma}$ has diagonal $1$ and $k$th off diagonal entries $\xi^k$ with $\xi=0.9$ and $k=1,\ldots,p-1$. The threshold $\rho$ is chosen according to \eqref{eqn:rhopformula} with $e_{n,\delta}=1$. Note that $\bm{\Sigma}^{-1}$ is tri-diagonal, which is sparse, unlike the covariance matrix.
As a result, we see that
$N_{V_\delta}^{(\bm{P})}$ approaches faster in distribution to the non-asymptotic compound Poisson distribution. The covariance matrix $\bm{\Sigma}$, though not sparse, but we know that the entries has $k$th off diagonal entries that are close to zero for $k$ large.
}
\label{fig:ar1}
\end{figure}
}
\item \mycomment{
The assumption $\tau_p = o(p)$ in Theorem II.4 implies that the matrix is almost diagonal.
Can this assumption be weakened by strengthening the assumptions on $\kappa, \mu(\bm{\Sigma})$ or $\rho$?}
\myresponse{ This comment is related to the previous comment and thus our response is similar.
For empirical correlation graphs, the answer is yes. Indeed in Lemma \ref{item:Poissonultrahigha} we confirm this conclusion under the weaker sparsity condition: row-$\kappa$ sparsity. Note that row-$\kappa$ sparsity is $(\tau,\kappa)$ sparsity with $\tau=p$.
For the empirical partial correlation screening, we do not yet know the answer. In terms of theory, our current proof does require that $\tau_p=o(p)$. For an informal argument underlying our current proof, please see \eqref{eqn:Bapprox}. The informal argument assumes that $\bm{\Sigma}$ is diagonal, and our proof is based on applying perturbation theory around the case that $\bm{\Sigma}$ is diagonal, which requires the perturbation to be small, e.g. $\tau_p=o(p)$. To further relax the $(\tau,\kappa)$ assumption with $\tau=o(p)$ is an important and interesting question we are currently pursuing, however we do not yet have a definite answer. However our numerical simulations (see the response to the previous comment) show that the same conclusion appear to hold for approximate $(\tau,\kappa)$ sparse $\bm{\Sigma}$, and even some particular non-sparse $\bm{\Sigma}$ (rank 1 spiked covariance model).
}
\item \mycomment{
I am a little confused by the Remark III.14 regarding the lower bound for $\rho$ in the non-asymptotic and asymptotic cases. I understand the Theorem III.11 provides a non-
asymptotic bound for every finite $p$, but still assumes the condition $2p^{1+1/\delta} P_n(r_\rho) < \gamma$,
which implies that
$$
\rho > 1- O(p^{-(1+\delta/2})^{2/(n-2)}.
$$
As $p\to\infty$, this is the same rate as in Theorem II.4. Hence, unless I am missing
something, Theorem III.11 also requires $\rho$ to increase to 1 as $p$ increases. So I think the phrase “Theorem III.11 does not require that $\rho$ approach 1” in Remark III.14 may
be a bit misleading.
}
\myresponse{ We agree that the phrase “Theorem III.11 does not require that $\rho$ approach 1” in Remark III.14 could be misleading. The main goal of Remark III.11 in the original submission was to compare the convergence rate between Theorem \ref{cor:Poissonlimit} and Theorem \ref{thm:Poissonultrahigh}. We have now moved the original Remark III.11 to Remark \ref{rem:comparisonsfinitelimit} in Section \ref{sec:proofoftheorem1} of the revision, which is a more suitable place to compare convergence rates. We have also updated our remark to avoid
any confusing or potentially misleading phrases:
\begin{quote}
\ONE{
In \eqref{eqn:slowupperbound} of the proof of Lemma \ref{lem:compoundpoissonlimit}, it is shown that part of the upper bound for its error rate is of the order $1-\rho$. This particular rate, however, decreases as $n$ increases. More concisely, if one chooses $\rho$ according to \eqref{eqn:rhopformula} then $1-\rho$ is of the order $p^{-(1+\frac{1}{\delta})\frac{2}{n-2}}$. Hence the convergence to $\operatorname{CP}(\lambda_{n,\delta}(e_{n,\delta}),\bm{\zeta}_{n,\delta})$ in Theorem \ref{cor:Poissonlimit} is only accurate for large $p$. On the other hand, the upper bound in Theorem \ref{thm:Poissonultrahigh} only depends on $\rho$ through $2p^{1+\frac{1}{\delta}}P_n(r_\rho)\leq \gamma$, which holds for small $p$, again if $\rho$ is chosen according to \eqref{eqn:rhopformula} (see the first paragraph of Remark \ref{rem:edgecor} for related discussion). Hence Theorem \ref{thm:Poissonultrahigh} provides an accurate approximation to $\bar{N}_{\delta}$ even for small $p$. The accuracy of these approximations for various values of $\rho$ and $n$ is numerically illustrated in Figure 8 in the Supplementary Material (\cite{hero2020unified}, Appendix K).
}
\end{quote}
}
\item \mycomment{
Given the above comment, it would be useful to provide a more detailed discussion of
why the approximation in Theorem III.11 is asymptotically better than Theorem II.4
(as illustrated in Figure 8 and Figure 9).
}
\myresponse{ We do not claim asymptotic superiority of Theorem \ref{thm:Poissonultrahigh} over Theorem \ref{cor:Poissonlimit} and apologize if it comes across this way. We only claim that when $p$ is finite and small, the non-asymptotic compound Poisson distribution approximates the random counts better than the limiting compound Poisson distributions. We are sorry that some of our phrases in the original submission misled the Referee. We have updated Remark \ref{rem:comparisonsfinitelimit} (Remark III.11 in the previous version) as already discussed in the response to your Comment \# 4, and we have also updated the caption of Figure 8 to avoid potentially misleading phrases.}
\item \mycomment{
While the illustrations in Figure 8 and Figure 9 are very informative, it will be useful to
have more illustrations which monitor the rates of convergence as the rate of increase
of $\tau$ and $\kappa$ with $p$ varies. Additionally, it will be interesting to see the empirical performance of the approximations even when some of the assumptions are violated. Finally,
some illustrations when the data is generated from non-Gaussian distributions (different choices of $g$) will also be informative of how the performance of the approximations
changes with $g$.
}
\myresponse{
For the benefit of the Referee, we briefly provide some additional numerical experiments to respond.
Figure \ref{fig:tau} illustrates the the compound Poisson approximation for different values of $\tau$, and the simulation indicates that the approximation accuracy is not sensitive to $\tau$.\\
\begin{figure}[h!]
\center
\includegraphics[scale=0.6]{kappa8tau0246.png}
\caption{The vertical axis is $d_{\operatorname{TV}}(\mathscr{L}\left({N_{V_\delta}^{(\bm{\Psi})}}\right),\operatorname{CP}(\lambda_{n,\delta}(1),\bm{\zeta}_{n,\delta}))$.
The samples are independently generated according to $\mathcal{N}(\bm{0},\bm{\Sigma})$ with $\bm{\Sigma}$ being $(\tau,\kappa)$ sparse. The threshold $\rho$ is chosen according to \eqref{eqn:rhopformula} with $e_{n,\delta}=1$. We fix $\kappa=p^{0.8}$ and vary $\tau$ in $p^{0.6},p^{0.4}, p^{0.2}, 0$. Note $\tau=0$ means that $\bm{\Sigma}$ is the identity matrix. The four choices of $\tau$ correspond to the four curves in the plot. But the four curves highly overlapped and thus from this particular experiment we see that the accuracy of the non-asymptotic compound approximation is not sensitive to $\tau$.}
\label{fig:tau}
\end{figure}
Figure \ref{fig:kappa} illustrates the the compound Poisson approximation for different values of $\kappa$, and the simulation indicates that the approximation accuracy is not sensitive to $\kappa$.
\begin{figure}[h!]
\center
\includegraphics[scale=0.6]{tau0.6kappa0258.png}
\caption{The vertical axis is $d_{\operatorname{TV}}(\mathscr{L}\left({N_{V_\delta}^{(\bm{\Psi})}}\right),\operatorname{CP}(\lambda_{n,\delta}(1),\bm{\zeta}_{n,\delta}))$.
The samples are independently generated according to $\mathcal{N}(\bm{0},\bm{\Sigma})$ with $\bm{\Sigma}$ being $(\tau,\kappa)$ sparse. The threshold $\rho$ is chosen according to \eqref{eqn:rhopformula} with $e_{n,\delta}=1$. We fix $\tau=p^{0.6}$ and vary $\kappa$ in $p^{0.8},p^{0.5}, p^{0.2}, 1$. Note $\kappa=1$ means that $\bm{\Sigma}$ is the identity matrix. The four choices of $\tau$ correspond to the four curves in the plot. But the four curves highly overlapped and thus from this particular experiment we see that the accuracy of the non-asymptotic compound approximation is not sensitive to $\kappa$.}
\label{fig:kappa}
\end{figure}
Furthermore, in our response to the Referee's Comment \# 2, we have provided numerical simulations for the case that sparsity assumption is violated.
Please refer there for the details.
Finally,
as mentioned in our response to the Referee's Comment \#1, our theoretical results are invariant to the choice of $\bm{\mu}$ and $g$, as proved and discussed in detail in Lemma \ref{lem:Uscoresdistribution} and Remark \ref{rem:Udisconsequence}.
As a result, we are in a position to respond to the reviewer's comments in the affirmative, that is our results go through regardless of $g$ (or $\bm{\mu}$).
Notwithstanding the interest of the affirming numerical results in Figures 1-4 in this response, we do not think that this theory paper is the appropriate venue for such empirical mathematical arguments.
}
\end{enumerate}
\newpage
\newpage
\section{Responses to Referee 3's comments (filename: IT-21-0152-rev.pdf)}
\begin{enumerate}
\item \mycomment{
The paper is very carefully written. It might even be a bit too pedantic in places, as the
core message is not too surprising for people that are familiar with the general area, and
this makes the paper difficult to read in many places (see below). Some of the comments,
however, lack depth, in that the paper describes what it does but does not make connections
with other areas of random graph theory where degree sequences are studied and, sometimes,
shown to converge to a Poisson or mixed Poisson distribution.
}
\myresponse{We are glad that the Referee found the paper to be very carefully written but we are sorry
to hear that the Referee thinks ``the core message is not too surprising for people that are familiar with the general area''. We apologize that we did not originally adopt an exposition that best conveyed our major contributions. To clarify our contributions, we have added a subsection in the introduction, which was quoted in our response to the AE. For the Referee's benefit we replicate this quote below:
\begin{quote}
\ONE{We summarize the principal contributions of the paper. As above $p$ denotes the number of variables and $n$ denotes the number of samples.
\begin{enumerate}
\item
The paper presents a unified and complete asymptotic analysis of the star subgraph counts, the counts of vertices of a given degree and the counts of vertices above a given degree in the random graphs obtained by thresholding the sample correlation and sample partial correlation. This unification of different types of random counts represents an important improvement over previous work \cite{hero2011large,hero2012hub} where only
the counts of vertices above a given degree are studied.
\item
We approximate the full distributions of the random counts for finite $p$ and as $p\to\infty$. In addition, we characterize the first and second moments of these random counts. This is a significant generalization of previous results \cite{hero2011large, hero2012hub} that only established approximations for the mean number of random counts and for the probabilities that these counts were positive.
\item
We obtains compound Poisson characterizations of the the distributions of the random counts.
The compound Poisson limit and approximation are well approximated by the standard Poisson limit when $n$ is moderately large (Section \ref{sec:limitingcompoundPoisson}). This result corrects and refines the claim in \cite{hero2012hub} that erroneously asserted a Poisson limit.
\item
The theory in this paper is developed under a novel sparsity condition on the population dispersion matrix. This sparsity condition, called $(\tau, \kappa)$ sparsity in Sec. \ref{sec:taukappaspa}, is significantly weaker than previously assumed conditions, which makes our theory more broadly applicable. Specifically, while the block sparsity condition in previous work \cite{hero2012hub,fan2008sure,firouzi2016two}
imposes that correlation can only occur locally in small blocks of variables, the $(\tau,\kappa)$ sparsity condition relaxes this condition to more general global correlation patterning.
\end{enumerate}
The implications of above technical contributions are far reaching. We elaborate on these implications below.
Previous conditions on population partial correlation networks assume they are of lower dimension.
In particular, conditions such as block sparsity do not allow for completely connected partial correlation graphs which involve all $p$ variables. Such restrictive assumptions are difficult to validate and rule out many realistic population (partial) correlation structures. Overcoming this hurdle has been an open problem for several years. The newly introduced $(\tau,\kappa)$ sparsity condition on the population covariance matrix settles this longstanding problem by successfully allowing for completely connected partial correlation graphs over the entire set of features. It therefore marks a significant theoretical breakthrough with important implications for correlation screening approaches in application domains.
Historically, the literature on correlation estimation and graphical models has separated the treatments of covariance graph models and undirected graphical models (or inverse covariance graph models) \cite{cox2014multivariate}. Unifying the two classes of statistical models has been an open problem for the better part of almost 3 decades. While this separate treatment may be appropriate in low dimensional settings when there are few variables, it is not immediately obvious which of the two frameworks is appropriate for a given data set in modern ultra-high dimensional regimes. To our knowledge, the framework in this paper is the first in the graphical model or correlation graph estimation literature to propose methodology which brings both approaches under one umbrella.
The results in the paper also have significant relevance to applications. Recall that our Poisson and compound Poisson expressions effectively describe the number of false discoveries and hence allow us to obtain results for the familywise error rate (FWER) or k-FWER, that is the probability of obtaining k or more false discoveries. Note that we can also obtain the marginal distributions of correlation estimation which in turn allow us to obtain expressions for p-values for testing correlation estimates. These marginal p-values allow us to establish FDR control too using either the Benjamini-Hochberg \cite{benjamini1995controlling} or Benjamini-Yakutelli procedures \cite{benjamini2001control}. In summary, the correlations screening framework is sufficiently rich that it allows us to undertake statistical error control in terms of FWER, k-FWER and FDR. This is one of the main strengths of our results: a rigorous inferential framework in the ultra-high dimensional setting.
}
\end{quote}
We are not sure why the Referee thinks our results are not too surprising.
The conclusions obtained in our paper were indeed quite surprising to us for a number of reasons. For example, the (incorrect) Poisson approximation from previous work \cite{hero2012hub} had been accepted as the correct limiting distribution - not just by ourselves, but by many others who had used the work. Had the Poisson result been obviously incorrect and the compound Poisson been the obvious alternative, it would have made sense for the community to challenge the incorrect result and immediately correct it - but since this did not happen, we and others did not see the appropriateness of the compound Poisson right away. It only came about after careful analysis. In summary, the results in the paper were not so obvious to us.
We next address your comment: ``Some of the comments, however, lack depth, in that the paper describes what it does but does not make connections with other areas of random graph theory where degree sequences are studied and, sometimes, shown to converge to a Poisson or mixed Poisson distribution.'' We are somewhat confused by this comment since we feel that we had specifically addressed the connections to random geometric graphs in a dedicated paragraph (quoted below) in Section \ref{sec:rangeo}. We feel that the connection and the comparisons between our models and the existing random geometric graphs are covered
in the sense that, while for the cognoscenti, existing results might motivate our conclusion, our model and hypotheses (and thus our proofs) are quite different from those used in random graph theory. We have added the content highlighted in red to further clarify these connections.
\begin{quote}
The random pseudo geometric graph in Definition \ref{def:pge} \ONE{is similar} to the random geometric graph introduced in \cite{penrose2003random}.
In particular, studied in the monograph \cite{penrose2003random} is the number of \emph{induced} subgraphs isomorphic to a given graph, typical vertex degrees, and other graphical quantities of random geometric graphs.
\ONE{As a specific example, the rate $a_n2^{n/2}p^{1+\frac{1}{\delta}} (1-\rho)^{\frac{n-2}{2}}\to e(n, \delta)$ as $p\to \infty$ in Theorem \ref{cor:Poissonlimit} is equivalent to $2a_np^{1+\frac{1}{\delta}} r_\rho^{{n-2}}\to e(n, \delta)$, which is consistent with existing Poisson approximation for random geometric graph \cite[ Theorem 3.4]{penrose2003random}.
}
The differences between our random pseudo geometric graph in Definition \ref{def:pge} and the random geometric graphs defined in \cite{penrose2003random} are: 1) our graph has vertices $\bm{u}_i$ lying on the unit sphere instead of on the entire Euclidean space; 2) our graph is induced by distance $\textbf{dist}(\cdot,\cdot)$ instead of the Euclidean distance. Another key difference is that vertices $\bm{u}_i$ are not necessarily independent in Definition \ref{def:pge}. Indeed in our model, the correlations between vertices $\bm{u}_i$ are encoded by a sparse matrix $\bm{\Sigma}$ (cf. Lemma \ref{lem:Uscoresdistribution} \ref{item:Uscoresdistributiona}), whereas in \cite{penrose2003random} the vertices associated with the random geometric graph are assumed to be i.i.d. In \cite{penrose2003random} it was \ONE{stated (Example after \cite[Corollary 3.6]{penrose2003random}) without proof} that the number of vertices with degree at least $3$ was approximately compound Poisson. A similar compound Poisson limit is established in Theorem \ref{cor:Poissonlimit}. \ONE{There is some recent work on testing whether a given graph is a realization of Erd\H o s–R\'enyi random graphs or a realization of random geometric graphs with vertices i.i.d. uniformly distributed on the sphere; see \cite{bubeck2016testing} and the reference therein.}
\end{quote}
Two related comments are the Referee's Other Comments \# 8 and \# 10, addressed in detail below, where some papers on random graphs are suggested by the Referee. For the Comment \# 8, we agree that it is related and therefore we added the last sentence of the above paragraph, colored in red. But we do not see this missing reference as a paper very closely related paper to support the Referee's claim in the comment. We address the relationship to Other Comments \# 10 in the next paragraph.
Finally, we point out that our results differ substantially from Chapter 6 of the van der Hofstadt book (in your Comment \# 10) that shows convergence of a different limit (Poisson or mixed Poisson) of a different class of random graphs. We provide more details below in our response to our response to the Referee's Comment \# 10.
}
\begin{comment}
The random pseudo geometric graph in Definition \ref{def:pge} bears some similarities to the random geometric graphs introduced in \cite{penrose2003random}.
In particular, the monograph \cite{penrose2003random} studies the number of \emph{induced} subgraphs isomorphic to a given graph, typical vertex degrees, and other graphical quantities of random geometric graphs. \ONE{As a specific example, the rate $a_n2^{n/2}p^{1+\frac{1}{\delta}} (1-\rho)^{\frac{n-2}{2}}\to e(n, \delta)$ as $p\to \infty$ in Theorem \ref{cor:Poissonlimit} is equivalent to $2a_np^{1+\frac{1}{\delta}} r_\rho^{{n-2}}\to e(n, \delta)$, which is consistent with existing Poisson approximation for random geometric graph \cite[ Theorem 3.4]{penrose2003random}.} The differences between our random pseudo geometric graph in Definition \ref{def:pge} and the random geometric graphs defined in \cite{penrose2003random} are: 1) our graph has vertices $\bm{u}_i$ lying on the unit sphere instead of on the entire Euclidean space; 2) our graph is induced by distance $\textbf{dist}(\cdot,\cdot)$ instead of the Euclidean distance. Another key difference is that vertices $\bm{u}_i$ are not necessarily independent in Definition \ref{def:pge}. Indeed in our model, the correlations between vertices $\bm{u}_i$ are encoded by a sparse matrix $\bm{\Sigma}$ (cf. Lemma \ref{lem:Uscoresdistribution} \ref{item:Uscoresdistributiona}), whereas in \cite{penrose2003random} the vertices associated with the random geometric graph are assumed to be i.i.d. In \cite{penrose2003random} it was stated (Example after \cite[Corollary 3.6]{penrose2003random}) without proof that the number of vertices with degree at least $3$ was approximately compound Poisson. A similar compound Poisson limit is established in Theorem \ref{cor:Poissonlimit}. \ONE{There is some recent work on testing whether a given graph is a realization of a Erd\H o s–R\'enyi random graph or a realization of a random geometric graph with vertices i.i.d. uniformly distributed on the sphere; see \cite{bubeck2016testing} and the reference therein.}
\end{comment}
\item \mycomment{
The mathematics are carefully laid out over many pages of calculations that develop a (com-
pound) Poisson approximation for, essentially, the degree sequence of a sparse correlation
graph. The paper, in Section III in particular, tries to give an intuition for the proof arguments, but I think fails very short of that. What is the difference with basic Poisson
approximation theory (e.g., [34])?
}
\myresponse{
We appreciate your comment that we had carefully laid our our mathematical arguments, and we did this so as to provide the most complete and reproducible presentation to the reader. In regards to your comments about subsection \ref{sec:closenessnumedge}, which attempted to provide intuition for the proof arguments, we have moved it to the Appendix (Sec. \ref{sec:derivationprop:edgecor}) in the revision.
In response to your comment on the difference of our model with respect to the basic Poisson approximation theory of [34],
as pointed out in our response to your Comment \# 1, in Section III-B of the submitted version we have included a paragraph (see our response to your Comment \# 1) dedicated to making comparisons between our random pseudo geometric graph models and the random geometric graphs studied in \cite{penrose2003random} ([34] in the original submission). Please see our response to your Comment 1 for the quote of that paragraph.
We point out that this informal derivation is in fact more closely related to the arguments used in \cite{barbour2001topics} ([33] in the submitted version) than it is to \cite{penrose2003random} ([34] in the submitted version)
and this is why it required introducing additional notation.
}
\item \mycomment{
My main concern is the novelty. The improvement over [19] (described bottom page 5 - top
page 6) is rather incremental. I do not doubt that the improvement is technical, but [19] was
already published in IEEE IT. Is the topic so important that it warrants this? In addition the
first paper in the line, [18], was published in JASA, a top statistics journal. If the contribution
with respect to these papers is purely technical, then perhaps a more mathematically-oriented
venue (a journal in probability theory) would be more appropriate.
}
\myresponse{Admittedly we did not do a good job at clearly bringing out our contribution and novelty with respect to prior work. In the revision we have made changes to emphasize the novelty, key contributions, and comparisons with the prior work [18,19]. Please see our response to your Comment 1 for details. We do not believe that our contributions are incremental.
Rather, we believe our manuscript represents significant progress on this important problem since it unifies and extends previous results, corrects previous results, has stronger conclusions than previous results, and weakens the assumptions necessary to obtain the previous results.
The implications of above technical contributions in our response to your Comment \# 1 are far reaching.
}
\end{enumerate}
\mycomment{So I am rather lukewarm about the paper, even though it's easy to recognize that a lot of careful
work has gone into it.}
\myresponse{
We trust that our careful summary of our significant contributions
will paint a positive picture of the novelty of the results in the paper. We take responsibility for not bring these out in our original submission of the paper and apologize for this. We have responded to all the comments of the Referee carefully and also modify our original submissions
accordingly, including but not limited to adding a new subsection to clarify the main contribution, moving some informal arguments to the appendix, adding some suggested references, improving some notations. We are hopeful that our responses and modifications will result in a warmer reaction by the Referee to our revised paper.
}
\\
\vspace{4mm}
\textbf{Other comments}
\begin{enumerate}
\item \mycomment{
The bibliography is perhaps a bit incomplete as few works on sparse covariance matrix estimation are cited. There are also works that are closely and very closely related, such as the following, that are missing: \\
Arias-Castro, E., Bubeck, S., & Lugosi, G. (2012). Detection of correlations. Annals of Statistics, 40(1). \\
Fan, J., Shao, Q. M., & Zhou, W. X. (2018). Are discoveries spurious? Distributions of
maximum spurious correlations and their applications. Annals of Statistics, 46(3). \\
The work here is instead presented as building on work by some of the authors [18, 19].
}
\myresponse{Thanks for suggesting the two related references. After reading both papers, we see that there are connections at a superficial level, i.e., they also consider sample covariance matrices in high dimensions. These papers provide additional evidence that the general problem of correlation mining is important and of broad interest, so in the revision we have included these references in the last sentence of the first paragraph on page 2:
\begin{quote}
\ONE{The reader is referred to \cite{anandkumar2009detection,castro2014detection, tan2014learning, heydari2017quickest, tarzanagh2018estimation, arias2012detection, fan2018discoveries} and the references therein for additional work related to high dimensional sample correlation matrices.}
\end{quote}
However, we do not see the close relation that the Referee claimed. We point out that the fact that these papers did not cite our previous work \cite{hero2011large} or \cite{hero2012hub} ([18,19] in the submitted version of our paper), indicating that these authors
did not themselves think their work was closely related to \cite{hero2011large,hero2012hub} ([18,19] in the submitted version of our paper).
Moreover, their models assumptions and results are very different.
}
\begin{comment}
\marginpar{to be updated}
\aoh{However, we point out the fact that these papers did not cite our previous work \cite{hero2011large} or \cite{hero2012hub} ([18,19] in the submitted version of our paper), indicating that these authors were either not aware of [18,19] or did not themselves think their work was closely related. \ONENEW{I do not think the awareness can be used as an explanation.} \br{ Would it not be better to state the second reply first and then thereafter mention this comment ? I thought this comment may be seen as less aggressive if it follows the second comment}
This said, after reading both papers, we do see that there are connections at a superficial level, i.e., they also consider sample covariance matrices in high dimensions. However, their models and assumptions are very different so we do not see the close relation to which you refer.
Nonetheless, these papers do provide additional evidence that the general problem of correlation mining is important and of broad interest,} so in the revision we have included these references in the last sentence of the first paragraph on page 2:
\begin{quote}
\ONE{The reader is referred to \cite{anandkumar2009detection,castro2014detection, tan2014learning, heydari2017quickest, tarzanagh2018estimation, arias2012detection, fan2018discoveries} and the references therein for additional work related to high dimensional sample correlation matrices.}
\end{quote}
\end{comment}
\item \mycomment{
I do not understand the sentence “We define the set ...” on page 3. The notation up to that point is strange. Why not $\bm{R}$ in place of $\bm{\Psi}^{(\bm{R})}$? That whole paragraph could be simpler. I
would define $\mathcal{G}_\rho(\bm{\Psi})$ for any n-by-n matrix $\bm{\Psi}$, for example.
}
\myresponse{ We agree with you that ``$\bm{R}$ in place of $\bm{\Psi}^{(\bm{R})}$'' is simpler. We have implemented these changes in the revision. Thanks for the suggestion.
}
\item \mycomment{
$\mathcal{VE}$ for vector-elliptical is rather confusing in the presence of $(\mathcal{V}, \mathcal{E})$ for a graph.
}
\myresponse{Thanks for pointing this out. We have changed $(\mathcal{V}, \mathcal{E})$ to be $(\mathrm{V}, \mathrm{E})$
}
\item \mycomment{
Again, the notation could be improved. For example, in $\Ge(\{v_i\}, r; n-2)$, unless I am missing
something, the ”n-2” part is already given implicitly in that it is the dimension of the space
where the points {$v_i$} reside.
}
\myresponse{ We agree with you and have implemented the changes. Thanks for the suggestion.
}
\item \mycomment{
In (7), a different summation variable needs to be used (other than $\ell$).
}
\myresponse{ We have updated the variable. Thanks for pointing this out.
}
\item \mycomment{
Again, $\delta$ in this part of the paper is less than optimal, having been recently used to denote
something very different.
}
\myresponse{We are confused by your comment. Do you mean the $\delta$ in (7)? If yes, the $\delta$ in (7) is the hub level $\delta$ and there is no conflict in notation that we can see.
}
\item \mycomment{
Section II.D is not particularly enlightening.
}
\myresponse{Section \ref{sec:upperboundjointdensity} presents some simple observations and examples for the new concept of local normalized determinant, which is used in Theorem \ref{cor:Poissonlimit}. To clarify the role of this section in the paper, we have added the following sentence in the first paragraph of Section \ref{sec:upperboundjointdensity}
\begin{quote}
\ONE
The normalized determinant measures the closeness between the identity matrix and the dispersion parameter $\bm{\Sigma}$. This plays an important role in Theorem \ref{cor:Poissonlimit}.}
\end{quote}
}
\item \mycomment{
In regards to Section III.B, the authors must know that (random) geometric graphs have
been considered in settings other than the Euclidean setting. In particular, there is a line of
work on testing a random graph (Erd\H o s–R\'enyi) vs a random geometric graph on the (high-
dimensional) sphere, which the authors might find interesting as it is somewhat related to
the topic at hand. See the following paper and others that cited it: \\
Bubeck, S., Ding, J., Eldan, R., & R\'acz, M. Z. (2016). Testing for high-dimensional geometry
in random graphs. Random Structures & Algorithms, 49(3).
}
\myresponse{
Thanks for bringing this interesting paper to our attention.
The suggested paper is clearly related since they also study random geometric graph with vertices i.i.d. uniformly distributed on the sphere. The difference is that our model as a random pseudo geometric graph is defined in terms of a pseudo distance $\textbf{dist}(\bm{v}_i,\bm{v}_j):= \min\left\{\|\bm{v}_i-\bm{v}_j\|_2,\|\bm{v}_i+\bm{v}_j\|_2 \right \} $ instead of the Euclidean distance, and our model allows possible correlation between vertices while theirs does not. We have added the following sentence in Section \ref{sec:rangeo} where we discuss the existing literature of random geometric graphs.
\begin{quote}
\ONE{There is some recent work on testing whether a given graph is a realization of a Erd\H o s–R\'enyi random graph or a realization of a random geometric graph with vertices i.i.d. uniformly distributed on the sphere; see \cite{bubeck2016testing} and the reference therein.
}
\end{quote}
Note that one could potentially choose any of the random counts studied in our paper as a test statistic in their problem setting using our results. One may be able to characterize the distribution of the test statistic under the alternative hypothesis.
Our paper can also potentially inspire people to study random geometric graphs with possible dependence among vertices of the random geometric graphs in their problem setting.
}
\item \mycomment{
Prop III.9 (b), do the constants in the bound also depend on $c$?
}
\myresponse
Indeed this is an insightful comment. The constants in the upper bound of Proposition \ref{thm:allclose} do not depend on $c$. As discussed in Remark \ref{rem:allclose}, the condition that involves $c$ is a quantitative way of saying that $p$ is sufficiently large. So if $c$ is sufficiently small, which implies that $p$ is sufficiently large, then the inequality in the conclusion of Proposition \ref{thm:allclose} \ref{item:allclosec} is valid. However the constants in the upper bound do not functionally depend on $c$.
To further clarify this point, we have written ``positive and \ONE{sufficiently} small
$c$'' in the statement of Proposition \ref{thm:allclose} \ref{item:allclosec}, and also added the following sentence in Remark \ref{rem:allclose}:
\begin{quote}
\ONE{
The inequalities (\ref{eqn:allclosec}) and (\ref{eqn:allclosed}) are valid as long as $c$ is less than finite constant, given respectively in \eqref{eqn:cdef} and \eqref{eqn:cuppbou2} in the supplementary material \cite{hero2020unified} to this paper.
}
\end{quote}
We have also updated other related places in the paper to be ``positive and \ONE{sufficiently} small $c$''.
}
\item \mycomment{
Section III is hard to read because of the large amount of notation introduced there... and
the payoff is not at all clear. I did not get any clear intuition on proof arguments beyond
how typical Poisson approximations work — which I am somewhat familiar with, at least at
the level of [34]. It would have been more instructive to understand how the setting differs
from the usual Poisson approximation setting, and perhaps draw some relations with what is
known for generalized random graphs (see the book by van der Hofstad, Ch 6).
}
\myresponse{ We think this is related to your major Comment \#2. We have discussed in great detail the differences between our informal arguments and the book [34] in our response to your major Comment 2. We also moved the informal derivation to the appendix. Please refer to our responses to your major Comment 2 for more details.
Relative to the last sentence of your comment,
Chapter 6 of the van der Hofstadt book covers a very different class of random graphs than the random pseudo-geometric graphs to which the theory in our paper applies. The class of random graphs in Ch. 6 of the reference are called generalized random graphs and, similarly to other inhomogeneous versions of the Erdos-Renyi graph, the edges are random and drawn indpendently with probabilities $p_{ij}$ of there existing an edge between vertices $i$ and $j$. Specifically, for the generalized random graph $p_{ij}=W_iW_j/(L_n+W_iW_j)$ where $W_i\geq 0$ is a scalar attribute (called a weight in Ch. 6 of the reference) and $L_n=\sum_{i=1} W_i$. On the other hand the random pseudo-geometric graph studied in our paper has random edges that are {\it dependent} with marginal probability $p_{ij}$ equal to threshold exceedance probability of the magnitude correlation or partial corelation between {\it vector-valued vertex attributes} $X_i$ that follow a elliptically contoured multivariate distribution. In Sec 6.3 van der Hofstadt obtains a mixed Poisson limiting distribution for the vertex degree.
The mixed Poisson and the compound Poisson are very different distributional limits for the vertex counts of two very different random graph models.
}
\item \mycomment{
Similarly, what is the utility of Section V? Isn’t the approximation of a compound Poisson
distribution with a Poisson distribution well-understood? I would have liked to see some
numerical experiments, perhaps using genetics data, on how accurate and useful the formulas
developed in the paper are. After all, the paper is motivated by such applications. The
discussion (Section VI) mentions some numerical experiments. (I am completely fine with a
purely theoretical paper, but then, what is the purpose of Section V?)
}
\myresponse{Yes, you are right that the paper is envisioned to be a purely theoretical paper.
As stated at the beginning of Section \ref{sec:limitingcompoundPoisson}, the purpose of this section is provide useful approximations to the parameters (\which are defined in terms of random geometric graphs) for the compound Poisson in the context of the sample correlation graph.
The main contribution in Section \ref{sec:limitingcompoundPoisson} lies in Lemma \ref{lem:randgeogra} and Lemma \ref{lem:randgeograconduppbou}, for which we apply concentration inequalities to random (pseudo) geometric graphs. To the best of our knowledge, these two lemmas are novel and we think they are quite interesting.
That we obtain a Poisson approximation to the compound Poisson distribution is not our goal/purpose in Section \ref{sec:limitingcompoundPoisson}, but is rather a consequence or a byproduct of the estimation/approximation to the complicated quantity in the random (pseudo) geometric graphs established in Lemma \ref{lem:randgeogra} and Lemma \ref{lem:randgeograconduppbou}.
However, if the Referee thinks Section \ref{sec:limitingcompoundPoisson} is well-understood in the literature, please provide a specific resource that has the same or similar results, and we would be happy to update the section accordingly.
}
\end{enumerate}
\newpage
\section{Responses to Referee 4's comments (filename: IEEE Trans Inf Theory report.pdf)}
\mycomment{I found the paper pretty interesting and felt that it definitely made novel contributions.
The paper was quite comprehensive and thorough in its theoretical treatment of correlation
and partial correlation screening. It definitely made interesting connections to random geometric graph theory. A lot of questions that I had while reading the paper were answered in
the remarks of the paper or in the Appendices. However, I have a few concerns that I hope
the authors can address, which I detail below.}
\myresponse{Thank Referee for positive comments on the comprehensiveness and the thoroughness of our paper. We try to be as precise as possible while try to make the results accessible by providing many remarks. We have provided responses to the Referee's specific concerns below.}
\begin{enumerate}
\item \mycomment{
Theorem II.4, the authors require the correlation threshold $\rho$ to grow at a very
specific function of $p$, i.e. $a_n2^{n/2}p^{1+\frac{1}{\delta}} (1-\rho)^{\frac{n-2}{2}}\to e(n, \delta)$ as $p\to \infty$. This seems to
be an artifact of the proof, and it is a bit troubling (even though the authors pointed
out later that the growth rate is not fast). Is this a sharp rate, i.e. can this condition
not be relaxed at all? Does the distributional limit result automatically fail when this
condition on the growth rate of $\rho$ does not hold? This deserves greater attention, and
the authors may want to prove that this rate is sharp when $p\to\infty$. Otherwise, it
raises a lot of questions and concerns.
}
\myresponse{This rate is not an artifact of the proof, it is a sharp rate. From Section \ref{sec:rangeo} we know that our model is a Random Psuedo Geometric Graph with radius parameter $r_\rho = \sqrt{2(1-\rho)}$. Thus the rate is equivalent to $2a_np^{1+\frac{1}{\delta}} r_\rho^{{n-2}}\to e(n, \delta)$, which is consistent with existing Poisson approximations for random geometric graphs \cite[ Theorem 3.4]{penrose2003random}. We have added such discussions in Section \ref{sec:rangeo}:
\begin{quote}
\ONE{ As a specific example, the rate $a_n2^{n/2}p^{1+\frac{1}{\delta}} (1-\rho)^{\frac{n-2}{2}}\to e(n, \delta)$ as $p\to \infty$ in Theorem \ref{cor:Poissonlimit} is equivalent to $2a_np^{1+\frac{1}{\delta}} r_\rho^{{n-2}}\to e(n, \delta)$, which is consistent with existing Poisson approximation for random geometric graph \cite[ Theorem 3.4]{penrose2003random}. }
\end{quote}
In Remark \ref{rem:remarkofthm1} we also added:
\begin{quote}
\ONE{This particular rate on $\rho$ is consistent with the rate
of the existing Poisson approximation results in
random geometric graphs \cite{penrose2003random} as will be
discussed in Section \ref{sec:rangeo}.}
\end{quote}
The rate is sharp in the sense that letting $\rho$ grow at a faster or slower rate the result will be a degenerate compound Poisson limit (a limit either having mean 0 or mean $\infty$). This was briefly discussed in Lemma \ref{lem:compoundpoissonlimit} in the submitted paper.
For the benefit of the Referee, we provide some additional details on the sharpness of this rate that we choose not to include in the revision.
Since we want to establish a result that holds for a certain type sparse dispersion parameter $\bm{\Sigma}$, we specialize to the special case that $\bm{\Sigma}$ is diagonal in the following derivation.
By Lemma \ref{lem:meanedge} with $\kappa=1$ (note that diagonal matrix is row-$\kappa$ sparse with $\kappa=1$), we have
$$
\mathbb{E} N_{E_{\delta}}^{(\bm{R})}=\mathbb{E} Z = \binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta},
$$
which behaves the same as $\frac{1}{\delta!}(2p^{1+\frac{1}{\delta}}P_n(r_\rho))^{\delta}$ when $p$ is large. By Lemma \ref{pderivative} \ref{item:pderivativeb}, we further have when $p$ is large
$$
\mathbb{E} N_{E_{\delta}}^{(\bm{R})} \approx \frac{1}{\delta!}\left(2p^{1+\frac{1}{\delta}}a_n (1-\rho)^{\frac{n-2}{2}}\right)^{\delta}.
$$
Thus in order to have a non-degenerate mean when $p\to \infty$, $2p^{1+\frac{1}{\delta}}a_n (1-\rho)^{\frac{n-2}{2}}$ needs to converges to some strictly positive and finite constant. Thus the rate we give in Theorem \ref{cor:Poissonlimit} is sharp. We hope this simple derivation addresses the concerns.
}
\item \mycomment{
As the authors point out, the compound Poisson characterization holds for the vertex counts in the empirical correlation graphs under the much weaker row-$\kappa$ sparsity
(Lemmas II.5 and Lemma III.12). So clearly the $(\tau,\kappa)$ assumption is an important one
to make for the other four random counts. Why is it that we need the $(\tau,\kappa)$ assumption for the counts besides the vertex counts of the empirical correlation graphs? This
should be clarified, so that we know why we need to consider $(\tau,\kappa)$ sparsity instead of
just the weaker row-$\kappa$ sparsity (which seems more realistic in the “large p” setting).
}
\myresponse{ We do not understand why you said ``So clearly the $(\tau,\kappa)$ assumption is an important one
to make for the other four random counts.'' There are totally $6$ random counts studied in our paper, and three random counts associated with the empirical correlation graph were treated in Lemma \ref{item:Poissonultrahigha} under the weaker row-$\kappa$ sparsity. The other $3$ counts associated with the empirical partial correlation graph are the ones that uses the $(\tau,\kappa)$ assumption.
So we are not sure why the Referee mentioned "the other four random counts" in the comment.
Currently we are not aware of any proof without $(\tau,\kappa)$ sparse assumption for the empirical partial correlation graph, although numerical experiments (see Figure \ref{fig:covasscom} and Figure \ref{fig:ar1} in our response to Referee 2's Comment \# 2)
suggest that the conclusion still holds beyond $(\tau,\kappa)$ sparse assumption.
}
\item \mycomment{
In modern high-dimensional problems, it is common not only for $p$ to be large but also
$n$. The authors only consider fixed n but allow $p$ to diverge to $\infty$. Can the results in
this paper be extended to the case where both $n$ and $p$ are growing? Do they still hold
if $n$ grows (much) slower than $p$? If this is not the case, then the authors could explain
what the technical challenges are with characterizing the limiting distributions when
we allow $n$ to also diverge with $p$.
}
\myresponse{This is an interesting question. The answer is yes, we think that the results still hold when $n$ grows (much) slower than $p$. Take Proposition \ref{prop:edgecor} for example. As mentioned in Remark \ref{rem:edgecor}, ``Respective expressions for $C'_{\delta,\gamma}$ and $C_{n,\delta,\rho}$ in Proposition \ref{prop:edgecor} are presented in equations \eqref{eqn:C'delgamdef} and \eqref{eqn:Cndelgamdef} in the Supplementary Material (\cite{hero2020unified})''. One could then just plug those formulae into the upper bound in Proposition \ref{prop:edgecor}, and one should be able to prove that the upper bound can go to $0$ with suitable condition on $\mu_{n,2\delta+2}\left(\bm{\Sigma}\right)$ (note that a sufficient condition to control $\mu_{n,2\delta+2}\left(\bm{\Sigma}\right)$ is given in Lemma \ref{lem:sufass}). So in principle $n$ can also approach infinity.
However, since the current manuscript is already quite long, we believe that the interesting case that $n$ and $p$ both approach infinity is best deferred to a future paper.
}
\item \mycomment{
I know that this is a theoretical paper, but I am curious if the authors could discuss
further some of the practical implications of their findings. The authors briefly mention
that their method can be used to approximate the family-wise error rate (FWER), but
this is presented rather lightly in the manuscript. The authors may want to highlight
this more, so it is not just seen as an interesting theoretical result, but also has some
important implications for practitioners who employ correlation or partial correlation
screening.
Is there anything else that the limiting compound Poisson distribution of these vertex
counts (or their Poisson approximations) tell us? For example, can we use these distributions for inference and hypothesis testing? Moreover, since FWER is known to be
very conservative, there seems to be more interest in modern settings for controlling
the false discovery rate (FDR) instead. Can the theory in this paper help to inform
efforts to control the false detection rate of vertices of degree exceeding $\delta$?}
\myresponse{
The reviewer makes very insightful remarks, and these insights are well placed: the results in our paper are not just of a technical nature, they also have tremendous practical implications, once again underscoring the substantive and non-incremental nature of the contributions in the paper. Our results can definitely be used to quantify the FWER, and as the reviewer points out, this multiple hypothesis testing error metric is sometimes viewed as overly conservative. This poses no problem for our results as we have expressions for k-FWER, the probability of obtaining k or more false discoveries. This multiple hypothesis testing error metric is naturally less conservative than simple FWER and addresses the reviewer’s concerns. In particular, our Poisson and compound Poisson expressions effectively describe the number of false discoveries and hence allow us to obtain results for k-FWER. Note that we can also obtain the marginal distributions of correlation estimation which in turn allow us to obtain expressions for p-values for correlation parameters. These marginal p-values allow us to establish FDR control too using either the Benjamini-Hochberg \cite{benjamini1995controlling} or Benjamini-Yakutelli procedures \cite{benjamini2001control}. In summary, the correlations screening framework is sufficiently rich that it allows us to undertake statistical error control in terms of FWER, k-FWER and FDR. This is one of the main strengths of our results: a rigorous inferential framework in the ultra-high dimensional setting. We have followed the reviewer’s advice and have highlighted the practical implications of the results in the paper as follows:
\begin{quote}
\ONE{The results in the paper also have significant relevance to applications. Recall that our Poisson and compound Poisson expressions effectively describe the number of false discoveries and hence allow us to obtain results for the familywise error rate (FWER) or k-FWER, that is the probability of obtaining k or more false discoveries. Note that we can also obtain the marginal distributions of correlation estimation which in turn allow us to obtain expressions for p-values for testing correlation estimates. These marginal p-values allow us to establish FDR control too using either the Benjamini-Hochberg \cite{benjamini1995controlling} or Benjamini-Yakutelli procedures \cite{benjamini2001control}. In summary, the correlations screening framework is sufficiently rich that it allows us to undertake statistical error control in terms of FWER, k-FWER and FDR. This is one of the main strengths of our results: a rigorous inferential framework in the ultra-high dimensional setting. }
\end{quote}
}
\end{enumerate}
\newpage
\section*{Responses to Associate Editor's comments (from email)}
It is a pleasure to inform you that your manuscript titled "A unified framework for correlation mining in ultra-high dimension" was accepted for publication in the IEEE Transactions on Information Theory in its present form. Additional comments of the reviewer(s), if any, are included at the foot of this letter. There may be additional comments from the reviewers as separate files available through your Author Center on the ScholarOne Manuscripts web site, so please check there also to make sure you have received all reviewer comments.
Specifically, there are two reviewers who still have some significant concerns and I think you should strongly consider revising your manuscript in light of their feedback. In submitting your final manuscript, please include a short response describing any changes you make in light of this feedback. However, I think that given the bulk of the feedback, I will not send this back for another full round of reviews.
\myresponse{We thank the AE for his decision to accept our paper. We have read the two reviews carefully and appreciate their close reading of our revision. We believe that the reviewer who provided a pdf is Reviewer 1 and and that the reviewer who provided a critical review in plaintext in the body of your email is Reviewer 3. In our final revision we have made additional revisions to address their comments, which are addressed below. To summarize:
\begin{enumerate}
\item Reviewer 3, who provided comments at the bottom of your email, makes two broad comments
\begin{itemize}
\item
Our use of overly strong words such as "important" in describing our work.
\vspace{0.1in}
We are sensitive to this comment and in the final version, where applicable, we have replaced or eliminated such wording to produce a more neutral presentation.
\item
Comments about novelty of the contribution and appropriateness of the IT Transactions for our work.
\vspace{0.1in}
We feel that we have addressed this as fully as possible in our first revision and in our response to reviews. Nonetheless, for the specific example given by the reviewer claiming inadequacy of comparisons to reference [6] ([38] in revised paper), in the revision we have provided specific discussion about the principal differences between the referenced work and the referenced paper.
{"\color{red} There is recent work on testing whether a given graph is a realization of an Erd\H{o}s–R\'enyi random graph or a realization of a random geometric graph with vertices i.i.d. uniformly distributed on the sphere [38] Therein, the authors study the asymptotics of a count statistic, the (signed) triangles, on a random geometric graph induced by thresholding the distances between $p$ i.i.d. uniformly distributed points on the $n$ dimensional sphere as $n$ goes to $\infty$. In contrast, this paper studies the large $p$ asymptotics of vertex degree counts for dependent non-uniform points on a sphere of fixed dimension $n$."}
\end{itemize}
\item Reviewer 1, who provided a pdf, has two comments: clarity of our statement of the objective; importance of the results. Both of these comments revolve around the stated condition on $\rho$, which specifies a rate of convergence of $\rho$ to $1$ as a function of $p$. It is interesting that Reviewer 1 did not bring up any specific concerns about the $\rho$-condition in the first round of review, while another reviewer did have questions about it (Reviewer 4) - to which we responded (see last page of this document).
Reviewer 1 makes the following statements: "A very particular $\rho$ is chosen\ldots," "What is special about the choice of $\rho$ that satisfies the following condition?", "This choice appears arbitrary." In fact, we do not simply {\em choose} the $\rho$ condition. This sharp condition is the result of our mathematical analysis as both {\it necessary and sufficient} to ensure that the number of counts $\tilde{N}_\delta$ have a non-trivial limit (not equal to $0$ or $\infty$).
\vspace{0.1in}
Given that this point was addressed in the last round we are surprised that Reviewer 1 has this misconception. In response to Reviewer 1's comments, we have made several additional changes to further clarify that the given $\rho$-condition is not arbitrary. This includes rewording the Theorems and Lemmas that invoke this condition in addition to further discussion about this after Lemma II.5. Specifically, we have added the following paragraph to {\em Remark} II.6 in the final revision:
"\color{red} As will be discussed in more detail in Remark IV.2
this rate is in fact both necessary and sufficient for the expected counts $\mathbb{E} \tilde{N}_\delta$ to converge to a non-trivial limit. If $\rho$ does not converge to $1$, or converges to $1$ at a slower rate, then $\mathbb{E}\tilde{N}_\delta$ diverges to $\infty$, while if $\rho$ converges to $1$ at a faster rate then $\mathbb{E} \tilde{N}_\delta$ converges to $0$."
{\color{blue} We have also added a remark that provides the detailed argument establishing sharpness of this rate.}
\noindent {\color{red} Remark IV.2:
Lemma \ref{lem:meanedge} implies that the condition on $\rho$ assumed in Theorem \ref{cor:Poissonlimit} is in fact necessary and sufficient for the mean to converge to a finite and strictly positive limit. For simplicity we specialize to the case that $\bm{\Sigma}$ is diagonal, which is row-$\kappa$ sparse with $\kappa=1$. By Lemma \ref{lem:meanedge} and (\ref{eqn:firstmomentofCP}),
$$
\mathbb{E} N_{E_{\delta}}^{(\bm{R})}=\mathbb{E} Z = \binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta},
$$
which via Stirling approximation behaves a
$\frac{1}{\delta!}(2p^{1+\frac{1}{\delta}}P_n(r_\rho))^{\delta}$ for large $p$.
By Lemma \ref{pderivative} \ref{item:pderivativeb}
we further have, again for large $p$,
$$
\mathbb{E} N_{E_{\delta}}^{(\bm{R})} \approx \frac{1}{\delta!}\left(2p^{1+\frac{1}{\delta}}a_n (1-\rho)^{\frac{n-2}{2}}\right)^{\delta}.
$$
Thus, in order that the mean count have a non-degenerate limit when $p\to \infty$, $2p^{1+\frac{1}{\delta}}a_n (1-\rho)^{\frac{n-2}{2}}$ must converge to some strictly positive and finite constant value. If $\rho$ does not converge to $1$, or if it converges to $1$ at a slower rate, then $\mathbb{E}\tilde{N}_\delta$ diverges to $\infty$, while if $\rho$ converges to $1$ at a faster rate then $\mathbb{E} \tilde{N}_\delta$ converges to $0$. This is reflected in the phase transition phenomenon discussed in \cite{hero2011large,hero2012hub}. Thus the rate on $\rho$ in Theorem \ref{cor:Poissonlimit}
is sharp.}
\end{enumerate}
}
\newpage
What follows is a replicate of our previous response to Reviewer 4's comment 1 (p. 19 of our round 1 response to reviewers) in which we specifically address the sharpness of our condition on $\rho$.
\begin{enumerate}
\item \mycomment{
Theorem II.4, the authors require the correlation threshold $\rho$ to grow at a very
specific function of $p$, i.e. $a_n2^{n/2}p^{1+\frac{1}{\delta}} (1-\rho)^{\frac{n-2}{2}}\to e(n, \delta)$ as $p\to \infty$. This seems to
be an artifact of the proof, and it is a bit troubling (even though the authors pointed
out later that the growth rate is not fast). Is this a sharp rate, i.e. can this condition
not be relaxed at all? Does the distributional limit result automatically fail when this
condition on the growth rate of $\rho$ does not hold? This deserves greater attention, and
the authors may want to prove that this rate is sharp when $p\to\infty$. Otherwise, it
raises a lot of questions and concerns.
}
\myresponse{This rate is not an artifact of the proof, it is a sharp rate. From Section III.B
we know that our model is a Random Psuedo Geometric Graph with radius parameter $r_\rho = \sqrt{2(1-\rho)}$. Thus the rate is equivalent to $2a_np^{1+\frac{1}{\delta}} r_\rho^{{n-2}}\to e(n, \delta)$, which is consistent with existing Poisson approximations for random geometric graphs \cite[ Theorem 3.4]{penrose2003random}. We have added such discussions in Section III.B:
\begin{quote}
\ONE{ As a specific example, the rate $a_n2^{n/2}p^{1+\frac{1}{\delta}} (1-\rho)^{\frac{n-2}{2}}\to e(n, \delta)$ as $p\to \infty$ in Theorem II.4
is equivalent to $2a_np^{1+\frac{1}{\delta}} r_\rho^{{n-2}}\to e(n, \delta)$, which is consistent with existing Poisson approximation for random geometric graph \cite[ Theorem 3.4]{penrose2003random}. }
\end{quote}
In Remark II.6
we also added:
\begin{quote}
\ONE{This particular rate on $\rho$ is consistent with the rate
of the existing Poisson approximation results in
random geometric graphs \cite{penrose2003random} as will be
discussed in Section III.B
}
\end{quote}
The rate is sharp in the sense that letting $\rho$ grow at a faster or slower rate the result will be a degenerate compound Poisson limit (a limit either having mean 0 or mean $\infty$). This was briefly discussed in Lemma III.16
in the submitted paper.
For the benefit of the Referee, we provide some additional details on the sharpness of this rate that we choose not to include in the revision.
Since we want to establish a result that holds for a certain type sparse dispersion parameter $\bm{\Sigma}$, we specialize to the special case that $\bm{\Sigma}$ is diagonal in the following derivation.
By Lemma IV.1
with $\kappa=1$ (note that diagonal matrix is row-$\kappa$ sparse with $\kappa=1$), we have
$$
\mathbb{E} N_{E_{\delta}}^{(\bm{R})}=\mathbb{E} Z = \binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta},
$$
which behaves the same as $\frac{1}{\delta!}(2p^{1+\frac{1}{\delta}}P_n(r_\rho))^{\delta}$ when $p$ is large. By Lemma A.22 B
we further have when $p$ is large
$$
\mathbb{E} N_{E_{\delta}}^{(\bm{R})} \approx \frac{1}{\delta!}\left(2p^{1+\frac{1}{\delta}}a_n (1-\rho)^{\frac{n-2}{2}}\right)^{\delta}.
$$
Thus in order to have a non-degenerate mean when $p\to \infty$, $2p^{1+\frac{1}{\delta}}a_n (1-\rho)^{\frac{n-2}{2}}$ needs to converges to some strictly positive and finite constant. Thus the rate we give in Theorem II.4
is sharp. We hope this simple derivation addresses the concerns.
}
\end{enumerate}
\newpage
\section{Conclusions and discussions}
\label{sec:conclusions}
In this paper, we studied the number of highly connected vertices in both the empirical correlation graph and the empirical partial correlation graph by adopting unified framework. More specifically, we showed that the distributions of the number of hubs $N_{V_\delta}^{(\bm{\Psi})}$ or $N_{\breve{V}_\delta}^{(\bm{\Psi})}$ and the star subgraph counts $N_{E_\delta}^{(\bm{\Psi})}$ are close to common compound Poisson, when $p$ is finite and when $p$ approaches infinity. We also establish that their first and second moments are close to that of the compound Poisson distribution. The parameters in the compound Poisson distributions are characterized in terms of random geometric graphs and random pseudo geometric graphs. The parameters are also approximated by simple formulae, which implies that the approximating compound Poisson distributions can be further approximated by Poisson distributions for reasonably large sample size $n$ or a reasonably hub degree $\delta$.
There are multiple avenues for future research. Numerical experiments suggest that the results in this paper hold beyond the $(\tau,\kappa)$ sparsity condition. A future line of work is to characterize the compound Poisson approximations for weaker sparsity conditions.
Generalization of the moment convergence results to beyond the second moments are also of interest.
\section{Discussions}
\subsection{Comparisons}
\end{comment}
\section{Convergence of moments}
\label{sec:convergenceofmoments}
Moment expressions are useful for a number of reasons, including characterizing the behavior of phase transition thresholds and the expected number of false discoveries \cite{hero2011large}. Theorems \ref{cor:Poissonlimit} and \ref{thm:Poissonultrahigh} only prescribe the distribution of $\bar{N}_\delta$ but not the moments. In this subsection, we present approximations to the first moment and second moment of $\bar{N}_\delta$ for finite $p$.
\begin{comment}
\YW{
Despite that we have shown in Theorem \ref{cor:Poissonlimit} that $\bar{N}_\delta \to \operatorname{CP}(\lambda_{n,\delta}(e_{n,\delta}),\bm{\zeta}_{n,\delta})$ in distribution and established in Theorem \ref{thm:Poissonultrahigh} an upper bound on $d_{\operatorname{TV}}(\bar{N}_\delta, \operatorname{CP}(\lambda_{n,\delta}(e_{n,\delta}),\bm{\zeta}_{n,\delta}))$, the convergence of the corresponding moments remains unknown.
Moreover, convergence or non-asymptotic approximation of moments is important. For instance, in \cite{hero2011large} an approximation formula of the first moment is used to derive a phase transition threshold by $d\mathbb{E}[\bar{N}_\delta]/d\rho=-1$. In this subsection, we present the non-asymptotic approximation of the first moment and second moment of $\bar{N}_\delta$.}
\end{comment}
Let $Z\sim \operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho})$. Then we can represent $ Z= \sum_{i=1}^N Z_i $, where $N$ is distributed as a Poisson with mean $\lambda_{p,n,\delta,\rho}$, $Z_i \overset{\text{i.i.d.}}{\sim} \bm{\zeta}_{n,\delta,\rho} $ and $N$ is independent of each $Z_i$. The first two moments of $Z$ are:
\begin{align*}
\mathbb{E} Z = & \mathbb{E} N \mathbb{E} Z_1 = \binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta}, \numberthis \label{eqn:firstmomentofCP}\\
\mathbb{E} Z^2 = & \mathbb{E} N \mathbb{E} Z_1^2 + (\mathbb{E} N \mathbb{E} Z_1)^2 \\
= &\binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta} \sum_{\ell =1}^{\delta+1} \ell \alpha(\ell,r_\rho) + \left( \binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta} \right)^2. \numberthis \label{eqn:secondmomentofCP}
\end{align*}
The next lemma provides an upper bound for the difference between
the first moment of $N_{E_{\delta}}^{(\bm{R})}$ and the first moment of the compound Poisson specified in \eqref{eqn:firstmomentofCP}.
\begin{lem} \label{lem:meanedge}
Let $p\geq n \geq 4$, $\delta\in [p-1]$ and $\gamma>0$ be given. Suppose $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$. Suppose $2p^{1+\frac{1}{\delta}}P_n(r_\rho)\leq \gamma$, and $\bm{\Sigma}$ is row-$\kappa$ sparse. Then
$$
\left| \mathbb{E} N_{E_{\delta}}^{(\bm{R})}-\mathbb{E} Z \right|\leq \frac{(\delta+1)}{2((\delta-1)!)}\gamma^{\delta} \mu_{n,\delta+1}\left(\bm{\Sigma}\right) \frac{\kappa-1}{p}.
$$
\end{lem}
\begin{rem}
\label{rem:rhorate}
{
Lemma \ref{lem:meanedge} implies that the condition on $\rho$ assumed in Theorem \ref{cor:Poissonlimit} is in fact necessary and sufficient for the mean to converge to a finite and strictly positive limit. For simplicity we specialize to the case that $\bm{\Sigma}$ is diagonal, which is row-$\kappa$ sparse with $\kappa=1$. By Lemma \ref{lem:meanedge} and (\ref{eqn:firstmomentofCP}),
$$
\mathbb{E} N_{E_{\delta}}^{(\bm{R})}=\mathbb{E} Z = \binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta},
$$
which via Stirling approximation behaves a
$\frac{1}{\delta!}(2p^{1+\frac{1}{\delta}}P_n(r_\rho))^{\delta}$ for large $p$.
By Lemma \ref{pderivative} \ref{item:pderivativeb}
we further have, again for large $p$,
$$
\mathbb{E} N_{E_{\delta}}^{(\bm{R})} \approx \frac{1}{\delta!}\left(2p^{1+\frac{1}{\delta}}a_n (1-\rho)^{\frac{n-2}{2}}\right)^{\delta}.
$$
Thus, in order that the mean count have a non-degenerate limit when $p\to \infty$, $2p^{1+\frac{1}{\delta}}a_n (1-\rho)^{\frac{n-2}{2}}$ must converge to some strictly positive and finite constant value. If $\rho$ does not converge to $1$, or if it converges to $1$ at a slower rate, then $\mathbb{E}\tilde{N}_\delta$ diverges to $\infty$, while if $\rho$ converges to $1$ at a faster rate then $\mathbb{E} \tilde{N}_\delta$ converges to $0$. This is reflected in the phase transition phenomenon discussed in \cite{hero2011large,hero2012hub}. Thus the rate on $\rho$ in Theorem \ref{cor:Poissonlimit}
is sharp.}
\end{rem}
By combining the preceding lemma and the portmanteau result of Proposition \ref{thm:allclose} one immediately obtains approximations for the first moment of all $6$ quantities $\{N_{E_\delta}^{(\bm{\Psi})},N_{\breve{V}_\delta}^{(\bm{\Psi})},N_{V_\delta}^{(\bm{\Psi})}: \bm{\Psi} \in \{\bm{R},\bm{P}\} \}$ for finite $p$.
We also characterize the second moment approximation of $N_{E_\delta}^{(\bm{R})}$ in the next lemma
under the same conditions as Lemma \ref{lem:meanedge}.
\begin{prop}[Second moment bounds for subgraph counts]
\label{lem:2ndmoment}
Let $p\geq n \geq 4$, $\delta\in [p-1]$ and $\gamma>0$ be given. Suppose $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$. Suppose $2p^{1+\frac{1}{\delta}}P_n(r_\rho)\leq \gamma$, and $\bm{\Sigma}$ is row-$\kappa$ sparse. Then
$$
\left| \mathbb{E} \left(N_{E_{\delta}}^{(\bm{R})}\right)^2-\mathbb{E} Z^2 \right|\leq C_{n,\delta,\gamma} \left(\mu_{n,2\delta+2}(\bm{\Sigma})\frac{\kappa}{p} +p^{-1/\delta} \right).
$$
\end{prop}
We extend the preceding proposition to
other quantities in $\{N_{E_\delta}^{(\bm{\Psi})},N_{\breve{V}_\delta}^{(\bm{\Psi})},N_{V_\delta}^{(\bm{\Psi})}: \bm{\Psi} \in \{\bm{R},\bm{P}\} \}$ by generalizing Proposition \ref{thm:allclose}
to $L^1$ distance between the square of the quantities.
\begin{prop}[Portmanteau result for $L^2$ distance]
\label{prop:allcloseL2}
{
Let $p\geq n \geq 4$ and $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$. Let $\delta \in [p-1]$. Suppose $2p^{1+\frac{1}{\delta}}P_n(r_\rho)\leq \gamma$.
\begin{enumerate}[label=(\alph*)]
\item
\label{item:allcloseL2a}
Suppose $\bm{\Sigma}$ is row-$\kappa$ sparse. Then for $\tilde{N}_\delta \in \left\{ N_{\breve{V}_\delta}^{(\bm{R})}, N_{V_\delta}^{(\bm{R})} \right\} $,
$$
\mathbb{E} \left| \left(\tilde{N}_\delta\right)^2 - \left(N_{\EE_\delta}^{(\bm{R})}\right)^2 \right|
\leq C_{n,\delta,\gamma} \left(1+\mu_{n,2\delta+3}(\bm{\Sigma})\frac{\kappa-1}{p}\right)p^{-1/\delta}. $$
\item \label{item:allcloseL2b}
Suppose $\bm{\Sigma}$, after some row-column permutation, is $(\tau,\kappa)$ sparse with $\tau\leq \frac{p}{2}$. Moreover, $\left(\sqrt{\frac{n-1}{p}}+\sqrt{\frac{\delta\ln p}{p}}\right) \leq c$ hold for some positive and \ONE{sufficiently} small constant $c$. Then
\begin{align*}
\mathbb{E}\left|\left(N_{\EE_\delta}^{(\bm{P})}\right)^2-\left(N_{\EE_\delta}^{(\bm{R})}\right)^2\right|
\leq C_{n,\delta,\gamma} \left(1+\mu_{n,2\delta+2}(\bm{\Sigma})\frac{\kappa-1}{p}\right) \left(\frac{\sqrt{\ln p}}{\sqrt{p}} + \frac{\tau}{p}\right).
\end{align*}
\item
\label{item:allcloseL2c}
Suppose the same conditions as in part \ref{item:allclosec} hold. Then for $\tilde{N}_\delta \in \left\{ N_{\breve{V}_\delta}^{(\bm{P})}, N_{V_\delta}^{(\bm{P})} \right\} $
\begin{align*}
\mathbb{E}\left|\left(\tilde{N}_{\delta}\right)^2-\left(N_{E_\delta}^{(\bm{P})}\right)^2\right|
\leq C_{n,\delta,\gamma} \left(1+\mu_{n,2\delta+3}(\bm{\Sigma})\frac{\kappa-1}{p}\right)p^{-1/\delta}.
\end{align*}
\end{enumerate}
}
\end{prop}
By applying triangle inequalities to Proposition \ref{prop:allcloseL2} \ref{item:allcloseL2b} and \ref{item:allcloseL2c}, one obtain for $\tilde{N}_{\delta}\in \{N_{V_\delta}^{(\bm{P})}, N_{\breve{V}_\delta}^{(\bm{P})} \}$
$$
\mathbb{E}\left|\left(\tilde{N}_{\delta}\right)^2-\left(N_{E_\delta}^{(\bm{R})}\right)^2\right|\leq C_{n,\delta,\gamma} \left(1+\mu_{n,2\delta+3}(\bm{\Sigma})\frac{\kappa-1}{p}\right)
\left(\frac{\sqrt{\ln p}}{\sqrt{p}} + \frac{\tau}{p}+p^{-1/\delta}\right).
$$
Thus we have established the $L^1$ distance between the square of each term in $\{N_{E_\delta}^{(\bm{\Psi})},N_{\breve{V}_\delta}^{(\bm{\Psi})},N_{V_\delta}^{(\bm{\Psi})}: \bm{\Psi} \in \{\bm{R},\bm{P}\} \}$ and $\left(N_{E_\delta}^{(\bm{R})}\right)^2$. By combining Proposition \ref{lem:2ndmoment} and Proposition \ref{prop:allcloseL2} one immediately obtains approximations to the second moment of $\bar{N}_\delta$ for finite $p$.
While this section focuses on the approximation to the first and second moments of $\bar{N}_\delta$ for finite $p$, their limits can also be obtained when $p\to\infty$ and $\rho\to 1$ at the rate specified in Theorem \ref{cor:Poissonlimit}.
\begin{comment}
\subsection{Beyond sparsity}
In both Theorem \ref{cor:Poissonlimit} and Theorem \ref{thm:Poissonultrahigh}, the row-$\kappa$ sparsity on $\bm{\Sigma}$ is imposed to study the random quantities in empirical correlation graph, while the stronger condition $(\tau,\kappa)$ sparsity is imposed to explore the random quantities in the empirical partial correlation graph. While these sparsity conditions are sufficient, it is unclear whether they can be relaxed. For instance, when studying the random quantities in the empirical partial correlation graph, it is more common to impose sparsity on the inverse of the covariance matrix $\bm{\Sigma}^{-1}$ instead of $\bm{\Sigma}$.
\begin{figure}[ht]
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=.8\linewidth]{cpvd_xigap07trial2000p1500.png}
\caption{small gap}
\end{subfigure}
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=.8\linewidth]{cpvd_xigap09trial2000p1500.png}
\caption{large gap}
\end{subfigure}
\caption{ The vertical axis of (a) is $d_{\operatorname{TV}}(N_{V_2}^{(\bm{\Psi})},\text{CP}(\lambda_{35,2}(1),\operatorname{Dirac}(1)))$ as in Theorem \ref{cor:Poissonlimit} and that of (b) is $d_{\operatorname{TV}}(N_{V_2}^{(\bm{\Psi})},\text{CP}(\lambda_{p,35,2,\rho},\operatorname{Dirac}(1)))$ as in Theorem \ref{thm:Poissonultrahigh}. Note the increment distributions $\bm{\zeta}_{35,2}$ in (a) and $\bm{\zeta}_{35,2,\rho}$ in (b) are both replaced by $\operatorname{Dirac}(1)$ since $n=35$ is sufficiently large for $\delta=2$ as indicated by Figure \ref{fig:incrementtodeltauppbou} (b). For both plots the samples are generated according to $\mathcal{N}(\bm{0},\bm{\Sigma})$ with $\bm{\Sigma}$ being a $(\tau=p^{0.6},\kappa=p^{0.8})$ sparse matrix for each $p$. The parameters are $n=35$, $\delta=2$ and the threshold $\rho$ is chosen according to \eqref{eqn:rhopformula} with $e_{n,\delta}=1$.
The blue curve is for the empirical correlation graph (\bm{\Psi} =$\bm{R}$) and the red curve is for the empirical partial correlation graph (\bm{\Psi} =$\bm{P}$). As demonstrated by the plots, for both empirical correlation and partial correlation graphs, the total variations in (a) decrease very slowly while the total variations in (b) converge to $0$ very fast.}
\label{fig:Poissonfinitevslimit}
\end{figure}
\begin{figure}
\center
\includegraphics[scale=0.3]{cpvd_xigap07trial2000p1500.png}
\caption{The vertical axis is the total variation distance between $N_{V_\delta}^{(\bm{\Psi})}$ to the predicted compound Poisson when $\delta=2$ and $n=4$. These are for the theoretical correlation matrix with diagonal $1$ and off diagonal $\xi=\rho-\text{gap}$. \textbf{Implication}: As we can see, when the $\text{gap}=0.7$, the partial correlation $N_{V_\delta}^{(\bm{P})}$ converges in distribution to the predicted compound Poisson while the $N_{V_\delta}^{(\bm{R})}$ doesn't. Note in this case the partial correlation matrix is of the same type but with smaller off-diagonal entries, which might account for that $N_{V_\delta}^{(\bm{P})}$ converges to the predicted compound Poisson but $N_{V_\delta}^{(\bm{R})}$ doesn't.}
\label{fig:covasscom}
\end{figure}
\begin{figure}
\center
\includegraphics[scale=0.3]{cpvd_xigap09trial2000p1500.png}
\caption{Similar to the previous one. \textbf{Implication}: For the current one, we have larger gap, which results small entry for both off-diagonal entries for both correlation and partial correlation matrix. In this case both $N_{V_\delta}^{(\bm{P})}$ and $N_{V_\delta}^{(\bm{R})}$ converge to the predicted compound Poisson.}
\label{fig:covasscom}
\end{figure}
\end{comment}
\begin{comment}
\begin{figure}
\center
\includegraphics[scale=0.3]{cpvd_arxi09trial6000p1500.png}
\caption{The theoretical correlation for this is the AR(1) model with $xi=0.9$. \textbf{Implication}: For the current one, the partial correlation is tridiagonal, which is sparser than correlation matrix. As a result, we see that
$N_{V_\delta}^{(\bm{P})}$ converges faster in distribution to the compound Poisson predicted. The correlation matrix, though not exactly sparser, but we know that the entries that are far away from the diagonal is of small magnitude, which might account for the convergence of $N_{V_\delta}^{(\bm{R})}$.}
\label{fig:covasscom}
\end{figure}
\end{comment}
\begin{comment}
\begin{figure}
\center
\includegraphics[scale=0.3]{tau6kappa8N4delta2.png}
\caption{$(\tau,\kappa)$ sparse covariance matrix with $\tau=p^{0.6}$ and $\kappa=p^{0.8}$. The yellow curve is the total variation distance between $N_{V_2}^{(\bm{R})}$ and the Poisson; while purple curve is the total variation distance between $N_{V_2}^{(\bm{P})}$ and the Poisson. We see that $n=4$ and $\delta=2$ the convergence is to compound Poisson and not to the Poisson.\YW{do we need to plot the total variation to Poisson as well?}}
\label{fig:covasscom}
\end{figure}
\begin{figure}
\center
\includegraphics[scale=0.3]{n4delta1difftrial.png}
\caption{Diagonal covariance matrix. The blue curve is the total variation distance between $N_{V_2}^{(\bm{R})}$ and the predicted compound Poisson; while the orange curve is the total variation distance between $N_{V_2}^{(\bm{P})}$ and the predicted compound Poisson. We see for $n=4$ and $\delta=1$ these total variation distances decreases as the number of the trails increases, due to the random effects from the samples.}
\label{fig:covasscom}
\end{figure}
\end{comment}
\section{Framework}
\section{Main results}
\label{sec:mainresult}
\subsection{Framework}
\label{Hubfra}
\noindent
Available is a data matrix consisting of multivariate samples
\begin{equation}
X= [x^{(1)},x^{(2)},\cdots, x^{(n)}]^\top= [x_1,x_2,\cdots, x_p]
\in \mathbb{R}^{n\times p}, \label{Xtildedef}
\end{equation}
where $\{x^{(i)}\}_{i=1}^{n} \subset \mathbb{R}^p$ are samples from a $p$-dimensional distribution. We assume that the $n\times p$ data matrix $X$ follows a vector elliptically contoured distribution \cite{dawid1977spherical, anderson1990theory,anderson1992nonnormal}.
A random matrix $X\in \mathbb{R}^{n\times p}$ is vector-elliptical
with positive definite covariance or dispersion parameter $\bm{\Sigma}\in \mathbb{R}^{p\times p}$ and location parameter $\bm{\mu}$ if its density satisfies
\begin{equation}
f_{X}(X) = \text{det}(\bm{\Sigma})^{-n/2}g(\text{tr}((X-\bm{1}\bm{\mu}^\top)\bm{\Sigma}^{-1}(X^\top-\bm{1}\bm{\mu}^\top))), \label{eqn:ellden}
\end{equation}
for a shape function $g: \mathbb{R}\to [0,\infty)$ such that $\int f_{X}(X) =1$. In \eqref{eqn:ellden}, $\bm{1}$ is a column vector with all elements equal to $1$.
We use the shorthand $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$ to denote that $X$ follows a vector elliptically contoured distribution with density \eqref{eqn:ellden}. Note that the rows $\{x^{(i)}\}_{i=1}^n $ of $X$ are uncorrelated but not necessarily independent \cite{anderson1992nonnormal}. An example of a vector-elliptical distributed is the matrix normal distribution, for which the rows $\{x^{(i)}\}_{i=1}^{n}\subset \mathbb{R}^p$ are i.i.d. samples from $\mathcal{N}(\bm{\mu},\bm{\Sigma})$. Specifically, the matrix normal density is obtained when, in \eqref{eqn:ellden}, $g(w)=g_0(w)=(2\pi)^{-\frac{n p}{2}}\exp(-\frac{1}{2}w)$ and in this case $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g_0)$.
Given a data matrix $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$\footnote{\ONE{In previous work \cite{hero2011large,hero2012hub} it was assumed that the samples $x^{(i)}$ are i.i.d. elliptical contoured distributed. This condition is in fact insufficient and the stronger vector elliptical contoured distribution condition (\ref{eqn:ellden})
is required.}
}, the sample mean $\bar{x}$ is given as a row vector
$$
\bar{x}=\frac{1}{n}\displaystyle\sum_{i=1}^{n}x^{(i)} = \frac{1}{n}X^\top\bm{1}
$$
The sample covariance matrix $\bm{S}$ is
\begin{equation}
\bm{S}=\frac{1}{n-1}\sum_{i=1}^{n} (x^{(i)}-\bar{x})(x^{(i)}-\bar{x})^\to
=\frac{1}{n-1}X^\top\left(\bm{I}_n -\frac{1}{n}\bm{1}\bm{1}^\top\right) X.\label{def:samcov}
\end{equation}
The sample correlation matrix $\bm{R}$ is defined as:
\begin{equation}
\bm{R}=\operatorname{{\bf diag } }(\bm{S})^{-\frac{1}{2}}\bm{S} \operatorname{{\bf diag } }(\bm{S})^{-\frac{1}{2}}, \label{defR}
\end{equation}
where $\operatorname{{\bf diag } }(\bm{A})$ for a matrix $\bm{A}\in \mathbb{R}^{n\times n}$ is the diagonal part of $\bm{A}$ and $\bm{B}^{-1/2}$ for a diagonal matrix $\bm{B}$ is a diagonal matrix formed by raising every diagonal element of $\bm{B}$ to the power $-1/2$. Since $\bm{R}$ is not invertible, we define the sample partial correlation matrix $\bm{P}$ by
\begin{equation}
\label{Pdef}
\bm{P} = \operatorname{{\bf diag } }(\bm{R}^\dagger)^{-\frac{1}{2}}\bm{R}^\dagger \operatorname{{\bf diag } }(\bm{R}^\dagger)^{-\frac{1}{2}}.
\end{equation}
where $\bm{R}^\dagger$ is the Moore-Penrose pseudo-inverse of $\bm{R}$.
Let $\bm{\Psi}=(\Psi_{ij})_{i,j\in [p]}$ be generic notation for a correlation-type matrix like $\bm{R}$ or $\bm{P}$. Given a threshold $\rho\in [0, 1)$ define the undirected graph induced by thresholding $\bm{\Psi}$, denoted by $\mathcal{G}_\rho(\bm{\Psi})$, as follows. The vertex set of graph $\mathcal{G}_\rho(\bm{\Psi})$ is $\mathrm{V}^{(\bm{\Psi})} = [p]:=\{1,2,\cdots,p\}$ and the edge set is $\mathrm{E}^{(\bm{\Psi})} \subset \mathrm{V}^{(\bm{\Psi})}\times\mathrm{V}^{(\bm{\Psi})}$, with $(i,j) \in \mathrm{E}^{(\bm{\Psi})}$ if $|\Psi_{ij} | \geq \rho$, where $(i,j)$ denotes an edge between $i$ and $j$ $(i\not =j)$. We call $\mathcal{G}_\rho(\bm{\Psi})$ the empirical correlation graph and the empirical partial correlation graph, respectively, when $\bm{\Psi} = \bm{R}$ and $\bm{\Psi} = \bm{P}$. Let $\bm{\Phi}^{(\bm{\Psi})}(\rho)$ be the adjacency matrices associated with the graph $\mathcal{G}_\rho(\bm{\Psi})$, with elements
$\Phi_{ij}^{(\bm{\Psi})}(\rho):={1}(|\Psi_{ij} | \geq \rho)$ for $i\not = j$, where ${1}(\cdot)$ is the indicator function.
The dependence of $\bm{\Phi}^{(\bm{\Psi})}(\rho)$ and $\Phi_{ij}^{(\bm{\Psi})}(\rho)$ on $\rho$ will be suppressed when it is clear from the context.
The focus of this paper is correlation and partial screening, which counts the number of
vertices of prescribed degree, the number of star subgraphs, or the number of edges in $\mathcal{G}_\rho(\bm{\Psi})$.
The objective is to characterize the
distributions of these counting statistics.
More specifically, for the graph $\mathcal{G}_\rho(\bm{\Psi})$ with $\bm{\Psi} =\bm{R}$ or $\bm{\Psi} =\bm{P}$, the degree of vertex $i$ is defined as $\sum_{j=1,j\not =i}^p \Phi_{ij}^{(\bm{\Psi})}(\rho)$.
For $1\leq \delta \leq p-1$, the total number of vertices with degree exactly $\delta$ (at least $\delta$), denoted by $N_{\breve{V}_\delta}^{(\bm{\Psi})}$ ($N_{V_\delta}^{(\bm{\Psi})}$), is of particular interest. Note that if a vertex has degree exactly $\delta$, then there exists a star subgraph with $\delta$ edges centered at that vertex. Consequently the number of star subgraphs are important in the analysis of $N_{\breve{V}_\delta}^{(\bm{\Psi})}$ and $N_{V_\delta}^{(\bm{\Psi})}$, hence our interest in the number of star subgraphs. For $2\leq \delta\leq p-1$, the number of subgraphs in $\mathcal{G}_\rho(\bm{\Psi})$ that are isomorphic to $\Gamma_\delta$ is denoted by
$N_{\EE_\delta}^{(\bm{\Psi})}$, where $\Gamma_\delta$ denotes a star shaped graph with $\delta$ edges. In the case when $\delta = 1$, we define $N_{E_1}^{(\bm{\Psi})}$ to be twice the number of edges in $\mathcal{G}_\rho(\bm{\Psi})$. $N_{\EE_\delta}^{(\bm{\Psi})}$ is referred as the \emph{star subgraph counts}.
\begin{exa}
\begin{figure}[ht!]
\labellist
\small \hair 2pt
\pinlabel {$2$} at 52 417
\pinlabel $1$ at 525 51
\pinlabel $5$ at 968 239
\pinlabel $3$ at 579 548
\pinlabel $4$ at 1049 826
\pinlabel \rotatebox{-35}{$e_1$} at 241 219
\pinlabel \rotatebox{20}{$e_2$} at 780 121
\pinlabel \rotatebox{80}{$e_3$} at 507 334
\pinlabel \rotatebox{-30}{$e_4$} at 810 437
\pinlabel \rotatebox{25}{$e_5$} at 801 725
\endlabellist
\centering
\includegraphics[width=0.3\linewidth]{graphs/example_of_definition}
\caption{A graph with $5$ vertices and $5$ edges.}
\label{fig:exaofdef}
\end{figure}
Figure \ref{fig:exaofdef} represents an empirical partial correlation graph. For this graph the number of vertices of degree $2$ is $N_{\breve{V}_2}^{(\bm{P})}=1$ and the number of vertices of degree at least $2$ is $N_{V_2}^{(\bm{P})}=3$. The number of subgraphs isomorphic to $\Gamma_3$ is $N_{E_3}^{(\bm{P})} = 2 $. The number of connected vertices is $N_{V_1}^{(\bm{P})}=5$, and $N_{E_1}^{(\bm{P})} = 10 $ as there are $5$ edges.
\end{exa}
Consider now the case where the sample size $n$ is fixed and there exists a sequence of data matrices $X \in \mathbb{R} ^{n\times p}$ with increasing dimension $p$. Following the procedure described in the paragraph after \eqref{Pdef}, we obtain a sequence of random graphs $G_{\rho}(\bm{\Psi})$ with increasing number of vertices. This paper derives finite sample compound Poisson characterizations of the distributions of the $6$ random quantities $\{N_{E_\delta}^{(\bm{\Psi})},N_{\breve{V}_\delta}^{(\bm{\Psi})},N_{V_\delta}^{(\bm{\Psi})}: \bm{\Psi} \in \{\bm{R},\bm{P}\} \}$, for $p$ finite and as $p\to\infty$, for suitably chosen $\rho$, under a sparsity assumption on the dispersion parameter $\bm{\Sigma}$. \ONE{Such characterizations can be used to test the sparsity structure of the dispersion parameter $\bm{\Sigma}$ or to guide the choice of the threshold $\rho$ \cite{hero2011large, hero2012hub}.}
Throughout the rest of the paper we use $\bar{N}_\delta$ to denote a generic random variable equal to one of the $6$ quantities $\{N_{E_\delta}^{(\bm{\Psi})},N_{\breve{V}_\delta}^{(\bm{\Psi})},N_{V_\delta}^{(\bm{\Psi})}: \bm{\Psi} \in \{\bm{R},\bm{P}\} \}$. By abuse of terminology, we refer to $\bar{N}_\delta$ generically as vertex counts. We reiterate that in this paper the number $n$ of samples is fixed and the number $p$ of variables could either be finite or tend to infinity.
\subsection{A unified theorem}
\label{sec:unifiedtheorem}
In this subsection we present a unified theorem that establishes that $\bar{N}_\delta$ converges in distribution to a compound Poisson distribution when $p\to \infty$. We begin by defining necessary quantities and then we state our main theorem.
For any positive number $\lambda$ and a probability distribution $\bm{\zeta}$ supported on positive integers, let $\operatorname{CP}(\lambda,\bm{\zeta})$ denote the corresponding compound Poisson distribution. Specifically, $\operatorname{CP}(\lambda,\bm{\zeta})$ is the distribution of $ Z= \sum_{i=1}^N Z_i $, where $N$ is distributed as a Poisson random variable with mean $\lambda$, $Z_i \overset{\text{i.i.d.}}{\sim} \bm{\zeta} $ and $N$ is independent of each $Z_i$. Here the random variable $N$ is the number of occurrences of increments and $\bm{\zeta}$ is the distribution of each increment. The parameter $\lambda$ and $\bm{\zeta}$ are often referred to as the arrival rate and the increment distribution, respectively.
As the parameters of the compound Poisson distribution in the next theorem involve a random geometric graph, we define relevant notation. Given a set of points $\{ \bm{v}_i \}_{i=1}^\delta$ in $\mathbb{R}^{n-2}$, denote by
$\textbf{Ge}\left(\{\bm{v}_i \}_{i=1}^\delta, r \right)$
the geometric graph with radius $r$, defined as follows. The vertex set of the graph is $\{ \bm{v}_i \}_{i=1}^\delta$, and there is an edge between $\bm{v}_i$ and $\bm{v}_j$ if $\|\bm{v}_i-\bm{v}_j\|_2 \leq r$. The graph is called a random geometric graph when the vertices of the geometric graph are random. A \emph{universal vertex} is a vertex of an undirected graph that is adjacent to all other vertices of the graph \cite{larrion2004clique}. Denote by $\operatorname{{\bf NUV}}\left(\{\bm{v}_i \}_{i=1}^\delta, r \right)$
the number of universal vertices in $\textbf{Ge}\left(\{\bm{v}_i \}_{i=1}^\delta, r \right)$. Denote $B^{n-2}$ the unit sphere in $\mathbb{R}^{n-2}$ and denote $\operatorname{{\bf unif}}(B^{n-2})$ the uniform distribution on $B^{n-2}$. Let $\{\tilde{\bm{u}}_i\}_{i=1}^\delta$ be i.i.d. from $\operatorname{{\bf unif}}(B^{n-2})$. For the random geometric graph $\textbf{Ge}\left(\{\tilde{\bm{u}}_i\}_{i=1}^\delta, 1 \right)$, we denote the probability that there are exactly $\ell-1$ universal vertices by
\begin{equation}
\alpha_\ell := P\left( \operatorname{{\bf NUV}}\left( \{\tilde{\bm{u}}_i \}_{i=1}^\delta, 1 \right)= \ell -1\right), \quad \forall \ell\in [\delta+1], \label{eqn:alphaelldef}
\end{equation}
and define a probability distribution $\bm{\zeta}_{n,\delta}$ on $[\delta+1]$:
\begin{equation}
\bm{\zeta}_{n,\delta}(\ell):=(\alpha_\ell/\ell)/\left(\sum_{s=1}^{\delta+1} (\alpha_s/s) \right), \quad \forall \ell\in [\delta+1]. \label{eqn:incrementsizelimitdef}
\end{equation}
As will be shown in the next theorem, $\bm{\zeta}_{n,\delta}$ is the increment distribution of a compound Poisson approximation to $\bar{N}_\delta$ when $p\to\infty$.
We also introduce the following sparsity conditions:
a matrix is said to be row-$\kappa$ sparse if every row has at most $\kappa$ nonzero elements. This is a weaker sparsity condition than the block sparsity condition of \cite{hero2012hub} (see also Subsection \ref{sec:taukappaspa}). The next definition is a stronger sparsity condition than row-$\kappa$ sparsity but remains weaker than block sparsity.
\begin{definition}[$(\tau,\kappa)$ sparsity]
A $p$ by $p$ dimensional symmetric matrix is said to be $(\tau,\kappa)$ sparse if it is row-$\kappa$ sparse and its lower $p-\tau$ by $p-\tau$ block is diagonal.
\end{definition}
Another relevant quantity is the normalized determinant defined as follows:
\begin{definition}[Normalized determinant]\label{def:nordet}
For any symmetric, positive definite matrix $\bm{A} \in \mathbb{R}^{p\times p}$, its normalized determinant $\mu(\bm{A})$ is defined by $$\mu(\bm{A}):= \prod\limits_{i=1}^p \frac{\lambda_i(\bm{A})}{\lambda_p(\bm{A})} = \frac{\text{det}(\bm{A})}{\left(\lambda_p(\bm{A})\right)^p},$$
where $\lambda_1(\bm{A})\leq \lambda_2(\bm{A})\leq \cdots \leq \lambda_p(\bm{A})$ are the eigenvalues of $\bm{A}$.
\end{definition}
For $\mathcal{I}\subset [p]$ denote by $\bm{A}_{\mathcal{I}}$ the set of all $|\mathcal{I}|\times |\mathcal{I}|$ submatrices of $\bm{A}\in \mathbb{R}^{p\times p}$ obtained by extracting the corresponding rows and columns indexed by $\mathcal{I}$. The set $\bm{A}_{\mathcal{I}}$ contains $|\mathcal{I}|!$ matrices that are all equivalent up to a permutation applied simultaneously to both rows and columns. Define the \textit{local normalized determinant of degree $m$} of a matrix $\bm{A}\in \mathbb{R}^{p\times p}$ to be $\mu_m(\bm{A}) = \min\{\mu(\bm{A}_{\mathcal{I}}) : \mathcal{I}\subset [p], |\mathcal{I}|=m \} $. Note that $\mu(\bm{A}_{\mathcal{I}})$ is well defined since $\mu(\cdot)$ is invariant to simultaneous application of a permutation to both rows and columns. For $\bm{A}\in \mathbb{R}^{p\times p}$ further define the \emph{inverse local normalized determinant}
\begin{equation}
\mu_{n,m} (\bm{A}) : =\begin{cases} [\mu_{m}(\bm{A})]^{-\frac{n-1}{2}}, & \bm{A} \text{ symmetric positive definite but not diagonal, } \\
1, & \bm{A} \text{ symmetric positive definite and diagonal. }
\end{cases} \label{eqn:munmA} \end{equation}
By definition $\mu(A)\in (0,1]$ and $\mu_{n,m} (\bm{A}) \in [1,\infty)$.
Denote $\Gamma(x)$ the gamma function and let $a_n := \frac{\Gamma((n-1)/2)}{(n-2)\sqrt{\pi}\Gamma((n-2)/2)}$. With the above definitions in place, we now state our main theorem: when $p
\to \infty$, if the threshold $\rho$
approaches $1$ at a particular rate, then the sequence of vertex counts $\bar{N}_\delta$ converges in distribution to a compound Poisson distribution.
\begin{thm}
[Compound Poisson limit] \label{cor:Poissonlimit}
Let $ n \geq 4$ and $\delta$ be fixed positive integers.
Let $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$.
Assume that the threshold $\rho$ is a function of $p$ that satisfies $a_n 2^{\frac{n}{2}}p^{1+\frac{1}{\delta}}(1-\rho)^{\frac{n-2}{2}}\to e_{n,\delta}$ as $p\to \infty$, where $e_{n,\delta}$ is some positive finite constant that possibly depends on $n$ and $\delta$. Denote $\lambda_{n,\delta}(e_{n,\delta}) = \frac{1}{\delta !} \left( e_{n,\delta}\right)^\delta \sum_{\ell=1}^{\delta+1}\frac{\alpha_\ell}{\ell}$. Suppose $\bm{\Sigma}$, after some row-column permutation, is $(\tau_p,\kappa_p)$ sparse with $\lim\limits_{p\to \infty} \frac{\tau_p}{p}+\mu_{n,2\delta+2}\left(\bm{\Sigma}\right)\frac{\kappa_p}{p} \to 0$. Then $\bar{N}_\delta$, a generic random variable in the set $\{N_{E_\delta}^{(\bm{\Psi})},N_{\breve{V}_\delta}^{(\bm{\Psi})},N_{V_\delta}^{(\bm{\Psi})}: \bm{\Psi} \in \{\bm{R},\bm{P}\} \}$, satisfies:
\begin{equation}
\bar{N}_\delta \overset{\mathcal{D}}{\to} \operatorname{CP}(\lambda_{n,\delta}(e_{n,\delta}),\bm{\zeta}_{n,\delta}) \text{ as } p\to \infty. \label{eqn:compoilimdis2}
\end{equation}
\end{thm}
If only the vertex counts in the empirical correlation graph is of interest, then the $(\tau,\kappa)$ sparsity assumption can be relaxed to row-$\kappa$ sparsity.
\begin{lem}[Compound Poisson limit in empirical correlation graph]
\label{item:Poissonlimita}
Let $ n \geq 4$ and $\delta$ be fixed positive integers.
Let $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$.
Assume that the threshold $\rho$ is a function of $p$ that satisfies $a_n 2^{\frac{n}{2}}p^{1+\frac{1}{\delta}}(1-\rho)^{\frac{n-2}{2}}\to e_{n,\delta}$ as $p\to \infty$, where $e_{n,\delta}$ is some positive finite constant that possibly depends on $n$ and $\delta$. Denote $\lambda_{n,\delta}(e_{n,\delta}) = \frac{1}{\delta !} \left( e_{n,\delta}\right)^\delta \sum_{\ell=1}^{\delta+1}\frac{\alpha_\ell}{\ell}$. Suppose $\bm{\Sigma}$ is row-$\kappa_p$ sparse with $\lim\limits_{p\to \infty} \mu_{n,2\delta+2}\left(\bm{\Sigma}\right)\frac{\kappa_p}{p} \to 0$. Then $\tilde{N}_\delta$, a generic random variable in the set $\{N_{E_\delta}^{(\bm{R})},N_{\breve{V}_\delta}^{(\bm{R})},N_{V_\delta}^{(\bm{R})} \}$, satisfies:
\begin{equation}
\tilde{N}_\delta \overset{\mathcal{D}}{\to} \operatorname{CP}(\lambda_{n,\delta}(e_{n,\delta}),\bm{\zeta}_{n,\delta}) \text{ as } p\to \infty. \label{eqn:compoilimdis}
\end{equation}
\end{lem}
\begin{rem} \label{rem:remarkofthm1}
The condition $a_n2^{\frac{n}{2}}p^{1+\frac{1}{\delta}}(1-\rho)^{\frac{n-2}{2}}\to e_{n,\delta}>0$ is equivalent to
$$
p^{\frac{2}{n-2}\left(1+\frac{1}{\delta}\right)}(1-\rho)\to \left(\frac{e_{n,\delta}}{a_n2^{\frac{n}{2}}}\right)^{\frac{2}{n-2}}= \frac{1}{2}\left(\frac{e_{n,\delta}}{2a_n}\right)^{\frac{2}{n-2}} ,
$$
which indicates that $\rho\to 1$ at rate $p^{-\frac{2}{n-2}\left(1+\frac{1}{\delta}\right)}$.
{ As will be discussed in more detail in Remark \ref{rem:rhorate}, this rate is in fact both necessary and sufficient for the expected counts $\mathbb{E} \tilde{N}_\delta$ to converge to a non-trivial limit. If $\rho$ does not converge to $1$, or converges to $1$ at a slower rate, then $\mathbb{E}\tilde{N}_\delta$ diverges to $\infty$, while if $\rho$ converges to $1$ at a faster rate then $\mathbb{E} \tilde{N}_\delta$ converges to $0$.
}
\ONE{This particular rate on $\rho$ is consistent with the rate
of the existing Poisson approximation results in
random geometric graphs \cite{penrose2003random} as will be
discussed in Section \ref{sec:rangeo}.} A sequence of correlation thresholds $\rho=\rho_p$ that satisfies this condition is
\begin{equation}
\rho_p=1-\frac{1}{2}\left(\frac{e_{n,\delta}}{2a_n p^{1+\frac{1}{\delta}}}\right)^{\frac{2}{n-2}}. \label{eqn:rhopformula}
\end{equation}
Observe that Theorem \ref{cor:Poissonlimit} and Lemma \ref{item:Poissonlimita} hold for any mean $\bm{\mu}$ and any shaping function $g$ when $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$. We will provide intuition for this invariance property in Remark \ref{rem:Udisconsequence}.
\end{rem}
The proofs of Theorem \ref{cor:Poissonlimit} and Lemma \ref{item:Poissonlimita} will be presented in Subsection \ref{sec:compoihighdim}.
We will call $\operatorname{CP}(\lambda_{n,\delta}(e_{n,\delta}),\bm{\zeta}_{n,\delta})$ the \emph{limiting compound Poisson distribution, approximation or characterization.} Since $\bar{N}_\delta$ and $\operatorname{CP}(\lambda_{n,\delta}(e_{n,\delta}),\bm{\zeta}_{n,\delta})$ are discrete, \eqref{eqn:compoilimdis2} is equivalent to:
$$
d_{\operatorname{TV}}\left(\mathscr{L}\left(\bar{N}_\delta\right),\operatorname{CP}(\lambda_{n,\delta}(e_{n,\delta}),\bm{\zeta}_{n,\delta})\right) \to 0 \text{ as } p\to \infty,
$$
where $\mathscr{L}\left(\cdot\right)$ represents the probability distribution of the argument,
and $d_{\operatorname{TV}}\left(\cdot,\cdot\right)$ is the total variation distance between two probability distributions.
A variant of Theorem \ref{cor:Poissonlimit} for finite $p$ (Theorem \ref{thm:Poissonultrahigh}), which establishes an upper bound on the total variation distance between $\mathscr{L}\left(\bar{N}_\delta\right)$ and a compound Poisson distribution, will be presented in Subsection \ref{sec:compoihighdim}. \ONE{Theorem \ref{cor:Poissonlimit} and Theorem \ref{thm:Poissonultrahigh} specify asymptotic compound Poisson limits and non-asymptotic
bounds on the full distribution of vertex counts. These limits correct and extend the Poisson limits that were falsely claimed to hold for all finite $n,\delta$,
although we note that the compound Poisson limit can be well approximated by the Poisson limit in the case of moderately large $n$ or large $\delta$ (see Sec. \ref{sec:limitingcompoundPoisson}).}
In Theorem \ref{cor:Poissonlimit} the $(\tau,\kappa)$ sparsity condition and the condition $\mu_{n,m}(\bm{\Sigma})<\infty$ are assumed. We elaborate on these two conditions in the next two subsections.
\subsection{$(\tau,\kappa)$ sparsity}
\label{sec:taukappaspa}
The matrix \eqref{examat} below is an example of a $(\tau,\kappa)$ sparse matrix
with $\tau=2, \kappa=3$. This $5\times5$ symmetric matrix is $(2,3)$ sparse since each of the first $2$ rows has at most $3$ nonzero elements and the lower $3\times3$ block is diagonal.
\begin{equation}
\begin{pmatrix}
5 & 0 & 2 & 0 & 1 \\
0 & 8 & 3 & 0 & 0 \\
2 & 3 & 6 & 0 & 0 \\
0 & 0 & 0 & 7 & 0 \\
1 & 0 & 0 & 0 & 8
\end{pmatrix} \label{examat}
\end{equation}
If the adjacency matrix of a graph $(\mathcal{V},\mathcal{E})$ is $(\tau,\kappa)$ sparse, then the vertices $\mathcal{V}$ can be partitioned into two disjoint subsets $\mathcal{V}_1$ and $\mathcal{V}_2$ with the following properties: 1) $|\mathcal{V}_1|\leq \tau$; 2) there is no edge between any two vertices in $\mathcal{V}_2$; 3) the degree of any vertex in $\mathcal{V}_1$ is no more than $\kappa-1$; 4) edges connecting vertex in $\mathcal{V}_1$ and $\mathcal{V}_2$ may exist.
When the dispersion parameter $\bm{\Sigma}$ is row-$\kappa$ sparse, the authors of \cite{hero2011large} studied the mean of the quantities $N^{(\bm{R})}_{E_1}$ and $N^{(\bm{R})}_{V_1}$ and they obtained limits of $P(N_{E_1}^{(\bm{R})}>0)$ or $P(N_{V_1}^{(\bm{R})}>0)$ when $p\to\infty$ while fixing $n$. In \cite{hero2012hub} these results were extended to empirical partial correlation graphs when the dispersion parameter $\bm{\Sigma}$ is assumed to be block-$\tau$ sparse up to a row-column permutation, i.e., there exists a permutation matrix $\bm{T}$ such that
\begin{equation}
\bm{T}\bm{\Sigma}\bm{T}^\top = \begin{pmatrix}
\bm{\Sigma}_{11} & \bm{\Sigma}_{12} \\
\bm{\Sigma}_{21} & \bm{D}_{p-\tau}
\end{pmatrix} \label{eqn:permutationresult}
\end{equation}
where $\bm{\Sigma}_{12} = \bm{\Sigma}_{21}^\top = \bm{0}\in \mathbb{R}^{\tau\times (p-\tau)}$ and $\bm{D}_{p-\tau}\in \mathbb{R}^{(p-\tau)\times (p-\tau)}$ is some diagonal matrix.
In Theorem \ref{cor:Poissonlimit} of this paper $\bm{\Sigma}$ is assumed to be $(\tau,\kappa)$ sparse after row-column permutation, i.e. there exists a permutation matrix $\bm{T}$ such that \eqref{eqn:permutationresult} holds with $\bm{D}_{p-\tau}\in \mathbb{R}^{(p-\tau)\times (p-\tau)}$ some diagonal matrix and with the first $\tau$ rows $(\bm{\Sigma}_{11}\ \bm{\Sigma}_{12})$ being row-$\kappa$ sparse.
It is clear that the $(\tau,\kappa)$ sparsity condition is more general than the block sparsity condition as that there is no restriction on $\bm{\Sigma}_{12}=0$.
Indeed, every block-$\tau$ sparse matrix is $(\tau,\kappa)$ sparse with $\kappa = \tau$. Nevertheless, the $(\tau,\kappa)$ sparsity with $\kappa=\tau$ allows non-zeros in the top-right submatrix, which permits more possible correlations between the variables relative to the block-$\tau$ sparsity in correlation graphical models. To see this, consider the associated graphical model $\mathcal{G}_0(\tilde{\bm{\Sigma}})$ for a correlation matrix $\tilde{\bm{\Sigma}}$. Recall in Section \ref{Hubfra} we define $\mathcal{G}_\rho(\cdot)$ as the graph with adjacency matrix obtained by thresholding a matrix with $\rho$.
In Figure \ref{fig:covasscom}, nodes represent the variables and edges represent the correlation between variables. The left panel is a graphical model associated with the block-$3$ sparse assumption, while the right panel satisfies $(\tau,\kappa)$ sparsity with $(\tau,\kappa)=(3,3)$. The later has more correlations (the red edges) across the two sets of variables in the $2$ circles. The $(\tau,\kappa)$ sparsity condition with $\kappa>\tau$ allows additional correlations between variables.
{ Notably, $(\tau,\kappa)$ sparsity allows the underlying graphical model to be connected as shown in the following example.}
\begin{figure}[ht]
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[width=.8\linewidth]{graphs/assumptionscomparision1.png}
\caption{block sparsity}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[width=.8\linewidth]{graphs/assumptionscomparision2.png}
\caption{$(\tau,\kappa)$ sparsity}
\end{subfigure}
\caption{Diagram of the correlation graph $\mathcal{G}_0(\tilde{\bm{\Sigma}})$ for a $p=7$ dimensional distribution with two different $7\times 7$ correlation matrices. The left panel is associated with a block-3 sparse assumption on $\tilde{\bm{\Sigma}}$. Only the $\tau = 3$ variables in the group inside the left circle are correlated: there are no correlations (edges) between the remaining $4$ variables in the right circle and there are no correlations across the two sets of variables in the different circles. The right panel is associated with $(\tau,\kappa)=(3,3)$ sparsity on $\tilde{\bm{\Sigma}}$, where two additional edges, representing correlations between variables, exist across the two groups.}
\label{fig:covasscom}
\end{figure}
\begin{exa}[A connected correlation graph that is $(\tau,\kappa)$ sparse]
\label{exa:connectedtaukappa}
Start with a set of $p$ vertices $[p]$ with no edges. Denote the neighborhood of vertex $i$ to be $\operatorname{NB}(i)$. We construct a connected $(\tau,\kappa)$ sparse graph on $[p]$ by creating neighborhoods of the first $\tau$ vertices according to the following rule:
\begin{align*}
\operatorname{NB}(i)= & \{ i+1, \tau+ (i-1) (\kappa-2) +1, \tau+ (i-1) (\kappa-2) +2 ,\ldots, \tau+ i (\kappa-2) \}, \quad \text{ for } i\in [\tau-1], \\
\operatorname{NB}(\tau) = & \{\tau+ (\tau-1) (\kappa-2) +1, \tau+ (\tau-1) (\kappa-2) +2,\ldots, \tau+ \tau (\kappa-2) + 1 \}.
\end{align*}
With this construction, each vertex among the first $\tau$ vertices is connected to $\kappa-1$ other vertices and the remaining $p-\tau$ vertices $[p]\setminus [\tau]$ do not connect to each other. As a result the associated adjacency matrix is $(\tau,\kappa)$ sparse. Moreover, as long as
\begin{equation}
\tau+ \tau (\kappa-2) + 1 = \tau (\kappa-1) + 1 \geq p, \label{eqn:connectedtaukappa}
\end{equation}
the graph is connected. Now, let $\bm{\Sigma}$ be a matrix with diagonal $1$ and the locations of the off-diagonal nonzero elements specified by the graph constructed. For simplicity, choose all the off-diagonal nonzero elements to be $\xi$. If $\xi$ is small enough, then $\bm{\Sigma}$ is close to the identity matrix, hence positive definite and is therefore a {\em bone fide} correlation matrix. We have thus constructed a $(\tau,\kappa)$ sparse correlation matrix whose associated graphical model is connected provided \eqref{eqn:connectedtaukappa} is true. Now observe that by specifying $\tau=p^{\alpha}$ and $\kappa=p^{\beta}$ with $\alpha+\beta>1$, \eqref{eqn:connectedtaukappa} is indeed satisfied for large $p$. Moreover, if $\alpha,\beta\in (0,1)$ then $\tau,\kappa=o(p)$. By choosing $\xi$ as a decreasing function of $p$, the condition number of $\bm{\Sigma}$ will not increase too fast w.r.t. $p$ so that the condition that $\lim\limits_{p\to \infty} \frac{\tau}{p}+\mu_{n,2\delta+2}\left(\bm{\Sigma}\right)\frac{\kappa}{p} \to 0$ in Theorem \ref{cor:Poissonlimit} is satisfied.
\end{exa}
On the other hand, $(\tau,\kappa)$ sparsity is a stronger assumption than row-$\kappa$ sparsit
, since every $(\tau,\kappa)$ sparse matrix is row-$\kappa$ sparse. $(\tau,\kappa)$ sparsity is thus an intermediate level of sparsity lying between block sparsity and row sparsity.
\begin{comment}
\begin{figure}
\centering
\includegraphics[scale=0.15]{sparsitypatternp60tao6kappa8_comparison.png}
\caption{These are sparse patterns of random generated $(\tau,\kappa)$ sparse matrices of dimension $60\times 60$, with blue dots representing nonzero elements. On the left is a random generated $(\tau,\kappa)$ sparse matrix with $\tau=11$ and $\kappa=26$. Out of 3600 entries, 422 are nonzero. On the right is a random generated block-$\tau$ sparse matrix with $\tau=11$. Only 170 out of 3600 entries are nonzero.}
\label{fig:spapat}
\end{figure}
\end{comment}
\begin{rem} \label{rem:firsttaurworemark}
In Theorem \ref{cor:Poissonlimit} we supposed that the dispersion parameter $\bm{\Sigma}$, after some row-column permutation, is $(\tau,\kappa)$ sparse. In this paper we are interested in $\bar{N}_\delta$, which is invariant under permutation of the $p$ variables. Since permutation of the variables is equivalent to row-column permutation of the dispersion parameter $\bm{\Sigma}$, without loss of generality, we can assume that the variables have been permuted such that $\bm{\Sigma}$ is $(\tau,\kappa)$ sparse.
\end{rem}
\subsection{Local normalized determinant}
\label{sec:upperboundjointdensity}
In this subsection, we elaborate on the normalized determinant $\mu(\cdot)$ and the inverse local normalized determinant defined in Subsection \ref{sec:unifiedtheorem}.
\ONE{The normalized determinant measures the closeness between the identity matrix and the dispersion parameter $\bm{\Sigma}$. This quantity plays an important role in Theorem \ref{cor:Poissonlimit}.}
Observe that $\mu(\bm{A})\in (0,1]$ and that, $\mu(\bm{A})=1$ if and only if $\bm{A}$ is a multiple of $\bm{I}_p$. Moreover $\mu(\bm{A})$ is close to $1$, and hence bounded away from $0$, as long as all eigenvalues concentrate around a positive number. Below is an example of a sequence of symmetric positive definite matrices with well-concentrated eigenvalues such that their normalized determinants are uniformly bounded away from $0$.
\begin{exa} \label{exa:normaliseddeterminant}
Let $\{\alpha_i\}_{i=1}^p$ be a positive real sequence. Let $\{\beta_i\}_{i=1}^\infty $ be a positive, decreasing sequence such that $\sum\limits_{i=1}^\infty \beta_i < \infty$. Consider $\bm{A} \in \mathbb{R}^{p\times p}$ a symmetric positive definite matrix with eigenvalues $\lambda_i = \alpha_p \exp(-\beta_i)$ for $1\leq i\leq p-1$ and $\lambda_p = \alpha_p$. Then
$$
\mu(\bm{A}) = \exp\left(-\sum\limits_{i=1}^{p-1} \beta_i\right).
$$
Consider now the case that $p$ is increasing and consider a sequence of matrices $\bm{A}$ of increasing dimension with the above properties. For this sequence $\mu(\bm{A}) \geq \exp\left(-\sum\limits_{i=1}^{\infty} \beta_i\right)>0$, i.e. $\mu(\bm{A})$ is bounded uniformly away from $0$.
\end{exa}
It follows by the interlacing property (see Theorem 8.1.7 in \cite{golub2012matrix}) that $\mu_m(\bm{A})$ is decreasing with respect to $m\in [p]$ for any symmetric positive definite matrix $\bm{A}\in \mathbb{R}^{p\times p}$. Thus the inverse local normalized determinant $\mu_{n,m}(\bm{A})$ is increasing with respect to $m \in [p]$.
It turns out that the inverse local normalized determinant of the dispersion matrix $\bm{\Sigma}$ will play an important role in our study of the distribution of $\bar{N}_\delta$. Indeed, when $\delta\geq 2$, $N_{\EE_\delta}^{(\bm{R})}$ can be represented as a sum of indicator functions of the event that a subgraph of $\delta+1$ vertices is isomorphic to $\Gamma_\delta$ (see \eqref{eqn:Nedgedef1} for the precise formula). Each term in the summation involves only $\delta+1$ variables, and thus each pair of two such terms involves at most $2(\delta+1)$ variables.
Hence $\mu_{2\delta+2}(\bm{\Sigma})$ or $\mu_{n,2\delta+2}(\bm{\Sigma})$ controls the correlation between two indicator terms in the summation of $N_{\EE_\delta}^{(\bm{R})}$, which determines the convergence of $N_{\EE_\delta}^{(\bm{R})}$ to a compound Poisson distribution.
In Theorem \ref{cor:Poissonlimit} we assume that $\mu_{n,2\delta+2}\left(\bm{\Sigma}\right)\frac{\kappa_p}{p} \to 0$, which holds when $\mu_{n,2\delta+2}\left(\bm{\Sigma}\right)$ is either bounded or increasing with rate $o(\frac{p}{\kappa_p})$. In the Section \ref{sec:localdeterminantvseigenvalues} of the Appendix, Lemma \ref{lem:sufass} provides a bound on $\mu_{n,2\delta+2}\left(\bm{\Sigma}\right)$ in terms of the condition number and the eigenvalues of $\bm{\Sigma}$.
\section{Analytical expressions for the compound Poisson parameters}
\label{sec:limitingcompoundPoisson}
Recall that the limiting compound Poisson distribution in Theorem \ref{cor:Poissonlimit} is in terms of $\alpha_\ell$, while the non-asymptotic compound Poisson distributions in Theorem \ref{thm:Poissonultrahigh} is in terms of $\alpha(\ell,r_\rho)$. Moreover the second moment approximation established in Proposition \ref{lem:2ndmoment} also involves the term $\alpha(\ell,r_\rho)$ due to \eqref{eqn:secondmomentofCP}. Since $\alpha_\ell$ and $\alpha(\ell,r_\rho)$ are expressed in terms of random (pseudo) geometric graphs, they are tedious to compute.
In subsection \ref{sec:approximationstoalphaell}, we obtain simple analytical approximations to these quantities for moderately large $n$ or $\delta$. In subsection \ref{sec:exacttoalphaell}, exact analytical formulae for these quantities are established for the special case $\delta=1$ and $\delta=2$.
\subsection{Approximations to $\alpha_\ell$ and $\alpha(\ell,r_\rho)$ when $\delta\geq 2$}
\label{sec:approximationstoalphaell}
In this subsection we show that when $n$ or $\delta$ is moderately large, $\alpha_1\approx 1$ and $\alpha_\ell\approx 0$ for $2\leq \ell\leq \delta+1 $ (parallel results for $\alpha(\ell,r_\rho)$ are $\alpha(1,r_\rho)\approx 1$ and $\alpha(\ell,r_\rho)\approx 0$ for $2\leq \ell\leq \delta+1$). These approximations yield simple formulae for parameters of the limiting and non-asymptotic compound Poisson distributions,
and for parameters of the second moment \eqref{eqn:secondmomentofCP}. Importantly, the compound Poisson distributions are well approximated by Poisson distributions for $n,\delta$ moderately large.
To show that $\alpha_\ell$ is small for $\ell\geq 2$ and $\alpha_1\approx 1$, it suffices to establish a vanishing upper bound on $\sum_{\ell=2}^{\delta+1}\alpha_\ell$, since $\sum_{\ell=1}^{\delta+1}\alpha_\ell=1$. By definition of $\alpha_\ell$, $\sum_{\ell=2}^{\delta+1}\alpha_\ell=P\left(\operatorname{{\bf NUV}}\left(\{\tilde{\bm{u}}_i \}_{i=1}^\delta, 1 \right) \geq 1 \right)$, where $\{\tilde{\bm{u}}_i\}_{i=1}^{\delta} \overset{\text{i.i.d.}}{\sim} \operatorname{{\bf unif}}(B^{n-2})$. The next lemma establishes an upper bound on the geometric quantity $P\left(\operatorname{{\bf NUV}}\left(\{\tilde{\bm{u}}_i \}_{i=1}^\delta, 1 \right) \geq 1 \right)$.
\begin{lem} \label{lem:randgeogra}
Let $n\geq 4$ and $\delta \geq 2$.
\begin{enumerate}[label=(\alph*)]
\item
\label{item:randgeograa}
Consider $\{\tilde{\bm{u}}_i\}_{i=1}^{\delta} \overset{\text{i.i.d.}}{\sim} \operatorname{{\bf unif}}(B^{n-2})$. Then
\begin{align*}
\sum_{\ell=2}^{\delta+1}\alpha_\ell = & P\left(\operatorname{{\bf NUV}}\left(\{\tilde{\bm{u}}_i \}_{i=1}^\delta, 1 ; n-2 \right) \geq 1 \right) \\
\leq & \delta (n-2)\int_{0}^1 \left(1-\frac{r^2}{4}\right)^{\frac{(n-2)(\delta-1)}{2}} r^{n-3}dr\\
= & \delta(n-2) 2^{n-3}B\left(\frac{1}{4};\frac{n-2}{2},\frac{(n-2)(\delta-1)}{2}+1\right), \numberthis \label{eqn:tolvarincrementdeltabetafunction}
\end{align*}
where $B\left(\cdot;\cdot,\cdot\right)$ is the incomplete beta function.
\item
\label{item:randgeograb}
$$\int_{0}^1 \left(1-\frac{r^2}{4}\right)^{\frac{(n-2)(\delta-1)}{2}} r^{n-3}dr \leq \begin{cases} \left(\frac{4}{5}\right)^{\frac{(n-2)\delta -1}{2}}+\left(1-\sqrt{\frac{4}{5}}\right)\left(\frac{3}{4}\right)^{\frac{(n-2)(\delta-1)}{2}}, & \delta=2,3, \\ \exp\left(\frac{1}{4}\right) \left(\frac{\delta-1}{\delta}\right)^{\frac{(n-2)(\delta-1)}{2}} \left(\frac{4}{\delta}\right)^{\frac{n-3}{2}}, & \delta \geq 4. \end{cases} $$
\end{enumerate}
\end{lem}
Lemma \ref{lem:randgeogra} \ref{item:randgeograa} establishes an upper bound for the probability that there is at least one universal vertex in the random geometric graph generated by the uniform distribution in the unit ball. Lemma \ref{lem:randgeogra} \ref{item:randgeograb} provides an explicit upper bound for part \ref{item:randgeograa}
and this upper bound provides the following insight
when $n$ or $\delta$ is large. In particular, when $\delta$ is fixed, $P\left(\operatorname{{\bf NUV}}\left(\{\tilde{\bm{u}}_i \}_{i=1}^\delta, 1 ; n-2 \right) \geq 1 \right)$ decays exponentially as $n$ increases. While $n$ is fixed, it decays at rate $\delta^{-{\frac{n-3}{2}}}$ as $\delta$ increases. Such decay rates also apply to $\alpha_\ell$ for $\ell\geq 2$ and $|\alpha_1-1|$.
\begin{figure}[ht]
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[width=.8\linewidth]{graphs/disincrementdelta2new.png}
\caption{Log-scale comparison of the decay when $\delta=2$}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[width=.8\linewidth]{familyupperbounew.png}
\caption{Family of upper bounds}
\end{subfigure}
\caption{ (a) is a comparison in the log-scale between the upper bound on $\sum_{\ell=2}^{\delta+1}\alpha_\ell$ by \eqref{eqn:tolvarincrementdeltabetafunction} with $\delta=2$ and the exact value of $\sum_{\ell=2}^{\delta+1}\alpha_\ell$ by Example \ref{lem:alphaldelta2}. (b) is the plot of the upper bound on $\sum_{\ell=2}^{\delta+1}\alpha_\ell$ as a function of $n$ for $\delta$ from $2$ to $7$.}
\label{fig:incrementtodeltauppbou}
\end{figure}
As illustrated in Figure \ref{fig:incrementtodeltauppbou} (a), when $\delta=2$, $\sum_{\ell=2}^{\delta+1}\alpha_\ell$ and the upper bound in \eqref{eqn:tolvarincrementdeltabetafunction} both decay to $0$ exponentially fast as $n$ increases. This diagram demonstrates that the upper bound in \eqref{eqn:tolvarincrementdeltabetafunction} captures the decay rate of $\sum_{\ell=2}^{\delta+1}\alpha_\ell$. Figure \ref{fig:incrementtodeltauppbou} (b) plots the upper bounds in \eqref{eqn:tolvarincrementdeltabetafunction} as functions of $n$ for fixed $\delta$. As is clear from the plot, as long as the number of samples $n$ is above $40$, $\sum_{\ell=2}^{\delta+1}\alpha_\ell\approx 0$, which yields $\alpha_\ell\approx0$ for $2\leq \ell \leq \delta+1$ and $\alpha_1\approx 1$.
Moreover, as $\delta$ increases, the number of samples $n$ required for $\sum_{\ell=2}^{\delta+1}\alpha_\ell\approx 0$ decreases.
Denote the Dirac Distribution at $a$ by $\operatorname{Dirac}(a)$. The next lemma establishes approximations to the limiting compound Poisson distribution in Theorem \ref{cor:Poissonlimit}, with the approximation errors upper bounded in terms of $\sum_{\ell=2}^{\delta+1}\alpha_\ell$.
\begin{lem}[Approximation to the limiting Compound Poisson distribution]
\label{prop:jumpsizedecay}
Consider $n\geq 4$ and $\delta\geq 2$. Let $e_{n,\delta}$ be the same as in Theorem \ref{cor:Poissonlimit}.
\begin{enumerate}[label=(\alph*)]
\item\label{item:jumpsizedecaya}
The increment distribution satisfies
$
d_{\operatorname{TV}}\left(\bm{\zeta}_{n,\delta}, \operatorname{Dirac}(1) \right)\leq \sum_{\ell=2}^{\delta+1}\alpha_\ell.
$
\item \label{item:jumpsizedecayb}
The arrival rate satisfies
$$
\left|\lambda_{n,\delta}(e_{n,\delta})-\frac{(e_{n,\delta})^\delta}{\delta!}\right|\leq \frac{3}{2}\frac{(e_{n,\delta})^\delta}{\delta!} \sum_{\ell=2}^{\delta+1}\alpha_\ell.
$$
\item \label{item:jumpsizedecayc}
The limiting compound Poisson distribution satisfies
$$
d_{\operatorname{TV}}\left(\operatorname{CP}(\lambda_{n,\delta},\bm{\zeta}_{n,\delta}),\operatorname{Pois}\left(\frac{(e_{n,\delta})^\delta}{\delta!}\right)\right) \leq \frac{5}{2}\frac{(e_{n,\delta})^\delta}{\delta!} \sum_{\ell=2}^{\delta+1}\alpha_\ell.
$$
\end{enumerate}
\end{lem}
The proof of Lemma \ref{prop:jumpsizedecay} \ref{item:jumpsizedecayc} relies on a general upper bound of the total variation distance between two compound Poisson distributions (Lemma \ref{eqn:cptv} in Section \ref{sec:auxlem} of the Appendix). Lemma \ref{eqn:cptv} may be of independent interest in compound Poisson approximation to non-Poissonian distributions (see \cite{barbour2001topics} and the references therein).
From Lemma \ref{prop:jumpsizedecay} \ref{item:jumpsizedecaya} and Lemma \ref{lem:randgeogra}, the total variation distance between the increment distribution and Dirac distribution at $1$ decays exponentially as $n$ increases and decays at rate $\delta^{-\frac{n-3}{2}}$ as $\delta$ increases.
Provided that the threshold $\rho$ is chosen such that $\frac{(e_{n,\delta})^\delta}{\delta!}$ is not large, the upper bounds in Lemma \ref{prop:jumpsizedecay} \ref{item:jumpsizedecayb} and \ref{item:jumpsizedecayc} have the same interpretation as part \ref{item:jumpsizedecaya}. In that case, the limiting compound Poisson are well approximated by the Poisson distribution $\operatorname{Pois}\left(\frac{(e_{n,\delta})^\delta}{\delta!}\right)$. By combining Lemma \ref{prop:jumpsizedecay} \ref{item:jumpsizedecayc} and Theorem \ref{cor:Poissonlimit}, we immediately obtain the following result on Poisson approximation to $\mathscr{L}(\bar{N}_\delta)$ as $p\to\infty$.
\begin{prop}[Poisson approximation as $p\to \infty$]
\label{prop:poiapplimit}
Suppose all the conditions in Theorem \ref{cor:Poissonlimit} hold. Then as $p\to \infty$,
$$
d_{\operatorname{TV}}\left(\mathscr{L}(\bar{N}_\delta), \operatorname{Pois}\left(\frac{(e_{n,\delta})^\delta}{\delta!}\right)\right) \leq \frac{5}{2}\frac{(e_{n,\delta})^\delta}{\delta!} \sum_{\ell=2}^{\delta+1}\alpha_\ell.
$$
\end{prop}
We now turn our attention to the parameter $\alpha(\ell,r_\rho)$ of the non-asymptotic compound Poisson distribution. Our results on $\alpha(\ell,r_\rho)$ are similar to those on $\alpha_\ell$ but the proofs are more technical. To establish that $\alpha(1,r_\rho)\approx 1$ and $\alpha(\ell,r_\rho)\approx 0$ for $2\leq \ell\leq \delta+1$, it suffices to obtain a vanishing upper bound on $\sum_{\ell=2}^{\delta+1}\alpha(\ell,r_\rho)$. By \eqref{eqn:alphalrhorgg}, when $\rho>3/5$, $\sum_{\ell=2}^{\delta+1}\alpha(\ell,r_\rho)=P\left(\operatorname{{\bf NUV}}\left(\{\bm{u}'_i \}_{i=1}^{\delta+1}, r ; n-2 \right) \geq 2 |\operatorname{deg}(\bm{u}'_{{\delta+1}})=\delta\right)$, where $\{\bm{u}'_i\}_{i=1}^{\delta+1} \overset{\text{i.i.d.}}{\sim} \operatorname{{\bf unif}}(S^{n-2})$. The next lemma is an analogous result to Lemma \ref{lem:randgeogra}.
\begin{lem} \label{lem:randgeograconduppbou}
Let $n\geq 4$ and $\delta \geq 2$.
\begin{enumerate}[label=(\alph*)]
\item \label{item:randgeograconduppboua}
Consider $\{\bm{u}'\}_{i=1}^{{\delta+1}} \overset{\text{i.i.d.}}{\sim} \operatorname{{\bf unif}}(S^{n-2})$. Then for $0<r<\sqrt{2}$
\begin{align*}
&P\left(\operatorname{{\bf NUV}}\left(\{\bm{u}'_i \}_{i=1}^{\delta+1}, r ; n-2 \right) \geq 2 |\operatorname{deg}(\bm{u}'_{{\delta+1}})=\delta\right) \\
\leq & \bar{h}\left(\frac{1}{\sqrt{1-r^2/4}},n,\delta\right) \delta (n-2) \int_0^1 \left(1 -\left(\frac{r_1}{2}\right)^2 \right)^{\frac{(n-2){(\delta-1)}}{2}} r_1^{n-3}dr_1\\
= & \bar{h}\left(\frac{1}{\sqrt{1-r^2/4}},n,\delta\right) \delta (n-2) 2^{n-3} B\left(\frac{1}{4};\frac{n-2}{2},\frac{(n-2)(m-1)}{2}+1\right) ,
\end{align*}
where $\bar{h}(x,n,\delta)=x^{n+\delta-5}x^{(n-2)(\delta-1)}$.
\item
\label{item:randgeograconduppboub}
When $r\leq \begin{cases} 2\sqrt{1-\sqrt{1-1/5}}, & \delta = 2,3 \\ 2\sqrt{1-\sqrt{1-1/\delta}}, & \delta\geq 4 \end{cases}$, the upper bound in part \ref{item:randgeograconduppboua} is upper bounded by
\begin{align*}
\begin{cases} \delta (n-2)\{\left({\frac{\sqrt{5}}{2}}\right)^{\frac{\delta-2}{2}} \left(\frac{2}{\sqrt{5}}\right)^{\frac{(n-2)\delta-1}{2}}+(1-\frac{2}{\sqrt{5}})\left(\frac{\sqrt{5}}{2}\right)^{\frac{n+\delta-5}{2}}\left(\frac{3\sqrt{5}}{8}\right)^{\frac{(n-2)(\delta-1)}{2}}\}, & \delta=2,3, \\ \delta (n-2)\exp\left(\frac{1}{4}\right) \left(\sqrt{\frac{\delta}{\delta-1}}\right)^{\frac{\delta-2}{2}} \left(\sqrt{\frac{\delta-1}{\delta}}\right)^{\frac{(n-2)(\delta-1)}{2}} \left(\frac{4}{\sqrt{\delta(\delta-1)}}\right)^{\frac{n-3}{2}}, & \delta \geq 4. \end{cases}.
\end{align*}
\item
\label{item:randgeograconduppbouc}
When $\rho> 3/5$, $\sum_{\ell=2}^{\delta+1} \alpha(\ell,r_\rho)$ is upper bounded by the upper bound in part \ref{item:randgeograconduppboua} with $r$ replaced by $r_\rho$. When $\rho \geq \begin{cases} 4/\sqrt{5}-1, & \delta = 2,3 \\ 2\sqrt{1-1/\delta}-1, & \delta\geq 4 \end{cases}$,
$\sum_{\ell=2}^{\delta+1} \alpha(\ell,r_\rho)$ is upper bounded by the upper bound in part \ref{item:randgeograconduppboub} with $r$ replaced by $r_\rho$.
\end{enumerate}
\end{lem}
Lemma \ref{lem:randgeograconduppbou} \ref{item:randgeograconduppboua} establishes an upper bound for the conditional probability that there are at least two universal vertices, conditioned on the existence of one universal vertex, in the random geometric graph over points generated by the uniform distribution on the sphere.
The interpretations of Lemma \ref{lem:randgeograconduppbou} \ref{item:randgeograconduppboub}, \ref{item:randgeograconduppbouc} are similar to those of Lemma \ref{lem:randgeogra} \ref{item:randgeograb} and $\sum_{\ell=2}^{\delta+1} \alpha_\ell$, and are therefore omitted.
The next lemma establishes approximations to the non-asymptotic compound Poisson distribution in Theorem \ref{thm:Poissonultrahigh} and to its second moment, with the approximation errors upper bounded in terms of $\sum_{\ell=2}^{\delta+1}\alpha(\ell,r_\rho)$.
Denote $\bar{\lambda}_{p,n,\delta,\rho} := \binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta}$.
\begin{lem}[Approximation to the non-asymptotic Compound Poisson distribution]
\label{prop:jumpsizedecaynonasymptotic}
Consider $n\geq 4$ and $\delta\geq 2$.
\begin{enumerate}[label=(\alph*)]
\item\label{item:jumpsizedecaynonasymptotica}
The increment distribution satisfies
$
d_{\operatorname{TV}}\left(\bm{\zeta}_{n,\delta,\rho}, \operatorname{Dirac}(1) \right)\leq \sum_{\ell=2}^{\delta+1}\alpha(\ell,r_\rho).
$
\item \label{item:jumpsizedecaynonasymptoticb}
The arrival rate satisfies
$
\left|\lambda_{p,n,\delta,\rho} - \bar{\lambda}_{p,n,\delta,\rho}\right|\leq \frac{3}{2}\bar{\lambda}_{p,n,\delta,\rho} \sum_{\ell=2}^{\delta+1}\alpha(\ell,r_\rho).
$
\item \label{item:jumpsizedecaynonasymptoticc}
The non-asymptotic compound Poisson distribution satisfies
\begin{align*}
d_{\operatorname{TV}}\left( \operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho}), \operatorname{Pois}\left(\bar{\lambda}_{p,n,\delta,\rho}\right) \right)
\leq & \frac{5}{2}\bar{\lambda}_{p,n,\delta,\rho} \sum_{\ell=2}^{\delta+1}\alpha(\ell,r_\rho).
\end{align*}
\item \label{item:jumpsizedecaynonasymptoticd}
Let $Z\sim \operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho})$. Then the second moment of the non-asymptotic compound Poisson $\mathbb{E} Z^2$ in \eqref{eqn:secondmomentofCP} satisfies
$$
\left|\mathbb{E} Z^2 - (\bar{\lambda}_{p,n,\delta,\rho}+(\bar{\lambda}_{p,n,\delta,\rho})^2) \right|\leq \frac{3}{2}\bar{\lambda}_{p,n,\delta,\rho} \sum_{\ell=2}^{\delta+1}\alpha(\ell,r_\rho).
$$
\end{enumerate}
\end{lem}
The proof of Lemma \ref{prop:jumpsizedecaynonasymptotic} and its interpretation are analogous to those of Lemma \ref{prop:jumpsizedecay}, and are therefore omitted. By combining Lemma \ref{prop:jumpsizedecaynonasymptotic} \ref{item:jumpsizedecaynonasymptoticc} and Theorem \ref{thm:Poissonultrahigh}, we immediately obtain the following result on Poisson approximation to $\mathscr{L}(\bar{N}_\delta)$ for finite $p$. One can derive an equivalent proof of the following proposition by first applying Poisson approximation from Chen-Stein's method (e.g. \cite[Theorem 1]{arratia1990poisson}) to obtain an upper bound $d_{\operatorname{TV}}(\mathscr{L}(N_{E_\delta}^{(\bm{R})}), \operatorname{Pois}\left(\bar{\lambda}_{p,n,\delta,\rho}\right))$, which is then combined with the portmanteau result Proposition \ref{thm:allclose}.
\begin{prop}[Poisson approximation for finite $p$]
\label{prop:poiappfinitep}
Suppose all the conditions in Theorem \ref{thm:Poissonultrahigh} hold. Then $\bar{\lambda}_{p,n,\delta,\rho}\leq \frac{\gamma^\delta}{\delta !}$ and
\begin{align*}
d_{\operatorname{TV}}\left(\mathscr{L}(\bar{N}_\delta), \operatorname{Pois}\left(\bar{\lambda}_{p,n,\delta,\rho}\right) \right)
\leq & C_{n,\delta,\gamma} \left(\mu_{n,2\delta+2}\left(\bm{\Sigma}\right)\frac{\kappa}{p} +p^{-\frac{1}{\delta}}+E(p,\tau)\right) + \frac{5}{2}\frac{\gamma^\delta}{\delta !} \sum_{\ell=2}^{\delta+1}\alpha(\ell,r_\rho),
\end{align*}
where $E(p,\tau)$ is defined in Theorem \ref{thm:Poissonultrahigh}.
\end{prop}
\begin{comment}
Relying on the results on random geometric graphs in part \ref{item:randgeograconduppboua} and \ref{item:randgeograconduppboub}, Lemma \ref{lem:randgeograconduppbou} \ref{item:randgeograconduppbouc} establishes that the total variation distance between the increment distribution and Dirac measure at $1$ decays exponentially as $n$ increases and decays at rate $\delta^{-\frac{n-3}{2}}$ as $\delta$ increases. Thus when either $n$ or $\delta$ is large, and when the threshold $\rho$ satisfies the condition in part \ref{item:randgeograconduppbouc}, the non-asymptotic compound Poisson distribution in Theorem \ref{thm:Poissonultrahigh} is approximately a Poisson distribution. Lemma \ref{lem:randgeograconduppbou} also establishes that
$\sum_{\ell=2}^{\delta+1}\alpha(\ell,r_\rho)\approx 0$, which yields
\begin{equation}
\label{eqn:lambdaappfinitep}
\lambda_{p,n,\delta,\rho} \approx \binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta}.
\end{equation}
Thus
$
\operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho})\approx \operatorname{Pois}\left(\binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta}\right).$
Moreover, the second moment of $\operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho})$ in \eqref{eqn:secondmomentofCP} approximately equals
\begin{equation}
\label{eqn:2ndmomentapproximation}
\binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta} + \left( \binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta} \right)^2.
\end{equation}
Analogous to Lemma \ref{lem:randgeograconduppbou} \ref{item:randgeograconduppbouc}, one can obtain errors of the approximations in the two expressions \eqref{eqn:lambdaappfinitep} and \eqref{eqn:2ndmomentapproximation}. Details are omitted for brevity.
\end{comment}
We have thus established that the limiting compound Poisson distribution $\operatorname{CP}(\lambda_{n,\delta}(e_{n,\delta}), \bm{\zeta}_{n,\delta})$ in Theorem \ref{cor:Poissonlimit} can be approximated
by $\operatorname{Pois}\left(\frac{(e_{n,\delta})^\delta}{\delta!}\right)$, and that the non-asymptotic compound Poisson distribution $\operatorname{CP}(\lambda_{p,n,\delta,\rho}, \bm{\zeta}_{n,\delta,\rho})$ in Theorem \ref{thm:Poissonultrahigh} can be approximated by $\operatorname{Pois}(\binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta})$ for sufficiently large $n$ or $\delta$. By combining these results with Theorem \ref{thm:Poissonultrahigh} and Theorem \ref{cor:Poissonlimit} we then obtain Poisson approximations to $\mathscr{L}(\bar{N}_\delta)$. In Section \ref{sec:simulation} of the Appendix, Figure \ref{fig:Poissonfinitevslimit} provides numerical simulations to demonstrate the accuracy of the Poisson approximation to $\mathscr{L}(N_\delta)$. See Figure \ref{fig:Poissonfinitevslimit} and the ensuing discussions for more details.
\begin{comment}
\begin{rem}[Poisson approximations]\label{rem:poissonapproximation}
\marginpar{New remark}
We have established that the limiting compound Poisson distribution $\operatorname{CP}(\lambda_{n,\delta}(e_{n,\delta}), \bm{\zeta}_{n,\delta})$ in Theorem \ref{cor:Poissonlimit} can be approximately by $\operatorname{Pois}\left(\frac{(e_{n,\delta})^\delta}{\delta!}\right)$, and that the non-asymptotic compound Poisson distribution $\operatorname{CP}(\lambda_{p,n,\delta,\rho}, \bm{\zeta}_{n,\delta,\rho})$ in Theorem \ref{thm:Poissonultrahigh} can be approximated by $\operatorname{Pois}(\binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta})$ for large $n$ or $\delta$. \YW{By combining these results with Theorem \ref{thm:Poissonultrahigh} and Theorem \ref{cor:Poissonlimit} we then obtain Poisson approximation results. For instance, the Poisson approximation result for finite $p$ is similar Theorem \ref{thm:Poissonultrahigh}, but now the new upper bound has an additional error term
$$
d_{\operatorname{TV}}\left(\operatorname{CP}(\lambda_{p,n,\delta,\rho}, \bm{\zeta}_{n,\delta,\rho}), \operatorname{Pois}\left(\binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^{\delta}\right)\right),
$$
which does not vanish as $p\to \infty$ but is negligible for large $n$ or $\delta$. YW: should we add a formal Proposition? }
Figure \ref{fig:Poissonfinitevslimit} in the Supplementary Material (\cite{hero2020unified}, Appendix \ref{sec:simulation}) provides numerical simulations to demonstrate the effect of using Poisson distributions to approximate the distributions of the quantities in $\{N_{E_\delta}^{(\bm{\Psi})},N_{\breve{V}_\delta}^{(\bm{\Psi})},N_{V_\delta}^{(\bm{\Psi})}: \bm{\Psi} \in \{\bm{R},\bm{P}\} \}$. See Figure \ref{fig:Poissonfinitevslimit} and the ensuing discussions for more details. \hfill $\Box$
\end{rem}
\end{comment}
\subsection{Exact formulae for $\alpha_\ell$ and $\alpha(\ell,\rho)$ when $\delta=1$ and $\delta=2$}
\label{sec:exacttoalphaell}
In this subsection we provide analytical expressions for $\alpha_\ell$ and $\alpha(\ell,r_\rho)$ when $\delta=1$ and expressions for $\alpha_\ell$ when $\delta=2$.
\begin{exa}[$\alpha_\ell$ and $\alpha(\ell,r_\rho)$ when $\delta=1$]\label{exa:limitcompoidelta1}
When $\delta=1$, $\alpha_2=1$ since in the random geometric graph $\textbf{Ge}(\{\tilde{\bm{u}}_i\}_{i=1}^1; 1, n-2)$ the number of universal vertices (vertices of degree $0$) is $1$. In this case $\alpha_1=1-\alpha_2=0$. Similarly, $\alpha(2,r_\rho)=1$ when $\delta=1$ since in the random pseudo geometric graph $\textbf{PGe}(\{\bm{u}'_i\}_{i=1}^2; r_\rho, n-1)$ the number of universal vertices is $2$ as long as there exists one universal vertex. In this case $\alpha(1,r_\rho)=1-\alpha(2,r_\rho)=0$.
\end{exa}
\begin{rem}[Compound Poisson approximations when $\delta=1$]\label{rem:limitcompoidelta1}
Using the results for $\alpha_\ell$ in Example \ref{exa:limitcompoidelta1}, we obtained $\bm{\zeta}_{n,1}=\operatorname{Dirac}(2)$ and $\lambda_{n,1}(e_{n,1}) = \frac{1}{2} e_{n,1}$. Then the limiting compound Poisson distribution in Theorem \ref{cor:Poissonlimit} is $\operatorname{CP}(\frac{1}{2} e_{n,1}, \operatorname{Dirac}(2))$. On the other hand, by the results for $\alpha(\ell,r_\rho)$ in Example \ref{exa:limitcompoidelta1}, $\bm{\zeta}_{n,1,\rho}=\operatorname{Dirac}(2)$ and $\lambda_{p,n,1,\rho}=p(p-1)P_n(r_\rho)$. Therefore the non-asymptotic compound Poisson distribution in Theorem \ref{thm:Poissonultrahigh} is $\operatorname{CP}(p(p-1)P_n(r_\rho), \operatorname{Dirac}(2))$. The intuition for the result that the increment distributions $\bm{\zeta}_{n,1}$ and $\bm{\zeta}_{n,1,\rho}$ are $\operatorname{Dirac}(2)$ is as follows. Recall that $N_{E_1}^{(\bm{\Psi})}$ is twice of the number of edges. Thus $N_{E_1}^{(\bm{\Psi})}$ always increases by $2$ whenever there is a new edge and $N_{{\breve{V}}_1}^{(\bm{\Psi})}$ always increases by $2$ since the increment is always a new pair of vertices of degree $1$.
Note that $N_{V_1}^{(\bm{\Psi})}$ has increment close to $2$ since $N_{V_1}^{(\bm{\Psi})}\approx N_{{\breve{V}}_1}^{(\bm{\Psi})}$ by Lemma \ref{lem:6quantitiesine}. Given that the increment distributions $\bm{\zeta}_{n,1}$ and $\bm{\zeta}_{n,1,\rho}$ are $\operatorname{Dirac}(2)$, we have the following equivalent Poisson approximations to $\bar{N}_1/2$: $\mathscr{L}(\bar{N}_1/2)\approx \operatorname{Pois}(p(p-1)P_n(r_\rho))$ and $\bar{N}_1/2\overset{\mathcal{D}}{\to} \operatorname{Pois}(\frac{1}{2} e_{n,1})$ as $p\to \infty$ and $a_n 2^{\frac{n}{2}}p^{2}(1-\rho)^{\frac{n-2}{2}}\to e_{n,1}$.
As a comparison, Proposition 1 and its proof in \cite{hero2011large} under row-$\kappa$ sparsity condition established that $N_{E_1}^{(\bm{R})}/2$ converges to a Poisson distribution and obtained the limits of $\mathbb{E} N_{V_1}^{(\bm{R})}$ and $P(N_{V_1}^{(\bm{R})}>0)$. Proposition 1 and Proposition 3 in \cite{hero2012hub} under block sparsity condition extended the results in \cite{hero2011large} to the corresponding versions in the empirical partial correlation graph, i.e. the same conclusions hold with $\bm{R}$ replaced by $\bm{P}$. Our results in Theorems \ref{thm:Poissonultrahigh}, \ref{cor:Poissonlimit} (and Lemmas \ref{item:Poissonultrahigha}, \ref{item:Poissonlimita}) with $\delta=1$ characterize the full distributions of the $6$ quantities $\{N_{E_\delta}^{(\bm{\Psi})},N_{\breve{V}_\delta}^{(\bm{\Psi})},N_{V_\delta}^{(\bm{\Psi})}: \bm{\Psi} \in \{\bm{R},\bm{P}\} \}$, and our results in Section \ref{sec:convergenceofmoments} characterize their first and second moments, which together contain and extend the aforementioned previous results. Moreover, our results are established with weaker sparsity assumptions and provide concise formulae for the parameters.
\begin{comment}
As a comparison, Proposition 1 and its proof in \cite{hero2011large} under row-$\kappa$ sparsity condition establishes that $N_{E_1}^{(\bm{R})}/2$ converges to a $\operatorname{Pois}(\lambda_{n,\delta}(e_{n,\delta}))$ and
$\mathbb{E} N_{V_1}^{(\bm{R})}\to 2\lambda_{n,\delta}(e_{n,\delta})$ and $P(N_{V_1}^{(\bm{R})}>0)\to 1-e^{-\lambda_{n,\delta}(e_{n,\delta})}$. Proposition 1 and Proposition 3 in \cite{hero2012hub} under block sparsity condition extend the preceding result to the corresponding version in empirical partial correlation graphs, i.e. the same conclusions hold with $\bm{R}$ replaced by $\bm{P}$. Our results in Theorems \ref{thm:Poissonultrahigh}, \ref{cor:Poissonlimit} (and Lemmas \ref{item:Poissonultrahigha}, \ref{item:Poissonlimita}) with $\delta=1$ characterize the full distribution of the $6$ quantities $\{N_{E_\delta}^{(\bm{\Psi})},N_{\breve{V}_\delta}^{(\bm{\Psi})},N_{V_\delta}^{(\bm{\Psi})}: \bm{\Psi} \in \{\bm{R},\bm{P}\} \}$, and our results in Section \ref{sec:convergenceofmoments} characterize their first and second moments, which together contain the aforementioned previous results. Moreover, our results are established with weaker sparsity assumptions and have more concise formulae for the parameters. \hfill $\Box$
\end{comment}
\end{rem}
We next explicitly characterize $\alpha_\ell$ when $\delta=2$.
\begin{lem}[$\alpha_\ell$ when $\delta=2$]
\label{lem:alphaldelta2}
When $\delta=2$, $\alpha_2=0$, $\alpha_3 = \frac{3}{2} I_{\frac{3}{4}}(\frac{n-1}{2},\frac{1}{2})$ and $\alpha_1=1-\alpha_3$, where $I_x(a,b)$ is the regularized incomplete Beta function.
\end{lem}
\begin{rem}[Limiting compound Poisson approximation when $\delta=2$]\label{exa:limitcompoidelta2}
When $\delta=2$,
by Lemma \ref{lem:alphaldelta2}, $\alpha_2=0$, $\alpha_3 = \frac{3}{2} I_{\frac{3}{4}}(\frac{n-1}{2},\frac{1}{2})$ and $\alpha_1=1-\alpha_3$. Then $\sum_{\ell=1}^3\alpha_\ell/\ell=1- I_{\frac{3}{4}}(\frac{n-1}{2},\frac{1}{2})$.
Thus, the parameters for $\operatorname{CP}(\lambda_{n,2}(e_{n,2}),\bm{\zeta}_{n,2})$ in Theorem \ref{cor:Poissonlimit} are
\begin{equation}
\bm{\zeta}_{n,2}(1)=\frac{1-\frac{3}{2} I_{\frac{3}{4}}(\frac{n-1}{2},\frac{1}{2})}{1- I_{\frac{3}{4}}(\frac{n-1}{2},\frac{1}{2})}, \quad \bm{\zeta}_{n,2}(2)=0, \quad \bm{\zeta}_{n,2}(3)=\frac{\frac{1}{2} I_{\frac{3}{4}}(\frac{n-1}{2},\frac{1}{2})}{1- I_{\frac{3}{4}}(\frac{n-1}{2},\frac{1}{2})} \label{eqn:incrementdelta2}
\end{equation}
and
$$
\lambda_{n,2}(e_{n,2}) = \frac{1}{2}(e_{n,2})^2\left(1- I_{\frac{3}{4}}\left(\frac{n-1}{2},\frac{1}{2}\right)\right).
$$
Note that Proposition 1 in \cite{hero2012hub} states that for any fixed $n$ and $\delta$, $P(N_{V_{\delta}}^{(\bm{R})}>0)$ converges to $P(N_\delta^*>0)$ when $p\to\infty$, where $N_\delta^*$ is the Poisson random variable specified in the proposition. Since the increment distribution $\bm{\zeta_{n,2}} \neq \operatorname{Dirac}(1)$ by \eqref{eqn:incrementdelta2}, the limit of $P(N_{V_{\delta}}^{(\bm{R})}>0)$ is $P(\operatorname{CP}(\lambda_{n,2}(e_{n,2}),\bm{\zeta}_{n,2})>0)$ and hence
\cite[Proposition 1]{hero2012hub} is incorrect.
{ However, as shown in Subsection \ref{sec:approximationstoalphaell}
the result \cite[Proposition 1]{hero2012hub} is still useful when $n$ or $\delta$ is large.}
Specifically, the distribution of
each of the count variables $\{N_{E_\delta}^{(\bm{\Psi})},N_{\breve{V}_\delta}^{(\bm{\Psi})},N_{V_\delta}^{(\bm{\Psi})}:i\in\{\bm{R},\bm{P}\}, \bm{\Psi} \in\{\bm{R},\bm{P}\} \}$ can be accurately approximated by a Poisson distribution but with an additional error term depending on $n$ and $\delta$. This error term is small
for large $n$ or large $\delta$ (see Proposition \ref{prop:poiapplimit} and Proposition \ref{prop:poiappfinitep}).
\end{rem}
\begin{comment}
The above discussion rules out the Poisson limit, a degenerate Compound Poisson limit, since the increment distribution $\bm{\zeta_{n,2}} \neq \operatorname{Dirac}(1)$. This corrects the incorrect statement of Proposition 1 in \cite{hero2012hub} where it states for any fixed $n$ and $\delta$, $P(N_{V_{\delta}}^{(\bm{R})}>0)$ converges to $P(N_\delta^*>0)$ when $p\to\infty$, where $N_\delta^*$ is a Poisson random variable specified there. Note that in Subsection \ref{sec:approximationstoalphaell} we do show that the distribution of any quantity in $\{N_{E_\delta}^{(\bm{\Psi})},N_{\breve{V}_\delta}^{(\bm{\Psi})},N_{V_\delta}^{(\bm{\Psi})}:i\in\{\bm{R},\bm{P}\}, \bm{\Psi} \in\{\bm{R},\bm{P}\} \}$ can be approximated by Poisson distribution but with an additional error term depending on $n$ and $\delta$, which goes to $0$ for moderately large $n$ or large $\delta$, as shown in Figure \ref{fig:Poissonfinitevslimit} in the Supplementary Material (\cite{hero2020unified}, Appendix \ref{sec:simulation}).
\end{comment}
\begin{comment}
We note that Proposition 1 in \cite{hero2012hub} claims that for any fixed $n$ and $\delta$, $P(N_{V_{\delta}}^{(\bm{R})}>0)$ converges to $P(N_\delta^*>0)$ when $p\to\infty$, where $N_\delta^*$ is a Poisson random variable specified there. Our results in Theorem \ref{cor:Poissonlimit}, Lemma \ref{lem:alphaldelta2} and Remark \ref{exa:limitcompoidelta2} however proves that for $\delta=2$ their claim is incorrect since $N_{V_{\delta}}^{(\bm{R})}$ converges to $\operatorname{CP}(\lambda_{n,2}(e_{n,2}),\bm{\zeta}_{n,2})$, but not a Poisson. Indeed their incorrect conclusion is due to one error in their proof when showing that $N_{E_{\delta}}^{(\bm{R})}$ for any $\delta\geq 2$ converges to a Poisson random variable in distribution. Note that in Subsection \ref{sec:approximationstoalphaell} we do show that the distribution of $\{N_{E_\delta}^{(\bm{\Psi})},N_{\ring{V}_\delta}^{(\bm{\Psi})},N_{V_\delta}^{(\bm{\Psi})}:\bm{\Psi} \in\{\bm{R},\bm{P}\} \}$ can be approximated by a Poisson distribution with an additional error term depending on $n$ and $\delta$, which is close to $0$ for large $n$ or large $\delta$. In this aspect, their result is incorrect in the sense that it neglects the additional error term depending on $n$ and $\delta$.
\end{comment}
\begin{comment}
\begin{exa}[$\alpha(\ell,\rho)$ when $\delta=1$]
When $\delta=1$, $\alpha(2,r_\rho)=1$ since the number of vertices of the degree $1$ is $2$ provided there exists one such vertex. Then $\alpha(1,r_\rho)=0$.
\hfill $\Box$
\end{exa}
A remark analogous to Remark \ref{rem:limitcompoidelta1} can be made for the non-asymptotic compound Poisson in Theorem \ref{thm:Poissonultrahigh} and is omitted for brevity.
\end{comment}
\section{Old versions abstracts}\YW{by Prof. Rajaratnam}
\section{Introduction}
\label{sec:intro}
This paper considers the problem of identifying high correlations and partial correlations in modern ultra high dimensional setting. In particular we study the problem of screening $n$ identically distributed $p$-variate samples for variables that have high correlation or high partial correlation with at least one other variable when the sample size $n \leq C_0\ln p$ for some constant $C_0$.
In the screening framework one applies a threshold to the sample correlation matrix or the sample partial correlation matrix to detect variables with at least one significant correlation. The threshold serves to separate signal from noise. Correlation and partial correlation screening in ultra-high dimensions with few samples arises frequently in applications where the per-sample cost of collecting high dimensional data is much more costly than the per-variable cost.
For example, in genomic correlation screening the cost of high throughput RNAseq assays is decreasing faster than the cost of biological samples \cite{hero2016Large}. In such situations $p$ is much larger than $n$.
The ultra-high dimensional regime when $n\leq C_0\ln p$ is challenging since the number of samples is insufficient to guarantee reliability of commonly applied statistical methods. For example, one way to undertake partial correlation screening is to first estimate the population covariance matrix, then obtain the inverse, from which a partial correlation matrix can be estimated. However, to get a reliable estimate of a general covariance matrix, the number of samples $n$ must be at least
$\Omega(p)$ as shown in \cite[Section 5.4.3]{vershynin2012introduction}. Even if the covariance matrix has a special structure like sparsity, covariance estimation requires the number of samples be of order $\Omega(\ln p)$
The reader is referred to
\cite{dawid1993hyper, letac2007wishart, anandkumar2009detection, castro2014detection, tan2014learning, khare2015convex, firouzi2016two, dalal2017sparse, heydari2017quickest, tarzanagh2018estimation, engle2019large, cao2019posterior, li2020adaptable, ledoit2020analytical} and the references therein for related work in modern high dimensional covariance selection and estimation.
While estimating the correlation matrix or partial correlation matrix is challenging in ultra-high dimensions, recent work \cite{hero2011large,hero2012hub} has shown that it is possible to accurately test the number of highly (partial) correlated variables under a false positive probability constraint; in particular the probability that a variable is spuriously (partially) correlated with at least one other variable. While correlation screening finds variables that have a high marginal correlation with at least another variable, partial correlation screening identifies variables that have high conditional correlations with one other variable conditioned on the rest.
In \cite{hero2011large}, the ultra-high dimensional correlation screening problem is studied under a row-sparsity assumption on the population covariance matrix. A phase transition in the number of false positive correlations was characterized as a function of the correlation threshold and the true covariance. In the case of block sparse covariance, the critical phase transition threshold becomes independent of the true covariance. In \cite{hero2012hub} the partial correlation screening problem was studied, and similar phase transition results as in correlation screening \cite{hero2011large} were obtained under the block-sparse assumption on the population covariance matrix. The survey \cite{hero2015foundational} reviews the correlation and partial correlation screening problems.
A follow on work \cite{zhang2017spherical} of \cite{hero2011large} applies a similar framework to the spurious correlations problem and the low-rank detection problem. \ONE{The reader is referred to \cite{anandkumar2009detection,castro2014detection, tan2014learning, heydari2017quickest, tarzanagh2018estimation, arias2012detection, fan2018discoveries} and the references therein \ONE{for additional work related to high dimensional sample correlation matrices.}
}
Despite these advances in correlation and partial correlation screening, the screening framework proposed in \cite{hero2011large, hero2012hub} has some serious methodological, theoretical and practical shortcomings. For instance, results for partial correlation screening impose a
restrictive block sparsity assumption on the true underlying correlation matrix. The block sparsity condition in \cite{hero2012hub} requires that only a small group of the variables have correlation within the blocks and have no correlations with variables outside the block. This assumption is severely restrictive for cases where variables have correlations within a group and also correlations with variables outside their respective groups. Furthermore, expressions for false discovery probabilities in \cite{hero2011large, hero2012hub} require estimating dependence functionals. Estimating such functionals lead to computationally prohibitive non-parametric estimation, rendering the screening methodology disconnected from the very setting it was designed for.
In this paper we propose a unifying framework for correlation and partial correlation screening that delivers a practical and scalable variable selection
in the ultra-high dimensional regime. By making novel connections to random
geometric graphs \cite{penrose2003random}, we demonstrate that the distribution of the number of discoveries beyond a certain threshold is approximated by a compound Poisson distribution, with different parameters in the regimes when $p$ is finite and when $p$ approaches $\infty$. To the best of our knowledge, such characterization has not previously appeared in the literature.
Furthermore, our results are proved in greater generality by relaxing the block-sparse assumption to a new sparsity condition, defined as $(\tau,\kappa)$ sparsity in Section \ref{sec:taukappaspa}, on the population covariance matrix.
The block-sparse assumption is a special case of the $(\tau,\kappa)$ sparsity assumption. The characterizations established in this paper depend on the covariance matrix only through the $(\tau,\kappa)$ sparsity condition.
Moreover, the assumptions of $(\tau,\kappa)$ sparsity allows us to formulate unified theorems, which covers both the cases of correlation and of partial correlation screening.
The theory in this paper is directly relevant to hypothesis testing concerning the empirical degree distribution of a correlation graph. This topic arises in a wide spectrum of areas including graph mining, network science, social science, and natural sciences \cite{chakrabarti2006graph,kolaczyk2014statistical,kolaczykstatistical}.
Variables having strong sample correlations will appear in the correlation graph as vertices having positive vertex degree. As one sweeps over vertex degree values, the histogram of vertex degrees specifies the empirical degree distribution of the graph.
From this perspective, this paper provides compound Poisson characterizations of the empirical degree distribution for large correlation graphs under more realistic sparsity conditions on the population covariance. The expressions that are derived from our theorems also provide approximations to family-wise error rates associated with false discoveries of vertices of degree exceeding a specified fixed degree.
Finite sample results for controlling the probability of discovering a
(false) partial correlation in high-dimensional thresholded covariance
settings have been
elusive for the better part of the last decade and a half. Indeed,
evaluating expressions for such probabilities in the fixed $n$ setting
are known to be a notoriously difficult problem. This difficulty is in
part attributed to the dependence of such false
positive probabilities on the
unknown covariance parameter. Previous work has instead provided
expressions for the probability that two distinct connectivity
components of the partial correlation graph are falsely joined (see
\cite{banerjee2008model} and the references therein for more detail).
Controlling such probabilities implicitly assumes that the covariance
parameter is block diagonal. Such an assumption is tantamount to
requiring that the true partial correlation graph is not fully connected, a
restrictive assumption in many application areas. In contrast, the $(\tau,\kappa)$ sparsity condition introduced in this paper allows the underlying graph to be fully connected.
Moreover, we provide finite sample results for
controlling the probability that a (partial) correlation is falsely
discovered when the population correlation matrix is $(\tau, \kappa)$ sparse.
\ONE{
\subsection{Contribution}
\label{sec:contribution}
\ONE{We summarize the principal contributions of the paper. As above $p$ denotes the number of variables and $n$ denotes the number of samples.
\begin{enumerate}
\item
The paper presents a unified and complete asymptotic analysis of the star subgraph counts, the counts of vertices of a given degree and the counts of vertices above a given degree in the random graphs obtained by thresholding the sample correlation and sample partial correlation. This unification of different types of random counts represents an
improvement over previous work \cite{hero2011large,hero2012hub} where only
the counts of vertices above a given degree are studied.
\item
We approximate the full distributions of the random counts for finite $p$ and as $p\to\infty$. In addition, we characterize the first and second moments of these random counts. This is a
generalization of previous results \cite{hero2011large, hero2012hub} that only established approximations for the mean number of random counts and for the probabilities that these counts were positive.
\item
We obtain a compound Poisson characterization of the distributions of the random counts.
The compound Poisson limit and approximation are well approximated by the standard Poisson limit when $n$ is moderately large (Section \ref{sec:limitingcompoundPoisson}). This result corrects and refines the claim in \cite{hero2012hub} that erroneously asserted a Poisson limit.
\item
The theory in this paper is developed under a novel sparsity condition on the population dispersion matrix. This sparsity condition, called $(\tau, \kappa)$ sparsity in Sec. \ref{sec:taukappaspa}, is
weaker than previously assumed conditions, which makes our theory more broadly applicable. Specifically, while the block sparsity condition in previous work \cite{hero2012hub,fan2008sure,firouzi2016two}
imposes that correlation can only occur locally in small blocks of variables, the $(\tau,\kappa)$ sparsity condition relaxes this condition to more general global correlation patterning.
\end{enumerate}
Some of the broader implications of the technical contributions of this paper are described below.
Previous conditions on population partial correlation networks assume they are of lower dimension.
In particular, conditions such as block sparsity do not allow for completely connected partial correlation graphs which involve all $p$ variables. Such restrictive assumptions are difficult to validate and rule out many realistic population (partial) correlation structures. Overcoming this hurdle has been an open problem for several years. The newly introduced $(\tau,\kappa)$ sparsity condition on the population covariance matrix settles this longstanding problem by successfully allowing for completely connected partial correlation graphs over the entire set of features.
Historically, the literature on correlation estimation and graphical models has separated the treatments of covariance graph models and undirected graphical models (or inverse covariance graph models) \cite{cox2014multivariate}. Unifying the two classes of statistical models has been an open problem for the better part of almost 3 decades. While this separate treatment may be appropriate in low dimensional settings when there are few variables, it is not immediately obvious which of the two frameworks is appropriate for a given data set in modern ultra-high dimensional regimes. To our knowledge, the framework in this paper is the first in the graphical model or correlation graph estimation literature to propose methodology which brings both approaches under one umbrella.
The results in the paper also have
relevance to applications. Recall that our Poisson and compound Poisson expressions effectively describe the number of false discoveries and hence allow us to obtain results for the familywise error rate (FWER) or k-FWER, that is the probability of obtaining k or more false discoveries. Note that we can also obtain the marginal distributions of correlation
estimates, which in turn allow us to obtain expressions for p-values for testing correlation estimates. These marginal p-values allow us to establish FDR control too using either the Benjamini-Hochberg \cite{benjamini1995controlling} or Benjamini-Yakutelli procedures \cite{benjamini2001control}. In summary, the correlations screening framework is sufficiently rich that it allows us to undertake statistical error control in terms of FWER, k-FWER and FDR. This is one of the main strengths of our results: a rigorous inferential framework in the ultra-high dimensional setting.
}
}
\begin{comment}
\marginpar{to be updated}
\ONE{\aoh{Please incorporate the edits I made to the following in the response to reviewers. THIS HAS NOT YET BEEN DONE as of Sat evening.} In this subsection we summarize our main contributions, including correcting some erroneous results in the literature, proving stronger conclusions, weakening sparsity assumptions and making connections to random graph theory.
\begin{enumerate}
\item
We establish unifying results that the number of hubs above a threshold (and other related random counts) converge to a compound Poisson limit.
The compound Poisson limit corrects previous results \cite{hero2012hub} that erroneously claimed a Poisson limit. When the number of samples is small, the compound Poisson limit and approximation in our paper behave
differently from the Poisson limit and approximation obtained in \cite{hero2012hub}. However when the number of samples are relatively large, we show that the compound Poisson limit and approximation can be well approximated by the Poisson limit and approximation obtained in \cite{hero2012hub}. This clarifies when the Poisson limit and approximation can be used in practical correlation mining.
\item
We establish the compound Poisson limit and approximation for the full distributions of the number of hubs above a degree, which is a
generalization of previous results \cite{hero2011large, hero2012hub}, where only the means of hub discoveries and the probabilities of positive number of hub discoveries are obtained. Our theories additionally establish compound Poisson limit and approximation for the number of hubs of exactly a degree and the number of subgraphs isomorphic to a star graph. We additionally establish second moment approximation, which is not studied in the previous papers \cite{hero2011large, hero2012hub}.
\item
The theories in this paper are established under weaker sparsity assumption on population dispersion matrix, which makes the theory more broadly applicable. While the block sparsity conditions in previous work \cite{hero2012hub} only allows a small set of variables to be possibly correlated, our sparsity condition allows more correlations beyond only a small group of variables. To the best of our knowledge, our paper is the first work to have a weaker assumption than block sparsity assumption in the literature. Details are available at Section \ref{sec:taukappaspa}.
\item
Finally, we demonstrate novel connection to random geometric graphs that have not been previously explored in previous papers.
Details are available at Section \ref{sec:rangeo}. \aoh{As mentioned in the response to reviewer edits of this subsection, I don't think this rises to the level of the principal contributions in the previous 3 items since we provide no formal theorems on these connections}.
\end{enumerate}
}
\end{comment}
The remainder of this paper is organized as follows. Section \ref{sec:mainresult} outlines the framework and presents our main theorem which characterizes the compound Poisson approximations when $p$ approaches $\infty$. In Section \ref{mainres} an approximating theorem when $p$ is finite is presented, based on which the main theorem follows. Section \ref{sec:convergenceofmoments} covers convergence of moments. Section \ref{sec:limitingcompoundPoisson} provides explicit expressions for the parameters of the compound Poisson characterizations.
Notation and symbols used in this paper are collected in the Section \ref{sec:symbols} of the Appendix.
Most of the technical proofs and auxiliary results are given in the Appendix.
\begin{comment}
\noindent {\bf Notation}
$\|\cdot\|_2$ for a vector represents its Euclidean distance to the origin. $C$ and $c$ denotes positive universal constants that might defer from line to line.
$C$ and $c$ with subscripts are positive finite constants depending only on the parameter in their subscripts and may differ from line to line.
\end{comment}
\section{Comparisons and Contributions}
Contributions:
\begin{enumerate}
\item
This paper contains all theoretical results from "Large scale correlation screening" and "Hub screening in partial correlation graph". This paper additionally characterize the full distribution converges in distribution to a compound Poisson in a unified framework as in Theorem \ref{cor:Poissonlimit} when $p$ goes to infinity and Theorem \ref{thm:Poissonultrahigh} for finite but large $p$.
\item This paper take a geometric perspective, and make connections to random geometric graph, see Subsection \ref{sec:rangeo}. With this view, it is more natural to see the number of subgraphs isomorphic to a star graph is approximately compound Poisson (cf. Subsection \ref{sec:closenessnumedge}). The geometric results (Lemmas \ref{prop:rangeo}, \ref{prop:rangeolimit} and \ref{lem:randgeogra}) are also of independent interest. With the 3 geometric Lemmas we go from the finite p compound Poisson approximation (cf. Theorem \ref{thm:Poissonultrahigh}) to a limiting Poisson approximation (cf. Theorem \ref{cor:Poissonlimit}) and further show for reasonably large $n$ or $\delta$, the limiting compound Poisson is approximately a Poisson (cf. Section \ref{sec:limitingcompoundPoisson}).
\item This paper relaxes the sparsity assumption from block-$\tau$ sparsity to $(\tau,\kappa)$ sparsity for empirical partial correlation graph, which allows more correlation between variables.
\end{enumerate}
\noindent
Comparisons:
\begin{enumerate}
\item
This paper correct the result $P(\bar{N}_\delta >0 )$ for $\bar{N}_\delta\in \{N_{E_\delta}^{(k)},N_{\breve{V}_\delta}^{(k)},N_{V_\delta}^{(k)}: k\in \{\bm{R},\bm{P}\} \}$ in "Hub screening in partial correlation graph" since it should converges to the probability of same event for a compound Poisson. This paper also correct a convergence rate in "Large scale correlation screening" (not essential).
\item Comparing to the two previous papers, theorems in the current paper are strengthened by weakening the assumptions made in the previous papers. On one hand, all the assumptions are on the model parameter $\bm{\Sigma}$ but not on some intermediate quantity show up in the proof. On the other hand, we relax some of the assumptions in the previous paper, like "the (conditional) joint density of $\bm{U}$-score are bounded"
\begin{comment}
\item (too detail. just for our own knowledge) This paper simplifies the assumptions. See the following tables for comparison of assumptions between our current paper and "Large scale correlation screening" in empirical correlation graph. As discussed, the matrix-elliptically contoured distribution assumption is to guarantee the fact "if $\bm{\Sigma}_{ij}=0$ then $\bm{u}_i$ and $\bm{u}_j$ are independent", which is used in "Large scale correlation screening" but not well justified. So the matrix elliptically contour assumption on data matrix is not really an additional assumption, and with this, we manage to drop many assumptions that is difficult to verify, as shown in the table. One additional assumption in this paper is "$\mu_{n,2\delta+2}(\bm{\Sigma})$ bounded by a constant independent of $p$", and this assumption can be easily verify by Lemma \ref{lem:sufass}. One goal for this paper is to only make assumption that can be easily verify and avoid making assumptions on intermediate quantities, joint density of $\bm{U}$-score for instance, and only make assumption on the initial parameters and model. Much efforts are put to achieve this goal.
The comparison between our current paper and "Hub screening paper in Partial Correlation Graph" for empirical partial correlation graph is similar and thus is omitted.
\\
\begin{tabular}{|c|c|c|}\hline
empirical correlation graph & large correlation paper & our current paper \\ \hline
matrix elliptically contour & \xmark & \cmark \\ \hline
row-$\kappa$ sparse & \cmark & \cmark \\ \hline
joint density of $\bm{U}$-scores multiplying by $|S^{n-2}|^p$ are bounded & \cmark & \xmark \\ \hline
derivative of joint density multiplying by $|S^{n-2}|^p$ are bounded & \cmark & \xmark \\ \hline
conditional density of $\bm{U}$-score multiplying by $|S^{n-2}|^p$ are bounded & \cmark & \xmark \\ \hline
$\mu_{n,2\delta+2}(\bm{\Sigma})$ bounded by a constant independent of $p$ & \xmark & \cmark \\ \hline
\end{tabular}
\end{comment}
\item Our model is random pseudo geometric graph on unit sphere. It's related to random geometric graph. In Penrose' book \cite{penrose2003random} in 2003 independent vertices is assumed, while our model allow some possible correlations between vertices (more comparisons are in Subsection \ref{sec:rangeo}). In the Example after Corollary 3.6 in \cite{penrose2003random}, the author makes comments on the number of vertices of degree $\delta$ is asymptotically a compound Poisson (but without characterizing the parameters of the limiting compound Poisson). Here in our paper we give a full characterization of the limiting compound Poisson distribution.
\end{enumerate}
\section{Old versions abstracts}\YW{by Prof. Rajaratnam}
\begin{abstract}
{\bf Proposed abstract}
An important problem in large scale inference is the identification of variables that have large correlation or partial correlation with at least one other variable. Recent work in correlation screening has yielded breakthroughs in the ultra-high dimensional setting when the sample size $n$ is fixed and the dimension $p \rightarrow \infty$ (see \cite{hero2012hub}). Despite these advances, the correlation screening framework suffers from some serious practical, methodological and theoretical deficiencies. For instance, theoretical safeguards for partial correlation screening requires that the population covariance matrix be block diagonal. This block sparsity assumption is however highly restrictive in numerous practical applications. As a second example, results for correlation and partial correlation screening framework requires the estimation of dependence measures or functionals, which can be highly prohibitive computationally, rendering the framework impractical and unappealing in the very setting it is designed for. In this paper, we propose a unifying approach to correlation and partial correlation screening which specifically allows for dependence in the underlying true correlation structure, thus yielding a methodology that is suitable for modern applications. Theoretical expressions for false discovery probabilities are obtained for finite $p$. The approach also obviates the need to estimate dependence measures rendering the framework readily scalable. The unifying framework also demonstrates an important duality between correlation and partial correlation screening with important theoretical and practical consequences.
\textbf{Contributions of this paper}:
\begin{enumerate}
\item Obtain the results on $(\tau,\kappa)$ sparsity, see Proposition \ref{thm:meannumdis}, \ref{thm:var} and Theorem \ref{falposratest}. $(\tau,\kappa)$ sparsity is a broader assumption than block sparsity assumption, see Section \ref{sec:taukappaspa} for detailed comparison.
\item have two sections (section \ref{sec:distancepreserving} and section 7) discussing why results for partial correlation screening is the same as correlation screening. Specifically, Proposition \ref{uyrelation} in Section \ref{sec:distancepreserving} shows the distance between columns of $Y$-score are preserved to associated columns of $\bm{U}$-score by coefficients depending on the condition number of a random matrix. Then Lemma \ref{proeig} in Section \ref{ranmatunisph} establishes for those type random matrix, the condition number is close to 1 with high probability.
\item have a new result (Theorem \ref{thm:var}) on the variance of number of discovery, which can be used to establish a non-trivial lower bound for discovery rate (Corollary \ref{cor:discoveryrate}) through 2nd moment method
\item
The approximations don't depend on the quantity $J$. The original results for correlation screening (Proposition 1 in \cite{hero2011large}) establish the results with dependence on $J$. In \cite{hero2012hub}, for both correlation and partial correlation screening, Proposition 1 and 2 establish results depending on $J$, then Proposition 3 establish that $J$ is close to 1 under block sparse condition. The current paper is similar the structure of \cite{hero2012hub}, but we just combine the fact that $J$ is close to 1 into the bounds so that the approximation is simpler, and is under the $(\tau,\kappa)$ sparsity. But in the proof, in fact, we actually establish results on $J$, and the fact that $J$ is close to $1$. So in appendix right before the proof instead of the main body, comes the definition of $J$.
\item The approximations are obtained for finite $p$. For example, Theorem \ref{falposratest} \ref{item:fprcor} establishes the approximation for discovery rate in correlation screening case for all $p$. While Theorem \ref{falposratest} \ref{item:fprpar} establish approximations for discovery rate in partial correlation screening case provided \eqref{eqn:plar2} holds. \eqref{eqn:plar2} is quantitative way of saying $p$ is large enough, as explained after Proposition \ref{thm:meannumdis}.
\end{enumerate}
\end{abstract}
\section{Introduction\YW{old versions by Prof. Rajaratnam}}
In this paper we consider the problem of screening $n$ independent and identically distributed $p$-variate samples for variables that have high correlation or high partial correlation with at least one other variable in the ultra-high dimensional regime when the sample size $n \leq C_0\ln p$.\footnote{Here $C_0$ is some universal constant satisfying $C_0\geq 1$. A ``universal constant" or ``absolute constant", is a constant that does not depend on any model parameter.} In the screening framework one applies a threshold to sample correlation matrix or the sample partial correlation matrix to detect variables with at least one significant correlation. This threshold aims to separate signal from noise. Correlation and partial correlation screening in ultra-high dimensions has become increasingly important in many modern applications as the per-sample cost of collecting high dimensional data is much more costly than per-variable cost. For example, in biomedical settings the cost of high throughput technology, like oligonucleotide gene microchips and RNAseq assays is decreasing, while the cost of biological samples isn't decreasing at the same rate \cite{hero201615}. In such situations $p$ is much larger than $n$.
The ultra-high dimensional regime when $n\leq C_0\ln p$ is very challenging since the number of samples is insufficient to apply many (if not most) reliable statistical methods. For example, one way to do the screening is firstly estimate the population covariance matrix, from which one estimates the inverse covariance, which is a proper estimation to the partial correlation matrix. However, to get a reliable estimate for general covariance matrix, the number of samples $n$ must be at least $O(p)$ as shown in Section 5.4.3. in \cite{vershynin2012introduction}. Even if the covariance matrix has special structure like sparsity, covariance estimation requires a number of samples of order $O(\ln p)$ \cite{rothman2008sparse}.
While estimating the covariance matrix or partial correlation matrix is challenging in ultra-high dimensions, it has been demonstrated in recent work that it is possible to accurately test the number of highly (partial) correlated variables under a false positive probability; in particular the probability that a variable is highly (partially) correlated with at least one other variable \cite{hero2011large,hero2012hub}. In \cite{hero2011large}, the ultra-high dimensional correlation screening problem is studied under a row-sparse assumption on the population covariance matrix. A phase transition in the number of false positive correlations was mathematically characterized as a function of the correlation threshold and the true covariance. In the case of block sparse covariance, the critical phase transition threshold becomes independent of the true covariance. While correlation screening finds the variables that have high marginal correlation with at least one other variable, partial correlation screening cares about the variables that have high conditional correlations with one other variable conditioned on the rest. In \cite{hero2012hub} the partial correlation screening problem was studied, and similar phase transition results as in correlation screening \cite{hero2011large} were obtained under the block-sparse assumption on the population covariance matrix. The survey \cite{hero2015foundational} reviews the correlation and partial correlation screening problem.
Despite these important advances in correlation and partial correlation screening, the screening framework proposed in \cite{hero2011large, hero2012hub} has some serious methodological, theoretical and practical shortcomings. For instance, results for partial correlation imposes a highly restrictive block sparsity assumption on the the true underlying correlation matrix. The block sparsity in \cite{hero2012hub} assumes only a small group of the variables are allowed to have correlation within the blocks and no correlations with variables outside the block. This assumption is severely restrictive for most modern applications since it is possible for variables to have correlations within a group and also correlations with variables outside their respective groups. Furthermore, expressions for false probabilities in \cite{hero2011large, hero2012hub} require estimating dependence functionals. Estimating such functionals lead to computationally prohibitive non-parametric estimation, rendering the screening methodology disconnected from the very setting it was designed for.
In this paper we propose a novel unifying framework for correlation and partial correlation screening that delivers a practical and scalable methodology in the ultra-high dimensional regime, which is simultaneously armed with theoretical safeguards. We derive results for the mean number of highly correlated or partial correlated variables, and quantify the discovery rate in \cite{hero2012hub} by relaxing block-sparse assumption to a weaker $(\tau,\kappa)$ sparsity assumption, defined in Section \ref{sec:taukappaspa}, on the population covariance matrix. The block-sparse assumption is a special case of the $(\tau,\kappa)$ sparsity assumption. Resulting approximations under this generalized covariance structure $(\tau,\kappa)$ also do not depend on dependence measures/functionals. Moreover, new results on the variance of the number of highly correlated or partial correlated variables are obtained in this paper, which in turn provide more information on the distribution (tail probability for instance) of the number of highly correlated or partial correlated variables, and provide a nontrivial bound on the discovery rate by second moment method. The results in this paper holds for both correlation and partial correlation screening. We show that the approximations of the mean and variance of the number of highly correlated variables, and false discovery probabilities, are the same as those quantities for partial correlation screening. This important duality naturally stems from new results relating the the score representation for correlation screening to that of partial correlation screening. The proofs of the generalized results in this paper are self-contained and are based on concentration of random matrices.
The remainder of this paper is organized as follows. We begin in Section \ref{Hubfra} by giving the framework. In Section \ref{sec:asspre} assumptions and preliminaries are presented. Section \ref{mainres} is devoted to the statement of our main result and discussion of its consequences. Section \ref{simulation} provide some numerical simulations to verify our results. Subsequent sections are devoted to proofs of the main results, including a key observation, the almost distance preserving property between $Y$-score for partial correlation screening and $\bm{U}$-score for correlation screening (Section \ref{sec:distancepreserving}), and new results on Random Matrix with columns distributed \emph {i.i.d} uniformly on the sphere (Section \ref{ranmatunisph}). A number of technical proofs and auxiliary results are given in the Appendix.
\section{Introduction\YW{Old Version by Yun}}
In this paper we consider the problem of screening $n$ independent and identically distributed $p$-variate samples for variables that have high correlation or partial correlation with at least one other variable in the ultra-high dimension regime $n \leq C_0\ln p$.\footnote{Here $C_0$ is some universal constant satisfying $C_0\geq 1$. Universal constant, or absolute constant, is a constant that doesn't depend on any parameter.} In the screening one applies a threshold to sample correlation matrix or the sample partial correlation matrix to detect variables with at least one significant correlation.
This problem of ultra dimension is important since in many applications the per-sample cost of collecting high dimensional data is much more costly than per-variable cost! For example, in biotech the cost of high throughput technology, like oligonucleotide gene microchips and RNAseq assays is decreasing, while the cost of biological samples isn't decreasing at the same rate \cite{hero201615}. In such a situation $p$ is much larger than $n$.
The ultra-high dimension regime $n\leq C_0\ln p$ is very challenging since the number of samples is insufficient to apply many (if not most) reliable statistical methods. For example, one way to do the screening is firstly estimate the population covariance matrix, from which one estimates the inverse covariance, which is a proper estimation to the partial correlation matrix. However, to get a reliable estimate for general covariance matrix, the number of samples $n$ must be at least $O(p)$ as shown in Section 5.4.3. in \cite{vershynin2012introduction}. Even if the covariance matrix has special structure like sparsity, the covariance estimation requires a number of samples of order $O(\ln p)$ \cite{rothman2008sparse}.
While estimating the covariance matrix or partial correlation matrix is challenging in ultra-high dimension, it is possible to accurately test the number of highly (partial) correlated variables under a false positive probability; specially and the probability of at least one highly (partial) correlated variable \cite{HerRajlar, HerRajhub}. In \cite{hero2011large}, the ultra-high dimensional correlation screening problem is studied under a row-sparse assumption on the population covariance matrix. A phase transition in the number of false positive correlations was mathematically characterized as a function of the correlation threshold and the true covariance. In the case of block sparse covariance, the critical phase transition threshold becomes independent of the true covariance. While correlation screening finds the variables that have high marginal correlation with at least one other variable, partial correlation screening cares about the variables that have high conditional correlations with one other variable conditioned on the rest. In \cite{hero2012hub} the partial correlation screening problem was studied, and similar phase transition results as in correlation screening \cite{hero2011large} was obtained under the block-sparse assumption on the population covariance matrix. The survey \cite{hero2015foundational} reviews the correlation and partial correlation screening problem.
This paper extends the results of mean number of highly correlated or partial correlated variables and the discovery rate in \cite{hero2012hub} by relaxing block-sparse assumption to a weaker $(\tau,\kappa)$ sparsity assumption, defined in Section \ref{sec:taukappaspa}, on the population covariance matrix. The block sparsity in \cite{hero2012hub} assumes only a small group of the variables are allowed to have correlation within the groups, and hence is restrictive since it's possible variables has both correlations within the group and correlations with a few variables outside the groups. The $(\tau,\kappa)$ sparsity assumption take those cases into consideration, and reduces to the block-sparse assumption as a special case. Moreover, an new result on approximation of the variance of number of highly correlated or partial correlated variables are obtained in this paper, which provide more information on the distribution (tail probability for instance) of the number of highly correlated or partial correlated variables, and provide a nontrivial bound on the discovery rate by second moment method.
The results in this paper holds for both correlation and partial correlation screening. We show that the approximations of mean, variance of number of highly correlated variables, and the probability that there exists some highly correlated pair are the same as those quantities for partial correlation screening. This fact come from the scores representation for correlation and partial correlation are almost the same. However, the results for partial correlation screening require the ultra dimensionality while it's not essential for correlation screening.
The remainder of this paper is organized as follows. We begin in Section \ref{Hubfra} by giving the framework. In Section \ref{sec:asspre} assumptions and preliminaries are presented. Section \ref{mainres} is devoted to the statement of our main result and discussion of its consequences. Section \ref{simulation} provide some numerical simulations to verify our results. Subsequent sections are devoted to proofs of the main results, including a key observation, the almost distance preserving property between $Y$-score for partial correlation screening and $\bm{U}$-score for correlation screening (Section \ref{sec:distancepreserving}), and new results on Random Matrix with columns distributed i.i.d. uniformly on the sphere (Section \ref{ranmatunisph}). A number of technical proofs and auxiliary results are given in the Appendix.
\section{Geometric Interpretation}
\subsection{Score representations of sample correlation and partial correlation}
\label{sec:scorerep}
In this subsection, the $\bm{U}$-score and the $Y$-score respectively for the sample correlation and the sample partial correlation are defined. These scores will serve as the vertices set on which the random geometric graphs are constructed in Subsection \ref{sec:rangeo}. We first present two useful canonical representations \myred{of} our model.
\myred{The first representation is to represent functions of $\bm{R}$ for $X$ from a vector elliptically contoured distribution $\mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$ as functions of $\bm{R}$ for $X$ from a centered matrix normal distribution $\mathcal{VE}(\bm{0},\bm{\Sigma},g_0)$. }
The following lemma is an immediate result of the theorem in Section 6 of \cite{anderson1992nonnormal}.
\begin{lem} \label{prop:anderinv}
Suppose $X \sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$. Then the distribution of $\bm{R}$ defined in \eqref{defR} is invariant to $g$ and $\bm{\mu}$.
\end{lem}
Since the quantities of interest in this paper are $\bar{N}_\delta$, functions of $\bm{R}$, their distribution are also invariant to $g$ and $\bm{\mu}$ when $X \sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$. Specifically, we may choose $ g(w)=g_0(w)=(2\pi)^{-\frac{n p}{2}}\exp(-\frac{1}{2}w)$ and $\bm{\mu} = \bm{0}$ such that $X$ follows the centered matrix normal distribution $\mathcal{VE}(\bm{0},\bm{\Sigma},g_0)$. That is, $\{x^{(i)}\}_{i=1}^{n}$, the rows of $X$, can be taken as i.i.d. samples from $\mathcal{N}(\bm{0},\bm{\Sigma})$, without changing the distribution of $\bar{N}_\delta$. \myred{Thus for the rest of the paper $X \sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$ implicitly means $\mathcal{VE}(\bm{0},\bm{\Sigma},g_0)$ by Lemma \ref{prop:anderinv} and hence we suppose $X$ follows the centered matrix normal distribution $\mathcal{VE}(\bm{0},\bm{\Sigma},g_0)$. As a consequence, $\Sigma_{ij}=0$ implies variable $i$ and variable $j$ are independent.} In particular, the different sparsity conditions discussed in Subsection \ref{sec:taukappaspa} imposed on $\bm{\Sigma}$ induces different independence structures between variables in the model.
The second representation represents the sample covariance matrix defined in \eqref{def:samcov} by a sample second moment of the projected data.
It is shown in Theorem 3.3.2 \cite{anderson2003introduction} that the sample covariance matrix of $n$ i.i.d. normal random vector can be represented as the sample second moment of $n-1$ i.i.d. zero-mean normal random vectors. Specifically, define
the orthogonal $n \times n$ matrix $\bm{H} = [n^{-\frac{1}{2}}\bm{1}, \bm{H}_{2:n}]^T$. The matrix $\bm{H}_{2:n}$ can be obtained by
Gramm-Schmidt orthogonalization and satisfies the properties
$$
\bm{1}^T\bm{H}_{2:n} = \bm{0}, \quad \bm{H}_{2:n}^T\bm{H}_{2:n} = \bm{I}_{n - 1}.
$$
Then
\begin{equation}
\tilde{X} = \bm{H}_{2:n} X \in \mathbb{R} ^{(n-1) \times p} \label{eqn:Xtildef}
\end{equation}
has i.i.d. rows $\{\tilde{x}^{(i)}\}_{i=1}^{n-1}$ drawn from $\mathcal{N}(\bm{0},\bm{\Sigma})$. Moreover the sample covariance matrix defined in \eqref{def:samcov}
\begin{equation}
\bm{S} = \frac{1}{n-1}\displaystyle\sum_{i=1}^{n-1}\tilde{x}^{(i)}(\tilde{x}^{(i)})^T=\frac{1}{n-1}\tilde{X}^T\tilde{X}. \label{eqn:sigdef}
\end{equation}
That is, the sample covariance matrix of the data matrix $X$, defined in \eqref{def:samcov}, is the same as the sample second moment of the data matrix $\tilde{X}$. It follows that $\Sigma_{ij}=0$ implies $\tilde{x}^{(i)}$ is independent of $\tilde{x}^{(j)}$.\\
\mypurple{ We now proceed to define and analyze the scores representation of $\bm{R}$ and $\bm{P}$. From the second representation above,}
$\{\tilde{x}^{(i)}\}_{i=1}^{n-1}\subset \mathbb{R}^p$, the rows of $\tilde{X}$, are i.i.d. copy from $\mathcal{N}(\bm{0},\bm{\Sigma})$. Let $\{\tilde{x}_i\}_{i=1}^{p}$ denote the columns of $\tilde{X}$. Then $\bm{u}_i\vcentcolon=\frac{\tilde{x}_i}{\|\tilde{x}_i\|_2}\in \mathbb{R}^{n-1}$ has distribution $\operatorname{{\bf unif}}(S^{n-2})$ since $\tilde{x}_i$ contains $n-1$ i.i.d. normal coordinates, where $\|\cdot\|_2$ for a vector represents its Euclidean distance to the origin. Denote
$$
\bm{U} = [\bm{u}_1, \bm{u}_2, \ldots, \bm{u}_p] \in \mathbb{R}^{(n -1) \times p}
$$
and it follows from \eqref{eqn:sigdef} and \eqref{defR} that
\begin{equation}
\bm{R}= \bm{U}^T\bm{U}. \label{Rfac}
\end{equation}
The $\bm{U}$ in equation \eqref{Rfac} is referred to as $\bm{U}$-score of the sample correlation matrix \cite{hero2011large, hero2012hub}. Note our formulation of $\bm{U}$-score is slightly different from the formulation in \cite{hero2011large, hero2012hub}, but it's an easy task to check the two different formulations are equivalent. \YW{Moreover, it follows that $\bm{u}_i$ and $\bm{u}_j$ are independent provided the population covariance $\Sigma_{ij}=0$. }
The normalized outer product of $\bm{U}$, defined by
\begin{equation}
\bm{B} =\frac{n-1}{p}\bm{U}\Uma^T\in \mathbb{R}^{(n-1)\times (n-1)} \label{eqn:Bdef}
\end{equation}
will play an important role in the analysis of empirical partial correlation graph.
\begin{lem}
\label{lem:Binv}
Let $X \sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$ and $p\geq n$. Then $\bm{B}$ is invertible with probability $1$.
\end{lem}
By Lemma 1 in \cite{hero2012hub}, provided $\bm{U}\Uma^T$ is invertible or, equivalently, $\bm{B}$ is invertible, \myred{the Moore-Penrose pseudo-inverse of $\bm{R}$ can be expressed as follows:}
\begin{equation}
\label{Rdagdef}
\bm{R}^\dagger = \bm{U}^T[\bm{U}\Uma^T]^{-2}\bm{U}=\left(\frac{p}{n-1}\right)^2\bm{U}^T\bm{B}^{-2}\bm{U}.
\end{equation}
It follows from Lemma \ref{lem:Binv} that equation \eqref{Rdagdef} holds $a.s.$. Define $\bm{A}=\bm{B}^{-1}$ , $\bar{Y}=\bm{A}\bm{U}$ and hence
\begin{equation}
\label{RdagYbar}
\bm{R}^\dagger = \left(\frac{p}{n-1}\right)^2 \bar{Y}^T\bar{Y} \quad a.s. .
\end{equation}
Further define
\begin{align*}
y_i &= \bar{y}_i/\|\bar{y}_i\|_2, \quad \forall\ i\in p, \\
Y & = [y_1,y_2,\ldots,y_p]\in \mathbb{R}^{(n-1)\times p},
\end{align*}
and thus
\begin{equation}
\bm{P} = Y^TY \quad a.s. \label{Pfac}
\end{equation}
by equation \eqref{Pdef} and \eqref{RdagYbar}. $Y$ in equation \eqref{Pfac} is referred to as the $Y$-score representations of the sample partial correlation matrix \cite{hero2012hub}. Similar to the $\bm{U}$-score, one can easily verify that our formulation of $Y$-score is equivalent to that in \cite{hero2012hub}.
Let $\sigma^{n-2}$ be the spherical measure on $S^{n-2}$, i.e. $\sigma^{n-2}$ is the probability measure corresponding to the uniform distribution on $S^{n-2}$. Denote by $f_{\bm{u}_{j_1},\bm{u}_{j_2},\cdots,\bm{u}_{j_m}}$ the joint density of $j_1$-th, $j_2$-th, $\ldots$, $j_m$-th column of $\bm{U}$ with respect to the product measure $\otimes^m \sigma^{n-2} := \underbrace{\sigma^{n-2}\otimes \sigma^{n-2} \otimes \cdots \otimes \sigma^{n-2}}_{m}$. The next lemma establishes that $f_{\bm{u}_{j_1},\bm{u}_{j_2},\cdots,\bm{u}_{j_m}}$ is bounded by $\mu_{n,m}(\bm{\Sigma})$. \myred{This highlights the role of $\mu_{n,m} (\bm{\Sigma})$ since the distribution of the three quantities $N_{\breve{V}_\delta}^{(\bm{R})}$ ,$N_{V_\delta}^{(\bm{R})}$, $N_{\EE_\delta}^{(\bm{R})}$, as discussed in Subsection \ref{sec:upperboundjointdensity}, have a local property in the sense that their distribution depends only on the joint density of $f_{\bm{u}_{j_1},\bm{u}_{j_2},\cdots,\bm{u}_{j_m}}$ for every possible collection of $\{j_i\}_{i=1}^m$ and some constant $m=2\delta+2$.}
\begin{lem} \label{lem:densitybou}
Let $X\sim \mathcal{VE}(\bm{\mu}, \bm{\Sigma}, g)$. Consider a set of distinct indexes $\mathcal{J}=\{j_i: 1\leq i \leq m\}\subset [p]$.
\begin{enumerate}[label=(\alph*)]
\item \label{item:densityboua}
$$
\mu_{n,m}(\bm{\Sigma}_{\mathcal{J}}) \leq \mu_{n,m} (\bm{\Sigma})
$$
\item \label{item:densityboub}
The joint density of any subset of $m$ columns of $\bm{U}$-score w.r.t. $\otimes^m \sigma^{n-1}$ is upper bounded by $\mu_{n,m} (\bm{\Sigma})$:
$$f_{\bm{u}_{j_1},\bm{u}_{j_2},\cdots,\bm{u}_{j_m}}(\bm{v}_1,\bm{v}_2,\cdots,\bm{v}_m) \leq \mu_{n,m}(\bm{\Sigma}_{\mathcal{J}}) \leq \mu_{n,m} (\bm{\Sigma}), \quad \forall \ \bm{v}_i\in S^{n-2}, \forall\ i\in [m]. $$
Here $\mu_{n,m}(\bm{\Sigma}_{\mathcal{J}})$ is well defined since $\mu_{n,m}(\cdot)$ is invariant to simultaneously applying a permutation to both rows and columns.
\item \label{item:densitybouc}
Let $h:\left(S^{n-2}\right)^m\to \mathbb{R}$ be a nonnegative Borel measurable function. Then
\[
\mathbb{E} h(\bm{u}_{j_1},\bm{u}_{j_2},\cdots,\bm{u}_{j_m})\leq \mu_{n,m} (\bm{\Sigma}_{\mathcal{J}}) \mathbb{E} h(\bm{u}_{j_1}',\bm{u}_{j_2}',\cdots,\bm{u}_{j_m}')\leq \mu_{n,m} (\bm{\Sigma}) \mathbb{E} h(\bm{u}_{j_1}',\bm{u}_{j_2}',\cdots,\bm{u}_{j_m}'),
\]
where $\{\bm{u}_{j_\ell}'\}_{\ell=1}^m$ are i.i.d. distributed as \operatorname{{\bf unif}}$(S^{n-2})$.
\end{enumerate}
\end{lem}
\begin{rem}
In Lemma \ref{lem:densitybou} \ref{item:densityboua}, since $\bm{\Sigma}_{\mathcal{J}}$ is an $m$ by $m$ symmetric positive definite matrix, by the definition in \eqref{eqn:munmA}, $$\mu_{n,m}(\bm{\Sigma}_{\mathcal{J}})=\begin{cases} [\mu(\bm{\Sigma}_{\mathcal{J}})]^{-\frac{n-1}{2}}, & \bm{\Sigma}_{\mathcal{J}} \text{ not diagonal, } \\
1, & \bm{\Sigma}_{\mathcal{J}} \text{ diagonal. }
\end{cases}$$
\hfill $\Box$
\end{rem}
The proof of Lemma \ref{lem:densitybou} \ref{item:densityboub} is deferred to Appendix \ref{sec:proofdensitybou}. Part \ref{item:densitybouc} immediately follows from part \ref{item:densityboub} by writing expectation as integrals. Part \ref{item:densityboua} follows trivially from $\mu_m(\bm{\Sigma}_{\mathcal{J}})=\mu(\bm{\Sigma}_{\mathcal{J}})\geq \mu_m(\bm{\Sigma})$.
Lemma \ref{lem:densitybou} \ref{item:densitybouc} is useful since when calculating expectation of nonnegative function of any $m$ columns of $\bm{U}$, one may always assume the associated columns $\{\bm{u}_j\}$ are independent $\operatorname{{\bf unif}}(S^{n-1})$ \YW{up to} an additional multiplicative factor $\mu_{n,m}(\bm{\Sigma})$.
\subsection{Random pseudo geometric graphs}
\label{sec:rangeo}
\mypurple{
In this subsection we define random pseudo geometric graphs and provide an equivalent formulation of our model, of which the vertex sets are the $\bm{U}$ and $Y$ scores defined in Subsection \ref{sec:scorerep}. We also define the increment distribution of the compound Poisson that non-asymptotically approximates the $\bar{N}_\delta$.
}
From equation \eqref{Rfac} and the fact that columns of $\bm{U}$ have Euclidean norm $1$,
\begin{equation}
R_{ij}=\bm{u}_i^T\bm{u}_j=1- \frac{\|\bm{u}_i-\bm{u}_j\|_2^2}{2} = \frac{\|\bm{u}_i+\bm{u}_j\|_2^2}{2} -1. \label{eqn:cortoycol}
\end{equation}
For a threshold $\rho\in [0,1)$, define $r_\rho:=\sqrt{2(1-\rho)} \in (0,\sqrt{2}]$. By equation \eqref{eqn:cortoycol},
$$
\{R_{ij} \geq \rho\} =\{\| \bm{u}_i-\bm{u}_j\|_2\leq \sqrt{2(1-\rho)} \}=\{ \|\bm{u}_i-\bm{u}_j\|_2 \leq r_\rho\}.
$$
and similarly,
$$
\{R_{ij} \leq -\rho\} = \{ \|\bm{u}_i+\bm{u}_j\|_2 \leq r_\rho\}.
$$
The two preceding expressions yield
\begin{equation}
\{ |R_{ij}| \geq \rho \} = \{\|\bm{u}_i+\bm{u}_j\|_2 \leq r_\rho\}\cup \{\|\bm{u}_i-\bm{u}_j\|_2 \leq r_\rho\}.
\label{eqn:corphiequzer}
\end{equation}
An analogous derivation yields the following for the empirical partial correlation graph,
\begin{equation}
\{ |P_{ij}| \geq \rho \} = \{\|y_i+y_j\|_2 \leq r_\rho\}\cup \{\|y_i-y_j\|_2 \leq r_\rho\}. \label{phiequzer}
\end{equation}
Based on \eqref{eqn:corphiequzer} and \eqref{phiequzer}, first presented in \cite{hero2011large,hero2012hub}, we now introduce some novel geometry connecting to random geometric graphs. Note \eqref{eqn:corphiequzer} indicates, $\{|R_{ij}|\geq \rho\}$, the event when sample correlation between $i$-th and $j$-th variables exceed the threshold $\rho$, or equivalently, the event when there exists a edge connecting the $i$-th and $j$-th vertices in the empirical correlation graphs $\mathcal{G}_{\rho}(\bm{\Psi}^{(\bm{R})})$, is the same as the event that $\bm{u}_i$ and $\bm{u}_j$, the associated $\bm{U}$-score, lie in some geometric set on $S^{n-2}\times S^{n-2}$. This insight provides an equivalent way to construct $\mathcal{G}_{\rho}(\bm{\Psi}^{(\bm{R})})$ through the $\bm{U}$-scores and similar interpretation can be drawn for \eqref{phiequzer}. Such equivalent construction is made formal in the next few paragraphs.
\begin{comment}
Recall given a set of points $\{ \bm{v}_i \}_{i=1}^m$ in $\mathbb{R}^{\mathscr{N}}$, denote by $\textbf{Ge}\left(\{\bm{v}_i \}_{i=1}^{m}, r ; m, \mathscr{N} \right)$ the geometric graph\index{geometric graph} with radius $r$, defined as follows. The vertex set of the graph is $\{ \bm{v}_i \}_{i=1}^m$, and there is an edge between $\bm{v}_i$ and $\bm{v}_j$ if $\|\bm{v}_i-\bm{v}_j\|_2 \leq r$.
Next, define a similar but different geometric graph, which is closely related to the empirical correlation and partial correlation graphs.
\end{comment}
\begin{definition}[Pseudo geometric graph\index{pseudo geometric graph}]
\label{def:pge}
Given $m\geq 2$ and a set of points $\{ \bm{v}_i \}_{i=1}^m$ in $\mathbb{R}^{\mathscr{N}}$, denote by $\textbf{PGe}\left(\{\bm{v}_i \}_{i=1}^m, r ; m, \mathscr{N} \right)$ the pseudo geometric graph with radius $r$, defined as follows. The vertex set of the graph is $\{ \bm{v}_i \}_{i=1}^m$, and there is an edge between $\bm{v}_i$ and $\bm{v}_j$ if $\textbf{dist}(\bm{v}_i,\bm{v}_j):= \min\left\{\|\bm{v}_i-\bm{v}_j\|_2,\|\bm{v}_i+\bm{v}_j\|_2 \right \} \leq r$.
\end{definition}
It's easy to verify that $\textbf{dist}(\cdot,\cdot)$ has the following properties: for $\forall \bm{v}_1,\bm{v}_2,\bm{v}_3\in \mathbb{R}^m$,
\begin{enumerate}
\item $\textbf{dist}(\bm{v}_1,\bm{v}_2)\geq 0$;
\item $\textbf{dist}(\bm{v}_1,\bm{v}_2) = 0$ if only if $\bm{v}_1 = \bm{v}_2$ or $\bm{v}_1 = - \bm{v}_2$;
\item $\textbf{dist}(\bm{v}_1,\bm{v}_2) = \textbf{dist}(\bm{v}_2,\bm{v}_1)$ and $\textbf{dist}(\bm{v}_1,\bm{v}_2) = \textbf{dist}(\bm{v}_1,-\bm{v}_2)$
\item $\textbf{dist}(\bm{v}_1,\bm{v}_2)\leq \textbf{dist}(\bm{v}_1,\bm{v}_3) + \textbf{dist}(\bm{v}_3,\bm{v}_2)$.
\end{enumerate}
That is, $\textbf{dist}(\cdot,\cdot)$ is a pseudo metric on $\mathbb{R}^\mathscr{N}$, which explains the name pseudo geometric graph in Definition \ref{def:pge}. Also, $\textbf{dist}(\cdot,\cdot)$ is indeed a metric on the quotient space of $\mathbb{R}^\mathscr{N}$ with any two points symmetric about origin identified.
If the set of points generating geometric graphs or pseudo geometric graphs are random, then the corresponding graphs are called random geometric graphs or random pseudo geometric graphs.
With the above definitions and by the discussions preceding Definition \ref{def:pge}, the empirical correlation graph $\mathcal{G}_{\rho}(\bm{\Psi}^{(\bm{R})})$ is isomorphic to $\textbf{PGe}\left(\{\bm{u}_i \}_{i=1}^p, r_\rho ; p, n-1 \right)$, the random pseudo geometric graph generated by $\bm{U}$-scores. Even though $\textbf{PGe}\left(\{\bm{u}_i \}_{i=1}^p, r_\rho ; p, n-1 \right)$ has additional geometric contents since each vertex in it is a specific point in $S^{n-2}$, it's not necessary to differentiate it from the empirical correlation graph as long as only the graph properties are of concerned. As an example, we may refer to $N_{V_\delta}^{(\bm{R})}$ the number of vertices with degree at least $\delta$ in $\textbf{PGe}\left(\{\bm{u}_i \}_{i=1}^p, r_\rho ; p, n-1 \right)$ as well. An entirely analogous analysis applies to empirical partial correlation graph and $\textbf{PGe}\left(\{y_i \}_{i=1}^p, r_\rho ; p, n-1 \right)$. This equivalent construction indicates that the distribution of each of the 3 quantities $\{N_{E_\delta}^{(k)},N_{\breve{V}_\delta}^{(k)},N_{V_\delta}^{(k)} \}$ with $k=\bm{R}$ ($k=\bm{P}$) depends only on the pairwise pseudo distances $\textbf{dist}(\cdot,\cdot)$ between columns of $\bm{U}$ ($Y$).
\myred{
The random pseudo geometric graphs analyzed above bear some similarities to the random geometric graphs in \cite{penrose2003random}.
In particular, the monograph \cite{penrose2003random} studies the number of \emph{induced} subgraphs isomorphic to a given graph, typical vertex degrees, and other graphical quantities of random geometric graphs. The difference between the random pseudo geometric graphs in this paper and the random geometric graphs in \cite{penrose2003random} is that our vertices $\bm{u}_i$ lie on the unit sphere instead of the whole Euclidean space and our distance is $\textbf{dist}(\cdot,\cdot)$ instead of Euclidean distance. Another key difference is that no independence among all vertices $\bm{u}_i$ are imposed in our model. Indeed in our model, the correlations between vertices $\bm{u}_i$ are encoded by a sparse matrix $\bm{\Sigma}$, where \cite{penrose2003random} impose the strong assumption that vertices of the random geometric graph are i.i.d.. In \cite{penrose2003random} an example concerning asymptotic compound Poisson limit for the number of vertices with degree at least $3$ was presented (cf. the Example after Corollary 3.6).
}
\begin{comment}
The differences between our model and the random geometric model are 1) our vertices $\bm{u}_i$ lie in the unit sphere instead of the whole Euclidean space; 2) the distance is $\textbf{dist}(\cdot,\cdot)$ instead of Euclidean distance; 3) no independence among all vertices $\bm{u}_i$ are imposed. Instead, their correlations are encoded by a sparse $\bm{\Sigma}$.
\end{comment}
Recall $\operatorname{{\bf NUV}}\left(\{\bm{v}_i \}_{i=1}^m, r ; m, \mathscr{N} \right)$ denotes the number of vertices of degree $m-1$ in the geometric graph $\textbf{Ge}\left(\{\bm{v}_i \}_{i=1}^m, r ; m, \mathscr{N} \right)$. When $m\geq 3$, $\operatorname{{\bf NUV}}\left(\{\bm{v}_i \}_{i=1}^m, r ; m, \mathscr{N} \right)$ is also the number of subgraphs of $\textbf{Ge}\left(\{\bm{v}_i \}_{i=1}^m, r ; m, \mathscr{N} \right)$ isomorphic to $\Gamma_{m-1}$. Define $\operatorname{ {\bf PNUV} }\left(\{\bm{v}_i \}_{i=1}^m, r ; m, \mathscr{N} \right)$ analogously for pseudo geometric graph.
Denote by $\text{deg}(\cdot)$ the degree of a given vertex in the graph. Consider $\{\bm{u}'_i\}_{i=1}^{\delta+1}\ \overset{\text{i.i.d.}}{\sim} \ \text{unif}(S^{n-2})$. Denote the conditional probability that there are $\ell$ vertices of degree $\delta$ in $\operatorname{ {\bf PNUV} }\left(\{\bm{u}'_i \}_{i=1}^{\delta+1}, r_\rho ; \delta+1, n-1 \right)$ by
\begin{equation}
\alpha_{n,\delta}(\ell,r_\rho) := P\left(\operatorname{ {\bf PNUV} }\left(\{\bm{u}'_i \}_{i=1}^{\delta+1}, r_\rho ; \delta+1, n-1 \right)=\ell|\text{deg}(\bm{u}'_1)=\delta\right),\quad \forall \ell\in [\delta+1]. \label{eqn:alphandeltarho}
\end{equation}
$\alpha_{n,\delta}(\ell,r_\rho)$ depends only on $n$, $\delta$ and the threshold $\rho$ and is abbreviated as $\alpha(\ell,r_\rho)$ when there is no confusion. Moreover, $\alpha(\ell,r_\rho)=0$ when $\ell \geq \delta+2$. Define a probability distribution $\bm{\zeta}_{n,\delta,\rho}$ supported on $[\delta+1]$ with
\begin{equation}
\bm{\zeta}_{n,\delta,\rho}(\ell) = \frac{\alpha(\ell,r_\rho)/\ell}{ \sum_{\ell=1}^{\delta+1} \left(\alpha(\ell,r_\rho)/\ell \right) }. \label{eqn:zetandelrhodef}
\end{equation}
It will be shown in the next three subsections that the probability distribution $\bm{\zeta}_{n,\delta,\rho}$ is the distribution of the increment of the compound Poisson that non-asymptotically approximates the $\mathscr{L}(\bar{N_\delta})$.
\section{Things To be done}
\begin{enumerate}
\item Simulation \YW{completed}
\begin{enumerate}
\item Validate the number of discovery for small sample is a compound Poisson but not a Poisson; this includes the corner cases $N=4,\delta=2$; and a family of curves for different $\delta$; and a graph of total variation between samples to the compound poisson (need to add a divergence to Poisson as a comparison); and a graph for $\delta=1$ to see the convergence rate for different $\delta$;
\item Comparison between result under different sparsity conditions on theorectical covariance matrix: $(\tau,\kappa)$ sparsity vs block sparsity vs diagonal matrix; vary the different sparsity level and see if there is a consequences
\item Use the result to test the sparsity structure of the theorectical covariance matrix in the hypothesis testing-> test on different dense matrix and see how the convergence behaves; vary the magnitude in a given type of dense matrix;
\item (might not be necessary for this paper) apply the result to some real data. Consider the familywise rate as in the hub screening paper.
\item (might not be necessary for this paper) closed form to approximate the phase transition for the compound Poisson result. or MCMC?
\end{enumerate}
\item Organize the paper \YW{we are here}
\item Introduction (look for possible new papers in the literature)
\end{enumerate}
\section{Compound Poisson characterizations for finite $p$}
\label{mainres}
In this section we establish a compound Poisson approximation for $\mathscr{L}(\bar{N}_\delta)$ for finite $p$, which is then used to prove the asymptotic Theorem \ref{cor:Poissonlimit}. In Subsection \ref{sec:scorerep} Z-score type representations are introduced, which are used in the subsequent development in Subsection \ref{sec:rangeo}. Subsection \ref{sec:rangeo} provides an equivalent formulation of the empirical correlation and partial correlation graphs in terms of random geometric graphs.
Against the backdrop of the first two subsections, Subsection \ref{sec:closenessnumedge} presents the compound Poisson approximation for star subgraph counts $N_{E_\delta}^{(\bm{R})}$. Subsection \ref{sec:allclose} demonstrates that all $6$ quantities in $\{N_{E_\delta}^{(\bm{\Psi})},N_{\breve{V}_\delta}^{(\bm{\Psi})},N_{V_\delta}^{(\bm{\Psi})}: k\in \{\bm{R},\bm{P}\} \}$ are close in $L^1$ distance. Combining results in Subsection \ref{sec:closenessnumedge} and Subsection \ref{sec:allclose}, a compound Poisson characterization for $\bar{N}_\delta$ for finite $p$ is obtained in Subsection \ref{sec:compoihighdim}, which is then used to deduce Theorem \ref{cor:Poissonlimit} in Subsection \ref{sec:proofoftheorem1}.
\subsection{Z-score type representations of sample correlation and partial correlation}
\label{sec:scorerep}
In this subsection, we define the U-scores and the Y-scores for the sample correlation and the sample partial correlation, respectively. These Z-type scores will serve as the vertex set on which random geometric graphs are constructed in Subsection \ref{sec:rangeo}.
The matrix of Z-scores $Z=[z_1,\ldots,z_p]\in \mathbb{R}^{n\times p}$ associated with the data matrix $X$ is defined by
$$
z_i := \frac{x_i-\bm{1}\bar{x}_i}{\sqrt{S_{ii}(n-1)}},
$$
where $\bar{x}_i$ is the $i$-th coordinate of the sample mean $\bar{x}$. The Z-scores specify the sample correlation matrix by the relation $\bm{R}= Z^\top Z$. Since $\bm{1}^\top z_i =0$ for every $i\in [p]$, we consider an equivalent but lower-dimensional type of Z-scores called the U-scores \cite{hero2011large, hero2012hub}, obtained by projecting each $z_i$ on the subspace $\{\bm{w}\in \mathbb{R}^n: \bm{1}^\top \bm{w} =0 \}$. Specifically, consider
an orthogonal $n \times n$ matrix $\bm{H} = [n^{-\frac{1}{2}}\bm{1}, \bm{H}_{2:n}]^\top$. The matrix $\bm{H}_{2:n}\in \mathbb{R}^{n\times(n-1)}$ can be obtained by
Gramm-Schmidt orthogonalization and satisfies the properties
$$
\bm{1}^\top\bm{H}_{2:n} := \bm{0}, \quad \bm{H}_{2:n}^\top\bm{H}_{2:n} = \bm{I}_{n - 1}.
$$
Define the matrix of U-scores as $\bm{U}=[\bm{u}_1,\ldots,\bm{u}_p]$ with $\bm{u}_i=\bm{H}_{2:n}^\top z_i\in \mathbb{R}^{n-1}$. It is clear that
\begin{equation}
\label{eqn:RUrelation}
\bm{R}=\bm{U}^\top \bm{U}.
\end{equation}
The normalized outer product of $\bm{U}$, defined by
\begin{equation}
\bm{B} :=\frac{n-1}{p}\bm{U}\Uma^\top\in \mathbb{R}^{(n-1)\times (n-1)} \label{eqn:Bdef}
\end{equation}
will play an important role in the analysis of the empirical partial correlation graph.
\begin{lem}
\label{lem:Binv}
Let $X \sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$ and $p\geq n$. Then $\bm{B}$ is invertible a.s.
\end{lem}
Proofs of all the lemmas and propositions in this paper can be found in the Appendix. By Lemma 1 in \cite{hero2012hub}, provided $\bm{B}$ is invertible, the Moore-Penrose pseudo-inverse of $\bm{R}$ can be expressed as follows:
\begin{equation}
\label{Rdagdef}
\bm{R}^\dagger = \bm{U}^\top[\bm{U}\Uma^\top]^{-2}\bm{U}=\left(\frac{p}{n-1}\right)^2\bm{U}^\top\bm{B}^{-2}\bm{U}.
\end{equation}
It follows from Lemma \ref{lem:Binv} that equation \eqref{Rdagdef} holds $a.s.$ Define
$\bar{Y}=\bm{B}^{-1}\bm{U}$ and observe that
\begin{equation}
\label{RdagYbar}
\bm{R}^\dagger = \left(\frac{p}{n-1}\right)^2 \bar{Y}^\top\bar{Y} \quad a.s.
\end{equation}
Further define $y_i := \bar{y}_i/\|\bar{y}_i\|_2$ for $i\in [p]$ and $Y := [y_1,y_2,\ldots,y_p]\in \mathbb{R}^{(n-1)\times p}$. Then
\begin{equation}
\bm{P} = Y^\topY \quad a.s. \label{Pfac}
\end{equation}
by equation \eqref{Pdef} and \eqref{RdagYbar}. $Y$ in \eqref{Pfac} is referred to as the Y-scores representation of the sample partial correlation matrix, as in \cite{hero2012hub}.
By \eqref{eqn:RUrelation} and \eqref{Pfac}, the set of U-scores and Y-scores summarize all the information about the sample correlations and partial correlations, and hence they capture all the information about the vertex counts $\bar{N}_\delta$. We will further elaborate on this point in Subsection \ref{sec:rangeo}. Observing that $Y$ is a function of $\bm{U}$, the distribution of $\bar{N}_\delta$ is uniquely determined by $\bm{U}$. The remainder of this subsection focuses on the distribution of the U-scores.
Denote $S^{n-2}$ the unit sphere in $\mathbb{R}^{n-1}$ and denote $\operatorname{{\bf unif}}(S^{n-2})$ the uniform distribution on $S^{n-2}$. Let $\{\tilde{x}^{(i)}\}_{i=1}^{n-1}\overset{\text{i.i.d.}}{\sim} \mathcal{N}(\bm{0},\bm{\Sigma})$ and
$$
\tilde{X}=[\tilde{x}^{(1)},\ldots,\tilde{x}^{(n-1)}]^\top =[\tilde{x}_1,\ldots,\tilde{x}_p] \in \mathbb{R}^{(n-1)\times p}.
$$
In the next lemma we characterize the distribution of $\bm{U}$ in terms of the distribution of $\tilde{X}$.
\begin{lem} \label{lem:Uscoresdistribution}
Assume $X \sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$.
\begin{enumerate}[label=(\alph*)]
\item \label{item:Uscoresdistributiona} The matrix $\bm{U}$ of U-scores has the same distribution as
$
\left[ \frac{\tilde{x}_1}{\|\tilde{x}_1\|_2}, \ldots, \frac{\tilde{x}_p}{\|\tilde{x}_p\|_2} \right] $.
\item \label{item:Uscoresdistributionb}
For each $i\in [p]$, $\bm{u}_i$ is distributed as $\operatorname{{\bf unif}}(S^{n-2})$. Moreover, if $\Sigma_{ij}=0$, then $\bm{u}_i$ and $\bm{u}_j$ are independent.
\end{enumerate}
\end{lem}
\begin{rem} \label{rem:Udisconsequence}
When $\Sigma_{ij}=0$, the $i$-th and the $j$-th columns of the data matrix are uncorrelated but are not necessarily independent. However Lemma \ref{lem:Uscoresdistribution} \ref{item:Uscoresdistributionb} establishes that in fact the corresponding U-scores $\bm{u}_i$ and $\bm{u}_j$ are independent. This is a consequence of the fact that the entire data matrix follows the vector elliptically contoured distribution $X \sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$. Another important implication of Lemma \ref{lem:Uscoresdistribution} is that the distribution of $\bm{U}$, hence the distribution of $\bar{N}_\delta$, is invariant to the mean $\bm{\mu}$ and the shaping function $g$.
\end{rem}
Let $\sigma^{n-2}$ be the spherical measure on $S^{n-2}$, that is $\sigma^{n-2}$ is the uniform probability measure corresponding to $\operatorname{{\bf unif}}(S^{n-2})$.
Denote by $f_{\bm{u}_{j_1},\bm{u}_{j_2},\cdots,\bm{u}_{j_m}}$ the joint density of the $j_1$-th, $j_2$-th, $\ldots$, $j_m$-th column of $\bm{U}$ with respect to the product measure $\otimes^m \sigma^{n-2} := \underbrace{\sigma^{n-2}\otimes \sigma^{n-2} \otimes \cdots \otimes \sigma^{n-2}}_{m}$. The next lemma establishes that $f_{\bm{u}_{j_1},\bm{u}_{j_2},\cdots,\bm{u}_{j_m}}$ is bounded by the inverse local normalized determinant $\mu_{n,m}(\bm{\Sigma})$.
\begin{lem} \label{lem:densitybou}
Assume $X\sim \mathcal{VE}(\bm{\mu}, \bm{\Sigma}, g)$. Consider a set of distinct indexes $\mathcal{J}=\{j_i: 1\leq i \leq m\}\subset [p]$.
\begin{enumerate}[label=(\alph*)]
\item \label{item:densityboua}
The inverse local normalized determinant satisfies
$$
\mu_{n,m}(\bm{\Sigma}_{\mathcal{J}}) \leq \mu_{n,m} (\bm{\Sigma}).
$$
\item \label{item:densityboub}
The joint density of any subset of $m$ columns of U-scores w.r.t. $\otimes^m \sigma^{n-1}$ is upper bounded by $\mu_{n,m} (\bm{\Sigma})$:
$$f_{\bm{u}_{j_1},\bm{u}_{j_2},\cdots,\bm{u}_{j_m}}(\bm{v}_1,\bm{v}_2,\cdots,\bm{v}_m) \leq \mu_{n,m}(\bm{\Sigma}_{\mathcal{J}}) \leq \mu_{n,m} (\bm{\Sigma}), \quad \forall \ \bm{v}_i\in S^{n-2}, \forall\ i\in [m]. $$
\item \label{item:densitybouc}
Let $h:\left(S^{n-2}\right)^m\to \mathbb{R}$ be a nonnegative Borel measurable function. Then
\begin{align*}
\mathbb{E} h(\bm{u}_{j_1},\bm{u}_{j_2},\cdots,\bm{u}_{j_m})
\leq & \mu_{n,m} (\bm{\Sigma}_{\mathcal{J}}) \mathbb{E} h(\bm{u}_{j_1}',\bm{u}_{j_2}',\cdots,\bm{u}_{j_m}') \\
\leq & \mu_{n,m} (\bm{\Sigma}) \mathbb{E} h(\bm{u}_{j_1}',\bm{u}_{j_2}',\cdots,\bm{u}_{j_m}'),
\end{align*}
where $\{\bm{u}_{j_\ell}'\}_{\ell=1}^m$ are i.i.d. uniformly distributed on $S^{n-2}$.
\end{enumerate}
\end{lem}
\begin{comment}
\begin{rem}
In Lemma \ref{lem:densitybou} \ref{item:densityboua}, since $\bm{\Sigma}_{\mathcal{J}}$ is an $m$ by $m$ symmetric positive definite matrix, by definition in \eqref{eqn:munmA}, $$\mu_{n,m}(\bm{\Sigma}_{\mathcal{J}})=\begin{cases} [\mu(\bm{\Sigma}_{\mathcal{J}})]^{-\frac{n-1}{2}}, & \bm{\Sigma}_{\mathcal{J}} \text{ not diagonal, } \\
1, & \bm{\Sigma}_{\mathcal{J}} \text{ diagonal. }
\end{cases}$$
\hfill $\Box$
\end{rem}
\end{comment}
According to Lemma \ref{lem:densitybou} \ref{item:densitybouc}, when calculating the expectation of nonnegative function of any $m$ columns of $\bm{U}$, one may always assume that the associated columns $\{\bm{u}_j\}$ are i.i.d. $\operatorname{{\bf unif}}(S^{n-1})$ up to an additional multiplicative factor $\mu_{n,m}(\bm{\Sigma})$.
\begin{comment}
We first present two useful canonical representations of our model.
The first representation characterizes the sample correlation matrix $\bm{R}$, with data matrix $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$, as an equivalent sample correlation matrix, with $X\sim\mathcal{VE}(\bm{0},\bm{\Sigma},g_0)$.
The following lemma is an immediate result of the theorem in Section 6 of \cite{anderson1992nonnormal}.
\begin{lem} \label{prop:anderinv}
Suppose $X \sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$. Then the distribution of $\bm{R}$ defined in \eqref{defR} is invariant to $g$ and $\bm{\mu}$.
\end{lem}
Since the quantities of interest in this paper are vertex counts $\bar{N}_\delta$, functions of the sample correlation matrix $\bm{R}$, their distributions are also invariant to $g$ and $\bm{\mu}$ when $X \sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$ by Lemma \ref{prop:anderinv}. As a consequence, without loss of generality, we may choose $ g(w)=g_0(w)=(2\pi)^{-\frac{n p}{2}}\exp(-\frac{1}{2}w)$ and $\bm{\mu} = \bm{0}$ such that $X$ follows the centered matrix normal distribution $\mathcal{VE}(\bm{0},\bm{\Sigma},g_0)$. That is, $\{x^{(i)}\}_{i=1}^{n}$, the rows of $X$, can be taken as i.i.d. samples from $\mathcal{N}(\bm{0},\bm{\Sigma})$, without changing the distribution of $\bar{N}_\delta$. In summary in the remainder of the paper without loss of generality we can suppose the data matrix $X$ follows the centered matrix normal distribution $\mathcal{VE}(\bm{0},\bm{\Sigma},g_0)$.
The second representation that we now present characterizes the sample covariance matrix defined in \eqref{def:samcov} by a sample second moment of a projected version of the data.
\YW{By the first representation discussed above, we suppose that $X\sim \mathcal{VE}(\bm{0},\bm{\Sigma},g_0)$.} It is shown in Theorem 3.3.2 \cite{anderson2003introduction} that the sample covariance matrix of $n$ i.i.d. normal random vector can be represented as the sample second moment of $n-1$ i.i.d. zero-mean normal random vectors. Specifically, define
the orthogonal $n \times n$ matrix $\bm{H} = [n^{-\frac{1}{2}}\bm{1}, \bm{H}_{2:n}]^\top$. The matrix $\bm{H}_{2:n}$ can be obtained by
Gramm-Schmidt orthogonalization and satisfies the properties
$$
\bm{1}^\top\bm{H}_{2:n} = \bm{0}, \quad \bm{H}_{2:n}^\top\bm{H}_{2:n} = \bm{I}_{n - 1}.
$$
Then
\begin{equation}
\tilde{X} = \bm{H}_{2:n} X \in \mathbb{R} ^{(n-1) \times p} \label{eqn:Xtildef}
\end{equation}
has i.i.d. rows $\{\tilde{x}^{(i)}\}_{i=1}^{n-1}$ drawn from $\mathcal{N}(\bm{0},\bm{\Sigma})$. Moreover the sample covariance matrix defined in \eqref{def:samcov}
\begin{equation}
\bm{S} = \frac{1}{n-1}\displaystyle\sum_{i=1}^{n-1}\tilde{x}^{(i)}(\tilde{x}^{(i)})^\top=\frac{1}{n-1}\tilde{X}^\top\tilde{X}. \label{eqn:sigdef}
\end{equation}
That is, the sample covariance matrix of the data matrix $X$, defined in \eqref{def:samcov}, is the same as the sample second moment of the projected data matrix $\tilde{X}$. \YW{We want to emphasize that the first and the second representations can be applied without changing the distribution of $\bm{R}$, hence distributions of $\bar{N}_\delta$.}
Then $\Sigma_{ij}=0$ implies variable $i$ and variable $j$ are independent. In particular, the different sparsity conditions discussed in Subsection \ref{sec:taukappaspa} and imposed on $\bm{\Sigma}$ induce different dependency structures between variables in the model.
It follows that $\Sigma_{ij}=0$ implies $\tilde{x}^{(i)}$ is independent of $\tilde{x}^{(j)}$.\\
We now proceed to define and analyze the Z-score type representation of the sample correlation $\bm{R}$ and partial correlation $\bm{P}$. From the second representation above, $\{\tilde{x}^{(i)}\}_{i=1}^{n-1}\subset \mathbb{R}^p$, the rows of $\tilde{X}$, are i.i.d. copies from $\mathcal{N}(\bm{0},\bm{\Sigma})$. Let $\{\tilde{x}_i\}_{i=1}^{p}$ denote the columns of $\tilde{X}$. Then $\bm{u}_i\vcentcolon=\frac{\tilde{x}_i}{\|\tilde{x}_i\|_2}\in \mathbb{R}^{n-1}$ has distribution $\operatorname{{\bf unif}}(S^{n-2})$ since $\tilde{x}_i$ contains $n-1$ i.i.d. normal coordinates, where $\|\cdot\|_2$ for a vector represents its Euclidean distance to the origin. Denote
$$
\bm{U} = [\bm{u}_1, \bm{u}_2, \ldots, \bm{u}_p] \in \mathbb{R}^{(n -1) \times p}
$$
and it follows from \eqref{eqn:sigdef} and \eqref{defR} that
\begin{equation}
\bm{R}= \bm{U}^\top\bm{U}. \label{Rfac}
\end{equation}
The $\bm{U}$ in equation \eqref{Rfac} is referred to as the U-scores of the sample correlation matrix \cite{hero2011large, hero2012hub}. Note our formulation of U-scores is slightly different from the formulation in \cite{hero2011large, hero2012hub}, but the two formulations are equivalent. Moreover, it follows that $\bm{u}_i$ and $\bm{u}_j$ are independent provided the population covariance $\Sigma_{ij}=0$.
The normalized outer product of $\bm{U}$, defined by
\begin{equation}
\bm{B} =\frac{n-1}{p}\bm{U}\Uma^\top\in \mathbb{R}^{(n-1)\times (n-1)} \label{eqn:Bdef}
\end{equation}
will play an important role in the analysis of empirical partial correlation graph.
\begin{lem}
\label{lem:Binv}
Let $X \sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$ and $p\geq n$. Then $\bm{B}$ is invertible a.s.
\end{lem}
By Lemma 1 in \cite{hero2012hub}, provided $\bm{U}\Uma^\top$ is invertible or, equivalently, $\bm{B}$ is invertible, the Moore-Penrose pseudo-inverse of $\bm{R}$ can be expressed as follows:
\begin{equation}
\label{Rdagdef}
\bm{R}^\dagger = \bm{U}^\top[\bm{U}\Uma^\top]^{-2}\bm{U}=\left(\frac{p}{n-1}\right)^2\bm{U}^\top\bm{B}^{-2}\bm{U}.
\end{equation}
It follows from Lemma \ref{lem:Binv} that equation \eqref{Rdagdef} holds $a.s.$ Define $\bm{A}=\bm{B}^{-1}$ , $\bar{Y}=\bm{A}\bm{U}$ and hence
\begin{equation}
\label{RdagYbar}
\bm{R}^\dagger = \left(\frac{p}{n-1}\right)^2 \bar{Y}^\top\bar{Y} \quad a.s.
\end{equation}
Further define
\begin{align*}
y_i &= \bar{y}_i/\|\bar{y}_i\|_2, \quad \forall\ i\in p, \\
Y & = [y_1,y_2,\ldots,y_p]\in \mathbb{R}^{(n-1)\times p},
\end{align*}
and thus
\begin{equation}
\bm{P} = Y^\topY \quad a.s. \label{Pfac}
\end{equation}
by equation \eqref{Pdef} and \eqref{RdagYbar}. $Y$ in equation \eqref{Pfac} is referred to as the Y-scores representations of the sample partial correlation matrix \cite{hero2012hub}. Similar to the U-scores, one can easily verify that our formulation of Y-scores is equivalent to that in \cite{hero2012hub}.
Let $\sigma^{n-2}$ be the spherical measure on $S^{n-2}$, i.e. $\sigma^{n-2}$ is the probability measure corresponding to the uniform distribution on $S^{n-2}$. Denote by $f_{\bm{u}_{j_1},\bm{u}_{j_2},\cdots,\bm{u}_{j_m}}$ the joint density of the $j_1$-th, $j_2$-th, $\ldots$, $j_m$-th column of $\bm{U}$ with respect to the product measure $\otimes^m \sigma^{n-2} := \underbrace{\sigma^{n-2}\otimes \sigma^{n-2} \otimes \cdots \otimes \sigma^{n-2}}_{m}$. The next lemma establishes that $f_{\bm{u}_{j_1},\bm{u}_{j_2},\cdots,\bm{u}_{j_m}}$ is bounded by $\mu_{n,m}(\bm{\Sigma})$. This highlights the role of $\mu_{n,m} (\bm{\Sigma})$ since the distribution of the three quantities $N_{\breve{V}_\delta}^{(\bm{R})}$ ,$N_{V_\delta}^{(\bm{R})}$, $N_{\EE_\delta}^{(\bm{R})}$, as discussed in Subsection \ref{sec:upperboundjointdensity}, have a local property in the sense that their distribution depends only on the joint density of $f_{\bm{u}_{j_1},\bm{u}_{j_2},\cdots,\bm{u}_{j_m}}$ for every possible collection of $\{j_i\}_{i=1}^m$ and some constant $m=2\delta+2$.
\begin{lem} \label{lem:densitybou}
Let $X\sim \mathcal{VE}(\bm{\mu}, \bm{\Sigma}, g)$. Consider a set of distinct indexes $\mathcal{J}=\{j_i: 1\leq i \leq m\}\subset [p]$.
\begin{enumerate}[label=(\alph*)]
\item \label{item:densityboua}
$$
\mu_{n,m}(\bm{\Sigma}_{\mathcal{J}}) \leq \mu_{n,m} (\bm{\Sigma})
$$
\item \label{item:densityboub}
The joint density of any subset of $m$ columns of U-scores w.r.t. $\otimes^m \sigma^{n-1}$ is upper bounded by $\mu_{n,m} (\bm{\Sigma})$:
$$f_{\bm{u}_{j_1},\bm{u}_{j_2},\cdots,\bm{u}_{j_m}}(\bm{v}_1,\bm{v}_2,\cdots,\bm{v}_m) \leq \mu_{n,m}(\bm{\Sigma}_{\mathcal{J}}) \leq \mu_{n,m} (\bm{\Sigma}), \quad \forall \ \bm{v}_i\in S^{n-2}, \forall\ i\in [m]. $$
Here $\mu_{n,m}(\bm{\Sigma}_{\mathcal{J}})$ is well defined since $\mu_{n,m}(\cdot)$ is invariant to simultaneously applying a permutation to both rows and columns of its argument.
\item \label{item:densitybouc}
Let $h:\left(S^{n-2}\right)^m\to \mathbb{R}$ be a nonnegative Borel measurable function. Then
\begin{align*}
\mathbb{E} h(\bm{u}_{j_1},\bm{u}_{j_2},\cdots,\bm{u}_{j_m})
\leq & \mu_{n,m} (\bm{\Sigma}_{\mathcal{J}}) \mathbb{E} h(\bm{u}_{j_1}',\bm{u}_{j_2}',\cdots,\bm{u}_{j_m}') \\
\leq & \mu_{n,m} (\bm{\Sigma}) \mathbb{E} h(\bm{u}_{j_1}',\bm{u}_{j_2}',\cdots,\bm{u}_{j_m}'),
\end{align*}
where $\{\bm{u}_{j_\ell}'\}_{\ell=1}^m$ are i.i.d. distributed as $\operatorname{{\bf unif}}(S^{n-2})$.
\end{enumerate}
\end{lem}
\begin{rem}
In Lemma \ref{lem:densitybou} \ref{item:densityboua}, since $\bm{\Sigma}_{\mathcal{J}}$ is an $m$ by $m$ symmetric positive definite matrix, by the definition in \eqref{eqn:munmA}, $$\mu_{n,m}(\bm{\Sigma}_{\mathcal{J}})=\begin{cases} [\mu(\bm{\Sigma}_{\mathcal{J}})]^{-\frac{n-1}{2}}, & \bm{\Sigma}_{\mathcal{J}} \text{ not diagonal, } \\
1, & \bm{\Sigma}_{\mathcal{J}} \text{ diagonal. }
\end{cases}$$
\hfill $\Box$
\end{rem}
The proof of Lemma \ref{lem:densitybou} \ref{item:densityboub} is deferred to the Supplementary Material (\cite{hero2020unified}, Appendix \ref{sec:proofdensitybou}). Part \ref{item:densitybouc} immediately follows from part \ref{item:densityboub} by writing expectations as integrals. Part \ref{item:densityboua} follows trivially from $\mu_m(\bm{\Sigma}_{\mathcal{J}})=\mu(\bm{\Sigma}_{\mathcal{J}})\geq \mu_m(\bm{\Sigma})$.
Lemma \ref{lem:densitybou} \ref{item:densitybouc} is useful since when calculating expectation of nonnegative function of any $m$ columns of $\bm{U}$, one may always assume that the associated columns $\{\bm{u}_j\}$ are independent $\operatorname{{\bf unif}}(S^{n-1})$ \YW{up to} an additional multiplicative factor $\mu_{n,m}(\bm{\Sigma})$.
\end{comment}
\subsection{Random pseudo geometric graphs}
\label{sec:rangeo}
In this subsection we define random pseudo geometric graphs and provide an equivalent formulation of the empirical correlation and partial correlation graphs in terms of these graphs, for which the vertex sets are, respectively, the U and Y scores. We also define the increment distribution of the compound Poisson that approximates $\bar{N}_\delta$ when $p$ is finite.
By the fact that $\bm{R} = \bm{U}^\top \bm{U}$ and the fact that columns of $\bm{U}$ have Euclidean norm $1$,
\begin{equation}
R_{ij}=\bm{u}_i^\top\bm{u}_j=1- \frac{\|\bm{u}_i-\bm{u}_j\|_2^2}{2} = \frac{\|\bm{u}_i+\bm{u}_j\|_2^2}{2} -1. \label{eqn:cortoycol}
\end{equation}
For a threshold $\rho\in [0,1)$, define $r_\rho:=\sqrt{2(1-\rho)} \in (0,\sqrt{2}]$.
\begin{comment}
By equation \eqref{eqn:cortoycol},
\begin{align*}
\{R_{ij} \geq \rho\} = & \{ \|\bm{u}_i-\bm{u}_j\|_2 \leq r_\rho\},\\
\{R_{ij} \leq -\rho\} = & \{ \|\bm{u}_i+\bm{u}_j\|_2 \leq r_\rho\}.
\end{align*}
\end{comment}
As shown in \cite{hero2011large,hero2012hub}, by \eqref{eqn:cortoycol},
\begin{equation}
\{ |R_{ij}| \geq \rho \} = \{\|\bm{u}_i+\bm{u}_j\|_2 \leq r_\rho\}\cup \{\|\bm{u}_i-\bm{u}_j\|_2 \leq r_\rho\}.
\label{eqn:corphiequzer}
\end{equation}
An analogous argument yields the following for the empirical partial correlation graph,
\begin{equation}
\{ |P_{ij}| \geq \rho \} = \{\|y_i+y_j\|_2 \leq r_\rho\}\cup \{\|y_i-y_j\|_2 \leq r_\rho\}. \label{phiequzer}
\end{equation}
Based on \eqref{eqn:corphiequzer}, we now introduce novel geometric connections between empirical correlation graphs and random geometric graphs. Recall $\{|R_{ij}|\geq \rho\}$ is the event that the magnitude of the sample correlation between the $i$-th and $j$-th variables exceeds the threshold $\rho$, or equivalently, the event that there exists a edge connecting the $i$-th and $j$-th vertices in the empirical correlation graph $\mathcal{G}_{\rho}(\bm{R})$. Equation \eqref{eqn:corphiequzer} indicates that $\{|R_{ij}|\geq \rho\}$ is the same as the event that the associated U-scores for the $i$-th and $j$-th variables, $\bm{u}_i$ and $\bm{u}_j$, lie in some geometric set on $S^{n-2}\times S^{n-2}$. This insight provides an equivalent way to construct $\mathcal{G}_{\rho}(\bm{R})$ through the U-scores. The empirical partial correlation graph $\mathcal{G}_{\rho}(\bm{P})$ may similarly be constructed through the Y-scores based on \eqref{phiequzer}. These remarks are formalized in the next few paragraphs.
\begin{comment}
Recall given a set of points $\{ \bm{v}_i \}_{i=1}^m$ in $\mathbb{R}^{\mathscr{N}}$, denote by $\textbf{Ge}\left(\{\bm{v}_i \}_{i=1}^{m}, r ; \mathscr{N} \right)$ the geometric graph\index{geometric graph} with radius $r$, defined as follows. The vertex set of the graph is $\{ \bm{v}_i \}_{i=1}^m$, and there is an edge between $\bm{v}_i$ and $\bm{v}_j$ if $\|\bm{v}_i-\bm{v}_j\|_2 \leq r$.
Next, define a similar but different geometric graph, which is closely related to the empirical correlation and partial correlation graphs.
\end{comment}
\begin{definition}[Pseudo geometric graph\index{pseudo geometric graph}]
\label{def:pge}
Given $m\geq 2$ and a set of points $\{ \bm{v}_i \}_{i=1}^m$ in $\mathbb{R}^{\mathscr{N}}$, denote by $\textbf{PGe}\left(\{\bm{v}_i \}_{i=1}^m, r ; \mathscr{N} \right)$ the pseudo geometric graph with radius $r$, defined as follows. The vertex set of the graph is $\{ \bm{v}_i \}_{i=1}^m$, and there is an edge between $\bm{v}_i$ and $\bm{v}_j$ if $\textbf{dist}(\bm{v}_i,\bm{v}_j):= \min\left\{\|\bm{v}_i-\bm{v}_j\|_2,\|\bm{v}_i+\bm{v}_j\|_2 \right \} \leq r$. The graph is called a random pseudo geometric graph when the vertices are random.
\end{definition}
It is easy to verify that $\textbf{dist}(\cdot,\cdot)$ has the following properties: for $\forall \bm{v}_1,\bm{v}_2,\bm{v}_3\in \mathbb{R}^m$,
\begin{enumerate}
\item $\textbf{dist}(\bm{v}_1,\bm{v}_2)\geq 0$;
\item $\textbf{dist}(\bm{v}_1,\bm{v}_2) = 0$ if only if $\bm{v}_1 = \bm{v}_2$ or $\bm{v}_1 = - \bm{v}_2$;
\item $\textbf{dist}(\bm{v}_1,\bm{v}_2) = \textbf{dist}(\bm{v}_2,\bm{v}_1)$ and $\textbf{dist}(\bm{v}_1,\bm{v}_2) = \textbf{dist}(\bm{v}_1,-\bm{v}_2)$
\item $\textbf{dist}(\bm{v}_1,\bm{v}_2)\leq \textbf{dist}(\bm{v}_1,\bm{v}_3) + \textbf{dist}(\bm{v}_3,\bm{v}_2)$.
\end{enumerate}
That is, $\textbf{dist}(\cdot,\cdot)$ is a pseudo metric on $\mathbb{R}^\mathscr{N}$, hence the name pseudo geometric graph in Definition \ref{def:pge}.
With the above definitions, and by the discussions preceding Definition \ref{def:pge}, the empirical correlation graph $\mathcal{G}_{\rho}(\bm{R})$ is isomorphic to $\textbf{PGe}\left(\{\bm{u}_i \}_{i=1}^p, r_\rho \right)$, the random pseudo geometric graph generated by U-scores. Consequently $\{N_{V_\delta}^{(\bm{R})},N_{\breve{V}_\delta}^{(\bm{R})},N_{E_\delta}^{(\bm{R})}\}$ are the numbers of vertices or subgraphs in $\textbf{PGe}\left(\{\bm{u}_i \}_{i=1}^p, r_\rho \right)$.
An analogous analysis applies to the empirical partial correlation graph and $\textbf{PGe}\left(\{y_i \}_{i=1}^p, r_\rho \right)$. This equivalent construction indicates that the distribution of each of the 3 quantities $\{N_{E_\delta}^{(\bm{\Psi})},N_{\breve{V}_\delta}^{(\bm{\Psi})},N_{V_\delta}^{(\bm{\Psi})} \}$ with $\bm{\Psi}=\bm{R}$ ($\bm{\Psi}=\bm{P}$) only depends on the pairwise pseudo distances $\textbf{dist}(\cdot,\cdot)$ between columns of $\bm{U}$ ($Y$).
The random pseudo geometric graph in Definition \ref{def:pge} \ONE{is similar} to the random geometric graph introduced in \cite{penrose2003random}.
In particular, studied in the monograph \cite{penrose2003random} is the number of \emph{induced} subgraphs isomorphic to a given graph, typical vertex degrees, and other graphical quantities of random geometric graphs.
\ONE{As a specific example, the rate $a_n2^{n/2}p^{1+\frac{1}{\delta}} (1-\rho)^{\frac{n-2}{2}}\to e(n, \delta)$ as $p\to \infty$ in Theorem \ref{cor:Poissonlimit} is equivalent to $2a_np^{1+\frac{1}{\delta}} r_\rho^{{n-2}}\to e(n, \delta)$, which is consistent with existing Poisson approximation for random geometric graph \cite[ Theorem 3.4]{penrose2003random}.
}
The differences between our random pseudo geometric graph in Definition \ref{def:pge} and the random geometric graphs defined in \cite{penrose2003random} are: 1) our graph has vertices $\bm{u}_i$ lying on the unit sphere instead of on the entire Euclidean space; 2) our graph is induced by distance $\textbf{dist}(\cdot,\cdot)$ instead of the Euclidean distance. Another key difference is that vertices $\bm{u}_i$ are not necessarily independent in Definition \ref{def:pge}. Indeed in our model, the correlations between vertices $\bm{u}_i$ are encoded by a sparse matrix $\bm{\Sigma}$ (cf. Lemma \ref{lem:Uscoresdistribution} \ref{item:Uscoresdistributiona}), whereas in \cite{penrose2003random} the vertices associated with the random geometric graph are assumed to be i.i.d. In \cite{penrose2003random} it was \ONE{stated (Example after \cite[Corollary 3.6]{penrose2003random}) without proof} that the number of vertices with degree at least $3$ was approximately compound Poisson. A similar compound Poisson limit is established in Theorem \ref{cor:Poissonlimit}. \ONE{There is recent work on testing whether a given graph is a realization of
an Erd\H{o}s–R\'enyi random graph or a realization of
a random geometric graph with vertices i.i.d. uniformly distributed on the sphere
\cite{bubeck2016testing}.
{ Therein, the authors study the asymptotics of a count statistic, the (signed) triangles, on a random geometric graph induced by thresholding the distances between $p$ i.i.d. uniformly distributed points on the $n$ dimensional sphere as $n$ goes to $\infty$. In contrast, this paper studies the large $p$ asymptotics of vertex degree counts for dependent non-uniform points on a sphere of fixed dimension $n$.}}
\begin{comment}
The differences between our model and the random geometric model are 1) our vertices $\bm{u}_i$ lie in the unit sphere instead of the whole Euclidean space; 2) the distance is $\textbf{dist}(\cdot,\cdot)$ instead of Euclidean distance; 3) no independence among all vertices $\bm{u}_i$ are imposed. Instead, their correlations are encoded by a sparse $\bm{\Sigma}$.
\end{comment}
Recall that $\operatorname{{\bf NUV}}\left(\{\bm{v}_i \}_{i=1}^m, r ; \mathscr{N} \right)$ denotes the number of universal vertices (vertices adjacent to all vertices except itself) in the geometric graph $\textbf{Ge}\left(\{\bm{v}_i \}_{i=1}^m, r ; \mathscr{N} \right)$.
Analogously, denote by $\operatorname{ {\bf PNUV} }\left(\{\bm{v}_i \}_{i=1}^m, r ; \mathscr{N} \right)$ the number of universal vertices in the pseudo geometric graph $\textbf{PGe}\left(\{\bm{v}_i \}_{i=1}^m, r ; \mathscr{N} \right)$.
Denote by $\operatorname{deg}(\cdot)$ the degree of a given vertex in the graph. Consider $\{\bm{u}'_i\}_{i=1}^{\delta+1}\ \overset{\text{i.i.d.}}{\sim} \ \operatorname{{\bf unif}}(S^{n-2})$. Denote the conditional probability that there are $\ell$ universal vertices in $\textbf{PGe}\left(\{\bm{u}'_i \}_{i=1}^{\delta+1}, r_\rho \right)$ by
\begin{equation}
\alpha_{n,\delta}(\ell,r_\rho) := P\left(\operatorname{ {\bf PNUV} }\left(\{\bm{u}'_i \}_{i=1}^{\delta+1}, r_\rho \right)=\ell|\operatorname{deg}(\bm{u}'_1)=\delta\right),\quad \forall \ell\in [\delta+1]. \label{eqn:alphandeltarho}
\end{equation}
The conditional probability $\alpha_{n,\delta}(\ell,r_\rho)$ depends on $n$, $\delta$ and the threshold $\rho$ and is abbreviated as $\alpha(\ell,r_\rho)$ when there is no risk of confusion.
Define a probability distribution $\bm{\zeta}_{n,\delta,\rho}$ supported on $[\delta+1]$ with
\begin{equation}
\bm{\zeta}_{n,\delta,\rho}(\ell) : = \frac{\alpha(\ell,r_\rho)/\ell}{ \sum_{\ell=1}^{\delta+1} \left(\alpha(\ell,r_\rho)/\ell \right) }. \label{eqn:zetandelrhodef}
\end{equation}
It will be shown in the next three subsections that the probability distribution $\bm{\zeta}_{n,\delta,\rho}$ is the increment distribution of the compound Poisson approximation to $\bar{N}_\delta$ for finite $p$.
\subsection{Closeness of the star subgraph counts
to a compound Poisson}
\label{sec:closenessnumedge}
\ONE{In this subsection we present a proposition that establishes an upper bound on the total variation between $\mathscr{L}\left(N_{E_\delta}^{(\bm{R})}\right)$ and \ONE{the} compound Poisson distribution.
We first introduce some notation.} Let $\operatorname{SC}(r,\bm{w})$ be the spherical cap with radius $r$ at the center $\bm{w}\in S^{n-2}$. Formally,
\begin{equation}
\operatorname{SC}(r,\bm{w}) = \{x\in S^{n-2}: \|x-\bm{w}\|_2 \leq r\}. \label{eqn:SCdef}
\end{equation}
Define $P_n(r):=\frac{{\text{Area}}(\operatorname{SC}(r,\bm{w}))}{\text{Area}(S^{n-2})}$, where $\text{Area}(\cdot)$ is the area of a subset of $S^{n-2}$. $P_n(r)$ is the normalized area of the spherical cap with radius $r$. As is shown in (2.6) in \cite{hero2011large},
\begin{equation}
P_n(r) = \frac{b_n}{2} \int_{1-\frac{r^2}{2}}^1 (1-u^2)^{\frac{n-4}{2}}du,\quad \text{when }r\in [0,\sqrt{2}], \label{spharefor}
\end{equation}
where $b_n = \frac{2\Gamma((n-1)/2)}{\sqrt{\pi}\Gamma((n-2)/2)}$. It follows by simple calculation that
$$
P_{n}(r)=1-\frac{{\text{Area}}(\operatorname{SC}(\sqrt{4-r^2},\bm{w}))}{\text{Area}(S^{n-2})}=1-P_n(\sqrt{4-r^2}) \quad \text{when} \sqrt{2}<r\leq 2,
$$
and $P_n(r)=1$ when $r>2$. Further properties of $P_n(r)$ are summarized in Lemma \ref{pderivative} in Section \ref{sec:auxlem} of the Appendix.
\ONE{Define the rate parameter
\begin{align}
\lambda_{p,n,\delta,\rho}:= \binom{p}{1}\binom{p-1}{\delta}(2P_n(r_\rho))^\delta\sum_{\ell=1}^{\delta+1} \frac{\alpha(\ell,r_\rho)}{\ell}. \label{eqn:lambdadeffinitep}
\end{align}
The next proposition \ONE{establishes that the distribution of $N_{\EE_\delta}^{(\bm{R})}$ is approximated by a compound Poisson distribution $\operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho})$ with rate parameter $\lambda_{p,n,\delta,\rho}$ and dispersion parameter $\bm{\zeta}_{n,\delta,\rho}$.}} Here $C$ and $c$, and their subscripted versions, denote generic positive constants that only depend on their subscripts.
\begin{prop}[Compound Poisson approximation for subgraph counts]
\label{prop:edgecor}
Let $p\geq n \geq 4$, $\delta\in [p-1]$ and $\gamma>0$ be given. Suppose $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$. Suppose $2p^{1+\frac{1}{\delta}}P_n(r_\rho)\leq \gamma$, and $\bm{\Sigma}$ is row-$\kappa$ sparse. Then
\begin{align*}
&d_{\operatorname{TV}}\left(\mathscr{L}\left(N_{E_\delta}^{(\bm{R})}\right), \operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho})\right) \\
\leq & C_{n,\delta,\gamma} \left(C'_{\delta,\gamma}\right)^{\mu_{n,\delta+1}\left(\bm{\Sigma}\right)\frac{\kappa-1}{p}}\left( \mu_{n,2\delta+2}\left(\bm{\Sigma}\right)\frac{\kappa}{p} \left(1 + \mu_{n,2\delta+2}\left(\bm{\Sigma}\right) \left(\frac{\kappa}{p}\right)^2 \right) + p^{-\frac{1}{\delta}} \right).
\end{align*}
where $C_{n,\delta,\gamma}$ and $C'_{\delta,\gamma}$ are two constants.
\end{prop}
\begin{rem} \label{rem:edgecor}
The condition $2p^{1+\frac{1}{\delta}}P_n(r_\rho)\leq \gamma$ specifies an implicit lower bound on the threshold $\rho$. To obtain an explicit lower bound, observe that the condition $2a_np^{1+\frac{1}{\delta}} \left(\sqrt{2(1-\rho)}\right)^{n-2} \leq \gamma $ is sufficient for $2 p^{1+\frac{1}{\delta}}P_n(r_\rho)\leq \gamma$, by Lemma \ref {pderivative} \ref{item:pderivativea} in Section \ref{sec:auxlem} of the Appendix. Solving for $\rho$, we then obtain
\begin{equation}
\label{eqn:rhononasymptotic}
\rho \geq 1 - \frac{1}{2} \left(\frac{\gamma}{2a_n p^{1+\frac{1}{\delta}}}\right)^{\frac{2}{n-2}}.
\end{equation}
The condition \eqref{eqn:rhononasymptotic} is a non-asymptotic version of \eqref{eqn:rhopformula}. \ONE{But the requirement on $\rho$ from $2p^{1+\frac{1}{\delta}}P_n(r_\rho)\leq \gamma$ is weaker than \eqref{eqn:rhononasymptotic} since \ONE{the bound} in Lemma \ref {pderivative} \ref{item:pderivativea} is not tight when $p$ is finite.}
%
For the upper bound in Proposition \ref{prop:edgecor} to be small, $p$ should be relatively large and, $\bm{\Sigma}$ should have relatively small $\mu_{n,2\delta+2}\left(\bm{\Sigma}\right)$, and $\bm{\Sigma}$ should be row-$\kappa$ sparse with relative small sparsity level $\kappa/p$.
In Lemma \ref{lem:sufass} in Section \ref{sec:localdeterminantvseigenvalues} of the Appendix, Lemma \ref{lem:sufass}) we bound $\mu_{n,2\delta+2}\left(\bm{\Sigma}\right)$ in terms of the condition number and eigenvalues of $\bm{\Sigma}$. In the special case when $\bm{\Sigma}$ is diagonal, $\mu_{n,2\delta+2}\left(\bm{\Sigma}\right) \kappa/p =1/p$.
%
In the case that
$\mu_{n,2\delta+2}\left(\bm{\Sigma}\right) \kappa/p$ is small, say $\mu_{n,2\delta+2}\left(\bm{\Sigma}\right) \kappa/p<1$, in the upper bound the term $\mu_{n,2\delta+2}\left(\bm{\Sigma}\right) \left(\kappa/p\right)^2$ can be dropped, resulting in an additional constant factor, since $\mu_{n,2\delta+2}\left(\bm{\Sigma}\right)\left(\kappa/p\right)^2<1$. In other words the effective upper bound is $\mu_{n,2\delta+2}\left(\bm{\Sigma}\right)\kappa/p+p^{-\frac{1}{\delta}}$, neglecting the coefficients depending on $n$, $\delta$ and $\gamma$.
%
Respective expressions for $C'_{\delta,\gamma}$ and $C_{n,\delta,\rho}$ in Proposition \ref{prop:edgecor} are presented in equations \eqref{eqn:C'delgamdef} and \eqref{eqn:Cndelgamdef} in the Appendix
\end{rem}
Proposition \ref{prop:edgecor} states that for given $n$, $p$, $\delta$ and $\gamma$, if the threshold $\rho$ is properly chosen, and $\bm{\Sigma}$ is row-$\kappa$ sparse and has small $\mu_{n,2\delta+2}(\bm{\Sigma})$, then the distribution of $N_{E_\delta}^{(\bm{R})}$ can be well approximated by the compound Poisson $\operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho})$. We will call $\operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho})$ the \emph{non-asymptotic compound Poisson distribution. }
\ONE{\ONE{In the Appendix we provide an informal argument (Section \ref{sec:derivationprop:edgecor}) to motivate Proposition \ref{prop:edgecor} in addition to a complete proof (Section \ref{sec:proofprop:edgecor})}. }
\subsection{A portmanteau result and bounds on pairwise total variations}
\label{sec:allclose}
In this subsection upper bounds for pairwise total variation distances and $L^1$ distances among $\{N_{E_\delta}^{(\bm{\Psi})},N_{\breve{V}_\delta}^{(\bm{\Psi})},N_{V_\delta}^{(\bm{\Psi})}: k\in \{\bm{R},\bm{P}\} \}$ are obtained.
\ONE{Some intuition is presented before stating the portmanteau result Proposition \ref{thm:allclose}.}
\begin{lem}
\label{lem:6quantitiesine}
Consider $\delta \in [p-2]$.
$$
N_{E_\delta}^{(\bm{R})} - (\delta+1)N_{E_{\delta+1}}^{(\bm{R})} \leq N_{\breve{V}_\delta}^{(\bm{R})} \leq N_{V_\delta}^{(\bm{R})} \leq N_{E_\delta}^{(\bm{R})},
$$
$$
N_{E_\delta}^{(\bm{P})} - (\delta+1)N_{E_{\delta+1}}^{(\bm{P})} \leq N_{\breve{V}_\delta}^{(\bm{P})} \leq N_{V_\delta}^{(\bm{P})} \leq N_{E_\delta}^{(\bm{P})}.
$$
\end{lem}
It follows directly from Lemma \ref{lem:6quantitiesine} that for $\tilde{N}_\delta \in \left\{ N_{\breve{V}_\delta}^{(\bm{R})}, N_{V_\delta}^{(\bm{R})} \right\} $,
\begin{equation}
\mathbb{E} \left|\tilde{N}_\delta - N_{E_\delta}^{(\bm{R})} \right| \leq (\delta+1) \mathbb{E} N_{E_{\delta+1}}^{(\bm{R})}. \label{eqn:corexaatupp}
\end{equation}
As a result, if $\mathbb{E} N_{E_{\delta+1}}^{(\bm{R})}$ is small, then $N_{\breve{V}_\delta}^{(\bm{R})}$ and $N_{V_\delta}^{(\bm{R})}$ are close to $N_{E_\delta}^{(\bm{R})}$ in $L^1$ norm.
To intuitively see why the vertex counts in the empirical partial correlation graph are close to those in the empirical correlation graph, consider large $p$ and suppose that $\bm{\Sigma}$ is diagonal. Then by Lemma \ref{lem:Uscoresdistribution} \ref{item:Uscoresdistributionb}, $\{\bm{u}_i\}_{i=1}^p$ are i.i.d. $\operatorname{{\bf unif}}(S^{n-2})$. According to the law of large numbers, for large $p$,
\begin{equation}
\bm{B} = \frac{n-1}{p}\sum_{i=1}^p\bm{u}_i\bm{u}_i^\top \approx (n-1)\mathbb{E}\bm{u}_i\bm{u}_i^\top = \bm{I}_{n-1}, \label{eqn:Bapprox}
\end{equation}
which implies that
\(
\bar{Y} = \bm{B}^{-1} \bm{U} \approx \bm{U}
\),
so that $\|\bar{y}_i\|_2\approx \|\bm{u}_i\|_2=1$, and thus
\begin{equation}
Y = \left[\frac{\bar{y}_1}{\|\bar{y}_1\|_2},\ldots,\frac{\bar{y}_p}{\|\bar{y}_p\|_2}\right] \approx \bar{Y} \approx \bm{U}. \label{eqn:UYclose}
\end{equation}
Recall in Subsection \ref{sec:rangeo} that $N_{E_\delta}^{(\bm{R})}$ and $N_{E_\delta}^{(\bm{P})}$ are the numbers of subgraphs isomorphic to $\Gamma_\delta$ associated with the random pseudo geometric graphs induced respectively by $\bm{U}$ and $Y$. Hence by \eqref{eqn:UYclose}, $N_{E_\delta}^{(\bm{R})} \approx N_{E_\delta}^{(\bm{P})}$.
Following the same reasoning, $N_{V_\delta}^{(\bm{R})}\approx N_{V_\delta}^{(\bm{P})}$ and \(N_{\breve{V}_\delta}^{(\bm{R})}\approx N_{\breve{V}_\delta}^{(\bm{P})}\). These informal arguments are formalized by Proposition \ref{thm:allclose}, which establishes that all $6$ quantities $\{N_{E_\delta}^{(\bm{\Psi})},N_{\breve{V}_\delta}^{(\bm{\Psi})},N_{V_\delta}^{(\bm{\Psi})}: k\in \{\bm{R},\bm{P}\} \}$ are mutually close in $L^1$ norm and their distributions are mutually close in total variation. Proposition \ref{thm:allclose} is therefore called a ``Portmanteau result''.
\begin{prop}[Portmanteau result] \label{thm:allclose}
{
Let $p\geq n \geq 4$ and $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$. Let $\delta \in [p-1]$. Suppose $2p^{1+\frac{1}{\delta}}P_n(r_\rho)\leq \gamma$.
\begin{enumerate}[label=(\alph*)]
\item \label{item:allclosea}
Suppose $\bm{\Sigma}$ is row-$\kappa$ sparse. Then for $\tilde{N}_\delta \in \left\{ N_{\breve{V}_\delta}^{(\bm{R})}, N_{V_\delta}^{(\bm{R})} \right\} $,
$$
d_{\operatorname{TV}}\left(\mathscr{L}(\tilde{N}_\delta ), \mathscr{L}\left(N_{\EE_\delta}^{(\bm{R})}\right) \right)
\leq
\mathbb{E} \left| \tilde{N}_\delta - N_{\EE_\delta}^{(\bm{R})} \right| \leq \frac{(\delta+1)^2}{\delta!}\gamma^{\delta+1}\left(1+\mu_{n,\delta+2}(\bm{\Sigma})\frac{\kappa-1}{p}\right)p^{-\frac{1}{\delta}}.$$
\item \label{item:allclosec}
Suppose $\bm{\Sigma}$, after some row-column permutation, is $(\tau,\kappa)$ sparse with $\tau\leq \frac{p}{2}$. Suppose the condition $\left(\sqrt{\frac{n-1}{p}}+\sqrt{\frac{\delta\ln p}{p}}\right) \leq c$ is satisfied for some positive and \ONE{sufficiently} small constant $c$. Then
\begin{align*}
d_{\operatorname{TV}}\left(\mathscr{L}\left(N_{\EE_\delta}^{(\bm{P})}\right),\mathscr{L}\left(N_{\EE_\delta}^{(\bm{R})}\right) \right) \leq & \mathbb{E}\left|N_{\EE_\delta}^{(\bm{P})}-N_{\EE_\delta}^{(\bm{R})}\right|\\
\leq & C_{E_\delta}^{(\bm{P})} \left(1+\frac{\kappa-1}{p}\mu_{n,\delta+1}(\bm{\Sigma})\right)\left(\sqrt{\frac{\ln p}{p}}+\frac{\tau}{p}\right), \numberthis \label{eqn:allclosec}
\end{align*}
where $C^{(\bm{P})}_{E_\delta} $ is a constant depending on only $n,\delta$ and $\gamma$.
\item \label{item:allclosed}
Suppose the same conditions as in part \ref{item:allclosec} hold. Then
\begin{align*}
d_{\operatorname{TV}}\left(\mathscr{L}\left(N_{\breve{V}_\delta}^{(\bm{P})}\right),\mathscr{L}\left(N_{\breve{V}_\delta}^{(\bm{R})}\right) \right)\leq & \mathbb{E}\left|N_{\breve{V}_\delta}^{(\bm{P})}-N_{\breve{V}_\delta}^{(\bm{R})}\right|\\
\leq & C_{\breve{V}_\delta}^{(\bm{P})} \left(1+\frac{\kappa-1}{p}\mu_{n,\delta+2}(\bm{\Sigma})\right)\left(\sqrt{\frac{\ln p }{p}}+\frac{\tau}{p}+p^{-\frac{1}{\delta}} \right), \numberthis \label{eqn:allclosed}
\end{align*}
where $C_{\breve{V}_\delta}^{(\bm{P})}$ is a constant depending on only $n,\delta$ and $\gamma$.
\end{enumerate}
}
\end{prop}
\begin{rem}
\label{rem:allclose}
In Proposition \ref{thm:allclose} \ref{item:allclosec} and \ref{item:allclosed}, the condition
\begin{equation}
\left(\sqrt{\frac{n-1}{p}}+\sqrt{\frac{\delta\ln p}{p}}\right) \leq c \label{eqn:plargendelta}
\end{equation}
explicitly specifies a lower bound on $p$. Specifically, observe that the left side of \eqref{eqn:plargendelta} is a decreasing function of $p$, and its limit is $0$ when $p\to\infty$. Thus the smallest positive integer value of $p$ satisfying \eqref{eqn:plargendelta} exists, denoted by $p_0$, which is a function \ONE{of $n, \delta$ and $c$}. \ONE{Therefore} \eqref{eqn:plargendelta} is equivalent to requiring $p\geq p_0$.
\ONE{The inequalities (\ref{eqn:allclosec}) and (\ref{eqn:allclosed}) are valid as long as $c$ is less than finite constant, given respectively in \eqref{eqn:cdef} and \eqref{eqn:cuppbou2} in the Appendix. }
%
The row-$\kappa$ sparsity condition on $\bm{\Sigma}$ guarantees that the vertex counts associated with the empirical correlation graph are close in $L^1$ norm to $N_{E_\delta}^{(\bm{R})}$ by Proposition \ref{thm:allclose} \ref{item:allclosea}. In Proposition \ref{thm:allclose} \ref{item:allclosec} and \ref{item:allclosed} the stronger condition of $(\tau,\kappa)$ sparsity suffices to establish that the vertex counts of $G_\rho(\bm{P})$ are close in $L^1$ norm to those of $G_\rho(\bm{R})$.
All $3$ upper bounds in the proposition are bounded by $ \left(1+\frac{\kappa-1}{p}\mu_{n,\delta+2}(\bm{\Sigma})\right)\left(\sqrt{\frac{\ln p }{p}}+\frac{\tau}{p}+p^{-\frac{1}{\delta}} \right)$, up to a multiplicative constant depending on only $n$, $\delta$ and $\gamma$. The interpretation of this upper bound is similar to the second paragraph in Remark \ref{rem:edgecor}.
%
%
%
%
%
Respective expressions for $C^{(\bm{P})}_{E_\delta}$ and $C_{\breve{V}_\delta}^{(\bm{P})}$ are presented in equations \eqref{eqn:CEatleadel} and \eqref{eqn:CVexcdel} in Appendix.
\end{rem}
\begin{figure}%
\labellist
\small \hair 2pt
\pinlabel {\tiny $N_{\breve{V}_\delta}^{(\bm{R})}$} at 83 602
\pinlabel {\tiny $N_{V_\delta}^{(\bm{R})}$} at 400 98
\pinlabel {\tiny $N_{E_\delta}^{(\bm{R})}$} at 427 1098
\pinlabel {\tiny $N_{E_\delta}^{(\bm{P})}$} at 980 1094
\pinlabel {\tiny $N_{\breve{V}_\delta}^{(\bm{P})}$} at 1287 555
\pinlabel {\tiny $N_{V_\delta}^{(\bm{P})}$} at 965 81
\pinlabel \rotatebox{0}{$(1+\frac{\kappa}{p}\mu_{n,\delta+1}(\bm{\Sigma}))(\sqrt{\frac{\ln p}{p}}+\frac{\tau}{p})$} at 711 1235
\pinlabel \rotatebox{0 }{$(1+\frac{\kappa}{p}\mu_{n,\delta+2}(\bm{\Sigma}))(\sqrt{\frac{\ln p }{p}}+\frac{\tau}{p}+p^{-\frac{1}{\delta}} )$} at 651 450
\pinlabel \rotatebox{0}{ {\footnotesize $(1+\frac{\kappa}{p}\mu_{n,\delta+2}(\bm{\Sigma}))(\frac{\sqrt{\ln p} }{\sqrt{p}}+\frac{\tau}{p}+p^{-\frac{1}{\delta}} )$ } } [l] at 1154 822
\pinlabel \rotatebox{0}{$(1+\mu_{n,\delta+2}(\bm{\Sigma})\frac{\kappa}{p})p^{-\frac{1}{\delta}}$} [r] at 150 880
\endlabellist
\centering
\includegraphics[width=0.45\linewidth]{graphs/relations1}%
\caption{
Illustration of the pairwise total variation distances between the vertex counts associated with the empirical correlation and partial correlation graph established by the portmanteau result Proposition \ref{thm:allclose}. The 4 solid edges correspond to existence of direct upper bounds on the total variation distances between two types of vertex counts, where the weights correspond respectively to the $4$ upper bounds (neglecting constant coefficients) in Proposition \ref{thm:allclose}. Dashed edges correspond to indirect upper bounds of the associated total variation distances obtained by applying the triangular inequalities, with weights computed from the solid path connecting the two vertices.
}
\label{fig:relations}
\end{figure}
\begin{comment}
\begin{figure}%
\labellist
\small \hair 2pt
\pinlabel $N_{\breve{V}_\delta}^{(\bm{R})}$ at 83 602
\pinlabel $N_{V_\delta}^{(\bm{R})}$ at 400 98
\pinlabel $N_{E_\delta}^{(\bm{R})}$ at 427 1098
\pinlabel $N_{E_\delta}^{(\bm{P})}$ at 980 1094
\pinlabel $N_{\breve{V}_\delta}^{(\bm{P})}$ at 1287 555
\pinlabel $N_{V_\delta}^{(\bm{P})}$ at 965 81
\pinlabel \rotatebox{0}{$\left(1+\frac{\kappa}{p}\mu_{n,\delta+1}(\bm{\Sigma})\right)\left(\sqrt{\frac{\ln p}{p}}+\frac{\tau}{p}\right)$} at 711 1250
\pinlabel \rotatebox{0 }{$\left(1+\frac{\kappa}{p}\mu_{n,\delta+2}(\bm{\Sigma})\right)\left(\sqrt{\frac{\ln p }{p}}+\frac{\tau}{p}+p^{-\frac{1}{\delta}} \right)$} at 651 480
\pinlabel \rotatebox{0 }{$(1+\frac{\kappa}{p}\mu_{n,\delta+2}(\bm{\Sigma}))(\sqrt{\frac{\ln p }{p}}+\frac{\tau}{p}+p^{-\frac{1}{\delta}} )$} [l] at 1194 722
\pinlabel \rotatebox{0}{$(1+\mu_{n,\delta+2}(\bm{\Sigma})\frac{\kappa}{p})p^{-\frac{1}{\delta}}$} [r] at 150 880
\endlabellist
\centering
\includegraphics[width=0.45\linewidth]{graphs/relations1}%
\caption{
This graph illustrates the pairwise total variation distances between the vertex counts associated with the empirical correlation and partial correlation graph. The 4 solid edges correspond to existence of an direct upper bound of the total variation between two vertices, where the weights correspond respectively to the $4$ upper bounds (neglecting constant coefficients) in Proposition \ref{thm:allclose}. Dashed edges correspond to indirect upper bounds of the total variations between vertices, with weights computed from the solid path connecting the two vertices.
}
\label{fig:relations}
\end{figure}
\end{comment}
Figure \ref{fig:relations} illustrates the relations established by the portmanteau result Proposition \ref{thm:allclose}.
Dashed edges correspond to indirect upper bounds of the total variation distances between vertices, with weights computed from solid path connecting the two vertices. For instance the weight of dash edge between $N_{E_\delta}^{(\bm{R})}$ and $N_{V_\delta}^{(\bm{P})}$ is computed from
\begin{align*}
&d_{\operatorname{TV}}\left(\mathscr{L}\left(N_{V_\delta}^{(\bm{P})}\right),\mathscr{L}\left(N_{E_\delta}^{(\bm{R})}\right) \right) \\
\leq & \mathbb{E} \left|N_{E_\delta}^{(\bm{R})}-N_{V_\delta}^{(\bm{P})}\right| \\
\leq & \mathbb{E} \left|N_{E_\delta}^{(\bm{R})}-N_{\breve{V}_\delta}^{(\bm{P})}\right|+\mathbb{E} \left|N_{E_\delta}^{(\bm{R})}-N_{E_\delta}^{(\bm{P})}\right| \\
\leq & \mathbb{E} \left|N_{E_\delta}^{(\bm{R})}-N_{\breve{V}_\delta}^{(\bm{R})}\right|+ \mathbb{E} \left|N_{\breve{V}_\delta}^{(\bm{R})}-N_{\breve{V}_\delta}^{(\bm{P})}\right| +\mathbb{E} \left|N_{E_\delta}^{(\bm{R})}-N_{E_\delta}^{(\bm{P})}\right|\\
\leq & \left(C_{E_\delta}^{(\bm{P})} + C_{\breve{V}_\delta}^{(\bm{P})}+\frac{(\delta+1)^2}{\delta!} \gamma^{\delta+1}\right) \left(1+\frac{\kappa-1}{p}\mu_{n,\delta+2}(\bm{\Sigma})\right)\left(\sqrt{\frac{\ln p }{p}}+\frac{\tau}{p}+p^{-\frac{1}{\delta}} \right),
\end{align*}
where the first step follows from Lemma \ref{prop:tvtomean} in Section \ref{sec:auxlem} in the Appendix, the second step follows from Lemma \ref{lem:6quantitiesine}, and the last step follows from Proposition \ref{thm:allclose}.
Figure \ref{fig:relations} graphically illustrates the implication of Proposition \ref{thm:allclose} that all $6$ quantities $\{N_{E_\delta}^{(\bm{\Psi})},N_{\breve{V}_\delta}^{(\bm{\Psi})},N_{V_\delta}^{(\bm{\Psi})}: k\in \{\bm{R},\bm{P}\} \}$ are mutually close in total variation.
As a result, the closeness of one quantity among the 6 to some distribution in total variation implies the closeness of all 6 quantities to that same distribution. By Proposition \ref{prop:edgecor}, $\mathscr{L}\left(N_{E_\delta}^{(\bm{R})}\right)$ is close to the non-asymptotic compound Poisson distribution $\operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho})$ in total variation, which implies all $6$ quantities $\{N_{E_\delta}^{(\bm{\Psi})},N_{\breve{V}_\delta}^{(\bm{\Psi})},N_{V_\delta}^{(\bm{\Psi})}: k\in \{\bm{R},\bm{P}\} \}$ are close in total variation to $\operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho})$. We state this formally as Theorem \ref{thm:Poissonultrahigh} in the next subsection.
\subsection{A unified theorem for finite $p$}
\label{sec:compoihighdim}
\begin{comment}
The following theorem is a non-asymptotic result that is used to prove our main theorem Theorem \ref{cor:Poissonlimit}. It states if the threshold $\rho$ is properly chosen, and $\bm{\Sigma}$ is row-$\kappa$ sparse, then any random quantity in $\{N_i^{(\bm{\Psi})}: \bm{\Psi}=\bm{R}, i\in \{E_\delta, \breve{V}_\delta, V_\delta \} \}$ is approximated by a compound Poisson distribution. If $\bm{\Sigma}$ satisfies $(\tau,\kappa)$ sparsity condition, then the distribution of $\bar{N}_\delta$ can be approximated by a compound Poisson distribution.
\end{comment}
The following theorem is a variant of Theorem \ref{cor:Poissonlimit} for finite $p$. It states that if the threshold $\rho$ is properly chosen, and $\bm{\Sigma}$ satisfies the $(\tau,\kappa)$ sparsity condition, then $\mathscr{L}(\bar{N}_\delta)$ can be approximated by a compound Poisson distribution.
\begin{thm}[Compound Poisson approximation for finite $p$] \label{thm:Poissonultrahigh}
Let $ n \geq 4$, $\delta\in [p-1]$, and $\gamma>0$ be given. Consider $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$.
Suppose $2p^{1+\frac{1}{\delta}}P_n(r_\rho)\leq \gamma$.
Suppose $\bm{\Sigma}$, after some row-column permutation, is $(\tau,\kappa)$ sparse with $\tau \leq \frac{p}{2}$ and $\mu_{n,2\delta+2}(\bm{\Sigma})\frac{\kappa}{p}<1$. Suppose the condition $\left(\sqrt{\frac{n-1}{p}}+\sqrt{\frac{\delta\ln p}{p}}\right) \leq c$ is satisfied for some positive and \ONE{sufficiently} small constant $c$.
Then $\bar{N}_\delta$, a generic random variable in the set $\{N_{E_\delta}^{(\bm{\Psi})},N_{\breve{V}_\delta}^{(\bm{\Psi})},N_{V_\delta}^{(\bm{\Psi})}: k\in \{\bm{R},\bm{P}\} \}$, satisfies:
\begin{equation}
d_{\operatorname{TV}}\left(\mathscr{L}\left(\bar{N}_\delta\right), \operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho}) \right) \leq C_{n,\delta,\gamma} \left(\mu_{n,2\delta+2}\left(\bm{\Sigma}\right)\frac{\kappa}{p} +p^{-\frac{1}{\delta}}+E(p,\tau)\right),
\label{eqn:mainthm2}
\end{equation}
where
$$
E(p,\tau) = \begin{cases}
0 & \text{if } \bar{N}_\delta = N_{E_\delta}^{(\bm{R})},\ N_{\breve{V}_\delta}^{(\bm{R})} \text{ or } N_{V_\delta}^{(\bm{R})}, \\
\sqrt{\frac{\ln p }{p}}+\frac{\tau}{p} & \text{if } \bar{N}_\delta = N_{E_\delta}^{(\bm{P})}, N_{\breve{V}_\delta}^{(\bm{P})} \text{ or } N_{V_\delta}^{(\bm{P})}.
\end{cases}
$$
\end{thm}
If only the vertex counts in the empirical correlation graph is of interest, then the $(\tau,\kappa)$ sparsity assumption can be relaxed to row-$\kappa$ sparsity.
\begin{lem}[Compound Poisson approximation in empirical correlation graph]
\label{item:Poissonultrahigha}
Let $ n \geq 4$, $\delta\in [p-1]$, and $\gamma>0$ be given. Consider $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$.
Suppose $2p^{1+\frac{1}{\delta}}P_n(r_\rho)\leq \gamma$.
Suppose $\bm{\Sigma}$ is row-$\kappa$ sparse with $\mu_{n,2\delta+2}(\bm{\Sigma})\frac{\kappa}{p}<1$. Then $\tilde{N}_\delta$, a generic random variable in the set $\{N_{E_\delta}^{(\bm{R})},N_{\breve{V}_\delta}^{(\bm{R})},N_{V_\delta}^{(\bm{R})} \}$, satisfies
\begin{equation}
d_{\operatorname{TV}}\left(\mathscr{L}\left(\tilde{N}_\delta\right), \operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho}) \right) \leq C_{n,\delta,\gamma} \left(\mu_{n,2\delta+2}\left(\bm{\Sigma}\right)\frac{\kappa}{p}+p^{-\frac{1}{\delta}}\right)
\label{eqn:mainthm}
\end{equation}
\end{lem}
\begin{rem}
\label{rem:nonasythm}
The assumption $\mu_{n,2\delta+2}(\bm{\Sigma})\frac{\kappa}{p}<1$ is only used to obtain simpler expressions for the upper bounds in \eqref{eqn:mainthm2} and \eqref{eqn:mainthm}. Without this assumption, similar inequalities hold with the upper bounds that are the sum the upper bounds in Proposition \ref{prop:edgecor} and Proposition \ref{thm:allclose}.
See the third paragraph in Remark \ref{rem:edgecor} for additional discussions.
%
\begin{comment}
Row-$\kappa$ sparsity on $\bm{\Sigma}$ suffices to guarantee \eqref{eqn:mainthm} holds, i.e. the vertex counts in the empirical correlation graph can be approximated by $\operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho}) $.
Neglecting the coefficient depending on only $n,\delta,$ and $\gamma$
the upper bound of \eqref{eqn:mainthm} is $ \left( \mu_{n,2\delta+2}(\bm{\Sigma})\frac{\kappa}{p}+ p^{-\frac{1}{\delta}} \right)$. Therefore, for the compound Poisson approximation in Lemma \ref{item:Poissonultrahigha} to be accurate, $p$ should be relatively large and, $\bm{\Sigma}$ should have small inverse local normalized determinant $\mu_{n,2\delta+2}(\bm{\Sigma})$ and small sparsity level $\frac{\kappa}{p}$. In the Supplementary Material (\cite{hero2020unified}, Lemma \ref{lem:sufass}) we provide conditions in terms of the condition number and eigenvalues of $\bm{\Sigma}$ to guarantee small $\mu_{n,\delta+2}\left(\bm{\Sigma}\right)$.
The stronger condition of $(\tau,\kappa)$ sparsity on $\bm{\Sigma}$ guarantees that \eqref{eqn:mainthm2} holds, i.e. the vertex counts in the empirical correlation and partial correlation graphs can be approximated by $\operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho}) $. For the compound Poisson approximation in Theorem \ref{thm:Poissonultrahigh} to be accurate, $p$ should be relatively large and, $\bm{\Sigma}$ should have small inverse local normalized determinant $\mu_{n,\delta+2}(\bm{\Sigma})$ and be $(\tau,\kappa)$ sparse with small sparsity level $\frac{\tau}{p}$, $\frac{\kappa}{p}$.
\end{comment}
Observe that
Theorem \ref{thm:Poissonultrahigh} and Lemma \ref{item:Poissonultrahigha} hold for any mean $\bm{\mu}$ and any shaping function $g$ when $X\sim \mathcal{VE}(\bm{\mu},\bm{\Sigma},g)$. This is a consequence of the invariance property of the U-scores distribution (see Remark \ref{rem:Udisconsequence}).
\ONE{In particular, none of the constants in Theorem \ref{thm:Poissonultrahigh} depend on $\bm{\mu}$ or $g$.}
\end{rem}
Theorem \ref{thm:Poissonultrahigh} and Lemma \ref{item:Poissonultrahigha} directly follow from Proposition \ref{prop:edgecor} and Proposition \ref{thm:allclose} and hence their proofs are omitted. Theorem \ref{thm:Poissonultrahigh} and Lemma \ref{item:Poissonultrahigha} provide an approximation for the family-wise error rate (FWER) \cite{hero2012hub}
$$
P(\bar{N}_\delta>0)\approx 1- e^{-\lambda_{p,n,\delta,\rho}}.
$$
We end this subsection by making a comparison between Theorem \ref{thm:Poissonultrahigh} and Theorem \ref{cor:Poissonlimit}.
Theorem \ref{thm:Poissonultrahigh} and Theorem \ref{cor:Poissonlimit} provide compound Poisson approximations to $\bar{N}_\delta$ respectively when $p$ is finite and when $p\to\infty$. By taking the limit as $p\to\infty$, we obtain simpler formulae for parameters of the approximation. Specifically the increment distribution $\bm{\zeta}_{n,\delta,\rho}$ of the non-asymptotic compound Poisson distribution in Theorem \ref{thm:Poissonultrahigh} depends on conditional probabilities in the random pseudo geometric graph as in \eqref{eqn:alphandeltarho}. On the other hand, the increment distribution $\bm{\zeta}_{n,\delta}$ of the limiting compound Poisson distribution in Theorem \ref{cor:Poissonlimit} depends on probabilities in the random geometric graph as in \eqref{eqn:alphaelldef}, which is relatively simpler. For instance, when $\delta=2$, an analytical formula for $\bm{\zeta}_{n,2}$ can be obtained (see Example \ref{exa:limitcompoidelta2}). Obtaining an analytical formula for $\bm{\zeta}_{n,2,\rho}$ does not seem straightforward.
\ONE{Despite the fact that the limiting compound Poisson distribution in Theorem \ref{cor:Poissonlimit} is relatively simpler than Theorem \ref{thm:Poissonultrahigh}, it has the disadvantage that the approximation is not accurate unless $p$ is large. Moreover,} Theorem \ref{thm:Poissonultrahigh} is stronger than Theorem \ref{cor:Poissonlimit} in the sense that Theorem \ref{thm:Poissonultrahigh} provides explicit upper bounds for the approximation errors, while Theorem \ref{cor:Poissonlimit} simply provides a limit but no convergence rates. \ONE{ \ONE{For further discussion see Remark \ref{rem:comparisonsfinitelimit} in the next subsection.}}
\begin{comment}
\begin{rem
\label{rem:comparisonsfinitelimit}
Despite the fact that the limiting compound Poisson distribution in Theorem \ref{cor:Poissonlimit} is relatively simpler than Theorem \ref{thm:Poissonultrahigh}, it has the disadvantage of requiring that $\rho\to 1$ at specific rate $p^{-\frac{2}{n-2}\left(1+\frac{1}{\delta}\right)}$ as discussed in Remark \ref{rem:remarkofthm1}. This particular rate, however, is very slow when $n$ is moderately large. Indeed if one chooses $\rho$ as in \eqref{eqn:rhopformula} and requires $\rho\geq 1-\epsilon$ for some $\epsilon\in (0,1/2)$, one obtains
$$
p \geq \left(\frac{e_{n,\delta}}{2a_n}\right)^{\frac{1}{1+1/\delta}} \left(\frac{1}{2\epsilon}\right)^{\frac{n-2}{2(1+1/\delta)}}.
$$
On the contrary, Theorem \ref{thm:Poissonultrahigh} does not require that $\rho$ approach $1$ and provides an approximation to $\bar{N}_{\delta}$ even for small $p$ \ONE{(see the first paragraph of Remark \ref{rem:edgecor} for a related discussion)}. The accuracy of these approximations are illustrated in Figure \ref{fig:finiteplimitcomparison} in the Supplementary Material (\cite{hero2020unified}, Appendix \ref{sec:simulation}).
\hfill $\Box$
\end{rem}
\end{comment}
\subsection{Proof of Theorem \ref{cor:Poissonlimit}}
\label{sec:proofoftheorem1}
In this subsection we present results on the limit of $\operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho})$ as $p\to\infty, \rho\to 1$, followed by a proof of Theorem \ref{cor:Poissonlimit}.
To study the limiting distribution of $\operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho})$, which requires the limit of the parameter $\alpha(\ell,r_\rho)$, the next two lemmas are useful.
\begin{lem}\label{prop:rangeo}
Consider $r < 2/\sqrt{5}$ and $\delta\geq 1$. Suppose $\{\bm{u}'_i\}_{i=1}^{{\delta+1}}\ \overset{\text{i.i.d.}}{\sim} \ \operatorname{{\bf unif}}(S^{n-2})$. Then for any $\ell \in [{\delta+1}]$,
\begin{align*}
&P\left(\operatorname{{\bf NUV}}\left(\{\bm{u}'_i \}_{i=1}^{\delta+1}, r \right)=\ell|\operatorname{deg}(\bm{u}'_{\delta+1})=\delta\right) \\
= & P\left(\operatorname{ {\bf PNUV} }\left(\{\bm{u}'_i \}_{i=1}^{\delta+1}, r \right)=\ell|\operatorname{deg}(\bm{u}'_{\delta+1})=\delta\right), \numberthis \label{eqn:newconditionalprobinvariant}
\end{align*}
where $\operatorname{deg}(\bm{u}'_{\delta+1})$ on the left (right) side is the degree of vertex $\bm{u}'_{\delta+1}$ in the corresponding random (pseudo) geometric graph.
\end{lem}
Lemma \ref{prop:rangeo} establishes that the conditional distributions of the number of universal vertices in the random geometric graph and the random pseudo geometric graph are identical.
By Lemma \ref{prop:rangeo} and \eqref{eqn:alphandeltarho},
\begin{equation}
\alpha(\ell,r_\rho) = P\left(\operatorname{{\bf NUV}}\left(\{\bm{u}'_i \}_{i=1}^{\delta+1}, r_\rho \right)=\ell|\operatorname{deg}(\bm{u}'_{\delta+1})=\delta\right) \label{eqn:alphalrhorgg}
\end{equation}
when $r_\rho<2/\sqrt{5}$ or equivalently $\rho>3/5$. Therefore to study the limit of $\alpha(\ell,r_\rho)$ when $r_\rho\to 0$ or equivalently $\rho\to 1$, the limiting form of the right hand side of \eqref{eqn:alphalrhorgg} is required, given by the following lemma.
\begin{lem}
\label{prop:rangeolimit}
Let $\delta\geq 1$ and $n\geq 3$. Suppose $\{\bm{u}'_i \}_{i=1}^{{\delta+1}}\overset{\text{i.i.d.}}{\sim} \operatorname{{\bf unif}}(S^{n-2})$ and $\{\tilde{\bm{u}}_i\}_{i=1}^{\delta} \overset{\text{i.i.d.}}{\sim} \operatorname{{\bf unif}}(B^{n-2})$. Then for any $\ell \in [{\delta+1}]$,
\begin{align*}
&\lim_{r\to 0}P\left(\operatorname{{\bf NUV}}\left(\{\bm{u}'_i \}_{i=1}^{\delta+1}, r \right)=\ell|\operatorname{deg}(\bm{u}'_{\delta+1})=\delta \right)\\
= & P\left(\operatorname{{\bf NUV}}\left(\{\tilde{\bm{u}}_i \}_{i=1}^{\delta}, 1 \right)=\ell-1\right). \numberthis \label{eqn:tobeproof2}
\end{align*}
\end{lem}
One immediate consequence of Lemma \ref{prop:rangeolimit} and \eqref{eqn:alphalrhorgg} is that $\lim_{\rho\to 1}\alpha(\ell,r_\rho)=\alpha_\ell$ for any $\ell\in [\delta+1]$. As $p\to \infty$, the condition $2p^{1+\frac{1}{\delta}}P_n(r_\rho)\leq \gamma$ in Theorem \ref{thm:Poissonultrahigh} entails $\rho\to 1$. The following lemma states if the rate of $\rho\to 1$ is coupled with the rate $p\to \infty$, then the non-asymptotic compound Poisson distribution $\operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho})$ converges \text{ in distribution} to the limiting compound Poisson distribution $\operatorname{CP}(\lambda_{n,\delta}(e_{n,\delta}),\bm{\zeta}_{n,\delta})$.
\begin{lem}
\label{lem:compoundpoissonlimit}
Suppose as $p\to \infty$, $\rho\to 1$ such that $a_n2^{\frac{n}{2}}p^{1+\frac{1}{\delta}}(1-\rho)^{\frac{n-2}{2}}\to e_{n,\delta}$, where $e_{n,\delta}$ is some positive constant that possibly depends on $n$ and $\delta$. Then
\begin{equation}
\operatorname{CP}(\lambda_{p,n,\delta,\rho},\bm{\zeta}_{n,\delta,\rho}) \overset{\mathcal{D}}{\to} \operatorname{CP}(\lambda_{n,\delta}(e_{n,\delta}),\bm{\zeta}_{n,\delta}). \label{eqn:jumpsizedislimit}
\end{equation}
\end{lem}
With the above results established we are in a position to prove Theorem \ref{cor:Poissonlimit} and Lemma \ref{item:Poissonlimita}.
\begin{proof}[Proof of Theorem \ref{cor:Poissonlimit} and Lemma \ref{item:Poissonlimita}]
Theorem \ref{cor:Poissonlimit} directly follows from Theorem \ref{thm:Poissonultrahigh} and Lemma \ref{lem:compoundpoissonlimit}. Lemma \ref{item:Poissonlimita} directly follows from Lemma \ref{item:Poissonultrahigha} and Lemma \ref{lem:compoundpoissonlimit}.
\end{proof}
\begin{rem
\label{rem:comparisonsfinitelimit}
\ONE{
In \eqref{eqn:slowupperbound} of the proof of Lemma \ref{lem:compoundpoissonlimit}, it is shown that part of the upper bound for its error rate is of the order $1-\rho$. This particular rate, however, decreases as $n$ increases. More concisely, if one chooses $\rho$ according to \eqref{eqn:rhopformula} then $1-\rho$ is of the order $p^{-(1+\frac{1}{\delta})\frac{2}{n-2}}$. Hence the convergence to $\operatorname{CP}(\lambda_{n,\delta}(e_{n,\delta}),\bm{\zeta}_{n,\delta})$ in Theorem \ref{cor:Poissonlimit} is only accurate for large $p$. On the other hand, the upper bound in Theorem \ref{thm:Poissonultrahigh} only depends on $\rho$ through $2p^{1+\frac{1}{\delta}}P_n(r_\rho)\leq \gamma$, which holds for small $p$, again if $\rho$ is chosen according to \eqref{eqn:rhopformula} (see the first paragraph of Remark \ref{rem:edgecor} for related discussion). Hence Theorem \ref{thm:Poissonultrahigh} provides an accurate approximation to $\bar{N}_{\delta}$ even for small $p$. The accuracy of these approximations for various values of $\rho$ and $n$ is numerically illustrated in Figure \ref{fig:finiteplimitcomparison} in Section \ref{sec:simulation} of the Appendix.
}
\end{rem}
|
\subsection{Overview of results}
This paper examines a model of RNA secondary structure in which secondary structures are modeled by plane trees. As defined more rigorously in Section~\ref{sec:preliminaries}, we consider the set of all plane trees with \(n\) edges under a Gibbs distribution, where the energy of each tree depends on its degree sequence and root degree. The energy function is also determined by 3 thermodynamic parameters, which we treat as fixed: \((\alpha , \beta , \gamma )\). In this paper, we treat these parameters as arbitrary real numbers. However, specific values of interest do arise from the thermodynamic models used in molecular biology; the biological motivation is discussed further in Section \ref{bioback}. For a wide class of properties of plane trees, we show a relationship in their asymptotic distributions under different values of \((\alpha , \beta , \gamma )\).
One such property is the path length. The path length of a plane tree is the sum of the distances from each vertex of the tree to the root. It is known that the path length of all plane trees with \(n\) edges is Airy distributed asymptotically as \(n\rightarrow \infty \) (see Theorem \ref{contactdistance - thm}). This case, with the plane trees being uniformly distributed, corresponds to the thermodynamic parameters \((0, 0, 0)\). Applying Theorem \ref{maintheorem3} to the path length property allows us to relate the asymptotic distribution of path length under arbitrary thermodynamic parameters \((\alpha , \beta , \gamma )\) to this known result. As shown in Corollary \ref{full dist corr CD}, we can explicitly state the asymptotic distribution of path length under arbitrary thermodynamic parameters.
For some properties, such as the path length, the asymptotic distribution in the uniform case is known. For many other properties, such as the sum of the distances from each leaf to the root, this distribution is not known. Therefore in addition to our main result relating asymptotic distributions of plane tree properties under different thermodynamic parameters, we also develop some tools which are useful for determining the asymptotic distributions of certain properties in the uniform case. Combining these tools with the main result discussed above allows us to obtain explicit forms for asymptotic distributions for a wide variety of plane tree properties under arbitrary thermodynamic parameters. More specifically, in Theorem \ref{mainSimpleAdditveTheorem1}, we relate the asymptotic distribution of a class of plane tree properties, which we call simple subtree additive properties, to the asymptotic distribution of path length, both in the uniform case. As shown in Corollary \ref{corollary P^LRD and P^IRD dist}, this theorem allows us to deduce that the total leaf to root distance is also distributed asymptotically as an Airy random variable.
\subsection{Biological background}
\label{bioback}
We now proceed with some exposition on RNA secondary structure and prior work in this area. The reader interested only in the mathematics may skip this section.
\subsubsection{RNA secondary structure}
RNA is an important biological polymer with roles including information transfer, regulation of gene expression, and catalysis of chemical reactions.
The \emph{primary structure} of an RNA molecule is the sequence of nucleotides in the polymer.
RNA nucleotides are adenine, cytosine, guanine, and urasil, which we frequently abbreviate as A, C, G, and U, respectively.
The primary structure, therefore, may simply be understood as a string of A's, C's, G's, and U's.
Because RNA is single-stranded, it has the capacity to form nucleotide-nucleotide bonds with itself.
The set of such bonds is the \emph{secondary structure} of an RNA molecule.
The bonds A-U, C-G, and G-U are considered canonical, and our the only bonds considered by the model presented here.
The \emph{tertiary structure} of an RNA molecule is its three-dimensional shape.
Though tertiary structure ultimately is most relevant to the determination of function, it is also very difficult to deduce with current laboratory techniques.
Therefore, secondary structure is often used as a first step in the process of predicting tertiary structure \cite{Doudna2000structural, Tinoco99how}.
In fact, secondary structure is often an input to tertiary structure prediction algorithms \cite{massire1998manip,seetin2000automated, yunjie2011improvements, zhao2012automated}.
One of the main computational tools for predicting RNA secondary structures is thermodynamic free energy minimization using Nearest Neighbor Thermodynamics Modeling (NNTM) \cite{Turner_89, mathews1999, mathews2004}. Under the NNTM, the free energy of a structure is computed as the sum of the free energy of its various substructures. This free energy is in turn used in algorithms to predict secondary structure given an RNA sequence, see, e.g., \cite{Ding_Chan_Lawrence_04_Sfold,Hofacker94_fastFolding, mathuriya09_gtfold}.
Though such algorithms perform reasonably well on short RNA sequences, performance rapidly degrades once sequence length exceeds a few hundred nucleotides.
\subsubsection{Multiloops and branching}
This paper investigates an aspect of RNA secondary structure that becomes more significant as sequence length grows: multiloops.
A \emph{multiloop} is a place where 3 or more helices meet in an RNA secondary structure.
Multiloops are not predicted well by the current NNTM energy assignments \cite{Doshi_04}.
The number and type of multiloops determines the branching behavior of an RNA secondary structure.
We study a model for RNA secondary structure first presented by Hower and Heitsch \cite{hower2011}.
This model isolates the multiloops and branching properties of secondary structure, allowing their study without the necessity to consider the identity of individual base pairs.
Under the model, secondary structures are placed in bijection with plane trees.
The minimum energy structures under the model were characterized by Hower and Heitsch in the original paper, but this leaves open the question of the full Gibbs distribution of possible structures, as well as the question of characterizing asymptotic behavior of the distribution.
Kirkpatrick et al. \cite{Kirkpatrick_et_al_20} have shown the existence of a polynomial-time Markov chain-based algorithm for sampling from the Gibbs distribution on structures of a fixed size.
Bakhtin and Heitsch~\cite{Bakhtin_Heitsch_09} analyzed a simplification of the model and determined degree sequence properties of the distribution of plane trees asymptotically.
Several properties are used to describe the overall branching behavior of an RNA secondary structure.
In particular, ladder distance and contact distance are used to characterize various aspects of a molecule's shape, size, and branching structure (see, e.g. \cite{borodavka2016sizes, Yoffe_08}).
As discussed in Section \ref{sec:examples}, ladder distance and contact distance correspond to Wiener index and path length, respectively, of plane trees.
We will examine the distribution of several plane tree properties asymptotically under the full version of the Hower and Heitsch model.
Under an assumption that the root degree is bounded, the theorems developed will allow us to characterize the asymptotic distribution of many properties of RNA secondary structures under the NNTM.
We will further show that altering the parameters \((\alpha, \beta, \gamma)\) only changes the property distribution by a constant multiple.
When the assumption that the root degree is bounded is removed, we will still obtain analogous results for parameter sets of the form \((\alpha , \beta , 0)\).
\subsection{Structure of this paper}
In Section~\ref{sec:preliminaries}, we give an overview the necessary mathematical preliminaries (including relevant known results). In Section \ref{Gen function count section}, we construct generating functions counting plane trees. In Section \ref{Distributions of Simple Subtree Additive Properties}, we relate the moments (and hence the distribution) of a class of subtree additive properties we call simple. In Section \ref{Dist of Subtree Additive Properties NNTM}, we apply the generating functions in Section \ref{Gen function count section} to relate the moments of a class of subtree additive properties under the uniformly weighted distribution on the plane trees to the same properties under the non-uniformly weighted distribution on the plane trees arising from the Nearest Neighbor Thermodynamic Model. We, hence, show that there exists a constant that translates the asymptotic random variable in the uniformly weighted case to the asymptotic random variable in the non-uniformly weighted case. In Section \ref{Counting trees - enumerations}, we provide some miscellaneous enumeration results on plane trees.
\section{Background and Introduction}
\input{Introduction.tex}
\section{Mathematical Preliminaries}
\input{Preliminaries}
\section{Results} \label{Resuls Section}
\input{Results_Sec_1.tex}
\input{Results_Sec_2.tex}
\input{Results_Sec_3.tex}
\input{Other_Results.tex}
\section{Discussion}
\label{sec:discussion}
\input{Discussion.tex}
\section{Acknowledgements}
\input{Acknowledgements.tex}
\bibliographystyle{plain}
\section{Other Results}
\subsection{Counting trees by leaves, internal nodes and root degree}
\label{Counting trees - enumerations}
In the study of the Nearest Neighbour Thermodynamic Model, the partition function, $\mathcal{Z}_{(n,\alpha,\beta,\gamma)}$, is of general interest. Thus, we provide enumeration results to assist with the computation this quantity. We define $\mathcal{T}_n(m,k,r)$ to be the number of trees on $n$ vertices with $m$ internal nodes, $k$ leaves and root degree $r$. We will omit any of the above parameters in $\mathcal{T}_n(m,k,r)$ to denote the number of trees summed over the omitted parameters.
It is well-known that $\mathcal{T}_n(k)$ is given by the Narayana numbers \cite{riordan1968combinatorial, narayana1955treillis}. Hence,
\begin{theorem} For $n \geq k$,
\label{count trees by leaves}
$$\mathcal{T}_n(k) = \frac{1}{n} \binom{n}{k} \binom{n}{k-1}.$$
\end{theorem}
From the work of Dershowitz and Zaks \cite{dershowitz1980enumerations}, we also know that
\begin{theorem} For $n \geq r$,
\label{count trees by root}
$$\mathcal{T}_n(r) = \frac{r}{n} \binom{2n-1-r}{n-1}.$$
\end{theorem}
From the work of Donaghey and Shapiro \cite{donaghey1977motzkin}, we know that the number of tree on $n$ edges with no internal nodes is given by the $(n-1)$th Motzkin number. Hence, we have the following result.
\begin{theorem} For $n > m$,
\label{count trees by intenral}
$$\mathcal{T}_n(m) = \binom{n-1}{m}M_{n-m-1},$$ where $$M_n = \sum_{k=0}^{\left\lfloor \frac{n}{2} \right\rfloor} \binom{n}{2k}C_k$$ are the Motzkin numbers.
\end{theorem}
The main result of this section is the closed form expression for $\mathcal{T}_n(m,k,r)$ given in the following theorem.
\begin{theorem}
\label{count trees by all params} For $n - m > k > r, $
$$\mathcal{T}_n(m,k,r) = \frac{r}{n} \binom{n}{k+m} \binom{k+m}{k}\binom{k-r-1}{n-m-k-1},$$ and for $n - m = k = r$,
$$\mathcal{T}_n(m,k,r) = \binom{n-1}{m}.$$
\end{theorem}
Before proving this theorem, we first observe the following useful lemmas.
\begin{lemma}
\label{psi translation lemma}
The number of trees with root degree $r$, $m$ internal nodes and $k$ leaves is equal to the number of trees with leftmost path with $r$ edge, $m$ leaves with a left sibling and $n+1-k$ leaves in total.
\end{lemma}
\begin{proof}
We briefly describe the following bijection from $\mathfrak{T}_n$ to $\mathfrak{T}_n$ given by Dershowitz and Zaks \cite{dershowitz1980enumerations}, which we will denote as $\psi$. We describe $\psi$ recursively. For $T = T_1 \ltimes T_2 \in \mathfrak{T}_{\geq 1}$, $\psi(T) = \psi(T_2) \ltimes \psi(T_1)$ and $\psi(T^*) = T^*$. (See top left of Figure \ref{fig:treebijectionimage} for illustration.)
Fix $T \in \mathfrak{T}_n$. We denote a leaf edge to be the edge incident to a leaf. Notice that leaf edges in $T$ become edges without a right sibling in $\psi(T)$ (the rightmost edge under each vertex). Notice that the number of rightmost edges in a tree is the number of vertices with at least 1 child, hence the number of non-leaf vertices.
To describe the effect of $\psi$ on internal node, for an internal node, $v$, of $T$, we denote its parent edge and child edge by $p(v)$ and $c(v)$. Since we assume the root cannot be internal, a vertex is an internal node if and only if it has a parent edge and a child edge which has no siblings. Notice that the edge $c(v)$ and $p(v)$ in $T$ becomes a leaf in $\psi(T)$ and a left sibling of that leaf respectively. Furthermore, every leaf and left sibling pair induce an internal node. Finally notice that the root degree of $T$ is the length of the path from the root to the leaf reached by moving down leftmost edges. (See Figure \ref{fig:treebijectionimage} for illustration.)
\end{proof}
\input{Figures/Figure_2.tex}
\begin{lemma}
\label{corrNumTree}
The number of trees with root degree $r$, no internal nodes and $k$ leaves is equal to the number of balanced parenthesis sequences with an initial run of $r$ opening parenthesis with $n+1 -k$ occurrences of $'()'$ with the restriction that no $'()'$ is preceded by a closing parenthesis.
\end{lemma}
\begin{proof}
We consider the standard bijection from plane trees to balances parenthesis sequences, denoted $\psi_2$. Fix $T \in \mathfrak{T}_n$. The length of the leftmost path of $T$ is the length of the initial run of opening parenthesis of $\psi_2(T)$. A leaf in $T$ corresponds to the sub-sequence $'()'$ in $\psi_2(T)$. A leaf of $T$ with a left sibling corresponds to the sub-sequence $')()'$ in $\psi_2(T)$. Thus for $T$ to have no leaves with left sibling, we must avoid the sub-sequence $')()'$. Applying Lemma \ref{psi translation lemma}, we achieve the result.
\end{proof}
\begin{lemma}[Cycle Lemma \cite{dershowitz1982cycle}]
For any sequence $p_0 p_1\cdots p_{m+n-1}$ of
$m$ open parentheses and $n$ close parentheses, where $m > n$, there exist exactly $m - n$ cyclic permutations
$$p_i p_{i+1}\cdots p_{m+n-1}p_0\cdots p_{i-1}$$
such that the number of opening parenthesis is always greater than the number of closing parenthesis. We will say these sequences are dominating.
\end{lemma}
We adapt the Cycle Lemma to aid the computation that will following.
\begin{corollary} \label{ratio cycle lemma}
For any sequence $p = p_0 p_1\cdots p_{m+n-1}$ of
$m$ open parentheses and $n$ close parentheses, where $m > n$, among all distinct cyclic permutation of $p$, the ratio between the number of dominating sequences and total number of such permutations is $\frac{m-n}{m+n}$.
\end{corollary}
\begin{proof}
We will denote the permutation $p_i p_{i+1}\cdots p_{m+n-1}p_0\cdots p_{i-1}$ by $p(i)$ (where $i$ is modulo $m+n$). Let $k \leq m + n$ be smallest natural number such that $p(0) = p(k)$. Notice that $p(0) = p(sk)$ for all $s \in \mathbb{Z}$.
We now show that for any $i, j \in \mathbb{Z}$, $p(i) = p(j)$ if and only if $i \equiv j \mod k$. It should be clear that $p(i) = p(i + sk)$ for all $s,i \in \mathbb{Z}$, since $p(0) = p(sk)$ and shifting $p(0)$ and $p(sk)$ by $i$, we get $p(i)$ and $p(i+sk)$, respectively. Alternatively, if we assume $i \not\equiv j \mod k$ (thus $i = j + sk + r$ where $0 <r < k$) and $p(i) = p(j)$, we see that $p(i) = p(j) = p(i + sk + r) = p(i + r)$. Shifting $p(i)$ and $p(i + r)$ by $-i$, we get that $p(0) = p(r)$, a contraction to the minimality of $k$.
Since $p(0) = p(m+n)$, $k$ divides $m+n$. We thus see that the set of distinct cyclic permutations of $p$ is $P = \{ p(i): 0 \leq i < k\} $ and each such permutation occurs $l = \frac{m+n}{k}$ times as a cyclic permutation of $p$. Let $q$ be the number of element of $P$ that are dominating sequences. By the Cycle Lemma, $\frac{m-n}{m+n} = \frac{lq}{lk} = \frac{q}{k}$, proving the result.
\end{proof}
\bigskip
\textit{Proof of Theorem \ref{count trees by all params}.}
We count the number of trees with root degree $r$, no internal nodes and $k$ leaves using the bijection in Corollary \ref{corrNumTree}. Let $l = n+1 -k$. The beginning of all such balanced parenthesis sequences is fixed as a run of $r$ opening parenthesis followed by a closing parenthesis. So we need to count the number of valid suffixes with $n-r$ opening parenthesis, $n-1$ closing parenthesis and $l-1$ occurrences of $'(()'$ and no other occurrences of $'()'$ or symmetrically, the number of valid prefixes with $n-1$ opening parenthesis, $n-r$ closing parenthesis and $l-1$ occurrences of $'())'$ and no other occurrences of $'()'$.
We count this via the cycle lemma. We count the number of dominating sequence with $n$ opening parenthesis, $n-r$ closing parenthesis and $k'-1$ occurrences of $'())'$ and no other occurrences of $'()'$. We first count all possible cyclic permutations of all such sequences:
\begin{enumerate}
\item \textit{Starting with $'('$ and ending in $')'$.}
All such sequence are achieved by partitioning the $n$ opening parenthesis into $l-1$ (ordered) parts of size 1 or more, denotes $x_1, \cdots, x_{l-1}$ and partitioning the $n-r$ closing parenthesis into $l-1$ parts of size 2 or more (since we want to guarantee that every $'()'$ is proceeded by a closing parenthesis), denotes $y_1, \cdots, y_{l-1}$. The formed sequence is then $x_1y_1 \cdots x_{l-1}y_{l-1}$. Thus the number of sequences is $\binom{n-1}{l-2} \binom{n-r-l}{l-2}$.
\item \textit{Starting with $'('$ and ending in $'('$.}
Similarly to (1) with the extra step of adding another run of opening parenthesis of size at least 1 at the end of the sequence. These are sequences of the form $x_1y_1 \cdots x_{l-1}y_{l-1}x_{l}$. The number of such sequences is $\binom{n-1}{l-1} \binom{n-r-l}{l-2}$.
\item \textit{Starting with $')'$ and ending in $'('$.}
We notice that we lose an occurrence of $'()'$ since it must be formed by the last opening parenthesis and the first closing parenthesis (in a cyclic shift). These sequences are similar to the sequences in (1) shifted so they are of the form $y_{l-1}x_1y_1 \cdots x_{l-1}$. Thus the number of such sequences is $\binom{n-1}{l-2} \binom{n-r-l}{l-2}$.
\item \textit{Starting with $')'$ and ending in $')'$.}
These are sequences of the form $y_{0}x_1y_1 \cdots x_{l-1}y_{l-1}$. Since, we are counting cyclic permutations of dominating sequences of the desired type, the restriction of each run of closing parenthesis sequences being of size 2 or more is loosen for the first and last runs (since in any cyclic permutation that would lead to a dominating sequences, these 2 runs will become 1 run). Thus the number of sequences is $\binom{n-1}{l-2} \binom{n-r-l+1}{l-1}$.
\end{enumerate}
Notice that we counted all the \emph{distinct} cyclic permutations of the desired dominating sequences. Using Corollary \ref{ratio cycle lemma}, for any sequence counted above, in the set of its distinct cyclic permutations, $\frac{r}{2n-r}$ of members of the set are valid dominating sequences. Thus the number of tree with $n$ edges, root degree of $r$ and $k$ leaves and no internal nodes is the sum of the terms derived for each of the above cases scaled by the term $\frac{r}{2n-r}$. After simplifying and setting $l = n+1-k$, we get this to be
$$\frac{r}{n}\binom{n}{k}\binom{k-r-1}{n-k-1}.$$
Adding internal nodes is now just a matter of deciding how many to put under each edge. Thus for each tree on $n-m$ edges with root degree of $r$ and $k$ leaves and no internal nodes, there are $\binom{n-1}{m}$ trees with $n$ edges with root degree of $r$ and $k$ leaves and $m$ internal nodes giving us the desired expression (after simplification).\\
\begin{corollary}
\label{count trees by leaves and internal} For $n - m > k $,
$$\mathcal{T}_n(m,k) = \frac{1}{n} \binom{n}{k+m} \binom{k+m}{k}\binom{k}{n-m-k+1},$$ and for $n - m = k $,
$$\mathcal{T}_n(m,k) = \binom{n-1}{m}.$$
\end{corollary}
\begin{proof}
Let $p, q \in \mathbb{N}$. We now evaluate the following sum.
\begin{eqnarray}
\sum_{q \geq 0}\sum_{r \geq 0}r\binom{q-r}{p} x^q &=& x\sum_{r \geq 0}rx^{r-1}\sum_{q \geq 0}\binom{q-r}{p}x^{q-r}\\
&=& \frac{x^{p+1}}{(1-x)^{p+3}} = \sum_{q \geq 0} \binom{q+1}{p+2} x^{q}.
\end{eqnarray}
Thus we get $$\sum_{r \geq 0}r\binom{q-r}{p} = \binom{q+1}{p+2}.$$ We now set $q = k-1$ and $p = n - m - k -1$ to achieve the desired result.\\
\end{proof}
\begin{corollary}
\label{count trees by leaves and root} For $n > k > r$,
$$\mathcal{T}_n(k, r) = \frac{r}{n} \binom{n}{k} \binom{n-r-1}{n-k-1},$$ and for $n = k = r$,
$$\mathcal{T}_n(k,r) = 1.$$
\end{corollary}
\begin{proof}
Let $p, q \in \mathbb{N}$. We now evaluate the following sum.
\begin{eqnarray}
\sum_{m \geq 0}\sum_{p\geq 0}\binom{n-k}{m}\binom{q}{p-m} x^p &=& \sum_{m \geq 0}\binom{n-k}{m}x^m\sum_{p\geq m}\binom{q}{p-m} x^{p-m}\\
&=& (1+x)^{n-k+q} = \sum_{p \geq 0} \binom{n-k+q}{p} x^{p}.
\end{eqnarray}
Thus we get $$\sum_{m \geq 0}\frac{r}{n} \binom{n}{k+m} \binom{k+m}{k}\binom{q}{p-m} = \frac{r}{n} \binom{n}{k} \sum_{m \geq 0}\binom{n-k}{m}\binom{q}{p-m} = \frac{r}{n} \binom{n}{k} \binom{n-k+q}{p} .$$
We now set $q = k-r-1$ and $p = n-k-1$ to achieve the desired result.\\
\end{proof}
\subsection{Plane Trees and their Properties}
A \emph{plane tree} is a rooted ordered tree. Let $\mathfrak{T}_n$ denote the set of plane trees on $n$ edges. Let $\mathfrak{T}_{\leq k} = \cup_{n\leq k} \mathfrak{T}_{n}$. It is well-known that $|\mathfrak{T}_n|$ is given by the $n$th Catalan number, $C_n = \frac{1}{n+1}\binom{2n}{n}$. We define the \emph{down degree} of a vertex to be the degree of the vertex when considering the root and one less than the degree of the vertex for all other vertices. We define a \emph{leaf} to be a non-root vertex with down degree 0 and an \emph{internal node} to be a non-root vertex with down degree 1. For a plane tree $T$, let $v(T)$, $n(T)$, ${d_0}(T)$, ${d_1}(T)$ and $r(T)$ be the number of vertices, edges, leaves, internal nodes and root degree of $T$, respectively. Let $\mathcal{V}(T)$ be the vertex set of $T$ and let $\overline{\mathcal{V}}(T)$ be the vertex set excluding the root vertex. For $v \in \mathcal{V}(T)$, let $T_v$ be the subtree of $T$ that contains all descendant of $v$ (including $v$).
For plane trees $T_1 \in \mathfrak{T}_n$ and $T_2 \in \mathfrak{T}_m$ for some $m,n \geq 0$, we define the join of $T_1$ and $T_2$, $T_1 \ltimes T_2$, to be the tree formed by adding a new edges to leftmost side of the root of $T_1$ and attaching $T_2$ to this new edge. Note that $T_1 \ltimes T_2 \in \mathfrak{T}_{n+m+1}$. Notice that for a tree $T \in \mathfrak{T}_{> 0}$, there is unique $T_1, T_2 \in \mathfrak{T}_{\geq 0}$ such that $T = T_1 \ltimes T_2$.
We define a \emph{property} of a plane tree to be a function $\mathcal{P} : \mathfrak{T}_{\geq 0} \rightarrow \mathbb{R}$. We define the 2 types of properties we will consider from this point forward: \emph{additive properties} and \emph{subtree additive properties}. A property $\mathcal{P}$ is \emph{additive} if, for $T \in \mathfrak{T}_{\geq 1}$ such that $T = T_1 \ltimes T_2$ for some $T_1, T_2$,
$$\mathcal{P}(T) = \mathcal{P}(T_1) + \mathcal{P}(T_2) + f(T_2),$$
where $f(T): \mathfrak{T}_{\geq 0} \rightarrow \mathbb{R}_{\geq 0}$. For a tree $T$, let $v_1, \cdots, v_d$ be the child vertices of the root vertex of $T$. By repeated use of the above definition, we see that
\begin{equation} \label{eq additive prop def}
\mathcal{P}(T) = \mathcal{P}(T^*) + \sum_{i=1}^d f(T_{v_i}) + \sum_{i=1}^d\mathcal{P}(T_{v_i}),
\end{equation}
where $d$ is the degree of the root of $T$ and $T^*$ is the tree on 1 vertex.
This is similar to the notion of an additive functional as described by Janson \cite{janson2016asymptotic}. An additive functional is a function $F: \mathfrak{T}_{\geq 0} \rightarrow \mathbb{R}$ such that
$$F(T) = f(T) + \sum_{i=1}^dF(T_{v_i}),$$
where $f:\mathfrak{T}_{\geq 0} \rightarrow \mathbb{R}$ is known as the toll function. Due to $(\ref{eq additive prop def})$, we note that an additive property (as we have defined it) is an additive functional with toll function $f$ where there exist a function $f^*: \mathfrak{T}_{\geq 0} \rightarrow \mathbb{R}_{\geq 0}$ such that $f(T) = c + \sum_{i=1}^df^*(T_{v_i})$ where $c = \mathcal{P}(T^*)$. We will borrow the terminology of additive functionals and call the tuple $(f, c)$ the toll of the additive property. It should be clear that a subtree additive property is uniquely determined by its toll. Thus, we will denote the additive property with a given toll by $\mathcal{P}^{(f,c)}$. We will call $f$, in the toll, the toll function of the subtree additive property. We will call an additive property non-negative integer valued if the co-domain of the toll function is a subset of $\mathbb{Z}_{\geq 0}$ (and $c \in \mathbb{Z}_{\geq 0}$).
It can also be shown inductively that, for $c, c_1, c_2 \in \mathbb{R}$, $f_1, f_2: \mathfrak{T}_{\geq 0} \rightarrow \mathbb{R}$ and $T \in \mathfrak{T}_{n}$,
\begin{equation} \label{additve prop linearity}
\mathcal{P}^{(c_1 \cdot f_1 + c_2 \cdot f_2, c)}(T) = c_1 \cdot \mathcal{P}^{(f_1, 0)}(T) + c_2 \cdot \mathcal{P}^{(f_2, 0)}(T) + c \cdot \mathcal{P}^{(0, 1)}(T).
\end{equation}
A property $\mathcal{P}$ is \emph{subtree additive} if, for $T \in \mathfrak{T}_{\geq 1}$,
$$\mathcal{P}(T) = \sum_{v \in \overline{\mathcal{V}}(T)} f(T_v, T).$$
where $f: \mathfrak{T}_{\geq 0} \times \mathfrak{T}_{\geq 0} \rightarrow \mathbb{R}$, and $\mathcal{P}(T^*) =0$. We will call such a property simple if
$$\mathcal{P}(T) = \sum_{v \in \overline{\mathcal{V}}(T)} \mathcal{P}'(T_v)$$
where $\mathcal{P}'$ is a non-negative integer valued additive property. In this case, we call $\mathcal{P}$ the subtree additive property induced by $\mathcal{P}'$.
\bigskip
For a plane tree $T$, the \emph{energy} of the tree is given by
\begin{equation}
E(T) = \alpha d_0(T) + \beta d_1(T) + \gamma r(T)
\label{energydef}
\end{equation} where $\alpha, \beta, \gamma \in \mathbb{R}$ are parameters of the energy function.
For fixed $n \in \mathbb{N}$ and parameters $\alpha, \beta, \gamma \in \mathbb{R}$ and property $\mathcal{P}$, we define the random variable $\mathcal{P}_{(\alpha, \beta, \gamma)}(\mathfrak{T}_n)$ to be $\mathcal{P}(T)$ for a plane tree $T \in \mathfrak{T}_n$ selected at random with probability
$\frac{e^{-E(T)}}{\mathcal{Z}_{(n,\alpha,\beta,\gamma)}}$
where $\mathcal{Z}_{(n,\alpha,\beta,\gamma)}$ is a normalizing constant given by
$\mathcal{Z}_{(n,\alpha,\beta,\gamma)} = \sum_{T \in \mathfrak{T}_n}e^{-E(T)}.$ For convenience, we will denote $\mathcal{P}_{(0,0,0)}(\mathfrak{T}_n)$ simply as $\mathcal{P}(\mathfrak{T}_n)$.
Let $\mathcal{P}$ and $\overline{\mathcal{P}}$ be properties and let $\alpha_1,\alpha_2, \beta_1, \beta_2, \gamma_1, \gamma_2 \in \mathbb{R}$ be parameters. We use $$\mathcal{P}_{(\alpha_1, \beta_1, \gamma_1)}(\mathfrak{T}_n) \overset{d}{\longleftrightarrow}\overline{\mathcal{P}}_{(\alpha_2, \beta_2, \gamma_2)}(\mathfrak{T}_n)$$ to imply that there exist a random variable $W$ such that as $n \rightarrow \infty$, $$\mathcal{P}_{(\alpha_1, \beta_1, \gamma_1)}(\mathfrak{T}_n) \overset{d}{\to} W \quad \quad \quad \quad \text{and} \quad \quad\quad \quad \overline{\mathcal{P}}_{(\alpha_2, \beta_2, \gamma_2)}(\mathfrak{T}_n) \overset{d}{\to} W,$$ where $\overset{d}{\to}$ denotes convergence in distribution. In this case, we will say that $\mathcal{P}_{(\alpha_1, \beta_1, \gamma_1)}(\mathfrak{T}_n)$ and $\mathcal{P}_{(\alpha_2, \beta_2, \gamma_2)}(\mathfrak{T}_n)$ are equivalent in distribution.
We now state the well-known Carleman's condition \cite{carleman1926fonctions} which tells us that, under certain conditions (which the random variables we will consider satisfy), to show equivalence in distribution, it is sufficient to show equality is asymptotic moments.
\begin{theorem}[Carleman's condition] \label{Carleman's condition}
Let $X$ be a random real-valued variable and let $m_k = \mathbb{E}[|X|^k] < \infty$ for all $k \geq 0$. If $$\sum_{k=1}^\infty |m_{2k}|^{-\frac{1}{2k}} = \infty,$$ then there exists a unique distribution with moments $m_k$.
\end{theorem}
\subsection{Examples}\label{sec:examples}
\begin{example} \label{Example number of leaves and edges} It can be shown inductively that the number of edges in a tree is given by the additive property with toll function $(f, 0)$ where $f(T) = 1$ for all $T \in\mathfrak{T}_{\geq 0}$. We will denote this by $\mathcal{P}^{e}$. Similarly, the number of vertices in a tree is given by the additive property with toll function $(0, 1)$ where $0$ is the zero function. We will denote this by $\mathcal{P}^{v}$.
\end{example}
\begin{example}
For $f(T) = 1$ when $T = T^*$ and $f(T) = 0$ otherwise, we observe that $\mathcal{P}^{(f,0)}(T)$ represents the number of leaves in $T$. We will denote this by $\mathcal{P}^{d_0}$. Similarly, for $f(T) = 1$ when $T$ has root degree 1, and $f(T) = 0$ otherwise, $\mathcal{P}^{(f,0)}(T)$ represents the number of internal nodes in $T$. We will denote this by $\mathcal{P}^{d_1}$. Note that for $T \in \mathfrak{T}_n$, $\mathcal{P}^{d_0}(T) \leq n$ and $\mathcal{P}^{d_1}(T) < n$.
\end{example}
\begin{example}
We define the \emph{path length} of $T$, $\mathcal{P}^{PL}(T)$, to be the sum of the edge distances from each vertex of $T$ to the root. In the biological context, this quantity is also known as the \emph{total contact distance}. We can observe that $\mathcal{P}^{PL}(T)$ is a simple subtree additive property since
$$\mathcal{P}^{PL}(T)= \sum_{v \in \overline{\mathcal{V}}(T)} \mathcal{P}^v(T_v) = \sum_{v \in \overline{\mathcal{V}}(T)} \mathcal{P}^e(T_v) + n$$ where $n = \left|\overline{\mathcal{V}}\right|$ is the number of edges in $T$. This holds because the number of paths from vertices to the root that utilize an edge $e$ is the number of vertices in the subtree directly below $e$.
\end{example}
It can be shown from the work of Tak{\'a}cs \cite{takacs1991bernoulli} or more directly from the work of Janson \cite{janson2003wiener} that the following result about the distribution of the path length holds.
\begin{theorem}
As $n \rightarrow \infty$,
\label{contactdistance - thm}
$$\frac{\mathcal{P}^{PL}(\mathfrak{T}_n)}{\sqrt{2n^3}} \overset{d}{\to} \int_0^1e(t) \: dt$$ where $e(t)$ is a normalized Brownian excursion on $[0, 1]$. Thus, it is Airy Distributed. Furthermore,
$$\lim_{n \rightarrow \infty}\mathbb{E}\left[\left(\frac{\mathcal{P}^{PL}(\mathfrak{T}_n)}{\sqrt{2n^3}}\right)^k\right] \: \sim \: \frac{6k}{\sqrt{2}}\left(\frac{k}{12e}\right)^{\frac{k}{2}}$$ as $k \rightarrow \infty$.
\end{theorem}
\begin{example}
We define the \emph{Wiener index} of $T$, $\mathcal{P}^{WI}(T)$, to be the sum of the edge distances between any 2 vertices. In the biological context, this quantity is also known as the \emph{total ladder distance}. We can observe that $\mathcal{P}^{WI}(T)$ is a subtree additive property since
$$\mathcal{P}^{WI}(T)= \sum_{v \in \overline{\mathcal{V}}(T)} \mathcal{P}^v(T_v)(\mathcal{P}^v(T) - \mathcal{P}^v(T_v)).$$ This holds because the number of paths between vertices that utilize an edge $e$ is the number of unordered pairs of vertices, one from the subtree below $e$ and the other not from that subtree.
\end{example}
From the work of Janson \cite{janson2003wiener}, the following result about the distribution of the Wiener index holds.
\begin{theorem} As $n \rightarrow \infty$,
\label{ladderdistance - thm}
$$\frac{\mathcal{P}^{WI}(\mathfrak{T}_n)}{\sqrt{2n^5}} \overset{d}{\to} \int \int_{0 < s < t < 1} (e(s) + e(t) - 2\min_{s \leq u \leq t}e(u)) \: ds \: dt$$ where $e(t)$ is a normalized Brownian excursion on $[0, 1]$.
\end{theorem}
\begin{example}
We define the \emph{total leaf to root distance} of $T$, $\mathcal{P}^{LR}(T)$, and the \emph{total internal node to root distance} of $T$, $\mathcal{P}^{IR}(T)$, to be the sum of the edge distances from every leaf vertex to the root and the sum of the edge distances from every internal node to the root, respectively. We can observe that $\mathcal{P}^{LR}(T)$ and $\mathcal{P}^{IR}(T)$ can be described in terms of a simple subtree additive properties as follows.
$$\mathcal{P}^{LR}(T)= \mathcal{P}^{d_0}(T) + \sum_{v \in \overline{\mathcal{V}}(T)} \mathcal{P}^{d_0}(T_v) = \sum_{v \in \overline{\mathcal{V}}(T)} \mathcal{P}^{d_0}(T_v) + O(n)$$
and
$$\mathcal{P}^{IR}(T)= \mathcal{P}^{d_1}(T) + \sum_{v \in \overline{\mathcal{V}}(T)} \mathcal{P}^{d_1}(T_v) = \sum_{v \in \overline{\mathcal{V}}(T)} \mathcal{P}^{d_1}(T_v) + O(n).$$ We see this as follows. The total leaf to root distance is the sum over all edges of the number of path that use that edge which is the number of leaves in the subtree below the edge, which is $T_v$, where $v$ is the vertex below the edge. We, however, notice that by our definition of a leaf, if $v$ is a leaf in $T$, it will not be counted as a leaf in $T_v$. Thus, overall, we under count by the number of leaves in $T$. A similar argument holds for the total internal node to root distance.
In Section \ref{Distributions of Simple Subtree Additive Properties}, we will show that $\frac{\mathcal{P}^{LR}(\mathfrak{T}_n)}{\sqrt{n^3}}$ and $\frac{\mathcal{P}^{IR}(\mathfrak{T}_n)}{\sqrt{n^3}}$ both converge weakly to an Airy random variable.
\end{example}
\subsection{Generating Functions and Analytic Combinatorics}
For a sequence $(a(n))$ for $n \in \mathbb{Z}_{\geq 0}$, we define the \emph{generating function} corresponding to the sequence to be the formal power series, $$F(x) = \sum_{n \geq 0} a(n) x^n.$$ Similarly, for a $k$-dimensional sequence, $(a(n_1, \cdots, n_k))$ for $n_1, \cdots, n_k \in \mathbb{Z}_{\geq 0}$, we define the \emph{multivariate generating function} corresponding to the sequence to be $$F(x_1, \cdots, x_k) = \sum_{n_1 \geq 0} \cdots \sum_{n_k \geq 0} a(n_1, \cdots, n_k) x_1^{n_1} \cdots x_k^{n_k}.$$
Conversely, we will use $\left[x_1^{n_1}\cdots x_k^{n_k}\right]F(x_1, \cdots, x_k)$ to denote the coefficient of $x_1^{n_1}\cdots x_k^{n_k}$ in the generating function $F(x_1, \cdots, x_k)$, namely $a(n_1, \cdots, n_k)$.
Fix $(a(n))$ for $n \in \mathbb{Z}_{\geq 0}$ and let $F(z)$ be the associated generating function. We can treat $F(z)$ as a function over the complex plane. We say $F(z)$ is \emph{analytic} at a point $z_0 \in \mathbb{C}$ if there exist a region around $z_0$ such that $F(z)$ is differentiable. We say $F(z)$ is analytic on a domain if it is analytic at all points in the domain.
A point, $z_0 \in \mathbb{C}$, is a \emph{singularity} of $F(z)$ if $F(z)$ is not analytic at $z_0$. Furthermore, that singularity is \emph{isolated} if there exist $\epsilon > 0$ such that $F(z)$ is analytic on the domain $\{z \in \mathbb{C} : 0<|z-z_0| < \epsilon \}$. Let $F(z)$ be such that all its singularities of are isolated. We define a \emph{dominant singularity} of $F(z)$ to be an isolated singularity with minimal distance from the origin.
From the work of Flajolet and Sedgewick \cite{flajolet2009analytic}, we now state a slightly simplified form of a so called Transfer Theorem that will allow us to deduce asymptotic information about $(a(n))$ using $F(z)$. For some $R>1$ and $0<\phi<\frac{\pi}{2}$, we define a $\Delta$-domain at $1$ to be the domain $$\Delta(\phi, R) = \{z\in \mathbb{C}:|z|<R, z\neq 1, |\text{arg}(z-1)|>\phi\}.$$
\begin{theorem}[Transfer Theorem] \label{transfer lemma}
Let $(a(n))$ be a sequence with associated generating functions $F(z)$. Let $F(z)$ be a function analytic at $0$ with a unique dominant singularity at $1$ and let $F(z)$ be analytic in a $\Delta$-domain at $1$, $\Delta_0$. Assume there exist $\sigma, \tau$ such that $\sigma$ is a finite linear combination of terms of the form $(1-z)^{-\alpha}$ for $\alpha \in \mathbb{C}$ and $\tau$ is a term of the form $(1-z)^{-\beta}$ for $\beta \in \mathbb{C}$ such that, for $z \in \Delta_0$,
$$F(z) = \sigma\left(z\right) + O\left(\tau\left(z\right)\right) \quad\quad \text{as} \quad\quad z \rightarrow 1.$$
Then, the following asymptotic estimation holds.
$$a(n) = [z^n]F(z) = [z^n]\sigma(z) + O(n^{\beta - 1})\,.$$
\end{theorem}
A basic application of the above result that we will be useful in Section \ref{Resuls Section} is as follows.
\begin{corollary} \label{analytic combo corollary 1} Let $(a(n))$ be a sequence with associated generating functions $F(z)$. Let $\frac{1}{\zeta} \in \mathbb{C}$ be the unique dominant singularity of $F(z)$ and assume $F(z) = (1 - \zeta z)^{-\alpha}g(z)$ where $\alpha \in \mathbb{C}$ and $g(z)$ is a complex-valued function that is analytic in the region $R = \left\{z \in \mathbb{C}: |z| \leq \left|\frac{1}{\zeta}\right| \right\}$. Then, the following asymptotic estimation holds.
$$a(n) = \frac{\zeta^n g\left(\frac{1}{\zeta}\right)}{\Gamma(\alpha)} n^{\alpha - 1} + O\left(\zeta^n n^{\alpha - 2}\right)\,.$$
\end{corollary}
\begin{proof}
We expand $g(z)$ about $z = \frac{1}{\zeta}$ using Taylor's Theorem to get $g(z) = g\left(\frac{1}{\zeta}\right) + O\left(1 - \zeta z\right).$ Thus $$F\left(\frac{z}{\zeta}\right) = g\left(\frac{1}{\zeta}\right)(1 - z)^{-\alpha} + O\left((1 - z)^{1-\alpha}\right) \quad\quad \text{as} \quad\quad z \rightarrow 1.$$ We then apply Theorem \ref{transfer lemma} and notice that $[z^n] F\left(\frac{z}{\zeta}\right) = \frac{1}{\zeta^n}[z^n]F(z)$ to achieve the desired result.
\end{proof}
\subsection{Generating functions for counting trees by leaves, internal nodes and root degree}
\label{Gen function count section}
Let $\mathcal{G}_n(d_0, d_1, r)$ be the set of plane trees on $n$ edges with $d_0$ leaves, $d_1$ internal nodes and root degree $r$. Let
$$G(x,a,b) = \sum_{n = 0}^\infty\sum_{d_0 = 0}^\infty\sum_{d_1 = 0}^\infty\sum_{r = 0}^\infty |\mathcal{G}_n(d_0, d_1, r)|x^na^{d_0}b^{d_1},$$
$$G_r(x,a,b) = \sum_{n = 0}^\infty\sum_{d_0 = 0}^\infty\sum_{d_1 = 0}^\infty |\mathcal{G}_n(d_0, d_1, r)|x^na^{d_0}b^{d_1}$$ and $$G(x,a,b,c) = \sum_{n = 0}^\infty\sum_{d_0 = 0}^\infty\sum_{d_1 = 0}^\infty \sum_{r = 0}^\infty |\mathcal{G}_n(d_0, d_1, r)|x^na^{d_0}b^{d_1}c^r.$$ As we will use it often from this point on, let $G^*(x,a,b) = G_1(x,a,b)$.
\begin{theorem} \label {G recurrence}
The following recurrences hold.
\begin{equation} \label{primary recurrence G}
G(x,a,b) = 1 + xG(x,a,b)^2 + (a-1)xG(x,a,b) + (b-1)xG(x,a,b)G^*(x,a,b),
\end{equation}
\begin{equation}\label{primary recurrence G with root}
G(x,a,b,c) = \frac{1}{1-cG^*(x,a,b)},
\end{equation}
and
\begin{equation}\label{primary recurrence G*}
G^*(x,a,b) = 1- \frac{1}{G(x,a,b)}.
\end{equation}
\end{theorem}
\begin{proof}
To achieve (\ref{primary recurrence G}), we apply the decomposition of a plane tree into 2 subtrees given by $T = T_1 \ltimes T_2$. This adds a new edge. The only tree not accounted for by this process is $T^*$. Note that this process of adjoining trees never creates any new leaves or internal nodes in $T_1$. Notice that when $T_2 = T^*$, $T_2$ has an extra leaf which is not accounted for (since it would be at the root of $T_2$). Thus we re-weight these trees. (Hence $(a-1)xG(x,a,b)$.) Also notice that when $T_2$ has root degree 1, $T_2$ has an extra internal node which is not accounted for (since it would be at the root of $T_2$). Thus we re-weight these trees. (Hence $(b-1)xG(x,a,b)G^*(x,a,b)$.)
To achieve (\ref{primary recurrence G with root}), we notice that a tree with root degree $r$ is equivalent to a sequence of $r$ trees with root degree 1 where we identify all the root vertices. This identification does not add or remove any edges, leaves or internal nodes. Thus the generating function for trees with root degree $r$, where we weight root degree, is $c^rG^*(x,a,b)^r$. Summing over all possible values for $r$, we get the desired expression.
To achieve (\ref{primary recurrence G*}), we set $c$ to 1 in (\ref{primary recurrence G with root}) to ignore root degree, and rearrange the expression.
\end{proof}
\begin{corollary}\label{eval G func}
The following generating functions hold.
\begin{equation}\label{G equation}
G(x,a,b) = \frac{1 + (2-a-b)x - \sqrt{(1+(2-a-b)x)^2 - 4x(1 - (b-1)x)}}{2x} \end{equation} and
\begin{equation}\label{G^* equation}
G^*(x,a,b) = \frac{1 + (a-b)x - \sqrt{(1+(2-a-b)x)^2 - 4x(1 - (b-1)x)}}{2(1-(b-1)x)}. \end{equation}
\end{corollary}
\begin{proof}
The result follows immediately by solving (\ref{primary recurrence G}) and (\ref{primary recurrence G*}) simultaneously.
\end{proof}
\begin{lemma}\label{dom sig lemma} For all $a, b \in \mathbb{R}_{> 0}$, the dominant singularity of $G(x,a,b)$ and $G^*(x,a,b)$ occurs at $\rho = \rho(a, b) = a + b + 2\sqrt{a}$.
\end{lemma}
\begin{proof}
Fix $a, b > 0$. Let $\Psi = (1+(2-a-b)x)^2 - 4x(1 - (b-1)x)$ and let $\overline{\rho} = \overline{\rho}(a,b) = a + b - 2\sqrt{a}$. The roots of $\Psi$ are $\frac{1}{\rho}$ and $\frac{1}{\overline{\rho}}$. Clearly, $0 < \frac{1}{\rho} < \frac{1}{\overline{\rho}}$. Notice that the only singularity caused by the numerator of (\ref{G^* equation}) and (\ref{G equation}) occurs when $\Psi = 0$. Thus the most significant of such singularities is $\frac{1}{\rho}$.
Notice that if there exists another singularity of $G(x,a,b)$, it must occur at $x = 0$ (caused by the denominator). We notice that when $x = 0$, the numerator of $(\ref{G equation})$ goes to 0. Thus, taking a Laurent expansion of $G(z,a,b)$ at $z = 0$, we get $$G(z,a,b) = \frac{1}{2z}\sum_{n \geq 1} d_nz^n = \sum_{n \geq 0} \frac{d_{n + 1}}{2}z^{n},$$ where $d_i$ are constants. Thus $G(z,a,b)$ is analytic at $z = 0$. Thus, the dominant singularity of $G(x,a,b)$ is $\frac{1}{\rho}$.
For $b = 1$, the denominator of $(\ref{G^* equation})$ is a constant, thus cannot cause another singularity. Assume $b \neq 1$. Notice that if there exists another singularity of $G^*(x,a,b)$, it must occur at $x = \frac{1}{b-1}$ (caused by the denominator). We notice that when $x = \frac{1}{b-1}$, the numerator of $(\ref{G^* equation})$ goes to $d'_0 = \frac{a - 1}{b-1} - \left| \frac{a-1}{b-1}\right|$. Notice that when the sign of $a-1$ and $b-1$ are the same (or $a = 1$), $d'_0 = 0$. Thus, taking a Laurent expansion of $G^*(z,a,b)$ at $z = \frac{1}{b-1}$, we get $$G^*(z,a,b) = \frac{1}{2(1 - (b-1)z)}\sum_{n \geq 0} d'_n\left(z - \frac{1}{b-1}\right)^n = \sum_{n \geq -1} \frac{d'_{n+1}}{2(1-b)}\left(z - \frac{1}{b-1}\right)^{n},$$ where $d'_i$ are constants. Notice that when $d'_0 = 0$, $G^*(z,a,b)$ is analytic at $z = \frac{1}{b-1}$. Thus, the dominant singularity of $G^*(x,a,b)$ is $\frac{1}{\rho}$. When $d'_0 \neq 0$, $G^*(z,a,b)$ has a singularity at $z = \frac{1}{b-1}$. For this to be the case, we must have that $a - 1$ and $b-1$ have the different signs. Notice that for $z = \frac{1}{b-1}$ to be the dominant singularity, $\rho^2 < (b-1)^2$, which implies that $a + 2b + 2\sqrt{a} < 1$. Since $a - 1$ and $b-1$ have different signs, one of $a$ and $b$ is at least 1, thus the inequality cannot hold. We thus conclude that the dominant singularity of $G^*(x,a,b)$ in this case is also $\frac{1}{\rho}$.
\end{proof}
\begin{corollary} \label{asymp normalizing const lemma}
Fix $\alpha, \beta \in \mathbb{R}$. Let $\rho = e^{-\alpha} + e^{-\beta} + 2e^{-\frac{\alpha}{2}}$. The following estimate holds.
$$\mathcal{Z}_{(n,\alpha,\beta,0)} = \frac{\sqrt{ e^{-\frac{\alpha}{2}} \rho }}{2\sqrt{\pi}} \cdot \rho^n \cdot n^{-\frac{3}{2}} + O\left(n^{-\frac{5}{2}}\right) $$
\end{corollary}
\begin{proof}
Let $\overline{\rho} = e^{-\alpha} + e^{-\beta} - 2e^{-\frac{\alpha}{2}}$. By definition, it should be clear that $\mathcal{Z}_{(n,\alpha,\beta,0)} = [x^n]G(x,e^{-\alpha}, e^{-\beta})$. Thus, from Corollary \ref{eval G func}, for $n \geq 2$, $$\mathcal{Z}_{(n-1,\alpha,\beta,0)} = [x^n]\left(-\frac{\sqrt{1 - \overline{\rho} x}}{2} \cdot \sqrt{1 - \rho x}\right).$$ By Lemma \ref{dom sig lemma}, for all $\alpha, \beta \in \mathbb{R}$, $\frac{1}{\rho} < \frac{1}{\overline{\rho}}$, thus $\sqrt{1 - \overline{\rho} x}$ is analytic on the disk $R = \left\{z \in \mathbb{C}: |z| \leq \frac{1}{\rho}\right\}$. We thus apply Corollary \ref{analytic combo corollary 1}, to see that
\begin{eqnarray}
\mathcal{Z}_{(n-1,\alpha,\beta,0)} &=& -\frac{\rho^n}{2\Gamma\left(-\frac{1}{2}\right)} \cdot \left( \sqrt{\frac{ 4e^{-\frac{\alpha}{2}}}{\rho}} \right)n^{-\frac{3}{2}} + O\left(n^{-\frac{5}{2}}\right) \nonumber\\
&=& \frac{\rho^n}{2\sqrt{\pi}} \cdot \sqrt{ \frac{ e^{-\frac{\alpha}{2}}}{\rho}} \cdot n^{-\frac{3}{2}} + O\left(n^{-\frac{5}{2}}\right).
\end{eqnarray}
We now set $n$ to $n + 1$ to get the desired result.
\end{proof}
\bigskip
We will now extract the generating function $G_n(a,b)$ defined by $$G_n(a,b) = \sum_{d_0 = 0}^\infty\sum_{d_1 = 0}^\infty\sum_{r = 0}^\infty |\mathcal{G}_n(d_0, d_1, r)|a^{d_0}b^{d_1}$$ using the following technical lemma. We will defer the proof of the lemma.
For a continuously differentiable function $F(x_1, \cdots, x_k)$ and $V$, a finite multiset with elements $v_1, \cdots, v_m \in \{x_1, \cdots, x_k\}$, define $$\frac{\partial F(x_1, \cdots, x_k)}{\partial V} = \frac{\partial^m F(x_1, \cdots, x_k)}{\partial v_1 \cdots \partial v_m}.$$ For a set $V$, let $\text{Part}(V)$ be the set of unordered partitions on $V$, $(V_i)$. For a multiset, $V$, we define $\text{Part}(V)$ by distinguishing all elements of $V$, taking the partitions of the induced set, then removing the distinction from each of the repeated elements. Note that $\text{Part}(V)$ is itself a multiset. For example, $\text{Part}(\{x_1, x_1\}) = \{\{\{x_1\}, \{x_1\}\}, \{\{x_1, x_1\}\}\}$.
\begin{lemma} \label{deriv lemma}
Let $F$ be a continuously differentiable function in $x$, $\Delta$ be a continuously differentiable function in $x_1, \cdots, x_k$ and $V$ be a non-empty finite set or multiset of the elements $x_1, \cdots, x_k$ with elements $v_1, \cdots, v_m$.
\begin{equation}
\frac{\partial F(\Delta)}{\partial V} = \sum_{(V_i) \in \text{Part}(V)} \frac{\partial \Delta}{\partial V_1} \cdots \frac{\partial \Delta}{\partial V_p} \cdot \left.\frac{\partial^p F(x)}{\partial x^p}\right|_{x = \Delta}
\end{equation}
\end{lemma}
We now achieve the following expression for $G_n(a,b)$.
We also compute the coefficients of the above expression explicitly via a combinatorial argument in Corollary \ref{count trees by leaves and internal}.
\begin{corollary} For $n \geq 2$,
$$G_n(a,b) = \frac{1}{2^{n+1}} \sum_{0 \leq k \leq \frac{n+1}{2}} C_{n-k}\binom{n-k+1}{k} \cdot \left(a+b\right)^{n-2k+1} \cdot \left(4a - (a+b)^2\right)^k.$$
\end{corollary}
\begin{proof}
We consider $xG(x,a,b)$. Notice that for $n \geq 2$, $$\frac{\partial^n xG(x,a,b)}{\partial x^n} = -\frac{1}{2} \cdot \frac{\partial \sqrt{\Delta}}{\partial V}$$ where $V$ is the multiset containing $n$ copies of $x$ and $\Delta = (1+(2-a-b)x)^2 - 4x(1 - (b-1)x)$. Thus we apply Lemma \ref{deriv lemma}. We notice that if $|V_i| > 2$, $\frac{\partial \Delta}{\partial V_i} = 0$. The number of elements in $\text{Part}(V)$ in which each part has size is at most 2 and there are $l$ parts of size $1$ and $k$ parts of size 2 is $\frac{n!}{2^k \cdot k! \cdot l!}.$ Thus, for $n \geq 2$,
\begin{eqnarray}\left.\frac{\partial^n xG(x,a,b)}{\partial x^n}\right|_{x =0} &=& \left.\frac{1}{2} \sum_{2k+l = n} \frac{n!}{2^k \cdot k! \cdot l!} \cdot \left(\frac{\partial \Delta}{\partial x}\right)^l \cdot \left(\frac{\partial^2 \Delta}{\partial x^2}\right)^k \cdot \frac{\left(-\frac{1}{2}\right)^{k+l}\cdot (2(k+l)-3)!!}{\Delta^\frac{2(k+l) - 1}{2}}\right|_{x =0} \nonumber\\
&=& \frac{n!}{2^n} \sum_{0 \leq k \leq \frac{n}{2}} C_{n-k-1}\binom{n-k}{k} \cdot \left(a+b\right)^{n-2k} \cdot \left(4a - (a+b)^2\right)^k.
\end{eqnarray}
Finally, notice that \begin{eqnarray}
\left.\frac{\partial^n xG(x,a,b)}{\partial x^n}\right|_{x =0} &=& n!\sum_{d_0 = 0}^\infty\sum_{d_1 = 0}^\infty\sum_{r = 0}^\infty |\mathcal{G}_{n-1}(d_0, d_1, r)|a^{d_0}b^{d_1}.
\end{eqnarray}
\end{proof}
\subsection{The Distributions of Simple Subtree Additive Properties}
\label{Distributions of Simple Subtree Additive Properties}
In this section, we will consider various additive properties. We will assume all the tolls in this section are of the form $(f, 0)$ since, from Example \ref{Example number of leaves and edges} and (\ref{additve prop linearity}), we see that for any $c \in \mathbb{Z}_{\geq 0}$ and $T \in \mathfrak{T}_{n}$,
\begin{equation}
\mathcal{P}^{(f, c)}(T) = \mathcal{P}^{(f, 0)}(T) + c \cdot \left(\mathcal{P}^{e}(T) + 1\right) = \mathcal{P}^{(f + c, 0)}(T) + c,
\end{equation} where $(f+c)(T) = f(T) + c$.
Let $\mathcal{P}$ be a non-negative integer valued additive property of plane trees. Let $\mathcal{P}^*$ be the subtree additive property induced by $\mathcal{P}$. Such a subtree additive property is simple as we have defined. The main result of this section is that if the toll function of $\mathcal{P}$ is bounded, the limiting distribution of $\mathcal{P^*}$ is determined by the limiting distribution of $\mathcal{P}$. Our primary result is stated as follows.
\begin{theorem} \label{mainSimpleAdditveTheorem1}
Let $\mathcal{P}_1$ and $\mathcal{P}_2$ be non-negative integer valued additive properties of plane trees with toll functions $f_1$ and $f_2$, respectively. Let the subtree additive properties induced by $\mathcal{P}_1$ and $\mathcal{P}_2$ be $\mathcal{P}^*_1$ and $\mathcal{P}^*_2$, respectively. Further assume that there exists $\zeta \in \mathbb{N}$ such that for all $T \in \mathfrak{T}_{\geq 0}$, $f_1(T) \leq \zeta$ and $f_2(T) \leq \zeta$. If, for all $m,n \in \mathbb{Z}$, \begin{equation} \label{main condition on theorem sec 2}
\sum_{T \in \mathfrak{T}_n} \mathcal{P}_1(T)^m = \mu^m \cdot \sum_{T \in \mathfrak{T}_n} \mathcal{P}_2(T)^m + O\left(n^{\frac{2m-4}{2}}4^n\right)
\end{equation} where $\mu \in \mathbb{R}$ is a constant, then as $n \rightarrow \infty$, $$\frac{\mathcal{P}^*_1(\mathfrak{T}_n)}{\sqrt{n^3}} \overset{d}{\longleftrightarrow} \mu \cdot \frac{\mathcal{P}^*_2(\mathfrak{T}_n)}{\sqrt{n^3}}.$$
\end{theorem}
\bigskip
Let $\mathcal{P}$ be a non-negative integer valued property of plane trees that is additive with toll function $f$. Let $\mathcal{F}(n, m)$ be the set of trees, $T$, on $n$ edges such that $\mathcal{P}(T) = m$. Let $$F(x,p) = \sum_{n \geq 0} \sum_{m \geq 0} |\mathcal{F}(n, m)| x^np^m.$$ We may also refer to $F(x,p)$ by $F_{\mathcal{P}}(x,p)$ where the property we are referring to is unclear.
We now let $\mathcal{H}^v(n, m)$ be the set of trees, $T$, on $n$ edges such that $\mathcal{P}(T) = m$ and $f(T) = v$. Let
$$H^v(x,p) = \sum_{n \geq 0} \sum_{m \geq 0} |\mathcal{H}^v(n, m)| x^np^m.$$
\begin{lemma} \label{F(x,p) recurrence}
For any fixed non-negative integer valued additive property of plane trees, the following recurrence holds.
\begin{equation}
F(x,p) = 1 + xF(x,p)\sum_{v\geq 0} p^v H^v(x,p)
\end{equation}
\end{lemma}
\begin{proof}
We apply the decomposition of a plane tree into 2 subtrees given by $T = T_1 \ltimes T_2$. This adds a new edge. Recall that $\mathcal{P}(T) = \mathcal{P}(T_1) + \mathcal{P}(T_2) + f(T_2)$. Thus, the tree $T$ gains an extra $p^{f(T_2)}$ in the weighting. When $T_2 \in \mathcal{H}^v(n, m)$, $T$ gains an extra $p^{v}$ in the weighting. The only tree not accounted for by this process is $T^*$.
\end{proof}
\begin{lemma} \label{bound on F(x,p) lemma}
Let $\mathcal{P}$ be a non-zero non-negative integer valued additive property with toll function $f$ such that $f(T) \leq \zeta$ for all $T \in \mathfrak{T}_{\geq 0}$ where $f$ achieve $\zeta$. For all $n, m \geq 0$, $$\sum_{T \in \mathfrak{T}_n} \mathcal{P}(T)^m = \Theta\left(n^{\frac{2m-3}{2}}4^n\right).$$
\end{lemma}
\begin{proof}
Fix $m \geq 0$. Assume $\zeta > 0$ (otherwise $\mathcal{P}(T) = 0$ for all $T \in \mathfrak{T}_{\geq 0}$). We first note that
\begin{eqnarray}
\frac{\partial F_{\mathcal{P}}(x,1)}{\partial p^m} &=& \sum_{n \geq 0}x^n \sum_{T \in \mathfrak{T}_n} \mathcal{P}(T)(\mathcal{P}(T) - 1) \cdots (\mathcal{P}(T) - m + 1) \nonumber \\
&=& \sum_{n \geq 0}x^n \sum_{T \in \mathfrak{T}_n} \mathcal{P}(T)^m + O\left(\mathcal{P}(T)^{m-1}\right) \label{eq 10}
\end{eqnarray}
Consider $\mathcal{P}_1$, the additive property with toll function $f_1(T) = \zeta$ for all $T \in \mathfrak{T}_{\geq 0}$. Let $T' \in \mathfrak{T}_N$ be a tree such that $f(T') = \zeta$. Consider $\mathcal{P}_2$, the additive property with toll function $f_2(T') = \zeta$ and $f_2(T) = 0$ for all other $T \in \mathfrak{T}_{\geq 0}$.
It should be clear that for any tree $T \in \mathfrak{T}_{\geq 0}$,
\begin{equation} \label{eq 11}
0 \leq \mathcal{P}_2(T) \leq \mathcal{P}(T) \leq \mathcal{P}_1(T).
\end{equation}
We now see that $F_{\mathcal{P}_1}(x,p) = 1 + xp^\zeta F_{\mathcal{P}_1}(x,p)^2$ and $F_{\mathcal{P}_2}(x,p) = 1 + xF_{\mathcal{P}_2}(x,p)^2 + (p^\zeta - 1)x^N$, thus $$F_{\mathcal{P}_1}(x,p) = \frac{1 - \sqrt{1 - 4xp^\zeta}}{2xp^\zeta} \quad\quad \text{and} \quad\quad F_{\mathcal{P}_2}(x,p) = \frac{1 - \sqrt{1 - 4x(1 +(p^\zeta - 1)x^N) }}{2x}.$$
Thus, for $m \geq 1$, $$\frac{\partial F_{\mathcal{P}_1}(x,1)}{\partial p^m} = c_1 \cdot (1 - 4x)^{\frac{1-2m}{2}} + O\left((1 - 4x)^{\frac{3 - 2m}{2}}\right)$$
and
$$\frac{\partial F_{\mathcal{P}_2}(x,1)}{\partial p^m} = c_2 \cdot (1 - 4x)^{\frac{1 - 2m}{2}} + O\left((1 - 4x)^{\frac{3 - 2m}{2}}\right),$$ where $c_1, c_2$ are constants (that depend on $m$). We now apply Corollary \ref{analytic combo corollary 1} to see that $$[x^n]\frac{\partial F_{\mathcal{P}_1}(x,1)}{\partial p^m} = c'_1 \cdot n^{\frac{2m - 3}{2}} 4^n (1 + o(1))\quad\quad \text{and} \quad\quad [x^n]\frac{\partial F_{\mathcal{P}_2}(x,1)}{\partial p^m} = c'_2 \cdot n^{\frac{2m - 3}{2}} 4^n (1 + o(1))$$ where $c'_1, c'_2$ are constants. From (\ref{eq 10}) and (\ref{eq 11}), we thus see that for all $n \geq 0$, $$c'_2 \cdot n^{\frac{2m - 3}{2}} 4^n (1 + o(1)) = \sum_{T \in \mathfrak{T}_n} \mathcal{P}_2(T)^m \leq \sum_{T \in \mathfrak{T}_n} \mathcal{P}(T)^m \leq \sum_{T \in \mathfrak{T}_n} \mathcal{P}_1(T)^m = c'_1 \cdot n^{\frac{2m - 3}{2}} 4^n (1 + o(1)).$$
\end{proof}
\begin{lemma}
Let $\mathcal{P}$ be a non-zero non-negative integer valued additive property with toll function $f$ such that $f(T) \leq \zeta$ for all $T \in \mathfrak{T}_{\geq 0}$ where $f$ achieve $\zeta$. As $n \rightarrow \infty$, the limiting distribution of $\frac{\mathcal{P}(\mathfrak{T}_n)}{n}$ is uniquely determined by its moments.
\end{lemma}
\begin{proof}
Consider $\mathcal{P}_1$, the additive property with toll function $f$ such that $f(T) = \zeta$ for all $T \in\mathfrak{T}_{\geq 0}$. We also see that for all $T \in\mathfrak{T}_{\geq 0}$, $\mathcal{P}(T) \leq \mathcal{P}_1(T)$. From Example \ref{Example number of leaves and edges} and (\ref{additve prop linearity}), for $T \in\mathfrak{T}_{n}$, $\mathcal{P}_1(T) = \zeta n$. Thus $\mathbb{E}\left[\left(\frac{\mathcal{P}(\mathfrak{T}_n)}{n}\right)^k\right] \leq \mathbb{E}\left[\left(\frac{\mathcal{P}_1(\mathfrak{T}_n)}{n}\right)^k\right] = \zeta^k$. The result follows immediately by the Carleman's condition (Theorem \ref{Carleman's condition}).
\end{proof}
\bigskip
The above result tells us that the condition in (\ref{main condition on theorem sec 2}) implies that $\frac{\mathcal{P}_1(\mathfrak{T}_n)}{n} \overset{d}{\longleftrightarrow} \mu \cdot \frac{\mathcal{P}_2(\mathfrak{T}_n)}{n}$. Thus Theorem \ref{mainSimpleAdditveTheorem1} is equivalent to the following corollary.
\begin{corollary} \label{mainSimpleAdditveTheorem2}
For non-negative integer valued additive properties, $\mathcal{P}_1$ and $\mathcal{P}_2$, with toll functions $f_1$ and $f_2$, respectively, such that there exists $\zeta \in \mathbb{N}$ such that for any $T \in \mathfrak{T}_{\geq 0}$, $f_1(T) \leq \zeta$ and $f_2(T) \leq \zeta$,
$$\frac{\mathcal{P}_1(\mathfrak{T}_n)}{n} \overset{d}{\longleftrightarrow} \mu \cdot\frac{\mathcal{P}_2(\mathfrak{T}_n)}{n} \Rightarrow \frac{\mathcal{P}^*_1(\mathfrak{T}_n)}{\sqrt{n^3}} \overset{d}{\longleftrightarrow} \mu \cdot\frac{\mathcal{P}^*_2(\mathfrak{T}_n)}{\sqrt{n^3}}$$ where $\mu \in \mathbb{R}$ is constant and $\mathcal{P}^*_1$ and $\mathcal{P}^*_2$ are the induced subtree additive properties.\\
\end{corollary}
\bigskip
Let $\mathcal{P^*}$ be the subtree additive property induced by $\mathcal{P}$. Let $\mathfrak{T}_ {n, m}$ be the set of plane trees, $T$, on $n$ edges such that $\mathcal{P}(T) = m$. We define
\begin{eqnarray}
M_{k,n, m} &=& \sum_{T \in \mathfrak{T}_{n, m}} \mathcal{P^*}(T)^k = \sum_{T \in \mathfrak{T}_{n, m}}\left(\sum_{v \in \overline{\mathcal{V}}(T)} \mathcal{P}(T_v)\right)^k \nonumber \\
&=& \sum_{T \in \mathfrak{T}_{n, m}}\sum_{(v_1, \cdots, v_k) \in \overline{\mathcal{V}}(T)^k} \mathcal{P}(T_{v_1}) \cdots \mathcal{P}(T_{v_k}) \label{M_{k,n, m} sum}
\end{eqnarray}
and let
\begin{equation}
M_{k}(x, p) = \sum_{n, m \geq 0} M_{k,n, m}x^np^m
\end{equation}
and $M_k(x) = M_{k}(x, 1)$. Note that $[x^n]M_k(x) = \sum_{T \in \mathfrak{T}_n }\mathcal{P^*}(T)^k = \mathbb{E}\left[\mathcal{P^*}(\mathfrak{T}_n)^k\right] \cdot C_n$, where $C_n$ is the $n$th Catalan number. \\
Fix $t_1, \cdots, t_k \in \mathbb{Z}_{\geq 0}$ and $n,m \geq 0$. We consider tuples of $(v_1, \cdots, v_k) \in \overline{\mathcal{V}}(T)$ for some $T \in \mathfrak{T}_{n,m}$ where $\mathcal{P}(T_{v_i}) = t_i$. Notice that the contribution to $M_{k,n, m}$ of any such tuples is $\prod_{i=1}^kt_i$. Also notice that the number of such tuples is $$\sum_{T \in \mathfrak{T}_{n, m}}\prod_{i=1}^k W(T, t_i)$$
where $W(T, t) = \left|\left\{v \in \overline{\mathcal{V}}(T): \mathcal{P}(T_v) = t \right\}\right|$. Let $\mathcal{F}_k(n, m, \vec{t}, \vec{s})$ be the set of trees, $T$, on $n$ edges such that $\mathcal{P}(T) = m$ and $W(T, t_i) = s_i$ where $\vec{s} = (s_1, \cdots, s_k) \in \mathbb{Z}_{\geq 0}^k$ and $\vec{t} = (t_1, \cdots, t_k) \in \mathbb{Z}_{\geq 0}^k$. We now let
$$F_k(\vec{t}|x, \vec{y}|m) = \sum_{n \geq 0} \sum_{s_1 \geq 0} \cdots \sum_{s_k \geq 0}|\mathcal{F}_k(n, m, \vec{t}, \vec{s})| x^n\prod_{i=1}^k y_i^{s_i}$$ and
$$F_k(\vec{t}|x, p, \vec{y}) = \sum_{m \geq 0} F_k(\vec{t}|x, \vec{y}|m) p^m.$$
Note that $$\frac{\partial F_k(\vec{t}|x, p, \vec{1})}{\partial y_1 \cdots \partial y_k} = \sum_{n \geq 0} \sum_{m \geq 0} \sum_{s_1 \geq 0} \cdots \sum_{s_k \geq 0}|\mathcal{F}_k(n, m, \vec{t}, \vec{s})| x^np^m\prod_{i=1}^k {s_i}$$ where $\vec{1}$ the vector of appropriate size for the context consisting of all 1s.
For $V = \{v_1, \cdots, v_l\} \subset \mathbb{N}$ and $F$, a function, we define $$\frac{\partial F}{\partial y(V)} = \frac{\partial F}{\partial y_{v_1} \cdots \partial y_{v_l}}.$$ We define $\frac{\partial F}{\partial z(V)}$ similarly. We also denote the set of integers from 1 to $k$ by $[k]$. We now show the following lemma that will be integral to the rest of our analysis.
\begin{lemma} For $k \in \mathbb{N}$ and $V = \{v_1, \cdots, v_l\} \subset [k]$ such that $|V| = l \leq k$, the following holds.
\begin{equation}
M_l(x, p) = \left.\frac{\partial}{\partial z(V)}\left(\sum_{t_{v_1} \geq 0} \cdots \sum_{t_{v_l} \geq 0}\frac{\partial F_k(\vec{t}|x, p, \vec{1})}{\partial y(V)} \prod_{i=1}^lz^{t_{v_i}}_{v_i}\right)\right|_{\vec{z} =\vec{1}}.
\end{equation}
\label{M_l(x, p) lemma}
\end{lemma}
\begin{proof} Let $W = [k]-V = \{w_1, \cdots, w_{k-l}\}$. We first see that \begin{eqnarray}
\frac{\partial F_k(\vec{t}|x, p, \vec{1})}{\partial y(V)} &=& \sum_{n \geq 0} \sum_{m \geq 0} \sum_{s_1 \geq 0} \cdots \sum_{s_k \geq 0}|\mathcal{F}_k(n, m, \vec{t}, \vec{s})| x^np^m\prod_{i=1}^l {s_{v_i}} \nonumber\\
&=& \sum_{n \geq 0} \sum_{m \geq 0} \sum_{s_{v_1} \geq 0} \cdots \sum_{s_{v_l}\geq 0} x^np^m\prod_{i=1}^l {s_{v_i}} \sum_{s_{w_1} \geq 0} \cdots \sum_{s_{w_{k-l}} \geq 0} |\mathcal{F}_k(n, m, \vec{t}, \vec{s})| \label{equat 1}\\
&=& \sum_{n \geq 0} \sum_{m \geq 0} \sum_{s_{v_1} \geq 0} \cdots \sum_{s_{v_l}\geq 0}|\mathcal{F}_l(n, m, \vec{t}, \vec{s})| x^np^m \prod_{i=1}^l {s_{v_i}} \label{equat 2}\\
&=& \frac{\partial F_l(\vec{t}|x, p, \vec{1})}{\partial y(V)} \label{equat 3}
\end{eqnarray}
We go from (\ref{equat 1}) to (\ref{equat 2}) by noticing that when we sum $|\mathcal{F}_k(n, m, \vec{t}, \vec{s})|$ over $s_{w_{1}} \cdots s_{w_{k-l}}$, we lose our dependency on $t_{w_1}, \cdots, t_{w_{k-l}} $. We go from (\ref{equat 2}) to (\ref{equat 3}) by relabeling the indices from $v_1, \cdots, v_l$ to $1,\cdots, l$. Thus to prove the lemma, we need only consider the case when $V = [k]$. We now see that
\begin{multline}
\left.\frac{\partial}{\partial z([k])}\left(\sum_{t_{v_1} \geq 0} \cdots \sum_{t_{v_l} \geq] 0}\frac{F_k(\vec{t}|x, p, \vec{1})}{\partial y([k])} \prod_{i=1}^lz^{t_{v_i}}_{v_i}\right)\right|_{\vec{z} =\vec{1}}\\
= \sum_{n \geq 0} \sum_{m \geq 0} \sum_{t_1 \geq 0}\sum_{s_1 \geq 0} \cdots \sum_{t_k \geq 0}\sum_{s_k \geq 0}|\mathcal{F}_k(n, m, \vec{t}, \vec{s})| x^np^m\prod_{i=1}^k s_it_i.
\end{multline}
For fixed $\vec{t}, \vec{s}$, we notice that trees in $\mathcal{F}_k(n, m, \vec{t}, \vec{s})$ are precisely the trees from which we can get $(v_1, \cdots, v_k) \in \overline{\mathcal{V}}(T)$ for some $T \in \mathfrak{T}_{n,m}$ where $\mathcal{P}(T_{v_i}) = t_i$. The contribution of each tuple to $M_{k,n, m}$ is $\prod_{i=1}^k t_i$. The number of tuples $(v_1, \cdots, v_k)$ that can be achieved from each tree $T \in \mathcal{F}_k(n, m, \vec{t}, \vec{s})$ is $\prod_{i=1}^k s_i$. Thus the total (weighted) contribution from tuples of the above form is $$|\mathcal{F}_k(n, m, \vec{t}, \vec{s})|x^np^m\prod_{i=1}^k s_it_i.$$ Hence, summing over $\vec{t}, \vec{s} \in \mathbb{Z}^k_{\geq 0}$ and $n,m \in \mathbb{Z}_{\geq 0}$, we get $M_k(x,p)$, proving the result.
\end{proof}
\bigskip
Let $\mathcal{H}_k^v(n, m, \vec{t}, \vec{s}) = \mathcal{H}^v(n, m) \cap \mathcal{F}_k(n, m, \vec{t}, \vec{s})$,
\begin{equation}
H^v_k(\vec{t}|x, \vec{y}|m) = \sum_{n \geq 0} \sum_{s_1 \geq 0} \cdots \sum_{s_k \geq 0}|\mathcal{H}_k^v(n, m, \vec{t}, \vec{s})| x^n\prod_{i=1}^k y_i^{s_i},
\end{equation}
\begin{equation}
H^v_k(\vec{t}|x, p, \vec{y}) = \sum_{m \geq 0} H^v_k(\vec{t}|x, \vec{y}|m) p^m
\end{equation} and
\begin{equation}
J^v_k(x, p) = \left.\frac{\partial}{\partial z([k])}\left(\sum_{t_{1} \geq 0} \cdots \sum_{t_{k} \geq 0}\frac{ H^v_k(\vec{t}|x, p, \vec{1})}{\partial y([k])} \prod_{i=1}^kz^{t_{i}}_{i}\right)\right|_{\vec{z} =\vec{1}}.
\end{equation}
Using a similar argument to Lemma \ref{M_l(x, p) lemma}, we get that for $V = \{v_1, \cdots, v_l\} \subset [k]$ such that $|V| = l \leq k$,
\begin{equation} \label{J^v_l def}
J^v_l(x, p) = \left.\frac{\partial}{\partial z(V)}\left(\sum_{t_{v_1} \geq 0} \cdots \sum_{t_{v_l} \geq 0}\frac{H^v_k(\vec{t}|x, p, \vec{1})}{\partial y(V)} \prod_{i=1}^lz^{t_{v_i}}_{v_i}\right)\right|_{\vec{z} =\vec{1}}.
\end{equation}
Note that \begin{equation} \label{equat 5}
\sum_{v \geq 0} H^v_k(\vec{t}|x, p, \vec{y}) = F_k(\vec{t}|x, p, \vec{y}) \quad\quad\quad\quad \text{and} \quad\quad\quad\quad \sum_{v \geq 0} J^v_k(x, p) = M_k(x, p).
\end{equation}
We define a partition of a set to be a set of disjoint subsets of the original set whose union is the original set. We call these subsets parts. We denote a partition of $S$ into $\lambda$ parts by $(S_i)_\lambda$, where $S_1, \cdots, S_\lambda$ are the parts of the partition. We say a partition of $S$, $(S_i)_\lambda$, refines another partition of $S$, $(S'_i)_\mu$ if for any $S_i$, there is a $S'_j$ such that $S_i \subset S'_j$. We denote this by $(S_i)_\lambda \subset (S'_i)_\mu$.
Let $\vec{t} \in \mathbb{Z}_{\geq 0}^k$. Notice that $\vec{t}$ induces a partition of $S = [k]$ as follows. Let $(S_i)^{\vec{t}}_\lambda$ be such that the numbers $i, j$ are in the same part if and only if $t_i = t_j$. For a fixed $(S_i)^{\vec{t}}_\lambda$, let $t_i^*=t_j$ such that $j \in S_i$.
\begin{lemma} The following recurrence holds.
\begin{eqnarray} \label{Full F_k recurrence}
F_k(\vec{t}|x, p, \vec{y}) &=& 1 + x\sum_{v\geq 0} p^v F_k(\vec{t}|x, p, \vec{y})H_k^v(\vec{t}|x,p, \vec{y}) \nonumber \\
&& \quad\quad\quad\quad + xF_k(\vec{t}|x, p, \vec{y})\sum_{v\geq 0} \sum_{i = 1}^\lambda \left(\prod_{j \in S_i} y_j - 1\right) \cdot p^{t_i^*} \cdot p^v \cdot H_k^v(\vec{t}|x, \vec{y}|t_i^*), \label{full expansion sum 1}
\end{eqnarray} where the $S_i$ are the parts in $(S_i)^{\vec{t}}_\lambda$.
\end{lemma}
\begin{proof}
We apply the decomposition of a plane tree into 2 subtrees given by $T = T_1 \ltimes T_2$. By similar argument to Lemma \ref{F(x,p) recurrence}, we properly weight with respect to $n$ and $m$. To properly weight with respect to $\vec{y}$, we notice that the number of non-root vertices $v$ with $\mathcal{P}(T_v) = t$, for some $t$, in $T$ is the sum of the number of such vertices in $T_1$ and $T_2$. Additionally, when $\mathcal{P}(T_2) = t$, we get an extra such vertex (the root of $T_2$). When $\mathcal{P}(T_2) = t^*_i$, we get an extra vertex, $v$, where $\mathcal{P}(T_v) = t_{j} = t^*_i$, where $j \in S_i$. Notice that these are the trees counted by $\sum_{v \geq 0}H_k^v(\vec{t}|x, \vec{y}|t_i^*)$. These trees should get an extra $\prod_{j \in S_i} y_j$ in the weighting. We however notice that the weight with respect to $p$ of such trees is $p^{t_i^*}$. Trees counted by $H_k^v(\vec{t}|x, \vec{y}|t_i^*)$ also get an extra $p^v$ (to properly weight the entire tree with respect to $p$).
\end{proof}
Differentiating both sides of (\ref{Full F_k recurrence}), we see that
\begin{multline} \label{full expansion sum 2}
\frac{\partial F_k(\vec{t}|x, p, \vec{1})}{\partial y([k])} = x\sum_{V \subset [k]} \frac{\partial F_k(\vec{t}|x, p, \vec{1})}{\partial y([k]-V)} \sum_{v \geq 0} p^v \cdot \frac{\partial H_k^v(\vec{t}|x,p, \vec{1})}{\partial y(V)}\\
+ x\sum_{V \subset [k]} \frac{\partial F_k(\vec{t}|x, p, \vec{1})}{\partial y([k]-V)} \cdot \left. \frac{\partial}{\partial y(V)}\left(\sum_{v \geq 0}\sum_{i = 1}^\lambda \left(\prod_{j \in S_i} y_j - 1\right) \cdot p^{t_i^*} \cdot p^v \cdot H_k^v(\vec{t}|x, \vec{y}|t_i^*)\right)\right|_{\vec{y} = \vec{1}}.
\end{multline}
For $V = \{v_1, \cdots, v_l\} \subset [k]$, we let $\vec{t}(V) = (t_{v_1}, \cdots, t_{v_{l}})$. Using Lemma \ref{M_l(x, p) lemma}, we simplify
$$\Phi_k^{(1)} = \frac{\partial}{\partial z([k])}\sum_{\vec{t} \in \mathbb{Z}_{\geq 0}^k}\left(\sum_{V \subset [k]} \frac{\partial F_k(\vec{t}|x, p, \vec{1})}{\partial y([k]-V)} \sum_{v \geq 0} p^v \cdot \frac{\partial H_k^v(\vec{t}|x,p, \vec{1})}{\partial y(V)}\right) \prod_{i \in [k]}z_i^{t_i}$$ as follows.
\begin{eqnarray}
\Phi_k^{(1)} &=& \sum_{V \subset [k]}\frac{\partial}{\partial z(V)}\left(\sum_{\vec{t}(V) \in \mathbb{Z}_{\geq 0}^k} \frac{\partial F_k(\vec{t}|x, p, \vec{1})}{\partial y(V)}\prod_{i \in V}z_i^{t_i}\right) \times \nonumber\\
&& \quad\quad\quad \sum_{v \geq 0} p^v \cdot \frac{\partial}{\partial z([k]-V)}\left(\sum_{\vec{t}([k]-V) \in \mathbb{Z}_{\geq 0}^k}\frac{\partial H_k^v(\vec{t}|x,p, \vec{1})}{\partial y([k]-V)} \prod_{i \in [k]-V}z_i^{t_i}\right) \\
&=& \sum_{V \subset [k]}M_{|V|}(x,p) \cdot \sum_{v \geq 0} p^v \cdot J_{k - |V|}^v(x,p)
\end{eqnarray} where we set $z_i$ to 1. To simplify
\begin{multline*}
\Phi_k^{(2)} = \frac{\partial}{\partial z([k])}\sum_{\vec{t} \in \mathbb{Z}_{\geq 0}^k}\left(\sum_{V \subset [k]} \frac{\partial F_k(\vec{t}|x, p, \vec{y})}{\partial y([k]-V)} \right. \times
\\ \left.\frac{\partial}{\partial y(V)}\left(\sum_{v \geq 0} \sum_{i = 1}^\lambda \left(\prod_{j \in S_i} y_j - 1\right) \cdot p^{t_i^*} \cdot p^v \cdot H_k^v(\vec{t}|x, \vec{y}|t_i^*)\right) \right) \prod_{i \in V}z_i^{t_i},
\end{multline*} we will use Lemma \ref{M_l(x, p) lemma} and the following lemma (whose proof is deferred to the Appendix).
\begin{lemma}
\label{lemma expansion}
Let $\emptyset \neq W = \{w_1, \cdots, w_m\}$ for some $m \in \mathbb{N}$. For any $x_{w_1}, \cdots, x_{w_m}$, \begin{equation}
\label{lemma expansion 1}
\prod_{i=1}^m x_{w_{i}} - 1 = \sum_{V = \{v_1, \cdots, v_l\} \subset W} (x_{v_1} - 1) \cdots (x_{v_l} - 1)
\end{equation}
\end{lemma}
\bigskip
For $V \subset [k]$ and $\vec{t} \in \mathbb{Z}^k_{\geq 0}$, we let $$\Phi(\vec{t}|V) = \frac{\partial}{\partial y(V)}\left(\sum_{v \geq 0}\sum_{i = 1}^\lambda \left(\prod_{j \in S_i} y_j - 1\right) \cdot p^{t_i^*} \cdot p^v \cdot H_k^v(\vec{t}|x, \vec{y}|t_i^*)\right).$$
We apply Lemma \ref{lemma expansion} to get
\begin{equation} \label{eq 2}
\Phi(\vec{t}|V) = \sum_{v \geq 0} p^v \cdot \frac{\partial}{\partial y(V)}\left(\sum_{i = 1}^\lambda \sum_{U = \{u_1, \cdots, u_q\} \subset S_i} (y_{u_1} - 1) \cdots (y_{u_q} - 1) \cdot p^{t_i^*} \cdot H_k^v(\vec{t}|x, \vec{y}|t_i^*)\right).
\end{equation}
Our goal is to take the sum
$$\sum_{\vec{t}(V) \in \mathbb{Z}_{\geq 0}^{|V|}}\Phi(\vec{t}|V) \prod_{i \in V}z_i^{t_i}.$$
For some $U = \{u_1, \cdots, u_q\} \subset [k]$, consider $(S'_i)$, the partition of $[k]$ where the elements $u_1, \cdots, u_q$ are in the same part and all other elements are in their own parts. Notice that
\begin{equation} \label{equat 4}
(y_{u_1} - 1) \cdots (y_{u_q} - 1) \cdot p^{t_i^*} \cdot H_k^v(\vec{t}|x, \vec{y}|t_i^*)
\end{equation}
is a term in the bracket of (\ref{eq 2}) if and only if the partition induced by $\vec{t}$ is such that $(S'_i)$ is a refinement of $(S_i)_{\lambda}^{\vec{t}}$, ie. $(S_i) \subset (S_i)_{\lambda}^{\vec{t}}$.
Assume $V$ is such that $U \subset V$ (otherwise the term we are considering will vanish when we set $y_i$ to 1). Let $V-U = \{j_1, \cdots, j_l\}$. We now consider the sum
\begin{eqnarray}
\Phi(V, U) &=& \sum_{v \geq 0} p^v \cdot \frac{\partial }{\partial y({V})} \sum_{t_{j_1} \geq 0} \cdots \sum_{t_{j_l} \geq 0}(y_{u_1} - 1) \cdots (y_{u_q} - 1) \times\\
&& \quad \quad \quad \quad \sum_{t^*_{i} \geq 0} H_k^v(\vec{t}|x, \vec{y}|t_i^*) \cdot p^{t^*_{i}} (z_{u_1} \cdots z_{u_q})^{t^*_{i}} \prod_{i = 1}^lz^{t_{j_i}}_{j_i} \nonumber\\
&=& \sum_{v \geq 0} p^v \cdot \sum_{t_{j_1} \geq 0} \cdots \sum_{t_{j_l} \geq 0} \frac{\partial H^v_{k}(\vec{t}|x, p z_{u_1} \cdots z_{u_q},\vec{y})}{\partial y({V-U})} \prod_{i = 1}^lz^{t_{j_i}}_{j_i}.
\end{eqnarray}
We see that as we sum up $\Phi(\vec{t}|V) \prod_{i \in V}z^{t_{i}}_{i}$ over $\vec{t}$, we actually take the sum above for all $U \subset V$. Thus \begin{eqnarray}
\left.\frac{\partial}{\partial z(V)}\sum_{\vec{t}(V) \in \mathbb{Z}_{\geq 0}^{|V|}} \Phi(\vec{t}|V) \prod_{i \in V}z^{t_{i}}_{i}\right|_{\vec{z} = \vec{1}} &=& \left.\sum_{\emptyset \neq U \subset V} \frac{\partial \Phi(V, U)}{\partial z(V)}\right|_{\vec{z} = \vec{1}}\\
&=& \left.\sum_{\emptyset \neq U \subset V} \sum_{v \geq 0} p^v \cdot \frac{\partial J^v_{|V|-|U|}(x, p z_{u_1} \cdots z_{u_q})}{\partial z(U)}\right|_{\vec{z} = \vec{1}}\\
&=& \sum_{\emptyset \neq U \subset V} \sum_{v \geq 0} p^v \cdot \mathcal{D}_p^{|U|} J^v_{|V|-|U|}(x, p)
\end{eqnarray} where for $F$, a function in some variables including $p$, we define the operator $\mathcal{D}_p^m$ recursively as
\begin{equation} \label{D_p definition}
\mathcal{D}_p^mF = p \cdot \frac{\partial \mathcal{D}_p^{m-1} F}{\partial p} \quad\quad\quad\quad \text{ and } \quad\quad\quad\quad \mathcal{D}_p^0 F = F.
\end{equation}
Thus, we simplify $\Phi_k^{(2)} $ as follows.
\begin{eqnarray}
\Phi_k^{(2)} &=& \sum_{V \subset [k]}\frac{\partial}{\partial z([k]-V)}\left(\sum_{\vec{t}([k]-V) \in \mathbb{Z}_{\geq 0}^{k-|V|}} \frac{\partial F_k(\vec{t}|x, p, \vec{1})}{\partial y([k]-V)}\prod_{i \in [k]-V}z_i^{t_i}\right) \times \nonumber\\
&& \quad\quad\quad\quad \frac{\partial}{\partial z(V)}\sum_{\vec{t}(V) \in \mathbb{Z}_{\geq 0}^{|V|}} \Phi_V(\vec{t}) \prod_{i \in V}z^{t_{i}}_{i} \\
&=& \sum_{V \subset [k]}M_{k-|V|}(x,p) \cdot \sum_{\emptyset \neq U \subset V} \sum_{v \geq 0} p^v \cdot \mathcal{D}_p^{|U|} J^v_{|V|-|U|}(x, p)
\end{eqnarray} where we set $z_i$ to 1.
We will denote $\left.\mathcal{D}_p^{m}\left(F(x,p)\right)\right|_{p=1}$ simply as $\mathcal{D}_p^{|U|}F(x)$.
\begin{theorem} For $k \geq 1$, the following recurrence holds.
\begin{equation}
M_k(x,p) = \sum_{U \subset V \subset [k]}M_{k-|V|}(x,p) \cdot \sum_{v \geq 0} p^v \cdot \mathcal{D}_p^{|U|} J^v_{|V|-|U|}(x, p).
\end{equation}
Furthermore, for $k \geq 1$ and $m \geq 0$, let $$S(k, m) = \{(a, b, c, d) \in \mathbb{Z}^4: 0 \leq a \leq b \leq m, 0 \leq c \leq d \leq k\}$$ and $$S'(k, m) = S(k, m) - \{(0,0,0,0), (0,m,0,k)\}.$$ The following recurrence also holds.
\begin{equation}
\mathcal{D}^m_p M_k(x) = \frac{x}{\sqrt{1-4x}} \cdot \sum_{(a, b, c, d) \in S'(k, m)} \binom{k}{d}\binom{d}{c}\sum_{v \geq 0} v^a \cdot \mathcal{D}^{m-b}_pM_{k-d}(x) \cdot \mathcal{D}_p^{c + b - a} J^v_{d-c}(x). \label{equation 10}
\end{equation}
\end{theorem}
\begin{proof}
From $(\ref{full expansion sum 2})$, we see that, for $k \geq 1$,
\begin{eqnarray}
M_k(x,p) &=& x\Phi_k^{(1)} + x\Phi_k^{(2)} \nonumber\\
&=& \sum_{V \subset [k]}M_{k-|V|}(x,p) \cdot \sum_{\emptyset \neq U \subset V} \sum_{v \geq 0} p^v \cdot \mathcal{D}_p^{|U|} J^v_{|V|-|U|}(x, p) \nonumber\\
&& \quad\quad\quad\quad + \sum_{V \subset [k]} \sum_{v \geq 0} p^v \cdot M_{k - |V|}(x,p) \cdot J_{|V|}^v(x,p) \nonumber\\
&=& \sum_{U \subset V \subset [k]}M_{k-|V|}(x,p) \cdot \sum_{v \geq 0} p^v \cdot \mathcal{D}_p^{|U|} J^v_{|V|-|U|}(x, p). \nonumber
\end{eqnarray}
Differentiating the above expression, we see that
\begin{eqnarray}
\mathcal{D}^m_p M_k(x,p) &=& x\sum_{U \subset V \subset [k]}\sum_{b=0}^m \sum_{a=0}^b \sum_{v \geq 0} \mathcal{D}^a_p(p^v) \cdot \mathcal{D}^{m-b}_pM_{k-|V|}(x,p) \times \nonumber\\
&& \quad\quad\quad\quad \mathcal{D}_p^{|U| + b - a} J^v_{|V|-|U|}(x, p) \nonumber\\
\mathcal{D}^m_p M_k(x) &=& x\sum_{(a, b, c, d) \in S(k, m)} \binom{k}{d}\binom{d}{c}\sum_{v \geq 0} v^a \cdot \mathcal{D}^{m-b}_pM_{k-d}(x) \cdot \mathcal{D}_p^{c + b - a} J^v_{d-c}(x) \nonumber\\
\mathcal{D}^m_p M_k(x)(1 - 2xM_0(x)) &=& x\sum_{(a, b, c, d) \in S'(k, m)} \binom{k}{d}\binom{d}{c}\sum_{v \geq 0} v^a \cdot \mathcal{D}^{m-b}_pM_{k-d}(x) \cdot \mathcal{D}_p^{c + b - a} J^v_{d-c}(x). \nonumber
\end{eqnarray}
Notice that $M_0(x)$ is the generating function counting plane trees by number of edges. Thus $1 - 2xM_0(x) = \sqrt{1-4x}$. Thus, we arrive at the desired result.
\end{proof}
\bigskip
For simple subtree additive properties $\mathcal{P}^*$ where the property in question is not clear, we will denote the generating function for $\sum_{T \in \mathfrak{T}_n}\mathcal{P}^*(T)^k$ by $M_{k}(\mathcal{P}^*, x)$. Let $[x^n]\mathcal{D}_p^mM_k(\mathcal{P}^*, x) = M^{(m)}_{k,n}(\mathcal{P}^*)$ (or simply $M^{(m)}_{k,n}$ if there is no ambiguity). Note that $M^{(m)}_{0,n}(\mathcal{P}^*) = \sum_{T \in \mathfrak{T}_n} \mathcal{P}(T)^m$ where $\mathcal{P}$ is the additive property from which $\mathcal{P}^*$ is derived.
We now prove the following lemmas from which the main theorem of this section will follows. To do so, we utilize the following lemma (whose proof is deferred to the Appendix).
\begin{lemma}\label{tech lemma 1}
Let $a_1, a_2 \in \mathbb{R}$ and $n \in \mathbb{N}$ be large. For $\min\{a_1, a_2\} > -1 $,
$$\sum_{\substack{n_1 + n_2 = n}\\n_1, n_2 \geq 1} n_1^{a_1} \cdot n_2^{a_2} = \Theta\left(n^{a_1 + a_2 + 1}\right)$$
and, for $\min\{a_1, a_2\} < -1 $,
$$\sum_{\substack{n_1 + n_2 = n}\\n_1, n_2 \geq 1} n_1^{a_1} \cdot n_2^{a_2} = \Theta\left(n^{\max\{a_1, a_2\}}\right).$$
\end{lemma}
\noindent
\begin{lemma} \label{bound M lemma}
Let $\mathcal{P}$ be a non-negative integer valued additive property with toll function $f$ and let their induced subtree additive property be $\mathcal{P}^*$. Further assume that there exists $\zeta \in \mathbb{N}$ such that for all $T \in \mathfrak{T}_{\geq 0}$, $f(T) \leq \zeta \in \mathbb{N}$. For all $m, k$, the following holds.
$$M^{(m)}_{k,n} = \Theta\left(n^{\frac{2m + 3k - 3}{2}} 4^n\right).$$
\end{lemma}
\begin{proof}
We show this result by induction on $k$ and $m$. The base case of $k = 0$ (and any $m$) holds by Lemma \ref{bound on F(x,p) lemma}. We now consider (\ref{equation 10}). Let
\begin{eqnarray}
\Psi(k,m,a,b,c,d) &=& \frac{x}{\sqrt{1-4x}}\binom{k}{d}\binom{d}{c}\sum_{v \geq 0} v^a \cdot \mathcal{D}^{m-b}_pM_{k-d}(x) \cdot \mathcal{D}_p^{c + b - a} J^v_{d-c}(x).
\end{eqnarray} We now take the coefficients on both sides of the equation. Let $[x^n]\mathcal{D}_p^mJ^v_k(x) = J^{(m)}_{k,n}(v)$.
\begin{eqnarray}
[x^n]\Psi(k,m,a,b,c,d) &=& \binom{k}{d}\binom{d}{c}\sum_{v \geq 0} v^a \sum_{\substack{n_1 + n_2 + n_3 = n - 1\\n_1, n_2, n_3 \geq 1}} M^{(m-b)}_{k-d,n_1} \cdot J^{(c + b - a)}_{d-c,n_2}(v) \cdot \Theta\left(n_3^{-\frac{1}{2}}4^{n_3}\right) \nonumber \label{equation 23}
\end{eqnarray}
where $(a,b,c,d) \in S'(k,m)$.
Fix $k \geq 1$, $m \geq 0$. Assuming the the theorem holds for all smaller $k$ and any $m$ as well as for equal $k$ and smaller $m$. Notice that the RHS of $(\ref{equation 23})$ depends precisely on terms for which the theorem holds. We now see using Lemma \ref{tech lemma 1} that
\begin{eqnarray}
[x^n]\Psi(k,m,a,b,c,d) &\leq& \binom{k}{d}\binom{d}{c} \zeta^a \sum_{\substack{n_1 + n_2 + n_3 = n - 1\\n_1, n_2, n_3 \geq 1}} M^{(m-b)}_{k-d,n_1} \cdot M^{(c + b - a)}_{d-c,n_2}(v) \cdot O\left(n_3^{-\frac{1}{2}}4^{n_3}\right) \nonumber\\
&\leq& O\left(4^{n} \sum_{\substack{n_1 + n_2 + n_3 = n - 1\\n_1, n_2, n_3 \geq 1}} n_1^{a_1} \cdot n_2^{a_2} \cdot n_3^{-\frac{1}{2}} \right)
\end{eqnarray} where $a_1 = \frac{2(m-b) + 3(k-d) - 3}{2}$ and $a_2 = \frac{2(c+b-a) + 3(d-c) - 3}{2}$. Notice that for $(a,b,c,d) \in S'(k,m)$, $a_1, a_2 \geq -\frac{3}{2}$. We now consider the 2 possible cases:
Case 1: $(a,b,c,d) \in S'(k,m)$ such that $d = k, m = b$ or $d = c, c + b = a$. We see that $\min\{a_1, a_2\} = -\frac{3}{2}$ and $\max\{a_1, a_2\} = \frac{2m + 3k -c -2a - 3}{2}$. Thus we apply Lemma \ref{tech lemma 1} twice, noting that $\min\{a_1, a_2\}= -\frac{3}{2} < -1$, to get
\begin{eqnarray}
[x^n]\Psi(k,m,a,b,c,d) &\leq& O\left(n^{\frac{2m + 3k -c -2a - 2}{2}} 4^{n}\right).
\end{eqnarray}
Note that we get the most significant upper bound when we minimize $c + 2a$. The minimal $c + 2a$ for which there is $(a,b,c,d) \in S'(k,m)$ for some $k,m$ in this case is when $c = 1, a = 0$.
Case 2: All other cases. We see that $\min\{a_1, a_2\} > -1$. Thus we apply Lemma \ref{tech lemma 1} twice to get
\begin{eqnarray}
[x^n]\Psi(k,m,a,b,c,d) &\leq& O\left(n^{\frac{2m + 3k - c - 2a - 3}{2}} 4^{n}\right).
\end{eqnarray}
Note that we get the most significant upper bound when we minimize $c + 2a$. The minimal $c + 2a$ for which there is $(a,b,c,d) \in S'(k,m)$ in this case is when $c = 0, a = 0$. We note that when $k = 1, m = 0$ there is no $(a,b,c,d) \in S'(k,m)$ in this case.
Thus, for all $(a,b,c,d) \in S'(k,m)$, $[x^n]\Psi(k,m,a,b,c,d) \leq O\left(4^{n} n^{\frac{2m + 3k - 3}{2}}\right)$. From (\ref{equation 10}), we hence get
\begin{eqnarray}
\mathcal{D}^m_p M_k( x) &=& \sum_{(a, b, c, d) \in S'(k, m)}\Psi(k,m,a,b,c,d)\\
M^{(m)}_{k,n} &\leq & O\left(4^{n} n^{\frac{2m + 3k - 3}{2}}\right).
\end{eqnarray}
Towards the lower bound, we see when $a = 0$,
\begin{eqnarray}
[x^n]\Psi(k,m,a,b,c,d) &\geq& \binom{k}{d}\binom{d}{c} \sum_{\substack{n_1 + n_2 + n_3 = n - 1\\n_1, n_2, n_3 \geq 1}} M^{(m-b)}_{k-d,n_1}\cdot M^{(c + b - a)}_{d-c,n_2} \cdot \Theta\left(n_3^{-\frac{1}{2}}4^{n_3}\right) \nonumber\\
&\geq& \Omega\left(4^{n} \sum_{\substack{n_1 + n_2 + n_3 = n - 1\\n_1, n_2, n_3 \geq 1}} n_1^{a_1} \cdot n_2^{a_2} \cdot n_3^{-\frac{1}{2}} \right)
\end{eqnarray} where $a_1 = \frac{2(m-b) + 3(k-d) - 3}{2}$ and $a_2 = \frac{2(c+b-a) + 3(d-c) - 3}{2}$. We break this into cases exactly as before. We notice that there is $(a,b,c,d) \in S'(k,m)$ where $a = 0$ and $[x^n]\Psi(k,m,a,b,c,d) \geq \Omega\left(4^{n} n^{\frac{2m + 3k - 3}{2}}\right)$.
Thus, from (\ref{equation 10}), we get
\begin{eqnarray}
M^{(m)}_{k,n}(\mathcal{P}^*) &\geq & \Omega\left(4^{n} n^{\frac{2m + 3k - 3}{2}}\right).
\end{eqnarray} The result follows by induction.
\end{proof}
\bigskip
\begin{observation} \label{bound on psi observation}
In the proof of Lemma \ref{bound M lemma}, for any $k,m \geq 0$ and $(a,b,c,d) \in S'(k,m)$, when $a \geq 1$,
$$[x^n]\Psi(k,m,a,b,c,d) \leq O\left(n^{\frac{2m + 3k -4}{2}} 4^{n}\right).$$
\end{observation}
\bigskip
\begin{lemma} \label{c^m+k lemma}
Let $\mathcal{P}_1$ and $\mathcal{P}_2$ be non-negative integer valued additive properties with toll functions $f_1$ and $f_2$, respectively. Let the induced subtree additive properties of $\mathcal{P}_1$ and $\mathcal{P}_2$ be $\mathcal{P}^*_1$ and $\mathcal{P}^*_2$, respectively. Further assume that there exists $\zeta \in \mathbb{N}$ such that for all $T \in \mathfrak{T}_{\geq 0}$, $f_1(T) \leq \zeta \in \mathbb{N}$ and $f_2(T) \leq \zeta \in \mathbb{N}$. Assume for all $m, n \geq 0$, $$\sum_{T \in \mathfrak{T}_n} \mathcal{P}_1(T)^m = \mu^m \cdot \sum_{T \in \mathfrak{T}_n} \mathcal{P}_2(T)^m + O\left(n^{\frac{2m-4}{2}}4^n\right)$$ where $\mu$ is a constant.
It then holds that for all $n, m, k \geq 0$,
$$M^{(m)}_{k,n}(\mathcal{P}_1^*) = \mu^{k+m} \cdot M^{(m)}_{k,n}(\mathcal{P}_2^*) + O\left(n^{\frac{2m + 3k-4}{2}}4^n\right).$$
\end{lemma}
\begin{proof}
We prove the result by induction on $k$ and $m$. The base case of $k = 0$ (and any $m$) is true by assumption. Fix $k \geq 1$, $m \geq 0$. Assuming the the theorem holds for all smaller $k$ and any $m$ as well as for equal $k$ and smaller $m$. We now apply Lemma \ref{bound M lemma} and Observation \ref{bound on psi observation} regarding $\Psi(k,m,a,b,c,d)$ to get
\begin{eqnarray}
M^{(m)}_{k,n}(\mathcal{P}_1^*) &=& \sum_{(a, b, c, d) \in S'(k, m)} \binom{k}{d}\binom{d}{c}\sum_{v \geq 0} v^a \sum_{n_1 + n_2 + n_3 = n - 1} M^{(m-b)}_{k-d,n_1}(\mathcal{P}_1^*) \cdot J^{(c + b - a)}_{d-c,n_2}(\mathcal{P}_1^*, v) \cdot \binom{2n_3}{n_3}\nonumber\\
&=& \mu^{m+k}\sum_{\substack{(a, b, c, d) \in S'(k, m)\\ a =0}} \binom{k}{d}\binom{d}{c} \sum_{n_1 + n_2 + n_3 = n - 1} M^{(m-b)}_{k-d,n_1}(\mathcal{P}_2^*) \cdot M^{(c + b - a)}_{d-c,n_2}(\mathcal{P}_2^*) \cdot \binom{2n_3}{n_3}\nonumber \\
&&\quad\quad\quad + \sum_{\substack{(a, b, c, d) \in S'(k, m)\\ a \geq 1}} \Psi(k,m,a,b,c,d) + O\left(n^{\frac{2m + 3k -4}{2}} 4^{n}\right)\\
&=& \mu^{m+k} \cdot M^{(m)}_{k,n}(\mathcal{P}_2^*) + O\left(n^{\frac{2m + 3k -4}{2}} 4^{n}\right)
\end{eqnarray}
Thus the result holds for all $k, m \in \mathbb{N}$ by induction.\\
\end{proof}
\begin{lemma} \label{ssap are uniquely detrmined by moments}
Let $\mathcal{P}$ be an non-negative integer valued additive properties of plane trees with toll function $f$. Let the subtree additive property induced by $\mathcal{P}$ be $\mathcal{P}^*$. Further assume that there exist $\zeta \in \mathbb{N}$ such that for any $T \in \mathfrak{T}_{\geq 0}$, $f(T) \leq \zeta$. The limiting distribution of $$\frac{\mathcal{P}^*(\mathfrak{T}_n)}{\sqrt{n^3}}$$ is unique determined by its moments. Specifically, it satisfies the Carleman's condition.
\end{lemma}
\begin{proof} Let $\mathcal{P}_1$ be the additive property with toll function $f$ where $f(T) = \zeta$ for all $T \in \mathfrak{T}_{\geq 0}$. Let $\mathcal{P}^*_1$ be the subtree additive property derived from $\mathcal{P}_1$. We see that $\mathcal{P}(T) \leq \mathcal{P}_1(T)$, and hence, $\mathcal{P}^*(T) \leq \mathcal{P}_1^*(T)$ for all $T \in \mathfrak{T}_{\geq 0}$. From Example \ref{Example number of leaves and edges}, $\mathcal{P}_1(T) = \zeta \cdot \mathcal{P}^v(T)$, thus $\mathcal{P}_1^*(T) = \zeta \cdot \mathcal{P}^{PL}(T)$. Applying Theorem \ref{contactdistance - thm}, we see that
$$\lim_{n \rightarrow \infty}\mathbb{E}\left[\left(\frac{\mathcal{P}^*(\mathfrak{T}_{n})}{\sqrt{2n^3}} \right)^k \right] \leq \lim_{n \rightarrow \infty}\mathbb{E}\left[\left(\frac{\zeta \cdot \mathcal{P}^{PL}(\mathfrak{T}_{n})}{\sqrt{2n^3}} \right)^k \right] \: \sim \: \frac{6k}{\sqrt{2}}\left(\frac{k}{12e}\right)^{\frac{k}{2}} \cdot \zeta^k$$ as $k \rightarrow \infty$. Thus applying the Carleman Condition (Theorem \ref{Carleman's condition}), we get the desired result.
\end{proof}
\begin{proof}[Proof of Theorem \ref{mainSimpleAdditveTheorem1}]
We recall that for all $n,k \geq 0$ and $\mathcal{P}^* \in \{\mathcal{P}_1^*, \mathcal{P}_2^*\}$, $M^{(0)}_{k,n}(\mathcal{P}^*) = \sum_{T \in \mathfrak{T}_n} \mathcal{P}^*(T)^k$, hence $\mathbb{E}[\mathcal{P}^*(\mathfrak{T}_n)^k] = \frac{M^{(0)}_{k,n}(\mathcal{P}^*)}{C_n}$. We apply the assumption of the Theorem \ref{mainSimpleAdditveTheorem1}, Lemma \ref{bound M lemma} and Lemma \ref{c^m+k lemma} to achieve
\begin{eqnarray}
\mathbb{E}\left[\left(\frac{\mathcal{P}_1^*(\mathfrak{T}_n)}{\sqrt{n^3}}\right)^k\right] &=& \mu^k \cdot \mathbb{E}\left[\left(\frac{\mathcal{P}_2^*(\mathfrak{T}_n)}{\sqrt{n^3}}\right)^k\right] + O\left(\frac{1}{\sqrt{n}}\right).
\end{eqnarray}
We now apply Lemma \ref{ssap are uniquely detrmined by moments} to see that the limiting distribution of the properties are unique determined by their moments. Thus, we arrive at the desired result.
\end{proof}
\bigskip
\begin{corollary} \label{corollary P^LRD and P^IRD dist}
The total leaf to root distance of a random plane tree on $n$ edges, $\mathcal{P}^{LR}(\mathfrak{T}_n)$, and the total internal node to root distance of a random plane tree on $n$ edges, $\mathcal{P}^{IR}(\mathfrak{T}_n)$ is asymptotically Airy distributed, up to a scaling factor. Specifically, as $n \rightarrow \infty$, $$2\cdot \frac{\mathcal{P}^{LR}(\mathfrak{T}_n)}{\sqrt{2n^3}} \quad \overset{d}{\longleftrightarrow} \quad 4 \cdot \frac{\mathcal{P}^{IR}(\mathfrak{T}_n)}{\sqrt{2n^3}} \quad \overset{d}{\longleftrightarrow} \quad \frac{\mathcal{P}^{PL}(\mathfrak{T}_n)}{\sqrt{2n^3}}.$$
\end{corollary}
\begin{proof}
Let $F(x, p)$ be the generating function for plane trees weighted by number of edges and vertices. Let $F_{d_0}(x, p)$ be the generating function for plane trees weighted by number of edges and leaves. Let $F_{d_1}(x, p)$ be the generating function for plane trees weighted by number of edges and internal nodes. From Corollary \ref{eval G func},
$$F_{d_0}(x, p) = \frac{1 + (1-p)x - \sqrt{(1+(1-p)x)^2 - 4x}}{2x}$$
and
$$F_{d_1}(x, p) = \frac{1 + (1-p)x - \sqrt{(1+(1-p)x)^2 - 4x(1 - (p-1)x)}}{2x}.$$
Thus, we see that $$[x^n]\frac{\partial F_{d_0}(x, 1)}{\partial p^m} = \frac{1}{2^m} \cdot [x^n]\frac{\partial F(x, 1)}{\partial p^m} + O\left((1- 4x)^{\frac{2-2m}{2}}\right)$$
and
$$[x^n]\frac{\partial F_{d_1}(x, 1)}{\partial p^m} = \frac{1}{4^m} \cdot [x^n]\frac{\partial F(x, 1)}{\partial p^m} + O\left((1- 4x)^{\frac{2-2m}{2}}\right).$$
Thus $$\sum_{T \in \mathfrak{T}} \mathcal{P}^{d_0}(T)^m = \frac{1}{2^m} \cdot \sum_{T \in \mathfrak{T}} \mathcal{P}^{v}(T)^m + O\left(n^{\frac{2m-4}{2}}4^n\right)$$ and $$\sum_{T \in \mathfrak{T}} \mathcal{P}^{d_1}(T)^m = \frac{1}{4^m} \cdot \sum_{T \in \mathfrak{T}} \mathcal{P}^{v}(T)^m + O\left(n^{\frac{2m-4}{2}}4^n\right).$$
We now apply Theorem \ref{mainSimpleAdditveTheorem1} to get that, as $n \rightarrow \infty$,
$$2 \cdot \frac{\mathcal{P}^{LR}(\mathfrak{T}_n)}{\sqrt{2n^3}} \quad \overset{d}{\longleftrightarrow} \quad 4 \cdot \frac{\mathcal{P}^{IR}(\mathfrak{T}_n)}{\sqrt{2n^3}} \quad \overset{d}{\longleftrightarrow} \quad \frac{\mathcal{P}^{PL}(\mathfrak{T}_n)}{\sqrt{2n^3}}.$$
From Theorem \ref{contactdistance - thm}, we know that $\frac{\mathcal{P}^{PL}(\mathfrak{T}_n)}{\sqrt{2n^3}}$ converges weakly to an Airy random variable, thus the same holds for $2 \cdot \frac{\mathcal{P}^{LR}(\mathfrak{T}_n)}{\sqrt{2n^3}}$ and $4 \cdot \frac{\mathcal{P}^{IR}(\mathfrak{T}_n)}{\sqrt{2n^3}}$.
\end{proof}
\subsection{The Distribution of Classes of Subtree Additive Properties under NNTM}
\label{Dist of Subtree Additive Properties NNTM}
Let $f: \mathfrak{T}_{\geq 0} \times \mathfrak{T}_{\geq 0} \rightarrow \mathbb{R}$ be a polynomial in the number of edges, leaves and internal nodes of its input trees. Hence, $f(T', T) \in \mathbb{R}[t, \lvl, \lvi][n, \gvl, \gvi]$ where $t, \lvl, \lvi$ represents the number of edges, leaves and internal nodes of $T'$ and $n, \gvl, \gvi$ represents the number of edges, leaves and internal nodes of $T$.
We now consider $\mathcal{P}^f$, the subtree additive property given by
\begin{equation} \label{f tree def}
\mathcal{P}^{f}(T)= \sum_{v \in \overline{\mathcal{V}}(T)} f(n(T_v), d_0(T_v), d_1(T_v), n(T), d_0(T), d_1(T)) = \sum_{v \in \overline{\mathcal{V}}(T)} f(T_v, T)
\end{equation} where, for $v \in \overline{\mathcal{V}}(T)$, the root vertex in $T_v$ is allowed to count as a leaf or internal node.
Let $f$ be a polynomial over $\mathbb{R}$ in the variables $x_1, \cdots, x_n$. We can write $f$ in the form of $\sum_{j=1}^L w_j \cdot v_j$ where $w_j \in \mathbb{R}$ and $v_j$ is a monic monomial in the variables such that the monomial are all distinct. We will call this the reduced form of $f$. It should be clear that $f$ can be written uniquely in this form. For any such $v_j$ and variables $x_1, \cdots, x_l$, let $\Delta(v_j|x_1, \cdots, x_l)$ be the degree of the monomial $v_j$ if we consider all variables other than $x_1, \cdots, x_l$ to be constant. For a vector of variables $\vec{x} = (x_1, \cdots, x_l)$, let $\Delta(v_j|\vec{x}) = \Delta(v_j|x_1, \cdots, x_l)$. We also define $\Delta(f|x_1, \cdots, x_l) = \max_{1 \leq j \leq L}\Delta(v_j|x_1, \cdots, x_l)$. \\
Fix parameters $\alpha,\beta, \gamma \in \mathbb{R}$. Let
$$M_{k,n}(f, \alpha,\beta, \gamma) = \sum_{T \in \mathfrak{T}_ n} \mathcal{P}^f(T)^ke^{-E(T)} = \mathbb{E}\left[\mathcal{P}^f_{(\alpha, \beta, \gamma)}(\mathfrak{T}_n)^k\right] \cdot \mathcal{Z}_{(n,\alpha,\beta,\gamma)}$$ and $$M_{k}(f, \alpha,\beta, \gamma)(x) = \sum_{n \geq 0}M_{k,n}(f, \alpha,\beta, \gamma)x^n.$$\\
We provide the following theorem that specifies the asymptotic form of $M_{k,n}(f, \alpha,\beta, 0)$.\\
\begin{theorem}
\label{mainTheorem01}
Fix $k \in \mathbb{N}$, $\alpha,\beta \in \mathbb{R}$ and $f \in \mathbb{R}[t, \lvl, \lvi][n, \gvl, \gvi]$. Let $\delta = \Delta(f| t, n, \lvl, \gvl, \lvi, \gvi)$. Let $V_k(f) = \frac{(2\delta+ 1)k -1}{2}$ if for every monomial of $f$, $u$, such that $\Delta(u| t, n, \lvl, \gvl, \lvi, \gvi) = \delta$, $\Delta(u| t, \lvl, \lvi) > 0$ and $V_k(f) = \frac{ 2(\delta+ 1)k - 1}{2}$ otherwise. We have that
$$M_{k}(f, \alpha,\beta, 0)(x) = \frac{W(\alpha, \beta, f)}{\left(1-\rho x\right)^{V_k(f)}} + O\left(\frac{1}{\left(1-\rho x\right)^{V_k(f) - \frac{1}{2}}}\right),$$ where $\rho = e^{-\alpha} + e^{-\beta} + 2e^{-\frac{\alpha}{2}}$ and $W(\alpha, \beta, f)$ is a constant that depends on $f, \alpha, \beta$.
Furthermore, assume every monomial of $f$, $u$, such that $\Delta(u| t, n, \lvl, \gvl, \lvi, \gvi) = \delta$ is such that $\Delta(u| t, n) = \delta_n$, $\Delta(u| \lvl, \gvl) = \delta_{d_0}$ and $\Delta(u| \lvi, \gvi) = \delta_{d_1}$ where $\delta_n$, $\delta_{d_0}$ and $\delta_{d_1}$ are constants that depend only on $f$. We have that
$$W(\alpha, \beta, f) = Q_k(f, \delta_{d_0}, \delta_{d_1}, \alpha,\beta) \cdot P_{k}(f)$$
where $Q_k(f, \delta_{d_0}, \delta_{d_1}, \alpha,\beta)$ depends on $\delta_{d_0}, \delta_{d_1}$ and $V_k(f)$, and $P_{k}(f)$ is a constant that depends on $f$ and is independent of $\alpha$ and $\beta$.
\end{theorem}
\bigskip
\input{Proof_of_Moment_Generalization_Theorem_1.tex}
Fix $h \in \mathbb{N}$. Let $\mathfrak{T}^h_ n$ for $h \in \mathbb{N}$ be the set of plane trees on $n$ edges with root degree at most $h$. We define $$M_{k,n}(\left. f, \alpha,\beta, \gamma\right| h) = \sum_{T \in \mathfrak{T}^h_n} \mathcal{P}^f(T)^ke^{-E(T)}$$ and
$$M_{k}(\left. f, \alpha,\beta, \gamma\right| h)(x) = \sum_{n \geq 0} M_{k,n}(\left. f, \alpha,\beta, \gamma\right| h) x^n.$$
\begin{theorem}
\label{mainTheorem02}
Fix $k \in \mathbb{Z}_{\geq 0}$ and $\alpha,\beta, \gamma \in \mathbb{R}$. Let $f \in \mathbb{R}[t, \lvl, \lvi][n, \gvl, \gvi]$. Let $\delta = \Delta(f| t, n, \lvl, \gvl, \lvi, \gvi)$. Let $V_k(f) = \frac{(2\delta+ 1)k -1}{2}$ if for every monomial of $f$, $u$, such that $\Delta(u| t, n, \lvl, \gvl, \lvi, \gvi) = \delta$, $\Delta(u| t, \lvl, \lvi) > 0$ and $V_k(f) = \frac{ 2(\delta+ 1)k - 1}{2}$ otherwise. We have
$$M_{k}(\left. f, \alpha,\beta, \gamma\right| h)(x) = \frac{J_h(\alpha,\beta, \gamma) \cdot W(\alpha, \beta, f)}{\left(1-\rho x\right)^{V_k(f)}} + O\left(\frac{1}{\left(1-\rho x\right)^{V_k(f) - \frac{1}{2}}}\right)$$ where $\rho = e^{-\alpha} + e^{-\beta} + 2e^{-\frac{\alpha}{2}}$, $J_h(\alpha,\beta, \gamma)$ is a constant independent of $f$ and $W(\alpha, \beta, f)$ is the same as in Theorem \ref{mainTheorem01}.
\end{theorem}
\bigskip
\input{Proof_of_Moment_Generalization_Theorem_2.tex}
\bigskip
Let $\mathcal{P}^f_{(\alpha, \beta, \gamma)}(\mathfrak{T}^h_n)$ be defined similarly to $\mathcal{P}^f_{(\alpha, \beta, \gamma)}(\mathfrak{T}_n)$ conditioned on the tree chosen having root degree at most $h$. From the above theorem, we can conclude the following.
\begin{theorem} \label{maintheorem3}
Let $\alpha,\beta, \gamma \in \mathbb{R}$ and $f \in \mathbb{R}[t, \lvl, \lvi][n, \gvl, \gvi]$. Assume every monomial of $f$, $u$, such that $\Delta(u| t, n, \lvl, \gvl, \lvi, \gvi) = \delta$ is such that $\Delta(u| t, n) = \delta_n$, $\Delta(u| \lvl, \gvl) = \delta_{d_0}$ and $\Delta(u| \lvi, \gvi) = \delta_{d_1}$ where $\delta_n$, $\delta_{d_0}$ and $\delta_{d_1}$ are constants that depend only on $f$. Let $V'(f) = \frac{2\delta+ 1}{2}$ if for every monomial of $f$, $u$, such that $\Delta(u| t, n, \lvl, \gvl, \lvi, \gvi) = \delta$, $\Delta(u| t, \lvl, \lvi) > 0$ and $V'(f) = \delta+ 1$ otherwise. We have
$$\frac{\mathcal{P}^f_{(\alpha, \beta, \gamma)}(\mathfrak{T}^h_ n)}{n^{V'(f)}} \overset{d}{\longleftrightarrow} \frac{\mathcal{P}^f_{(\alpha, \beta, 0)}(\mathfrak{T}_ n)}{n^{V'(f)}} \overset{d}{\longleftrightarrow} \frac{Q(f, \alpha, \beta) \cdot \mathcal{P}^f_{(0,0,0)}(\mathfrak{T}_ n)}{n^{V'(f)}} $$
where, for $V'(f) = \frac{2\delta+ 1}{2}$, $$Q(f, \alpha, \beta) = 2^{\delta_{d_0} + 2\delta_{d_1}} \cdot \frac{ (e^{-\frac{\alpha}{2}}+1)^{\delta_{d_0} - 1} \cdot e^{-\frac{\alpha}{4}(2 \delta_{d_0}-1)} \cdot e^{-\beta \delta_{d_1}}}{\sqrt{\rho^{2\delta_{d_0} + 2\delta_{d_1} - 1}}}$$
and, for $V'(f) = \delta+ 1$, $$Q(f, \alpha, \beta) = 2^{\delta_{d_0} + 2\delta_{d_1}} \cdot \frac{(e^{-\frac{\alpha}{2}}+1)^{\delta_{d_0}} \cdot e^{-\frac{\alpha \delta_{d_0}}{2}} \cdot e^{-\beta \delta_{d_1}}}{\rho^{\delta_{d_0} + \delta_{d_1}}}.$$
\end{theorem}
\bigskip
\begin{proof}
We notice that $M_{k,n}(1,\alpha, \beta, \gamma) = n^k\mathcal{Z}_{(n, \alpha, \beta, \gamma)}$. For a random variable $X$, let $\text{Mom}_k(X)$ the $k$th moment of $X$. Fixed $n \in \mathbb{Z}_{\geq 0}$. By definition, $$\text{Mom}_k(\mathcal{P}^f_{(\alpha, \beta, 0)}(\mathfrak{T}_ n)) = \frac{M_{k,n}(f,\alpha, \beta, 0)}{\mathcal{Z}_{(n, \alpha, \beta, 0)}} = \frac{n^k \cdot M_{k,n}(f,\alpha, \beta, 0)}{M_{k,n}(1,\alpha, \beta, 0)}$$
We now apply Theorem \ref{mainTheorem01} and use the standard expression for the asymptotic coefficients of the Taylor series expansion of $(1-x)^k$ for $k \in \mathbb{C}$. We use the Transfer Theorem from \cite{flajolet2009analytic}, to get $$M_{k,n}(f, \alpha, \beta, 0) = \frac{Q_k(f, \delta_{d_0}, \delta_{d_1}, \alpha,\beta) \cdot P_{k}(f) \cdot n^{V_k(f)-1} \cdot (e^{-\alpha} + e^{-\beta}+ 2e^{-\frac{\alpha}{2}})^n}{\Gamma(V_k(f))}$$ where we only consider the most significant terms.
Notice that $V_k(1) = \frac{2k-1}{2}$ and every maximal monomial of $1$ is independent of $t, \lvl, \lvi$. Thus
$$\text{Mom}_k(\mathcal{P}^f_{(\alpha, \beta, 0)}(\mathfrak{T}_ n)) = \frac{Q_k(f, \delta_{d_0}, \delta_{d_1}, \alpha,\beta) \cdot P_{k}(f) \cdot \Gamma(V_k(1))}{Q_k(1, 0,0, \alpha,\beta) \cdot P_{k}(1) \cdot \Gamma(V_k(f)) } \cdot n^{V_k(f)+ \frac{1}{2}} + O\left(n^{V_k(f)}\right)$$
Notice that $V'(f)k = V_k(f)+ \frac{1}{2}$. Thus
$$\lim_{n\rightarrow \infty}\text{Mom}_k\left(\frac{\mathcal{P}^f_{(\alpha, \beta, 0)}(\mathfrak{T}_ n)}{n^{V'(f)}}\right) = \frac{Q_k(f, \delta_{d_0}, \delta_{d_1}, \alpha,\beta) \cdot P_{k}(f) \cdot \Gamma(V_k(1))}{Q_k(1, 0,0, \alpha,\beta) \cdot P_{k}(1) \cdot \Gamma(V_k(f)) }.$$
We now see that
$$\text{Mom}_k(\mathcal{P}^f_{(\alpha, \beta, \gamma)}(\mathfrak{T}_n^h)) = \frac{n^k \cdot M_{k,n}(f,\alpha, \beta, \gamma|h)}{M_{k,n}(1,\alpha, \beta, \gamma|h)}$$
Using Theorem \ref{mainTheorem02} and nearly identical computation to above, we see that
$$\lim_{n\rightarrow \infty}\text{Mom}_k\left(\frac{\mathcal{P}^f_{(\alpha, \beta, \gamma)}(\mathfrak{T}^h_ n)}{n^{V'(f)}}\right) = \frac{Q_k(f, \delta_{d_0}, \delta_{d_1}, \alpha,\beta) \cdot P_{k}(f) \cdot \Gamma(V_k(1))}{Q_k(1, 0,0, \alpha,\beta) \cdot P_{k}(1) \cdot \Gamma(V_k(f)) }.$$
Finally, we set that \begin{eqnarray*}
\lim_{n\rightarrow \infty}\text{Mom}_k\left(\frac{\mathcal{P}^f_{(\alpha, \beta, 0)}(\mathfrak{T}_ n)}{n^{V'(f)}}\right) &=& \frac{Q_k(f, \delta_{d_0}, \delta_{d_1}, \alpha,\beta) \cdot P_{k}(f) \cdot \Gamma(V_k(1))}{Q_k(1, 0,0, \alpha,\beta) \cdot P_{k}(1) \cdot \Gamma(V_k(f))}\\
&=& \frac{Q_k(f, \delta_{d_0}, \delta_{d_1}, \alpha,\beta) \cdot Q_k(1, 0,0,0,0)}{Q_k(f, \delta_{d_0}, \delta_{d_1}, 0,0) \cdot Q_k(1, 0,0, \alpha,\beta)} \cdot \lim_{n\rightarrow \infty}\text{Mom}_k\left(\frac{\mathcal{P}^f_{(0, 0, 0)}(\mathfrak{T}_ n)}{n^{V'(f)}}\right)\\
&=& \lim_{n\rightarrow \infty}\text{Mom}_k\left(Q(f, \alpha, \beta) \cdot \frac{\mathcal{P}^f_{(0, 0, 0)}(\mathfrak{T}_ n)}{n^{V'(f)}}\right)
\end{eqnarray*} where $Q(f, \alpha, \beta) = \sqrt[k]{\frac{Q_k(1, 0,0,0,0)}{Q_k(f, \delta_{d_0}, \delta_{d_1}, 0,0)} \cdot \frac{Q_k(f, \delta_{d_0}, \delta_{d_1}, \alpha,\beta)}{Q_k(1, 0,0, \alpha,\beta)}}$ is a constant independent of $k$.
\end{proof}
\bigskip
\begin{corollary} \label{full dist corr CD}
Let $\alpha,\beta, \gamma \in \mathbb{R}$. For any $h \in \mathbb{N}$, we have
$$\frac{\mathcal{P}^{PL}_{(\alpha, \beta, \gamma)}(\mathfrak{T}^h_ n)}{\sqrt{2n^3}} \quad \overset{d}{\longleftrightarrow} \quad \frac{\mathcal{P}^{PL}_{(\alpha, \beta, 0)}(\mathfrak{T}_ n) }{\sqrt{2n^3}} \quad \overset{d}{\to} \quad \frac{\sqrt{\rho}}{ (e^{-\frac{\alpha}{2}}+1) \cdot e^{-\frac{\alpha}{4}}} \int_0^1e(t) \: dt$$ where $e(t)$ is a normalized Brownian excursion on $[0, 1]$.
\end{corollary}
\begin{proof}
We take $f = t+1$ then apply Theorem $\ref{maintheorem3}$ and Theorem \ref{contactdistance - thm}.
\end{proof}
\bigskip
\begin{corollary}
Let $\alpha,\beta, \gamma \in \mathbb{R}$. For any $h \in \mathbb{N}$, we have
\begin{multline*}
\frac{\mathcal{P}^{WI}_{(\alpha, \beta, \gamma)}(\mathfrak{T}^h_ n)}{\sqrt{2n^5}} \quad \overset{d}{\longleftrightarrow} \quad \frac{\mathcal{P}^{WI}_{(\alpha, \beta, 0)}(\mathfrak{T}_ n) }{\sqrt{2n^5}}\\
\overset{d}{\to} \quad \frac{\sqrt{\rho}}{ (e^{-\frac{\alpha}{2}}+1) \cdot e^{-\frac{\alpha}{4}}} \int \int_{0 < s < t < 1} (e(s) + e(t) - 2\min_{s \leq u \leq t}e(u)) \: ds \: dt
\end{multline*}
where $e(t)$ is a normalized Brownian excursion on $[0, 1]$.
\end{corollary}
\begin{proof}
We take $f = (t+1)(n - t)$ then apply Theorem $\ref{maintheorem3}$ and Theorem \ref{ladderdistance - thm}.
\end{proof}
\bigskip
\begin{corollary} \label{full dist corr LRD/IRD}
Let $\alpha,\beta, \gamma \in \mathbb{R}$. For any $h \in \mathbb{N}$, we have
$$\frac{\mathcal{P}^{LR}_{(\alpha, \beta, \gamma)}(\mathfrak{T}^h_ n)}{\sqrt{2n^3}} \quad \overset{d}{\longleftrightarrow} \quad \frac{\mathcal{P}^{LR}_{(\alpha, \beta, 0)}(\mathfrak{T}_ n) }{\sqrt{2n^3}} \quad \overset{d}{\to} \quad \frac{e^{-\frac{\alpha}{4}}}{\sqrt{\rho}} \int_0^1e(t) \: dt$$
and
$$\frac{\mathcal{P}^{IR}_{(\alpha, \beta, \gamma)}(\mathfrak{T}^h_ n)}{\sqrt{2n^3}} \quad \overset{d}{\longleftrightarrow} \quad \frac{\mathcal{P}^{IR}_{(\alpha, \beta, 0)}(\mathfrak{T}_ n) }{\sqrt{2n^3}} \quad \overset{d}{\to} \quad \frac{ e^{-\beta}}{\sqrt{\rho} \cdot e^{-\frac{\alpha}{4}} \cdot (e^{-\frac{\alpha}{2}}+1)} \int_0^1e(t) \: dt$$
where $e(t)$ is a normalized Brownian excursion on $[0, 1]$.
\end{corollary}
\begin{proof}
For $\mathcal{P}^{LR}$, we take $f = \lvl$. For $\mathcal{P}^{IR}$, we take $f = \lvi$. We then apply Theorem $\ref{maintheorem3}$ and Corollary \ref{corollary P^LRD and P^IRD dist}.
\end{proof}
\bigskip
|
\section{Introduction}
\label{Sec_Intro}
As a generalization of the conventional 0-1 knapsack problem (KP) \cite{kellerer2003knapsack}, the disjunctively constrained knapsack problem (DCKP) is defined as follows. Let $V = \{1,\ldots, n\}$ be a set of $n$ items, where each item $i = \{1,\ldots, n\}$ has a profit $p_i > 0$ and a weight $w_i > 0$. Let $G = (V, E)$ be a conflict graph, where $V$ is the set of $n$ items and an edge $\{i, j\} \in E$ defines the incompatibility of items $i$ and $j$. Let $C > 0$ be the capacity of a given knapsack. Then the DCKP involves finding a subset $S$ of pairwisely compatible items of $V$ to maximize the total profit of $S$ while ensuring that the total weight of $S$ does not surpass the knapsack capacity $C$. Formally, the DCKP can be stated as follows.
\begin{equation}\label{MAX}
(DCKP) \quad \quad \mathrm{Maximize} \quad f(S) = \sum\limits_{i = 1}^n p_ix_i
\end{equation}
\begin{equation}\label{constraint1}
\mathrm{subject \ to} \quad W(S) = \sum\limits_{i = 1}^n w_ix_i \leq C, \ S \subseteq V,
\end{equation}
\begin{equation}\label{constraint2}
x_i + x_j \leq 1, \forall(i, j) \in E,
\end{equation}
\begin{equation}\label{constraint3}
x_i \in \{0,1\}, i=1,\ldots,n.
\end{equation}
Objective function (\ref{MAX}) commits to maximize the total profit of the selected item set $S$. Constraint (\ref{constraint1}) ensures that the knapsack capacity constraint is satisfied. Constraints (\ref{constraint2}), called disjunctive constraints, guarantee that two incompatible items are never selected simultaneously. Constraints (\ref{constraint3}) force that each item is selected at most once.
It is easy to observe that the DCKP reduces to the NP-hard KP when $G$ is an empty graph. The DCKP is equivalent to the NP-hard maximum weighted independent set problem \cite{johnson1979computers} when the knapsack capacity is unbounded. Moreover, the DCKP is closely related to other combinatorial optimization problems, such as the multiple-choice knapsack problem \cite{kellerer2003knapsack}, and the bin packing problem with conflicts \cite{jansen1999approximation}. In addition to its theoretical significance, the DCKP is a useful model for practical applications where the resources with conflicts cannot be used simultaneously while a given budget envelope cannot be surpassed.
Given the importance of the DCKP, a number of solution methods have been developed including exact, approximation and heuristic algorithms. As the literature review shown in Section \ref{Sec_Rela}, considerable progresses have been continually made since the introduction of the problem. Meanwhile, given the NP-hard nature of the problem, more powerful algorithms are still needed to push the limits of existing methods.
In this work, we investigate for the first time the population-based memetic framework \cite{moscato1999memetic} for solving the DCKP and design an effective algorithm mixing threshold based local optimization and crossover based solution recombination. The threshold search procedure ensures the main role of search intensification by finding high quality local optimal solutions. The specialized backbone crossover generates promising offspring solutions for search diversification. The algorithm uses also a distance-and-quality strategy for population management. The algorithm has the advantage of avoiding the difficult task of parameter tuning.
From a perspective of performance assessment, we apply the proposed algorithm to solve the two sets of DCKP benchmark instances in the literature. The results show that for the 100 instances of Set I (optimality still unknown) which were commonly tested by heuristic algorithms, our algorithm discovers 24 new best-known results (new lower bounds) and matches the best-known results for the 76 remaining instances. For the 6240 instances of Set II which were tested by exact algorithms, our algorithm finds 354 improved best lower bounds on the difficult instances whose optimal values are unknown and attains the known optimal results on most of the remaining instances.
The rest of the paper is organized as follows. Section \ref{Sec_Rela} provides a literature review on the DCKP. Section \ref{Sec_TSBMA} presents the proposed algorithm. Section \ref{Sec_Result} shows computational results of our algorithm and provides comparisons with the state-of-the-art algorithms. Section \ref{Sec_AD} analyzes essential components of the algorithm. Finally, Section \ref{Sec_Conclu} summarizes the work and provides perspectives for future research.
\section{Related work}
\label{Sec_Rela}
The DCKP has attracted considerable attentions in the past two decades. In this section, we review related literature for solving the DCKP. Existing solution methods can be roughly classified into two categories as follows.
\begin{enumerate}
\item \textit{Exact and approximation algorithms}: These algorithms are able to guarantee the quality of the solutions they find. In 2002, Yamada et al. \cite{yamada2002heuristic} introduced the DCKP and proposed the first implicit enumeration algorithm where the disjunctive constraints are relaxed. In 2007, Hifi and Michrafy \cite{hifi2007reduction} introduced three versions of an exact algorithm based on a local reduction strategy. In 2009, Pferschy and Schauer \cite{pferschy2009knapsack} proposed a pseudo-polynomial time and space algorithm for solving three special cases of the DCKP and proved the DCKP is strongly NP-hard on perfect graphs. In 2016, Salem et al. \cite{salem2018optimization} developed a branch-and-cut algorithm that combines a greedy clique generation procedure with a separation procedure. In 2017, Bettinelli et al. \cite{bettinelli2017branch} presented a branch-and-bound algorithm by combining a upper bounding procedure that considers both the capacity constraint and the disjunctive constraints with a branching procedure that employs a dynamic programming to presolve the 0-1 KP. They generated 4800 DCKP instances with conflict graph densities between 0.1 and 0.9 (see Section \ref{Subsec_Bench}). Also in 2017, Pferschy and Schauer \cite{pferschy2017approximation} applied the approximation methods of modular decompositions and clique separators to the DCKP, and showed complexity results on special graph classes. In 2019, Gurski and Rehs \cite{gurski2019solutions} designed a dynamic programming algorithm and achieved pseudo-polynomial solutions for the DCKP. In 2020, Coniglio et al. \cite{coniglio2020new} presented another branch-and-bound algorithm based on an $n$-ary branching scheme and solved the integer linear programming formulations of the DCKP by the CPLEX solver. They introduced 1440 new and challenging DCKP instances (see Section \ref{Subsec_Bench}).
\item \textit{Heuristic algorithms}: These algorithms aim to find good near-optimal solutions with a given time. In 2002, Yamada et al. \cite{yamada2002heuristic} proposed a greedy algorithm to generate an initial solution and a 2-opt neighborhood search algorithm to improve the obtained solution. In 2006, Hifi and Michrafy \cite{hifi2006reactive} reported a local search algorithm, which combines a complementary constructive procedure to improve the initial solution and a degrading procedure to diversify the search. They generated a set of 50 DCKP instances with 500 and 1000 items (see Section \ref{Subsec_Bench}), which was widely tested in later studies. In 2012, Hifi and Otmani \cite{hifi2012algorithm} studied two scatter search algorithms. In 2014, Hifi \cite{hifi2014iterative} devised an iterative rounding search-based algorithm that uses a rounding strategy to perform a linear relaxation of the fractional variables. In 2017, Salem et al. \cite{salem2017probabilistic} designed a probabilistic tabu search algorithm (PTS) that operates with multiple neighborhoods. In the same year, Quan and Wu investigated two parallel algorithms: the parallel neighborhood search algorithm (PNS) \cite{quan2017design} and the cooperative parallel adaptive neighborhood search algorithm (CPANS) \cite{quan2017cooperative}. They also designed a new set of 50 DCKP large instances with 1500 and 2000 items (see Section \ref{Subsec_Bench}).
\end{enumerate}
Existing studies have significantly contributed to better solving the DCKP. According to the computational results reported in the literature, the parallel neighborhood search algorithm \cite{quan2017design}, the cooperative parallel adaptive neighborhood search algorithm \cite{quan2017cooperative}, and the probabilistic tabu search algorithm \cite{salem2017probabilistic} can be regarded as the state-of-the-art methods for the instances of Set I. For the instances of Set II, the branch-and-bound algorithms presented in \cite{bettinelli2017branch,coniglio2020new} and the integer linear programming formulations solved by the CPLEX solver \cite{coniglio2020new} showed the best performance.
In this work, we aim to advance the state-of-the-art of solving the problem by proposing the first threshold search based memetic approach, which proves to be effective on the two sets of DCKP instances tested in the literature.
\section{Threshold search based memetic algorithm for the DCKP}
\label{Sec_TSBMA}
Our threshold search based memetic algorithm (TSBMA) for the DCKP is a population-based algorithm combining evolutionary search and local optimization. In this section, we first present the general procedure of the algorithm and then describe its components.
\subsection{General procedure}
\label{Subsec_Intro}
The TSBMA algorithm relies on the general memetic algorithm framework \cite{moscato1999memetic} and follows the design principles recommended in \cite{hao2012memetic}. The flowchart of TSBMA and its pseudo-code are shown in Figure \ref{Fig_FC} and Algorithm \ref{Algo_TSBMA}, respectively.
\begin{figure}[h]\centering
\includegraphics[width=3.0in]{Flow_chart.eps}
\caption{Flowchart of the proposed TSBMA algorithm.}
\label{Fig_FC}
\end{figure}
The algorithm starts from a set of feasible solutions of good quality that are generated by the population initialization procedure (line 4, Alg. \ref{Algo_TSBMA}, and Section \ref{Subsec_PI}). The best solution is identified and recorded as the overall best solution $S^*$ (line 5, Alg. \ref{Algo_TSBMA}). Then the algorithm enters the main `while' loop (lines 6-15, Alg. \ref{Algo_TSBMA}) to perform a number of generations. At each generation, two solutions are randomly picked and used by the crossover operator to create an offspring solution (line 7-8, Alg. \ref{Algo_TSBMA}, and Section \ref{Subsec_CO}). Afterwards, the threshold search procedure is triggered to perform local optimization with three neighborhoods $N_1$, $N_2$ and $N_3$ (line 9, Alg. \ref{Algo_TSBMA}, and Section \ref{Subsec_TS}). After conditionally updating the overall best solution $S^*$ (line 11-13, Alg. \ref{Algo_TSBMA}), the diversity-based pool updating procedure is applied to decide whether the best solution $S_b$ found during the threshold search should be inserted into the population (line 14, Alg. \ref{Algo_TSBMA}, and Section \ref{Subsec_PU}). Finally, when the given time limit $t_{max}$ is reached, the algorithm returns the overall best solution $S^*$ found during the search and terminates.
\begin{algorithm}
\footnotesize
\caption{Main framework of threshold search based memetic algorithm for the DCKP}\label{Algo_TSBMA}
\begin{algorithmic}[1]
\STATE \textbf{Input}: Instance $I$, cut-off time $t_{max}$, population $P$, the maximum number of iterations $IterMax$, neighborhoods $N_1$, $N_2$, $N_3$.
\STATE \textbf{Output}: The overall best solution $S^*$ found.
\STATE $S^* \gets \emptyset$ \hfill /* Initialize $S^*$ (i.e., $f(S^*) = 0$)*/
\STATE $POP = \{S^1, \ldots, S^{|P|}\} \gets Population\_Initialization(I)$ \hfill /* Section \ref{Subsec_PI} */
\STATE $S^* \gets arg max\{f(S^k)|k = 1, \ldots, p\}$
\WHILE {$Time \leq t_{max}$}
\STATE Randomly pick two solutions $S^i$ and $S^j$ from the population POP
\STATE $S^{o} \gets Crossover\_Operator(S^i, S^j)$ \hfill /* Section \ref{Subsec_CO} */
\STATE $S_b \gets Threshold\_Search(S^{o}, N_{1-3}, IterMax)$ \hfill /* Section \ref{Subsec_TS} */
\STATE /* Record the best solution $S_b$ found during threshold search */
\IF {$f(S_b) > f(S^*)$}
\STATE $S^* \gets S_b$ \hfill /* Update the overall best solution $S^*$ found so far */
\ENDIF
\STATE $POP \gets Pool\_Updating(S_b, POP)$ \hfill /* Section \ref{Subsec_PU} */
\ENDWHILE
\RETURN $S^*$
\end{algorithmic}
\end{algorithm}
\subsection{Solution representation, search space, and evaluation function}
\label{Subsec_SSE}
The DCKP is a subset selection problem. Thus, a candidate solution for a set $V = \{1,\ldots, n\}$ of $n$ items can be conveniently represented by a binary vector $S = (x_1, \ldots, x_n)$, such that $x_i = 1$ if item $i$ is selected, and $x_i = 0$ otherwise. Equivalently, $S$ can also be represented by $S = <A,\bar{A}>$ such that $A = \{q : x_q = 1 \ $in$ \ S\}$ and $\bar{A} = \{p : x_p = 0 \ $in$ \ S\}$.
Let $G = (V, E)$ be the given conflict graph and $C$ be the knapsack capacity. Our TSBMA algorithm explores the following feasible search space $\Omega^F$ satisfying both the disjunctive constraints and the knapsack constraint.
\begin{equation}\label{omega}
\Omega^F = \{x \in \{0,1\}^n : \sum\limits_{i = 1}^n w_ix_i \leq C; x_i + x_j \leq 1, \forall\{i, j\} \in E, 1 \leq i,j \leq n, i \neq j\}
\end{equation}
The quality of a solution $S$ in $\Omega^F$ is determined by the objective value $f(S)$ of the DCKP (Equation \ref{MAX}).
\subsection{Population initialization}
\label{Subsec_PI}
The TSBMA algorithm builds each of the $|P|$ initial solutions of the population $P$ in two steps. First, it randomly adds one by one non-selected items into an individual solution $S^i$ ($i = 1, \ldots, |P|$) until the capacity of the knapsack is reached, while keeping the disjunctive constraints satisfied. Second, to obtain an initial population of reasonable quality, it improves the solution $S^i$ by a short run of the threshold search procedure (Section \ref{Subsec_TS}) by setting $IterMax = 2n$.
It is worth mentioning that the population size $|P|$ is determined according to the number of candidate items $n$ of the given instance, i.e., $|P| = n/100 + 5$. This strategy is based on two considerations. First, since the TSBMA algorithm is powerful enough to solve the instances of small size, a smaller population size can help to reduce the initialization time. Second, the instances of large size are more challenging, a larger population size helps to diversify the search.
\subsection{Local optimization using threshold search}
\label{Subsec_TS}
The local optimization procedure of the TSBMA algorithm relies on the threshold accepting method \cite{DUECK1990161}. To explore a given neighborhood, the method accepts both improving and deteriorating neighbor solutions so long as the solution satisfies a quality threshold. One notices that this method has been successfully applied to solve several knapsack problems (e.g., quadratic multiple knapsack problem \cite{chen2015iterated}, multi-constraint knapsack problem \cite{dueck1991threshold} and multiple-choice knapsack problem \cite{zhou2008algorithm}) and other combinatorial optimization problems (e.g., \cite{castelino1996tabu,tarantilis2004threshold}). In this work, we adopt for the first time this method for solving the DCKP and devise a multiple neighborhood threshold search procedure reinforced by an operation-prohibiting mechanism.
\subsubsection{Main scheme of the threshold search procedure}
\label{Subsec_MTSP}
As shown in Algorithm \ref{Algo_TS}, the threshold search procedure (TSP) starts its process from an input solution and three empty hash vectors (used for the operation-prohibiting mechanism, lines 3-5, Alg. \ref{Algo_TS}). It then performs a number of iterations to explore three neighborhoods (Section \ref{Subsec_MN}) to improve the current solution $S$. Specifically, for each `while' iteration (lines 9-25, Alg. \ref{Algo_TS}), the TSP procedure explores the neighborhoods $N_{1}$, $N_{2}$ and $N_{3}$ in a deterministic way as explained in the next section. Any sampled non-prohibited neighbor solution $S'$ is accepted immediately if the quality threshold $T$ is satisfied (i.e., $f(S') \geq T$). Then the hash vectors are updated for solution prohibition and the best solution found during the TSP procedure is recorded in $S_b$ (line 18-20, Alg. \ref{Algo_TS}). The main search (`while' loop) terminates when 1) no admissible neighbor solution (i.e., non-prohibited and satisfying the quality threshold) exists in the neighborhoods $N_{1}$, $N_{2}$ and $N_{3}$, or 2) the best solution $S_b$ cannot be further improved during $IterMax$ consecutive iterations. Specifically, the quality threshold $T$ is determined adaptively by $f(S_b) - n/10$ ($n$ is the number of items of each instance) while $IterMax$ is set to $(n/500 +5) \times 10000$.
\begin{algorithm}
\footnotesize
\caption{Threshold search procedure}\label{Algo_TS}
\begin{algorithmic}[1]
\STATE \textbf{Input}: Input solution $S^{o}$, threshold $T$, the maximum number of iterations $IterMax$, hash vectors $H_1$, $H_2$, $H_3$, hash functions $h_1$, $h_2$, $h_3$, length of hash vectors $L$, neighborhoods $N_1$, $N_2$, $N_3$.
\STATE \textbf{Output}: The best feasible solution $S_b$ found by threshold search procedure.
\FOR {$i \gets 0$ \textbf{to} $L-1$}
\STATE $H_1[i] \gets 0$; $H_2[i] \gets 0$; $H_3[i] \gets 0$; \hfill /* Initialization of hash vectors */
\ENDFOR
\STATE $S_b \gets S^{o}$ \hfill /* $S_b$ record the best solution found */
\STATE $S \gets S^{o}$ \hfill /* $S$ record the current solution */
\STATE $iter \gets 0$
\WHILE {$iter \leq IterMax$}
\STATE Examine the neighborhoods $N_1(S)$, $N_2(S)$, $N_3(S)$ in turn; /* Section \ref{Subsec_MN} */\\
/* Each non-prohibited neighbor solution $S'$ satisfies $H_1[h_1(S{'})] \wedge H_1[h_1(S{'})] = 0 \wedge H_1[h_1(S{'})] = 0$ */
\FOR {Each non-prohibited $S'$ of $N_1(S)$ or $N_2(S)$ or $N_3(S)$}
\IF {$f(S') \geq T$}
\STATE $S \gets S'$
\STATE /* Update the hash vectors with $S$, Section \ref{Subsec_OP} */
\\ $H_1[h_1(S)] \gets 1$; $H_2[h_2(S)] \gets 1$; $H_3[h_3(S)] \gets 1$
\STATE \textbf{break};
\ENDIF
\ENDFOR
\IF {$f(S) > f(S_b)$}
\STATE $S_b \gets S$ \hfill /* Update the best solution $S_b$ found during threshold search */
\STATE $iter \gets 0$
\ELSE
\STATE $iter \gets iter + 1$
\ENDIF
\ENDWHILE
\RETURN $S_b$
\end{algorithmic}
\end{algorithm}
\subsubsection{Neighborhoods and their exploration}
\label{Subsec_MN}
The TSP procedure examines candidate solutions by exploring three neighborhoods induced by the popular move operators: $add$, $swap$ and $drop$. Let $S$ be the current solution and $mv$ is one of these operators. We use $S' = S \oplus mv$ to denote a feasible neighbor solution obtained by applying $mv$ to $S$ and $N_x$ ($x=1,2,3$) to represent the resulting neighborhoods. To avoid the examination of unpromising neighbor solutions, TSP employs the following dynamic neighborhood filtering strategy inspired by \cite{lai2019two,wei2019iterated}. Let $S'$ be a neighbor solution in the neighborhood currently under examination, and $S_c$ be the best neighbor solution encountered during the current neighborhood examination. Then $S'$ is excluded for consideration if it is no better than $S_c$ (i.e., $f(S') \leq f(S_c)$). By eliminating the unpromising neighbor solutions, TSP increases the efficiency of its neighborhood search.
Specifically, the associated neighborhoods induced by $add$, $swap$ and $drop$ are defined as follows.
\begin{itemize}[leftmargin = 25pt]
\item $add(p)$: This move operator expands the selected item set $A$ by one non-selected item $p$ from the set $\bar{A}$ such that the resulting neighbor solution is feasible. This operator induces the neighborhood $N_1$.
\begin{equation}\label{N_1}
N_1(S) = \{S': S'=S \oplus add(p), p \in \bar{A} \}
\end{equation}
\item $swap(q,p)$: This move operator exchanges a pair of items $(q,p)$ where item $q$ belongs to the selected item set $A$ and $p$ belongs to the non-selected item set $\bar{A}$ such that the resulting neighbor solution is feasible. This operator induces the neighborhood $N_2$.
\begin{equation}\label{N_2}
N_2(S) = \{S': S'=S \oplus swap(q,p), q \in A, p \in \bar{A}, f(S') > f(S_c)\}
\end{equation}
\item $drop(q)$: This operator displaces one selected item $q$ from the set $A$ to the non-selected item set $\bar{A}$ and induces the neighborhood $N_3$.
\begin{equation}\label{N_3}
N_3(S) = \{S': S'=S \oplus drop(q), q \in A, f(S') > f(S_c)\}
\end{equation}
\end{itemize}
One notices that the $add$ operator always leads to a better current solution with an additional eligible item, and thus the neighborhood filtering is not needed for $N_1$. The $drop$ operator always deteriorates the quality of the current solution, and the feasibility of a neighbor solution is always ensured. The $swap$ operator may either increase or decrease the objective value and the feasibility of a neighbor solution needs to be verified. For $N_2$ and $N_3$, neighborhood filtering excludes uninteresting solutions that can in no way be accepted during the TSP process.
The TSP procedure examines the neighborhoods $N_1, N_2,$ and $N_3$ in a token-ring way \cite{di2006neighborhood} to explore different local optimal solutions. For $N_1$, as long as there exists a non-prohibited neighbor solution, TSP selects such a neighbor solution to replace the current solution (ties are broken randomly). Once $N_1$ becomes empty, TSP moves to $N_2$, if there exists a non-prohibited neighbor solution $S'$ satisfying $f(S') \geq T$, TSP selects $S'$ to become the current solution and immediately returns to the neighborhood $N_1$. When $N_2$ becomes empty, TSP continues its search with $N_3$ and explores $N_3$ exactly like with $N_2$. When $N_3$ becomes empty, TSP terminates its search and returns the best solution found $S_b$. TSP may also terminate if its best solution remains unchanged during $IterMax$ consecutive iterations.
\subsubsection{Operation-prohibiting mechanism}
\label{Subsec_OP}
During the TSP procedure, it is important to prevent the search from revisiting a previously encountered solution. For this purpose, TSP utilizes an operation-prohibiting (OP) mechanism that is based on the tabu list strategy \cite{GloverLagunaBook}. To implement the operation-prohibiting (OP) mechanism, we adopt the solution-based tabu search technique \cite{woodruff1993hashing}. Specifically, we employ three hash vectors $H_v$ $(v=1, 2, 3)$ of length $L$ ($|L| = 10^8$) to record previously visited solutions. Given a solution $S = (x_1, \ldots , x_n)$ ($x_i \in \{0,1\}$), we pre-compute for each item $i$, the weight $\mathcal{W}_i = i^{\gamma_v}$ $(v=1, 2, 3)$, where $\gamma_v$ is equal to $1.2, 1.6, 2.0$, respectively. Then the hash functions $h_v$ {$(v=1, 2, 3)$} are defined as follows.
\begin{equation}\label{Hash_f}
\begin{aligned}
h_v(S) = (\sum_{i=1}^{n} \lfloor \mathcal{W}_i \times x_i \rfloor) \bmod L
\end{aligned}
\end{equation}
The hash value of a neighbor solution $S'$ from the $add$, $swap$ or $drop$ operator can be efficiently computed as follows ($x \in A, y \in \bar{A}$, Section \ref{Subsec_SSE}).
\begin{equation}\label{Hash_f2}
h_v(S') =
\begin{cases}
h(S) + \mathcal{W}_y, & $for\ the$ \ add \ $operator$ \\
h(S) - \mathcal{W}_x + \mathcal{W}_y, & $for\ the$ \ swap \ $operator$ \\
h(S) - \mathcal{W}_x, & $for\ the$ \ drop \ $operator$
\end{cases}
\end{equation}
Starting with the hash vectors set to 0, the corresponding positions in the three hash vectors $H_v$ is updated by 1 whenever a new neighbor solution $S'$ is accepted to replace the current solution $S$ (line 12-16, Alg. \ref{Algo_TS}). For each candidate neighbor solution $S'$, its hashing value $h_v(S')$ is calculated with Equation (\ref{Hash_f2}) in $O(1)$. Then, this neighbor solution $S'$ is previously visited if $H_1[h_1(S')] \wedge H_2[h_2(S')] \wedge H_3[h_3(S')] = 1$ and is prohibited from consideration by the TSP procedure.
\subsection{Crossover operator}
\label{Subsec_CO}
The crossover operator generally creates new solutions by recombining two existing solutions. For the DCKP, we adopt the idea of the double backbone-based crossover (DBC) operator \cite{zhou2018memetic} and adapt it to the problem.
Given two solutions $S^i$ and $S^j$, we use them to divide the set of $n$ items into three subsets: the common items set $X_1 = S^i \cap S^j$, the unique items set $X_2 = (S^i \cup S^j) \setminus (S^i \cap S^j)$ and the unrelated set $X_3 = V \setminus (S^i \cup S^j)$. The basic idea of the DBC operator is to generate an offspring solution $S^{o}$ by selecting all items in set $X_1$ (the first backbone) and some items in set $X_2$ (the second backbone), while excluding items in set $X_3$.
As shown in Algorithm \ref{Algo_CO}, from two randomly selected parent solutions $S^i$ and $S^j$, the DBC operator generates $S^{o}$ in three steps. First, we initialize $S^{o}$ by setting all the variables $x_a^{o}$ ($a = 1, \ldots, n$) to 0 (line 3, Alg. \ref{Algo_CO}). Second, we identify the common items set $X_1$ and the unique items set $X_2$ (line 4-10, Alg. \ref{Algo_CO}). Third, we add all items belonging to $X_1$ into $S^{o}$ and randomly add items from $X_2$ into $S^{o}$ until the knapsack constraint is reached (line 11-17, Alg. \ref{Algo_CO}). Note that the knapsack and disjunctive constraints are always satisfied during the crossover process.
Since the DCKP is a constrained problem, the DBC operator adopted for TSBMA has several special features to handle the constraints, which is different from the DBC operator introduced in \cite{zhou2018memetic}. First, we iteratively add an item into $S^{o}$ by selecting one item from the unique items set $X_2$ randomly until the knapsack constraint is reached, while each item in $X_2$ is considered with a probability $p_0$ ($0 < p_0 < 1$) in \cite{zhou2018memetic}. Second, unlike \cite{zhou2018memetic} where a repair operation is used to achieve a feasible offspring solution, our DBC operator ensures the satisfaction of the problem constraints during the offspring generation process.
\begin{algorithm}
\footnotesize
\caption{The double backbone-based crossover operator}\label{Algo_CO}
\begin{algorithmic}[1]
\STATE \textbf{Input}: Two parent solutions $S^i = (x_1^i, x_2^i, \ldots, x_n^i)$ and $S^j = (x_1^j, x_2^j, \ldots, x_n^j)$.
\STATE \textbf{Output}: An offspring solution $S^{o} = (x_1^{o}, x_2^{o}, \ldots, x_n^{o})$.
\STATE $S^{o} \gets \emptyset$ \hfill /* Initialize $S^{o}$ (i.e., $f(S^{o}) = 0$)*/
\FOR {$a \gets 1$ \textbf{to} $n$}
\IF {$x_a^i = 1$ and $x_a^j = 1$}
\STATE $X_1 \gets a$ \hfill /* $X_1$ is the common items set *
\ELSIF {$x_a^i = 1$ or $x_a^j = 1$}
\STATE $X_2 \gets a$ \hfill /* $X_2$ is the unique items set */
\ENDIF
\ENDFOR
\STATE $S^{o} \gets X_1$ \hfill /* Add all items belonging to $X_1$ into $S^{o}$ */
\STATE Randomly shuffle all items in $X_2$;
\FOR {each $a \in$ $X_2$}
\IF {$S^{o} \cup (x_a^{o} = 1)$ is a feasible solution}
\STATE $x_a^{o} \gets 1$ \hfill /* The second backbone */
\ENDIF
\ENDFOR
\RETURN $S^{o}$
\end{algorithmic}
\end{algorithm}
\subsection{Population updating}
\label{Subsec_PU}
Once a new offspring solution is obtained by the DBC operator in the last section, it is further improved by the threshold search procedure presented in Section \ref{Subsec_TS}. Then we adopt a diversity-based population updating strategy \cite{lai2018two} to decide whether the improved offspring solution should replace an existing solution in the population. This strategy is beneficial to balance the quality of the offspring solution and its distance from the population.
To accomplish this task, we temporarily insert the improved offspring solution into the population and compute the distance (Hamming distance) between any two solutions in the population. Then we obtain the goodness score of each solution in the same way as proposed in \cite{lai2018two}. Finally, the worst solution in the population is identified according to the goodness score and deleted from the population.
\subsection{Time complexity of TSBMA}
\label{Subsec_TC}
As shown in Section \ref{Subsec_PI}, the population initialization procedure includes two steps. Given a DCKP instance with $n$ items, the first step of random selection takes time $O(n)$. Given an input solution $S=<A,\bar{A}>$ (see Section \ref{Subsec_SSE}), the complexity of one iteration of the TSP procedure is $O((n + |A| \times |\bar{A}|))$. Then the second step of the initialization procedure can be realized in $O([(n + |A| \times |\bar{A}|)] \times IterMax)$, where $IterMax$ is set to $2n$ in the initialization procedure. The complexity of the population initialization procedure is $O(n^3)$.
Now we consider the four procedures in the main loop of the TSBMA algorithm: parent selection, crossover operator, the TSP procedure and population updating. The parent selection procedure is realized in $O(1)$. The crossover operator takes time $O(n)$. The complexity of the TSP procedure is $O([(n + |A| \times |\bar{A}|)] \times IterMax)$, where $IterMax$ is determined in Section \ref{Subsec_MTSP}. The population updating procedure can be achieved in $O(n|P|)$, where $|P|$ is the population size. Then, the complexity of one iteration of the main loop of the TSBMA algorithm is $O(n^2 \times IterMax)$.
\section{Computational results and comparisons}
\label{Sec_Result}
In this section, we assess the proposed TSBMA algorithm by performing extensive experiments and making comparisons with state-of-the-art DCKP algorithms. We report computational results on two sets of 6340 benchmark instances.
\subsection{Benchmark instances}
\label{Subsec_Bench}
The benchmark instances of the DCKP tested in our experiments were widely used in the literature, which can be divided into two sets (see Tables \ref{Intro_insI} and \ref{Intro_insII} for the main characteristics of these instances).
\textbf{Set I (100 instances)}: These instances are grouped into 20 classes (each with 5 instances) and named by $xIy$ ($x = \{1,\ldots, 20\}$ and $y = \{1,\ldots, 5\}$). The first 50 instances ($1Iy$ to $10Iy$) were introduced in 2006 \cite{hifi2006reactive} and have the following features: number of items $n=500$ or $1000$, capacity $C=1800$ or $2000$, and density $\eta$ going from 0.05 to 0.40. Note that the density is given by $2m / n(n-1)$, where $m$ is the number of disjunctive constraints (i.e., the number of edges of the conflict graph). These instances have an item weight $w_i$ uniformly distributed in $[1,100]$ and a profit $p_i = w_i + 10$. For the instance classes $11Iy$ to $20Iy$ introduced in 2017 \cite{quan2017cooperative}, the number of items $n$ is set to 1500 or 2000, the capacity $C$ is set to 4000, and the density $\eta$ ranges from 0.04 to 0.20. These instances have an item weight $w_i$ uniformly distributed in $[1,400]$ and a profit $p_i$ equaling $w_i + 10$.
\textbf{Set II (6240 instances)}: This set of instances was introduced in 2017 \cite{bettinelli2017branch} and expanded in 2020 \cite{coniglio2020new}. For the four correlated instance classes $C1$ to $C15$ (denoted by $CC$) and four random classes $R1$ to $R15$ (denoted by $CR$), the number of items $n$ is from 60 to 1000, the capacity $C$ is from 150 to 15000, and the density $\eta$ is from 0.10 to 0.90. Each of these eight classes contains 720 instances. For the correlated instance class $SC$ and the random instance class $SR$ of the sparse graphs, the number of items $n$ is from 500 to 1000, the capacity $C$ is from 1000 to 2000, and the density $\eta$ is from 0.001 to 0.05. Each of these two classes contains 240 DCKP instances. More details about this set of instances can be found in \cite{coniglio2020new}.
\renewcommand{\baselinestretch}{1}\large
\begin{table}[!htbp]\centering
\caption{Summary of main characteristics of the 100 DCKP instances of Set I.}
\begin{scriptsize}
\setlength{\tabcolsep}{4mm}{
\begin{tabular}{llllllllll}
\toprule[0.75pt]
Class & Total & $n$ & $C$ & $\eta$ & Class & Total & $n$ & $C$ & $\eta$ \\
\hline
$1Iy$ & 5 & 500 & 1800 & 0.10 & $11Iy$ & 5 & 1500 & 4000 & 0.04 \\
$2Iy$ & 5 & 500 & 1800 & 0.20 & $12Iy$ & 5 & 1500 & 4000 & 0.08 \\
$3Iy$ & 5 & 500 & 1800 & 0.30 & $13Iy$ & 5 & 1500 & 4000 & 0.12 \\
$4Iy$ & 5 & 500 & 1800 & 0.40 & $14Iy$ & 5 & 1500 & 4000 & 0.16 \\
$5Iy$ & 5 & 1000 & 1800 & 0.05 & $15Iy$ & 5 & 1500 & 4000 & 0.20 \\
$6Iy$ & 5 & 1000 & 2000 & 0.06 & $16Iy$ & 5 & 2000 & 4000 & 0.04 \\
$7Iy$ & 5 & 1000 & 2000 & 0.07 & $17Iy$ & 5 & 2000 & 4000 & 0.08 \\
$8Iy$ & 5 & 1000 & 2000 & 0.08 & $18Iy$ & 5 & 2000 & 4000 & 0.12 \\
$9Iy$ & 5 & 1000 & 2000 & 0.09 & $19Iy$ & 5 & 2000 & 4000 & 0.16 \\
$10Iy$ & 5 & 1000 & 2000 & 0.10 & $20Iy$ & 5 & 2000 & 4000 & 0.20 \\
\bottomrule[0.75pt]
\end{tabular}}
\label{Intro_insI}
\end{scriptsize}
\end{table}
\renewcommand{\baselinestretch}{1}\large\normalsize
\renewcommand{\baselinestretch}{1}\large
\begin{table}[!htbp]\centering
\caption{Summary of main characteristics of the 6240 DCKP instances of Set II.}
\begin{scriptsize}
\setlength{\tabcolsep}{4mm}{
\begin{tabular}{llllllll}
\toprule[0.75pt]
\multirow{2}{*}{Class} &\multirow{2}{*}{Total} & \multicolumn{2}{c}{$n$} & \multicolumn{2}{c}{$C$} & \multicolumn{2}{c}{$\eta$} \\
\cmidrule(lr){3-4} \cmidrule(lr){5-6} \cmidrule(lr){7-8}
& & Min & Max & Min & Max & Min & Max \\
\hline
$C1$ & 720 & 60 & 1000 & 150 & 1000 & 0.10 & 0.90 \\
$C3$ & 720 & 60 & 1000 & 450 & 3000 & 0.10 & 0.90 \\
$C10$ & 720 & 60 & 1000 & 1500 & 10000 & 0.10 & 0.90 \\
$C15$ & 720 & 60 & 1000 & 15000 & 15000 & 0.10 & 0.90 \\
$R1$ & 720 & 60 & 1000 & 150 & 1000 & 0.10 & 0.90 \\
$R3$ & 720 & 60 & 1000 & 450 & 3000 & 0.10 & 0.90 \\
$R10$ & 720 & 60 & 1000 & 1500 & 10000 & 0.10 & 0.90 \\
$R15$ & 720 & 60 & 1000 & 15000 & 15000 & 0.10 & 0.90 \\
$SC$ & 240 & 500 & 1000 & 1000 & 2000 & 0.001 & 0.05 \\
$SR$ & 240 & 500 & 1000 & 1000 & 2000 & 0.001 & 0.05 \\
\bottomrule[0.75pt]
\end{tabular}}
\label{Intro_insII}
\end{scriptsize}
\end{table}
\renewcommand{\baselinestretch}{1}\large\normalsize
\subsection{Experimental settings}
\label{Subsec_Setting}
\textbf{Reference algorithms.} For the 100 DCKP instances of Set I that were widely tested by heuristic algorithms, we adopt as our reference methods three state-of-the-art heuristic algorithms: parallel neighborhood search algorithm (PNS) \cite{quan2017design}, cooperative parallel adaptive neighborhood search algorithm (CPANS) \cite{quan2017cooperative}, and probabilistic tabu search algorithm (PTS) \cite{salem2017probabilistic}. Note that PTS only reported results of the 50 instances $1Iy$ to $10Iy$, since the other 50 instances of $11Iy$ to $20Iy$ were designed later. For the 6240 DCKP instances of Set II that were only tested by exact algorithms until now, we cite the results of three best performing methods: branch-and-bound algorithms BCM \cite{bettinelli2017branch} and CFS \cite{coniglio2020new}) as well as the integer linear programming formulations solved by the CPLEX solver (ILP) \cite{coniglio2020new}.
\textbf{Computing platform.} The proposed TSBMA algorithm was written in C++\footnote{The code of our TSBMA algorithm will be available at: \url{http://www.info.univ-angers.fr/pub/hao/DCKP_TSBMA.html.}\label{foot}} and compiled using the g++ compiler with the -O3 option. All experiments were carried out on an Intel Xeon E5-2670 processor (2.5 GHz CPU and 2 GB RAM) under the Linux operating system. The results of the main reference algorithms have been obtained on computing platforms with the following features: an Intel Xeon processor with 2$\times$3.06 GHz for CPANS and PNS, an Intel Pentium i5-6500 processor with 3.2 GHz and 4 GB RAM for PTS, and an Intel Xeon E5-2695 processor with 3.00GHz for CFS. Note that the parallel algorithms PNS and CPANS used 10 to 400 processors to obtain the results.
\textbf{Parameter settings.} The TSBMA algorithm does not require parameter tuning (it is parameter-free). However, for the 6240 instances of Set II (with a wide range of densities and number of items), we adjusted the threshold $T$ (see Section \ref{Subsec_MTSP}) to $T = MinP + rand(20)$, where $MinP$ is the minimum profit value for each instance tested.
\textbf{Stopping condition.} For the 100 DCKP instances of Set I, the TSBMA algorithm adopted the same cut-off time as the reference algorithms (PNS, CPANS and PTS), i.e., 1000 seconds. Note that for the instances $11Iy$ to $20Iy$, PNS used a much longer limit of 2000 seconds. Given its stochastic nature, TSBMA was performed 20 times independently with different random seeds to solve each instance. For the 6240 instances of Set II, the cut-off time was set to 600 seconds as in the CFS algorithm and the number of repeated runs was set to 10.
\subsection{Computational results and comparisons}
\label{Subsec_Result}
In this section, we first present summarized comparisons of the proposed TSBMA algorithm against each reference algorithm on the 100 instances of Set I, and then show the comparative results on the 6240 DCKP instances of Set II. The detailed computational results of our algorithm and the reference algorithms on the instances of Set I are shown in the Appendix, while our solution certificates for these 100 instances are available at the webpage indicated in footnote \ref{foot}. For the 6240 instances of Set II, we report their objective values at the same website.
\subsubsection{Comparative results on the 100 benchmark instances of Set I}
\label{Compara_I}
The comparative results of the TSBMA algorithm and each reference algorithm are summarized in Table \ref{Summary_insI}. Column 1 indicates the pairs of compared algorithms and column 2 gives the names of instance class. Column 3 shows the quality indicators: the best objective value ($f_{best}$) and the average objective value ($f_{avg}$) (when the average results are available in the literature). The following columns \#Wins, \#Ties and \#Losses present the number of instances for which TSBMA achieves a better, equal and worse result according to the indicators. To further analyze the performance of our algorithm, we carried out the Wilcoxon signed-rank test to verify the statistical significance of the compared results between TSBMA and each compared algorithm in terms of the $f_{best}$ and $f_{avg}$ values (when the average results are available in the literature). The outcomes of the Wilcoxon tests are shown in the last column where `NA' means that the two sets of compared results are exactly the same.
From Table \ref{Summary_insI}, one observes that the TSBMA algorithm competes very favorably with all the reference algorithms by reporting improved or equal results on all the instances. Compared to the probabilistic tabu search algorithm (PTS) \cite{salem2017probabilistic} which reported results only on the first 50 instances of classes $1Iy$ to $10Iy$, TSBMA finds 8 (45) better $f_{best}$ ($f_{avg}$) values, while matching the remaining results. Compared to the two parallel algorithms (PNS) \cite{quan2017design} and (CPANS) \cite{quan2017cooperative} that reported only the $f_{best}$ values, TSBMA obtained 35 and 29 better $f_{best}$ results, respectively. The small $p$-$values$ ($< 0.05$) from the Wilcoxon tests between TSBMA and its competitors indicate that the performance differences are statistically significant. Finally, it is remarkable that our TSBMA algorithm discovered 24 new lower bounds on the instances $11Iy$ to $20Iy$ (see the detailed results shown in the Appendix).
\begin{table}[!htp]\centering
\caption{Summarized comparisons of the TSBMA algorithm against each reference algorithm with the $p$-$values$ of the Wilcoxon signed-rank test on the 100 DCKP instances of Set I.}
\renewcommand{\baselinestretch}{1.0}\large\normalsize
\begin{scriptsize}
\setlength{\tabcolsep}{1mm}{
\begin{tabular}{l|cccccc}
\toprule[0.75pt]
Algorithm pair & Instance & Indicator & \#Wins & \#Ties & \#Losses & $p$-$value$ \\
\hline
TSBMA vs. PTS \cite{salem2017probabilistic} & $1Iy - 10Iy$ (50) & $f_{best}$ & 8 & 42 & 0 & 1.40e-2 \\
& & $f_{avg}$ & 45 & 5 & 0 & 5.34e-9 \\
\hline
TSBMA vs. PNS \cite{quan2017design} & $1Iy - 10Iy$ (50) & $f_{best}$ & 9 & 41 & 0 & 8.91e-3 \\
& $11Iy - 20Iy$ (50) & $f_{best}$ & 26 & 24 & 0 & 8.25e-6 \\
\hline
TSBMA vs. CPANS \cite{quan2017cooperative} & $1Iy - 10Iy$ (50) & $f_{best}$ & 0 & 50 & 0 & NA \\
& $11Iy - 20Iy$ (50) & $f_{best}$ & 29 & 21 & 0 & 2.59e-6 \\
\bottomrule[0.75pt]
\end{tabular}}
\label{Summary_insI}
\end{scriptsize}
\end{table}
\renewcommand{\baselinestretch}{1.0}\large\normalsize
To complete the assessment, we provide the performance profiles \cite{dolan2002benchmarking} of the four compared algorithms on the 100 instances of Set I. Basically, the performance profile of an algorithm shows the cumulative distribution for a given performance metric, which reveals the overall performance of the algorithm on a set of instances. In our case, the plots concern the best objective values ($f_{best}$) of the compared algorithms since the average results of some reference algorithms are not available in the literature. Given a set of algorithms (solvers) $\mathcal{S}$ and an instance set $\mathcal{P}$, the performance ratio is given by $r_{p,s} = \frac{f_{p,s}}{min\{f_{p,s} : s \in \mathcal{S} \}}$, where $f_{p,s}$ is the $f_{best}$ value of instance $p$ of $\mathcal{P}$ obtained by algorithm $s$ of $\mathcal{S}$. The performance profiles are shown in Figure \ref{Fig_PP}, where the performance ratio and the percentage of instances solved by each compared algorithm are displayed on the $X$-$axis$ and $Y$-$axis$, respectively. When the value of $X$-$axis$ is 1, the corresponding value of $Y$-$axis$ indicates the fraction of instances for which algorithm $s$ can reach the best $f_{best}$ value of the set $\mathcal{S}$ of the compared algorithms.
From Figure \ref{Fig_PP}, we observe that our TSBMA algorithm has a very good performance on the 100 benchmark instances of Set I compared to the reference algorithms. For the 50 instances $1Iy$ to $10Iy$, TSBMA and CPANS are able to reach 100\% best $f_{best}$ values on these 50 instances, while PTS and PNS fail on around 15\% of the instances. When considering the 50 instances $11Iy$ to $20Iy$, the plot of TSBMA strictly runs above the plots of PNS and CPANS, revealing that our algorithm dominates the reference algorithms on these 50 instances. These outcomes again confirm the high performance of our TSBMA algorithm.
\begin{figure}[H]
\begin{minipage}{0.44\textwidth}
\includegraphics[width=3.1in]{PP1.eps}
\end{minipage}
\hfill
\begin{minipage}{0.44\textwidth}
\includegraphics[width=3.1in]{PP2.eps}
\end{minipage}
\caption{Performance profiles of the compared algorithms on the 100 DCKP instances of Set I.}
\label{Fig_PP}
\end{figure}
\subsubsection{Comparative results on the 6240 benchmark instances of Set II}
\label{Compara_II}
Table \ref{Summary_insII} summarizes the comparative results of our TSBMA algorithm on the 6240 instances of Set II, together with the three reference algorithms mentioned in \ref{Subsec_Setting}. Note that three ILP formulations were studied in \cite{coniglio2020new}, we extracted the best results of these formulations in Table \ref{Summary_insII}, i.e., the results on instances $CC$ and $CR$ (conflict graph density from 0.10 to 0.90) with ILP$_2$ and the results on very sparse instances $SC$ and $SR$ (conflict graph density from 0.0001 to 0.005) with ILP$_1$. Columns 1 and 2 of Table \ref{Summary_insII} identify each instance class and the total number of instances of the class. Columns 3 to 5 indicate the number of instances solved to optimality by the three reference algorithms. Column 6 shows the number of instances for which our TSBMA algorithm reaches the optimal solution proved by exact algorithms. The number of new lower bounds (denoted by NEW LB in Table \ref{Summary_insII}) found by TSBMA is provided in column 7. The best results of the compared algorithms are highlighted in bold. In order to further evaluate the performance of our algorithm, we summarize the available comparative results between MSBTS and the main reference algorithm CFS in columns 8 to 10. The last three rows provide an additional summary of the results for each column.
From Table \ref{Summary_insII}, we observe that TSBMA performs globally very well on the instances of Set II. For the 5760 $CC$ and $CR$ instances, TSBMA reaches most of the proved optimal solutions (5381 out of 5389) and discovers new lower bounds for 323 difficult instances whose optima are still unknown. For the 240 very sparse $SC$ instances, TSBMA matches 195 out of 200 proved optimal solutions and finds 24 new lower bounds for the remaining instances. Although TSBMA successfully solves only 9 out of the 229 solved very sparse $SR$ instances, it discovers 7 new lower bounds. The high performance of TSBMA is further evidenced with the comparison with the best exact algorithm CFS (last three columns).
Notice that the performance of CPLEX with ILP$_1$ is better than TSBMA as well as the two reference algorithms BCM and CFS on the two classes of very sparse instances ($SC$ and $SR$). As analyzed in \cite{coniglio2020new}, one of the main reasons is that the LP relaxation of ILP$_1$ provides a very strong upper bound, which makes the ILP$_1$ formulation very suitable for solving very sparse instances. The disjunctive constraints become very weak when the conflict graph is very sparse. For these two classes of instances, the pure branch-and-bound CFS algorithm is more effective on extremely sparse instances with densities up to 0.005. On the contrary, our TSBMA algorithm is more suitable for solving sparse instances with densities between 0.01 and 0.05. In fact, the new lower bounds found by TSBMA all concern instances with a density of 0.05. Finally, the TSBMA algorithm remains competitive on the 240 correlated sparse instances $SC$, even if the density is the smallest (0.001), which means that only the random sparse instance class $SR$ is challenging for TSBMA.
In summary, our TSBMA algorithm is computational efficient on a majority of the 6240 benchmark instances of Set II and is able to discover new lower bounds on 354 difficult DCKP instances, whose optimal solutions are still unknown.
\renewcommand{\baselinestretch}{1}\large
\begin{table}[!htbp]\centering
\caption{Summarized comparisons of the TSBMA algorithm against each reference algorithm on the 6240 DCKP instances of Set II.}
\begin{scriptsize}
\setlength{\tabcolsep}{1.2mm}{
\begin{tabular}{llllllllll}
\toprule[0.75pt]
\multirow{2}{*}{Class} & \multirow{2}{*}{Total} & \multicolumn{1}{c}{ILP$_{1,2}$ \cite{coniglio2020new}} & \multicolumn{1}{c}{BCM \cite{bettinelli2017branch}} & \multicolumn{1}{c}{CFS \cite{coniglio2020new}} & \multicolumn{2}{c}{TSBMA (this work)} & \multicolumn{3}{c}{TSBMA vs. CFS} \\
\cmidrule(lr){3-3} \cmidrule(lr){4-4} \cmidrule(lr){5-5} \cmidrule(lr){6-7} \cmidrule(lr){8-10}
& & Solved & Solved & Solved & Solved & New LB & \#Wins & \#Ties & \#Losses \\
\hline
$C1$ & 720 & \textbf{720} & \textbf{720} & \textbf{720} & \textbf{720} & 0 & 0 & 720 & 0 \\
$C3$ & 720 & 584 & \textbf{720} & \textbf{720} & 716 & 0 & 0 & 716 & 4 \\
$C10$ & 720 & 446 & 552 & \textbf{617} & \textbf{617} & \textbf{91} & 91 & 629 & 0 \\
$C15$ & 720 & 428 & 550 & \textbf{600} & \textbf{600} & \textbf{117} & 117 & 603 & 0 \\
$R1$ & 720 & \textbf{720} & \textbf{720} & \textbf{720} & 717 & 0 & 0 & 717 & 3 \\
$R3$ & 720 & 680 & \textbf{720} & \textbf{720} & \textbf{720} & 0 & 0 & 720 & 0 \\
$R10$ & 720 & 508 & 630 & \textbf{670} & 669 & \textbf{37} & 37 & 681 & 2 \\
$R15$ & 720 & 483 & 590 & \textbf{622} & \textbf{622} & \textbf{78} & 78 & 641 & 1 \\
$SC$ & 240 & \textbf{200} & 109 & 156 & 195 & \textbf{24} & 70 & 165 & 5 \\
$SR$ & 240 & 229 & 154 & 176 & 9 & \textbf{7} & 43 & 8 & 189 \\
\hline
Total on $CC$ and $CR$ & 5760 & 4569 & 5201 & 5389 & 5381 & 323 & 323 & 5427 & 10 \\
Total on $SC$ and $SR$ & 480 & 429 & 263 & 332 & 204 & 31 & 113 & 173 & 194 \\
\hline
Grand total & 6240 & 4998 & 5424 & 5721 & 5585 & 354 & 436 & 5600 & 204 \\
\bottomrule[0.75pt]
\end{tabular}}
\label{Summary_insII}
\end{scriptsize}
\end{table}
\renewcommand{\baselinestretch}{1}\large\normalsize
\section{Analysis and discussions}
\label{Sec_AD}
In this section, we analyze two essential components of the TSBMA algorithm: the importance of the threshold search and the contribution of the operation-prohibiting mechanism. The studies in this section are based on the 50 benchmark instances $11Iy$ to $20Iy$ of Set I.
\subsection{Importance of the threshold search}
\label{Analy_TS}
The threshold search procedure of the TSBMA algorithm is the first adaptation of the threshold accepting method to the DCKP. To assess the importance of this component, we compare TSBMA with two TSBMA variants by replacing the TSP procedure with the $first$-$improvement$ descent procedure and $best$-$improvement$ descent procedure. In other words, these variants (named as MA1 and MA2) use, in each iteration, the first and the best improving solution $S'$ in the neighborhood to replace the current solution, respectively. We carried out an experiment by running the two variants to solve the 50 instances $11Iy$ to $20Iy$ with the same experimental settings of Section \ref{Subsec_Setting}. The performance profiles of TSBMA and these TSBMA variants are shown in Figure \ref{Fig_PP2} based on the best objective values (left sub-figure) and the average objective values (right sub-figure).
From Figure \ref{Fig_PP2}, we can clearly observe that TSBMA dominates MA1 and MA2 according to the cumulative probability obtained by the $f_{best}$ and $f_{avg}$ values. The plots of TSBMA strictly run above the plots of MA1 and MA2, indicating TSBMA performs always better than the two variants. This experiment implies that the adopted threshold search procedure of TSBMA is relevant for its performance.
\begin{figure}[H]
\begin{minipage}{0.44\textwidth}
\includegraphics[width=3.1in]{PP_best.eps}
\end{minipage}
\hfill
\begin{minipage}{0.44\textwidth}
\includegraphics[width=3.1in]{PP_avg.eps}
\end{minipage}
\caption{Performance profiles of the compared algorithms on the 50 DCKP instances $11Iy$ to $20Iy$.}
\label{Fig_PP2}
\end{figure}
\subsection{Contribution of the operation-prohibiting mechanism}
\label{Analy_OP}
\renewcommand{\baselinestretch}{0.8}\large\normalsize
\begin{table}[!htbp]
\centering
\caption{Comparison between TSBMA$^-$ (without the OP mechanism) and TSBMA (with the OP mechanism) on the instances $11Iy$ to $20Iy$.}
\begin{scriptsize}
\setlength{\tabcolsep}{3mm}{
\begin{tabular}{l|lll|lll}
\toprule[0.75pt]
\multirow{2}{*}{Instance} & \multicolumn{3}{c|}{TSBMA$^-$} & \multicolumn{3}{c}{TSBMA} \\
\cmidrule(lr){2-4} \cmidrule(lr){5-7}
& \textit{$f_{best}$} & \textit{$f_{avg}$} & \textit{$std$}
& \textit{$f_{best}$} & \textit{$f_{avg}$} & \textit{$std$} \\
\hline
$11I1$ & 4960 & 4960 & 0.00 & 4960 & 4960 & 0.00 \\
$11I2$ & 4940 & 4940 & 0.00 & 4940 & 4940 & 0.00 \\
$11I3$ & 4950 & 4949.45 & 2.18 & 4950 & \textbf{4950} & 0.00 \\
$11I4$ & 4930 & 4924 & 4.42 & 4930 & \textbf{4930} & 0.00 \\
$11I5$ & 4920 & 4916.35 & 4.68 & 4920 & \textbf{4920} & 0.00 \\
$12I1$ & 4685 & 4676.95 & 4.99 & \textbf{4690} & \textbf{4687.65} & 2.22 \\
$12I2$ & 4670 & 4668.70 & 3.10 & \textbf{4680} & \textbf{4680} & 0.00 \\
$12I3$ & 4690 & 4685.45 & 4.20 & 4690 & \textbf{4690} & 0.00 \\
$12I4$ & 4680 & 4669.80 & 6.36 & 4680 & \textbf{4679.50} & 2.18 \\
$12I5$ & 4670 & 4664.50 & 4.57 & 4670 & \textbf{4670} & 0.00 \\
$13I1$ & 4525 & 4511.20 & 8.55 & \textbf{4539} & \textbf{4534.80} & 3.60 \\
$13I2$ & 4521 & 4509.25 & 7.29 & \textbf{4530} & \textbf{4528} & 4.00 \\
$13I3$ & 4520 & 4515.40 & 4.55 & \textbf{4540} & \textbf{4531} & 3.00 \\
$13I4$ & 4520 & 4507.10 & 6.94 & \textbf{4530} & \textbf{4529.15} & 2.29 \\
$13I5$ & 4530 & 4513.65 & 6.51 & \textbf{4537} & \textbf{4534.20} & 3.43 \\
$14I1$ & 4429 & 4413.55 & 7.41 & \textbf{4440} & \textbf{4440} & 0.00 \\
$14I2$ & 4420 & 4413.55 & 4.47 & \textbf{4440} & \textbf{4439.40} & 0.49 \\
$14I3$ & 4420 & 4415.20 & 4.70 & \textbf{4439} & \textbf{4439} & 0.00 \\
$14I4$ & 4420 & 4412.40 & 4.57 & \textbf{4435} & \textbf{4431.50} & 2.06 \\
$14I5$ & 4420 & 4413.85 & 4.27 & \textbf{4440} & \textbf{4440} & 0.00 \\
$15I1$ & 4359 & 4346.15 & 5.06 & \textbf{4370} & \textbf{4369.95} & 0.22 \\
$15I2$ & 4359 & 4344.10 & 6.22 & \textbf{4370} & \textbf{4370} & 0.00 \\
$15I3$ & 4359 & 4341.85 & 6.54 & \textbf{4370} & \textbf{4369.25} & 1.84 \\
$15I4$ & 4350 & 4341.05 & 7.78 & \textbf{4370} & \textbf{4369.85} & 0.36 \\
$15I5$ & 4360 & 4346.10 & 5.47 & \textbf{4379} & \textbf{4373.15} & 4.29 \\
$16I1$ & 5020 & 5013.75 & 4.93 & 5020 & \textbf{5020} & 0.00 \\
$16I2$ & 5010 & 5003.30 & 5.60 & 5010 & \textbf{5010} & 0.00 \\
$16I3$ & 5020 & 5010.65 & 5.33 & 5020 & \textbf{5020} & 0.00 \\
$16I4$ & 5020 & 5008.95 & 8.24 & 5020 & \textbf{5020} & 0.00 \\
$16I5$ & 5060 & 5052.85 & 8.37 & 5060 & \textbf{5060} & 0.00 \\
$17I1$ & 4730 & 4707.50 & 7.51 & 4730 & \textbf{4729.70} & 0.64 \\
$17I2$ & 4716 & 4704.50 & 6.27 & \textbf{4720} & \textbf{4719.50} & 2.18 \\
$17I3$ & 4720 & 4705.10 & 6.68 & \textbf{4729} & \textbf{4723.60} & 4.41 \\
$17I4$ & 4722 & 4701.20 & 9.68 & \textbf{4730} & \textbf{4730} & 0.00 \\
$17I5$ & 4720 & 4706.20 & 8.37 & \textbf{4730} & \textbf{4726.85} & 4.50 \\
$18I1$ & 4555 & 4539.75 & 6.31 & \textbf{4568} & \textbf{4565.80} & 3.40 \\
$18I2$ & 4540 & 4532.20 & 4.64 & \textbf{4560} & \textbf{4551.40} & 3.01 \\
$18I3$ & 4570 & 4545.20 & 8.58 & 4570 & \textbf{4569.40} & 2.20 \\
$18I4$ & 4550 & 4539.30 & 6.75 & \textbf{4568} & \textbf{4565.20} & 3.12 \\
$18I5$ & 4550 & 4542.50 & 5.32 & \textbf{4570} & \textbf{4567.95} & 3.46 \\
$19I1$ & 4432 & 4424.65 & 4.71 & \textbf{4460} & \textbf{4456.65} & 3.48 \\
$19I2$ & 4443 & 4430.85 & 6.06 & \textbf{4460} & \textbf{4453.25} & 4.17 \\
$19I3$ & 4440 & 4428.15 & 6.01 & \textbf{4469} & \textbf{4462.05} & 4.04 \\
$19I4$ & 4450 & 4431.25 & 5.63 & \textbf{4460} & \textbf{4453.20} & 3.89 \\
$19I5$ & 4449 & 4435.65 & 5.42 & \textbf{4466} & \textbf{4460.75} & 1.61 \\
$20I1$ & 4364 & 4358.95 & 2.80 & \textbf{4390} & \textbf{4383.20} & 3.36 \\
$20I2$ & 4360 & 4356.85 & 4.25 & \textbf{4390} & \textbf{4381.80} & 3.78 \\
$20I3$ & 4370 & 4360.45 & 5.11 & \textbf{4389} & \textbf{4387.90} & 2.77 \\
$20I4$ & 4370 & 4359.75 & 5.78 & \textbf{4389} & \textbf{4380.40} & 1.98 \\
$20I5$ & 4366 & 4357.45 & 4.78 & \textbf{4390} & \textbf{4386.40} & 4.05 \\
\hline
\#Avg & 4603.08 & 4593.13 & 5.56 & \textbf{4614.14} & \textbf{4611.83} & 1.80 \\
\#Best & 15/50 & 2/50 & - & 50/50 & 50/50 & - \\
$p$-$values$ & 2.51e-7 & 1.68e-9 & - & - & - & - \\
\bottomrule[0.75pt]
\end{tabular}}
\label{Result_OP}
\end{scriptsize}
\end{table}
\renewcommand{\baselinestretch}{1}\large\normalsize
TSBMA avoids revisiting previously encountered solutions with the OP mechanism introduced in Section \ref{Subsec_OP}. To assess the usefulness of the OP mechanism, we created a TSBMA variant (denoted by TSBMA$^-$) by disabling the OP component and keeping the other components unchanged. We ran TSBMA$^-$ to solve the 50 $11Iy$ to $20Iy$ instances according to experimental settings given in Section \ref{Subsec_Setting} and reported the results in Table \ref{Result_OP}. The first column gives the name of each instance and the remaining columns show the best objective values ($f_{best}$), the average objective values ($f_{avg}$) and the standard deviations ($std$). Row \#Avg presents the average value of each column and row \#Best indicates the number of instances for which an algorithm obtains the best values between the two sets of results. The last row shows the $p$-$values$ from the Wilcoxon signed-rank test. The best results of the compared algorithms are highlighted in bold.
From Table \ref{Result_OP}, we observe that TSBMA$^-$ performs worse than TSMBA. TSBMA$^-$ obtains worse $f_{best}$ values for 35 out of the 50 instances and worse $f_{avg}$ values for 48 instances. Considering the $std$ values, TSBMA$^-$ shows a much less stable performance than TSMBA. Moreover, the small $p$-$values$ ($ < 0.05$) from the Wilcoxon tests confirm the statistically significant difference between the results of TSMBA and TSBMA$^-$. This experiment demonstrates the effectiveness and robustness of the operation-prohibiting mechanism employed by the TSMBA algorithm.
\section{Conclusions}
\label{Sec_Conclu}
The disjunctively constrained knapsack problem is a well-known NP-hard model. Given its practical significance and intrinsic difficulty, a variety of exact and heuristic algorithms have been designed for solving the problem. We proposed the threshold search based memetic algorithm that combines for the first time threshold search with the memetic framework.
Extensive evaluations on a large number of benchmark instances in the literature (6340 instances in total) showed that the algorithm performs competitively with respect to the state-of-the-art algorithms. Our approach is able to discover 24 new lower bounds out of the 100 instances of Set I and 354 new lower bounds out of the 6240 instances of Set II. These new lower bounds are useful for future studies on the DCKP. The algorithm also attains the best-known or known optimal results on most of the remaining instances. We carried out additional experiments to investigate the two essential ingredients of the algorithm (the threshold search technique and the operation-prohibiting mechanism). The disjunctively constrained knapsack problem is a useful model to formulate a number of practical applications. The algorithm and its code (that we will make available) can contribute to solving these problems.
There are at least two possible directions for future work. First, TSBMA performed badly on most random sparse instances of $SR$. It would be interesting to improve the algorithm to better handle such instances. Second, given the good performance of the adopted approach, it is worth investigating its underlying ideas to solve related problems discussed in the introduction.
\section*{Declaration of competing interest}
The authors declare that they have no known competing interests that could have appeared to influence the work reported in this paper.
\section*{Acknowledgments}
We would like to thank Dr. Zhe Quan, Dr. Lei Wu, Dr. Pablo San Segundo and their co-authors for sharing the instances of the DCKP and the detailed results of their algorithms reported in \cite{quan2017cooperative}, \cite{quan2017design}, and \cite{coniglio2020new}.
\bibliographystyle{elsart-num-sort}
|
\section{Introduction}\label{sec:introduction}}
\IEEEPARstart{I}{mage} segmentation has been studied over many decades using conventional methods, and in the past few years using deep learning.
Several different conventional approaches, such as interactive methods, active contour (level-set methods), graph-theoretic approaches, perceptual grouping methods and so on, have been studied for image segmentation over the past decades.
Yet automatic methods fail where boundaries are complex. Interactive methods let humans resolve the complex cases. Interactive methods,~\cite{DBLP:journals/ijcv/RussellTMF08, DBLP:journals/cviu/WangHC14, DBLP:conf/cvpr/CastrejonKUF17, qin2018bylabel}, are usually able to produce accurate and robust results, but with significant time costs.
Active contour \cite{osher1988fronts, RC30ActiveContour, RC3ActiveContour, RC63ActiveContour, RC8ActiveContour, caselles1997geodesic, RC67ActiveContour}, graph-theoretic \cite{wu1993optimal, cox1996ratio, shi2000normalized, sarkar2000supervised, wang2003image} and perceptual grouping \cite{GestaltLaws, elder1993effect, mahamud2003segmentation, amir1998generic, elder2003contour, DBLP:conf/iccv/RenFM05, qin2017bmvc, qin2017iros}, methods require almost no human interventions so that they are faster than interactive methods. However, they are relatively less robust.
In recent years, to achieve accurate, robust and fast performance, many deep learning models \cite{minaee2020image} have been developed for image segmentation. Semantic image segmentation \cite{long2015fully, huang2017densely} is one of the most popular topics, which aims at labeling every pixel in an image, with one of the several predefined class labels. It has been widely used in many applications, such as scene understanding \cite{lin2014microsoft, zhou2017scene}, autonomous driving \cite{Geiger2013IJRR, Cordts2016Cityscapes}, \emph{etc}. The targets in these applications are usually large in size, so most existing methods focus on achieving robustness performance with high regional accuracy. Less attention has been paid to the high spatial accuracy of boundaries and fine structures. However, many other applications, \emph{e.g.}~image segmentation/editing \cite{qin2018accurate,DBLP:journals/ijcv/KadirB01,goferman2012context,qin2018bylabel} and manipulation \cite{DBLP:conf/iccv/Jagersand95,DBLP:conf/wacv/MechrezSZ18}, visual tracking \cite{DBLP:conf/wacv/LeeK18,DBLP:conf/iros/QinHQSDJ17,DBLP:conf/bmvc/QinHZDJ17}, vision guided robot hand manipulation \cite{qin2017iros} and so on, require highly accurate object boundaries and fine structures.
There are two main challenges in accurate image segmentation:
\textbf{Firstly}, large-scale features play important roles in classifying pixels since they can provide more semantic information compared with local features. However, large-scale features are usually obtained from deep low-resolution feature maps or by large size kernels and the spatial resolution is sacrificed.
Simple upsampling of the low-resolution feature maps to high resolution is not able to recover the fine structures \cite{long2015fully}.
Thus, many encoder-decoder architectures \cite{ronneberger2015u} have been developed for segmenting edges or thin structures.
Their skip connections and gradual upsampling operations play important roles in recovering the high resolution probability maps.
Additionally, different cascaded or iterative architectures \cite{xu2017deep,wang2018detect,deng2018r3net,DBLP:conf/iccv/WangBZZL17} have been introduced to further improve the segmentation accuracy by gradually refining the coarse predictions, which sometimes leads to complicated network architectures and computational bottleneck.
\textbf{Secondly}, most of the image segmentation models use cross entropy (CE) loss to supervise the training process.
CE loss usually gives greater penalties on these seriously erroneous predictions (\emph{e.g.}~predict ``1'' as ``0.1'' or predict ``0'' as ``0.9'').
Therefore, deep models trained with CE loss compromise andes prefer to predict ``hard'' samples with a non-committed ``0.5''. In image segmentation tasks, the boundary pixels of targets are usually the hard samples, so this will lead to blurry boundaries in predicted segmentation probability maps.
Other losses, such as intersection-over-union (IoU) loss \cite{rahman2016optimizing,mattyus2017deeproadmapper,DBLP:conf/bmvc/NagendarSBJ18}, F-score loss \cite{DBLP:journals/corr/abs-1805-07567} and Dice-score loss \cite{DBLP:conf/miccai/FidonLGEKOV17a},
have also been introduced to image segmentation tasks for handling biased training sets.
These are sometimes able to achieve higher (regional) evaluation metrics, \emph{e.g.}~IoU, F-score, since their optimization targets are consistent with these metrics.
However, they are not specifically designed for capturing fine structures and often produce ``biased'' results, which tend to emphasize the large structures while neglecting fine details.
To address the above issues, we propose a novel but simple \textbf{B}oundary-\textbf{A}ware \textbf{S}egmentation \textbf{Net}work (\textbf{BASNet}), which consists of a predict-refine network and a hybrid loss, for highly accurate image segmentation.
The predict-refine architecture is designed to predict and refine the predicted probability maps sequentially.
It consists of a U-Net-like \cite{ronneberger2015u} deeply supervised \cite{lee2015deeply,xie2015holistically} ``heavy'' encoder-decoder network and a residual refinement module with ``light'' encoder-decoder structure. The ``heavy'' encoder-decoder network transfers the input image to a segmentation probability map, while the ``light'' refinement module refines the predicted map by learning the residuals between the coarse map and ground truth (GT).
In contrast to \cite{peng2017large, islamsalient, deng2018r3net}, which iteratively use refinement modules on saliency predictions or intermediate feature maps at multiple scales, our refinement module is used only once on the original scale of the segmentation maps. Overall, our predict-refine architecture is concise and easy to use.
The hybrid loss combines the binary cross entropy (BCE) \cite{DBLP:journals/anor/BoerKMR05}, structural similarity (SSIM) \cite{wang2003multiscale} and IoU losses \cite{mattyus2017deeproadmapper}, to supervise the training process in a three-level hierarchy: pixel-, patch- and map- level.
Instead of using \textit{explicit} boundary loss (NLDF+ \cite{luo2017non}, C2S \cite{DBLP:conf/eccv/LiYCLS18}, BANet\cite{su2019selectivity}),
we \textit{implicitly} inject the goal of accurate boundary prediction in the hybrid loss, contemplating that it may help reduce spurious error from cross propagating the information learned from the boundaries and other regions on the image (see \figref{fig:intro}).
\begin{figure}[t!]
\centering
\begin{overpic}[width=\columnwidth]{./Imgs/TeaserImage-min.pdf}
\put(7,-3) {(a) Image}
\put(45,-3) {(b) GT}
\put(75,-3) {(c) BASNet}
\end{overpic}
\vspace{2pt}
\caption{Sample results of our BASNet on salient object detection (Top) and camouflaged object detection (Bottom).
}\label{fig:intro}
\end{figure}
In addition to proposing novel segmentation techniques, developing novel segmentation applications also plays a very important role in advancing the segmentation field.
Therefore, we developed two novel BASNet-based applications: \textbf{AR COPY \& PASTE}\footnote{\url{https://clipdrop.co/}} and \textbf{OBJECT CUT}\footnote{\url{https://objectcut.com/}}.
AR COPY \& PASTE is a mobile app built upon our BASNet model and the Augmented Reality techniques.
By using cellphones, it provides a novel interactive user experience where users can ``COPY'' the real-world targets and ``PASTE'' them into desktop software .
Specifically, AR COPY \& PASTE allows users to take a photo of an object using a mobile device. Then the background removed object returned by our remote BASNet server will be shown in the camera view. In this view, the ``COPIED“ object is overlapped with real scene video stream.
Users can move and target the mobile camera at a specific position on the desktop screen.
Then tapping the screen of the mobile device will trigger the ``PASTE'' operation, which transmits the object from the mobile device to the software opened in the desktop. Meanwhile, OBJECT CUT provides a web-based service for automatic image background removal based on our BASNet model. An image can be uploaded from a local machine or through an URL. This application greatly facilitates the background removal for users who have no image editing experience or software.
The main contributions can be summarized as:
\begin{itemize}[noitemsep]
\item
We develop a novel boundary-aware image segmentation network, BASNet, which consists of a deeply supervised encoder-decoder and a residual refinement module, and a novel hybrid loss that fuses BCE, SSIM, and IoU to supervise the training process of accurate image segmentation on three levels: pixel-level, patch-level and map-level.
\item
We conduct thorough evaluations of the proposed method including a comparison with 25 state-of-the-art (SOTA) methods on six widely used public salient object segmentation datasets, a comparison with 16 models on the SOC (Salient Object in Clutter) dataset and a comparison with 13 camouflaged object detection (COD) models on three public COD datasets. BASNet~achieves very competitive performance in terms of regional evaluation metrics, while outperforms other models in terms of boundary evaluation metrics.
\item
We develop two (close to) commercial applications, AR COPY \& PASTE and OBJECT CUT, based on our BASNet. These two applications further demonstrate the simplicity, effectiveness and efficiency of our model.
\end{itemize}
Compared with the CVPR version~\cite{qin2019basnet} of this work, the following extensions are made. First, deeper theoretical explanations of the hybrid loss design are added. Second, more comprehensive and thorough experiments on different datasets, including salient objects in clutter (SOC) and COD, are included. Third, two (close to) commercial applications, AR COPY \& PASTE and OBJECT CUT, are developed.
\section{Related Works}
\subsection{Traditional Image Segmentation Approaches}
Watershed \cite{vincent1991watersheds}, graph cut \cite{kumar2005obj,ladicky2010graph}, active contour \cite{osher1988fronts}, perceptual grouping \cite{qin2017bmvc} as well as the interactive methods based on these approaches mainly rely on well-designed handcrafted features, objective functions and optimization algorithms.
Watershed and graph cut approaches segment images based on the regional pixel similarities, so they are less effective in segmenting very fine structures and achieving smooth and accurate segmentation boundaries.
Active contour and perceptual grouping methods can be considered as boundary based approaches. Active contour methods represent the 2D segmentation contour by the level set of a 3D function. Instead of directly evolving the 2D contour, this group of approaches evolves the 3D function to find the optimal segmentation contour, which avoids complicated 2D contour splitting and merging issues. Perceptual grouping methods segment images by selecting and grouping subsets of the detected edge fragments or line segments from given images to formulate closed or open contours of the targets to be segmented. However, although these methods are able to produce relatively accurate boundaries, they are very sensitive to noise and local minima, which usually leads to less robust and unreliable performance.
\subsection{Patch-wise Deep Models}
To improve the robustness and accuracy, deep learning methods have been widely introduced to image segmentation \cite{qin2020visual}. Early deep methods use existing image classification networks as feature extractors and formulate the image segmentation tasks as patch-wise image pixel (super-pixel) \cite{li2015visual, liu2015predicting, wang2015deep, zhao2015saliency, li2016visual} classification problems. These models greatly improve the segmentation robustness in some tasks due to the strong fitting capability of deep neural networks. However, they are still not able to produce high spatial accuracy, let alone segmenting fine structures. The main reason is probably that the pixels in patch-wise models are classified independently based on the local features inside each patch and larger-scale spatial contexts are not used.
\subsection{FCN and its Variants}
With the development of fully convolutional network (FCN) \cite{long2015fully}, deep convolutional neural networks have become a standard solution for image segmentation problems.
Large number of deep convolutional models \cite{minaee2020image} have been proposed for image segmentation. FCN adapts classification backbones, such as VGG \cite{simonyan2014very}, GoogleNet \cite{szegedy2015going}, ResNet \cite{he2016deep} and DenseNet \cite{huang2017densely}, by discarding the fully connected layers and directly upsampling the output features of certain convolutional layers with specific scales to build a fully convolutional image segmentation model. However, the direct upsampling from low resolution fails in capturing accurate structures. Therefore, The DeepLab family \cite{chen2014semantic, chen2017deeplab, chen2017rethinking} replaces the pooling operations by atrous convolutions to avoid degrading the feature map resolution. Besides, they also introduce a densely connected Conditional Random Field (CRF) to improve the segmentation results. However, applying atrous convolutions on high-resolution maps leads to larger memory costs and CRF usually yields noisy segmentation boundaries.
Holistically Edge Detection (HED) \cite{xie2015holistically}, RCF \cite{girshick2014rich} and CASENet \cite{yu2017casenet} are proposed to directly segment edges by making full use of the features from both the shallow and deep stages of the image classification backbones.
Besides, many variants \cite{li2016deep, kruthiventi2016saliency,hu2017deep} of FCN have been proposed for salient object detection (binary-class image segmentation) \cite{wang2019salient}.
Most of these works are focusing on either developing novel multi-scale feature aggregation strategies or designing new multi-scale feature extraction modules.
Zhang {\em et al.~} (Amulet) \cite{DBLP:conf/iccv/ZhangWLWR17} developed a generic framework for aggregating multi-level convolutional features of the VGG backbone.
Inspired by HED \cite{xie2015holistically}, Hou {\em et al.~} (DSS+) \cite{hou2017deeply} introduced short connections to the skip-layer structures of HED to better use the deep layer features.
Chen {\em et al.~} (RAS) \cite{DBLP:conf/eccv/ChenTWH18} developed a reverse attention model to iteratively refine the side-outputs from a HED-like architecture.
Zhang {\em et al.~} (LFR) \cite{DBLP:conf/ijcai/ZhangLLS18} designed a symmetrical fully convolutional network which takes images and their reflection as inputs to learn the saliency features from the complementary input spaces.
Instead of passing the information with single direction (deep to shallow or shallow to deep), Zhang {\em et al.~} (BMPM) \cite{zhang2018bi} proposed to have the information passed between the shallow and deep layers by a controlled bi-directional passing module.
\subsection{Encoder-decoder Architectures}
Rather than directly upsampling features from deep layers of the backbones, SegNet \cite{badrinarayanan2017segnet} and U-Net \cite{ronneberger2015u} employ encoder-decoder like structures to gradually up-sample the deep low-resolution feature maps. Combined with skip connections, they are able to recover more details.
One of the main characteristics of these models is the symmetrical downsampling and upsampling operations.
To reduce the checkerboard artifacts in the prediction, Zhang {\em et al.~} (UCF) \cite{DBLP:conf/iccv/ZhangWLWY17} reformulated the dropout and developed a hybrid module for the upsampling operation.
To better use the features extracted by backbones, Liu {\em et al.~} (PoolNet) \cite{liu2019simple} constructed the decoder part using their newly developed feature aggregation, pyramid pooling and global guidance modules.
In addition, stacked HourglassNet \cite{newell2016stacked}, CU-UNet \cite{tang2018quantized}, UNet++ \cite{zhou2018unet++} and U$^2$-Net \cite{qin2020u2} further explore diverse ways of improving the encoder-decoder architectures by cascaded or nested stacking.
\begin{figure*}[t!]
\centering
\begin{overpic}[width=\textwidth]{./Imgs/EncoderDecoder.jpg}
\end{overpic}
\caption{Architecture of the proposed boundary-aware segmentation network: BASNet. See \secref{sec:Methodology} for details.
}\label{fig:arc}
\end{figure*}
\subsection{Deep Recurrent Models}
Recurrent techniques have been widely applied in image segmentation models.
Kuen {\em et al.~} \cite{kuen2016recurrent} proposed to achieve the completely segmented results by sequentially segmenting image sub-regions using a recurrent framework.
Liu {\em et al.~} (PiCANetR) \cite{liu2018picanet} deployed a bidirectional LSTM along the row and column of the feature maps respectively to generate the pixel-wise attention maps for salient object detection.
Hu {\em et al.~} (SAC-Net) \cite{hu2020sac} developed similar strategies to~\cite{liu2018picanet} to capture spatial attenuation context for image segmentation.
Zhang {\em et al.~} (PAGRN) \cite{zhang2018progressive} proposed to transfers global information from deep to shallower layers via a multi-path recurrent connection.
Wang {\em et al.~} (RFCN) \cite{wang2018salient} built a cascaded network by stacking multiple encoder-decoders to recurrently correcting the prediction errors of the previous stages.
Instead of iteratively refining the segmentation results \cite{wang2018salient}, Hu {\em et al.~} (RADF+) \cite{DBLP:conf/aaai/HuZQFH18} recurrently aggregated and refined multi-layer deep features to achieve accurate segmentation results.
However, due to the serial connections between each recurrent step, models using the ``recurrent'' techniques are relatively less efficient in terms of time costs.
\subsection{Deep Coarse-to-Fine Models}
This group of models aims at improving the segmentation results by gradually refining the coarse predictions.
Lin {\em et al.~} (RefineNet) \cite{lin2017refinenet} developed a multi-path refinement segmentation network, which uses long-range residual connections to exploit the information along the down-sampling process.
Liu {\em et al.~} (DHSNet) \cite{liu2016dhsnet} proposed a hierarchical recurrent
convolutional neural network (HRCNN), which hierarchically and progressively refines the segmentation results in a coarse-to-fine manner.
Wang {\em et al.~} (SRM) \cite{DBLP:conf/iccv/WangBZZL17} developed a multi-stage framework for segmentation map refinement, in which each stage takes the input image and the segmentation maps (lower resolution) from the last stage to produce higher-resolution results.
Deng {\em et al.~} (R$^3$Net+) \cite{deng2018r3net} proposed to alternatively refine the segmentation results based on the shallow, high-resolution and deep low-resolution feature maps.
Wang {\em et al.~} (DGRL) \cite{wang2018detect} developed a global-to-local framework which first localizes the to-be-segmented targets globally and then refines these targets using a local boundary refinement module.
The coarse to fine models reduce the probability of overfitting and show promising improvements in accuracy.
\subsection{Boundary-assisted Deep Models}
Region and boundaries are mutually determined. Therefore, many models introduce boundary information to assist segmentation.
Luo {\em et al.~} (NLDF) \cite{luo2017non} proposed to supervise a 4$\times$5 grid structure adapted from VGG-16 by fusing the cross entropy and the boundary IoU inspired by Mumford-Shah \cite{mumford1989optimal}.
Li {\em et al.~} (C2S) \cite{DBLP:conf/eccv/LiYCLS18} tried to recover the regional saliency segmentation from segmented contours.
Su {\em et al.~} (BANet) \cite{su2019selectivity} developed a boundary-aware segmentation network with three separate streams: a boundary localization stream, an interior perception stream and a transition compensation stream for boundary, region and boundary/region transition prediction, respectively.
Zhao {\em et al.~} (EGNet) \cite{zhao2019egnet} proposed an edge guidance network for salient object segmentation by explicitly modeling and fusing complementary region and boundary information.
Most of models in this category explicitly use boundary information as either an additional supervision loss or a assisting prediction stream for inferring the region segments.
In this paper, we propose a simple predict-refine architecture which takes advantage of both the encoder-decoder architecture and the coarse-to-fine strategy. Besides, instead of explicitly using boundary loss or additional boundary prediction streams, we design a simple hybrid loss which implicitly describes the dissimilarity between the segmentation prediction and the ground truth at three levels: pixel-, patch- and map-level. The predict-refine architecture together with the hybrid loss provides a simple yet powerful solution for image segmentation and some close to commercial applications.
\section{Methodology}\label{sec:Methodology}
\subsection{Overview}
Our BASNet architecture consists of two modules as shown in Fig. \ref{fig:arc}.
The prediction module is a U-Net-like densely supervised Encoder-Decoder network \cite{ronneberger2015u}, which learns to predict segmentation probability maps from input images.
The multi-scale Residual Refinement Module (RRM) refines the resulting map of the prediction module by learning the residuals between the coarse map and the GT.
\subsection{Prediction Module}\label{sec:pm}
Share same spirit of U-Net \cite{ronneberger2015u} and SegNet \cite{badrinarayanan2017segnet}, we design our segmentation prediction module as an encoder-decoder fashion, since this kind of architectures is able to capture high-level global contexts and low-level details at the same time.
To reduce over-fitting, the last layer of each decoder stage is supervised by the GT, inspired by HED \cite{xie2015holistically} (see Fig. \ref{fig:arc}).
The encoder has an input convolutional layer and six stages comprised of basic res-blocks.
The input convolutional layer and the first four stages are adopted from ResNet-34 \cite{he2016deep}.
The difference is that our input layer has 64 convolutional filters with a size of 3$\times$3 and stride of 1 rather than a size of 7$\times$7 and stride of 2.
Additionally, there is no pooling operation after the input layer.
This means that the feature maps before the second stage have the same spatial resolution as the input image.
This is different from the original ResNet-34, which has a quarter of the resolution in the first feature map.
This adaptation enables the network to obtain higher resolution feature maps in earlier layers, while decreasing the overall receptive fields.
To achieve the same receptive field as ResNet-34 \cite{he2016deep}, we add two more stages after the fourth stage of ResNet-34.
Both stages consist of three basic res-blocks with 512 filters after a non-overlapping max pooling layer of size 2.
To further capture global information, we add a bridge stage between the encoder and decoder.
It consists of three convolutional layers with 512 dilated (dilation=2)~\cite{yu2015multi} 3$\times$3 filters.
Each of these convolutional layers is followed by a batch normalization \cite{ioffe2015batch} and a ReLU activation function \cite{hahnloser2001permitted}.
\begin{figure}[t!]
\centering
\begin{overpic}[width=\columnwidth]{./Imgs/coarse-min.pdf}
\put(25.5,48){(a)}
\put(75.5,48){(b)}
\put(25.5,10){(c)}
\put(75.5,10){(d)}
\end{overpic}
\caption{Illustration of different aspects of coarse prediction in one-dimension: (a) Red: probability plot of GT, (b) Green: probability plot of coarse boundary not aligning with GT, (c) Blue: coarse region having too low probability, (d) Purple: real coarse predictions usually have both (b\&c) problems.
}\label{fig:coarse}
\end{figure}
Our decoder is almost symmetrical to the encoder.
Each stage consists of three convolution layers followed by a batch normalization and a ReLU activation function.
The input of each stage is the concatenated feature maps of the up-sampled output from its previous stage and its corresponding stage in the encoder.
To achieve the side-output maps, the multi-channel output of the bridge stage and each decoder stage is fed to a plain $3\times 3$ convolution layer followed by a bilinear upsampling and a sigmoid function.
Therefore, given an input image, our prediction module produces seven segmentation probability maps in the training process.
Although every predicted map is up-sampled to the same size with the input image, the last one has the highest accuracy and hence is taken as the final output of the prediction module. This output is passed to the refinement module.
\subsection{Residual Refinement Module}\label{sec:rm}
Refinement Modules (RMs) \cite{islamsalient, deng2018r3net} are usually designed as a residual block \cite{xu2017deep} that refines the coarse segmentation maps $S_{coarse}$ by learning the residuals $S_{residual}$ between the coarse maps and the GT, as
\begin{equation}
\centering
S_{refined}=S_{coarse}+S_{residual}.
\label{equ:rrm}
\end{equation}
Before introducing our refinement module, the term ``coarse'' has to be determined.
Here, ``coarse'' includes two aspects.
One is blurry and noisy boundaries (see the one-dimension illustration in \figref{fig:coarse}(b))
The other one is the unevenly predicted regional probabilities (see \figref{fig:coarse}(c)).
As shown in \figref{fig:coarse}(d), real predicted coarse maps usually contain both coarse cases.
The residual refinement module based on local context (RRM\_LC), \figref{fig:rf1}, was originally designed for boundary refinement \cite{peng2017large}.
Since its receptive field is small, Islam {\em et al.~} \cite{islamsalient} and Deng {\em et al.~} \cite{deng2018r3net} iteratively or recurrently used it for refining segmentation probability maps on different scales.
Wang {\em et al.~} \cite{DBLP:conf/iccv/WangBZZL17} adopted the pyramid pooling module from \cite{he2014spatial}, in which three-scale pyramid pooling features are concatenated.
To avoid losing details caused by pooling operations, RRM\_MS
( \figref{fig:rf2}) uses convolutions with different kernel sizes and dilations \cite{yu2015multi, zhang2018bi} to capture multi-scale contexts.
However, these modules are shallow thus hard to capture high-level information for refinement.
To refine inaccuracies in coarse segmentation maps of image reigons and boundaries,
we develop a novel residual refinement module.
Our RRM employs the residual encoder-decoder architecture, RRM\_Ours (see Figs. \ref{fig:arc} and \ref{fig:rf5}).
Its main architecture is similar but simpler than our prediction module.
It contains an input layer, an encoder, a bridge, a decoder and an output layer.
Different from the prediction module, both the encoder and decoder have four stages.
Each stage only has one convolutional layer.
Each layer has 64 filters of size $3\times 3$ followed by a batch normalization and a ReLU activation function.
The bridge stage also has a convolutional layer with 64 filters of size $3\times 3$ followed by a batch normalization and ReLU activation.
Non-overlapping max pooling is used for downsampling in the encoder and bilinear interpolation is utilized for upsampling in the decoder.
The output of this RM module is used as the final generating segmentation results of BASNet.
\subsection{Hybrid Loss}
Our training loss is defined as the summation over all outputs:
\begin{equation}
\centering
\mathcal{L} = {\textstyle \sum}_{k=1}^K{\alpha_k \ell^{(k)}},
\label{equ:loss}
\end{equation}
where $\ell^{(k)}$ is the loss of the $k$-th side output, $K$ denotes the total number of the outputs and $\alpha_k$ is the weight of each loss.
As described in Sec.~\ref{sec:pm} and Sec.~\ref{sec:rm}, our segmentation model is deeply supervised with eight outputs, i.e. $K=8$, including seven outputs from the prediction module and one output from the refinement module.
To obtain high quality regional segmentation and clear boundaries, we propose to define $\ell^{(k)}$ as a hybrid loss:
\begin{equation}
\centering
\ell^{(k)} = \ell_{bce}^{(k)} + \ell_{ssim}^{(k)} + \ell_{iou}^{(k)},
\label{equ:lside}
\end{equation}
where $\ell_{bce}^{(k)}$,~$\ell_{ssim}^{(k)}$, and $\ell_{iou}^{(k)}$ denote BCE loss \cite{DBLP:journals/anor/BoerKMR05}, SSIM loss \cite{wang2003multiscale} and IoU loss \cite{mattyus2017deeproadmapper}, respectively.
BCE \cite{DBLP:journals/anor/BoerKMR05} loss is the most widely used loss in binary classification and segmentation. It is defined as:
\begin{equation}
\centering
{\scriptstyle
\ell_{bce} = -\sum\limits_{(r,c)}[G(r,c)\log(S(r,c))+(1-G(r,c))\log(1-S(r,c))]},
\label{equ:bce_loss}
\end{equation}
where $G(r,c)\in\{0,1\}$ is the GT label of the pixel $(r,c)$ and $S(r,c)$ is the predicted probability of segmented object.
SSIM~\cite{wang2003multiscale} was originally devised for image quality assessment.
It captures the structural information in an image.
Hence, we integrated it into our training loss to learn the structural information of the GT.
Let $\mathbf{x}=\{x_j:j=1,...,N^2\}$ and $\mathbf{y}=\{y_j:j=1,...,N^2\}$ be the pixel values of two corresponding patches (size: $N\times N$) cropped from the predicted probability map $S$ and the binary GT mask $G$, respectively. The SSIM of $\mathbf{x}$ and $\mathbf{y}$ is defined as:
\begin{equation}
\centering
\small{\ell_{ssim}=1 - \frac{(2\mu_x\mu_y+C_1)(2\sigma_{xy}+C_2)}{(\mu_x^2+\mu_y^2+C_1)(\sigma_x^2+\sigma_y^2+C_2)}}
\label{equ:ssim_loss}
\end{equation}
where $\mu_x$, $\mu_y$ and $\sigma_x$, $\sigma_y$ are the mean and standard deviations of $\mathbf{x}$ and $\mathbf{y}$ respectively, $\sigma_{xy}$ is covariance, $C_1=0.01^2$ and $C_2=0.03^2$ are used to avoid dividing by 0.
IoU was originally proposed for measuring the similarity between two sets \cite{jaccard1912distribution} and has become a standard evaluation measure for object detection and segmentation.
Recently, it has been used as a training loss \cite{rahman2016optimizing,mattyus2017deeproadmapper}.
To ensure its differentiability, we adopted the IoU loss used in \cite{mattyus2017deeproadmapper}:
\begin{equation}
\centering
\ell_{iou} = 1 - \tfrac{\sum\limits_{r=1}\limits^H\sum\limits_{c=1}\limits^WS(r,c)G(r,c)}{\sum\limits_{r=1}\limits^H\sum\limits_{c=1}\limits^W[S(r,c)+G(r,c)-S(r,c)G(r,c)]}
\label{equ:liou}
\end{equation}
where $G(r,c)\in\{0,1\}$ is the GT label of the pixel $(r,c)$ and $S(r,c)$ is the predicted probability of segmented object.
\figref{fig:hybrid_loss} illustrated the impact of each of the three losses.
Fig.~\ref{fig:hybrid_loss} (a) and (b) are the input image and its ground truth segmentation mask.
It is worth noting that the probability maps in Fig.~\ref{fig:hybrid_loss} are generated by fitting a single pair of image and its ground truth (GT) mask. Hence, after a certain number of iterations, all the losses are able to produce perfect results due to over-fitting.
Here, we ignore the final fitting results and aim to observe the different characteristics and problems of these losses in the fitting process.
The (c), (d), (e) and (f) columns show changes of the intermediate probability maps as the training progresses.
The BCE loss is computed pixel-wise. It does not consider the labels of the neighborhood and it weights both the foreground and background pixels equally.
This helps with the convergence on all pixels and guarantee a relatively good local optima.
Since significantly erroneous predictions (predicting 0 as 0.9 or predicting 1 as 0.1) produce large BCE loss, the models trained with BCE loss suppress these errors by giving prediction values around 0.5 around the boundaries, which often leads to blurred boundaries and fine structures, as we can see from the second row of column (c), where the contour of the whole foreground region is blurring, and the third row, in which the cable below the backpack is with low probability values.
\begin{figure}[t!]
\vspace{-5pt}
\begin{center}
\subfigure[RRM\_LC]{\includegraphics[height=0.42\linewidth]{./figures/rrm/RM1.png}
\label{fig:rf1}}
\hfill
\subfigure[RRM\_MS]{\includegraphics[height=0.42\linewidth]{./figures/rrm/RM2.png}
\label{fig:rf2}}
\hfill
\subfigure[RRM\_Ours]{\includegraphics[height=0.42\linewidth]{./figures/rrm/RM5_1.png}
\label{fig:rf5}}
\vspace{-5pt}
\end{center}
\caption{Illustration of different Residual Refine Modules (RRM): (a) local boundary refinement module RRM\_LC; (b) multi-scale refinement module RRM\_MS; (c) our encoder-decoder refinement module RRM\_Ours.}
\label{fig:rf}
\end{figure}
The SSIM loss is a patch-level measure, which considers a local neighborhood of each pixel.
It assigns higher weights to pixels located in the transitional buffer regions between foregrounds and backgrounds, \emph{e.g.}~boundaries, fine structures, so that the loss is higher around the boundary, even when the predicted probabilities on the boundary and the rest of the foreground are the same.
It is worth noting that the loss for the background region is similar or sometimes even higher than the foreground region. However, the background loss does not contribute to the training until the prediction of background pixel becomes very close to the GT, where the loss drops rapidly from one to zero. Because $\mu_y$, $\sigma_{xy}$, $\mu_x\mu_y$ and $\sigma_y^2$ in the SSIM loss (Equ. \ref{equ:ssim_loss}) are all zeros in the background regions, so the SSIM loss can be approximated by:
\begin{equation}
\centering
\small{\ell^{bg}_{ssim}=1 - \frac{C_1C_2}{(\mu_x^2+C_1)(\sigma_x^2+C_2)}}.
\label{equ:ssim_loss_aprx}
\end{equation}
Since $C_1=0.01^2$ and $C_2=0.03^2$, only if the prediction $x$ is close to zero, the SSIM loss (Equ. \ref{equ:ssim_loss_aprx}) will become the dominant term.
The second and third rows of column (d) in Fig. \ref{fig:hybrid_loss} illustrate that the model trained with the SSIM loss is able to predict correct results on the foreground region and boundaries while neglecting the background accuracy in the beginning of the training process.
This characteristic of the SSIM loss helps the optimization to focus on the boundary and foreground region.
As the training progresses, the SSIM loss for the foreground is reduced and the background loss becomes the dominant term.
This is helpful since the prediction typically goes close to zero only late in the training process, where BCE loss becomes flat.
The SSIM loss ensures that there is still enough gradient to drive the learning process.
Hence, the background prediction looks cleaner since the probability is pushed to zero.
\begin{figure}[t!]
\centering
\begin{overpic}[width=\columnwidth]{./Imgs/Hybrid-loss-min.pdf}
\put(1,51){\small (a) Image}
\put(4,-4){\small (b) GT}
\put(29.5,-4){\small (c) $\ell_{bce}$}
\put(47,-4){\small (d) $\ell_{ssim}$}
\put(67,-4){\small (e) $\ell_{iou}$}
\put(86,-4){\small (f) $\ell_{bsi}$}
\end{overpic}
\vspace{5pt}
\caption{Intermediate predictions of our BASNet when fitting with different losses.
}\label{fig:hybrid_loss}
\end{figure}
The IoU is a map-level measure.
Larger ares contribute more to the IoU, so models trained with IoU loss emphasizes more on the large foreground regions and are thus able to produce relatively homogeneous and more confident (whiter) probabilities for these regions. However, these models often produce false negatives on fine structures. As shown in the column (e) of Fig.~\ref{fig:hybrid_loss}, the human head in the second row and the backpack cord in both the second and third rows are missing.
To take advantage of the above three losses, we combine them together to formulate the hybrid loss. BCE is used to maintain a smooth gradient for all pixels, while IoU is employed to put more focus on the foreground.
SSIM is used to encourage the prediction to respect the structure of the original image, by employing a larger loss near the boundaries, as well as further push the backgrounds predictions to zero.
\section{Experiments}
In this paper, we are focusing on improving the spatial accuracy of segmentation results. Therefore, experiments are conducted on two reverse
binary class image segmentation tasks: salient object segmentation \cite{wang2017learning} and camouflaged object segmentation \cite{fan2020camouflaged}.
Salient object segmentation is a popular task in computer vision, which aims at segmenting the salient regions against their backgrounds. In this task, the targets are usually with high contrast against their backgrounds.
However, camouflaged object segmentation is the most challenging one because the camouflaged objects usually have similar appearance to their backgrounds, which means they are difficult to be perceived and segmented. In addition, many of the camouflaged objects have very complex structures and boundaries.
\subsection{Implementation and Setting}
We implement our network using the publicly available Pytorch 1.4.0 \cite{paszke2017automatic}. An 16-core PC with an AMD Threadripper 2950x 3.5 GHz CPU (with 64GB 3000 MHz RAM, ) and an RTX Titan GPU (with 24GB memory) is used for both training and testing.
During training, each image is first resized to 320$\times$320 and randomly cropped to 288$\times$288. Some of the encoder parameters are initialized from the ResNet-34 model \cite{he2016deep}. Other convolutional layers are initialized by Xavier \cite{DBLP:journals/jmlr/GlorotB10}. We use the Adam optimizer \cite{kingma2014adam} to train our network and its hyperparameters are set to the default values, where the initial learning rate lr=1e-4, betas=(0.9, 0.999), eps=1e-8, weight\_decay=0. We train the network until the loss converges, without using the validation set. The training loss converges after 400k iterations with a batch size of eight and the whole training process takes about 110 hours.
During testing, the input image is resized to 320$\times$320 and fed into the network to obtain its segmentation probability map. Then, the probability map (320$\times$320) is resized back to the original size of the input image. Both resizing processes use bilinear interpolation. The inference for a 320$\times$320 image only takes 0.015s (\textbf{70 fps}, different from that reported in our CVPR version~\cite{qin2019basnet}, in which IO time is included).
\subsection{Evaluation Metrics}\label{sec:metrics}
Five measures are used to evaluate the performance of the proposed model.
\textbf{(1) Weighted F-measure} $F^w_\beta$ \cite{Margolin2014HowTE} gives a comprehensive and balanced evaluation on both precision and recall, which is able to better leverage the interpolation, dependency and equal-importance flaw.
\textbf{(2) Relax boundary F-measure} $F^b_\beta$ \cite{ehrig2005relaxed} is adopted to quantitatively evaluate the boundary quality of the predicted maps. \textbf{(3) Mean absolute error} $M$ \cite{perazzi2012saliency} reflects the average per-pixel difference between the probability map and the GT.
\textbf{(4) Mean structural measure} $S_\alpha$ \cite{DBLP:conf/iccv/FanCLLB17} quantizes the structural similarity between the predicted probability map and the GT mask.
\textbf{(5) Mean enhanced-alignment measure} $E^m_\phi$ \cite{fan2018enhanced} takes both global and local similarity into consideration. Evaluation code: \url{https://github.com/DengPingFan/CODToolbox}.
\begin{figure*}[t!]
\centering
\begin{overpic}[width=\textwidth]{./Imgs/ab_arch_loss-min.pdf}
\put(3,18){\small (a) Image}
\put(15,18){\small (b) GT}
\put(23.5,18){\small (c) U-Net$_{bce}$}
\put(36,18){\small (d) ED$_{bce}$}
\put(46,18){\small (e) EDS$_{bce}$}
\put(56,18){\small (f) ED\_LC$_{bce}$}
\put(67,18){\small (g) ED\_MS$_{bce}$}
\put(77.5,18){\small (h) BASNet$_{bce}$}
\put(89,18){\small (i) BASNet$_{bsi}$}
%
\put(3,0.5){\small (j) Image}
\put(15,0.5){\small (k) GT}
\put(26,0.5){\small (l) $\ell_{bce}$}
\put(35.5,0.5){\small (m) $\ell_{ssim}$}
\put(48,0.5){\small (n) $\ell_{iou}$}
\put(59.5,0.5){\small (o) $\ell_{bs}$}
\put(70,0.5){\small (p) $\ell_{bi}$}
\put(81,0.5){\small (q) $\ell_{si}$}
\put(92,0.5){\small (r) $\ell_{bsi}$}
\end{overpic}
\caption{Qualitative comparison of different configures in the ablation study. The first row show the predicted probability maps of different architectures trained with BCE loss and our BASNet trained with $\ell_{bsi}$ loss. The second row show the segmentation maps of our proposed prediction-refinement architecture trained with different losses. The corresponding quantitative results can be found in Table~\ref{tab:ablation}.
}\label{fig:ab_arch_loss}
\end{figure*}
\begin{figure*}[t!]
\centering
\begin{overpic}[width=\textwidth]{./Imgs/QualiSOD-min.pdf}
\put(3,0){\small (a) Image}
\put(17,0){\small (b) GT}
\put(27,0){\small (c) BASNet}
\put(39.5,0){\small (d) PoolNet}
\put(53,0){\small (e) CPD}
\put(65,0){\small (f) AFNet}
\put(76,0){\small (g) PiCANet}
\put(90,0){\small (h) NLDF}
\end{overpic}
\caption{Qualitative comparison on salient object segmentation datasets.
}\label{fig:qual_SOD}
\end{figure*}
\subsection{Experiments on Salient Object Segmentation}
\subsubsection{Datasets}
For salient object segmentation task\footnote{The camouflaged object segmentation task use the same augmentation strategies.}, we train our network using the DUTS-TR \cite{wang2017learning} dataset, which has 10553 images. Before training, the dataset is augmented by horizontal flipping to 21106 images. For salient object segmentation tasks, we evaluate our method on six commonly used salient object segmentation benchmark datasets: SOD \cite{movahedi2010design}, ECSSD \cite{yan2013hierarchical}, DUT-OMRON \cite{yang2013saliency}, PASCAL-S \cite{li2014secrets}, HKU-IS \cite{li2015visual}, DUTS-TE \cite{wang2017learning}.
\textbf{DUT-OMRON} has 5,168 images with one or multiple objects. The majority of these objects are structurally complex.
\textbf{PASCAL-S} was originally created for semantic image segmentation and consists of 850 challenging images.
\textbf{DUTS} is a relatively large salient object segmentation dataset. It has two subsets: DUTS-TR and DUTS-TE. There are 10,553 images in \textbf{DUTS-TR} for training and 5,019 images in \textbf{DUTS-TE} for testing. In our experiments, DUTS-TR is used for training the model for salient object segmentation.
\textbf{HKU-IS} contains 4,447 images, many of which contain multiple foreground objects.
\textbf{ECSSD} contains 1,000 semantically meaningful images. However, the structures of the foreground objects in these images are complex.
\textbf{SOD} contains 300 very challenging images. These images have either single complicated large foreground objects which overlap with the image boundaries or multiple salient objects with low contrast.
\subsubsection{Ablation Study}
In this section, we validate the effectiveness of each key components used in our model.
The ablation study is divided into two parts: an architecture ablation and loss ablation.
For simplify, the ablation experiments are conducted on the ECSSD dataset. The same hyper-parameters to that described in Sec. 4.1 are used here.
\begin{table}[tbp]
\centering
\footnotesize
\setlength{\tabcolsep}{3.0pt}
\renewcommand{\arraystretch}{0.5}
\caption{Ablation study on different architectures (Arch.) and losses: ED: encoder-decoder, EDS: encoder-decoder + side output supervision; $\ell_{b}$, $\ell_{s}$ and $\ell_{i}$ denote the BCE, SSIM and IoU loss, respectively, $\ell_{bi}=\ell_{b}+\ell_{i}$, $\ell_{bs}=\ell_{b}+\ell_{s}$,
$\ell_{si}=\ell_{s}+\ell_{i}$, $\ell_{bsi}=\ell_{b}+\ell_{s}+\ell_{i}$.}
\label{tab:ablation}
\begin{tabular}{c|c|ccccc}
\hline
Ablation & Configurations & $F^w_\beta\uparrow$ & $F^{b}_\beta\uparrow$ & $~M~\downarrow$ & $S_{\alpha}\uparrow$ & $E_{\phi}^{m}\uparrow$ \\
\hline
\multirow{6}{*}{\rotatebox[origin=c]{90}{\parbox[c]{2 cm}{\centering {\footnotesize Arch. }}}}
& \multicolumn{1}{r|}{U-Net \cite{ronneberger2015u} + $\ell_{b}$} & 0.827 & 0.669 & 0.064 & 0.867 & 0.897\\
& \multicolumn{1}{r|}{ED + $\ell_{b}$} & 0.871 & 0.786 & 0.045 & 0.908 & 0.923\\
& \multicolumn{1}{r|}{EDS + $\ell_{b}$} & 0.891 & 0.819 & 0.041 & 0.920 & 0.935\\
& \multicolumn{1}{r|}{EDS+RRM\_LC + $\ell_{b}$} & 0.900 & 0.804 & 0.038 & 0.915 & 0.935\\
& \multicolumn{1}{r|}{EDS+RRM\_MS + $\ell_{b}$} & 0.890 & 0.816 & 0.041 & 0.919 & 0.934\\
& \multicolumn{1}{r|}{EDS+RRM\_Ours + $\ell_{b}$} & 0.900 & 0.827 & 0.037 & 0.923 & 0.943\\
\hline
\multirow{6}{*}{\rotatebox[origin=c]{90}{\parbox[c]{2 cm}{\centering {\footnotesize Loss}}}}
& \multicolumn{1}{l|}{EDS+RRM\_Ours + $\ell_{s}$} & 0.886 & 0.814 & 0.044 & 0.904 & 0.932\\
& \multicolumn{1}{l|}{EDS+RRM\_Ours + $\ell_{i}$} & 0.902 & 0.820 & 0.037 & 0.911 & 0.943\\
& \multicolumn{1}{l|}{EDS+RRM\_Ours + $\ell_{bs}$} & 0.903 & 0.823 & 0.037 & 0.920 & 0.942\\
& \multicolumn{1}{l|}{EDS+RRM\_Ours + $\ell_{bi}$} & 0.909 & 0.832 & 0.035 & 0.921 & \bf{0.947}\\
& \multicolumn{1}{l|}{EDS+RRM\_Ours + $\ell_{si}$} & 0.894 & 0.812 & 0.041 & 0.906 & 0.938\\
& \multicolumn{1}{l|}{EDS+RRM\_Ours + $\ell_{bsi}$} & \bf{0.912} & \bf{0.840} & \bf{0.034} & \bf{0.925} & \bf{0.947} \\
\hline
\end{tabular}
\end{table}
\begin{table*}[t!]
\caption{Comparison of the proposed method and 25 other methods on three salient object segmentation datasets: DUT-OMRON, DUTS-TE and HKU-IS. $\uparrow$ and $\downarrow$ indicate the higher the score the better and the lower the score the better, respectively. ``*'' indicates results post-processed by CRF. \textbf{Bold} font denotes the best performance.}
\centering
\footnotesize
\renewcommand{\arraystretch}{0.5}
\setlength\tabcolsep{5.5pt}
\begin{tabular}{l|ccccc|ccccc|ccccc}
\hline
\multirow{2}{*}{\textbf{Models}} & \multicolumn{5}{|c|}{\textbf{DUT-OMRON}\cite{yang2013saliency}} & \multicolumn{5}{|c|}{\textbf{DUTS-TE}\cite{wang2017learning}} & \multicolumn{5}{|c}{\textbf{HKU-IS}\cite{li2015visual}}\\
\cline{2-16}
& $F^w_\beta\uparrow$ & $F^{b}_\beta\uparrow$ & $~M~\downarrow$ & $S_{\alpha}\uparrow$ & $E_{\phi}^{m}\uparrow$ &
$F^w_\beta\uparrow$ & $F^{b}_\beta\uparrow$ & $M\downarrow$ & $S_{\alpha}\uparrow$ & $E_{\phi}^{m}\uparrow$ &
$F^w_\beta\uparrow$ & $F^{b}_\beta\uparrow$ & $M\downarrow$ & $S_{\alpha}\uparrow$ & $E_{\phi}^{m}\uparrow$\\
\hline
\textbf{MDF}$_\text{TIP16}$ & 0.565 & 0.406 & 0.142 & 0.721 & 0.759 & 0.543 & 0.447 & 0.099 & 0.723 & 0.764 & 0.564 & 0.594 & 0.129 & 0.81 & 0.742 \\
\textbf{UCF}$_\text{ICCV17}$ & 0.573 & 0.48 & 0.12 & 0.76 & 0.761 & 0.596 & 0.518 & 0.112 & 0.777 & 0.776 & 0.779 & 0.679 & 0.062 & 0.875 & 0.887 \\
\textbf{Amulet}$_\text{ICCV17}$ & 0.626 & 0.528 & 0.098 & 0.781 & 0.794 & 0.658 & 0.568 & 0.084 & 0.796 & 0.817 & 0.817 & 0.716 & 0.051 & 0.886 & 0.910 \\
\textbf{NLDF$^*$}$_\text{CVPR17}$ & 0.634 & 0.514 & 0.08 & 0.77 & 0.799 & 0.71 & 0.591 & 0.065 & 0.805 & 0.851 & 0.838 & 0.694 & 0.048 & 0.879 & 0.914 \\
\textbf{DSS$^*$}$_\text{CVPR17}$ & 0.697 & 0.559 & 0.063 & 0.79 & 0.831 & 0.755 & 0.606 & 0.056 & 0.812 & 0.877 & 0.867 & 0.706 & 0.04 & 0.878 & 0.925 \\
\textbf{LFR}$_\text{IJCAI18}$ & 0.647 & 0.508 & 0.103 & 0.78 & 0.799 & 0.689 & 0.556 & 0.083 & 0.799 & 0.833 & 0.861 & 0.731 & 0.04 & 0.905 & 0.934 \\
\textbf{C2S}$_\text{ECCV18}$ & 0.661 & 0.565 & 0.072 & 0.798 & 0.823 & 0.713 & 0.607 & 0.062 & 0.829 & 0.859 & 0.829 & 0.717 & 0.048 & 0.883 & 0.859 \\
\textbf{RAS}$_\text{ECCV18}$ & 0.695 & 0.615 & 0.062 & 0.814 & 0.844 & 0.74 & 0.656 & 0.059 & 0.828 & 0.871 & 0.843 & 0.748 & 0.045 & 0.887 & 0.92 \\
\textbf{RADF$^*$}$_\text{AAAI18}$ & 0.723 & 0.579 & 0.061 & 0.815 & 0.857 & 0.748 & 0.608 & 0.061 & 0.814 & 0.869 & 0.872 & 0.725 & 0.039 & 0.888 & 0.935 \\
\textbf{PAGRN}$_\text{CVPR18}$ & 0.622 & 0.582 & 0.071 & 0.775 & 0.772 & 0.724 & 0.692 & 0.055 & 0.825 & 0.843 & 0.82 & 0.762 & 0.048 & 0.887 & 0.900 \\
\textbf{BMPM}$_\text{CVPR18}$ & 0.681 & 0.612 & 0.064 & 0.809 & 0.831 & 0.761 & 0.699 & 0.048 & 0.851 & 0.883 & 0.859 & 0.773 & 0.039 & 0.907 & 0.931 \\
\textbf{PiCANet}$_\text{CVPR18}$ & 0.691 & 0.643 & 0.068 & 0.826 & 0.833 & 0.747 & 0.704 & 0.054 & 0.851 & 0.873 & 0.847 & 0.784 & 0.042 & 0.906 & 0.923 \\
\textbf{MLMS}$_\text{CVPR19}$ & 0.681 & 0.612 & 0.064 & 0.809 & 0.831 & 0.761 & 0.699 & 0.048 & 0.851 & 0.883 & 0.859 & 0.773 & 0.039 & 0.907 & 0.931 \\
\textbf{AFNet}$_\text{CVPR19}$ & 0.717 & 0.635 & 0.057 & 0.826 & 0.846 & 0.785 & 0.714 & 0.046 & 0.855 & 0.893 & 0.869 & 0.772 & 0.036 & 0.905 & 0.935 \\
\textbf{MSWS}$_\text{CVPR19}$ & 0.527 & 0.362 & 0.109 & 0.756 & 0.729 & 0.586 & 0.376 & 0.908 & 0.749 & 0.742 & 0.685 & 0.438 & 0.084 & 0.818 & 0.787 \\
\textbf{R}$^3$\textbf{Net$^*$}$_\text{IJCAI18}$ & 0.728 & 0.599 & 0.063 & 0.817 & 0.853 & 0.763 & 0.601 & 0.058 & 0.817 & 0.873 & 0.877 & 0.74 & 0.036 & 0.895 & 0.939 \\
\textbf{CapSal}$_\text{CVPR19}$ & 0.482 & 0.396 & 0.101 & 0.674 & 0.659 & 0.691 & 0.605 & 0.072 & 0.808 & 0.849 & 0.782 & 0.654 & 0.062 & 0.85 & 0.883 \\
\textbf{SRM}$_\text{ICCV17}$ & 0.658 & 0.523 & 0.069 & 0.798 & 0.808 & 0.722 & 0.592 & 0.058 & 0.824 & 0.853 & 0.835 & 0.68 & 0.046 & 0.887 & 0.913 \\
\textbf{DGRL}$_\text{CVPR18}$ & 0.697 & 0.584 & 0.063 & 0.810 & 0.845 & 0.76 & 0.656 & 0.051 & 0.836 & 0.887 & 0.865 & 0.744 & 0.037 & 0.897 & 0.939 \\
\textbf{CPD}$_\text{CVPR19}$ & 0.719 & 0.655 & \textbf{0.056} & 0.825 & 0.847 & 0.795 & 0.741 & 0.043 & 0.858 & 0.898 & 0.875 & 0.795 & 0.034 & 0.905 & 0.939 \\
\textbf{PoolNet}$_\text{CVPR19}$ & 0.729 & 0.675 & \textbf{0.056} & 0.836 & 0.854 & 0.807 & 0.765 & \textbf{0.040} & 0.871 & 0.904 & 0.881 & 0.811 & 0.033 & 0.917 & 0.940 \\
\textbf{BANet}$_\text{ICCV19}$ & 0.719 & 0.611 & 0.061 & 0.823 & 0.861 & 0.781 & 0.687 & 0.046 & 0.861 & 0.897 & 0.869 & 0.760 & 0.037 & 0.902 & 0.938 \\
\textbf{EGNet}$_\text{ICCV19}$ & 0.728 & 0.679 & \textbf{0.056} & 0.836 & 0.853 & 0.797 & 0.761 & 0.043 & 0.879 & 0.898 & 0.875 & 0.802 & 0.035 & 0.910 & 0.938 \\
\textbf{MINet}$_\text{CVPR20}$ & 0.719 & 0.640 & 0.057 & 0.822 & 0.846 & 0.813 & 0.747 & \textbf{0.040} & 0.875 & 0.906 & 0.889 & 0.799 & 0.032 & 0.912 & 0.944 \\
\textbf{GateNet}$_\text{ECCV20}$ & 0.703 & 0.625 & 0.061 & 0.821 & 0.840 & 0.786 & 0.722 & 0.045 & 0.871 & 0.892 & 0.872 & 0.783 & 0.036 & 0.910 & 0.934 \\
\hline
\textbf{BASNet} (Ours) & \textbf{0.760} & \textbf{0.703} & 0.057 & \textbf{0.841} & \textbf{0.868} & \textbf{0.825} & \textbf{0.786} & 0.042 & \textbf{0.881} & \textbf{0.907} & \textbf{0.900} & \textbf{0.821} & \textbf{0.030} & \textbf{0.918} & \textbf{0.948} \\
\hline
\end{tabular}
\label{tab:compSOA}
\end{table*}
\begin{table*}[t!]
\caption{Comparison of the proposed method and 25 other methods on three salient object detection datasets: ECSSD, PASCAL-S and SOD. See \tabref{tab:compSOA} for details.
}
\centering
\footnotesize
\renewcommand{\arraystretch}{0.5}
\setlength\tabcolsep{5.2pt}
\begin{tabular}{l|ccccc|ccccc|ccccc}
\hline
\multirow{2}{*}{\textbf{Baseline Models}} & \multicolumn{5}{|c|}{\textbf{ECSSD}\cite{yan2013hierarchical}} & \multicolumn{5}{|c|}{\textbf{PASCAL-S}\cite{li2014secrets}} & \multicolumn{5}{|c}{\textbf{SOD}\cite{movahedi2010design}}\\
\cline{2-16}
& $F^w_\beta\uparrow$ & $F^{b}_\beta\uparrow$ & $~M~\downarrow$ & $S_{\alpha}\uparrow$ & $E_{\phi}^{m}\uparrow$ &
$F^w_\beta\uparrow$ & $F^{b}_\beta\uparrow$ & $M\downarrow$ & $S_{\alpha}\uparrow$ & $E_{\phi}^{m}\uparrow$ &
$F^w_\beta\uparrow$ & $F^{b}_\beta\uparrow$ & $M\downarrow$ & $S_{\alpha}\uparrow$ & $E_{\phi}^{m}\uparrow$\\
\hline
\textbf{MDF}$_\text{TIP16}$ & 0.705 & 0.472 & 0.105 & 0.776 & 0.796 & 0.589 & 0.343 & 0.142 & 0.696 & 0.706 & 0.508 & 0.311 & 0.192 & 0.643 & 0.607 \\
\textbf{UCF}$_\text{ICCV17}$ & 0.806 & 0.669 & 0.069 & 0.884 & 0.891 & 0.694 & 0.493 & 0.115 & 0.805 & 0.809 & 0.675 & 0.471 & 0.148 & 0.762 & 0.773 \\
\textbf{Amulet}$_\text{ICCV17}$ & 0.84 & 0.711 & 0.059 & 0.894 & 0.909 & 0.734 & 0.541 & 0.100 & 0.818 & 0.835 & 0.677 & 0.454 & 0.144 & 0.753 & 0.776 \\
\textbf{NLDF$^*$}$_\text{CVPR17}$ & 0.839 & 0.666 & 0.063 & 0.897 & 0.900 & 0.737 & 0.495 & 0.098 & 0.798 & 0.839 & 0.709 & 0.475 & 0.125 & 0.755 & 0.777 \\
\textbf{DSS$^*$}$_\text{CVPR17}$ & 0.872 & 0.696 & 0.052 & 0.882 & 0.918 & 0.759 & 0.499 & 0.093 & 0.798 & 0.845 & 0.710 & 0.444 & 0.124 & 0.743 & 0.774 \\
\textbf{LFR}$_\text{IJCAI18}$ & 0.858 & 0.694 & 0.052 & 0.897 & 0.923 & 0.737 & 0.499 & 0.107 & 0.805 & 0.835 & 0.734 & 0.479 & 0.123 & 0.773 & 0.813 \\
\textbf{C2S}$_\text{ECCV18}$ & 0.851 & 0.708 & 0.055 & 0.893 & 0.917 & 0.766 & 0.543 & 0.082 & 0.836 & 0.864 & 0.700 & 0.457 & 0.124 & 0.760 & 0.785 \\
\textbf{RAS}$_\text{ECCV18}$ & 0.857 & 0.741 & 0.056 & 0.893 & 0.914 & 0.736 & 0.560 & 0.101 & 0.799 & 0.830 & 0.720 & 0.544 & 0.124 & 0.764 & 0.788 \\
\textbf{RADF$^*$}$_\text{AAAI18}$ & 0.883 & 0.720 & 0.049 & 0.894 & 0.929 & 0.755 & 0.515 & 0.097 & 0.802 & 0.840 & 0.729 & 0.476 & 0.126 & 0.757 & 0.801 \\
\textbf{PAGRN}$_\text{CVPR18}$ & 0.834 & 0.747 & 0.061 & 0.889 & 0.895 & 0.738 & 0.594 & 0.090 & 0.822 & 0.830 & - & - & - & - & - \\
\textbf{BMPM}$_\text{CVPR18}$ & 0.871 & 0.770 & 0.045 & 0.911 & 0.928 & 0.779 & 0.617 & 0.074 & 0.845 & 0.872 & 0.726 & 0.562 & 0.108 & 0.786 & 0.799 \\
\textbf{PiCANet}$_\text{CVPR18}$ & 0.865 & 0.784 & 0.046 & 0.914 & 0.924 & 0.772 & 0.612 & 0.078 & 0.848 & 0.866 & 0.722 & 0.572 & 0.103 & 0.789 & 0.796 \\
\textbf{MLMS}$_\text{CVPR19}$ & 0.871 & 0.770 & 0.045 & 0.911 & 0.928 & 0.779 & 0.62 & 0.074 & 0.844 & 0.875 & 0.726 & 0.562 & 0.108 & 0.786 & 0.799 \\
\textbf{AFNet}$_\text{CVPR19}$ & 0.887 & 0.776 & 0.042 & 0.914 & 0.936 & 0.798 & 0.626 & 0.070 & 0.849 & 0.883 & 0.723 & 0.545 & 0.111 & 0.774 & 0.79 \\
\textbf{MSWS}$_\text{CVPR19}$ & 0.716 & 0.411 & 0.096 & 0.828 & 0.791 & 0.614 & 0.289 & 0.133 & 0.768 & 0.731 & 0.573 & 0.231 & 0.167 & 0.700 & 0.656 \\
\textbf{R}$^3$\textbf{Net$^*$}$_\text{IJCAI18}$ & 0.902 & 0.759 & 0.040 & 0.910 & 0.944 & 0.761 & 0.538 & 0.092 & 0.807 & 0.843 & 0.735 & 0.431 & 0.125 & 0.759 & 0.796 \\
\textbf{CapSal}$_\text{CVPR19}$ & 0.771 & 0.574 & 0.077 & 0.826 & 0.849 & 0.786 & 0.527 & 0.073 & 0.837 & 0.872 & 0.597 & 0.404 & 0.148 & 0.695 & 0.699 \\
\textbf{SRM}$_\text{ICCV17}$ & 0.853 & 0.672 & 0.054 & 0.895 & 0.913 & 0.758 & 0.509 & 0.084 & 0.834 & 0.853 & 0.670 & 0.392 & 0.128 & 0.741 & 0.744 \\
\textbf{DGRL}$_\text{CVPR18}$ & 0.883 & 0.753 & 0.042 & 0.906 & 0.938 & 0.787 & 0.569 & 0.074 & 0.839 & 0.877 & 0.731 & 0.502 & 0.106 & 0.773 & 0.807 \\
\textbf{CPD}$_\text{CVPR19}$ & 0.898 & 0.811 & 0.037 & 0.918 & 0.942 & 0.800 & 0.639 & 0.071 & 0.848 & 0.878 & 0.714 & 0.556 & 0.112 & 0.767 & 0.778 \\
\textbf{PoolNet}$_\text{CVPR19}$ & 0.896 & 0.813 & 0.039 & 0.921 & 0.940 & 0.798 & 0.644 & 0.075 & 0.832 & 0.876 & 0.759 & 0.606 & 0.102 & \textbf{0.797} & 0.818 \\
\textbf{BANet}$_\text{ICCV19}$ & 0.890 & 0.758 & 0.041 & 0.913 & 0.940 & 0.792 & 0.589 & 0.078 & 0.840 & 0.875 & 0.750 & 0.589 & 0.109 & 0.782 & 0.813 \\
\textbf{EGNet}$_\text{ICCV19}$ & 0.892 & 0.814 & 0.041 & 0.920 & 0.936 & 0.793 & 0.650 & 0.077 & 0.848 & 0.873 & 0.737 & 0.586 & 0.112 & 0.784 & 0.798 \\
\textbf{MINet}$_\text{CVPR20}$ & 0.905 & 0.805 & 0.037 & 0.920 & 0.943 & \textbf{0.813} & 0.648 & \textbf{0.065} & 0.854 & \textbf{0.889} & - & - & - & - & - \\
\textbf{GateNet}$_\text{ECCV20}$ & 0.886 & 0.782 & 0.042 & 0.917 & 0.933 & 0.803 & 0.623 & 0.068 & \textbf{0.857} & 0.882 & - & - & - & - & - \\
\hline
\textbf{BASNet} (Ours) & \textbf{0.912} & \textbf{0.840} & \textbf{0.034} & \textbf{0.925} & \textbf{0.947} & 0.808 & \textbf{0.674} & 0.072 & 0.847 & 0.878 & \textbf{0.762} & \textbf{0.640} & \textbf{0.102} & 0.793 & \textbf{0.822} \\
\hline
\end{tabular
\label{tab:compSOB}
\end{table*}
\textbf{Architecture:} To demonstrate the effectiveness of our BASNet, we report quantitative comparison results of our model against other related architectures.
We take U-Net \cite{ronneberger2015u} as our baseline network.
Then we start with our proposed encoder-decoder network and progressively extend it with dense side output supervision and different residual refinement modules, including RRM\_LC, RRM\_MS and RRM\_Ours.
The top part of Table~\ref{tab:ablation} and the first row of Fig.~\ref{fig:ab_arch_loss} illustrate the qualitative and quantitative results of the architecture ablation study, respectively.
As we can see, our BASNet architecture achieves the best performance among all configurations.
\begin{figure*}[t!]
\centering
\begin{overpic}[width=\textwidth]{./Imgs/QualiSOC-min.pdf}
\put(3,1){\small (a) Image}
\put(17,1){\small (b) GT}
\put(27,1){\small (c) BASNet}
\put(39.5,1){\small (d) SCRN}
\put(53,1){\small (e) BANet}
\put(65,1){\small (f) EGNet}
\put(76,1){\small (g) PoolNet}
\put(90,1){\small (h) CPD}
\end{overpic}
\caption{Qualitative comparison on typical samples from the SOC dataset. Images from top to bottom are from attributes SO (Small Object), OV(Out-of-View), OC (Occlusion) and SC (Shape Complexity) respectively.
}\label{fig:qual-soc}
\end{figure*}
\begin{figure}[t!]
\centering
\begin{overpic}[width=.9\columnwidth]{./Imgs/FailSOD-min.pdf}
\put(9,-3){\small (a) Image}
\put(45,-3){\small (b) GT}
\put(75,-3){\small (c) BASNet}
\end{overpic}
\vspace{2pt}
\caption{Failure cases on salient object segmentation datasets.
}\label{fig:fail_sod}
\end{figure}
\textbf{Loss:} To demonstrate the effectiveness of our proposed fusion loss, we conduct a set of experiments over different losses based on our BASNet architecture.
The results in Table \ref{tab:ablation} indicate that the proposed hybrid $\ell_{bsi}$ loss greatly improves the performance, especially in terms of the boundary quality. It is clear that our hybrid loss achieves superior qualitative results, as shown in the second row of Fig.~\ref{fig:ab_arch_loss}.
\subsubsection{Comparison with State-of-the-Arts}
We compare our method with 25 state-of-the-art models, including MDF \cite{li2016visual},
UCF \cite{DBLP:conf/iccv/ZhangWLWY17},
Amulet \cite{DBLP:conf/iccv/ZhangWLWR17},
NLDF \cite{luo2017non},
DSS \cite{hou2017deeply},
LFR \cite{DBLP:conf/ijcai/ZhangLLS18},
C2S \cite{DBLP:conf/eccv/LiYCLS18},
RAS \cite{DBLP:conf/eccv/ChenTWH18},
RADF \cite{DBLP:conf/aaai/HuZQFH18},
PAGRN \cite{zhang2018progressive},
BMPM \cite{zhang2018bi},
PiCANet \cite{liu2018picanet},
MLMS \cite{MLMS},
AFNet \cite{AFNet},
MSWS \cite{MSWS},
R$^3$-Net \cite{deng2018r3net},
CapSal \cite{CapSal},
SRM \cite{DBLP:conf/iccv/WangBZZL17},
DGRL \cite{wang2018detect},
CPD \cite{wu2019cascaded},
PoolNet \cite{liu2019simple},
BANet \cite{su2019selectivity},
EGNet \cite{zhao2019egnet},
MINet \cite{pang2020multi} and
GateNet \cite{GateNet}, on the salient object segmentation task.
For fair comparison, we either use the segmentation maps released by the authors or run their publicly available models with their default settings.
\textbf{Quantitative Evaluation:}
Tables \ref{tab:compSOA} and \ref{tab:compSOB} provide quantitative comparisons on six salient object segmentation datasets. Our BASNet outperforms other models on the DUT-OMRON, DUTS-TE, HKU-IS, ECSSD and SOD datasets in terms of nearly all metrics, except for the $M$ measures on DUT-OMRON and DUTS-TE and the $S_\alpha$ on SOD. On the PASCAL-S dataset, MINet performs the best in terms of three metrics: $F^w_\beta$, $MAE$ and $E^m_\phi$. It is worth noting that BASNet achieves the highest relax boundary F-measure $F^b_\beta$ on all of the six datasets, which indicates its strong capability in capturing boundaries and fine structures.
\textbf{Qualitative Evaluation:}
Fig.~\ref{fig:qual_SOD} shows the qualitative comparison between our BASNet and 5 other typical models. As we can see, our BASNet is able to handle different challenging cases, such as small target with relatively low contrast (1st row), large object with complicated boundaries (2nd row), object with hollow structures (3rd row) and target with very fine structures (4th row). The third and fourth row of Fig.~\ref{fig:qual_SOD} show inspiring results, in which the segmentation maps predicted by our BASNet contain more details than the GT. These details reveals the possible inconsistency between the labels of training and testing datasets. Although detecting of these details usually leads to the deterioration of the quantitative evaluation scores of our model, it is more practically useful than good scores.
\subsubsection{Failure Cases}
Fig. \ref{fig:fail_sod} shows three typical failure cases of our BASNet on SOD datasets. For instance, the model sometimes fails in very complicated scenarios, in which there seems no salient objects, as show in the first row of Fig. \ref{fig:fail_sod}.
The second row gives an exemplary failure case of ``saliency confusing'', where the scene contains multiple independent ``salient'' targets, But only one of them is labeled. Our BASNet sometimes fails in these cases due to lack of the ability of recognizing the tiny saliency differences between multiple connected targets. The recent uncertainty model~\cite{zhang2020uncertainty} may be one of the solutions.
\begin{table*}[tbp]
\caption{Comparison of the proposed method and other SOTA methods on the SOC test set. $\uparrow$ and $\downarrow$ indicate the higher score the better and the lower score the better respectively. \textbf{Bold} font indicates the best performance. \textbf{Avg.} denotes the average of all the attribute-based metric scores.}
\centering
\centering
\footnotesize
\renewcommand{\arraystretch}{0.5}
\setlength\tabcolsep{0.2pt}
\begin{tabular}{l|c|cccccccccccccccc|cc}
\hline
\textbf{Attr} & \textbf{Metr.} & \tabincell{c}{\textbf{Amulet}\\\cite{DBLP:conf/iccv/ZhangWLWR17}} & \tabincell{c}{\textbf{DSS}\\\cite{hou2017deeply}} & \tabincell{c}{\textbf{NLDF}\\\cite{luo2017non}} & \tabincell{c}{\textbf{C2SNet}\\\cite{DBLP:conf/eccv/LiYCLS18}} & \tabincell{c}{\textbf{SRM}\\\cite{DBLP:conf/iccv/WangBZZL17}} & \tabincell{c}{\textbf{R3Net}\\\cite{deng2018r3net}} & \tabincell{c}{\textbf{BMPM}\\\cite{zhang2018bi}} & \tabincell{c}{\textbf{DGRL}\\\cite{wang2018detect}} & \tabincell{c}{\textbf{PiC(R)}\\\cite{liu2018picanet}} & \tabincell{c}{\textbf{RANet}\\\cite{chen2020reverse}} & \tabincell{c}{\textbf{AFNet}\\\cite{AFNet}} & \tabincell{c}{\textbf{CPD}\\\cite{wu2019cascaded}} & \tabincell{c}{\textbf{PoolNet}\\\cite{liu2019simple}} & \tabincell{c}{\textbf{EGNet}\\\cite{zhao2019egnet}} & \tabincell{c}{\textbf{BANet}\\\cite{su2019selectivity}} & \tabincell{c}{\textbf{SCRN}\\\cite{Wu_2019_ICCV}} &
\tabincell{c}{\textbf{Ours}\\(DUTS)} &
\tabincell{c}{\textbf{Ours}\\(SOC)} \\
\hline
\multirow{5}{*}{\textbf{AC}} & $F^w_\beta\uparrow$ & 0.620 & 0.629 & 0.620 & 0.647 & 0.690 & 0.593 & 0.680 & 0.718 & 0.682 & 0.603 & 0.712 & 0.727 & 0.713 & 0.731 & 0.740 & 0.724 & 0.735 & \textbf{0.792} \\
& $F^{b}_\beta\uparrow$ & 0.448 & 0.384 & 0.374 & 0.408 & 0.410 & 0.387 & 0.531 & 0.457 & 0.489 & 0.448 & 0.569 & 0.626 & 0.578 & 0.597 & 0.562 & 0.588 & 0.659 & \textbf{0.696} \\
& $~M~\downarrow$ & 0.120 & 0.113 & 0.119 & 0.109 & 0.096 & 0.135 & 0.098 & 0.081 & 0.093 & 0.132 & 0.084 & 0.083 & 0.094 & 0.085 & 0.086 & 0.078 & 0.087 & \textbf{0.060} \\
& $S_{\alpha}\uparrow$ & 0.752 & 0.753 & 0.737 & 0.755 & 0.791 & 0.713 & 0.780 & 0.790 & 0.792 & 0.708 & 0.796 & 0.799 & 0.795 & 0.806 & 0.806 & 0.809 & 0.805 & \textbf{0.831} \\
& $E_{\phi}^{m}\uparrow$ & 0.791 & 0.788 & 0.784 & 0.807 & 0.824 & 0.753 & 0.815 & 0.853 & 0.815 & 0.765 & 0.852 & 0.843 & 0.846 & 0.854 & 0.858 & 0.849 & 0.844 & \textbf{0.885} \\
\hline
\multirow{5}{*}{\textbf{BO}} & $F^w_\beta\uparrow$ & 0.612 & 0.614 & 0.622 & 0.730 & 0.667 & 0.456 & 0.670 & 0.786 & 0.799 & 0.453 & 0.741 & 0.739 & 0.610 & 0.585 & 0.720 & 0.778 & 0.747 & \textbf{0.808} \\
& $F^{b}_\beta\uparrow$ & 0.274 & 0.213 & 0.218 & 0.362 & 0.274 & 0.229 & 0.400 & 0.392 & 0.466 & 0.231 & 0.450 & 0.481 & 0.323 & 0.319 & 0.360 & 0.453 & 0.519 & \textbf{0.572} \\
& $~M~\downarrow$ & 0.346 & 0.356 & 0.354 & 0.267 & 0.306 & 0.445 & 0.303 & 0.215 & 0.200 & 0.454 & 0.245 & 0.257 & 0.353 & 0.373 & 0.271 & 0.224 & 0.253 & \textbf{0.166} \\
& $S_{\alpha}\uparrow$ & 0.574 & 0.561 & 0.568 & 0.654 & 0.614 & 0.437 & 0.604 & 0.684 & 0.729 & 0.421 & 0.658 & 0.647 & 0.561 & 0.528 & 0.645 & 0.698 & 0.666 & \textbf{0.723} \\
& $E_{\phi}^{m}\uparrow$ & 0.551 & 0.537 & 0.539 & 0.661 & 0.616 & 0.419 & 0.620 & 0.725 & 0.741 & 0.404 & 0.698 & 0.665 & 0.554 & 0.528 & 0.650 & 0.706 & 0.677 & \textbf{0.775} \\
\hline
\multirow{5}{*}{\textbf{CL}} & $F^w_\beta\uparrow$ & 0.663 & 0.617 & 0.614 & 0.655 & 0.665 & 0.546 & 0.678 & 0.714 & 0.692 & 0.542 & 0.696 & 0.724 & 0.681 & 0.677 & 0.726 & 0.717 & 0.700 & \textbf{0.730} \\
& $F^{b}_\beta\uparrow$ & 0.374 & 0.275 & 0.292 & 0.342 & 0.327 & 0.315 & 0.432 & 0.393 & 0.420 & 0.344 & 0.465 & 0.553 & 0.488 & 0.493 & 0.461 & 0.506 & 0.552 & \textbf{0.579} \\
& $~M~\downarrow$ & 0.141 & 0.153 & 0.159 & 0.144 & 0.134 & 0.182 & 0.123 & 0.119 & 0.123 & 0.188 & 0.119 & 0.114 & 0.134 & 0.139 & 0.117 & 0.113 & 0.121 & \textbf{0.110} \\
& $S_{\alpha}\uparrow$ & 0.763 & 0.722 & 0.713 & 0.742 & 0.759 & 0.659 & 0.761 & 0.770 & 0.787 & 0.624 & 0.768 & 0.773 & 0.760 & 0.757 & 0.784 & \textbf{0.795} & 0.774 & 0.785 \\
& $E_{\phi}^{m}\uparrow$ & 0.789 & 0.763 & 0.764 & 0.789 & 0.793 & 0.710 & 0.801 & 0.824 & 0.794 & 0.715 & 0.802 & 0.821 & 0.801 & 0.790 & 0.824 & 0.820 & 0.807 & \textbf{0.826} \\
\hline
\multirow{5}{*}{\textbf{HO}} & $F^w_\beta\uparrow$ & 0.688 & 0.660 & 0.661 & 0.668 & 0.696 & 0.633 & 0.684 & 0.722 & 0.704 & 0.626 & 0.722 & 0.751 & 0.739 & 0.720 & 0.754 & 0.743 & 0.746 & \textbf{0.764} \\
& $F^{b}_\beta\uparrow$ & 0.465 & 0.347 & 0.378 & 0.398 & 0.392 & 0.383 & 0.496 & 0.447 & 0.462 & 0.425 & 0.527 & 0.618 & 0.570 & 0.560 & 0.542 & 0.577 & 0.627 & \textbf{0.639} \\
& $~M~\downarrow$ & 0.119 & 0.124 & 0.126 & 0.123 & 0.115 & 0.136 & 0.116 & 0.104 & 0.108 & 0.143 & 0.103 & 0.097 & 0.100 & 0.106 & 0.094 & 0.096 & 0.099 & \textbf{0.093} \\
& $S_{\alpha}\uparrow$ & 0.791 & 0.767 & 0.755 & 0.768 & 0.794 & 0.740 & 0.781 & 0.791 & 0.809 & 0.713 & 0.798 & 0.803 & 0.815 & 0.802 & 0.819 & \textbf{0.823} & 0.809 & 0.814 \\
& $E_{\phi}^{m}\uparrow$ & 0.810 & 0.796 & 0.798 & 0.805 & 0.819 & 0.782 & 0.813 & 0.833 & 0.819 & 0.777 & 0.834 & 0.845 & 0.846 & 0.829 & \textbf{0.850} & 0.842 & 0.843 & \textbf{0.850} \\
\hline \multirow{5}{*}{\textbf{MB}} & $F^w_\beta\uparrow$ & 0.561 & 0.577 & 0.551 & 0.593 & 0.619 & 0.489 & 0.651 & 0.655 & 0.637 & 0.576 & 0.626 & 0.679 & 0.642 & 0.649 & 0.672 & 0.690 & 0.678 & \textbf{0.725} \\
& $F^{b}_\beta\uparrow$ & 0.435 & 0.396 & 0.397 & 0.450 & 0.395 & 0.348 & 0.561 & 0.464 & 0.520 & 0.476 & 0.537 & 0.619 & 0.592 & 0.584 & 0.539 & 0.595 & 0.635 & \textbf{0.674} \\
& $~M~\downarrow$ & 0.142 & 0.132 & 0.138 & 0.128 & 0.115 & 0.160 & 0.105 & 0.113 & 0.099 & 0.139 & 0.111 & 0.106 & 0.121 & 0.109 & 0.104 & 0.100 & 0.115 & \textbf{0.072} \\
& $S_{\alpha}\uparrow$ & 0.712 & 0.719 & 0.685 & 0.720 & 0.742 & 0.657 & 0.762 & 0.744 & 0.775 & 0.696 & 0.734 & 0.754 & 0.751 & 0.762 & 0.764 & 0.792 & 0.755 & \textbf{0.797} \\
& $E_{\phi}^{m}\uparrow$ & 0.739 & 0.753 & 0.740 & 0.778 & 0.778 & 0.697 & 0.812 & 0.823 & 0.813 & 0.761 & 0.762 & 0.804 & 0.779 & 0.789 & 0.803 & 0.817 & 0.805 & \textbf{0.836} \\
\hline \multirow{5}{*}{\textbf{OC}} & $F^w_\beta\uparrow$ & 0.607 & 0.595 & 0.593 & 0.622 & 0.630 & 0.520 & 0.644 & 0.659 & 0.638 & 0.527 & 0.680 & 0.672 & 0.659 & 0.658 & 0.678 & 0.673 & 0.672 & \textbf{0.707} \\
& $F^{b}_\beta\uparrow$ & 0.395 & 0.310 & 0.335 & 0.382 & 0.343 & 0.323 & 0.456 & 0.396 & 0.439 & 0.382 & 0.503 & 0.545 & 0.510 & 0.505 & 0.466 & 0.514 & 0.573 & \textbf{0.601} \\
& $~M~\downarrow$ & 0.143 & 0.144 & 0.149 & 0.130 & 0.129 & 0.168 & 0.119 & 0.116 & 0.119 & 0.169 & 0.109 & 0.115 & 0.119 & 0.121 & 0.112 & 0.111 & 0.115 & \textbf{0.101} \\
& $S_{\alpha}\uparrow$ & 0.735 & 0.718 & 0.709 & 0.738 & 0.749 & 0.653 & 0.752 & 0.747 & 0.765 & 0.641 & 0.771 & 0.750 & 0.756 & 0.754 & 0.765 & 0.775 & 0.760 & \textbf{0.780} \\
& $E_{\phi}^{m}\uparrow$ & 0.763 & 0.760 & 0.755 & 0.784 & 0.780 & 0.706 & 0.800 & 0.808 & 0.784 & 0.718 & 0.820 & 0.810 & 0.801 & 0.798 & 0.809 & 0.800 & 0.806 & \textbf{0.829} \\
\hline \multirow{5}{*}{\textbf{OV}} & $F^w_\beta\uparrow$ & 0.637 & 0.622 & 0.616 & 0.671 & 0.682 & 0.527 & 0.701 & 0.733 & 0.721 & 0.529 & 0.723 & 0.721 & 0.697 & 0.707 & \textbf{0.752} & 0.723 & 0.730 & 0.749 \\
& $F^{b}_\beta\uparrow$ & 0.405 & 0.311 & 0.339 & 0.420 & 0.368 & 0.336 & 0.494 & 0.434 & 0.490 & 0.383 & 0.524 & 0.592 & 0.526 & 0.541 & 0.509 & 0.545 & 0.617 & \textbf{0.630} \\
& $~M~\downarrow$ & 0.173 & 0.180 & 0.184 & 0.159 & 0.150 & 0.216 & 0.136 & 0.125 & 0.127 & 0.217 & 0.129 & 0.134 & 0.148 & 0.146 & 0.119 & 0.126 & 0.132 & \textbf{0.114} \\
& $S_{\alpha}\uparrow$ & 0.721 & 0.700 & 0.688 & 0.728 & 0.745 & 0.624 & 0.751 & 0.762 & \textbf{0.781} & 0.611 & 0.761 & 0.748 & 0.747 & 0.752 & 0.779 & 0.774 & 0.764 & \textbf{0.781} \\
& $E_{\phi}^{m}\uparrow$ & 0.751 & 0.737 & 0.736 & 0.790 & 0.779 & 0.663 & 0.807 & 0.828 & 0.810 & 0.664 & 0.817 & 0.803 & 0.795 & 0.802 & \textbf{0.835} & 0.808 & 0.809 & 0.828 \\
\hline \multirow{5}{*}{\textbf{SC}} & $F^w_\beta\uparrow$ & 0.608 & 0.599 & 0.593 & 0.611 & 0.638 & 0.550 & 0.677 & 0.669 & 0.627 & 0.594 & 0.696 & 0.708 & 0.695 & 0.678 & 0.706 & 0.691 & 0.728 & \textbf{0.746} \\
& $F^{b}_\beta\uparrow$ & 0.481 & 0.407 & 0.414 & 0.433 & 0.423 & 0.427 & 0.561 & 0.455 & 0.492 & 0.504 & 0.572 & 0.627 & 0.613 & 0.597 & 0.562 & 0.603 & 0.654 & \textbf{0.672} \\
& $~M~\downarrow$ & 0.098 & 0.098 & 0.101 & 0.100 & 0.090 & 0.114 & 0.081 & 0.087 & 0.093 & 0.110 & 0.076 & 0.080 & 0.075 & 0.083 & 0.078 & 0.078 & 0.074 & \textbf{0.072} \\
& $S_{\alpha}\uparrow$ & 0.768 & 0.761 & 0.745 & 0.756 & 0.783 & 0.716 & 0.799 & 0.772 & 0.784 & 0.724 & 0.808 & 0.793 & 0.807 & 0.793 & 0.807 & 0.809 & 0.812 & \textbf{0.820} \\
& $E_{\phi}^{m}\uparrow$ & 0.794 & 0.799 & 0.788 & 0.806 & 0.814 & 0.765 & 0.841 & 0.837 & 0.799 & 0.792 & 0.854 & 0.858 & 0.856 & 0.844 & 0.851 & 0.843 & 0.861 & \textbf{0.872} \\
\hline \multirow{5}{*}{\textbf{SO}} & $F^w_\beta\uparrow$ & 0.523 & 0.524 & 0.526 & 0.531 & 0.561 & 0.487 & 0.567 & 0.602 & 0.566 & 0.518 & 0.596 & 0.623 & 0.626 & 0.594 & 0.621 & 0.614 & 0.634 & \textbf{0.684} \\
& $F^{b}_\beta\uparrow$ & 0.386 & 0.325 & 0.341 & 0.353 & 0.334 & 0.342 & 0.442 & 0.382 & 0.417 & 0.412 & 0.468 & 0.533 & 0.523 & 0.494 & 0.457 & 0.506 & 0.551 & \textbf{0.612} \\
& $~M~\downarrow$ & 0.119 & 0.109 & 0.115 & 0.116 & 0.099 & 0.118 & 0.096 & 0.092 & 0.095 & 0.113 & 0.089 & 0.091 & 0.087 & 0.098 & 0.090 & 0.082 & 0.092 & \textbf{0.075} \\
& $S_{\alpha}\uparrow$ & 0.718 & 0.713 & 0.703 & 0.706 & 0.737 & 0.682 & 0.732 & 0.736 & 0.748 & 0.682 & 0.746 & 0.745 & 0.768 & 0.749 & 0.755 & 0.767 & 0.758 & \textbf{0.787} \\
& $E_{\phi}^{m}\uparrow$ & 0.745 & 0.756 & 0.747 & 0.752 & 0.769 & 0.732 & 0.780 & 0.802 & 0.766 & 0.759 & 0.792 & 0.804 & 0.814 & 0.784 & 0.801 & 0.797 & 0.800 & \textbf{0.835} \\
\hline \multirow{5}{*}{\textbf{Avg.}} & $F^w_\beta\uparrow$ & 0.613 & 0.604 & 0.600 & 0.636 & 0.650 & 0.533 & 0.661 & 0.695 & 0.674 & 0.552 & 0.688 & 0.705 & 0.674 & 0.667 & 0.708 & 0.706 & 0.708 & \textbf{0.745} \\
& $F^{b}_\beta\uparrow$ & 0.407 & 0.330 & 0.343 & 0.394 & 0.363 & 0.343 & 0.486 & 0.424 & 0.466 & 0.401 & 0.513 & 0.577 & 0.525 & 0.521 & 0.495 & 0.543 & 0.599 & \textbf{0.631} \\
& $~M~\downarrow$ & 0.156 & 0.157 & 0.161 & 0.142 & 0.137 & 0.186 & 0.131 & 0.117 & 0.117 & 0.185 & 0.118 & 0.120 & 0.137 & 0.140 & 0.119 & 0.112 & 0.121 & \textbf{0.096} \\
& $S_{\alpha}\uparrow$ & 0.726 & 0.713 & 0.700 & 0.730 & 0.746 & 0.653 & 0.747 & 0.755 & 0.774 & 0.647 & 0.760 & 0.757 & 0.751 & 0.745 & 0.769 & 0.782 & 0.767 & \textbf{0.791} \\
& $E_{\phi}^{m}\uparrow$ & 0.748 & 0.743 & 0.739 & 0.775 & 0.775 & 0.692 & 0.788 & 0.815 & 0.793 & 0.706 & 0.803 & 0.806 & 0.788 & 0.780 & 0.809 & 0.809 & 0.806 & \textbf{0.837} \\
\hline
\end{tabular
\label{tab:compSOC}
\end{table*}
\subsubsection{Attribute-base Analysis}
In addition to the most frequently used salient object segmentation datasets, we also test our model on another dataset, SOC \cite{fan2018SOC}. The SOC dataset contains complicated scenarios, which are more challenging than those in the previous six SOD datasets. Besides, the SOC dataset categorizes images into nine different groups including AC (Appearance Change), BO (Big Object), CL (Clutter), HO (Heterogeneous Object), MB (Motion Blur), OC (Occlusion), OV (Out-of-View), SC (Shape Complexity), and SO (Small Object), according to their attributes. We train our BASNet on both DUTS-TR and the training set (1,800 images with salient objects) of SOC dataset \cite{fan2018SOC} and evaluate their performance on the testing set of SOC-Sal. There are totally 600 images with salient objects in the testing set. Each image may be categorized into one or multiple attributes (\emph{e.g.} AC and BO).
\textbf{Quantitative Evaluation:}
Tab. \ref{tab:compSOC} illustrates a comparison between our BASNet and 16 other state-of-the-art models, including Amulet \cite{DBLP:conf/iccv/ZhangWLWR17}, DSS \cite{hou2017deeply}, NLDF \cite{luo2017non}, C2S-Net \cite{DBLP:conf/eccv/LiYCLS18}, SRM \cite{DBLP:conf/iccv/WangBZZL17}, R3Net \cite{deng2018r3net}, BMPM \cite{zhang2018bi}, DGRL \cite{wang2018detect}, PiCANet-R (PiC(R)) \cite{liu2018picanet}, RANet \cite{chen2020reverse}, AFNet \cite{AFNet}, CPD \cite{wu2019cascaded}, PoolNet \cite{liu2019simple}, EGNet \cite{zhao2019egnet}, BANet \cite{su2019selectivity} and SCRN \cite{Wu_2019_ICCV} in terms of attribute-based performance. As we can see, our BASNet achieves obvious improvements against the existing methods. Particularly, our BASNet advances the boundary measure $F^b_\beta$ by large margins (over $5\%$ and sometimes over $10\%$) on different attributes.
\textbf{Qualitative Evaluation:} Fig.~\ref{fig:qual-soc} provides a qualitative comparison of our BASNet and other baseline models. As we can see, BASNet is able to handle different challenges, including small objects (1st row), out-of-view objects (2nd row), occluded targets (3rd row) and objects with complicated shapes (4th row).
\subsection{Experiments on Camouflaged Object Segmentation}
To further evaluate the performance of the proposed BASNet, we also tested it on the camouflaged object segmentation (COS) task \cite{Jinnan-2020,ltnghia-AAAI2021,fan2020camouflaged}. Compared with salient object segmentation, COS is a relatively newer and more challenging task. Because the contrast between the camouflaged targets and their backgrounds is sometimes extremely low. Besides, the targets usually have similar color and texture to their backgrounds. In addition, their shape or structure of these targets can sometimes be very complex.
\begin{table*}[t!]
\caption{Comparison of the proposed method and 13 other methods on three camouflaged object segmentation datasets: CHAMELEON, CAMO-Test and COD10K-Test. $\uparrow$ and $\downarrow$ indicate the higher score the better and the lower the score the better, respectively. \textbf{Bold} font indicates the best performance.}
\centering
\centering
\footnotesize
\renewcommand{\arraystretch}{0.5}
\setlength\tabcolsep{5.0pt}
\begin{tabular}{l|ccccc|ccccc|ccccc}
\hline
\multirow{2}{*}{\textbf{Baseline Models}} & \multicolumn{5}{|c|}{\textbf{CHAMELEON}\cite{chameleon}} & \multicolumn{5}{|c|}{\textbf{CAMO-Test}\cite{le2019anabranch}} & \multicolumn{5}{|c}{\textbf{COD10K-Test}\cite{fan2020camouflaged}}\\
\cline{2-16}
& $F^w_\beta\uparrow$ & $F^{b}_\beta\uparrow$ & $~M~\downarrow$ & $S_{\alpha}\uparrow$ & $E_{\phi}^{m}\uparrow$ &
$F^w_\beta\uparrow$ & $F^{b}_\beta\uparrow$ & $M\downarrow$ & $S_{\alpha}\uparrow$ & $E_{\phi}^{m}\uparrow$ &
$F^w_\beta\uparrow$ & $F^{b}_\beta\uparrow$ & $M\downarrow$ & $S_{\alpha}\uparrow$ & $E_{\phi}^{m}\uparrow$\\
\hline
\textbf{FPN}$_\text{CVPR17}$ & 0.590 & 0.246 & 0.075 & 0.794 & 0.784 & 0.483 & 0.232 & 0.131 & 0.684 & 0.677 & 0.411 & 0.195 & 0.075 & 0.697 & 0.692 \\
\textbf{MaskRCNN}$_\text{CVPR17}$ & 0.518 & 0.128 & 0.099 & 0.643 & 0.778 & 0.430 & 0.117 & 0.151 & 0.574 & 0.715 & 0.402 & 0.110 & 0.081 & 0.613 & 0.748 \\
\textbf{PSPNet}$_\text{CVPR17}$ & 0.555 & 0.207 & 0.085 & 0.773 & 0.756 & 0.455 & 0.191 & 0.139 & 0.663 & 0.659 & 0.377 & 0.166 & 0.080 & 0.678 & 0.681 \\
\textbf{UNet++}$_\text{DLMIA18}$ & 0.501 & 0.246 & 0.094 & 0.695 & 0.763 & 0.392 & 0.232 & 0.149 & 0.599 & 0.654 & 0.350 & 0.195 & 0.086 & 0.623 & 0.674 \\
\textbf{PiCANet}$_\text{CVPR18}$ & 0.536 & 0.200 & 0.084 & 0.769 & 0.749 & 0.356 & 0.166 & 0.155 & 0.609 & 0.584 & 0.322 & 0.173 & 0.083 & 0.649 & 0.643 \\
\textbf{MSRCN}$_\text{CVPR19}$ & 0.443 & 0.074 & 0.091 & 0.637 & 0.686 & 0.454 & 0.128 & 0.133 & 0.618 & 0.669 & 0.419 & 0.101 & 0.073 & 0.641 & 0.706 \\
\textbf{PFANet}$_\text{CVPR19}$ & 0.378 & 0.096 & 0.139 & 0.679 & 0.648 & 0.391 & 0.130 & 0.169 & 0.659 & 0.622 & 0.286 & 0.107 & 0.118 & 0.636 & 0.618 \\
\textbf{HTC}$_\text{CVPR2019}$ & 0.204 & 0.071 & 0.129 & 0.517 & 0.489 & 0.174 & 0.076 & 0.172 & 0.477 & 0.442 & 0.221 & 0.099 & 0.088 & 0.548 & 0.520 \\
\textbf{PoolNet}$_\text{CVPR2019}$ & 0.555 & 0.151 & 0.079 & 0.777 & 0.779 & 0.494 & 0.155 & 0.128 & 0.703 & 0.698 & 0.416 & 0.126 & 0.07 & 0.705 & 0.713 \\
\textbf{ANet-SRM}$_\text{CVIU19}$ & - & - & - & - & - & 0.484 & 0.217 & 0.126 & 0.682 & 0.686 & - & - & - & - & - \\
\textbf{CPD}$_\text{CVPR2019}$ & 0.706 & 0.383 & 0.052 & 0.853 & 0.868 & 0.550 & 0.306 & 0.115 & 0.726 & 0.730 & 0.508 & 0.286 & 0.059 & 0.747 & 0.771 \\
\textbf{EGNet}$_\text{ICCV19}$ & 0.702 & 0.289 & 0.050 & 0.848 & 0.871 & 0.583 & 0.264 & 0.104 & 0.732 & 0.768 & 0.509 & 0.209 & 0.056 & 0.737 & 0.779 \\
\textbf{SINet}$_\text{CVPR20}$ & 0.740 & 0.410 & 0.044 & 0.869 & 0.893 & 0.606 & 0.334 & 0.100 & \textbf{0.752} & 0.772 & 0.551 & 0.311 & 0.051 & 0.771 & 0.809 \\
\hline
\textbf{BASNet}~(Ours) & \textbf{0.866} & \textbf{0.650} & \textbf{0.022} & \textbf{0.914} & \textbf{0.954} & \textbf{0.646} & \textbf{0.420} & \textbf{0.096} & 0.749 & \textbf{0.796} & \textbf{0.677} & \textbf{0.546} & \textbf{0.038} & \textbf{0.802} & \textbf{0.855} \\
\hline
\end{tabular
\label{tab:compCOD}
\end{table*}
\subsubsection{Datasets}
We test our model on the CHAMELEON \cite{chameleon}, CAMO-Test \cite{le2019anabranch} and COD10K-Test datasets \cite{fan2020camouflaged}. CHAMELEON \cite{chameleon} contains 76 images taken by independent photographers. These images are marked as good examples of camouflaged animals by the photographers.
CAMO \cite{le2019anabranch} contains both camouflaged and non-camouflaged subsets. We use the camouflaged subset, which comprises two further subsets: CAMO-Train (1,000 images) and CAMO-Test (250 images).
COD10K \cite{fan2020camouflaged} is currently the largest camouflaged object detection dataset. It comprises 10,000 images of 78 object categories in various natural scenes. There are 5,066 images densely labeled with accurate (matting-level) binary masks. COD10K consists of 3,040 images for training (COD10K-Train) and 2,026 images for testing (COD10K-Test). For fair comparison, we use the same training sets as SINet \cite{fan2020camouflaged}.
\subsubsection{Comparison with State-of-the-Arts}\label{sec:CmpCOD}
To validate the performance of the proposed BASNet on the camouflaged object segmentation task, we compared BASNet
against 13 sate-of-the-art models, including FPN\cite{lin2017feature}, MaskRCNN\cite{he2017mask}, PSPNet\cite{zhao2017pyramid}, UNet++\cite{zhou2018unet++}, PiCANet\cite{liu2018picanet}, MSRCN\cite{huang2019mask}, PFANet\cite{zhao2019pyramid}, HTC\cite{chen2019hybrid}, PoolNet\cite{liu2019simple}, ANet-SRM\cite{le2019anabranch}, CPD\cite{wu2019cascaded}, EGNet\cite{zhao2019egnet} and SINet\cite{fan2020camouflaged}. For fair comparison, the results of different models are either provided by the authors or obtained by re-training the model with the default settings with same training data.
\begin{figure}[t!]
\centering
\begin{overpic}[width=.9\columnwidth]{./Imgs/FailCOD-min.pdf}
\put(9,2){\small (a) Image}
\put(45,2){\small (b) GT}
\put(75,2){\small (c) BASNet}
\end{overpic}
\vspace{-5pt}
\caption{Failure cases on camouflaged object segmentation task. The first row shows the typical false negative artifacts. The second row illustrates the false positive phenomenon.
}\label{fig:fail_cod}
\end{figure}
\textbf{Quantitative Evaluation:}
The quantitative evaluation results are illustrated in Table \ref{tab:compCOD}. As we can see that our BASNet achieves the best performance in nearly all metrics with great advantages. SINet is the second best model. EGNet and CPD are competitive with each other and can be ranked as the third and fourth best. Our BASNet improves the weighted F-measure $F^w_\beta\uparrow$ with large margins ($12.6\%$, $4.0\%$ and $12.6\%$ on CHAMELEON, CAMO-Test and COD10K-Test respectively). Particularly, our BASNet outperforms the second best model SINet by $24.0\%$, $8.6\%$ and $23.5\%$ in terms of the relax boundary F-measure $F^b_\beta\uparrow$ on CHAMELEON, CAMO-Test and COD10K-Test datasets. This reveals the effectiveness and accuracy of our BASNet in capturing boundaries and fine structures. In terms of the $M\downarrow$, our BASNet reduces the metric by $50.0\%$, $4.0\%$ and $25.5\%$ on the three datasets, respectively. For the structure measures, $S_{\alpha}\uparrow$, the improvements of our BASNet against the second best model are also significant ($4.5\%$ and $3.1\%$ on CHAMELEON and COD10K-Test datasets) but there is a $0.3\%$ $S_{\alpha}$ decrease on CAMO-Test. Compared with SINet, $E^m_{\phi}\uparrow$ takes both local and global structure similarity into consideration. As we can see, our BASNet achieves even greater improvements ($6.1\%$, $2.4\%$ and $4.6\%$ on the three datasets, respectively) against the second best model in $E^m_{\phi}\uparrow$ than in $S_{\alpha}\uparrow$.
\begin{figure*}[t!]
\centering
\begin{overpic}[width=.9\textwidth]{./Imgs/QualiCOD-min.pdf}
\put(3,1){\small (a) Image}
\put(16,1){\small (b) GT}
\put(27,1){\small (c) BASNet}
\put(39.5,1){\small (d) SINet}
\put(52,1){\small (e) EGNet}
\put(64,1){\small (f) PoolNet}
\put(78,1){\small (g) CPD}
\put(89,1){\small (h) PiCANet}
\end{overpic}
\caption{Qualitative comparison on camouflaged object segmentation datasets. See \secref{sec:CmpCOD} for details.
}\label{fig:qual-cod}
\end{figure*}
\textbf{Qualitative Evaluation:}
Qualitative comparisons against several of the baseline models are illustrated in Fig. \ref{fig:qual-cod}. As we can see, our BASNet (the 3rd column) is able to handle different types of challenging camouflaged cases including complex foreground targets with low contrast (the 1st row), targets with very thin foreground structures (the 2nd and 5th row), targets occluded by fine objects (the 3rd row), targets with complicated boundaries (the 4th row), targets with extremely complex hollow structures (the 5th row), multiple objects with low contrast (the 6th row), \emph{etc}.
Compared with the results of other models, the results of our BASNet demonstrates its excellent abilities of perceiving fine structures and complicated boundaries, which also explains why our BASNet is able to achieve such high boundary evaluation scores $F^b_\beta\uparrow$ on camouflage object segmentation datasets (see \tabref{tab:compCOD}).
\subsubsection{Failure Cases}
Although our BASNet outperforms other camouflaged object segmentation (COS) models and rarely produces completely incorrect results, there are still some false negative (the 1st row in Figure \ref{fig:fail_cod}) and false positive predictions (the 2nd row in Fig. \ref{fig:fail_cod}) in many of the COS cases. It is worth noting that other models usually have the same or even worse results on these challenging cases. Although these failure cases may not have a huge impact on evaluation metrics, they will somehow limit the applications and degrade the user experiences.
\section{Applications}
Thanks to the high accuracy, fast speed and simple architecture of our network, we developed two real-world applications based on BASNet: \textbf{AR COPY \& PASTE} and \textbf{OBJECT CUT}. These two applications further demonstrate the effectiveness and efficiency of our BASNet.
\subsection{Application I: AR COPY \& PASTE}
Introduced in HCI by Larry Tesler in the early 70s \cite{tesler2012personal}, cut/copy-paste has become essential to many applications of modern computing. In this section, we explore how BASNet can be used to apply this principle to the mobile phone camera and seamlessly transfer visual elements between the physical space and the digital space. AR COPY \& PASTE is a prototype that we built upon our BASNet to conveniently capture real-world items using the camera of a mobile phone (\emph{e.g.}~objects, people, drawings, schematics, \emph{etc}), automatically remove the background, and insert the result in an editing software on the computer by pointing the camera at it, as shown in Fig. \ref{fig:arcopypaste_steps}. Specifically, AR COPY \& PASTE first removes the background of the photo and only shows the foreground salient target on the mobile phone screen. Then users can ``paste'' the segmented target by moving the cellphone to point the mobile camera at a specific location of a document opened on a computer. The whole process of AR COPY \& PASTE makes it seem like the real-world target is ``copied'' and ``pasted'' into a document, which provides a novel and inspiring interactive experience. A demonstration video\footnote{\url{https://twitter.com/cyrildiagne/status/1256916982764646402}} of the prototype has been released along with the source code.\footnote{\url{https://github.com/cyrildiagne/ar-cutpaste}} Both have received world-wide attention (millions of views for the video, tens of thousands of Github stars for the source code). Hundreds of thousands of people have subscribed to the private beta.
\begin{figure}[t!]
\centering
\begin{overpic}[width=.9\columnwidth]{./Imgs/arcopypaste_steps-min.pdf}
\put(9,-1){\small (a) Copy}
\put(45,-1){\small (b) Move}
\put(76,-1){\small (c) Paste}
\end{overpic}
\vspace{0.5pt}
\caption{Screenshots from the video demonstration. (a) \textbf{Copy}: Point and tap to ``copy'' the object by masking its background out using BASNet. (b) \textbf{Move}: Move the mobile phone, where the ``copied'' object is shown, to target at the computer screen at a specific location. (c) \textbf{Paste}: Tap to ``paste'' the ``copied'' object into the current document.
}\label{fig:arcopypaste_steps}
\end{figure}
\subsubsection{Workflow of AR COPY \& PASTE}
\begin{table*}[t!]
\centering
\renewcommand{\arraystretch}{0.5}
\setlength\tabcolsep{18pt}
\caption{Number of operations of different methods for image capturing and masking out.}
\begin{tabular}{c|cc}
\hline
Methods & Number of steps (on mobile) & Number of steps (on desktop) \\
\hline
\tabincell{c}{Cross-platform \\(Android v11 and macOS v10.15)} & \tabincell{l}{6\\
\textcircled{1} Take a photo, \\ \textcircled{2} Tap the thumbnail, \\ \textcircled{3} Tap share, \\ \textcircled{4} Tap more, \\ \textcircled{5} Select Bluetooth, \\ \textcircled{6} Tap the destination on a device.} & \tabincell{l}{ 5 \\
\textcircled{1} Click ``Open'' on the Bluetooth notification, \\
\textcircled{2} Export image, \\
\textcircled{3} Select destination software, \\
\textcircled{4} Use ``Select Object'' Tool, \\
\textcircled{5} Apply mask.} \\
\hline
\tabincell{c}{Constructor-specific \\ (iOS v13 and macOS v10.15)} &
\tabincell{l}{5 \\
\textcircled{1} Take a photo, \\
\textcircled{2} Tap the thumbnail, \\
\textcircled{3} Tap share, \\
\textcircled{4} Tap “AirDrop”, \\ \textcircled{5} Tap the destination on a device.} &
\tabincell{l}{5 \\
\textcircled{1} Click “Open” on the AirDrop notification, \\
\textcircled{2} Export image, \\ \textcircled{3} Select destination software, \\
\textcircled{4} Use ``Select Object'' Tool, \\
\textcircled{5} Apply mask.} \\
\hline
AR COPY \& PASTE (Ours) & \tabincell{l}{\hspace{-0.2in}2 \\
\hspace{-0.2in}\textcircled{1} Take a photo, \\
\hspace{-0.2in}\textcircled{2} Tap toward the destination \\
software.} & \tabincell{l}{0 \\
} \\
\hline
\end{tabular}
\label{tab:arcopypaste_comp}
\end{table*}
From the perspective of users, our AR COPY \& PASTE only consists of two main steps: ``copy'' and ``paste'', as shown in Fig. \ref{fig:arcopypaste}. \textbf{(1) Copy.} The first step consists in pointing the mobile camera at a subject and tapping the screen to take a picture. BASNet is then used to hide all the pixels that are not part of the main foreground subject. The remaining pixels keep floating on top of the camera and provide a preview of the paste result. Compared to other methods like image segmentation \cite{minaee2020image}, BASNet can produce very accurate segmentation results with sharp and high-quality boundaries, which is essential in many image composition workflows. \textbf{(2) Paste.} The second step consists of pointing the mobile phone at a specific location on the computer screen and tapping to ``paste'' the ``copied'' subject. SIFT \cite{lowe2004distinctive} (implemented in OpenCV \cite{opencv_library}) is used to find the corresponding computer screen coordinates targeted by the center of the mobile phone camera. The image containing the background removed target is finally imported in an image editing software at the computed screen coordinates to create the final composition.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\linewidth]{./figures/arcopypaste/arcopypasteflow.png}
\end{center}
\caption{Schematic of the AR COPY \& PASTE flow.}
\label{fig:arcopypaste}
\end{figure}
\begin{figure}[thbp]
\begin{center}
\includegraphics[width=\linewidth]{./figures/arcopypaste/ar_pipeline.png}
\end{center}
\caption{Overall pipeline of the AR COPY \& PASTE.}
\label{fig:arcopypaste_pip}
\end{figure}
\begin{figure}[t!]
\centering
\begin{overpic}[width=\columnwidth]{./Imgs/ObjectCut-min.pdf}
\end{overpic}
\vspace{-1pt}
\caption{Sample results given by the OBJECT CUT API: the first row shows the input images and the second row shows the background removed results.
}\label{fig:object_cut_results}
\end{figure}
\subsubsection{Implementation Details}
Fig. \ref{fig:arcopypaste_pip} illustrates the overall implementation pipeline of the AR COPY \& PASTE prototype, which consists of three main parts: Kubernetes cluster, mobile application and desktop application. The AR COPY \& PASTE prototype was built using our BASNet model trained on DUTS-TR \cite{wang2017learning}. To make sure that it runs smoothly on mobile device, it has been wrapped as an HTTP service/container image\footnote{\url{https://github.com/cyrildiagne/basnet-http}} so that it can be deployed easily on remote GPUs using Kubernetes \footnote{\url{https://github.com/kubernetes/kubernetes}}. Hence, photos taken by mobile devices in AR COPY \& PASTE are sent to the remote server to obtain their segmentation masks. The desktop application is a python HTTP server which takes three files from the mobile application as input (original picture, BASNet mask, photo of the screen) and runs SIFT feature matching and perspective transformation based on the photo of the screen and the screenshot to determine the destination coordinates. Finally, the desktop application is responsible for sending javascript commands to desktop apps like Photoshop\footnote{\url{https://www.photoshop.com/en}} in order to import the image into the document at the right position.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\linewidth]{./figures/applications/object_cut/figure.png}
\end{center}
\caption{OBJECT CUT pipeline.}
\label{fig:object_cut}
\end{figure}
\subsubsection{Comparison with Other Methods}
Our AR COPY \& PASTE prototype applies BASNet in a novel human-computer-interaction setting, which makes the process easier and faster than other methods (two operations for our method versus 10 or 11 operations for other methods). Table \ref{tab:arcopypaste_comp} illustrates examples of typical user flow to clip and import an object from a mobile devices to an desktop image editing software, such as Adobe Photoshop. As we can see, our prototype greatly reduces the numbers of operations and simplifies the process. Besides, our AR COPY \& PASTE allows users to delegate some of the lower-level decisions (how visible each pixel should be), and focus on the higher-level objectives (how do they want the object to look). Removing these tasks lowers the barrier to entry (there is no need to learn how to paint masks in an image editing software), saves a significant amount of time, and ultimately leads to better end results by removing the cognitive load of the low-level tasks \cite{carter2017using}.
\subsection{Application II: OBJECT CUT}
OBJECT CUT is an online image background removal service that uses BASNet. Removing the background from an image is a common operation in the daily work of professional photographers and image editors. This process is usually a repeatable and manual task that requires a lot of effort and time. However, thanks to BASNet, one of the most robust and fastest performing deep learning models for image segmentation, OBJECT CUT was able to turn it into an easy and automatic process.
The program was built as an API to make it as easy as possible to integrate. APIs, also known as Application Programming Interfaces, are already commonly used to integrate different types of solutions to improve systems without actually knowing what is happening inside. For instance, RESTful APIs are a standard in the software engineering field for designing and specifying APIs. Making it substantially easier to adapt desired APIs to specific workflows.
Our system is based on three well-distinguished parts, as shown in Fig. \ref{fig:object_cut}:
1) the API Handler, responsible for receiving and validating clients requests, downloading and preprocessing input images and sending those to the model;
2) BASNet, responsible for performing salient object detection.
3) Once the output from the network is generated, the postprocessing module applies an unsharp masking algorithm and morphological operations to improve the quality of the output. Afterward, OBJECT CUT uploads the cropped image to the Google Cloud Storage and returns its public URL to the client. This is structured as-is in order to isolate different parts, such as the BASNet component, removing all the API parameter validations as well as image download and upload processes, as much as possible. In this scenario, OBJECT CUT maximizes the operations running on the BASNet thread. The whole stack from the API is running under Docker containers, all managed by the cloud native application proxy called Traefik. The usage of Traefik here allows us to have a production-ready deployment making easy, from the containers' perspective, to communicate with other processes. In addition, we have a Load Balancer system in place to enable each of the components to scale more easily. The source code for this pipeline can be found under the OBJECT CUT GitHub repository: \url{https://github.com/AlbertSuarez/object-cut}.
To ensure easy integration, it is publicly available at RapidAPI\footnote{https://rapidapi.com/objectcut.api/api/background-removal}, the biggest API marketplace in the world, and it has been effectively utilized by people and companies from 86 different countries around the globe, including China, United States, Canada, India, Spain, Russia, \emph{etc}. OBJECT CUT was born to power up the designing and image editing process for the people who work with images daily. Integrating the OBJECT CUT API removes the necessity of understanding the complex inner workings behind it and automates the process of removing the background from images in a matter of seconds. See examples in Fig. \ref{fig:object_cut_results}.
\section{Conclusion}
In this paper, we proposed a novel end-to-end boundary-aware model, BASNet, and a hybrid fusion loss for accurate image segmentation.
The proposed BASNet is a predict-refine architecture, which consists of two components: a prediction network and a refinement module.
Combined with the hybrid loss, BASNet is able to capture both large-scale and fine structures, \emph{e.g.}~thin regions, holes, and produce segmentation probability maps with highly accurate boundaries.
Experimental results on six salient object segmentation datasets, one salient object in clutter dataset and three camouflaged object segmentation datasets demonstrate that our model achieves very competitive performance in terms of both region-based and boundary-aware measures against other models.
Additionally, the proposed network architecture is modular. It can be easily extended or adapted to other tasks by replacing either the prediction network or the refinement module.
The two (close to) commercial applications, AR COPY \& PASTE and OBJECT CUT, based on our BASNet not only prove the effectiveness and efficiency of our model but also provide two practical tools for reducing the workload in real-world production scenarios. The world-wide impacts of these two applications indicates the huge demand for highly accurate segmentation approaches, which motivates us to explore more accurate and reliable segmentation models.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{ieee_fullname}
|
\section{Introduction}
For robots to work safely and effectively, they must be aware of their environment. One aspect of this is the estimation of the pose of the objects in the scene to be able to avoid collisions and allow robust grasping and manipulation of the components.
6D object pose estimation (OPE) and grasp planning in highly cluttered scenes based on a single depth image is
challenging because of sensor noise, incomplete information, and uncertainties about the state of the environment. Furthermore, the robot has to reason
on how to manipulate the objects because selecting the wrong object and grasp pose can result in failed grasps.
\begin{figure}[tb]
\centering
\includegraphics[scale=0.25]{fig_result_real_world_data.png}
\caption{Estimated object poses of our approach on real-world data after ICP refinement. The greener the object, the more certain the model is that the object can be grasped safely. The gripper (blue) indicates the top ranked grasp based on our
policy.
}
\label{fig:result_real_world_data}
\end{figure}
Works such as~\cite{DOPE,IanLenz.2015,JosephRedmon.2015,Jacquard} focus on robotic grasping and manipulation tasks in scenarios with limited clutter
which do not require a defined picking order of the objects.
Simply selecting collision-free and kinematically feasible grasps in highly cluttered scenes~\cite{FPS_GPC_1,FPS_GPC_2}, might lead to a movement of the object relative to the gripper which prevents a precise placement without additional in-hand localization and entanglements with other objects for complex object geometries as visualized in Fig.~\ref{fig:Motivation}.
In this paper, we tackle these challenges by providing a novel learning-based approach for grasp pose evaluation in
scenes of many parts in bulk.
\begin{figure}[tb]
\centering
\input{fig_motivation_V2.pdf_tex}
\caption{Failure cases
for picking tasks in cluttered scenes:
(left) The robot picks an object (IPABar) which moves relative to the gripper during
lifting due to overlaps with other objects. The object cannot be placed precisely anymore. The objects in the bin have to be picked in a defined order.
(right) The robot picks an object (IPAUBolt) which is entangled with another object.
Therefore, a goal is to select objects
which do
do not entangle
with other objects.}
\label{fig:Motivation}
\end{figure}
Approaches to robotic grasping and manipulation usually rely on datasets consisting of human-labeled grasps~\cite{IanLenz.2015,JosephRedmon.2015,GG-CNN}, which are tedious to get, or on physical grasp outcomes where data collection can take several months~\cite{Google_2016_Learning_Hand-Eye_Coordination,Google_2018_Learning_Hand-Eye_Coordination,Supersizing_Self_Supervision,QT_Opt}.
In this work, we execute each predefined grasp pose in a physics simulation and
transfer the gained
experience
to the real world using domain randomization~\cite{Domain_Randomization} to increase generalization.
Our approach directly transfers from simulation to the real world (see Fig.~\ref{fig:result_real_world_data}).
Learning-based approaches to robotic grasping~\cite{Dex-Net_1.0,Dex-Net_2.0,Dex-Net_3.0,Google_2016_Learning_Hand-Eye_Coordination,Google_2018_Learning_Hand-Eye_Coordination,QT_Opt} usually rely on top-down grasps and cannot be used for bin-picking due to collisions when attempting to grasp objects close to the border of the bin. Analytical approaches for \mbox{pick-and-place} tasks in cluttered scenes require an object-specific configuration and tuning until a satisfactory
system performance is reached, which limits the scalability~\cite{FPS_GPC_1,Config_Effort_Bin_Picking,ReviewArticleSpringer_KBK,paper_TEM,IPA_Diss_ThomasLedermann.2012,IPA_Diss_MatthiasPalzkill.2014}. \mbox{PQ-Net} configures automatically based on a given object model and does not require any human intervention.
Inspired by the success of single shot approaches~\cite{YOLOv1,YOLOv2,SSD,Tekinetal.},
we go further and extend OP-Net~\cite{OP-Net}, a single shot approach for OPE outperforming the winner of the ``\textit{Object Pose Estimation Challenge for Bin-Picking}'' at IROS 2019\footnote{\url{http://www.bin-picking.ai/en/competition.html}} on the Sil\'{e}ane dataset~\cite{Bregier_SymmetryAwareEvaluation}. To the best of our knowledge, we are the first extending a single shot approach for OPE to
grasp success prediction in a joint framework.
Based on a single depth image, \mbox{PQ-Net} predicts the object pose $P \in \mathrm{SE}(3)$ relative to the camera coordinate system and outputs a success estimate for a set of predefined grasps $\mathcal{G}$ defined relative to the object coordinate system.
Furthermore, we introduce graspability metrics
which allow a
gentle component removal
and avoid entanglements.
In summary, the main contributions of this work are:
\begin{itemize}
\item Extension of a single shot approach
for object pose estimation to grasp pose success prediction suitable for precise object placement
\item Novel metrics for the assessment of the graspability of objects in highly cluttered scenes
\item Scalable system, which enables a robot to learn how to place objects precisely on the basis of an object model of arbitrary symmetry (automatic configuration)
\item Extension of the Fraunhofer IPA~\cite{FraunhoferIPABinPickingDataset} and Sil\'{e}ane~\cite{Bregier_SymmetryAwareEvaluation} datasets with grasp annotations
and provide data for two new objects. All datasets are publicly available at
\url{http://www.bin-picking.ai/en/dataset.html}.
\end{itemize}
The paper is structured as follows. The next section reviews related work. In Section~\ref{sec:approach} the proposed approach is described. Experimental evaluations are provided in Section~\ref{sec:evaluation}. Pros and cons of our approach are discussed in Section~\ref{sec:discussion}. The paper closes with a conclusion.
\section{Related Work}
Methods for robotic grasping can roughly be categorized in analytical
and
data-driven methods~\cite{Sahbani_2012,Bohg_2014,ReviewArticleSpringer_KBK}.
\subsection{Analytical Approaches}
Analytical approaches (model-based approaches) use an object model with predefined grasps. First off, they localize the object in the scene~\cite{BinPicking_6D_OPE_IROS_2019}. Based on this, they try to find a collision-free and kinematically feasible path
for grasping and placing~\cite{FPS_GPC_1,FPS_GPC_2}.
Especially for highly cluttered scenes, they require significant effort for manually tuning suitable grasp poses and grasp priorities to reach a satisfactory
system performance,
limiting the scalability to new
objects~\cite{FPS_GPC_1,Config_Effort_Bin_Picking,ReviewArticleSpringer_KBK,paper_TEM,IPA_Diss_ThomasLedermann.2012,IPA_Diss_MatthiasPalzkill.2014}. Usually, the grasp poses are prioritized independent of the object pose. Furthermore, zones on the object can be specified where no measurement point of the 3D point cloud should be contained in order to pick the candidate next. This can be used to specify a
picking order of the localized objects in the scene.
\subsection{Learning-based Approaches}
Approaches to Robotic Grasp Detection estimate oriented rectangles in the input image which represent a grasp configuration for parallel jaw grippers~\cite{rectangle_representation_ICRA_2011}.
Public datasets are the Cornell Grasping Dataset~\cite{IanLenz.2015} providing 1,035 manually annotated samples of 280 objects and the Jacquard Dataset~\cite{Jacquard}
with over 50,000 synthetic samples on a large diversity of objects (11,000), each with multiple labeled grasps.
MultiGrasp~\cite{JosephRedmon.2015} uses the Cornell dataset to train a neural network to predict oriented rectangles (bounding boxes) in an image together with a confidence and
makes local predictions based on global information by discretizing the output in $S \times S$ grid cells.
This work led to the
YOLO~\cite{YOLOv1,YOLOv2} approach for object detection.
With a two-stage system that first samples grasp candidates and ranks them using neural networks, Lenz et al.~\cite{IanLenz.2015} demonstrated that this parameterization (oriented rectangles) can be used for real-world robotic grasping tasks.
GG-CNN~\cite{GG-CNN, GG-CNN_2} predicts a quality and configuration of grasps at every
pixel of the input image using a lightweight convolutional neural network trained on the Cornell and Jacquard dataset~\cite{Jacquard}.
The generated antipodal grasps that are executed closed-loop and allow grasping in cluttered scenes and non-static environments.
Dex-Net makes use of large scale synthetic data collection for learning grasping policies for parallel jaw~\cite{Dex-Net_1.0, Dex-Net_2.0} and suction grippers~\cite{Dex-Net_3.0}
using analytic metrics. The sampled grasps are ranked using a neural network which gets a cropped depth image and grasp candidate as input.
Dex-Net observes a local image patch, and is not designed to execute grasps in a defined order or avoid entangled objects due to missing global scene information.
Levine et al.~\cite{Google_2016_Learning_Hand-Eye_Coordination,Google_2018_Learning_Hand-Eye_Coordination} parallelizes the real-word data collection to up to 14 robot and collect 800,000 samples in two months for robust grasping.
QT-Opt~\cite{QT_Opt} makes use of reinforcement learning
to train robotic grasping and manipulation policies based on self-supervision on real-world systems.
Because of the time-consuming and hardware demanding data collection procedure, works such as GraspGAN~\cite{KonstantinosBousmalis.2018} or RCAN~\cite{RCAN} focus on reducing the need of real-world data collection.
While all these aforementioned model-free approaches to robotic grasping show promising results, they do not provide a solution for a precise placement of the objects and only consider pick-and-drop tasks using top-down grasps (4D). Using grasps in this grasp representation
has limitations, e.g., for bin-picking due to collisions with the bin when attempting to grasp objects at the border. Therefore, works such as
\cite{Google_2016_Learning_Hand-Eye_Coordination,Google_2018_Learning_Hand-Eye_Coordination,Dex-Net_1.0,Dex-Net_2.0,QT_Opt,Morrison_MultiViewPicking,Berscheid.2019} use bins with slanted or no high
bin walls to ensure that the top-down grasps
work.
Furthermore, it is not possible to blindly move into the bin for data collection due to damaging the gripper because of unknown fill levels of the bins. Picking multiple objects is
often
also considered as a successful grasp~\cite{Google_2016_Learning_Hand-Eye_Coordination,Google_2018_Learning_Hand-Eye_Coordination,QT_Opt}.
\section{Placement Quality Network}
\label{sec:approach}
In this section, we describe the routine for automatically generating grasp poses for object models, the process of data generation for training our neural network using a physics simulation, the proposed definitions for the graspability of objects, the parameterization of the network's output, the loss function, the network architecture, and the training procedure together with the technique for a robust transfer of the model from simulation to the real world. Fig.~\ref{fig:Overview} shows an overview of our approach.
\subsection{Automatic Grasp Pose Generation}
To avoid the need for manually defining grasp poses $G \in \mathrm{SE}(3)$ on the object model, we provide a method that
automatically generates a set of grasp poses $\mathcal{G}$ for common gripper types such as parallel jaw, suction, and magnetic grippers based on a given 3D object model.
Each grasp $G \in \mathcal{G}$ is represented by $(R; t) \in \mathrm{SE}(3)$ where $R \in \mathrm{SO}(3)$ and $t \in \ensuremath{\mathds{R}}^3$ are the rotation and translation of grasp $G$.
As a first step of our technique, points are sampled on the surface of the object. For parallel jaw grippers, we check the distance between all pairs of points to verify whether it is smaller than the opening distance of the gripper,
filter the candidates using the normal information of the 3D points, discretize the rotation around the straight line between any two points in
\ang{20} steps, and
finally filter the candidates with a collision check using the gripper model. For suction and magnetic grippers, we sample grasp poses by evaluating the flatness of the object locally. Depending on the shape of the gripper, we define cylinder or cuboid elements, which should and should not contain points of the object model while also taking surface normals into account.
The proposed procedure results in a high number of grasp poses.
We make use of unsupervised learning to reduce the amount of data while keeping a high diversity in terms of position and orientation. We apply partitioning around medoids (PAM)~\cite{PAM} clustering
to reduce the number of grasps to approximately 500. Fig.~\ref{fig:GraspPoses} exemplary shows automatically generated grasp poses using our technique.
\begin{figure}[tb]
\centering
\input{fig_grasps.pdf_tex}
\caption{Automatically generated
grasp poses exemplary visualized for the IPARingScrew (left) and IPAGearShaft (right) for a parallel jaw gripper. Objects are taken from the Fraunhofer IPA dataset~\cite{FraunhoferIPABinPickingDataset}.}
\label{fig:GraspPoses}
\end{figure}
\subsection{Physics Simulation for Data Generation}
\begin{figure*}[thpb]
\begin{footnotesize}
\centering
\input{overview_GQ-Net.pdf_tex}
\caption{Overview of our approach: (a) 3D object model with automatically generated grasp poses (b) Physics simulation for scene generation (c) Physics simulation for grasp execution with a robot
(d) We train a deep neural network for 6D object pose estimation and grasp pose success estimation to transfer the knowledge gained in simulation to the real world using domain randomization~\cite{Domain_Randomization}. The output of our network is a 3D tensor comprising estimates of the probability $\hat{p}$, visibility $\hat{v}$, positions $\hat{x}$, $\hat{y}$, $\hat{z}$, Euler angles $\hat{\varphi}_1$, $\hat{\varphi}_2$, $\hat{\varphi}_3$, graspabilies $\hat{g}_\mathrm{a}$, $\hat{g}_\mathrm{u}$, $\hat{g}_\mathrm{e}$, and success $\hat{s}_j$ for each grasp pose $G_j \in \mathcal{G}$.}
\label{fig:Overview}
\end{footnotesize}
\end{figure*}
\subsubsection{Scene Generation}
We use the physics simulation \mbox{V-REP} / CoppeliaSim~\cite{V-REP} to create scenes with a high amount of clutter.
These scenarios are challenging because the robot has to avoid collisions with other objects in the scene and carefully select which object to pick next.
Analogous to the Sil\'{e}ane~\cite{Bregier_SymmetryAwareEvaluation} and Fraunhofer IPA~\cite{FraunhoferIPABinPickingDataset} datasets, we drop objects in a random position and orientation above a bin to generate chaotic scenes typical for bin-picking.
We save the RGB image, depth image, and segmentation masks together with the visibility $v \in [0,1]$ and pose $P \in \mathrm{SE}(3)$ for each object in the scene. The number of objects that are dropped is increased iteratively until a predefined drop limit is reached, resulting in a uniform distribution over different fill levels of the bin (see also~\cite{FraunhoferIPABinPickingDataset}).
\subsubsection{Grasping}
Using the filled bins, we loop over each (automatically generated) grasp pose for all objects in the scene.
First off, we check the collision of the gripper at every grasp pose with the environment (other objects and the bin).
In case no collision occurs, we try to find a kinematically feasible robot configuration and plan a collision-free path to the grasp pose using the OMPL~\cite{OMPL} module integrated in \mbox{V-REP} / CoppeliaSim~\cite{V-REP}.
In case a suitable path was found, we execute the grasp and place the object at the defined target pose. We log whether an object is in the gripper after lifting and after placement of the object.
Furthermore, we log the pose difference after grasping and lifting the object (chosen grasp pose relative to the gripper TCP) and placement (current object pose relative to defined placement pose). We consider an object as successfully lifted / grasped or placed precisely enough if the distance between the pose representatives based on~\cite{Bregier_PoseDistance} is less than 0.1 times the diameter of the smallest bounding sphere of the object. This is analogous to the metric used for object pose estimation in computer vision proposed by Br\'{e}gier et al.~\cite{Bregier_PoseDistance,Bregier_SymmetryAwareEvaluation} and allows to properly consider all possible kinds of object symmetry.
Furthermore, we log for each grasp pose whether an entanglement with other objects in the scene occurred.
Since the grasp poses are defined relative to the object coordinate system, pose ambiguities due to object symmetries result in convergence issues during neural notwork training. To avoid this, we introduce a unique object pose definition.
For discrete symmetries, we ensure to pick the pose where the $z$-component of a non-symmetry axis ($x$- or $y$-axis) is maximal (assuming the axis of symmetry is the $z$-axis). For continuous symmetries, we maximize the $z$-component of a non-symmetry axis by rotating around the axis of symmetry.
\subsection{Graspability Metrics}
Using the results from all executed grasps, we define instance based metrics to assess
the graspability of each object in the scene.
The graspability of an object based on the accessibility of the grasp poses $g_\mathrm{a} \in [0,1]$ is defined as the ratio between the number of collision-free grasps and
the number of total grasps $J$.
Fig.~\ref{fig:Graspability}~(a) shows the ground truth labels $g_\mathrm{a}$ for the IPARingScrew. Some fully visible objects (which are easy to localize) cannot be grasped because other objects are in the way, demonstrating that visibility and graspability of objects are not fully correlated.
For a removal that is gentle on the components
and to avoid movements of the grasped object relative to the gripper due to overlapping objects, entanglements, or jams preventing a precise object placement, we log the movement in terms of position $\mat{x} \in \ensuremath{\mathds{R}}^3$ of all other objects in the scene before grasping ($t_0$) and after lifting ($t_1$). This information is used
to define
the graspability of object $k$ based on the unrest caused in the bin
during grasping ("mikado metric")
\begin{equation}
g_{\mathrm{u},k}=1-\min \Bigg(\sum_{n=1,n\neq k}^{N}\lVert\mat{x}_{n,t_0}-\mat{x}_{n,t_1}\lVert,1\Bigg)
\end{equation}
with $\lVert.\rVert$ being the $L^2$ norm and $N$ being the number of objects in the scene without the picked object $k$~\cite{MA_Moosmann}.
Fig.~\ref{fig:Graspability}~(b) shows exemplary ground truth $g_\mathrm{u} \in [0,1]$ labels for the IPABar object. It can be seen that the objects at the top of the bin have a high graspability value regarding the unrest.
The graspability of object $k$ based on the entanglement with other objects
is
\begin{equation}
g_{\mathrm{e},k}=\begin{cases}
0,~\text{if an entanglement occurred for any grasp pose}\\
1,~\text{otherwise}.
\end{cases}
\end{equation}
Fig.~\ref{fig:Graspability}~(c) shows the ground truth $g_\mathrm{e}$ labels for the IPAUBolt. A goal for the robot is to avoid picking objects which can potentially entangle.
\begin{figure*}[thpb]
\centering
\input{fig_graspability.pdf_tex}
\caption{Exemplary ground truth samples: (a) Graspability $g_\mathrm{a} \in [0,1]$ of the IPARingScrew based on the accessibility of the grasp poses (each collision-free grasp pose is indicated by a small coordinate system). (b) Graspability $g_\mathrm{u} \in [0,1]$ of the IPABar based on the unrest caused in the bin
during grasping. (c) Graspability $g_\mathrm{e} \in \{0,1\}$ of the IPAUBolt based on entanglements with other objects after grasping. Objects shown in green have a high rating for tangibility. The more difficult an object is to grasp, the more the colour changes from green to yellow to red with increasing transparency.
}
\label{fig:Graspability}
\end{figure*}
\subsection{Parameterization of the Output}
Similar to~\cite{OP-Net}, we introduce a spatial discretization of the 3D scene into $S \times S$ volume elements (see white grid in Fig.~\ref{fig:Graspability}) and solve a regression problem locally, i.e., individually for each volume element. Each volume element comprises an $(11+J)$-dimensional vector containing the probability $p$, visibility $v$, positions $x$, $y$, $z$, Euler angles $\varphi_1$, $\varphi_2$, $\varphi_3$, graspabilities $g_\mathrm{a}$, $g_\mathrm{u}$, $g_\mathrm{e}$ of the object, and a $J$-dimensional vector with a success label $s \in \{0,1\}$ for each grasp pose $G$ for the considered task (grasping, precise placement).
For the ground truth generation, the objects are assigned to the volume element
which contains the origin of the object coordinate system.
In case multiple objects fall into the same volume element, we assign the object with the highest visibility $v$ as ground truth.
All volume elements not containing an object are filled with a zero vector. The output of the network is a $S \times S \times (11+J)$ tensor as depicted in Fig.~\ref{fig:Overview} (d).
\subsection{Loss Function}
To train the network, the multi-task
loss function
\begin{equation}
\mathcal{L} = \sum_{i=1}^{S^2} \bigg( \lambda_1\mathcal{L}_\mathrm{p}+\Big[ \lambda_2\mathcal{L}_\mathrm{v} + \lambda_3\mathcal{L}_\mathrm{pose}+\lambda_5\mathcal{L}_\mathrm{g} + \lambda_6\mathcal{L}_\mathrm{gp_s} \Big] p_i \bigg)
\end{equation}
is optimized. The $\lambda$-factors
are manually tuned weights for the different loss terms. While $\lambda_1=0.1$, $\lambda_2=0.1$, $\lambda_4=1$, $\lambda_5=1$, and $\lambda_6=1/J$ are constant, $\lambda_3=(g_\mathrm{a}+g_\mathrm{u}+g_\mathrm{e})^3$ is a function of the ground truth graspabilities $g_\mathrm{a}, g_\mathrm{u}, g_\mathrm{e}$ to make the network focus on the relevant objects for grasping.
For the loss of the pose
\begin{equation}
\mathcal{L}_\mathrm{pose}=\mathcal{L}_\mathrm{pos} + \lambda_4 \mathcal{L}_\mathrm{ori}
\end{equation}
we use
\begin{equation}
\mathcal{L}_\mathrm{pos} = \lVert\mat{x}-\hat{\mat{x}} \rVert^2
\end{equation}
with $\mat{x}=[x, y, z]^\top$ and
\begin{equation}
\mathcal{L}_\mathrm{ori} = \lVert\bm{\varphi}-\hat{\bm{\varphi}}\rVert^2
\end{equation}
with $\bm{\varphi}=[\varphi_1, \varphi_2, \varphi_3]^\top$
and $\varphi_1,\varphi_2 \in [0,2\pi)$ and $\varphi_3 \in [0,2\pi/k)$,
where $k \in \mathbb{N}$ represents the order of the cyclic symmetry.
To stabilize the training, the position $\mat{x} \in \ensuremath{\mathds{R}}^3$ of the object is estimated relative to the volume element, i.e., $x, y, z \in (0,1)$, while $z$ is the position between the near and far clipping plane of the 3D sensor, and the angles are bounded, i.e.,
$\varphi_1,\varphi_2,\varphi_3$
are mapped to $[0,1)$. For objects with a revolution symmetry, we omit the respective output feature-map.
We use the binary cross-entropy loss for the probability channel, the visibility channel, the three graspability channels, and the grasp pose result channels to compute $\mathcal{L}_\mathrm{p}$, $\mathcal{L}_\mathrm{v}$, $\mathcal{L}_\mathrm{g}$, and $\mathcal{L}_\mathrm{gp_s}$, respectively, while only backpropagating the loss for the elements that contain ground truth
by multiplying each channel element-wise with the ground truth probability channel.
\subsection{Network Architecture}
The input of our model is a single normalized depth image
which is
processed with a fully convolutional architecture and mapped to a 3D output tensor as shown in Fig.~\ref{fig:Overview}~(d). In the experiments, we use an input resolution of $128 \times 128$ pixel,
a DenseNet-BC~\cite{DenseNet} with 40 layers and a growth rate of 50, which represents the number of feature-maps being added per layer, and $S=16$. We choose a DenseNet-BC because it promotes gradient propagation by introducing direct connections between any two layers with the same feature-map size and has a high parameter efficiency.
The
network architecture consists of four dense blocks and downsampling is performed three times via $2 \times 2$ average pooling to reduce the size of the feature-maps from $128 \times 128$ to $16 \times 16$ and preserve the spatial information.
ReLU activation functions are employed in the dense blocks and sigmoid functions for
the 3D output tensor.
With this architecture, forward passes
are performed with a frame rate of 92~fps on a Nvidia Tesla V100.
\subsection{Training}
\label{sec:training}
During training, the error of the entire probability channel and the error of the remaining elements from the 3D output tensor that contain ground truth are backpropagated.
Since annotating the grasps is time-consuming, we do not annotate the whole training dataset
and only backpropagate
$\mathcal{L}_\mathrm{g}$ and $\mathcal{L}_\mathrm{gp_s}$ for the samples, where ground truth annotations are available.
In our experiments, we only annotate 100 out of 750 cycles from the Fraunhofer IPA~\cite{FraunhoferIPABinPickingDataset} and our newly provided dataset, resulting in 100 uniformly distributed samples over different fill levels of the bin.
We augment the training data by rotating around the $z$-axis of the camera coordinate system and mirroring the images if the object is symmetric with respect to a plane while adjusting the ground truth pose annotations accordingly. To not lose the information of the robot placement relative to the bin,
we only backpropagate $\mathcal{L}_\mathrm{gp_s}$ for the non-augmented samples.
For a robust Sim-to-Real Transfer, we use domain randomization~\cite{Domain_Randomization}. To allow the model generalizing on real-world data, we apply different augmentations with varying intensity to the rendered training images, e.g., adding noise, blurring, elastic transformations, dropout, etc. This allows \mbox{PQ-Net} generalizing to different 3D sensor technologies.
We use the Adam optimizer with an initial learning rate of 0.01, monitor the validation loss, reduce the learning rate by a factor of 10 if the loss did not improve for three epochs, and train the network for about 50 epochs on the synthetic data.
\subsection{Policy}
Based on a single
depth image $I$ with global scene information, the neural network $f$ with weights $\theta$ outputs a 3D tensor $\hat{T}$. Our policy $\pi$ uses the network output to select
the highest quality grasp weighted with $\hat{p}$, $\hat{v}$, $\hat{g_\mathrm{a}}$, $\hat{g_\mathrm{u}}$, and $\hat{g_\mathrm{e}}$ from all $S^2$ volume elements for execution
\begin{equation}
\pi(f_\theta(I)) = \argmax_{i,j}(\hat{s}_{j,i} \cdot \hat{p}_i \cdot \hat{v}_i \cdot \hat{g_\mathrm{a}}_i \cdot \hat{g_\mathrm{u}}_i \cdot \hat{g_\mathrm{e}}_i)
\end{equation}
with $i=1,...,S^2$, $j=1,...,J$ with $J$ being the number of predefined grasp poses, and $\hat{s}_{j,i}$ being the success estimate of grasp pose $G_j$ at volume element $i$
for the considered task (grasping, precise placement).
\section{Experimental Evaluation}
\label{sec:evaluation}
In this paper, we focus on parallel jaw grippers
using a RG2 gripper~\cite{RG2_gripper}. Given a proper physics simulation, our approach can easily be transferred to other gripper types.
\subsection{Sim-to-Real Transfer}
\begin{table*}[h]$ $
\begin{tiny}
\caption{Prediction of collision-free reachability of grasp poses of our approach on real-world / noisy data for different objects with different kinds of object symmetry form the Sil\'{e}ane~\cite{Bregier_SymmetryAwareEvaluation} and Fraunhofer IPA~\cite{FraunhoferIPABinPickingDataset} datasets.}
\label{table:collision_free_reachability}
\begin{center}
\begin{tabular}{|l||c|c|c|c|c|c|c|}
\hline
object & SileaneBunny & SileaneCandlestick & SileanePepper & SileaneGear & SileaneTLess20 & IPARingScrew & IPAGearShaft \\
\hline
object symmetry based on~\cite{Bregier_PoseDistance,Bregier_SymmetryAwareEvaluation} & no proper & revolution & revolution & revolution & cyclic & cyclic & revolution \\
& symmetry & & & & (order 2) & (order 2) & \\
\hline
\hline
PQ-Net success rate of policy & 0.98 & 0.97 & 0.98 & 0.99 & 0.98 & 0.98 & 0.99 \\
\hline
PQ-Net precision (all grasp poses) & 0.57 & 0.70 & 0.71 & 0.73 & 0.77 & 0.75 & 0.83 \\
\hline
PQ-Net recall (all grasp poses) & 0.52 & 0.67 & 0.66 & 0.64 & 0.65 & 0.45 & 0.65 \\
\hline
\hline
PQ-Net success rate OPE for chosen object & 0.89 & 0.92 & 0.95 & 0.98 & 0.98 & 0.98 & 0.99 \\
\hline
PQ-Net success rate OPE for chosen object with ICP & 0.91 & 0.93 & 0.95 & 0.99 & 0.99 & 0.99 & 0.99 \\
\hline
PQ-Net AP (OPE) whole scene & 0.86 & 0.88 & 0.92 & 0.74 & 0.82 & 0.86 & 0.98 \\
\hline
OP-Net~\cite{OP-Net} AP (OPE) whole scene & 0.92 & 0.95 & 0.98 & 0.82 & 0.85 & 0.88 & 0.99 \\
\hline
\end{tabular}
\end{center}
\end{tiny}
\end{table*}
For demonstrating a robust transfer of
\mbox{PQ-Net} to the real world, we extend the Sil\'{e}ane~\cite{Bregier_SymmetryAwareEvaluation} and Fraunhofer IPA~\cite{FraunhoferIPABinPickingDataset} datasets with annotations for the collision-free reachability of approximately 500 densely sampled grasp poses (examples see Fig.~\ref{fig:GraspPoses}). In Table~\ref{table:collision_free_reachability} we report the success rate of our grasping policy and the precision and recall over all
grasp poses in the scene.
Applying randomizations on the synthetic images during training (see Section~\ref{sec:training}) allows \mbox{PQ-Net} providing robust pose estimates and very high success rates of the policy on real-world data recorded with different 3D sensors.
Table~\ref{table:collision_free_reachability} gives the average precision (AP) results for the object pose estimation based on the metric provided by Br\'{e}gier et al.~\cite{Bregier_PoseDistance, Bregier_SymmetryAwareEvaluation} from OP-Net~\cite{OP-Net} with exactly the same depth image resolution, network architecture, and output discretization as used for \mbox{PQ-Net}. Even with a success estimate for approximately 500 grasp poses together with the graspability $g_\mathrm{a}$, \mbox{PQ-Net} only loses very few points in terms of AP.
In addition to the datasets, videos of real-world experiments are available at
\url{http://www.bin-picking.ai/en/dataset.html}.
\subsection{Benchmarking in Simulation}
For evaluation, we compare the performance of three approaches
in simulation on two very challenging objects. Each approach gets to observe the same 250 scenes for both the IPABar (see Fig.~\ref{fig:Motivation} left) and IPAUBolt (see Fig.~\ref{fig:Motivation} right), respectively, and executes one grasp per scene to ensure a comparison under the same conditions (the approaches face the exactly same scenarios). Table~\ref{table:grasping_and_precise_placement} reports the success rates for each method.
A robust picking of the IPABar objects from a cluttered bin is challenging because they require a defined picking order. If an occluded object is chosen for grasping, the grasp trial might fail completely or the object might move relative to the gripper which hinders a precise placement of the object.
Therefore, the right object and grasp pose has to be chosen from the highly cluttered scene. This is especially important for friction grasps (force closure) because for a form closure it is unlikely that object moves relative to the gripper.
The IPAUBolt is challenging because it can potentially entangle with other objects in the bin.
In case a wrong object and grasp pose for picking is chosen,
the robot might lift multiple objects resulting in failed grasp because no collision-free placement of the object is possible.
GG-CNN~\cite{GG-CNN, GG-CNN_2}
and other model-free approaches~\cite{Dex-Net_1.0,Dex-Net_2.0,Dex-Net_3.0,JosephRedmon.2015,IanLenz.2015,Google_2016_Learning_Hand-Eye_Coordination,Google_2018_Learning_Hand-Eye_Coordination,Morrison_MultiViewPicking,QT_Opt} focus on generalization performance, use top-down grasps for pick-and-drop of the objects, and cannot solve precise \mbox{pick-and-place} tasks.
To compare our approach with these methods, we use the logged
grasping success labels for training.
Table~\ref{table:grasping_and_precise_placement} reports the success rates for each approach.
Our approach outperforms GG-CNN
because of operating in 6D and being specifically configured to the object.
Furthermore, GG-CNN collides with the bin when attempting to grasp object close to the border due to the limited flexibility in the gripper orientation.
\begin{table*}[hbtp]$ $
\caption{Comparison of the performance of \mbox{PQ-Net} with other approaches and performance evaluation.
}
\label{table:grasping_and_precise_placement}
\begin{center}
\begin{tabular}{|l||c|c|}
\hline
object & IPABar & IPAUBolt \\
\hline
object symmetry based on~\cite{Bregier_PoseDistance,Bregier_SymmetryAwareEvaluation} & finite non trivial & cyclic (order 2) \\
\hline
\hline
GG-CNN~\cite{GG-CNN, GG-CNN_2} success rate for grasping & 0.78 & 0.67 \\
\hline
GG-CNN~\cite{GG-CNN, GG-CNN_2} success rate for grasping without bin & 0.81 & 0.72 \\
\hline
PQ-Net (ours) success rate for grasping & 0.99 & 0.87 \\
\hline
PQ-Net (ours) precision (all grasp poses) & 0.63 & 0.59 \\
\hline
PQ-Net (ours) recall (all grasp poses) & 0.65 & 0.57 \\
\hline
\hline
analytical approach~\cite{FPS_GPC_1} success rate for precise placement & 0.85 & 0.80 \\
\hline
PQ-Net (ours) success rate for precise placement & 0.89 & 0.81 \\
\hline
PQ-Net (ours) success rate OPE & 0.96 & 0.85 \\
\hline
PQ-Net (ours) precision (all grasp poses) & 0.57 & 0.66 \\
\hline
PQ-Net (ours) recall (all grasp poses) & 0.52 & 0.60 \\
\hline
\end{tabular}
\end{center}
\end{table*}
The analytical approach considers the collision-free reachability and kinematical feasibility of the grasp pose and the path only and does not give an estimate on the actual physical outcome of the grasp, e.g., whether the object might move relative to the gripper in the given scenario,
jamming,
or entanglements with other objects in the bin.
With our simulation-driven and learning-based approach, we let our system autonomously learn how to localize and grasp the objects
and transfer the automatically gained experience from the simulation to the real world without any time-consuming object-specific manual tuning.
\iffalse
Since the success labels from the physics simulation are annotated with respect to the ground truth pose, the accuracy of the pose estimation is very important.
Tables~\ref{table:collision_free_reachability} and \ref{table:grasping_and_precise_placement} show the success rates for the OPE based on the metric by Br\'{e}gier et al.~\cite{Bregier_PoseDistance,Bregier_SymmetryAwareEvaluation} for the chosen object
with and without ICP refinement. Due to the high success rates, the robot can reliably grasp the object for precise placement.
\fi
\section{Discussion}
\label{sec:discussion}
In the following, we summarize strengths and discuss limitations of our approach.
\subsection{Strengths}
\mbox{PQ-Net} gets to observe the whole depth image and selects highly robust grasps on a global level because of not looking at local patches of the image only.
Our approach can operate in a closed-loop fashion with 92~fps for the forward pass (for OPE and grasp planning) and is therefore suitable for grasping in non-static environments.
Our graspability metrics allow a gentle removal of the components and avoiding to grasp entangling objects.
Furthermore, our approach can be extended to prioritizing grasp poses which allow a precise placement without re-grasping (e.g., important for objects without symmetries) which allows to reduce cycle times.
\mbox{PQ-Net} provides robust estimates on real-world data independent of the actual 3D technology being used
and does not require any human labeled data or grasping trials on the real-world
system, facilitating scalability. Furthermore, it
automatically configures
for new object geometries using simulation and machine learning for precise
\mbox{pick-and-place} tasks in highly cluttered scenes by providing an object model only.
Our approach properly considers all possible kinds of object symmetries during data generation in the physics simulation and in the loss function for the regression of the angles.
Furthermore, our provided simulated scenes can be used to benchmark further approaches.
\subsection{Limitations}
\mbox{PQ-Net} is a model-based approach and, therefore, does not generalize to unseen objects. Instead, it configures for novel objects without any human intervention.
The approach requires a large dataset to train on, where the process of
data generation (grasp execution) can take long, especially, when increasing the number of grasp poses. Still the process for data generation can run much faster than real time and can easily be parallelized across multiple machines.
Furthermore, our method cannot do pre-grasp manipulations on the objects to change their poses
in order to more robustly grasp them.
While \mbox{PQ-Net} can avoid entangled object situations,
we do not propose a solution to unhook very complex object geometries, for which no general solution has been proposed so far~\cite{Potentially_Tangled_Objects,paper_MAM}.
\section{Conclusion and Future Work}
In this paper, we proposed a novel learning-based approach
for grasping in highly cluttered scenes and precise object placement based on depth images.
Our approach outputs the 6D object poses together with a graspability and quality estimate for each automatically generated grasp pose for multiple objects simultaneously in a single forward pass in a joint framework (running at 92~fps).
All densely discretized and automatically generated grasp poses are executed in a physics simulation and the gained experience is transferred from simulation
to the real world.
Our approach outperforms model-free approaches in terms of grasping success rate
and does contrary to analytical approaches not require any human involvement (automatic configuration).
We demonstrate that our approach can be used for precise real-world robotic \mbox{pick-and-place} tasks, although being entirely trained on simulated data.
In future work, we plan to extend the approach to mixed bins
and study how to reduce the time for data generation and training to allow a faster deployment of our solution. Furthermore, we want to study whether the generated data (grasping trials) can be used for model-free robotic grasping approaches in 6D.
\section*{Acknowledgment}
This work was partially supported by the
Federal Ministry of Education and Research (Deep Picking -- Grant No. 01IS20005C) and
the Ministry of Economic Affairs of the state Baden-W\"urttemberg (Center for Cognitive Robotics –- Grant No. 017-180004 and Center for Cyber Cognitive Intelligence (CCI) -- Grant No. 017-192996). We would like to thank our colleagues for helpful discussions
and D. Unruh for the support with experiments.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:introduction}
Underground human activity, including fluid injection in rocks and mining operations, can cause microseismic events \citep{Majer07, Ellsworth13}. The monitoring of both human-induced and natural microseismicity is critical for understanding seismic hazard \citep[][and references therein]{Brueckl08, Shapiro10, Mukuhira16, Das17}.
Accurate seismic event locations in space and in time are of paramount importance for reliable seismic monitoring efforts, and are mainly obtained from seismograms recorded on land and/or at the seafloor.
Various methods for locating seismic events are available in the literature, dating back to the work of \citet{Geiger10}, and up to today \citep[see e.g.][and references therein, for a recent review]{Vasco19}. One of the most common approaches relies on using the Eikonal equation to determine the theoretical travel time of first seismic arrivals \citep[see e.g.][]{Noack17, Smith20}, which is compared with real travel times through a direct grid search or more sophisticated inverse modelling techniques \citep{Wuestfeld18}. More accurate source location estimations can be obtained exploiting methods that use the full waveform, even though they generally require heavier computational resources (\citealt{Song11, Li13, Angus14, Li16, Willacy19, Vasco19}; see also \citealt{Li20} for a recent review of waveform-based inversion methods).
A Bayesian approach can also be adopted to solve the location inverse problem \citep{Lomax00, Tarantola05, Staehler14, Staehler16, Pugh16}. In this framework, a posterior distribution of the model parameters (i.e., the event's location and/or the moment tensor) is estimated and used to determine the optimal model parameters and their associated uncertainty.
Markov Chain Monte Carlo \citep[MCMC; see e.g.][for a review]{Craiu14} and nested sampling \citep{Skilling06} techniques are among those employed to sample the posterior distribution. However, these approaches become prohibitive when dealing with a high number of parameters, or when the forward model is computationally expensive to simulate \citep[see e.g.][]{Rajaratnam15, Conrad16, Alsing18}. For these reasons, being able to cheaply and accurately simulate the theoretical waveforms of microseismic events given their location has become paramount in recent years. The forward modelling of microseismic events typically involves solving the elastic wave equation in a wide frequency range given a 3-D heterogeneous density and velocity model for the propagating medium \citep{Das17}, which can be prohibitively expensive.
Machine learning generative models have gained considerable attention in recent years, with applications to many fields ranging from computer vision \citep{Goodfellow14, Gulrajani16} to astrophysics \citep{Auld07, Auld08, SpurioMancini21}, as well as climate science, nuclear physics and drug selection \citep[see e.g.][]{Kasim20, Chenthamarakshan20}. These advances have been enabled by both an increased accessibility to computational resources, as well as by a significant growth in the amount of available data.
In seismology, machine learning techniques have been successfully applied to a wide range of problems \citep[see e.g.][for a recent review]{Bergen19}. For instance, \citet{Zheng18} used a recurrent neural network to pick the arrival times of microseismic (or acoustic) events, while \citet{Zhu18}, \citet{Ross18} and \citet{Zhu19} trained convolutional neural networks (CNNs) to measure P- and S-wave arrival times and determine high-level features with high precision, often outperforming the measurements performed manually or semi-automatically by expert seismologists. \citet{Mousavi20} further proposed a deep-learning model for simultaneous earthquake detection and phase picking based on an attention mechanism, while \citet{Li22} trained multiple CNNs both to detect whether there is an acoustic emission data event in some recorded data, and to pick the arrival time of the P-wave.
Looking in particular at generative models, \citet{Das17} developed an optimised approach to simulate microseismic event propagation that, for each event location and given a physical model for the propagating medium, produces the corresponding seismogram in $\mathcal{O}(1 \ \rm{h})$ using a Tesla graphics processing unit (GPU) and the software \textsc{k-wave} \citep{Treeby14}. Subsequently, \citet{Das18} and \citet{ASM20} (D18 and SM20 hereafter, respectively) showed the limitations of this direct approach, and presented an alternative whereby the mapping is learnt using machine learning techniques. In particular, D18 showed how Gaussian Processes \citep[GPs,][]{Rasmussen05} can be used to learn an accurate surrogate model, while SM20 demonstrated the effectiveness of a variety of machine learning algorithms as emulators, and showed how their surrogate model yields an accurate estimate of the posterior distribution of an event's location in a fraction of the time required by the D18 method.
D18 and SM20, however, only applied their methodologies to microseismic events with an isotropic source mechanism. It is well known that any source mechanism can be mathematically decomposed into three components: isotropic (ISO), double couple (DC), and compensated linear vector dipole (CLVD) \citep[see e.g.][]{Knopoff70, Vavrycuk01, Vavrycuk05, Vavricuk15}. The pure ISO source is associated with implosive or explosive forces, while the pure DC source is caused by shear faulting.
In this paper, we present an approach that aims at learning the direct mapping between event locations and seismograms for any microseismic source type (ISO, DC and CLVD). We show that it is sufficient to consider the power spectrum of the recorded pressure waves in order to distinguish different seismograms (as already investigated e.g. in \citealt{Pratt99a, Pratt99b, Tao13,Jakobsen15}), and we train a simple machine learning algorithm to learn this mapping efficiently. Moreover, we demonstrate how our method allows the accurate inference of the posterior distribution of the coordinates of a single source in $\mathcal{O}(0.1 \ \rm{h})$ on a commercial laptop, thus paving the way for fast and computationally cheap joint determinations of event locations and seismic moment tensors.
Finally, we show how we can use the trained emulators to identify the source mechanism through Bayesian evidence estimation, thus demonstrating the versatility of our Bayesian approach.
Regarding the structure of this paper, in Sect.~\ref{sec:data} we describe the data we use in this work. In Sect.~\ref{sec:pgi} we explain our inversion approach, describe what preprocessing steps we perform and recall the details of the generative method we employ. We show its performance at both training and inference time in Sect.~\ref{sec:results}, where we also compare our approach to standard arrival time analysis \citep{Lomax09} and include an experiment with real noise from Groningen field data \citep{Smith19thesis}, thus showing the robustness of our method. Finally, in Sect.~\ref{sec:conclusions} we discuss our results and provide an outlook on possible extensions of this work.
\section{Data}
\label{sec:data}
\begin{figure*}
\centering
\includegraphics[scale=0.07]{plots/p-velocity_new.png}
\includegraphics[scale=0.195]{plots/s-velocity_new.png}
\includegraphics[scale=0.07]{plots/density_new.png}
\caption{P-wave velocity ($V_p$), S-wave velocity ($V_s$) and density ($\rho$) models of the simulated domain we consider in this work. The models are specified as 3-D grids of voxels, with size 81 $\times$ 81 $\times$ 301 points, corresponding to a real geological model of size 1 km $\times$ 1 km $\times$ 3 km. We observe that our model has a layered structure, with variation along the vertical dimension more marked than along horizontal planes. The plots were adapted from figure 1 in \citet{ASM20}.}
\label{fig:velocity_model}
\end{figure*}
In this work, we consider the same data framework as D18 and SM20, starting from 3-D heterogeneous density and velocity models for the propagating medium, which we show in Fig.~\ref{fig:velocity_model}. The model, which is discretised on a 3-D grid of voxels, specifies the values of the density $\rho$ of the propagating medium, as well as the propagation velocities for P- and S-waves ($V_p$, $V_s$). We assume that sensors are placed at the seabed, and that they record both pressure and three-component particle velocity of the propagating medium (even though we will use only the former, as we explain later on). As anticipated, our aim is to apply our method to any source mechanism, so we will consider a more general generation procedure than previous work. Unlike D18 and SM20, who only considered isotropic sources, we take the microseismic moment tensor to be one of three types, which we denote as $\textbf{M}_{\rm{ISO}}$, $\textbf{M}_{\rm{DC}}$ and $\textbf{M}_{\rm{CLVD}}$. Following \citet{Vavrycuk05} and \citet{Li15}, we define these quantities as:
\begin{align}
&\textbf{M}_{\rm{ISO}}=
\begin{bmatrix}
M_{11} & 0 & 0 \\
0 & M_{22} & 0 \\
0 & 0 & M_{33}
\end{bmatrix} \label{eq:M_ISO}, \\
&\textbf{M}_{\rm{DC}}=
\begin{bmatrix}
0 & M_{12} & 0 \\
M_{21} & 0 & 0 \\
0 & 0 & 0
\end{bmatrix}, \label{eq:M_DC} \\
&\textbf{M}_{\rm{CLVD}}=
\begin{bmatrix}
M_{11} & 0 & 0 \\
0 & M_{22} & 0 \\
0 & 0 & -2M_{33}
\end{bmatrix},
\label{eq:M_CLVD}
\end{align}
where each $M_{ij}$ represents a different couple of forces. We additionally assume $M_{11}=M_{22}=M_{33}=M_{12}=M_{21}=1$ MPa, which is a realistic assumption following \citet{Collettini02}; we anticipate that while in this work we fix the $M_{ij}$ coefficients, we will explore the joint determination of moment tensor components and event coordinates in future work. In practice, isotropic (ISO) events are characterised by a single (explosive or implosive) P-wave, while double couple (DC) events are linked to shear stress and are characterised by both a P- and S-wave, with comparable amplitudes. Similarly to isotropic events, compensated linear vector dipole (CLVD) events display an often dominant principal wave, whose amplitude is however much smaller than the isotropic one, and a non-negligible S-wave.
\begin{figure}
\centering
\includegraphics[width=\columnwidth, keepaspectratio]{plots/receivers_positions.pdf}
\caption{Projection of the positions of the 23 receivers on the $x$-$y$ plane; $z=2.43$ km corresponds to the seabed, where all sensors lie. Each receiver records the acoustic pressure wave and particle velocity generated by a microseismic event below the seabed. The red crosses indicate all 23 receivers, while green circles and blue squares (9 and 5 receivers, respectively) refer to subsets of receivers we used to test the robustness of our method, as shown in Sect.~\ref{sec:inference_results}.}
\label{fig:rec_pos}
\end{figure}
We generate forward simulations using the GPU implementation of \citet{Das17}, employing different types of GPUs with memory size ranging from 5 GB to 11 GB. The precise GPU cards used to generate the data are Tesla K20, Tesla C2075, GeForce RTX 2080 Ti and GeForce GTX 1080 Ti, which carry a different number of CUDA cores each, ranging from 500 to 5000. We observe that the speed of the generation scales linearly with the number of cores available; however, given that we produce the data on a shared cluster, we cannot always choose which card to employ. Note that the software with which we work is optimised for GPUs, and therefore cannot be expected to scale similarly when running on Central Processing Units (CPUs) or Tensor Processing Units (TPUs).
For each source type, we produce 10\,000 events corresponding to different source locations, which are randomly sampled using Latin Hypercube Sampling on a 3-D grid of 81 $\times$ 81 $\times$ 301 points, corresponding to a real geological model of size 1 km $\times$ 1 km $\times$ 3 km. The total time to generate these data using the hardware specified above is about 150 h for each source type.
We consider 23 receivers in total, whose position is depicted in Fig.~\ref{fig:rec_pos}, even though we stress that we will not focus on finding the optimal geometry of the sensors in this work. Each receiver records each of the 10\,000 events independently, thus producing a set of 10\,000 waveforms for each source type at each receiver. In Sect.~\ref{sec:inference_results} we will present the results of a full analysis of the dependence of the posterior distribution on the number of training points, the number of receivers and the noise level, in order to demonstrate the robustness of our approach.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{plots/data_comparison.pdf}
\caption{\textit{Left column:} Example acoustic pressure wave for each different type of moment tensor: isotropic (ISO), double couple (DC) and compensated linear vector dipole (CLVD). These seismograms correspond to a source location of $(x,y,z)=(0.55 \ \rm{km},0.73 \ \rm{km},1.8 \ \rm{km})$ as recorded by a receiver in $(x,y,z)=(0.13 \ \rm{km},0.38 \ \rm{km},2.43 \ \rm{km})$. The seismograms' amplitude is measured in arbitrary units of pressure. Note the different scales for each source mechanism. The vertical dashed black line in the top panel indicates a cut we perform for the isotropic sources only, based on the left panel of Fig.~\ref{fig:cuts}. \textit{Right column:} The corresponding power spectra, calculated as described in Sect.~\ref{sec:preprocessing}. No noise is added when training the emulator, while some noise is introduced when doing inference, as described in Sec.~\ref{sec:inference}. The vertical dashed lines indicate a frequency cut we perform to further reduce the number of features and to be robust to noise, based on the right panel of Fig.~\ref{fig:cuts}.}
\label{fig:data_comparison}
\end{figure*}
The time interval (sampling rate) for the solution of the elastic wave equation is $0.5$ ms (2 kHz), and the total length of each seismic event is $5$ s. After generation, all seismograms are downsampled to a time resolution of $2.5$ ms (400 Hz) to reduce computational storage. In this way, each seismic trace is ultimately a time series composed of $N_t = 2001$ time samples: we show an example for each source mechanism in the left column of Fig.~\ref{fig:data_comparison}.
Finally, note that we consider the seismograms to be noiseless at training time, while some noise is added to the simulated recorded event when performing inference on the coordinates' posterior distribution, as we are working in a Bayesian framework, detailed in Sect.~\ref{sec:inference}. We will discuss and show the effect of the noise level in Sect.~\ref{sec:preprocessing} and Sect.~\ref{sec:inference_results}, respectively.
\section{Inversion approach}
\label{sec:pgi}
Our goal is to perform Bayesian inference on the source location of a microseismic event. In order to do so, we will employ a neural network (NN) as the forward model to parametrise the mapping between the coordinates and the principal components of the power spectra of the seismograms. Our inversion approach is summarised in Fig.~\ref{fig:workflow}.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{plots/workflow.pdf}
\caption{Workflow of our inversion approach. The observed event, indicated in coral in the top left corner, is preprocessed according to the steps described in Sect.~\ref{sec:preprocessing} in order to obtain a logarithmic power spectrum. Our goal is to obtain the posterior distribution of the coordinates $\theta$ of the event's location beneath the surface, as detailed in Sect.~\ref{sec:inference}. In order to sample from the posterior distribution, we specify a data likelihood, and use a generative model $g(\boldsymbol{\theta})$ to accelerate the evaluation of the likelihood. As our generative model, which we describe in Sect.~\ref{sec:gen_model}, we employ a feedforward neural network that maps source coordinates to the principal components of the logarithmic power spectra. The neural network sketch has been drawn using \textsc{NN-SVG} \citep{LeNail19}. Note that the form of Bayes' theorem in this picture is a simplified version of Eq.~\ref{eq:bayes}.}
\label{fig:workflow}
\end{figure*}
\subsection{Inference}
\label{sec:inference}
We recall here the basic assumptions of our Bayesian analysis. Given a set of parameters $\boldsymbol{\theta}$ (the coordinates, in our case), their posterior distribution given some data $\boldsymbol{D}$ and some hypothesis $\mathcal{H}$ can be written using Bayes' theorem \citep[see e.g.][]{Bishop06}:
\begin{align}
\label{eq:bayes}
\mathrm{Pr} \left( \boldsymbol{\theta} | \boldsymbol{D}, \mathcal{H} \right) = \frac{ \mathrm{Pr} \left( \boldsymbol{D} | \boldsymbol{\theta}, \mathcal{H}\right) \mathrm{Pr} \left( \boldsymbol{\theta} | \mathcal{H} \right) }{ \mathrm{Pr} \left( \boldsymbol{D} | \mathcal{H}\right) } \ ,
\end{align}
which expresses the posterior distribution $\mathrm{Pr} \left( \boldsymbol{\theta} | \boldsymbol{D}, \mathcal{H} \right)$ as the product of the likelihood $\mathrm{Pr} \left( \boldsymbol{D} | \boldsymbol{\theta}, \mathcal{H} \right)$ and the parameters' prior $\mathrm{Pr} \left( \boldsymbol{\theta} | \mathcal{H}\right)$, divided by the evidence $\mathrm{Pr} \left( \boldsymbol{D} | \mathcal{H}\right)$. For the purposes of inference, we will ignore this last term, as it is just a normalisation factor independent of $\boldsymbol{\theta}$; however, in Sect.~\ref{sec:evidence} we show how the evidence can be used to perform model selection, which is another advantage of working in a Bayesian framework. In order to sample the posterior distribution of the source coordinates, we employ nested sampling \citep{Skilling06}, as implemented in \textsc{PyMultiNest}\footnote{\url{https://github.com/JohannesBuchner/PyMultiNest}}~\citep{Buchner14}, the Python interface to \textsc{MultiNest} \citep{Feroz08}. We choose nested sampling over Metropolis-Hastings sampling or other MCMC techniques as it generally converges faster \citep{Allison14} and provides an estimate of the evidence. For the prior, we assume a uniform distribution in the range of the physical model ($[0, 1] \times [0, 1] \times [0, 2.43]$, with units in kilometres).
To perform Bayesian inference on a given seismogram, we first randomly choose an event's coordinates from the test set. For this set of coordinates, we simulate the observation of a microseismic event for each source mechanism, and generate the noiseless trace as it would be recorded by each of the 23 receivers. We add random Gaussian noise to each component of the noiseless trace. We define the signal-to-noise ratio (SNR) as \citep{Li18, Zhang20}:
\begin{align}
\label{eq:snr}
\mathrm{SNR} = 10\log_{10}{\frac{\sum^{\rm{N}}_i \sum^{\rm{N_{keep}}}_j s^2_{ij}}{\sum^{\rm{N}}_i \sum^{\rm{N_{keep}}}_j (s_{ij} - \tilde{s}_{ij})^2}} \ ,
\end{align}
where $s_{ij}$ refers to the $j$-th sample of the $i$-th trace and $\tilde{s}_{ij}$ to the corresponding noisy trace, $\rm{N}=8000$ is the number of training data and $\rm{N_{keep}}=2001$ (1000 in the ISO case) is the number of time samples. As we explain in Sect.~\ref{sec:gen_model}, we consider 8000 events out of the total 10\,000 for training, and reserve the remainder for validation and testing purposes. Following \citet{Li18}, we set SNR$=33$ dB, which corresponds to a standard deviation of the Gaussian noise of $\sigma$ = 10.0, 0.3 and 3.5 for ISO, DC and CLVD, respectively, in the same arbitrary units as the seismograms' amplitude. We show examples of noiseless and noisy signals in Fig.~\ref{fig:compare_all_noises}.
\begin{figure}
\centering
\includegraphics[width=\columnwidth, keepaspectratio]{plots/compare_noises.pdf}
\caption{Comparison of the signal without (solid blue) and with (dashed coral) noise for each source mechanism: isotropic (ISO), double couple (DC) and compensated linear vector dipole (CLVD). As explained in Sect.~\ref{sec:inference}, after training the emulator on the noiseless traces, we add Gaussian noise to the observed signal to infer its coordinates. In this case, the seismogram corresponds to event 3 in Table~\ref{tab:inference} as recorded by a receiver in $(x,y,z)=(0.13 \ \rm{km},0.38 \ \rm{km},2.43 \ \rm{km})$, and the signal-to-noise ratio (SNR) is 33 dB; we explore higher and lower SNR values in Sect.~\ref{sec:inference_results} and Fig.~\ref{fig:compare_noise}. The vertical dashed black line in the top panel indicates a cut we perform for the isotropic sources only, based on the left panel of Fig.~\ref{fig:cuts}.}
\label{fig:compare_all_noises}
\end{figure}
We note that the choice of Gaussian noise has a quantifiable consequence on the power spectra of the signals, as white noise has an expected constant power in Fourier space \citep[see e.g.][]{Haykin01, Papoulis02}. This is reflected on the right-hand side of Fig.~\ref{fig:cuts}, where the mean noisy signal is shifted up with respect to the noiseless signals due to the noise addition. We argue that in general any information about the noise power (even if more complicated than Gaussian noise) can be easily accounted for when preprocessing the data: knowing the behaviour of the noise signal in the frequency domain allows one to fully take into account its effect on the relevant signal. For this reason, we decide to transform our data into the Fourier space, as we describe in Sect.~\ref{sec:preprocessing}. We will show in Sect.~\ref{sec:inference_results} that this choice makes our proposed approach robust to noise; moreover, our approach lends itself to the extension to coloured noise, which is most likely in realistic seismic data, especially for events with a low SNR \citep{Liu17}.
The noisy seismogram is further preprocessed as described in Sect.~\ref{sec:preprocessing}. At each likelihood evaluation of \textsc{PyMultiNest}, the proposed coordinates are mapped to the predicted preprocessed seismograms by means of the generative model $g( \boldsymbol{\theta})$ described in Sect.~\ref{sec:gen_model}: by evaluating the likelihood in multiple points of the prior space, \textsc{PyMultiNest} can sample from the posterior distribution of the coordinates, thus yielding the required credibility regions in parameter space. Note that, similarly to D18 and SM20, we assume a Gaussian likelihood, i.e.\ we write:
\begin{align}
\label{eq:glike}
\mathrm{Pr} \left( \boldsymbol{D} | \boldsymbol{\theta}, \mathcal{H}\right) \propto \exp \left( -\frac{1}{2} \left(\boldsymbol{D}-g(\boldsymbol{\theta})\right)^T \boldsymbol{C}^{-1} \left(\boldsymbol{D}-g(\boldsymbol{\theta})\right) \right) \ ,
\end{align}
where $\boldsymbol{C}$ indicates the covariance matrix of the preprocessed seismograms, estimated from the training data.
We note that our choice of a Gaussian likelihood comes without loss of generality as our method can be easily extended to more complicated likelihood models. It is worth stressing that adding Gaussian noise to the seismograms does not necessarily imply that the distribution of the preprocessed seismograms will also be Gaussian; however, we verified experimentally that the distribution of each preprocessed seismogram component is unimodal and symmetric, thus supporting our assumption.
\subsection{Preprocessing}
\label{sec:preprocessing}
Learning a mapping between coordinates and seismograms directly would be hard for at least two reasons. First, each signal has features with different amplitudes: this means that e.g.\ a neural network (which will be described in detail in Sect.~\ref{sec:gen_model}) would likely just focus on the main peak and ignore the other components, thus losing useful information for the source location purpose. Secondly, given the complexity of the seismograms and the high number of features, the amount of data required to train an accurate emulator without overfitting would be at least an order of magnitude higher than what we consider in this work \citep[see e.g.][and references therein]{Bishop06, Zhu15}.
In this sense, we $\textit{have to}$ preprocess the seismograms in order to extract only the relevant information that is needed to locate an event, while discarding all the noisy or redundant features of the signal. Both D18 and SM20 showed the importance of preprocessing, employing GPs in order to select only the components of each seismogram that are essential for inference. However, their methods fail on more complicated sources like the ones we consider in this work. While it could be argued that employing more training data could improve the results, it is also well-known that GPs do not scale well with the number of training points \citep{Liu18}, thus it is likely that the D18 method would struggle to generalise to more complicated source mechanisms. Additionally, the method proposed in SM20 is applied directly to the complicated seismic traces like the ones in the left panel of Fig.~\ref{fig:data_comparison}, thus making it more difficult for any algorithm to capture the most useful features for event location. Therefore, we follow a different preprocessing procedure, based on translating the data to the Fourier domain, and we outline the steps in the next paragraphs.
The left panel of Fig.~\ref{fig:cuts} shows the mean and standard deviation of all seismograms in the training set (8000 seismic traces): based on these distributions, we keep all samples of the DC and CLVD signals, and only keep the first half of the ISO traces since they sharply vanish after about one third of the trace. In other words, we keep only $\rm{N_{keep}}=1000$ time samples for the ISO traces, and $\rm{N_{keep}}=\rm{N_t}=2001$ time samples for the other mechanisms.
The next preprocessing step we implement is applying the one-dimensional discrete Fourier Transform \citep{Cooley65} to each seismogram, using the version of \textsc{NumPy}\footnote{\url{https://numpy.org/doc/stable/reference/routines.fft.html}}. Since the amplitudes at each time sample are real numbers, the Fourier Transform returns $(\floor{\rm{N_{keep}}/2}+1)$ frequency components: this means that for DC and CLVD sources we are left with 1001 components (501 in the ISO case) in the Fourier domain. We then take the square of the absolute value of these complex numbers: this is usually referred to as a power spectrum. In the right column of Fig.~\ref{fig:data_comparison} we report the power spectra corresponding to the individual seismograms in the left column of the same figure. We further take the decimal logarithm of the power spectra at each frequency value, and refer to it as logarithmic power spectra in the rest of the paper.
As anticipated in Sect.~\ref{sec:inference}, we shall add some noise to the observed seismogram whose source location coordinates will be inferred. In the right panel of Fig.~\ref{fig:cuts} we show the mean power spectra for each source mechanism with and without noise.
We calculate the ratio between the noiseless and the noisy signals, and filter out the frequencies for which this ratio is less than 0.99. We experimented with different thresholds, and chose 0.99 as a good balance between retaining enough features to locate a seismogram and being insensitive to noise. In other words, we additionally cut each power spectrum in the ranges $[1$ Hz, $62.4$ Hz$]$, $[0$ Hz, $52.2$ Hz$]$ and $[0$ Hz, $45.8$ Hz$]$ for ISO, DC and CLVD, respectively, to keep the parts of each signal that are less affected by noise. We observe that translating the seismograms to the Fourier domain has allowed us to obtain smoother signals, as well as to reduce the number of features by a factor of $10$. Moreover, this allows our proposed method to be robust to noise: any effect due to noise can be translated into some information of the noise power, and hence readily accounted for in the analysis, e.g.\ by selecting a band-limited frequency signal as we discussed in Sect.~\ref{sec:inference}.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{plots/error_bars.pdf}
\caption{\textit{Left column:} Mean (blue line) and standard deviation (grey area) of all the seismic traces in the training set, for each source mechanism: isotropic (ISO), double couple (DC) and compensated linear vector dipole (CLVD). The seismograms' amplitude is measured in arbitrary units of pressure. The training set is made of 8000 traces for each source mechanism. We cut the isotropic sources at $2.5$ s, as indicated by the vertical dashed black line. \textit{Right column:} Noiseless (solid blue) and noisy (dashed coral) mean of the corresponding power spectra, calculated as described in Sect.~\ref{sec:preprocessing}. We consider a signal-to-noise ratio of 33 dB, as described in Sect.~\ref{sec:inference}. We filter the power spectra between the vertical dashed lines selecting only the frequencies where the ratio between the noiseless and noisy signals is more than 99$\%$.}
\label{fig:cuts}
\end{figure*}
To further reduce the number of features, we apply principal component analysis (PCA). PCA is a standard linear compression technique where the data is projected along the eigenvectors of the data covariance matrix. Considering only the components that carry more variance (the so-called `principal components', corresponding to the largest eigenvalues), it is possible to reduce the number of features while maintaining the relevant information for inference \citep{Bishop06}. We fit PCA to the training data, and use it to compress the whole dataset. After applying PCA to the logarithmic power spectra, we retain 10 principal components for each signal when training the generative model; note that inference is done at the level of the logarithmic power spectra instead. We verified experimentally that varying the number of retained PCA components does not impact the final results significantly.
\subsection{Generative model}
\label{sec:gen_model}
As our generative model $g(\boldsymbol{\theta})$, we choose to employ a feedforward neural network (NN). A feedforward NN is a set of subsequent layers, each made of a certain number of neurons, that allow for the parametrisation of any measurable function between finite dimensional spaces \citep{Hornik89}. Each neuron is associated to a weight, and each layer is additionally associated to a bias (i.e.,\ an offset): weights and biases constitute the parameters that we wish to learn. Additionally, activation functions can be introduced after each layer to model non-linear mappings. Training the neural network consists of feeding some data through all the layers, and then updating the value of the parameters in order to optimise a chosen loss function.
We employ a neural network made of three layers with 256 neurons each, to provide enough flexibility to the parametrisation without consuming too much memory. We set Leaky ReLU \citep{Maas13} as the activation function for all layers except the last one, where we keep a linear activation function. We recall here that Leaky ReLU acts on the output of each layer $\mathbf{O}$ as follows:
\begin{align}
\rm{Leaky ReLU}(\mathbf{O}) = \bigg \{\begin{array}{lr}
\mathbf{O} & \mathrm{if} \, \mathbf{O} > 0 \\
\alpha \mathbf{O} & \mathrm{otherwise} \\
\end{array} \ ,
\end{align}
where we set the hyperparameter $\alpha = 0.3$; Leaky ReLU is usually preferred over the standard ReLU (Rectified Linear Unit) because of non-vanishing gradients \citep{Kolen01}. We also experimented with the ELU (Exponential Linear Unit, \citealt{Clevert16}) activation function, and found no significant improvements in the overall results with respect to using Leaky ReLU. We choose the Mean Squared Error (MSE) between the network output and the principal components of the training data as our loss function to minimise.
For each source mechanism and each receiver, we train the emulator using 8000 traces; we reserve 1000 seismograms for validation purposes and 1000 seismograms for testing purposes. We remark that we train a single neural network for each receiver: for a given source mechanism and underground location, the training data for each receiver includes the waveforms as observed by that particular station. In this way, we are able to include the information on the receiver's position in our generative model.
To train the neural network, we use the Adam optimiser \citep{Kingma14} with default parameters; moreover, we choose a learning rate of $0.001$ and a batch size of $256$: the former controls the step size of the parameters' update, while the latter indicates the number of training points that are fed through the network at each iteration. We additionally set a patience of $50$ to early-stop \citep{Yao07} based on the validation loss: this means that if the loss calculated on the validation set has not decreased in the last 50 epochs, we stop training and take the model corresponding to the minimum validation loss as the best model, as we interpret the model to have achieved its minimum error on unseen data\footnote{We also tested a dynamic learning rate decreasing by a factor of ten every time the validation loss did not decrease for 50 epochs, but found no significant improvement over a constant learning rate of 0.001, which we therefore chose for our analysis.}. We report the typical trend for the training and validation loss curves in Fig.~1 in the supplementary material.
The test set is used to randomly sample events on which we perform our Bayesian inversion analysis, as described in Sect.~\ref{sec:inference}. We also explore the behaviour of the posterior distribution as a function of the number of training events, number of receivers and noise scale, which we show in Sect.~\ref{sec:inference_results}. However, we do not perform a full grid search amongst the hyperparameters (e.g.\ number of layers, number of neurons, activation function and learning rate), as we observe the results are not significantly affected by them; we defer a more complete grid search to future work.
\begin{figure*}%
\subfloat[][ISO]{\includegraphics[scale=0.2]{plots/relative_difference_iso.pdf}}
\subfloat[][DC]{\includegraphics[scale=0.2]{plots/relative_difference_dc.pdf}}
\subfloat[][CLVD]{\includegraphics[scale=0.2]{plots/relative_difference_clvd.pdf}}
\caption{Accuracy of the generative model described in Sect.~\ref{sec:gen_model}, for isotropic (ISO), double couple (DC) and compensated linear vector dipole (CLVD). The dark red, red and salmon areas enclose the 68, 95 and 99 percentiles of the absolute value of the relative error between the predictions of the model and the target waveforms from the test set.}
\label{fig:gen_model_accuracy}
\end{figure*}
\subsection{Model selection}
\label{sec:evidence}
As detailed in Sect.~\ref{sec:inference}, we ignore the denominator in Eq.~\ref{eq:bayes} when inferring the coordinates of a microseismic event. However, we can use the evidence $\mathrm{Pr} \left( \boldsymbol{D} | \mathcal{H}\right)$ to perform model selection, thus showing another advantage of our proposed Bayesian approach \citep[see e.g.][]{Knuth15}. The quantity $\mathrm{Pr} \left( \boldsymbol{D} | \mathcal{H}\right)$ can be interpreted as the likelihood of a given signal under a certain hypothesis, with the constraint that the hypotheses form a set of $n_{\rm{hyp}}$ pairwise disjoint events whose union is the entire possibility space - i.e.\ $\mathrm{Pr} \left( \mathcal{H}_i \cap \mathcal{H}_j \right)$ = 0 for $i \neq j, \ \forall i, j = \{ 1, \dots, n_{\rm{hyp}} \}$, and $\sum_{i=1}^{n_{\rm{hyp}}} \mathrm{Pr} \left( \mathcal{H}_i \right) = 1$.
To compare two hypotheses, and thus perform model selection, we define the Bayes factor BF as:
\begin{align}
\label{eq:bayes_f}
\mathrm{BF} \equiv \frac{\mathrm{Pr} \left( \boldsymbol{D} | \mathcal{H}_i\right)}{\mathrm{Pr} \left( \boldsymbol{D} | \mathcal{H}_j\right)} = \frac{\mathrm{Pr} \left( \mathcal{H}_i | \boldsymbol{D}\right) \mathrm{Pr} \left( \mathcal{H}_j \right) }{\mathrm{Pr} \left( \mathcal{H}_j | \boldsymbol{D}\right) \mathrm{Pr} \left( \mathcal{H}_i \right)} \ ,
\end{align}
where the second equality has been obtained using Bayes' theorem, $\mathrm{Pr} \left( \mathcal{H}_i \right)$ is the prior distribution of the hypothesis $\mathcal{H}_i$, and $\mathrm{Pr} \left( \mathcal{H}_i|\boldsymbol{D}\right)$ is the posterior distribution of hypothesis $\mathcal{H}_i$ given $\boldsymbol{D} $. If the hypotheses are equiprobable \textit{a priori}, we can write $\mathrm{Pr} \left( \mathcal{H}_i \right) = \mathrm{Pr} \left( \mathcal{H}_j \right)$, which allows us to express the Bayes factor as the ratio of the posterior distribution of one hypothesis over the other. Hence, if the Bayes factor as defined in Eq.~\ref{eq:bayes_f} is greater than 1, we can interpret it as hypothesis $\mathcal{H}_i$ being more favoured than hypothesis $\mathcal{H}_j$ under the observed data $\boldsymbol{D}$ \citep{Knuth15}.
Translating this into practice, after training the emulators, given an observed signal $\boldsymbol{D}$ as described in Sect.~\ref{sec:inference}, we can compare the three following equiprobable hypotheses: the source mechanism is isotropic ($\mathcal{H}_{\rm{ISO}}$), the source mechanism is double couple ($\mathcal{H}_{\rm{DC}}$), or the source mechanism is compensated linear vector dipole ($\mathcal{H}_{\rm{CLVD}}$). The advantage of using nested sampling is that the evidence is calculated while sampling the posterior distribution. Consequently, by feeding the observation $\boldsymbol{D}$ to each emulator it is straightforward to obtain the evidences $\mathrm{Pr} \left( \boldsymbol{D} | \mathcal{H}_{\rm{ISO}}\right)$, $\mathrm{Pr} \left( \boldsymbol{D} | \mathcal{H}_{\rm{DC}}\right)$ and $\mathrm{Pr} \left( \boldsymbol{D} | \mathcal{H}_{\rm{CLVD}}\right)$.
By looking at the hypothesis that maximises the evidence, we can select the model that best describes the given observation, thus identifying the source type for a given observation. We show the results in Sect.~\ref{sec:model_sel_res} and Table~\ref{tab:evidence}.
\section{Results}
\label{sec:results}
\setlength{\tabcolsep}{6pt}
\subsection{Speed performance}
\label{sec:speed}
We first report on the speed performance of our method. We recall here that if we were to solve the elastic wave equation at each likelihood evaluation, inference would be severely compromised, as a single event's source inversion would take thousands of hours on a High Performance Computing (HPC) cluster, if at all possible. In contrast, our method requires only $\sim10^4$ simulations to be produced once, and the emulator to be trained once - an overhead of $\mathcal{O}(100 \ \rm{h})$ and $\mathcal{O}(1 \ \rm{h})$, respectively - and then it allows for the complete source inversion of any event in $\mathcal{O}(0.1 \ \rm{h})$ on a commercial laptop. In other words, most of the time taken by our approach is spent training the emulator, which needs to be done only once, provided the density and velocity models remain unchanged\footnote{In general, the stability of a given velocity model is not well known - see e.g.\ \citet{Thornton13, Usher13, Gesret13, Gesret14, Das18} and references therein for a discussion on the uncertainties of velocity models, and their consequences on location errors.}; after the training is complete, performing inference on a given recorded seismogram takes less than 10 minutes on a commercial laptop. As a reference, conventional full-waveform inversion techniques can take several hours on CPUs, and a comparable amount of time to our method when running on GPUs \citep[see e.g.][]{AbreoCarrillo15}.
\subsection{Generative model accuracy}
\label{sec:gen_model_res}
In this section, we briefly look at the performance of the generative model that we described in Sect.~\ref{sec:gen_model}. We consider the 1000 events in the test set, and measure the relative difference between the prediction of our model and the target waveforms. We show the 68, 95 and 99 percentiles of the relative difference in Fig.~\ref{fig:gen_model_accuracy} for all source mechanisms. We observe that there are some significant discrepancies, especially at low frequency for the DC and CLVD, which however we attribute to low values in the logarithmic power spectra, and which do not seem to compromise inference, as we show in the next section.
\subsection{Inference results}
\label{sec:inference_results}
\begin{table*}
\centering
\caption{Prior range and marginalised mean and 68 percent credibility intervals on the coordinates $(x,y,z)$ for each source mechanism - isotropic (ISO), double couple (DC) and compensated linear vector dipole (CLVD). The three events are randomly sampled from the test set. These results are obtained by considering all 23 receivers, and training on 8000 simulated events. The noise level is set to 10.0, 0.3 and 3.5 respectively, which corresponds to a signal-to-noise ratio of $33$ dB.}
\label{tab:inference}
\begin{tabular}{@{}ccccccc}
\textbf{Event} &\textbf{Coordinate} & \textbf{Prior range [km]} & \textbf{Ground truth [km]} & \textbf{ISO [km]} & \textbf{DC [km]} & \textbf{CLVD [km] } \\
\hline \hline
\multirow{3}{*}{1} & $x$ & $[0, 1]$ & $0.71$ & $0.63^{+0.13}_{-0.12}$ & $0.72^{+0.13}_{-0.11}$ & $0.71^{+0.11}_{-0.10}$ \\&$y$& $[0, 1]$ & $0.25$ & $0.28^{+0.13}_{-0.14}$ & $0.25^{+0.10}_{-0.11}$ & $0.24^{+0.11}_{-0.11}$ \\&$z$ & $[0,2.43]$ & $2.10$ & $2.04^{+0.21}_{-0.47}$ & $2.09^{+0.11}_{-0.13}$ & $2.10^{+0.22}_{-0.25}$ \\
\hline
\multirow{3}{*}{2} & $x$ & $[0, 1]$ & $0.46$ & $0.48^{+0.19}_{-0.19}$ & $0.43^{+0.12}_{-0.13}$ &$0.46^{+0.15}_{-0.12}$ \\&$y$& $[0, 1]$ & $0.34$ & $0.32^{+0.16}_{-0.19}$ & $0.33^{+0.09}_{-0.08}$ &$0.33^{+0.15}_{-0.14}$ \\&$z$ & $[0, 2.43]$ & $1.48$ & $1.60^{+0.23}_{-0.30}$ & $1.54^{+0.18}_{-0.15}$ & $1.44^{+0.22}_{-0.24}$ \\
\hline
\multirow{3}{*}{3} & $x$ & $[0, 1]$ & $0.20$ & $0.28^{+0.20}_{-0.15}$ & $0.20^{+0.10}_{-0.10}$ & $0.22^{+0.25}_{-0.13}$ \\&$y$& $[0, 1]$ & $0.43$ & $0.46^{+0.18}_{-0.16}$ & $0.41^{+0.10}_{-0.13}$ & $0.54^{+0.17}_{-0.21}$ \\&$z$ & $[0, 2.43]$ & $0.99$ & $0.93^{+0.17}_{-0.19}$ & $1.09^{+0.28}_{-0.17}$ & $0.98^{+0.21}_{-0.20}$ \\
\end{tabular}
\end{table*}
\begin{figure}%
\begin{center}
\includegraphics[width=0.41\textwidth]{plots/ISO_23_8000_10.0_-996_156_10_log_final_km.pdf}
\includegraphics[width=0.41\textwidth]{plots/DC_23_8000_0.3_-996_261_10_log_km.pdf}
\end{center}
\begin{center}
\includegraphics[width=0.41\textwidth]{plots/CLVD_23_8000_3.5_-996_229_10_log_final_km.pdf}
\end{center}
\vspace{-2.4em}\caption{Marginalised 68 and 95 per cent credibility contours obtained with our method for a source located at $(x,y,z)=(0.71 \ \rm{km},0.25 \ \rm{km},2.10 \ \rm{km})$, indicated by the dashed black lines. We compare 3 source mechanisms: isotropic (ISO), double couple (DC), and compensated linear vector dipole (CLVD). The event corresponds to event 1 in Table~\ref{tab:inference}. Note that we are considering 23 receivers, the signal-to-noise ratio is 33 dB and the emulator was trained on 8000 training points.}
\label{fig:inference_1}
\end{figure}
\begin{figure}%
\begin{center}
\includegraphics[width=0.41\textwidth]{plots/ISO_23_8000_10.0_-16_156_10_log_final_km.pdf}
\includegraphics[width=0.41\textwidth]{plots/DC_23_8000_0.3_-16_261_10_log_km.pdf}
\end{center}
\begin{center}
\includegraphics[width=0.41\textwidth]{plots/CLVD_23_8000_3.5_-16_229_10_log_final_km.pdf}
\end{center}
\vspace{-2.4em}\caption{Same as Fig.~\ref{fig:inference_1} for a source located at $(x,y,z)=(0.46 \ \rm{km},0.34 \ \rm{km},1.48 \ \rm{km})$. The event corresponds to event 2 in Table~\ref{tab:inference}.}
\label{fig:inference_2}
\end{figure}
\begin{figure}%
\begin{center}
\includegraphics[width=0.41\textwidth]{plots/ISO_23_8000_10.0_-2_156_10_log_final_km.pdf}
\includegraphics[width=0.41\textwidth]{plots/DC_23_8000_0.3_-2_261_10_log_km.pdf}
\end{center}
\begin{center}
\includegraphics[width=0.41\textwidth]{plots/CLVD_23_8000_3.5_-2_229_10_log_final_km.pdf}
\end{center}
\caption{Same as Fig.~\ref{fig:inference_1} for a source located at $(x,y,z)=(0.20 \ \rm{km},0.43 \ \rm{km},0.99 \ \rm{km})$. The event corresponds to event 3 in Table~\ref{tab:inference}.}
\label{fig:inference_3}
\end{figure}
We then turn our attention to the accuracy of the inferred posterior distribution of the coordinates. For each source mechanism, we report the inference results for 3 different coordinates in Figs.~\ref{fig:inference_1}, \ref{fig:inference_2} and \ref{fig:inference_3}: each shows the posterior contour plots obtained with our methodology, considering all 23 receivers, SNR=33 dB and using 8000 seismograms as the training set for the emulator. The numerical results are summarised in Table~\ref{tab:inference}, reporting the prior ranges and marginalised mean and 68 percent credibility interval on the coordinates. We additionally report the mean and the standard deviation of the absolute value of the difference between the ground truth and the maximum of the retrieved posterior distribution for 100 randomly-picked test events: this is $(0.17 \pm 0.29)$ km. This difference indicates good agreement, despite being skewed by a small fraction (less than 5\%) of test cases for which the inference results were totally unconstrained, or for which the width of the posterior distribution was particularly broad: we remark that these occurrences generally corresponded to events for which the ground truth had at least one coordinate lying at the border of the prior space, thus complicating the sampling procedure. We argue that it is safe to ignore these limit cases, given their particular location; in a real scenario, one would train the neural network on events coming from a larger prior volume than the one that is being investigated here, thus avoiding this problem. We therefore conclude that with our method we can accurately retrieve the correct value of the coordinates across almost the entire prior parameter space.
Additionally, we observe that the $x$ and $y$ source coordinates are usually less constrained than the $z$ coordinate for ISO and CLVD, while this behaviour is less prominent in the DC case, for which all coordinates are always tightly constrained. We attribute this effect to two possible causes. On one hand, it can be related to the specific density model we are considering in this work, which has a layered structure. On the other hand, we note that when translating the seismic traces to the Fourier domain we ignored the phase signal (since we only considered the amplitude power spectra), thus possibly losing useful information for the source location purpose \citep[e.g.,][]{Ferreira07}; we argue that this also results in larger uncertainties in the retrieved posterior contours with respect to a full-waveform approach, which however would be too computationally expensive to run. We additionally note that retaining the phase information would yield a generative model, as by combining the power spectra with the phase one could in principle reconstruct a full seismogram from the source coordinates (after training the emulator). We defer the study of phase information to future work, as we anticipate that given the oscillatory behaviour of the phase signals it will be harder to train an emulator on them.
We then study the dependence of the posterior contours on the SNR, the number of training data and the number of receivers used. In Fig.~\ref{fig:compare_noise}, we first show the effect of different noise levels; in particular, for each source mechanism we vary the SNR from 13 dB to 54 dB. We note that, in order to increase the robustness to noise, we cut different frequency windows based on the noise levels: given the Gaussian noise model we assume in this work, a smaller SNR corresponds to a higher noise power, and hence to a smaller number of retained power spectrum principal components. We observe that while higher noise levels can lead to small biases in some of the 1-D marginalised coordinates' distributions, in general no significant variations in the shape of the 2-D posterior contours are present.
In Fig.~2 in the supplementary material we show the posterior contours when training the emulator with 2000, 5000 and 8000 data points. Again, we observe no significant differences for the ISO and CLVD sources, while very small deviations appear when using fewer training data in the DC case. In general, $\mathcal{O} (10^3)$ training data are enough to obtain accurate posterior contours for all source mechanisms. Finally, in Fig.~3 in the supplementary material we vary the geometry of the receivers used for recording the microseismic traces. While we are not interested in a full study of the optimal geometry of the receivers, we observe that using fewer receivers leads to broader posterior contours, while still allowing accurate event locations for all source mechanisms.
In summary, our results are very robust to the noise injected into the observed seismogram. Additionally, very few receivers - $\mathcal{O}(10)$ - are needed to obtain accurate results, and a number of training points of order $\mathcal{O} (10^3)$ is sufficient to locate any event.
\begin{figure*}
$\begin{array}{cc}
\includegraphics[width=0.41\textwidth]{plots/ISO_compare_noise.pdf} &
\includegraphics[width=0.41\textwidth]{plots/DC_compare_noise.pdf} \\
\rm{ISO}& \rm{DC}
\end{array}$
$\begin{array}{c}
\includegraphics[width=0.41\textwidth]{plots/CLVD_compare_noise.pdf} \\
\rm{CLVD}
\end{array}$
\caption[]{\label{fig:compare_noise}Marginalised 68 per cent credibility contours obtained with our method for a source located at $(x,y,z)=(0.71 \ \rm{km},0.25 \ \rm{km},2.10 \ \rm{km})$, indicated by the dashed black lines, comparing different levels of noise. In particular, in each panel we show signal-to-noise ratios of 13 dB, 33 dB and 54 dB for the 3 source mechanisms: isotropic (ISO), double couple (DC), and compensated linear vector dipole (CLVD). Note that we are considering 8000 training data for the emulator and 23 receivers. This figure is best viewed in colour.}
\end{figure*}
\begin{figure}%
\begin{center}
\includegraphics[width=0.4\textwidth]{plots/arrival_time_iso.pdf}
\includegraphics[width=0.4\textwidth]{plots/arrival_time_dc.pdf}
\end{center}
\begin{center}
\includegraphics[width=0.4\textwidth]{plots/arrival_time_clvd.pdf}
\end{center}
\caption{Comparison of the marginalised 68 and 95 per cent credibility contours obtained with our method and with the arrival time approach described in Sect.~\ref{sec:arrival_time}. The event corresponds to event 2 in Table~\ref{tab:inference}. Note that we are considering 23 receivers in both cases, and for our approach the signal-to-noise ratio is 33 dB and the emulator was trained on 8000 training points.}
\label{fig:edt_compare}
\end{figure}
\subsection{Comparison with arrival time techniques}
\label{sec:arrival_time}
In order to establish a connection with existing standard earthquake location methods, we compare our results with a non-linear probabilistic location technique based on arrival times. We consider the \textsc{NonLinLoc} algorithm \citep{Lomax00}, which implements the LS-L2 approach \citep{Tarantola82, Moser92, Wittlinger93}. In this framework, the likelihood of the arrival time for a single observed event for a single receiver is:
\begin{equation}
\mathcal{L}_{\rm{arrival \ time}} \propto \exp \left(-\frac{1}{2} \frac{\left(T_{\rm{obs}} - T_{\rm{calc}}\right)^2}{\sigma_T^2} \right)
\label{eq:edt}
\end{equation}
where $T_{\rm{obs}}$ is the observed arrival time, $T_{\rm{calc}}$ is the theoretical estimate for the travel time, and $\sigma_T$ is the error on the manual picking of arrival times. We use a fast marching method \citep{Sethian96} as implemented in \textsc{Pykonal} \citep{White20} to estimate theoretical arrival times given a set of coordinates for the geological model described in Sect.~\ref{sec:data}. We further replace the posterior sampler available in \textsc{NonLinLoc} with a \textsc{PyMultiNest} implementation for the likelihood in Eq.~\ref{eq:edt}.
We compare the location inference results for event 2 in Fig.~\ref{fig:edt_compare}. In this instance, we are interested in showing that there exists a reasonable value of the picking error $\sigma_T$ that yields comparable constraints to our method; this error is usually arbitrary and heavily depends on the SNR and prior information \citep{Smith19thesis, Abakumov20}. We remark that we are not interested in showing that the constraints obtained with our method are tighter than the ones obtained with the time-arrival information: while we start from the full waveform, we subsequently discard the phase information and part of the power spectrum, as we described in Sect.~\ref{sec:preprocessing}. We additionally note that the arrival time estimates are typically obtained an order of magnitude faster than our method, taking $\mathcal{O}(0.01 \ \rm{h})$ on the same commercial laptop used in Sect.~\ref{sec:speed}; this is not surprising, given the simple form of the likelihood in Eq.~\ref{eq:edt} and the efficiency of \textsc{Pykonal}.
Nevertheless, we show that with a picking error of $\sigma_T=0.05$ s, we obtain results which are generally in agreement with our approach. In particular, the constraints on the $z$ source coordinate are tighter in the arrival-time case, while the $x$ and $y$ coordinates are equally or better constrained by our approach; moreover, as in our method, the arrival-time analysis yields $z$ coordinates that are more constrained than the $x$ and $y$ ones. We speculate that this is due to the particular layered structure of the geological model described in Sect.~\ref{sec:data}, and, as anticipated in Sect.~\ref{sec:inference_results}, we expect the inference power of our approach to improve significantly by including the phase information into the analysis. We stress that our work is a step towards the fast joint inversion of coordinates and moment tensor components, for which arrival time techniques alone are not sufficient \citep{Dahm14, Pugh16, Alvizuri18}; therefore, we leave a full comparison to other techniques to future work. However, we further discuss the joint inversion approach and how to deal with more complicated noise types in Sect.~\ref{sec:conclusions}.
\subsection{Model selection results}
\label{sec:model_sel_res}
In general, a microseismic event can be described as a linear combination of the three source mechanisms described in Sect.~\ref{sec:data} \citep{Vavricuk15}. While we considered the three sources separately in this work, we show how the proposed Bayesian approach additionally allows for the identification of a source type given an observed seismogram. We consider event 1 as reported in Table~\ref{tab:inference}, and produce an observation for each source mechanism (ISO, DC and CLVD). As described in Sect.~\ref{sec:evidence}, we can run nested sampling for each event and for every trained emulator, and obtain 9 evidence values in total, whose logarithm we report in Table~\ref{tab:evidence}. As expected, the evidence is maximal in correspondence of the source type that generated the given event, which indicates that we are capable of correctly identifying the source type for a given observation. What is more, this selection is also very fast, as after training the emulators each evidence calculation takes less than 10 minutes on a commercial laptop.
\setlength{\tabcolsep}{3pt}
\begin{table}
\centering
\caption{Natural logarithm of the evidence for a source located at $(x,y,z)=(0.71 \ \rm{km},0.25 \ \rm{km},2.10 \ \rm{km})$ and source mechanism isotropic (ISO), double couple (DC) and compensated linear vector dipole (CLVD), as described in Sect.~\ref{sec:evidence} and Sect.~\ref{sec:model_sel_res}. The hypotheses correspond to an ISO ($\mathcal{H}_{\rm{ISO}}$), DC ($\mathcal{H}_{\rm{DC}}$) and CLVD ($\mathcal{H}_{\rm{CLVD}}$) source mechanism, respectively. We highlighted in bold the highest evidence in each line, which correctly corresponds to the known source mechanism. Note that the natural logarithm of the evidence is returned by \textsc{PyMultiNest}, and in this instance we ignored its associated error (which is very small).}
\begin{tabular}{@{}cccc}
\textbf{Event type} &$\ln{ \mathrm{Pr} \left( \boldsymbol{D} | \mathcal{H}_{\rm{ISO}}\right)} $ & $\ln{ \mathrm{Pr} \left( \boldsymbol{D} | \mathcal{H}_{\rm{DC}}\right)} $ & $\ln{ \mathrm{Pr} \left( \boldsymbol{D} | \mathcal{H}_{\rm{CLVD}}\right)} $ \\
\hline \hline
ISO & \textbf{2601} & $-$608 & $-$79 \\
\hline
DC & $-$17688 & \textbf{$-$1165} & $-$2232 \\
\hline
CLVD & $-$6620 & $-$1052 & \textbf{$-$394} \\
\end{tabular}
\label{tab:evidence}
\end{table}
\subsection{Realistic noise and network configuration}
\label{sec:real}
\begin{figure}%
\begin{center}
\includegraphics[width=0.41\textwidth]{plots/ISO_23_8000_-996_156_10_log_final_km_real.pdf}
\includegraphics[width=0.41\textwidth]{plots/DC_23_8000_-996_261_10_log_final_km_real.pdf}
\end{center}
\begin{center}
\includegraphics[width=0.41\textwidth]{plots/CLVD_23_8000_-996_229_10_log_final_km_real.pdf}
\end{center}
\caption{Same as Fig.~\ref{fig:inference_1}, considering real noise from Groningen field data instead of Gaussian noise. As in our fiducial analysis, the signal-to-noise ratio is 33 dB, and we consider 23 receivers.}
\label{fig:real_noise}
\end{figure}
In this section, we explore the application of our approach to a more realistic scenario. We consider the same 23-receiver configuration, which is uniformly spread on the plane at $z=2.43$ km; however, instead of Gaussian noise, we add real noise traces extracted from the freely available Groningen field data set \citep{Smith19thesis}.
Our approach to obtain real noise traces is as follows. We cluster the receivers and synthetic traces in cubic volumes of side 0.5 km; this returns 100 clusters in total. We therefore take 100 different real waveforms among Groningen data, and cut the first 5 s window, well before the first wave arrival. We interpolate the signal in these windows in order to match the sampling rate of the synthetic signals; then, in order to reach the required SNR of 33 dB, we tune the amplitude of the noise signal we add to each synthetic waveform, and subtract the average signal so that the resulting noisy waveform has zero mean. We found that in some cases the SNR of the resulting noisy signals is lower than 33 dB: these correspond to waveforms originating at the edge of our simulated field, which, however, do not impact the inference analysis, as we show in this section.
Real noise is added to event 1 in Table~\ref{tab:inference}, as well as to all training waveforms to calculate the covariance matrix in Eq.~\ref{eq:glike}. We re-train the neural networks in the DC and CLVD cases removing the zero-frequency value of the power spectrum, in order to be insensitive to the scale of the noisy signals in the time domain. In Fig.~\ref{fig:real_noise} we show the posterior contours of our inference analysis using real noise: the results show little difference with respect to the case with Gaussian noise in Fig~\ref{fig:inference_1}, thus further demonstrating the robustness of our approach.
\section{Conclusions}
\label{sec:conclusions}
In this paper, we proposed a method that allows for the fast and accurate retrieval of the source coordinates for any microseismic source mechanism: isotropic (ISO), double couple (DC), and compensated linear vector dipole (CLVD). This offers an efficient technique to both locate an event and identify its source type, exploiting the power of machine learning and Bayesian tools to extract the information contained in seismic waveforms.
Our proposed method is based on a physically motivated preprocessing of the raw signals, using Fourier analysis and principal component compression, followed by the use of a neural network to learn the mapping between coordinates and principal components. Using the learnt forward model in combination with Bayesian techniques, we showed that we can retrieve an accurate estimate of any microseismic event coordinates, for any source mechanism, in less than $10$ minutes on a commercial laptop. Therefore, we demonstrated for the first time that machine learning techniques allow for a fast and accurate Bayesian analysis on microseismic traces, yielding competitive results on ISO sources and state-of-the-art results on DC and CLVD sources.
We showed that $\mathcal{O} (10^3)$ events for each source mechanism are enough to train a representative emulator, when using the data coming from $\mathcal{O}(10)$ receivers placed at the seabed as indicated in Fig.~\ref{fig:rec_pos}. We explored the effect of the noise level, and how the number of receivers and the number of training data for the emulator impact the accuracy of the coordinates' posterior distribution, demonstrating the robustness of our approach, which was also compared with a standard arrival-time inversion technique, showing good agreement for an appropriate choice of the picking error. We also showed that our method works in the presence of realistic noise statistics and network configuration. Finally, we demonstrated the utility of our Bayesian approach by calculating the Bayesian evidence for a given observation and three hypotheses, and showed that this correctly identifies the source type of any given event.
In conclusion, our work lays the foundations for the fast and reliable location of microseismic events with any source mechanism, given a minimal amount of computing resources. We also foresee that recent improvements in solving the forward model even outside the boundary of the training data, like physics-informed neural networks \citep[PINNs,][]{Raissi19, Xu19, Costa19, Moseley20}, could be combined with our proposed approach to make it even more robust, especially when different velocity models have to be employed.
Our work represents a step forward towards the fast Bayesian characterisation of microseismic events: after training our models on noiseless simulation data, it is possible to apply our method to each individual seismic trace as recorded by the receivers in order to obtain a fast and accurate source location estimate. We plan to integrate our approach into a joint inversion analysis, where both moment tensor components and coordinates are inferred: such analysis is usually slow and computationally expensive \citep[e.g.,][]{Pugh16}, and we therefore expect our approach to significantly accelerate it. Some straightforward extensions of our approach then include the following three points. First, a microseismic event is in general described by a linear mixture of components of the moment tensor, which we considered separately in this work. As a consequence, this implies that a larger set of parameters, including two more free parameters measuring the relative strength of the source types, has to be considered. We anticipate that this increase in the total number of parameters to infer
will require a larger dataset to train the emulator; however, we note that our proposed method scales well with the number of training data, and therefore we anticipate that performing a Bayesian analysis with a larger parameter space is an attainable goal using our approach. Second we note that in order for this method to be deployed in a realistic scenario, the noise associated with the recorded seismograms should be modelled more carefully: a lower SNR may have to be considered, a more complicated likelihood distribution might have to be implemented, or a ``likelihood-free" approach should be investigated \citep{Sunnaker13}. Last, the errors in the 3-D density and velocity models should be incorporated into the analysis, in order to account for all sources of uncertainty \citep{Gesret13, Gesret14}. This will be addressed in future work.
\begin{acknowledgments}
We thank Saptarshi Das
for useful discussions, as well as Jonathan Smith and the anonymous reviewers for their helpful comments. DP was supported by the STFC UCL Centre for Doctoral Training in Data Intensive Science. Generation of the synthetic data used in this work has been performed in part on the Wilkes High Performance GPU computer cluster at the University of Cambridge, and in part on the Beaker cluster at UCL.
This work was partially enabled by funding from Royal Dutch Shell plc, and used facilities provided by the UCL Cosmoparticle Initiative. We acknowledge the use of $\textsc{NumPy}$ \citep{Harris20}, $\textsc{Matplotlib}$ \citep{Hunter07}, $\textsc{TensorFlow}$ \citep{Abadi15}, $\textsc{SciPy}$ \citep{Virtanen20}, and $\textsc{ChainConsumer}$ \citep{Hinton16}.
\end{acknowledgments}
\begin{dataavailability}
The code to reproduce this work is publicly available in the GitHub repository at this link: \url{https://github.com/alessiospuriomancini/seismoML}. The data underlying this article will be shared on reasonable request to the corresponding author.
\end{dataavailability}
\newcommand{}{}
\bibliographystyle{apalike}
\section{Loss curves}
\label{app:loss}
In Fig.~\ref{fig:loss} we report the training and validation loss curves for one receiver in the CLVD case. Note this is a typical example, as the trend in the loss curves is similar for all receivers and all source types.
\begin{figure*}
\centering
\includegraphics[width=\textwidth, keepaspectratio]{plots/loss_curves.pdf}
\caption{Typical example of training (solid blue) and validation (dashed coral) loss curves for one receiver in the CLVD case. The vertical dashed black line indicates the epoch with minimum validation loss, which we interpret as corresponding to the best model.}
\label{fig:loss}
\end{figure*}
\section{Additional studies on the posterior distribution}
\label{app:more_studies}
In this section, we show complementary experiments to those presented at the end of Sect.~4.3 in the main text. In order to demonstrate the robustness of our inference pipeline, we show the dependence of the posterior distribution retrieved with our method as a function of the number of training points (Fig.~\ref{fig:compare_training}) and of the number of receivers (Fig.~\ref{fig:compare_recs}). We further comment on these results at the end of Sect.~4.3 in the main text.
\begin{figure*}
$\begin{array}{cc}
\includegraphics[width=0.41\textwidth]{plots/ISO_compare_training.pdf} &
\includegraphics[width=0.41\textwidth]{plots/DC_compare_training.pdf} \\
\rm{ISO}& \rm{DC}
\end{array}$
$\begin{array}{c}
\includegraphics[width=0.41\textwidth]{plots/CLVD_compare_training.pdf} \\
\rm{CLVD}
\end{array}$
\caption[]
{\label{fig:compare_training}Marginalised 68 per cent credibility contours obtained with our method for a source located at $(x,y,z)=(0.20 \ \rm{km},0.43 \ \rm{km},0.99 \ \rm{km})$, indicated by the dashed black lines, comparing different numbers of training data for the emulator (2000, 5000 and 8000). We show these results for 3 source mechanisms: isotropic (ISO), double couple (DC), and compensated linear vector dipole (CLVD). Note that we are considering 23 receivers and a signal-to-noise ratio of 33 dB. This figure is best viewed in colour.}
\end{figure*}
\begin{figure*}
$\begin{array}{cc}
\includegraphics[width=0.41\textwidth]{plots/ISO_compare_recs.pdf} &
\includegraphics[width=0.41\textwidth]{plots/DC_compare_recs.pdf} \\
\rm{ISO}& \rm{DC}
\end{array}$
$\begin{array}{c}
\includegraphics[width=0.41\textwidth]{plots/CLVD_compare_recs.pdf} \\
\rm{CLVD}
\end{array}$
\caption[]
{\label{fig:compare_recs}Marginalised 68 per cent credibility contours obtained with our method for a source located at $(x,y,z)=(0.46 \ \rm{km},0.34 \ \rm{km},1.48 \ \rm{km})$, indicated by the dashed black lines, comparing different dispositions of the receivers. In particular, 5 receivers refer to the blue squares in Fig.~2 in the main text, and 9 receivers refer to the green circles in Fig.~2 in the main text. We show these results for 3 source mechanisms: isotropic (ISO), double couple (DC), and compensated linear vector dipole (CLVD). Note that we are considering 8000 training data for the emulator and a signal-to-noise ratio of 33 dB. This figure is best viewed in colour.}
\end{figure*}
\label{lastpage}
\end{document} |
\subsection{Numerical Comparison using RMSE}
We provide numerical comparison to further verify the advantage of VKRE. Tables \ref{tab1}, \ref{tab2}, and \ref{tab3} demonstrate the difference between NWE and VKRE for the second regression function in (\ref{reg_fun}) using the measure,
$$RMSE= \sqrt{\frac{1}{n}\sum_{i=1}^n \big(r(X_i) - \hat{r}(X_i)\big)^2},$$
where $r(t)$ is the real regression function and $\hat{r}(t)$ is NWE or VKRE. Each entry of the tables is the average of RMSE from $N=250$ samples.
In Table \ref{tab1}, we compare the RMSE for NWE and VKRE with the random errors from different distributions, $X$ from the T-distribution with degree of freedom $4$, and the sample size of $5000$. The table shows that as the bounds for the uniform distribution increases, the RMSE increases for both NWE and VKRE.
\begin{table}[H
\center
\caption{ Comparing RMSE of VKRE and NWE with the random errors from different distributions.}
\label{tab1}
\bigskip
\begin{tabular}{| c| c| c| }
\hline
\textbf{\;\;Errors\;\;} & \textbf{\;\;NWE RMSE\;\;} & \textbf{\;\;VKRE RMSE\;\;} \\
\hline
$U[-0.5,0.5]$ & 0.01165791& 0.008649485\\
\hline
$U[-1,1]$ & 0.02135706& 0.01474327\\
\hline
$U[-2,2]$ & 0.03912875& 0.02785478\\
\hline
\end{tabular}
\end{table}
In Table \ref{tab2}, we list the RMSE for NWE and VKRE with $X$ generated from different distributions, $\varepsilon$ generated from the uniform distribution on the interval $[-0.5,0.5]$, and $n=5000$ with $N=250$ repetitions. Recall that the T-distribution with degree of freedom $1$ is same as the standard Cauchy distribution. It is interesting to see that as the degrees of freedom increase for T-distribution, the RMSE decreases for both of the estimators.
\begin{table}[H
\center
\caption{ Comparing RMSE of VKRE and NWE with $X$ from different distributions.}
\label{tab2}
\bigskip
\begin{tabular}{| c| c| c| }
\hline
\textbf{\;\;$X$\;\;} & \textbf{\;\;NWE RMSE\;\;} & \textbf{\;\;VKRE RMSE\;\;} \\
\hline
$T(df=1)$ & 0.01973383& 0.01215381\\
\hline
$T(df=4)$ & 0.01165791& 0.008649485\\
\hline
$T(df=8)$ & 0.006720047 & 0.004981605 \\
\hline
$Cauchy (3,4)$ & 0.02445313 &0.01189971 \\
\hline
$Cauchy (5,7)$ & 0.02656666 & 0.01461736 \\
\hline
$N(0,1)$ & 0.007069801 &0.005971244 \\
\hline
$N(5,10)$ & 0.01349415 & 0.01335264 \\
\hline
\end{tabular}
\end{table}
In Table \ref{tab3}, we compare the RMSE for NWE and VKRE with different sample sizes, $\varepsilon$ generated from the uniform distribution on the interval $[-0.5,0.5]$, and $X$ generated from the T-distribution with degree of freedom $4$. The table shows that the RMSE decreases as the sample size increases for both of the estimators.
\begin{table}[H
\center
\caption{ Comparing RMSE of VKRE and NWE with different sample sizes.}
\label{tab3}
\bigskip
\begin{tabular}{| c| c| c| }
\hline
\textbf{\;\;$n$\;\;} & \textbf{\;\;NWE RMSE\;\;} & \textbf{\;\;VKRE RMSE\;\;} \\
\hline
$500$ & 0.01866795 & 0.01520967\\
\hline
$1000$ & 0.01598541 & 0.01162556 \\
\hline
$2000$ & 0.01358658 & 0.01012766\\
\hline
$5000$ & 0.01165791& 0.008649485\\
\hline
$8000$ & 0.009052834 & 0.00564806 \\
\hline
$10000$ & 0.007793899 &0.004173846 \\
\hline
\end{tabular}
\end{table}
Moreover, the results from all three tables indicate that the RMSE of VKRE is smaller than that of NWE in each case. Thus, in all situations considered, VKRE outperforms NWE.
\subsection{Numerical Comparison using Monte-Carlo Estimation of MSE}
This subsection covers Monte-Carlo simulation and approximation of the following mean square error,
$$MSE= \mathbb{E}[ \big(\hat{r}(t) - r(t)\big)^2],$$
to compare NWE and VKRE estimators. We compare the regression estimators for the second regression function (Bounded RF) and the third regression function (Unbounded RF) in (\ref{reg_fun}) in the following tables. Each entry in the tables is the average of $(\hat{r}(t) - r(t))^2$ from $N=250$ random samples each with sample size $n= 5000$. In each table, we list the results at $10$ points which are evenly selected from the range of $nN$ simulated $X$ values with the specified T-distribution or the standard normal distribution. \\
Tables \ref{tab4} and \ref{tab5} compares NWE and VKRE for the bounded and the unbounded regression functions in (\ref{reg_fun}) where $X$ is generated from the T-distribution with degrees of freedom 4, whereas in Tables \ref{tab6} and \ref{tab7}, $X$ is generated from the standard normal distribution. In Tables \ref{tab4} and \ref{tab6} the errors are from the standard normal distribution whereas the errors in Tables \ref{tab5} and \ref{tab7} are generated from the uniform distribution on the interval [-1,1].
\begin{table}[H
\center
\caption{ Comparing MSE of VKRE and NWE with $X \sim T(4)$ and $\varepsilon \sim N(0,1)$.}
\label{tab4}
\bigskip
\begin{tabular}{|c|l|l|l|l|l|l|}
\hline
&
\multicolumn{2}{c}{ \textbf{Bounded RF}} &
\multicolumn{2}{c|}{ \textbf{Unbounded RF}} \\
\hline
\textbf{\;\;$t$\;\;} & \textbf{\;\;NWE \;\;} & \textbf{\;\;VKRE \;\;} & \textbf{\;\;NWE \;\;} & \textbf{\;\;VKRE \;\;}\\
\hline
-7.161518 & 0.082425068 & 0.022757438 & 0.952882068 & 0.012199782\\
\hline
-5.593896 & 0.038621453 &0.029278125 & 0.181799310 & 0.010549010 \\
\hline
-4.026274 & 0.001274921 &0.001691833 & 0.059748939 & 0.000630284 \\
\hline
-2.458652& 0.000693960 & 0.000953009 & 0.028731527 & 0.000743185 \\
\hline
-0.89103& 0.000095033 & 0.000350961 & 0.003600552 & 0.000450090 \\
\hline
0.676592& 0.000470135 &0.000042570 & 0.000808568 & 0.000039578 \\
\hline
2.244214 & 0.006266251 & 0.002081689 & 0.027084148 & 0.002896018 \\
\hline
3.811836 & 0.006611614 & 0.000129696 & 0.049252162 & 0.002134087 \\
\hline
5.379458 & 0.021020283 & 0.001573274 & 0.067687352 & 0.014088619 \\
\hline
6.94708 &0.013454439 & 0.004420131& 0.709749776 & 0.042635434 \\
\hline
\end{tabular}
\end{table}
\begin{table}[H
\center
\caption{ Comparing MSE of VKRE and NWE with $X \sim T(4)$ and $\varepsilon \sim U[-1,1]$.}
\label{tab5}
\bigskip
\begin{tabular}{|c|l|l|l|l|l|l|}
\hline
&
\multicolumn{2}{c}{ \textbf{Bounded RF}} &
\multicolumn{2}{c|}{ \textbf{Unbounded RF}} \\
\hline
\textbf{\;\;$t$\;\;} & \textbf{\;\;NWE \;\;} & \textbf{\;\;VKRE \;\;} & \textbf{\;\;NWE \;\;} & \textbf{\;\;VKRE \;\;}\\
\hline
-7.161518 & 0.037688485 & 0.015996216 & 1.4541232 & 0.0888033 \\
\hline
-5.593896 & 0.017563571 & 0.003636569 & 0.2238683 & 0.0268285 \\
\hline
-4.026274 & 0.000725586 & 0.000042321 & 0.0240457 & 0.0029750 \\
\hline
-2.458652 & 0.000536810 & 0.000052783 & 0.0103590 & 0.0001132 \\
\hline
-0.89103 & 0.000083700 & 0.000015774 & 0.0020801 & 0.0000299\\
\hline
0.676592 & 0.000015691 & 0.000047977 & 0.0019164 & 0.0000508 \\
\hline
2.244214 & 0.000151535 & 0.000032906 & 0.0111435 & 0.0002437 \\
\hline
3.811836 & 0.004435139 & 0.000604593 & 0.0250448 & 0.0005697 \\
\hline
5.379458 & 0.016278344 & 0.004865297 & 0.1074176 & 0.0271930 \\
\hline
6.94708 & 0.024189858 & 0.011530802 & 1.4184324 & 0.0756486\\
\hline
\end{tabular}
\end{table}
In Tables \ref{tab4}-\ref{tab7}, VKRE outperforms NWE for the unbounded regression function at all selected $t$ values. For the bounded regression function, VKRE has better performance than NWE at most of the $t$ values.
On the other hand, for the $t$ values where NWE has better performance than VKRE, the difference between the corresponding MSEs is relatively very small. In summary, the Monte-Carlo approximation of MSE also shows that VKRE outperforms NWE in general.
\begin{table}[H
\center
\caption{ Comparing MSE of VKRE and NWE with $X \sim N(0,1)$ and $\varepsilon \sim N(0,1)$.}
\label{tab6}
\bigskip
\begin{tabular}{|c|l|l|l|l|l|l|}
\hline
&
\multicolumn{2}{c}{ \textbf{Bounded RF}} &
\multicolumn{2}{c|}{ \textbf{Unbounded RF}} \\
\hline
\textbf{\;\;$t$\;\;} & \textbf{\;\;NWE \;\;} & \textbf{\;\;VKRE \;\;} & \textbf{\;\;NWE \;\;} & \textbf{\;\;VKRE \;\;}\\
\hline
-3.166296000 & 0.044502685 & 0.014599755 & 0.162391619 & 0.001478955 \\
\hline
-2.476748778 & 0.002788421 & 0.001986454 & 0.052629439 & 0.000617446 \\
\hline
-1.787201556 & 0.000879162 & 0.000831587 & 0.008172456 & 0.000803430 \\
\hline
-1.097654333 & 0.000227731 & 0.000011475 & 0.001634166 & 0.000075463 \\
\hline
-0.408107111 & 0.000120885 & 0.000182900 & 0.001949065 & 0.000539604 \\
\hline
0.281440111 & 0.000061373 & 0.000206260 & 0.000934502 & 0.000086808 \\
\hline
0.970987333 & 0.000007520 & 0.000083182 & 0.001005907 & 0.000117938 \\
\hline
1.660534556 & 0.003158112 & 0.001136445 & 0.017263192 & 0.000629518 \\
\hline
2.350081778 & 0.000985980 & 0.000025147 & 0.035076566 & 0.003724841 \\
\hline
3.039629000 & 0.003365033 & 0.000184329 & 0.026700199 & 0.025010083 \\
\hline
\end{tabular}
\end{table}
\begin{table}[H
\center
\caption{ Comparing MSE of VKRE and NWE with $X \sim N(0,1)$ and $\varepsilon \sim U[-1,1]$.}
\label{tab7}
\bigskip
\begin{tabular}{|c|l|l|l|l|l|l|}
\hline
&
\multicolumn{2}{c}{ \textbf{Bounded RF}} &
\multicolumn{2}{c|}{ \textbf{Unbounded RF}} \\
\hline
\textbf{\;\;$t$\;\;} & \textbf{\;\;NWE \;\;} & \textbf{\;\;VKRE \;\;} & \textbf{\;\;NWE \;\;} & \textbf{\;\;VKRE \;\;}\\
\hline
-3.166296000 & 0.030718309 & 0.002079258 & 0.067146594 & 0.051639374 \\
\hline
-2.476748778 & 0.000201962 & 0.000035443 & 0.018435422 & 0.006108237 \\
\hline
-1.787201556 & 0.000143322 & 0.000067562 & 0.003757715 & 0.000101261 \\
\hline
-1.097654333 & 0.000091673 & 0.000189295 & 0.002869324 & 0.000185627 \\
\hline
-0.408107111 & 0.000030893 & 0.000017899 & 0.001339005 & 0.000052513 \\
\hline
0.281440111 & 0.000013158 & 0.000017005 & 0.001477281 & 0.000069223 \\
\hline
0.970987333 & 0.000205468 & 0.000050455 & 0.001344766 & 0.000136607 \\
\hline
1.660534556 & 0.000036136 & 0.000004868 & 0.004016354 & 0.000039407 \\
\hline
2.350081778 & 0.001747355 & 0.000059404 & 0.019114022 & 0.001437631 \\
\hline
3.039629000 & 0.013097496 & 0.000167359 & 0.041389033 & 0.035245030 \\
\hline
\end{tabular}
\end{table}
\section{Conclusion}\label{conclusion}
In this article we propose a variable bandwidth kernel regression estimator. With this estimator, the bandwidth is proportional to $1/\sqrt{f(X_i)}$, $1\le i\le n$, the inverse of the square root of the marginal density function value of the independent variable at the observation which is same as the square root law in the variable bandwidth kernel density estimation (\cite{Abramson}, \cite{HM1988}, \cite{HallHuMarron}, \cite{TerrellScott}, \cite{McKay a}, \cite{McKay b}, \cite{JonesMcKayHu}, \cite{GineSang}, \cite{GS2013} and \cite{NS2016}). The bandwidth is also proportional to the inverse of the absolute value of the fourth root of the derivative of the regression function. It intuitively sounds since there are much more observations in the area with a large marginal density of the independent variable or with a large derivative of the regression function and we therefore shall take a relatively small bandwidth. On the other hand, this variable bandwidth method also selects a larger bandwidth in the sparse area to pick up more observations in the estimation. In the area where the marginal density of the independent variable is extremely small or the regression is very flat, the clipping procedure there will take a relative constant bandwidth to avoid over estimation since the the clipping function is bounded below.
Under some regular conditions on the kernel function and bandwidth sequence, we study the bias and mean squared error for both the ideal estimator and the true estimator. The order of the bias is $h_n^4$ instead of $h_n^2$ as in the classical Nadaraya-Watson kernel regression estimator. In consequence, the mean squared error has order of $n^{-8/9}$ instead of $n^{-4/5}$. We also obtain central limit theorem for the ideal estimator and the true estimator. The advantage of this variable bandwidth estimator over the classical Nadaraya-Watson kernel regression estimator is demonstrated by a simulation study.
It is also interesting to study the case when $X_i$'s are random variables on $\mathbb{R}^d$. In \cite{GS2013}, the authors studied variable bandwidth kernel density estimation (\ref{GSV}) in $d$-dimensional case. However, the true bandwidth in multidimensional variable bandwidth regression estimation should also involve the regression function $r$. It is an interesting problem to find the right variable bandwidth which can remove the term with $h_n^2$. We leave this part of work for future research.
\section {Proofs}\label{proof}
For convenience, we slightly modify the proofs of the theorems in \cite{McKay b}, \cite{JonesMcKayHu} or \cite{GS2013} and provide a proof for Proposition \ref{bias0rx}. The proof of Proposition \ref{bias0} is similar. \\
\noindent\begin{proof}[Proof of Proposition \ref{bias0rx}]
Since $\xi\geq c$ and $\xi'/\xi$ is bounded in a neighborhood of $t$, there exists $\delta>0$ such that $(v\xi(t-v))'=\xi(t-v)-v \xi'(t-v)>0$ for $v \in [-\delta, \delta]$. Hence the function $U_t(v):= v \xi(t-v)$ is invertible for $v \in [-\delta, \delta]$. The inverse function $V_t(u)$ is $l+1$ times differentiable with continuous derivatives. Since $K((t-s)\xi(s)/h_n)=0$ unless $|t-s|\leq T h_n/\xi(s) \leq T h_n/c$, the change of variables
\begin{eqnarray*}\label {hz}
h_n z = (t-s)\xi(t-(t-s)), \ \text{or} \ t-s = V_t(h_nz)
\end{eqnarray*}
in the following integral is valid
\begin{eqnarray*}\label {Vhz}
&& \frac{1 }{h_n}\int K\left ( \frac{t-s}{h_n} \xi(s) \right ) \xi(s) \eta(s) ds \nonumber\\
&=&-\int K(z) \xi \left (t-V_t(h_n z)\right)\eta \left (t-V_t(h_nz)\right)\frac{dV_t(h_nz)}{d (h_nz)}dz.
\end{eqnarray*}
Developing $\xi \left (t-V_t(h_n z)\right)\eta \left (t-V_t(h_nz)\right)\frac{dV_t(h_nz)}{d (h_nz)}$ into powers of $h_nz$, and the first statement of the proposition follows.
Let $\psi$ be an infinitely differentiable function of finite support. Changing variable $t=s+h_n u$, developing $\psi$, changing variable $\omega = u \xi(s)$, and integrating by parts, we have
\begin{eqnarray}\label {23aa}
&& \frac{1 }{h_n}\int \psi(t) K\left ( \frac{t-s}{h_n} \xi(s) \right ) \xi(s) \eta(s) ds \nonumber\\
&=& \int \psi(s)\eta(s)ds + \sum_{k=1}^l (-1)^k \frac{\tau_k h_n^k}{k!}\int \psi(s) D_k \left (\frac{\eta(s)}{\xi(s)} \right ) ds +o(h_n^l)
\end{eqnarray}
where $\tau_k=0$ when $k$ is odd by symmetry. By (\ref{P1}),
\begin{eqnarray}\label {24aa}
\frac{1 }{h_n}\int \psi(t) K\left ( \frac{t-s}{h_n} \xi(s) \right ) \xi(s) \eta(s) ds =\sum_{k=0}^l h_n^k \int \psi(t) a_k(t)dt +o(h_n^l).
\end{eqnarray}
Comparing the coefficients of (\ref{23aa}) and (\ref{24aa}), we obtain (\ref{P2}).
\end{proof}
We shall use the following formula to estimate the expectation of a quotient of two random variables:
\begin{eqnarray}\label {VB16ag}
\frac{1}{z}= 1-(z-1)+\cdots +(-1)^p (z-1)^p +(-1)^{p+1} \frac{(z-1)^{p+1}}{z}.
\end{eqnarray}
\noindent\begin{proof}[Proof of Theorem \ref{Thm01}]
By (\ref{rnt}) and (\ref{VB16ag}) with $p=1$ and $z=\bar{f}(t;h_n)/E \bar{f}(t;h_n)$,
\begin{eqnarray} \label {21ta}
E \bar{r} (t; h_n)= \frac{E \bar{g}(t; h_n)}{E \bar{f}(t; h_n)} \label {VB32a} +\frac{-I_1+I_2}{(E \bar{f} (t; h_n))^2}
\end{eqnarray}
where
\begin{eqnarray}
I_1 &=& E\left \{\bar{g}(t;h_n) \left (\bar{f}(t;h_n)-E \bar{f}(t;h_n) \right) \right\}, \label {I_1}\\
I_2 &=& E \left \{\bar{r}(t;h_n) \left (\bar{f}(t;h_n)-E \bar{f}(t;h_n)\right)^2 \right\}.\label {I_2}
\end{eqnarray}
\textbf{Step 1}. We estimate $E \bar{g}(t;h_n)/E\bar{f}(t;h_n)$. Let $g(t)=f(t)r(t)$. Then
\begin{eqnarray}\label {29L}
\frac{E \bar{g}(t; h_n)}{E \bar{f}(t; h_n)}-r(t)= \frac{f(t) E \bar{g}(t; h_n)-g(t)E \bar{f}(t; h_n)}{f(t)E \bar{f}(t; h_n)}.
\end{eqnarray}
Note that
\begin{eqnarray*}\label {28wv}
E \bar{g}(t; h_n)= \frac{1}{h_n} \int K \left ( \frac{t-s}{h_n} \alpha \left (q(s) \right) \right )\alpha(q(s))g(s) ds.
\end{eqnarray*}
For $t \in {\cal D}_{rf}$, by Proposition \ref{bias0rx} and Remark \ref{xicon},
\begin{eqnarray*} \label {BV14a}
E \bar{g}(t; h_n)= g(t) + \frac{\mu_{2,1} h_n^2}{2} D_2 \left ( \frac{g(t)}{ q(t)} \right ) + \frac{\mu_{4,1} h_n^4}{24} D_4 \left ( \frac{g(t)}{ q^2(t)} \right )+ o(h^4_n).
\end{eqnarray*}
Similarly,
\begin{eqnarray} \label {VB22ca}
E \bar{f}(t; h_n)= f(t) + \frac{\mu_{2,1} h_n^2}{2} D_2 \left ( \frac{ f(t)}{ q(t)} \right ) +\frac{\mu_{4,1} h_n^4}{24} D_4 \left ( \frac{ f(t)}{ q^2(t)} \right )+ o(h^4_n) .
\end{eqnarray}
Hence
\begin{eqnarray}\label {32L}
&& f(t) E \bar{g}(t; h_n)-g(t)E \bar{f}(t; h_n) \nonumber\\
&=&\frac{f(t)\mu_{2,1} h_n^2}{2 } \left \{D_2 \left ( \frac{g(t)}{ q(t)} \right ) - r(t) D_2 \left ( \frac{ f(t)}{ q(t)} \right ) \right \} \nonumber\\
&& +\frac{f(t)\mu_{4,1} h^4_n}{24 }\left \{ D_4 \left ( \frac{ g(t)}{ q^2(t)}\right ) -r(t) D_4 \left ( \frac{ f(t)}{q^2(t)} \right ) \right \}+o(h_n^4).
\end{eqnarray}
We next show that the term with $h^2_n$ is zero. Let $\phi(t)=f(t)/ q(t)$. Then
\begin{eqnarray*}
D_2 \left ( \frac{g(t)}{ q(t)} \right ) - r(t) D_2 \left ( \frac{ f(t)}{ q(t)} \right )&=& D_2 \left ( r(t) \phi(t) \right )- r(t)D_2 (\phi(t)) \nonumber\\
&=& r''(t) \phi(t) +2 r'(t)\phi'(t) \nonumber\\
&=& \frac{\left ( r'(t) \phi^2(t) \right )'}{\phi(t)} .
\end{eqnarray*}
Since $\phi(t)= f(t)/q(t)= |r'(t)|^{-1/2}$. Then $r'(t)\phi^2(t)=-1$ or $1$. Hence
\begin{eqnarray}\label {VB27ar}
D_2 \left ( \frac{g(t)}{ q(t)} \right ) - r(t) D_2 \left ( \frac{ f(t)}{ q(t)} \right ) = 0.
\end{eqnarray}
By (\ref{32L}) and (\ref{VB27ar}), for $\theta (t)$ defined in (\ref{VB12a}),
\begin{eqnarray}\label {37svb}
f(t) E \bar{g}(t; h_n)-g(t)E \bar{f}(t; h_n)= f^2(t) \theta(t)h^4_n +o(h^4_n).
\end{eqnarray}
Since $E \bar{f}(t; h_n)=f(t)+o(1)$ by (\ref{VB22ca}), then by (\ref{29L}) and (\ref{37svb}),
\begin{eqnarray}\label {VB21a}
\frac{E \bar{g}(t; h_n)}{E \bar{f}(t; h_n)} = r(t) + \theta(t) h^4_n +o(h^4_n).
\end{eqnarray}
\textbf{Step 2}. We estimate $I_1$ in (\ref{I_1}). Denote $W_i=(t-X_i) \alpha(q(X_i))/h_n$ and let $F_i=K(W_i)\alpha(q(X_i))$. By Propositons \ref{bias0rx} and \ref{bias0},
\begin{eqnarray}\label {30cb}
I_1 &=& \frac{1}{n^2 h_n^2} E \sum_{i \neq j} F_iY_i \left \{ F_j - EF_j \right \} +\frac{1}{n^2 h_n^2} E \sum_{i=1}^n F_iY_i \left \{ F_i - EF_i \right \} \nonumber \\
&=& \frac{1}{n h_n^2} E \left( F^2_1Y_1 \right)- \frac{1}{n h_n} E \left(F_1Y_1\right)\frac{1}{h_n}E F_1 \nonumber \\
&=&\frac{\sqrt{q(t)}g(t)\mu_{0,2}}{n h_n} +O\left ( \frac{1}{n}\right ).
\end{eqnarray}
\textbf{Step 3}. We estimate $I_2$ in (\ref{I_2}). By Taylor expansion of the function $\gamma(y)=(1+y)^{-1}$,
\begin{eqnarray}\label {barfrec}
\frac{1}{\bar{f}(t; h_n)}&=&\frac{1}{E\bar{f}(t; h_n)\left \{1+\bar{f}(t; h_n)/E\bar{f}(t; h_n)-1 \right\}}\nonumber\\
&=&\frac{1}{E\bar{f}(t; h_n)}\left ( 1+\gamma'(\xi_t)\frac{\bar{f}(t; h_n)-E\bar{f}(t; h_n)}{E\bar{f}(t; h_n)}\right)
\end{eqnarray}
where $\xi_t$ is between $0$ and $\bar{f}(t; h_n)/E\bar{f}(t; h_n)-1$. Then
\begin{eqnarray}\label {40vbb}
I_2 &=& \frac{1}{E \bar{f}(t; h_n)}E \left \{ \bar{g}(t; h_n) \left (\bar{f}(t; h_n) - E \bar{f}(t; h_n)\right)^2 \right \} \nonumber \\
&& +\frac{1}{(E \bar{f}(t; h_n))^2}E \left \{ \gamma'(\xi_t) \bar{g}(t; h_n) \left (\bar{f}(t; h_n) - E \bar{f}(t; h_n)\right)^3 \right \} \nonumber\\
&:=& \frac{1}{E \bar{f}(t; h_n)} I_{2,1}+\frac{1}{(E \bar{f}(t; h_n))^2}I_{2,2}.
\end{eqnarray}
Similar to the estimate in (\ref{30cb}),
\begin{eqnarray} \label {37cba}
I_{2,1}&= &\frac{1 }{n^3 h^3_n } E \sum_{i\neq j} F_i Y_i \left \{F_j- EF_j\right \}^2 +\frac{1 }{n^3 h^3_n } E \sum_{i=1}^n F_iY_i \left \{F_i - EF_i \right \}^2\nonumber\\
&=&\frac{1}{n h_n} E \bar{g}(t; h_n) \left \{ \frac{1}{h_n}E F_1^2- h_n\left (\frac{1}{h_n} E F_1\right )^2 \right \}+ O\left (\frac{1}{n^2 h^2_n} \right)\nonumber\\
&=&\frac{\sqrt{ q(t)}g(t)f(t)\mu_{0,2}}{n h_n} +o\left ( \frac{1}{n h_n} \right).
\end{eqnarray}
By H\a"{o}lder's inequality,
\begin{eqnarray*}
|I_{2,2}| \leq \|\gamma'(\cdot)\|_{\infty}\left(E \bar{g}^2(t; h_n)\right)^{1/2}\left ( E \left (\bar{f}(t; h_n) - E \bar{f}(t; h_n)\right)^6\right)^{1/2}.
\end{eqnarray*}
Recall that $F_i = K(W_i) \alpha (q(X_i)))$. Then
\begin{eqnarray}\label {Ebarf}
&& E \left (\bar{f}(t; h_n) - E \bar{f}(t; h_n)\right)^6\nonumber\\
&=& \frac{1}{n^6 h^6_n}E \sum_{i,j,k \ are \ different}(F_i-EF_i)^2 (F_j-EF_j)^2 (F_k-EF_k)^2 \nonumber\\
&&+ \frac{1}{n^6 h^6_n}E \sum_{i \neq j} \left \{ (F_i-EF_i)^4 (F_j-EF_j)^2+ (F_i-EF_i)^3 (F_j-EF_j)^3\right\} \nonumber\\
&&+ \frac{1}{n^6 h^6_n}E \sum_{i=1}^n (F_i-EF_i)^6 \nonumber\\
&=& O \left ( \frac{1}{n^3 h^3_n}\right ).
\end{eqnarray}
Hence
\begin{eqnarray}\label {40cb}
I_{2,2} = o \left ( \frac{1}{n h_n} \right ).
\end{eqnarray}
Since $E \bar{f}(t; h_n)=f(t)(1+o(1))$, then by (\ref{40vbb}), (\ref{37cba}) and (\ref{40cb}),
\begin{eqnarray}\label {42bb}
I_2 = \frac{\sqrt{q(t)}g(t)\mu_{0,2}}{n h_n} + o\left ( \frac{1}{n h_n} \right ).
\end{eqnarray}
By (\ref{21ta}), (\ref{VB21a}), (\ref{30cb}) and (\ref{42bb}), we obtain (\ref{Thm1(1)}).
\vskip5pt
\textbf{Step 4}.
Now we prove (\ref{Thm1(2)}). By (\ref{barfrec}),
\begin{eqnarray*}\label {46wt}
\bar{r}(t; h_n)-r(t) &=& \frac{f(t)\bar{g}(t; h_n)- g(t)\bar{f}(t; h_n)}{f(t)\bar{f}(t; h_n)}\nonumber\\
&=&\frac{f(t)\bar{g}(t; h_n)- g(t)\bar{f}(t; h_n)}{f(t)E\bar{f}(t; h_n) }\nonumber\\
&& + \frac{\gamma'(\xi_t) \{f(t)\bar{g}(t; h_n)- g(t)\bar{f}(t; h_n)\}\{ \bar{f}(t; h_n)-E\bar{f}(t; h_n) \}}{f(t)(E\bar{f}(t; h_n))^2 }\nonumber\\
&:=& \frac{J_1}{f(t)E\bar{f}(t; h_n) }+ \frac{J_2}{f(t)(E\bar{f}(t; h_n))^2}.
\end{eqnarray*}
Since $E \bar{f}(t; h_n)=f(t)(1+o(1))$, then
\begin{eqnarray}\label {51svb}
E \left \{ \bar{r}(t; h_n)-r(t) \right \}^2 = \left (\frac{E J_1^2}{f^4(t) }+ \frac{2 E (J_1 J_2)}{f^5(t) } + \frac{E J_2^2}{f^6(t) }\right)(1+o(1)).
\end{eqnarray}
Recall that $F_i=K(W_i)\alpha(q(X_i))$. Then
\begin{eqnarray}\label {45cd}
EJ_1^2&=& \frac{1}{n^2 h_n^2}E \sum_{i \neq j} \left\{f(t) F_i Y_i-g(t) F_i \right\} \left\{f(t) F_j Y_j-g(t) F_j \right\} \nonumber\\
&& + \frac{1}{n^2 h_n^2}E \sum_{i=1}^n \left\{f(t) F_i Y_i-g(t) F_i \right\}^2.
\end{eqnarray}
By (\ref{37svb}),
\begin{eqnarray}\label {46cd}
\frac{1}{h_n}E \left\{f(t) F_1 Y_1-g(t) F_1 \right\}&=& f(t) E \bar{g}(t; h_n) -g(t) E \bar{f}(t; h_n)\nonumber\\
&=& f^2(t) \theta(t)h^4_n +o(h^4_n).
\end{eqnarray}
Recall that $m(t)=E(Y_1^2|X_1=t)$ and $\sigma^2(t)=m(t)-r^2(t)$. By Proposition \ref{bias0},
\begin{eqnarray}\label {47cd}
&&\frac{1}{h_n}E \left\{f(t) F_1 Y_1-g(t) F_1 \right\}^2 \nonumber\\
&=& \frac{f^2(t)}{h_n}E (F_1 Y_1)^2 - \frac{2 f(t)g(t)}{h_n}E F^2_1Y_1 +\frac{g^2(t)}{h_n} E F_1^2 \nonumber\\
&=& f^2(t) \sqrt{q(t)} m(t)f(t) \mu_{0,2}- 2 f(t)g(t) \sqrt{q(t)}g(t)\mu_{0,2}\nonumber\\
&& + g^2(t) \sqrt{q(t)} f(t)\mu_{0,2}+O(h_n^2)\nonumber\\
&=& f^3(t) \sqrt{q(t)} \sigma^2(t) \mu_{0,2}+O(h^2_n).
\end{eqnarray}
Applying (\ref{46cd}) and (\ref{47cd}) to (\ref{45cd}), we have
\begin{eqnarray} \label {55svb}
EJ_1^2= \frac{f^{7/2}(t)|r'(t)|^{1/4} \sigma^2(t) \mu_{0,2}}{n h_n} +f^4(t)\theta^2(t)h^8_n +o(h^8_n)+o\left ( \frac{1}{nh_n}\right).
\end{eqnarray}
By H\a"{o}lder's inequality,
\begin{eqnarray}\label {52wb}
E J^2_2 \leq \|\gamma'(\cdot)\|_{\infty}^2 (E J_1^4)^{1/2} \left ( E \left\{ \bar{f}(t; h_n)-E\bar{f}(t; h_n) \right\}^4\right)^{1/2}.
\end{eqnarray}
Similar to the estimate in (\ref{45cd})-(\ref{47cd}),
\begin{eqnarray}\label {53wb}
&&E J_1^4=\frac{1}{n^4 h_n^4}E \sum_{i_1, i_2, i_3, i_4 \ are \ different} \prod_{k=1}^4 \left\{f(t) F_{i_k} Y_{i_k}-g(t) F_{i_k} \right\}\nonumber\\
&& + \frac{1}{n^4 h_n^4}E\sum_{i_1, i_2, i_3 \ are \ different} \left\{f(t) F_{i_1} Y_{i_1}-g(t) F_{i_1} \right\}^2\prod_{k=2}^3 \left\{f(t) F_{i_k} Y_{i_k}-g(t) F_{i_k} \right\} \nonumber\\
&&+\frac{1}{n^4 h_n^4}E \sum_{i \neq j}\left\{f(t) F_i Y_i-g(t) F_i \right\}^2 \left\{f(t) F_j Y_j-g(t) F_j \right\}^2\nonumber\\
&&+ \frac{1}{n^4 h_n^4}E \sum_{i \neq j} \left\{f(t) F_i Y_i-g(t) F_i \right\}^3 \left\{f(t) F_j Y_j-g(t) F_j \right\}\nonumber\\
&&+ \frac{1}{n^4 h_n^4}E \sum_{i =1}^n \left\{f(t) F_i Y_i-g(t) F_i \right\}^4 \nonumber\\
&=&f^8(t) \theta^4 (t)h^{16}_n +o(h^{16}_n) + O \left (\frac{h_n^8}{n h_n} \right )+O \left ( \frac{1}{n^2 h_n^2} \right ).
\end{eqnarray}
Similar to (\ref{Ebarf}),
\begin{eqnarray}\label {54wb}
E \left (\bar{f}(t; h_n) - E \bar{f}(t; h_n)\right)^4= O \left ( \frac{1}{n^2 h^2_n}\right ).
\end{eqnarray}
Applying (\ref{53wb}) and (\ref{54wb}) to (\ref{52wb}), we have
\begin{eqnarray} \label {55wa}
E J_2^2 = o(h^8_n)+ o \left (\frac{1}{n h_n} \right ).
\end{eqnarray}
By (\ref{55svb}) and (\ref{55wa}), and by H\a"{o}lder's inequality, $|E(J_1J_2)| = o(h^8_n)+o(1/(n h_n))$. Then by (\ref{51svb}), (\ref{55svb}) and (\ref{55wa}), we obtain (\ref{Thm1(2)}).
\end{proof}
\bigskip
\begin{proof}[Proof of Theorem \ref{central1}]
By (\ref{Thm1(1)}) and (\ref{Thm1(2)}),
\begin{eqnarray*}
E \bar{r}(t; h_n) - r(t)= \theta(t) h_n^4 +o(h_n^4)+o\left (\frac{1}{n h_n} \right )
\end{eqnarray*}
and
\begin{eqnarray*}
Var\left ( \bar{r}(t; h_n)- r(t) \right )= \frac{\mu_{0,2}|r'(t)|^{1/4}\sigma^2(t)}{ nh_n \sqrt{f(t)}} +o(h_n^8)+o\left (\frac{1}{n h_n} \right).
\end{eqnarray*}
If $h_n^4 \sqrt{n h_n} \rightarrow \lambda$, then by the Lindeberg's central limit theorem,
\begin{eqnarray*}
\sqrt{n h_n} \left \{\bar{r}(t; h_n)-r(t) \right \}\xrightarrow {D}N \left (\lambda \theta(t), \frac{\mu_{0,2}|r'(t)|^{1/4}\sigma^2(t)}{\sqrt{f(t)}} \right ).
\end{eqnarray*}
\end{proof}
To prove Theorem \ref{Thm2}, we first establish the following two lemmas. Note that $\hat{q}(t; h_n)$ is a function of all observations $X_1,..., X_n$.
\begin{lemma}\label {Lemma1}
Under the condition of Theorem \ref{Thm2}, for any integer $a \geq 1$,
\begin{eqnarray*}\label {46vbaa}
&& E \left \{\left (\hat{q}^4 (X_1; h_{1,n})- q^4 (X_1 ) \right) \big |X_1,...,X_a \right \} \nonumber\\
&=& \Psi(X_1, h_{1,n}) h_{1,n}^2+ \frac{\phi(X_1)}{n h^3_{1,n}} +O\left ( \frac{1}{n h^2_{1,n}}\right)+o(h^6_{2,n})
\end{eqnarray*}
for some functions $\Psi$ and $\phi$, where $\Psi(x, h_{1,n})=b(x)+d(x)h^2_{1,n}$ for some functions $b(x)$ and $d(x)$.
\end{lemma}
\begin{proof}
Fix $X_1=x_1$. Since $\hat{r}(x_1; h_{1,n})=\hat{g}(x_1;h_{1,n})/\hat{f}(x_1; h_{1,n}) $, then by (\ref{hatq}),
\begin{eqnarray*}
\hat{q}^4(x_1; h_{1,n})&=& \left \{\hat{f}^2(x_1; h_{1,n})\hat{r}'(x_1; h_{1,n}) \right \}^2 \nonumber\\
&=& \left \{ \hat{f}(x_1; h_{1,n})\hat{g}'(x_1; h_{1,n}) - \hat{g}(x_1; h_{1,n}) \hat{f}'(x_1; h_{1,n})\right \}^2,
\end{eqnarray*}
where
\begin{eqnarray*}
\hat{g}'(x_1; h_{1,n}) :=\frac{d\hat{g}(x; h_{1,n}) }{dx}\Big|_{x=x_1} = \frac{1}{n h^2_{1,n}}\sum_{i=1}^n K'\left ( \frac{x_1-X_i}{h_{1,n}} \right ) Y_i, \label {VB58aw} \\
\hat{f}'(t; h_{1,n}):=\frac{d\hat{f}(x_1; h_{1,n}) }{dt}\Big|_{x=x_1} = \frac{1}{n h^2_{1,n}}\sum_{i=1}^n K'\left ( \frac{x_1-X_i}{h_{1,n}} \right ) . \label {VB59aw}
\end{eqnarray*}
For $m \in [1,n]$, denote
\begin{eqnarray}\label {Nts}
U^{(m)}_i &=& (X_m-X_i)/h_{1,n} \nonumber\\
G^{(m)}_{i,j} &=&K(U^{(m)}_i) K'(U^{(m)}_j)Y_j - K(U^{(m)}_i)Y_i K'(U^{(m)}_j) \nonumber\\
H^{(m)}_{i,j,k,l}&=&G^{(m)}_{i,j}G^{(m)}_{k,l} - h^6_{1,n} q^4(X_m)
\end{eqnarray}
Then $G^{(m)}_{i,i}=0$ for any $m,i \in [1,n]$, and
\begin{eqnarray}\label {64st}
\hat{q}^4(x_1; h_{1,n})-q^4(x_1) &=&\frac{1}{n^4 h^6_{1,n}}\sum_{ i,j,k, l } \left (G^{(1)}_{i,j}G^{(1)}_{k,l} -h^6_{1,n}q^4(x_1)\right)\nonumber\\
&=& \frac{1}{n^4 h^6_{1,n}}\sum_{ i,j,k, l } H^{(1)}_{i,j,k,l} .
\end{eqnarray}
Denote $E^*$ as the expectation of $\{ X_{a+1},..., X_n\}$. Write
\begin{eqnarray}\label {E*}
E^*[ \hat{q}^4(x_1; h_{1,n})-q^4(x_1)] =J_1+J_2+J_3,
\end{eqnarray}
where
\begin{eqnarray}
J_1&=&E^* \Bigg(\frac{1}{n^4 h^6_{1,n} } \sum_{1, i, j, k, l \;\;are \;\;different } H^{(1)}_{i,j, k, l} \Bigg), \nonumber \\
J_2&=& E^* \Bigg(\frac{1}{n^4 h^6_{1,n} } \sum_{ exactly \ two \ of \ 1, i, j, k, l \ are \ equal } H^{(1)}_{i,j, k , l} \Bigg), \nonumber\\
J_3&=& E^* \Bigg(\frac{1}{n^4 h^6_{1,n} } \sum_{\substack{ other \ cases }} H^{(1)}_{i,j, k , l} \Bigg). \nonumber
\end{eqnarray}
First we estimate $E^* G^{(1)}_{i,j}/h^3_{1,n}$ for $i \neq j$ and $i,j \in [a+1, n]$. Denote $w_j=(x_1-x_j)/h_{1,n}$. Then
\begin{eqnarray}\label {63vbag}
\frac{1}{h_{1,n}^3} E^* G^{(1)}_{i,j}= \frac{1}{h_{1,n}^3} \int K'(w_3) \int K(w_2) (r(x_3)-r(x_2))f(x_2)dx_2 f(x_3)dx_3.
\end{eqnarray}
Since $g(x_2)=r(x_2)f(x_2)$, by Proposition \ref{bias0rx} with $\xi(s)=1$,
\begin{eqnarray}\label {64vbag}
&& \frac{1}{h_{1,n}}\int K(w_2) (r(x_3)-r(x_2))f(x_2)dx_2 \nonumber\\
&&=r(x_3) \left \{f(x_1) +\frac{f''(x_1)\mu_{2,1}}{2} h^2_{1,n} \right \}\nonumber\\
&& -\left \{ g(x_1)+\frac{g''(x_1)\mu_{2,1}}{2}h^2_{1,n} \right\}+O(h_{1,n}^4). \qquad
\end{eqnarray}
By (\ref{63vbag}) and (\ref{64vbag}),
\begin{eqnarray}\label {65vbag}
&&\frac{1}{h_{1,n}^3} E^* G^{(1)}_{i,j}=\left \{f(x_1) +\frac{f''(x_1)\mu_{2,1}}{2} h^2_{1,n} \right \}\frac{1}{h_{1,n}^2}\int K'(w_3)g(x_3)dx_3 \nonumber\\
&& - \left \{ g(x_1)+\frac{g''(x_1)\mu_{2,1}}{2}h^2_{1,n} +O(h_{1,n}^4)\right\}\frac{1}{h_{1,n}^2}\int K'(w_3)f(x_3)dx_3 . \qquad
\end{eqnarray}
Note that $\int K'(w)w^{2k}dw=0$ for $k =0, 1,\cdots$, $\int K'(w)wdw=-1$ and $\int K'(w)w^3 dw=-3\mu_{2,1}$. Then by Taylor expansion,
\begin{eqnarray}\label {66vbag}
&& \frac{1}{h_{1,n}^2}\int K'(w_3)g(x_3)dx_3 \nonumber\\
&=& \frac{1}{h_{1,n}}\int K'(w_3) \bigg \{g(x_1)-g'(x_1)w_3 h_{1,n} +\frac{g''(x_1)}{2}w_3^2 h_{1,n}^2\nonumber\\
&&-\frac{g'''(x_1)}{6} w^3_{3}h_{1,n}^3 +\frac{g^{(4)}(x_1)}{24}w_3^4 h_{1,n}^4 \bigg \}dw_3 +O(h_{1,n}^4) \nonumber\\
&=&g'(x_1)+\frac{g'''(x_1)\mu_{2,1}}{2}h_{1,n}^2 +O(h_{1,n}^4).
\end{eqnarray}
Similarly,
\begin{eqnarray}\label{67vbag}
\frac{1}{h_{1,n}^2}\int K'(w_3)f(x_3)dx_3= f'(x_1)+\frac{f'''(x_1)\mu_{2,1}}{2}h_{1,n}^2 +O(h_{1,n}^4).
\end{eqnarray}
Applying (\ref{66vbag}) and (\ref{67vbag}) to (\ref{65vbag}), we have, for some function $b_0$,
\begin{eqnarray*}\label {64vbr}
\frac{1}{h_{1,n}^3} E^* G^{(1)}_{i,j}&=& f(x_1)g'(x_1)-g(x_1)f'(x_1)+b_0(x_1)h^2_{1,n}+O(h_{1,n}^4). \qquad
\end{eqnarray*}
Since $q^4(x_1)=( f(x_1)g'(x_1)-g(x_1)f'(x_1))^2$, then for some function $b$,
\begin{eqnarray}\label {64tt}
\left ( \frac{1}{ h^3_{1,n}}E^* G^{(1)}_{i,j} \right)^2 - q^4(x_1) = b (x_1)h^2_{1,n}+ O(h^4_{1,n}).
\end{eqnarray}
If we apply the above Taylor expansion further, $O(h_{1,n}^4)$ in (\ref{64tt}) can be expressed as $d(x_1)h^4_{1,n}+O(h^6_{1,n})$ for some funtion $d$. Then \\
\begin{eqnarray} \label {65vbr}
\left ( \frac{1}{ h^3_{1,n}}E^*G^{(1)}_{i,j} \right)^2 - q^4(x_1)= \Psi(x_1, h_{1,n})h^2_{1,n} +O(h^6_{1,n})
\end{eqnarray}
where $\Psi(x_1, h_{1,n})=b(x_1) +d(x_1)h^2_{1,n}$.
Next we estimate $J_1$ in (\ref{E*}) with three cases. Case 1: $i,j,k,l \in [a+1, n]$. By (\ref{65vbr}),
\begin{eqnarray*} \label {57vt}
&&E^* \Bigg(\frac{1}{n^4 h^6_{1,n} } \sum_{\substack{ i,j,k, l \in [a+1, n] \\ 1, i, j, k, l \;\;are \;\;different }} H^{(1)}_{i,j, k, l} \Bigg) \nonumber\\
&&= \Psi(x_1, h_{1,n}) h^2_{1,n}+O(h^6_{1,n}) + O \left ( \frac{1}{n}\right).\qquad
\end{eqnarray*}
Case 2: One of $i,j,k,l$ is in $[2,a]$. If $j \in [2,a]$, for example, then
\begin{eqnarray}\label {72vt}
&& \frac{1}{n^4 h^6_{1,n}} E^* \Bigg(\sum_{\substack{i,k,l \in [a+1, n], \ j \in [2,a]\\ 1, i,k, l \ are \ different}}H^{(1)}_{i,j,k,l} \Bigg) \nonumber\\
&=& \frac{ K'(\frac{x_1-X_j}{h_{1,n}})r(X_j)E^* K(U^{(1)}_i) -K'(\frac{x_1-X_j}{h_{1,n}})E^* \left \{K(U^{(1)}_i)r(X_i) \right \}}{n h^3_{1,n}}\times \frac{E^* G^{(1)}_{k,l}}{h^3_{1,n}} -\frac{q^4(x_1)}{n}\nonumber
\end{eqnarray}
where $E^* K(U^{(1)}_i)/h_{1,n}=E^*\hat{f}(x_1; h_{1,n})$ and $E^* \left \{K(U^{(1)}_i)r(X_i) \right\}/h_{1,n}=E^*\hat{g}(x_1, h_{1,n})$. Hence
\begin{eqnarray*}
\frac{1}{n^4 h^6_{1,n}} E^* \Bigg(\sum_{\substack{i,k,l \in [a+1, n], \ j \in [2,a]\\ i,k, l \ are \ different}}H^{(1)}_{1, i,j,k,l} \Bigg)= O\left ( \frac{1}{n h^2_{1,n}}\right) .
\end{eqnarray*}
If $i,k$ or $l$ is in $[2,a]$, the results are similar. Hence
\begin{eqnarray*} \label {65ww}
E^* \Bigg(\frac{1}{n^4 h^6_{1,n} } \sum_{\substack{one \ of \ i,j,k, l \ in \ [2, a] \\1, i, j, k, l \;\;are \;\;different }} H^{(1)}_{i,j, k, l} \Bigg) =O\left ( \frac{1}{n h^2_{1,n}}\right).
\end{eqnarray*}
Case 3: Two or more of $i,j,k,l$ are in $[2,a]$. The result is $O(1/(n^2 h^4_{1,n}))=o(h_{2,n}^{10})$. Combining the three cases, we have
\begin{eqnarray}\label {E*1}
J_1 = \Psi(x_1, h_{1,n}) h_{1,n}^2+o(h^6_{2,n}) +O\left ( \frac{1}{n h^2_{1,n}}\right).
\end{eqnarray}
Now we estimate $J_2$ with two cases. Case 1: Two of $i,j,k,l$ are equal, and $i,j,k,l \in [a+1, n]$. For example, if $j=l$ and $i,j,k \in [a+1, n]$ are different, then
\begin{eqnarray*}\label {69st}
&& \frac{1}{n^4 h^6_{1,n}}E^* \Bigg(\sum_{\substack{i,j,k \in [a+1, n]\\ i , j, k \ are \ different}}H^{(1)}_{i,j,k,j}\Bigg)\nonumber\\
&=& \frac{1}{n h^6_{1,n}} E^* \left \{K(U^{(1)}_i)K'(U^{(1)}_j) (Y_j-Y_i)K(U^{(1)}_k)K'(U^{(1)}_j) (Y_j-Y_k)\right\}\nonumber\\
&&- \frac{q(x_1)}{n}+O\left( \frac{1}{n}\right)\qquad
\end{eqnarray*}
where the first term on the right-hand side turns out to be a function of $x_1$ multiplied by $1/(n h^3_{1,n})$. The other situations have similar results. Case 2: Two of $i,j,k,l$ are equal, and some $i,j,k,l$ are in $[1,a]$. The result is $O(1/(n^2 h^6_{1,n}))=o(h_{2,n}^{8})$. Then
\begin{eqnarray}\label {E*2}
J_2 = \frac{\phi(x_1)}{n h^3_{1,n}}+ o(h_{2,n}^{8}).
\end{eqnarray}
for some function $\phi$. Also
\begin{eqnarray}\label {E*3}
J_3=O\left ( \frac{1}{n^2 h^4_{1,n}} \right)=o(h_{2,n}^{10}).
\end{eqnarray}
Applying (\ref{E*1}), (\ref{E*2}) and (\ref{E*3}) to (\ref{E*}), we obtain the lemma.
\end{proof}
\bigskip
\begin{lemma}\label {Lemma2}
Denote $W_i=(t-X_i)\alpha(q(X_i))/h_{2,n}$ and let $L(w)= K(w)+w K'(w)$. Assume that $f'/f$ and $r''/r'$ are bounded in a neighborhood of $t$.
For any function $H$ with bounded and continuous second order derivtive,
\begin{eqnarray*}
\frac{1}{h_{2,n}} E \left \{L(W_1)H(X_1) \right \} = O(h^{2}_{2,n}).
\end{eqnarray*}
\end{lemma}
\begin{proof}
We have
\begin{eqnarray*}\label {98Vt}
\frac{1}{h_{2,n}} E \left \{L(W_1)H(X_1) \right \} =\frac{1}{h_{2,n}} \int L \left ( (t-s)\alpha(q(s))/h_{2,n} \right ) H(s) ds.
\end{eqnarray*}
Let $U_t(v)=v\alpha(q(t-v))$. By a similar argument as in Remark \ref{xicon}, $U'_t(v)=\alpha(q(t-v))-v\alpha'(q(t-v))q'(t-v)>0$ in a small neighborhood of $0$. Then $U_t(v)$ is invertible in a small neighborhood of $0$. Denote $V_t(u)$ as the inverse function of $U_t(v)$. Let $zh_{2,n}= (t-s)\alpha (q(s))= (t-s)\alpha (q(t-(t-s)))$. Then $z h_{2,n}=U_t(t-s)$ and hence $t-s = V_t(zh_{2,n})$. The change of variables from $s$ to $z$ gives
\begin{eqnarray*}\label{G}
\frac{1}{h_{2,n}} E L(W_1) H(X_1)&=& - \int L \left ( z \right ) H \left (t- V_t(zh_{2,n})\right) \frac{dV_t}{du}_{u=zh_{2,n}}dz \nonumber\\
&:=& - \int L \left ( z \right ) G_t(zh_{2,n})dz.
\end{eqnarray*}
By Taylor expansion, $G_t(zh_{2,n})=G_t(0)+G_t'(0)zh_{2,n}+ G_t''(\tau)(z^2h_{2,n}^2)/2$ where $\tau$ is between $0$ and $zh_{2,n}$. By the condition on $H$, $G_t$ has bounded and continuous second order derivtive. Also notice that $\int L(z)dz=0$, $\int z L(z)dz=0$. Then
\begin{eqnarray*}\label{Lq'}
\frac{1}{h_{2,n}} E\left \{ L(W_1) H (X_1)\right\} = O(h^2_{2,n}).
\end{eqnarray*}
\end{proof}
\begin{proof}[Proof of Theorem \ref{Thm2}]
By (\ref{VB16ag}) with $p=1$ and $z=\hat{f}(t; h_{1,n}, h_{2,n})/E \hat{f}(t; h_{1,n}, h_{2,n})$,
\begin{eqnarray} \label {75vbt}
E \hat{r} (t; h_{1,n}, h_{2,n}) = \frac{E \hat{g}(t; h_{1,n}, h_{2,n})}{E \hat{f}(t; h_{1,n}, h_{2,n})} +\frac{-M_1+M_2}{(E \hat{f} (t; h_{1,n}, h_{2,n}))^2}
\end{eqnarray}
where
\begin{eqnarray}\label {I1}
M_1=E\left \{\hat{g}(t; h_{1,n}, h_{2,n}) \left (\hat{f}(t; h_{1,n}, h_{2,n})-E \hat{f}(t; h_{1,n}, h_{2,n})\right) \right\}
\end{eqnarray}
and
\begin{eqnarray}\label {I2}
M_2=E \left \{\hat{r}(t; h_{1,n}, h_{2,n}) \left (\hat{f}(t; h_{1,n}, h_{2,n})-E \hat{f}(t; h_{1,n}, h_{2,n})\right)^2 \right\}.
\end{eqnarray}
\textbf{Step 1}. We estimate $E \hat{g}(t; h_{1,n}, h_{2,n})/E\hat{f}(t; h_{1,n}, h_{2,n})$. Define the function $\beta(y)=\alpha( y^{1/4})$ for $y> 0$. By Taylor expansion,
\begin{eqnarray}\label {65vbc}
\alpha (\hat{q}(X_i; h_{1,n})) -\alpha(q(X_i))&= & \beta(\hat{q}^4(X_i; h_{1,n}))-\beta(q^4(X_i))\nonumber\\
&=& \beta'(q^4(X_i)) \left \{ \hat{q}^4(X_i; h_{1,n}) - q^4(X_i) \right \} \nonumber\\
&&+\frac{\beta''(\hat{\nu_i})}{2} \left \{\hat{q}^4(X_i; h_{1,n}) -q^4(X_i) \right \}^2 \qquad
\end{eqnarray}
where $\hat{\nu_i}$ is between $\hat{q}^4(X_i; h_{1,n})$ and $q^4(X_i)$. Denote $W_i=(t-X_i) \alpha (q(X_i))/h_{2,n}$. By Taylor expansion and (\ref{65vbc}),
\begin{eqnarray} \label {66vbc}
&& K \left ( \frac{t-X_i}{h_{2,n}}\alpha (\hat{q}(X_i; h_{1,n})) \right ) \nonumber\\
&=&K(W_i)+\sum_{k=1}^2 \frac{K^{(k)} \left (W_i\right) }{k!} \left (\frac{t-X_i}{h_{2,n}}\left \{ \beta(\hat{q}^4(X_i; h_{1,n}))-\beta(q^4(X_i)) \right \}\right )^k\nonumber\\
&& + \frac{K''' (\hat{\xi}_i ) }{6} \left (\frac{t-X_i}{h_{2,n}}\left \{ \beta(\hat{q}^4(X_i; h_{1,n}))-\beta(q^4(X_i)) \right \}\right )^3 \nonumber\\
&=& K \left (W_i\right)+K' \left (W_i\right) \frac{t-X_i}{h_{2,n}}\beta'(q^4(X_i))\left \{ \hat{q}^4(X_i; h_{1,n}) - q^4(X_i) \right \} + \hat{\delta}_{i}
\end{eqnarray}
where $\hat{\xi}_i$ is between $ (t-X_i)\alpha( \hat{q}(X_i; h_{1,n}) )/h_{2,n}$ and $W_i$, and
\begin{eqnarray}\label{delta}
\hat{\delta}_{i}&=&K' \left (W_i\right) \frac{t-X_i}{h_{2,n}}\Bigg \{ \frac{1}{2}\beta''(\hat{\nu}_i) \left \{ \hat{q}^4(X_i; h_{1,n}) - q^4(X_i) \right \}^2 \Bigg \} \nonumber\\
&&+ \frac{K''\left (W_i\right) }{2} \left (\frac{t-X_i}{h_{2,n}}\left \{ \beta(\hat{q}^4(X_i; h_{1,n}))-\beta(q^4(X_i))\right \}\right )^2\nonumber\\
&& + \frac{K''' (\hat{\xi}_i ) }{6} \left (\frac{t-X_i}{h_{2,n}}\left \{ \beta(\hat{q}^4(X_i; h_{1,n}))-\beta(q^4(X_i))\right \}\right )^3.
\end{eqnarray}
Let $L(w)= K(w)+w K'(w)$. Then by (\ref{65vbc}) and (\ref{66vbc}),
\begin{eqnarray}\label {71vbd}
&& K \left ( \frac{t-X_i}{h_{2,n}}\alpha (\hat{q}(X_i; h_{1,n})) \right )\alpha (\hat{q}(X_i; h_{1,n})) \nonumber\\
&=&K \left (W_i\right) \alpha( q(X_i))+ L(W_i) \beta'(q^4(X_i)) \left \{\hat{q}^4(X_i; h_{1,n}) -q^4(X_i) \right \}+ \hat{\eta}_{i} \qquad
\end{eqnarray}
where
\begin{eqnarray}\label {83vbs}
\hat{\eta}_{i}&=& K(W_i) \left \{ \frac{1}{2}\beta''(\hat{\nu}_i) \left \{ \hat{q}^4(X_i; h_{1,n}) - q^4(X_i) \right \}^2 \right \} \nonumber\\
&& + K' \left (W_i\right) \frac{t-X_i}{h_{2,n}}\beta'(q^4(X_i))\left \{ \hat{q}^4(X_i; h_{1,n}) - q^4(X_i) \right \} \nonumber\\
&&\times\left \{\beta(\hat{q}^4(X_i; h_{1,n}))-\beta(q^4(X_i))\right \} \nonumber\\
&&+\hat{\delta}_i \alpha (\hat{q}(X_i; h_{1,n})).
\end{eqnarray}
Similar to the proof of Lemma \ref{Lemma1}, by (\ref{delta}), (\ref{83vbs}), and using the second equation of (\ref{65vbc}),
\begin{eqnarray}\label {81ce}
\frac{1}{ h_{2,n}} E \left\{ E (\hat{\eta}_{1}Y_1 |X_1)\right\}= o(h^4_{2,n}) .
\end{eqnarray}
Let
\begin{eqnarray} \label {lgxi}
\hat{\theta}_i = L(W_i) \beta'(q^4(X_i))\left \{\hat{q}^4(X_i; h_{1,n}) -q^4(X_i) \right \}+ \hat{\eta}_{i} .
\end{eqnarray}
By (\ref{71vbd}),
\begin{eqnarray}\label {94Eg}
\hat{g}(t; h_{1,n}, h_{2,n})= \bar{g}(t; h_{2,n})+ \frac{1}{n h_{2,n}} \sum_{i=1}^n \hat{\theta}_i Y_i
\end{eqnarray}
and
\begin{eqnarray}\label {123vbaw}
\hat{f}(t; h_{1,n}, h_{2,n})= \bar{f}(t; h_{2,n})+ \frac{1}{n h_{2,n}} \sum_{i=1}^n \hat{\theta}_i .
\end{eqnarray}
By Lemma \ref{Lemma1} and Lemma \ref{Lemma2},\\
\begin{eqnarray} \label {91pa}
&& \frac{1}{h_{2,n}}E \left \{ L(W_1) \beta'(q^4(X_1))Y_1 E\left( \left \{\hat{q}^4(X_1; h_{1,n}) -q^4(X_1) \right \}|X_1 \right) \right\}\nonumber\\
&=& \frac{1}{h_{2,n}}E \left \{L(W_1)\beta'(q^4(X_1)) r(X_1) \left (\Psi(X_1, h_{1,n}) h^2_{1,n} +\frac{\phi(X_1)}{n h^3_{1,n}}+ O\left ( \frac{1}{n h^2_{1,n}} \right)+o(h^6_{2,n})\right) \right\} \nonumber\\
&=& O (h^2_{1,n}h^2_{2,n})+\frac{O(h^2_{2,n})}{n h^3_{1,n}}\nonumber\\
&&+\left \{ O \left ( \frac{1}{n h^2_{1,n}}\right )+o(h^6_{1,n})\right\}\frac{1}{h_{2,n}}E \left | L(W_1)\beta'(q^4(X_1)) r(X_1)\right|\nonumber\\
&=& o(h^4_{2,n}).
\end{eqnarray}
Applying (\ref{81ce}) and (\ref{91pa}) to (\ref{lgxi}), we have
\begin{eqnarray}\label {92pa}
\frac{1}{ h_{2,n}}E \{E ( \hat{\theta}_1 Y_1 |X_1) \} = o(h^4_{2,n}).
\end{eqnarray}
By (\ref{94Eg}) and (\ref{92pa}),
\begin{eqnarray}\label {94vbt}
E \hat{g}(t; h_{1,n}, h_{2,n})= E \bar{g}(t; h_{2,n})+ o(h^4_{2,n}).
\end{eqnarray}
Similarly,
\begin{eqnarray}\label {95vbt}
E \hat{f}(t; h_{1,n}, h_{2,n})= E \bar{f}(t; h_{2,n})+ o(h^4_{2,n}) .
\end{eqnarray}
By (\ref{VB21a}) in the proof of Theorem \ref{Thm01},
\begin{eqnarray}\label {96vbt}
\frac{E \hat{g}(t; h_{1,n}, h_{2,n})}{E \hat{f}(t; h_{1,n}, h_{2,n})} = \frac{E\bar{g}(t; h_{2,n})}{E\bar{f}(t; h_{2,n})}+o(h^4_{2,n})= r(t) + \theta(t) h^4_{2,n }+ o(h^4_{2,n}).
\end{eqnarray}
\textbf{Step 2}. We estimte $M_1$ in (\ref{I1}). Applying (\ref{94vbt}) and (\ref{95vbt}) to (\ref{I1}), we have
\begin{eqnarray}\label {I1a}
M_1 = E \left\{\hat{g}(t; h_{1,n}, h_{2,n}) \hat{f}(t; h_{1,n}, h_{2,n})\right\}- E \bar{g}(t; h_{2,n}) E \bar{f}(t; h_{2,n})+o(h^4_{2,n}) .
\end{eqnarray}
By (\ref{94Eg}) and (\ref{123vbaw}),
\begin{eqnarray}\label {124vbaw}
&& E\left\{\hat{g}(t; h_{1,n}, h_{2,n}) \hat{f}(t; h_{1,n}, h_{2,n})\right\}\nonumber\\
&=& E\left\{ \bar{g}(t; h_{2,n})\bar{f}(t; h_{2,n})\right\}+ E\left\{\bar{f}(t; h_{2,n})\frac{1}{n h_{2,n}} \sum_{i=1}^n \hat{\theta}_i Y_i\right\}\nonumber\\
&&+E \left \{\bar{g}(t; h_{2,n})\frac{1}{n h_{2,n}} \sum_{i=1}^n \hat{\theta}_i \right\} + E \left \{\frac{1}{n^2 h^2_{2,n}}\sum_{i=1}^n \hat{\theta}_iY_i \sum_{i=1}^n \hat{\theta}_i \right\}.
\end{eqnarray}
Similar to (\ref{92pa}),
\begin{eqnarray}\label {124Ld}
&& E\left\{\bar{f}(t; h_{2,n})\frac{1}{n h_{2,n}} \sum_{i=1}^n \hat{\theta}_i Y_i\right\} \nonumber\\
&=&\frac{1}{n^2 h^2_{2,n}} E \sum_{i\neq j}K(W_j)\alpha(q(X_j)) \hat{\theta}_iY_i + \frac{1}{n^2 h^2_{2,n}}E\sum_{i=1}^nK(W_i)\alpha(q(X_i)) \hat{\theta}_iY_i \nonumber\\
&=& o(h^4_{2,n}) .
\end{eqnarray}
Similarly, the third term on the right-hand side of (\ref{124vbaw}) is $o(h^4_{2,n})$. Note that
\begin{eqnarray}\label {90vt}
E \left \{\frac{1}{n^2 h^2_{2,n}}\sum_{i=1}^n \hat{\theta}_iY_i \sum_{i=1}^n \hat{\theta}_i \right\}= \frac{1}{n^2 h^2_{2,n}}E\sum_{i\neq j}\hat{\theta}_iY_i \hat{\theta}_j + \frac{1}{n^2 h^2_{2,n}}E\sum_{i}\hat{\theta}^2_iY_i .
\end{eqnarray}
For later use, we now show that it is $o(h^8_{2,n})+o(1/(n h_{2,n}))$. By (\ref{lgxi}) and letting $\hat{\lambda}_i=L(W_i) \beta'(q^4(X_i))\left \{\hat{q}^4(X_i; h_{1,n}) -q^4(X_i) \right \}$, we have
\begin{eqnarray}\label {101tt}
\frac{1}{n^2 h^2_{2,n}}E\sum_{i\neq j}\hat{\theta}_iY_i \hat{\theta}_j = \frac{1}{n^2 h^2_{2,n}}E\sum_{i\neq j}\left ( \hat{\lambda}_i Y_i + \hat{\eta}_{i}Y_i\right) \left ( \hat{\lambda}_j + \hat{\eta}_{j}\right).
\end{eqnarray}
By (\ref{Nts}) and (\ref{64st}),
\begin{eqnarray}\label {102tt}
\frac{1}{n^2 h^2_{2,n}}E\sum_{i\neq j} \hat{\lambda}_i Y_i \hat{\lambda}_j &=& \frac{1}{n^2 h^2_{2,n}}E\sum_{i\neq j}\bigg\{ L(W_i) \beta'(q^4(X_i))L(W_j) \beta'(q^4(X_j)) \nonumber\\
&& \times \frac{1}{n^4 h^6_{1,n}}\sum_{k, l , p, q} H^{(i)}_{k,l, p, q} \frac{1}{n^4 h^6_{1,n}}\sum_{k', l' , p', q'} H^{(j)}_{k',l', p', q'} \bigg \}.\qquad
\end{eqnarray}
Write
\begin{eqnarray}
\frac{1}{n^4 h^6_{1,n}}\sum_{k, l , p, q} H^{(i)}_{k,l, p, q} \frac{1}{n^4 h^6_{1,n}}\sum_{k', l' , p', q'} H^{(j)}_{k',l', p', q'}=T_1 +T_2 +T_3
\end{eqnarray}
where
\begin{eqnarray}
T_1 &=& \frac{1}{n^8 h^{12}_{1,n}}\sum_{i, j , k,l,p,q, k', l' , p', q' \ are \ different}H^{(i)}_{k,l, p, q} H^{(j)}_{k',l', p', q'}, \nonumber\\
T_2 &=& \frac{1}{n^8 h^{12}_{1,n}}\sum_{exactly \ two \ of \ i, j , k,l,p,q, k', l' , p', q' \ are \ equal} H^{(i)}_{k,l, p, q} H^{(j)}_{k',l', p', q'}, \nonumber\\
T_3 &=& \frac{1}{n^8 h^{12}_{1,n}}\sum_{Other \ cases} H^{(i)}_{k,l, p, q} H^{(j)}_{k',l', p', q'}.
\end{eqnarray}
Similar to the estimate of $J_1$ in Lemma \ref{Lemma1}, and by Lemma \ref{Lemma2},
\begin{eqnarray}\label {105tt}
&& \frac{1}{n^2 h^2_{2,n}}E\sum_{i\neq j}\left\{ L(W_i) \beta'(q^4(X_i))L(W_j) \beta'(q^4(X_j)) T_1 \right\} \nonumber\\
&=& \frac{1}{n^2 h^2_{2,n}}\sum_{i\neq j}E\left\{ L(W_i) \beta'(q^4(X_i)) \left ( \Psi(X_i, h_{1,n}) h_{1,n}^2\right ) \right\}\nonumber\\
&& \times E\left\{ L(W_j) \beta'(q^4(X_j)) \left ( \Psi(X_j, h_{1,n}) h_{1,n}^2 \right ) \right\}+o(h^8_{2,n})+O\left (\frac{1}{n} \right )\nonumber\\
&=& o(h^8_{2,n})+o\left ( \frac{1}{n h_{2,n}}\right ).
\end{eqnarray}
We consider two cases when replacing $T_1$ by $T_2$ in the above analysis. \\
Case 1: The two equal indices are both in $\{i, k, l, p, q \}$ or both in $\{ j, k',l', p', q'\}$. Suppose they are in $\{i, k, l, p, q \}$. Similar to the estimates of $J_2$ and $J_1$ in Lemma \ref{Lemma1}, the result is
\begin{eqnarray} \label {106tt}
&& \frac{1}{n^2 h^2_{2,n}}\sum_{i\neq j}E\left\{ L(W_i) \beta'(q^4(X_i)) \left ( \frac{\phi(X_i)}{n h^3_{1,n}} \right ) \right\}\nonumber\\
&& \times E\left\{ L(W_j) \beta'(q^4(X_j)) \left ( \Psi(X_j, h_{1,n}) h_{1,n}^2 \right ) \right\}+o(h^8_{2,n})+o\left ( \frac{1}{n h_{2,n}} \right )\nonumber\\
&=&o(h^8_{2,n})+o\left ( \frac{1}{n h_{2,n}} \right ).
\end{eqnarray}
Case 2: One of the equal indices is in $\{i, k, l, p, q \}$ and the other in $\{ j, k',l', p', q'\}$. Suppose that $k=k'$. In this case we first fix $X_k$ and take conditional expectation of the other variables, and then take expectation of $X_k$. The result is $O(h^4_{2,n}/(n h^3_{1,n}))=o(h^8_{2,n})$. Combining the two cases, we conclude
\begin{eqnarray*}
\frac{1}{n^2 h^2_{2,n}}E\sum_{i\neq j}\left\{ L(W_i) \beta'(q^4(X_i))L(W_j) \beta'(q^4(X_j)) T_2 \right\}=o(h^8_{2,n})+o\left ( \frac{1}{n h_{2,n}} \right ).
\end{eqnarray*}
If $T_1$ is replaced by $T_3$ in (\ref{105tt}), similar to the estimate of $J_3$ in Lemma \ref{Lemma1}, the result is $o(h^{10}_{2,n})$. Together with (\ref{102tt})-(\ref{106tt}), we have
\begin{eqnarray*}
\frac{1}{n^2 h^2_{2,n}}E\sum_{i\neq j} \hat{\lambda}_i Y_i \hat{\lambda}_j =o(h^8_{2,n})+o\left ( \frac{1}{n h_{2,n}} \right ).
\end{eqnarray*}
Similarly, the other terms in (\ref{101tt}) are also $o(h^8_{2,n})+o(1/(nh_{2,n}))$. Hence
\begin{eqnarray*}
\frac{1}{n^2 h^2_{2,n}}E\sum_{i\neq j}\hat{\theta}_iY_i \hat{\theta}_j = o(h^8_{2,n})+o\left ( \frac{1}{n h_{2,n}} \right ).
\end{eqnarray*}
Then by (\ref {90vt}),
\begin{eqnarray}\label {109tt}
E \left \{\frac{1}{n^2 h^2_{2,n}}\sum_{i=1}^n \hat{\theta}_iY_i \sum_{i=1}^n \hat{\theta}_i \right\}= o(h^8_{2,n})+o\left ( \frac{1}{n h_{2,n}} \right ) .
\end{eqnarray}
By (\ref{I1a})-(\ref{124Ld}) and (\ref{109tt}),
\begin{eqnarray*}
M_1 = E\left\{ \bar{g}(t; h_{2,n}) \bar{f}(t; h_{2,n})\right\}- E \bar{g}(t; h_{2,n}) E \bar{f}(t; h_{2,n})+o(h^4_{2,n})+ o\left ( \frac{1}{n h_{2,n}} \right ).
\end{eqnarray*}
Then by (\ref{I_1}) and (\ref{30cb}),
\begin{eqnarray}\label {128vbaw}
M_1= \frac{ \sqrt{q(t)}g(t)\mu_{0,2}}{n h_{2,n}} +o(h^4_{2,n}) +o\left ( \frac{1}{n h_{2,n}} \right).
\end{eqnarray}
\textbf{Step 3}. We estimate $M_2$ in (\ref{I2}). Let $\gamma(y)=(1+y)^{-1}$. By Taylor expansion,
\begin{eqnarray*}
\frac{1}{\hat{f}(t; h_{1,n}, h_{2,n})}= \frac{1}{E\hat{f}(t; h_{1,n}, h_{2,n}) }\left ( 1 + \gamma'(\hat{\rho}_t)\left (\frac{\hat{f}(t; h_{1,n}, h_{2,n})}{E \hat{f}(t; h_{1,n}, h_{2,n})} -1\right ) \right)
\end{eqnarray*}
where $\hat{\rho}_t$ is between $0$ and $\hat{f}(t; h_{1,n}, h_{2,n})/E\hat{f}(t; h_{1,n}, h_{2,n}) -1$. Hence
\begin{eqnarray}\label {122cb}
&&M_2 = \frac{1}{E \hat{f}(t; h_{1,n}, h_{2,n})}E \left \{\hat{g}(t; h_{1,n}, h_{2,n})\left(\hat{f}(t; h_{1,n}, h_{2,n})- E \hat{f}(t; h_{1,n}, h_{2,n})\right)^2\right\}\nonumber \\
&&+\frac{1}{(E \hat{f}(t; h_{1,n}, h_{2,n}))^2}E \left \{\gamma'(\hat{\rho}_t) \hat{g}(t; h_{1,n}, h_{2,n}) \left (\hat{f}(t; h_{1,n}, h_{2,n})- E \hat{f}(t; h_{1,n}, h_{2,n})\right)^3\right\} \nonumber\\
&&:= \frac{1}{E \hat{f}(t; h_{1,n}, h_{2,n})}M_{2,1}+\frac{1}{(E \hat{f}(t; h_{1,n}, h_{2,n}))^2}M_{2,2}.
\end{eqnarray}
By (\ref {94Eg}) and (\ref{123vbaw}),
\begin{eqnarray*}\label {101de}
M_{2,1}=M_{2,1,1}+M_{2,1,2}+M_{2,1,3} +M_{2,1,4}
\end{eqnarray*}
where
\begin{eqnarray}\label {99cu}
M_{2,1,1} &=& E\left\{\bar{g}(t; h_{2,n}) \left ( \bar{f}(t; h_{2,n}) -E\ \bar{f}(t; h_{2,n}) \right )^2\right\} \nonumber\\
M_{2,1,2}&=& 2 E \left \{\bar{g}(t; h_{2,n}) \left( \bar{f}(t; h_{2,n}) -E \bar{f}(t; h_{2,n}) \right)\frac{1}{n h_{2,n}}\sum_{i=1}^n (\hat{\theta}_i-E \hat{\theta}_i) \right\}\nonumber\\
M_{2,1,3}&=& E\left\{ \bar{g}(t; h_{2,n}) \left ( \frac{1}{n h_{2,n}}\sum_{i=1}^n (\hat{\theta}_i-E \hat{\theta}_i)\right)^2\right\}\nonumber\\
M_{2,1,4}&=& E\left \{ \frac{1}{n h_{2,n}}\sum_{i=1}^n \hat{\theta}_i Y_i \left( \left( \bar{f}(t; h_{2,n}) -E \bar{f}(t; h_{2,n}) \right)+\frac{1}{n h_{2,n}}\sum_{i=1}^n (\hat{\theta}_i-E \hat{\theta}_i)\right )^2\right\} . \nonumber
\end{eqnarray}
Note that $M_{2,1,1}$ is $I_{2,1}$ in (\ref{40vbb}) with $h_n$ replaced by $h_{2,n}$. Then by (\ref{37cba}),
\begin{eqnarray}\label {M211}
M_{2,1,1} = \frac{\sqrt{ q(t)}g(t)f(t)\mu_{0,2}}{n h_{2,n}} +o\left(\frac{1}{n h_{2,n}}\right).
\end{eqnarray}
By Lemma \ref{Lemma1} and Lemma \ref{Lemma2}, and similar to (\ref{92pa}),
$$E ( E \{ (\hat{\theta}_i-E \hat{\theta}_i)|X_i, X_j, X_k \}| X_j, X_k)=o(h^4_{2,n}).$$
Denote $F_i = K(W_i)\alpha(q(X_i)).$
Then
\begin{eqnarray}\label {106pa}
M_{2,1,2} &=& \frac{2}{n^3 h^3_{2,n}} \sum_{i , j, k \ are \ not \ equal} E\bigg \{F_k Y_k \left \{F_j -E F_j \right \}\times E \left \{ E (\hat{\theta}_i-E \hat{\theta}_i)| X_j, X_k \right \}\bigg\} \nonumber\\&& + o\left ( \frac{1}{n h_{2,n}}\right )\nonumber\\
&=& \frac{o(h^4_{2,n})}{h_{2,n}} E \left |F_k Y_k \right| \frac{1}{h_{2,n}} E\left |F_j-EF_j \right | + o\left ( \frac{1}{n h_{2,n}}\right )\nonumber\\
&=&o(h^4_{2,n})+ o\left ( \frac{1}{n h_{2,n}}\right ).
\end{eqnarray}
Similarly,
\begin{eqnarray}\label {M213}
M_{2,1,3}+M_{2,1,4}=o(h^4_{2,n})+ o\left ( \frac{1}{n h_{2,n}}\right ).
\end{eqnarray}
By (\ref{M211})-(\ref{M213}),
\begin{eqnarray}\label {109deh}
M_{2,1}= \frac{\sqrt{ q(t)}g(t)f(t)\mu_{0,2}}{n h_{2,n}}+ o(h^4_{2,n})+ o\left ( \frac{1}{n h_{2,n}}\right ).
\end{eqnarray}
For $M_{2,2}$ defined in (\ref{122cb}), by H\a"{o}lder's inequality,\\
\begin{eqnarray}\label {124tt}
|M_{2,2}| &\leq& \|\gamma'(\cdot)\|_{\infty} \left \{E \left( \hat{g}(t; h_{1,n}, h_{2,n})\right)^2\right\}^{1/2} \nonumber\\
&& \times \left \{E \left ( \bar{f}(t; h_{2,n}-E \bar{f}(t; h_{2,n})+\frac{1}{n h_{2,n}}\sum_{i=1}^n (\hat{\theta}_i-E \hat{\theta}_i)\right)^6\right \}^{1/2}. \qquad
\end{eqnarray}
By (\ref{Ebarf}),
\begin{eqnarray}\label {99vt}
E \left ( \bar{f}(t; h_{2,n}-E \bar{f}(t; h_{2,n}) \right )^6= O\left (\frac{1}{n^3 h^3_{2,n}} \right).
\end{eqnarray}
Similar to the estimate of (\ref{90vt}), \\
\begin{eqnarray}\label {100vt}
E \left ( \frac{1}{n h_{2,n}}\sum_{i=1}^n (\hat{\theta}_i-E \hat{\theta}_i)\right)^6= o(h^{8}_{2,n})+ o\left ( \frac{1}{n^2 h^2_{2,n}} \right).
\end{eqnarray}
Applying (\ref{99vt}) and (\ref{100vt}) to (\ref{124tt}),
\begin{eqnarray}\label{112de}
M_{2,2}= o (h^4_{2,n}) + o\left ( \frac{1}{n h_{2,n}} \right ).
\end{eqnarray}
Since $E \bar{f}(t; h_{1,n}, h_{2,n})=f(t)+o(1)$, by (\ref{122cb}), (\ref{109deh}) and (\ref{112de}),
\begin{eqnarray}\label {126cb}
M_2=\frac{\sqrt{q(t)}g(t)\mu_{0,2}}{n h_n} +o(h^4_{2,n})+ o\left ( \frac{1}{n h_{2,n}} \right ).
\end{eqnarray}
By (\ref{75vbt}), (\ref{96vbt}), (\ref{128vbaw}) and (\ref{126cb}), we obtain (\ref{Thm2(1)}). \\
Now we prove (\ref{Thm2(2)}). By the Taylor expansion of $1/\hat{f}(t; h_{1,n}, h_{2,n})$ at the beginning of Step 3,
\begin{eqnarray}\label {153svb}
&& \hat{r}(t; h_{1,n}, h_{2,n})- r(t)\nonumber\\
&=& \frac{f(t)\hat{g}(t; h_{1,n}, h_{2,n})- g(t)\hat{f}(t; h_{1,n}, h_{2,n}) }{f(t)\hat{f}(t; h_{1,n}, h_{2,n})} \nonumber\\
&=& \frac{N_1}{f(t) E\hat{f}(t; h_{1,n}, h_{2,n}) }+ \frac{N_2}{f(t) (E\hat{f}(t; h_{1,n}, h_{2,n}))^2 }
\end{eqnarray}
where
\begin{eqnarray}\label {J1}
N_1 &=& f(t)\hat{g}(t; h_{1,n}, h_{2,n})- g(t)\hat{f}(t; h_{1,n}, h_{2,n}) \nonumber\\
N_2 &=& \left \{ f(t)\hat{g}(t; h_{1,n}, h_{2,n})- g(t)\hat{f}(t; h_{1,n}, h_{2,n}) \right \}\nonumber\\
&& \times \left \{ \gamma'(\hat{\rho}_t) \left ( \hat{f}(t; h_{1,n}, h_{2,n}) -E\hat{f}(t; h_{1,n}, h_{2,n}) \right )\right \}. \label {J2}
\end{eqnarray}
Since $E\hat{f}(t; h_{1,n}, h_{2,n})=f(t)(1+o(1))$, then
\begin{eqnarray}\label {127et}
E \left \{\hat{r}(t; h_{1,n}, h_{2,n})- r(t) \right \}^2= \left (\frac{E N_1^2}{f^4(t) }+ \frac{2 E (N_1 N_2)}{f^5(t) } + \frac{E N_2^2}{f^6(t) }\right)(1+o(1)).
\end{eqnarray}
By (\ref{94Eg}) and (\ref{123vbaw}),
\begin{eqnarray}\label {117ua}
E N^2_1&=& E \left \{ f(t) \bar{g}(t; h_{2,n}) - g(t)\bar{f}(t; h_{2,n}) \right \}^2 + E \left \{ \frac{1}{n h_{2,n}}\sum_{i=1}^n \hat{\theta}_i \left \{f(t)Y_i -g(t)\right\} \right \}^2\nonumber\\
&& +2 E\left \{ f(t) \bar{g}(t; h_{2,n}) - g(t) \bar{f}(t; h_{2,n}) \right \} \left \{ \frac{1}{n h_{2,n}}\sum_{i=1}^n \hat{\theta}_i\left \{f(t)Y_i -g(t)\right\} \right \} \nonumber\\
&:=&E N_{1,1}+E N_{1,2}+2 E N_{1,3}.
\end{eqnarray}
By (\ref{55svb}),
\begin{eqnarray} \label {120eg}
E N_{1,1}= \frac{f^{7/2}(t)|r'(t)|^{1/4} \sigma^2(t) \mu_{0,2}}{n h_n} +f^4(t)\theta^2(t)h^8_n +o(h^8_n)+o\left ( \frac{1}{nh_n}\right).
\end{eqnarray}
Similar to the estimate of (\ref{90vt}) above,
\begin{eqnarray}\label {121eg}
E N_{1,2}= o(h^8_{2,n})+o\left (\frac{1}{n h_{2,n}} \right ).
\end{eqnarray}
By H\a"{o}lder's inequality, (\ref{120eg}) and (\ref{121eg}),
\begin{eqnarray}\label {120ua}
|E N_{1,3}| \leq (E N_{1,1})^{1/2}(E N_{1,2})^{1/2} = o(h^8_{2,n})+ o\left (\frac{1}{n h_{2,n}} \right ).
\end{eqnarray}
By (\ref{117ua})-(\ref{120ua}),
\begin{eqnarray}\label {J1a}
E N^2_1=\frac{f^{7/2}(t)|r'(t)|^{1/4} \sigma^2(t) \mu_{0,2}}{n h_n} +f^4(t)\theta^2(t)h^8_n +o(h^8_n)+o\left ( \frac{1}{nh_n}\right).
\end{eqnarray}
Next we estimate $EN^2_2$. By H\a"{o}lder's inequality,
\begin{eqnarray}\label {130vw}
E N^2_{2}\leq \|\gamma'(\cdot)\|^2_{\infty} \left ( E N_1^4 \right )^{1/2} \left ( E \left ( \hat{f}(t; h_{1,n}, h_{2,n}) -E\hat{f}(t; h_{1,n}, h_{2,n}) \right )^4 \right )^{1/2}.
\end{eqnarray}
By (\ref{94Eg}), (\ref{123vbaw}), (\ref{53wb}) and similar to the estimate of (\ref{90vt}),
\begin{eqnarray}\label {136vk}
E N_1^4 &=&E \left \{f(t) \bar{g}(t; h_{2,n}) -g(t) \bar{f}(t; h_{2,n} )+ \frac{1}{n h_{2,n}}\sum_{i=1}^n \hat{\theta}_i \left \{f(t)Y_i -g(t)\right\} \right \}^4\nonumber\\
&\leq&8 E \left \{ f(t) \bar{g}(t; h_{2,n} ) - g(t) \bar{f}(t; h_{2,n} ) \right \}^4 +8 E \left \{ \frac{1}{n h_{2,n}}\sum_{i=1}^n \hat{\theta}_i \left \{f(t)Y_i -g(t)\right\} \right \}^4\nonumber\\
&=& o(h^8_{2,n})+ o\left ( \frac{1}{n h_{2,n}} \right ).
\end{eqnarray}
By (\ref{123vbaw}), (\ref{54wb}) and similar to the estimate of (\ref{90vt}),
\begin{eqnarray}\label {137vk}
&&E \left ( \hat{f}(t; h_{1,n}, h_{2,n}) -E\hat{f}(t; h_{1,n}, h_{2,n}) \right )^4\nonumber\\
&\leq& 8 E \left \{\bar{f}(t; h_{2,n} )-E \bar{f}(t; h_{2,n} )\right \}^4+ 8 E \left \{ \frac{1}{n h_{2,n}}\sum_{i=1}^n (\hat{\theta}_i-E \hat{\theta}_i)\right \}^4 \nonumber\\
&=& o(h^8_{2,n})+ o\left ( \frac{1}{n h_{2,n}}\right).
\end{eqnarray}
Applying (\ref{136vk}) and (\ref{137vk}) to (\ref{130vw}), we have
\begin{eqnarray}\label {127eh}
E N^2_2= o(h^8_{2,n})+ o\left (\frac{1}{n h_{2,n}} \right ).
\end{eqnarray}
Since $|E (N_1 N_2)| \leq (E N_1^2)^{1/2}(E N_2^2)^{1/2}$, then applying (\ref{J1a}) and (\ref{127eh}) to (\ref{127et}), we obtain (\ref{Thm2(2)}). The integrated mean squared error\\
\begin{eqnarray*}
&&\int_{t \in D_{rf}} E \left ( \hat{r}(t; h_{1,n}, h_{2,n})- r(t) \right )^2 dt\nonumber\\
&&=h^8_{2,n} \int_{t \in D_{rf}} \theta^2(t)dt + \frac{\mu_{0,2}\sigma^2}{n h_{2,n}} \int_{t \in D_{rf}} \frac{|r'(t)|^{1/4}}{\sqrt{f(t)}}dt + o(h^8_{2,n})+o\left (\frac{1}{n h_{2,n}} \right ).\qquad
\end{eqnarray*}
Taking derivative with respect to $h_{2,n}$ and letting the result equal to $0$, we obtain the optimal bandwidth
\begin{eqnarray*}
h_n^*=\left (\frac{1}{n} \right )^{1/9}\left ( \frac{\mu_{0,2}\sigma^2 }{8 \int_{t \in D_{rf}} \theta^2(t)dt} \int_{t \in D_{rf}} \frac{|r'(t)|^{1/4}}{\sqrt{f(t)}}dt\right )^{1/9}.
\end{eqnarray*}
\end{proof}
\begin{proof}[Proof of Theorem \ref{central2}]
The proof is similar to the proof of Theorem \ref{central1} and use the results in Theorem \ref{Thm2}.
\end{proof}
|
\section{Introduction}
The growth of dust grains by accretion of gas-phase metals is widely assumed to be an important process in the interstellar medium (ISM). Observations such as the correlation of depletion with gas density \citep{jenkins2009} are naturally explained by this mechanism, and studies of galaxy evolution have typically found it a necessary addition to models in order to reproduce both present-day \citep{popping2017,devis2019,triani2020} and high-redshift \citep{mancini2015,graziani2020} dust masses, although some recent works have challenged this \citep{gall2018,delooze2020,nanni2020}. Despite this, our understanding of how grain growth actually works is limited. In dense gas, where accretion is efficient, dust grains form icy mantles rather than directly accreting refractory elements, which are rapidly photodesorbed when exposed to ultraviolet (UV) radiation \citep{barlow1978}. \citet{ferrara2016} and \citet{ceccarelli2018} have argued that this makes any actual increase in bulk dust mass via accretion impossible.
\citet{zhukovska2016,zhukovska2018} identified a potential way to avoid this issue by considering grain growth in the cold neutral medium (CNM). While the low density makes accretion highly inefficient for traditional models (e.g. \citealt{hirashita2011}), under these conditions sufficiently small dust grains become negatively charged \citep{weingartner2001}. As most dust-forming elements exist as singly-charged positive ions when exposed to UV radiation, Coulomb attraction can significantly enhance the accretion rate \citep{weingartner1999}, leading to much more efficient grain growth than otherwise expected. \citet{zhukovska2016,zhukovska2018} implemented this mechanism into a hydrodynamical simulation of the ISM, finding that the observed patterns of silicon and iron depletion with respect to gas density were well-reproduced, although requiring a somewhat smaller minimum grain size than the typical \citet{mathis1977} (MRN) distribution.
The models in \citet{zhukovska2016,zhukovska2018} assumed a constant power law grain size distribution, the properties of which are treated as input parameters. However, grain growth, by definition, involves an evolution of the size distribution. This results in small grains rapidly becoming significantly larger \citep{hirashita2011}, to the point where the \citet{weingartner1999} model predicts they become positively charged and repel, rather than attract, positive ions. In this paper, {we show that the inclusion of a consistently evolving size distribution significantly reduces the effiency of grain growth in the diffuse ISM. Observed elemental depletions in the CNM can only be reproduced by assuming either very efficient grain shattering, or a further reduction in the minimum grain size from the \citet{zhukovska2016} model, both of which result in extinction laws in conflict with Galactic measurements. We suggest that the assumption of rapid dust destruction by supernovae in the CNM is at fault. Relaxing this assumption, the observed depletion patterns can be reproduced as long as the material injected into the CNM is already significantly depleted (${\rm [Si/H]_{gas}} \sim -1.5$).}
\section{Method}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{coulomb.pdf}
\caption{Coulomb focusing factor versus grain size for $T=100 \, {\rm K}$, $x_e = 0.0015$ and $n_{\rm H} = 5$ (solid line), $30$ (dashed line) and $100 \,{\rm cm}^{-3}$ (dotted line). The thin solid line marks $D(a) = 1$, i.e. no net attraction or repulsion.}
\label{fig:coulomb}
\end{figure}
\begin{table}
\centering
\caption{Model parameters}
\begin{tabular}{ccc}
\hline
Parameter & Value & Unit \\
\hline
Gas density $n_{\rm H}$ & $30$ & $\,{\rm cm}^{-3}$ \\
Gas temperature $T$ & $100$ & $\, {\rm K}$ \\
Electron fraction $x_e$ & $0.0015$ & - \\
Grain density $\rho$ & $3.13$ & $\,{\rm g \, cm}^{-3}$ \\
Silicon mass fraction $f_{\rm Si}$ & $0.165$ & - \\
Silicon elemental abundance $\epsilon_{\rm Si}$ & $3.24 \times 10^{-5}$ & - \\
Initial depletion ${\rm [Si/H]_{gas}}$ & $-0.5$ & - \\
Dust destruction timescale $\tau_{\rm dest}$ & $350$ & $\, {\rm Myr}$ \\
Maximum grain size $a_{\rm max}$ & $0.25$ & $\, {\rm \mu m}$ \\
Minimum grain size $a_{\rm min}$ & $0.005$ & $\, {\rm \mu m}$ \\
\hline
\end{tabular}
\label{tab:params}
\end{table}
We follow, as closely as possible, the formalism and parameter values used by \citet{zhukovska2016}. We consider silicate grains, with a density of $\rho = 3.13 \,{\rm g \, cm}^{-3}$ and a silicon mass fraction $f_{\rm Si} = 0.165$, and assume that growth is limited by the availability of silicon atoms. We take the elemental abundance of silicon to be $\epsilon_{\rm Si} = 3.24 \times 10^{-5}$ \citep{lodders2009}. In the absence of grain charge, the rate of change of grain radius is then
\begin{equation}
\frac{da}{dt} = \frac{n_{\rm Si} m_{\rm Si} <v_{\rm Si}>}{4 \rho f_{\rm Si}}
\label{eq:gr}
\end{equation}
where $n_{\rm Si}$, $m_{\rm Si} = 28 m_{\rm H}$ and $<v_{\rm Si}>$ are the gas-phase number density, mass and mean thermal velocity of silicon atoms respectively. In principle, this equation may be modified by a sticking probability, dependent on gas/grain temperature among other properties. \citet{zhukovska2016} find different implementations of the sticking probability have little impact on their results, so we assume the sticking probability is unity under CNM conditions.
Equation \ref{eq:gr} is independent of grain radius. The introduction of grain charge, which varies with size, modifies the growth rate by a Coulomb focusing factor $D(a)$ \citep{weingartner1999}, the ratio of the actual cross-section for collision to the geometric cross section. We calculate the grain charge distributions for sizes between $0.001-0.5 \, {\rm \mu m}$ following \citet{weingartner2001}, and convert these into focusing factors assuming singly-charged positive ions using the formulae presented in \citet{weingartner1999}. Figure \ref{fig:coulomb} shows the variation of $D(a)$ with grain size for a range of densities, for gas with temperature $T = 100 \, {\rm K}$ and electron fraction $x_e = 0.0015$. For typical CNM conditions, small grains tend to be negatively charged and thus attract positive ions, but beyond a radius of $\sim 0.01 \, {\rm \mu m}$ grains instead become positively charged and $D(a) < 1$. This transition radius increases with density as electron attachment becomes more effective (for constant $x_e$), which also reduces the magnitude of the repulsion effect, but the qualitative behaviour is the same.
We note that while \citet{weingartner1999}, and by extension \citet{zhukovska2016,zhukovska2018}, assume an electron fraction in the CNM of $0.0015$, this requires the majority of electrons to come from hydrogen or helium, and including grain-assisted recombination for these elements has been shown to reduce the CNM electron fraction to $\sim 10^{-4}$ \citep{weingartner2001b,liszt2003}. We explore the effect of this lower value in Appendix \ref{sec:lowe}, finding that it significantly reduces the accretion efficiency due to generally more positively-charged grains. Nonetheless, we continue to use the higher value for consistency with previous work.
We assume an initial MRN grain size distribution with $a_{\rm min} = 0.005 \, {\rm \mu m}$, $a_{\rm max} = 0.25 \, {\rm \mu m}$ and a power law index of $-3.5$, and an initial silicon depletion value ${\rm [Si/H]_{gas}} = \log_{10} \left( \frac{n_{\rm Si}}{\epsilon_{\rm Si} n_{\rm H}} \right) = -0.5$, again following \citet{zhukovska2016}. The initial ${\rm [Si/H]_{gas}}$ allows us to determine the initial number density of grains, $n_g(a)$. We divide the size distribution into 100 logarithmically-distributed bins, and use Equation \ref{eq:gr} combined with $D(a)$ to calculate the growth rate for each size bin. For a time interval $dt$ we can then calculate the increase in grain radius, $da$, and the new grain size, $a_1 = a_0 + da$. As we later implement processes which do not conserve grain number, rather than simply updating $a$ for each bin, we redistribute the grains in bin $i$ between bins $j$ and $j+1$ where the new radius $a_1$ falls between $a_j$ and $a_{j+1}$, such that
\begin{equation}
dn_g(j) = n_g(i)\left(1 - \frac{a_1 - a(j)}{a(j+1)-a(j)}\right)
\end{equation}
and
\begin{equation}
dn_g(j+1) = n_g(i)\left(1 - \frac{a(j+1)-a_1}{a(j+1)-a(j)}\right).
\end{equation}
The new mass in grains due to the increase in total volume is then used to update the remaining gas-phase abundance of silicon $n_{\rm Si}$.
Note that as we used a fixed grid of grain sizes, the largest grain size does not grow, and over time smaller grains `pile up' in the last bin as they reach this size. For an MRN distribution, the accretion rate is always dominated by the smallest grain sizes, particularly when charge is included (as large grains are positively charged and so have $D(a) < 1$). Even without grain charge, the difference in the mass accreted when accounting for this effect is negligible, as the fractional increase in grain radius (and thus volume) for the largest grains is tiny. In any case, this has little effect on our overall argument.
\citet{zhukovska2016} model the growth of dust mass via the equation
\begin{equation}
\frac{df_d}{dt} = \frac{1}{\tau_{\rm acc}} f_d (1 - f_d)
\label{eq:gr2}
\end{equation}
where $f_d$ is the fraction of silicon locked up in dust grains (i.e. $f_d = 1 - 10^{\rm [Si/H]_{gas}}$) and $\tau_{\rm acc}$ is the accretion timescale, defined as
\begin{equation}
\tau_{\rm acc}^{-1} = \frac{3 \epsilon_{\rm Si} m_{\rm Si} v_{\rm Si} n_{\rm H}}{\rho f_{\rm Si} <a>_3}.
\label{eq:tscale}
\end{equation}
The quantity $<a>_3$ is the average grain radius accounting for Coulomb focusing, given by
\begin{equation}
<a>_3 = <a^3>/<D(a)a^2>.
\end{equation}
In \citet{zhukovska2016} this quantity, and thus $\tau_{\rm acc}$, is constant for a given set of physical conditions. {We use the initial value of $\tau_{\rm acc}$ and Equation \ref{eq:gr2} to track the depletion in the case of a constant size distribution.}
\citet{zhukovska2016} implement dust destruction via two mechanisms: direct destruction in gas particles affected by supernovae in the simulation, using a prescription for the mass of gas `cleared' of dust \citep{jones1994,jones1996,dwek2007}, and additional destruction in the diffuse ISM representing supernovae originating from field OB stars, rather than those near their birth molecular clouds. The latter is treated as a destruction timescale $\tau_{\rm dest}$ in gas below a density threshold of $n_{\rm H} = 1 \,{\rm cm}^{-3}$, with \citet{zhukovska2016} choosing $\tau_{\rm dest} = 100 \, {\rm Myr}$. { From the evolution of the total destruction timescale presented in \citet{zhukovska2016}, we note that a) the second process appears to be dominant over destruction by individual supernova events, and b) $\tau_{\rm dest}$ remains between $\sim 300-400 \, {\rm Myr}$, in approximate balance with the dust production rate, for the majority of the simulation.} We thus set $\tau_{\rm dest} = 350 \, {\rm Myr}$, with the dust destruction then being given by
\begin{equation}
\frac{df_d}{dt} = -\frac{f_d}{\tau_{\rm dest}}.
\end{equation}
This treatment does not account for grain size, with larger grains expected to be more resilient to destruction via sputtering. However, grain shattering processes, which are more efficient for larger grains, can redistribute the mass to smaller grain sizes, so that the post-shock size distribution is in general a complex function of many input parameters and assumptions \citep{slavin2015,kirchschlager2019}. As a full treatment of dust destruction is beyond the scope of this paper, we assume all grain sizes are affected equally, i.e. with the same $\tau_{\rm dest}$.
We initially investigate dust growth under typical CNM conditions of $n_{\rm H} = 30 \,{\rm cm}^{-3}$ and $T = 100 \, {\rm K}$ \citep{weingartner1999}. {We follow the evolution for $10 \, {\rm Myr}$, typical of grain residence times in the CNM \citep{peters2017}, with a timestep of $10^4 \, {\rm yr}$.} Model parameters are listed in Table \ref{tab:params}.
\section{Results}
\begin{figure*}
\centering
\subfigure{\includegraphics[width=\columnwidth]{n30t100.pdf}}\quad
\subfigure{\includegraphics[width=\columnwidth]{n30t100_size.pdf}}
\caption{Silicon depletion (left) and final grain size distribution (right) for models with an evolving size distribution (blue solid lines) or for a constant MRN distribution (red dashed lines), for $n_{\rm H} = 30 \,{\rm cm}^{-3}$ and $T = 100 \, {\rm K}$.}
\label{fig:n30t100}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{obsdep.pdf}
\caption{Silicon depletion after $10 \, {\rm Myr}$ versus gas density for models with an evolving size distribution (blue solid line) or for a constant MRN distribution (red dashed line). Observational values from \citet{savage1996} for the warm and cool disc are shown as black triangles at representative densities of $n_{\rm H} = 0.5$ and $30 \,{\rm cm}^{-3}$.}
\label{fig:obsdep}
\end{figure}
\begin{figure*}
\centering
\subfigure{\includegraphics[width=\columnwidth]{obsdep_a3.pdf}}\quad
\subfigure{\includegraphics[width=\columnwidth]{obsdep_a1.pdf}}
\caption{Silicon depletion after $10 \, {\rm Myr}$ versus gas density for models with an evolving size distribution (blue solid line) or for a constant MRN distribution (red dashed line), with $a_{\rm min} = 3$ (left) and $1 \, {\rm nm}$ (right). Observational values from \citet{savage1996} for the warm and cool disc are shown as black triangles at representative densities of $n_{\rm H} = 0.5$ and $30 \,{\rm cm}^{-3}$.}
\label{fig:obsdepamin}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{sizeext.pdf}
\caption{Extinction curves, using \citet{laor1993} silicate optical properties, for the initial MRN distribution (solid black line) and the grain size distributions after $10 \, {\rm Myr}$ for our fiducial model, with $a_{\rm min} = 5 \, {\rm nm}$ (dashed black line), and for $a_{\rm min} = 1 \, {\rm nm}$ (dotted black line). {The values of $A_{\rm V}/N_{\rm H}$ are $1.2$, $1.2$ and $1.1 \times 10^{-22} \, {\rm mag \, cm^2}$ respectively. The value of $A_{\lambda}/A_{\rm V}$ has been reduced by half to account for the typical silicate/carbon ratio of grains in the ISM.} The \citet{cardelli1989} Galactic extinction curve, with $R_{\rm V} = 3.1$, is also shown for comparison (red dotted line).}
\label{fig:sizeext}
\end{figure}
\begin{figure}
\centering
\subfigure{\includegraphics[width=\columnwidth]{shatter.pdf}}
\caption{Silicon depletion for models with an evolving size distribution (blue lines) or for a constant MRN distribution (red dashed line), for $n_{\rm H} = 30 \,{\rm cm}^{-3}$ and $T = 100 \, {\rm K}$, and a shattering efficiency of $0$ (solid blue line), $0.01$ (dashed blue line) and $0.05$ (dotted blue line).}
\label{fig:shatter}
\end{figure}
\begin{figure*}
\centering
\subfigure{\includegraphics[width=\columnwidth]{shatter_size.pdf}}\quad
\subfigure{\includegraphics[width=\columnwidth]{shatter_ext.pdf}}
\caption{\textit{Left:} Final grain size distribution (blue solid line) and initial MRN distribution (red dashed line) after $10 \, {\rm Myr}$ for $n_{\rm H} = 30 \,{\rm cm}^{-3}$, $T = 100 \, {\rm K}$, and a shattering efficiency of $0.05$. \textit{Right:} Extinction curves, using \citet{laor1993} silicate optical properties, for the initial MRN distribution (solid black line) and the grain size distributions after $10 \, {\rm Myr}$ for a shattering efficiency of $0.01$ (dashed black line) and $0.05$ (dotted black line). {The values of $A_{\rm V}/N_{\rm H}$ are $1.2$, $0.28$ and $0.12 \times 10^{-22} \, {\rm mag \, cm^2}$ respectively. The value of $A_{\lambda}/A_{\rm V}$ has been reduced by half to account for the typical silicate/carbon ratio of grains in the ISM.} The \citet{cardelli1989} Galactic extinction curve, with $R_{\rm V} = 3.1$, is also shown for comparison (red dotted line).}
\label{fig:shatterext}
\end{figure*}
\begin{figure*}
\centering
\subfigure{\includegraphics[width=\columnwidth]{bigdeplete.pdf}}\quad
\subfigure{\includegraphics[width=\columnwidth]{bigsize.pdf}}
\caption{Silicon depletion (left) and final grain size distribution (right) for models with an evolving (blue solid lines) and constant size distribution (red dashed lines), for $n_{\rm H} = 30 \,{\rm cm}^{-3}$, $T = 100 \, {\rm K}$ and a shattering efficiency of $0.05$, and an initial log-normal size distribution centred at and with a width of $0.1 \, {\rm \mu m}$.}
\label{fig:bigdust}
\end{figure*}
{Figure \ref{fig:n30t100} shows the evolution of ${\rm [Si/H]_{gas}}$ and the grain size distribution for $n_{\rm H} = 30 \,{\rm cm}^{-3}$, $T = 100 \, {\rm K}$, with and without an evolving grain size distribution. For a constant MRN distribution, the level of silicon depletion increases by $0.2$ dex over $10 \, {\rm Myr}$, whereas when the increase in grain radii is accounted for the increase is negligible. The final size distribution develops a peak at $\sim 0.01 \, {\rm \mu m}$, where the typical grain charge starts to transition from negative to positive. Grain smaller than this can accrete gas-phase material efficiently, but at larger radii this is increasingly prevented as the Coulomb focusing factor decreases, and large, positively-charged grains are almost completely unable to grow via accretion. This effect occurs regardless of gas properties - Figure \ref{fig:obsdep} shows ${\rm [Si/H]_{gas}}$ after $10 \, {\rm Myr}$ for varying gas density, compared to representative observational values taken from \citet{savage1996}. The increase in the size of the smallest grains, and the corresponding reduction in the accretion efficiency, results in less depletion at all densities, but this is particularly noticeable at the higher densities, where the timescales are short enough for significant accretion to occur over the $10 \, {\rm Myr}$ timespan.}
{Even for a constant size distribution, our model returns values of ${\rm [Si/H]_{gas}}$ significantly higher than those observed in the CNM. \citet{zhukovska2016} found the same issue for the typical MRN minimum grain size of $5 \, {\rm nm}$, motivating them to investigate smaller values. Figure \ref{fig:obsdepamin} shows ${\rm [Si/H]_{gas}}$ versus gas density for minimum grain radii of $3$ and $1 \, {\rm nm}$. With $a_{\rm min} = 3 \, {\rm nm}$ and a constant MRN distribution, the \citet{savage1996} CNM depletion can be reproduced, as found by \citet{zhukovska2016}, but with an evolving size distribution this value ($-1.3$ dex) is only reached for a gas density of $n_{\rm H} = 100 \,{\rm cm}^{-3}$, at which point several of our model assumptions, such as the temperature and electron density, are likely to become inappropriate. Reducing the minimum grain size further to $1 \, {\rm nm}$ allows ${\rm [Si/H]_{gas}}$ to reach the observed value within $10 \, {\rm Myr}$, but the resulting size distribution is significantly overabundant in $\sim 0.01 \, {\rm \mu m}$ grains compared to the typical MRN case. This causes a corresponding increase in the far-UV extinction which is incompatible with values seen along Galactic sightlines, shown in Figure \ref{fig:sizeext} (we have reduced the value of $A_{\lambda}/A_{\rm V}$ by half to account for the fact that silicates make up only $\sim 50\%$ of the ISM dust budget).} {The model ratios of visual extinction to column density, $A_{\rm V}/N_{\rm H}$, are $\sim 10^{-22} \, {\rm mag \, cm^2}$, lower than expected compared to Galactic values ($\sim 5 \times 10^{-22} \, {\rm mag \, cm^2}$; \citealt{bohlin1978}) if silicates make up half of the total $A_{\rm V}$. This suggests that large grains are underabundant, even for the initial size distribution, due to the low initial depletion and lack of a mechanism to produce grains larger than $\sim 0.01 \, {\rm \mu m}$.}
{The reduction in accretion efficiency is caused by the bottlenecking of small grains as they reach radii where they become positively charged, at which point further growth becomes very slow. In principle, this could be mitigated by grain shattering, redistributing mass from large grains into more numerous, smaller ones. However, we find this either fails to resolve the issue, or raises new ones. A full treatment of shattering requires knowledge of both the dynamics and bulk properties of the dust grains \citep{hirashita2009,kirchschlager2019}, and is beyond the scope of this paper, but we can approximate its effect by calculating the rate of grain-grain collisions and assuming that a given fraction (hereafter the `shattering efficiency') result in a shattering event. Following \citet{hirashita2009}, we assume a typical turbulent velocity dispersion in the CNM of $2 \, {\rm km \, s^{-1}}$, and that the mass of shattered grains is redistributed into a MRN power law size distribution, with the maximum mass equal to the size of the shattered grain.}
{Figure \ref{fig:shatter} shows the evolution of ${\rm [Si/H]_{gas}}$ for models with varying shattering efficiency. Reproducing the static size distribution values requires a shattering efficiency of $\gtrsim 0.01$, but the size distributions and extinction curves produced with these values, shown in Figure \ref{fig:shatterext}, are in even greater tension with those observed than the reduced $a_{\rm min}$ models discussed above - {the extinction in the far-UV is even higher, and $A_{\rm V}/N_{\rm H}$ even lower, due to the rapid redistribution of mass from large to small grains}. We also consider this level of shattering to be unrealistic - the threshold velocity for shattering of silicates in \citet{hirashita2009} is $2.7 \, {\rm km \, s^{-1}}$, larger than the CNM turbulent velocity, so the number of collisions resulting in a shattering event is likely to be very small. The CNM size distribution presented in \citet{hirashita2009} is almost unchanged after $50 \, {\rm Myr}$ of evolution, and does not display the large overabundance of small grains required for efficient grain growth.}
{The inability of grain growth models to reproduce the observed elemental depletions without also producing unrealistic extinction curves is tied to the initial size distribution - the amount of possible growth is limited by the number of small grains, which can only accrete up to a certain radius before becoming positively charged, and for an MRN distribution the amount of additional growth required to reach ${\rm [Si/H]_{gas}} \sim -1.3$ (as required by \citealt{savage1996}) results in implausible levels of far-UV extinction. It is possible that these issues could be circumvented by an initial size distribution with fewer small grains, and subsequent reprocessing of larger grains into smaller ones via shattering. However, we are unable to find a situation where this occurs. Figure \ref{fig:bigdust} shows the evolution of ${\rm [Si/H]_{gas}}$ and the grain size distribution for the extreme case of a top-heavy size distribution (a log-normal with centre and width $0.1 \, {\rm \mu m}$). Such a size distribution could plausibly be produced by the injection of dust by core-collapse supernovae, which are established as primarily producing large grains \citep{gall2014,wesson2015,bevan2016,priestley2020}; from hotter phases of the ISM where only the largest grains are expected to survive; or from coagulation in molecular clouds. Even with a shattering efficiency of $0.05$, accretion of gas-phase material is still not efficient enough to reach observed depletion levels in the CNM after $10 \, {\rm Myr}$, and the final size distribution is skewed far enough to small grain sizes to result in unrealistic levels of far-UV extinction. It may be possible to reproduce the \citet{savage1996} CNM depletions via grain growth without violating constraints on the far-UV extinction with a suitable choice of initial conditions, but we are unable to find a physically motivated scenario for which this is the case.}
\section{Discussion}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{lowdep.pdf}
\caption{Silicon depletion after $10 \, {\rm Myr}$ versus gas density for models with an evolving size distribution (blue solid line) or for a constant MRN distribution (red dashed line), assuming an initial depletion of ${\rm [Si/H]_{gas}} = -1.5$ and a dust destruction timescale $\tau_{\rm dest} = 1 {\rm \, Gyr} \, (n_{\rm H}/30 \,{\rm cm}^{-3})$. Observational values from \citet{savage1996} for the warm and cool disc are shown as black triangles at representative densities of $n_{\rm H} = 0.5$ and $30 \,{\rm cm}^{-3}$.}
\label{fig:lowdep}
\end{figure}
\begin{figure*}
\centering
\subfigure{\includegraphics[width=\columnwidth]{lowdep_size.pdf}}\quad
\subfigure{\includegraphics[width=\columnwidth]{lowdep_ext.pdf}}
\caption{\textit{Left:} Final grain size distribution (blue solid line) and initial MRN distribution (red dashed line) after $10 \, {\rm Myr}$ for $n_{\rm H} = 30 \,{\rm cm}^{-3}$, $T = 100 \, {\rm K}$, {${\rm [Si/H]_{gas}} = -1.5$ and $\tau_{\rm dest} = 1 \, {\rm Gyr}$}. \textit{Right:} Extinction curves, using \citet{laor1993} silicate optical properties, for the initial (solid black line) {and final (dashed black line) size distributions}. {The values of $A_{\rm V}/N_{\rm H}$ are $1.7$ and $1.7 \times 10^{-22} \, {\rm mag \, cm^2}$ respectively.} The value of $A_{\lambda}/A_{\rm V}$ has been reduced by half to account for the typical silicate/carbon ratio of grains in the ISM. The \citet{cardelli1989} Galactic extinction curve, with $R_{\rm V} = 3.1$, is also shown for comparison (red dotted line).}
\label{fig:lowdepext}
\end{figure*}
{If observed elemental depletions in the CNM cannot be reproduced by efficient accretion onto small grains, they must be caused by some other mechanism. We suggest that the assumed dust destruction timescales are the most likely avenue. As noted above, \citet{zhukovska2016} account for both destruction by supernovae captured in their hydrodynamical simulations, and additional destruction in diffuse gas ascribed to supernovae occuring at a distance from their birth molecular clouds. The latter mechanism is introduced in order to prevent overdepletion in the CNM, which may not be necessary if the assumed grain-growth timescales are too low as we argue, and the former is based on one-dimensional theoretical models which appear to be in conflict with at least some observed supernova remnants \citep{koo2016,chawner2020b,priestley2021}. If either or both of the assumed destruction timescales are too low, the low gas-phase silicon abundance in the CNM is not due to efficient accretion of gas-phase material, but is indicative of highly-depleted gas injected into this phase of the ISM. Observations of both the Crab Nebula \citep{owen2015,delooze2019} and Cassiopeia A \citep{delooze2017,laming2020} have found that supernova ejecta dust masses are comparable to those of the gas-phase dust-forming elements, or possibly even higher. High initial depletions could also be due to grain accretion in molecular clouds, although this would require some way of growing the refractory mass rather than ice mantles.}
{Figure \ref{fig:lowdep} demonstrates an attempt at constructing a model based on these hypotheses. We assume that an initial value of ${\rm [Si/H]_{gas}} = -1.5$, and a dust destruction timescale scaling as $\tau_{\rm dest} = 1 {\rm \, Gyr} \, (n_{\rm H}/30 \,{\rm cm}^{-3})$. The normalisation of $\tau_{\rm dest}$ is motivated by the values in \citet{zhukovska2016} without the additional diffuse ISM destruction, while the scaling is consistent with previous theoretical work on the density dependence of dust destruction \citep{draine1990,hu2019}. With or without an evolving size distribution, the model broadly reproduces the depletions in both the CNM and warmer gas, as well as the trend of increasing depletion at higher densities. {The size distribution and extinction curve, shown in Figure \ref{fig:lowdepext}, are nearly unchanged from the initial conditions after $10 \, {\rm Myr}$ at $n_{\rm H} = 30 \,{\rm cm}^{-3}$, and with $A_{\rm V}/N_{\rm H} = 1.7 \times 10^{-22} \, {\rm mag \, cm^2}$ are in better agreement with \citet{bohlin1978}, particularly as we expect carbon grains to contribute slightly more to the total $A_{\rm V}$ (using amorphous carbon optical properties from \citet{zubko1996} instead of silicates, we find $A_{\rm V}/N_{\rm H} = 4.1 \times 10^{-22} \, {\rm mag \, cm^2}$).} While this is a somewhat artificial example due to the choice of the initial depletion, and neglects the cycling of gas between different phases of the ISM, it does show that phenomena commonly attributed to grain growth can be equally well explained by variations in the dust destruction efficiency.}
\section{Conclusions}
{Efficient accretion of gas-phase metals by small, negatively-charged dust grains has been proposed as an explanation for elemental depletion patterns observed in the CNM \citep{zhukovska2016,zhukovska2018}. We have demonstrated that once the evolution of the size distribution is properly accounted for, this becomes impossible, as the growth in dust mass is halted once the small grains grow large enough to become positively charged. Increasing or replenishing the number of small grains, such as by altering the initial size distribution or invoking efficient grain shattering, results in far-UV extinction curves incompatible with anything observed along Galactic sightlines. We suggest that relatively high levels of depletion in the CNM, rather than being a sign of efficient grain growth, are actually indicative of {the survival of dust grains in initially highly-depleted material from a (presumably) denser phase of the ISM.}}
\section*{Acknowledgements}
{We thank the referee for a useful report which significantly improved this paper.} We are grateful to Luisa Lucie-Smith for constructive discussions on this topic. FDP is funded by the Science and Technology Facilities Council. IDL acknowledges support from European Research Council (ERC) starting grant 851622 DustOrigin. MJB acknowledges support from the ERC grant SNDUST ERC-2015-AdG-694520.
\section*{Data Availability}
The data underlying this article will be made available upon request. The Fortran code used to generate the data is available at \url{www.github.com/fpriestley/growth}.
\bibliographystyle{mnras}
|
\section{Introduction}
In an interesting recent work, \citeauthor{kuzborskij2019efron}
derived a confidence bound for the random variable
\[
\Delta = f(S) - \E[f(S)
\]
where $S = (Z_1, Z_2, \ldots, Z_n)$ is a size-$n$ random sample composed of independent $\cZ$-valued random elements $Z_i$, and $f : \cZ^n \to \R$ is a measurable function.
Notice, however, that the components
are not required to be identically distributed: each $Z_i$ may be distributed according to a different%
\footnote{$\cM_1(\cA,\Sigma_{\cA})$ denotes the family of probability measures defined on a measurable space $(\cA,\Sigma_{\cA})$. When $\Sigma_{\cA}$ is clear from the context, we write simply $\cM_1(\cA)$ for simplicity.}
$\mu_i \in \cM_1(\cZ)$.
Accordingly, the distribution of the size-$n$ random sample $S$ is
$P_n = \mu_1 \otimes\cdots\otimes \mu_n$.
Their confidence bound
is based on an estimator of the variance of $f(S)$.
Recall that McDiarmid's inequality, which is based on the bounded differences property, relates concentration of $\Delta$ around zero (its mean) to
the sensitivity of $f$ to coordinatewise perturbations (``first-order'').
By contrast, the bound of \citeauthor{kuzborskij2019efron} relates concentration to squared perturbations (``second-order''), which leads to an inequality based on a variance estimator.
The latter has a resemblance with a well-known estimator, recalled next.
\paragraph{The variance estimator used in the Efron-Stein inequality.}
This is defined as follows:
\begin{equation}
\label{eq:VES}
V^{\text{\scshape{es}}} = \sum_{k=1}^n \E\left[(f(S) - f(S\rep{k}))_+^2 \,\middle|\, S \right]\,,
\end{equation}
where $(s)_+ = \max\{0,s\}$ is the positive part, and the notation $S\rep{k}$ indicates that the $k$th element of $S$ is replaced with $Z_k'$, where $S'= (Z_1', Z_2', \ldots, Z_n')$ is an independent copy of $S = (Z_1, Z_2, \ldots, Z_n)$.
Further details about this estimator, with context and references,
can be found in \citet{boucheron2013book}.
Problem: In order to prove a confidence bound for $\Delta$ based on $V^{\text{\scshape{es}}}$, one needs
\emph{a priori} assumptions on the moments of $V^{\text{\scshape{es}}}$.
To avoid this limitation, \citeauthor{kuzborskij2019efron} used a modified variance estimator.
\paragraph{The variance estimator used in the Kuzborskij-Szepesv{\'a}ri inequality.}
This is defined as follows:
\begin{equation}
\label{eq:VKS}
V^{\text{\scshape{ks}}} = \sum_{k=1}^n \E\left[(f(S) - f(S\rep{k}))^2 \,\middle|\, Z_1, \ldots, Z_k \right]\,.
\end{equation}
\citeauthor{kuzborskij2019efron} called it a ``semi-empirical'' estimator, because of its dependence on both the sample and the distribution of the sample.
The main result of \citeauthor{kuzborskij2019efron} is the following high-confidence bound:
For any $y > 0$ and $x \geq 1$,
with probability at least $1-e^{-x}$ one has
\begin{equation}
\label{eq:intro:es_1}
|\Delta| \leq \sqrt{2 (V^{\text{\scshape{ks}}} + y) \left[ 1 + \frac{1}{2} \log\pr{1 + \frac{V^{\text{\scshape{ks}}}}{y}} \right] x}\,.
\end{equation}
\paragraph{Remark:}
Inequality~\eqref{eq:intro:es_1} does not require boundedness of random variables $Z_i$, nor of the function $f$;
the only crucial assumption is independence of elements in the sample $S$.
Observe that inequality~\eqref{eq:intro:es_1} basically depends on $V^{\text{\scshape{ks}}}$ and a positive free parameter $y$, which must be selected by the user.
For instance, choosing $y = 1/n^2$ gives:
For any $x \geq 1$, with probability at least $1-e^{-x}$ one has
\[
|\Delta|
\leq \sqrt{2 (V^{\text{\scshape{ks}}} + 1/n^2) [1 + \tfrac{1}{2}\log(1 + n^2 V^{\text{\scshape{ks}}}) ] x}\,.
\]
Paraphrasing \citeauthor{kuzborskij2019efron}:
With this particular choice of $y$, the resulting inequality shows a Bernstein-type behavior, in the sense that the upper-bound is dominated by the lower-order term whenever $V^{\text{\scshape{ks}}}$
is small enough;
and the price for such a simple choice of $y$ is in the logarithmic term.
\paragraph{Remark:}
In addition to inequality~\eqref{eq:intro:es_1}, \citeauthor{kuzborskij2019efron} showed a bound that does not involve $y$ and, in particular, is scale-free:
For any $x > 0$, with probability at least $1-\sqrt{2} e^{-x}$ one has
$|\Delta| \leq 2 \sqrt{(V^{\text{\scshape{ks}}} + \E[V^{\text{\scshape{ks}}}]) x}$.
\bigskip
The remaining of this note's content is as follows.
The confidence bound of \citeauthor{kuzborskij2019efron} is presented and proved in \cref{sec:main}; and the `PAC-Bayes-ified' version of this bound is presented and proved in \cref{sec:pac_bayes}.
\section{The main result and its proof}
\label{sec:main}
\begin{theorem}
\label{thm:main}
Let $f : \cZ^n \to \R$ be a measurable function,
let $\Delta = f(S) - \E[f(S)]$ be the random gap with $S$ randomly chosen from a distribution $P_n \in \cM_1(\cZ^n)$,
and let $V = V^{\text{\scshape{ks}}}$ be the variance estimator defined in \cref{eq:VKS}. \\[1mm]
(i) For any $x>0$,
\[
\P\left(|\Delta| > 2 \sqrt{(V + \E[V]) x}\right)
\leq \sqrt{2} e^{-x}\,.
\]
(ii) For any $y > 0$, and any $x \geq 1$,
\[
\P\left(
|\Delta| > \sqrt{2 (V + y) [1 + \tfrac{1}{2}\log(1 + V/y) ] x}
\right)
\leq e^{-x}\,.
\]
\end{theorem}
To discuss the proof of \cref{thm:main}, the following definition will be convenient: A pair of random variables $(A,B)$ is called a \textbf{canonical pair} if $B \ge 0$ and
\begin{align}
\label{eq:canonical_pair}
\sup_{\lambda\in \R}
\E\br{ \exp\pr{\lambda A - \frac{\lambda^2}{2} B^2 }}
\leq 1\,.
\end{align}
See \citet[Section 10.2]{delapena2009} for further discussion on this condition, and its connection with the so-called self-normalized processes.
A key step of the proof of \cref{thm:main} consists of establishing that $(\Delta,\sqrt{V})$ is a canonical pair. We state this as a lemma for convenient reference:
\begin{lemma}
\label[lemma]{lem:canonical}
$(\Delta, \sqrt{V})$ is a canonical pair.
\end{lemma}
The rest of the proof of \cref{thm:main} relies on following technical result, which essentially gives subgaussian tail probabilities for some functions of a canonical pair
(cf. \citet[Theorem 12.4 \& Corollary 12.5]{delapena2009}):
\begin{lemma
\label[lemma]{lem:self_norm_concentration}
Suppose $(A,B)$ is a canonical pair. Then: \\[1mm]
(i) For any $t > 0$,
\[
\P\pr{\frac{|A|}{\sqrt{B^2 + (\E[B])^2}} \geq t}
\leq \sqrt{2} e^{-\frac{t^2}{4}}\,.
\]
(ii) For any $y > 0$ and $t \geq \sqrt{2}$,
\[
\P\pr{\frac{|A|}{
\sqrt{(B^2 + y) \br{1 + \frac{1}{2} \log\pr{1 + \frac{B^2}{y}}}}
} \geq t}
\leq e^{-\frac{t^2}{2}}\,.
\]
\end{lemma}
The proof of \cref{thm:main} is then merely by combining \cref{lem:canonical} and \cref{lem:self_norm_concentration}.
Hence,
it remains to prove \cref{lem:canonical}.
This uses the martingale method, which is at the core of the proofs of McDiarmid's/Azuma-Hoeffding's inequalities.
\begin{proof}[Proof of \cref{lem:canonical}]
Let $\E_k[\cdot]$ stand for $\E[\,\cdot\,|\, Z_1,\dots,Z_k]$.
Using the martingale difference decomposition,
the gap $\Delta = f(S) - \E[f(S)]$ can be written as
\begin{align*}
\Delta = \sum_{k=1}^n D_k
\end{align*}
where $D_k = \E_k[f(S)] - \E_{k-1}[f(S)]$.
Notice that $D_k = \E_k[f(S)-f(S\rep{k})]$, which follows from the elementary identity
$\E_{k-1}[f(S)] =\E_k[f(S\rep{k})]$.
The variance estimator $V = V^{\text{\scshape{ks}}}$ (cf. \cref{eq:VKS}) can be rewritten as
\[
V = \sum_{k=1}^n V_k
\]
where $V_k = \E_k\br{ \pr{f(S) - f(S\rep{k})}^2 }$.
This is just a convenient notation.
Assume for now that for every $k\in [n]$ the following holds:
\begin{align}
\E_{k-1}\br{
\exp\pr{ \lambda D_k - \frac{\lambda^2}{2} V_k }
}
\leq 1\,.
\label{eq:dkbkcp}
\end{align}
Then, using a recursive argument and \cref{eq:dkbkcp}, we get
\begin{align*}
&\E\br{\exp\pr{\lambda \Delta - \frac{\lambda^2}{2} V}}
=
\E\br{\prod_{k=1}^n \exp\pr{
\lambda D_k - \frac{\lambda^2}{2} V_k
}}
\\
&\hspace*{7mm} =
\E\br{
\underbrace{
\E_{n-1}\br{
\exp\pr{ \lambda D_n - \frac{\lambda^2}{2} V_n
}}}_{\le 1}
\prod_{k=1}^{n-1}
\exp\pr{ \lambda D_k - \frac{\lambda^2}{2} V_k}
}
\\
&\hspace*{7mm} \le
\E\br{
\underbrace{\E_{n-2}\br{
\exp\pr{ \lambda D_{n-1} - \frac{\lambda^2}{2} V_{n-1}}
}}_{\le 1}
\prod_{k=1}^{n-2}
\exp\pr{ \lambda D_k - \frac{\lambda^2}{2} V_k}
}
\\
&\hspace*{7mm} \le \dots \le 1\,.
\end{align*}
Thus, it remains to prove \cref{eq:dkbkcp}.
Fix $k\in [n]$ and let $\varepsilon \in \{-1,+1\}$
be a random variable independent of $S,S'$
such that
$\P(\varepsilon = +1) = \P(\varepsilon = -1) = 1/2$.
Let $\Delta_k = f(S) - f(S\rep{k})$.
Notice that
$\lambda D_k - \frac{\lambda^2}{2} V_k = \E_k[\lambda \Delta_k - \frac{\lambda^2}{2} \Delta_k^2 ]$,
and by Jensen's inequality
\[
\exp\pr{\E_k\br{\lambda \Delta_k - \frac{\lambda^2}{2} \Delta_k^2 }}
\leq \E_k\br{\exp\pr{\lambda \Delta_k - \frac{\lambda^2}{2} \Delta_k^2}}\,.
\]
Let $\E_{-k}[\cdot]$ denote conditioning on $S$ without $Z_k$.
Then we have
\begin{align*}
\MoveEqLeft
\E_{k-1}\br{
\exp\pr{ \lambda D_k - \frac{\lambda^2}{2} V_k }
}
\leq
\E_{k-1}\br{
\exp\pr{ \lambda \Delta_k - \frac{\lambda^2}{2} \Delta_k^2 }
} \nonumber\\
& =
\E_{k-1}\br{
\E_{-k}\E\br{
\exp\pr{ \varepsilon \lambda \Delta_k - \frac{\lambda^2}{2} \pr{\varepsilon \Delta_k }^2 }
\,\Big|\, S,S' }
}\,.
\end{align*}
The last equality follows from the assumption on the distributions, that is,
given $Z_1,\dots,Z_{k-1},Z_{k+1},\dots,Z_n$, the random variables
$Z_k$ and $Z_k'$ are identically distributed, hence so are
$\Delta_k$ and $-\Delta_k$.
Since $x \varepsilon$ is subgaussian (for any $x \in \R$),
the innermost expectation
in the last display is upper-bounded by one.
\end{proof}
\paragraph{Remark:}
The proof makes it clear that this inequality holds in the slightly more general setting in which $f : \cZ_1\times\cdots\times\cZ_n \to \R$ and $S = (Z_1,\ldots,Z_n)$ has independent components, where each $Z_i$ is a $\cZ_i$-valued random variable with distribution $\mu_i \in \cM_1(\cZ_i)$.
\section{PAC-Bayes-ification}
\label{sec:pac_bayes}
We adapt the notation for $\Delta$ and $V=V^{\text{\scshape{ks}}}$ to make explicit their dependence on $f$, and see them as being defined over $f$'s from some function class $\cF$:
\begin{align}
\Delta(f) &= f(S) - \E[f(S)]\,, \tag{1'} \\
V(f) &= \sum_{k=1}^n \E_k\left[(f(S) - f(S\rep{k}))^2 \right]\,. \tag{2'}
\end{align}
It might be convenient to make explicit the dependence of $\Delta$ and $V$ on the sample $S$ as well;
to do so, we may write $\Delta_S(f)$ and $V_S(f)$.
Recall that the distribution of the (size-$n$) random sample $S$ is $P_n = \mu_1 \otimes\cdots\otimes \mu_n$.
Notice that for a fixed nonrandom $s=(z_1,\ldots,z_n) \in \cZ^n$,
the gap is
\[
\Delta_s(f) = f(s) - \int_{\cZ^n} f(s') P_n(ds')\,.
\]
The expression for $V_s(f)$ is longer to write, but easy to imagine.
The point is that $\Delta$ and $V$ are real-valued functions defined over $\cZ^n \times \cF$.
Let $\cF = \{ f_\theta \}_{\theta\in\Theta}$ be a parametric family of functions $f_\theta : \cZ^n \to \R$.
For each $\theta \in \Theta$, define
$\Delta_S(\theta)$ and $V_S(\theta)$, the gap and the variance estimator for $f_\theta(S)$. Then $(\Delta_S(\theta),\sqrt{V_S(\theta)})$ is a canonical pair, for each $\theta$, by \cref{lem:canonical}.
Given a probability kernel $Q$ from $\cZ^n$ to $\Theta$ and $s \in \cZ^n$, we write expectations with respect to the distribution $Q_s$ as $Q_s[\Delta_s] = \int_{\Theta} \Delta_s(\theta) Q_s(d\theta)$, and similarly $Q_s[V_s] = \int_{\Theta} V_s(\theta) Q_s(d\theta)$.
If $S \sim P_n$ is the random sample, then expectations with respect to the random measure $Q_S$ are conditional expectations:
\begin{align*}
Q_S[\Delta_S] = \E[\Delta_S(\theta) \,|\, S]\,,
\qquad\text{and}\qquad
Q_S[V_S] = \E[V_S(\theta) \,|\, S]\,.
\end{align*}
The joint distribution over $\cZ^n\times\Theta$ defined by $P_n$ and the probability kernel $Q$, denoted $P_n \otimes Q$, is so that choosing a random pair $(S,\theta) \sim P_n \otimes Q$ corresponds to choosing $S \sim P_n$ and then choosing $\theta \sim Q_S$. Accordingly, integrals
under $P_n \otimes Q$ correspond to the `total expectation' with respect to the random choice of $S \sim P_n$ and $\theta \sim Q_S$. For instance,
\begin{align*}
(P_n \otimes Q)[V]
= \int_{\cZ^n} \int_{\Theta} V_s(\theta) Q_s(d\theta) P_n(ds)
= \E[\E[V_S(\theta) \,|\, S]]
= \E[V_S(\theta)]\,.
\end{align*}
With a slight abuse of notation, we may write $P_n[Q_S[V_S]]$ instead of $(P_n \otimes Q)[V]$.
\pagebreak
The `PAC-Bayes-ification' of \cref{thm:main} is as follows.
\begin{theorem}
\label{thm:pac_bayes}
Fix an arbitrary `data-free' probability distribution $Q^0$ over $\Theta$,
and an arbitrary probability kernel $Q$ from $\cZ^n$ to $\Theta$.
%
Then\\[1mm]
(i) For any $x > 0$,
with probability at least $1-2 e^{-x}$
we have
\begin{equation}
\label{eq:pac_bayes_mix_3}
|Q_S[\Delta_S]|
\leq
\sqrt{
2 (\E[V_S({\theta})] + Q_S[V_S]) \pr{\KL\pr{Q_S \,\|\, Q^0} + 2 x}
}\,.
\end{equation}
%
(ii) For all $y > 0$ and $x \geq 1$,
with probability at least $1-e^{-x}$ we have
\begin{equation}
\label{eq:pac_bayes_mix_4}
|Q_S[\Delta_S]|
\leq
\sqrt{
2 \pr{Q_S[V_S] + y}
\br{\KL\pr{Q_S \,\|\, Q^0} + x + \frac{x}{2} \log\pr{1 + \frac{Q_S[V_S]}{y} }}
}\,.
\end{equation}
\end{theorem}
The statement of this theorem uses the language of probability kernels for representing data-dependent distributions (cf. \citet{rivasplata2020}).
In the remaining of this note, we switch back to the usual notation in terms of conditional expectations: $Q_S[\Delta_S] = \E[\Delta_S(\theta) \,|\, S]$ and $Q_S[V_S] = \E[V_S(\theta) \,|\, S]$.
Also recall that
$\E[V_S(\theta)] = \E[\E[V_S(\theta) \,|\, S]] = P_n[Q_S[V_S]]$
is the total expectation.
The proof of \cref{thm:pac_bayes} is based on the following lemma.
\begin{lemma}
\label[lemma]{lem:pac_bayes_mgf}
Under the same conditions as in \cref{thm:pac_bayes}. \\[1mm]
(i) For all $x \geq 0$,
\begin{equation}
\label{eq:pac_bayes_mix_2}
\E\br{
\exp\cbr{
x \sqrt{
\left(
\frac{\E[\Delta_S(\theta) \,|\, S]^2}{\E[V_S(\theta)] + \E[V_S(\theta) \,|\, S]} - 2 \KL\pr{Q_S \,\Vert\, Q^0}
\right)_+
}
}
}
\leq 2 e^{x^2}\,.
\end{equation}
(ii) For any $y > 0$, we have
\begin{equation}
\label{eq:pac_bayes_mix_1}
\E\br{\frac{y}{\sqrt{y^2 + \E[V_S(\theta) \,|\, S]}} \, \exp\cbr{\frac{\E[\Delta_S(\theta) \,|\, S]^2}{2 (y^2 + \E[V_S(\theta) \,|\, S])} - \KL\pr{Q_S \,\Vert\, Q^0}}}
\leq
1\,.
\end{equation}
\end{lemma}
\begin{proof}[Proof of \cref{lem:pac_bayes_mgf}]
For convenience, we start with the proof of \cref{eq:pac_bayes_mix_1}.
Recall the following change of measure, which is the basis of the PAC-Bayesian analysis:
Let $\pi$ and $\pi^0$ be probability measures on $\Theta$, and let the induced expectation operators be $\E$ and $\E^0$, respectively.
Let $X$ be a $\Theta$-valued random variable.
Then, for any measurable function $f \,:\, \Theta \to \R$ we have
\[
\E[f(X)] \leq \KL(\pi\,\Vert\,\pi^0) + \log \E^0\br{e^{f(X)}}\,.
\]
Below we use this with $\pi = Q_S$, $\pi^0 = Q^0$,
and $f_S(\theta) = \lambda \Delta_S(\theta) - \frac{\lambda^2}{2} V_S(\theta)$.
Let $\E$ and $\E^0$ be the expectation
with respect to $P_n \otimes Q$ and $P_n \otimes Q^0$, respectively.
Conditioning on the random sample $S$ we have:
\begin{align*}
\E\br{ \lambda \Delta_S(\theta) - \frac{\lambda^2}{2} V_S(\theta) \,\middle|\, S}
&\leq
\KL\pr{Q_S \,\Vert\, Q^0}
+
\log \E^0\br{ e^{\lambda \Delta_S(\theta) - \frac{\lambda^2}{2} V_S(\theta)} \,\middle|\, S}\,.
\end{align*}
Subtracting the KL term, and taking exponential on both sides gives
\begin{align*}
e^{\E\br{ \lambda \Delta_S(\theta) - \frac{\lambda^2}{2} V_S(\theta) \,\middle|\, S} - \KL\pr{Q_S \,\Vert\, Q^0}}
&\leq
\E^0\br{ e^{\lambda \Delta_S(\theta) - \frac{\lambda^2}{2} V_S(\theta)} \,\middle|\, S}\,.
\end{align*}
Then, taking expectation over the random sample $S$ on both sides, and keeping in mind that $(\Delta_S(\theta), \sqrt{V_S(\theta)})$ is a canonical pair for any fixed $\theta$, we have
\begin{align*}
\E\br{e^{\E\br{ \lambda \Delta_S(\theta) - \frac{\lambda^2}{2} V_S(\theta) \,\middle|\, S} - \KL\pr{Q_S \,\Vert\, Q^0}}}
&\leq
\E^0\br{ e^{\lambda \Delta_S(\theta) - \frac{\lambda^2}{2} V_S(\theta)}} \\
&=
\E^0\br{ \E^0\br{ e^{\lambda \Delta_S(\theta) - \frac{\lambda^2}{2} V_S(\theta)} \, \middle| \, \theta } }
\leq
1\,,
\end{align*}
The equality is by swapping the order of expectation, which is possible since $Q^0$ is a data-free distribution (cf. \cite{rivasplata2020}).
Next, multiplying both sides by $e^{-\lambda^2 y^2/2} y / \sqrt{2 \pi}$ for some fixed $y > 0$, integrating with respect to $\lambda \in \R$, and applying Fubini's theorem, gives%
\footnote{
This is inspired by the proof of \cite[Theorem 12.4]{delapena2009}, which uses the method of mixtures with a Gaussian distribution.
}
\begin{align*}
\E\br{
e^{- \KL\pr{Q_S \,\Vert\, Q^0}}
\frac{y}{\sqrt{2 \pi}} \int_{-\infty}^{\infty} e^{
\lambda \E\br{\Delta_S(\theta) \,|\, S } - \frac{\lambda^2}{2} \E\br{V_S(\theta) \,\middle|\, S} - \frac{\lambda^2}{2} y^2
} \diff \lambda
}
\leq 1\,.
\end{align*}
Carrying out the Gaussian integration we arrive at
\begin{equation*}
\E\br{\frac{y}{\sqrt{y^2 + \E[V_S(\theta) \,|\, S]}} \, \exp\cbr{\frac{\E[\Delta_S(\theta) \,|\, S]^2}{2 (y^2 + \E[V_S(\theta) \,|\, S])} - \KL\pr{Q_S \,\Vert\, Q^0}}}
\leq
1\,,
\end{equation*}
which finishes the proof of~\cref{eq:pac_bayes_mix_1}.
For the other part of the lemma, we consider the following:
\begin{claim}
\label[claim]{lem:subG_from_exp}
Let $U$ be a non-negative random variable, and for $\alpha>0$ define $C(\alpha) =\E\br{\exp(\alpha U^2)}$. Then, for any $x\ge 0$,
$\E\br{\exp(x U )} \le C(\alpha) e^{x^2/4\alpha}$.
\end{claim}
The proof of this claim is as follows.
Fix $\alpha>0$ and $x \geq 0$.
Using the inequality $ab \leq (a^2 + b^2)/2$ with $a = x/\sqrt{2 \alpha}$ and $b=\sqrt{2\alpha}U$ we have
\begin{align*}
x U =
\frac{x}{\sqrt{2\alpha}} \sqrt{2\alpha}U \le \frac{x^2}{4\alpha} + \alpha U^2\,.
\end{align*}
Then take exponential on both sides, and take expectations.
Next, we see the proof of part (i) of the lemma.
Consider the random variable
\[
U=\sqrt{
\left(\frac{\E[\Delta_S(\theta) \,|\, S]^2}{\E[V_S(\theta)] + \E[V_S(\theta) \,|\, S]} - 2 \KL\pr{Q_S \,\Vert\, Q^0}\right)_+
}
\]
and notice that \cref{eq:pac_bayes_mix_2} follows from the claim with $\alpha = 1/4$, provided that we show that $C(1/4) = \E\br{\exp\pr{U^2/4}} \le 2$.
For this, consider an arbitrary $y>0$,
and consider the abbreviations
\[
A=\frac{y}{\sqrt{y^2 + \E[V_S(\theta) \,|\, S]}}\,,
\hspace{9mm}
B=\frac{\E[\Delta_S(\theta) \,|\, S]^2}{2 (y^2 + \E[V_S(\theta) \,|\, S])} - \KL\pr{Q_S \,\Vert\, Q^0}\,.
\]
We need to upper-bound $\E\br{\exp\pr{U^2/4}} = \E\br{\exp(\pr{B}_+/2)}$.
Keeping in mind that $A>0$ (in fact, $A \in (0,1]$),
by Cauchy-Schwarz,
\begin{align*}
\E\br{\exp(\pr{B}_+/2)}
&=\E\br{\exp\pr{\pr{B}_+/2} A^{1/2} A^{-1/2} } \nonumber \\
&\le
\sqrt{\E\br{A \exp\pr{\pr{B}_+} }} \, \sqrt{ \E\br{ A^{-1} } } \,.
\end{align*}
Observe that $\E[A \exp(B)] \leq 1$ by~\cref{eq:pac_bayes_mix_1},
and $A \in (0,1]$.
Now, we have
\begin{align*}
&\sqrt{\E\br{A \exp\pr{(B)_+}} \E\br{\frac{1}{A}}} \\
&\hspace*{5mm}=
\sqrt{\Big(
\E\br{A \, \mathds{1}\{B \geq 0\} \exp\pr{B}} + \E\br{A \, \mathds{1}\{B < 0\}}
\Big)
\E\br{\frac{1}{A}}
}
\leq
\sqrt{\E\br{\frac{2}{A}}}\,.
\end{align*}
Finally, by subadditivity of the square root function and Jensen's inequality,
\[
\sqrt{\E\br{\frac{2}{A}}}
=
\sqrt{2 \E\br{\sqrt{\frac{y^2 + \E[V_S(\theta) \,|\, S]}{y^2}}}}
\leq
\sqrt{2 + 2 \frac{\sqrt{\E[V_S(\theta)]}}{y}}
\leq
2\,,
\]
where the last inequality is by taking any $y \geq \sqrt{\E[V_S(\theta)]}$.
Thus, $C(1/4) \leq 2$ for the chosen $U$.
Applying \cref{lem:subG_from_exp} with $\alpha = 1/4$ completes the proof.
\end{proof}
To complete the argument, the proof of \cref{thm:pac_bayes} is given next.
\begin{proof}[Proof of \cref{thm:pac_bayes}]
Applying Chernoff's bounding technique with~\cref{eq:pac_bayes_mix_2} gives
\begin{align*}
\P\pr{
\sqrt{\pr{\frac{\E[\Delta_S(\theta) \,|\, S]^2}{\E[V_S(\theta)] + \E[V_S(\theta) \,|\, S]} - 2 \KL\pr{Q_S \,\Vert\, Q^0}}_+}
\geq t
}
&\leq
2 \inf_{x \geq 0} e^{x^2 - t x}\,.
\end{align*}
The infimum is $e^{-t^2/4}$.
Thus, with probability at least $1-2 e^{-x}$ we have
\begin{align*}
\pr{\frac{\E[\Delta_S(\theta) \,|\, S]^2}{\E[V_S(\theta)] + \E[V_S(\theta) \,|\, S]} - 2 \KL\pr{Q_S \,\Vert\, Q^0}}_+
\leq
4 x\,.
\end{align*}
With some algebra, this event implies
\begin{align*}
|\E[\Delta_S(\theta) \,|\, S]|
\leq
\sqrt{(\E[V_S(\theta)] + \E[V_S(\theta) \,|\, S]) \pr{2 \KL\pr{Q_S \,\Vert\, Q^0} + 4 x}}\,.
\end{align*}
The last display is equivalent to~\cref{eq:pac_bayes_mix_3}.
Hence \cref{thm:pac_bayes}(i) is proved.
Next, observe that for any $y > 0$ and any $t \geq \sqrt{2}$,
\begin{align*}
\MoveEqLeft
\P\pr{ \frac{\E[\Delta_S(\theta) \,|\, S]^2}{2 (y^2 + \E[V_S(\theta) \,|\, S])} - \KL\pr{Q_S \,\Vert\, Q^0} \geq \frac{t^2}{2} \br{1 + \frac{1}{2} \log\pr{1 + \frac{\E[V_S(\theta) \,|\, S]}{y^2}}} }\\
&\hspace*{-5mm}\leq
\P\pr{ \frac{\E[\Delta_S(\theta) \,|\, S]^2}{2 (y^2 + \E[V_S(\theta) \,|\, S])} - \KL\pr{Q_S \,\Vert\, Q^0} \geq \frac{t^2}{2} + \frac{1}{2} \log\pr{1 + \frac{\E[V_S(\theta) \,|\, S]}{y^2}} }\\
&\hspace*{-5mm}=
\P\pr{ \frac{\E[\Delta_S(\theta) \,|\, S]^2}{2 (y^2 + \E[V_S(\theta) \,|\, S])} - \KL\pr{Q_S \,\Vert\, Q^0} - \frac{1}{2} \log\pr{1 + \frac{\E[V_S(\theta) \,|\, S]}{y^2}} \geq \frac{t^2}{2} }\\
&\hspace*{-5mm}\leq
\E\br{\sqrt{\frac{y^2}{\E[V_S(\theta) \,|\, S] + y^2}} \, \exp\cbr{\frac{\E[\Delta_S(\theta) \,|\, S]^2}{2 (y^2 + \E[V_S(\theta) \,|\, S])} - \KL(Q_S\,\Vert\,Q^0)} } \, e^{-\frac{t^2}{2}}\\
&\hspace*{-5mm}\leq
e^{-\frac{t^2}{2}}\,,
\end{align*}
where the last two inequalities follow from Markov's inequality and~\cref{eq:pac_bayes_mix_1}.
This implies that for all $x \geq 1$,
with probability at least $1-e^{-x}$, one has
\begin{align*}
\frac{\E[\Delta_S(\theta) \,|\, S]^2}{2 (y^2 + \E[V_S(\theta) \,|\, S])} - \KL\pr{Q_S \,\Vert\, Q^0}
\leq
x\pr{1 + \frac{1}{2} \log\pr{1 + \frac{\E[V_S(\theta) \,|\, S]}{y^2}}}\,.
\end{align*}
Notice that $y^2$ may be replaced with $y$, since $y>0$ is a free variable.
Doing this replacement, and rearranging the terms,
we get the equivalent of~\cref{eq:pac_bayes_mix_4}.
Hence \cref{thm:pac_bayes}(ii) is proved.
\end{proof}
\paragraph{Closing remarks.}
\citeauthor{kuzborskij2019efron} deserve fair credit for showing that the pair $(\Delta,\sqrt{V})$ meets \citet{delapena2009}'s `canonical condition' (\cref{lem:canonical}), which enabled powerful tools for bounding exponential moments.
Of course, this was possible with their variance estimator $V = V^{\text{\scshape{ks}}}$.
Apart from that, the main part of the work of \citeauthor{kuzborskij2019efron} is in the proofs of \cref{lem:pac_bayes_mgf} and \cref{thm:pac_bayes}, which cleverly use the techniques of \citeauthor{delapena2009}.
In the next iteration of this note (provided that enough readers cared about it) I intend to add discussions about \cref{thm:main} \& \cref{thm:pac_bayes}, and applications.
|
\section{Introduction}
\paragraph{}The Maldacena conjecture \cite{Maldacena:1997re} and its extensions motivate the search for AdS backgrounds preserving different amounts of Supersymmetry (SUSY), in different dimensions.
The half-maximal SUSY case is especially fructiferous. The correspondence between linear quiver conformal field theories preserving half-maximal SUSY and half-maximal BPS solutions with an AdS factor, leads to a precise map between infinite families of string backgrounds and their dual super-conformal field theories (SCFTs). Indeed, various works have developed the dictionary between $d$-dimensional SCFTs, the associated Hanany-Witten brane set-ups \cite{Hanany:1996ie} and the dual AdS$_{d+1}$ string theory backgrounds.
For the case $d=6$, for which the strongly coupled conformal point is reached at high energies, the papers \cite{Apruzzi:2013yva, Gaiotto:2014lca, Cremonesi:2015bld, Nunez:2018ags} have outlined the holographic dictionary and many other works have developed it. For $d=5$, the works \cite{Brandhuber:1999np,Bergman:2012kr,Lozano:2012au,DHoker:2016ujz,DHoker:2016ysh,DHoker:2017zwj,Lozano:2018pcp} presented backgrounds with an AdS$_6$ factor and their UV-dual SCFTs. The dictionary for the case of four dimensional ${\cal N}=2$ linear quiver SCFTs and their AdS$_5$ dual backgrounds was studied in \cite{Gaiotto:2009gz, ReidEdwards:2010qs, Aharony:2012tz, Nunez:2019gbg} among other works. The case of $d=3$ SCFTs (arising at low energies after a RG flow) and the dual AdS$_4$ backgrounds is studied in \cite{DHoker:2007zhm, DHoker:2008lup, Assel:2011xz} among other works. The correspondence for the case of two-dimensional (half-maximal BPS) low-energy SCFTs is particularly rich and has received a lot of attention recently. With the lens described above (linear quivers, Hanany-Witten set-ups and dual backgrounds), we encounter the works \cite{Couzens:2017way,Lozano:2019emq,Lozano:2019jza,Lozano:2019zvg,Lozano:2019ywa,Lozano:2020bxo,Faedo:2020nol,Faedo:2020lyw,Dibitetto:2020bsh} among various other papers.
The study of AdS$_2$ backgrounds in string/M-theory has a long and illustrious history. With the point of view described above, partial aspects of the correspondence between super-conformal quantum mechanics theories (SCQMs) of the quiver type and half-maximal BPS backgrounds containing an AdS$_2$ factor were initially studied in \cite{Dibitetto:2018gtk, Gauntlett:2006ns, Kim:2013xza, Chiodaroli:2009yw, Chiodaroli:2009xh, Corbino:2018fwb, Corbino:2020lzq}.
The recent works \cite{Dibitetto:2019nyz, Lozano:2020txg, Lozano:2020sae}
made precise and concrete the viewpoint advertised above for different infinite families of string backgrounds containing an AdS$_2$ factor.
This work presents a new infinite family of backgrounds with an AdS$_2$ factor. We focus our presentation mostly on geometrical aspects of the new type IIB solutions. The contents of this paper are distributed as follows.
In Section \ref{B-type}, we present the new backgrounds preserving eight supersymmetries (four Poincar\'e and four conformal SUSYs). We study the conserved brane charges and
deduce the associated brane set-up, consisting on D1 and D5 'colour' branes (dissolved into fluxes) with D3 and D7 'source' branes (present in the background and violating Bianchi identities). NS-five branes and fundamental strings complete this configuration. We define the holographic central charge following the procedure and physical meaning advanced in \cite{Lozano:2020txg}. The section is closed with a brief discussion of the dual SCQM. In Section \ref{CGKsection} we connect our backgrounds with those presented in \cite{Chiodaroli:2009yw, Chiodaroli:2009xh}. We point out that the presence of sources in our solutions extend (for the AdS$_2$ fixed point) the results of \cite{Chiodaroli:2009yw, Chiodaroli:2009xh}. We also link the solutions in \cite{Lozano:2020txg} with those of \cite{Chiodaroli:2009yw, Chiodaroli:2009xh} (under the above mentioned restrictions). These links require a zoom-in procedure that we discuss in detail.
In Section \ref{NATD} we uncover a new and explicit infinite family of solutions of cohomogeneity-two, by applying non-Abelian T-duality on the AdS$_3$ backgrounds of \cite{Lozano:2019emq, Lozano:2019jza, Lozano:2019zvg, Lozano:2019ywa}. The study of these backgrounds and their 'completion' following the ideas of \cite{Lozano:2016kum, Lozano:2016wrs, Lozano:2017ole, Itsios:2017cew, Lozano:2018pcp}
is reserved for a future study. We extend to the families of backgrounds discussed in this work a relation uncovered in \cite{Lozano:2020txg},\cite{Lozano:2020sae} between the Ramond-Ramond sector of the backgrounds and the holographic central charge. Such relation is discussed In Section \ref{maxwellcentralhol}. A functional whose extremisation yields the central charge is also presented.
Finally, Section \ref{conclusions} gives a short summary of the work, together with some ideas to work on the future.
\section{New AdS$_2\times \text{S}^2\times \text{CY}_2$ backgrounds}\label{B-type}
\paragraph{}In this section we present a new family of AdS$_2$ solutions with $\mathcal{N}=4$ Poincar\'e supersymmetry in Type IIB supergravity. These geometries are foliations of $\text{AdS}_2\times$S$^2\times$CY$_2\times$S$^1$ over an interval. Alternatively, they can be considered as foliations of $\text{AdS}_2\times$S$^2\times$CY$_2$ over a 2d Riemann surface $\Sigma$ with the topology of an annulus.
The NS-NS sector of our solutions reads,
\begin{equation}\label{NS sector-B}
\begin{split}
\text{d}s_{st}^2 &= \frac{u \sqrt{\widehat{h}_4 h_8} }{4 \widehat{h}_4 h_8-(u')^2} \text{d}s^2_{\text{AdS}_2} + \frac{u}{4\sqrt{\widehat{h}_4 h_8}} \text{d}s^2_{\text{S}^2} + \sqrt{\frac{\widehat{h}_4}{h_8}} \text{d}s^2_{\text{CY}_2} + \frac{\sqrt{\widehat{h}_4 h_8}}{u} (\text{d} \psi^2+\text{d} \rho^2 )\, , \\
e^{- 2 \phi}&= \frac{h_8}{4\widehat{h}_4} \big(4 \widehat{h}_4 h_8 - (u')^2 \big) \, , \\ H_{3} &=- \frac{1}{2} \text{d} \bigg( \rho + \frac{u u'}{4 \widehat{h}_4 h_8 - (u')^2} \bigg) \wedge \text{vol}_{\text{AdS$_2$}}+ \frac{1}{h_8^2} \text{d} \rho \wedge H_2+\frac{1}{2} \text{vol}_{\text{S$^2$}} \wedge \text{d} \psi \, .
\end{split}
\end{equation}
Here $\phi$ is the dilaton, $H_3$ the NS-NS three-form and the metric is given in string frame. A prime denotes a derivative with respect to $\rho$. The two-form $H_2$ is defined on the $\text{CY}_2$.
The coordinate $\psi$ ranges in $[0, 2\pi]$, while the $\rho$ coordinate describes an interval that we will take to be bounded between $0$ and $2\pi(P+1)$ (see below). Note that $u\ge 0$ and $4 \widehat{h}_4 h_8 - (u')^2 \ge 0$ must be imposed to have a positive definite metric.
The background in eq.(\ref{NS sector-B}) is supported by the RR fluxes,
\begin{equation}\label{RR sector lower rank fluxes}
\begin{split}
F_{1} &= h_8' \text{d} \psi \, ,\\ F_{3} &=- \frac{1}{h_8}H_2\wedge \text{d} \psi -\frac{1}{2} \Big( h_8 + \frac{h_8' u' u}{4 h_8 \widehat{h}_4 - (u')^2} \Big) \text{vol}_{\text{AdS$_2$}} \wedge \text{d} \psi + \frac{1}{4} \left(- \text{d} \bigg(\frac{u'u}{2 \widehat{h}_4} \bigg) + 2 h_8 \text{d} \rho \right) \wedge \text{vol}_{\text{S$^2$}} \, , \\
F_{5} &= -(1+\star_{10})\left(\partial_{\rho} \widehat{h}_4 \text{vol}_{\text{CY$_2$}} +\frac{h_8}{u} \widehat{\star}_4 \text{d}_4 h_4 \wedge d \rho - \frac{u' u}{2h_8(4 \widehat{h}_4 h_8 - (u')^2)} H_2 \wedge \text{vol}_{\text{AdS$_2$}} \right)\wedge\text{d}\psi.\\
F_{7} & =\left( \frac{1}{2}\left(\widehat{h}_4+\frac{uu'\widehat{h}_4'}{4 h_8 \widehat{h}_4-(u')^2}\right)\text{vol}_{\text{AdS$_2$}}\wedge\text{d} \psi-\frac{1}{4}\left( 2\widehat{h}_4\text{d}\rho-\text{d}\left(\frac{uu'}{2h_8}\right)\right)\wedge\text{vol}_{\text{S}^2}\right)\wedge\text{vol}_{\text{CY$_2$}} \, \\&-\star_{10}\left(\frac{1}{h_8}H_2\wedge \text{d} \psi\right), \\
F_{9} &=\frac{\widehat{h}_4 h_8' u^2}{4 \widehat{h}_8 (4 \widehat{h}_4 h_8 - (u')^2)} \text{vol}_{\text{AdS$_2$}} \wedge\text{vol}_{\text{CY$_2$}} \wedge \text{vol}_{\text{S$^2$}} \wedge \mathrm{d} \rho \, .
\end{split}
\end{equation}
Supersymmetry holds whenever,
\begin{eqnarray}
\label{susyC}
u''=0,~~~~ H_2+ \widehat{\star}_4 H_2=0,
\end{eqnarray}
where $\widehat{\star}_4$ is the Hodge dual on CY$_2$. In turn, the Bianchi identities of the fluxes impose--away from localised sources--that,
\begin{eqnarray}
\begin{split}
h_8''=0\, , \qquad \mathrm{d}H_2=0, \qquad\frac{h_8}{u}\nabla_{\text{CY}_2}^2\widehat{h}_4+\partial_\rho^2\widehat{h}_4-\frac{1}{h_8^3}H_2\wedge H_2=0.
\end{split}
\end{eqnarray}
In what follows we will concentrate on backgrounds for which $H_2=0$ and $\widehat{h}_4=\widehat{h}_4(\rho)$. These backgrounds are supersymmetric solutions of the Type IIB equations of motion if the warping functions satisfy (away from localised sources),
\begin{eqnarray}
\widehat{h}_4''=0,\qquad h_8''=0,\qquad u''=0,
\end{eqnarray}
which makes them linear functions of $\rho$.
We focus on the solutions defined by the piecewise linear functions $\widehat{h}_4, h_8$ considered in \cite{Lozano:2019jza,Lozano:2019zvg}. These are continuous functions with discontinuous derivatives. These imply discontinuities in the RR-sector that are interpreted as generated by sources in the background. The solutions in \cite{Lozano:2019jza,Lozano:2019zvg} have well-defined 2d dual CFTs. This requires a global definition of the $\rho$-interval. We achieve this imposing that $\widehat{h}_4$ and $h_8$ vanish at both ends of the $\rho$-interval, that we take at $\rho=0,2\pi (P+1)$. Ending the space in this fashion, introduces extra source branes in the configuration. For the backgrounds to be trustable (in view of holographic applications), we need to impose that the sources are 'sparse', namely that they occur separated enough in the $\rho$-interval. This imposes that $P$ (the length of the $\rho$-interval) is large.
The functions $\widehat{h}_4$ and $h_8$ are then defined as,
\begin{eqnarray} \label{profileh4sp}
\widehat{h}_4(\rho)\!=\!\Upsilon\! \,h_4(\rho)\!=\!\!
\Upsilon\;\!\!\left\{ \begin{array}{cccrcl}
\frac{\beta_0 }{2\pi}
\rho & 0\leq \rho\leq 2\pi, &\\
\alpha_k\! +\! \frac{\beta_k}{2\pi}(\rho-2\pi k) &~~ 2\pi k\leq \rho \leq 2\pi(k+1),& ~~k=1,...,P-1\\
\alpha_P- \frac{\alpha_P}{2\pi}(\rho-2\pi P) &~~ 2\pi P\leq \rho \leq 2\pi(P+1),&
\end{array}
\right.\\
\label{profileh8sp}
h_8(\rho)
=\left\{ \begin{array}{cccrcl}
\frac{\nu_0 }{2\pi}
\rho & 0\leq \rho\leq 2\pi, &\\
\mu_k+ \frac{\nu_k}{2\pi}(\rho-2\pi k) &~~ 2\pi k\leq \rho \leq 2\pi(k+1),& ~~k=1,...,P-1\\
\mu_P- \frac{\mu_P}{2\pi}(\rho-2\pi P) &~~ 2\pi P\leq \rho \leq 2\pi(P+1).&
\end{array}
\right.
\end{eqnarray}
The choice of constants is imposed by continuity of the metric and dilaton. This implies that
\begin{equation}
\alpha_k=\sum_{j=0}^{k-1} \beta_j, \qquad \mu_k=\sum_{j=0}^{k-1} \nu_j.
\end{equation}
In turn, $\beta_k$ and $\nu_k$ must be integer numbers to give well defined quantised charges (see the next subsection).
In (\ref{profileh4sp}) the number $\Upsilon$ is chosen such that,
\begin{equation}
\Upsilon \text{Vol}_{\text{CY$_2$}}= 16\pi^4.\label{upsilonnorm}
\end{equation}
In most of our analysis in this paper we will concentrate on solutions for which $u=u_0=$ constant. In that case the behaviour of the metric and dilaton at both ends of the $\rho$-interval is
\begin{equation}\label{O3O7brane-singularity}
\begin{split}
\text{d}s^2 \sim \frac{1}{x}(\text{d}s^2_{\text{AdS}_2} + \text{d}s^2_{\text{S}^2}) + \; \text{d}s^2_{\text{CY}_2} + x\; (\text{d} x^2+\text{d} \psi^2 )\, , \qquad
e^{-\phi}\sim x\, ,
\end{split}
\end{equation}
where $x=\rho$ close to $\rho=0$ and $x=2\pi (P+1)-\rho$ close to $\rho=2\pi (P+1)$.
This corresponds to a superposition of D3-branes, extended on AdS$_2\times$S$^2$ and smeared on $\psi$ and the $\text{CY}_2$, and D7-branes, extended on AdS$_2\times$S$^2\times$CY$_2$ and smeared on $\psi$.
The backgrounds in eqs.\eqref{NS sector-B}-\eqref{RR sector lower rank fluxes} can be obtained applying the usual T duality rules over the Hopf fibre of the three sphere of the $\text{AdS}_2\times$S$^3$ backgrounds in \cite{Lozano:2020sae}. Additionally, these solutions have the same structure as the geometries in \cite{Lozano:2020txg}, namely $\text{AdS}_2\times$S$^2\times$CY$_2\times$S$^1$ foliated over an interval. The relation with the backgrounds in \cite{Lozano:2020txg} is through an analytic continuation,
\begin{equation}
\begin{split}
&\text{d}s_{\text{AdS}_2}^2\to-\text{d}s_{\text{S}^2}^2 ,\qquad \text{d}s_{\text{S}^2}^2\to-\text{d}s_{\text{AdS}_2}^2, \qquad e^{\phi}\to ie^{\phi},\qquad F_{i}\to -iF_{i},\\
&u\to -iu,\qquad \widehat{h}_4\to i\widehat{h}_4, \qquad h_8\to ih_8, \qquad\rho\to i\rho,\qquad\psi\to-i\psi, \qquad g_{i}\to i g_{i}.\label{diagramaeq}
\end{split}
\end{equation}
These relations are summarised in Figure \ref{diagrama}.
\begin{figure}[h]
\centering
\includegraphics[scale=0.73]{Figures/diagrama.pdf}
\caption{Relations between the infinite family of AdS$_3$ backgrounds to massive IIA constructed in \cite{Lozano:2019emq} (top left), the IIB AdS$_2$ backgrounds studied in \cite{Lozano:2020txg} (bottom left), the IIA AdS$_2$ backgrounds constructed in \cite{Lozano:2020sae} (top right), and the new AdS$_2$ solutions in Type IIB given by eqs.(\ref{NS sector-B})-(\ref{RR sector lower rank fluxes}) (bottom right).}
\label{diagrama}
\end{figure}
Next we study the charges associated with the backgrounds in eqs. \eqref{NS sector-B}-\eqref{RR sector lower rank fluxes} and the associated brane set-up.
\subsection{Brane charges and brane set-up}
\paragraph{}
We compute the charges associated to our backgrounds using that the magnetic charge for a Dp brane is given by,
\begin{eqnarray}\label{chargemagnetic}
Q_{\text{Dp}}^{m
=\frac{1}{(2\pi)^{7-p} }\int_{\mathcal{M}_{8-p}} \widehat{F}_{8-p},\end{eqnarray}
where $\mathcal{M}_{8-p}$ is any $(8-p)$-dimensional compact manifold transverse to the branes.
In turn, the electric charge of Dp branes is defined by,
\begin{eqnarray}
Q_{\text{Dp}}^e=\frac{1}{(2\pi)^{p+1}}\int_{\text{AdS}_2\times \Sigma_{p}} \widehat{F}_{p+2},\label{electric-charges}
\end{eqnarray}
where $\Sigma_p$ is the $p$-dimensional manifold on which the brane extends.
In both expressions we have set $\alpha'= g_s = 1 $. For the electric charges we need to regularise the volume of the AdS$_2$ space. We take it to be the analytical continuation of the volume of the two-sphere,
\begin{equation} \label{regpres}
\text{Vol}_{\text{AdS}_2}=4\pi\, .
\end{equation}
In the previous expressions $\widehat{F}$ are the Page fluxes, defined as $\widehat{F}=F\wedge e^{-B_{2}}$.
They read, for our backgrounds
\begin{equation}\label{fluxes23}
\begin{split}
\widehat{F}_{1} &= h'_8 \, \mathrm{d} \psi \, , \\
\widehat{F}_{3} &= \frac{1}{2} \left(h'_8(\rho-2\pi k) - h_8\right) \, \text{vol}_{\text{AdS$_2$}} \wedge \mathrm{d} \psi + \frac{1}{4} \bigg(2h_8+\frac{u'(u\widehat{h}_4'-\widehat{h}_4u')}{2\widehat{h}_4^2} \bigg)\text{vol}_{\text{S$^2$}}\wedge \text{d}\rho \, , \\
\widehat{F}_{5} &=\frac{1}{4}\left(h_8(\rho-2\pi k)- \frac{\left(u - (\rho-2\pi k) u' \right) ( u \widehat{h}_4' - \widehat{h}_4 u' )}{4 \widehat{h}_4^2}\right)
\text{vol}_{\text{AdS$_2$}} \wedge \text{vol}_{\text{S$^2$}} \wedge \mathrm{d} \rho - \widehat{h}_4' \text{vol}_{\text{CY$_2$}} \wedge \mathrm{d} \psi \, , \\
\widehat{F}_{7} &= \frac{1}{2} (\widehat{h}_4 - (\rho-2\pi k) \widehat{h}_4')\; \text{vol}_{\text{AdS$_2$}} \wedge \text{vol}_{\text{CY$_2$}} \wedge \mathrm{d} \psi - \frac{1}{4}\left( 2\widehat{h}_4+\frac{u'(uh_8'-h_8u')}{2h_8^2} \right) \text{vol}_{\text{S$^2$}} \wedge \text{vol}_{\text{CY$_2$}}\wedge\text{d}\rho \, ,\\
\widehat{F}_{9}&= -\frac{1}{4}\left((\rho-2\pi k) \widehat{h}_4 -\frac{\left(u - (\rho-2\pi k) u' \right) ( u h_8' - h_8 u')}{4h_8^2}\right)
\text{vol}_{\text{AdS$_2$}} \wedge \text{vol}_{\text{S$^2$}}\wedge \text{vol}_{\text{CY$_2$}}\wedge d\rho.
\end{split}
\end{equation}
In these expressions we have allowed for large gauge transformations of the $B_2$-field, $B_2\to B_2+ k \pi \text{vol}_{\text {AdS$_2$}}$, as in \cite{Lozano:2020sae} (see this reference for more details).
Before calculating the quantised charges associated to these fluxes it is useful to compute the following quantities,
\begin{equation}\label{caxa2}
\begin{split}
\text{d}\widehat{F}_{1} &= h_8'' \text{d}\rho\wedge \text{d}\psi \,, \quad
\text{d}\widehat{F}_{3} = \frac{1}{2}h_8'' (\rho-2\pi k) \text{vol}_{\text{AdS$_2$}} \wedge\mathrm{d} \rho\wedge \mathrm{d} \psi\, , \quad
\text{d}\widehat{F}_{5} = -h_4'' \; \text{vol}_{\text{CY$_2$}} \wedge \mathrm{d}\rho\wedge\mathrm{d} \psi \, , \\
\text{d}\widehat{F}_{7} &= -\frac{1}{2}h_4'' (\rho-2\pi k) \text{vol}_{\text{AdS$_2$}} \wedge \text{vol}_{\text{CY$_2$}} \wedge\mathrm{d}\rho \wedge \mathrm{d} \psi\, , \quad
\text{d}\widehat{F}_{9} = 0 \, .
\end{split}
\end{equation}
In these expressions $\widehat{h}_4''$ and $h_8''$ are the ones that follow from eqs.\eqref{profileh4sp}-\eqref{profileh8sp},
\begin{eqnarray}
\label{nana}
& & \widehat{h}_4''=\frac{1}{2\pi}\sum_{k=1}^P (\beta_{k-1}-\beta_{k}) \delta(\rho-2\pi k),\;\;\;\; h_8''=\frac{1}{2\pi}\sum_{k=1}^P (\nu_{k-1}-\nu_{k} )\delta(\rho-2\pi k),\\
& & \widehat{h}_4'' \times(\rho-2\pi k) = h_8'' \times(\rho-2\pi k) =x \delta(x)=0.\nonumber
\end{eqnarray}
We then obtain
\begin{equation}
\text{d}\widehat{F}_{3}=\text{d}\widehat{F}_{7}=0,
\end{equation}
and
\begin{eqnarray}
\text{d}\widehat{F}_{1} &=&\frac{1}{2\pi}\sum_{k=1}^P (\nu_{k-1}-\nu_{k} )\delta(\rho-2\pi k) \,\text{d}\rho\wedge \text{d}\psi \\
\text{d}\widehat{F}_{5} &=& -\frac{1}{2\pi}\sum_{k=1}^P (\beta_{k-1}-\beta_{k}) \delta(\rho-2\pi k)\, \text{vol}_{\text{CY$_2$}} \wedge \mathrm{d}\rho\wedge\mathrm{d} \psi.
\end{eqnarray}
These results can be put in correspondence with the brane set-up summarised in Table \ref{Table brane web set type IIA2}.
\begin{table}[ht]
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|}
\hline
& $x^0=t$ & $x^1$ & $x^2$ & $x^3$ & $x^4$ & $x^5=\rho$ & $x^6=r$ & $x^7=\theta_1$ & $x^8=\theta_2$ & $x^9=\psi$\\
\hline
D1 & x & & & & & $$ & & &$$ &x \\
\hline
D3 & x & & & & & & x & x & x & \\
\hline
D5 & x & x & x & x & x & $$ & & & &x \\
\hline
D7 & x & x & x & x & x & & x & x & x & \\
\hline
NS5 & x & x & x & x & x & & x& & & $$ \\
\hline F1 &x & $$ & $$ & $$ & $$ & $$& x & & & $$ \\ \hline
\end{tabular}
\caption{Brane set-up associated to our solutions. $x^0$ corresponds to the time direction of the ten dimensional spacetime, $x^1, \dots , x^4$ are the coordinates spanned by the CY$_2$ and $ x^7 , x^8$ are the coordinates parametrising the $S^2$.}
\label{Table brane web set type IIA2}
\end{center}
\end{table}
The fact that $\text{d}\widehat{F}_{3}=0$ and $\text{d}\widehat{F}_{7}=0$ indicates that the D5 and D1 branes play the role of colour branes (dissolved in fluxes) in the brane set-up. On the other hand, $\text{d}\widehat{F}_{1}$ and $\text{d}\widehat{F}_{5}$ being nonzero indicate that the D7 and D3 branes are flavour branes, that is, explicit sources with dynamics described by the Born-Infeld-Wess-Zumino action.
Substituting $\widehat{h}_4$ and $h_8$ as defined by eqs.(\ref{profileh4sp}) and (\ref{profileh8sp}), together with eqs.(\ref{upsilonnorm}) and (\ref{regpres}), we find, in each $\rho$-interval $[2\pi k,2\pi(k+1)]$
\begin{eqnarray}
\begin{split}
Q_{\text{D1}}^{e}&=
\frac{1}{(2\pi)^2}\int_{\text{AdS}_2\times\text{S}_\psi} \widehat{F}_3=\left(\frac{ \text{Vol}_{\text{AdS$_2$} }}{4\pi} \right)\left(\frac{ \text{Vol}_{\psi}}{\pi} \right)\frac{h_8-h'_8(\rho-2\pi k)}{2}=\mu_k,\\
Q_{\text{D3}}^{m}&=\frac{1}{16\pi^4}\int_{\text{CY}_2\times\text{S}_{\psi}} \widehat{F}_5= \frac{1}{16\pi^4}\int_{\text{CY}_2\times\text{S}_{\psi}\times\text{I}_\rho} \text{d}\widehat{F}_5=\left(\frac{ \Upsilon \text{Vol}_{\text{CY$_2$} }}{16\pi^4} \right) \times \text{Vol}_\psi~ \int h_4''\text{d}\rho=(\beta_{k-1}-\beta_{k}),\\
Q_{\text{D5}}^{e}&=\frac{1}{(2\pi)^6}\int_{\text{AdS}_2\times\text{CY}_2\times\text{S}_{\psi}} \widehat{F}_7=\left(\frac{ \text{Vol}_{\text{AdS$_2$} }}{4\pi} \right)\left(\frac{ \Upsilon \text{Vol}_{\text{CY$_2$} }}{16\pi^4} \right)\left(\frac{ \text{Vol}_{\psi}}{\pi} \right)\frac{h_4-h'_4(\rho-2\pi k)}{2}=\alpha_k,\\
Q_{\text{D7}}^{m}&=\int_{\text{S}_\psi} \widehat{F}_1=\int_{\text{S}_{\psi}\times\text{I}_\rho} \text{d} \widehat{F}_1= \text{Vol}_\psi ~\int h_8'' \text{d}\rho=(\nu_{k-1}-\nu_{k}).
\label{cargaspageB}
\end{split}
\end{eqnarray}
Further, in the brane set-up the F1-strings are electrically charged with respect to the NS-NS 3-form $H_3$ while the NS5 branes are magnetically charged,
\begin{eqnarray}
\begin{split}
Q_{\text{F1}}^{e}&=\frac{1}{(2\pi)^2}\int_{\text{AdS}_2\times\text{I}_\rho} H_3=\left(\frac{ \text{Vol}_{\text{AdS$_2$} }}{4\pi} \right)\left(\frac{1}{2\pi} \right)\int_{2\pi k}^{2\pi (k+1)} \text{d}\rho=1,\\
Q_{\text{NS5}}^{m}&=\frac{1}{(2\pi)^2}\int_{\text{S}^2\times\text{S}_{\psi}} H_3=\left(\frac{ \text{Vol}_{\text{S$^2$} }}{4\pi} \right)\left(\frac{ \text{Vol}_{\psi}}{2\pi} \right)=1 .
\end{split}
\end{eqnarray}
\subsection{Holographic central charge}\label{hccharge}
\paragraph{} To close this part of our study we compute the holographic central charge associated to our solutions. Being the field theory zero-dimensional, the previous quantity should be interpreted as the number of vacuum states in the dual superconformal quantum mechanics (see \cite{Lozano:2020txg, Lozano:2020sae} for a further discussion of the physical meaning of this quantity). We follow the prescription in \cite{Macpherson:2014eza, Bea:2015fja}.
We get for the internal volume,
\begin{equation}\label{Vint-Btype}
\begin{split}
V_{\text{int}} &= \int \text{d}^8x\sqrt{e^{-4\phi}\;\text{det}\;g_{8,ind}} = \frac{\text{Vol}_{\text{CY}_2} \text{Vol}_{\text{S}^2} \text{Vol}_{\psi}}{4^2} \int_0^{2\pi(P+1)}(4\widehat{h}_4h_8-(u')^2)\;\text{d}\rho \, ,
\end{split}
\end{equation}
and, finally, for the central charge
\begin{equation}\label{hccharge-B}
c_{\text{hol,1d}}=\frac{3 V_{\text{int}}}{4\pi G_N} =\frac{3}{\pi}\int_0^{2\pi(P+1)} \left({h}_4h_8 - \frac{(u')^2}{4\Upsilon} \right) \text{d} \rho \, .
\end{equation}
We have used that $G_N=8\pi^6$ and set units so that $\alpha'=g_s=1$.
We would like to stress that in the usual calculations, such as the previous one, giving rise to the holographic central charge, only the NS-NS sector of the backgrounds needs to be taken into account. We will point out an interesting relation between the holographic central charge and the RR sector of our AdS$_2$ solutions in Section \ref{maxwellcentralhol}. Such relation has been previously encountered in the AdS$_2$ solutions constructed in \cite{Lozano:2020txg,Lozano:2020sae}.
\subsection{Aspects of the dual Conformal Quantum Mechanics }
\paragraph{}
Whilst the main focus of this work in not the Quantum Mechanical analysis of the duals to the backgrounds in eqs.(\ref{NS sector-B})-(\ref{RR sector lower rank fluxes}), we add below some thoughts along this direction.
In the papers \cite{Lozano:2019jza,Lozano:2019zvg,Lozano:2020txg,Lozano:2020sae} concrete quivers were proposed
as UV-descriptions of weakly coupled 2d QFTs or 1d Quantum Mechanics. It was conjectured that these quivers become strongly coupled at low energies and a conformal fixed point arises. Checks for these proposals were presented in each of the works \cite{Lozano:2019jza,Lozano:2019zvg,Lozano:2020txg,Lozano:2020sae}, for the different systems under study. These checks deal with RG-invariant quantities that can be well-identified in the UV and IR descriptions.
As we indicated around eq.(\ref{diagramaeq}) and summarised in Figure \ref{diagrama}, the backgrounds of Section \ref{B-type} arise after an Abelian T-duality on the backgrounds of \cite{Lozano:2020sae}.
This suggests that the quantum mechanical system proposed in \cite{Lozano:2020sae} should also apply here.
We are in fact T-dualising across a non-R-symmetry-direction, hence we expect the amount of SUSY to be the same. The R-symmetry of the quivers in \cite{Lozano:2020sae} is $SU(2)_R$, and there is also
a global $SU(2)_g$ symmetry. We are choosing a $U(1)_g$ inside $SU(2)_g$ for our dualisation. Therefore, our dual quantum mechanical system should have $SU(2)_R\times U(1)_g$ symmetry. This is in fact
geometrically realised by the presence of the round S$^2$ and the circle S$^1_\psi$ in the backgrounds of Section \ref{B-type}.
Since the string sigma model in a background and in its T-dual is the same, we expect the same dual quantum mechanical systems
for our backgrounds as those for the backgrounds \cite{Lozano:2020sae} (only that perhaps it will be written in a different language).
Using this reasoning, we may think about the SCQM as that arising in the very low energy limit of a system of D3-D7 branes---dual to a four dimensional ${\cal N}=2$ QFT. This system is 'polluted' by
one-dimensional defects. These are Wilson loops (arising from F1-D5) and 't Hooft loops (arising from NS5-D1) added to the background, see for example \cite{Assel:2019iae}. Note that both the D1's and the D5's
extend on the $\psi$-isometric direction.
From the discussion above, it follows that the dual SCQM to our backgrounds is the description of these one-dimensional defects inside a four dimensional ${\cal N}=2$
QFT. In fact, in the IR the gauge symmetry on both D7 and D3 branes should become global. This implies that these branes must be sources/flavours, as it occurs in the backgrounds of Section \ref{B-type}.
By the same token we have two lines of one dimensional gauge groups: $\Pi_{i=1}^{P} U(\alpha_i)$ and $\Pi_{i=1}^{P} U(\mu_i)$ realised on D5 and D1 branes in each $\rho$-interval.
This is reflected by the counting of branes of eq.(\ref{cargaspageB}). The nodes in the $[2\pi k, 2\pi(k+1)]$ interval will have $SU(\beta_k-\beta_{k+1})$ and $SU(\nu_k-\nu_{k+1})$ flavour groups, realised
on the D3 and D7 branes,
as also reflected by eq.(\ref{cargaspageB}). The brane set-up is the one described in Table \ref{Table brane web set type IIA2}.
As was found in \cite{Lozano:2020sae}, our 1d system should also have Wilson lines (in an antisymmetric representation) inserted in the different gauge nodes of the quiver. These Wilson lines arise from the
massive fermionic strings that stretch between D1s in the $k$-th interval and D7s in all other intervals. The Wilson lines would be in the $(\nu_0,\dots, \nu_{k-1}$) antisymmetric representation of the $U(\mu_k)$ gauge group. The same applies to the massive D3-D5 fermionic strings and the antisymmetric Wilson lines
on the $U(\alpha_k)$ groups. As in \cite{Lozano:2020sae}, this information can be encoded in Young diagrams.
We would also have a dynamical CS-term of each gauge group. This comes from the massless fermionic strings
stretched between D1-D7 and D5-D3 branes. The coefficient can be extracted studying the WZ action for a D1 along $[t,\psi]$ and a D5 along $[t, \text{CY}_2, \psi]$. As expected, these coefficients are quantised.
The field content of the UV-quantum mechanical quiver
follows directly from the analysis of Appendix B in \cite{Lozano:2020sae}. In fact, each node contains a $(4,4)$ vector multiplet and a $(4,4)$ adjoint hyper, $(0,4)$ bifundamental hypers join the two types of colour, D5 and D1, branes, (4,4) bifundamental hypers join the D7 sources with D5-branes, and the
D3 sources with D1 branes, respectively. Finally, $(0,2)$ Fermi multiplets join source D7 with colour D1s and source D3 with colour D5 branes. The quiver diagram is
depicted in Figure \ref{quiverprop}.
\begin{figure}[h]
\centering
\includegraphics[scale=0.55]{Figures/III.pdf}
\caption{The proposed quantum mechanical quiver. This follows from the analysis of open strings in the Hanany-Witten set-up.}
\label{quiverprop}
\end{figure}
\section{Connection with the $\text{AdS}_2\times \text{S}^2\times \text{CY}_2\times \Sigma$ backgrounds of Chiodaroli-Gutperle-Krym} \label{CGKsection}
\paragraph{} In this section we relate our backgrounds to the general class of $\text{AdS}_2\times \text{S}^2\times \text{CY}_2\times \Sigma$ solutions to Type IIB supergravity with 8 supercharges found by Chiodaroli, Gutperle and Krym (CGK) in \cite{Chiodaroli:2009yw}. We show that our solutions fit locally in their classification in the absence of D3 and D7 brane sources (in this sense our backgrounds extend those in \cite{Chiodaroli:2009yw} at the AdS$_2$ point). A similar analysis shows that the family of $\text{AdS}_2$ solutions to Type IIB supergravity recently found in \cite{Lozano:2020txg} also fits in their general class.
\subsection{Review of the CGK geometries}\label{Gutperlething}
\paragraph{} The CGK backgrounds are dual to one dimensional conformal interfaces inside the two dimensional CFT associated to the D1-D5 system. These solutions (unlike ours) interpolate between AdS$_2$ in the IR (at the interface) and the AdS$_3\times \text{S}^3\times \text{CY}_2$ solution of Type IIB supergravity in the UV. We shall focus on the AdS$_2$ fixed points and compare them with both the backgrounds discussed in section \ref{B-type} and the solutions found in \cite{Lozano:2020txg}.
In \cite{Chiodaroli:2009yw}, the authors used techniques developed in \cite{DHoker:2007zhm, DHoker:2007mci} to find half BPS solutions that preserve eight of the sixteen supersymmetries of the AdS$_3\times$S$^3\times$CY$_2$ vacuum, and are locally asymptotic to this vacuum solution. They provided an ansatz for the bosonic fields in Type IIB supergravity for a foliation of AdS$_2\times$S$^2\times $CY$_2$ over a two-dimensional Riemann surface $\Sigma$ with a boundary, and found that
the local solutions of the BPS equations can be written in terms of two harmonic and two holomorphic functions defined on $\Sigma$. The solutions corresponds to a D1-D5 configuration with extra NS5 branes and fundamental strings, but vanishing D3 and D7 brane charges. We will see that our solutions fit locally within this class of solutions in the absence of D3 and D7 brane sources. Our D3 and D7 sources are localised in $\rho$ and smeared in the $\psi$-coordinate. The mapping explained below is valid at points in $\rho$ where the sources are not present.
We start summarising the local solutions constructed in \cite{Chiodaroli:2009yw}. The metric for the ten-dimensional spacetime is given, in Einstein frame, by
\begin{equation}\label{mmaacc}
\text{d}s^2=f_1^2\text{d}s_{\text{AdS}_2}^2+f_2^2\text{d}s_{\text{S}^2}^2+f_3^2\text{d}s_{\text{CY}_2}^2+\widetilde{\rho}^2\text{d}{z}\text{d}{\overline{z}} \, ,
\end{equation}
where the warping factors $f_i$ ($i=1, \dots, 3$) and $\widetilde{\rho}$ are functions of $z$ and $\overline{z}$, the local holomorphic coordinates of $\Sigma$. The orthonormal frames can be written as,
\begin{alignat}{2}
\begin{split}
f_1^2 \text{d}s_{\text{AdS}_2}^2 = \eta_{i_1 i_2} e^{i_1} \otimes e^{i_2} \, ,& \quad \quad \quad i_{1,2} = 0, 1\, ,\\
f_2^2 \text{d}s_{\text{S}^2}^2 = \delta_{j_1 j_2}e^{j_1} \otimes e^{j_2} \, ,& \quad \quad \quad j_{1, 2} = 2, 3 \, , \\
f_3^2 \text{d}s_{\text{CY}_2}^2 = \delta_{k_1 k_2} e^{k_1} \otimes e^{k_2} \, ,& \quad \quad \quad k_{1, 2} = 4, 5, 6, 7\, , \\
\widetilde{\rho}^2 \text{d}{z} \text{d}{\overline{z}} = \delta_{a b}e^{a} \otimes e^{b} \, ,& \quad \quad \quad a, b = 8, 9\, .
\end{split}
\end{alignat}
The NS-NS and RR three-forms are written as a complex three-form, defined as $G=e^{\Phi}H_{3} + i e^{-\Phi}(F_{3} - \chi \;H_{3})$. This form is given by,
\begin{equation}\label{3-formComplex}
G=g^{(1)}_ae^{a01}+g^{(2)}_ae^{a23} \, .
\end{equation}
In turn, the self-dual five-form flux is,
\begin{equation}\label{5-form}
F_{5} = h_ae^{a0123} + \widetilde{h}_ae^{a4567} \, , \quad a = z, \overline{z} \, ,
\end{equation}
where the self-duality condition implies $h_a=-\epsilon_a\,^b\widetilde{h}_b$.
The local solutions of the BPS equations and Bianchi identities admit a description in terms of four functions, $A$, $B$, $H$ and $K$. The analysis in \cite{Chiodaroli:2009yw} shows that the functions $A$ and $B$ must be holomorphic on the Riemann surface $\Sigma$, whilst $H$ and $K$ must be harmonic. The supergravity fields can be written in terms of these functions as,
\begin{equation}
\begin{split}
\label{mmbbx}
f_1^2&=\frac{e^{-2\Phi}|H|}{2f_3^2K}\left((A+\overline{A})K-(B-\overline{B})^2\right) \, , \quad\quad
f_2^2=\frac{e^{-2\Phi}|H|}{2f_3^2K}\left((A+\overline{A})K-(B+\overline{B})^2\right) \, , \\
f_3^4&=4\frac{e^{2\Phi}K}{A+\overline{A}} \, , \quad\quad
e^{4\Phi} = \frac{1}{4K^2}\left( ( A + \overline{A}) K - ( B + \overline{B})^2 \right) \left( ( A + \overline{A}) K - (B - \overline{B})^2 \right) \, ,\\
\chi &=\frac{1}{2 i K} \left( B^2 - \overline{B}^2 - ( A - \overline{A}) K \right) \, , \quad\quad \widetilde{\rho}^4=e^{2\Phi}K\frac{(A+\overline{A})}{H^2}\frac{|\partial_zH|^4}{|B|^4} \, .
\end{split}
\end{equation}
Here $\Phi=-\phi/2$, where $\phi$ is the dilaton.
For the five-form field strength, we define a four-form potential, along CY$_2$,
\begin{equation}\label{four-potential}
C_{\text{CY}_2} = -\frac{i}{2}\frac{ B^2 - \overline{B}^2}{ A + \overline{A}} - {2} \widetilde{K}, \qquad \qquad \partial_z C_{\text{CY}_2} = f_3^4 \widetilde{\rho}\;\widetilde{h}_z \, ,
\end{equation}
where $\widetilde{K}$ is the harmonic function conjugate\footnote{\label{conjugatefoot}The harmonic conjugate of $g$ is denoted as $\widetilde{g}$ and satisfies $i\partial_z \widetilde{g}=\partial_z g$.} to $K$.
The potentials for the field strengths in equation \eqref{3-formComplex} are written in terms of the holomorphic and harmonic functions as
\begin{equation}
\begin{split}
\label{potential-b1}
b^{(1)}=-\frac{H(B+\overline{B})}{(A+\overline{A})K-(B+\overline{B})^2}-h_{1},\qquad &h_1=\frac{1}{2}\int\frac{\partial_zH}{B}+\text{c.c.}\;,\\
b^{(2)}=-i\frac{H(B-\overline{B})}{(A+\overline{A})K-(B-\overline{B})^2}+\widetilde{h}_{1},\qquad &\widetilde{h}_1=-\frac{i}{2}\int\frac{\partial_zH}{B}+\text{c.c.}\;,\\
c^{(1)}=-i\frac{H(A\overline{B}-\overline{A}B)}{(A+\overline{A})K-(B+\overline{B})^2}+\widetilde{h}_{2},\qquad &\widetilde{h}_2=-\frac{i}{2}\int\frac{A\;\partial_zH}{B}+\text{c.c.}\;,\\
c^{(2)}=-\frac{H(A\overline{B}+\overline{A}B)}{(A+\overline{A})K-(B-\overline{B})^2}+h_{2},\qquad &h_2=\frac{1}{2}\int\frac{A\;\partial_zH}{B}+\text{c.c.},\;
\end{split}
\end{equation}
where $\widetilde{h}_i$ and $h_i$ are harmonic functions conjugate to each other. In the previous expression $b^{(1)}$ and $b^{(2)}$ are the potentials of the NS-NS three-form $H_{3}$ and $c^{(1)}$ and $c^{(2)}$ are the potentials related to the RR three-form $F_{3}$. These read,
\begin{equation}
\begin{split}
H_3=&\text{d}b^{(1)}\wedge {\mathrm{vol}}_{{\mathrm{AdS}}_2}+\text{d}b^{(2)}\wedge {\mathrm{vol}}_{\text{S}^2}\\
F_3=&\mathrm d C_2-\chi H_3=(\mathrm d c^{(1)}-\chi\text{d}b^{(1)})\wedge {\mathrm{vol}}_{{\mathrm{AdS}}_2}+(\mathrm d c^{(2)}-\chi\text{d}b^{(2)})\wedge {\mathrm{vol}}_{\text{S}^2}.
\end{split}
\end{equation}
The existence of sensible regular solutions imposes the following conditions on the functions $A$, $B$, $H$ and $K$,
\begin{itemize}
\item The harmonic functions $A+\overline{A},B+\overline{B}$ and $K$ must have common singularities.
\item No singular points should appear in the bulk of the Riemann surface $\Sigma$.
\item The functions $A+\overline{A},K$ and $H$ cannot have any zero in the bulk of the Riemann surface.
\item The holomorphic functions $B$ and $\partial_zH$ must have common zeros.
\end{itemize}
The previous conditions guarantee a non-vanishing and finite everywhere $f_1$ (except at isolated singular points at the boundary), a finite $f_2$ in the interior of the Riemann surface and vanishing at the boundary, and, finally, finite and non-vanishing $f_3$ and $e^{2\Phi}$ functions everywhere on the Riemann surface, including the boundary.
The equations in \eqref{mmbbx} can be inverted to find $A$, $B$, $H$ and $K$ in terms of $f_i$ ($i=1,\dots, 3$), $\chi$ and $\Phi$. One finds two possibilities, that we will refer as the ``plus and minus solutions''\footnote{The ``plus solution'' corresponds to our AdS$_2$ backgrounds and the `` minus solutions'' to the AdS$_2$ geometries of \cite{Lozano:2020txg}. Both these solutions are related through an analytical continuation, as explained around eq.(\ref{diagramaeq}).},
\begin{eqnarray}
\label{kp-gen}
\text{Sol}_+:\; &&H = f_1f_2f_3^2,\quad K_{+}=\frac{f_1f_3^4}{2f_2},\quad A_{+}=\frac{f_1}{f_2}e^{2\Phi}-i\chi,\quad B_{+}=\frac{e^{\Phi}f_3^2}{2f_2}\sqrt{f_1^2-f_2^2}\;,\\
\label{km-gen}
\text{Sol}_-:\; &&H = f_1f_2f_3^2, \quad K_{-}=\frac{f_2f_3^4}{2f_1},\quad A_{-}=\frac{f_2}{f_1}e^{2\Phi}-i\chi,\quad B_{-}=i\;\frac{e^{\Phi}f_3^2}{2f_1}\sqrt{f_1^2-f_2^2}.
\end{eqnarray}
Inserting the ``plus-solution'', equation \eqref{kp-gen}, or the ``minus-solution'', equation \eqref{km-gen}, in the first expression of \eqref{four-potential} one obtains, in both cases, the function associated to the 4-form potential,
\begin{eqnarray}
C_{\text{CY}_2}=-2\widetilde{K},
\end{eqnarray}
where $\widetilde{K}$ is the harmonic function conjugate to $K$, according the footnote \ref{conjugatefoot}.
In the next subsection we obtain the harmonic and holomorphic functions that give rise to our backgrounds in eqs. \eqref{NS sector-B}-\eqref{RR sector lower rank fluxes}, as well as to the geometries in \cite{Lozano:2020txg}.
\subsection{Our AdS$_2$ geometries and the ``plus-solution''}\label{local-matching}
\paragraph{}
In order to compare the generic backgrounds given by eqs.\eqref{NS sector-B}-\eqref{RR sector lower rank fluxes} with the solutions in \cite{Chiodaroli:2009yw, Chiodaroli:2009xh} we express our solutions in Einstein frame, to agree with their conventions. We obtain,
\begin{eqnarray}
\begin{split}
\label{R-MetricB}
f_1^2=\frac{u}{\sqrt{2}}\left(\frac{\widehat{h}_4h_8^3}{(4\widehat{h}_4h_8-(u')^2)^3}\right)^{1/4},\qquad f_2^2=\frac{u}{\sqrt{2^{5}}}\left(\frac{4\widehat{h}_4h_8-(u')^2}{\widehat{h}_4^3h_8}\right)^{1/4},\\
f_3^2=\left(\frac{\widehat{h}_4(4\widehat{h}_4h_8-(u')^2)}{2^2h_8}\right)^{1/4},\qquad e^{2\Phi}=e^{-\phi}=\frac{1}{2}\sqrt{\frac{h_8}{\widehat{h}_4}} \sqrt{4\widehat{h}_4h_8-(u')^2},\\
\chi=h_8'\psi, \qquad \widetilde{\rho}^2=\frac{1}{\sqrt{2}u}
\left(\widehat{h}_4h_8^3(4\widehat{h}_4h_8-(u')^2)\right)^{1/4},\qquad C_{\text{CY}_2}=-h_4'\psi.
\end{split}
\end{eqnarray}
We emphasise that these expressions are valid at the points where $h_8''=\widehat{h}_4''=0$.
We take the $\rho$ and $\psi$ coordinates to define the real and imaginary parts of the $z$ variable. With this parametrisation, $\Sigma$ is an annulus, defined in the complex plane (see Figure \ref{Strip-ABack}),
\begin{equation}\label{StripA}
z=\psi + i \rho \qquad \text{with} \qquad \psi \in [0,2\pi] \qquad \text{and} \qquad \rho \in [0,2\pi(P+1)] \, .
\end{equation}
\begin{figure}[h]
\centering
\includegraphics[scale=0.55]{Figures/Riemann1.pdf}
\caption{Riemann surface associated to our AdS$_2$ geometries. Given the periodicity of $\psi$ it defines an annulus.}
\label{Strip-ABack}
\end{figure}
Locally, our solutions are defined by the three functions $u, \widehat{h}_4, h_8$, which must be linear in $\rho$. We take
\begin{equation}
\label{linearfunctions}
u=u_0+u_1\rho, \qquad h_8 = \mu+\nu\rho, \qquad \widehat{h}_4= \alpha+\beta\rho .
\end{equation}
Substituting \eqref{R-MetricB} in \eqref{kp-gen} and taking into account (\ref{linearfunctions}), we find
for the functions $A, B, H, K$,
\begin{equation}
\begin{split}
\label{A-BT}
A=&h_8-i\psi h_8'=\mu-i\nu z , \qquad
B=\frac{u'}{4}=\frac{u_1}{4} \, ,\\
H=&\frac{u}{4}=\frac{u_0}{4}-i \frac{u_1}{8}(z-\overline{z}) \, ,\qquad
K=\frac{\widehat{h}_4}{2}=\frac{\alpha}{2}-i\frac{\beta}{4}(z-\overline{z}) \, .
\end{split}
\end{equation}
It is easy to check that $H$, $K$, $A+\overline{A}$ and $B+\overline{B}$ are harmonic functions and $A$ and $B$ are holomorphic.
The harmonic function $ \widetilde{K}$ reads, in turn,
\begin{equation}
\label{k-ATHC}
\widetilde{K}=-\frac{\widehat{h}_4'}{4}(z+\overline{z})= -\frac{\beta}{{4}}(z+\overline{z}) \,,
\end{equation}
which is the harmonic function conjugate to the expression for $K$ in \eqref{A-BT}.
From the equations \eqref{potential-b1}, and using \eqref{A-BT}, we can then obtain the harmonic functions and potentials associated with the NS-NS three-form,
\begin{equation}
\begin{split}
&h_1=-\frac{i}{4}(z-\overline{z}) \, , \qquad \widetilde{h}_1 = - \frac{1}{4}(z +\overline{z}) \, ,\\
&b^{(1)} = \frac{u_1(2u_0-i u_1(z-\overline{z}))}{u_1^2+(2i\alpha+(z-\overline{z})\beta)(2i\mu+(z-\overline{z})\nu)} - h_1 \, , \qquad b^{(2)}=-\frac{1}{4}(z+\overline{z}) \, ,
\end{split}
\end{equation}
as well as those associated with the RR three-form,
\begin{equation}
\begin{split}
&h_2=-\frac{\nu}{8}(z^2+\overline{z}^2)-\frac{i}{4}\mu(z-\overline{z}) \, , \qquad \widetilde{h}_2 = i \frac{ \nu}{8}(z^2 - \overline{z}^2) -\frac{\mu}{4}(z+\overline{z})\,,\\
&c^{(1)} = \frac{u_1(2u_0-i u_1(z-\overline{z}))(z+\overline{z})\nu}{8(u_1^2+(2i\alpha+(z-\overline{z})\beta)(2i\mu+(z-\overline{z})\nu))} + \widetilde{h}_2 \, , \qquad c^{(2)} =-\frac{u_1 (2i u_0+u_1(z-\overline{z}))}{8 (\beta (z-\overline{z}) + 2 i\alpha)}+h_2 \, .
\end{split}
\end{equation}
From these expressions we can recover $H_3$ and $F_3$ as given in eqs. \eqref{NS sector-B}-\eqref{RR sector lower rank fluxes}. Note that $h_i$ and $\widetilde{h}_i$ are harmonic functions conjugate to each other.
We have thus shown that our solutions can be obtained, locally, from the class of solutions constructed in \cite{Chiodaroli:2009yw}. Note that in our analysis we have implicitly assumed that $h_8''=0$ and $\widehat{h}_4''=0$ also hold globally. This is necessary in order to match the axion and the 4-form RR potential given in (\ref{R-MetricB}). This assumption --translated to our geometries-- indicates that we are not allowing for D7 and D3 brane sources, according to equations \eqref{caxa2}-\eqref{nana}. This agrees with the analysis in \cite{Chiodaroli:2009yw}, which does not include either these types of branes.
We will show in subsection \ref{annulus1} that D3-brane sources can be included in the two boundaries of the annulus following the formalism for the annulus derived in \cite{Chiodaroli:2009xh}. This allows to recover the solutions in our class where D3-branes terminate the space at $\rho=2\pi (P+1)$. Quite surprisingly, we will also see that, even if not included in the analysis in \cite{Chiodaroli:2009xh}, D7-brane sources can also be allowed at the end of the space. We will show that they also manifest as (smeared) singularities of the basic harmonic function defined in the annulus in \cite{Chiodaroli:2009xh}.
Before that, we show in the next subsection that the AdS$_2$ geometries found in \cite{Lozano:2020txg}, that we will refer as LNRS geometries, fit as well in the CGK class.
\subsection{The LNRS geometries and the ``minus-solution''}
\label{LNRS-section}
\paragraph{} As we already mentioned in section \ref{B-type}, our class of geometries can be obtained through a double analytic continuation from the AdS$_2$ solutions studied in \cite{Lozano:2020txg}. In this section we show that the latter fit within the class of solutions referred as ``minus solutions'' in \cite{Chiodaroli:2009yw}.
The warping factors, dilaton, axion and RR 4-form potential associated to the AdS$_2$ geometries constructed in \cite{Lozano:2020txg} (in Einstein frame) are given by,
\begin{eqnarray}
\begin{split}
\label{WF-A}
f_1^2=\frac{u}{\sqrt{2^5}}\left(\frac{4\widehat{h}_4h_8+(u')^2}{\widehat{h}_4^3h_8}\right)^{1/4} \, , \qquad f_2^2=\frac{u}{\sqrt{2}}\left(\frac{\widehat{h}_4h_8^3}{(4\widehat{h}_4h_8+(u')^2)^3}\right)^{1/4} \, ,\\
f_3^2=\frac{1}{\sqrt{2}}\left(\frac{\widehat{h}_4(4\widehat{h}_4h_8+(u')^2)}{h_8}\right)^{1/4} \, , \qquad e^{2\Phi}=e^{-\phi}=\frac{1}{2}\sqrt{\frac{h_8}{\widehat{h}_4}} \sqrt{4\widehat{h}_4h_8+(u')^2} \, , \\
\chi=h_8'\psi \, , \qquad \widetilde{\rho}^2=\frac{1}{\sqrt{2}u} \left(\widehat{h}_4h_8^3(4\widehat{h}_4h_8+(u')^2)\right)^{1/4} ,\qquad C_{\text{CY}_2}=-h_4'\psi\, .
\end{split}
\end{eqnarray}
The Riemann surface is the same one defined in equation \eqref{StripA} and Figure \ref{Strip-ABack}, and, as in the previous subsection, we are also taking $h_8''=0$ and $\widehat{h}_4''=0$ globally, i.e. solutions without D7 and D3 brane sources. This is needed to obtain the axion and RR 4-form potential of the previous equations.
In this case the matching with the solutions in \cite{Chiodaroli:2009yw} is with the
``minus-solutions'' defined by equation \eqref{km-gen}. Taking into account \eqref{linearfunctions}, the harmonic and holomorphic functions read,
\begin{equation}
\begin{split}
\label{A-AT}
A=h_8-i\psi h_8'=\mu-i\;\nu z \, ,&\qquad
B=i\;\frac{u'}{4}=i\;\frac{u_1}{4} \, ,\\
H=\frac{u}{4}=\frac{u_0}{4}-i\frac{u_1}{8}(z-\overline{z}) \, ,&\qquad
K=\frac{\widehat{h}_4}{2}=\frac{\alpha}{2}-i\;\frac{\beta}{4}(z-\overline{z}) \, .
\end{split}
\end{equation}
As in the previous subsection, the functions $H$, $K$, $A + \overline{A}$ $B + \overline{B}$ are harmonic and $A$ and $B$ holomorphic. The harmonic function $\widetilde{K}$ reads exactly as in \eqref{k-ATHC}.
In turn, the harmonic functions that give rise to the NS-NS and RR three-forms read,
\begin{eqnarray}
\begin{split}
&h_1=-\frac{1}{4}(z+\overline{z}), \qquad \widetilde{h}_1=\frac{i}{4}(z-\overline{z}),\\
&h_2=-\frac{\mu}{4}(z+\overline{z})+i\frac{\nu}{8}(z^2-\overline{z}^2),\qquad \widetilde{h}_2=i\frac{\mu}{4}(z-\overline{z})+\frac{\nu}{8}(z^2+\overline{z}^2),\\
&b^{(1)} = \frac{1}{4}(z + \overline{z}) \, ,\qquad b^{(2)}=\frac{u_1(2u_0-iu_1(z-\overline{z}))}{4(u_1^2-(2i\mu+\nu(z-\overline{z}))(2i\alpha+\beta(z-\overline{z})))} + \widetilde{h}_1 \, ,\\
&c^{(1)} = -\frac{u_1 (u_1(z-\overline{z}) +2 i u_0)}{8(2i\alpha+\beta (z-\overline{z} ))} +\widetilde{h}_2 \, , \qquad c^{(2)} = \frac{\nu u_1(2u_0-iu_1(z-\overline{z}))(z+\overline{z})}{8(u_1^2-(2i\mu+\nu(z-\overline{z}))(2i\alpha+\beta(z-\overline{z})))} + h_2 \, .
\end{split}
\end{eqnarray}
From these expressions we recover the NS-NS and RR field strengths, $H_3$ and $F_3$, of the solutions in \cite{Lozano:2020txg}.
\subsection{The annulus}\label{annulus1}
\paragraph{}
As we have already mentioned, the class of solutions constructed in \cite{Chiodaroli:2009yw} have vanishing D3 and D7-brane charges. Those solutions have a Riemann surface with a single boundary component.
In the follow-up paper \cite{Chiodaroli:2009xh}, the authors constructed solutions in which the Riemann surface $\Sigma$ has an arbitrary number of boundaries and non-vanishing D3 brane charges. The D3-branes occur as poles of a basic harmonic function at the boundaries.
In this section we consider the simplest case of a Riemann surface with two disconnected boundary components, namely the annulus. We will then see in subsection \ref{annulus2} that we can recover the solutions with D3-brane sources at $\rho=2\pi (P+1)$, the end of the space. Quite surprisingly, we will see that D7-branes seem also allowed at the end of the space.
The annulus is defined as,
\begin{eqnarray}\label{annulus}
\Sigma\equiv\left\{w\in C,0\leq \text{Re}(w) \leq 1,0\leq \text{Im}(w) \leq\frac{t}{2}\right\}
\end{eqnarray}
with $t\in \mathbb{R}^+$. The points $w+1$ and $w$ are identified, thus giving the topology of an annulus.
Its two boundaries, $\partial\Sigma_{1,2}$, are located at Im$(w)=0$ and Im$(w)=\frac{t}{2}$. The annulus can be constructed from a double surface $\widehat{\Sigma}$, which is defined as a rectangular torus with periods 1 and $\tau$, where $\tau$ is a purely imaginary parameter, $\tau=i t$. The original surface $\Sigma$ is obtained as the quotient $\Sigma=\widehat{\Sigma}/\mathcal{J}$ where $\mathcal{J}(z)=\overline{z}$.
The construction of the solutions for the annulus in \cite{Chiodaroli:2009xh} proceeds in three steps. First, a basic harmonic function with singularities and suitable boundary conditions is constructed. Second, the harmonic functions, $A+\overline{A}$, $H$ and $K$, are expressed as linear superpositions of the basic harmonic function, evaluated at the various poles in the two boundaries. Finally, the meromorphic function $B$ is constructed such that it satisfies certain regularity conditions. Some of these conditions come from imposing that the solutions asymptote locally to the AdS$_3\times \text{S}^3\times\text{CY}_2$ background. These regularity conditions will not be satisfied by our solutions, first because they do not asymptote to this geometry and, second, because the D3-branes (also the D7-branes) are smeared in the $\psi$ direction. This introduces significant changes in the regularity analysis. For this reason we will not give a detailed account of the regularity conditions imposed in \cite{Chiodaroli:2009xh}. We will see however in the next subsection that our solutions can still be recovered from the general formalism in \cite{Chiodaroli:2009xh} in an appropriate limit.
The construction in \cite{Chiodaroli:2009xh} of the basic harmonic function is carried out in terms of elliptic functions and their related Jacobi theta function of the first kind,
\begin{eqnarray}\label{theta1}
\theta_{1}(w|\tau)=2\sum_{n=0}^{\infty}(-1)^n e^{i\pi\tau(n+\frac{1}{2})^2}\sin[(2n+1)w],
\end{eqnarray}
as follows,
\begin{eqnarray}\label{basic}
h_0(w,\overline{w})=i\left(\frac{\partial_{w}\theta_{1}(\pi w|\tau)}{\theta_{1}(\pi w|\tau)}+\frac{2\pi iw}{\tau}\right)+\text{c.c.}
\end{eqnarray}
This function has the following simple properties: it has a single simple pole on $\partial \Sigma$, it satisfies Dirichlet conditions away from the pole, and it is positive in the interior of $\Sigma$.
Notice that $h_0(w,\overline{w})$ has a singularity at $w=0$, on the first boundary. This pole can be shifted to any point on $\partial\Sigma_1$ by a real translation, so that $h_0(w-x,\overline{w}-x)$ has a singularity at $w=x$. Instead, to obtain the harmonic functions with singularities at $\partial\Sigma_2$ one needs to define,
\begin{eqnarray}\label{wp}
w'\equiv\frac{\tau}{2} -w.
\end{eqnarray}
Then the function $h_0(w'+y,\overline{w}'+y)$ has a pole at $w'=-y$, on the second boundary, for a real $y$. In other words the pole is localised at $w=y+it/2$.
In the annulus the harmonic functions $A+\overline{A}$, $B+\overline{B}$, $H$ and $K$ are expressed as linear combinations of $h_0$ harmonic functions with poles on both boundaries,
\begin{equation}
\begin{split}
A+\overline{A}=&\sum_{\ell_A=1} ^{M_A}r_{\ell_A}h_0(w-x_{\ell_A},\overline{w}-x_{\ell_A})+\sum_{j_A=1} ^{M'_A}r'_{j_A}h_0(w'+y_{j_A},\overline{w}'+y_{j_A})
\\
B+\overline{B}=&\sum_{\ell_B=1} ^{M_B}r_{\ell_B}h_0(w-x_{\ell_B},\overline{w}-x_{\ell_B})+\sum_{j_B=1} ^{M'_B}r'_{j_B}h_0(w'+y_{j_B},\overline{w}'+y_{j_B})
\\
H=&\sum_{\ell_H=1} ^{M_H}r_{\ell_H}h_0(w-x_{\ell_H},\overline{w}-x_{\ell_H})+\sum_{j_H=1} ^{M'_H}r'_{j_H}h_0(w'+y_{j_H},\overline{w}'+y_{j_H})
\\
K=&\sum_{\ell_K=1} ^{M_K}r_{\ell_K}h_0(w-x_{\ell_K},\overline{w}-x_{\ell_K})+\sum_{j_K=1} ^{M'_K}r'_{j_K}h_0(w'+y_{j_K},\overline{w}'+y_{j_K})\label{kann}.
\end{split}
\end{equation}
Each harmonic function is taken to have $M_i$ poles $x_{\ell_i}$ with $\ell_{i}=1,...,M_i$ on $\partial\Sigma_1$, and $M_i'$ poles $y_{j_i}$ with $j_{i}=1,...,M'_i$ on $\partial\Sigma_2$. The corresponding residues are $r_{\ell_i}$ and $r_{j_i}$.
In addition to the regularity conditions given in subsection \ref{Gutperlething}, the harmonic functions \eqref{kann} satisfy an extra condition coming from the requirement that $e^{4\Phi}>0$. Namely, $(A+\overline{A})K-(B+\overline{B})^2>0$ must be obeyed throughout $\Sigma$. Furthermore, in this language the first regularity condition can be written in terms of the residues as $r_Ar_K=r_B^2$.
\subsection{Zoom-in to our solutions}\label{annulus2}
\paragraph{} In this subsection we show that it is possible to recover well-defined global solutions with source branes at the ends of the space from the general analysis above for the annulus. These solutions do not satisfy most of the regularity conditions imposed in \cite{Chiodaroli:2009yw,Chiodaroli:2009xh}, and, moreover, contain not only D3 but also D7-brane sources at the ends of the space. Still, we will be able to recover them in a particular limit from the formalism in \cite{Chiodaroli:2009xh}.
As we have already stressed, the choice of constants in the general $\widehat{h}_4$ and $h_8$ functions defined by equations (\ref{profileh4sp}) and (\ref{profileh8sp}) allows for discontinuities in the RR sector of our backgrounds at each $\rho=2\pi k$ value, with $k=1,\dots, P$. The discontinuities in $\widehat{h}_4'$ are interpreted as generated by D3-brane sources, while the discontinuities in $h_8'$ are interpreted as generated by D7-branes. Both types of branes are smeared in the $\psi$ direction. The space is terminated in the $\rho$ direction by imposing that $\widehat{h}_4= h_8=0$ at $\rho=0, 2\pi (P+1)$.
When $u=$ constant the closure of the space by setting $\widehat{h}_4=h_8=0$ generates D3 and D7 sources, in the boundary of the space, as explained around eq.(\ref{O3O7brane-singularity}).
Instead, in the general discussion for the annulus in \cite{Chiodaroli:2009xh} the D3-branes occur as poles of a basic
harmonic function at its two boundaries. The basic harmonic function must however be regular in the interior. Therefore, in order to fit in the discussion for the annulus we need continuous $\widehat{h}_4'$ and $h_8'$ functions. This is imposed taking
\begin{equation}
\label{increasing2}
\beta_k\equiv \beta, \qquad \nu_k\equiv \nu, \qquad \text{for} \qquad k=0,1,\dots, P,
\end{equation}
in (\ref{profileh4sp}), (\ref{profileh8sp}), which implies
\begin{equation}
\label{increasing}
\alpha_k=k\, \beta, \qquad \mu_k=k\, \nu, \qquad \text{for} \qquad k=0,\dots, P.
\end{equation}
The solutions are then defined by the functions
\begin{equation}
\widehat{h}_4=\frac{\beta}{2\pi}\, \rho, \qquad h_8=\frac{\nu}{2\pi}\,\rho
\end{equation}
at all $\rho$-intervals. Yet, the closure of the space at $\rho=2\pi (P+1)$ requires that $(P+1)\beta$ D3-branes and $(P+1)\nu$ D7-branes are present at the end of the space. Instead of closing the space by introducing sources as we did with the choice of $\widehat{h}_4$ and $h_8$ functions given by (\ref{profileh4sp}) and (\ref{profileh8sp}), these branes will be automatically present at the end of the space in the annulus construction.
Let us now see how these solutions arise from the general formalism in \cite{Chiodaroli:2009xh}. We take the annulus in \eqref{annulus} as defined from,
\begin{eqnarray}\label{w}
\begin{split}
&w=\frac{z}{2\pi}=\widetilde{\psi}+i\widetilde{\rho}, \quad &\text{with}\qquad\widetilde{\psi}=\frac{\psi}{2\pi}, \quad \widetilde{\rho}=\frac{\rho}{2\pi}.
\end{split}
\end{eqnarray}
Then $\widetilde\psi\in [0,1]$ and the parameter $t$ in the definition of the annulus is $t=2(P+1)$.
As recalled in section \ref{B-type}, our class of solutions is valid when $P$ is large. This allows us to approximate the Jacobi theta function introduced in (\ref{theta1}) by its
asymptotic expansion when $t\to\infty$,
\begin{eqnarray}\label{theta-ex1}
\theta_{1}(\pi w|\tau)|_{t\to \infty}\approx 2e^{-\frac{\pi}{4}t}\sin{\pi w}\approx i e^{-\frac{\pi}{4}t}e^{-i\pi w}.
\end{eqnarray}
This approximation will be key in showing the matching with our solutions. Indeed,
in this approximation it is easy to see that the basic harmonic function defined by \eqref{basic} reads,
\begin{eqnarray}\label{h_0-exp}
h_0(w,\overline{w})\approx 2\pi+\frac{i\pi}{P+1}(w-\overline{w}).
\end{eqnarray}
This gives, at the two boundaries $\partial\Sigma_1$ and $\partial\Sigma_2$,
\begin{eqnarray}\begin{split}\label{h_0boundaries}
& h_0(w-x_{\ell_i},\overline{w}-x_{\ell_i})\approx 2\pi+\frac{i\pi}{P+1}(w-\overline{w}),\\
& h_0(w'+y_{j_i},\overline{w}'+y_{j_i})\approx-\frac{i\pi}{P+1}(w-\overline{w}),
\end{split}
\end{eqnarray}
respectively, where for the second boundary we have used the relation \eqref{wp}. These expressions are thus independent of the positions of the poles at both boundaries. This is in agreement with the fact that our D3/D7 branes are smeared in the $\psi$-direction. We then
get for the harmonic functions in eq.\eqref{kann},
\begin{equation}
\begin{split}
A+\overline{A}=&2\pi\sum_{\ell_A=1} ^{M_A}r_{\ell_A}-\frac{i\pi}{P+1}(w-\overline{w})\left(\sum_{j_A=1}^{M'_A}r'_{j_A}-\sum_{\ell_A=1} ^{M_A}r_{\ell_A}\right)
\\
B+\overline{B}=&2\pi\sum_{\ell_B=1} ^{M_B}r_{\ell_B}-\frac{i\pi}{P+1}(w-\overline{w})\left(\sum_{j_B=1} ^{M'_B}r'_{j_B}-\sum_{\ell_B=1} ^{M_B}r_{\ell_B}\right)
\\
H=&2\pi\sum_{\ell_H=1} ^{M_H}r_{\ell_H}-\frac{i\pi}{P+1}(w-\overline{w})\left(\sum_{j_H=1} ^{M'_H}r'_{j_H}-\sum_{\ell_H=1} ^{M_H}r_{\ell_H}\right)
\\
K=&2\pi\sum_{\ell_K=1} ^{M_K}r_{\ell_K}-\frac{i\pi}{P+1}(w-\overline{w})\left(\sum_{j_K=1} ^{M'_K}r'_{j_K}-\sum_{\ell_K=1} ^{M_K}r_{\ell_K}\right)\label{kann2}.
\end{split}
\end{equation}
In order to match these expressions with the expressions for $A+\overline{A}$ and $K$ given in eq.(\ref{A-BT}) we take into account that $w=z/(2\pi)$, and we obtain
\begin{equation}\label{Aannulus}
\sum_{\ell_A=1} ^{M_A}r_{\ell_A}=0 \qquad \text{and} \qquad \sum_{j_A=1}^{M'_A}r'_{j_A}=\frac{(P+1)\nu}{4\pi},
\end{equation}
for the matching of $A+\overline{A}$, and
\begin{equation}\label{kannulus}
\sum_{\ell_K=1} ^{M_K}r_{\ell_K}=0 \qquad \text{and} \qquad \sum_{j_K=1}^{M'_K}r'_{j_K}=\frac{(P+1)\beta}{\pi},
\end{equation}
for the matching of $K$. Rescaling the residues as\footnote{Note that a rescaling is also necessary in order to interpret the residues of the solutions in \cite{DHoker:2017mds} as charges of $(p,q)$ 5-branes.}
\begin{equation}
r'_{j_A}\to 2r'_{j_A},\qquad r'_{j_K}\to \frac{r'_{j_K}}{4}.
\end{equation}
and replacing the sums by
\begin{equation}
\sum_{j_A=1}^{M'_A}r'_{j_A}\rightarrow \frac{1}{2\pi}\int_0^{2\pi} dr'_{j_A},
\end{equation}
as implied by the smearing of the branes in the $\psi$-direction, we can finally interpret the residues as the charge-densities of D7 and D3 brane sources at both boundaries of the annulus. We would like to stress that even if the general formalism in \cite{Chiodaroli:2009xh} does not account for D7-branes at the boundaries of the annulus, we have associated these to the (smeared) poles of the basic harmonic function for $A+\overline{A}$. The analysis goes in complete parallelism to the analysis of the residues and poles of the $K$ function, associated to the D3-brane sources at both boundaries of the annulus. It is unclear to us the precise reason why this seems to work in the presence of D7-branes.
Finally, from the matching of the $B+\overline{B}$ and $H$ functions we find
\begin{equation}
\sum_{\ell_B=1} ^{M_B}r_{\ell_B}=\sum_{j_B=1} ^{M'_B}r'_{j_B}=0 \qquad \text{and} \qquad \sum_{\ell_H=1} ^{M_H}r_{\ell_H}=\sum_{j_H=1} ^{M'_H}r'_{j_H}=\frac{u_0}{8\pi}.
\end{equation}
These expressions do not seem to have however a direct interpretation in terms of charges of our solutions.
The previous analysis holds true as well for the LNRS backgrounds discussed in \cite{Lozano:2020txg}. The matching of the $A+\overline{A}$, $H$ and $K$ functions is valid for both solutions, while the harmonic function $B+\overline{B}$ vanishes. Again, there are smeared D3 and D7-branes at the end of the space with the same relations between residues and charges.
\section{A new class of AdS$_2\times \text{S}^2\times \text{CY}_2\times \Sigma$ solutions with $\Sigma$ an infinite strip}\label{NATD}
\paragraph{}
In this section we construct a new class of AdS$_2$ solutions to Type IIB supergravity with 8 supercharges by acting with non-Abelian T-duality (NATD) on the AdS$_3\times$S$^2\times$CY$_2\times$I$_\rho$ solutions obtained in \cite{Lozano:2019emq}. The non-Abelian T-duality transformation is performed with respect to a freely acting SL$(2,\mathbb{R})$ isometry group of the AdS$_3$ subspace. This transformation gives rise to a new class of solutions in which the AdS$_3$ subspace is replaced by AdS$_2$ times an interval. These solutions fit in the classification of \cite{Chiodaroli:2009yw} for a Riemann surface with a single boundary, equivalent to an infinite strip.
\subsection{NATD of the AdS$_3\times$S$^2\times$CY$_2$ solutions}
\paragraph{}The study of NATD as a solution generating technique of supergravity was initiated in \cite{Sfetsos:2010uq}. Further works include \cite{Lozano:2011kb, Itsios:2012dc, Itsios:2012zv, Itsios:2013wd}. In all these examples the dualisation took place with respect to a freely acting SU(2) subgroup of the entire symmetry group of the solutions. Instead, in this section we perform the non-Abelian T-duality transformation with respect to one of the freely acting SL$(2,\mathbb{R})$ isometry groups of the AdS$_3$ subspace.
In order to perform the dualisation with respect to the SL$(2,\mathbb{R})$ isometry group we follow the derivation in \cite{Alvarez:1993qi}. We take the $sl(2,\mathbb{R})$ generators analytically continuing the $su(2)$ generators, as $t^a=\tau_a/\sqrt{2}$, with
\begin{eqnarray}
\tau_1=\left(\begin{matrix}
&0 & i\\
&i& 0
\end{matrix}\;\;\right),\qquad \tau_2=\left(\begin{matrix}
&0 & -i\\
&i& 0
\end{matrix}\;\;\right), \qquad \tau_3=\left(\begin{matrix}
&i & 0\\
&0& -i
\end{matrix}\;\;\right).
\end{eqnarray}
These generators satisfy,
\begin{equation}
{\rm Tr}(t^at^b)=(-1)^a\delta^{ab},\quad\quad\quad [t^1,t^2]=i\sqrt{2}t^3,\quad [t^2,t^3]=i\sqrt{2}t^1,\quad [t^3,t^1]=-i\sqrt{2}t^2.
\end{equation}
A group element in the Euler parametrisation is given by,
\begin{equation}
g=e^{\frac{i}{2}\phi\tau_3}e^{\frac{i}{2}\theta\tau_2}e^{\frac{i}{2}\psi\tau_3},
\qquad\text{where} \qquad 0\leq\theta\leq\pi,\;\; 0\leq\psi< \infty,\;\; 0\leq\phi< \infty,
\end{equation}
from which we write the left invariant one forms, $L^a=-i{\rm Tr}(t^ag^{-1}\text{d}g)$, in the following fashion,
\begin{eqnarray}
\begin{split}
\label{Maurer}
L^1=&\sinh{\psi}\text{d}\theta-\cosh{\psi}\sin{\theta}\text{d}\phi\\
L^2=&\cosh{\psi}\text{d}\theta-\sinh{\psi}\sin{\theta}\text{d}\phi\\
L^3=&-\cos{\theta}\text{d}\phi-\text{d}\psi.
\end{split}
\end{eqnarray}
The backgrounds in \cite{Lozano:2019emq} support an SL$(2,\mathbb{R})$ isometry such that the metric, the Kalb-Ramond field and the dilaton can be written as\footnote{We have taken $g^{\mu\nu}=-{\rm Tr}(t^\mu t^\nu)$ to have signature $(+,-,+)$.},
\begin{equation}
\begin{split}
ds^2&=\frac{1}{4}g_{\mu\nu}(x)L^\mu L^\nu+G_{ij}(x)dx^idx^j,\qquad
B_2
B_{ij}(x)dx^i\wedge dx^j, \qquad \phi=\phi(x),
\end{split}
\end{equation}
where $x^i$ are the coordinates in the internal manifold, for $i,j=1,2,...,7$, and $L^{\mu}$ are the forms given by \eqref{Maurer}. All the coordinate dependence on the SL$(2,\mathbb{R})$ group is contained in these forms. The subsequent details on how to technically compute the NATD transformation have been developed extensively in the literature \cite{Sfetsos:2010uq, Itsios:2013wd} (see these reference for more details).
The geometries obtained through NATD with respect to a freely acting SL$(2,\mathbb{R})$ group on the AdS$_3$ of the solutions in \cite{Lozano:2019emq} are given by,
\begin{equation}
\begin{split}
\text{d}{s}_{st}^2&=\frac{u\sqrt{\widehat{h}_4 h_8}}{4 r^2\widehat{h}_4 h_8-u^2} r^2\text{d}s_{\text{AdS}_2}^2+\sqrt{\frac{\widehat{h}_4}{h_8}}\text{d}s^2_{\text{CY}_2}+\frac{u\sqrt{\widehat{h}_4 h_8} }{4 \widehat{h}_4 h_8+u'^2}\text{d}s^2_{\text{S}^2}+\frac{\sqrt{\widehat{h}_4 h_8} }{u}(\text{d}\rho ^2+\text{dr}^2),\\
e^{-2\phi}&=\frac{\left(4 r^2\widehat{h}_4 h_8-u^2\right) \left(4 \widehat{h}_4 h_8+u'^2\right)}{2^8\widehat{h}_4^2},\\
B_2&=-\frac{2 r^3\widehat{h}_4 h_8}{4 r^2\widehat{h}_4 h_8-u^2}\text{vol}_{\text{AdS}_2}-\frac{4 \rho \widehat{h}_4 h_8-u'(u-\rho u')}{2 \left(4 \widehat{h}_4 h_8+u'^2\right)} \text{vol}_{\text{S}^2}\label{B2NATD}.
\end{split}
\end{equation}
Additionally, the background is supported by the RR fluxes,
\begin{equation}
\begin{split}
F_1=&-\frac{r h'_8}{4}\text{d}r+\frac{1}{2^4}\left(4h_8+\partial_\rho \left[\frac{u u'}{\widehat{h}_4}\right]\right)\text{d}\rho,\\
F_3=&\frac{h_8}{8 \left(4 \widehat{h}_4 h_8+u'^2\right)}\left(\frac{\widehat{h}_4'u^2}{\widehat{h}_4}\text{d}\rho+rh_8\left(4\widehat{h}_4+\partial_\rho \left[\frac{u u'}{h_8}\right]\right)\text{d}r\right)\wedge \text{vol}_{\text{S}^2}\\
&+\frac{r^2h_8}{8 \left(4 r^2\widehat{h}_4 h_8-u^2\right)}\left(\frac{h_8'u^2}{h_8}\text{d}r-r\widehat{h}_4\left(4h_8+\partial_\rho \left[\frac{u u'}{\widehat{h}_4}\right]\right)\text{d}\rho\right)\wedge \text{vol}_{\text{AdS}_2},\label{F3NATD}\\
F_5=&\frac{1}{2^4}\left(4r\widehat{h}_4'\text{d}r-\left(4\widehat{h}_4+\partial_\rho \left[\frac{u u'}{h_8}\right]\right)\text{d}\rho\right)\wedge \text{vol}_{\text{CY}_2}\\
&-\frac{r^2u^2h_8^2}{2^4\left(4 r^2\widehat{h}_4 h_8-u^2\right) \left(4 \widehat{h}_4 h_8+u'^2\right)}\left(4 r\widehat{h}_4'\text{d$\rho $}+\left(4\widehat{h}_4+\partial_\rho \left[\frac{u u'}{h_8}\right]\right)\text{d$r$}\right)\wedge\text{vol}_{\text{AdS}_2}\wedge \text{vol}_{\text{S}^2},\\
F_7=&-\frac{\widehat{h}_4}{8 \left(4\widehat{h}_4 h_8+u'^2\right)}\left(\frac{h_8'u^2}{h_8}\text{d}\rho+r\widehat{h}_4\left(4h_8+\partial_\rho \left[\frac{u u'}{\widehat{h}_4}\right]\right)\text{d}r\right)\wedge \text{vol}_{\text{S}^2}\wedge \text{vol}_{\text{CY}_2}\\
&-\frac{r^2\widehat{h}_4}{8 \left(4 r^2\widehat{h}_4 h_8-u^2\right)}\left(\frac{\widehat{h}_4'u^2}{\widehat{h}_4}\text{d}r-rh_8\left(4\widehat{h}_4+\partial_\rho \left[\frac{u u'}{h_8}\right]\right)\text{d}\rho\right)\wedge \text{vol}_{\text{AdS}_2}\wedge \text{vol}_{\text{CY}_2},\\
F_9=&\frac{r^2u^2\widehat{h}_4^2}{4\left(4 r^2\widehat{h}_4 h_8-u^2\right)\left(4 \widehat{h}_4 h_8+u'^2\right)}\left(rh_8'\text{d}\rho+\frac{1}{4}\left(4h_8+ \partial_\rho \left[\frac{u u'}{\widehat{h}_4}\right]\right)\text{d}r\right)\wedge \text{vol}_{\text{AdS}_2}\wedge \text{vol}_{\text{CY}_2}\wedge \text{vol}_{\text{S}^2}.
\end{split}
\end{equation}
The previous background is a solution to the Type IIB supergravity EOM whenever $4 r^2\widehat{h}_4 h_8-u^2>0$. Namely we get a well-defined geometry for
\begin{equation}
r>r_0=\frac{u}{2\sqrt{\widehat{h}_4 h_8}}.\label{r0natd}
\end{equation}
In the next section we show that a subset of the solutions defined by \eqref{B2NATD} and \eqref{F3NATD} fit in the general classification of AdS$_2\times$S$^2\times$CY$_2\times\Sigma$ geometries given in \cite{Chiodaroli:2009yw} with $\Sigma$ an infinite strip.
\subsection{The NATD solution as a CGK geometry}
\paragraph{} In this section we discuss how the solutions given by \eqref{B2NATD}-\eqref{F3NATD} fit in the class of CGK. Going to Einstein frame we get the warp factors of the metric, dilaton and axion,
\begin{equation}
\begin{split}
&f_1^2=\frac{ur^2\sqrt{h_8}(4 \widehat{h}_4 h_8+u'^2)^{1/4}}{4(4 \widehat{h}_4 h_8r^2-u^2)^{3/4}},\quad
f_2^2=\frac{u\sqrt{h_8}(4 \widehat{h}_4 h_8r^2-u^2)^{1/4}}{4(4 \widehat{h}_4 h_8+u'^2)^{3/4}},\\
&f_3^2=\frac{(4 \widehat{h}_4 h_8r^2-u^2)^{1/4}(4 \widehat{h}_4 h_8+u'^2)^{1/4}}{4\sqrt{h_8}},\quad
e^{2\Phi}=e^{-\phi}=\frac{\sqrt{(4 \widehat{h}_4 h_8r^2-u^2)(4 \widehat{h}_4 h_8+u'^2)}}{2^4\widehat{h}_4},\\
&\chi =\frac{1}{2^4}\left(2\nu(\rho^2-r^2)+4\mu\rho+\frac{u u'}{\widehat{h}_4}\right),\quad \widetilde{\rho}^2=\frac{\sqrt{h_8}(4 \widehat{h}_4 h_8r^2-u^2)^{1/4}(4 \widehat{h}_4 h_8+u'^2)^{1/4}}{2^2u
.
\end{split}
\end{equation}
In $\chi$, the axion field, we have taken $h_8=\mu+\nu\rho$, with $\mu$, $\nu$ constants.
This choice corresponds to backgrounds without D7-branes,
as those constructed in \cite{Chiodaroli:2009yw}.
The 2d Riemann surface associated to the solutions is the strip depicted in Figure \ref{strip2}, parametrised as,
\begin{eqnarray}
\label{StripD}
z=\rho+i\;r\qquad
\text{where}\qquad \rho\in[0,2\pi(P+1)]\qquad\text{and}\qquad r\in[r_0,\infty],
\end{eqnarray}
where the value of $r_0$ is determined in eq.(\ref{r0natd}).
\begin{figure}[t]
\centering
\includegraphics[scale=0.6]{Figures/riemann2.pdf}
\caption{Infinite strip associated to the NATD solution.}
\label{strip2}
\end{figure}
Taking the 'plus-solution' defined by equation \eqref{kp-gen} we obtain the $A$, $B$, $H$ and $K$ functions in terms of the defining functions of our backgrounds, $\widehat{h}_4$, $h_8$ and $u$,
\begin{equation}
\begin{split}
A=\frac{1}{2^4}\left(4\mu(r-i\rho)+2i\nu(r-i\rho)^2+\frac{u'}{\widehat{h}_4}(ru'-iu)\right), \quad
B=\frac{1}{2^5}&\frac{\sqrt{4 \widehat{h}_4 h_8+u'^2}\sqrt{u^2+r^2u'^2}}{\sqrt{\widehat{h}_4 h_8}},\\
H=\frac{ru}{2^4},\qquad
K=\frac{r(4 \widehat{h}_4 h_8+u'^2)}{2^5h_8}.
\end{split}
\end{equation}
We anticipate these functions are neither harmonic nor holomorphic. In order to ensure harmonicity -in $H$ and $K$- and holomorphicity -in $A$ and $B$- we need to choose $u'=0$. In that case we obtain,
\begin{equation}
\begin{split}
&A=\frac{4\mu(r-i\rho)+2i\nu(r-i\rho)^2}{2^4}=-iz\frac{2\mu+z\nu}{8}, \qquad
B=\frac{u}{2^4}=\frac{u_0}{2^4},\\
&H=\frac{ru}{2^4}=-i\frac{u_0}{2^5}(z-\overline{z}),\qquad
K=\frac{r \widehat{h}_4}{2^3}=-i\frac{(z-\overline{z})}{2^5}(\beta(z+\overline{z})+2\alpha),
\end{split}
\end{equation}
where we have used \eqref{linearfunctions}. The harmonic function conjugated to $K$ is,
\begin{eqnarray}
\widetilde{K}=-\frac{1}{2^5}(\beta(z^2+\overline{z}^2)+2\alpha(z+\overline{z})),\qquad\text{with}\qquad C_{\text{CY}_2}=\frac{1}{2^3}(\beta(r^2-\rho^2)+2\alpha\rho).
\end{eqnarray}
Note that we have taken $\widehat{h}_8=\alpha+\beta\rho$, with $\alpha$, $\beta$ constants, which corresponds to backgrounds without D3-branes, as those constructed in \cite{Chiodaroli:2009yw}.
The functions associated to the complex three-form are,
\begin{equation}
\begin{split}
&h_1=-\frac{i}{4}(z-\overline{z}),\qquad \widetilde{h}_1=-\frac{1}{4}(z+\overline{z}),\\
&h_2=-\frac{1}{2^5}\left(\mu(z^2+\overline{z}^2)+\frac{\nu}{3}(z^3+\overline{z}^3)\right),\quad \widetilde{h}_2=\frac{i}{2^5}\left(\mu(z^2-\overline{z}^2)+\frac{\nu}{3}(z^3-\overline{z}^3)\right).
\end{split}
\end{equation}
Notice that $h_i$ and $\widetilde{h}_i$ are harmonic functions conjugate to each other. The potentials given in \eqref{potential-b1} are,
\begin{equation}
\begin{split}
& b^{(1)}=-i\frac{u_0^2(z-\overline{z})}{(z-\overline{z})^2(\beta(z+\overline{z})+2\alpha)(\nu(z+\overline{z})+2\mu)+4u_0^2}-h_1,\qquad b^{(2)}=-\frac{1}{4}(z+\overline{z}),\\
& c^{(1)}=-i\frac{u_0^2(z-\overline{z})(2\mu(z+\overline{z})+\nu(z^2+\overline{z}^2))}{16((z-\overline{z})^2(\beta(z+\overline{z})+2\alpha)(\nu(z+\overline{z})+2\mu)+4u_0^2)}+\widetilde{h}_2,\\
& c^{(2)}=-\frac{u_0^2}{16(\beta(z+\overline{z})+2\alpha)}+h_2,
\end{split}
\end{equation}
which agree with the expressions \eqref{B2NATD} and \eqref{F3NATD} for $u'=0$.
The previous analysis shows that the new class of solutions constructed through non-Abelian T-duality provide an explicit example of CGK geometries where the Riemann surface is an infinite strip. We will provide a more detailed global study of these solutions in a future publication.
\section{Electric-magnetic charges and a minimisation principle}\label{maxwellcentralhol}
\paragraph{}
In this section we extend two results discussed in \cite{Lozano:2020txg, Lozano:2020sae} to our new infinite family of AdS$_2$ solutions.
\\
The first result is a relation between the holographic central charge in eq.(\ref{hccharge-B}) and an integral of the product of the electric and magnetic fluxes of the Dp-branes present in the background. This relates the holographic central charge in Section \ref{hccharge}, computed purely in terms of the NS-NS sector of the background, with a calculation purely in terms of the Ramond-Ramond sector.
\\
Furthermore, in section \ref{functional}, we explore this relation from a geometrical point of view. We define a quantity in terms of geometric forms in our geometries and through an extremisation principle relate it to the holographic central charge in eq. (\ref{hccharge-B}). In summary, in this section we present a connection between the holographic central charge, the product of the electric and magnetic charges and an extremised functional.
\subsection{A relation between the holographic central charge and the RR fluxes}
\paragraph{} We provide a relation between the holographic central charge found in eq.(\ref{hccharge-B}) and the fluxes of the Ramond-Ramond sector in eq \eqref{RR sector lower rank fluxes}.
Consider a Dp brane and the associated electric $\widehat{F}_{p+2}$ and magnetic $\widehat{F}_{8-p}$ Page field strengths. We define the ``density of electric and magnetic charges'', $\rho^e_{\text{Dp}}$ and $\rho^m_{\text{Dp}}$, as follows,
\begin{eqnarray}
& & \rho_{\text{Dp}}^{e}=\frac{1}{(2\pi)^p} \widehat{F}_{p+2} ,\;\;\;\;\; \rho_{\text{Dp}}^{m}=\frac{1}{(2\pi)^{7-p}} \widehat{F}_{8-p}.\label{mc}
\end{eqnarray}
From these we construct the quantity,
\begin{equation}
\begin{split}
\label{sumdensities}
& \int \sum_{p=1,3,5,7} \rho_{\text{D}p}^e \rho_{\text{D}p}^m=\,\\
& = \frac{1}{\pi}\text{Vol}_{\text{AdS}_2}\left(\frac{\text{Vol}_{\text{CY$_2$}}} {16 \pi^4}\right) \int \text{d} \rho\left[\frac{4\widehat{h}_4h_8-(u')^2}{8} + \frac{1}{16}\partial_\rho\left(u^2\frac{(h_4 h_8)'}{h_4h_8} \right)-\frac{u^2}{16}\left(\frac{\widehat{h}_4''}{\widehat{h}_4}+\frac{h_8''}{h_8}\right) \right]
.
\end{split}
\end{equation}
In the absence of sources $\widehat{h}_4''= h_8''=0$ and, up to a boundary term, this is proportional to the expression for the holographic central charge in equation \eqref{hccharge-B}. We explore below the contribution of the sources to this expression. Notice that eq.(\ref{sumdensities}) links the holographic central charge in eq.\eqref{hccharge-B}---a calculation purely in terms of the NS-NS sector---with one purely in terms of the Ramond-Ramond sector.
\subsection{An action functional for the central charge}\label{functional}
\paragraph{} Following the ideas of \cite{Couzens:2018wnk, Gauntlett:2018dpc} and the lead of the works \cite{Lozano:2020txg, Lozano:2020sae}, we construct a functional in terms of an integral of forms defined in the internal space. Once such functional is extremised the holographic central charge in eq.(\ref{hccharge-B}) is recovered, up to a boundary term.
We define forms $J_i$ and ${\mathcal F}_i$ (for $i=1,3,5,7$) on the internal space $X_8=$[S$^2$, CY$_2$, S$_\psi$, I$_\rho$]. These forms are inherited from the Page fluxes \eqref{fluxes23}\footnote{The same result can be obtained considering the Maxwell fluxes in \eqref{RR sector lower rank fluxes}.}. As explained in \cite{Lozano:2020txg, Lozano:2020sae}, they are the {\it restriction} of the fluxes to the internal space. Writing the Page fluxes in eqs.(\ref{fluxes23}) in terms of forms $J_i$ and ${\mathcal F}_i$ as,
\begin{equation}
\begin{split}
\widehat{F}_1&=J_1,\qquad \widehat{F}_3={\mathcal F}_{1}\wedge\text{vol}_{\text{AdS}_2}+ J_3,\qquad \widehat{F}_5={\mathcal F}_{3}\wedge\text{vol}_{\text{AdS}_2}+J_5,\\
\widehat{F}_7&={\mathcal F}_{5}\wedge\text{vol}_{\text{AdS}_2}+J_7,\qquad \widehat{F}_9={\mathcal F}_{7}\wedge\text{vol}_{\text{AdS}_2}.
\end{split}
\end{equation}
The forms $J_i$ and ${\mathcal F}_i$ are,
\begin{equation}\label{forms}
\begin{split}
J_1 &= h'_8 \, \mathrm{d} \psi \, , \quad\quad
J_{3} = \frac{1}{4} \bigg(2h_8+\frac{u'(u\widehat{h}_4'-\widehat{h}_4u')}{2\widehat{h}_4^2} \bigg) \text{vol}_{\text{S$^2$}}\wedge\text{d}\rho,\quad\quad
J_{5} = - \widehat{h}_4' \text{vol}_{\text{CY$_2$}} \wedge \mathrm{d} \psi \,\\
J_7&=- \frac{1}{4}\left( 2\widehat{h}_4+\frac{u'(uh_8'-h_8u')}{2h_8^2} \right) \text{vol}_{\text{CY$_2$}}\wedge\text{vol}_{\text{S$^2$}}\wedge \text{d}\rho,\;\;\;\;
{\mathcal F}_1= \frac{1}{2} \left(h'_8(\rho-2\pi k) - h_8\right)\text{d}\psi\, ,\\
{\mathcal F}_{3} &=\frac{1}{4}\left((\rho-2\pi k)h_8- \frac{\left(u - (\rho-2\pi k) u' \right) ( u \widehat{h}_4' - \widehat{h}_4 u' )}{4 \widehat{h}_4^2}\right)\text{vol}_{\text{S$^2$}} \wedge \mathrm{d} \rho, \\
{\mathcal F}_{5}&= \frac{1}{2} (\widehat{h}_4 - (\rho-2\pi k) \widehat{h}_4')\; \text{vol}_{\text{CY$_2$}} \wedge \mathrm{d} \psi \, ,\\
{\mathcal F}_{7} &= -\frac{1}{4}\left((\rho-2\pi k) \widehat{h}_4 -\frac{\left(u - (\rho-2\pi k) u' \right) ( u h_8' - h_8 u')}{4h_8^2}\right) \text{vol}_{\text{CY$_2$}}\wedge\text{vol}_{\text{S$^2$}}\wedge \text{d}\rho.
\end{split}
\end{equation}
With the forms in eqs.\eqref{forms}, we construct the functional,
\begin{equation}\label{functionalcB}
\begin{split}
{\mathcal C} &= \int_{X_8} {\mathcal F}_{1}\wedge J_7+{\mathcal F}_{3}\wedge J_5-(J_1\wedge{\mathcal F}_{7}+J_3\wedge{\mathcal F}_{5})
\\%&(J_1+{\mathcal F}_{1})\wedge (J_7+{\mathcal F}_{7})-(J_3+{\mathcal F}_{3})\wedge (J_5+{\mathcal F}_{5}) \\
&=\int_{X_8} \left(\frac{4\widehat{h}_4 h_8-(u')^2}{8} -\frac{u^2}{16} \left( \frac{\widehat{h}_4'^2}{\widehat{h}_4^2} + \frac{h_8'^2}{h_8^2} \right) + \frac{u u'}{8} \left( \frac{\widehat{h}_4'}{\widehat{h}_4} + \frac{h_8'}{h_8} \right) \right) \text{vol}_{\text{CY}_2} \wedge \text{vol}_{\text{S}^2}\wedge \text{d}\psi\wedge \text{d}\rho,
\end{split}
\end{equation}
We minimise the functional ${\mathcal C}$ by imposing the Euler-Lagrange equation for $u(\rho)$,
\begin{eqnarray}
2u''=u\left(\frac{\widehat{h}_4''}{\widehat{h}_4}+\frac{h_8''}{h_8}\right).
\end{eqnarray}
This equation of motion is solved if,
\begin{equation}
\begin{split}
h_8''=0,\qquad \widehat{h}_4''=0,\qquad u''=0,
\end{split}
\end{equation}
the first two are Bianchi identities for the background and the last is a BPS equation.
The functional in eq.\eqref{functionalcB} can be rewritten as,
\begin{equation}
\begin{split}
{\mathcal C} &=\frac{1}{8}\int_{X_8}\left(4\widehat{h}_4 h_8-(u')^2+\partial_\rho\left[\frac{u^2}{2}\left( \frac{\widehat{h}_4'}{\widehat{h}_4}+\frac{h_8'}{h_8} \right)\right]-\frac{u^2}{2}\left( \frac{\widehat{h}_4''}{\widehat{h}_4}+\frac{h_8''}{h_8} \right)
\right) \text{vol}_{\text{CY}_2} \wedge \text{vol}_{\text{S}^2}\wedge \text{d}\psi\wedge \text{d}\rho.
\end{split}
\end{equation}
The last term (that would vanish in the absence of sources), is proportional to the quotient of the number of flavours by the number of colours in each node. Using the condition that the flavours are sparse, as explained below eq.(\ref{nana}), we see that its contribution is subleading in front of the other terms. Furthermore, the boundary term gives a divergent contribution. Indeed, for the case $u=u_0$ and $\widehat{h}_4, h_8 $ in eqs.\eqref{profileh4sp}-\eqref{profileh8sp} the boundary term reads,
\begin{equation}
\label{diver}
\begin{split}
\int_0^{2\pi (P+1)} &\partial_\rho\left[\frac{u^2}{2}\left( \frac{\widehat{h}_4'}{\widehat{h}_4}+\frac{h_8'}{h_8} \right)\right]\text{vol}_{\text{CY}_2} \wedge \text{vol}_{\text{S}^2}\wedge \text{d}\psi\wedge \text{d}\rho=-\lim_{\epsilon\to 0}\frac{2\pi u_0^2}{\epsilon}(\alpha_P+\mu_P+\beta_0+\nu_0)\text{Vol}_{\text{CY}_2}\\
&=-\lim_{\epsilon\to 0}\frac{2\pi u_0^2}{\epsilon}(Q_{\text{D}3}^{total}+Q_{\text{D}7}^{total})\text{Vol}_{\text{CY}_2},
\end{split}
\end{equation}
where we regularised $\widehat{h}_4(0)=h_8(0)=\widehat{h}_4(2\pi(P+1))=h_8(2\pi(P+1))=\epsilon$. The divergence in eq.\eqref{diver} is associated with the presence of sources in the background as was found in \cite{Lozano:2020txg,Lozano:2020sae}.
In summary, the functional in eq.(\ref{functionalcB}) is proportional to the holographic central charge of eq.(\ref{hccharge-B}), plus a subleading contribution and a boundary term. For our infinite family of backgrounds, we have linked a calculation purely in terms of the NS-NS sector---eq.(\ref{hccharge-B}), with a calculation purely in terms of the Ramond-Ramond sector---eq.(\ref{sumdensities}), with the extremisation of a functional constructed as a restriction of the Ramond-Ramond forms to the internal space---eq.(\ref{functionalcB}). We believe that this may be a generic feature, worth exploring in backgrounds dual to various SCFTs in different dimensions.
\section{Conclusions}\label{conclusions}
\paragraph{}We close this paper by presenting a short summary of the contents of this work and proposing future lines of investigation.
This work presents two new infinite families of backgrounds with an AdS$_2$ factor. The presentation focuses mostly on geometrical aspects of the new solutions. The new family of backgrounds in Section \ref{B-type} can be obtained by analytically continuing the backgrounds of \cite{Lozano:2020txg} or via T-duality, on the Hopf-fibre of the S$^3$, from the solutions in \cite{Lozano:2020sae}. These connections are summarised in Figure \ref{diagrama}. A precise brane set-up was proposed for these backgrounds and the holographic central charge was calculated. We used the brane set-up to argue for a precise quiver. The IR dynamics of such quivers should be the SCQMs dual to our backgrounds.
The family of AdS$_2$ backgrounds in Section \ref{B-type} and that in the paper \cite{Lozano:2020txg} have been shown to be connected to the solutions of \cite{Chiodaroli:2009yw, Chiodaroli:2009xh}. In fact, under certain circumstances they extend this class of solutions. The connection between these qualitatively different backgrounds requires of a subtle zoom-in procedure that we explained in detail in Section \ref{CGKsection}.
A second family of new backgrounds is presented in Section \ref{NATD}. These interesting solutions depend explicitly on two coordinates (labelled as $\rho$ and $r$ in Section \ref{NATD}) and were obtained by the application of non-Abelian T-duality on the AdS$_3$ factor of the backgrounds in \cite{Lozano:2019emq}. We leave for future work to discuss the associated brane set-up, though it seems clear that the ideas described in \cite{Lozano:2016kum, Lozano:2016wrs, Lozano:2017ole, Itsios:2017cew, Lozano:2018pcp} will play an essential role in the global-definition of these solutions. By the same token, it would be interesting to study the integrability (or not) of the backgrounds presented here, as well as those in \cite{Lozano:2020txg,Lozano:2020sae}. Integrable string backgrounds dual to field theories described by linear quivers in dimensions $d=2,4,6$, have been found in \cite{Filippas:2019ihy, Rigatos:2020igd, Nunez:2018qcj,Filippas:2019puw}. Similar techniques should probably apply for the $d=1$ case.
Finally, in Section \ref{maxwellcentralhol} the holographic central charge defined in Section \ref{B-type}---a quantity computed solely in terms of the NS-NS sector of the backgrounds, has been connected with a calculation purely in terms of the Ramond-Ramond sector of our solutions. A functional whose extremisation yields the holographic central charge was also discussed. It should be interesting to find out if a similar structure occurs generically for other AdS$_{d+1}$ backgrounds.
\section*{Acknowledgements} We would like to thank Niall Macpherson and Salomon Zacarias for very useful discussions.
\\
The work of CN is supported by STFC grant ST/T000813/1. Y.L. and A.R. are partially supported by the Spanish government grant
PGC2018-096894-B-100. AR is supported by CONACyT-Mexico.
|
\section{Introduction}
Classical force fields are essential tools for the atomistic simulation of biomolecular systems. They can be used to probe the structure and biochemical interactions of peptides, small organic molecules, proteins and their ligands, nucleic acids, and larger structures such as lipid membranes and ion channels. The most widely-used empirical force fields are those based on the CHARMM \cite{mackerell1998,brooks2009,vanommeslaeghe2010}, AMBER\cite{cornell1995,wang2004,maier2015} and OPLS-AA\cite{jorgensen1996} families. Solvation effects---critical to an accurate description of protein structure, dynamics, and function\cite{bellissent2016}---are described via implicit or explicit (atomic) water models,\cite{guillot2002,vega2011,onufriev2018} the latter often employing fixed charges, as in the widely-used TIP \cite{jorgensen1983} and SPC\cite{berendsen1987} models. The additive energy function describing the interactions among atoms includes harmonic bond and angle terms, bonded torsional energies, and electrostatic and non-bonded (van der Waals) interactions, with fixed, atom-centered partial charges derived from electrostatic models or quantum mechanical calculations.\cite{jorgensen2005} This general class of empirical force fields is also used as a jumping-off point for developing polarizable force fields\cite{demerdash2018,jing2019} and other variants,\cite{nerenberg2018} and are regularly revisited, refined, reparameterized, and extended to encompass new classes of biomolecules and their interactions.\cite{lindorff2010,vanommeslaeghe2012a,nerenberg2018,he2020} Efforts are underway to extend force fields and associated water models\cite{shabane2019} to describe increasingly complex biochemical and biophysical processes, such as the fluctuations of intrinsically-disordered proteins among multiple structural minima,\cite{wang2017,huang2017,robustelli2018} and the impact of ionic interactions on nucleic acid catalysis,\cite{yang2006} RNA folding, \cite{Nguyen2020} and the mechanochemistry and catalytic cycles of molecular motor proteins\cite{kiani2016} such as kinesin \cite{Parke2010,hariharan2009,hwang2008,cross2016} and myosin.\cite{lawson2004,ovchinnikov2010,li2013}
While considerable biophysical and structural insights can be gained using all-atom techniques based on empirical classical force fields, in many instances a quantum-level description of atomic interactions is required. The atomistic ReaxFF force field\cite{chenoweth2008,senftle2016} reframes the energy function in terms of dynamically-computed bond order and charge polarization contributions. However, due to its empirical design, and the intrinsically complex nature of reactive chemistry, parameter transferability can be limited even between nominally similar problem subclasses.\cite{senftle2016} Fragmentation methods,\cite{gordon2012,pruitt2014} which partition a system into interacting atom-groupings, each of which is treated quantum-mechanically, are a promising approach for modeling biochemical systems at scale, but require coarse-grained partitioning into functional groups. The pioneering QM/MM\cite{warshel1976,field1990} method can be used to focus attention on a smaller, predefined, region using quantum mechanics, while treating the rest of the system classically. This technique requires careful consideration of boundary and embedding effects, including interactions between the QM and MM regions, but is a valuable tool for studying free energies, catalysis\cite{van2013} including ATP hydrolysis,\cite{mcgrath2013,kiani2014,kiani2015,baldo2020} and photochemistry.\cite{boulanger2018}. {\it Ab initio} molecular dynamics\cite{car1985,carloni2002} integrates the quantum mechanical evaluation of the forces on each atom---typically done using DFT---with the classical dynamical evolution of the system itself. However, since quantum calculations are performed as part of the simulation, this method is limited in the size of the systems (hundreds of atoms) and timescale of simulations (${\cal O}(100$~ps)) that can be studied.\cite{zhang2018}
An alternative, and increasingly popular, approach is to utilize machine-learned representations of the force field. This is done by constructing neural network classifiers\cite{behler2007,behler2011} for predicting individual atomic contributions to the overall energy function: the parameters of the neural network are determined by training the predictions of the force field energy against a pre-computed, quantum mechanical database. At a given timestep in a simulation, the input to each atom's classifier is a set of symmetry-respecting \textit{features} (also known as {\it descriptors}) characterizing the local conformational and chemical environment. Examples include the ANI-1 potential for small organic molecules,\cite{smith2017} and the DeepMD potential, in which symmetry functions with a pre-defined form are replaced by a local coordinate representation.\cite{han2018,zhang2018}
In all of these methods, there is an inevitable tradeoff between system size, diversity of atomic composition, and level of quantum mechanical description that can be modeled. This limits current ability to describe nonlocal, quantum-mediated effects in large systems---for instance, allosteric protein interactions and charge-transfer-driven structural changes,\cite{Wei2016} or biochemical processes such as motor protein procession along a microtubule in a cell,\cite{hancock2016} which are driven by the coupling of chemical and mechanical transitions occurring over long (millisecond to second) time scales. For such problems, chemically-accurate atomistic potentials are also needed in order to identify the relevant substructures---switches, salt bridges, disordered regions---and Markov states\cite{noe2017} for constructing atomistically-informed multiscale kinetic models incorporating an accurate description of the underlying chemistry.\cite{jacobson2014SA,jacobson2014load,jacobson2017}
Here, we present a physics-based approach for holistically incorporating quantum effects within an atomistic force field. This is accomplished by expressing the potential functional in terms of dynamically-evolving atom-in-molecule electron densities,\cite{Atlas2020} which are further represented as statistical ensembles of isolated neutral atom, excited state, and ionic state densities. These {\it basis densities} serve as fundamental constructs of the model.\cite{Amokwao2020} Since the basis densities are those of isolated atoms, they can be precomputed at an arbitrarily high level of theory, including spin polarization, electron correlation, and even relativistic effects in the case of heavier atoms. The framework we present here---a formulation of the charge-transfer embedded-atom method (CT-EAM)\cite{valone2006Kos,muralidharan2007ED} to facilitate large-scale simulations of biomolecular systems---has its roots in the empirical embedded-atom method (EAM) for metallic \cite{daw1983,daw1984,foiles1986} and covalent\cite{baskes1987,baskes1989} systems, and the density functional theory (DFT) of quantum systems as elucidated by Hohenberg and Kohn\cite{hohenberg1964} and Kohn and Sham,\cite{kohn1965} with subsequent extensions to ensembles of atomic charge states\cite{perdew1982} and ensembles of atomic excited states.\cite{kohn1986,gross1988var,gross1988DFT,oliveira1988} We derive a general force field expression---the ensemble DFT charge-transfer EAM (CT-EAM)---for coupling quantum mechanical densities to atomistic energies, including both local and global correlations. The quantum mechanical coupling is effected through the medium of the electron density, a physical observable, in order to flexibly and consistently describe reactive chemistry, without local bias, across large biomolecular systems.
\section{Theoretical Background and Methods}
This section provides the theoretical and conceptual background for establishing systematic linkages between the electronic and atomistic length scales through DFT and the DFT-based EAM model. EAM was originally developed for simulating elemental materials systems,\cite{daw1983,daw1984,foiles1986,baskes1987,baskes1989} and is extended here to incorporate charge polarization and charge transfer effects in molecular systems. We begin by reviewing the related valence bond, fluctuating Hamiltonian (FH) model\cite{valone2011IM,valone2011WF,valone2014} to highlight the essential role played by an ensemble atom-in-molecule picture in both wavefunction- and density-based formulations of a general charge-transfer EAM model.\cite{valone2006Kos} Energy-density duality is presented as a unifying concept for understanding the tradeoffs involved in constructing a DFT-based force field. We then describe the specific energy and density foundations of the ensemble DFT CT-EAM potential: the EAM method itself, which provides the DFT-based scaffold incorporating charge polarization and charge transfer effects within ensemble DFT CT-EAM; and a density functional theory of the atom-in-molecule, couched in terms of statistical ensembles of atomic charge and excited states.
\subsection{Valence bond origins: spectral representation and the atom-in-molecule}
Potential models based on a valence bond wavefunction picture\cite{aaqvist1993,schmitt1998,morales2001,morales2004,valone2004,valone2008} have long provided important insights into the interplay between energy, charge, and chemical potential for accurately describing bond formation and breaking in charge-transfer systems, even at the level of diatomic models.\cite{valone2006Kos,valone2011IM}
Using a simple two-state valence bond model, Valone and Atlas\cite{valone2006ED} showed that by applying two key concepts---(i) transformation from a valence-bond picture (with relative ionic and covalent contributions characterized by an ionicity $\gamma$) to a spectral representation; and (ii) the introduction of an atom-in-molecule decomposition characterized by a charge transfer $q$---it was possible to extend the the ensemble model of charge transfer of Perdew {\it et al.}\cite{perdew1982} from weakly-interacting to strongly-interacting subsystems.
Starting from these ideas, Valone proposed and developed the fluctuating Hamiltonian (FH) model\cite{valone2011IM,valone2011WF,valone2014,valone2016} which uses an ensemble wavefunction picture and an explicit Hamiltonian formulation to describe fluctuating charge states of an atom interacting with its environment. The model yields generalizations of familiar chemical concepts---Mulliken electronegativity, Parr-Pearson hardness, ionization potential, electron affinity, and charge-transfer gap energy---expressed as combinations of Hamiltonian matrix elements and density matrix elements, together with an expression for the energy of an atomic fragment, incorporating charge-transfer dependence through the ionicity. Furthermore, by mapping the charge-state-dependent fluctuating Hamiltonian picture onto Moffitt's Hamiltonian representation for the atom-in-molecule,\cite{moffitt1951} a charge-transfer EAM embedding energy\cite{valone2006Kos} can be defined.\cite{valone2014,valone2016} The remaining two-body terms in the atom-in-molecule energy expression correspond to electrostatic interactions, as in the original EAM ({\it cf.}~Eq.~(\ref{eq:EAM}) below).\cite{daw1983,daw1984} The FH model thus integrates familiar chemical concepts within an atom-in-molecule EAM framework, using wavefunction-based reasoning.
\subsection{Energy-density duality}
That an intimate relationship exists between the total energy of a many-electron system and its spatial electron density distribution $\rho({\bf r})$ was postulated almost a century ago, as what is now known as Thomas-Fermi (TF) theory.\cite{lieb1977} The TF picture of the electron cloud in an atom is that of an inhomogeneous electron gas, whose spatial distribution can be described by a mean field theory. The kinetic energy is approximated using a local density approximation\cite{hohenberg1964}, i.e., the local application of the expression for the kinetic energy of a homogeneous electron gas, and the total energy is given by:\cite{lundqvist1983}
\begin{equation}
E_{\rm TF}[\rho({\bf r})] = T_{\rm TF}[\rho({\bf r})] + U_{\rm H}[\rho({\bf r})] + \int \rho ({\bf r})\, v({\bf r})\, d{\bf r},
\label{eq:TF}
\end{equation}
where $T_{\rm TF} [\rho({\bf r})] = C_{\rm K} \int [\rho({\bf r})]^{5/3} d{\bf r}$ ($C_K$ is a constant),
$U_{\rm H}[\rho({\bf r})] = 1/2 \int \int d{\bf r}\, d{\bf r}' \rho({\bf r})\, \rho({\bf r}')/|{\bf r} - {\bf r}'|$ is the classical Hartree electrostatic energy, and $v({\bf r})$ is the external potential due to the positive nuclear charge of the atom.
Some forty years later, this appealing picture found formal expression in the celebrated theorem of Hohenberg and Kohn (HK),\cite{hohenberg1964} which established the foundation for modern density functional theory (DFT). The proof of the Kohn-Sham (KS) theorem and coupled single-particle equations followed one year later.\cite{kohn1965} The KS equations provide a Hartree-Fock-like framework for the practical application of the HK theorem to electronic structure calculations, including statistical and spatial electron correlation,\cite{kohn1965} and the relative simplicity of their application has transformed quantum chemical calculation of molecular systems. In this discussion, we focus exclusively on the conceptual implications of the HK theorem, since our objective is to derive a purely density-dependent force field, without reliance on wavefunction theory or Kohn-Sham orbitals.\cite{kohn1965}
The Hohenberg-Kohn theorem consists of three parts:\cite{dreizler1990} (i) the statement that knowledge of the ground state electron density $\rho({\bf r})$ for a many-electron system uniquely determines the configuration (magnitude and locations of nuclear charges---equivalently, the external potential $v({\bf r})$) for the interacting atoms that gave rise to $\rho({\bf r})$; (ii) formal proof of the existence of an energy functional $E_v[\rho({\bf r})]$ that provides an exact and complete representation of the quantum mechanics giving rise to the system ground state; and (iii) a corresponding density-based variational principle for determining the ground state density $\rho_{gs}({\bf r})$ by minimizing $E_v[\rho({\bf r})]$ over the space of all possible densities.
Taken together, statements (ii) and (iii) imply that the solution of the many-electron Schr\"{o}dinger equation, including a complete description of quantum mechanical exchange and correlation, can be replaced by the variational minimization of a simple energy functional, depending only on the electron density and external potential $v({\bf r})$ of the system:
\begin{equation}
E_v[\rho({\bf r})] = F_{\rm HK}[\rho({\bf r})] + \int \rho({\bf r})\, v({\bf r})\, d{\bf r},
\label{eq:HK}
\end{equation}
with
\begin{equation}
\rho_{gs}({\bf r}) = \argmin_{\rho({\bf r})} E_v[\rho({\bf r})].
\end{equation}
$v({\bf r})$, the system-specific external potential due to all $N_A$ atoms in the system, is given by:
\begin{equation}
v({\bf r}) = -\sum_{i=1}^{N_A} \frac{Z_i}{|{\bf r} - {\bf R}_i|},
\end{equation}
and $F_{\rm HK}[\rho({\bf r})]$ is the {\it universal} energy density functional. $F_{\rm HK}[\rho({\bf r})]$ consists of the sum of the total kinetic energy and quantum-mechanical electron-electron interaction energies of the system. Comparing Eqs.~(\ref{eq:TF}) and (\ref{eq:HK}), it is easy to see how the HK energy expression generalizes the original intuition of TF theory.
The functional $F_{\rm HK}[\rho({\bf r})]$ is universal because, unlike $v({\bf r})$, it is system-independent: the same $F_{\rm HK}[\rho({\bf r})]$ applies equally to atoms, molecules, and solids. The other remarkable feature of $F_{\rm HK}[\rho({\bf r}$)] (or equivalently, the exchange-correlation energy functional $E_{\rm xc}[\rho({\bf r})]$ appearing in the KS equations\cite{fn1}), is that, by definition, it embodies the full quantum mechanical complexity of the system, including statistical (exchange) and spatial electron correlations. These correlations would ordinarily require wavefunction quantum mechanics (e.g., configuration interaction or coupled-cluster solutions of the Schr\"{o}dinger equation to determine the full many-electron wavefunction $\Psi({\bf r}_1, {\bf r}_2, \cdots {\bf r}_N)$), at many orders of magnitude greater computational cost. Instead, these correlations are completely folded into the universal functional expression. However, since the HK theorem only proves the existence of the universal functional, the functional must be approximated in practice.\cite{cohen2012,sun2015,medvedev2017,williams2020}
It is important to note that the HK theorem not only opened the door to practical, accurate and efficient electronic structure calculations using the Kohn-Sham equations;\cite{kohn1965} it has also led to a fundamental shift in thinking about how to construct models of atomic interactions to directly incorporate quantum mechanical effects without the need to explicitly solve for a many-body wavefunction. We refer to this profound---and in principle, exact---relationship between the energy and density as \textit{energy-density duality}:\cite{fn2} the energy functional implies the ground state density through the HK theorem variational principle; conversely, the ground state density implies the energy, through the existence of the universal functional.
Energy-density duality has inspired the development of numerous models in which atomic-like densities are used to construct an energy model for a larger system. Early examples include the Gordon-Kim model of rare gas interactions\cite{gordon1972} and its extensions\cite{lacks1993}; the Harris functional;\cite{jharris1985} Cortona's ``atoms in solids'' method;\cite{cortona1991} frozen DFT;\cite{wesolowski1993,wesolowski2015} and the self-consistent atomic deformation method.\cite{boyer2008} Over the past decade, there has been renewed interest in DFT fragment-based approaches descended from these earlier ideas, including partition DFT,\cite{elliott2010} DFT embedding theory, \cite{huang2011,huang2011a,manby2012} and subsystem DFT\cite{jacob2014,sun2016}. These approaches implement a variety of theoretical techniques to impose a degree of quantum mechanical self-consistency between a smaller, embedded system---an atom, or cluster of atoms---treated at a different level of theory, and its environment. This enables a subsystem to be studied at a higher level of theory than its surroundings. The methods all require some form of subsystem registration with the computed wavefunction or density of the larger system---often accomplished through a Kohn-Sham potential---and as such, they are primarily aimed at computing the static electronic structures of large systems at fixed geometries. As with any fragment-based method, the choice of how to define the subsystem and environment can have a significant impact on results.\cite{sun2016}
Since the energy is a functional of the density, the flip side of energy-density duality is to focus attention first on the electron density, in addition to, or lieu of, a combined energy-density model. A notable early effort in this direction---like the present work, statistical in nature, and aimed at computing dynamically-evolving forces for large systems---is the elegant path integral formulation of Pratt and co-workers.\cite{harris1985,pratt1988,hoffman1988,pratt1990} Harris and Pratt used a discretized propagator method to show\cite{harris1985} that the electron density $\rho({\bf r})$ of a system could be expressed in terms of an effective single-particle potential\cite{fn3} via a $p$-dimensional integral over the coordinates of all possible $p$-segment paths initiating and terminating at ${\bf r} = {\bf r}_0 = {\bf r}_p$. The expression is remarkably simple:\cite{hoffman1988}
\begin{equation}
\rho({\bf r}) = 2\int d{\bf r}_1 \ldots \int d{\bf r}_{p-1} \left ( \frac{pk_p}{2\pi l_p}\right)^{3p/2} J_{3p/2} (k_pl_p)\, \eta(k_p)^2,
\label{eq:HP}
\end{equation}
where
\[
l_p^2 \equiv p \sum_{j=0}^{p-1} ({\bf r}_{j+1} - {\bf r}_j )^2,
\]
and
\[
\frac{\hbar^2 k_F^2}{2m} = \epsilon_F - \frac{1}{p} \sum_{j=1}^p v({\bf r}_j; p).
\]
Here $\eta$ is the Heaviside function; $J_\nu(z)$ is the Bessel function of order $\nu$, and $\epsilon_F$ is the Fermi energy.\cite{harris1985,hoffman1988} For the case $p = 1$, and suitable choice of $v({\bf r})$, this result reduces to Thomas-Fermi theory; for $p\rightarrow\infty$, it becomes exact.\cite{harris1985} The expression can thus be seen as interpolating beween the semi-classical Thomas-Fermi and fully-quantum mechanical regimes. While technically challenging to implement in practice due to the requisite high-dimensional numerical integration,\cite{pratt1988,hoffman1988} the Harris-Pratt formulation represents an important conceptual advance in focusing attention squarely on the electron density as a preferred target for modeling large systems. In particular, it embodies several features shared by the present work: it is orbital-free; it takes a bottom-up approach to evaluating the electron density scalar field for the complete system; global information from the larger system is incorporated into the evaluation of $\rho({\bf r})$, through contributions to argments in the integrand from the potential along sampled paths in the neighborhood of ${\bf r}$; the expression is statistical in nature; and it trains focus on the electron density as a more compact embodiment of the quantum substructure of a larger system than wavefunction-based approaches.\cite{pratt1990}
As detailed below, the statistical modeling of $\rho({\bf r})$ in the present work is accomplished through an ensemble, rather than path integral, formulation. Sampling of the local chemical environment is done through an ensemble spectral representation of atom-in-molecule densities\cite{Atlas2020} and EAM energies, and global information about the system is transmitted via chemical potential equalization adjustment of the corresponding ensemble weights.
\subsection{The embedded-atom method and its density functional foundations}
The embedded-atom method (EAM) for metals,\cite{daw1983,daw1984} the modified EAM (MEAM) for covalent materials,\cite{baskes1987,baskes1989} and their many extensions\cite{fn7} are among the most widely-used interaction potentials for simulating materials structure, defects, and phase diagrams.\cite{lee2010,foiles2012} The physical picture underlying the EAM {\it ansatz} was motivated by the development of density functional theory, through the quasiatom theory of Stott and Zaremba\cite{stott1980} and concurrently-developed effective medium theory of Norskov and Lang.\cite{norskov1980} These works analyzed the energy of an atom embedded in a uniform electron gas with compensating positive background. The empirical EAM model was designed to address the more complicated scenario of many interacting atoms, each viewed as an impurity embedded within a locally-defined host (electron density).
In a seminal work published several years after the introduction of the original EAM, Daw\cite{daw1989} demonstrated that the EAM could be derived---modulo an error term shown to be small under certain well-defined conditions---directly from Hohenberg-Kohn density functional theory. Although not presented in multiscale terms, this work demonstrated a direct link between the quantum and atomistic length scales through the electron density. It also suggested\cite{valone2006Kos} the tantalizing prospect of developing a theoretically-grounded charge-transfer extension of the EAM based on extensions to DFT itself. Demonstrating that this is indeed possible is the central result of this work.
The EAM expression for the cohesive energy $E_{\rm coh}$ (equivalent to the interaction potential $V(\{{\bf R}_i\})$ for a configuration of atomic nuclei located at $\{{\bf R}_i$\}) is remarkably simple. It is written as the sum of individual atomic embedding contributions, and a term consisting of pairwise electrostatic interactions between nucleii:
\begin{equation}
E_{{\rm coh}} = \sum_{i=1}^{N_{\rm at}} E_i,
\end{equation}
where
\begin{equation}
E_i \equiv F_i(\bar \rho_i(\textbf{R}_i)) + \frac{1}{2} \sum_{j\ne i}
\phi_{ij}(R_{ij}). \label{eq:EAM}
\end{equation}
$F_i$ is an element-dependent {\it embedding function} of the effective local background electron density $\overline{\rho}_i({\bf R}_i)$ at atomic site $i$, and represents the collective many-body effects of the other atoms in the host system, and $\phi_{ij}$ is the effective electrostatic pair potential between atoms $i$ and $j$. The inclusion of the many-body term $F_i$ in the energy expression differentiates the EAM from earlier two-body potentials used in materials modeling, including ionic systems.
In practice, the constant background value ${\bar \rho_i}$ at the site of each embedded nucleus must be approximated. In the standard approach, static model electron densities are associated with each nucleus, and parameterized along with the other functions in the model. $\overline{\rho}_i({\bf R}_i)$ is approximated as the sum of the tails of the $n_i$ nearest-neighbor atom electron densities at site $i$:
\begin{equation}
\overline{\rho}_i({\bf R}_i) \simeq \sum_{\stackrel{j=1}{j\ne
i}}^{n_i}{\rho_j^a({\bf R}_{ij})}. \label{rhobar}
\end{equation}
where $\rho_j^a$ corresponds to the isolated atom electron density of neighbor $j$. Second-nearest-neighbor methods have also been developed.\cite{lee2010}
The analysis of Daw showed that for the original EAM, with no charge polarization and no charge transfer, the optimal (energy-minimizing) embedding density could be chosen to minimize the error term resulting from the definition of the embedding term as a difference of DFT energy functionals.\cite{daw1989} This point will be discussed in greater detail below. Daw also suggested the possibility of computing the embedding density as a weighted average over the contributions of neighboring atoms within a suitable radius.\cite{daw1989} It has been noted that different physical interpretations of the background density can lead to very different EAM parameterization schemes, and can also limit the extent to which additional energy scales, associated with charge transfer, can be incorporated into the baseline model.\cite{valone2014} This is yet another manifestation of energy-density duality.
\subsection{Density functional theory of the atom-in-molecule}
The concept of the atom-in-molecule\cite{fn6} has a long history, dating back to the ideas of Lewis,\cite{lewis1916} Moffitt,\cite{moffitt1951} Hirshfeld,\cite{hirshfeld1977} Parr and coworkers,\cite{parr1978,nalewajski2000} and Bader.\cite{bader1985,bader1990} In the modern formulation, one asks how to decompose a given electron density distribution $\rho({\bf r})$ for a molecule or a material into ``chemically-reasonable'' atom-like components $\rho_i^*({\bf r})$:
\begin{equation}
\rho({\bf r}) = \sum_{i=1}^{N_{at}} \rho_i^*({\bf r}),
\label{rhoMap}
\end{equation}
where $N_{at}$ is the number of atoms in the system, and * is used to denote atom-in-molecule quantities. In principle, there are an infinity of ways to accomplish the decomposition, and the diversity of available schemas reflect differing theoretical and practical {\it desiderata}. Since the decomposition can be used to compute an effective charge $q_i$ for the $i$th atom-in-molecule
\begin{equation}
q_i = Z_i - \int \rho_i^*({\bf r})\, d{\bf r},
\end{equation}
any atom-in-molecule method allows comparison to be made with charge partitioning\cite{lowdin1950,mulliken1955} or other atom-in-molecule decomposition techniques.\cite{bader1990,heidar2017} Effective charges derived from atom-in-molecule decompositions can also be used to compute parametrization data for fixed-charge empirical force fields, or as serve target values for training empirical or machine-learned potentials.\cite{verstraelen2016,nerenberg2018,zubatyuk2019}
Historically, many atom-in-molecule decomposition strategies have been designed with the goal of transferability, e.g., the construction of reusable chemical ``Legos.''\cite{walker1993} However, since the objective in this work is to forge a systematic theoretical link between the quantum and atomistic length scales, we have developed an approach,\cite{Atlas2020} summarized in the following subsections, that does not impose a transferability bias; instead, the \textit{chemical environment itself} dynamically determines the atoms-in-molecule, consistent with structural and electronic changes occuring both locally and globally. The decomposition is a formal consequence of the Hohenberg-Kohn theorem,\cite{hohenberg1964} and there is no requirement of external theoretical\cite{bader1990} or reference state\cite{parr1978,nalewajski2000,bultinck2007} constraints. Since it follows directly from the HK theorem, the decomposition also avoids the need for self-consistent solution of a set of Kohn-Sham equations coupling the atoms-in-molecule to the larger system.\cite{cohen2007,atlas2007} The price paid for this apparent simplicity is that the decomposition, while unique, is not exact. Nevertheless, it has been shown to reproduce Bader charges as a function of internuclear separation for two challenging chemical systems, LiF and CO,\cite{Atlas2020} a surprising result given the very different nature of the two approaches. Whereas Bader's method\cite{bader1985,bader1990} partitions a molecular density into topologically-distinct ``basins'' of electronic charge, the density deconstruction procedure allows charge clouds of contributing atoms to overlap and intermix over all space. The agreement in effective charges derived from the two methods can be seen as reflecting the overriding importance of the external potential---as emphasized by the HK theorem---in determining the fate of electronic charge transfer for a given molecular geometry.
The next section summarizes the two ensemble DFT theories required by the density deconstruction procedure.\cite{Atlas2020} The first is a formulation of ensemble DFT for excited states,\cite{gross1988var,gross1988DFT,oliveira1988} to enable a description of atomic charge density polarization; the second is Perdew {\it et al.}'s ensemble DFT for fractional charge states,\cite{perdew1982} to enable a description of charge transfer. Both of these formalisms are valid only for isolated atoms; their application to the interacting atom case is discussed at the end of the section.
\subsubsection{Ensemble DFT formulations of charge polarization and charge transfer}
{\it Charge polarization}. Gross {\it et al.} proved an extended Rayleigh-Ritz variational principle\cite{gross1988var} for ensembles, and used this result to prove an ensemble generalization of the Hohenberg-Kohn theorem.\cite{gross1988DFT} Consider an atom A with external nuclear potential $v_{\rm A}({\bf r}) = -Z_{\rm A}/|{\bf R}_{\rm A} - {\bf r}|$, energy spectrum $\{E_i\}$, and a chosen set of $N_{ens}$ non-zero weights $\{\omega_i\},\ \ i=1,\ldots N_{ens}.$\cite{gross1988DFT} The energies are ordered as $E_1 \le E_2 \cdots E_{N_{ens}}$, and the weights are ordered such that $\omega_1 \ge \omega_2 ... \ge \omega_{N_{ens}} > 0.$ The weights are normalized to 1. Note that the ensemble state of atom A correspnds to an excitation only; the total number of electrons remains constant. In this case, Gross {\it et al.} showed that the ensemble energy is given by:
\begin{equation}
E_v[\{\omega_i\};\rho({\bf r})] = \sum_{i=1}^{N_{ens}} \omega_i E_i,
\end{equation}
and the DFT ensemble-energy-minimizing electron density is:
\begin{equation}
\rho({\bf r};\{\omega_i\}) = \sum_{i=1}^{N_{ens}} \omega_i \rho_i({\bf r}),
\end{equation}
where $\rho_i({\bf r})$ is the electron density of the state with energy $E_i$. This ensemble excited state picture has an alternative interpretation in terms of an electronic canonical (thermal) ensemble.\cite{kohn1986}
{\it Charge transfer.} For fractional charge state ensemble DFT, Perdew {\it et al.} considered an isolated open-system atom with externally-imposed charge $q$ ($0 \le q \le 1$). Using the constrained-search formulation of DFT\cite{levy1979}, they showed that the minimizing energy of the system takes an ensemble form, but with only two integer charge states $M=Z$ and $M+1$ of the atom contributing. The expression for the ensemble energy is:\cite{perdew1982}
\begin{equation}
E_v = (1-q)E_M + qE_{M+1},
\label{eq:PPLB-1}
\end{equation}
and the DFT ensemble-energy-minimizing electron density $\rho({\bf r})$ is given by:
\begin{equation}
\rho({\bf r}) = (1-q)\rho_M({\bf r}) + q\rho_{M+1}({\bf r}).
\label{eq:PPLB-2}
\end{equation}
Dreizler and Gross have noted that the constrained-search formulation for the isolated atom can be extended to include a manifold of states with arbitrary integer numbers of electrons, although in the isolated atom case the minimum energy is expected to be attained for the two-state case.\cite{dreizler1990} As with the excited state theory, there is an alternative thermodynamic interpretation of this result, in terms of an electronic grand canonical ensemble picture.\cite{gyftopoulos1968,perdew1982}
{\it Strong interactions and the atom-in-molecule.} The excited state and charge state ensemble expressions apply only to an isolated atom experiencing a ``{\it deus ex machina}'' fractional excited state occupation $\omega$ or fractional charge transfer $q$\cite{valone2006ED}---{\it i.e.,} the external imposition of a perturbation on the isolated system. This restriction is implicit in the assumption of ordered weights and states in the work of Gross {\it et al.}, and in the assumption of a weak interaction between an atom and its charge reservoir, in the work of Perdew {\it et al.} A general atomistic force field, however, must be able to describe an atom experiencing \textit{significant} charge polarization or charge transfer---including the possibility of fluctuating charge states exceeding unit charge---resulting from chemical bonding or reactive interactions. That is, it must be able to describe an arbitrary atom-in-molecule.
The charge transfer or charge polarization induced in the strong coupling case can be seen as equivalent to imposing an external interaction potential $v_{int}({\bf r})$ on the atom, in addition to its isolated-atom nuclear potential,\cite{cohen2006,cohen2007,scheffler2007} leading to an effective ``external potential'' $\tilde{v}({\bf r}) = v_{\rm A}({\bf r}) + v_{int}({\bf r})$ for the atom-in-molecule. Of course, this potential can be determined only as the result of a self-consistent electronic structure problem for the full system, which, by design, is not possible for a classical potential. Our solution, therefore, is to construct a representation of the atom-in-molecule density that utilizes the isolated atom ensemble states as {\it basis densities}. This representation, while approximate, can be shown to be a unique consequence of the Hohenberg-Kohn theorem,\cite{Atlas2020} and is summarized in the next section.
The basis density representation implements two changes relative to the weakly-interacting ensemble formalisms: the manifolds of contributing states can now include any physically-accessible states of the isolated atoms or ions; and the ensemble weights are no longer required to be ordered according to energy\cite{gross1988var,gross1988DFT} or charge state.\cite{perdew1982} The impact of the effective potential is detected indirectly, through modified ensemble weights appearing in both the density and energy contributions to the atomistic model. In fact, the appearance of charge-dependent weights in an ensemble representation of an atom-in-molecule can be regarded as a signature of the presence of an interaction potential.\cite{valone2006ED,kraisler2013,senjean2018} Recent work has shown that this observation can be turned to advantage in devising self-consistent exchange-correlation functionals that address the well-known derivative discontinuity and band gap problems through modified ensemble representations.\cite{kraisler2013,senjean2018} In contrast to the expanded ladder of states that will be used to construct the atomistic potential, these electronic structure methods are designed to retain compact two-\cite{kraisler2013} and three-state\cite{senjean2018} ensemble forms, while folding in the effect of $v_{int}({\bf r})$ through renormaliztion of the contributing integer-electron densities\cite{kraisler2013} or introduction of a weight-dependent exchange-correlation functional.\cite{senjean2018} These extensions to the isolated atom ensemble theories provide additional theoretical support for using ensemble weights to encode the effects of electronic interactions, as an alternative to determining the self-consistent densities and energies of a more restricted set of ensemble states.
\subsubsection{The ensemble DFT atom-in-molecule decomposition}
We now summarize the ensemble DFT atom-in-molecule decomposition\cite{Atlas2020} to be used in constructing the ensemble DFT CT-EAM.
The contribution of the $i$th atom-in-molecule to the density $\rho({\bf r})$ of a larger system is expressed as dual expansions in terms of both excited state and charge state densities of isolated atoms having the same, fixed nuclear charge $Z_i$ as the atom of interest. The decomposition $\tilde{\rho}({\bf r})$ is written as:
\begin{equation}
\rho({\bf r}) \rightarrow \tilde{\rho}({\bf r}) \equiv \sum_{i=1}^{N_{at}} \rho_i^*({\bf r}),
\label{eq:DD}
\end{equation}
where the atom-in-molecule densities $\rho_i^*({\bf r})$ are given by
\begin{equation}
\rho_i^*({\bf r}) = \sum_{j= -\infty}^{Z_i - 1} \alpha_{ij} \varrho_{ij}({\bf r}),
\label{eq:AIM-expans}
\end{equation}
with weights $\alpha_{ij} \ge 0$ $\forall$ $i$,$j$, and
the excited state ensemble densities $\varrho_{ij}({\bf r})$ for atom $i$ and charge state $j$ defined as
\begin{equation}
\varrho_{ij}({\bf r}) \equiv \sum_{k=0}^{\infty} \beta_{ijk} \rho_{ijk}({\bf r}),
\label{eq:excit-expans}
\end{equation}
with $\beta_{ijk}$ $\ge 0$ $\forall$ $i$, $j$, $k$. The \textit{basis density} $\rho_{ijk}({\bf r})$ is thus the density of the $k$th eigenstate of the $j$th ion of atom $i$ with number of electrons $N_{ij} = Z_i + j$. $j$ labels the charge state, and $k$, the excitation state of the atom. The typical range of $j$ is between the lowest and highest oxidation states of the atom-in-molecule. $N_i \equiv N_{i0} = Z_i$ is the number of electrons in the neutral atom. The charge ensemble and excitation ensemble weights satisfy separate sum rules: $\sum_{j= -\infty}^{Z_i - 1} \alpha_{ij} = 1$ for each atom $i$, and $\sum_{k=0}^{\infty} \beta_{ijk} = 1$ for each atom $i$ in ionic state $j$.
The decomposition of Eq.~(\ref{eq:DD}) can be understood intuitively from a consideration of the respective atomic and molecular potentials. Note that the external nuclear potential of the $i$th atom-in-molecule is the same for all atomic states contributing to its ensemble decomposition: $v_{i}({\bf r}) \equiv -Z_{i}/|\textbf{R}_i - \emph{\textbf{r}}|$. The decomposition uses a collection of electronic structure problems, for as many ionic and excited state excitations above the ground state as one chooses to select, to characterize the changing chemical environment of the atom-in-molecule (see Fig.~\ref{fig:ESProblems}).
\begin{figure}[htbp]
\includegraphics[scale=.595
]{figures/ESProblems.jpg}
\caption{Schematic illustration of an ensemble of electronic structure problems contributing to the characterization of an atom-in-molecule ${\rm A}^*$ with density $\rho_{\rm A}^*({\bf r})$, in a large biomolecular system. From left to right: isolated, neutral atom ground state density $\rho_{\rm A}^0({\bf r})$; isolated cation density with more compact electron density $\rho_{\rm A}^+({\bf r})$; isolated anion with diffuse electron density $\rho_{\rm A}^{-}({\bf r})$; isolated excited state density ${\rho}^{\prime}({\bf r})$ displaying charge polarization. (${\rho}^{\prime}({\bf r})$ is displayed as a non-spherical density for illustration purposes only. In practice, basis densities are spherically-averaged as noted in the text.) All four electronic structure problems derive from a common atomic potential, $v_{\rm A}({\bf r}) = -Z_{\rm A}/|{\bf r} - {\bf R}_{\rm A}|$, where ${\bf R}_{\rm A}$ is the location of nucleus A. An unnormalized equiensemble is illustrated at far right. During simulations, the ensemble weights for all states contributing to $\rho_{\rm A}^*({\bf r})$ will readjust according to chemical potential equalization (system-wide energy minimization with respect to the ensemble weights). There is no restriction on the number of states and corresponding basis densities that can be included in the ensemble for a given atom---these are chosen based on the particular integer charge and polarization states that are expected to be encountered in simulating a given system. Excited states of ions (not illustrated) may also be included.
}
\label{fig:ESProblems}
\end{figure}
The potential for the molecule is just the sum of the individual atom-in-molecule potentials, {\it i.e.}, $v({\bf r}) \equiv \sum_{i=1}^{N_{\rm A}} -Z_{i}/|\textbf{R}_i - \emph{\textbf{r}}|$, which is the same as the potential that would appear in the Hamiltonian for solving the exact molecular electronic structure problem. Instead, through the dual ensemble decomposition, an approximate molecular density $\tilde{\rho}({\bf r}) \approx \rho({\bf r})$ is given by the sum of the $N_A$ ensemble atoms-in-molecule.
The basis densities can be computed at a high level of quantum mechanical theory using standard electronic struture codes, and, as proposed elsewhere,\cite{Amokwao2020} converted to analytical radial basis functions satisfying a set of formal quantum mechanical constraints. These constraints govern the short-, medium-, and long-range behavior of atomic densities.\cite{Amokwao2020} THis provides a further mechanism for incorporating formal chemical and physical knowledge into the structure of the force field, since it is the tails of the atom-in-molecule densities that will contribute to the embedding density in the ensemble DFT CT-EAM (see Eq.~(\ref{eq:AIM-dens}) below).
Since the atom-in-molecule densities are expressed in terms of pre-computed, correlated basis densities, electron correlation effects are automatically propagated upward through the atom-in-molecule expansion to the atomic level. This is illustrated schematically in Fig.~\ref{fig:lengthscales}.
\begin{figure}[htbp]
\includegraphics[scale=.50
]{figures/LengthScales.jpg}
\caption{Particle correlations at the electronic and atomistic length scales. (a) Electrons interacting quantum mechanically. Beyond-Hartree-Fock electron correlations are computed by including determinants with increasing degrees of virtual orbital excitation within a configuration interaction (CI) expansion. Figure after Zgid {\it et al.}\cite{zgid2012} (b) Atoms interacting via an ensemble DFT classical force field. Effects of the chemical environment on the baseline isolated atom are computed by including contributions of excited (charge-distortion) and ionic (charge-transfer) atomic states within the force field. Illustration for carbon is schematic only (not to scale). Ionic states are indicated in blue; excited states in red. Relative energies from \citet{sasaki1974} and the NIST Atomic Spectra Database\cite{NIST2020}.}
\label{fig:lengthscales}
\end{figure}
Note that by combining Eqs.~(\ref{eq:AIM-expans}) and (\ref{eq:excit-expans}), $\rho_i^*({\bf r})$ may also be written as
\begin{equation}
\rho_i^*({\bf r}) = \sum_{jk} w_{ijk} \rho_{ijk}({\bf r}),
\label{eq:w-wts}
\end{equation}
where $w_{ijk} \equiv \alpha_{ij} \beta_{ijk}$, and as a consequence of the $\alpha_{ij}$ and $\beta_{ijk}$ sum rules, the $w_{ijk}$ satisfy the sum rule $\sum_{jk} w_{ijk} =1$ for each $i$.
\section{Results and Discussion}
\subsection{A DFT-based charge-transfer force field}
This section presents the main result of this work, the derivation of the ensemble DFT CT-EAM from density functional theory. The initial development and notation of Eqs.~(\ref{II.15})--(\ref{II.23}) closely follow that of Daw for the original EAM.\cite{daw1989} We then show how to incorporate the ensemble DFT energy and atom-in-molecule representations from the previous section to derive the CT-EAM generalization.
Daw's derivation of the EAM starts from the DFT expression for the cohesive energy of a configuration $\{{\bf R}_i\}$ of $N_A$ atoms in terms of the total self-consistent electron density $\rho({\bf
r)}$:
\begin{eqnarray}
E_{{\rm coh}} &=& G[\rho] + \frac{1}{2}\
{\sum_{i,j}}^{\prime}{\frac{Z_i Z_j }{|{\bf R}_i - {\bf R}_j|}} -
\sum_i \int d{\bf r}\ {\frac{Z_i\, \rho({\bf r} ) }{|\,{\bf r} -
{\bf R}_i |}} \nonumber \\
&+& \frac{1}{2} \int d{\bf r}\, d{\bf r^{\prime}}\,
{\frac{\rho({\bf r})\, \rho({\bf r^{\prime}}) }{|{\bf r} - {\bf
r^{\prime}}|}} - \sum _i E_i.
\label{II.15}
\end{eqnarray}
$Z_i$ and $E_i$ are the atomic number and total energy of the $i$th atom, respectively, and the prime indicates that the $i=j$ terms are omitted from the summation. $G[\rho]$ is given by:
\begin{equation}
G[\rho] = E_{{\rm xc}} [\rho] + T_s[\rho],
\label{eq:G}
\end{equation}
where the DFT exchange-correlation energy $E_{{\rm xc}} [\rho]$ and non-interacting kinetic energy $T_s[\rho]$ have been defined previously.\cite{fn1} The DFT expression for the isolated-atom energies $E_i$ is:
\begin{eqnarray}
E_i &=& E[\rho_i^a] = G[\rho_i^a] - \int d{\bf r}\ {\frac{Z_i\,
\rho_i^a({\bf r} - {\bf R}_i) }{|\,{\bf r} - {\bf R}_i |}}
\nonumber \\
&+& \frac{1}{2} \int d{\bf r}\, d{\bf r^{\prime}}\,
{\frac{\rho_i^a({\bf r} - {\bf R_i})\, \rho_i^a({\bf r^{\prime} -
{\bf R_i}}) }{|{\bf r} - {\bf r^{\prime}}|}},
\label{eq:isol-atom}
\end{eqnarray}
where the superscript $a$ denotes an isolated (undistorted) atomic density.
To simplify notation, it is helpful to establish the following definitions. The difference density
$\tilde{\rho}_{i}^{a}({\bf r})$ between the local electronic and
nuclear densities of the $i$th atom is defined as:
\begin{equation}
\tilde{\rho}_{i}^{a}({\bf r})\equiv \rho _{i}^{a}({\bf r}-{\bf
R}_{i})-Z_{i}\delta ({\bf r}-{\bf R}_{i}).
\label{eq:diff-den}
\end{equation}
The effective classical electrostatic interaction energy is:
\begin{equation}
U_{ij}^a \equiv \int d{\bf r}\, d{\bf r^{\prime}}\, {\frac{\tilde{\rho}_i^a(
{\bf r})\, \tilde{\rho}_j^a({\bf r^{\prime}}) }{|{\bf r} - {\bf r^{\prime}}|}
}. \label{eq:EScl}
\end{equation}
The {\it embedding energy} ${\cal G}_{i}(\bar{\rho})$ is defined as the net exchange-correlation and non-interacting kinetic energy required to
embed the $i$th atom in an electron gas of constant density $\bar{\rho}$:
\begin{equation}
{\cal G}_{i}(\bar{\rho})\equiv G[\rho _{i}^{a}+\bar{\rho}]-G[\rho
_{i}^{a}]-G[\bar{\rho}] \, .
\label{II.13}
\end{equation}
Note that electrostatic contributions are specifically excluded from the definition of ${\cal G }_{i}$, and $\bar{\rho}$ is implicitly assumed to be
neutralized by the compensating positive background density provided by a smeared-out positive background.\cite{daw1989}
It is now straightforward to show that the cohesive energy of Eq.\ (\ref{II.15}) may be rewritten in the following form:
\begin{equation}
E_{{\rm coh}} = G[\rho({\bf r})] - \sum_i G[\rho_i^a({\bf r})] +
\frac{1}{2} \,{\sum_{i,j}}^{\prime}U_{ij}^a.
\label{label:Ecoh-EAM}
\end{equation}
$U_{ij}^a$ clearly corresponds to the electrostatic functions $\phi _{ij}$
in Eq.\ (\ref{eq:EAM}); approximations are required in order to identify the
analogs of the embedding functions $F_i$.
The first assumption is that the total electron density $\rho ({\bf r})$
at site $i$ can
be approximated by the sum of the undistorted atomic densities $\rho_j^a(
{\bf r} - {\bf R}_i)$:
\begin{equation}
\rho({\bf r}) \approx \sum _j \rho_j^a({\bf r} - {\bf R}_i). \label{II.17}
\end{equation}
Note that this assumes a {\it different approximation to $\rho({\bf r})$ at each atomic site $i$}, which we indicate by $\varrho_i({\bf r})$. In constructing each local approximation, the undistorted atomic
density of atom $i$ removed, and the remainder is identified as the background density $\rho_{b,i}({\bf r})$ appropriate to that site:
\begin{equation}
\sum _j \rho_j^a({\bf r} - {\bf R}_i) = \rho_i^a({\bf r} - {\bf
R}_i) + \rho_{b,i}({\bf r}) \, ,
\label{II.18}
\end{equation}
where
\begin{equation}
\rho_{b,i}({\bf r}) \equiv \sum_{j \ne i} \rho_j^a({\bf r} - {\bf
R}_i) \, .
\label{II.19}
\end{equation}
The next step is to assume that the variation in $\rho_{b,i}({\bf r})$ is small in the vicinity of site $i$, relative to the dominant local
atomic density $\rho_i^a({\bf r})$, so that $\rho_{b,i}({\bf r})$ may be approximated by a constant $\bar{\rho _i}$:
\begin{equation}
\rho_{b,i}({\bf r}) \approx \bar{\rho _i} \, .
\label{II.19a}
\end{equation}
Eq.\ (\ref{II.19}) suggests taking
$\bar{\rho _i}$ to be the sum of electron density tails from neighboring atoms, evaluated at site $i$:
\begin{equation}
\bar{\rho _i} \approx \rho_{b,i}({\bf r}){\big \vert}_{{\bf r} =
{\bf R}_i} = \sum _{j \ne i} \rho_j^a({\bf R}_i - {\bf R}_j) \, .
\label{II.19b}
\end{equation}
This is the well-known background density approximation of EAM. It is equivalent to a weighted-density approximation to $\rho_{b,i}({\bf
r})$, in the limit of delta-function localization to the nucleus\cite{daw1989}
\begin{equation}
\bar{\rho _i} \approx \int d{\bf r} \sum_{j\ne i} \rho_j({\bf r} -
{\bf R} _j)\ w_i({\bf r}) \, ,
\label{II.19c}
\end{equation}
with $w_i({\bf r}) = \delta ({\bf r} - {\bf R}_i)$.
Combining Eqs.\ (\ref{II.17})--(\ref{II.19a}), the site-specific local
approximation $\varrho^{(i)} ({\bf r})$ to $\rho({\bf r})$ becomes:
\begin{equation}
\varrho^{(i)} ({\bf r}) \approx \rho_i^a({\bf r} - {\bf R}_i) +
\bar{\rho _i} \, .
\label{II.19d}
\end{equation}
We now perform a series of straightforward algebraic manipulations with $G[\rho ]$, making use of Eq.\ (\ref{II.19d}), and dropping explicit spatial dependencies for clarity. We have:
\begin{eqnarray}
G[\rho ] &=&G[\rho ]+\sum_{i}G[\varrho^{(i)}]-\sum_{i}G[\varrho^{(i)}] \nonumber \\
&=&\sum_{i}G[\rho _{i}^{a}+\bar{\rho _{i}}]+\left( G[\rho
]-\sum_{i}G[\varrho^{(i)}]\right) \nonumber \\
&=&\sum_{i}G[\rho _{i}^{a}+\bar{\rho _{i}}]-\sum_{i}G[\bar{\rho
_{i}}] \nonumber \\
&\ &{\rm \hspace*{0.1in}}+\left( G[\rho ]-\sum_{i}G[\varrho^{(i)}]+\sum_{i}G[ \bar{\rho _{i}}]\right) \nonumber \\
&=&\sum_{i}(G[\rho _{i}^{a}+\bar{\rho _{i}}]-G[\bar{\rho
_{i}}])+E_{{\rm err} } \, ,
\label{II.20}
\end{eqnarray}
with
\begin{equation}
E_{{\rm err}}\equiv G[\rho ]-\sum_{i}(G[\varrho^{(i)}]-G[\bar{\rho
_{i}}]) \, .
\label{II.21}
\end{equation}
Subtracting $\sum_{i}G[\rho _{i}^{a}]$ from each side of Eq.~(\ref{II.20}) and using the definition of the embedding energy from Eq.~(\ref{II.13}), one finds that
\begin{equation}
G[\rho ]-\sum_{i}G[\rho _{i}^{a}]=\sum_{i}{\cal G}_{i}[\bar{\rho
_{i}}]+ E_{{\rm err}} \, ,
\label{II.22}
\end{equation}
so that the cohesive energy (\ref{II.15}) becomes
\begin{equation}
E_{{\rm coh}}=\sum_{i}{\cal G}_{i}[\bar{\rho _{i}}]+
\frac{1}{2}\,{\sum_{i,j}}^{\prime }U_{ij}^{a}+E_{{\rm err}} \, .
\label{II.23}
\end{equation}
This is the EAM expression derived by Daw.\cite{daw1989} Although the focus of Daw's work was to assess the DFT-EAM connection numerically for a particular case (\textit{fcc} Ni) where electron density redistribution was expected to be small, Daw also recapitulated his derivation allowing for a charge redistribution (polarization) $\Delta\rho_i({\bf r})$ at the location of each atom, {\it i.e.}, $\rho_i^a({\bf r}) \rightarrow \rho_i^a({\bf r}) + \Delta\rho_i({\bf r}) \equiv \varrho_i^*({\bf r})$. He showed that an expression analogous to Eq.~(\ref{II.23}) could be derived, with $\rho_i^a({\bf r})$ replaced by $\varrho_i^*({\bf r})$ in the embedding density and electrostatic energy terms. Although the derivation does not apply to the charge-transfer case, it nevertheless presages the atom-in-molecule replacement of the static $\rho_i^a({\bf r})$ by the atom-in-molecule density $\rho_i^*({\bf r})$ introduced in the present work.
Eq.~(\ref{II.23}) is precisely the EAM form Eq.~(\ref{eq:EAM}), with an additional error term, $E_{{\rm err}}.$ Daw further showed that under certain well-defined approximations, including the use of a local gradient approximation for the density functional $G[\rho]$, minimizing $E_{{\rm err}}$ corresponds to the determination of an optimal background embedding density. While his derivation and analysis cannot be applied directly to situations with significant charge distortion and charge transfer, due to various localization and small-perturbation assumptions, it identified the core components that must be generalized in order to extend a DFT formulation of EAM to the strong interaction regime: (i) the density-dependent embedding functions embodying nonlocal, many-body quantum effects through their origins in the DFT functional $G[\rho]$; (ii) a pair interaction term describing electrostatic interactions between atom-centered densities; and (iii) an embedding density expressed in terms of neighboring atom-centered densities.
To extend EAM to the general case we therefore propose the following mappings, starting from Eq.~(\ref{eq:EAM}) and using the basis-density ensemble expression for the atom-in-molecule density (Eq.~(\ref{eq:w-wts})) and corresponding energy ensemble weights:
\begin{itemize}
\item[1.] Embedding function for atom $i$: $F_i \Rightarrow \sum_{M_i} \omega_{iM_i} F_{iM_i}$
\item[2.] Electrostatic energy for atom $i$: $\phi_{ij} \Rightarrow \sum_{M_iP_j} \omega_{iM_i} \omega_{jP_j} \phi_{ij,M_iP_j}$
\item[3.] Atom-centered density: $\rho_i^a({\bf r}) \Rightarrow \rho_i^*({\bf r})$
\item[4.] Embedding density: $\bar{\rho _i} \Rightarrow \bar{\rho} _i^*$
\end{itemize}
In the expressions above, since excited states of atoms with integer charge can also be included in the ensemble expansion, we have combined the excited-state and charge state indices into a single index $M_i$ looping over all $N_M$ charge and excitation ensemble states (a second combined index $P_j$ loops over the $N_P$ charge and excitation ensemble states of interacting atom $j$). With the mappings above, the final result for the general ensemble DFT CT-EAM is given by:\cite{fn8}
\begin{eqnarray}
E &=& \sum_{i} \bigg[ \sum_{M_i=1}^{N_M} \omega_{i,M_i} F_{i,M_i}
[\overline{\rho}_i^*] \nonumber \\
&+& \frac{1}{2}{\sum_{i,j}}^{\prime}
\sum_{M_i=1}^{N_M} \sum_{P_j=1}^{N_P} \omega_{i,M_i} \omega_{j,P_j}\Phi_{ij,M_iP_j}\bigg].
\label{eq:ensFF}
\end{eqnarray}
The background embedding electron density $\overline{\rho}_i^*$ is given by:
\begin{equation}
\overline{\rho}_i^*({\bf R}_i) \simeq \int d{\bf r}\, \tilde{\omega}({\bf r})\, \sum_{\stackrel{j=1}{j\ne
i}}^{n_i}{\rho_j^*({\bf r}-{\bf R}_j)}.
\label{eq:AIM-dens}
\end{equation}
$\tilde{\omega}({\bf r})$ is a weighting function that averages over the contributions of atom-in-molecule tails in a neighborhood of atom $i$. As noted above, the standard EAM formulation assumes $\tilde{\omega}({\bf r}) = \delta({\bf r} - {\bf R}_i)$.\cite{daw1989}
Several comments are in order here. A unique feature of the ensemble DFT CT-EAM is the requirement of internal self-consistency---between the density models appearing in the embedding and electrostatic components, and between the densities and energy functionals, which are both of ensemble form. The density consistency requirement implies that the atom-centered densities used to compute the effective background embedding density seen by the embedding functions, are identical to the atom-in-molecule densities used to compute effective electrostatic energies in the pairwise interaction term. This is because the $\rho_i^*({\bf r})$ have precise physical meaning as atom-in-molecule densities, rather than serving as mathematical proxies for a density-like quantity in a purely empirical model. In addition, the ensemble DFT-imposed consistency requirement between the embedding energies and atom-in-molecule densities requires that their respective state-dependent ensemble weights---in the atom-in-molecule ensemble expression (\ref{eq:DD})--(\ref{eq:excit-expans}) and in the EAM expression (\ref{eq:ensFF})---correspond.
We emphasize again that while we have presented arguments linking the various components of the EAM model---the embedding and electrostatic terms, atom-centered densities, and the embedding density---to formal DFT concepts, the relationship is not exact. In the end, Eq.~(\ref{eq:ensFF}) is an atomistic model, not a theory. The decomposition of the total density into atom-in-molecule ensembles omits interatomic correlations; the embedding density requires a weighted local averaging of a local superposition of atom-in-molecule tails; and the embedding and electrostatic interaction functions require parameterization against a set of coarse-grained physical characteristics (e.g. the energies of surfaces and defects in materials or metastable states of amino-acid conformers in proteins). From a practical standpoint, it will be necessary to develop a stable and efficient algorithm for computing the dynamically-changing weights $\omega_{iM}$ via global chemical potential equalization. Nevertheless, the hope is that by grounding the model in physical constructs that translate to the quantum level, it will be possible to attack a much broader range of chemical problems with greater overall fidelity than is currently feasible with more empirical approaches.
\subsection{Relation to subsystem and machine learning approaches}
The ensemble DFT formulation of Eq.~(\ref{eq:ensFF}) represents a particular approach to implementing energy-density duality within a quantum-informed force field. In light of the Hohenberg-Kohn theorem, one way to understand this class of models is by considering how the balance between descriptions of density and energy are implemented.
For comparison, the two methods most closely related to the present approach---subsystem embedding and machine learning---are illustrated in Fig.~\ref{fig:energy-density}, together with ensemble DFT CT-EAM.
\begin{figure}
\includegraphics[scale=.19
]{figures/Energy-Density.jpg}
\caption{Three approaches to energy-density duality. (a) Subsystem embedding: The subsystems of interest are embedded within the larger molecular environment. The energy $E_i^{\rm frag}$ of each subsystem---an atom or cluster of atoms---contributes to the total energy, in addition to any fragment-fragment interactions. This method can be used to study a subsystem at a higher level of theory than its surroundings, e.g. wavefunction-in-DFT. (b) Machine learning: The total energy is the sum of atomic energies computed via the application of local classifiers, typically implemented as neural networks. Environment-dependent effects and interatomic interactions are incorporated implicitly, through changes to the local descriptors input to the network as the system evolves in time. (c) Ensemble DFT CT-EAM (present work): The total energy is the sum of atomic energies computed from the ensemble DFT CT-EAM model described in the text. Relative weights of pre-computed basis densities for each atom are determined via global chemical potential equalization of an energy- and density-dependent functional. The weights determine the atomic contributions $E_i^{\rm ens}$ to the total energy. Only the weights vary over the course of a simulation. Here, the evolution of weights over time is represented as a notional audio spectrogram: the horizontal axis denotes time evolution of the system; the vertical axis corresponds to the discrete ensemble energy states of each atom in the system (neutral, ionic, excited, labeled by frequency $\Omega$), and color corresponds to the amplitude (ensemble weight) of a given energy state's contribution to the total. Image credits: (a) and (c), JSMol visualizations of kinesin-1 motor domain, PDB ID 5LT0\cite{cao2017}; (c), illustrative spectrogram generated using https://academo.org/demos/spectrum-analyzer/.}
\label{fig:energy-density}
\end{figure}
In all three approaches, there is a notion of quantum mechanical embedding of a molecular subsystem, be it a cluster or a single atom, within a broader chemical environment. Importantly, this chemical environment is also modeled quantum mechanically, and regardless of the details of the embedding process, there is a mechanism for quantum mechanical registration between the subsystem(s) and their environment. In the present approach, this is accomplished by imposing \textit{theoretical} self-consistency between the ensemble density and ensemble energy representations, as in the earlier isolated-atom ensemble DFT formulations, and \textit{computational} self-consistency, equivalent to global energy minimization through chemical potential equalization,\cite{sanderson1951,parr1978,rappe1991,rick1994} for determining the dynamically-changing ensemble weights. In all three models, the total energy is expressed as the sum of component energies (for subsystem DFT there are additional contributions from coupling to the surrounding environment; for ensemble DFT there is a an additional summation over pairwise electrostatic interactions).
It is instructive to further compare the two atom-centric methods---machine learning and ensemble DFT CT-EAM---to appreciate the conceptual differences in their formulations, and identify potential opportunties for cross-fertilization between their respective data-driven and theory-driven approaches.
{\it The (meta-)density as descriptor.} In keeping with energy-density duality, the present work utilizes a physical atom-in-molecule density as input to the energy functional of each atom. In the first generation of neural-network potentials, the emphasis was on demonstrating the feasibility of data-driven, statistical machine learning approaches for reproducing quantum-computed energies, or even improving upon them.\cite{ramakrishnan2015} Input descriptors were therefore constructed based on local geometry and modeling intuition, and descriptor and (neural network) classifier designs emphasized the use of mathematical basis functions to facilitate adherence to symmetry constraints.\cite{behler2007,behler2011}
As a result, many of these atomic descriptor models can now be seen as specific instances of a common, abstract structural representation.\cite{bartok2013,willatt2019} This has led to the development of the bispectrum family of descriptors\cite{bartok2013} and related potentials.\cite{bartok2010,thompson2015,wood2019} In parallel with these advances, and motivated by the goal of generating potential energy surfaces with quantum-level accuracy, there is growing interest in developing more realistic models of the electron-density-as-descriptor.\cite{ghasemi2015,grisafi2018, sinitskiy2018,dick2019} These include DFT \cite{bogojeski2020} and machine-learned ``atom-in-molecule''\cite{fn5} models.\cite{zubatyuk2019,huang2020} Such data-driven approaches are complementary to the physical modeling of the atom-in-molecule described here. A potentially fruitful avenue for future research will be to explore combinations of these very different perspectives.
{\it Energy-density duality and model generalizability.} Any theoretical model that is not exact introduces implicit bias into its predictions. In DFT, the bias enters in the form of approximations to the exchange-correlation functional: some functionals are oriented toward accurate descriptions of molecular structure; others, the study of chemical reactions; and still others, the study of extended systems (including functionals tailored for various sub-categories: metals, insulators, strongly-correlated materials, etc.) These biases are mirrored in the design of force field models, where they manifest themselves in the form of energy-density duality.\cite{fn4} However cleverly a model is designed, there are inevitable tradeoffs in how particular physico-chemical effects are built into model input (density, density descriptor vector) or energy predictor (model functional, machine learning classifier). Indeed, the act of model parameterization not only sets the values of hyperparameters, it also serves to compensate for missing correlations in the structure of the model itself.\cite{valone2006Kos} That is why physically-motivated, compact representations such as the one described here have a potential advantage over purely mathematical models containing thousands of parameters. In complex systems where competing chemical environments often co-exist (for instance, a solvated protein interacting with a surface), achieving a good balance between energy and density models, and the number of model parameters, can have a significant impact on generalizability.
In the present approach, the propagation of information from the quantum to atomistic length scale illustrated in Fig.~\ref{fig:lengthscales} is not perfect: there is missing interatomic correlation due to the approximate nature of the ensemble atom-in-molecule decomposition and corresponding ensemble representation of the energy. To preserve the simplicity of the model and avoid the need to parameterize angular-dependent basis densities, the atom-in-molecule basis densities are sphericalized:\cite{Amokwao2020} theoretical rationales for this strategy come from DFT\cite{theophilou2018} and the theory of function approximation. Nevertheless, it will be important to explore systematically the extent to which the ensemble representation, in combination with introduction of the charge- and polarization-dependent embedding energy functionals, is able to compensate for this missing information within the force field.
{\it Energy-density duality and length scale coupling.} The $\Delta$-DFT approach,\cite{bogojeski2020} motivated by the work of \citet{ramakrishnan2015}, uses a two-level machine learning approach to first learn a coupled-cluster-accuracy electron density from an input DFT density, followed by a second machine learning model which uses learned DFT densities to predict the total energy. This approach is of particular interest in the context of the present work, since it ties the construction of a machine-learned energy functional designed for atomistic simulation, to a machine-learned model for the true electron density functioning as a chemical descriptor, and links the two through the medium of DFT. A second notable feature, as discussed further below, is the explicit recognition of the importance of incorporating a range of excitation energies above the ground state in designing and training the density representation against a conformer training set.
{\it Reference states and training data.} We have previously noted an analogy between the zero-temperature electronic excitations characterizing spatial electron correlations, and atomic charge state excitations characterizing dynamic correlations among atoms interacting in a finite-temperature MD simulation. Correlations-as-electronic excitations are mirrored upward in length scale, to become correlations-as-atomic-charge-and-excited-states. The construction of the DFT ensemble force field makes explicit use of this analogy. At the next length scale up, the existence of stable and metastable structures that are free-energy-accessible at finite temperature can inform the selection of reference states and configurations for parameterization.
Our goal in this work has been to present a principled, DFT-based force field that can help explore not only the known states of the system, but also discover and explore the unexpected ones. An important application will be the construction of kinetic models based on quantum-informed, reactive-MD explorations of biomolecular potential energy surfaces.\cite{klus2018} In order to discover new states contributing to mechanistic dynamics, it is essential to be able to sample all relevant (and potentially unanticipated) atomic excitations arising from chemical interactions and reactive dynamics. In the machine learning case, this translates to the question of generalizability, and whether the force field can accurately extrapolate predictions from a limited dataset of randomly-sampled chemical bonding environments. For the ensemble DFT CT-EAM force field, the ensemble representation will guide the choice of parameterization strategy to emphasize the selection of a compact set of physically- and chemically-motivated structures.
\section{Conclusion}
In the language of modern machine learning, DFT is Natures's classifier for electronic structure. By formulating the CT-EAM in terms of ensemble DFT, we have designed a principled force field that describes both polarization and charge-transfer excitations relative to the ground states of the interacting atoms. The respective excited state and ionic charge densities further function as Nature's descriptors for characterizing local changes in electronic structure due to each atom's dynamically-changing environment. In this way, the ensemble-based atom-in-molecule formulation has a direct translation, through density functional theory, from the electronic to the atomic level. The coupling between length scales is mediated by the electron density---a quantum mechanical scalar field that is also an experimental observable. The new force field constitutes an atom-in-molecule ensemble charge-transfer generalization of the original embedded atom method (EAM) force field,\cite{daw1983,daw1984} originally developed for atomistic simulations of elemental materials assuming fixed atomic density distributions.\cite{daw1989} The foundation in EAM makes it possible to take advantage of a rich body of experience from condensed phase modeling: for example, information-theoretic\cite{baskes1987,muralidharan2007ED} and physics-based\cite{valone2014,valone2016} functional forms for the embedding energy; strategies for the judicious selection of structural reference systems\cite{valone2006Kos,valone2014,valone2016} to be used in parametrization, such as metastable (excited-state) amino acid conformers,\cite{Amokwao2012,monti2013} structures derived from symmetric dilation of vibrational modes,\cite{muralidharan2007ED,muralidharan2007Si} ionic fragments of mechanistic importance for a given system (e.g. ${\rm OH}^-$, ${\rm H}_3{\rm O}^+$, ${\rm H}_5{\rm O}_2^+$ and ${\rm H}_9{\rm O}_4^+$ for water);\cite{muralidharan2007ED,schmitt1999,day2002} and chemical insights derived from the wavefunction-based fragment Hamiltonian (FH) approach for characterizing atomic charge-transfer excitations within the embedding function.\cite{valone2011IM,valone2011WF,valone2014,valone2016} The atom-in-molecule formulation in terms of quantum-mechanical, physics-constrained atomic and ionic basis densities\cite{Amokwao2020} enables a straightforward interpretation of the dynamically-evolving chemistry effected through chemical potential equalization of ensemble weights. This is in contrast to machine-learned quantum chemical force fields, where it can be difficult to tease apart descriptor and descriptor-energy redundancies, and the use of automated descriptor discovery or empirically-chosen functions present ongoing challenges to interpretability.\cite{du2019}
The rigorous foundation of the force field in DFT, and its formulation in terms of atoms rather than larger fragments, affords a number of key advantages. While this work was motivated by the need for scalability to large, heterogeneous systems with many different elements, and the study of reactive biomolecular chemistry, it is equally applicable to materials problems such as charge transfer in the vicinity of interfaces and for complex alloys.\cite{streitzPRB1994,haftel2001,zhou2004} Since the characteristics of the local chemical environment are encoded through an atom's canonical set of basis densities and embedding energy functions, there is no need for multiple ``oxygens'' or ``carbons'' in order to parameterize and apply a force field for a particular atom under differing chemical conditions: the set of basis states is unique and immutable for a given element.
The foundation in DFT also provides a mechanism for a principled coupling between the quantum mechanical, atomistic, and kinetic modeling regimes, through the determination of Markov states extracted from chemically-informed MD trajectories.\cite{noe2017,klus2018} Carrying out such a multiscale programme may enable simulations to finally reach biophysically-relevant timescales without the loss of
quantum- and atomic-level details critical to describing charge-transfer-mediated (catalytic) dynamics. We are currently proceeding to applications of the ensemble DFT CT-EAM force field to the energetics of the amino acids, as fundamental building blocks for the study of proteins and their interactions, and as the foundation for constructing chemically-informed kinetic models of protein dynamics.
\begin{acknowledgements}
It is a pleasure to acknowledge Steve Valone for
numerous stimulating and insightful discussions of this problem. The author is grateful to the members of her research group, past and present, for their valuable contributions.
This work was supported in part by the National Science Foundation and by the DoD/DTRA CB Basic Research Program (Grant No. HDTRA1-09-1-0018).
\end{acknowledgements}
|
\subsection{The full interaction Hamiltonian}
We consider the Coulomb interactions
\begin{equation}
H_C=\frac{1}{2}\int d\,\mathbf{r}\,d\,\mathbf{r}^{\prime} \sum _{\sigma, \sigma ^{\prime}}\hat{\psi}_\sigma ^{\dagger}(\mathbf{r})\hat{\psi}_{\sigma ^{\prime}}^{\dagger}(\mathbf{r} ^{\prime})\frac{e^2}{4\pi \epsilon _0\vert \mathbf{r}-\mathbf{r}^{\prime}\vert}\hat{\psi}_{\sigma ^{\prime}}(\mathbf{r} ^{\prime}) \hat{\psi}_{\sigma}(\mathbf{r})
\label{eq:coulomb}
\end{equation}
where $\hat{\psi}_{\sigma}(\mathbf{r})$ is real-space electron annihilation operator at $\mathbf{r}$ site with spin $\sigma$. In the graphene system, this interaction can be written as
\begin{equation}
H_C=\frac{1}{2}\sum _{i i' j j'}\sum _{\alpha \alpha '\beta \beta '}\sum _{\sigma \sigma '} \hat{c}^{\dagger}_{i\alpha \sigma}\hat{c}^{\dagger}_{i'\alpha '\sigma '}U^{\alpha \beta \sigma , \alpha ' \beta ' \sigma '} _{ij,i'j'}\hat{c}_{j'\beta '\sigma '}\hat{c}_{j\beta \sigma}\;,
\end{equation}
where
\begin{align}
U^{\alpha \beta \sigma , \alpha ' \beta ' \sigma '} _{ij,i'j'}=&\int d \mathbf{r}\,d\mathbf{r}^{\prime}\,\frac{e^2}{4\pi \epsilon _0\vert \mathbf{r}-\mathbf{r}^{\prime}\vert}\,\phi ^*_\alpha (\mathbf{r}-\mathbf{R}_i-\tau _\alpha)\,\phi _\beta (\mathbf{r}-\mathbf{R}_j-\tau _\beta)\;\nonumber\\
&\times\phi ^*_{\alpha '}(\mathbf{r}-\mathbf{R}_i'-\tau _{\alpha '})\phi _{\beta '}(\mathbf{r}-\mathbf{R}_j'-\tau _{\beta '})\chi ^{\dagger}_\sigma \chi ^{\dagger}_{\sigma '}\chi _{\sigma '}\chi _{\sigma}\;.
\end{align}
Here $i$, $\alpha$, and $\sigma$ refer to Bravis lattice vectors, layer/sublattice index, and spin index. $\phi$ is Wannier function and $\chi$ is the two-component spinor wave function. We further assume that the "density-density" like interaction is dominant in the system, i.e., $U^{\alpha \beta \sigma , \alpha ' \beta ' \sigma '} _{ij,i'j'}\approx U^{\alpha \alpha \sigma , \alpha ' \alpha ' \sigma '} _{ii,i'i'}\equiv U_{i\alpha \sigma ,i'\alpha '\sigma '}$, then the Coulomb interaction is simplified to
\begin{align}
H_C=&\frac{1}{2}\sum _{i i'}\sum _{\alpha \alpha '}\sum _{\sigma \sigma '}\hat{c}^{\dagger}_{i\alpha \sigma}\hat{c}^{\dagger}_{i'\alpha '\sigma '}U_{i\alpha \sigma ,i'\alpha '\sigma '}\hat{c}_{i'\alpha '\sigma '}\hat{c}_{i\alpha \sigma}\;\nonumber\\
=&\frac{1}{2}\sum _{i\alpha \neq i'\alpha '}\sum _{\sigma \sigma '}\hat{c}^{\dagger}_{i\alpha \sigma}\hat{c}^{\dagger}_{i'\alpha '\sigma '}U_{i\alpha,i'\alpha '}\hat{c}_{i'\alpha '\sigma '}\hat{c}_{i\alpha \sigma}\;\nonumber\\
&+\sum _{i\alpha}U_0 \hat{c}^{\dagger}_{i\alpha\uparrow}\hat{c}^{\dagger}_{i\alpha\downarrow}\hat{c}_{i\alpha\downarrow}\hat{c}_{i\alpha\uparrow}
\end{align}
Here we can see that the Coulomb interaction can be divided into intersite Coulomb interaction and on-site Coulomb interaction. Given that the electron density in a typical moir\'e graphene system is extremely low, i.e., a few electrons per moir\'e supercell, the chance that two electrons meet at the same atomic site is very low. The Coulomb correlations between two electrons in the moir\'e system are mostly contributed by the inter-site Coulomb interactions.
Therefore, in most of the previous studies, the on-site Hubbard interaction has been neglected. In this work, we first consider the effects of the dominant inter-site Coulomb interactions, then we discuss the effects of the atomic on-site Hubbard interactions, which turn out to be crucial in lifting the (quasi-)degeneracy between the valley polarized and spin polarized states.
In order to model the screening effects to the electron-electron Coulomb interactions from the dielectric environment, we introduce two screening parameters: the background dielectric constant $\epsilon$ and the inverse screening length $\kappa$, and we assume the Coulomb interaction takes the Thomas-Fermi screened form:
\begin{equation}
V(\vert\mathbf{r}-\mathbf{r}'\vert)=\frac{e^2}{4\pi \epsilon \epsilon _0\vert \mathbf{r}-\mathbf{r}^{\prime}\vert}e^{-\kappa \vert \mathbf{r}-\mathbf{r}^{\prime}\vert}
\end{equation}
Then the Fourier transform is expressed as
\begin{align}
V(\mathbf{q})&=\frac{1}{\Omega _M}\int d\mathbf{r}\frac{e^2}{4\pi \epsilon \epsilon _0\vert \mathbf{r}\vert}e^{-\kappa \vert \mathbf{r}\vert} e^{-i\mathbf{q}\cdot \mathbf{r}}\;\nonumber\\
&=\frac{e^2}{2\Omega _M\epsilon _0\epsilon \sqrt{\vert\mathbf{q}\vert^2+\kappa ^2}}\;,
\end{align}
where $\Omega _M$ is the area of a moir\'e primitive cell.
At small twist angles, the intersite Coulomb interactions can be divided into the intra-valley term and the inter-valley term \cite{ashvin-double-bilayer-nc19}. The intra-valley term $H_{C}^{\rm{intra}}$ can be expressed as
\begin{equation}
H_{C}^{\rm{intra}}=\frac{1}{2N_s}\sum_{\alpha\alpha '}\sum_{\mu\mu ',\sigma\sigma '}\sum_{\mathbf{k}\mathbf{k} '\mathbf{q}}\,V(\mathbf{q})\,
\hat{c}^{\dagger}_{\mathbf{k}+\mathbf{q},\mu \sigma \alpha} \hat{c}^{\dagger}_{\mathbf{k}'-\mathbf{q},\mu '\sigma '\alpha '}\hat{c}_{\mathbf{k}',\mu '\sigma '\alpha '}\hat{c}_{\mathbf{k},\mu \sigma \alpha}\;,
\label{eq:h-intra}
\end{equation}
and the inter-valley term $H_{C}^{\rm{inter}}$ is expressed as
\begin{equation}
H_{C}^{\rm{inter}}=\frac{1}{2N_s}\sum_{\alpha\alpha '}\sum_{\mu ,\sigma\sigma '}\sum_{\mathbf{k}\mathbf{k} '\mathbf{q}}\,V(\vert\mathbf{K}-\mathbf{K}'\vert)\hat{c}^{\dagger}_{\mathbf{k}+\mathbf{q},\mu \sigma \alpha} \hat{c}^{\dagger}_{\mathbf{k}'-\mathbf{q},-\mu \sigma '\alpha '}\,
\hat{c}_{\mathbf{k}',\mu \sigma '\alpha '}\hat{c}_{\mathbf{k},-\mu \sigma \alpha}\;.
\label{eq:h-inter}
\end{equation}
Here $V(\mathbf{q})$ denotes the screened Coulomb interaction $V(\mathbf{q})\!=\!e^2/(\,2\Omega _M\epsilon \epsilon _0\sqrt{q^2+\kappa ^2}\,)$, where $\Omega _M$ is the area of moir\'e supercell, $\kappa$ is the inverse screening length and $\epsilon$ denotes background dielectric constant. $\epsilon$ and $\kappa$ will be treated as two free parameters in this work.
$H_{C}^{\rm{intra}}$ includes the Coulomb scattering processes of two electrons created and annihilated in the same valley, and $H_{C}^{\rm{inter}}$ includes the processes that two electrons are created in $\mu$ and $-\mu$ and get annihilated in $-\mu$ and $\mu$ valleys. Here the atomic wavevector $\mathbf{k}$ is expanded around the valley $K^{\mu}$ in the big Brillouin zone of the monolayer graphene, which can be decomposed as $\mathbf{k}=\widetilde{\mathbf{k}}+\mathbf{G}$, where $\widetilde{\mathbf{k}}$ is the moir\'e wavevector in the moir\'e BZ, and $\mathbf{G}$ denotes a moir\'e reciprocal lattice vector. We note that the typical intravalley interaction energy $V_M\approx e^2/(4\pi\epsilon_0\epsilon L_s)\approx 25\,$meV for twist angle $\theta\approx 1.2^{\circ}$ and $\epsilon\approx 5$; while the intervalley interaction $V(\vert\mathbf{K}-\mathbf{K}'\vert)\sim 0.35\,$meV for $\theta\approx 1.2^{\circ}$ and $\epsilon\approx 5$, which is two orders of magnitudes smaller than the intravalley interaction, thus we neglect the intervalley term (Eq.~(\ref{eq:h-inter}) in our calculations.
The electron annihilation operator can be transformed from the original basis to the band basis:
\begin{equation}
\hat{c}_{\mathbf{k},\mu\alpha\sigma}=\sum _n C_{\mu \alpha \mathbf{G},n}(\widetilde{\mathbf{k}})\,\hat{c}_{\mu\sigma , n\widetilde{\mathbf{k}}}\;,
\label{eq:transform}
\end{equation}
where $C_{\mu \sigma \alpha \mathbf{G},n}(\widetilde{\mathbf{k}})$ is the expansion coefficient in the $n$th Bloch eigenstate at $\widetilde{\mathbf{k}}$ of valley $\mu$:
\begin{equation}
\vert \Psi _{n\widetilde{\mathbf{k}}}^{\mu}\rangle=\sum _{\alpha \mathbf{G}}C_{\mu \alpha \mathbf{G},n}(\widetilde{\mathbf{k}})\,\vert \mu \sigma \alpha \mathbf{G},\widetilde{\mathbf{k}} \rangle\;.
\end{equation}
We note that the non-interacting Bloch functions are spin degenerate due to the separate spin rotational symmetry ($SU(2)\otimes SU(2)$ symmetry) of each valley. Using the transformation given in Eq.~(\ref{eq:transform}), the intravalley Coulomb interaction (Eq.~(\ref{eq:h-intra}) can be re-written in the band basis:
\begin{equation}
H^{\rm{intra}}=\frac{1}{2N_s}\sum _{\widetilde{\mathbf{k}} \widetilde{\mathbf{k}}'\widetilde{\mathbf{q}}}\sum_{\substack{\mu\mu' \\ \sigma\sigma'}}\sum_{\substack{nm\\ n'm'}}\left(\sum _{\mathbf{Q}}\,V(\mathbf{Q}+\mathbf{\widetilde{q}})\,\Omega^{\mu \sigma,\mu'\sigma'}_{nm,n'm'}(\widetilde{\mathbf{k}},\widetilde{\mathbf{k}}',\widetilde{\mathbf{q}},\mathbf{Q})\right)\hat{c}^{\dagger}_{\mu\sigma,n\widetilde{\mathbf{k}}+\widetilde{\mathbf{q}}} \hat{c}^{\dagger}_{\mu'\sigma',n'\widetilde{\mathbf{k}}'-\widetilde{\mathbf{q}}}\,\hat{c}_{\mu'\sigma',m'\widetilde{\mathbf{k}}'}\,\hat{c}_{\mu\sigma,m\widetilde{\mathbf{k}}}
\label{eq:Hintra-band}
\end{equation}
where the form factor $\Omega ^{\mu \sigma,\mu'\sigma'}_{nm,n'm'}$ is written as
\begin{equation}
\Omega ^{\mu \sigma,\mu'\sigma'}_{nm,n'm'}(\widetilde{\mathbf{k}},\widetilde{\mathbf{k}}',\widetilde{\mathbf{q}},\mathbf{Q})
=\sum _{\alpha\alpha'\mathbf{G}\mathbf{G}'}C^*_{\mu\sigma\alpha\mathbf{G}+\mathbf{Q},n\widetilde{\mathbf{k}}+\widetilde{\mathbf{q}}}C^*_{\mu'\sigma'\alpha'\mathbf{G}'-\mathbf{Q},n'\widetilde{\mathbf{k}}'-\widetilde{\mathbf{q}}}C_{\mu'\sigma'\alpha'\mathbf{G}',m'\widetilde{\mathbf{k}}'}C_{\mu\sigma\alpha\mathbf{G},m\widetilde{\mathbf{k}}}
\end{equation}
We make Hartree-Fock approximation to Eq.~(\ref{eq:Hintra-band}) such that the two-particle Hamiltonian is decomposed into a superposition of the Hartree and Fock single-particle Hamiltonians, where the Hartree term is expressed as
\begin{equation}
\begin{split}
H_H^{\rm{intra}}=&\frac{1}{2N_s}\sum _{\widetilde{\mathbf{k}} \widetilde{\mathbf{k}}'}\sum _{\substack{\mu\mu'\\ \sigma\sigma'}}\sum_{\substack{nm\\ n'm'}}\left(\sum _{\mathbf{Q}} V(\mathbf{Q})\Omega ^{\mu \sigma,\mu'\sigma'}_{nm,n'm'}(\widetilde{\mathbf{k}},\widetilde{\mathbf{k}}',0,\mathbf{Q})\right)\\
&\times \left(\langle \hat{c}^{\dagger}_{\mu\sigma,n\widetilde{\mathbf{k}}}\hat{c}_{\mu\sigma,m\widetilde{\mathbf{k}}}\rangle \hat{c}^{\dagger}_{\mu'\sigma',n'\widetilde{\mathbf{k}}'}\hat{c}_{\mu'\sigma',m'\widetilde{\mathbf{k}}'} + \langle \hat{c}^{\dagger}_{\mu'\sigma',n'\widetilde{\mathbf{k}}'}\hat{c}_{\mu'\sigma',m'\widetilde{\mathbf{k}}'}\rangle \hat{c}^{\dagger}_{\mu\sigma,n\widetilde{\mathbf{k}}}\hat{c}_{\mu\sigma,m\widetilde{\mathbf{k}}}\right)
\end{split}
\end{equation}
and the Fock term is expressed as:
\begin{equation}
\begin{split}
H_F^{\rm{intra}}=&-\frac{1}{2N_s}\sum _{\widetilde{\mathbf{k}} \widetilde{\mathbf{k}}'}\sum _{\substack{\mu\mu'\\ \sigma\sigma'}}\sum_{\substack{nm\\ n'm'}}\left(\sum _{\mathbf{Q}} V(\widetilde{\mathbf{k}}’-\widetilde{\mathbf{k}}+\mathbf{Q})\Omega ^{\mu \sigma,\mu'\sigma'}_{nm,n'm'}(\widetilde{\mathbf{k}},\widetilde{\mathbf{k}}',\widetilde{\mathbf{k}}’-\widetilde{\mathbf{k}},\mathbf{Q})\right)\\
&\times \left(\langle \hat{c}^{\dagger}_{\mu\sigma,n\widetilde{\mathbf{k}}'}\hat{c}_{\mu'\sigma',m'\widetilde{\mathbf{k}}'}\rangle \hat{c}^{\dagger}_{\mu'\sigma',n'\widetilde{\mathbf{k}}}\hat{c}_{\mu\sigma,m\widetilde{\mathbf{k}}} + \langle \hat{c}^{\dagger}_{\mu'\sigma',n'\widetilde{\mathbf{k}}}\hat{c}_{\mu\sigma,m\widetilde{\mathbf{k}}}\rangle \hat{c}^{\dagger}_{\mu\sigma,n\widetilde{\mathbf{k}}'}\hat{c}_{\mu'\sigma',m'\widetilde{\mathbf{k}}'}\right)\;.
\end{split}
\end{equation}
Since the inter-site interaction $V(\textbf{q})\!=\!e^2/(2\epsilon\epsilon_0\Omega_M\sqrt{q^2+\kappa ^2})$, the intervalley section of the intersite Coulomb interaction is much weaker than intravalley section as discussed above. However, in the atomic on-site Hubbard interaction, the characteristic interaction strength is independent of wavevector $\mathbf{q}$, thus both intervalley and intravalley sections of the on-site Hubbard interaction have to be taken into account. In particular, the intravalley section of on-site Hubbard interaction with Hartree-Fock approximation in the band basis can be expressed as:
\begin{equation}
\begin{split}
H_{\rm{on-site}}^{\rm{intra}}=&\frac{U_0a^2}{L_s^2N_s}\sum _{\widetilde{\mathbf{k}} \widetilde{\mathbf{k}}'}\sum _{\substack{\mu\mu'\\ \sigma\sigma'}}\sum_{\substack{nm\\ n'm'}}\left(\sum _{\mathbf{Q}} \sum _{\alpha\alpha'\mathbf{G}\mathbf{G}'}C^*_{\mu\uparrow\alpha\mathbf{G}+\mathbf{Q},n\widetilde{\mathbf{k}}}C^*_{\mu'\downarrow\alpha'\mathbf{G}'-\mathbf{Q},n'\widetilde{\mathbf{k}}'}C_{\mu'\downarrow\alpha'\mathbf{G}',m'\widetilde{\mathbf{k}}'}C_{\mu\uparrow\alpha\mathbf{G},m\widetilde{\mathbf{k}}} \right)\\
&\times \left(\langle \hat{c}^{\dagger}_{\mu\uparrow,n\widetilde{\mathbf{k}}}\hat{c}_{\mu\uparrow,m\widetilde{\mathbf{k}}}\rangle \hat{c}^{\dagger}_{\mu'\downarrow,n'\widetilde{\mathbf{k}}'}\hat{c}_{\mu'\downarrow,m'\widetilde{\mathbf{k}}'} + \langle \hat{c}^{\dagger}_{\mu'\downarrow,n'\widetilde{\mathbf{k}}'}\hat{c}_{\mu'\downarrow,m'\widetilde{\mathbf{k}}'}\rangle \hat{c}^{\dagger}_{\mu\uparrow,n\widetilde{\mathbf{k}}}\hat{c}_{\mu\uparrow,m\widetilde{\mathbf{k}}}\right)
\end{split}
\label{eq:hubbard1}
\end{equation}
and the intervalley section of the on-site interaction with Hartree-Fock approximation in the band basis is expressed as
\begin{equation}
\begin{split}
H_{\rm{on-site}}^{\rm{inter}}=&\frac{U_0a^2}{L_s^2N_s}\sum _{\widetilde{\mathbf{k}} \widetilde{\mathbf{k}}'}\sum _{\substack{\mu\\ \sigma\sigma'}}\sum_{\substack{nm\\ n'm'}}\left(\sum _{\mathbf{Q}} \sum _{\alpha\alpha'\mathbf{G}\mathbf{G}'}C^*_{\mu\uparrow\alpha\mathbf{G}+\mathbf{Q},n\widetilde{\mathbf{k}}}C^*_{-\mu\downarrow\alpha'\mathbf{G}'-\mathbf{Q},n'\widetilde{\mathbf{k}}'}C_{\mu\downarrow\alpha'\mathbf{G}',m'\widetilde{\mathbf{k}}'}C_{-\mu\uparrow\alpha\mathbf{G},m\widetilde{\mathbf{k}}} \right)\\
&\times \left(\langle \hat{c}^{\dagger}_{\mu\uparrow,n\widetilde{\mathbf{k}}}\hat{c}_{-\mu\uparrow,m\widetilde{\mathbf{k}}}\rangle \hat{c}^{\dagger}_{-\mu\downarrow,n'\widetilde{\mathbf{k}}'}\hat{c}_{\mu\downarrow,m'\widetilde{\mathbf{k}}'} + \langle \hat{c}^{\dagger}_{-\mu\downarrow,n'\widetilde{\mathbf{k}}'}\hat{c}_{\mu\downarrow,m'\widetilde{\mathbf{k}}'}\rangle \hat{c}^{\dagger}_{\mu\uparrow,n\widetilde{\mathbf{k}}}\hat{c}_{-\mu\uparrow,m\widetilde{\mathbf{k}}}\right)
\end{split}
\label{eq:hubbard2}
\end{equation}
Since the full interacting Hamiltonian of twisted graphene system preserves spin rotational symmetry ($SU(2)$ symmetry), one can fix the spin quantization axis (the spin ``$z$" axis), then the spin-flip density matrix $\langle \hat{c}_{\uparrow}^{\dagger} \hat{c}_{\downarrow}$ vanishes. Therefore, the Fock term vanishes for the Hubbard interactions as shown in Eq.~(\ref{eq:hubbard1}) and Eq.~(\ref{eq:hubbard2}).
We continue to discuss the symmetries of the full interacting Hamiltonian of twisted multilayer graphene systems. First, all the twisted graphene systems have time-reversal ($\mathcal{T}$) symmetry and $C_{3z}$ symmetry. Some of the twisted multilayer graphene systems may have additional crystalline symmetries. For example, in free-standing TBG, there is $C_{2z}$ symmetry; in $AB$-$AB$ stacked TDBG there is $C_{2x}$ symmetry. These additional crystalline symmetries are important in the determining the topological properties of the energy bands. For example, when $C_{2z}$ symmetry is present in TBG, the $C_{2z}$ combined with $\mathcal{T}$ symmetry ($C_{2z}\mathcal{T}$) enforces that the Berry curvature has to vanish at every $\widetilde{\mathbf{k}}$ point in the moir\'e BZ for each valley; and the $C_{2x}$ symmetry of $AB$-$AB$ stacked TDBG enforces that the total valley Chern number of the two flat bands have to be zero \cite{koshino-tdbg-prb19,jpliu-prx19}.
Moreover, if one only considers the non-interacting continuum model Eq.~(\ref{eq:HMN}) and the intravalley inter-site Coulomb interaction Eq.~(\ref{eq:h-intra}), then all the twisted graphene systems have $U(1)\times U_v(1)\times SU(2)\times SU(2)$ symmetry which can be interpreted as follows. The Bloch states of graphene around $K$ and $K'$ valleys can be separately folded into the moir\'e BZ. For small twist angles the separation between the low-energy states around the $K$ and $K'$ valleys $\sim\!\vert\mathbf{K}-\mathbf{K}' \vert=4\pi/(3a)$ is much greater than the size of the moir\'e reciprocal lattice vector $\vert\mathbf{g}_1\vert=4\pi/(\sqrt{3}L_s)$. Therefore the coupling between the low-energy states around the two valleys can be neglected, because the Fourier components of the corresponding moir\'e potential is vanishingly small given that the potential is smooth on the moir\'e length scale and that $L_s\!\gg\!a$. As a result, the charge is separately conserved for each valley, and the low-energy states of the system has an emergent valley $U(1)$ symmetry \cite{po-prx18} (dubbed as $U_v(1)$). Moreover, as spin-orbit coupling is negligible in graphene and at the non-interacting level the two valleys are approximately decoupled for small twist angles, approximately there is separate spin $SU(2)$ symmetry for each valley. Therefore, all the moir\'e graphene systems have the approximate continuous $U(1)\times U_v(1)\times SU(2)\times SU(2)$ symmetry \cite{po-prx18}, where $U(1)$ stands for global charge conservation symmetry. Such a symmetry can be re-written as $U(2)\otimes U(2)$ symmetry, which means that there is separate charge-conservation and spin rotational symmetry for each valley. Such $U(2)\times U(2)$ symmetry is preserved if one only includes the intravalley inter-site Coulomb interaction Eq.~(\ref{eq:h-intra}). If one includes the atomic Hubbard interaction, the two valleys are coupled by the intervalley component of the Hubbard interaction (Eq.~(\ref{eq:hubbard2})), which reduces the $U(2)\times U(2)$ symmetry to a global $U(2)$ symmetry, i.e., there is only global charge conservation and global spin rotational symmetry if atomic Hubbard interaction is included.
First let us neglect the atomic Hubbard interactions ,then a generic twisted multilayer graphene system \textit{at least} has $U(2)\times U(2)$, $\mathcal{T}$, and $C_{3z}$ symmetry. The generators of the $U(2)\times U(2)$ symmetry are $\{\tau^{0,z}\, s^{a}, a=0,x,y,z\}$, the $C_{3z}$ operator can be represented as $C_{3z}=e^{-i(2\pi/3)\tau_z\sigma_z}$, and the time-reversal operator $\mathcal{T}=\tau_x\mathcal{K}$, where $\tau^{i}$, $s^{j}$, and $\sigma^{k}$ denote Pauli matrices in the valley, spin, and sublattice spaces respectively. $\mathcal{K}$ denotes complex conjugation operation.
Under a symmetry operation $g$, the electron annihilation operator $\hat{c}_{\mathbf{k},\mu\sigma\alpha}$ is transformed to:
\begin{equation}
g\,\hat{c}_{\mathbf{k},\mu\sigma\alpha}\,g^{-1}=\sum_{\mu'\sigma'\alpha'}\,(\hat{O}_g)_{\mu'\sigma'\alpha',\mu\sigma\alpha}\,\hat{c}_{g^{-1}\mathbf{k},\mu'\sigma'\alpha'}\;,
\label{eq:c-symmetry}
\end{equation}
Then one can easily verify that the intravalley Coulomb interaction Eq.~(\ref{eq:h-intra}) is invariant under $U(2)\otimes U(2)$, $\mathcal{T}$ and $C_{3z}$ symmetries.
Including the atomic Hubbard interactions would break the $U(2)\times U(2)$ symmetry to a global $U(2)$ symmetry, i.e., there is only total charge conservation global spin rotational symmetry.
In our calculations we project the interaction Hamiltonian given onto a few energy bands around the charge-neutrality point (CNP), assuming that interaction effects are negligible for energy bands at higher energies. In particular, we have performed Hartree-Fock calculations with interactions projected onto 2, 4, and 6 low-energy bands around the CNP for each valley each spin. The results are qualitatively consistent for all the twisted graphene systems.
\vspace{12pt}
\begin{center}
\textbf{\large {\uppercase\expandafter{\romannumeral 5}}\ Symmetry analysis on the order parameters}
\end{center}
In this section we discuss how the order parameters defined in the valley, spin, and sublattice space are transformed under symmetry operations.
First we define the order parameter $\hat{O}^{a b c}=\tau^{a}\, s^{b} \,\mathbb{I}_{N_l\times N_l}\, \sigma^{c}$ in the valley-spin-sublattice space, where $\mathbf{\tau}$, $\mathbf{s}$, and $\mathbf{\sigma}$ represent the Pauli matrices in the valley, spin, and sublattice spaces, with $a, b, c=0,x,y,z$ (the zeroth component of the Pauli matrix is a $2\times 2$ identity matrix). $\mathbb{I}_{N_l\times N_l}$ is an identity matrix of dimension $N_l=(M+N)$ defined in the layer space ($N_l=M+N$ is the number of layers). Here we only consider the order parameters in the valley, spin, and sublattice space, and take the average over the layer degrees of freedom.
The expectation value of the order parameter $\hat{O}^{a, b, c}$ at moir\'e wavevector $\widetilde{\mathbf{k}}$ can be expressed as
\begin{equation}
\Delta_{a b c}(\widetilde{\mathbf{k}})=\sum _n \left\langle \Psi _{n\widetilde{\mathbf{k}}}\vert \hat{O}^{a b c} \vert \Psi _{n\widetilde{\mathbf{k}}} \right\rangle\theta(\varepsilon _F-E_{n\widetilde{\mathbf{k}}})\;,
\label{eq:Delta}
\end{equation}
where $\theta(\varepsilon _F-E _{n\widetilde{\mathbf{k}}})$ refers to Fermi-Dirac distribution at zero temperature and
\begin{equation}
\vert \Psi _{n\widetilde{\mathbf{k}}} \rangle_{\rm{HF}} = \sum _{\mu \sigma l \alpha \mathbf{G}}\,C_{\mu \sigma l \alpha \mathbf{G},n\widetilde{\mathbf{k}}}^{\rm{HF}}\,\vert \mu \sigma l\alpha,\mathbf{G}+\widetilde{\mathbf{k}}\rangle
\end{equation}
is the Bloch eigenstate of a Hartree-Fock Hamiltonian expressed in the original basis of the continuum model, and $\mu$, $\sigma$, $l$, and $\alpha$ refer to the valley, spin, layer, and sublattice indices.
The basis function $\langle \mathbf{r} \vert \mu s l \sigma,\mathbf{G}+\widetilde{\mathbf{k}}\rangle = \chi _{\mu sl\sigma,\widetilde{\mathbf{k}}+\mathbf{G}}(\mathbf{r})\,e^{i(\widetilde{\mathbf{k}}+\mathbf{G})\cdot\mathbf{r}}$ transforms as follows under symmetry operation $g$:
\begin{align}
\chi _{\mu\sigma l\alpha,\widetilde{\mathbf{k}}+\mathbf{G}}(\mathbf{r})e^{i(\widetilde{\mathbf{k}}+\mathbf{G})\cdot\mathbf{r}} \longrightarrow \begin{cases}\sum _{\mu'\sigma' l'\alpha'}\,(O_{\mathit{g}})_{\mu'\sigma' l'\alpha',\mu \sigma l \alpha}\,\chi _{\mu'\sigma' l' \alpha',g^{-1}(\widetilde{\mathbf{k}}+\mathbf{G})}(\mathbf{r})\,e^{i\mathit{g}^{-1}(\widetilde{\mathbf{k}}+\mathbf{G})\cdot \mathbf{r}}\; & \hbox{if no $\mathcal{T}$ operation involved}\nonumber\\
\sum _{\mu'\sigma' l'\alpha'}\,(O_{\mathit{g}})_{\mu'\sigma' l'\alpha',\mu \sigma l\alpha}\,\chi^{*} _{\mu'\sigma' l'\alpha', -g^{-1}(\widetilde{\mathbf{k}}+\mathbf{G})}(\mathbf{r})\,e^{-i g^{-1}(\widetilde{\mathbf{k}}+\mathbf{G})\cdot \mathbf{r}}\; & \hbox{if $\mathcal{T}$ operation involved}
\end{cases}
\end{align}
Therefore, under symmetry operation $g$, the matrix element of the order parameter $\hat{O}^{a b c}$ in the original basis of the continuum model $\langle \mu'\sigma' l'\alpha',\mathbf{G}+\widetilde{\mathbf{k}} \vert\, \hat{O}^{a b c}\, \vert\mu\sigma l\alpha,\mathbf{G}+\widetilde{\mathbf{k}}\rangle$ becomes
\begin{equation}
\langle \mu'\sigma'l'\alpha',\mathbf{G}+\widetilde{\mathbf{k}} \vert \hat{O}^{a b c} \vert \mu\sigma l \alpha,\mathbf{G}+\widetilde{\mathbf{k}}\rangle \longrightarrow (\,O^{\dagger}_g(g^{-1}\widetilde{\mathbf{k}},g^{-1}\mathbf{G})\,\hat{O}^{a b c}\, O_g(g^{-1}\widetilde{\mathbf{k}}, g^{-1}\mathbf{G})\,)_{\mu'\sigma' l'\sigma',\mu\sigma l\sigma}
\end{equation}
where ``$(g^{-1}\widetilde{\mathbf{k}},g^{-1}\mathbf{G})$" means that the symmetry operation transforms $\widetilde{\mathbf{k}}$ and $\mathbf{G}$ to $g^{-1}\widetilde{\mathbf{k}}$ and $g^{-1}\mathbf{G}$.
Thus the expectation value $\Delta _{a b c}$ under $\mathit{g}$ operation becomes
\begin{equation}
\Delta _{a b c d}(\widetilde{\mathbf{k}}) \to \begin{cases}
&\Delta ^{\mathit{g}}_{a b c}(\mathit{g}^{-1}\widetilde{\mathbf{k}})=\sum _n\langle\Psi_{n\mathit{g}^{-1}\widetilde{\mathbf{k}}}\vert\, (\,O^{\dagger}_{\mathit{g}}\,\hat{O}^{a b c}\,O_{\mathit{g}})\,\vert \Psi_{n\mathit{g}^{-1}\widetilde{\mathbf{k}}}\rangle\,\theta(\varepsilon _F-E_{n\widetilde{\mathbf{k}}})\;,\hbox{if $\mathcal{T}$ operation is not involved in $g$}\;,\nonumber\\
&\Delta ^{\mathit{g}}_{a b c d}(-\mathit{g}^{-1}\widetilde{\mathbf{k}})=\sum _n\,\langle\Psi_{n-\mathit{g}^{-1}\widetilde{\mathbf{k}}}\vert\, (\,O^{\dagger}_{\mathit{g}}\,(\hat{O}^{a b c})^*\, O_{\mathit{g}}\,)\,\vert \Psi_{n-\mathit{g}^{-1}\widetilde{\mathbf{k}}}\rangle\,\theta(\varepsilon _F-E_{n\widetilde{\mathbf{k}}})\;, \hbox{if $\mathcal{T}$ operation is involved in $g$}\;.
\end{cases}
\end{equation}
If $g$ is a symmetry, it is required that $\Delta _{abc}(\widetilde{\mathbf{k}}) = \Delta^{g}_{a b c}(g^{-1}\widetilde{\mathbf{k}})$ or $\Delta _{abc}(\widetilde{\mathbf{k}}) = \Delta^{g}_{a b c}(-\mathit{g}^{-1}\widetilde{\mathbf{k}})$.
The order parameters $\sigma^x$, $\tau^z\sigma^y$, $\sigma_y$ and $\tau_z\sigma_x$ transform as follows under $C_{3z}$ operation:
\begin{eqnarray}
\sigma_x &\longrightarrow &\cos(2\theta)\sigma_x - \sin(2\theta)\tau_z\sigma_y \;\nonumber\\
\tau_z\sigma_y &\longrightarrow &\sin(2\theta)\sigma_x+\cos(2\theta)\tau_z\sigma_y \;\nonumber\\
\sigma_y &\longrightarrow &\cos(2\theta)\sigma_y + \sin(2\theta)\tau_z\sigma_x \;\nonumber\\
\tau_z\sigma_x &\longrightarrow &-\sin(2\theta)\sigma_y+\cos(2\theta)\tau_z\sigma_x
\end{eqnarray}
Here $\theta = 2\pi/3$. Thus if the system preserves $C_{3z}$ symmetry, these order parameters should satisfy the following relationship:
\begin{eqnarray}
\Delta _{0x}(\widetilde{\mathbf{k}})=\cos(2\theta)\Delta _{0x}(C^{-1}_3\widetilde{\mathbf{k}})+\sin(2\theta)\Delta _{zy}(C^{-1}_3\widetilde{\mathbf{k}})\;\nonumber\\
\Delta _{zy}(\widetilde{\mathbf{k}})=-\sin(2\theta)\Delta _{0x}(C^{-1}_3\widetilde{\mathbf{k}})+\cos(2\theta)\Delta _{zy}(C^{-1}_3\widetilde{\mathbf{k}})\;\nonumber\\
\Delta _{0y}(\widetilde{\mathbf{k}})=\cos(2\theta)\Delta _{0y}(C^{-1}_3\widetilde{\mathbf{k}})-\sin(2\theta)\Delta _{zx}(C^{-1}_3\widetilde{\mathbf{k}})\;\nonumber\\
\Delta _{zx}(\widetilde{\mathbf{k}})=\sin(2\theta)\Delta _{0y}(C^{-1}_3\widetilde{\mathbf{k}})+\cos(2\theta)\Delta _{zx}(C^{-1}_3\widetilde{\mathbf{k}})\;\nonumber\\
\label{eq:order-C3z}
\end{eqnarray}
In Fig.~\ref{figss1} and Fig.~\ref{figss3} we present the distribution of the dominant order parameters in moir\'e Brillouin zone at 1/2 filling of TBMG under $U_d=0.0536\,$eV and $U_d=-0.0536\,$eV respectively ($\theta\!=\!1.25\,^{\circ}$). The dominant order parameters in the spin polarized, nematic insulator states involve the $\tau^{0,z}s^{0,z}\sigma^{x,y}$ orders, which spontaneously break $C_{3z}$ symmetry and spin rotational symmetry. It is interesting to note that the $\tau^{z}\sigma^x$ and $\sigma^{y}$ order break time-reversal symmetry $\mathcal{T}=\tau^{x}\mathcal{K}$, but preserve a ``Kramers" time-reversal symmetry introduced in Ref.~\onlinecite{zaletel-tbg-hf-prx20} $\widetilde{\mathcal{T}}=\tau^{z}\mathcal{T}$, which is the combination of valley $U(1)$ symmetry operation and the physical time-reversal operation. On the other hand, the operators $\tau^{z}\sigma^{y}$ and $\sigma^{x}$ preserve the physical time-reversal symmetry. Either the $\mathcal{T}$ or $\widetilde{\mathcal{T}}$ symmetry would guarantee that the Chern number and the net orbital magnetization of the system vanish.
\begin{figure}[htbp]
\includegraphics[width=3.5in]{opt-Udp.pdf}
\caption{~\label{figss1} The reciprocal-space distributions of main order parameters at half filling under --0.4\,V/nm field ($U_d=0.0536$\,eV) with $\epsilon$ = 9.6 and $\kappa$ = 0.005\,\AA{}$^{-1}$. The reciprocal-space coordinates are limited to -0.0644$\sim$ 0.0644\,\AA{}$^{-1}$.}
\end{figure}
\begin{figure}[htbp]
\includegraphics[width=3.5in]{opt-Udp-sigma.pdf}
\caption{~\label{figss2} The reciprocal-space distributions of main order parameter symmetrical deviations ($\Delta _{sym}-\Delta$) at half filling under --0.4\,V/nm field ($U_d=0.0536$\,eV) with $\epsilon$ = 9.6 and $\kappa$ = 0.005\,\AA{}$^{-1}$. The reciprocal-space coordinates are limited to -0.0644$\sim$ 0.0644\,\AA{}$^{-1}$.}
\end{figure}
In Fig.~\ref{figss2} and Fig.~\ref{figss4} we present the differences between the actual order parameters obtained from self-consistent Hartree-Fock calculations (denoted by $\Delta$) from the ``symmetrized" order parameters that are enforced to obey $C_{3z}$ symmetry (denoted by $\Delta_{symm}$) for $U_d=\pm 0.0536\,$eV at 1/2 filling. For ($\sigma_x$, $\tau_z\sigma_y$) and ($\sigma_y$, $\tau_z\sigma_x$) order parameters, the conditions that these order parameters obey $C_{3z}$ symmetry are given by Eqs.~(\ref{eq:order-C3z}). ($\sigma_x$, $\tau_z\sigma_y$) and ($\sigma_y$, $\tau_z\sigma_x$) form two pairs of two dimensional irreducible representations of $C_{3z}$ operation.
From Fig.~\ref{figss2} and Fig.~\ref{figss4} we can see the deviation from $C_{3z}$ symmetry can be as large as 0.1\,meV at some $\widetilde{\mathbf{k}}$ points in the moir\'e Brillouin zone; while from Fig.~\ref{figss1} and Fig.~\ref{figss3} we see that the maximal magnitudes of the dominant order parameters $\sim 0.5\rm{-}1$\,meV at half filling under both positive and negative displacement fields. Thus we conclude that the order parameters at 1/2 filling of TBMG under finite displacement fields at $\theta=1.25\,^{\circ}$ strongly break $C_{3z}$ symmetry, which give rise to the nematic insulator state as discussed in main text.
\begin{figure}[htbp]
\includegraphics[width=3.5in]{opt-Udn.pdf}
\caption{~\label{figss3} The reciprocal-space distributions of main order parameters at half filling under 0.4\,V/nm field ($U_d=-0.0536$\,eV) with $\epsilon$ = 9.6 and $\kappa$ = 0.005\,\AA{}$^{-1}$. The reciprocal-space coordinates are limited to -0.0644$\sim$ 0.0644\,\AA{}$^{-1}$.}
\end{figure}
\begin{figure}[htbp]
\includegraphics[width=3.5in]{opt-Udn-sigma.pdf}
\caption{~\label{figss4} The reciprocal-space distributions of main order parameter symmetrical deviations ($\Delta _{sym}-\Delta$) at half filling under 0.4\,V/nm field ($U_d=-0.0536$\,eV) with $\epsilon$ = 9.6 and $\kappa$ = 0.005\,\AA{}$^{-1}$. The reciprocal-space coordinates are limited to -0.0644$\sim$ 0.0644\,\AA{}$^{-1}$.}
\end{figure}
\vspace{12pt}
\begin{center}
\textbf{\large {\uppercase\expandafter{\romannumeral 6}}\ Competition between spin-polarized and valley-polarized states}
\end{center}
\vspace{12pt}
\begin{center}
\textbf{\large VIA. Quasi-degeneracy between spin-polarized and valley-polarized states}
\end{center}
\label{sec:form}
We first discuss the degeneracy between spin-polarized and valley-polarized states in the twisted multilayer graphene systems under finite displacement fields. First, we define
\begin{equation}
\lambda _{\mu \sigma m,\mu \sigma n}(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})=\sum _{\alpha \mathbf{G}}C^*_{\mu \alpha \mathbf{G}+\mathbf{Q},n\widetilde{\mathbf{k}} + \widetilde{\mathbf{q}}}C_{\mu \alpha \mathbf{G}, m\widetilde{\mathbf{k}}}
\label{eq:lambda}
\end{equation}
where $C_{\mu \alpha \mathbf{G}, m\widetilde{\mathbf{k}}}$ is the non-interacting wavefunction (see Eq.~(17)), and $\mu$, $\sigma$, and $\alpha$ refer to the valley, spin, and layer/sublattice degrees of freedom. Then the form factor can be re-written as
$\Omega ^{\mu \sigma, \mu '\sigma '}_{nm,n'm'}(\widetilde{\mathbf{k}},\widetilde{\mathbf{k}} ',\widetilde{\mathbf{q}},\mathbf{Q})=\lambda _{\mu\sigma m,\mu \sigma n}(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})\lambda ^*_{\mu ' \sigma 'n',\mu ' \sigma ' m'}(\widetilde{\mathbf{k}} '-\widetilde{\mathbf{q}},\widetilde{\mathbf{q}},\mathbf{Q}).$
It should be noted that the conduction flat band is usually isolated from other bands in the twisted multilayer graphene under finite displacement fields. Therefore, in such case of isolated flat band, the band index can be dropped, and $\lambda _{\mu\sigma,\mu\sigma}(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})$ becomes a $4\times 4$ diagonal matrix defined in the valley-spin space at this time. Thus the Coulomb interaction projected onto single conduction band becomes
\begin{equation}
\hat{H}_{single-band}=\frac{1}{N_s}\sum _{\widetilde{\mathbf{k}},\widetilde{\mathbf{k}}',\widetilde{\mathbf{q}}}\sum_{\mu\mu'\sigma\sigma'}\left(\sum_{\mathbf{Q}}V(\widetilde{\mathbf{q}}+\mathbf{Q})\lambda _{\mu\sigma ,\mu \sigma }(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})\lambda ^*_{\mu ' \sigma ',\mu ' \sigma '}(\widetilde{\mathbf{k}} '-\widetilde{\mathbf{q}},\widetilde{\mathbf{q}},\mathbf{Q})\right)\hat{c}^{\dagger}_{\mu\sigma,\widetilde{\mathbf{k}}+\widetilde{\mathbf{q}}}\hat{c}^{\dagger}_{\mu ' \sigma ',\widetilde{\mathbf{k}} '-\widetilde{\mathbf{q}}}\hat{c}_{\mu ' \sigma ' ,\widetilde{\mathbf{k}} '}\hat{c}_{\mu \sigma,\widetilde{\mathbf{k}}}
\end{equation}
Under the Hartree-Fock approximation, the interaction energy can be divided into the Hartree energy $E^H$ and the Fock energy $E^F$,
\begin{equation}
E^H=\frac{1}{2N_s}\sum_{\widetilde{\mathbf{k}} \widetilde{\mathbf{k}} '}\sum_{\mathbf{Q}}\sum_{\mu \mu '\sigma \sigma '}V(\mathbf{Q})\lambda _{\mu \sigma ,\mu\sigma}(\widetilde{\mathbf{k}} ,0,\mathbf{Q})\lambda^*_{\mu '\sigma ',\mu '\sigma '}(\widetilde{\mathbf{k}} ' ,0,\mathbf{Q})\Delta_{\mu\sigma,\mu\sigma}(\widetilde{\mathbf{k}})\Delta_{\mu '\sigma ',\mu '\sigma '}(\widetilde{\mathbf{k}} '),
\end{equation}
\begin{equation}
E^F=-\frac{1}{2N_s}\sum_{\widetilde{\mathbf{k}} \widetilde{\mathbf{q}}}\sum_{\mathbf{Q}}\sum_{\mu \mu '\sigma \sigma '}V(\widetilde{\mathbf{q}} + \mathbf{Q})\lambda _{\mu \sigma ,\mu\sigma}(\widetilde{\mathbf{k}} ,\widetilde{\mathbf{q}}, \mathbf{Q})\lambda^*_{\mu '\sigma ',\mu '\sigma '}(\widetilde{\mathbf{k}} ,\widetilde{\mathbf{q}}, \mathbf{Q})\Delta_{\mu\sigma,\mu '\sigma '}(\widetilde{\mathbf{k}}+\widetilde{\mathbf{q}})\Delta_{\mu '\sigma ',\mu \sigma}(\widetilde{\mathbf{k}}).
\end{equation}
Here the density operator $\Delta_{\mu\sigma,\mu '\sigma '}(\widetilde{\mathbf{k}})$ is defined as
\begin{equation}
\Delta_{\mu\sigma,\mu '\sigma '}(\widetilde{\mathbf{k}})=\langle \hat{c}^{\dagger}_{\mu\sigma,\widetilde{\mathbf{k}}}\hat{c}_{\mu '\sigma ',\widetilde{\mathbf{k}}} \rangle\;.
\end{equation}
At 1/2 filling and when the system is in an insulator state, the trace of the density operator at every $\widetilde{\mathbf{k}}$ point equals to 2, thus one can decompose the $4\times 4$ matrix $\hat{\Delta}(\widetilde{\mathbf{k}})$ defined in the valley-spin space as
\begin{equation}
\hat{\Delta}(\widetilde{\mathbf{k}})=(\mathbbm{1}+\hat{\Sigma}(\widetilde{\mathbf{k}}))/2\;,
\label{eq:sigma}
\end{equation}
where $\mathbbm{1}$ is the $4\times 4$ identity matrix, and $\hat{\Lambda}(\widetilde{\mathbf{k}})=\tau^{a}\, s^{b}$ ($a, b =0,x,y,z$) is a traceless $4\times 4$ matrix, which can be written as the tensor product of two Pauli matrices $\tau^{a}$ and $s^{b}$ defined in the valley and spin space respectively ($\hat{\Sigma}(\widetilde{\mathbf{k}})$ cannot be identity).
We first discuss the Hartree energy $E^H$. The Hartree energy $E^H$ can be re-written in the matrix form
\begin{equation}
E^H=\frac{1}{2N_s}\sum _{\mathbf{Q}}V(\mathbf{Q})\sum _{\widetilde{\mathbf{k}}}\rm{tr}\left[\hat{\lambda}(\widetilde{\mathbf{k}},0,\mathbf{Q})\hat{\Delta}(\widetilde{\mathbf{k}})\right]\sum _{\widetilde{\mathbf{k}} '}\rm{tr}\left[\hat{\lambda} ^{\dagger}(\widetilde{\mathbf{k}} ',0,\mathbf{Q})\hat{\Delta}(\widetilde{\mathbf{k}} ')\right]\;,
\end{equation}
where ``$\rm{tr}$" means taking the trace in the valley-spin space.
Clearly the dominant Hartree energy is $\mathbf{Q}=0$ term. In this term, the matrix $\hat{\lambda}(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}}=0,\mathbf{Q}=0)$ becomes the identity matrix in the valley-spin space. So the dominant Hartree energy is a constant for given filling, which is independent of the flavor symmetry breaking.
\begin{figure*}[!htbp]
\includegraphics[width=5in]{lambda0z-23.pdf}
\caption{~\label{figs-lambda} The reciprocal space distributions of $ |\lambda_0(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q}=0)|^2+ |\lambda_z(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q}=0)|^2$, (a)-(b) for twisted bilayer-monolayer graphene (TBMG), and (c)-(d) for twisted double bilayer graphene (TDBG). The reciprocal space distributions of $2$Im$[\lambda_0(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q}=0)\lambda^*_z(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q}=0)]$, (e)-(f) in TBMG, and (g)-(h) in TDBG system. In the subfigures (a,c,e,g), $\widetilde{\mathbf{q}}=\mathbf{G}_1/2$, and in (b,d,f,h), $\widetilde{\mathbf{q}}=\mathbf{G}_1/3+2\mathbf{G}_2/3$.}
\end{figure*}
As for the Fock energy $E^F$, it can be re-written in the matrix form as follows:
\begin{equation}
E^F=-\frac{1}{2N_s}\sum_{\widetilde{\mathbf{k}} \widetilde{\mathbf{q}}}\sum _{\mathbf{Q}} V(\widetilde{\mathbf{q}}+\mathbf{Q})\rm{tr}\left[ \hat{\Delta}(\widetilde{\mathbf{k}})\hat{\lambda}(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})\hat{\Delta}(\widetilde{\mathbf{k}}+\widetilde{\mathbf{q}})\hat{\lambda}^{\dagger}(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q}) \right]
\end{equation}
The matrix $\hat{\lambda}(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})$ is explicitly written as
\begin{equation}
\hat{\lambda}(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})=\begin{pmatrix}
\lambda _{+,+} & 0 & 0 & 0 \\
0 & \lambda _{+,+} & 0 & 0 \\
0 & 0 & \lambda _{-,-} & 0 \\
0 & 0 & 0 & \lambda _{-,-}
\end{pmatrix}
\end{equation}
in which $\mp$ refers to the $K/K'$ valley, $\lambda_{\mu,\mu}$ (with implicit wavevector dependence) is the single-band form factor as defined in Eq.~(\ref{eq:lambda}). Because spin-orbit coupling is negligibly weak in graphene, the non-interacting wavefunctions of spin-up and spin-down electrons are identical, thus $\lambda_{\mu,\mu}$ is independent of spin. This matrix $\hat{\lambda}(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})$ can be further written as
\begin{equation}
\hat{\lambda}(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})=\lambda _0(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})\mathbbm{1} +\lambda _z(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})\tau _z\;,
\end{equation}
where
\begin{align}
\lambda _0(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})&=[\lambda_{+,+}(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})+\lambda_{-,-}(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})]/2,\;\nonumber\\
\lambda _z(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})&=[\lambda_{+,+}(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})-\lambda_{-,-}(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})]/2.
\end{align}
With these notations, the Fock energy can be written as
\begin{align}
E^F=& -\frac{1}{2N_s}\sum_{\widetilde{\mathbf{k}},\widetilde{\mathbf{q}}}\sum_{\mathbf{Q}}V(\widetilde{\mathbf{q}}+\mathbf{Q})\left[ |\lambda_0(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})|^2\,\rm{tr}[\hat{\Delta}(\widetilde{\mathbf{k}})\hat{\Delta}(\widetilde{\mathbf{k}}+\widetilde{\mathbf{q}})] +\lambda_0(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})\lambda^*_z(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})\rm{tr}[\hat{\Delta}(\widetilde{\mathbf{k}})\hat{\Delta}(\widetilde{\mathbf{k}}+\widetilde{\mathbf{q}})\tau_z]\right.\;\nonumber\\
&+\left.\lambda^*_0(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})\lambda_z(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})\,\rm{tr}[\hat{\Delta}(\widetilde{\mathbf{k}})\tau_z\hat{\Delta}(\widetilde{\mathbf{k}}+\widetilde{\mathbf{q}})]+ |\lambda_z(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})|^2\,\rm{tr}[\hat{\Delta}(\widetilde{\mathbf{k}})\tau_z\hat{\Delta}(\widetilde{\mathbf{k}}+\widetilde{\mathbf{q}})\tau_z]\right]\;.
\label{fockenergy}
\end{align}
Now we focus on the symmetry-breaking ground states with zero magnetic field. The second and third term in the Eq.~(\ref{fockenergy}) either favors a strongly $\widetilde{\mathbf{k}}$-dependent IVC (intervalley coherent) ordered state with $\hat{\Sigma}(\widetilde{\mathbf{k}})=\tau_{x,y}$ ($\hat{\Sigma}(\widetilde{\mathbf{k}})$ defined in Eq.~(\ref{eq:sigma})), or favors a $\widetilde{\mathbf{k}}$-independent valley polarized state with $\hat{\Sigma}(\widetilde{\mathbf{k}})=\tau_{z}$. This is because $\rm{tr}[\hat{\Delta}(\widetilde{\mathbf{k}})\hat{\Delta}(\widetilde{\mathbf{k}}+\widetilde{\mathbf{q}})\tau_z] \neq 0$ and $\rm{tr}[\hat{\Delta}(\widetilde{\mathbf{k}})\tau_z\hat{\Delta}(\widetilde{\mathbf{k}}+\widetilde{\mathbf{q}})] \neq 0$ in two cases: (\lowercase\expandafter{\romannumeral 1}) $\hat{\Sigma}(\widetilde{\mathbf{k}}) \neq \hat{\Sigma}(\widetilde{\mathbf{k}}+\widetilde{\mathbf{q}}) \neq \tau_z$, and (\lowercase\expandafter{\romannumeral 2}) $\hat{\Sigma}(\widetilde{\mathbf{k}})=\hat{\Sigma}(\widetilde{\mathbf{k}}+\widetilde{\mathbf{q}})=\tau_z$.
However, the form factors in the second and the third terms ($\lambda^*_0(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})\lambda_z(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})$ and $\lambda_0(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})\lambda_z^*(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})$) are much smaller than those of the first term and fourth terms ($|\lambda_0(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})|^2$ and $|\lambda_z(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})|^2$) in Eq.\ref{fockenergy}. In Fig.~\ref{figs-lambda}(a)-(b) we present the calculated $|\lambda_0(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q}=0)|^2+ |\lambda_z(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q}=0)|^2$ for the conduction flat band for twisted bilayer-monolayer graphene at $\theta\!=\!1.25^{\circ}$ and $U_d\!=\!0.0536\,$eV, with $\widetilde{\mathbf{q}}$ at $M_s$ and $K_s$ points respectively; and in Fig.~\ref{figs-lambda}(e)-(f) we show $2$Im$[\lambda_0(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q}=0)\lambda^*_z(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q}=0)]$ also for TBMG system with the same choice of parameters. Similarly, in Fig.~\ref{figs-lambda}(c)-(d) we show the calculated $|\lambda_0(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q}=0)|^2+ |\lambda_z(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q}=0)|^2$ for the isolated conduction flat band of twisted double bilayer graphene with $\theta=1.28^{\circ}$ and $U_d=0.04$\,eV; whereas in Fig.~\ref{figs-lambda}(g)-(h) we present $2$Im$[\lambda_0(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q}=0)\lambda^*_z(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q}=0)]$ for TDBG with the same parameter choice. For both systems,
the average value of $ |\lambda_0(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q}=0)|^2+ |\lambda_z(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q}=0)|^2$ reaches $0.13\!\sim\!0.25$, but the average value of $2$Im$[\lambda_0(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q}=0)\lambda^*_z(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q}=0)]\sim 10^{-3}$ for $\widetilde{\mathbf{q}}$ at $K_s$ and is zero for $\widetilde{\mathbf{q}}$ at $M_s$. Therefore, we rule out the IVC states as ground-state candidates due to the small form factor Im$[\lambda_0(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q}=0)\lambda^*_z(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q}=0)]$. We do not rule out the valley polarized state at this moment, since the valley polarized state is favored by the dominant first and fourth terms in Eq.~(\ref{fockenergy}) which we explain below.
Since the form factors in the first term and fourth terms (($\lambda^*_0(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})\lambda_z(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})$ and $\lambda_0(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})\lambda_z^*(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})$)) in Eq.\ref{fockenergy} are dominating, now we only consider these two terms. First, we note that the first term in Eq.~\ref{fockenergy} favors a $\widetilde{\mathbf{k}}$-independent order parameter, because \rm{tr}$[\hat{\Delta}(\widetilde{\mathbf{k}})\hat{\Delta}(\widetilde{\mathbf{k}}+\widetilde{\mathbf{q}})]=2$ (see Eq.~(\ref{eq:sigma})) for all $\widetilde{\mathbf{k}}$ for an insulator state at 1/2 filling of conduction flat band. Thus, for all types of $\widetilde{\mathbf{k}}$-independent flavor ordered states $\hat{\Sigma}(\widetilde{\mathbf{k}})=\tau^{a} s^{b}$, the term in Eq.~(\ref{fockenergy}) contributes to the same Fock energy $-\sum_{\widetilde{\mathbf{k}} \widetilde{\mathbf{q}}}\sum_\mathbf{Q} V(\widetilde{\mathbf{q}}+\mathbf{Q})|\lambda_0(\widetilde{\mathbf{k}},\widetilde{\mathbf{q}},\mathbf{Q})|^2/N_s$.
As for the fourth term in the Eq.~\ref{fockenergy}, for convenience, we let $\hat{\Delta}(\widetilde{\mathbf{k}})=\hat{A}$, $\tau_z\hat{\Delta}(\widetilde{\mathbf{k}}+\widetilde{\mathbf{q}})\tau_z=\hat{B}$. Then we use the Cauchy-Schwarz inequality $(\rm{tr}{\hat{A}\hat{B}})^2\leq (\rm{tr} \hat{A})^2 (\rm{tr} \hat{B})^2 $, from which we obtain
\begin{equation}
\rm{tr}[\hat{\Delta}(\widetilde{\mathbf{k}})\tau_z\hat{\Delta}(\widetilde{\mathbf{k}}+\widetilde{\mathbf{q}})\tau_z]\leq\sqrt{(\rm{tr} [\hat{\Delta}(\widetilde{\mathbf{k}})])^2(\rm{tr} [\tau_z\hat{\Delta}(\widetilde{\mathbf{k}}+\widetilde{\mathbf{q}})\tau_z])^2}
\end{equation}
the equality condition is satisfied if and only if
\begin{equation}
\hat{\Delta}(\widetilde{\mathbf{k}})=\tau_z\hat{\Delta}(\widetilde{\mathbf{k}}+\widetilde{\mathbf{q}})\tau_z.
\end{equation}
We note that similar trick has been used in Ref.~\onlinecite{liu-prr21} in the analysis of ground state at charge neutrality point of TBG.
With this condition, the Fock energy contributed by the fourth term in the Eq.~\ref{fockenergy} reaches maximum magnitude. Therefore, the fourth term favors certain types of $\widetilde{\mathbf{k}}$-independent order parameters that commute with $\tau_z$. It gives us three order parameters: $\tau_z$ (valley polarized), $s_z$ (spin polarized) and $\tau_zs_z$ (valley spin locking) where we choose the spin-ordering direction to be the ``$z$" direction. We see that the spin polarized and valley polarized states are degenerate if only considers the dominant component of the Hartree and Fock energies, but the sub-leading terms may lead to slight energy difference.
This explains why we have the quasi-degeneracy between the valley polarized and spin polarized states. Inclusion of atomic Hubbard interactions would split such quasi degeneracy and favors a spin polarized state. Therefore, the ground state at 1/2 filling is spin polarized in twisted multilayer graphene systems under finite displacement fields with isolated conduction flat bands. On the other hand, applying a vertical magnetic field generates nonzero valley polarization thus favors a valley polarized state, which will be explained in detail in the following two subsections.
\vspace{12pt}
\begin{center}
\textbf{\large VIB. Orbital magnetic Zeeman effects in twisted graphene systems}
\label{sec:zeeman}
\end{center}
The effects of magnetic fields can be separated into two parts: the spin Zeeman effects and the orbital magnetic effects. The former can be trivially described by the spin Zeeman splitting, $H^{s}_{\rm{Zeeman}}=\mu_B\mathbf{s}\cdot\mathbf{B}$, where $\mathbf{s}$ represents the Pauli matrix in spin space, and $\mathbf{B}$ is the external magnetic field. The orbital magnetic effects deserve careful discussions.
First, the vertical magnetic field tends to recombine the flat bands into a series of recurring Landau levels (LLs), i.e., the Hofstadter butterfly spectra \cite{hofstadter-prb76}, which are dependent on the number of magnetic fluxes in each moir\'e primitive cell. Second, aside from the formation of LLs, the magnetic
field also induces splitting between the flat bands from the opposite valleys, because when $C_{2z}$ (and $C_{2z}\mathcal{T}$) symmetry is broken, the states in the two valleys have opposite orbital magnetizations which couple linearly to vertical magnetic fields. Such valley-contrasting orbital magnetizations are giant in twisted graphene systems because the orbital angular momenta of the electrons circulating on the moir\'e length scale are large \cite{jpliu-prx19}. For example, the orbital magnetization contributed by the flat bands of is on the order of $\pm 10\,\mu_{\textrm{B}}$ per moir\'e cell in TBMG system.
Therefore, the vertical magnetic field tends to drive the system into a valley polarized, time-reversal broken state even without the necessity of forming LLs. For weak vertical magnetic fields, the magnetic flux for each moir\'e supercell is small in the TBMG, e.g., for $B\!=\!2\,$T, the flux per supercell $\Phi/\Phi_0\!=\!0.053\!\approx\!7/132$ ($\Phi_0\!=\!h/e$), which is far from forming notable Hofstadter bands. Thus we can neglect the effects of LL quantization for $B\!\lessapprox\!2\,$T, and only consider the orbital magnetic Zeeman effect in the subspace of the two flat bands for each valley, which can be conveniently described using the orbital $g$ factor defined in the subspace of the flat bands \cite{song-zeeman-arxiv15,Koshino-orbital-prb11,ashvin-double-bilayer-nc19, wu-tdbg-arxiv20, Sun-zeeman-prb20}, the matrix element of the orbital magnetic $g$ factor in the flat-band basis is expressed as:
the matrix element is expressed as
\begin{align}
\hat{g}^{\mu}_{mm^{\prime}}(\widetilde{\mathbf{k}})=&\frac{-im_e}{2\hbar ^2}\sum _{l}\left(\frac{1}{E^{\mu}_{m\widetilde{\mathbf{k}}}-E^{\mu}_{l\widetilde{\mathbf{k}}}}+\frac{1}{E^{\mu}_{m'\widetilde{\mathbf{k}}}-E^{\mu}_{l\widetilde{\mathbf{k}}}}\right)\;\nonumber\\
&\times (\hat{v}^{x,\mu}_{ml}\hat{v}^{y,\mu}_{lm^{\prime}}-\hat{v}^{y,\mu}_{ml}\hat{v}^{x,\mu}_{lm^{\prime}}),
\label{eq:g-factor}
\end{align}
where $m, m'$ refer to the indices of the flat bands, and $l$ refers to remote band index, $E^{\mu}_{m\widetilde{\mathbf{k}}}$ represents the non-interacting flat-band energy of valley $\mu$ at moir\'e wavevector $\widetilde{\mathbf{k}}$ , and $v^{a,\mu}=\partial H(\mathbf{k})/(\hbar\partial k_a)$ ($a=x,y$) is the velocity operator for valley $\mu$.
Then the orbital magnetic Zeeman effects can be described by $(H^{\mu}_{\rm{Zeeman}})_{m m'}(\widetilde{\mathbf{k}})=\mu_B \hat{g}^{\mu}_{mm^{\prime}}(\widetilde{\mathbf{k}})\,B_z$, where $B_z$ is the $z$ component of the external magnetic field. In Fig.~\ref{figs9}(a)-(b) we show the orbital magnetic $g$ factors of the valence flat band and conduction flat band from the $K$ valley in the TBMG at $\theta=1.25\,^{\circ}$ with $U_d=0.0536\,$eV, we see that the maximal amplitude $\sim 10\rm{-}15$, and they are of opposite signs for the $K'$ valley. Similarly, in Fig.~\ref{figs9}(c)-(d) we show the orbital $g$ factor distributions of the valence flat band and conduction flat band of TBMG with $U_d\!=\!-0.0536\,$eV and $\theta\!=\!1.25\,^{\circ}$, which have maximal amplitudes around $\Gamma_s$ $\sim 15$. In Fig.~\ref{figs10}(a) and (b) we show the orbital $g$ factor distributions of the valence flat band and conduction flat band in TDBG with $U_d=0.04\,$eV and $\theta\!=\!1.28\,^{\circ}$. The orbital $g$ factor in TDBG is as large as -30 around $K_s$ point in the moir\'e for the valence flat band (Fig.~\ref{figs10}(a)), and is maximal around $\Gamma_s$ point ($\sim -35$) for the conduction flat band (Fig.~\ref{figs10}(b)).
Such giant orbital magnetic $g$ factor corresponds to a significant orbital Zeeman splitting $\sim 1\rm{-}2\,$meV between the flat bands from the $K$ and $K'$ valleys for vertical magnetic field $B_z=1\,$T, which completely dominates over LL spacing. For example, for TBMG with $U_d=0.0536\,$eV, for the Fermi velocity around 1/2 filling, $\hbar\omega_c\!=\!74\,\mu$\,eV with $B_z=1\,$T, which is much smaller than the orbital magnetic Zeeman splitting.
\begin{figure}[!htbp]
\includegraphics[width=3.5in]{figss-gfactor.pdf}
\caption{~\label{figs9} The orbital g-factors of valence flat band (a) and conduction flat band (b) under $U_d$ = 0.0536\,eV field and those of valence flat band (c) and conduction flat band (d) under $U_d$ = -0.0536\,eV field in twisted bilayer-monolayer graphene at $\theta=1.25\,^{\circ}$. The reciprocal-space coordinates are limited to -0.032$\sim$ 0.032\,\AA{}$^{-1}$. }
\end{figure}
\begin{figure}[!htbp]
\includegraphics[width=3.5in]{figss-gfactor-tdbg.pdf}
\caption{~\label{figs10} The orbital g-factors of valence flat band (a) and conduction flat band (b) under $U_d$ = 0.04\,eV field in the twisted double-bilayer graphene with $\theta=1.28\,^{\circ}$ . The reciprocal-space coordinates are limited to -0.032$\sim$ 0.032\,\AA{}$^{-1}$. }
\end{figure}
\vspace{12pt}
\begin{center}
\textbf{\large VIC. Berry-curvature correction to the density of states under vertical magnetic fields}
\end{center}
Magnetic field does not only couples to the electrons through the orbital Zeeman effect, for topological bands with nonzero Chern numbers, magnetic field also changes the density of the Chern bands. To be specific, the change of particle number per cell $\delta n$ is described by the Streda formula $\delta n = B_z \Omega_M C/\Phi_0$, where $\Omega_M$ is the area of the moir\'e primitive cell, $C$ is the Chern number of the occupied bands, and $\Phi_0=h/e$ is the flux quantum. The change of particle number in the Chern band induced by magnetic field can be well characterized by introducing a Berry-curvature correction to density of states under vertical magnetic field \cite{xiao-prl05}. Following the semi-classical treatment introduced in Ref.~\onlinecite{xiao-prl05}, when the magnetic field is present, the density matrix at $\widetilde{\mathbf{k}}$ for each spin species $\rho_{\mu n,\mu m}(\widetilde{\mathbf{k}})$ ($m$ and $n$ are band indices and $\mu$ is valley index) is multiplied by a factor of $(\mathbbm{1}+eB_z\hat{\Omega}(\widetilde{\mathbf{k}})/\hbar)$,
i.e.
\begin{equation}
\hat{\rho}(\widetilde{\mathbf{k}},B_z)=\hat{\rho}(\widetilde{\mathbf{k}},0)\cdot (\mathbbm{1}+eB_z\hat{\Omega}(\widetilde{\mathbf{k}})/\hbar)
\label{eq:rho-B}
\end{equation}
where $\mathbbm{1}$ is the identity matrix defined in the valley-band space, and $\hat{\Omega}(\widetilde{\mathbf{k}})$ is the Berry curvature matrix at $\widetilde{\mathbf{k}}$ defined in the valley-band basis, and $\hat{\rho}(\widetilde{\mathbf{k}},B_z)$ is the Berry-curvature-corrected density matrix at $\widetilde{\mathbf{k}}$ with magnetic field $B_z$. The ``$\cdot$" symbol means a matrix product. Here we only the density matrices of the valley and/or spin polarized states since the IVC states are energetically unfavored as argued in Sec.~\ref{sec:form}.
From Eq.~(\ref{eq:rho-B}) it follows that the density under magnetic field $B_z$ is
\begin{align}
&\rho(B_z)\;\nonumber\\
=&\frac{1}{N_k \Omega_M}\sum_{\widetilde{\mathbf{k}}}\rm{tr}[\hat{\rho}(\widetilde{\mathbf{k}},0)\cdot(\mathbbm{1}+eB_z\hat{\Omega}(\widetilde{\mathbf{k}})/\hbar)]\;\nonumber\\
=&\rho(0)+ \frac{e}{\hbar}B_z\,\frac{1}{N_k\Omega_M}\sum_{\widetilde{\mathbf{k}}}\rm{tr}[\hat{\rho}(\widetilde{\mathbf{k}},0)\hat{\Omega}(\widetilde{\mathbf{k}})]\;\nonumber\\
=&\rho(0)+\frac{e B_z}{h} C
\end{align}
where $\rho(0)$ is the density with zero magnetic field, and $C$ is the Chern number of the occupied bands, which is expressed as
\begin{align}
C&=\frac{1}{2\pi}\int d^2\widetilde{\mathbf{k}} \,\rm{tr}[\hat{\Omega}(\widetilde{\mathbf{k}}) \cdot \hat{\rho}(\widetilde{\mathbf{k}})]\;\nonumber\\
&=\frac{1}{2\pi}\sum_{\widetilde{\mathbf{k}}} \frac{(2\pi)^2}{N_k\Omega _M}\, \rm{tr}[\hat{\Omega}(\widetilde{\mathbf{k}}) \cdot \hat{\rho}(\widetilde{\mathbf{k}})]\;\nonumber\\
&=\frac{2\pi}{\Omega _M N_k}\sum_{\widetilde{\mathbf{k}}} \,\rm{tr}[\hat{\Omega}(\widetilde{\mathbf{k}}) \cdot \hat{\rho}(\widetilde{\mathbf{k}})].
\end{align}
Therefore, with the Berry-corrected density matrix as given in Eq.~(\ref{eq:rho-B}), Streda formula $\delta\rho(B_z)=\rho(B_z)-\rho(0)=eB_z C/h$ immediately follows. All the self-consistent Hartree-Fock calculations under finite magnetic fields reported in main text are performed with such a Berry-curvature correction to density operator. Since the Chern numbers of the flat bands for the $K$ and $K'$ valleys are opposite, the field-induced change of densities for the $K$ and $K'$ valleys have opposite sign, thus favors a valley polarized state over a spin polarized state.
In particular, if the valley polarized state around 1/2 filling has a Chern number of the same sign as the magnetic field, it gains more Fock energy and is more easily to open up a gap and becomes a Chern-insulator state.
Our calculations indicate that such effects is even more dramatic than the valley polarization induced by the orbital Zeeman splitting. The results presented in Fig. 4 of main text are obtained through self-consistent Hartree-Fock calculations including both orbital magnetic Zeeman effects and the Berry-curvature correction to the density of Chern bands. Since the charge density would be changed under magnetic fields for states with nonzero Chern numbers, we have performed self-consistent Hartree-Fock calculations at three filling factors around filling 2: (a) $n(C\!=\!0)=2$, (b) $n(C\!=\!2)=2+2\Omega_M e B/h$, and (c) $n(C\!=\!-2)=2-2\Omega_M e B/h$, where $\Omega_M$ is the area of the moir\'e primitive cell. For each magnetic field $B$, we have performed self-consistent Hartree-Fock calculations for both spin polarized and valley polarized for all the three filling factors $n(C\!=\!0$), $n(C\!=\!2)$, and $n(C\!=\!-2)$, and compare the energies of the different symmetry-breaking states for each filling with increasing magnetic field.
\vspace{12pt}
\begin{center}
\textbf{\large {\uppercase\expandafter{\romannumeral 7}}\ More results of Hartree-Fock calculations for TBMG and TDBG systems}
\end{center}
\begin{figure}[htbp]
\includegraphics[width=3.5in]{tdbg-phase-band.pdf}
\caption{~\label{figssHF} The HF energy bands of twisted double bilayer graphene with $U_d\!=\!0.04$\,eV at 1/2 filling ($\epsilon\!=\!9.6$ and $\kappa\!=\!0.005$\,\AA{}$^{-1}$). The blue and red lines represent the energy bands from different valleys.}
\end{figure}
\vspace{12pt}
\begin{center}
\textbf{\large VIIA. Correlated insulators at 1/2 filling}
\end{center}
In this section we present more results for the Hartree-Fock calculations of TBMG and TDBG systems. First we present the Hartree-Fock phase diagram of TDBG at 1/2 filling with $U_d\!=\!0.04\,$eV and $\theta\!=\!1.28\,^{\circ}$ in Fig.~\ref{figssHF}(a), where $\epsilon$ is the background dielectric constant, and $\kappa$ is the inverse screening length. The color coding indicates the Chern number of the Hartree-Fock ground states. We see that in most of the parameter space the system stays in a correlated insulator state with Chern number zero. A more detailed analysis reveals that these states are spin polarized, $C_{3z}$-breaking, and quantum valley Hall insulator states with valley Chern numbers $\pm 2$. The nonzero valley Chern numbers also result from the nontrivial band topology as shown in Fig.~\ref{figstdbg}(a)-(b): the Chern numbers ($C$) for the valence flat band and the conduction flat band from the $K^{\prime}$ valley are $+2$ and $-2$ for respectively for $U_d\!=\!0.04\,$eV and $\theta\!=\!1.28\,^{\circ}$. As a result, in a spin polarized state at 1/2 filling with six of the eight flat bands being occupied, there would be one $C=-2$ ($C=2$) conduction flat band from the $K'$ ($K$) valley being unoccupied, giving rise to nonzero valley Chern numbers $\pm 2$ of the occupied bands. Such spin polarized, nematic, and quantum valley Hall insulator states are also
characterized by the order parameters $\tau^{0,z}s^{0,z}\sigma^{x,y}$, which are similar to those found at 1/2 filling of TBMG. The energy bands of the Hartree-Fock ground state at 1/2 filling of TDBG under $U_d\!=\!0.04$\,eV at $\theta\!=\!1.28\,^{\circ}$ are shown in Fig.~\ref{figssHF}(b), with the dielectric constant $\epsilon=9.6$, and the inverse screening length $\kappa=0.005\,\mbox{\normalfont\AA}^{-1}$. We see that the band gaps are on the order of 10\,meV.
In Fig.~\ref{figs2}(a)-(b) we show the indirect gaps of the Hartree-Fock ground states in the parameter space of $\epsilon$ and $\kappa$ at 1/2 filling of the TBMG system with $U_d=\mp0.0536\,$eV at $\theta=1.25\,^{\circ}$. This figure should be compared with Fig.~2 of the main text which shows the Chern numbers of the gapped states. Moreover, we note that the $C\!=\!-1$ phases in Fig.~2(a) and (b) of main text are states with one $C\!=\!2$ band and one $C\!=\!-1$ band being occupied from the majority spin of the $K'$ valley, and two $C\!=\!-2$ bands and two $C\!=\!1$ bands being occupied from both spins of the $K$ valley. Such a state is a valley- and spin-polarized insulator state at 1/2 filling, and the resulted Chern number -1 is inconsistent with the experimental observation with zero Chern number \cite{young-monobi-nature20}.
\begin{figure}[!htbp]
\includegraphics[width=3.5in]{figss2.pdf}
\caption{~\label{figs2} The indirect gaps of the Hartree-Fock ground states including atomic Hubbard interactions at 1/2 filling of twisted bilayer-monolayer graphene at $\theta=1.25\,^{\circ}$: (a) $U_d=-0.0536\,$eV, and (b) $U_d=0.0536\,$eV.}
\end{figure}
In Fig.~\ref{figs3} (a) and (b) we show the Chern numbers of the Hartree-Fock ground states at $U_d\!=\!-0.0402\,$eV ($D\!=\!0.3\,$V/nm) and $U_d\!=\!-0.067\,$eV ($D\!=\!0.5\,$V/nm) at 1/2 filling. The Chern-number-zero states in Fig.~\ref{figs3} are the same state as that explained in the main text: they are spin polarized, nematic states stabilized by atomic Hubbard interactions. Such zero-Chern-number states are more robust for larger displacement fields as they occupy larger area in the phase diagram shown in Fig.~\ref{figs3}(b); while such a zero-Chern-number state is less robust for weaker displacement fields, which only survives for relatively weak interactions as shown in Fig.~\ref{figs3}(a). The blank in Fig.~\ref{figs3}(a) indicates that at these points the Hartree-Fock ground states are metallic such that Chern numbers of occupied states are ill defined.
\begin{figure}[!htbp]
\includegraphics[width=3.5in]{figss3.pdf}
\caption{~\label{figs3} The Chern numbers and indirect gaps of the Hartree-Fock (HF) ground states at 1/2 filling of TBMG at $\theta=1.25\,^{\circ}$ under different displacement fields: (a) Chern numbers of the HF ground states, $D=0.3\,$V/nm ($U_d=-0.0402\,$eV), and (b) Chern numbers of the HF ground states, $D=0.5\,$V/nm ($U_d=-0.067\,$eV). }
\end{figure}
\vspace{12pt}
\begin{center}
\textbf{\large VIIB. Including more active bands into the Hartree-Fock calculations}
\end{center}
To verity our conclusion, we also project the interactions onto six active bands (including two flat bands and four remote bands) for each valley each spin in the TBMG, and perform Hartree-Fock calculations at 1/2 filling of the flat bands with $U_d\!=\!0.0536\,$eV and $\theta\!=\!1.25\,^{\circ}$.
We find that the ground state is still the spin polarized, nematic, and quantum valley Hall insulator state, which is consistent with the calculations with interactions only projected to the two flat bands. The energy difference between the VP and SP states at the half filling for such six-band (per spin per valley) Hartree-Fock calculations are shown in the Fig.~\ref{nflat3} (a), from which we can see that the energy difference becomes larger while the on-site Hubbard interaction $U_0$ is increasing. This is consistent with the results obtained from the calculations with interactions only projected to the two flat bands.
\begin{figure}[!htbp]
\includegraphics[width=3.5in]{Nflat3.pdf}
\caption{~\label{nflat3} (a) The energy difference between VP and SP states at the half filling in the TBMG with six-band-projected interactions. (b) The Hartree-Fock energy bands with $\epsilon$ = 9.6 and $\kappa$ = 0.005\,\AA{}$^{-1}$. }
\end{figure}
\vspace{12pt}
\begin{center}
\textbf{\large VIIC. Quantum anomalous Hall states at 1/4 and 3/4 fillings}
\end{center}
Now we discuss the quantum anomalous Hall states at 1/4 filling and 3/4 filling of TBMG. In Fig.~3(a) and (b) of the main text we have presented the Chern numbers of the Hartree-Fock ground states at 1/4 filling of TBMG with $D\!=\!0.5\,$V/nm ($U_d\!=\!-0.067\,$eV) and at 1/4 filling with $D\!=\!0.3\,$V/m ($U_d=-0.0402\,$eV), from which we see that in most regions of the phase diagrams, the system stays in the $\vert C\vert=2$ quantum anomalous Hall states. Here In Fig.~\ref{figs6} (a)-(b) we show the indirect gaps of the Hartree-Fock ground states at 1/4 filling with $D\!=\!0.5\,$V/nm (Fig.~\ref{figs6}(a)) and at 3/4 filling with $D\!=\!0.3\,$V/nm (Fig.~\ref{figs6}(b)).
\begin{figure}[!htbp]
\includegraphics[width=3.5in]{figss6-new.pdf}
\caption{~\label{figs6} The Hartree-Fock phase diagram of Chern number (a) and bandgap (b) at 1/4 filling under 0.5\,V/nm field ($U_d = -0.0670$\,eV) for twisted bilayer-monolayer graphene at $\theta=1.25\,^{\circ}$. The ground-state phase diagram of Chern number (c) and bandgap (d) at 3/4 filling under 0.3\,V/nm field ($U_d = -0.0402$\,eV).}
\end{figure}
It is interesting to note that the valley Chern numbers of the valence flat band and conduction flat band are interchanged under opposite displacement fields (see Fig.~1(c)-(d)) in the main text): when $U_d>0$, for $\theta\gtrapprox 1.05\,^{\circ}$, the Chern numbers of the conduction (valence) flat band from the $K'$ valley is -1 (2); for $U_d<0$, the Chern numbers of the conduction (valence) flat band from the $K'$ valley becomes 2 (-1). Therefore, when $U_d<0$ ($D>0$ in our definition), the system favors a spin and valley polarized state with one conduction flat band with $\vert C\vert=2$ being occupied (unoccupied) at 1/4 (3/4) filling, which gives rise to the $\vert C\vert=2$ quantum anomalous Hall effects observed in experiments \cite{young-monobi-nature20}.
Following the above argument, it is naturally expected that when the displacement field is flipped, i.e., when $U_d>0$ ($D<0$), the ground state at 1/4 (3/4) filling would be a valley and spin polarized state with $\vert C\vert=1$, since the Chern number of the conduction flat band of the $K$ valley has been changed from $2$ to $-1$ due to the flip of displacement field. This has been verified by Hartree-Fock calculations at 1/4 filling with $D=-0.5$\,V/nm and 3/4 filling with $D=-0.3$\,V/nm as shown by the calculated Chern numbers of the Hartree-Fock ground states in Fig.~\ref{figs8}(a) and (c) respectively. We see tha the $C=1$ quantum anomalous Hall state is extremely robust at 3/4 filling with $D=-0.3\,$V/nm. Basically the system stays at the $C=1$ QAH state for the entire parameter space we have explored. The $C=1$ QAH state survives for weaker interactions at 1/4 filling with $D=-0.5\,$V/nm ($U_d=0.0536\,$eV) as shown in Fig.~\ref{figs6}(a).
\begin{figure}[!htbp]
\includegraphics[width=3.5in]{figss8.pdf}
\caption{~\label{figs8} The ground-state phase diagram of Chern number (a) and bandgap (b) at 1/4 filling under -0.5\,V/nm field ($U_d = 0.0670$\,eV) for twisted bilayer-monolayer graphene at $\theta=1.25\,^{\circ}$. The ground-state phase diagram of Chern number (c) and bandgap (d) at 3/4 filling under -0.3\,V/nm field ($U_d = 0.0402$\,eV).}
\end{figure}
\vspace{12pt}
\begin{center}
\textbf{\large {\uppercase\expandafter{\romannumeral 8}}\ Hartree-Fock calculations for hBN-aligned twisted bilayer graphene system}
\end{center}
\begin{figure}[!htbp]
\includegraphics[width=3.5in]{tbg-BN.pdf}
\caption{~\label{fig-tbg-BN} (a) The non-interacting energy bands of hBN-aligned twisted bilayer graphene. The solid (dashed) lines represent the energy bands of K (K$^{\prime}$) valley. (b) The Hartree-Fock phase diagram of hBN-aligned twisted bilayer graphene at the half filling. (c) The Hartree-Fock energy bands at the half filling in the hBN-aligned twisted bilayer graphene. (d) The nematic charge density of spin-polarized phase at the half filling.}
\end{figure}
Now we consider the hBN-aligned magic-angle twisted bilayer graphene (TBG) system where we assume the aligned hBN substrate introduces a staggered sublattice potential $\approx\!15\,$meV to the bottom graphene layer. The staggered sublattice potential breaks the $C_{2z}$ symmetry, thus opens a gap at the Dirac point. The non-interacting energy bands of hBN-aligned TBG system are present in the Fig.~\ref{fig-tbg-BN} (a). We can see that a gap $\sim 4\,$meV opens up at the Dirac points.
We present Hartree-Fock phase diagram of this system at half filling in the Fig.~\ref{fig-tbg-BN} (b). Similar to TBMG and TDBG, there are two degenerate states of VP and SP state at the half filling if we only consider the inter-site Coulomb interaction. However, further calculations including both intersite Coulomb interaction and on-site Hubbard interaction reveal that the SP state becomes the only ground state with zero Chern number as shown in Fig.~\ref{fig-tbg-BN} (b). Here, the atomic Hubbard interactions lowers the energy of the SP state by about 0.6\,meV per electron. The order parameters of the correlated insulator state at 1/2 filling of hBN-aligned TBG are also $\tau^{0,z}s^{0,z}\sigma^{x,y}$, which are qualitatively the same as those found in TBMG and TDBG. Similarly, the local charge density of flat bands at the half filling as shown in Fig.~\ref{fig-tbg-BN} (d) also exhibits the nematicity. We also show the Hartree-Fock energy bands at the half filling with $\epsilon$ = 9.6 and $\kappa$ = 0.005\,\AA{}$^{-1}$ in Fig.~\ref{fig-tbg-BN} (c).
\vspace{12pt}
\begin{center}
\textbf{\large {\uppercase\expandafter{\romannumeral 9}}\ Hartree-Fock calculations for other twisted multilayer graphene systems}
\end{center}
\begin{figure}[!htbp]
\includegraphics[width=3.5in]{tmg-3and.pdf}
\caption{~\label{fig3and} The Hartree-Fock energy bands at the half filling: (a) twisted $(3+1)$-layer graphene with $\theta\!=\!1.4^{\circ}$, and $U_d=-0.06\,$eV, and (b) twisted $(3+2)$-layer graphene with $\theta\!=\!1.28\,^{\circ}$ and $U_d=0.04\,$eV. The local charge densities of twisted $(3+1)$-layer graphene (c) and twisted $(3+2)$-layer graphene (d) are also present.}
\end{figure}
In this section, we present the Hartree-Fock energy bands of other twisted multilayer graphene system, for example, the twisted $(3+1)$-layer graphene with $\theta\!=\!1.4^{\circ}$, and $U_d=-0.06\,$eV, and twisted $(3+2)$-layer graphene with $\theta\!=\!1.28\,^{\circ}$ and $U_d=0.04\,$eV in Fig.~\ref{fig3and} (a,b).
The ground states at half filling in these systems are also spin-polarized (SP) states which are stabilized by atomic Hubbard interactions. The inclusion of atomic Hubbard interactions lowers the energy of the SP state by about 0.5\,meV per electron in the twisted $(3+1)$-layer graphene or 0.4\,meV per electron in the twisted $(3+2)$-layer graphene. The order parameters are $\tau^{0,z}s^{0,z}\sigma^{x,y}$, which are same with those at 1/2 filling of the TBMG and TDBG systems. Because the distributions of these order parameters in the moir\'e Brillouin zone break $C_{3z}$ symmetry, the local charge density contributed by the low-energy bands at half filling also exhibits nematicity as shown in Fig.~\ref{fig3and} (d).
\end{document}
|
\section{Intersection stringy E-polynomial}\label{sec:intersectionstringyE-pol}
The intersection cohomology of a complex variety $X$ with compact support, middle perversity and rational coefficients is denoted by $IH^*_c(X)$. Recall that $IH^*_c(X)$ carries a canonical mixed Hodge structure, and so we can define the intersection E-polynomial of $X$ as
\begin{equation}\label{defn:IEpol}
IE(X) \coloneqq \sum_{r,s,d}(-1)^d \dim ( \operatorname{Gr}^W_{r+s} IH^d_{c}(X, \mathbb{C}))^{r,s} u^rv^s.
\end{equation}
Suppose that $X$ is endowed with the action of a finite abelian group $\Gamma$, and denote the group of characters of $\Gamma$ by $\hat{\Gamma}$. The intersection cohomology of $X$ decomposes under the action of $\Gamma$ into isotypic components:
\[IH^*_c(X)= \bigoplus_{\kappa \in \hat{\Gamma}} IH^*(X)_{\kappa}.\]
Then, if we pose
\begin{align*}
IE(X)_{\kappa} & \coloneqq \sum_{r,s,d}(-1)^d \dim ( \operatorname{Gr}^W_{r+s} IH^d_{c}(X, \mathbb{C})_{\kappa})^{r,s} u^rv^s,
\end{align*}
we obtain $IE(X)= \sum_{\kappa \in \hat{\Gamma}} IE(X)_{\kappa}$.
Define also the intersection stringy E-polynomial by
\[IE_{\mathrm{st}}(X) \coloneqq \sum_{\gamma \in \Gamma} IE(X_{\gamma}/\Gamma; u,v)(uv)^{F(\gamma)},\]
where
\begin{itemize}
\item $X_{\gamma}$ is the fixed-point set of $\gamma \in \Gamma$.
\item $F(\gamma)$ is the Fermionic shift, defined as $F(\gamma)=\sum_j w_j$, where $\gamma$ acts on the normal bundle of $X_{\gamma}$ in $X$ with eigenvalues $e^{2\pi i w_j}$ with $w_j \in (0,1)$.
\end{itemize}
\section{Topological mirror symmetry}
Let $\Gamma \coloneqq \mathrm{Pic}^0(C)[r]\simeq (\mathbb{Z}/r\mathbb{Z})^{2g}$ be the group of $r$-torsion line bundles over the compact Riemann surface $C$ of genus $g$, endowed with the canonical flat connection. The group $\Gamma$ acts by tensorisation on ${M}_{\mathrm{Dol}}^d(C, \operatorname{SL}_r)$ and $M_{\mathrm{DR}}^d(C, \operatorname{SL}_r)$. Via the non-abelian Hodge correspondence, the action corresponds to the algebraic action of the characters
$\Gamma \simeq \operatorname{Hom}(\pi_1(C), \pm 1)$ which acts on $M_{\mathrm{B}}^d(C, \operatorname{SL}_r)$ by multiplication. The quotient of $M^d(C, \operatorname{SL}_r)$ by the action of $\Gamma$ is isomorphic to $M^d(C, \operatorname{PGL}_r)$.
We identify $w\colon \Gamma \to \hat{\Gamma}$ through Poincar\'{e} duality (also known as Weil pairing)
\[\Gamma \times {\Gamma} \simeq H_1(C, \mathbb{Z}/r\mathbb{Z})\times H_1(C, \mathbb{Z}/r\mathbb{Z}) \to \mathbb{Z}/r\mathbb{Z}.\]
\begin{conj}[Topological mirror symmetry in degree zero]\label{conj:topmirr} For $\kappa \in \hat{\Gamma}$ we have
\begin{equation}\label{eq:refinedtopologicalmirrorsymmetry}
IE(M(C, \operatorname{SL}_r))_{\kappa} =IE(M(C, \operatorname{SL}_r)_{\gamma}/\Gamma; u,v)(uv)^{F(\gamma)}
\end{equation}
where $\gamma = w(\kappa)$. In particular, we obtain
\[IE(M(C, \operatorname{SL}_r))=IE_{\mathrm{st}}(M(C, \operatorname{PGL}_r)).\]
\end{conj}
\begin{thm}[\cref{mainthm}]\label{thm:topologicalmirrorsymmetryrank2}
\cref{conj:topmirr} holds for $r=2$.
\end{thm}
\begin{proof}
Without loss of generality we can suppose $\gamma\neq 0$, or equivalently $\kappa \neq 1$. Indeed, \[IE(M(C, \operatorname{SL}_r))_{1} =IE(M(C, \operatorname{SL}_r)/\Gamma; u,v);\]
see for instance the proof of \cite[Proposition 3]{GottscheSoergel93}.
For any $\gamma \in \Gamma \setminus \{0\}$, we have an associated 2-torsion line bundle $L_{\gamma}$. Consider the \'{e}tale double cover $\pi_\gamma\colon C_{\gamma} \to C$ consisting of the square root of a non-zero section of $L_{\gamma}^{\otimes 2}\simeq \mathcal{O}_C$ in the total space of $L_{\gamma}$, and let $\iota$ be its deck transformation.
For any $L \in M(C_{\gamma}, \operatorname{GL}_1)$, the rank-two vector bundle $L \oplus \iota^* L$ is a $\iota$-invariant object in $M(C_{\gamma}, \operatorname{GL}_2)$, which descends to an object $L_{\iota} \in M(C, \operatorname{GL}_2)$. Hence, the pushforward morphism
\begin{align*}
\pi_{\gamma, *}\colon M(C_{\gamma}, \operatorname{GL}_1) \to M(C, \operatorname{GL}_2), \quad L \mapsto L_{\iota},
\end{align*}
descends to a $\Gamma$-invariant embedding
\[j\colon M(C_{\gamma}, \operatorname{GL}_1)/\mathbb{Z}/2\mathbb{Z} \hookrightarrow M(C, \operatorname{GL}_2).\]
The determinant map $\det\pi_{\gamma, *}$ can be identified with the norm map
\begin{align*}
\mathrm{Nm}_{C_{\gamma}/C}\colon M(C_{\gamma}, \operatorname{GL}_1)/\mathbb{Z}/2\mathbb{Z} \to M(C, \operatorname{GL}_1), \quad L \mapsto L\otimes \iota^*L,
\end{align*}
Therefore, the fixed-point set $M(C, \operatorname{SL}_2)_{\gamma}$ admits the following geometric characterization:
\[M(C, \operatorname{SL}_2)_{\gamma} = \mathrm{Im}j \cap M(C, \operatorname{SL}_2) \simeq \ker \mathrm{Nm}_{C_{\gamma}/C}^{\circ},\]
where the last term is the connected component of $\mathrm{Nm}_{C_{\gamma}/C}^{-1}(\mathcal{O}_C)$ containing $\mathcal{O}_{C_{\gamma}}$.
On the Dolbeault side, ${M}_{\mathrm{Dol}}(C, \operatorname{SL}_2)_{\gamma}$ is isomorphic to the quotient by $\mathbb{Z}/2\mathbb{Z}$ of the cotangent bundle of an abelian variety of dimension $g-1$, as ${M}_{\mathrm{Dol}}(C_{\gamma}, \operatorname{GL}_1)$ is isomorphic to $T^*\operatorname{Pic}^0(C_{\gamma})$; see also the proof of \cref{thm:perversetopologicalmirrorsymmetryrank2}. On the Betti side, we have
\[M_{\mathrm{B}}(C_{\gamma}, \operatorname{GL}_1) \simeq (\mathbb{C}^*)^{4g-2}, \text{ and so } M_{\mathrm{B}}(C, \operatorname{SL}_2)_{\gamma} \simeq (\mathbb{C}^*)^{2g-2}/{\mathbb{Z}/2\mathbb{Z}}.\]
The involution defining the $\mathbb{Z}/2\mathbb{Z}$-quotient is the inverse of the group law.
Since the $\Gamma$-module $IH^*_c(M(C, \operatorname{SL}_2))$ is a direct sums of copies of the trivial and regular representations by \cite[Remark 4.4]{Mauri20}, we have
\[IE(M(C, \operatorname{SL}_2))_{\kappa} = IE(M(C, \operatorname{SL}_2))_{\kappa'}\]
for any $\kappa, \kappa' \in \hat{\Gamma}\setminus \{1\}$. Thanks to \cite[Corollary 1.11, Equations (23) and (25)]{Mauri20} we have
\begin{align}
IE({M}_{\mathrm{Dol}}(C, \operatorname{SL}_2))_{\kappa} = & \frac{1}{2} (uv)^{3g-3} ((u+1)^{g-1}(v+1)^{g-1} +(u-1)^{g-1}(v-1)^{g-1}) \nonumber\\
= & IE({M}_{\mathrm{Dol}}(C, \operatorname{SL}_2)_{\gamma}/\Gamma)(uv)^{2g-2},\nonumber\\
IE(M_{\mathrm{B}}(C, \operatorname{SL}_2))_{\kappa} = & \frac{1}{2} (uv)^{2g-2} ((uv+1)^{2g-2}+(uv-1)^{2g-2}) \label{eq:intersectioncohomokappa}\\
= & IE((\mathbb{C}^*)^{2g-2}/{\mathbb{Z}/2\mathbb{Z}})(uv)^{2g-2} \nonumber\\
= & IE(M_{\mathrm{B}}(C, \operatorname{SL}_2)_{\gamma}/\Gamma)(uv)^{2g-2}. \nonumber
\end{align}
Note that the Fermionic shift $F(\gamma)$ equals half of the codimension of ${M}_{\mathrm{Dol}}(C, \operatorname{SL}_2)_{\gamma}$ in ${M}_{\mathrm{Dol}}(C, \operatorname{SL}_2)$, since $\gamma$ is an involution. Hence, for $\gamma \neq 0$ we have indeed
\[F(\gamma)= \frac{1}{2}\operatorname{codim} {M}_{\mathrm{Dol}}(C, \operatorname{SL}_2)_{\gamma} = 2g-2.\]
Finally, the same argument of \cite[\S 6]{HauselThaddeus03}, together with \cite[Theorem 3.2]{FelisettiMauri2020}, implies that Conjecture \eqref{eq:refinedtopologicalmirrorsymmetry} for the Dolbeault moduli spaces yields \eqref{eq:refinedtopologicalmirrorsymmetry} for the de Rham moduli spaces.
\end{proof}
\begin{rmk}[Failure of topological mirror symmetry for ordinary cohomology] In general the equality \eqref{eq:refinedtopologicalmirrorsymmetry} fails for ordinary cohomology. For instance, for $\kappa \neq 1$, $\gamma=w(\kappa)$ and $q=uv$ we have
\begin{align*}
E(M_B(C, \operatorname{SL}_2))_{\kappa} & =\frac{1}{2}q^{2g-2}((q+1)^{2g-2}+(q-1)^{2g-2}-2)\\
& \neq \frac{1}{2}q^{2g-2}((q+1)^{2g-2}+(q-1)^{2g-2})= E(M_B(C, \operatorname{SL}_2)_{\gamma}/\Gamma)q^{F(\gamma)},
\end{align*}
where the first equality follows from \cite[Theorem 2]{MartinezMunoz16} or \cite[Theorem 1.3]{BaragliaHekmati17}, together with \cite[Remark 4.3]{Mauri20}, while the last equality comes from \eqref{eq:intersectioncohomokappa}, since $M_B(C, \operatorname{SL}_2)_{\gamma}/\Gamma$ has only quotient singularities. This shows that there is a non-negligible contribution of the singularity of $M(C, \operatorname{SL}_r)$ to the agreement \eqref{eq:refinedtopologicalmirrorsymmetry} of Hodge numbers.
\end{rmk}
\begin{rmk}
The proof of \cref{thm:topologicalmirrorsymmetryrank2} relies on the computation of $\sum_{\kappa \neq 1} IE(\newline M(C, \operatorname{SL}_2))_{\kappa} $ in \cite[Corollary 1.11]{Mauri20}, and ultimately on the explicit construction of a desingularization of $M(C, \operatorname{SL}_2)$ in \cite[\S 3]{Mauri20}. To the best of the author's knowledge, this is not available in higher rank, and so it is unclear if the arguments above extend in higher rank. However, remarkable progress in this direction have been made in \cite{MaulikShen2020I} and \cite{MaulikShen2020II}.
\end{rmk}
\color{red}
\color{black}
\section{Perverse topological mirror symmetry}
The intersection cohomology of ${M}_{\mathrm{Dol}}(C, \operatorname{SL}_r)$ and ${M}_{\mathrm{Dol}}(C, \operatorname{SL}_2)_{\gamma}$ are filtered by the weight filtration $W$ of Deligne's mixed Hodge structure, and by the perverse filtration $P$ associated to the Hitchin fibrations
\begin{align*}
\chi \colon {M}_{\mathrm{Dol}}(C, \operatorname{SL}_r) & \to \Lambda \coloneqq \bigoplus^n_{i=2} H^0(C, K^{\otimes i}_C) \\
\chi_{\gamma}\coloneqq \chi|_{{M}_{\mathrm{Dol}}(C, \operatorname{SL}_r)_{\gamma}} \colon {M}_{\mathrm{Dol}}(C, \operatorname{SL}_r)_{\gamma} & \to \Lambda_{\gamma}\coloneqq \mathrm{Im}(\chi|_{{M}_{\mathrm{Dol}}(C, \operatorname{SL}_r)_{\gamma}}) \subseteq \Lambda,
\end{align*}
which assigns to the Higgs bundle $(E, \phi)$ the characteristic polynomial of $\phi$; see \cite[\S 2.2]{Mauri20} for a brief account on the perverse filtration.
Define the perverse intersection E-polynomial
\[PIE({M}_{\mathrm{Dol}}(C, \operatorname{SL}_r);u,v,q)\coloneqq \sum_{r,s,d}(-1)^d \dim ( \operatorname{Gr}^W_{r+s} \operatorname{Gr}^P_{k}IH^d_{c}(X, \mathbb{C}))^{r,s} u^rv^sq^k\]
and the stringy perverse intersection E-polynomial
\[PIE_{\mathrm{st}}({M}_{\mathrm{Dol}}(C, \operatorname{PGL}_r); u,v,q)\coloneqq \sum_{\gamma \in \Gamma} PIE({M}_{\mathrm{Dol}}(C, \operatorname{SL}_r)_{\gamma}/\Gamma; u,v)(uvq)^{F(\gamma)}.\]
By Definition \eqref{defn:IEpol} and the last paragraph of the proof of \cref{thm:topologicalmirrorsymmetryrank2}, we have
\[PIE({M}_{\mathrm{Dol}}(C, \operatorname{SL}_r);u,v,1) = IE({M}_{\mathrm{Dol}}(C, \operatorname{SL}_r);u,v)=IE(M_{\mathrm{DR}}(C, \operatorname{SL}_r);u,v).\]
Further, Relative Hard Lefschetz \cite[Theorem 2.1.1]{deCataldoMigliorini05} implies
\[PIE({M}_{\mathrm{Dol}}(C, \operatorname{SL}_r);u,v,q)=(uvq)^{\dim}PIE\bigg({M}_{\mathrm{Dol}}(C, \operatorname{SL}_r);u,v,\frac{1}{uvq}\bigg),\]
where $\dim = 2(r^2-1)(g-1)$.
We conjecture the exchange of the perverse Hodge numbers.
\begin{conj}[Perverse topological mirror symmetry in degree zero]\label{conj:pervtopmirr}
\[PIE({M}_{\mathrm{Dol}}(C, \operatorname{SL}_r);u,v,q)=(uvq)^{\dim}PIE_{\mathrm{str}}\bigg({M}_{\mathrm{Dol}}(C, \operatorname{PGL}_r);u, v, \frac{1}{uvq}\bigg).\]
\end{conj}
For $q=1$, \cref{conj:pervtopmirr} specialises to
\[IE({M}_{\mathrm{Dol}}(C, \operatorname{SL}_r)) = IE_{\mathrm{st}}({M}_{\mathrm{Dol}}(C, \operatorname{PGL}_r)).\]
Further, the PI=WI conjecture \cite[Question 4.1.7]{deCataldoMaulik2018} would imply
\[PIE({M}_{\mathrm{Dol}}(C, \operatorname{SL}_r);1,1,q) = IE(M_{\mathrm{B}}(C, \operatorname{SL}_r); q),\]
and together with \cref{conj:pervtopmirr} it would give
\[IE(M_{\mathrm{B}}(C, \operatorname{SL}_r)) = IE_{\mathrm{st}}(M_{\mathrm{B}}(C, \operatorname{PGL}_r)).\]
\begin{thm}\label{thm:perversetopologicalmirrorsymmetryrank2}
\cref{conj:pervtopmirr} holds for $r=2$.
\end{thm}
\begin{proof}
By Relative Hard Lefschetz, it is enough to show
\begin{equation*}\label{eq:PIEPIESTR}
PIE({M}_{\mathrm{Dol}}(C, \operatorname{SL}_2))_{\kappa} = PIE({M}_{\mathrm{Dol}}(C, \operatorname{SL}_2)_{\gamma}/\Gamma; u,v)(uvq)^{F(\gamma)}
\end{equation*}
for any $\kappa \in \hat{\Gamma}$ and $\gamma = w(\kappa)$.
As in \cref{thm:topologicalmirrorsymmetryrank2}, the case $\kappa=1$, alias $\gamma=0$, is trivial. Suppose then $\kappa \neq 1$ and $\gamma \neq 0$. The perverse filtration on $IH^d({M}_{\mathrm{Dol}}(C, \operatorname{SL}_2))_{\kappa}$ is concentrated in degree $d-2g+2$ by \cite[Theorem 5.5]{Mauri20}. Moreover, the Hitchin map $\chi_{\gamma}$ is a quotient by the inverse of the group law of the projection
\[{M}_{\mathrm{Dol}}(C_{\gamma}, \operatorname{GL}_1)\supset T^*\mathrm{Prym} = \mathrm{Prym} \times \mathbb{C}^{g-1} \to \mathbb{C}^{g-1},\]
where $\mathrm{Prym}$ is the connected component of the identity of the kernel of the norm map $\mathrm{Nm}\colon \operatorname{Pic}^0(C_{\gamma}) \to \operatorname{Pic}^0(C)$, given by $\mathrm{Nm}(L)=L \otimes \iota^*L$. Hence, the perverse filtration on $IH^d({M}_{\mathrm{Dol}}(C, \operatorname{SL}_2)_{\gamma}/\Gamma)$, with $\gamma \neq 0$, is concentrated in degree $d$; cf proof of \cite[Theorem 6.6]{FelisettiMauri2020}.
Then one easily see that \cref{conj:pervtopmirr} for $r=2$ is equivalent to \cref{thm:topologicalmirrorsymmetryrank2}.
\end{proof}
\textbf{Acknowledgement.}
This work have been supported by the Max Planck Institute for Mathematics.
\bibliographystyle{plain}
|
\section{A COVID-19 decision support framework}
In this section, we describe how to construct a framework for a DSS to be used for examining the efficacy of different COVID-19 countermeasures within a Bayesian decision analysis. We assume that the DC satisfies the following key requirements: (1) they must agree on a single agreed rationale for their stated beliefs, and (2) the preferences and any elicited expert judgement used is adopted as their own.
In this paper, we use the definition of a DSS from \cite{french2009decision} that, ``A decision support system is a computer-based system that supports the decision-making process, helping [DCs] to understand the problem before them and to form and explore the implications of their judgements, and hence to make a decision based upon understanding''.
It is convenient to break the process of creating a DSS down into four phases:
\\
\noindent \textbf{1) Elicitation of the class of strategies we might consider and the attributes of the utility function.}
We start by considering the types of strategy whose effectiveness we need to assess. In this framework, for the purpose of this paper, a strategy is defined by the regimes that might be imposed (such as a lockdown) and the thresholds (based on the system, which when reached leads to a switching of regimes). These strategies are designed to control the behaviour of the population and are typically tiered by the level of stringency these represent, i.e. the amount of disruption they cause to the normal life of the population.
The \emph{attributes} of the utility function are the different features of interest that each strategy affects. We define $\textbf{A}=\{A_1,A_2,\dots,A_m\}$ as the set of attributes of interest, with $a_i$ the value that attributes $A_i$. In the context of decision support for COVID-19, attributes which may be of interest include: the years of life saved across the population from avoiding infection by the virus; the years saved through timely medical examinations, for example through the detection of cancer and the encouragement to report to Accident and Emergency departments (A\&E) when exhibiting symptoms of a stroke or heart attack. A further threat to survival will be the response to increased poverty as categorised by the distribution of life expectancy given movements in the social class across the population induced by, for example, less effective education, unemployment or reduced employment activities. Note that all these attributes can be measured in units of years of life saved. On top of these, we may need to consider further attributes that define the quality of life of the population or its wellbeing, such as measures of depression and anxiety or economic hardship.
\\
\noindent\textbf{2) Elicitation of a quantification of the DC's marginal utilities of these attributes and the criterion weights.}
A subjective expected utility analysis consists of two components: a utility function $U(\boldsymbol{a})$ over a set of measured attributes $\boldsymbol{a}=\left( a_{1},a_{2},\ldots,a_{m}\right)$ and a set of
multivariate probability densities $\left\{p_{s}(\boldsymbol{a}):s\in S\right\}$ into the future associated with each possible countermeasure strategy. Whilst the latter would be provided by the appropriate domain experts working with statisticians and mathematical modelers, the utility function $U(\boldsymbol{a})$ needs to be elicited from the DC to reflect how
they intend to frame their objectives and prioritise them. More precisely, if $\boldsymbol{a}^{-}$ and $\boldsymbol{a}^{+}$ denote what the DC perceive to be respectively the worst and best credible outcomes then for each vector of outcomes $\boldsymbol{a}$, $\left\{ \boldsymbol{a:a}^{-}%
\leq\boldsymbol{a}\leq\boldsymbol{a}^{+}\right\} $, $U(\boldsymbol{a})$ is an increasing linear function of the probability $q(\boldsymbol{a})$ where the DC finds the outcome $\boldsymbol{a}$ with certainty equally preferable to a hypothetical situation where they are faced with obtaining the best possible outcome $\boldsymbol{a}^{+}$ with probability $q(\boldsymbol{a}%
)$ and the worst $\boldsymbol{a}^{-}$ with probability $1-q(\boldsymbol{a})$. In this way, aspects of the DC's risk aversion can be captured (see \citet{smith2010bayesian}, \citet{papamichail2003explaining}.
On the basis of certain basic axioms (see \citet{smith2010bayesian}), various ways have been devised to indirectly elicit these preferences efficiently, effectively and specifically with less biases. One assumption that is often made, and can be checked against a DC's expressed preferences, is that the DC has value independent attributes. The attributes $\boldsymbol{a}$ are said to be \emph{value independent} if the DC always finds two strategies $s_{1}$ and $s_{2}$ (leading to densities over outcomes $\pi _{1}(\boldsymbol{a})$ and $\pi_{2}(\boldsymbol{a})$ respectively) equally preferable whenever the marginal distributions over $\pi_{1}(\boldsymbol{a})$ and $\pi_{2}(\boldsymbol{a})$ for each component attribute in $\boldsymbol{a}$ are the same. It can then be shown that we can find positive criterion weights
$\left( k_{1},k_{2},\ldots,k_{m}\right) $, $\sum_{i=1}^{m}k_{i}=1$ such that
$U(\boldsymbol{a})$ can be written
\[
U(\boldsymbol{a})=\sum_{i=1}^{m}k_{i}U_{i}(a_{i})
\]
where $U_{i}(a_{i})$ -- called the \emph{marginal utility} of $a_{i}$ -- is an
increasing function of its argument, and $k_{i}\geq0$, $i=1,2,\ldots,m$. Whether the attributes satisfy value independence is not an inherent feature of the attributes but is a subjective judgement of the DC. It has been found that in practice, provided the attribute vector $\boldsymbol{a}$ is carefully defined, this simple form well approximates the DC's actual utility function in the vast majority of analyses. The advantage of making this assumption is that $U(\boldsymbol{a})$ is then much easier to elicit and the output of the analysis much more transparent and easy to explain: see \citep{smith2010bayesian} for a long discussion of these points. In particular, the criterion weights of the different component attributes can be chosen to reflect their relative importance. On the other hand, the form of the marginal utilities can be chosen to reflect the extent to which a DC considers an outcome good or bad relative to its extremes for each single attribute in turn. Henceforth, in this paper, we will assume the DC has value independent attributes.
\\
\noindent \textbf{3) Building a probabilistic model of each attribute and performing a Bayesian analysis which combines available data with probabilistic expert judgements.}
This is the process of obtaining $p(a_i|s)$: the probability of attribute $i$ having value $a_i$ given strategy $s$. In the context of COVID-19, this involves obtaining a probabilistic model for the number of deaths based on the strategy taken. This requires some sophistication since the spread of the virus at any time is complex and uncertain as it is a function of both the regime that a strategy is in at that time and the latent state of the disease. Therefore, any model informing the DC needs to be able to accommodate expert judgements, epidemiological modelling and data, all synthesised through Bayesian techniques; where in complex cases these techniques would tend to be computational.
\\
\noindent\textbf{4) Calculating a final score on each attribute as an expectation of its utility function.}
These scores are weighted by the elicited criterion weight to provide expected utility scores for each of the strategies considered. Note that there is always an explanatory rationale to whether or not a strategy scores well based on each of the different individual attribute scores and the predictions of the uncertain consequences of each strategy.
As the attributes chosen are value independent,(\cite{keeney1993decisions, insua_french_2010}), the subjective expected utility score $\overline{U}(s)$ of each strategy $s$ over the $m$ attributes $a_{1},a_{2},\ldots,a_{m}$ with respective marginal utilities $U_{1},U_{2},\ldots,U_{m}$ and the associated criterion weights $k_{1},k_{2},\ldots,k_{m}$ elicited in phase 2, is given by %
\begin{equation}
\overline{U}(s)=\sum_{i=1}^{m}k_{i}\overline{U}_{i}(s).
\label{eq:utility}
\end{equation}
Here, the marginal expected utility scores for each attribute $j$ given strategy $s$ are given by
\begin{equation}
\overline{U}_{i}(s)=\int U_{i}(a_{i})dp(a_{i}|s)
\end{equation}
where $p(a_{i}|s)$ is the DC's subjective probability distribution built in phase 3 and $U_i(a_i)$ is the marginal utility function elicited in phase 2. Note here that in this context it is natural to define the attributes to extend to a point in time when the virus is no longer a threat. In this sense, the eventual utility scores associated with a given attribute is always uncertain throughout the decision making process.
\section{A simplified Bayesian analysis of COVID-19 strategies}
In this section, we give a simple illustration of how a formal Bayesian multi-criteria decision analysis, as described in Section \ref{sec:COVID-19DSS}, could have been used to help DCs weigh the efficacy of different options open to them. Typically, marginal utility functions would need to be elicited. We have mentioned above that these would typically be risk-averse, which would have the effect of reducing expected utility scores when the outcomes of associated strategies are more uncertain. However, because our main focus in this example is how the Bayesian multi-criteria decision analysis balances the efficacy achieved associated with different options, we shall in the first instance make the common assumption that each of the marginal utilities is linear in its attribute.
By making this assumption, the analysis becomes much simpler because the DC's uncertainty only enters into calculations \textit{through the expectations of the random variables defining the future}. For simplicity, we simulate the effects of each strategy up to 40 weeks \footnote{The code used to generate this example is available here: https://github.com/peterrhysstrong/COVID-19-DSS}.
In this example we include an analysis of the effects of policies on health issues. For simplicity, we do not include the effects of the policies on the economy or the political environment. However, in a real-world analysis, these would also need to be assessed.
\subsection{Class of strategies and attributes}
\subsubsection{Regimes}
In practice, there is a large set of different regimes to consider, including potentially all possible combinations of preventative measures such as: closing schools, making mask wearing compulsory and promoting working from home. Here, we have chosen to limit our example to 3 regimes that the strategies will be able to switch between for simplicity. These loosely correspond to a regime implemented so far by the UK government. The three regimes are:
\begin{itemize}
\item $r_{0}$ -- \textbf{No lockdown}: No restrictions.
\item $r_{1}$ -- \textbf{Partial lockdown}: Return to work and school, non-essential businesses open with 1m+ distancing.
\item $r_{2}$ -- \textbf{Complete lockdown}: Work from home in effect, ban on non-essential interactions, complete closure of schools and non-essential businesses.
\end{itemize}
The first two regimes are analogous to the UK policies in effect on 26th March and 4th July respectively. A strategy determines when and how the transitions between these regimes occur, with the possibility of no transitions allowed.
\subsubsection{Strategies}
As the DC will be responding to the spread of the disease in a population which is susceptible to shocks such as large outbreaks, strategies considered should naturally include those which include switching between these three regimes.
In this example, we will consider several strategies that have different thresholds that give the transitions between regimes. At the time of writing, the thresholds on what measures the government was using to inform their strategies were not publicly available. In lieu of this, the transition thresholds for our example were based around the decisions made by England if they were based on the metrics defined in our example.
In practice, a DC would likely be interested in a much wider range of strategies. However in our example, for ease of analysis, we will narrow down the set of strategies that we are interested in by putting the following conditions on them:
\begin{itemize}
\item At most, one switch between regimes can take place in a week.
\item Within the timescale of the simulation, once restrictions have began, we do not return to no restrictions.
\item The transitions between regimes here are set such that they are dependent on the cumulative reported deaths, which is assumed to be known and the reported proportion of the population infected, which is assumed to be known with a delay of a week. In practice, this could be estimated from the proportion of positive tests from a sample of the population with the information being made available with a delay.
\end{itemize}
Thus, the state-space diagram for each countermeasure strategy considered corresponds to one of the figures in Figure \ref{fig:exregimes}.
\begin{figure}[H]
\centering
\includegraphics[width=0.5\textwidth]{tikz_figure.jpg}
\caption{State-space diagram of transitions between the three regimes for the two different types of strategies. (a) Strategies including lockdowns. (b) Strategies without lockdowns. }
\label{fig:exregimes}
\end{figure}
Below, we outline the different conditions that a strategy might have for switching between regimes. \\
\noindent \textbf{Initial lockdown} ($r_0 \rightarrow r_1 / r_0 \rightarrow r_2$): Given that we are in regime $r_0$, a transition to $r_1$ or $r_2$ occurs when the cumulative deaths are larger than some chosen critical value $L$. We consider three choices of $L$:
\begin{itemize}
\item Earlier lockdown: $L_1=100$;
\item lockdown in line with when the UK decided to go into lockdown: $L_2=300$;
\item Later lockdown: $L_3=500$.
\end{itemize}
Note that since the population can only transition out of $r_0$ at most once, there is no need to set out different conditions for transitioning to $r_1$ and $r_2$. Instead, each strategy specifies whether its transition out of $r_0$ is to $r_1$ or $r_2$. \\
\noindent \textbf{Tightening of lockdown} ($r_1 \rightarrow r_2$): When a population is under regime $r_1$, a 5$\%$ rise in the observed number of cases in a week will cause a transition into $r_2$. As no tightening of lockdown had occurred at the time of this study, the choice of a 5\% rise was an arbitrary one for illustrative purposes and, in practice, a range of values would be considered. \\
\noindent \textbf{Easing of lockdown} ($r_2 \rightarrow r_1$): Given that a population is currently under regime $r_2$, a transition into $r_1$ occurs when the proportion of infected individuals falls below a certain critical proportion $E$ of the peak proportion of infected individuals since $r_2$ has been in effect. We consider the following choices of $E$:
\begin{itemize}
\item No easing: $E_0=0$;
\item Easing in line with when the UK decided to ease lockdown: $E_1=0.12$;
\item Quicker easing: $E_2=0.3$;
\item Very quick easing: $E_3=0.5$.\\
\end{itemize}
Based on these transition rules, we assume that the initial transition out of $r_0$ is to $r_2$ for all but three of the strategies we consider. A strategy denoted as $L_i\_E_j$ has initial transition from $r_0$ to $r_2$ under critical value $L_i$ and allows for easing from $r_2$ to $r_1$ under critical proportion $E_j$ where $i \in \{1,2,3\}$ and $j \in \{0,1,2,3\}$. Further, analogous to ``no complete lockdown" strategies, we define three strategies denoted by $L_i\_E*$ where the initial transition is from $r_0$ to $r_1$ under critical value $L_i$, $i \in \{1,2,3\}$. Under these three strategies, once regime $r_1$ is effected, it remains in place until the end of the simulations with no transition into $r_2$ permitted.
\subsubsection{Attributes}
We will consider the attributes shown in Figure \ref{fig:atttree} that shows the attribute tree for our example. Since we are only considering health-related attributes, all the attributes can be measured in terms of aggregate expected life-years lost by the population. In this example, we are interested in the attributes of:
\begin{itemize}
\item $a_1$: Life-years lost due to COVID-19.
\item $a_2$: Excess life-years lost due to poverty.
\item $a_3$: Excess life-years lost due to delayed cancer diagnosis.
\end{itemize}
This attribute tree can be easily adapted to add other value independent attributes (see Section \ref{sec:COVID-19DSS}) of interest for the DC such as cost of strategy implementation, economic impacts and social impacts such as quality of life.
\begin{figure}[H]
\centering
\begin{tikzpicture}[sibling distance=20em,
every node/.style = {shape=rectangle, rounded corners,
draw, align=center,
top color=white, bottom color=blue!20}]]
\node {Attributes}
child { node {Short to medium term\\ $\Bar{a_1}$}
child { node {Covid-19 Deaths\\$a_1$} }
child { node {Delayed\\ cancer diagnoses \\$a_2$} }}
child { node {Long term\\ $\Bar{a_2}$}
child { node {Poverty\\$a_3$} }};
\end{tikzpicture}
\caption{Example of an attribute tree for COVID-19.}
\label{fig:atttree}
\end{figure}
\subsection{Marginal utilities and criterion weights}
\subsubsection{Marginal utilities}
As in this example each of our attributes is in terms of life-years lost, we will take our marginal utility function to be the negative identity function, $U_i(a_i)=-a_i \quad \forall i$.
The values of $a_i$ are, of course, uncertain. However, because of the assumption of preference independence and this linear utility function for the evaluation of the score, we need only elicit the subjective expectation of this quantity under each strategy. Here, again for simplicity, we have chosen to identify this expectation with the output a stochastic model of the process. We assume that the centre adopts as their expectation those delivered by relevant experts. For a theory justifying when this is an appropriate protocol for a DC to adopt, see \cite{leonelli2015bayesian} and \cite{smith2015decision}.
\subsubsection{Criterion weights}
The criterion weights for these attributes reflect how the DC prioritises them. We compare the effects of different setting of the attribute weights $k = (k_1, k_2, k_3)$ on the aggregate weighted life-years lost where weight $k_i$ is for attribute $a_i$, $i = 1,2,3$. For this example, we will consider the effect of several choices of criterion weights:
\begin{itemize}
\item (1, 0, 0): The DC only cares about life-years lost due to COVID-19
\item (0.5, 0.5, 0): The DC only cares about life-years lost due to COVID-19 and delayed cancer diagnosis and weights them equally.
\item (1/3, 1/3, 1/3):The DC cares about all 3 causes of life-years lost and cares about them all equally.
\item (0.45, 0.45, 0.1): The DC cares about all 3 causes of life-years lost and but life-years lost caused by short and medium term attributes are more important.
\end{itemize}
\subsection{Probabilistic model of each attribute}
For each attribute, we will give a probabilistic model giving us the probability of an attribute obtaining each value for each strategy, $p(a_i|s)$.
\subsubsection{COVID-19 deaths ($a_1$)}
To estimate the deaths from COVID-19 and the proportion of infected individuals in the population, we use a simple discrete-time SIRD compartmental epidemiological model. The states in this model are ``Susceptible" (S), ``Infectious" (I), ``Recovered" (R) and ``Dead" (D). The dynamics of this model can be seen in Figure \ref{sird} and are described in the supplementary material S1. Clearly, D is an absorbing state and we assume that those that have recovered from COVID-19 are immune to it. We stratify our population in this model by region and age.
\begin{figure}[H]
\centering
\includegraphics[width=0.8\textwidth, trim={0cm 2cm 0 0cm}] {sird.jpeg}
\caption{SIRD disease dynamics}
\label{sird}
\end{figure}
In order to capture the uncertain nature of the future, we have modelled uncertainty around the rate of infection with detail given in supplementary material S1.2. By running our stochastic simulation 1000 times we obtained expected values for both the attribute of number of deaths due to COVID-19 and the number of weeks spent in each regime. One of these simulations for each strategy is shown in Figure \ref{deaths_evolve}. Note here that in practice the expected numbers used are those provided by epidemiologists using more sophisticated models.
\begin{figure}[H]
\centering
\centering
\includegraphics[width=\textwidth]{lockdown_example.jpg}
\caption{Plots showing how the daily deaths evolve over time for one simulation within the stochastic model.(a) Strategies including lockdowns. (b) Strategies without lockdowns.}
\label{deaths_evolve}
\end{figure}
\subsubsection{Delayed cancer diagnoses ($a_2$)}
In this example, deaths due to delayed diagnoses of cancer are used here as a proxy for deaths that would not have normally occurred, as a composite of missed treatments and patients who have not presented in hospital. This was chosen as a proxy as information was easily available. In practice, this could be improved on by either creating attributes for other excess deaths due to missed treatments and eliciting models of each attribute or, when that is not feasible, updating the utility function for this attribute. The latter could be done simply by assuming that deaths due to delayed cancer diagnosis will form a certain proportion of all of the deaths due to missed treatment. Estimates like this are often necessary in the early stages of a decision support system for an emergency response when there is sparse information about the underlying science.
\citet{sud2020effect} evaluate how delays to cancer diagnoses due to the COVID-19 outbreak impact survival and life-years lost. This study gives a nearly linear relationship for delay in cancer diagnoses to extra deaths per age group. Using this relationship, we are able to produce plausible values for the expected total number of deaths arising from this delay. These can be used as benchmark values the DC might use for its subjective expectations of this attribute.
As at the time of this report there is little research done on the impact of partial lockdown on cancer referrals, we will assume that the number of deaths in a partial lockdown is half of those in a total lockdown. In practice, the DC would elicit expert judgement about this value where possible and perform a sensitivity analysis.
\subsubsection{Poverty ($a_3$)}
Here, we notice that the centre has other experts available who can assess the impact on years of life lost on poverty induced by a strategy.
For example, \citet{decerf2020lives} evaluate the impact of the COVID-19 outbreak in terms of number of poverty-years. Poverty-years are calculated using COVID-19's predicted effect on GDP and estimates the number of additional people in poverty as a result (in the UK, 4.37 million). \citet{decerf2020lives} conservatively assumes that these individuals will remain poor only for a single year.
We use \citet{decerf2020lives}'s calculation that 8.8 poverty-years equate to 1 life-year to inform our DSS, which gives us an estimate for the impact of lockdown on life-years lost due to poverty. Further information is given in the supplementary material section S2.
\subsection{Final scores calculated}
We can now combine this all together to get Equation \ref{eq:raw_attribute} and assess which strategy maximises the utility score given the criterion weights.
\begin{equation}
\overline{U}(s)=\sum_{i=1}^{3}k_{i}\overline{U}_{i}(s)=\sum_{i=1}^{3}k_{i} (-\overline{a_i(s)})
\label{eq:raw_attribute}
\end{equation}
Here $\overline{a_i(s)}$ is the expected value of attribute $i$ under strategy $s$. In this example, this is the magnitude of the utility function, $\overline{U}_{i}(s)$.
\subsection{Results}
\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{weighted_att.jpeg}
\caption{Aggregate weighted life-years lost under different attribute weightings for each strategy. Top left: COVID-19 deaths only. Top right: COVID-19 and delayed cancer diagnoses deaths equally weighted. Bottom left: COVID-19, delayed cancer diagnoses and poverty deaths equally weighted. Bottom right: COVID-19, delayed cancer diagnoses and poverty deaths custom weighting.}\label{attweight}
\end{figure}
From Figure \ref{attweight}, we can see that under $k = (1, 0, 0)$ where only life-years lost due to COVID-19 are of interest, strategies that do not involve a lockdown perform orders of magnitudes worse. Looking at strategies that do involve lockdown, ``earlier lockdown" and ``slower easing of lockdown" lead to fewer life-years lost. The second scenario under $k = (0.5, 0.5, 0)$, in which life-years lost due to delayed cancer diagnoses are also considered, we can begin to see the drawbacks of the lockdown with reduced disparity between strategies that include lockdown and those that do not. However, the lockdown strategies remain more desirable. This example also shows little difference between the lockdown strategies that only differed in their easing of lockdown with more significant differences depending on when the lockdown conditions were implemented.
Under $k = (1/3, 1/3, 1/3)$, which includes life-years lost due to poverty with equal weighting for all attributes, the non-lockdown strategies perform the best. It is interesting to see how small the weight on our long term impacts on health might need to be before we impose the most stringent of lockdown strategies. For $k=(0.45,0.45,0.1)$, we treat short and medium term years lost as more important than long term years lost. While lockdown strategies generally still perform better, the disparity among all strategies is much less.
\begin{figure}[H]
\centering
\includegraphics[width=1\textwidth]{front.jpeg}
\caption{Pareto plots showing the trade-off between life-years lost (LYL) due to poverty (deprivation) on one hand and COVID-19 and delayed cancer diagnoses on the other. The dashed lines represent examples of different attribute weightings.}
\label{paretototal}
\end{figure}
We further explore these trade-offs in Figure \ref{paretototal}, which shows a Pareto front plotting life-years lost due to poverty against those lost from COVID-19 and cancer combined, assuming equal weighting. We can see how the trade-off in weights leads to different strategies being the most effective, with the weights giving the gradient of the line. This plot can help to reduce the number of strategies considered, as in this example strategies with no easing of lockdown are strictly dominated by strategies with quick easing of lockdown. Thus, there is no weighting of attributes in which we would rather have no easing. We can also find the value $c$ in $k^*=(c,c,1-2c)$ for which the DC would swap preferences between strategies that involve lockdown and those that don't. In this example, that point is at $c=0.4.19$, where short and medium term attributes are weighted as 2.58 times more important than long term attributes. At $c \geq 0.4.19$, the optimal weighting would be strategies that involve lockdown. For comparing trade-offs between more attributes at the same time, an n-dimensional Pareto front with each axis on its own label should be used. Further analyses of the trade-offs between the attributes using Pareto fronts is given in the supplementary materials S3-S4.
\section{Discussion}
In this paper, we presented how a DC can create a DSS under which countermeasure strategies for COVID-19 could be evaluated in real-time, using the best expert judgements associated with various strategies that it might be feasible for the DC to enact. This DSS enables the DC to input the attributes they consider to be vital, with attribute weightings reflecting their priorities, to evaluate and choose an appropriate metric by which these attributes can be measured.
The novelty of the framework lies in demonstrating how to provide multi-attribute Bayesian decision support in evolving decision contexts. We have shown that Bayesian decision analysis is an appropriate and holistic approach for addressing uncertainty in dynamic environments as we move from short-term to longer-term considerations. Indeed, the results of a Bayesian subjective utility approach are informative and help decision makers devise action plans \citep{french_2020}.
In the results from our example, we observed the effect of different criterion weights demonstrating their effect on the optimal decision. For illustrative purposes, the analysis presented here was simplified. In order to be operationalised, it would need to be refined with additional attributes as well as a utility function and attribute weightings that reflect the priorities of the DC. Firstly, more attributes -- such as hospital admissions, well-being, economic viability and public acceptability -- could be considered.
Further, such a DSS could be extended by considering the impact of other countries' management of COVID-19, caused by a spillover effect. For example, the UK's economy is likely to be affected by COVID-19 spreading in other countries regardless of the UK's state of lockdown. Our model implicitly assumes that such further causes of deprivation are additive, whereas the relationship between impacts to domestic employment and international economic variables on the differential scores associated with different options is likely to be more subtle. However, we note that, by embarking on this decision analysis, we are drawn into trying to quantify such issues and to fold these important considerations into the analysis; this DSS gives a framework for introducing such studies as these become available.
Another critical issue that could be incorporated would be the capacity of the health provider, e.g. the National Health Service (NHS) in the UK, both with respect to the number of available beds and the number of doctors and nurses available to treat COVID-19 as well as non-COVID-19 patients. Hospitals functioning close to full capacity may lead to increase in the mortality rate of hospitalised COVID-19 patients \citep{wilde2021association} as well as temporary closure of all non-urgent hospital procedures \citep{covidsurg2020global}. Thus, if the NHS were to be overwhelmed, the underlying model would need to be adapted to reflect the increased life years lost due to the added pressure. Indeed, in the UK, COVID-19 policy decisions have been influenced by the capacity of the NHS \citep{guardian_nhs_overwhelm}, and measures such as introducing additional capacity through construction of temporary hospital facilities (known in the UK as Nightingale Hospitals) \citep{nightingale_nhs} were effected to prevent overwhelming the NHS. These capacities could be incorporated (1) in deciding the thresholds for changes in regimes and (2) by choosing hospital admissions as an attribute. The utility function of this attribute could be an indicator function that indicates when the hospital capacity is exceeded. Assigning a very large criterion weight to such an attribute would have the effect of making any strategy that would lead to the threshold being passed strictly dominated by other strategies and therefore, not considered.
We also assumed that all the marginal utilities were linear. In contexts like the one above, the DC tends to be risk-averse. The mathematics means that the scores we assign to different options which lead to more uncertain outcomes are penalised. There are two elements that come into play here. The first issue is that the impacts of previously untried strategies are likely to be more uncertain. A risk-averse marginal utility would therefore tend to down-weight the scores of less well-tried options. The second is that attributes which are intrinsically more uncertain will tend to be given less weight across all strategies. For example, the implicit assumption of giving negligible weight to non-poverty related economic effects could be justified if the differential economic effects of one strategy against another were extremely uncertain. However, once the marginal utilities have been elicited, the framework is able to score the different options in light of this and the type of analysis above can be adapted. Further, compliance of the population to flip-flopping between regimes could be accounted for by considering it within the probabilistic predictions of the attributes.
Finally, we have used a na{\"i}ve predictive model of the progress of the disease. This dynamic model could obviously be improved; see \citet{keeling2021predictions, aguilar2020investigating}. In particular, we could choose a stochastic disease model such as \citet{abrams2020modeling}. Assumptions on immunity after recovery might also need to be revised as relevant studies become available; for example \citet{long2020clinical, liu2020prevalence}.
In conclusion, further research is needed to extend on this work in all the ways mentioned above. However, we have shown, using as simple example, how a DSS could be created and analysed to assess different countermeasure strategies for COVID-19. We have also shown how this could be done in a real time setting -- where there is only basic scientific understanding of such a disease and the impacts of its countermeasures, as witnessed for COVID-19 -- where DCs may need to rely on only coarse descriptions given by the experts when many judgements are unknown or unsubstantiated by evidence.
\section{Introduction}
Major disasters and incidents such as pandemics, nuclear disasters, volcano eruptions and tsunamis can impact the short-term and long-term well-being and health of people, nature and economies (see e.g. \citet{boyd2014emergency, jorda_singh_taylor_2020, ohtsuru2015nuclear, tamura2000modeling}). Decision-making in such crises involves balancing several complex factors, where each factor's priority may be a subjective judgement \citep{smith2010bayesian}. Such decision-making is further complicated by the decision centre (DC) -- an individual or a group of people, such as a government committee, who must have the authority to enact the decisions made -- receiving information from multiple sources. Such information is either factual or narrative in nature, and also may be noisy, uncertain and incomplete \citep{bakker2019decision}. Bayesian decision analysis has been shown to enable principled decision-making in such complex, multi-faceted problems \citep{smith2010bayesian, moffat2002bayesian, french1996multi, kurth2017trends}.
This paper focuses on implementing the well-developed principles of Bayesian multi-criteria decision analysis on the ongoing COVID-19 pandemic which first emerged in Wuhan, China in late 2019 \citep{who}. Using the framework presented in this paper, a DC, such as the government, can combine streaming uncertain information they receive about the current and potential effects of the pandemic on the various aspects of society in the form of expected utility scores to decide among a finite set of countermeasure strategies. The study within this paper was undertaken between June and October 2020.
In March 2020, most countries across Europe, including the UK, imposed lockdowns on regional or national levels in an effort to control the spread of the COVID-19 virus, prevent overburdening their healthcare systems and to buy more time to enable researchers to learn more about the disease \citep{sridhar_2020}. However, lockdowns have been economically damaging, particularly for hard-hit sectors such as aviation, tourism and hospitality \citep{coronavirus_impact}. The UK was in recession in the first two quarters of 2020 with early signs of recovery in the third quarter when lockdown restrictions were relaxed \citep{pwc_report}. As of 18th October 2020, 9.6 million jobs have been furloughed under the UK government's job retention scheme \citep{hmrc_statistics}. Following the relaxation of initial national lockdown measures in summer of 2020, there have been several localised outbreaks of COVID-19 across the UK, resulting in local lockdowns. The government, keen to avoid another national lockdown to protect the economy, has been under scrutiny regarding the efficacy and frequent changes of the measures introduced under local lockdowns \citep{bbc_local_lockdown}. In this ``second wave", the government has again faced a challenging balancing act: managing the public health impacts of COVID-19 on one hand, its economic and social impacts on the other. The research presented in this paper was completed against this backdrop; it reached completion in October 2020, when COVID-19 cases were increasing rapidly and the government was on the brink of announcing a second lockdown.
To compare the health, social and economic impacts of the candidate countermeasure strategies, it is necessary to evaluate them on a comparable scale. Typically, health and social impacts are measured in terms of health-adjusted life-years (QALYs) \citep{zala2020costing, miles2020stay} or wellbeing-adjusted life-years (WELLBYs) \citep{layard2020release, de2020taking} as caused directly by the virus itself or indirectly by the countermeasures implemented to tackle it; the economic impacts are measured in terms of difference in realised GDP and empirical or projected GDP, or sometimes as number of jobs in various sectors, consumer spending growth or business investment growth \citep{pwc_report}. Further, these impacts often move in opposite directions in response to any given countermeasure. Hence, a \textit{common metric} is essential for their comparison \citep{layard2020release}.
Within a Bayesian decision analysis, this common metric is defined through an \textit{expected utility score} given to each of the health, social and economic attributes. For each attribute, we define a function to convert the actual recorded or estimated measurements (e.g. number of deaths or first order difference in the GDP) into a common metric (here, utilities). Next, by combining these expected utility scores using a specified utility function, we arrive at an expected utility score for each countermeasure strategy. However, the exercise of arriving at an expected utility score for each strategy is complicated by two factors: (1) the specification of the DC's utility function may involve uncertain estimates of future events (e.g. predicted number of COVID-19 deaths under a complete lockdown), and (2) individuals within a DC may be inclined to prioritise the importance of the various impacts in different ways (e.g. some individuals may view the impact on GDP as more important than others). Consequently, there may be no consensus within a DC as to how efficacious each policy may be.
In this paper, we argue that, in light of the aforementioned challenges, it is essential that any decisions made by policymakers are done after due consideration and analysis within a systematic decision support framework with proper treatment of the associated uncertainties \citep{french1995uncertainty}. Such a framework is a defensible and powerful continuous assessment tool and allows policymakers to feed information (for example, the statistics and estimates the UK government obtains from various sources, including its Scientific Advisory Group for Emergencies (SAGE)), and their priorities into the model when it comes to decision making. At the time of writing, the COVID-19 decision support tools in the literature (discussed in Section \ref{sec:related_research}) fall short on providing such a systematic and statistically sound framework that any DC could utilise to compare their options under their self-determined constraints. However, established Bayesian methodologies have been developed over a number of decades to support DCs, capable of respecting both the preferences of the centre and the uncertainties around predictions that inform such decisions, including methods that help compare a variety of different priority weightings. This paper outlines how such a Bayesian decision analysis can help evaluate the efficacy of different COVID-19 countermeasure strategies.
\section{Related research}
Bayesian decision analyses have been successfully adapted and implemented for various applications, including nuclear disaster support \citep{geldermann2009multi}, food security management \citep{barons2020decision}, industrial risk management \citep{rikalovic2014gis} and environmental planning \citep{mattiussi2014decision}. For more Bayesian and frequentist applications of decision support systems, see \citet{eom2006survey}. The COVID-19 threat is fast moving, driven not only by the stochastic spread of the virus but also the dynamic control exerted by government on the activities of its agents and the general population to the unfolding crisis. Any support tool for managing this crisis needs to acknowledge that, while governments may need to switch between lockdown regimes to increase or decrease the severity of social restrictions in response to the virus prevalence, the general population may tire of the uncertainty in their lives attributable to constantly changing restrictions. In this paper, we draw from previous work \citep{leonelli2015bayesian} to demonstrate how to create a multi-attribute decision support system (DSS) under a Bayesian approach to address COVID-19 decision making challenges.
We note here that a Bayesian approach has been used for other COVID-19 studies \citep{dehning2020inferring, verma2020global, neil2020bayesian, mbuvha2020Bayesian}, although not for the purpose of providing a decision support framework.
Finally, note that in many situations, historical data alone may be inadequate for estimating aggregate utilities from the various attributes. Continuously generated data for our recent and evolving COVID-19 situation, along with empirical data from past epidemics and pandemics are useful to a certain extent but are not sufficient. In this case, it is useful to elicit the required estimates and their associated uncertainties through discussions among a panel of domain experts. The iterative nature of such discussions and the development of the corresponding DSS are detailed in \cite{barons2018eliciting}. The structure and estimates of such a DSS would need to be repeatedly reviewed as more information comes to light (e.g. development of a vaccine or new information on immunity from the disease). Typically, the iterative improvements to the DSS are performed until it is deemed to be requisite, i.e. the DC is content that the structure of the DSS is as required \citep{phillips1984theory}. Given the urgency of the issue at hand, it is prudent to begin the process of establishing such a requisite DSS by first developing its framework to address the new challenges of the COVID-19 decision making problem. As more information and data come in, the DSS can be appropriately modified within the framework and the data can be fed into it. Thus our contribution here is the framework for a multi-attribute Bayesian DSS for COVID-19.
There have been several studies focused upon specific aspects of the COVID-19 decision-making problem such as impacts on mortality and poverty \citep{decerf2020lives}, a cost-benefit analyses of a lockdown \citep{miles2020stay, layard2020release}, and the impact of specific countermeasure strategies \citep{peto2020universal, karnon2020simple, lander2020uk}. However, none of these studies adopted a Bayesian approach which we believe is essential in this case as it not only supports uncertainty handling but also gives the government a more transparent and auditable tool. Additionally, these studies look at specific cases but do not describe their general framework. This makes it hard to adapt the existing work as new information comes in. We note here that clinical decision support tools for COVID-19 (e.g. \cite{mcrae2020clinical, wu2020development, liu2020covid, reeves2020rapid}) are beyond the scope of discussion for this paper.
In the OR literature, Bayesian approaches have been used to model probabilities (see e.g. \cite{zafari_soyer_2020}; \cite{vargo_cogill_2015}). There have been recent calls however, for developing Bayesian subjective utility models to address deep uncertainty in evolving decision contexts (French, 2015; French, 2020). This paper addresses this by developing a framework for providing decision support in an evolving and complex decision context.
\section{Introduction}
\label{sec:intro}
\input{intro_revised}
\section{Related research}
\label{sec:related_research}
\label{practical considerations}
\input{litreview}
\section{A COVID-19 Decision Support Framework}
\label{sec:COVID-19DSS}
\input{COVID-19DSS}
\section{A Simplified Bayesian Analysis of COVID-19 Strategies}
\label{sec:experiments}
\input{Experiments}
\section{Discussion}
\label{sec:discussion}
\input{dicsussion}
\printbibliography[title = References]
\newpage
\begin{appendices}
\input{sup}
\end{appendices}
\end{document}
\section*{S1: Disease dynamic model}
\label{ddm supp}
In our example, the UK will be split up into 14 regions and the 7 age groups used by the ONS, each with population $N_{i}$ which we will assume to be constant. Table \ref{sup:tab:regions_ages} shows the regions and age groups. We further split up these populations into age groups giving us $N_{i,a}$. This model has the constraint:
\begin{equation}
S_{i,a}(t)+I_{i,a}(t)+R_{i,a}(t)+D_{i,a}(t)=N_{i,a}
\label{N}
\end{equation}
where $i= 1,2,\dots,14$ are the locations and $a=1,2, \dots,7$ are the age classifications. For any time $t$, the sum of individuals in each state in location $i$ in age group $a$ is equal to the number of individuals in the sub population $i$ in age group $a$. The disease dynamics are shown in Equations (\ref{N}-\ref{D}).
\begin{equation}
\Delta S_{i,a}(t) = - \beta_{a} \frac{I_{i}(t)}{N_i}S_{i,a}(t)
\label{S}
\end{equation}
\begin{equation}
\Delta I_{i,a}(t) = \beta_{a} \frac{I_{i}(t)}{N_i}S_{i,a}(t)-(\gamma_a+\lambda_a)I_{i,a}(t)
\label{I}
\end{equation}
\begin{equation}
\Delta R_{i,a}(t) = \gamma_a I_{i,a}(t)
\label{R}
\end{equation}
\begin{equation}
\Delta D_{i,a}(t) = \lambda_a I_{i,a}(t)
\label{D}
\end{equation}
Each time step in this model represents a day, with $\beta_{a}$, $\gamma_a$ and $\lambda_a$ being the retrospective daily rate of infection, recovery and death. We will assume each of these rates are independent of location and time but are age-dependent (hence, these are indexed only by age classification $a$ and not indexed by location $i$ or time $t$). $\beta_{a}$ is dependent on what strategy is currently being enforced to fight the infection. $I_i(t)=\sum_a I_{i,a}(t)$ is the total number of infectious people in location $i$. This is used in Equations \ref{S} and \ref{I} as we are assuming uniform mixing between age groups.
\begin{table}[H]
\centering
\begin{tabular}{|c|c|}
\hline
\textbf{Regions} & \textbf{Age groups} \\ \hline
North East & Under 1 \\
North West & 1-14 \\
Yorkshire and the Humber & 15-44 \\
East Midlands & 45-64 \\
West Midlands & 65-74 \\
East & 75-84 \\
London & 85 and over \\
South East & \\
South West & \\
Wales & \\
Scotland & \\
Northern Ireland & \\\hline
\end{tabular}
\caption{Table showing regions and age groups used in our analysis}
\label{sup:tab:regions_ages}
\end{table}
\subsubsection*{S1.1: $\lambda_a$ and $\gamma_a$}
To estimate $\gamma_a$ and $\lambda_a$, we will use estimates of infection fatality ratios in each age group $D_a$ from \cite{ferguson2020report}. We will assume that
28 days is sufficient time after a positive test for an individual's death not to be thought of as due to coronavirus. We will therefore model that after 28 days, the probability of still being infected is 0.05. Using this, we can estimate $\gamma_a$ and $\lambda_a$ by solving the following:
\begin{equation}
\sum_{t=0}^{27} (\gamma_a+\lambda_a)(1-\gamma_a-\lambda_a)^t=0.95
\label{notI}
\end{equation}
\begin{equation}
\sum_{t=0}^{\infty} \lambda_a(1-\gamma_a-\lambda_a)^t=D_a
\label{Dead}
\end{equation}
Equations \ref{notI} and \ref{Dead} give the probability of either recovering or dying in 28 days since infection and the probability of dying after becoming infected respectively. $(1-\gamma_a-\lambda_a)^t$ is the probability that an infected individual of age $a$ is still infectious at time step $t$.
Equation \ref{notI} can be solved algebraically using the geometric series formula, giving $\lambda_a+\gamma_a=0.101466$. Combining this with Equation \ref{Dead} can be used to give each $\lambda_a$ and $\gamma_a$ values.
\subsubsection*{S1.2: $\beta_{a}$}
We can calculate an estimate for $\beta_{a}$ by splitting it into $p$ (the probability of becoming infected after contact with an infected individual) multiplied by $c_{a}$ (the average number of contacts an individual in age group $a$ has per day while strategy $s$ is being enforced). Under normal conditions, the average number of contacts for individuals of each age can be found from \cite{mossong2008social}. We assume under partial lockdown that the number of contacts will be half of this and within lockdown individuals of any age group will have 3 contacts. We can work out an estimate of $p$ for our example by considering the value of $R_0$, the average number of people an infectious person will infect in a fully susceptible population. There have been many estimates of $R_0$ for COVID-19. \cite{liu2020reproductive} gives a review of these and gives a median value of 2.79. We can calculate $p$ as follows:
\begin{equation}
R_0=\frac{\sum_{a \in A}n_a\sum_{t=0}^{\infty}c_a p(1-\gamma_a-\lambda_a)^t}{\sum_{a \in A}n_a}
\end{equation}
Here $\sum_{t=0}^{\infty}c_a p(1-\gamma_a-\lambda_a)^t$ is the average number of infections that an infectious individual in age group $a$ will cause. The population weighted average of this gives $R_0$.
This can be rewritten to give $p$ and solved using the previously described values.
\begin{equation}
p=\frac{R_0\sum_{a \in A}n_a}{\sum_{a \in A}\sum_{t=0}^{\inf}n_ac_a (1-\gamma_a-\lambda_a)^t}
\end{equation}
This gives $p$=0.023.
To add stochasticity into our model, we ran the simulated SIRD model 1000 times with $p$ distributed from a log normal with mean 0.023 and 0.1 standard deviation on the log scale.
\subsubsection*{S1.3: Initial conditions}
The model was initialised on the 2nd March: the first date when all regions in the UK had at least one coronavirus case.
\newpage
\section*{S2: Other attributes} \label{other attributes supp}
\subsection*{S2.1: Delayed cancer diagnoses}
\cite{sud2020effect} use 10-year age- and stage-stratified cancer survival estimates in England for 20 different common tumour types present in adults using information from Public Health England. Using these estimates and annual numbers of early-stage cancers diagnosed via the `2-week-wait’ pathway, they calculate probable lives and life-years lost in England dependant on the length of a lockdown and how many symptomatic patients’ assessments are delayed by lockdown.
The research primarily focuses on early stage I-III cancers; they assume delays to stage IV cancer treatment will not affect 10 year survival. They also focus on diagnosis delays instead of treatment delays; it is assumed that there is no delay to treatment, which the authors accept is unlikely in reality. The authors calculate the attributable lives lost by length of delay due to the pausing of the `2-week-wait’ pathway for 1-6 months for 20 different tumour types. They also age-stratify each cancer type to denote the proportion of tumours prevalent in each age group. By combining this information, we can identify how many lives are lost for each cancer in each age group dependant on how long the `2-week-wait’ pathway is not operating at 100\% capacity.
\subsection*{S2.2: Deaths due to poverty}
\citet{decerf2020lives} estimates the number of additional poverty years for the UK at 4,370,000 and estimate that 8.8 poverty years is equivalent 1 life-year lost. In Table \ref{tab:poverty_estimates}, we use the estimates of the population living under the poverty threshold in the UK, stratified by age (`children', `working-age' and `pension-age') as given in \citet{social2019} and assume that the additional poverty years are distributed the same way.
\begin{table}[hb]
\begin{tabular}{c|c|c|cl}
Age Group & \begin{tabular}[c]{@{}c@{}}2017/18 Estimate of number \\ of people in poverty\end{tabular} & \begin{tabular}[c]{@{}c@{}}Additional \\ poverty years\end{tabular} & \begin{tabular}[c]{@{}c@{}}Equivalent \\ life-years lost\end{tabular} & \\ \cline{1-4}
Children & 4.6 & 1.41 & 0.16 & \\
Working-age & 8.3 & 2.54 & 0.29 & \\
Pension-age & 1.3 & 0.42 & 0.048 & \\ \cline{1-4}
Total & 14.2 & 4.37 & 0.498 &
\end{tabular}
\caption{Additional Poverty years in 2020 in the UK. All estimates are in millions.}
\label{tab:poverty_estimates}
\end{table}
We assume that children comprise of those under 16 years of age, working-age as those between 16 and 65 years of age, and pension-age as those over 65 years. \citet{decerf2020lives} evaluated the effect of the COVID-19 pandemic on poverty years as of early June 2020. For the UK, this period corresponds to the time that the UK was more or less under complete lockdown. In our paper, we shall assume -- for illustrative purposes -- that the impacts on poverty years were caused as a direct effect of the lockdown and whose alleviation would be proportional to the removal of lockdown restrictions. As with cancer deaths, at the time of this report there is little research done on the impact of partial lockdown on poverty and therefore we will assume that the number of additional years in poverty as a result of a partial lockdown is half of those in a total lockdown.
\cite{decerf2020lives} state that their methodology gives a conservative estimate by providing a lower bound for additional years spent in poverty due to COVID-19 countermeasure strategies. They make the assumption that the condition of individuals who are faced with poverty will improve within a single year and will not have further lasting effects. Thus, the premature mortality effects of poverty are not actually considered \citep{baland2019too}. Additionally, it is assumed that the GDP reduction caused by the economic contraction is distribution neutral, i.e. the inequality in the society remains constant. However, this is known not to be the case \citep{van2020covid}. While there are several shortcomings to these estimates, we use them in our paper for illustrative purposes and for the lack of better estimates regarding deprivation caused by COVID-19 and the lockdown at this point of the pandemic \citep{miles2020stay}.
\subsection*{S2.3: Age-stratification}
Occasionally, the age-stratification from our data sources is not the same as the age-stratification in this study. Where possible, the age-stratifications were combined. Where this was not possible, the group was split proportionally; for example, the 60-69 age group in \cite{sud2020effect} was split 50/50 into the appropriate age ranges for this research of 45-64 and 65-74. This assumes that the proportion of cases is uniformly distributed across the age group.
\newpage
\section*{S3: Life-years lost under each attribute}
\begin{sidewaysfigure}
\includegraphics[width=\textwidth]{lyl3att.jpeg}
\caption{Diagram showing the Life-Years Lost under each attribute due to each strategy.}
\end{sidewaysfigure}
\newpage
\section*{S4: Age comparison of strategies}
\begin{figure}[H]%
\centering
\centering
\includegraphics[width=10cm]{paretocomp.jpeg}
\caption{Pareto plots showing the trade-offs of life-years lost due to COVID-19 and delayed cancer diagnoses to life-years lost due to poverty (deprivation) comparing the effects of different trade-offs on the two age classifications of over and under 45 year olds. The dashed lines represent examples of different attribute weightings with the one in red representing the critical weighting between lockdown and non lockdown strategies.}%
\label{age_parate}%
\end{figure}
We explore how the strategy impacts different age groups in Figure \ref{age_parate}. We look at the critical weightings $k^*_o=(c_o,c_o,(1-2c_o))$ and $k^*_u=(c_u,c_u,(1-2c_u))$ for switching between strategies that involve lockdown and those that don't for groups of over and under 45 years old respectively. We get $c_u= 0.474$ and $c_o=0.338$, which means that when selecting strategies that focus on people over the age of 45, you would have to value short and medium term attributes as 1.05 times as important than those strategies that do not involve lockdown. When optimising to benefit people under the age of 45, you would have to value short and medium term attributes as 9.26 times as important to consider strategies than those which involve lockdowns.
|
\section{Yellow Post-AGB Stars in Globular Clusters}
The visually brightest objects in globular clusters (GCs) and old stellar
populations are stars in their final evolution from the tip of the
asymptotic giant branch (AGB) toward higher temperatures in the color-magnitude
diagram (CMD)\null. As these post-AGB (PAGB) stars pass at nearly constant
bolometric luminosity through spectral types early G, F, and late A, they reach
their brightest visual absolute magnitudes, because of the dependence of the
bolometric correction upon effective temperature.
It has been suggested \citep{Bond1997a, Bond1997b} that these luminous ``yellow'' PAGB stars
may be useful Population~II standard candles. In populations containing only old stars there should be fairly
sharp upper and lower limits to their brightnesses. In addition, yellow PAGB stars have
conspicuously large Balmer discontinuities in their spectral-energy
distributions (SEDs), making them easy to recognize using a single set of suitable photometric
observations. They should be detectable in early-type galaxies that do not
contain Cepheids, and in the halos of spirals, where there are fewer issues of
interstellar extinction than there are for Population~I distance indicators.
Because of the rapid evolutionary timescales for PAGB stars, however, these objects are very rare. The previously known luminous yellow PAGB stars in the Galactic GC system number only five: (1)~HD~116745 (``Fehrenbach's star,'' ROA~24; \citealt{Gonzalez1992} and references therein) and the RV~Tauri variable V1 in $\omega$~Cen tauri \citep{Jones1968}; (2)~two luminous yellow PAGB stars in NGC~5986 \citep{Bond1977, Alves2001}; and (3)~a yellow PAGB star in M79, discovered by \citet[][hereafter B16]{Bond2016}. There are very few, if any, yellow stars in the CMDs of Galactic GCs that lie above the horizontal branch (HB) and within $\sim$1~magnitude of the brightness of these objects (see \citealt{Davis2021}, hereafter D21). Analogs of the GC objects are also known in the field; an example is the Galactic halo star BD+14$^\circ$3061, along with a few other similar field objects (\citealt{Bond2020}, hereafter B20, and references therein).
In this paper, we report our discovery of a luminous yellow PAGB star in the
Galactic GC M19 (NGC~6273). As described in detail in D21, this object was the only new luminous yellow PAGB star found in our photometric survey of nearly the entire known sample of Galactic
GCs. Thus, it is likely to be the final member of this class in the Galactic GCs.
As PAGB stars continue to evolve, they reach high effective temperatures, and
arrive at the top of the white-dwarf cooling sequence in the CMD\null. These
hot, luminous objects are conspicuous at short wavelengths, especially in the
ground- and space-based ultraviolet (UV). There is a substantial literature on
these prominent, but rare, ``UV-bright'' objects in GCs, beginning with \citet[][hereafter ZNG]{Zinn1972}, and summarized recently by \citet{Moehler2019}, who list about three dozen known or candidate members of the class. Our observations of M19 also revealed a previously unrecognized luminous UV-bright PAGB star, and we briefly discuss this object as well.
\section{Observations \label{sec:observations}}
In the mid-1990s, H.E.B. and collaborators began to develop a ground-based
photometric system optimized for efficient discovery of luminous, low-gravity A,
F, and G-type stars having large Balmer jumps. This
``{\it uBVI\/}'' system combines the \citet{Thuan1976} $u$ filter, whose
bandpass lies almost entirely shortward of the Balmer
discontinuity,\footnote{The Thuan-Gunn $u$ bandpass is distinct from, and
shortward of, that of the $u'$ filter of the Sloan Digital Sky Survey, which
partially overlaps the Balmer jump.} with the
broad-band {\it BVI\/} filters of the standard Johnson-Kron-Cousins system. The
astrophysical motivations and design principles of the {\it uBVI\/} system were
presented by \citet[][hereafter Paper~I]{Bond2005}. \citet[][hereafter Paper~II]{Siegel2005} established a network of equatorial {\it uBVI}\/ standard stars, based on
extensive CCD observations with 0.9-, 1.5-, and 4-m telescopes at Kitt Peak
National Observatory (KPNO) and Cerro Tololo Inter-American Observatory (CTIO).
Over the years 1994 to 2001, H.E.B. made {\it uBVI}\/ observations at KPNO and CTIO
of nearly the entire known sample of Galactic GCs. The primary aim of this
survey was to search systematically for yellow PAGB stars, with a goal of testing
their potential as standard candles and to establish a photometric zero-point.
A forthcoming paper (D21) will present a complete census of stars in the Galactic GC system lying above the HB, bluer than the red-giant branch and AGB, and redder than
$(B-V)_0=-0.1$ in the cluster CMDs. Further papers will describe the
zero-point calibration for yellow PAGB stars, and the results of searches for
them in Local Group galaxies. This paper focuses on the luminous yellow PAGB
star that the survey revealed in M19, as well as a luminous hot UV-bright member of the cluster.
{\it uBVI}\/ survey observations of M19 were made with the CTIO 0.9-m telescope
on 1998 April~21 ($2\times2$ grid centered on the cluster; exposure times at each pointing were $2\times800$, 75, 40, and 45~s, respectively) and April~22 (cluster center; exposure times
400, 75, 40, and 45~s). The field of view of the 0.9-m CCD camera was
$13'\times13'$. Exposure times in the {\it uBVI}\/ filters on the second night were
chosen so as to reach a signal-to-noise ratio of about 200--300 for stars about
2~mag brighter than the HB, allowing for interstellar extinction. The first
night was not photometric, and the exposure times in $u$ were lengthened.
M19 is a populous GC, one of the 10 or so most massive clusters in the Galactic GC system. \citet{Kruijssen2020}, \citet{Pfeffer2020}, and others have argued that it is the remnant nuclear star cluster of a dwarf galaxy (``Kraken''), which was disrupted and accreted by the Milky Way. M19 lies in the Galactic bulge in Ophiuchus ($l=356\fdg9, b=+9\fdg4$), and is overlain by a considerable number of field stars. Early work (e.g., \citealt{Harris1976}) showed that M19 is affected by substantial differential reddening, and that it has a very blue HB, consistent with low metallicity.
The \citet[][hereafter H10]{Harris2010} catalog of GC parameters\footnote{Online version of 2010 December, at \url{http://physwww.mcmaster.ca/~harris/mwgc.dat}} gives a metal content of $\rm[Fe/H]=-1.74$. A recent spectroscopic study \citep{Johnson2017} of over 300 red giants and AGB stars in the cluster showed that there is a range of iron contents among the members, from about $\rm[Fe/H]=-2$ to $-1$, but concentrated around $-1.75$ and $-1.5$. These findings are indicative of considerable self-enrichment in this massive GC.
\section{Data Analysis and Selection of PAGB Candidates \label{sec:data}}
The CCD frames were reduced as described in detail by B16 and D21, using
standard tasks in IRAF\footnote{IRAF was distributed by the National Optical
Astronomy Observatory, operated by the Association of Universities for Research
in Astronomy (AURA) under a cooperative agreement with the National Science
Foundation.} for bias subtraction and flat-fielding. Instrumental stellar
magnitudes were measured on the frames with the {\tt ALLSTAR} and {\tt DAOGROW} tasks in
{\tt DAOPHOT} \citep{Stetson1987}. These data were corrected for atmospheric extinction,
and then the 1998 April~22 frames were calibrated to the {\it uBVI}\/ system using
measurements of standard fields obtained during this observing run. We used the
$u$ magnitudes of standard stars from Paper~II, and $BVI$ magnitudes for the standards of
\citet{Landolt1992}. Since the 1998 April~21 night was not photometric, its
zero-points were determined by scaling to the M19 frames obtained on the
photometric night. Finally, all of the measurements were combined into a single
catalog of mean calibrated magnitudes. For the bright PAGB candidates
considered here, the photometry provides essentially complete stellar samples
into the cluster center (apart from rare cases of badly blended bright stars;
see D21 for further discussion of the sample completeness for our GC survey).
We then searched for candidate yellow PAGB stars by choosing objects that are
simultaneously bright, have large Balmer jumps, and have spatial locations and
proper motions consistent with cluster membership. Figure~\ref{fig:CMDs} illustrates the
selection process. The top two frames show two CMDs for the cluster, after
correcting for a nominal cluster-averaged interstellar extinction of $E(B-V)=0.38$ (from H10), and assuming $R_V=3.1$. (For the extinction corrections in $V-I$, we used the
formula of \citealt{Dean1978}.) The top left panel plots $V_0$ vs.\ $(B-V)_0$,
and the top right plots $V_0$ vs.\ $(V-I)_0$.
As noted in \S\ref{sec:observations}, M19 lies in the Galactic bulge, and there
are substantial numbers of field stars in our CCD frames. To remove most of them
from our {\it uBVI}\/ catalog, we applied three constraints, based on spatial location and on astrometry of the stars given in the {\it Gaia}\/ Data Release~2
(DR2; \citealt{Gaia2016, Gaia2018})\footnote{\url{http://vizier.cfa.harvard.edu/viz-bin/VizieR-3?-source=I/345/gaia2}}:
(1)~angular distance from the cluster center less than three times the half-light radius
(i.e., $3\times79\farcs2$, using the half-light radius given by H10); (2)~DR2
proper motions in right ascension and declination within $1.1\rm\,mas\,yr^{-1}$
of the cluster mean of $(\mu_\alpha,\mu_\delta)=(-3.22,
+1.61)\rm\,mas\,yr^{-1}$ \citep{GaiaGC2018}; and (3)~DR2 parallax less than 0.7~mas (to exclude objects almost certainly in the foreground). The top panels
in Figure~\ref{fig:CMDs} show that these criteria result in a nearly pure sample of cluster
members, especially brighter than $V_0\simeq14$; but there do appear to be a few
remaining field stars below that level.
Also shown in the top left panel is the approximate location of the RR~Lyrae,
Cepheid, and RV~Tauri instability strip (e.g., \citealt{Harris1983}; D21). M19 contains seven
objects listed in the \citet{Clement2001} catalog\footnote{Updated version available online at \url{http://www.astro.utoronto.ca/~cclement/cat/listngc.html}} of variable stars in GCs. However, two of them (V6 and V7) appear to be non-members, based on their proper motions in {\it Gaia}\/ DR2. The positions of the remaining five variables in the CMDs in Figure~\ref{fig:CMDs}, based on our photometry, are plotted as filled green circles. They consist of four Type~II Cepheids\footnote{\citet{Clement1978} point out that the only Galactic GCs having more Type~II Cepheids than M19 are $\omega$~Cen\ and M14.} and one RR~Lyrae variable. Our data are comprised of only a few measurements on two successive nights, made at random pulsation phases, so they should not be regarded as representing the time-averaged locations of the
variables in the CMDs. Nevertheless, the RR~Lyr variable, and three of the
Cepheids, do lie inside, or on the border of, the schematic instability strip.
The red-giant branch (RGB) and HB of the cluster's CMD in Figure~\ref{fig:CMDs} have
appreciable width (compare with, for example, the very narrow RGB and HB of the
nearly unreddened GC M79 in our team's data, shown in Figure~1 of B16). This
spread is partially due to substantial differential interstellar reddening across the face of M19---as we noted above, and has been discussed in several more recent studies of the cluster,
including \citet{Piotto1999}, \citet{Alonso2012}, \citet{Johnson2017},
and references therein to earlier literature.
\begin{figure*}
\begin{center}
\includegraphics[width=2.5in]{m19v-bv.eps}
\hfil
\includegraphics[width=2.5in]{m19v-vi.eps}
\vskip0.2in
\includegraphics[width=3.5in]{m19c2_vi.eps}
\figcaption{\footnotesize
Top panels: Color-magnitude diagrams for M19 in $V,B-V$ and $V,V-I$\null. Our
photometry has been corrected for foreground reddening of $E(B-V)=0.38$, and
{\it Gaia}\/ DR2 proper motions and parallaxes have been used to reject
field stars (see text). The filled orange and blue
circles show the locations of the two luminous PAGB stars belonging to the
cluster. Dashed lines indicate the location of the pulsational instability strip. Filled green circles mark one RR~Lyrae variable (the faintest object) and four Type~II Cepheids.
Bottom panel: Gravity-sensitive $(u-B)-(B-V)$ color index, corrected for
extinction and plotted against the temperature-sensitive $V-I$ color, for stars
brighter than $V_0=14$ (black points). The red curve is a polynomial fit to the location of the HB in a sample of GCs with redder HBs than M19's, taken from D21. The yellow PAGB star (filled orange circle), and the Cepheids, stand out from the HB locus because of their lower gravities and larger Balmer jumps. The blue PAGB star (filled blue circle) is hotter, and lies in the regime where the size of the Balmer jump depends on temperature rather than surface gravity.
\label{fig:CMDs}}
\end{center}
\end{figure*}
As our index of the strength of the Balmer discontinuity, we use the color
difference $(u-B)-(B-V)$, which is a broad-band analog of the $c_1$ Balmer-jump index of
the Str\"omgren $uvby$ system.\footnote{Use of a color difference has the
advantage of only a weak dependence on interstellar extinction; a formula for
the extinction correction as a function of $E(B-V)$ is given in Paper~I.} In the
bottom panel of Figure~\ref{fig:CMDs}, we plot this color difference vs.\ the $(V-I)_0$
temperature index, for the M19 members in the top panels that are brighter than
$V_0=14.0$. At these bright magnitudes, nearly all of the M19 stars are red
giants or AGB stars. The selection criterion for yellow PAGB stars is
that they have a Balmer-jump index significantly larger (i.e., redder) than do
HB stars at the same color, and that they are at least 3~magnitudes brighter than
the HB at the same color. As the top panels show, the HB in M19 is extremely blue, and almost
entirely lacks stars on the cooler ``horizontal'' locus of the HB\null.
Therefore, as an indicator of the HB's location, we use a polynomial fit to the
mean location of the HB in a sample of GCs with redder HBs, taken from D21. This
relation is plotted as a red line in the bottom panel of Figure~\ref{fig:CMDs}.
The peak in this sequence on the left-hand side of the
diagram, at $(V-I)_0\simeq0.1$, is due to the maximum size of the Balmer jump at this color on the HB.
The single RR~Lyrae star in M19 (included in the bottom panel of Figure~\ref{fig:CMDs} in
spite of being fainter than $V_0=14.0$) does lie almost exactly on the mean HB
locus. The four Cepheids fall above the HB relation, consistent with their
higher luminosities, lower surface gravities, and consequent higher
$(u-B)-(B-V)$ color differences. There also appears to be an above-the-HB (AHB)
member of M19 at $V_0\simeq12.7$ with a high value of the color difference;
curiously, although this star apparently lies within the instability strip, it
is not a known variable. This object, and other AHB stars in Galactic GCs, will
be discussed in D21.
The most conspicuous star in the three panels of Figure~\ref{fig:CMDs} is the very bright
object plotted as an orange filled circle. It has a Balmer jump larger than
those of HB stars, as shown in the bottom panel, making it a strong candidate
for a yellow PAGB star.
Also conspicuous in the Figure~\ref{fig:CMDs} CMD is a bright, very blue star. For stars this
hot, the $(u-B)-(B-V)$ color difference is no longer sensitive to surface
gravity (see Paper~I, Figures~4 and~5), becoming instead a temperature index. This
star is a candidate luminous blue PAGB star.
Both objects had been marked as candidate UV-bright stars in the
classical ZNG study; our yellow PAGB star is M19 ZNG~4, and the blue one
is M19 ZNG~2. To our knowledge, the present study is the first to present evidence that both of them are luminous members of the cluster, rather than foreground objects.
The left panel in Figure~2 presents a finding chart for the yellow and blue PAGB
stars in M19, made from one of our $B$-band frames. To illustrate how conspicuous
these two stars are in the ground-based UV, the right panel in
Figure~\ref{fig:charts} shows a $u$-band frame. The two PAGB stars dominate this image. ({The bright star to the NW of the blue PAGB star is the brightest Type~II Cepheid in M19, V1, which was marked as the UV-bright object ZNG~3 in the ZNG study. The
bright star to the NE of the blue PAGB star is ZNG~1, a foreground field
object, as indicated by its {\it Gaia}\/ parallax and proper motion. Two more UV-bright candidates, further from the cluster center, designated ZNG~5 and 6, are also non-members, according to {\it Gaia}\/ astrometry.})
\begin{figure*}[hbt]
\begin{center}
\includegraphics[height=2.7in]{m19_pagb_chart.eps}
\hfil
\includegraphics[height=2.7in]{m19_uband.eps}
\figcaption{\footnotesize
Left panel: finding chart for the yellow PAGB star (orange circle) and the blue
PAGB star (blue circle) in the globular cluster M19. Made from a $B$-band CCD
frame obtained with the CTIO 0.9-m telescope. North is at the top, east on the
left; frame is $2\farcm5$ high.
Right panel: $u$-band frame from same telescope, height $4\farcm8$,
illustrating how conspicuous both stars are in the
ground-based ultraviolet. ({ The bright star to the NE of the blue
PAGB star is a foreground object, ZNG~1. The bright star to the NW is the Type~II Cepheid V1 = ZNG~3.})
\label{fig:charts}
}
\end{center}
\end{figure*}
Table~\ref{table:basicdata} presents basic data for our two PAGB candidates. The {\it Gaia}\/ Early Data Release~3 (EDR3; \citealt{Gaia2020}) became available as we were completing this paper, and is the source for the astrometry given in rows 1 through 5 of the table. The radial velocity (RV) for the yellow PAGB star in row 6 is from {\it Gaia}\/ DR2. Rows 7 through 10 give our {\it uBVI}\/ photometry. The
interstellar reddenings in row~11 are taken from the high-resolution M19
extinction maps of \citet{Alonso2012} and \citet{Johnson2017}. The
former map is based on stellar photometry with the Magellan Telescope, and the
latter on stellar photometry using {\it Hubble Space Telescope\/} ({\it HST})
images.\footnote{The Alonso-Garcia et al.\ reddening map is presented
pictorially in their Figure~13, and in tabular form at
\url{https://vizier.u-strasbg.fr/viz-bin/VizieR-4}. C.~Johnson kindly sent us a
numerical table giving the extinction map depicted in Figure~2 of \citet{Johnson2017}. Unfortunately the existing {\it HST}\/ frames cannot be used for optical photometry of the
bright PAGB stars because their images are saturated.} The reddening values from
these two maps agree well [the differences in $E(B-V)$ at the locations of the two stars are 0.006 and 0.013 mag,
respectively], and we give their means in Table~\ref{table:basicdata}. However, it should be noted
that there are appreciable variations in $E(B-V)$ from cell to cell in both
extinction maps, and it is difficult to estimate the uncertainty in the
reddening for an individual star. We adopt $\pm$0.02 mag as a reasonable guess.
\begin{deluxetable*}{lccc}[ht!]
\tablewidth{0 pt}
\tablecaption{Basic Data for the Luminous Post-AGB Stars in M19
\label{table:basicdata}
}
\tablehead{
\colhead{Parameter} &
\colhead{``Yellow'' PAGB} &
\colhead{``Blue'' PAGB} &
\colhead{Source\tablenotemark{a}} \\
\colhead{} &
\colhead{(M19 ZNG~4)} &
\colhead{(M19 ZNG~2)} &
\colhead{}
}
\startdata
R.A. [J2000] & 17:02:35.186 & 17:02:39.155 & (1) \\
Dec. [J2000] & $-26$:15:24.14 & $-26$:15:29.36& (1) \\
Parallax [mas] & $+0.119\pm0.016$ & $+0.158\pm0.024$ &(1) \\
R.A. proper motion [$\rm\,mas\,yr^{-1}$] & $-2.878\pm0.019$ & $-2.990\pm0.025$ &(1) \\
Dec. proper motion [$\rm\,mas\,yr^{-1}$] & $+1.146\pm0.012$ & $+1.454\pm0.018$ &(1) \\
Radial velocity [${\>\rm km\>s^{-1}}$] & $+142.3\pm1.4$ & $\dots$ & (2) \\
$V$ & $12.512\pm0.006$ & $13.291\pm0.006$ & (3) \\
$u-B$ & $1.471\pm0.012$ & $0.608\pm0.015$ & (3) \\
$B-V$ & $0.795\pm0.010$ & $0.211\pm0.010$ & (3) \\
$V-I$ & $1.021\pm0.011$ & $0.390\pm0.010$ & (3) \\
Reddening, $E(B-V)$ & $0.344\pm0.020$ & $0.342\pm0.020$ & (4) \\
Absolute magnitude, $M_V$ & $-3.39\pm0.09$ & $-2.61\pm0.09$ & (5) \\
Absolute luminosity, $\log L/L_\odot$ & 3.24 & 3.22 & (6)
\enddata
\tablenotetext{a}{Sources: (1)~{\it Gaia}\/ EDR3; (2)~{\it Gaia}\/ DR2; (3)~This paper (not corrected for extinction); note that the zero-point for the $u$ magnitude is such that $u=1.0$ for Vega; (4)~From the reddening maps of \citet{Alonso2012} and \citet{Johnson2017} (see the text, \S\ref{sec:data}); (5)~This paper, calculated from data in this table, $R_V=3.1$, and a distance modulus of $(m-M)_0=14.84\pm0.06$~mag (see the text, \S\ref{sec:absmag}); (6)~This paper, calculated from data in this table, and bolometric corrections from \citet{Castelli2004} (see the text, \S\ref{sec:seds}).}
\end{deluxetable*}
\bigbreak
\section{Cluster Distance and Visual Absolute Magnitudes \label{sec:absmag}}
To convert the apparent magnitudes in row~7 of Table~\ref{table:basicdata} to absolute magnitudes, we need the distance to the cluster. For M19, distance determinations from photometric methods are unusually problematic, for at least two reasons. First, the relatively large and spatially variable reddening makes the correction for extinction difficult. Second, distance methods based on the HB suffer because, in the case of M19 (as shown in Figure~\ref{fig:CMDs}), the cluster's ``horizontal'' branch is nearly vertical in CMDs made at optical wavelengths.
With the recent availability of {\it Gaia}\/ EDR3, it is also possible to obtain a direct geometric distance estimate. We considered the sample of 202 RGB cluster members brighter than $V_0=14$, which we had chosen for the bottom panel in Figure~\ref{fig:CMDs}. The mean EDR3 parallax for this sample, and its standard error, are $0.0934\pm0.0062$~mas. To this we applied the global zero-point offset of +0.017~mas from the analysis of \citet{Lindegren2020}.
In Table~\ref{table:distance} we list six distance determinations, made using a variety of essentially independent techniques, as summarized in the notes to the table. We see no compelling reason to adopt one of these determinations, so we will use a weighted mean of the four determinations for which uncertainties are given (and which, as it happens, span the range of distances found by these studies). This results in our adopted distance modulus of $(m-M)_0=14.84\pm0.06$ ($d=9.29\pm0.26$~kpc).
Row 12 in Table~\ref{table:basicdata} gives the visual absolute magnitudes of both stars,
calculated from our photometry, the interstellar extinction given in row~11
(using $R_V=3.1$), and the distance modulus adopted in this section. The stated formal uncertainties are likely somewhat optimistic, given the possibilities of systematic errors in the extinction and distance.
\begin{deluxetable}{ccl}[htb]
\tablewidth{0 pt}
\tablecaption{M19 Distance Determinations
\label{table:distance}
}
\tablehead{
\colhead{Distance [kpc]} &
\colhead{$(m-M)_0$} &
\colhead{Source}
}
\startdata
$8.99\pm0.83$ & $14.77\pm0.20$ & \citet{Piotto1999}\tablenotemark{a} \\
$9.95\pm0.37$ & $14.99\pm0.08$ & \citet{Recio2005}\tablenotemark{b} \\
8.24 & 14.58 & \citet{Valenti2007}\tablenotemark{c} \\
8.79 & 14.72 & \citet{Harris2010}\tablenotemark{d} \\
$8.13\pm0.47$ & $14.55\pm0.13$ & \citet{Baumgardt2019}\tablenotemark{e} \\
$9.06\pm0.51$ & $14.78\pm0.12$ & This paper, {\it Gaia}\/ EDR3\tablenotemark{f} \\
\enddata
\tablenotetext{a}{Distance determined from luminosity of red-giant bump in {\it HST}-based $BV$ color-magnitude diagram.}
\tablenotetext{b}{Distance calculated by us by subtracting $3.1\,E(B-V)$ from the value of $(m-M)_{\rm F555W}=16.04\pm0.08$, and adopting $E(B-V)=0.34$, as given by \citet{Recio2005}.}
\tablenotetext{c}{Distance determined from near-IR photometry of stars on the red-giant branch. Uncertainty not given.}
\tablenotetext{d}{Distance calculated by us by subtracting $3.1\,E(B-V)$ from the value of $(m-M)_V=15.90$, and adopting $E(B-V)=0.38$, as given by \citet{Harris2010}. Uncertainty not given.}
\tablenotetext{e}{``Kinematic'' distance derived from cluster-member proper-motion and radial-velocity dispersions and $N$-body modeling.}
\tablenotetext{f}{Calculated by us based on the mean {\it Gaia}\/ EDR3 parallax of a sample of cluster red giants, as described in \S\ref{sec:absmag}.}
\end{deluxetable}
\bigbreak \bigskip
\section{Cluster-Membership Tests}
Both of the PAGB candidates lie well within the spatial boundaries of the cluster (see Figure~\ref{fig:charts}): the yellow and blue stars fall $50''$ and $36''$ from the cluster center, respectively, which are within the half-light radius of $79\farcs2$. Moreover, they have {\it uBVI}\/ colors that are extremely unusual for field stars. Thus it is already highly probable that they are members of the
cluster.
Two further membership tests are possible: (1)~RV, and (2)~proper
motion. Parallax is a less useful criterion for individual stars at the large adopted distance of M19, which
corresponds to a parallax of only 0.108~mas. Nevertheless, the {\it Gaia}\/ EDR3 parallaxes for
both stars, given in row~3 of Table~\ref{table:basicdata}, corrected for the +0.017 zero-point offset \citep{Lindegren2020}, agree with the nominal value to within 1.8 and 2.8$\sigma$, respectively.
{\it Gaia}\/ EDR3 did not give RVs for either star, but the earlier DR2 listed a RV of $+142.3\pm1.4\,{\>\rm km\>s^{-1}}$ for the yellow PAGB
star. This agrees extremely well with the mean cluster RV of
$+145.54\pm0.59\,{\>\rm km\>s^{-1}}$ \citep{Baumgardt2019}; the velocity dispersion in the cluster is $11.0{\>\rm km\>s^{-1}}$ \citep{Baumgardt2018}. {\it Gaia}\/ DR2 and EDR3 did not list a RV
for the early-type blue PAGB candidate.
To make proper-motion membership tests for the two stars, we selected nearly pure samples of M19 members from
{\it Gaia}\/ EDR3. We chose stars lying within $30''$ of each object, brighter than magnitude
$G=18$, redder than $BP-RP=1.2$, and having a parallax less than 1~mas. The proper motions for these
samples are plotted as black points in the two panels in Figure~\ref{fig:pm}, with the proper motions of
the PAGB candidates themselves (rows~4 and 5 in Table~\ref{table:basicdata}) marked with orange and
blue open circles. The {\it Gaia}\/ CMDs of these two samples indicate that a large
majority of the stars are cluster members lying on the RGB\null. In both cases, the motions of the
PAGB candidates are well within the distributions of the cluster members. To illustrate the proper-motion distribution in the surrounding field, we selected stars from EDR3 in a nearby field, with a radius of $200''$ and the same criteria. These are plotted as red points in the two figures; they show a wide range of proper motions, most of them in fact outside the range plotted.
In summary, the available evidence strongly confirms that both stars are
physical members of M19.
\begin{figure*}[hbt]
\begin{center}
\includegraphics[width=3.0in]{yellow_pm.eps}
\hfil
\includegraphics[width=3.0in]{blue_pm.eps}
\figcaption{\footnotesize
Left panel: Proper motions from {\it Gaia}\/ EDR3 for stars within $30''$ of the yellow PAGB star in M19, with parallax less than 1~mas, brighter than $G=18$, and $BP-RP$ color redder than 1.2 (black points). The color-magnitude diagram of these stars (not shown) indicates that nearly all of them are cluster members on the red-giant branch. The proper motion of the yellow PAGB star is marked with an orange circle.
Right panel: Proper motions from EDR3 for stars within $30''$ of the blue PAGB star in M19, selected with the same criteria (black points). The proper motion of the blue PAGB star is marked with a blue circle. In both panels the red points show the proper motions of stars in a nearby field, selected using the same criteria. The proper motions of both stars are well within the distributions of cluster members.
\label{fig:pm}
}
\end{center}
\end{figure*}
\section{Variability}
As shown in the top left panel in Figure~\ref{fig:CMDs}, the M19 yellow PAGB star lies close
to the instability strip in the CMD\null. Several field analogs of the yellow
PAGB stars in GCs are known to be low-amplitude semi-regular variables,
including HD~46703 and BD+14$^\circ$3061. As discussed in B20 (and references therein), both
field stars have typical pulsation periods of $\sim$29--32~days, and
peak-to-peak amplitudes of $\sim$0.1--0.3~mag. The pulsation amplitudes are variable; in
the case of BD+14$^\circ$3061, the variations can actually drop below detectability for
extended intervals.
The most recent extensive search for variable stars in M19 of which we are aware
is the photographic study by \citet{Clement1978}; this investigation
did not mark the yellow PAGB star as variable. Our own {\it uBVI}\/ data are of
limited value, since we only have the two epochs in 1998 April separated by one
day, plus an earlier lower-quality CCD observation obtained with the CTIO 1.5-m
in 1995, which we did not include in our calibrated photometric reductions. We
see no convincing evidence for variability in this limited material at a level
of more than a few hundredths of a magnitude. The available data from current
all-sky monitoring programs with small telescopes are generally of limited use,
because of the crowding within the cluster. The uncertainty given for the $G$ magnitude in {\it Gaia}\/ EDR3 is very small, consistent with little or no variability.
Thus there is no evidence that the yellow PAGB star is variable, but additional
monitoring observations with sufficient spatial resolution would be useful.
(There is also no evidence that the blue PAGB star is variable, based on the
same material.)
\section{Spectral-Energy Distributions \label{sec:seds}}
We determined SEDs for both PAGB stars by combining our optical photometry from Table~\ref{table:basicdata} with public data from the following sources:
(1)~The {\it Galaxy Evolution Explorer\/} ({\it GALEX}) \citep{Morrissey2007} imaged M19 in its all-sky UV survey,\footnote{\url{https://galex.stsci.edu/GalexView/}} but only in its far-UV (FUV) bandpass. Although our blue PAGB star is prominent at FUV wavelengths, it is not contained in the {\it GALEX}\/ source catalog (presumably because of crowding near the cluster center).
We obtained the FUV image and, because the blue PAGB
is so much brighter than its neighboring objects at FUV wavelengths, we measured its
brightness using simple aperture photometry. We then translated these raw magnitudes
into the standard flux system by scaling our measurements to similar aperture
photometry of nearby isolated field stars with magnitudes given in the {\it GALEX}\/
source catalog.
The yellow PAGB star was too faint and blended to be measured in the FUV.
(2)~The {\it Hubble Space Telescope\/} ({\it HST}) has imaged M19 on a few occasions,\footnote{\url{https://archive.stsci.edu/}} but in nearly all of the frames the images of both PAGB stars are heavily saturated. We found an exposure of the blue PAGB star obtained with the Wide Field Planetary Camera~2 (WFPC2) in the near-UV (NUV) F255W filter (program GO-8718; PI G.~Piotto), in which only the central pixel was saturated; we performed aperture photometry to obtain a lower limit on the star's flux. For the yellow PAGB star, there are 10 WFPC2 frames in F255W (GO-10815; PI T.~Brown) with unsaturated images, on which we also performed aperture photometry. We used the {\tt PHOTFLAM} keyword in the image headers to convert the counts to absolute fluxes, and applied the standard 0.10~mag correction to an infinite aperture. The {\it HST}\/ images are virtually unaffected by source crowding.
(3)~The UV/Optical Telescope (UVOT) on the Neil Gehrels {\it Swift\/} Observatory obtained images\footnote{\url{https://archive.stsci.edu/swiftuvot/}} of M19 in 2009 and 2010 in its three NUV/FUV bandpasses, $uvw1$, $uvm2$, and $uvw2$. We combined these images, and then performed PSF-fitting photometry of both PAGB stars using {\tt DAOPHOT}\null. Corrections for time-dependent sensitivity loss, coincidence losses, and exposure time were applied, as detailed in \citet{Siegel2014}; the data were then calibrated to the absolute photometric system of \citet{Breeveld2011}. The $uvw1$ and $uvw2$ filters have substantial red leaks, which are problematic for cool stars; because of this, exacerbated by the considerable interstellar reddening, we did not include the observations of the yellow PAGB star in these two bandpasses in our analysis.
(4)~The 2MASS near-infrared (NIR) sky survey \citep{Skrutskie2006} obtained images of M19 in $J$, $H$, and $K_s$. Photometry of our PAGB stars is contained in the 2MASS source catalog,\footnote{\url{https://irsa.ipac.caltech.edu/cgi-bin/Gator/nph-dd}} but is clearly affected by the stellar crowding near the cluster center. We therefore downloaded the 2MASS frames, and performed {\tt DAOPHOT} PSF photometry on the images, calibrating to the 2MASS photometric zero-point using measurements of nearby isolated stars contained in the source catalog. {\it HST}\/ images in the $I$ (F814W) band show that the yellow PAGB star, at the 2MASS resolution, is blended with four nearby red giants. We attempted to de-blend the images, using the {\it HST}\/ frame as a prior to define the precise locations of the PAGB stars in the 2MASS frames, but with limited success.
(5)~Images of M19 were obtained in 2013 with the warm {\it Spitzer Space Telescope\/} and its Infrared Array Camera (IRAC), in the 3.6 and $4.5\,\mu$m bandpasses; the observations were part of a program aimed at RR~Lyrae stars (PI W.~Freedman). We downloaded a selection of these images\footnote{\url{https://sha.ipac.caltech.edu/applications/Spitzer/SHA/}} and performed {\tt DAOPHOT} photometry to de-blend the stellar images, using PSF fitting radii of the order of the FWHM\null.
Aperture corrections were then applied, based on the total magnitudes of isolated
field stars in the frames.
(6)~We examined NIR and far-IR images of M19 from the {\it Wide-Field Infrared Survey Explorer\/} ({\it WISE}\/) sky survey, which we downloaded from SkyView.\footnote{\url{https://skyview.gsfc.nasa.gov/current/cgi/query.pl}} However, the spatial resolution of these frames is far too poor to provide useful photometry for our targets lying in very crowded regions of the cluster.
We then corrected all of the measured fluxes for interstellar extinction, using the values of $E(B-V)$ in Table~\ref{table:basicdata}, and for most bandpasses applying the formulae of \citet{Cardelli1989}, calculated at the effective wavelength of each bandpass, and assuming $R_V=3.1$. For the UVOT photometry, because of the complex structure of the extinction curve around 2200~\AA, we determined the extinction corrections by integrating the Cardelli et al.\ formula convolved with the system throughput curves and a blackbody function having the approximate temperature of each star. For the two IRAC bandpasses, whose wavelengths are beyond the range of the Cardelli et al.\ formulae, we determined the extinction at the $K_s$ band, and scaled it to [3.6] and [4.5] according to the relations given by \citet{Indebetouw2005}. The resulting extinction-corrected fluxes must be regarded as only approximate, given the very large correction factors in the UV, the uncertainty whether $R_V=3.1$ is appropriate for this line of sight, and the source crowding in many of the bandpasses.
The two panels in Figure~\ref{fig:sed} plot the SEDs for the two PAGB stars. We superpose model-atmosphere SEDs selected from the ATLAS9 grid\footnote{\url{http://wwwuser.oats.inaf.it/castelli/grids.html}} of \citet{Castelli2004}. The metal abundances for these SEDs are $\rm[M/H]=-1.5$, with $\alpha$-element enhancements. We adopted surface gravities of $\log g=0.5$ and 2.0 for the yellow and blue PAGB stars, respectively. The best fits to the SEDs are found for effective temperatures of 6250 and 11750~K, respectively, but we caution that there is some degeneracy between the adopted $T_{\rm eff}$ values and reddenings.
Using the bolometric corrections for the two Castelli \& Kurucz models, we find the absolute luminosities given in the final row of Table~\ref{table:basicdata}, $\log L/L_\odot=3.24$ and 3.22 for the yellow and blue PAGB stars, respectively. Thus their luminosities are nearly identical, and they are likely to be on very similar post-AGB evolutionary tracks.
\begin{figure*}[hbt]
\begin{center}
\includegraphics[width=3.3in]{yellow_sed.eps}
\hfill
\includegraphics[width=3.3in]{blue_sed.eps}
\figcaption{\footnotesize
Left panel: Spectral-energy distribution for the M19 yellow PAGB star (filled black circles), corrected for interstellar reddening of $E(B-V)=0.344$ as described in the text. The orange curve is a model-atmosphere SED for a star with the parameters indicated in the figure. Note the conspicuously large Balmer discontinuity.
Right panel: SED for the M19 blue PAGB star (filled black circles), corrected for $E(B-V)=0.342$. The {\it HST}\/ observation at F255W (cyan plotting point and label) is a lower limit (see text). The blue curve is a model-atmosphere SED with parameters indicated in the figure. See the text for details of the various public sky surveys used to assemble these SEDs, and caveats about uncertainties due to blending and the extinction corrections.
\label{fig:sed}
}
\end{center}
\end{figure*}
A primary aim in investigating the SEDs was to search for evidence of circumstellar dust, ejected during the AGB phase. In general, circumstellar dust has not been found to be prominent in yellow and blue PAGB stars in old populations, such as those in GCs (e.g., B16 and B20), possibly because the PAGB evolutionary timescales are slow enough for any ejecta to have dissipated, combined with the difficulty of forming dust in a low-metallicity environment. As Figure~\ref{fig:sed} shows, we likewise see no evidence for warm circumstellar dust in the two M19 PAGB stars. Unfortunately, the available observations only go out to the {\it Spitzer}\/ [4.5] bandpass; the stellar crowding in the cluster precludes photometry at longer wavelengths with instruments such as {\it WISE}\null. Thus the constraints on dust are not very tight.
\goodbreak
\section{Discussion and Future Studies}
\subsection{Yellow PAGB Stars as Standard Candles}
The main goal of our {\it uBVI}\/ survey was to search the Galactic GC system for yellow PAGB stars and to test them as potential Population~II standard candles. The yellow PAGB star in M19 reported in this paper joins the small number of known objects of this class. We find it to have a visual absolute magnitude of $M_V=-3.39\pm0.09$, based on several recent determinations of the distance to M19. The main contributors to the error are the uncertainties in the cluster's distance, and in the foreground extinction.
B16 listed (their Table~4) the visual absolute magnitudes of the four other known non-variable yellow PAGB stars in Galactic GCs; they range from $M_V=-3.10$ to $-3.46$. Adding our new result for the M19 star, and including also the absolute magnitude for the field yellow PAGB star BD+14$^\circ$3061\ from B20, we find that these six objects have a mean absolute magnitude of $M_V=-3.35\pm0.06$, with a standard deviation of 0.14~mag. This is a very narrow luminosity function, compared for example to that of Population~I Cepheids at a given pulsation period (e.g., \citealt{Clementini2019}). And the discovery and measurement of yellow PAGB stars requires only a single observation epoch. Thus, we believe that these stars continue to be potential extragalactic distance indicators.\footnote{If we had used its yellow PAGB star to estimate the distance to M19, using the zero-point of $M_V=-3.38\pm0.05$ found earlier by B16, we would have obtained $(m-M)_0=14.83\pm0.08$. This agrees extremely well with the mean distance modulus from several independent methods of $(m-M)_0=14.84\pm0.06$, which we discussed in \S\ref{sec:absmag}.} In several forthcoming papers, we will further explore this possibility.
We also note that the five yellow PAGB stars in Galactic GCs all belong to clusters with ``intermediate'' metallicities, of $\rm[Fe/H]=-1.59$ (NGC\,5986), $-1.60$ (M79), $-1.74$ (M19), and $-1.53$ ($\omega$~Cen), as tabulated by H10. (However, M19 [see \S\ref{sec:observations}] and $\omega$~Cen\/ have fairly wide ranges of [Fe/H] among their members.) Moreover, all four clusters contain very blue HB stars. We will discuss these and other clues to the evolutionary status of these stars in more detail in papers now in preparation.
\subsection{Future Work}
There are several desirable future investigations of these two new and relatively bright PAGB stars in M19. First, moderate-resolution spectra should be obtained in order to confirm even more definitively that both stars are indeed low-metallicity cluster members, consistent with the interpretations in this paper. High-resolution spectroscopic abundance studies of both of them would be of considerable interest, given the peculiarities seen in other PAGB stars. For example, \citet{Sahin2009} found an anomalously low iron abundance in the yellow PAGB star in M79, and several PAGB stars in younger populations show extreme depletions of refractory chemical elements with high condensation temperatures (see \citealt{Oomen2019} and references therein). Several field PAGB stars appear to be long-period spectroscopic binaries (B20 and references therein), so RV monitoring would be useful to investigate whether binary interactions play a role in the evolution of these objects. Effective temperatures of the stars determined from model-atmosphere fitting, compared to the observed colors, would help refine the estimates of reddening, and thus of their absolute magnitudes. High-precision photometric monitoring would be useful to test whether the yellow PAGB star is a low-amplitude pulsating variable. High-spatial-resolution MIR photometry would provide tighter limits on circumstellar dust than was possible in the study reported here.
\acknowledgments
The observational work in 1998 was partially supported by NASA grant NAG 5-6821 under the ``UV, Visible, and Gravitational Astrophysics Research and Analysis'' program, and by the Director's Discretionary Research Fund at STScI\null. H.E.B. thanks the staff at Cerro Tololo for their support over the years.
This work has made use of data from the European Space Agency (ESA) mission
{\it Gaia\/} (\url{https://www.cosmos.esa.int/gaia}), processed by the {\it Gaia\/}
Data Processing and Analysis Consortium (DPAC,
\url{https://www.cosmos.esa.int/web/gaia/dpac/consortium}). Funding for the DPAC
has been provided by national institutions, in particular the institutions
participating in the {\it Gaia\/} Multilateral Agreement.
Based in part on observations made with the NASA Galaxy Evolution Explorer.
{\it GALEX}\/ was operated for NASA by the California Institute of Technology under NASA
contract NAS 5-98034.
Based in part on observations made with the NASA\slash ESA {\it Hubble Space Telescope}, obtained from the data archive at the Space Telescope Science Institute. STScI is operated by the Association of Universities for Research in Astronomy, Inc.\ under NASA contract NAS 5-26555.
We acknowledge the use of public data from the Neil Gehrels {\it Swift}\/ Observatory data archive.
This publication makes use of data products from the Two Micron All Sky Survey,
which is a joint project of the University of Massachusetts and the Infrared
Processing and Analysis Center/California Institute of Technology, funded by the
National Aeronautics and Space Administration and the National Science
Foundation.
This work is based in part on observations made with the {\it Spitzer Space Telescope}, operated by the Jet Propulsion Laboratory, California Institute of Technology under a contract with NASA.
It also makes use of data products from the {\it Wide-field Infrared Survey Explorer}, which is a joint project of the University of California, Los Angeles, and the Jet Propulsion Laboratory/California Institute of Technology, funded by the National Aeronautics and Space Administration.
\facilities{CTIO 0.9m, Gaia, GALEX, 2MASS, Swift, HST, Spitzer, WISE}
|
\section{Motivation}
\label{sec:intro}
Knowledge of galaxy redshifts $z$ is essential for understanding their evolutionary processes and for utilizing them as cosmological tracers of large-scale structure.
Spectral observations of absorption and emission lines enable the high-confidence measurement of spectroscopic redshifts (spec-$z$}%\nospace s), however, such measurements are costly and preclude large samples of high-redshift galaxies.
When high-resolution spectra are unavailable, broad-band photometry that is sensitive to the spectral continuum can be used to estimate galaxy redshifts \citep{baum_photoelectric_1962}.
Though inherently noisy, the relationship between redshift and photometry can be established from a library of spectral energy distribution (SED) templates or a representative sample of galaxies with known redshifts.
By boosting the cosmological sample size, especially at high redshift, photo-$z$}%\nospace s have directly enabled the era of precision cosmology derived by weak gravitational lensing tomography and baryon acoustic oscillation peak measurements.
However, photo-$z$}%\nospace s are by their very nature imprecise and inaccurate compared to their spectroscopic counterparts.
Per-galaxy photo-$z$}%\nospace\ probability density functions (PDFs), defined over all possible redshifts and usually denoted as $\mathrm{P}(z)$, better encapsulate the nontrivial uncertainty landscape \citep{koo_photometric_1999}.
The most common application of photo-$z$}%\nospace\ PDFs is their use in estimating the redshift distribution function $\mathcal{N}(z)$\ of a sample of galaxies, a quantity essential to cosmological parameter constraints via the power spectra of weak gravitational lensing and large-scale structure \citep{mandelbaum_precision_2008, sheldon_photometric_2012, bonnett_redshift_2016, hildebrandt_kids-450_2017}.
When photo-$z$}%\nospace\ PDFs are available instead of true redshifts $z^{\dagger}$, the simplest approach reduces each $\mathrm{P}(z)$ to a point estimate $\hat{z}$ of redshift by using $\delta(z, \hat{z})$ as a substitute for the unknown (and unknowable) $\delta(z, z^{\dagger})$, where $\delta(z, z') \equiv \{\infty, z = z';\ 0, z \neq z'$ is the Dirac delta function between a random variable $z$ and a particular value $z'$ thereof, normalized such that $\int \delta(z, z') dz = 1$.
An intuitive approach to circumventing the loss of valuable knowledge of redshift uncertainty resulting from reduction of photo-$z$}%\nospace\ PDFs to point estimates is to average their photo-$z$}%\nospace\ PDFs according to the \textit{stacked estimator} of the redshift distribution function \citep{lima_estimating_2008},
\begin{align}
\label{eqn:stack}
\hat{\mathcal{N}}(z) &\equiv \sum_{i = 1}^{N} \mathrm{P}(z_{i}) ,
\end{align}
of a sample of $N$ galaxies $i$.
A number of popular extensions to \Eq{eqn:stack} are addressed in \Sect{sec:math}.
Unfortunately, the stacked estimator $\hat{\mathcal{N}}(z)$ of the redshift distribution is mathematically incorrect \citep{hogg_data_2012}.
Even under simplifying assumptions, stacking yields a biased estimator of $N(z)$.
Consider a scenario in which the true redshift distribution $\mathcal{N}^{\dagger}(z) = \delta(z, z^{\dagger})$ is a delta function at $z^{\dagger}$ and each photo-$z$}%\nospace\ PDF $\mathrm{P}(z_{i}) = \mathrm{N}(\hat{z}_{i}, \sigma^{2})$ is a Gaussian distributions with a shared variance $\sigma^{2}$ and a mean $\hat{z}_{i} \sim \mathrm{N}(z^{\dagger}, \sigma^{2})$ drawn from a Gaussian distribution centered at the shared true redshift $z^{\dagger}$ with the same shared variance $\sigma^{2}$;
in such a situation, $\hat{\mathcal{N}}(z)$ cannot be equal to $\mathcal{N}^{\dagger}(z)$.
Mathematically principled methodologies for recovering the redshift distribution, including those that simultaneously infer the photo-$z$}%\nospace\ PDFs directly from photometry \citep{leistedt_hierarchical_2016, leistedt_hierarchical_2019} and those that use an existing set of photo-$z$}%\nospace\ PDFs \citep{malz_how_2020, rau_composite_2021}, have been proposed and validated.
Despite the availability of robust methodological alternatives, however, \Eq{eqn:stack} and its close cousins have remained prevalent in modern cosmological analysis pipelines.
In addition to the understandable inertia and expected technical challenges of adopting new approaches, stacking is bolstered by pervasive misconceptions about the causal structure of the problem of redshift inference and photo-$z$}%\nospace\ PDFs as probabilistic objects overall \citep{gruen_combining_2017, jarvis_open_2018, malz_re:_2018}.
This letter is pedagogical in nature and serves as a companion to \citep{malz_how_2020}, which quantifies the detrimental consequences of stacking, derives a mathematically self-consistent alternative methodology for constraining the redshift distribution from a sample of photo-$z$}%\nospace\ PDFs, and provides a public implementation thereof.
Here, we complementarily employ a series of thought experiments to motivate the paradigm shift whose necessity is demonstrated in \citep{malz_how_2020}.
We present a flexible mathematical framework in \Sect{sec:preamble}, derive the general form of the redshift distribution in \Sect{sec:math}, explore limiting cases of the general form in \Sect{sec:results}, and interpret the implications thereof in \Sect{sec:disco}.
\section{Definitions}
\label{sec:preamble}
To answer the question of when stacking can recover the true redshift distribution, we must first dust off our knowledge of probability and explicitly connect these classic concepts to the problem at hand.
We begin by considering the chance that a single galaxy's true redshift $z^{\dagger}_{i}$ takes some reference value $z'$.
\begin{definition}\label{def:binarystatespace}
The \textit{outcome space} $\Omega(z^{\dagger}_{i})$ of the true redshift $z_{i}^{\dagger}$ of a single-galaxy $i$ is $-\infty < z^{\dagger}_{i} < \infty$, although in cosmology we can safely assume $z \geq 0$.
\end{definition}
\begin{definition}\label{def:event}
An \textit{event} is a collection of possible outcomes;
in our case, the experiment may be the natural occurence of the true redshift $z_{i}^{\dagger}$ of a galaxy $i$, though it may be unknown to us.
\end{definition}
\begin{definition}\label{def:disjoint}
The possible outcomes in $\Omega(z_{i}^{\dagger})$ are \textit{disjoint} if one occurring means that all others cannot occur;
our single galaxy $i$ cannot satisfy both ${z^{\dagger}_{i} = z'}$ and ${z^{\dagger}_{i} = \lnot z'}$.
\end{definition}
\begin{definition}\label{def:pdens}
The \textit{probability density function (PDF)} ${\mathrm{P}}(z^{\dagger}_{i} = z') \geq 0$ is the chance that a galaxy's redshift $z_{i}^{\dagger}$ takes the value $z'$ if $z$ is continuous.
\end{definition}
Though redshift $z$ is continuous in reality, we assume, for the pedagogical purposes of this letter, that redshift $z$ is a discrete random variable ${z_{i}^{\dagger} \in \{z', \lnot z'\}}$ that can take either the value $z'$ of some reference redshift or any other value of redshift $\lnot z'$.
This assumption of a binary discrete outcome space is made without loss of generality, as all results may be straightfowardly extended to the limit of continuous redshift;
guidance to relevant proofs for such extensions is provided throughout \Sect{sec:math}.
\begin{definition}\label{def:pmass}
In our pedagogical example of the binary discrete outcome space, ${\pr{z^{\dagger}_{i} = z'} \equiv \int_{z'-\epsilon}^{z'+\epsilon} \mathrm{P}(z^{\dagger} = z) \mathrm{d}z \geq 0}$, for a very small $0 \leq \epsilon \ll 1$, is instead a \textit{probability mass function (PMF)};
it must be noted that this integral over a continuous PDF $\mathrm{P}(z)$ respects the physical units of redshift, whereas the resulting discrete PMF $\pr{z}$ is unitless \cite{hogg_data_2012}.
\end{definition}
\begin{definition}\label{def:normalization}
A PMF must satisfy the \textit{normalization} condition ${\sum_{z} \pr{z^{\dagger}_{i} = z}} = 1$;
the PMF over our binary outcome space $\Omega(z^{\dagger}_{i})$ satisfies ${\pr{z^{\dagger}_{i} = z'} = 1 - \pr{z^{\dagger}_{i} = \lnot z'}}$.
\end{definition}
The above definitions pertain to the redshift of a single galaxy $i$, but this letter concerns the distribution $\mathcal{N}(z)$\ of redshifts of an ensemble $S$ of $|S| = N$ galaxies, which is some pre-defined sample that need not include all galaxies in the universe.
A small window about a single redshift $z' \pm \epsilon$ for small $0 < \epsilon \ll 1$ contains the true redshift of some whole number ${\mathcal{N}(z') = K^{\dagger}}$ of those galaxies.
The discrete outcome space ${\Omega(K^{\dagger}) = \{K^{\dagger} = 0, \dots, K^{\dagger} = N\}}$ of the true number $K^{\dagger}$ of galaxies with true redshift $z'$ thus has $N + 1$ elements, and there is some probability ${\pr{K^{\dagger} = K}}$ of each value of $K$ that must be related to the \pz\ PMF s ${\{\pr{z^{\dagger}_{i} = z'}\}_{i = 1, \dots, N}}$.
\begin{definition}\label{def:conditional}
We may consider the \textit{conditional} probability that an event occurs given that another event occurs;
for example, if ${z^{\dagger}_{i} = z'}$, then ${\pr{K^{\dagger} = 0 \mid z^{\dagger}_{i} = z'} = 0}$.
\end{definition}
\begin{definition}\label{def:independence}
Multiple events are \textit{statistically independent} if the probability of each occurring does not affect the probability that the others occur;
for this work, we broadly assume the statistical independence ${\pr{z^{\dagger}_{i}} \perp \pr{z^{\dagger}_{h \neq i}}}$ of the redshifts of individual galaxies $h, i \in S$ from an arbitrary set $S$ of galaxies.
\end{definition}
\begin{definition}\label{def:intersection}
In a universe with an ensemble $S$ of $|S| = N$ galaxies whose redshifts are statistically independent, the \textit{intersection} of the probability that they all have a specific redshift $z'$, which is also the probability that the number of galaxies with redshift $z'$ is equal to the total number of galaxies ${\pr{K^{\dagger} = N}}$, is the product of the probabilities that each has that redshift: ${\pr{z^{\dagger}_{1} = z' \cap \dots \cap z^{\dagger}_{N} = z'} = \pr{z^{\dagger}_{1} = z'} \times \dots \times \pr{z^{\dagger}_{N} = z'}}$.
\end{definition}
Definition~\ref{def:intersection} implies that the stacked estimator $\hat{\mathcal{N}}(z)$ of the redshift distribution given by Equation~\ref{eqn:stack} can be equal to the true redshift distribution if ${\pr{z^{\dagger}_{1} = z' \cap \dots \cap z^{\dagger}_{N} = z'}} = {\pr{z^{\dagger}_{1} = z'} + \dots + \pr{z^{\dagger}_{N} = z'}}$, which does not directly follow from any combination of the above definitions and in fact may well violate Definition~\ref{def:normalization}.
Because Equation~\ref{eqn:stack} yields a single estimate $\hat{K}$ rather than the probability distribution $\pr{K}$ over the value of the redshift distribution, we work with a summary statistic of $\pr{K}$ to make an apples to apples comparison.
\begin{definition}\label{def:expected}
The \textit{expected value} of a continuous random variable such as $z$ is its first moment ${\langle z_{i}^{\dagger} \rangle \equiv \int_{-\infty}^{\infty}z'\mathrm{P}(z_{i}^{\dagger} = z')\mathrm{d}z'}$;
the expected value of a discrete random variable such as $K$ is defined as ${\langle K^{\dagger} \rangle \equiv \sum_{K' = 0}^{N} K' \pr{K^{\dagger} = K'}}$, which is not restricted to discrete values.
\end{definition}
Now, we may derive the tools necessary to determine when the stacked estimator of the redshift distribution is equal to the mathematically self-consistent expected value of the redshift distribution and thus valid.
\section{Derivation}
\label{sec:math}
Here we compare the stacked estimator $\hat{\mathcal{N}}(z')$ of the redshift distribution to the expected value of the PMF over all possible values of the redshift distribution, derived using the tools of combinatorics.
We consider a generic set ${S_{j} = \{i_{1}, \dots, i_{K}\}}$ of ${|S_{j}| = K}$ galaxies $i \in S_{j}$ and its complement
$\lnot S_{j}$ of ${|\lnot S_{j}| = N - K}$ galaxies $i \notin S_{j}$.
All $N$ galaxies belong to one set or the other, and no galaxy can be in both $S_{j}$ and $\lnot S_{j}$.
One can imagine sets such as these being galaxies with true redshifts equal to $z'$ and $\lnot z'$.
Since galaxies have true redshifts in nature, there is a true number $K^{\dagger}$ of galaxies with $z_{i}^{\dagger} = z'$.
If we aim to recover $\mathcal{N}(z')$, we only care about the size of the set, not its membership.
Thus we seek sets $\{S_{j}\}$ with ${|S_{j}| = K^{\dagger}}$ galaxies out of all $[N]^{K}$ possible subsets of $N$ galaxies.
However, to find the probability that $\mathcal{N}(z')$ takes its true value $K^{\dagger}$, we will need to acknowledge that there is exactly one set $S_{j^{\dagger}}$ out of the set $[N]^{K^{\dagger}}$ of all sets $S_{j}$ with ${|S_{j}| = K^{\dagger}}$ galaxies that contains all galaxies of true redshift ${z_{i}^{\dagger} = z'}$ and no galaxies that have redshift ${z_{i}^{\dagger} = \lnot z'}$, for each possible redshift $z'$.
Constructing all these sets for a cosmologically relevant sample of $N$ galaxies is unfortunately impractical due to the following two theorems, presented here as definitions, whose proofs by induction may be found in \cite{pitman_probability_1999}.
\begin{definition}\label{lem:permutations}
The number of possible orderings or \textit{permutations} of $N$ galaxies is ${N! \equiv N \cdot (N - 1) \cdot \dots \cdot 1}$.
\end{definition}
\begin{definition}\label{lem:combinations}
The number of possible unordered subsets or \textit{combinations} of $K$ galaxies that can be selected from a set of $N \geq K$ galaxies is ${\binom{N}{K} \equiv \frac{N!}{K! (N - K)!}}$.
\end{definition}
Because there are so many possible sets of $K^{\dagger}$ galaxies in question and we do not even know the value of $K^{\dagger}$ in the first place, it is not possible to check them by hand, but this thought experiment mandates that we consider them.
We are now prepared to evaluate the generic probability ${\pr{K^{\dagger} = K'}}$ that $\mathcal{N}(z')$ takes any particular value $K'$ at a fixed $z'$, where ${\mathcal{N}(z') \equiv \sum_{i \in S_{j}} \begin{cases}
1, & z'-\epsilon < z_{i}^{\dagger} < z'+\epsilon\\
0, & \mathrm{else}
\end{cases}}$.
We begin with a toy case of simplifying assumptions and strip them away one by one to obtain a general result.
\subsection{A universe of identical galaxies}
\label{sec:tworedshift}
It is instructive to first consider the special case in which there is one \pz\ PMF\ ${\pr{z'} \equiv \pr{z_{i}^{\dagger} = z'}}$ shared among each galaxy $i$ in the sample of $N$ galaxies.
\begin{lemma}\label{lem:identical}
In a universe in which all galaxies have the same \pz\ PMF\ with probability $\pr{z'}$ of having redshift $z'$, the PMF over the redshift distribution is
\begin{equation}
\label{eqn:binomial}
\pr{K^{\dagger} = K'} = \binom{N}{K'} \pr{z'}^{K'} (1 - \pr{z'})^{N - K'} .
\end{equation}
\end{lemma}
\begin{proof}
In terms of the discrete binary outcome space of Definition~\ref{def:binarystatespace}, the problem is equivalent to that of counting $K$ successes after flipping a single coin with $\pr{\mathrm{success}} = \pr{z'}$ a total of $N$ times.
The $N$ coin flips build a set $S_{j}$ of ${|S_{j}| = K}$ galaxies with redshift $z'$ and a set $\lnot S_{j}$ of ${|\lnot S_{j}| = N - K}$ galaxies with redshift $\lnot z$'.
If we want the probability of $K'$ successes, we sum over the set $[N]^{K'}$ of all combinations of galaxies that could define these sets.
Bearing in mind that our galaxies have ${\pr{\lnot z'} = 1 - \pr{z'}}$, due to Definitions \ref{def:binarystatespace} and \ref{def:disjoint}, the probability of $K'$ galaxies having redshift $z'$ is given by the binomial theorem, which is \Eq{eqn:binomial}.
\end{proof}
One may generalize the two-redshift universe to a multi-valued redshift universe by following the classic proof extending the binomial theorem to the ``birthday problem'' seeking the probability that $K'$ people in a room of $N$ people share the same birthday $z'$.
To instead estimate the redshift density ${n(z') \equiv \mathcal{N}(z') / N}$, a normalized version of the redshift distribution $\mathcal{N}(z')$, for all galaxies ${N \to \infty}$, not just those observed in a given sample, one may follow the classic derivation of the Poisson distribution from the binomial distribution under the limit of $N\to\infty$.
\subsection{A universe of unique galaxies}
\label{sec:unique}
Of course galaxies do not share the same \pz\ PMF, so we must next extend the toy case of identical galaxies to a more realistic set of galaxies with unique \pz\ PMF s ${\{\pr{z'_{i}} \equiv \pr{z_{i}^{\dagger} = z'}\}}$.
\begin{lemma}\label{lem:unique}
In a universe where galaxies do not necessarily all have the same \pz\ PMF, the probability that $\mathcal{N}(z')$ takes the value $K'$ is
\begin{align}
\label{eqn:general}
\pr{K^{\dagger} = K'} &= \sum_{j \backepsilon |S_{j}| = K'}^{\binom{N}{K'}} \left[ \prod_{i \in S_{j}}^{K'} \pr{z'_{i}} \prod_{i \notin S_{j}}^{N - K'} 1 - \pr{z'_{i}} \right] .
\end{align}
\end{lemma}
\begin{proof}
By Definition \ref{def:independence}, the probability of obtaining set $S_{j}$ of ${|S_{j}| = K'}$ galaxies is the product of the probabilities that all the galaxies in the set $S_{j}$ have redshift $z'$ and the probabilities that all the galaxies outside that set, in $\lnot S_{j}$, have redshift $\lnot z'$.
By Definition \ref{def:independence} again, the probability that $S_{j}$ corresponds to the true set of galaxies of redshift $z'$ is the product of the $K'$ probabilities ${\{\pr{z'_{i}}\}_{i \in S_{j}}}$ of the galaxies in $S_{j}$ having redshift $z'$, and the probability that all galaxies outside $S_{j}$ have redshift $\lnot z'$ is the product of the $N - K'$ probabilities ${\{\pr{\lnot z'_{i}}\}_{i \notin S_{j}}}$.
By Definition \ref{def:binarystatespace}, $\pr{\lnot z'_{i}} = 1 - \pr{z'_{i}}$, so the probability $\pr{S_{j}}$ of obtaining a set of galaxies $S_{j}$ with ${|S_{j}| = K'}$ members is ${\pr{S_{j}} = {\prod_{i \in S_{j}} \pr{z'_{i}} \prod_{i \notin S_{j}} 1 - \pr{z'_{i}}}}$.
Invoking Definition \ref{def:disjoint}, we know that ${\pr{K^{\dagger} = K'} = \sum_{j \backepsilon |S_{j}| = K'}^{\binom{N}{K'}} \pr{S_{j}}}$.
\end{proof}
Though \Eq{eqn:general} is computationally intractable for any nontrivial number of galaxies (and even moreso in the limit of continuous redshift), we answer the question at hand by comparing it with \Eq{eqn:stack} in the limiting cases of data- and prior-dominated \pz\ PMF s in Sections~\ref{sec:informative} and \ref{sec:uninformative} below.
\section{Results}
\label{sec:results}
We are now able to derive the general form of the expected value $\langle \mathcal{N}(z')\rangle$ of the number of galaxies with reference redshift $z'$, which we then compare to the stacked estimator of the redshift distribution $\hat{\mathcal{N}}(z')$ given by Equation~\ref{eqn:stack}.
\begin{theorem}\label{thm:general}
The expected value of the number of galaxies $\mathcal{N}(z')$ with redshift $z'$ in the general case of a universe of unique galaxies is given by
\begin{align}
\label{eqn:complete}
\langle K^{\dagger} \rangle &= \sum_{K' = 0}^{N} K' \sum_{j \backepsilon |S_{j}| = K'}^{\binom{N}{K'}} \left[ \prod_{i \in S_{j}}^{K'} \pr{z_{i}^{\dagger} = z'} \prod_{i \notin S_{j}}^{N - K'} 1 - \pr{z_{i}^{\dagger} = z'} \right].
\end{align}
\end{theorem}
\begin{proof}
Apply Definition \ref{def:expected} to Lemma \ref{lem:unique} to arrive at \Eq{eqn:complete}.
\end{proof}
Equation~\ref{eqn:complete} gives the key quantity against which we may compare the stacked estimator of the redshift distribution given in Equation~\ref{eqn:stack}.
In \Sect{sec:informative} and \Sect{sec:uninformative}, we expose the true nature of \pz\ PMF s by iteratively correcting our oversimplified notation.
By confronting what ``$\pr{z_{i}}$'' sweeps under the rug, we identify the only conditions under which Equation~\ref{eqn:stack} can yield a result consistent with \Eq{eqn:complete}.
\subsection{Perfectly informative data}
\label{sec:informative}
In Section~\ref{sec:unique}, we assert that \pz\ PMF s are in general different for each galaxy, and we now justify that claim and consider its implications.
Across the galaxy sample, redshifts are always defined over the same dimension, so the \pz\ PMF s are all defined over $z$, not $z_{i}$ as if it were another dimension unique to each galaxy $i$.
Rather, what differs between the \pz\ PMF s of the galaxy sample is the photometric data $\data_{i}$ upon which our knowledge of each galaxy's redshift is conditioned.
We correct our shorthand of ${\pr{z_{i}^{\dagger} = z'}}$ for the probability that a particular galaxy's true redshift takes a reference value with a more complete ${\pr{z = z' \mid \data_{i}}}$ for the probability of a redshift being equal to a reference value conditional on the data observed from a particular galaxy in preparation to explore the consequences of this clarification.
The dependence of \pz\ PMF s on data transforms \Eq{eqn:stack} into
\begin{align}
\label{eqn:stackwithdata}
\hat{\mathcal{N}}(z) &= \sum_{i = 1}^{N} \pr{z \mid \data_{i}} .
\end{align}
Crucially, \Eq{eqn:stackwithdata} is not correct.
While it may look like Definition~\ref{def:normalization}, we can never integrate over the fixed value on the right side of a conditional;
it is only mathematically valid to integrate over the free variable on the left side of a conditional \citep{hogg_data_2012}.
If the photometric data $\data_{i}$ is optimistically informative, meaning ${\pr{z \mid \data_{i}} \approx \delta(z,\ z_{i}^{\dagger})}$, then \Eq{eqn:stackwithdata} approaches $K^{\dagger}$.
Though \pz\ PMF s cannot be delta functions because photometry is inherently noisy,
the PMF of an idealized errorless spectroscopic redshift could be considered a delta function like this.
We now assess whether \Eq{eqn:complete} yields an equivalent result to \Eq{eqn:stackwithdata} in the idealized case of errorless spectroscopic-quality observations.
\begin{theorem}
\label{thm:informative}
The mathematically valid expected value $\langle K \rangle$ of the redshift distribution at a reference redshift $z'$ given by \Eq{eqn:complete} is equivalent to the stacked estimator $\hat{\mathcal{N}}(z')$ at the reference redshift and the true number $K^{\dagger}$ of galaxies at the reference redshift if the data is perfectly informative, with ${\{\pr{z' \mid \data_{i}} = \delta(z',\ z^{\dagger}_{i})\}}$.
\end{theorem}
\begin{proof}
First, we replace all instances of ${\pr{z_{i}^{\dagger} = z'}}$ in \Eq{eqn:complete} with $\pr{z' \mid \data_{i}}$, yielding
\begin{align}
\label{eqn:proofwithdata1}
\langle K \rangle &= \sum_{K' = 0}^{N} K' \sum_{j \backepsilon |S_{j}|=K'}^{\binom{N}{K'}} \left[ \prod_{i \in S_{j}} \pr{z' \mid \data_{i}} \prod_{i \notin S_{j}} 1 - \pr{z' \mid \data_{i}} \right] .
\end{align}
For galaxies $i$ with perfectly informative data, each of their individual \pz\ PMF s ${\pr{z' \mid \data_{i}}}$ becomes a delta function $\delta(z,\ z^{\dagger}_{i})$ centered at their true redshift, making \Eq{eqn:proofwithdata1} into
\begin{align}
\label{eqn:proofwithdata2}
\langle K \rangle &= \sum_{K' = 0}^{N} K' \sum_{j \backepsilon |S_{j}| = K'}^{\binom{N}{K'}} \left[ \prod_{i \in S_{j}} \delta(z', z^{\dagger}_{i}) \prod_{i \notin S_{j}} 1 - \delta(z', z^{\dagger}_{i}) \right] .
\end{align}
Each of the delta function terms in \Eq{eqn:proofwithdata2} will be $1$ when ${z^{\dagger}_{i} = z'}$ and $0$ otherwise.
Only sets of galaxies with ${|S_{j}| = K^{\dagger}}$ contribute to the outer sum, and there is only one set of galaxies $S_{j^{\dagger}}$ that results in a nonzero product within the inner sum.
This single contributing set of galaxies is the one containing all ${K = K^{\dagger}}$ galaxies with ${z^{\dagger}_{i} = z'}$ and no interlopers with ${z^{\dagger}_{i} = \lnot z'}$.
Thus the sole contributing term to the nested summations is
\begin{align}
\label{eqn:proofwithdata3}
\langle K \rangle &= K^{\dagger} \left[ \prod_{i \in S_{j^{\dagger}}} \delta(z', z^{\dagger}_{i}) \prod_{i \notin S_{j^{\dagger}}} 1 - \delta(z', z^{\dagger}_{i}) \right] .
\end{align}
Since the entire bracketed quantity in \Eq{eqn:proofwithdata3} is $1$, we arrive at the desired ${\langle K \rangle = K^{\dagger}}$, the same as what stacking yields.
\end{proof}
To reiterate, the mathematically complete expected value of the redshift distribution is equivalent to the stacked estimator of the redshift distribution in the case of delta function \pz\ PMF s with perfect accuracy and precision, as would result from an idealized spectroscopic survey.
Ongoing and upcoming broad-band photometric surveys are unlikely to achieve the requisite photometric precision and accuracy for the validity of the stacked estimator.
If the \pz\ PMF s are not delta functions but can be well-described by a specific parametric form, it may also be possible to recover the true redshift distribution via deconvolution \citep{padmanabhan_calibrating_2005}.
Though it is outside the scope of this letter, a quantification of the information loss, on $N(z)$ and on recovered cosmological parameter constraints, due to using the stacked estimator rather than a mathematically self-consistent estimator under a generic parameterization of \project{LSST}-like photo-$z$}%\nospace\ requirements can be found in \cite{malz_how_2020}.
\subsection{Perfectly uninformative data}
\label{sec:uninformative}
\Sect{sec:informative} concerns \pz\ PMF s perfectly informed by photometry, but that picture is still overly simplistic.
If \pz\ PMF s were conditioned solely on photometry, there would be no disagreement about how to derive them from photometric catalogs, because every approach would yield the same result.
Thus the differences between \pz\ PMF s derived by each of dozens of methods indicate that \pz\ PMF s must be conditioned not only on data unique to each galaxy but also on prior information $\tilde{\theta}$ corresponding to a model for the relationship between data $\data$ and redshift $z$.
This prior information may come from a template library or training set, as well as the way each algorithm combines those pieces of information with the actual observations to arrive at an estimated \pz\ PMF\ \citep{schmidt_evaluation_2020}.
The prior information is projected into the space of $\mathcal{N}(z')$ as some interim guess $\tilde{K}$ for the number of galaxies with reference redshift $z'$.
If \pz\ PMF s are more completely written as ${\pr{z \mid \data_{i}, \tilde{K}}}$, \Eq{eqn:stackwithdata} becomes
\begin{align}
\label{eqn:stackwithprior}
\hat{\mathcal{N}}(z) &= \sum_{i = 1}^{N} \pr{z \mid \data_{i}, \tilde{K}} .
\end{align}
If the photometric data were totally uninformative, as could occur pessimistically in an infinitely deep, completely noise-dominated photometric survey, each galaxy $i$ would have the same ${\pr{z \mid \data_{i}, \tilde{K}} \approx \pr{z \mid \tilde{K}}}$.
Furthermore, though the prior $\tilde{K}$ may take any possible value for $K$, it is baked into the \pz\ PMF s on the right side of the conditional and cannot in general be modified at will.
Now, we consider what happens to \Eq{eqn:complete} under perfectly uninformative data with a prior $\tilde{K}$.
\begin{theorem}
\label{thm:uninformative}
The mathematically valid expected value $\langle K \rangle$ of the number of galaxies with reference redshift $z'$ is equivalent to the stacked estimator $\hat{\mathcal{N}}(z')$ at the reference redshift; both $\langle K \rangle$ and $\hat{\mathcal{N}}(z')$ are equal to the true number $K^{\dagger}$ of galaxies with redshift $z'$ under a special case of perfectly uninformative photometry and perfect prior information ${\tilde{K} = K^{\dagger}}$.
\end{theorem}
\begin{proof}
If every galaxy has the same \pz\ PMF, then every galaxy also has the same probability ${\pr{z' \mid \tilde{K}} = \frac{\tilde{K}}{N}}$ of having the reference redshift.
The stacked estimator is thus ${\hat{\mathcal{N}}(z) = N \cdot \pr{z \mid \tilde{K}} = \tilde{K}}$.
To start deriving the complete, correct result, we first update \Eq{eqn:complete} to reflect our new understanding of the role of the prior $\tilde{K}$ in estimated \pz\ PMF s.
\begin{align}
\label{eqn:proofwithprior1}
\langle K \rangle &= \sum_{K' = 0}^{N} K' \sum_{j \backepsilon |S_{j}|=K'}^{\binom{N}{K'}} \left[ \prod_{i \in S_{j}} \pr{z' \mid \tilde{K}} \prod_{i \notin S_{j}} 1 - \pr{z' \mid \tilde{K}} \right] .
\end{align}
Without having to worry about the true galaxy redshifts $z^{\dagger}_{i}$, we can rewrite this in the form of \Eq{eqn:binomial} as
\begin{align}
\label{eqn:proofwithprior2}
\langle K \rangle &= \sum_{K' = 0}^{N} K' \sum_{j \backepsilon |S_{j}| = K'}^{\binom{N}{K'}} \left[ \left(\frac{\tilde{K}}{N}\right)^{K'} \left(1 - \frac{\tilde{K}}{N}\right)^{N - K'} \right] .
\end{align}
Since the \pz\ PMF s are identical to one another, the bracketed terms are the same for a given $K'$, meaning each term in the inner sum is the same, so we can eliminate the inner sum as
\begin{align}
\label{eqn:proofwithprior3}
\langle K \rangle &= \sum_{K' = 0}^{N} K' \binom{N}{K'} \left(\frac{\tilde{K}}{N}\right)^{K'} \left(1 - \frac{\tilde{K}}{N}\right)^{N - K'} .
\end{align}
Noting that the first term with $K = 0$ always vanishes, canceling factors of $K$, and factoring out a quantity motivated by the coin-flipping analogy of \Sect{sec:tworedshift}, we rewrite \Eq{eqn:proofwithprior3} as
\begin{align}
\label{eqn:proofwithprior4}
\langle K \rangle &= N \frac{\tilde{K}}{N} \sum_{K' = 1}^{N} \binom{N - 1}{K' - 1} \left(\frac{\tilde{K}}{N}\right)^{K' - 1} \left(1 - \frac{\tilde{K}}{N}\right)^{(N - 1) - (K' - 1)} .
\end{align}
We recognize this as a binomial expansion of
\begin{align}
\label{eqn:proofwithprior5}
\langle K \rangle &= \tilde{K} \left(\frac{\tilde{K}}{N} + 1 - \frac{\tilde{K}}{N}\right)^{N - 1} .
\end{align}
Since the term in parentheses is $1$, we arrive at ${\langle K \rangle = \tilde{K}}$ for this case.
As $\langle K \rangle = \tilde{K} = \hat{\mathcal{N}}(z')$, the only way for the result of stacking and the expected value of the redshift distribution to be equal to the true value $K^{\dagger}$ of the redshift distribution is if $\tilde{K} = K^{\dagger}$, i.e. if the prior is equal to the truth.
\end{proof}
Thus stacking is valid if the data is uninformative and if the true $N(z)$ is known \textit{a priori}.
Previous low-redshift galaxy surveys with a sufficient sample size to overcome cosmic variance could serve as an appropriate prior for one another's redshift distributions because they would have complete support over the redshift range of their galaxy samples.
However, at higher redshifts or in atypical galaxy environments, a lucky guess of ${\tilde{K} \approx K^{\dagger}}$ is tremendously unlikely;
in fact, if the truth were known independently of the redshift survey itself, it would not be necessary to collect the photometric data at all.
Though stacking works in the case of uninformative data with a prior equal to the truth, this situation is not realistic for upcoming nor ongoing galaxy surveys.
\section{Conclusion}
\label{sec:disco}
This letter explores the common assumptions and misconceptions that have inadvertently disguised the stacked estimator of the redshift distribution as valid throughout the cosmological literature.
We establish a complete and consistent way to frame the redshift distribution $\mathcal{N}(z)$\ in terms of photo-$z$}%\nospace\ PDFs and provide proofs comparing the stacked estimator to the provably correct but computationally prohibitive expected value of the distribution of possible redshift distributions.
We find that equivalence is only achievable in two idealized limiting cases, thereby addressing the dual questions of how stacking earned its place among established methodologies in observational cosmology and why it can no longer be trusted in future applications.
The only cases in which stacking is able to recover the true $\mathcal{N}(z)$\ are when the photo-$z$}%\nospace\ PDFs are conditioned on perfect data or conditioned on perfect prior information.
The former case may have been approximately true for high-fidelity spectroscopic or narrow-band photometric redshift surveys of bright sources, however it is unlikely to hold for broad-band photometric surveys probing faint galaxies, such as those used in cosmological weak lensing.
The latter case may have been approximately true for galaxy surveys whose prior-inspiring predecessors covered the same redshift range with comparable depth and sufficient area to overcome cosmic variance, but upcoming surveys will explore higher redshift ranges than have been previously studied, violating the perfect prior condition.
Finally, while it is possible to imagine a circumstance in which realistically informative and uninformative photo-$z$}%\nospace\ PDFs and priors conspire to produce a stacked $\mathcal{N}(z)$\ that faithfully recovers the truth, such a solution would necessarily have to be very finely tuned and thus is not generically achievable.
While we do not address the numerical degree to which the requirements of stacking were met by previous and ongoing studies, we argue that they will certainly not be met in future cosmological studies.
Therefore, the message is clear;
for the accurate estimation of the redshift distribution in next-generation broad-band photometric surveys, we must not stack!
\section*{Acknowledgments}
I thank Michael Blanton, Johann Cohen-Tanugi, Daniel Gruen, Alan Heavens, David Hogg, Jeff Kuan, Rachel Mandelbaum, Eduardo Rozo, Ted Singer, Anze Slosar, James Stevenson, and Angus Wright for helpful feedback on this letter as it developed.
I further thank all of the above, as well as David Alonso, Will Hartley, Mike Jarvis, Boris Leistedt, Tom Loredo, Mark Manera, Phil Marshall, Jeff Newman, and Josh Speagle, for the conversations that inspired this work.
Finally, I express sincere gratitude to the anonymous referee for suggestions that significantly improved this manuscript.
AIM acknowledges support from the Max Planck Society and the Alexander von Humboldt Foundation in the framework of the Max Planck-Humboldt Research Award endowed by the Federal Ministry of Education and Research.
During the completion of this work, AIM was advised by David W. Hogg and supported by National Science Foundation grant AST-1517237.
This work was initiated while AIM was supported by the U.S. Department of Energy, Office of Science, Office of Workforce Development for Teachers and Scientists, Office of Science Graduate Student Research (SCGSR) program, administered by the Oak Ridge Institute for Science and Education for the DOE under contract number DE‐SC0014664.
|
\section{Introduction}
Observations have shown that the Circumgalactic Medium (CGM) of galaxies is a complex, multi-phase structure where cold ($T\sim10^4$ K) gas clouds co-exist with an ionised, hot ($T\sim10^6$ K) diffuse halo \citep{Putman_2012, Tumlinson_2017}. The CGM is thought to be maintained by flows of gas away from and towards galaxies, which cause a strong interplay between the hot and cold gas phases. The properties of these hot and cold phases and how they interact with each other are poorly understood. Nevertheless, these interactions are expected to be very important for the process of gas accretion and the feeding of star formation in galaxies.
The idea of a hot halo around the Milky Way arose as a hypothesis by \citet{Spitzer_1956} to provide pressure support to observed gas clouds high above the Galactic disc \citep{Muller_1963}. The existence of such hot halos around galaxies was later theoretically predicted as gas accreted from the intergalactic medium (IGM), that is shock-heated to the virial temperature \citep[][]{White_1978, White_1991, Birnboim_2003}. Hot halos are furthermore consistently formed through accretion in Lambda Cold Dark Matter ($\Lambda$CDM) cosmological simulations \citep[e.g.][]{Fumagalli_2014, Nelson_2016, Fielding_2017}. They are expected to extend to the virial radius ($\sim250$ kpc for the Milky Way), and contain a significant fraction of the baryonic mass of galaxies \citep{Fukugita_2006}, which might be the source of gas for the prolonged star formation observed in typical galactic discs \citep[e.g.][]{Bauermeister_2010}.
The hot phase of the CGM is difficult to detect directly, due to its low X-ray surface brightness \citep[e.g.][]{Bregman_2007}. An X-ray excess around some massive early-type galaxies is known to exist for a long time \citep[e.g.][]{Forman_1979, Forman_1985}. Recently, using the new generation of X-ray telescopes they have also been detected around late-type galaxies \citep[][]{Anderson_2011, Humphrey_2011, Yamasaki_2009, Bogdan_2012, Hodges-Kluck_2014, Anderson_2016}. Evidence in support of a hot galactic halo also comes from absorption line studies towards distant active galactic nuclei (AGN) \citep[e.g.][]{Miller_2013, Stocke_2013}. Additional evidence is obtained indirectly from dwarf satellite galaxies that are devoid of gas, which indicates ram-pressure stripping by a diffuse medium \citep[e.g.][]{Grcevich_2009, Gatto_2013}.
The high-density cold gas phase in the CGM of the Milky Way is detected using H {\small{I}} 21-cm emission, and has revealed a population of extra-planar HVCs \citep[e.g.][]{Muller_1963, Woerden_1991, Wakker_1997, Wakker_2001}. These clouds are characterised by velocities of the order of 100 km/s with respect to the local standard of rest that are inconsistent with the rotation of the Galactic disc. The largest population of extraplanar clouds have deviating velocities with respect to the disc below 100 km/s. These are called Intermediate Velocity Clouds (IVCs). Both HVCs and IVCs in the Milky Way form the extraplanar HI layer \citep{Marasco_2011} which is also observed in external galaxies \citep{Sancisi_2008, Marasco_2019}. HVCs are typically metal-poor, while IVCs have metallicities close to solar \citep{Wakker_2001}. This difference hints at different origins: accretion for the HVCs and feedback for the IVCs.
It is thought that
supernovae create hot, ionised super-bubbles that eject gas clouds into the halo regions at velocities on the order of 100 km/s \citep{Maclow_1989}. The ejected gas consists of metal-rich disc material and is likely warm \citep[$\sim10^5$ K,][]{Houck_1990}. As it moves out of the disc it is slowed down by drag forces while cooling and ultimately falls back onto the disc, in a continuous circulation process called the galactic fountain \citep{Shapiro_1976, Fraternali_2006, Grand_2019}. The metal-rich gas clouds from the disc interact with the metal-poor halo material, and gas is stripped from the cloud primarily by Kelvin-Helmholtz instability \citep[KH,][]{Helmholtz_1868, Kelvin_1871}. If the stripped gas mixes efficiently with the hot halo gas it can easily move down the temperature of the latter to where the cooling function peaks ($T\approx$ few $10^5$ K). This causes the cooling time of the mixed gas to become one or more orders of magnitude lower than that of the original hot gas making condensation possible \citep{Marinnaci_2010, Armillotta_2016, Gronke_2020}. The net effect is thus that ejected gas clouds trigger the condensation of hot material from the halo, generally referred to as fountain accretion \citep{Fraternali_2017}. This process can provide the required cold gas to justify the observed star formation rates in galaxies \citep{Fraternali_2008, Marasco_2012, Hobbs_2020}.
Current knowledge suggests that both the ISM and the CGM of spiral galaxies are magnetised. While much of the precise structure and magnitudes remain unknown, Faraday rotation measures show that the spiral arms of the Milky Way have amplified magnetic fields \citep{Beck_2003, Brown_2007}. The Galactic magnetic field can indeed be of the order of tens of $\mu$G in the spiral arms and the bulge region, while having an average value of $\approx3\mu$G \citep{Beck_2009, Beck_2016}.
Magnetic fields are also observed outside the discs of spiral galaxies, in the halo region \citep[e.g.][]{Ekers_1977, Sancisi_1979, Irwin_2012}. It remains however very hard to infer the magnetic field orientation and strength in the halo of the Milky Way. Based on combined Wilkinson Microwave Anisotropy Probe 5 \citep[WMAP5, ][]{Komatsu_2009} polarisation data and Faraday rotation measures, it was shown by \citet{Jansson_2009} that most Galactic field models predict magnetic fields that are inconsistent with observations. The halo region likely requires its own component in a galactic field model, rather than being a simple extension of the disc field. Observations of external galaxies show that an additional out-of-plane component is also required \citep{Beck_2009, Krause_2009}. Turbulent gas flows in galactic halos also give rise to an amplified small scale random magnetic field component \citep[e.g.][]{Beck_2016}.
Magnetic fields can suppress the formation of KH instabilities \citep{Chandrasekhar_1961}. This also occurs in cloud-wind simulations \citep{Maclow_1994, Jones_1996} and generally extends the lifetimes of the clouds. Additionally, under typical CGM conditions the magnetic field drapes around the cloud leading to locally amplified magnetic fields that can become dynamically important and shield the cloud from destruction \citep[e.g.][]{Dursi_2008, Banda-Barragan_2016, Gronnow_2017, Gronnow_2018}. The magnetic draping effect is also seen in simulations that include random magnetic fields \citep[e.g.][]{Asai_2007, Sparre_2020}. The field lines in these simulations are draped along the cloud and are mostly ordered. Turbulent magnetic fields thus become mostly ordered locally around IVCs and HVCs. This is supported by recent highly detailed Faraday rotation measures near the Smith HVC, that also show signs of magnetic draping \citep{Betti_2019}.
The primary effects on the evolution of fast moving cold clouds in hot halos are expected to be from radiative cooling, thermal conduction, magnetic fields, and in some cases self-gravity \citep{Li_2020}. However, the inclusion of radiative cooling, magnetic fields and thermal conduction in simulations is complicated and has only recently been investigated \citep[][but see also \citealp{Orlando_2008} for the related cloud-shock case]{Li_2020, Liang_2020}. Thermal conduction is a diffusive heat exchange process that takes place in the presence of strong temperature gradients \citep{Spitzer_1962}. It is therefore expected to play a fundamental role in the evolution of the CGM due to the large temperature difference between the cold cloud and the hot halo, and has been included by several authors \citep[e.g.][]{Vieser_2007, Bruggen_2016, Armillotta_2016, Armillotta_2017, Li_2020, Liang_2020}. The thermal conduction heat flux at high temperatures is dominated by the contributions of free electrons, that preferentially move along magnetic field lines. The heat flux in the presence of magnetic fields is thus highly anisotropic, being strongly suppressed perpendicularly to field lines (see Section~\ref{sec:numerical_simulations} for a detailed description). However, running full MHD simulations with thermal conduction can be very costly and several authors have resorted to include thermal conduction in hydrodynamical simulations in the form of an isotropic conduction, while mimicking the effect of magnetic fields using a suppression factor. This later point is crucial because even in a weak and dynamically unimportant magnetic field, thermal conduction perpendicular to field lines is still dramatically suppressed.
\citet{Bruggen_2016} shows indeed that isotropic unsuppressed thermal conduction evaporates clouds on timescales that are too short as compared to observational constraints. Thermal conduction at Spitzer value is thus clearly unphysical. For this reason, thermal conduction is generally assumed to operate at a certain efficiency $f<1$. \citet{Narayan_2001} show that the efficiency of thermal conduction in a tangled magnetic field can be reduced to $20\%$ ($f=0.20$) of the classical Spitzer value. Other authors find that thermal conduction is much less efficient at $0.1-1\%$ ($f\approx0.001-0.01$) of the Spitzer value \citep{Chandran_1998}. As we previously mentioned, the fast moving clouds will experience a draped magnetic field instead of a tangled magnetic field. It is therefore unclear how efficient thermal conduction is in the CGM, and thus also its importance. Typical values that are used in literature are $10\%$ \citep[$f=0.1$, ][]{Armillotta_2016, Armillotta_2017}.
In this paper we investigate the approximation of the efficiency $f$-factor by running a large suite of fully 3D HD and MHD simulations of cloud-wind systems representative of the conditions in the lower (close to the galaxy) CGM of Milky Way-like galaxies. Our main goal is to compare the evolution between HD and MHD simulations, focusing specifically on the evolution of the cold gas mass. Since we include radiative cooling in all of our simulations, the net amount of cold gas will generally increase \citep[condensation, e.g.][]{Marinnaci_2010}, as opposed to decrease (evaporation). However, given that both thermal conduction \citep{Armillotta_2016}, and magnetic fields \citep{Gronnow_2018} have been shown to suppress condensation, the quantification of this reduction is a primary goal in this paper.
This paper is organised as follows. In Section~\ref{sec:Methods} we describe the code used for our simulations and the numerical setup. In Section~\ref{sec:Results} we show our results, which are discussed in Section~\ref{sec:discussion}. We conclude in Section~\ref{sec:conclusions}.
\section{Methods} \label{sec:Methods}
\subsection{Simulation Code}
We use \textsc{pluto} version 4.3 \citep{Mignone_2007_pluto, Mignone_2011_amr} to solve the system of equations for hydrodynamical (HD) and magnetohydrodynamical (MHD) flow in 3 spatial dimensions. \textsc{pluto} is a Eulerian Godunov-type \citep{Godunov_1959} code. For both HD and MHD flows we close the system of equations with an ideal equation of state. For the internal energy we assume that both the cloud and hot halo are monatomic and accordingly set the adiabatic index $\gamma = 5/3$. In MHD runs we approximately enforce the solenoidal constraint ($\nabla \cdot \bmath{B} = 0$) using the hyperbolic divergence cleaning scheme by \citet{divclean_Dedner_2002}, implemented in \textsc{pluto} by \citet{divclean_mignone_2010}. For the flux computations there is a trade-off to be made between accuracy and numerical stability, as more accurate Riemann solvers are generally also less stable. As a compromise, we employ the approximate Riemann solver HLLC \citep{HLLC_Toro_1994} for HD simulations and HLLD \citep{HLLD_Miyoshi_2005} for MHD simulations. Note that the HLLD Riemann solver is merely an MHD extension of the HLLC Riemann solver, and does not introduce artificial differences between HD and MHD simulations. We use second-order Runge-Kutta time integration and linear (2nd order) reconstruction for all simulations.
In order to run our simulations at sufficient spatial resolutions while keeping computational costs feasible, we employ adaptive mesh refinement (AMR). We use the gradient of the density as refinement variable, and to ensure numerical stability we set the Courant-Friedrichs-Lewy \citep[CFL,][]{Courant_1928} number to $C_\alpha$ = 0.3 for all simulations. We start our simulations from a base grid of $N_x \times N_y \times N_z = 60\times20\times20$ on a physical grid of $L_x\times L_y\times L_z = 6\times2\times2$ kpc. For each level of refinement we increase the resolution in every dimension by a factor 2. In the fiducial setup we refine 5 times to an equivalent resolution of $1920\times640\times640$. In this way, there are 32 cells per cloud radius ($\mathcal{R}_{32}$), which gives a spatial resolution of $\sim$3 pc/cell at the highest refinement level.
\subsection{Numerical simulations}
\label{sec:numerical_simulations}
We initialise our simulations according to the typical "cloud-wind" problem, and follow the general setup as described in \citet{Gronnow_2018}. In particular, we assume the same steep, smooth density profile given by
\begin{equation}
\rho(r) = \rho_{\rm hot} + \frac{1}{2}(\rho_{\rm cloud} - \rho_{\rm hot})\times\Bigl(1 - \tanh\Bigl[s\Bigl(\frac{r}{r_0}-1\Bigr)\Bigr]\Bigr),
\end{equation}
where $\rho_{\rm hot}$ is the density of the hot halo, $\rho_{\rm cloud}$ is the density of the cloud, $r_0$ is the cloud radius, and $s=10$ is the steepness parameter. Thus, the cloud radius $r_0$ for this density profile is defined to be at the halfway point between the central cloud density and the halo density. Turbulent velocities with Mach-number $\mathcal{M}\approx1$ are added to the cloud to make the simulations more realistic \citep{Marasco_2019}, and to reduce artifacts arising from symmetry. Similar to \cite{Armillotta_2016}, we pick random values from a Gaussian distribution of velocities with a dispersion of 10 km/s, and a mean of 0 km/s. These velocities are then assigned to the region consisting of only cloud material i.e. $r<r_{\rm 0}$, where $r_0$ is the cloud radius.
\begin{table*}
\centering
\parbox{10.4cm}{
\caption{The cloud and hot halo parameters for our fiducial setup. }
}
\label{tab:default_parameters}
\begin{tabular}{p{1cm}p{1cm}p{1cm}p{1cm}p{1cm}p{1cm}p{1cm}p{0.3cm}}
\hline
$v_\text{rel}$\ \ \ \ (km s$^{-1}$) & $T_\text{cloud}$ (K) & $T_\text{hot}$\ \ \ \ \ (K) & $Z_{\text{cloud}}$ ($Z_{\odot}$) & $Z_{\text{hot}}$ ($Z_{\odot}$) & $n_{\rm cloud}$ (cm$^{-3}$) & $n_\text{hot}$ (cm$^{-3}$) & $r_{0}$\ \ \ \ \ \ \ \ \ (pc) \\
\hline
75 & 10$^4$ & 2$\times$10$^6$ & 1 & 0.1 & 0.2 & $10^{-3}$ & 100 \\
\hline
\end{tabular}
\parbox{10.4cm}{
\textit{Notes.} The parameters are the relative velocity between cloud and halo ($v_\text{rel}$), cloud temperature ($T_\text{cloud}$), halo temperature ($T_\text{hot}$), cloud/halo metallicity ($Z_\text{cloud/hot}$), halo number density ($n_\text{hot}$), and cloud radius ($r_0$). Note that the pressure equilibrium between cloud and halo makes the cloud number density a fixed value.
}
\end{table*}
We include radiative cooling in every simulation using the tabulated cooling module, based on the collisional ionization equilibrium cooling tables of \citet{Sutherland_1993}. The variation in mean molecular weight $\mu$ with temperature $T$ is accounted for, and we interpolate for the variation in metallicity ($Z$) with $T$ between 3 tables with metallicities 0.1$Z_\odot$, 0.3$Z_\odot$, $Z_\odot$, similarly to \citet{Marinacci_2011} and \citet{Gronnow_2018phd}. We keep track of the metallicity using a passive scalar $C$ that does not affect the flow, which we advect conservatively as
\begin{equation}\label{eqn:passive_scalar}
\frac{\partial (\rho C)}{\partial t} + \nabla\cdot(\rho C\bmath{v}) = 0.
\end{equation}
We initialise the passive scalar $C$ in simulations at a cloud metallicity ($Z_\text{cloud}$) for $r<r_\text{0}$, and at a hot halo metallicity $Z_\text{hot}$ elsewhere. We finally assume a cooling floor at $T=10^4$ K below which no cooling occurs. This assumption is justified as the cooling rate drops drastically below $10^4$ K, and most of the CGM "cold" material is observed at these temperatures \citep{Putman_2012, Werk_2013}.
Thermal conduction is included in our simulations using the readily available thermal conduction module in \textsc{pluto}, but with the following modifications. We assume that thermal conduction happens exclusively through free electrons, and we account for this by multiplying the thermal conduction heat flux by the ionisation fraction $x_{\rm i}$. The ionisation fraction depends on the temperature, which we calculate using the same procedure as described for radiative cooling.
The full equation for the thermal conduction heat flux becomes
\begin{equation} \label{eqn:conduction}
\bmath{q} = -f\frac{\kappa_{\rm Sp}\nabla T}{1 + \sigma}x_{\rm i},
\end{equation}
where $\bmath{q}$ is the heat flux, $x_{\rm i}$ is the ionisation fraction, $\nabla T$ is the gradient of the temperature, and $\kappa_\text{Sp}$ is the classical Spitzer \citep{Spitzer_1962} conductivity given by
\begin{equation}
\kappa_\text{Sp}= 1.84\times10^{-5}\frac{T_{\rm e}^{5/2}}{\text{ln}\ \Psi}\ \ \ \ \text{ergs/s/K/cm},
\end{equation}
where $\nabla T$ is the gradient of the temperature, $T_{\rm e}$ is the electron temperature which we assume to be equal to the fluid temperature, and ln $\Psi$ is the Coulomb logarithm given by
\begin{equation}
\text{ln}\ \Psi = 29.7 + \text{ln} \frac{1}{\sqrt{n_{\rm e}/\text{cm}^{-3}}}\frac{T_{\rm e}}{10^6 \text{K}},
\end{equation}
where $n_{\rm e}$ is the electron number density.
When the mean free path of electrons becomes similar to or greater than the temperature scale length, the classical Spitzer flux is no longer accurate and the heat flux is said to be saturated. We account for saturated heat flux according to the description as given by \citet{Cowie_1977}, and we set
\begin{equation}
\bmath{q}_\text{sat} = 5\phi_\text{sat}\rho c_{\rm s}^3,
\end{equation}
where $\bmath{q}_\text{sat}$ is the saturated heat flux, $\rho$ is the gas density, $c_{\rm s}$ is the local sound speed, and $\phi_\text{sat}$ is a factor of order unity that accounts for uncertainties regarding the flux-limited treatment. Following the same arguments as presented in \citet{Armillotta_2016} we set $\phi_\text{sat} = \sqrt{f}$. To ensure a smooth transition between the classical and saturated regimes the factor $1 + \sigma$ is included in Eq.~\ref{eqn:conduction}, defined as
\begin{equation}
\sigma = \frac{\kappa_\text{Sp}{\lVert}\nabla T{\rVert}}{\bmath{q}_\text{sat}},
\end{equation}
where ${\lVert}\nabla T{\rVert}$ is the magnitude of the temperature gradient. Finally, the efficiency factor $f$ is a dimensionless parameter between 0 and 1 that is typically used to approximate the suppression due to magnetic fields as discussed above, and is the focus of this work.
In a subset of our simulations we include magnetic fields. We vary between weak (0.1 $\mu$G) and strong (1.0 $\mu$G) magnetic fields, as appropriate for the lower part of the CGM \citep{Gronnow_2018}. For each field strength we run the simulations for orientations of the magnetic field parallel ($B_\parallel$) or perpendicular ($B_\bot$) to the relative velocity. In simulations where we include both thermal conduction and a magnetic field, we calculate the anisotropic thermal conduction heat flux, which is natively included in \textsc{pluto}. In this case the heat flux is split up into parallel and perpendicular components with respect to the direction of the magnetic field as follows \citep{Spitzer_1962, Braginski_1965}.
\begin{equation}\label{eq:anisotopric_conduction}
\bmath{q}_{\rm class} = \kappa_{\parallel}\bmath{\hat{b}}(\bmath{\hat{b}}\cdot{\nabla} T) + \kappa_{\bot}[{\nabla} T - \bmath{\hat{b}}(\bmath{\hat{b}}\cdot{\nabla} T)],
\end{equation}
where $\bmath{q}_{\rm class}$ is the classical conduction heat flux, $\bmath{\hat{b}} = \bmath{B}/|\bmath{B}|$ is a unit vector denoting the direction of the magnetic field, and $\kappa_\bot$ and $\kappa_\parallel$ are the perpendicular and parallel components of the conductivity, where $\kappa_\parallel=\kappa_{\rm Sp}$. The perpendicular conductivity is given by
\begin{equation}
\kappa_{\bot} = \frac{8\sqrt{\pi m_{\rm H}k_{\rm B}}n_{\rm H}^2e^2c^2\ln \Psi}{3|\bmath{B}^2|\sqrt{T}},
\end{equation}
where $m_{\rm H}$ is the mass of hydrogen, $k_{\rm B}$ is Boltzmann's constant, $n_{\rm H}$ is the hydrogen number density, and $c$ is the speed of light in vacuum. For the parameters used in our analysis, $\frac{\kappa_\bot}{\kappa_\parallel} \lesssim 10^{-9}$. The conductivity perpendicular to field lines is thus much smaller than that parallel to the field lines, and we set $\kappa_\bot = 0$.
To assess the effect of isotropic thermal conduction, we perform HD simulations with either only radiative cooling, or radiative cooling and isotropic thermal conduction at several efficiencies by varying $f$, as listed in Table~\ref{tab:variation_parameters}. To examine the effect of anisotropic thermal conduction in the presence of magnetic fields, we run MHD simulations at the field strengths and orientations as mentioned above. In this case, we set anisotropic thermal conduction either off, or at full Spitzer efficiency ($f = 1$). Our main goal is to compare the evolution between isotropic thermal conduction runs (no magnetic field) and anisotropic thermal conduction runs (with magnetic fields).
Additionally, we explore several different cloud and halo parameters. We investigate a lower cloud metallicity, a higher relative velocity, and lower cloud and halo densities. For every different set of parameters, we perform the same suite of HD and MHD simulations as described above (Table~\ref{tab:variation_parameters}).
\begin{table}
\caption{The varied simulation parameters.}
\label{tab:variation_parameters}
\begin{tabular}{ |p{1.5cm}|p{1.5cm}|p{4cm}| }
\multicolumn{3}{|c|}{} \\
\hline
Parameter & Setup & Values considered \\
\hline
$f$ & HD & 0.01, 0.05, 0.1, 0.15, 0.2 \\
$v_\text{rel}$ $^a$ & HD,MHD & $\mathbf{75}$, 150 [km/s] \\
$Z_\text{cloud}$ & HD,MHD & 0.1, $\mathbf{1.0}$ [$Z_\odot$] \\
$n_\text{cloud}$,$n_\text{hot}$ $^b$ & HD,MHD & ($\mathbf{0.2}, \mathbf{10^{-3}}$), (0.1, 5$\times10^{-4}$) [cm$^{-3}$] \\
$|\bmath{B_0}|$ $^c$ & MHD & 0.1, 1.0 [$\mu$G] \\
$B_\text{orientation}$ & MHD & $B_\bot$, $B_\parallel$, $B_{\rm ob}$ \\
\hline
\end{tabular}
\footnotesize{\textit{Notes.} The 'Setup' column describes for which module(s) we perform the variation. The values used in the fiducial simulation setup are highlighted in bold.
$^a$This corresponds to Mach numbers of the hot gas $\mathcal{M}\approx0.35$ and 0.70 for relative velocities 75 and 150 km/s, respectively. $^b$We assume pressure equilibrium between the cloud and the hot halo, thus the density contrast between cloud and halo is in both cases $\chi = \frac{n_\text{cloud}}{n_\text{hot}}=200$. $^c$The plasma-$\beta$ parameter, defined as the ratio between the thermal pressure of the gas over the magnetic pressure, is for the higher densities $\beta$=7 and 700 for field strengths 1 and 0.1 $\mu$G, respectively. For the lower density setup, the plasma-$\beta$ values are a factor 2 smaller.}\\
\end{table}
\subsection{Analysis of the simulations} \label{sec:analysis}
To follow the evolution of simulations with different parameter setups equally, we denote time in units of cloud-crushing time $t_\text{cc}$ \citep{Jones_1996},
\begin{equation}
t_\text{cc} = \sqrt{\chi}\frac{2r_{\rm 0}}{v_\text{rel}},
\end{equation}
where $\chi = \frac{n_\text{cloud}}{n_\text{hot}}$ is the density contrast, $r_{\rm 0}$ is the radius of the cloud, and $v_\text{rel}$ is the relative velocity between the cloud and the halo. For our fiducial setup, $t_\text{cc} \approx 18$ Myr.
We diagnose our data using volume averaged, mass weighted quantities as described in \citet{Klein_1994} and \citet{Shin_2008}. We focus on the evolution of the amount of cold gas in the simulation domain according to
\begin{equation}\label{eqn:coldgas}
\Delta \text{M}_\text{cold} = \frac{\int \rho_\text{cold} dV}{M_\text{cloud, $t$=2 Myrs}},
\end{equation}
where, similar to \citet{Armillotta_2016}, we denote gas as "cold" if $T < 2\times10^4$ K, and we normalise by the cloud mass at $t=2$ Myrs $M_\text{cloud, $t$=2 Myrs}$. We use the cloud mass after 2 Myrs as opposed to the initial cloud mass since the transition region between cloud and halo consists of intermediate temperature ($T\approx10^5$ K) material that quickly cools down to $10^4$ K which leads to an early jump in the cold gas mass that is not due to condensation.
We furthermore calculate the mass of mixed gas similar to \citet{Xu_1995}. We define a passive scalar $C$ over the simulation volume, where we set $C=1$ inside the cloud, $C=0$ elsewhere. This is advected using Eqn.~\ref{eqn:passive_scalar}, and we define gas to be 'mixed' if $0.1 < C < 0.9$.
\section{Results}\label{sec:Results}
\subsection{Fiducial simulation setup}
Our fiducial simulation setup focusses on cool clouds travelling through the lower CGM of a Milky Way-like galaxy. We list the parameters used in our fiducial simulation setup in Table~\ref{tab:default_parameters}. We assume parameters for a typical intermediate velocity cloud, that are consistent with earlier work \citep[e.g.][]{Marinnaci_2010, Armillotta_2016, Gronnow_2018}. We furthermore assume a halo temperature typical of the Milky Way \citep[e.g.][]{Bregman_2007}, and a halo metallicity ($Z_{\rm hot}$) as estimated in Milky Way-like galaxies \citep[e.g.][]{Bogdan_2017, Hodges-Kluck_2018}. The clouds considered in this work exceed the critical "mass growth" radius of \citet{Gronke_2018, Gronke_2020} and are thus in the mass growth regime, this agrees with the fact that we find condensation.
\citet{Li_2020} propose a qualitatively different threshold for which the column density of the cloud can be used to distinguish between cloud destruction and mass growth.
In our fiducial setup the clouds are {\em just} within their “classical cloud destruction” regime and thus the discrepancy with this criterion does not seem significant; in this respect see also \citet{Kanjilal_2020}.
In our simulations with a lower cloud column density (Section~\ref{sec:lowrho}) that more clearly belong to the cloud destruction regime of \citet{Li_2020} we do not see a discrepancy, as they indeed do not show condensation.
\subsubsection{HD simulations}
We investigate the effects of isotropic thermal conduction in 3D HD simulations by varying the efficiency of conduction through the $f$-factor. We show slices of the temperature distribution through part of the simulation domain after $60$ Myrs in Fig.~\ref{fig:solar_CTC}. Similar to previous work, e.g. \citet{Bruggen_2016, Armillotta_2016}, we find that thermal conduction decreases the size of the wake through the evaporation of stripped cloudlets. The amount of cold gas in the simulation domain increases (i.e. there is condensation) in all cases, which is displayed in Fig.~\ref{fig:solar_condensation}, top panel. However, the amount of condensation decreases with stronger (higher value of $f$) thermal conduction. Additionally, we show the mixed gas evolution in Fig.~\ref{fig:solar_condensation}, bottom panel. All runs except $f=0.2$ are fully mixed ($M_\text{mixed}/M_\text{cloud, $t$=2 Myrs} = 1$) after $60$ Myrs. This noticeable decrease in mixing efficiency for the $f=0.2$ run shows that thermal conduction can have an additional effect, which we expand on later in this section.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{imgs/Slices_CTC_temp.pdf}
\caption{Slices through part of the simulation domain of the gas temperature at $t=60$ Myr for the fiducial simulation setup (Table \ref{tab:default_parameters}). We vary the efficiency of isotropic thermal conduction with the $f$-factor from $f=0.0$ (no thermal conduction) to $f=0.20$ (strong thermal conduction) of the Spitzer value. All gas below $2\times10^4$ K has been coloured dark blue.}
\label{fig:solar_CTC}
\end{figure*}
According to the work by \citet{Begelman_1990}, the thermal instability is suppressed on scales smaller than the Field length $\lambda_{\rm F}$, named after the work on thermal instability by \citet{Field_1965}:
\begin{equation}\label{eqn:Field}
\lambda_{\rm F} = \sqrt{\frac{f\kappa_\text{Sp}T_\text{hot}}{n^2_\text{cold}\Lambda(T_\text{cold})}},
\end{equation}
where $n_{\rm{cold}}$ is the number density of cold (in our case $\sim10^4$ K) material, and $\Lambda(T_{\rm{cold}})$ is the cooling rate at this temperature\footnote{Since we implement a cooling floor at $T=10^4$ K, in principle there is no cooling at this temperature. Instead, we consider the cooling rate slightly above $10^4 K$ and take $\Lambda(T_{\rm cold}) = 10^{-24}$ ergs cm$^3$/s.}. For our fiducial simulations the Field length varies in the range $\sim10-60$ pc for $f=0.01-0.20$, respectively. Hence, for weak thermal conduction ($f=0.01$) the Field length is slightly below the size of stripped cloudlets ($\sim10$ pc) and they do not evaporate. However, stronger thermal conduction increases the Field length and thus quickly dominates over cooling making the cloudlets evaporate. We also see the formation of a smooth, intermediate temperature wake. Stripped cloudlets inside this intermediate temperature wake are more stable against evaporation since $T_{\rm hot}$ is smaller, which decreases the Field length. It is for this reason that some cloudlets are seen to survive in the wake even for strong ($f=0.2$) thermal conduction. Note also that in the strong thermal conduction cases ($f=0.15-0.2$), the Field length is $\approx 50$ pc, which evidently has a strong effect on the evolution of the cloud itself. We argue that in this case, thermal conduction can form and maintain a smooth temperature gradient between the head of the cloud and the hot halo. This subsequently creates a less steep density gradient that can suppress the formation of KH-instabilities \citep{Vieser_2007}, which extends the lifetime of the cloud \citep{Armillotta_2017}.
Therefore the primary effect of thermal conduction is the evaporation of stripped cloudlets. However, in cases where thermal conduction is very strong, a secondary effect can become dominant that suppresses KH instabilities and delays mixing and condensation.
\begin{figure}
\includegraphics[width=0.5\textwidth]{imgs/solar_coldgas_mixing.pdf}
\caption{We show the cold gas mass evolution (top) for our fiducial simulation setup (Table~\ref{tab:default_parameters}) for HD runs. We vary only the efficiency of isotropic thermal conduction through the parameter $f$. On the bottom, we show the mixing fraction for the same simulations, where the colour coding is the same as for the top panel.}
\label{fig:solar_condensation}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{imgs/Slice_Bperp_temp_default_weak.pdf}
\caption{Slices through part of the simulation domain of the temperature in the weak, perpendicular magnetic field runs for our fiducial simulation setup (Table~\ref{tab:default_parameters}). To illustrate the anisotropic morphology we show slices through the y- and z-axes. The magnetic field is initially parallel to the y-axis. We show the simulation without thermal conduction on the left, and the simulation with anisotropic conduction on the right. Same as for Fig.~\ref{fig:solar_CTC}, all gas below $2\times10^4$ K has been coloured dark blue.}
\label{fig:mag_field_temperatures}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{imgs/B_MAG_PERP_FIDUCIAL.pdf}
\caption{The same simulations as shown in Fig.~\ref{fig:mag_field_temperatures}, but here we show slices of the magnetic field strength (colours) and orientation (arrows). We overlay the normalised magnetic field vectors on the top panels. We do not include the field vectors in the bottom panels since the field is mostly oriented into the paper. }
\label{fig:mag_field_perp}
\end{figure*}
\subsubsection{MHD simulations}
We perform MHD simulations with and without anisotropic thermal conduction, and we show temperature slices for a weak (0.1 $\mu$G), magnetic field perpendicular to the motion of the cloud ($B_{x} = B_{z} = 0$, and $B_{y}= B_0$) in Fig.~\ref{fig:mag_field_temperatures}. The dynamical evolution of the system is affected by the magnetic field even in the weak field setup. By a magnetic draping effect \citep{Dursi_2008, Gronnow_2018}, field lines are swept up and stretched by the cloud leading to significant field amplification as we show in Fig.~\ref{fig:mag_field_perp}. Magnetic pressure becomes dynamically important and squeezes cloud material along one axis (see e.g. the top panels of Fig.~\ref{fig:mag_field_temperatures} and \ref{fig:mag_field_perp}), which causes the cloud to expand along the axis perpendicular to it (bottom panels of Fig.~\ref{fig:mag_field_temperatures} and \ref{fig:mag_field_perp}). The overall morphology of the cloud gas is not strongly affected by the inclusion of anisotropic thermal conduction in this case. However, we note that thermal conduction creates a wake with temperatures of $T \approx1-1.5\times10^6$ K. Due to the magnetic draping effect the field lines are wrapped around the head of the cloud and cloudlets. Since thermal conduction is only efficient along the field lines, a steep temperature gradient between cloud and hot halo can be maintained for at least $60$ Myrs. Additionally, we notice that small stripped cloudlets can survive even with thermal conduction included, which is consistent with the findings of \citet{Liang_2020} in their 2D MHD simulations.
We show temperature and magnetic field slices with a weak field parallel to the motion of the cloud ($B_{x} = B_0$, and $B_{y}, B_{z} = 0$) in Fig.~\ref{fig:slice_bpar}. While the elongated morphology is more akin to the runs without a magnetic field, we notice that there is significantly less stripping in this case. The parallel magnetic field efficiently suppresses KH instabilities \citep{Sur_2014}, leading to a narrow wake and little stripping. In this case we notice a morphological change with the inclusion of thermal conduction. The magnetic field lines enter the cold cloud directly from the hot halo, leading to highly efficient thermal conduction at the cloud-halo interface. Similar to the HD setups with strong ($f = 0.15-0.2$) thermal conduction, this creates a temperature and density gradient at the cloud-halo interface. In this case, both the magnetic field and the temperature gradient act to suppress stripping by KH instability.
\begin{figure*}
\centering
\includegraphics[width=1.0\textwidth]{imgs/Bpar_Temp+B.pdf}
\caption{Slices through part of the simulation domain for simulations with a weak, parallel magnetic field without thermal conduction (top) and with anisotropic thermal conduction (bottom) for the fiducial simulation setup (Table~\ref{tab:default_parameters}). We show the temperature on the left, and the magnetic field strength (colours), and orientation (arrows) on the right. In the temperature slices we colour all gas below $2\times10^4$ K dark blue.}
\label{fig:slice_bpar}
\end{figure*}
We show the cold gas evolution for all magnetic field setups in Fig.~\ref{fig:mag_coldgas_default}. As found previously by \citet{Gronnow_2018}, we notice that all magnetic field setups have less condensation than the HD 'only cooling' run (as shown in Fig.~\ref{fig:solar_CTC}). In the perpendicular field setups the condensation is further decreased by $\sim10\%$ when thermal conduction is included. Due to the magnetic draping effect thermal conduction does not operate efficiently on the cloud-halo interface, and stripping is mostly unhindered. Hence, the decrease in condensation is due to the evaporation of cloudlets in the wake. In contrast, for the parallel field setups the condensation changes little with the inclusion of thermal conduction. In this case the amount of gas stripped from the cloud is already very limited due to the magnetic field orientation, such that evaporation of cloudlets is negligible. The dominant way for the cloud to condense material is thus directly onto the cloud, as opposed to in the wake. The figure also shows an oblique orientation of the field ($B_{\rm ob}$, where $B_{x}= B_{y} = \frac{B_0}{\sqrt{2}}$, and $B_{z} = 0$). In this case the magnetic field is also seen to drape around the cloud, leading to very similar results as the perpendicular field setup. This shows that if magnetic draping can occur the results will be similar to the perpendicular field setup. Thus, the perpendicular magnetic field setup can be considered representative of most field orientations.
\begin{figure*}
\centering
\includegraphics[width=1.0\textwidth]{imgs/compgas_MHD.pdf}
\caption{The cold gas evolution for our fiducial MHD simulation runs with thermal conduction turned off (left), and for runs with anisotropic thermal conduction (TC) turned on (right). The different curves refer to different strengths \& orientations of the magnetic field. Note that all simulations include radiative cooling. For comparison, we show the cold gas evolution when no magnetic field is included as seen in Fig.~\ref{fig:solar_condensation} in grey. The grey band in the right hand side plot shows the range of cold gas masses for simulations with $f=0.01$ to $f=0.2$.}
\label{fig:mag_coldgas_default}
\end{figure*}
\subsubsection{The efficiency of thermal conduction}
The $f$-factor only approximates the suppression effect in the thermal conduction heat flux, so in order to isolate the effect of adding thermal conduction to the simulations we compare the ratio of cold gas masses as follows. We calculate
\begin{equation}\label{eq:delta_HD}
\Delta M_\text{HD} = \frac{M_\text{cold, HD}}{M_\text{cold, HD + conduction}},
\end{equation}
for HD runs, and
\begin{equation}\label{eq:delta_MHD}
\Delta M_\text{MHD} = \frac{M_\text{cold, MHD}}{M_\text{cold, MHD + conduction}},
\end{equation}
for MHD runs. In this way, the majority of the difference in the cold gas mass between HD and MHD simulations that comes from the magnetic field hindering the KH instability is filtered out. With these ratios we can thus obtain an estimate of suitable values for $f$ by comparing the approximated suppression of thermal conduction in HD simulations to the 'true' suppression in the MHD simulations.
In Fig.~\ref{fig:all_suppressions}, we show the results for Eqn.~\ref{eq:delta_HD} on the left, and the results for Eqn.~\ref{eq:delta_MHD} on the right, for all simulation setups. The first row contains the results for the fiducial simulation setup, seen until now. The results show percentage wise how much cold gas is formed by runs without thermal conduction, compared to runs with thermal conduction. We notice a clear distinction between field orientations, where as mentioned before, perpendicular fields are more strongly suppressed in terms of the cold gas formed than the parallel fields. For perpendicular magnetic fields we find that a thermal conduction efficiency of $f\approx0.10-0.15$ fits well, and for parallel fields the effect of thermal conduction is near negligible at $f\approx0.02$.
\begin{figure*}
\centering
\includegraphics[width=0.73\textwidth]{imgs/all_suppressions.pdf}
\caption{We show the ratio of cold gas masses with and without isotropic thermal conduction for HD simulations at a certain $f$ (left), and for MHD simulations with $f=1$, but with anisotropic thermal conduction (right). Every row represents a different simulation setup. We show from top to bottom the fiducial, 10\% solar cloud, lower density, and the high velocity setups, respectively.}
\label{fig:all_suppressions}
\end{figure*}
\subsection{Varying cloud metallicity}\label{sec:subsolar}
Typically IVC's are observed to have metallicities close to solar \citep[see e.g.][]{Richter_2001}, indicating that their origin was likely close to the metal-rich galactic disc, and were possibly ejected by the galactic fountain process \citep{Shapiro_1976, Bregman_1980, Spitoni_2008,Fraternali_2006, Fraternali_2017}. However, some IVC's \citep[see e.g.][]{Hernandez_2013, Fukui_2018} show signs of sub-solar metallicity. Here we examine the effect of lowering the cloud metallicity to 10$\%$ solar (low $Z$). This can be more typical of HVC's \citep[e.g.][]{Wakker_2007}, which we further consider in Section~\ref{sec:wind_velocity}.
It is well known that gas with higher metallicity have a shorter cooling time than lower metallicity gas, due to the rapid cooling by metal lines as compared to hydrogen and helium. However, this does not necessarily lead to a significant increase in condensation in "cloud-wind" systems. An increased cooling rate has been found, for instance, to suppress the stripping of cloudlets \citep{Cooper_2009}. We also find that
the cloudlets that are stripped and that mix with the halo gas do cool more efficiently at high $Z$ with respect to low $Z$, which in this case leads to very similar amounts of condensation, as can be seen in the left panel of Fig.~\ref{fig:coldgas_metal}. When isotropic thermal conduction is included the runs with lower cloud metallicity produce significantly less cold gas than their higher metallicity counterparts. From the definition of the Field length (Eqn.~\ref{eqn:Field}), we know that a decrease in cooling rate increases the Field length. At the peak of the cooling curve ($T\sim10^5$ K), the difference between solar and $10\%$ solar can be up to a factor 10, which increases the Field length by a factor $\sqrt{10}$. This difference in cooling rate converges around $T=10^4$ K, such that the overall change to the Field length is small, but not negligible. Thus, cloudlets are more prone to evaporate for the lower metallicity setup leading to less condensation. In addition, the mixing fraction in the right panel of Fig.~\ref{fig:coldgas_metal}, suggests that the mixing efficiency is also suppressed more strongly for a lower cloud metallicity.
We show the cold gas evolution for the MHD simulations with anisotropic thermal conduction in Fig.~\ref{fig:coldgas_tc} on the left panel, and the suppression effect in Fig.~\ref{fig:all_suppressions}, second row. The amount of cold gas formed is less than in the simulations with a solar metallicity cloud in all cases. The suppression effect in the MHD simulations is very similar to that in fiducial simulation setups. However, the suppression effect for the HD runs does change significantly, which suggests an overall value of $f\approx 0.05-0.10$, lower than what is found in the fiducial (high $Z$) case.
\begin{figure}
\includegraphics[width=0.5\textwidth]{imgs/metal_comp_mix_cold.pdf}
\caption{On the top panel we show the cold gas evolution as compared between a solar metallicity cloud (fiducial) and a 10\% solar metallicity cloud. On the bottom panel we show the mixing fraction, where the colours and markers correspond to the same simulations as in the left panel.}
\label{fig:coldgas_metal}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{imgs/Coldgas_TC_par_variation.pdf}
\caption{Cold gas evolution for the non-fiducial simulation setups including radiative cooling, magnetic fields, and anisotropic thermal conduction (TC). From left to right we show the lower cloud metallicity setup, the lower cloud/halo density setup, and the higher velocity setup, respectively. In each panel we include the fiducial simulation results as dashed lines, where the colours correspond to the same magnetic field strength/orientation.}
\label{fig:coldgas_tc}
\end{figure}
\subsection{Varying cloud and halo density}\label{sec:lowrho}
We investigate a factor of 2 lower density for cloud and halo with $n_{\rm cloud}=0.1\ {\rm cm}^{-3}$ and $n_{\rm hot}=5\times10^{-4}\ {\rm cm}^{-3}$. We keep the density contrast $\chi = \frac{n_\text{cloud}}{n_\text{hot}}$ at 200, such that the pressure equilibrium between cloud and hot halo is maintained without changing their temperatures. We make the same comparison between cold gas mass and mixing rate as for the lower metallicity setup in Fig.~\ref{fig:coldgas_lowrho}. Lowering the density has little effect on the simulations with only radiative cooling. This can be understood as follows. Gas at a temperature of $10^5$ K is at the peak of the cooling curve, for which the cooling time is of the order of a few thousand years and thus immediately cools the gas to $\sim 10^4$ K. The metallicity or gas density thus does not significantly alter the evolution of the gas below $10^5$ K. Since the cooling time of gas at the halo temperature of $2\times10^6$ K is of the order of a few billion years \citep{Fraternali_2017}, the only way to obtain gas at the intermediate temperatures of $10^5$ K is to mix the cloud and halo gas. Therefore the amount of condensation is determined by the mixing efficiency. However, with the inclusion of thermal conduction a drastic decrease in condensation is seen. For this reason we have not performed simulations with $f > 0.1$, since the condensation for $f=0.1$ was already very small. For this lower density setup, the Field length increases by a factor 2. For $f> 0.1$ the Field length is of the order of the size of the cloud, which drastically alters its evolution. Besides rapid evaporation of any stripped cloudlets, there is another effect at play similar to the strong ($f=0.2$) conduction runs of the fiducial setup. Qualitative examination shows that the cloud stays mostly intact over its full evolution, with little to no stripping. The formation of KH instabilities is thus much more strongly suppressed than for the fiducial setup. The combined effect of less efficient cooling and a larger Field length is that condensation is near completely quenched for the simulation with thermal conduction at $f=0.1$. Since we see similar effects for the simulation setup with lower metallicity, it could be possible that there is a turn-off point in parameter space where thermal conduction completely inhibits condensation.
\begin{figure}
\includegraphics[width=0.5\textwidth]{imgs/lowrho_comp_def.pdf}
\caption{Cold gas evolution (top) and the mixing fraction (bottom), as compared between the fiducial simulation setup and the low-density ($n_{\rm cloud} = 0.1\ \text{cm}^{-3}$, $n_{\rm hot} = 5\times10^{-4}\ \text{cm}^{-3}$) setup.}
\label{fig:coldgas_lowrho}
\end{figure}
In MHD simulations there is an additional effect as the plasma-$\beta$ parameter ($\beta=\frac{P_{\rm Th}}{P_{\rm mag}}$, where $P_{\rm Th}$ is the thermal pressure, and $P_{\rm mag}$ is the magnetic pressure) is now a factor 2 lower with respect to the fiducial setup, which means that magnetic fields are more dynamically important in this setup. The combined effect of this, and less efficient cooling is clear: almost no condensation occurs at all field strengths and orientations except a weak, parallel field, as shown in the middle panel of Fig.~\ref{fig:coldgas_tc}. We have run the strong field simulations for this setup only until $\sim0.8t_\text{cc}$, due to numerical instabilities, likely associated with the dynamically important magnetic field. Since the strong field runs have shown a consistent trend in the other simulation setups, we argue that 0.8$t_{\rm cc}$ gives a good indication of its behavior at 1.6$t_{\rm cc}$. The simulation with a strong, transverse magnetic field is not only suppressed in condensation, but shows a decrease in cold gas mass with time (evaporation).
Similar to the 10\% solar metallicity setup, the efficiency of thermal conduction is lower as compared to the fiducial setup. In this case the suppression effect due to thermal conduction is stronger for both the HD and MHD runs (Fig.~\ref{fig:all_suppressions}, third row). We find for the lower density setup that perpendicular magnetic fields can be approximated with a thermal conduction efficiency of $f\approx0.05-0.10$. For the parallel weak field, $f\approx 0.01$.
\subsection{Varying relative velocity} \label{sec:wind_velocity}
In- or out-flowing gas clouds have been observed to have a wide range of velocities \citep[e.g.][]{Wakker_2007, Boomsma_2008}, and a typical distinction is made between IVC's with typical galactic fountain velocities of $\lesssim100$ km/s \citep{Fraternali_2008, Marasco_2012} with respect to the disc gas, and HVCs with higher velocities. To some extent, we can consider this distinction valid also for relative velocities between clouds and a hot halo \citep{Marinacci_2011b, Pezzulli_2016, Pezzulli_2017, Tepper-Garci_2019}. We assess the effect of the relative velocity on the evolution of our clouds by increasing it to 150 km/s ($\mathcal{M}\approx0.70$). Cloud-wind simulations with higher Mach numbers are generally harder to run numerically, but their evolution is also faster. Note that $t_\text{cc}$ is a factor 2 smaller for this HVC setup, thus these simulations are performed until $t=30$ Myrs. We show the cold gas mass evolution and mixing fractions for HD simulations in Fig.~\ref{fig:hvc_comp}. Note that similar to the lower density setup we have not performed HD simulations with $f>0.1$ since the condensation for the $f=0.1$ simulation was already very small. There is a substantial delay in the onset of condensation for the high-velocity setup. This is due to increased adiabatic heating, as shown in \citet{Gronnow_2018}.
\begin{figure}
\includegraphics[width=0.5\textwidth]{imgs/hvc_comp_def.pdf}
\caption{Cold gas mass evolution (top) and mixing fraction (bottom) for a HVC ($v_{\rm rel} = 150$ km/s) as compared to the fiducial setup. Note that $t_{\rm cc}$ is a factor 2 smaller in this HVC setup with respect to the fiducial case.}
\label{fig:hvc_comp}
\end{figure}
For the MHD simulations, we show the suppression effect in Fig.~\ref{fig:all_suppressions} on the fourth row, and the cold gas evolution in Fig.~\ref{fig:coldgas_tc} on the third row. In terms of the cloud-crushing time, we notice that there is a significant delay in the typical exponential growth rate of cold gas seen in other simulation setups \citep[see also][]{Fraternali_2015}. Even after $1.6t_\text{cc}$ the simulation with a strong, transverse magnetic field shows a net decrease in the amount of cold gas (evaporation). However, after $\approx 1.2t_\text{cc}$ this seems to increase again indicating that condensation is not indefinitely suppressed.
\section{Discussion}\label{sec:discussion}
We have shown and compared the evolution of HD simulations with isotropic thermal conduction to MHD simulations with anisotropic thermal conduction. In particular, we have focused on the evolution of the cold gas mass (condensation). We isolated the effect of thermal conduction on the condensation by dividing the condensation of runs without thermal conduction by the condensation of runs with thermal conduction. We then used the resulting plots (Fig.~\ref{fig:all_suppressions}) to find the HD thermal conduction efficiency $f$ by direct comparison. We now show a summary plot of $f$ as a function of time where we have compared the real suppression from the MHD results to the approximate suppression from HD results in Fig.~\ref{fig:summary_plot}. We show values after $\sim0.8t_{\rm cc}$ to allow the system to move away from its idealised initial conditions. To obtain these values we interpolate the MHD results to the curves in the HD results predicting a certain $f$. An upper limit of $f=0.15$ seems to be valid for all simulations, consistent with the upper limit of $f=0.20$ as found by \citet{Narayan_2001} for a tangled magnetic field. A clear dichotomy is also seen in different field orientations: perpendicular fields average to $0.07\pm0.04$, whereas parallel fields average to $0.03\pm0.02$, where the uncertainties are the 1$\sigma$ bounds calculated from the data points shown in Fig.~\ref{fig:summary_plot}. The difference is mainly due to the strongly suppressed stripping in the parallel field case as opposed to the perpendicular field, since thermal conduction can efficiently evaporate the stripped cloudlets. Despite the large spread in thermal conduction efficiencies, these values confirm that magnetic fields have an important effect in suppressing thermal conduction to $\lesssim10\%$ of the Spitzer value. Below we discuss the effect of resolution on our results and other potential limitations.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{imgs/summary_plot.pdf}
\caption{Linearly interpolated values for the $f$-factor sampled at regular time intervals. We do not plot the first $\sim0.8 t_\text{cc}$ since the system needs time to avoid a bias from its idealised initial conditions. There is a clear dichotomy between perpendicular and parallel magnetic fields, where thermal conduction is much less efficient in the parallel field case.}
\label{fig:summary_plot}
\end{figure*}
\subsection{Convergence tests}
We have verified for a subset of our simulations whether the results in terms of their cold gas mass evolution are converged. We perform full convergence tests for runs with only radiative cooling, radiative cooling and thermal conduction at $f=0.1$, and for radiative cooling and a weak, transverse magnetic field. We have not run the convergence test for simulations with radiative cooling, thermal conduction \textit{and} magnetic fields to 60 Myrs due to computational constraints.
It has been estimated that runs with only radiative cooling shatter into cold cloudlets with size $\sim0.1{\rm pc}/n$ $\approx 0.5$ pc, where $n$ is the number density of the cold material \citep{Mccourt_2018}. These sub-parsec scales are far out of our reach in terms of computational power for 3D simulations. However, thermal conduction effectively evaporates instabilities smaller than the Field length ($\lambda_{\rm F} \approx 40$ pc for $f=0.1$). For simulations with thermal conduction the resolution necessary for convergence is likely more achievable, which was also noticed by \citet[][see their Fig. 4]{Armillotta_2016}. We show our resolution studies in terms of condensation in Fig.~\ref{fig:convergence}. Note that the turbulent velocities that were added to the cloud can affect the condensation by $\sim 5\%$.
As expected, the simulation with only radiative cooling is not converged in terms of condensation. However, with the inclusion of thermal conduction at $f=0.1$ the condensation seems to be converged to a good degree at $\mathcal{R}_{32}$. We have not performed convergence tests for runs with other values of $f$. Since our run with $f=0.1$ seems to evaporate the small scales that inhibit convergence for the cooling only run, we argue that our runs with even stronger thermal conduction with $f=0.15$ and $f=0.2$ are also converged. By the same argument, we expect that our runs with $f=0.01$ and $f=0.05$ are likely only marginally converged.
Magnetic fields can also suppress the formation of hydrodynamical instabilities \citep{Sur_2014}. \citet{Gronnow_2018} found that increasing the resolution for MHD simulations \textit{decreases} the amount of condensation, which was attributed to the width of the wake being unresolved for lower resolution. Such wider wakes could make mixing occur more efficiently. In this work, we do not see this effect. Instead, we find all resolutions predict a $\approx20\%$ increase of cold gas after $1.6t_\text{cc}$. This thin wake effect could however be obfuscated since we have added turbulent velocities to the cloud.
Finally, considering that our resolution study seems to suggest convergence for runs including thermal conduction \textit{or} a magnetic field, we might naively expect that it is likely that simulations including both thermal conduction \textit{and} magnetic fields are also converged at $\mathcal{R}_{32}$. However, the effects of anisotropic thermal conduction could affect the results differently. For this setup we have run the $\mathcal{R}_{64}$ simulation to 12 Myrs ($\approx 0.4t_{\rm cc}$), after which it was deamed computationally unfeasible. While 12 Myrs is too short to draw a reliable conclusion, we notice that the evolution is closest to that of the $\mathcal{R}_{32}$ run.
\begin{figure*}
\centering
\includegraphics[width=1.0\textwidth]{imgs/convergence_all.pdf}
\caption{Convergence tests in our primary diagnostic for our fiducial simulation setup for radiative cooling (top left), radiative cooling and conduction at $f=0.1$ (top right), radiative cooling and a weak ($0.1\mu G$) perpendicular magnetic field with (right) and without (left) thermal conduction (bottom plots).}
\label{fig:convergence}
\end{figure*}
\subsection{Limitations and Future Work} \label{sec:Limitations}
While our simulations contain the typically dominant physical processes: radiative cooling, thermal conduction and magnetic fields, we have not taken into account some physics. Notably, we have not included self-gravity. Our cloud mass is $2.6\times10^4\ M_\odot$, which is much less than the Jeans mass ($M_{\rm J}\approx1.6\times10^8\ M_\odot$), so self-gravity is likely not important in this case. This conclusion is consistent with the findings of \citet{Li_2020}. Other physics not included in this work that were investigated by \citet{Li_2020} are viscosity and self-shielding of cloud material, which are also likely not important for our cloud and halo parameters.
Our cooling function assumes collisional ionisation equilibrium tables from \citet{Sutherland_1993}, which are slightly outdated. The collisional ionisation equilibrium assumes that the medium is optically thin for its own radiation, and that all ions are in their ground states \citep[e.g.][]{Dopita_2003}. These assumptions are very basic and in principle a chemical network based approach should be used to properly account for radiative cooling \citep[see e.g.][]{Salz_2015}. However, we do not expect significant changes to our results for different cooling functions. Furthermore, we implemented an artificial cooling floor to account for heating by a UV background. This prevents any gas from cooling to very low temperatures, but since we classify gas as cold when T $< 2\times10^4$ K the effect on the condensation will likely be negligible.
We have used ordered magnetic fields for the results presented in this work. This description is idealised, since the real magnetic field in galactic halos likely has a random component \citep[e.g.][]{Klessen_2010}. However, previous work on cold clouds interacting with a wind with a random magnetic field indicates that the magnetic field still drapes around the cloud \citep{Asai_2007, Sparre_2020}. As such, the cloud will still locally experience a mostly ordered field, similar to our idealised setup. Therefore we do not expect significantly different results for setups including a random magnetic field with respect to our perpendicular field setup, which we consider to be our most realistic setup.
\section{Conclusions}\label{sec:conclusions}
Fully 3D MHD simulations including anisotropic thermal conduction of "cloud-wind" systems have not been investigated thoroughly in the literature. In this work we presented such simulations for several magnetic field strengths, orientations and parameter setups. We furthermore presented 3D HD simulations with isotropic thermal conduction at a certain efficiency $f$. Our main goal was to compare the evolution of MHD simulations with anisotropic thermal conduction to HD simulations with isotropic thermal conduction at a certain $f$.
We used the cold gas mass evolution as our primary diagnostic, and found that thermal conduction suppresses condensation in almost every case. For both HD and MHD simulations we isolated the effects of thermal conduction by dividing the cold gas mass evolution for runs with no thermal conduction to runs with thermal conduction, and directly compared the curves to find a suitable efficiency $f$.
We can draw the following conclusions:
\begin{enumerate}
\item In HD simulations, isotropic thermal conduction always suppresses condensation, where the suppression is stronger for more efficient thermal conduction (larger $f$). This was found to be primarily due to the evaporation of small, stripped cloudlets. Condensation was found to occur most prominently close to the galactic disc ($n_{\rm hot} = 10^{-3}$ cm$^{-3}$), and was found to be substantially smaller farther away from the galaxy ($n_{\rm hot} < 5\times10^{-4}$ cm$^{-3}$). Therefore, condensation is inefficient in the outer CGM.
\item In MHD simulations, anisotropic thermal conduction was found to suppress condensation more prominently for a magnetic field perpendicular to the relative velocity than for a parallel magnetic field. This was found to be due to the lack of stripped material in the parallel magnetic field setup as compared to perpendicular fields. The simulation with an oblique field produced very similar results to the perpendicular field setup, which showed that the perpendicular field setup is the most realistic.
\item There is a dichotomy in efficiencies between parallel and transverse magnetic field orientations due to the distinct magnetic field interactions between the cloud and the hot medium. Transverse fields were found to have a value of $f$ in the range $0.03-0.15$, whereas parallel fields have a value of $f$ in the range $0.01-0.06$. Since perpendicular fields have been found to be the most representative field setup, an efficiency of $f$ in the range $0.03-0.15$ is likely most realistic.
\item In essentially all simulation setups there is condensation of hot halo gas. This condensation in the lower CGM environments of Milky Way-like galaxies seems unavoidable. The picture emerges that fountain accretion likely plays an important role in the recycling of halo material and can be an important regulator of the observed star formation rate in Milky Way-like galaxies.
\end{enumerate}
Our resolution study shows that both thermal conduction and magnetic fields can aid in reaching convergence in terms of the total amount of cold gas formed as a function of time. This is a significant result: the shattering hypothesis \citep{Mccourt_2018, Sparre_2018, Liang_2020} argues that thermal instability can shatter cold gas to very small scales. However, thermal conduction actively evaporates the smallest scales, which is likely why our run with thermal conduction appears to reach convergence. It should be noted that this is in the context of clouds moving with significant velocities with respect to the surrounding medium, and our simulations do not exclude that static or slow moving clouds in a tangled magnetic field will exhibit shattering.
We can conclude that \textit{thermal conduction plays an important role in the CGM}. In almost all simulations thermal conduction was shown to reduce condensation. Since thermal conduction is highly temperature dependent ($F \propto T^{5/2}\nabla T$), we expect that its effect on the condensation will be greater for higher temperatures. This means that galaxies with a higher virial temperature ($T>2\times10^6$ K), i.e. higher virial mass, likely experience less effective, or no fountain accretion \citep{Armillotta_2016}.
While the approximation of a global efficiency of thermal conduction can in principle be used to simulate suppression due to magnetic fields, we urge caution. Magnetic fields have strong effects on cloud morphology, condensation, and in some cases the momentum transfer between cloud and halo. Therefore we stress that to obtain reliable estimates of accretion rates into galaxies and other properties linked to cloud-halo interactions in the CGM one must include magnetic fields and fully anisotropic thermal conduction.
\section*{Acknowledgements}
We thank an anonymous referee for their insightful comments and suggestions.
AG and FF acknowledge support from the Netherlands Research School for Astronomy (Nederlandse Onderzoekschool voor Astronomie, NOVA), Phase-5 research programme Network 1, Project 10.1.5.7.
The authors thank F. Marinacci, L. Armillotta for useful dicussions. We made use of \textsc{visit} \citep{VisIt}, and \textsc{paraview} \citep{Ahrens_2005} for 3D rendering. We have also made extensive use of \textsc{python} \citep{vanRossum_2009}, and \textsc{python} packages \textsc{yt} \citep{Turk_2010}, \textsc{numpy} \citep{Harris_2020}, and \textsc{matplotlib} \citep{Hunter_2007}, for the analysis and visualisation in this work.
We would furthermore like to thank the Center for Information Technology of the University of Groningen for their support and for providing access to the Peregrine high performance computing cluster.
\section*{Data Availability}
Data available on request.
\bibliographystyle{mnras}
|
\section{The lattice Green function and its decay}
\subsection{Introduction}
\label{sec:intro}
The \emph{lattice Green function}
is defined to be the Fourier integral
\begin{equation}
\label{e:Cax}
C_{a}(x) =
\int_{[-\pi,\pi]^d} \frac{e^{ik\cdot x}}{a^2 +1 - \hat D(k)} \frac{\mathrm{d} k}{(2\pi)^d}
\end{equation}
with $x \in \bb Z^d$, $a \ge 0$, and
\begin{equation}
\hat D(k) = \frac 1d \sum_{j=1}^d \cos k_j \qquad (k = (k_1,\ldots,k_d) ).
\end{equation}
The integral \eqref{e:Cax}
converges for all $a >0$ in all dimensions $d \ge 1$, but converges for
$a=0$ only for the transient case
$d>2$ since the denominator of the integrand is quadratic in small $k$ when $a=0$.
The integral $C_0(0)$ in dimension $d=3$ is a Watson integral that has been evaluated
explicitly
in \cite{Wats39}; see \cite{Zuck11,Gutt10} for the interesting history and further
developments.
The lattice Green function derives its name from the fact that it is equal to the $0,x$ matrix element
of $(-\Delta+a^2)^{-1}$
(the inverse is as an operator on $\ell_2(\bb Z^d)$).
Here $\Delta$ is the Laplacian on $\bb Z^d$, i.e.,
\begin{equation}
\label{e:DeltaJ}
\Delta = D- I
\end{equation}
where $D$ is the $\bb Z^d \times \bb Z^d$ matrix with $D_{xy}=\frac{1}{2d}$ if $\|x-y\|_1=1$,
and otherwise $D_{xy}=0$, and $I$ denotes the identity matrix.
Thus $C_a$ is the resolvent of the lattice Laplacian.
In the physics literature,
$C_a$ is often called the \emph{Euclidean lattice scalar propagator}.
Our purpose is to study the precise asymptotic behaviour of $C_a(x)$ as $x$ goes
to infinity, with emphasis on how this behaviour depends on values of $a$
close to the critical value $a =0$, and on how Euclidean invariance is restored in the
small $a$ limit.
Let $m_a$ denote the unique nonnegative solution to $\cosh m_a = 1+da^2$.
For $a>0$, an elementary proof that $m_a$ is the
exponential rate of decay of $C_{a}(x)$ when $x\to\infty$ along a coordinate axis,
and that
$C_{a}(x) \le C_{a}(0)e^{-m_a\|x\|_\infty}$ for all $x \in \bb Z^d$, is given in
\cite[Theorem~A.2]{MS93} (with a change of notation $2dz=\frac{1}{1+a^2}$).
On the other hand, for the critical case $a=0$ it is well-known
that there is instead polynomial
decay $\|x\|_2^{-(d-2)}$ \cite{LL10}.
We will prove, in a unified way and with precise constants for
the amplitudes in the asymptotic formulas, that there are the following four regimes of decay
for dimensions $d \ge 1$ (with the restriction $d>2$ for regime (IV)):
\begin{alignat*}{4}
& {\rm (I)} &\qquad \text{fixed }a >0
& \qquad\text{anisotropic OZ}
& \qquad m_a^{(d-3)/2} |x|_a^{-(d-1)/2}e^{-m_a|x|_a}
\\
&{\rm (II)}
&\qquad a\|x\|_2\to \infty, \; a^3\|x\|_2\to 0
&\qquad \text{isotropic OZ}
&\qquad a^{(d-3)/2} \|x\|_2^{-(d-1)/2}e^{-\sqrt{2d} a\|x\|_2}
\\
&{\rm (III)}
&\qquad \text{fixed }a\|x\|_2 >0
&\qquad \text{massive continuum}
&\qquad \|x\|_2^{-(d-2)}
\\
&{\rm (IV)}
&\qquad a = 0 \;\;(d>2)
&\qquad \text{massless continuum}
&
\qquad \|x\|_2^{-(d-2)}
.
\end{alignat*}
The decay in regime (I) is called \emph{Ornstein--Zernike decay}.
The norm $|\cdot|_a$ is an explicitly defined
$a$-dependent anisotropic norm on $\bb R^d$ (not an $\ell^p$ norm).
Ornstein--Zernike decay is widely studied and has been proved for a variety
of subcritical statistical mechanical models, e.g., \cite{CIV08,CC86b,CCC91}.
These proofs for much
more difficult models than the lattice Green function
show decay of the form $|x|_a^{-(d-1)/2}e^{-m_a|x|_a}$,
but do not however reveal
the factor $m_a^{(d-3)/2}$ for small $a>0$. Indeed, a solution to the latter
problem would be tantamount to a control of the critical behaviour of those models, a topic with
difficult unsolved problems of great current interest.
The decay in regime (II) is also Ornstein--Zernike decay, but the mass in regime (I)
is now replaced by its asymptotic form $\sqrt{2d} a$ as $a \to 0$, and Euclidean
invariance is restored since the norm
$|x|_a$ from regime (I) is replaced by the Euclidean norm.
This is natural: we will prove that $\lim_{a \to 0}|x|_a
=\|x\|_2$.
The decay in regimes (III) and (IV) is in fact expressed in terms of the
continuum Green function for the Laplacian on $\bb R^d$: the massive Green function
in regime (III) and massless Green function in regime (IV).
These continuum Green functions appear explicitly in the full asymptotic formulas
in these regimes. In both cases,
the decay is Euclidean invariant and is expressed in terms of the $\ell_2$ norm.
The transition from regime (II) to (III) can be
anticipated
by replacing $a$ in (II)
by $\|x\|_2^{-1}$, which corresponds to $x$ on the order of the
\emph{correlation length} $m_a^{-1}$.
This replacement causes the asymptotic formula in (II) to transform
into the formula in (III).
More generally, for real numbers $q\in (0,\infty)$ we consider the decay of
\begin{equation}
\label{e:Cxq}
C^{(q)}_{a}(x)=
\int_{[-\pi,\pi]^d}\frac{e^{ik\cdot x}}{(a^2+1-\hat{D}(k))^q}\frac{\mathrm{d} k}{(2\pi)^d}
\end{equation}
in the above four regimes.
When $q \ge 2$ is a positive integer, $C^{(q)}_{a}$ is the $q$-fold convolution of
$C_a$ with itself.
For $q=2,3,4$, $C_a^{(q)}(0)$ is known respectively as the bubble, triangle and square
diagram. These diagrams play an important role in the study of various statistical mechanical
models above their upper critical dimensions, especially when $a=0$; see, e.g., \cite{Slad06}.
For integers $q \ge 2$, $C^{(q)}_{0}(x)$ is the critical lattice polyharmonic
Green function. Polyharmonic functions have been widely studied, especially on $\bb R^d$
rather than on the lattice $\bb Z^d$ (e.g., \cite{ACL83}).
We note in passing
that the lattice Green function has the following probabilistic interpretation.
Let $X_1,X_2,\ldots$ be independent and identically distributed
random variables with each $X_i$
equally likely to be any one of the $2d$ unit vectors (positive or negative) in $\bb Z^d$,
for any fixed integer $d \ge 1$.
For $a \in [0,\infty)$,
let $N$ be a geometric random variable with
\begin{equation}
\P(N=n)=
\Big(\frac{1}{1+a^2}\Big)^{n}
\frac{a^2}{1+a^2} \qquad (n\ge 0),
\end{equation}
with
$N$ independent of the $X_i$. Then $\P(N \ge n)=(\frac{1}{1+a^2})^n$.
Let $S_0=0$, and
consider the nearest-neighbour
random walk $S_n=X_1+\cdots +X_n$ on $\bb Z^d$ subjected to $a$-dependent killing, i.e.,
the walk takes $N$ steps and then dies.
Let $p_n(x)$ denote the probability that the random walk without killing makes a transition
from $0$ to $x$ in $n$ steps.
The expected number
of visits of the random walk to a point $x\in \bb Z^d$ is
\begin{align}
\label{e:Gkappadef}
\bb E \Bigg(\sum_{n=0}^N \mathbbm{1}_{S_n =x} \Bigg)
=
\sum_{n=0}^\infty \Big(\frac{1}{1+a^2}\Big)^{n} p_n(x)
\qquad (x\in \bb Z^d).
\end{align}
The expectation in \eqref{e:Gkappadef}
is equal to $(1+a^2) C_a^{(1)}(x)$, as
a consequence of the fact that the Fourier transform
of $p_n(x)$ is simply
\begin{equation}
\sum_{x\in \bb Z^d}p_n(x) e^{ik\cdot x} = \hat D(k)^n
.
\end{equation}
We do not consider more
general random walks, which would correspond to operators
other than the Laplacian.
We expect that our results should extend to
the Green function
for random walks taking finite-range symmetric steps, but as can be seen in \cite{AIOV21}
the nature of the decay can change when arbitrarily long steps are permitted.
Our motivation to study the decay of the lattice
Green function originates from statistical mechanics.
The long-distance asymptotic behaviour of the two-point function is
an essential feature in the analysis of
critical phenomena in lattice statistical mechanical models such as percolation, the Ising
model, or the self-avoiding walk.
In high dimensions, $\|x\|_2^{-(d-2)}$ decay of the critical two-point function
has been proved in several cases, including \cite{HHS03,Hara08,BHH21,Saka07}.
However, the near-critical behaviour, which merges the subcritical
exponential decay and the power-law critical decay, has received scant
attention despite the fact that it has the potential
to reveal important and hitherto unstudied aspects of the critical behaviour,
particularly for models defined on a torus. Recently progress has been made
in this direction for weakly self-avoiding walk for dimensions $d>4$
\cite{MS22,Slad20_wsaw} and percolation for $d>6$
\cite{HMS22}.
In high dimensions, where mean-field behaviour is known to occur,
the near-critical two-point function is conjectured to have similar decay to that of
the lattice Green function.
It is therefore important to have a detailed understanding of the
long-distance behaviour of the lattice Green function
as a prototype. In this paper, we provide a comprehensive account of the decay
of the lattice Green function.
\subsection{The anisotropic norm}
Lattice effects play a significant role
in the asymptotic behaviour of $C_a^{(q)}(x)$ when
$a>0$ is fixed, and lead to anisotropy in the decay.
The
following definition enters into the description of the anisotropy.
\begin{defn}
\label{def:norm}
Let $d \ge 1$ and $a \ge 0$.
We define the \emph{mass}, or \emph{inverse correlation length}, to be the unique solution
$m_a\ge 0$ of
\begin{equation}
\label{e:m0def}
\cosh m_a = 1+ da^2 .
\end{equation}
For nonzero $x \in \bb R^d$, let $u=u_a(x)\ge 0$ be the unique solution of
\begin{equation}
\label{e:udef1}
\frac{1}{d}\sum_{i=1}^d\sqrt{1+x_i^2u^2} = 1+a^2,
\end{equation}
which exists since the left-hand side of \eqref{e:udef1} is a strictly
increasing function of $u \in [0,\infty)$ onto $[1,\infty)$.
Finally, with the restriction now that
$a>0$, we define
$|0|_a=0$ and for nonzero $x\in\bb R^d$ define
\begin{equation}
\label{e:normdef1}
|x|_a = \frac{1}{m_a}\sum_{i=1}^d x_i \arcsinh (x_i u_a(x)).
\end{equation}
\end{defn}
It follows from \eqref{e:m0def} and Taylor's theorem that, as $a \to 0$,
\begin{equation}
\label{e:maasy}
m_a = \sqrt{2d}a(1+O(a^2)).
\end{equation}
Equation \eqref{e:normdef1} defines a norm on $\bb R^d$ whose properties are
indicated in Proposition~\ref{prop:norm}. In particular, the norm $|\cdot|_a$ interpolates
between the $\ell_1$ norm when $a=\infty$ and the $\ell_2$ norm when $a=0$.
The norm's unit ball in dimensions $d=2,3$ is depicted in Figure~\ref{fig:norm}.
\begin{figure}[ht]
\centering{
\includegraphics[width=35mm, height=35mm]{norm_dim_2_a_01_green.PNG}
\hspace{5mm}
\includegraphics[width=35mm, height=35mm]{norm_dim_2_a_1_green.PNG}
\hspace{5mm}
\includegraphics[width=35mm, height=35mm]{norm_dim_2_a_10_green.PNG}
\vspace{5mm}
\includegraphics[width=4cm, height=4cm]{norm_dim_3_a_1_iter_40.PNG}
\hspace{5mm}
\includegraphics[width=4cm, height=4cm]{norm_dim_3_a_2_iter_40.PNG}
\hspace{5mm}
\includegraphics[width=4cm, height=4cm]{norm_dim_3_a_20_iter_40.PNG}
\caption{Unit ball for the norm $|\cdot |_a$ in dimensions $d=2,3$.}
\label{fig:norm}}
\end{figure}
\begin{prop}
\label{prop:norm}
Let $d \ge 1$ and $a>0$.
The function $|\cdot|_a$ defines a norm on $\bb R^d$ which is monotone increasing
in $a$ and for all $x\in\bb R^d$ obeys
\begin{equation}
\lim_{a\to 0}|x|_a = \|x\|_2, \qquad \lim_{a\to \infty}|x|_a = \|x\|_1,
\end{equation}
in fact $|x|_a = \|x\|_2(1+O(a^2))$ uniformly in $x \neq 0$.
In particular,
\begin{equation}
\label{e:normineq}
\|x\|_2 \le |x|_a \le \|x\|_1.
\end{equation}
\end{prop}
By definition, $m_a(0)=0$ and
$m_a$ is a strictly positive strictly increasing function of $a>0$.
To understand why the factor $m_a^{-1}$ appears on the right-hand side
of \eqref{e:normdef1}, we note that for any $a>0$ and for any unit vector $e_j\in\bb R^d$,
\begin{align}
\label{e:ue}
u_a(e_j) &= \sqrt{(1+da^2)^2-1} = \sqrt{\cosh^2 m_a -1} = \sinh m_a,
\end{align}
and hence, for all $a > 0$,
\begin{equation}
\label{e:norm1}
|e_j|_a = 1.
\end{equation}
The $|\cdot |_a$ norm originated in the analysis of the 2-dimensional Ising model
\cite[pp.~302--303]{MW73},
although it was not identified there
as a norm. A proof that it defines a norm was given in \cite[Lemma~6.5]{Pfis99};
there the proof of the triangle inequality was based on the second Griffiths inequality
applied to the 2-dimensional Ising model.
We provide a simple alternate proof based on a random walk argument.
We do not know of any direct proof of the triangle inequality
based on the definition of the norm.
Neither are we aware of any prior proof of the monotonicity of the norm.
\subsection{The continuum Green function}
In Appendix~\ref{sec:continuumGreen} we consider and interpret the
integral
\begin{equation}
\label{e:Gx}
G_s^{(q)}(x)
= \int_{\bb R^d} \frac{e^{ik\cdot x}}{(\frac{1}{2d}\|k\|_2^2+s^2)^{q}} \frac{\mathrm{d} k}{(2\pi)^d}
\qquad
(x \in \bb R^d \setminus \{0\}),
\end{equation}
which in the case $q=1$ is the Green function
for the (normalised) continuum Laplace operator on $\bb R^d$.
It follows from \eqref{e:Gx} that there is a scaling relation
\begin{equation}
\label{e:Gscaling}
G_s^{(q)}(x) = s^{d-2q} G_1^{(q)}(s x)
\qquad
(s>0).
\end{equation}
In Appendix~\ref{sec:continuumGreen}, we
recall the elementary proof that
for $s >0$, integers $d \ge 1$, and nonzero $q\in \bb R$, the massive and massless
continuum Green functions are given explicitly
(in the sense of tempered distributions) by \eqref{e:Gscaling} together with
\begin{alignat}{2}
\label{e:Gaqx-2d}
G_1^{(q)}(x)
& =
\frac{2d^q}{\Gamma(q)(2\pi)^{d/2}}
\Bigg( \frac{\sqrt{2d} }{\|x\|_2} \Bigg)^{(d-2q)/2}
\!\!\!\!\!
K_{(d-2q)/2}(\sqrt{2d} \|x\|_2)
,
\\
\label{e:G0qx}
G_0^{(q)}(x)
& =
\frac{d^q \Gamma(\frac{d-2q}{2})}{2^{q}\pi^{d/2}\Gamma(q)}
\frac{1}{\|x\|_2^{d-2q}},
\end{alignat}
where $K_\alpha$ is the modified Bessel
function of the second kind
and for \eqref{e:G0qx} we restrict to $d> 2q$.
For $\alpha >0$ the asymptotic behaviour of $K_\alpha$ is known to be
\begin{align}
\label{e:Kasy}
K_\alpha(z) &\sim \frac{\Gamma(\alpha)}{2} \Big(\frac{2}{z}\Big)^\alpha
\quad (z \to 0),
\qquad
K_\alpha(z) \sim \sqrt{\frac{\pi}{2z}} e^{-z}
\quad (z \to \infty).
\end{align}
\subsection{Asymptotic behaviour}
\label{sec:ab}
\subsubsection{Main result}
The following theorem gives a precise statement of the asymptotic decay of the lattice
Green function for arbitrary dimension $d \ge 1$ and for $a\ge 0$ (possibly $n$-dependent).
The norm $|\cdot|_a$ plays a key role in the anisotropic limit \eqref{e:mr}, for
which lattice effects persist when $a$ is independent of $n$.
Recall that $m_a=\arccosh (1+da^2)$ for $a \ge 0$.
We write $f(n) \sim g(n)$ to mean $\lim_{n\to\infty}f(n)/g(n)=1$.
\begin{theorem}
\label{thm:green-asy}
Let $d \ge 1$ and
$q\in (0,\infty)$ (not necessarily integer). Fix $x \in \bb Z^d \setminus \{0\}$.
\\
{\bf (i) (Ornstein--Zernike decay).}
Let $a_n\in (0,\infty)$ and suppose that $a_nn \to \infty$ in such a manner that
$a_n$ remains bounded
(this includes in particular the case of fixed $a_n=a\in (0,\infty)$).
There exists $c_{a,q,\hat x}>0$ depending on $a$, $q$, and the direction
$\hat x = x/|x|_a$ (and on the dimension $d$), such that, as $n \to \infty$,
\begin{equation}
\label{e:mr}
C_{a_n}^{(q)}(nx)
\sim
c_{a_n,q,\hat x}
\frac{m_{a_n}^{(d-1-2q)/2}}{(n|x|_{a_n})^{(d+1-2q)/2} }
\;
e^{-m_{a_n}n|x|_{a_n}} .
\end{equation}
The ratio of the above left- and right-hand sides
converges to $1$ uniformly in nonzero $x$, and
the constant $c_{a,q,\hat x}$ has the explicit $\hat x$-independent limit
\begin{equation}
\label{e:calim}
c_{0,q}
=
\lim_{a \to 0}c_{a,q,\hat x}
=
\frac{d^q}{(2\pi)^{(d-1)/2}\Gamma(q)}.
\end{equation}
{\bf (ii) (Critical decay).}
Let
$a_n = s /n$ with $s \in [0,\infty)$, with $d>2q$ if $s=0$.
Then, as $n \to \infty$,
\begin{equation}
\label{e:mr-critical}
C_{a_n}^{(q)}(nx)
\sim
\frac{1}{n^{d-2q}} G_{s}^{(q)}(x)
.
\end{equation}
\end{theorem}
The asymptotic formula \eqref{e:mr} encompasses both regimes (I) and (II)
mentioned in Section~\ref{sec:intro}.
The anisotropic OZ regime (I) is the case of fixed $a_n=a>0$, for which the
anisotropic norm $|x|_a$ plays a role.
For the isotropic OZ regime (II), we
are interested in the case where $a_n\to 0$ in such a way that $a_nn\to\infty$
and $a_n^3n\to 0$.
Recall from \eqref{e:maasy}
and Proposition~\ref{prop:norm} that $m_a=\sqrt{2d}a(1+O(a^2))$ and
$|x|_a=\|x\|_2(1+O(a^2))$. Consequently, as $a_n \to 0$ we have
\begin{equation}
m_{a_n} n |x|_{a_n} = \sqrt{2d} a_n n \|x\|_2(1+O(a_n^2))
\end{equation}
and hence it follows from \eqref{e:mr} that if $a_nn \to\infty$ then
\begin{equation}
\label{e:mr-OZ-error}
C_{a_n}^{(q)}(nx)
\sim
c_{0,q}
\frac{(\sqrt{2d}a_n)^{(d-1-2q)/2}}{(n\|x\|_{2})^{(d+1-2q)/2}}
e^{-\sqrt{2d} a_nn \, \|x\|_2[1+O(a_n^2)]}.
\end{equation}
If we now assume additionally that $a_n^3n\to 0$ then the error term in the
exponential can be neglected and we obtain the result claimed for regime (II), namely
\begin{equation}
\label{e:mr-OZ-a}
C_{a_n}^{(q)}(nx)
\sim
c_{0,q}
\frac{(\sqrt{2d}a_n)^{(d-1-2q)/2}}{(n\|x\|_{2})^{(d+1-2q)/2}} e^{-\sqrt{2d} a_nn \, \|x\|_2}.
\end{equation}
If the condition $a_n^3n\to 0$ is violated then we see from \eqref{e:mr-OZ-error} that
modifications to the exponential decay will occur from the error term in the exponent.
The massive critical regime (III) and the massless critical regime (IV)
are respectively the $s >0$ and $s =0$ cases of \eqref{e:mr-critical}.
There is coherence between regimes (II)
and (III) in the sense that if $a_n$ in (II) is replaced by $s/n$ then the exponential
factor becomes a constant and the powers $a_n^{(d-1-2q)/2} n^{-(d+1-2q)/2}$ reduce
to an $s$-dependent multiple of $n^{-(d-2q)}$. The continuum Green function
$G_s^{(q)}(x)$ provides the amplitude for the asymptotic decay in the critical regimes.
There is no statement of uniformity in $x$ in \eqref{e:mr-critical} because uniformity is
impossible for $s>0$:
e.g., if $x=n^2y$ with $y$ independent of $n$ then $\frac{s}{n} x
=s ny \to \infty$ as $n \to \infty$ and we are actually in regime (I), not regime (III).
\subsubsection{Previous results}
The proof of Theorem~\ref{thm:green-asy} is based on the representation
\begin{align}
\label{e:BesselC}
C^{(q)}_{a}(x)
&=
\frac{1}{\Gamma(q)}\int_0^\infty t^{q-1}e^{-(a^2+1)t}
\prod_{j=1}^{d}I_{x_j}(t/d) \mathrm{d} t
\end{align}
in terms of the modified Bessel function of the first kind.
Much of Theorem~\ref{thm:green-asy} has been proved previously by other authors,
and we now describe what was done previously and how our approach simplifies, extends and
unifies earlier work.
For $q=1$ and for fixed $a_n=a>0$, the asymptotic formula \eqref{e:mr} is proved
in \cite[Theorem~3.2]{MY12} for $d \ge 1$, and for $d=2$ in
\cite[Proposition~13]{Mess06}.
Neither of those references identified the role of the anisotropic norm in
\eqref{e:mr}, and the norm
makes the statement significantly more transparent.
In \cite[Theorem~3.3]{MY12},
\eqref{e:mr-OZ-a} is stated to hold in
the limit in which $a_n \to 0$ with $a_nn\to \infty$; in fact
the further restriction $a_n^3n\to 0$ is necessary for the
simplification of the exponential in \eqref{e:mr} to yield the isotropic form \eqref{e:mr-OZ-a}.
Our method of proof is based on the Laplace method as in
\cite{MY12} but it is simplified by our appeal to well-established properties of the modified
Bessel function rather than deriving them as part of the proof as in \cite{MY12}.
Also, unlike the separate proofs for the anisotropic and isotropic cases in \cite{MY12},
we give one unified proof.
The massive critical regime was considered in \cite{PS99}
(indeed these authors computed higher-order terms as well), but the arguments used in \cite{PS99} do not constitute a proof.
The formula \eqref{e:mr-critical} for $q=1$ and
$s > 0$ can be inferred from
the statement of \cite[Proposition~3.1]{DGGZ22}, which is proved via
the local central limit theorem.
Our proof, which again uses known properties of the modified Bessel function,
involves a straightforward application of the dominated convergence theorem and does
not involve the Laplace method.
For the massless critical regime (IV),
the asymptotic behaviour of the
critical lattice polyharmonic Green function is given in \cite{Mang67} as
\begin{equation}
\label{e:Cq0}
C^{(q)}_{0}(x) \sim
\frac{d^q\Gamma(\frac{d-2q}{2})}{2^{q}\pi^{d/2}\Gamma(q)} \frac{1}{\|x\|_2^{d-2q}}
\qquad
(q=1,2,3,\ldots; \; d>2q)
\end{equation}
with explicit higher-order correction term. Since higher-order terms
are known we make no effort
here to compute them, as our focus in the proof is on simplicity. We prove \eqref{e:Cq0}
as the $s=0$ case of \eqref{e:mr-critical} for arbitrary
real $q\in (0,\infty)$ when $d>2q$. This special case of
our proof of \eqref{e:mr-critical} in the entire critical regime $s \ge 0$
does not require separate attention.
When $q=1$, \eqref{e:Cq0}
gives the well-known decay of
the critical lattice Green function.
In fact, the $\|x\|_2^{-(d-2)}$ decay in \eqref{e:Cq0} for $q=1$
holds more generally
under a second-moment condition for $D_{xy}$ (recall \eqref{e:DeltaJ}), with error term
of order $\|x\|_2^{-d}$ with known coefficient. For $q=1$ see, e.g.,
\cite[p.~82]{LL10} or \cite[Theorem~3.4]{MY12}, or \cite[p.~308]{Spit76} for $d=3$, and for
further error terms see
\cite{Uchi98}. A version of \eqref{e:Cq0} for $q=1$ holds under certain conditions even when
the transition matrix $D_{xy}$ is permitted to assume negative values \cite{Hara08}.
\subsubsection{Explicit calculation for $d=1$}
For $d=1$ and integers $q\ge 1$, the
condition $d=1>2q$ is violated and $C_0^{(q)}=\infty$, so regime (IV) does not apply.
The computation of $C_a^{(q)}(x)$ for $d=1$, integer $q\ge 1$, and $a>0$ can be
done explicitly
with the result that
\begin{align}
\int_{-\pi}^\pi \frac{e^{ikx}}{(1+a^2-\cos k)^q} \frac{dk}{2\pi}
& =
\frac{e^{-m_a|x|}}{\sinh^q m_a}
\sum_{l=0}^{q-1}
\binom{|x|+q-1}{q-1-l} \binom{q-1+l}{l}
\Big( \frac{e^{-m_a}} {2\sinh m_a}
\Big)^l
\label{e:d1generalq}
\end{align}
with $m_a = \arccosh (1+a^2)$.
The above formula can be verified by residue calculus or by
an appropriate rewriting of the formula \cite[(3.616.7)]{GR07}.
In detail, the cases $q=1$ and $q=2$ are
\begin{align}
\label{e:d1q1}
C_a^{(1)}(x)
&=
\frac{e^{-m_a|x|}}{\sinh m_a}
\qquad\qquad\qquad\qquad\qquad\qquad\qquad\;\;
(d=1),
\\
\label{e:d1q2}
C_a^{(2)}(x)
&=
\frac{|x| e^{-m_a|x|}}{\sinh^2 m_a}
\Bigg[ 1 + \frac{1}{|x|} \Big(1 + \frac{e^{-m_a}} {\sinh m_a} \Big) \Bigg]
\qquad\quad
(d=1) .
\end{align}
Both of the formulas \eqref{e:d1q1}--\eqref{e:d1q2} refine and are consistent with
\eqref{e:mr} and \eqref{e:mr-critical} from Theorem~\ref{thm:green-asy}.
In particular, for $a=s/n$ with fixed $s>0$, \eqref{e:d1q1} gives
\begin{equation}
C_{s/n}^{(1)}(nx) \sim \frac{n}{\sqrt{2}s}e^{-\sqrt{2} s |x|},
\end{equation}
and since $K_{-1/2}(y) = K_{1/2}(y) = \pi^{1/2}(2y)^{-1/2}e^{-y}$ for $y >0$ (see \cite[8.432.8]{GR07})
this agrees with \eqref{e:mr-critical}.
\subsubsection{Ornstein--Zernike vs critical decay}
In the physics literature, the inverse mass $\xi_a=m_a^{-1}$ is known as the
\emph{correlation length}.
With $a=s/n$ and $s >0$,
Theorem~\ref{thm:green-asy} can then be interpreted informally as identifying the
following decay of the lattice Green function:
\begin{alignat}{3}
&s >0 &\qquad n\|x\|_2 \asymp \xi_a
&\qquad \text{massive continuum limit},
\\
& s \to \infty &\qquad n\|x\|_2 \gg \xi_a
&\qquad \text{Ornstein--Zernike decay}.
\end{alignat}
For the latter case, we see the Euclidean invariance if $s=o(n)$ but not for
$s = an$
with fixed $a>0$.
The Ornstein--Zernike and critical regimes occur in general dimensions
in lattice statistical mechanical
models such as the self-avoiding walk, percolation, and the Ising model
\cite{CIV08,CC86b,CCC91}.
This perspective is standard in the physics literature but a mathematical description
of the near-critical behaviour which crosses over between the two regimes
is lacking in most examples, even in high dimensions where the lace expansion applies.
The asymptotic formula \eqref{e:mr} provides a prototype for what can be expected
for the near-critical two-point functions of
the high-dimensional statistical mechanical models.
The bounds in regimes (I)--(II) in general do \emph{not} hold uniformly in all
$a > 0$, $n \ge 1$, and nonzero $x\in\bb Z^d$. This is evident in the
explicit formula \eqref{e:d1q2} for $d=1$ and $q=2$,
where
the first term $|x| \frac{e^{-m_a|x|}}{\sinh^2 m_a}$
dominates when $x \to \infty$ with fixed $a$, in agreement with \eqref{e:mr},
whereas with fixed $x$, in the limit $a\to 0$
we have $m_a\sim \sqrt{2}a \to 0$, the exponentials become insignificant, and \eqref{e:d1q2}
is dominated by the factor $\sinh^{-3}m_a \sim (\sqrt{2}a)^{-3}$
arising from its last term.
This shows the impossibility for this case
of an upper bound of the form $|x|a^{-2} e^{-m_a|x|}$
that is uniform in both $x$ and $a$.
Similarly, for $d>3$ and $q=1$, there can be no upper bound
on $C_a^{(1)}$ of the form
\begin{equation}
\label{e:OZnotuniform}
m_a^{(d-3)/2}\frac{1}{|x|_a^{(d-1)/2}}e^{-m_a|x|_a}
\end{equation}
that is uniform in all $a>0$ and nonzero $x \in \bb Z^d$,
because \eqref{e:OZnotuniform} vanishes as $ a\to 0$ with fixed $x$
due to the factor $m_a^{(d-3)/2}$, whereas if $|x|_a$ grows like $m_a^{-1}$ then
$C_a^{(1)}$ is in regime (III) and decays as a multiple of $\|x\|_2^{-(d-2)}$.
It remains an open problem to determine for which values of $d,q$ the formula
\eqref{e:mr} in fact gives a bound which is unform in $a >0$ and nonzero $x$.
On the other hand,
for $q=1$ and $d>2$ an upper bound that \emph{is} uniform in $a \ge 0$ and in $x$
is given in \cite[Proposition~2.1]{Slad20_wsaw}, which asserts that
there are constants $\kappa_1>0$ and $\kappa\in (0,1)$ such that
for all $a \ge 0$ and all $x \neq 0$,
\begin{equation}
\label{e:srwbd}
C^{(1)}_a(x) \le \kappa_1 \frac{1}{|x|_a^{d-2}}e^{-\kappa m_a|x|_a}
.
\end{equation}
(By changing the constants, another norm than $|x|_a$ could be used in the above.)
As in \cite[Lemma~3.3]{MS22}, the inequality \eqref{e:srwbd} easily implies that
for general integers $q\ge 1$ and dimensions $d>2q$,
\begin{equation}
\label{e:srwbd-q}
C_a^{(q)}(x) \le \kappa_q \frac{1}{|x|_a^{d-2q}}e^{-\kappa m_a|x|_a}
.
\end{equation}
The uniform upper bound \eqref{e:srwbd} combines
the critical $|x|_a^{-(d-2)}$ decay with the exponential decay for $a>0$.
The relaxation of the exponential decay via $\kappa<1$ compensates for the
differing power laws in \eqref{e:srwbd-q} and in regime (I).
Bounds of the form \eqref{e:srwbd} have been proved and applied to analyse the
critical behaviour of weakly self-avoiding
walk in dimensions $d>4$ \cite{MS22,Slad20_wsaw} and of percolation in dimensions
$d>6$ \cite{HMS22}.
\subsection{Organisation}
The remainder of the paper is organised as follows.
In Section~\ref{sec:Bessel}, we give the elementary derivation of the
representation \eqref{e:BesselC} of $C_a^{(q)}(x)$ in terms of the modified
Bessel function $I_\nu$. This representation in terms of a $1$-dimensional integral
is the basis for all of our analysis. We then recall properties of $I_\nu$ which
enable the asymptotic evaluation of the integral \eqref{e:BesselC}.
In Section~\ref{sec:subcrit}, we prove Theorem~\ref{thm:green-asy}(i), pertaining
to the Ornstein--Zernike regime. In this regime,
the Bessel integral \eqref{e:BesselC} has an exponential
factor in the integrand which makes it amenable to application of the Laplace method.
The norm $|\cdot|_a$ emerges naturally from a computation involving
the critical point which dominates the behaviour arising in the Laplace method.
In Section~\ref{sec:contlim}, we prove Theorem~\ref{thm:green-asy}(ii), pertaining
to the critical regime.
In the critical regime, there is no longer any exponential behaviour in the integrand
of the Bessel integral \eqref{e:BesselC} and there is no need for the Laplace method.
Given the well-known asymptotics for $I_\nu$ recalled in Section~\ref{sec:Bessel},
the proof follows quickly from the dominated convergence theorem.
Finally, in Appendix~\ref{sec:continuumGreen} we provide an elementary proof that
the formulas \eqref{e:Gaqx-2d}--\eqref{e:G0qx} for the continuum Green function are
equal to the integral \eqref{e:Gx} over $\bb R^d$ in the sense of tempered distributions,
and in Appendix~\ref{sec:Besselpf} we discuss properties of the modified Bessel function.
\section{Bessel representation}
\label{sec:Bessel}
For any integer $\nu \geq 0$ and $ t \in \bb R$ the modified Bessel function of the first kind $I_\nu(t)$ is given by
\begin{equation}
I_\nu(t) = \frac{1}{\pi}\int_0^\pi e^{t \cos\theta+i\nu\theta}\mathrm{d} \theta.
\end{equation}
For our purposes it is more useful to consider
\begin{equation}
\label{e:Ibardef}
\bar I_\nu(t)
=
e^{-t} I_\nu(t)
=
\frac{1}{\pi}\int_0^\pi e^{-t (1-\cos\theta)+i\nu\theta}\mathrm{d} \theta
\end{equation}
which has the exponential growth of $I_\nu(t)$ cancelled.
The following lemma provides the well-known integral representation that
is the foundation for the proof of Theorem~\ref{thm:green-asy}.
\begin{lemma}
\label{lem:CBessel}
For $d \ge 1$, $a \ge 0$, $q>0$, $x \in \bb Z^d$, and with the restriction $d>2q$ when $a=0$,
\begin{align}
\label{e:Besselrep0}
C^{(q)}_{a}(x)
&=
\frac{1}{\Gamma(q)}\int_0^\infty t^{q-1}e^{-a^2t}
\prod_{j=1}^{d}\bar I_{x_j}(t/d) \mathrm{d} t.
\end{align}
\end{lemma}
\begin{proof}
Let $\hat F(k) =a^2+1-\hat D(k)$.
We use the identity
\begin{align}
\label{e:reciprocal}
\frac{1}{v^{q}} &= \frac{1}{\Gamma(q)}\int_0^\infty t^{q-1}e^{-tv}\mathrm{d} t
\qquad
(v>0)
\end{align}
in the definition \eqref{e:Cxq} to obtain
\begin{align}
\label{e:Besselrep}
C^{(q)}_{a}(x)
&= \int_{[-\pi,\pi]^d}\frac{e^{ik\cdot x}}{\hat F(k)^q}\frac{\mathrm{d} k}{(2\pi)^d}
\nnb
&= \frac{1}{\Gamma(q)}\int_{[-\pi,\pi]^d}
\int_0^\infty t^{q-1}e^{- t\hat F(k)}\mathrm{d} t\;
e^{ikx} \frac{\mathrm{d} k}{(2\pi)^d}\nnb
&= \frac{1}{\Gamma(q)}\int_0^\infty t^{q-1}e^{-a^2t}
\prod_{j=1}^{d}\bar I_{x_j}(t/d) \mathrm{d} t,
\end{align}
and the proof is complete. Note that there is no issue with convergence of this last
integral at $t=0$, and for large $t$ convergence is guaranteed
(assuming $d>2q$ when $a=0$)
by the fact that
$\bar I_\nu(z) \sim (2\pi z)^{-1/2}$ as $z \to \infty$. In particular, this justifies
the above application of Fubini's Theorem.
\end{proof}
To study the Ornstein--Zernike regime we apply the change of variable $t=dnv$
to the integral representation \eqref{e:Besselrep0} to obtain
\begin{align}
\label{e:BesselOZ}
C^{(q)}_{a}(nx)
&=
\frac{d^{q}n^q}{\Gamma(q)}\int_0^\infty v^{q-1}e^{-dna^2v}
\prod_{j=1}^{d} \bar I_{nx_j}(nv) \mathrm{d} v
.
\end{align}
For the continuum regime we will also make the replacement $v=nt/d$ in \eqref{e:BesselOZ}
and use
\begin{align}
C_{{s}/n}^{(q)}(nx)
&=
\frac{n^{2q}}{\Gamma(q)} \int_0^\infty t^{q-1}
e^{-s^2 t} \prod_{j=1}^{d}\bar I_{n x_j} (n^2t/d) \, \mathrm{d} t
.
\label{e:Besselcont}
\end{align}
To
study the integrals \eqref{e:BesselOZ}--\eqref{e:Besselcont}
we will make use of well-established asymptotic
properties for $I_\nu$.
To state these properties, for $\nu > 0$ and $t >0$ we define
\begin{equation}
\label{e:L_psi_def}
L_\nu(t) = \frac{1}{(2\pi\nu)^{1/2}} \frac{e^{\nu \psi(t)}}{(1+t^2)^{1/4}},
\qquad
\psi(t) =
-t+
\sqrt{1+t^2} + \log\Big( \frac{t}{1+\sqrt{1+t^2}}\Big).
\end{equation}
The identity $\log( \frac{t}{1+\sqrt{1+t^2}}) = -\arcsinh(t^{-1})$ gives a
useful alternate representation for $\psi$.
The first three derivatives of $\psi$ are:
\begin{align}
\label{e:psi_deriv}
\psi'(t)
&= -1+ \sqrt{1+t^{-2}},\\
\label{e:psi_deriv2}
\psi''(t)
&= -\frac{t^{-3}}{\sqrt{1+t^{-2}}},\\
\label{e:psi_deriv3}
\psi'''(t)
&= \frac{2t^{-6}+3t^{-4}}{(1+t^{-2})^{3/2}}.
\end{align}
The following lemma
gives asymptotic representations
of the Bessel function of large argument and large order.
The proof of the lemma is deferred to Appendix~\ref{sec:Besselpf}.
We use \eqref{e:ILasy} for the OZ regime and \eqref{e:Inunu}--\eqref{e:Lclaim}
for the continuum regime.
\begin{lemma}
\label{lem:Inunu2}
As $\nu \to \infty$,
\begin{equation}
\label{e:ILasy}
\bar I_\nu(\nu t) = L_\nu (t)(1+o(1))
\end{equation}
where the $o(1)$ is uniform in $t>0$.
Also, as $\nu \to \infty$, for any $s >0$,
\begin{align}
\label{e:Inunu}
\bar I_\nu(\nu^2 s) &\sim \frac{e^{-1/2s}}{\nu (2\pi s)^{1/2}},
\end{align}
with an error that is not uniform in $s$.
Finally, there exist $C,\delta, \nu_0 >0$ such that
\begin{align}
\label{e:Lclaim}
\bar I_{\nu}(\nu^2 s)
&\le C
\Big(
e^{ - \delta\nu} \mathbbm{1}_{2\nu s < 1}
+
\nu^{-1} s^{-1/2} e^{- \delta/s} \mathbbm{1}_{2\nu s \ge 1}
\Big)
\qquad
(\nu \ge \nu_0,\; s>0).
\end{align}
\end{lemma}
\section{Ornstein--Zernike regime: Proof of Theorem~\ref{thm:green-asy}(i)}
\label{sec:subcrit}
In this section, we prove Theorem~\ref{thm:green-asy}(i).
Let $x$ be a vector in $\bb Z^d \setminus \{0\}$, and without loss of generality assume that
$x_1 \ge x_2 \ge \cdots \ge x_d \ge 0$. We write $r$ for the number of nonzero components of $x$. Throughout this section, we consider a bounded sequence $a_n \in (0, a_{\max}]$ with $a_nn\to\infty$.
To lighten the notation, we write simply $a$ in place of $a_n$.
In particular, $a$ can be independent of $n \to \infty$, or we can have $a \to 0$ as long
as $an \to \infty$.
We start with \eqref{e:BesselOZ}, which states that
\begin{align}
C^{(q)}_{a}(nx)
&=
\frac{d^{q}n^q}{\Gamma(q)}\int_0^\infty v^{q-1}e^{-dna^2v}
(\bar I_0(nv))^{d-r}
\prod_{j=1}^{r} \bar I_{nx_j}(nv) \mathrm{d} v
.
\end{align}
With the asymptotic formula for $\bar I_\nu(\nu t)$ from \eqref{e:ILasy} together
with the definitions of $L_\nu$ and $\psi$
from \eqref{e:L_psi_def}, after some algebra this leads to
\begin{align}
\label{e:CqOZI}
C^{(q)}_{a}(nx)
&= (1+\delta_n)
\alpha_q n^{q-r/2}
\int_{0}^\infty h_{n,x}(v)
e^{-ng_{a,x}(v)}
\mathrm{d} v
,
\end{align}
where $\delta_n \to 0$ (uniformly in nonzero $x$) and
\begin{align}
\label{e:alphaqdef}
\alpha_q &= \frac{d^q}{(2\pi)^{d/2}\Gamma(q)}, \\
\label{e:hdef}
h_{n,x}(v) &= v^{q-1}
(\sqrt{2\pi}\bar I_0(nv))^{d-r}
\prod_{j=1}^{r} \frac{1}{(x_j^2+v^2)^{1/4}}, \\
\label{e:gdef}
g_{a,x}(v) &=
da^2v - \sum_{j = 1}^r x_j \psi(v/x_j)
.
\end{align}
We first solve $g_{a,x}'(v) = 0$. By definition of $g_{a,x}$ and by \eqref{e:psi_deriv},
\begin{align}
\label{e:gder}
g_{a,x}'(v) &= d(1+a^2) - \sum_{j=1}^d \sqrt{1+v^{-2}x^2_j} .
\end{align}
By the definition of $u_a(x)$ in \eqref{e:udef1}, we see that
the unique solution of the equation $g_{a,x}'(v) = 0$
is $v_a(x)=u_a^{-1}$, where for notational convenience we write $u_a(x)$ simply as $u_a$.
We will soon see that this solution is
the location of the unique minimum of $g_{a,x}$.
Since we are allowing the variable $a$ to go to zero,
which sends $v_a(x)$ to infinity, it is convenient to relocate this minimum to $1$.
We therefore rescale
the representation \eqref{e:CqOZI} via $v = y/u_a$ and obtain
\begin{align}
\label{e:CqOZII}
C^{(q)}_{a}(nx)
&= (1+\delta_n)
\alpha_q \Big(\frac{n}{u_a}\Big)^{q-d/2}
\int_0^\infty \bar h_{n,a,x}(y)
e^{-n\bar g_{a,x}(y)}\mathrm{d} y
\end{align}
with
\begin{align}
\label{e:bardefs}
\bar h_{n,a,x}(y) =
h_{n}(y/u_a)n^{(d-r)/2}u_a^{q-1-d/2}, \quad
\bar g_{a,x}(y) =
g_{a,x}(y/u_a)
.
\end{align}
The minimum of $\bar{g}_{a,x}$ is located exactly at $1$, as
is illustrated in Figure~\ref{fig:g}.
\begin{figure}
\center
\includegraphics[scale = 0.5]{g_bar.PNG}
\caption{Plot of $\bar g_{a,x}$ for different values of $a$, for $d=1$ and $x=1$.
\label{fig:g}
}
\end{figure}
Recall the norm $|x|_a$ from Definition~\ref{def:norm}.
We write $\hat x = x/|x|_a$ and $\hat u_a = u_a(\hat x)$.
The scaling relation $\lambda u_a(\lambda x) = u_a(x)$ for all $\lambda> 0$
follows from the definition of $u_a(x)$ in \eqref{e:udef1},
and implies that $u_a x_j = \hat u_a \hat x_j$ and
$\hat u_a = |x|_a u_a$.
The definitions lead to
\begin{align}
\label{e:gbarxhat}
\bar g_{a,x}(y) & = |x|_a \, \bar g_{a,\hat x}(y),
\\
\label{e:bar_h_new}
\bar h_{n,a,x}(y) & =
y^{q-1}
\Big(\sqrt{2\pi n/u_a}\bar I_0(ny/u_a) \Big)^{d-r}
\prod_{j=1}^r \frac{1}{(y^2+ \hat u_a^2 \hat x_j^2)^{1/4}}
.
\end{align}
The $a$-dependence of $\bar{g}_{a,x}''(y)$
hinders an immediate application of a standard theorem for the Laplace method
such as \cite[Theorem~7.1, p.~127]{Olve97},
so we prove Theorem~\ref{thm:green-asy}(i) by analysing the integral in \eqref{e:CqOZII} directly.
To do so, we require the detailed understanding of
the $\bar g_{a,x}$ that we present next.
As a preliminary, we note that
it follows from the definition of $u_a$
in \eqref{e:udef1} that $\hat u_a = O(a)$ as $a\to 0$ uniformly in $x \neq 0$, and moreover
that
\begin{align}
\label{e:asymptotics_for_u}
\hat u_a = \sqrt{2d} \frac{a}{\|\hat x\|_2}
(1+O(a^2 ))
\quad
\text{uniformly in } x \neq 0 .
\end{align}
\begin{lemma}
\label{lem:ga}
Let $a>0$.
The function $\bar g_{a,x}$ is convex and attains its unique minimum on $(0 ,\infty)$ at $1$,
with $\bar g_{a,x}(1)= m_a|x|_a$. Also
\begin{align}
\label{e:gbarprimeprime}
\bar g_{a,x}''(y) &
=
| x|_a \, \bar g_{a,\hat x}''(y)
=
| x|_a \hat u_a
\sum_{j=1}^d \frac{\hat x_j^2y^{-3}}{\sqrt{1+\hat u_a^2 \hat x_j^2 y^{-2}}}
,
\\
\label{e:gbarppp}
\bar g'''_{a,x}(y)
&
=
| x|_a \, \bar g_{a,\hat x}'''(y)
= -
|x|_a \hat u_a
\sum_{j=1}^d
\frac{3\hat x_j^2y^{-4}+2\hat u_a^2 \hat x_j^4y^{-6}}
{(1+\hat u_a^2 \hat x_j^2 y^{-2})^{3/2}}
.
\end{align}
In addition,
for any $\alpha \in \bb R$ and any $n \ge 1$,
\begin{align}
\label{e:g_at_infty}
\lim_{y \to \infty} y^{\alpha} e^{-n \bar g_{a,x}(y)} = 0.
\end{align}
Finally, if $\alpha<0$ then
the maximum of $y \mapsto y^{\alpha} e^{-n\bar g_{a,x}(y)}$ for $y\in (0,1]$
is uniquely attained and lies in the interval $[\frac 12,1]$ provided that
$a$ is bounded and $an$ is
sufficiently large (depending on $\alpha$ but not on nonzero $x$).
\end{lemma}
\begin{proof}
By definition, $\bar g_{a,x}(1)=g_{a,x}(1/u_a)$ and
\begin{align}
g_{a,x}(1/u_a) &=
u_a^{-1} \Big[d(a^2+1)- \sum_{j = 1}^d \sqrt{1+u_a^{2}x_j^{2}} \,\Big]
+ \sum_{j = 1}^d x_j\arcsinh(u_ax_j) \nnb
&=\sum_{j = 1}^d x_j\arcsinh(u_ax_j) = m_a|x|_a,
\end{align}
with the last equality due to the definition of the norm in \eqref{e:normdef1}.
For the derivatives we use \eqref{e:gbarxhat} together with \eqref{e:gdef} and
the expressions
for $\psi''$ and $\psi'''$ given in \eqref{e:psi_deriv2}--\eqref{e:psi_deriv3}.
In particular,
it follows from \eqref{e:gbarprimeprime} that
$\bar g_{a,x}$ is convex and therefore the unique critical point at $1$ is
the location of the unique minimum.
For \eqref{e:g_at_infty},
it suffices to consider $g_{a,x}$ since there is no claim of
uniformity in $a$.
It can be seen from the definition of $\psi$ in \eqref{e:L_psi_def} that
$\psi(t) \to 0$ as $t \to \infty$.
With the definition of $g_{a,x}$ in \eqref{e:gdef}, this implies that
$g_{a,x}(y) \sim da^2y$ as $y \to \infty$,
so \eqref{e:g_at_infty} holds for any $n \geq 1$.
Finally, and most substantially, we let $\alpha<0$ and
prove that the maximum of $y \mapsto y^{\alpha} e^{-n\bar g_{a,x}(y)}$ for $y\in [0,1]$
is uniquely attained and lies in the interval $[\frac 12,1]$, provided that $a=a_n$ is
bounded and $an$ is
sufficiently large (depending on $\alpha$ but not on nonzero $x$).
We write
\begin{equation}
y^{-|\alpha|} e^{-n\bar g_{a,x}(y)} = \exp[-n \varphi_{a,x,\alpha}(y)]
\quad \text{with} \quad
\varphi_{a,x,\alpha,n}(y) = \bar g_{a,x}(y) + \frac{|\alpha|}{n}\log y.
\end{equation}
To find a critical point of $\varphi=\varphi_{a,x,\alpha,n}(y)$ we first observe, as in \eqref{e:gder}, that
\begin{align}
\varphi'(y)
= \frac{1}{u_a}F(y/u_a)
\quad \text{with} \quad
\label{e:tilde_g_prime}
F(t) = d(1+a^2) - \sum_{j=1}^d \sqrt{1+t^{-2}x^2_j} + \frac{|\alpha|}{n}t^{-1}.
\end{align}
Note that
$F(u_a^{-1})=|\alpha|u_a/n>0$, and that $F(t) \sim -t^{-1}(\|x\|_1-|\alpha|/n)$
as $t\to 0$ so $F(t)\to -\infty$ uniformly in $n \geq |\alpha|+1$ and $x \neq 0$.
To prove that $\varphi$ has a unique critical point in $[0,1]$, it therefore suffices to
prove that $F(t)$ is increasing on $t\in [0,u_a^{-1}]$.
The derivative of $F$ is
\begin{align}
F'(t)
&= t^{-2}G(t)
\quad \text{with} \quad
G(t)=\sum_{j=1}^d \frac{t^{-1}x_j^2}{\sqrt{1+t^{-2}x_j^2}}-\frac{|\alpha|}{n}.
\end{align}
By multiplying by $t$ in the numerator and denominator of the above sum, we see that
$G$ is decreasing.
As $t \to 0$, $G(t) \to \|x\|_1- |\alpha|/n \ge 1- |\alpha|/n>0$ uniformly in $x\neq0$ and in $n \geq |\alpha|+1$. Also, since $u_ax_j=\hat u_a \hat x_j$,
\begin{align}
G(1/u_a)
&
= \sum_{j=1}^d \frac{\hat u_a\hat x_j x_j}{\sqrt{1+\hat u_a^2 \hat x_j^2}}- \frac{|\alpha|}{n}.
\end{align}
Recall \eqref{e:asymptotics_for_u} and \eqref{e:normineq}.
The square root on the right-hand side is bounded above since $a$ is bounded, so
there is a constant $c_0>0$ such that, uniformly in nonzero $x$,
\begin{align}
G(1/u_a)
&\geq c_0 \hat u_a \frac{\|x\|_2^2}{|x|_a} - \frac{|\alpha|}{n}
\ge \hat u_a\Big(\frac{c_0\|x\|_2^2}{\|x\|_1}-\frac{|\alpha|}{n \hat u_a}\Big) .
\end{align}
This proves that $G(1/u_a)>0$ for $an $ large enough (independent of $x \neq 0$).
Therefore $G(t)>0$ for all $t \in [0, u_a^{-1}]$, which
completes the proof that $F$ is increasing on $[0, u_a^{-1}]$.
As noted previously,
this proves that there exists a unique $t^*(a,n,x)\in [0,u_a^{-1}]$ such that $F(t^*) = 0$.
To conclude, we now verify that $t^*\in[(2u_a)^{-1}, u_a^{-1}]$.
It suffices to show that $F(1/(2u_a))<0$ if $an$ is large
enough (independent of $x\neq 0$).
By definition of $u_a$,
\begin{align}
F(1/(2u_a))
&= d(1+a^2)-\sum_{j=1}^d\sqrt{1+4u_a^2x_j^2} + 2u_a \frac{|\alpha|}{n} \nnb
&=-\sum_{j=1}^d \Big( \sqrt{1+4\hat u_a^2\hat x_j^2}- \sqrt{1+\hat u_a^2\hat x_j^2}\Big)
+ 2\hat u_a \frac{|\alpha|}{n|x|_a}.
\end{align}
If $a$ is bounded below away from zero then the last term on the right-hand side is as
small as desired by taking $n$ large, whereas the difference in the first term is bounded
below by a positive constant, so $F(1/(2u_a))<0$ in this case. It therefore suffices to
consider small $a$, for which we see that
\begin{align}
F(1/(2u_a))
&\le - c\sum_{j=1}^d \hat u_a^2\hat x_j^2
+ 2\hat u_a \frac{|\alpha|}{n|x|_a}
\le -c' \hat u_a^2
+ 2\hat u_a \frac{|\alpha|}{n|x|_a}
\le - \hat u_a^2 \big(c' - \frac{2|\alpha|}{\hat u_a n } \big) .
\end{align}
For $an$ sufficiently large (independent of $x\neq 0$)
we conclude that $F(1/(2u_a))<0$.
This completes the proof.
\end{proof}
Next, we establish further properties of
the functions $\bar g_{a,x}$ and $\bar h_{n,a,x}$.
Let $\varepsilon>0$ and set $A_2=[1-\varepsilon,1+\varepsilon]$.
In the following, we are interested in the limit $\varepsilon \to 0$ and we write
$o(1)$ for error terms that go to zero in this limit.
\medskip \noindent \emph{Properties of $\bar g_{a,x} =|x|_a\bar g_{a,\hat x}$.}
By Lemma~\ref{lem:ga}, $\bar g_{a,x}$ is convex and has unique minimum
$\bar g_{a,x}(1)=m_a|x|_a$. In particular, $\bar g_{a,\hat x}(1)=m_a$.
Taylor expansion of $\bar g_{a,\hat x}$ about $1$ gives
\begin{equation}
\label{e:gTay}
\bar g_{a,\hat x}(y)
=
m_a + \frac{1}{2!} \bar g_{a,\hat x}''(1) (y-1)^2
+ \frac{1}{3!}g_{a,\hat x}'''(y^*)(y-1)^3
\end{equation}
for some $y^*$ between $1$ and $y$.
We see from \eqref{e:asymptotics_for_u} and Lemma~\ref{lem:ga} that as
$\varepsilon\to 0$
we have
\begin{align}
\label{e:gA2_bis}
\bar g_{a,\hat x}''(y) &= \bar g_{a,\hat x}''(1) (1+O(\varepsilon))
,
\\
\bar g_{a,\hat x}'''(y) &= O(a)
,
\end{align}
uniformly in $y \in A_2$,
in $a\le a_{\max}$, and in $x \neq 0$.
By \eqref{e:gbarprimeprime}, $\bar g_{a,\hat x}''(1) \asymp a$
uniformly in $a\le a_{\max}$ and in $x$.
For the endpoints of $A_2$, the above implies that there exists a constant $\gamma>0$
such that, for $\varepsilon$ sufficiently small
\begin{align}
\label{e:gA3}
\bar g_{a,\hat x}(1 \pm \varepsilon)
&= \bar g_{a,\hat x}(1) + \frac{1}{2} \bar g''_{a,\hat x}(1) \varepsilon^2
+ O(a)\varepsilon^3
\ge m_a + \gamma a \varepsilon^2
,
\\
\label{e:gpA3}
\bar g'_{a,\hat x}(1 \pm \varepsilon)
&= \pm \bar g''_{a,\hat x}(1)\varepsilon_n +O(a \varepsilon^2)
= O(a \varepsilon)
,
\end{align}
uniformly in $a\le a_{\max}$ and in $x \neq 0$.
\medskip \noindent \emph{Properties of $\bar h_{n,a,x}$.}
We first prove that, as $\varepsilon\to 0$,
\begin{align}
\label{e:hA2a}
\bar h_{n,a,x}(y)
&= (1+o(1))\prod_{j=1}^d(1+\hat u_a^2 \hat x_j^2)^{-1/4}
\quad
\text{uniformly in $y \in A_2$, in $a\le a_{\max}$, and in $x\neq 0$}
.
\end{align}
When $y \in A_2$, the ratio $y/u_a = y|x|_a/\hat u_a$ is bounded away
from zero uniformly in $a\le a_{\max}$ and $x \neq 0$.
The estimate \eqref{e:hA2a} then follows from
\eqref{e:bar_h_new} and the fact (see \cite[p.~83]{Olve97}) that
\begin{equation}
\sqrt{2\pi n t} \bar I_0(nt) = 1+o(1)
\text{ uniformly in $t$ bounded away from zero.}
\end{equation}
Next, we claim that there is a $C >0$ such that
\begin{align}
\label{e:hub}
\bar h_{n,a,x}(y) &\le Cy^{q-1-d/2}
\quad
\text{ uniformly in $y> 0$, in $a\le a_{\max}$, and in $x\neq 0$}.
\end{align}
To obtain \eqref{e:hub}, we use \eqref{e:bar_h_new} and
the fact that
$\bar I_0(t) \leq O(t^{-1/2})$ which can also be seen from \cite[p.~83]{Olve97}.
Finally,
we use \eqref{e:gbarprimeprime} and \eqref{e:hA2a} to see, after some algebra, that
as $\varepsilon\to 0$ we have
\begin{align}
\label{e:hgratio}
\frac{\bar h_{n,a,x}(1)}{\sqrt{\bar g_{a,x}''(1)}}
&
=
\frac{1+o(1)}
{(|x|_a \bar g_{a,\hat x}''(1)
\prod_{j=1}^d (1+ \hat u_a^2 \hat x_j^2)^{1/2})^{1/2}}
= \frac{\kappa_a(\hat x)}{ \sqrt{|x|_a \hat u_a}}(1+o(1))
,
\end{align}
with
\begin{equation}
\label{e:kappadef}
\kappa_a(\hat x)
=
\Big( \sum_{j=1}^d \hat x_j^2 \prod_{i\neq j} (1+\hat u_a^2 \hat x_i^2)^{1/2} \Big)^{-1/2}
,
\end{equation}
and where the $o(1)$ term goes to zero as
$\varepsilon\to 0$
uniformly in $a\le a_{\max}$ and $x \neq 0$.
\medskip
\begin{proof}[Proof of Theorem~\ref{thm:green-asy}(i)]
Recall from \eqref{e:CqOZII} that
\begin{align}
\label{e:Jsufficient}
C^{(q)}_{a}(nx)
&= (1+\delta_n)
\Big(\frac{n}{u_a}\Big)^{q-d/2}
\alpha_q
\int_0^\infty \bar h_{n,a,x}(y)
e^{-n\bar g_{a,x}(y)}\mathrm{d} y
.
\end{align}
With $c_{0,q} = \sqrt{2\pi}\,\alpha_q$
as in \eqref{e:calim} (recall \eqref{e:alphaqdef}), we define
\begin{equation}
\label{e:caqx}
c_{a,q,\hat x}
=
c_{0,q}
\kappa_a(\hat x)
\Big(\frac{\hat u_a}{m_a}\Big)^{(d-1-2q)/2}
.
\end{equation}
By definition,
$c_{a,q,\hat x}$ depends on $x$ only via its direction $\hat x$. Also, as stated in \eqref{e:calim},
$\lim_{a\to 0}c_{a,q,\hat x} = c_{0,q}$ due to \eqref{e:asymptotics_for_u},
\eqref{e:kappadef}, the relation $m_a \sim \sqrt{2d}\, a$, and the fact
that $\|\hat x\|_2 = \|x\|_2/|x|_a \to 1$ by Proposition~\ref{prop:norm}.
Our goal is to prove that \eqref{e:mr} holds,
which by \eqref{e:Jsufficient} will follow if we prove that,
uniformly in $x \neq 0$ and in $a\le a_{\max}$, as $an \to \infty$ we have
\begin{align}
\alpha_q
\int_0^\infty \bar h_{n,a,x}(y)
e^{-n\bar g_{a,x}(y)}\mathrm{d} y
& \sim
\Big(\frac{u_a}{n} \Big)^{q-d/2}
c_{a,q,\hat x}
\frac{m_a^{(d-1-2q)/2}}{(n|x|_{a})^{(d+1-2q)/2} }
\;
e^{-m_an|x|_a}
\nnb
& =
c_{0,q}
\frac{1}{\sqrt{n |x|_a \hat u_a }}\kappa_a(\hat x)
e^{-m_an|x|_a}
\label{e:J2want}
\end{align}
(the equality holds by definition---recall that $u_a = \hat u_a/ |x|_a$).
We set $\varepsilon_n =(an)^{-1/4}$,
which does obey $\varepsilon_n \to 0$ as we imposed below
\eqref{e:asymptotics_for_u}, and we
divide the interval of integration in \eqref{e:J2want} into three subintervals:
\begin{equation}
A_1 = [0,1-\varepsilon_n],
\qquad
A_2= [1-\varepsilon_n,1+\varepsilon_n]
\qquad
A_3 = [1-\varepsilon_n, \infty).
\end{equation}
Then we set
\begin{equation}
J_i =
\alpha_q
\int_{A_i} \bar h_{n,a,x}(y) e^{-n\bar g_{a,x}(y)} dy \qquad (i=1,2,3).
\end{equation}
We will prove that $J_2$ gives the main contribution
to \eqref{e:J2want}, with $J_1$ and $J_3$ relatively small.
\medskip\noindent
\emph{The integral $J_2$.}
By \eqref{e:gTay}, \eqref{e:gA2_bis}, and \eqref{e:hA2a},
\begin{align}
J_2
&= (1+o(1))
\alpha_q
\bar h_{n,a,x}(1)
e^{-n m_a|x|_a}
\int_{-\varepsilon_n}^{\varepsilon_n}\exp\Big(-\frac{n}{2} \bar g_{a,x}''(1)(1+o(1))y^2 \Big)\mathrm{d} y,
\end{align}
with the $o(1)$
(as $\varepsilon_n\to 0$)
uniform in $y$, in $a\le a_{\max}$, and in $x \neq 0$.
We make the change of variables $v=y(n\bar g_{a,x}''(1))^{1/2}$ and obtain,
with $M_n=\varepsilon_n(n\bar g_{a,x}''(1))^{1/2}$,
\begin{align}
J_2
&= (1+o(1))
\alpha_q
\frac{\bar h_{n,a,x}(1)}{\sqrt{n \bar g_{a,x}''(1)}}
e^{-n m_a|x|_a}
\int_{-M_n}^{M_n}
\exp\Big(-\frac{1}{2} (1+o(1))v^2 \Big)\mathrm{d} v
.
\end{align}
By our choice of $\varepsilon_n$, and by
the fact that $\bar g_{a,x}''(1)= |x|_a \bar g_{a,\hat x}''(1) \asymp a|x|_a$ (as noted above
\eqref{e:gA3}), there exists a $c>0$ such that
\begin{equation}
M_n \geq c\varepsilon_n (n|x|_aa)^{1/2}
\ge c(na)^{1/4} \to \infty.
\end{equation}
Since $\alpha_q (2\pi)^{1/2} = c_{0,q}$, this gives
\begin{equation}
J_2
=
(1+o(1))
c_{0,q}
\frac{\bar h_{n,a,x}(1)}{\sqrt{n \bar g_{a,x}''(1)}} e^{-nm_a|x|_a}.
\end{equation}
To obtain the desired right-hand side of \eqref{e:J2want}, we
replace the ratio in the above using \eqref{e:hgratio}.
It remains to show that the contributions from
the integrals $J_1$ and $J_3$ are relatively small.
\medskip \noindent \emph{The integral $J_1$.}
To show that $J_1$ is relatively small compared to $J_2$, it suffices to prove that as $an \to \infty$
\begin{equation}
\label{e:J1sufficient}
\int_0^{1-\varepsilon_n} \bar h_{n,a,x}(y) e^{-n\bar g_{a,x}(y)} \mathrm{d} y
\le \frac{o(1)}{ \sqrt{na|x|_a}} e^{-n\bar g_{a,x}(1)}
\end{equation}
uniformly in $x\neq 0$.
Let $\alpha = q-1-d/2$.
By the upper bound
$\bar h_{n,a,x}(y) \leq Cy^{\alpha}$ of \eqref{e:hub}, the above integral is at most
\begin{equation}
C \int_0^{1-\varepsilon_n} y^{\alpha} e^{-n\bar g_{a,x}(y)} \mathrm{d} y.
\end{equation}
If $\alpha > -1$ then we simply bound the exponential by its maximum value
to obtain an upper bound proportional to
$\exp[-n\bar g_{a,x}(1-\varepsilon_n)]$. By \eqref{e:gA3}, this gives an
upper bound
(with $\gamma >0$)
\begin{equation}
e^{-nm_a|x|_a} e^{-n \gamma a\varepsilon_n^2 |x|_a}
= e^{-nm_a|x|_a} e^{-\gamma (na)^{1/2}|x|_a}
= o((na|x|_a)^{-1/2}) e^{-nm_a|x|_a} ,
\end{equation}
which is sufficient.
If instead $\alpha \le -1$, then we apply Lemma~\ref{lem:ga} to bound
$y^{\alpha} e^{-n\bar g_{a,x}(y)}$ by its maximum which is attained
on $[\frac 12, 1]$ and hence is at most
$2^{|\alpha|}\exp[-n\bar g_{a,x}(1-\varepsilon_n)]$, and this is again sufficient
to obtain \eqref{e:J1sufficient}. This proves that $J_1$ is negligible compared to $J_2$.
\medskip \noindent \emph{The integral $J_3$.}
We bound $\bar h(y)$ by $y^{\alpha}$ with $\alpha = q-1-d/2$.
If $\alpha <-1$, so that $y^\alpha$ is integrable,
then we simply
extract additional
exponential decay (compared to $J_2$) using \eqref{e:gA3} again.
Then we integrate $y^\alpha$ over $[1,\infty]$ and obtain an upper bound of the form
\begin{equation}
\label{e:J3bd}
Ce^{-\gamma (na)^{1/2}|x|_a} = o\Big( \frac{1}{\sqrt{na|x|_a}}\Big),
\end{equation}
which is sufficient for the case $\alpha<-1$.
If instead $\alpha \ge -1$ then we use integration by parts to reduce the power.
For example, if $\alpha \in (-1,0]$ then, with $t=1+\varepsilon_n$
and $f(y)=n\bar g_{a,x}(y)$,
we use the facts that
by \eqref{e:gpA3}
$f'(t) \ge c(na)^{3/4}|x|_a$
which eventually exceeds $1$, that
$f'$ is positive and increasing on $[t,\infty)$,
and that $\lim_{y\to\infty} y^\alpha e^{-f(y)}=0$
by \eqref{e:g_at_infty}, to obtain
\begin{align}
\int_{t}^{\infty} y^{\alpha}e^{-f(y)}\mathrm{d} y
&\leq \frac{1}{f'(t)}\int_{t}^{\infty} y^{\alpha}f'(y)e^{-f(y)}\mathrm{d} y
\leq t^\alpha e^{-f(t)}+ |\alpha| \int_t^\infty y^{\alpha-1}e^{-f(y)} \mathrm{d} y .
\end{align}
The term $t^\alpha e^{-f(t)}$ is bounded above by a multiple of
$e^{-f(1+\varepsilon_n)}$, which is bounded as in \eqref{e:J3bd}.
This process can be iterated to reduce the power of $y$ to below $-1$,
which we have seen to be sufficient.
\medskip \noindent
This completes the proof.
\end{proof}
\section{Continuum regime: Proof of Theorem~\ref{thm:green-asy}(ii)}
\label{sec:contlim}
In this section, we prove Theorem~\ref{thm:green-asy}(ii). The method of proof is different from the proof of Theorem~\ref{thm:green-asy}(i) and relies
instead on a dominated convergence argument which applies simultaneously
for both $s >0$ and $s = 0$.
We define the \emph{heat kernel} (for the normalised Laplacian $\frac {1}{2d}\Delta_{\bb R^d}$)
\begin{equation}
\label{e:ptx}
p_t(x) =
\Big(\frac{d}{2\pi t}\Big)^{d/2} e^{-d\|x\|_2^2 /2t}
\qquad
(x \in \bb R^d, \; t >0).
\end{equation}
Recall the definitions of $G_s(x)$ and $G_0(x)$ in \eqref{e:Gaqx-2d} and \eqref{e:G0qx}. The following representations of the continuum Green function will be useful.
For the case $a=0$ (with $d>2q$), we observe that the change of variables
$s=d\|x\|_2^2/2t$ leads to
\begin{align}
\label{e:G0int}
&
\frac{1}{\Gamma(q)}
\int_0^\infty \mathrm{d} t \, t^{q-1}
p_t(x)
=
\frac{d^q \Gamma(\frac{d-2q}{2})}{2^{q}\pi^{d/2}\Gamma(q)}
\frac{1}{\|x\|_2^{d-2q}} = G_0^{(q)}(x).
\end{align}
For $a>0$, we use
\begin{align}
\label{e:ptK}
\frac{1}{\Gamma(q)}
\int_0^\infty \mathrm{d} t \, t^{q-1} e^{-ta^2}
p_t(x)
& =
\Big(\frac d2 \Big)^{d/2} \frac{1}{\Gamma(q)}
\frac{a^{d-2q}}{2^d \pi^{d/2}}
\int_0^\infty \mathrm{d} s \frac{1}{s^{(d-2q)/2}}
e^{-s - d(a\|x\|_2)^2/2s}
=
G_a(x),
\end{align}
where we applied the formula
\begin{equation}
\label{e:Kint}
K_\alpha(z) = \frac 12 \Big(\frac z2\Big)^\alpha \int_0^\infty \frac{1}{t^{\alpha+1}}e^{-t-z^2/4t}\mathrm{d} t
\qquad
(\alpha \in \bb R,\; z>0)
\end{equation}
from \cite[8.432.6]{GR07} for the last equality.
\begin{proof}[Proof of Theorem~\ref{thm:green-asy}(ii)]
Fix $x \neq 0$; without loss of generality we may assume that $x_j \ge 0$ for all $j$.
We are interested in $a = s/n$ with some fixed $s\ge 0$.
We rewrite \eqref{e:Besselcont} as
\begin{align}
n^{d-2q} C_{{s}/n}^{(q)}(nx)
&=
\frac{1}{\Gamma(q)} \int_0^\infty \mathrm{d} t\, t^{q-1}
e^{-s^2 t} f_n(t)
\label{e:cl1}
\end{align}
where we define (with $x$-dependence suppressed in the notation)
\begin{equation}
\label{e:fnu}
f_n(t) = n^{d}
\prod_{j=1}^{d}\bar I_{n x_j} (n^2t/d)
.
\end{equation}
Our goal is to prove that \eqref{e:cl1} has limit $G_{s}^{(q)}(x)$ as $n\to\infty$.
We prove below in Lemmas~\ref{lem:flim}--\ref{lem:fbd} that
\begin{equation}
\label{e:flim}
\lim_{n\to\infty} f_n(t) = p_t(x),
\end{equation}
and that there are positive constants $C,\delta$ (depending on $x,d$) such that
\begin{equation}
\label{e:fbd0}
f_n(t) \le
C
\Big(
\mathbbm{1}_{t \le 1}
+
t^{-d/2}
\mathbbm{1}_{t > 1}
\Big)
\quad \text{uniformly in $n \ge 0$ and $t \ge 0$.}
\end{equation}
Once \eqref{e:flim}--\eqref{e:fbd0} are proved, since the upper bound in \eqref{e:fbd0}
is integrable after insertion in the integral on the right-hand side of \eqref{e:cl1}
(assuming $d>2q$ if $s=0$),
the dominated convergence theorem
can be applied.
For $s>0$, this gives
\begin{align}
\lim_{n \to\infty}
\frac{1}{\Gamma(q)} \int_0^\infty \mathrm{d} t\, t^{q-1}
e^{-ts^2} f_n(t)
&=
\frac{1}{\Gamma(q)} \int_0^\infty \mathrm{d} t\, t^{q-1} e^{-ts^2}
p_t(x),
\end{align}
and we have seen in \eqref{e:ptK} that the right-hand side is equal to $G_{s}^{(q)}(x)$.
Similarly, if $s=0$ and $d>2q$, we instead obtain
\begin{align}
\lim_{n\to\infty}
\frac{1}{\Gamma(q)} \int_0^\infty \mathrm{d} t\, t^{q-1}
f_n(t)
&=
\frac{1}{\Gamma(q)} \int_0^\infty \mathrm{d} t\, t^{q-1}
p_t(x) ,
\end{align}
which is the integral identified as $G_0(x)$ in \eqref{e:G0int}.
This completes the proof.
\end{proof}
It remains to prove \eqref{e:flim}--\eqref{e:fbd0}. We do this in Lemmas~\ref{lem:flim}--\ref{lem:fbd}, whose proofs rely on
the asymptotic form of the modified Bessel function from Lemma~\ref{lem:Inunu2}.
\begin{lemma}
\label{lem:flim}
For $d\ge 1$, for $x \in \bb Z^d$ with $x_j \ge 0$, and for $t>0$,
\begin{equation}
\label{e:flim-lem}
\lim_{n\to\infty} f_{n}(t) = p_t(x) .
\end{equation}
\end{lemma}
\begin{proof}
Recall the definition of $f_n(t)$ in \eqref{e:fnu}.
When $x_j >0$, it follows from \eqref{e:Inunu}
that
\begin{align}
\label{e:Lasy}
\bar I_{n x_j} (n^2t/d)
= \bar I_{n x_j}((n x_j)^2 t/ d x_j^2)
&\sim
\frac{1}{n} \Big(\frac{d}{2\pi t}\Big)^{1/2}e^{-dx^2_j/2t}
.
\end{align}
If $x_j =0$ then, since
$\bar I_0(z)\sim (2\pi z)^{-1/2}$ as $z \to \infty$,
\begin{align}
\label{e:I0asy}
\bar I_0 (n^2t/d)
&\sim \frac{1}{n} \Big(\frac{d}{2\pi t}\Big)^{1/2}
,
\end{align}
which is the same formula as \eqref{e:Lasy} but with $x_j$ set equal to zero.
Substitution of \eqref{e:Lasy}--\eqref{e:I0asy} into \eqref{e:fnu},
together with the definition of $p_t(x)$ in \eqref{e:ptx}, then gives
\begin{equation}
\lim_{n \to\infty} f_n(t) = p_t(x)
\end{equation}
and the proof is complete.
\end{proof}
\begin{lemma}
\label{lem:fbd}
Let $d \ge 1$ and
let $x \in \bb Z^d$ be nonzero with $x_1 \ge x_2 \ge \cdots \ge x_d \ge 0$.
There
are constants $C,\delta,n_0 >0$ (depending only on $d$)
such that
\begin{equation}
\label{e:fbd}
f_n(t) \le
C
\Big(
\mathbbm{1}_{t \le 1}
+
t^{-d/2}
\mathbbm{1}_{t > 1}
\Big)
\quad \text{uniformly in $n \ge n_0$ and $t \ge 0$.}
\end{equation}
\end{lemma}
\begin{proof}
We use $C$ to denote a constant that may depend on $d$ and may
change value from line to line.
By hypothesis, $x_1 \ge 1$.
Since $I_{\alpha'}(z) < I_\alpha(z)$ for any $z \ge 0$ and any $\alpha' > \alpha \ge 0$
\cite{Coch67}, we can bound each factor with $j \ge 2$ in \eqref{e:fnu} above by
$\bar I_{0}(n^2t/d)$ to obtain
\begin{equation}
f_n(t) \le n^{d-1}\Big(\bar I_{0}(n^2t/d)\Big)^{d-1} n \bar I_{n}(n^2t/d).
\end{equation}
Since $\bar I_0(z)\sim (2\pi z)^{-1/2}$ as $z \to \infty$,
and since $\bar I_0(z) \le 1$ for all $z \ge 0$, we see that
\begin{equation}
\label{e:fgbd}
f_n(t) \le
C \min (n^{d-1}, t^{-(d-1)/2}) \;
n \bar I_{n} (n^2t/d ).
\end{equation}
By \eqref{e:Lclaim}, there exist $C,\delta,n_0 >0$ such that
\begin{align}
\label{e:fgbd2}
n\bar I_{n}(n^2 t/d)
&\le C
\Big(
n e^{ - \delta n} \mathbbm{1}_{ t < d/2n}
+
t^{-1/2} e^{- \delta/t}
\mathbbm{1}_{t \geq d/2n} \Big)
\qquad
(n \ge n_0,\; t>0).
\end{align}
We insert \eqref{e:fgbd2} into \eqref{e:fgbd}, using $n^{d-1}$ for the first
term and $t^{-(d-1)/2}$ for the second, and obtain
\begin{equation}
f_n(t) \le C \Big(
n^d e^{ - \delta n} \mathbbm{1}_{t < d/2n}
+
t^{-d/2} e^{- \delta/t}
\Big)
\qquad
(n \ge n_0,\; t>0).
\end{equation}
The second term on the right-hand side is bounded for $t \le 1$ and is less than
$t^{-d/2}$ for $t >1$. Also, $n^d e^{ - \delta n}$ is bounded
as a function of $n$, and $\mathbbm{1}_{t<d/2n} \le \mathbbm{1}_{t \le 1}$ once
$n \ge d/2$
so
the first term is bounded by a multiple of $\mathbbm{1}_{t \le 1}$.
This completes the proof.
\end{proof}
\section{Properties of the norm: Proof of Proposition~\ref{prop:norm}}
\label{sec:norm}
In this section, we prove Proposition~\ref{prop:norm}. We assume throughout that $d \ge 1$
and $a>0$.
Recall the definition
\begin{equation}
\label{e:normdef5}
|x|_a = \frac{1}{m_a} \sum_{i=1}^d x_i \arcsinh(x_iu)
\qquad (x \neq 0).
\end{equation}
Proposition~\ref{prop:norm} asserts that
$|\cdot|_a$ defines a norm on $\bb R^d$ which is monotone increasing
in $a$ and for all $x\in\bb R^d$ obeys
\begin{equation}
|x|_a=\|x\|_2(1+O(a^2)),
\qquad \lim_{a\to \infty}|x|_a = \|x\|_1,
\end{equation}
with the error term in the first equality uniform in nonzero $x$ as $a \to 0$.
From this, we conclude immediately that $\|x\|_2 \le |x|_a \le \|x\|_1$.
For the limit $a \to 0$,
it follows from the relation $u_a(x) = \frac{\sqrt{2d}a}{\|x\|_2}(1+O(a^2))$
from \eqref{e:asymptotics_for_u}, together
with $m_a=\sqrt{2d}a(1+O(a^2))$ from \eqref{e:maasy}
and the definition \eqref{e:normdef5},
that
\begin{align}
|x|_a
&= \frac{1}{m_a}\sum_{i=1}^d x_i^2\frac{\sqrt{2d}a}{\|x\|_2}
\big(1+O(a^2)+O(a^2 x_i^2\|x\|_2^{-2})\big)
= \|x\|_2(1+O(a^2)).
\label{e:a2}
\end{align}
To see that $\lim_{a\to\infty}|x|_a = \|x\|_1$, we first observe that $m_a = \arccosh(1+da^2)\sim \log a^2$ as $a\to\infty$. Also, it follows from \eqref{e:udef1} that
$u=u_a(x) \sim \|x\|_1^{-1}da^2$ as $a \to \infty$, and therefore
\begin{equation}
|x|_a
\sim \frac{1}{\log a^2} \sum_{i=1}^d |x_i| \log a^2 = \|x\|_1.
\end{equation}
Thus, to complete the proof of Proposition~\ref{prop:norm}, it suffices to prove
that $|\cdot|_a$ defines a norm on $\bb R^d$, and that $|x|_a$ is monotone increasing in $a$
for each fixed $x$. We prove these two items in Lemmas~\ref{lem:normpf}--\ref{lem:normincreasing}.
To lighten the notation, we will write $C_a(x)$ instead of $C_a^{(1)}(x)$.
The following elementary lemma is the basis for our proof of the triangle inequality for $|\cdot|_a$.
\begin{lemma}
\label{lem:Cxy}
For $d \ge 1$, for $x,y\in \bb Z^d$ and for $a > 0$ (also for $a=0$ if $d >2$),
\begin{equation}
\label{e:CCCC}
C_a(0)C_a(x) \ge C_a(y) C_a(x-y).
\end{equation}
\end{lemma}
\begin{proof}
Let $p_n(x)$ be the $n$-step transition probability for simple random walk (without killing)
to travel from
$0$ to $x$ in $n$ steps, and
let $P_\kappa(x) = \sum_{n=0}^\infty (1-\kappa)^{n} p_n(x)$.
We have seen below \eqref{e:Gkappadef} that $(1+a^2)C_a(x)=P_\kappa(x)$ with
$\kappa = \frac{a^2}{1+a^2}$,
so it suffices to prove \eqref{e:CCCC} instead for $P_\kappa$.
Let $q_n(x)$ be the $n$-step transition probability for simple random walk to travel from
$0$ to $x$ in $n$ steps without revisiting $0$, and let
$Q_\kappa(x) = \sum_{n=0}^\infty (1-\kappa)^{n} q_n(x)$.
By considering only walks from $0$
to $x$ which pass through a fixed $y \in \bb Z^d$ and visit $y$ for the last time at
the $m^{\rm th}$ step, we obtain
\begin{equation}
p_n(x) \ge \sum_{m=0}^n p_m(y)q_{n-m}(x-y).
\end{equation}
This inequality gives
\begin{equation}
P_\kappa(x) \ge P_\kappa(y)Q_\kappa(x-y).
\end{equation}
Also, with $m$ the time of the last return to $0$,
\begin{equation}
p_n(x) = \sum_{m=0}^n p_m(0)q_{n-m}(x),
\end{equation}
and by replacing $x$ with $x-y$ we similarly obtain
\begin{equation}
P_\kappa(x-y) = P_\kappa(0)Q_\kappa(x-y).
\end{equation}
Therefore,
\begin{equation}
P_\kappa(0)P_\kappa(x) \ge P_\kappa(y)P_\kappa(x-y),
\end{equation}
and the proof is complete.
\end{proof}
\begin{lemma}
\label{lem:normpf}
For $d \ge 1$ and $a>0$, $|\cdot|_a$ is a norm on $\bb R^d$.
\end{lemma}
\begin{proof}
By its definition in \eqref{e:normdef5}, $|\cdot|_a$ is non-negative and homogeneous
(recall that $u_a(\lambda x) = |\lambda|^{-1}u_a(x)$),
with $|x|_a=0$ if and only if $x=0$. It remains only to prove the triangle inequality.
To prove the triangle inequality first for points in $\bb Z^d$, we conclude
from Lemma~\ref{lem:Cxy} that
\begin{equation}
C_a(0)C_a(nx) \ge C_a(ny)C_a(nx-ny)
\qquad
(x,y\in\bb Z^d).
\end{equation}
The asymptotic formula \eqref{e:mr}
(whose proof did not use the triangle inequality we are now proving)
implies that
\begin{equation}
- \lim_{n \to \infty} \frac 1n \log C_a(nx) = m_a |x|_a.
\end{equation}
From this, we obtain
\begin{equation}
m_a|x|_a \le m_a|y|_a + m_a |x-y|_a,
\end{equation}
and hence the triangle inequality does hold when the norm is evaluated at points in $\bb Z^d$.
For $x,y\in \bb R^d$, we write $\floor{x} = (\floor{x_1},\cdots, \floor{x_d})$. The triangle inequality holds for $\floor{2^nx},\floor{2^ny}$, for all $n\in \bb N$. By homogeneity, it also holds for $\frac{\floor{2^nx}}{2^n},\frac{\floor{2^ny}}{2^n}$. Since $x \mapsto u(x)$ is a continuous function on $\bb R^d\setminus\{0\}$, $x\mapsto|x|_a$ is a continuous function on $\bb R^d$ which is extended continuously at $0$ by $|0|_a = 0$. Thus by letting $n\to\infty$ we obtain the triangle inequality for all $x,y\in\bb R^d$.
\end{proof}
\begin{lemma}
\label{lem:normincreasing}
For $d \ge 1$, $a>0$, and $x\in \bb R^d$,
the norm $|x|_a$ is a monotone increasing function of $a$.
\end{lemma}
\begin{proof}
We fix $x\in \bb R^d$ and prove that the function $f(a)=|x|_a$
is increasing in $a$. It is convenient to introduce the notation
\begin{equation}
\sigma_i = \sigma_i(x) = x_i^2 (1+x_i^2u^2)^{-1/2},
\qquad
\|\sigma\|_1 = \sum_{i=1}^d \sigma_i.
\end{equation}
Implicit differentiation of $\cosh m_a=1+da^2$ with respect to $a$ gives
\begin{equation}
m_a'(a) = \frac{2da}{\sinh m_a},
\end{equation}
and differentiation of \eqref{e:udef1} leads to
\begin{equation}
u_a' = \frac{2da}{u \|\sigma\|_1}.
\end{equation}
Therefore, by \eqref{e:normdef5},
\begin{align}
f'(a) &=
- \frac{m_a'}{m_a^2}\sum_{i=1}^d x_i \arcsinh(x_iu)
+
\frac{u'}{m_a} \|\sigma\|_1
\label{e:derivative_a_norm}
=
\frac{2da}{um_a }\Big(1-\frac{|x|_a u }{\sinh m_a }\Big).
\end{align}
Let
\begin{equation}
U(x) = |x|_a u_a(x) \quad (x\in \bb R^d \setminus \{0\}),
\end{equation}
and note that when $x=e_i$
is a unit vector, it follows from \eqref{e:ue}--\eqref{e:norm1} that
\begin{equation}
U(e_i) = u_a(e_i) = \sinh m_a.
\end{equation}
Thus it suffices to show that $U$ is maximal at $e_i$, as this implies $f'(a) \ge 0$.
Since $U(\lambda x)=U(x)$ for all $\lambda >0$, and since
$U$ is continuously differentiable and bounded on $\bb R^d\setminus \{0\}$,
the maximum exists and
will be attained along lines through the origin. There will be several lines
since $U(x)$ is invariant under permutation or sign changes of the coordinates of $x$, so we may restrict attention to nonzero $x$ with $x_1 \geq \cdots \geq x_d$.
We first argue that any critical point $x^*$ of $U$ must have all its nonzero coordinates
equal. A critical point of $U$ obeys
\begin{align}
\label{e:xcrit}
\frac{\partial U}{\partial x_i}(x^*) &= \frac{u\arcsinh(x^*_iu)}{m_a}+|x^*|_a\frac{\partial u}{\partial x_i}(x^*) = 0 \qquad (i=1,\ldots,d).
\end{align}
Differentiation of \eqref{e:udef1} with respect to $x_i$ gives
\begin{align}
x_i \frac{\partial u}{\partial x_i} &=
-u\frac{\sigma_i}{\|\sigma\|_1}
.
\end{align}
Thus, with $\sigma_i^*=\sigma_i(x^*)$ and $u^*=u_a(x^*)$, \eqref{e:xcrit} can be rephrased as
\begin{align}
\label{e:crit_point_U_norm}
\frac{\sigma^*_i}{\|\sigma^*\|_1}
&= \frac{x^*_i \arcsinh u^* x^*_i}{m_a |x^*|_a}
\qquad (i=1,\ldots,d).
\end{align}
Let $k\ge 1$ denote the largest subscript $i$ such that $x^*_i>0$.
From \eqref{e:crit_point_U_norm}, we see that
\begin{align}
\label{e:crit_point_U_norm_bis}
\frac{\sigma_i^*u^{*2}}{u^*x^*_i\arcsinh u^* x^*_i}
&= \frac{\sigma_j^*u^{*2}}{u^*x^*_j\arcsinh u^* x^*_j}
\qquad ( i,j \le k ).
\end{align}
An elementary calculation shows that the
function
\begin{equation}
t \mapsto \frac{t}{\sqrt{1+t^2} \arcsinh t}
\end{equation}
is a bijection from $\bb R^+$ onto $[0,1]$. Thus \eqref{e:crit_point_U_norm_bis} implies the equality $u^*x^*_i = u^*x^*_j$, so indeed all nonzero coordinates of any critical vector $x^*$ must be equal.
Let $v_k = \sum_{i=1}^k e_i$, for $k=1,\ldots,d$. It remains only to determine
which value of $k$ maximises $U(v_k)$.
The explicit values of $u_a(v_k)$ and $|v_k|_a$ can be computed from \eqref{e:udef1}--\eqref{e:normdef5}, with the result that
\begin{align}
u_a(v_k) & = \sqrt{(1+da^2/k)^2-1} = \sinh(\arccosh(1+da^2/k)),
\\
|v_k| &= \frac{1}{m_a}k\arcsinh u_a(v_k) = \frac{1}{m_a}k \arccosh(1+da^2/k).
\end{align}
From this, we find that
\begin{align}
U(v_k) & = \frac{da^2}{m_a} \psi(1+da^2/k),
\qquad
\psi(x) = \Big(\frac{x+1}{x-1} \Big)^{1/2} \arccosh x.
\end{align}
A computation gives
\begin{align}
\psi'(x) &= \frac{1}{x-1}\Big(1-\frac{\arccosh x}{\sqrt{x^2-1}}\Big)
\ge 0,
\end{align}
with the inequality due to the fact that
$\arccosh x \leq \sqrt{x^2-1}$ for all $x\ge 1$.
Therefore $U(v_k)$ is decreasing in $k$ and the maximum of $U(v_k)$ is attained at $v_{1}=e_1$. We have noted previously that this suffices, so the proof is
complete.
\end{proof}
|
\section{Introduction}
\input{real_intro.tex}
\section{The standard model at low energies}
\input{intro.tex}
\section{Physical definitions of the strong coupling}
\input{physical_couplings.tex}
\section{Lattice field theory}
\label{sec:lattice-field-theory}
\input{lattice.tex}
\section{Decoupling of heavy quark and matching accross thresholds}
\label{sec:deco-heavy-quark}
\input{decoupling.tex}
\section{Convenient observables for a coupling definition}
\input{observables.tex}
\input{fss.tex}
\section{Present and future of lattice determinations of $\alpha_s$}
\label{sec:pres-future-latt}
\input{future.tex}
\section{Conclusions}
\input{conclusions.tex}
\section*{Acknowledgments}
\addcontentsline{toc}{section}{Acknowledgments}
The authors want to thank G.
Salam for interesting discussions and a critical reading of an earlier
version of this manuscript.
We also thank him for hospitality in Oxford.
AR has a large debt with the members of the ALPHA collaboration for
a fruitful collaboration in several works, spanning several years,
about many issues covered in this review.
AR also wants to show his gratitude to Rainer Sommer and Stefan Sint
for the many discussions on several topics covered in this review.
We warmly thank Peter Petreczky for sharing his data
and scripts to produce figures.~\ref{fig:hqcorr}, the authors
of~\cite{Zafeiropoulos:2019flq}, in special Jose
Rodriguez Quintero and Savvas Zafeiropoulos, for sharing and explaining
the data used to produce figure~\ref{fig:alpha_taylor} (b).
Rainer Sommer for sharing figure~\ref{fig:sign}, and Christopher Kelly
and the RBC/UKQCD collaboration for permission to reproduce
figure~\ref{fig:omega}.
LDD is supported by an STFC Consolidated Grant, ST/P0000630/1, and a
Royal Society Wolfson Research Merit Award, WM140078.
AR is partially supported by the Generalitat Valenciana
(CIDEGENT/19/040) and was partially supported by the H2020 program in
the Europlex training network, grant agreement No. 813942.
\subsection{Topology freezing and large autocorrelation times}
\begin{figure}[h!]
\centering
\begin{subfigure}{0.47\textwidth}
\includegraphics[width=\textwidth]{fig/Qtop_coarse_hist.pdf}
\caption{$32^4$ with $a=0.08$ fm.}
\end{subfigure}
\begin{subfigure}{0.47\textwidth}
\includegraphics[width=\textwidth]{fig/Qtop_fine_hist.pdf}
\caption{$48^4$ with $a=0.03$ fm.}
\end{subfigure}
\begin{subfigure}{\textwidth}
\includegraphics[width=\textwidth]{fig/Qtop_coarse.pdf}
\caption{$32^4$ with $a=0.08$ fm.}
\end{subfigure}
\begin{subfigure}{\textwidth}
\includegraphics[width=\textwidth]{fig/Qtop_fine.pdf}
\caption{$48^4$ with $a=0.03$ fm.}
\end{subfigure}
\caption{The topological charge is not quantized on the lattice
(here we use the definition based on the Gradient flow), but
in practice values for $Q$ in simulations with very fine lattice
spacing cluster very close to integer values. At the same time the
transition between different topological sectors becomes less and
less frequent in simulation time. }
\label{fig:freeze}
\end{figure}
Most definitions of the topological charge are not quantized on the
lattice. It is only
when the lattice spacing of a simulation is small enough
($a< 0.05$ fm), that the values of the topological charge measured on the
lattice cluster close to integer values. The different
topological sectors emerge as the simulation approaches the continuum
limit. As was first realized in~\cite{DelDebbio:2004xh}, at the same
time the \emph{transition} between
different topological sectors also becomes less and less
frequent at small lattice spacings (see
fig.~\ref{fig:freeze})\footnote{This is expected for the kind of
algorithms used in all lattice QCD simulations, since they are based
on a continous change of the fields (i.e. the HMC). Local update
algorithms (such as those used in the simulations of
fig.~\ref{fig:freeze}) still suffer from topology freezing at small
lattice spacings.}.
This has implications for the error estimation of observables that
couple strongly with the topological charge. Lattice measurements of
these observables are correlated for very long simulation times, since
they feel the topological sector in which they have been measured. Getting
a solid estimate of the statistical error of such observables is
really difficult since the topological charge is not well sampled.
The most clean solutions to the problem improve the sampling of
different topological sectors by changing the boundary conditions of
the lattice simulation in Euclidean time~\cite{Luscher:2011kk,
Luscher:2014kea}. In these cases the topological charge is no longer
quantized (even in the continuum), and can fluctuate. In other cases
the problem can be bypassed. One can perform the simulations at fixed
topological sector and deal with power law finite volume effects if
necessary~\cite{Brower:2003yx}. Recently
simulations on very
large physical volumes so that the finite volume effects of a fixed
topology become irrelevant have been performed
in the pure gauge theory\cite{Giusti:2018cmp}.
All in all it is pretty uncomfortable that no algorithm is able to
sample correctly the theory at fine lattice spacings. It remains an
algorithmic challenge to find an efficient algorithm for the
simulation of QCD at fine lattice spacings. In the meantime, the
accuracy of the statistical errors quoted for simulations at fine lattice
spacings without the use of some variant of open boundary conditions
has to be taken with a grain of salt.
\subsection{Signal to noise problem}
\begin{figure}[h!]
\centering
\includegraphics[width=\textwidth]{fig/sn.pdf}
\caption{The variance of a meson at large Euclidean times is
dominated by two mesons propagating. On the other hand the
variance of a nucleon
two-point functions at large Euclidean times is not dominated
by the propagation of two baryons, but by the propagation of
three meson states. }
\label{fig:sign}
\end{figure}
\begin{figure}[h!]
\includegraphics[width=\textwidth]{fig/plateaux.pdf}
\caption{Several meson and baryon correlators, as well as the
static potential at different distances (shifted vertically).
Meson correlators, do
not have a signal to noise problem,
and its value can be determined with high precision at large
Euclidean times.
(Source: $m_p$~\cite{Jager:2013kha},
$m_{\Omega}$~\cite{Capitani:2011fg},
$V(\approx r_0), V(\approx r_1)$~\cite{Fritzsch:2012wq},
$f_\pi$~\cite{Lottini:2013rfa}).}
\label{fig:plateaux}
\end{figure}
\begin{figure}[h!]
\includegraphics[width=\textwidth]{fig/omega_48I.pdf}
\caption{Precise values of baryon masses require to fit the
correlators at small Euclidean times, where there are
significant contributions from excited states.
(Source~\cite{Blum:2014tka})
}
\label{fig:omega}
\end{figure}
Every observable ($O$) in lattice QCD is computed as an avegare over
configurations generated by Monte Carlo sampling
\begin{equation}
\langle O \rangle = \frac{1}{N}\sum_{t=1}^N O[U_t] + \mathcal
O({1}/{\sqrt N})\,.
\end{equation}
Here $U_t$ represents the configuration at Monte Carlo time $t$ and
$O[U_t]$ is the value of the observable measured on such a configuration. All
observables carry an statistical uncertainty $\mathcal
O({1}/{\sqrt N})$ due to this stochastic estimation. The error in
the observable is proportional to the variance
\begin{equation}
(\delta O)^2 \propto \langle O^2 \rangle - \langle O \rangle^2\,.
\end{equation}
The signal to noise problem refers to the particular behavior of the
error in some correlators. If we examine the case of a meson
correlator, like the pion, we have
\begin{eqnarray}
\langle [\bar u\gamma_5d(x)\, \bar d\gamma_5u(0)] \rangle^2
&\simas{x_0\to \infty}& A e^{-2M_\pi x_0} + \dots \\
\langle [\bar u\gamma_5d(x)\, \bar d\gamma_5u(0)]^2 \rangle
&\simas{x_0\to \infty}& A' e^{-2M_\pi x_0} + \dots
\end{eqnarray}
The first equation is clear, since this is just the squared $\pi$
propagator. The second equation becomes clear when one realizes that
the Wick contractions in the expectation value
$\langle O^2 \rangle$ can be understood as 2 $\pi$ mesons propagating
between Euclidean times $0$ and $x_0$ (see fig.~\ref{fig:sign}). On the other
hand for the case of a baryon, like the proton $p$, the correlators
have the form ($a,b,c$ are color indices and $C$ the charge
conjugation matrix)
\begin{eqnarray}
\langle u^T_a C\gamma_5u_b d_c \epsilon^{abc}(x)\, [u^T_a C\gamma_5u_b d_c \epsilon^{abc}(0)]^\dagger \rangle^2
&\simas{x_0\to \infty}& A e^{-2M_p x_0} + \dots \\
\langle \left\{u^T_a C\gamma_5u_b d_c \epsilon^{abc}(x)\, [u^T_a C\gamma_5u_b d_c \epsilon^{abc}(0)]^\dagger\right\}^2 \rangle
&\simas{x_0\to \infty}& A' e^{-3M_\pi x_0} + \dots
\end{eqnarray}
In this case the dominant contribution at large Euclidean times for
the expectation value $\langle O^2 \rangle$ comes from three $\pi$
meson propagating between Euclidean times 0 and $x_0$ (fig.~\ref{fig:sign}).
The term with
two protons propagating also contributes to this expectation value,
but with a term $\propto e^{-2M_p x_0}$ that decays much faster.
See also~\cite{Luscher:2010ae} for more details.
In summary the ratio signal-to-noise shows a quite different behavior
in both cases (see fig.~\ref{fig:sign})
\begin{eqnarray}
\frac{C_\pi(x_0)}{\delta C_\pi(x_0)} &\simas{x_0\to\infty}& 1\,,\\
\frac{C_p(x_0)}{\delta C_p(x_0)} &\simas{x_0\to\infty}& e^{\left(M_p- \frac{3}{2}M_\pi\right)x_0 }\,.\\
\end{eqnarray}
This exponential increase in the signal to noise has in practice
important implications: it is almost impossible to determine the value
of a baryon correlator with high precision at distances of 1 fm and
larger. Precise determinations of baryon masses extract such
quantities from the values of the correlator at relatively small
Euclidean times, where there is significant contamination of excited
estates. The use of different interpolating operators (with different
excited state contamination) and the use of several numerical
techniques are common to deal with this situation (see
figs.~\ref{fig:plateaux} and~\ref{fig:omega}).
Once more, the situation is far from ideal. Such basic quantities as baryon
masses ideally should be determined without dealing with complicated
systematic effects in the determination of the effective mass
plateau, but currently there no known techniques that allow the
precise extraction of baryon masses at distances $\sim 1$ fm.
\subsection{Decoupling Theorem}
\label{sec:hist-remarks-will}
Following a Wilsonian approach to field theory, it is natural to imagine that
the low-energy dynamics is not sensitive to the details of the theory at high
energies -- the high-energy degrees of freedom are {\em integrated out}. More
precisely, when considering observables defined at some low-energy scale $\mu$,
the effects of heavy states of mass $m$ are expected to be encoded in a
redefinition of the couplings, or suppressed by powers of $\mu/m$ in the limit
where $\mu/m\ll 1$.
\newcommand{\bar{m}_\mathrm{os}}{\bar{m}_\mathrm{os}}
\newcommand{\bar{g}_\mathrm{os}}{\bar{g}_\mathrm{os}}
The original idea of decoupling of massive states dates back to the seminal
paper by Appelquist \& Carazzone~\cite{Appelquist:1974tg}, where decoupling was
proven by analysing the behaviour of Feynman diagrams containing heavy quark
loops in perturbation theory. In order to be able to discuss decoupling, the
relevant quantities need to be precisely defined. It is therefore useful to
summarise, briefly, the results in Ref.~\cite{Appelquist:1974tg}, paying
specific attention to the subtleties due to the regularization and
renormalization of the theory. Since we are ultimately interested in QCD, we can
follow the arguments in~\cite{Appelquist:1974tg} starting from a non-Abelian
gauge theory coupled to a massive fermion as defined in Eq.~\ref{eq:QCDAction}.
Working in the Landau gauge, we choose an {\em on-shell} renormalization scheme,
and denote the renormalized mass and coupling constant $\bar{m}_\mathrm{os}$ and
$\bar{g}_\mathrm{os}$ respectively. The gluon propagator, the three-gluon vertex
and the fermion propagator are given respectively by
\begin{align}
\label{eq:GluonProp}
& D_{\mu\nu}^{ab}(k) = \delta^{ab} \,
\frac{1}{k^2} \left(g_{\mu\nu}-\frac{k_\mu k_\nu}{k^2}\right)\,
d\left(\frac{k^2}{\mu^2},\frac{\bar{m}_\mathrm{os}^2}{\mu^2},\bar{g}_\mathrm{os}(\mu)\right)\, ,\\
\label{eq:ThreeGluonVert}
& i \Gamma^{abc}_{\mu\nu\sigma}(p,q,r) =
f^{abc}\, \left[
\left(p-q\right)_\mu g_{\nu\sigma} +
\left(q-r\right)_\nu g_{\sigma\mu} +
\left(r-p\right)_\sigma g_{\mu\nu}
\right] \times \nonumber \\
&\quad \times
G\left(\frac{k^2}{\mu^2},\frac{\bar{m}_\mathrm{os}^2}{\mu^2},\bar{g}_\mathrm{os}(\mu)\right)\, , \\
\label{eq:FermionProp}
& S(p) = \frac{1}{p^2-m^2} \left[
a\left(\frac{p^2}{\mu^2},\frac{\bar{m}_\mathrm{os}^2}{\mu^2},\bar{g}_\mathrm{os}(\mu)\right)
\slashed{p} +
b\left(\frac{p^2}{\mu^2},\frac{\bar{m}_\mathrm{os}^2}{\mu^2},\bar{g}_\mathrm{os}(\mu)\right) m
\right]\, ,
\end{align}
where in Eq.~\ref{eq:ThreeGluonVert}, the 1PI vertex is evaluated at a
symmetric point $p^2=q^2=r^2=k^2$. The mass counterterm is adjusted by
imposing that the fermion propagator has a pole for $p^2=\bar{m}_\mathrm{os}^2$, and the
remaining counterterms are defined by providing renormalization
conditions at the scale $\mu$, {\em e.g.}
\begin{align}
\label{eq:AFieldRenorm}
&d\left(-1,\frac{\bar{m}_\mathrm{os}^2}{\mu^2},\bar{g}_\mathrm{os}(\mu)\right) = 1\, , \\
\label{eq:gmuRenorm}
&G\left(-1,\frac{\bar{m}_\mathrm{os}^2}{\mu^2},\bar{g}_\mathrm{os}(\mu)\right) = 1\, , \\
\label{eq:PsiFieldRenorm}
&a\left(-1,\frac{\bar{m}_\mathrm{os}^2}{\mu^2},\bar{g}_\mathrm{os}(\mu)\right) = 1\, .
\end{align}
The running coupling constant at a generic scale $k^2$ is defined as
\begin{align}
\label{eq:AppRunningCoupling}
&\bar{g}_\mathrm{os}\left(\frac{k^2}{\mu^2},\frac{m^2}{\mu^2},\bar{g}_\mathrm{os}(\mu)\right) =
\bar{g}_\mathrm{os}(\mu) \,
G\left(\frac{k^2}{\mu^2},\frac{m^2}{\mu^2},\bar{g}_\mathrm{os}(\mu)\right)
\,
d\left(\frac{k^2}{\mu^2},\frac{m^2}{\mu^2},\bar{g}_\mathrm{os}(\mu)\right)^{3/2}\, ,
\end{align}
and its dependence on the scale is described by the beta function:
\begin{align}
\label{eq:FullBetaFunc}
k^2 \frac{d}{dk^2} \bar{g}_\mathrm{os} =
\beta\left(\frac{\bar{m}_\mathrm{os}^2}{-k^2},\bar{g}_\mathrm{os}\right)\, .
\end{align}
The specific value of the coupling at the scale $\mu$ that was used in the
renormalization process serves as the initial condition for integrating the beta
function. Note that, as pointed out in the original derivation in
Ref.~\cite{Appelquist:1974tg}, the beta function depends on the renormalized
mass $\bar{m}_\mathrm{os}$ through the ratio $\bar{m}_\mathrm{os}^2/k^2$ -- this is a direct consequence
of working in an on-shell scheme, where the mass of the particle enters
explicitly in the renormalization conditions and therefore in the running of the
coupling constant.
In this specific scheme, Appelquist \& Carazzone show that diagrams containing
heavy propagators are suppressed by powers of $k/\bar{m}_\mathrm{os}$ or $\mu/\bar{m}_\mathrm{os}$,
where $k$ is the typical scale of the external momenta of the diagram. The
second, important, result is that, in the decoupling limit, the beta function of
the theory in Eq.~\ref{eq:FullBetaFunc} reduces to the beta function of the
theory where the heavy particle has been integrated out, {\em i.e.} in this
particular case the beta function of pure Yang-Mills theory; the effects of the
heavy states simply show up as power corrections that interpolate between the
two theories.
\subsection{Matching Theories}
\label{sec:matching-theories}
It was noted in Refs.~\cite{Binetruy:1979hc,Binetruy:1980xn} that the decoupling
theorem does not apply in minimal subtraction (MS), since all loops contribute
to the beta function independently of the mass of the state. The same problem
exists for generic mass-independent schemes~\cite{Weinberg:1951ss}. The solution
to this problem is found by matching at low energies the theory with heavy
particles to an effective theory containing only the light degrees of freedom,
{\em i.e.} by tuning the couplings of the effective theory so that it reproduces
the field correlators of the full theory at low energies up to corrections that
are suppressed by the ratio of the scales that characterise the light and heavy
degrees of freedom respectively. Matching gauge theories across thresholds is
first discussed in Ref.~\cite{Weinberg:1980wa}, then analysed in detail in
Refs.~\cite{Wetzel:1981qg,Bernreuther:1981sg}. It has become the method of
choice to define the coupling constant at energies that span a wide range and
hence cross several mass thresholds.
Once again we find it useful to give a pedagogical review of the main
steps in the procedure. We consider a theory with $n_f$ light fermions
and only 1 heavy fermion~\footnote{The argument can be readily
extended to the case where more than one particle is integrated
out.}, whose dynamics is specified by a lagrangian $\mathcal{L}$ and
a set of bare couplings and fields:
\begin{align}
\label{eq:EqandDfields}
& g, m, m_h, \xi, \psi, A_\mu, c\, .
\end{align}
Using a familiar notation for QCD, $g$ is the gauge coupling, $m$ the
mass of the light fermions, $m_h$ the mass of the heavy fermion, and
$\xi$ is the gauge parameter. The fields $\psi, A_\mu, c$ describe the
fermions, the gauge field and the ghosts respectively.
The effective low-energy theory will be a theory defined by an effective
lagrangian $\mathcal{L}'$ which only involves the light degrees of freedom, and
a set of rescaled couplings and fields. Following Ref.~\cite{Chetyrkin:1997un},
the bare couplings and fields of the effective theory are denoted by primed
letters, and are connected to the bare couplings and fields of the full theory
through the so-called {\em decoupling constants} $\zeta_i$:
\begin{align}
\label{eq:DecConst}
&g' = \zeta_g g\, , \quad m' = \zeta_m m\, , \quad \xi'-1 =
\zeta_3\left(\xi-1\right)\, , \\
&\psi' = \sqrt{\zeta_2} \psi\, , \quad A'_\mu = \sqrt{\zeta_3}
A_\mu\, \quad c' = \sqrt{\tilde\zeta_3} c\, .
\end{align}
It is easy to argue on symmetry grounds that $\mathcal{L}'$ must have
the same form as $\mathcal{L}$, but contain only the light degrees of
freedom:
\begin{align}
\label{eq:MatchLPrime}
\mathcal{L}'\left(g,m,\xi,\psi,A_\mu,c; \zeta_i\right)
= \mathcal{L}\left(g',m',\xi',A'_\mu, c'\right)\, .
\end{align}
Higher-dimensional operators can appear in $\mathcal{L}'$, but are
suppressed by inverse powers of the heavy mass. The decoupling
constants $\zeta_i$ are determined by computing field correlators in
both theories, and matching them up to power contributions.
The matching procedure yields the relation between the renormalised coupling $\bar\alpha$ in
the two theories:
\begin{align}
\label{eq:AlphaMatching}
&\bar\alpha'(\mu) = \left(\frac{Z_g}{Z'_g} \zeta_g\right)^2
\bar\alpha(\mu) = \zeta_{Rg}^2 \bar\alpha(\mu)\, ,
\end{align}
where $Z_g$ and $Z'_g$ are respectively the renormalization constants for the
coupling in the full and in the effective theory. The decoupling constant
$\bar\zeta_{g}$ has a perturbative expansion
\begin{align}
\label{eq:ZetaGExp}
\bar\zeta_{Rg} = 1 + \sum_{\ell=1}^\infty \bar\alpha(\mu)^\ell C_\ell(x)\, ,
\end{align}
where the coefficients $C_\ell$ are functions of the logarithm of the
ratio of scales $x=\log\left(\mu^2/\bar{m}_h(\mu)^2\right)$, and
$\bar{m}_h(\mu)$ is the mass of the heavy fermion in $\overline{\rm MS}$.
It is interesting to recall here that the functional dependence of the
coefficients $C_\ell$ on $x$ is dictated by the renormalization group equations,
as discussed {\em eg.} in Ref.~\cite{Bernreuther:1981sg}. Keeping in mind that
\begin{align}
\mu^2 \frac{d}{d\mu^2} x = 1 + \gamma = 1 + \sum_{k=1}^\infty
\gamma_k \bar\alpha(\mu)^k\, ,
\end{align}
where $\gamma$ is the mass anomalous dimension, we can take the derivative of
Eq.~\ref{eq:AlphaMatching} with respect to the logarithm of $\mu^2$, and solve
the resulting equation order by order in $\bar\alpha(\mu)$. This procedure
yields a set of differential equations that the functions $C_\ell$ must satisfy,
{\em viz.}
\begin{align}
\label{eq:DiffEqCone}
&\frac{d}{dx}C_1(x) = \beta_0'-\beta_0\, ,\\
\label{eq:DiffEqCTwo}
&\frac{d}{dx}C_2(x) = 2\left(\beta_0'-\beta_0\right) C_1 +
\beta_1'-\beta_1 -\gamma_0 \frac{d}{dx}C_1\, , \\
&\dots
\end{align}
The structure of these equations implies that $C_\ell$ is a polynomial
of degree $\ell$. The coefficients of these polynomials are functions
of the coefficients $\beta_k$, $\beta_k'$ and $\gamma_k$. On top of
that, for each differential equation, an integration constant
$C_{\ell,0}$ appears, which is determined by matching a vertex
function at $\ell$ loops. The simplest example is the integration of
Eq.~\ref{eq:DiffEqCone}, which yields
\begin{align}
\label{eq:DiffEqOneIntegrate}
C_1(x) = \left(\beta_0'-\beta_0\right) x + C_{1,0}\, .
\end{align}
As discussed above, $C_1$ is a linear function of $x$, the slope of the function
is given by the difference of the first coefficients of the beta functions in
the two theories, $\beta_0$ and $\beta_0'$, while the integration constant
$C_{1,0}$ needs to be computed from a one-loop matching. Recent matching
calculations up to four loops are
available~\cite{Schroder:2005hy,Chetyrkin:2005ia}.
The final result for the matching of couplings across mass thresholds
is reported in the PDG~\cite{Patrignani:2016xqp}; using the PDG
notation and for the particular case $\mu=m_{\rm h}^\star $, we have
\begin{equation}
\label{eq:MassThreshMatching}
\alpha_s^{(n_f+1)}(\mu^2) = \alpha_s^{(n_f)}(\mu^2)
\left(
1 + \sum_{\ell=2}^\infty c_\ell \left[\alpha_s^{(n_f)}(\mu^2)\right]^\ell
\right)\, ,
\end{equation}
here $\alpha_s^{(n_f)}$ is the coupling in the $\overline{\rm MS}$\ scheme with $n_f$ massless
fermions, and $m_{\rm h}^\star$ is the mass of the heavy fermion, also defined in the
$\overline{\rm MS}$\ scheme, at the energy scale given by the mass itself. Note that for this
particular choice of scales the $\mathcal O(\alpha^2)$ term in the relation
between couplings vanish. The coefficients $c_{n}$ for $n\leq 4$ and the
physically relevant cases $3\to 4$ and $4\to 5$ are available in
table~\ref{tab:cdec}. We also quote the effect that the last term of the series
$c_4\left[\alpha_s^{(n_f)}(\mu^2)\right]^4$ has in the determination of the strong
coupling. As the reader can see, the truncation of the perturbative series in
the decoupling relations has a completely negligible effect on the extraction of
$\alpha_s$ ($\lesssim 0.2\%$ for the case of the charm quark). Note that this
analysis does not exclude potentially large non-perturbative corrections in the
matching between theories (see next section).
\begin{table}
\centering
\begin{tabular}{lllllc}
\toprule
& $m_{\rm h}$ [GeV]& $c_2\times 10^2$ &$c_3\times 10^2$ &$c_4\times 10^2$ & $\delta\alpha_{\overline{\rm MS} }(M_Z)$ [\%] \\
\midrule
$3\to 4$ & $m_{\rm c}\approx 1.3$& -1.547963 & -2.315990 & -1.922535 & 0.18\% \\
$4\to 5$ &$m_{\rm b}\approx 4.2$ & -1.547963 & -2.042976 & -0.7278004 & 0.02\% \\
\bottomrule
\end{tabular}
\caption{Coefficients for the decoupling of the charm quark ($3\to
4$) and the bottom quark ($4\to 5$) for the case $\mu=m_{\rm h}$
(see
Eq.~(\ref{eq:MassThreshMatching}))~\cite{Chetyrkin:2005ia,Schroder:2005hy,Herren:2017osy}.
The last column quotes the effect of the last known term in the
series in the value of the strong coupling at the scale $M_Z$. }
\label{tab:cdec}
\end{table}
\subsection{Nonperturbative decoupling}
\label{sec:nonp-deco}
As noted before in Sect.~\ref{eq:QCDScale}, the running of the
coupling constant with the energy scale is determined by the knowledge
of the beta function and the $\Lambda$-parameter of a given theory. Hence
the matching of the coupling between theories described in the
sections above can be reformulated in terms of the matching of the
$\Lambda$-parameters, which leads naturally to a framework where
decoupling can be discussed beyond perturbation theory. We present
here a brief summary of the ideas that were originally developed in
Ref.~\cite{Bruno:2014ufa}.
Following the notation introduced above, we denote quantities in the
theory with heavy particles with unprimed variables, while primed
variables always refer to the effective theory that includes the light
degrees of freedom only. The relation between the $\Lambda$-parameters
is fixed by requiring that low-energy quantities are matched up to
power corrections; without loss of generality we can write
\begin{align}
\Lambda' = f\left(\Lambda, M\right)\, ,
\end{align}
where $M$ is the RGI mass of the heavy particle in the full
theory. Furthermore, we can argue on dimensional grounds that
\begin{align}
\label{eq:LamRatioMassThres}
\Lambda'/\Lambda = P\left(M/\Lambda\right)\, .
\end{align}
The dependence of $P$ on the mass $M$ is encoded in
\begin{align}
\label{eq:DefEtaFun}
\eta(M) = \left. \frac{1}{P} M \frac{\partial}{\partial M} P
\right|_\Lambda\, ,
\end{align}
which can be reliably evaluated in perturbation theory only for large
values of $M$
\begin{align}
\label{eq:EtaPertTh}
\eta(M) \simas{M\to\infty} \eta_0 + \eta_1 \bar{g}^2(M) + \ldots \, .
\end{align}
Introducing the variable $\tau=\log\left(M/\Lambda\right)$, and
integrating Eq.~\ref{eq:EtaPertTh}, yields
\begin{align}
P\left(M/\Lambda\right) =
\frac{1}{k}\, \exp\left(\eta_0\tau\right)\,
\tau^{\eta_1/(2b_0)} \times
\left(1 + O\left(\frac{\log\tau}{\tau}\right)\right)\, ,
\end{align}
where $k$ is a constant that can be computed given the conventions for
$\Lambda$ and $M$.
A nonperturbative matching condition requires that some hadronic scale
remains the same in the two theories, {\em ie.}
\begin{align}
\label{eq:NPMatch}
m_\mathrm{had}'\left(\Lambda'\right) =
m_\mathrm{had}\left(\Lambda, M\right) + O\left(\frac{\Lambda^2}{M^2}\right)\, .
\end{align}
In the equation above $\Lambda$ and $M$ are dimensionful quantities that {\em
define} the coupling and the masses in the full theory, and are given. The
matching condition then determines $\Lambda'$. As a consequence we find:
\begin{align}
\frac{m_\mathrm{had}\left(\Lambda, M\right)}{\Lambda} =
\frac{m_\mathrm{had}'\left(\Lambda'\right)}{\Lambda'}
\frac{\Lambda'}{\Lambda}
+ O\left(\frac{\Lambda^2}{M^2}\right)\, ,
\end{align}
and hence
\begin{align}
\label{eq:MassRatio}
\frac{m_\mathrm{had}\left(\Lambda,
M\right)}{m_\mathrm{had}\left(\Lambda, 0\right)} =
Q_\mathrm{had} \, P\left(M/\Lambda\right)
+ O\left(\frac{\Lambda^2}{M^2}\right)\, ,
\end{align}
where
\begin{align}
Q_\mathrm{had} = \frac{m_\mathrm{had}'\left(\Lambda'\right)}{\Lambda'}
\frac{\Lambda}{m_\mathrm{had}\left(\Lambda, 0\right)} \, .
\end{align}
The left-hand side of Eq.~\ref{eq:MassRatio} can be measured by
performing MC simulations of the full theory with different values of
the mass $M$, while $Q_\mathrm{had}$ requires an extra simulation in
the effective theory.
References~\cite{Bruno:2014ufa, Athenodorou:2018wpk} study the effect
of heavy quarks along the lines described above.
Making a long story short, they estimate the non-perturbative effects
in the matching between theories with the conclusions that for the
most important case of the charm quark, they affect the extraction of
$\alpha_s$ below the $0.4\%$ level. The interested reader is invited to
consult the original works.
\subsection{Finite size scaling}
\label{sec:finite-size-scaling}
Here we will describe a theoretical idea to overcome the fundamental
limitation of any of the previous lattice determination of the strong
coupling~\footnote{The ideas of finite size scaling are able to deal
with many other multiscale problems, like the description of heavy
quarks in lattice simulations, see Ref.~\cite{Sommer:2010ic} for
example, or the renormalization of composite operators.}, namely the
compromise between reaching large energy scales to control
perturbative truncation effects, and and having a clear separation
between these energy scales and the lattice cutoff so that cutoff
effects in observables measured at small distances are kept under
control. This fundamental compromise present in any single lattice
computation -- the scales $\mu$ that can be probed have to
obey $1/a\gg \mu \gg m_\pi$ -- originates from the fact that we want to
accommodate in a single lattice computation the scales used to match
with perturbation theory $\mu_{\rm PT}$ and the scales used to
describe hadronic physics. Finite size scaling solves this problem by
adopting a different strategy. Only a finite range of scales is
resolved in any single lattice computation, and a recursive procedure
allows to relate the energy scales explored with different
simulations.
This idea is implemented by using a \emph{finite volume renormalization
scheme}~\cite{Luscher:1991wu}. The observable used to define the coupling
$O(\mu)$ is defined at a scale linked with the finite volume of the simulation
$\mu \propto 1/L$ and the coupling will run with the size of the system. Very
small physical volumes -- the so-called \emph{femtouniverse} -- are simulated in
order to reach high energy scales. The only constraint for these simulations is
that the energy scale explored has to be far away from the cutoff ($1/a\gg \mu
\sim 1/L$). This is easily achieved by using lattices of moderate size $L/a\sim
10-50$.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{fig/fss_scheme.pdf}
\caption{ A product of dimensionless ratios together with a single
dimensionful experimental input is used to determine $\Lambda$.
At low energies an experimental quantity ($M_\Omega$ in
the example) is computed in terms of $\mu_{\rm had}$, a hadronic
scale defined in a
massless finite volume scheme (see section~\ref{sec:match-an-exper}).
The finite volume scheme is used to connect non perturbatively the
hadronic ($\mu_{\rm had}$) and perturbative ($\mu_{\rm PT}$) regimes
of QCD by determining the step
scaling function (see section~\ref{sec:solving-rg-equations},
Eq.~(\ref{eq:ston}) and Eq.~(\ref{eq:ratio_beta})).
Finally using perturbation theory one can determine the $\Lambda$
parameter in units of the high energy perturbative scale ($\mu_{\rm
PT}$).
}
\label{fig:fss_scheme}
\end{figure}
There are three key ingredients in any finite size scaling study:
\begin{itemize}
\item We need a way to relate our finite volume simulations
with experimental data.
\item We need a way to match simulations in
different volumes.
\item We need to match our results in
very small volumes (high energies) with a perturbative
computation.
\end{itemize}
Figure~\ref{fig:fss_scheme} shows a schematic representation of the procedure.
Note that we are effectively separating the problem of solving
non-perturbatively the RG equations and determining an overall hadronic scale.
But before explaining in detail these three ingredients, let us discuss some
technical details of QCD in small physical volumes.
\subsubsection{QCD in a finite volume simulations}
Typical lattice simulations aim to explore QCD in the infinite volume
limit. One is usually interested in the determination of hadronic
quantities like the spectrum or some decay form factors, which can be
reliably computed only using large volume simulations, with say
$L m_\pi > 4$. In these simulations the choice of boundary conditions
is not very relevant, since we are supposed to describe the infinite
volume physics within a percent accuracy.
On the other hand, when studying QCD on a finite volume we are exploring a
completely unphysical regime of the theory with then aim of solving
the RG equations. The size of the system is just
part of the renormalization prescription, and we are free to choose it at will.
But in this particular situation the boundary conditions play a fundamental
role. Different choices of boundary conditions have to be understood
as different renormalization schemes, and not as small corrections. The
choice of boundary conditions is part of the definition of the observable
itself.
One would naturally think that periodic boundary conditions are a natural
choice, but it was soon found~\cite{GonzalezArroyo:1981vw} that QCD observables
computed in a finite periodic box have complicated perturbative expansions.
Perturbative expansions derive from a saddle point expansion of the path
integral
\begin{equation}
\label{eq:path_int}
\mathcal Z = \int\mathcal D A_\mu\, e^{-S[A_\mu]}
\end{equation}
around the minimum. In infinite volume this minimum ($A_\mu=0$) is
unique, up to gauge transformations. On the other hand fields with
periodic boundary conditions on a finite volume do not have a unique
minimum of the action. All zero momentum fields
$A_\mu = \text{constant}$ have zero action. Note that these
configurations, in a finite volume, are not all related to each other
by gauge transformations. Gauge transformations
\begin{equation}
\Omega(x) = \exp\left(\omega^a(x) T^a\right)
\end{equation}
have to be single valued functions, imposing the condition
$\omega^a(x+2\pi) = \omega^a(x)$. It is easy to see that this implies
that gauge transformation can only shift the zero mode of the gauge
fields by a multiple of $2\pi$,
\begin{equation}
A_\mu^a \to A_\mu^a + 2\pi n\,.
\end{equation}
Naively these zero momentum field configurations produce flat
directions in the path integral Eq.~(\ref{eq:path_int}), making the
integral divergent. Expectation values are finite, but in
general the perturbative expansion can no longer be written as a power
series in $\alpha$. Fractional powers (like $\alpha^{3/2}$) or even
logarithmic contributions ($\alpha\log\alpha$) appear in the
perturbative expansion\footnote{The concrete form of these terms
depends on the number of periodic directions and the rank of the
group.}. These non-analytic terms in the perturbative expansion of
observables defined in a periodic box result in coupling definitions
(defined via Eq.~(\ref{eq:O2gsq})) that generically do not even share
the universal coefficients of the $\beta$-function. The
$\Lambda$-parameter can not be defined in schemes defined from these
observables (readers interested in this topic can consult the
review~\cite{vanBaal:1988qm}).
Of course one could still match these observables with the perturbative
expansion, including the non-analytic terms\footnote{The matching with the
asymptotic perturbative behavior for these kind of observables might be
delicate, and require access to substantially larger energy scales.}, but
fortunately there are better solutions. Since these non-analytic perturbative
expansions are just a consequence of our choice of boundary conditions, we can
choose the latter wisely in order to avoid these complications, and there exists
several options to accomplish this goal.
\begin{description}
\item[Twisted boundary conditions] Demanding physical quantities to be periodic
does not necessarily require a periodic gauge field $A_\mu(x)$. It is enough
for $A_\mu(x)$ to be
\emph{periodic modulo a gauge transformation}~\cite{tHooft:1981sz}
\begin{equation}
A_\mu(x+L_\mu\hat \mu) = \Omega_\mu(x)A_\mu\Omega^\dagger_\mu(x) +
\imath \Omega_\mu(x)\partial_\mu\Omega^\dagger_\mu(x)\,.
\end{equation}
The matrices $\Omega_\mu(x)$, known as transition matrices, can be chosen in
order to guarantee that the action has a unique minimum up to gauge
transformations~\cite{GonzalezArroyo:1981vw}.
In some sense, twisted boundary conditions are the most natural choice, since
translational invariance is fully preserved. Fermions can be added without
major conceptual problems. However, for fermions in the fundamental
representation of the gauge group, boundary conditions impose that the number
of flavors to be included in the model have to be a multiple of the rank of
the group ({\em i.e.}\ $N_{\rm f} \propto 3$ for the case of $SU(3)$). For QCD
applications this might be a problem, since only the three and six flavor
theories can be formulated with this choice of boundary conditions.
The first applications of twisted boundary conditions in finite size scaling
used a ratio of Polyakov loops as definition of the renormalized
coupling~\cite{deDivitiis:1994yz}. This coupling definition suffers from large
statistical errors, due to the large variance of this particular observable.
Most recent works use the gradient flow, which we introduced in
section~\ref{sec:theory-scales}, to define the renormalized
coupling~\cite{Ramos:2014kla}.
\item[Schr\"odinger Functional boundary conditions] The most common choice of
boundary conditions to study QCD on a finite volume are called Schr\"odinger
Functional boundary conditions~\cite{Luscher:1992an, Sint:1993un}. Dirichlet
boundary conditions are imposed on the spatial components of the gauge field
at Euclidean times $x_0=0,T$
\begin{equation}
\label{eq:sfbcs}
A_i(x)\Big|_{x_0=0} = C_i(\mathbf x);\qquad
A_i(x)\Big|_{x_0=0} = C'_i(\mathbf x)\,.
\end{equation}
The time component of the gauge field inherits its boundary
conditions from the gauge fixing condition. If the boundary fields
$C_i(\mathbf x), C'_i(\mathbf x)$ are chosen appropriately, the
minimum of the action is unique up to gauge transformations. There
are two common choices in the literature. First one can choose
$C_i(\mathbf x), C'_i(\mathbf x)$ to be constant diagonal
matrices. For the case of $SU(3)$, which has two diagonal generators,
the generic form of the the background field is
\begin{eqnarray}
\label{eq:CCp}
C_k &=&
\frac{i}{L}{\rm diag}
\bigg\{\eta-{\pi\over 3},\,
\eta\Big(\nu-{1\over2}\Big),\,
-\eta\Big(\nu+{1\over 2}\Big)+{\pi\over 3}\bigg\}\,,\\
C_k' &=& \frac{i}{L}{\rm diag}\bigg\{
-\eta-\pi,\,
\eta\Big(\nu+{1\over 2}\Big)+{\pi\over 3},\,
-\eta\Big(\nu-{1\over 2}\Big)+{2\pi\over 3}\bigg\}\,,
\end{eqnarray}
where the parameters $\eta,\nu$ can be chosen at will. An important
advantage of this setup is that derivatives of the effective action
with respect to these boundary parameters can be used to define a
renormalized coupling at a scale given by the volume of the system:
\begin{equation}
\label{eq:alphasf}
\frac{12\pi}{\bar g^2_{{\rm SF},\nu}(\mu)} =
\left\langle \frac{\partial S}{\partial
\eta}\right\rangle\bigg|_{\eta=0},
\qquad
(\mu= 1/L)\,.
\end{equation}
Different choices of $\nu$ represent different renormalization schemes
(different coupling definitions). Conveniently, the values of $\bar g_{{\rm
SF},\nu}$ for all values of $\nu$ can be calculated from expectation values
evaluated at $\nu=\eta = 0$. These couplings have been the
preferred choice for finite size scaling studies in QCD
(see~\cite{Capitani:1998mq,DellaMorte:2004bc,Aoki:2009tf,Tekin:2010mm}), but
they are gradually being replaced by the new coupling definitions based on the
gradient flow~\cite{Fritzsch:2013je}. These are more conveniently defined with
zero background field (i.e. $C_i(\mathbf x) = C'_i(\mathbf x) = 0$ in
eq.~(\ref{eq:CCp})).
Schr\"odinger functional boundary conditions can be easily simulated on the
lattice, but breaking translational invariance has the unpleseant effect of
producing linear $\mathcal O(a)$ cutoff effects (even in the pure gauge
theory~\cite{Luscher:1992an}). These can in principle be removed by tuning
boundary counterterms. They are only known in perturbation theory and the
effect of the higher order corrections have to be studied in detail in any
step scaling study.
\item[Open-SF boundary conditions] For a long time, topology freezing was
thought not be an issue in small volume simulations since non-trivial
topological sectors are highly suppressed on small volumes. But nowadays it is
clear~\cite{Fritzsch:2013yxa} that these simulations are also suffering from
this effect when the physical volume is $\sim 1\,{\rm fm}$ and either
twisted or SF bondary conditions are used~\footnote{There exists an index
theorem with twisted boundary conditions that guarantees that \emph{when
simulating massless quarks} only topologically trivial sectors contribute to
the path integral. This opens the door for three flavor QCD to actually
determine the running coupling without having to worry about any topology
freezing problems.\label{foot:twisted}}. A way to overcome this issue is just
to use definitions in the sector of zero topological charge, as suggested in
Ref.~\cite{Fritzsch:2013yxa}.
Using open boundary conditions in Euclidean time provides a solution to the
problem: the topological charge is no longer quantized and transitions between
different topological sectors are allowed~\cite{Luscher:2014kea}. As discussed previously, the breaking of translational invariance leads also in this case to linear $\mathcal O(a)$ cutoff effects near the Euclidean time boundaries.
\end{description}
Whatever the choice of boundary conditions, simulations on small physical
volumes have one more crucial ingredient that makes them attractive: the
possibility to directly simulate massless quarks on the lattice. This is
possible because in this regime the size of the system acts as IR cutoff ($\sim
1/L$). The Dirac operator has a gap, even at $m=0$. This possibility removes a
source of systematic effect in the lattice determinations of the strong
coupling, since the matching with the perturbative regime is usually only known
to high accuracy in massless renormalization schemes. There is no need to
perform an extrapolation to zero mass.
\subsubsection{Solving the RG equations}
\label{sec:solving-rg-equations}
Finite volume renormalization schemes work by providing coupling
definitions that depend on a single scale given by the volume of the
system ($\mu\sim 1/L$). It is clear that large energy scales can be
easily reached simulating a small physical volume, requiring only a
modest lattice size ($L/a\sim 10-30$). What is not so clear is how to
relate different finite volume simulations in order to determine the
actual running of the strong coupling.
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{fig/step.pdf}
\caption{In finite volume renormalization schemes the coupling $\bar
g ^2(\mu)$ is defined at a scale given by the physical size of
the system ($\mu = 1/L$). If at some scale $\mu_1 = 1/L_1$ the coupling has
value $u = g ^2(\mu_1)\Big|_{\mu_1 = 1/L_1}$, and we measure the
coupling in a volume two times larger ($2L_1$), we obtain the step
scaling function $\sigma_2(u)$ (see Eq.~(\ref{eq:ssf})). This is a
discrete version of the $\beta$-function that measures how much
the coupling changes when the scale is changed by a factor 2.}
\label{fig:step}
\end{figure}
The \emph{step scaling function} $\sigma_s(u)$ is the key quantity that makes it
possible solving the RG equations (see figure~\ref{fig:step}). It measures the
change in the coupling when the renormalization scale changes by a factor $s$,
and therefore can be understood as a discrete version of the $\beta$-function,
\begin{equation}
\label{eq:ssf}
\sigma_s(u) = \bar g ^2(\mu)\Big|_{\bar g ^2(\mu/s) = u}\,,\qquad
(\mu = 1/L)\,.
\end{equation}
The values $s=2,3/2$ are typical in the literature. The advantage of the step
scaling function is that it can be easily computed on the lattice. We have to
recall that there is a one-to-one relationship between the bare parameters used
in a lattice simulation (i.e. $g_0^2,am_0$) and the lattice spacing $a$. In
massless schemes we simulate with $\bar m=0$, and the lattice spacing $a$ is set
solely by the value of the bare coupling $g_0^2$. This means that if we keep the
same bare simulation parameters and just multiply the number of lattice points
by a factor $s$, we will obtain a lattice estimate of the step scaling
function
\begin{equation}
\label{eq:ssf_comp}
L/a \to sL/a \Longrightarrow u=\bar g^2 \to \Sigma_s(u,a/L)\, .
\end{equation}
Of course the previous process still depends on the value of the cutoff $a$, but
this result can be repeated with several pairs of lattices in order to obtain a
continuum result
\begin{equation}
\label{eq:sigma_cont}
\lim_{a\to 0} \Sigma_s(u,a/L) = \sigma_s(u)\,.
\end{equation}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{fig/1607.06423/sigma.pdf}
\caption{The bare coupling $g_0^2$ is fixed on several lattice sizes
$L/a=8,10,12,16$ so that the renormalized coupling is equal to $\bar g ^2 =
u_i$ for nine values $u_i \approx 2.12, 2.39, 2.73, 3.20, 3.86, 4.49, 5.30,
5.86, 6.54$. This means that all simulations have the same physical volume
$L$, up to scaling violations. By computing the coupling on lattices twice
as large, one determines a lattice approximation $\Sigma_s(u,a/L)$ of the
step scaling function $\sigma_2(u)$. This plot shows the continuum
extrapolation of the lattice step scaling function $\Sigma_s(u,a/L)$ (see
Eq.~(\ref{eq:sigma_cont})). The continuum values can be used to parameterize
$\sigma_2(u)$ for $u\in[2.12,6.54]$, or to determine the $\beta$-function.
(Source~\cite{DallaBrida:2016kgh}).}
\label{fig:ssf}
\end{figure}
By repeating a series of continuum extrapolations at different values
of the coupling $u$, the function $\sigma_s(u)$ can be determined.
Once the step scaling function is known, the RG problem is solved
non-perturbatively. Note that the relation
\begin{equation}
\label{eq:basic_ssf}
\log s = -\int_{\sqrt{u}}^{\sqrt{\sigma_s(u)}}\, \frac{{\rm d} x}{\beta(x)}\,,
\end{equation}
is exact. The determination of the $\Lambda$-parameter
Eq.~(\ref{eq:lam}) uses the previous relation to break the fundamental
integral of the $\beta$-function
\begin{equation}
\label{eq:basic_int}
\int_{0}^{\bar g(\mu_{\rm had})}\, \frac{{\rm d}x}{\beta(x)}\,,
\end{equation}
into several pieces. The scale $\mu_{\rm had}$ is a low-energy scale, of the
same order as the reference scale used to ``set the scale'', as discussed in
section~\ref{sec:scale-sett-latt}. With the help of the step-scaling function
one can produce a series of couplings $\bar g_k$ such that
\begin{equation}
\bar g_0^2 \equiv \bar g^2 (\mu_{\rm PT});\qquad
\bar g_{k-1}^2 = \sigma_s(\bar g^2 _{k})\,.
\end{equation}
Note that since the step scaling function changes the scale by a factor $s$, we
have $\bar g^2 _k = \bar g^2 (s^{k}\mu_{\rm had})$. Recalling the basic relation
Eq.~(\ref{eq:basic_ssf}) the integral of Eq.~(\ref{eq:basic_int}) can now be
written as
\begin{equation}
\label{eq:split_int}
\int_{0}^{\bar g(\mu_{\rm had})}\, \frac{{\rm d}x}{\beta(x)} =
\int_{0}^{\bar g(s^n\mu_{\rm had})}\, \frac{{\rm d}x}{\beta(x)} +
n\log s\,.
\end{equation}
With moderate values of $n$ ({\em i.e.} $\, n\sim 10$) one can reach high energy scales, since
$s^n\mu_{\rm had}\sim 100$ GeV. The remaining integral in
Eq.~(\ref{eq:split_int}) can be very well approximated by using perturbation
theory (cf.
section~\ref{sec:syst-extr-alph}).
\begin{equation}
\int_{0}^{\bar g(s^n\mu_{\rm had})}\, \frac{{\rm d}x}{\beta(x)}
\simas{n \to \infty}
\int_{0}^{\bar g(s^n\mu_{\rm had})}\, \frac{{\rm d}x}{\beta_{\rm
PT}(x)} + \dots\,.
\end{equation}
The scale $s^n\mu_{\rm had}$ is the scale at which the result is
matched with perturbation theory
\begin{equation}
\label{eq:ston}
\mu_{\rm PT} = s^n\mu_{\rm had}\,.
\end{equation}
One of the big advantages is that
finite size scaling allows to vary this scale substantially: it can be
pushed to very large values with a moderate computational effort.
The $\Lambda$-parameter can be determined by matching with
perturbation theory at different energy scales. We have already seen
an example of such analysis in figure~\ref{fig:msbar}.
\bigskip
\noindent
{\em Direct determination of the $\beta$-function}
\medskip
Recent works use the step scaling function to directly determine the
$\beta$-function. This approach has some technical advantages like making the
determination of arbitrary ratios of scales $\mu_1/\mu_2$ possible, or allowing
different scales factors $s$ to be used to fix the same set of coefficients.
Several parametrizations are possible. At high energies the most natural one is
just to write
\begin{equation}
\beta(x) = -x^3\sum_{n=0}^N b_nx^{2n}\,,
\end{equation}
with the first few coefficients fixed by the perturbative
prediction. In this case, once the $\beta$-function is known, one can
use
\begin{equation}
\label{eq:ratio_beta}
\frac{\mu_{\rm PT}}{\mu_{\rm had}} = \exp \left\{ -\int_{g(\mu_{\rm PT})}^{g(\mu_{\rm had})}\, \frac{{\rm d} x}{\beta(x)} \right\}
\end{equation}
to connect the perturbative and hadronic scales similarly to
Eq.~(\ref{eq:ston}).
Another advantage is that it allows a direct
comparison of the data with perturbation theory by comparing the
non-perturbatively determined $\beta$-function and its perturbative
expansion (see~\cite{Brida:2016flw, DallaBrida:2016kgh} for some
examples).
\subsubsection{Matching to an experimental quantity}
\label{sec:match-an-exper}
In the previous section we have seen how finite size scaling techniques are able
to solve non-perturbatively the RG equations: ratios of scales defined
in a massless scheme, like
$\mu_{\rm PT}/\mu_{\rm had}$ can be determined precisely. But what we
really need are the values of these scales in physical units.
Following the discussion of section~\ref{sec:scale-sett-latt}, we need
to determine $\mu_{\rm had}/M_{\rm ref}$, where $M_{\rm ref}$ is a
reference scale (for example the $\Omega$ mass) determined in large
volumes and with physical values of the quark masses ({\em i.e.}\ in a setup
that can be matched with an experimental input).
In detail, the procedure is as follows: first one fixes the value of
the coupling in a given massless finite volume renormalization scheme
to some particular value
\begin{equation}
\bar g ^2(\mu_{\rm had}) = \text{fixed}\,,
\end{equation}
for several values of $L/a$. Since the coupling depends only
on one scale $\mu_{\rm had}\propto 1/L_{\rm had}$, the physical volume
of all these simulations is the same, up to scaling violations
(i.e. the different values of $L/a\propto 1/(a\mu)$ are really
different values of $a$ at the same $L=L_{\rm had}$).
Second, one determines in large volume simulations the value of some low-energy
scale \emph{at the same values of the bare coupling $g_0$} and for physical
values of the quark masses. This low energy scale is typically whatever
reference quantity is used to set the scale. For example let us assume that it
is the mass of the $\Omega$ baryon $M_{\Omega}$. The large volume lattice
simulations yield values of the mass in lattice units, we denote this
dimensionless quantity $\hat{M}_{\Omega}(a)$ where we have written explicitly its
dependence on the lattice spacing $a$. Since the bare coupling has been kept the
same, the values of $a$ in our determination of $a\mu_{\rm had} = a/L_{\rm had}$
are the same, up to scaling violations, as the values of $a$ in our
determinations of $\hat{M}_{\Omega}(a)$. This allows to determine the ratio
\begin{equation}
\frac{\mu_{\rm had}}{M_{\Omega}} = \lim_{a\to 0} \frac{a\mu_{\rm
had}}{\hat{M}_{\Omega}(a)}\,.
\end{equation}
Note that the fact that different values of the quark masses or physical volumes
are used for the determination of $a\mu_{\rm had}$ and $aM_{\Omega}$ is not an
issue once the continuum extrapolation is performed~\cite{Luscher:1996sc}\footnote{If one uses a
fermionic action that violates chiral symmetry, like Wilson fermions, the
scaling violations might naively be $\mathcal O(a)$, unless the bare coupling
$g_0$ is shifted by an term $\propto am_q$. This is a technical issue only of
practical importance, that does not change the validity of the above
statements.}. Once this ratio is known in the continuum, the experimental value
of $M_{\Omega}$ can be used to determine $\mu_{\rm had}$ in physical units
\begin{equation}
\mu_{\rm had} = M_{\Omega}^{\rm exp}
\times \frac{\mu_{\rm had}}{M_{\Omega}}\,.
\end{equation}
We suggest the reader to look again to Fig.~\ref{fig:fss_scheme} for a schematic
summary of the full procedure: the $\Lambda$-parameter, and therefore the strong
coupling constant, is determined from just {\em one} experimental dimensionful
quantity (like $M^{\rm exp}_\Omega$). Perturbation theory is only needed at
scales larger than $\mu_{\rm PT} = s^n\mu_{\rm had}$. This scale can be made
(almost) arbitrarily large with a modest (but dedicated) computational effort.
\subsection{Methods not entering the FLAG average}
\label{sec:methods-not-entering}
FLAG applies several quality criteria to determine which works enter in the
average. These quality criteria aim at covering all the possible sources of
systematic effects in the calculation. Data sets that are unlikely to allow for
a reasonable control of systematic effects are excluded from final averages. The
methods that do not qualify to enter in the final average are discussed in this
subsection.
\begin{description}
\item[QCD vertices] There are three lattice works in the latest FLAG review that
extract the value of the strong coupling from QCD
vertices~\cite{Blossier:2011tf,Blossier:2012ef,Blossier:2013ioa}. In all cases
the data set does not allow for a controlled continuum extrapolation, which
prevents these works from contributing to the average. Another
work~\cite{Zafeiropoulos:2019flq} has appeared since the FLAG review was
published, but the extraction of the strong coupling is still done from
simulations at a single lattice spacing.
The limiting factor in these determinations is the ability to
reach energy scales that are sufficiently high to make contact with perturbation
theory while having the continuum extrapolation under control.
\item[HVP] Two works extract the strong coupling from the vacuum
polarization~\cite{Shintani:2010ph,Hudspith:2018bpz}. Nevertheless these works
do not qualify for the FLAG average because the continuum extrapolation is not
under control.
{Here we face the same issues: the difficulty to reach the
perturbative region while having the continuum extrapolation under
control.}
\item[Eigenvalues of the Dirac operator] There is only a single work that
determines the strong coupling from the eigenvalues of the Dirac
operator~\cite{Nakayama:2018ubk}. The extraction is performed at a very low
energy scale, where $\alpha\approx 0.6 - 0.4$. According to the FLAG criteria,
this does not allow to control the matching with perturbation theory.
\end{description}
All in all, these methods fail to show convincing evidence for a safe contact
with the asymptotic perturbative behavior and/or fail to show that the continuum
limit can be reached at the energy scales needed to extract the value of
$\alpha_s$. At the time of writing, we think that it would be better to use the
value of the strong coupling as an input to investigate the physics related with
these phenomena, instead of using these physics effects to perform a precision
determination of the fundamental parameters of the SM. Of course this situation
might change in the future. More powerful computers might allow to reach higher
energies. Eventually a safe contact with perturbation theory in the continuum
can lead to precise values of $\alpha_s$, but this is not the case at the
moment.
\subsection{Methods that enter the FLAG average}
\label{sec:methods-that-enter}
Next we will comment critically on the methods that actually enter in the final
FLAG average (see Fig.~\ref{fig:summary}). These methods show compelling
evidence of producing robust results in the continuum limit (by using several
lattice spacing to extrapolate their results) and reach energy scales where the
perturbative matching is convincing.
A crucial point in the following discussion is the estimate of the
truncation error, {\em i.e.}\ the uncertainty that is introduced in the
determination of the coupling by working at a given order in
perturbation theory. As we shall see below, this is the main source of
uncertainty in most of the determinations of the strong
coupling~\footnote{The FLAG estimate of the truncation uncertainties
for the static potential, heavy quark correlators and small Wilson
loops determinations are larger that the uncertainties quoted by
some of the works that enter in the average, in some cases by more
than a factor two. This has in fact produced some controversies (see
for example~\cite{dEnterria:2015kmd}).}. Here we will look at these
uncertainties in detail using the scale variation method. In
particular we will examine how a given measurement of an observable in
three flavor QCD at a physical scale $Q$ produces different
determinations of $\alpha_ {\overline{\rm MS} }(M_Z)$ when the ratio
between the renormalization and physical scales is varied. Without
loss of generality, we will assume that the observable is determined
in $N_{\rm f} = 3$ QCD and normalized
as a coupling (as discussed at the beginning of
section~\ref{sec:conv-observ-coupl})
\begin{equation}
\label{eq:Pinpt}
\alpha(Q) = \alpha_{\overline{\rm MS} }(sQ) + \sum_{k=2}^n c_k(s)\alpha_{\overline{\rm MS} }^k(sQ)\,.
\end{equation}
The procedure that we employ to estimate the uncertainty in our
determination of $\alpha_{\overline{\rm MS}}(M_Z)$
is detailed in appendix~\ref{sec:scale-vari-estim}.
The reader interested in numerical values should consult the freely
available package~\url{https://igit.ific.uv.es/alramos/scaleerrors.jl}.
In summary we proceed as follows:
\begin{enumerate}
\item Fix the value $\alpha(Q)$ using a canonical value
$\Lambda^{(3)}_{\overline{\rm MS} } = 341\, {\rm MeV}$ (equivalent
to $\alpha_{\overline{\rm MS} }(M_Z) \approx 0.1185$\footnote{Note that since we assume that
the observable is determined in three-flavor QCD, we need to cross
the charm and bottom thresholds.}) for some value of the
scale $\mu=s_{\rm ref} Q$. This is achieved by {\em
reverse engineering}, {\em i.e.}\ by computing
\begin{equation}
\alpha(Q) = \alpha_{\overline{\rm MS} }(\mu) +
\sum_{k=2}^n c_k(s_{\rm ref})\alpha_{\overline{\rm MS} }^k(\mu)\,,
\qquad (\mu = s_{\rm ref}Q)\,.
\end{equation}
where $\alpha_{\overline{\rm MS} }(\mu)$ is the value of the three
flavor coupling at scale $s_{\rm ref}Q$ obtained from
$\Lambda^{(3)}_{\overline{\rm MS} } = 341$ MeV.
\item Use Eq.~(\ref{eq:Pinpt}) again, in order to determine the values of
$\alpha_{\overline{\rm MS} }(sQ)$, by solving
\begin{equation}
\alpha(Q) = \alpha_{\overline{\rm MS} }(sQ) +
\sum_{k=2}^n c_k(s)\alpha_{\overline{\rm MS} }^k(sQ)\,.
\end{equation}
at the values $s = s_{\rm ref}/2, 2s_{\rm ref}$.
\item Use the 4- and 5-loop $\beta$-function to run the values of
$\alpha_{\overline{\rm MS} }(sQ)$ obtained in step 2. to the
reference scale $M_Z$ (crossing the charm and bottom thresholds). A
comparison between the values of
$\alpha_{\overline{\rm MS}}(M_Z)$ is a measure of the truncation
uncertainty due to the scale variation.
\end{enumerate}
\begin{table}[t!]
\centering
\begin{tabular}{lllcrrr}
\toprule
Observable & loops &\(Q\) [GeV] & FLAG error [\%] & $\delta_{(4)}^\star\, [\%]$ & $\delta_{(2)}\, [\%]$& \(\delta_{(2)}^\star\, [\%]\)\\
\midrule
& & 1.5 & 1.4 & & 2.6 & 2.7\\
Potential & 4 & 2.5 & & 0.9 & 1.5 & 1.5\\
& & 5.0 & & 0.4 & 0.8 & 0.8\\
\midrule
HQ \(r_4\) & & \(m_{\rm c}\) & 1.3 & & 2.7 & 2.8\\
HQ \(r_4\) & 3 & \(2m_{\rm c}\) & &1.2 & 1.5 & 1.6\\
HQ \(r_6\) & & \(2m_{\rm c}\) & & & 2.3 & 1.2\\
HQ \(r_8\) & & \(2m_{\rm c}\) & & & 2.8 & 4.8\\
\midrule
$- \log W_{11}$ & 3 & 4.4 & 1.0 & 2.8 & 3.3 & 2.5 \\
$-\log W_{12}/u_0^6$ & &4.4 & & 3.5 & 3.2 & 3.1 \\
\midrule
FSS & 3 & 80 & & 0.1 & 0.2 & 0.2\\
\bottomrule
\end{tabular}
\caption{\label{tab:scale_truncation}
Summary of truncation uncertainties on $\alpha_{\overline{\rm MS} }(M_Z)$ estimated by varying the scale. We compare the error quoted by flag with a change of scale by factors two and four around \(s=1\) or \(s=s^\star\) (the value of fastest apparent convergence).
For each method we quote the number of known loops in the relation
between the observable and the $\overline{\rm MS} $ scheme according
to h counting in section~\ref{sec:conv-observ-coupl}.
Details on the different types of extractions can be found in
section~\ref{sec:static-potential-1} (potential),
section~\ref{sec:heavy-quark-corr-1} (HQ),
section~\ref{sec:observ-at-cutoff} ($\log W_{11}, \log W_{12}$) and
section~\ref{sec:finite-size-scaling-1} (FSS).}
\end{table}
The usual procedure in phenomenology is to vary the renormalization
scale by a factor 2 above and below some reference scale. Estimates of
the truncation uncertainties that use renormalization scales below 1
GeV tend not to be reliable. Since many extractions of the strong
coupling are performed at relatively low scales, the above mentioned
procedure might lead to unreasonable uncertainties. For this reason,
the uncertainty resulting from comparing the change
$s_{\rm ref}\to 2s_{\rm ref}$ provides complementary information on
the size of the truncation uncertainties, specially if one explores
the dependence with $s_{\rm ref}$ in a range $1-2$. In order to get a
more quantitative understanding of these effects, we will use the
following quantities.
\begin{description}
\item [$\delta_{(4)}(s_{\rm ref})$:] Change the renormalization scale by a factor
two above and below some reference scale $s_{\rm ref}Q$. We quote a symmetric
error by averaging the difference between the scales $s_{\rm ref}Q$ and
$2s_{\rm ref}Q$, and the difference between the scales $s_{\rm ref}/2Q$ and
$s_{\rm ref}Q$. Note however that in some cases the error is markedly
asymmetric.
\item[$\delta_{(2)}(s_{\rm ref})$:] Change the renormalization scale by a factor
two above the reference scale $s_{\rm ref}Q$ only.
\end{description}
We will show explicitly in the computations below how the two measures
$\delta_{(4)}(s_{\rm ref})$ and $\delta_{(2)}(s_{\rm ref})$ depend on
the choice of $s_{\rm ref}$. In principle any number $\mathcal O(1)$ is a
reasonable choice for $s_{\rm ref}$, but there can be significant
differences in the results depending on its actual value. For this
reason we will explore two common choices.
\begin{itemize}
\item Take $s_{\rm ref} = 1$.
{\em i.e.}\, the renormalization and physical scales are the same.
In this case the uncertainties will be labeled $\delta_{(2)}, \delta_{(4)}$.
\item Take $s_{\rm ref} = s^\star$ as \emph{the value of fastest apparent
convergence}. This value is determined with the condition
\begin{equation}
c_2(s^\star) = 0
\end{equation}
{\em i.e.}\ the NLO coefficient in the relation between $\alpha(Q)$ and
$\alpha_s$ vanishes (see Eq.~(\ref{eq:Pinpt})).
In this case the uncertainties will be labeled $\delta^\star_{(2)},
\delta^\star_{(4)}$.
\end{itemize}
A summary of the results is presented in table~\ref{tab:scale_truncation}. One
can readily see that the truncation uncertainties obtained with this method are
in the same ballpark as the FLAG uncertainties, except for the case of
the Wilson loops, where our estimates are substantially larger. Once
again we would like to end
with a warning: estimates of the truncation uncertainties based on the scale
variation can (and have been shown to) fail in some cases (see discussion in
section~\ref{sec:syst-extr-alph} and specifically figure~\ref{fig:msbar}).
\subsubsection{Finite size scaling}
\label{sec:finite-size-scaling-1}
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/fss.pdf}
\caption{}
\end{subfigure}
\begin{subfigure}[t]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/fss_scale_err.pdf}
\caption{}
\end{subfigure}
\caption{Truncation effects in $\alpha_s$ extractions from finite
size scaling.
Note that in this methods the extraction is performed at very high
energies $Q\approx 80$ GeV.
The scale dependence of the coupling at $Q=8$ GeV in plot (a) is
just for reference.
Plot (b) shows that truncation uncertainties are typically much
smaller than the quoted uncertainties.}
\label{fig:fsstruncation}
\end{figure}
The FLAG average is the result of Refs.~\cite{Aoki:2009tf,Bruno:2017gxd}, wich
are in good agreement with each other. Perturbation theory is used at very high
energies ($Q \approx 80$ GeV), where perturbative estimates of the truncation
uncertainties are reliable. They affect the extraction of $\alpha_s$
only at the $0.1-0.2\%$ level (see figures~\ref{fig:fsstruncation}), well below
the quoted uncertainties.
The most recent work~\cite{Bruno:2017gxd} explores the dependence of
the extractions of $\Lambda$ on the physical energy scale over a large
range of energy scales $Q\sim 4 - 140$ GeV (see Figure~\ref{fig:msbar}
and the related discussion). This study compares their extraction of
the coupling $\alpha_s$ with the extraction performed with several
observables after extrapolating the renormalization scale at which
perturbation theory is used $\mu \to \infty$ (see discussion in
section~\ref{sec:syst-extr-alph}). All in all, results based on finite
size scaling do not depend on estimates of the perturbative
uncertainties done in perturbation theory, although
figure~\ref{fig:fsstruncation} show that they tend to be rather small,
as expected given that the electroweak scale is reached by
non-perturbative simulations.
We should also point to the detailed study in~\cite{DallaBrida:2018rfy}.
The continuum extrapolation is the main source of systematic
uncertainty. {According to our
discussions in sections~\ref{sec:cont-limit-scale}
and~\ref{sec:finite-size-scaling}} the method allows to
to perform a controlled continuum extrapolation by using several
values of the lattice spacing at each energy scale. In particular, the
most delicate continuum extrapolations in~\cite{DallaBrida:2016kgh}
uses three values of the lattice
spacing that vary by a factor two at each value of the scale.
The statistical precision of the non-perturbative running is the main limiting
factor of these determinations. Note that coupling definitions using the
gradient flow -- see section~\ref{sec:theory-scales}
and~\ref{sec:finite-size-scaling} -- have a very small variance. These couplings
have been used in Ref.~\cite{Bruno:2017gxd}, but not in Ref.~\cite{Aoki:2009tf}
(these techniques were not known at the time), explaining the large difference
in the error between the two works: the result in~\cite{Aoki:2009tf} quotes a
2.5\% error in $\alpha_{\overline{\rm MS} }(M_Z)$, while the
result~\cite{Bruno:2017gxd} quotes a 0.7\% uncertainty.
\subsubsection{Static potential}
\label{sec:static-potential-1}
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/pot.pdf}
\caption{}
\end{subfigure}
\begin{subfigure}[t]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/pot_scale_err.pdf}
\caption{}
\end{subfigure}
\caption{Truncation effects in $\alpha_s$ extractions from the
static potential.
Note that the high energy scale $Q=8$ GeV is only reached in the
last work~\cite{Bazavov:2019qoo}.
The FLAG error is determined for }
\label{fig:pottruncation}
\end{figure}
The FLAG average is basically the result of Ref.~\cite{Bazavov:2014soa} with an
added uncertainty because of the perturbative truncation errors. Since the
publication of the FLAG report, two new determination of the strong coupling by
the same group has been published~\cite{Bazavov:2019qoo,
Ayala:2020odx}. In these last works the
determination of the strong coupling is extracted at energy scales $Q\approx 2.5-8$
GeV\footnote{The same work also extract the
strong coupling at shortes distances in a finite temperature setup,
finding good agreement.}, improving significantly the previous
determination, which used $Q\in
1.4-4.9$ GeV.
Energy scales $\approx 5$ GeV can be reached with several values of
the lattice spacing.
On the other hand the largest energy scales are reached at a single
lattice spacing, preventing a proper continuum extrapolation.
The strong coupling is typically extracted from fits using
energy scales about $2.5$ GeV and above.
We should recall that the simulations at very fine lattice
spacing have a relatively small volume and are performed at fixed topology (see
discussion in section~\ref{sec:static-potential}), and in relatively
small physical volumes.
In what follows we will focus on the truncation uncertainties.
In our analysis we ignore the logarithmic corrections due the IR
divergences of the static potential. This is partially justified: if
one takes the logarithms as constants, truncation uncertainties as
estimated below show only a mild variation (see appendix~\ref{sec:static-potential-app}).
Nevertheless there is some concern here, since the natural scale at
which the terms $\log\alpha$ are evaluated (the \emph{ultra-soft}
scale) is significantly smaller (see discussion in
section~\ref{sec:static-potential}).
Figure~\ref{fig:pottruncation} shows the scale dependence of the strong coupling
and the truncation errors $\delta_{(2)}^{\rm ref}, \delta_{(4)}^{\rm ref}$ as a
function of $s_{\rm ref}$. First it is important to point out that the error
estimate by FLAG, even if it comes from a completely different argument is in
the same ballpark as our estimates. For the relevant energy scales used in
present works~\cite{Bazavov:2014soa,Bazavov:2019qoo} ($Q\in 1.4-5$ GeV)
the perturbative estimates of the truncation uncertainties are in
the range $\delta\alpha_{\overline{\rm MS} }(M_Z) \approx 2.6\%-0.8\%$, while
the FLAG estimate of the truncation uncertainties is 1.4\% (the most recent
work~\cite{Bazavov:2019qoo} had not appeared at the time the FLAG review was
published).
It is interesting to consider in detail the analysis of
Refs.~\cite{Bazavov:2014soa, Bazavov:2019qoo, Ayala:2020odx}.
\begin{description}
\item [Ref.~\cite{Bazavov:2014soa}:] The perturbative truncation error is
estimated by changing the renormalization scale a factor $\sqrt{2}$ above and
below the physical scale. Their result
\begin{equation}
\label{eq:pot_res1}
\alpha_{\overline{\rm MS} }(M_Z) = 0.1166_{-0.0008}^{+0.0012}\,,
\end{equation}
has an uncertainty between $-0.7\%$ and $+1\%$, that is actually dominated by
the perturbative truncation uncertainty. This uncertainty is smaller than the
quoted uncertainty by FLAG ($1.4\%$).
\item[Ref.~\cite{Bazavov:2019qoo}:] The estimate comes from a similar analysis,
but this time the renormalization scale is varied a factor 2 above and below
the physical scale. Moreover they also include in their estimate the effect of
the different treatment of the logarithmic corretions in the perturbative
expansion (that we have ignored in our analysis). This is very similar to our
approach, and the uncertainty is similar to our quoted $\delta_{(4)}$ in
table~\ref{tab:scale_truncation}, with the exception that they do not
symmetrize the error. Their result
\begin{equation}
\label{eq:pot_res2}
\alpha_{\overline{\rm MS} }(M_Z) = 0.1166_{-0.0006}^{+0.0011}\,,
\end{equation}
has a very similar uncertainty than the previous work, see
Eq.~(\ref{eq:pot_res1}), despite the fact that they reach significantly higher
energies. Obviously this is a result of the more conservative approach to the
estimate of the truncation uncertainties.
Reference~\cite{Bazavov:2019qoo} claims that if they would have
followed the same recipe as in reference~\cite{Bazavov:2014soa},
their updated result Eq.~(\ref{eq:pot_res2}) would have the
uncertainty reduced by a factor two.
\item[Ref.~\cite{Ayala:2020odx}:] This work uses the same data as
reference~\cite{Bazavov:2019qoo}, {but they use the known terms in
the perturbative series to fix the normalisation of the renormalon
ambiguity and subtract some non-perturbative ({\em i.e.}\, power)
corrections}. Their final result
\begin{equation}
\label{eq:pot_res3}
\alpha_{\overline{\rm MS} }(M_Z) = 0.1181(9)\,,
\end{equation}
quote a similar uncertainty as~\cite{Bazavov:2019qoo}, but their
central value is significantly larger.
\end{description}
In summary, the estimate by FLAG of these uncertainties (1.4\%) is reasonable.
It is basically the same as the difference in central
values between the two most recent works~\cite{Bazavov:2019qoo} and
~\cite{Ayala:2020odx}, that use the same data but different
strategies to match with perturbation theory.
This uncertainty is also larger than the total quoted uncertainty in
both works~\cite{Ayala:2020odx, Bazavov:2019qoo}.
Determinations of the strong coupling from the static potential have recently
improved significantly. Also the estimate of perturbative
uncertainties is more conservative compared with previous works.
These determinations are in very good shape, but the following points
should be better understood:
\begin{itemize}
\item The most important point to be understood is {if the
perturbative region is reached in current extractions.
The significant
difference in central values between extractions using the same
dataset but different power corrections ({\em i.e.}\,
Refs.~\cite{Ayala:2020odx, Bazavov:2019qoo}) needs a better understanding}.
This difference between central values is about $1.4\%$, similar to
the uncertainties that we obtained from a simple scale variation approach.
We emphasize that our estimate \emph{does not address the
issue} of the logarithmic corrections to the perturbative series,
related with the IR divergences of the static potential.
\item Another manifestation of the same problem is the strong dependence on the value of
$\Lambda^{(N_{\rm f} =0)}$ that has been observed for extractions based on the static
potential for values of the coupling $\alpha^3_s \lesssim
0.01$~\cite{Husung:2017qjz}.
This effect has been studied in~\cite{Bazavov:2019qoo} and they
observe a much milder effect.
Note however that in this last reference $\Lambda$ is extracted as a
fit over a range of energy scales. This procedure makes it much more
difficult to see any dependence in the extracted value of $\Lambda$,
and in fact means that all the different extractions of $\Lambda$
(with the exception of one point) use data with $\alpha^3> 0.01$.
\item The effect of the bad sampling of the topological charge has still to be
investigated in detail. The ensembles at the finest lattice spacing have the
topology completely frozen (see~\cite{Bazavov:2017dsy}). The impact of frozen
topology can be studied by computing the scale $r_1$ from ensembles that are
stuck in distinct topological sectors. The numerical results so far are all
compatible within errors. Note however that the values of the topological charge
simulated (basically 2 and 0) are small. A dedicated quenched study could shed
some light on these issues.
\item A related problem is that the physical volumes are relatively
small (with $m_\pi L \approx 2.2$), and the effect of this small
volumes on the determination of the scale and the static force is
unclear. On general grounds, the static force is a short distance
observable and one expects finite volume effects to be very small.
Probably the largest source of finite volume effects are in the
scale determination $r_1$, but one should keep in mind that the
uncertainty in the scale has a small effect on the uncertainty of
$\alpha_s(M_Z)$. It would be interesting to quantify at which level
of precision finite volume effects start to be a concern for these
determinations.
\end{itemize}
\subsubsection{Heavy quark correlators}
\label{sec:heavy-quark-corr-1}
\begin{figure}
\centering
\begin{subfigure}[t]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/hq_comp_mc.pdf}
\caption{}
\end{subfigure}
\begin{subfigure}[t]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/HQr4_scale_err.pdf}
\caption{}
\end{subfigure}
\caption{Truncation effects in $\alpha_s$ extractions from heavy
quark correlators.}
\label{fig:HQtruncation}
\end{figure}
The average in figure~\ref{fig:summary} is the result of the
works~\cite{McNeile:2010ji,Chakraborty:2014aca,Nakayama:2016atf}. Once
again the most delicate sources of uncertainty are the perturbative
error and controlling the continuum extrapolation. It is instructive to
see how different works estimate that
particular systematic error.
\begin{description}
\item [JLQCD collaboration~\cite{Nakayama:2016atf}:] The authors use
a scale variation method, very similar to our procedure (see
figure~\ref{fig:HQtruncation}). A crucial difference is that they
vary independently both the renormalization scale in the quark mass
and in the strong coupling in the range $2-4$ GeV (this typically
increases the estimates of the uncertainties). Also this work does
not use the moment $r_4$. This is the quantity with the best
perturbative behavior (see figure~\ref{fig:HQtruncation}), but they
find the continuum extrapolation very challenging. Their result
\begin{equation}
\alpha_{\overline{\rm MS} }(M_Z) = 0.1177(26)\,,
\qquad [2.2\%]\,,
\end{equation}
claims a $\sim 2\%$ error, mostly dominated by the perturbative
truncation uncertainties.
\item [Ref.~\cite{Maezawa:2016vgv}:] They estimate the truncation uncertainty by
estimating the $\alpha^4(\mu)$ term in the perturbative expansion of $r_n$.
They use a range of values $c_3(1) = \pm 2c_2(1)$ (see Eq.~(\ref{eq:Pinpt})),
which yields a perturbative truncation uncertainty that is much smaller than
the one obtained by the scale variation method. The truncation uncertainty
represents a negligible contribution to the uncertainty of their final result
\begin{equation}
\alpha_{\overline{\rm MS} }(M_Z) = 0.11622(84)\,,
\qquad [0.7\%]\,,
\end{equation}
which is 3 times more precise than the JLQCD result.
\item[Ref.~\cite{Petreczky:2019ozv}:] This work can be considered an update
of~\cite{Maezawa:2016vgv}. Again the perturbative uncertainty is computed by
estimating the size of the $\mathcal \alpha^4(\mu)$ term in the perturbative
expansion of $r_n$, but this time they allow a larger range of values for the
unknown coefficient $c_3(1) = \pm 5c_2(1)$ (see Eq.~(\ref{eq:Pinpt})). Their
updated result
\begin{equation}
\alpha_{\overline{\rm MS} }(M_Z) = 0.1159(12)\,,
\qquad [1.0\%]\,,
\end{equation}
has in fact a larger uncertainty.
\item[Ref~\cite{Chakraborty:2014aca}:] The HPQCD collaboration analyze
data close to the physical charm quark mass ($Q=m_{\rm c}$). Their
analysis includes higher order terms in the perturbative expansion
of their data amongst the fitted parameters. These unknown terms (up
to powers $\alpha^{15}$) are constrained using Bayesian priors. The
perturbative truncation errors are the main source of uncertainty,
but their final result
\begin{equation}
\alpha_{\overline{\rm MS} }(M_Z) = 0.11822(74)\,,
\qquad [0.6\%]\,,
\end{equation}
claims an uncertainty 4 times smaller than our estimates of the truncation
uncertainties at the scale $Q=m_{\rm c}$. This uncertainty is estimated by
varying the number of terms added to the fit. It is not clear to the authors
why this estimate should be a reliable estimate of the truncation
uncertainties. In particular these error estimates are substantially
smaller than the usual estimates coming from scale variation.
\end{description}
FLAG quotes a 1.2\% truncation uncertainty for extractions performed at the
scale $m_{\rm c}$. Our scale variation method tends to point to even larger
values for the truncation uncertainty (see figures~\ref{fig:HQtruncation}):
about 2-3\% for extractions at $m_{\rm c}$ and between 1-2\% for extractions at
$2m_{\rm c}$.
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{fig/HQc3.pdf}
\caption{Truncation effects in $\alpha_s$ extractions using the last
term in the perturbative series.
The horizontal dashed line represent the estimate of
Ref.~\cite{Petreczky:2019ozv} $\delta\alpha = 5c_2(1)\alpha^4(m_{\rm c})$.
This uncertainty is significantly smaller than the last known term in
the series (dashed curve) unless the renormalization scale is chosen
very close to the physical scale ({\em i.e.}\, $s\approx 1$).
This estimate is also smaller than the naive estimate coming
\emph{only} from the scale dependence of $c_3$ (solid curve).}
\label{fig:HQc3}
\end{figure}
It is easy to see why the estimates based on the value of the unknown
coefficient $c_3(1)$ lead to small uncertainties.
The last known coefficient in the series is given by
\begin{equation}
c_2(s) = 0.0796 + 0.588 \log(s) + 2.052 \log^2(s)\,.
\end{equation}
This coefficient is anomalously small for $s=1$.
Even when multiplied by a factor 5, it leads to a small estimate for the
coefficient $|c_3|\approx 0.4$.
On the other hand $c_3(s)$ can be estimated by other means.
The scale dependence of $c_3(s)$ is fully predicted by the RG
equations
\begin{equation}
c_3(s) = c_3(1) + 0.865\log(s) + 2.425\log^2s + 2.939\log^3 s\,.
\end{equation}
({\em i.e.}\, only the non-logarithmic dependence $c_3(1)$ is unknown). This
logarithmic dependence alone predicts a coefficient
$|c_3(s)|\approx 3$ for even modest values of the scale $s=2$. This
value for the $c_3$ term is 7.5 times larger than the estimate used in
the last work Ref.~\cite{Petreczky:2019ozv} and almost 20 times larger
than the estimate of~\cite{Maezawa:2016vgv}. There is an extra
suppression $\alpha_s^4(sQ)$ that makes the uncertainty smaller when
$s>1$, but this effect can only account for a factor between 2.8 (for
$Q=m_{\rm c}$) and 4.5 (for $Q=2m_{\rm c}$): clearly insufficient to
compensate the factor 7.5 or 20 in the estimate of $c_3$. Moreover,
note that this simple estimate of $c_3(s)$ neglects completely
$c_3(1)$. (See Figure~\ref{fig:HQc3}). This explains why uncertainties
based on scale variation are generally larger.
Even assuming that one order more is known, and that
$c_3(1) = 0$, the scale variation approach results in $\approx 1.5\%$
error for $r_4$ at the charm scale.
Estimates of the
truncation uncertainties based on varying the number of fit parameters
constrained by priors and using Bayesian methods lacks a solid
theoretical basis. We suggest that estimates based on scale variation
should be preferred.
These considerations, together with the complicated scaling violations (see
discussion in section~\ref{sec:heavy-quark-corr}) make it very challenging to
obtain $\alpha_s$ with less than a 1.5\% uncertainty using these methods. Quark
masses significantly larger than the physical charm quark mass would be required
(and one would need to deal with complicated continuum extrapolations).
There are two interesting directions for future research. First, the next order
in the perturbative relation of the heavy quark moments could be very useful in
future extractions. Second a dedicated pure gauge study, where significantly
larger energy scales could be explored, would shed some light on the
difficulties associated with the truncation uncertainties and the approach to
the continuum in this type of lattice determinations.
\subsubsection{Observables at the cutoff scale}
\label{sec:observ-at-cutoff}
\begin{figure}
\centering
\begin{subfigure}[t]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/Wloop.pdf}
\caption{}
\end{subfigure}
\begin{subfigure}[t]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/Wloop_scale_err.pdf}
\caption{}
\end{subfigure}
\caption{Truncation effects in $\alpha_s$ extractions from
quantities at the scale of the cutoff.}
\label{fig:Wlooptruncation}
\end{figure}
There are two studies that contribute to this
average~\cite{Maltman:2008bx,McNeile:2010ji}\footnote{Reference~\cite{McNeile:2010ji}
simply updates the analysis of~\cite{Davies:2008sw} with a more
precise determination of the scale.}. The main contribution to the
uncertainty in these determinations is purely systematic, with the
perturbative uncertainties playing a leading role. Note that an
advantage for these observables is that there is no need to perform a
continuum extrapolation.
It is very difficult to obtain an independent estimate of the truncation
uncertainties for these observables. The reason is that the data
\emph{does not follow the known perturbative prediction} in the range
of energy scales reached in the numerical
simulations~\cite{Maltman:2008bx,McNeile:2010ji}. Several terms are
added to the perturbative expansion, up to terms $\alpha^{10}_{\rm V}$ (see
Eq.~(\ref{eq:alphaW})), and the higher-order coefficients are
fixed by fitting the lattice data to the expression
\begin{equation}
\label{eq:alphaWpt}
\alpha_{W_{nm}}(1/a) =
\alpha_{{\rm V} }(\mu) + c_2\alpha_{{\rm V} }^2(\mu)
+ c_3\alpha_{{\rm V} }^3(\mu) +
\sum_{k=4}^{10} d_i \alpha_{{\rm V} }^i(\mu)\,.
\end{equation}
Here $\alpha_{\rm V}$ is the coupling in the potential scheme (see
Ref.~\cite{Lepage:1992xa}) and $\mu \approx \pi/a$ with the exact
value depending on the observable). A crucial point in these extractions is the
discretization used.
We will focus in the same discretization as used in the
works~\cite{Davies:2008sw,McNeile:2010ji}.
At least one extra term, with an unknown coefficient, is necessary in
order to obtain a satisfactory description of the data. Moreover, the
data is not precise enough to determine all the extra coefficients, so
Bayesian priors are used in order to constrain the size of the
coefficients $d_i \approx 0 \pm 2.5$. The HPQCD
analyses~\cite{McNeile:2010ji} estimate the perturbative truncation error
by varying the number of terms in expression
Eq.~(\ref{eq:alphaWpt}). Firstly let us note that the perturbative
expansion of $\alpha_{W_{nm}}$ is only expected to be asymptotic. In
particular the coefficients in the perturbative expansion should
eventually grow. Constraining the size of all coefficients to the same
size $\mathcal O(1)$ is not justified based on theoretical arguments.
Following the exact same procedure that we used for the other observables, we
can estimate the truncation error by performing a scale variation analysis. For
this purpose, we describe the observable using the \emph{known} perturbative
coefficients in the perturbative expansion in terms of the renormalized $\overline{\rm MS} $ coupling:
\begin{equation}
\label{eq:alphaWptMSbar}
\alpha_{W_{nm}}(1/a) =
\alpha_{{\overline{\rm MS}} }(\mu) + c_2\alpha_{{\overline{\rm MS}} }^2(\mu)
+ c_3\alpha_{{\overline{\rm MS}} }^3(\mu) + \dots\, ,
\end{equation}
with $\mu \approx 2.4/a$ (the exact relation depending on the concrete
quantity). For illustration, let us focus here on the shortest
distance object (the plaquette), and the $1\times 2$ Wilson loop with
tadpole improvement:
\begin{align}
\alpha_{W_{11}}(a) &= -\frac{1}{c_1^{(11)}}\log W_{11} =
\alpha_V(d/a) + \frac{c_2^{(11)}}{c_1^{(11)}}\alpha_V^2(d/a)
+ \frac{c_3^{(11)}}{c_1^{(11)}} \alpha_V^3(d/a) + \dots \\
\alpha_{W_{12}}(a) &= -\frac{1}{c_1^{(12)}}\, \frac{\log W_{12}}{u_0^6} = \alpha_V(d/a) + \frac{c_2^{(12)}}{c_1^{(12)}}\alpha_V^2(d/a)
+ \frac{c_3^{(12)}}{c_1^{(12)}} \alpha_V^3(d/a) + \dots \, .
\end{align}
These quantities show a very strong dependence on the scale (see
figures~\ref{fig:Wlooptruncation}), pointing to truncation uncertainties of the
order 3\%.
FLAG uses the HPQCD fit result, $d_4\approx 2$, to estimate the truncation
uncertainty as $2\alpha^4$. This procedure results in a smaller value than the
one obtained form the scale variation approach
\begin{equation}
\delta\alpha_{\overline{\rm MS} }(M_Z) \approx 0.0012\,,
\qquad [1\%]\,.
\end{equation}
This last uncertainty is about 2.5 times larger than the estimate of HPQCD.
A last piece of information comes from comparing the results in
Refs.~\cite{Maltman:2008bx,Davies:2008sw}. They use basically the same dataset
(reference~\cite{Maltman:2008bx} uses a subset of the 22 quantities used
in~\cite{Davies:2008sw}), but analyse it using different perturbative
expressions and deal with the non-perturbative corrections in different ways.
Their results for $\alpha_{\overline{\rm MS} }(M_Z)$ differ by a 0.6\%--1.2\%.
The overall conclusion is that truncation uncertainties based on varying the
number of terms constrained with Bayesian priors underestimate significantly the
true uncertainties. The same phenomena have been observed in the case of
determinations based on currents of heavy quark correlators. Moreover it is
important to recall that the lattice data for these short distance quantities do
not follow the perturbative prediction if only the known coefficients are used,
making it mandatory to fit several additional terms in the perturbative
expansion constrained by Bayesian priors. We think that this method needs
further study. In particular it is mandatory to find a solid estimate of the
truncation uncertainties based on the same techniques that are common in the
estimates of the perturbative QCD uncertainties. A detailed study in pure gauge
could shed some light on the issue. Without these insights the claimed
uncertainties of some of these computations (0.6\%) seem an underestimate of the
true uncertainties.
\subsection{An opinionated and critical summary}
What can we conclude on the status of the lattice determinations of the strong
coupling? First we should only consider extractions with a dataset that allows a
proper control over the different sources of systematic effects. This is the
approach taken for example by the FLAG review. For the particular case of the
determination of the strong coupling, the most delicate points in the
extractions of the strong coupling is the estimate of the uncertainties related
with the truncation of the perturbative series and the large scaling violations
in short distance quantities. Note that controlling both sources of systematic
uncertainties is challenging: cutoff effects are larger for short distance
quantities, while perturbative truncation errors are larger for low energy
quantities, \emph{the window problem} described in
section~\ref{sec:chall-determ-alph} is again the limiting factor. This is why we
have insisted on these points along the review.
Many lattice methods do not allow a simultaneous control of these two
sources of systematic effects with current computer resources.
Typically in these cases the extractions of the strong coupling are
performed at a single lattice spacing, and/or the extraction is
performed at energy scales where significant contribution from
non-perturbative effects are present. Our analytic understanding of
these effects is, at best, very limited. In practical terms we have
to deal with them by fitting these contributions. This is very
delicate, since distinguishing the perturbative running from the
non-perturbative contributions when the data is available only in a
restricted range of scales is far from easy. We therefore prefer to
focus on methods where the data follows the perturbative prediction
and non-perturbative corrections are smaller than the uncertainties.
Still, with this ambitious aim, several methods that allow a reliable
extraction the strong coupling have been developed in lattice
QCD. These methods differ in their control over the systematic errors
associated with the continuum extrapolation and the truncation
uncertainties.
\begin{itemize}
\item Finite size scaling is the only method that offers a \emph{solution} to
the window problem instead of trying to find a \emph{compromise}. Arbitrarily
large energy scales can be reached, and the continuum extrapolation can be
performed by using several lattice spacing at each constant renormalization
scale.
This strategy trades the systematic errors associated with the
truncation of the perturbative series at relatively low scales with
the statistical error accumulated when computing the
non-perturbative running. It remains challenging to obtain precise
results, but thanks to several recent developments, a sub-percent
precision has been reached by these kind of determinations. The
truncation uncertainties are negligible, since perturbation theory
is typically applied at the electroweak scale. Several observables
have been studied non perturbatively and in some cases agreement
with perturbation theory is achieved over a large range of scales
(from 4 to 140 GeV)~\cite{DallaBrida:2018rfy,Bruno:2017gxd}.
Our reservation with this approach is that until now only two groups have used
it to determine the strong coupling, and with very similar setups. A new
independent determination would be welcome.
\item Determinations based on the static potential have several advantages over
most other determinations. First, the relevant perturbative relations are
known to N$^3$LO, one order more than what is typically known for other
observables. Second, the non-perturbative lattice data seems to follow the
perturbative prediction for energy scales as low as 2 GeV. This energy scale
can be reached at several values of the lattice spacing and the data can be
extrapolated to the continuum.
At the moment of writing, two new studies~\cite{Bazavov:2019qoo,
Ayala:2020odx} have been published.
One of them can be
considered an update of some of the works evaluated by FLAG. This
new determination improves significantly in the
energy scales that they are able to reach. What is more important, they treat
the perturbative uncertainties more conservatively than in previous works
and claim a sub-percent precision in the strong
coupling. The other recent work (Ref.~\cite{Ayala:2020odx}) uses
the same dataset but a different treatment of the lattice data and
matching with perturbation theory.
They obtain a result for $\alpha_{\overline{\rm MS} }(M_Z)$ about
1.3\% larger, rising some doubts on the claimed sub-percent accuracy
of these works.
According to our analysis using the scale variation approach, the
uncertainty quoted by FLAG is reasonable, although our method
neglets the delicate issue of logarithmic corrections to the
perturbative series.
As with all large volume determinations, there are reservations with
this method beyond the estimates of perturbative uncertainties,
related with the compromises
that are made. Determinations reach high energy scales (up to
8 GeV), but the energy scales where the continuum limit can be taken
with several values of the lattice spacing are substantially lower.
Some of the volumes simulated are relatively small
($m_\pi L \approx 2.2$). Moreover some of these simulations are
affected by the problem of topology freezing, potentially making
difficult to asses the statistical errors correctly. The range of
scales where the methodology can be tested is limited, and we should
always remember that truncation uncertainties estimated within
perturbation theory can be misleading (see
section~\ref{sec:syst-extr-alph}), \emph{even when a conservative
approach is taken}. These points should (and will) be investigated
further, but beyond any doubts determinations coming from the static
potential have reached a considerable level of maturity and
precision.
\item Some extractions based on currents of heavy quark correlators
are among the most precise determinations of the strong
coupling. Scaling violations have a complicated functional form, but
at least the continuum limit can be explored with several lattice
spacing. On the other hand, and compared with extractions based on
the static potential, our current perturbative knowledge in these
extractions is one order less and cutoff effects seem larger. In
summary, this method is more challenging
than the static potential from both sides of the window
problem. Most of the determinations are performed at the charm quark
mass $M_{\rm c}$ GeV. At these low scales, truncation
uncertainties estimated using the scale variation method are about
$2-3\%$ in $\alpha_{\overline{\rm MS} }(M_Z)$. The FLAG 2019 review
quotes a smaller uncertainty ($1.5\%$). Some
works claim a sub-percent precision, but we find it difficult to
consider these estimates reliable.
There are several directions in which these extractions can improve. First, it
should be possible to extend the current perturbative knowledge by one order
more. Second, this method has never been studied in detail in the pure gauge
theory, where one could reach energy scales significantly larger than $M_{\rm
c}$ with full control over the continuum extrapolations. In any case, until
the perturbative knowledge is known and larger energy scales have been
studied, it seems difficult to claim a smaller uncertainty than what the FLAG
review assigns to these extractions.
\item Finally, there is still work to do in order to better understand the
determinations based on lattice observables defined at the cutoff scale.
Despite these methods claiming the smallest uncertainties, this claim is not
backed up by a solid analysis of the truncation uncertainties. A scale
variation approach points to significantly larger uncertainties. In fact the
truncation uncertainties in these extractions seem to be significantly larger
when using several different methods. Our numerical investigation
suggests that even the FLAG error is still underestimating this source of
systematic uncertainty.
One should also point out that in pure gauge there is a significant
discrepancy between some recent extractions based on observables defined at
the cutoff scale and a recent extraction using finite size scaling
(see~\cite{DallaBrida:2019wur}).
\end{itemize}
Let us end this section with a general comment about figure~\ref{fig:summary}.
Basically every method to extract the strong coupling using lattice QCD seems
to be able to reach a better precision than any phenomenological
extraction. This seems at least in some cases, to be fully justified based on
the general principles on which we have insisted along the review: the most
precise phenomenological determination, the extraction of $\alpha_s$ from
$\tau$ decays, is performed at a \emph{fixed} energy scale
$m_{\tau}\approx 1.7$ GeV. Larger energy scales cannot be probed in these
extractions, while the perturbative expansion of the observable is known at
the same order as in the extractions based on the static potential, where
energy scales $\mu \approx 5$ GeV can be explored at several values of
the lattice spacing. In finite size
scaling the perturbative knowledge is one order less, but perturbation theory
is applied at the electroweak scale and consistency of the results is checked
in the energy range $4-140$ GeV for a one-parameter family of observables.
\subsection{The future of lattice determinations of $\alpha_s$}
Now that we understand the main limitations of the different lattice
methods to extract the strong coupling we are in a good position to
discuss what is needed in order to substantially reduce the current
uncertainty in the strong coupling.
From a general point of view we must realize that, with the exception
of finite size scaling methods, the limitations of the lattice
determinations of the strong coupling are a direct consequence of the
\emph{window problem}: the fundamental compromise between reaching
large energy scales, where the perturbative behavior is better, and
using low renormalization scales, where the continuum extrapolation is
well under control.
One should also be aware that truncation uncertainties decrease with
powers of the coupling, and therefore slowly ({\em i.e.}\, logarithmically)
with the scale (see figure~\ref{fig:scale_errors}).
This makes the window problem hard to solve by brute force.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{fig/scale_errors.pdf}
\caption{Scale truncation uncertainties for different lattice
methods (we quote $\delta_{(2)}$, see section~\ref{sec:methods-that-enter}).
The dark shaded region is the current accessible range for large
volume lattice simulations nowadays.
The light shaded region represents the accessible scales if a
reduction by a factor two in the lattice spacing is possible in the future.
Note that most methods require a continuum extrapolation with several
lattice spacings: in practice the accessible renormalization scales
for all methods except finite size scaling (SF) is at least factor two/four
smaller than the lattice spacing.}
\label{fig:scale_errors}
\end{figure}
What progress can we expect on this front? Since the \emph{window problem} is
basically a computational limitation we expect to see improvements in the
future by just waiting. Computer power has been steadily improving during the
last 50 years, and most probably will continue to do so, with exascale
machines looming on the horizon. Pushing the UV cutoff of a simulation ({\em i.e.}\
reducing the lattice spacing of the simulation) by a factor two keeping the
physical parameters fixed multiplies the computational cost by a
factor~\footnote{Naively the simulation scales with the lattice volume
$\propto (L/a)^4$. At the same time simulations at smaller lattice spacing
show larger autocorrelations, that are expected to scale like $a^2$ in
absence of topology freezing. On top of these six powers of $a$, the
integration of the molecular dynamic equations requires to reduce the step
size at smaller lattice spacing, which brings down another power of $a$ to
the scaling. } $2^7 = 128$. We can therefore expect that on exascale
machines the state of the art lattice simulations will be performed on lattice
spacing reaching down to $a\approx 0.02$~fm, assuming that topology freezing
is still under control on such fine lattices. Beyond the trivial observation
that smaller lattice spacing would allow to check the extrapolations
thoroughly, the expected increase in computing power would allow the following
improvements.
\begin{enumerate}
\item Determinations based on QCD vertices and the hadron vacuum
polarization, that are nowadays limited by the size of scaling
violations, will be able to perform a continuum extrapolation at fixed
renormalization scale with several lattice spacing. This will bring
the scaling violations under control.
As of today these methods require to include power corrections as fit
parameters to describe the lattice data. Having access to shorter
distances would clarify whether the perturbative region can be
accessed without having to include these power corrections.
\item Pushing the UV cutoff a factor two higher would allow to match
with the perturbative regime at larger renormalization scales.
For determinations based on the static potential, whose $\beta$-function is
known up to 4-loops, one expects truncation errors $\mathcal O(\alpha^3)$ to
be reduced. Note however that the latest works have reached high energy
scales $Q\approx 5$ GeV. At these high energies the running of the coupling
is really slow, and a factor two in scales only reduces the truncation
effects in the determination of $\Lambda$ by a factor 2. But more computer
power would allow to check the compromises done in the most recent
computations thoroughly, and reach solid determinations at the
percent level.
For the determinations based on heavy quark correlators, we also expect a
reduction of the perturbative truncation errors by a factor of two. Note
that in this case the $\beta$-function is only known to 3-loops (with
truncation errors formally $\mathcal O(\alpha^2)$), but the scale typically
used in the extractions is significantly lower (we use
$Q \approx \bar m_c$). At these low scales the strong coupling runs
faster. Using our estimates for the perturbative truncation effects, these
methods could reach a precision $\sim 1\%$ with a solid and
conservative estimate of the truncation uncertainties.
Determinations based on observables defined at the cutoff scale have a
more uncertain future. Truncation uncertainties estimated using the
scale variation approach point to a not very well behaved
perturbative series for these observables.
The main benefit would come from the
possibility of reaching the perturbative domain \emph{without} having to add
any additional terms to the known perturbative expansion ({\em i.e.}\,
reaching comfortably the perturbative region).
\item Determinations based on finite size scaling are not limited by the
systematic error associated with the continuum extrapolation or the
truncation of the perturbative series, but by statistics. An increase in
computer power by a factor $2^7$ would help in reducing the statistical
uncertainty dramatically. Naively one expects that the non-perturbative
data would become an order of magnitude more precise~\footnote{As with all
Monte Carlo methods, the statistical uncertainties of a lattice
computation decrease $\propto1/\sqrt{N}$ where $N$ is the number of
configurations. The computer requirements are proportional to $N$.},
decreasing the uncertainty to the level of $0.07\%$.
This precision clearly is exagerated. Such level of precision would require
finer lattices to control the continuum extrapolations. An increase
in computer power by $2^7$ seems impressive, but this will be eaten by
just simulating values of the lattice spacing two times smaller.
Some compromise between reducing the statistical errors, and
improving the continuum extrapolations would be needed.
Moreover some ingredients of
the determination ({\em i.e.}\, the scale setting), that nowadays represent a very
small contribution to the total uncertainty, would need to be computed on
large volumes where the uncertainties are not merely statistical.
Nevertheless, it is clear that such an increase in computer power would
allow to reduce the uncertainty in $\Lambda$ by a factor two to three.
A determination of the strong coupling with an uncertainty $\lesssim 0.3 \%$
would in principle be possible.
Of course at such level of precision one would have to think about
other problems, like electromagnetic contributions to the scale and
the running.
\end{enumerate}
Beyond the improvement coming from the increase in computational power, those
determinations that are limited by the truncation of the perturbative series
could benefit from a better perturbative knowledge. This is especially true
for the case of the moments of heavy quark correlators and the observables
defined at the cutoff scale, since the relation to the $\overline{\rm MS} $
scheme is only known to 2-loops, and increasing this to 3-loops seems
feasible\footnote{Note however that for the case of observables defined at the
cutoff scale, this requires a perturbative computation in \emph{Lattice
perturbation theory}, that is substantially more complicated.}. This would
naively suppress the perturbative truncation effects by an extra power of
$\alpha$, so potentially a reduction of the error by a factor $3-4$ could be
achieved \emph{if the size of the perturbative coefficients does not
increase}. However this assumption is not completely innocent. One should
never forget the asymptotic nature of the perturbative series: \emph{higher
order is never as good as higher renormalization scales}.
All in all, experience shows that progress in Lattice QCD comes always from
two fronts: computer power, and the new methods and better understanding of
the problems by the community. We can speculate the improvements that
more computer power would bring, but the most interesting and potentially
better improvements coming from the new methods are much more difficult to
predict.
{On this front there is a recent proposal~\cite{DallaBrida:2019mqg}
that allows to determine the $N_{\rm f}$-flavor $\Lambda$-parameter
from the pure gauge one.
The connection between QCD and the pure gauge theory is based on
non-perturbatively decoupling hevy quarks (see
section~\ref{sec:deco-heavy-quark}) using lattice simulations.
Since the RG equations are solved in the pure gauge theory
(computationally easier), it allows to reach higher energy scales and
improve the statistical precision.
This idea is discussed in detail in the recent review~\cite{DallaBrida:2020pag}.
This proposal is promosing, but at the time of writing this review
there are no results for $\alpha_s$ using this method yet.}
\subsubsection{The pure gauge theory as a perfect laboratory}
We have discussed that the limitations usually found in the determination of
the strong coupling are not the same as those of other state of the art
lattice QCD computations. As a consequence of the window problem, the
truncation of the perturbative series and the large scaling violations present
in short distance observables represent the limitation of all lattice methods
to determine the strong coupling, with the exception of finite size
scaling. Statistical accuracy is the main limitation of finite size scaling.
The determination of the $\Lambda$ parameter in the pure gauge theory faces
\emph{exactly} the same limitations. On the other hand, computationally the
pure gauge theory is much more tractable than QCD. Simulations algorithms for
the pure gauge theory are much more efficient, and smaller physical volumes
are acceptable, since finite volume effects are suppressed by the larger mass
of the glueballs.
We do not need to wait 10 years to simulate the pure gauge theory at a lattice
spacing of $a=0.02$ fm. Testing what precision in the strong coupling could
potentially be achieved with 100 times more computer time can be done just by
determining the $\Lambda$ parameter in pure gauge.
In fact the situation in pure gauge is not that clear, with some recent
determinations claiming an uncertainty $\approx 1.5\%$ in the $\Lambda$
parameter, but disagreeing by about a $5\%$~\footnote{The
determinations~\cite{Gockeler:2005rv,Asakawa:2015vta,Kitazawa:2016dsl},
based on observables defined at the cutoff scale point to a value
$r_0\Lambda\approx 0.620$, while the recent
determination~\cite{DallaBrida:2019wur} based on finite size scaling points
to a larger value $\approx 0.660$.}. A very detailed and precise study is
been performed using the static potential~\cite{Husung:2017qjz,sommer:19lat},
but at the moment the preliminary results are inconclusive. There is not a
single pure gauge determination using heavy quark correlators.
Given the similarities in the challenges, we hope that the lattice community
takes this discrepancy seriously and does not neglect the issue due to the
lack of phenomenological interest. Pure gauge studies do not enter in the
world average, but in our opinion they will be crucial to help us
understand the problems that we have to face and how to solve them.
\subsection{QCD and the scale of the strong interactions}
\label{eq:QCDScale}
The strong interactions are described by a relativistic quantum field theory. It
describes the interactions between color charged particles: the 6 quark flavors
and the gluons. It is a non-abelian gauge theory with symmetry group
$SU(3)$.
Matter content and symmetries is all that is needed to write down the action of
QCD, that reads~\footnote{We are going to work in 4-dimensional Euclidean space.
The gauge field $A_\mu(x)$ lives in the Lie algebra $\mathfrak{su}(3)$, and
therefore, for matter in the fundamental representation of the gauge group, it is an anti-hermitian
traceless $3\times 3$ matrix.}
\begin{equation}
\label{eq:QCDAction}
S[A] =\int {\rm d}^4x\, \left\{
-\frac{1}{2g^2} {\rm Tr}\left( F_{\mu\nu} F_{\mu\nu}\right) +
\sum_{i=1}^6\bar \psi_i(\gamma_\mu D_\mu + m_i)\psi_i
\right\}\,,
\end{equation}
where $D_\mu = \partial_\mu + A_\mu$, $m_i$ is the bare mass of quark
flavor $i$, and $g$ is the bare gauge coupling. The field strength is
defined by
\begin{equation}
\label{eq:FieldStrengthDef}
F_{\mu\nu} = \partial_\mu A_\nu - \partial_\nu A_\mu + [A_\mu,A_\nu]
\end{equation}
It is worth noting that quark masses are the only dimensionful
parameters of the QCD action, since the gauge coupling $g$ is
dimensionless in 4 dimensions. At the classical level quark masses are
the only source of breaking of scale invariance.
QCD predictions are made by computing expectations values
of fields in the Euclidean theory as path integral averages with
partition function
\begin{equation}
\label{eq:PathIntegralDef}
\mathcal Z = \int \mathcal{D}A\, e^{-S[A]}\,.
\end{equation}
All physical information is then extracted from these correlators. The
path integral written above is, naively, ill defined. A simple
perturbative calculation for instance shows that the path integral is
plagued by ultraviolet (UV) divergences, {\em i.e.}\ divergences that arise
when summing over the high-energy modes in the
theory. Expectation values can be made finite by modifying
the theory at short distances. There are several possibilities for such a
\emph{regularization} of the theory, the most natural consists in
defining the theory on a four dimensional Euclidean lattice with
spacing $a$. When performing Fourier transforms in a discretized
spacetime, momenta are limited to the first Brillouin zone, which
implies that the inverse lattice spacing provides a UV
cutoff. There are other possibilities to make expectation
values finite, like defining the theory in an arbitrary number of
dimensions (dimensional regularization), that are more convenient in the
context of perturbative computations.
Independently of the details of the regularization procedure, any
physical quantity $P(Q)$, measured at a typical scale $Q$, computed
from some expectation value in the regularized theory, will depend not
only on $Q$ and the particular values of the gauge coupling and quark
masses ($g,m_i$), but also on the short distance scale (denoted $a$)
at which QCD is modified. Denoting the mass dimension of $P$ by $d_P$,
we have:
\begin{equation}
\label{eq:PhysRegulatedQuantity}
a^{d_P} P(Q) = {\mathcal P}(aQ,g,am_i)\,.
\end{equation}
Note that the quantity on the left-hand side of
Eq.~\eqref{eq:PhysRegulatedQuantity} is the dimensionless product
$a^{d_P} P(Q)$, and that accordingly the function ${\mathcal P}$ only
depends on dimensionless quantities. The problem is how to make any
solid prediction when the arbitrary value of the short distance $a$
appears in all determinations of physical quantities. The answer comes
under the name of \emph{renormalization}. Even if determinations in
the regularized theory depend on the particular choice of ultraviolet
cutoff ($a$), the physics at large distances compared with the cutoff
(the regime $aQ\ll 1$) is universal if it is parametrized in terms of
the \emph{renormalized} coupling ($\bar g(\mu)$) and
\emph{renormalized} quark masses ($\bar m_i(\mu)$). The
renormalization scale $\mu$ is an arbitrary scale that is introduced
in the renormalization procedure. A more precise relation would then
read
\begin{equation}
\label{eq:PhysRenormQuantity}
\mathcal P(aQ,g,am_i)
= \bar{\mathcal P}(Q/\mu,\bar g(\mu),\bar m_i(\mu)/\mu) +
\mathcal O\left((aQ)^p, (a\mu)^p, (am)^p\right) + \dots\, .
\end{equation}
Note that the arbitrary scale $a$ does not show up in the first term
on the right-hand side. Moreover in the limit where the short-distance
scale $a$ is much smaller than the physical ($Q$) and renormalization
($\mu$) scales a \emph{precise} prediction for any physical observable
emerges
\begin{equation}
\label{eq:PhysRenormLimit}
\frac{P(Q)}{M^{d_P}} = \frac{1}{(aM)^{d_P}} \,
\bar{\mathcal P}(Q/\mu,\bar g(\mu),\bar m_i(\mu)/\mu) \,.
\end{equation}
In the equation above we have expressed $P(Q)$ in units of some
physical mass scale $M$, which in turn can be obtained from a lattice
simulation in units of the cutoff $a$ -- this the quantity in the
denominator in the RHS of the expression above.
The renormalized quantities $\bar g(\mu), \bar m_i(\mu)$ are functions
of the quark masses and coupling constant of the finite theory (the
bare parameters $g,m_i$), the cut-off $a$ and the renormalization
scale $\mu$. The physics content of this renormalization process is
that at low energies the theory is sensitive to the particular choice
of cutoff \emph{only} via the relation between bare and renormalized
parameters. This relation is not observable and remains an arbitrary
choice needed in order to make physical predictions. The set of
prescriptions that are necessary to fully specify the relation between
bare and renormalized quantities is called a {\em renormalization
scheme}.
Note that in the renormalization procedure, we have introduced a new
scale $\mu$. This is not an accident, and is unavoidable,
independently of the chosen regularization and/or renormalization
schemes. The renormalization scale $\mu$ is arbitrary and physical
quantities must be independent on $\mu$. This requirement can be
expressed as a set of mathematical conditions, which go under the name of
Callan-Symanzik~\cite{Callan:1970yg,Symanzik:1970rt} equations:
\begin{equation}
\label{eq:CallanSymanzikObs}
\mu \frac{{\rm d}}{{\rm d}\mu}
\bar{\mathcal P}(Q/\mu,\bar g(\mu),\bar m_i(\mu)/\mu) = 0\, .
\end{equation}
These equations can be used to determine how the renormalized coupling $\bar
g(\mu)$ and the renormalized quark masses $\bar m_i(\mu)$ change (``run'') with
the renormalization scale. One of the main characters of this review is the
$\beta$-function, which dictates the dependence of the renormalized coupling on
the renormalization scale~\cite{Politzer:1973fx, Gross:1973id}\footnote{In this
section we will use massless renormalization schemes, where the $\beta$ function
is independent of the values of quark masses. See
section~\ref{sec:deco-heavy-quark} for a discussion of massive renormalization
schemes.}
\begin{equation}
\label{eq:beta}
\mu\frac{{\rm d}}{{\rm d}\mu} \bar g(\mu) = \beta(\bar g)\,.
\end{equation}
This renormalization group (RG) equation is a first order equation,
and therefore its solution depends on exactly one integration
constant. Moreover the solution to this equation has to respect the
correct boundary condition given by the asymptotic behavior of the
$\beta$-function determined in perturbation theory~\footnote{In
general perturbative expansions in quantum field theories are
asymptotic. Through this work a function $f(x)$ having an asymptotic
expansion will be denoted by $f(x) \simas{x\to 0}$ \dots}
\begin{equation}
\label{eq:BetaFunAsymp}
\beta(\bar g) \simas{\bar g\to 0} -\bar g^3 \sum_{k=0}^\infty b_k
\bar g^{2k}\, ,
\end{equation}
where
\begin{subequations}
\label{eq:beta_univ}
\begin{eqnarray}
b_0 &= \frac{1}{(4\pi)^2}\left(11-\frac{2N_f}{3} \right) \, ,\\
b_1 &= \frac{1}{(4\pi)^4}\left(102-\frac{38N_f}{3} \right) \, ,
\end{eqnarray}
\end{subequations}
and $N_f$ is the number of fermions in the fundamental representation ({\em i.e.}\
quarks). Note that for $\bar g \to 0$ the $\beta$-function is
\emph{negative} (at least for $N_f<17$), which implies asymptotic
freedom, {\em i.e.}\ the decrease of the coupling with increasing energy.
It is instructive to discuss in some detail the integration of the RG
equation, Eq.~\eqref{eq:beta}. We can readily see that
\begin{equation}
\label{eq:IntStepOne}
\frac{d\mu}{\mu} = \frac{d\bar{g}}{\beta(\bar{g})}
\quad \Longrightarrow \quad
\log\left(\frac{\mu_1}{\mu_2}\right) =
\int_{\bar{g}_2}^{\bar{g}_1} \frac{dx}{\beta(x)}\, ,
\end{equation}
where $\bar{g}_1=\bar{g}(\mu_1)$, $\bar{g}_2=\bar{g}(\mu_2)$. The
logarithmic divergence on the left-hand side of
Eq.~\eqref{eq:IntStepOne} when $\mu_1$ (resp. $\mu_2$) tends to
infinity is reflected in the divergence of the integral on the
right-hand side when $\bar{g}_1$ (resp. $\bar{g}_2$) tend to zero. The
asymptotic behaviour of the integrand is
\begin{align}
\frac{1}{\beta(x)}
&= -\frac{1}{b_0 x^3} \frac{1}{1 + \frac{b_1}{b_0} x^2 + O(x^4)} \\
& \underset{x\to 0}{\simeq} -\frac{1}{b_0 x^3}
\left[
1 - \frac{b_1}{b_0} x^2 + O(x^4)
\right]\, ,
\end{align}
and therefore the integral can be rewritten as
\begin{align}
\int_{\bar{g}_2}^{\bar{g}_1} \frac{dx}{\beta(x)}
&= \left[
\frac{1}{2b_0\bar{g}_1^2} - \frac{1}{2b_0\bar{g}_2^2}
+ \frac{b_1}{b_0^2} \log \bar{g}_1 -
\frac{b_1}{b_0^2} \log \bar{g}_2
\right] + \nonumber \\
& \quad + \int_{\bar{g}_2}^{\bar{g}_1}{\rm d}x\, \left[\frac{1}{\beta(x)} +
\frac{1}{b_0x^3} - \frac{b_1}{b_0^2x}\right]\, .
\end{align}
Note that, when rewritten in this form, the integrand that appears on
the right-hand side is $\mathcal O(x)$ when $x\to 0$, and hence the integral is
finite when the integration limit tends to zero. After some algebraic manipulations
Eq.~\eqref{eq:IntStepOne} yields
\begin{align}
&\mu_1
\left[b_0\bar{g}_1^2\right]^{-\frac{b_1}{2b_0^2}}\,
e^{-\frac{1}{2b_0\bar{g}_1^2}}\,
\exp\left\{-
\int_{0}^{\bar{g}_1}{\rm d}x\, \left[\frac{1}{\beta(x)} +
\frac{1}{b_0x^3} - \frac{b_1}{b_0^2x}\right]\right\} = \nonumber
\\
&\quad \mu_2
\left[b_0\bar{g}_2^2\right]^{-\frac{b_1}{2b_0^2}}\,
e^{-\frac{1}{2b_0\bar{g}_2^2}}\,
\exp\left\{-
\int_{0}^{\bar{g}_2}{\rm d}x\, \left[\frac{1}{\beta(x)} +
\frac{1}{b_0x^3} - \frac{b_1}{b_0^2x}\right]\right\} \, .
\label{eq:IntConst}
\end{align}
The equality holds for any value of $\mu_1$ and $\mu_2$, showing that
the combination in Eq.~\eqref{eq:IntConst} has units of mass, and is
independent of $\mu$. It is called the $\Lambda$-parameter and can be
understood as the \emph{intrinsic scale} of QCD that we were looking
for. Note that the integration of the renormalization group equation
Eq.~\eqref{eq:beta} is exact, and the $\Lambda$-parameter can be
defined as:
\begin{equation}
\label{eq:lam}
\Lambda = \mu
\left[b_0\bar g^2(\mu)\right]^{-\frac{b_1}{2b_0^2}}\,
e^{-\frac{1}{2b_0\bar g^2(\mu)}}\,
\exp\left\{-
\int_{0}^{\bar g(\mu)}{\rm d}x\, \left[\frac{1}{\beta(x)} +
\frac{1}{b_0x^3} - \frac{b_1}{b_0^2x}\right]\right\}\,.
\end{equation}
This expression is valid beyond perturbation theory. Hadron masses, meson decay
constants, or any other dimensionful quantity in QCD, can be measured in units
of $\Lambda$, and are given by dimensionless functions of the renormalized
coupling, and the renormalized quark masses (also expressed in units of
$\Lambda$). It is in this respect that we like to think of the $\Lambda$
parameter as an {\em intrinsic scale}\ of QCD.
The renormalized theory is defined by specifying the value of
the renormalized coupling at a given scale, or equivalently by
specifying the value of the $\Lambda$ parameter. Note that
Eq.~\eqref{eq:lam} is an implicit equation for $\bar g (\mu)$,
and therefore the running coupling is a function of $\Lambda/\mu$; at
high energies compared with $\Lambda$ (i.e. $\mu/\Lambda\gg 1$) the
running of the coupling is given by
\begin{equation}
\label{eq:RGEHighEnergy}
\bar g^2(\mu) \simas{\mu/\Lambda\gg 1}
\frac{1}{2 b_0 \log(\mu/\Lambda) + b_1 \log\log(\mu/\Lambda)}
+ \dots\,.
\end{equation}
At scales \emph{much larger than} $\Lambda$, $\bar g(\mu)$ is small,
QCD is weakly coupled and quarks behave as almost free particles.
\subsection{The determination of the intrinsic scale of QCD}
There is quite some freedom when renormalizing QCD. In the framework of
perturbative computations there are many valid ways to subtract the divergent
parts of Feynman diagrams that differ by finite terms. Non-perturbatively there
are also multiple conditions to use as a definition for renormalized coupling
and quark masses. This freedom is called \emph{choice of scheme}.
The value of the strong coupling constant at high energies is a
necessary input for the study of all QCD cross sections at the Large
Hadron Collider (LHC) and many other high-energy experiments. For this
reason it is convenient to quote its value in a scheme that can be
easily used for phenomenological input. The so-called \emph{modified
minimal subtraction} ($\overline{\rm MS}$)
scheme~\cite{Bardeen:1978yd} is by far the most widely-used
choice. This scheme is defined in the context of perturbative
computations; however the $\Lambda$-parameter extracted in this
convenient scheme still has a non-perturbative meaning, as discussed
below.
\subsubsection{Scheme dependence}
Most of the time we are going to deal with \emph{mass-independent}
renormalization schemes. Any modification of the theory that is performed in
order to regularize and renormalize QCD can always be made at energies much
larger than the quark masses.\footnote{In some cases massive
renormalization schemes might be more convenient, like for example
heavy quarks regulated on the lattice. In practice the lattice spacings that are
currently accessible to simulations provide a UV cutoff that is not much larger
than the mass of the heavy quarks $c$ and $b$. In this context mass-dependent
renormalization schemes might have some advantageous properties. See
{\em e.g.}~\cite{Boyle:2016wis,Fritzsch:2018kjg}.} From a perturbative point of view
we can say that the UV divergences of Feynman diagrams are independent of any
quark mass. A nice property of mass-independent schemes is that the RG functions
(like the $\beta$-function Eq.~\eqref{eq:beta}) are independent of the quark
masses. Of course there is nothing fundamentally wrong with renormalization
schemes that are not mass-independent, and we will study in detail the
relation between mass-independent and mass-dependent renormalization schemes in
chapter~\ref{sec:decoupling}, but first let us state some basic relations
between massless renormalization schemes.
By convention we normalize couplings in different schemes so that they
agree to leading order. This implies that renormalized couplings in
two schemes $s$ and $s'$ are related perturbatively by
\begin{equation}
\label{eq:gssp}
\bar g^2_{s'}(\mu)\, \simas{\bar g_s\to 0}\, \bar g^2_s(\mu) + c_{ss'}\bar g^4_s(\mu) +\dots
\end{equation}
with $c_{ss'}$ a finite number. The $\beta$-function for the couplings $\bar
g_{s}(\mu)$ and $\bar g_{s'}(\mu)$ are different (i.e. $\beta$-functions are
scheme dependent), but it is easy to check that the two leading terms in its
asymptotic expansion~\eqref{eq:beta_univ} are scheme independent: $b_0$ and
$b_1$ are universal. Higher order coefficients $b_n$ with $n>1$ are scheme
dependent. For the case of the $\overline{\rm MS}$ scheme, the $\beta$-function
is known up to five
loops~\cite{vanRitbergen:1997va,Czakon:2004bu,Baikov:2016tgj,Luthe:2016ima,Herzog:2017ohr}
(see table~\ref{tab:bnms}).
\input{tab_bnms.tex}
On the other hand the $\Lambda$-parameter, defined in Eq.~\eqref{eq:lam}, is
also scheme dependent. It is easy to see by using the one-loop relation between
couplings, Eq.~\eqref{eq:gssp}, that
\begin{subequations}
\begin{eqnarray}
\left[\frac{\bar g_{s'}^2(\mu)}{\bar g_s^2(\mu)}\right]^{-\frac{b_1}{2b_0}}
&\, \simas{\bar g_s\to 0}\, &
1 - \frac{b_1}{2b_0}c_{ss'}\bar g_{s}^2(\mu) + \dots\,, \\
\frac{1}{\bar g_{s'}^2(\mu)} - \frac{1}{\bar g_s^2(\mu)}
&\, \simas{\bar g_s\to 0}\, &
c_{ss'} + \dots\,.
\end{eqnarray}
\end{subequations}
Since the integral in Eq.~\eqref{eq:lam} is $\mathcal O(\bar
g^2)$, one can obtain an exact relation between $\Lambda$-parameters
by taking the limit $\bar g_s\to 0$~\cite{Hasenfratz:1980kn}
\begin{equation}
\frac{\Lambda_{s'}}{\Lambda_s} =
\exp\left(\frac{-c_{ss'}}{2b_0}\right)\,.
\end{equation}
In other words, the relation of $\Lambda$-parameters in different
schemes is \emph{exactly} known via a one-loop computation, as
reported in Eq.~\eqref{eq:gssp}. This observation, together with
Eq.~\eqref{eq:lam} allows a precise non-perturbative definition of the
$\Lambda$-parameter even for schemes that are intrinsically defined in
a perturbative context: even if $\overline{\rm MS}$ is a
``perturbative scheme'', $\Lambda_{\overline{\rm MS}}$ is a meaningful
quantity beyond perturbation theory.
\subsubsection{Quark thresholds}
Quarks are not massless particles. Every physical process in QCD
depends not only on the intrinsic scale of the strong interactions,
but on the quark masses. In particular we expect that if the process
takes place at a some energy scale much lower than the mass of some
quark, this quark should ``decouple'' from all physical processes.
How this decoupling takes place in mass-independent renormalization
schemes is in fact not trivial. The RG functions (and therefore the
renormalized parameters of the theory) do not depend on any
quark mass: at any scale, the top and the up quarks give the very same
contribution to the $\beta$-function and therefore to the running
coupling. Still, physical observables written in terms of these
renormalized parameters should ``know'' when the energy scale of the
physical process is large or small compared with some quark masses.
We will return to these problems in detail in section~\ref{sec:decoupling}, here
it is sufficient to mention that decoupling in mass-independent renormalization
schemes can be understood as a matching between different theories. At energy
scales below the top quark mass, it is more convenient to use an effective
5-flavor QCD theory, without the top quark. The effects of the top quark at low
energies can be conveniently reabsorbed in a redefinition of the coupling and
quark masses of the 5-flavor theory (these can be computed perturbatively), with
further corrections being power suppressed $\sim\mathcal O\left((\Lambda/m_{\rm
t})^2\right), \mathcal O\left((Q/m_{\rm
t})^2\right)$.
In a similar way, at energies much below the bottom (respectively charm)
quark mass threshold, 4- (respectively 3-) flavor QCD is an excellent
description of nature. Each theory has its own set of fundamental
parameters, so {\em e.g.}\ the 4-flavor theory is completely defined by the
values of the 4-flavor coupling constant and of the quark
masses. These effective theories can be used to describe any physical
process at energy scales much below the corresponding thresholds
$m_{\rm b}\sim 4\, {\rm GeV}$ and $m_{\rm c}\sim 1.4\, {\rm GeV}$.
Following the notation in Ref.~\cite{Chetyrkin:2005ia}, the coupling in the
effective theory with $N_{\rm f} -1$ active flavors is related to the coupling
in the fundamental theory with $N_{\rm f} $ active flavors by the relation
\begin{equation}
\label{eq:gdec}
\bar g^2_{\overline{\rm MS}, N_{\rm f}-1 }(m^\star) = \bar g^2_{\overline{\rm
MS},N_{\rm f}}(m^\star) \times \xi(\bar g^2_{\overline{\rm MS},N_{\rm f}}(m^\star))\,.
\end{equation}
This expression neglects power corrections in the matching between
theories, and $\xi(x)$ is a just a polynomial.
In Eq.~(\ref{eq:gdec}) $m^\star = m_{\overline{\rm MS} }(m^\star)$ is
the $\overline{\rm MS} $ mass at its own scale. In this case the
one-loop term vanish and we have
\begin{equation}
\label{eq:xi}
\xi(\bar g ) = 1 + c_1 \bar g ^4 + c_2 \bar g ^6 + c_3 \bar g ^8 + \mathcal O(\bar g ^{10})\,.
\end{equation}
with the first 4 terms
known~\cite{Chetyrkin:2005ia,Weinberg:1980wa,Bernreuther:1981sg,Grozin:2011nk,Schroder:2005hy}.
Eq.~(\ref{eq:gdec}) allows to relate the values of the
$\Lambda$ parameters with a different number of flavors.
For example, for processes at energy scales above the top mass
threshold we would need the value of $\Lambda^{(6)}_{\overline{\rm
MS}}$. This can be obtained from $\Lambda^{(5)}_{\overline{\rm
MS}}$ by first determining the value of the five flavor coupling
at the top mass threshold\footnote{We will not discuss any subtleties in the determination
of the top quark mass here.} ($m_{\rm t}^\star \approx 163$
GeV)
\begin{equation}
\bar g _{\rm t,5} = \bar g_{\overline{\rm MS},5 }(m^\star_{\rm t})\,,
\end{equation}
from the implicit equation
\begin{equation}
\frac{\Lambda^{(5)}_{\overline{\rm MS}}}{m^\star_{\rm t}} =
\left[b_0\bar g _{\rm t,5}^2\right]^{-\frac{b_1}{2b_0^2}}\,
e^{-\frac{1}{2b_0\bar g _{\rm t,5}^2}}\,
\exp\left\{-
\int_{0}^{\bar g _{\rm t,5}}{\rm d}x\, \left[\frac{1}{\beta^{(N_{\rm f} =5)}(x)} +
\frac{1}{b_0x^3} - \frac{b_1}{b_0^2x}\right]\right\}\,.
\end{equation}
Now this value of the coupling is transformed into the 6 flavor
coupling by using the decoupling relations Eq.~(\ref{eq:gdec}), to
obtain
\begin{equation}
\bar g^2_{\overline{\rm MS}, 6}(m^\star_{\rm t}) =
\frac{\bar g_{\rm t, 5}^2}{\xi(\bar g_{\rm t, 5})}\,.
\end{equation}
Finally the value of $\bar g _{\rm t, 6} = \bar g _{\overline{\rm MS}, 6}(m^\star_{\rm
t})$ can be used to determine the value of the 6-flavor $\Lambda$
parameter
\begin{equation}
\Lambda^{(6)}_{\overline{\rm MS} } = m^\star_{\rm t}
\left[b_0\bar g _{\rm t, 6}^2\right]^{-\frac{b_1}{2b_0^2}}\,
e^{-\frac{1}{2b_0\bar g _{\rm t, 6}^2}}\,
\exp\left\{-
\int_{0}^{\bar g _{\rm t, 6}^2}{\rm d}x\, \left[\frac{1}{\beta^{(N_{\rm f} =6)}(x)} +
\frac{1}{b_0x^3} - \frac{b_1}{b_0^2x}\right]\right\}\,.
\end{equation}
All this procedure can be summarized by defining
\begin{equation}
\varphi_{(N_{\rm f}) }(t) =
\left[b_0t^2\right]^{-\frac{b_1}{2b_0^2}}\,
e^{-\frac{1}{2b_0t^2}}\,
\exp\left\{-
\int_{0}^{t}{\rm d}x\, \left[\frac{1}{\beta^{(N_{\rm f} )}(x)} +
\frac{1}{b_0x^3} - \frac{b_1}{b_0^2x}\right]\right\}\,.
\end{equation}
and using
\begin{equation}
\label{eq:lamrat}
\frac{\Lambda^{(N_{\rm f} +1)}}{\Lambda^{(N_{\rm f})}} =
\frac{\varphi_{(N_{\rm f} +1)}\left( \frac{\bar g ^\star}{\xi(\bar g^\star)} \right) }
{\varphi_{(N_{\rm f})}(\bar g ^\star)}\,.
\end{equation}
Of course such a conversion suffers from several uncertainties.
The expressions for the $\beta$-functions ({\em i.e.}\, $\beta^{(N_{\rm f}
=5,6)}$) and the decoupling relations Eq.~(\ref{eq:gdec}) are only
known to a certain order in perturbation theory.
This implies that the conversion of $\Lambda$-parameters
Eq.~(\ref{eq:lamrat}) carries a perturbative uncertainty.
On top of that there are power corrections that have been neglected in
the matching between the effective $N_{\rm f}$ and fundamental
$N_{\rm f} + 1$ theories (see
section~\ref{sec:deco-heavy-quark}). However, we have now strong
numerical
evidence~\cite{Athenodorou:2018wpk, Korzec:2016eko} showing that both
the perturbative and power corrections are very small in the ratio
Eq.~(\ref{eq:lamrat}).
Even for the case of the decoupling of the charm quark (at a rather
low energy scale $m^\star_{\rm c} \approx 1.4$ GeV), these
effects seem to be too small to affect the current determinations of
$\Lambda$. The interested reader is encouraged to read
section~\ref{sec:deco-heavy-quark} and consult the original
reference~\cite{Athenodorou:2018wpk} where these issues are discussed
in detail.
\subsubsection{Challenges in the determination of
$\alpha_{\overline{\rm MS}}(M_Z)$}
\label{sec:chall-determ-alph}
From the previous discussion it seems logical to quote the intrinsic scale of
QCD by giving the value of $\Lambda_{\overline{\rm MS}}^{(5)}$, which is a well
defined quantity, even beyond perturbation theory. Together with
Eq.~\eqref{eq:lam} and the coefficients of the $\beta$-function in the
$\overline{\rm MS}$ scheme reported in table~\eqref{tab:bnms}, it can be used
for high energy phenomenology. Moreover if one is interested in a process at an
energy scale above the top quark threshold (or below the bottom/charm
thresholds), the procedure described in the previous section can be used to
determine the 3,4 or 6 flavor $\Lambda$-parameters.
For historical reasons it is now standard to quote the intrinsic
scale of the strong interactions in an indirect way by referring to
the value of the strong coupling in the $\overline{\rm MS}$ scheme at
a reference scale $\mu=M_Z$ ({\em i.e.}\ the mass of the $Z$ vector boson
$M_Z\approx 91.19\, {\rm GeV}$). The current world
average for
\begin{equation}
\alpha_{\overline{\rm MS} }(\mu) = {\bar
g^2_{\overline{\rm MS}}(\mu)}/{4\pi}
\end{equation}
quoted in the PDG~\cite{pdgtbp:2020} is:
\begin{equation}
\label{eq:alpha_pdg}
\alpha_{\overline{\rm MS} }(M_Z) = 0.1179(10)\, ,
\end{equation}
with an uncertainty of around $\sim 1\%$ obtained by combining the uncertainty
in the determination from several processes. Note that this coupling refers to
the 5 flavor theory since $m_{\rm b} < M_Z < m_{\rm t}$. By using the five-loop
asymptotic expansion of the $\beta$-function, the world average
Eq.~\eqref{eq:alpha_pdg} is equivalent to\footnote{Note that due to the
logarithmic running of the strong coupling a $\sim 6\%$ uncertainty in
$\Lambda^{(5)}_{\overline{\rm MS}}$ translates into an $\sim 1\%$ uncertainty in
$\alpha_{\overline{\rm MS} }(M_Z)$.}
\begin{equation}
\Lambda_{\overline{\rm MS}}^{(5)} = 0.207(12)\, {\rm GeV}\,.
\end{equation}
But, what are the challenges in a precise determination of the strong
coupling? To understand this subtle point, we have to look carefully
at the fundamental equation used to determine $\Lambda$ in units of
some reference scale $\mu_{\rm ref}$:
\begin{align}
\label{eq:lam_again}
\Lambda = \mu_{\rm ref}
\left[b_0\bar g^2(\mu_{\rm ref})\right]^{-\frac{b_1}{2b_0^2}}\,
& e^{-\frac{1}{2b_0\bar g^2(\mu_{\rm ref})}} \times \nonumber \\
& \times \exp\left\{-
\int_{0}^{\bar g(\mu_{\rm ref})}{\rm d}x\, \left[\frac{1}{\beta(x)} +
\frac{1}{b_0x^3} - \frac{b_1}{b_0^2x}\right]\right\}\,.
\end{align}
As already discussed above, this is the solution of a first-order differential
equation, and $\Lambda$ can be understood as an integration constant. In other
words, knowing the value of the coupling at some reference scale is sufficient
to determine $\Lambda$ and hence to fix the coupling value at all energies
according to the RG running. In principle only {\em one}\ number is needed to
fully determine the value of the strong coupling. It could be for instance the
value of $\bar g(\mu_{\rm ref})$. The systematic errors in the determination of
the strong coupling are better understood by discussing the determination of
$\Lambda$. As we can see, \eqref{eq:lam_again} involves the integral of the
$\beta$-function from $0$ to $\bar g(\mu_{\rm ref})$. Since
\begin{equation}
\lim_{\mu\to\infty}\bar g (\mu) = 0\,,
\end{equation}
the lower limit of the integral corresponds to an infinite energy scale.
Therefore the determination of $\Lambda$ requires the knowledge of the
nonperturbative beta function for all energies between $\mu_\mathrm{ref}$ and
infinity. In practice this limit is never reached, neither in experimental
processes nor in lattice simulations.
At most we can
determine the non-perturbative running in a limited range of scales,
say from $\mu_{\rm ref}$ to $\mu_{\rm PT}$.
At energies higher than $\mu_{\rm PT}$ one uses the perturbative
approximation of the non-perturbative $\beta$-function.
If we denote $\beta^{(\ell)}$ the perturbative $\beta$
function to $\ell$-loops, we have that
\begin{align}
\label{eq:PTerrorOne}
\int_{0}^{\bar g(\mu_{\rm ref})}{\rm d}x\, \left[\frac{1}{\beta(x)} -
\frac{1}{\beta^{(\ell)}(x)} \right] =
\mathcal O\left(\left(\bar g^2(\mu_{\rm ref})\right)^{\ell-1}\right)\,.
\end{align}
And therefore this uncertainty propagates to the determination of $\Lambda$:
\begin{align}
\nonumber
\int_{0}^{\bar g(\mu_{\rm ref})}{\rm d}x\, \left[\frac{1}{\beta(x)} +
\frac{1}{b_0x^3} - \frac{b_1}{b_0^2x}\right]
\simas{\bar g (\mu_{\rm PT}) \to 0}
& \int_{\bar g(\mu_{\rm PT})}^{\bar g(\mu_{\rm ref})}{\rm d}x\,
\left[\frac{1}{\beta(x)} +
\frac{1}{b_0x^3} - \frac{b_1}{b_0^2x}\right] \\
\nonumber
& + \int_{0}^{\bar g(\mu_{\rm PT})}{\rm d}x\,
\left[\frac{1}{\beta^{(\ell)}(x)} +
\frac{1}{b_0x^3} - \frac{b_1}{b_0^2x}\right] \\
& + \mathcal O(\bar g^{2\ell -2}(\mu_{\rm PT}))\,.
\label{eq:intpt}
\end{align}
As shown in table~\ref{tab:bnms} the $\beta$-function is known up to five loops
in the $\overline{\rm MS}$ scheme. Note, nevertheless, that in practice one
never reaches this level of accuracy. In order to apply Eq.~(\ref{eq:intpt}) in
the $\overline{\rm MS} $ scheme, the value of the coupling $\bar
g_{\overline{\rm MS} } (\mu_{\rm PT})$ is needed. The latter is determined by
matching an experimental quantity with its asymptotic perturbative expansion,
typically known up to 3-4 loops. In this case the accuracy in the extraction of
$\Lambda_{\overline{\rm MS}}$ will be limited by the limited knowledge in the
perturbative expression of the physical observable, and not by the perturbative
knowledge in $\beta_{\overline{\rm MS} }(\bar g)$.
This phenomenon is present in one form or another in any extraction of the
strong coupling, not only the ones from lattice QCD, but also in
phenomenological extractions. Even if $\Lambda_{\overline{\rm MS} }^{(5)}$ is
defined non-perturbatively, perturbation theory is needed for its determination.
Of course this does not mean that a truly non-perturbative determination of the
$\Lambda$-parameter is impossible. The situation is conceptually very similar to
many other systematic effects present in any lattice determination; for example
lattice calculations are always performed at non-zero lattice spacing (and on a
finite volume) and this does not prevent us to obtain values in the continuum
(and in infinite volume). We need to simulate several lattice spacing (and
several volumes) and \emph{perform an extrapolation}. The situation here is very
similar: the determination of $\Lambda$ has to be understood as an extrapolation
in $\bar g^{2n}(\mu_{\rm PT})$, with perturbation theory as a guide.
The scale $\mu_{\rm PT}$ is usually called the \emph{scale of matching with
perturbation theory}. One crucial point to note is that the size of the missing
terms is $\mathcal O(\bar g^{2(\ell-1)}(\mu_{\rm PT}))$, where $\ell$ is the
number of loops included in the computation of the beta function. In order to
have a significant change in the contribution of the missing terms, the matching
scale $\mu_{\rm PT}$ has to be changed substantially due to the slow logarithmic
running of the strong coupling (cf. Eq.~\eqref{eq:RGEHighEnergy}). These issues
play a central role in the determination of the systematic error presented in
section~\ref{sec:syst-extr-alph}.
\subsection{Gauge fields on the lattice}
A naive discretization of the pure gauge action, obtained by
substituting the derivatives in the continuum action by finite
differences, results in discretization effects that break gauge
invariance. The way to construct lattice actions for gauge theories is
rooted in the geometric interpretation of gauge invariance, and was
first proposed by Wilson~\cite{PhysRevD.10.2445}. Since the gauge
field acts as an affine connection in the continuum theory, its
lattice counterpart is the parallel transporter along the links of
discretized spacetime. Hence the key idea is to work with link
variables
\begin{equation}
\label{eq:LinkVarsDef}
U(x,\mu) = e^{aA_\mu(x)} \,,
\end{equation}
where the pair $(x,\mu)$ uniquely identifies the link that originates
from point $x$ in the positive $\mu$ direction. These link variables
can be seen as a discretization of a continuum Wilson line, the
parallel transporter mentioned above, linking the points $x$ and
$x+a\hat \mu$~\footnote{The four-vector $\hat \mu$ has all components
equal to zero, except the coordinate $\mu$, that has a value 1.}
\begin{equation}
\label{eq:ParallelTransp}
U(x,\mu) = \mathcal P\exp \left\{ a\int_0^1{\rm d}t\,
A_\mu(\gamma(t)) \right\} + \mathcal O(a^2)\,.
\end{equation}
Here $\gamma(t)$ is a path that links the point $x$ with $x+a\hat\mu$,
{\em e.g.}
\begin{equation}
\label{eq:SimplePath}
\gamma(t) = x + at\hat \mu\, , \quad t\in [0,1]\, .
\end{equation}
A product of link variables along a closed loop is called a {\it
Wilson loop}.
\begin{figure}[h!]
\centering
\begin{subfigure}{0.3\textwidth}
\centering
\includegraphics[height=0.2\textheight]{fig/plaquette.pdf}
\caption{$\mathcal S_0$}
\end{subfigure}\hspace{0.2\textwidth}
\begin{subfigure}{0.3\textwidth}
\centering
\includegraphics[height=0.2\textheight]{fig/2x1.pdf}
\caption{$\mathcal S_1$}
\end{subfigure}
\caption{Traces of the plaquette ($\mathcal S_0$) and the $2\times 1$ Wilson
loops ($\mathcal S_1$) can be used to construct different lattice
actions.}
\label{fig:wloops}
\end{figure}
Link variables transform under a gauge transformation $\omega(x)\in SU(3)$ as
\begin{equation}
U(x,\mu) \to \omega(x) U(x,\mu) \omega(x+a\hat \mu)^\dagger\,,
\end{equation}
which implies that the trace of a Wilson loop is gauge invariant. Lattice
actions are constructed by combining these traces, with different choices
yielding the same low-energy physics (compared to the cutoff scale), with
different lattice artefacts. The simplest choice, first proposed by Wilson,
consists in using the smallest possible loop (the plaquette)
\begin{equation}
\label{eq:Sw}
S_W = \frac{1}{g_0^2}\sum_{\mathcal W \in \mathcal S_0}{\rm
tr}\{1-U(\mathcal W)\}\,,
\end{equation}
where the sum runs over all oriented Wilson loops of type $\mathcal S_0$ (see
figure~\ref{fig:wloops}).
It is easy to check that for a given plaquette $\mathcal W_x$ in the $\mu,\nu$
plane with left lower corner at point $x$, we have
\begin{equation}
{\rm Re}\, {\rm tr}\{1-U(\mathcal W_x)\} = \frac{1}{2} a^4{\rm
tr}\{F_{\mu\nu}(x)F_{\mu\nu}(x)\} + \mathcal O(a^5)
\end{equation}
And therefore the Wilson action reduces to the continuum YM action in the naive
limit $a\to 0$.
\paragraph{Improved actions}
This is not the only option. By including in the definition of the lattice
action larger loops the rate of convergence to the continuum can be
improved, {\em i.e.}\ the size of lattice artefacts can be reduced. This is a
general programme that goes under the name of {\em improvement}, and
extends to the fermionic part of the action. Improved actions play an
important role in reducing lattice artefacts and therefore providing more
precise extrapolations to the continuum limit. Although the literature
covers a wide range of lattice actions, most lattice simulations are
performed using some particular choice of the one-parameter family that can
be constructed from the plaquette and the $2\times 1$ loops (see
figure~\ref{fig:wloops})
\begin{equation}
\label{eq:ImprovedGaugeAction}
S_{\rm latt} = \frac{1}{g_0^2}\sum_{i=0,1} c_i \sum_{\mathcal W \in \mathcal S_i}{\rm
tr}\{1-U(\mathcal W)\}\,.
\end{equation}
The constants $c_0,c_1$ have to obey the constraint
\begin{equation}
c_0+8c_1 = 1\, ,
\end{equation}
in order to recover the classical continuum limit, but otherwise can be chosen
at will. Clearly the simple Wilson action, $S_W$ in Eq.~\eqref{eq:Sw}, is
recovered by choosing $c_0=1, c_1=0$. Other popular choices include the Symanzik
tree-level improved action ($c_0=5/3, c_1=-1/12$), or the Iwasaki action
($c_0=3.648, c_1=-0.331$). A detailed discussion of the improvement of
lattice actions is beyond the scope of this review. However the reader
should keep in mind that there are discretization effects when
reading the lattice literature. These discretization effects induce a
systematic error in the lattice observables, which in turn impacts on
the determination of the strong coupling. We shall return to this
issue later in this paper, when discussing the different lattice
extractions of $\alpha$.
For the path integral to be fully specified, it is also necessary to
define the integration measure of the gauge link variables
$U_\mu(x)$. In order to ensure the gauge invariance of the path
integral, the measure ${\rm d} U$ of each link variable needs to be
invariant under both left and right multiplication by elements of the
group:
\begin{equation}
\label{eq:MeasureInv}
{\mathrm d} U = \mathrm{d}(g U) = \mathrm{d}(U g)\, ,
\end{equation}
where $g$ is a generic element of the gauge group. Imposing the
normalization condition
\begin{equation}
\label{eq:MeasureNorm}
\int dU = 1\, ,
\end{equation}
the integration measure is uniquely defined to be the Haar measure on
the group. The reader interested in more details can consult any
standard reference on compact topological groups. When integrating
over all the lattice link variables, we will use the shorthand
notation
\begin{equation}
{\rm d} U = \prod_{x,\mu} {\rm d} U(x,\mu) \,.
\end{equation}
\subsection{Fermions on the lattice}
\subsubsection{Fermionic Path Integral}
\label{sec:grassmann-variables}
Fermions in the functional integral language are represented by
Grassmann (anti-commuting) variables
\begin{equation}
\{\psi_i,\psi_j\} = \{\psi_i,\bar \psi_j\} = \{\bar \psi_i,\bar
\psi_j\} = 0\,.
\end{equation}
Obviously $\psi_i^2 = \bar \psi_i^2 = 0$, so that any function of
Grassmann variables is defined by its Taylor expansion up to second
order. Integrals over Grassmann variables are defined by
\begin{equation}
\begin{split}
\int\, {\rm d} \psi = \int\, {\rm d} \bar \psi = 0\,, \\
\int\, {\rm d} \psi \psi = \int\, {\rm d} \bar \psi \bar\psi = 1\,, \\
\end{split}
\end{equation}
Computing integrals of a function of Grassmann variables
$f(\psi,\bar\psi)$ is therefore a problem in combinatorics.
When computing the integral over several Grassmann variables we will use the
shorthand notation
\begin{equation}
{\rm d} \psi = \prod_i {\rm d} \psi_i\, , \qquad
{\rm d} \bar \psi = \prod_i {\rm d} \bar \psi_i\,.
\end{equation}
A key role is played by the integrals
\begin{eqnarray}
\label{eq:grassmann_int}
\int {\rm d} \psi{\rm d} \bar \psi\,
e^{-\bar \psi_i M_{ij}\psi_j} &=& (-1)^{\frac{n(n-1)}{2}}\det M\,,\\
\nonumber
\int {\rm d} \psi{\rm d} \bar \psi\,
\psi_{k_1}\psi_{k_2}\cdots\bar\psi_{l_1}\bar\psi_{l_2}\cdots
e^{-\bar \psi_i M_{ij}\psi_j} &=&
\sum_{P}(-1)^{\sigma_P}
(M^{-1})_{k_{a_1} l_{b_1}}\cdots(M^{-1})_{k_{a_2} l_{b_2}}
\end{eqnarray}
with the sum is over all permutation of the $k,l$ indices.
As discussed above, matter fields on the lattice are associated to sites,
denoted by the suffices in the equations above. The fermionic action is
quadratic in the fermion fields, and different discretizations can be cast into
different choices for the matrix $M$.
A discretized version of the derivative
\begin{eqnarray}
\partial_\mu f(x) = \frac{f(x+a\hat\mu) - f(x)}{a}\,,\\
\partial_\mu^* f(x) = \frac{f(x) - f(x-a\hat\mu)}{a}\,,
\end{eqnarray}
enters in the non-diagonal elements of this matrix $M$. Moreover, for fermions
minimally coupled to the gauge field, the discrtized derivative needs to be
replaced by its covariant version. Hence in QCD the matrix $M$ depends on the gauge field
configuration $U_{\mu}(x)$ and the mass of the fermions $m_\mathrm{f} $. We will
use the notation
\begin{equation}
\label{eq:LatDiracOp}
D_\mathrm{f} = D\left[U, m_\mathrm{f} \right]\, ,
\end{equation}
to denote the lattice Dirac operator for one fermion species, the latter being
identified by the index $\mathrm{f}$.
Although the simulation of Grassmann variables on a computer is
possible, it is computationally very inefficient. Instead, the
previous relation is used to directly define the path integral of
lattice QCD as (cf.
Eq.~(\ref{eq:grassmann_int}))
\begin{equation}
\mathcal Z = \int {\rm d} U\,
\, \left[\prod_{\mathrm{f}=1}^{N_{\rm f}}\det(D_{\rm f})\right]
e^{-S_G[U]}\,.
\end{equation}
Note that by integrating out the fermions fields exactly, one is
effectively simulating a non-local theory.
\subsubsection{Chiral symmetry and lattice fermions}
The Euclidean action for a single free fermion in the continuum reads
\begin{equation}
S_F[\psi,\bar\psi] =
\int {\rm d}^4x\,
\bar\psi (\gamma_\mu\partial_\mu + m)\psi\,.
\end{equation}
A naive attempt to discretize this action leads to the so-called \emph{doubling
problem:} instead of describing a single fermion, the lattice action describes
$2^4$ fermion flavors. In fact this phenomenon is intimately related with chiral
symmetry. In the absence of a mass term the fermion action is invariant under
chiral transformations
\begin{equation}
\psi \mapsto e^{\imath \theta \gamma_5}\psi\, , \quad
\bar{\psi} \mapsto \bar{\psi} e^{\imath \theta \gamma_5}\, .
\end{equation}
This is just a consequence of the kernel of the fermion bilinear being
proportional to $\gamma_\mu$. The Nielsen-Ninomiya
theorem~\cite{Nielsen:1981hk,Karsten:1981gd,Pelissetto:1987ad} shows that any
local lattice hermitian action that preserves translational invariance and
chiral symmetry describes an equal number of positive- and negative-chirality
fermions. In the case of the naive fermion action, the Nielsen-Ninomiya theorem
is satisfied because of the 16 fermions, 8 have positive chirality and the
remaining 8 have negative chirality. It is possible to reduce the number of
\emph{doublers} from 15 to just 1, but in order to describe a single fermion one
has to break one of the hypotheses of the Nielsen-Ninomiya theorem. Giving up
locality leads to serious difficulties for the renormalization of the theory.
Therefore most efforts have focused on four particular approaches.
\begin{description}
\item[Wilson fermions] One of the most popular choices follows
Wilson's original proposal to break chiral symmetry at finite
lattice spacing by an irrelevant operator~\cite{Wilson:1975id}. This
is implemented by adding a dimension 5 term to the Lagrangian that
is just a suitable discretization of
\begin{equation}
\label{eq:wterm}
\mathcal L_5 = \frac{r}{2} \bar \psi \partial^2 \psi\, .
\end{equation}
The addition of this irrelevant operator has nevertheless an
important impact in the spectrum of the theory by removing all the
doublers.
There are two {\em unpleasant}\ effects of this extra term in the
action. Firstly the fermion mass is no longer protected by chiral
symmetry in the regularised theory, and therefore acquires an
additive renormalization. As a consequence the massless theory can
only be obtained by fine tuning the bare mass in the
action. Secondly the scaling violations are linear in the lattice
spacing $\mathcal O(a)$. The massless theory can be
\emph{non-perturbatively improved} by adding the so called
Sheikholeslami-Wohlert term~\cite{Sheikholeslami:1985ij}
\begin{equation}
\label{eq:csw}
\mathcal L'_5 = c_{\rm SW} \bar \psi F_{\mu\nu}[\gamma_\mu,\gamma_\nu]\psi\,.
\end{equation}
If the coefficient $c_{\rm SW}$ is chosen appropriately
(see~\cite{Luscher:1996sc}), all remaining
linear cutoff effects are proportional to the quark masses
$\mathcal O(am)$.~\footnote{These terms can be further
eliminated. See~\cite{Sommer:1997xw} for a comprehensive review of the
improvement programme.} This last term is usually called \emph{clover}
term in the lattice jargon, and the discretization is referred as
Wilson-clover fermion action.
\item[Twisted mass fermions] A close relative of Wilson clover
fermions are twisted mass fermions. In this case one uses the same
5-dimensional operator to break chiral symmetry, except that in this
case the mass term is of the form
\begin{equation}
m + \imath \mu \gamma_5\tau_3\,,
\end{equation}
where $\tau_3$ is the third Pauli matrix acting in flavor
space. Twisted mass lattice QCD always describes multiples of two fermion
flavors with a mass given by a combination of $m$ and
$\tau$. Note that in the continuum one can always set
$\mu=0$ (with the help of a non-anomalous chiral transformation),
recovering the usual mass term.
On the other hand, at non-zero values of the lattice spacing, the
twisted mass term cannot be reduced to the standard Wilson form,
because of the explicit breaking of chiral symmetry of the Wilson
term Eq.~(\ref{eq:wterm}).
The main advantage of this formulation is that, for a specific
choice of the mass parameter $m$ (called \emph{maximally twisted})
all \emph{physical observables} are automatically $\mathcal
O(a)$-improved~\cite{Frezzotti:2003ni}. In
twisted mass lattice QCD there is no need of tuning the $c_{\rm SW}$
term of Eq.~\eqref{eq:csw}. On the other hand, parity and flavor
symmetries are broken at
finite lattice spacing, and, as already mentioned, one can only
simulate an even number of quarks.
The reader might be interested in the review~\cite{Shindler:2007vp}.
\item[Staggered fermions] One can live with some of the doublers and
in fact use them in one's favor~\cite{Kogut:1974ag,
Susskind:1976jm}. This is the approach taken by the
staggered formulations of QCD: some of the doublers are used to
represent the 4 spin components of the fermion. The staggered
fermion formulation therefore reduces the amount of doublers to
4. The main advantage of this approach is that it preserves an
exact $U(1)$ symmetry that is enough to guarantee, among other
things, that scaling violations are proportional to $a^2$. Moreover
they are computationally cheap.
The main drawback of this particular fermion formulation is that the 4
remaining doublers are degenerate in mass, while the 4 lightest quarks in
nature have very different masses. In order to describe single flavors, the
staggered fermion formulation uses a \emph{rooting prescription:} the fermion
determinant that describes the 4 doublers is replaced with its fourth root
with the hope that this will describe a single
flavor~\cite{Marinari:1981qf}. This rooting
prescription has the unpleasant effect of breaking locality. It has been
argued that locality is recovered in the continuum.
If this is actually the case or not has been the subject of many heated
discussions in the past, although the issue has never been completely
resolved~\cite{Sharpe:2006re,Bernard:2006vv,Bernard:2007eh,
Creutz:2007yg}.
\item[Domain Wall fermions] One way to circumvent the Nielsen-Ninomiya
theorem is to require that the fermion action, at finite lattice
spacing, is invariant under a set of modified chiral
transformations,
\begin{equation}
\label{eq:ModifiedChiralTransf}
\psi \mapsto e^{i \alpha \gamma_5 [1 - 2\zeta aD]} \psi\, ,
\quad \mathrm{and}\ \quad
\bar\psi \mapsto \bar\psi e^{i \alpha \gamma_5 [1 - 2(1-\zeta) aD]}\, ,
\end{equation}
where $\zeta$ is a free parameter~\cite{Luscher:1998pqa}. A lattice
Dirac operator that is invariant under these transformation
satisfies
\begin{equation}
\label{eq:GWRelation}
\left\{\gamma_5, D\right\} = 2 a D \gamma_5 D\, .
\end{equation}
This relation is known as Ginsparg-Wilson relation, and was derived
from renormalization group arguments in Ref.~\cite{Ginsparg:1981bj}
long before the symmetry above was suggested.
One solution of Eq.~\eqref{eq:GWRelation} is the overlap operator
obtained in Ref.~\cite{Neuberger:1997fp}:
\begin{equation}
\label{eq:NeubOp}
a D_N = \frac12 \left[
1 + \gamma_5 \mathrm{sgn}\left(H\right)
\right]\, ,
\end{equation}
where $H = \gamma_5 (D_W - M)$, where $D_W$ is a lattice Dirac
operator that has the correct naive continuum limit, and $M\sim 1/a$ is a
mass parameter.
The operator $D_N$ is a representation in terms of four-dimensional fields of
five-dimensional domain-wall fermions (DWF)~\cite{Kaplan:1992bt}. In DWF
formulations, a fermionic zero-mode with definite chirality, localised on the
boundary of the (semi-infinite) extra dimension, plays the role of a
four-dimensional chiral fermion, while the other states decouple from the
low-energy dynamics. As it turns out, the five-dimensional formulations of
DWFs presented in Refs.~\cite{Shamir:1993zy,Brower:2012vk} has become the
method of choice for simulating chiral fermions on the lattice. Without
entering into a detailed explanation of the DWF construction on the lattice,
we write down explicitly the action for DWF, and discuss some of its features,
while referring the interested reader to the literature for detailed
discussions~\cite{Kaplan:2009yg,Kennedy:2006ax}. Denoting the coordinate in
the fitfth direction by $s \in [0,\infty)$, the DWF action in
Ref.~\cite{Shamir:1993zy} reads
\begin{align}
\label{eq:ShamirDWF}
S =& \sum_{x,s,\mu} \bar\psi(x,s) \gamma_\mu \partial_\mu \psi(x,s)
+ M \bar\psi(x,s) \psi(x,s) + \frac{r}{2} \sum_{x,s,\mu}
\bar\psi(x,s) \partial_\mu^2 \psi(x,s) \, + \nonumber \\
&+ \sum_{x,s>0} \bar\psi(x,s) \gamma_5\partial_5 \psi(x,s) +
\frac{r}{2} \sum_{x,s>0} \bar\psi(x,s) \partial_5^2
\psi(x,s) \, + \nonumber \\
&+ \frac12 \sum_x \bar\psi(x,0) \gamma_5 \psi(x,1) +
\frac{r}{2} \sum_x \bar\psi(x,0) \left[
\psi(x,1) - 2 \psi(x,0)
\right]\, .
\end{align}
The index $\mu$ runs over the first four Euclidean directions, and
the derivatives $\partial_\mu$ become covariant derivatives when the
fermions are coupled to a four-dimensional gauge field. In actual
simulations the fifth dimension has a finite size, while the lattice chiral
symmetry is only recovered when the size of the fifth dimension goes
to infinity. In practice a balance must be found between the breaking
of chirality due to the finite fifth dimension and the growing cost
of the simulations as the size of this dimension is increased.
\end{description}
\subsection{Masses, correlators and all that}
Quantum Field Theories provide the machinery to evaluate field
correlators, {\em i.e.}\ the expectation value of functions of the elementary
fields that appear in the definition of the path integral. Specific
physical properties can then be extracted from these correlators by
means of dedicated analyses. It is interesting to discuss a few
explicit examples in order to introduce some of the quantities that
are used later in this review.
Let us begin with a two-point correlator
\begin{equation}
\label{eq:TwoPtCorr}
C_{\Gamma\Gamma'}(t) = \sum_{\mathbf{x},\mathbf{y}}
\langle O_\Gamma(t,\mathbf{x})\,
O_{\Gamma'}^\dagger(t,\mathbf{y}) \rangle\, ,
\end{equation}
where $O_\Gamma$ is a quark bilinear
\begin{equation}
\label{eq:QuarkBilinear}
O_\Gamma(t,\mathbf{x}) = \bar{q}(t,\mathbf{x}) \Gamma
q'(t,\mathbf{x})\, .
\end{equation}
The matrix $\Gamma$ determines the spin structure of the bilinear, which
we assume to be a color singlet. In Eq.~\eqref{eq:QuarkBilinear} we have
suppressed the flavor indices, for simplicity, we assume the bilinear to
be a non-singlet with respect to flavor transformations.
Using Wick's theorem, the correlator can be rewritten in terms of
quark propagators:
\begin{equation}
\label{eq:TwoPtProps}
C_{\Gamma\Gamma'}(t) = \sum_{\mathbf{x},\mathbf{y}}
\mathrm{tr}\
\left[
\Gamma' S(0,\mathbf{y}; t,\mathbf{x})
\Gamma S'(t,\mathbf{x}; 0,\mathbf{y})
\right]\, ,
\end{equation}
where $S(x;y)$ and $S'(x,y)$ are the propagators of the quarks $q$
and $q'$ respectively, computed as the inverse of the Dirac operator
in the gauge background. The expectation value is computed by
averaging the trace above over an ensemble of gauge configurations
generated by Monte Carlo methods.
In order to extract physical quantities from a two-point function, we
insert a complete set of hadronic states,
\begin{equation}
\label{eq:UnitarityId}
1 = \sum_n \int \frac{d^3\mathbf{p}}{(2\pi)^3}
\frac{1}{2E_n} |n,\mathbf{p}\rangle \langle n,\mathbf{p}|\, .
\end{equation}
A few lines of algebra show that the sum over the spatial coordinates
implements a projection over zero-momentum states, and therefore
\begin{equation}
\label{eq:TwoPtSpectral}
C_{\Gamma\Gamma'}(t) =
\sum_n \frac{1}{2E_n}\,
\langle 0 | O_\Gamma(0) | n, {\mathbf p}=0\rangle
\langle n, {\mathbf p}=0 | O_{\Gamma'}(0)^\dagger | 0\rangle e^{-E_n t}\, .
\end{equation}
The two-point function can be trivially extended to project on eigenstates of the spatial momentum.
Eq.~\eqref{eq:TwoPtSpectral} shows the main features that allow the
extraction of physical observables from field correlators. In Euclidean
space the time dependence of correlators is a sum of exponentials, whose
decay rates are determined by the energies of the states that have a
non-vanishing matrix element
$\langle 0 | O_\Gamma(0) | n, {\mathbf p}=0\rangle$. For this reason these
operators are often called {\em interpolating operators}. For large time
separations, the correlators are dominated by the lowest energy state, and
the time dependence becomes a simple exponential. The prefactors
multiplying the exponential yield various combinations of matrix elements
of interest.
The so-called effective mass is defined as the time-derivative of the
two-point function:
\begin{equation}
\label{eq:EffMassDef}
M_\mathrm{eff}(t) = -\frac{d}{dt} \log C_{\Gamma\Gamma'}(t) =
E_0 + c\, e^{-(E_1-E_0)t} + \ldots\, .
\end{equation}
For large times $t$, $M_\mathrm{eff}(t)$ tends to a constant, which is the mass
of lowest-energy state. However for the range of separations that can be
achieved in practice, it is always important to check that the contamination
from excited states is sufficiently small, or otherwise under control, since
this is one of sources of systematic error that affect the computation of
physically interesting observables.
Another interesting example is the computation of the PCAC mass. Specialising
the above interpolating operators to the case of two degenerate light quarks
($u$ and $d$), and following the notation in Ref.~\cite{DelDebbio:2007pz}, we
define
\begin{equation}
\label{eq:PandA}
P^{ud}(x) = \bar{u}(x) \gamma_5 d(x)\, , \quad \mathrm{and} \quad
A^{ud}_\mu(x) = \bar{u}(x) \gamma_\mu \gamma_5 d(x)\, ,
\end{equation}
and the two-point correlators
\begin{align}
C_{PP}(t) &= \sum_\mathbf{x} \langle P^{ud}(t,\mathbf{x})
P^{du}(0)\rangle\, ,\\
C_{AP}(t) &= \sum_\mathbf{x} \langle A^{ud}_0(t,\mathbf{x})
P^{du}(0)\rangle\, .
\end{align}
Following the arguments above, $C_{PP}(t)$ at large distances is dominated by a
a single pseudoscalar state. Denoting by $M_\mathrm{PS}$ the mass of the
pseudoscalar state, and by $G_\mathrm{PS}$ its vacuum-to-meson matrix element,
we obtain
\begin{align}
C_{PP}(t) &= - \frac{G_\mathrm{PS}^2}{M_\mathrm{PS}}\,
e^{-M_\mathrm{PS} t} + \ldots\, .
\end{align}
Interestingly the ratio
\begin{align}
\label{eq:mPCAC}
m_\mathrm{\tiny PCAC} = \left(
\frac12 (\partial_0 + \partial_0^*) C_{AP}(t) + c_A a \partial_0^*
\partial_0 C_{PP}(t)
\right) / C_{PP}(t)\, ,
\end{align}
tends to the PCAC mass defined in the continuum theory through the
axial Ward identity. The interest in the PCAC mass is two-fold. For
fermionic formulations that break explicitly chiral symmetry at finite
lattice spacing, like {\em e.g.}\ the Wilson fermions described above, the
bare parameters in the action need to be tuned to approach the chiral
limit. In particular an additive renormalization of the bare fermion
mass is required. The chiral theory is defined by requiring the PCAC
mass to vanish. After renormalization, the rate of convergence of
$m_\mathrm{\tiny PCAC}$ will be proportional to the lattice spacing
$a$ if the theory is not improved, while it becomes proportional to
$a^2$ when $c_{\rm SW}$ in Eq.~\eqref{eq:csw} and $c_A$ in
Eq.~\eqref{eq:mPCAC} are properly tuned.
Note that the decay constant of the PS state, defined as
\begin{equation}
\label{eq:DecayConst}
\langle 0 | A_\mu(0) | \mathrm{PS}\rangle = i Z_AF_\mathrm{PS} p_\mu\, ,
\end{equation}
can be computed directly from fitting\footnote{The renormalization
constants $Z_A$ and $Z_m$ need also to be computed.
We point the reader to the reviews~\cite{Sommer:1997xw,Vladikas:2011bp}
for more details on the topic.} $C_{AP}(t)$ and $C_{PP}(t)$,
which yield the matrix elements
$\langle 0 | A_0(0) | \mathrm{PS}\rangle$ and
$\langle 0 | P(0) | \mathrm{PS}\rangle$. Alternatively, the decay
constant can be evaluated from the quantities defined above using
\begin{equation}
\label{eq:DecayConstTwo}
F_\mathrm{PS} = Z_m\frac{m_\mathrm{\tiny PCAC}}{M_\mathrm{PS}^2}\,
G_\mathrm{PS}\, .
\end{equation}
It is interesting to remark that the two definitions differ by lattice
artefacts. While in general this is not necessarily a cause for
concern, it shows that definitions that are equivalent in the
continuum limit do differ at finite lattice spacing. This is
something to keep in mind in choosing observables when aiming for high
precision measurements.
\subsection{Systematic effects in lattice QCD}
Any lattice QCD computation has several sources of systematic uncertainties
that have to be kept under control in order to be able to quote accurate
results. Since lattice QCD is a first principle definition of QCD, these
sources of systematic uncertainties reflect the current limitations in
computer power, or in our knowledge of efficient algorithms. Since both
computer power and our knowledge of efficient algorithms are constantly
improving, lattice QCD is able to solve now problems that were basically
impossible just a few years ago.
The most basic limitation that computer power sets in any lattice
computation is related to the number of points $(L/a)^3\times (T/a)$
simulated (the lattice volume). The computer cost increases \emph{at
least} linearly with this lattice volume, and sets a basic
compromise between simulating a large physical volume, and a small
lattice spacing. Computing resources also limit the range of quark
masses that can be simulated, even though it has become common to have
simulations with physical quark masses. In summary we have the
following main sources of systematic error.
\begin{description}
\item[Finite volume corrections: ] QCD quantities determined on a
large but finite box suffer from finite volume effects. These are
exponentially suppressed with the smallest mass present in the
spectrum of the theory (i.e. the pion $M_\pi$)~\cite{Luscher:1985dn}
\begin{equation}
\propto e^{-M_\pi L}\,.
\end{equation}
Usually $M_\pi L > 4$ is sufficient for finite volume effects to be a
small sub-percent correction to the quantities determined on a finite
volume box. $M_\pi L>3$ is the bare minimum to keep these effects under
control.
For some hadronic quantities ({\em e.g.}\ meson decay constants) chiral
perturbation theory yields an estimate of the size of the finite
volume corrections, allowing to subtract them from the data in some
cases.
Since one typically uses a short distance observable to determine
the strong coupling, these determinations are normally affected very
little by finite volume effects. Nevertheless every determination of
the strong coupling needs a determination of the scale (see
section~\ref{sec:scale-sett-latt}), where finite volume corrections
can be substantial.
\item[Continuum extrapolation: ] All determinations on the lattice require
a continuum extrapolation to reproduce QCD results. Lattice artefacts are
small only if we can achieve a significant separation between the scales
at which observables are defined and the lattice cutoff. We will examine
in detail the process of taking the continuum limit in
section~\ref{sec:cont-limit-scale}. Here we just mention that this point
is particularly delicate for the determinations of the strong
coupling. As we try to use observables computed at high energies in order
to define the strong coupling, we necessarily need to face the issue of
larger cutoff effects. Typical current large volume simulations use
$a\approx 0.04-0.1~{\rm fm}$.
\item[Chiral extrapolation: ] Many lattice QCD computations used to be
performed at nonphysically heavy values of the quark masses. There
are two reasons for this. First, lattice QCD simulations become more
expensive at lighter quark masses. The gap in the spectrum of the
Dirac operator depends on the mass of the lightest quark in the
simulation. This has the effect of making simulations close to
physical values of the quark masses computationally very
expensive. Second, close to physical values for the quark masses
finite volume effects are larger, and therefore there is an extra
cost due to the need to simulate in larger physical volumes.
For example, simulating physical values for the quark masses
($M_\pi \approx 135\, {\rm MeV}$) with sub-percent finite volume
effects $M_\pi L = 4$ and a very fine lattice spacing
$a\approx 0.05\, {\rm fm}$ requires a lattice with $L/a\approx 120$
points in each direction. At the time of writing this report this is
right at the edge of current capabilities for most choices of lattice
action. Algorithmic developments have made it possible to simulate
directly at the physical point, but these physical regimes of the
parameters are usually simulated on coarser lattices, making the
chiral extrapolation a crucial ingredient of any lattice QCD
computation.
\end{description}
\subsection{The continuum limit and scale setting}
\label{sec:cont-limit-scale}
Any lattice action has $N_{\rm f}+1$ free parameters: the bare quark
masses in lattice units $am_0$, and the bare coupling $g_0$ (usually
the lattice community uses $\beta = 6/g_0^2$ as input parameter for the
simulations). While the role of the bare quark masses is clear (they
directly affect the values of the quark masses), the role of the bare
coupling is less obvious. In fact the bare coupling is tuned in order
to approach the continuum limit. Naively the continuum limit amounts
to take $a\to 0$. But in the lattice action there is nowhere any
reference to the lattice spacing $a$ (or any other parameters with
dimensions), raising the question about how to actually take the
continuum limit.
Since all lattice input parameters are dimensionless, lattice QCD by itself only
gives predictions of dimensionless quantities. For example, the study of the
proton correlator yields the proton mass in lattice units $aM_p$. The key idea
in order to make contact with physical, dimensionful, quantities is to choose
\emph{one} quantity as a reference scale. Every other dimensionful quantity is
computed in units of this reference scale. For example, one can take as
reference scale the proton mass $M_p$. Any other quantity, say for instance the
$\Omega$ baryon mass is measured in units of this reference mass -- in practice
in a lattice simulation we determine the dimensionless ratio $(aM_\Omega)/(aM_p)
= M_\Omega/M_p$.
If we focus on the case of $N_{\rm f}=2+1$ simulations (two degenerate
light quark masses plus the strange quark), a prediction for the value
of $M_{\Omega}$ would conceptually proceed as follows:
\begin{enumerate}
\item Choose a value of the bare coupling $g_0$. Measure the values of the
masses of the $\pi$ and $K$ mesons and the reference scale in lattice units
(i.e. $aM_\pi, aM_K, aM_p$). Tune the bare quark masses such that the ratios
of the pseudo goldstone bosons masses to the reference scale $(aM_\pi)/(aM_p)$
and $(aM_K)/(aM_p)$ are equal to the physical values $M_{\pi}^{\rm
exp}/M_p^{\rm exp}$ and $M_k^{\rm exp}/M_p^{\rm exp}$, see {\em e.g.}\ the values
reported by the PDG~\cite{Eidelman:2004wy}. This procedure fixes the values of
the bare quark masses $am_0$ for each choice of $g_0$. The lattice spacing is then $a=(a M_p)/M_p^{\rm exp}$, where the numerator is the output of the numerical simulation, and the denominator is the reference scale.
\item Repeat the process for several values of $g_0$. The final prediction has
to be taken as the limit where the reference scale in lattice units is much
smaller than one, {\em i.e.}\ $aM_p \ll 1$:
\begin{equation}
\label{eq:cont_limit}
\frac{M_\Omega}{M_p} = \lim_{aM_p\to 0} \frac{aM_\Omega}{aM_p}\,.
\end{equation}
By using the experimental value of the proton mass $M_p \approx 940$ MeV, this
last prediction of a dimensionless ratio can be translated in a prediction for
$M_\Omega$.
\end{enumerate}
The first step sets up a \emph{line of constant physics} (LCP). It requires
one experimental input per quark mass. Pseudo-goldstone bosons are the
natural candidates, since their mass depends strongly on the values of
quark masses. The value of the reference scale ($M_p$ in the example) is an
extra input, used to convert lattice dimensionless predictions in
dimensionful quantities. Note that although usually one takes the values of
the quantities to fix from experiment, one can follow the same procedure to
set up a line of constant physics at arbitrary non-physical values of the
quark masses (for example to investigate a world with mass degenerate
$u,d,s$ quarks): lattice QCD also allows to make \emph{unambiguous}
predictions of dimensionless quantities for \emph{unphysical} values of the
quark masses. The second step \emph{takes the continuum limit}. Different
LCP (for example by choosing a different reference scale) would result in
different approaches to \emph{the same universal continuum values}.
These two steps together define a non-perturbative renormalization scheme for
the theory. The bare parameters are tuned in order to reproduce some physical
world (values of the quark masses and reference scale) and quantities are
computed in the limit where the UV cutoff ($1/a$) is much larger than the energy
scales of interest ($M_p, M_\Omega, \dots$). Asymptotic freedom can be used
as a guidance in order to take this continuum limit. At short distances the
theory is weakly coupled, which suggests that a series of decreasing values of
the bare coupling $g_0$ will successively approach the continuum limit.
Even though the procedure sketched above is perfectly correct from a conceptual
point of view, it misses some fine details that are crucial in order to obtain
the precision achieved nowadays. Let us briefly mention them here.
\begin{itemize}
\item As mentioned above lattice QCD simulations become very expensive at the
bare parameters that correspond to the physical values of the quark masses.
There are two reasons for this. First is the increasing numerical cost of
simulating quarks at small values of the bare quark mass. Second, lighter
values of the quark masses result in lighter values of the $\pi$ meson masses.
Since $M_\pi L$ is the quantity that dictates the size of finite volume
effects, simulations at lighter quark masses requires to simulate larger
physical volumes. In many practical situations the physical point is only
reached by extrapolating from simulations at heavier quark masses, although
this situation is changing fast with the increasing computer power and
improved simulation algorithms.
\item The experimental values of the physical hadronic quantities are affected
by the electromagnetic interactions. This has implications for the
determination of the LCP, since the use of the experimental values as input
for a lattice QCD computation has to be done with care, usually correcting
them for isospin breaking effects. One has to make sure that the size of the
electromagnetic effects has a negligible effect in the determination of
physical quantities. When this is not the case, a first-principles prediction
requires to simulate both the strong and the electromagnetic interactions in
order to make contact with the physical world.
We anticipate here that for the case of the the
determination of $\alpha_s$, isospin breaking effects are
not particularly relevant at the current level of precision
(cf. section~\ref{sec:pres-future-latt}).
\end{itemize}
\subsubsection{Systematics in the continuum extrapolation}
Since short-distance observables are more susceptible to show large
cutoff effects, the continuum extrapolation plays a key role in the
extraction of the strong coupling. Here we want to show in some detail
how to asses the quality of these continuum extrapolations.
Most lattice QCD simulations choose an improved discretization where the
leading scaling violations that we discussed above are $\mathcal
O(a^2)$. This is achieved either by choosing a fermion formulation where
any $\mathcal O(a)$ violations are forbidden by symmetry arguments (like
domain-wall fermions, twisted mass at maximum twist, or staggered
fermions), or by tuning the parameters of the action (the so-called
Wilson-clover fermions). In this situation any
dimensionless quantity $D$ has an asymptotic expansion of the form
\begin{equation}
D \simas{a\to 0} D_{\rm cont} + \dots\,,
\end{equation}
where the dots represent the scaling violations, with an asymptotic
expansion with leading term $\mathcal
O(a^2)$. But the functional form of these scaling violations is in
general very complicated. The corrections include
logarithmic terms of the form (i.e.
are of the form $a^n\log^k a$). How does
these generic statements affect in practice the extrapolation of
lattice data? It is clear that corrections need to be
small in order for the precise functional form to have a negligible
effect on the extrapolation.
\begin{figure}[ht!]
\centering
\includegraphics[width=\textwidth]{fig/cont.pdf}
\caption{Continuum extrapolation of a dimensionless quantity. Without
entering in the details of the definition of the specific quantity in
the plot, it is worth noting that these data correspond to actual lattice
simulations, see~\cite{ramos:19lat} for more details.}
\label{fig:cont}
\end{figure}
In order to be more precise it is best to look at one example.
Figure~\ref{fig:cont} shows the continuum extrapolation of a dimensionless
quantity
using two different discretizations. These are labeled 1 and 2 respectively
in the legend. The extrapolations of both datasets have very different
scaling violations, and this has an impact in the accuracy of the continuum
value. These differences are:
\begin{description}
\item[Range of lattice spacings: ] Dataset 1 includes simulations that
span a factor
\[
a_{\rm max}/a_{\rm min}\approx 2.5
\]
in lattice spacings, while dataset 2 has data that span only a factor
\[
a_{\rm
max}/a_{\rm min}\approx 1.7\, .
\]
\item[Size of the extrapolation: ] Dataset 1 has very small scaling violations:
the finest lattice spacing is compatible within errors with the continuum
value. On the other hand dataset 2 shows large scaling violations: the data at
the finest lattice spacing is about 5 standard deviations away from the
extrapolated value.
\end{description}
These two points are crucial to understand the assumptions that are behind the
extrapolation of these two datasets. What we need is a way to measure the impact
of the terms \emph{beyond} the leading $\mathcal O(a^2)$ scaling violations. In
order to do so, it is useful to consider not only the extrapolation of $D$, but
also the extrapolation of different functions $f(D)$. In principle this function
seems completely redundant, since
\begin{equation}
f(D) \simas{a\to 0} f_{\rm cont} + \dots\,,\qquad (f_{\rm cont} =
f(D_{\rm cont}))\,,
\end{equation}
one can, in principle, recover \emph{the same} $D_{\rm cont}$ from any choice of
function by using
\begin{equation}
D_{\rm cont} = f^{-1}(f_{\rm cont})\,.
\end{equation}
However, when the range of lattice spacings available is not very large and/or
the extrapolation is quite sizable, the choice of function can change
substantially the result of the extrapolation. This is illustrated in
figure~\ref{fig:cont}, with two simple choices $f(x) = x$ (i.e. extrapolating
the data itself) and $f(x)=1/x$ (i.e. extrapolating the inverse of the data). In
set 1, these different choices result in practically identical values for the
extrapolated value. On the other hand in set 2 the different choices of $f$
result in extrapolations that differ by several standard deviations. Obviously
the difference between these two choices for the extrapolation are higher order
terms, since
\begin{equation}
\frac{1}{A + Ba^2} = \frac{1}{A} - \frac{B}{A^2}a^2 +
\frac{B^2}{A^3} a^4 + \dots \,.
\end{equation}
These results suggest that \emph{the dependence of the extrapolation
on the choice of function $f$ is equivalent to a dependence on the
non-leading scaling violations}. It is important to note that
discriminating among different functions cannot always be done by
looking at the quality of the fit. In the quoted example of
figure~\ref{fig:cont} the fits for the two choice of $f$ are equally good.
Another important point worth mentioning is that this kind of analysis
usually only explores the non-leading power corrections ({\em i.e.}\ $a^4$
in the example). Logarithmic corrections are much more difficult to
estimate, although one should be very careful with neglecting them (the
reader interested in this topic will enjoy the puzzle described
in Ref.~\cite{Balog:2009np}, and the recent discussion in the context
of lattice QCD~\cite{Husung:2019ytz}).
These topics are also very nicely discussed in a recent contribution to
the Lattice Field Theory Symposium~\cite{DallaBrida:2020pag}, a
reference that we recommend to the reader will enjoy.
Summarising these results, it is crucial to remember that any extrapolation
is based on some assumptions. In the case of the continuum extrapolations,
these assumptions depend crucially on three characteristics of the dataset:
the number of lattice spacings simulated, the ratio of the coarser and the
finest lattice spacings, and the size of the extrapolation.
\subsubsection{Scale setting in lattice QCD}
\label{sec:scale-sett-latt}
This process of taking the continuum limit is usually seen by the lattice
community from an equivalent perspective, which comes under the name of
\emph{scale setting}. As discussed above some reference scale~\footnote{For the
sake of the presentation we will assume that the reference quantity has units of
mass.} $M_{\rm ref}$ is used to determine the value of the lattice spacing $a$.
This is simply done by setting up a line of constant physics and declaring that
\begin{equation}
\label{eq:a_determination}
a \equiv \frac{\left(aM_{\rm ref}\right)}{M_{\rm ref}^{\rm exp}}\,,
\end{equation}
where the parentheses emphasise that $\left(aM_{\rm ref}\right)$ is
the quantity that is actually computed in a lattice simulation, and
$M_{\rm ref}^{\rm exp}$ is the experimental value of the reference
scale. Now any other quantity $Q$ with units of mass can be determined
by the expression
\begin{equation}
Q = \lim_{a\to 0} \frac{(aQ)}{a}\,.
\end{equation}
Obviously this is just a rephrasing of Eq.~\eqref{eq:cont_limit}. Note that
determinations of $a$ are intrinsically entangled with the LCP and the
experimental quantity that is used to set the scale (the particular choice of
$M_{\rm ref}$). Different LCP and/or physical quantities will result in
different values for the lattice spacing. This is however not a problem since
\emph{any} such determinations will give the same predictions for any physical
quantity \emph{after the continuum limit is taken}. This procedure explains the
usual jargon of the field: one experimental input is used to determine the
lattice spacing.
Scale setting is a key ingredient in any lattice determination of the
strong coupling. The $\Lambda$ parameter is a quantity with units of
mass, and therefore the error in the scale translates into an equal
relative error in the determination of $\Lambda$. What is more
important \emph{any unaccounted systematic in the determination of the
scale propagates into an effect of the same relative size to}
$\Lambda$. What quantities are used as reference scales? Ideally one would like
some quantity that has a clean and precise experimental determination,
and that can be determined with high precision and accuracy on the
lattice. The reader might be surprised to read that no such clean
quantity exist. For example the above mentioned proton mass, that
would look as a natural quantity, has a large signal to noise problem
(see appendix~\ref{ap:challenges}), especially close to the physical
point. In practice different quantities are used, with each choice
having pros and cons. Let us briefly review the characteristics of
some of them.
\bigskip
\noindent
{\em Meson decay constants ($F_\pi, F_K$)}
\medskip
These quantities are clean from a lattice point of view, which explains why they
are a popular choice in the lattice community. They are determined from meson
2-point functions, as discussed above. For example $F_\pi$ is obtained thanks to
the relation
\begin{equation}
\label{eq:fp_corr}
\left\langle 0|\bar u\gamma_5\gamma_0 d(0)[\bar u\gamma_5\gamma_0 d(x)]^\dagger |0 \right\rangle \sim
aM_\pi(aF_\pi)^2e^{-aM_\pi\, x_0} + \dots\, ,
\end{equation}
which is free from the infamous signal-to-noise problem, leading to these
decay constants in lattice units $aF_\pi, aF_K$ being determined with a
precision of a few permille. One problem is that the chiral
corrections on the decay constants (especially $F_\pi$)
is not small. Recent
determinations are performed at values of the quark masses very close to
its physical values, so in principle this has become a lesser issue in
state of the art calculations.
On the other hand decay constants are not that clean from the
experimental point of view. First, there is a theoretical issue in the
definition of these quantities. They are unambiguous in QCD, but the
electromagnetic interactions render these quantities ill defined --
the quantity in Eq.~\eqref{eq:fp_corr} is not even invariant under a
$U(1)$ gauge transformation. The actual experimental observable is the
photon-inclusive decay rate $\Gamma_{Pl2}$, defined as
\begin{equation}
\Gamma_{Pl2} = \Gamma(P\rightarrow l\nu) +
\Gamma(P\rightarrow l\nu\gamma)\, ,
\end{equation}
with $P=K^\pm,\pi^\pm$. Most lattice QCD calculations~\footnote{... albeit not
all!} compute the isospin symmetric quantity $F_P$ which parametrizes the decay
of the pseudoscalar in a world where electromagnetic interactions are switched
off, and isospin is unbroken:
\begin{equation}
\Gamma(P\rightarrow l\nu)\Big|_{\alpha_{\rm EM} = 0, m_u = m_d} = \frac{G_F
|V_P|^2 F_{P}^2}{4\pi} \,
M_P m_l^2\left[ 1-\frac{m_l^2}{M_P^2}\right]^2\,.
\end{equation}
In order to relate these two quantities we need to estimate the EM
effects. The master formula, as discussed in Ref.~\cite{Marciano1993},
is
\begin{eqnarray}
\Gamma_{Pl2} = \Gamma(P\rightarrow l\nu)
&\times&
\left[1 + \delta_{\rm EW}
\right] \times \left( 1 + \delta_{\rm EM}^P \right)\, ,
\end{eqnarray}
where the first bracket term is the universal electroweak correction
\begin{equation}
\delta_{\rm EW} =
0.0232 \approx
\frac{2\alpha_{\rm EM}}{\pi}\log\left(\frac{m_Z}{m_\rho}\right)\,,
\end{equation}
a short distance contributions that affects all semileptonic charged current
amplitudes when expressed in terms of the Fermi constant. Finally the
$\delta_{\rm EM}^P$ piece can be written as
\begin{equation}
\begin{split}
\delta_{\rm EM}^P = -\frac{\alpha_{\rm EM}}{\pi}&\left\{
- F\left(\frac{m_l^2}{m_P^2}\right) +
\frac{3}{2}\log\left(\frac{m_\rho}{m_P}\right) + c_1^P
\right.\\
&+\left.
\frac{m_l^2}{m_\rho^2}\left[c_2^P
\log\left(\frac{m_\rho}{m_l}\right) +
c_3^P + c_4^P\left(\frac{m_l}{m_P} \right)\right]
- \tilde c_2^P \frac{m_P^2}{m_\rho^2}\log\left(\frac{m_\rho^2}{m_l^2}\right)
\right\}\,.
\end{split}
\end{equation}
In this expression, the first term is the universal long distance
electromagnetic correction, computed to one loop assuming that the $\pi,K$
are point-like particles, while the terms proportional to
$c_{1,2,3,4}^P, \tilde c_2^P$ parametrize the structure dependent part,
$c_1^P$ being the leading one. A conservative estimate of the
electromagnetic uncertainties consists in taking the full difference
between the point particle approximation -- where all $c_i^P, \tilde c_i^P$
are zero -- and the best phenomenological determination available in the
literature, based on $\chi$PT. This gives an uncertainty $\sim 0.27\%$ for
the case of $F_\pi$ and $\sim 0.21\%$ for $F_K$. These figures set a
conservative limit on the precision that can be quoted for these decay
constants, and therefore a limit on the precision of the scale
determination for any lattice computation that relies on them to set the
scale. Going beyond this precision requires the inclusion of
electromagnetic effects on the lattice and a direct computation of the
decay rate, something that nowadays is challenging from the theoretical
point of view, but an impressive progress has been achieved
recently~\cite{Lubicz:2016xvp, DiCarlo:2019thl}.
For the particular case of $F_K$ one has also to take into account the
strong isospin breaking effects, which vanish at leading order for the case
of $F_\pi$. This is only a practical problem, since the leading corrections
$\propto (m_u-m_d)$ can in principle be determined. Finally, the relation
between the experimentally measured decay rates and the decay constants
involves some CKM matrix elements. In the case of $F_\pi$ the relevant term
is $V_{ud}$, that is very well determined experimentally from super-allowed
$\beta$-decays, but the case of $F_K$ needs a determination of $V_{us}$,
which usually involves lattice input and assumes CKM unitarity relations.
All in all, the pion decay rate ($F_\pi$ in pure QCD) remains an attractive
quantity for scale setting, especially nowadays that we can simulate quark
masses close to their physical values. It can be determined very accurately
on the lattice. The model dependent electromagnetic corrections are below
the 0.3\%, the leading strong isospin breaking effects vanish, and the
necessary CKM matrix element is cleanly determined experimentally.
\bigskip
\noindent
{\em The $\Omega$ mass}
\medskip
The mass of the $\Omega^{-}$ baryon is also a common choice for scale
setting. This particle is stable under the strong interactions and its mass
is known very precisely. Being made of three strange valence quarks, the
dependence on the value of the light quark masses is only induced via loop
effects, which translates in a mild chiral extrapolation if simulations are
performed at constant value of the strange quark mass.
Contrary to the case of the decay constant, what is measured by the
experiment is directly related with what is measured on the
lattice. Strong isospin breaking and electromagnetic corrections on
the $\Omega$ mass are also small, with some recent lattice studies
pointing to corrections below the 0.3\%.
Unfortunately the determination of the
$\Omega^-$ mass on the lattice is challenging. Like
all baryons, the $\Omega^-$ is affected by the signal-to-noise
problem (see appendix~\ref{ap:challenges}).
The current precision in the scale determined from $M_\Omega$ ranges from
$2\%$ to $0.3\%$. The main difference in the claimed precision stems from
the time range used to extract the mass from the two-point function along
the lines of what we presented above. The most precise determinations use
the values of the correlator at small Euclidean times, where the signal to
noise problem is less severe. On the other hand at this early Euclidean
times there is substantial contamination from excited states, that has to
be disentangled from the real signal of the $\Omega$ mass.
We can summarise the state of affairs by saying that the $\Omega$ mass is a
theoretically clean quantity with a mild dependence on the values of the
light quark masses, and $\le 0.5\%$ isospin breaking corrections, which makes it
an attractive quantity for scale setting. Nevertheless, very precise
determinations of the scale require to control the excited states
contamination in a correlator affected from a strong signal to noise
problem. How to achieve this in practice without assumptions on these
excited states is currently another hot research topic.
\subsubsection{Theory scales}
\label{sec:theory-scales}
An ideal quantity to be used as a reference scale must have some particular
characteristics. First it must have a weak dependence on the quark
masses. Having a simple chiral dependence is crucial for those lattice QCD
simulations that are performed at unphysical values of the quark masses,
and reach the physical point only by extrapolation. Second, the quantity
must be clean from the computational point of view. A quantity that is
complicated to compute on the lattice, as a result of several
extrapolations or some involved fits of the lattice raw data are better
left as predictions, and not as reference scales. Third, the quantity must
have a clear experimental determination, ideally with a weak dependence on
strong isospin breaking effects.
In the previous sections we have seen the typical quantities used for
scale setting (decay constants and $M_\Omega$), and the pros and cons
of each choice. There exists interesting alternatives, with a very
weak chiral dependence and that are straightforward to compute from
the lattice point of view. The drawback is that they are not
quantities that can be accessed by experiments (hence the name
``theory scales''). Nevertheless they are
very useful as intermediate reference scales (i.e. to ``determine the
lattice spacing'' as in Eq.~\eqref{eq:a_determination}).
\bigskip
\noindent
{\em Scales derived from the static potential}
\medskip
One example is the theory scale $r_0$, which is derived from the force
between static quarks, as discussed in
section~\ref{sec:static-potential}. This force $F(r)$ has dimension of
mass squared, and therefore the quantity
\begin{equation}
r^2 F(r)\,,
\end{equation}
is a dimensionless function of the distance between the static
quarks. This suggests to define a reference scale $r_0$ by the
condition~\cite{Sommer:1993ce}
\begin{equation}
\label{eq:rZeroDef}
r^2F(r)\Big|_{r=r_0} = 1.65\,.
\end{equation}
The particular value $1.65$ is chosen so that $r_0\sim 0.5$ fm,
although the shorter distance version $r_1 \sim 0.3$ fm, defined by
$r_1^2F(r_1^2) = 1$ is also commonly used in order to improve the
precision, at the cost of larger cutoff
effects~\cite{Bernard:2000gd}.
The extraction of the scales $r_0,r_1$ from lattice data is not completely free
of challenges. The quantity from which $r_0,r_1$ is extracted has a divergent
signal-to-noise ratio when one approaches the continuum, these challenges have
been overcome by a combination of several techniques
(see~\cite{Sommer:2014mea} for an overview). At the current values of the
simulation parameters a precision $<1\%$ can be achieved in these quantities.
The advantages of scales derived from the static potential are clear. Being
gluonic quantities, their dependence on the value of the quark masses is very
small. The chiral dependence of these quantities is very mild. Even if its
extraction is not completely trivial\footnote{In fact the literature has seen
discrepancies in the values of these scales, although it is not clear that the
fault was the evaluation of the static force, and not the conversion of
$r_0,r_1$ to physical units. See~\cite{Sommer:2014mea}.}, the lattice community
has vast experience determining the static potential.
\bigskip
\noindent
{\em Scales derived from the gradient flow}
\medskip
Recently even better theory scales have been proposed. They are
derived from the gradient flow~\cite{Luscher:2010iy,
Narayanan:2006rf}, a diffusion like process for the gauge field in a
fictitious time coordinate $t$ called flow time. The gauge field
evolves in flow time according to the equation
\begin{equation}
\partial_t B_\mu(t,x) = D_\nu G_{\nu\mu}(t,x),\qquad
B_\mu(0,x) = A_\mu(x),
\label{eq:YMflow}
\end{equation}
where $D_\mu = \partial_\mu + [B_\mu,\cdot]$ is the
covariant derivative with respect to the field $B_\mu$, and
\begin{equation}
G_{\mu\nu}= \partial_\mu B_\nu - \partial_\nu B_\mu + [B_\mu,B_\nu],
\end{equation}
is the corresponding field strength tensor. Note that here $x$ denotes the
four-dimensional spacetime coordinates, while the flow time $t$ has units
of length squared. The field $B_\mu(t,x)$ can be seen as a smoothed version
of the original gauge field $A_\mu(x)$ over a length scale
$\sim\sqrt{8t}$. Gauge invariant quantities constructed from the flow field
$B_\mu(t,x)$ do not need renormalization at $t>0$, beyond the usual
renormalization of the bare parameters of the
Lagrangian~\cite{Luscher:2011bx}. For example, the action density
\begin{equation}
\langle E(t,x) \rangle =
- \frac{1}{4} \langle {\rm tr}\{G_{\mu\nu}G_{\mu\nu}(t,x)\} \rangle
\end{equation}
is a renormalized observable - {\em i.e.}\ it has a finite continuum
limit. By dimensional analysis, the quantity
$t^2 \langle E(t,x) \rangle$ is dimensionless, but its value depends
on the scale $\sqrt{8t}$. Similar to what is done for $r_0$, a
convenient scale can be defined by the condition
\begin{equation}
t^2 \langle E(t,x) \rangle \Big|_{t=t_0} = 0.3\,,
\end{equation}
which results in a hadronic scale ($\sqrt{8t_0} \sim 0.5$ fm). A close
relative to $t_0$ is the $w_0$ scale~\cite{Borsanyi:2012zs}, initially
introduced with the aim of reducing the size of the cutoff effects in
$t_0$ and defined by the condition
\begin{equation}
t\frac{{\rm d} }{{\rm d}t} t^2 \langle E(t,x) \rangle \Big|_{t=w_0^2} = 0.3\,.
\end{equation}
Flow scales have many advantages. First, as for the
case of $r_0,r_1$ they are
simple gluonic observables, so that their chiral dependence is very
mild. But in contrast with the scales $r_0,r_1$
they are given directly by an expectation
value. Their computation in lattice simulations only involves
integrating the flow equation~(\ref{eq:YMflow}), something that can be
done in practice with arbitrary precision. There is no need to look
to the large Euclidean time behavior of a correlator, to perform any
fit or to deal with any signal to noise issue. Moreover flow
observables have a very small variance, making the
statistical errors in the computation of such quantities very
small. Recent scale determination of $t_0$ in the pure gauge case have
reached a precision $\sim 0.2\%$ in very fine lattice
spacings~\cite{Giusti:2018cmp}.
Of course the drawback of any of these theory reference scales is that
ultimately they need to be computed in terms of a real experimental
observable if one aims at making a full prediction. But they are
invaluable as intermediate reference scales, especially if one takes
into account that they can be quoted at quite unphysical values of the
quark masses.
\subsection{Data analysis in lattice QCD}
\label{sec:analysis-lattice-qcd}
As was discussed at the beginning of this section, numerical lattice QCD is
based on the fact that the path integral, after discretization, is an integral
in a large, but finite, dimensional space
\begin{equation}
\mathcal Z_{\rm latt} = \int\, e^ {-S[U]}\, {\rm d} U \,.
\end{equation}
In typical state of the art current simulations, expectation values
are integrals in $d \approx 10^9$ dimensions. They are computed with a
sub-percent precision using a few ($N \sim \mathcal O(1000)$) gauge fields
($U^{(1)}, \dots, U^{(N)}$ ) that are drawn with probability distribution
\begin{equation}
{\rm d}\mathcal P(U^{(k)}) \sim \frac{e^{-S[U]}}{\mathcal Z_{\rm latt}}{\rm d}U\,,
\end{equation}
where ${\rm d}U$ is the Haar measure on $SU(3)$.
Drawing representative ensembles in lattice QCD uses the techniques of
Markov chain Monte Carlo.
For the case of the pure gauge theory very efficient local link update
algorithms exists, but almost every lattice QCD simulation is
performed with some variant of the Hybrid Monte Carlo (HMC)
algorithm~\cite{Duane:1987de}.
Once a representative ensemble $\{U^{(k)} \}_{k=1}^N$ is available,
estimates of any observable are determined by averaging over the
ensemble
\begin{equation}
\langle O \rangle = \frac{1}{N}\sum_{k=1}^NO(U^{(k)}) + \mathcal O \left( 1/\sqrt{N} \right)\,.
\end{equation}
A crucial step in any lattice QCD work is the estimate of the
statistical uncertainty.
{\em i.e.}\, how much does the estimate
\begin{equation}
\bar O = \frac{1}{N}\sum_{k=1}^NO(U^{(k)})
\end{equation}
deviates from the exact value of the expectation value $\langle O
\rangle$. An estimate of this uncertainty $\delta \bar O$ is given by
the variance of the mean
\begin{equation}
\label{eq:varobs}
(\delta \bar O)^2 = {\rm Var}\left[ \frac{1}{N}\sum_{k=1}^NO(U^{(k)}) \right]\,.
\end{equation}
There are two key points in estimating this uncertainty
\begin{enumerate}
\item The properties of Markov chains ensure that the variance of each
of the terms in the sum Eq.~(\ref{eq:varobs}) is the same for all
samples $U^{(k)}$ and in fact given by an expectation value with the
same probability distribution
\begin{equation}
{\rm Var}\left[ O(U^{(k)}) \right] = \langle (O - \langle O \rangle)^2 \rangle = \sigma^2\,.
\end{equation}
\item Any Markov chain Monte Carlo algorithm works by producing the
\emph{next} sample ($U^{(k+1)}$) from the \emph{current} one ($U^{(k)}$).
Subsequent measurements of an observable $O(U^{(k)})$
are \emph{correlated}.
Note that this correlations have the unpleasant effect of increasing
the uncertainties:
the error estimate of an observable (Eq.~(\ref{eq:varobs})) is given by
\begin{equation}
\label{eq:varobs2}
(\delta \bar O)^2 = \frac{\sigma^2}{N}\left[ 1 + \frac{2}{N}\sum_{i>j} \frac{\Gamma(i-j)}{\sigma^2} \right]\,.
\end{equation}
where
\begin{equation}
\label{eq:acf}
\Gamma(i-j) = {\rm Cov}\left( O(U^{(i)}), O(U^{(j)}) \right)\,.
\end{equation}
The first term in the bracket of Eq.~(\ref{eq:varobs2}) accounts for
the error estimate if the data were uncorrelated.
The second term is due to the correlations.
Usually the previous formula for the error of an observable is
written using the \emph{integrated autocorrelation time}
\begin{equation}
\label{eq:tau_rho}
\tau_{\rm int} = \frac{1}{2} + \frac{1}{N}\sum_{i>j} \rho(i-j)\,,\qquad \left( \rho(t) = \frac{\Gamma(t)}{\sigma} \right)\,,
\end{equation}
as
\begin{equation}
\label{eq:varobs_tauint}
(\delta \bar O)^2 = \frac{\sigma}{N}(2\tau_{\rm int})\,.
\end{equation}
\end{enumerate}
It is clear that $\tau_{\rm int} = 1/2$ characterizes uncorrelated
data (see Eq.~(\ref{eq:varobs_tauint})).
The larger the value of $\tau_{\rm int}$ for a particular observable
$O$, the larger the uncertainty \emph{for some ensemble of fixed
length} $N$. The problem is that $\tau_{\rm int}$ has to be estimated from the data itself.
The sum in Eq.~(\ref{eq:tau_rho}) has no upper limit $(i-j) \to \infty$,
while in practice it has to be \emph{truncated} at some finite value.
This means that the estimated value of $\tau_{\rm int}$ will naively
be systematically lower than the correct value (see Fig.~\ref{fig:auto}).
This is not a mere academic observation.
The properties of Markov chain Monte Carlo ensures that the
autocorrelation function Eq.~(\ref{eq:acf}) is a sum of exponentials
\begin{equation}
\Gamma (t) = \sum_{k} A_k e^{-t/\tau_k }\, \xrightarrow[t\to \infty]{}\,
A' e^{-t/\tau_{\rm exp}}\,.
\end{equation}
At large MC times $t\to\infty$, the dominant contribution to the
autocorrelation function is given by the slowest mode of the Markov operator.
This is usually called \emph{exponential autocorrelation time}
($\tau_{\rm exp}$).
Clearly the number of measurements must be large compared with
$\tau_{\rm exp}$ in order to have sensible error estimates and ensure
the ergodicity of the simulation.
What do we know about $\tau_{\rm exp}$ that is relevant for the
lattice determinations of the strong coupling?
\begin{enumerate}
\item At fixed physical volume, one expects $\tau_{\rm exp}$ to
increase proportional to $1/a^2$.
At fine lattice spacing, one need large statistics in order to
estimates the uncertainties correctly\footnote{In practice the
situation might be even more delicate due to a phenomena called
\emph{topology freezing}.
see appendix~\ref{ap:challenges} and the original
works~\cite{DelDebbio:2004xh, Schaefer:2010hu}.}.
\item The values of $\tau_{\rm exp}$ are not very sensitive to the
fermion masses.
In fact even pure gauge simulations show similar values of
$\tau_{\rm exp}$ as simulations with dynamical fermions when similar
algorithms are used.
\item A reasonable estimate of the order of magnitude for $\tau_{\rm
exp}$ can be made by taking $\tau_{\rm exp} \sim 70$
MDU\footnote{MDU stands for Molecular Dynamics Unit, and measure
simulation ``Monte Carlo'' time in HMC simulations.
The typical spacing between measurements in realistic simulations is
between 1 and 2 MDU's. A simulation of 1000 MDUs would allow to have
between 500-1000 measurements.
} at $a\approx$ 0.065 fm for \emph{simulations without topology
freezing} (see~\cite{Bruno:2016plf, Bruno:2014jqa}).
Values at finer lattice spacing can be estimated using an approximate
$a^2$ scaling.
This simple estimate of the order of magnitude is already telling us
that estimating statistical uncertainties for $a\sim 0.03$ fm, where
$\tau_{\rm exp}\sim 350$ requires substantial statistics ({\em i.e.}\,
one would not feel comfortable with less than 4000MDU's).
\end{enumerate}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{fig/auto.pdf}
\caption{Two \emph{replica} of a Monte Carlo simulation with
$\tau_{\rm exp} = 100$. Ensemble 1 has length 20000, while
ensemble 2 has length 500. A long Monte Carlo run allows to
determine the normalized autocorrelation function $\rho(t)\equiv
\Gamma(t)/\sigma^2$ more precisely (see Eq.~(\ref{eq:acf})). This
results in a more solid estimate of the statistical uncertainties.}
\label{fig:auto}
\end{figure}
Let us end this section commenting that the most common analysis
techniques in lattice QCD involves \emph{binning} the data: the Monte
Carlo measurements are averaged in groups of size $N_{\rm bin}$. The
data bins are treated as independent measurements and a naive
error estimate is performed (usually by resampling). This
alternative analysis method does not improve the determinations of the
statistical uncertainties over the methods that directly determine the
autocorrelation function.
It is clear that bins of data are less correlated than the data
itself, but it has been shown that the
decrease in the correlations is slow $\sim 1/N_{\rm
bin}$~\cite{Wolff:2003sm}.
Moreover, in the fairly common case that $N_{\rm bin}$ cannot be taken
much larger than the exponential autocorrelation time $\tau_{\rm
exp}$, there are no known methods to explicitly include the slow
modes of the Markov operator in the binning analysis.
On the other hand a direct analysis of the autocorrelation function
allows to include these effects in the error
estimates~\cite{Virotta2012Critical, Schaefer:2010hu} (see also the
summary in~\cite{Ramos:2018vgu}).
In summary, it is important to point out that in contrast with other
numerical fields where the number of MC samples is very large, lattice
QCD simulations are performed in the uncomfortable situation that the
number of samples is not much larger than the relaxation time of the
Markov operator. In this situation the estimates of statistical
uncertainties can be challenging.
This observation is especially relevant for the determinations of the
strong coupling, since $\tau_{\rm exp}$ scales like $1/a^2$, and fine
lattice spacing are needed in order to study observables at short
distances.
\subsection{Eigenvalues of the Dirac operator}
\label{sec:eigenv-dirac-oper}
Recently a novel approach to extract the strong coupling has been
proposed.
It uses the spectral density of the continuum Dirac operator
\begin{equation}
\rho(\lambda) = \frac{1}{V} \left\langle
\sum_k \left[ \delta(\lambda - \imath \lambda_k) + \delta(\lambda + \imath \lambda_k) \right]
\right\rangle\,,
\end{equation}
and its perturbative expansion
\begin{equation}
\label{eq:rho_pt}
\rho(\lambda) = \frac{3\lambda^3}{4\pi^2} \left(
1 - \rho_1(s) \alpha_{\overline{\rm MS} }(\mu)
- \rho_2(s) \alpha_{\overline{\rm MS} }^2(\mu)
- \rho_3(s) \alpha_{\overline{\rm MS} }^3(\mu)
+ \mathcal O(\alpha_{\overline{\rm MS} }^4)
\right)\,.\qquad
(s = \mu/\lambda)\,.
\end{equation}
A non-perturbative coupling definition can be defined by using
\begin{equation}
\alpha_D(\lambda) = \rho_1(s)^{-1}\left[\frac{4\pi^2}{3\lambda^3}\rho(\lambda) -
1\right]\,.
\end{equation}
with a $\beta$-function known up to 3-loops.
Alternatively one can use the derivative
\begin{equation}
F(\lambda) = \frac{\partial\rho(\lambda)}{\partial\log\lambda}
= 3-F_1(s) \alpha_{\overline{\rm MS} }(\mu)
- F_2(s) \alpha_{\overline{\rm MS} }^2(\mu)
- F_3(s) \alpha_{\overline{\rm MS} }^3(\mu)
- F_4(s) \alpha_{\overline{\rm MS} }^4(\mu)
+ \mathcal O(\alpha_{\overline{\rm MS} }^4)
\end{equation}
to also define the strong coupling whith a perturbative expansion
known up to 4-loops.
Only one work in the literature uses this method to extract the strong
coupling~\cite{Nakayama:2018ubk}. Naively the truncation error at the
energy scales used to extract $\alpha_s$ ({\em i.e.}\
$\lambda \approx 0.8-1.2$ GeV) turns out to be very large ($\sim 20\%$
in $F(\lambda)$). The renormalization scale is pushed to higher
values by using $\mu/\lambda = 5$. For these values of $\mu$
truncation effects are expected to be reduced to a percent level,
while the perturbative expansion for the strong coupling still shows
good convergence properties.
It is important to note that the dominant uncertainty from the
truncation of the perturbative series is determined by an estimate of
the leading missing coefficient in the perturbative series
Eq.~(\ref{eq:rho_pt}). Due to the low energy scale of the
determination, the usual procedure of varying the renormalization
scale by a factor 2 below and above $\lambda$ would result in a
substantially larger truncation error. On the other hand the
work~\cite{Nakayama:2018ubk} does not seem to need any power
corrections to describe the data at energy scales $0.8 - 1.2$ GeV.
The main issue with this novel approach seems to be the very low energy
scales at which the extraction is performed.
These low energy scales are needed in order to be able to extrapolate
the data to the continuum.
The work~\cite{Nakayama:2018ubk} shows that a cut $a\lambda < 0.5$
is needed in order to avoid a substantial
deviation from the leading $\mathcal O(a^2)$ scaling violations.
This restricts the energy scales that can be reached with their data-set
(with lattice spacing $a^{-1} = 2.5, 3.6$ and $4.5$ GeV) to $\lambda <
1.2$ GeV.
This novel method needs further study, in particular at finer lattice
spacing, in order to convincingly show that contact with the
perturbative running is done and to be able to perform a robust
estimate of the truncation effects.
\subsection{The ghost-ghost-gluon vertex}
\label{sec:quark-gluon-vertex}
The definition of the strong coupling using QCD vertices is the one that is more
similar to the type of computation that is usually done in the context of
perturbation theory. There are several issues in the extraction of these QCD
vertices from lattice QCD simulations. In practice the coupling is extracted
from the gluon/ghost two-point functions, but these are not gauge invariant, and
therefore this scheme can only be implemented by fixing the gauge of the lattice
configurations. The problems of gauge fixing \emph{beyond} perturbation theory
({\em i.e.}\ Gribov ambiguities~\cite{Gribov:1977wm}), have been discussed at length in the
literature (see for example~\cite{Vandersickel:2012tz}), and we can
add very little to the discussion, except
pointing out that the issue of Gribov ambiguities is also present in other
lattice QCD calculations. One of the most widely used methods of
non-perturbative renormalization (i.e. ``RI/MOM'' schemes,
see~\cite{Martinelli:1994ty}) also
requires to fix the configurations (typically the
Landau gauge is used). It is believed that at the relatively high
energy scales where $\alpha_s$ is extracted, this is not a serious issue.
In principle there are several options to extract the strong coupling, since it
can be defined from different three- and four-point functions. All the methods
define the coupling by requiring that some vertex is equal to its tree-level
value.
The momenta entering in the vertex are part of the definition of the
scheme. The most popular choices set one of the momenta to zero, and
are usually labeled $\widetilde{\rm MOM}$ schemes.
Nowadays the most common coupling definition uses the un-renormalized
ghost-ghost-gluon vertex, which can be constructed from the gluon and
ghost two-point functions. The non-perturbative coupling definition
reads
\begin{equation}
\label{eq:alpha_Taylor}
\alpha_T(\mu) = \lim_{a\to 0} F_{\rm lat}(p,a)
D_{\rm lat}(p,a)\frac{g_0^2}{4\pi} \Big|_{\mu = p}\,.
\end{equation}
This is usually referred to as the Taylor scheme, as indicated by the suffix.
Here $D_{\rm latt}(p,a)$ and $F_{\rm latt}(p,a)$ are the ``dressing functions''
of the lattice gluon and ghost two-point functions~\footnote{In the continuum
the relation of the propagators $F^{ab}(p), D^{ab}_{\mu\nu}(p)$ with the
dressing functions $F, D$ is
\begin{eqnarray}
F^{ab}(p) &=& -\delta^{ab} \frac{F(p)}{p^2}
\,,\\
D^{ab}_{\mu\nu}(p) &=& -\delta^{ab} \left( \delta_{\mu\nu} -
\frac{p_\mu p_\nu}{p^2} \right)
D(p)\,.
\end{eqnarray}
On the lattice the relation is similar, but the momentum $p$ is substituted by
a function of $p$ and the lattice spacing $a$ that
only reduces to $p$ in the continuum limit and
depends on the particular choice of discretization.} and $g_0^2$ is the bare
coupling used for the simulation. The main advantage of this scheme is that one
does not need to determine any three- or four-point function, since the coupling
is directly defined from the computation of the propagators. We are going to
focus the discussion on this particular choice, although most of what we are
going to state is also valid for other, similarly defined schemes. The
perturbative expansion for $\alpha_T(\mu)$,
\begin{equation}
\label{eq:alphaTpt}
\alpha_T(\mu) \simas{\mu\to\infty} \alpha_{\overline{\rm MS} }(\mu) +
t_1\alpha_{\overline{\rm MS} }^2(\mu) + t_2\alpha_{\overline{\rm MS} }^3(\mu)
+ t_3\alpha_{\overline{\rm MS} }^4(\mu)
+ \mathcal O(\alpha_{\overline{\rm MS} }^5)\, ,
\end{equation}
is known up to three-loops~\cite{Chetyrkin:2000dq}.
There are two important issues that affect generally these extractions and
that are worth mentioning.
\begin{description}
\item[Non perturbative corrections] Several studies have concluded that
including non perturbative corrections in the analysis is mandatory to find
consistent results among observables and range of scales used to determine the
$\Lambda$-parameter (see reference~\cite{Boucaud:2005xn} for a detailed study
in pure gauge and energy scales in the range $3-5$ GeV).
The leading correction comes from
\begin{equation}
\sim \frac{g_0^2 \langle A^2 \rangle }{p^2}\,.
\end{equation}
These corrections are included in the analysis either by using an estimate for
$\langle A^2 \rangle$ (like for example~\cite{Zafeiropoulos:2019flq}) or by
fitting their data including such a term. Higher order non-perturbative
corrections (i.e. $\sim p^{-x}$ for different values of $x$) are also
typically needed to match the lattice data with the perturbative
running~\cite{Blossier:2011tf,Blossier:2012ef,
Blossier:2013ioa,Zafeiropoulos:2019flq} (see Fig.~\ref{fig:alpha_taylor}).
\item[Cutoff effects] The range of scales where $\alpha_T(\mu)$ can be described
by its perturbative expansion (and the non-perturbative contributions) is
typically about $\mu \sim 3$ GeV. Since the typical lattice spacings in
state-of-the-art lattice simulations are in the range $2-5$ GeV, it is really
challenging to make a continuum extrapolation with several lattice spacings in
order to extract the value of the coupling at some scale $\mu$. Therefore the
common approach is to include several terms to either subtract or fit the
cutoff effects (see~\cite{Blossier:2011tf,Blossier:2012ef,
Blossier:2013ioa,Boucaud:2018xup}).
\end{description}
Figure~\ref{fig:alpha_taylor} shows these two points exemplified in the results
of one particular work~\cite{Blossier:2013ioa}. Panel (a) shows that the raw
measurements for the Taylor coupling at two different values of the lattice
spacing differ by approximately $50\%$ after subtracting the $H(4)$ breaking
cutoff effects (see~\cite{Blossier:2013ioa} for details). The remaining scaling
violations have an asymptotic expansion that includes terms $\sim\mathcal O(a^2
p^2)$ and are noticeable compared with the statistical precision of the data.
Panel (b) shows the extraction of the $\Lambda$-parameter after matching with
its perturbative expansion at the scale $\mu$~\cite{Zafeiropoulos:2019flq}. The
plot shows the values of $\Lambda_{\overline{\rm MS} }$ as a function of
$\alpha_{\overline{\rm MS} }^3$ (the leading correction to the extraction). The
data with only $1/\mu^2$ corrections subtracted shows a correction compatible
with a large $\alpha^3$ perturbative term. But if another non-perturbative term
$1/\mu^6$ is included as a fit parameter, the data seem to be well described by
the perturbative expression in Eq.~(\ref{eq:alphaTpt}).
\begin{figure}
\centering
\begin{subfigure}[t]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/1310.3763/raw.pdf}
\caption{}
\end{subfigure}
\begin{subfigure}[t]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/Lambda_vs_alphaT_all.pdf}
\caption{}
\end{subfigure}
\caption{Extraction of the strong coupling from QCD vertices.
(a) Determination of the strong coupling in the Taylor
scheme. The figure shows the raw values of the coupling for
two different lattice spacing ($a\approx
0.089$ fm for $\beta=1.90$ and $a\approx 0.06$ fm for
$\beta=2.10$) and different values
of the quark masses. As the reader can see, the chiral
dependence is mild, but the cutoff effects are substantial.
(source~\cite{Blossier:2013ioa}).
(b)
Extraction of the $\Lambda$-parameter in the Taylor scheme.
Including only $1/\mu^2$ power corrections is not enough to
to reach the perturbative running.
In this particular work~\cite{Zafeiropoulos:2019flq} , the
extraction of $\Lambda$ seems
consistent in the region $0.012<\alpha^3<0.025$ by fitting the
data with an additional $1/\mu^6$ power correction term.}
\label{fig:alpha_taylor}
\end{figure}
In summary, the extraction of $\alpha_s$ from QCD vertices is hampered
by a slow rate of convergence to the perturbative behavior. Several
non-perturbative corrections need to be fitted at the same time, since
they are noticeable. Even after fitting for the non-perturbative
corrections the range of energies that can be used to determine the
strong coupling is limited. Large energies scales are needed, where
most data come typically from a single lattice spacing.
These extractions also show that distinguishing the perturbative and
non-perturbative corrections is, in practice, very difficult.
Figure~\ref{fig:alpha_taylor} (b) shows the difficulty in
distinguishing a correction of order $\alpha^3(\mu)$ from a $1/\mu^6$
non perturbative correction when we have only access to a limited
range of scales. Indeed we see that a variation of the order of 10 MeV
in $\Lambda$ can be reabsorbed by higher-order power corrections.
We believe that a dedicated study in pure gauge theory with the aim of reaching
energy scales where the non-perturbative data is described by the perturbative
prediction \emph{without} fitting any non-perturbative terms would be very
interesting. A pure gauge simulation would also allow a detailed investigation
of the continuum extrapolations using several fine lattice spacing.
\subsection{The static potential}
\label{sec:static-potential}
The force between static color charges has been traditionally one of
the first observables to be studied in lattice
QCD~\cite{Wilson:1974sk}. The potential at distance $r$ can be
extracted from Wilson loops, that behave asymptotically as
\begin{equation}
\mathcal W_{r\times T} \sim \lambda_0^2e^{-V(r)T} + \sum_k
\lambda_k^2 e^{-V_n(r)T}\,.
\end{equation}
The potential $V(r)$, given by the ground state
(i.e. the leading decaying exponential), is formally computed as
\begin{equation}
V(r) = \lim_{T\to \infty} \frac{1}{T} \log \langle \mathcal
W_{r\times T} \rangle\,.
\end{equation}
In practice it is extracted at large, but finite, values of $T$, and therefore
several techniques are needed in order to enhance the overlap with the ground
state and to distinguish the leading exponential from the excited state
contamination. When computed on the lattice, the static potential is power
divergent $\sim 1/a$. This is related to the ambiguity in the overall magnitude
of $V(r)$: only energy differences are physical. The cleanest way to deal with
this linear divergence is to define the coupling via the static force
\begin{equation}
F(r) = \frac{{\rm d} V(r)}{{\rm d} r}\,.
\end{equation}
The derivative with respect to $r$ removes the linear divergence but requires to
perform a numerical derivative of the potential. This is implemented by some
finite difference expression. It is convenient to define the force by
\begin{equation}
F(r_I) = \frac{V(r) - V(r-a)}{a}\,,
\end{equation}
with $r_I$ chosen so that the force has no cutoff effects to leading
order in perturbation theory. This has been shown to reduce the cutoff
effects in the force~\cite{Necco:2001xg}.
A renormalized coupling constant can be defined non-perturbatively
using the static force. The non-perturbative definition of the
coupling and its perturbative expansion in powers of the coupling in
the $\overline{\rm MS}$-scheme reads
\begin{eqnarray}
\label{eq:qqpt}
\nonumber
\alpha_{qq}(\mu) = \frac{4}{3} r^2F(r)\Big|_{\mu = 1/r}
&\simas{r\to 0}&\,
\alpha_{\overline{\rm MS} }(\nu) +
c_{qq}^{(1)}(s)\alpha_{\overline{\rm MS} }^2(\nu)
+ c_{qq}^{(2)}(s)\alpha_{\overline{\rm MS} }^3(\nu)\\
&+&
c_{qq}^{(3)}(s)\alpha_{\overline{\rm MS}
}^4(\nu) + \dots \qquad (s=\mu/\nu)\,.
\end{eqnarray}
The relation of this coupling to the $\overline{\rm MS} $ scheme is known up to
three loops, but the observable suffers from IR divergences, that manifest
themselves in the naive perturbative expansion of eq.~(\ref{eq:qqpt}) being
divergent. These so-called \emph{soft} and \emph{ultra-soft} divergences can be
re-summed and produce logarithmic corrections to the perturbative series. The
leading one is $\propto \alpha_{\overline{\rm MS} }^4(\nu)\log \alpha_{\overline{\rm
MS}}(\nu_{\rm us})$. This re-summation process introduces
an arbitrary energy scale (so called \emph{ultra-soft scale}), its
natural value being $\nu_{\rm us} = \alpha(\nu)/r$.
In principle this means that not only the scale $\nu\approx 1/r$ has
to be large, but also $\nu_{\rm us} = \alpha(\nu)/r$ has
to be large. The additional factor $\alpha$ is not negligible
taking into account that these extractions take place at a few GeV.
All in all, the perturbative expansion, including terms
$\alpha^4,\alpha^4\log\alpha,\alpha^5\log\alpha,\alpha^5\log^2\alpha$, is
known~\cite{Fischler:1977yf,Billoire:1979ih,Peter:1997me,Schroder:1998vy,Brambilla:1999qa,Smirnov:2009fh,Anzai:2009tm,Brambilla:2009bi}
(see~\cite{Tormo:2013tha} for a summary on the perturbative expressions).
Regarding the perturbative behavior, figure~\ref{fig:beta} shows that
the $\beta$-function in this scheme is well behaved, with a
2-loop coefficient of a similar size as in the $\overline{\rm MS} $
scheme.
Several works extract the value of the strong coupling not
from the force, but directly from the potential. In particular they
examine the dimensionless quantity $rV(r)$. Basically the same
considerations apply for these works: the perturbative expansion in
the $\overline{\rm MS} $ scheme is known up to order $\mathcal
O(\alpha_{\overline{\rm MS} }^4)$, and the \emph{soft} and
\emph{ultra-soft} gluons give rise to logarithmic corrections in the
perturbative expansion, starting at order $\alpha_{\overline{\rm MS}
}^4$. The main difference is in the role of the linear divergence in
the potential $V(r)\sim 1/a$.
\begin{figure}
\centering
\begin{subfigure}[t]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/1811.11801/comp_beta7825.pdf}
\caption{}
\end{subfigure}
\begin{subfigure}[t]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/1811.11801/comp_beta7825_diff.pdf}
\caption{}
\end{subfigure}
\caption{(a) Comparison of the perturbative prediction and the
lattice data at the finest lattice spacings ($a\approx 0.04$
fm) for the static energy $E_0(r)$. (b) Residuals of the fit.}
\label{fig:alpha_stat}
\end{figure}
Of course there are many different ways to use the perturbative expansion
Eq.~(\ref{eq:qqpt}) (or a similar expression for the potential
$rV(r)$) to extract $\alpha_s$. Different works, although similar in
spirit, use different approaches to fit the lattice data and deal
with the additive renormalization (in case they use $V(r)$ for the
extraction). Here we will focus on one particular
work~\cite{Bazavov:2014soa} to show the details of such
analysis. Ref.~\cite{Bazavov:2014soa} uses the expression of
the static force and determines the integral up to a reference
distance $r_{\rm ref}$,
\begin{equation}
\label{eq:E0}
E_0(r) = \int_{r_{\rm ref}}^{r}{\rm d} x\, F(x)\,.
\end{equation}
The perturbative expansion for the force gives a similar perturbative expansion
for $E_0(r)$. Note that this quantity is similar to the static potential $V(r)$,
with the difference that it is free of the linear divergence $\sim 1/a$ (the
relation $E_0(r_{\rm ref}) = 0$ is exact for all lattice spacings), and that
depends now on two scales ($r$ and $r_{\rm ref}$). The perturbative expression
for $E_0(r)$ is fitted to the lattice data. Fig.~\ref{fig:alpha_stat} shows the
result of such an analysis. Non-perturbative corrections do not seem to be
needed to describe the data at least for distances in the range $r\lesssim 0.2$
fm. This distance is, nevertheless, large enough to allow the analysis with
several lattice spacings, so the lattice data can be extrapolated to the
continuum. Other recent works include non-perturbative
corrections in the analysis of the lattice data either to claim a
better convergence~\cite{Ayala:2020odx} or to extend the range described by
their fit to $r\lesssim 0.3$ fm~\cite{Takaura:2018lpw}.
In summary, extracting the strong coupling from the static potential
is, in the opinion of the authors, one of the most promising
approaches. The N$^3$LO perturbative knowledge, together with the fact
that data in this scheme seem to follow the perturbative predictions
at scales as low as $1.5$ GeV means that a precise determination can
be achieved. In particular, contrary to other approaches, there is no
need to fit or parametrize any non-perturbative corrections. At least
in theory, the dependence on the lattice spacing, can be accounted
for. Conceptually, the fact that the observable
used is not IR safe is not ideal.
An additional (lower) energy scale enters into the game, and a few
works discuss the best methods to deal with these
corrections. The main criticism to
such extractions, is that their results
depends sensibly on the physics at a scale of just a few GeV, in
particular the range of scales where their assumptions can be checked
(role of the IR divergences, size of cutoff effects,
matching with perturbation theory) is limited. It
would be interesting to check the agreement with the
perturbative running down to higher energies, at least in the pure
gauge theory, where very fine lattice spacings can be
simulated (see reference~\cite{sommer:19lat} for a recent study). We
postpone to section~\ref{sec:pres-future-latt} a more detailed study.
\subsection{Heavy quark correlators}
\label{sec:heavy-quark-corr}
The idea of using correlators of heavy quarks to extract the value of
the strong coupling has its origins in a phenomenological
determination of $\alpha_s$: moments of quarkonium correlators in the
vector channel can be compared with experimental data for $e^+e^- \to
\text{hadrons}$.
As was first noted by the HPQCD collaboration~\cite{Allison:2008xk}, the strong
coupling and the charm quark mass can be extracted on the lattice from
correlators of the pseudoscalar density involving two heavy quarks,
\begin{equation}
G(x_0) = a^6(am_0)^2 \sum_{\mathbf x}
\langle \overline \psi\gamma_5\psi(\mathbf{x},x_0)
\, \overline \psi\gamma_5\psi (\mathbf{0},0)
\rangle\,,
\end{equation}
where $am_0$ is the bare quark mass. This correlator has a short
distance divergence $\sim 1/x_0^3$, but if one uses a fermion formulation
that preserves some chiral symmetry (like staggered or domain wall
fermions), the PCAC relation ensures that moments of
$G(x_0)$, defined as
\begin{equation}
G_n = \sum_{x_0} \left( \frac{x_0}{a} \right)^n G(x_0)\,,
\end{equation}
are dimensionless quantities with a well defined continuum limit~\footnote{With
Wilson fermions an extra finite renormalization for the axial current would be
needed.} for $n\ge 4$. The main contribution to these moments comes from
Euclidean times $x_0 \sim 1/m$.
The extraction of the strong coupling is performed using the reduced
even moments
\begin{eqnarray}
r_4 &=& \frac{G_4}{G_4^{(0)}}\,,\\
r_n &=& \bar m(\mu) \frac{am_{\eta}}{2am_0}\left[
\frac{G_n}{G_n^{(0)}} \right]\,,\qquad (n\ge 6)
\end{eqnarray}
where $G_n^{(0)}$ denotes the leading order prediction for $G_n$ in bare lattice
perturbation theory. This normalization is introduced in order to reduce cutoff
effects (i.e. to leading order, $r_n$ is free of lattice artifacts). Here
$m_\eta$ denotes the mass of the $\eta_c$ meson, and $r_n$ admits a perturbative
expansion that allows an extraction of the strong coupling constant:
\begin{equation}
r_n \simas{\alpha_{\overline{\rm MS}}(\mu)\to 0} 1 +
r_{n,1} \alpha_{\overline{\rm MS}}(\mu) +
r_{n,2}(s) \alpha_{\overline{\rm MS}}^2(\mu) +
r_{n,3}(s) \alpha_{\overline{\rm MS}}^3(\mu) + \dots
\end{equation}
where
\begin{equation}
s = \frac{\mu}{\bar m(\mu)}\,.
\end{equation}
A non-perturbative definition of the strong coupling at the scale $\bar m(\bar
m)$ is given by
\begin{equation}
\label{eq:alphahq}
\alpha_{\rm HQ, n}(\mu)\Big|_{\mu = \bar m(\bar m)} =
\frac{r_n-1}{r_{n,1}} \simas{\alpha_{\overline{\rm
MS}}(\mu)\to 0}
\alpha_{\overline{\rm MS}}(\nu) +
\frac{r_{n,2}(s)}{r_{n,1}} \alpha_{\overline{\rm MS}}^2(\nu) +
\frac{r_{n,3}(s)}{r_{n,1}} \alpha_{\overline{\rm MS}}^3(\nu) + \dots\,,
\end{equation}
with $s=\nu/\bar m(\bar m)=\nu/\mu$. The first three coefficients
($r_{n,1}, r_{n,2}(s), r_{n,3}(s)$) are analytically known
(see~\cite{Maier:2009fz, Maier:2007yn, Boughezal:2006px, Chetyrkin:2006xg}.
Reference~\cite{McNeile:2010ji} has values tabulated for $N_{\rm f} = 3$).
There are two crucial points in these determinations: the estimate of the
truncation uncertainties, and controlling the continuum extrapolations. Note
that these two points have competing interests: the continuum extrapolation is
more easily kept under control for a quantity measured at larger distances, and
therefore the ``high'' moments $r_{6,8,10}$ have milder continuum
extrapolations. On the other hand the truncation uncertainties are smaller for a
short distance quantity: the first moment $r_4$ has a better behavior. This is
just another manifestation of the ``window problem'' (see
figure~\ref{fig:LQCDscales}).
The estimates for these uncertainties vary significantly across different
studies, and we will comment in detail on the issue of the truncation
uncertainties in section~\ref{sec:pres-future-latt}. Here we will focus on the
more technical issue of the continuum extrapolation. Figure~\ref{fig:hqcorr}
shows the continuum extrapolation of $r_4$ at scale $m_{\rm c}$ of the
works~\cite{Maezawa:2016vgv,Petreczky:2019ozv}. As the reader can see, the scaling
violations are significant and have a complicated functional form. Different
works in the literature deal with these complicated cutoff effects in very
different ways
\begin{description}
\item [JLQCD collaboration~\cite{Nakayama:2016atf}:] In this case they
prefer to only perform extrapolations linearly in $a^2$.
Their data for $r_4$ does not allow such an extrapolation, and
therefore it is excluded from their analysis.
\item [HPQCD collaboration~\cite{McNeile:2010ji,Chakraborty:2014aca}:]
In these works all moments are used, and masses above the charm quark mass are
used, including data with $am\sim 0.9$. Cutoff effects are large and the data
is contaminated by effects $\sim (am_{\rm c})^{2p}$. Therefore their fit
Ansatz includes terms $a^{2p}$ with $p$ up to 10. These fits typically have
more terms than data, and require to include an estimate of the size of these
coefficients as Bayesian priors.
\item [Ref.~\cite{Petreczky:2019ozv}:] Here energy scales
larger than the physical charm quark mass are explored, but the continuum
extrapolations are difficult and the data usually has associated large
uncertainties (see figure~\ref{fig:hqcorr}).
\end{description}
\begin{figure}
\centering
\begin{subfigure}[t]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/1901.06424/r4.pdf}
\caption{}
\end{subfigure}
\begin{subfigure}[t]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/1901.06424/lambda.pdf}
\caption{}
\end{subfigure}
\caption{(a) Continuum extrapolation of $r_4$ at the lower energy
scale $\mu\sim \bar m_c$. Scaling violations are significant, even
at the smaller quark masses used in the study. The extrapolations
is performed using both a 5$^{\underline{th}}$ degree polynomial,
or a second degree one with a restricted fitting window
(source~\cite{Petreczky:2019ozv}). (b)
Dependence of the
$\Lambda_{\overline{\rm MS} }$
parameter extacted using heavy quark correlators on the value of
the coupling at the matching scale with perturbation theory
($\alpha_{\overline{\rm MS} }(\mu)$) (source~\cite{Petreczky:2019ozv}).}
\label{fig:hqcorr}
\end{figure}
The main drawback of this approach is the large cutoff effects that
affect the quantity used to extract the strong coupling. This makes it
very challenging to explore energy scales larger than the physical
charm quark mass $m_c\sim 1.4$~GeV, which is not particularly
large. The recent work in Ref.~\cite{Petreczky:2019ozv} explores
different energy scales in the range $\bar m_c - 3\bar m_c$, but the
continuum extrapolation is very challenging already at
$\mu\gtrsim 2m_c$~\footnote{Note that the leading cutoff effects in
this approach are $\mathcal O(a^2\alpha_s)$, while in other
extractions (like for instance the one based on the static
potential) they are suppressed by an extra factor of
$\alpha_s$.}. Even at the scale of the charm quark mass, scaling
violations are significant and have a complicated functional form (see
Figure~\ref{fig:hqcorr}). Together with the fact that the perturbative
relation is known only to NNLO the situation is far from ideal:
truncation uncertainties at the energy scales reached by current
simulations are not small (see detailed discussion in
section~\ref{sec:heavy-quark-corr-1}). This might change in the
future, as smaller lattice spacings can be simulated, allowing a
reduction of the discretization effects. A detailed study in pure gauge and
reaching energy scales significantly larger than the charm quark mass,
would probably give very important information on the systematics of
this method.
\subsection{Observables defined at the scale of the cutoff}
\label{sec:observ-defin-at}
Lattice QCD offers the interesting possibility of extracting the
strong coupling from expectation values computed at non-zero lattice
spacing. This approach is fundamentally different from the strategies
outlined above, where a quantity is computed in QCD (i.e. extrapolated
to the continuum), and then compared with a perturbative
prediction. The observables that we are going to discuss in this
subsection are defined at a scale given by the lattice spacing
$1/a$. Lattice bare perturbation theory is able to relate these purely
lattice observables with a power series in the renormalized
coupling. The usual problem that a naive approach has to face is that
bare lattice perturbation theory is just terrible. Absurdly small
values of the lattice coupling $\alpha_{\rm latt} = g_0^2/4\pi$ have
to be used in order to reach the domain of apparent convergence. It
has been argued that this apparent failure of lattice perturbation
theory is just a due to the choice of the bare coupling $g_0^2$ as
expansion parameter~\cite{Lepage:1992xa}. If lattice quantities are
expressed as a perturbative series in a \emph{renormalized} coupling,
like $\alpha_{\overline{\rm MS} }(\mu)$, their perturbative behavior
improves substantially (see~\cite{Lepage:1992xa}).
The HPQCD collaboration has pursued the systematic study of several Wilson loops
of size $n\times m$ (denoted $\langle W_{nm} \rangle$) and used them to extract
the value of the strong coupling. In these analyses lattice quantities are
expressed as a perturbative series in terms of the renormalized coupling
$\alpha_V(\mu)$; for an SU(3) gauge theory coupled to $N_f$ fermions in the
fundamental representation, the latter coupling is defined
by~\cite{Schroder:1998vy}
\begin{equation}
\label{eq:alphaV}
\alpha_{\rm V}(\mu) = \alpha_{\overline{\rm MS}}(\bar \mu) + \frac{2.6 - 0.3 N_{\rm f} }{\pi}\alpha_{\overline{\rm MS}}^2(\bar \mu)
+ \frac{53.4 - 7.2 N_{\rm f} + 0.2 N_{\rm f} ^2}{\pi^2}
\alpha_{\overline{\rm MS}}^3(\bar \mu)\,,
\end{equation}
with $\bar \mu = \exp(\gamma-5/6)\mu \approx 0.774\times \mu$. Wilson
loops have a perturbative expansion
\begin{equation}
\label{eq:wmn}
- \log \langle W_{nm} \rangle \simas{a\to 0}
w_1 \alpha_{\rm V}(\mu) +
w_2 \alpha_{\rm V}^2(\mu) +
w_3 \alpha_{\rm V}^3(\mu) + \dots \,.
\end{equation}
Alternatively one
can use Creutz ratios or tadpole improved Wilson loops. This latter choice
\begin{equation}
\label{eq:wmnu}
- \log \left( \frac{\langle W_{nm} \rangle}{\sqrt{\langle W_{11}
\rangle^{n+m}}} \right) \simas{a\to 0}
w_1^{\rm b} \alpha_{\rm V}(\mu) +
w_2^{\rm b} \alpha_{\rm V}^2(\mu) +
w_3^{\rm b} \alpha_{\rm V}^3(\mu) + \dots \,.
\end{equation}
is supposed to lead to smaller truncation uncertainties. In all cases
the perturbative coefficients
are known for several choices of $n, m$~\cite{Mason:2005zx}. The scale is given
by $\mu = d/a$, where $d\approx \pi$ with the exact value depending on the
choice of Wilson loop. Note that non-perturbative couplings can be defined by
expressions
\begin{equation}
\label{eq:alphaW}
\alpha_{W_{nm}}(1/a) = - \frac{\log \langle W_{nm} \rangle }{w_1}
\simas{a\to 0} \alpha_{\rm V}(\mu) +
\frac{w_2 }{w_1} \alpha_{\rm V}^2(\mu) +
\frac{w_3 }{w_1} \alpha_{\rm V}^3(\mu) + \dots \,.
\end{equation}
Several quantities are fitted to the previous perturbative
expressions, with the value of $\alpha_{\rm V}(\mu_{\rm ref})$ as fit
parameter. The values of $\alpha_{\rm V}(\mu)$ at other scales are
obtained from $\alpha_{\rm V}(\mu_{\rm ref})$ and the RG equation
\begin{equation}
\mu^2 \frac{{\rm d} \alpha_{\rm V}(\mu)}{{\rm d} \mu^2} =
-\alpha_{\rm V}^2\sum_{k = 0}^3 \beta_k\alpha_{\rm V}^k\,.
\end{equation}
The final value of $\alpha_{\rm V}(\mu_{\rm ref})$ can be converted to
the more convenient $\overline{\rm MS} $ scheme using Eq.~(\ref{eq:alphaV}).
An example of such extractions are the HPQCD works
(see~\cite{Allison:2008xk}). A
total of 22 quantities (all like Eq.~(\ref{eq:wmn})) are fitted to
their perturbative expression. Unfortunately the truncated
perturbative expression Eq.~(\ref{eq:wmn}) does not describe the data
well, and several extra terms (up to $\alpha_{{\rm
V}}^{10}$) are necessary in order to obtain a sensible fit. The
coefficients of these terms are constrained with Gaussian priors,
which eventually lead to stable fits, with a consistent determination
of the strong coupling using any of the 22 quantities (although some
of them require to also fit several power corrections).
The main advantage of methods based on observables defined at the
cutoff scale, is that high energies can be reached without having to
worry about the continuum extrapolation. The statistical accuracy is
excellent, since the observables entering the determination
Eq.~(\ref{eq:wmn}) have a very small variance. On the other hand the
uncertainty in these determinations is dominated by the truncation of
the perturbative series. The fact that several higher order terms have
to be fitted (and constrained with Gaussian priors) in order to
describe the data is not ideal. It is clear that expressing lattice
quantities as a power series in \emph{renormalized couplings}, as
suggested in~\cite{Lepage:1992xa} greatly improves the predictive
power of perturbation theory (bare perturbation theory is just useless).
Still these lattice observables are far from ideal from a perturbative
point of view: even if energy scales $1/a\approx 4$ GeV are reached,
perturbation theory does not predict the lattice data and truncation
uncertainties are not small (see detailed discussion in
section~\ref{sec:observ-at-cutoff}).
Another delicate point in this approach is that cutoff
effects have the same functional form as the non-perturbative effects
(power corrections) in the expansion of $\alpha_P(1/a)$. This can be
easily understood by noting that
\begin{equation}
a^2 \sim \exp\left\{- \frac{4\pi}{2b_0\alpha_W(1/a)} \right\}\,,
\end{equation}
Of course in any extraction of the strong coupling based on these methods,
these effects are not parametrically the leading ones, since the
truncation of the perturbative series Eq.~(\ref{eq:alphaW}) misses
terms of order $\mathcal O(\alpha_W^n) \sim \log^n a$. However in
practice it
is not clear which effects dominate (the $\mathcal O(a^2)$ cutoff effects
or the $\log^n a$ from the truncation of the perturbative series), and
this might even depend on the particular observable used to set the scale.
\subsection{The hadron vacuum polarization}
The hadronic vacuum polarization function (HVP) is defined from two-point
functions of the vector and axial-vector currents
\begin{eqnarray}
V_\mu^a(x) &=& \bar \psi_a\gamma_\mu \psi_a(x)\,, \\
A_\mu^a(x) &=& \bar \psi_a\gamma_5\gamma_\mu \psi_a(x)\,,
\end{eqnarray}
after a decomposition in Fourier space (with $J_\mu= V_\mu, A_\mu$)
\begin{equation}
\int \mathrm{d} ^4x\, e^{\imath p x}\, \langle J_\mu^a(x) J_\nu^a(0)
\rangle =
(\delta_{\mu\nu}p^2 - p_\mu p_\nu)\Pi_J^{(1)}(p^2) - p_\mu
p_\nu\Pi_J^{(0)}(p^2) \,.
\end{equation}
The quantity
\begin{equation}
\Pi(p^2) = \Pi_{V}^{(0)}(p^2) + \Pi_{V}^{(1)}(p^2) + \Pi_{A}^{(0)}(p^2) + \Pi_{A}^{(1)}(p^2)\,.
\end{equation}
is dimensionless and has a perturbative expansion
\begin{equation}
\label{eq:hvppt}
\Pi(p^2) \simas{p\to\infty} c_0
+ \sum_{k=1}^4c_k(s)\alpha^k_{\overline{\rm MS}}(\mu) + \mathcal O(\alpha^5_{\overline{\rm MS} })\,.
\qquad (s=p/\mu)\,.
\end{equation}
known up to 5-loops. The constant term $c_0(s)$ is divergent, so that the strong
coupling is usually extracted from the difference $\Pi(p^2) - \Pi(p^2_{\rm
ref})$, or the Adler function
\begin{equation}
D(p^2) = p^2 \frac{{\rm d} \Pi(p^2)}{{\rm d} p^2}\,.
\end{equation}
The recent work~\cite{Hudspith:2018bpz} determines the finite
difference
\begin{equation}
\Delta(p^2,p^2_{\rm ref}) = \frac{\Pi(p^2) -\Pi(p^2_{\rm
ref})}{\log(p/p_{\rm ref})} \,.
\end{equation}
at high energies in order to make contact with the perturbative running. They
use several values of $p\sim 2-4$ GeV and different fit procedures, ranges of
$p$ and values of $p_{\rm ref}$ to extract $\alpha_{\overline{\rm MS} }(M_Z)$.
The main issue with extractions based on the HVP is that power
corrections are significant even for large
momenta~\cite{Shintani:2010ph}. In fact, as discussed in detail in
Ref.~\cite{Hudspith:2018bpz}, these corrections show a very poor
convergence. One would expect that higher power corrections become
negligible as the momentum increases, but this is not the case due to
accidental cancellations between condensates of different dimensions
(up to $1/\mu^8$). Ref.~\cite{Hudspith:2018bpz} pushes the
determination to high energies, so that the data can be described
without any power corrections, but then cutoff effects become larger
and the window of scales to obtain the strong coupling decreases.
Despite the impressive perturbative knowledge in this scheme (5
loops), the authors think that more work is needed in order to
convincingly show that contact with the perturbative running has been
made, and that the continuum extrapolations are under control. Being a
relatively new technique, there is not a single work for the pure
gauge theory. Once again, we would like to stress that a detailed
study in this simpler case, where very fine lattice spacings can be
simulated, would shed some light on many of these issues.
\input{obs_dirac.tex}
\subsection{Determinations of $\alpha_{\overline{\rm MS}}(M_Z)$}
How is the value of the strong coupling constant extracted from experimental
data? The generic procedure can be sketched as follows. Broadly speaking, the
experimental results for a physical process $P(Q)$ at high energies $Q$ are
compared with the perturbative prediction (typically available up to some order
$n$),
\begin{equation}
\label{eq:Pseries}
P(Q) = \sum_{k=0}^n c_k(s) \alpha_{\overline{\rm MS}}^k(\mu) + \mathcal
O(\alpha_{\overline{\rm MS}}^{n+1}(\mu)) +
\mathcal O \left( \frac{\Lambda^p}{Q^p} \right) \,,
\qquad (s=\mu/Q)\,.
\end{equation}
Several subtle points are involved in this comparison. First we
should notice that the coefficients $c_k(s)$ grow logarithmically with
$s$, and therefore the renormalization scale $\mu$ has to be chosen
close to the physical scale of the process $Q$, in order to avoid
large logarithms and a poorly converging perturbative series. We
should also note that once $\alpha_{\overline{\rm MS} }(\mu)$ is known
at some energy scale $\mu\sim Q$, one can use the 5-loop
$\beta$-function in the $\overline{\rm MS}$ scheme to ``run'' this
result either to a common reference scale (i.e. $M_Z$), or up to
infinite energy and quote the value of the
$\Lambda_{\overline{\rm MS}}$ parameter. The considerations raised in
section~\ref{sec:chall-determ-alph} also apply to the determinations
that follow this approach. In this case the renormalization scale
$\mu = sQ$ plays the role of $\mu_{\rm PT}$: the energy scale at which
we match with perturbation theory. One would like to extract
$\alpha_{\overline{\rm MS} }(M_Z)$ (or $\Lambda_{\overline{\rm MS}}$)
by using data at several values of $\mu = sQ$, and take as the final
result a suitable extrapolation $sQ\to\infty$. Since the value of $s$
cannot be taken to be arbitrarily large, such a procedure requires
data at several values of the physical scale $Q$ in order to have a
real constraining power on the value of the coupling.
In Eq.~\eqref{eq:Pseries} we show the two types of corrections present in the
perturbative expansion of a physical quantity. First the {\em missing higher
orders}, due to the fact that we only know a finite (typically $n = 2, 3$)
number of terms in the perturbative expansion of the observable. Second,
non-perturbative corrections (usually called \emph{power corrections}). These
are of the form $e^{-A/\alpha_{\overline{\rm MS}}(Q)} \sim \mathcal
O\left(\frac{\Lambda^p}{Q^p} \right)$ with $p=2A\, b_0$ and decrease faster
than any power of $\alpha_{\overline{\rm MS}}$. In order to keep the truncation
and non-perturbative corrections small, the chosen process should be ideally
inclusive and defined at high enough energies. High energy scales ensure that
$\alpha_{\overline{\rm MS}}(Q)$ is small. Inclusive measurements do not require
a quantitative description of the strong interactions of hadronic states and
therefore are less affected by systematic errors coming from models of
hadronization and parton showers.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{fig/alpha_err_scale.pdf}
\caption{Error of the coupling at a scale $\mu$ compared with the
error propagated to the reference scale $M_Z$.
Note that when the strong coupling is determined at low energy
scales the result at the reference scale $M_Z$ becomes more precise.
For example, the error in $\alpha_{\overline{\rm MS} }(M_\tau)$ is
reduced by almost one order of magnitude when the result is
evolved to the scale $M_Z$ (solid purple line).
This effect is \emph{not} only due to the reduction in the coupling
itself: the relative error in the coupling is reduced
approximately by a factor three (dashed green line).}
\label{fig:alpha_err_scale}
\end{figure}
Obtaining a precise value for the strong coupling with high energy
experimental input has its own challenges. At high energies the strong
coupling is small, which is just what is needed to have the truncation
and power corrections under control, but at the same time the effect
that one is trying to measure is small. This usually translates in
larger uncertainties in $\alpha_{\overline{\rm MS}}(M_Z)$ from
determinations based on data at high energies (see
figure~\ref{fig:alpha_err_scale}). Extracting the value of
the strong coupling at lower energies usually leads to smaller
uncertainties, although the estimation of the truncation uncertainties
and the non-perturbative effects become more challenging: clearly the
extrapolation $sQ\to\infty$ is more difficult without data at large
$Q$.
Another point to take into account is that in contrast with the
perturbative computations, quarks are not the observed final states of
any physical process. Hadronization and other non-perturbative effects
have to be taken into account when comparing experimental data with
perturbative predictions, usually by using Monte Carlo
generators.
\paragraph{Extraction from data}
A well-know example of the extraction of the strong coupling from
experimental data is the extraction of
$\alpha_{\overline{\rm MS}}(M_Z)$ using data for $\tau$ decaying into
hadrons. We briefly summarise the procedure here, in order to
highlight the main steps and the sources of uncertainties, we refer
the reader to an extensive review like e.g. Ref.~\cite{Pich:2013lsa}
for a detailed discussion. The physical processes considered in this
case are the decays of $\tau$ leptons. More specifically, the ratio of
the hadronic and leptonic decay widths can be written as
\begin{equation}
\label{eq:tau_decays}
R_{\tau, V+A} = \frac{\Gamma(\tau \to \nu_\tau + \text{hadrons})}
{\Gamma(\tau \to \nu_\tau e^-\bar\nu_e)} =
3|V_{ud}|^2 S_{\rm EW}\left(1 + \delta_{P} + \delta_{\rm NP}\right)
\, .
\end{equation}
In this case the typical energy scale of the process is set by the
$\tau$ mass $Q=M_\tau = 1.77682(16)$ GeV. In
Eq.~\eqref{eq:tau_decays} $S_{\rm EW}(Q)$ is the electroweak
contribution to $R_{\tau, V+A}$, and $\delta_{\rm P}, \delta_{\rm NP}$
are the QCD perturbative and non-perturbative corrections to the
process respectively. The non-perturbative (i.e. power) corrections
are estimated to be very small
$\delta_{\rm NP}\sim 10^{-4}$~\cite{Pich:2013lsa}. On the other hand
the perturbative prediction
\begin{equation}
\label{eq:Rpt}
\delta_{\rm P} = \sum_{k=1}^4 r_n(s)\,
\left(\frac{\alpha_{\overline{\rm MS}}(\mu)}{\pi}\right)^k +
\mathcal O(\alpha_{\overline{\rm MS}}^5)
\, , \qquad (s=\mu/M_\tau)\,.
\end{equation}
is known up to four
loops~\cite{Baikov:2008jh,Baikov:2012er,Baikov:2012zn}. The impressive
perturbative knowledge in
the ratio $R_{\tau, V+A}$ makes this quantity a good candidate to
determine $\alpha_{\overline{\rm MS}}$. In fact such determinations
are one of the most precise phenomenological determinations. On the other
hand the scale at which $\alpha_{\overline{\rm MS}}$ is determined is
relatively low, and it cannot be changed, since $M_\tau$ is what it
is.
The procedure with other observables is basically the same, although
some details, like the number of known terms in the perturbative
expansion or the size of the non-perturbative effects
(i.e. $\delta_{\rm NP}$ in Eq.~\eqref{eq:tau_decays}), change from one
observable to another.
Combining multiple collider observables in a global fit provides a
better lever-arm to constrain $\alpha_s$ together with the parton
distribution functions (PDFs). Global fits that include the wider
ranges of data provide determinations of the strong coupling constant
with good statistical accuracy, see {\em e.g.}
Refs.~\cite{Ball:2011us,Ball:2018iqk,Harland-Lang:2015nxa}. The
challenges here stem from controlling the systematic errors (both
theoretical and experimental) in fits that involve very large and
diverse datasets and the relatively low energies involved (see for
example the recent review\cite{Salam:2017qdl}). Moreover,
as recently discussed in
Ref.~\cite{Forte:2020pyp}, determinations of the strong coupling from
hadronic processes should entail a simultaneous determination of the
parton distribution functions.
\paragraph{Extraction from lattice simulations}
Lattice QCD offers an interesting alternative to phenomenological
determinations. Being a non-perturbative formulation of QCD, one can
combine input from well-measured QCD quantities -- like for example
the proton mass, or a meson decay constant -- with the perturbative
expansion of a short distance observable that does not need to be
directly observable (like the quark anti-quark force). The advantage
of this approach is that the experimental input comes from the hadron
spectrum with a negligible uncertainty. Hadronization corrections are
not needed, since we are working directly in a non-perturbative
framework.
\begin{figure}[t!]
\centering
\includegraphics[width=\textwidth]{fig/LQCDscales.pdf}
\caption{Scales in a typical state of the art lattice
computation. The volume of the simulation is a few fm, while the
cutoff $1/a$ is a few GeV. Lattice QCD can resolve observables in
this window of scales, where the strong coupling is $\alpha_{\overline{\rm MS}} \sim
0.25$.}
\label{fig:LQCDscales}
\end{figure}
Despite being very different approaches, both the phenomenological and
the lattice QCD methods have to overcome similar challenges. Lattice
extractions of the strong coupling must be done at sufficiently high
energies so that truncation and power corrections are well under
control when matching to perturbative expansions. Due to the finite
nature of computer resources, every lattice QCD simulation has two
intrinsic scales: the total physical volume simulated $L$ (IR cutoff),
usually of a few fermi in order to keep finite-volume corrections well
under control, and the lattice spacing $a$ (the UV cutoff $\sim 0.04$
fm in the most challenging present day simulations, which corresponds
roughly to a cutoff of 5~Gev in energies). Any lattice QCD simulation
can only resolve a process if it is defined at a scale between these
IR and UV cutoffs (see figure~\ref{fig:LQCDscales}). The number of
lattice points in each direction is given by the ratio $L/a$, viz. the
separation of the UV and IR cutoffs determines the memory footprint
and computing power,
and hence the computational cost, of the corresponding simulations,
putting in practice a limit on the energy scales that can be studied
in any lattice simulation. While in principle lattice techniques can
be used to compute non-perturbatively the running of the coupling
until the perturbative regime is reached, in practice the range of
scales that can be studied \emph{in a single lattice simulation} is
limited by computer resources. Reaching scales higher than a few GeV
requires a dedicated approach.
\subsection{Systematics in the extraction of
$\alpha_{\overline{\rm MS}}$}
\label{sec:syst-extr-alph}
The truncation performed in Eq.~\eqref{eq:Pseries} neglects higher
order terms (i.e. perturbative corrections) and non-perturbative power
corrections. When performing an extrapolation to $Q\to\infty$, these
effects only affect \emph{how fast} we approach the extrapolated
value. An example of this behaviour can be seen in
figure~\ref{fig:msbar}, where different observables (labeled by $\nu$)
are used to estimate $\Lambda_{\overline{\rm MS} }$ by matching with
perturbation theory at different physical scales $Q$, which are
translated in different values of $\alpha_{\overline{\rm MS} }$ in the
plot. Different observables predict compatible results for
$\Lambda_{\overline{\rm MS} }$ when the extrapolation $Q\to\infty$,
corresponding to $\alpha\to 0$ in the plot, is performed. The results
of these extrapolations agree well within errors with the result
quoted in Ref.~\cite{Brida:2016flw} (gray error band in the
plot). Note however that some observables ({it viz.}\, $\nu=-0.5$)
show a slow approach to the extrapolated value, with significant
discrepancies even at energy scales $Q\sim 8-10$ GeV.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{fig/msbar2.pdf}
\caption{Determination of the $\Lambda$-parameter in units of a
scale $L_0\sim 1/(4 {\rm GeV})$. Different values of $\nu$
represent different choices of observable, and each point
corresponds to a determination of the dimensionless product
$\Lambda_{\overline{\rm MS}} L_0$ using a perturbative expansion
(like Eq.~(\ref{eq:Pseries})).
The horizontal axes label the scale of matching with perturbation
theory ($\mu$ in Eq.~(\ref{eq:Pseries})) parametrized in terms of
the leading corrections $\mathcal O(\alpha_{\overline{\rm MS}
}^2(\mu))$. The first error bar shows the
statistical uncertainty. The second error bar shows the total
systematic plus statistic added in quadratures. The systematic
uncertainty is determined by varying the renormalization scale a
factor of two above and below some specific value. For $\nu=-0.5$
the systematic uncertainty determined with this method is unable
to account for the difference with the final results at values of
$\alpha\sim 0.13-0.2$, while in other cases ($\nu = 0.3$) the
systematic uncertainty overestimates the true
difference. See text for more details.
(source~\cite{Brida:2016flw, DallaBrida:2018rfy}). }
\label{fig:msbar}
\end{figure}
In practice performing the extrapolation $Q\to\infty$ is very
difficult.
Data over a large range of energy scales is required in order to
perform such an analysis.
For example the data in figure~\ref{fig:msbar} involves precise lattice
determinations of the target observables for energy scales $Q\in
2-140$ GeV. This is only possible with a dedicated approach (see
section~\ref{sec:finite-size-scaling}).
How do we estimate the systematics effects in the
extraction of $\alpha_{\overline{\rm MS}}$ when the data does not
allow an extrapolation $Q\to \infty$?
Of course this is a complex subject in itself, that is of much
relevance not only for the extraction of $\alpha_{\overline{\rm MS}}$,
but also in the interpretation of many experimental data from hadron
colliders. A possible estimate of the uncertainty due to the
missing terms is given by the last known term in the series
$c_n(s) \alpha_{\overline{\rm MS}}^n(\mu)$. Generally this results in
significant theoretical uncertainties, and, perhaps more
interestingly, correlations between experimental data.
A \emph{common} approach to estimate these uncertaintites
exploits the fact that the truncated perturbative expansion to $n^{\rm
th}$ order
\begin{equation}
\label{eq:TruncObsP}
P^{(n)}(Q, s) = \sum_{k}^n c_k(s) \alpha_{\overline{\rm MS}}^k(\mu)\,,
\qquad (s=\mu/Q)\,.
\end{equation}
still depends on $\mu$, while the true value of the observable $P(Q)$
does not. The truncation of the perturbative expansion introduces a
spurious dependence on the renormalization scale, which is in general
an unphysical, arbitrary quantity. Higher-order effects are estimated
by looking at the variation of $P^{(n)}(Q, s)$ when the
renormalization scale $\mu$ is changed by a factor two around some
preferred value (for example $\mu=Q$). In principle the relation
between a variation in the renormalization scale $\mu$ and the size of
the missing higher-order terms given by
$\delta_n=|P(Q) - P^{(n)}(Q,s)|$ is unclear, beyond the fact that the
scale dependence in Eq.~\eqref{eq:TruncObsP} is due to the truncation
of the perturbative expansion. Under some assumptions on the size of
the coefficients of the perturbative expansion
($c_{n+2}(s)\alpha(\mu)\ll c_{n+1}(s)$), it is possible to show
that the scale variation yields a sensible estimate of $\delta_n$ (see
for example Ref.~\cite{Cacciari:2011ze}). Formally,
\begin{equation}
\label{eq:truncation_1}
\mu \frac{{\rm d} P^{(n)}(Q,s)}{{\rm d} \mu} \propto \alpha_{\overline{\rm MS}}^{n+1}(\mu)\,,
\end{equation}
which implies that, at least parametrically, changes in $\mu$ capture the
correct size of the missing terms. As an example, a recent comprehensive study
of the theoretical uncertainties for numerous observables, based on scale
variations, can be found in
Refs.~\cite{AbdulKhalek:2019bux,AbdulKhalek:2019ihb}.
What about power corrections? they are not captured by this kind of analysis.
Estimating them requires to have access to different physical scales $Q$.
Ideally one would like to work at sufficiently high energies so that they are
negligible compared with the accuracy of the data. In practice this is not
always the case. Note that the perturbative running is logarithmic, and
distinguishing this perturbative running from a power-like behaviour requires
data that span large energy ranges.
The assumptions that underlie the scale variation procedure constrain both the
non-perturbative effects and the character of the perturbative series. In
particular, the assumption that the first unknown term of the
perturbative series is smaller than the last known one is implicit in any
estimate that uses Eq.~\eqref{eq:truncation_1}. Also the value of
$\alpha_{\overline{\rm MS}}(\mu)$ is assumed to be small enough so that these
uncertainties are meaningful. These assumptions might seem reasonable and mild,
and often yield sensible estimates, but there are examples in the literature
where they have been shown not to be accurate. Let us mention here
three relevant cases.
\begin{itemize}
\item The convergence of the perturbative series in practice is not as
good as we would like. Due to the asymptotic nature of the PT
series, one expects that at some point the coefficients in
the perturbative series will grow factorially, see
Ref.~\cite{Beneke:1998ui} for a review.
\item Extractions of the strong coupling from $\tau$ decays can be done by
applying two frameworks in perturbation theory, called fixed order
perturbation theory (FOPT) and contour improved perturbation theory (CIPT).
Using $\alpha_{\overline{\rm MS}}(m_\tau)=0.34$ as the typical value of the
strong coupling at the scale set by the mass of the $\tau$, the contributions
to both perturbative series look as follows\footnote{The
perturbative series has the form
\begin{equation}
\delta = \sum_n(K_n + g_n)\left(\frac{\alpha}{\pi}\right)^n
\end{equation}
where the $K_n$ coefficients are the same in FOPT and CIPT, while
the $g_n$ coefficients are different in both formulations.
The coefficient $K_5$ is unknown. Here we use an estimate $K_5=275$.
Note that this does not affect the difference in $\delta$
between both formulations.}
\begin{eqnarray}
\delta_{\rm FOPT} &=&
0.1082 + 0.0609 + 0.0334 + 0.0174 + 0.0087 = 0.2286\,,\\
\delta_{\rm CIPT} &=&
0.1479 + 0.0297 + 0.0122 + 0.0086 + 0.0037 = 0.2021\,.
\end{eqnarray}
The terms in both series decrease, and each of the expansions by themselves
seem reasonable. Taking the last term as a measure of the uncertainty in the
truncation, these values should be accurate with a precision $\sim 0.01$. But
both approaches result in values that differ by more than twice this amount.
What is more worrisome, for the highest orders the difference
between both estimates \emph{grows} as more terms are included in
the expansion.
\item The scale variation approach to estimate truncation uncertainties
(changing the value of the renormalization scale $\mu$ by a factor two around
a preferred value) has been compared with non-perturbative data in a careful
study for values of $\alpha_{\overline{\rm MS}}\sim
0.1-0.2$~\cite{DallaBrida:2018rfy} (see figure~\ref{fig:msbar}). The error
bars in figure~\ref{fig:msbar} include an estimate of the truncation
uncertainty using this approach. It is clear from the plot that for $\nu=-0.5$
the scale variation underestimates the systematic error. Note that the known
terms in the perturbative series for all these observables suggest an apparent
good perturbative behavior.
\end{itemize}
These examples show that estimating the truncation uncertainties within
perturbation theory is difficult. In the absence of a theorem, our attempts to
quantify these uncertainties remain exploratory, and caution should be exercised
in interpreting the results. One should never forget the asymptotic nature of
perturbative series in QCD~\cite{tHooft:1977xjm}. Eventually a \emph{factorial
growth} of the size of the terms in the series is expected, which is deeply
related to the non-perturbative effects of the theory. Note however that the
structure of the corrections have been investigated at length (see
{\em e.g.}~Ref.~\cite{Beneke:1998ui}). It is clear
that there are at least two ingredients in the quality of any
extraction of the strong coupling.
\begin{enumerate}
\item The value of $\alpha_{\overline{\rm MS}}$ at which perturbation theory is
used. Non-perturbative (power) corrections decrease very quickly with
$\alpha_{\overline{\rm MS}}$. In order to make them negligible one needs to
have access to high energy scales, hence, small values of
$\alpha_{\overline{\rm MS}}$.
\item The extraction has to be performed over a range of values of
$\alpha_{\overline{\rm MS}}$. This allows the unknown terms in the series to
vary substantially, so that one can check that indeed they are negligible. A
reasonable requirement would be that the first unknown term
$\alpha_{\overline{\rm MS}}^{n+1}$ varies significantly, say by a factor four.
\end{enumerate}
These are two criteria that are usually relevant in any extrapolation. As
shown in the examples above, they will impact the quality of any extraction of
the strong coupling. Of course some determinations cannot really change the
value of the momentum scale at which perturbation theory is used. A
good example is the above
mentioned extraction from $\tau$ decays, since the mass of the $\tau$ is what
it is and sets the overall energy scale to the process. For the case of
lattice simulations, changing the values of $\alpha_{\overline{\rm MS}}$ at
which perturbation theory is used is challenging, yet feasible. We shall keep
these two criteria in mind when describing any lattice computation/method, and
not only the quoted theoretical uncertainties.
We would like to end this section recommending the reader the recent
contribution to the Lattice Field Theory Symposium by M.
Dalla Brida~\cite{DallaBrida:2020pag}, where these issues are also
discussed in detail.
\subsection{The case of the Schr\"odinger Funcional coupling in full detail}
In this section we describe in full detail our procedure for estimating the
impact of scale variations in the case of works based on finite size
scaling. The
implementation for the other observables is sketched briefly in the rest of the
appendix.
Our starting point is the value of the three flavor theory, that we
assume to be
\begin{equation}
\label{eq:lam3ref}
\Lambda^{(3)}_{\overline{\rm MS} } = 341\, {\rm MeV}\,,
\end{equation}
and the perturbative expansion of the Schr\"odinger Functional (SF)
coupling, that is measured on a finite volume at the scale $\mu =
1/L$. For $N_{\rm f} =3$ QCD we have
\begin{eqnarray}
\label{eq:sfptnf3}
\alpha_{\rm SF}(\mu) &=& \alpha_{\overline{\rm MS} }(s\mu) - [1.37520970 - 1.43239449 \log(s)]\alpha_{\overline{\rm MS} }^2(s\mu) \\
\nonumber
&+& [0.57120172 - 3.12911612 \log(s) + 2.05175397 \log(s)^2]\alpha_{\overline{\rm MS} }^3(s\mu)\,.
\end{eqnarray}
where the $\log$ terms can be computed from
expression~(\ref{eq:Otruncated}) and the perturbative
computation~\cite{Bode:1998hd, Bode:1999sm}. We point that the
routine \texttt{scale\_errors} from the package
\texttt{ScaleErrors.jl} (see footnote~\ref{ftn:scaleerrors}),
takes as input the perturbative coefficients
\begin{equation}
c_k (1) = \{1.0,-1.3752097,0.571202\} \,,
\end{equation}
and determines the truncation uncertainties exactly as described below.
We proceed as follows:
\begin{enumerate}
\item Chose a reference value for the scale $s_{\rm ref}$ and solve
the non linear equation
\begin{equation}
\frac{\Lambda^{(3)}_{\overline{\rm MS}} }{s_{\rm ref}\mu}=
\left[b_0\bar g^2(s_{\rm ref}\mu)\right]^{-\frac{b_1}{2b_0^2}}\,
e^{-\frac{1}{2b_0\bar g^2(s_{\rm ref}\mu)}}\,
\exp\left\{-
\int_{0}^{\bar g(s_{\rm ref}\mu)}{\rm d}x\, \left[\frac{1}{\beta_{\overline{\rm MS} }(x)} +
\frac{1}{b_0x^3} - \frac{b_1}{b_0^2x}\right]\right\}\,.
\end{equation}
in order to obtain $\alpha_{\overline{\rm MS} }(s_{\rm ref}\mu)
\equiv \bar g^2(s_{\rm ref}\mu)/(4\pi)$.
For example, using $\mu = 80$ GeV, $s_{\rm ref} = 2$ and the 5-loop
$\beta_{\overline{\rm MS} }$ function together
with our reference value of $\Lambda^{(3)}_{\overline{\rm MS} } =
341$ MeV we get
\begin{equation}
\label{eq:alms1}
\alpha_{\overline{\rm MS} }(s_{\rm ref}\mu) = 0.09703895\,.
\end{equation}
\item This value is plugged in Eq.~(\ref{eq:sfptnf3}) in order to
obtain
\begin{equation}
\label{eq:alphasfref}
\alpha_{\rm SF}(80\, {\rm GeV}) = 0.09287934\,.
\end{equation}
\item We can now solve the polynomial equation~(\ref{eq:sfptnf3})
for different values of $s$ and the l.h.s fixed to the value in
Eq.~(\ref{eq:alphasfref}). {\em i.e.}\, Using $s=4$ ($s\mu = 320$ GeV), we solve
\begin{eqnarray}
0.09287934 &=& \alpha_{\overline{\rm MS} }(s\mu) - [1.37520970 - 1.43239449 \log(s)]\alpha_{\overline{\rm MS} }^2(s\mu) \\
\nonumber
&+& [0.57120172 - 3.12911612 \log(s) + 2.05175397 \log(s)^2]\alpha_{\overline{\rm MS} }^3(s\mu)\,,
\end{eqnarray}
To obtain
\begin{equation}
\label{eq:alms2}
\alpha_{\overline{\rm MS} }(320\, {\rm GeV}) = 0.08802818\,.
\end{equation}
\item The two values of the coupling
Eqs.~(\ref{eq:alms1}),~(\ref{eq:alms2}) should be equivalent,
except for the truncation uncertainties.
In order to compare them, we run the two values to a common scale
($M_Z$), crossing the charm and bottom thresholds.
We get
\begin{subequations}
\label{eq:alvalues}
\begin{eqnarray}
\alpha_{\overline{\rm MS} }(80\, {\rm GeV}) = 0.09703895 &\Longrightarrow& \alpha^{(N_{\rm f} = 5)}_{\overline{\rm MS} }(M_Z) = 0.11851821\,,\\
\alpha_{\overline{\rm MS} }(320\, {\rm GeV}) = 0.08802818 &\Longrightarrow& \alpha^{(N_{\rm f} = 5)}_{\overline{\rm MS} }(M_Z) = 0.11843281\,.
\end{eqnarray}
\end{subequations}
Let us give a few details on how this is done.
Using the 5-loop three flavor beta function and the charm quark
mass $m_{\rm c}^\star = m_{\rm c}(m_{\rm c}) = 1275.0$ MeV, we
solve the non-linear equation
\begin{equation}
\label{eq:rung}
\log \left( \frac{80\,{\rm GeV}}{m_{\rm c}^\star} \right) =
\int_{\bar g(m_{\rm c})}^{\bar g(80\, {\rm GeV})} \frac{{\rm d} x }{\beta_{\overline{\rm MS} }^{(N_{\rm f} = 3)}(x)}\,.
\end{equation}
in order to obtain the value of the three flavor coupling at the
charm scale.
We obtain
\begin{equation}
\alpha_{\overline{\rm MS} }(80\, {\rm GeV}) = 0.09703895 \Longrightarrow
\alpha_{\overline{\rm MS} }(m_{\rm c}^\star) = 0.39566020\,.
\end{equation}
We now ``cross the charm threshold'' by using the decoupling
relations~\cite{Chetyrkin:2005ia,Weinberg:1980wa,Bernreuther:1981sg,Grozin:2011nk,Schroder:2005hy} to determine the four flavor coupling at the scale
$m_{\rm c}^\star$.
\begin{equation}
\alpha_{\overline{\rm MS} }(m_{\rm c}^\star) = 0.39566020 \Longrightarrow
\alpha_{\overline{\rm MS} }^{(N_{\rm f} = 4)}(m_{\rm c}^\star) = 0.393947409 \,.
\end{equation}
A similar procedure is used to run the coupling to the bottom
quark threshold $m_{\rm b}^\star = m_{\rm b}(m_{\rm b}) = 4198.0$
MeV, and convert to the five flavor coupling
\begin{displaymath}
\alpha_{\overline{\rm MS} }^{(N_{\rm f} = 4)}(m_{\rm c}^\star) = 0.393947409 \Longrightarrow
\alpha_{\overline{\rm MS} }^{(N_{\rm f} = 4)}(m_{\rm b}^\star) = 0.226549731 \Longrightarrow
\alpha_{\overline{\rm MS} }^{(N_{\rm f} = 5)}(m_{\rm b}^\star) = 0.226311579\,.
\end{displaymath}
Finally, one uses Eq.~(\ref{eq:rung}) using the five flavor
$\beta_{\overline{\rm MS} }$ function to run the coupling to the
scale $M_Z$.
\item Finally, the difference between values of
$\alpha_{\overline{\rm MS} }^{(N_{\rm f} = 5)}(M_Z)$ in Eqs.~(\ref{eq:alvalues})
\begin{equation}
\delta \alpha_{\overline{\rm MS} }^{(N_{\rm f} = 5)}(M_Z) = 8.5\times 10^{-5}
\qquad [0.07\%]\,.
\end{equation}
can be used as an estimate of the truncation uncertainties.
\end{enumerate}
\subsection{The static potential}
\label{sec:static-potential-app}
We start from the perturbative expression of the potential $V(r)$
\begin{equation}
V(r) = -\frac{4}{3r} \sum_{n=0} P_n \left( \frac{\alpha_s}{4\pi} \right)^n\,,
\end{equation}
where $\alpha_s = \alpha_{\overline{\rm MS} }(1/r)$ is the strong coupling.
Note that we ignore the logarithmic corrections $\propto \log\alpha$
due to the IR divergent nature of $V(r)$.
See section~\ref{sec:static-potential} for more details.
Using the RG equations, we have
\begin{equation}
\frac{{\rm d} \alpha_s}{{\rm d} r} = \frac{2\alpha_s^2}{r}
\sum_{i=0}(4\pi)^i b_i\alpha_s^i\, ,
\end{equation}
which we can use to evaluate the derivative of the static potential with respect
to the scale $r$:
\begin{equation}
F(r) = \frac{{\rm d} V(r)}{{\rm d} r} =
\frac{4}{3r^2} \sum_{n=0}\frac{P_n}{(4\pi)^n} \alpha_s^{n+1} -
\frac{4}{3r^2} \sum_{n=0} \frac{2(n+1)P_n}{(4\pi)^n} \alpha_s^{n+2}
\left[ \sum_{j=0}(4\pi)^{j+1}b_j\alpha_s^j \right]\,.
\end{equation}
Collecting the coefficients in the equation above, we get an expression for the
perturbative expansion of the force, as a function of the number of flavors
$N_f$. We use the force as the observable that determines the strong coupling
constant, with the typical scale associated to the observable being $\mu=1/r$.
In $N_{\rm f} =3$ QCD the known terms in the perturbative series~\cite{Fischler:1977yf, Peter:1996ig, Smirnov:2009fh, Smirnov:2008pn}
together with expression Eq.~(\ref{eq:Otruncated}) allows to write
\begin{equation}
\begin{split}
\alpha_{\rm qq}(\mu) \simas{\mu\to\infty}
&\alpha_{\overline{\rm MS} }(s\mu) + [-0.0485502 + 1.43239449 \log(s)] \alpha^{2}_{\overline{\rm MS} }(s\mu) \\
+\, & [ 0.687447 + 0.67148339 \log(s) + 2.05175397 \log(s)^2] \alpha^{3}_{\overline{\rm MS} }(s\mu)\\
+\, & [ 0.818808 + 3.52427341 \log(s) + 2.6037989 \log(s)^2 + 2.93892108 \log(s)^3 ] \alpha^{4}_{\overline{\rm MS} }(s\mu)\\
+\, &\dots\,.
\end{split}
\end{equation}
The scale of fastest apparent convergence is reached at $s^\star =
1.034475$, and the maximum scale reached in current state of the art
determinations is 8 GeV, albeit at a single value of the lattice
spacing.
IR divergences affect the term $\propto \alpha_{\overline{\rm MS} }^4(Q)$.
If we assume a fixed value for the ultra-soft scale $\nu_{\rm us} = 2$ GeV,
the last term would read instead:
\begin{equation}
\dots + \alpha^{4}_{\overline{\rm MS} }(sQ) [ 0.246933 + 3.52427341 \log(s) + 2.6037989 \log(s)^2 + 2.93892108 \log(s)^3 ]\,.
\end{equation}
This has an effect in the truncation uncertainties at the $20\%$ level.
\subsection{HQ correlators}
The perturbative expansions for the ratios of moments $\alpha_{\rm HQ,
n}$ is given by\cite{Chetyrkin:2006xg, Chetyrkin:1997mb, Broadhurst:1991fi}:
\begin{eqnarray}
\alpha_{\rm HQ, 4}(\mu) &\simas{\mu \to \infty}& \alpha_{\overline{\rm MS} }(s\mu) - [0.07762325 - 1.43239449 \log(s)]\alpha_{\overline{\rm MS} }^2(s\mu) \\
\nonumber
&+& [0.07957445 + .58819524 \log(s) + 2.05175397 \log(s)^2]\alpha_{\overline{\rm MS} }^3(s\mu) + \dots\,.\\
\alpha_{\rm HQ, 6}(\mu) &\simas{\mu \to \infty}& \alpha_{\overline{\rm MS} }(s\mu) + [0.77386542 + 1.43239449 \log(s)]\alpha_{\overline{\rm MS} }^2(s\mu) \\
\nonumber
&-& [0.08560363 - 3.02753059 \log(s) - 2.05175397 \log(s)^2]\alpha_{\overline{\rm MS} }^3(s\mu) + \dots\,.\\
\alpha_{\rm HQ, 8}(\mu) &\simas{\mu \to \infty}& \alpha_{\overline{\rm MS} }(s\mu) + [1.08917060 + 1.43239449 \log(s)]\alpha_{\overline{\rm MS} }^2(s\mu) \\
\nonumber
&+& [0.20034888 + 3.93081340 \log(s) + 2.05175397\log(s)^2]\alpha_{\overline{\rm MS} }^3(s\mu) + \dots\,.\\
\alpha_{\rm HQ, 10}(\mu) &\simas{\mu \to \infty}& \alpha_{\overline{\rm MS} }(s\mu) + [1.44848150 + 1.43239449 \log(s)]\alpha_{\overline{\rm MS} }^2(s\mu) \\
\nonumber
&+& [0.66519861 + 4.96016330 \log(s) + 2.05175397 \log(s)^2]\alpha_{\overline{\rm MS} }^3(s\mu) + \dots\,.
\end{eqnarray}
\subsection{Wilson loops}
For the couplings defined in Eqs~(\ref{eq:alphaWptMSbar}), defined
from Wilson loops, the perturbative expansion reads~\cite{Davies:2008sw}:
\begin{eqnarray}
\alpha_{\rm W_{11}}(\mu) &\simas{\mu \to \infty}& \alpha_{\overline{\rm MS} }(s\mu) - [0.87811924 - 1.43239449 \log(s)]\alpha_{\overline{\rm MS} }^2(s\mu) \\
\nonumber
&+& [4.20161085 - 1.70505684 \log(s) + 2.05175397 \log(s)^2]\alpha_{\overline{\rm MS} }^3(s\mu) + \dots\,.\\
\alpha_{\rm W_{12}}(\mu) &\simas{\mu \to \infty}& \alpha_{\overline{\rm MS} }(s\mu) + [0.79128076 + 1.43239449 \log(s)]\alpha_{\overline{\rm MS} }^2(s\mu) \\
\nonumber
&+& [3.18658638 + 3.07742188 \log(s) + 2.05175397 \log(s)^2]\alpha_{\overline{\rm MS} }^3(s\mu) + \dots\,.
\end{eqnarray}
The scale of fastest apparent convergence is reached at $s^\star =
1.4252357$, and the maximum scale reached in current state of the art
determinations is $4.4$ GeV.
|
\section{Introduction}\label{sec:introduction}
We present here a new framework for group decision making in which a group of
individuals, or judges, collectively ranks all of the objects in a universal
set. This framework takes into consideration the pairwise comparisons implied
by the individuals' evaluations, and furthermore, it is first to combine
ordinal rankings with cardinal ratings so as to achieve an aggregate ranking
that represents as well as possible the individuals' assessments, as measured
by pre-set penalty functions.
Group-ranking problems are differentiated by whether the evaluations are given
in ordinal or cardinal scales. An ordinal evaluation, or \textit{ranking}, is
one where the objects are ordered from ``most preferred'' to ``least
preferred'' in the form of an ordered list (allowing ties). On the other hand,
a cardinal evaluation, or \textit{rating}, is an assignment of scalars, which
are cardinal scores/grades, to the objects evaluated. In a rating, the
difference between the scores of two objects indicates the magnitude of
separation between such objects. Depending on the type of evaluations to be
aggregated, group-ranking problems are referred to as \textit{ranking
aggregation problem} or \textit{rating aggregation problems}.
Previous work addressed either the rankings alone aggregation problem (e.g.
\citealt{KS62,Arr63,BTT89,HL06}), or the ratings alone aggregation problem
(e.g. \citealt{Kee76,Saa77}), but not both. One of the primary contributions
here is the technique that permits to jointly aggregate rankings and ratings
into a collective evaluation.
The individual evaluations input to a group-ranking problem can be
\textit{complete} or \textit{incomplete}. In cases when each individual in the
group ranks (rates) all of the objects in the universal set, then the ranking
(rating) is said to be complete, or {\em full list}; otherwise, it is said to
be incomplete, or {\em partial list}.
The framework developed here is applicable when the judges' ratings and
rankings are incomplete.
The power of the framework developed here is illustrated in ranking the
participants of the 2007 MSOM's student paper competition (\textbf{SPC}). This
SPC aggregate ranking problem poses challenges that are unique to that
scenario:
\begin{compenum}
\item The judges provided both ratings and rankings of the papers they
reviewed. This requires to reconcile the possibly conflicting two types of
evaluations.
\item The incompleteness of the evaluations was extreme: Each judge evaluated
fewer than a tenth of the papers, and each paper was reviewed by fewer than
a tenth of the judges. This caused the aggregation to be subject to the
``incomplete evaluation'' phenomenon bias, in which the individual scales
used by the judges affect the average scores, even if the preference
ordering of all the judges agree with each other. Also, outlier scores that
are too low or too high tend to dominate the aggregate score of the papers.
\end{compenum}
The issue of subjective scales is well recognized within the aggregate ranking
literature. \citet{Fre88}\rem{Ch. 8, page 296} argues that, the \emph{value
difference functions} (the rating scales) of two individuals involve an
arbitrary choice of scale and origin and thus the same numeric score from two
different judges generally do not have the same meaning. Similarly, in the
context of international surveys, a large number of studies (see, for example,
\citealt{BS01,Smi04,Har06}) show that the responses across different countries
do not have the same meaning. In particular, these studies showed that even
when asking respondents to rate each object using a simple 5-point rating
scale, there are significant differences in the \textit{response styles}
between countries. One example of a difference in response style, that arises
even when using a simple 5-point rating scale, is that in some countries there
is a tendency to use only the extreme categories while in others there is a
tendency to use only the middle categories. Another example of a difference in
response style, is that in some countries there is a tendency to use only the
top categories while in others there is a tendency to use only the bottom
categories.
In a decision making set up when the judges provide scores, one can generate
{\em implied pairwise comparisons} that reflect the intensity of the
preference. This is done by letting this intensity be the difference in the
scores for the two respective objects (these are called {\em additive}
comparisons further discussed later). \citet{HL06} demonstrated that an
aggregate rating that minimizes the penalties for differing from the individual
judges' implied {\em pairwise comparisons} overcomes the issue of using
different parts of the scale and is less sensitive to subjective scales than
the use of cardinal scores alone. These type of penalties are called {\em
separation penalties}, and the optimization problem that seeks to assign scores
that minimize the total separation penalties is called the {\em separation
problem}. The separation-deviation (SD) model, proposed in
\citep{Hoc04,Hoc06,HL06}, considers an aggregate rating scenario where the
input to the rating process is given as \emph{separation gaps} and
\emph{point-wise scores}. A separation gap is a quantity that expresses the
intensity of the preference of one object $i$ over another $j$ by one
particular judge. A point-wise score is a cardinal score of an object. The SD
optimization problem combines the objective of minimizing the penalties of the
deviation of the assigned scores to the point-wise scores assigned by judges to
each object and the minimization of the separation penalties. For any choice of
penalty functions the aggregate rating obtained by solving the SD model is a
complete-rating that minimizes the sum of penalties on deviating from the given
point-wise scores and separation gaps. The SD model is solved in polynomial
time if the penalty functions are convex. It is NP-hard otherwise.
In our problem setting the judges provided only point-wise scores. Therefore
there are no pairwise comparisons provided directly. Instead we use here the
pairwise comparisons {\em implied} by the scores. The mechanism we propose
here uses the SD model for both the rankings and ratings provided by the
judges. For the rankings the penalty functions proposed are not convex. We
``convexify" those functions and attain an optimization model that {\em
combines} the separation and deviation penalties for deviating from the
rankings and from the ratings of all judges. This is the first aggregate
decision model that combines both ordinal and cardinal inputs.
The advantages of the mechanism proposed are obvious in comparison to standard
approaches. It is easy to recognize a discrepancy in scores given to the same
object by different judges. However, it is possible that the scores given are
very close, yet each one is assigned from a different subjective scale. For
one judge the score of 7 out of 10 can indicate the top evaluation, whereas for
another it may mean the very bottom. Such scale differences cannot be
identified by considering the variance of the scores alone. Our optimal
solution to the SD problem, with the given penalty functions, allow to identify
immediately the largest penalty pairs which, if large enough, indicate that
different judges disagreed significantly on the comparison between such pairs
of objects. This permits to identify inconsistencies and outliers that could
be judges who are too lenient or too strict, or for other reasons had intensity
of preference substantially different from the others. As such the methodology
proposed not only provides an aggregate ranking, but also clarifies the
disagreements and inconsistencies that allow to go back and possibly
investigate the reasons for those outliers.
The paper is organized as follows: Section \ref{sec:literature} provides a
literature review on some relevant aggregate group-decision making techniques
for rankings and ratings aggregation. Section \ref{sec:data} describes the
evaluation methodology used in the 2007 MSOM's SPC, and gives examples where
the differences in scale used by the judges are evident. Section
\ref{sec:method} reviews the models and distance metrics used to construct the
penalty functions and defines the notions of consensus ranking and consensus
rating used here. Section \ref{sec:jointAggregation} describes the methodology
for the combined use of the given ratings and rankings in order to obtain the
aggregate ranting-ranking pair. Section \ref{sec:results} uses the methodology
presented in Section \ref{sec:method} to rank the contestants in the 2007
MSOM's SPC and analyzes the obtained results. Finally, Section
\ref{sec:conclusion} provides comments on our group-decision making framework
and its usefulness for different applications and decision-making scenarios.
\section{Literature Review}\label{sec:literature}
The ranking aggregation problem has been studied extensively, especially in the
social choice literature. In this context, one of the most celebrated results
is Arrows's impossibility theorem \citep{Arr63}, which states that there is no
``satisfactory'' method to aggregate a set of rankings. Kenneth Arrow defined a
satisfactory method as one that satisfies the following properties: universal
domain, no imposition, monotonicity, independence of irrelevant alternatives,
and non-dictatorship.
\citet{KS62} proposed a set of axioms that a distance metric between two
complete rankings should satisfy. They proved that these axioms were jointly
satisfied by a unique metric distance. This distance between two rankings is
measured by the number of {\em rank reversals} between them. A rank reversal is
incurred whenever two objects have a different relative order in the given
rankings. Similarly, {\em half} a rank reversal is incurred whenever two
objects are tied in one ranking but not in the other. Kemeny and Snell defined
the consensus ranking as the ranking that minimizes the sum of the distances to
each of the input rankings. \citet{BTT89} showed that the optimization problem
that needs to be solved to find the Kemeny-Snell consensus ranking is NP-hard.
Following the work of Kemeny-Snell, several axiomatic approaches have been
developed to determine consensus. For instance, \citet{Bog73} developed an
axiomatic distance between partial orders. One of the applications of Bogart's
distance is to determine a consensus partial order from a set of partial
orders. \citet{MC10} developed an axiomatic distance
between incomplete rankings that is used here.
The difficulties presented by Arrow's impossibility theorem and the NP-hardness
of finding the Kemeny-Snell's consensus ranking can be overcome by replacing
ordinal rankings by (cardinal) ratings. Following this direction, \citet{Kee76}
proved that the \emph{averaging method} satisfied all of Arrow's desirable
properties. In the averaging method, the consensus rating of each object is the
average of the scores it received. The most immediate drawback of this approach
is that the averaging method implicitly requires that all judges use the same
rating scale; that is, that all individuals are equally strict or equally
lenient in their score assignments. This work also ignores the aspect of
pairwise comparisons, which is essential to the Kemeny-Snell model.
Pairwise comparisons intensities are the input to Saaty's Analytic Hierarchy
Process technique \citep{Saa77}. There, the optimal scores are found by the
principal eigenvector technique. The readers are referred to \citep{Hoc10} for
an analysis of the principal eigenvector method in the context of aggregate
decision making.
The separation-deviation model of \citep{Hoc04,Hoc06,HL06} addresses the
computational shortcomings of the Kemeny-Snell model, and the decision quality
inadequacies of the principal eigenvector method. This model takes point-wise
scores and potentially also pairwise comparison as inputs. It is the building
block of the mechanism proposed here. As pointed out above, the respective
separation-deviation optimization problem is solvable in polynomial time if all
the penalty functions are convex \citep{HL06}.
The rating aggregation problem has also been studied in the context of
multi-criteria decision making literature. \citet{HL06} showed the equivalence
between the rating aggregation problem and the multi-criteria decision making
problem. In this context, the non-axiomatic ELECTRE \citep{BRV75} and PROMETHEE
\citep{BV85} methods (and their extensions) solve the rating aggregation
problem that arises from a multi-criteria decision problem by transforming it
in some sense to a ranking aggregation problem. This transformation is claimed
to be needed because each criterion is evaluated on a different scale.
\section{The Data}\label{sec:data}
The data used here is the evaluations for the 2007 MSOM's SPC. There were 58
papers submitted to the competition and 63 judges participated in the
evaluation process. Each of the 63 judges evaluated only three to five out of
the 58 papers; and each of the 58 papers was evaluated by only three to five
out of the 63 judges. Each judge reviewed and evaluated the assigned papers on
the attributes (scale):
\renewcommand{\labelenumi}{\Alph{enumi})}
\begin{compenum}
\item Problem importance/interest (1--10),
\item Problem modeling (0--10),
\item Analytical results (0--10),
\item Computational results (0--10),
\item Paper writing (1--10), and
\item Overall contribution to the field (Field contribution, for short)
(1--10).
\end{compenum}
\renewcommand{\labelenumi}{\arabic{enumi}.}
On each attribute, the judges assigned scores according to the score guidelines
provided (see Table \ref{tab:interpretations}). In addition, each judge also
provided an ordinal ranking of the papers he/she reviewed (1 = best, 2 = second
best, etc.).
\begin{table}[H]
\centering \caption{Interpretation of each numerical score. The journals
considered are: MSOM, Operations Research (OR) and Management Science
(MS).}\label{tab:interpretations} \footnotesize\begin{tabular}{c|l} {\bf Score}
& {\bf Definition / Interpretation} \\ \hline
10 & Attribute considered is comparable to that of the best papers published in the journals. \\
8,9 & Attribute considered is comparable to that of the average papers published in the journals. \\
7 & Attribute considered is at the minimum level for publication in the journals. \\
5,6 & Attribute considered independently would require a minor revision before publication in the journals. \\
3,4 & Attribute considered independently would require a major revision before publication in the journals. \\
1,2 & Attribute considered would warrant by itself a rejection if the paper were submitted to the journals. \\
0 & Attribute considered is not relevant or applicable to the paper being
evaluated.
\end{tabular}
\end{table}
Although precise score interpretations were provided to the judges (Table
\ref{tab:interpretations}), they nevertheless appeared to have differed
significantly in their evaluation and must have interpreted the scale
differently. Examples of this phenomenon are illustrated for paper 43, in
Table \ref{tab:notcomp43}, and paper 26, in Table \ref{tab:notcomp26}. To
maintain the anonymity of judges and papers the judge and paper identification
numbers were assigned randomly.
\begin{table}[H]
\centering \caption{\small Evaluations received on paper
43.}\label{tab:notcomp43} \tiny\begin{tabular}{c|cccccccc}
{\bf Judge} & {\bf Problem} & {\bf Problem} & {\bf Analytical} & {\bf Computational} & {\bf Paper} & {\bf Field} & {\bf Paper} \\
{\bf } & {\bf Importance} & {\bf Modeling} & {\bf Results} & {\bf Results} & {\bf Writing} & {\bf Contribution} & {\bf Ranking} \\ \hline
47 & 9 & 8 & 8 & 8 & 9 & 9 & 1 \\
\nn6 & 6 & 4 & 2 & 4 & 4 & 4.5 & 1 \\
55 & 9 & 6 & 0 & 9 & 8 & 6 & 2 \\
\nn2 & 7 & 7 & 2 & 6 & 7.5 & 4 & 3
\end{tabular}
\end{table}
A detailed examination of Table \ref{tab:notcomp43} illustrates that paper 43
received in the Problem Modeling category a score of 8 by one judge (meaning
that the Problem Modeling in the paper is comparable to that in an average
paper published in MSOM, OR and MS), and a score of 4 by other judge (meaning
that the problem modeling in the paper requires a major revision before
publication in MSOM, OR and MS). These score differences are not insignificant.
Another example of the differences between the judges' evaluations is found on
the Analytical Results category. In this category, a judge gave a score of 8
(meaning that the analytical results in the paper are comparable to those in an
average paper published in MSOM, OR and MS), two judges gave a score of 2
(meaning that the analytical results in the paper are so bad that the paper
should be rejected by MSOM, OR and MS), and the remaining judge considered that
the category was not applicable to the paper (thus assigned the value of zero).
\begin{table}[H]
\centering \caption{\small Evaluations received on paper
26.}\label{tab:notcomp26} \tiny\begin{tabular}{c|cccccccc}
{\bf Judge} & {\bf Problem} & {\bf Problem} & {\bf Analytical} & {\bf Computational} & {\bf Paper} & {\bf Field} & {\bf Paper} \\
{\bf } & {\bf Importance} & {\bf Modeling} & {\bf Results} & {\bf Results} & {\bf Writing} & {\bf Contribution} & {\bf Ranking} \\ \hline
21 & 8 & 10 & 8 & 8 & 5 & 8 & 3 \\
24 & 8 & 9 & 8 & 10 & 7 & 8 & 1 \\
14 & 7 & 2 & 3 & 2 & 2 & 2 & 5 \\
26 & 8 & 8 & 7 & 8 & 8 & 7 & 3 \\
49 & 10 & 7 & 6 & 9 & 9 & 8 & 1
\end{tabular}
\end{table}
In Table \ref{tab:notcomp26} the data shows that judge 14's evaluations were
not on the same scale as the evaluations of the other judges. In particular,
in all attributes (with the exception of Problem Importance) judge 14 gave a
score indicating that the paper would be rejected by MSOM, OR and MS; on the
other hand in every attribute all of the other judges considered the paper is
worth of publishing (some of their evaluations even indicate that the paper
would be among the best papers published in MSOM, OR and MS!). Such
discrepancies in the judges' evaluations are quite common throughout the data.
Henceforth, we use as the input point-wise ratings the (cardinal) scores only
on the attribute ``Overall Contribution to the Field'' (``Field Contribution'',
for short). This is because the authors and the head judge of the 2007 MSOM's
SPC, believe that, among all the attributes that were scored according to the
cardinal scale in Table \ref{tab:interpretations} (i.e., excluding the ordinal
paper ranking), this attribute is the single most important attribute
evaluated.
\section{Preliminaries}\label{sec:method}
This section gives the notation used throughout the rest of the paper, and
reviews the concepts of: the separation-deviation (SD) model, distance between
incomplete ratings, and distance between incomplete rankings.
\subsection{Notation}
Let $V$ be the ground set of $n$ objects to be rated; without loss of
generality, we assign a unique identifier to each element so that
$V=\set{1,2,...,n}$. The judges are $K$ individuals. Each judge $k$,
$k\in\set{1,2,...,K}$, provides a set of scores, or ratings vector, $\a^{(k)}$
of the objects in a subset $\sA^{(k)}$ of $V$. Thus $a^{(k)}_j$ is the score
of object $j$ by the k\th individual, and $a^{(k)}_j$ is undefined if the k\th
individual did not rate object $j$. Without loss of generality, we assume that
the scores are integers contained in a pre-specified interval $[\ell,u]$. The
\emph{range} of the ratings is defined as $R\equiv u-\ell$.
We say that judge's $k$ {\em implied pairwise comparison}, or {\em separation
gap} of $i$ to $j$ is $p^{(k)}_{ij}$ where
\[p^{(k)}_{ij}=\begin{cases}
a^{(k)}_i-a^{(k)}_j & \text{if }i\in \sA^{(k)} \text{ and } j\in \sA^{(k)} \\
\text{undefined} & \text{otherwise.}\end{cases}\]
Analogously, in the ordinal setting of the incomplete-ranking aggregation
problem, each judge $k$ provides an incomplete ranking $\b^{(k)}$ of the
objects in $\sB^{(k)}$, a subset of $V$. Here $b^{(k)}_i$ is the rank (an
ordinal number) of object $i$ in the ranking provided by the k\th individual,
and $b^{(k)}_i$ is undefined if individual $k$ did not rank object $i$.
The implied separation gaps for ordinal rankings are
$\operatorname{sign}(b^{(k)}_i-b^{(k)}_j)$ for $i,j\in \sB^{(k)}$, where the $\operatorname{sign}$ function
is defined as:
\[\operatorname{sign} (x)=\begin{cases}
-1 & \text{if } x<0 \\
0 & \text{if } x=0 \\
1 & \text{if } x>0.
\end{cases}\]
For a vector of scores, or ratings, $\a$ of a set of objects, we denote by
$\rank{\a}$ the complete ranking of those objects obtained by sorting the
objects according to their scores in $\a$. For example, the vector of scores
$(4.5, 5, 3, 2.7, 3)$ corresponds to the ranking $(2,1,3,5,3)$.
\subsection{Review of the Separation-Deviation Model}\label{sec:sepdev}
The SD model can be applied to group-decision making problems where the input
is given as pairwise comparisons and/or point-wise scores.
In the model formulation, the variable $x_i$ is the aggregate score of the i\th
object, and the variable $z_{ij}$ is the aggregate separation gap of the i\th
over the j\th object. The separation gaps must be consistent. A set of
separation gaps, $p_{ij}$, is said to be {\em consistent} if and only if for
all triplets $i,j,k$, $p_{ij}+p_{jk}=p_{ik}$. In \citep{Hoc10,HL06} it was
proved that the consistency of a set of separation gaps is equivalent to the
existence of a set of scores $\o_i$ for $i=1,\dots,n$ so that
$p_{ij}=\o_i-\o_j$.
The mathematical programming formulation of the
SD model is:
\begin{subequations}\label{eqn:sepdev}
\begin{optprog}
(SD)$\qquad$ \optaction[\bm{x},\bm{z}]{min} & \objective{\sum_{k=1}^K \sum_{i=1}^n \sum_{j=1}^n f^{(k)}_{ij}(z_{ij}-p^{(k)}_{ij}) + \sum_{k=1}^K \sum_{i=1}^n g^{(k)}_i(x_i-a^{(k)}_i)} \\
subject to & z_{ij}& = & x_i-x_j & i=1,\dots,n; & j=1,\dots,n \label{eqn:c1} \\
& \ell & \leq & x_i \leq u & i=1,\dots,n \\
& x_i & \in & \Z & i=1,\dots,n.
\end{optprog}
\end{subequations}
The function $f^{(k)}_{ij}(\cdot)$ penalizes the difference between the
aggregate separation gap of the object pair $(i,j)$ and the k\th reviewer's
separation gap of the object pair $(i,j)$. The function $g^{(k)}_{i}(\cdot)$
penalizes the difference between the aggregate score of object $i$ and the k\th
reviewer's score of object $i$. In order to ensure polynomial-time solvability,
the functions $f^{(k)}_{ij}(\cdot)$ and $g^{(k)}_i(\cdot)$ must be
\textbf{convex}. In the context of rating aggregation, the penalty functions
assume the value $0$ for the argument $0$; meaning that if the output
separation gap for $i$ an $j$, $z_{ij}$ agree with $p^{(k)}_{ij}$ then
$f^{(k)}_{ij}(z_{ij}-p^{(k)}_{ij})=$ $f^{(k)}_{ij}(0)=$ $0$. If $i\not\in
\sB^{(k)}$, then $g^{(k)}_i(\dot)$ is set to the constant function $0$;
similarly, if at least one of $i$ or $j$ $\not\in \sB^{(k)}$, then
$f^{(k)}_{ij}(\dot)$ is set to the constant function $0$. Constraints
\eqref{eqn:c1} enforce the consistency of the aggregate separation gaps
conforming to the aggregate rating.
It was proved in \citep{Hoc04,Hoc06,HL06} that problem (SD) is a special case
of the convex dual of the minimum cost network flow (CDMCNF) problem. The most
efficient algorithm known for the CDMCNF has a running time of
$O(mn\log\frac{n^2}{m}\log(u-\ell))$ \citep{AHO03}, where $m$ is the total
number of given separation gaps, and $n$ is the number of objects.
\citet{AHO04} presented an alternative algorithm that uses a minimum-cut
algorithm as a subroutine.
\subsection{Distance between Incomplete-Ratings}\label{sec:IncompleteRatingAggregation}
Defining a penalty function on separation gaps is equivalent to quantifying the
distance between them. \citet{CK85} proposed a distance between complete
ratings. This distance function was adapted to incomplete ratings in
\citep{MC10}. It was shown that for a set of desirable properties this
adaption, called \textit{normalized projected Cook-Kress distance} (NPCK), is
the only one that satisfies all those properties.
Given two incomplete ratings $\a^{(1)}$ and $\a^{(2)}$, the NPCK distance
between the implied separation gaps is
\begin{equation}\label{eqn:npck}
d_{NPCK}(\a^{(1)},\a^{(2)})=\mathcal{C}\sum_{i\in \sA^{(1)}\bigcap \sA^{(2)}}\sum_{j\in \sA^{(1)}\bigcap \sA^{(2)}}
|p^{(1)}_{ij}-p^{(2)}_{ij}|,
\end{equation}
where
\begin{equation}\label{eqn:CteRate}
\mathcal{C}=\left(4\cdot R\cdot \ceil{\frac{\abs{\sA^{(1)}\bigcap
\sA^{(2)}}}{2}}\cdot\floor{\frac{\abs{\sA^{(1)}\bigcap \sA^{(2)}}}{2}}\right)^{-1}.
\end{equation}
$\mathcal{C}$ is a normalization constant that guarantees that $0\leq
d_{NPCK}(\cdot,\cdot)\leq 1$ and $R$ is the \emph{range} of the ratings,
$R\equiv u-\ell$. We note that $d_{NPCK}(\a^{(1)},\a^{(2)})=0$ indicates a
total agreement between the ratings $\a^{(1)}$ and $\a^{(2)}$, and
$d_{NPCK}(\a^{(1)},\a^{(2)})=1$ indicates a total disagreement between the
ratings $\a^{(1)}$ and $\a^{(2)}$. The normalization is important so that the
distances in problem \eqref{eqn:minnpck} are comparable to each other even when
the individuals rate a different number of objects. The normalization constant
$\mathcal{C}$ was chosen to address the following difficulties: (a) Each of the
distances in problem \eqref{eqn:minnpck} are between a complete rating $\xc$
and an incomplete rating. (b) The number of objects rated by each incomplete
rating are different; therefore the distances in problem \eqref{eqn:minnpck}
are over different dimensional spaces (the distance only considers the objects
rated by the incomplete rating). (c) Distances in higher dimensional spaces
tend to be bigger than distances in lower dimensional spaces; specifically,
observe that the number of summands in equation \eqref{eqn:npck} is the square
of the number of objects rated by the incomplete rating.
In \citep{MC10} the \textit{consensus rating}, $\xc$, is the optimal solution
to the following optimization problem:
\begin{equation}\label{eqn:minnpck}
\min_{\bm{x}} \sum_{k=1}^K d_{NPCK}(\a^{(k)},\bm{x}).
\end{equation}
The problem of finding the consensus rating is as a special case of the SD
model and therefore solvable in polynomial time.
\subsection{Distance between Incomplete-Rankings}\label{sec:IncompleteRankingAggregation}
Given a set of incomplete rankings, $\set{\b^{(k)}}_{k=1}^K$, the consensus
ranking is defined as the complete ranking closest to the given incomplete
rankings. \citet{KS62} proposed a distance between complete rankings. This
distance function was adapted to incomplete rankings in \citep{MC10}. It was
shown that for a set of desirable properties this adaption, called
\textit{normalized projected Kemeny-Snell distance} (NPKS), is the only one
that satisfies all those properties.
Given two incomplete rankings $\b^{(1)}$ and $\b^{(2)}$, the NPKS distance
between them is calculated as follows:
\begin{equation}\label{eqn:npks}
d_{NPKS}(\b^{(1)},\b^{(2)})=\mathcal{D}\sum_{i\in \sB^{(1)}\bigcap \sB^{(2)}}\sum_{j\in \sB^{(1)}\bigcap \sB^{(2)}}
\frac{1}{2}|\operatorname{sign}(b^{(1)}_i-b^{(1)}_j)-\operatorname{sign}(b^{(2)}_i-b^{(2)}_j)|,
\end{equation}
where $\mathcal{D}=\left(\abs{\sB^{(1)}\bigcap
\sB^{(2)}}^{2}-\abs{\sB^{(1)}\bigcap \sB^{(2)}}\right)^{-1}$. \rem{ERICK: This
is correct and not 'choose 2' because of the double sum instead of sum on
objects pairs. (The $\frac{1}{2}$ is needed to count each reversal as 1 and
each half reversal as 1/2).} $\mathcal{D}$ is a normalization constant that
guarantees that $0\leq d_{NPKS}(\cdot,\cdot)\leq 1$. When
$d_{NPKS}(\b^{(1)},\b^{(2)})=0$ there is a total agreement between $\b^{(1)}$
and $\b^{(2)}$, and when $d_{NPKS}(\b^{(1)},\b^{(2)})=1$ there is a total
disagreement between $\b^{(1)}$ and $\b^{(2)}$. The normalization is important
so that the distances in problem \eqref{eqn:minnpks} are comparable to each
other even when the individuals rank a different number of objects. The
normalization constant $\mathcal{D}$ was chosen to address the following
difficulties: (a) Each of the distances in problem \eqref{eqn:minnpks} are
between a complete ranking $\xo$ and an incomplete ranking. (b) The number of
objects ranked by each incomplete ranking are different; therefore the
distances in problem \eqref{eqn:minnpks} are over different dimensional spaces
(the distance only considers the objects ranked by the incomplete ranking). (c)
Distances in higher dimensional spaces tend to be bigger than distances in
lower dimensional spaces; specifically, observe that the number of summands in
equation \eqref{eqn:npks} is the square of the number of objects ranked by the
incomplete ranking.
The distance $d_{NPKS}(\b^{(1)},\b^{(2)})$ has the following natural
interpretation: The distance between two incomplete rankings is proportional to
the number of {\em rank reversals} between them. Where a rank reversal is
incurred whenever two objects have a different relative order in the rankings
$\b^{(1)}$ and $\b^{(2)}$. Similarly, \textit{half} a rank reversal is incurred
whenever two objects are tied in one ranking but not in the other ranking.
In \citep{MC10} the \textit{consensus ranking}, $\xo$, is the optimal solution
to
\begin{equation}\label{eqn:minnpks}
\min_{\bm{x}} \sum_{k=1}^K d_{NPKS}(\b^{(k)},\bm{x}).
\end{equation}
\subsection{Convexifying the Rankings Penalty Function}
In contrast to problem \eqref{eqn:minnpck}, problem \eqref{eqn:minnpks} is
NP-hard. We propose here to convexify the nonlinear sign functions in
$d_{NPKS}(\cdot,\cdot)$ as suggested in \citep{MC10}:
\begin{equation} \label{eqn:sepkemrelaxH}
h^{(k)}_{ij}(z_{ij}) = \begin{cases}
\max\set{0,\frac{z_{ij}+1}{2}} & \text{if sign$(b^{(k)}_i-b^{(k)}_j)=-1$} \\
\max\set{\frac{-z_{ij}}{2},\frac{z_{ij}}{2}} & \text{if sign$(b^{(k)}_i-b^{(k)}_j)=0$} \\
\max\set{\frac{1-z_{ij}}{2},0} & \text{if sign$(b^{(k)}_i-b^{(k)}_j)=1$}
\end{cases}
\end{equation}
The following formulation is then a convex version of problem
\eqref{eqn:minnpks}:
\begin{subequations}\label{eqn:sepkemrelax}
\begin{optprog}
\optaction[\bm{x},\bm{z}]{min} & \objective{\sum_{k=1}^{K}\ \mathcal{D}_k\sum_{i\in \sB^{(k)}}\ \sum_{j\in \sB^{(k)}}\ h^{(k)}_{ij}(z_{ij})} \label{eqn:sepkemrelax_obj} \\
subject to & z_{ij} & = & x_i-x_j & i=1,\dots,n; & j=1,\dots,n\text{.}
\rem{ERICK: This is correct according to derivation in green research notebook}
\end{optprog}
\end{subequations}
We conclude this section by observing that, for the rankings given by the
judges in the 2007 MSOM's SPC, the optimal solution to convexified problem
\eqref{eqn:sepkemrelax} is a good approximation to the optimal solution of
problem \eqref{eqn:minnpks}. That is, the distance $d_{NPKS}(\cdot,\cdot)$
between the optimal solution to problem \eqref{eqn:minnpks}, (obtained using
the implicit hitting set approach problem of \citealt{KM}), and the optimal
solution to the convex approximation, problem \eqref{eqn:sepkemrelax}, is only
$0.1606$. This is further discussed in Section \ref{sec:results}.
\section{Joint Aggregation of Ratings and Rankings}\label{sec:jointAggregation}
This section describes the model to jointly aggregate the ratings and the
rankings. The goal of this model is not only to fairly represent the judges'
rating and the judges' rankings, but also to balance the cardinal and ordinal
evaluations. We refer to this optimization model as the \textit{Combined
Aggregate raTing} problem, or (CAT).
The input to (CAT) is a set of ratings $\set{\a^{(k)}}_{k=1}^K$ and a set of
rankings $\set{\b^{(k)}}_{k=1}^K$. (CAT) is a combination of the rating
aggregation problem \eqref{eqn:minnpck} and the ranking aggregation problem
\eqref{eqn:minnpks}. In order to guarantee that ratings rankings weigh equally
in the optimization problem (CAT), both distance functions, $d_{NPCK}$ and
$d_{NPKS}$, are normalized. Note that one can weigh these distances differently
if justified by the circumstances of the decision context. Also, the choice of
$d_{NPCK}$ and $d_{NPKS}$ as penalty functions, or distances, can be replaced
by other distances between incomplete ratings and between incomplete rankings,
respectively.
\begin{optprog
${\mbox {\rm (CAT)}}\qquad$ \optaction[\bm{x}]{min} & \objective{\sum_{k=1}^K
d_{NPCK}\Bigl(\a^{(k)},\bm{x}\Bigr)+\sum_{k=1}^K
d_{NPKS}\Bigl(\b^{(k)},\rank{\bm{x}}\Bigr)}
\end{optprog}
We next establish that (CAT) is NP-hard by reducing problem \eqref{eqn:minnpks}
(which is NP-hard) to it.
\begin{lemma}
(CAT) is NP-hard.
\end{lemma}
\begin{proof}
Given an instance of problem \eqref{eqn:minnpks}, a set of incomplete rankings
$\set{\b^{(k)}}_{k=1}^K$), one can transform it (in polynomial time) to an
instance of (CAT) as follows. Keep unchanged $\set{\b^{(k)}}_{k=1}^K$, and
create a set of ratings $\set{\a^{(k)}}_{k=1}^K$ such that each rating
evaluates exactly one object (the choice of object is irrelevant; moreover all
of the ratings can evaluate the same object). From the definition of $d_{NPCK}$
(equation \eqref{eqn:npck}), it follows that, for every $\bm{x}$, the first summand
in (CAT) will be equal to $0$. Therefore, with this choice of ratings,
$\rank{\bm{x}^*}$, where $\bm{x}^*$ is the optimal solution to (CAT), will be the
optimal solution to problem \eqref{eqn:minnpks}.
\end{proof}
The (nonlinear, nonconvex) mathematical programming formulation of (CAT) is
\begin{subequations}\label{eqn:simultnonlinear}
\begin{optprog}
\optaction[\bm{x},\bm{z}]{min} & \objective{\sum_{k=1}^{K}\ \mathcal{C}_k\sum_{i\in \sA^{(k)}}\ \sum_{j\in \sA^{(k)}}\ \abs{z_{ij}-p^{(k)}_{ij}} + }\nonumber \\
& \objective{\sum_{k=1}^{K}\ \mathcal{D}_k\sum_{i\in \sB^{(k)}}\ \sum_{j\in \sB^{(k)}}\ \frac{1}{2}|\operatorname{sign}(z_{ij})-\operatorname{sign}(b^{(k)}_j-b^{(k)}_i)| } \label{eqn:nlobj} \\
subject to & z_{ij} & = & x_i-x_j & i=1,\dots,n; & j=1,\dots,n \label{eqn:nl1} \\
& \ell & \leq & x_i \leq u & i=1,\dots,n \label{eqn:nlbounds} \\
& x_i & \in & \Z & i=1,\dots,n. \label{eqn:nlintegraliy}
\end{optprog}
\end{subequations}
The convexification of the objective of problem \eqref{eqn:simultnonlinear}, as
described in Section \ref{sec:IncompleteRankingAggregation}, results in the
convex formulation:
\begin{subequations}\label{eqn:simultconvex}
\begin{optprog}
\optaction[\bm{x},\bm{z}]{min} & \objective{\sum_{k=1}^{K}\ \mathcal{C}_k\sum_{i\in
\sA^{(k)}}\
\sum_{j\in \sA^{(k)}}\ \abs{z_{ij}-p^{(k)}_{ij}} + \sum_{k=1}^{K}\ \mathcal{D}_k\sum_{i\in \sB^{(k)}}\ \sum_{j\in \sB^{(k)}}\ h^{(k)}_{ij}(z_{ij}) } \label{eqn:scobj} \\
subject to & z_{ij} & = & x_{i}-x_{j} & i=1,\dots,n; & j=1,\dots,n \label{eqn:sc1} \\
& \ell & \leq & x_i \leq u & i=1,\dots,n \label{eqn:scbounds} \\
& x_i & \in & \Z & i=1,\dots,n \label{eqn:scintegraliy} \\
where, & \objective{
h^{(k)}_{ij}(z_{ij}) = \begin{cases}
\max\set{0,\frac{z_{ij}+1}{2}} & \text{if sign$(b^{(k)}_j-b^{(k)}_i)=-1$} \\
\max\set{\frac{-z_{ij}}{2},\frac{z_{ij}}{2}} & \text{if sign$(b^{(k)}_j-b^{(k)}_i)=0$} \\
\max\set{\frac{1-z_{ij}}{2},0} & \text{if sign$(b^{(k)}_j-b^{(k)}_i)=1$.}
\end{cases}}\label{eqn:scH}
\rem{ERICK: This is correct according to derivation in green research notebook}
\end{optprog}
\end{subequations}
Problem \eqref{eqn:simultconvex} is a special case of the convex SD model and
thus solvable in polynomial time.
\textbf{Remark: } Note that in equations \eqref{eqn:nlobj} and \eqref{eqn:scH},
the argument of the $\operatorname{sign}$ function is $b^{(k)}_j-b^{(k)}_i$ and not
$b^{(k)}_i-b^{(k)}_j$ as in equations \eqref{eqn:npks} and
\eqref{eqn:sepkemrelaxH}. This is because of the classical convention that in
the given ratings \textbf{high cardinal numbers} are assigned to the
\textbf{most} preferred objects; while in the given rankings \textbf{high
ordinal numbers} are assigned to the \textbf{least} preferred objects.
The optimal solution to (CAT) is a combined aggregate rating-ranking pair which
is denoted by $\xdc$, and its implied ranking is denoted by $\xdo$.
Next, we propose two mechanisms to identify inconsistencies in the given
evaluations (e.g. outliers, judges that are too lenient or too strict, etc.).
This information is helpful so that (say) the lead decision maker initiates an
investigation of the nature of the discrepancies and acts appropriately (for
example, by discussing these inconsistencies with the judges and promote a
discussion with the objective of alleviating them).
The first mechanism is to use the solution $\xdc$ to identify (a) judges whose
evaluations differ the most with the rest of the evaluations and (b) objects
such that the judges evaluating them had particularly divergent evaluations.
These judges (objects) are those that assigned (received) scores that disagree
the most with $\xdc$. Specifically, we use the separation penalty to identify
the judges whose evaluations are at the farthest distance from $\xdc$ (i.e.,
have the highest separation penalty). Specifically, the contribution of judge
$k$ to the separation penalty is
\begin{equation}
\sum_{i\in \sA^{(k)}} \sum_{j\in \sA^{(k)}} \mathcal{C}_k\abs{(\xdc_i-\xdc_j)-(\a^{(k)}_i-\a^{(k)}_j)}.
\end{equation}
Similarly, we use the separation penalty to identify the objects such that the
judges evaluating them had particularly divergent evaluations. These objects
are those with the highest contribution to the separation penalty. The
contribution of object $i$ to the separation penalty is
\begin{equation}
\sum_{k|i\in \sA^{(k)}} \sum_{j\in \sA^{(k)}} \mathcal{C}_k\abs{(\xdc_i-\xdc_j)-(\a^{(k)}_i-\a^{(k)}_j)}.
\end{equation}
The second mechanism to identify inconsistencies in the given evaluations is
based on Brans and Vincke's PROMETHEE method \citep{BV85}. The mechanism is to
aggregate the consensus rating $\xc$ (solution to problem \eqref{eqn:minnpck})
and the consensus rating $\xo$ (solution to problem \eqref{eqn:minnpks}) into a
partial order $(P,T,I)$ as follows:
\begin{subequations}\label{eqn:partial}
\begin{align}
& a \text{ is preferred to } b\ (a\ P\ b) && \text{ if } && \begin{cases}
\xc(a)>\xc(b) \text{ and } \xo(a)\geq\xo(b)\\
\xc(a)\geq\xc(b) \text{ and } \xo(a)>\xo(b)\end{cases}\\
& a \text{ and } b \text{ are tied } (a\ T\ b) && \text{ if } && \xc(a)=\xc(b) \text{ and } \xo(a)=\xo(b)\\
& a \text{ and } b \text{ are incomparable } (a\ I\ b) && \text{ } && \text{otherwise.}
\end{align}
\end{subequations}
Thus, by construction, the partial order $(P,T,I)$ summarizes the agreement (or
lack thereof) between the consensus rating $\xc$ and the consensus ranking
$\xo$.
Section \ref{sec:results} illustrates these mechanisms and their usefulness for
identifying objects whose evaluations deserve special attention/further
discussion.
\section{Results}\label{sec:results}
We illustrate here how to use the proposed mechanism in the ranking of the
contestants of the 2007 MSOM's SPC. These results are compared to those
obtained by aggregating only the cardinal evaluations, and those obtained by
aggregating only the ordinal evaluations.
Table \ref{tab:results} gives the consensus rating (optimal solution to problem
\eqref{eqn:minnpck}) $\xc$; the (approximate) consensus ranking (optimal
solution to problem \eqref{eqn:sepkemrelax}) $\xo$; and, the combined aggregate
rating $\xdc$ and ranking $\xdo$ (optimal solutions to problem
\eqref{eqn:simultconvex}).
\begin{table}[H]
\centering \caption{Aggregate ratings and rankings for the 2007 MSOM's
SPC.}\label{tab:results}
\tiny\begin{tabular}{r|rrrrrr|rrrr}%
{\bf Paper} & {\bf $\xc$} & {\bf $\xo$} & {\bf $\xdc$} & {\bf $\xdo$} &
& {\bf Paper} & {\bf $\xc$} & {\bf $\xo$} & {\bf $\xdc$} & {\bf
$\xdo$}\\\cline{1-5}\cline{7-11}
1 & 3 & 41 & 4 & 41 & & 30 & 5 & 24 & 5 & 23 \\
2 & 5.5 & 24 & 5 & 23 & & 31 & 6 & 24 & 5 & 23 \\
3 & 5 & 41 & 4 & 41 & & 32 & 5.5 & 24 & 5 & 23 \\
4 & 5.5 & 24 & 5 & 23 & & 33 & 5.5 & 41 & 4 & 41 \\
5 & 4.5 & 41 & 4 & 41 & & 34 & 6.5 & 3 & 7 & 2 \\
6 & 6.5 & 9 & 6 & 8 & & 35 & 5 & 41 & 4 & 41 \\
7 & 6.5 & 24 & 5 & 23 & & 36 & 5.5 & 24 & 5 & 23 \\
8 & 6 & 9 & 6 & 8 & & 37 & 5 & 24 & 5 & 23 \\
9 & 5.5 & 9 & 6 & 8 & & 38 & 3 & 41 & 4 & 41 \\
10 & 6 & 3 & 7 & 2 & & 39 & 7 & 9 & 6 & 8 \\
11 & 6 & 3 & 7 & 2 & & 40 & 5 & 41 & 4 & 41 \\
12 & 5 & 24 & 5 & 23 & & 41 & 6.5 & 3 & 6 & 8 \\
13 & 6.5 & 9 & 6 & 8 & & 42 & 6 & 24 & 5 & 23 \\
14 & 7.5 & 9 & 6 & 8 & & 43 & 6 & 9 & 6 & 8 \\
15 & 5 & 41 & 4 & 41 & & 44 & 4.5 & 9 & 5 & 23 \\
16 & 4 & 53 & 3 & 53 & & 45 & 5.5 & 24 & 5 & 23 \\
17 & 6.5 & 9 & 6 & 8 & & 46 & 6 & 9 & 6 & 8 \\
18 & 3.5 & 53 & 3 & 53 & & 47 & 6 & 9 & 6 & 8 \\
19 & 5.5 & 24 & 5 & 23 & & 48 & 6.5 & 9 & 6 & 8 \\
20 & 2.5 & 53 & 2 & 58 & & 49 & 7.5 & 2 & 7 & 2 \\
21 & 4.5 & 41 & 4 & 41 & & 50 & 4.5 & 53 & 3 & 53 \\
22 & 4 & 41 & 4 & 41 & & 51 & 5.5 & 9 & 6 & 8 \\
23 & 4.5 & 41 & 4 & 41 & & 52 & 4.5 & 41 & 4 & 41 \\
24 & 5.5 & 24 & 5 & 23 & & 53 & 5.5 & 24 & 5 & 23 \\
25 & 5 & 24 & 5 & 23 & & 54 & 7 & 3 & 7 & 2 \\
26 & 6.5 & 24 & 5 & 23 & & 55 & 4.5 & 53 & 3 & 53 \\
27 & 7.5 & 9 & 6 & 8 & & 56 & 6.5 & 3 & 7 & 2 \\
28 & 4.5 & 53 & 3 & 53 & & 57 & 7 & 1 & 8 & 1 \\
29 & 6 & 9 & 6 & 8 & & 58 & 6 & 24 & 5 & 23 \\
\end{tabular}%
\end{table}
In Table \ref{tab:results}, the consensus rating $\xc$ is non-integral because
some of the judges assigned fractional scores (in particular they assigned
grades that are multiple of $1/2$). To appropriately handle the judges'
fractional grades, we decided to set the `grading unit' to 1/2. From an
optimization point of view, this represents no problem, since the
separation-deviation problem can be solved in any pre-specified precision.
Next we give a specific example of objects/papers whose ratings and ranking are
in conflict with several other objects/papers. In particular, paper 14 has the
highest consensus score, however this conflicts with several papers (e.g.,
paper 54) that have a lower consensus score but a higher consensus rank. The
evaluations received by papers 14 and 54 are given in Table
\ref{tab:scores14and54}. The number of papers reviewed by each judge and the
average Field Contribution (FC) they gave are given in Table
\ref{tab:judges14and54}. The adjusted FC, obtained by dividing the paper's FC
by the judge's average FC, is given in Table \ref{tab:adjusted14and54}. From
these tables we observe the following:
\begin{compenum}
\item The ordinal evaluations of paper 54 seem better than those of paper 14.
This explains in part why paper 54 has a better consensus rank than paper
14.
\item The average FC of paper 14 (5.6) is only slightly bigger than that of
paper 54 (5.5). This explains in part why paper 14 has a better consensus
score than paper 54.
\item It seems that judge 44, who evaluated paper 14, was
remarkably lenient, while judge 30, who evaluated paper 14, was
remarkably strict. This suggests that the FC of '5' given by
these two judges is not comparable. Note that the adjusted FC of
paper 14-judge 44 is of 0.71; while the adjusted FC of
paper 54-judge 30 is of 1.39. Moreover, the average adjusted FC of
paper 14 and 54 are 1.10 and 1.28, respectively.
\item All of this suggests that paper 54 deserves a collective
evaluation better than that of paper 14.
\end{compenum}
In the combined aggregate rating-ranking pair, $\xdc$, which is the optimal
solution to (CAT), and in its implied ranking $\xdo$, paper 54 is rated and
ranked higher than paper 14; this, as discussed previously, seems appropriate.
In contrast, the consensus rating $\xc$ ranks paper 14 higher than 54. This
provides some evidence that indeed the combined aggregate rating-ranking pair
better represents the judges' evaluations/opinions than the consensus rating
$\xc$, which takes into consideration only the ratings provided by the judges.
\begin{table}[H]
\begin{minipage}[t]{0.5\linewidth}
\centering
\vspace{3ex}
\caption{Evaluations of papers 14 and~54.} \label{tab:scores14and54}
\vspace{3ex}
\small\begin{tabular}{cc|cc}%
{\bf }& {\bf } & {\bf Field} & {\bf } \\
{\bf }&{\bf } & {\bf Contribution} & {\bf Paper} \\
{\bf Paper}&{\bf Judge} & {\bf Score } & {\bf Ranking} \\\hline
14 & 35 & 6 & 1 \\
14 & 23 & 6 & 1 \\
14 & 48 & 7 & 1 \\
14 & 57 & 4 & 4 \\
14 & 44 & 5 & 4 \\\hline
54 & 30 & 5 & 1 \\
54 & 32 & 4 & 4 \\
54 & 25 & 6 & 1 \\
54 & 22 & 7 & 1
\end{tabular}%
\end{minipage}\hspace{0.04\linewidth}
\begin{minipage}[t]{0.45\linewidth}
\centering
\caption{Evaluation statistics of the judges that evaluated papers 14 and 54.}\label{tab:judges14and54}
\small\begin{tabular}{c|cc}%
{\bf } & {\bf Number} & {\bf Average} \\
{\bf } & {\bf of Papers} & {\bf Field} \\
{\bf Judge} & {\bf Evaluated} & {\bf Contribution} \\ \hline
35 & 4 & 4.50 \\
23 & 4 & 4.25 \\
48 & 4 & 5.25 \\
57 & 4 & 5.75 \\
44 & 5 & 7.00 \\\hline
30 & 5 & 3.60 \\
32 & 4 & 5.25 \\
25 & 5 & 4.00 \\
22 & 4 & 4.75
\end{tabular}%
\end{minipage}
\end{table}
\begin{table}[H]
\centering
\begin{minipage}[t]{0.45\linewidth}
\centering
\caption{Adjusted Field Contribution received by papers 14 and 54.}\label{tab:adjusted14and54}
\small\begin{tabular}{cc|c}%
{\bf } & {\bf } & {\bf Adjusted} \\
{\bf Paper} & {\bf Judge} & {\bf Field Contribution} \\\hline
14 & 35 & 1.33 \\
14 & 23 & 1.41 \\
14 & 48 & 1.33 \\
14 & 57 & 0.70 \\
14 & 44 & 0.71 \\\hline
54 & 30 & 1.39 \\
54 & 32 & 0.76 \\
54 & 25 & 1.50 \\
54 & 22 & 1.47 \\
\end{tabular}%
\end{minipage}
\end{table}
Next, we use the partial order $(P,T,I)$ (created as described in Section
\ref{sec:jointAggregation}) to highlight the discrepancies between the
consensus rating $\xc$ and the consensus ranking $\xo$. Figure
\ref{fig:conflicts} gives a graphical representation of the partial order that
highlights the pairs of objects where $\xc$ and $\xo$ disagree on their
relative order (that is, those object pairs that are members of the set $I$ in
the partial order $(P,T,I)$).
From Figure \ref{fig:conflicts} we observe the following: (a) Paper 14 has the
highest consensus score, however this conflicts with several papers (e.g.,
paper 54) that have a lower consensus score but a higher consensus rank (this
agrees with the analysis given above). (b) Paper 20 (lower left corner of
Figure \ref{fig:conflicts}) should receive the lowest consensus evaluation. (c)
Although the agreement between $\xc$ and $\xo$ is not perfect, there are
subsets of papers should receive a lower (or higher) collective evaluation than
others. For example, the papers $\set{1,38,18,16,22,28,50,55}$ should receive a
collective evaluations higher than that of paper 20; lower than or equal to
that of papers $\set{5,21,23,52}$; and lower than the rest of the papers.
\begin{landscape}
\begin{figure}[H]
\centering
\caption{The papers (circled) are ordered
(top to bottom) in decreasing consensus score. There is an arc
between two papers whenever the lower rated paper has a better
ranking than a higher rated paper.}
\label{fig:conflicts}
\includegraphics[width=8.62in]{2013-07-02.SPC_conflicts.pdf}\\
\end{figure}
\end{landscape}
In the 2007 MSOM's SPC, papers 38, 14, 10, 1 and 42 had the highest
contributions to the separation penalty. As noted previously, this indicates
that these papers are those whose evaluations are not consistent/deserve
further discussion. For example, paper 38---a very low rated paper in the
consensus rating---received scores from 2 to 5 and was ranked by all but one of
the judges as their least preferred paper (see Tables \ref{tab:eval38} and
\ref{tab:judges38}). In particular, paper 38 was the second most preferred
paper of judge 9; perhaps because this judge received other papers with less
quality than paper 38? We believe this is not the case since, as shown in Table
\ref{tab:judge9}, the paper ranked last by judge 9 was paper 10. As noted
above, paper 10 is also among the highest contributors to the separation
penalty. Paper 10 received three high evaluations and 2 very low evaluations
(see Table \ref{tab:paper10}). Therefore, we believe that, in order to get a
better consensus, the scores/ranks of paper 38 and paper 10 should be discussed
by the judges assigned to these two papers.
\begin{table}[H]
\begin{minipage}[t]{0.5\linewidth}
\centering
\caption{Evaluations of paper 38.} \label{tab:eval38}
\small\begin{tabular}{c|cc}%
{\bf } & {\bf Field} & {\bf } \\
{\bf } & {\bf Contribution} & {\bf Paper} \\
{\bf Judge} & {\bf Score } & {\bf Ranking} \\\hline
30 & 3 & 5 \\
41 & 2 & 5 \\
44 & 3 & 5 \\
9 & 5 & 2 \\
20 & 5 & 4
\end{tabular}%
\end{minipage}
\hspace{0.04\linewidth}
\begin{minipage}[t]{0.45\linewidth}
\centering
\caption{Evaluation statistics of the judges that evaluated paper 38.}\label{tab:judges38}
\small\begin{tabular}{c|cc}%
{\bf } & {\bf Number} & {\bf Average} \\
{\bf } & {\bf of Papers} & {\bf Field} \\
{\bf Judge} & {\bf Evaluated} & {\bf Contribution} \\ \hline
30 & 5 & 3.60 \\
41 & 5 & 5.00 \\
44 & 5 & 7.00 \\
9 & 5 & 4.60 \\
20 & 4 & 7.25
\end{tabular}%
\end{minipage}
\end{table}
\begin{table}[H]
\begin{minipage}[t]{0.5\linewidth}
\centering
\caption{Evaluations of judge 9.}\label{tab:judge9}
\small\begin{tabular}{cc|c}%
{\bf } & {\bf Field} & {\bf } \\
{\bf } & {\bf Contribution} & {\bf Paper} \\
{\bf Paper} & {\bf Score } & {\bf Ranking} \\\hline
10 & 3 & 5 \\
19 & 4 & 3 \\
38 & 5 & 2 \\
50 & 4 & 3 \\
58 & 7 & 1
\end{tabular}%
\end{minipage}
\hspace{0.04\linewidth}
\begin{minipage}[t]{0.45\linewidth}
\centering
\caption{Evaluations of paper 10.}\label{tab:paper10}
\small\begin{tabular}{cc|c}%
{\bf } & {\bf Field} & {\bf } \\
{\bf } & {\bf Contribution} & {\bf Paper} \\
{\bf Judge} & {\bf Score } & {\bf Ranking} \\\hline
33 & 7 & 1 \\
41 & 7 & 1 \\
19 & 2 & 3 \\
15 & 6 & 1 \\
9 & 3 & 5
\end{tabular}%
\end{minipage}
\end{table}
Next we analyze the combined aggregate rating $\xdc$ and ranking $\xdo$
(solution to problem \eqref{eqn:simultconvex}). We make the following
observations:
\begin{compenum}
\item The consensus rating, $\xc$, has a total rating
distance (equation \eqref{eqn:minnpck}) of 7.3611.
\item The consensus ranking, $\xo$, has a total ranking
distance (equation \eqref{eqn:minnpks}) of 13.8500.
\item
\begin{compenum}
\item The combined aggregate rating, $\xdc$, has a total rating
distance (equation \eqref{eqn:minnpck}) of 8.16667.
\item The combined aggregate ranking, $\xdo$, has a total ranking
distance (equation \eqref{eqn:minnpks}) of 13.9333.
\end{compenum}
\end{compenum}
This shows that, in this case, the combined aggregate rating $\xdc$ and ranking
$\xdo$ achieve a very good compromise. In particular, $\xdc$ remains almost as
close as the consensus rating $\xc$ to the judges' ratings, and $\xdo$ remains
almost as close as the consensus ranking $\xdo$ to the judges' rankings.
\section{Concluding Remarks}\label{sec:conclusion}
We propose here a new framework for group decision making that aggregates both
cardinal and ordinal input evaluations (referred to as ratings and rankings,
respectively). Our framework consists on finding the rating-ranking pair that
minimizes the sum of the rating-distances from the rating to the given ratings
plus the sum of the ranking-distances from the ranking to the given rankings.
The effectiveness of the new framework is illustrated by ranking the
contestants of the 2007 MSOM's student paper competition. We provide evidence
that obtaining a combined aggregate cardinal and ordinal evaluations better
represents the judges' opinions as compared to a rating that aggregates only
the judges' cardinal evaluations or only the judges' ordinal evaluations.
Aggregating incomplete evaluations is challenging because the aggregate
evaluation is prone to be biased by the judges' subjective scales; for example,
objects assigned to a particularly strict (lenient) judge have an advantage
(disadvantage) compared to those objects not assigned to this specific judge.
Our framework identifies these inconsistencies in the given evaluations. This
information is helpful so that the lead decision maker can initiate an
investigation of the nature of the conflicts and act accordingly (for example,
by having the specific judges discuss, and possibly resolve, these
inconsistencies).
The problem of aggregating complete evaluations (in which all judges evaluate
all objects) is a special case of the problem of aggregating incomplete
evaluations (in which the judges are allowed to evaluate only some of the
objects). Therefore our framework is also applicable to aggregating complete
evaluations.
\section*{Acknowledgements}
The authors gratefully acknowledge J\'er\'emie Gallien, head judge of the 2007
MSOM's student paper competition, for using our methodology to aggregate the
judges evaluations in the competition.
The research of the first author is supported in part by NSF award No. CMMI-1760102.
The research of the second author is supported in part by NSF award No. OAC-1835499.
\bibliographystyle{chicago}
|
\section{Introduction}
Procedural reasoning by following several steps to achieve a goal is an essential part of our daily tasks. However, this is challenging for machines due to the complexity of instructions and commonsense reasoning required for understanding the procedure~\cite{mishra2018tracking,yagcioglu2018recipeqa,bosselut2017simulating}.
In this paper, we tackle the task of procedural reasoning in a multimodal setting for understanding cooking recipes.
The RecipeQA dataset~\cite{yagcioglu2018recipeqa} contains recipes from internet users. Thus, understanding the text is challenging due to the different language usage and informal nature of user-generated texts. The recipes are along with images provided by users which are taken in an unconstrained environment. This exposes a level of difficulty similar to real-world problems.
The tasks proposed with the dataset include textual cloze, visual cloze, visual ordering, and visual coherence.
Here, we focus on textual cloze. An example of this task is shown in Figure \ref{fig:sample}. The input to the task is a set of multimodal instructions, three textual items from the question and a placeholder to be filled by the answer. The answer has to be chosen from four options. The three question items and the correct answer make a sequence which correctly describes the steps of the recipe.
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{sample.png}
\caption{A sample of textual cloze task}
\label{fig:sample}
\end{figure*}
To design our model, we rely on the intuition that given question items, each answer describes exactly one step of the recipe.
Hence, we design a model to make explicit alignments between the candidate answers and each step and use those alignment results, given the question information. This alignment space is latent due to not having any direct supervision based on provided annotations.
Using multimodal information and representations by making a joint space for comparison has been broadly investigated in the recent research ~\cite{hessel-lee-mimno-2019unsupervised,UnifiedVisual,li2019visualbert,su2020vl-bert,yu2019multimodal,fan2018stacked,tan-bansal-2019-lxmert,nam2017dual}. Our work differs from those as we do not have direct supervision on multimodal alignments. Moreover, the task we are solving uses the sequential nature of visual and textual modality as a weak source of supervision to build a neural model to compare the textual representation of context and the answers for a given question representation.
Procedural reasoning has been investigated on different tasks~\cite{amac-etal-2019-procedural,CRCN}. While PRN~\cite{amac-etal-2019-procedural} is proposed on RecipeQA, their model does not apply to the textual cloze task. \cite{CRCN} is using procedural reasoning on multimodal information to generate a story from a sequence of images. However, the textual cloze task is about filling a blank in a sequence given a set of textual options.
Our model exploits the latent alignment space and the positional encoding of questions and answers while applying a novel approach for constraining the output space of the latent alignment. Moreover, we exploit cross-modality representations based on cross attention to investigate the benefits from information flow between images and instructions. We compare our results to the provided baselines in~\cite{yagcioglu2018recipeqa} and achieve the state-of-the-art by improving over 19\%.
\section{Proposed Model}
We design a model to solve a structured output prediction on the textual cloze task. The intuition of our model is that the correct answer option should describe precisely one instruction, and this instruction should not be already described with other items in the question. Hence, our model assumes the instruction and question as the context and candidate answers as an additional input to the alignment process.
Moreover, to incorporate the order of the sequence in question items and the placeholder, we utilize a one-hot encoding vector of positions to be concatenated with the candidate answers and question items' representations.
We give the instructions to a sentence splitter using Stanford Core NLP library~\cite{manning2014stanford}. The output is then tokenized by Flair data structure~\cite{akbik2018coling} and embedded with BERT~\cite{devlin-etal-2019-bert}. The words' embeddings are passed to an LSTM layer and the last layer is used as the instruction representation. We propose two different approaches to include images representations. These proposals are described in Section \ref{sec:ablation}. An overview of our approach is shown in Figure \ref{fig:model}.
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{model.png}
\caption{An overview of proposed model}
\label{fig:model}
\end{figure*}
Question representation is the last layer of an LSTM on question items. The representation of each question item is the concatenated vector of a one-hot position encoding and word embedding obtained from BERT. The candidate answers' representations are computed using the same approach. We concatenate the question representation to each instruction. Then, the similarity of each candidate answer and instruction is computed using the cosine similarity and form a similarity matrix. We use $S$ to denote the similarity matrix. The rows of this matrix are candidate answers and the columns represent the recipe steps. The value of $S_{ij}$ indicates the similarity score of candidate $i$ and step $j$.
For training the model, we define two different objectives directly applied to the similarity matrix. The textual cloze task does not have the direct supervision required for the alignment between candidates and steps, and our objective is designed to use the answer of the question to train this latent space of alignments. For imposing the constraint of the alignment to be disjoint between steps and candidates, one way is to simply compute the maximum of each row in the similarity matrix and use that as the aligned step for each candidate answer; However, we introduce constrained max-pooling which is a more sophisticated approach as shown in Figure~\ref{fig:approach}. We compare these two alternatives in the experimental results. We apply an iterative process to select the most related pair of instruction (a column) and answer candidate (a row) while removing the related column and row each time until all candidate answers find their aligned instruction. We denote the final selected maximum scores by $m = (S_{1i_{1}}, S_{2i_{2}}, S_{3i_{3}}, S_{4i_{4}})$, where $i_{c} \in [1, number\_of\_steps]$ is the index of the step with maximum alignment score with candidate $c$and for all pairs of candidates $c$ and $d$, $c \neq d \implies i_{c} \neq i_{d}$.
Respectively, we define two following objectives. The first objective maximizes the distance between the maximum score of the correct answer and the maximum score of another random wrong answer candidate. Furthermore, by fixing the instruction with the maximum alignment with the correct answer, it decreases the score of the other candidates alignments with that instruction. The second objective, increases the maximum similarity score of the answer to approach to 1 while decreasing the other maximum scores to be lower than $0.1$.
\begin{figure}[!h]
\centering
\includegraphics[width=\linewidth]{matrix.png}
\caption{The matrix operation for constrained max-pooling}
\label{fig:approach}
\end{figure}
\begin{align}
\begin{split}
\label{formula:one}
Loss = \max(0, S_{ri_{r}} - S_{ai_{a}} + 0.1)+ \\
\sum^{4}_{c \neq a}\max(0, S_{ci_{a}} - S_{ai_{a}} + 0.1)
\end{split} \\
\begin{split}
\label{formula:two}
Loss = (1 - S_{ai_{a}}) + \sum^{4}_{c \neq a}\max(0, S_{ci_{c}} - 0.1)
\end{split}
\end{align}
Where $a \in \{1, 2, 3, 4\}$ is the correct answer number and $r$ is a random index from $\{1, 2,, 3, 4\} - \{a\}$. The main difference in objective~\ref{formula:one} and objective~\ref{formula:two} is the regularization term on the selected instruction column in the alignment matrix.
\section{Experiment}
\subsection{Baselines}
\textbf{Hasty Student}~\cite{tapaswi2016movieqa} is a simple approach considering only the similarity between elements in question and candidate answers. This baseline fails to get good results due to the intrinsic of the task.
\noindent \textbf{Impatient Reader}~\cite{hermann2015teaching} computes attention from answers to the recipe for each candidate and despite being a complicated approach, yet it fails to get good results on the task. Moreover, multimodal Impatient reader approach uses both instructions and corresponding images.
\subsection{Results}
The RecipeQA textual cloze task contains $7837$ training, $961$ validation, and $963$ test examples. A learning rate of $4-e1$ is used for the first half and then $8-e2$ for the second half of training iterations. We use the momentum of $0.9$ for all variations of our model. We train for $30$ iterations with a batch size of 1 and optimize the weights using an SGD optimizer.
For word embedding, the pre-trained BERT embedding in Flair framework is used. For the image representations, ResNet50~\cite{he2016deep} pre-trained on Imagenet~\cite{russakovsky2015imagenet} using PyTorch library~\cite{NEURIPS2019_9015} is applied.
Table~\ref{tab:results} presents the experimental results. We call the model variations which use the loss objective in Equation (\ref{formula:one}) as Model-obj~\ref{formula:one} and the ones that use the loss in Equation (\ref{formula:two}) as Model-obj~\ref{formula:one}. Using the objective in Formula (\ref{formula:one}) yields better results in all experiments. This indicates the benefit of using the column-wise disjoint constraint on the similarity matrix. Also, using multimodal information yields $1.12\%$ improvement. We elaborate further on the comparison between multimodal and unimodal results in Section \ref{sec:discussion}.
\begin{table*}[h]
\centering
\begin{tabular}{lcc}
\hline
Models & Accuracy & p@2\\ \hline
Human & 73.6 & - \\
Hasty Student & 26.89 & - \\
Impatient Reader & 28.03 & - \\
Impatient Reader (multimodal) & 29.07 & - \\ \hline
Model-Obj~\ref{formula:one} & 46.35 & \textbf{78.7}\\
Model-Obj~\ref{formula:two} & 43.36 & - \\
Model-Obj~\ref{formula:one} (multimodal) & 45.41 \\
Model-Obj~\ref{formula:one} (multimodal) + LXMERT& \textbf{47.5} & 77.5\\
Model-Obj~\ref{formula:one} (multimodal) + LXMERT - ConstrainedMaxPooling& 46.9 & 76.3 \\
\hline
\end{tabular}
\caption{Evaluation on the test set}
\label{tab:results}
\end{table*}
We provide our Pytorch implementation publicly available on Github~\footnote{https://github.com/HLR/LatentAlignmentProcedural}.
\subsection{Multimodal Results}
\label{sec:ablation}
In order to investigate the usefulness of the images in solving the textual cloze task, we propose two different models that incorporate the image representation in addition to the textual information of recipe steps. The first variation receives ResNet50 representations of the images and, after applying an LSTM layer, pulls the last layer as image representation. Finally, it concatenates the image representation to the question and instruction representation in the main architecture before applying the MLP and computing the cosine similarities.
The second variation as shown in figure \ref{fig:lxmert}, uses a more complex architecture introduced in LXMERT~\cite{tan-bansal-2019-lxmert}. We modify the architecture of LXMERT and apply it to the word embedding and image representations to flow the information from each to another. The updated word embedding and image representations are passed to an LSTM, and its last layer is used to represent the visual and textual information of a step. In the end, these representations are concatenated to each other and the question representation to build the instruction vector representation.
We report the results of these model variations in Table~\ref{tab:results}. Using the cross modality representations based on LXMERT provided extensive way to flow the information from text and image to each other and yields the best results.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{LXMERT.png}
\caption{Using LXMERT for integrating multimodal information on steps}
\label{fig:lxmert}
\end{figure}
\section{Discussion and Analysis}
\label{sec:discussion}
We did qualitative analysis using some examples and their results to better understand the behaviour of the proposed model.
Our model is almost able to detect all matched candidates with the instructions (in case that there exist multiple matches) but fails to choose the one that completes the sequence of the question items.
This indicates the shortage of procedural hints inside our architecture while the latent alignment is proven to be practical.
By analysing the results, we found interesting cases where either multimodal or unimodal architectures could yield more accurate predictions.
\noindent\textbf{Multimodal - , Unimodal +}:
\begin{compactitem}
\item Images contain misleading information~(see example in Figure \ref{fig:samplev-}).
\item Image quality is low.
\item Images are not showing the steps correctly.
\item Text contains direct mentions of candidate answers.
\end{compactitem}
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{sample_v-.png}
\caption{The image is misleading the multimodal setting to choose apple slices rather than cutting option}
\label{fig:samplev-}
\end{figure}
\noindent\textbf{Multimodal + , Unimodal -}:
\begin{compactitem}
\item The sequence of the images provide detailed steps and good quality.
\item The entities in candidates answers are shown in the pictures but not in the text.
\item The recipes instructions are very short and the images provide more information.
\end{compactitem}
In some cases, the multimodal information can fix the errors resulted from not considering the order of events in the proposed architecture. Our intuition is that, although, the textual model does not contain information from previous steps, the images carry useful information on what has been already done. An example of this is shown in Figure \ref{fig:samplev+}, where co-reference resolution is required to answer the question correctly.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{sample-v+.png}
\caption{The images lead the model to understand that "it" refers to bread rather than sandwich}
\label{fig:samplev+}
\end{figure}
Furthermore, we have tested our multimodal architecture with representations of ResNet101 and the results dropped. We confirmed this experiment by re-implementing Hasty Student approach on visual coherence task~(that has 68\% accuracy with ResNet50) and obtained 35\% lower than ResNet50. This can be due to the lack of quality of images resulting in extra noise when using a more complicated network. Thus, ResNet50 achieves better accuracy by producing more abstract representations of the images.
\section{Conclusion and Future Work}
We proposed a model for RecipeQA textual cloze task which exploits the latent alignment of question items with instructions. Moreover, we investigated the benefit of using multimodal information in this task by comparing three different architectures and provided qualitative analysis on some examples to justify the results. Our model exceeded the baselines and improved the SOTA by over 19\%.
As a future direction, we will investigate the usage of the latent alignment in other tasks. We will apply more complex methods on textual abstractions and attention mechanisms to link the candidate answers with the recipe instructions. Investigating how to incorporate the question order in the architecture is another direction.
\section*{Acknowledgments}
This work is (partially) supported by the Office of Naval Research grant N00014-19-1-2308.
|
\subsection{Compute array}
\label{subsec:bypass-only-array}
\textbf{Traditional MAC units.}
In \autoref{fig:traditional-mac-array}(a) we show a traditional systolic array constructed by laying down Multiply-and-Accumulate (MAC) (\autoref{fig:traditional-mac-array}(b)) units in a 2D grid.
Each MAC unit is designed to get an operand from either both (\textit{Left in, Top in}) ports or from either of the ports, and perform multiplication and addition operation.
In the next cycle the operand data received is sent to its neighbour over the peer-to-peer links.
The internal registers, and multiplexers enable the array to work in output stationary (OS), weight stationary (WS), and input stationary (IS) modes of operation \cite{eyeriss}.
This simple mechanism of data movement results in high wire reuse, but at the same time restricts the mapping of compute only to those operations which require same set of operands to be mapped along a row or a column.
\begin{comment}
\insertWideFigure{traditional-and-smart-mac-array}{
(a) A systolic array of traditional MAC units,
(b) the architecture of a traditional MAC unit,
(c) addition of bypass paths to a traditional MAC to create a SMART MAC unit to
(d) enable flexible mapping in a systolic array by creating `islands' of MAC units capable of receiving data directly from SRAM banks by bypassing its peers. \TK{remove Fig 5c and 5d --> no need for SMART MAC. Just say we want to create islands inside the array to mimic distributed configuration. }
}
\end{comment}
\insertFigure{traditional-mac-array}{
(a) A systolic array of traditional MAC units,
(b) the architecture of a traditional MAC unit
}
\begin{comment}
\textbf{SMART MAC units.}
Any mechanism to enable MAC units to accept data from sources other than the neighbouring MAC unit will ease the restrictions of mapping compute and therefore enable high utilization.
In \autoref{fig:traditional-and-smart-mac-array}(c) we depict a modified MAC unit, which is augmented with multiplexers in the input and output ports.
\textit{We call this, a SMART MAC unit}.
Note that SMART MAC denotes a MAC unit with multiplexers on one or more input/output ports to enable bypass, not necessarily on all ports.
These multiplexers enable reading and writing to wires other than the peer-to-peer links, thus enabling it to recieve and work on operands unrelated to the ones forwarded by its peers.
Akin to the multi-hop bypass path employed in the SMART-NoC, we can allocate bypass busses from memory to the array, to provide extra channels to move data.
Connecting one of the inputs of the multiplexers with these bypass busses from memory therefore enables us to create `\textit{islands}' of MAC units.
These `\textit{islands}' are a collection of neighbouring MAC units forming a rectangular group which can act as independent smaller arrays.
We can map computations on such `\textit{islands}' independent of the other neighbouring MACs thus improving mapping flexibility (see \autoref{fig:traditional-and-smart-mac-array}(d)).
However, note that this design allows for arbitrary reconfiguration which is an overkill and makes the design costlier than necessary.
\end{comment}
\insertWideFigure{high-bw-smart-systolic-array}{
(a) Construction of a $4\times4$ \textit{systolic-cell} with bypass muxes and bypass links.
(b) A $8\times8$ SMART systolic array operating in scale-up configuration. Each $4\times4$ \textit{systolic-cell} is connected to its neighbor with the peer-to-peer links as the bypass muxes are turned off. The SRAM ports connected to bypass links are unused.
(c) Configuration of bypass muxes to enable the $8\times8$ SMART systolic array to work as a scaled-out distributed collection of systolic arrays. The bypass muxes are turned on to allow systolic-cells to directly connect to the SRAM ports which are all active.
(d) Possible monolithic and distributed configurations possible in the reconfigurable Smart-Systolic Array(SSA) using $2\times2$ \textit{systolic-cell} s
}
\textbf{Systolic Cells.}
The mapping flexibility in systolic arrays can be improved by allowing adjacent MAC units to work on different operands.
To enable this, the architecture needs to provision for additional links from the SRAM to the MACs.
Providing such links to each MAC however is costly in terms of area as well as energy since the spatial reuse over wires is compromised.
To simultaneously achieve mapping flexibility and the advantages of spatial reuse in systolic arrays, we propose a design called \textit{systolic-cell}.
A \textit{systolic-cell}~ is a small grid of traditional MAC units augmented with multiplexers at the edges.
This enables them to chose the operands from the neighbouring MAC units or a separate set of operands available via bypass links.
The MACs within a \textit{systolic-cell}~ are connected using peer-to-peer links similar to that of a traditional systolic array.
\autoref{fig:high-bw-smart-systolic-array}(a) shows a $4\times4$ \textit{systolic-cell}~ example.
Please note that choice of the size of a \textit{systolic-cell}~ is implementation dependent.
In general, the smaller the cell size, higher the mapping flexibility, which comes at a cost of slightly increased area and power.
\textbf{Scale-up and Scale-out using Systolic Cells.}
Larger arrays can be created by arranging and connecting the \textit{systolic-cell} s as depicted in \autoref{fig:high-bw-smart-systolic-array}(b) using the peer-to-peer links.
At the edge of each \textit{systolic-cell} ~the muxes can be configured to connect to the bypass links.
Please note that dedicated bypass links are allocated to each \textit{systolic-cell}{} to allow concurrency.
Attaining flexible mapping in such a design is a matter of configuring the multiplexers of the \textit{systolic-cell} s.
Depending on the mapping , an user can chose not to use the bypass paths at all and use the entire array as a single monolithic unit by setting the multiplexers to accept data only to/from the peer-to-peer links, (this is the case depicted in \autoref{fig:high-bw-smart-systolic-array}(b)), which is equivalent to a \emph{scaled-up} configuration.
One the other hand, the user can set all the multiplexers to accept and deliver data solely to the bypass links, therefore operating as a cluster of arrays, each the size of a \textit{systolic-cell}.
This configuration, depicted in \autoref{fig:high-bw-smart-systolic-array}(c) is equivalent to a \emph{scaled-out} configuration.
\autoref{fig:high-bw-smart-systolic-array}(d) illustrates some of the possible configurations constructed using a 64 MAC units with $2\times2$ \textit{systolic-cell} s.
As can be observed in this figure, not only can the array be configured to work in fully monolithic or fully distributed configurations, but also in any of the configurations in between.
By setting the appropriate muxes in either pass-through or bypass modes, sub-arrays larger than \textit{systolic-cell} ~size can be constructed (eg. $4\times4$, $8\times4$ etc in this example).
Each of the sub-arrays have access to the scratchpad memory using the bypass links.
Please note that when fully utilized, a larger systolic array improves energy efficiency over a distributed configuration of same number of MAC units by exploiting wire reuse and reducing SRAM reads.
The availability of such variety of choices for reconfiguration leads to flexible and efficient mapping, hence improving the utilization and energy efficiency of the design.
\begin{comment}
\revised{
\textbf{Scalability.}
For larger arrays, the bypass links can be pipelined into multiple-cycles via latches after a set of systolic cells.
The proposed architecture \textit{does not} need a NoC with switches between \textit{systolic-cell} s even for larger arrays since the traffic flow is completely deterministic once the muxes get configured.
Also, \textit{systolic-cell} s with dimensions larger than $4\times4$ can be used to reduce the number of bypass links, but at the expense of mapping flexibility.
}
\end{comment}
\vspace{-2mm}
\subsection{Bypass links}
\label{subsec:bypasslinks}
Adding a dedicated bypass link from the SRAM bank to each \textit{systolic-cell}~ along that row/column is necessary to attain full throughput from the array.
Given the nature of the data movements in systolic arrays, we recognize that the vertical links can be used for both second input and the output operands.
In \autoref{table:wire-scaling}, we examine the bandwidth requirements from the bypass links for the three systolic dataflows in a distributed setting, by contrasting it to the requirements of the operands.
These requirements clearly dictate that high bandwidth bypass are necessary.
Another addition in our proposed architecture are the switches at the edges of the \textit{systolic-cell}{}s.
However, these switches are simple multiplexers, which are configured statically for a given workload, without the need for any additional logic.
\textbf{Scalability via Pipelining.}
On-chip wire scalability studies such as SMART\cite{smart} have shown that it is possible to traverse a few millimeters (9mm to 11mm) of wire length in 1ns before latching the signal.
The authors in SMART achieved this using conventional asynchronous repeaters (a pair of inverters) placed 1mm apart.
In \textit{RSA}{}, repeated wires offer an opportunity to not only cross a single-systolic cell in a cycle, but in fact bypass multiple systolic cells within a single-cycle.
In our reference architecture \mys{}, we perform place-and-route to determine the number of systolic cells per pipeline stage of the bypass links.
At 28nm, we find that 8 systolic cells can be bypassed at 1GHz, as we demonstrate later in \autoref{subsec:eval-impl}, \autoref{fig:rebuttal_pnr_chart}(h).
Note that pipelining the bypass links only adds a few cycles of fill time to the \textit{RSA}{}, and does not impact the internal timing of the systolic array within each systolic cell (which is itself pipelined at each MAC unit).
\input{tables/wire-scaling-table}
\subsection{Scratch pad memory}
\vspace{-1mm}
The array constructed from \textit{systolic-cells} is backed by SRAM scratchpad memories, which are constructed as two individual buffers.
Each of these buffers is dedicated to one of the operand matrices.
Such scratchpad SRAM buffers are common in accelerators, and are designed to reduce the number of off chip accesses and facilitate temporal reuse.
Each operand buffer is operated in a double buffered fashion, so that the prefetch latency can be minimized.
The system also contains a third buffer which is used to store generated outputs elements.
To support the bandwidth of bypass links, we provision for this extra bandwidth by increasing the number of memory banks in the scratchpad SRAM buffers.
\begin{comment}
In a traditional systolic array based design, each row and column of the array is connected to one dedicated SRAM port to supply one element per cycle each.
In a similar fashion we allocate one port for each row, column \textit{and the individual bypass links}.
To reduce the complexity of multiplexing data within the SRAM,
we increase the number of SRAM banks to support the increased number of ports.
For example, the compute arrays shown in \autoref{fig:high-bw-smart-systolic-array}(c) is backed by two scratchpad memories each with 16 ports and would be constructed as collection of 16 SRAM banks.
%
The SRAM banks in a traditional 8$\times$ systolic array on the other hand would have 8 port per buffer; one per each row/column.
We evaluate the overheads for such a design in \autoref{subsec:eval-impl}.
\end{comment}
Despite having the same number of SRAM ports as in a distributed configuration, this approach has a couple of advantages over the latter.
\textit{First}, there is no replication of data required, which otherwise reduces the effective capacity of the system therefore adversely affecting reuse.
In our design by eliminating replication we inherently improve the temporal reuse of operands.
\textit{Second}, each of the \textit{systolic-cells} can access data in the entire operand buffer. Due to unified memory control of each buffer, operation like multi-cast are implicit in form of read collation, which improves energy efficiency without impacting performance.
We describe the impact on reads and energy efficiency in detail in \autoref{subsec:sim-eval}.
\vspace{-1mm}
\subsection{Control}
\vspace{-1mm}
\insertFigure{pseudocode}{Psuedocode depicting the control logic}
\autoref{fig:pseudocode} shows the control logic executed for each GEMM workload or DNN layer.
\begin{comment}
The control logic of our proposed system is similar to the control of a distributed systolic array based system.
However, unlike other systems, in a \textit{systolic-cell} based design, the number of distributed units is a variable and is determined at runtime based on the data-flow and operand shapes.
\end{comment}
The following steps describe the logic. \\
\textbf{1.} \textcolor{brown}{recNetInference}(): In this work we use a recommendation system based described in \autoref{sec:airchitect}. The model takes in the layer parameters and recommends a configuration, which is the most efficient for the workload.
\textbf{2.} \textcolor{brown}{setBypassMuxes}(): Next, the bypass muxes are set in the compute hardware to realize the partitioned configuration.
This is accomplished by writing select values to a register, whose individual bits drives the select lines. These configurations stay static throughout the GEMM computation.
\textbf{3.} \textcolor{brown}{partitionWorkload}(): The control logic, then partitions the original workload by marking portions of the original operand arrays to be used by each individual partition.
\textbf{4.} \textcolor{brown}{systolicController}(): Finally, for each partition, an instance of systolic array controller is initiated to drive the GEMM operations to completion and orchestrate the required data movement. Please note that in contrast to a traditional systolic array like TPU, multiple control units are required to work in parallel.
\begin{comment}
\begin{itemize}
\item{\textbf{1.} \textcolor{brown}{recNetInference}(): In this work we use a recommendation system based described in \autoref{sec:airchitect}. The model takes in the layer parameters and recommends a configuration, which is the most efficient for the workload.}
\item{\textbf{2.} \textcolor{brown}{setBypassMuxes}(): Next, the bypass muxes are set in the compute hardware to realize the partitioned configuration.
This is accomplished by writing select values to a register, whose individual bits drives the select lines. These configurations stay static throughout the GEMM computation.}
\item{\textbf{3.} \textcolor{brown}{partitionWorkload}(): The control logic, then partitions the original workload by marking portions of the original operand arrays to be used by each individual partition.}
\item{\textbf{4.} \textcolor{brown}{systolicController}(): Finally, for each partition, an instance of systolic array controller is initiated to drive the GEMM operations to completion and orchestrate the required data movement. Please note that in contrast to a traditional systolic array like TPU, multiple control units are required to work in parallel. }
\end{itemize}
\end{comment}
\section{Architecture Recommendation Model}
\section{Recommendation Model}
\label{sec:airchitect}
\insertWideFigure{adaptnet-config-arch-wide}{
(a) Size of the configuration space wrt number of MAC units for a \textit{systolic-cell}~ based flexible array
(b) Example of configurations predicted by \recnet~ indexed by category ID
(c) Performance and energy consumption by various configurations when running layer 19 of FasterRCNN
(d) Chart showing the name and description and name of the various classifiers used in this work
(e) The accuracies obtained by classifiers on predicting the architecture parameters for out dataset of RSA configurations with $2^{14}$ MAC units
(f) Architecture of our proposed recommendation network
}
\begin{comment}
Extracting the performance gains from a reconfigurable architecture critically depends on finding the best configuration for a given workload.
The configuration space in this case encompasses both the optimal architecture configuration as well as the mapping strategy.
\TK{We've alrady mentioned these point in the intro -- i feel like a short para is sufficient. 2 paras motivating why we do this not needed. Can be shortened}
Finding optimal hardware configuration and mapping is the key to extract maximum performance from a flexible architecture.
Current reconfigurable architectures \cite{maeri, sigma} are completely reliant on compiler support to find the best architecture configuration and the optimal mapping strategy for a given workload.
This introduces a major bottleneck for deploying the accelerator, as significant effort is required to develop a new compilation framework \cite{zhao2019mrna} or to port an existing compiler to support the new architecture.
Furthermore, finding optimal configuration (both architecture configuration and mapping) in software is expensive.
Analytical model based tools, like Timeloop\cite{timeloop}, MARVEL\cite{maestro} take from seconds to minutes to search for optima for recent reconfigurable architectures.
Accelerating this search is a major area of thrust in compiler research, and techniques like using machine learning based cost predictors have been proposed in state-of-the-art compilers\cite{autotvm}.
Still the search times remain high enough to discourage real-time reconfiguration.
\end{comment}
This section describes a neural network based recommendation unit which can simultaneously predict the optimal architecture configuration and mapping strategy, when a workload arrives.
This system solves two problems.
First it minimizes the changes required in a compiler for configuration and mapping search, thus easing deployment.
Second, it enables real time reconfigurability.
Given that a large reconfigurable array is most likely be deployed in data-center like use cases, the capability to adapt in real time will help achieve improved resource allocation and consequently meeting tight service-level-agreements (SLA).
\vspace{-2mm}
\subsection{Architecture design as ML problem}
To facilitate learning the design space we have to frame the search problem into a ML task framework like classification or regression.
We found that framing this as a classification or recommendation task works the best.
This abstraction lets us leverage the existing works and models which have been invented by the ML community.
An important step in solving this problem is to define the output space of the model.
It is natural to assign bins for the each of the design parameters and independently predict the optimal values for each parameter of interest.
However, this would require a separate model to be trained and queried for each design parameter.
We show that multiple parameters can be combined into a single output class and consequently can capture the design space using a single model.
In our case, the output space comprises of (i) The number and logical layout of the partitions, (ii) The dimensions of the arrays in each partition, and (iii) the mapping/dataflow to be used eg. output stationary (OS), weight stationary (WS),and input stationary (IS).
\autoref{fig:adaptnet-config-arch-wide}(b) shows this output space captured as categories of architecture configurations, indexed by the class ID.
The learned classifier is expected to select an architecture configuration and corresponding mapping strategy which provides the optimal performance for the workload.
To better visualize the complexity of the design space, in \autoref{fig:adaptnet-config-arch-wide}(c) we depict the runtime and energy consumption for computation and SRAM reads when running layer 19 of FasterRCNN (see \autoref{subsec:sim-eval}) for the different architecture configurations and dataflows.
We observe that determining the optima is a non-trivial task; and the likelihood to chose a sub-optimal configuration is high, when naive methods are used, resulting in significant performance and energy costs.
Moreover, the best configuration for this layer is using 256 partitions laid out as a $8\times32$ grid of $16\times4$ arrays using WS dataflow, which does not conform to conventional practice of using square or near-square layouts.
\insertFigure{adaptnet-performance}{
(a) The training and validation accuracies obtained during the training step in \recnet{}
(b) Test accuracies obtained on test sets for \recnet s trained on \textit{systolic-cell}{} based flexible array with various MAC units
}
\subsection{Recommendation Neural Network}
\textbf{Dataset generation.} We generated a dataset of about 2 million workloads, by sampling M, N, and K dimensions from a uniform distribution of positive integers $<= 10^4$. For each such workload dimension we searched through the configuration space of the reconfigurable array design using $2^{12}$ MAC units to find the optimal (minimum runtime) configuration using SCALE-Sim simulator, modified for fast runtime estimation.
When using a server cluster with about 200 Xeon cores, it takes about a week to obtain all the samples.
\textbf{Choosing the classifier.} Abstracting the problem in the form of a classification naturally opens up the choice of using existing classification algorithms.
We explored a handful of pre-existing classifiers, some of which are listed in \autoref{fig:adaptnet-config-arch-wide}(d).
The Support Vector Classifiers and the XGBoost models we use are standard implementations provided in scikit-learn~\cite{sklearn} and xgboost\cite{xgboost} python packages respectively.
We implement the MLPs in keras subpackage in tensorflow and train them for 20 epochs.
In \autoref{fig:adaptnet-config-arch-wide}(e), we show the prediction accuracy of these models on a test set of 200K points, after the model has been trained on 90\% of the dataset.
It is interesting to observe that among all the models only XGBoost was able to reasonably learn the design space and achieve about 87\% prediction accuracy.
\textbf{Recommendation Model:} The performance of XGBoost model is encouraging and demonstrates that the design space can be learnt.
To further improve the prediction performance of the model, we hand designed a recommendation neural network.
We take inspiration from typical neural network based recommendation systems like DLRM\cite{dlrm}, which is constructed by augmenting embedding lookups with MLP based classification.
The presence of trainable embeddings help in mapping the input data from the raw input space to a latent space, which is observed to improve the classification performance.
Given our use-case, there are two main requirements we need to satisfy.
First, we need our network to have high accuracy in predicting the best runtime configuration which maximizes performance.
Second, given that the recommendation network needs to be queried at runtime, the network should be small keep the inference latency and implementation costs low.
In our use case, the recommendation inference for a given layer is run concurrent to the execution of a previous layer whenever possible.
Lower inference latency therefore moves the recommendation step out of the critical path.
Moreover, a smaller network has fewer computation and storage requirements and hence minimizes the overheads.
Honoring these requirements, we propose a network as depicted in \autoref{fig:adaptnet-config-arch-wide}(f).
The network, called \recnet, is simple, where we lookup the embedding entries for the input features, and then use a classifier with single hidden layer with 128 nodes and softmax activation at the output.
\textbf{Training, Performance, and Generalizability.}
To train our recommendation network we use one Titan RTX GPU with 84 SMs.
When training on the dataset for $2^{14}$ MAC based RSA, for 30 epochs with a mini-batch size of 32, it takes about an hour to converge.
\autoref{fig:adaptnet-performance}(a) shows the accuracy progression as the training proceeds.
We obtain a high accuracy of 95\% of the test dataset of 200K points, which is compared against other classifiers in \autoref{fig:adaptnet-config-arch-wide}(e).
We also test the robustness of our design by generating similar datasets of 2M points
each for RSA's with varying number of MAC units (eg $2^{12}$, $2^{13}$ etc).
The aim is to test the performance of different \recnet~ with different output configuration space.
In \autoref{fig:adaptnet-performance}(b) we plot the test accuracies obtained for each such \recnet~ trained for 30 epochs with 90:10 training-testing split. Please note that the data points in test datasets are unknown at training time.
We observe that the networks all achieve high accuracies over 90\%.
To distinguish the \recnet's among themselves we use the size of the configuration space as a suffix.
For example, the design space of $2^{14}$ MAC has 858 possible configuration, therefore we call the corresponding network \recnet-858.
\subsection{Alternatives to \recnet}
Memoization, in form of caching is one alternative to \recnet{} to attain constant time configuration lookup.
However, caching only works for a limited number of previously computed workloads.
For any workload which does not hit in the cache, search has to be performed at runtime.
The large configuration space of \textit{RSA}{} as depicted in \autoref{fig:adaptnet-config-arch-wide}(a) makes it a non scalable solution.
One the other hand, \recnet{}, owing to learned parameters, can generalize configuration recommendation to any query having workload dimensions generated from the distribution of its training dataset.
\begin{comment}
Conventional alternatives to \recnet{} would be searching for the configurations in-situ or caching previously generated optimal configuration.
Searching for optimal configurations for each layer is expensive.
\autoref{fig:adaptnet-config-arch-wide}(a) shows the configuration space of our flexible \textit{systolic-cell}~ based arrays for varying MAC units.
The configuration space for a TPU v2 ($2^{14}$ MACs) and TPU v1 ($2^{16}$ MAC) like system comprises of 858 and 1400 points respectively, searching through which in runtime is costly and is not scalable as the array sizes increase.
%
Similarly using cached configurations don't scale as well.
\recnet~ when fully trained is capable of recommending optimal configurations even for the network dimensions which has not been previously seen.
Given that a scaled accelerator would be used in a data-center or enterprise setting, which is likely to process a wide variety of layers, \recnet's generalizability provides some degree of future proofing.
The same cannot be said about caching. For caching based approach to provide similar guarantees, one needs to compute the best configuration for all possible layer dimensions, which naturally is not feasible.
However, a cache can be augmented with \recnet~ to aid faster look ups for frequently encountered workloads.
\end{comment}
\insertFigure{adaptnet-x-case}{
(a) Cycles needed to run \recnet -858 on
an array of \textit{systolic-cell} s and on the custom hardware unit (\core) as a function of number of multipliers.
(b) Architecture of the custom 1-D unit hardware for \core
(c) Relative performance of the configurations predicted by \recnet-858 for \mys~ for $2\times10^{5}$ test samples when compared to the runtime of best possible configurations
}
\begin{comment}
\textbf{Network Execution Runtime.}
%
%
As mentioned before, \recnet ~is run concurrently with the workload execution eg. while a current workload is run, the recommendation lookup for the next workload is performed.
The most intuitive option is to allocate some \textit{systolic-cell} s to run \recnet.
However, this choice will lead to either fewer MAC units left for the actual workloads, or to allocate additional \textit{systolic-cell} s for \recnet~ leading to an additional overhead.
An alternative to adding more \textit{systolic-cell} s will be to add a custom hardware dedicated for running \recnet.
We explore both the \textit{systolic-cell}~and custom hardware options below.
\autoref{fig:recomagneto-combined}(a) shows the cycles required for a single inference of the \recnet-858 as a function of multipliers used in $4\times4$ \textit{systolic-cell} ~based array and a custom architecture.
Understandably, the runtime decreases proportional to the increase in number of multipliers as we increase the number of \textit{systolic-cell} s, achieving the best runtime of 1134 cycles when using 1024 multipliers or 64 cells.
When both the workloads and the recommendation engine is run on a same array; for a TPU equivalent machine with $2^{14}$ MAC units, about 6.25\% of the array needs to be allocated for running the \recnet-858.
Another choice could be allocating more hardware resources in terms of extra 64 \textit{systolic-cell} s dedicated to run the recommender network.
However, given that \recnet~ has exclusively dense layers processing the embedding lookups, a systolic execution turns out to be sub-optimal.
A custom design tuned for \recnet~ layer parameters should be more efficient.
For efficient execution of the dense layers, we chose a 1-D multiplier unit with a binary tree based reduction as shown in \autoref{fig:recomagneto-combined}(b).
We found Input stationary (IS) dataflow to be the most performant for our use case.
In this mapping the elements of the input vector is buffered near the multipliers, while elements of the weight matrix are streamed through to generate one output element/partial sum, with a sustained throughput of 1 element per cycle.
Throughput can be further increased by adding more such 1-D units.
We name the custom core with one or more such 1-D units as \core.
In \autoref{fig:recomagneto-combined}(a) we depict the variation of runtime of \recnet-858 inference on \core~ with two 1-D units as a function of multipliers.
We find the 512 multipliers result in best runtime of 576 cycles, when running \recnet-858.
\end{comment}
\section{Reconfigurable array architecture}
\section{Reconfigurable array design}
\label{sec:arch}
\input{architecture-1-compute}
\input{architecture-2-build}
\section{Conclusions}
\label{sec:conc}
\begin{comment}
\TK{nobody reads conclusion. Cane be shotened .. this also sees seems to be re-expanding the acronyms. not needed}
We develop SARA by augmenting a reconfigurable array with a hardware running a neural network recommender system.
The recommender system can provide optimal configuration of the array at runtime when as the workload arrives.
This makes the array self sufficient to run optimally.
We design a novel, highly accurate and fast recommendation network called \recnet, and a specialized hardware accelerator \core{}.
We also present a new design approach for building scalable GEMM accelerator architectures by augmenting traditional systolic array units with bypass links called `\textit{systolic-cells}'.
This \textsc{Reconfigurable Systolic Array} can
be configured to operate in any regime between a monolithic \textit{scaled-up} and distributed \textit{scaled-out} architecture.
Using a reference design named \mys, we show that we can simultaneously achieve high mapping flexibility and exploit spatio-temporal reuse, therefore achieving both performance and energy efficiency.
\end{comment}
This work shows that the mapping and configuration space of reconfigurable accelerator can be learnt using ML.
We demonstrate this by developing a recommendation model called \recnet{} which learns and predicts the optimal configurations for \textit{RSA}{} with high accuracy.
\textit{RSA}{} is a flexible, scalable GEMM accelerator constructed using \textit{systolic-cell} s and pipelined bypass paths.
\subsection{Architectural evaluations}
\label{subsec:sim-eval}
\begin{comment}
\insertWideFigure{runtime-DNN-chart}{
Simulated runtimes for monolithic $128\times128$ baseline, distributed 1024 $4\times4$ baseline, and \mys ~for
(a) all layers in AlphaGoZero,
(b) layers in DeepSpeech2, and
(c) first 10 layers in FasterRCNN.
(d) The total and average runtime obtained for each of network
}
\end{comment}
\insertWideFigure{sim-eval-combined}{
(a) Simulated runtimes for monolithic $128\times128$ baseline, distributed 1024 $4\times4$ baseline, and \mys{} for layers in AlphaGoZero, DeepSpeech2, and first 10 layers of FasterRCNN
(b) SRAM reads for the same workloads for \mys{} and baseline configurations
(c) Speedup of \mys{} and distributed baseline as compared to the monolithic baseline
(d) Energy consumption breakdown for our workloads in \mys{} and baselines
(e) Energy delay product(EDP) of \mys{} and baselines, normalized to EDP for monolithic baseline
(f-g) Sensitivity analysis on various networks for runtime and SRAM reads
}
\input{tables/baseline-config-table}
\textbf{Methodology.}
For our architecture level studies we chose to use SCALE-Sim \cite{scalesim-arxiv}.
SCALE-Sim is a cycle accurate simulator for systolic array, which generates per cycle data accesses to and from various memories.
This enables us to estimate and compare performance, energy consumption, power etc. of systolic array based components to a certain degree of accuracy.
We created in-house scripts to generate SCALE-sim input files to perform the workload partitioning for the configurations recommended by \recnet-858.
\textbf{Workloads.}
For our evaluations we choose FasterRCNN\cite{fasterrcnn}, DeepSpeech2\cite{deepspeech2}, and AlphaGoZero\cite{alphagozero}, as our workloads as a representative of convolution neural networks, language modelling network, and DNNs for reinforcement learning respectively.
\autoref{fig:sim-eval-combined}(f-g) shows our sensitivity analysis
using a few other well known networks.
\textbf{Baselines.}
\begin{comment}
\TK{at a high level it seems like the baselines should move to a table to shorten this entire para. Since table already exists, we should be able to cut this significantly}
We have two baselines, a large monolithic systolic array and a collection of small systolic arrays to work as a single distributed system (see \autoref{table:baseline-config}).
We modelled the monolithic systolic array with the same dimensions as Google's TPU v2, with $128\times128$ MAC units.
However unlike TPU v2 which supports floating point MAC operations, our model assumes byte long words, which is the accepted size for operands for inference operations.
We allocated a total 3MB of SRAM memory to the entire array divided into 3 operand buffers, one for each input operand matrix and the output matrix.
Although we would like to point out that, SCALE-Sim assumes that there is sufficient DRAM bandwidth available for ideal prefetching of operands, and therefore the computation runs in a stall free mode.
Therefore, for runtime generated by simulation the SRAM sizes have no impact.
Nevertheless, previous work~\cite{scalesim-arxiv} has shown that buffer sizes of this scale lead to reasonable off-chips requests.
For our second baseline, we chose a collection of $4\times4$ systolic arrays.
This choice is motivated by the sizes of the Nvidia tensor cores, which are small matrix multiplication units available as special functional units Nvidia's GPUs from their Volta architecture onward.
Although the largest data-center GPUs have 84 streaming multiprocessors \cite{nvidia2017nvidia} per unit, we choose a collection of 1024 such $4\times4$ array as our system, to keep the compute capability of both our baselines the same.
We also keep the total amount of memory in the system equal to the monolithic baseline.
\end{comment}
We chose a $128\times128$ monolithic systolic array and distributed array of 1024 $4\times4$ arrays as our baselines as depicted in \autoref{table:baseline-config}.
Both the arrays have same number of MAC units as TPUv2. Each array in distributed configuration resembles the tensor cores in Nvidia GPUs.
Both that baselines have the same total SRAM memory capacity of 3MB divided into buffers for staging two operand and one output matrix.
\textbf{Performance Analysis.}
We model both of the baseline systems and \mys~in SCALE-Sim and compare the performance for our workloads.
In \autoref{fig:sim-eval-combined}(a) we depict the cycles taken to run all the layers in AlphaGoZero, DeepSpeech2, and the first 10 layers of FasterRCNN networks.
Among the baselines, the distributed configuration mostly results in faster runtime owing to higher mapping flexibility.
However \mys, owing to reconfigurability is capable of matching the better baseline configuration.
Naturally, this flexibility leads to lower aggregated runtime for \mys ~than either of the baselines.
We see this trend generalizing in \autoref{fig:sim-eval-combined}(f) as well.
\input{tables/gemm-workloads-table}
\insertFigure{best-config-distribution}{
Distribution of favorable array sizes for a 16384 MAC distributed system which attain the lowest runtime when run for each layer in
(a) synthetic GEMM workloads
(b) AlphaGoZero,
(c) DeepSpeech2, and
(d) FasterRCNN.
}
\textbf{\textit{systolic-cell}{} Design Space Exploration.}
\mys ~is also capable of realizing configurations which are out of scope of either of baselines.
This allows \mys ~to achieve higher performance than both the baselines on certain layers.
For example, consider the synthetic GEMM operands depicted in \autoref{table:gemm-workloads}.
\autoref{fig:best-config-distribution}(a) depicts the histogram of the best configuration for these layers obtained from simulation.
The layers favouring $8\times8$ or $32\times32$ configurations constitute about 40\% of the set.
Neither of these configurations can be realized a fixed array configuration like the baselines.
In \autoref{fig:best-config-distribution}(b,c,d) we show the histogram of a similar experiment conducted on our DNN workloads.
For these specific workloads, the $4\times4$ configuration works the best for majority of the layers.
This observation also explains our findings in \autoref{fig:sim-eval-combined}(a) on why \mys's performance is identical to the $4\times4$ baseline.
Nevertheless, for layers which favor configurations like $8\times8$, $32\times32$ etc. \mys ~will lead to lower runtime than both the baselines.
This is depicted by \autoref{fig:sim-eval-combined}(c), where we see that \mys{} achieves about $>10\times$ speedup over monolithic when distributed configurations are preferred. While in cases where monolithic is preferred it runs faster than both the baselines.
\begin{comment}
\insertWideFigure{sram-reads-DNN-chart}{
Number of SRAM reads reported in simulation, for monolithic $128\times128$ baseline, distributed 1024 $4\times4$ baseline, and \mys for
(a) all layers in AlphaGoZero,
(b) layers in DeepSpeech2, and
(c) first 10 layers in FasterRCNN.
(d) The total and average of SRAM reads for all layers in the representative networks for the three compute systems}
\end{comment}
\textbf{SRAM reads and Energy efficiency.}
In general, due to the loss of reuse, distributed configurations with smaller array sizes have more SRAM reads resulting in lower energy efficiency.
We observe this trend in action in \autoref{fig:sim-eval-combined}(b) where we depict the number of SRAM reads performed for layers when running our workloads on the two baselines and on \mys.
The distributed $4\times4$ system has much higher number of reads as compared to \mys ~and the monolithic baseline.
In \mys ~this efficiency loss in reuse is mitigated by using bypassing links.
As shown in \autoref{fig:sim-eval-combined}(b), across all layers in our workloads, \mys ~incurs SRAM reads close to that in the monolithic baseline.
In the case of DeepSpeech2, \textit{\mys}, owing to efficient mapping, incurs reads even fewer than that of the monolithic baseline.
Similar trends are also reflected in other networks as well (\autoref{fig:sim-eval-combined}(g)).
\begin{comment}
\insertFigure{rebuttal-compute-sram-reads-energy-DNN-chart}{
(a) Energy consumed by computations and SRAM reads normalized to energy consumption of baseline monolithic array,
(b) the breakdown of the energy consumption in terms of computation energy and energy consumed for SRAM reads, and
(c) the energy-delay product normalized to the monolithic baseline;
when running the reference workloads for our baseline configurations and \mys
\TODO{Dont use red blue green, since it means different in the prev graphs}\TK{replace Mystique with SAGAR}
}
\end{comment}
To further quantify the efficiency of \mys, we estimated the energy spent by the three configurations on the workloads by taking into account the cycle counts and the SRAM reads and scaling the counts by typical energy consumed per operation computed from RTL PnR flows.
For all the workloads, the wire energies calculated using 100 fJ/bit-mm at 14nm \cite{dally2020domain}, come to be about 0.1\% (maximum being 0.11\% or 0.8uJ in AlphaGoZero), which is negligible.
\begin{comment}
\revised{
Using the block dimension of 3.93mm2 of the compute array, the total length of wires from SRAM to systolic cells comes to be 15491.072 mm, even when naive placement is employed.
Using the wire energy of 100 fJ/bit-mm at 14nm \cite{dally2020domain}, the total energy consumed in SAGAR using 8 bit words comes out to be 0.8uJ (0.11\%) in AlphaGoZero, 3.4uJ (0.09\%) in DeepSpeech2, and 4.9uJ (0.08\%) in FasterRCNN, which is negligible.
}
\end{comment}
In \autoref{fig:sim-eval-combined}(d) we plot the energy consumed for the three workloads on the baselines and \mys.
We observe that for workloads amenable to monolithic array (ie. FasterRCNN and DeepSpeech2), \mys's energy consumption is almost identical to the monolithic baseline.
The distributed baseline on the other hand consumes an order of magnitude higher energy for all the three workloads, while supporting the same mapping configurations as \mys.
The difference in energies are a direct consequence of utilization.
Since fine grained power or clock gating is impractical, the arrays with poor utilization consume same amount of power as the arrays with better utilization. However, these arrays take longer to complete resulting in higher energy consumption.
\begin{comment}
\revised{
The arrays which have poor utilization take longer to finish a given amount of computation as compared to an otherwise equivalent array with better utilization.
Since the MAC units in the arrays are active and consume power till all the computation is done. The array with poorest utilization stays on the longest and hence ends up consuming the most energy.}
\end{comment}
For AlphaGoZero, which favours a distributed configuration, \mys ~consumes about 20\% of the energy consumed by the monolithic baseline, while almost one order of magnitude lower than that of the distributed baseline.
\autoref{fig:sim-eval-combined}(d) also shows that \mys's energy consumption for SRAM is close to that of consumed by the monolithic array for all the three workloads.
The computation energy consumption in \mys ~equivalent to the better of the two baselines.
The combined effect of improved latency and reuse is perhaps better represented by the energy-delay product (EDP) depicted by \autoref{fig:sim-eval-combined}(e).
In this figure we plot the EDP for \mys ~and the two baselines normalized to the values corresponding to the monolithic configuration.
We observe that \mys ~results in about 92\% to 80\% less EDP compared to the monolithic baseline.
This further demonstrates the efficiency of our proposed architecture, resulting from preserving reuse while simultaneously decreasing latency due to improved mapping.
\subsection{Implementation evaluations}
\label{subsec:eval-impl}
\insertWideFigure{rebuttal_pnr_chart}{
Design-space exploration and final architecture of \mys.
(a) The post PnR floor-plan diagram of \mys 's compute array,
(b) A table detailing architecture configuration of \mys, the implementation parameters, and post PnR area and power of \mys.
(c) The comparison and breakdown of post synthesis area for distributed systolic array based designs, the monolithic systolic baseline, \mys, and SIGMA
(d) The corresponding breakdown for power consumed by various components in distributed systolic array based designs, the monolithic systolic baseline, \mys, and SIGMA
(e) The variation of total area footprint of SRAM banks in various distributed systolic array and monolithic configuration juxtaposed with the variation in bank sizes and the number of banks required,
(f) A similar variation in the power consumption by the SRAM banks in distributed systolic array and monolithic configurations, and
(g) the the area and power of a $128\times128$ array when constructed using different sized of ``\textit{systolic-cells}'' normalized to the area and power of an array constructed with traditional MAC units.
(h) The maximum attainable frequency vs the number of $4\times4$ \textit{systolic-cell} s to bypass at 28nm.
}
\textbf{Methodology.}
We implemented \mys ~in RTL as a $32\times32$ array of $4\times4$ \textit{systolic-cells} and ran ASIC flow till Place-and-Route (PnR) to obtain area and power.
We used 28nm library for implementing the logic.
We also implemented the SRAM buffers as a collection of 1024 1KB cells with the SAED32 education library from Synopsis, to quantify the power and area overheads, and then scaled down to 28nm equivalent by using Dennard's scaling \cite{dennard1974design}.
\autoref{fig:rebuttal_pnr_chart}(a) depicts the post PnR floorplan of \mys's compute logic.
\autoref{fig:rebuttal_pnr_chart}(b) lists the array configuration, area, and power consumption reported after PnR by synthesizing the \textit{RSA}{} and memory at a operating frequency of 1 GHz.
At 32.768 TOPs (with 1 MAC being two operation) at 1 GHz \mys ~takes 81.90 $mm^2$ of real estate while consuming 13.01 W of power. \core ~consumes 8.65\% of area and 1.36\% of power.
\textbf{Baselines.}
We implement the baseline monolithic $128\times128$ systolic array and distributed $4\times4$ array in RTL.
The distributed array is implemented using 1024 identical $4\times4$ traditional systolic arrays connected together by a mesh interconnect.
We used the OpenSMART \cite{opensmart} tool to generate and synthesize the mesh topologies for these systems.
The total memory capacity of both the monolithic and the distributed configurations are kept the same at 3MB.
As discussed in \autoref{subsec:sim-eval} the monolithic array has two input operand buffer of 1MB each and an output buffer also with the 1MB capacity.
In our implementation, we opted for one bank per row or column of the array.
This choice ensures that each incoming link to the array will have full bandwidth from SRAM provided that bank conflicts are negligible.
Therefore each buffer in the monolithic baseline is constructed using 128, 8KB banks.
For the distributed configuration, for each $4\times4$ array we end up with 1MB for each operand buffer.
Using the same design approach as above, we end up with each buffer being constructed using 4 banks of 256 words each.
In \autoref{table:baseline-config} we extend the same design principle for designing the memory for various other cell sizes and for \mys.
In \mys, in addition to the links going directly from the SRAM to the edge MAC units of the array, we have to consider the bypass links as well.
To get full bandwidth on these links we need to consider additional buffers.
Extending the design described in \autoref{fig:high-bw-smart-systolic-array}, each row and column of \mys ~has 31 bypass links and one link to the first MAC unit, we need 32 banks per row/column.
Therefore each SRAM buffer is constructed with 1024, 1KB banks.
\textbf{Area Analysis.}
In \autoref{fig:rebuttal_pnr_chart}(c) we depict the break down of area overheads for SRAM buffers, mesh NoC, and the compute array for various distributed configurations, the monolithic array, \mys ~and SIGMA~\cite{sigma}.
We observe that the monolithic configuration is the most efficient in terms of area, where it is about 5$\times$ more compact than the distributed $4\times4$ array configuration.
The breakdown suggests that the bloating in the distributed $4\times4$ configuration is caused predominantly by the Mesh NoC (contributing to 40.5\%), followed by the SRAM buffers.
\mys ~on the other hand takes about 8\% more area than the monolithic array, while consuming about 3.2$\times$ lower area than the distributed $4\times4$ configuration.
Considering both \mys ~and the distributed configuration provides same mapping flexibility, the proposed design is strictly more efficient.
\begin{comment}
Across the various systolic-array configurations in \autoref{fig:rebuttal_pnr_chart}(c), the SRAM area appears to remain fairly constant.
This is a direct consequence of the buffer capacity and construction of the array.
In \autoref{fig:rebuttal-area_power_compute_arr-sram_breakdown}(a) we depict the total area obtained for various configurations depicted in \autoref{table:baseline-config}.
We observe that, the various configurations vary in the bank capacity and the number of banks.
Since the total capacity remains the same across the configurations, these factor counter balance each other leading to observed trends.
\end{comment}
\textbf{Power Consumption.}
In \autoref{fig:rebuttal_pnr_chart}(d) we depict the post PnR power consumption for various array configuration.
The Mesh NoC stands out as the major contributor, which naturally makes the $4\times4$ distributed configuration about 5.3$\times$ more expensive than the monolithic configuration, with the NoC contributing to about 78\% of the power.
Considering the power of the compute array alone, all the systolic-array based configurations appear to consume similar power.
We also depict the trend in power consumed by SRAM banks across various systolic-array based configurations in
\autoref{fig:rebuttal_pnr_chart}(f).
Similar to the trends observed in area breakdown, the counter balancing affects of increasing the bank sizes and lowering of number of banks lead to similar powers across various distributed and monolithic configurations.
\textit{RSA}{} however consumes about 50\% more power than the monolithic configuration, owing to the bypass links.
However this extra cost results in achieving the same mapping flexibility of the $4\times4$ distributed configuration, which is about 3.5$\times$ more expensive.
\textbf{Scalability Analysis.} (i) \autoref{fig:rebuttal_pnr_chart}(g) we show the overhead of using smaller \textit{systolic-cell}{} sizes in terms of area and power normalized to monolithic configuration.
For specific use cases with relaxed requirements for flexibility larger sized \textit{systolic-cell} s can be used to improve the implementation costs.
(ii) \autoref{fig:rebuttal_pnr_chart}(h) we depict the max frequency that can be met as a function of number of $4\times4$ \textit{systolic-cell} s that can be bypassed at 28 nm.
Since we target 1GHz, we need to pipeline the bypass paths by inserting flops after 8 \textit{systolic-cell} s as we discuss in \autoref{subsec:bypasslinks}.
\subsection{Comparison with SIGMA}
\label{subsec:sigma-comp}
\insertFigure{sigma-compare-revised}{
Runtimes obtained for
(a) running dense workloads for monolithic baseline, \mys, ~and \textit{compute normalized} configuration of Sigma (SIGMA\_C),
(b) dense workloads for monolithic baseline, \mys, ~and \textit{area normalized} configuration of Sigma (SIGMA\_A); and
(c) \mys ~and SIGMA\_C configuration by increasing levels of sparsity (decreasing density) in DeepSpeech2,
(d) \mys ~and SIGMA\_A configuration by varying levels of sparsity in AlphaGoZero
}
\textbf{Implementation Comparison.}
\begin{comment}
We compare the area of \mys~ with the published area and power numbers of a state-of-the-art flexible accelerator SIGMA \cite{sigma}.
The paper reports the breakdown of area consumed by compute and NoC components but not the SRAM.
%
Given that both SIGMA and \mys~ have same number of MAC units, we assume the same amount of SRAM is sufficient for SIGMA as well.
For our evaluation we consider both SIGMA and \mys ~get 1024B/cycle from the SRAM banks.
From \autoref{fig:rebuttal_pnr_chart}(c) we observe that in SIGMA about 80\% of area is consumed by SRAM and NoC.
This is because SIGMA uses a switch and wire-heavy benes network for streaming operands from the SRAMs to the multipliers, and a custom tree-based
topology between the adders for fast reduction.
In \mys, on the other hand,
the SMART links are used to get to a specific `\textit{systolic-cell}',
but within that
the operands are distributed and the outputs reduced via a systolic
neighbor-to-neighbor communication.
\mys~ has about 70\% less overhead as compared to SIGMA.
%
\autoref{fig:rebuttal_pnr_chart}(c) contrasts the power consumption between \mys ~and SIGMA.
We observe that NoC is SIGMA consumes about 1.8$\times$ more power than \mys, where the NoC contributes to 45\% of the total power.
\end{comment}
We compare the area of \mys~ with the published area and power numbers of a state-of-the-art flexible accelerator SIGMA \cite{sigma}.
SIGMA allocates a significant portion of area for NoC, which together with SRAM comprise about 80\% of the total area \autoref{fig:rebuttal_pnr_chart}(c).
In \mys, simple bypass links are used to achieve the flexibility, which saves about 30\% of the area in comparison.
From \autoref{fig:rebuttal_pnr_chart}(d), we observe that NoC is SIGMA consumes about 1.8$\times$ more power than \mys, with NoC consuming 45\% of total power.
\textbf{Performance Comparisons.}
We use the analytical model used in the original paper
\footnote{We thank the authors of SIGMA for their gracious support}
to estimate performance of SIGMA~\cite{sigma}, which accounts for the time taken to stream, compute, and add partial sums as per the functionality described in their paper.
%
In \autoref{fig:sigma-compare-revised}(a) we plot the simulated runtimes for \mys, monolithic baseline, and SIGMA with equal number of MAC units (denoted as SIGMA\_C) for our representative workloads
and the ten layers reported in the SIGMA paper.
%
SIGMA\_C outperforms \mys ~in all workloads.
This is due to the fact that the operands are directly streamed to the multiplier over the heavy Benes network, whereas in \mys, the store-and-forward operation takes up some cycles.
The gap in performance further widens with the increase in sparsity as shown in \autoref{fig:sigma-compare-revised}(c).
As SIGMA implementation takes more area than \mys, we also compare against the area normalized configuration \revised{(2734 MACs)} of SIGMA (denoted as SIGMA\_Ain \autoref{fig:sigma-compare-revised}) for fairness.
In this case, SIGMA\_A consumes about an order of magnitude more number of cycles for each workload as compared to compute normalized configuration, therefore rendering \mys ~as the best performer (\autoref{fig:sigma-compare-revised}(b)).
Even when considering workloads with sparse operands, SIGMA\_A is able to surpass \mys ~only at operand sparsity values above 70\% (see \autoref{fig:sigma-compare-revised}(d)).
\section{Evaluations}
\label{sec:eval}
\begin{comment}
To showcase the capabilities of our proposed design, we evaluate \mys~in two settings.
To show the benefits that arise solely from the architecture aspects, we present results obtained from analysis done with a high level simulator.
Then, to capture the implementation dependent aspects of the design, we implement \mys{} and the baselines in RTL and capture the PPA numbers by running Place-and-Route (PnR) flow.
We also compare \mys ~with a state-of-the-art flexible accelerator architecture SIGMA\cite{sigma}.
The following subsections describe our findings in details.
\end{comment}
We evaluate \mys~ in two settings.
To capture the merits of the architecture, we present results obtained from simulation. While the implementation aspects are captured by reporting PPA number obtained from Place-and-Route (PnR).
\input{evaluations-arch.tex}
\input{evaluations-impl.tex}
\input{evaluations-sigma}
\begin{comment}
\textit{\textbf{Summary.}}
\textit{Considering our findings from architectural simulations and physical implementation, we conclude that the proposed Self Adaptive Reconfigurable hardware enables achieving both high performance and energy efficiency simultaneously.
The \textsc{Reconfigurable Systolic Array} compute unit enables high mapping efficiency of a fine-grained flexible architecture, while retaining the scalability of monolithic systolic array.
The novel \recnet~ensures optimal configuration at runtime with high accuracy, while minimizing performance, power, and area overheads when run on the proposed \core.}
\end{comment}
\section{Introduction}
This document provides instructions for submitting papers to HPCA
2022. In an effort to respect the efforts of reviewers and in the
interest of fairness to all prospective authors, we request that all
submissions to HPCA 2022 follow the formatting and submission rules
detailed below. Submissions that violate these instructions may not be
reviewed, at the discretion of the Program Chair, in order to
maintain a review process that is fair to all potential authors. This
document is itself formatted using the HPCA 2022 submission format. The
content of this document mirrors that of the submission instructions
that appear on the conference website. All questions regarding paper
formatting and submission should be directed to the Program Chair.
\subsection{Format Highlights}
Here are the format highlights in a nutshell:
\begin{itemize}
\item Paper must be submitted in printable PDF format.
\item Text must be in a minimum 10pt Times font, see Table~\ref{table:formatting}.
\item Papers must be at most 11 pages (not including references) in a
two-column format.
\item No page limit for references.
\item Each reference must specify {\em all} authors, i.e., no {\em et al.}
\end{itemize}
\subsection{Paper Evaluation Objectives}
The committee will make every effort to judge each submitted paper on
its own merits. There will be no target acceptance rate. We expect to
accept a wide range of papers with appropriate expectations for
evaluation --- while papers that build on significant past work with
strong evaluations are valuable, papers that open new areas with less
rigorous evaluation are equally welcome and especially encouraged. We
also acknowledge the wide range of evaluation methodologies including
modeling, simulation, prototyping, experimental implementation, real
product evaluation, etc.
\section{Paper Preparation Instructions}
\subsection{Paper Formatting}
Papers must be submitted in printable PDF format and should contain a
{\em maximum of 11 pages} of single-spaced two-column text, {\bf not
including references}. You may include any number of pages for
references, but see below for more instructions. If you are using
\LaTeX~\cite{lamport94} to typeset your paper, then
you should use the template used to prepare this document, which you can find
on the HPCA 2022 website. If you use a different
software package to typeset your paper, then you should adhere to the
guidelines given in Table~\ref{table:formatting}.
\begin{scriptsize}
\begin{table}[h!]
\centering
\caption{Formatting guidelines for submission.}
\label{table:formatting}
\begin{tabular}{|l|l|}
\hline
\textbf{Field} & \textbf{Value}\\
\hline
\hline
File format & PDF \\
\hline
Page limit & 11 pages, {\bf not including}\\
& {\bf references}\\
\hline
Paper size & US Letter 8.5in $\times$ 11in\\
\hline
Top margin & 1in\\
\hline
Bottom margin & 1in\\
\hline
Left margin & 0.75in\\
\hline
Right margin & 0.75in\\
\hline
Body & 2-column, single-spaced\\
\hline
Space between columns & 0.25in\\
\hline
Line spacing (leading) & 12pt \\
\hline
Body font & 10pt, Times\\
\hline
Abstract font & 10pt, Times\\
\hline
Section heading font & 12pt, bold\\
\hline
Subsection heading font & 10pt, bold\\
\hline
Caption font & 9pt (minimum), bold\\
\hline
References & 8pt, no page limit, list \\
& all authors' names\\
\hline
\end{tabular}
\end{table}
\end{scriptsize}
{\em Please ensure that you include page numbers with your
submission}. This makes it easier for the reviewers to refer to
different parts of your paper when they provide comments. Please
ensure that your submission has a banner at the top of the title page,
similar to this document, which contains the submission number and the
notice of confidentiality. If using the template, just replace `NaN'
with your submission number.
\subsection{Content}
Reviewing will be {\em double blind} (no author list); therefore,
please do not include any author names on any submitted documents
except in the space provided on the submission form. You must also
ensure that the metadata included in the PDF does not give away the
authors. If you are improving upon your prior work, refer to your
prior work in the third person and include a full citation for the
work in the bibliography. For example, if you are building on {\em
your own} prior work in the
papers~\cite{nicepaper1,nicepaper2,nicepaper3}, you would say
something like: ``While the authors of~\cite{nicepaper1,nicepaper2,nicepaper3} did X,
Y, and Z, this paper additionally does W, and is therefore much
better.'' Do NOT omit or anonymize references for blind review. There
is one exception to this for your own prior work that appeared in IEEE
CAL, arXiv, workshops without archived proceedings, etc.\ as
discussed later in this document.
\noindent\textbf{Figures and Tables:} Ensure that the figures and
tables are legible. Please also ensure that you refer to your figures
in the main text. Many reviewers print the papers in
gray-scale. Therefore, if you use colors for your figures, ensure that
the different colors are highly distinguishable in gray-scale.
\noindent\textbf{References:} There is no length limit for
references. {\em Each reference must explicitly list all authors of
the paper. Papers not meeting this requirement will be rejected.}
Since there is no length limit for the number of pages
used for references, there is no need to save space here.
\section{Paper Submission Instructions}
\subsection{Guidelines for Determining Authorship}
IEEE guidelines dictate that authorship should be based on a {\em
substantial intellectual contribution}. It is assumed that all
authors have had a significant role in the creation of an article that
bears their names. In particular, the authorship credit must be
reserved only for individuals who have met each of the following
conditions:
\begin{enumerate}
\item Made a significant intellectual contribution to the theoretical
development, system or experimental design, prototype development,
and/or the analysis and interpretation of data associated with the
work contained in the article;
\item Contributed to drafting the article or reviewing and/or revising
it for intellectual content; and
\item Approved the final version of the article as accepted for publication, including references.
\end{enumerate}
A detailed description of the IEEE authorship guidelines and
responsibilities is available online.\footnote{\url{https://www.ieee.org/publications_standards/publications/rights/Section821.html}} Please keep these guidelines in
mind while determining the author list of your paper.
\subsection{Declaring Authors}
Declare all the authors of the paper upfront. Addition/removal of
authors once the paper is accepted will have to be approved by the
Program Chair, since it potentially undermines the goal of eliminating
conflicts for reviewer assignment.
\subsection{Areas and Topics}
Authors should indicate specific topics covered by the paper on the
submission page. If you are unsure whether your paper falls within the
scope of the conference, please check with the Program Chair --- HPCA is a broad,
multidisciplinary conference and encourages new topics.
\subsection{Declaring Conflicts of Interest}
Authors must register all their conflicts on the paper submission
site. Conflicts are needed to ensure appropriate assignment of
reviewers. If a paper is found to have an undeclared conflict that
causes a problem OR if a paper is found to declare false conflicts in
order to abuse or ``game'' the review system, the paper may be rejected.
Please declare a conflict of interest with the following people for any author of your paper.
A conflict occurs in the following cases:
\begin{enumerate}
\item Between advisors and advisees, forever.
\item Between family members, forever (if they might be potential reviewers).
\item Between people who have collaborated in the last 5 years. This collaboration can consist of a joint research or development project, a joint paper, or when there is direct funding from the potential reviewer (as opposed to company funding) to an author of the paper. Co-participation in professional activities, such as tutorials or studies, is not a cause for conflict. When in doubt, the author should check with the Program Chair.
\item Between people from the same institution, who were in the same institution in the last 5 years, or where one of them is being actively considered for employment (e.g. invitation for interview) by the other person's institution.
\item When there is funding involving the potential reviewer (as opposed to merely funding from the reviewer's company).
\item Between people who are members of the same research center where unpublished work is regularly discussed (e.g., at review meetings).
\item Between people whose relationship prevents the reviewer from being objective in their assessment.
\item There may be others not covered by the above with whom you believe a COI exists, for example, an ongoing collaboration which has not yet resulted in the creation of a paper or proposal. Please report such COIs; however, you may be asked to justify them. Please be reasonable. For example, you cannot declare a COI with a reviewer just because that reviewer works on topics similar to or related to those in your paper. The Program Chair may contact co-authors to explain a COI whose origin is unclear.
\end{enumerate}
The following scenarios do not constitute a conflict:
\begin{enumerate}
\item Authors of previously-published, closely related work on that basis alone.
\item ``Service'' collaborations such as co-authoring a report for a professional organization, serving on a program committee, or co-presenting tutorials.
\item Co-authoring a paper that is a compendium of various projects, community-wide tools (e.g., gem5), non-research articles, and working groups (e.g. RISC-V) with no true collaboration among the co-authors.
\item People who work on topics similar to or related to those in your papers.
\item People under the same umbrella funding award where there is no close collaboration, no discussion of unpublished work, and no joint benefit in the paper being published.
\end{enumerate}
Most reviews will be solicited among the members of the PC and the ERC, but other members from the community may also write reviews. Please declare all your conflicts (not just restricted to the PC and ERC) on the submission form. When in doubt, contact the Program Chair.
\subsection{Revision of Previously-Reviewed Manuscript}
If the manuscript has been previously reviewed and rejected and is now being re-submitted to HPCA, the authors are encouraged to provide, by the paper deadline, an optional letter explaining how the paper has been revised for this current submission. We expect this revision information to both improve submissions and the review process.
We encourage you to keep this letter concise and to the point: you are highlighting the differences to previous versions that improve the paper and address prior reviewer concerns. This letter is not required for submission. Authors who wish to submit one have control about who and when this letter will be shared with, by specifying one of the following options:
\begin{enumerate}
\item First submission (no letter to upload)
\item Shared immediately with all reviewers
\item Shared along with rebuttal
\item Shared upon explicit request naming prior reviewed venue (upload for PC chair only)
\item Shared only with paper discussion lead (upload for PC chair only)
\item Do not wish to share
\end{enumerate}
If the authors select Option 6, meaning they do not wish to share this information, a minimalistic letter can be uploaded for the PC Chair only.
The deadline for the letter is the same as the paper deadline. If you fail to upload a revision letter by the paper deadline, contact the PC Chair.
\subsection{ArXiv Submissions}
Authors may submit their work to HPCA 2022 for publication even if the paper has been submitted to or currently appears on Arxiv. However, please do recognize that Arxiv submissions cause serious issues with the double-blind review process. The PC chair has already advised the PC/ERC members not to consider Arxiv submissions in their evaluation. However,
we encourage the authors to reduce the possibility of their name being disclosed. While there is no magical solution to hide from the power of web search in locating Arxiv papers, we encourage authors to avoid substantially similar titles abstract etc., to reduce this match.
\subsection{Concurrent Submissions and Workshops}
By submitting a manuscript to HPCA 2022, the authors guarantee that
the manuscript has not been previously published or accepted for
publication in a substantially similar form in any conference,
journal, or the archived proceedings of a workshop (e.g., in the
ACM/IEEE digital libraries) --- see exceptions below. The authors also
guarantee that no paper that contains significant overlap with the
contributions of the submitted paper will be under review for any
other conference or journal or an archived proceedings of a workshop
during the HPCA 2022 review period. Violation of any of these
conditions will lead to rejection.
The only exceptions to the above rules are for the authors' own papers
in (1) workshops without archived proceedings such as in the ACM/IEEE
digital libraries (or where the authors chose not to have their paper
appear in the archived proceedings), or (2) venues such as IEEE CAL or
arXiv where there is an explicit policy that such publication does not
preclude longer conference submissions. In all such cases, the
submitted manuscript may ignore the above work to preserve author
anonymity. This information must, however, be provided on the
submission form --- the Program Chair will make this information available
to reviewers if it becomes necessary to ensure a fair review. As
always, if you are in doubt, it is best to contact Program Chair.
Finally, the ACM/IEEE Plagiarism Policies\footnote{\url{http://www.acm.org/publications/policies/plagiarism_policy}\\
\url{https://www.ieee.org/publications_standards/publications/rights/plagiarism_FAQ.html}}
cover a range of ethical issues concerning the misrepresentation of
other works or one's own work.
\section*{Acknowledgements}
This document is derived from two previous conferences, in particular HPCA 2021 and MICRO 2021, which, in turn, are derived from past MICRO, HPCA, ISCA, and ASPLOS conferences.
\bibliographystyle{IEEEtranS}
\section{Introduction}
General Matrix to Matrix Multiplication (GEMM) is at the heart of Deep Learning (DL) training and inference and thus has been the target application of many recent accelerator designs~\cite{eyeriss, tpu, brainwave, zhang-fpga-2015, tetris, tangram}.
As DL becomes pervasive and models become larger, scaling the size of accelerators is a natural next research step.
In general, there are two ways to scale GEMM accelerators. The first, often called ``scale-up", involves adding more compute and memory to a monolithic array. The second, often called ``scale-out", involves creating a distributed set of small arrays connected by a network-on-chip/network-on-package. Unfortunately, neither approach is optimal across all workloads~\cite{scalesim-ispass}.
\autoref{fig:scale-out-example-intro-new} illustrates this with two toy GEMM workloads.
Scaling-up rigid arrays (depicted as \textit{Rigid Monolithic} in \autoref{fig:scale-out-example-intro-new})
can often lead to under-utilization when the dimensions of the
GEMM do not align with the aspect ratio of the array (\textit{irregular workload} in \autoref{fig:scale-out-example-intro-new}).
Current solutions to address this issue have advocated for expensive non-blocking
networks-on-chip (e.g.,
fat-trees, butterflies and benes)
betweeen the
SRAMs and PEs~\cite{maeri, sigma} to provide flexibility (depicted as \textit{Flexible Monolithic} in \autoref{fig:scale-out-example-intro-new}).
Unfortunately, while this can help enhance mapping utilization, it comes at a significant area and power cost\footnote{non-blocking topologies are known to scale as O(NlogN) to O(N$^2$) with the number of end-points} and limits scalability.
Yet another problem with reconfigurable
flexible architectures
is the cost to find optimal architecture configurations for the given workload.
All prior works on reconfigurable accelerators rely on a compiler/mapper to determine the optimal configuration via exhaustive, heuristic or optimization algorithm-based searches~\cite{timeloop, gamma, zhao2019mrna}.
Deploying such a design therefore has a couple of challenges: (i) a cost model has be to developed and integrated into the software stack to help find optimal mappings, without which the flexible design loses utility,
(ii) an expensive configuration and mapping search has to be performed at compile-time before scheduling any workload. Usually mapping search in software takes about a few seconds to hours~\cite{timeloop, gamma, zhao2019mrna}, even with sophisticated ML assisted frameworks like autoTVM \cite{autotvm}.
As an alternative, scaling out (shown as \textit{Distributed} in \autoref{fig:scale-out-example-intro-new})
can help mitigate the under-utilization challenge, since irregular GEMMs can be tiled on to the smaller arrays as \autoref{fig:scale-out-example-intro-new} shows.
However, this leads to loss in
spatial reuse (i.e., direct data-forwarding) that scale-up designs provide,
and also requires data replication across the SRAMs of the individual arrays.
Data replication leads to a decrease in overall on-chip storage capacity, leading to a loss of temporal reuse due to smaller tiles.
Moreover, distributed arrays can exacerbate the mapping search problem~\cite{simba, timeloop}.
We propose a solution to address
all the challenges listed above. Specifically,
we make three contributions.
\insertFigure{scale-out-example-intro-new}{Comparison of scalability, utilization, and operand reuse in traditional monolithic and distributed accelerators, and the position of the proposed architecture}
\insertFigure{sara-concept}{
The constitution and interactions of the self adaptive (SA) and reconfigurable array (RA) components to make up the SARA accelerator called \mys{} in this work.
}
\textit{First contribution:} A novel accelerator architecture called \textsc{Reconfigurable Systolic Array} (\textit{RSA}) that can be configured to operate in both scale-up and scale-out settings by adding light-weight configurability within a traditional systolic array.
The \textit{RSA}~is capable of realizing all mappings as in distributed configuration, without losing spatial reuse.
Moreover, unlike a distributed configuration with fixed array dimensions and fixed parts, the \textit{RSA}~can be configured to have variable array dimensions and number of sub-arrays as depicted later in \autoref{fig:high-bw-smart-systolic-array}(d).
In practice, \textit{RSA}~closely approximates a flexible monolithic design, with a fraction of area cost.
\autoref{fig:scale-out-example-intro-new} shows our \textit{RSA} provides utilization, reuse, and scalability.
\textit{Second contribution:} We develop a novel ML model called \recnet{} which we train to learn and generalize the configuration space of \textit{RSA}{}.
In contrast to existing methods, \recnet{} enables a O(1) recommendation of the accelerator configuration given any workload, avoiding any search.
The recommendations from \recnet{} on average (GeoMean), achieve 99.93\% of the best possible runtime (oracle) in our tests with 200K samples (\autoref{fig:sara-system}(c)).
\textit{Third contribution}: We present a custom hardware engine for running \recnet, called \core. The custom design of \core~ enables it to
generate both the optimal mapping and architecture configuration for any workload being run on \textit{RSA}{}
in approximately 600 cycles, which is at least about 6 orders of magnitude faster than software.
Furthermore, \core{} consumes the same hardware real-estate and roughly the same on-chip memory capacity
\footnote{The only change in \recnet{} between various \textit{RSA}{} is the weight of the output layer, which is small in comparison to the embedding table which takes most of the on-chip space}
for different arrays and with number of workloads,
thus proving to be a scalable solution in contrast to approaches like using configuration caches.
With \core{} the configuration lookup using \recnet{} can be performed at runtime, without involving the software stack.
Together, these three components enable us to develop
a new class of accelerators that we call \textit{Self Adaptive Reconfigurable Array (SARA)} (\autoref{fig:sara-concept}).
SARA accelerators can self adapt at runtime to optimized configurations for the target workload, without
requiring compile-time analysis.
We demonstrate an instance of SARA that we name `Shape Adaptive GEMM AcceleratoR (\mys{} \footnote{Sagar is a Sanskrit word that means Ocean, reflecting the ability of our accelerator to have flexible shapes})' as shown in \autoref{fig:sara-concept}
and evaluate its performance across various
configurations.
%
We show that \mys{} has
3.2$\times$ higher compute density and
3.5$\times$ improved power efficiency, over equivalent scaled-out systolic array.
The extra flexibility costs $<$10\% in area and
50\% in power, compared to equivalent scaled-up systolic array.
Compared to an area normalized state-of-the-art flexible scalable accelerator \cite{sigma}, \mys~ incorporates
45\% more compute.
When comparing compute-equivalent configurations,
\mys~ consumes 43\% less power and 30\% less area.
To summarize, we make the following contributions.
(i) We propose \textbf{\textit{RSA}}, a reconfigurable architecture for scalable GEMM acceleration, simultaneously achieving high utilization and reuse.
(ii) We develop \textbf{\recnet}, a lean recommendation neural net which suggests optimized configuration and dataflow
with high accuracy.
(iii) We implement \textbf{\core}, a hardware capable of running \recnet{} in constant time and configuring \textit{RSA}{} at runtime.
(iv) We integrate the above components into a SARA accelerator called \textbf{\mys{}}, which achieves optimal runtime at lower power and area than SOTA.
\begin{comment}
General Matrix to Matrix Multiplication (GEMMs) is at the heart of Deep Learning (DL) training and inference and thus has been the target application of many accelerator designs~\cite{eyeriss, tpu, brainwave, zhang-fpga-2015, tetris, tangram}.
However these individual devices work on small matrix tiles, and do not have enough computation power to work on larger networks without multiple costly passes.
Recent proposals \cite{tangram}\cite{simba} have demonstrated the need for scaling the DNN computation engines to meet the computation demands of contemporary workloads.
Despite extensive research and product development on architectures for small-tile GEMMs, designing efficient architectures for performing \textit{GEMMs at scale} is still non-trivial.
The crux of the problem is that there exists a pernicious trade-off between \textit{scalability} and \textit{utilization} (or mapping efficiency).
Scalability is a direct consequence of simplicity and regularity of a particular design.
For instance, regular designs like the TPU~\cite{tpu} (systolic array) can pack large number of MAC units (e.g., $128\times128$ in TPUv3)
in a reasonable power and area budget.
Utilization, on the other hand, requires
mapping flexibility since GEMMs are often irregular in practice~\cite{sigma} and may
not match the dimensions of the physical array.
There are two ways of achieving such flexibility.
The first is to provide rich connectivity
for operand distribution and output collection within the array. This has been demonstrated in proposals like MAERI~\cite{maeri} and SIGMA~\cite{sigma} that use all-to-all interconnects to enhance utilization.
This, however, comes at the cost of high area-power costs,
which is at direct odds with scalability.
The alternate approach is to design
an accelerator as a
distributed \emph{scaled out}\footnote{
We use scale-up/scale-out interchangeably with monolithic/distributed.
} collection of smaller sub-accelerators, which is further corroborated by recent architecture proposals \cite{simba, tangram}. However, this leads to loss of operand reuse opportunities between the sub-accelerators.
\end{comment}
\begin{comment}
\insertFigure{scale-out-example-intro}{Challenges in efficiently scaling GEMM arrays.
(i) In case of irregular matrix dimensions, a monolithic configuration under performs due to mapping inefficiency, while a distributed configuration achieves max performance.
(ii) For regular matrix dimensions, both the arrays achieve similar performance, but the distributed configuration suffers from poor operand reuse. \TK{I think it'll be better to show monolithic on left and distributed on right instead of organizing based on workload. Also the text inside the brown and blue/green boxes is not visible at all -- make the color lighter and increase font size. Scale-up (Rigid) :
Scale-up (Flexible):
Scale-out
Scale-* (This work). And add boxes for Wire Scalability, Utilization and Operand Reuse. Come up with three examples, one that is bad for distributed.
}
}
\end{comment}
\begin{comment}
\insertFigure{scale-out-example-intro-new}{Caption TBD}
\TK{lets discuss three choices here: Scale-up (Rigid) :
Scale-up (Flexible), and scale-out. And lead this up to our proposal which is Scale-*?}
\autoref{fig:scale-out-example-intro} sheds light into the trade-off space.
In the case of irregular matrix dimensions (case (i)), the monolithic array is unable to extract full performance, due to mapping inefficiencies, while the distributed configuration has no such limitation.
However, the distributed configuration is unable to extract the benefits of operand reuse.
This makes the choice of distributed configurations inferior to a monolithic array, where matrices with regular dimensions are involved, as shown in case (ii).
The loss of spatio-temporal operand reuse is a consequence of two factors.
First, spatial reuse via wires is lost since communication paths between two distinct compute units are fundamentally lower-bandwidth than within the array---in the worst case requiring off-chip access~\cite{simba}.
Second, some operand data is needed to be replicated in the local storage of each device (case (ii)), which reduces the effective on-chip capacity.
This reduced capacity leads to loss of temporal reuse and results in increased off-chip accesses.
These factors make a distributed configuration inferior in energy efficiency.
\insertFigure{sara-concept}{
The constitution and interactions of the self adaptive (SA) and reconfigurable array (RA) components to make up the SARA accelerator called \mys{} in this work. \TK{change SMART systolic Array to Reconfigurable Systolic Array}
}
\textbf{Contribution 1:} In this paper, we propose a novel accelerator micro-architecture organization called \textsc{SmartSystolic} Array (\textit{RSA}{}) aimed to attain the benefits of monolithic (i.e., scale-up) and distributed (i.e., scale-out) designs for efficient GEMM computation in a single unified substrate.
We choose systolic arrays as our scale-up building blocks, since the simplicity of these arrays leads to low area and power overheads, maximizing local bandwidth while minimizing communication distance.
To mitigate the under-utilization problem, we propose augmenting the traditional systolic design with a bypass interconnection network inspired by SMART \cite{smart}.
These links permits us to emulate similar mapping to a distributed cluster of accelerator, within a monolithic compute array. The bypass interconnect can also be configured to emulate distributed systems of different granularities, eg. a system with $128\times128$ MAC (multiply and accumulate units) can be configured to be used as 4 $64\times64$ units or 16 $32\times32$ units or even 32 $32\times16$ units.
\textbf{Contribution 2: }
Determining the optimal architecture configuration and mapping strategy for each workload (eg. each layer in a DNN) is critical to obtain the performance and energy benefits from a reconfigurable array.
Currently such architectures solely rely on software for this task.
The software methods~\cite{timeloop, gamma, autotvm} are constrained to operate in a static setting, since their runtimes are in the order of a few seconds to hours.
Furthermore, the complete reliance on software also hinders the deployability of such new architectures, as a significant effort is needed to update the software stack to incorporate the cost model and configuration search capability.
In this work, we aim to reduce this cross-stack dependency by incorporating a reconfiguration unit called \textit{Self Adaptive (SA)} unit (see \autoref{fig:sara-concept}), which is capable of predicting the optimal configuration as the workloads arrive at runtime.
We developed a novel light weight neural recommendation model called \recnet, to recommend the optimal configuration of the array depending on the incoming workload at runtime.
The recommender is capable of predicting attain about 99.93\% of the best runtime on average (GeoMean) in our tests with 200K samples.
\textbf{Contribution 3: }
We also design a custom hardware for running \recnet, called \core, and augment it with the reconfigurable accelerator.
The custom design of \core~ enables the SA unit to generate both the optimal mapping and hardware configuration in approximately 600 cycles, which is at least about 6 orders of magnitude faster than software.
The resulting design is self sufficient for providing optimal performance.
Furthermore, the SA unit consume the same hardware real-estate and on-chip memory capacity for different arrays and with number of workloads thus proving to be a scalable solution in contrast to approaches like using configuration caches.
\textbf{Contribution 4: }
We integrate these three components into an accelerator which we call `Shape Adaptive GEMM AcceleratoR (\mys{} \footnote{Sagar is a Sanskrit word that means Ocean, reflecting the ability of our accelerator to have flexible shapes})' as shown in \autoref{fig:sara-concept}
and evaluate its performance across various
configurations.
%
We show that \mys{} has
3.2$\times$ higher compute density and
3.5$\times$ improved power efficiency, over equivalent scaled-out systolic array.
The extra flexibility costs $<$10\% in area and
50\% in power, compared to equivalent scaled-up systolic array.
Compared to an area normalized state-of-the-art flexible scalable accelerator \cite{sigma}, \mys~ incorporates
45\% more compute,
while when comparing compute-equivalent configurations,
\mys~ consumes 43\% less power
and saves 30\% area.
We believe our proposed accelerator is the first in a class of designs we name \textit{Self Adaptive Reconfigurable Array (SARA)} (\autoref{fig:sara-concept}).
To summarize, we make the following contributions.
(i) We propose \textbf{\textit{RSA}}, a reconfigurable architecture for scalable GEMM acceleration, simultaneously achieving high utilization and reuse.
(ii) We develop \textbf{\recnet}, a lean recommendation neural net which suggests optimized configuration and dataflow
with high accuracy.
(iii) We implement \textbf{\core}, a hardware capable of running \recnet{} in constant time and configuring \textit{RSA}{} at runtime.
(iv) We integrate the above components into a SARA accelerator called \textbf{\mys{}}, which achieves optimal runtime at lower power and area than SOTA.
\end{comment}
\begin{comment}
\item{We develop a neural network recommendation engine, which can predict the most performant architecture and mapping configuration per workload, at runtime with very high accuracy.
To the best of our knowledge this is the first work to do so.}
\item{We demonstrate the first instance of a Self Adaptive Reconfigurable Architecture (SARA); a class of architectures capable of independently selecting and implementing the most optimal configuration for a given workload at runtime.}
\squishend
\end{comment}
\section{Introduction}
\label{sec:introduction}
\insertFigure{scale-out-example-intro-new}{Comparison of scalability, utilization, and operand reuse in traditional monolithic and distributed accelerators, and the position of the proposed architecture}
\insertFigure{sara-concept}{
The constitution and interactions of the self adaptive (SA) and reconfigurable array (RA) components to make up the SARA accelerator called \mys{} in this work.
}
Custom architecture design enables us to achieve high performance and energy efficiency for a given class of workloads in post Moore's Law era.
Highly specialized architectures however are inflexible to any variation in the nature of workload and thus can easily be rendered obsolete.
To mitigate this limitation, there has been an increasing interest in developing flexible architectures which have additional components (interconnects, buffers, and configuration registers) to support changing workload requirements.
For popular applications like DNN acceleration, several such flexible architectures have been proposed \cite{brainwave, zhang-fpga-2015, tetris, tangram, maeri}.
In all of the prior works on flexible DNN accelerators, however,
the onus of finding and setting the best configuration lies on the software stack, typically using a compiler/mapper~\cite{gamma, timeloop, mindmappings, zhao2019mrna}.
This dependence causes a few deployment challenges: (i) a cost model has be to developed and integrated as an optimizer into the compilation stack to help find optimal mappings, without which the flexible design loses utility,
(ii) an expensive configuration and mapping search has to be performed at compile-time before scheduling any workload.
Usually mapping search in software is performed via exhaustive, heuristic or optimization algorithm-based approaches which take about a few seconds to hours~\cite{timeloop, gamma, zhao2019mrna}, even with sophisticated ML assisted frameworks like autoTVM \cite{autotvm}.
(iii) the search-time overhead also eliminates opportunities for deploying such flexible accelerators for domains/applications with soft or hard-real time inference targets.
In this work, we demonstrate that the mapping and configuration space of a reconfigurable accelerator can be \textit{learnt} by a machine learning (ML) model,
which can then be used to query for optimal parameters for any workload at constant time.
Dependence on software stack can be eliminated by incorporating this learnt model into the hardware itself and querying it in runtime.
We illustrate this via two contributions:
\textbf{\textit{First,}} we design a scalable reconfigurable hardware optimized for GEMM workloads called \textsc{Reconfigurable Systolic Array} (\textit{RSA}{}).
\textit{RSA}{} is developed upon the intuition that flexible accelerators often need to trade-off utilization, data reuse, and hardware
complexity (i.e., scalability).
This is illustrated in \autoref{fig:scale-out-example-intro-new}.
\textit{Rigid Monolithic} arrays (e.g., TPU's systolic array~\cite{tpu}),
are simple to construct but offer no flexibility leading to high under -utilization for many workloads~\cite{sigma, scalesim-ispass}.
\begin{comment} offers no flexibility. It works well for regular GEMM workloads whose dimensions fit on the array, but can lead to high under-utilization when running GEMMs with irregular (non-square) dimensions as several prior works have shown~\cite{sigma, scalesim-ispass}.\end{comment}
\textit{Flexible Monolithic} arrays (e.g., MAERI~\cite{maeri}, Eyeriss\_v2~\cite{eyerissv2}, SIGMA~\cite{sigma}) provide flexbility via clever use of interconnects and configuration logic, enabling high utilization for a majority of workloads.
However, the increased hardware complexity hinders scaling,
and the design requires external software support to exploit the benefits of reconfigurability~\cite{timeloop, mindmappings, gamma}.
Distributed architectures (e.g., Tangram~\cite{tangram}, Simba~\cite{simba})
\begin{comment} build accelerators as a collection of
smaller accelerator cores connected via a network-on-chip.
This \end{comment}
help address the utilization challenge, since irregular workloads can be tiled on to these smaller arrays.
However, this architecture leads to loss in spatial reuse (i.e., direct data-forwarding) that monolithic designs provide,
and also requires data replication across the SRAMs of the individual arrays.
Data replication leads to a decrease in overall on-chip storage capacity, leading to a loss of temporal reuse due to smaller tiles.
Moreover, distributed arrays can exacerbate the mapping search problem~\cite{simba, timeloop}.
\textit{RSA}{} aims to address the shortcomings of all three design strategies. It is a flexible accelerator capable of supporting mappings that can be realized by monolithic as well as distributed arrays by configuring to variable array dimensions and number of sub-arrays (as depicted later in \autoref{fig:high-bw-smart-systolic-array}(d)), thereby enhancing both utilization and reuse.
In practice, \textit{RSA}~closely approximates a flexible monolithic design, with a fraction of area cost.
\textbf{\textit{Second,}} we present a systematic mechanism to cast the architecture configuration as a ML classification problem and discuss considerations for optimal model design, training, and performance of the model at inference. Specifically, we develop a custom ML recommendation system model called \recnet{} that achieves a recommendation accuracy of 95\% on a dataset of 200K GEMM workloads, and on average(GeoMean) 99.93\% of the best attainable performance (Oracle).
We also design a custom hardware unit to run \recnet{} called \core{}.
\core{} enables to get a recommendation response for any query in about 600 cycles which is at least about 6 orders of magnitude faster than software.
Furthermore, \core{} consumes the same hardware real-estate and roughly the same on-chip memory capacity
\footnote{The only change in \recnet{} between various \textit{RSA}{} is the weight of the output layer, which is small in comparison to the embedding table which takes most of the on-chip space}
for different arrays, thus proving to be a scalable solution in contrast to approaches like using configuration caches.
With \core{} the configuration lookup using \recnet{} can be performed at runtime, without involving the software stack.
Together, these two components enable us to develop
a new class of accelerators that we call \textit{Self Adaptive Reconfigurable Array (SARA)} (\autoref{fig:sara-concept}).
SARA accelerators can self adapt at runtime to optimized configurations for the target workload, without
requiring compile-time analysis.
We demonstrate an instance of SARA that we name `Shape Adaptive GEMM AcceleratoR (\mys{}
\footnote{means Ocean in Sanskrit, reflecting the shape flexibility of our accelerator.})
as shown in \autoref{fig:sara-concept}
and evaluate its performance across various
configurations.
%
We show that \mys{} has
3.2$\times$ higher compute density and
3.5$\times$ improved power efficiency, over equivalent scaled-out systolic array.
The extra flexibility costs $<$10\% in area and
50\% in power, compared to equivalent scaled-up systolic array.
Compared to an area normalized state-of-the-art flexible scalable accelerator \cite{sigma}, \mys~ incorporates
45\% more compute.
When comparing compute-equivalent configurations,
\mys~ consumes 43\% less power and 30\% less area.
\section{Background and Motivation}
\insertFigure{motivation-data-scaleup-scaleout}{
The trade-off between improved runtime and lost operand reuse in compute equivalent monolithic and distributed systolic array configurations.
(a) the theoretical minimum runtime, and the runtime obtained for stall free operation of monolithic and compute normalized distributed systolic array settings; and
(b) the corresponding SRAM reads, normalized to theoretical minimum reads required when multiplying a $256\times64$ matrix with another $64\times256$ matrix.
}
\vspace{-2mm}
\subsection{Motivation}
\label{subsec:motivation}
To help understand the trade offs involved in choosing a performant configuration, and the associated loss of reuse we perform a simple experiment.
We run one GEMM operation, involving operand matrices of sizes sizes $256 \times 64$ and $64 \times 256$ on various systolic array configurations.
These are, a $128 \times 128$ monolithic array, and five distributed scale-out configurations viz. 4 $64 \times 64$ arrays, 16 $32 \times 32$ arrays, 64 $16 \times 16$ arrays, 256 $8 \times 8$ arrays, and 1024 $4 \times 4$ arrays.
We obtain the runtime and memory accesses for running this workload on all the array configurations using SCALE-Sim \cite{scalesim-arxiv} (see \autoref{subsec:sim-eval}).
In \autoref{fig:motivation-data-scaleup-scaleout}(a) we show the runtime normalized to the theoretical minimum cycles required.
Please note that with the chosen matrix dimensions, the systolic arrays in all the configurations are mapped 100\% with useful computation.
The differences in runtime in various arrays under 100\% mapping efficiency is attributed to the array filling and draining at each serialization step (see sec III in \cite{scalesim-ispass}).
We observe that the configuration with $32\times32$ array is the most performant, beating the monolithic configuration by about 2$\times$.
In \autoref{fig:motivation-data-scaleup-scaleout}(b) we depict the SRAM read accesses performed by all the array configurations, normalized
to the theoretical minimum number of reads possible.
From this figure we observe that the $32\times32$ configuration performs about 4$\times$ more memory accesses then the monolithic.
The excess memory accesses, which lead to reduced energy efficiency, result from the loss of wire reuse.
From the discussion above we make two observations.\\
(i) Distributed arrays are more performant than an equivalent monolithic array, even when mapping efficiency is 100\% on both.
However, the optimal size of each device in a distributed setting is workload dependent.
(ii) Monolithic configurations are strictly more energy efficient than distributed arrays, due to loss the of spatio-temporal reuse in the latter.
\input{tables/related-works-table}
In \autoref{table:related-works} we inspect a few well known accelerator proposals in terms of scalability and potential to maximize utilization.
We notice that simple architectures that are easy to scale in size, under perform on extracting operand reuse. On the other hand, architectures with sufficient flexibility are not scalable.
None of the architectures, including the ones with multiple arrays and NoC support, can create variable sized arrays or flexible array dimensions which can help simultaneously achieve high mapping efficiency and data reuse.
\section{Self Adaptive Reconfigurable Arrays}
\section{\textbf{\textsc{Self Adaptive Reconfigurable Arrays}}}
By coupling \recnet{} with a reconfigurable array, we can create a self adaptive system which can be conceptually viewed as a combination of two units, a Self Adaptive unit (SA), and a Reconfigurable Array (RA) unit as shown in \autoref{fig:sara-concept}.
The SA unit encompasses the software and hardware components which recommend the optimal configurations.
The RA unit is the hardware unit capable of flexibly configuring to the recommended configurations and hence run the workloads.
It is worth pointing out that this design class is not specific to a reconfigurable core for running GEMM workloads.
Instead any Coarse Grained Reconfigurable Array (CGRA) unit, configurable at runtime, can be augmented with a suitable SA, to ensure optimal performance.
We believe this results in a new class of designs, which we name Self Adaptive Reconfigurable Array (SARA).
\vspace{-2mm}
\subsection{Hardware to run \recnet}
In the context of our use case, an intuitive option is to allocate a few \textit{systolic-cell} s from the main array to run \recnet.
However, this choice will lead to either fewer MAC units left for the actual workloads, or to allocate additional \textit{systolic-cell} s for \recnet~ leading to an additional overhead.
An alternative to adding more \textit{systolic-cell} s will be to add a custom hardware dedicated for running \recnet.
We explore both the \textit{systolic-cell}~and custom hardware options below for \recnet-858.
\textbf{\recnet{} Runtime on \textit{systolic-cell} s.}
\autoref{fig:adaptnet-x-case}(a) shows the cycles required for a single inference of the \recnet ~as a function of multipliers used in $4\times4$ \textit{systolic-cell} ~based array.
Understandably, the runtime decreases proportional to the increase in number of multipliers as we increase the number of \textit{systolic-cell} s, achieving the best runtime of 1134 cycles when using 1024 multipliers or 64 cells.
When both the workloads and the recommendation engine is run on a same array; for a TPU equivalent machine with $2^{14}$ MAC units, about 6.25\% of the array needs to be allocated for running the \recnet.
Another choice could be allocating more hardware resources in terms of extra 64 \textit{systolic-cell} s dedicated to run the recommender network.
However, given that \recnet~ has exclusively dense layers processing the embedding lookups, a systolic execution turns out to be sub-optimal.
\textbf{\recnet{} Runtime on \core.} We found a custom design tuned for \recnet~ layer parameters to be more efficient.
For efficient execution of the dense layers, we chose a 1-D multiplier unit with a binary tree based reduction as shown in \autoref{fig:adaptnet-x-case}(b).
We found Input stationary (IS) dataflow to be the most performant for our use case.
In this mapping the elements of the input vector is buffered near the multipliers, while elements of the weight matrix are streamed through to generate one output element/partial sum, with a sustained throughput of 1 element per cycle.
Throughput can be further increased by adding more such 1-D units.
We name the custom core with one or more such 1-D units as \core.
In \autoref{fig:adaptnet-x-case}(a) we depict the variation of runtime of \recnet~ inference on \core~ with two 1-D units as a function of multipliers.
We find the 512 multipliers result in best runtime of 576 cycles, when running \recnet{} for $2^{14}$ MAC unit \textit{systolic-cell}{} design.
We also examine the cost of misprediction of \recnet{} in \autoref{fig:adaptnet-x-case}(c), where we plot the runtime of the predicted configurations from \recnet-858 normalized to best possible runtime.
We see that most mispredictions are benign and only a few misprediction lead to catastrophic performance losses, leading to a geometric mean of 99.93\% of the best possible performance.
\vspace{-2mm}
\subsection{\mys{} Accelerator}
\mys~ is constructed by augmenting the $2^{14}$ MAC \textit{RSA}{} unit, laid out as $32\times32$ grid of \textit{systolic-cell} s, with \core ~running \recnet-858 (see \autoref{fig:sara-system}).
We chose this configuration as it has the same compute as the TPU v2, and the $4\times4$ \textit{systolic-cell} ~size works the best for our workloads (see \autoref{subsec:sim-eval}).
Since each row and column in this configuration has 31 bypass links and one link to MAC, each buffer is constructed as a collection of 1024 1KB banks.
\textbf{Real-time Reconfiguration.}
The \core ~uses an additional SRAM bank of 512KB to store the embedding table and the weight matrices for \recnet-858.
Each configuration corresponds to a 3968 bit vector which sets the bypass muxes, once the layer is ready to be mapped.
\insertFigure{sara-system}{Schematic of \mys{}, an instance of a SARA accelerator.}
\section{Related Works}
\label{sec:related-work}
\textbf{Flexible DNN Accelerator}.
\autoref{table:related-works} depicts the standing of various such accelerators in term of native operation supported, mapping capability and flexibility.
Designs like\cite{neurocube}, \cite{flexflow}, \cite{zhang-fpga-2015}, \cite{planaria}, \cite{brainwave}, \cite{sigma}, \cite{maeri}, \cite{cascades}, \cite{zhang-fpga-2015, alwani-2016-MICRO-fused, tangram} have limited flexibility in either reconfigurability or dataflow.
The \textsc{Reconfigurable Systolic Array} enables both mapping flexibility and reconfigurability.
\textbf{Dataflow and Accelerator Design Space Search.}
Contemporary tools \cite{scalesim-ispass}, \cite{maestro}, \cite{tetris}, \cite{timeloop}, \cite{dmazerunner} etc enable DSE by fast cost estimation or heuristics.
SARA systems like \mys{} on the other hand obtain optimized configuration at runtime in one-shot using \recnet.
\textbf{ML assisted system configuration.}
Recent work \cite{gamma, confuciux, nautilus} show using GAs for efficient search.
RL and recommendation has been used for chip PnR \cite{mirhoseini2017device, kwon2019learning}.
AutoTVM \cite{autotvm} use ML models for cost prediction to improve compilation time.
It is worth noting that these approaches mostly enhance search for the optimal configuration, while \recnet~ replaces search.
\section{Related Works}
\label{sec:related-work}
\textbf{Flexible DNN Accelerator}.
\autoref{table:related-works} depicts the standing of various such accelerators in term of native operation supported, mapping capability and flexibility.
To efficiently execute a variety of workloads, DNN accelerator designs generally come with two tiers of flexibility, architecture and dataflow.
Designs like Neurocube\cite{neurocube}, Flexflow\cite{flexflow}, and by FPGA based designs\cite{zhang-fpga-2015} enable flexible mapping by supporting multiple dataflow.
On the other hand proposals like Planaria \cite{planaria}, Brainwave\cite{brainwave}, SIGMA \cite{sigma}, MAERI\cite{maeri}, Cascades\cite{cascades} and others \cite{zhang-fpga-2015, alwani-2016-MICRO-fused, tangram} enable reconfiguration at the hardware level.
\textit{RSA}{} enables both mapping flexibility and reconfigurability.
\textbf{Dataflow and Accelerator Design Space Search.} Several architecture and mapping space exploration tools have been proposed in the recent past to take advantage of flexibilities in the design.
Tool like SCALE-Sim\cite{scalesim-ispass}, MAESTRO\cite{maestro}, Tetris\cite{tetris} etc. provide analytical models for fast cost estimation of specific configurations.
While Timeloop\cite{timeloop}, dMazeRunner\cite{dmazerunner} etc are tools which perform heuristic or exhaustive search for architecture configuration or mapping strategy.
SARA systems like \mys{} on the other hand use a trained recommender like \recnet{} to circumvent the search and obtain the optimal configuration and dataflow in one shot at runtime.
\textbf{ML assisted system configuration.} Recent works have demonstrated the use of ML algorithms to assist in system configuration.
Gamma\cite{gamma} and ConfuciuX\cite{confuciux} perform architecture mapping and design space configuration search using genetic algorithm and reinforcement learning (RL).
On more systems size, work by
Mirhoseni et al\cite{mirhoseini2017device} use RL for task placement on a heterogenous system, while modern compilers like AutoTVM\cite{autotvm} use ML models for cost prediction to improve compilation time.
Nautilus\cite{nautilus} uses genetic algorithm to improve FPGA place and route.
It is worth noting that these approaches mostly enhance search for the optimal configuration, and this unlike \recnet{} do not replace search.
Perhaps the closest to our approach is work by Kwon et al\cite{kwon2019learning}, who use online tensor-based recommender systems to aid place and route in chip design.
|
\section{Introduction}\label{sec1}\setcounter{equation}{0}
The Burgers-Huxley equation is a special type of nonlinear advection-diffusion-reaction problems that are of importance in applications in mechanical engineering, material sciences, and neurophysiology. Some examples include, for instance, particle transport \cite{JS}, dynamics of ferroelectric materials \cite{OYNA}, action potential propagation in nerve fibers \cite{XYW}, wall motion in liquid crystals \cite{wang85}, and many others (see also \cite{VJE,MTAR} and the references therein).
Our starting point is the following stationary form of the generalized Burgers-Huxley equation with Dirichlet boundary conditions
\begin{equation}\label{8.1}
\left\{
\begin{aligned}
-\nu\Delta u+\alpha u^{\delta}\sum\limits_{i=1}^d\frac{\partial u}{\partial x_i}-\beta u (1-u^{\delta} )(u^{\delta} -\gamma)&=f, \ \text{ in }\ \Omega, \\
u&=0,\ \text{ on }\ {\partial}\Omega,
\end{aligned}\right.
\end{equation}
where it is assumed that $\Omega\subset \mathbb{R}^d \ (d=2,3)$ is an open bounded and simply connected domain with Lipschitz boundary $\partial\Omega$. Here $\nu>0$ is the constant diffusion coefficient,
$\alpha>0$ is the advection coefficient, and $\beta>0$, $\delta\geq 1$, $\gamma\in(0,1)$ are model parameters modulating the interplay between non-standard nonlinear advection, diffusion, and nonlinear reaction (or applied current) contributions.
The global solvability of the one-dimensional Burgers-Huxley equation has been recently established in \cite{MTAR}. In this paper we extend the analysis to the multi-dimensional case. Drawing inspiration from the techniques usually employed for the analysis of steady Navier-Stokes equations (cf. \cite[Ch. 10]{Te}), we use a Faedo-Galerkin approximation, Brouwer's fixed-point theorem, and compactness arguments to derive the existence and uniqueness of weak solutions to the two- and three-dimensional stationary generalized Bur\-gers-Huxley equation in bounded domains with Lipschitz boundary and under a minimal regularity assumption. For the case of domains that are convex or have $C^2-$boundary, we employ the elliptic regularity results available in, e.g., \cite{brezis_book11,grisvard_book85}, and establish that the weak solution of \eqref{8.1} satisfies $u\in H^2(\Omega)\cap H^1_0(\Omega)$.
The recent literature relevant to the construction and analysis of discretizations for \eqref{8.1} and closely related problems is very diverse. For instance, numerical methods specifically designed to capture boundary layers in
singularly perturbed generalized Bur\-gers-Huxley equations have been studied in \cite{kumar11},
different types of finite differences have been used in \cite{sari11,macias18,shukla20,verma20}, spectral, B-spline and Chebyshev wavelet collocation methods
have been advanced in \cite{alinia19,javidi06,wasim18,celik16}, numerical solutions obtained with the so-called adomain decomposition were analyzed in \cite{hashim06}, homotopy perturbation techniques were used in
\cite{maurya19}, Strang splittings were proposed in \cite{cicek16}, meshless radial basis functions were studied in \cite{khattak09},
generalized finite differences and finite volume schemes have been analyzed in \cite{chen03,zhou19} for the restriction of \eqref{8.1} to the diffusive Nagumo (or bistable) model, and a finite element method satisfying a discrete maximum principle was introduced in \cite{VJE} (the latter reference is closer to the present study).
Although there is a growing interest in developing numerical techniques for the generalized Bur\-gers-Huxley equation, it appears that
the aspects of error analysis for finite element discretizations have not been yet thoroughly addressed. Then, somewhat differently from the methods listed above (where we stress that such list is far from complete), here we propose a family of schemes consisting of conforming finite elements (CFEM), non-conforming finite elements (NCFEM) and discontinuous Galerkin methods (DGFEM). Following the assumptions adopted for the continuous problem, we rigorously derive a priori error estimates indicating first-order convergence of the CFEM. In contrast, for NCFEM and DGFEM the solvability of the discrete problem does not follow from the continuous problem, but separate conditions are established to ensure the existence of discrete solutions in these cases. The minimal assumptions on the domain are also used to prove first-order a priori error bounds for NCFEM and DGFEM, and we briefly comment about $L^2-$estimates. We also include a set of computational tests that confirm the theoretical error bounds and which also show some properties of the model equation.
We have organized the remainder of the paper as follows: Section \ref{sec2} contains notational conventions and it presents the well-posedness and regularity analysis of (\ref{8.1}), discussing also some possible modifications to the proofs of existence and uniqueness of weak solutions. The numerical discretizations are introduced and then a priori error estimates are derived for CFEM, NCFEM and DGFEM in Section \ref{sec3}. Finally, Section \ref{sec4} has a compilation of numerical tests in 2D and 3D that serve to illustrate our theoretical results.
\section{Solvability of the stationary generalized Burgers-Huxley equation}\label{sec2}\setcounter{equation}{0}
\subsection{Preliminaries}
Throughout this section we will adopt the usual notation for functional spaces.
In particular, for $p \in [1,\infty)$ we denote the Banach space of Lebesgue $p-$integrable functions by
$$ L^{p}(\Omega) :=\left\{u: \int_{\Omega}|u(x)|^{p}d x <\infty \right\},$$
whereas for $p = \infty$, $L^{\infty}(\Omega)$ is the space conformed by essentially bounded measurable functions
on the domain. Moreover, for integers $s\ge0$, by $H^s(\Omega)$ we denote the standard Sobolev spaces $W^{s,2}(\Omega)$, endowed with the norm
$\|u\|_{s,\Omega}^2 = \| u \|^2_{0,\Omega} + \sum_{|i|\leq s} \|\partial^i u\|^2_{0,\Omega}$. For $s=0$, we adopt the convention $H^0(\Omega)=L^2(\Omega)$, and recall the definition of the closure of all $C^\infty$ functions with
compact support in $H^1(\Omega)$
$H^1_0(\Omega) :=\{u\in H^1(\Omega): u|_{\partial \Omega}=0\ \text{a.e.}\}$.
If $Y(M)$ denotes a generic normed space of functions over the spatial domain $M$, then the associated norm will be
at some instances denoted as $\|\cdot\|_{Y}$ (omitting the domain specification whenever clear from the context).
In addition, let $H^{-1}(\Omega)$ be the dual space of the Sobolev space $H^1_0(\Omega)$ with the following norm
\begin{align*}
\|u\|_{H^{-1}(\Omega)}:=\sup_{0\neq v\in H^{1}_0(\Omega)}\frac{\langle u,v\rangle}{\|v\|_{1,\Omega}},
\end{align*}
where $\langle\cdot,\cdot\rangle$ denotes the duality pairing between $H_0^1(\Omega)$ and $H^{-1}(\Omega)$. In the sequel, we use the same notation for the duality pairing between $L^p(\Omega)$ and its dual $L^{\frac{p}{p-1}}(\Omega)$, for $p\in(2,\infty)$.
We proceed to rewrite problem (\ref{8.1}) in the following abstract form:
\begin{align}\label{8p2}
\nu Au +\alpha B(u )-\beta C(u )=f,
\end{align}
where the involved operators are
\begin{align*}
Au =-\Delta u, \quad B(u)= u^{\delta}\sum\limits_{i=1}^d\frac{\partial u}{\partial x_i},\quad \text{ and }\ C(u)=u (1-u^{\delta} )(u^{\delta} -\gamma).
\end{align*}
For the Dirichlet Laplacian operator $A$, it is well-known that $D(A)=H^2(\Omega)\cap H_0^1(\Omega)\subset L^p$, for $p\in[1,\infty)$ and $1\leq d\leq 4$, using the Sobolev Embedding Theorem (see, e.g., \cite{grisvard_book85}) and also $A:H_0^1(\Omega)\to H^{-1}(\Omega)$. Since $\Omega$ is bounded, the embedding $H_0^1(\Omega)\subset L^2(\Omega)$ is compact, and hence using the spectral theorem, there exists a sequence $0<\lambda_1\leq\lambda_2\leq\ldots\to\infty$ of eigenvalues of $A$ and an orthonormal basis $\{w_k\}_{k=1}^{\infty}$ of $L^2(\Omega)$ consisting of eigenfunctions of $A$ \cite[p. 504]{RDJL}. Furthermore, we have the following Friedrichs-Poincar\'e inequality: $\sqrt{\lambda_1}\|u\|_0\leq \|\nabla u\|_0$.
Testing \eqref{8.1} against a smooth function $v$, integrating by parts, and applying the boundary
condition, we end up with the following problem in weak form:
Given any $f\inH^{-1}(\Omega)$, find $u \in H_0^1(\Omega)$ such that
\begin{align}\label{8p3}
\nu(\nabla u ,\nabla v)+\alpha b(u,u,v) -\beta\langle C(u ),v\rangle=\langle f,v\rangle, \quad \text{ for all }\ v\inH_0^1(\Omega),
\end{align}
where $b(u,u,v)=\langle B(u ),v\rangle$.
\subsection{Existence of weak solutions}
Let us first address the well-posedness of \eqref{8.1} in two dimensions.
\begin{theorem}[Existence of weak solutions]\label{thm6.1}
For a given $f\inH^{-1}(\Omega)$, there exists at least one solution to the Dirichlet problem \eqref{8.1}.
\end{theorem}
\begin{proof}
We prove the existence result using the following steps.
\vskip 0.2 cm
\noindent
\textbf{Step 1: Finite dimensional system.} We formulate a Faedo-Galerkin approximation method. Let the functions $w_k=w_k(x),$ $k=1,2,\ldots,$ be smooth, the set $\{w_k(x)\}_{k=1}^{\infty}$ be an orthogonal basis of $H_0^1(\Omega)$ and orthonormal basis of $L^2(\Omega)$. One can take $\{w_k(x)\}_{k=1}^{\infty}$ as the complete set of normalized eigenfunctions of the operator $-\Delta$ in $H_0^1(\Omega)$.
For a fixed positive integer $m$, we look for a function $u_m\inH_0^1(\Omega)$ of the form
\begin{equation}\label{8p4}
u_m=\sum\limits_{k=1}^m\xi_m^kw_k,\ \xi_m^k\in\mathbb{R},
\end{equation}
and
\begin{equation}\label{8p5}
\nu(\nabla u_m,\nabla w_k)+\alpha b(u_m,u_m,w_k)-\beta\langle C(u_m),w_k\rangle=\langle f,w_k\rangle,
\end{equation}
for $k=1,\ldots,m$. The set of equations in \eqref{8p5} is equivalent to
\[
\nu Au_m+\alpha P_mB(u_m)-\beta P_mc(u_m)=P_mf.
\]
Equations \eqref{8p4}-\eqref{8p5} constitute a nonlinear system for $\xi_m^1,\ldots,\xi_m^m$. We invoke \cite[Lem. 1.4]{Te} (an application of Brouwer's fixed point theorem) to prove the existence of solution to such a system. Let us consider the space $W=\text{Span}\left\{w_1,\ldots,w_m\right\}$ and the associated scalar product $[\cdot,\cdot]=(\nabla\cdot,\nabla\cdot)$. We define the map $P=P_m$ as
\[
[P_m(u),v]=(\nabla P_m(u),\nabla v)=\nu(\nabla u,\nabla v)+\alpha b(u,u,v)-\beta \langle C(u),v\rangle -\langle f,v\rangle,
\]
for all $u,v\in W$. The continuity of $P_m$ can be verified in the following way
\begin{align*}
& |[P_m(u),v]|\\&\leq\left(\nu\|\nabla u\|_0+\frac{\alpha}{\delta+1}\|u\|_{L^{2(\delta+1)}}^{\delta+1}\right)\|\nabla v\|_0+\beta\left[(1+\gamma)\|u\|_{L^{2(\delta+1)}}^{\delta+1}+\gamma\|u\|_{0}\right]\|v\|_0\\&\quad+\|u\|_{L^{2(\delta+1)}}^{2\delta+1}\|v\|_{L^{2\delta+1}}\nonumber\\&\leq \left[\left(\nu+\frac{\beta\gamma}{\lambda_1^2}\right)\|\nabla u\|_0+\left(\frac{\alpha}{\delta+1}+\frac{\beta(1+\gamma)}{\lambda_1}\right)\|u\|_{L^{2(\delta+1)}}^{\delta+1}+\|u\|_{L^{2(\delta+1)}}^{2\delta+1}\right]\|\nabla v\|_0,
\end{align*}
for all $v\in H_0^1(\Omega)$. Using Sobolev's embedding, we know that $H_0^1(\Omega)\subset L^p(\Omega)$, for all $p\in[2,\infty)$, and hence the continuity follows. In order to apply \cite[Lem. 1.4]{Te}, we need to show that $$[P_m(u),u]>0, \ \text{ for } \ [u]=k>0,$$ where $[\cdot]$ denotes the norm on $W$, which is in turn the norm induced by $H_0^1(\Omega)$. We can then use Poincar\'e's, H\"older's and Young's inequalities to estimate $[P_m(u),u]$ as
\begin{align*}
& [P_m(u),u]\nonumber\\&=\nu\|\nabla u\|_{0}^2+\beta\gamma\|u\|_{0}^2+\beta\|u\|_{L^{2\delta+2}}^{2\delta+2}-\beta(1+\gamma)(u^{\delta+1},u)-(f,u)\nonumber\\
&\geq \nu\|\nabla u\|_{0}^2+\beta\gamma\|u\|_{0}^2+\beta\|u\|_{L^{2\delta+2}}^{2\delta+2}-\beta(1+\gamma)\|u\|_{L^{2\delta+2}}^{\delta+1}\|u\|_{0}-\|f\|_{H^{-1}}\|\nabla u\|_{0}\nonumber\\
&\geq \frac{\nu}{2}\|\nabla u\|_{0}^2+\beta\gamma\|u\|_{0}^2+\frac{\beta}{2}\|u\|_{L^{2\delta+2}}^{2\delta+2} -\frac{\beta\delta(1+\gamma)^{\frac{2(\delta+1)}{\delta}}}{2(\delta+1)}\left(\frac{\delta+2}{\delta+1}\right)^{\frac{\delta+2}{\delta}}|\Omega|-\frac{1}{2\nu}\|f\|_{H^{-1}}^2 \nonumber\\
&\geq\frac{\nu}{2}\|\nabla u\|_{0}^2-\frac{\beta\delta(1+\gamma)^{\frac{2(\delta+1)}{\delta}}}{2(\delta+1)}\left(\frac{\delta+2}{\delta+1}\right)^{\frac{\delta+2}{\delta}}|\Omega|-\frac{1}{2\nu}\|f\|_{H^{-1}}^2,
\end{align*}
where $|\Omega|$ is the Lebesgue measure of $\Omega$.
It follows that $[P_m(u),u]$ $> 0,$ for $\|u\|_{1}=\kappa,$ where $\kappa$ is sufficiently large. More precisely, the
analysis requires $$\kappa>\sqrt{\frac{2}{\nu}\left(\frac{\beta\delta(1+\gamma)^{\frac{2(\delta+1)}{\delta}}}{2(\delta+1)}\left(\frac{\delta+2}{\delta+1}\right)^{\frac{\delta+2}{\delta}}|\Omega|+\frac{1}{2\nu}\|f\|_{H^{-1}}^2\right)}. $$ Thus the hypotheses of \cite[Lem. 1.4]{Te} are satisfied and a solution $u_m$ to \eqref{8p5} exists.
\bigskip
\noindent\textbf{Step 2: Uniform boundedness.} Next we need to show that the solution $u_m$ is bounded. Multiplying \eqref{8p5} by $\xi_m^k$ and then adding from $k=1,\ldots,m$, we find
\begin{align}\label{8.9}
&\nu\|\nabla u_m\|_{0}^2+\beta\|u_m\|_{L^{2\delta+2}}^{2\delta+2}+\beta\gamma\|u_m\|_{0}^2\nonumber\\&=\beta(1+\gamma)(u_m^{\delta+1},u_m)+\langle f,u_m\rangle \nonumber\\&\leq\beta(1+\gamma)\|u_m\|_{L^{2\delta+2}}^{\delta+2}|\Omega|^{\frac{\delta}{2(\delta+1)}}+\|f\|_{H^{-1}}\|u_m\|_{1}\nonumber\\&\leq\frac{\beta}{2}\|u_m\|_{L^{2\delta+2}}^{2\delta+2}+\frac{\beta\delta(1+\gamma)^{\frac{2(\delta+1)}{\delta}}}{2(\delta+1)}\left(\frac{\delta+2}{\delta+1}\right)^{\frac{\delta+2}{\delta}}|\Omega|+\frac{\nu}{2}\|u_m\|_{1}^2+\frac{1}{2\nu}\|f\|_{H^{-1}}^2,
\end{align}
where we have used H\"older's and Young's inequalities. From \eqref{8.9}, we deduce that
\begin{align}\label{2p13}
\nu\|u_m\|_{1}^2+\beta\|u_m\|_{L^{2\delta+2}}^{2\delta+2}\leq\frac{\beta\delta(1+\gamma)^{\frac{2(\delta+1)}{\delta}}}{\delta+1}\left(\frac{\delta+2}{\delta+1}\right)^{\frac{\delta+2}{\delta}}|\Omega|+\frac{1}{\nu}\|f\|_{H^{-1}}^2.
\end{align}
\bigskip
\noindent\textbf{Step 3: Passing to the limit.}
We have bounds for $\|u_m\|_{1}^2$ and $\|u_m\|_{L^{2\delta+2}}^{2\delta+2}$ that are uniform and independent of $m$. Since $H_0^1(\Omega)$ and $L^{2\delta+2}(\Omega)$ are reflexive, using the Banach-Alaoglu theorem, we can extract a subsequence $\{u_{m_k}\}$ of $\{u_m\}$ such that
\[
\begin{cases}
u_{m_k}&\xrightarrow{w} u , \ \text{ in }\ H_0^1(\Omega), \ \text{ as }\ k\to\infty, \\
u_{m_k}&\xrightarrow{w} u , \ \text{ in }\ L^{2\delta+2}(\Omega), \ \text{ as }\ k\to\infty.
\end{cases}
\]
In two dimensions we have that $H_0^1(\Omega)\subsetL^{2\delta+2}(\Omega)$, thanks to the Sobolev embedding theorem. Since the embedding of $H_0^1(\Omega)\subsetL^2(\Omega)$ is compact, one can extract a subsequence $\{u_{m_{k_j}}\}$ of $\{u_{m_k}\}$ such that
\begin{align}\label{2p15}
u_{m_{k_j}}\to u , \ \text{ in }\ L^2(\Omega), \ \text{ as }\ j\to\infty.
\end{align}
Passing to limit in \eqref{8p5} along the subsequence $\{m_{k_j}\}$, we find that $u $ is a solution to \eqref{8p3}, provided one can show that
\begin{align*}
B(u_{m_{k_j}})\xrightarrow{w} B(u) , \ \text{ and }\ C(u_{m_{k_j}})\xrightarrow{w} C(u )\ \text{ in }\ H^{-1}(\Omega), \ \text{ as } \ j\to\infty.
\end{align*}
In order to do this, we first show that $b(u_{m_{k_j}},u_{m_{k_j}},v)\to b(u,u,v),$ for all $v\in C_0^{\infty}(\Omega)$. Then, using a density argument, we obtain that $ B(u_{m_{k_j}})\xrightarrow{w} B(u) \ \text{ in }\ H^{-1}(\Omega)$, as $j\to\infty$. Using an integration by parts, Taylor's formula \cite[Th. 7.9.1]{ciarlet_book13}, H\"older's inequality, the estimate \eqref{2p13}, and convergence \eqref{2p15}, we obtain
\begin{align}\label{215}
& |b(u_{m_{k_j}},u_{m_{k_j}},v)- b(u,u,v)|\nonumber\\&= \left|\frac{1}{\delta+1}\sum_{i=1}^2\int_{\Omega}(u_{m_{k_j}}^{\delta+1}(x)-u^{\delta+1}(x))\frac{\partial v(x)}{\partial x_i}d x\right|\nonumber\\&=\left|\sum_{i=1}^2\int_{\Omega}(\theta u_{m_{k_j}}(x)+(1-\theta)u(x))^{\delta}(u_{m_{k_j}}(x)-u(x))\frac{\partial v(x)}{\partial x_i}d x\right|\nonumber\\&\leq\|u_{m_{k_j}}-u\|_{0}\left(\|u_{m_{k_j}}\|_{L^{2(\delta+1)}}^{\delta}+\|u\|_{L^{2(\delta+1)}}^{\delta}\right)\|\nabla v\|_{L^{2(\delta+1)}}\nonumber\\&\to 0\ \text{ as } \ j\to\infty, \ \text{ for all } \ v\in C_0^{\infty}(\Omega).
\end{align}
Making use again of Taylor's formula, interpolation and H\"older's inequalities, we find
\begin{align}\label{216}
& |(C(u_{m_{k_j}})-C(u),v)|\nonumber\\&\leq(1+\gamma)\left|\int_{\Omega}(u_{m_{k_j}}^{\delta+1}(x)-u^{\delta+1}(x))v(x)d x\right|+\left|\int_{\Omega}(u_{m_{k_j}}(x)-u(x))v(x)d x\right| \nonumber\\
&\quad+\left|\int_{\Omega}(u_{m_{k_j}}^{2\delta+1}(x)-u^{2\delta+1}(x))v(x)d x\right|\nonumber\\&\leq (1+\gamma)(\delta+1)\int_{\Omega}\left|(u_{m_{k_j}}(x)-u(x))(\theta u_{m_{k_j}}(x)+(1-\theta)u(x))^{\delta}v(x)\right|d x\nonumber\\
&\quad+\int_{\Omega}\left|(u_{m_{k_j}}(x)-u(x))v(x)\right|d x\nonumber\\
&\quad+(1+2\delta)\int_{\Omega}\left|(u_{m_{k_j}}(x)-u(x))(\theta u_{m_{k_j}}(x)+(1-\theta)u(x))^{2\delta}v(x)\right|d x\nonumber\\
&\leq(1+\gamma)(\delta+1)\|u_{m_{k_j}}-u\|_{0}\left(\|u_{m_{k_j}}\|_{L^{2(\delta+1)}}^{\delta}+\|u\|_{L^{2(\delta+1)}}^{\delta}\right)\|v\|_{L^{2(\delta+1)}}\nonumber\\
&\quad+\|u_{m_{k_j}}-u\|_{0}\|v\|_{0}
+(1+2\delta)\|u_{m_{k_j}}-u\|_{L^{\delta+1}}\left(\|u_{m_{k_j}}\|_{L^{2(\delta+1)}}^{2\delta}+\|u\|_{L^{2(\delta+1)}}^{2\delta}\right)\|v\|_{L^{\infty}}\nonumber\\
&\leq \left((1+\gamma)(\delta+1
\left(\|u_{m_{k_j}}\|_{L^{2(\delta+1)}}^{\delta}+\|u\|_{L^{2(\delta+1)}}^{\delta}\right)\|v\|_{L^{2(\delta+1)}}+\|v\|_0\right)\|u_{m_{k_j}}-u\|_{0}\nonumber\\
&\quad+(1+2\delta)\|u_{m_{k_j}}-u\|_{0}^{\frac{1}{\delta}}\left(\|u_{m_{k_j}}\|_{L^{2(\delta+1)}}^{1-\frac{1}{\delta}}+\|u\|_{L^{2(\delta+1)}}^{1-\frac{1}{\delta}}\right)\times\nonumber\\
&\quad\quad\left(\|u_{m_{k_j}}\|_{L^{2(\delta+1)}}^{2\delta}+\|u\|_{L^{2(\delta+1)}}^{2\delta}\right)\|v\|_{L^{\infty}}\to 0\ \text{ as } \ j\to\infty, \ \text{ for all } \ v\in C_0^{\infty}(\Omega).
\end{align}
Moreover, $u $ satisfies \eqref{8p3} and
\begin{align}\label{8.14}
\nu\|u\|_{1}^2+\beta\|u\|_{L^{2\delta+2}}^{2\delta+2}\leq\frac{\beta\delta(1+\gamma)^{\frac{2(\delta+1)}{\delta}}}{\delta+1}\left(\frac{\delta+2}{\delta+1}\right)^{\frac{\delta+2}{\delta}}|\Omega|+\frac{1}{\nu}\|f\|_{H^{-1}}^2=:\widetilde{K},
\end{align}
which completes the existence proof.
\end{proof}
\subsection{Uniqueness of weak solution}
\begin{theorem}[Uniqueness]\label{thm6.2}
Let $f\inH^{-1}(\Omega)$ be given. Then, for
\begin{align}\label{8pp5}
\nu>\max\left\{\frac{4^{\delta}\alpha^2}{\beta},\frac{\beta}{\lambda_1}\left[4^{\delta}(1+\gamma)^2(1+\delta)^2-{2\gamma}\right]\right\},
\end{align}
where $\lambda_1$ is the first eigenvalue of the Dirichlet Laplacian operator, the solution of \eqref{8p3} is unique.
\end{theorem}
\begin{proof}
We assume $u $ and $v$ are two weak solutions of \eqref{8p3} and define $w:=u -v$. Then $w$ satisfies:
\begin{align}\label{8p17}
\nu(\nabla w,\nabla v)+\alpha\langle B(u )-B(v),v\rangle -\beta\langle C(u )-C(v),v\rangle =0,
\end{align}
for all $v\inH_0^1(\Omega)$. Taking $v=w$ in \eqref{8p17}, we have
\begin{align}\label{8p18}
\nu\|\nabla w\|_{0}^2&=-\alpha\langle B(u)-B(v),w\rangle+\beta \langle C(u )-C(v),w\rangle.
\end{align}
Then it can be readily seen that
\begin{align}\label{3.24}
&\beta\left[\langle u(1-u^{\delta})(u^{\delta}-\gamma)-v(1-v^{\delta})(v^{\delta}-\gamma),w\rangle \right]\nonumber\\&=-\beta\gamma\|w\|_{0}^2-\beta(u^{2\delta+1}-v^{2\delta+1},w)+\beta(1+\gamma)(u^{\delta+1}-v^{\delta+1},w).
\end{align}
Let us take the term $-\beta(u^{2\delta+1}-v^{2\delta+1},w)$ from \eqref{3.24} and estimate it using H\"older's and Young's inequalities as
\begin{align}\label{3.25}
-\beta(u^{2\delta+1}-v^{2\delta+1},w)&=-\beta(|u|^{2\delta}(u-v)+|u|^{2\delta}v-|v|^{2\delta}u,w+|v|^{2\delta}(u-v),w)\nonumber\\
&=-\beta\|u^{\delta}w\|_{0}^2-\beta\|v^{\delta}w\|_0^2-\beta(|u|^{2\delta}+|v|^{2\delta},uv)+\beta(|u|^2,|v|^{2\delta})\nonumber\\&\quad+\beta(|v|^2,|u|^{2\delta})\nonumber\\&= -\frac{\beta}{2}\|u^{\delta}w\|_{0}^2-\frac{\beta}{2}\|v^{\delta}w\|_0^2-\frac{\beta}{2}((|u|^{2\delta}-|v|^{2\delta}),(|u|^2-|v|^2))\nonumber\\&\leq -\frac{\beta}{2}\|u^{\delta}w\|_{0}^2-\frac{\beta}{2}\|v^{\delta}w\|_0^2.
\end{align}
Next, we take the term $\beta(1+\gamma)(u^{\delta+1}-v^{\delta+1},w)$ from \eqref{3.24} and estimate it using Taylor's formula, H\"older's and Young's inequalities as
\begin{align}\label{3.26}
& \beta(1+\gamma)(u^{\delta+1}-v^{\delta+1},w)\nonumber\\&=\beta(1+\gamma)(\delta+1)((\theta u+(1-\theta)v)^{\delta}w,w)\nonumber\\&\leq \beta(1+\gamma)(\delta+1)2^{\delta-1}(\|u^{\delta}w\|_{0}+\|v^{\delta}w\|_{0})\|w\|_{0}\nonumber\\&\leq\frac{\beta}{4}\|u^{\delta}w\|_{0}^2+\frac{\beta}{4}\|v^{\delta}w\|_{0}^2+\frac{\beta}{2}2^{2\delta}(1+\gamma)^2(\delta+1)^2\|w\|_{0}^2.
\end{align}
Combining \eqref{3.25}-\eqref{3.26} and substituting the result back into \eqref{3.24}, we obtain
\begin{align}\label{327}
&\beta\left[(u(1-u^{\delta})(u^{\delta}-\gamma)-v(1-v^{\delta})(v^{\delta}-\gamma),w)\right]\nonumber\\
&\leq -\beta\gamma\|w\|_{0}^2-\frac{\beta}{4}\|u^{\delta}w\|_{0}^2-\frac{\beta}{4}\|v^{\delta}w\|_{0}^2+\frac{\beta}{2}2^{2\delta}(1+\gamma)^2(\delta+1)^2\|w\|_{0}^2.
\end{align}
On the other hand, we derive a bound for $-\alpha\langle B(u)-B(v),w\rangle$ using an integration by parts, Taylor's formula, H\"older's and Young's inequalities. This gives
\begin{align}\label{6.49}
-\alpha\langle B(u)-B(v),w\rangle&=\frac{\alpha}{\delta+1} \left((u^{\delta+1}-v^{\delta+1})\left(\begin{array}{c}1\\1\end{array}\right),\nabla w\right)\nonumber\\&=\alpha \left((u-v)(\theta u+(1-\theta)v)^{\delta}\left(\begin{array}{c}1\\1\end{array}\right),\nabla w\right)\nonumber\\&\leq 2^{\delta-1}\alpha\|\nabla w\|_{0}\left(\|u^{\delta} w\|_{0}+\|v^{\delta} w\|_{0}\right)\nonumber\\
&\leq\frac{\nu}{2}\|\nabla w\|_{0}^2+\frac{2^{2\delta}\alpha^2}{4\nu}\|u^{\delta} w\|_{0}^2+\frac{2^{2\delta}\alpha^2}{4\nu}\|v^{\delta} w\|_{0}^2. \end{align}
Combining \eqref{327}-\eqref{6.49}, and substituting that back in \eqref{8p18}, we further have
\begin{align}\label{8.19}
&\left[ \frac{ \nu}{2}+\frac{1}{\lambda_1}\left(\beta\gamma-\frac{\beta}{2}2^{2\delta}(1+\gamma)^2(\delta+1)^2\right)\right]\|\nabla w\|_{0}^2\nonumber\\&\quad+\left(\frac{\beta}{4}-\frac{2^{2\delta}\alpha^2}{4\nu}\right)\|u^{\delta} w\|_{0}^2+\left(\frac{\beta}{4}-\frac{2^{2\delta}\alpha^2}{4\nu}\right)\|v^{\delta} w\|_{0}^2\leq 0.
\end{align}
It should also be noted that
\begin{align*}
\|u-v\|_{L^{2\delta+2}}^{2\delta+2}&=\int_{\Omega}|u(x)-v(x)|^{2\delta}|u(x)-v(x)|^2d x\nonumber\\&\leq 2^{2\delta-1}(\|u^{\delta}(u-v)\|_{0}^2+\|v^{\delta}(u-v)\|_{0}^2).
\end{align*}
Thus from \eqref{8.19}, it is immediate to see that
\begin{align*}
&\left[ \frac{ \nu}{2}+\frac{1}{\lambda_1}\left(\beta\gamma-\frac{\beta}{2}4^{\delta}(1+\gamma)^2(\delta+1)^2\right)\right]\|\nabla w\|_{0}^2+\frac{1}{2^{2\delta+1}}\left(\beta-\frac{4^{\delta}\alpha^2}{\nu}\right)\|w\|_{L^{2\delta+2}}^{2\delta+2}\leq 0,
\end{align*}
and for the condition given in \eqref{8.19}, the uniqueness readily follows.
\end{proof}
\subsection{Possible modifications in the proofs, and a regularity result}
\begin{remark}
If one uses Gagliardo-Nirenberg interpolation inequality to estimate the term $-\alpha\langle B(u)-B(v),w\rangle$, then it can be easily seen that
\begin{align}\label{2.31}
-\alpha\langle B(u)-B(v),w\rangle&\leq \alpha\|\nabla w\|_{0}\|w\|_{L^{2(\delta+1)}}\left(\|u\|_{L^{2(\delta+1)}}^{\delta}+\|v\|_{L^{2(\delta+1)}}^{\delta}\right)\nonumber\\
&\leq C\alpha\|\nabla w\|_{0}^{\frac{2\delta+1}{\delta+1}}\left(\|u\|_{L^{2(\delta+1)}}^{\delta}+\|v\|_{L^{2(\delta+1)}}^{\delta}\right)\|w\|_{0}^{\frac{1}{\delta+1}}\nonumber\\
&\leq \frac{C\alpha}{\lambda_1^{\frac{1}{2(\delta+1)}}}\left(\|u\|_{L^{2(\delta+1)}}^{\delta}+\|v\|_{L^{2(\delta+1)}}^{\delta}\right)\|\nabla w\|_{0}^2\nonumber\\
&\leq\frac{2C\alpha}{\lambda_1^{\frac{1}{2(\delta+1)}}}\sqrt{\frac{\widetilde{K}}{\beta}}\|\nabla w\|_{0}^2,
\end{align}
where $C$ is the constant appearing in the Gagliardo-Nirenberg inequality. Combining \eqref{327} and \eqref{2.31}, and substituting it in \eqref{8p18}, we get
\begin{align*}
&\left[\nu+\frac{1}{\lambda_1}\left(\beta\gamma-\frac{\beta}{2}2^{2\delta}(1+\gamma)^2(\delta+1)^2\right)-\frac{2C\alpha}{\lambda_1^{\frac{1}{2(\delta+1)}}}\sqrt{\frac{\widetilde{K}}{\beta}}\right]\|\nabla w\|_{0}^2\leq 0,
\end{align*}
Thus the uniqueness follows provided
\begin{align}\label{234}
\nu+\frac{\beta\gamma}{\lambda_1}>\frac{\beta}{\lambda_1}2^{2\delta-1}(1+\gamma)^2(\delta+1)^2+\frac{2C\alpha}{\lambda_1^{\frac{1}{2(\delta+1)}}}\sqrt{\frac{\widetilde{K}}{\beta}},
\end{align}
where $\widetilde{K}$ is defined in \eqref{8.14}.
\end{remark}
\begin{remark}
For $\delta=1$ (that is, for the classical Burgers-Huxley equation), we obtain a simpler condition than \eqref{8pp5} for the uniqueness of weak solution. In this case, the estimate \eqref{327} becomes (see \cite{MTAR})
\begin{align}\label{231}
&\beta\left[(u(1-u)(u-\gamma)-v(1-v)(v-\gamma),w)\right]\nonumber\\
&\leq -\beta\|uw\|_{0}^2-\beta\|vw\|_{0}^2+\beta(1+\gamma+\gamma^2)\|w\|_{0}^2.
\end{align}
Similarly, we estimate the term $-\alpha\langle B(u)-B(v),w\rangle$ as
\begin{align}\label{235}
-\alpha\langle B(u)-B(v),w\rangle&=-\alpha [b(w,w,w) +b(w,v,w)+b(v,w,w)]\nonumber\\&=\alpha b(v,w,w)\leq
\frac{\nu}{2}\|\nabla w\|_{0}^2+\frac{\alpha^2}{2\nu}\|vw\|_{0}^2.
\end{align}
Thus, as an immediate consequence we have that
\[
\left[\frac{\nu}{2}-\frac{\beta(1+\gamma+\gamma^2)}{\lambda_1}\right]\|\nabla w\|_{0}^2+ \beta\|uw\|_{0}^2+\left(\beta-\frac{\alpha^2}{2\nu}\right)\|uw\|_{0}^2\leq 0,
\]
and hence for $$\nu>\max\left\{\frac{2\beta(1+\gamma+\gamma^2)}{\lambda_1},\frac{\alpha^2}{2\beta}\right\},$$ the uniqueness of weak solution holds. To conclude, one can use the Ladyzhenskaya inequality to estimate $-\alpha\langle B(u)-B(v),w\rangle$. Then,
the bound \eqref{235} becomes
\begin{align}\label{237}
-\alpha\langle B(u)-B(v),w\rangle&=\alpha b(v,w,w)=\alpha\sum_{i=1}^2\int_{\Omega}\frac{\partial v(x)}{\partial x_i}w^2(x)dx \nonumber\\&\leq \alpha\|w\|_{L^4}^2\|\nabla v\|_{0}\leq \sqrt{2}\alpha\|w\|_{0}\|\nabla w\|_{0}\|\nabla v\|_{0}\nonumber\\
&\leq\sqrt{\frac{2}{\lambda_1}}\alpha\|\nabla v\|_{0}\|\nabla w\|_{0}^2\leq\sqrt{\frac{2\widetilde{K}}{\lambda_1\nu}}\alpha\|\nabla w\|_{0}^2,
\end{align}
where $\widetilde{K}$ is defined in \eqref{8.14}. Thus, combining \eqref{231} and \eqref{237}, we have
\[
\left[\nu-\sqrt{\frac{2\widetilde{K}}{\lambda_1\nu}}\alpha-\frac{\beta}{\lambda_1}(1+\gamma+\gamma^2)\right]\|\nabla w\|_{0}^2+ \beta\|uw\|_{0}^2+\beta\|uw\|_{0}^2\leq 0,
\]
and hence the uniqueness follows in this case for $\nu>\sqrt{\frac{2\widetilde{K}}{\lambda_1\nu}}\alpha+\frac{\beta}{\lambda_1}(1+\gamma+\gamma^2)$.
\end{remark}
\begin{remark}\label{rem2.3} For the three-dimensional case, the existence of weak solution to \eqref{8.1} can be
established for $1\leq \delta<\infty$. Since the proof of Theorem \ref{thm6.1} involves only interpolation inequalities (see \eqref{215} and \eqref{216}), we infer that \eqref{8.1} has a weak solution for all $1\leq\delta<\infty$. An application of Sobolev's inequality yields $H_0^1(\Omega)\subsetL^{2\delta+2}(\Omega)$, for all $1\leq\delta\leq 2$ and hence, in three dimensions, the definition of weak solution given in \eqref{8p3} makes sense for all $v\inH_0^1(\Omega)\capL^{2\delta+2}(\Omega)$, for $2<\delta<\infty$. For the condition given in \eqref{8pp5}, the uniqueness of weak solution follows verbatim as in the proof of Theorem \ref{thm6.2}, since we are only invoking an interpolation inequality (see \eqref{3.26}).
For $1\leq\delta\leq 2$, the condition given in \eqref{234} needs to be replaced by
\[
\nu+\frac{\beta\gamma}{\lambda_1}>\frac{\beta}{\lambda_1}2^{2\delta-1}(1+\gamma)^2(\delta+1)^2+\frac{2C\alpha}{\lambda_1^{\frac{2-\delta}{4(\delta+1)}}}\sqrt{\frac{\widetilde{K}}{\beta}},
\]
where $\widetilde{K}$ is defined in \eqref{8.14}. This change is needed since the estimate \eqref{2.31} should be replaced by
\begin{align*}
-\alpha\langle B(u)-B(v),w\rangle&\leq \alpha\|\nabla w\|_{0}\|w\|_{L^{2(\delta+1)}}\left(\|u\|_{L^{2(\delta+1)}}^{\delta}+\|v\|_{L^{2(\delta+1)}}^{\delta}\right)\nonumber\\
&\leq C\alpha\|\nabla w\|_{0}^{\frac{5\delta+2}{2(\delta+1)}}\|w\|_{0}^{\frac{2-\delta}{2(\delta+1)}}\left(\|u\|_{L^{2(\delta+1)}}^{\delta}+\|v\|_{L^{2(\delta+1)}}^{\delta}\right)\nonumber\\&\leq\frac{C\alpha}{\lambda_1^{\frac{2-\delta}{4(\delta+1)}}}\left(\|u\|_{L^{2(\delta+1)}}^{\delta}+\|v\|_{L^{2(\delta+1)}}^{\delta}\right)\|\nabla w\|_{0}^2\nonumber\\
&\leq \frac{2C\alpha}{\lambda_1^{\frac{2-\delta}{4(\delta+1)}}} \sqrt{\frac{\widetilde{K}}{\beta}}, \qquad \text{for $1\leq\delta\leq 2$},
\end{align*}
where we have applied Holder's, Gagliardo-Nirenberg's and Young's inequalities.
\end{remark}
\begin{theorem}[Regularity]
If $\Omega\subset\mathbb{R}^d,d=2,3,$ is either convex, or a domain with $C^2$-boundary and $f\in L^2(\Omega)$, then the weak solution of \eqref{8.1} belongs to $ H^2(\Omega)$.
\end{theorem}
\begin{proof}
Let us first assume that $f\inL^2(\Omega)$. Proceeding to multiply \eqref{8p5} by $u_m^{2\delta}\xi_m^k$ and then adding from $k=1,\ldots,m$, we get
\begin{align*}
&\nu(2\delta+1)\|u_m^{\delta}\nabla u_m\|_{0}^2+\beta\gamma\|u_m\|_{L^{2\delta+2}}^{2\delta+2}+\beta\|u_m\|_{L^{4\delta+2}}^{4\delta+2}\nonumber\\&=\beta(1+\gamma)(u_m^{\delta+1},|u_m|^{2\delta}u_m)+(f,|u_m|^{2\delta}u_m)\nonumber\\
&\leq\frac{\beta}{2}\|u_m\|_{L^{4\delta+2}}^{4\delta+2}+\beta(1+\gamma)^2\|u_m\|_{L^{2\delta+2}}^{2\delta+2}+\frac{1}{\beta}\|f\|_{0}^2,
\end{align*}
where we used the Cauchy-Schawrz and Young inequalities. Thus, using \eqref{2p13}, it is immediate to see that
\begin{align}\label{242}
&\nu(2\delta+1)\|u_m^{\delta}\nabla u_m\|_{0}^2+\frac{\beta}{2}\|u_m\|_{L^{4\delta+2}}^{4\delta+2}\leq(1+\gamma+\gamma^2)\widetilde{K}+\frac{1}{\beta}\|f\|_{0}^2.
\end{align} Multiplying \eqref{8p5} by $\lambda_k\xi_m^k$ and then adding from $k=1,\ldots,m$, we can assert that
\begin{align}\label{8p14}
\nu\|Au_m \|_{0}^2&=-\alpha (B(u_m ),Au_m )+\beta(C(u_m),Au_m)+(f,Au_m).
\end{align}
Let us take the term $-\alpha (B(u_m ),Au_m )$ from \eqref{8p14} and estimate it using \eqref{242}. Then, H\"older's and Young's inequalities give the following bound
\begin{align}\label{243}
\alpha |(B(u_m ),Au_m )|&\leq\alpha \|B(u_m)\|_{0}\|A u_m\|_{0}\leq\alpha\|u_m^{\delta}\nabla u_m\|_{0}\|Au_m\|_{0}\nonumber\\
&\leq\frac{\nu}{4}\|Au_m\|_{0}^2+\frac{\alpha^2}{\nu}\|u_m^{\delta}\nabla u_m\|_{0}^2.
\end{align}
Integrating by parts and applying H\"older's and Young's inequalities, we find
\begin{align*}
\beta(C&(u_m),Au_m)\nonumber\\
&= -\beta\gamma\|\nabla u_m\|_{0}^2-\beta(2\delta+1)\|u_m^{\delta}\nabla u_m\|_{0}^2+\beta(1+\gamma)(\delta+1)(u_m^{\delta}\nabla u_m,\nabla u_m)\nonumber\\
&\leq -\beta\gamma\|\nabla u_m\|_{0}^2-\frac{\beta(2\delta+1)}{2}\|u_m^{\delta}\nabla u_m\|_{0}^2+\frac{\beta(1+\gamma)^2(\delta+1)^2}{2(2\delta+1)}\|\nabla u_m\|_{0}^2.
\end{align*}
Then we use the Cauchy-Schwarz and Young's inequalities to estimate $|(f,Au_m)|$ as
\begin{align}\label{2.22}
|(f,Au_m)|\leq\|f\|_{0}\|Au_m\|_{0}\leq\frac{\nu}{4}\|Au_m\|_{0}^2+\frac{1}{\nu}\|f\|_{0}^2.
\end{align}
Combining \eqref{243}-\eqref{2.22} and substituting the outcome back in \eqref{8p14}, we obtain
\begin{align*}
& \frac{\nu}{2}\|Au_m \|_{0}^2+\frac{\beta(2\delta+1)}{2}\|u_m^{\delta}\nabla u_m\|_{0}^2\nonumber\\&\leq \frac{\alpha^2}{\nu}\|u_m^{\delta}\nabla u_m\|_{0}^2+\frac{\beta((1+\gamma^2)(\delta+1)^2+2\gamma\delta^2)}{2(2\delta+1)}\|\nabla u_m\|_{0}^2+\frac{1}{\nu}\|f\|_{0}^2.
\end{align*}
From the estimates \eqref{2p13} and \eqref{242}, we infer that $u_m\in D(A)$. Once again invoking the Banach-Alaoglu theorem, we can extract a subsequence $\{u_{m_k}\}$ of $\{u_m\}$ such that
\begin{equation*}
\left\{\begin{aligned}
u_{m_k}&\xrightarrow{w}u\ \text{ in }\ L^{4\delta+2}(\Omega)\ \text{ as }\ k\to\infty,\\
u_{m_k}&\xrightarrow{w}u\ \text{ in }\ D(A)\ \text{ as }\ k\to\infty,
\end{aligned}
\right.\end{equation*}
since the weak limit is unique. Using the compact embedding of $H^2(\Omega)\subset H^1(\Omega)$, along a subsequence, we further have
$$u_{m_{k_j}}\to u\ \text{ in }\ H^1(\Omega), \ \text{ as }\ j\to\infty. $$
Proceeding similarly as in the proof of Theorem \ref{thm6.1}, we obtain that $u\in D(A)$ satisfies $$\nu Au +\alpha B(u )-\beta C(u )=f,\ \text{ in }\ L^2(\Omega), $$ and
$$\|Au\|_{0}^2+\|u^{\delta}\nabla u\|_0^2+\|u\|_{L^{4\delta+2}}^{4\delta+2}\leq C(\|f\|_0,\nu,\alpha,\beta,\gamma,\delta).$$
But, we know that
\begin{align*}
& \|f-\alpha B(u )+\beta C(u )\|_0\\&\leq \|f\|_0+\alpha\|u^{\delta}\nabla u\|_0+\beta\gamma\|u\|_0+\beta(1+\gamma)\|u\|_{L^{2\delta+2}}^{\delta+1}+\beta\|u\|_{L^{4\delta+2}}^{2\delta+1}<\infty,
\end{align*}
and hence an application of \cite[Th. 9.25]{brezis_book11} (for a domain with $C^2$ boundary) or \cite[Th. 3.2.1.2]{grisvard_book85} (for convex domains) yields $u\in H^2(\Omega)$.
\end{proof}
\section{Numerical schemes and their a priori error estimates}\label{sec3}\setcounter{equation}{0}
Let the domain $\Omega$ be partitioned into a mesh (consisting of shape-regular triangular or rectangular
cells $K$) denoted by $\mathcal{T}_h$. We use the symbols $\mathcal{E}_h$, $\mathcal{E}^i_h$ and $\mathcal{E}^{\partial}_h$ to denote the set of edges, interior edges and
boundary edges of the mesh, respectively. For a given $\mathcal{T}_h$,
the notations $C^{0}(\mathcal{T}_h)$ and $H^s(\mathcal{T}_h)$ indicate broken spaces associated with
continuous and differentiable function spaces, respectively.
\subsection{Conforming method}
Let $V_h$ be a finite dimensional subspace of $H_0^1(\Omega)$ associated with the mesh parameter $h$. Numerical solutions are sought in the family $\{V_h\}\subsetH_0^1(\Omega),$ (where one additionally assumes that $h$ is sufficiently small) satisfying the following approximation
property (see \cite{VTh})
\begin{align*}
\inf_{\chi\in V_h}\left\{\|u-\chi\|_{0}^2+h\|\nabla(u-\chi)\|_{0}^2\right\}\leq Ch^k\|u\|_{k},
\end{align*}
for all $u\inH^r(\Omega)\capH_0^1(\Omega)$, $ 1\leq k\leq r$, where $r$ is the order of accuracy of the family $\{V_h \}$.
The CFEM for \eqref{8p2} reads: find $u_h\in V_h$ such that
\begin{equation}\label{7p1}
\nu a(u_h,\chi)+\alpha b(u_h,u_h,\chi)=\beta\langle C(u_h),\chi\rangle+\langle f,\chi\rangle, \qquad \forall \chi \in V_h.
\end{equation}
\begin{theorem}[Existence of a discrete solution]\label{excgfem}
Equation (\ref{7p1}) admits at least one solution $u_h\in V_h$.
\end{theorem}
\begin{proof}
It follows as a direct consequence of Theorem \ref{thm6.1}.
\end{proof}
Let $R^h$ be the elliptic or Ritz projection onto $V_h$ (see \cite{VTh}), defined by
\begin{align*}
(\nabla R^hv,\nabla\chi)=(\nabla v,\nabla\chi), \text{ for all }\ \chi\in V_h \ \text{ for }\ v\inH_0^1(\Omega).
\end{align*}
By setting $\chi=R^hv$ above, we readily obtain that the Ritz projection is stable, that is, $\|\nabla R^hv\|_{0}\leq\|\nabla v\|_{0}$, for all $v\inH_0^1(\Omega)$. Moreover, using \cite[Lem. 1.1]{VTh}, we have
\begin{align}\label{7a1}
\|R^hv-v\|_{0}+h\|\nabla(R^hv-v)\|_{0}\leq Ch^s\|v\|_{s},
\end{align}
for all $ v\inH^s(\Omega)\capH_0^1(\Omega)$, $1\leq s\leq r$.
\begin{theorem}[Energy estimate]\label{thm7.1}
Let $V_h$ be a finite dimensional subspace of $H_0^1(\Omega)$. Assume that \eqref{234} holds true and that $u\in D(A)=H_0^1(\Omega)\cap H^2(\Omega)$ satisfies \eqref{8p2}. Then the error incurred by the Galerkin approximation satisfies
\[
\|u_h-u\|_{1}\leq Ch,
\]
where $C$ is a constant possibly depending on $\nu,\alpha,\beta,\gamma,\delta$, $\|f\|_{0}$, but independent of $h$.
\end{theorem}
\begin{proof}
Using triangle inequality we can write
\begin{equation}\label{7p7}
\|u_h-u\|_{1}\le \|u_h-W\|_{1}+\|W-u\|_{1},
\end{equation}
where $W\in V_h$. We need to estimate $\|u_h-W\|_{1}$. First we note that from \eqref{7a1}, the second term in the RHS of \eqref{7p7} satisfies
\[
\|W-u\|_{1}\le Ch.
\]
Next, and using \eqref{8p3} and \eqref{7p1}, we can assert that $u^h-u$ satisfies
\begin{align}\label{7p2}
\nu a(u_h-u,\chi)= -\alpha[b(u_h,u_h,\chi)-b(u,u,\chi)]+\beta[\langle C(u_h),\chi\rangle -\langle C(u),\chi\rangle],
\end{align}
for all $\chi\in V_h$. Let us choose $\chi=u_h-W\in V_h$ in \eqref{7p2}, to eventually obtain
\begin{align}\label{7p3}
\nu a(u_h-u,u_h-W)&= -\alpha[b(u_h,u_h,u_h-W)-b(u,u,u_h-W)]\nonumber\\
&\quad+\beta[\langle C(u_h),u_h-W\rangle -\langle C(u),u_h-W\rangle ].
\end{align}
On the other hand, we can write $u_h-u$ as $u_h-W+W-u$ in \eqref{7p3} to find
\begin{align*}
\nu\|\nabla(u_h-W)\|_{0}^2&=-\nu(\nabla(W-u),\nabla \chi)-\alpha[b(u_h,u_h,\chi)-b(W,W,\chi)]\\
&\quad-\alpha[b(W,W,\chi)-b(u,u,\chi)]+\beta[\langle C(u_h),\chi\rangle-\langle C(W),\chi\rangle ]\\
&\quad+\beta[\langle C(W),\chi\rangle -\langle C(u),\chi\rangle ].
\end{align*}
Thus, following \eqref{327} and \eqref{6.49}, we can establish the bound
\begin{align}\label{7p5}
\frac{\nu}{2}\|\nabla \chi\|_{0}^2+\left(\frac{\beta}{4}-\frac{4^{\delta}\alpha^2}{4\nu}\right)&\|{u_h}^{\delta}\chi\|_{0}^2+\left(\frac{\beta}{4}-\frac{4^{\delta}\alpha^2}{4\nu}\right)\|W^{\delta}\chi\|_{0}^2\nonumber\\
+(\beta\gamma-C(\beta,\alpha,\delta))\|\chi\|_{0}^2&
\leq \nu(\nabla(u-W),\nabla \chi)-\alpha\sum_{i=1}^2\left({W}^{\delta}\frac{\partial W}{\partial x_i}-u^{\delta}\frac{\partial u}{\partial x_i},\chi\right)\nonumber\\
&\quad+\beta(W(1-{W}^{\delta})({W}^{\delta}-\gamma)-u(1-u^{\delta})(u^{\delta}-\gamma),\chi),
\end{align}
where we have introduced the constant $C(\beta,\alpha,\delta)= \beta 2^{2\delta-1}(1+\gamma)^2(\delta+1)^2$.
Using an integration by parts, Taylor's formula, H\"older's and Young's inequalities, we can rewrite the first term on
the RHS of \eqref{7p5} as
\begin{align}\label{7p6}
-\frac{\alpha}{\delta+1}&\sum_{i=1}^d\left(\frac{\partial}{\partial x_i}({W}^{\delta+1}-u^{\delta+1}),\chi\right)
=\frac{\alpha}{\delta+1}\sum_{i=1}^d({W}^{\delta+1}-u^{\delta+1},\frac{\partial}{\partial x_i}\chi)\nonumber\\
&=\alpha\sum_{i=1}^d\left((\theta W+(1-\theta)u)^{\delta}(W-u),\frac{\partial}{\partial x_i}\chi\right)\nonumber\\
&\leq 2^{\delta-1}\alpha\left(\|{W}^{\delta}(W-u)\|_{0}+\|{u}^{\delta}(W-u)\|_{0}\right)\|\nabla \chi\|_{0}\nonumber\\
&\leq 2^{\delta-1}\alpha\left(\|{W}^{2\delta}\|_{0}^{1/2}+\|{u}^{2\delta}\|_{0}^{1/2}\right)\|W-u\|_{L^4}\|\nabla \chi\|_{0}.
\end{align}
And we can also rewrite the second term on the RHS of \eqref{7p5} as
\[
\beta(1+\gamma)({W}^{\delta+1}-u^{\delta+1},\chi)-2\beta\gamma(W-u,\chi)
-2\beta({W}^{2\delta+1}-u^{2\delta+1},\chi):=\sum_{i=1}^3J_i,
\]
where
\begin{gather*}
J_1=\beta(1+\gamma)({W}^{\delta+1}-u^{\delta+1},\chi),\qquad
J_2 =-2\beta\gamma(W-u,\chi),\\
J_3=-2\beta({W}^{2\delta+1}-u^{2\delta+1},\chi).
\end{gather*}
We estimate $J_1$ using Taylor's formula, H\"older's and Young's inequalities as
\begin{align*}
J_1&=
\beta(1+\gamma)(\delta+1)((\theta W+(1-\theta)u)^{\delta}(W-u),\chi)\\
&\leq 2^{\delta-1}\beta(1+\gamma)(\delta+1)\left(\|{W}^{\delta}(W-u)\|_{0}+\|{u}^{\delta}(W-u)\|_{0}\right)\|\chi\|_{0}\\
&\leq 2^{\delta-1}\beta(1+\gamma)(\delta+1)\left(\|{W}^{2\delta}\|_{0}^{1/2}+\|{u}^{2\delta}\|_{0}^{1/2}\right)\|W-u\|_{L^4}\|\chi\|_{0}.
\end{align*}
In turn, using Cauchy-Schwarz and Young's inequalities, an estimate for $J_2$ reads
\[
J_2 \leq 2\beta\gamma\|W-u\|_{0}\|\chi\|_{0},
\]
while a bound for $J_3$ results from applying
Taylor's formula together with H\"older's and Young's inequalities
\begin{align}\label{7p11}
J_3&=-(2\delta+1)\beta((\theta W+(1-\theta)u)^{2\delta}(W-u),\chi)\nonumber\\
&\leq 2^{2\delta-1}(2\delta+1)\beta \left(\|{W}^{\delta}(W-u)\|_{0}\|{W}^{\delta}\chi\|_{0}+\|{u}^{\delta}(W-u)\|_{0}\|{u}^{\delta}\chi\|_{0}\right)\nonumber\\
&\leq 2^{2\delta-1}(2\delta+1)\beta \left(\|{W}^{2\delta}\|_{0}+\|{u}^{2\delta}\|_{0}\right)\|W-u\|_{L^4}\|\chi\|_{L^4}.
\end{align}
Combining \eqref{7p6}-\eqref{7p11}, substituting the result back into \eqref{7p5}, and then using
\eqref{7a1} and \eqref{7p7}, implies the desired result.
\end{proof}
\subsection{Non-conforming finite element method}
Let $\mathbb{P}_1$ denote the space of polynomials which have degree at most $1$, and let us recall the
definition of the Crouzeix-Raviart (CR) non-conforming finite element space
\begin{equation}\label{CRFEM11}
{V}_{h}^{CR} =\left\{v\in L^2(\Omega) :\ \text{ for all }\ K\in \mathcal{T} \;v_{|_K}\in\mathbb{P}_1 \;\mbox{and}\; \int_E [|v|]=0\quad E\in\mathcal{E}\right\}.
\end{equation}
It is useful to introduce the piecewise gradient operator $\nabla_h: H^1(\mathcal{T}_h)\rightarrow L^2(\Omega;\mathbb{R}^2)$ with $(\nabla_h v)|_K = \nabla v|_K,$ for all $K\in \mathcal{T}_h$.
The discrete weak formulation of (\ref{8.1}) in this context reads: find $u^{CR}_h\in V_h^{CR}$ such that
\begin{align}\label{ncweakform}
A_{NC}(u_h^{CR},\chi)=(f,\chi), \quad \ \text{ for all }\ \chi\in V_h^{CR},
\end{align}
with
\begin{gather*}
A_{NC}(v,v)=\nu a_{NC}(v,v)+\alpha b_{NC}(v;v,v)-\beta(C(v),v),\\
a_{NC}(v,v) = (\nabla_h v, \nabla_h v), \quad b_{NC}(v;v,v)= ((v^{\delta},v^{\delta})^T\cdot\nabla_h v,v),
\end{gather*}
and we define the associated discrete energy norm $\vertiii{v}_{NC}:=\sqrt{a_{NC}(v,v)}$.
\begin{lemma}\label{crnclem11}
For any $v\in V_h^{CR}$, we have
\begin{align}\label{317}
A_{NC}(v,v)\ge \bar{C} \vertiii{v}_{NC}^2,
\end{align}
provided $\nu > \max\{{\beta}(1+\gamma^2)C_{\Omega}^{NC},\frac{2\alpha^2}{\beta}\}$.
\end{lemma}
\begin{proof}
Owing to Young's and Poincar\'{e}-Friedrichs's inequalities, it readily follows that
\begin{align*}
A_{NC}(v,v)&=\nu\|\nabla_h v\|^2_{0,\mathcal{T}_h}+\beta\gamma\|v\|_{0}^2+\beta\|v\|_{L^{2\delta+2}}^{2\delta+2}-\beta(1+\gamma)(v^{\delta+1},v)-b_{NC}(v;v,v)\nonumber\\
&\geq \nu\|\nabla_h v\|_{0,\mathcal{T}_h}^2+\beta\gamma\|v\|_{0}^2+\beta\|v\|_{L^{2\delta+2}}^{2\delta+2}-\beta(1+\gamma)\|v\|_{L^{\delta+1}}^{\delta+1}\|v\|_{0}\nonumber\\
&\qquad-\alpha\|v\|_{L^{\delta+1}}^{\delta+1}\|\nabla_h v\|_{0,\mathcal{T}_h}\nonumber\\
&\geq {\nu}\|\nabla_h v\|_{0,\mathcal{T}_h}^2+\beta\gamma\|v\|_{0}^2+\frac{\beta}{4}\|v\|_{L^{2\delta+2}}^{2\delta+2}-\frac{\beta}{2}(1+\gamma)^2\|v\|_{0}^2-\frac{\alpha^2}{\beta}\|\nabla_hv\|_{0,\mathcal{T}_h}^2\nonumber\\
&\geq\nu\|\nabla_h v\|_{0,\mathcal{T}_h}^2-\frac{\beta}{2}(1+\gamma^2)\|v\|_{0}^2-\frac{\alpha^2}{\beta}\|\nabla_hv\|_{0,\mathcal{T}_h}^2\nonumber\\
&\geq\left(\frac{\nu}{2}-\frac{\beta}{2}(1+\gamma^2)C_{\Omega}^{NC}+\frac{\nu}{2}-\frac{\alpha^2}{\beta}\right)\|\nabla_h v\|_{0,\mathcal{T}_h}^2,
\end{align*}
and the estimate \eqref{317} follows.
\end{proof}
\begin{theorem}[Existence of a discrete solution]\label{excrfem}
Let $\|u_h^{CR}\|_0=k_{CR}$ and
\begin{align*}
k_{CR}> \frac{(C_{\Omega}^{CR})}{\nu\sqrt{\nu+\beta\gamma C_{\Omega}^{CR}-\beta(1+\gamma)^2C_{\Omega}^{CR}-\frac{2\alpha^2}{\beta}}}\|f\|_{0},
\end{align*}
provided $\nu +\beta \gamma C_{\Omega}^{CR}> \beta(1+\gamma)^2C_{\Omega}^{CR}+\frac{2\alpha^2}{\beta}$. Then, problem (\ref{ncweakform}) admits at least one solution $u_h^{NC}\in V_h^{NC}$.
\end{theorem}
\begin{proof}
We introduce the Crouzeix-Raviart operator $P_{CR}:V_h^{CR}\rightarrow V_h^{CR}$ as
\[
(P_{CR}(u_{h}^{CR}),v)=A_{NC}(u_{h}^{CR},v)-(f,v),
\]
which is well defined and continuous on $V_h^{CR}$. Choosing $v=u_h^{CR}$ and using Lemma \ref{crnclem11}, we have
\begin{align}\label{eqcrnc12}
(P_{CR}(u_h^{CR}),&u_h^{CR})\nonumber\\
&\ge \nu\|\nabla_h v\|_{0,\mathcal{T}_h}^2-\frac{\beta}{2}(1+\gamma^2)\|v\|_{0}^2-\frac{\alpha^2}{\beta}\|\nabla_hv\|_{0,\mathcal{T}_h}^2+\beta\gamma\|v\|_{0}^2\nonumber\\
&\quad-\frac{C_{\Omega}^{CR}}{2\nu}\|f\|_0^2-\frac{\nu}{2C_{\Omega}^{CR}}\|u^{CR}_h\|_0^2,\nonumber\\
&\ge\frac{1}{C_{\Omega}^{CR}} \left(\frac{\nu}{2}-\frac{\beta}{2}(1+\gamma^2)C_{\Omega}^{CR}-\frac{\alpha^2}{\beta}+\beta\gamma C_{\Omega}^{CR}\right)\| u_{h}^{CR}\|_{0}^2-\frac{C_{\Omega}^{CR}}{2\nu}\|f\|_0^2.
\end{align}
Let $\|u_h^{CR}\|_0=k_{CR}$ and
\[
k_{CR}> \frac{(C_{\Omega}^{CR})}{\nu\sqrt{\nu+\beta\gamma C_{\Omega}^{CR}-\beta(1+\gamma)^2C_{\Omega}^{CR}-\frac{2\alpha^2}{\beta}}}\|f\|_{0},
\]
provided $\nu +\beta \gamma C_{\Omega}^{CR}> \beta(1+\gamma)^2C_{\Omega}^{CR}+\frac{2\alpha^2}{\beta}$. Then the RHS in (\ref{eqcrnc12}) is non-negative. Finally,
Brouwer's fixed-point theorem implies that $P_{CR}(u_h^{CR})=0$.
\end{proof}
Next we denote by $I_h$ the usual finite element interpolation \cite{john1998}. Then the following estimates hold
\begin{align}\label{ncapprox11}
|v-I_hv|_{m,k}&\le Ch^{2-m}_K\|v\|_{2,K}\quad v\in H^2(K),\\
\|v-(I_hv)\|_{0,E}&\le Ch^{3/2}\|v\|_{2,K}\quad v\in H^2(K)\quad E\in\mathcal{E}(\mathcal{T}_h). \label{ncapprox12}
\end{align}
Regarding the edge projection $P_E:L^2(E)\rightarrow P_0(E)$, where $P_0(E)$ is a constant on $E$, we have
\begin{align}\label{P0projection}
\|v-P_E v\|_{0,E}\le Ch^{1/2}_K|v|_{1,K},\ \text{ for all }\ v\in H^1(K),\ E\in\mathcal{E}(\mathcal{T}_h).
\end{align}
\begin{lemma}\label{crnclem111}
There holds:
\begin{align*}
\alpha[b_{NC}(v_1,v_1,w)-b_{NC}(v_2,v_2,w)]&\le \frac{\nu}{2}\|\nabla_h w\|_{0,\mathcal{T}_h}^2+\frac{2^{2\delta}C_{\star}\alpha^2}{4\nu}(\|v^{\delta}_1 w\|_{0}^2+\|v^{\delta}_2 w\|_{0}^2),\\
A_{NC}(v_1,w)-A_{NC}(v_2,w)&\ge \frac{\nu}{2}\|\nabla_h w\|_{0,\mathcal{T}_h}^2+(\beta\gamma-C(\beta,\alpha,\delta))\|w\|_{0}^2\nonumber\\
&+\left(\frac{\beta}{4}-\frac{2^{2\delta}C_{\star}\alpha^2}{4\nu}\right)(\|{v}^{\delta}_1w\|_{0}^2+\|v^{\delta}_2w\|_{0}^2),
\end{align*}
where $v_1,v_2\in V_{h}^{NC}$, $w=v_1-v_2$ and $C_{\star}$ is a postive constant.
\end{lemma}
\begin{proof}
To prove the first estimate, we use the definition of $b_{NC}(\cdot, \cdot)$. Then
\begin{align*}
\alpha[b_{NC}(v_1,v_1,w)-b_{NC}(v_2,v_2,w)]&=\alpha \sum_{K\in\mathcal{T}_h}\sum_{i=1}^{d}\int_{K}\left(v_1^{\delta}\frac{\partial v_1}{\partial x_i} -v_2^{\delta}\frac{\partial v_2}{\partial x_i}\right)w dx\\
&=\frac{\alpha}{\delta+1}\sum_{K\in\mathcal{T}_h}\sum_{i=1}^{d}\int_{K}\left(\frac{\partial (v_1^{\delta+1}- v_2^{\delta+1})}{\partial x_i} \right)w dx.
\end{align*}
Using Cauchy-Schwarz and inverse inequalities, Taylor's formula,
H\"{o}der's and Young's inequalities, implies the first stated result.
To prove the second inequality, we write
\begin{align*}
A_{NC}(v_1,w)-A_{NC}(v_2,w)&= \nu a_{NC}(v_1-v_2,w)+\alpha[b_{NC}(v_1,v_1,w)-b_{NC}(v_2,v_2,w)]\\
&\quad-\beta[(C(v_1),w)-(C(v_2),w)].
\end{align*}
Applying the first estimate and \eqref{327} leads to the second estimate.
\end{proof}
\begin{theorem}\label{ncthm11}
Let $V_h^{CR}$ be the non-conforming space defined in (\ref{CRFEM11}). Assume that \eqref{234} holds true and that $u\in D(A)=H_0^1(\Omega)\cap H^2(\Omega)$ satisfies \eqref{8p2}. Then the error incurred by the NCFEM approximation satisfies
\[
\vertiii{u^{CR}_h-u}_{NC}\leq Ch,
\]
where the constant $C$ is independent of $h$ and $C$ depends on $\nu,\alpha,\beta,\gamma,\delta$, $\|f\|_{0}$, etc.
\end{theorem}
\begin{proof} Similarly as before, we split the error and use triangle inequality to write
\[
\vertiii{u_h^{CR}-u}_{NC}\le \vertiii{ u_h^{NC}-W}_{NC}+\vertiii{ W-u}_{NC}.
\]
From \eqref{ncapprox11}, the following estimate is valid for the second term on the RHS
\[
\vertiii{ W-u}_{NC}\le Ch.
\]
Using (\ref{ncweakform}), we have
\[
A_{NC}(u_h^{CR},\chi)=(f,\chi), \ \text{ for all }\ \chi\in V_h^{CR}.
\]
If $u\in D(A)=H_0^1(\Omega)\cap H^2(\Omega)$ satisfies \eqref{8p2}, then it readily
follows that
\[
A_{NC}(u,\chi)=(f,\chi)+\sum_{K\in\mathcal{T}}\int_{K}\nu\frac{\partial u}{\partial n_K} \chi, \ \text{ for all }\ \chi\in V_h^{CR}.
\]
We can then use Lemma \eqref{crnclem111}, which leads to
\begin{align*}
&\frac{\nu}{2}\|\nabla_h \chi\|_{0,\mathcal{T}_h}^2+(\beta\gamma-C(\beta,\alpha,\delta))\|\chi\|_{0}^2+\left(\frac{\beta}{4}-\frac{2^{2\delta}C_{\star}\alpha^2}{4\nu}\right)(\|{u}^{CR}_h\chi\|_{0}^2+\|W^{\delta}\chi\|_{0}^2)\\
&\leq A_{NC}(u,\chi)-A_{NC}(W,\chi)-\sum_{K\in\mathcal{T}}\int_{K}\nu\frac{\partial u}{\partial n_K} \chi.
\end{align*}
To estimate the consistency error, it suffices to exploit the CR approximation
\[
\sum_{K\in\mathcal{T}}\int_{\partial K}\nu\frac{\partial u}{\partial n_K} \chi=-\sum_{E\in\mathcal{E}}\int_{E}\nu\frac{\partial u}{\partial n_E} [\chi]
=-\sum_{E\in\mathcal{E}}\int_{E}\nu\left(\frac{\partial u}{\partial n_E} -P\left(\frac{\partial u}{\partial n_E}\right)\right)[\chi].
\]
Consequently, we can invoke estimate (\ref{P0projection}), which yields
\[
\left|\sum_{K\in\mathcal{T}}\int_{\partial K}\nu\frac{\partial u}{\partial n_K} \chi\right| \le C \left(\sum_{K\in\mathcal{T}}\nu h_K^2\|u\|_{2,K}^2\right)^{1/2}\vertiii{ \chi}_{NC},
\]
and the remainder of the proof follow similarly to that of Theorem \ref{thm7.1}.
\end{proof}
\subsection{Discontinuous Galerkin method}
In addition to the mesh notation used so far, we also require the
following preliminaries.
Let $E=K_+\cap K_-\in\mathcal{E}^i_h$ be the common edge that is shared by the
two mesh cells $K_\pm$. We use the symbol $w_{\pm}$ to denote
the traces of functions $w\in C^0(\mathcal{T}_h)$ on $E$ from $K_\pm$,
respectively. In addition, we denote
the sum (which in turn translates into the jump operator) over an edge as
\begin{align*}
[\![w]\!]=w_+ + w_-,
\end{align*}
and if ${w}\in C^1(\mathcal{T}_h)$ we also define
\begin{align*}
[\![\partial{w}/\partial\bm{n}]\!]=\nabla({w}_+ -{w}_-)\bm{n}_+,
\quad\text{and}\quad[\![{w}\otimes{n}]\!]=({w}_+ - {w}_-)\otimes\bm{n}_+,
\end{align*}
where $\bm{n}_\pm$ denote the unit outward normal vectors to $K_\pm$, respectively. In case of
boundary edges $E=K_+\cap\partial\Omega$, we take $[\![{v}]\!]={w}_+$. The exterior trace of $u$ taken over the edge under consideration is denoted by $u^e$ and we chose $u^e =0$ for boundary edges.
We recall the definition of
the local gradient
$\nabla_h$
satisfying $(\nabla_h{w})|_K = \nabla({w}|_K)$ on each $K\in\mathcal{T}_h$.
We will use the discrete subspace of $L^2(\Omega)$
\begin{align}\label{dgsubspace1}
{V}_h^{DG}=\{{v}\in L^2(\Omega): \text{ for all }\ K\in \mathcal{T}_h : {v}|_K \in \mathcal{P}_1(K)\}.
\end{align}
where $\mathcal{P}_1(K)$ is the space of polynomials on $K$ having partial degree $1$.
The discrete weak formulation of (\ref{8.1}) reads now: find $u^{DG}_h\in V_h^{DG}$ such that
\begin{equation}\label{dgweakform}
A_{DG}(u_h^{DG},\chi)=(f,\chi), \ \text{ for all }\ \chi\in V_h^{DG},
\end{equation}
where, for ${u},{v}\in {V}^{DG}_h$, the bilinear form
\begin{equation}\label{ADG}
A_{DG}(v,v)=\nu a_{DG}(v,v)+\alpha b_{DG}(\bm{v},v,v)-\beta(C(v),v),\end{equation}
is defined with the following contributions
\begin{gather*}
a_{DG}({u},{v})=(\nabla_h {u}, \nabla_h {v})+a^{i}_{h}({u},{v})+a^{\partial}_{h}({u},{v}),\\
a^{i}_{h}({u},{v})=a^{i}_p({u},{v})-a^{i}_c({u},{v})-a^{i}_c({v},{u}),\quad
a^{\partial}_{h}({u},{v})=a^{\partial}_p({u},{v})-a^{\partial}_c({u},{v})-a^{\partial}_c({v},{u}),\\
a^{i}_c({u},{v})=\frac{1}{2}\sum_{E\in\mathcal{E}^i_h}\int_E [\![\nabla_h {u}]\!]\!\cdot\![\![{v}\otimes \bm{n}]\!]d{s},\quad
a^{i}_p({u},{v})=\sum_{E\in\mathcal{E}^i_h}\int_E\gamma_h [\![{u}\otimes \bm{n}]\!]\!\cdot\![\![{v}\otimes \bm{n}]\!]d{s},\\
a^{\partial}_c({u},{v})= \sum_{E\in\mathcal{E}^{\partial}_h}\int_E \nabla {u}\!\cdot\!({v}\otimes \bm{n})d{s},
\quad a^{\partial}_p({u},{v})=2\sum_{E\in\mathcal{E}^{\partial}_h}\int_E\gamma_h ({u}\otimes \bm{n})\!\cdot\!({v}\otimes \bm{n})d{s},\\
b_{DG}(\bm{w};u,v)= \!\sum_{K\in \mathcal{T}_h} \!\!\int_{K} \!\!\bm{w}\cdot \nabla u {v} dx
+\!\!\!\sum_{K\in \mathcal{T}_h} \frac{1}{2}\int_{\partial K} \left[\bm{w}\cdot\bm{n}_K (u^e \! -\! u)\!-|\bm{w}\cdot \bm{n}_K|(u^e\!-\!u)\right]{v} ds,
\end{gather*}
with $\bm{w}=(w,w)^T$ and $\gamma_h=\frac{\gamma}{h_E}$, where $h_E$ is the length of the edge $E$ and $\gamma$ is a penalty parameter chosen sufficiently large to guarantee the stability of the formulation (see, e.g., \cite{ADN}).
It is also convenient to rewrite $b_{DG}(\cdot;\cdot,\cdot)$, after integration by parts, as follows
\begin{align*}
b_{DG}(\bm{w};u,v)&= \sum_{K\in \mathcal{T}_h} \int_{K} (-u \bm{w}\cdot \nabla {v} -\nabla\cdot \bm{w} u {v})dx\\
&\quad+\sum_{K\in \mathcal{T}_h} \int_{\partial K} \left[\frac{1}{2}\bm{w}\cdot\bm{n}_K [\![u]\!]-\frac{1}{2}|\bm{w}\cdot \bm{n}_K|(u^e-u)\right]{v} ds.
\end{align*}
For the subsequent error analysis, we adopt the following discrete norm
\[
\vertiii{ v}^2:= \sum_{K\in\mathcal{T}_h}\|\nabla_h v\|_{0,K}^2+\sum_{E\in\mathcal{E}(\mathcal{T}_h)}\|[\![v]\!]\|_{0,E}^2.\]
\begin{lemma}\label{dglem11}
Coercivity of $a_{DG}$ and continuity of $b_{DG}$ hold in the following sense
\[
a_{DG}(v,v) \ge \alpha_a\vertiii{ v}^2, \quad
\alpha b_{DG}(\bm{v};v,v)\le \frac{\beta}{4}\|v\|_{L^{2\delta+2}}^{2\delta+2}+\frac{2\alpha^2}{\beta}\vertiii{v}^2, \qquad \forall
v\in V_{h}^{DG}.\]
\end{lemma}
\begin{proof}
The first estimate follows from \cite{ADN}.
Using Cauchy-Schwarz, inverse trace and Young's inequalities in $b_{DG}$, implies the second stated result.
\end{proof}
\begin{lemma}
For any $v\in V_h^{DG}$, the bilinear form $A_{DG}$ defined in \eqref{ADG} satisfies
\[
A_{DG}(v,v)\ge \bar{C} \vertiii{v}^2.
\]
\end{lemma}
\begin{proof}Owing to Young's inequality and Lemma \ref{dglem11}, we have
\begin{align*}
A_{DG}(v,v)&\ge \alpha_a\nu \vertiii{ v} ^2+\beta\gamma\|v\|_{0}^2+\beta\|v\|_{L^{2\delta+2}}^{2\delta+2}-\beta(1+\gamma)(v^{\delta+1},v)-\alpha b_{DG}(\bm{v};v,v)\nonumber\\
&\geq \alpha_a{\nu}\vertiii{ v} ^2+\beta\gamma\|v\|_{0}^2+\frac{\beta}{4}\|v\|_{L^{2\delta+2}}^{2\delta+2}-\frac{\beta}{2}(1+\gamma)^2\|v\|_{0}^2-\frac{2\alpha^2}{\beta}\vertiii{ v}^2\nonumber\\
&\geq \alpha_a\nu\vertiii{ v}^2-\frac{\beta}{2}(1+\gamma^2)\|v\|_{0}^2-\frac{2\alpha^2}{\beta}\vertiii{ v}^2\nonumber\\
&\geq\left(\frac{\alpha_a\nu}{2}-\frac{\beta}{2}(1+\gamma^2)C_{\Omega}+\frac{\alpha_a\nu}{2}-\frac{2\alpha^2}{\beta}\right)\vertiii{ v}^2.
\end{align*}
\end{proof}
\begin{theorem}[Existence of a discrete solution]
Let $\|u^{DG}_h\|_0=k_{DG}$ and
\begin{align*}
k_{DG}> \frac{(C_{\Omega}^{DG})}{\nu\sqrt{\nu+\beta\gamma C_{\Omega}^{DG}-\beta(1+\gamma)^2C_{\Omega}^{DG}-\frac{2\alpha^2}{\beta}}}\|f\|_{0},
\end{align*}
provided $\nu +\beta \gamma C_{\Omega}^{DG}> \beta(1+\gamma)^2C_{\Omega}^{DG}+\frac{2\alpha^2}{\beta}$. Then equation (\ref{dgweakform}) admits at least one solution $u_h^{DG}\in V_h^{DG}$.
\end{theorem}
\begin{proof}
Proceeding as before, we introduce the map
$P_{DG}:V_h^{DG}\rightarrow V_h^{DG}$ with
\[
(P_{DG}(u_{h}^{DG}),v)=A_{DG}(u_{h}^{DG},v)-(f,v),
\]
which is well-defined and continuous.
Choosing $v=u_h^{DG}$ in Lemma \ref{dglem11} yields
\begin{align}\label{eqdg12}
(P_{DG}(u_h^{DG}),&u_h^{DG})\nonumber\\
&\ge \alpha_a\nu \vertiii{ u_h^{DG}}^2-\frac{\beta}{2}(1+\gamma^2)\|u^{DG}_h\|_{0}^2-\frac{2\alpha^2}{\beta}\vertiii{ u_h^{DG}} ^2+\beta\gamma\|u^{DG}_h\|_{0}^2\nonumber\\
&\quad-\frac{C_{\Omega}^{DG}}{2\nu}\|f\|_0^2-\frac{\nu}{2C_{\Omega}^{DG}}\|u^{DG}_h\|_0^2,\nonumber\\
&\ge\frac{\alpha_a}{C_{\Omega}^{DG}} \left(\frac{\nu}{2}-\frac{\beta(1+\gamma^2)C_{\Omega}^{DG}}{2\alpha_a}-\frac{\alpha^2}{\beta\alpha_a}+\frac{\beta\gamma C_{\Omega}^{DG}}{\alpha_a}\right)\| u_{h}^{DG}\|_{0}^2-\frac{C_{\Omega}^{DG}}{2\nu}\|f\|_0^2.
\end{align}
Next, let us define $\|u_h^{DG}\|_0=k_{DG}$, and note that
\begin{align*}
k_{DG}> \frac{(C_{\Omega}^{DG})}{\nu\sqrt{\alpha_a\nu+2\beta\gamma C_{\Omega}^{DG}-\beta(1+\gamma)^2C_{\Omega}^{DG}-\frac{2\alpha^2}{\beta}}}\|f\|_{0},
\end{align*}
provided that $\nu +2\beta \gamma C_{\Omega}^{DG}> \beta(1+\gamma)^2C_{\Omega}^{DG}+\frac{2\alpha^2}{\beta}$. Then the RHS in (\ref{eqdg12}) is non-negative. Finally,
Brouwer's fixed point theorem implies that $P_{DG}(u_h^{DG})=0$.
\end{proof}
On the other hand, we can
establish the following result, whose proof is similar to \eqref{crnclem111}.
\begin{lemma}\label{dglem111}
There holds:
\[
A_{DG}(v_1,w)-A_{DG}(v_2,w)\ge \tilde{C}_{DG}\vertiii{ w} ,
\]
where $v_1,v_2\in V_{h}^{DG}$ and $w=v_1-v_2$.
\end{lemma}
Finally, we can state an a priori error estimate in the following theorem.
\begin{theorem}\label{dgerr11}
Let $V_h^{DG}$ be as in (\ref{dgsubspace1}), and let us assume \eqref{234} and that $u$ satisfies \eqref{8p2}. Then, there exists $\tilde{C}$ is independent of $h$ such that
\[
\vertiii{ u^{DG}_h-u|} \leq \tilde{C}h.
\]
\end{theorem}
\begin{proof}
Using triangle inequality readily gives
\[
\vertiii{ u_h^{DG}-u} \le \vertiii{ u_h^{DG}-W} +\vertiii{W-u}.
\]
Proceeding again as in the conforming and non-conforming cases, we have the bound
\[
\vertiii{ W-u} \le Ch.
\]
Using the formulation \eqref{dgweakform}, we have
\[
A_{DG}(u_h^{DG},\chi)=(f,\chi), \quad \ \text{ for all }\ \chi\in V_h^{DG},
\]
and if $u\in D(A)=H_0^1(\Omega)\cap H^2(\Omega)$ satisfies \eqref{8p2}, then we
immediately have that
\[
A_{DG}(u,\chi)=(f,\chi), \quad \ \text{ for all }\ \chi\in V_h^{DG}.
\]
Finally, recalling Lemma \eqref{dglem111}, can write
\[
\tilde{C}\vertiii{\chi} \le A_{DG}(u_h^{DG},\chi)-A_{DG}(W,\chi)= A_{DG}(u,\chi)-A_{DG}(W,\chi),
\]
and the rest of the proof follows much in the same way as in Theorems \ref{thm7.1} and \ref{ncthm11}.
\end{proof}
\begin{remark}
Note that we can drive the following $L^2$-error estimates, essentially as a direct consequence of Theorems \ref{thm7.1}, \ref{ncthm11} and \ref{dgerr11}
\[
||u-u_h||_0\le C\,h,\quad ||u-u_h^{CR}||_0\le C\,h, \quad ||u-u_h^{DG}||_{0}\le C\,h,
\]
where the constant $C$ is independent of $h$. These $L^2$-error estimates are however sub-optimal. We nevertheless provide in Section \ref{sec4} numerical evidence that all three numerical methods achieve optimal convergence also in the $L^2-$norm.
\end{remark}
\section{Numerical results}\label{sec4}
In this section, we present a few computational results that confirm the
theoretical results advanced in Section~\ref{sec3}. All examples have been
implemented with the help of the open-source finite element library FEniCS \cite{alnaes_ans15}.
\subsection{Example 1: Accuracy verification against smooth solutions}
First we consider problem (\ref{8.1}) defined on the domain $\Omega=(0,1)^d$, where $d=2,3$. The two expressions of the exact solution $u$ are as follows:
\[
\mbox{Case } 1: u= \Pi_{i=1}^{d}(x_i-x_i^2),\qquad
\mbox{Case } 2: u= \frac{1}{16}\Pi_{i=1}^{d}\sin(\pi x_i).
\]
We choose the values of parameters as follows: $\alpha=0.2$, $\beta=0.1$, $\nu=2$ and $\gamma=0.5$, and the right-hand side
datum $f$ is manufactured using these closed-form solutions. A sequence of successively refined uniform meshes is constructed
and the error history (decay of errors measured in the energy and $L^2-$norm as well
as corresponding convergence rates) for the numerical solutions constructed with CGFEM, NCFEM and DGFEM are reported in what follows. Table \ref{table11-12} presents the convergence results related to Case 1 for 2D and 3D, whereas Table \ref{table13-14} shows the results pertaining to Case 2. In all tables we can observe that errors in the energy and $L^2-$norms decrease with the mesh size at rates $O(h)$ and $O(h^2)$, respectively. We have used in all simulations a first-order polynomial degree. Other sets of computations performed after modifying the values of the parameter $\delta$ to $3$ and $5$ (not reported here) also show optimal convergence. We can also see that the number of Newton iterations
required to reach the prescribed tolerance of $10^{-6}$ is at most three.
\begin{table}[ht!]
\caption{Example 1, case 1. Errors, iteration count, and convergence rates for the numerical solutions $u_h$, $u_h^{CR}$ and $u_h^{DG}$.}
\label{table11-12}
\begin{center}
{\small
\begin{tabular}{| c | c | c | c | c | c | c |}
\hline
\multicolumn{7}{|c|}{Error history in 2D}\\
\hline
\multirow{6}{*}{CGFEM}&\multicolumn{1}{|c|}{mesh}&\multicolumn{1}{|c|} {Newton it.} &\multicolumn{1}{|c|} {$H^1$-error}&\multicolumn{1}{|c|}{$O(h)$}&\multicolumn{1}{|c|}{$L^2$-error}&\multicolumn{1}{|c|}{$O(h^2)$}\\
\cline{2-7}
&{$4\times 4$}&$3$ &$5.90(-02)$ &$-$ &$5.38(-03)$ & $-$\\
\cline{2-7}
&{$8\times 8$}&$3$ &$3.01(-02)$ &$0.9709$ &$1.42(-03)$ & $1.9217$\\
\cline{2-7}
&{$16\times 16$}&$3$ &$1.51(-02)$ &$0.9952$ &$3.60(-04)$ & $1.9798$\\
\cline{2-7}
&{$32\times 32$}&$3$ &$7.60(-03)$ &$0.9904$ &$9.03(-05)$ & $1.9951$\\
\cline{1-7}
\multirow{5}{*}{NCFEM}
&{$4\times 4$}&$3$ &$4.62(-02)$ &$-$ &$2.32(-03)$ & $-$\\
\cline{2-7}
&{$8\times 8$}&$3$ &$2.35(-02)$ &$0.9752$ &$6.10(-04)$ & $2.1026$\\
\cline{2-7}
&{$16\times 16$}&$3$ &$1.18(-02)$ &$0.9938$ &$1.54(-04)$ & $1.9858$\\
\cline{2-7}
&{$32\times 32$}&$3$ &$5.91(-03)$ &$0.9975$ &$3.88(-05)$ & $1.9888$\\
\cline{1-7}
\multirow{5}{*}{DGFEM}
&{$4\times 4$}&$3$ &$5.83(-02)$ &$-$ &$5.27(-03)$ & $-$\\
\cline{2-7}
&{$8\times 8$}&$3$ &$2.94(-02)$ &$0.9876$ &$1.36(-03)$ & $1.9541$\\
\cline{2-7}
&{$16\times 16$}&$3$ &$1.46(-02)$ &$1.0098$ &$3.40(-04)$ & $2.0000$\\
\cline{2-7}
&{$32\times 32$}&$3$ &$7.25(-03)$ &$1.0099$ &$8.43(-05)$ & $2.0119$\\
\hline
\multicolumn{7}{|c|}{Error history in 3D}\\
\hline
\multirow{6}{*}{CGFEM}&\multicolumn{1}{|c|}{mesh}&\multicolumn{1}{|c|} {Newton it.} &\multicolumn{1}{|c|} {$H^1$-error}&\multicolumn{1}{|c|}{$O(h)$}&\multicolumn{1}{|c|}{$L^2$-error}&\multicolumn{1}{|c|}{$O(h^2)$}\\
\cline{2-7}
&{$4\times 4\times 4$}&$2$ &$1.63(-02)$ &$-$ &$1.52(-03)$ & $-$\\
\cline{2-7}
&{$8\times 8\times 8$}&$2$ &$8.54(-03)$ &$0.9325$ &$4.22(-04)$ & $1.8487$\\
\cline{2-7}
&{$16\times 16\times 16$}&$2$ &$4.32(-03)$ &$0.9832$ &$1.08(-04)$ & $1.9662$\\
\cline{2-7}
&{$32\times 32\times 32$}&$2$ &$2.16(-03)$ &$1.0000$ &$2.73(-05)$ & $1.9840$\\
\cline{1-7}
\multirow{5}{*}{NCFEM}
&{$4\times 4\times 4$}&$2$ &$1.06(-02)$ &$-$ &$5.42(-04)$ & $-$\\
\cline{2-7}
&{$8\times 8\times 8$}&$2$ &$5.39(-03)$ &$0.9757$ &$1.41(-04)$ & $1.9426$\\
\cline{2-7}
&{$16\times 16\times 16$}&$2$ &$2.70(-03)$ &$0.9973$ &$3.64(-05)$ & $1.9573$\\
\cline{2-7}
&{$32\times 32 \times 32$}&$2$ &$1.35(-03)$ &$1.0000$ &$8.99(-05)$ & $2.0175$\\
\cline{1-7}
\multirow{5}{*}{DGFEM}
&{$4\times 4\times 4$}&$3$ &$1.59(-02)$ &$-$ &$1.44(-03)$ & $-$\\
\cline{2-7}
&{$8\times 8\times 8$}&$3$ &$8.05(-03)$ &$0.9820$ &$3.85(-04)$ & $1.5409$\\
\cline{2-7}
&{$16\times 16\times 16$}&$3$ &$3.94(-03)$ &$1.0308$ &$9.49(-05)$ & $2.0204$\\
\cline{2-7}
&{$32\times 32 \times 32$}&$3$ &$1.93(-03)$ &$1.0296$ &$2.31(-05)$ & $2.0385$\\
\hline
\end{tabular}}
\end{center}
\end{table}
\begin{table}[ht!]
\caption{Example 1, case 2. Errors, iteration count, and convergence rates for the numerical solutions $u_h$, $u_h^{CR}$ and $u_h^{DG}$.}
\label{table13-14}
\begin{center}
{\small
\begin{tabular}{| c | c | c | c | c | c | c | }
\hline
\multicolumn{7}{|c|}{Error history in 2D}\\
\hline
\multirow{5}{*}{CGFEM}&\multicolumn{1}{|c|}{mesh}&\multicolumn{1}{|c|} {Newton it.} &\multicolumn{1}{|c|} {$H^1$-error}&\multicolumn{1}{|c|}{$O(h)$}&\multicolumn{1}{|c|}{$L^2$-error}&\multicolumn{1}{|c|}{$O(h^2)$}\\
\cline{2-7}
& $4\times 4$&$3$ &$1.26(-01)$ &$-$ &$1.08(-02)$& $-$\\
\cline{2-7}
&{$8\times 8$}&$3$ &$6.84(-02)$ &$0.8814$ &$3.21(-03)$ & $1.7504$\\
\cline{2-7}
&{$16\times 16$}&$3$ &$3.49(-02)$ &$0.9708$ &$8.45(-04)$ & $1.9256$\\
\cline{2-7}
&{$32\times 32$}&$3$ &$1.75(-02)$ &$0.9959$ &$2.14(-04)$ & $1.9813$\\
\cline{1-7}
\multirow{4}{*}{NCFEM}& $4\times 4$&$3$ &$1.22(-01)$ &$-$ &$7.62(-02)$& $-$\\
\cline{2-7}
&{$8\times 8$}&$3$ &$6.44(-02)$ &$0.9217$ &$2.09(-03)$ & $1.8663$\\
\cline{2-7}
&{$16\times 16$}&$3$ &$3.26(-02)$ &$0.9822$ &$5.38(-04)$ & $1.9578$\\
\cline{2-7}
&{$32\times 32$}&$3$ &$1.63(-02)$ &$0.9912$ &$1.35(-04)$ & $1.9946$\\
\cline{1-7}
\multirow{5}{*}{DGFEM}&{$4\times 4$}&$3$ &$1.23(-01)$ &$-$ &$1.01(-02)$& $-$\\
\cline{2-7}
&{$8\times 8$}&$3$ &$6.58(-02)$ &$0.9025$ &$2.99(-03)$ & $1.7561$\\
\cline{2-7}
&{$16\times 16$}&$3$ &$3.34(-02)$ &$0.9782$ &$7.86(-04)$ & $1.9275$\\
\cline{2-7}
&{$32\times 32$}&$3$ &$1.68(-02)$ &$0.9914$ &$1.99(-04)$ & $1.9818$\\
\hline
\multicolumn{7}{|c|}{Error history in 3D}\\
\hline
\multirow{6}{*}{CGFEM}&\multicolumn{1}{|c|}{mesh}&\multicolumn{1}{|c|} {Newton it.} &\multicolumn{1}{|c|} {$H^1$-error}&\multicolumn{1}{|c|}{$O(h)$}&\multicolumn{1}{|c|}{$L^2$-error}&\multicolumn{1}{|c|}{$O(h^2)$}\\
\cline{2-7}
& $4\times 4 \times 4$&$3$ &$1.07(-01)$ &$-$ &$9.25(-03)$& $-$\\
\cline{2-7}
&{$8\times 8\times 8$}&$3$ &$5.98(-02)$ &$0.7650$ &$2.97(-03)$ & $1.4731$\\
\cline{2-7}
&{$16\times 16\times 16$}&$3$ &$3.08(-02)$ &$0.9325$ &$8.04(-04)$ & $1.8487$\\
\cline{2-7}
&{$32\times 32\times 32$}&$3$ &$1.55(-02)$ &$0.9832$ &$2.05(-04)$ & $1.9662$\\
\cline{1-7}
\multirow{5}{*}{NCFEM}&{$4\times 4\times 4$}&$3$ &$8.79(-02)$ &$-$ &$5.09(-03)$& $-$\\
\cline{2-7}
&{$8\times 8\times 8$}&$3$ &$4.54(-02)$ &$0.9159$ &$1.39(-03)$ & $1.7789$\\
\cline{2-7}
&{$16\times 16\times 16$}&$3$ &$2.29(-02)$ &$0.9757$ &$3.56(-04)$ & $1.9426$\\
\cline{2-7}
&{$32\times 32\times 32$}&$3$ &$1.14(-02)$ &$0.9973$ &$8.97(-05)$ & $1.9573$\\
\cline{1-7}
\multirow{5}{*}{DGFEM}&{$4\times 4\times 4$}&$3$ &$1.00(-01)$ &$-$ &$8.03(-03)$& $-$\\
\cline{2-7}
&{$8\times 8\times 8$}&$3$ &$5.38(-02)$ &$0.8943$ &$2.51(-03)$ & $1.6777$\\
\cline{2-7}
&{$16\times 16\times 16$}&$3$ &$2.74(-02)$ &$0.9734$ &$6.74(-04)$ & $1.8969$\\
\cline{2-7}
&{$32\times 32\times 32$}&$3$ &$1.37(-02)$ &$1.0000$ &$1.71(-04)$ & $1.9788$\\
\hline
\end{tabular}}
\end{center}
\end{table}
\subsection{Example 2: Stationary wave solution}
Next we consider (\ref{8.1}) endowed with non-homogeneous Dirichlet boundary conditions.
The domain is again as in Example 1, and the setup of the problem has been adopted
from \cite{VJE}, where the exact solution is
\[
u= 0.5-0.5\tanh(z/(r-\bar{\alpha})),
\]
with $r= \sqrt{\bar{\alpha}^2+8}$ and $\bar{\alpha}=\alpha\sqrt{2}$. The values of the
model parameters are now $\alpha=0.2$, $\beta=1$, $\nu=16$ and $\gamma=0.5$. In Table \ref{table15-16} we present
the convergence rates associated with the errors in the energy norm as well as $L^2$-norm for CGFEM, NCFEM and DGFEM. Again we observe optimal convergence in all instances.
\begin{table}[ht!]
\caption{Example 2. Errors, iteration count,
and convergence rates for the numerical solutions $u_h$, $u_h^{CR}$ and $u_h^{DG}$.}
\label{table15-16}
\begin{center}
{\small
\begin{tabular}{| c | c | c | c | c | c | c | }
\hline
\multicolumn{7}{|c|}{Error history in 2D}\\
\hline
\multirow{6}{*}{CGFEM}&\multicolumn{1}{|c|}{mesh}&\multicolumn{1}{|c|} {Newton it.} &\multicolumn{1}{|c|} {$H^1$-error}&\multicolumn{1}{|c|}{$O(h)$}&\multicolumn{1}{|c|}{$L^2$-error}&\multicolumn{1}{|c|}{$O(h^2)$}\\
\cline{2-7}
&{$4\times 4$}&$3$ &$1.16(-02)$ &$-$ &$8.99(-04)$ & $-$\\
\cline{2-7}
&{$8\times 8$}&$3$ &$5.83(-03)$ &$0.9926$ &$2.26(-04)$ & $1.9920$\\
\cline{2-7}
&{$16\times 16$}&$3$ &$2.91(-03)$ &$1.0025$ &$5.67(-05)$ & $1.9949$\\
\cline{2-7}
&{$32\times 32$}&$3$ &$1.45(-03)$ &$1.0050$ &$1.41(-05)$ & $2.0077$\\
\cline{1-7}
\multirow{5}{*}{NCFEM}
&{$4\times 4$}&$3$ &$7.96(-03)$ &$-$ &$3.91(-04)$ & $-$\\
\cline{2-7}
&{$8\times 8$}&$3$ &$3.98(-03)$ &$1.0000$ &$9.80(-05)$ & $1.9963$\\
\cline{2-7}
&{$16\times 16$}&$3$ &$1.99(-03)$ &$1.0000$ &$2.45(-05)$ & $2.0000$\\
\cline{2-7}
&{$32\times 32$}&$3$ &$9.96(-04)$ &$0.9986$ &$6.13(-06)$ & $1.9988$\\
\cline{1-7}
\multirow{5}{*}{DGFEM}
&{$4\times 4$}&$3$ &$1.13(-02)$ &$-$ &$8.84(-04)$ & $-$\\
\cline{2-7}
&{$8\times 8$}&$3$ &$5.57(-03)$ &$1.0206$ &$2.19(-04)$ & $2.0131$\\
\cline{2-7}
&{$16\times 16$}&$3$ &$2.76(-03)$ &$1.0130$ &$5.47(-05)$ & $2.0013$\\
\cline{2-7}
&{$32\times 32$}&$3$ &$1.37(-03)$ &$1.0105$ &$1.36(-05)$ & $2.0079$\\
\hline
\multicolumn{7}{|c|}{Error history in 3D}\\
\hline
\multirow{6}{*}{CGFEM}&\multicolumn{1}{|c|}{mesh}&\multicolumn{1}{|c|} {Newton it.} &\multicolumn{1}{|c|} {$H^1$-error}&\multicolumn{1}{|c|}{$O(h)$}&\multicolumn{1}{|c|}{$L^2$-error}&\multicolumn{1}{|c|}{$O(h^2)$}\\
\cline{2-7}
&{$4\times 4\times 4$}&$3$ &$2.39(-02)$ &$-$ &$1.98(-03)$ & $-$\\
\cline{2-7}
&{$8\times 8\times 8$}&$3$ &$1.19(-02)$ &$1.0060$ &$5.01(-04)$ & $1.9826$\\
\cline{2-7}
&{$16\times 16\times 16$}&$3$ &$5.98(-03)$ &$0.9927$ &$1.25(-04)$ & $2.0029$\\
\cline{2-7}
&{$32\times 32\times 32$}&$3$ &$2.99(-03)$ &$1.0000$ &$3.14(-05)$ & $1.9931$\\
\cline{1-7}
\multirow{5}{*}{NCFEM}
&{$4\times 4\times 4$}&$3$ &$1.35(-02)$ &$-$ &$7.07(-04)$ & $-$\\
\cline{2-7}
&{$8\times 8\times 8$}&$3$ &$6.75(-03)$ &$1.0000$ &$1.77(-04)$ & $1.9980$\\
\cline{2-7}
&{$16\times 16\times 16$}&$3$ &$3.37(-03)$ &$1.0021$ &$4.42(-05)$ & $2.0016$\\
\cline{2-7}
&{$32\times 32 \times 32$}&$3$ &$1.68(-04)$ &$1.0043$ &$1.10(-05)$ & $2.0065$\\
\cline{1-7}
\multirow{5}{*}{DGFEM}
&{$4\times 4\times 4$}&$3$ &$2.30(-02)$ &$-$ &$1.95(-03)$ & $-$\\
\cline{2-7}
&{$8\times 8\times 8$}&$3$ &$1.11(-02)$ &$1.0511$ &$4.84(-04)$ & $2.0104$\\
\cline{2-7}
&{$16\times 16\times 16$}&$3$ &$5.47(-03)$ &$1.0209$ &$1.19(-04)$ & $2.0240$\\
\cline{2-7}
&{$32\times 32\times 32$}&$3$ &$2.70(-03)$ &$1.0186$ &$2.96(-05)$ & $2.0073$\\
\hline
\end{tabular}}
\end{center}
\end{table}
\subsection{Example 3: Application to nerve pulse propagation}
To conclude this section, and as a qualitative illustration of the differences between a classical bistable
equation (without advection and with a simplified cubic nonlinearity induced by $\delta = 1$)
and the generalized Burgers-Huxley equation, we conduct a simple simulation of a transient
problem where also an additional ODE (governing the dynamics of a gating variable $v$)
is considered so that self-sustained patterns are possible (see, e.g., \cite{murray,bini10}).
The system reads
\begin{equation}\label{transientBH}
\partial_t u + \alpha u^\delta \sum_{i=1}^d\partial_i u - \nu \Delta u - \beta u (1-u^\delta)(u^\delta-\gamma) + v = 0, \qquad \partial_t v = \varepsilon(u-\rho v).\end{equation}
Setting $\delta = 1$ and $\alpha = 0$, one recovers the well-known FitzHugh-Nagumo equations
$$ \partial_t u - \nu \Delta u - \beta u (1-u)(u-\gamma) + v = 0, \qquad \partial_t v = \varepsilon(u-\rho v).$$
We apply a simple backward Euler time discretization with constant time step $\Delta t = 0.2$, after which we recover a discrete formulation
resembling \eqref{7p1} for the CFEM (and similarly for the other two methods). The domain $\Omega = (0,300)^2$ is discretized into
a uniform triangular mesh with 25K elements, and the model parameters are taken as
$\alpha = 0.1, \delta = 1.5, \beta = \nu = 1, \varepsilon = \gamma = 0.01, \rho = 0.05$ (see also \cite{buerger10} for the classical FitzHugh-Nagumo parameters, whereas the modified terms adopt here very mild values).
For this example we prescribe Neumann boundary conditions for $u$ on $\partial\Omega$. Figure~\ref{fig:ex3}
depicts three snapshots of the evolution of $u$ (representing the action potential propagation in a
piece of nerve tissue, cardiac muscle, or any excitable media) for the classical FitzHugh-Nagumo system
vs. the modified generalized Burgers-Huxley system \eqref{transientBH}, all numerical solutions computed using the DGFEM setting $\gamma = 2$. The differences in spiral dynamics (initiated with a cross-shaped and shifted initial condition for $u$ and $v$) seem to be more sensitive to
the amount of additional nonlinearity (encoded in $\delta$), rather than to the intensity of the additional advection (modulated by
$\alpha$).
\begin{figure}[ht!]
\caption{Example 3. Snapshots at $t = 80, 200, 650$ of $u_h^{DG}$ for the FitzHugh-Nagumo model using $\delta = 1$, $\alpha = 0$ (top panels) and
for the modified generalized Burgers-Huxley system \eqref{transientBH} with $\delta = 1$, $\alpha = 0.1$ (middle row) and
with $\delta = 1.5$, $\alpha = 0.1$ (bottom).}\label{fig:ex3}
\begin{center}
\includegraphics[width=0.325\textwidth]{ex03a}
\includegraphics[width=0.325\textwidth]{ex03b}
\includegraphics[width=0.325\textwidth]{ex03c}\\
\includegraphics[width=0.325\textwidth]{ex03d}
\includegraphics[width=0.325\textwidth]{ex03e}
\includegraphics[width=0.325\textwidth]{ex03f}\\
\includegraphics[width=0.325\textwidth]{ex03g}
\includegraphics[width=0.325\textwidth]{ex03h}
\includegraphics[width=0.325\textwidth]{ex03i}
\end{center}
\end{figure}
\section{Conclusion}
In this paper we have addressed two main contributions. First, we have proved the well-posedness for the stationary generalized Burgers-Huxley equation. Moreover, we have established a new regularity result that only uses minimal theoretical requirements.
Secondly, we have introduced three types of finite element approximations (CFEM, NCFEM and DGFEM) for \eqref{8.1}. We have rigorously derived a priori error estimates for all of these discretizations. Finally, computational results are given to validate the theoretical first-order convergence of the methods.
As a next step we are extending the theory to cover the transient case, and we will also construct efficient and reliable residual-based a posteriori error estimators and adaptive schemes. We also plan to address the formulation of other conservative discretizations using adequate mixed methods.
\medskip\noindent
\bibliographystyle{siam}
|
\section*{Acknowledgments}
This research is supported by the Ministry of Education and Science of Ukraine (Project No. 0119U103982).
\section{Introduction}
\label{intro}
The process of proton-proton elastic scattering is the subject of both theoretical and experimental research for the last few decades. The humanity have accumulated a lot of experimental data \cite{Cafagna:2020izf}, but we still not have a dynamic theory from the first principles \cite{Dremin:2017eng} yet. The descriptions of such processes are usually built with either the phenomenological approaches \cite{Csorgo:2018ruk} which are different variations of Regge theory \cite{Collins1977, Ryskin:2009qf}, or with the geometrical models \cite{Dremin:2019pza,Islam:2017muy,UgoAmaldi}. All these approaches are based on some assumptions that are not the corollaries of fundamental physical principles, but may even be in conflict with these principles. In particular, the Reggeized models \cite{FADIN197550,Kuraev:1977eng,Lipatov:2008eng} use the assumption that multi-Regge region \cite{KRFeng,MultiRggeBFKLeng,Duhr2019,Hentschinski:2020rfx} provides the dominant contribution to the integrals for observables. This region of the phase space contains the points corresponding to the significantly different values of energy-momentum of the secondary particles in the final state of the scattering \cite{Sharph:2011wm}, which violates the energy-momentum conservation law.
Among the various experimental data regarding the proton-proton(antiproton) scattering the most theoretical investigations have centred on the elastic scattering \cite{Dremin:2013eng}. The later is justified by expectation for the description of elastic processes to be significantly simpler compared to the description of inelastic ones. In particular, the scattering amplitude for elastic process is the function of just two Lorentz-invariants \cite{Collins1977}. The description of inelastic processes is then obtained from the description of elastic processes with the Abramovsky-Gribov-Kancheli (AGK) cutting rules \cite{AGKeng,Levin:1900tt,Bartels:2008ru,Pelicer:2018xzu}. However, this approach requires additional assumptions regarding the form of multi-Regge vertex \cite{Gribov_Reggeonnaja_technica_eng,BAKER19761} and introduction of unobservable quantity such as cross-section of scattering with an exchange of a certain number of Reggeons. All these assumptions are directed to establish the elastic scattering amplitude dependence on Mandelstam variable $s$. At the same time, the dependence on another variable $t$ remains undefined. In the simplest one-Reggeon model $t$-dependence is included in both Regge pole trejectory and residue. While the trajectory is defined by the masses of $t$-channel resonances, the residue remains completely undefined. In the context of Regge diagrams this residue is explained as the product of vertices corresponding to the interaction of Reggeon with the hadrons undergoing elastic scattering \cite{Nikitin1980,Shabelski:2014yba,Kaidalov:2003eng}. The certain $t$-dependence of such vertices is postulated in such a way that allows one to describe the experimentally measured differential cross-section ${d\sigma}/{dt}$ at low values of $t$. This dependence is usually postulated to be $ \exp \left( -{{R}^{2}}\left| t \right| \right) $, where $R$ is a fitted parameter also known as the Regge radius of hadron \cite{Nikitin1980}. As being said, such assumption provides the desirable $t$-dependence of the differential cross-section $d\sigma/dt$ of elastic scattering at low $t$ which is close to linear one in logarithmic scale applied to the cross-section axis \cite{Ter-Martirosyan}. However, this dependence is not linear on the whole range of measurements, but is non-monotonic and has maxima and minima \cite{NAGY1979221,PhysRevLett.54.2180,Antchev:2018edk,Abazov:2020rus}. The models with the multi-Regge exchange of simple poles and the model of quasi-eikonal multi-Regge vertices \cite{Ter-Martirosyan,Kaidalov:1983vn_eng, Kaidalov:1983ew} are failed to explain these features. Indeed, this non-monotonic behavior has been described within the additive quark model \cite{Shabelski:2014yba} due to the interference contributions from processes with an exchange of various number of Reggeons. As for the phenomenological approaches, this dependence is also reproduced within the models with a Regge multiple poles \cite{Dipol_eng,PhysRevD.87.114018,PhysRevD.76.074030}. However, these models have high uncertainty in $t$-dependence, since the increasing in the order of pole requires increasing in the number of terms in Laurent series whose forms are also need postulating.
The non-monotonic dependency of differential cross-section ${d\sigma}/{dt}(t)$ was also reproduced in the models with simple Regge poles of various signatures and with the two-Reggeon cuts \cite{refId0}.
The present research is devoted to the description of effect of non-monotonic $t$-dependence of the elastic scattering differential cross-section $d\sigma/dt$ built upon a purely dynamical model. In other words, we use the model that is based on the fundamental physical principles, beginning from the Lagrangian and its corresponding dynamical equations, quantization, and solution of the equations describing the dynamics of corresponding relativistic quantum system in Fock space. We do that in the framework of the multi-particle fields model.
\subsection{The second pair of denominators}
Let us now consider the region of the integration domain in \eqref{tensor_ab_short}, where the divergence of the integrand arises due to the second pair of denominators
\begin{equation}
\begin{split}
&\frac{1}{M_{G}^{2}-{{\left( {{k}^{0}} \right)}^{2}}+{{{\vec{k}}}^{2}}-i\varepsilon }
\\
\times
&\frac{1}{M_{G}^{2}-{{\left( {{k}^{0}} \right)}^{2}}+{{\left( \left( {{{\vec{P}}}_{1}}-{{{\vec{P}}}_{3}} \right)-\vec{k} \right)}^{2}}-i\varepsilon },
\end{split}
\label{second_denom_pair}
\end{equation}
\noindent
If we let both the energy component ${{k}^{0}}$ of four-momentum $k$ and the length of its spatial part $| {\vec{k}} |$ tend to infinity simultaneously, the difference of their squares ${{\left( {{k}^{0}} \right)}^{2}}-{{| {\vec{k}} |}^{2}}$ may still be finite. In contrast, the sum ${{\left( {{k}^{0}} \right)}^{2}} + {{| {\vec{k}} |}^{2}}$ takes infinite values even when either ${{k}^{0}}$ or $| {\vec{k}}|$ tends to infinity. Note that the first pair of denominators contains ${{k}^{0}}$ in first power, which provides the decreasing of the integrand as ${{k}^{0}}$ tends to infinity. At the same time, in the center of mass system the energy component of ${{P}_{1}}-{{P}_{3}}$ is equal to zero, and the denominators \eqref{second_denom_pair} do not contain the first power of ${{k}^{0}}$, but only the difference ${{\left( {{k}^{0}} \right)}^{2}}-{{| {\vec{k}} |}^{2}}$ which, as we have already mentioned, may take finite values as ${{k}^{0}} \to \infty$ and $| {\vec{k}} | \to \infty$. In this case the whole integrand goes to zero due to the first pair of denominators containing the first power of ${{k}^{0}}$.
If we consider the component $t_{00}$ of the tensor \eqref{tensor_ab_short}, the numerator ${{f}_{ab}}\left( {{k}^{0}},\vec{k} \right)=s-{{\left( {{k}^{0}} \right)}^{2}}$, where $s$ is the Mandelstam invariant, tends to infinity as the denominator of the integrand. In this case the whole integral diverges. However, this divergence of \eqref{tensor_ab_short} is only
delusive. Let us show that.
As mentioned above, if the expression contained the sum instead of the difference ${{\left( {{q}^{0}} \right)}^{2}}-{{\left| {\vec{q}} \right|}^{2}}$, there would be no problem with the divergence at all. It is possible to turn this difference into a sum if we succeeded by transforming the integration contour in the complex plane from the real axis to the imaginary one, which can be done by the Wick rotation. The possibility of this transformation depends on the location of the integrand poles. Les us take a closer look at them.
First, we rewrite the expression \eqref{tensor_t_through_q} in the form
\begin{widetext}
\begin{equation}\label{tensor10}
\begin{split}
{{t}_{ab}}= & \int\limits_{-\infty }^{\infty }{d{{q}_{0}}}\int\limits_{-\infty }^{\infty }{d\vec{q}}
\left[
{{f}_{ab}}\left( {{q}_{0}},\vec{q} \right)
\frac{1}{{{q}^{0}}-\left( -\frac{\sqrt{s}}{2}+ \eta \right)}
\frac{1}{{{q}^{0}}-\left( -\frac{\sqrt{s}}{2}- \eta \right)}
\right.
\\
& \times
\left.
\frac{1}{{{q}^{0}}-\left( \frac{\sqrt{s}}{2}+ \eta \right)}
\frac{1}{{{q}^{0}}-\left( \frac{\sqrt{s}}{2}- \eta \right)}
\frac{1}{\left(z_3 - i\varepsilon\right)\left(z_4 - i\varepsilon\right)}
\right] ,
\end{split}
\end{equation}
\end{widetext}
\noindent
where $\eta = \sqrt{M_{p}^{2}+{{\left( {{q}^{1}} \right)}^{2}}+{{\left( {{q}^{2}} \right)}^{2}}+{{\left( \frac{\sqrt{4{{P}^{2}}-\left| t \right|}}{2}+{{q}^{3}} \right)}^{2}}-i\varepsilon }$.
Let us analyze the placement of quantities $\eta$ and $-\eta$ on the complex plane. It depends on the variables $ {{q}^{1}},{{q}^{2}},{{q}^{3}}$, so it varies across different subsets of the ${{q}^{1}},{{q}^{2}},{{q}^{3}}$ integration domain. Thus the major question here is which of the two conditions
\begin{equation}
\begin{aligned}[c]
\eta {\vert_{\varepsilon = 0}} > \frac{\sqrt{s}}{2}
\end{aligned}
\qquad
\begin{aligned}[c]
\text{or}
\end{aligned}
\qquad
\begin{aligned}[c]
\eta {\vert_{\varepsilon = 0}} <\frac{\sqrt{s}}{2}
\end{aligned}
\label{quadrants_ineq}
\end{equation}
\noindent
is met for the particular subset. Whether the first or the second condition is met, determines the distribution of the poles among the quadrants of the complex plane.
In case the second inequality \eqref{quadrants_ineq} holds, each quadrant of the complex plane contains a pole of the integrand \eqref{tensor10}. It means that one cannot do Wick rotation, because one would cross the poles during rotation.
In the case of the first inequality, the poles are located in the second and fourth quadrants, so it is possible to rotate the integration path in the first and third quadrants. Thus the integration over ${{q}^{0}}$ can be transformed into the integration along the imaginary axis (Wick rotation), and the expression ${{\left( {{q}^{0}} \right)}^{2}}-{{\left| {\vec{q}} \right|}^{2}}$ transforms into $\left( -{{\left( {{q}^{0}} \right)}^{2}}-{{\left| {\vec{q}} \right|}^{2}} \right)$ as the integration variable changes ${{q}^{0}} \rightarrow i{{q}^{0}}$. If one of these components tends to infinity, their sum also tends to infinity, which solves the problem. The integrand tends to zero as ${{\left( {{\left( {{q}^{0}} \right)}^{2}}+{{\left| {\vec{q}} \right|}^{2}} \right)}^{-4}}$, which guarantees the convergence of the integral.
The possibility to apply Wick rotation arises in the region of high values of $ \left|\vec{q}\right|$, i.e. in that part of the integration domain where the integral can diverge. Therefore, selecting this region and applying corresponding Wick rotation, we obtain the convergent integral. The rest of the integration domain is finite and does not impact the convergence of the integral, since there are no singularities of the integrand in this region at $ \varepsilon \neq 0$. This was actually the reason for calling the divergence of the integral ``delusive'' earlier in this section.
\section{Problem statement and literature review}
We described the model of multi-particle fields in \cite{Korotca_statta_v_UJP, Ptashynskiy2019MultiparticleFO,Sharph:2020gki} and showed that hadrons can be considered as the quanta of field defined on the subsets of simultaneous events. These subsets are extracted from the tensor product of two (for mesons) and three (for barions) Minkowski spaces. The codomains of these fields are invariant subspaces of the tensor product of bispinors, where the scalar (for mesons) and bispinor (for protons) representations of Lorentz group act. The trivial representation of $SU_c\left(3\right)$ group acts also on these subsets, which corresponds to the fact that hadrons have no color. As usually, we build a Lagrangian for these fields and apply gauge principle. The local $SU_c\left(3\right)$-invariance of Lagrangian is provided by the gauge fields transforming by the tensor representation of the Lorentz group and internal symmetry groups. In this case, the common way of providing the gauge invariance through the introduction of covariant derivatives is the special case of the multi-particle approach for obtaining the local invariance. We have shown \cite{Korotca_statta_v_UJP, Ptashynskiy2019MultiparticleFO,Sharph:2020gki} that these tensor fields can be used for description of the creation and annihilation processes of glueballs, i.e. the bound states of confined gluons. The same fields provide both the confinement of quarks within the protons and mesons, and interaction between the quarks of different protons through the exchange of glueballs. As the result, we have the dynamical model for the three-particle bispinor fields that interact via the two-particle glueball fields. The calculation of observable quantities within this model can be performed with Feynman diagram technique, since such nonperturbative effects as the confinement of quarks and gluons are already accounted for in the internal dynamics of multi-particle field quanta. The non-zero masses of glueballs are also obtained from the dynamical equations in a natural way. This leads to the finite value of the elastic scattering amplitude at $t = 0$ due to the strong interaction. It is known that the total proton-(anti)proton scattering cross-section is finite after exclusion of electromagnetic interaction \cite{Cartiglia:2013jya,UgoAmaldi}. As the result of optical theorem \cite{Collins1977}, the scattering amplitude must have a finite value at $t = 0$. This finiteness is simply postulated in the mentioned phenomenological models. So it is necessary to develop a model that would describe the mentioned features of scattering amplitude without simply postulating them. In QCD perturbation theory one have an infrared singularity for the scattering with a massless gluon exchange. It is clearly problematic to use perturbative QCD at $t = 0$, where finiteness of the elastic scattering amplitude may be caused by nonperturbative effects. However, it is still unclear which nonperturbative effects and how they lead to the finiteness of the scattering amplitude. Meanwhile, in the model of multi-particle fields it is the non-zero mass of the glueball (which is actually the consequence of nonperturbative effects) that leads to the finite value of the scattering amplitude.
The qualitative description of the inelastic scattering may be done considering tree-level diagrams only \cite{PhysRevD.101.076021}. However, the tree-level diagrams (see \fref{fig:polusnaya}) are not enough to describe the elastic scattering. Nevertheless, the non-monotonicity of the cross-section dependence appears even in such a simple model, except that it reproduces only the minimum, but not maximum.
\begin{figure}[tbp!]
\centering
\includegraphics[width=1\linewidth]{figs/pole_tree_diagram.pdf}
\caption{
Pole tree-level diagrams of elastic proton-proton scattering. ${{P}_{1}},{{P}_{2}}$ are the four-momenta of initial protons, ${{P}_{3}},{{P}_{4}}$ are the four-momenta of outgoing protons. Double lines correspond to the glueballs (bound state of gluons). The \textit{minus} between the diagrams reflects the fact that protons are described by Fermi-Dirac statistics.
}
\label{fig:polusnaya}
\end{figure}
Therefore, the description of experimental data requires the calculation of more complex loop diagrams.
The calculation of such diagrams is reduced to the calculation of multidimensional integral over virtual four-momenta.
Similar integrals have been considered in \cite{Eden:98637} with an application of Feynman parametrization, where, however, only the location of the singularities was studied but not the calculation method. The Feynman identity simplifies the integrand due to the fact that instead of the product of multiple Feynman denominators we obtain the power of single denominator, while, unfortunately, it also leads to the essentially more complex integration domain. For this reason, here we consider an approximate method of calculation starting from the expression for the loop diagram contribution to the elastic scattering amplitude. There is also another approach to the similar integrals \cite{PhysRev.186.1656} based on the eikonal approximation. However, such approximation significantly changes the pole structure of the integrand. While the eikonal approximation allows one to reduce the dimension of the integral for the eikonal, such approach does not solve the problem of calculating the integrals themselves. The main problem here is to calculate the limit of multidimensional integral as the parameters that shift the integrand poles tend to zero. Note that taking the limit inside the integral is not allowed in this case, because the poles will move inside the integration domain so the integral will diverge. This puts restrictions on the application of numerical integration methods, because taking the limit in the end requires the calculation at small parameters, which makes the poles close to the integration domain and thus complicates the numerical calculations.
To solve the outlined problem, we use the Laplace's method \cite{DeBruijn:225131} which worked well for the description of inelastic scattering processes \cite{Sharf:2011ujp,Sharf:2006it, Sharf:2007cf, Sharf:2009yy}. This method allowed us to effectively calculate the integrals with dimension up to 100. However, in those works the limit could be taken before the integration, because the poles of integrands were outside of the integration domain, which is not the case here.
Now let us figure out how to apply Laplace's method in the present situation. We know that the problem arises from the Feynman denominators that correspond to the lines of a diagram. Let us number these lines in an arbitrary order and put the expression of each denominator in the form ${{\left( {{z}_{a}}-i\varepsilon \right)}^{-1}}$, where $a$ is the line number, $z_a$ is the expression corresponding to the $a$-th line, $ \varepsilon $ is the parameter which should be made zero after the integral is calculated. Selecting a subset of $k$ lines we can equate the corresponding expressions $z_a$ to zero and consider the obtained system of equations. If the obtained system of equations is consistent, it defines the subset of integration domain where the integrand is equal to ${{\varepsilon }^{-k}}$. As a result, the major contribution to the integral comes from the region in which the greatest number of denominators is equal to zero. Let us denote this number by $l$ and the total number of integration variables by $n$. From the corresponding system of equations we can express the $l$ variables through the rest $n - l$ ones. Then it is convenient to change the first $l$ integration variables. These new variables are the deviations from the values that satisfy the equation system for $l$ denominators. It means that the absolute value of the integrand now has a distinct maximum at zero values of the first $l$ variables regardless of the next $n - l$ variables. Then we can apply the Laplace's method to integrate over the first $l$ variables. As we have already mentioned, the absolute value of the integrand at the maximum point is equal to ${{\varepsilon }^{-l}}$. At the same time, when applying Laplace's method, there comes the gaussian integral which leads to the factor ${{\varepsilon }^{-l}}$. It removes the $\varepsilon$ from expression and allows one to turn $\varepsilon$ to zero before the integration. Then the obtained integral can be calculated using numerical methods.
In the present paper we apply the described idea to calculate the sum of simplest single-loop diagrams for the elastic scattering of protons (\fref{fig:simplest_loops}).
\begin{figure}[tbp!]
\includegraphics[width=1\linewidth]{figs/simplest_loop_diagrams}
\caption{\label{fig:simplest_loops} The simplest loop diagrams for the elastic scattering of protons. ${{P}_{1}},{{P}_{2}}$ and ${{P}_{3}},{{P}_{4}}$ are the four-momenta of the incoming and outgoing protons correspondingly, $k$ is the four-momentum of the virtual particle. Double gluon lines correspond to the glueballs (bound state of gluons).}
\end{figure}
We calculate the elastic proton-proton scattering differential cross-section within the multi-particle fields approach. We consider the contributions of the diagrams in Figs.~\ref{fig:polusnaya},~\ref{fig:simplest_loops}, including the diagrams obtained from those depicted on \fref{fig:simplest_loops} by interchanging the final-state particle lines. Finally, we compare the calculation results with the experimental data.
\section{Calculation of the second maximum with the Wick rotation}
\subsection{Calculation of the second contribution with the Wick rotation}
Let us implement the idea explained in the previous section. We split the integral \eqref{tensor_t_through_q} into two integrals. The first integral over the finite region, where the Wick rotation cannot be applied, we denote as
\begin{equation}\label{tensordovika}
\begin{split}
& {{t}_{ab}^{<}}=\int\limits_{-\infty }^{\infty }{d{{q}^{0}}}\int\limits_{\left| {\vec{q}} \right|<cP}{d\vec{q}}{{f}_{ab}}\left( {{q}^{0}},\vec{q} \right)
\prod_{j=1}^4\frac{1}{\left(z_j - i\varepsilon\right)},
\end{split}
\end{equation}
\noindent
where $c > 1$. The value of $c$ has no impact on the result of integration. However, if $c=1$ and $\left| {\vec{q}} \right|<P$, then the poles of integrand \eqref{tensor10} lie on the integration path and the integral diverges, which imposes the lower bound on $c$.
The second contribution to the tensor ${{t}_{ab}}$, i.e. the integral over the region where the Wick rotation can be applied, we denote as
\begin{equation}\label{tensorposlevika}
\begin{split}
& {{t}_{ab}^{>}}=\int\limits_{-\infty }^{\infty }{d{{q}^{0}}}\int\limits_{\left| {\vec{q}} \right|>cP}{d\vec{q}}{{f}_{ab}}\left( {{q}^{0}},\vec{q} \right)
\prod_{j=1}^4\frac{1}{\left(z_j - i\varepsilon\right)}.
\end{split}
\end{equation}
\noindent
Here we also use the notations \eqref{zpm} for denominators with $\vec{P}_1$ and $\vec{P}_3$ expressed through $t$ \eqref{cherez_t}
\begin{equation}\label{zpm2}
\begin{split}
z_1^{\pm}=M_{p}^{2}-{{\left( \frac{\sqrt{s}}{2}\pm{{q}^{0}} \right)}^{2}}+{{\left( {{q}^{1}} \right)}^{2}}+{{\left( {{q}^{2}} \right)}^{2}}
\\
\hspace{2cm}
+{{\left( \frac{\sqrt{4P^2 - \vert t \vert}}{2}+{{q}^{3}} \right)}^{2}},
\\
z_3^{\pm} = M_{G}^{2}-{{\left( {{q}^{0}} \right)}^{2}}+{{\left( \frac{\sqrt{|t|}}{2} \pm {{q}^{1}} \right)}^{2}}
\hphantom{11+1}
\\
\hspace{3cm}
+{{\left( {{q}^{2}} \right)}^{2}}+{{\left( {{q}^{3}} \right)}^{2}} .
\end{split}
\end{equation}
Next we consider the system of equations for the second pair of denominators similar to what we did for the first pair of denominators. First, we set the real parts equal to zero and then find the second conditional maxima
\begin{equation}\label{vtoroymax}
\left\{ \begin{split}
& z_3^- = M_{G}^{2}-{{\left( {{q}^{0}} \right)}^{2}}+{{\left( \frac{\sqrt{\left| t \right|}}{2}-{{q}^{1}} \right)}^{2}}+{{\left( {{q}^{2}} \right)}^{2}}+{{\left( {{q}^{3}} \right)}^{2}}=0 \\
& z_3^+ = M_{G}^{2}-{{\left( {{q}^{0}} \right)}^{2}}+{{\left( \frac{\sqrt{\left| t \right|}}{2}+{{q}^{1}} \right)}^{2}}+{{\left( {{q}^{2}} \right)}^{2}}+{{\left( {{q}^{3}} \right)}^{2}}=0. \\
\end{split}\right.\
\end{equation}
\noindent
Let $\chi = \sqrt{M_{G}^{2}+\frac{\left| t \right|}{4}+{{\left( {{q}^{2}} \right)}^{2}}+{{\left( {{q}^{3}} \right)}^{2}}}$. It is easily seen that ${{q}^{1}}=0$, ${{q}^{0}}=\pm \chi$ is the solution of \eqref{vtoroymax}. Note that ${{q}^{0}}=\pm \chi$ can take both positive and negative values, but we consider only positive solutions, since the integrand is an even function of ${{q}^{0}}$. We can now proceed the calculation of the integral \eqref{tensordovika} analogously to what we did in Section \ref{first_pair_section}. First, we represent the second pair of denominators in exponential form
\begin{equation}\label{z34_exp_form}
\begin{split}
\frac{1}{z_3^\pm - i\varepsilon}
= &
\frac{1}{
\sqrt{
\left(z_3^\pm\right)^2 + \varepsilon^2
}
}
\\
& \times
\exp{
\left(
i
\arccos{
\frac
{z_3^\pm}
{\sqrt{\left(z_3^\pm\right)^2 + \varepsilon^2 }}
}
\right)
}.
\end{split}
\end{equation}
\noindent
In order to cancel the $\varepsilon$ in denominators in \eqref{z34_exp_form}, we make the following variable change (analogously to \eqref{zamina1})
\begin{equation}\label{zamina_2}
\begin{aligned}[c]
{{q}^{0}}= \chi + \varepsilon E
\end{aligned}
\qquad \qquad
\begin{aligned}[c]
{{q}^{1}}=\varepsilon x,
\end{aligned}
\end{equation}
\noindent
so after passing to the limit when $\varepsilon \to 0$, the right-hand side of the expression \eqref{z34_exp_form} takes the form
\begin{equation}\label{z3pm_t1}
\begin{split}
\frac{1}{z_3^\pm - i\varepsilon}\bigg|_{\varepsilon = 0}
= &
\frac{1}{\sqrt{\omega_\pm^2 + 1}}
\\
& \times
\exp{
\left(
i \arccos {
\frac{\omega_\pm}{\sqrt{\omega_\pm^2 + 1}}
}
\right)
},
\end{split}
\end{equation}
\noindent
where $\omega_\pm = -2E\chi \pm \sqrt{|t|}x$.
Then we can rewrite the integral \eqref{tensordovika} as
\begin{widetext}
\begin{equation}\label{tensordovika2}
\begin{split}
{{t}_{ab}^{<}}=
2 & \int\limits_{-\infty }^{\infty }{dE}
\int\limits_{-\infty }^{\infty }{dx}
\int\limits_{D^<}{d{{q}^{2}}d{{q}^{3}}}
\left[ \vphantom{\frac{1^1}{1^1}}
{{f}_{ab}}\left( {{q}^{2}},{{q}^{3}} \right)
\frac{1}{\left(z_1^+ z_1^-\right)}
\frac{1}{\sqrt{\omega_-^2 + 1}}
\right.
\\
& \times \left.
\frac{1}{\sqrt{ \omega_+^2 + 1}}
\exp{
\left( i \arccos {
\frac{\omega_-}{\sqrt{\omega_-^2 + 1}}
}
+
i \arccos {
\frac{\omega_+}{\sqrt{\omega_+^2 + 1}}
}
\right)
}
\right],
\end{split}
\end{equation}
\end{widetext}
\noindent
where $D^< = \lbrace{ \left(q^2, q^3\right) \in R^2 \; \vert \; {{\left( {{q}^{2}} \right)}^{2}}+{{\left( {{q}^{3}} \right)}^{2}}<{c}^{2}{{P}^{2}} \rbrace}$ and ${{f}_{ab}}\left( {{q}^{2}},{{q}^{3}} \right)$ denotes ${{f}_{ab}}\left(q^0 = \chi, q^1 = 0, {{q}^{2}},{{q}^{3}} \right)$.
According to Laplace's method \cite{DeBruijn:225131}, we transform the square roots in the denominators in \eqref{tensordovika2} containing $\omega_\pm$ as
\begin{equation}\label{omega_laplace1}
\begin{split}
\frac{1}{\sqrt{ \omega_+^2 + 1}\sqrt{ \omega_-^2 + 1}} =
\exp
&
\left[
-\frac{1}{2}\log
\left(
\omega_+^2 + 1
\right)
\right.
\\
&
\left.
-\frac{1}{2}\log
\left(
\omega_-^2 + 1
\right)
\right],
\end{split}
\end{equation}
\noindent
and take the second order Taylor approximation of the exponent at $E = 0, \; x = 0$
\begin{equation}\label{omega_laplace2}
\frac{1}{\sqrt{ \omega_+^2 + 1}\sqrt{ \omega_-^2 + 1}}
\approx
\exp \left( -4{{E}^{2}} \chi ^ 2 -\left| t \right|{{x}^{2}} \right).
\end{equation}
\noindent
Similarly, for the other two terms in these denominators
\begin{equation}\label{exp_omega_laplace1}
\begin{split}
\exp{
\left( i \arccos {
\frac{\omega_-}{\sqrt{\omega_-^2 + 1}}
}
+
i \arccos {
\frac{\omega_+}{\sqrt{\omega_+^2 + 1}}
}
\right)
}
\approx
\\
\approx
-
\exp{
\left(
-4i
E\chi
\right).
}
\end{split}
\end{equation}
Substituting \eqref{omega_laplace2} and \eqref{exp_omega_laplace1} into \eqref{tensordovika2}, the integration with respect to $E$ and $x$ is now reduced to the Poisson integrals
\begin{equation}\label{puason1}
\begin{split}
&\int\limits_{-\infty }^{\infty }{dx}\exp \left( -\left| t \right|{{x}^{2}} \right)=\frac{\sqrt{\pi }}{\sqrt{\left| t \right|}}
\\
& \int\limits_{-\infty }^{\infty }{dE}\exp \left( -4{{E}^{2}}\chi^2 + 4iE\chi \right) =
\frac{ \sqrt{\pi } } { 2 e \chi }.
\end{split}
\end{equation}
After these transformations we can rewrite the expression \eqref{tensordovika2} as
\begin{equation}\label{tensordovika3}
\begin{split}
{{t}_{ab}^<}
\approx &
-\frac{\pi }{2e\sqrt{\left| t \right|}}
\int\limits_{D^<}{d{{q}^{2}}d{{q}^{3}}}
\left[
\frac{{{f}_{ab}}\left( {{q}^{2}},{{q}^{3}} \right)}{\chi}
\right.
\\
& \times
\left.
\frac{1}{{{\left( M_{G}^{2}+\frac{\left| t \right|}{2}-{{q}^{3}}\sqrt{4{{P}^{2}}-\left| t \right|} \right)}^{2}}-s\chi^2}
\right].
\end{split}
\end{equation}
It is convenient to use the polar coordinates for further calculations
\begin{equation}\label{polyarnye}
\begin{split}
& {{q}^{2}}=q\cos \left( \alpha \right), \\
& {{q}^{3}}=q\sin \left( \alpha \right). \\
\end{split}
\end{equation}
Then the expression \eqref{tensordovika3} takes the form
\begin{widetext}
\begin{equation}\label{tensordovika4}
\begin{split}
& {{t}_{ab}^{<}}=-\frac{\pi }{2e\sqrt{\left| t \right|}}\int\limits_{0}^{cP}{qdq\frac{1}{\sqrt{M_{G}^{2}+\frac{\left| t \right|}{4}+{{q}^{2}}}}}\times \\
& \times \int\limits_{0}^{2\pi }{d\alpha }\frac{{{f}_{ab}\left( q,\alpha \right)}}{{{\left( M_{G}^{2}+\frac{\left| t \right|}{2}-q\sin \left( \alpha \right)\sqrt{4{{P}^{2}}-\left| t \right|} \right)}^{2}}-s\left( M_{G}^{2}+\frac{\left| t \right|}{4}+{{q}^{2}} \right)}.
\end{split}
\end{equation}
\end{widetext}
As a result, we obtain the two-dimensional integral with respect to the variables $q$ and $\alpha$. This integral can be calculated numerically and allows us to calculate each component of the tensor ${{t}_{ab}}$ separately.
Let us get back to the expression \eqref{tensorposlevika}. Applying Wick rotation, we change the variable ${{q}^{0}}\to i{{q}^{0}}$ in the integral \eqref{tensorposlevika}, so it can be rewritten as
\begin{equation}\label{tensorposlevika1}
\begin{split}
{{t}_{ab}^{>}}=
i
\int\limits_{-\infty }^{\infty }{d{{q}^{0}}}
\int\limits_{-\infty }^{\infty }{d{{q}^{2}}}
\int\limits_{D^>}{d{{q}^{1}}}d{{q}^{3}}
\left[
\vphantom{\frac{1^1}{1^1}}
{{f}_{ab}}\left( {{q}^{0}},{{q}^{1}},{{q}^{2}},{{q}^{3}} \right)
\right.
\\
\left.
\times
\frac{1}{\left(z_1 - i\varepsilon\right)}
\frac{1}{\left(z_2 - i\varepsilon\right)}
\frac{1}{\left(z_3 - i\varepsilon\right)}
\frac{1}{\left(z_4 - i\varepsilon\right)}
\right]
,
\end{split}
\end{equation}
where $D^> = \lbrace{ \left(q^1, q^3\right) \in R^2 \; \vert \; {{\left( {{q}^{1}} \right)}^{2}}+{{\left( {{q}^{3}} \right)}^{2}}>{c}^{2}{{P}^{2}} \rbrace}$.
Note that $\varepsilon $ in denominators can be set to zero now, because the integration is performed along the imaginary axis and we obtain the convergent integral. Rewriting the denominators in \eqref{tensorposlevika1}, we obtain
%
\begin{equation}\label{tensorposlevika2}
\begin{split}
{{t}_{ab}^{>}}=
i
& \int\limits_{-\infty }^{\infty }{d{{q}^{0}}}
\int\limits_{-\infty }^{\infty }{d{{q}^{2}}}
\int\limits_{D^>}{d{{q}^{1}}}d{{q}^{3}}
\left[
\vphantom{\frac{1^1}{1^1}}
{{f}_{ab}}\left( {{q}^{0}},{{q}^{1}},{{q}^{2}},{{q}^{3}} \right)
\right.
\\
& \times
\frac{1}{{{\left(M_{p}^{2} + \xi^2 + {{\left(\frac{\sqrt{4P^2 - |t|}}{2} + {{q}^{3}} \right)}^{2}}-\frac{s}{4} \right)}^{2}}+\left(q^0\right)^2s} \\\
& \left.
\times
\frac{1}{{{\left( M_{G}^{2}+ \frac{|t|}{4}+ \xi^2 +{{\left( {{q}^{3}} \right)}^{2}} \right)}^{2}}-{{\left( {{q}^{1}} \right)}^{2}} |t|}
\right]
,
\end{split}
\end{equation}
\noindent
where $\xi^2 = {{\left( {{q}^{0}} \right)}^{2}} + {{\left( {{q}^{1}} \right)}^{2}} + {{\left( {{q}^{2}} \right)}^{2}}$.
It can be seen that the product of two fractions in \eqref{tensorposlevika2} is even function with regard to $q^0$ and $q^2$ and
has the maximum at ${{q}^{0}}=0, {{q}^{2}}=0$. Increasing the absolute values of ${{q}^{0}}$ and ${{q}^{2}}$ results in increasing the denominators, i.e. in decreasing the whole fraction. Therefore we return back to \eqref{tensorposlevika1} and separate the integration with respect to ${{q}^{0}}, {{q}^{2}}$ and ${{q}^{1}}, {{q}^{3}}$. For the convenience and simplicity of the further calculations, we introduce the following polar coordinates
\begin{equation}\label{polyarnye1}
\begin{aligned}[c]
{{q}^{1}}=q\cos \left( \alpha \right)
\end{aligned}
\quad \quad
\begin{aligned}[c]
{{q}^{3}}=q\sin \left( \alpha \right).
\end{aligned}
\end{equation}
\noindent
The expression \eqref{tensorposlevika1} then can be rewritten as
\begin{equation}\label{tensorposlevika3}
\begin{split}
{{t}_{ab}^{>}}=i\int\limits_{cP}^{\infty }{qdq}\int\limits_{0}^{2\pi }{d\alpha }\int\limits_{-\infty }^{\infty }{d{{q}^{0}}}
& \int\limits_{-\infty }^{\infty }{d{{q}^{2}}}
\left[ \vphantom{\frac{1}{1}}
{{f}_{ab}}\left( q,\alpha ,{{q}^{0}},{{q}^{2}} \right)
\right.
\\
& \times
\left.
\frac{1}{z_1^+z_1^-}
\frac{1}{z_3^+z_3^-}
\right],
\end{split}
\end{equation}
\noindent
where
\begin{equation}\label{z1z3_wick}
\begin{split}
z_1^\pm = M_{p}^{2}+
&
{{\left( {{q}^{0}} \mp i\frac{\sqrt{s}}{2} \right)}^{2}}+{{q}^{2}}{{\cos }^{2}}\left( \alpha \right)+{{\left( {{q}^{2}} \right)}^{2}}
\\
&
\hphantom{{{\left( 1 \right)}^{2}}}
+{{\left( \frac{\sqrt{4P^2 - |t|}}{2}+q\sin \left( \alpha \right) \right)}^{2}},
\\
z_3^\pm = M_{G}^{2}+
&
{{\left( {{q}^{0}} \right)}^{2}}+{{\left( \frac{\sqrt{|t|}}{2}\mp q\cos \left( \alpha \right) \right)}^{2}}
\\
&
\hphantom{{{\left({{q}^{0}} {{q}^{0}} \right)}^{2}}}
+{{\left( {{q}^{2}} \right)}^{2}}+{{q}^{2}}{{\sin }^{2}}\left( \alpha \right).
\end{split}
\end{equation}
The expression \eqref{tensorposlevika3} contains the integration with respect to $q$ from $cP$ to $+\infty$. We can change the variable $q \to x = 1 / q$ in order to make the integration limits finite. At this point the integrand in \eqref{tensorposlevika3} becomes large enough if we substitute the expressions for \eqref{z1z3_wick} for $z_1^\pm$ and $z_3^\pm$ which depend on every integration variable $x, \alpha, q^0, q^2$. However, introducing the following notations
\begin{equation}\label{omega_xalpha}
\begin{split}
& {{w}_{1}}\left( x,\alpha \right)=1-{{x}^{2}}\frac{\left| t \right|}{4}+x\sqrt{4{{P}^{2}}-\left| t \right|}\sin \left( \alpha \right) \\
& {{w}_{2}}\left( x,\alpha \right)={{x}^{2}}M_{G}^{2}+{{x}^{2}}\frac{\left| t \right|}{4}-x\sqrt{\left| t \right|}\cos \left( \alpha \right)+1 \\
& {{w}_{3}}\left( x,\alpha \right)={{x}^{2}}M_{G}^{2}+{{x}^{2}}\frac{\left| t \right|}{4}+x\sqrt{\left| t \right|}\cos \left( \alpha \right)+1,
\end{split}
\end{equation}
\noindent
and changing the integration variables again $q^0 \to E = x{{q}^{0}}$ and $q^2 \to y = x{{q}^{2}}$, we can finally rewrite \eqref{tensorposlevika3} using notations \eqref{omega_xalpha}
\begin{equation}\label{tensorposlevika6}
\begin{split}
{{t}_{ab}^{>}}=&\hphantom{a}i
\int\limits_{0}^{\frac{1}{cP}}{xdx}
\int\limits_{0}^{2\pi }{d\alpha }
\int\limits_{-\infty }^{\infty }{dE}
\int\limits_{-\infty }^{\infty }{dy}
\left[\vphantom{\frac{1^1}{1^1}}
{{f}_{ab}}\left( x,E,\alpha ,y \right)
\right.
\\
& \times
\frac{1}{{{\left( {{w}_{1}}\left( x,\alpha \right) \right)}^{2}}{{w}_{2}}\left( x,\alpha \right){{w}_{3}}\left( x,\alpha \right)}
\\
& \times
\left.
\frac{1}{1+{{W}_{1}}\left( x,\alpha \right)\left( {{E}^{2}}+{{y}^{2}} \right)+{{W}_{2}}\left( x,\alpha \right){{E}^{2}}}
\right],
\end{split}
\end{equation}
\noindent
where
\begin{equation}\label{poznachenya2}
\begin{split}
& {{W}_{1}}\left( x,\alpha \right)=\frac{2}{{{w}_{1}}\left( x,\alpha \right)}+\frac{1}{{{w}_{2}}\left( x,\alpha \right)}+\frac{1}{{{w}_{3}}\left( x,\alpha \right)}, \\
& {{W}_{2}}\left( x,\alpha \right)=\frac{{{x}^{2}}s}{{{\left( {{w}_{1}}\left( x,\alpha \right) \right)}^{2}}}.
\end{split}
\end{equation}
The further calculation of \eqref{tensorposlevika6} depends on the values of indices $a$ and $b$. We demonstrate the calculation workflow for the component $t_{00}^{>}$ only, while the calculation of other components is completely similar. In this case ($a = 0, \; b = 0$) the function $f_{ab}$ in the integrand of \eqref{tensorposlevika6} is equal to ${{f}_{00}} = \left( {{E}^{2}}+{{x}^{2}}s \right)$, so the expression for $t_{00}^>$ has the form
\begin{widetext}
\begin{equation}\label{tensorposlevika7}
\begin{split}
& {{t}_{00}^{>}}=
i
\int\limits_{0}^{\frac{1}{cP}}{xdx}
\int\limits_{0}^{2\pi }{d\alpha }
\left[\vphantom{\int\limits_a^b}
\frac{1}{{{\left( {{w}_{1}}\left( x,\alpha \right) \right)}^{2}}{{w}_{2}}\left( x,\alpha \right){{w}_{3}}\left( x,\alpha \right)}
\right.
\\ & \times
\left.
\int\limits_{-\infty }^{\infty }{dE}\int\limits_{-\infty }^{\infty }{dy}\left( {{E}^{2}}+{{x}^{2}}s \right)\frac{1}{1+{{W}_{1}}\left( x,\alpha \right)\left( {{E}^{2}}+{{y}^{2}} \right)+{{W}_{2}}\left( x,\alpha \right){{E}^{2}}}
\right]
.
\end{split}
\end{equation}
\noindent
The integration with respect to $E$ and $y$ can be reduced to the calculation of the Poisson integrals
\begin{equation}\label{puason}
\begin{split}
& \int\limits_{-\infty }^{\infty }{dE}\int\limits_{-\infty }^{\infty }{dy}\left( {{E}^{2}}+{{x}^{2}}s \right)\frac{1}{1+{{W}_{1}}\left( x,\alpha \right)\left( {{E}^{2}}+{{y}^{2}} \right)+{{W}_{2}}\left( x,\alpha \right){{E}^{2}}}= \\
& =\frac{\pi }{\sqrt{W\left( x,\alpha \right)}\sqrt{{{W}_{1}}\left( x,\alpha \right)}}\left( \frac{1}{2}\frac{1}{\left| W\left( x,\alpha \right) \right|}+{{x}^{2}}s \right),
\end{split}
\end{equation}
\noindent
where $W\left( x,\alpha \right) = {{W}_{1}}\left( x,\alpha \right)+{{W}_{2}}\left( x,\alpha \right)$. Substituting \eqref{puason} into \eqref{tensorposlevika7} yields
\begin{equation}\label{tensorposlevika8}
\begin{split}
& {{t}_{00}}^{>}=\pi i\int\limits_{0}^{\frac{1}{cP}}{xdx}\int\limits_{0}^{2\pi }{d\alpha }\frac{1}{{{\left( {{w}_{1}}\left( x,\alpha \right) \right)}^{2}}{{w}_{2}}\left( x,\alpha \right){{w}_{3}}\left( x,\alpha \right)}\times \\
& \times \frac{1}{\sqrt{{{W}_{3}}\left( x,\alpha \right)}\sqrt{{{W}_{1}}\left( x,\alpha \right)}}\left( \frac{1}{2}\frac{1}{\left| {{W}_{3}}\left( x,\alpha \right) \right|}+{{x}^{2}}s \right).
\end{split}
\end{equation}
\end{widetext}
Finally, we obtain the two-dimensional integral with finite limits which can be calculated numerically. The other non zero components of $t_{ab}^{>}$ have to be calculated in the same way.
After the components $t^<_{ab}$ and $t^>_{ab}$ get calculated, we can finally calculate their sum $t_{ab} = t^<_{ab} + t^>_{ab}$ and return to the analytical expression \eqref{vyraz_diagramma33} for the diagram \fref{fig:simplest_loops}a.
So far, we have calculated the analytical expression for the simplest one-loop diagram shown in \fref{fig:simplest_loops}a. However, we should also consider all possible contractions of bispinor and glueball field operators. In the simplest case of the elastic scattering of two protons there is another diagram \fref{fig:simplest_loops}b added to the one-loop diagram \fref{fig:simplest_loops}a. The analytical expression for this diagram is pretty similar to \eqref{vyraz_diagramma33} and the integral in this expression can be calculated in the same way as for $t_{ab}$.
\section{Results and discussion}
The analytic calculations presented in the previous sections allowed us to calculate the differential cross section of elastic proton-proton scattering $d{{\sigma }_{el}}/dt\left( t \right)$. This calculation includes the contributions from the tree-level (pole) (\fref{fig:polusnaya}) and one-loop (\fref{fig:simplest_loops}) diagrams, as well as contributions from the diagrams with the ${{P}_{3}}$ and ${{P}_{4}}$ interchanged. The model used for calculation contains two parameters -- ${{M}_{G}}$ -- the mass of glueball, and $G$ -- the effective coupling of proton-glueball interaction. All the quantities were expressed in the units of the proton mass ${{M}_{P}} = 0.938$ GeV. The obtained dependency $d\sigma_{el}/dt\left( t \right)$ is presented in Fig.\ref{fig:dif_sech23}-\ref{fig:dif_sech_t} at different energies $\sqrt{s}$ and different values of $M_G$ and $G$.
\begin{figure*}
\centering
\includegraphics[width=0.74\linewidth]{figs/diff_cs_23d4GeV.pdf}
\caption[Рис.\ref{fig:dif_sech23}]{
The differential cross-section $d\sigma_{el}/dt\left( t \right)$ of elastic $pp$ scattering at $\sqrt{s}=23.4$ GeV. The solid curve is the calculated dependency ($M_G=0.116$ and $G = 1.8$), the dots with error bars are experimental data \cite{NAGY1979221}.}
\label{fig:dif_sech23}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.74\linewidth]{figs/diff_cs_30d5GeV.pdf}
\caption[Рис.\ref{fig:dif_sech30}]{
The differential cross-section $d\sigma_{el}/dt\left( t \right)$ of elastic $pp$ scattering at $\sqrt{s}=30.5$ GeV. The solid curve is the calculated dependency ($M_G=0.13$ and $G = 2.2$), the dots with error bars are experimental data \cite{NAGY1979221}.}
\label{fig:dif_sech30}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.74\linewidth]{figs/diff_cs_44d6GeV.pdf}
\caption[Рис.\ref{fig:dif_sech44}]{The differential cross-section $d\sigma_{el}/dt\left( t \right)$ of elastic $pp$ scattering at $\sqrt{s}=44.6$ GeV. The solid curve is the calculated dependency ($M_G=0.099$ and $G = 1.8$), the dots with error bars are experimental data \cite{NAGY1979221}.}
\label{fig:dif_sech44}
\end{figure*}
As can be seen in Fig.\ref{fig:dif_sech23}, the obtained dependency is non-monotonic and qualitatively describes the first minimum of the experimental curve. The values of $M$ and $G$ were chosen in each case (energy) differently to better reproduce the experiment. However, the observed non-monotonic behavior of the theoretical dependency is preserved in all three cases even for the same values of $M$ and $G$.
Although the results obtained are only in qualitative agreement with the experimental data, we expect that inclusion of the diagrams with the higher number of loops may help to achieve the quantitative description of the experiment.
In order to demonstrate the qualitative behavior of the dependency $d\sigma_{el}/dt\left( t \right)$ (including the second fall), we plotted it separately in a wider range of $t$ up to $30$ GeV$^2$ (\fref{fig:dif_sech_t}).
\begin{figure*}
\centering
\includegraphics[width=0.74\linewidth]{figs/plot_theory_only.pdf}
\caption{
The calculated dependency of differential cross-section $d\sigma_{el}/dt\left( t \right)$ of elastic $pp$ scattering at $\sqrt{s}=44.6$ GeV ($M_G=0.088$ and $G = 1.8$). }
\label{fig:dif_sech_t}
\end{figure*}
We also performed the analogous calculations within another model $\phi^3$ (phi-cubed) with the scalar field. However, within that model we have not obtained the effects of non-monotonic $d\sigma_{el}/dt\left( t \right)$ dependencies similar to the presented here. It suggests that the physical mechanisms responsible for this non-monotonicity are associated with the spin effects. The spin flip effects observed in the protons scattering have been calculated within the Regge theory \cite{Bence:2020usl}. However, our results are different in that within the Regge approach this non-monotonicity arises due to an exchange of reggeons having different signatures, i.e. the reggeon contributions to the scattering amplitude with different signs. In \cite{Bence:2020usl} the spin effects provides only quantitative changes in the behavior of differential cross-section dependency $d\sigma_{el}/dt\left( t \right)$. In our approach it is the spin effects that are responsible for the appearance of non-monotonicity in the differential cross-section dependency $d\sigma_{el}/dt\left( t \right)$ .
\section{Summary}
We demonstrated that the multi-particle fields approach can be used for a development of dynamic models which describe the scattering of multi-quark systems. Using this approach, we considered the elastic proton-proton scattering and calculated the differential cross-section. As a result, we succeeded to obtain the qualitative agreement with the experiment.
Laplace's method allowed us to approximately calculate the loop diagrams. Unfortunately, the calculation is lengthy enough and requires a further improvement.
We found that the experimentally observed non-monotonicity of the differential cross-section dependency $d\sigma_{el}/dt\left( t \right)$ is caused in our model by the spin effects. Taking these effects into account in the tree-level and simplest loop diagrams allowed us to obtain the qualitative description of experiment. The obtained results suggest that inclusion of the more complex loop diagrams to the calculation will help achieve the quantitative description of experimental data.
\section{Laplace's method and passage to the limit $\varepsilon \to 0$}
An analytical expression for the diagram in \fref{fig:simplest_loops}a has the following form
\begin{equation}\label{vyraz_diagramma33}
\begin{split}
A=\frac{{{\left( ig \right)}^{4}}}{{{\left( 2\pi \right)}^{6}}}&{{\left( \bar{v}_{{{v}_{3}}}^{+}\left( {{P}_{4}} \right) \right)}_{{{s}_{4}}}}\gamma _{{{s}_{4}}{{s}_{2}}}^{a}{{\left( v_{{{v}_{2}}}^{-}\left( {{P}_{2}} \right) \right)}_{{{s}_{2}}}} \\
\times &{{\left( \bar{v}_{{{v}_{4}}}^{+}\left( {{P}_{3}} \right) \right)}_{{{s}_{3}}}}\gamma _{{{s}_{3}}{{s}_{1}}}^{b}{{\left( v_{{{v}_{1}}}^{-}\left( {{P}_{1}} \right) \right)}_{{{s}_{1}}}}\\
\times & \delta \left( \left( {{P}_{3}}+{{P}_{4}} \right)-\left( {{P}_{1}}+{{P}_{2}} \right) \right) t_{ab},
\end{split}
\end{equation}
\noindent
where $g$ is the effective coupling, ${{M}_{p}}$ and ${{M}_{G}}$ are the masses of the proton and the glueball respectively,
${{\left( \bar{v}_{{{v}_{3}}}^{+}\left( {{P}_{4}} \right) \right)}_{{{s}_{4}}}},{{\left( v_{{{v}_{2}}}^{-}\left( {{P}_{2}} \right) \right)}_{{{s}_{2}}}}, {{\left( \bar{v}_{{{v}_{4}}}^{+}\left( {{P}_{3}} \right) \right)}_{{{s}_{3}}}}, {{\left( v_{{{v}_{1}}}^{-}\left( {{P}_{1}} \right) \right)}_{{{s}_{1}}}}$ - are the solutions of the Dirac equations, $\gamma _{{{s}_{4}}{{s}_{2}}}^{a},\gamma _{{{s}_{3}}{{s}_{1}}}^{b}$ - are the elements of the Dirac matrices, and $t_{ab}$ - is the tensor whose components are defined by
\begin{equation}\label{tensor_ab}
\begin{split}
& {{t}_{ab}}=\int{{{d}^{4}}}k\left( {{k}_{a}}+2{\left(P_2\right)_a} \right)\left( 2{\left(P_1\right)_b}-{{k}_{b}} \right)
\\
& \times \frac{1}{M_{p}^{2}-{{\left( {{P}_{1}}-k \right)}^{2}}-i\varepsilon }\frac{1}{M_{p}^{2}-{{\left( {{P}_{2}}+k \right)}^{2}}-i\varepsilon } \\
& \times \frac{1}{M_{G}^{2}-{{k}^{2}}-i\varepsilon }\frac{1}{M_{G}^{2}-{{\left( {{P}_{1}}-{{P}_{3}}-k \right)}^{2}}-i\varepsilon },
\end{split}
\end{equation}
\noindent
where $a, b = 0 .. 3$.
We consider the problem in the center of mass reference frame with the right hand coordinate system whose $z$ axis is oriented along ${{\vec{P}}_{1}}$, and the $x$ axis is perpendicular to $z$ and lies the plane containing vectors ${{\vec{P}}_{1}}$ and ${{\vec{P}}_{3}}$. All quantities are expressed in the units of the proton mass, ${{M}_{G}}$ and $g$ are considered as the model parameters.
First, we need to calculate the components of the tensor ${{t}_{ab}}$. After the tensor components $t_{ab}$ have been calculated, we use the squared absolute value of the scattering amplitude to calculate the elastic scattering differential cross-section $d\sigma/dt$ taking into account the contributions from the pole diagram from \fref{fig:polusnaya}.
To calculate the integral in \eqref{tensor_ab} we use Laplace's method. According to the calculation method described in the previous section, for a small value of $\varepsilon$ we determine which region makes the major contribution to the integral. If the $\varepsilon$ is small and non-zero, the absolute value of the integrand reaches its maximum in a region where the maximal number of real parts of the denominators (i.e. the parts that do not contain $\varepsilon$) turn to zero. Let's denote the real parts of the denominators in \eqref{tensor_ab} as
\begin{equation}\label{znamenyky}
\begin{aligned}[c]
& {{z}_{1}}=M_{p}^{2}-{{\left( {{P}_{1}}-k \right)}^{2}} \\
& {{z}_{2}}=M_{p}^{2}-{{\left( {{P}_{2}}+k \right)}^{2}} \\
& {{z}_{3}}=M_{G}^{2}-{{k}^{2}} \\
& {{z}_{4}}=M_{G}^{2}-{{\left( {{P}_{1}}-{{P}_{3}}-k \right)}^{2}}.
\end{aligned}
\end{equation}
Taking into account \eqref{znamenyky}, expression \eqref{tensor_ab} can be rewritten as
\begin{equation}\label{tensor_ab_short}
\begin{split}
& {{t}_{ab}}=\int{{{d}^{4}}}k
{{f}_{ab}}\left( {{k}_{0}},\vec{k} \right)
\prod_{j=1}^4\frac{1}{\left(z_j - i\varepsilon\right)}.
\end{split}
\end{equation}
The squared absolute value of the denominator in \eqref{tensor_ab} is the sum of the real part squared and the $\varepsilon^2$. When the real part in \eqref{znamenyky} is equal to zero and $\varepsilon$ is non-zero, then the integral converges and has the maximum. Assume that one of the denominators \eqref{znamenyky} is zero, which yields some subset of the integration domain where the real part of the denominator is zero. Again, considering the integration over this subset, the main contribution will be provided by that region where the real parts of some of the other denominators turn to zero. So a natural question to ask is how many expressions in \eqref{znamenyky} can be turned to zero at the same time? It has been shown \cite{sharph2020laplace} that either the first pair of expressions ${{z}_{1}}$ and ${{z}_{2}}$ corresponding to the
horizontal lines in Feynman diagram (\fref{fig:simplest_loops}a), or another pair ${{z}_{3}}$ and ${{z}_{4}}$ corresponding to the vertical lines, can be turned to zero simultaneously. Thus, both the horizontal and vertical lines cannot be turned to zero at the same time. Next we present the calculations of each denominator in more detail.
Let us consider the tensor \eqref{tensor_ab} and denote the numerator as ${{f}_{ab}}\left( {{k}_{0}},\vec{k} \right)=\left( {{k}_{a}}+2{{P}_{2a}} \right)\left( 2{{P}_{1b}}-{{k}_{b}} \right)$. Taking into account that in the center of mass reference frame ${{\vec{P}}_{2}}=-{{\vec{P}}_{1}}$, the denominator $z_2$ in the tensor \eqref{tensor_ab} may be represented as follows
\begin{equation}\label{z2_cm}
{{z}_{2}}= M_{p}^{2}-{{\left( \frac{\sqrt{s}}{2}-{{k}^{0}} \right)}^{2}}+{{\left( \vec{k}-{{{\vec{P}}}_{1}} \right)}^{2}}.
\end{equation}
Tensor \eqref{tensor_ab_short} contains now 7 non-zero terms, and each of them has to be calculated separately. Note that they differ only in numerators and have the same denominators, which means they all can be calculated in the same way. It is then convenient to change the coordinate system as shown in \fref{fig:osi_1}.
\begin{figure}[tbp!]
\centering
\includegraphics[width=0.95\linewidth]{figs/k_rotation.pdf}
\caption{
The coordinates ${{k}^{1}}$ and ${{k}^{3}}$ of the vector $\vec{k}$ in initial coodrinate system and its coordinates ${{{k}'}^{1}}$ and ${{{k}'}^{3}}$ after transformation of the coordinate system.}
\label{fig:osi_1}
\end{figure}
The vector $\vec{k}$ has the coordinates ${{k}^{1}}$ along $x$ axis and ${{k}^{3}}$ along $z$ axis. We transform the axes so that ${x}'$ becomes parallel to the vector ${{\vec{P}}_{1}}-{{\vec{P}}_{3}}$. The coordinates of vector $\vec{k}$ expressed through the new coordinates ${{{k}'}^{1}}$ and ${{{k}'}^{3}}$ in the transformed system are
\begin{equation}\label{k1k3}
\begin{split}
& {{k}^{1}}=-{{{{k}'}}^{1}}\cos \left( \frac{\theta }{2} \right)+{{{{k}'}}^{3}}\sin \left( \frac{\theta }{2} \right), \\
& {{k}^{3}}={{{{k}'}}^{1}}\sin \left( \frac{\theta }{2} \right)+{{{{k}'}}^{3}}\cos \left( \frac{\theta }{2} \right). \\
\end{split}
\end{equation}
The numerator ${{f}_{ab}}\left( {{k}_{0}},\vec{k} \right)=\left( {{k}_{a}}+2{{P}_{2a}} \right)\left( 2{{P}_{1b}}-{{k}_{b}} \right)$ in \eqref{tensor_ab_short} is also expressed through the coordinates \eqref{k1k3}. For a further calculation of \eqref{tensor_ab_short} we also change the variables as follows
\begin{equation}\label{zamina}
\begin{aligned}[c]
{{q}^{0}} = {{k}^{0}}
\end{aligned}
\qquad \qquad
\begin{aligned}[c]
\vec{q} = \frac{{{{\vec{P}}}_{1}}-{{{\vec{P}}}_{3}}}{2}-\vec{k},
\end{aligned}
\end{equation}
\noindent
so the tensor \eqref{tensor_ab_short} can be rewritten as
\begin{equation}\label{tensor_t_through_q}
\begin{split}
& {{t}_{ab}}=\int\limits_{-\infty }^{\infty }{d{{q}^{0}}}\int{d\vec{q}}{{f}_{ab}}\left( {{q}^{0}},\vec{q} \right)
\prod_{j=1}^4\frac{1}{\left(z_j - i\varepsilon\right)},
\end{split}
\end{equation}
where the denominators \eqref{znamenyky} take the following form
\begin{widetext}
\begin{equation}\label{z_through_q}
\begin{split}
& z_1 = M_{p}^{2}-{{\left( \frac{\sqrt{s}}{2}-{{q}^{0}} \right)}^{2}}+{{\left( {{q}^{1}} \right)}^{2}}+{{\left( {{q}^{2}} \right)}^{2}}+{{\left( \frac{\left| {{{\vec{P}}}_{1}}+{{{\vec{P}}}_{3}} \right|}{2}+{{q}^{3}} \right)}^{2}} \\
& z_2 = M_{p}^{2}-{{\left( \frac{\sqrt{s}}{2}+{{q}^{0}} \right)}^{2}}+{{\left( {{q}^{1}} \right)}^{2}}+{{\left( {{q}^{2}} \right)}^{2}}+{{\left( \frac{\left| {{{\vec{P}}}_{1}}+{{{\vec{P}}}_{3}} \right|}{2}+{{q}^{3}} \right)}^{2}} \\
& z_3 = M_{G}^{2}-{{\left( {{q}^{0}} \right)}^{2}}+{{\left( \frac{\left| {{{\vec{P}}}_{1}}-{{{\vec{P}}}_{3}} \right|}{2}-{{q}^{1}} \right)}^{2}}+{{\left( {{q}^{2}} \right)}^{2}}+{{\left( {{q}^{3}} \right)}^{2}} \\
& z_4 = M_{G}^{2}-{{\left( {{q}^{0}} \right)}^{2}}+{{\left( \frac{\left| {{{\vec{P}}}_{1}}-{{{\vec{P}}}_{3}} \right|}{2}+{{q}^{1}} \right)}^{2}}+{{\left( {{q}^{2}} \right)}^{2}}+{{\left( {{q}^{3}} \right)}^{2}}.
\end{split}
\end{equation}
\end{widetext}
Note that $z_1$ and $z_2$ differ only in sign before $q^0$. At the same time the only difference between $z_3$ and $z_4$ is the sign before $q^1$. It allows us to shorten the calculations by introducing the following notations
\begin{equation}\label{zpm}
\begin{split}
& z_1^{\pm}=
M_{p}^{2}-{{\left( \frac{\sqrt{s}}{2}\pm{{q}^{0}} \right)}^{2}}
+{{\left( {{q}^{1}} \right)}^{2}}
+{{\left( {{q}^{2}} \right)}^{2}}
\\
& \hphantom{\hspace{3cm}}
+{{\left( \frac{\left| {{{\vec{P}}}_{1}}+{{{\vec{P}}}_{3}} \right|}{2}+{{q}^{3}} \right)}^{2}}
\\
& z_3^{\pm} =
M_{G}^{2}-{{\left( {{q}^{0}} \right)}^{2}}
+{{\left( \frac{\left| {{{\vec{P}}}_{1}}-{{{\vec{P}}}_{3}} \right|}{2} \pm {{q}^{1}} \right)}^{2}}
\\
& \hphantom{\hspace{4cm}}
+{{\left( {{q}^{2}} \right)}^{2}}+{{\left( {{q}^{3}} \right)}^{2}},
\end{split}
\end{equation}
\noindent
so that $z_1 = z_1^-$, $z_2 = z_1^+$, and $z_3 = z_3^-$, $z_4 = z_3^+$.
Since the differential cross-section depends on the transmitted four-momentum $t$, we express the vectors ${{\vec{P}}_{1}}$ and ${{\vec{P}}_{3}}$ through Mandelstam variable $t$
\begin{equation}\label{cherez_t}
\begin{split}
& \left| {{{\vec{P}}}_{1}}-{{{\vec{P}}}_{3}} \right|=\sqrt{\left| t \right|}
\\
& \left| {{{\vec{P}}}_{1}}+{{{\vec{P}}}_{3}} \right|=\sqrt{4{{P}^{2}}-\left| t \right|},
\end{split}
\end{equation}
\noindent
where $P = | \vec{P}_1 | = | \vec{P}_3 |$.
Applying one more change of variable $${{q}^{3}} \rightarrow \frac{1}{2}\sqrt{4{{P}^{2}}-\left| t \right|}+{{q}^{3}}$$ and taking into account \eqref{cherez_t}, we obtain the new expressions for \eqref{zpm}:
\begin{equation}\label{zpm_after_q3_change}
\begin{split}
& z_1^{\pm} = M_{p}^{2}-{{\left( \frac{\sqrt{s}}{2} \pm {{q}^{0}} \right)}^{2}}+{{\left( {{q}^{1}} \right)}^{2}}+{{\left( {{q}^{2}} \right)}^{2}}+{{\left( {{q}^{3}} \right)}^{2}}
\\
& z_3^{\pm} = M_{G}^{2}-{{\left( {{q}^{0}} \right)}^{2}}+{{\left( \frac{\sqrt{\left| t \right|}}{2} \pm {{q}^{1}} \right)}^{2}}+{{\left( {{q}^{2}} \right)}^{2}}
\\
&\hphantom{\hspace{3cm}}+ {{\left( {{q}^{3}}-\frac{\sqrt{4{{P}^{2}}-\left| t \right|}}{2} \right)}^{2}}
.
\end{split}
\end{equation}
Finally, introducing the spherical coordinates $\left( q, \theta, \phi \right)$:
\begin{equation}\label{sferychni_koordynaty}
\begin{split}
& {{q}^{1}}=q\sin \left( \theta \right)\cos \left( \varphi \right), \\
& {{q}^{2}}=q\sin \left( \theta \right)\sin \left( \varphi \right), \\
& {{q}^{3}}=q\cos \left( \theta \right),
\end{split}
\end{equation}
\noindent
where $0 \leq \theta \leq \pi$ and $0 \leq \phi < 2\pi$, we can rewrite \eqref{tensor_t_through_q} as follows
\begin{equation}\label{tensor_in_sphere}
\begin{split}
{{t}_{ab}}=&\int\limits_{-\infty }^{\infty }{d{{q}^{0}}}\int\limits_{0}^{\infty }{{{q}^{2}}dq}\int\limits_{0}^{\pi }{\sin \left( \theta \right)}d\theta \int\limits_{0}^{2\pi }{d\varphi }{{f}_{ab}}\left( {{q}^{0}},q,\theta ,\varphi \right) \\
& \times
\frac{1}{\left(z_1^{+} - i\varepsilon\right)}
\frac{1}{\left(z_1^{-} - i\varepsilon\right)}
\frac{1}{\left(z_3^{-} - i\varepsilon\right)}
\frac{1}{\left(z_3^{+} - i\varepsilon\right)},
\end{split}
\end{equation}
\noindent
where
\begin{equation}\label{zpm_in_sphere}
\begin{split}
& z_1^{\pm} = M_{p}^{2}-{{\left( \frac{\sqrt{s}}{2} \pm {{q}^{0}} \right)}^{2}}+{{q}^{2}}
\\
& z_3^{\pm} = M_{G}^{2}-{{\left( {{q}^{0}} \right)}^{2}}
+ {\left(\frac{\sqrt{\left| t \right|}}{2} \mp q\sin \left( \theta \right)\cos \left( \varphi \right)\right)}^{2} \\
& + {\left( q\sin \left( \theta \right)\sin \left( \varphi \right) \right)}^{2}
+ {\left( q\cos \left( \theta \right)-\frac{\sqrt{4{{P}^{2}}-\left| t \right|}}{2} \right)}^{2}
\end{split}
\end{equation}
As mentioned in the problem statement, we are going to use the Laplace's method to calculate the multidimensional integral \eqref{tensor_in_sphere}.
The essential idea behind this method is that the integral whose integrand has a single maximum point in the integration domain, can be approximated nicely by the corresponding Gaussian integral. We have already seen above that either the first pair of denominators (associated with the horizontal lines of the diagram in \fref{fig:simplest_loops}a) or the second pair (two vertical lines) can be turned to zero simultaneously. Thus, we examine the first and the second pairs of denominators separately. We find the corresponding regions of the integration domain where either the first or the second pair of denominators turn into zero, and calculate the contribution of these regions to the integral.
\subsection{Contribution of the first region}
\subsection{The first pair of denominators} \label{first_pair_section}
Let us consider the system of equations for the first two denominators $\left(z_1^{+} - i\varepsilon\right)$ and $\left(z_1^{-} - i\varepsilon\right)$ in \eqref{tensor_in_sphere}. If we set the real parts of these denominators to zero, their product reduces to ${\varepsilon }^{2}$, which provides the maximal contribution to the integrand in \eqref{tensor_in_sphere}.
\begin{equation}\label{systema_rivnan}
\left\{ \begin{split}
& z_1^+ = M_{p}^{2}-{{\left( \frac{\sqrt{s}}{2}+{{q}^{0}} \right)}^{2}}+{{q}^{2}}=0 \\
& z_1^- = M_{p}^{2}-{{\left( \frac{\sqrt{s}}{2}-{{q}^{0}} \right)}^{2}}+{{q}^{2}}=0 \\
\end{split} \right.\
\end{equation}
It is easy to verify that ${{q}^{0}}=0,\;q=P$ satisfies the equation system \eqref{systema_rivnan}. This solution describes the $2$-dimensional sphere and thus provides some subset of the $4$-dimensional integration domain where the absolute value of the first two denominators in \eqref{tensor_in_sphere} has a single minimum.
Next, let us apply Laplace's method to the first pair of denominators
\begin{equation}\label{exponenta_vid_log}
\begin{split}
&\frac{1}{\left(z_1^+ - i\varepsilon\right)\left(z_1^- - i\varepsilon\right)}
=
\\
& =
\exp{\left[-\ln{\left(z_1^+-i\varepsilon\right)} -\ln{\left(z_1^--i\varepsilon\right)} \right]} =
\\
& =
-\frac{1}{\varepsilon^2}
\exp{
\left[
- \ln{\left(1- \frac{z_1^+}{i\varepsilon}\right)}
- \ln{\left(1 - \frac{z_1^-}{i\varepsilon}\right)}
\right].
}
\end{split}
\end{equation}
Considering the second order Taylor approximation of the exponent in \eqref{exponenta_vid_log} and changing the variable $q = P + x$ in \eqref{zpm_in_sphere}, we obtain
\begin{equation}\label{exponenta_vid_log1}
\begin{split}
\frac{1}{\left(z_1^+ - i\varepsilon\right)\left(z_1^- - i\varepsilon\right)}
\approx
-\frac{1}{\varepsilon^2}
\exp \left[
\frac{s{{\left( {{q}^{0}} \right)}^{2}}+4{{P}^{2}}{{x}^{2}}}{-{{\varepsilon }^{2}}} \right.
\\
\hphantom{\hspace{2cm}}
\left.
+\frac{-2{{\left( {{q}^{0}} \right)}^{2}}+4Px+2{{x}^{2}}}{i\varepsilon }
\right].
\end{split}
\end{equation}
The convenience of variable $x = q - P$ is that together with $q^0$ they are the offset from the point $(q^0 = 0,\; q = P)$ where the real parts of denominators in the left-hand side of \eqref{exponenta_vid_log1} take minimum value. Consequently, the absolute value of exponent in \eqref{exponenta_vid_log1} has maximum at $\left(q^0 = 0,\; x = 0 \right)$.
Changing the variable $q \to x = q - P$ in the integral \eqref{tensor_in_sphere} and substituting \eqref{exponenta_vid_log1}, we get the approximation for the integral \eqref{tensor_in_sphere}
\begin{widetext}
\begin{equation}\label{tensor_for_z12}
\begin{split}
{{t}_{ab}} \approx &
-\frac{1}{\varepsilon^2}
\int\limits_{-\infty }^{\infty }{d{{q}^{0}}}
\int\limits_{-P}^{\infty }{{{\left(P + x\right)}^{2}}dx}
\int\limits_{0}^{\pi }{\sin \left( \theta \right)}d\theta
\int\limits_{0}^{2\pi }{d\varphi }
\left[
\vphantom{\frac{1}{\left(z_3^{-} - i\varepsilon\right)\left(z_3^{+} - i\varepsilon\right)}}
{{f}_{ab}}\left( {{q}^{0}},x,\theta ,\varphi \right) \right.
\\
& \left.
\times
\exp{\left(
\frac{-2{{\left( {{q}^{0}} \right)}^{2}}+4Px+2{{x}^{2}}}{i\varepsilon }+\frac{s{{\left( {{q}^{0}} \right)}^{2}}+4{{P}^{2}}{{x}^{2}}}{-{{\varepsilon }^{2}}}
\right)}
\frac{1}{\left(z_3^{-} - i\varepsilon\right)\left(z_3^{+} - i\varepsilon\right)}
\right],
\end{split}
\end{equation}
\end{widetext}
\noindent
where the same variable change $q = P + x$ is also performed for $f_{ab}, \; z_3^+$ and $z_3^-$.
\begin{figure*}
\centering
\includegraphics[width=0.44\linewidth]{figs/a_qx_44p6GeV.png}
\includegraphics[width=0.44\linewidth]{figs/a_ey_44p6GeV.png}
\caption{Colormap plot of $\vert a_\varepsilon\left(q^0, x\right) \vert $ in $q^0x$ plane (a), and
$\vert a\left(E, y\right) \vert$ in $Ey$ plane (b) for $\varepsilon = 1$. The values of parameters: $\sqrt{s} = 44.6$ GeV, $M_p = 0.938$ GeV.}
\label{fig:a_abs}
\end{figure*}
We next turn to analyzing the integrand in \eqref{tensor_for_z12}. We denote it shortly by $a_\varepsilon\left(q^0, x, \theta, \phi\right)$. The exponent \eqref{exponenta_vid_log1} in the integrand depends only on the integration variables $q^0$ and $x$. The absolute value of this exponent has maximum at $\left(q^0 = 0, \; x = 0\right)$, and the \textit{width} of its peak tends to zero as $\varepsilon \to 0$. Since the other part of the integrand \eqref{tensor_for_z12} containing the second pair of denominators does not have singularities at $\left(q^0 = 0, \; x = 0\right)$, we can use such the exponential suppression of the integrand to cut the integration domain
\begin{equation}\label{domain_cut}
\int^\infty_{-\infty}dq^0\int^\infty_{-P}dx \to \int^{q^0_{cut}}_{-q^0_{cut}}dq^0\int^{x_{cut}}_{-x_{cut}}dx,
\end{equation}
\noindent
where $q^0_{cut} > 0$ and $0 < x_{cut} < P$ .
This idea is demonstrated in \fref{fig:a_abs}a. First, we found that the function $a_\varepsilon\left(q^0 = 0, x = 0, \theta, \phi\right)$ has maximum at $(\theta = 0, \phi = 0)$. Then we considered a restriction of $a_\varepsilon\left(q^0, x, \theta, \phi \right)$ to $a_\varepsilon\left(q^0, x, \theta = 0, \phi = 0\right)$ which we denoted by $a_\varepsilon\left(q^0, x\right)$ and plotted its absolute value in the $q^0x$ plane.
In this way we consider the contribution to the original integral \eqref{tensor_in_sphere} supplied by the first pair of denominators $z_1$ and $z_2$. Note that the actual value of this contribution does not depend on the selection of $q^0_{cut}$ and $x_{cut}$, which becomes clear as soon as we make the following change of variables
\begin{equation}\label{zamina1}
\begin{aligned}[c]
x=\varepsilon y
\end{aligned}
\qquad \qquad
\begin{aligned}[c]
{{q}^{0}}=\varepsilon E.
\end{aligned}
\end{equation}
\noindent
It makes the integration \eqref{domain_cut} to take the form
\begin{equation}\label{domain_cut_yE}
\int^{q^0_{cut}}_{-q^0_{cut}}dq^0
\int^{x_{cut}}_{-x_{cut}}dx
\to
\varepsilon^2
\int^{q^0_{cut}/\varepsilon}_{-q^0_{cut}/\varepsilon}dE
\int^{x_{cut}/\varepsilon}_{-x_{cut}/\varepsilon}dy
\end{equation}
\noindent
and also cancels all the $\varepsilon$ in denominators in \eqref{tensor_for_z12}. Since there is no $\varepsilon$ in the denominator now, we can finally pass to the limit when $\varepsilon \to 0$. At this point it is clear that the resulting integration limits will not depend on the particular selection of $q^0_{cut}$ and $x_{cut}$ because both $q^0_{cut}/\varepsilon$ and $x_{cut}/\varepsilon$ tend to infinity as $\varepsilon \to 0$ (see \fref{fig:a_abs}b), while the \textit{width} of the peak of
$\vert a\left(E, y\right) \vert$ remains unchanged.
Now \eqref{tensor_for_z12} can be rewritten as
\begin{equation}\label{tensor7}
\begin{split}
{{t}_{ab}}
\approx
&-{{P}^{2}}
\int\limits_{-\infty }^{\infty }{dE}
\int\limits_{-\infty }^{\infty }{dy}
\int\limits_{0}^{\pi }{\sin \left( \theta \right)}d\theta
\int\limits_{0}^{2\pi }{d\varphi }
\left[
\vphantom{
\frac{1}{\left(z_3^{-} - i\varepsilon\right)\left(z_3^{+} - i\varepsilon\right)}
}
{{f}_{ab}}\left(\theta ,\varphi \right)
\right.
\\
&\left.
\times
\exp \left( -4iPy-s{{E}^{2}}-4{{P}^{2}}{{y}^{2}} \right)
\frac{1}{\left(z_3^{-} z_3^{+} \right)}
\right],
\end{split}
\end{equation}
\noindent
where $f_{ab}\left(\theta, \varphi\right)$ now denotes $f_{ab}\left(q^0 = 0, x = 0, \theta, \varphi\right)$. The integrals over $E$ and $y$ in \eqref{tensor7} are now reduced to Poisson integrals
\begin{equation}\label{rozrahunok_puason}
\begin{split}
& \int\limits_{-\infty }^{\infty }{dE}\exp \left( -s{{E}^{2}} \right)=\frac{\sqrt{\pi }}{\sqrt{s}}, \\
& \int\limits_{-\infty }^{\infty }{dy}\exp \left( -4{{P}^{2}}{{y}^{2}}-4iPy \right)= \frac{\sqrt{\pi }}{2 e P},
\end{split}
\end{equation}
\noindent
where $e$ is the Euler's number (i.e. the base of the natural logarithm).
Combining \eqref{tensor7} with \eqref{rozrahunok_puason} and substituting expressions for $z_3^\pm$
we obtain
\begin{widetext}
\begin{equation}\label{tensor8}
\begin{split}
{{t}_{ab}}\approx & -\frac{\pi P}{2e\sqrt{s}}\int\limits_{0}^{\pi }{\sin \left( \theta \right)d\theta }\int\limits_{0}^{2\pi }{d\varphi }
\left[
\vphantom{\frac{a^2}{b^2}}
{{f}_{ab}}\left( \theta ,\varphi \right) \right. \\
& \left. \times \frac{1}{{{\left( M_{G}^{2}+2{{P}^{2}}-P\cos \left( \theta \right)\sqrt{4{{P}^{2}}-\left| t \right|} \right)}^{2}}-{{P}^{2}}\left| t \right|{{\sin }^{2}}\left( \theta \right){{\cos }^{2}}\left( \varphi \right)}
\right].
\end{split}
\end{equation}
\end{widetext}
Finally, there remains only two-dimensional integral over $\theta$ and $\varphi$ for each component $t_{ab}$. Moreover, we can calculate the integral over $\varphi$ for each pair of indices $(a, b)$ analytically (these calculations are quite long but yet straightforward). So we end up with one-dimensional integral over $\theta$ which can be calculated numerically.
|
\section{Introduction}
\begin{figure*}[t]
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/super8tab_block5200_1800_32kHz.png}
\end{subfigure}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/super8tab_block5200_1800_32kHz_spectrogram.png}
\end{subfigure}
\caption{On top, the MDCT amplitude representation $A_k(m)$ of a 14 second music sample with a 32,768Hz sample rate. This representation has 256 filter bands on the y-axis ($k: 0 \rightarrow 255$) and shows 1800 blocks on the x-axis. Negative MDCT amplitudes are represented as dark grey and positive amplitudes as light grey. To make the broad range of amplitudes present in the audio signal visible to the eye, the amplitudes are converted to the dB-scale before plotting, otherwise large part of the image would be plain grey as the amplitudes are small.
On the bottom, we show the MDCT spectrogram $A^2_k(m)$ of that same music sample. The intensity is again show on a dB-scale and the color is added as a visual aid to better distinguish the intensity of each pixel in the tensor. Visually, the spectrogram is much easier to interpret compared to the MDCT amplitudes. One can for example clearly make out several harmonics as the set of parallel horizontal lines in the spectrogram.}
\label{fig:spectrogram}
\end{figure*}
The impressive progress in the field of image generation has resulted in generative models which are able to generate high-resolution images indistinguishable from real pictures \cite{progan, stylegan, selfattention}. The field of generative models for music has long focused on generation of music scores.
For such models, the final conversion to sound relies on synthesized musical sounds (e.g. MIDI files) to convert the symbolic music into an audio signal. This approach reduces the dimensionality of the audio generation problem, but limits the output to a combination of the set of synthesized sounds.
With the increase in computing power in recent years, generative audio models have moved to generate raw audio directly \cite{wavenet, wavegan, googlespecgan, jukebox}. An high-quality stereo sample of 20 seconds has an uncompressed size of 3.4MiB when sampled at 44.1kHz with a 16-bit resolution. This is similar to the high-resolution 1024x1024 images with 3 8-bit color channels from \citet{progan} which have a 3.0MiB size in its uncompressed representation. \citet{jukebox} show that it is now possible to generate diverse and coherent audio samples several minutes long. However, these generative model require large amounts of computing power.
On the other hand, audio is an integral part of consumer electronic devices, which have much less computing power. This is in large part due to audio compression techniques such as MP3 and Vorbis compressing the audio data into much smaller data formats \cite{mp3explained}. MP3 relies on two key features to achieve a reduction in audio file size, often by as much as 75-93\% \cite{mp3compressionrate}. First, it leverages the psychoacoustic properties of human hearing \cite{pafilteroldarticle}. In particular, if a audio signal is below the threshold for hearing, the human ear will not perceive it. Additionally, certain louder sounds will mask more quiet sounds, even if the more quiet sound would have been audible when played by itself. The second key feature behind MP3 compression is Huffman coding which further decreases the data format size. Huffman coding would not be an easy data format for (convolutional) neural networks to parse, given the codes have variable length. The psychoacoustic properties, on the other hand, allow us to remove a significant amount of data from the audio signal without an audible impact, hence reducing the computing power required to generate raw audio samples with a generative model.
The psychoacoustic properties of the ear are expressed in the frequency domain, so MP3 encoding first converts an audio signal using the modified discrete cosine transformation (MDCT). Compared to other Fourier-related transformations, such as the Short-Time Fourier Transforms (STFT), the MDCT transformation has the added benefit that it is a real transformation. All phase information of the raw audio signal is hence encoded, without the need to use complex numbers, such as in \citet{googlespecgan}, or recreate the phase data with a spectrogram inversion algorithm as used in \citet{melnet}. Additionally, MDCT compacts the original audio signal in fewer amplitude components compared to other Fourier-related transforms.
In this work, we use the MDCT amplitude as the data representation for raw audio in a deep 2D convolutional Generative Adverserial Network (GAN). In the first layer of the discriminator, we add inaudible psychoacoustic noise to the representation, similar to the quantization noise in MP3 encoding. This widens the limited support of the real distribution, hence stabilizing the training process \cite{stablegan_withnoise2}.
The architecture of our network is inspired by the ProGAN model \cite{progan} with the noticeable difference that we don't increase/decrease the pixel density along the frequency axis with each successive model-block in the generator/discriminator. Instead, each model-block adds/removes an octave along the frequency axis.
MP3net does not require spectrogram inversion since the phase information is fully contained in the data representation. By training the top layers of the model with deeper weights frozen, we can eliminate noise and scratchiness in the audio.
As the deeper layers of the convolutional network have the full context of the entire sample, MP3net lends itself more naturally to long-range coherence.
Moreover, the model CNN-based architecture allows for almost instantaneous generation of new samples.
\section{Model}
\subsection{Modified Discrete Cosine Transform (MDCT) amplitudes as data representation}
Given a signal $x(t)$ sampled over $t \in [0, T[$ with sample rate $f_s = 1 / t_0$, we can group the sampled datapoints in $M$ blocks of length $N$:
\begin{equation} \label{eq:sampling}
x\Big((mN + n)t_0 \Big) \; \text{with} \; \left\{
\begin{array}{rcl}
n & = & 0, \dots, N-1 \\
m & = & 0, \dots, M-1
\end{array}
\right.
\end{equation}
where $T = N M t_0$.
By applying on each pair of consecutive blocks the following linear transformation, we can compute the MDCT amplitudes $A_k(m)$:
\begin{multline}
A_k(m) = \sum_{n=0}^{2N-1} x\Big((mN+n)t_0\Big) \; w_n \\
\cos \left[ \frac{\pi}{N} \left(n + \frac{1}{2} + \frac{N}{2}\right) \left(k + \frac{1}{2} \right) \right]
\end{multline}
where $k = 0, \dots, N - 1$ is the frequency index corresponding with the following frequency filter band:
\begin{equation} \label{eq:freq}
[f^0_k, f^0_k + \frac{f_s}{2N} [ = \left[ \frac{f_s}{2N} k, \frac{f_s}{2N} \left( k + 1 \right) \right[
\end{equation}
The discretized window function $w_n$ satisfies
\begin{eqnarray}
w_n = w_{2N-1-1} \\
w_n^2 + w_{n+N}^2 = 1
\end{eqnarray}
These latter conditions ensure that the MDCT transformation is invertible. As such, MDCT is a lossless linear transformation of the raw audio signal. Multiple choices exist for the window function, in our implementation we opted for the window function used by the Vorbis project behind the Ogg data format:
\begin{equation}
w_n = \sin\left(\frac{\pi}{2} \sin^2 \left[\frac{\pi}{2N}\left(n+\frac{1}{2}\right)\right]\right)
\end{equation}
The MDCT is similar to the Short-Time Fourier Transform (STFT) which has been applied in several recent music generation models \cite{wavegan, melnet, googlespecgan},
\begin{equation}
\tilde{A}_k(m) = \sum_{n=-\infty}^{\infty} x(nT) w_{n-m} \big( \cos kn - i \sin kn \big)
\end{equation}
where $w_n$ is a discretization of a window function with compact support.
Compared to the STFT, the MDCT transformation amplitudes $A_k(m)$ are real-valued whereas the STFT amplitudes $\tilde{A}_k(m)$ are complex. The spectrogram of the STFT is defined as $|\tilde{A}_k(m)|^2$. Taking the spectrum of the STFT amplitudes removes the phase information of the original signal. Hence, generative models producing a spectrogram also need to generate the corresponding phase to produce the audio signal. The Griffin-Lim algorithm \cite{griffin} and other approaches exist to reconstruct the phase \cite{invphase, phaselearning}. Additionally, MDCT has strong energy compaction properties compared to other discrete transformations such as STFT \cite{DCTbook} meaning that a lower number of non-zero amplitudes are needed to carry the same amount of information. And finally, the MDCT representation allows to leverage the psychoacoustic filtering effect (see section \ref{pafilter})
MP3net uses the MDCT representation directly, hence not removing the implicit phase information and benefiting from the compaction of the audio signal in few amplitudes. See figure \ref{fig:spectrogram} for an example of the MDCT representation and the corresponding spectrogram. Note that we observed that it is beneficial for convergence and expressivity of the model not to convert the amplitudes to a (signed) log-scale (dB) when representing the audio signal in our model. Since MDCT is a linear transformation, superposition of two audio samples is achieved by adding the respective amplitudes. We hypothesize this helps the model to easily superimpose different sounds from different feature channels to generate the output. Another benefit is not switching to the (signed) dB is that it allows for the application of progressive training \cite{progan} to reduce the training time. Indeed, if one trains the deeper layers in the progressive training process, one needs to train the model on downsampled (blurred) data. However, if one were to blur the dB-rescaled amplitudes, the blurred data tends to average out at 0, leaving no information in the blurred sample (see section \ref{section:relatedwork}).
\subsection{Psychoacoustic filter} \label{pafilter}
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{figures/psychoacoutic_filter.png}
\caption{Example spectrum of a pure tone at 440Hz with 3 additional harmonics. The spectrum is plotted in blue. The green shading indicates the region of inaudible amplitudes due to the limit of perception of the human ear. Amplitudes in the area shaded in yellow are also inaudible since the are "over-powered" by the sound represented by the blue spectrum. Note that the first harmonic at 880Hz is entirely inaudible.}
\label{fig:psychoacoutic_filter}
\end{figure}
The human ear is not able to distinguish between all the different possible raw audio signals. This phenomena is known as the psychoacoustic filter \cite{pafilterbook}. Mathematically this filter is expressed in the frequency spectrum, making the MDCT or STFT representation of the audio data the natural representation to apply this filter on.
The psychoacoustic filter consists of two key features. First, the filter leverages the fact that the human ear cannot hear certain quiet sounds. The level of perception depends on the frequency of the tone. Based on auditory experiments, one can approximate the amplitude intensity of this absolute threshold as \cite{pafilterbook}:
\begin{equation}
I_{\text{absolute threshold}, j} = 10^{\frac{1}{10} L_{\text{absolute threshold}}(f_j)}
\end{equation}
with
\begin{multline}
L_{\text{absolute threshold}}(f) [\mathrm{dB}] = \\ 3.64 f^{-0.8} - 6.5 e^{-0.6(f - 3.3)^2} + 10^{-3} f^4
\end{multline}
where $f$ is the frequency expressed in kHz. This function is commonly approximated as a stepwise function where $ L_{\text{absolute thresholdf}} $ is constant over a certain frequency range. Experimentally, these frequency ranges have been established as critical bands or Bark bands, with $f_j$ the mid point of each band. In figure \ref{fig:psychoacoutic_filter}, the absolute threshold is plotted in green. Amplitudes in the region below the absolute threshold (shaded in light green) are inaudible.
A second pyschoacoustic effect is that louder noises can render more quiet sounds inaudible, even if they are above the absolute threshold. This masking effect is also frequency dependent. The farther a frequency $j$ of an amplitude $A_j$ is from the frequency $i$ of the masking amplitude $A_i$, the weaker the masking effect:
\begin{equation} \label{eq:bark_start}
I_{\text{mask}, j} = \left( \sum_i \left( A^{2}_i \right)^{\alpha} 10^{\frac{\alpha}{10} \left( f_{ij} - O_j \right)} \right)^{1 / \alpha}
\end{equation}
with $f_{ij}$ the spreading function and $O_j$ the offset:
\begin{eqnarray}
f_{ij} & = & 15.81 + 7.5(i - j + 0.474) \nonumber \\
& & \quad - 17.5 \sqrt{1 + (i - j + 0.474)^2}\\
O_j & = & \tau (14.5 + j) + (1 - \tau) 5.5
\end{eqnarray}
Here, $\alpha$ is a fixed non-linear superposition coefficient and $\tau$ is the tonality of the amplitude spectrum $A_k$ given by the spectral flatness measure:
\begin{equation} \label{eq:bark_end}
\tau = \min \left(1, \frac{10}{-60\mathrm{dB}} \log_{10} \frac{ \exp \left( \frac{1}{N} \sum_k \ln A_k \right) }{ \frac{1}{N} \sum_k A_k} \right)
\end{equation}
which is $0$ for white noise and $1$ for a pure tone. The frequency indices in formulas (\ref{eq:bark_start})-(\ref{eq:bark_end}) refer to the mid-point frequencies of the respective Bark bands. In figure \ref{fig:psychoacoutic_filter}, the masking threshold of the spectrum (in blue) is plotted in yellow. Note that the first harmonic in the figure is below the masking threshold and hence inaudible.
In MP3 compression, all amplitudes are quantized as integer multiples $a_k(m)$ of the auditory threshold:
\begin{equation}
A_k(m) \approx a_k(m) \max(I_{\text{absolute threshold}, k}, I_{\text{mask}, k})
\end{equation}
This discretization process introduces an inaudible quantization error of size:
\begin{equation} \label{eq:quanterror}
\frac{1}{2} \max(I_{\text{absolute threshold}, k}, I_{\text{mask}, k})
\end{equation}
In the MP3 format, the auditory threshold is stored for each Bark band, together with the integer multiples using Huffman coding.
In generative models, this psychoacoustic auditive equivalence between two different audio signals can be leveraged to improve training stability and convergence. Alternatively it can be used to reduce the dimensionality of the problem. In MP3net, we opted to add gaussian noise with a standard deviation proportional to the psychoacoustic quantization error of equation (\ref{eq:quanterror}) to the MDCT amplitudes for both real and generated samples before they enter the discriminator. As studied by \citet{stablegan_withnoise} and \citet{stablegan_withnoise2}, this process improves the stability of the GAN training process, since it smooths the distribution of both the true and generated distributions, extending the support of both distributions. In earlier versions of our model, we implemented the psychoacoustic filter as a differentiable projection similar to a Leaky ReLU. This projection operator projected both real and generated samples onto a lower dimensional manifold by attenuating inaudible frequencies. However, the model yields better results with the gaussian noise approach.
We believe that the pyschoacoustic equivalence might also be useful for VAE-based generative models. The autoencoder can be trained to project out the psychoacoustic equivalence.
\section{Network architecture}
\subsection{2D convolutions to up- and downscale the MDCT amplitude representation}
Our network architecture is based on the architecture of the ProGAN network \cite{progan} with successive model blocks which scale up/down the image using strided 2D convolutions in the generator and discriminator respectively. Each model block in the generator takes the activation tensor from the previous layer as input and computes an activation tensor with double the resolution as output. The structure of the output tensor is such that each original input pixel is replaced with 2x2 pixels. Similarly, the model blocks of the discriminator halve the resolution of the input activation tensor.
To build a similar convolutional network which doubles the audio resolution, we need to consider the two different ways we can increase the resolution given our MDCT amplitude tensor representation. First, we can double the sampling rate $f_s \rightarrow 2 f_s$ of $x(t)$ by doubling the number of blocks $M \rightarrow 2M$. Using the notation of equation (\ref{eq:sampling}), we get:
\begin{equation}
\!\! x \! \left( \tilde{m}N t_0 + n \frac{t_0}{2} \right)
\end{equation}
with,
\begin{eqnarray}
n & = & 0, \ldots, N-1 \\
\tilde{m} & = & 0, \underbrace{\frac{1}{2}}_{\text{\ldots}}, 1, \ldots, \underbrace{M-\frac{3}{2}}_{N \text{ new blocks}}, M-1, \underbrace{M-\frac{1}{2}}_{\text{\ldots}}
\end{eqnarray}
We see that, similar to the image, each block $m$ of the input activation tensor $A_k(m)$ is replaced by two blocks $\tilde{m} = m$ and $m + 1/2$ in the output activation tensor $\tilde{A}_k(\tilde{m})$. We also note that while the number of filter bands $N$ has not changed, the corresponding frequencies of the filter bands have doubled to:
\begin{equation}
\tilde{f}^0_k = \frac{(2f_s)}{2N} k \quad \text{with} \; k = 0, \dots, N - 1
\end{equation}
Alternatively, we can double the sampling rate of $x(t)$ by doubling the size of the blocks $N \rightarrow 2N$:
\begin{equation}
\!\! x \! \left( (m 2N + \tilde{n}) \frac{t_0}{2} \right) \text{with} \left\{
\begin{array}{rcl}
\tilde{n} & = & 0, \ldots, 2N-1 \\
m & = & 0, \ldots, M-1
\end{array}
\right.
\end{equation}
This time there is no change to the block structure $m$ of the output activation tensor $\tilde{A}_{\tilde{k}}(m)$, but the filter bands corresponding with $\tilde{k}$ are now given by
\begin{equation}
\tilde{f}^0_{\tilde{k}} = \frac{(2f_s)}{2(2N)} \tilde{k}
\end{equation}
with,
\begin{equation}
\tilde{k} = 0, \dots, N - 1, \underbrace{N, \dots, 2N - 1}_{N \text{ new output frequencies}}
\end{equation}
Here, we see that each frequency component $k$ is not replaced by two corresponding frequencies $\tilde{k} = k$ and $k + 1/2$, as was the case for the block structure $m$ when doubling the number of blocks. Instead, we first have the original $N$ frequencies from the input tensor and then we add $N$ new higher frequencies. Intuitively, we can understand this as adding a new higher-pitched octave with $N$ frequencies to our MDCT amplitude representation. Indeed, these new frequencies with even index $\tilde{k}$ are exactly double the frequencies of the highest octave already present in the input activation tensor:
\begin{eqnarray}
\tilde{f}^0_{2k} &=& 2 f^0_k \; \text{where} \; k = \underbrace{\frac{N}{2}, \dots, N - 1}_{N/2 \text{ input frequencies}}
\end{eqnarray}
Given this relation, we can take the MDCT amplitudes of the highest octave of the input tensor (frequencies $k = N/2, \dots, N-1$) and apply a transposed convolution with stride 2 to generate the new higher-pitched octave with $N$ amplitudes (frequencies $\tilde{k} = N, \dots, 2N - 1$). We then concatenate the newly generated octave with the input amplitudes.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{figures/modelblock.png}
\caption{Design of the generator model block. The activation tensor of MDCT amplitudes $A_k(m)$ is represented by the horizontal bar, where the width illustrates the number of filter bands $k$. The (unbatched) dimensions of the tensors are indicated as $\text{block length} \times \text{filter bands} \times \text{channels}$. The notation for the transposed convolutions indicates the strides for the $\text{block} \times \text{filter band}$ directions.}
\label{fig:modelblock}
\end{figure}
Using the above method we build a generator model block (see figure \ref{fig:modelblock}) which doubles the number of frequency components ($N \rightarrow 2N$) and quadruples the number of blocks ($M \rightarrow 4M$). To allow the model to balance the MDCT amplitudes of the newly created octave with the lower octaves, we introduce a convolutional layer with strides $1 \times 1$ and kernel $1 \times 1$ and without activation before concatenating.
The discriminator model block has the same structure as the generator model block, with the layer order reversed except for the position of the octave balancing layer. In the generator, this layer is located just before the downsampled highest octave is summed with the lower octaves.
\subsection{Overall model architecture}
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{figures/model.png}
\caption{Model architecture of MP3net with 4 model blocks. The model to create 1m35s long samples with a 22,016Hz sample rate has 6 model blocks.}
\label{fig:modelarchitecture}
\end{figure}
The model architecture of MP3net is shown in figure \ref{fig:modelarchitecture}. The model is a WGAN with gradient penalty. Similar to \citet{progan}, the discriminator loss function has a small drift term $\epsilon_{\text{drift}} \mathbb{E}_{x \in \mathbb{P}_r} [D^2(x)]$.
Both real and generated MDCT amplitudes pass through the psychoacoustic layer where gaussian noise is added. The standard deviation of the gaussian is proportional to the quantization error computed in equation (\ref{eq:quanterror}).
The model also contains the minibatch standard deviation layer as described in \citet{progan}. This layer is added just before the last model block in the discriminator. It computes the variance over the minibatch for each of the elements in the activation tensor. This variance is added as an extra feature to the discriminator to force the generator to produce a diversity of generated samples. If the generator were only to produce a low diversity of samples, then the resulting low variance over the minibatch would help the discriminator to identify the samples as generated.
No batch normalization or pixelwise feature normalization layers (see \citet{progan}) are included in the generator. Note that we observed that standard batch normalization, where each feature is rescaled independently, destroys the training process as the model is unable to learn patterns which require calibration along the feature dimension.
The benefit of using a deep convolutional network is that since the deepest layers of our MP3net have the full context of the generated sample, they are able to produce samples with long-range coherence. Additionally, once the model is trained, inference is almost instantaneous.
Given that 2D convolutional networks are well studied due to their importance in image generation, we can borrow some of the tools developed in that context to further improve model preformance. In particular progressive training of the deeper layers could significantly reduce the training time \cite{progan} and self-attention layers such as discussed in \cite{selfattention} can further improve long-range coherence of the generated audio sample.
\section{Experiments with the MAESTRO dataset}
\subsection{Dataset description}
Our experiments are based on the MAESTRO-V2.0.0 dataset \cite{maestro}. This dataset contains over 200h of classical piano music, recorded over nine years of the International Piano-e-Competition. To train MP3net, we use the WAV audio files from the competition years 2004, '06, '08 and '09 (51GiB of WAV files). These audio files were recorded with conventional recording equipment and hence contain some background noises. We resampled the audio to 22,016Hz and sliced the pieces into fixed-length audio samples for training. As such, the start of the training samples does not usually coincide with the start of the music piece. We did not condition MP3net on starting position of the sample in the song as was done in \citet{jukebox}
\subsection{Training details}
We ran two experiments. In the first experiment, the model generates 95-second audio samples to evalutate long-range coherence, musicality and diversity of the generated samples. In the second experiment, we reconfigure the model to produce 5-second samples to study audio quality and timbre.
\subsubsection{95-second model}
We used a model with a 512 dimensional latent space and 6 subsequent model blocks (see figure \ref{fig:modelblock}). The dimension of the generated output is 16,384 $\times$ 128 $\times$ 2 in the MDCT representation of $\text{block length} \times \text{filter bands} \times \text{channels}$. We capped the number of feature channels at 512 in the deeper layers of the network. The dimensions of our model was memory-bound by the 8GiB HBM of each TPUv2 core. We used a batch-size of 8 and split convolutions over activations with large block length into multiple convolutions along the batch dimension to avoid TPU padding overhead on the batch dimension. The generator and discriminator each have 59 million parameters. We used the Adam optmizer with learning rate $0.0001$, $\beta_1 = 0.5$ and $\beta_2 = 0.9$.
We trained the model for 250h on a single Cloud TPUv2, corresponding with 750,000 iterations of the discriminator (batch size 8), with two discriminator updates for each generator update. While we have not benchmarked the resulting generated samples of MP3net against other multi-minute generative audio models, such as jukebox \cite{jukebox}, the training time of MP3net is orders of magnitude smaller than jukebox which requires approximately 400,000h on a single V100\footnote{Depending on the model type and implementation details 1 to 3 NVIDIA V100s are roughly equivalent to 1 Cloud TPUv2, which has 8 cores \cite{hardwarebenchmarking}}.
\subsubsection{5-second model}
\begin{figure*}[ht]
\centering
\begin{subfigure}[b]{\textwidth}
\centering
\includegraphics[width=0.48\linewidth]{figures/tonality_generated2.png}
\hfill
\includegraphics[width=0.48\linewidth]{figures/tonality_real2.png}
\end{subfigure}
\caption{Evolution of the tonality as function of the discriminator iteration during the training process. The left chart shows the tonality of the generated samples. The right graph displays the tonality of the real samples used for training the model}
\label{fig:tonality}
\end{figure*}
To study audio quality and timbre, we increased the number of features in the most shallow layer of the model to 128. We offset this increase in memory consumption by shortening the sample to 5s (1,024 $\times$ 128 $\times$ 2) with a model consisting of 5 subsequent model blocks. This model configuration has 64 million parameter for both the generator and the discriminator. We used the same Adam optimizer hyperparameters as for the 95s model and lowered $\beta_1$ to zero for the last 18,000 iterations to improve convergence.
The 5s model was trained for 120h on a single Cloud TPUv2, corresponding with 160,000 iterations of the discriminator (batch size 64).
\subsection{Observations}
A set of generated samples together with the source code of our model is available on the web page\footnote{\href{https://korneelvdbroek.github.io/mp3net/}{\url{https://korneelvdbroek.github.io/mp3net}}}.
\subsubsection{Audio quality and timbre}
A key parameter to compare the short-range performance of an audio model is the tonality $\tau = \mathbb{E}_m [\tau(m)]$ with $ \tau(m)$ defined in equation (\ref{eq:bark_end}). In figure \ref{fig:tonality}, we see that the tonality of the 95s samples quickly increases during initial training iterations, but then levels off in the 40-45\% range. Comparing with the real samples used for training which have a much higher tonality of around 72-73\% we see that a tonality gap of 30\% remains. The 5s samples exhibit a similar tonality gap.
Both the 95s model as well as the 5s model reproduce the piano timbre in the generated samples. However, the additional feature depth of the shallow layers of the 5s model results in a piano timbre which is clearly superior to that of the 95s model. The generated samples of the 5s model have a clearer and brighter piano sound, often with a better defined start of each note reproducing characteristic sound of the piano hammer hitting the string. Indeed, in the spectrogram of the 5s model one can more clearly identify the characteristic triangular structure of a piano note, where all harmonics start at exactly the same time and the higher harmonics fade out faster than the lower harmonics. The timbre of the 95s samples sometimes resembles more woodwind- or string-like timbres.
During the training process, we often observe a characteristic humming sound both for the 5s and the 95s generated samples. The spectrograms of the samples with hum exhibit a checkerboard pattern with a periodicity along the block (time) direction linked with the more shallow model blocks. Similar checkerboard patterns have also been observed in images generated using (transposed) convolutions. In \citet{checkerboard}, this phenomena is described, together with techniques to avoid these artefacts. When freezing the weights of all but the two most shallow model blocks in both the generator and the discriminator, the tonality improves and the humming noise disappears after training for an additional 10-20k iterations. However, this comes at the expense of less interesting musicality in the generated samples.
We also note that the expressivity of the highest octave in the samples produced by our model is limited. The model generates much less well defined harmonics in the highest octave, while many training samples do exhibit such structure in the highest octave.
\subsubsection{Musicality}
The rhythmic structure of the generated sample exhibits coherence from start to end of most 95s samples, with the tempo remaining constant throughout the sample.
The harmonic progressions in many of the generated samples follow the patterns common for western classical music. The chords structure is consistent between the start and end of these sample. Some other samples exhibit an atonal structure, without a clear tonal center. This latter is to be expected since ${\sim}4\%$ of the training dataset contains atonal pieces (mostly from Alexander Scriabin).
Some of the generated samples, exhibit clear melody. Even some of the 5s samples contain short musical phrases consisting of interesting motifs.
While the 95s clearly do not have the form and structure of a humanly composed piece, one can identify musical phrases starting softly (piano), building up with a crescendo towards a louder (forte) section. We have not identied any recurrent melodies (chorus) in any of the generated pieces.
\subsubsection{Sample diversity}
The MAESTRO dataset consists of music from the Baroque era over the Classical, Romantic and Impressionist styles, to the Expressionist period. Even though the generated samples are unprimed, samples resembling each of these styles are generated. The tempo of the generated samples also varies from slow pieces to samples with fast bravura. Harmony, chord progressions and melody are also very varied amongst the generated samples.
\section{Related Work} \label{section:relatedwork}
The field of synthetizing sound and music using computers is as old as computer science itself. With the advent of artificial neural networks new methods have become available to create audio. One of these techniques, Generative Adversarial Networks (GAN), uses two competing neural networks to generate a distribution of generated samples which closely approximates the true distribution of real samples. This technique was first described in the seminal paper by \citet{goodfellow2014}.
A lot of research in recent years has contributed to improving the stability of the training process of GANs. \citet{wasserstein} introduced the WGAN model with a loss function based on the Wasserstein distance between distributions. This loss function helps to improve stability, in particular for distributions which have a low dimensional support compared to the full dimensionality of the data representation space. \citet{gradientpenalty} introduced a further key improvement to the WGAN technique by replacing the weight clipping with a gradient penalty in the loss function. Multiple other GAN flavors exist with different loss functions and other features to stabilize the training process \cite{zerogp, hingeloss, ganscan}. Recent research on GAN stability has identified the key role played by the singular values of the network kernels \cite{singularvalues}.
GAN-based models have produced impressive results in the field of image generation. \citet{progan} introduced the ProGAN model. It can generate $1024 \times 1024$ images of faces in full color which are hard to distinguish from real pictures. The StyleGAN model \cite{stylegan}, is a further modification to the ProGAN model allowing one to tune the style of the generated image at each level of detail going from fine-grained features over middle-level styles, such as eyes, hair and lighting of the picture to high-level styles like hair style and face shape. The SAGAN model \cite{selfattention} uses a self-attention layer to increase the long-range coherence of convolutional neural nets and boost model performance for images which contain geometric structures.
As these models become very large with millions to billions of parameters, new challenges in stabilizing the trainings process present themselves. \citet{biggan} explore these issues and give an extensive hyper-parameter scan for hinge-loss based model such as SAGAN.
Part of the research into audio generation has focus on symbolic music generation such as in \citet{deepbach} and \citet{musegan}. Direct generation of the raw audio is often more computationally intensive but required for key applications such as authentically reproducing human speech in text-to-speech. Several different neural network techniques have been applied to generate raw audio. WaveNet \cite{wavenet} uses an autoregression model to generate speech which mimics human voices. Using Variational Autoencoder (VAE) based model, Jukebox \cite{jukebox} produces impressive quasi-realistic multi-minute songs primed on music genre, artist and lyrics. GAN-based models on raw audio were introduced in WaveGAN \cite{wavegan}.
Most closely related to the work presented here is GANsynth \cite{googlespecgan} and MelNet \cite{melnet}. GANsynth produces the STFT spectrograms for 4s audio samples. The GANsynth architecture is based on ProGAN. The resulting GANsynth audio samples of musical notes from different instruments are consistently judged by human evaluators of better fidelity compared to the similar samples generated by WaveNet. MetNet combines an RNN-based autoregressive model with a multi-scale generation procedure to generate STFT amplitude spectrograms which capture both the local as well as more long-range structures of spoken language and music.
\section{Future work}
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{figures/arvo_spectrogram_blur_long2.png}
\caption{Reduced spectrogram of the first 10 bars Arvo P\"art's \textit{Spiegel im Spiegel}. The spectrogram corresponds closely to the score of the piece. The technique of reducing the spectrogram can potentially be used to speed up the training process with a progressive training approach.}
\label{fig:blurring}
\end{figure}
Since the model architecture of MP3net is very similar to the convolutional networks well studies in the field of image generation, we can borrow many of the techniques of image generation. In particular, using progressive training \cite{progan} we might be able to reduce training time significantly and improve convergence. Progressive training, leverages the layered structure of the model where deeper layers are trained first. This training process is much faster since the memory footprint of the features in the deeper layers is much smaller allowing for larger batch sizes and since less weights have to be optimized.
To apply progressive training, one needs to guess (or impose) how the data is represented in the deeper layers. In the case of images, the authors of ProGAN take blurred versions of the real images and present these when training the deep layers. While the MP3net architecture is similar in many respects to ProGAN, the characteristics of the data representation is rather different. Images consist of regions where adjacent pixels vary smoothly, where our MDCT amplitude representation of audio signals exhibits periodic functions that oscillate between -1 and 1 (see figure \ref{fig:spectrogram}).
Initial experiments show that if we were to rescale the MDCT amplitudes to the dB-scale as is customary, blurring the images would very quickly remove all the information content from the image as the oscillating amplitudes would cancel out to zero. Our non-rescaled amplitudes don't exhibit this disadvantage. Experimentation with blurring the highest octave and subsequently folding (summing) it with the second highest octave has the advantage that harmonics in different octaves are summed together. Repeating this procedure gives a reduced spectrogram which could be used as the training data for the deeper layers. Visually, such a reduced spectrogram resembles music scores since harmonics are reduced/folded on top of the fundamental frequency. Note that this technique to reduce the spectrogram does not give the same results as merely downsampling the audio and then converting to a spectrogram. Figure \ref{fig:blurring} shows an example of a reduced spectrogram where the pixels that light up correspond clearly to the notes of the music piece.
Another prerequisite to successfully apply progressive training is that the data representation in each of the layers of the generator is properly normalized. Often, this is accomplished with batch normalization. In \citet{progan}, the authors apply a pixelwise feature normalization to obtain the same result.
We would like to explore the long-range coherence on a more divers set of training data. In particular, training on the Blizzard \cite{blizzard2011} and VoxCeleb2 datasets \cite{voxceleb2} would test the model's ability to synthesis speech. Transformer techniques as used in \citet{selfattention} can further improve the long-range coherence of the generated samples.
Another direction to explore is conditioning the network. \citet{melnet} conditioned MelNet on text for text-to-speech synthesis. In \citet{jukebox}, the model is conditioned on the lyrics of songs in a lyrics-to-song synthesis. Another approach to controlling the generated samples is to upgrade the network structure in line with StyleGAN \cite{stylegan}. This modified architecture would allow us to control the generated audio at different scales.
\section{Conclusion}
In this paper, we introduce MP3net, a 2D convolutional GAN with an architecture similar to some of the most successful image generation GANs \cite{progan, selfattention, stylegan}. MP3net borrows techniques from the field of audio compression. In particular, we use the MDCT as data representation since it includes all phase information of the original audio sample and we leverage the pyschoacoustic properties of human ears to simplify the training problem by widening our data and generated distributions.
Our model is able to produce high-quality, stereo audio with relative limited computational power. The samples produced exhibit long-range coherence over the full 95s of the samples produced. This ability is explained in part by the fact that the deepest convolutional layers of the model do have the full context of the sample. Hence they are able to generate music features that are coherent from the start to the end of the sample.
Compared to other generational models generating multi-minute samples, training times of MP3net are much shorter and inference is quasi-instantaneous given the inherent CNN-model architecture.
\section{Acknowledgement}
We would like to thank Werner Van Geit for providing feedback on the initial draft of this paper.
We also extend gratitude to Google Colab for making powerful compute available for everyone.
|
\section*{}
Standard one-dimensional stellar models require a boundary condition
that summarises the transition from their opaque depths to their
observable outer layers. These are usually provided by either (a) tables of
surface pressure and temperature computed using separate codes that
model the complicated physics of stellar atmospheres or (b) $T(\tau)$
relations that specify the stratification of the temperature $T$ as a function
of the optical depth $\tau$. These $T(\tau)$ relations follow the general form
\begin{equation}
\left(\frac{T(\tau)}{T\st{eff}}\right)^4=\frac{3}{4}\left(\tau+q(\tau)\right)
\label{e:hopf}
\end{equation}
where $T\st{eff}$ is the effective temperature and $q(\tau)$ is a
\emph{Hopf function}.
We denote the optical depth at which $T=T\st{eff}$ by $\tau\st{eff}$.
Hopf functions can be derived theoretically given some assumptions
(e.g. the Eddington model $q(\tau)=2/3$) or fit to data or detailed atmosphere models.
Two common $T(\tau)$ relations, besides the Eddington relation, are
those of \citet{krishna-swamy1966} and analytic fits to Model C by
\citet{valc}, known as VAL-C \citep[see e.g.][]{paxton2013,sonoi2019}.
In the study of solar-like oscillations, the atmosphere's structure
can affect the mode frequencies appreciably and must therefore be
included in the equilibrium stellar models. This can be done
by integrating the equations of the atmosphere's structure and
appending the integrated structure to the interior model. Alternatively,
one can modify the stellar structure equations and extend the model
to smaller optical depths. Specifically, one
multiplies the radiative temperature gradient
$\nabla\st{rad}=(\partial\ln T/\partial\ln P)\st{rad}$
by $1+\mathrm{d}q/\mathrm{d}\tau$
\citep[see e.g.][]{mosumgaard2018}. The first method requires the Hopf function;
the second requires its gradient.
\citet{trampedach2014a} presented Hopf functions for a set of
three-dimensional radiation-coupled hydrodynamics (3D RHD) simulations of
near-surface convection, as well as routines that allow stellar
modellers to interpolate the Hopf functions as a function of the
surface gravity $\log g$ and effective temperature $T\st{eff}$. The
simulated Hopf functions are most similar to that of VAL-C but do not
tend to a constant temperature at small optical depths.
While we encourage modellers to use the routines to interpolate
in the full suite of Hopf functions, we
present here an analytic function that allows relatively simple
implementation of the solar Hopf function, which is itself somewhat
representative of all the Hopf functions in the grid of models.
The gradient of the Hopf function for the simulation
with solar parameters ($T\st{eff}=5775\,\mathrm{K}$, $\log g=4.438$)
can be approximated by the function
\begin{equation}
\frac{\mathrm{d}q}{\mathrm{d}x}=\frac{c_1+e^\frac{x-a}{v}}{1+e^\frac{x-b}{w}}\mathrm{,}
\label{e:dq_dx}
\end{equation}
where $x=\log_{10}\tau$. This motivates fitting the Hopf function using the integral
of eq.~\ref{e:dq_dx} \citep{wolfram_atm_rnaas2021},
\begin{equation}
q(x)=c_0 + c_1\left(x-w\ln\left(e^\frac{b}{w}+e^\frac{x}{w}\right)\right)
+ v\,e^\frac{x-a}{v}\,{}_2F_1\left(1,\frac{w}{v};1+\frac{w}{v}; -e^\frac{x-b}{w}\right)\mathrm{,}
\label{e:q}
\end{equation}
where ${}_2F_1$ is the hypergeometric function. We found best-fitting parameters $c_0=0.6887302$,
$c_1=0.0668698$, $a=0.9262126$, $b=0.1148743$, $v=0.7657857$ and $w=0.0514999$,
with which the fit reproduces the data to within
$0.82$ per cent over the full range $-4.5<x<2.0$. The fit is also fairly representative
of all the simulations away from the low-temperature and low-gravity edges of
the grid. If we exclude simulations with $T\st{eff}<4400\,\mathrm{K}$
or $\log g < T\st{eff}/1000\,\mathrm{K} - 2.2$,
the fit reproduces all the remaining Hopf functions within $13$ per cent.
The hypergeometric function in eq.~\ref{e:q} is not always practical but the term that
contains it does not contribute to the function for $x\lesssim0$. Ignoring this term
is equivalent to ignoring the denominator in eq.~\ref{e:dq_dx}, in which case the integral
is
\begin{equation}
q(x)=c_0+c_1(x-b)+v\,e^\frac{x-a}{v}\mathrm{.}
\label{e:q_approx}
\end{equation}
This is also accurate to within $0.82$ per cent up to $x=0.0741$ for the solar model
and to the
same accuracy up to $\tau\st{eff}=0.5147929$. Thus, if integrating an
atmosphere using $q(x)$, where one usually terminates at or below
$\tau\st{eff}$, the approximate formula in eq.~\ref{e:q_approx} can be
used. If including the atmosphere's structure by modifying the
structure equations, then the full equation of the gradient
(eq.~\ref{e:dq_dx}) must be used because we require
$\mathrm{d}q/\mathrm{d}\tau\to0$ for $\tau\gg1$.
Fig.~\ref{f:fit} shows the Hopf functions we have discussed: the data
from \citet{trampedach2014a}, our fits of eqs~\ref{e:q} and
\ref{e:q_approx} to their solar simulation, and the three widely-used
$T(\tau)$ relations. Fig.~1 and most of the preceding analysis can be
generated by a publicly available Python
script\footnote{\url{https://github.com/warrickball/atm_rnaas2021}}
\citep{zenodo_atm_rnaas2021}.
\begin{figure*}
\centering
\includegraphics[width=\columnwidth]{fit.pdf}
\caption{All the Hopf functions $q(x)$ discussed, with the
corresponding values of $\tau\st{eff}$ indicated by filled circles. The
full set of simulations by \citet{trampedach2014a} is shown in
light grey and their solar simulation in black. The solid orange
line is our fit of eq.~\ref{e:q} to the data for the solar simulation
and the dashed orange line the
approximate function in eq.~\ref{e:q_approx} using the same
parameters. The blue, green and red lines are the popular Hopf
functions for the standard Eddington atmosphere, the relation by
\citet{krishna-swamy1966} and a fit to VAL-C
\citep{valc,paxton2013,sonoi2019}.}
\label{f:fit}
\end{figure*}
\acknowledgements
WHB thanks the UK Science and Technology Facilities Council
(STFC) for support under grant ST/R0023297/1.
Funding for the Stellar Astrophysics Centre is provided by The
Danish National Research Foundation (Grant agreement no.: DNRF106).
\software{NumPy\footnote{\url{http://www.numpy.org}} \citep{numpy2020},
SciPy\footnote{\url{http://www.scipy.org}} \citep{scipy2020},
Matplotlib\footnote{\url{http://matplotlib.org}} \citep{matplotlib}
}
\input{atm.bbl}
\end{document}
|
\section{INTRODUCTION}
\vspace{-3mm}
Advances in constraints programming have opened several venues for control system synthesis and verification of hybrid systems. For instance, linear programming and convex optimization are heavily used in a multitude of control system design and analysis tools. Recent surveys [\cite{surveuNumTool1,surveuNumTool2}] showed that such numerical tools had changed the control system design philosophy.
Nevertheless, linear and convex programming are limited in their ability to problems with specific structures. In several hybrid system design and verification problems, constraints are neither linear nor convex. This calls for efficient solvers that can reason about \textit{general multivariate polynomial constraints}. In that regard, Cylindrical Algebraic Decomposition (CAD) has long been one of the most influential algorithms capable of solving general multivariate polynomial constraints. The first CAD algorithm was introduced by [\cite{collins}]. However, the use of CAD is often limited by the number of variables in the input polynomials, a reflection of its worst-case complexity that grows in a doubly exponential fashion in the number of variables [\cite{complexityproblem1}].
To alleviate the CAD's doubly exponential issue, we introduce PolyAR, a highly parallelizable solver that uses convex programming and abstraction refinement to solve general multivariate polynomial inequality constraints. The main novel contributions of this work can be summarized as follows:
\begin{itemize}
\item PolyAR is a highly parallelizable solver that uses a combination of convex programming and abstraction refinement to solve multivariate polynomial inequality constraints.
\item PolyAR uses a novel convex abstraction refinement process where the original problem is iteratively relaxed into a series of convex programming problems with the aim to find the solution and prune the search space. Second, it refines such abstraction where it becomes tighter with each iteration of the algorithm. Finally, it examines in parallel all the identified small volume regions left from the abstraction using off-the-shelf solvers (e.g., Z3 and Yices) to search for a solution in these regions.
\item We validate our approach by comparing the scalability of the proposed PolyAR solver with respect to the latest versions of state-of-art non-linear real arithmetic solvers, such as Z3 8.9 and Yices 2.6, on synthesizing stabilizing static output feedback controller (SOF) for linear time-invariant (LTI) continuous systems and designing non-parametric controller for the non-linear Duffing oscillator.
\item We demonstrate the performance of PolyAR on the problem of designing switching signals for continuous-time linear switching systems.
\end{itemize}
\vspace{-1mm}
\textbf{Related work:} The original CAD algorithm that was introduced by \cite{collins} was the first algorithm that solves general polynomial inequality constraints. However, due to Collins CAD's high time complexity, there have been improvements to this algorithm. \cite{Hong} proposed an improvement of the projection operator in Collins CAD. However, the execution time of the modified version of the algorithm is still limited by the number of variables. \cite{McCalum} introduced a new projection operator which is a subset of Hong projection operator, removing redundant polynomials. However, McCallum proved that lifting over a sign-invariant CAD with this projection set is not sufficient to guarantee sign-invariance which makes the algorithm prone to error. The ABsolver tool proposed by \cite{ABsolver} leverages a generic nonlinear optimization tool for solving non-linear constraints. However, generic optimization tool may produce incomplete results, and possibly incorrect, due to the local nature of the solver. \cite{Z3} introduced Z3 which is another solver that implements an efficient nonlinear real arithmetic solver, that provide support for nonlinear polynomial arithmetic. However, it is still affected a lot by the increase in the number of variables in the polynomials. Because of the high complexity of existing approaches, we propose a highly parallelizable, efficient, and complete solver that uses the advantage and the simplicity of convex optimizations and abstraction refinement to solve higher order polynomial inequality constraints. To the best of our knowledge, this approach is new and has not been highlighted before.
\section{Problem Formulation}
\vspace{-3mm}
\subsection{Notation}
\vspace{-3mm}
We denote by $x=\big(x_1,x_2,\cdots,x_n\big) \in \mathbb{R}^n$ the set of real-valued variables, where $x_i \in \mathbb{R}$. We denote by $I_n=\big[\underline{d}_1,\overline{d}_1\big] \times \cdots \times$ $\big[\underline{d}_n,\overline{d}_n\big] \subset \mathbb{R}^{n}$ the $n$-dimensional region. We denote the space of polynomials with $n$ variables and coefficients in $\mathbb{R}$ by $\mathbb{R}[\left(x_1,x_2,\cdots,x_n\right)]$. We denote by $\wedge$ the Boolean conjunction. A set of the form $L^{-}_{0}\left(f\right)=\{\big(x_1,\cdots,x_n\big)\big|f\big(x_1,\cdots,x_n\big)\leq 0\}$ ($L^{+}_{0}\left(f\right)=\{\big(x_1,\cdots,x_n\big)\big|f\big(x_1,\cdots,x_n\big)\geq 0\}$) is called zero sublevel (superlevel) set of $f$, respectively.
\subsection{Main Problem}
In this paper, we focus on \textit{polynomial inequality constraints} with input ranges as closed boxes which are described in the following definition:
\begin{definition}
A polynomial inequality constraint $F=I^n~\wedge~P_m$ consists of:
\begin{itemize}
\item a set of interval constraints:
\begin{align}\label{pic1}
I^n=\bigwedge\limits_{i=1}^{n}x_i\in[\underline{d}_i,\overline{d}_i],
\end{align}
\item a polynomial constraint:
\begin{align}\label{pic2}
P_m=\bigwedge\limits_{i=1}^{m}~p_i\left(x_1,\cdots,x_n\right)~\leq~0,
\end{align}
\end{itemize}
where $p_i\left(x\right)=p_i\left(x_1,\cdots,x_n\right) \in \mathbb{R}[\left(x_1,x_2,\cdots,x_n\right)]$ is a polynomial over variables $x_1,\cdots,x_n$. Without loss of generality, $\bigwedge\limits_{i=1}^{m}~p_i\big(x\big)$ $\geq~0$ and $\bigwedge\limits_{i=1}^{m}~p_i\left(x\right)~=~0$ can be encoded in constraint number \eqref{pic2}.
\end{definition}
We are now in a position to state the problem that we will consider in this paper.
\begin{problem}
$\exists x=\left(x_1,\cdots,x_n\right)$ subject to $F=I^n~\wedge~P_m$.
\end{problem}
\section{Abstraction Refinement of Higher Order Polynomials Using Quadratic Polynomials} \label{sec:abstraction}
\vspace{-3mm}
Traditional techniques for solving Problem 1 focus on finding all the $n$ roots of the $m$ polynomials and check all the regions between two successive roots to assign a positive/negative sign for each of these regions. Therefore, solving Problem 1 is known to be a doubly combinatorial problem in $n$ with a total running time that is bounded by $\left(md\right)^{2^n}$ \cite{complexityproblem1}, where $d$ is the maximum degree among polynomials in $P_m$.
In problems that are doubly exponential in the input space $n$, it is beneficial to isolate subsets of the search space in which the solution is guaranteed not to exist. Recall that Problem 1 asks for an $x$ in $\mathbb{R}^n$ for which all the polynomials are negative. Therefore, a solution does not exist in subsets of $\mathbb{R}^n$ at which one of the polynomials is always positive. Similarly, isolating regions of the input space for which some of the polynomials are negative is also beneficial to finding the solution faster.
Our tool's main novelty is to use ``convex abstractions'' of the polynomials to find subsets of $L^{+}_{0}\left(p\right)$ and $L^{-}_{0}\left(p\right)$ efficiently. Indeed such ``abstractions'' may not be able to identify all regions for which the polynomial is positive or negative, which calls for an ``abstraction refinement'' process in which these ``convex abstractions'' become tighter with each iteration of the algorithm.
Figure~\ref{F1}(top) visualizes the proposed abstraction refinement process. Starting from a polynomial $p(x) \in \mathbb{R}[x]$ and an interval $I_n \subset \mathbb{R}^n$, we compute two quadratic polynomials:
\begin{align*}
O^p_1(x) &\ge p(x) \qquad \forall x \in I_n,\\
U^p_1(x) &\le p(x) \qquad \forall x \in I_n.
\end{align*}
where $O$ and $U$ stands for Over-approximate and Under-approximate quadratic polynomials, respectively, and the subscript in $O^p_1(x)$ and $U^p_1(x)$ encodes the iteration index of the abstraction refinement process. Computing such upper and lower abstractions can be carried out efficiently using Taylor approximation. Please refer to the example depicted in Figure~\ref{F1} (top) for a visualization of $O^p_1(x)$ and $U^p_1(x)$ for one dimensional higher order polynomial (order $\geq$ 3) defined in the closed interval $[\underline{d},~\overline{d}]~\subset~\mathbb{R}$.
The next step is to use the quadratic abstractions to isolate subsets of $L^{+}_{0}\left(p\right)$ and $L^{-}_{0}\left(p\right)$. It is particularly direct to show that the zero superlevel set of $U^p_1(x)$ is a subset of $L^{+}_{0}\left(p\right)$, i.e., $L^{+}_{0}\left(U^p_1\right) \subseteq L^{+}_{0}\left(p\right)$. Similarly, the zero sublevel set of $O^p_1(x)$ is a subset of $L^{-}_{0}\left(p\right)$, i.e., $L^{-}_{0}\left(O^p_1\right) \subseteq L^{-}_{0}\left(p\right)$. Thanks to the fact that $O^p_1(x)$ and $U^p_1(x)$ are quadratic polynomials, finding their zero superlevel and zero sublevel sets, respectively, can be computed efficiently. Referring to the example in Figure~\ref{F1}(top), these zero superlevel and sublevel sets are $L^{+}_{0}\left(U^p_1\right)=[x_1, \overline{d}]$ and $L^{-}_{0}\left(O^p_1\right)=[\underline{d},x_0]$, respectively.
It is clear from Figure~\ref{F1}(top) that the abstractions $O^p_1(x)$ and $U^p_1(x)$ fails to identify all subsets of $L^{-}_{0}\left(p\right)$ and $L^{+}_{0}\left(p\right)$. Therefore, the next step is to compute tighter over and under approximations of $p(x)$. Such a refinement process can be carried out by removing the zero superlevel and the zero sublevel sets, i.e., $L^{+}_{0}\left(U^p_1\right)$ and $L^{-}_{0}\left(O^p_1\right)$, identified using the previous abstraction and computing new over and lower approximation, as shown in Figure~\ref{F1}(bottom). The process of abstraction refinement can continue until the remaining subsets of the search space, in which case we call them ambiguous regions, and with some abuse of notation, denoted them by $L^{+/-}_{0}\left(p\right)$, are \emph{small} enough to be analyzed using off-the-shelf solvers. More details about the proposed abstraction refinement process are given in the next section.
\begin{figure}
\includegraphics[width=0.9\columnwidth, trim=0 50mm 0 0, clip]{AbstRefi.eps} \\
\includegraphics[width=0.9\columnwidth, trim=0 70mm 0 0, clip]{AbstRefi1.eps}
\caption{Abstraction Refinement of higher order polynomial using quadratic approximations: (top) first iteration and (bottom) second iteration.}
\label{F1}
\end{figure}
\vspace{-3mm}
\section{Algorithm Architecture}
\vspace{-3mm}
\begin{figure}[!t]
\includegraphics[width=1.0\columnwidth]{PolyAR.eps}
\caption{Framework of PolyAR.}
\label{fig:PolyAR}
\end{figure}
In this section, we describe the different steps used by our solver PolyAR to solve Problem 1.
Our design methodology for the PolyAR tool aims to reduce the number of the required abstraction refinement and tries to find a solution early on in the process. To that end, the tool starts by computing a set of convex (quadratic or linear) polynomials $O_0^{p_i}(x), i = 1,\ldots,m$, that over approximate the original polynomials. The next step is to solve a convex feasibility problem aiming to find a solution that satisfy the constraints:
$$ \exists x = (x_1,\ldots,x_n) \qquad \text{s.t.} \qquad O_0^{p_i}(x) \le 0, \quad i = 1,\ldots,m.$$
Indeed, if such a convex problem is feasible, the tool terminates and returns the solution found by the convex feasibility problem above (\textbf{Conv\_Solver}, Line~\ref{alg:line:conv} in Algorithm~\ref{alg:PolyAR}). If not, then the tool selects one polynomial $p_j$ (\textbf{Select\_Poly}, Line \ref{alg:line:selectPoly}) to perform the abstraction refinement process. Indeed, several heuristics can be applied to select which polynomial will be selected. In the PolyAR tool, we opt-out to select the polynomial with the highest Lipschitz constant. Our intuition is that the higher the Lipschitz constant, the harder to obtain a tight over-approximation that can be used to find the solution.
Once a polynomial $p_j$ is selected, the next step is to apply the abstraction refinement process on $p_j$ (\textbf{Abst\_Refin}, Line~\ref{alg:line:abst_ref} in Algorithm~\ref{alg:PolyAR}). The objective of the abstraction refinement process is to identify subsets of the positive regions $L^{+}_{0}\left(p_j\right)$ and negative regions $L^{-}_{0}\left(p_j\right)$. Indeed, such abstraction refinement may not be able to identify all positive and negative regions, and hence a remaining portion of the search space may not be identified to belong to either $L^{+}_{0}\left(p_j\right)$ or $L^{-}_{0}\left(p_j\right)$ in which case it belongs to the ambiguous region $L^{+/-}_{0}\left(p_j\right)$. The abstraction refinement process of the polynomial $p_j$ ensure that the volume of such ambiguous regions are below a certain user defined threshold.
The process of using the convex solver to find the solution and abstracting one polynomial continues. Since a solution of Problem 1 needs to lie in a negative region for all the polynomials, we confine the tool attention to the negative regions identified by the abstraction refinement in the previous iterations (Line~\ref{alg:line:neg} in Algorithm~\ref{alg:PolyAR}) to accelerate the process of searching for the solution.
While excluding the positive regions identified in previous iterations does not affect the tool (since a solution is guaranteed not to exist in such regions), excluding the ambiguous regions from the next iterations may affect the correctness of the tool. Therefore, the last step in the PolyAR tool is to examine all the identified ambiguous regions using off-the-shelf solvers (e.g., Z3 and Yices) to search for a solution in these regions (\textbf{Solver\_Parallel}, Line~\ref{alg:line:solver} in Algorithm~\ref{alg:PolyAR}). Because the volume of these ambiguous regions is smaller than a user-defined threshold, the execution time of running off-the-shelf tools on such small volume regions is shorter than solving the original problem. This reflects that the number of roots for each polynomial is limited in small regions. Moreover, searching for a solution in these ambiguous regions can be highly parallelized, leading to an extra level of efficiency. This process is summarized in Algorithm 1 and Figure~\ref{fig:PolyAR}. We describe in detail each block algorithm that constitutes Algorithm 1 in the next subsections.
\begin{algorithm}
\caption{PolyAR$\left(F\right)$} \label{alg:PolyAR}
\begin{flushleft}
\textbf{Input:} $F = I^n \wedge P_m$\\
\textbf{Output}: STATUS, $x_{\text{Sol}}$
\end{flushleft}
\begin{algorithmic}[1]
\STATE $Neg= \{I_n \}$
\STATE $Ambig= \{\}$
\STATE $\text{List\_pols}= \{p_1,\ldots,p_m\}$
\WHILE{$\text{List\_pols}~\neq~ \emptyset$}
\STATE $x_{\text{Sol}}:=\textbf{Conv\_Solver}\left(Neg,\text{List\_pols}\right)$ \label{alg:line:conv}
\IF{$x_{\text{Sol}}~\neq~\text{None}$}
\STATE STATUS=SAT
\RETURN STATUS, $x_{\text{Sol}}$
\ENDIF
\STATE $p_j=\textbf{Select\_Poly}\left(\text{List\_pols}\right)$ \label{alg:line:selectPoly}
\STATE $L^{-}_{0}\left(p_j\right),L^{+}_{0}\left(p_j\right),L^{+/-}_{0}\left(p_j\right)$ \\ \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad $:= \textbf{Abst\_Refin}\left(Neg,p_j\right)$ \label{alg:line:abst_ref}
\STATE $Ambig.\text{add}\left(L^{+/-}_{0}\left(p_j\right)\right)$
\IF{$L^{-}_{0}\left(p_j\right)~==~\emptyset$}
\STATE break
\ENDIF
\STATE $Neg=L^{-}_{0}\left(p_j\right)$ \label{alg:line:neg}
\STATE $\text{List\_pols}=\text{List\_pols}\setminus{p_j}$
\ENDWHILE
\IF{$\text{List\_pols}~\neq~ \emptyset$}
\STATE $\text{STATUS}, x_{\text{Sol}}:=\textbf{Solver\_Parallel}\left(Ambig,P_m\right)$ \label{alg:line:solver}
\RETURN $\text{STATUS}, x_{\text{Sol}}$
\ELSE
\STATE $\text{STATUS=SAT}$
\STATE $x_{\text{Sol}}=\text{center}\left(Neg\right)$
\RETURN $\text{STATUS}, x_{\text{Sol}}$
\ENDIF
\end{algorithmic}
\end{algorithm}
\vspace{-2mm}
\subsection{Early Termination Using \textbf{Conv\_Solver}:}
The objective of the \textbf{Conv\_Solver} (Algorithm 2) is to search for a solution to Problem 1 using the information of (i) a set of closed convex regions $Neg$ identified by the previous iterations of the abstraction refinement process and (ii) a list of polynomials (List\_pols) that have not yet been processed by the abstraction refinement process.
Our approach is to compute a convex over-approximation of the polynomials in List\_pols using Taylor approximation. To that end, we recall the definition of Taylor polynomials:
\begin{definition}
Let $f:\mathbb{R}^n\rightarrow\mathbb{R}$ be two times differentiable in open interval around a point $a \in \mathbb{R}^n$, then $f\left(x\right)$ can be written in terms of first and second order Taylor polynomials , $T_1\left(x\right)$ and $T_2\left(x\right)$, around the neighborhood of $a$, as follows:
\vspace{-2mm}
\begin{eqnarray}\label{eq6}
f\left(x\right) &=& \left(x-a\right)^{T}D_{f}\left(a\right)+R_1\left(c_1\right)\nonumber\\
&=&T_1\left(x\right)+R_1\left(c_1\right), \\
f\left(x\right) &=& \left(x-a\right)^{T}D_{f}\left(a\right)\nonumber\\
&+&\frac{1}{2}\left(x-a\right)^{T}H_{f}\left(a\right)\left(x-a\right)+R_2\left(c_2\right)\nonumber\\
&=&T_2\left(x\right)+R_2\left(c_2\right),
\end{eqnarray}
\end{definition}
\vspace{-2mm}
\noindent where $T_1\left(x\right)=\left(x-a\right)^{T}D_{f}\left(a\right)$ and $T_2\left(x\right)=\left(x-a\right)^{T}D_{f}\left(a\right)+\frac{1}{2}\left(x-a\right)^{T}H_{f}\left(a\right)\left(x-a\right)$. $D_{f}\left(a\right)$ and $H_{f}\left(a\right)$ denote the Gradient vector and the Hessian matrix of $f$ at the point $a$. $R_1\left(c_1\right)$ and $R_2\left(c_2\right)$ are reminders that depend on $a$ and two points $c_1 \in \mathbb{R}^n$ and $c_2 \in \mathbb{R}^n$ that are located in the neighborhood of $a$.
We upper-bound $R_1\left(c_1\right)$ and $R_2\left(c_2\right)$ by $M_1 \in \mathbb{R}_{+}$ and $M_2 \in \mathbb{R}_{+}$, i.e., $\abs{R_1\left(c_1\right)} \leq M_1$ and $\abs{R_2\left(c_2\right)} \leq M_2$. Hence:
\begin{align*}
T_1(x) - M_1 &\le f(x) \le T_1(x) + M_1 \\
T_2(x) - M_2 &\le f(x) \le T_2(x) + M_2
\end{align*}
Next, we check the convexity of the obtained second Taylor approximation and use it to compute the over-approximation function $O^{p_i}$ whenever it is convex (Line~\ref{line:alg:taylor2} in Algorithm~\ref{alg:conv}). Otherwise, we use the first Taylor approximation instead (Line~\ref{line:alg:taylor1} in Algorithm~\ref{alg:conv}). Finally, for each $region$ in the set of negative regions ($Neg$), we solve the following convex feasibility problem:
\begin{align}\label{eq10}
x_{\text{Sol}}:=&\argmin\limits_{x \in region}^{}1 ~~ \text{s.t.} ~~ O^{p_i}\left(x\right) \leq~0,~~i \in \text{List\_pols}.&
\end{align}
\begin{algorithm}[t]
\caption{$\textbf{Conv\_Solver}\left(Neg,\text{List\_pols}\right)$} \label{alg:conv}
\begin{flushleft}
\textbf{Input:} $Neg$, $\text{List\_pols}$\\
\textbf{Output}: $x_{\text{Sol}}$
\end{flushleft}
\begin{algorithmic}[1]
\FOR{$region \in Neg$}
\FOR{$i \in \text{List\_pols}$}
\IF{$\texttt{Taylor}_{over}\left(p_i\left(x\right), 2\right)~\text{is convex}$}
\STATE $O^{p_i}\left(x\right)=\texttt{Taylor}_{over}\left(p_i\left(x\right), 2\right)$ \label{line:alg:taylor2}
\ELSE
\STATE $O^{p_i}\left(x\right)=\texttt{Taylor}_{over}\left(p_i\left(x\right), 1\right)$ \label{line:alg:taylor1}
\ENDIF
\ENDFOR
\STATE $x_{\text{Sol}}:=\argmin \limits_{x \in region}^{}1 \quad s.t. \quad O^{p_i}\left(x\right)\leq~0$ (see eq. \eqref{eq10}))
\RETURN $x_{\text{Sol}}$
\ENDFOR
\end{algorithmic}
\end{algorithm}
\vspace{-5mm}
\subsection{Abstraction Refinement Using \textbf{Abst\_Refin}:}
Given the set of negative regions $Neg$ identified by the previous abstraction refinement process along with the polynomial $p_j$ selected by the \textbf{Select\_Poly} algorithm, the objective of the \textbf{Abst\_Refin} algorithm is to find subsets of the zero sublevel sets of $p_j$ that lie inside $Neg$. The output of this algorithm are subsets of $L_0^-(p_j)$ and $L_0^+(p_j)$. The remainder of $Neg$ is then considered to be part of the ambiguous regions $L_0^{+/-}(p_j)$. To do so, for every $region$ in $Neg$, the tool initiates a list of ambiguous regions $List\_Ambig\_reg$, which will contain all the ambiguous regions from the abstraction refinement (Line 4 in Algorithm 3). Next, it selects one element from these ambiguous regions (Line 5 in Algorithm 3) and performs the abstraction refinement on this region iteratively until the volume of the remaining ambiguous region is smaller than a user-defined threshold (Line 6 in Algorithm 3). During the iterative abstraction refinement, all the identified zero sublevel and superlevel subsets are stored in the sets $L_0^-(p_j)$ and $L_0^+(p_j)$, respectively.
While the zero sublevel (superlevel) sets of the quadratic over-approximation (under-approximation) are ellipsoid or hyperboloid in general, we opt to represent all the subsets of $L_0^-(p_j)$ and $L_0^+(p_j)$ as $n-$dimensional hypercubes. This choice reflects the fact that off-the-shelf solvers (e.g., Z3 and Yices) can exploit the geometry of hypercubes to accelerate their computations. The process of finding these hypercubes can be summarized as follows:
\begin{enumerate}
\item \textbf{Step 1:} Compute the largest polytope inside the ellipsoid or hyperboloid representing the zero sublevel (superlevel) sets of the quadratic over-approximation (under-approximation) of $p_j$. To that end, we use a set of user-defined templates for the polytope.
\item \textbf{Step 2:} The previous step uses user-defined templates to find the polytope, such templates may fail and return an infeasible solution. In such scenarios, we split the ambiguous region into two (along the longest dimension) until a polytope is found.
\item \textbf{Step 3:} Finally, we under approximate the computed polytope with hypercubes.
\end{enumerate}
This process is visualized in Figure~\ref{undpol}. The details of each of these steps are given in the following subsections.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.7\columnwidth]{underapproxpolyt.eps}
\caption{\small{Polytopic under-approximation of a $2-$dimensional ellipse sublevel set $L^{-}_{0}\left(p_j\right)$. $\mathcal{P}^{N}$ presents the under-approximate polytope inscribed in $L^{-}_{0}\left(p_j\right)$, and $\mathcal{B}^{N}$ represents the axis-aligned box of maximum volume inscribed in $\mathcal{P}^{N}$.}}
\label{undpol}
\end{figure}
\subsubsection{\textbf{Step 1: Computing the largest polytope subset of $L_0^-(p_j)$ and $L_0^+(p_j)$}}
Given the over-approximation $O^{p_j}$ computed using Taylor polynomials (detailed in Section 4.1) and a convex ambiguous region (Ambig\_reg), we start by computing a set of $n+1$ vertices $v^{N}_1, \ldots, v^{N}_{n+1}$ that are inscribed in the ambiguous region $Ambig\_reg$. Each vertex can be computed by solving the following convex optimization problem:
\begin{align}\label{vertn}
v^{N}_i=&\argmin \limits_{v_i \in Ambig\_reg} \left(l^T_iv_i\right)
\qquad \text{s.t.} \qquad O^{p_j}(v_i)~\leq~0,&
\end{align}
where $l_i$ is a user defined normal vector (or template) (see Figure~\ref{undpol} for graphical representation of such normal vectors). Using these vertices, we can obtain the polytope $\mathcal{P}^{N}$ as:
$$\mathcal{P}^{N} = \textbf{Convex\_Hull}\left(v^{N}_1, \ldots, v^{N}_{n+1} \right).$$
Thanks to the constraints in the optimization problem~\eqref{vertn} along with the convexity of $L_0^-(p_j)$, it is direct to conclude that the polytope $\mathcal{P}^{N}$ satisfy $\mathcal{P}^{N} \subset L_0^-(p_j)$. We compute the polytope $\mathcal{P}^{P} \subset L_0^+(p_j)$ in a similar fashion using the under-approximation $U^{p_j}$ (Line 23 in Algorithm 3).
\begin{algorithm}[t]
\caption{$\textbf{Abst\_Refin}\left(Neg,p_j\right)$}
\begin{flushleft}
\textbf{Input:} $Neg$, $p_j$ \\
\textbf{Output}: $L^{-}_{0}\left(p_j\right)$, $L^{+}_{0}\left(p_j\right)$, $L^{+/-}_{0}\left(p_j\right)$
\end{flushleft}
\begin{algorithmic}[1]
\STATE $L^{-}_{0}\left(p_j\right)=\{~\}$, $L^{+}_{0}\left(p_j\right)=\{~\}$, $L^{+/-}_{0}\left(p_j\right)=\{~\}$
\FOR{$region \in Neg$}
\STATE $\text{vertices}^{N}=\{~\}$, $\text{vertices}^{P}=\{~\}$
\STATE $List\_Ambig\_reg=\{region\}$
\STATE $Ambig\_reg=\textbf{Select\_region}\left(List\_Ambig\_reg\right)$
\WHILE{$\text{Volume}\left(Ambig\_reg\right)~>~\text{Vol}_{\text{threshold}}$}
\STATE $List\_Ambig\_reg=List\_Ambig\_reg\setminus{Ambig\_reg}$
\FOR{$i~\in~\big(1,\cdots,n+1\big)$}
\STATE $v^{N}_i=\argmin\limits_{v_i \in Ambig\_reg} \left(l^T_iv_i\right) \quad s.t. \quad O^{p_j}(v_i) ~\leq~0.$
\IF{$v^{N}_i~\neq~\text{None}$}
\STATE $\text{vertices}^{N}.\text{add}\left(v^{N}_i\right)$
\ENDIF
\STATE $v^{P}_i=\argmin\limits_{v_i \in Ambig\_reg} \left(l^T_iv_i\right) \quad s.t. \quad U^{p_j}(v_i)~\leq~0.$
\IF{$v^{P}_i~\neq~\text{None}$}
\STATE $\text{vertices}^{P}.\text{add}\left(v^{P}_i\right)$
\ENDIF
\ENDFOR
\IF{$\big(\text{vertices}^{N}==\emptyset ~ \text{and} ~ \text{vertices}^{P}==\emptyset\big)$}
\STATE $Ambig\_reg_{1}$, $Ambig\_reg_{2}$ \\ \quad \quad \quad \quad \quad \quad \quad \quad \quad $:=\text{\textbf{Half\_Div}}\left(Ambig\_reg\right)$
\STATE $List\_Ambig\_reg.add\left(Ambig\_reg_{1}, Ambig\_reg_{2}\right)$
\ELSIF{$\big(\text{vertices}^{N}\neq \emptyset ~ \text{and} ~ \text{vertices}^{P}\neq \emptyset\big)$}
\STATE $\mathcal{P}^N=\textbf{Convex\_Hull}\left(\text{vertices}^{N}\right)$
\STATE $\mathcal{P}^P=\textbf{Convex\_Hull}\left(\text{vertices}^{P}\right)$
\STATE $\mathcal{B}^{N}=\textbf{Box}\left(\mathcal{P}^N\right)$; $\mathcal{B}^{P}=\textbf{Box}\left(\mathcal{P}^P\right)$
\STATE $L^{-}_{0}\left(p_j\right).\text{add}\left(\mathcal{B}^{N}\right)$; $L^{+}_{0}\left(p_j\right).\text{add}\left(\mathcal{B}^{P}\right)$
\STATE $Ambig\_reg=Ambig\_reg\setminus{\left(\mathcal{B}^{N} \cup \mathcal{B}^{N}\right)}$
\STATE $List\_Ambig\_reg.\text{add}\big(Ambig\_reg\big)$
\ENDIF
\STATE $Ambig\_reg=\textbf{Select\_region}\left(List\_Ambig\_reg\right)$
\ENDWHILE
\STATE $L^{+/-}_{0}\left(p_j\right).add\left(List\_Ambig\_reg\right)$
\ENDFOR
\RETURN $L^{-}_{0}\left(p_j\right)$, $L^{+}_{0}\left(p_j\right)$, $L^{+/-}_{0}\left(p_j\right)$
\end{algorithmic}
\end{algorithm}
\subsubsection{\textbf{Step 3: Under approximate the polytopes with axis aligned boxes:}}
To compute the largest axis-aligned hypercube $\mathcal{B}^{N}$ inscribed inside the polytope $\mathcal{P}^{N}$, we solve the following convex optimization problem~\cite{InscribRec}:
\begin{align}\label{eq14}
& \argmax\limits_{(l_1^N, u_1^N, \ldots, l_n^N,u_n^N)\in \mathbb{R}^{2n}}^{} \sum\limits_{k=1}^{n}\log\left(u^N_k-l^N_k\right)\nonumber\\
& \text{s.t.} \sum\limits_{k=1}^{n}\left(p^{N,+}_{ik}u^N_{k}-p^{N,-}_{ik}l^N_{k}\right)\leq c^N_i,~i=1,\cdots,n_p,
\end{align}
where $(l_1^N, u_1^N, \ldots, l_n^N,u_n^N) \in \mathbb{R}^{2n}$ is the representation of the box $\mathcal{B}^{N}$ with $(l_k,u_k)$ is the lower/upper limit of the box in the $k$th dimension, $p_{ik}^{N,+}=\max\{p^N_{ik},0\}$, $p_{ik}^{N,-}=\max\{-p^N_{ik},0\}$, and $p^N_{ik},c^N_i$ are the rows of the half-space matrix/vector representation of the polytope $\mathcal{P}^{N}$.
\subsection{Highly Parallelizable Analysis of Ambiguous Regions using \textbf{Solver\_Parallel}}
\vspace{-2mm}
Once all the ambiguous regions are identified, the next step is to analyze all of them using off-the-shelf solvers. In particular, PolyAR supports the use of the latest versions Z3 8.9 and Yices 2.6 solvers. Thanks to the fact that all the ambiguous regions are hypercubes, both these solvers can exploit the geometry of the region to accelerate their computations. Also, thanks to the fact that the volume of all ambiguous regions is lower than a user-defined threshold, the CAD algorithm can run efficiently. To that end, PolyAR tool runs multiple instances of Z3 or Yices to analyze all these ambiguous regions in parallel as summarized in Algorithm~\ref{alg:solveparllel}.
\begin{algorithm}[t]
\caption{$\textbf{Solver\_Parallel}\left(Ambig,P_m\right)$} \label{alg:solveparllel}
\begin{flushleft}
\textbf{Input:} $Ambig$, $P_m$\\
\textbf{Output}: $\text{STATUS}$, $x_{\text{Sol}}$
\end{flushleft}
\begin{algorithmic}[1]
\STATE The tool runs off-the-shelf solvers such as Z3 or Yices on small-volume ambiguous regions in $Ambig$ in parallel:
\STATE $\text{STATUS}, x_{\text{Sol}}:=\textbf{Z3/Yices\_Parall}\left(Ambig,P_m\right)$
\RETURN $\text{STATUS}, x_{\text{Sol}}$
\end{algorithmic}
\end{algorithm}
\section{Extension to SMT solving}
We extend the PolyAR solver described in the previous sections to account for combinations of Boolean and Polynomial inequality constraints of the form:
\begin{align}
\exists& (b_1,\ldots,b_o, x_1,\ldots,x_n) \in \mathbb{B}^o \times \mathbb{R}^n, \nonumber \\
&\text{subject to:} \nonumber \\
& p_i(x_1,\ldots,x_n) \le 0, && i = 1,\ldots,m \\
&x_k \in [\underline{d}_k, \overline{d}_k], && k = 1,\ldots,n\\
&\varphi_j(b_1,\ldots,b_o) \; \longleftrightarrow \;\texttt{TRUE}, && j = 1,\ldots,r\\
&b_l \longleftrightarrow \big( p_{l+m}(x_1,\ldots,x_n) \; \le \; 0 \big), && l = 1,\ldots,h \label{eq:hybrid}
\end{align}
where $\varphi_j(b_1,\ldots,b_o)$ is any combinations of Boolean and pseudo-Boolean predicates.
We can create a Satisfiability Modulo Theory (SMT) solver by combining a SAT solver for Boolean and pseudo-Boolean constraints and a theory solver (PolyAR) for interval and polynomial constraints on real numbers by following the lazy SMT paradigm \cite{lazysmt}. The SAT solver solves the combination of Boolean and pseudo-Boolean constraints using the David-Putnam-Logemann-Loveland (DPLL) algorithm and suggests satisfying assignments for the Boolean variables $b$ and thus suggesting which polynomial constraints should jointly satisfied (or unsatisfied). The theory solver (PolyAR) checks the validity of the given assignments and provides an explanation of the conflict, i.e., an \textit{UNSAT certificate}, whenever a conflict is found. Each certificate is a new Boolean constraint that will be used by the SAT solver to prune the search space.
While in the lazy SMT paradigm, the PolyAR solver needs to be executed multiple times with a different set of polynomial constraints, we modify the PolyAR solver to perform all the abstraction refinement for all the polynomials as a pre-processing step. This eliminates the need to re-compute the same abstraction refinement every time the PolyAR solver is executed.
Whenever the SAT solver assigns one of the Boolean variables $b_l$ in~\eqref{eq:hybrid} to zero, then the PolyAR solver needs to guarantee that the corresponding polynomial $p_{l+m}$ satisfy $p_{l+m}(x) > 0$ or equivalently $-p_{l+m}(x) \le 0$. To eliminate the need to apply the convex abstraction refinement process for both $p_{l+m}(x)$ and $-p_{l+m}(x)$, the PolyAR solver computes the negative and positive boxes ($\mathcal{B}^N$ and $\mathcal{B}^P$) only for $p_{l+m}(x)$ and flips their usage for $-p_{l+m}(x)$.
\vspace{-3mm}
\section{NUMERICAL RESULTS}
\vspace{-3mm}
In this section, we compare the performance of PolyAR to the state-of-the-art solvers Z3 8.9 and Yices 2.6. The objective of this comparison is to study the performance on:
\begin{itemize}
\item Problems that appear naturally in parametric controller synthesis. In particular, we focus on the problem of designing stabilizing SOF controllers for LTI systems~\cite{SOFdesign}.
\item Problems that appear in non-parametric controller synthesis for non-linear systems. In particular, we focus on the problem of designing a controller for the nonlinear Duffing oscillator~\cite{MPCdesign}.
\item Additionally, we demonstrate the performance of PolyAR on designing a hybrid switching system; a problem which state-of-the-art tools are incapable of handling.
\end{itemize}
All the experiments were executed on an Intel Core i7 2.6-GHz processor with 16 GB of memory.
\vspace{-2mm}
\subsection{Static Output Feedback Controller Synthesis for Linear Time Invariant Systems}
\vspace{-2mm}
In this subsection, we assess the scalability of the PolyAR solver compared to state-of-the-art solvers on control synthesis problems. In particular, we consider the problem of synthesizing a parametric controller for the following continuous LTI system:
$$ \dot{x} = A x + B u, \qquad y = C x,$$
where $x \in \mathbb{R}^{n_A}$ is the system state, $u \in \mathbb{R}^{n_B}$ is the system control input, $y \in \mathbb{R}^{n_C}$ is the system output, and the matrices $A \in \mathbb{R}^{n_{A}\times n_{A}}$, $B \in \mathbb{R}^{n_A \times n_B}$ and $C \in \mathbb{R}^{n_{C} \times n_A}$ are the system matrices. We are interested in designing a static output feedback controller of the form:
$$ u = K y,$$
such that the resulting closed loop system:
$$ \dot{x} = (A + BKC) x,$$
is stable, i.e., the matrix $A+BKC$ is Hurwitz.
We follow the steps detailed in~\cite{SOFdesign} to pose the problem of designing the static output feedback controller as a set of polynomial constraints using the Routh-Hurwitz stability criteria. The Routh-Hurwitz stability criteria result in a set of $n_A$ polynomials in the elements of the controller matrix $K$. We consider five instances of the controller synthesis problem with the following parameters:
\begin{itemize}
\item \textbf{Example 1}: $n_A = 3, n_B = 4, n_C = 4$ which results in 3 polynomial constraints with 16 variables and max polynomial order of $4$. We restrict the elements of the controller matrix to be inside $[-4,7]$.
\item \textbf{Example 2}: $n_A = 3, n_B = 5, n_C = 5$ which results in 3 polynomial constraints with 25 variables and max polynomial order of $3$. We restrict the elements of the controller matrix to be inside $[-0.5,1]$.
\item \textbf{Example 3}: $n_A = 2, n_B = 6, n_C = 6$ which results in 2 polynomial constraints with 36 variables and max polynomial order of $3$. We restrict the elements of the controller matrix to be inside $[0,5]$.
\item \textbf{Example 4}: $n_A = 2, n_B = 7, n_C = 7$ which results in 2 polynomial constraints with 49 variables and max polynomial order of $2$. We restrict the elements of the controller matrix to be inside $[-10,0]$.
\item \textbf{Example 5}: $n_A = 5, n_B = 4, n_C = 4$ which results in 5 polynomial constraints with 16 variables and max polynomial order of $4$. We restrict the elements of the controller matrix to be inside $[-4,7]$. In addition, we want to enforce the following controller structure:
$$k_{21}\times k_{22} \times k_{23} <~0, \quad k_{21}+ k_{22} + k_{23} <-1$$,
which can be encoded using the additional SMT constraints:
\begin{align*}
&b_1 \wedge b_2 \longleftrightarrow \text{True},\\
& b_1 \rightarrow k_{21}\times k_{22} \times k_{23} < 0,\\
& b_2 \rightarrow k_{21}+ k_{22} + k_{23} < -1,
\end{align*}
where $k_{ij}$ are the elements of the controller matrix $K$.
\end{itemize}
For each of these examples, we generate random system matrices from a zero-mean normal distribution and feed them to four versions of our solver PolyAR:
\begin{itemize}
\item PolyAR + Z3 (1 thread): This version uses one instance of Z3 to analyze all the ambiguous regions.
\item PolyAR + Z3 (max threads): This version uses a separate instance of Z3 to analyze each of the ambiguous regions. All Z3 instances are running in parallel.
\item PolyAR + Yices (1 thread): This version uses one instance of Yices to analyze all the ambiguous regions.
\item PolyAR + Yices (max thread): This version uses a separate instance of Yices to analyze each of the ambiguous regions. All Yices instances are running in parallel.
\end{itemize}
We compare the execution times of these four solvers with Z3 8.9 and Yices 2.6. Table~1 shows the execution time for all the solvers. As evident by the results in Table~1, off-the-shelf solvers are incapable of solving all the five examples and they time out after one hour. On the other hand, and thanks to the abstraction refinement process, the PolyAR solver is able to solve all the instances in a few seconds, leading to $240X$ speed up in the total execution time in the PolyAR+Yices (max threads) case, evidence of the scalability of the proposed approach.
\begin{table}[t!]
Table 1: Experiment results for SOF design. The timeout is set by $3600~s$.
\begin{adjustbox}{width=\columnwidth,center}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
\multirow{3}{*}{Example} & \multicolumn{6}{c|}{Times (seconds)} \\
\cline{2-7}
& Z3 8.9 & Yices 2.6 & PolyAR+Z3 & PolyAR+Z3 & PolyAR+Yices & PolyAR+Yices\\
& & & (1 thread) & (max threads) & (1 thread) & (max threads)\\
\hline
\hline
1 & \textit{timeout} & \textit{timeout} & \textit{timeout} & $7.552$ & $\mathbf{2.405}$ & $2.442$\\
\hline
2 & \textit{timeout} & \textit{timeout} & $83.776$ & $114.453$ & $timeout$ & $\mathbf{3.766}$ \\
\hline
3 & \textit{timeout} & \textit{timeout} & $23.551$ & $23.970$ & \textit{timeout} & $\mathbf{8.725}$ \\
\hline
4 & \textit{timeout} & \textit{timeout} & $0.718$ & $0.729$ & $\mathbf{0.416}$ & $0.432$\\
\hline
5 & \textit{timeout} & \textit{timeout} & $3.636$ & $3.768$ & $0.621$ & $\mathbf{0.498}$ \\
\hline
\hline
\# Problems & $0$ & $0$ & $4$ & $5$ & $3$ & $\mathbf{5}$ \\
Solved & & & & & & \\
\hline
Total Time & \textit{timeout} & \textit{timeout} & $111.681$ & $150.472$ & $3.442$ & $\mathbf{15.863}$ \\
(seconds) & & & & & &\\
\hline
\end{tabular}
\end{adjustbox}
\vspace{-2mm}
\end{table}
In the following, we give the stabilizing controller matrices $K_{1}$, $K_{2}$, $K_{3}$, $K_{4}$, $K_{5}$, and the two Boolean variables $b_1$ and $b_2$ that PolyAR (Yices) returned for Examples 1, 2, 3, 4, and 5:
\begin{align*}
K_1&=\begin{bmatrix}
-4 & -2 & -2 & 1 \\
1 & 1 & 1 & 1 \\
1 & 1 & 1 & 1 \\
1 & 4 & 1 & 1
\end{bmatrix}, K_2=\begin{bmatrix}
0 & 1 & 0 & 0 & 0\\
0 & 0 & 0 & 0 & 0\\
0 & 0 & 0 & 1 & 0\\
0 & 0 & 0 & 0 & 0\\
0 & 0 & 0 & 0 & 0
\end{bmatrix},\\
K_3&=\begin{bmatrix}
1 & 1 & 3 & 3 & 2 & 2 \\
2 & 2 & 2 & 2 & 2 & 2 \\
2 & 0 & 0 & 2 & 2 & 5 \\
0 & 0 & 2 & 2 & 0 & 2 \\
2 & 2 & 0 & 0 & 2 & 5 \\
5 & 2 & 2 & 2 & 2 & 2
\end{bmatrix}, K_4=\begin{cases}
-8, & i=1,j=1,\\ -5, & 2\leq i,j\leq 7,
\end{cases}\\
K_5&=\begin{bmatrix}
-2 & 4 & -2 & 7 \\
1 & -3.5 & 1 & 1 \\
5 & 6 & 1 & 1 \\
1 & 6.99 & 1 & 1
\end{bmatrix},~b_1=\text{True},~b_2=\text{True}.
\end{align*}
It is easily to note that the solutions given by PolyAR (Yices) satisfies the two Boolean constraints, i.e., $k_{21}\times k_{22} \times k_{23}=-3.5 <~0$ and $k_{21}+ k_{22} + k_{23}=-1.5 <-1$.
In conclusion, PolyAR+Yices (max thread) solver outperforms all the other solvers due to the effectiveness of Yices in reasoning about problems with small volumes.
\subsection{Non-Linear Controller Design for a Duffing Oscillator}
In this subsection, we assess the scalability of PolyAR solver compared to state-of-the-art solvers on synthesizing a non-parametric controller for a Duffing oscillator reported by~\cite{MPCdesign}. The dynamics of the oscillator is given by the higher-order differential equation:
\begin{align}\label{duff1eq}
y^{(n)}\!\!\left(t\right)\!+\!\cdots\!+\!y^{(2)}\!\!\left(t\right)\!+\!2\zeta y^{(1)}\!\!\left(t\right) \!+\! y\!\left(t\right) \!+\! y\!\left(t\right)^3 \!\!=\!\! u\left(t\right),
\end{align}
where $y \in \mathbb{R}$ is the continuous state variable and $u \in \mathbb{R}$ is the control input. The parameter $\zeta$ is the damping coefficient. The objective of the control is to regulate the state to the origin. To derive the discrete-time model, forward difference approximation is used (with sampling period of $h=0.05$ time units). The resulting state space model with discrete state vector $x=[x_1,x_2,\cdots,x_n]^T=[y,y^{(1)},\cdots,y^{(n-1)}]^T \in \mathbb{R}^{n-1}$ and input $u \in \mathbb{R}$ is:
\begin{align}\label{duff2eq}
\begin{bmatrix}
x_1\\
x_2\\
\vdots\\
x_n
\end{bmatrix}^{+}&\!\!\!\!\!\!=\!\!\begin{bmatrix}
\!1\! & \!\!\!h\! & \!\!\!0\! & \!\!\!\cdots\! & \!\!\!\cdots\! & \!\!\!0\!\\
\!0\! & \!\!\!1\! & \!\!\!h\! & 0 & \!\!\!\cdots\! & \!\!\!0\!\\
\!\vdots\! & \!\!\!\vdots\! & \!\!\!\vdots\! & \!\!\!\vdots\! & \!\!\!\vdots\! & \!\!\!\vdots\!\\
\!-h\! & \!\!\!-2\zeta h\! & \!\!\!-h\! & \!\!\!\hdots\! & \!\!\!-h\! & \!\!\!1\!\!-\!\!h\!
\end{bmatrix}\!\!\!\!
\begin{bmatrix}
x_1 \\
x_2 \\
\vdots\\
x_2
\end{bmatrix}
\!\!\!+\!\!\!\begin{bmatrix}
0 \\
0 \\
\vdots\\
h
\end{bmatrix}\!\!u\!+\!\!\!\begin{bmatrix}
0 \\
0 \\
\vdots\\
\!\!-hx_1^3\!
\end{bmatrix}.
\end{align}
The previous equation
is written in the form of $x\left(k+1\right)=Ax\left(k\right)+Bu\left(k\right)+E\left(x\right)$, which includes a nonlinear term $E\left(x\right)=\begin{bmatrix} 0, \cdots, -hx_1^3\left(k\right) \end{bmatrix}^T$. Our objective is to design a non-parametric controller. To that end, we encode the controller as the solution of a feasibility problem of several constraints that capture the system dynamics, state/input constraints, and stability constraints as discussed below.
First, to enforce the stability of the resulting non-parametric controller, we consider the candidate quadratic Lyapunov function $V\left(x\right)=x^TPx$ with the symmetric positive definite matrix $P$ is a solution of the discrete-time Lyapunov equation $APA^T+P+Q=0$ and is a positive definite matrix. Thanks to the fact that $E(x)$ satisfies $\text{lim}_{\norm{x} \rightarrow 0}\frac{\norm{E\left(x\right)}}{\norm{x}}=0$ along with the Lyapunov's indirect method in~\cite{khalil}, one can directly conclude that $V(x)$ is indeed a Lyapunov function.
For simplicity, we pick $Q=I_n$, where $I_n$ is the identity matrix of size $n$.
Moreover, to ensure the smoothness of the resulting controller signals, we add additional filters in the form of high order polynomial $L(x,u) \leq 0$.
In addition, we consider the state-constraints of the form $\norm{x\left(k\right)}_{\infty} \leq 0.6$.
The final non-parametric controller is then encoded as the solution of the following feasibility problem:
\begin{align}\label{mpc4}
&\exists x_1(k),\ldots x_n(k), x_1(k+1),\ldots x_n(k+1), u(k) \nonumber\\
&\text{subject to}: \nonumber\\
&\qquad x\left(k+1\right)=Ax\left(k\right)+Bu\left(k\right)+E\left(x\right), \nonumber\\
&\qquad V\left(x\left(k+1\right)\right)-V\left(x\left(k\right)\right) \le -\epsilon, \nonumber\\
&\qquad L\left(x\left(k\right), u(k)\right) \leq 0, \nonumber \\
&\qquad \norm{x\left(k\right)}_{\infty} \leq 0.6.
\end{align}
Since the PolyAR solver only handles polynomial inequalities, hence, we transform the equality constraint
$x\left(k+1\right)=Ax\left(k\right)+Bu\left(k\right)+E\left(x\right) $ above into two inequalities $x\left(k+1\right)-Ax\left(k\right)+Bu\left(k\right)+E\left(x\right) \leq\epsilon~\wedge~x\left(k+1\right)-Ax\left(k\right)+Bu\left(k\right)+E\left(x\right) \geq -\epsilon$, where $\epsilon \in \mathbb{R}$ is a small value.
We consider three instances of the controller synthesis problem for the Duffing oscillator with the following parameters:
\begin{itemize}
\item $n=2$, $\zeta=0.3$, $x\left(0\right)=[0.4,0.1]^T$, $L\left(x\left(k\right),u\left(k\right)\right)=x_1^{11}\left(k\right)+x_2^{11}\left(k\right)-u^{10}\left(k\right)$, which results in $6$ polynomial constraints with $3$ variables and max polynomial order of $11$.
\item $n=3$, $\zeta=1.0$, $x\left(0\right)=[0.1,0.1,0.1]^T$, $L\left(x\left(k\right),u\left(k\right)\right)=x_1^{5}\left(k\right)+x_2^{5}\left(k\right)+x_3^{5}\left(k\right)+u^{5}\left(k\right)$, which results in $8$ polynomial constraints with $4$ variables and max polynomial order of $5$.
\item $n=4$, $\zeta=1.75$, $x\left(0\right)=[0.1,0.1,0.01, 0.1]^T$, $L\left(x\left(k\right),u\left(k\right)\right)=x_1^{4}\left(k\right)+x_2^{4}\left(k\right)+x_3^{4}\left(k\right)+x_4^{4}\left(k\right)-u^{4}\left(k\right)$, which results in $10$ polynomial constraints with $5$ variables and max polynomial order of $4$.
\end{itemize}
\begin{figure}
\resizebox{.45\textwidth}{!}
\centering
\begin{tabular}{ c | c | c |}
$n$ & State Space & Execution Time Evolution over time\\\hline
2 &
\raisebox{-0.5\totalheight}{\includegraphics[width=0.6\columnwidth,trim=8mm 0 15mm 0, clip]{Figure_1n2.eps}} &
\raisebox{-0.5\totalheight}{\includegraphics[width=0.6\columnwidth,trim=8mm 0 15mm 0, clip]{Figure_2n2.eps}} \\ \hline
3 &
\raisebox{-0.5\totalheight}{\includegraphics[width=0.6\columnwidth,trim=8mm 0 15mm 0, clip]{Figure_1n3.eps}} &
\raisebox{-0.5\totalheight}{\includegraphics[width=0.6\columnwidth,trim=8mm 0 15mm 0, clip]{Figure_2n3.eps}} \\ \hline
4 &
\raisebox{-0.5\totalheight}{\includegraphics[width=0.6\columnwidth,trim=8mm 0 15mm 0, clip]{Figure_1n4.eps}} &
\raisebox{-0.5\totalheight}{\includegraphics[width=0.6\columnwidth,trim=8mm 0 15mm 0, clip]{Figure_2n4.eps}} \\ \hline
\end{tabular}
}
\caption{\small{Results of controlling the Duffing oscillator with different $n$ (left) evolution of the states $x_1(k)$ and $x_2(k)$ for the solvers in the state-space, (right) evolution of the execution time of solvers during the $20$ seconds. The timeout is equal to $1s$. Trajectories are truncated once the solver exceeds the timeout limit.}}
\label{fig:duff}
\vspace{-1mm}
\end{figure}
We feed the resultant polynomial inequality constraint to PolyAR$+$Yices, PolyAR$+$Z3, Yices, and Z3. We solve the feasibility problem for $n=2$, $n=3$, and $n=4$. We set the timeout to be $1s$. Figure~\ref{fig:duff} (left) shows the state-space evolution of the controlled Duffing oscillator for different solvers for number of variables $n$ of $2,3,$ and $4$. Figure~\ref{fig:duff} (right) shows the evolution of the execution time of the solvers during the $20$ seconds. As it can be seen from Fig. 4, our solver PolyAR$+$ Yices succeeded to find a control input $u$ that regulates the state to the origin for all $n$. However, off-the-shelf solvers are incapable of solving all the three instances and they early time out after one second out of the simulated $20~seconds$.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.85\columnwidth]{Figexpmatr.eps}
\caption{\small{The trajectory that starts from an initial state $x\left(0\right)=[40,30]^T$ and reaching a final state $x\left(3\right) \in Goal$ while avoiding the obstacles. The goal and the obstacles are represented with a red and yellow rectangle, respectively.}}
\label{expmatr}
\end{figure}
\subsection{Designing Switching Signals for Continuous-Time Linear Switching Systems}
\vspace{-3mm}
In this subsection, we show how to use the PolyAR solver to successfully design a controller for a continuous-time linear switching system. In particular, we consider the following switching dynamics:
$$ \dot{x} = A_{\sigma(t)} x, \qquad \sigma(t) = \{1,2,3\},$$
with $x(t) \in \mathcal{X} \subset \mathbb{R}^2$ is the system state at time $t$ and the matrices $A_1$, $A_2$, and $A_3 \in \mathbb{R}^{2\times2}$ represents three modes for the switching system. Consider the state space in Figure~\ref{expmatr}. The objective is to design a switching signal $\sigma(t)$ that can steer the state of the system to the goal set $Goal~\subset~\mathcal{X}$ while avoiding entering the obstacle set $Obstacle~\subset\mathcal{X}$. For simplicity, we confine our attention to step-wise switching signals $\sigma(t)$. That is, we assume the switching signal $\sigma(t)$ will be constant for some amount of time $t_1, t_2, \ldots, t_L$. Our objective is then to design the switching times and the associated system mode that leads to the satisfaction of the reach-avoid specifications. To that end, we define a set of Boolean variables $b_{ij}$ such that $b_{ij}$ is equal to 1 whenever the $j$th mode is active during $t_i$. Given the initial condition of the system $x(0)$, we can use these Boolean variables to encode the problem of designing the switching signal as the following SMT constraints:
\begin{align}\label{switch1}
&\exists b_{11},\ldots,b_{13},\ldots,b_{L1}, \ldots, b_{L3}, x(1), \ldots, x(L), t_1, \ldots, t_L \nonumber\\
&\text{subject to:} \nonumber\\
&\qquad b_{11}\rightarrow x\left(1\right)=\exp{\left(A_1t_1\right)}~x\left(0\right),\nonumber\\
&\qquad b_{12}\rightarrow x\left(1\right)=\exp{\left(A_2t_1\right)}~x\left(0\right),\nonumber\\
&\qquad b_{13}\rightarrow x\left(1\right)=\exp{\left(A_3t_1\right)}~x\left(0\right),\nonumber\\
&\qquad b_{11}+b_{12}+b_{13}=1,\nonumber\\
&\qquad \qquad\qquad \vdots\nonumber\\
&\qquad b_{L1}\rightarrow x\left(L\right)=\exp{\left(A_1t_L\right)}~x\left(L-1\right),\nonumber\\
&\qquad b_{L2}\rightarrow x\left(L\right)=\exp{\left(A_2t_L\right)}~x\left(L-1\right),\nonumber\\
&\qquad b_{L3}\rightarrow x\left(L\right)=\exp{\left(A_3t_L\right)}~x\left(L-1\right),\nonumber\\
&\qquad b_{L1}+b_{L2}+b_{L3}=1,\nonumber\\
&\qquad x\left(1\right),\ldots,x\left(L-1\right) \notin Obstacle,\nonumber\\
&\qquad x\left(L\right) \in Goal,
\end{align}
where the constraint $b_{i1} + b_{i2} + b_{i3} = 1$ is a pseudo-Boolean constraint that enforces the consistency between the Boolean variables such that only one of the three modes $A_1,A_2$ and $A_3$ can be selected during the period $t_{j-1} < t \le t_j$.
Since PolyAR solver only handles polynomial inequalities, we approximate the exponential matrix $\exp\left(A_it_j\right)\approx I_2+t_jA_i+\frac{t_j^2A_i^2}{2}+\frac{t_j^3A_i^3}{6}$, $i,j=1,\cdots,L$, where $I_2$ is the identity matrix of size $2$ and $A_i^n=A_i\times \cdots \times A_i$. Furthermore, we transform the equality $x\left(i\right)=\exp{\left(A_it_j\right)}~x\left(i-1\right)$, $i,j=1,\cdots,L$, into two inequalities $x\left(i\right)-\exp{\left(A_it_j\right)}~x\left(i-1\right) \leq \epsilon \wedge x\left(i\right)-\exp{\left(A_it_j\right)}~x\left(i-1\right) \geq -\epsilon$, where $\epsilon \in \mathbb{R}$ is a small value.
In our experiments, we pick the horizon $L = 3$ and the modes $A_1=\begin{bmatrix}-1, 2\\-2,-2 \end{bmatrix}$, $A_2=\begin{bmatrix}-1, 3\\-3,-1 \end{bmatrix}$, and $A_3=\begin{bmatrix}0, 2\\-2,0 \end{bmatrix}$, and we start with an initial state $x\left(0\right)=[40,30]^T$. We restrict the states to be inside $\mathcal{X} \in [-100,100]$. We feed the resultant polynomial inequality constraint to PolyAR+Yices, and as it can be seen from Fig.\ref{expmatr}, the solver succeeded to find the right modes ($b_{11}=b_{22}=b_{33}=1$) and the necessary times $t_1=0.391s$, $t_2=0.5s$, and $t_3=0.25s$ that ensures that $x\left(3\right)$ reaches a $Goal$ while the intermediate states $x\left(2\right),x\left(1\right)$ avoid $Obstacles$. In addition, we remark that the trajectory between $x\left(0\right)$ and $x\left(2\right)$ is making its way to the equilibrium point $[0,0]^T$. This is explained by the fact that the matrices $A_1$ and $A_2$ are stables. Our solver computes the necessary time $t_3$ that ensures that the final state $x\left(3\right) \in~Goal$ and does not converge to the equilibrium point.
\vspace{-3mm}
|
\section{Introduction}
\label{s.intro}
The standard parton correlation functions of QCD, such as collinear parton distribution functions (PDFs) and fragmentation functions (FFs), are being utilized in an increasingly diverse range of phenomenological applications.
Beyond their traditional role in predicting new high energy phenomena, they also enter frequently into the study of more complex and extended objects like transverse momentum dependent (TMD) PDFs and FFs and generalized parton distributions (GPDs), where they are needed to understand the transition between different factorization regions.
Both TMDs and GPDs are central to the study of the nonperturbative parton structure of hadrons, and understanding how they encapsulate their longitudinal and transverse features will be critical to current experimental programs at Jefferson Lab and elsewhere, as well as to the future Electron-Ion Collider.
These considerations provide one of the main motivations for the study of collinear PDFs and FFs in this paper.
The great value of PDFs and FFs extracted from global QCD data analysis lies with their predictive power, or ``universality''.
However, the translation from experimental data to quark and gluon operator structures is a challenging inverse problem.
It is not possible to exactly constrain parton correlations from data alone since this connection involves nontrivial convolution integrals in a factorization formalism (whose accuracy itself is difficult to quantify in any given instance), and because of the limited quantity of available data.
The complexity of the inverse problem is also magnified by the number of flavor degrees of freedom involved.
Nevertheless, assessing and maximizing the universality of collinear PDFs and FFs is crucial given the increasingly broad scenarios where they are used.
A major focus in the current effort by the Jefferson Lab Angular Momentum (JAM) Collaboration is therefore to both test and broaden the predictive power of parton correlation functions.
This is achieved through a Bayesian inference procedure in which PDFs and FFs are extracted simultaneously, and the uncertainty quantification associated with particular parametrizations of parton correlation functions is given in terms of a Bayesian posterior distribution.
To test universality, the system of equations relating observables to parton correlation functions must of course exceed the total number of correlation functions involved --- a minimum requirement is that the parton correlation functions be overconstrained by the data in the fit.
Of course, realizing this in practical analyses requires that all parton correlation functions be truly fitted simultaneously.
This is a major numerical and technological challenge, and traditionally PDFs and FFs have thus been extracted in separate procedures.
However, simultaneous fits can be achieved with the Bayesian Monte Carlo approach, and have been implemented recently in the JAM17~\cite{Ethier:2017zbq} analysis of helicity PDFs, and in the JAM19~\cite{Sato:2019yez} analysis of unpolarized PDFs and FFs.
The same basic methodology was also applied in the three-dimensional JAM3D20~\cite{Cammarota:2020qcw} study, in the first combined analysis of TMD observables that satisfies the overconstraining criterion.
In this paper, we extend the previous work by performing the first simultaneous and overconstrained fit of unpolarized PDFs and FFs that utilizes both charged hadron production in semi-inclusive deep-inelastic scattering (SIDIS) and single-inclusive $e^+ e^-$ annihilation (SIA).
This is partly motivated by a number of recent observations associated with the study of TMD PDFs.
For example, significant tension has recently been found between fits performed with standard sets of PDFs and FFs and fixed order perturbative QCD calculations in processes including SIDIS~\cite{Gonzalez-Hernandez:2018ipj, Wang:2019bvb}, Drell-Yan (DY)~\cite{Bacchetta:2019tcu}, and SIA into wide-angle hadron pairs~\cite{Moffat:2019pci}.
A number of suggested solutions and explanations have been proposed to account for this, including a possible need for power suppressed corrections~\cite{Liu:2019srj} at the moderate scales of most SIDIS experiments.
However, more tests of the limits of applicability of standard collinear factorization are needed before it is possible to draw firm conclusions.
Given that the majority of data used to constrain collinear correlation functions (both PDFs and FFs) are either highly inclusive or exist are at very high scales, or both, it is perhaps not surprising that tension arises when these are evolved downward and used to make predictions at lower scales and for highly differential observables.
Indeed, there have been few tests that $Q^2$-scaling, a hallmark of the collinear perturbative regime, actually holds to a reasonable approximation in SIDIS measurements at moderate $Q^2$.
Our hope is that the new combined fit, which we refer to as ``JAM20-SIDIS'', will help to shed light on this and similar issues in the future.
In Sec.~\ref{s.method} we begin the discussion by summarizing the methodology used in our simultaneous Monte Carlo analysis, including the parametrizations used for the distributions and the multi-step Bayesian inference algorithm.
Details of the data sets included in the fit are summarized in \sref{datasets}, while in \sref{uni} we discuss the criteria for universality and how these are met in this analysis.
A detailed discussion of the numerical results is given in \sref{analysis}, where we present the fitted PDFs and FFs, as well as detailed comparisons of data to theory.
Finally, in \sref{conclusion} we summarize our conclusions and discuss the implications of our analysis.
Some formulas for SIDIS cross sections and structure functions are collected in \aref{sidisrev}.
\section{Theoretical framework}
\label{s.method}
In this section we give an overview of the theoretical framework on which our analysis is based, including the observables to be fitted, the parametrizations used for the PDFs and FFs, details of the perturbative QCD setup, and Bayesian inference strategy employed.
\subsection{Observables and factorization}
In this analysis we work in standard collinear factorization~\cite{Collins:1989gx, Ellis:1991qj, Collins:2011qcdbook}, in which QCD cross sections are separated into perturbatively calculable partonic hard factors convoluted with nonperturbative PDFs and/or FFs.
We perform calculations of all observables consistently to order $\alpha_s$ in the QCD coupling.
Details of the basic theoretical setups for the inclusive DIS, inclusive Drell-Yan lepton-pair production and SIA reactions are provided in the literature~\cite{Ellis:1991qj, Devenish:2004pb}, and will not be repeated here.
However, since SIDIS is a comparatively novel addition to global QCD analyses, we review it in more detail in \aref{sidisrev}.
The processes considered in the present analysis can be summarized as follows:
\begin{align}
\ell + N &{} \to \ell + X ,
\qquad &{}\text{inclusive DIS} \,
\nonumber\\
\ell + N &{} \to \ell + h^\pm + X ,
\qquad &{}\text{semi-inclusive DIS} \,
\nonumber\\
N_1 + N_2 &{} \to \ell^+ + \ell^- + X ,
\qquad &{} \text{Drell-Yan lepton-pair production}\,
\nonumber\\
\ell^+ + \ell^- &{} \to h^\pm + X ,
\qquad &{}\text{single-inclusive annihilation}\,
\nonumber
\end{align}
where $h^\pm$ represent charged pions, kaons, or unidentified hadrons, and the nucleon $N$ (or~$N_{1,2}$) in the initial state can be either a proton or a neutron (in practice, deuteron).
Within the framework of collinear factorization, the cross sections for each of these processes can be written schematically as convolutions of hard functions and the nonperturbative parton distribution and fragmentation functions,
\begin{align}
\frac{\diff{\sigma_\text{\tiny DIS}}}{\diff{Q^2}{} \diff{\ensuremath{x_{\mbox{\tiny \rm Bj}}}}{}}
&{}= \sum_{i}
\mathcal{H}_i^\text{\tiny DIS} \otimes f_i \, ,
\qquad &{}
\label{e.DIS} \\
\frac{\diff{\sigma_\text{\tiny SIDIS}}{}}{\diff{Q^2}{} \diff{\ensuremath{x_{\mbox{\tiny \rm Bj}}}}{} \diff{z_h}{} }
&{}= \sum_{ij}
\mathcal{H}_{ij}^\text{\tiny SIDIS} \otimes f_i \otimes D_j^h \, ,
\qquad &{}
\label{e.SIDIS} \\
\frac{\diff{\sigma_\text{\tiny DY}}{}}{\diff{Q^2}{} \diff{\ensuremath{x_{\mbox{\tiny \rm F}}}}{}}
&{}= \sum_{ij}
\mathcal{H}_{ij}^\text{\tiny DY} \otimes f_i \otimes f_j\, ,
\qquad &{}
\label{e.DY} \\
\frac{\diff{\sigma_\text{\tiny SIA}}{}}{\diff{Q^2}{} \diff{z_h}{}}
&{}= \sum_{j}
\mathcal{H}_j^\text{\tiny SIA} \otimes D_j^h\, ,
\qquad &{}
\label{e.SIA}
\end{align}
where the symbols $\otimes$ represent the convolution integrals in longitudinal momentum fractions of the hard scattering functions $\mathcal{H}_{ij}$ and the PDFs $f_i$ and FFs $D_j^h$ for parton flavors $i, j$ (see the Appendix).
In each process, $Q$ represents the hard scale given by the photon virtuality, $Q \gg$ hadron masses, which allows the observables to be factorized into the short-distance perturbative and long-distance nonperturbative parts.
For the inclusive DIS and SIDIS processes,
\begin{equation}
\ensuremath{x_{\mbox{\tiny \rm Bj}}} = \frac{Q^2}{2 p \cdot q}
\end{equation}
is the usual Bjorken scaling variable, while for the DY process the analogous scaling variables are defined as
\begin{equation}
x_1 = \frac{Q^2}{2 p_1 \cdot q}\, , \qquad
x_2 = \frac{Q^2}{2 p_2 \cdot q}\, ,
\end{equation}
where $p_1$ and $p_2$ denote the incoming hadron momenta, with the Feynman scaling variable given by
\begin{equation}
\ensuremath{x_{\mbox{\tiny \rm F}}} = x_1 - x_2.
\end{equation}
In the DY center of mass frame, and in the limit of negligible hadron masses ($\ll Q$), the virtual photon rapidity can be written in terms of $x_1$ and $x_2$ as
\begin{equation}
y = \frac12 \ln\frac{x_1}{x_2}.
\end{equation}
For the processes involving fragmentation to a hadron $h$ in the final state, we have
\begin{equation}
\hspace*{4cm}
z_h = \frac{p_h \cdot p}{q \cdot p}
\hspace*{4cm} \textrm{[SIDIS]}
\end{equation}
for SIDIS in \eref{SIDIS}, while
\begin{equation}
\hspace*{3.6cm}
z_h = \frac{2 p_h \cdot q}{Q^2}
\hspace*{3.8cm} \textrm{[SIA]}
\end{equation}
for SIA in \eref{SIA}.
\begin{comment}
All perturbative hard coefficients $\mathcal{H}_i$ in Eqs.~(\ref{e.DIS})--(\ref{e.SIA}) are calculated to order $\alpha_s(\mu)$ and in the $\ensuremath{\overline{\rm MS}}$ renormalization scheme, with $\mu$ fixed to the photon virtuality $Q$ in all cases.
The DGLAP $Q^2$ evolution equations are implemented to next-to-leading logarithmic accuracy, and in the running coupling we set $\alpha_s(M_Z) = 0.118$ at the $Z$-boson mass.
The flavor transitions are implemented according to the flavor scheme discussed in Ref.~\cite{Vogt:2004ns}, with the charm and bottom quark masses $m_c=1.28$~GeV and $m_b=4.18$~GeV, respectively.
\end{comment}
\subsection{Perturbative QCD and numerical setups}
For our numerical analysis we make use of Mellin space techniques to enable fast evaluations of observables needed for the Bayesian analysis.
In particular, we solve the DGLAP evolution equations analytically in Mellin space~\cite{Vogt:2004ns}, which allows one to effectively render high-dimensional momentum space convolutions from process-specific factorization theorems, along with the integrals in the DGLAP equations, in the form of lower-dimensional inverse Mellin transforms.
For example, for the inclusive DIS observables one can write schematically
\begin{align}
\frac{\diff \sigma_\text{\tiny DIS}}{\diff Q^2 \diff \ensuremath{x_{\mbox{\tiny \rm Bj}}}}
= \sum_{i}
\frac{1}{2\pi i} \int dN\, \ensuremath{x_{\mbox{\tiny \rm Bj}}}^{-N}\,
\widetilde{\cal H}^\text{\tiny DIS}_i(N,\mu)\,
U^{\rm S}_{ij}(N,\mu,\mu_0)\,
\widetilde{f}_j(N,\mu_0),
\end{align}
where $N$ here is the conjugate variable to $\ensuremath{x_{\mbox{\tiny \rm Bj}}}$, $\widetilde{f}_j(N,\mu_0)$ is the Mellin moment of the PDF $f_j(x,\mu_0)$, defined by
\begin{equation}
\label{e.fNdef}
\widetilde{f}_j(N,\mu_0)
= \int_0^1 dx\, x^{N-1}\, f_j(x,\mu_0),
\end{equation}
and $\widetilde{\cal H}^\text{\tiny DIS}_i(N,\mu)$ is the corresponding moment of the partonic DIS cross section.
The analytic solution for the DGLAP evolution is entirely encoded in the evolution matrix $U^{\rm S}_{i,j}$ that evolves the moments $\widetilde{f}_j(N,\mu_0)$ of the PDFs from a given input scale $\mu_0$ to the relevant DIS hard scale $\mu=Q$.
A similar expression can be written for the SIA cross section,
\begin{align}
\frac{\diff \sigma_\text{\tiny SIA}}{\diff Q^2 \diff z_h}
= \sum_{ij}
\frac{1}{2\pi i} \int dM\, z_h^{-M}\,
\widetilde{\cal H}^\text{\tiny SIA}_i(M,\mu)\,
U^{\rm T}_{ij}(M,\mu,\mu_0)\,
\widetilde{D}_j^h(M,\mu_0)
\end{align}
where $M$ is the Mellin conjugate variable for $z_h$, $\widetilde{D}_j^h(M,\mu_0)$ is the moment of the FF, and $\widetilde{\cal H}^\text{\tiny SIA}_i$ is the moment of the partonic SIA cross section.
The superscripts $\rm S$ and $\rm T$ in the evolution matrix distinguish between the spacelike and timelike evolution for the PDFs and FFs, respectively, which are encoded in the corresponding DGLAP splitting kernels.
The same procedure can be extended for the case of SIDIS, which gives
\begin{align}
\frac{\diff \sigma_\text{\tiny SIDIS}}{\diff Q^2 \diff \ensuremath{x_{\mbox{\tiny \rm Bj}}} \diff z_h}
= &\sum_{ijkl}
\frac{1}{(2\pi i)^2}
\int dN\, \ensuremath{x_{\mbox{\tiny \rm Bj}}}^{-N}
\int dM\, z_h^{-M}\,
\widetilde{\cal H}^\text{\tiny SIDIS}_{ik}(N,M,\mu)
\notag\\
\times
&\ U^{\rm S}_{ij}(N,\mu,\mu_0)\,
\widetilde{f}_j(\mu_0)\,
U^{\rm T}_{kl}(M,\mu,\mu_0)\,
\widetilde{D}_j^h(M,\mu_0).
\end{align}
For the case of the Drell-Yan process, a special treatment is required since the Mellin moments for the partonic cross sections are not known.
For this we employ the strategy developed by Stratmann and Vogelsang~\cite{Stratmann:2001pb}, where by the Mellin moments are numerically pre-calculated and used as lookup tables during the analysis.
The resulting expression can be written schematically as
\begin{align}
\frac{\diff \sigma_\text{\tiny DY}}{\diff Q^2 \diff \ensuremath{x_{\mbox{\tiny \rm F}}}}
= &\sum_{ijkl}
\frac{1}{(2\pi i)^2}
\int dN\,
\int dM\,
\widetilde{\cal H}^\text{\tiny DY}_{ik}(N,M,\mu)
\notag\\
\times
&\ U^{\rm S}_{ij}(N,\mu,\mu_0)\,
\widetilde{f}_j(\mu_0)\,
U^{\rm S}_{kl}(M,\mu,\mu_0)\,
\widetilde{f}_l(\mu_0),
\end{align}
where the relevant inverse Mellin factors $x_1^{-N}$ and $x_2^{-M}$ arising from the scaling variables $x_1$ and $x_2$ for the incident nucleons $N_1$ and $N_2$, respectively, in \eref{DY} are integrated numerically with the hard scattering cross section and contained inside $\widetilde{\cal H}^\text{\tiny DY}_{ik}(N,M,\mu)$.
The analytic solutions for the evolution matrices are computed at next-to-leading logarithmic accuracy using splitting kernels up to ${\cal O}(\alpha_s^2)$ and the truncated solution for the single evolution operators (see Ref.~\cite{Vogt:2004ns} for details).
We employ the zero-mass variable flavor scheme for solving the DGLAP evolution equations, setting the input scale for the PDFs and FFs at $\mu_0=m_c$.
The numerical values for the mass thresholds are taken from the PDG values in the $\overline{\rm MS}$ scheme~\cite{Zyla:2020zbs}:
$m_c = 1.28$~GeV and $m_b = 4.18$~GeV.
The strong coupling is evolved numerically using the QCD beta function up to ${\cal O}(\alpha_s^2)$, using the boundary condition $\alpha_s(M_Z) = 0.118$ at the $Z$ boson mass, $M_Z=91.18$~GeV.
Finally, all the process specific hard coefficients are computed at fixed next-to-leading order in pQCD, which are available in the literature~\cite{Floratos:1981hs, Kretzer:2000yf, Stratmann:2001pb, Anastasiou:2003yy}.
\subsection{Parametrization of nonperturbative functions}
\label{ss.param}
For the nonperturbative parton distribution and fragmentation functions we use standard parametrizations that have been utilized in the literature.
Namely, for the dependence on the parton momentum fraction $x$ of the PDF $f(x)$ we use the template function
\begin{equation}
f(x,\mu_0)\ \to\ T\parz{x;\boldsymbol{a}} = {\cal M} \frac{x^\alpha\parz{1-x}^\beta\parz{1+\gamma\sqrt{x}+\delta{x}}}{\int_0^1\diff{x}\, x^{\alpha+1}\parz{1-x}^\beta\parz{1+\gamma\sqrt{x}+\delta{x}}},
\label{e.template1}
\end{equation}
where $\boldsymbol{a}=\{{\cal M},\alpha,\beta,\gamma,\delta\}$ is a vector containing the shape parameters ($\alpha$, $\beta$, $\gamma$, and $\delta$) and a normalization coefficient (${\cal M}$) to be fitted.
The integral in the denominator ensures that the value of the normalization coefficient ${\cal M}$ is equal to the second moment ($x$-weighted integral) of the function $T(x;\boldsymbol{a})$.
For fitting the PDFs, we assume isospin symmetry to relate the PDFs in the neutron, $f_{i/n}(x)$, to those in the proton, $f_{i/p}(x) \equiv f_i(x)$, switching the $u \leftrightarrow d$ and $p \leftrightarrow n$ labels for the light quark flavors, and taking the PDFs for other flavors equal for the proton and neutron.
In practice, we parametrize the valence $u$ and $d$ quark distributions,
$u_v \equiv f_u - f_{\bar u}$
and $d_v \equiv f_d - f_{\bar d}$,
directly using the template function (\eref{template1}). The gluon distribution, $g \equiv f_g$, is also directly parametrized per \eref{template1}.
For the sea quark and antiquark distributions, we use five functions parametrized as in \eref{template1}.
These are a flavor symmetric sea function ($S$) that dominates at very low $x$ and flavor specific functions ($q_0(\bar{q}_0)$) for the $s$, $\bar{u}$, $\bar{d}$, and $\bar{s}$ that take into account the possible nonperturbative origin of the sea.
The distributions for $s$, $\bar{u}$, $\bar{d}$, and $\bar{s}$ are constructed from these according to:
$q(\bar q) \equiv f_{q(\bar q)} = S + q_0(\bar q_0)$.
Note that $s$ and $\bar{s}$ are parametrized separately because their contributions to the $K^+$ and $K^-$ SIDIS cross sections differ.
We do not fit the charm and bottom PDFs, and their contributions are generated purely from the DGLAP evolution.
In total there are 8 parametrized PDF functions being fitted.
For the valence quark PDFs $u_v$ and $d_v$ and the nonperturbative sea components $\bar{u}_0$ and $\bar{d}_0$, we use the four shape parameters as in \eref{template1}; for all other distributions we set the $\gamma$ and $\delta$ parameters to zero. This gives 24 free shape parameters and 8 free normalization parameters.
The number of free parameters is further reduced by valence number sum rules, which constrain the normalization parameters $\cal M$ for the $u_v$, $d_v$, and $s-\bar s$ distributions, whose lowest moments are required to be 2, 1, and zero, respectively.
The normalization for the gluon PDF is determined using the momentum sum rule.
With these constraints, there is a total of 28 free parameters for the PDFs.
For the $z$ dependence of FFs, the functional form follows a similar template,
\begin{equation}
D(z,\mu_0)\ \to\ T\parz{z;\boldsymbol{a}} = {\cal M} \frac{z^\alpha\parz{1-z}^\beta\parz{1+\gamma\sqrt{z}+\delta{z}}}
{\int_0^1\diff{z}\, z^{\alpha+1}\parz{1-z}^\beta\parz{1+\gamma\sqrt{z}+\delta{z}}},
\label{e.template2}
\end{equation}
where again the integral in the denominator ensures that the coefficient ${\cal M}$ corresponds to the second moment ($z$-weighted integral) of the function.
In addition to the fragmentation to pions and kaons studied in earlier JAM analyses of SIA and SIDIS data~\cite{Sato:2016wqj, Sato:2019yez}, here we consider also the inclusive production of unidentified charged hadrons, $h^\pm$.
Accounting for unidentified hadrons can be implemented in two ways.
First, the hadron FFs can be fit independently from those for pions and kaons, as preferred by the NNPDF Collaboration~\cite{Bertone:2018ecm}.
Alternatively, one can take advantage of existing knowledge of specified hadron FFs and add a fitted residual correction to their sum.
Such an approach was adopted by de Florian, Sassot, and Stratmann (DSS)~\cite{deFlorian:2007ekg}, for example, in which a residual correction was fitted to the sum of previously obtained pion, kaon, and proton fragmentation functions.
In our analysis we follow the latter approach, but include only the pion and kaon FFs, so that the residual term $D_i^{\rm res^+}$ parametrizes the difference between the total hadron FF $D_i^{h^+}$ and the $D_i^{\pi^+}$ and $D_i^{K^+}$ functions,
\begin{equation}
D_i^{h^+} = D_i^{\pi^+} + D_i^{K^+} + D_i^{\rm res^+}.
\end{equation}
To reduce the total number of residual FFs being fit, we assume SU(3) flavor symmetry for light quarks and antiquarks,
\begin{subequations}
\begin{eqnarray}
D_q^{\rm res^+}=D_u^{\rm res^+} &=& D_d^{\rm res^+} = D_s^{\rm res^+}, \ \ \ \ \ \ \
\label{e.rprel} \\
D_{\bar{q}}^{\rm res^+}=D_{\bar{u}}^{\rm res^+} &=& D_{\bar{d}}^{\rm res^+} = D_{\bar{s}}^{\rm res^+}, \label{e.rmrel}
\end{eqnarray}
\end{subequations}
where $D_q^{\rm res^+}$ and $D_{\bar{q}}^{\rm res^+}$ are parametrized per the template (\eref{template2}). To allow for differentiation between the residual FFs for light quarks and antiquarks, we leave ${\cal M}$ and $\beta$ for $D_{\bar{q}}^{\rm res^+}$ as free parameters, but fix $\alpha$, $\gamma$, and $\delta$ to be the same as for $D_{q}^{\rm res^+}$.
This achieves a similar constraint on the parameters as the condition used by DSS~\cite{deFlorian:2007ekg},
$2D_{\bar{q}}^{\rm res^+} = \parz{1-z}^{\beta'} D_{q+\bar{q}}^{\rm res^+}$.
For the pion FFs, $D_i^{\pi^+}$, we reduce the number of fitted functions by grouping the light quarks into ``favored'' (valence) and ``unfavored'' (non-valence) flavors,
\begin{subequations}
\begin{align}
D_{\rm fav}^{\pi^+} &= D_u^{\pi^+} = D_{\bar{d}}^{\pi^+},
\label{e.pifav}
\\
D_{\rm unf}^{\pi^+} &= D_d^{\pi^+} = D_s^{\pi^+} = D_{\bar{u}}^{\pi^+} = D_{\bar{s}}^{\pi^+},
\label{e.piunfav}
\end{align}
\end{subequations}
where $D_{\rm fav}^{\pi^+}$ and $D_{\rm unf}^{\pi^+}$ are parametrized as in \eref{template2}. For the parameters of the kaon FFs, $D_i^{K^+}$, we equate the ``unfavored'' flavors,
\begin{equation}
D_{\rm unf}^{K^+} = D_d^{\pi^+} = D_s^{\pi^+} = D_{\bar{u}}^{\pi^+} = D_{\bar{d}}^{\pi^+},
\label{e.Kunfav}
\end{equation}
but leave the favored FFs $D_u^{K^+}$ and $D_{\bar{s}}^{K^+}$ independent.
Here $D_{\rm unf}^{K^+}$, $D_u^{K^+}$, and $D_{\bar{s}}^{K^+}$ are parametrized per \eref{template2}.
For the heavier flavors, we assume the charm and bottom quark and antiquark FFs to be equivalent, $D_c^{h^+} = D_{\bar{c}}^{h^+}$ and $D_b^{h^+} = D_{\bar{b}}^{h^+}$ for $h=\pi,K,{\rm res}$, with $D_c^{h^+}$ and $D_b^{h^+}$ parametrized per \eref{template2}.
Finally, the gluon FFs $D_g^{h^+}$ for $h=\pi,K,{\rm res}$ are also parametrized according to \eref{template2}.
We use charge conjugation symmetry to relate FFs for opposite charges by
\begin{equation}
D_q^{h^+} = D_{\bar{q}}^{h^-},
\end{equation}
where $h=\pi,K,{\rm res}$.
This results in 5 fitted functions for pions and residual hadrons, and 6 for kaons.
At this point, there are 17 shape parameters and 5 normalization parameters for residual hadrons, 20 shape parameters, and 5 normalization parameters for pions, and 24 shape parameters and 6 normalization parameters for kaons.
The number of shape parameters is reduced further because throughout the fitting procedure, the parameters $\gamma$ and $\delta$ for the gluon, charm, and bottom FFs are fixed at zero.
In the end there are 16 free parameters to be fitted for residual charged hadron FFs, 19 free FF pion parameters, and 24 free parameters for the kaon FFs.
Together with the 28 PDF parameters, we have a total of 87 free parameters for the fitted functions.
In addition, there are also 42 free parameters associated with normalization of various data sets, making for a total of 129 free parameters to be fitted in the analysis.
\subsection{Bayesian inference}
\label{ss.MCfit}
Our methodology for extracting nonperturbative PDFs and FFs is based on the general premise of Bayesian inference.
Namely, we use Bayes' theorem to define a multivariate probability distribution ${\cal P}$ for the shape parameters characterizing the PDFs and FFs (the posterior) at a given input scale $\mu_0$,
\begin{align}
\label{e.Pdef}
{\cal P}(\boldsymbol{a}|{\rm data}) \sim {\cal L} (\boldsymbol{a},{\rm data})\,
\pi(\boldsymbol{a}),
\end{align}
where $\cal L$ is a standard Gaussian likelihood function,
\begin{align}
{\cal L}\parz{\boldsymbol{a},{\rm data}}
= \exp\parz{-\frac{1}{2}\chi^2\parz{\boldsymbol{a},{\rm data}}},
\end{align}
with the $\chi^2$ function defined by
\begin{equation}
\chi^2(\boldsymbol{a}) = \sum_{i,e}
\bigg( \frac{d_{i,e} - \sum_k r_e^k \beta^k_{i,e} - T_{i,e}(\boldsymbol{a})/N_e}
{\alpha_{i,e}}
\bigg)^2
+ \sum_k \big( r^k_e \big)^2
+ \left( \frac{1-N_e}{\delta N_e} \right)^2.
\label{e.chi2}
\end{equation}
Here, $d_{i,e}$ is the value of the $i$-th data point for the experimental dataset $e$, with $T_{i,e}$ the theoretical prediction for the data point; $\alpha_{i,e}$ is the uncorrelated systematic and statistical uncertainty for each data point added in quadrature; $\beta^k_{i,e}$ is the $k$-th source of point-to-point correlated systematic uncertainties for the $i$-th bin of dataset $e$, with $r^k_e$ the related weight; and $N_e$ and $\delta{N}_e$ are the normalization and normalization uncertainty for each data set, respectively.
In Eq.~(\ref{e.Pdef}), $\pi(\boldsymbol{a})$ is the prior distribution for the set of parameters $\boldsymbol{a}$, which is used as input for a given fit to the data.
In principle, given the Bayesian posterior distribution, one can estimate confidence regions for a generic observable ${\cal O}$ (such as a PDF or a function of PDFs or FFs) by integrating over an $d$-dimensional parameter space,
\begin{subequations}
\label{e.exp_var1}
\begin{align}
{\rm E}[\mathcal{O}]
&=\int\diff^d\boldsymbol{a}\, \mathcal{P}\parz{\boldsymbol{a}|{\rm data}}
\mathcal{O}\parz{\boldsymbol{a}},
\label{e.exp1} \\
{\rm V}[\mathcal{O}]
&=\int\diff^d\boldsymbol{a}\, \mathcal{P}\parz{\boldsymbol{a}|{\rm data}}
\parz{\mathcal{O}\parz{\boldsymbol{a}}-{\rm E}[\mathcal{O}]}^2,
\label{e.var1}
\end{align}
\end{subequations}
where E and V are the expectation value and variance of the observable ${\cal O}$, respectively.
Due to the significant numerical expense of evaluating the likelihood function, the explicit usage of Eqs.~(\ref{e.exp_var1}) is often not practical.
Instead, a more efficient option is to build Monte Carlo parameter samples $\{\boldsymbol{a}_k; k=1,\ldots,n\}$, which contain all parameters, including the $N_e$ from \eref{chi2}, that are faithfully distributed according to the posterior distribution.
These can in turn be used to evaluate the integrals in Eqs.~(\ref{e.exp_var1}) as Monte Carlo sums,
\begin{subequations}
\label{e.exp_var2}
\begin{align}
{\rm E}[\mathcal{O}]&=\frac{1}{n}\sum_{k=1}^n\mathcal{O}\parz{\boldsymbol{a}_k},
\label{e.exp2} \\
{\rm V}[\mathcal{O}]&=\frac{1}{n}\sum_{k=1}^n\parz{\mathcal{O}\parz{\boldsymbol{a}_k}-{\rm E}[\mathcal{O}]}^2.
\label{e.var2}
\end{align}
\end{subequations}
\begin{figure}[t!]
\centering
\includegraphics[width=\textwidth]{workflow.pdf}
\vspace*{-2.5cm}
\caption{Schematic illustration of the multi-step workflow employed in our simultaneous Monte Carlo analysis. Each box represents a collection of Monte Carlo samples associated with a specific nonperturbative hadronic structure (PDFs, FFs). The vertical arrows indicate the inclusion of additional datasets from which new optimized Monte Carlo samples (posteriors) are generated as input (priors) for the next step.\\}
\label{f.workflow}
\end{figure}
Our Monte Carlo sampling strategy is based on data resampling methodology, whereby multiple maximum likelihood optimizations are carried out.
Each optimization consists of taking a random point in parameter space and fitting the parameters to data that have been distorted away from the central values by Gaussian shifts within the quoted uncertainties.
To build the Monte Carlo samples, we use the multi-step strategy developed in Ref.~\cite{Sato:2019yez}, where the PDF and FF parameters are pre-optimized to minimize evaluating the likelihood in parameter regions that are strongly disfavored.
To that end we start by first considering PDF and FF parameters separately using flat priors, with the resulting samples from each type of hadron structure combined at a later stage to build new prior samples for the final runs.
The workflow is illustrated in \fref{workflow}, where each step is represented as vertical arrows that accumulate additional experimental data from the previous step, with the posterior samples at each step becoming the priors for the subsequent step.
This strategy allows the samples to become more optimized and avoids unnecessary likelihood evaluations in regions of parameters space by disfavoring those regions in earlier stages of the multi-step chain.
\section{Data Sets}
\label{s.datasets}
The data sets used in the present analysis include the primary electromagnetic processes that traditionally have been used in global QCD analyses, namely, inclusive DIS, Drell-Yan lepton-pair production (which constrain PDFs), SIA (which constrains FFs), and SIDIS (which constrains both PDFs and FFs).
The inclusive DIS data are measurements of the $F_2(\ensuremath{x_{\mbox{\tiny \rm Bj}}},Q^2)$ structure function performed by the BCDMS~\cite{Benvenuti:1989rh, Benvenuti:1989fm} and New Muon Collaborations~\cite{Arneodo:1996qe, Arneodo:1996kd} at CERN, and from experiments at SLAC~\cite{Whitlow:1991uw}, as well as from reduced electron and positron cross sections from the H1 and ZEUS Collaborations~\cite{Abramowicz:2015mha} at DESY.
These include both proton~\cite{Benvenuti:1989rh, Whitlow:1991uw, Arneodo:1996qe} and deuteron~\cite{Benvenuti:1989fm, Whitlow:1991uw, Arneodo:1996kd} targets, and with both neutral and charged current probes~\cite{Abramowicz:2015mha}.
For the kinematics we implement cuts of $W^2 > 10$~GeV$^2$ and $Q^2 > m_c^2$,
where $W^2 = M^2 + Q^2 (1-\ensuremath{x_{\mbox{\tiny \rm Bj}}})/\ensuremath{x_{\mbox{\tiny \rm Bj}}}$, in order to select DIS data that can be fitted within leading power factorization.
For Drell-Yan lepton-pair production data we use differential cross section measurements $\diff^2\sigma_{\mbox{\tiny DY}}/\diff{Q}\diff{\ensuremath{x_{\mbox{\tiny \rm F}}}}$ by the E866/NuSea Collaboration~\cite{Hawker:1998ty, Towell:2001nh, Webb:2003bj} at Fermilab, which include proton scattering from proton and deuteron targets.
We include data in the range $Q^2 > 36$~GeV$^2$.
Excluding lower $Q^2$ data is recommended by Ref.~\cite{Alekhin:2006zm}, which demonstrated that inclusion of the lower $Q^2$ data results in deteriorated prediction quality with no reduction in uncertainty when compared with fits to DIS data alone.
\begin{figure}[t!]
\centering
\includegraphics[width=\textwidth]{kinematics}
\caption{Kinematic coverage of data used in this analysis, with $Q^2$ versus the Bjorken scaling variable $\ensuremath{x_{\mbox{\tiny \rm Bj}}}$ for inclusive DIS~\cite{Benvenuti:1989rh, Benvenuti:1989fm, Whitlow:1991uw, Arneodo:1996qe, Arneodo:1996kd, Abramowicz:2015mha} and SIDIS data~\cite{Adolph:2016bga, Adolph:2016bwc} (left panel), fragmentation variable~$z$ for SIDIS and SIA data~\cite{Brandelik:1980iy, Althoff:1982dh, Braunschweig:1988hv, Lu:1986mc, Aihara:1983ic, Aihara:1988su, Cowan:1988gz, Derrick:1985wd, Itoh:1994kb, Abe:2003iy, Buskulic:1994ft, Akers:1994ez, Abreu:1998vq, Albrecht:1989wd, Leitgab:2013qh, Leitgab:2013dva, Lees:2013rqd} (central panel), and momentum fractions $x_1, x_2, \ensuremath{x_{\mbox{\tiny \rm F}}}$ for Drell-Yan data~\cite{Hawker:1998ty, Towell:2001nh, Webb:2003bj} (right panel).}
\label{f.kinematics}
\end{figure}
All SIA measurements are of the normalized differential cross sections $(\diff\sigma_{\mbox{\tiny SIA}}/\diff{z_h}) / \sigma_{\rm tot}$ for the reaction $e^+ e^- \to (\pi^\pm, K^\pm, h^\pm)\, X$.
The data are from experiments performed by the TASSO~\cite{Brandelik:1980iy, Althoff:1982dh, Braunschweig:1988hv} and ARGUS~\cite{Albrecht:1989wd} Collaborations at DESY, by the TPC~\cite{Lu:1986mc, Aihara:1983ic, Aihara:1988su, Cowan:1988gz}, HRS~\cite{Derrick:1985wd}, SLD~\cite{Abe:2003iy} and BaBar~\cite{Lees:2013rqd} Collaborations at SLAC, by the OPAL~\cite{Akers:1994ez, Abbiendi:1999ry}, ALEPH~\cite{Buskulic:1994ft} and DELPHI~\cite{Abreu:1998vq} Collaborations at CERN, and by the TOPAZ~\cite{Itoh:1994kb} and Belle~\cite{Leitgab:2013qh,Leitgab:2013dva} Collaborations at KEK.
As shown in \fref{kinematics}, the SIA data cover the large-$Q^2$ region where a leading power description in terms of FFs should be accurate.
Approximately half of the SIA data points have $Q \approx M_Z$, while the Belle and BaBar $B$ factories have lower $Q \approx 10.5$~GeV.
To ensure applicability of the leading power formalism, the SIA data in our fits are restricted to the range $0.2 < z_h < 0.9$.
Identification of heavy quark flavors for some of the SIA datasets is achieved through measurement of the total energy and momentum in secondary vertices.
The flavor tagged cross sections for a specific flavor $q=c$ or $b$ are particularly sensitive to the $D_q^h$, $D_{\bar q}^h$ and $D_g^h$ fragmentation functions into the observed hadron $h$.
In general, however, care needs to be taken with the precise method for separating primary quark flavors, and there are ongoing discussions regarding the optimal approach to this.
For more in-depth discussion see, for example, Ref.~\cite{Sato:2016wqj}.
Finally, the critical addition in this work compared with the previous JAM19 analysis~\cite{Sato:2019yez} is the inclusion of unidentified charged hadron data, along with charged pions and kaons, in the SIDIS off deuterium targets from the COMPASS Collaboration~\cite{Adolph:2016bga, Adolph:2016bwc} at CERN.
Since the SIDIS data
$\diff{\sigma_{\mbox{\tiny SIDIS}}^{h^\pm}}/\diff{Q^2}\diff{\ensuremath{x_{\mbox{\tiny \rm Bj}}}}\diff{z_h}$
are differential in $\ensuremath{x_{\mbox{\tiny \rm Bj}}}$ and $z_h$, they combine information on both PDFs and FFs, which appear in the description of SIA, Drell-Yan, and DIS data.
Furthermore, as illustrated in \fref{kinematics}, the SIDIS data have significant overlap in $\ensuremath{x_{\mbox{\tiny \rm Bj}}}$ and $z_h$ with the $\ensuremath{x_{\mbox{\tiny \rm Bj}}}$ and $\ensuremath{x_{\mbox{\tiny \rm F}}}$ range of inclusive DIS and Drell-Yan data, respectively, and the $z_h$ range of SIA data, so that the combined analysis constitutes a genuine test of their universality.
For the COMPASS SIDIS data we use the same kinematic cuts on $W^2$ and $Q^2$ as for inclusive DIS, and restrict the fragmentation variable to $0.2 < z_h < 0.8$ in order to exclude data from the target fragmentation region and avoid large-$z$ threshold corrections.
\section{Assessing universality}
\label{s.uni}
Before proceeding to the results of our numerical analysis, we briefly discuss the criteria for universality of the PDFs and FFs and how these are implemented in our analysis.
Extracting parton correlation functions, and using the extractions to test models of parton structure, is a nontrivial inverse problem, the detailed examination of which is beyond the scope of the present paper.
However, a claim that the success of a fit is a measure of the predictive power of the PDFs and FFs requires a number of basic minimal conditions to be met:
\begin{enumerate}
\item The system of unknown correlation functions must be over-constrained, by which we mean that the constraints on unknown correlations imposed by data (or other theoretical constraints such as sum rules) must be greater than the total number of functions involved.
\item Each unknown correlation function must appear at least twice within the set of factorization formulas relating the correlation functions to physical observables.
\item There must be reasonable kinematical overlap between the observables so that correlation functions can be compared within similar ranges of parton momentum fractions.
\end{enumerate}
\begin{comment}
\begin{table}
\scriptsize
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
& & \\ [-0.8em]
{\bf Observable} & {\bf Reactions} & {\bf Nonperturbative Functions} \\ [0.5ex]
\hline\hline
& & \\ [-0.8em]
$\frac{1}{\sigma_{\rm tot}} \frac{\diff\sigma^{\pi^\pm}}{\diff{z}}$ & $e^++e^-\rightarrow(\gamma,Z)\rightarrow{\pi^\pm}+X$ & $D^{\pi^\pm}_i\parz{z}$ \\ [1.5ex]
$\left.\frac{1}{\sigma_{\rm tot}} \frac{\diff\sigma^{\pi^\pm}}{\diff{z}}\right|_c$ & & $D^{\pi^\pm}_{c,\bar{c},g}\parz{z}$ \\ [1.5ex]
$\left.\frac{1}{\sigma_{\rm tot}} \frac{\diff\sigma^{\pi^\pm}}{\diff{z}}\right|_b$ & & $D^{\pi^\pm}_{b,\bar{b},g}\parz{z}$ \\ [1.5ex]
$\frac{1}{\sigma_{tot}}\frac{\diff\sigma^{K^\pm}}{\diff{z}}$ & $e^++e^-\rightarrow(\gamma,Z)\rightarrow{K^\pm}+X$ & $D^{K^\pm}_i\parz{z}$ \\ [1.5ex]
$\left.\frac{1}{\sigma_{\rm tot}} \frac{\diff\sigma^{K^\pm}}{\diff{z}}\right|_c$ & & $D^{K^\pm}_{c,\bar{c},g}\parz{z}$ \\ [1.5ex]
$\left.\frac{1}{\sigma_{\rm tot}} \frac{\diff\sigma^{K^\pm}}{\diff{z}}\right|_b$ & & $D^{K\pm}_{b,\bar{b},g}\parz{z}$ \\ [1.5ex]
$\frac{1}{\sigma_{tot}}\frac{\diff\sigma^{h^\pm}}{\diff{z}}$ & $e^++e^-\rightarrow(\gamma,Z)\rightarrow{h^\pm}+X$ & $D^{h^\pm}_i\parz{z}$ \\ [1.5ex]
$\left.\frac{1}{\sigma_{\rm tot}} \frac{\diff\sigma^{h\pm}}{\diff{z}}\right|_c$ & & $D^{h^\pm}_{c,\bar{c},g}\parz{z}$ \\ [1.5ex]
$\left.\frac{1}{\sigma_{tot}}\frac{\diff\sigma^{h\pm}}{\diff{z}}\right|_b$ & & $D^{h^\pm}_{b,\bar{b},g}\parz{z}$ \\ [2ex]
\hline
& & \\ [-0.8em]
$\frac{\diff{M}^{\pi^+}}{\diff{z}}$ & $\mu^++d\rightarrow{\pi^+}+X$ & $f_{i/p}\parz{x},D^{\pi^+}_i\parz{z}$ \\ [1ex]
$\frac{\diff{M}^{\pi^-}}{\diff{z}}$ & $\mu^++d\rightarrow{\pi^-}+X$ & $f_{i/p}\parz{x},D^{\pi^-}_i\parz{z}$ \\ [1ex]
$\frac{\diff{M}^{K^+}}{\diff{z}}$ & $\mu^++d\rightarrow{K^+}+X$ & $f_{i/p}\parz{x},D^{K^+}_i\parz{z}$ \\ [1ex]
$\frac{\diff{M}^{K^-}}{\diff{z}}$ & $\mu^++d\rightarrow{K^-}+X$ & $f_{i/p}\parz{x},D^{K^-}_i\parz{z}$ \\ [1ex]
$\frac{\diff{M}^{h^+}}{\diff{z}}$ & $\mu^++d\rightarrow{h^+}+X$ & $f_{i/p}\parz{x},D^{h^+}_i\parz{z}$ \\ [1ex]
$\frac{\diff{M}^{h^-}}{\diff{z}}$ & $\mu^++d\rightarrow{h^-}+X$ & $f_{i/p}\parz{x},D^{h^-}_i\parz{z}$ \\ [1ex]
\hline
& & \\ [-0.8em]
$F_2^p$ & $(e,\mu)+p\rightarrow{X}$ & $f_{i}(x)$ \\ [1ex]
$F_2^d,F_2^d/F_2^p$ & $(e,\mu)+d\rightarrow{X}$ & $\vdots$ \\ [1ex]
$\sigma_{r,{\rm NC}}^+$ & $e^++p\rightarrow(\gamma,Z)\rightarrow{X}$ & $\,$ \\ [1ex]
$\sigma_{r,{\rm NC}}^-$ & $e^-+p\rightarrow(\gamma,Z)\rightarrow{X}$ & $\,$ \\ [1ex]
$\sigma_{r,{\rm CC}}^+$ & $e^++p\rightarrow{W^+}\rightarrow{X}$ & $\,$ \\ [1ex]
$\sigma_{r,{\rm CC}}^-$ & $e^-+p\rightarrow{W^-}\rightarrow{X}$ & $\,$ \\ [1ex]
\hline
& & \\ [-0.8em] $M^3\frac{\diff^2\sigma^{pp}}{\diff{M}\diff{\ensuremath{x_{\mbox{\tiny \rm F}}}}}$ & $p+p\rightarrow\mu^++\mu^-+X$ & $f_{i}(x)$ \\ [1ex]
$M^3\frac{\diff^2\sigma^{pd}}{\diff{M}\diff{\ensuremath{x_{\mbox{\tiny \rm F}}}}}$ & $p+d\rightarrow\mu^++\mu^-+X$ & $f_{i}(x)$ \\ [1ex]
\hline
\end{tabular}
\end {center}
\caption{Summary of processes included in our fit $i \in \left\{u,\bar{u},d,\bar{d},s,\bar{s},c,\bar{c},b,\bar{b},g \right\}$.}
\normalsize
\label{t.thesummary}
\end{table}
\end{comment}
Using isospin invariance to relate the PDFs in the proton to those in the neutron, we have seven independent PDFs:
$f_u$, $f_d$, $f_s$, $f_{\bar{u}}$, $f_{\bar{d}}$, $f_{\bar{s}}$ and $f_g$,
with PDFs for heavy flavors generated perturbatively.
For the FFs, there are five functions for $\pi^+$ production:
$D^{\pi^+}_u$, $D^{\pi^+}_{\bar{u}}$, $D^{\pi^+}_c$, $D^{\pi^+}_b$ and $D^{\pi^+}_g$,
assuming that for equal $u$ and $d$ quark masses we can equate
$D^{\pi^+}_{\bar{d}} = D^{\pi^+}_u$.
Charge symmetry allows all the FFs for $\pi^-$ production to be related to those for $\pi^+$ production.
For $K^+$ production, there are six independent FFs:
$D^{K^+}_u$, $D^{K^+}_{\bar s}$, $D^{K^+}_{\bar{u}}$, $D^{K^+}_c$, $D^{K^+}_b$ and $D^{K^+}_g$,
where we differentiate between the $u$ and $\bar{s}$ functions.
Again, using charge symmetry the FFs for $K^-$ can be obtained from these six $K^+$ FFs.
Finally, for the unidentified charged hadrons $h^\pm$ or residual FFs, we identify five independent functions:
$D^{\rm res^+}_u$, $D^{\rm res^+}_{\bar{u}}$, $D^{\rm res^+}_c$, $D^{\rm res^+}_b$ and $D^{\rm res^+}_g$.
This makes then a total of 23 functions to be determined.
The quark and gluon PDFs are constrained by their appearance in several sum rules; in particular, the number sum rules,
\begin{equation}
\label{e.numberSR}
\int_0^1\diff{x}\, \big( f_q(x) - f_{\bar q}(x) \big)
= n_q,
\end{equation}
where $n_u = 2$, $n_d = 1$ and $n_s = 0$, and the momentum sum rule,
\begin{equation}
\label{e.momentumSR}
\sum_{i=q,\bar q, g} \int_0^1\diff{x}\, x f_i\parz{x} = 1.
\end{equation}
Note that in \ssref{param} these constraints were specifically used to fix the values of the normalization parameters for several fitted functions. However, for the purpose of assessing universality, they are simply counted as additional independent equations which include and thus constrain the PDFs.
The data sets discussed in \sref{datasets} also constrain the light quark and gluon PDFs since they appear in expressions for multiple independent observables.
Counting these and also the four sum rules (\ref{e.numberSR}) and (\ref{e.momentumSR}),
\begin{equation}
f_i(x) \stackrel{i \neq c,b}{\ \longrightarrow\ }\
\begin{cases}
6 \qquad \text{DIS} \\
2 \qquad \text{Drell-Yan} \\
6 \qquad \text{SIDIS} \\
4 \qquad \text{sum rules}
\end{cases}
\end{equation}
there is a total of 18 relations between the light quark PDFs.
The heavy quarks appear in an even greater number of observables.
The light quark fragmentation functions appear in at least one SIA observable and, because of charge conjugation invariance, in 2 SIDIS observables,
\begin{equation}
D_i^{\pi^+}(z) \stackrel{i \neq c,b,g}{\ \longrightarrow\ }
\begin{cases}
1 \qquad \text{SIA} \\
2 \qquad \text{SIDIS}
\end{cases}
\end{equation}
and similarly for the kaon and charged hadron fragmentation functions.
For a robust stress-test of universality, there should be reasonable overlap of the ranges in parton momentum fraction for both the PDFs and the FFs.
An indication for how well this is achieved in the current fit can be be gleaned from the kinematical coverage plots shown in \fref{kinematics}.
To lowest order in $\alpha_s$, the kinematical variables $\ensuremath{x_{\mbox{\tiny \rm Bj}}}$, $x_1$, $x_2$ and $z_h$ approximate the parton momentum fractions $x$ and $z$, respectively, while QCD evolution relates all values of $Q^2$.
Figure~\ref{f.kinematics} confirms that PDFs and FFs are both constrained by multiple processes in overlapping regions of momentum fractions.
In summary, our analysis does indeed fulfill the basic criterion for qualifying as a test of universality, and retaining predictive power for the PDFs and FFs more generally.
Note, however, that the momentum sum rule for FFs has not been imposed in the analysis.
Instead, this will be used as a consistency check for the final fit in \sref{analysis}.
\section{Numerical analysis}
\label{s.analysis}
In this section we present the results of our simultaneous Monte Carlo analysis of PDFs and FFs.
We begin with a survey of the fitted cross sections for the various global datasets used in this study, focusing especially on the quality of agreement with the SIDIS and SIA data on $\pi^\pm$ and $K^\pm$, as well as unidentified $h^\pm$ production.
We then present our final fitted PDFs and FFs, and discuss the vital role played by the SIDIS and SIA datasets in particular in constraining the strange quark distribution in the proton.
\subsection{Data and theory agreement}
\begin{figure}[t!]
\centering
\vspace*{-0.5cm}
\includegraphics[width=\textwidth]{chi2.pdf}
\caption{Reduced $\chi^2_{\rm red}$ values for each DIS (red), DY (green), SIDIS (orange) and SIA (blue) experiment considered in this analysis (left column), along with the corresponding mean and standard deviation of the residuals for each experiment, E\,[residual] (right column).}
\label{f.chi2}
\end{figure}
To assess the agreement of the fitted results with the various datasets, in \fref{chi2} we show the reduced $\chi^2$ for each individual experiment, which is defined by
\begin{equation}
\chi^2_{\rm red}
= \frac{1}{N}\sum_{i,e}
\frac{1}{\alpha_{i,e}^2}
\bigg( d_{i,e} - {\rm E}\Big[\sum_k r_e^k \beta^k_{i,e} + T_{i,e}/N_e\Big]
\bigg)^2.
\label{e.chi2red}
\end{equation}
Here, the expectation value ${\rm E}[...]$, as defined in \eref{exp2}, represents the mean theory, including optimized multiplicative and additive corrections to match the data, with $N$ the total number of data points.
In \fref{chi2} we show the mean and standard deviation of the Monte Carlo residuals for each experiment $e$, where the residual per data point is defined as
\begin{equation}
{\rm residual}\,(e,i)
=
\frac{1}{\alpha_{i,e}}
\bigg( d_{i,e} - {\rm E}\Big[\sum_k r_e^k \beta^k_{i,e} + T_{i,e}/N_e\Big]
\bigg).
\label{e.chi2red}
\end{equation}
For the inclusive DIS, Drell-Yan and SIDIS datasets we find excellent overall agreement between data and theory, with $\chi^2_{\rm red}$ values close to 1.
The $\chi^2_{\rm red}$ for the SIA datasets are slightly higher, but nonetheless the overall fit is very good, giving a total reduced $\chi^2_{\rm red}$ = 1.26 for almost 5000 data points.
The values of $\chi^2_{\rm red}$ for each type of dataset and for each specific hadron in the final state are summarized in Table~\ref{t.chi2}, along with the number of data points for each dataset.
\begin{table}[b]
\centering
\caption{Reduced $\chi^2_{\rm red}$ values for each type of dataset (DIS, Drell-Yan, SIDIS, SIA) considered in this analysis, together with the number of data points $N_{\rm dat}$ for each dataset.\\}
\begin{tabular}{ll|c|r} \hline
\multicolumn{2}{l|}{~~reaction} & ~~~$\chi^2_{\rm red}$~~~ & ~~~$N_{\rm dat}$~~ \\
\hline
~~DIS~ & & 1.29 & 2680~~ \\
~~DY~ & & 1.52& 250~~ \\
~~SIDIS~ & $\pi^\pm$~& 1.39 & 498~~ \\
& $K^\pm$~ & 1.38 & 494~~ \\
& $h^\pm$~ & 0.85 & 498~~ \\
~~SIA & $\pi^\pm$~& 1.09 & 231~~ \\
& $K^\pm$~ & 1.37 & 213~~ \\
& $h^\pm$~ & 1.15 & 120~~ \\
\hline
~~{\bf total} & & 1.26 & 4984~~ \\ \hline \\
\end{tabular}
\label{t.chi2}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[width=0.8\textwidth]{sidis-pi.pdf}\vspace*{-0.3cm}
\caption{Comparison of the multiplicities $dM^h/dz_h$ for $h=\pi^+$ (dashed lines) and $\pi^-$ (dotted lines) production with the COMPASS data~\cite{Adolph:2016bga, Adolph:2016bwc} in various bins of $\ensuremath{x_{\mbox{\tiny \rm Bj}}}$ and $y$ (offset by a factor $2^i$).}
\label{f.sidis-pi}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=0.8\textwidth]{sidis-K.pdf}\vspace*{-0.3cm}
\caption{As in Fig.~\ref{f.sidis-pi}, but for $K^\pm$ COMPASS SIDIS data~\cite{Adolph:2016bga, Adolph:2016bwc}.}
\label{f.sidis-K}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.8\textwidth]{sidis-h.pdf}
\caption{As in Fig.~\ref{f.sidis-pi}, but for unidentified hadron $h^\pm$ COMPASS SIDIS data~\cite{Adolph:2016bga, Adolph:2016bwc}.}
\label{f.sidis-h}
\end{figure}
The residuals profile for the DIS, Drell-Yan and SIDIS datasets is well centered around zero, with variances $\sim 1$, indicating an average Gaussian behavior of their associated likelihood function.
The variance for the SIDIS $h^-$ data from COMPASS, however, is found to be up to $\approx 50\%$ below unity, suggesting a deviation from a Gaussian likelihood.
This may be due to the fact that these data are dominated by systematic uncertainties, which is also reflected by the relatively small reduced $\chi^2_{\rm red}$ values, especially for the COMPASS $h^-$ data relative to the rest of DIS and SIDIS data sets.
A more detailed comparison with the COMPASS SIDIS is made in Figs.~\ref{f.sidis-pi}, \ref{f.sidis-K} and \ref{f.sidis-h}, where we show the $z_h$ dependence of the $\pi^\pm$, $K^\pm$ and $h^\pm$ multiplicities, respectively, which are defined as ratios of SIDIS to inclusive DIS cross sections at the same $\ensuremath{x_{\mbox{\tiny \rm Bj}}}$ and $Q^2$,
\begin{equation}
\frac{dM^h}{dz_h}
= \frac{\diff{\sigma_{\mbox{\tiny SIDIS}}^{h}}/\diff{Q^2}\diff{\ensuremath{x_{\mbox{\tiny \rm Bj}}}}\diff{z_h}}
{\diff{\sigma_{\mbox{\tiny DIS}}}/\diff{Q^2}\diff{\ensuremath{x_{\mbox{\tiny \rm Bj}}}}}.
\label{e.multiplicity}
\end{equation}
The agreement between theory and the experimental $z_h$ spectrum is quite remarkable, given that it spans some 2 orders of magnitude, which suggests that at these kinematics a leading power perturbative QCD factorization at next-to-leading order provides sufficient accuracy to describe the data.
Interestingly, the differences between the multiplicities for positively and negatively charged hadron species increase with $\ensuremath{x_{\mbox{\tiny \rm Bj}}}$, especially for kaons, and in the valence region these can differ by an order of magnitude for low values of $Q^2$.
Such differences can enhance our ability to extract flavor dependent effects in nonperturbative PDFs and parton to kaon FFs from the data.
The new data set included for the first time in the present JAM analysis, namely the unidentified charged hadron data shown in \fref{sidis-h}, are also well described by our nonperturbative ansatz for the corresponding FFs.
In contrast to the excellent agreement with the $z_h$ dependence of the data in Figs.~\ref{f.sidis-pi}--\ref{f.sidis-h}, we note that analysis of the same data differential in the hadron transverse momentum using existing PDFs and FFs within TMD factorization results in poor agreement between predictions and data~\cite{Gonzalez-Hernandez:2018ipj, Wang:2019bvb}, indicating that further work is needed to understand the SIDIS transverse momentum spectra.
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{pion-sia.pdf}
\caption{Data to theory ratios for SIA $\pi^\pm$ production cross sections versus $z_h$, with the bands indicating the uncertainty on the fitted result.}
\label{f.sia-pi-dot}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=\textwidth]{kaon-sia.pdf}
\caption{As in \fref{sia-pi-dot}, but for SIA $K^\pm$ production.}
\label{f.sia-k-dot}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=\textwidth]{hadron-sia.pdf}
\caption{As in \fref{sia-pi-dot}, but for SIA unidentified charged hadron $h^\pm$ production.}
\label{f.sia-h-dot}
\end{figure}
For the SIA data sets, there is a somewhat wider spread in the data versus theory comparisons, as seen in Figs.~\ref{f.sia-pi-dot}, \ref{f.sia-k-dot} and \ref{f.sia-h-dot} for the $\pi^\pm$, $K^\pm$ and unidentified charged hadron $h^\pm$ final states, respectively.
Generally, the $\pi^\pm$ data have the best agreement among the SIA datasets, with a reduced $\chi^2_{\rm red} = 1.09$, followed by the hadron data with $\chi^2_{\rm red} = 1.15$, and lastly the kaon data, which have an overall reduced $\chi^2_{\rm red} = 1.37$.
For about 3/4 of the $\approx 40$ SIA datasets, we find very good agreement with the global fit, with $\chi^2_{\rm red} \approx 1$ or below.
For the remaining datasets that have larger $\chi^2_{\rm red}$ values, to better understand the reasons for some of the tensions between data and theory we discuss in the following some individual cases ranked by the reduced $\chi^2_{\rm red}$ values.
Starting with the datasets that have the largest $\chi^2_{\rm red}$ values, namely, $\chi^2_{\rm red} \gtrsim 3$, we identify the OPAL ($\pi^\pm$ and $c \to K^\pm$), TPC ($K^\pm$), SLD ($\pi^\pm$ and $c \to K^\pm$), DELPHI ($K^\pm$), and TASSO ($\pi^\pm$ and $h^\pm$ at 35 GeV) datasets.
For the inclusive OPAL ($\pi^\pm$) data, we observe in \fref{sia-pi-dot} that for $z_h < 0.5$ the data are indeed in tension with the corresponding inclusive ALEPH and SLD results, and the overall trend of the data/theory ratio suggests a possible normalization issue with this dataset. This can also be said for the DELPHI ($K^\pm$) which appears to have some tension with the corresponding inclusive OPAL and ALEPH results.
Similarly, from \fref{sia-k-dot} we find that the TPC ($K^\pm$) spectrum lies below the theory, suggesting again a normalization problem with these data.
The situation for the TASSO ($\pi^\pm$) data is less clear, as only the $Q=14$~GeV dataset seems to give a bad fit, while data at other energies can be described fairly well.
This again hints at a problem with the overall normalization for this dataset.
The same behavior appears also in the TASSO ($h^\pm$) data in \fref{sia-h-dot}, where both the $Q=35$ and 45~GeV datasets are above the theoretical cross sections.
The case of SLD and OPAL ($c \to K^\pm$) data in \fref{sia-k-dot} shows a clear overestimation of the $z_h$ spectra.
While one can argue that this problem could be a reflection of the need for a more sophisticated heavy quark treatment in our theory, the description of $b$-tagged data from SLD, DELPHI and OPAL is relatively good, so that an explanation in terms of a normalization uncertainty in the SLD and OPAL ($c \to K^\pm$) data may be more relevant.
For SIA datasets that have smaller, but still large, $\chi^2_{\rm red}$ values, $2 \lesssim \chi^2_{\rm red} \lesssim 3$, we identify the $b$-tagged TPC ($b \to \pi^\pm$), OPAL ($b \to \pi^\pm$), and TASSO ($h^\pm$ at 44 GeV).
For the case of the TPC ($b \to \pi^\pm$) data, we see from \fref{sia-pi-dot} that for the largest $z_h$ bin the theory overestimates the data.
On the other hand, good agreement is found for the SLD ($b \to \pi^\pm$) data at the same kinematics.
It is possible that at the smaller $Q$ values of TPC relative to SLD, the range in $z_h$ where leading power factorization is applicable is narrower, in particular for the $b$-tagged data.
The $z_h$ dependence of the OPAL ($b \to \pi^\pm$) data appear to be clearly different from the theory, even within the large uncertainties.
We note here that the OPAL data are presented as truncated moments as a function of the lower limit of the integration, $z_h^{\rm min}$, and the inclusion of the very high $z_h$ bins may be problematic for the validity of factorization theorems at $z_h \to 1$.
Lastly, as with TASSO ($h^\pm$ at 35 GeV), the somewhat large $\chi^2_{\rm red}$ values for the 44 GeV data is likely attributable to a problem with overall normalization.
For datasets that have $\chi^2_{\rm red} \lesssim 2$, we consider the agreement to be generally acceptable.
Indeed, the vast majority of datasets in this category have $\chi^2_{\rm red} \approx 1$ or below.
These include all of the recent high-statistics $B$-factory data from BaBar ($\pi^\pm$, $K^\pm$) and Belle ($\pi^\pm$, $K^\pm$), most of the TASSO ($\pi^\pm$, $K^\pm$), TPC ($\pi^\pm$, $c \to \pi^\pm$) and SLD ($h^\pm$, $b \to \pi^\pm$, $b \to h^\pm$) datasets, all of the ALEPH ($\pi^\pm$, $K^\pm$, $h^\pm$) and most of the DELPHI ($\pi^\pm$, $K^\pm$, $b \to K^\pm$, $h^\pm$, $b \to h^\pm$) data, along with the older ARGUS ($K^\pm$), TOPAZ ($\pi^\pm$, $K^\pm$) and OPAL ($K^\pm$, $h^\pm$, $c \to h^\pm$, $b \to h^\pm$) data.
Slightly higher, but still reasonable, $\chi^2_{\rm red}$ values are obtained for the ARGUS ($\pi^\pm$), TPC ($h^\pm$), DELPHI ($b \to \pi^\pm$), and SLD ($K^\pm$, $c \to \pi^\pm$, $c \to h^\pm$) datasets.
Finally, we note that most of the large $\chi^2_{\rm red}$ values found in this analysis were absent in the previous JAM Monte Carlo analysis of fragmentation functions~\cite{Sato:2016wqj}.
The main reason is the restriction of the SIA datasets here to the range $0.2 < z_h < 0.8$, chosen to coincide with the range over which the SIDIS data in this work are able to be described within collinear factorization.
For the LEP data in particular there are many data points at $z_h < 0.2$ which can be well fitted within the current framework, and which would reduce the overall $\chi^2_{\rm red}$.
A~careful point by point comparison of the individual $\chi^2_{\rm red}$ values for the various datasets indeed confirms that similar discrepancies also occurred in Ref.~\cite{Sato:2016wqj}.
However, for consistency in our joint analysis of PDFs and FFs, we restrict the kinematic range to the region where both SIA and SIDIS can be simultaneously described.
The same choice for the $z_h$ range was made in the recent JAM19 analysis, which required SIDIS data to be restricted to $z_h \gtrsim 0.2$ to ensure separation of the target and current fragmentation regions.
\subsection{Parton distributions and fragmentation functions}
\label{ss.pdfsff}
The proton PDFs from our simultaneous fit are displayed in \fref{pdfs} at a scale $\mu^2=10$~GeV$^2$, where we focus on the kinematic region of parton momentum fractions $x \gtrsim 0.01$ that is constrained by the SIDIS data.
For comparison, we contrast our results with other next-to-leading order PDF parametrizations, namely, from the CJ15~\cite{Accardi:2016qay} and NNPDF3.1~\cite{Ball:2017nwa} global analyses.
Compared with the other fits, our valence $u$ and $d$ quark distributions have slightly larger magnitude in the intermediate-$x$ region, $x \sim 0.1$, with a compensating stronger suppression at small $x$ needed to ensure that the valence number sum rules are respected.
The ratio $d/u$ is quite compatible with the results from the other groups, on the other hand, but has a significantly larger uncertainty at large $x$ compared with the CJ15 result, reflecting the Monte Carlo nature of our analysis.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{pdfs.pdf}
\caption{Proton PDFs from the present JAM20-SIDIS analysis (red bands) versus $x$ at a scale $\mu=10$~GeV$^2$, compared with the CJ15~\cite{Accardi:2016qay} (blue bands) and NNPDF3.1~\cite{Ball:2017nwa} (green bands) parametrizations. The bands shown represent the mean $\pm 1\sigma$.}
\label{f.pdfs}
\end{figure}
The intermediate-$x$ enhancement in the valence PDFs in our fit is correlated with the slightly smaller $\bar d+\bar u$ light antiquark sea compared with the CJ15 and NNPDF3.1 parametrizations.
This in turn is correlated with the behavior of the strange quark sea, as seen in the ratio
\begin{equation}
R_s = \frac{s+\bar s}{\bar d+\bar u}
\end{equation}
of the strange to nonstrange sea quark PDFs.
In \fref{pdfs} this ratio is generally larger in our analysis than for the other parametrizations, with a somewhat bigger uncertainty.
This is understood from the fact that in the CJ15 fit $R_s$ is fixed to be 0.4 at the input scale, with deviations from the constant value arising only from DGLAP evolution.
For the NNPDF3.1 fit the uncertainties are smaller because of their inclusion of the neutrino DIS data, which we do not include in our analysis because of unknown nuclear corrections in neutrino scattering~\cite{Accardi:2009qv, Kovarik:2010uv, Kalantarians:2017mkj}.
Our light antiquark asymmetry $\bar d-\bar u$ is also compatible with the other groups, but again with a larger uncertainty, which may be related to the absence of collider $W$ and lepton asymmetry data in our fit.
Finally, for the gluon distribution, the magnitude and uncertainties are very similar across all the analyses, even though our fit does not include jet production data from hadron colliders.
This reflects the fact that the HERA DIS data, which are included here, provide strong constraints on the shape of the gluon PDF via scaling violations.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{ff.pdf}
\caption{Parton to hadron FFs versus $z$ at $\mu^2=100$~GeV$^2$ from the JAM20-SIDIS analysis for various parton flavors fragmenting to $\pi^+$ (red bands), $K^+$ (blue bands), unidentified hadrons $h^+$ (green bands), and residual hadrons $\delta h^+$ (yellow bands), defined as the difference between $h^+$ and the sum of $\pi^+$ and $K^+$. The bands shown are mean $\pm 1\sigma$.}
\label{f.ffs}
\end{figure}
For the parton to hadron FFs, we show in \fref{ffs} the $z$ dependence of the FFs at a scale $\mu^2=100$~GeV$^2$ for the positively charged $\pi^+$, $K^+$ and unidentified hadrons $h^+$, as well as for the residual hadrons $\delta h^+$, defined as the difference between $h^+$ and the sum of $\pi^+$ and $K^+$ (so that the total is given by $h^+ = \pi^+ + K^+ + \delta h^+$).
For most of the flavors we find that the quark $\to$ $\pi^+$ fragmentation dominates, as expected from the pion being the lightest hadron in the QCD spectrum.
Exceptions to this are for $\bar{s} \to K^+$ and $c \to K^+$ at intermediate $z$ values, and for $b$ quark fragmentation into residual hadrons $\delta h^+$.
For gluon fragmentation, pion production dominates for $z$ up to $\sim 0.5-0.6$, above which kaon fragmentation becomes as sizeable as the pion.
This is consistent with the findings of previous FF analyses~\cite{Hirai:2007cx, Sato:2016wqj}, which observed that the production of heavier particles such as kaons requires larger momentum fractions from the fragmenting gluon compared to the production of lighter particles.
The production of hadrons heavier than kaons, as indicated in \fref{ffs} by the residual hadrons $\delta h^+$, can be sizable and comparable to that of kaons, especially for the $d$ and $s$ quarks and at large values of $z$.
The relatively large $d \to \delta h^+$ FF can be understood in terms of the fragmentation into protons.
Note that we have imposed flavor symmetry for the residual hadron fragmentation, so that $D_d^{\delta h^+} = D_s^{\delta h^+}$.
In principle, the presence of hyperons such as $\Sigma^+$ should brake this relation, but we leave analysis of such effects for future work.
As the case for the $g \to K^+$, the fragmentation of gluons into heavier particles peaks at large $z$, where larger momentum fractions from the fragmenting gluons are need for the production of heavier particles.
For production of hadrons initiated by heavy quarks, we find similar fragmentation of charm quarks into pions and kaons, but a rather different pattern for the fragmentation of bottom quarks.
Some of this difference can be explained by the flavor-changing properties of $u$-type quarks decaying into $d$-type quarks.
While the charm quark can decay into strange quarks and hence enhance $K^+$ production, the same does not occur for bottom quarks, which suppresses kaon production relative to pion production due to the mass difference.
Interestingly, the production of other species of charged hadrons is much larger for $b$ quarks than for $c$ quarks, which may be understood from the greater phase space available for $b$ quarks to decay into heavier hadrons to which charm quarks cannot transition.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{moms.pdf}
\caption{Normalized yield of truncated moments
$\langle z \rangle_i^h$ of the $i \to h$ FFs $z D_i^h$, for the favored $\pi^+$ (red) and favored $K^+$ (blue), unfavored $\pi^+$ (light red) and unfavored $K^+$ (light blue), the total hadron $h^+$ (green) and residual hadron $\delta h^+$ (yellow) FFs, at a scale of $\mu^2=100$~GeV$^2$.}
\label{f.sumrule}
\end{figure}
In \fref{sumrule} we present truncated moments
\begin{equation}
\label{e.ztrunc}
\langle z \rangle_i^h
= \int_{z_{\rm min}}^1\, dz\, z D_i^h(z),
\end{equation}
for each flavor $i$ and final state hadron $h$, where we take the lower limits on the $z$ integration $z_{\rm min} = 0.2$ to restrict the moment to the region of SIDIS kinematics.
The truncated moment indicate how energetic is the production different a hadron species $h$ relative to the parent parton $i$.
In general, we find that the production of hadron species heavier than pions and kaons is typically produced with lower energies, which is consistent with the physical picture whereby more energy is required to produce heavier hadrons than lighter hadrons.
As expected, the favored fragmentation of $\bar{d}$ quarks is predominantly into highly energetic pions, while for the antistrange $\bar s$ the production rate of energetic kaons is slightly higher than that of pions.
The unfavored fragmentation of $d, s$ and $\bar u$ quarks follows a similar pattern, with the lightest (pion) state produced at the highest energies followed by kaons and other heavier charged hadrons.
An exception to this behavior is for charm and bottom quark fragmentation: for $c$ quarks kaons are produced with energies comparable to those of pions, while for $b$ quarks kaon production is suppressed with heavier mass hadrons produced at similar energies as pions.
Interestingly, the production of hadrons from gluons follows the same pattern as for $u$-quark fragmentation.
While the latter can be explained in terms of mass differences between the produced hadron species, the fact that $u$ quarks and gluons give a similar average energy profile across hadron species is intriguing.
On perturbative grounds one can argue that gluon fragmentation is enhanced because of the $C_A=3$ factor in the the gluon splitting function, $P_{gg}$, relative to quark splitting functions, $P_{qq}$ and $P_{gq}$, which are proportional to $C_F=4/3$.
The absence of direct constraints on the gluon FF beyond scaling violations, however, anything drawing more than speculative conclusions at present.
\begin{figure}
\centering
\includegraphics[width=0.9\textwidth]{Rs.pdf}
\caption{Monte Carlo samples for the $R_s$ ratio (left) and $zD_{\bar s}^{K^+}$ FF (right) at $\mu^2=10$~GeV$^2$, color coded according to the scaled $\chi^2_{\rm red}$ for the SIA ($K^\pm$) (top row) and SIDIS ($K^+, K^-$) (bottom row) datasets.}
\label{f.PDFstrange}
\end{figure}
We conclude the discussion of our numerical results by focusing on the correlation between the strange to nonstrange PDF ratio $R_s$ and the strange to kaon fragmentation function $D_{\bar s}^{K^+}$.
In \fref{PDFstrange} we show $R_s$ and the $\bar{s}\to K^+$ FF, with individual Monte Carlo samples color coded by the scaled $\chi^2_{\rm red}$ intensity (with darker replicas indicating higher likelihoods) computed for the specific cases of SIA ($K^\pm$) and SIDIS ($K^+, K^-$) datasets.
The SIA datasets have a clear preference for a smaller $R_s$ and enhanced $D_{\bar{s}}^{K^+}$, as was found in the previous JAM19 analysis~\cite{Sato:2019yez}.
Interestingly, the SIDIS ($K^+, K^-$) data, which have smaller $\chi^2_{\rm red}$, have a slight tendency to favor solutions with a larger $R_s$ and smaller $D_{\bar{s}}^{K^+}$, however, this preference is much weaker than the preference of the SIA data for smaller $R_s$ values.
We also note that in the current analysis we have extended the flexibility of the PDF and FF parametrizations, which allowed us to obtain a more uniform Monte Carlo distribution of $R_s$ compared JAM19, where a more restricted parametrization gave rise to multiple solutions.
Our new analysis confirms that the most probable solutions found in JAM19 did not result from parametrization bias, and corroborates the need for a suppressed strange quark PDF in the proton in order to simultaneously describe both the SIA and SIDIS datasets within leading power QCD factorization.
\section{Conclusion}
\label{s.conclusion}
In this paper we have presented the results of a simultaneous Monte Carlo analysis of PDFs and FFs constrained by a diverse array of data from inclusive and semi-inclusive DIS, Drell-Yan lepton-pair production, and SIA in $e^+ e^-$ collisions.
The analysis extends the previous JAM19~\cite{Sato:2019yez} simultaneous fit by including in addition unidentified charged hadrons in the final states of SIDIS and SIA, and increasing the flexibility of the PDF and FF parametrizations.
The analysis --- referred to as ``JAM20-SIDIS'' --- represents the most comprehensive determination of parton to hadron ($\pi^\pm$, $K^\pm$, $h^\pm$) FFs fitted concurrently with spin-averaged parton distributions, broadening the test of universality of parton correlation functions to more observables.
The more thorough exploration of the parameter space and reduced $\chi^2_{\rm red}$ values for each of the $\approx 70$ datasets fitted in this study confirmed the previous finding~\cite{Sato:2019yez} that the combination of SIA and SIDIS datasets have a strong preference for a smaller strange to nonstrange PDF ratio, $R_s$, correlated with an enhanced $D_{\bar{s}}^{K^+}$ FF.
As further tests of this scenario, we plan in future to extend the experimental datasets to include weak-boson and jet production in hadronic collisions, from both Tevatron and LHC data, as well as to relax the $W^2$ cuts for inclusive DIS to incorporate more fixed-target DIS data at high $\ensuremath{x_{\mbox{\tiny \rm Bj}}}$ values~\cite{Cocuzza21}.
An important application of the current results will be in benchmark calculations of transverse momentum dependent cross sections, and in particular for the small transverse momentum region where the transition from collinear factorization to TMD factorization is expected to set in.
One motivation for the present project was to assess the possible role of limitations in collinear PDF and FF fits in explaining discrepancies between theory and data in the range of intermediate and large transverse momentum across a number of transversely differential processes~\cite{Gonzalez-Hernandez:2018ipj, Wang:2019bvb, Bacchetta:2019tcu, Moffat:2019pci}.
For this, a truly simultaneous analysis of parton distribution and fragmentation functions across the standard set of electromagnetic processes, integrated over all transverse momentum, is necessary.
The general success of the collinear fits for transverse momentum integrated cross sections that we have examined here, and their evident predictive power, suggests that factors unique to the transverse momentum differential treatment are responsible for the tension with data.
We plan to address this also in future work. \\
\begin{acknowledgments}
This work was supported by the US Department of Energy contract DE-AC05-06OR23177, under which Jefferson Science Associates, LLC operates Jefferson Lab.
The work of N.S. was supported by the DOE, Office of Science, Office of Nuclear Physics in the Early Career Program.
E.M. and T.R. were supported by the U.S. Department of Energy, Office of Science, Office of Nuclear Physics, under Award Number DE-SC0018106.
\end{acknowledgments}
|
\section{Related Work and Conclusion}
\label{sec:conclusion}
We conclude with a brief discussion of proximately-related work that
lies at the synthesis of reverse engineering and static / dynamic
analysis, and contextualize this work in terms of our aspirations for
the future of \textsc{D\textsuperscript{3}RE}{}. There has been extensive work using logic
programming, and in particular Datalog, for static analysis of
higher-level langauges such as
Java~\cite{smaragdakis:2014,jordan:2016,scholz:2016}. The success of
the Souffl\'e Datalog engine has inspired recent adoption of logic
programming within the binary analysis community. For example, Datalog
Disassembly uses Souffl\'e to achieve both faster and more-precise
disassembly than the state-of-the-art disassembler
Ramblr~\cite{floresmontoya:2020}. Similarly, OOAnalyzer uses
XSB-Prolog, a version of Prolog implemented as a
library~\cite{schwartz:2018}. We are currently reimplementing
OOAnalyzer in \textsf{d3re}{} targeting \textsf{C++} Linux binaries. We feel
particularly excited about this direction because we believe Souffl\'e
will be immediately more scalable than XSB-Prolog.
While there are a broad range of plugins for Ghidra and IDA Pro to
load the \emph{results} of static analyses, we believe \textsf{d3re}{} is the
first to focus on the combination of open-ended deductive logical
inference and rapid interactivity (enabled by our metadatabase). We
believe the most closely related work is Ponce~\cite{ponce}, which
enables GUI-based symbolic execution. We plan to integrate symbolic
execution into \textsf{d3re}{} as a long-term goal, inspired by the recent
work of Formulog~\cite{formulog}.
Our goal in this work was to introduce a new vision for reverse
engineering, \textsc{D\textsuperscript{3}RE}{}, wherein expert users rapidly query
high-performance logical inference engines to help them accomplish
their day-to-day work in RE, vulnerability construction, and
penetration testing. Visualization-based tools such as Ghidra are of
immense value in understanding a binary, but have fundamentally
different design considerations than high-performance logical
inference enginges (such as Souffl\'e). Recent work in compiling
Datalog to parallel relational algebra (e.g., Gilray et
al.~\cite{hipc}) has enabled a new frontier in scale of Datalog-based
analyses. We hope that developments such as these will someday enable
realizing fully the vision of \textsc{D\textsuperscript{3}RE}{} to help reverse engineers
perform powerful static binary analyses at unprecedented scale.
\section{Design and Implementation}
\label{sec:design}
\begin{figure}[t]
\centering
\includegraphics[width=.45\textwidth]{design.png}
\caption{High-level components and their interactions in \textsf{d3re}{}.}
\label{fig:design}
\end{figure}
In this section, we present both a formal semantics for \textsc{D\textsuperscript{3}RE}{} and
describe our implementation of \textsf{d3re}{}. The high-level architecture of
\textsf{d3re}{} is outlined in Figure~\ref{fig:design}. Conceptually, the key
idea of our semantics is to maintain a \emph{metadatabase} to allow
efficient incremental reuse of previously-computed databases. In
\textsf{d3re}{}, this metadatabase takes the form of a server which accepts
Datalog programs to run to a fixed-point.
The metadatabase (server) interacts with both the REPL process and
Ghidra to render output databases into view annotations (e.g.,
highlights or comments) in the Ghidra UI based on REPL commands. Our
visualization is currently limited to printing to Ghidra's console,
highlighting a set of lines (typically some output relation), or
annotating a line with a comment (whose contents may be dynamically
determined via a Datalog query). We plan to investigate adding
comments to other Ghidra UI elements (such as inferred classes) and
other visual integration as future work.
\subsection{Formal semantics of \textsc{D\textsuperscript{3}RE}{}}
\label{sec:formal}
Due to space restrictions, we present only a sketch of a formal
semantics for \textsc{D\textsuperscript{3}RE}{}. Semantics of Datalog programs are typically
phrased in terms of an extensional database (EDB), an
extensionally-enumerated set of ground facts, and intensional database
(IDB), the set of rules defining the program~\cite{datalog}. Datalog's
semantics is given by a least-fixed-point of an ``immediate
consequence'' operator over the rules for the program. Because Datalog
programs have a finite Herbrand base (sets of atoms), this fixed-point
necessarily exists (though in practice Datalog engines allow
extra-logical behavior such as arithmetic). Datalog's conventional
semantics is \emph{monotonic}, in the sense that strictly more facts
are accumulated as the fixed-point computation evolves---negation is
allowed only when it may be stratified.
We define an \emph{EDB metadatabase} as a graph of EDBs with labeled
edges, $(\Delta, \stackrel{P}{\rightarrow})$, where $\Delta$ is a set
of EDBs, each EDB enumerating tuples for a given set of relations, and
$\stackrel{P}{\rightarrow}$ is a relation in $\Delta \times
\textit{Rules} \times \Delta$. When we process a program, $P$, using
an input EDB, we traverse the graph $(\Delta,
\stackrel{P}{\rightarrow})$ to find the most optimal,
\emph{compatible} EDB to start execution of $P$. Aided by Datalog's
monotonicity, we define an EDB as compatible if it was produced by a
subset of rules (or facts) from the input program / EDB. We conclude
our formalism sketch by remarking that $(\Delta,
\stackrel{P}{\rightarrow})$, given our usage, also forms a lattice.
\subsection{Implementation of \textsf{d3re}{}}
\textsf{d3re}{} is implemented in two parts: a REPL that communicates with
Ghidra's GUI and a background service to manage the metadatabase and
run the Datalog engine. The REPL currently communicates with Ghidra
via a third-party extension named
\textsf{ghidra\_bridge}~\cite{bridge}, which we plan to replace
imminently with an extension using protocol buffers.
To execute a Datalog program P, \textsf{d3re}{} analyzes the file using the
logic sketched in the above section to determine an optimal compatible
EDB to use. In the common case, a user will gradually accumulate a
stream of programs $P$, $P'$, $P'''$ consisting of a mix of rules and
assumptions. In the future, we envision that certain assumptions
(e.g., about calling conventions) may be implemented as GUI extensions
rather than, e.g., manually-enumerated facts. After each run, the
metadatabase will index the output facts and associate them with the
program $P$, establishing an edge in the aforementioned graph. In our
experiments, we refer to this as ``caching.''
\section{Evaluation}
\label{sec:evaluation}
\newcolumntype{n}{X}
\newcolumntype{s}{>{\centering\hsize=.4\hsize}X}
\newcolumntype{t}{>{\centering\hsize=.5\hsize}X}
\newcolumntype{l}{>{\centering\hsize=1\hsize}X}
\begin{table}[t]
\centering
\normalsize
\caption{Script size (lines of code) of Ghidra script (Python) vs. \textsf{d3re}{} Datalog}
\label{tab:loc}
\begin{tabularx}{.45\textwidth}{@{}nnn@{}} \toprule
& {Ghidra Python} & {\textsf{d3re}{} Datalog} \\ \midrule
{non-xor} & 33 & 8 \\
{overflow} & 60 & 18 \\
{basicblk} & 37 & 4 \\
{findcrypto} & 166 & 45 \\ \bottomrule
\end{tabularx}
\end{table}
\begin{table}[t]
\footnotesize
\centering
\caption{Running time of Ghidra scripts vs. equivalent
implementation in \textsf{d3re}{} (all numbers in seconds).}
\label{tab:ghidravsdddre}
\begin{tabularx} {.5\textwidth}{@{}nssssss@{}} \toprule
& {bison} & {souffle} & {gzip} & {re2c} & {redis} & {rsync} \\ \midrule
{non-xor Ghidra} & 3.569 & 107.5 & 2.205 & 3.903 &
10.52 & 3.050 \\
{non-xor \textsf{d3re}{}} & 0.518 & 6.515 & 0.097 & 0.756 &
1.306 & 0.486 \\ \midrule
{overflow Ghidra} & 0.370 & 0.247 & 0.600 & 0.240 &
0.760 & 0.180 \\
{overflow \textsf{d3re}{}} & 0.617 & 0.319 & 0.051 & 0.094 &
0.095 & 0.044 \\ \midrule
{basicblk Ghidra} & 340.6 & {--} & 4.664 & 472.1 &
1806 & 107.4 \\
{basicblk \textsf{d3re}{}} & 0.539 & 7.13 & 0.094 & 0.812 &
1.433 & 0.571 \\ \midrule
{findcrypt Ghidra} & 0.207 & 1.033 & 0.224 & 0.214 & 0.475 & 0.289 \\
{findcrypt \textsf{d3re}{}} & 1.287 & 14.53 & 0.224 & 1.701 & 2.938 & 1.186 \\
\bottomrule
\end{tabularx}
\end{table}
\begin{table}[t]
\footnotesize
\centering
\caption{Runtime of successive invocations to \textsf{d3re}{} with (C) and without (S) rule caching.}
\label{tab:cached}
\begin{tabularx} {.5\textwidth}{@{}nlllll@{}} \toprule
& {ddisasm} & {stack\_var} & {heap\_var} & {static\_var} & {unl\_static} \\ \midrule
{\textsf{souffle} C} & 170 & 11.88 & 58.35 & 5.008 & 0.039 \\
{\textsf{souffle} S} & 170 & 11.79 & 66.02 & 67.00 & 66.52 \\ \midrule
{\textsf{bison} C} & 7 & 0.932 & 1.409 & 0.545 & 0.022 \\
{\textsf{bison} S} & 7 & 0.934 & 1.916 & 2.122 & 2.075 \\ \midrule
{\textsf{re2c} C} & 9 & 1.457 & 4.417 & 0.704 & 0.025 \\
{\textsf{re2c} S} & 9 & 1.494 & 5.257 & 5.449 & 5.458 \\ \midrule
{\textsf{redis} C} & 11 & 1.918 & 2.544 & 1.302 & 0.025 \\
{\textsf{redis} S} & 11 & 1.919 & 3.525 & 3.712 & 3.726 \\ \midrule
{\textsf{rsync} C} & 8 & 0.766 & 0.908 & 0.481 & 0.028 \\
{\textsf{rsync} S} & 8 & 0.783 & 1.325 & 1.423 & 1.384 \\ \bottomrule
\end{tabularx}
\end{table}
We evaluated \textsf{d3re}{} both qualitatively, by implementing several
queries, and quantitatively by measuring its performance in
benchmarks. While \textsf{d3re}{} is still a work in progress, we had several
hypotheses we aimed to test as we designed and conducted these
experiments. First, we wanted to understand whether \textsf{d3re}{} provided
the necessary building blocks to enable replacing currently-existing
Ghidra scripts. Second, we wanted to understand whether \textsf{d3re}{} could
offer performance competitive with the kinds of Ghidra scripts that
reverse engineers typically use. Last, we wanted to understand the
performance of Ghidra for performing several repeated queries that
might mirror a realistic end-to-end workload using \textsf{d3re}{}.
\paragraph*{Ghidra Script Replication Study}
we wanted to determine whether \textsf{d3re}{} could realistically be used to
accomplish the kinds of tasks that reverse engineers face on a
day-to-day basis. This is an admittedly challenging question, which we
plan to eventually evaluate in several ways including user
studies. However, as initial work in this direction we arbitrarily
selected four Ghidra scripts listed in the \texttt{awesome-ghidra}
GitHub repository\cite{awesomeghidra}. The scripts we chose are listed
in Table~\ref{tab:loc}, along with their corresponding lines of code
in Python / Datalog. While Ghidra scripts may consist of a mix of
Python and Java, our experience is that most scripts use a small
subset of the Python API. The first three are relatively small and
find instructions that match a specific template, e.g.,
\textsf{non-xor} finds \textsf{xor} instructions that aren't zeroing
registers, and \textsf{overflow} heuristically searches for potential
overflows in calls to common functions such as \textsf{strcpy}. Our
largest was \textsf{findcrypto}, which looks for common cryptographic
constants.
\paragraph*{Qualitative Results of our Replication Study}
Our experience using \textsf{d3re}{} to replace Ghidra scripts must be
understood in the context that we are expert users and the developers
of \textsf{d3re}{}. However, we are pleasantly surprised that \textsf{d3re}{} enabled
us to succinctly write equivalent implementations of each Ghidra
script: we rewrote each script in substantially less Datalog
code. This is because the declarative nature of Datalog eliminates the
need for much of the conventional ceremony around, e.g., looping over
instructions and checking against a type that we found in our
evaluation scripts. Key to \textsc{D\textsuperscript{3}RE}{}'s success, we believe, is its
ability to directly use relations from \textsf{ddisasm}{}: we found that much
of the necessary work of, e.g., filtering instructions by their type
or operand was very useful at achieving succinct Datalog in
practice. We are in the initial planning stages of developing a
reverse engineering tutorial (or mini-course) around \textsf{d3re}{}, and are
hoping to use this to recruit developers to get more realistic
assessment of \textsf{d3re}{}'s usability by professional REs.
\paragraph*{Quantitative Results of our Replication Study}
We hoped that \textsf{d3re}{}, being based on a high-performance Datalog solver, would
offer performance competitive with Ghidra's current scripts. Each of our
evaluation scripts processed the entire binary and would highlight or label
certain instructions. To test the Ghidra scripts, we used Python's standard
\textsf{time} function before and after the script's work finished. We evaluated
the corresponding Datalog program by using Souffl\'e's internal performance
timers. We then benchmark Ghidra vs. \textsf{d3re}{} on a corpus of six binaries (all
sized less than 10MB), five from \textsf{ddisasm}{}'s test suite and Souffl\'e, shown at
the top of Table~\ref{tab:ghidravsdddre}. We used the latest versions of each
pre-built in the latest Arch Linux, but we used a pre-built version of
Souffl\'e. For each script, we waited for all of Ghidra's typical analyses to
finish, and similarly we ran \textsf{ddisasm}{} to build up the initial input database
for \textsf{d3re}{}.
The body of Table~\ref{tab:ghidravsdddre} compares the runtime of each
Ghidra script versus its corresponding implementation in \textsf{d3re}{}. The
single occurrence of -- indicates that Ghidra did not finish within an
hour. Broadly, we found that \textsf{d3re}{} outperformed Ghidra for each of
the scripts in our replication study. As we had hoped, \textsf{d3re}{}'s
design allowed us to leverage useful relations from \textsf{ddisasm}{}. We
found that many scripts do things like naive loops over sets of
functions or symbols to locate some property. By contrast, the
declarative style of \textsf{d3re}{} allowed us to write these not only more
succinctly (e.g., Datalog naturally aggregates results) but also more
efficiently---Souffl\'e optimally compiles input programs to efficient
relational algebra kernels that loop only when necessary. We did
observe various ways in which \textsf{d3re}{}'s limitations could cause
performance issues. For example, the \textsf{findcrypto} script scans
the binary for 256-segments of code. \textsf{d3re}{} is built on Souffl\'e,
which supports 64-bit primitive ints, but not 256-byte sequences. Thus,
we had to build up sequences via a set of Datalog rules, causing
inefficient memory representation due to the necessary duplication due
to representing subsequences as Datalog facts.
\paragraph*{Evaluating End-to-End Behavior in Subsequent Invocations}
To understand the effect of caching via repeated calls to \textsf{d3re}{}, we
ran four subsequent analysis queries in a row using both our
caching-based approach and without caching (wherein we started only
with the results of \textsf{ddisasm}{}). Our results are shown in
Table~\ref{tab:cached}: the time of the cached run (C) is shown above
the time for the correspond sequential run (S). As each query builds
on the previous, we expect caching to reduce the amount of work and
commensurately reduce the runtime. \text{stack\_var} finds
stack-allocated variables, while \textsf{heap\_var} calculates stack
variables holding pointers to heap values based on
\textsf{stack\_var}. \textsf{static\_var} and \textsf{unl\_static}
attempt to find uninitialized global variables. Overall, we found rule
caching was especially important on larger binaries versus sequential
runs, justifying our choice to structure the metadatabase as a graph.
\section{Introduction}
\label{sec:intro}
Binary reverse engineering (henceforth RE) is the process by which we
start with some input binary (sequence of bytes) and employ various
reasoning principles to explicate its behavior when executed as
code. While RE tasks are often partially automated (e.g., via
decompilation), full automation is often impossible: the extreme
semantic expressivity afforded to binaries (including encrypted code,
stripped symbol tables, etc..) often necessitates open-ended
exploration and case-specific reasoning. Recent literature suggests
that many practicioners follow an iterative approach involving several
rounds of hypothesis formation and validation/falsification, often
assisated via a combination of static and dynamic
analysis~\cite{votipka:2020,smith:2015,johnson:2013}.
To rapidly interact with a binary, RE practicioners often use reverse
engineering tools such as Ghidra~\cite{ghidra}, IDA Pro~\cite{idapro},
or Radare2~\cite{Radare2}. The goal of these tools is to allow an
RE\footnote{When unambiguous, we will use the term RE both to mean the
process of reverse engineering and a reverse engineering
practicioner} to quickly explore the binary and visualize it
(typically interactively, via a GUI or CLI) in a variety of ways. For
example, a reverse engineer looking for a time bomb may first search
for calls to the system's \textsf{time} function, and then walk
backwards to understand whether each call is associated with
legitimate or malicious behavior. In doing so, the RE may need to
reason about, e.g., indirect control flow, or even identify the
\textsf{time} function (in a stripped binary). Because REs are expert
users, and often skilled programmers, RE tools provide programmatic
interfaces that enable REs to systematize reasoning tasks via
extensions. A broad range of popular extensions exist for several
tools which perform such tasks as loading the results of static
analyses~\cite{schulte:2019,gtirb}, interacting with
debuggers~\cite{retsync}, and identifying common
cryptographic-relevant code~\cite{findcrypt}.
In this paper, we argue that \emph{deductive databases} (e.g.,
Datalog) serve as a natural abstraction boundary between RE tools and
logical inference tasks over binaries. We envision a future in which a
reverse engineer interactively explores a binary using an RE tool
while simultaneously querying arbitrarily-complex logical properties
written in a terse declarative style. We call this Declarative
Demand-Driven Reverse Engineering (henceforth \textsc{D\textsuperscript{3}RE}{}). In \textsc{D\textsuperscript{3}RE}{},
an RE interacts with a deductive database by giving inputs (e.g., the
currently highlighted address) to a rule-based deductive inference
system written in a declarative language such as Datalog. Rules
inductively compute \emph{relations} over facts about the binary. As
an example, consider a relation \textsf{direct\_call} $\in
\textit{Addr} \times \textit{Addr}$ which relates callsite addresses
(offsets within the binary) to procedure invocation target
addresses. In our vision, \textsc{D\textsuperscript{3}RE}{} allows REs to interactively compute
with and visualize the results of queries over these deductive rules.
We see \textsc{D\textsuperscript{3}RE}{} as a natural extension of several observations about
the state of the start. First, many existing RE tools assemble
databases to index various properties (e.g., addresses, symbols,
etc...) of binaries for quick exploration. Deductive databases further
allow REs to write arbitrary logical queries which are computed
maximally efficiently via, e.g., compilation to relational algebra
kernels as done in Souffl\'e. Deductive databases have also enabled
several recent advances in binary analysis demonstrating both
efficiency and robustness over conventional techniques. For example,
the Datalog-based disassembler \textsf{ddisasm} achieves both faster
and more-precise disassembly than other state-of-the-art
disassemblers, and OOAnalyzer uses Prolog to enable declarative
recovery of classes from compiled \textsf{C++} code.
In this short paper we describe our progress in implementing a
prototype tool, \textsf{d3re}{}, which we are building to realize the \textsc{D\textsuperscript{3}RE}{}
vision. \textsf{d3re}{} allows REs to interactively define and calculate
queries of arbitrary complexity over large production binaries and
then visualize their results using Ghidra. To implement \textsf{d3re}{} we
have designed an interface, which we call the \emph{mediator}, that
sits between a traditional Datalog solver and an RE tool. We briefly
formalize this interaction between the RE tool and logic solver in
Section~\ref{sec:design}, and go on to describe our prototype Ghidra
extension that enables visualizing the results of binary analyses in
our tool. Using this formalism, we describe how \textsf{d3re}{} readily
enables a broad range of binary analyses and sketch a vision for how
we believe \textsc{D\textsuperscript{3}RE}{} will prove to be a natural ergonomic for reverse
engineering.
We have measured the robustness of \textsf{d3re}{} in several ways. First, we
wanted to know whether \textsf{d3re}{} could truly live up to our vision of
being a natural replacement for the kinds of scripts REs already use
in their day-to-day work. To evaluate this, we reimplemented a set of
currently-existing Ghidra scripts. We happily observed that \textsf{d3re}{}
was not only an ergonomic advantage (allowing us to write succinct but
obviously-correct queries) but also a performance advantage. For
example, many Ghidra scripts play tricks to avoid unnecessarily
complexity that would arise in a straightforward implementation, e.g.,
iterating over a set of functions in a loop to check a property
resulting in super-linear complexity. In \textsf{d3re}{}, the Datalog solver
was naturally able to compile and organize work in an optimal way. We
discuss this and other results in Section~\ref{sec:evaluation}. We
conclude with a brief overview of related work and our outlook on future directions in Section~\ref{sec:conclusion}
Specifically, we claim the following three contributions:
\begin{itemize}
\item A formalization of our metadatabase as a database of databases
used to optimize subsequent invocations of the Datalog solver
(Section~\ref{sec:design}).
\item A prototype tool, \textsf{d3re}{}, consisting of a server which wraps
\textsf{ddisasm}{} with logic to enable chaining multiple subsequent calls
via the metadatabase. Also included in \textsf{d3re}{} is an extension to
the Ghidra RE toolkit to enable visualizing results computed using
\textsf{d3re}{}.
\item An evaluation of \textsf{d3re}{} on a set of benchmarks demonstrating
positive initial results indicating that \textsf{d3re}{} could replace
present-day binary analysis infrastructure (e.g., Ghidra scripts)
and directly enable more efficient and succinct implementation.
\end{itemize}
\section{Overview}
\label{sec:overview}
In this section, we demonstrate the vision and application of \textsc{D\textsuperscript{3}RE}{}
by illustrating how a reverse engineer might explicate a vulnerability
due to an uninitialized global variable. We consider a particular
binary, \texttt{CROMU\_00038}, from DARPA's Cyber Grand Challenge
which contains a function pointer which is uninitialized when an
invalid flag is set in the metadata portion of an input
file~\cite{cgc,example}. We demonstrate how our prototype tool,
\textsf{d3re}{}, can be used to build a declarative query to find
uninitialized function entry points and visualize them within
Ghidra. We do not claim that \textsf{d3re}{} can immediately or automatically
discover vulnerabilities---in this section we try to focus on how its
declarative reasoning instead enables rapidly exploring a binary to
uncover some property.
The vulnerable segment of code is a use-before-definition bug shown in
Figure~\ref{fig:sourcecode}. The \textsf{swap\_word} function is
initialized inside of the \textsf{main} function based on a value
parsed in a TIFF header---if the flag does not match \textsf{0x4949}
or \textsf{0x4d4d} the function is left uninitialized and the call on
line 17 crashes.
\begin{figure}
\scriptsize
\begin{lstlisting}[numbers=left]
// swap_short and swap_word only initialized within if
if (tiff_hdr->Byte_Order == 0x4949) {
printf("Intel formatted integers\n");
swap_word = intel_swap_word;
}
else if (tiff_hdr->Byte_Order == 0x4d4d) {
printf("Motorola formatted integers\n");
swap_word = motorola_swap_word;
}
#ifdef PATCHED
else {
printf("Invalid header values\n");
_terminate(-1);
}
#endif
// might cause an uninitialized variable bug here
offset = swap_word(tiff_hdr->Offset_to_IFD);
\end{lstlisting}
\caption{Uninitialized variable vulnerability in CROMU0038 source code}
\label{fig:sourcecode}
\end{figure}
\begin{figure}
\footnotesize
\begin{lstlisting}
>>> load dl/use_def_global.dl
>>> run dl/use_def_global.dl
>>> load dl/uninitialized.dl
>>> run dl/uninitialized.dl
>>> highlight
>>> comment
>>> query use_before_def_global
00004feb 0000a180 swap_short
00005017 0000a188 swap_word
...
0000515e 0000a180 swap_short
\end{lstlisting}
\caption{\textsf{d3re}{} REPL session used in this overview.}
\label{fig:result}
\end{figure}
\paragraph*{Loading the binary}
To begin an analysis of a binary, an RE will load the binary into a
reverse engineering tool. In our current implementation of \textsf{d3re}{}, a
user opens two processes silmultaneously: a GUI-based instance of
Ghidra, and a terminal running \textsf{d3re}{}'s REPL. The user can explore
the binary using all of the normal features of Ghidra and use all of
its conventional analyses (e.g., to recover entrypoints). However,
\textsf{d3re}{}'s REPL communicates with Ghidra so that when \textsf{d3re}{}'s
analysis finishes Ghidra's views update as appropriate.
\paragraph*{Initial processing}
It is conventional that reverse engineering tools will apply a set of
analyses to a binary to disassemble it and index various items such as
entrypoints and callsites In \textsf{d3re}{}, the user builds queries in
Datalog starting from a large initial set of Datalog rules that build
on top of \textsf{ddisasm}{}, a Datalog-based disassembly
engine~\cite{floresmontoya:2020}. Analogously to the indexing and
analysis operations provided by Ghidra (and other RE tools), \textsf{d3re}{}
invokes \textsf{ddisasm}{} once to build an initial database.
Building on top of \textsf{ddisasm}{} was initially a strategic
choice---\textsf{ddisasm}{} already includes facilities to parse object files
and transform them into input databases in the style required by
Souffl\'e. Initially, we extended \textsf{ddisasm}{}'s set of rules with
additional user-specific queries---a slow process, as \textsf{ddisasm}{} can
take several minutes to run on large binaries. This was at odds with
our goal of enabling rapid real-time feedback to users of \textsf{d3re}{}.
\textsc{D\textsuperscript{3}RE}{} builds upon a key observation that we have found crucial to
enable efficient interactive binary analyses in practice: because
Datalog is monotonic, we can evaluate an extended program (i.e., a
program extended with a set of additional rules or queries) by using
the database resulting from the calculation of the previous
program. Thus, running \textsf{ddisasm}{} \emph{once} allows pre-populating a
large set of inferred relations for a wide range of interesting facts
about binaries, including intraprocedural reachability and calling
conventions.
When a binary is loaded, \textsf{d3re}{} invokes \textsf{ddisasm}{} with one slight
modification: every Datalog relation in \textsf{ddisasm}{}'s rule database
(used by \textsf{ddisasm}{} to build a disassembly) is modified to be an
output relation. In \textsf{ddisasm}, only dissassembly-relevant relations are
output, rather than internal relations (e.g., those that relate to
intraprocedural reachability). By marking all \textsf{ddisasm}{}'s relations
as output relations, \textsf{d3re}{} provides them to the user as primitives
with which to build queries over binaries\footnote{A relevant analogy
might be that \textsf{ddisasm}{} is the standard library of \textsf{d3re}{}}. After
the binary is loaded, all rules declared in \textsf{ddisasm}{} will be
available for querying. Additionally, \textsf{ddisasm}{} will be run only
once, even if the user uploads the same binary several times. All
facts generated in this step will be stored in a temporary folder on
disk managed by the metadatabase (described in
Section~\ref{sec:design}).
\begin{figure}[t]
\begin{lstlisting}
def_global(EA,dest) :-
code(EA), instruction_get_dest_op(EA,Index,_),
pc_relative_operand(EA,Index,dest),
defined_symbol(dest,_,"OBJECT","GLOBAL",_,_).
used_global(EA,dest,Index) :-
code(EA), instruction_get_src_op(EA,Index,_),
pc_relative_operand(EA,Index,dest),
defined_symbol(dest,_,"OBJECT","GLOBAL",_,_).
def_used_global(EA_def,GA,EA_used,Index) :-
used_global(EA_used,GA,Index),
block_last_def_global(EA_used,EA_def,GA).
def_used_global(EA_def,GA, EA_used, Index) :-
last_def_global(Block,EA_def,GA),
code_in_block(EA_used, Block),
used_global(EA_used, GA, Index),
!block_last_def_global(EA_used,_,GA),.
\end{lstlisting}
\caption{Global Var Use-Def analysis}
\label{fig:usedef}
\end{figure}
\paragraph*{Designing a query to explicate use-before-define}
In the \textsc{D\textsuperscript{3}RE}{} approach, REs interactively build queries to highlight
various portions of the program that match certain properties. They
then manually inspect the results of their queries and use their
intuition to build subsequent queries. Along the way, the RE may
choose to add comments to various instructions, functions, or other
forms and browse those instructions in Ghidra. In \textsf{d3re}{}, the
communication between the logical rules and the state of the RE tool
is reconciled by input and output tables---RE users can write queries
that consume the state of the RE tool (such as
\textsf{currentAddress}, the currently-selected address) as input
relations, perform logical inference, and leave their output in
relations such as \textsf{comment(addr,``vuln'')}.
Like \textsf{ddisasm}{}, \textsf{d3re}{} uses the Souffl\'e Datalog engine to perform
logical inference over binaries. Users of \textsf{d3re}{} can incrementally
build up more rules in the interactive REPL (shown in
Figure~\ref{fig:result}). Currently, our REPL allows loading rules by
loading new files---we plan on adding direct support for new rules,
along with error-reporting feedback soon.
Knowing there was an uninitialized global function pointer being used,
a user of \textsf{d3re}{} might first define a set of relations to build up
def-use-chains of global variables. Datalog code to implement these
queries is illustrated in Figure~\ref{fig:usedef}. The last two rules
build up a relation
\textsf{def\_used\_global(EA\_def,GA,EA\_used,Index)}, which infers
that at address \textsf{EA\_def}, the global variable (at address
\textsf{GA}) is defined and used at address \textsf{EA\_used} at
operand index \textsf{Index}. While this is a relatively coarse query,
we envision the user could run the query on the binary to visualize a
large answer set. In our setting, this can be done using the
\textsf{highlight} or \textsf{comment} commands, which display the
data marked to be highlighted by the most recent result computation.
\begin{figure}
\begin{lstlisting}
use_before_def_global(EA_used,GA,Name) :-
used_global(EA_used,GA,Index),
!def_used_global(_,GA,EA_used,_),
defined_symbol(GA,_,"OBJECT","GLOBAL",_,Name).
use_before_def_global(EA_used,GA,Name) :-
used_global(EA_used,GA,Index),
def_used_global(EA_def,GA,EA_used,_),
!def_null_global(EA_def,GA),
defined_symbol(GA,_,"OBJECT","GLOBAL",_,Name).
\end{lstlisting}
\caption{uninitialized variables}
\label{fig:uninitvar}
\end{figure}
Based on our definitions in Figure~\ref{fig:usedef}, we can define a relation
for variables which are possibly used before they are defined. We demonstrate
this in Figure~\ref{fig:uninitvar}. In the first rule, we say that if there is
some usage of a global variable at some address, but in that address, we can't
find any definition related to it, then we will consider variable there as an
uninitialized variable; The second clause says that for some usage of a global
variable even if it has some definition associated with it, if that definition
is nullptr, we will still consider that there is a use-before-def vulnerability
here.
\paragraph*{Refining the query}
In \textsf{d3re}{}, users can easily access the result of the rule and all
facts generated by \textsf{ddisasm}{} through the GUI by writing into output
tables using \textsf{d3re}{} rules. Unfortunately, our above query produces
over 50 possible results---checking each occurrence would still be a
timely endeavor. Next, we narrow down the query space to the range of
just the main function. We use an auxiliary predicate,
\textsf{code\_in\_range}, which we seed with constants for the
beginning and end of the \textsf{main} function we gain from
inspecting the binary in Ghidra.
\begin{lstlisting}
code_in_range(19490,21704).
use_before_def_global(EA_used, GA, Name) :-
code_in_range(from, to), EA_used >= from, EA_used < to,
used_global(EA_used,GA,Index),
!def_used_global(_,GA,EA_used,_),
defined_symbol(GA,_,"OBJECT","GLOBAL",_,Name).
\end{lstlisting}
After new rules are applied, the output of the program becomes empty:
however, this does not specify the program is free from the
vulnerability. First, because of our constraint, only the
\textsf{main} function is searched, bugs may still hide in other
functions. Secondly, if all usage of a variable is before it’s
definition, null pointer error can still appear: programmers may
initialize a variable to \textsf{NULL} and use several non-total
branches to initialize the pointer, leaving the pointer uninitialized
at the join point when no switch fires. We modify our rules to account
for this:
\begin{lstlisting}
def_null_global(EA,GA) :-
def_global(EA,GA), instruction_get_src_op(EA,_,Op),
op_immediate(Op,offset), offset=0.
use_before_def_global(EA_used, GA, Name) :-
code_in_range(from,to), EA_used >= from, EA_used < to,
used_global(EA_used,GA,Index),
def_used_global(EA_def,GA,EA_used,_),
!def_null_global(EA_def,GA),
defined_symbol(GA,_,"OBJECT","GLOBAL",_,Name).
\end{lstlisting}
This change results in 19 addresses to search, and combining these
results with use-def information in the previous step and
intra-procedural control-flow graph in Ghidra, we can fairly easily
infer that the global variable \texttt{swap\_word} is initialized to
\textsf{0} at address \texttt{0x4c2a}, that both conditional jumps
\texttt{0x4f80} and \texttt{0x4fb8} fail, and observe a subsequent
usage of \texttt{swap\_word} at \texttt{0x5017} which will trigger a
crash. At any stage in our process, we can sync Ghidra's UI with the
current database using several REPL commands (an example is shown in
Figure~\ref{fig:codeview}). In a fully-fledged implementation of
\textsf{d3re}{}, we hope to have UI gadgets (or templates) to help users
interactively build queries. For example, we may allow the user to
select a region of the binary and build a rule that applies only to
that region, or right-click on a function and build a rule specific to
callers of that function. We believe this will need to be informed by
a combination of interviews with expert users, participatory design,
and (perhaps) user studies. This is work we plan to undertake now that
we have proven initial success to ourselves with \textsf{d3re}{}.
\begin{figure}[b]
\centering \includegraphics[width=.45\textwidth]{img/codeview.png}
\caption{Ghidra with highlights and comments declaratively specified
to output results inferred via \textsf{d3re}{} for our example.}
\label{fig:codeview}
\end{figure}
We conclude this section by remarking upon the nature of our
analyses. Our analyses would be considered na\"ive by the standards of
industrial static analyses. Indeed, our reasoning is not even
sound---we can restrict ourselves to looking at results for only one
function or ignore complex behavior. Still, we believe that this
iterative ad-hoc reasoning is a technique many reverse engineers
already employ---the vision of \textsc{D\textsuperscript{3}RE}{} is to harmoniously leverage
state-of-the-art deductive reasoning engines while performing
human-guided RE tasks.
\section{Related Work}
\label{sec:related}
\subsection{Logic program in reverse engineering}
Few work is done in combining logic programming and reverse engineering. Unlike
\textsf{d3re}{} most existing work still mainly focused on some specific problems, but not
a general tool that can deal with a broad kind of reverse engineering problem.
OOAnalyzer\cite{schwartz:2018} is one of them, it is a tool designed to recover
C++ abstraction in a binary using logic programming language. It also provides a
Ghidra plugin currently, but it will only work for programs originally written
in C++. Most of the rules in OOAnalyzer are based on win32 thiscall calling
convention, so it is also not easy to be directly used it for binaries in linux
and other platforms.
XSB-Prolog is chosen as the logic engine of OOAnalyzer. In order to achieve good
performance and make sure forward reasoning always terminates, all rules in
OOAnalyzer are running in tabling mode\cite{swift:2010}. In \textsf{d3re}{}, we choose
datalog as our logic engine. Datalog is similar to prolog with tabling, the
semantic of datalog guaranteed termination of a program for free. In performance
perspective, nowadays datalog engines like souffle have much better performance
compared to prolog implementations like XSB-prolog, SWI-prolog.
\subsection{Program Analysis Using Datalog}
There is lots of good research in high performance datalog compilers. Due to
proper data structure and good index selection algorithm \cite{subotic:2018},
the speed of souffle is really fast. Compared to other datalog implementations,
souffle has a good debugger and abundant language features like arithmetic and
component. These features are very useful for design program analysis algorithms
using datalog. In souffle’s source code repository, it gives lots of examples of
how to do small size of program analysis, but it is also applicable for large
size program analysis .Most famous program analysis framework using datalog is
Doop \cite{smaragdakis:2014}. Doop is a pointer analysis framework that allows
users to do object sensitive analysis on large size java programs like tomcats.
And in the binary analysis area, OOAnalyzer also shows using logic program
language like datalog and prolog can make analysis both more readable and
faster.
|
\section{Introduction}
Concavity is a central tool in mathematical economics and optimization theory, but in practice the widely used functions in these areas are considered to be quasiconcave instead of concave.
In Arrow-Enthoven $(1961)$ \cite{A-E}, the concave optimization problem was extended to the quasiconcave programming, and sufficient optimality conditions in differentiable case were obtained. Later, several authors have studied the quasiconvex optimality conditions by means of various generalized gradients (see for instance Hassouni \cite{H}, Hiriart-Urruty \cite{HU}, Martinez-Legaz \cite{MLeg}).
In consummer theory, the functions studied in many cases are considered to be separable of nature. Sufficient and necessary condition on separable utility functions to be quasiconcave was given in Yaari \cite{Y}, Debreu and Koopmans \cite{D-K}, Crouzeix and Lindberg \cite{C-L}, Berdi and Hassouni \cite{B-H}.
In this paper, the unconstrained optimization problem
$$ \rm{(USQP)}:\qquad \rm{\max_{x\in X} f(x)}$$
and the constrained optimization problem
$$\rm{(CSQP)}:\qquad \rm{ \displaystyle{ \max_{x\in X}f(x) \quad subject\ to\ h_j(x)\geqslant 0 \ (j=1,...,p) }} $$
are investigated, where
$\rm{f}$ and $\rm{h_j}\ $ $(j=1,...,p)$ are quasiconcave on $X$, and
$$\rm{f(x) = \prod_{i=1}^{m} f_{i}(x_i), \qquad x=(x_1,...,x_m)\in X=X_1 \times X_2 \times...\times X_m}$$
with $\rm{f_{i}}$ is a positive non constant real valued function defined on the nonempty open convex set $X_i$ of $\mathbb{R}^{n_i}$.
The scope of the paper is to obtain sufficient optimality conditions for the unconstrained separable quasiconcave problem $ \rm{(USQP)}$, and first order sufficient optimality conditions of K.K.T type for the constrained separable quasiconcave problem $ \rm{(CSQP)}$ with quasiconcave constraint functions, in both differentiable and non differentiable cases.
The paper is organized as follows: In section $2$, we recall some definitions and properties that will be needed, such as some characterizations of quasiconcavity and pseudoconcavity in the differentiable case.
In section $3$, we recall the definition and some properties of a multiplicatively separable $\rm{(MS)}$ function. The basic tool is the multiplicative concavity index introduced by Crouzeix and Kebbour in \cite{C-K}. Under the separability condition $\rm{(MS)}$, quasiconcave differentiable functions become more regular (in convex sense), and then nice results of the problems $ \rm{(USQP)}$ and $ \rm{(CSQP)}$ are obtained.
In section $4$, we define a generalized superdifferential to extend the notion of pseudoconcavity and some of its fundamental properties without differentiability assumption.
Some useful and important generalizations of pseudoconcavity in non differentiable case were introduced and studied under various assumptions by Aussel \cite{Aus}, Hassouni and Jaddar \cite{H-J}, and Koml\'osi \cite{K} for related works.
Again, under the separability condition $\rm{(MS)}$, the problems $ \rm{(USQP)}$ and $ \rm{(CSQP)}$ will be studied in the non differentiable case. An appropriate variant of K.K.T conditions of $ \rm{(CSQP)}$ will be provided.
\section{Preliminaries and notations.}
We recall some definitions and properties that will be needful in the sequel of this paper.
\begin{deft}\label{deft 2.1.}
Let $C$ be a nonempty convex set in $\mathbb{R}^n$ and let $g$ be a real valued function defined on $C$.\\
The hypograph of $g$ is the set
$$hyp(g):=\left\lbrace (x,\mu)\in C\times \mathbb{R}:g(x)\geqslant \mu \right\rbrace$$
For $\alpha\in \mathbb{R}$, the upper-level set $U(g,\alpha)$ and the strict upper-level set $U^{s}(g,\alpha)$ of $g$ are defined as follows:
$$U(g,\alpha):=\left\lbrace x\in C: g(x)\geqslant \alpha \right\rbrace.$$
$$U^{s}(g,\alpha):=\left\lbrace x\in C: g(x) > \alpha \right\rbrace.$$
\begin{itemize}
\item $g$ is said to be \textbf{ concave} on $C$ if $hyp(g)$ is convex, or equivalently if
$$g((1-\lambda)x+\lambda y) \geqslant (1-\lambda)g(x)+\lambda g(y) $$
for all $x,y\in C$ and for all $\lambda \in \left[0;1\right].$
\item $g$ is said to be \textbf{ quasiconcave} on $C$ if the upper-level set $U(g,\alpha)$ is convex for all $\alpha \in \mathbb{R}$, or equivalently if
$$g((1-\lambda)x+\lambda y) \geqslant \min\left[g(x);g(y)\right] $$
for all $x,y\in C$ and for all $\lambda \in \left[0;1\right].$
\item $g$ is said to be \textbf{ semi-strictly quasiconcave} on $C$ if
$$ \forall x_1, x_2 \in C, x_1 \not = x_2 :\ \ g(x_2) > g(x_1) \Rightarrow g((1-\lambda) x_ 1+ \lambda x_ 2) > g(x_ 1) $$
for all $ \lambda $ in $ (0,1) $
For a full description of concavity and quasiconvexity we refer to \cite{C-M,C-F,D-A-Z,R}.
%
\item $g$ is said to be convex (quasiconvex, semi-strictly quasiconvex) if $-g$ is concave (quasiconcave, semi-strictly quasiconcave).
\item If $g$ is positive, it is said to be \textbf{logarithmically concave} (log-concave for short) if
$\ln \circ g$ is concave.
\end{itemize}
The following properties are rather direct consequences of the definitions.
\begin{itemize}
\item $g$ concave $\Longrightarrow g$ log-concave $\Longrightarrow g$ quasiconcave.\\
Let $\varphi:g(C)\rightarrow \mathbb{R}$;
\item If $g$ is quasiconcave on $C$ and $\varphi$
is nondecreasing then $\varphi \circ g$ is quasiconcave.
\item If $g$ is convex (concave) and $\varphi$ is nondecreasing and convex (concave) then $\varphi \circ g$ is convex(concave).
\item If $g$ is convex (concave) and $\varphi$ is nonincreasing and concave (convex) then $\varphi \circ g$ is concave (convex).\\
In particular, if $g$ is positive and concave on $C$, then $\displaystyle \frac{1}{g}$ is convex on $C$.
\item Given $x \in C$ and $d\in \mathbb{R}^{n}$, let us define
$$I_{x,d} := \{t\in \mathbb{R} : x+td \in C \}$$
$$ g_{x,d}(t)=g(x+td), \qquad t\in I_{x,d} $$
Then, $g$ is concave (quasiconcave, log-concave) if and only if for every
$x\in C$ and $d\in \mathbb{R}^n$, the function $g_{x,d}$ is concave (quasiconcave, log-concave) on the interval $I_{x,d}.$
\end{itemize}
\end{deft}
Let's recall the well known characterization of quasiconcavity under differentiability assumption.
\begin{pro}\label{pro 2.1.}
(Arrow-Enthoven \cite{A-E})
Let $g$ be a differentiable function defined on a nonempty open convex set $X$ of $\mathbb{R}^n$.
Then, $g$ is quasiconcave on $X$ if and only if for all $x_1,x_2 \in X:$
$$\displaystyle{\left\langle \nabla g(x_1),x_2-x_1 \right\rangle< 0 \Rightarrow g(x_2)< g(x_1) }$$
\end{pro}
Under differentiability assumption, the following proposition gives a necessary condition of quasiconcavity by means of the strict upper level-set.
\begin{pro}\label{pro 2.2.}
Let $g$ be a differentiable quasiconcave function defined on a convex set $X$ of $\mathbb{R}^n$, and $x,\bar{x}\in X$. Then,
$$\displaystyle{x\in clU^s(g, g(\bar{x})) \Rightarrow \left\langle \nabla g(\bar{x}), x-\bar{x}\right\rangle \geq 0 }$$
\end{pro}
By extending the inequality in Proposition \ref{pro 2.1.}, the notion of pseudoconcavity was introduced as a generalized concavity which plays an important role in applied mathematics such as, optimization theory and mathematical economics. Let's recall the definition and some properties of pseudoconcavity.
\begin{deft}\label{deft 2.2.}
Let $g$ be a differentiable function defined on an open nonempty convex set $X$ of $\mathbb{R}^n$. \\
$g$ is said to be pseudoconcave on $X$ if for all $x_1,x_2 \in X$
$$\displaystyle{ \left\langle\nabla g(x_1), x_2-x_1 \right\rangle \leq 0 \Rightarrow g(x_2) \leq g(x_1) }$$
$g$ is said to be pseudoconvex on $X$ if $(-g)$ is pseudoconcave
\end{deft}
\begin{pro}\label{pro 2.3.}(Mangasarian \cite{M})
Let $g$ be differentiable on $X$. Then,
if $ g $ is pseudoconcave on $ X $, then it is semi-strictly quasiconcave on $X$.
\end{pro}
\begin{pro}\label{pro 2.4.}
Let $g$ be a differentiable function on an open convex set $X$ of $\mathbb{R}^n$. Then, $g$ is pseudoconcave on $X$ if and only if the restriction of $g$ to any line segment in $X$ is pseudoconcave. (see \cite{D-A-Z}).
\end{pro}
It has been known that a differentiable pseudoconcave function $g$ is quasiconcave and has a maximum at $x$ whenever $\nabla g(x) =0$. In \cite{C-F}, Crouzeix and Ferland have shown that this property is a necessary and sufficient condition for pseudoconcavity.
\begin{pro}\label{pro 2.5.}(Theorem 2.2. \cite{C-F})
Let $g$ be a differentiable and quasiconcave function on an open convex set $X \subset \mathbb{R}^n$ . Then $g$ is pseudoconcave on $X$ if and only if $g$ has a local maximum at $x\in X$ whenever
$\nabla g(x) = 0$.
\end{pro}
\begin{deft}
Let $f$ be a real function defined on a convex set $C \subset \mathbb{R}^n$. For $r\in \mathbb{R}\setminus \left\lbrace0\right\rbrace$, let $f_r$ be defined by: \qquad $f_r(x)=e^{rf(x)}$.\\
$f$ is said to be $r$-concave if $f_r$ is concave whenever $r>0$ and convex whenever $r<0$.\\
$f$ is $0$-concave if it is concave. (For more details on $r$-concavity/convexity see Avriel \cite{Avr}).
\end{deft}
\begin{pro}\label{pro 2.6.}(Theorem 6.1. \cite{Avr})
Let $r$ be any real number and let $f$ be a differentiable
$r$-concave function on a convex set $C \subset \mathbb{R}^n$. Then $f$ is pseudoconcave on $C$.
\end{pro}
\section{Optimality conditions under Separability: The differentiable case.}
In this section and the next one, we will study sufficient optimality conditions for separable quasiconcave programming when the objective function is multiplicatively decomposed.\\
First, we recall the definition and some properties of a separable product function.
\begin{deft}\label{deft 3.1.}
Let $X$ be a subset of $\mathbb{R}^n$.
A function $f:X \rightarrow \mathbb{R}$ is said to be multiplicatively separable if it satisfies the following condition:\\
$(MS):$ there exist subsets $X_i$ of $\mathbb{R}^{n_i}$,
and functions $f_i:X_i \rightarrow \mathbb{R}$ $\left( i=1,...,m \right)$ such that
$$ f(x)=\prod_{i=1}^m f_i(x_i) $$
where $\displaystyle{ x=(x_1,...,x_m)\in X=X_1 \times \cdots\times X_m}$
and $\displaystyle{\sum_{i=1}^m n_i=n}$ \\
\end{deft}
First, we recall a necessary condition for the function $f$ to be quasiconcave. See \cite{B-H}
\begin{pro}\label{pro 3.1.}(Lemma 3.4. \cite{B-H})
Let $f$ be a real function defined on a convex set $X \subset \mathbb{R}^n$ verifying the condition $(MS)$.
If $f$ is quasiconcave on $X$, then all $f_i \ (i=1,...,m)$ are quasiconcave on $X_i$.
\end{pro}
Now we recall the definition and some properties of the multiplicative concavity index of a function introduced by Crouzeix and Kebbour in \cite{C-K}. Such an index was the basic tool to study the (generalized)concavity of a function.
\begin{deft} {(Crouzeix-Kebbour \cite{C-K})}\label{deft 3.2.}
The multiplicative concavity index $i_{cv}(f)$ of a function $f:X\rightarrow (0,\infty)$ is defined as follows:
\begin{itemize}
\item If there exists $\lambda < 0$ such that $f^{\lambda}$ is not convex then
$$i_{cv}(f) = \displaystyle \sup \{\mu < 0 : f^{\mu} \ \textrm{ is convex} \}$$
\item If $f^{\lambda} $\ is convex for every $\lambda < 0 $, then
$$i_{cv}(f) = \displaystyle\sup\{\mu > 0 : f^{\mu}\ \textrm{is concave} \}$$
\end{itemize}
\end{deft}
The following proposition is an immediate consequence of Definition \ref{deft 3.2.}. (See \cite{C-K,C-L}).
\begin{pro}\label{pro 3.2.}Let $X$ and $f$ as in Proposition \ref{pro 3.1.}. Then,
\item[(a)] If $\displaystyle {i_{cv}(f) > -\infty}$, then $f$ is quasiconcave ;
\item[(b)] $f$ is log-concave if and only if \ $\displaystyle{ i_{cv}(f) \geq 0}$ ;
\item[(c)] $f$ is concave if and only if \ $\displaystyle{i_{cv}(f) \geq 1}$ ;
\item[(d)] $f$ is constant if and only if \ $\displaystyle{i_{cv}(f) = +\infty}$ and $X$ is open ;
\item[(e)] Let $\displaystyle{\alpha > 0}$, then \ $\displaystyle{i_{cv}(f^{\alpha}) = \frac{1}{\alpha} i_{cv}(f)}$ ;
\item[(f)] $\displaystyle{i_{cv}(f) = \inf \{i_{cv}(f_{x,d}) : x\in X ,d \in \mathbb{R}^{n} \setminus \{ 0 \} \}}$.
\end{pro}
The following proposition, which reveals an interesting property of separable quasiconcave functions, will be used frequently in the sequel of this paper (see \cite{B-H}).
\begin{pro}\label{pro 3.3.}(Theorem 3.12.\cite{B-H})
For $i = 1,...,m$, let $X_{i}$ be a non-empty open convex subset of $\mathbb{R}^{n_{i}}$, and let $f_{i}$ be a positive non constant function defined on $X_{i}$, and $f$ be the function defined on the product space $X = X_{1} \times X_{2} \times ... \times X_{m}$ by $$f(x_{1} ,...,x_{m}) = \displaystyle\prod_{i=1}^{m}f_{i}(x_{i})$$
\begin{itemize}
\item[i)] The function $f$ is quasiconcave if and only if one of the following holds:
\begin{itemize}
\item[a)] all functions $f_{i}$ are log-concave.
\item[b)] all functions $f_{i}$ except one are log-concave and
\begin{equation}\label{eq 1}
\displaystyle\sum_{i = 1}^{n} \frac{1}{i_{cv}(f_{i})} \leq 0
\end{equation}
\end{itemize}
\item[ii)] If $f$ is quasiconcave then
\begin{equation}\label{eq 2}
\frac{1}{i_{cv}(f)} = \displaystyle\sum_{i = 1}^{n} \frac{1}{i_{cv}(f_{i})}
\end{equation}
\end{itemize}
with the convention $\displaystyle{ \frac{1}{0} = \infty }$.
\end{pro}
\begin{rem}\label{rem 3.1}
Notice that if all $f_{i} \ (i=1,\cdots,m) $ are differentiable log-concave, then so is the function $f$, and then it is pseudoconcave. If not, there exists $i_0\in \left\lbrace1,\cdots,m\right\rbrace$ such that $f_{i_o}$ is not log-concave. It is clear that $f_{i_0}$ is $i_{cv}(f_{i_0})$-concave with $i_{cv}(f_{i_0})<0$, then, from Proposition \ref{pro 2.6.}., $f_{i_0}$ is pseudoconcave on $X$, hence, $f$ is also pseudoconcave on $X$.(For more details see the proof of Theorem $5.3.$ in \rm{Crouzeix-Hassouni} \cite{C-H} ).
\end{rem}
\subsubsection*{Unconstrained problem.}
Consider the unconstrained problem
$$\displaystyle{(USQP)}:\qquad \displaystyle{ \max_{x\in X} f(x)} $$
where $f$ satisfies the condition $(MS)$ in Definition \ref{deft 3.1.}.
\begin{theo}\label{Theo 3.1.}
Assume that all $f_i$ are differentiable, and $f$ is quasiconcave on $X$. Then,
\item[i)] If $\bar{x}$ is a critical point of $f$ then $\bar{x}$ is a global maximum of $(USQP)$.
\item[ii)] If $\bar{x}$ is a local maximum of $(USQP)$ then it is a global maximum.
\end{theo}
\begin{proof}
$i)$ Suppose that $\bar{x}$ is a critical point, that is $\nabla f(\bar{x}) =0$, since $\displaystyle{ \left\langle\nabla f(\bar{x}), x-\bar{x}\right\rangle=0 \ \forall x\in X}$ then, by the pseudoconcavity of $f$ , one has $ f(x)\leq f(\bar{x}) $ for all $x\in X.$ \\
$ii)$ Let $ \bar{x} $ be a local maximum of $f$, then there is a neighbourhood $\mathcal{N}(\bar{x}) $ of $ \bar{x} $, such that for all $ x \in \mathcal{N} ( \bar{x}) \cap X $ we have $ f(x) \leq f (\bar{x})$. \\
Let $ x \in X $ such that $ x \not \in \mathcal{N} ( \bar{x}) $. There exists $ \tilde{ \lambda} \in (0,1) $ such that $ \tilde{x} = (1- \tilde{ \lambda}) \bar{x} + \tilde{ \lambda} x \in \mathcal{N} ( \bar{x}) \cap X $. \\
By Proposition \ref{pro 2.4.}. the restriction of $f$ to the line segment $\left[\bar{x},x \right]$ is pseudoconcave.
If $ f (x) > f( \bar{x}) $ then by the semi-strict quasiconcavity of $ f $ (see Proposition \ref{pro 2.3.}.), one has $ f ( \tilde {x}) > f( \bar{x }) $ which is a contradiction. Thus $ \bar{x} $ is a global maximum .
\end{proof}
\subsubsection*{Constrained problem.}
Now consider the constrained problem
\begin{center}
$(CSQP)$:
\begin{tabular}{lcr}
& $\displaystyle{ \max_{x\in X} f(x)}$ & \\
subject to & $h_j(x)\geq 0$ & $j=1,\cdots, p$
\end{tabular}
\end{center}
where $f$ satisfies the condition $(MS)$ in Definition \ref{deft 3.1.}.\\
Define the feasible set $F=\left\lbrace x\in X :h_j(x) \geq 0, \ j=1,...,p \right\rbrace $.
\begin{theo}\label{Theo 3.2.}
Assume that $f$ and all $h_j \ (j=1,...,p)$ are differentiable and quasiconcave on $X$.
Let $\bar{x}$ be a feasible point such that $\displaystyle{ \nabla h_j(\bar{x}) \neq 0 }$ for all $j$ and $h_j(\bar{x}) >0$ for some $j$.\\ If there exist $\lambda_j\in \mathbb{R}, j = 1,\cdots,p,$ such that
\begin{equation}\label{eq 3}
\nabla f (\bar{x})+\sum_{j=1}^p\lambda_j \nabla h_j(\bar{x})=0,
\end{equation}
\begin{equation}\label{eq 4}
\lambda_jh_j(\bar{x})=0,\ j=1,\cdots,p,
\end{equation}
\begin{equation}\label{eq 5}
\lambda_j\geq 0,\ j=1,\cdots,p,
\end{equation}
then $\bar{x}$ is a global solution of $(CSQP).$
\end{theo}
\begin{proof}
Assume, by contradiction, that there exists $x_0 \in F$ such that
$f(x_0) > f(\bar{x})$. By the pseudoconcavity of $f$ we have $\left\langle\nabla f(\bar{x}), x_0- \bar{x}\right\rangle > 0$.\\
Since $h_j(x_0)\geq 0 = h_j(\bar{x}),\ j\in J(\bar{x})=\left\lbrace j / h_j(\bar{x})=0\right\rbrace$, the quasiconcavity of $h_j$ implies $\left\langle\nabla h_j(\bar{x}), x_0-\bar{x}\right\rangle\geq 0,\ j \in J(\bar{x})$.\\
From the complementarity condition (\ref{eq 4}), we have
$\lambda_j = 0, \forall j \not\in J(\bar{x})$, hence:\\ $\displaystyle{\left\langle\nabla f(\bar{x}), x_0-\bar{x}\right\rangle +\sum_{j=1}^p\left\langle\lambda_jh_j(\bar{x}), (x_0-\bar{x}\right\rangle > 0}$,
which contradicts (\ref{eq 3}).
\end{proof}
\begin{ex}
Consider the Cobb-Douglas utility function defined by:
$$ u(x_{1},x_{2},...,x_{n}) = \prod_{i=1}^{n}u_{i}(x_{i}) $$
where $u_{i}(x_{i}) = x_{i}^{\alpha_{i}} $ with $x_{i} > 0$ and $\alpha_{i} > 0$ ; $i=1,...,n$ \\
Consider the problem of maximization of $u$ with constraint budget:
\begin{center}
$(P)$:
\begin{tabular}{lcr}
& $\displaystyle{\max u(x)}$ & \\
subject to & $G(x) \leq B$ &
\end{tabular}
\end{center}
where $G(x)= p_1x_1+p_2x_2+...+p_nx_n $.
From Proposition \ref{pro 3.3.}., $u$ is quasiconcave , and since it is differentiable with $\nabla u(x) \neq~0$ for all $x$, then, by Proposition \ref{pro 2.5.}., $u$ is pseudoconcave.\\
Let $H(x)=B-p_1x_1-p_2x_2-...-p_nx_n$. It is clear that
$H$ is quasiconcave and differentiable.\\
Let $\bar{x}=(\bar{x}_1,...,\bar{x}_n)$ a feasible point, that is $H(\bar{x})\geq0$, and let $\lambda \in \mathbb{R}$ such that $(\bar{x},\lambda)$ satisfies the $KKT$ conditions (\ref{eq 3}), (\ref{eq 4}) and (\ref{eq 5}) in Theorem \ref{Theo 3.2.}. \\
By (\ref{eq 3}) one has:
$ \displaystyle{ \alpha_i \bar{x}_1^{\alpha_1}\bar{x}_2^{\alpha_2}\cdots \bar{x}_i^{\alpha_i -1}\cdots \bar{x}_n^{\alpha_n}=\lambda p_i }$ for all $i=1 \cdots n$.
Since $\bar{x}_i>0, \alpha_i>0, p_i>0, (i=1,...,n)$ then $\lambda \neq 0$, and $\displaystyle{\frac{\displaystyle{\prod_{j=1}^{n}}(\bar{x}_{j})^{\alpha_{j}}}{\lambda}=\frac{p_i}{\alpha_{i}}\bar{x}_i}$,
thus $\displaystyle{ \bar{x}_i= \frac{\alpha_i p_1}{\alpha_1 p_i}\bar{x}_1 }$ for all $i=1,...,n$.\\
By (\ref{eq 4}) and since $\lambda \neq 0$ one has $H(\bar{x}_1,...,\bar{x}_n)=0$, \rm{i.e.} $p_1\bar{x}_1+...+p_n\bar{x}_n=B$,
then;\\ $\displaystyle{ \bar{x}_1=\frac{\alpha_1 B}{(\alpha_1+...+\alpha_n)p_1} }$,
thus $\displaystyle{ \bar{x}=\big(\frac{\alpha_1 B}{(\alpha_1+...+\alpha_n)p_1},...,\frac{\alpha_n B}{(\alpha_1+...+\alpha_n)p_n} \big)}$ is a solution of $(P)$.
\end{ex}
\section{Optimality conditions under Separability: The non differentiable case.}
In this section we will study the problems $\rm{(USQP)}$ and $\rm{(CSQP)}$ studied in the previous section when the objective function and the constrained functions are not necessarily differentiable.
First, notice that in Proposition \ref{pro 3.3.}., any assumption of differentiability of $f$ is required. \\
Secondary, we recall that the notion of pseudoconcavity can be extended to non differentiable case by means of a generalized superdifferential instead of the classical gradient.\\
Let's define an abstract superdifferential in the same sense as the abstract subdifferential defined by Aussel et al in \cite{A-C-L}.
\begin{deft}\label{deft 4.1}
We call superdifferential, denoted by $\stackrel{\frown}{\partial}$, any operator which associates
a subset $\stackrel{\frown}{\partial}f(x) $ of $\mathbb{R}^n$ to any upper semi-continuous function $f : X \rightarrow \mathbb{R}\cup \left\lbrace -\infty\right\rbrace$
and any $x\in X$ such that $f(x)$ is finite, and satisfies the following properties:
\item[(P1)]$\stackrel{\frown}{\partial} f(x) = \left\lbrace x^*\in \mathbb{R}^n:\left\langle x^*,y-x\right\rangle + f(x) \geq f(y), \forall y \in X \right\rbrace $ whenever $f$ is concave;
\item[(P2)]$0 \in \stackrel{\frown}{\partial}f(x) $ whenever $x$ is a local maximum of $f$;
\item[(P3)]$ \stackrel{\frown}{\partial}(f+g)(x)\subset \stackrel{\frown}{\partial} f(x) + \stackrel{\frown}{\partial} g(x)$ whenever $g$ is a real-valued concave continuous function which is $\stackrel{\frown}{\partial}$-differentiable at $x$.\\
where $g$ is $\stackrel{\frown}{\partial}$-differentiable at $x$ means that both $ \stackrel{\frown}{\partial}g(x)$ and $\stackrel{\frown}{\partial}(-g)(x)$ are
non-empty.
\end{deft}
\begin{exs}
Let's recall the Clarcke-Rockafellar subdifferential $\partial^{CR}$ and the upper-Dini subdifferential $\partial^{D+}$ for a lower-semicontinuous function
$f : X \rightarrow \mathbb{R}\cup \left\lbrace +\infty\right\rbrace$:
$$\displaystyle{\partial^{CR}f(x):=\left\lbrace x^*\in X^* : \left\langle x^*,v\right\rangle\leq f^{\uparrow} (x;v) \ \forall v\in X\right\rbrace} $$
$${\rm with}\qquad \displaystyle{ f^{\uparrow}(x;v) = \sup_{\varepsilon>0}\limsup_{x'\rightarrow_f x\atop {t\searrow 0}}\inf_{v'\in B_{\varepsilon}(v)}\frac{f(x'+tv')-f(x')}{t},}$$
and
$$\displaystyle{\partial^{D^+}f(x):=\left\lbrace x^*\in X^* : \left\langle x^*,v\right\rangle\leq f^{D^+}(x,v),\ \forall v\in X\right\rbrace}$$
$${\rm with}\qquad\displaystyle{f^{D^+}(x,v)=\limsup_{t\searrow 0}\frac{f(x+tv)-f(x)}{t}}$$
where the notation $\displaystyle{x'\rightarrow_f x} $ means that $x'\rightarrow x$ and $f(x')\rightarrow f(x)$.\\
Let's define the Clarcke-Rockafellar superdifferential $\stackrel{\frown}{\partial}^{CR}$ and the upper-Dini superdifferential $\stackrel{\frown}{\partial}^{D+}$ for an upper-semicontinuous function $f : X \rightarrow \mathbb{R}\cup \left\lbrace -\infty\right\rbrace$:
$$\stackrel{\frown}{\partial}^{CR}f(x):=-\partial^{CR}(-f)(x) $$
$$\stackrel{\frown}{\partial}^{D+}f(x):=-\partial^{D+}(-f)(x)$$
These two superdifferentials check, among others, the above abstract superdifferential's properties in Definition \ref{deft 4.1}.
\end{exs}
We recall that $\partial^{CR}$ and $\partial^{D+}$ contain the best known subdifferentials such as the lower Hadamard subdifferential $\partial ^{H-}$, the Fr\'echet subdifferential $\partial ^F$ and the lipschitz subdifferential $\partial ^{LS}$.\\
In the sequel, we will use the symbol $\stackrel{\frown}{\partial}$ to mean either $\stackrel{\frown}{\partial}^{CR}$ or $\stackrel{\frown}{\partial}^{D+}$.\\
The following proposition, which extend Proposition \ref{pro 2.1.}., is an immediate consequence of Definition \ref{deft 4.1}. and Theorem $2.1.$ in \cite{Aus}.
\begin{pro}\label{pro 4.1}
Let $X$ be a nonempty convex subset of $\mathbb{R}^n $ and let $f: X \longrightarrow \mathbb{R}$ be an upper-semicontinuous function. Then, the following assertions are
equivalent:
\item[i)] $f$ is quasiconcave;
\item[ii)] $\displaystyle{ \left(\exists x^*\in \stackrel{\frown}{\partial}f(x) : \left\langle x^*,y-x\right\rangle < 0\right) \Rightarrow f(y) > f(z) \qquad \forall z\in \left[x;y\right])} $
\end{pro}
From Definition \ref{deft 4.1}. and Proposition \ref{pro 4.1}., the Proposition \ref{pro 2.2.}. can be extended as follows:
\begin{pro}\label{pro 4.2.}
Let $f$ be an upper semi-continuous and quasiconcave real valued function on a convex set $X$ of $\mathbb{R}^n$, and let $x,\bar{x}\in X$. Then,
$$x\in clU^s(f,f(\bar{x})) \Rightarrow \left\langle \bar{x}^*,x-\bar{x}\right\rangle \geq 0 \qquad \forall \bar{x}^*\in \stackrel{\frown}{\partial}f(\bar{x})$$
\end{pro}
Now, we extend the definition and some properties of pseudoconcavity to the non differentiable case.
\begin{deft}\label{deft 4.2} Let $X$ be a nonempty convex subset of $\mathbb{R}^n $ and let $f: X \longrightarrow \mathbb{R}$ be an upper-semicontinuous function.
$f$ is said to be pseudoconcave with respect to $ \stackrel{\frown}{\partial}$ (in short $\stackrel{\frown}{\partial}$-pseudoconcave) if, for any $x, y\in X$, one has
$$\displaystyle{ \left( \exists x^*\in \stackrel{\frown}{\partial} f(x) : \left\langle x^*,y-x\right\rangle \leq 0\right) \Rightarrow f(y)\leq f(x)}.$$
\end{deft}
As in the differentiable case, every $\stackrel{\frown}{\partial}$-pseudoconcave function satisfies the following fundamental properties:
\begin{pro}\label{pro 4.3.}
Let $f$ be an upper-semicontinuous function on $X$. Then, $f$ is $\stackrel{\frown}{\partial}$-pseudoconcave on $X$ if and only if the restriction of $f$ to any line segment in $X$ is $\stackrel{\frown}{\partial}$-pseudoconcave.
\end{pro}
The proof follows from the radial property of the generalized derivatives $f^{\uparrow}$ and $f^{D+}$.
\begin{pro}\label{pro 4.4.}
Let $f: X\rightarrow\mathbb{R}$ be an upper-semicontinuous and radially continuous function. Then, the following assertions are equivalent:
\item[(i)]$f$ is $\stackrel{\frown}{\partial}$-pseudoconcave;
\item[(ii)]$f$ is quasiconcave and ($0\in \stackrel{\frown}{\partial} f(x) \Rightarrow f$ has a global maximum at $x$).
\end{pro}
The proof is a direct consequence of Definition \ref{deft 4.2}. and Proposition \ref{pro 4.1}.
\subsubsection*{Unconstrained problem}
Let $f$ be a real valued function defined on a non empty open convex set $X$ of $\mathbb{R}^n$ satisfying the separability condition $(MS)$ in Definition \ref{deft 3.1.}. Consider the unconstrained problem:
$$\displaystyle{ \rm{(USQP)}: \qquad \max_{x\in X} f(x) }$$
\begin{theo}\label{theo 4.1.}
Assume that $f$ is upper-semicontinuous, quasiconcave and radially continuous on $X$. Then,
if $\bar{x}$ solves $(USQP)$ locally, then it is a global solution.
\end{theo}
\begin{proof}
By the separability condition $(MS)$ and from Proposition \ref{pro 3.3.}. and Remark \ref{rem 3.1}., the function $f$ is either log-concave or $r$-concave, and then it is $\stackrel{\frown}{\partial}$-pseudoconcave. If $\bar{x}$ is a local maximum of $f$, then by $(P2)$ in Definition \ref{deft 4.1}., $\displaystyle{0\in \stackrel{\frown}{\partial}f(\bar{x})}$, thus from Proposition \ref{pro 4.4.}., $\bar{x}$ is a global maximum of $f$.
\end{proof}
\begin{coro}
Let $f$ as in Theorem \ref{theo 4.1.}. Assume that $\displaystyle{\bar{x}\notin clU^s(f,f(\bar{x}))}$. Then $\bar{x}$ is a global solution of $(USQP)$.
\end{coro}
\begin{proof}
If $\displaystyle{\bar{x}\notin clU^s(f,f(\bar{x}))}$, then there exists a neighbourhood $\displaystyle{\mathcal{N}(\bar{x})}\subset X$ of $\bar{x}$ such that $\displaystyle{ \mathcal{N}(\bar{x})\cap clU^s(f,f(\bar{x}))=\emptyset}$, thus $\bar{x}$ is a local maximum of $f$, hence by the $\stackrel{\frown}{\partial}$-pseudoconcavity of $f$ and Theorem \ref{theo 4.1.}., $\bar{x}$ is a global solution of $(USQP)$.
\end{proof}
\subsubsection*{Constrained problem.}
Let $f$ be an upper-semicontinuous and quasiconcave real valued function defined on $X$ satisfying the separability condition $(MS)$ in Definition \ref{deft 3.1.}. Consider the constrained problem:
\begin{center}
$(CSQP)$:
\begin{tabular}{lcr}
& $\displaystyle{ \max_{x\in X} f(x)}$ & \\
subject to & $h_j(x)\geq 0$ & $j=1,\cdots, p$
\end{tabular}
\end{center}
where $h_j \ (j=1,...,p)$ are quasiconcave and upper-semicontinuous on $X$.\\
Let's define the feasible set $\displaystyle{\mathcal{F}=\left\lbrace x\in X :h_j(x) \geqslant 0, \ j=1,...,p \right\rbrace }$.\\
For $x\in \mathcal{F}$, denote $\displaystyle{ J(x)=\left\lbrace j:h_j(x)=0 \right\rbrace }$.
\begin{pro}\label{pro 4.5.}
Let $\bar{x}$ be a feasible point of $(CSQP)$. For $j\in J(\bar{x})$, assume that $\displaystyle{0 \notin \stackrel{\frown}{\partial}h_j(x)}$ whenever $x\in \mathcal{F}$ and $h_j(x)=0$. Then, $\displaystyle{ \left\langle \bar{x}^*,x-\bar{x}\right\rangle \geq 0 \ \forall \bar{x}^*\in \stackrel{\frown}{\partial}h_j(\bar{x})}$.
\end{pro}
\begin{proof}
For $j\in J(\bar{x})$, and $x\in \mathcal{F}$ such that $h_j(x)=0$, if $\displaystyle{0 \notin \stackrel{\frown}{\partial}h_j(x)}$, then, by $(P_2)$ in Definition \ref{deft 4.1}., $x$ is not a local maximum of $h_j$. Suppose that there exists $\displaystyle{\bar{x}^*\in \stackrel{\frown}{\partial}h_j(\bar{x})} $, such that $\displaystyle{ \left\langle \bar{x}^*,x-\bar{x}\right\rangle < 0 }$. Since $\displaystyle{h_j(x)=h_j(\bar{x})=0}$, Proposition \ref{pro 4.2.}. yields $\displaystyle{x\notin clU^s(h_j, h_j(x))}$. Thus $x$ is a local maximum of $h_j$, which is a contradiction.
\end{proof}
Let's define a variant of the well known K.K.T. conditions that we show as a sufficient optimality conditions for the constrained problem $(CSQP)$.
\begin{deft}\label{deft 4.3.}
We say that a pair $(\bar{x},\bar{\lambda}) \in X \times \mathbb{R}^p$ satisfies the
modified Karush-Kuhn-Tucker conditions (m-K.K.T. conditions) if it satisfies the super-gradient condition:
\begin{equation}\label{eq 6}
0\in \stackrel{\frown}{\partial}f(\bar{x})+\sum_{j=1}^p\lambda_j \stackrel{\frown}{\partial} h_j(\bar{x})-N_X(\bar{x})
\end{equation}
where $\displaystyle{N_X(\bar{x})}$ is the normal cone of $X$ at $\bar{x}$,
and also the usual complementary slackness conditions:
\begin{equation}\label{eq 7}
\lambda_jh_j(\bar{x})=0,\ j=1,\cdots,p
\end{equation}
\begin{equation}\label{eq 8}
h_j(\bar{x})\geq 0,\ j=1,\cdots,p
\end{equation}
\begin{equation}\label{eq 9}
\lambda_j\geq 0,\ j=1,\cdots,p
\end{equation}
\end{deft}
Let's recall the {\itshape{Slater constraint qualification}}: there exists $\tilde{x}$ in $X$, called a Slater point for $\rm{(CSQP)}$, such that
$h_j(\tilde{x})>0$ for some $\displaystyle{j\in \left\lbrace1,\cdots,p\right\rbrace}$.
\begin{theo}\label{Theo 4.2.}
Let $\bar{x}$ be a feasible point of $(CSQP)$. Assume that $(CSQP)$ has a Slater point, and $0 \not\in \stackrel{\frown}{\partial}h_j(x)$ whenever $x\in \mathcal{F}$ and $h_j(x)=0 $. If there exists $\bar{\lambda}\in \mathbb{R}^p$ such that $(\bar{x},\bar{\lambda})$ satisfies the m-K.K.T. conditions, then $\bar{x}$ is a solution of $(CSQP)$.
\end{theo}
\begin{proof}
Assume, by contradiction, that there exists a feasible point $x_0$ such that $f(x_0) > f(\bar{x})$.\\
By Proposition \ref{pro 3.3.}, because of separability condition, $f$ is actually $\stackrel{\frown}{\partial}$-pseudoconcave and then, for all $\displaystyle{\bar{x}^{*}\in\stackrel{\frown}{\partial}f(\bar{x})}$ one has $\displaystyle{\left\langle \bar{x}^*, x_0- \bar{x}\right\rangle > 0}$.\\
Since $N_X(\bar{x})$ coincides with the normal cone of convex analysis when $X$ is convex (see \cite{Cl}), then for all $v\in N_X(\bar{x})$ one has
$\left\langle v,x_0-\bar{x}\right\rangle \leq 0$. Thus, for all $\bar{x}^*\in \stackrel{\frown}{\partial} f (\bar{x})$, $\bar{x}_j^*\in \stackrel{\frown}{\partial} h_j(\bar{x})$ and $v\in N_X(\bar{x})$ one has:
\begin{equation}\label{eq 10}
\left\langle \bar{x}^*-v , x_0-\bar{x} \right\rangle > 0
\end{equation}
If $j\in J(\bar{x})$, by Proposition \ref{pro 4.5.}., one has for all $\bar{x}_j^* \in \stackrel{\frown}{\partial} h_j(\bar{x})$:
\begin{equation}\label{eq 11}
\left\langle \bar{x}_j^*, x_0-\bar{x}\right\rangle \geq 0
\end{equation}
From the condition (\ref{eq 7}),
$\lambda_j = 0$ for all $j \not\in J(\bar{x})$.
Adding (\ref{eq 11}) for $j=1,\cdots,p$, and combining with (\ref{eq 10}) we get:
$$\displaystyle{ \left\langle \bar{x}^*, x_0- \bar{x}\right\rangle + \sum_{j=1}^p \lambda_j\left\langle \bar{x}_j^*, x_0-\bar{x}\right\rangle - \left\langle v,x_0-\bar{x} \right\rangle >0}$$
for all $\displaystyle{\bar{x}^*\in \stackrel{\frown}{\partial} f (\bar{x})}$, $\displaystyle{ \bar{x}_j^* \in \stackrel{\frown}{\partial} h_j(\bar{x}), \ (j=1,...,p) }$ and $\displaystyle{v\in N_X(\bar{x})}$, which contradicts (\ref{eq 6}).
\end{proof}
|
\section{Introduction}
\label{sec:intro}
Several groups have reported an excess of $\gamma$ rays in the data collected by the {\it Fermi} Large Area Telescope ({\it Fermi}-LAT) in the direction of the Galactic center region, known as the Galactic center excess \citep[GCE, e,.g.][]
{Goodenough:2009gk,Hooper:2010mq,Boyarsky:2010dr,Hooper:2011ti,Abazajian:2012pn,Gordon:2013vta,Abazajian:2014fta,Daylan:2014rsa,Calore:2014xka,TheFermi-LAT:2015kwa,TheFermi-LAT:2017vmf}.
This GCE has been significantly detected when modeling the $\gamma$-ray sky using a range of interstellar emission models (IEMs), source catalogs, data selections and techniques.
Most of the references cited above agree on the fact that the GCE spectral energy distribution (SED, $E^2 dN/dE$) is peaked at few GeV.
However, other properties are much more uncertain and still under debate.
For example, the GCE has been claimed in \cite{Daylan:2014rsa} to be spherically symmetric and centered around the Galactic center.
Moreover, the GCE energy spectrum can be well modeled as $\gamma$ rays produced by dark matter (DM) particles annihilating into $b\bar{b}$ with a thermal annihilation cross section, which is the proper cross section to explain the observed density of DM in the Universe \cite{Aghanim:2018eyx}.
All these characteristics make the GCE very appealing for indirect DM searches.
In other publications (e.g., Refs.~\cite{Calore:2014xka,TheFermi-LAT:2015kwa,TheFermi-LAT:2017vmf}) it has been shown that, using different IEMs and source catalogs, the measurement of some GCE properties, in particular the normalization of the spectrum, are too uncertain to conclude that it is of DM origin.
For example, Refs.~\cite{Macias:2016nev,Bartels:2017vsx} claimed that the spatial shape of the GCE is better modeled with the distribution of stars in the bulge of our Galaxy.
In particular they demonstrated that fitting the GCE with two emission templates associated with the galactic bulge provides a much better fit than using a DM template.
This result implies that the GCE is not spherically symmetric since the boxy bulge has an ellipsoidal shape.
Refs.~\cite{Bartels:2015aea,Lee:2015fea} likewise raise doubts on a DM origin of the GCE.
By applying wavelet analysis and non-Poissonian template fitting to {\it Fermi}-LAT data, they published compelling evidence for the existence of a faint population of sources located in the Galactic center with properties that can explain the GCE.
The presence of these sources could be interpreted as a population of millisecond pulsars located around the bulge of our Galaxy.
However, very recently, Refs.~\cite{Leane:2019uhc,Chang:2019ars} have shown that the non-Poissonian template fitting method can misattribute un-modeled point sources or imperfections in the IEM to a signal of a faint population of sources or DM.
In Ref.~\cite{Buschmann:2020adf} they rerun the non-Poissonian template fitting with an improved description of the Galactic diffuse emission and found that this method continues to robustly favor the interpretation that the GCE is due, in part, to unresolved astrophysical point sources. Finally, Ref.~\cite{List:2020mzd} uses Bayesian graph convolutional neural networks and reports that the GCE is almost entirely attributed to smooth emission.
All these results challenge the robustness of the results presented in \cite{Lee:2015fea} and the conclusion that the GCE is due to a population of pulsars.
Correspondingly, Ref.~\cite{Zhong:2019ycb} has applied wavelet analysis, similarly to what has been done in \cite{Bartels:2015aea}, to about 10 years of LAT data using the latest 4FGL catalog released by the {\it Fermi}-LAT Collaboration \cite{Fermi-LAT:2019yla}. They find that the GCE is still present, although they do not find any compelling evidence for the existence of a faint population of un-modeled sources.
Alternative interpretations for the GCE are associated to $\gamma$ rays produced, through inverse Compton scattering, bremsstrahlung or $\pi^0$ decays, by cosmic rays injected from the Galactic center during recent outbursts \cite{Carlson:2014cwa,Petrovic:2014uda,Gaggero:2015nsa}.
These mechanisms, however, provide $\gamma$-ray signals not fully compatible with the GCE properties.
Specifically, the hadronic scenario (i.e., mostly cosmic-ray protons) predicts a $\gamma$-ray signal that is distributed along the Galactic plane, since the $\pi^0$ decay process is correlated with the distribution of gas present in the Milky Way disk \cite{Petrovic:2014uda}.
On the other hand, a leptonic outburst would lead to a signal that is approximatively spherically symmetric.
However, this model requires a complicated scenario with at least two outbursts
to explain the morphology and the intensity of the excess with the older
outbursts injecting more-energetic electrons.
The puzzle of the GCE origin is thus far from being solved.
Indeed, some of the GCE properties, for example its spectrum and spatial morphology, are still under debate.
Moreover, the choice of the IEM significantly changes the characteristics of the GCE, making it very challenging at the moment to make significant progress in the interpretation.
In Ref.~\cite{Dimaurosim} we have studied this problem using simulated {\it Fermi}-LAT data of the Galactic center region.
We have shown that the systematic uncertainties due to the IEM are at the level of 15\% in the normalization of the energy spectrum and 5\% in the value of the DM density profile slope used to model the GCE spatial morphology.
Finally, we have found that an additional component in the Galactic center with a flux of the order of the {\it Fermi}-LAT bubbles can modify the results for the GCE spectrum and spatial extension at roughly the $20\%$ level.
In order to make progress on this problem, we analyze $11$ years of {\it Fermi}-LAT data using the state of the art source catalog and IEMs and new analysis techniques and we provide accurate measurements for several properties of the GCE.
We list below the main novel results with respect to previous papers on the same subject.
We use a new analysis method called weighted likelihood, recently developed by the {\it Fermi}-LAT Collaboration to mitigate the systematic uncertainties due to the mismodeling of the IEM in the Galactic plane and at energies lower than 1 GeV. As we have demonstrated in Ref.~\cite{Dimaurosim}, this technique yields more robust results for the GCE at low energies\footnote{In order to reduce significantly the systematic uncertainties due to the modeling of the IEM improved models should be studied and used in the analysis. However, as we will explain in Sec.~\ref{sec:analysistec}, we use the weighted maximum likelihood technique that effectively reduces the differences in the results obtained with different IEMs. See Ref.~\cite{Fermi-LAT:2019yla} for a detailed discussion of this method.}.
We employ the newest 4FGL catalog of sources that contains almost twice the number of sources with respect to previous {\it Fermi}-LAT catalogs used for previous analyses.
We utilize the recently created event class {\tt SOURCEVETO} which increases the effective area of the data selections while minimizing the CR background contamination.
We make a systematic study of several aspects of the GCE such as the flux between 0.1 GeV to 1 TeV, its position, sphericity and the spatial morphology employing the state of the art IEMs provided in \cite{TheFermi-LAT:2017vmf}.
In order to derive the spatial morphology we use specific models with DM or leptonic outbursts. We also employ a model independent technique (i.e., with no dependence on the specific GCE physics interpretation) that measures, for the first time, the GCE surface brightness (i.e., the flux as a function of the angular distance) from $0^{\circ}$ up to $15^{\circ}$ and the energy dependence of its spatial morphology.
Utilizing these new techniques and analyses, we find new clues to the origin of the GCE.
The paper is organized as follow: in Sec.~\ref{sec:analysis} we describe the selection we apply to {\it Fermi}-LAT data, the choice of IEM and the analysis technique.
In Sec.~\ref{sec:refinement} we inspect how many and which components of the IEMs should be left free to provide a good fit to the data;
in Sec.~\ref{sec:spectrum} we report our measurements for the spectrum of the GCE;
in Sec.~\ref{sec:spacial} we show the results we obtain for the spatial morphology with a model independent and with a model dependent technique;
in Sec.~\ref{sec:quadrants} we study the GCE with a DM template divided into quadrants in order to study the spatial symmetry of the excess;
and in Sec.~\ref{sec:position} we measure the GCE position relative to the Galactic center and we test its sphericity.
\section{Analysis Framework}
\label{sec:analysis}
In this section we provide the details of the {\it Fermi}-LAT data selection, background models and analysis pipeline.
\subsection{Data selection}
\label{sec:analysisdata}
We use 11 years of {\it Fermi}-LAT data from 2008 August 4 to 2019 August 4 passing standard data quality selection criteria\footnote{\url{https://fermi.gsfc.nasa.gov/ssc/data/analysis/documentation/Cicerone/Cicerone_Data_Exploration/Data_preparation.html}}.
We select an energy range between $0.1-1000$ GeV to study the GCE flux and spatial morphology. Then, since the GCE is significantly detected between $1-10$ GeV we focus on this energy range for analyzing its position and sphericity.
We consider a region of interest (ROI) of $40^{\circ} \times 40^{\circ}$ centered at the Galactic center.
The size of the ROI is motivated by the extension of the GCE that can be measured significantly up to $\sim10^{\circ}$ (e.g., \cite{Calore:2014xka,TheFermi-LAT:2017vmf}).
However, we also run the analysis selecting data from a $30^{\circ} \times 30^{\circ}$ region, in order to test how the results depend on the size of the ROI.
We select photon data belonging to the Pass~8 {\tt SOURCEVETO} event class and the corresponding instrument response functions {\tt P8R3\_SOURCEVETO\_V2}.
{\tt SOURCEVETO} has the same background rate of the {\tt SOURCE} class up to 10 GeV. Above 50 GeV it is the same as the {\tt ULTRACLEANVETO}, while having 15\% more acceptance\footnote{\url{https://fermi.gsfc.nasa.gov/ssc/data/analysis/documentation/Cicerone/Cicerone_Data/LAT_DP.html}.}.
This data selection is thus ideal to analyze diffuse emission such as the GCE.
We will also show results derived with {\tt SOURCE} and {\tt ULTRACLEANVETO} data and instrument response functions to show how the GCE properties change with a different selection of the data.
The data is binned using 8 energy bins per decade and a pixel size of $0.08^{\circ}$.
\subsection{Model components}
\label{sec:analysisiem}
The data measured in the direction of the Galactic center are modeled with the fluxes from individual sources, isotropic and IEM templates, and DM emission.
Point-like and extended sources are taken from the 4FGL {\it Fermi}-LAT catalog \cite{Fermi-LAT:2019yla} selecting all those that are in a region $48^{\circ}\times48^{\circ}$ centered at the Galactic center.
The source Sagittarius A$^{\star}$, that is usually considered to be the dynamical center of the Galaxy, is included in our background model with the source 4FGL J1745.6-2859. For a recent analysis where the spectrum, position of this source is analyzed and its cosmic-ray production is inspected see \cite{sagprep}.
We include sources up to $4^{\circ}$ outside of the ROI in order to account for the point spread function of the detector.
The isotropic template is taken from \cite{TheFermi-LAT:2017vmf} and we leave free in the fit its normalization.
We include the $\gamma$-ray emission from the Sun, the Moon, and Loop I as in \cite{TheFermi-LAT:2017vmf}.
The spatial morphology of $\gamma$ rays produced from DM particle interactions is modeled with a DM density profile ($\rho$) parametrized using a generalized Navarro-Frenk-White (NFW) model \cite{1997ApJ...490..493N}:
\begin{equation}
\rho = \frac{\rho_0}{\left( \frac{r}{r_s} \right)^{\gamma} \left( 1 + \frac{r}{r_s} \right)^{3-\gamma}},
\label{eq:NFW}
\end{equation}
where $\rho_0$ is a normalization, $r_s$ is the scaling radius and $\gamma$ is the slope of the DM density profile.
The analysis of the Galactic center region is insensitive to the value of the scaling radius $r_s$, which we fix to 20 kpc.
Instead, the value of $\gamma$ will be inferred from the analysis by fitting the GCE.
The $\gamma$-ray flux produced by DM annihilation is usually calculated as (see, e.g., \cite{1990NuPhB.346..129B}):
\begin{equation}
\frac{dN}{dE} (E) = \frac{1}{4\pi} \frac{\langle\sigma v\rangle}{m_\chi^2} \frac{dN_\gamma}{d E} \frac{1}{2}\mathcal{J},
\label{eq:flux_gamma}
\end{equation}
where the factor $1/2$ comes from considering self-conjugate DM particles, $m_\chi$ is the DM particle mass, $\langle\sigma v\rangle$ defines the annihilation cross section times the relative
velocity, averaged over the Galactic velocity distribution function, and $dN_\gamma/d E$ is the $\gamma$-ray production source spectrum per DM annihilation event which depends on the elementary processes ruling the annihilation of DM particles. $\mathcal{J}$ is the geometrical factor and is calculated with the integral performed in the solid angle $\Delta \Omega$ and along the line of sight (l.o.s., $s$) of the squared DM density distribution (see Eq.~\ref{eq:NFW}):
\begin{equation}
\mathcal{J} = \int_{\Delta \Omega} d\Omega \int_{l.o.s.} \rho^2 ds.
\label{eq:geomfactor}
\end{equation}
Finally, we place the center of the DM template at $(l,b)=(0^{\circ},0^{\circ})$ unless otherwise differently stated.
In this paper we are not interested in a specific DM particle physics theory interpretation of the GCE, and we accordingly
adopt a phenomenological GCE spectrum given by a log-parabola (LP)\footnote{See this page for the exact definition of this SED shape \url{https://fermi.gsfc.nasa.gov/ssc/data/analysis/scitools/source_models.html#spectralModels}.}:
\begin{equation}
\frac{dN_{{\rm{LP}}}}{dE} (E) = N_0 \left( \frac{E}{E_0} \right)^{-\alpha -\beta \log{(E/E_0)}},
\label{eq:logparabola}
\end{equation}
with the normalization $N_0$, spectral index $\alpha$, and curvature index $\beta$ free to vary in the fit. The pivot energy $E_0$ is fixed to 1 GeV.
As we will demonstrate in Sec.~\ref{sec:spectrum}, this SED shape reasonably fits the GCE spectrum.
The choice of a log-parabola model does not affect our results for the estimation of the GCE SED as for that analysis the spectral normalization is free to vary independently in each energy bin and the initial shape of the model is not important.
Moreover, when analyzing the spatial morphology, position and sphericity we select energies between 1--10 GeV and a slight change in the SED shape does not affect the results.
When we model the GCE with a DM template, we calculate its spatial morphology with the geometrical factor $\mathcal{J}$ evaluated at different longitudes and latitudes and normalized to the solid angle of the ROI ($\Delta \Omega$): $\bar{\mathcal{J}} = \mathcal{J}/\Delta \Omega$. Indeed, this is the way the {\tt Fermitools} deal with the template of extended sources\footnote{\url{https://fermi.gsfc.nasa.gov/ssc/data/analysis/scitools/extended/extended.html}}.
Finally, we compute the GCE flux with the product of a log-parabola SED (Eq.~\ref{eq:logparabola}) and the normalized geometrical factor:
\begin{equation}
\frac{dN}{dE} = \frac{dN_{{\rm{LP}}}}{dE} \times \bar{\mathcal{J}}.
\label{eq:fluxDM}
\end{equation}
The choice of the IEM is central in modeling such a complex region as the Galactic center.
Therefore we employ the templates created in \cite{TheFermi-LAT:2017vmf} which have been optimized to model the Galactic center region.
We refer to \cite{TheFermi-LAT:2017vmf} for the full details of these models and we summarize below the main characteristics.
The templates have been created with the {\tt Galprop} code\footnote{\url{http://galprop.stanford.edu}} \cite{1998ApJ...493..694M,2000ApJ...537..763S,2004ApJ...613..962S}, which calculates the propagation and interactions of cosmic rays in the Galaxy by numerically solving the transport equations given a model for the cosmic-ray source distribution, injection spectrum, and interaction targets.
We consider 11 models generated by assuming different source distributions, interstellar gas maps\footnote{In \cite{TheFermi-LAT:2017vmf} the contribution of HI and H2 gas is considered while the one from ionized gas is neglected.}, inverse Compton components, and {\it Fermi}-LAT bubble templates.
Therefore, these templates provide an appropriate framework, even if not exhaustive, to study how the GCE properties depend on the choice of the IEM.
We list below and in Tab.~\ref{tab:iem} the characteristics of each IEM.
{\bf Baseline}: The reference model, labelled as {\tt Baseline}, is taken from one of the models in \cite{2012ApJ...750....3A}.
It assumes a cosmic-ray source distribution traced by the distribution of pulsars reported in \cite{Lorimer:2006qs}.
The cosmic-ray confinement volume has a height of 10 kpc and a radius of 20 kpc.
The {\tt Baseline} model assumes HI column densities derived from the 21-cm line intensities for a spin temperature of 150 K.
The dust reddening map of \cite{1998ApJ...500..525S} is used to correct the HI maps to account for the presence of dark neutral gas not traced by the combination of HI and CO surveys \cite{2012ApJ...750....3A}.
Moreover, it includes the inverse Compton model reported in \cite{Porter_2008} and divided into the starlight, infrared and cosmic microwave background (CMB) components.
Finally, the model contains the Loop I, Sun, Moon emission and the {\it Fermi} bubbles divided into the low-latitude (closer to the Galactic center) and the high-latitude components.
The {\bf Baseline} model we use in this paper differs from the one used in \cite{TheFermi-LAT:2017vmf} for the inclusion of the low-latitude bubbles template.
{\bf Source distribution}: In order to account for different tracers of cosmic rays we will use different source distributions. We use an alternative model, labeled {\tt Yusifov}, generated using the pulsar distribution reported in \cite{2004AA...422..545Y}, the model {\tt SNR} created with distribution of supernova remnants published in \cite{1998ApJ...504..761C} (labelled as {\tt SNR}) and finally with the distribution of OB stars from \cite{2000AA...358..521B} ({\tt OBstars}).
Finally, we test a model labeled as {\tt Pulsars} where we include the bremsstrahlung and $\pi^0$ emission divided into the neutral atomic (HI) and molecular hydrogen (H2) components.
{\bf Gas and inverse Compton}: We test, in a model labeled {\tt ICS combined}, an inverse Compton template where the starlight, infrared and CMB components are merged in a unique template.
In order to account for different gas models we use templates generated from maps of the starlight extinction due to interstellar dust taken from the Variables in the Via Lactea survey \cite{2010NewA...15..433M} ({\tt SLext}) and using the high-resolution maps from the GASS survey \cite{2010AA...521A..17K} and the dust extinction map from extinction map from \cite{2014AA...571A..11P} that is built using IRAS and Planck data (labelled as {\tt PlanckGASS})
\textbf{\textit{Fermi}} {\bf Bubbles}: The {\it Fermi} bubbles were discovered in \cite{2010ApJ...724.1044S}
Recently, \cite{TheFermi-LAT:2017vmf,Herold:2019pei} modeled this component dividing the emission in two parts: a low-latitude and a high-latitude component.
The low-latitude component is the one that has the largest impact on the GCE. Therefore, in the model {\tt No low-lat Bubbles} we decide to neglect the presence of this component to study how this affects the results.
{\bf Additional cosmic rays in the Galactic center}: We test the presence of an additional population of cosmic rays injected from the Galactic center considering the following two processes: cosmic-ray protons injected in the central molecular zone and electrons and positrons emitted from the Galactic bulge.
In the former, $\gamma$ rays are produced for bremsstrahlung and $\pi^0$ from protons interacting with the gas in the central molecular zone.
In the latter, photons are generated from electrons and positrons emitted from the Galactic bulge and inverse Compton scattering on the interstellar radiation field in the inner part of the Galaxy.
These additional cosmic-ray electrons and positrons and protons are considered both with the model labeled as {\tt CMZ}.
Instead, in the model named as {\tt IC bulge} only the electrons and positrons produced from the Galactic bulge are added.
For the $\gamma$-ray emission produced from protons, the tracer of the cosmic-ray production in the central molecular zone is taken from the distribution of molecular gas in Equation 18 from \cite{2007AA...467..611F}.
In the model with electrons and positrons the population of MSPs in the Galactic bulge is modeled assuming that their distribution is traced by the old stellar population in the bulge from \cite{2012AA...538A.106R}.
For both these models we tried different vertical sizes $z$ of the propagation halo from 4 and 8 kpc finding for all these values similar results.
In Fig.~\ref{fig:fluxmap} we show the flux map at 1 GeV for the $\gamma$ rays produced for bremsstrahlung and $\pi^0$ from cosmic-ray protons and for inverse Compton scattering from electrons and positrons and assuming $z=8$ kpc.
The signal produced for bremsstrahlung and $\pi^0$ decay from protons follows the distribution of interstellar gas, thus the signal is elongated on the Galactic plane and far from being spherically symmetric.
The $\gamma$-ray emission from electrons and positrons traces the distribution of the interstellar radiation field in the inner part of the Galaxy and the shape of the Galactic bulge. The morphology is not spherically symmetric but rather similar to an ellipsoid with the ratio between the major and minor axis equal to about 1.7, similarly to what considered for this component in \cite{Macias:2016nev}.
\begin{figure*}[t]
\includegraphics[width=0.49\textwidth]{figures/fluxmap_galprop_pi0brem.pdf}
\includegraphics[width=0.49\textwidth]{figures/fluxmap_galprop_ICS.pdf}
\caption{Map of the flux calculated at 1 GeV for the $\gamma$-ray emission produced for bremsstrahlung and $\pi^0$ from cosmic-ray protons (left panel) and for inverse Compton scattering from electrons and positrons injected from the Galactic center (right panel). The color bar represents the intensity of the emission in the different directions of the ROI in units of 1/MeV/cm$^2$/s/sr. We consider the case with a vertical size of the Galactic halo of $z=8$ kpc. The maps reported here have been produced in Ref.~\cite{TheFermi-LAT:2017vmf}.}
\label{fig:fluxmap}
\end{figure*}
{\bf Analysis setup}: In addition to the choice of different IEMs we also test our analysis for different data selections.
We consider the {\tt SOURCE} and {\tt ULTRACLEANVETO} data and correspondent instrument response functions.
The LAT point spread function (PSF) is a function of the incident photon's energy and inclination angle, and the event class. Using an event-level quantity indicating the quality of the reconstructed direction, the data is divided into quartiles, from the lowest quality quartile (PSF0) to the best quality quartile (PSF3)\footnote{Visit the following page for further details \url{https://fermi.gsfc.nasa.gov/ssc/data/analysis/documentation/Cicerone/Cicerone_Data/LAT_DP.html}}.
We test the selection of data for event type PSF 2 and 3, i.e., with the best reconstructed direction, and {\tt SOURCEVETO} instrument response functions ({\tt PSF23}).
Moreover, we make a test with a smaller ROI of size $30^{\circ}\times 30^{\circ}$ ({\tt ROI $30\times30$}).
Finally, we run the analysis both with and without ({\tt no weights}) the weighted maximum likelihood analysis, as described in Sec.~\ref{sec:analysistec}.
\begin{table*}
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
IEM name & Assumption tested & Model used \\
\hline
{\tt Yusifov} & Source distribution & Pulsar distribution from \cite{2004AA...422..545Y} \\
{\tt SNR} & Source distribution & SNR distribution from \cite{1998ApJ...504..761C} \\
{\tt OBstars} & Source distribution & OB stars distribution from \cite{2000AA...358..521B} \\
{\tt Pulsars} & Source distribution & Pulsar distribution from \cite{Lorimer:2006qs} and divided into HI, H2 \\
\hline
\hline
{\tt ICS combined} & Gas and inverse Compton & unique template for the inverse Compton component \\
{\tt SLext} & Gas and inverse Compton & gas model derived from starlight extinction maps \cite{2010NewA...15..433M} \\
{\tt PlanckGASS} & Gas and inverse Compton & gas model derived from Planck and GASS data \cite{2014AA...571A..11P,2010AA...521A..17K} \\
\hline
\hline
{\tt No low-lat Bubbles} & {\it Fermi} Bubbles & no low-latitude {\it Fermi} Bubbles \\
\hline
\hline
{\tt CMZ} & Additional cosmic rays & protons injected from the CMZ and electrons from the Galactic bulge \\
{\tt IC bulge} & Additional cosmic rays & electrons injected from the Galactic bulge \\
\hline
\hline
{\tt SOURCE} & Analysis setup & {\tt SOURCE} data and instrument response function \\
{\tt ULTRACLEANVETO} & Analysis setup & {\tt ULTRACLEANVETO} data and instrument response function \\
{\tt PSF23} & Analysis setup & {\tt SOURCEvet} data and event type PSF23 \\
{\tt ROI $30\times 30$} & Analysis setup & $30^{\circ}\times 30^{\circ}$ ROI \\
{\tt no weights} & Analysis setup & no weighted likelihood analysis \\
\hline
\hline
\end{tabular}
\caption{This table summarizes the IEM and analysis setup and techniques used in the paper. For each model we report the name we use throughout the paper and the assumption that we test with it. The complete description of each model is reported in Sec.~\ref{sec:analysisiem} .}
\label{tab:iem}
\end{center}
\end{table*}
\subsection{Analysis technique}
\label{sec:analysistec}
Our analysis pipeline is entirely based on {\tt FermiPy}, a Python package that automates analyses with the {\tt Fermitools} \cite{2017ICRC...35..824W}\footnote{See \url{http://fermipy.readthedocs.io/en/latest/}.}.
{\tt FermiPy} includes tools to perform high-level analysis of {\it Fermi}-LAT data.
We employ versions {\tt 18.0.0} of {\tt Fermipy} and {\tt 1.2.3} of the {\tt Fermitools}.
The pipeline we apply is based on the following steps.
We run the {\tt gta.setup()} tool which makes cuts on the data according to the time and energy range, spatial and energy bins considered and the instrument response functions selected.
Then, we apply the {\tt gta.optimize()} tool that fits the model to the data with an iterative strategy.
First, it simultaneously fits the normalizations of the brightest components and sources.
Then, it individually fits the normalizations of all sources that are not included in the first step.
Finally, it fits the shape and normalization parameters of each source simultaneously.
This tool is a fast method to reach a good agreement between model and data.
In the absence of strong degeneracies between the different components, i.e.~high latitudes and selection of data above 1 GeV, it provides a model that is basically identical to the one found with a complete likelihood fit (i.e., with all the parameters free to vary at the same time).
However, since the Galactic center includes many sources and components, many degeneracies are present and we have to perform a complete fit.
This is done by running the {\tt gta.fit()} tool which is a Python wrapper of the pyLikelihood fit method implemented in the {\tt Fermitools}.
This tool returns the best-fit SED parameters and the full covariance matrix.
After this first step, we remove from the model the sources detected with a Test Statistic ($TS$)\footnote{The Test Statistic ($TS$) is defined as twice the difference in maximum log-likelihood between the null hypothesis (i.e., no source present) and the test hypothesis: $TS = 2 ( \log\mathcal{L}_{\rm test} -\log\mathcal{L}_{\rm null} )$~\cite{1996ApJ...461..396M}.} lower than 25. This is indeed the usual cut in $TS$ that is used to include sources in {\it Fermi}-LAT catalogs. A $TS$ of 25 corresponds roughly to a detection at $5\sigma$ significance.
We then re-optimize the likelihood to obtain the final model.
In the last step, we calculate the GCE SED using the {\tt gta.sed} command.
This tool fits the data by varying the flux normalization independently in each energy bin. It assumes a power-law SED in each energy bin with a fixed spectral index of 2.0.
This is a good approximation with small enough bins as we have in our analysis.
During this step of the analysis we leave free in the fit the SED parameters of the GCE, IEM templates and sources detected with a $TS>1000$ and located at a distance smaller than $10^{\circ}$ from the Galactic center.
Therefore, the results obtained with {\tt gta.sed} are independent from the SED model assumed initially for the DM component.
Our results for the GCE SED are thus model independent and can be applied a posteriori to search possible interpretations of the GCE.
In \cite{Dimaurosim} we have tested this pipeline with simulated data of the Galactic center.
We have verified that it recovers properly the injected signal of DM (specifically its spatial morphology and flux) and the correct SED parameters of background sources.
Finally, we have shown, by simulating the data with one model and using the same model in the analysis, the pipeline does not leave significant residuals. Indeed, the $TS$ map is compatible with the $\chi^2/2$ distribution for 1 degree of freedom, as it should be in case of a perfect knowledge of the model (see the right panel of Fig.~1 in \cite{Dimaurosim}).
In order to characterize the spatial morphology of the GCE, we avoid the use of specific spatial templates and instead model the excess as a set of concentric, uniform annuli.
This method consists of including in the model concentric uniform annuli and fitting them to the data simultaneously.
The SED of each annulus is modeled with a power law, i.e., the free parameters for each annulus are the normalization and the power-law index.
Since we will apply this analysis to small energy bins the power-law approximation is appropriate.
The fitting procedure we employ is the same explained above.
After performing the fit, we extract the energy flux ($S$) of each annulus in units of MeV cm$^{-2}$ s$^{-1}$ and we compute the surface brightness of the GCE as $dN/d\Omega$, i.e., by dividing the annulus energy fluxes by their solid angles $d\Omega$.
In particular in \cite{Dimaurosim} we have demonstrated that we are able to recover the standard deviation of an injected 2D Gaussian signal and the value of $\gamma$ for an injected DM signal.
We also use a model-dependent technique to find the GCE spatial morphology by fitting the data with DM templates generated using a NFW density profile (see Eq.~\ref{eq:NFW}) with different values for $\gamma$.
The background components are the same for all the test cases.
These likelihood values, obtained by optimizing with respect to all other parameters, yield the profile likelihood $\rm{Log} (\mathcal{L})(\gamma)$, from which we can obtain estimates of $\gamma$ and its uncertainties with standard maximum likelihood methods.
Finally, by maximizing $\rm{Log} (\mathcal{L})(\gamma)$ we find the best-fit value and errors for $\gamma$.
We have demonstrated in \cite{Dimaurosim} that, in case the GCE is simulated with a DM template, the model independent technique with the annuli and the results of the $\rm{Log} (\mathcal{L})(\gamma)$ method provide very similar results for the $\gamma$ parameter.
The region where the uncertainty of the IEM is the largest is along the Galactic plane ($b\sim0^{\circ}$).
In order to mitigate these uncertainties in our analysis, we apply the weighted likelihood technique that has been recently included in the {\tt Fermitools}.
This technique introduces weights for every pixel of the sky according to the number of counts present in the data.
These weights are then multiplied by the $\rm{Log} (\mathcal{L})$ found in each pixel \cite{Fermi-LAT:2019yla}.
Effectively this procedure penalizes pixels with a very large number of photons in which the systematics for the choice of the IEM are larger\footnote{A technical document explaining the weighted likelihood is available at this link \url{https://fermi.gsfc.nasa.gov/ssc/data/analysis/scitools/weighted_like.pdf}.}.
We use as, in the 4FGL catalog paper, a systematic level of $\epsilon=3\%$. This value is motivated by the study performed in Ref.~\cite{Fermi-LAT:2019yla} with the relative spatial and spectral residuals in the Galactic plane where the diffuse emission is strongest.
We show in Fig.~\ref{fig:weights} the weight maps derived at energies 0.27 GeV and 1.1 GeV in our ROI.
At low energy the weights for $|b|<3^{\circ}$ are very small. This implies that all these pixels constrain much less the fitting procedure than the higher latitude ones where the weights are much larger.
Instead, at 1.1 GeV most of the ROI has weights close to 1 meaning that most of pixels have the same weights.
The effect of the weighted likelihood is thus important below 1 GeV and in the inner few degrees from the Galactic plane where the IEMs differ the most.
\begin{figure*}
\includegraphics[width=0.49\textwidth]{figures/weightmap_027GeV.pdf}
\includegraphics[width=0.49\textwidth]{figures/weightmap_11GeV.pdf}
\caption{Weight maps generated for $E=0.27$ GeV (left panel) and $1.1$ GeV (right panel). The color bar represents the $\rm{Log}_{10}$ of the weight values. See the main text for further details on the weighted likelihood technique.
\label{fig:weights}
\end{figure*}
Finally, we apply the energy dispersion to all the components of our model using the method implemented in the {\tt Fermitools}\footnote{For a complete description see \url{https://fermi.gsfc.nasa.gov/ssc/data/analysis/documentation/Pass8_edisp_usage.html}.}.
Specifically, we select {\tt apply\_edisp=true} and {\tt edisp\_bins=-1} which applies the energy dispersion only on the spectrum accounting for one extra bin.
\section{Results}
\subsection{Interstellar emission components}
\label{sec:refinement}
Previous papers on the GCE have made different choices for the number of IEM components to leave free in the analysis.
For example in \cite{Goodenough:2009gk,Hooper:2010mq,Hooper:2011ti,Daylan:2014rsa} the IEM was modeled with a single template, and Ref.~\cite{Daylan:2014rsa} used the IEM released with Pass 6 data.
Instead Refs.~\cite{Calore:2014xka,TheFermi-LAT:2015kwa,TheFermi-LAT:2017vmf} divided the IEM in multiple templates generated with {\tt Galprop}.
Specifically, Ref.~\cite{Calore:2014xka} uses 60 different IEMs, each divided into the components we reported in sec.~\ref{sec:analysisiem}.
On the other hand Refs.~\cite{TheFermi-LAT:2015kwa,TheFermi-LAT:2017vmf} consider the inverse Compton, bremsstrahlung and $\pi^0$ components divided into rings.
\begin{figure*}
\includegraphics[width=0.49\textwidth]{figures/Resid_30x30.pdf}
\includegraphics[width=0.49\textwidth]{figures/E2dNdE_trycompt_1-10.pdf}
\caption{Left Panel: value of the $\rm{Log}(\mathcal{L})$ and sum of the absolute value of the residuals obtained when we run the analysis pipeline to {\tt Model i} with $i=[1,11]$ (see Tab.~\ref{tab:models} for further details.).
Right Panel: SED of the GCE obtained with the different {\tt Models}. In parenthesis of the legend we report the number of components present for each {\tt Model}.
\label{fig:trycomp}
\end{figure*}
\begin{figure}
\includegraphics[width=0.49\textwidth]{figures/E2dNdE_components.pdf}
\caption{Energy spectrum between 1 and 10 GeV of the different components included in our analysis with {\tt Model 2}. Here, we use the components generated with the {\tt Baseline} IEM (see Sec.~\ref{sec:analysisiem} for further details).
\label{fig:spectrumcomp}
\end{figure}
\begin{figure*}[t]
\includegraphics[width=0.49\textwidth]{figures/residmap_sigma3_GCEmodel2_optimized.pdf}
\includegraphics[width=0.49\textwidth]{figures/residmap_sigma3_GCEmodel2.pdf}
\caption{We show in the left panel the map of the residuals obtained using the {\tt Model 2} and the {\tt Baseline} IEM to fit the ROI in the energy range 1-10 GeV. Instead, in the right panel we have added to the residuals the counts of the GCE. The different colors represent the fractional residual, i.e., the residual counts (data minus model) divided by the total counts.
\label{fig:residmap}
\end{figure*}
In order to define the appropriate number of IEM components, we apply the following analysis.
We consider different cases labeled as {\tt Model i} with {\tt i} ranging from 1 to 11.
Models with increasing {\tt i} include a decreasing number of IEM components.
The DM template is always separated from all the other background sources.
We consider for all the {\tt Models} the same list of point and extended sources which has been found by using the analysis pipeline described in Sec.~\ref{sec:analysistec}.
In particular we use the templates generated with the case labeled as {\tt Baseline} in Sec.~\ref{sec:analysisiem} and we assume different choices for the number of components left free in the fit.
In Tab.~\ref{tab:models}, we report a summary of the templates that are left free for each {\tt Model}.
In the next sections we will fix the number and type of components left free in the fit and will test the different IEMs and analysis setup listed in Sec.~\ref{sec:analysisiem}.
\begin{table*}
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
{\tt Model} & N$^{\circ}$ comp. & list templates used for each component \\
{\tt 1} & 17 & 5 rings for bremsstrahlung (Brem.) and $\pi^{0}$ emission, 3 for inverse Compton (CMB, SL. IR),\\
& & 2 for low latitude bubbles, isotropic (ISO), 1 template for Loop I, the Sun and the Moon (LoopMoonSun). \\
{\tt 2} & 9 & 1 Brem., 1 $\pi^{0}$, 3 for inverse Compton (IC), 2 for low latitude bubbles, 1 ISO, 1 LoopMoonSun. \\
{\tt 3} & 8 & 1 Brem. and $\pi^{0}$, 3 IC, 2 for low latitude bubbles, 1 ISO, 1 LoopMoonSun. \\
{\tt 4} & 8 & 1 Brem., 1 for $\pi^{0}$, 3 IC, 1 bubbles, 1 ISO, 1 LoopMoonSun. \\
{\tt 5} & 8 & 2 for Brem. and $\pi^{0}$ divided into H1 and H2, 3 IC, 1 bubbles, 1 ISO, 1 LoopMoonSun. \\
{\tt 6} & 7 & 1 Brem., 1 for $\pi^{0}$, 1 for IC, 2 for bubbles, 1 ISO, 1 LoopMoonSun. \\
{\tt 7} & 7 & 1 Brem. and $\pi^{0}$, 3 IC, 1 bubbles, 1 ISO, 1 LoopMoonSun. \\
{\tt 8} & 5 & 1 Brem. and $\pi^{0}$, 1 IC, 1 bubbles, 1 ISO, 1 LoopMoonSun. \\
{\tt 9} & 4 & 1 Brem. and $\pi^{0}$, 1 IC, 1 bubbles, 1 ISO and LoopMoonSun. \\
{\tt 10} & 4 & 1 Brem. and $\pi^{0}$ and IC, 1 bubbles, 1 ISO, 1 LoopMoonSun. \\
{\tt 11} & 1 & 1 unique template for all components. \\
\hline
\hline
\end{tabular}
\caption{This table summarizes the templates that we leave free for each {\tt Model}.}
\label{tab:models}
\end{center}
\end{table*}
In this analysis, we select data between 1 and 10 GeV since at these energies the GCE is brighter and more significant.
We apply the analysis pipeline explained in Sec.~\ref{sec:analysistec} to all the {\tt Models} reported before.
We then save the value of the log-likelihood ($\rm{Log}(\mathcal{L})$) obtained in the end of the analysis and we generate a residual map to check how well each model fits the data.
In the left panel of Fig.~\ref{fig:trycomp} we show the value of $\rm{Log}(\mathcal{L})$ obtained for each {\tt Model}.
Moreover, we calculate the sum of the absolute values of the residuals in all the pixels of the ROI. This is shown also shown in Fig.~\ref{fig:trycomp}.
The figure demonstrates that we obtain similar $\rm{Log}(\mathcal{L})$ values and sum of the residuals if we use {\tt Model $i$} with $i<6$.
Instead, for models with a smaller number of components ($i>6$) the $\rm{Log}(\mathcal{L})$ decreases and the residuals increase significantly (meaning that the fit becomes poorer).
We also calculate the GCE SED obtained with the different {\tt Models} and we show it in the right panel of Fig.~\ref{fig:trycomp}.
The SED obtained is similar at energies $1-2$ GeV while at higher energies {\tt Model} 6 to 11 seem to give slightly steeper SED compared to all the other cases.
The difference in the SED obtained for {\tt Model} 1 to 6 is at most 5\% which is much smaller than the differences obtained when using different IEMs as we will see in Sec.~\ref{sec:spectrum}. Therefore, the systematics we find due to the choice of the number of components to leave free in the fit (assuming {\tt Model i} with $i<6$) is negligible with respect to the systematics due to the choice of the specific IEM physics.
We decide to use in our analysis {\tt Model 2} that provides, as well as the others with $i<6$, a good fit to the data, small residuals, has a number of components much smaller than {\tt Model 1} and about the same number of components for the ones with $2<i<6$. {\tt Model 2} is divided into the following parts: two separate templates for the bremsstrahlung and $\pi^{0}$ emission, three templates for the inverse Compton emission, one for each interstellar radiation field (starlight, infrared and CMB), two components for the {\it Fermi} bubbles (low and high-latitude as in \cite{TheFermi-LAT:2017vmf}), the isotropic component and one unique template for the emission from the Moon, Sun and Loop I.
In Fig.~\ref{fig:spectrumcomp} we show the flux of the different components as a function of energy between 1 and 10 GeV.
The component with the largest flux is due to the bremsstrahlung and $\pi^{0}$ processes which contributes roughly 60\% of the observed flux.
The inverse Compton is at about 30\% of the total flux. The isotropic emission and the cumulative flux of all sources contribute together at most about 10\%.
The GCE flux modeled with a DM template with $\gamma=1.2$ is at about a factor of 1\% of the total flux.
The {\it Fermi} bubbles also have a similar flux.
In Fig.~\ref{fig:residmap} we show the fractional residuals, i.e., the residuals divided by the total number of counts, and the residuals plus the GCE counts obtained with {\tt Model 2}
The residuals are of the order of $20-25\%$ of the counts, similarly to that found in \cite{Calore:2014xka,TheFermi-LAT:2015kwa,TheFermi-LAT:2017vmf}.
\begin{figure}
\includegraphics[width=0.49\textwidth]{figures/GPsurvey_fluxexcess_summary.pdf}
\caption{Flux, integrated between $1-10$ GeV, absorbed by the DM template when it is placed at different positions along the Galactic plane. We show the results obtained with three different IEMs.
\label{fig:GPsurvey}
\end{figure}
The presence of the GCE is robust over different data selection analysis techniques, IEMs and point source catalogs used to analyze the data (see, e.g., the results in this paper and \cite{Calore:2014xka,TheFermi-LAT:2017vmf}).
It is thus interesting to investigate if similar excesses are present in other directions towards the Galactic plane.
In order to test this we use a DM template generated with $\gamma=1.2$
and we select different ROIs of size $40^{\circ}\times40^{\circ}$ and centered at different directions on the Galactic plane from $l=0^{\circ}$ to $360^{\circ}$ with a step size of $2^{\circ}$.
This gives 180 ROIs for which we apply the same analysis explained in Sec.~\ref{sec:analysis}.
After performing a fit to the ROI we save the flux of the DM template integrated between 1 and 10 GeV.
We show in Fig.~\ref{fig:GPsurvey} the flux absorbed by the DM template located at the different longitudes. We display the results obtained with the {\tt Baseline}, {\tt SLext} and {\tt SNR} IEM, but we find similar outcomes with the other models.
The next highest flux is detected at $l=\pm20^{\circ}$, with a decreased amplitude by a factor of $\sim 2$.
These two excesses have been already found in \cite{Calore:2014xka} by performing a similar analysis, and they are probably associated to the $\gamma$-ray emission from molecular clouds. Our results are very similar to results found in Ref.~\cite{Calore:2014xka}.
\subsection{Energy spectrum}
\label{sec:spectrum}
\begin{figure}
\includegraphics[width=0.49\textwidth]{figures/E2dNdE_GCE_gasics_ebins.pdf}
\includegraphics[width=0.49\textwidth]{figures/E2dNdE_GCE_sources_ebins.pdf}
\includegraphics[width=0.49\textwidth]{figures/E2dNdE_GCE_seldata_ebins.pdf}
\caption{GCE SED measured when using the {\tt Baseline}, {\tt Yusifov}, {\tt SNR}, {\tt OBstars}, {\tt Pulsars}, {\tt SLext}, {\tt ICS combined} and {\tt PlanckGASS} IEMs. We also show the results obtained with different selections of the data ({\tt SOURCE}, {\tt ULTRACLEANVETO} and {\tt PSF23}), with a smaller ROI ({\tt $30\times30$ ROI}) and without using the weight map in the maximum likelihood analysis ({\tt no weights}).
\label{fig:spectrumcases}
\end{figure}
\begin{figure*}
\includegraphics[width=0.60\textwidth]{figures/E2dNdE_GCE_band_ebins_comparison.pdf}
\caption{Comparison between the results for the GCE SED obtained in our analysis and in \cite{Calore:2014xka,TheFermi-LAT:2017vmf}. The bands represent the variation of the GCE SED obtained by using all the IEMs and analysis techniques shown in Fig.~\ref{fig:spectrumcases} and the results found in \cite{Calore:2014xka,TheFermi-LAT:2017vmf} when using different IEMs. See the text for further details on the conversion of the GCE SED found in our analysis and in \cite{Calore:2014xka,TheFermi-LAT:2017vmf} into flux per solid angle (i.e., in units of MeV/cm$^2$/s/sr). We also display the best-fit to the GCE SED, obtained with the {\tt Baseline} IEM, by using a log-parabola function.
\label{fig:spectrumconv}
\end{figure*}
\begin{figure}
\includegraphics[width=0.49\textwidth]{figures/E2dNdE_GCE_addcomp_ebins.pdf}
\caption{GCE SED obtained in case we use the {\tt Baseline}, {\tt CMZ 4kpc}, {\tt CMZ 8kpc}, {\tt IC bulge} and {\tt no low-lat bubbles} IEMs.
\label{fig:spectrumaddcomp}
\end{figure}
The first characteristic of the GCE that we investigate is the energy spectrum.
In order to calculate the GCE SED, we consider the energy range $0.1-1000$ GeV and we apply the analysis pipeline presented in Sec.~\ref{sec:analysistec}.
The GCE is detected with a maximum $TS$ of $1000-3000$, depending on the IEM, in the energy range between $0.8-3$ GeV.
These $TS$ values correspond to a significance of the GCE of about $40\sigma$, i.e., larger than that reported in \cite{Daylan:2014rsa}, where they find $30\sigma$ using the interstellar emission modeled with a unique template.
We find a larger significance for the GCE because we select a larger data sample (11 versus 5.5 years) and more refined models for the background components with respect to \cite{Daylan:2014rsa}.
At lower and higher energies the GCE $TS$ drops quickly, and for energies smaller than 0.3 GeV and larger than $30-50$ GeV the value goes below 25 for most of the IEMs.
We show in Fig.~\ref{fig:spectrumcases} the GCE SED found for the following IEMs: {\tt Baseline}, {\tt Yusifov}, {\tt SNR}, {\tt OBstars}, {\tt Pulsars}, {\tt SLext}, {\tt ICS combined} and {\tt PlanckGASS}.
We also test different selections of the data ({\tt SOURCE}, {\tt ULTRACLEANVETO}, {\tt PSF23}), a smaller ROI ({\tt ROI $30\times30$}) and the exclusion of the weight maps in the maximum likelihood analysis ({\tt no weights}).
All the results obtained with these IEMs, data selections, and techniques share the same general behavior. The GCE has a bumpy SED, with a peak at around 1-3 GeV, and a low and high-energy tail similarly to what found in Refs.~\cite{Abazajian:2012pn,Daylan:2014rsa,Calore:2014xka,TheFermi-LAT:2015kwa,TheFermi-LAT:2017vmf}.
Most of the data points below 500 MeV are upper limits.
This is due both to the falling SED of the GCE below 1 GeV and to the weighted likelihood technique, which de-emphasizes the statistical weight of low-energy photons. This paper represents the first use of this technique to account properly for the systematic differences resulting from differing IEMs, which for this region and at these energies is by far the dominant component (see Fig.~\ref{fig:spectrumaddcomp}).
In Fig.~\ref{fig:spectrumconv} we show the envelope of the GCE SEDs for all of the above IEMs compared to the results found in \cite{Calore:2014xka,TheFermi-LAT:2017vmf}.
The normalization of the GCE SED changes by roughly 60\% when using different IEMs, data selections and analysis techniques.
While these results are similar to those obtained in Refs.~\cite{2015PhRvD..91l3010Z,Calore:2014xka}, they are more robust due to using twice as much data, a newer catalog with almost twice as many sources, and IEMs designed specifically for the Galactic center region.
The SED is shown in this figure in units of MeV/cm$^2$/s/sr, i.e., we have divided the SED by the ROI solid angle, since the selection of the ROI is different in our analysis with respect to \cite{Calore:2014xka,TheFermi-LAT:2017vmf}. In particular, Ref.~\cite{Calore:2014xka} selected a region given by $2^{\circ}<|b|<20^{\circ}$ and $|l|<20^{\circ}$, while Ref.~\cite{TheFermi-LAT:2017vmf} used an ROI defined with an angular distance $<10^{\circ}$ from the Galactic center and a mask of bright sources (which implies an effective mask of $2^{\circ}$ from the Galactic center). Considering the systematic band due to the choice of the IEM, our results are compatible with the ones reported in \cite{Calore:2014xka,TheFermi-LAT:2017vmf}.
We fit different analytical functions to the GCE SED. In particular we test a power-law, a power-law with an exponential cutoff (PLE) and a log-parabola (LP)\footnote{\url{https://fermi.gsfc.nasa.gov/ssc/data/analysis/scitools/source_models.html#spectralModels}.}.
We apply the fits to the SED measured with the {\tt Baseline} IEM. Similar results are found by fitting the SED obtained with the other models.
We find a much better match with the data using the log-parabola shape and we find best-fit values for the spectral index of $-2.0$ and curvature index of 0.27.
In particular the $TS$ calculated as $2(\rm{Log}{\mathcal{L}_{\rm{LP}}-\rm{Log}\mathcal{L}_{\rm{PLE}}})$, where $\rm{Log}{\mathcal{L}_{\rm{LP}}}$ ($\rm{Log}{\mathcal{L}_{\rm{PLE}}}$) is the fit obtained by using the LP (PLE) SED, is 380.
Ref.~\cite{Abazajian:2012pn} also found a preference for the fit with a LP with a roughly similar value for the spectral and curvature indexes.
We display in Fig.~\ref{fig:spectrumconv} the comparison between the data and the LP best-fit.
A LP shape is able to reproduce the GCE spectrum between $0.1-10$ GeV but is not able to properly capture the high-energy tail. However, this tail is not significant for all the IEMs (see Fig.~\ref{fig:spectrumconv} where the convolution of the results obtained for all the IEMs is displayed).
Possible interpretations of the GCE are associated with the $\gamma$-ray emission from cosmic-ray protons and/or electrons and positrons injected from the Galactic center.
We test these possibilities using the {\tt CMZ 4kpc}, {\tt CMZ 8kpc}, and {\tt IC bulge} models.
We also run the analysis for the case without the presence of the low-latitude bubbles component ({\tt no low-lat bubbles}).
The results obtained with these models are presented in Fig.~\ref{fig:spectrumaddcomp}.
The case {\tt no low-lat bubbles}, as expected, provides a $20-30\%$ larger SED because the GCE absorbs part of the low-latitude bubble emission. This model fits much worse the ROI giving a $\rm{Log}{\mathcal{L}}$ lower by 2100 with respect to the {\tt Baseline} model.
In the cases {\tt CMZ 4kpc}, {\tt CMZ 8kpc}, {\tt IC bulge} we measure a smaller GCE flux since the additional cosmic-ray components take part, but not all, the GCE emission.
The GCE SED changes significantly with these latter models, but an excess peaked at a few GeV still remains with a high significance.
In particular, the model for which the GCE spectrum decreases the most, roughly by a factor of two, is the {\tt IC bulge} case.
This model represents the possible flux of a population of pulsars located around the bulge of our Galaxy.
This result demonstrates the viability of the millisecond pulsar interpretation for the GCE \cite{Bartels:2015aea,Lee:2015fea}.
Since the cases {\tt CMZ 4kpc}, {\tt CMZ 8kpc}, {\tt IC bulge} absorb a significant fraction of the GCE and they have been considered in the past as possible interpretations to the GCE (see, e.g., \cite{Carlson:2014cwa,Petrovic:2014uda,Gaggero:2015nsa}), we have tested these model without including the DM template. Therefore, we try to fully explain the GCE with the $\gamma$-ray emission produced for inverse Compton scattering, bremsstrahlung or $\pi^0$ decays, by cosmic rays injected from the Galactic center during recent outbursts.
The fits provide differences of likelihood with respect to the {\tt Baseline} that are $+420$ and $+540$ for the models {\tt CMZ 4kpc}, {\tt CMZ 8kpc}, (the fit improves) and $-230$ for the model {\tt IC bulge} (the fit worsens).
The models labeled as {\tt CMZ 4kpc} and {\tt CMZ 8kpc} consider two more components with respect to {\tt Baseline} because they substitute the DM component with three templates associated to the $\gamma$-ray emission for inverse Compton, Bremsstrahlung and $\pi^0$ decay.
Therefore, it is difficult to compare the improved values of the $\rm{Log}(\mathcal{L})$ obtained with the {\tt CMZ 4kpc} and {\tt CMZ 8kpc} with respect to the {\tt Baseline} model that uses the DM template.
Moreover, the additional components present in the {\tt CMZ 4kpc} and {\tt CMZ 8kpc}, associated with the Bremsstrahlung and $\pi^0$ decay emissions, follow the distribution of interstellar gas that is not spherically symmetric around the Galactic center but rather elongated on the Galactic plane.
However, as we will see in the next sections, the GCE is spherically symmetric.
So we think the improvement in the $\rm{Log}(\mathcal{L})$ with these models is due to fitting better the emission from the Galactic plane rather than accounting better for the GCE emission.
We tested different scenario for the {\tt IC bulge} model by assuming different sizes for the diffusive halo with the vertical size $z$ varied between 4 to 10 kpc and the radius $R$ from 10 to 20 kpc. However, the result is very similar for all the tested cases: the {\tt IC bulge} model performs always worse than the {\tt Baseline} model with the DM template with a difference of $\rm{Log}(\mathcal{L})$ between -250 and -230.
The two main reasons for this result are the following. First, $\gamma$ rays produced from electrons and positrons injected from the Galactic bulge (this is the component that in the {\tt IC bulge} model absorbs the GCE) has a much flatter dependence with the angular distance from the Galactic center in the inner few degrees with respect to the GCE (see Sec.~\ref{sec:spacial} and Fig.~\ref{fig:SB}). Second, the inverse Compton emission from the Galactic bulge does not have a spherically symmetric morphology as we measure for the GCE (see Fig.~\ref{fig:fluxmap} and also the model used in \cite{Macias:2016nev}).
To conclude the {\tt IC bulge} model can fit reasonably well the GCE but with a worse overall fit compared to a DM profile generated with a NFW density with $\gamma\sim 1.25$.
\subsection{Spatial morphology}
\label{sec:spacial}
\begin{figure}
\includegraphics[width=0.49\textwidth]{figures/SB_DM_morebins_1-10GeV_Pulsars_realdata_Horiuchi.pdf}
\caption{Surface brightness data ($dN/d\Omega$) obtained with the analysis in annuli of size $1^{\circ}$ and with the {\tt Baseline} IEM in the energy range $1-10$ GeV. We also show the best fit obtained with a NFW density profile for $\gamma=1.27$ (dashed blue line) and for the {\tt IC bulge} model at $E=0.7$ (dotted green line) and 20 GeV (dot-dashed green line) for which we fit the GCE data with the $\gamma$ rays emitted for inverse Compton scattering by cosmic-ray electrons injected from the Galactic bulge (see Sec.~\ref{sec:analysisiem} for a complete description of this model). Red data points show the results obtained in \cite{Daylan:2014rsa} at 2.67 GeV.
\label{fig:SB}
\end{figure}
\begin{figure}
\includegraphics[width=0.49\textwidth]{figures/SBsummary_DM_simdata_moreEbins_1p0deg_gasics_sizeEnergy.pdf}
\includegraphics[width=0.49\textwidth]{figures/SBsummary_DM_simdata_moreEbins_1p0deg_sources_sizeEnergy.pdf}
\includegraphics[width=0.49\textwidth]{figures/SBsummary_DM_simdata_moreEbins_1p0deg_seldata_sizeEnergy.pdf}
\caption{Best fit for $\gamma$ obtained by fitting the surface brightness data obtained in the following energy bins: $0.6-1.0$, $1.0-1.8$, $1.8-3.2$, $3.2-5.6$, $5.6-10$, $10-32$ GeV. We show the results found using following IEMs: {\tt Baseline}, {\tt Yusifov}, {\tt SNR}, {\tt OBstars}, {\tt Pulsars}, {\tt SLext}, {\tt ICS combined} and {\tt PlanckGASS} IEM. We also display the values of $\gamma$ we obtained with different selections of the data ({\tt SOURCE}, {\tt ULTRACLEANVETO} and {\tt PSF23}), with a smaller ROI ({\tt $30\times30$ ROI}) and without using the weight map in the maximum likelihood analysis ({\tt no weights}).
\label{fig:gammafit}
\end{figure}
In this paper we apply several methods to study the spatial morphology of the GCE.
In this section we employ a model dependent and a model independent technique.
In the model independent technique we substitute the DM template with concentric and uniform annuli.
Then, we fit the annuli to the data, using the pipeline presented in Sec.~\ref{sec:analysistec}, and we extract the energy flux of each annulus.
Finally, we divide the annulus energy fluxes for the solid angles and we obtain the surface brightness ($dN/d\Omega)$) of the GCE.
As demonstrated in \cite{Dimaurosim} with simulated data, the optimal annulus size is between $0.75^{\circ}-1.5^{\circ}$.
Indeed, angular widths of these orders are similar to the angular resolution of the LAT for $E\sim1$ GeV\footnote{\url{https://www.slac.stanford.edu/exp/glast/groups/canda/lat_Performance.htm}} and they are small enough to capture the right spatial distribution of the GCE.
We show here the results for an annulus size of $1^{\circ}$, but our conclusions do not change by using $0.75^{\circ}$ or $1.5^{\circ}$.
In Fig.~\ref{fig:SB} we show the surface brightness data obtained with an analysis in the energy range $1-10$ GeV using the {\tt Baseline} IEM.
The surface brightness data are very precise in the inner $10^{\circ}$ where the annuli are detected with at least $10\sigma$ significance and the precision of the data is between $2-10\%$.
The GCE extends with a significant flux roughly up to $12^{\circ}$. This demonstrates that our choice of an ROI with a size of $40^{\circ}\times40^{\circ}$ is appropriate.
We can fit well these data with a NFW DM profile with $\gamma=1.27$ (see Eq.~\ref{eq:NFW}).
Refs.~\cite{Abazajian:2012pn,Daylan:2014rsa,Calore:2014xka,TheFermi-LAT:2017vmf} found similar best-fit values for $\gamma$. However, most of those references only provide the value of $\gamma$ and not the data for the flux of the GCE as a function of angular distance from the Galactic center as we do.
The main new result of this paper is that we provide the spatial distribution of the GCE for a wide region and with a method that does not depend on the specific DM model. We report the surface brightness data that can be used to find which astrophysical interpretation is more suitable to explain the GCE spatial distribution. Refs.~\cite{Daylan:2014rsa,Horiuchi:2016zwu} are the only two publications that published results in a similar way as ours. However, Ref.~\cite{Daylan:2014rsa} provided the results for a limited region between $2.5^{\circ}-10^{\circ}$, with data that are not as precise as ours and tested only one IEM that was not designed for the Galactic center region. We compare our results with those obtained in \cite{Daylan:2014rsa} at 2.67 GeV in Fig.~\ref{fig:SB}. The surface brightness data are compatible between $2^{\circ}-7^{\circ}$, while smaller angular distances are not considered by Ref.~\cite{Daylan:2014rsa} and at larger angles their surface brightness deviates significantly both from our result and from the DM template predictions.
Instead, the result in Ref.~\cite{Horiuchi:2016zwu} have been derived using almost one half of the data and the older Pass 7 data selection. The authors have considered three IEMs taken from \cite{Calore:2014xka}. The surface brightness they obtain for the energy range $E\in[1.9,10]$ GeV is shown in Fig.~\ref{fig:SB}. We have rescaled their measurements to the energy range 1-10 GeV to be comparable with ours\footnote{We have calculated the LogParabola best fit to the GCE SED (see Fig.~\ref{fig:spectrumcases}) and integrated it between $E\in[1.9,10]$ GeV and $E\in[1.0,10]$ GeV. Then, we taken the ratio between the two integrals to rescale the results in Ref.~\cite{Horiuchi:2016zwu}, that are given for $E\in[1.9,10]$ GeV, to our energy range.}. The GCE spatial distribution found in Ref.~\cite{Horiuchi:2016zwu} is significantly different from ours in the inner few degrees from the Galactic center. In fact it is compatible with a NFW profile with $\gamma\sim 1.0$, i.e.~smaller than most of the $\gamma$ values found in several other references.
We also test the model {\tt IC bulge} for which we eliminate the DM template and we fit the surface brightness data with $\gamma$ rays emitted for inverse Compton scattering by cosmic-ray electrons injected from the Galactic bulge (see Sec.~\ref{sec:analysisiem} for a complete description of this model). We perform the fit using the model at $E=0.7$ and 20 GeV to demonstrate that this emission mechanism would produce a signal with a spatial morphology that evolves with energy. In particular at 0.7 GeV the signal is flatter and it decreases by a factor of about 30 between $0^{\circ}$ and $10^{\circ}$ instead at 20 GeV changes by roughly 100 times.
As we will see next in this section, the spatial morphology of the GCE does not change significantly between 0.6 and 30 GeV so the variation of the Galactic bulge emission as a function of energy is problematic for the interpretation of the GCE with this mechanism.
We now explore a novel analysis, never made before, and derive the surface brightness data in different energy bins between $0.1-1000$ GeV to investigate if there is an energy dependence of the spatial morphology of the GCE. We do so by performing the same analysis presented before in the following energy bins : $0.1-0.3$, $0.3-0.6$, $0.6-1.0$, $1.0-1.8$, $1.8-3.2$, $3.2-5.6$, $5.6-10$, $10-32$, $32-100$, $100-1000$ GeV.
Exactly as for the energy range $1-10$ GeV, we find a good fit to the data with a DM profile for all the energies between $0.6-30$ GeV for which we detect significantly the GCE.
Instead, at energies lower than 0.6 GeV and higher than 30 GeV the GCE is not significant and so the analysis provides surface brightness data with upper limits for most of the annuli.
Therefore, the fit to the $dN/d\Omega$ at these energies is prohibitive.
For the analyses at energies $0.6-30$ GeV, we decide to study the possible energy dependence of the spatial morphology by fitting with a DM template the surface brightness data found into different energy bins. Indeed, we find that the surface brightness data are well fit with an NFW profile with $\gamma$ free to vary (similarly to the fit depicted in Fig.~\ref{fig:SB} for energies between $1-10$ GeV).
In Fig.~\ref{fig:gammafit} we report the best-fit $\gamma$ values obtained for different IEMs and selections of the data and ROI size.
We do not find an energy evolution of $\gamma$ and its value at energies between $1-10$ GeV ranges between $1.1-1.2$.
This scatter is larger than the error on the single measurements which is of about a few \%.
This implies that the systematics on the choice of the IEM or data selection and analysis technique dominates the uncertainty on the value of $\gamma$.
On the other hand, the scatter in the average values obtained for each case considered is between $1.1-1.3$.
The global average value calculated considering all the IEM and analysis setup is 1.25.
We can interpret the differences for the $\gamma$ values we obtain for the different cases as the systematic uncertainty for the GCE spatial distribution.
In \cite{Dimaurosim} we have studied this extensively with simulations, finding that the systematic uncertainties on $\gamma$ are on the order of $5\%$ from the injected value.
Using the real data we find roughly $10\%$ (obtained as $100\cdot 0.5\cdot(1.08-1.32)/1.25$), thus the systematics are twice larger than the one obtained in \cite{Dimaurosim}.
This difference is probably due to the fact that it is difficult to bracket with simulations all the possible systematic effects present in the real data.
We also use a model dependent technique to find the spatial morphology of the GCE that consists of fitting the GCE with DM templates generated with different values of $\gamma$. For each run with a different $\gamma$ value the background components (i.e., sources and IEM components) are the same so we are able to produce a profile of $\rm{Log}(\mathcal{L})$ as a function of $\gamma$.
By maximizing the function $\rm{Log}(\mathcal{L})(\gamma)$ we then find best-fit value for $\gamma$.
We apply this analysis to all the cases presented in Fig.~\ref{fig:gammafit}. The values of $\gamma$ are consistent with the ones derived with the fit to the surface brightness data (see Fig.~\ref{fig:gammafit}).
To conclude, we do not find any significant evolution in the value of $\gamma$, i.e., in the GCE spatial morphology.
There is a scatter of the average values between $1.2-1.3$ that is due to different results obtained for the models and analysis setup considered in the analysis.
A change in the $\gamma$ value due to an evolution of the GCE spatial morphology is not excluded for some models, as for the {\tt Baseline} for which $\gamma$ increases from 0.6 GeV to 30 GeV from 1.15 to 1.4. However, this evolution is not present for the other models tested.
Considering all the cases reported in Fig.~\ref{fig:gammafit}, the change of $\gamma$ is constrained to be at most $\sim 10\%$ from the average value.
The scatter on the value of $\gamma$ and a maximum evolution of $10\%$ in its value is smaller than the evolution expected if the GCE is due to electrons and positrons injected from the Galactic bulge (this component is included in the {\tt IC bulge} model). As we have demonstrated in \cite{Dimaurosim}, with this mechanism the GCE spatial morphology would significantly evolve with energy and $\gamma$ would increase by about a factor of $30\%$ from 0.6 to 30 GeV.
Moreover, as we have reported in Fig.~\ref{fig:SB}, it is challenging with an electron and positron burst signal to produce a surface brightness compatible with the GCE data in the inner few degrees from the Galactic center.
Our results for the value of $\gamma$ are compatible with the ones published in \cite{Calore:2014xka,Daylan:2014rsa}, while in Ref.~\cite{TheFermi-LAT:2017vmf} they find slightly smaller values of the order of $1.0-1.1$.
\subsection{Analysis in quadrants}
\label{sec:quadrants}
\begin{figure}
\includegraphics[width=0.49\textwidth]{figures/E2dNdE_realdata_AltIEM_Pulsars_quadrant_1-10.pdf}
\includegraphics[width=0.49\textwidth]{figures/E2dNdE_realdata_AltIEM_Pulsars_quadranto_1-10.pdf}
\caption{SED of the quadrants for the {\tt quad$+$} configuration in the top panel and for {\tt quadx} in the bottom panel. We also display the GCE SED, obtained with the DM template considered as a whole, divided by 4. The data have been obtained with the {\tt Baseline} IEM.
\label{fig:quadrantsSED}
\end{figure}
\begin{figure}
\includegraphics[width=0.49\textwidth]{figures/SBsummary_GC_realdata_1p0deg_quadrants_all.pdf}
\caption{Best-fit for $\gamma$ found by fitting the surface brightness of each quadrant and using different IEMs. We show here the results obtained with the quadrant configuration {\tt quad$+$}.
\label{fig:quadrantsSB}
\end{figure}
We apply another method to study the spatial morphology of the GCE.
We divide the DM template into quadrants and we leave free to vary in the fit their SED parameters.
We focus this analysis in the energy range between $1-10$ GeV since we have not found any energy dependence of the value of $\gamma$, and in this energy range the GCE is detected with the largest significance.
We use a LP SED for each quadrant, thus the GCE is fit in this case with 12 SED parameters.
We test two different orientations of the quadrants.
The first, labeled as {\tt quad$+$}, has quadrants separated by the Galactic plane direction and by $b=\pm90^{\circ}$. The first quadrant is located at positive longitudes and latitudes, and the other quadrants move from the first in a counterclockwise direction.
The second configuration, labeled as {\tt quadx} is rotated by $+45^{\circ}$ with respect to the former, with the first quadrant located at positive longitudes and latitudes in the range $-45^{\circ}<b<45^{\circ}$. The second, third and fourth quadrants are found by moving from the first in a counterclockwise direction. Therefore, in the {\tt quad$+$} configuration all the quadrants share a portion of the Galactic plane while in the {\tt quadx}, only the first and third quadrants include the Galactic plane.
We use both these quadrant configurations to investigate a possible change in the results due to the choice of the IEM.
In Fig.~\ref{fig:quadrantsSED} we show the SED we find for each quadrant compared to the SED obtained when we use a unique DM template for the GCE and divide the flux value by four.
We report the result for the {\tt Baseline} IEM, but similar conclusions are found when taking the other models and different data selections.
All the quadrants share a similar SED and the difference in normalization is roughly $30\%$.
The data are well fit with a LP with values for the spectral index $\alpha$ and curvature index $\beta$ compatible with the values found for the GCE when fitting it with a unique DM template (i.e., $\alpha=2.0$ and $\beta=0.27$).
There is not a relevant difference in the results considering the {\tt quad$+$} or {\tt quadx} configuration.
Finally, the SED of DM considered as a whole and divided by four is compatible with the quadrant SED considering the scatter.
In Ref.~\cite{Dimaurosim} we have performed a similar analysis with simulated data finding that the scatter is roughly $15\%$.
Therefore, also in this case the scatter that we find in the real data is larger, roughly by a factor of 2, compared to simulations .
In Ref.~\cite{TheFermi-LAT:2017vmf} the authors performed a similar analysis finding that, when accounting for the asymmetric emission from the {\it Fermi} bubbles, the SED of the four quadrants are similar to each other. Our result is thus similar to theirs.
Ref.~\cite{Calore:2014xka} inspected the GCE SED by considering ten different portions of the region around the Galactic center and they found similar spectra in all regions.
Therefore, we confirm the results found by previous publications: the GCE has the same spectrum also when it is divided into multiple spatial components.
We also study the spatial morphology of the quadrants by substituting the DM template with uniform annuli divided into quadrants.
Then, by fitting all the annuli quadrants to the data we find the surface brightness data of each quadrant.
Finally, we fit the surface brightness data of each quadrant with a DM template with $\gamma$ free to vary.
We perform this analysis using the {\tt quad$+$} quadrant configuration.
In Fig.~\ref{fig:quadrantsSB} we show the best-fit $\gamma$ we find for each quadrant if we run the analysis with different IEMs.
The results found are similar among the IEMs employed.
We find $\gamma\sim 1.15$ for the quadrants 2,3,4 and $\gamma\sim 1.3$ for the first quadrant.
The only IEM that gives practically the same $\gamma$ for all quadrants is the {\tt SL ext}.
The differences we see in Fig.~\ref{fig:quadrantsSB} are due to larger negative residuals present in the first quadrant with respect to the others, that gives also a different spatial morphology for the DM template in this region of the Galactic center.
However, the difference in $\gamma$ is of the same order as that found when performing the analysis in annuli with different IEMs (see Fig.~\ref{fig:gammafit}).
Moreover, DM profiles with $\gamma$ of 1.15 and 1.3, if we normalize both cases to have the same flux at $1^{\circ}$, produce a difference at an angular distance of $2^{\circ}$ ($4^{\circ}$) from the Galactic center of roughly 20\% (40\%) between the two cases.
Thus these results do not indicate any significant asymmetry in the GCE.
We will make a further test of the GCE symmetry in Sec.~\ref{sec:position}.
\subsection{Position}
\label{sec:position}
\begin{figure}
\includegraphics[width=0.49\textwidth]{figures/LogLike_position_finer_Baseline.pdf}
\caption{Value of the likelihood as a function of the Galactic longitude and latitude $\rm{Log}(\mathcal{L})(l,b)$ found with the {\tt Baseline} IEM for the energy range between $1-10$ GeV. We show with a cyan star the position for which we maximize the $\rm{Log}(\mathcal{L})(l,b)$.
\label{fig:pos}
\end{figure}
\begin{figure*}
\includegraphics[width=0.49\textwidth]{figures/LogLike_sphericity_finer_Alt1.pdf}
\includegraphics[width=0.49\textwidth]{figures/LogLike_sphericity_finer_Alt2.pdf}
\includegraphics[width=0.49\textwidth]{figures/LogLike_sphericity_finer_Alt3.pdf}
\includegraphics[width=0.49\textwidth]{figures/LogLike_sphericity_finer_Alt6.pdf}
\caption{Value of the likelihood as a function of the parameters {\it ratio} and $\gamma$ $\rm{Log}(\mathcal{L})(ratio,\gamma)$ found with the {\tt Baseline} (top left), {\tt SNR} (top right), {\tt Yusifov} (bottom left) and {\tt SL ext} (bottom right) IEM for the energy range between $1-10$ GeV.
\label{fig:sph}
\end{figure*}
\begin{table*}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|}
\hline
$\rm{Log}(\mathcal{L})-\rm{Log}(\mathcal{L}_{\rm{DM}})$ & Baseline & ICS combined & OB stars & Pulsars& SL ext. & SNR & Yusifov \\
\hline
BB & -1139 & -1192 & -797 & -1434 & -543 & -826 & -1043 \\
DM+NB & +179 & +217 & +38 & +261 & +84 & +135 & +205 \\
BB+NB & +55(-124) & +21(-196) & -34(-72) & +36(-225) & -51(-135) & +15(-120) & +9(-196) \\
\hline
\end{tabular}
\caption{This table represents the difference of $\rm{Log}(\mathcal{L})$ obtained in the case of boxy bulge (BB), DM plus nuclear bulge (DM+NB) or boxy bulge and nuclear bulge (BB+NB) with respect to the case where we use only the DM template (DM). A positive number implies that the fit improves with respect to the case of the use of the DM template. In the last row we also report in parenthesis the difference of the likelihood of the boxy bulge and nuclear bulge with respect to DM plus nuclear bulge ($\rm{Log}(\mathcal{L_{\rm{BB+NB}}})-\rm{Log}(\mathcal{L}_{\rm{DM+NB}})$). In this case negative values imply that the fit with the boxy bulge and nuclear bulge is worse than the case with DM and the nuclear bulge.}
\label{tab:like}
\end{center}
\end{table*}
The last two GCE characteristics we investigate are the position and sphericity.
We still consider an energy range between $1-10$ GeV.
In order to find the position of the GCE we move the center of the DM template from $(l,b)=(0^{\circ},0^{\circ})$ as assumed before, to different locations around the Galactic center.
We choose a square of side $0.8^{\circ}$ and a grid of width $0.1^{\circ}$.
We do not choose a finer grid because of the {\it Fermi}-LAT angular resolution.
In the analysis we use the same background model (i.e., sources and IEM components) and we only move the position of the DM template center.
For each DM template position we make a fit as explained in Sec.~\ref{sec:analysistec} and we save the value of $\rm{Log}(\mathcal{L})$.
This analysis provides the likelihood as a function of the DM template position (longitude $l$ and latitude $b$) $\rm{Log}(\mathcal{L})(l,b)$.
The best-fit position is at about $l=-0.1^{\circ},b=-0.0^{\circ}$ for the {\tt Baseline} IEM.
The best fit position, considering all the IEMs and analysis setup, changes between $l=[-0.3^{\circ},0.0^{\circ}]$ and $b=[-0.1^{\circ},0.0^{\circ}]$.
Considering that the grid resolution is $0.1^{\circ}$ and that the best-fit position changes due to the IEMs and analysis setup more than the grid size, we conclude that the GCE position is compatible with the dynamical center, of the Milky Way that is considered to be Sagittarius A$^{\star}$, whose position is $(l,b) = (359.94,-0.05)$.
We remind that Sagittarius A$^{\star}$ is included in our background model with the source 4FGL J1745.6-2859.
We show in Fig.~\ref{fig:pos} the $\rm{Log}(\mathcal{L})$ as a function of the coordinates of the DM template obtained with the {\tt Baseline} model.
We also run this analysis for the other IEMs used in Fig.~\ref{fig:spectrumcases}, finding similar results for all of them.
Refs.~\cite{Daylan:2014rsa,Linden:2016rcf} performed a similar analysis with only one IEM and found that the position is compatible with the Galactic center.
On the other hand, Ref.~\cite{TheFermi-LAT:2017vmf} analyzed the position at different latitudes and found that the best fit for $b\in[-1^{\circ},1^{\circ}]$ is at $l \sim 1^{\circ}$. Moreover, the grid used in Ref.~\cite{TheFermi-LAT:2017vmf} to make the likelihood analysis as a function of the position seems to be of the order of $0.5^{\circ}$, i.e., much wider than what we assume in our paper.
Finally, Ref.~\cite{Karwin:2016tsw} found a best-fit position for the GCE template at roughly $-0.4^{\circ}$ of longitude and $0^{\circ}$ in latitude. They perform a likelihood analysis placing the DM template at different locations with a grid of $0.2^{\circ}$ spacing. Their results are thus similar to ours considering the larger grid size.
\subsection{Sphericity}
\label{sec:sphericity}
The sphericity of the GCE is one of the main features that can help to disentangle among the different interpretations.
Indeed, if the GCE is due to DM we expect a spherically symmetric signal.
In case the GCE is due to bremsstrahlung or $\pi^0$ decay we should have a signal elongated on the Galactic plane since these processes trace the distribution of gas in the Galaxy.
Finally, in \cite{Macias:2016nev,Macias:2019omb} the authors claim that using a boxy template, tracing old stars in the Galactic bulge, plus a template that follows the nuclear bulge, they are able to find a much better fit with respect to a DM template.
The preference for the boxy bulge over the DM template should be demonstrated by finding that the GCE is not spherically symmetric.
Very recently, some of the authors of \cite{Macias:2016nev} have generated in \cite{Coleman:2019kax} a new model for the boxy bulge template that better models the stellar bulge and provides a better fit to the GCE. The authors of \cite{Bartels:2017vsx} reach similar conclusions for the preference of fitting the GCE with a template that follows the stellar bulge.
In order to study the GCE sphericity and investigate which of the above interpretations is more compatible with the GCE, we modify the DM template by introducing an ellipsoid to model its spatial distribution. In particular we introduce a parameter, called {\it ratio}, that parametrizes the ellipticity.
This is defined as the ratio between the axis along $l=(180^{\circ},0^{\circ})$ and $b=0^{\circ}$ and the axis on the perpendicular direction, i.e., defined along $l=0^{\circ}$ and $b=\pm90^{\circ}$.
Therefore, an ellipsoid with a ratio larger than one implies that the template is elongated along the Galactic plane.
The boxy bulge model used in \cite{Macias:2016nev,Macias:2019omb,Coleman:2019kax} is very similar to an ellipsoid with ratio equal to 1.8.
We generate templates with different {\it ratio} values between 0.5 to 2.4 and for $\gamma$ from 0.8 to 1.7.
For each ellipsoid template, defined for a given {\it ratio} and $\gamma$ values, we perform a fit to the data.
We show in Fig.~\ref{fig:sph} the contour plots for the likelihood as a function of ratio and $\gamma$ ($\rm{Log}(\mathcal{L})(ratio,\gamma)$) that we obtain when running the analysis with the {\tt Baseline}, {\tt SNR}, {\tt Yusifov} and {\tt SL ext} IEMs.
With all the other IEMs and selections of the data and analysis techniques we reach similar conclusions.
We obtain best-fit values for the parameters of ratio $=0.8-1.2$ and $\gamma = 1.15-1.25$ meaning that there is not a clear preference for a GCE that is not spherically symmetric.
More importantly, the value of the ratio compatible with the boxy bulge used in \cite{Coleman:2019kax} is roughly 1.8.
This is very far from the maximum value we find using all the IEMs, which is 1.2.
Therefore, a spherically symmetric morphology seems to be preferred over a non-spherical template such as the boxy bulge model used in \cite{Macias:2016nev,Macias:2019omb,Coleman:2019kax}.
Our results for the ellipticity are consistent with the ones reported in \cite{Calore:2014xka,Daylan:2014rsa}.
However, the exercise done before does not provide any evidence that the spherically symmetric DM template performs better than the nuclear and boxy bulge templates. In fact, these latter are not exactly elliptical and the angular profile does not follow a NFW.
In order to test directly the interpretation published in Refs.~\cite{Macias:2016nev,Macias:2019omb,Coleman:2019kax}, we perform a fit in the energy range between $1-10$ GeV considering a DM template or the boxy bulge as reported in \cite{Coleman:2019kax}.
We also test in both cases the addition of the nuclear bulge template reported in \cite{Macias:2016nev}.
We use the same sources and IEM components for the DM and stellar bulge scenarios so the likelihood value of the fit is a direct information of which of the two scenarios fit better the GCE.
We generate the DM template using the best fit values for $\gamma$ found in Sec.~\ref{sec:spacial} with the fit to the GCE surface brightness data.
In Tab.~\ref{tab:like} we report the difference between the $\rm{Log}(\mathcal{L})$ obtained with the DM template (labelled as DM) and the following cases: with the boxy bulge (BB), DM plus the nuclear bulge (DD+NB) and the boxy bulge plus the nuclear bulge (BB+NB).
We show the results for different IEMs. For the models not reported in the table we find similar results.
The first thing we notice is that in the DM case the addition of the nuclear bulge template improves the fit by roughly $\Delta \rm{Log}(\mathcal{L})$ of $100-200$.
On the other hand, in the case of the boxy bulge the improvement is much larger $\sim500-1200$.
Finally, the case with DM only is preferred over the boxy bulge case for all the IEMs roughly for $\Delta \rm{Log}(\mathcal{L}) = 500-1200$.
The same is also true for the case DM+NB and BB+NB but with much smaller $\Delta \rm{Log}(\mathcal{L})$ values.
To conclude even if the addition of the nuclear bulge improves significantly the fit in the stellar bulge scenario, the DM interpretation still fits better the GCE for all IEMs considered in the analysis.
We have reported a similar test in Sec.~\ref{sec:spectrum} by performing a fit to the ROI using the {\tt IC bulge} model after having deleted from it the DM template.
In this case, we were fitting the GCE with the inverse Compton emission from electrons and positrons injected from the Galactic bulge using the model in \cite{TheFermi-LAT:2017vmf}. The spatial template we used is reported in Fig.~\ref{fig:fluxmap} for a $\gamma$-ray energy of 1 GeV.
This model is more complex than the one reported in Ref.~\cite{Macias:2016nev} because it is energy dependent since it is calculated fully solving the propagation equation of electrons and positrons in the Galactic bulge environment.
We tested different vertical sizes of the Galactic halo and for all the cases the fit with this template was worse than the DM template by $\Delta \rm{Log}(\mathcal{L}) \sim 230-250$ depending on the value for $z$. So the Galactic bulge model published in \cite{TheFermi-LAT:2017vmf} performs better than the one in Ref.~\cite{Macias:2016nev}.
The different results we obtain with respect to Refs.~\cite{Macias:2016nev,Macias:2019omb,Coleman:2019kax} are probably due to the different assumptions we make for the background components.
Probably, the most relevant ones are related to the choices for the IEM. Refs.~\cite{Macias:2016nev,Macias:2019omb,Coleman:2019kax} use interstellar gas components divided in rings (four each for the HI and H2 related emission) and include two dust correction templates too. These components are modeled with an energy independent morphology. The ICS emission is divided into 6 rings and the low-latitude component of the {\it Fermi} bubbles is also different with respect to the one we use. Finally, they include sources from the 2FIG catalog taken from Ref.~\cite{Fermi-LAT:2017yoi} while we use the 4FGL catalog.
There are other differences that are expected to provide a mild effect in the results such as selecting 7 years of LAT data instead of 11 as we do, using a smaller ROI of $30^{\circ}\times30^{\circ}$ and not using the weighted likelihood technique.
We will work on a follow-up analysis dedicated entirely to test the nuclear and boxy bulge models used in \cite{Macias:2016nev,Macias:2019omb,Coleman:2019kax} and on the investigation of the effect that the choice of different IEMs has on the preference of the bulge model over the DM scenario.
\section{Conclusions}
\label{sec: conclusions}
We analyze 11 years of {\it Fermi}-LAT data detected from the Galactic center region and we provide updated measurements of the characteristics of the GCE.
We utilize the state of the art IEMs created in \cite{TheFermi-LAT:2017vmf} for different source distributions, gas maps and inverse Compton configurations and test different selections of the data and analysis techniques.
Employing this strategy, we derive for the first time the systematics on several GCE properties due to the choice of several IEMs, data selections and analysis techniques.
First, we measure the energy spectrum which is peaked at a few GeV and is compatible with a log-parabola with spectral index of $-2.0$ and curvature index of 0.27.
We also find that the GCE SED normalization systematics are roughly a factor of 50\% at 2 GeV.
Our results are compatible with previous papers such as \cite{Calore:2014xka,TheFermi-LAT:2017vmf}.
We also verify that by including the emission of $\gamma$ rays produced by cosmic rays injected from the Galactic center in the IEM, the GCE spectrum is significantly modified, however, it is still significantly detected.
We investigate the spatial morphology of the GCE using a model independent technique based on an analysis using concentric and uniform annuli.
Fitting the annuli to the data we find the surface brightness of the GCE and we verify that it is compatible with a DM template generated for a generalized NFW with $\gamma=1.1-1.2$ and an average value calculated among all the tested cases of 1.25.
The GCE extends with a significant flux up to roughly $12^{\circ}$ and its spatial morphology is well determined in the energy range between $0.6-30$ GeV.
We also use a model dependent technique where we use directly a DM template created with a generalized NFW and we confirm the previous results for the best-fit values of $\gamma$.
We also study the GCE taking a DM template and dividing it into quadrants. We derive the SED and the spatial distribution of each quadrant.
The SED is similar for each quadrant and well fit with the same SED found for the GCE modeled with a unique DM template.
The spatial distribution is also well fit with a DM template with $\gamma=1.1-1.3$, where the first quadrant (defined for $(l,b)>0$), has slightly larger values of $\gamma$.
We determine the best-fit position of the GCE centroid which is found to be $l=[-0.3^{\circ},0.0^{\circ}]$ and $b=[-0.1^{\circ},0.0^{\circ}]$ using the different IEMs and analysis setup of our analysis. Considering the variation in its value and the resolution of the grid we used in the analysis, which is $0.1^{\circ}$, the GCE position is thus compatible with the dynamical Milky Way center Sagittarius A$^{\star}$.
Finally, we demonstrate that the GCE is compatible with a spherically symmetric template considering the systematic uncertainties associated with the choice of the IEM.
Indeed, if we use a DM template modeled with an ellipsoid, we find that the ratio of the horizontal (i.e., along the Galactic plane) and vertical axis is between $0.8-1.2$.
Our results for the ellipticity are consistent with the ones reported in \cite{Calore:2014xka,Daylan:2014rsa}.
The best-fit region we find for the DM spatial morphology is not compatible with the one used in \cite{Macias:2016nev,Coleman:2019kax} for the boxy bulge which roughly similar to an ellipsoid with a ratio $\sim 1.8$.
We tested this also with a fit to the data finding that the DM template provides a much better log-likelihood value with respect to the boxy bulge template derived in \cite{Macias:2016nev,Coleman:2019kax} or the model reported in \cite{TheFermi-LAT:2017vmf} that we label as {\tt IC bulge} and calculated for $\gamma$ rays produced for inverse Compton from electrons and positrons injected from the Galactic stellar bulge.
Our results differ from the ones in Refs.~\cite{Macias:2016nev,Macias:2019omb,Coleman:2019kax} where the boxy and nuclear bulge templates are preferred over the DM scanario. The difference is very likely due to different background components used in the analysis.
This paper thus provides evidence that the GCE SED has the following properties:
\begin{itemize}
\item {\bf DM SED}: the spectrum has a bump at around 2 GeV and is well described with a log-parabola with spectral index of $-2.0$ and curvature index of 0.27.
\item {\bf Spatial morphology}: the GCE is well modeled with a DM template parametrized with a generalized NFW density profile with $\gamma=1.1-1.2$.
\item {\bf Energy evolution of spatial shape}:
The energy evolution in the energy range $0.6-30$ GeV is constrained to be less than $10\%$ of the $\gamma$ value from its average value that is 1.25.
\item {\bf Position}: The GCE is centered at $l=[-0.3^{\circ},0.0^{\circ}]$ and $b=[-0.1^{\circ},0.0^{\circ}]$ and thus compatible with the dynamical Milky Way center.
\item {\bf Sphericity}: Considering an ellipsoid, the ratio between the horizontal, aligned along the Galactic plane, and vertical axis is between $0.8-1.2$, so the GCE is roughly spherically symmetric.
\end{itemize}
We will present specific interpretations of the GCE in a companion paper.
However, in Ref.~\cite{Dimaurosim} we have studied the properties of the GCE in case it is produced by the different mechanisms discussed in the introduction.
Therefore, we can already try to draw some conclusions.
The GCE from cosmic-ray protons would not be spherically symmetric but rather significantly elongated on the Galactic plane (see left panel of Fig.~\ref{fig:fluxmap}).
The signal produced from cosmic-ray electrons would have a spatial extension that evolves with the energy and a spatial morphology that is not spherically symmetric (see right panel of Fig.~\ref{fig:fluxmap}).
Indeed, we have found in Ref.~\cite{Dimaurosim} that the value of $\gamma$ would change by a factor of 2 in the energy range $1-100$ GeV and in this paper that the surface brightness changes shape significantly between 0.7 GeV and 20 GeV (see Fig.~\ref{fig:SB}).
Finally, a millisecond pulsar population generated by the old stars present in the Galactic bulge would create a GCE that is compatible with all the properties we measured in this paper except for the morphology that, if modeled with the templates introduced in \cite{Macias:2016nev,Coleman:2019kax}, would not be spherically symmetric.
Indeed, a DM template is a better fit to the GCE than the boxy bulge scenario in \cite{Macias:2016nev,Coleman:2019kax}.
A millisecond pulsar population that is spherically symmetric would result in a fit that is as good as the one of DM.
Finally, a DM signal would be perfectly compatible with the GCE properties. However, the annihilation cross section needed to fit the GCE flux is constrained by the non-detection of $\gamma$-rays from Milky Way dwarf spheroidal galaxies (see, e.g., \cite{Hoof:2018hyn}).
Indeed, the GCE SED is roughly compatible with DM particles with a mass of 50 GeV and annihilating into a $b\bar{b}$ channel with an annihilation cross section of about $\langle \sigma v \rangle \sim 3-4 \times 10^{-26}$ cm$^3$/s\footnote{To obtain this result we have simply rescaled the results in \cite{Calore:2014xka}}. The upper limits found from Milky Way dwarf spheroidal galaxies are for the same DM candidate in the range $\langle \sigma v \rangle < 2-3 \times 10^{-26}$ cm$^3$/2.
In order to further improve our knowledge of the GCE characteristics, a significant improvement in the modeling of the IEM is necessary.
Indeed, as we have demonstrated in this paper and in Ref.~\cite{Dimaurosim}, this component is the main source of systematic in the analysis of this complicated region of the sky. New gas maps, interstellar radiation field data and refined source distribution models need to be derived in order to achieve this goal.
\begin{acknowledgments}
The author thanks Michela Negro, Dmitry Malyshev, Chris Karwin for insightful discussions.
The author's research is supported by Fellini - Fellowship for Innovation at INFN, funded by the European Union's Horizon 2020 research programme under the Marie Sklodowska-Curie Cofund Action, grant agreement no.~754496.
He acknowledges support by the NASA Fermi Guest Investigator Program Cycle 12 through the Fermi Program N. 121119 (P.I.~MDM).
The {\it Fermi} LAT Collaboration acknowledges generous ongoing support from a number of agencies and institutes that have supported both the development and the operation of the LAT as well as scientific data analysis. These include the National Aeronautics and Space Administration and the Department of Energy in the United States, the Commissariat\'a l'Energie Atomique and the Centre National de la Recherche Scientifique / Institut National de Physique Nucl\'eaire et de Physique des Particules in France, the Agenzia Spaziale Italiana and the Istituto Nazionale di Fisica Nucleare in Italy, the Ministry of Education, Culture, Sports, Science and Technology (MEXT), High Energy Accelerator Research Organization (KEK) and Japan Aerospace Exploration Agency (JAXA) in Japan, and the K. A. Wallenberg Foundation, the Swedish Research Council and the Swedish National Space Board in Sweden.
Additional support for science analysis during the operations phase is gratefully acknowledged from the Istituto Nazionale di Astrofisica in Italy and the Centre National d'Etudes Spatiales in France. This work performed in part under DOE Contract DE- AC02-76SF00515.
\end{acknowledgments}
|
\section*{Nomenclature}
{\renewcommand\arraystretch{1.0}
\noindent\begin{longtable*}{@{}l @{\quad=\quad} l@{}}
$z$ & latent feature vectors of airfoils \\
$\hat{z}$ & random noise sampled from Gaussian distribution \\
$x$ & airfoils \\
$\tilde{x}$ & airfoils reconstructed from $z$ \\
$\hat{x}$ & airfoils synthesized from $\hat{z}$ \\
$\mathcal{L}_{prior}$ & prior loss \\
$\mathcal{L}_{recon}$ & reconstruction loss \\
$\mathcal{L}_{layer}$ & layer like loss \\
$\mathcal{L}_{GAN}$ & adversarial GAN loss \\
$C_l$ & lift coefficient \\
$C_d$ & drag coefficient \\
\end{longtable*}}
\section{Introduction}
\noindent
\lettrine{R}{ecent} years have witnessed the success of deep learning \cite{lecun2015deep} in many fields like computer vision \cite{voulodimos2018deep}, natural language process \cite{cambria2014jumping} and robotics \cite{pierson2017deeprobot} \cite{carrio2017review}. Such data-driven methods can automatically learn compact and comprehensive representations from samples. However, most of the prevalent deep learning models are based on supervised learning, meaning the samples are paired with manually tagged labels. Such supervision makes the model hard to generalize since both the amount of labeled data and the information contained in the label are limited. Hence, self-supervised learning is proposed to learn the features directly from data. Variational Autoencoder (VAE) \cite{kingma2013auto} \cite{rezende2014stochastic} follows the insight via an encoder-decoder structure, where the encoder down-samples high-dimensional input into a latent feature domain while the decoder reconstructs the sample from learned low-dimensional feature. By minimizing the difference between the reconstructed sample and the original one, VAE automatically learns features without any labels. Generative Adversarial Network (GAN) \cite{goodfellow2014gan} pushes learning from self-supervision even further via a min-max game between a generator and a discriminator. The discriminator works as a classifier to determine real samples from synthesized fake ones; meanwhile, the generator which synthesizes samples from random noise is intended to cheat the discriminator. By jointly training the two components, GAN can generate super high-quality realistic samples \cite{karras2019style}, which VAE fails to achieve. Through learning to reconstruct or synthesizing samples, the self-supervised models automatically encode high-dimensional input into informative features, which can be generalized to different tasks without restrictions from human labels.
Such a self-supervised learning manner can help the parameterization of various geometries using the learned representations. Geometry parameterization plays an important role in shape design \cite{samareh2001survey,chang20113d,salunke2014airfoil}, and geometry heavily influences the performance, especially in the design of aerodynamic products like airfoils. A practical and effective airfoil design must meet certain aerodynamic requirements, like lift, drag, pitching moment, and critical-speed characteristics \cite{abbott1945summary}.
However, due to the curse of dimensionality, design optimization based on CFD simulations can be hard or even infeasible on the airfoil geometry domain, especially for gradient-free methods \cite{rajnarayan2008multifidelity}. Therefore, parameterization or dimension-reduction is required to define the design domain before any design optimization is applied.
Traditional parameterization or dimension-reduction techniques rely on manually selected design parameters like control points of B\'ezier curves \cite{sederberg1992bezier} or B-splines \cite{derksen2010bezierparsec}, which places restrictions on the generalization to various shapes as well as synthesizing novel geometries. Implementation of self-supervised deep learning methods on shape parameterization, like VAE and GAN, can overcome the limitations of traditional techniques and synthesize shapes with great novelty, which can provide insights for future geometric design.
In this work, VAEGAN \cite{larsen2015vaegan}, which is a combination of Variational Autoencoder (VAE) and Generative Adversarial Network (GAN), is utilized to learn feature vectors and generate novel airfoils without any human prior. VAEGAN takes advantage of both VAE and GAN. With the encoder-decoder architecture from VAE, the model learns to explicitly encode an existing airfoil shape into a low-dimensional feature domain and reconstruct the shape with little error. With the discriminator from GAN, our model can generate a large number of high-quality novel airfoils from random noise with no manually designed parameters.
Our experiments show that generated airfoils are smooth even without any smoothing post-process. K-means clustering in the learned feature domain demonstrates that feature vectors encode essential shape information in a way that each cluster represents various shape patterns. Further test on learned latent features illustrates that different geometry information is encoded in each dimension of the representation. The performance of the model in synthesizing novel airfoils is examined as well. By either interpolation or extrapolation of feature vectors, a synthesized airfoil inherits features of parent samples, while generated airfoils from sampled Gaussian noise show great novelty in a way that it is not a simple combination of two existing airfoils. A further experiment on the aerodynamic properties of synthesized airfoils, either by interpolating, extrapolating, or sampling, indicates the synthesized airfoils can possess competitive aerodynamic properties, and some even surpass the existing ones. With a genetic algorithm \cite{mitchell1998introduction}, airfoil geometries can be optimized on the feature domain and evolve to possess specific aerodynamic properties. Our model proves its ability to parameterize the existing airfoil shape as well as generating novel and practical airfoils; both lead to designing the new generation of airfoils more intelligently and efficiently without intensely relying on experimental experience or manually design parameters.
\section{Related Works}
\noindent
We propose to use a generative model to synthesize novel airfoils without any predefined design parameters. Through training via the gradient-based method, our deep learning model also automatically learns to parameterize airfoils into latent feature vectors. To better address the insight of our work, this section reviews previous work on shape parameterization, especially on aerodynamic geometries and the implementation of deep learning on geometry design and synthesis.
\subsection{Geometry Parameterization and Dimension Reduction}
\noindent
A lot of work has been done in parameterizing complex shapes and reducing geometric dimensions. Some research focuses on analytically expressing the curves. In \cite{hicks1978wing}, it is introduced that by adding analytic shape functions to the baseline shape, a compact formulation for parameterization can be obtained. The design variables, in this case, are the coefficients correlated with shape functions. By this means, the analytical function of the airfoil curve is formulated. Also, \cite{hager1992multi} follows the same formulation strategy but with different shape functions. Another common and efficient method is PARSEC \cite{sobieczky1997parsec}, which defines eleven geometric parameters to thoroughly express the airfoil shape, including upper and lower curvature, thickness, leading-edge radius, etc. With the defined parameters, a linear combination of shape functions is introduced to describe the airfoil shape as well. Such methods work well for a specific set of curves but may fail to express complex geometries since they rely on manually designed parameters.
Polynomial and spline are also utilized to help dimension reduction. With different orders of polynomials as the basis, the airfoil shape can be described as a linear combination of the basis \cite{elliott1997practical,taylor1991sensitivity}. However, high-order terms can overfit to high-frequency noise, especially when coefficients are of different magnitudes. Besides polynomial, B\'ezier curve \cite{sederberg1992bezier}, which is built upon the Bernstein polynomials, is another mathematical formulation of curves. In detail, $n+1$ control points of B\'ezier are needed to define an $n$-degree B\'ezier curve. Although B\'ezier and polynomial curves are mathematically equivalent, B\'ezier usually perform better in controlling a curve since control points are closely related to the curve position and shape. To mitigate the rounding error, De Casteljau \cite{boehm1999casteljau}, a recursive algorithm, is introduced to compute the Bernstein polynomials numerically. Besides, B\'ezier-PARSEC \cite{derksen2010bezierparsec}, which combines B\'ezier and PARSEC, uses PARSEC parameters to define B\'ezier curves. The B-spline curve with B-spline basis functions is also utilized to describe the airfoil shape. Yet B-spline formulation fails to represent implicit conic sections accurately. That is why Non-uniform rational B-spline (NURBS) is introduced \cite{farin2014curves}. NURBS can accurately represent both standard geometric objects like lines, circles, ellipses, and cones, as well as free‑form geometry, which are prevalent in industrial design.
Another popular technique in shape parameterization is free-form deformation (FFD), which utilizes high-level shape deformation instead of lower-level geometric entities to represent a shape. Based on this insight, \cite{sederberg1986free} presents a technique which can apply deformation either locally or globally based on trivariate Bernstein polynomials. Deformation is manipulated by control points of trivariate B\'ezier volumes. In \cite{coquillart1990extended}, an extended free-form deformation (EFFD) method is presented, which allows arbitrarily shaped deformations by using non-parallelepiped lattices. Research presented in \cite{yeh1998applying} incorporates FFD and sensitivity analysis where geometry changes and structural responses are correlated, and the shape satisfying deformation or stress constraints can be found easily.
Other methods like leveraging camber and thickness mode shapes derived from existing airfoils are also used to parameterize the airfoil shapes \cite{li2019data}. Also, the linear reduction method like the SVD is utilized to extract airfoil representations and optimize shape design \cite{poole2019efficient}. The conventional dimension reduction or parameterization techniques have been implemented in different scenarios and successfully represented the existing airfoil shapes. However, these methods usually require pre-defining the design space as well as the boundary of design space, like the design parameters in NACA, shape functions, etc., which can degrade the synthesis of novel/new airfoils and optimization towards the desired design.”
\subsection{Deep Learning in Geometry Design}
\noindent
In recent years, deep learning has been a great success in extracting informative features from data \cite{lecun2015deep}. Especially in a supervised visual learning manner, with the invention of convolutional neural network (CNN) \cite{lecun1989lenet,lecun1998gradient}, deep learning has been general solutions in many fields, like image classification \cite{szegedy2015going} \cite{he2015delving}, object detection \cite{ren2015fasterrcnn,redmon2016yolo}, and segmentation \cite{li2017fully,he2017maskrcnn}. Also, with the introduction to GAN \cite{goodfellow2014gan}, deep learning in a self-supervised manner has been widely used in synthesizing realistic samples, and some sophisticated GAN-based model can generate high-resolution images which are even hard to distinguish by humans \cite{karras2019style}. With the ability to extract representative features and synthesize realistic samples, deep learning has been widely used to make geometric design more systematic and efficient. This section will introduce some work of deep learning in geometric design, especially in aerodynamic shape design.
In the work of \cite{norgaard1997neural}, a multi-layer perceptron (MLP), which takes the angle of attack and flap setting as input, is trained to predict multiple aerodynamic coefficients, including lift coefficient, drag coefficient, and moment of inertia. Further, \cite{zhang2018application} utilizes CNN with airfoil images as input to learn the lift coefficients of different airfoils in multiple flow Mach numbers, Reynolds numbers, and diverse angles of attack. Similarly, in \cite{yilmaz2017convolutional}, CNN is implemented to predict the pressure coefficient value at the test point. Moreover, \cite{yilmaz2018deep} presents a CNN-based method to learn the correlation between airfoil geometry and pressure distribution. The model can also conduct an inverse airfoil design given the pressure.
On the other hand, the GAN model \cite{goodfellow2014gan}, as a self-supervised learning framework, has been a prevalent and great success in generating realistic samples.
GAN, with a discriminator and a generator, is designed to learn features and generate samples without manually tagged labels. The discriminator is a classifier telling the true input from the synthesized fake one, while the generator is intended to generate plausible samples to cheat the discriminator. Since first introduced \cite{goodfellow2014gan}, many pieces of research have been dedicated to pushing the edge of GAN. In Wasserstein GAN \cite{arjovsky2017wasserstein}, by introducing Wasserstein distance, the quality of generated samples can be well measured during training. ConditionalGAN \cite{mirza2014conditional} and InfoGAN \cite{chen2016infogan} extend GAN to generate a sample from various categories within one model. Also, DCGAN \cite{radford2015unsupervised} and StyleGAN \cite{karras2019style} can generate high-quality realistic samples with sophisticated architectures.
With the ability to generate plausible samples, GAN provides a powerful architecture to learn representations that can help shape design as well. Based on this insight, \cite{chen2018beziergan,chen2019aerodynamic} introduce B\'ezierGAN, which uses a GAN to generate B\'ezier curve control points and then uses the control points to formulate the boundary of airfoils. Such pipeline guarantees generated airfoils to be smooth, and further shape optimization can be conducted on the feature domain \cite{chen2019aerodynamic}. Such a method is, however, restricted to B\'ezier curves and fails to encode existing airfoil shapes explicitly. Further, GAN can be implemented in generating three-dimensional samples, as shown in \cite{wu20163dgan,huang2015analysis,sinha2017surfnet}. \cite{wu20163dgan} proposes to use a three-dimensional convolutional layer to generate volumetric objects. While \cite{huang2015analysis} trains a generative model of three-dimensional shape surfaces, which directly encodes surface geometry and shape structure, \cite{kalogerakis2012probabilistic} further proposes a model to represent probabilistic relationships between properties of shape components and relates them to learned underlying causes of structural variability within the domain.
In our work, we propose to use a VAEGAN-based model \cite{larsen2015vaegan} to extract features of airfoil shapes and synthesize new airfoil designs. The learned latent features encode the airfoil shape and can be utilized to synthesize new designs through interpolation or extrapolation. In comparison to conventional parameterization methods, our model can generate a wider variety of new airfoils, and some show promising aerodynamic properties. By applying the genetic algorithm, the VAEGAN synthesized airfoils can be optimized to desired aerodynamic properties.
\section{Proposed Method}
\subsection{Data Pre-processing}
\noindent
The UIUC Coordinates Database \cite{uiucdataset}, which contains more than 1,600 2-dimensional airfoils, is used to train the generative model. Each airfoil in the database is represented by varying numbers of points with x, y coordinates. Such variation restricts data to be fed directly into a neural network model, which requires a homogeneous input. To deal with it, we first scale the x coordinates of all airfoils to $[0,1]$. Then all airfoils are interpolated by splines, and $N$ points are selected with $x$ given in Eq.~\ref{eq:x}, where $i$ represents the index of each point.
\begin{equation}
\begin{split}
\theta_i &= \frac{\pi (i-1)}{N} \\
x_i &= 1 - \cos(\theta_i)
\end{split}
\label{eq:x}
\end{equation}
In our case, $N$ is set to be 200 with 100 points from the upper boundary and the other 100 from the lower boundary. By this means, all the interpolated airfoils share the same x coordinates. Therefore, only the y coordinates of each airfoil are fed into the model, which reduces the dimensionality of the data. Finally, all y coordinates are scaled to $[-1, 1]$ by multiplying a normalization coefficient. As illustrated in Fig~\ref{fig:preprocess}, the first row shows the original airfoils from the UIUC Coordinate Database, while the second row shows the corresponding processed airfoils.
\begin{figure}[hbt!]
\centering
\includegraphics[width=.8\textwidth]{figures/new_preprocessing.png}
\caption{Data pre-processing: the first row shows the origin airfoil coordinates from the UIUC database, the second row shows the corresponding processed airfoils}
\label{fig:preprocess}
\end{figure}
\subsection{VAEGAN}
\noindent
Our model is based upon VAEGAN \cite{larsen2015vaegan}, which takes advantage of both VAE \cite{kingma2013auto} \cite{rezende2014stochastic} and GAN \cite{goodfellow2014gan}. VAE contains two components: an encoder and a decoder. The former encodes a high-dimensional sample, $x$, into a low-dimensional latent representation, $z$. While the decoder takes as input the latent vector, $z$, and upsamples from the representation domain to the original data domain, $\tilde{x}$. The encoder and decoder are given as:
\begin{equation}
z \sim \text{Enc}(x) = q(z|x), \: \tilde{x} \sim \text{Dec}(z) = p(\tilde{x}|z).
\label{eq:vae}
\end{equation}
To regularize the encoder, VAE takes into consideration a prior distribution of the latent vector, $p(z)$. Here it is assumed that $z \sim \mathcal{N}(0,\mathbf{I})$, which follows an isotropic Gaussian distribution. The loss function for VAE to minimize is given by:
\begin{equation}
\mathcal{L}_{VAE} = \mathcal{L}_{recon} + \mathcal{L}_{prior},
\label{eq:loss_vae}
\end{equation} with
\begin{equation}
\begin{aligned}
\mathcal{L}_{recon} &= ||\tilde{x} - x||_2^2, \text{ and} \\
\mathcal{L}_{prior} &= D_{KL}(q(z|x)||p(z)),
\label{eq:loss_recon_prior}
\end{aligned}
\end{equation}
where $\mathcal{L}_{recon}$ measures how well the reconstructed data, $\tilde{x}$, is comparing to the original $x$ by the mean square error (MSE), and $\mathcal{L}_{prior}$ is the Kullback Leibler divergence (KL divergence), which measures the difference between encoded representation vectors and Gaussian distribution.
VAE learns the representation of samples and can reconstruct them from $z$ with the encoder-decoder architecture. However, it suffers from poor performance in generating novel samples, which have not been seen before \cite{goodfellow2014gan} \cite{larsen2015vaegan}. To this end, a generative adversarial network (GAN) \cite{goodfellow2014gan} is introduced, which contains a discriminator, $\mathcal{D}$, and a generator, $\mathcal{G}$, competing with each other in a self-supervised manner. $\mathcal{G}$ tries to generate plausible samples to fool $\mathcal{D}$, while $\mathcal{D}$ keeps sharpening its decision boundary to determine synthesized fake samples from real ones. In detail, the generator, $\mathcal{G}$, is fed with random noise $\hat{z} \sim p(\hat{z})$ and maps the noise to the data sample domain to generate fake $\hat{x}$. The discriminator takes both $x$ and $\hat{x}$ to predict whether the input is from the real dataset or generated by $\mathcal{G}$. The cross-entropy loss function for the min-max game is given as:
\begin{equation}
\min_G \max_D \mathcal{L}_{GAN} =\log \mathcal{D}(x) + \log(1-\mathcal{D}(\mathcal{G}(\hat{z}))).
\label{eq:loss_gan}
\end{equation}
\begin{figure}[hbt!]
\centering
\includegraphics[width=.7\textwidth]{figures/vaegan.png}
\caption{VAEGAN combines the encoder-decoder structure from VAE and discriminator, $\mathcal{D}$, from GAN}
\label{fig:vaegan}
\end{figure}
As one may expect, GAN generates samples purely from random noise, making it hard to obtain an explicit mapping from the data domain to the feature domain. Therefore, we build our model upon VAEGAN, which combines VAE and GAN. Namely, the generator is replaced by an encoder-decoder structure from VAE, as shown in Figure \ref{fig:vaegan}. Notice that in VAEGAN, the model generates a reconstructed sample, $\tilde{x}$, given a real sample, $x$, and meanwhile generates a fake sample, $\hat{x}$, directly from noise, $\hat{z}$. Both $\tilde{x}$ and $\hat{x}$ should be classified as fake by the discriminator, $\mathcal{D}$, and only $x$ is recognized as the real sample. Hence GAN loss function $\mathcal{L}_{GAN}$ from Eq.~\ref{eq:loss_gan} is modified to:
\begin{equation}
\mathcal{L}_{GAN} = \log (\mathcal{D}(x)) + \log(1-\mathcal{D}(\text{Dec}(z))) + \log (1 - \mathcal{D}(\text{Dec}(\text{Enc}(x)))),
\end{equation}
which takes into consideration the real sample, $x$, reconstructed sample, $\tilde{x}$, and fake sample, $\hat{x}$. Besides, to stabilize the training process and sharpen the decision boundary of $\mathcal{D}$, another loss function, $\mathcal{L}_{layer}$, is introduced when training the encoder and decoder. $\mathcal{L}_{layer}$, as given in Eq.~\ref{eq:loss_layer}, measures the $l_1$ distance between the values of the neurons in one particular layer of $\mathcal{D}$ when fake samples are fed and the values when real samples are fed.
\begin{equation}
\mathcal{L}_{layer} = ||\mathcal{D}_l (x) - \mathcal{D}_l (\text{Dec}(\hat{z}))||_1.
\label{eq:loss_layer}
\end{equation} The complete loss function for VAEGAN is a weighted combination of all the loss terms given by:
\begin{equation}
\mathcal{L} = \lambda_0 \mathcal{L}_{prior} + \lambda_1 \mathcal{L}_{recon} + \lambda_2 \mathcal{L}_{layer} + \lambda_3 \mathcal{L}_{GAN}.
\label{eq:loss_vaegan}
\end{equation} The three components, encoder, decoder and discriminator are trained jointly, and each term of the loss function is assigned with different weights $\lambda$ when training each component.
\subsection{Airfoil Synthesis}
\noindent
By training on the UIUC Database, the VAEGAN model automatically learns to encode airfoils into latent features and reconstruct airfoils from the feature domain. The learned latent features can be directly utilized for dimension reduction and shape parameterization. Moreover, the VAEGAN model is intended to synthesize novel airfoils which are different from samples in the training dataset. To this end, we propose three synthesis methods: interpolation, extrapolation, and sampling, all of which are conducted on the latent feature domain.
More specifically, in interpolation or extrapolation, two airfoils from the UIUC Database are first mapped to latent feature vectors, $z_1$ and $z_2$, via a well-trained encoder. A new feature vector $\bar{z}$, which is an affine combination of $z_1$ and $z_2$, is calculated as given in Eq.~\ref{eq:inter_2}:
\begin{equation}
\bar{z} = \nu z_1 + (1-\nu) z_2,
\label{eq:inter_2}
\end{equation}
where $\nu$ is the coefficient controlling the weight between $z_1$ and $z_2$. When $0 \leq \nu \leq 1$, $\bar{z}$ is an interpolated feature vector, else it is an extrapolation between $z_1$ and $z_2$. The interpolated/extrapolated feature vector, $\bar{z}$, is then fed into the decoder to synthesize an airfoil. Also, such interpolation/extrapolation between two airfoils can be directly extended to a triplet case. Given $z_1$, $z_2$, and $z_3$ are three feature vectors mapped from three different airfoils via the encoder, the expression of triplet interpolation/extrapolation is shown in Eq.~\ref{eq:inter_3}:
\begin{equation}
\bar{z} = \alpha z_1 + \beta z_2 + \gamma z_3,
\text{ where } \alpha+\beta+\gamma=1.
\label{eq:inter_3}
\end{equation}
Similarly, when $0 \leq \alpha,\beta,\gamma \leq 0$, $\bar{z}$ is an interpolation of the three feature vectors, and an extrapolation otherwise.
Besides interpolation and extrapolation, sampling is another method to synthesize novel airfoils. Unlike interpolation or extrapolation, which relies on feature vectors from existing airfoils, sampling generates airfoils directly from random noise. A feature vector $\hat{z}$ is randomly sampled from an isotropic Gaussian distribution $\mathcal{N}(0,\mathbf{I})$ and then mapped to an airfoil via the decoder. By this means, synthesized airfoils from sampling are less restricted since sampled latent vectors are not constrained by features extracted from airfoils in the UIUC Database and are more likely to introduce novelty to the synthesized shapes.
\subsection{Aerodynamic-aware Shape Optimization}
\noindent
So far, how the VAEGAN model is built and used to generate novel airfoils has been introduced. However, the novelty in shape does not guarantee a better airfoil design. To design engineering effective airfoils, aerodynamic properties are supposed to be considered. To this end, we propose to use a genetic algorithm (GA) \cite{poon1995genetic} \cite{mitchell1998introduction} to optimize airfoil shapes by controlling feature vectors learned from the VAEGAN model so that the airfoils can evolve to have the desired aerodynamic properties. Specifically, lift coefficient, $C_l$, and drag coefficient, $C_d$, which measure the aerodynamic force perpendicular and horizontal to the direction of motion, are considered to evaluate the aerodynamic performance of the synthesized airfoils.
As a non-gradient optimization technique, GA is inspired by natural selection and is intended to force individuals to gradually evolve to the optimal. Assume the GA has $N$ generations in total and $M$ individuals in each generation. In our case, individuals are feature vectors. We use $z_i$ to represent all individuals in the $i$\textsuperscript{th} generation, and $z_{i,j}$ for the $j$\textsuperscript{th} individual in the $i$\textsuperscript{th} generation; also the airfoil decoded from $z_{i,j}$ is annotated as $a_{i,j}$. Similarly, $C_l^{i,j}$ and $C_d^{i,j}$ represents lift and drag coefficients of $a_{i,j}$, respectively. The fitness score, $s_{i,j}$, is used to measure the aerodynamic performance of the individual, $z_{i,j}$, as shown in Eq.~\ref{eq:score}:
\begin{equation}
s_{i,j} = - (\frac{C_l^{i,j} - C_l^t}{C_l^t})^2 - (\frac{C_d^{i,j} - C_d^t}{C_d^t})^2,
\label{eq:score}
\end{equation}
where the square of the difference between the target and current aerodynamic coefficients is calculated and normalized by the squared target $C_l^t$ and $C_d^t$. The fitness score is supposed to approach zero as individuals evolve on each generation. As shown in Algorithm~\ref{alg:ga}, the initial generation, $z_0$, is randomly sampled from an isotropic Gaussian distribution, $\mathcal{N}(0,\mathbf{I})$. The GA starts with the selection from the initial generation by randomly picking two individuals and comparing their fitness scores. The one with a higher fitness score wins the tournament and becomes one of the parents. $p_1^i$ and $p_2^i$ denote all the parents 1 and parents 2 in the $i$\textsuperscript{th} generation respectively. Single-point crossover is then implemented to generate offspring from parents 1 and 2. Namely, a crossover point on the parent vector is randomly selected, and all elements after that point are swapped between the two parents. Mutation in the natural selection process is also imitated with additive Gaussian noises.
\begin{algorithm}
\caption{Aerodynamic-aware Shape Optimization via GA}
\begin{algorithmic}[1]
\Procedure{GA}{$N, M, C_l^t, C_d^t$, $p$}
\State Initialize $i := 0$
\State Sample first generation $z_{0j} \sim \mathcal{N}(0, \mathbf{I})$, for $0<j<M$
\While{$i < N$}
\State Synthesize airfoils $a_{i,j}$ from $z_{i,j}$ via decoder
\State Compute $C_l^{i,j}$, $C_d^{i,j}$ and fitness score $s_{i,j} = - (\frac{C_l^{i,j} - C_l^t}{C_l^t})^2 - (\frac{C_d^{i,j} - C_d^t}{C_d^t})^2$
\State Select $M$ parent 1, $p_1^i$, and $M$ parent 2, $p_2^i$, from $z_i$ by tournament
\State Generate next generation, $z_{i+1}$, through single-point crossover
\State With probability $p$, $z_{i+1,j}$ will add a Gaussian noise $\mathcal{N}(0,\mathbf{I})$
\State $i := i+1$
\EndWhile
\Return The individual with the highest score from $z_{i}$
\EndProcedure
\end{algorithmic}
\label{alg:ga}
\end{algorithm}
\section{Experiments}
\noindent
Our VAEGAN model consists of 3 components: an encoder, a decoder, and a discriminator, which are all built on multi-layer perceptron (MLP). As illustrated in Fig~\ref{fig:vaegan}, the encoder encodes 200-dimensional airfoil coordinates into a 32-dimensional feature domain while the decoder maps the feature back to the airfoil. The discriminator is a classifier examining whether the input is a real airfoil from the UIUC Database, or a fake one reconstructed from the decoder, or synthesized from random noises. In detail, the encoder is modeled by a 3-layer MLP with the number of neurons $[256, 128, 32]$ in each layer, and LeakyReLU \cite{maas2013leakyrelu} is implemented as the activation function in each layer. The decoder is also a 3-layer MLP with the number of neurons $[128, 256, 200]$ in each layer. A hyperbolic tangent (Tanh) function works as the activation function in the output layer to scale all outputs into $[-1,1]$. Similarly, the discriminator contains three layers with the number of neurons $[256, 128, 1]$, and outputs the probability of whether the input is real or fake through a Sigmoid activation function.
To automatically learn the latent features and synthesize airfoils, the VAEGAN model is trained on the UIUC Database for 5000 epochs, and each epoch goes through all the samples in the dataset. Initial learning rates for all three components: encoder, decoder, and discriminator are set to be $0.0005$ and decay to $0.00005$ after 2500 epochs. The batch size is set to be 16, which is approximately $1/100$ of the database size. Adam optimizer \cite{kingma2014adam} is utilized to update all the parameters in the model. As mentioned in Section \RNum{3} C, different coefficients are assigned to each term in the loss function Eq.~\ref{eq:loss_vaegan}; also different components, namely the encoder, decoder, and discriminator, have different coefficients, respectively. Coefficients of different loss terms and components are shown in Eq.~\ref{eq:concrete_loss}:
\begin{equation}
\begin{split}
\mathcal{L}_{Enc} &= 0.1 \mathcal{L}_{prior} + 0.1 \mathcal{L}_{layer} + 10 \mathcal{L}_{recon}, \\
\mathcal{L}_{Dec} &= 0.1 \mathcal{L}_{prior} + 0.1 \mathcal{L}_{layer} + 10 \mathcal{L}_{recon} + 5 \mathcal{L}_{GAN}, \\
\mathcal{L}_{\mathcal{D}} &= \mathcal{L}_{GAN},
\end{split}
\label{eq:concrete_loss}
\end{equation}
where $\mathcal{L}_{Enc}$, $\mathcal{L}_{Dec}$, and $\mathcal{L}_{\mathcal{D}}$ represent loss functions for the encoder, decoder, and discriminator, respectively.
To better investigate our VAEGAN-based model, we compare the performance with two other paramterization methods, principle component analysis (PCA) and variational autoencoder (VAE) \cite{kingma2013auto}. PCA conducts a linear transformation from the pre-processed airfoil point coordinates into prioritized latent variables. In our case, the top 32 dimensions are kept as the feature. The VAE follows the same encoder-decoder architecture as the VAEGAN, but lacks the discriminator. The latent feature dimension is also set to 32, and the loss function is given in Eq.~\ref{eq:VAE_loss}:
\begin{equation}
\mathcal{L}_{Enc} = 0.1 \mathcal{L}_{prior} + 10 \mathcal{L}_{recon}.
\label{eq:VAE_loss}
\end{equation}
\subsection{Airfoil Reconstruction via Encoder-decoder}
\begin{figure}[hbt!]
\centering
\includegraphics[width=\textwidth]{figures/new_recon.png}
\caption{Reconstructed airfoils: the first row shows the airfoils from UIUC Coordinate Database, the second row shows the reconstructed airfoils with VAE, and the third row shows the reconstructed airfoils with smoothness}
\label{fig:recon}
\end{figure}
\noindent
The VAEGAN model can automatically learn feature vectors, namely mapping the high-dimension airfoils into low-dimension representations. To estimate whether or not the feature vector fully encodes the geometric information of the original airfoil,we first feed airfoils from UIUC Coordinate Database into the encoder to obtain the encoded feature vectors. The decoder then takes the vectors as input and outputs the reconstructed airfoils. Also, the Savitzky-Golay filter \cite{schafer2011savitzky}, a moving polynomial fitting, is implemented to smoothen the boundary of reconstructed airfoils. In our case, the second-order polynomial is used in the Savitzky-Golay filter, and the length of the moving window is set to be 7. In Fig~\ref{fig:recon}, the first row illustrates samples from the UIUC Database, and the second row shows reconstructed airfoils from corresponding feature vectors, with an MSE, 3.65345$\times 10^{-4}$, between the reconstructed and original airfoils. This small error indicates the learned features well represent the shape of airfoils. The third row shows reconstructed airfoils with the Savitzky-Golay filter with an MSE, 3.65054$\times 10^{-4}$, comparing to the original airfoils. These results further demonstrate that the encoder-decoder can reconstruct airfoils that are smooth and realistic without smoothing filters. Also, our VAEGAN-based model is compared with PCA and VAE as shown in Table~\ref{tab:MSE}. PCA reaches the lowest MSE since it provides a close form solution, whereas VAE and VAEGAN are optimized numerically via the gradient-based method. With the discriminator and adversarial loss from GAN, VAEGAN model performs slightly better than VAE in reconstruction. It should be pointed out that all the three parameterization methods have small reconstruction MSEs of magnitude $10^{-4}$, meaning all the features extracted well encodes the airfoil shapes from the UIUC database.
\begin{table}[htb!]
\caption{\label{tab:MSE} Mean squared error of airfoil reconstruction via different featurization techniques}
\centering
\begin{tabular}{lccc}
\hline
Featurization & PCA & VAE & VAEGAN \\ \hline
MSE & 1.29208$\times 10^{-4}$ & 3.70018$\times 10^{-4}$ & 3.65345$\times 10^{-4}$ \\
\hline
\end{tabular}
\end{table}
\subsection{Clustering in Feature Domain}
\noindent
The encoded features obtained from the encoder in our method can help better understand the shape of current airfoils. All airfoils from the UIUC Coordinate Database are first mapped to feature vectors, and an unsupervised learning algorithm, K-Means \cite{jain2010kmeans}, is used to cluster these airfoils in the feature domain. To visualize the 32-dimensional feature domain, we use Parametric t-distributed Stochastic Neighbor Embedding (parametric t-SNE) \cite{maaten2008tsne, van2009ptsne} as a visualization tool. Parametric t-SNE is modeled by MLP, which maps the high dimensional feature vector $z_i$ into a low-dimension embedding $y_i$, while keeps the similarity between points. It converts similarities between data points to joint probabilities and minimizing the KL divergence between the joint probabilities of embedding $y_i$ and the original feature vector $z_i$. Eq.~\ref{eq:tsne} shows the cost function $C$, which t-SNE is expected to minimize:
\begin{equation}
\begin{aligned}
p_{j|i} = \frac{\exp(-\|z_i-z_j\|^2/2\sigma_i^2)}{\sum_{k \neq i} \exp (-\|z_i-z_k\|^2/2\sigma_i^2) }, & \;
q_{j|i} = \frac{\exp(-\|y_i-y_j\|^2/2\sigma_i^2)}{\sum_{k \neq i} \exp (-\|y_i-y_k\|^2/2\sigma_i^2) }, \\
C = KL(P || Q) = & \sum_i \sum_j p_{j|i} \log \frac{p_{j|i}}{q_{j|i}},
\end{aligned}
\label{eq:tsne}
\end{equation}
where $\sigma_i$ is calculated by a binary search given a fixed perplexity that is specified by the user \cite{maaten2008tsne}. Fig.~\ref{fig:cluster} shows the K-means clustering results visualized with parametric t-SNE, where different colors represent different clusters, and the centroid of each cluster is also shown. The centroid of each cluster is different from each other in symmetry, height, camber, etc., and each represents the geometric pattern of each cluster. Features from close clusters represent similar airfoil shapes. The distance between feature points intuitively reflects the difference between the two airfoil shapes. This indicates that our VAEGAN-based model learns features that maintain the similarity of input airfoil shapes.
\begin{figure}[hbt!]
\centering
\includegraphics[width=\textwidth]{figures/cluster_airfoils.png}
\caption{Parametric t-SNE visualization for clustering on feature domain}
\label{fig:cluster}
\end{figure}
\subsection{What is Encoded in the Feature Domain}
\noindent
Also, experiments are conducted to investigate what geometric features are encoded in each dimension of the learned representation. A series of manually designed feature vectors are fed into the decoder, where all the elements are set to zero except for one specific dimension. That particular element is changed gradually from $-10$ to $10$, and the designed feature vectors are mapped to the airfoil coordinate domain by the decoder. Changes of generated airfoils are illustrated in Fig~\ref{fig:change}, and the 2D embedding of feature vectors using parametric t-SNE is shown in Fig~\ref{fig:change_cluster}. Here, only four dimensions are chosen for analysis purposes.
The 1\textsuperscript{st} dimension encodes the height of the upper boundary, as illustrated in Fig~\ref{fig:change_1}. As the 1\textsuperscript{st} dimension increasing from $-10$ to $10$, the height of the front half airfoil increases while the tail becomes thinner. The 8\textsuperscript{th} dimension encodes the camber of both the upper boundary and the lower boundary. It is shown that by tuning the 8\textsuperscript{th} dimension, the upper boundary of the airfoil changes from a concave curve to a horizontal straight line, while the lower boundary evolves from a concave to a convex curve. Interestingly, the 22\textsuperscript{nd} dimension encodes quite similar representations as the 1\textsuperscript{st} dimension while in the opposite direction. In other words, generated airfoils from feature vectors whose 22\textsuperscript{nd} dimension change from $-10$ to $10$ are like those with the 1\textsuperscript{st} dimension change from $10$ to $-10$ as illustrated in Fig~\ref{fig:change_cluster_1} and Fig~\ref{fig:change_cluster_22}. Besides, Fig~\ref{fig:change_32} shows how the last dimension is connected to the camber of the lower boundary. In detail, the curvature of the lower boundary decreases as the 32\textsuperscript{nd} dimension increases.
In comparison to the features learned by our VAEGAN-based model, Fig.~\ref{fig:vae_change} shows the VAE-synthesized airfoils when changing only one feature dimension. As shown in ~\ref{fig:vae_change_0}, the first dimension of VAE features fails to encode any shape representations. Even in the dimensions where representations are learned as dimension 3 and 31 shown in Fig.~\ref{fig:vae_change_2} and Fig.~\ref{fig:vae_change_30}, the feature does not change continuously as we observe in the VAEGAN results. Our VAEGAN-based model learns more representative and thorough features than the VAE model. Also, the representations in each dimension are entangled, like the 3\textsuperscript{rd} dimension encodes both the upper bound and lower bound. These results indicate that, without manually designed parameters, our VAEGAN-based model learns geometrically meaningful features, and each dimension of the learned feature domain encodes informative and different geometry features.
\begin{figure}[hbt!]
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/change_1.png}
\caption{}
\label{fig:change_1}
\end{subfigure}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/change_8.png}
\caption{}
\label{fig:change_8}
\end{subfigure}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/change_22.png}
\caption{}
\label{fig:change_22}
\end{subfigure}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/change_32.png}
\caption{}
\label{fig:change_32}
\end{subfigure}
\caption{VAEGAN-generated airfoils by gradually changing only one dimension of the feature domain: (a) changes the 1\textsuperscript{st} dimension, (b) changes the 8\textsuperscript{th} dimension, (c) changes the 22\textsuperscript{nd} dimension, and (d) changes the 32\textsuperscript{nd} dimension}
\label{fig:change}
\end{figure}
\begin{figure}[hbt!]
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=.95\linewidth]{figures/vae_feat_tune_0.png}
\caption{}
\label{fig:vae_change_0}
\end{subfigure}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=.95\linewidth]{figures/vae_feat_tune_2.png}
\caption{}
\label{fig:vae_change_2}
\end{subfigure}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=.95\linewidth]{figures/vae_feat_tune_30.png}
\caption{}
\label{fig:vae_change_30}
\end{subfigure}
\caption{VAEGAN-generated airfoils by gradually changing only one dimension of the feature domain: (a) change the 1\textsuperscript{st} dimension, (b) changes the 3\textsuperscript{rd} dimension, and (c) changes the 31\textsuperscript{st} dimension}
\label{fig:vae_change}
\end{figure}
\begin{figure}[hbt!]
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/change_cluster_1.png}
\caption{}
\label{fig:change_cluster_1}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/change_cluster_8.png}
\caption{}
\label{fig:change_cluster_8}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/change_cluster_22.png}
\caption{}
\label{fig:change_cluster_22}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/change_cluster_32.png}
\caption{}
\label{fig:change_cluster_32}
\end{subfigure}
\caption{Path visualization of gradually changed feature on 2D embedded space using parametric t-SNE: (a) changing the 1\textsuperscript{st} dimension, (b) changing the 8\textsuperscript{th} dimension, (c) changing the 22\textsuperscript{nd} dimension, and (d) changing the 32\textsuperscript{nd} dimension}
\label{fig:change_cluster}
\end{figure}
\subsection{Synthesizing Novel Airfoils}
\noindent
To make use of the encoder-decoder architecture to synthesize novel airfoils, we conduct experiments on interpolation and extrapolation of feature vectors obtained from the UIUC Database airfoils as well as sampling from random Gaussian noises following the method introduced in Section \RNum{3} C.
As given in Eq.~\ref{eq:inter_2}, the affine combination of two feature vectors, $z_1$ and $z_2$, are computed with $\nu=0.5$. Fig~\ref{fig:inter} shows the interpolated airfoils from two different clusters. The labels under each airfoil indicate which two clusters are $z_1$ and $z_2$ come from. As illustrated in Fig~\ref{fig:inter}, the interpolated airfoil inherits features from both clusters. For instance, Cluster 6 and Cluster 11 both represent symmetric airfoil but with variant heights. The interpolation between these two clusters synthesizes a symmetric airfoil with a medium height, as shown in the last airfoil of Fig~\ref{fig:inter}. Also, Clusters 3 and 4 both encode thin airfoils. However, the lower boundary is concave in Cluster 3, while Cluster 4 represents a convex lower boundary making the airfoil symmetric in shape. The interpolation between these two generates a thin airfoil with a flat lower boundary, which is a combination of concave and convex curves. By interpolation, novel airfoils with features from different clusters can be generated. Extrapolation between airfoils from different clusters is conducted as well. Following Eq.~\ref{eq:inter_2}, two feature vectors, $z_1$ and $z_2$, are encoded from two different airfoils, and coefficient $\nu$ is set to be $2$. Fig~\ref{fig:extra} shows the generated results from the extrapolation. Similar to interpolation, extrapolated airfoils inherit features from $z_1$ and $z_2$.
\begin{figure}[hbt!]
\centering
\includegraphics[width=\textwidth]{figures/new_inter.png}
\caption{Interpolation of airfoils from different clusters}
\label{fig:inter}
\end{figure}
\begin{figure}[hbt!]
\centering
\includegraphics[width=\textwidth]{figures/new_extra.png}
\caption{Extrapolation of airfoils from different clusters}
\label{fig:extra}
\end{figure}
Besides interpolation and extrapolation, the performance of the sampling synthesis method is also estimated. In sampling, a Gaussian noise, $\hat{z} \sim \mathcal{N}(0,\mathbf{I})$, is directly fed into the decoder to generate novel airfoils. Shown in Fig.~\ref{fig:VAEGAN_sample} are airfoils synthesized by our VAEGAN model through sampling. Besides, sampled airfoils using PCA and VAE are also included in Fig.~\ref{fig:PCA_sample} and Fig.~\ref{fig:VAE_sample}. PCA, as a linear projection technique, fails to synthesize realistic airfoils through random sampling. Both deep-learning-based generative models, VAE and VAEGAN, can synthesize different smooth airfoil shapes. To quantitatively measure the synthesized airfoils from VAE and VAEGAN, we here introduce Fréchet Inception Distance (FID) \cite{heusel2017gans}, which is used to evaluate the quality of samples from deep learning-based generative models. FID is calculated by computing the Fréchet distance between two feature representations. Generally, lower FID indicates higher generative sample quality. In our case, we feed the synthesized airfoils and the UIUC airfoils into the well-trained discriminator from VAEGAN and extract the second hidden layer as the representation. The FID for VAE and VAEGAN are 1.38788 and 0.65366, respectively, meaning VAEGAN synthesizes more realistic airfoils. Also, airfoils synthesized via VAEGAN possess more novelty while maintains the general geometric pattern of airfoils. For instance, in Fig.~\ref{fig:VAEGAN_sample}, the first airfoils in the fourth row and the third one in the second row are different from existing samples in the UIUC database. Though such novelty does not guarantee better aerodynamic properties, some airfoils are likely to have negative lift coefficients, which are infeasible in practice. The VAEGAN-based model can synthesize a wide variety of airfoils that serve as candidates for further optimization through CFD simulation as we will investigate in Section \RNum{4} F.
\begin{figure}[hbt!]
\centering
\begin{subfigure}{0.3\textwidth}
\centering
\setlength{\fboxsep}{0.3pt}
\setlength{\fboxrule}{0.3pt}
\fbox{\includegraphics[width=\linewidth]{figures/PCA_sample.png}}
\caption{PCA}
\label{fig:PCA_sample}
\end{subfigure}
\hfill
\begin{subfigure}{0.3\textwidth}
\centering
\setlength{\fboxsep}{0.3pt}
\setlength{\fboxrule}{0.3pt}
\fbox{\includegraphics[width=\linewidth]{figures/VAE_sample.png}}
\caption{VAE}
\label{fig:VAE_sample}
\end{subfigure}
\hfill
\begin{subfigure}{0.3\textwidth}
\centering
\setlength{\fboxsep}{0.3pt}
\setlength{\fboxrule}{0.3pt}
\fbox{\includegraphics[width=\linewidth]{figures/VAEGAN_sample.png}}
\caption{VAEGAN}
\label{fig:VAEGAN_sample}
\end{subfigure}
\caption{Generated airfoils by sampling with different featurization techniques.}
\label{fig:sample}
\end{figure}
\subsection{Aerodynamic Properties of Synthesized Airfoil}
\begin{figure}[hbt!]
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/inter_aerodyn.png}
\caption{}
\label{fig:aerodyn_inter}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/sample_aerodyn.png}
\caption{}
\label{fig:aerodyn_sample}
\end{subfigure}
\caption{Lift coefficient $C_l$ v.s. drag coefficient $C_d$ of: (a) interpolated/extrapolated airfoils, and (b) sampled airfoils}
\label{fig:aerodyn}
\end{figure}
\noindent
Aerodynamic properties of the synthesized airfoils are also tested since airfoils have to meet certain aerodynamic properties to make the design feasible and effective. XFoil\footnote{\url{https://web.mit.edu/drela/Public/web/xfoil/}} is utilized to compute the lift coefficient, $C_l$, and the drag coefficient, $C_d$. The experiments on XFoil is set for a low speed condition: Reynolds number $Re=2 \times 10^6$, Mach number $Ma=0.02$, and attack angle $\alpha=\ang{0}$. As illustrated in Fig~\ref{fig:aerodyn_inter}, $C_l$ and $C_d$ are tested on three airfoils: NACA1412, NACA2424 and NACA4415 from the UIUC database. Following Eq.~\ref{eq:inter_3}, triplet interpolation/extrapolation is conducted with feature vectors, $z_1$, $z_2$, and $z_3$, encoded from the three NACA airfoils. The interpolated airfoils, marked by green dots, possess $C_l$ and $C_d$ in between the three NACA airfoils. While extrapolated airfoils, marked by black crossings, have significantly different aerodynamic properties from the interpolated airfoils. Some airfoils synthesized by extrapolation have high $C_l$ with relatively low $C_d$, located at the upper part of Fig.~\ref{fig:aerodyn_inter}. This demonstrates that by interpolation/extrapolation in the feature domain, novel airfoils with promising aerodynamic properties can be synthesized. Also, aerodynamic properties of synthesized airfoils by sampling are tested in comparison with some airfoils from the UIUC database, as shown in Fig~\ref{fig:aerodyn_sample}. Though the generated airfoils from Gaussian noises are not guaranteed to have good aerodynamic properties, like red dots lying on the bottom left. Some promising airfoils can be synthesized, as shown on the top right, with a high lift coefficient and a low drag coefficient. By sampling, the VAEGAN-based model can synthesize airfoils with a wide variety of aerodynamic properties. Such a variety provides abundant candidates to explore in design space for airfoil shape optimization.
\subsection{Shape Optimization on Aerodynamic Properties via Genetic Algorithm}
\begin{figure}[hbt!]
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/avg_score_new.png}
\caption{}
\label{fig:avg_score}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/genetic_alg_new.png}
\caption{}
\label{fig:genetic_alg_clcd}
\end{subfigure}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/airfoil_evolve.png}
\caption{}
\label{fig:airfoil_evole}
\end{subfigure}
\caption{Shape optimization via genetic algorithm: (a) average score of each generation; (b) $C_l$ and $C_d$ of synthesized airfoils for different generations; (c) synthesized airfoil geometry for different generation}
\label{fig:genetic_alg}
\end{figure}
\noindent
The VAEGAN model has been proven to be able to parameterize existing airfoils to latent feature vectors and synthesize novel airfoils automatically. However, whether or not the learned features and synthesized airfoils can be optimized to possess desired aerodynamic properties remains untested. To this end, this section demonstrates that with the VAEGAN model, airfoil shapes can be optimized to realize the target $C_l$ and $C_d$ value via a genetic algorithm (GA). The lift and drag coefficients, $C_l$ and $C_d$, calculated are under the same condition as Section \RNum{4} E. As illustrated in Fig.~\ref{fig:genetic_alg}, the target lift coefficient is $C_l^t=0.6$, and target drag coefficient is $C_d^t=0.06$. In our case, the total number of generations, $N$, is set to be 60, and the number of populations on each generation, $M$, to be 25. Fig.~\ref{fig:avg_score} and Fig.~\ref{fig:genetic_alg_clcd} show how the average score and $C_l,C_d$ change with generation, respectively. As Fig.~\ref{fig:airfoil_evole} illustrates, the airfoil shape gradually evolves to the target $C_l$ and $C_d$. Also, we compare the performance of airfoil optimization using different featurization techniques, PCA, and VAE. The genetic algorithm with the same objective function and settings are conducted. The lift coefficient $C_l$, drag coefficient $C_d$, and fitness score of the last generation in the genetic algorithm is reported in Table~\ref{tab:GA}. The VAEGAN-synthesized airfoils reach an averaged lift coefficient of 0.5857 and an averaged drag coefficient of 0.0061. The coefficients are close to the desired aerodynamic properties. Whereas PCA and VAE fail to synthesize desired airfoils within the same number of generations and population size. This is because our VAEGAN-based model generates a wider variety of airfoils that serves as potential candidates in design optimization.
Such experiments prove that a simple optimization technique like GA and the well-trained VAEGAN model can synthesize airfoils with desired aerodynamic properties, which can guide designing effective and efficient aerodynamic products.
\begin{table}[htb!]
\caption{\label{tab:GA} Airfoil design optimization results with different featurization techniques}
\centering
\begin{tabular}{lcccccccc}
\hline
& \multicolumn{2}{c}{$C_l$} & & \multicolumn{2}{c}{$C_d$} & & \multicolumn{2}{c}{Fitness score} \\
\cline{2-3} \cline{5-6} \cline{8-9}
Featurization & mean & std & & mean & std & & mean & std \\ \hline
PCA & 0.41953 & 0.012672 & & 0.0069430 & 6.9231 $\times 10^{-4}$ & & -0.12893 & 0.046250 \\
VAE & 0.52746 & 0.0017073 & & 0.0056662 & 1.4631 $\times 10^{-5}$ & & -0.017819 & 0.0014068 \\
VAEGAN & 0.58570 & 0.0028946 & & 0.0061030 & 7.0711 $\times 10^{-6}$ & & -8.5312 $\times 10^{-4}$ & 0.00089211 \\
\hline
\end{tabular}
\end{table}
\section{Conclusion}
\noindent
In this work, a data-driven method is proposed to achieve three goals: (1) automatically featuring airfoil geometries from the UIUC Database \cite{uiucdataset} without manually designed parameters, (2) synthesizing novel airfoils by either interpolating or extrapolating the encoded features, as well as generating from random noise, and (3) optimizing the features to synthesize airfoils with desired aerodynamic properties. Our model is built upon VAEGAN, which combines the encoder-decoder architecture from VAE \cite{kingma2013auto} \cite{rezende2014stochastic} and the discriminator from GAN \cite{goodfellow2014gan}. With the encoder-decoder structure, our model learns explicit mappings from airfoil coordinates to latent feature domain as well as from feature vectors to airfoils, while with the discriminator, the model can automatically synthesize realistic samples. Also, our model is trained in a self-supervised manner. Namely, the model learns compact and informative features directly from airfoil shapes without manually tagged labels or designed parameters. Optimized on the learned feature domain via GA, the synthesized airfoils can evolve to have desired aerodynamic properties.
Experiments show that our model learns compact and comprehensive features encoding shape information of airfoils and can automatically generate novel airfoils. First, airfoils can be reconstructed via decoding the learned features with minor error compared to the origin coordinates. Second, K-Means \cite{jain2010kmeans} clustering on the feature domain of the UIUC Coordinate Database further demonstrates the learned representations are meaningful in a way that the centroid of each cluster represents different shapes. It is also investigated what is encoded in each dimension of the feature domain by gradually changing the feature vector on one specific dimension with all other dimensions fixed. Without human prior, each dimension encodes different geometric information like height, camber, symmetry, and even coupled features. Moreover, novel airfoils are synthesized by interpolating and extrapolating learned features from different airfoils as well as directly generated from random noise. By interpolating or extrapolating, the synthesized airfoil inherits and blends features from existing airfoils, which provides insights for designing new airfoils. On the other hand, airfoils generated from Gaussian noise are more aggressive in a way that they follow a less geometrical format of existing airfoils, and more novelty is introduced to the airfoil design. Finally, the synthesized airfoils can be optimized via GA to possess competitive or even better aerodynamic properties in comparison to existing ones, indicating the synthesized geometries are not only plausible in shape but also practical in aerodynamic performance.
\section*{Nomenclature}
{\renewcommand\arraystretch{1.0}
\noindent\begin{longtable*}{@{}l @{\quad=\quad} l@{}}
$z$ & latent feature vectors of airfoils \\
$\hat{z}$ & random noise sampled from Gaussian distribution \\
$x$ & airfoils \\
$\tilde{x}$ & airfoils reconstructed from $z$ \\
$\hat{x}$ & airfoils synthesized from $\hat{z}$ \\
$\mathcal{L}_{prior}$ & prior loss \\
$\mathcal{L}_{recon}$ & reconstruction loss \\
$\mathcal{L}_{layer}$ & layer like loss \\
$\mathcal{L}_{GAN}$ & adversarial GAN loss \\
$C_l$ & lift coefficient \\
$C_d$ & drag coefficient \\
\end{longtable*}}
\section{Introduction}
\noindent
\lettrine{R}{ecent} years have witnessed the success of deep learning \cite{lecun2015deep} in many fields like computer vision \cite{voulodimos2018deep}, natural language process \cite{cambria2014jumping} and robotics \cite{pierson2017deeprobot} \cite{carrio2017review}. Such data-driven methods can automatically learn compact and comprehensive representations from samples. However, most of the prevalent deep learning models are based on supervised learning, meaning the samples are paired with manually tagged labels. Such supervision makes the model hard to generalize since both the amount of labeled data and the information contained in the label are limited. Hence, self-supervised learning is proposed to learn the features directly from data. Variational Autoencoder (VAE) \cite{kingma2013auto} \cite{rezende2014stochastic} follows the insight via an encoder-decoder structure, where the encoder down-samples high-dimensional input into a latent feature domain while the decoder reconstructs the sample from learned low-dimensional feature. By minimizing the difference between the reconstructed sample and the original one, VAE automatically learns features without any labels. Generative Adversarial Network (GAN) \cite{goodfellow2014gan} pushes learning from self-supervision even further via a min-max game between a generator and a discriminator. The discriminator works as a classifier to determine real samples from synthesized fake ones; meanwhile, the generator which synthesizes samples from random noise is intended to cheat the discriminator. By jointly training the two components, GAN can generate super high-quality realistic samples \cite{karras2019style}, which VAE fails to achieve. Through learning to reconstruct or synthesizing samples, the self-supervised models automatically encode high-dimensional input into informative features, which can be generalized to different tasks without restrictions from human labels.
Such a self-supervised learning manner can help the parameterization of various geometries using the learned representations. Geometry parameterization plays an important role in shape design \cite{samareh2001survey,chang20113d,salunke2014airfoil}, and geometry heavily influences the performance, especially in the design of aerodynamic products like airfoils. A practical and effective airfoil design must meet certain aerodynamic requirements, like lift, drag, pitching moment, and critical-speed characteristics \cite{abbott1945summary}.
However, due to the curse of dimensionality, design optimization based on CFD simulations can be hard or even infeasible on the airfoil geometry domain, especially for gradient-free methods \cite{rajnarayan2008multifidelity}. Therefore, parameterization or dimension-reduction is required to define the design domain before any design optimization is applied.
Traditional parameterization or dimension-reduction techniques rely on manually selected design parameters like control points of B\'ezier curves \cite{sederberg1992bezier} or B-splines \cite{derksen2010bezierparsec}, which places restrictions on the generalization to various shapes as well as synthesizing novel geometries. Implementation of self-supervised deep learning methods on shape parameterization, like VAE and GAN, can overcome the limitations of traditional techniques and synthesize shapes with great novelty, which can provide insights for future geometric design.
In this work, VAEGAN \cite{larsen2015vaegan}, which is a combination of Variational Autoencoder (VAE) and Generative Adversarial Network (GAN), is utilized to learn feature vectors and generate novel airfoils without any human prior. VAEGAN takes advantage of both VAE and GAN. With the encoder-decoder architecture from VAE, the model learns to explicitly encode an existing airfoil shape into a low-dimensional feature domain and reconstruct the shape with little error. With the discriminator from GAN, our model can generate a large number of high-quality novel airfoils from random noise with no manually designed parameters.
Our experiments show that generated airfoils are smooth even without any smoothing post-process. K-means clustering in the learned feature domain demonstrates that feature vectors encode essential shape information in a way that each cluster represents various shape patterns. Further test on learned latent features illustrates that different geometry information is encoded in each dimension of the representation. The performance of the model in synthesizing novel airfoils is examined as well. By either interpolation or extrapolation of feature vectors, a synthesized airfoil inherits features of parent samples, while generated airfoils from sampled Gaussian noise show great novelty in a way that it is not a simple combination of two existing airfoils. A further experiment on the aerodynamic properties of synthesized airfoils, either by interpolating, extrapolating, or sampling, indicates the synthesized airfoils can possess competitive aerodynamic properties, and some even surpass the existing ones. With a genetic algorithm \cite{mitchell1998introduction}, airfoil geometries can be optimized on the feature domain and evolve to possess specific aerodynamic properties. Our model proves its ability to parameterize the existing airfoil shape as well as generating novel and practical airfoils; both lead to designing the new generation of airfoils more intelligently and efficiently without intensely relying on experimental experience or manually design parameters.
\section{Related Works}
\noindent
We propose to use a generative model to synthesize novel airfoils without any predefined design parameters. Through training via the gradient-based method, our deep learning model also automatically learns to parameterize airfoils into latent feature vectors. To better address the insight of our work, this section reviews previous work on shape parameterization, especially on aerodynamic geometries and the implementation of deep learning on geometry design and synthesis.
\subsection{Geometry Parameterization and Dimension Reduction}
\noindent
A lot of work has been done in parameterizing complex shapes and reducing geometric dimensions. Some research focuses on analytically expressing the curves. In \cite{hicks1978wing}, it is introduced that by adding analytic shape functions to the baseline shape, a compact formulation for parameterization can be obtained. The design variables, in this case, are the coefficients correlated with shape functions. By this means, the analytical function of the airfoil curve is formulated. Also, \cite{hager1992multi} follows the same formulation strategy but with different shape functions. Another common and efficient method is PARSEC \cite{sobieczky1997parsec}, which defines eleven geometric parameters to thoroughly express the airfoil shape, including upper and lower curvature, thickness, leading-edge radius, etc. With the defined parameters, a linear combination of shape functions is introduced to describe the airfoil shape as well. Such methods work well for a specific set of curves but may fail to express complex geometries since they rely on manually designed parameters.
Polynomial and spline are also utilized to help dimension reduction. With different orders of polynomials as the basis, the airfoil shape can be described as a linear combination of the basis \cite{elliott1997practical,taylor1991sensitivity}. However, high-order terms can overfit to high-frequency noise, especially when coefficients are of different magnitudes. Besides polynomial, B\'ezier curve \cite{sederberg1992bezier}, which is built upon the Bernstein polynomials, is another mathematical formulation of curves. In detail, $n+1$ control points of B\'ezier are needed to define an $n$-degree B\'ezier curve. Although B\'ezier and polynomial curves are mathematically equivalent, B\'ezier usually perform better in controlling a curve since control points are closely related to the curve position and shape. To mitigate the rounding error, De Casteljau \cite{boehm1999casteljau}, a recursive algorithm, is introduced to compute the Bernstein polynomials numerically. Besides, B\'ezier-PARSEC \cite{derksen2010bezierparsec}, which combines B\'ezier and PARSEC, uses PARSEC parameters to define B\'ezier curves. The B-spline curve with B-spline basis functions is also utilized to describe the airfoil shape. Yet B-spline formulation fails to represent implicit conic sections accurately. That is why Non-uniform rational B-spline (NURBS) is introduced \cite{farin2014curves}. NURBS can accurately represent both standard geometric objects like lines, circles, ellipses, and cones, as well as free‑form geometry, which are prevalent in industrial design.
Another popular technique in shape parameterization is free-form deformation (FFD), which utilizes high-level shape deformation instead of lower-level geometric entities to represent a shape. Based on this insight, \cite{sederberg1986free} presents a technique which can apply deformation either locally or globally based on trivariate Bernstein polynomials. Deformation is manipulated by control points of trivariate B\'ezier volumes. In \cite{coquillart1990extended}, an extended free-form deformation (EFFD) method is presented, which allows arbitrarily shaped deformations by using non-parallelepiped lattices. Research presented in \cite{yeh1998applying} incorporates FFD and sensitivity analysis where geometry changes and structural responses are correlated, and the shape satisfying deformation or stress constraints can be found easily.
Other methods like leveraging camber and thickness mode shapes derived from existing airfoils are also used to parameterize the airfoil shapes \cite{li2019data}. Also, the linear reduction method like the SVD is utilized to extract airfoil representations and optimize shape design \cite{poole2019efficient}. The conventional dimension reduction or parameterization techniques have been implemented in different scenarios and successfully represented the existing airfoil shapes. However, these methods usually require pre-defining the design space as well as the boundary of design space, like the design parameters in NACA, shape functions, etc., which can degrade the synthesis of novel/new airfoils and optimization towards the desired design.”
\subsection{Deep Learning in Geometry Design}
\noindent
In recent years, deep learning has been a great success in extracting informative features from data \cite{lecun2015deep}. Especially in a supervised visual learning manner, with the invention of convolutional neural network (CNN) \cite{lecun1989lenet,lecun1998gradient}, deep learning has been general solutions in many fields, like image classification \cite{szegedy2015going} \cite{he2015delving}, object detection \cite{ren2015fasterrcnn,redmon2016yolo}, and segmentation \cite{li2017fully,he2017maskrcnn}. Also, with the introduction to GAN \cite{goodfellow2014gan}, deep learning in a self-supervised manner has been widely used in synthesizing realistic samples, and some sophisticated GAN-based model can generate high-resolution images which are even hard to distinguish by humans \cite{karras2019style}. With the ability to extract representative features and synthesize realistic samples, deep learning has been widely used to make geometric design more systematic and efficient. This section will introduce some work of deep learning in geometric design, especially in aerodynamic shape design.
In the work of \cite{norgaard1997neural}, a multi-layer perceptron (MLP), which takes the angle of attack and flap setting as input, is trained to predict multiple aerodynamic coefficients, including lift coefficient, drag coefficient, and moment of inertia. Further, \cite{zhang2018application} utilizes CNN with airfoil images as input to learn the lift coefficients of different airfoils in multiple flow Mach numbers, Reynolds numbers, and diverse angles of attack. Similarly, in \cite{yilmaz2017convolutional}, CNN is implemented to predict the pressure coefficient value at the test point. Moreover, \cite{yilmaz2018deep} presents a CNN-based method to learn the correlation between airfoil geometry and pressure distribution. The model can also conduct an inverse airfoil design given the pressure.
On the other hand, the GAN model \cite{goodfellow2014gan}, as a self-supervised learning framework, has been a prevalent and great success in generating realistic samples.
GAN, with a discriminator and a generator, is designed to learn features and generate samples without manually tagged labels. The discriminator is a classifier telling the true input from the synthesized fake one, while the generator is intended to generate plausible samples to cheat the discriminator. Since first introduced \cite{goodfellow2014gan}, many pieces of research have been dedicated to pushing the edge of GAN. In Wasserstein GAN \cite{arjovsky2017wasserstein}, by introducing Wasserstein distance, the quality of generated samples can be well measured during training. ConditionalGAN \cite{mirza2014conditional} and InfoGAN \cite{chen2016infogan} extend GAN to generate a sample from various categories within one model. Also, DCGAN \cite{radford2015unsupervised} and StyleGAN \cite{karras2019style} can generate high-quality realistic samples with sophisticated architectures.
With the ability to generate plausible samples, GAN provides a powerful architecture to learn representations that can help shape design as well. Based on this insight, \cite{chen2018beziergan,chen2019aerodynamic} introduce B\'ezierGAN, which uses a GAN to generate B\'ezier curve control points and then uses the control points to formulate the boundary of airfoils. Such pipeline guarantees generated airfoils to be smooth, and further shape optimization can be conducted on the feature domain \cite{chen2019aerodynamic}. Such a method is, however, restricted to B\'ezier curves and fails to encode existing airfoil shapes explicitly. Further, GAN can be implemented in generating three-dimensional samples, as shown in \cite{wu20163dgan,huang2015analysis,sinha2017surfnet}. \cite{wu20163dgan} proposes to use a three-dimensional convolutional layer to generate volumetric objects. While \cite{huang2015analysis} trains a generative model of three-dimensional shape surfaces, which directly encodes surface geometry and shape structure, \cite{kalogerakis2012probabilistic} further proposes a model to represent probabilistic relationships between properties of shape components and relates them to learned underlying causes of structural variability within the domain.
In our work, we propose to use a VAEGAN-based model \cite{larsen2015vaegan} to extract features of airfoil shapes and synthesize new airfoil designs. The learned latent features encode the airfoil shape and can be utilized to synthesize new designs through interpolation or extrapolation. In comparison to conventional parameterization methods, our model can generate a wider variety of new airfoils, and some show promising aerodynamic properties. By applying the genetic algorithm, the VAEGAN synthesized airfoils can be optimized to desired aerodynamic properties.
\section{Proposed Method}
\subsection{Data Pre-processing}
\noindent
The UIUC Coordinates Database \cite{uiucdataset}, which contains more than 1,600 2-dimensional airfoils, is used to train the generative model. Each airfoil in the database is represented by varying numbers of points with x, y coordinates. Such variation restricts data to be fed directly into a neural network model, which requires a homogeneous input. To deal with it, we first scale the x coordinates of all airfoils to $[0,1]$. Then all airfoils are interpolated by splines, and $N$ points are selected with $x$ given in Eq.~\ref{eq:x}, where $i$ represents the index of each point.
\begin{equation}
\begin{split}
\theta_i &= \frac{\pi (i-1)}{N} \\
x_i &= 1 - \cos(\theta_i)
\end{split}
\label{eq:x}
\end{equation}
In our case, $N$ is set to be 200 with 100 points from the upper boundary and the other 100 from the lower boundary. By this means, all the interpolated airfoils share the same x coordinates. Therefore, only the y coordinates of each airfoil are fed into the model, which reduces the dimensionality of the data. Finally, all y coordinates are scaled to $[-1, 1]$ by multiplying a normalization coefficient. As illustrated in Fig~\ref{fig:preprocess}, the first row shows the original airfoils from the UIUC Coordinate Database, while the second row shows the corresponding processed airfoils.
\begin{figure}[hbt!]
\centering
\includegraphics[width=.8\textwidth]{figures/new_preprocessing.png}
\caption{Data pre-processing: the first row shows the origin airfoil coordinates from the UIUC database, the second row shows the corresponding processed airfoils}
\label{fig:preprocess}
\end{figure}
\subsection{VAEGAN}
\noindent
Our model is based upon VAEGAN \cite{larsen2015vaegan}, which takes advantage of both VAE \cite{kingma2013auto} \cite{rezende2014stochastic} and GAN \cite{goodfellow2014gan}. VAE contains two components: an encoder and a decoder. The former encodes a high-dimensional sample, $x$, into a low-dimensional latent representation, $z$. While the decoder takes as input the latent vector, $z$, and upsamples from the representation domain to the original data domain, $\tilde{x}$. The encoder and decoder are given as:
\begin{equation}
z \sim \text{Enc}(x) = q(z|x), \: \tilde{x} \sim \text{Dec}(z) = p(\tilde{x}|z).
\label{eq:vae}
\end{equation}
To regularize the encoder, VAE takes into consideration a prior distribution of the latent vector, $p(z)$. Here it is assumed that $z \sim \mathcal{N}(0,\mathbf{I})$, which follows an isotropic Gaussian distribution. The loss function for VAE to minimize is given by:
\begin{equation}
\mathcal{L}_{VAE} = \mathcal{L}_{recon} + \mathcal{L}_{prior},
\label{eq:loss_vae}
\end{equation} with
\begin{equation}
\begin{aligned}
\mathcal{L}_{recon} &= ||\tilde{x} - x||_2^2, \text{ and} \\
\mathcal{L}_{prior} &= D_{KL}(q(z|x)||p(z)),
\label{eq:loss_recon_prior}
\end{aligned}
\end{equation}
where $\mathcal{L}_{recon}$ measures how well the reconstructed data, $\tilde{x}$, is comparing to the original $x$ by the mean square error (MSE), and $\mathcal{L}_{prior}$ is the Kullback Leibler divergence (KL divergence), which measures the difference between encoded representation vectors and Gaussian distribution.
VAE learns the representation of samples and can reconstruct them from $z$ with the encoder-decoder architecture. However, it suffers from poor performance in generating novel samples, which have not been seen before \cite{goodfellow2014gan} \cite{larsen2015vaegan}. To this end, a generative adversarial network (GAN) \cite{goodfellow2014gan} is introduced, which contains a discriminator, $\mathcal{D}$, and a generator, $\mathcal{G}$, competing with each other in a self-supervised manner. $\mathcal{G}$ tries to generate plausible samples to fool $\mathcal{D}$, while $\mathcal{D}$ keeps sharpening its decision boundary to determine synthesized fake samples from real ones. In detail, the generator, $\mathcal{G}$, is fed with random noise $\hat{z} \sim p(\hat{z})$ and maps the noise to the data sample domain to generate fake $\hat{x}$. The discriminator takes both $x$ and $\hat{x}$ to predict whether the input is from the real dataset or generated by $\mathcal{G}$. The cross-entropy loss function for the min-max game is given as:
\begin{equation}
\min_G \max_D \mathcal{L}_{GAN} =\log \mathcal{D}(x) + \log(1-\mathcal{D}(\mathcal{G}(\hat{z}))).
\label{eq:loss_gan}
\end{equation}
\begin{figure}[hbt!]
\centering
\includegraphics[width=.7\textwidth]{figures/vaegan.png}
\caption{VAEGAN combines the encoder-decoder structure from VAE and discriminator, $\mathcal{D}$, from GAN}
\label{fig:vaegan}
\end{figure}
As one may expect, GAN generates samples purely from random noise, making it hard to obtain an explicit mapping from the data domain to the feature domain. Therefore, we build our model upon VAEGAN, which combines VAE and GAN. Namely, the generator is replaced by an encoder-decoder structure from VAE, as shown in Figure \ref{fig:vaegan}. Notice that in VAEGAN, the model generates a reconstructed sample, $\tilde{x}$, given a real sample, $x$, and meanwhile generates a fake sample, $\hat{x}$, directly from noise, $\hat{z}$. Both $\tilde{x}$ and $\hat{x}$ should be classified as fake by the discriminator, $\mathcal{D}$, and only $x$ is recognized as the real sample. Hence GAN loss function $\mathcal{L}_{GAN}$ from Eq.~\ref{eq:loss_gan} is modified to:
\begin{equation}
\mathcal{L}_{GAN} = \log (\mathcal{D}(x)) + \log(1-\mathcal{D}(\text{Dec}(z))) + \log (1 - \mathcal{D}(\text{Dec}(\text{Enc}(x)))),
\end{equation}
which takes into consideration the real sample, $x$, reconstructed sample, $\tilde{x}$, and fake sample, $\hat{x}$. Besides, to stabilize the training process and sharpen the decision boundary of $\mathcal{D}$, another loss function, $\mathcal{L}_{layer}$, is introduced when training the encoder and decoder. $\mathcal{L}_{layer}$, as given in Eq.~\ref{eq:loss_layer}, measures the $l_1$ distance between the values of the neurons in one particular layer of $\mathcal{D}$ when fake samples are fed and the values when real samples are fed.
\begin{equation}
\mathcal{L}_{layer} = ||\mathcal{D}_l (x) - \mathcal{D}_l (\text{Dec}(\hat{z}))||_1.
\label{eq:loss_layer}
\end{equation} The complete loss function for VAEGAN is a weighted combination of all the loss terms given by:
\begin{equation}
\mathcal{L} = \lambda_0 \mathcal{L}_{prior} + \lambda_1 \mathcal{L}_{recon} + \lambda_2 \mathcal{L}_{layer} + \lambda_3 \mathcal{L}_{GAN}.
\label{eq:loss_vaegan}
\end{equation} The three components, encoder, decoder and discriminator are trained jointly, and each term of the loss function is assigned with different weights $\lambda$ when training each component.
\subsection{Airfoil Synthesis}
\noindent
By training on the UIUC Database, the VAEGAN model automatically learns to encode airfoils into latent features and reconstruct airfoils from the feature domain. The learned latent features can be directly utilized for dimension reduction and shape parameterization. Moreover, the VAEGAN model is intended to synthesize novel airfoils which are different from samples in the training dataset. To this end, we propose three synthesis methods: interpolation, extrapolation, and sampling, all of which are conducted on the latent feature domain.
More specifically, in interpolation or extrapolation, two airfoils from the UIUC Database are first mapped to latent feature vectors, $z_1$ and $z_2$, via a well-trained encoder. A new feature vector $\bar{z}$, which is an affine combination of $z_1$ and $z_2$, is calculated as given in Eq.~\ref{eq:inter_2}:
\begin{equation}
\bar{z} = \nu z_1 + (1-\nu) z_2,
\label{eq:inter_2}
\end{equation}
where $\nu$ is the coefficient controlling the weight between $z_1$ and $z_2$. When $0 \leq \nu \leq 1$, $\bar{z}$ is an interpolated feature vector, else it is an extrapolation between $z_1$ and $z_2$. The interpolated/extrapolated feature vector, $\bar{z}$, is then fed into the decoder to synthesize an airfoil. Also, such interpolation/extrapolation between two airfoils can be directly extended to a triplet case. Given $z_1$, $z_2$, and $z_3$ are three feature vectors mapped from three different airfoils via the encoder, the expression of triplet interpolation/extrapolation is shown in Eq.~\ref{eq:inter_3}:
\begin{equation}
\bar{z} = \alpha z_1 + \beta z_2 + \gamma z_3,
\text{ where } \alpha+\beta+\gamma=1.
\label{eq:inter_3}
\end{equation}
Similarly, when $0 \leq \alpha,\beta,\gamma \leq 0$, $\bar{z}$ is an interpolation of the three feature vectors, and an extrapolation otherwise.
Besides interpolation and extrapolation, sampling is another method to synthesize novel airfoils. Unlike interpolation or extrapolation, which relies on feature vectors from existing airfoils, sampling generates airfoils directly from random noise. A feature vector $\hat{z}$ is randomly sampled from an isotropic Gaussian distribution $\mathcal{N}(0,\mathbf{I})$ and then mapped to an airfoil via the decoder. By this means, synthesized airfoils from sampling are less restricted since sampled latent vectors are not constrained by features extracted from airfoils in the UIUC Database and are more likely to introduce novelty to the synthesized shapes.
\subsection{Aerodynamic-aware Shape Optimization}
\noindent
So far, how the VAEGAN model is built and used to generate novel airfoils has been introduced. However, the novelty in shape does not guarantee a better airfoil design. To design engineering effective airfoils, aerodynamic properties are supposed to be considered. To this end, we propose to use a genetic algorithm (GA) \cite{poon1995genetic} \cite{mitchell1998introduction} to optimize airfoil shapes by controlling feature vectors learned from the VAEGAN model so that the airfoils can evolve to have the desired aerodynamic properties. Specifically, lift coefficient, $C_l$, and drag coefficient, $C_d$, which measure the aerodynamic force perpendicular and horizontal to the direction of motion, are considered to evaluate the aerodynamic performance of the synthesized airfoils.
As a non-gradient optimization technique, GA is inspired by natural selection and is intended to force individuals to gradually evolve to the optimal. Assume the GA has $N$ generations in total and $M$ individuals in each generation. In our case, individuals are feature vectors. We use $z_i$ to represent all individuals in the $i$\textsuperscript{th} generation, and $z_{i,j}$ for the $j$\textsuperscript{th} individual in the $i$\textsuperscript{th} generation; also the airfoil decoded from $z_{i,j}$ is annotated as $a_{i,j}$. Similarly, $C_l^{i,j}$ and $C_d^{i,j}$ represents lift and drag coefficients of $a_{i,j}$, respectively. The fitness score, $s_{i,j}$, is used to measure the aerodynamic performance of the individual, $z_{i,j}$, as shown in Eq.~\ref{eq:score}:
\begin{equation}
s_{i,j} = - (\frac{C_l^{i,j} - C_l^t}{C_l^t})^2 - (\frac{C_d^{i,j} - C_d^t}{C_d^t})^2,
\label{eq:score}
\end{equation}
where the square of the difference between the target and current aerodynamic coefficients is calculated and normalized by the squared target $C_l^t$ and $C_d^t$. The fitness score is supposed to approach zero as individuals evolve on each generation. As shown in Algorithm~\ref{alg:ga}, the initial generation, $z_0$, is randomly sampled from an isotropic Gaussian distribution, $\mathcal{N}(0,\mathbf{I})$. The GA starts with the selection from the initial generation by randomly picking two individuals and comparing their fitness scores. The one with a higher fitness score wins the tournament and becomes one of the parents. $p_1^i$ and $p_2^i$ denote all the parents 1 and parents 2 in the $i$\textsuperscript{th} generation respectively. Single-point crossover is then implemented to generate offspring from parents 1 and 2. Namely, a crossover point on the parent vector is randomly selected, and all elements after that point are swapped between the two parents. Mutation in the natural selection process is also imitated with additive Gaussian noises.
\begin{algorithm}
\caption{Aerodynamic-aware Shape Optimization via GA}
\begin{algorithmic}[1]
\Procedure{GA}{$N, M, C_l^t, C_d^t$, $p$}
\State Initialize $i := 0$
\State Sample first generation $z_{0j} \sim \mathcal{N}(0, \mathbf{I})$, for $0<j<M$
\While{$i < N$}
\State Synthesize airfoils $a_{i,j}$ from $z_{i,j}$ via decoder
\State Compute $C_l^{i,j}$, $C_d^{i,j}$ and fitness score $s_{i,j} = - (\frac{C_l^{i,j} - C_l^t}{C_l^t})^2 - (\frac{C_d^{i,j} - C_d^t}{C_d^t})^2$
\State Select $M$ parent 1, $p_1^i$, and $M$ parent 2, $p_2^i$, from $z_i$ by tournament
\State Generate next generation, $z_{i+1}$, through single-point crossover
\State With probability $p$, $z_{i+1,j}$ will add a Gaussian noise $\mathcal{N}(0,\mathbf{I})$
\State $i := i+1$
\EndWhile
\Return The individual with the highest score from $z_{i}$
\EndProcedure
\end{algorithmic}
\label{alg:ga}
\end{algorithm}
\section{Experiments}
\noindent
Our VAEGAN model consists of 3 components: an encoder, a decoder, and a discriminator, which are all built on multi-layer perceptron (MLP). As illustrated in Fig~\ref{fig:vaegan}, the encoder encodes 200-dimensional airfoil coordinates into a 32-dimensional feature domain while the decoder maps the feature back to the airfoil. The discriminator is a classifier examining whether the input is a real airfoil from the UIUC Database, or a fake one reconstructed from the decoder, or synthesized from random noises. In detail, the encoder is modeled by a 3-layer MLP with the number of neurons $[256, 128, 32]$ in each layer, and LeakyReLU \cite{maas2013leakyrelu} is implemented as the activation function in each layer. The decoder is also a 3-layer MLP with the number of neurons $[128, 256, 200]$ in each layer. A hyperbolic tangent (Tanh) function works as the activation function in the output layer to scale all outputs into $[-1,1]$. Similarly, the discriminator contains three layers with the number of neurons $[256, 128, 1]$, and outputs the probability of whether the input is real or fake through a Sigmoid activation function.
To automatically learn the latent features and synthesize airfoils, the VAEGAN model is trained on the UIUC Database for 5000 epochs, and each epoch goes through all the samples in the dataset. Initial learning rates for all three components: encoder, decoder, and discriminator are set to be $0.0005$ and decay to $0.00005$ after 2500 epochs. The batch size is set to be 16, which is approximately $1/100$ of the database size. Adam optimizer \cite{kingma2014adam} is utilized to update all the parameters in the model. As mentioned in Section \RNum{3} C, different coefficients are assigned to each term in the loss function Eq.~\ref{eq:loss_vaegan}; also different components, namely the encoder, decoder, and discriminator, have different coefficients, respectively. Coefficients of different loss terms and components are shown in Eq.~\ref{eq:concrete_loss}:
\begin{equation}
\begin{split}
\mathcal{L}_{Enc} &= 0.1 \mathcal{L}_{prior} + 0.1 \mathcal{L}_{layer} + 10 \mathcal{L}_{recon}, \\
\mathcal{L}_{Dec} &= 0.1 \mathcal{L}_{prior} + 0.1 \mathcal{L}_{layer} + 10 \mathcal{L}_{recon} + 5 \mathcal{L}_{GAN}, \\
\mathcal{L}_{\mathcal{D}} &= \mathcal{L}_{GAN},
\end{split}
\label{eq:concrete_loss}
\end{equation}
where $\mathcal{L}_{Enc}$, $\mathcal{L}_{Dec}$, and $\mathcal{L}_{\mathcal{D}}$ represent loss functions for the encoder, decoder, and discriminator, respectively.
To better investigate our VAEGAN-based model, we compare the performance with two other paramterization methods, principle component analysis (PCA) and variational autoencoder (VAE) \cite{kingma2013auto}. PCA conducts a linear transformation from the pre-processed airfoil point coordinates into prioritized latent variables. In our case, the top 32 dimensions are kept as the feature. The VAE follows the same encoder-decoder architecture as the VAEGAN, but lacks the discriminator. The latent feature dimension is also set to 32, and the loss function is given in Eq.~\ref{eq:VAE_loss}:
\begin{equation}
\mathcal{L}_{Enc} = 0.1 \mathcal{L}_{prior} + 10 \mathcal{L}_{recon}.
\label{eq:VAE_loss}
\end{equation}
\subsection{Airfoil Reconstruction via Encoder-decoder}
\begin{figure}[hbt!]
\centering
\includegraphics[width=\textwidth]{figures/new_recon.png}
\caption{Reconstructed airfoils: the first row shows the airfoils from UIUC Coordinate Database, the second row shows the reconstructed airfoils with VAE, and the third row shows the reconstructed airfoils with smoothness}
\label{fig:recon}
\end{figure}
\noindent
The VAEGAN model can automatically learn feature vectors, namely mapping the high-dimension airfoils into low-dimension representations. To estimate whether or not the feature vector fully encodes the geometric information of the original airfoil,we first feed airfoils from UIUC Coordinate Database into the encoder to obtain the encoded feature vectors. The decoder then takes the vectors as input and outputs the reconstructed airfoils. Also, the Savitzky-Golay filter \cite{schafer2011savitzky}, a moving polynomial fitting, is implemented to smoothen the boundary of reconstructed airfoils. In our case, the second-order polynomial is used in the Savitzky-Golay filter, and the length of the moving window is set to be 7. In Fig~\ref{fig:recon}, the first row illustrates samples from the UIUC Database, and the second row shows reconstructed airfoils from corresponding feature vectors, with an MSE, 3.65345$\times 10^{-4}$, between the reconstructed and original airfoils. This small error indicates the learned features well represent the shape of airfoils. The third row shows reconstructed airfoils with the Savitzky-Golay filter with an MSE, 3.65054$\times 10^{-4}$, comparing to the original airfoils. These results further demonstrate that the encoder-decoder can reconstruct airfoils that are smooth and realistic without smoothing filters. Also, our VAEGAN-based model is compared with PCA and VAE as shown in Table~\ref{tab:MSE}. PCA reaches the lowest MSE since it provides a close form solution, whereas VAE and VAEGAN are optimized numerically via the gradient-based method. With the discriminator and adversarial loss from GAN, VAEGAN model performs slightly better than VAE in reconstruction. It should be pointed out that all the three parameterization methods have small reconstruction MSEs of magnitude $10^{-4}$, meaning all the features extracted well encodes the airfoil shapes from the UIUC database.
\begin{table}[htb!]
\caption{\label{tab:MSE} Mean squared error of airfoil reconstruction via different featurization techniques}
\centering
\begin{tabular}{lccc}
\hline
Featurization & PCA & VAE & VAEGAN \\ \hline
MSE & 1.29208$\times 10^{-4}$ & 3.70018$\times 10^{-4}$ & 3.65345$\times 10^{-4}$ \\
\hline
\end{tabular}
\end{table}
\subsection{Clustering in Feature Domain}
\noindent
The encoded features obtained from the encoder in our method can help better understand the shape of current airfoils. All airfoils from the UIUC Coordinate Database are first mapped to feature vectors, and an unsupervised learning algorithm, K-Means \cite{jain2010kmeans}, is used to cluster these airfoils in the feature domain. To visualize the 32-dimensional feature domain, we use Parametric t-distributed Stochastic Neighbor Embedding (parametric t-SNE) \cite{maaten2008tsne, van2009ptsne} as a visualization tool. Parametric t-SNE is modeled by MLP, which maps the high dimensional feature vector $z_i$ into a low-dimension embedding $y_i$, while keeps the similarity between points. It converts similarities between data points to joint probabilities and minimizing the KL divergence between the joint probabilities of embedding $y_i$ and the original feature vector $z_i$. Eq.~\ref{eq:tsne} shows the cost function $C$, which t-SNE is expected to minimize:
\begin{equation}
\begin{aligned}
p_{j|i} = \frac{\exp(-\|z_i-z_j\|^2/2\sigma_i^2)}{\sum_{k \neq i} \exp (-\|z_i-z_k\|^2/2\sigma_i^2) }, & \;
q_{j|i} = \frac{\exp(-\|y_i-y_j\|^2/2\sigma_i^2)}{\sum_{k \neq i} \exp (-\|y_i-y_k\|^2/2\sigma_i^2) }, \\
C = KL(P || Q) = & \sum_i \sum_j p_{j|i} \log \frac{p_{j|i}}{q_{j|i}},
\end{aligned}
\label{eq:tsne}
\end{equation}
where $\sigma_i$ is calculated by a binary search given a fixed perplexity that is specified by the user \cite{maaten2008tsne}. Fig.~\ref{fig:cluster} shows the K-means clustering results visualized with parametric t-SNE, where different colors represent different clusters, and the centroid of each cluster is also shown. The centroid of each cluster is different from each other in symmetry, height, camber, etc., and each represents the geometric pattern of each cluster. Features from close clusters represent similar airfoil shapes. The distance between feature points intuitively reflects the difference between the two airfoil shapes. This indicates that our VAEGAN-based model learns features that maintain the similarity of input airfoil shapes.
\begin{figure}[hbt!]
\centering
\includegraphics[width=\textwidth]{figures/cluster_airfoils.png}
\caption{Parametric t-SNE visualization for clustering on feature domain}
\label{fig:cluster}
\end{figure}
\subsection{What is Encoded in the Feature Domain}
\noindent
Also, experiments are conducted to investigate what geometric features are encoded in each dimension of the learned representation. A series of manually designed feature vectors are fed into the decoder, where all the elements are set to zero except for one specific dimension. That particular element is changed gradually from $-10$ to $10$, and the designed feature vectors are mapped to the airfoil coordinate domain by the decoder. Changes of generated airfoils are illustrated in Fig~\ref{fig:change}, and the 2D embedding of feature vectors using parametric t-SNE is shown in Fig~\ref{fig:change_cluster}. Here, only four dimensions are chosen for analysis purposes.
The 1\textsuperscript{st} dimension encodes the height of the upper boundary, as illustrated in Fig~\ref{fig:change_1}. As the 1\textsuperscript{st} dimension increasing from $-10$ to $10$, the height of the front half airfoil increases while the tail becomes thinner. The 8\textsuperscript{th} dimension encodes the camber of both the upper boundary and the lower boundary. It is shown that by tuning the 8\textsuperscript{th} dimension, the upper boundary of the airfoil changes from a concave curve to a horizontal straight line, while the lower boundary evolves from a concave to a convex curve. Interestingly, the 22\textsuperscript{nd} dimension encodes quite similar representations as the 1\textsuperscript{st} dimension while in the opposite direction. In other words, generated airfoils from feature vectors whose 22\textsuperscript{nd} dimension change from $-10$ to $10$ are like those with the 1\textsuperscript{st} dimension change from $10$ to $-10$ as illustrated in Fig~\ref{fig:change_cluster_1} and Fig~\ref{fig:change_cluster_22}. Besides, Fig~\ref{fig:change_32} shows how the last dimension is connected to the camber of the lower boundary. In detail, the curvature of the lower boundary decreases as the 32\textsuperscript{nd} dimension increases.
In comparison to the features learned by our VAEGAN-based model, Fig.~\ref{fig:vae_change} shows the VAE-synthesized airfoils when changing only one feature dimension. As shown in ~\ref{fig:vae_change_0}, the first dimension of VAE features fails to encode any shape representations. Even in the dimensions where representations are learned as dimension 3 and 31 shown in Fig.~\ref{fig:vae_change_2} and Fig.~\ref{fig:vae_change_30}, the feature does not change continuously as we observe in the VAEGAN results. Our VAEGAN-based model learns more representative and thorough features than the VAE model. Also, the representations in each dimension are entangled, like the 3\textsuperscript{rd} dimension encodes both the upper bound and lower bound. These results indicate that, without manually designed parameters, our VAEGAN-based model learns geometrically meaningful features, and each dimension of the learned feature domain encodes informative and different geometry features.
\begin{figure}[hbt!]
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/change_1.png}
\caption{}
\label{fig:change_1}
\end{subfigure}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/change_8.png}
\caption{}
\label{fig:change_8}
\end{subfigure}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/change_22.png}
\caption{}
\label{fig:change_22}
\end{subfigure}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/change_32.png}
\caption{}
\label{fig:change_32}
\end{subfigure}
\caption{VAEGAN-generated airfoils by gradually changing only one dimension of the feature domain: (a) changes the 1\textsuperscript{st} dimension, (b) changes the 8\textsuperscript{th} dimension, (c) changes the 22\textsuperscript{nd} dimension, and (d) changes the 32\textsuperscript{nd} dimension}
\label{fig:change}
\end{figure}
\begin{figure}[hbt!]
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=.95\linewidth]{figures/vae_feat_tune_0.png}
\caption{}
\label{fig:vae_change_0}
\end{subfigure}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=.95\linewidth]{figures/vae_feat_tune_2.png}
\caption{}
\label{fig:vae_change_2}
\end{subfigure}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=.95\linewidth]{figures/vae_feat_tune_30.png}
\caption{}
\label{fig:vae_change_30}
\end{subfigure}
\caption{VAEGAN-generated airfoils by gradually changing only one dimension of the feature domain: (a) change the 1\textsuperscript{st} dimension, (b) changes the 3\textsuperscript{rd} dimension, and (c) changes the 31\textsuperscript{st} dimension}
\label{fig:vae_change}
\end{figure}
\begin{figure}[hbt!]
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/change_cluster_1.png}
\caption{}
\label{fig:change_cluster_1}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/change_cluster_8.png}
\caption{}
\label{fig:change_cluster_8}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/change_cluster_22.png}
\caption{}
\label{fig:change_cluster_22}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/change_cluster_32.png}
\caption{}
\label{fig:change_cluster_32}
\end{subfigure}
\caption{Path visualization of gradually changed feature on 2D embedded space using parametric t-SNE: (a) changing the 1\textsuperscript{st} dimension, (b) changing the 8\textsuperscript{th} dimension, (c) changing the 22\textsuperscript{nd} dimension, and (d) changing the 32\textsuperscript{nd} dimension}
\label{fig:change_cluster}
\end{figure}
\subsection{Synthesizing Novel Airfoils}
\noindent
To make use of the encoder-decoder architecture to synthesize novel airfoils, we conduct experiments on interpolation and extrapolation of feature vectors obtained from the UIUC Database airfoils as well as sampling from random Gaussian noises following the method introduced in Section \RNum{3} C.
As given in Eq.~\ref{eq:inter_2}, the affine combination of two feature vectors, $z_1$ and $z_2$, are computed with $\nu=0.5$. Fig~\ref{fig:inter} shows the interpolated airfoils from two different clusters. The labels under each airfoil indicate which two clusters are $z_1$ and $z_2$ come from. As illustrated in Fig~\ref{fig:inter}, the interpolated airfoil inherits features from both clusters. For instance, Cluster 6 and Cluster 11 both represent symmetric airfoil but with variant heights. The interpolation between these two clusters synthesizes a symmetric airfoil with a medium height, as shown in the last airfoil of Fig~\ref{fig:inter}. Also, Clusters 3 and 4 both encode thin airfoils. However, the lower boundary is concave in Cluster 3, while Cluster 4 represents a convex lower boundary making the airfoil symmetric in shape. The interpolation between these two generates a thin airfoil with a flat lower boundary, which is a combination of concave and convex curves. By interpolation, novel airfoils with features from different clusters can be generated. Extrapolation between airfoils from different clusters is conducted as well. Following Eq.~\ref{eq:inter_2}, two feature vectors, $z_1$ and $z_2$, are encoded from two different airfoils, and coefficient $\nu$ is set to be $2$. Fig~\ref{fig:extra} shows the generated results from the extrapolation. Similar to interpolation, extrapolated airfoils inherit features from $z_1$ and $z_2$.
\begin{figure}[hbt!]
\centering
\includegraphics[width=\textwidth]{figures/new_inter.png}
\caption{Interpolation of airfoils from different clusters}
\label{fig:inter}
\end{figure}
\begin{figure}[hbt!]
\centering
\includegraphics[width=\textwidth]{figures/new_extra.png}
\caption{Extrapolation of airfoils from different clusters}
\label{fig:extra}
\end{figure}
Besides interpolation and extrapolation, the performance of the sampling synthesis method is also estimated. In sampling, a Gaussian noise, $\hat{z} \sim \mathcal{N}(0,\mathbf{I})$, is directly fed into the decoder to generate novel airfoils. Shown in Fig.~\ref{fig:VAEGAN_sample} are airfoils synthesized by our VAEGAN model through sampling. Besides, sampled airfoils using PCA and VAE are also included in Fig.~\ref{fig:PCA_sample} and Fig.~\ref{fig:VAE_sample}. PCA, as a linear projection technique, fails to synthesize realistic airfoils through random sampling. Both deep-learning-based generative models, VAE and VAEGAN, can synthesize different smooth airfoil shapes. To quantitatively measure the synthesized airfoils from VAE and VAEGAN, we here introduce Fréchet Inception Distance (FID) \cite{heusel2017gans}, which is used to evaluate the quality of samples from deep learning-based generative models. FID is calculated by computing the Fréchet distance between two feature representations. Generally, lower FID indicates higher generative sample quality. In our case, we feed the synthesized airfoils and the UIUC airfoils into the well-trained discriminator from VAEGAN and extract the second hidden layer as the representation. The FID for VAE and VAEGAN are 1.38788 and 0.65366, respectively, meaning VAEGAN synthesizes more realistic airfoils. Also, airfoils synthesized via VAEGAN possess more novelty while maintains the general geometric pattern of airfoils. For instance, in Fig.~\ref{fig:VAEGAN_sample}, the first airfoils in the fourth row and the third one in the second row are different from existing samples in the UIUC database. Though such novelty does not guarantee better aerodynamic properties, some airfoils are likely to have negative lift coefficients, which are infeasible in practice. The VAEGAN-based model can synthesize a wide variety of airfoils that serve as candidates for further optimization through CFD simulation as we will investigate in Section \RNum{4} F.
\begin{figure}[hbt!]
\centering
\begin{subfigure}{0.3\textwidth}
\centering
\setlength{\fboxsep}{0.3pt}
\setlength{\fboxrule}{0.3pt}
\fbox{\includegraphics[width=\linewidth]{figures/PCA_sample.png}}
\caption{PCA}
\label{fig:PCA_sample}
\end{subfigure}
\hfill
\begin{subfigure}{0.3\textwidth}
\centering
\setlength{\fboxsep}{0.3pt}
\setlength{\fboxrule}{0.3pt}
\fbox{\includegraphics[width=\linewidth]{figures/VAE_sample.png}}
\caption{VAE}
\label{fig:VAE_sample}
\end{subfigure}
\hfill
\begin{subfigure}{0.3\textwidth}
\centering
\setlength{\fboxsep}{0.3pt}
\setlength{\fboxrule}{0.3pt}
\fbox{\includegraphics[width=\linewidth]{figures/VAEGAN_sample.png}}
\caption{VAEGAN}
\label{fig:VAEGAN_sample}
\end{subfigure}
\caption{Generated airfoils by sampling with different featurization techniques.}
\label{fig:sample}
\end{figure}
\subsection{Aerodynamic Properties of Synthesized Airfoil}
\begin{figure}[hbt!]
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/inter_aerodyn.png}
\caption{}
\label{fig:aerodyn_inter}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/sample_aerodyn.png}
\caption{}
\label{fig:aerodyn_sample}
\end{subfigure}
\caption{Lift coefficient $C_l$ v.s. drag coefficient $C_d$ of: (a) interpolated/extrapolated airfoils, and (b) sampled airfoils}
\label{fig:aerodyn}
\end{figure}
\noindent
Aerodynamic properties of the synthesized airfoils are also tested since airfoils have to meet certain aerodynamic properties to make the design feasible and effective. XFoil\footnote{\url{https://web.mit.edu/drela/Public/web/xfoil/}} is utilized to compute the lift coefficient, $C_l$, and the drag coefficient, $C_d$. The experiments on XFoil is set for a low speed condition: Reynolds number $Re=2 \times 10^6$, Mach number $Ma=0.02$, and attack angle $\alpha=\ang{0}$. As illustrated in Fig~\ref{fig:aerodyn_inter}, $C_l$ and $C_d$ are tested on three airfoils: NACA1412, NACA2424 and NACA4415 from the UIUC database. Following Eq.~\ref{eq:inter_3}, triplet interpolation/extrapolation is conducted with feature vectors, $z_1$, $z_2$, and $z_3$, encoded from the three NACA airfoils. The interpolated airfoils, marked by green dots, possess $C_l$ and $C_d$ in between the three NACA airfoils. While extrapolated airfoils, marked by black crossings, have significantly different aerodynamic properties from the interpolated airfoils. Some airfoils synthesized by extrapolation have high $C_l$ with relatively low $C_d$, located at the upper part of Fig.~\ref{fig:aerodyn_inter}. This demonstrates that by interpolation/extrapolation in the feature domain, novel airfoils with promising aerodynamic properties can be synthesized. Also, aerodynamic properties of synthesized airfoils by sampling are tested in comparison with some airfoils from the UIUC database, as shown in Fig~\ref{fig:aerodyn_sample}. Though the generated airfoils from Gaussian noises are not guaranteed to have good aerodynamic properties, like red dots lying on the bottom left. Some promising airfoils can be synthesized, as shown on the top right, with a high lift coefficient and a low drag coefficient. By sampling, the VAEGAN-based model can synthesize airfoils with a wide variety of aerodynamic properties. Such a variety provides abundant candidates to explore in design space for airfoil shape optimization.
\subsection{Shape Optimization on Aerodynamic Properties via Genetic Algorithm}
\begin{figure}[hbt!]
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/avg_score_new.png}
\caption{}
\label{fig:avg_score}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/genetic_alg_new.png}
\caption{}
\label{fig:genetic_alg_clcd}
\end{subfigure}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/airfoil_evolve.png}
\caption{}
\label{fig:airfoil_evole}
\end{subfigure}
\caption{Shape optimization via genetic algorithm: (a) average score of each generation; (b) $C_l$ and $C_d$ of synthesized airfoils for different generations; (c) synthesized airfoil geometry for different generation}
\label{fig:genetic_alg}
\end{figure}
\noindent
The VAEGAN model has been proven to be able to parameterize existing airfoils to latent feature vectors and synthesize novel airfoils automatically. However, whether or not the learned features and synthesized airfoils can be optimized to possess desired aerodynamic properties remains untested. To this end, this section demonstrates that with the VAEGAN model, airfoil shapes can be optimized to realize the target $C_l$ and $C_d$ value via a genetic algorithm (GA). The lift and drag coefficients, $C_l$ and $C_d$, calculated are under the same condition as Section \RNum{4} E. As illustrated in Fig.~\ref{fig:genetic_alg}, the target lift coefficient is $C_l^t=0.6$, and target drag coefficient is $C_d^t=0.06$. In our case, the total number of generations, $N$, is set to be 60, and the number of populations on each generation, $M$, to be 25. Fig.~\ref{fig:avg_score} and Fig.~\ref{fig:genetic_alg_clcd} show how the average score and $C_l,C_d$ change with generation, respectively. As Fig.~\ref{fig:airfoil_evole} illustrates, the airfoil shape gradually evolves to the target $C_l$ and $C_d$. Also, we compare the performance of airfoil optimization using different featurization techniques, PCA, and VAE. The genetic algorithm with the same objective function and settings are conducted. The lift coefficient $C_l$, drag coefficient $C_d$, and fitness score of the last generation in the genetic algorithm is reported in Table~\ref{tab:GA}. The VAEGAN-synthesized airfoils reach an averaged lift coefficient of 0.5857 and an averaged drag coefficient of 0.0061. The coefficients are close to the desired aerodynamic properties. Whereas PCA and VAE fail to synthesize desired airfoils within the same number of generations and population size. This is because our VAEGAN-based model generates a wider variety of airfoils that serves as potential candidates in design optimization.
Such experiments prove that a simple optimization technique like GA and the well-trained VAEGAN model can synthesize airfoils with desired aerodynamic properties, which can guide designing effective and efficient aerodynamic products.
\begin{table}[htb!]
\caption{\label{tab:GA} Airfoil design optimization results with different featurization techniques}
\centering
\begin{tabular}{lcccccccc}
\hline
& \multicolumn{2}{c}{$C_l$} & & \multicolumn{2}{c}{$C_d$} & & \multicolumn{2}{c}{Fitness score} \\
\cline{2-3} \cline{5-6} \cline{8-9}
Featurization & mean & std & & mean & std & & mean & std \\ \hline
PCA & 0.41953 & 0.012672 & & 0.0069430 & 6.9231 $\times 10^{-4}$ & & -0.12893 & 0.046250 \\
VAE & 0.52746 & 0.0017073 & & 0.0056662 & 1.4631 $\times 10^{-5}$ & & -0.017819 & 0.0014068 \\
VAEGAN & 0.58570 & 0.0028946 & & 0.0061030 & 7.0711 $\times 10^{-6}$ & & -8.5312 $\times 10^{-4}$ & 0.00089211 \\
\hline
\end{tabular}
\end{table}
\section{Conclusion}
\noindent
In this work, a data-driven method is proposed to achieve three goals: (1) automatically featuring airfoil geometries from the UIUC Database \cite{uiucdataset} without manually designed parameters, (2) synthesizing novel airfoils by either interpolating or extrapolating the encoded features, as well as generating from random noise, and (3) optimizing the features to synthesize airfoils with desired aerodynamic properties. Our model is built upon VAEGAN, which combines the encoder-decoder architecture from VAE \cite{kingma2013auto} \cite{rezende2014stochastic} and the discriminator from GAN \cite{goodfellow2014gan}. With the encoder-decoder structure, our model learns explicit mappings from airfoil coordinates to latent feature domain as well as from feature vectors to airfoils, while with the discriminator, the model can automatically synthesize realistic samples. Also, our model is trained in a self-supervised manner. Namely, the model learns compact and informative features directly from airfoil shapes without manually tagged labels or designed parameters. Optimized on the learned feature domain via GA, the synthesized airfoils can evolve to have desired aerodynamic properties.
Experiments show that our model learns compact and comprehensive features encoding shape information of airfoils and can automatically generate novel airfoils. First, airfoils can be reconstructed via decoding the learned features with minor error compared to the origin coordinates. Second, K-Means \cite{jain2010kmeans} clustering on the feature domain of the UIUC Coordinate Database further demonstrates the learned representations are meaningful in a way that the centroid of each cluster represents different shapes. It is also investigated what is encoded in each dimension of the feature domain by gradually changing the feature vector on one specific dimension with all other dimensions fixed. Without human prior, each dimension encodes different geometric information like height, camber, symmetry, and even coupled features. Moreover, novel airfoils are synthesized by interpolating and extrapolating learned features from different airfoils as well as directly generated from random noise. By interpolating or extrapolating, the synthesized airfoil inherits and blends features from existing airfoils, which provides insights for designing new airfoils. On the other hand, airfoils generated from Gaussian noise are more aggressive in a way that they follow a less geometrical format of existing airfoils, and more novelty is introduced to the airfoil design. Finally, the synthesized airfoils can be optimized via GA to possess competitive or even better aerodynamic properties in comparison to existing ones, indicating the synthesized geometries are not only plausible in shape but also practical in aerodynamic performance.
|
\section{Introduction}
\label{Introduction}
Deep Reinforcement Learning (DRL) has led to recent advancements that allow autonomous agents to solve complex tasks in a wide range of fields ({\cite{DBLP:journals/corr/SchulmanLMJA15}}, \cite{1509.02971}, \cite{DBLP:journals/corr/LevineFDA15}). However, traditional approaches in DRL learn a separate policy for each unique task, requiring large amounts of samples. Meta-Reinforcement learning (meta-RL) algorithms provide a solution by teaching agents to implicitly learn a shared structure among a batch of training tasks so that the policy for unseen similar tasks can quickly be acquired ({\cite{DBLP:journals/corr/FinnAL17}}). Recent progress in meta-RL has shown efforts being made in improving the sample complexity of meta-RL algorithms ({\cite{DBLP:journals/corr/abs-1903-08254}}, {\cite{DBLP:journals/corr/abs-1810-06784}}), along with the out-of-distribution performance of meta-RL algorithms during adaptation ({\cite{1910.00125, mendonca2020meta}}). However, most of the existing meta-RL algorithms prioritize sample efficiency at the sacrifice of computational complexity in adaptation, making them infeasible to adapt to fast-changing environments in real-world applications such as robotics.
In this paper, we present \textbf{Fast Linearized Adaptive Policy} (FLAP), an off-policy meta-RL method with great generalization ability and fast adaptation speeds.
FLAP is built on the assumption that similar tasks share a common linear (or low-dimensional) structure in the representation of the agent's policy, which is usually parameterized by a neural network. During training, we learn the shared linear structure among different tasks using an actor-critic algorithm. A separate \textit{adapter} net is also trained as a supervised learning problem to learn the weights of the output layer for each unique train task given by the environment interactions from the agent. Then when adapting to a new task, we fix the learned linear representation (shared model layers) and predict the weights for the new task using the trained adapter network. An illustration of our approach is highlighted in Figure \ref{overview method}.
We highlight our main contributions below:
\begin{itemize}
\item \textbf{State of the Art Performance.} We propose an algorithm based on learning and predicting the shared linear structures within policies, which gives the strongest results among the meta-RL algorithms and the fastest adaptation speeds.
FLAP is the state of the art in all these areas including performance, run-time, and memory usage.
As is shown in Figure~\ref{introduction.results}, the FLAP algorithm outperforms most of the existing meta-RL algorithms including MAML ({\cite{DBLP:journals/corr/FinnAL17}}), PEARL (\cite{DBLP:journals/corr/abs-1903-08254}) and MIER ({\cite{mendonca2020meta}}) in terms of both adaptation speed and average return. Further results from our experiments show that FLAP acquires adapted policies that perform much better on out-of-distribution tasks at a rapid run-time adaptation rate up to 8X faster than prior methods.
\item \textbf{Prediction rather than optimization.} We showcase a successful use of prediction via adapter network rather than optimization with gradient steps ({\cite{DBLP:journals/corr/FinnAL17}}) or the use of context encoders (\cite{DBLP:journals/corr/abs-1903-08254}) during adaptation. This ensures that different tasks would have policies that are different from each other, which boosts the out-of-distribution performance, while gradient-based and context-based methods tend to produce similar policies for all new tasks. Furthermore, the adapter network learns an efficient way of exploration such that during adaptation, only a few samples are needed to acquire the new policy. To our knowledge, this is the first meta-RL method that directly learns and predicts a (linearly) shared structure successfully in adapting to new tasks.
\end{itemize}
\begin{figure}[!htb]
\centering
\begin{minipage}[b]{0.1\textwidth}\hspace{1cm}
\end{minipage}
\begin{minipage}[b]{0.4\textwidth}
\centering
\includegraphics[width=\textwidth]{graphs/intro-return-ood.png}
\label{fig:introreturn}
\end{minipage}\hspace{0.1cm}
\begin{minipage}[b]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{graphs/intro-runtime.png}
\label{fig:introruntime}
\end{minipage}
\caption{\textbf{Strong Experimental Results:} We showcase the performance of meta-RL methods on tasks that are very different from the training tasks to assess the generalization ability of methods. We also analyze the adaptation run-time speed of these methods on tasks that are similar (in-distribution) and tasks that are not very similar (out-of-distribution) to further evaluate these models. Flap presents significantly stronger results compared to prior meta-RL methods.}
\label{introduction.results}
\end{figure}
\begin{figure}[!htbp]
\centering
\includegraphics[scale=0.2]{figures/overview.png}
\caption{\textbf{Overview of our approach:} In training, for different tasks $\{\mathcal{T}_i\}$, we parametrize their policy as $\pi_i = \phi\cdot w_i$, where $\phi\in\mathbb{R}^d$ is the shared linear representation we hope to acquire. In testing (adaptation), we fix the acquired linear representation $\phi$ and directly alter the weights $w_{test}$ by using the output of the feed-forward adapter network.}
\label{overview method}
\end{figure}
\subsection{Related work}
The idea behind learning a shared linear structure is motivated by
the great progress of representation learning and function approximation in domains such as robotics (\cite{robotics-survey}) and natural language processing ({\cite{DBLP:journals/corr/abs-1805-09461}}). Representation learning has seen success in cases of transfer learning RL in different tasks, and multi-agent reinforcement learning (MARL) where different agents may share common rewards or dynamics (\cite{JMLR:v10:taylor09a}, {\cite{ijcai2018-774}}). The idea of learning shared information (embeddings) across different tasks has been investigated deeply in transfer learning, including for example, universal value function approximators (\cite{pmlr-v37-schaul15}), successor features (\cite{NIPS2017_350db081}, \cite{DBLP:conf/icml/HuntBLH19}), and invariant feature spaces (\cite{DBLP:journals/corr/GuptaDLAL17})).
Using a prediction model during meta-testing presents a new idea into meta-RL. The current trend in meta-reinforcement learning has been to either use an encoder based method through the use of context captured in recurrent models (\cite{1910.00125}) or variational inference (\cite{DBLP:journals/corr/abs-1903-08254}), or use a variation of model-agnostic meta-learning (MAML) (\cite{DBLP:journals/corr/FinnAL17}), which aims at finding some common parametrization of the policy such that it suffices to take one gradient step in adaption to acquire a good policy.
In contrast, FLAP aims at predicting the weights for the policy network directly in adaptation without the need of taking gradient steps or using context and comparing with existing data. This gives better performance in (a) faster adaptation speeds due to the minimal amount of computation required; (b) smaller sample complexity during adaptation due to the better exploration strategy provided by adapter network; (c) strong out-of-distribution generalization ability resulting from the fact that the policy for each new task can be completely different in the last layer (in contrast, the new policies produced by gradient-based algorithms like MAML only differ in a single gradient update, which makes them highly similar, hurting the generalization ability). We believe the idea of predicting instead of fine-tuning via gradient descent can shed light to broader questions in meta-learning.
\section{Background}
We consider the standard Markov Decision Process (MDP) reinforcement learning setting defined by a tuple $\mathcal{T} = (\mathcal{S},\ \mathcal{A},\ \mu,\ \mu_0,\ r,\ T)$, where $\mathcal{S}$ is the state space, $\mathcal{A}$ is the action space, $\mu(s_{t+1} |s_t,\ a_t)$ is the unknown transition distribution with $\mu_0$ being the initial state distribution, $r(s_t, a_t)$ is the reward function, and $T$ is the episode length. Given a task $\mathcal{T}$, an agent's goal is to learn a policy $\pi:\mathcal{S} \rightarrow \mathcal{A}$ to maximize the expected (discounted) return:
\begin{equation}
q_{\gamma}(\pi, \mathcal{T}) = \mathbb{E}[\sum_{t=0}^{T-1} \gamma^tr(s_t, a_t)], s_0 \sim \mu_0, a_t \sim \pi(s_t), s_{t+1} \sim \mu(s_t, a_t)
\end{equation}
We now formalize the meta-RL problem. We are given a set of training tasks $\mathcal{T}_{train} = \{\mathcal{T}_i\}_{i=1}^n$ and test tasks $\mathcal{T}_{test} = \{\mathcal{T}_j\}_{j=1}^m$. In meta-training, we are allowed to sample in total $k$ trajectories $\mathcal{D}_{train}^k$ from the tasks in $\mathcal{T}_{train}$ and obtain some information $\phi(\mathcal{D}_{train}^k)$ that could be utilized for adaptation. The choice of $\phi(\mathcal{D}_{train}^k)$ varies for different algorithms. In meta-testing (or adaptation), we are allowed to sample $l$ trajectories $\mathcal{D}_{test}^l$ from each $\mathcal{T}_i\in \mathcal{T}_{test}$. We would like to obtain a policy $\pi(\phi(\mathcal{D}_{train}^k), \mathcal{D}_{test}^l)$ for each testing task $ \mathcal{T}_i$, such that the average expected return for all testing tasks is maximized:
\begin{align}\label{eqn.avg_exp_return}
\frac{1}{m} \sum_{\mathcal{T}_i \in \mathcal{T}_{test}}q_{\gamma}(\pi(\phi(\mathcal{D}_{train}^k), \mathcal{D}_{test}^l), \mathcal{T}_i).
\end{align}
Here the policy only depends on the training data through the information $\phi(\mathcal{D}_{train}^k)$, and can be different for different test tasks. We are interested in getting the expected return in \eqref{eqn.avg_exp_return} as large as possible with the smallest sample complexity $k, l$. Furthermore, we hope that the information $\phi(\mathcal{D}_{train}^k)$ is reasonably small in size such that the memory needed for testing is also controlled.
We also state that in-distribution tasks typically refer to the test tasks being drawn from the same distribution as the training distribution of tasks whereas out-of-distribution tasks represent the test distribution as very different from the training distribution (in our case, all \textit{out-of-distribution tasks} are completely disjoint from the training set).
\section{Fast Linearized Adaptive Policy (FLAP)}
\label{3}
In this section, we describe the FLAP algorithm. We first assert and justify our assumption behind the goal of the FLAP algorithm (Section \ref{motivation}), and then discuss the high-level flow of the algorithm (Section \ref{overview}).
\subsection{Assumption and Motivation}
\label{motivation}
Our FLAP algorithm is based on the following assumption on the shared structure between different tasks during training and testing:
\begin{assump}\label{assumption}
We assume that there exists some function $\phi: \mathcal{S} \rightarrow \mathbb{R}^{d_1}$ such that for any task in the meta-RL problem
$\mathcal{T}_i \in \{ \mathcal{T}_\textnormal{train} \cup \mathcal{T}_\textnormal{test} \} $, there exists a policy $\pi_i$ of the following form that maximizes the expected return $q_\gamma(\pi, \mathcal{T}_i)$:
\begin{equation}\label{eqn.rep_pi}
\pi_{i}= \sigma(\langle \phi, w_i \rangle + b_i).
\end{equation}
Here $w_i\in\mathbb{R}^{d_1 \times d_2}, \ b_i\in\mathbb{R}^{d_2}$. $d_1$ is the latent space dimension and $d_2$ is the dimension of the action space. The function $\phi$ represents a shared structure among all tasks, and $w_i$ and $b_i$ are the unique weight and bias for task $\mathcal{T}_i$. The function $\sigma$ is the activation function depending on the choice of the action space.
\end{assump}
Similar assumptions that constrain the policy space also appear in literature, e.g.~\cite{levine2016end, lillicrap2015continuous, schmidhuber2019reinforcement, mou2020sample} and references therein.
One can also extend the linear representation in~\eqref{eqn.rep_pi} to quadratic or higher-order representations for more representational ability and looser asumptions. We compare the performance between linear structures and non-linear structures in Section~\ref{nonlinear output layers}.
\subsection{Algorithm overview}
\label{overview}
Given assumption~\ref{assumption}, the overall workflow of the algorithm is quite straightforward: in meta-training, we would like to learn the shared feature $\phi$ from the training tasks $\mathcal{T}_{train}$; in meta-testing, we utilize the learned linear feature $\phi$ to derive the parameters $w_i, b_i$ for each new task. To speed up the adaptation, we also train a separate feed-forward adapter network simultaneously that takes the trajectories of each task as input and predicts the weights that are learned during training. The full pseudo-code for both meta-training and meta-testing is provided in Appendix \ref{pseudo code}.
\subsubsection{Meta-training}
In order to learn the linear representation $\phi$ from Assumption~\ref{assumption}, we use a deep neural network for the policy, sharing all but the output layer among all the train tasks in the training set. Each train task will have a unique output layer during meta-training, and the shared neural network is the feature $\phi$ we hope to learn during training.
Mathematically, we parametrize the policy via a neural network as $\pi(\phi, w_i, b_i)= \sigma(\langle \phi, w_i \rangle + b_i)$, where $w_i$ and $b_i$ are the weights and bias unique to task $i$, and $\sigma$ represents the linear activation for the output layer since for all the tasks we considered in this paper, the action space takes values in the range of $[-\infty, +\infty]$.
FLAP uses the general multi-task objective to train the policy over all the $n$ training tasks during meta-training. We maximize the average returns over all training tasks by learning the policy parameters represented by $\{\phi, {\{w_i, b_i\}}_{i=1}^n\} $. We would like to find the meta-learned parameters via solving the following optimization problem:
\begin{align}\label{eqn.meta_train}
\{\phi, {\{w_i, b_i\}}_{i=1}^n\} = \argmax_{ \{\phi, {\{w_i, b_i\}}_{i=1}^n\} } \frac{1}{n} \sum_{i=1}^{n} q_\gamma(\pi(\phi, w_i, b_i), \mathcal{T}_i).
\end{align}
Here $\phi$ is parameterized by a neural network, and $\pi_i$ is parameterized by the parameters $ \{\phi, {\{w_i, b_i\}}_{i=1}^n\}$.
\paragraph{Actor-Critic. }
For solving the optimization problem in~\eqref{eqn.meta_train}, we apply the off-policy soft actor-critic (SAC) ({\cite{DBLP:journals/corr/abs-1801-01290}}) for sample efficiency.
\begin{figure}[h]
\begin{center}
\centering
\includegraphics[scale=0.18]{figures/linear-rep.png}
\end{center}
\caption{
Each task shares a feature mapping in the critic and a different shared feature mapping in the policy network. In this illustration, for the training task 1 input data $x_1$, only the output $y_1$ from its corresponding output layer will be used in the meta-loss function. Similarly for task 2 input data $x_2$, only the output $y_2$ from its corresponding output layer will be used in the meta-loss. This is extended up to $n$ points where $n$ is the number of training tasks used.}
\label{Figure 1}
\end{figure}
In the actor-critic model, besides the policy network, we also have a critic network which learns a $Q$ function for evaluating the current policy. We linearize the critic network in a similar fashion to the policy by fixing a set of shared layers in the critic network. We show the shared layer structure for both policy and critic network in Figure~\ref{Figure 1}. Linearizing the critic network, though not explicitly required, does not incur any loss in performance from experimental results, and drastically reduces the number of parameters needed to be learned during training for the critic portion of the actor-critic method. We discuss in more detail about the use of the soft actor-critic in Appendix \ref{Actor-critic appendix}.
\paragraph{Adapter network.}
Learning the shared structure is a standard approach in multi-task transfer learning (\cite{pan2009survey, Asawa2017UsingTL,D'Eramo2020Sharing}). The key contribution and distinction of our algorithm is the use of an adapter network for rapid adaptation. We train a separate adapter network simultaneously with the shared structure during meta-training. We define the \textbf{SARS} tuple as a tuple containing information from one time step taken on the environment by an agent given a single task, including the current state, action taken, reward received from taking the action given the current state, and the next state the agent enters ($s_t, a_t, s_{t+1}, r(s_t, a_t)$). The adapter net is fed a single tuple or a sequence of the most recent {SARS} tuples from a task and predicts the unique weights ($\hat{w_i}$) needed to acquire the new policy for that task. We train the adapter network with mean squared error between the output of the adapter net ($\hat{w}_i$) obtained from the input {SARS} tuples, and the set of weights obtained from meta-training contained in each output layer. The process of training the adapter net is shown in Figure \ref{adapter method}.
\begin{figure}[!htbp]
\centering
Training the Adapter Network\par\medskip
\includegraphics[scale=0.2]{figures/meta-train.png}
\caption{Illustration of training the adapter network. The weights of the corresponding output layer for the training task are flattened and concatenated to form a 1-D target for the output of the adapter model.}
\label{adapter method}
\end{figure}
\subsubsection{Meta-testing}
In meta-testing, the FLAP algorithm adapts the meta-learned policy parameters $w_i, b_i$ to a new task ($\mathcal{T}_{i}\in\mathcal{T}_{test}$) with very little data sampled from the new task. Following Assumption~\ref{assumption}, we assume that the shared structure $\phi$ is already learned well during meta-training. Thus in meta-testing, we fix the structure $\phi$ obtained in training, and only adjust the weights for each test task.
Mathematically, we run the following optimization problem to obtain the policy for each new test task $\mathcal{T}_{new} \in \mathcal{T}_{test}$: The objective for the new task ($\mathcal{T}_{\textnormal{new}}$) is to find the set of weights that with the shared structure $\phi$ form the adapted policy
\begin{equation}\label{eqn.meta_test}
w_{new}, b_{new} = \argmax_{w, b} q_\gamma(\pi(\phi, w, b), \mathcal{T}_{new}).
\end{equation}
We then output $\pi(\phi, w_{new}, b_{new})$ as the policy for the new task.
\begin{figure}[!htbp]
\centering
Online Adaptation\par\medskip
\includegraphics[scale=0.2]{figures/meta-test.png}
\caption{The first step taken is arbitrary. Each step following the first is fed into the feed-back loop where the SARS pair is inputted to the adapter network providing a new set of weights for the output layer of the policy.}
\label{adapter}
\end{figure}
In our practical implementation and results, we show our contribution that requiring an objective function to be optimized during meta-testing does not necessarily have to be the only approach for meta-RL, though FLAP can still have the optimization in Equation \ref{eqn.meta_test} easily integrated in meta-testing. FLAP performs rapid online adaptation by using the adapter network to predict the weights of the output layer as opposed to performing the optimization step. The first time-step taken by the agent is based off the most recent weights learned from the adapter network. This then initiates a feed-back loop where now each time-step gives the adapter network a single tuple or a sequence of the most recent \textbf{SARS} tuples which inputted to the adapter network, immediately produces a set of predicted weights to use with the linear feature mapping ($\phi$) to acquire an adapted policy ($\pi_{\textnormal{adapt}} = \phi^T\cdot w_i + b_i$). Only a fixed number of online adaptation steps are needed to acquire the fully adapted policy. Once the last time-step of weights is predicted, FLAP keeps \textit{only} the last set of predictive weights for the rest of the episode. The meta-testing loop is shown in Figure \ref{adapter}.
\section{Experiments}
\label{experiment section}
This section presents the main experients of our method. We evaluate FLAP in Section \ref{4.1} by comparing it to prior meta-learning approaches on in-distribution tasks. We then examine how the FLAP algorithm is able to generalize to very different test tasks using out-of-distribution tasks in Section \ref{sec.outofdistribution}. We take a holistic view of FLAP as a meta-RL method and analyze its run-time and memory usage on an autonomous agent during the meta-testing stage in Section \ref{sec.model_analysis}. We leave the ablation studies and sparse reward experiments to Appendix~\ref{appendix.more_experiments}. Finally, we provide thorough discussions on the empirical performance in Section~\ref{sec.discussion_experiments}.
\subsection{Sample Efficiency and Performance Compared to Prior Methods}
\label{4.1}
\begin{figure}[h]
\centering
\begin{minipage}[b]{0.33\textwidth}
\includegraphics[width=\textwidth]{graphs/cheetah-fb.png}
\end{minipage}
\begin{minipage}[b]{0.33\textwidth}
\includegraphics[width=\textwidth]{graphs/Ant-FB.png}
\end{minipage}
\begin{minipage}[b]{0.33\textwidth}
\includegraphics[width=\textwidth]{graphs/Humanoid.png}
\end{minipage}
\begin{minipage}[b]{0.33\textwidth}
\includegraphics[width=\textwidth]{graphs/Walker.png}
\end{minipage}\par\vspace{-0.2\baselineskip}
\begin{center}
\begin{minipage}[b]{0.7\textwidth}
\includegraphics[width=\textwidth]{graphs/in-distribution_legend.png}
\end{minipage}
\end{center}
\caption{
Validation Test-Task performance over the course of the meta-training process on standard in-distribution meta-RL benchmarks. Our algorithm is comparable in all environments to prior meta-RL methods.}
\end{figure}
We evaluate FLAP on standard meta-RL benchmarks that use in-distribution test tasks (training and testing distributions are shared). These continuous control environments are simulated via the MuJoCo simulator ({\cite{6386109}}) and OpenAI Gym ({\cite{DBLP:journals/corr/BrockmanCPSSTZ16}}). \cite{DBLP:journals/corr/abs-1903-08254} constructed a fixed set of meta-training tasks ($\mathcal{T}_{train}$) and a validation set of tasks ($\mathcal{T}_{test}$). These tasks have different rewards and randomized system dynamics (Walker-2D-Params). To enable comparison with these past methods, we follow the evaluation code published by \cite{DBLP:journals/corr/abs-1903-08254} and follow the exact same evaluation protocol. In all the tasks we consider, the adapter network is fed with a single SARS tuple during training and adaptation. Other hyper-paramters of our method are given in Appendix \ref{hyper} and we report the result of FLAP on each environment from 3 random seeds.
We compare against standard baseline algorithms of MAML ({\cite{DBLP:journals/corr/FinnAL17}}), RL2 ({\cite{DBLP:journals/corr/DuanSCBSA16}}), ProMP ({\cite{DBLP:journals/corr/abs-1810-06784}}), PEARL (\cite{DBLP:journals/corr/abs-1903-08254}), MQL ({\cite{1910.00125}}), and MIER (and MIER-wR) ({\cite{mendonca2020meta}}). We obtained the training curves for the first four algorithms from \cite{DBLP:journals/corr/abs-1903-08254} and the latter two from {\cite{mendonca2020meta}}.
\subsection{Generalization to Out-of-Distribution Tasks}\label{sec.outofdistribution}
To evaluate generalization ability, we test on the extrapolated out-of-distribution (OOD) environments, where the train and test distributions are highly dissimilar and completely disjoint.
\begin{figure}[!htb]
\centering
\begin{minipage}[b]{0.35\textwidth}
\includegraphics[width=\textwidth]{graphs/Cheetah-Med.png}
\end{minipage}
\begin{minipage}[b]{0.35\textwidth}
\includegraphics[width=\textwidth]{graphs/Cheetah-Hard.png}
\end{minipage}
\begin{minipage}[b]{0.35\textwidth}
\includegraphics[width=\textwidth]{graphs/Ant-Dir.png}
\end{minipage}
\begin{minipage}[b]{0.35\textwidth}
\includegraphics[width=\textwidth]{graphs/Ant-Negated.png}
\end{minipage}\par\vspace{-0.2\baselineskip}
\begin{center}
\begin{minipage}[b]{0.7\textwidth}
\includegraphics[width=\textwidth]{graphs/out-of-dist-legend.png}
\end{minipage}
\end{center}
\caption{Validation Test-Task performance over the course of the meta-training process on \textbf{out-of-distribution} meta-RL benchmarks. Our algorithm is comparable if not better to prior methods in all environments except for the Ant-Negated task.}
\end{figure}
These environments were created by {\cite{mendonca2020meta}} and {\cite{1910.00125}}, and we closely follow their evaluation protocols for fair comparisons. The environments include changes in rewards and dynamics. We run FLAP on 3 random seeds to evaluate the performance. In all the tasks we consider, the adapter network is fed with a single \textbf{SARS} tuple during training and adaptation. Hyperparameters are listed in Appendix \ref{hyper}.
\subsection{Overall Model Analysis}\label{sec.model_analysis}
In this section, we assess the run-time of FLAP. We compare our method to MIER-wr and PEARL in meta-testing speeds on benchmarks from both in-distribution and out-of-distribution tasks. MIER and MQL both reuse the training buffer during testing, causing adaptation speeds and memory usage that are far worse than prior methods (MIER took 500+ seconds for meta-testing). We only compare with the most competitive algorithms speed and memory-wise to showcase FLAP's strength. We defer the analysis of storage efficiency to Appendix (\ref{memory analysis}).
To calculate the run-time speeds, we use the Python Gtimer package and a NVIDIA GEFORCE GTX 1080 GPU and time each of the Meta-RL methods on four MuJoCo environments (Ant, Cheetah, Humanoid, Walker) on 30 validation tasks after allowing each algorithm to achieve convergence during meta-training.
\begin{figure}[!h]
\label{runtime}
\centering
\includegraphics[width=5cm]{graphs/Speed.png}
\caption{ Our method clearly outperforms all other meta-RL algorithms in runtime during adaptation to new tasks.
}
\end{figure}
\subsection{Discussion}\label{sec.discussion_experiments}
Beyond the experiments on in-distribution and out-of-distribution tasks, we did more experiments and analysis to justify the superiority of FLAP, including ablation studies and extra experiments on MDPs with sparse rewards. We briefly describe the results and take-aways in Section~\ref{sec.ablation} and~\ref{sec.sparse}, and leave the details on experiments to Appendix~\ref{appendix.more_experiments}. Lastly, we analyze all the experimental results and summarize the reason for the validity and generalization ability of the adapter network in Section~\ref{sec.success_reasons}
\subsubsection{Ablation studies} \label{sec.ablation}
We validate the effectiveness of FLAP via a large amount of ablation studies. In particular, we compare the performance under the following different settings:
\begin{enumerate}
\item \textbf{With adapter net vs Without adapter net:} One may wonder whether the adapter network really boosts the performance compared to the approach of directly learning the linear weights via soft actor-critic during adaptation without the help of the adapter network. As is shown in Appendix~\ref{appendix.without_adapter}, although both algorithms converge to approximately the same average reward in the end, the adapter network drastically decreases the sample complexity during the adaptation process and accelerates the speed (at least 30X faster among all the experiments) compared to the algorithm without the adapter.
\item \textbf{Shared linear layer vs Non-linear layers:} One may wonder why we assume the policies share only the very last layer instead of the last two or more layers as embeddings. Indeed, experiments in Appendix~\ref{nonlinear output layers} show that when sharing two (or more) layers, the policies may have too much degree of freedom to converge to the global minimum. The shared linear layer turns out to be the best assumption that balances the performance with the speed.
\item \textbf{Single SARS tuple vs a Sequence of SARS tuple as Input of Adapter Net:} All the experiments above work with the adapter net with a single (the most recent) SARS tuple as input. We show in Appendix~\ref{appendix.input_adapter} that if we let the adapter net take a sequence of the most recent SARS tuples as input, the average return would not be affected on the tasks we considered above. However, we see a reduction in variance with a slightly slower adaptation speed. Overall, taking a longer sequence of SARS tuples as input tends to \textbf{stabilize the adaptation} and \textbf{reduce the variance} in the sacrifice of runtime. However, for some specific tasks with sparse rewards and similar SARS tuples, it also helps improve the average return, as is discussed in the next section.
\end{enumerate}
\subsubsection{Experiments on Sparse Reward} \label{sec.sparse}
We show in Appendix~\ref{appendix.sparse} that for tasks with sparse rewards and similar dynamics (e.g. the navigation tasks), the performance of FLAP may deteriorate a little since the resulting policy highly relies on the chosen transition tuple, which may lead to higher variance. However, it is still comparable to PEARL if we let the adapter net take a sequence of SARS tuples as input. This is due to fact that a sequence of SARS tuples provides more information on the context for the adapter, making the resulting policy less dependent on the single chosen transition tuple.
However, for most of the other tasks we considered above, one single transition tuple suffices to provide enough information for the success of adapter network with small variance and strong generalization ability. We also validate it in Appendix~\ref{loss anlysis} by showing that the loss for adapter net converges in a reliably fast and stable way.
\subsubsection{On the validity and generalization ability of FLAP and the Adapter Network}\label{sec.success_reasons}
To summarize the experimental results above, we see that FLAP guarantees a much smaller sample complexity during adaptation, improved generalization ability and faster adaptation speeds. We provide justification as to why FLAP can achieve all the properties separately.
\begin{enumerate}
\item \textbf{Small Sample Complexity:} From our ablation study in Appendix~\ref{appendix.without_adapter}, it is clear that the small sample complexity during meta-testing comes from the introduction of adapter net. We conjecture that during training, the adapter net learns the structure of optimal policies shared by all the training tasks; and during adaptation, the predicted weights encourage the most efficient exploration over potential structures of optimal policies such that the average return converges in a fast and stable way.
\item \textbf{Strong generalization: } From our ablation study in Appendix~\ref{appendix.without_adapter}, we see that the algorithms with and without adapter converge to a similar average return. Thus we conclude that the generalization ability is not related to the specific design of adapter net, but comes from the shared linear structure. In traditional gradient-based meta-RL approaches like MAML, the new policies are generated from taking one gradient step from the original weights in training, which makes the new policies very similar to the old ones. In contrast, the new policy in FLAP can have a completely different last layer, which guarantees more flexibility in varying policies, and thus a better generalization to new tasks.
\item \textbf{Fast adaptation:} Compared to other meta-RL approaches, FLAP only requires evaluating the output of a fixed adapter net, which provides instant adaptation when the adapter net is small enough.
\end{enumerate}
The successful experiments also justify the validity of Assumption~\ref{assumption}: for all the tasks we consider, they do have some near-optimal policies with shared embedding layers. We believe this could also provide interesting theoretical directions as well.
\section{Conclusion}
In this paper, we presented \textbf{FLAP}, a new method for meta-RL with significantly faster adaptation run-time speeds and performance.
Our contribution in the FLAP algorithm presents a completely new view into meta reinforcement learning by applying linear function approximations and predicting the adaptation weights directly instead of taking gradient steps or using context encoders. We also believe the idea behind the adapter network during meta-testing can be extended to other meta learning problems outside of reinforcement learning.
|
\section{Introduction}
A \textit{strong edge coloring} of a graph $G$ is a proper edge
coloring with no bichromatic path of length three; in other
words, each color class is an induced matching. The minimum
number of colors for which $G$ admits a strong edge coloring is
called the \textit{strong chromatic index}, and is denoted by
$\chis{G}$. In 1985, Erd\H{o}s and Ne\v{s}et\v{r}il proposed
the following conjecture which was later published
in~\cite{Erd88} and updated by Faudree et
al.~\cite{FauSchGyaTuz90} to fit the graphs with an even or odd
maximum degree.
\begin{conjecture}[Erd\H{o}s, Ne\v{s}et\v{r}il, 1988]\label{conj:Erdos}
The strong chromatic index of an arbitrary graph $G$ satisfies
$$
\arraycolsep=1.4pt\def1.2{1.2}
\chi_s'(G) \le \left \{
\begin{array}{cl}
\tfrac{5}{4} \Delta(G)^2\,, &\quad \textrm{if }
\Delta(G) \textrm{ is even}\\
\tfrac{1}{4}(5\Delta(G)^2 - 2\Delta(G) +1)\,, &\quad \textrm{if }
\Delta(G) \textrm{ is odd.}
\end{array} \right.
$$
\end{conjecture}\noindent
Despite many efforts, this conjecture is still widely open and
the best current upper bound $1.772\Delta(G)^2$ (provided that
$\Delta(G)$ is large enough) is due to Hurley et
al.~\cite{HurJoaKan20}.
The motivation for this note comes from one of the two extant
cases of the conjecture of Faudree et
al.~\cite[Section~4]{FauSchGyaTuz90} about strong edge
colorings of subcubic graphs, that is, graphs with maximum
degree $3$.
\begin{conjecture}[Faudree, Schelp, Gy\'{a}rf\'{a}s, Tuza, 1990]\label{conj:main}
Let $G$ be a graph with maximum degree~$3$. Then
\begin{itemize}
\item[$(1)$] $\chis{G} \le 10$,
\item[$(2)$] $\chis{G} \le 9$ if $G$ is bipartite,
\item[$(3)$] $\chis{G} \le 9$ if $G$ is planar,
\item[$(4)$] $\chis{G} \le 6$ if $G$ is bipartite and for
every edge the sum of degrees of its endvertices is at
most $5$,
\item[$(5)$] $\chis{G} \le 7$ if $G$ is bipartite with
girth at least $6$,
\item[$(6)$] $\chis{G} \le 5$ if $G$ is bipartite with
large girth.
\end{itemize}
\end{conjecture}
The first four cases of Conjecture~\ref{conj:main} have already
been resolved. Case~$(1)$, which~is just a special case of the
conjecture of Erd\H{o}s and Ne\v{s}et\v{r}il, has been
confirmed by Andersen~\cite{And92} and by Hor\'{a}k et
al.~\cite{HorQinTro93}. Case~$(2)$ was proved in 1993 by Steger
and Yu~\cite{SteYu93}, and Case~$(3)$, just recently, by
Kostochka et al.~\cite{KosLiRukSanWanYu16}. Case~$(4)$ was
established by Wu and Lin~\cite{WuLin08}; it easily follows
also from a result of Maydanskiy~\cite{May05}. Up to our best
knowledge, Cases~$(5)$ and~$(6)$ are still open, although
several partial results confirming Case~$(6)$ are known
\cite{BorIva13,Lidetal18}.
Our aim is to show that Case~$(6)$ of
Conjecture~\ref{conj:main} is false. In order to be able to
produce infinitely many counterexamples, in
Theorem~\ref{thm:main} we characterize $k$-regular graphs with
strong chromatic index $2k-1$ as those which admit a covering
projection onto the Kneser graph $K(2k-1,k-1)$. In particular,
when $k=3$, a cubic graph is strongly $5$-edge-colorable if and
only if it covers the Kneser graph $K(5,2)$. However, the
latter is nothing but the Petersen graph. Subsequently, in
Theorem~\ref{thm:disproof}, we construct bipartite cubic graphs
of arbitrarily large girth that do not cover the Petersen
graph. By the former result, their strong chromatic index must
be at least $6$.
The last three sections of this paper are devoted to further
aspects of strong $(2k-1)$-colorings of $k$-regular graphs,
with emphasis on the cubic case. In Section~\ref{sec:equiv}, we
reflect on the fact, established in Section~\ref{sec:main},
that the Kneser graph $K(2k-1,k-1)$ has a unique strong
$(2k-1)$-coloring up to automorphism. With the help of a result
borrowed from the theory of voltage graphs we are able to
provide an example of a cubic graph on 40 vertices which covers
the Petersen graph and admits two substantially different
strong $5$-edge-colorings. In Section~\ref{sec:normal}, we
explain the relationship of strong $5$-colorings of cubic
graphs to the famous Petersen coloring conjecture, and in the
final section we present several open problems. Among them, we
propose a strengthening of Case~(5) of
Conjecture~\ref{conj:main}, its only remaining open case.
\section{Main results}\label{sec:main}
Let $\phi$ be a proper edge coloring of a graph $G$. An edge
$e$ of $G$ is said to be \textit{rich} with respect to $\phi$
if all the edges adjacent to $e$ receive pairwise distinct
colors. If $\phi$ is a strong edge coloring, then each edge
must obviously be rich, and vice versa. In particular, every
strong edge coloring of a $k$-regular graph requires at least
$2k-1$ colors. It has been shown in
\cite[Theorem~8]{FauSchGyaTuz90} that this minimum is also
sufficient if $G$ is the Kneser graph $K(2k-1,k-1)$. Recall
that the \textit{Kneser graph} $K(m,n)$, with $m\ge 2n+1$ and
$n\ge 2$, is the graph whose vertices are the $n$-element
subsets of a ground set of $m$ elements, say $\{1,2,\ldots,
m\}$, and where two vertices are adjacent if and only if the
two corresponding sets are disjoint. The Kneser graphs
$K(2k-1,k-1)$ are commonly known as the \textit{odd graphs}
$O_k$ and have been subject to numerous investigations (see
for example the work of Biggs \cite{Biggs}). The smallest odd
graph $K(5,2)$ is isomorphic to the Petersen graph.
Every odd graph $K(2k-1,k-1)$ has a natural -- or
\textit{canonical} -- strong $(2k-1)$-edge-coloring, which we
denote by $\sigma_k$. It can be described as follows: for any
edge $uv$ of $K(2k-1,k-1)$ the set $u\cup
v\subseteq\{1,2,\ldots, 2k-1\}$ contains precisely $2k-2$
elements, so we can set $\sigma_k(uv)$ to be the missing
element of the ground set. It is easy to see that this coloring
is indeed strong. The canonical strong $5$-edge-coloring
$\sigma_3$ of the Petersen graph is represented in
Figure~\ref{fig:pet}.
Observe that every strong $(2k-1)$-edge-coloring $\sigma$ of
any $k$-regular graph $G$ induces a vertex coloring $\sigma'$
of $G$ where every vertex $v$ is colored with the
$(k-1)$-element set of colors that do not occur on the edges
incident with $v$. We call $\sigma'$ the \textit{derived vertex
coloring}. Since $\sigma$ is strong, the colors of any two
adjacent vertices of $G$ are disjoint $(k-1)$-subsets; in
particular, $\sigma'$ is a proper vertex coloring. For the
Petersen graph the derived coloring $\sigma_3'$ is again
indicated in Figure~\ref{fig:pet}.
It is quite remarkable that the edge coloring $\sigma$ can be
uniquely reconstructed from the vertex coloring $\sigma'$: the
edge $uv$ is colored with the element of the ground set not
occurring in the set $\sigma'(u)\cup\sigma'(v)$. This fact
readily implies that the canonical coloring is a unique strong
$(2k-1)$-edge-coloring of $K(2k-1,k-1)$ up to automorphism of
$K(2k-1,k-1)$. Indeed, consider an arbitrary strong
$(2k-1)$-edge-coloring $\tau$ of $K(2k-1,k-1)$. The derived
vertex coloring $\tau'$ associates with each vertex $v$ of
$K(2k-1,k-1)$ -- which is a $(k-1)$-element subset of
$\{1,2,\ldots,2k-1\}$ -- another $(k-1)$-element subset
$\tau'(v)$ of the same set. In other words, $\tau'$ sends a
vertex of $K(2k-1,k-1)$ to another such vertex. Since $\tau$ is
strong, the assignment $v\mapsto \tau'(v)$ is
adjacency-preserving and injective on the neighbors of $v$. It
means that $\tau'$ determines a degree-preserving endomorphism
$\alpha$ of $K(2k-1,k-1)$, which necessarily must be an
automorphism. The way how the canonical $(2k-1)$-edge-coloring
$\sigma_k$ was defined implies that $\alpha$ transforms
$\sigma_k'$ to $\tau'$, and consequently $\sigma_k$ to $\tau$.
Summing up, any two strong $(2k-1)$-edge-colorings of
$K(2k-1,k-1)$ are equivalent under the action of its
automorphism group. Taking into account the fact that every
automorphism of $K(2k-1,k-1)$ is induced by a permutation of
the ground set \cite[Statement~3.1]{Biggs} we can conclude that
any two strong $(2k-1)$-colorings of $K(2k-1,k-1)$ can be
obtained from each other by a permutation of colors.
\begin{figure}[ht]
\begin{center}
\includegraphics{fig_pet_new}
\caption{The unique strong $5$-edge-coloring of the Petersen graph
along with the derived proper vertex coloring}
\label{fig:pet}
\end{center}
\end{figure}
We aim to prove that all $k$-regular graphs whose strong
chromatic index equals $2k-1$ are closely related to the Kneser
graph $K(2k-1,k-1)$, the relationship being a covering
projection onto $K(2k-1,k-1)$. The pertinent definitions are
now in order.
A surjective graph homomorphism $f\colon \tilde G\to G$ is
called a \textit{covering projection} if for every vertex
$\tilde v$ of $\tilde G$ the set of edges incident with $\tilde
v$ is bijectively mapped onto the set of edges incident with
$f(\tilde v)$. (If $G$ is permitted to contain loops, then the
definition has to be applied to the half-edges incident with
$v$ rather than the edges themselves.) The graph $G$ is usually
referred to as the \textit{base graph} and $\tilde G$ as a
\textit{covering graph} or a \textit{lift} of $G$. A graph
$\tilde G$ \textit{covers} $G$ if there exists such a covering
projection.
It is well known (see \cite{GrossTucker}, Section 2.3) that for
every covering projection $f\colon\tilde G\to G$ there exists a
positive integer $d$ such that every vertex of $G$ has exactly
$d$ preimages and every edge of $G$ has exactly $d$ preimages;
such a cover is said to be \textit{$d$-fold}. For example, the
Petersen graph admits a $5$-fold covering projection onto the
dumbbell graph (which consists of two adjacent vertices and a
loop attached to each of them).
Covering graphs have been useful in numerous parts of graph
theory, especially when a locally defined structure on the base
graph can be `lifted' to the covering graph. This is true, for
example, for edge colorings, flows, embeddings on surfaces, and
other graph structures. The following easy fact is in a similar
vein.
\begin{lemma}\label{lm:lift}
Let $f \colon \tilde{G}\to G$ be a covering projection of
graphs. If $G$ is strongly $n$-edge-colorable for
some integer $n$, then so is $\tilde G$.
\end{lemma}
\begin{proof}
Let $\phi$ be a strong $n$-edge-coloring of $G$. Define an edge
coloring $\tilde\phi$ of $\tilde G$ by setting
$\tilde\phi(x)=\phi(f(x))$ for each edge $x$ of $\tilde G$. As
previously mentioned, every edge of $G$ with respect to $\phi$
is rich. The definition of a covering immediately implies that
the same holds for each edge of $\tilde G$ with respect to
$\tilde\phi$. Therefore $\tilde\phi$ is a strong
$n$-edge-coloring of $\tilde G$.
\end{proof}
Now we are ready for our main results.
\begin{theorem}\label{thm:main}
The strong chromatic index of a $k$-regular graph $G$ equals
$2k-1$ if and only if $G$ covers the Kneser graph
$K(2k-1,k-1)$.
\end{theorem}
\begin{proof}
The backward implication is a direct consequence of
Lemma~\ref{lm:lift}, so we are left with the forward
implication.
Let $\sigma$ be a strong edge coloring of a $k$-regular graph
$G$ with $2k-1$ colors from the set $\{1,2,\ldots, 2k-1\}$.
Without loss of generality we may assume that $G$ is connected.
To define a covering projection $f\colon G\to K(2k-1,k-1)$ we
use the derived vertex coloring $\sigma'$ of $G$. Recall that
for each vertex $v$ of $G$ the color $\sigma'(v)$ is a
$(k-1)$-subset of $\{1,2,\ldots, 2k-1\}$. Thus there is a
unique vertex $\bar v$ of $K(2k-1,k-1)$ such that
$\sigma'(v)=\bar v$. Define $f$ by sending $v$ to $\bar v$. The
mapping is clearly correctly defined.
We first observe that $f\colon G\to K(2k-1,k-1)$ is a
homomorphism. To see this, note that the colors of adjacent
vertices in $G$ are disjoint $(k-1)$-elements sets. It follows
that $f$ sends adjacent vertices $u$ and $v$ to disjoint sets
$\bar u$ and $\bar v$. However, in $K(2k-1,k-1)$ such vertices
are adjacent. Therefore $f$ sends adjacent vertices to adjacent
vertices.
To show that $f$ is a covering projection we need to check that
$f$ takes the neighborhood of every vertex bijectively to
$K(2k-1,k-1)$, and that $f$ is surjective. Consider an
arbitrary vertex $v$ of $G$, and note that the $k$ neighbors
$u_i$ of $v$, where $1 \le i \le k$, receive from $\sigma'$
pairwise distinct colors $\sigma'(u_i)$. Since $f$ sends each
$u_i$ to the vertex $\bar u_i=\sigma'(u_i)$ in $K(2k-1,k-1)$,
it takes the $k$ neighbors of $v$ to $k$ distinct neighbors of
$f(v)$, as required.
Finally, to check that $f$ is surjective it is sufficient to
realize that $f(G)$ is a connected subgraph of $K(2k-1,k-1)$
and that $f(G)$ is $k$-regular. Therefore $f(G)=K(2k-1,k-1)$,
which proves that $f$ is a covering projection.
\end{proof}
For cubic graph the previous theorem amounts to the following.
\begin{corollary}\label{cor:main}
The strong chromatic index of a cubic graph $G$ equals $5$ if
and only if $G$ covers the Petersen graph.
\end{corollary}
Our next theorem refutes Case~$(6)$ of
Conjecture~\ref{conj:main}.
\begin{theorem}\label{thm:disproof}
There exist bipartite cubic graphs with arbitrarily large girth
and strong chromatic index at least $6$.
\end{theorem}
\begin{proof}
We construct an infinite sequence $(G_n)_{n\ge 1}$ of connected
bipartite cubic graphs such that $G_n$ has girth $2^n$ and
order a power of $2$. Since the order of any graph that covers
the Petersen graph is a multiple of $10$, it follows that $G_n$
does not cover the Petersen graph for any $n\ge 1$. From
Corollary~\ref{cor:main} we get that $\chis{G_n}\ge 6$ for each
$n\ge 2$.
We now construct the sequence by induction on $n$. For the
starting graph $G_1$ we take the cubic graph consisting of two
vertices and three parallel edges, which is connected,
bipartite, and has girth $2$. Assume that we have already
constructed the graph $G_n$ for some $n\ge 1$. By the induction
hypothesis, $G_n$ is connected and bipartite with girth $2^n$
and order a power of $2$. To create $G_{n+1}$, we employ the
construction of Exoo and Jajcay described in the proof of
Theorem~3.1 of \cite{ExoJaj11}. Their method uses a
$\mathbb{Z}_2$-homology voltage assignment on a connected graph
$H$ of girth $g$ to produce a $2^{\beta(H)}$-fold covering
projection $\tilde H\to H$ with $\tilde H$ connected of girth
$2g$, where $\beta(G)$ denotes the cycle rank (Betti number)
of~$G$. (We refer the reader for details to \cite{ExoJaj11}.)
If we apply this construction to $G_n$, we obtain a connected
graph $G_{n+1}$ of girth $2^{n+1}$ and order
$2^{\beta(G_n)}m_n$, where $m_n$ is the order of $G_n$. Since
$m_n$ is a power of $2$, so is the order of $G_{n+1}$.
Furthermore, $G_{n+1}$ is bipartite because any covering graph
over a bipartite base graph is bipartite. This concludes the
construction of $(G_n)_{n\ge1}$ as well as the entire proof.
\end{proof}
\begin{remark}
{\rm The reader can check that the graph $G_2$ constructed in
the previous proof is isomorphic to the graph of the $3$-cube
$Q_3$, whose strong chromatic index equals $6$. Since the
composition of covering projections is again a covering
projection, each $G_n$ with $n\ge 3$ covers $G_2$, and
therefore it is strongly $6$-edge-colorable by
Lemma~\ref{lm:lift}. Theorem~\ref{thm:main} now implies that
$\chis{G_n}=6$ for each $n\ge 2$. }
\end{remark}
\section{Equivalence of coverings and colorings}\label{sec:equiv}
We have proved that the odd graph $K(2k-1,k-1)$ admits a unique
strong $(2k-1)$-edge-coloring up to automorphism. It is
therefore natural to ask whether the same holds for the graphs
that cover it. As we shall see in this section, the answer is
negative, which at the first glance might seem to be rather
counter-intuitive.
We first show that the problem of finding two essentially
different strong $(2k-1)$-edge-colorings of a $k$-regular graph
is closely related to the problem of finding two non-equivalent
covering projections of the same $k$-regular graph onto the odd
graph $K(2k-1,\penalty0 k-1)$.
We call two edge colorings $\phi_1$ and $\phi_2$ of a
graph $G$ \textit{equivalent} if there exists an automorphism
$\alpha$ of $G$ such that $\phi_2=\phi_1\alpha$. Similarly, we
say that two covering projections $f_1\colon G_1\to G$ and
$f_2\colon G_2\to G$ are \textit{equivalent} if there exists an
isomorphism $\xi\colon G_2\to G_1$ such that $f_2= f_1\alpha$.
The following theorem shows that for strong
$(2k-1)$-edge-colorings of $k$-regular graphs equivalence of
colorings coincides with equivalence of coverings.
\begin{theorem}\label{thm:equivcol}
Every strong $(2k-1)$-edge-coloring $\sigma$ of a $k$-regular
graph $G$ determines a unique covering projection
$f_{\sigma}\colon G\to K(2k-1,k-1)$. Moreover, two such
colorings $\sigma$ and $\tau$ are equivalent if and only if the
corresponding covering projections $f_{\sigma}$ and
$f_{\tau}$ of $G$ are equivalent.
\end{theorem}
\begin{proof}
Recall that given a strong $(2k-1)$-edge-coloring $\sigma$ of
an arbitrary graph $G$ we have defined a covering projection
$f$ by sending an arbitrary vertex $v$ of $G$ to the vertex
$\sigma'(v)$, where $\sigma'$ is the derived vertex coloring of
$G$ with colors being the $(k-1)$-element subsets of the
$(2k-1)$-element set. This is the required covering projection
$f_{\sigma}$ corresponding to the coloring~$\sigma$.
Assume that $\sigma$ and $\tau$ are equivalent strong
$(2k-1)$-edge-colorings of $G$, and let $\alpha$ be an
automorphism of $G$ such that $\tau=\sigma\alpha$. It follows
that $\tau'=\sigma'\alpha$ and therefore, by the definition of
the covering projection corresponding to a strong
$(2k-1)$-edge-coloring, $f_{\tau}=f_{\sigma}\alpha$.
Conversely, if covering projections $f_{\sigma}, f_{\tau}\colon
G\to K(2k-1,k-1)$ are equivalent, then there exists an
automorphism $\beta$ of $G$ such that
$f_{\tau}=f_{\sigma}\beta$. The latter can be rewritten as
$\tau'=\sigma'\beta$. However, the way how the original strong
$(2k-1)$-edge-coloring can be reconstructed from the derived
vertex coloring implies that $\tau=\sigma\beta$, which means
that the colorings $\sigma$ and $\tau$ are equivalent.
\end{proof}
What remains to be done is to find a $k$-regular graph $G$ that
admits two non-equivalent covering projections on the Kneser
graph $K(2k-1,k-1)$. We do it for $k=3$, in which case
$K(2k-1,k-1)$ coincides with the Petersen graph. For this
purpose we need to recall several notions pertaining to the
theory of graph covers.
First of all, it will be convenient to regard each edge
(including the loops) as a pair of oppositely oriented
\textit{darts}. Each dart $x$ directed from $u$ to $v$ has its
unique \textit{inverse} $x^{-1}$ directed from $v$ to $u$. The
set of all darts of a graph $K$ is denoted by $D(K)$. The
\textit{symmetric group} of all permutations of the $d$-element
set $\{1,2,\ldots,d\}$ is denoted by $S_d$; it acts on
$\{1,2,\ldots,d\}$ on the right.
A \textit{permutation voltage assignment} on a graph $K$ is a
mapping $\kappa\colon D(K)\to S_d$ such that
$\kappa(x^{-1})=\kappa(x)^{-1}$ for each dart $x\in D(K)$. For
convenience, we often denote $\kappa(x)$ by $\kappa_x$. Every
permutation voltage assignment $\kappa$ on $K$ gives rise to
the \textit{derived graph} $K^{\kappa}$ for $K$, or the
\textit{lift} of $K$, which is defined as follows. Set
$V(K^{\kappa})=V(K)\times\{1,2,\ldots,d\}$,
$D(K^{\kappa})=D(K)\times\{1,2,\ldots,d\}$, and for each dart
$x=uv$ and $i\in\{1,2,\ldots,d\}$ let the lifted dart $(x,i)$
join $(u,i)$ to $(v,(i)\kappa_x)$. It is easy to see that the
natural projection $p_{\kappa}\colon K^{\kappa}\to K$ which
erases the second coordinate is a covering projection.
Moreover, a classical result of the theory of voltage graphs
states that every $d$-fold covering projection $\tilde K\to K$
is equivalent to the natural projection $K^{\kappa}\to K$ for a
suitable permutation voltage assignment $\kappa$ on $K$ with
values in $S_d$, see \cite[Theorem~2.4.5]{GrossTucker}.
Let us henceforth assume that the base graph $K$ is connected.
Pick a spanning tree $T$ of $K$ and let $r$ be an arbitrary
vertex of $K$, the \textit{root}. For any vertex $w$ of $K$ let
$T(w)$ denote the unique directed path from the root to $w$,
encoded as a sequence of darts, and let $T(w)^{-1}$ be the
inverse path. Further, for any dart $z=uv$ whose underlying
edge is not contained in $T$ define the permutation
$\theta(z)\in S_d$ by taking the product of voltages (that is,
values of $\kappa$) on the closed walk $T(u)zT(v)^{-1}$, rooted
at $r$, in the order determined by the walk and starting from
the root. If we set $\kappa'(x)=\theta(x)$ whenever $x$ is a
cotree dart and $\kappa'(x)={\mathrm{id}}$ otherwise, we obtain a new
voltage assignment $\kappa'$ on $K$ called the
\textit{$(T,r)$-reduction} of $\kappa$.
The concept of a $(T,r)$-reduction of a voltage assignment is
quite useful. For example, it can be shown that the derived
graph $K^{\kappa}$ is connected if and only if the voltages of
$\kappa'$ generate a transitive subgroup of $S_d$. What is more
important for us, it can be used to determine whether or not
two covering projections are equivalent. The respective result
is taken from \cite[Theorem~2]{Sko}.
\begin{theorem}\label{thm:equivcp}
Let $\kappa$ and $\lambda$ be permutation voltage assignments
on a connected graph $K$, both having their values in the
symmetric group $S_d$, and let $\kappa'$ and $\lambda'$ be
their $(T,r)$-reductions. The natural projections $p_{\kappa}$
and $p_{\lambda}$ are equivalent if and only if there exists an
inner automorphism $\gamma$ of $S_d$ such that
$\lambda'=\gamma\kappa'$.
\end{theorem}
Now we are prepared to describe an example of a cubic graph
that covers the Petersen graph and has two non-equivalent
strong $5$-edge-colorings. It is depicted in
Figure~\ref{fig:lift}.
\begin{example}\label{ex:non-equiv}{\rm
Consider the permutation voltage assignments $\kappa$ and
$\lambda$ on the Petersen graph with values in $S_4$ which are
represented in Figure~\ref{fig:pva}; the edges not labelled
carry the trivial voltage ${\mathrm{id}}$ (in both directions). The
values attached to all the edges are involutions, which means
that they unambiguously represent the respective voltage
assignments.
\begin{figure}[htp!]
$$
\includegraphics{fig_pva-1} \quad \quad \quad
\includegraphics{fig_pva-2}
$$
\caption{Permutation voltage assignments $\kappa$ and
$\lambda$ on the Petersen graph with values in $S_4$.}
\label{fig:pva}
\end{figure}
The corresponding lifts are isomorphic graphs as can be easily
detected from Figure~\ref{fig:lift}; for sim\-plicity, a vertex
$(v,i)$ is denoted by $v_i$.
\begin{figure}[htp!]
$$
\includegraphics{fig_lift-1} \quad \quad
\includegraphics{fig_lift-2}
$$
\caption{The lifts of the Petersen graph corresponding to the
permutation voltage assignments $\kappa$ and $\lambda$.}
\label{fig:lift}
\end{figure}
We claim that
the natural projections $p_{\kappa}\colon P^{\kappa}\to P$ and
$p_{\lambda}\colon P^{\lambda}\to P$, where $P$ denotes the
Petersen graph, are not equivalent. To see it, pick the
spanning tree $T$ indicated in Figure~\ref{fig:pva} by bold
lines and choose the central vertex $a$ to be the root.
Clearly, the $(T,a)$-reduction $\kappa'$ of $\kappa$ coincides
with $\kappa$. The $(T,a)$-reduction $\lambda'$ of $\lambda$ is
shown in Figure~\ref{fig:reduced}; the value $\lambda'(gf)$ is
not an involution and holds for the indicated direction.
\begin{figure}[htp!]
$$
\includegraphics{fig_reduced}
$$
\caption{The $(T,a)$-reduction $\lambda'$ of $\lambda$.}
\label{fig:reduced}
\end{figure}
By comparing $\kappa'$ and $\lambda'$ we immediately conclude
that no inner automorphism $\gamma$ of $S_d$ such that
$\lambda'=\gamma\kappa'$ can exist, because every inner
automorphism preserves the cycle structure of permutations.
Theorem~\ref{thm:equivcp} now implies that the covering
projections $p_{\kappa}$ and $p_{\lambda}$ are not equivalent.
Further, from Theorem~\ref{thm:equivcol} we conclude that there
exist strong $5$-edge-colorings $\sigma$ and $\tau$ of the
graph shown in Figure~\ref{fig:lift} such that $f_{\sigma}$ is
equivalent to $p_{\kappa}$ and $f_{\tau}$ is equivalent to
$p_{\lambda}$. They can be constructed simply by lifting the
strong $5$-edge-coloring $\sigma_3$ of the Petersen graph via
$p_{\kappa}$ and $p_{\lambda}$, respectively. By the same
theorem, the colorings $\sigma$ and $\tau$ are not equivalent.
}%
\end{example}
\section{Strong, normal, and Petersen colorings}\label{sec:normal}
Corollary~\ref{cor:main} links strong edge colorings of cubic
graphs to two other interesting types of edge colorings of
cubic graphs -- normal colorings and Petersen colorings -- and
through them to the outstanding Petersen coloring conjecture of
Jaeger \cite{Jae88}.
\begin{conjecture}[Jaeger, 1988]\label{conj:peter}
Every bridgeless cubic graph admits a Petersen coloring.
\end{conjecture}
For a cubic graph $G$ a mapping $\xi\colon E(G)\to E(P)$ is
said to be a \textit{Petersen coloring} if any two adjacent
edges of $G$ are mapped to adjacent edges of the Petersen
graph. As a consequence, for every vertex $v$ of $G$ the three
edges incident with $v$ are mapped to three edges incident with
a vertex of $P$ (as $P$ is triangle-free); in particular, $\xi$
is a proper edge coloring. Nevertheless, a Petersen coloring
need not be a homomorphism $G\to P$ because the induced mapping
between the vertex sets need not send adjacent vertices of $G$
to adjacent vertices of $P$ (for example, it can send them to
the same vertex). If $\xi$ does map adjacent vertices to
adjacent ones, then it determines a covering projection $G\to
P$. Conversely, every covering projection $G\to P$ gives rise
to a Petersen coloring of $G$.
In \cite[Section~5]{Jae85} Jaeger proved that a cubic graph
admits a Petersen coloring if and only if it has a
$5$-edge-coloring which he termed `normal'. A proper
$5$-edge-coloring $\phi$ of a cubic graph $G$ is said to be
\textit{normal} if for every edge $e$ of $G$ the number of
colors on the edges adjacent to $e$ is either $2$ or $4$, but
never $3$. In the former case, $e$ is called \textit{poor}, as
opposed to \textit{rich} (introduced in
Section~\ref{sec:main}), which corresponds to the latter case.
Clearly, a normal $5$-edge-coloring with no poor edges is
strong.
By combining these observations with Corollary~\ref{cor:main}
we obtain the following four equivalent statements.
\begin{theorem}\label{thm:equiv}
Let $G$ be cubic graph. Then the following statements are
equivalent:
\begin{itemize}
\item[$(i)$] $\chi_s'(G) = 5$;
\item[$(ii)$] $G$ covers the Petersen graph;
\item[$(iii)$] $G$ admits a normal $5$-edge-coloring in which
every edge is rich;
\item[$(iv)$] $G$ admits a Petersen coloring which is a
homomorphism.
\end{itemize}
\end{theorem}
\section{Conclusion}\label{sec:conclusion}
Coloring subcubic graphs with five colors without creating a
bichromatic path of length $3$ is very restrictive. For
example, we already need five colors when there are two
adjacent vertices of degree $3$ in a graph. Therefore, the
existence of infinitely many cubic graphs admitting a strong
$5$-edge-coloring is quite surprising.
\begin{figure}[htp!]
$$
\includegraphics{fig_K33} \quad \quad \quad
\includegraphics{fig_C6c} \quad \quad \quad
\includegraphics{fig_cub8}
$$
$$
\includegraphics{fig_Wag} \quad \quad \quad
\includegraphics{fig_K33s}
$$
\caption{Subcubic graphs with high strong chromatic indices;
the three graphs above need nine colors and the two graphs below
ten colors for any strong edge coloring.}
\label{fig:9}
\end{figure}
What if we have more available colors? It is already known that
the Heawood graph, a cubic graph on $14$ vertices of girth $6$,
has strong chromatic index $7$. Computational results reveal
that there are two and six cubic graphs of girth at least $6$
on $16$ and $18$ vertices, respectively, for which six colors
are not sufficient. It is even not enough to slightly increase
the girth. The Tutte $8$-cage, a cubic graph of girth $8$ on
$30$ vertices, does not admit any strong $6$-edge-coloring, nor
do five cubic graphs of girth $8$ on 38 vertices. This
encourages us to ask the following.
\begin{question}
Is there a constant $C$ such that every cubic graph of girth at
least $C$ admits a strong $6$-edge-coloring?
\end{question}
Perhaps one can find a characterization of cubic graphs
admitting a strong $6$-edge-coloring similar to the one
presented in this paper for five colors. However, it is likely
that the condition of being bipartite will not play any role in
answering the above question, because Tutte's $8$-cage is
bipartite.
Having seven colors available, it seems that cubic graphs
become strongly colorable as soon as they do not contain short
cycles. Our computational experiments show that all cubic
graphs of girth at least $5$ on at most $26$ vertices and all
bridgeless subcubic graphs of girth at least $5$ on at most
$18$ vertices admit a strong $7$-edge-coloring. We therefore
propose the following conjecture, which strengthens Case
$(5)$ of Conjecture~\ref{conj:main}.
\begin{conjecture}\label{conj:new}
Let $G$ be a subcubic graph of girth at least $5$. Then
$$
\chis{G} \le 7\,.
$$
\end{conjecture}
We remark that without restricting girth eight colors are still
not sufficient to color all bridgeless subcubic graphs (see
Figure~\ref{fig:9} for some examples of graphs that need more
colors). However, as suggested in \cite{HocLajLuz20} on the basis of
computational evidence, it seems that nine colors should
suffice to color all bridgeless subcubic graphs that are not
isomorphic to either $K_{3,3}$ with one subdivided edge or to
the $8$-vertex M\"obius ladder known as the Wagner graph.
\begin{conjecture}[Hocquard, Lajou, Lu\v{z}ar, 2020]
Every bridgeless subcubic graph not isomorphic to the Wagner
graph and the complete bipartite graph $K_{3,3}$ with one edge
subdivided admits a strong $9$-edge-coloring.
\end{conjecture}
In fact, exhaustive computational search indicates that
there are no bridgeless subcubic graphs on more than 12
vertices and strong chromatic index at least $9$ (apart from
the graphs in Figure~\ref{fig:9} and four graphs on 12
vertices). Thus we can strengthen the above conjecture to the
following.
\begin{conjecture}
If $G$ is a bridgeless subcubic graph on at least $13$
vertices, then
$$
\chis{G} \le 8\,.
$$
\end{conjecture}
\paragraph{Acknowledgment.}
The first author was partially supported by the Slovenian
Research Agency Program P1--0383 and the project J1--1692. The
second and the third author received partial support from
APVV--19--0308 and VEGA 1/0813/18. The fourth author was
supported by APVV--19--0153.
|
\section{Introduction}
\label{sec:Intro}
The understanding of the physical conditions that lead to the formation of new stars in galaxies is essential to understand their formation and evolution. There are basically two physical scenarios that drive the star formation in disk galaxies \citep[see a review in][]{Kennicutt_2012}. In the first scenario, the star formation is mostly controlled by the properties and amount of the interstellar medium (ISM). In this so-called 'bottom-up' picture, the local star formation rate is controlled completely by the amount of dense gas and structure of the molecular clouds \citep[e.g.,][]{Krumholz_2005}. On the other hand, in the so-called 'top-down' scenario, local star-formation is controlled largely by global dynamical events and dynamical timescales \citep[e.g.,][]{Silk_1997}. In the latter scenario, variations of star-formation is controlled by gravitational instabilities in the disk rather than cooling of molecular clouds, with no distinction between densities regimes of the molecular clouds that can affect the amount of newly formed stars. Another model that emerge from these two views is the self-regulation star-formation scenario in which the hydrostatic pressure from the baryonic component balances the feedback from newly formed massive stars reaching an equilibrium \citep[e.g., ][]{Cox_1981,Dopita_1985,Silk_1997}.
Using data from the PHANGS survey, \cite{Sun_2020} found that the dynamical pressure (i.e., the pressure due to self-gravity and external disk gravity) is in equilibrium with the turbulent pressure for most of their sampled molecular clouds located in nearby star-forming galaxies. Different mechanisms can be considered that could explain such equilibrium including momentum flux injection to the interstellar medium (ISM) from stellar feedback \citep[including supernovae explosions, stellar winds and radiation, e.g.,][]{Thompson_2005, Ostriker_Shetty_2011,Faucher-Giguere_2013} or/and gravitational instabilities \citep[e.g., ][]{Ibanez_2017,Krumholz_2016}. The injection of momentum due to stellar feedback into the ISM is a rather sporadic and very localized event that may not occur in every single location of the galaxy where molecular gas is available. Therefore, in the self-regulated framework, a star-forming galaxy can be considered as a quasi-steady-state system \citep{Ostriker_2010,Ostriker_Shetty_2011}. The equilibrium between the pressure and the star-formation feedback has to be considered on spatial scales significantly larger than the typical size of giant molecular clouds (few tenths of pc) and temporal scales larger than a cycle of star formation. Numerical simulations suggest that these scales are of the order of few kpc and few hundreds of Myr \citep[e.g.,][]{Kim_2017,Semenov_2017, Orr_2018}.
Recently, the relation between the star formation surface density and the pressure at kpc scales has been investigated extensively in different surveys including star-forming galaxies. Using a sample of 23 dwarf and disk galaxies included in the HERACLES survey, \cite{Leroy_2008} found a strong correlation between these two observables for radial bins. \cite{Herrera-Camus_2017} and \cite{Sun_2020} found similar results using spatially resolved dataset from samples of 31 and 28 star-forming galaxies in the nearby Universe ($D \lesssim 30$ Mpc) included in the KINGFISH and PHANGS surveys, respectively. In general, the strong correlation between the star formation surface density (\Ssfr) and the pressure is in agreement with a linear relation (i.e., power-law index close to one). Theory and numerical simulations suggest that this could be the case in the scenario in which supernovae feedback is the main supplier of pressure against the hydrostatic pressure \citep[e.g.,][]{Ostriker_Shetty_2011, Kim_2013}.
At those kpc scales \Ssfr exhibits a strong correlation with the surface gas density (\Sgas) in star-forming galaxies \citep[also known as Schmidt-Kennicutt relation, see, rSK,][ and reference therein]{Kennicutt_2012}. On the other hand, Integral Field Spectroscopy (IFS) used in large samples of star-forming galaxies has make clear the strong relation between \Ssfr and the stellar component of the baryonic mass at kpc scales, \Sstar\, known as the resolved star-formation main sequence \citep[rSFMS, see a review in ][and references therein]{Sanchez_2020ARAA}. Both baryonic components, tracing the local gravitational potential or similarly the hydrostatic pressure, may provide a better estimate of \Ssfr at kpc scales (e.g., Barrera-Ballesteros et al., submitted). Therefore we would like to explore whether a star-forming scaling relation that uses both components of the baryonic mass (like the hydrostatic mid-plane pressure, \Phyd) provides a better description of \Ssfr than those relations using individual components of the baryonic mass.
Among the recent efforts to explore the local properties of galaxies in the nearby Universe, the CARMA Extragalactic Database for Galaxy Evolution (EDGE) survey \citep{Bolatto_2017} has mapped the molecular gas in 126 galaxies observed IFS data from the Calar Alto Legacy Integral Field Area (CALIFA) survey \citep{Sanchez_2012}. This yields a dataset that allow us not only to study the interplay of the molecular gas with other observables derived from the optical spectra at kpc scales, but also to explore the impact of the global/integrated properties in the derived local scaling relations. The EDGE-CALIFA survey overcomes the so-called "cosmic variance" problems, meaning that it samples a sufficient volume of the Universe to truly represent the relations at local scales \citep[e.g.,][]{Diemer_2019}. Among other results, this survey has improved our understanding of how the molecular gas depletion time changes across the extension of galaxies \citep{Utomo_2017}, on how different local and global parameters affect \Ssfr using a multi-linear approach \citep{Dey_2019}, and on how we can characterize the molecular gas at kpc scales using the optical extinction \citep{BB2020}. Using the EDGE-CALIFA spatially resolved dataset, we explore in this article the relation between \Ssfr and \Phyd. We find that these two parameters are strongly correlated, suggesting that star-formation at kpc scales, in a significant fraction of regions located in galaxies in the nearby Universe, is consistent with the self-regulation scenario \citep[e.g., ]{Elmegreen_1989,Silk_1997}. We explore the impact of local and global observables in this relation as well as its role in comparison to other star-forming scaling relations. This paper is organized as follows. In Sec.~\ref{sec:Data} we present the spatially-resolved and ancillary data used in this study which is available in the \texttt{edge\_pydb} database (Wong et al. in prep.). In Sec.~\ref{sec:Results} we show the main results of this paper. In Sec.~\ref{sec:Disc} we discuss these results. Finally, in Sec.~\ref{sec:Conclusions} we present the main conclusions of this article.
\section{Data and Analysis}
\label{sec:Data}
\begin{figure*}
\includegraphics[width=\linewidth]{figures/Sample_comp_v1.pdf}
\caption{Comparison of the 96 EDGE-CALIFA galaxies used in this study (empty circles) with different samples where the spatially-resolved \mbox{\Ssfr - \Phyd} relation has been derived. Blue diamonds, and green squares represent the samples from the THINGS \citep{Leroy_2008}, and PHANGS surveys \citep{Sun_2020}, respectively. The samples are presented in the SFR-M$_\mathrm{\ast}$ (left panel), the SFR-M$_\mathrm{mol}$ (middle panel), and the M$_\mathrm{\ast}$-M$_\mathrm{mol}$ planes (right panel). The solid lines represent the best linear fits to the data points in our sample. The inset in the middle panels shows the distribution of morphologies for this sample (empty bins), the THINGS+HERACLES (blue bins) and the PHANGS (green bins) surveys, respectively. Our sample covers a significant dynamic range in global properties for a variety of morphological types, which is essential to study the impact of global parameters on the \mbox{\Ssfr - \Phyd} relation at kpc scales.
}
\label{fig:Sample}
\end{figure*}
\subsection{The CALIFA and EDGE surveys}
\label{sec:CALIFA-EDGE}
Here we provide a brief description of the optical Integral Field Spectroscopy (IFS) CALIFA and the molecular gas EDGE surveys, respectively. Observations from both surveys comprise the core datasets that we use in this study to perform our analysis.
The CALIFA (Calar Alto Legacy Integral Field Area) survey \citep{Sanchez_2012} provides IFS data for more than 600 galaxies in the nearby Universe (0.005$< z < $0.03) using the PMAS/PPAK Integral Field Unit (IFU) instrument \citep{Roth_2005} mounted at the 3.5 m telescope of the Calar Alto Observatory. The instrument consists of 331 fibers of 2\farcs7 diameter each, concentrated in a single hexagon bundle covering a field-of-view (FoV) of $\sim$ 1 arcsec$^2$ with a filling factor of $\sim$ 60\%\,. To provide a full coverage of the FoV a three-point dithering is performed. The average resolution of this instrument is $\lambda$/$\Delta\lambda$\,$\sim$\,850 at $\sim$5000\AA\, with a typical wavelength range from 3745 to 7300\AA. The isophotal diameter of CALIFA galaxies are in the range 45 $\lesssim D_{25}\lesssim$\,80 arcsec in the SDSS (Sloan Digital Sky Survey) $r$-band \citep{Walcher_2014}. The data reduction is performed by a pipeline designed specifically for the CALIFA survey. The final data cube for each galaxy consists of more than 5000 spectra with a sampling of 1$\times$1 arcsec$^2$ per spaxel. The reduction process is described in detail in \cite{Sanchez_2012}, improvements on the reduction pipeline as well as extensions to the original sample (reaching a total of 834 galaxies) are presented by \cite{Husemann_2013,GarciaBenito_2015,Sanchez_2016}.
The EDGE survey obtained millimeter-wave interferometric observations for a sample of 126 galaxies included in the CALIFA survey. These observations were carried out at the Combined Array for Millimeterwave Astronomy \citep[CARMA, ][]{Bock_2006}. The EDGE survey provides the first effort to combine resolved CO data with IFS optical data for a significant sample of galaxies representative of the local Universe. We present a brief description of the survey here, see \cite{Bolatto_2017} for a detailed description. Galaxies were observed using half-beam-spaced seven-point hexagonal mosaics yielding a half-power field-of-view of radius $\sim$ 50\arcsec. Each galaxy has been observed in both the E and D-array configuration with integration times typically of 40 min and 3.5 hr, respectively. The typical resolution for each configuration is 8 and 4 arcsecs, respectively. The final maps combined the E and D array observations resulting in a velocity resolution of 20 km s$^{-1}$ with a typical angular resolution of 4.5\arcsec and typical rms sensitivity of 30 mK at the velocity resolution. Assuming a Milky-Way constant CO-to-\htwo\, conversion factor \citep[$\alpha_{\rm CO} = 4.3\,{\rm M_{\odot} (K\,km\,s^{-1}\,pc^{2})^{-1}} $,][]{Bolatto_2013}, the survey is sensitive to an \htwo\, surface mass density of $\sim\,4-110\,\,\mathrm{M_{\odot}\,\,pc^{-2}}$ (averaged over a $\sim$ 1.5 kpc scale). The data cubes are smoothed and then masked in order to distinguish CO signal from noise and to reach higher signal to noise \citep[see more details in][]{Bolatto_2017}.
\subsection{The \texttt{edge\_pydb} database}
\label{sec:edgepy}
As we show above, the spatially resolved dataset from the IFU CALIFA survey and the dataset from the CO millimeter obtained from CARMA array are relatively different in terms of field-of-view coverage, spatial resolutions and spatial sampling. The \texttt{edge\_pydb} database has been created as a homogeneous source of optical and millimeter maps and data for the 126 EDGE galaxies to be used in a flexible python environment that allows a simple yet robust exploration of the EDGE-CALIFA dataset. This database also provides an integration of external properties with the spatially resolved information. A detailed description of the database can be found in Wong et al. (in prep.). Here we highlight its main features and the data used in this analysis.
This database provides different estimates of the CO moments from the CARMA observations. The database also provides a smoothed and masked version of the CARMA CO datacubes. The CO datacubes are integrated in the velocity axis to obtain the surface brightness maps (\texttt{smo} table). Both CALIFA and EDGE datasets are convolved to the same spatial resolution (i.e., 7\arcsec). Then the \textsc{Pipe3D} data analysis pipeline \citep{Sanchez_2015, Sanchez_2016_pipe3d} is run over the convolved optical datacubes resulting in two dimensional maps of optical properties with the same resolution as the CO surface brightness maps. By fitting the stellar continuum (using a single stellar population fitting, SSP, adopting a \cite{Salpeter_1955} initial Mass Function, IMF) and the emission lines for each of the spaxels in each datacube, this pipeline extracts two dimensional maps of a given stellar or ionized gas observable. The maps are sampled on a square grid spaced by 3\arcsec\, in RA and DEC after interpolation to the CARMA WCS. The grid is shifted so that the reference pixel is retained. The tables with ancillary data include information from the LEDA and NED databases, IR photometry from the WISE survey among others. The database also include the information from integrated properties derived from the CO and optical datacubes. For each of the galaxies in this study, we use the values from the database of: the total molecular gas mass, (M$_\mathrm{mol}$), the total stellar mass (M$_\mathrm{\ast}$), the integrated star formation rate (SFR), the effective radius (R$_{\rm eff}$), the stellar scale height ($l_s$), the minor-major axis ratio ($b/a$), and the morphology. The reader is addressed to Wong et al. (in prep.) for a detailed description on how these observables are derived or from which database they have been obtained.
\subsection{Derived Quantities}
\label{sec:Quants}
We use the maps of the molecular gas density, \Smol, for each galaxy from the \texttt{edge\_pydb} database. These maps are obtained by converting the CO surface brightness maps from the \texttt{smo} list into molecular gas mass density maps using a constant CO-to-\htwo\, conversion factor following \cite[][$\alpha_{\rm CO} = 4.3\,{\rm M_{\odot} (K\,km\,s^{-1}\,pc^{2})^{-1}}$]{Bolatto_2013}. This factor includes the mass contribution from helium (below we also estimate a variable conversion factor).
The \texttt{edge\_pydb} database provides all the maps from both the stellar continuum and the ionized gas components derived from the \textsc{Pipe3D} pipeline. From the fitting of the stellar continuum we use the stellar surface mass density map (\Sstar), the stellar age and metallicity ([Z/H]), and the stellar velocity dispersion ($\sigma_{\ast}$) for each of the sampled regions. From the analysis of emission lines of the ionized gas we use the integrated flux maps of \ha, \hb, \mbox{\rm [O{\small III}]} and \mbox{\rm [N{\small II}]} emission lines. We also use the equivalent width map of the \ha\, emission line \citep[EW(\ha), see details in ][]{Sanchez_2016_pipe3d}. From these emission line fluxes we derive the Balmer decrement ratio (\ha/\hb). From this ratio and the \ha\, emission line, we obtain the extinction-corrected star formation rate surface density map, \Ssfr \citep{Kennicutt_1998_SFR}. All the surface densities are corrected by the galaxy's inclination following \cite{BB_2016}. For this study, we adapted both the \Ssfr, and \Sstar densities to a Kroupa IMF. This is equivalent to multiply these quantities in the database (Salpeter IMF) by a factor of 0.61 \citep[see][review]{Madau_Dickinson_2014}. As proxy for the inclination we use the $b/a$ axis ratio. The typical relative error of \Smol, \Sstar, and \Ssfr\, are $\sim$ 0.28, 0.15, and 0.20 dex, respectively.
We follow \cite{Elmegreen_1989} to derive for each region the mid-plane hydrostatic pressure\footnote{Note that we use the traditional term `hydrostatic', even though the gas in the ISM is not static, with the majority of the pressure associated with turbulence} (\Phyd):
\begin{equation}
\label{eq:Phyd}
\Phyd = \frac{\pi}{2}\, G\, \Sgas \left( \Sgas + \frac{\sigma_{\rm gas}}{\sigma_{\rm \ast,z}}\, \Sstar \right),
\end{equation}
where $G$ is the gravitational constant, \Sgas and \mbox{$\sigma_{\rm gas}$} are the total gas surface mass density (molecular and atomic \Sgas = \Smol+\Shi) and the total gas velocity dispersion, respectively; and \mbox{$\sigma_{\rm \ast,z}$} is the stellar velocity dispersion in the axis perpendicular to the disk. Since we do not have direct observations of the atomic gas density distribution for these galaxies, to derive the best fit of the \Ssfr - \Phyd\, relation in Sec.~\ref{sec:SPrel} we run a Monte Carlo simulation assuming 1000 realizations with different values of \Shi within a normal distribution centered at \Shi = 7 \msunperpcsq and with a standard deviation of 2 \msunperpcsq. These range of values are representative of the atomic gas densities found in normal star-forming galaxies \citep[e.g.,][]{Bigiel_2008}. The value of \Phyd\, for each region presented in this study is thus the average of the above realizations. We adopt a constant value of the total gas velocity dispersion \mbox{$\sigma_{\rm gas}$ = 11 km s$^{-1}$}. This value is within the range of typical velocity dispersions found in disk galaxies for both components of the cold gas in the nearby Universe \citep[e.g.,][]{Caldu-Primo_2013, Levy_2018} and have been adopted in different studies of the \mbox{\Ssfr - \Phyd} relation\citep[e.g., ][]{Leroy_2008,Sun_2020}. To estimate \mbox{$\sigma_{\rm \ast,z}$}, we follow \cite{Leroy_2008, Zheng_2013}. By assuming a relation between this dispersion and the disk scale height ($h_s$) and the stellar mass surface density for an isothermal disk \citep{van_der_Kruit_1988} we have
\begin{equation}
\label{eq:sz}
\sigma_{\rm \ast,z} = \sqrt{2\,\pi\,G\,\Sstar h_s},
\end{equation}
here we assume that the EDGE-CALIFA galaxies follow the ratio \mbox{$l_s/h_s$} = 7.3, where $l_s$ is the disk stellar scale length \citep{Kregel_2002, Sun_2020}. In turn, we relate the stellar scale length $l_s$ of the disk to the effective radius of the galaxies in units of pc by adopting \mbox{$l_s$ = R$_{\rm eff}$/1.68} \cite[i.e., assuming a Sersic profile with $n$ = 1, ][]{Graham_2005}. We adopt this relation in order to provide an estimation of $l_s$ in larger samples of galaxies where only \mbox{R$_{\rm eff}$} has been determined. In Appendix \ref{app:Phyd} we compare the \mbox{$\sigma_{\rm \ast,z}$} ratio assuming different estimations of $l_s$.
In Sec.~\ref{sec:caveats} we study the impact of a variable \mbox{$\mathrm{\alpha_{CO}}$} conversion factor in the estimation of the hydrostatic mid-plane pressure. We use Eq.~7 from \cite{Colombo_2020} to estimate this variable conversion factor:
\begin{equation}
\mathrm{\alpha_{CO}}(Z^{\prime},\Sstar) = 2.9 \exp \left(\frac{0.4}{Z^{\prime}}\right) \left( \frac{\Sstar}{100 \msunperpcsq} \right)^{-\gamma},
\label{eq:aCO}
\end{equation}
where \mbox{$\mathrm{\alpha_{CO}}(Z^{\prime},\Sstar)$} is in units of \mbox{${\rm M_{\odot} (K\,km\,s^{-1}\,pc^{2})^{-1}}$}, \mbox{$Z^{\prime}$} is the ionized gas metallicity relative to the solar one, and $\gamma$ = 0.5 where \mbox{$\Sstar\, > 100 \msunperpcsq$} and $\gamma$ = 0 otherwise. This is a variation of the variable conversion factor derived in Eq.~31 by \cite{Bolatto_2013}. Following \cite{Colombo_2020}, we assume that the total density in our regions is dominated by \Sstar ($\Sigma_{\rm total} \sim \Sstar$), also that the Giant Molecular Cloud (GMC) molecular gas surface density in units of 100 \msunperpcsq\, is equal to one. The ionized gas metallicity is obtained by using the strong-lines calibrator derived by \cite{Marino_2013}:
\begin{equation}
12 + \log({\rm O/H}) = 8.533 - 0.214 \left( \frac{{\rm [O{\small III}]}}{\hb} \times \frac{\ha}{{\rm [N{\small II}}]} \right).
\end{equation}
To provide a reliable comparison with the literature, in Sec.\ref{sec:comp} instead of \Phyd we use a slightly different estimate of the dynamical equilibrium pressure, \Pde. Following \citet{Kim_2011}, and \citet[][]{Sun_2020} \Pde\, is given by:
\begin{equation}
\label{eq:Pde}
\Pde = \frac{\pi G}{2} \Sigma^2_{\rm gas} +\Sgas\sqrt{2G\rho_{\ast}}\,\sigma_{\rm gas,z} ,
\end{equation}
where $\rho_{\ast}$ is the mid-plane stellar volume density \citep[][see their Eq.~13]{Sun_2020}. Here, $\sigma_{\rm gas, z}$ is the velocity dispersion of the gas perpendicular to the disk, and we adopt a value 11 km/s as above. \cite{Sun_2020} noted that this estimate can slightly increases the measurement of \Pde\, by a factor at most of $\sim$ 0.2 dex in comparison to those dynamical pressures derived directly from their observations. We perform a similar analysis as in Sec.~\ref{sec:SPrel}, this is, we assume 1000 different realizations of \Shi. We presented in this article the averaged values of \Pde\, from those realizations.
The \texttt{edge\_pydb} database provides the information derived from the CALIFA and EDGE data for over $\sim$16000 individual regions located in the 126 galaxies included the survey. For the analysis in this article we select those star-forming regions with the most reliable estimations of the considered observables. This is, regions where \Smol, \ha, and \hb\, have values larger than three times their errors as well as non-zero estimates of their \Sstar\, and their \Ssfr and EW(\ha)$>$ 6\AA. This selection yields a total of 4260 regions located in 96 galaxies. In Fig.~\ref{fig:Sample} we compare the EDGE-CALIFA galaxies used in this study with other angular-resolved samples used to estimate the \Ssfr - \Pde\, relation in the SFR-M$_\mathrm{\ast}$, in the SFR-M$_\mathrm{mol}$, and in the M$_\mathrm{\ast}$-M$_\mathrm{mol}$ planes. The dynamical range of our sample in these three observables is wide, covering $\sim$ 2 orders of magnitude for each of them. In comparison to our sample, the THINGS sample \citep{Leroy_2008} covers a range of low-mass galaxies whereas the 28 galaxies from the PHANGS sample from \cite{Sun_2020} covers a similar range of properties as our sample \footnote{estimates of M$_\mathrm{mol}$ for the PHANGS survey have been kindly provided by A. K. Leroy (private communication). Full details will be available in Leroy et al. (in prep.). SFR, and M$_\mathrm{\ast}$ values are taken from the website: \texttt{https://sites.google.com/view/phangs/home/sample}}, however the EDGE-CALIFA galaxies includes a wide range in morphologies (see inset in the middle panel of Fig.~\ref{fig:Sample}). In Sec.~\ref{sec:global} we explore how the radial distribution of the \Ssfr - \Phyd\, relation varies for galaxies with different global star-formation activity and global gas fraction derived from the best fits of the above relations (black lines in each panel of Fig.~\ref{fig:Sample}). For our sample of galaxies in the SFR-M$_\mathrm{\ast}$, the SFR-M$_\mathrm{mol}$, and in the M$_\mathrm{\ast}$-M$_\mathrm{mol}$ planes, these fits correspond to: \mbox{$\mathrm{\log(SFR/M_{\odot}\,yr^{-1}) = -7.0 + 0.68\, \log(M_{\ast}/M_{\odot})}$}, \mbox{$\mathrm{\log(SFR/M_{\odot}\,yr^{-1}) = -6.3 + 0.70\, \log(M_{mol}/M_{\odot})}$}, and \mbox{$\mathrm{\log(M_{mol}/M_{\odot}) = 0.88\,\log(M_{\ast}/M_{\odot})}$}, respectively. Furthermore, in Sec.~\ref{sec:comp} we compare the estimate of the \Ssfr - \Pde\, relation using our sample and those derived using samples with resolved measurements.
\section{Results}
\label{sec:Results}
\subsection{The \Ssfr - \Phyd\, relation at kpc scales for the EDGE-CALIFA galaxies.}
\label{sec:SPrel}
\begin{figure}
\includegraphics[width=\linewidth]{figures/Pism_SFR_EDGE_v2.pdf}
\caption{
The \Ssfr - \Phyd\, relation for 4260 star forming regions included in 96 galaxies from the EDGE-CALIFA survey. The data points are color-coded according to the density of points. Inwards, the black contours enclose 90\%, 80\% and 50\% of the sample. The solid line represents the best ODR linear fit. The shaded area represents the uncertainty of the fit due to the assumed values of \Shi. The tightness of the relation is indicated by the small standard deviation ($\sigma$), whereas $r$ represents the Pearson correlation coefficient.
}
\label{fig:SPrel}
\end{figure}
In Fig.~\ref{fig:SPrel} we show the \Ssfr - \Phyd relation for our dataset that includes 4260 regions located in 96 EDGE-CALIFA galaxies. The hydrostatic mid-plane pressure covers $\sim$ four orders of magnitude whereas \Ssfr covers $\sim$ 3 orders of magnitude. On a log-log scale, \Ssfr increases linearly with \Phyd, resulting in a strong correlation (Pearson correlation coefficient of \mbox{$r$ = 0.85}). Similar but smaller correlation coefficients have also been observed in other star-forming scaling relation at kpc scales \citep[e.g., the rSFMS or the rSK, ][]{Cano-Diaz_2016, Lin_2019,Cano-Diaz_2019}.
To obtain the best parameters that represent the \Ssfr - \Phyd\, relation we fit the following relation to our dataset using an orthogonal distance regression (ODR) fitting technique:
\begin{equation}
\label{eq:bestfit}
\frac{\Ssfr}{10^{-3} \msunperyrkpcsq} = A \left( \frac{\Phyd}{10^{4}\,\, k_{\rm B}\,\, {\rm K\,\, cm^{-3}}} \right)^{b},
\end{equation}
the fitting procedure was repeated 1000 times for each of the different realizations performed to take into account our lack of knowledge of the \Shi distribution (see Sec.~\ref{sec:Quants}). The shaded area in Fig.~\ref{fig:SPrel} represents the best fits from these realizations and the black solid line represents their median. Therefore the reported values of the fit are obtained from this average while their uncertainties come from the scatter of these realizations (\mbox{$b$ = 0.95$\pm$0.05} and \mbox{$A$ = 0.3$\pm$0.1}). We also perform an ordinary least-square (OLS) fit to the average values. We find a slightly flatter relation in comparison to the one derived using the ODR fit (\mbox{$b$ = 0.83$\pm$0.04} and \mbox{$A$ = 0.37$\pm$0.05}). In Sec.~\ref{sec:comp}, we compare these slopes with recent results for regions at sub-kpc scales. The scatter of the residuals -- measured from their standard deviation, $\sigma$ -- is small compared to other star-forming scaling relations ($\sigma$ = 0.27, see a comparison in Sec.~\ref{sec:local}). We are interested in exploring how other local and global parameters affect this relation. In Secs.~\ref{sec:local} and \ref{sec:global} we study how the residuals from the best fit of this relation correlate with other observables. In Sec.~\ref{sec:Disc} we discuss the possible different scenarios that can explain the sub-linear slope exhibited by this relation.
\begin{figure*}
\includegraphics[width=\linewidth]{figures/res_Local_v3.pdf}
\caption{The correlation between the residuals of different scaling relations with those derived from the \Ssfr - \Phyd\, relation, \mbox{$\Delta {\rm SFHP}$}. ({\it left}) The residuals of the Schmidt-Kennicutt law (\mbox{$\Delta {\rm SK}$}). ({\it middle}) The residuals of the resolved star formation main sequence (\mbox{$\Delta {\rm SFMS}$}). ({\it right}) The residuals of the molecular gas main sequence (\mbox{$\Delta {\rm MGMS}$}). The number in each panel represents the Pearson correlation coefficient. Solid lines in left and middle panels represent a one-to-one relation. In each panel the blue solid contour encloses $\sim$ 64\% of the sample while the black dashed contour represents the same fraction for those residuals estimated from mock data assuming a constant scatter around best-fit scaling relations (see details in Sec.~\ref{sec:local}). This comparison suggests that correlations between residuals are mostly explained by the strong covariance of the quantities defined in each axis.
}
\label{fig:SPrel_local}
\end{figure*}
\subsection{Impact of Local parameters in the \Ssfr - \Phyd\, relation}
\label{sec:local}
In the last decade it has become evident that most of the star-forming scaling relations derived for integrated properties are also observable at kpc scales \citep[for a recent review see][]{Sanchez_2020ARAA}. Recent studies have also shown that scatter of the star-forming scaling relations is modulated by different local observables \citep{Ellison_2018,Ellison_2020}. In the previous section we showed that \Ssfr strongly correlates with \Phyd. Furthermore this relation is tight, exhibiting a similar scatter in comparison to other star-forming scaling relations ($\sim$ 0.25 dex). In this section, we explore how the \Ssfr - \Phyd\, relation compares with the three scaling local relations that correlate \Ssfr, \Sstar, and \Smol among them at kpc scales. Then we explore how the \mbox{\Ssfr - \Phyd} relation is modulated by other stellar properties such as the age, metallicity, and their velocity dispersion.
Using a sample of galaxies included in the ALMAQuest survey, \cite{Lin_2019} found that within star-forming galaxies \Ssfr, \Sstar, and \Sgas closely correlate with each another. On the one hand, \Ssfr correlates with \Sstar \citep[resolved star-formation main sequence, rSFMS; see also ][]{Sanchez_2012, Wuyts_2011,Cano-Diaz_2016, Cano-Diaz_2019}, and with \Smol \citep[resolved Schmidt-Kennicutt relation, rSK; see also ][]{Bigiel_2008}. On the other hand, \Smol\, also correlates with \Sstar \citep[resolved molecular gas main sequence, rMGMS, see also ][]{BB2020}. We derive these scaling relations using our dataset from the EDGE-CALIFA survey in S\'anchez et al. (in prep.). The following are the best relations presented in S\'anchez et al. (in prep.) for this survey:
the resolved Schmidt-Kennicutt relation (rSK):
\begin{equation}
\label{eq:rSK}
\log \left(\frac{\Ssfr}{\rm M_\odot~yr^{-1} pc^{-2}} \right) = 0.98\,\, \log \left( \frac{\Smol}{\msunperpcsq} \right) -9.01,
\end{equation}
the resolved star formation main sequence (rSFMS):
\begin{equation}
\label{eq:rSFMS}
\log \left(\frac{\Ssfr}{\rm M_\odot~yr^{-1} pc^{-2}}\right) = 1.02\,\, \log \left( \frac{\Sstar}{\msunperpcsq} \right) - 10.10,
\end{equation}
and the resolved molecular gas main sequence (rMGMS):
\begin{equation}
\label{eq:rMGMS}
\log \left( \frac{\Smol}{\msunperpcsq} \right) = 0.93\,\, \log \left( \frac{\Sstar}{\msunperpcsq} \right) -0.91,
\end{equation}
their typical scatter is of the order of $\sigma \sim$ 0.25 dex (see details in S\'anchez et al., in prep.). For each of these relation we derive their residuals (i.e., for each star-forming region, the distance in the $y$-direction between the best fit and the data point). In each of the panels in Fig.~\ref{fig:SPrel_local} we plot these residuals against the one derived from the best fit of the \mbox{\Ssfr - \Phyd} relation (\mbox{$\Delta {\rm SFHP}$}): the residuals of the rSK (\mbox{$\Delta {\rm SK}$}, left panel), the residuals of the rSFMS (\mbox{$\Delta {\rm SFMS}$}, middle panel), and the residuals of the rMGMS (\mbox{$\Delta {\rm MGMS}$}, left panel). In principle, \mbox{$\Delta {\rm SK}$}, and \mbox{$\Delta {\rm SFMS}$}\, quantify the lack (or excess) of \Ssfr for a given \Sgas, and \Sstar, respectively; whereas the \mbox{$\Delta {\rm MGMS}$}\, measures the lack or excess of \Smol for a given \Sstar
We find that \mbox{$\Delta {\rm SFMS}$}, and \mbox{$\Delta {\rm SK}$}\, correlate with \mbox{$\Delta {\rm SFHP}$}. This is supported by their Pearson correlation coefficients ($r$ = 0.94, and 0.54, respectively). Even more, these two residuals increases linearly with \mbox{$\Delta {\rm SFHP}$}\, following a one-to-one relation (dashed lines in Fig.~\ref{fig:SPrel_local}). This may suggest, that for a given \Phyd the variations in \Ssfr are tightly correlated to those variations expected from the rSK and rSFMS. The smaller scatter observed in the left panel of Fig.~\ref{fig:SPrel_local} compare to the one observed in middle panel may indicate that \mbox{$\Delta {\rm SK}$}\, has a larger impact in setting \Ssfr for a given \Phyd. In comparison to the previous correlations the residuals of the molecular gas main sequence, \mbox{$\Delta {\rm MGMS}$}, weakly anti-correlates with \mbox{$\Delta {\rm SFHP}$} ($r$ = -0.31, right panel of Fig.~\ref{fig:SPrel_local}). The lack of a strong correlation between these residuals suggests the small impact that the gas fraction has in modulating the star formation at kpc scales in comparison to the stellar and gas surface densities.
Mathematically speaking, the observables from which the previous scaling relations are made of (i.e., the rSK, the rSFMS, and the rMGMS), are the same observables used to derive the \mbox{\Ssfr - \Phyd} relation. Therefore, the correlations among the residuals we find in Fig.~\ref{fig:SPrel_local} can be induced by the strong covariance among those observables. To test the impact of their covariances in the relations of the residuals, in Appendix~\ref{app:Mock} we build the same relations among residuals using a mock dataset by considering the best fits of the scaling relations (i.e., Eqs., \ref{eq:rSK}, \ref{eq:rSFMS}, and \ref{eq:rMGMS}) and assuming the typical scatter from the observables. In each panel of Fig.~\ref{fig:SPrel_local} the blue contour encloses $\sim$64\% of the observed correlation by the residuals whereas the black dashed contours encloses the same fraction from the relations derived from the mock dataset. The comparison between these two distributions shows that the correlation we find between \mbox{$\Delta {\rm SK}$}\, and \mbox{$\Delta {\rm SFHP}$}\, is mostly driven by the strong covariance of the observables (left panel of Fig.~\ref{fig:SPrel_local}). The Pearson correlation coefficient derived from the mock dataset is similar as the one obtained from observations ($r$ = 0.85). On the other hand, the distribution of the observed relation between \mbox{$\Delta {\rm SFMS}$}\, and \mbox{$\Delta {\rm SFHP}$}\, is wider in comparison to one obtained using the mock dataset (middle panel of Fig.~\ref{fig:SPrel_local}). This suggests that the observed correlation between these two residuals is not entirely driven by the covariance of the observables. Finally, the distribution of the observed relation between \mbox{$\Delta {\rm MGMS}$}\, and \mbox{$\Delta {\rm SFHP}$}\, is slightly tighter than the one obtained from the mock dataset (right panel of Fig.~\ref{fig:SPrel_local}). The Pearson correlation coefficient derived from the mock dataset shows a larger anti-correlation between residuals is larger as the one obtained from observations ($r$ = -0.53). The lower correlation coefficient and the smaller scatter observed between \mbox{$\Delta {\rm MGMS}$}\, and \mbox{$\Delta {\rm SFHP}$}\, -- compare with those derived from the mock data -- indicate that \Phyd is an appropriate parameter to describe \Ssfr than the gas fraction. In Sec.~\ref{sec:Disc} we further discuss the impact of individual components of the baryonic mass in the \mbox{\Ssfr - \Phyd} relation.
\begin{figure*}
\includegraphics[width=\linewidth]{figures/Ssfr_Ph_StellarProp_hc1.pdf}
\caption{The impact of various stellar properties on the \Ssfr - \Phyd\, relation shown in Fig.~\ref{fig:SPrel}. In the three panels, the insets show the scatter of the relation as a function of the stellar parameter. Left, middle, and right panels show the \Ssfr - \Phyd\, relation color coded by the average stellar age, average stellar age ([Z/H]), and stellar velocity dispersion ($\sigma_{\ast}$), respectively. The residuals of the \Ssfr - \Phyd\, relation (\mbox{$\Delta {\rm SFHP}$}) are anti-correlated with both stellar properties; however \mbox{$\Delta {\rm SFHP}$}\, does not seem to be strongly affected by the stellar velocity dispersion.
}
\label{fig:SPrel_Z}
\end{figure*}
Thanks to the \texttt{edge\_pydb} we are able to estimate how stellar properties affect the \mbox{\Ssfr - \Phyd} relation. In the left, middle and right panels of Fig.~\ref{fig:SPrel_Z} we color code Fig.~\ref{fig:SPrel} with respect to the luminosity-weighted stellar age, stellar metallicity ([Z/H]), and the line-of-sight stellar velocity dispersion ($\sigma_{\ast}$), respectively. We find a significant anti-correlation of the stellar age and metallicity with \mbox{$\Delta {\rm SFHP}$}\, ($r$ = -0.25, and -0.40, respectively). For a given \Phyd, \Ssfr decreases as the age or [Z/H] increases. This suggests that for those regions where the hydrostatic mid-plane pressure is similar, those with young/metal-poor stellar population tend to have large \Ssfr in comparison to those with old/ metal-rich population. These two anti-correlations could emerge from the stellar velocity dispersion. Regions with large/low star formation rate for a given \Phyd\, could be dynamically cold/hot (i.e., low/high $\sigma_{\ast}$). However, in the right panel of Fig.~\ref{fig:SPrel}, when we color code the \mbox{\Ssfr - \Phyd} relation with respect to $\sigma_{\ast}$, we do not find similar patterns as those observed with the other two stellar properties in the other panels. Even more, the residuals of this relation do not seem to show a correlation with $\sigma_{\ast}$ ($r$ = 0.01). The spectral resolution of the CALIFA survey only allows to have a reliable estimation of $\sigma_{\ast}$ at high values (i.e.,$\sim$ 60 km s$^{-1}$), but even for large velocity dispersions we do not find a variation of the \mbox{$\Delta {\rm SFHP}$}\, against $\sigma_{\ast}$.
In Sec.~\ref{sec:global} we find that the scatter of the \mbox{\Ssfr - \Phyd} relation apparently (anti-)correlates with the (total stellar mass) and morphology. In other words, at a fixed \Phyd, \Ssfr\, is higher, for lower ages and metallicities, mainly in low-mass galaxies. Since the stellar metallicity and age are features of the underlying stellar populations, both hint past properties of the ISM, however the \Ssfr\, is a feature of the current ISM. Therefore, it seems that the galactic areas where the younger and metal-poor \hii\, regions are embedded, behave as the outer disk of early-type spiral galaxies (Sa-Sb) or as the inner disk of late-type spiral galaxies (Sc - Sd). These behaviors can be explained by the inside-out formation scenario for the former, and by the outside-in scenario for the latter. In both cases, their star formation histories could be roughly represented as an increasing exponential function. It can also be the case that the observed \mbox{$\Delta {\rm SFHP}$} is affected by our estimation of \Ssfr. The IMF varies with SFR and metallicity, such that very massive stars form when the SFR is high, regardless of the stellar metallicity, but for super solar metallicities the formation of low mass stars dominates, regardless of the SFR. The \Ssfr\, considered in this work is computed using the Kennicutt \Ssfr–\ha\, relation \citep{Kennicutt_1998}. This calibration was obtained by assuming an invariant IMF for stars of solar metallicity with masses between 0.1 and 100 M$_{\odot}$. Therefore this fact may reduce the dispersion of the \mbox{\Ssfr - \Phyd} relation and incidentally it could also modify the slope of other scaling relations such as the rSK and rSFMS \citep[e.g.,][]{Jerabkova_2018}.
We cannot rule out whether these secondary trends could be induce by expected relations between the components of \Phyd and \Ssfr. Although we do not find strong correlations among \Ssfr, \Sstar, \Smol\, and the stellar age and metallicity, we find significant anti-correlations among the specific star formation rate (sSFR = \Ssfr/\Sstar), the star formation efficiency (SFE = \Ssfr/\Smol) and the stellar age and metallicity, respectively. On the other hand, by construction of the \mbox{\Ssfr - \Phyd} relation, \mbox{$\Delta {\rm SFHP}$}\, strongly correlates with both sSFR, and SFE. These inverse relations between sSFR age and metallicity are expected, since low (high) values of sSFR (or equivalently the EW(\ha)) usually trace old/metal-rich (young/metal-poor) stellar populations \citep[e.g.,][]{Mejia-Narvaez_2020}.
\subsection{Impact of Global parameters on the \Ssfr - \Phyd\, relation}
\label{sec:global}
\begin{figure}
\includegraphics[width=\linewidth]{figures/Pism_SFR_EDGE_Mass_v2.pdf}
\includegraphics[width=\linewidth]{figures/Pism_SFR_EDGE_Morph_v2.pdf}
\caption{({\it top}) The \Ssfr - \Phyd\, relation color coded by the total stellar mass of the host galaxy. The circles in the inset show the average \mbox{$\Delta {\rm SFHP}$}\, for each galaxy against its total stellar mass while the associated error bars represent the standard deviation of \mbox{$\Delta {\rm SFHP}$}\, for each galaxy; the black solid line represents a linear fit to the averaged residuals. ({\it bottom}) The \Ssfr - \Phyd\, relation color coded by the morphological type of the host galaxy. The circles in the inset shows the average \mbox{$\Delta {\rm SFHP}$}\, for a given morphological type. The violin-histograms in the inset represent the distribution of \mbox{$\Delta {\rm SFHP}$}\, for the different morphological types. Although there seems to be a correlation between \mbox{$\Delta {\rm SFHP}$}\, and the total stellar mass, these global properties does not seem to have a large impact in the \Ssfr - \Phyd\, relation (see details in Sec.~\ref{sec:global}.)
}
\label{fig:Mass_Morph}
\end{figure}
Having a homogeneous spatially-resolved dataset for a significant large sample of galaxies in the nearby Universe allow us to explore the impact of global properties on the \mbox{\Ssfr - \Phyd} relation. As reviewed by \cite{Sanchez_2020ARAA}, most of the scaling relations derived on kpc scales in the local Universe are affected by the structural/integrated properties of their host galaxy. In particular, the total stellar mass and the morphology of a galaxy can modulate most of these relations at local scales. In Fig.~\ref{fig:Mass_Morph} we color code the \mbox{\Ssfr - \Phyd} relation according to these two global observables. The top panel indicates that for regions located in low-mass galaxies \Ssfr is larger in comparison to regions with similar \Phyd\, located in more massive galaxies. This trend is more evident when we plot the average residuals for each galaxy of the \mbox{\Ssfr - \Phyd} relation against the total stellar mass (see inset). We find that the average \mbox{$\Delta {\rm SFHP}$}\, for each galaxy anti-correlates with the total stellar mass ($r = $ -0.3). For most of low-mass galaxies ($\log({\rm M_{\odot}/M_{\ast}} \lesssim 10.5)$) the best fit of the \mbox{\Ssfr - \Phyd} relation underestimates the observed \Ssfr\, whereas for more massive galaxies \Ssfr\, is slightly overestimated. This is quantified by the slope of the black line in the inset ($\sim -0.17$~dex/M$_{\ast}$). We note that even though this trend seems to be systematic, due to the large scatter, observations for a wider range of galaxies is required to test the robustness of the impact of the total stellar mass in shaping the local \mbox{\Ssfr - \Phyd} relation.
We explore whether the inclusion of the total stellar mass as a secondary parameter in the \mbox{\Ssfr - \Phyd} relation could induce a reduction in its scatter. If so, this would be strong evidence of the importance of the role that the potential well could have in shaping \Ssfr at local scales. We measure the scatter of the relation between \Ssfr and a parameter that includes the stellar mass as a secondary parameter of the form \mbox{$x = \log(\mathrm{P_{h}}\,/ \,k_\mathrm{B}\,\mathrm{K\,\,cm^{-3} }) - \alpha \,\,\log(\mathrm{M_{\ast}/ 10^{10} M_{\odot}}$)} with \mbox{$-1 < \alpha < 1$}. We find that the value that yields the smallest scatter is $\alpha$ = -0.07. The scatter of this relation is very similar to the one where the stellar mass is not included as a secondary parameter (i.e., $\sigma = $ 0.28 dex). We suggest that even though the overall gravitational potential may have an impact in shaping \Ssfr, its impact is relatively mild in comparison to the local pressure. Further studies including larger samples of low-mass galaxies are thus required to quantify the actual impact of the potential well in this relation.
On the other hand, when we color code this relation according to the morphology of the host galaxies (bottom panel), regions located in late-type galaxies have a slightly large \Ssfr\, than regions in more early-type galaxies for a similar \Phyd. We also identify this mild trend when plotting the average residual of the \mbox{\Ssfr - \Phyd} relation for a given morphological type. For early-type galaxies (E-S0) the best \mbox{\Ssfr - \Phyd} relation overestimates \Ssfr by a factor of $\sim$ 0.2 dex, whereas for late type galaxies (Sc,Sd-Irr) \Ssfr is underestimated by an average of $\sim$ 0.1 dex. Despite this apparent trend we should note that the standard deviation of the residuals for each morphological type (error bars for the white points) are consistent with no change in the residual of the \mbox{\Ssfr - \Phyd} relation with respect to the morphology (i.e., \mbox{$\Delta {\rm SFHP}$} = 0). Thus to either confirm or rule out these trends we would require a larger sample of galaxies that includes more early-type galaxies than those currently available (see inset in the middle panel of Fig.~\ref{fig:Sample}).
\begin{figure}
\includegraphics[width=\linewidth]{figures/res_Global_v1.pdf}
\caption{The radial distribution of the residuals of the \mbox{\Ssfr - \Phyd} relation for three global scaling relations. For each of the global scaling relation we divide the sample in galaxies with low, average and high bins (red, green, and blue lines) referring to the vertical distance of the galaxy with respect to the best fit of the global scaling relation (see black solid lines in each panel of Fig.~\ref{fig:Sample}). In the top, middle and bottom panel we show the radial gradients of \mbox{$\Delta {\rm SFHP}$}\, for the star-formation main sequence, the Schmidt-Kennicutt law and the stellar-molecular mass relation, respectively. These radial trends highlight the impact of global properties in the local \mbox{\Ssfr - \Phyd} relation.
}
\label{fig:Rad}
\end{figure}
In Fig.~\ref{fig:Rad} we plot the radial distribution of the residuals of the \mbox{\Ssfr - \Phyd} relation according to the position of the host galaxy in different global scaling relations presented in Fig.~\ref{fig:Sample}.
In the top panel of Fig.~\ref{fig:Rad} we plot the radial gradient of \mbox{$\Delta {\rm SFHP}$}\, dividing galaxies in three groups (low, average, and high $\Delta$SFMS) according to their location respect the star formation main sequence (below -0.5, between -0.5 and 0.5, and above 0.5 times the standard deviation of the scatter from the black solid line in the left panel of Fig.~\ref{fig:Sample}, respectively).
Although the radial distributions are clustered around the zero residual, we find that in galaxies with higher and average SFR for their stellar mass have a flat radial gradient of their \mbox{$\Delta {\rm SFHP}$}.
For galaxies with low $\Delta$SFMS, \mbox{$\Delta {\rm SFHP}$}\, is negative for all radii, decreasing with radii ($\sim$ -0.07 dex/R$_{\rm eff}$). In other words, for global low star-forming galaxies we are overestimating \Ssfr when deriving it from \Phyd, particularly at they outskirts.
Similar as above, in the middle panel of Fig.~\ref{fig:Rad} we classify galaxies according to their vertical distance with respect to the black solid line in the middle panel of Fig.~\ref{fig:Sample} (low, average, and high $\Delta$SK). Galaxies with low and average SFR for their total gas mass have similar mildly increasing gradients ($\sim$ 0.03 dex/R$_{\rm eff}$). On the other hand, for those galaxies in the large $\Delta$SK bin, \mbox{$\Delta {\rm SFHP}$}\, is positive for all radii ($\sim$ 0.1 dex), even more it shows a positive gradient ($\sim$ 0.07 dex/R$_{\rm eff}$). This suggests that for those galaxies with larger SFR, with respect to the SK-law, the \Ssfr is underestimated from the local \Phyd, particularly at they outskirts.
Finally, we plot in the bottom panel of Fig.~\ref{fig:Rad} the radial gradient of \mbox{$\Delta {\rm SFHP}$}\, for galaxies classified according to their vertical distance with respect to the black line in right panel of Fig.~\ref{fig:Sample} (low, average, and high $\Delta$F$_{\rm gas}$; in other words the lack or excess of M$_{\rm mol}$ for a given M$_{\ast}$). For the average $\Delta$F$_{\rm gas}$ the radial distribution of \mbox{$\Delta {\rm SFHP}$}\, is flat and close to zero. For galaxies in the low $\Delta$F$_{\rm gas}$ bin, we find a significant negative radial gradient (\mbox{$\sim$ -0.14 dex/R$_{\rm eff}$}), suggesting that in the outskirts of these galaxies \Ssfr is overestimated by \Phyd. We note a dip of \mbox{$\Delta {\rm SFHP}$}\, at central regions for galaxies in the high $\Delta$F$_{\rm gas}$ bin, inducing a positive gradient ($\sim$ 0.07 dex/R$_{\rm eff}$). These radial trends show the impact of global properties in the local scaling relations such as the \mbox{\Ssfr - \Phyd} one.
\subsection{Testing the impact of systematic in the \mbox{\Ssfr - \Phyd} relation}
\label{sec:caveats}
\begin{figure}
\includegraphics[width=\linewidth]{figures/Pism_SFR_EDGE_lowIncl_v1.pdf}
\caption{The \mbox{\Ssfr - \Phyd} relation, with regions located in low inclination galaxies ($b/a$ $>$ 0.45) shown with blue crosses and high inclination ones ($b/a$ $<$ 0.45) shown as orange plus symbols. Similar to Fig.~\ref{fig:SPrel}, the solid line representing and ODR fit to the regions in low-inclined galaxies show a good agreement with the best fit for all the sample in Fig.~\ref{fig:SPrel}. The inset shows the impact of $b/a$ in each axis of the \mbox{\Ssfr - \Phyd} relation.
}
\label{fig:low-inc}
\end{figure}
In this section we explore the impact of the inclination and a variable \mbox{$\alpha_{\rm CO}$} conversion factor on the \mbox{\Ssfr - \Phyd} relation. Although all the surface densities in this study are corrected by inclination effects, to further explore the impact of galaxy projection on this relation, in Fig.~\ref{fig:low-inc} we plot the \mbox{\Ssfr - \Phyd} relation as in Fig.~\ref{fig:SPrel}, but color-coding galaxies according to inclination: regions located in low inclination galaxies ($b/a$ $>$ 0.45) are indicated with blue crosses and regions located in high inclination galaxies $b/a$ $<$ 0.45) are indicated with orange plus symbols. We note that on average for low inclined galaxies \Phyd\, and \Ssfr are larger compared to regions located in highly inclined galaxies. We perform a similar analysis as in Sec.~\ref{sec:SPrel} for the low-inclined sample. We find a smaller Pearson correlation coefficient in comparison to the entire sample ($r = 0.80$).
The ODR fitting shows a similar slope by comparison to the entire sample (\mbox{$b$ = 0.95$\pm$0.05, $A$ = 0.2$\pm$0.1}). The residual of the fit is slightly larger than the one derived from the entire sample ($\sigma$ = 0.30). These results show that inclination does not play a major impact in shaping the slope of the \mbox{\Ssfr - \Phyd} relation. The impact of inclination measured by the $b/a$ ratio for \Ssfr and \Phyd is shown in the inset of Fig.~\ref{fig:low-inc}. This further indicates the impact of inclination in the observed \mbox{\Ssfr - \Phyd} relation.
\begin{figure}
\includegraphics[width=\linewidth]{figures/Pism_SFR_EDGE_aCO_v1.pdf}
\caption{The red-dashed contours enclose the \mbox{\Ssfr - \Phyd} relation using a variable $\mathrm{\alpha_{CO}}$ conversion factor. The best ODR fit is represented by a red-dashed line while the the red shaded area shows the uncertainty of assuming different \Shi densities (see details in Sec.~\ref{sec:Data}). The black contours and black lines are the same as those in Fig.~\ref{fig:SPrel}. By allowing a variable $\mathrm{\alpha_{CO}}$ conversion factor, the slope of the \mbox{\Ssfr - \Phyd} relation is steeper than when adopting a constant factor.
}
\label{fig:aCO}
\end{figure}
Along this article, we adopt a constant Milky-Way value of the conversion factor between the CO luminosity and the molecular gas mass density ($\mathrm{\alpha_{CO}}$). Since $\mathrm{\alpha_{CO}}$ can vary, in particular decreasing in regions of high density and temperature where the CO excitation is higher, which tend to be associated with high-pressure inner-galaxy regions \citep[e.g.,][]{Bolatto_2013, Gong_2020}, here we explore how the \mbox{\Ssfr - \Phyd} relation varies when including a variable $\mathrm{\alpha_{CO}}$ (\mbox{$\mathrm{\alpha_{CO}}(Z,\Sstar)$}, see Eq.~\ref{eq:aCO}). In Fig.~\ref{fig:aCO}, datapoints and red-dashed contours represent the \mbox{\Ssfr - \Phyd} relation using \mbox{$\mathrm{\alpha_{CO}}(Z,\Sstar)$} whereas the black contours are the same as those presented in Fig.~\ref{fig:SPrel}. We follow a similar procedure as in Sec.~\ref{sec:SPrel} to derive the best ODR fit to the data. We find that this fit is steeper (red-dashed line, \mbox{$b$ = 1.14$\pm$0.07, $A$ = 0.0$\pm$0.1}) in comparison to the one derived adopting a constant $\mathrm{\alpha_{CO}}$ (black solid line). The slope of the best fit agrees with derivation from numerical simulations of feedback from SN explosions and photoelectric heating \citep[][]{Kim_2013}. However, the scatter of the relation is larger (and the Pearson correlation is smaller, $r =$ 0.83) in comparison to the relation presented in Sec.~\ref{sec:SPrel}.
\subsection{Comparison with the literature}
\label{sec:comp}
\begin{figure}
\includegraphics[width=\linewidth]{figures/Pde_SFR_LitComp_v1.pdf}
\caption{
The \Ssfr - \Pde\, relation derived for the EDGE-CALIFA galaxies (contours enclosing 90\%, 80\%, and 50\% of the sample) compared to spatially resolved data from previous studies. Orange plus symbols represent the data from the THINGS survey \citep{Leroy_2008}. Blue crosses represent the data from the PHANGS survey \citep{Sun_2020}. The solid black line represents the best fit to our data. The dashed line shows the prediction from hydrodynamic simulations \citep{Kim_2013}. Despite the differences in observables and samples size, the \Ssfr - \Pde\, relation for the EDGE-CALIFA sample is in agreement with previous estimates from smaller samples of galaxies.
}
\label{fig:Pde}
\end{figure}
Using spatially resolved data from surveys including small yet significant samples of galaxies, different studies have shown the tight correlation between \Ssfr and the so-called dynamical equilibrium pressure, \Pde\, (see Eq.~\ref{eq:Pde}). In this section we compare the \Ssfr-\Pde\, relation derived for our dataset of spatially resolved observations of 96 star-forming galaxies from the EDGE-CALIFA survey with those relations derived in the literature.
In Fig.~\ref{fig:Pde} we plot the \Ssfr-\Pde\, relation for the regions sampled by the EDGE-CALIFA survey. Inwards, the contours enclose 90\%, 80\%, and 50\% of the distribution in this relation. The correlation coefficient of this relation is similar to the one derived for the \mbox{\Ssfr - \Phyd} relation ($r =$ 0.85). Despite the assumptions to derive \Pde, the trend observed of the \mbox{\Ssfr - \Pde} relation with our dataset is in good agreement with those derived in the literature for spatially resolved measurements from the HERACLES+THINGS \citep{Leroy_2008} and PHANGS surveys \citep[][blue x-symbols and orange crosses, respectively]{Sun_2020}. Our sample covers similar values of \Ssfr and \Pde\, than those derived from the PHANGS multi-wavelength dataset, the distribution of our estimate of pressure is slightly shifted towards lower values of \Ssfr\, ($\sim$ 0.2 dex). Although small, this difference could be due, among others to (i) the significant difference between the datasets \citep[][used the photometric data at different wavelengths such as UV, and IR, whereas here we use IFS dataset only in the optical regime]{Sun_2020} in particular this could lead to an underestimation of $\sim$ 0.2 dex when using the dust-corrected \ha luminosity to derive \Ssfr \citep[][although see \citealt{Catalan_Torrecilla_2015}]{Hirashita_2003}, and (ii) we study a large galaxy sample \citep[][used a sample of 28 star-forming galaxies whereas we use a sample of 96 galaxies including massive galaxies with a wider range of morphological types, see Fig.~\ref{fig:Sample}]{Sun_2020}. Furthermore, in Sec.~\ref{sec:PS_ratio} we find an excellent agreement with the best-fit parameters derived by \cite{Sun_2020} when selecting regions with large \ha\, equivalent width ( EW(\ha) $\gtrsim$ 20 \AA).
The HERACLES-THINGS survey, on the other hand, traces the \mbox{\Ssfr - \Pde} relation mainly for galaxies with lower stellar mass than our sample (see Fig.~\ref{fig:Sample}). For a fraction of radial bins (those with similar values of \Pde), the \mbox{\Ssfr - \Pde} relation derived from the this survey is in agreement with those derived in this study. However, there is a significant fraction of radial bins that have smaller values of \Pde\, in comparison to those estimated from the EDGE-CALIFA or PHANGS surveys. Those radial bins exhibit a sharp drop in their values of \Ssfr as those expected from the previous surveys. It can be the case that for regions with \mbox{$\Pde/ k_{\rm B}\, \lesssim $ 10$^{4}$ $\rm{K\,\,cm^{-3}}$} the \Ssfr does not follows a power-law relation (or with a different index) as the one described for the bulk of our observations.
Following the same procedure as in Sec.~\ref{sec:SPrel}, we derive the best fit for the \mbox{\Ssfr - \Pde} relation. The best relation from an ODR fit is similar -- within uncertainties -- to the one derived in Sec.~\ref{sec:SPrel} for the \mbox{\Ssfr - \Phyd} relation (\mbox{$b$ = $ 0.96 \pm 0.08$}, \mbox{$\log(A)$ = $ 0.2\pm 0.1$}). The scatter of this relation is similar to the one derived for the \mbox{\Ssfr - \Phyd} relation (\mbox{$\sigma$ = 0.28}). The slope of this fit is in agreement with estimations using CO staked spectra from the EDGE-CALIFA survey (Villanueva et al. in prep.). On the other hand, an OLS fitting of this relation is in good agreement with the best-fit derived from the PHANGS dataset (\mbox{$b$ = $ 0.84 \pm 0.04$}, \mbox{$\log(A)$ = $ 0.4\pm 0.1$}). The trend observed for our derivation of the \mbox{\Ssfr - \Phyd} relation agrees with the prediction from a hydrodynamic simulation \citep[$b$ = 1.13, and \mbox{$\log(A)$ = 0.26}, ][]{Kim_2013}. As mentioned by \cite{Sun_2020} using the PHANGS data, our observations also show a shallower slope in comparison to the value expected from the simulation, in particular at large values of \Pde. This difference in slope could reflect systematic effects of the ISM at different locations of the galaxies or the change of the properties for massive galaxies. In Sec.\ref{sec:PS_ratio} we further explore these possibilities. In summary, despite the samples sizes and their differences in their measurement, our estimation of the \mbox{\Ssfr - \Pde} relation is in good agreement with those presented previously in the literature.
\subsection{The \Phyd/\Ssfr ratio}
\label{sec:PS_ratio}
\begin{figure}
\includegraphics[width=\linewidth]{figures/P_SFR_ratio_v3.pdf}
\caption{The 4 P/\Ssfr ratio against P. The points are color-coded by the EW(\ha) representing the \mbox{4 \Phyd/\Ssfr} vs \Phyd\, relation for the regions sampled in this study. The gray contour encloses 80\% of the 4 \Pde/\Ssfr vs \Pde\, relation derived for regions included in 28 star-forming PHANGS galaxies \citep[][]{Sun_2020}. The empty circle and large blue cross represent the locations of the median ratio and pressure for the previous distributions, respectively. The triangles with error bars represent the integrated measurements from highly turbulent galaxies included in the DYNAMO survey \citep[][]{Fisher_2019}. The dashed green line represents the typical value adopted as the feedback-driven momentum injection per unit mass of stars formed \citep[3000 km s$^{-1}$, ][]{Ostriker_Shetty_2011,Fisher_2019}. The dashed violet line represents the sum of the aforementioned momentum injection with the one derived from the stellar winds and radiation pressure produced by a typical starburst population \citep[$\sim$ 4200 km s$^{-1}$][]{Heckman_2015, Heckman_2017}. Active star-forming regions are in agreement with both the feedback-driven momentum injection and the median values from the literature.
}
\label{fig:P_SFRratio}
\end{figure}
From star-formation theory, the ratio between the pressure (P) and the \Ssfr is proportional to the momentum injection per unit mass mostly from supernovae to the ISM, $p_{\ast}/m_{\ast}$ \citep[e.g.,][]{Ostriker_Shetty_2011, Shetty_2012, Faucher-Giguere_2013, Kim_2017}. These studies usually adopt or derive a constant value for this ratio that ranges between \mbox{$\sim 10^3$ - $10^4$ km s$^{-1}$} depending on the adopted conditions for the ISM, the clustering of the supernovae and losses due to interphase mixing \citep[e.g.,][]{Iffrig_2015, Kim_Ostriker_2015, Martizzi_2015, Walch_Naab_2015,Kim_2017, El-Badry_2019, Gentry_2019}. In Fig.~\ref{fig:P_SFRratio} we compare the 4 P/\Ssfr ratio against P, usually this ratio is expressed in units of km s$^{-1}$. The factor 4 comes from assuming that a spherical injection of the momentum flux is centered in the disk mid-plane and that this momentum is converted directly to turbulent pressure \citep{Ostriker_Shetty_2011}; this factor has been directly verified in numerical simulations of disk galaxies \citep[see Fig. 15 of][]{Kim_2013}. In this figure we plot the $4\,\Phyd/\Ssfr$ - \Phyd\, distribution adopting a constant CO conversion factor. Our sample covers a large dynamical range of the $4\,\Phyd/\Ssfr$ ratio (from 5.1 $\times$ 10$^2$ to 1.6 $\times$ 10$^5$ km s$^{-1}$). The median of this ratio is 4.3 $\times$ 10$^4$ km s$^{-1}$ (see empty circle in Fig.~\ref{fig:P_SFRratio}). We also derive this ratio assuming a variable CO conversion factor (see Eq.\ref{eq:aCO}). We find larger values for the $4\,\Phyd/\Ssfr$ ratio when adopting a variable CO conversion factor with a median of $\sim$ 6.4 $\times$ 10$^4$ km s$^{-1}$. On the other hand, we color-coded each region according to their \ha\, equivalent width, EW(\ha). Since this parameter captures the contrast between the adjacent stellar continuum and the \ha\, emission lines, it has been considered as a tracer of the star-formation activity \citep[e.g.,][]{Lacerda_2018, Sanchez_2020ARAA}. For a given \Phyd we find a clear trend, regions with large EW(\ha) have lower values of the $4\,\Phyd/\Ssfr$ ratio. This is particularly evident for regions with the median \Phyd ($\sim $ 10$^{4.4}$ $\rm{K\,\,cm^{-3}}$).
Using the data from \cite{Sun_2020}, we derive the $4\,\Pde/\Ssfr$ - \Pde\, distribution from the PHANGS survey (gray contour enclosing 80\% of their sample). The average $4\,\Pde/\Ssfr$ is smaller than those derived from our regions using \Phyd ($\sim$ 3.0 $\times$ 10$^3$ km s$^{-1}$, see thick large cross in Fig.~\ref{fig:P_SFRratio}). As we mention in Sec.~\ref{sec:comp}, these differences can be expected due to the equations used to estimate \Pde\, and \Phyd, the observables used to derive \Ssfr, and the observed samples of galaxies. Despite these differences, the distribution agrees with our sample, in particular with those regions that show large star-formation activity. In fact the observed trend for the blue regions (i.e., regions with high star-formation activity, EW(\ha) $\gtrsim$ 20 \AA) agrees with the distribution of the $4\,\Pde/\Ssfr$ - \Pde\, ratio from the PHANGS survey. The P/\Ssfr ratio increases with P. A similar trend is also observed for integrated properties of highly-turbulent galaxies included in the DYNAMO survey \citep[][although a shift of $\sim$ -0.2 dex in both axis can be expected due to improved estimations of their \Phyd\, (Girard et al., submitted), the observed trend holds]{Fisher_2019}.
Following \cite{Kim_Ostriker_2015}, we adopt a fiducial value of \mbox{$\sim$ 3000 km s$^{-1}$} as the value of the $p_{\ast}/m_{\ast}$ expected from supernovae (dashed green horizontal line). Based on the stellar models from \texttt{STARBURST99} \citep[][]{Leitherer_2014} we add to this estimation the momentum injection to the ISM produced by a combination of stellar winds and radiation pressure \citep[$\sim$ 1200 km s$^{-1}$$ $, dashed violet horizontal line,][]{Heckman_2017}. This value should be considered as an upper limit from the stellar models since the momentum flux due to stellar winds for a typical starburst population is smaller and can vary depending on the assumptions of the models \citep[$\sim$ 400 - 700 km s$^{-1}$$ $, e.g.,][]{Ostriker_Shetty_2011,Heckman_2015}. On the one hand, we find that a significant fraction of regions in our sample ($\sim$ 50\%) have a $4\,\Phyd/\Ssfr$ ratio larger than the expected values of momentum injection from models and simulations of star-formation. On the other hand we find that, when we only consider active star-forming regions (EW(\ha) $\gtrsim$ 20 \AA), the median ratio is consistent with the momentum injection due to SNe explosions ($\sim$ 3.3 $\times$ 10$^3$ km s$^{-1}$). We note that the median $4\,\Pde/\Ssfr$ from PHANGS is in good agreement with the expectation that momentum injection to the ISM is most likely driven by supernovae.
We recall that one of our selection criteria to ensure that we are only considering star-forming regions in our sample is that their EW(\ha) $>$ 6 \AA\, (see Sec~\ref{sec:Data}). However, we find that for those regions considered as active star-forming (i.e., EW(\ha) $\gtrsim$ 20 \AA), the $4\,\Phyd/\Ssfr$ ratio could indeed be represented by the momentum injection per mass unit expected from SNe, stellar winds and radiation pressure. The trend observed between $4\,\Phyd/\Ssfr$ and \Phyd\, is similar as those derived using (un-)resolved measurements in other star-forming galaxies \citep[][]{Fisher_2019,Sun_2020}. Furthermore, when selecting regions with EW(\ha) $\gtrsim$ 20 \AA, we find a tighter \mbox{\Ssfr - \Phyd} relation ($\sigma$= 0.22 dex) with a strong correlation coefficient ($r$= 0.90) and best-fit ODR parameters very similar as those derived for the PHANGS survey (\mbox{$b$ = $ 0.88 \pm 0.05$}, \mbox{$\log(A)$ = $ 0.5\pm 0.1$}).
There could be different reasons that explain why we find a significant fraction of star-forming regions with low EW(\ha) with large $4\,\Phyd/\Ssfr$ ratios. We recall that we are convolving the optical maps -- including the \ha map -- to have a common spatial resolution as the CO maps ($\sim$ 7\arcsec). This may induce a dilution of both the \Ssfr and the EW(\ha) for a given value of \Phyd. Thus, for a given sampled area with a low value of EW(\ha) we may be including some active star-forming regions and some others with no significant star-formation (regions with EW(\ha) $<$ 6 \AA, e.g., diffuse ionized gas). On the other hand, in comparison to previous studies exploring the \Ssfr - P relation at (sub-)kpc scales \citep[e.g.][]{Leroy_2008, Sun_2020}, our sample covers a large number of galaxies. This implies that we are probing different regimes of star-formation, including those where star formation is not as intense as those probed previously. This maybe the case for the massive galaxies that we sampled (see Sec.~\ref{sec:global}). It could also be the case that our measurements (i.e., Eq.~\ref{eq:Phyd}) over-estimates the pressure in those regions with EW(\ha) $<$ 20 \AA. This is certainly the case for regions located in structures of galaxies others than the disk like a bulge or a bar. We find that although central regions (i.e., those with R/R$_{\rm eff} < 0.5$) own the highest pressures ($\sim $ 10$^{5.5}$ $\rm{K\,\,cm^{-3}}$) they exhibit a wide range of $4\,\Phyd/\Ssfr$ ratios and EW(\ha) even more they represent only $\sim$ 17\% of our sample. Thus, those regions with large $4\,\Phyd/\Ssfr$ ratios and low EW(\ha) are not usually located in central regions of galaxies where the bulge dominates. Furthermore, in Sec.~\ref{sec:Disc} we show that barred galaxies have a similar distribution of the $4\,\Phyd/\Ssfr$ ratio in comparison to disk galaxies. From the observational data we suggest that for those regions with small EW(\ha) -- but still considered star-forming regions, 6 $<$ EW(\ha) $<$ 20 \AA -- the feedback provided from \Ssfr may not be sufficient to balance the pressure estimated from the stellar and gas mass densities leading to the observed large $4\,\Phyd/\Ssfr$ ratios. Finally, we note that this analysis assumes that the pressure produced by star formation feedback is mainly due to the specific momentum injected by SNe, $ p_{\ast}/(4m_{\ast})$. However, \cite{Ostriker_2010} argued that in regions of low shielding the thermal pressure and magnetic pressure (both driven by feedback) are expected to be comparable to the turbulent pressure, and this has been verified in solar neighborhood simulations by \cite{Kim_Ostriker_2015,Kim_Ostriker_2017}. Other potential sources of pressure associated with star formation, including cosmic rays and radiation, could also contribute to increasing the ratio of pressure to star formation \citep[see e.g.,][]{Ostriker_Shetty_2011, Diesing_2018}. It could be the case that the observed \Phyd/\Ssfr ratios for regions with low EW(\ha) are affected by more sources of feedback other than SNe.
\section{Discussion}
\label{sec:Disc}
In this study we describe and explore the \mbox{\Ssfr - \Phyd} relation at kpc scales for star forming regions located in a sample of 96 galaxies included in the EDGE-CALIFA survey. In Sec.~\ref{sec:SPrel} we find that this is a tight relation (i.e., with a small scatter, $\sim$ 0.2 dex), with a significant correlation coefficient ($r$ = 0.85, see Fig.~\ref{fig:SPrel}). This highlights the impact of \Phyd\, in shaping \Ssfr at local scales suggesting a scenario in which star-formation activity is self-regulated. As we mention in Sec.~\ref{sec:Intro}, averaged on scales of kpc and Myr, the feedback from SNe and stellar winds from massive stars counteracts the pressure from the gravity produced by the baryonic mass content \citep[e.g.,][]{Thompson_2005, Ostriker_Shetty_2011,Faucher-Giguere_2013}. In this section we discuss the implications of the slope we find for the \mbox{\Ssfr - \Phyd} relation (Sec.~\ref{sec:sub-linear_b}). We also discuss possible explanations for the correlations (or lack thereof) we observe between the residuals of other star-forming scaling relations and those derived from this relation as well as their anti-correlation with stellar properties (Sec.~\ref{sec:resSFR}). Finally, we discuss the impact of global properties on the \mbox{\Ssfr - \Phyd} relation (Sec~\ref{sec:Integ}).
\subsection{The slope of the \mbox{\Ssfr - \Phyd} relation}
\label{sec:sub-linear_b}
Despite the uncertainty in deriving the best \mbox{\Ssfr - \Phyd} relation due to the lack of information regarding the atomic gas distribution for our sampled galaxies, the range of slopes that best describe this relation are slightly sub-linear. In Sec.~\ref{sec:SPrel} we derive the \mbox{\Ssfr - \Phyd} relation assuming an expected range of \Shi densities. The average value of these slopes after 1000 realizations with different \Shi densities is $b$ $\sim$ 0.9. As we find in Sec.~\ref{sec:caveats}, this slope depends on different parameters including the CO conversion factor.
From a theoretical point of view, there are analytical models and numerical simulations of star formation in disk galaxies suggesting that star formation is a self-regulated process. In this scenario, the hydrostatic pressure is balanced by different feedback sources such as turbulent (from stellar winds, supernovae), thermal, magnetic and radiative pressure produced from young stellar objects \citep[e.g.,][]{Ostriker_2010,Ostriker_Shetty_2011, Shetty_2012, Kim_2013, Krumholz_2018}. The \mbox{\Ssfr - \Phyd} relation predicted from these studies is given by $\Sigma_{\rm SFR} = P/\eta_{\rm tot}$, where $\eta_{\rm tot}$ is the total feedback yield associated with young stars, and the contribution from turbulent pressure, $\eta_{\rm turb} = p_*/(4 m_*)$ is typically the largest single term. (see Sec.\ref{sec:PS_ratio}). Allowing for a dependence of $\eta$ on environment, this yields $\Sigma_{\rm SFR} \propto P^b$, where the power-law index, $b$ is either unity \citep[for analyical models, e.g., ][]{Ostriker_Shetty_2011} or slightly supra linear \citep[$\sim$ 1.13, for numerical simulations,][]{Kim_2013}. Within the uncertainties, we consider that the \mbox{\Ssfr - \Phyd} relation derived from EDGE-CALIFA galaxies is in agreement with the predicted linear slope from theory of star formation. In Secs.~\ref{sec:SPrel} and \ref{sec:comp} we find that for both estimates of the pressure (\Phyd\, and \Pde) a linear relation describe the bulk of our dataset. However, we observe differences in the slopes depending on the fitting technique we consider. The slopes derived from the OLS fit ($b \sim$ 0.84) are flatter than those derived using a ODR fit ($b \sim$ 0.95) for both the \mbox{\Ssfr - \Phyd} and \mbox{\Ssfr - \Pde} relations.
As we mention in Sec.~\ref{sec:comp} similar slopes have been observed using different samples and datasets. The slope derived from the spatially resolved dataset from the PHANGS survey using an OLS bisector method the fit is $b \sim$ 0.84 \citep{Sun_2020} while from unresolved measurements from DYNAMO galaxies the slope is also sub-linear \citep[$b \sim $0.76,][]{Fisher_2019}. \cite{Sun_2020} suggested that one possible reason for which they found a sub-linear slope in the \mbox{\Ssfr - \Pde} relation could be the fact that they sampled galaxies/regions for typical star-forming galaxies where the most extreme cases of star formation have not yet been tested. Thus, these author suggested that in order to fully explore the slope of the \mbox{\Ssfr - \Pde} relation, extreme star-forming regions (or "starburst" regime) such as the central regions of Ultra Luminous Infrared galaxies (ULIRGs) should be also included in the previous analysis.
From their results using unresolved measurements, \cite{Fisher_2019} suggested different scenarios that could explain the sub-linear slope they found in their sample of highly turbulent galaxies. They consider that the \mbox{\Ssfr - \Phyd} relation can be truly sub-linear, the $p_{\ast}/m_{\ast}$ ratio increases with the pressure, or/and there are other mechanisms that sustain the pressure in disk galaxies. Numerical simulation of star formation regulated by feedback have shown a qualitatively sub-linear relation between the \Ssfr and the pressure \citep{Benincasa_2016}. These simulations suggest that the feedback has a non-linear impact in the scale height of the galaxy. In turn, this affects the second term of Eq.~\ref{eq:Phyd}, specifically the \mbox{$\mathrm{\sigma_{mol}/\sigma_{stars,z}}$} ratio. To estimate \Phyd\, we have adopted -- as most of the observational studies in this regard -- a constant stellar scale height across the galaxy disk which in turn is proportional to their stellar scale length (see Sec.~\ref{sec:Data}). Although the scale height may vary for different position of the disk, we consider that this does not strongly affect the estimation of \Phyd\, and thus the slope \mbox{\Ssfr - \Phyd} relation. In Appendix \ref{app:Phyd} we show that adopting different estimations of the stellar length yield very similar values of the \mbox{$\mathrm{\sigma_{mol}/\sigma_{stars,z}}$} ratio. Furthermore, \cite{Sun_2020} have found that the stellar scale height and length are tightly correlated for their sample of star-forming galaxies, supporting that indeed the scale height can be considered as constant across the galaxy.
In Sec.~\ref{sec:PS_ratio} we show that the median value of the $4\,\Phyd/\Ssfr$ ratio (an observational proxy for $p_{\ast}/m_{\ast}$) is larger than the fiducial value expected from the momentum flux injection from SNe (see Fig.~\ref{fig:P_SFRratio}). Nevertheless, for most of the regions with high star formation activity (i.e., EW(\ha) $>$ 20\AA), the $4\,\Phyd/\Ssfr$ ratio is in agreement with a momentum injection per unit mass of star formed produced by SNe. For those active star-forming regions we also find that this ratio increases with \Phyd. This trend is also suggested by the $4\,\Pde/\Ssfr$ - \Pde reported by the PHANGS survey \citep[gray contour in Fig.~\ref{fig:P_SFRratio}][]{Sun_2020} as well as the unresolved measurement from DYNAMO galaxies \citep[red triangles with error bars in Fig.~\ref{fig:P_SFRratio}][]{Fisher_2019}. In fact, using those unresolved measurement \cite{Fisher_2019} suggested that this ratio is not constant across the disks of star-forming galaxies increasing with pressure, leading to the sub-linearity found in the \mbox{\Ssfr - \Pde} relation. For our sample we find this is the case, in particular, for regions with large EW(\ha). This suggests that for those regions the pressure can be inducing a variation of $p_{\ast}/m_{\ast}$ which in turn may play a significant role in shaping the \mbox{\Ssfr - \Pde} relation. Our results from Sec.~\ref{sec:PS_ratio} also suggest that there could be other processes that can induce departures from the $4\,\Phyd/\Ssfr$ ratio expected from the momentum injection from SNe, in particular for regions with low star-formation. Those physical processes which could include magnetic, and/or thermal pressure or cosmic rays may be very relevant to balance the mid plane hydrostatic pressure in those low star-formation regions.
Different studies have suggested the relevance of other processes than supernovae explosions which can maintain the pressure support in disk galaxies. From their measurements of the relation between the turbulent pressure and \Ssfr, \cite{Sun_2020} suggested that radial inflows induced by structures such as bars or bulges could be another source of turbulent pressure. On the other hand, models that include momentum flux injection to the ISM from other sources such as radiation pressure, photoionization and winds can contribute to the pressure \citep[e.g.,][]{Hopkins_2011,Hopkins_2014,Murray_2011}, although the ray-tracing simulations of \citealt[][see their Fig.12,]{Kim_2018} show that the specific momentum injection from radiation is small compared to that from SNe, and simulations from Lancaster et al (2021, submitted) found that wind momentum contributions are also much smaller than from SNe. In the same direction, recent models of star formation suggest that radial transport and feedback from supernovae can have similar impact in regulating star formation in disk galaxies \citep{Krumholz_2018}. In any case, these models/simulations are still needed to explain the observed relation between \Ssfr and the pressure at kpc scales for the sample of galaxies in this study.
Finally, we note that the slope of the \mbox{\Ssfr - \Phyd} relation can vary depending on the assumption to derive the physical quantities. In Sec.~\ref{sec:caveats} we show that the slope of this relation can vary when assuming a variable CO conversion factor ($b \sim$ 1.15). Furthermore, in Sec.~\ref{sec:PS_ratio} we find that the observed correlation between $4\,\Phyd/\Ssfr$ and \Phyd is reduced once we consider a variable CO conversion factor. We suggest that a larger sample of galaxies (covering a wider range of chemical abundances) is require to further understand the impact of the assumption in the derived quantities.
\subsection{The impact of \Phyd\ on other star-forming relations}
\label{sec:resSFR}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/SFrelations_v1.pdf}
\caption{Comparison between the resolved star formation main sequence (top-left panel), the resolved Schmidt-Kennicutt law (top-right panel), and the \mbox{\Ssfr - \Phyd} relation (bottom panels). The relations in the left panels are color-coded by the molecular gas mass surface density, \Smol, whereas the right panels are color-coded by the stellar mass surface density. In contrast to the single-variable star-forming scaling relations, the color coding illustrate the minimal impact that either \Sgas or \Sstar has in shaping \Ssfr once \Phyd\, is considered as the independent variable.
}
\label{fig:SFrelations}
\end{figure}
In Sec.~\ref{sec:local} we find strong correlations between the residuals of the \mbox{\Ssfr - \Phyd} relation, \mbox{$\Delta {\rm SFHP}$}, and those derived for the resolved star-formation main sequence (\mbox{$\Delta {\rm SFMS}$}, see Eq.~\ref{eq:rSFMS}) and those from the resolved Schmidt-Kennicutt relation (\mbox{$\Delta {\rm SK}$}, see~\ref{eq:rSK}). In contrast, we do not find a strong correlation between the residuals of the resolved molecular gas main sequence (\mbox{$\Delta {\rm MGMS}$}) and \mbox{$\Delta {\rm SFHP}$} (see Fig.~\ref{fig:SPrel_local}). The main goal of this comparison is to quantify whether individual components of the baryonic mass density are driving star formation at kpc scales in the EDGE-CALIFA galaxies.
\cite{Ellison_2020} found a significant correlation between the residuals of the resolved star-forming scaling relation (\mbox{$\Delta {\rm SK}$}\, and \mbox{$\Delta {\rm SFMS}$}) and a secondary correlation with the residuals of the resolved molecular gas main sequence (\mbox{$\Delta {\rm MGMS}$}) and the \mbox{$\Delta {\rm SFMS}$}. From their results, they suggested that star formation at kpc scales is primarily regulated by the amount of molecular gas, \Smol, with a secondary role for the star formation efficiency, SFE. In the top panels of Fig.~\ref{fig:SFrelations} we illustrate these results. The left panel shows the rSFMS color coded by \Smol. The right panel shows the rSK relation color coded by \Sstar. In both panels the effect that \Smol and \Sstar\, has on each of these relations is clear. For a given value of \Sstar\, (\Sgas), \Ssfr changes with respect to \Sgas (\Sstar). Therefore it is expected to find significant correlation between the residuals of these relations.
In the bottom panels of Fig.~\ref{fig:SFrelations} we color code the regions in the \mbox{\Ssfr - \Phyd} plane according to their gas and stellar mass densities (left and right panels, respectively). Contrary to the relations in the top panels, these plots show that for a given range of pressures there is no significant change in the values of either \Sgas\, or \Sstar. By construction, \Phyd\, is a combination of both \Sstar\, and \Sgas (see Eq.~\ref{eq:Phyd}). Thus, by including these two terms, the variations in \mbox{$\Delta {\rm SFHP}$}\, are going to scale in a similar way as those observed in the rSFMS or the rSK relation. The vertical distribution of both \Sgas and \Sstar in the \mbox{\Ssfr - \Phyd} plane could also explain the lack of relation between \mbox{$\Delta {\rm SFHP}$}\, and \mbox{$\Delta {\rm MGMS}$}. Previous studies suggested that the \Sgas\Sstar product (or their linear combination in logarithmic scales) better describes \Ssfr than each of them \citep[the so called extended Schmidt-Kennicutt relation, e.g.,][]{Shi_2011,Shi_2018}. Although the expected correlation from these studies has not yet been corroborated using IFS datasets \citep{Lin_2019,Barrera-Ballesteros_2021}, the fact that we find stronger correlation coefficients for the \mbox{\Ssfr - \Phyd} relation (see Sec.\ref{sec:SPrel}), small scatter in comparison to the rSFMS and the rSK (S\'anchez et al., in prep.), and the relations presented in Sec.\ref{sec:local} suggest that the primary driver for the star-formation at kpc scales rather than the individual components of the baryonic mass density is the hydrostatic pressure, \Phyd.
\subsection{The impact of global properties on the \mbox{\Ssfr-\Phyd} relation}
\label{sec:Integ}
In Sec.~\ref{sec:global} we find that the residuals of the \mbox{\Ssfr - \Phyd} relation apparently correlate with the total stellar mass (see top panel of Fig.~\ref{fig:Mass_Morph}). However, we did not find a strong reduction of the scatter of this relation when we include the total stellar mass as a secondary parameter. We consider that even though for a given \Phyd those regions with high/low \Ssfr tend to be in low-mass/massive galaxies, the pressure is the main parameter that modulates \Ssfr at local scales.
\begin{figure}
\includegraphics[width=\linewidth]{figures/P_SFR_ratio_Bars_v3.pdf}
\caption{Similar to Fig.~\ref{fig:P_SFRratio}. The color of the data points varies according to the galactocentric distance, darker data points indicate regions located closer to center of each galaxy. The circles represent the regions located in barred galaxies whereas the diamonds represent the regions located in disk galaxies. Central regions located in barred galaxies tend to have larger $4\,\Phyd/\Ssfr$ ratios.
}
\label{fig:P_SFRratioBars}
\end{figure}
We also explore how the \mbox{\Ssfr - \Phyd} relation is affected by the morphology of the host galaxy (see bottom panel of Fig.~\ref{fig:Mass_Morph}). Our analysis shows a mild variation of the residual for different morphological types. We indicate in Sec.~\ref{sec:Data} that most of the targets in our sample are late type galaxies and very few are early-type (see inset in the middle panel of Fig.~\ref{fig:Sample}). For those few early-type galaxies we note that the best \mbox{\Ssfr - \Phyd} slightly overestimated \Ssfr. This may be a hint of the so-called morphological quenching where \Ssfr is halted due to the presence of a bulge rather than the absence of molecular gas \citep[e.g.,][]{Martig_2009,Colombo_2018}. In order to further explore this very interesting possibility, we require a larger sample of galaxies with significant bulge fraction than the one provided in this study. On the other hand, we remind ourselves that the estimation of \Phyd has been derived under the assumption of a thin disk (see Eq.~\ref{eq:Phyd}), therefore it may be not valid for a bulge-dominated galaxies which in turn may lead to this overestimation.
In Sec.~\ref{sec:sub-linear_b} we suggest that alternative venues that could sustain the pressure other than feedback from supernovae in disk galaxies could be the presence of a bulge or a bar \citep[e.g.][]{Sun_2020}. A significant fraction of our sample includes barred galaxies ($\sim$ 44\%, 45/101) allowing us to test statistically the impact of bars in the estimation of the \mbox{\Ssfr - \Phyd} relation. In Fig.~\ref{fig:P_SFRratioBars}, we compare the regions located in bars and disk galaxies in the \mbox{$4\,\Phyd/\Ssfr$ - \Phyd} plane. Data points are colored by their galactocentric distance with darker points representing regions closer to the center. It is expected that central regions are those with the highest pressures. According to \cite{Sun_2020}, we would expect that those regions with large values of the $4\,\Phyd/\Ssfr$ ratio are located preferentially in barred galaxies. Fig.~\ref{fig:P_SFRratioBars} shows that regions with large values of $4\,\Phyd/\Ssfr$ ratio are located in both barred and disk galaxies. However, there are more regions with \mbox{$4\,\Phyd/\Ssfr$ $\gtrsim$ 10$^{4.5}$ km s$^{-1}$} in barred galaxies than in disk galaxies (20 vs 6). Regions with the highest pressures and the largest \mbox{$4\,\Phyd/\Ssfr$} ratios are located in barred galaxies. Also, the median value of the \mbox{$4\,\Phyd/\Ssfr$} ratio is slightly larger for regions located in barred galaxies than for those located in disk galaxies ($\sim$ 5015 vs. 4500 km s$^{-1}$). Overall, we suggest that bars have a rather mild impact in setting the pressure at kpc scales. Detailed numerical simulations exploring the role of the radial motions as source of pressure are required to quantify these trends.
\section{Conclusions}
\label{sec:Conclusions}
The spatially resolved dataset from the EDGE-CALIFA survey \citep{Bolatto_2017} allow us to estimate the relation between the star formation rate density, \Ssfr and the hydrostatic mid-plane pressure, \Phyd\, for a sample of 4260 star-forming regions located in 96 galaxies of the nearby Universe. This sample covers a significant range of properties which is essential to test the impact of global observables on this spatially resolved scaling relation. The main results of this study are as follows:
\begin{itemize}
\item We find that \Ssfr strongly correlates with \Phyd\, (Pearson correlation coefficient, $r = 0.84$). This correlation is tight (scatter $\sim$ 0.2 dex). The bulk of this relation is in agreement with a linear relation suggesting that star-formation is an auto-regulated process.
\item From the \mbox{$4\,\Phyd/\Ssfr$} ratios, we suggest that one of the main source of momentum flux injection to the ISM comes from supernovae explosions, in particular for those regions considered as actively star-forming ones (i.e., EW(\ha) $>$ 20 \AA). For those regions with EW(\ha) $<$ 20 \AA, we suggest that either it is required to invoke other sources of pressure such as magnetic, and/or thermal pressure or cosmic rays or that the measured \Ssfr in those regions from the \ha emission line may be polluted with emission not corresponding to pure star-forming regions (e.g., diffuse ionized gas).
\item The strong correlation coefficient of the \mbox{\Ssfr - \Phyd} relation in comparison to other star-forming scaling relations (such as the rSFMS and the rSK), the fact that its scatter is very similar to the scatter of those scaling relations, and that its residuals do not correlate with the residuals of the molecular gas main sequence, indicate that \Phyd is probably the main driver of \Ssfr at kpc scales rather than individual components of the baryonic mass.
\item Total stellar mass may play a role in shaping the local \Ssfr. For a given \Phyd, \Ssfr decreases as stellar mass increases. However, when imposing a secondary relation with the stellar mass we do not find a strong reduction in the scatter of the \mbox{\Ssfr - \Phyd} relation. If the potential of the host galaxy affects the production of stars at kpc scales, its effect is rather mild.
\item The \mbox{\Ssfr - \Phyd} relation does not seem to be affected by the host galaxy's morphological type. Furthermore, central regions in barred galaxies have similar \mbox{$4\,\Phyd/\Ssfr$} ratios than those located in the center of disk galaxies. In the framework of feedback from recently formed stars, this suggest that bars may play a secondary role as a source of pressure support in late type galaxies.
\end{itemize}
Our results indicate that thanks to the self-regulation of the star formation, the mid-plane pressure plays a paramount role in shaping the creation of newly born stars at kpc scales in disk galaxies. Injection of momentum flux from supernovae explosions to the ISM is apparently one of the main process that induces this self-regulation. However our analysis, in agreement with previous studies, also suggests that there can be other process that can support the pressure in disk galaxies. Numerical simulations exploring these different channels are thus required as well as spatially resolved observations in actively star-forming regions.
\section*{Acknowledgements}
J.B-B and SFS acknowledge support from the grants IA-100420 and IN100519 (DGAPA-PAPIIT ,UNAM) and funding from the CONACYT grants CF19-39578, CB-285080, and FC-2016-01-1916. SV, ADB, RCL, and VVL acknowledge partial support from NSF-AST1615960. T.W., Y.C., and Y.L. acknowledge support from the NSF through grant AST-1616199. DC acknowlegdes support from the \emph{Deut\-sche For\-schungs\-ge\-mein\-schaft, DFG\/} project number SFB956A.
\section*{Data Availability}
As we mention in Sec.\ref{sec:edgepy}, the data used to derive the physical quantities presented in this article are those available in the \texttt{edge\_pydb} database. A detailed description of the database can be found in Wong et al. (in prep.).
\bibliographystyle{mnras}
|
\section{Introduction} \label{sec:intro}
Stellar-mass black holes (with masses between 5--15 solar masses [$M_{\odot}$]) can produce outbursts of radiation triggered by a sudden influx of material. The emission from these outbursts encodes critical information about the physics of accretion onto a compact object, nature's most efficient way to convert mass to energy. Broadly speaking, three accretion states are observed during these outbursts: the quiescent, the soft, and the hard state \citep{Homan05,Remillard06}. These states are defined by the relative importance of the two main physical components of the inner accretion flow: the accretion disk, and a hot, tenuous plasma known as the corona.
In quiescence and at the start of an outburst, stellar-mass black holes typically produce relatively hard emission (dominated by higher energy photons). As it nears peak brightness, the emission becomes soft, i.e., dominated by relatively lower energy photons from the accretion disk, with a weak coronal contribution and low X-ray flux variability. This state is referred to as the soft state. As the outburst intensity declines, the emission becomes harder again, dominated by higher energy X-ray photons from the corona. This hard state is also characterised by large X-ray variability. As the brightness decreases further, these systems return to the quiescent state. This last is poorly understood because the system is intrinsically very faint, but quiescent emission is typically softer than in the hard state. Evolution across states, which represents a framework to understand the process of accretion onto black holes, is thought to correlate with the overall mass accretion rate, although other factors (e.g. black hole spin, magnetic flux) likely also play an important role.
A long-standing question in compact object accretion physics is whether supermassive black holes (SMBHs; with masses $\gtrsim$ a few$\times$10$^{5}$ $M_{\odot}$) undergo the same accretion cycle, i.e., with similar states and mechanisms that trigger transitions between the states. This is interesting because state transitions signal changes to the fundamental physics governing the accretion processes \citep{Abramowicz13}. These include the dominant cooling mechanism, geometry, interplay between the emitting regions, jet formation and accretion efficiency. A key question is if and how these mechanisms scale with the black hole mass.
For actively accreting SMBHs or active galactic nuclei (AGN), large changes in mass accretion rate similar to stellar-mass black hole outbursts are thought to occur on timescales of hundreds to thousands of years. Observations of multiple state transitions in individual AGN are therefore very rare \citep{Mcelroy16, Parker19}, and the observed timescales ($\sim$ 10s of years) highlight the uncertainties in our understanding of the mechanism responsible for such state changes. Observational constraints for SMBHs in different accretion states are largely statistical in nature (\citealt{Merloni03, Falcke04, Mchardy06}, although some work on individual systems exists -- e.g. \citealt{Gezari17clagn, Noda18, Frederick19, Trakhtenbrot19}), which severely complicates the development of a detailed, holistic theoretical framework.
Tidal disruption events (TDEs) are episodes in which passing stars are ripped apart by tidal forces in the vicinity of SMBHs within distant galaxies \citep{Hills1975,Rees1988}. They have long been heralded as ideal systems to study accretion states and transitions in SMBHs, as they evolve from a dormant (quiescent) state to a high accretion rate phase, and back again, on $\sim$year timescales. However, many TDE accretion disks appear to be stable for at least 5--10 yr after disruption \citep{vanvelzen2019}. Moreover, only a small fraction of TDEs show persistently bright X-ray emission, without which a direct comparison to X-ray binary (XRB) spectral and timing properties is challenging. As a result, TDEs with multiple state transitions have not yet been found (see \citealt{Komossa04, Maksym14, Jonker20, Wevers20} for work in this regard).
The transient ASASSN--18ul/AT2018fyk was discovered by the All-Sky Automated Survey for Supernovae \citep{Shappee14} on 8 September 2018 in the nucleus of a galaxy at a redshift of 0.059 (luminosity distance of 264 Mpc). Based on the blue optical spectrum with broad H and He emission lines, hot (T$\sim$35000 K) UV/optical blackbody emission that does not cool significantly over time, and the lack of AGN-like emission lines, it was classified as a TDE by a black hole with a (independently derived) mass of log$_{10}$(M$_{\rm BH}$) = 7.7 $\pm$ 0.4 M$_{\odot}$\citep{Wevers20, Wevers19b}.\\
In this manuscript we present an in-depth analysis of both archival and new radio, optical, UV and X-ray observations of AT2018fyk taken up to 2 years after the initial discovery. These data provide three key diagnostics to characterise the accretion flow properties following the TDE, which together enable detailed comparison with stellar-mass black holes:
\begin{itemize}
\item The evolution of the UV--X-ray spectral slope {$\alpha_{\rm ox}$\,} with bolometric Eddington ratio.
\item The ratio of the power-law spectral component flux to the total emission in X-rays, and the ratio of power-law emission to the total bolometric luminosity (including the accretion disk, soft excess and corona).
\item Photometric variability properties of the X-ray emission derived from the {\it Swift} and {\it XMM-Newton} light curves.
\end{itemize}
Using these diagnostic tools, we investigate the properties of AT2018fyk, and find remarkable similarities to the properties of accreting stellar-mass black holes.\\
Section \ref{sec:reduction} details the observations and data reduction. We present the results of host galaxy modelling and SED, X-ray energy spectral and timing analysis in Section \ref{sec:analysis}. The main results are discussed in the framework of accretion state transitions in Section \ref{sec:discussion}, including similarities and differences between stellar-mass and SMBH systems, and a comparison between AT2018fyk and AGNs. We summarise our findings in Section \ref{sec:summary}.
\section{Observations and data reduction}
\label{sec:reduction}
\subsection{XMM-Newton stare observations}
AT2018fyk was observed by the {\it XMM-Newton} \citep{Jansen01} European Photon Imaging Camera (EPIC) on three occasions. The first observation (obsID: 0831790201; hereafter referred to as XMM1), for a total of 32 kiloseconds (ks), was observed shortly after the initial disk transition began (between states A and B) on 2018 December 9 (MJD 58461.75), $\approx$ 92 days after discovery.
The second observation (obsID: 0853980201; XMM2) was performed $\approx$ 413 days after discovery (in state D), for a total of 54 ks on 2019 October 26 (MJD 58782.21).
The third observation, totalling 17 ks (obsID: 0854591401; XMM3), was taken on 2020 May 12 (MJD 58981.27), roughly 612 days after optical discovery and after the second state transition (in state E).
In all the observations EPIC was operating in the imaging mode and the source did not suffer from pile-up\footnote{\url{https://heasarc.gsfc.nasa.gov/docs/xmm/sas/USG/epicpileup.html}}.
We visually inspect the EPIC (pn+MOS) images from all three observations. A point source coincident with the optical position ($\alpha$=22:50:16.090, $\delta$=--44:51:53.50, \citealt{Wevers19b}) is present in both XMM1 and XMM2 images. However, no source is visible in the X-ray image of XMM3.
We start our data reduction with the raw Observation Data Files (ODFs) and process them using the {\it XMM-Newton} Standard Analysis System (SAS) version 17.0.0 tools {\tt epproc} and {\tt emproc}. Using the most recent version of the calibration database (CALDB) files this procedure results in ``cleaned'' event files, which are used for deriving scientific products. We extract good time intervals (GTIs) for each of the three observations by screening for periods of background flaring and considering only times when all detectors (pn+MOS1+MOS2) are active.
Source events are extracted from a circular aperture of radius 33 arcsec, which corresponds to roughly 90\% of light as estimated from the encircled energy function of the EPIC instruments. For each observation, background events are extracted from two circular regions of radius 50 arcsec close to the source and on the same CCD. The energy spectra and light curves are corrected for the different source and background areas. Standard data filters of {\it (PATTERN $<=$ 4) } and {\it (PATTERN $<=$ 12)} are applied for the pn and MOS data, respectively; we only use events in the 0.3--10 keV band for further analysis. Corresponding response files were generated using {\tt XMM SAS} tools {\tt rmfgen} and {\tt arfgen}.
For the XMM3 observation, we estimate the X-ray flux upper limit (3$\sigma$) using the EPIC/pn data by following the SAS documentation\footnote{\url{https://www.cosmos.esa.int/web/xmm-newton/sas-thread-src-find}}. Background flaring during XMM3 reduced the effective exposure from 17 ks to about 10 ks. Running the {\it XMMSAS} tool {\tt edetect\_chain} on a 0.3--10 keV pn image does not yield a point source at the location of AT2018fyk. We estimate the 0.3--10 keV count rate upper limit at the source location to be 0.0065 counts s$^{-1}$ using the sensitivity map generated by the {\tt edetect\_chain} task. We then use the {\tt fakeit} tool in the X-ray spectral fitting program {\it XSPEC} \citep{Arnaud96} and the {\it XMM-Newton} response files generated using the {\tt arfgen} and {\tt rmfgen} commands to estimate the flux upper limit. Because no constraints on the X-ray spectral shape are available, we estimate the upper limit using the best fit spectral model parameters from XMM1 and XMM2 with mean 0.3--10 keV count rates of 0.87 and 1.13 counts s$^{-1}$, respectively (see Table \ref{tab:xrayspectra}). Assuming the flux scales linearly in an energy-independent manner, 0.3--10 (0.01--10) keV 3$\sigma$ upper limits are 8.4$\times$10$^{-15}$ (7.8$\times$10$^{-14}$) and 1.2$\times$10$^{-14}$ (2.2$\times$10$^{-14}$) erg cm$^{-2}$ s$^{-1}$ for XMM1 and XMM2, respectively. We adopt the more conservative estimates of the XMM2 spectral model in our analysis.
\subsection{Chandra stare observation}
Following the non detection in the latest {\it XMM-Newton} exposure (XMM3) we obtained a deep {\it Chandra}/ACIS-S Director’s Discretionary Time (DDT) observation. The 46.8 ks exposure started 660 days after the optical discovery (MJD 59029.22). For better sensitivity at low energies, we used the back-illuminated S3 chip in timed exposure mode with the telemetry set to very faint format. Similar to the {\it XMM-Newton}/EPIC data reduction, we start our ACIS analysis with the level-1 (secondary) data files which we reprocess using the {\it Chandra} data analysis software (CIAO version 4.10). We reduce the level-1 data with the {\tt chandra\_repro} script with default parameter values.
We analyse the level-2 event files by first extracting a broadband (0.5--7 keV) X-ray image of the field of view. We use the {\tt fluximage} task with a {\it binsize} of 1 (0.492 arcsec). Visual inspection of the resulting exposure-corrected image does not show a source at the optical position of AT2018fyk. Assuming Poisson statistics and using the CIAO task {\tt srcflux}, we estimate the 0.3--10 keV flux upper limit using the best fit {\tt bremss+pow} model parameters from XMM1 and XMM2. Both model parameters give a similar value for the unabsorbed 3$\sigma$ flux upper limit in the 0.3--10 keV energy range of 2$\times$10$^{-15}$ erg cm$^{-2}$ s$^{-1}$ .
\subsection{NICER monitoring observations}\label{sec:nicerdata}
The {\it NICER} \citep{Gendreau16} X-ray Timing Instrument (XTI) onboard the International Space Station (ISS) began monitoring AT2018fyk on MJD 58388.15; 242 observations were made with individual exposures lasting between a few tens and two thousand seconds. We consider all publicly available observations taken before 2020 January 13 (MJD 58861.34; up to obsID 2200370264). These data are processed using the standard {\it NICER} Data Analysis Software (NICERDAS) task {\tt nicerl2} with default filtering and the gain file {\it nixtiflightpi20170601v004.fits}. Consecutive observations are combined to ensure a minimum of 20 ks of cleaned exposure. This yields a total of 31 energy spectra over the entire monitoring period.
All source spectra are extracted, and background spectra are estimated, as follows. Two empirical background spectrum libraries are constructed from observations of source-free areas of the sky. These include the cosmic X-ray (astrophysical) background associated with those fields as well as the {\it NICER}/XTI instrumental Non-X-ray-Background (NXB). The ISS night background library spectra are collected and categorised into discrete cells according to two background proxies. The first is the count rate of focused events in the 15--18 keV band, where the low effective area of the {\it NICER} optics assures that these are not X-rays from the source. The second is the rate of spatially extended (with respect to where X-rays are focused on the detector) events, as identified by their location in the plane of energy versus pulse invariant (PI) ratio of slow chain to fast chain detected events. These are the so-called ``trumpet’’ rejected events\footnote{\url{https://heasarc.gsfc.nasa.gov/docs/nicer/mission$\_$guide/}}. Corresponding rates in the source data are extracted using the input cleaned event file GTIs, subdivided into
intervals no larger than 120 s, and a match to one of the background library cells identified. The total night background is calculated as a sum over the night library spectra, appropriately weighted by fractional exposure and scaled by the 15--18 keV count rate in each
interval relative to average value in the cell. A supplemental ISS day residual background library is constructed from the same source-free observation database to account for an additional background associated with an optical light leak. This is used to derive an additional background component identified and scaled according to the rate in the 0--0.25 keV energy band.
The total spectrum extracted from the input event list is based on these subdivided GTIs, with the additional excision of data in intervals with background proxies that lie outside of the background library bounds. Also, subdivided GTIs where the absolute value of the estimated net 13--15 keV count rate exceeds 0.1 counts per second, or the estimated background rate exceeds the total rate, are excluded. Data from noisy Focal Plane Modules (FPMs) 14 and 34 are filtered out, as well.
\subsection{Swift monitoring observations}
{\it Swift} started monitoring AT2018fyk on MJD 58383.69, just two weeks after its discovery. Between MJDs 58383.69 -- 59031.24 a total of 161 snapshots were made, which accumulated 198 ks of exposure. All these observations were performed as part of multiple Target of Opportunity (ToO) requests over the $\sim$650 day period. All the data used here are publicly available and can be downloaded from NASA's High Energy Astrophysics Science Archive (HEASARC).
To measure the UV and optical photometry, a circular aperture of 7" is used on images taken by the {\it Swift} UV and Optical Telescope (UVOT; \citealt{Roming05}) images. We use the task \texttt{uvotsource} of HEASOFT v6.24. Sub-exposures of a given observation are combined before extracting the flux.
For the X-ray Telescope (XRT) data, we start our analysis with the raw level-1 data. These are reduced using {\tt xrtpipeline} with the source optical coordinates. We only use data taken in photon counting mode with event grades between 0 and 12. Source events are extracted from a circular region with a radius of 47''. This value corresponds to roughly 90\% (at 1.5 keV) of the light from a point source (as estimated from the XRT fractional encircled energy function). Background events are extracted from an annular region with inner and outer radii of 70'' and 150'', respectively. These values were chosen to avoid point sources in the background annulus. Spectra are binned to a minimum of 20 counts per bin for spectral analysis.
X-ray light curves are then extracted using the {\tt xrtlccorr} task, which properly takes bad pixels and columns into account. For each exposure we estimate the source and background count rates separately. Finally, background corrected source rates are estimated by subtracting the area-scaled background rates from the source rates. These steps are repeated for both the soft (0.3--1.5 keV) and the hard (1.5--10 keV) X-ray bands to extract the soft and the hard X-ray light curves.
To obtain a deeper constraint on the late time X-ray emission with {\it Swift}, we combine the individual exposures ranging from MJDs 58930 -- 58975. A 3$\sigma$ upper limit is derived on the combined image, taking into account the combined exposure map, using the {\it sosta} task in {\tt ximage}. The light curve data on a per obsID basis can be retrieved from the attached supplementary files ({\it XRT\_lightcurve\_30\_150.dat} and {\it XRT\_lightcurve\_150\_1000.dat}).
\subsection{Australia Telescope Compact Array observations}
\begin{table*}
\centering
\caption{ATCA observations of AT2018fyk. States are labelled as in Figure \ref{fig:alphaox}. Upper limits are reported for the stacked image made from the joint deconvolution of the 5.5 and 9.0\,GHz data. The corresponding luminosity upper limits are calculated assuming a central frequency of 7.25 GHz.}
\begin{tabular}{ccccccc}
\hline
State & MJD & Phase & On-source time & Config. & $3\sigma$ limit & L$_{\rm radio}$ \\
& (days) & (MJD - 58369) & (min) & & ($\mu$Jy\,beam$^{-1}$) & (erg s$^{-1}$)\\
\hline
C &$58594.14$ & 225 & 63 & 750C & $<104$ & $<$6.3$\times$10$^{37}$\\ %
C &$58646.99$ & 278 & 187 & 6A & $<27$ & $<$1.6$\times$10$^{37}$\\%
E &$58952.20$ & 583 & 145 & 6A & $<46$ & $<$2.8$\times$10$^{37}$\\%
\hline
\end{tabular}
\label{tab:atca}
\end{table*}
Radio observations were obtained with the Australia Telescope Compact Array (ATCA). Observations were performed at early times (25, 52 and 89 days after discovery; see \citealt{Wevers19b} for details of these observations) as well as at late times (225, 278 and 583 days after discovery), as detailed in Table~\ref{tab:atca}. For the new late time observations, we observed in the 4\,cm band to maximise sensitivity, with 2\,GHz of bandwidth in each of two frequency bands centred at 5.5 and 9.0\,GHz. We used the standard calibrator source PKS 1934-638 to determine the band pass solutions and to set the amplitude scale. We used the nearby calibrator source J2230-4416 to determine the complex gain solutions, which were interpolated onto the target. Data were reduced using the Common Astronomy Software Application (CASA; \citealt{McMullin2007}). AT2018fyk was not detected in either frequency band at any of the three late-time epochs, and we jointly deconvolved the data from both frequency bands to get the deepest possible upper limits on the source flux density, as reported in Table~\ref{tab:atca}.
\section{Analysis and results}
\label{sec:analysis}
\subsection{Host galaxy SED modelling}
\begin{figure}
\includegraphics[width=\columnwidth, keepaspectratio]{AT2018fyk_parasfh_wduste.pdf}
\caption{ Samples from the posterior distribution of host galaxy SED models. The open squares (and triangle) show detections (and one upper limit) from GALEX, DECaLS, and WISE. Open circles show the synthetic host galaxy magnitude, including predictions for the three bluest bands of the {\it Swift}/UVOT instrument.}
\label{fig:sed}
\end{figure}
To estimate the host galaxy brightness in the {\it Swift}/UVOT bands, we follow the approach of \citet{vanvelzen20}. We use the flexible stellar population synthesis code \citep{Conroy09} and pre-flare host galaxy photometry, including UV (GALEX), optical (DeCALs; \citealt{Dey19}) and IR (WISE; \citealt{Wright10}) observations to model the SED (see Table \ref{tab:hostmags}). The GALEX flux is extracted using the gPhoton software \citep{Million16}, using the same aperture (7") as applied to the UVOT images. For the WISE photometry we adopt the forced photometry values measured by the Legacy survey \citep{Dey19}.
\begin{table}
\centering
\caption{Measurements of the host galaxy brightness in UV, optical and IR filters, used to model the host SED. The model predictions are also presented. Apertures are matched (7 arcsec) for consistency.}
\begin{tabular}{ccc}
\hline
Filter & Measured magnitude & Model magnitude\\
& (mag) & (mag) \\
\hline
GALEX FUV & 23.17$\pm$0.24& 23.26$\pm$0.25\\
GALEX NUV & 22.15$\pm$0.15& 21.78$\pm$0.15\\
DeCALs g & 17.10$\pm$0.04& 16.94$\pm$0.03\\
DeCALs r & 16.34$\pm$0.03& 16.27$\pm$0.03\\
DeCALs z & 15.70$\pm$0.03& 15.66$\pm$0.03\\
WISE W1 & 16.00$\pm$0.03& 16.15$\pm$0.04\\
WISE W2 & 16.61$\pm$0.04& 16.81$\pm$0.04\\
WISE W3 & 16.9$\pm$0.5& 17.0$\pm$0.5\\
Swift UVW2 & ---& 21.98$\pm$0.10\\
Swift UVM2 & ---& 22.00$\pm$0.15\\
Swift UVW1 & ---& 20.56$\pm$0.08\\
Swift U & ---& 18.80$\pm$0.05\\
Swift B & ---& 17.38$\pm$0.04\\
Swift V & ---& 16.60$\pm$0.03\\
\hline
\end{tabular}
\label{tab:hostmags}
\end{table}
Markov chain Monte Carlo samples of the posterior distribution are used to constrain the model parameters: stellar mass, stellar population age, metallicity, star formation history e-folding time, and the optical depth of the dust (following the extinction law of \citealt{Calzetti2000a}). The best fit SED model (Figure~\ref{fig:sed}) is used to synthesise host galaxy magnitudes, which are subsequently subtracted from the measured photometry. The uncertainty on the host flux is propagated into the uncertainty on this difference photometry.
\subsection{UV/optical blackbody analysis}
\label{sec:uvblackbody}
\begin{figure}
\centering
\includegraphics[width=\columnwidth, keepaspectratio]{at2018fyk_bbevolution.pdf}
\caption{ Evolution of the blackbody radius (top panel) and temperature (bottom panel) over time. We show the full time resolution measurements (grey), as well as estimates corresponding to the binned data (red).}
\label{fig:bbrad}
\end{figure}
The UV/optical emission of TDEs can be well described as thermal blackbody emission \citep{vanvelzen11, Hung17}. We therefore measure the blackbody temperature and luminosity by fitting an evolving temperature blackbody model to the light curve. This also yields the blackbody radius, assuming isotropic emission. To smooth out variations induced by poorly constrained measurements, we use a 40 day smoothing length for the temperature measurements, and a 20 day window for the luminosity. Values are linearly interpolated in between grid points.
The blackbody radius and temperature evolution are shown in Figure \ref{fig:bbrad}. Individual epoch measurements are shown in grey; binned values are shown in red. The long term behaviour of the radius evolves slowly (decreasing only by a factor of $\sim$2) up to 450 days. Similarly, the temperature does not cool significantly over the entire light curve, with variations within the error budget. This is somewhat atypical of other TDEs, although there are sources that show a similar evolution \citep{vanvelzen20}. We do note that during the soft state, there is some temperature evolution consistent with an L $\propto$ T$^4$ behaviour (Figure \ref{fig:templum}). At very late times (560 days after discovery) we find that the radius has decreased significantly, while the temperature (although relatively uncertain) does not show any evidence for significant change.
We interpret this lack of significant temperature and radius evolution during the first 500~d as suggestive evidence that the UV emission originates in a stable, rapidly formed accretion disk. At late times, the luminosity decreases significantly, leading to the inferred decrease in radius.
\begin{figure}
\centering
\includegraphics[width=\columnwidth, keepaspectratio]{templum_allstates.pdf}
\caption{TDE UV luminosity as a function of blackbody temperature. The black solid line shows L$\propto$T$^4$ behaviour (with arbitrary normalisation), which roughly describes the behaviour in the soft state (A; blue markers) well. The different states are coloured according to Figure \ref{fig:alphaox}.}
\label{fig:templum}
\end{figure}
\subsection{The UV--X-ray spectral slope {$\alpha_{\rm ox}$\,} and bolometric Eddington fraction}
\label{sec:alphaox}
\begin{figure*}
\includegraphics[width=\textwidth, keepaspectratio]{alphaox_fedd_lbol_recent_4panel_grey.pdf}
\caption{Light curves and {$\alpha_{\rm ox}$\,} evolution of AT2018fyk. Top left: {\it Swift}/XRT 0.3--10 keV (black triangles) and {\it Swift}/UVW1 (green circles) light curves. The black star, square and diamond represent {\it Swift}, XMM3 and {\it Chandra} X-ray upper limits, respectively. The vertical coloured lines denote transitions in source properties as detailed in Section \ref{sec:alphaox} and Table \ref{tab:phases}, while vertical black dashed lines indicate long stare X-ray observations.
Top right: {$\alpha_{\rm ox}$\,} as a function of the bolometric Eddington fraction. A representative error bar is shown in black in the top left; the uncertainty in {f${_{\rm Edd, bol}}$\,} is dominated by the intrinsic scatter in the M\,--$\sigma$ relation. The typical transition luminosity of stellar-mass black holes is shown as a grey shaded band. Bottom panels: XMM high cadence (bin size of 200 s) lightcurves during the XMM1 (left) and XMM2 (right) observations. Background rates are shown in grey, offset (for display purposes) by +0.4 and +0.5 for XMM1 and XMM2, respectively.}
\label{fig:alphaox}
\end{figure*}
To probe the broad band long term evolution of the accretion flow, we employ the UV to X-ray spectral index {$\alpha_{\rm ox}$\,} \citep{Tananbaum79}:
\begin{equation}
\alpha_{\rm ox} = 1 - \frac{log_{10}(\lambda L_{2500}) - log_{10}(\lambda L_{2 keV})}{log_{10}(\nu_{2500}) - log_{10}(\nu_{2 keV})}
\end{equation}
where $\lambda$L$_{2500}$ and $\lambda$L$_{2 keV}$ are the monochromatic luminosities at 2500 \AA\ and 2 keV, respectively, and $\nu_{2500}$ and $\nu_{2 keV}$ are the frequencies at those wavelengths. We use the {\it Swift} UVW1 host-subtracted luminosities ($\lambda_{\rm cen}$ = 2629 \AA) as a proxy for L$_{2500}$. These data are included as supplementary data files. Based on the observed behaviour of {$\alpha_{\rm ox}$\,} with bolometric Eddington ratio (Figure \ref{fig:alphaox}, top right panel), we divide the TDE evolution into 5 phases (A through E). These phases are defined based on a combination of {$\alpha_{\rm ox}$\,}, X-ray spectral and timing properties as well as X-ray and UV variability properties.
We summarise the observational properties for each phase in Table \ref{tab:phases}, but defer a detailed discussion to Section \ref{sec:discussion}.
\begin{table*}
\centering
\caption{Observational properties for each phase labelled in Figure \ref{fig:alphaox}. The phase is referenced with respect to the discovery epoch. f$_{\rm Edd}$ is the bolometric Eddington fraction of emission. BB stands for blackbody; PL for power-law.}
\begin{tabular}{lcccccccc}
\hline
Phase / State & MJD & Phase & {$\alpha_{\rm ox}$\,} & f$_{\rm Edd}$ & X-ray spectral state & X-ray PSD power & UV var. \\
\hline
A / Soft & 58383 -- 58446 & 14 -- 77 & 2.0 & $\sim$0.1 & BB dominated & Low $\nu$ ($<$10$^{-5}$ Hz) & Yes \\
B / Transition & 58447 -- 58574 & 78 -- 205&1.6 & $\sim$0.05 & --- & Low $\nu$ ($<$10$^{-5}$ Hz) & Yes \\
C / Hard & 58576 -- 58717 & 207 -- 348&1.2 & $\sim$0.03 & PL dominated & Low+high $\nu$ ($<$10$^{-3}$ Hz) & No \\
D / Hard & 58721 -- 58858 & 352 -- 489&1.2 & $\sim$0.03 & PL dominated &Low+high $\nu$ ($<$10$^{-3}$ Hz) & Yes \\
E / Quiescent & 58930 -- ... & 561 -- ... & $>$1.9 & $<$0.0004 & --- & --- & ---\\
\hline
\end{tabular}
\label{tab:phases}
\end{table*}
The UV blackbody temperature does not cool significantly over the 650 day lightcurve, with an average temperature of T$\sim$35000 K.
The UVW1 filter is on the Rayleigh-Jeans tail of the blackbody spectrum with log$_{10}$(T)=4.63 K. Similar to \cite{vanvelzen2019}, we apply a temperature correction to convert the observed luminosity L$_{UVW1}$ to a bolometric UV luminosity estimate L$_{\rm UV}$, integrated from 0.03 to 3 $\mu$m. This correction amounts to a factor of $\sim$6.1, which is adopted for the entire light curve because of the lack of temperature evolution. We further include the integrated X-ray emission in the 0.01--10 keV energy range (including the EUV band) by extrapolating the spectral model from the observed band to define the bolometric Eddington fraction of emission {f${_{\rm Edd, bol}}$\,} $\equiv$ (L$_{0.01-10 \rm keV}$ + L$_{\rm UV}$) / L$_{\rm Edd}$.
A high value of {$\alpha_{\rm ox}$\,} indicates a UV (disk) dominated system (also dubbed a soft state), while a low value of {$\alpha_{\rm ox}$\,} indicates that the X-ray (corona) emission dominates the energy output (i.e., a hard state).
We use the spectral fits (Section \ref{sec:xrayfitting}) to determine the monochromatic 2 keV X-ray flux L$_{2 keV}$. Given the observed spectral variability with time, we use a spectral model for each of the states A (soft), B (transition) and C+D (hard), and take the best fit power-law index and the appropriate power-law fraction to compute L$_{2 keV}$ for each epoch (Table \ref{tab:xrayspectra}). Because the temperature of the thermal X-ray component is low, the power-law dominates the X-ray flux $>1.5$ keV at all times, and we ignore the thermal contribution in our estimate of L$_{2 keV}$ (i.e. we only consider the power-law component). Uncertainties are propagated using the standard rules.
We note that in the final epoch of {\it Swift} UV observations, the UVM2 brightness is consistent with the SED model prediction for the host galaxy (Table \ref{tab:hostmags}). The UVW1 and UVW2 magnitudes are still slightly elevated when compared to the prediction for the host galaxy. This could suggest that the SED model magnitudes are slight underestimates of the true host brightness. We assess the impact of this difference, in particular for the UVW1 filter, by repeating all calculations while assuming a host galaxy brightness equal to that of the last {\it Swift} UVW1 observation, rather than that of the best fit SED model. The impact on both {$\alpha_{\rm ox}$\,} and {f${_{\rm Edd, bol}}$\,} is small, with changes of up to $\approx 0.15$ and 0.05, respectively, which doesn't influence the main conclusions of our work.
The left panel of Figure \ref{fig:alphaox} shows the light curves that were used to compute {$\alpha_{\rm ox}$\,} and {f${_{\rm Edd, bol}}$\,}; note that epochs with only an X-ray or UVW1 measurement are discarded.
The source spends approximately 65 days in state A, 125 days in state B (although this is an upper limit due to lack of observations in a seasonal gap), and $\sim$280 days in the hard state (C and D) afterwards. State C and D are spectrally identical in X-rays, but the latter is characterised by large amplitude variations in the UV band. At very late times (state E), the UV luminosity decreases by a factor $\sim$15, while the X-ray flux drops by a factor of $>$5000 compared to the last detection with {\it Swift}.
The right panel shows the corresponding behaviour in {$\alpha_{\rm ox}$\,} and {f${_{\rm Edd, bol}}$\,}.
\subsection{X-ray spectral analysis}
\label{sec:xrayfitting}
\subsubsection{X-ray spectral state transition}
We employ a two-component phenomenological spectral model to characterise the X-ray spectral evolution following the TDE. We use the 0.3--10 keV band, unless stated otherwise. Our focus is on the evolution of the relative strengths of these two model components during the different phases of the TDE. We start our analysis with the XMM1 and XMM2 EPIC/pn spectra, which contain $\approx$12\,000 and 28\,000 counts for the early and late epochs, respectively. We also tabulate the results of the spectral fitting of the XRT stacked spectra (Table \ref{tab:xrayspectra}) and time-resolved {\it NICER} stacked spectra (0.4-2 keV; Supplementary data file). All analysis is performed using \textsc{xspec} version 12.10.0 in HEASOFT v6.24. Best-fit model parameter uncertainties are estimated using the {\tt error} command in \textsc{xspec}.
We find no evidence for additional intrinsic absorption in the earliest {\it Swift} spectrum, which might be expected if the disk is initially in a slim state \citep{Wen20} (recalling that we find a peak mass accretion rate $\sim$ the Eddington rate). None of the later spectra show evidence for additional neutral absorption either. Therefore, we fix the hydrogen column density to the Galactic foreground value of n$_H$ = 1.15$\times$10$^{20}$ cm$^{-2}$ \citep{hi4pi} for all subsequent analysis.
For spectral analysis we only used the EPIC-pn data. We bin the two EPIC/pn spectra to a minimum of 25 counts per spectral bin, and oversample by a factor of 3.
We first attempt to fit the spectra using a single component: a blackbody model ({\tt bbody}), as well as multi-colour blackbody ({\tt diskbb}), power-law ({\tt powerlaw}) and thermally Comptonised continuum models ({\tt Nthcomp}). None of these describe the data well (reduced chi-squared value $\chi_r > 1.6$). Strong systematic residuals are apparent in all cases.
\begin{figure*}
\includegraphics[width=\textwidth, keepaspectratio]{XMM_model_ratio.pdf}
\caption{XMM-Newton EPIC/pn X-ray spectra and best fit models. {\it XMM-Newton} PN spectra (black), overlaid by two best fit models (diskbb+powerlaw in orange, bremsstrahlung+powerlaw in blue). Dashed lines represent the power-law model, while dotted lines show the thermal component and solid lines are the sum of the two. The bottom panels show the ratio of data and model. The power-law spectral component is stronger in XMM2 than during XMM1.}
\label{fig:xmmspectra}
\end{figure*}
We next employ a thermal model to describe the soft energies ($<$1.5 keV) and a power-law model to describe the higher energies (see Figure \ref{fig:xmmspectra}). Fitting a power-law to higher energies only, we find a very large soft excess below 1.5 keV for XMM1 (factor $>$30 at 0.3 keV), suggesting that relativistic reflection models cannot account for the spectral shape, although we did not explore detailed reflection spectral modelling. We note that the soft excess we refer to is soft X-ray emission in addition to power-law X-ray emission, and does not refer to the pure blackbody spectra observed in some TDEs. We next add a thermal component to describe the soft energies. We first try {\tt TBabs$\times$zashift$\times$(diskbb+powerlaw)} in {\it XSPEC}, which yields temperatures $kT$ = 123 and 146 eV for XMM1 and XMM2, respectively. The power-law index $\Gamma \sim$3.3$\pm$1.0 is steep at early times (although the errors are large), but at late times becomes well constrained $\Gamma$ = 2.1$\pm$0.1. We note that the reduced $\chi^2$ is marginally statistically acceptable for both fits (reduced chi square of $\chi_r = 1.29$ and 1.32, respectively; see Table \ref{tab:xrayspectra}). We find significant evolution in the power-law contribution to the total X-ray flux between XMM1 and XMM2, increasing from $\sim$0.25 to 0.64. Using a more physically motivated model, {\tt TBabs$\times$zashift$\times$(diskbb + nthcomp)}, where the Comptonising seed photon temperature is linked to the accretion disk temperature, yields similar results.
Employing instead a bremsstrahlung model ({\tt bremss}) for the thermal component provides better quality of fit. Using this model, we find plasma temperatures of $kT$ = 210 and 290 eV, while the power-law indices remains constant at $\Gamma \sim$ 1.9, for $\chi_r$ = 1.08 and 1.07, respectively. In agreement with the {\it diskbb} model results, we find that the power-law fraction of emission increases from near negligible (0.05$\pm$0.03) to providing a dominant (0.54$\pm$0.04) contribution to the total X-ray flux.
\subsubsection{Power-law fraction of emission}
\begin{figure*}
\includegraphics[width=\textwidth, keepaspectratio]{PLfraction_2panel.pdf}\\
\caption{Long-term evolution of the X-ray corona. Left: the fractional contribution of the power-law component to the total X-ray flux, i.e., a measure of X-ray corona's strength, derived from {\it NICER} (squares), {\it Swift} (stars) and {\it XMM-Newton} (triangles) spectral fits in the 0.4--2 keV band. Colours are the same as in Figure \ref{fig:alphaox}. The results are consistent with spectral fits in the 0.3--10 keV energy range of {\it Swift} and {\it XMM-Newton} (right panel and Table \ref{tab:xrayspectra}).
Right: fractional contribution of the 2--10 keV power law emission to the total bolometric luminosity, as derived from {\it Swift} measurements. Errors are added in quadrature, but exclude the uncertainty in the M--$\sigma$ relation ($\sim$0.4 dex). The corona contribution is negligible at early times but becomes very significant after +200 days, indicating the strengthening of the corona over time.}
\label{fig:corfrac}
\end{figure*}
Time-resolved spectra from {\it NICER} over the first $\sim$ 450 d provide further clear evidence for significant spectral evolution. Background subtracted, unbinned spectra are fit to the same thermal + power-law models employed for the XMM analysis in {\tt xspec}, using Cash statistics \cite{Cash79}. The fits are conducted in the 0.4--2 keV band, and reported luminosities refer to this energy range. This particular band pass was chosen to minimise the background contamination (see Section \ref{sec:nicerdata} for more details). The effects of imperfect background subtraction can become significant when the power-law flux is low. For power-law fluxes $<$10$^{-13}$ erg cm$^{-2}$ s$^{-1}$, the power-law index is therefore fixed at the best fit value of the composite ($\sim$310 ks) spectrum to mitigate these effects.
The fractional contribution of the power-law component with respect to the total flux is estimated from each of these fits. The evolution of the power-law fraction of X-ray emission, i.e. the relative strength of the corona, are shown in Figure \ref{fig:corfrac} (left panel) for the {\tt diskbb+powerlaw} model. For consistency, we also refit the {\it Swift} and {\it XMM-Newton} spectra in the 0.4--2 keV energy range; we note that these results are consistent with those obtained from 0.3--10 keV spectral fits for {\it Swift} and {\it XMM-Newton} observations. The spectra at early times (state A and early state B) are dominated by a thermal component, which provides 50-–95 \% of the X-ray flux (0.3--10 keV), indicating the presence of a weak corona. In the hard states the corona strengthens to provide $\sim 60-80$ per cent of the X-ray flux. A similar increase in power-law fraction from near negligible ($<15$ per cent) in the soft state to very significant (20--75 per cent) in the hard state is also found for the {\tt bremss+powerlaw} model. To remain agnostic about the nature of the soft excess emission, we perform a similar calculation but this time taking the ratio of power-law flux over the bolometric emission (Figure \ref{fig:corfrac} right panel). This yields a qualitatively very similar picture, with an initially weak corona that subsequently strengthens over time. This evolution is further corroborated by the {\it Swift} and {\it XMM-Newton} data (Figure \ref{fig:swift_hr}), which shows a steady increase in the fraction of hard X-ray count rate to total X-ray emission. \\
We conclude that there is significant spectral evolution in AT2018fyk, where the corona is energetically unimportant at early times, while it produces more than half of the observed X-ray flux, and a significant fraction of the bolometric emission, at late times. This evolution is consistent with a spectral state transition from a soft state to a hard state, analogous to those seen in XRBs, AGN \citep{Done05, Remillard06} and a sample of X-ray bright TDEs \citep{Wevers20}.
\begin{figure}
\includegraphics[width=\linewidth, keepaspectratio]{Swift_hr.pdf}
\includegraphics[width=\linewidth, keepaspectratio]{XMM_hardtotal.pdf}
\caption{Hard to total count rate ([1.5--10 keV]/[0.3--10 keV]) as a function of time as measured by {\it Swift}/XRT (top) and XMM-Newton/EPIC (bottom). The evolution is consistent with results from time-resolved spectra shown in Figure \ref{fig:corfrac}, i.e., hard X-rays dominate in states C and D.}
\label{fig:swift_hr}
\end{figure}
\subsubsection{Rapid spectral variability}
In terms of rapid ($\sim$minutes to tens of hours timescale) spectral variability, the hardness ratio of the X-ray emission as observed by {\it Swift} and {\it XMM-Newton} can vary on short timescales (Figure \ref{fig:swift_hr}). However, this variability does not obviously correlate with the X-ray flares.
We do not have high quality spectroscopic X-ray data to assess spectral variability on rapid timescales of tens of minutes (e.g. as seen in quasi-periodic eruptions \citealt{Miniutti19, Giustini20}). Analysis of {\it Swift} spectra stacked based on their X-ray count rate does not reveal significant spectral changes.
The {\it XMM-Newton} hardness fraction is constant during the soft state (XMM1) with an average value of 0.016$\pm$0.015, i.e. consistent with 0. However, the hardness fraction varies smoothly between 0.07 and 0.22, with an average value of 0.15$\pm$0.03 during the hard state (XMM2). Interestingly, roughly 27000 s after the start of the exposure there is a hardness flare that coincides with a flux enhancement in the light curve (see Figures \ref{fig:alphaox} and \ref{fig:swift_hr}).
\begin{deluxetable*}{cccccccccccc}
\tablecaption{Best fit parameters obtained from X-ray spectral modelling of {\it Swift} and {\it XMM-Newton} data. The states are as labelled in Figure \ref{fig:alphaox}; states C and D are spectrally identical, and only 1 Swift stacked spectrum is produced covering both. The mean count rate for each spectrum is given in the second column. The effective exposure time t$_{\rm exp}$ is given in kiloseconds. The spectral model used is {\tt TBabs$\times$zashift$\times$(model + powerlaw)}, where model is listed in the table; kT is the temperature of the thermal component, while $\Gamma$ denotes the power-law spectral index. Normalisations for the thermal and power-law components are listed in the norm(kT) and norm($\Gamma$) columns. The flux is integrated from 0.3--10 keV. PL frac denotes the fractional contribution of the power-law component to the total X-ray flux. The final column lists the reduced $\chi^2$ and degrees of freedom (dof).\label{tab:xrayspectra}}
\tablewidth{700pt}
\tabletypesize{\scriptsize}
\tablehead{
\colhead{Spectrum} & \colhead{Count rate} & \colhead{State} & \colhead{t$_{\rm exp}$} &
\colhead{Model} & \colhead{kT} & \colhead{norm(kT)} & \colhead{$\Gamma$} &
\colhead{log$_{10}$(norm ($\Gamma$))} & \colhead{log$_{10}$(flux)} & \colhead{PL frac} & \colhead{$\chi^2$ (dof)} \\
\colhead{} & \colhead{(cts s$^{-1}$)} & \colhead{} & \colhead{(ks)} &
\colhead{} & \colhead{(eV)} & \colhead{} &
\colhead{} & \colhead{} & \colhead{(erg cm$^{-2}$ s$^{-1}$)} & \colhead{}
}
\startdata
XRT & 0.037 & A & 37.6 & diskbb & 142$\pm 10$ & 260$\pm$150& 2.8$^{+1.1}_{-1.0}$ & --4.2$\pm$0.3 & --11.89$\pm$0.03 & 0.25$\pm 0.13$ & 1.30 (40) \\\vspace{1.5mm}
& & & & bremss & 260$\pm 25$ & 0.0037$\pm$0.0005 & 1.8$^{+1.4}_{-1.1}$ & --4.7$\pm$0.5& --11.87$\pm 0.03$ & 0.10$^{+0.12}_{-0.04}$ & 1.24 (40) \\\vspace{1.5mm}
EPIC/pn & 0.87 & B & 13.8 & diskbb & 123$^{+8}_{-4}$ & 506$^{+103}_{-238}$ & 3.3$^{+1.1}_{-1.0}$ & --4.5$^{+0.2}_{-0.3}$ & --11.96$\pm$0.02 & 0.27$^{+0.26}_{-0.20}$ & 1.29 (26) \\\vspace{1.5mm}
& & & & bremss & 210$\pm 8$ & 0.0050$\pm$0.0004 & 1.8$^{+0.8}_{-0.7}$ & --5.0$\pm$0.3& --11.94$\pm 0.01$ & 0.05$\pm 0.03$ & 1.08 (26) \\\vspace{1.5mm}
XRT & 0.097 & B & 30.0 & diskbb &166$\pm 15$ & 160$^{+80}_{-50}$ & 2.11$\pm 0.17$ & --3.37$\pm$0.06& --11.44$\pm 0.02$ & 0.63$\pm 0.07$ & 0.82 (95) \\\vspace{1.5mm}
& & & & bremss &330$\pm 50$ & 0.0034$\pm$0.0006& 2.0$\pm 0.2$ &--3.43$\pm$0.03 & --11.43$\pm 0.02$ & 0.43$^{+0.05}_{-0.10}$ & 0.86 (95) \\\vspace{1.5mm}
XRT & 0.13 & C+D & 110.4 & diskbb &179$\pm 8$ & 131$\pm$25 & 2.19$\pm 0.08$ & --3.19$\pm$0.05 & --11.31$\pm 0.01$ & 0.68$\pm 0.04$ & 0.92 (214) \\\vspace{1.5mm}
& & & & bremss &372$\pm 26$ & 0.0036$\pm$0.0003 & 2.06$\pm 0.11$ & --3.27$\pm$0.05 & --11.31$\pm 0.01$ & 0.59$\pm$0.05 & 0.98 (214) \\\vspace{1.5mm}
EPIC/pn & 1.13 & D & 24.6 & diskbb & 146$\pm 7$ & 171$^{+40}_{-30}$ & 2.10$\pm 0.08$ & --3.60$\pm$0.03 & --11.69$\pm 0.01$ & 0.64$\pm 0.03$ & 1.32 (90) \\\vspace{1.5mm}
& & & & bremss & 292$\pm 19$ & 0.0024$\pm$0.0002 & 1.93$\pm 0.09$ & --3.69$\pm$0.05 & --11.67$\pm 0.02$ & 0.54$\pm 0.04$ & 1.07 (90) \\
\enddata
\end{deluxetable*}
\subsection{X-ray timing analysis}
\label{sec:xraytiming}
\begin{figure*}
\includegraphics[width=0.5\textwidth, keepaspectratio]{psdfit_lf1soft_nuPnu.pdf}
\includegraphics[width=0.475\textwidth, keepaspectratio]{contours_lf1soft2.pdf}
\includegraphics[width=0.5\textwidth, keepaspectratio]{psdfit_lf1_nuPnu.pdf}
\includegraphics[width=0.46\textwidth, keepaspectratio]{contours_lf1.pdf}
\includegraphics[width=0.5\textwidth, keepaspectratio]{psdfit_lf0_nuPnu.pdf}
\includegraphics[width=0.49\textwidth, keepaspectratio]{contours_lf0.pdf}\\
\\
\caption{Power spectra and confidence contours in the soft (top panels) and hard states (middle/lower panels). Left: power density spectra. Observations are shown as solid black lines; the best fit distorted model in red; the undistorted model in blue; and Poisson noise levels are indicated in grey. The soft state is shown in the upper panel; no bend in the power-law model is required to reproduce the data. For high frequencies ($\gtrsim$10$^{-6}$ Hz) the data (black) is consistent with Poisson noise (grey curve). The middle and lower panels shows the hard state, with the low frequency slope fixed to $\alpha_{LF}$ = --1 and $\alpha_{LF}$ = 0, respectively. Variability on frequencies between 10$^{-7}$ to 10$^{-3}$ Hz is evident; a bend is required to reproduce the data. Right panels: corresponding confidence contours between the bend frequency and the high-frequency slope of the power spectrum for 1, 2 and 3-sigma confidence regions (solid black, red dashed, blue dotted lines), corresponding to $\delta \chi^{2} \simeq 2.3$, 6.18 and 11.83 respectively. Black crosses mark the best fit set of parameters.}
\label{fig:psresp_comparison}
\end{figure*}
To constrain the power spectra for the intervals of the light curve identified with the soft (state A) and hard (state C+D) states (including the corresponding {\it XMM-Newton} light curves) we fit the power spectra using the power-spectral response (PSRESP) method \citep{Uttley02}. PSRESP compares the observed power spectra with those of simulated light curves for a given power-spectral shape, resampled to match the gaps in the sampling of the real light curves, so that the distorting effects of sampling on the power spectrum are properly accounted for.
For the {\it Swift}/XRT light curves we first rebin the light curves to give one flux measurement per observation and remove a small amount of data with signal to noise (S/N) $<$1. The light curves are then further rebinned into 2-day bins (slightly larger than the average sampling interval of 1.8 d in states C+D and 1.2 d in state A) and gaps are filled by linear interpolation. The {\it XMM-Newton} light curves are binned at 200 s resolution and gaps are also filled by linear interpolation. We calculate the observed power spectra from the resulting light curves, binning the log-power values geometrically in frequency by a factor of 1.5 (with a minimum of 2 data points per frequency bin).
The Monte Carlo simulated underlying time-series are generated using a sharply bending power-law model, parameterised by low frequency index $\alpha_{\rm LF}$, high frequency index $\alpha_{\rm HF}$ and bend frequency $\nu_{\rm bend}$. For the {\it Swift}/XRT data the simulation time-resolution is set to 5760 s to roughly match the Swift orbital period and ensure that there are 30 simulated data points per 2-day bin. We account for the missing simulated power on time-scales from 5670 s down to the duration of each Swift snapshot ($\sim$1 ks) by calculating the rms of the model on those time-scales and adding a Gaussian random number with the same standard deviation to each sampled simulated data point. By accounting for random variations, this is an advance over the original PSRESP, which accounted for this effect by adding a constant component to the simulated power spectrum and assuming that the log-power errors scaled with those of the simulation.
To match the observed data, 200s sampling is used for the {\it XMM-Newton} light curve simulations. In the PSRESP software, the simulations are resampled to match the sampling of the unbinned light curves, and rebinned and interpolated in the same way as the data, so the effects on the power spectrum are accounted for.
For each observed light curve, we generate 1024 simulated light curves as segments cut from a single very long light curve (to account for low frequency 'red-noise leak', \citealt{Uttley02}). Unlike in the original PSRESP, which added observational ('Poisson') noise as a constant power spectral component, we include the effects of observational noise by simulating separate noise data sets from Gaussian errors which match the observed errors on each data point, resampling, interpolating and adding the resulting noise power spectrum to the simulated light curve power spectrum. By separating the noise contribution in this way, we can use an arbitrary normalisation to fit the simulated intrinsic power spectra to the data, without needing to specify the observed S/N as an additional variable to be stepped through in the parameter search.
As in the original PSRESP, the simulated light curves are used to calculate the mean and standard deviation on the log-power in each frequency bin, and for a given set of model parameters we can compare this with the observed power spectrum to yield a `pseudo' $\chi^{2}$ value for the distorted power spectrum, $\chi^{2}_{\rm dist}$ of a given pair (Swift/XRT and XMM-Newton) of light curves. We minimise $\chi^{2}_{\rm dist}$ to obtain the best fit normalisation for the given shape parameters of the power-spectral model (power-law indices and bend frequency). By comparing the $\chi^{2}_{\rm dist}$ of the observed power spectrum with those of the simulations for a given model, we can estimate a goodness-of-fit for that model.
In the original PSRESP, contours of goodness-of-fit are used to obtain errors on the power-spectral parameters. However as noted by \cite{Marshall15}, this approach can lead to acceptance regions which are too small when the best fit goodness-of-fit is itself low. From the comparison of $\chi^{2}_{\rm dist}$ of the best fit models with the goodness-of-fit derived from the simulations, we are able to identify that for this particular set of light curve samples, $\chi^{2}_{\rm dist}$ behaves close to an undistorted $\chi^{2}$ distribution. Therefore, we estimate confidence contours on the model parameters using the standard $\Delta \chi^{2}$ approach (which is equivalent to likelihood ratio, for normally distributed errors).
To limit the parameter space covered, we consider two `typical' values of low frequency slope, $\alpha_{\rm LF}=0$ (corresponding to the low frequency shape of the `band-limited noise' seen in black hole XRB hard states) and $\alpha_{\rm LF} = -1$, corresponding to the intermediate frequency slope of `broadband noise' seen in faint black hole XRB hard states, or the low frequency slope that is seen in some soft states seen of black hole XRBs (e.g. see \citealt{Heil15} for a comparison of the range of XRB power-spectral shapes). For each of these two low frequency slopes, we search a parameter space covering a wide range of bend frequency (searched in equally-spaced intervals of log(bend-frequency) and high frequency slope (searched in equal intervals of the power-law index). We obtain confidence intervals on the PDS by Monte Carlo likelihood sampling of the fitted model parameters, the bend frequency and the high frequency slope.
Both low frequency shapes, $\alpha_{\rm LF}=0$ and $\alpha_{\rm LF} = -1$, provide adequate fits to the hard state data over some of their parameter space, with best fit goodness-of-fits of 0.28 and 0.19 respectively. The confidence contours for the two types of model considered and plots of the best fit model (distorted and underlying) versus the data, are shown in Figure \ref{fig:psresp_comparison}. The best fit parameters, their 1-dimensional (1-$\sigma$) errors and fractional rms (integrated from the model over $10^{-7}$--$10^{-3}$ Hz) are provided in Table \ref{tab:psresp}. The high frequency slopes we measure are similar to those measured for AGN samples, which have slopes of $\alpha_{\rm HF} \approx -2$ (if no bend is required for the PDS) to --3 (if the PDS requires a bending power-law; \citealt{Gonzalez12}). Note that the high frequency slope is not well constrained for large negative indices due to red-noise leak effects \citep{Uttley02}.
Since either model provides a good fit to the hard state interval, we cannot distinguish the type of hard state solely on the basis of the power-spectral shape, but we note that the inferred broadband fractional rms of 55~per~cent is much more consistent with black hole X-ray binary hard states than with the soft states (which show rms $<$a few per cent).
\begin{deluxetable*}{cccccccc}
\tablecaption{Results of the power-spectral fitting. The rms values are reported in various frequency bands. The black hole mass is calculated using the bend frequency in combination with the relation of \cite{Mchardy06}. Values marked with an asterisk have negative error bars bounded by the parameter range. $^{\dagger}$99\% confidence upper limit. ${^\ddagger}$99\% confidence lower limit (the upper range of rms is not well constrained since the low frequency power-spectral shape is not well constrained by the soft state light curve).}
\label{tab:psresp}
\tablewidth{700pt}
\tablehead{
\colhead{State} & \colhead{log($\nu_{\rm bend}$)} & \colhead{$\alpha_{\rm HF}$} & \colhead{rms (\%)} &
\colhead{rms (\%)} & \colhead{rms (\%)} & \colhead{rms (\%)} & \colhead{log(M$_{BH}$)} \\
\colhead{} & \colhead{(Hz)} & \colhead{(\footnotesize{10$^{-7}$--10$^{-3}$ Hz})} & \colhead{(\footnotesize{10$^{-7}$--10$^{-6}$ Hz})} &
\colhead{(\footnotesize{10$^{-6}$--10$^{-5}$ Hz})} & \colhead{(\footnotesize{10$^{-5}$--10$^{-3}$ Hz})} & \colhead{(M$_{\odot}$)}
}
\startdata
Hard ($\alpha_{\rm LF}=0$) & --5.00$_{-0.27}^{+0.33}$ & --2.65 + 0.5$^{*}$ & 54$\pm$1 & 13$_{-3}^{+4}$ & 39$^{+5}_{-8}$ & 35$^{+7}_{-8}$ & 7.25$\pm$0.55\\%\
Hard ($\alpha_{\rm LF}=-1$) & --4.26$_{-0.20}^{+0.33}$ & --2.75 + 0.5$^{*}$ & 55$_{-3}^{+1}$ & 32$^{+4}_{-3}$ & 32$^{+5}_{-8}$ & 31$^{+7}_{-8}$ & 6.9$\pm$0.55\\%\
Soft & $<$ --5.5$^{\dagger}$ & --2.95 + 0.5$^{*}$ & 21$^{\ddagger}$ & 32$_{-7}^{+4}$ & 7$^{+2}_{-1}$ & 1$^{+1}_{-0.5}$ & ---\\\hline
\enddata
\end{deluxetable*}
\section{Discussion}
\label{sec:discussion}
The emission of the outer accretion disk around a 10$^{7.7}$ M$_{\odot}$ black hole peaks in the UV band, which we monitored using the Neil Gehrels {\it Swift} satellite. {\it Swift} also performed monitoring observations in X-rays (0.3--10 keV), which contains information about the poorly understood components referred to as the soft excess (dominating the 0.3--1.5 keV band) and the X-ray corona (emitting primarily between 2 and 10 keV). The nature of the former component is still unclear, but it could originate in the inner accretion disk either directly (although the observed temperature may be too high for this in AT2018fyk) or from a low temperature Comptonisation component in the inner regions, while the standard corona likely originates in a hot plasma above the inner disk. We also monitored for radio emission to search for signatures of a newly launched jet.
These data provide three key diagnostics to characterise the accretion properties following the TDE, which together enable detailed comparison with stellar-mass black holes.
First, the long term light curves (top-left panel of Figure \ref{fig:alphaox}) allow us to track the evolution of the spectral energy distribution (SED), using the UV to X-ray spectral slope {$\alpha_{\rm ox}$\,} (Section \ref{sec:alphaox}). This quantity represents the relative strength of the accretion disk (in the UV at 2500 \AA) and the corona (at 2 keV), and serves as a proxy for accretion state. In stellar-mass black holes, the strength of the disk relative to the corona changes as a function of the Eddington ratio {f${_{\rm Edd, bol}}$\,} \citep{Fender04, Remillard06}. Appropriate mass scaling of the relative component strengths, represented by {$\alpha_{\rm ox}$\,} in SMBHs, also reproduces the observed properties of large samples of AGN and relatively fast evolving changing-look AGN \citep{Sobolewska11a,Ruan19a}; we reiterate that there is potentially a significant (systematic) uncertainty in the black hole mass estimate for AT2018fyk, which should be taken into account when making the comparison to (CL-)AGN behaviour. Given that the uncertainty is systematic, the shape of the evolution throughout the diagram will not change as a result.
Second, to characterise the X-ray spectra in the different phases, we use a two component spectral model (Section \ref{sec:xrayfitting}), including a thermal component to describe the soft energies (0.3--2 keV) and a power-law to account for the harder X-rays (2--10 keV). The key diagnostics are the ratio of the power-law spectral component flux to the total emission in X-rays, and the ratio of power-law emission to the total bolometric luminosity (including the accretion disk, soft excess and corona). These quantities are indicators of accretion state, and their evolution is presented in Figure \ref{fig:corfrac}.
Third, photometric variability properties of the X-ray emission are analysed by using power density spectra (PDS) derived from the {\it Swift} and {\it XMM-Newton} light curves in the different phases (Figure \ref{fig:psresp_comparison}, see also Section \ref{sec:xraytiming}). The {\it Swift} light curves provide constraints on long timescales (1--25 days), while the {\it XMM-Newton} light curves probe short timescales (minutes -- hours).
In Figure \ref{fig:psresp_comparison} we compare the soft and hard spectral state power spectra. The best fit model parameters are constrained using Monte Carlo light curve simulations, which are fit using a bending power law model (see Section \ref{sec:xraytiming}).
Using these diagnostic tools, we investigate the properties of AT2018fyk in each of the phases defined in Figure \ref{fig:alphaox}, and find remarkable similarities to the properties of accreting stellar-mass black holes. The {$\alpha_{\rm ox}$\,} evolution with {f${_{\rm Edd, bol}}$\,} is strikingly similar to predictions from scaling a stellar-mass black hole outburst (\citealt{Sobolewska11a} and Fig. 3 of \citealt{Ruan19a}); we show a direct comparison with XRB outburst evolution across the hardness - intensity diagram (HID) in Figure \ref{fig:hid}, which further illustrates the similarities between these systems.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth, keepaspectratio]{HID_ann.png}
\caption{AT2018fyk in the hardness - intensity diagram; hardness is represented by {$\alpha_{\rm ox}$\,}. Note that we have reversed the horizontal axis to reproduce the classical HID. The grey bands show the typical turtle-head evolution of stellar-mass black holes in outburst. The TDE was discovered after peak light in the soft state; the observations missed the equivalent of the rising phase and hard to soft transition in X-ray binaries (segments marked with grey arrows).}
\label{fig:hid}
\end{figure}
\subsection{Soft-to-hard accretion state transition}
At the time of discovery, we find AT2018fyk at high {$\alpha_{\rm ox}$\,}, with an X-ray spectrum dominated by thermal emission with a temperature of $kT$ = 120--140 eV. The bolometric luminosity peaks around L$_{\rm bol}\approx$ 0.1L$_{\rm Edd}$ and is already declining, suggesting that we missed the rise to peak. The corona is present but weak, contributing only 5--25 \% of the X-ray flux and a few per cent of the bolometric luminosity. This is consistent with the power-law fraction observed in outbursting stellar-mass black holes in the soft state ($\sim 1-60 \%$; e.g. \citealt{Dunn10}), but inconsistent with the properties of soft state AGN where the corona is typically dominant (e.g. even for highly accreting narrow-line Seyfert 1 AGN the PL fraction is on average $70 \%$, see e.g \citealt{Vaughan99, Gliozzi20}; see also Section \ref{sec:agncomparison} for a more elaborate comparison to AGNs). The source is not detected at radio wavelengths (radio luminosity/X-ray luminosity = L$_{\rm radio}$ / L$_X \lesssim 10^{-6}$, \citealt{Wevers19b}), and we find no significant photometric variability on short timescales (frequencies $\gtrsim$10$^{-6}$ Hz; see Figure \ref{fig:psresp_comparison}). These properties are analogous with the soft states of stellar-mass black holes \citep{Homan05,Remillard06}.\\
When {f${_{\rm Edd, bol}}$\,} drops below $\sim$ 0.05, roughly consistent with the typical transition luminosities of stellar mass black holes (ranging from a few to $\sim$10 \%, \citealt{Maccarone03}), AT2018fyk transitions into an intermediate state (phase B). Both the SED (i.e., {$\alpha_{\rm ox}$\,}, top right panel of Figure \ref{fig:alphaox}) and the X-ray spectrum harden significantly over time (Figure \ref{fig:corfrac}). The observations show an increase in the X-ray power-law fraction (hardening), as well as an increasingly significant contribution of the power-law to the total bolometric luminosity. This is consistent with a scenario where the inner accretion disk becomes very cool or evaporates and the relative strength of the corona increases.
When the source reaches phase C, {$\alpha_{\rm ox}$\,} stabilises, and the power-law component dominates the X-ray spectrum (providing $>$50\% of the X-ray flux), and the X-ray flux varies dramatically. Rapid, large amplitude variability is present on both short (a few tens of minutes; bottom right panel of Figure \ref{fig:alphaox}) and long (days; top left panel of Figure \ref{fig:alphaox}) timescales. The onset of fast variability on roughly ten minute timescales is indicative of emission dominated by a compact emitting region, viz., an X-ray corona (see Section \ref{sec:xraytiming} and \ref{sec:timingdiscussion}). We note that the power-law contribution to the bolometric output in this state is similar to that seen in type 1 AGN with Eddington ratios $\sim$0.01 (\citealt{Lusso10}, see also Figure \ref{fig:plfrac_bol}).
The ratio of radio to X-ray luminosity is constrained to $\lesssim$ 4.5$\times$10$^{-7}$ in AT2018fyk. This rules out the appearance of a powerful radio jet, which is generally present in XRB hard states. Except for the absence of a radio jet, discussed in more detail in Section \ref{sec:radiodiscussion}, the rest of the properties are consistent with typical hard states found in XRBs and AGNs \citep{Homan05, Remillard06, Koerding06}. While the spectral hardening can, in principle, be due to variable absorption columns, the contemporaneous change in {$\alpha_{\rm ox}$\,} as well as the change in timing properties cannot be explained in this way. Instead, this indicates that the disk transitioned from a thermal dominated soft state to a radiatively inefficient, power-law/non-thermal dominated hard state.
X-ray flaring follows the soft-to-hard transition, when the thermal X-ray component has decreased in flux by $\sim$50 \%, while the power-law brightens by more than an order of magnitude. Rapid flares occur only in the X-ray band, indicating that they are confined to the hot, inner part of the accretion flow. While the X-ray spectral and photometric behaviour remain unchanged in the hard states (state C+D), the UV light curve in phase C is roughly constant whereas in phase D there are significant brightness variations (Figure \ref{fig:alphaox}, top left panel). The final phase (state D) of the UV light curve is remarkable, showing 2 cycles of {\it modulations} on $\sim$50 day timescales. There is no significant correlation between the UV and X-ray variability, which suggests physically distinct regions for these two phenomena.
\subsection{Changes in the timing properties: appearance of a band-limited high frequency component}
\label{sec:timingdiscussion}
The soft state power spectrum (Figure \ref{fig:psresp_comparison}) is very different to that seen in the hard state, as unlike the hard state power spectrum, it can be well fitted (goodness-of-fit = 0.7) with a single steep power-law at low frequencies, with no power-spectral break to flatter slope required. The parameter confidence contours were obtained by assuming a low frequency slope of $\alpha_{\rm LF} = -1$ to have the best chance of matching the low frequency variability seen in the light curve, but a steeper slope extending to low frequencies is still preferred.
\begin{figure*}
\centering
\includegraphics[width=\textwidth, keepaspectratio]{compare_psds_1e5scaling.pdf}
\caption{Power density spectral models of the soft (state A) and hard (state C+D) states. The median (solid curve) and 95\% confidence intervals (shaded regions) of the models are shown. The confidence interval includes the effects of the uncertainty in high frequency slope and bend frequency. Sampling effects significantly distort the PDS from the {\it Swift} data, so we assume two different low frequency slopes (0 and --1) covering the range seen in AGN and X-ray binaries. Most of the variability/power comes from low frequencies in the soft state, while a high frequency component ($\gtrsim$10$^{-5}$ Hz) appears in the hard state and dominates the overall variability (see Table \ref{tab:psresp}). This supports the idea that a compact corona dominates emission in the hard state. The black and grey datapoints show the PDS of a well-observed X-ray binary, Cyg X-1, scaled up to a 10$^7$ M$_{\rm BH}$ black hole (see text).}
\label{fig:psresp_models}
\end{figure*}
The soft state shows a smaller fractional rms than the hard state, but this is concentrated entirely at low frequencies ($<10^{-6}$~Hz) while in the hard state a significant fraction of variability occurs above $10^{-5}$ Hz, as is apparent from the much more variable {\it XMM-Newton} light curve.
In Figure \ref{fig:psresp_comparison} we compare the soft and hard state power spectra directly. The lower bend frequency in the $\alpha_{LF} = 0$ case makes it a closer match to the low frequency constraint on the soft state bend and a more conservative comparison with the soft state power spectrum. However, the overall shapes of the distorted power spectra should not depend too strongly on the low frequency index, since both fit the data well.
It is interesting to note that the low frequency power spectra for both soft and hard states are relatively consistent in amplitude, with a strong divergence in power at higher frequencies. Therefore, we can speculate that the main difference between the soft and hard state light curves is the addition of a band-limited high frequency component during the hard state. An additional broad high frequency component associated with the harder state is reminiscent of the transition between soft and hard states in X-ray binaries, where band-limited noise at higher frequencies is one of the first components associated with hard-intermediate states immediately prior to or after a transition to/from the soft state (e.g. see the power-spectral evolution presented in \citealt{Heil15}).
We measure the break frequency of the red-noise PDS in the hard state and find log($\nu_{bend}$)$\approx$--4.5. This is on the lower end of the distribution of typical break frequencies found in AGN, which typically occur around $\nu_{bend}$ $\sim$ a few $\times$ 10$^{-4}$ Hz (although a lower frequency break similar to AT2018fyk has also been observed, \citealt{Gonzalez12}).
Assuming we can use the break frequency to black hole mass scaling relation derived for AGN \citep{Mchardy06} here, we derive a black hole mass of log(M$_{\rm BH}$) = 6.9--7.2 $\pm$ 0.55 M$_{\odot}$. This is in agreement to within the uncertainties with the estimate from the M--$\sigma$ relation \citep{Wevers20}.
It is evident from the XMM2 light curve (Figure \ref{fig:alphaox}, bottom right panel) that the brightness sometimes changes by $\sim$50\% on timescales as short as 1250 seconds. Using light travel time arguments, this corresponds to a very compact emission region of $<$5 R$_g$ for the observed X-ray variability, assuming M$_{\rm BH}$ = 10$^{7.7\pm0.4}$ M$_{\odot}$ (with an upper limit of R$<$13 R$_g$, allowing for the uncertainty in the M$_{\rm BH}$ measurement). These results support a scenario where the fastest X-ray variability originates in a very compact X-ray corona.
For comparison with X-ray binary PDS, we used NICER observations of Cyg X-1 in a hard (ObsID 2636010101) and soft state (ObsID 1100320122), obtaining PDS in respectively the 2--10 keV band (representing the hard power-law component variability) and 0.6--1 keV band (representing the disk blackbody component variability). Although Cyg X-1 shows uncharacteristically high variability amplitude in its soft state compared to other black hole X-ray binaries ($\sim10$\% rms vs. $<1$\% rms, e.g. \citealt{Heil15}), the fact that it accretes from a stellar wind (and hence has a relatively small accretion disk) may make it a better comparison with AGN, and with TDEs in particular.
PDS characteristic time-scales scale inversely with black hole mass and linearly with accretion rate. Assuming the $\sim 10^{7}$~M$_{\odot}$ black hole mass we infer from the measured break time-scale of the hard state TDE power spectrum, and a black hole mass of $\simeq$20~$M_{\odot}$ for Cyg X-1 \citep{Miller-Jones21}, combined with the larger (factor 5) accretion rate of the TDE hard state compared to Cyg X-1, we scale the Cyg~X-1 PDS by a factor $10^{-5}$ in frequency for a better comparison with the soft and hard TDE PDS in Figure \ref{fig:psresp_models}. This figure illustrates the similarity between the X-ray binary and TDE PDS in the soft and hard states.
\subsection{Constraints on the presence of a jet}
\label{sec:radiodiscussion}
TDEs are, in principle, excellent opportunities to study jet formation, as the newly formed accretion flow settles into a steady state. Follow up campaigns at radio wavelengths are indeed revealing a growing population of radio-bright TDEs \citep{Alexander20}. It is interesting to note that several (though not all) radio-bright TDEs launch their non-relativistic jets/outflows already at early times, while they are in the soft state \citep{vanvelzen2016, Stein20}. This is at odds with the standard paradigm in XRBs, where the jet is quenched during disk dominated accretion states \citep{Corbel02, Russell11} and the transition of the accretion flow from the soft into the hard state is generally accompanied by the emergence of a compact radio jet \citep{Fender03}.
While our early observations of AT2018fyk are consistent with quenched jets, if the analogy holds we would expect the emergence of a radio counterpart following the soft-to-hard state transition.
However, our radio observations, covering the soft, hard and quiescent states, indicate that no bright radio jet is launched. Using the mean X-ray luminosity (3$\times$10$^{43}$ erg s$^{-1}$) in the hard state, the fundamental plane \citep{Koerdingplane} predicts a radio luminosity of $\sim$few $\times$ 10$^{39}$ erg s$^{-1}$. Our radio non detections constrain the luminosity to at least two orders of magnitude lower. We do note that a black hole mass that is lower by about 1 dex ($\sim$10$^{6.7}$ M$_{\odot}$) would bring the radio upper limits in line with the scatter observed in the fundamental plane.
While the presence of a correlation among radio luminosity, X-ray luminosity and black hole mass is well established for black holes accreting in a radiatively inefficient state \citep{Merloni03, Falcke04}, the situation for systems in the process of changing accretion state is less clear.
We briefly speculate which factors might be important to explain the apparent discrepancy with the XRB paradigm.
The emergence of a compact radio jet following the soft-to-hard state transition has been well documented in X-ray binaries \citep{Fender01}, although there is a delay of one to several weeks between the transition in the X-ray spectrum and the onset of a radio source \citep{Kalemci13, Vahdat18}. The cause of this delay is currently unclear, as is any potential scaling with black hole mass, because multiple conditions need to be satisfied to successfully launch a jet. \citep{Kalemci13} found that for compact jets to be produced in XRBs, the disk luminosity Eddington fraction needs to be below 10$^{-4}$. This condition is not fulfilled for AT2018fyk, where the bolometric disk emission Eddington ratio is $> 10^{-3}$; other factors may also play a role. Once a strong corona forms, magnetic fields need to be generated and transported, which is one plausible explanation for the delay seen in XRBs. Alternatively, the corona has to be large enough to efficiently collimate the outflow into a compact radio jet.
If launching a jet requires a significant magnetic field, this will take time to grow via the magnetorotational instability. In contrast to AT2018fyk, several TDEs in the soft state do launch powerful outflows \citep{vanvelzen2016, Alexander20}, suggesting that the jet launching mechanism may be driven by other factors \citep{Pasham18}, such as the availability of magnetic flux. Since the disrupted star is not expected to carry significant magnetic flux, the build-up through dynamos or magnetorotational instabilities may be too slow for most TDEs, unless a fossil magnetic flux reservoir is present \citep{Kelley14}.
We note that the host galaxy spectra of several radio-bright TDEs show possible evidence for nuclear activity. The narrow emission lines, when plotted in a diagnostic emission line diagram, fall in the regions indicative of a low ionisation narrow emission line region (LINER), AGN or composite AGN/star formation \citep{Nikolajuk13, French17}. While LINER-like emission can also be excited by other emission mechanisms, if instead this indicates recent or on-going low level AGN activity, the pre-existing jet cone and/or relic magnetic fields may facilitate the renewed launching of an outflow. AT2018fyk, on the other hand, does not show indications of recent AGN or star forming activity, so the absence of relic magnetic fields may also help explain the lack of a radio jet.
\subsection{Hard state to quiescence transition}
The strong UV and X-ray variability ends with a transition into a different regime (the quiescent state, E), evidenced by a factor $\sim$15 drop in the UV luminosity and a factor $>$5000 in X-rays. As a result, the SED significantly softens, i.e., {$\alpha_{\rm ox}$\,} increases, and the UV emission becomes relatively more important when compared to the hard state.
These dramatic changes constitute the {\it second} accretion state transition. Analogous, although much less dramatic, behaviour is also seen in stellar-mass black holes, which move towards a baseline, quiescent state at very low accretion rates. For stellar-mass black holes this state is defined empirically by L$_X <$ 10$^{-5}$ L$_{\rm Edd}$ \citep{Plotkin13}, when the anti-correlation between the photon index of the X-ray spectrum and L$_X$ plateaus or inverts \citep{Homan13, Yang15}. Qualitatively similar behaviour is also observed in (changing-look) AGN \citep{Yang15, Ruan19a}.
The transition towards quiescence in stellar-mass black holes and AGN is more gradual and much less dramatic than observed for AT2018fyk. This warrants a discussion of alternative scenarios (see Section \ref{sec:thermalinstability}). For now, we note an important difference between X-ray binaries and AGN, and TDEs: the fallback rate of material at the outer disk edge. In stellar-mass black holes and AGN there is a steady supply of material, whereas in TDEs the supply is expected to decrease rapidly with time (following a t$^{-5/3}$ power-law behaviour \citealt{Rees1988}). This may trigger much more dramatic changes in the structural properties of the accretion flow in TDEs.
We constrain L$_X < 2.6 \times 10^{-6}$ L$_{\rm Edd}$ for AT2018fyk, indicating that the accretion flow has indeed reached the quiescent state and the mass supply rate from the stellar debris has dropped to very low levels.
The increase in {$\alpha_{\rm ox}$\,} (softening of emission) at low Eddington ratio has also been observed in stellar-mass black hole outbursts \citep{Homan13,Plotkin17},
as well as in a sample of changing-look AGN \citep{Ruan19a}. Such a softening cannot be explained by the progressive disk truncation/evaporation thought to be responsible for the hardening of {$\alpha_{\rm ox}$\,} from the soft to the hard state \citep{Sobolewska11b}. Instead, it suggests that the X-ray component is dominated by an entirely different emission mechanism, although observations are sparse and this state is poorly understood \citep{Plotkin15}.
Remarkably, the {$\alpha_{\rm ox}$\,} value in the quiescent state (state E) is at least as high as that observed in the soft state (state A), indicating that the X-ray corona is extremely weak or possibly non existent.
\subsection{Alternative explanation for the second state transition}
\label{sec:thermalinstability}
Several peculiar features in the hard state (specifically, state D), including those discussed above, warrant a closer look to potential alternatives for the standard accretion state transition scenario for the total collapse of the system.
Thermal and viscous instabilities are an inherent property of 1D accretion disk theory at sub-Eddington accretion rates \citep{Lightman74,Shakura76}.
In particular, \citet{Shen14} predict that a thermal instability occurs when the fallback rate drops below the advection dominated regime. When this happens, the disk rapidly transitions into a gas pressure dominated, radiatively cooled state. Furthermore, in the presence of a sufficiently high rate of material falling back onto the outer disk, they predict limit cycle oscillations between this gas pressure dominated and the initial radiation pressure dominated states. More specifically, changes to both the disk radius as well as thermal state in the presence of fallback are expected, where the addition of low angular momentum material leads to a shrinking of the disk radius. This increases the disk surface density, and can cause the disk to jump back to the advective state for brief periods of time.
The observed behaviour in AT2018fyk shows some similarities to theoretical predictions by \citet{Shen14}. First, the predicted timescale for the instabilities to manifest is $\sim$ 1 year, as observed. Second, the blackbody radius inferred in Figure \ref{fig:bbrad} shows an overall decreasing trend over time, as expected. Some oscillations are observed after $\sim$300 days, marking the onset of the UV photometric variability. \citet{Shen14} predict that the thermal cycles will end when the mass fallback rate drops below a critical threshold, which typically occurs after a few (2--3) cycles. AT2018fyk goes through 2 oscillatory cycles with a typical timescale of $\sim$50 days before the UV component (the accretion disk) eventually decreases by a large factor. This scenario provides a self-consistent explanation of the long term behaviour, the UV variability, the blackbody radius evolution, the subsequent state transition and the absence of a radio jet, providing tantalising evidence that thermal instabilities might occur in nature.
\subsection{The unusually rapid evolution of the accretion flow}
AT2018fyk is only the fourth UV/optical discovered TDE that is persistently X-ray bright for $\sim$500 days, the others being ASASSN--14li \citep{Holoien201614li}, ASASSN--15oi \citep{Gezari17} and AT2019azh \citep{Hinkle20}. It shows evidence for a much more rapid disk formation and subsequent evolution when compared to typical TDEs, including the early emergence of Fe\,\textsc{ii} emission lines \citep{Wevers19b} and a pronounced second maximum early in the light curve \citep{Dong16, Leloudas16}. The sudden drop in UV luminosity also occurs relatively early when compared to other TDEs, with \citet{vanvelzen2019} finding only 1 TDE (SDSS TDE1) that may have experienced a similar drop in UV brightness 5--10 years after disruption. Similarly, the sharp decrease of the X-ray emission occurs much more rapidly after disruption than observed in other TDEs \citep{Jonker20}.
These seemingly peculiar properties suggest that the post disruption evolution proceeds more rapidly than in the bulk of the TDE population. Such a rapid evolution is expected if the debris streams can violently self-collide upon their return to pericenter, efficiently removing orbital angular momentum and energy. It is, however, unclear (both theoretically and observationally) whether this circularisation phase can indeed happen efficiently in the majority of TDEs \citep{Ulmer1999, Shiokawa2015, vanvelzen2019}.
The dominant mechanism to achieve strong shocks between streams is relativistic apsidal precession, which becomes increasingly important as the black hole mass increases, or alternatively when the star penetrates deeply into the SMBH potential well (high $\beta$ = R$_{p}$/R$_{t}$).
Both AT2018fyk and SDSS TDE1 have comparatively high inferred black hole masses in excess of 10$^7$ M$_{\odot}$ \citep{Wevers17, Wevers20} and are UV faint at late times. At the inferred black hole mass of 1--5$\times$10$^7$ M$_{\odot}$, relativistic effects are important for all stellar encounters \cite{Stone19}, providing a natural explanation for the rapid formation and evolution of the accretion flow, which shines brightly in soft X-rays and at UV wavelengths.
The rapid subsequent evolution is a natural consequence of a high black hole mass. The Eddington limit in AT2018fyk is $\sim$10-50 times higher than in other well-observed TDEs, which typically have correspondingly lower black hole masses \citep{Wevers17, Wevers19a, Mockler19}. These results indicate that black hole mass could be an important factor setting the timescales of the post-disruption debris evolution.
An alternative explanation for rapid evolution could be a particularly low reservoir of debris to accrete from, i.e., a comparatively low mass star. As a result, the system will run out of material more quickly. However, stars with a mass $\lesssim$ 0.4 M$_{\odot}$ would be swallowed whole by a non-rotating SMBH in AT2018fyk (e.g. \citealt{Stone19}), while a dimensionless spin of $a > 0.7$ is required to produce an observable disruption for a 0.1 M$_{\odot}$ star (e.g. \citealt{Leloudas16}). Therefore, given the large mass of the black hole in AT2018fyk, it is unlikely that the disrupted stellar mass is significantly smaller when compared to typical TDEs \citep{Mockler19, vanvelzen2019, Wen20}.
A conservative lower limit for the accreted stellar mass can be obtained by making an assumption for the conversion of accretion power to radiation. We calculate the bolometric energy output by piece wise integration of the entire light curve to find a total radiated energy of E$_{\rm rad}$ = 9.4$ \pm$ 0.2 $\times$ 10$^{51}$ ergs. Assuming a standard radiative efficiency of $\eta = 0.1$ and taking into account that only half of the stellar debris is available for accretion, the corresponding stellar mass is M$_{\star}\approx$0.1 M$_{\odot}$. We consider this calculation a lower limit to the true stellar mass. First, because of the assumption of radiatively efficient accretion throughout the evolution, which may not be valid when the source transitions into the hard state, where we may have $\eta \ll 0.1$. This could lead to an increase in the estimate of M$_{\star}$ of a factor 2--3, accounting for the observed duration of the soft and hard states respectively. Second, there are no constraints on the potentially large bolometric correction (up to a factor 10 or more, \citet{vanvelzen2016b}) due to dust absorption and re-emission on sub-parsec scales, which is the dominant contributor in the uncertainty for the radiated energy.
\subsection{Comparison of the corona properties with AGN}
\label{sec:agncomparison}
It is still unclear whether the soft excess spectral component that is ubiquitously observed in AGN and TDEs is produced at the inner edge of the compact accretion disk, as an additional low temperature Comptonisation component, or elsewhere. This complicates the analogy and interpretation of the power-law fraction of emission as a proxy for accretion state, as is done for X-ray binaries.
To mitigate this uncertainty, an alternative approach that is directly analogous to the one employed for stellar-mass black hole systems is to consider not only the X-ray emission, but also include the UV emission where the accretion disk is expected to dominate for SMBHs. In Figure \ref{fig:corfrac} (right panel) we show the ratio of the 2--10 keV power-law emission to the total, bolometrically corrected (optical+UV+X-ray) emission (see Section \ref{sec:alphaox} for more details on how this correction is applied).
We illustrate how AT2018fyk compares to AGN samples in Figure \ref{fig:plfrac_bol}. The values in the hard state, at low Eddington ratios ($\sim$10$^{-2}$) cover a very similar parameter space to type 1 AGN (grey triangles, \citealt{Lusso10}) and other AGN samples including many NLS1 sources (black stars, \citealt{Vasudevan07}). However, in the soft state the power-law emission in AT2018fyk is on the low side compared to AGN with similar Eddington ratios. This could be resolved by taking into account the uncertainties on our black hole mass estimate $\sim$0.5 dex, and may indicate that the lower end of the mass range is favoured. Alternatively, we do not include the full energy ranges used to compute the bolometric corrections in the comparison samples of \citet{Lusso10} and \citet{Vasudevan07}; in particular, we don't have observational constraints for the 10--200 keV band. However, given the steep power-law indices we find, this should result in only small corrections, suggesting that the difference is physical in nature.
We recall that the power-law fraction in the soft state of AT2018fyk is comparable to the values seen in stellar-mass black holes, but much lower than in soft state AGN. The size of the accretion flow in TDEs (normalised to the black hole gravitational radius) is much more compact than in AGN. As a result, irradiation and/or radiation pressure are much higher in TDEs (a situation much more like X-ray binaries than in AGNs), which may prevent the formation of an efficient corona shortly after disruption, when the mass accretion rates are highest. Typical soft state AGN have been actively accreting material for long periods of time, and hence the corona has had time to build up and stabilise, whereas in the newly formed accretion flow following AT2018fyk we witnessed the formation of the corona in near real-time, strengthening as the source moves into the hard state. In the future, a larger TDE sample will help better understand the formation of the corona, and the nature (and role) of the poorly understood soft excess emission component seen in both TDEs and AGN.
\begin{figure}
\includegraphics[width=\linewidth, keepaspectratio]{fedd_plfrac_agncomp.pdf}
\caption{Comparison of the evolution of the corona with AGN. Colour-coded filled circles show the evolution of the power law fraction to the bolometric emission as a function of the Eddington ratio. Grey and black triangles and stars represent AGN samples.}
\label{fig:plfrac_bol}
\end{figure}
\section{Summary}
\label{sec:summary}
We have presented a very rich, multi-wavelength set of observations of the tidal disruption event AT2018fyk, from peak light (at an Eddington ratio of $\sim$0.1) into quiescence (Eddington ratio $<$ 10$^{-3.4}$).
We find that around peak light, the source exhibits several of the hallmark features of highly accreting supermassive black holes and stellar-mass black holes in the soft state, including a thermal-dominated X-ray spectrum, a soft (high {$\alpha_{\rm ox}$\,}) SED, no bright radio emission and exclusively low-frequency ($<$10$^{-5}$ Hz) X-ray variability. After a transition phase, dramatic changes to these properties are observed. The power-law component becomes dominant in the X-ray spectrum, {$\alpha_{\rm ox}$\,} hardens significantly and a band-limited high-frequency ($\sim$10$^{-3}$ Hz) time-variable component appears in the X-ray band. These changes are consistent with an accretion state transition from a soft to a hard state, and a strengthening of the X-ray corona. No bright radio jet accompanies this transition, marking an apparent deviation of the stellar-mass black hole accretion state paradigm. We discuss several plausible explanations for the absence of a radio jet around TDEs, and conclude that TDEs showing accretion state transitions provide a promising avenue to better understand the necessary conditions for successful jet formation around SMBHs with future observations.
At late times, a second dramatic transition occurs, with the X-ray luminosity dropping by a factor $>$5000 and the UV luminosity decreasing by a factor $\sim$15. The SED significantly hardens during this transition, similar to the behaviour seen in changing-look AGN and other TDEs. This constitutes the second state transition, after which the source reaches the quiescent state at a low Eddington ratio $<$0.0004.
These transitions between the soft, hard and quiescent states, as well as the main drivers of disk (in)stabilities in accreting black holes remain poorly understood, especially so for SMBHs. This work shows that TDEs can provide a unique new opportunity to observationally constrain accretion state properties and transitions in individual SMBHs. In comparison to changing-look AGN \citep{Gezari17clagn, Ruan19a, Trakhtenbrot19, Frederick19} TDEs evolve faster, facilitating coordinated observational studies of the full evolution of accretion cycles and transitions. Furthermore, TDEs are readily discovered at increasing rates in wide-field optical and X-ray photometric surveys without the need for archival observations such as spectroscopy (to determine whether the source has changed state). We have demonstrated that a detailed characterisation of both the soft and hard states, as well as transition timescales and changes in the interplay between the different emitting regions, are possible. In summary, TDEs enable direct studies of the apparent scale invariance of accretion processes across seven orders of magnitude in black hole mass. In the future, statistical samples of TDEs with UV and radio monitoring observations, as well as soft and hard X-ray observations covering both short (stare) and long (monitoring) timescales will allow in-depth comparisons between stellar-mass and supermassive black hole accretion flow properties.
\acknowledgments
We thank D.J. Walton, W. Alston, A.C. Fabian, C. Reynolds, E. Coughlin, K. Hayasaki and M.I. Saladino for discussions, and the anonymous referee for constructive comments and suggestions.
We are grateful to the {\it Swift}, {\it XMM-Newton}, {\it NICER} and {\it Chandra} operation teams and PIs for awarding and scheduling the ToO/DDT observations. TW was funded in part by European Research Council grant 320360 and by European Commission grant 730980. SvV was supported by the James Arthur Postdoctoral Fellowship. This research was partially supported by the Australian Government through the Australian Research Council's Discovery Projects funding scheme (project DP200102471). Raw optical/UV/X-ray observations are available in the NASA/{Swift} archive (\url{http://heasarc.nasa.gov/docs/swift/archive}, Target Names: AT2018fyk, ASASSN-18UL); the XMM-Newton Science Archive (\url{http://nxsa.esac.esa.int}, obsIDs: 0831790201, 0853980201, 0854591401); and the Chandra data archive (\url{https://cxc.harvard.edu/cda/}, obsID: 23289). {\it NICER} data is publicly available through the HEASARC: \url{https://heasarc.gsfc.nasa.gov/cgi-bin/W3Browse/w3browse.pl}.
\vspace{5mm}
\facilities{XMM-Newton, Swift(XRT and UVOT), NICER, CXO, Magellan, ATCA}
\software{astropy \citep{astropy},
Heasoft \citep{heasoft},
CIAO \citep{ciao},
gPHOTON \citep{Million16},
CASA \citep{McMullin2007}
}
\clearpage
|
\section{Introduction}
The estimation of physical quantities is an important task in many branches of science and technology. In recent decades, the field of quantum metrology and sensing has emerged, in which the objective is to exploit quantum coherence or entanglement to give enhanced sensitivity in such parameter estimation tasks \cite{Gio-11, Deg-17, Pez-18}. Proposed applications include magnetometry \cite{Bud-07, Tay-08, Tan-15}, electrometry \cite{Dol-11, Fac-16}, quantum clocks \cite{Lud-15} and, perhaps most famously, gravitational wave detection \cite{Cav-81, Aas-13}.
One of the most widely used approaches to quantum sensing is Ramsey interferometry, or one of its variants. After preparing the probe system in some quantum superposition state, it is allowed to interact with the parameter-of-interest, followed by a measurement of the probe to extract information about the parameter \cite{Yur-86b, Lee-02, Gio-11, Deg-17, Pez-18}. Precise sensing then usually relies on maintaining quantum coherence in the probe system for long times. However, in many realistic sensing devices, other unwanted interactions are expected to degrade the achievable sensitivity by limiting the useful coherence time. Even if the probe is completely isolated from any external environment, internal interactions between the consituent particles of the probe can lead to decoherence and thermalisation. Moreover, a high density of probe particles will likely result in stronger interactions, and more rapid decoherence and thermalisation. Several schemes to overcome this limitation have been proposed \cite{Doo-18a, Cho-17, Rag-18}, or implemented experimentally \cite{Zho-20}.
Recently, a new mechanism was discovered in which thermalisation can be slowed down, or even completely avoided, despite strong interactions in a non-integrable many-body system. This mechanism -- dubbed quantum many-body scarring -- was shown to be responsible for long-lived oscillations in an experiment on a chain of interacting Rydberg atoms \cite{Ber-17, Tur-18a}. Since the long-lived oscillations are associated with long coherence times, this is suggestive of a possible advantage in quantum sensing \cite{Ser-20}. Despite intensive work on scars and their properties \cite{Mou-18a,Mou-18b,Ok-19,Bul-19,Ho-19,Mar-20,Shi-20,Mou-20,Mic-20,Iad-20,Bul-20,Kun-20,McC-20,Ban-20, Doo-20b}, this possibility has not yet been explored.
In this paper, we examine the potential for robust quantum sensing via quantum scarring. We begin in section \ref{sec:general} with a brief discussion of quantum sensing, and its connection with many-body scars. Then, through two examples, we demonstrate the connection more concretely. First, in section \ref{sec:spin_1}, a spin-1 model with a Dzyaloshinskii-Moriya interaction (DMI). We find that despite the non-integrability of the model, robust quantum sensing is possible. This is associated with a diverging coherence time caused by of a set of scars with perfectly harmonic energy gaps. Then, in section \ref{sec:MFIM}, we consider a mixed-field Ising model. Usually, strong Ising interactions between the spins are expected to lead to fast decoherence and thermalisation. Counterintuitively, in this example the stronger couplings can extend the coherence time and enhance the quantum sensing. We show that this is due to the emergence of a set of quantum many-body scars in the ``PXP'' limit of the mixed-field Ising model. Finally, in section \ref{sec:noise} we briefly discuss some non-ideal perturbations, and the possibility of suppressing their effect on sensing by periodic driving.
\section{Quantum sensing, ETH, and many-body scars}\label{sec:general}
Consider an $N$-particle probe system, whose purpose is to estimate a parameter $\omega$ that appears in its Hamiltonian $\hat{H} = \omega \sum_{n=0}^{N-1}\hat{h}_n + \hat{H}_\text{int}$. Here, $\hat{h}_n$ is a local operator for the $n$'th particle and $\hat{H}_\text{int}$ generates interactions between the particles. Typically, the estimation scheme involves initialising the probe in some easily prepared state $\ket{\psi(t=0)}$, and extracting information about the parameter $\omega$ from a measurement of the time-evolved state $\ket{\psi(t)} = e^{-it\hat{H}}\ket{\psi(0)}$. For small probe systems (e.g., a single qubit) arbitrary measurements of the final state may be possible. For larger systems, however, the measurement may be restricted to observables $\hat{\mathcal{O}} \in \mathcal{M}$ from some set of experimentally accessible measurements $\mathcal{M}$. This set will depend on the details of the physical implementation, but is often a subset of local operators. To accumulate statistics about the parameter-of-interest, the prepare-evolve-measure sequence is repeated during a total available time $T$, giving $T/t$ independent repetitions of the measurement (where the time needed for state preparation and readout is assumed to be negligible \cite{Doo-16b, Hay-18}). Employing method-of-moments estimation, the final error can be calculated by the propagation-of-error formula \cite{Win-94, Pez-18}: \begin{equation} \delta\omega = \min_{\hat{\mathcal{O}} \in \mathcal{M}} \frac{ \Delta \hat{\mathcal{O}}(t)}{| \partial_{\omega} \langle \hat{\mathcal{O}} (t) \rangle | \sqrt{T/t}} , \label{eq:error_formula} \end{equation} where the numerator is the uncertainty $\Delta\hat{\mathcal{O}} \equiv \sqrt{\langle \hat{\mathcal{O}}^2\rangle - \langle\hat{\mathcal{O}} \rangle^2}$ of the measured observable $\langle\hat{\mathcal{O}}\rangle$ and the factor $|\partial_\omega \langle\hat{\mathcal{O}}\rangle|$ in the denominator quantifies its response to small changes in the parameter $\omega$.
In the setup described above, the dynamics are unitary and the time-evolved state $\ket{\psi(t)}$ is always pure. However, for a non-integrable many-body system the interactions $\hat{H}_\text{int}$ between particles can lead to decoherence and thermalisation with respect to the expectation values $\langle\hat{\mathcal{O}}\rangle$. This is because the evolution generates entanglement in the system that makes information about the initial conditions inaccessible to the experimental observables $\hat{\mathcal{O}} \in \mathcal{M}$.
Instead of waiting for the probe to thermalise, it is usually better to measure the system before it thermalises, at some optimal sensing time $t = t_*$ giving the optimised error $\delta\omega_* = \min_t \delta\omega$. For separable initial states $\ket{\psi(0)}$ the optimised error typically has the form $\delta\omega_* \propto 1/\sqrt{N t_* T}$, with fast decoherence and thermalisation corresponding to a short $t_*$, and poor quantum sensor performance \cite{Deg-17}. To estimate $\omega$ precisely, one should try to engineer a probe system that has both a large coherence time $t_*$ and a large number of particles $N$. One approach is to design the probe Hamiltonian so that the interacting part is suppressed as much as possible, $\hat{H}_\text{int} \approx 0$. This gives a long coherence time $t_*$, but usually means that the particle density must be low, to suppress the interactions. For high density quantum sensing, with long coherence times $t_*$, we should look for mechanisms for avoiding decoherence and thermalisation, even in the presence of strong interactions.
The process of thermalisation in closed quantum systems is often framed in terms of the eigenstate thermalisation hypothesis (ETH) \cite{Deu-91, Sre-94, Rig-08}. Let $\hat{H} = \sum_j E_j \ket{E_j}\bra{E_j}$ be the spectral decomposition of the Hamiltonian. An eigenstate $\ket{E_j}$ is said to be thermal if, for all $\hat{\mathcal{O}} \in \mathcal{M}$, we have $\bra{E_j}\hat{\mathcal{O}}\ket{E_j} \approx \mbox{Tr}[\hat{\mathcal{O}}\hat{\rho}_\text{th}]$, where $\hat{\rho}_\text{th}$ is the thermal state at the temperature corresponding to the energy $E_j$. One can show that if \emph{all} eigenstates around a given finite energy density are thermal, the observables $\hat{\mathcal{O}}\in\mathcal{M}$ will thermalise for any initial state at that energy density \cite{DAl-16}. Conversely, the system can fail to thermalise if there are some non-thermal eigenstates that have a large overlap with the initial state \cite{Bir-10}. Recently, such ETH-violating systems were discovered where most, but not all, Hamiltonian eigenstates are thermal \cite{Shi-17}. The non-thermal eigenstates were dubbed quantum many-body scars (QMBS), and were found to be responsible for long-lived coherence in an experiment with a chain of Rydberg atoms \cite{Ber-17, Tur-18a, Ser-20}.
Based on the foregoing discussion, some necessary conditions for robust quantum sensing via many-body scars are:
\begin{enumerate}
\item A Hamiltonian with a set of QMBS.
\item An easily prepared initial state, with a large component in the QMBS subspace.
\item Dynamics in the QMBS subspace that are sensitive to the Hamiltonian parameter to be estimated.
\item An experimentally accessible observable $\hat{\mathcal{O}} \in \mathcal{M}$ that can extract the parameter information from the time-evolved state.
\end{enumerate}
In sections \ref{sec:spin_1} and \ref{sec:MFIM} we present two example models in which these criteria are satisfied, giving robust quantum sensing despite strong interactions in the many-body probe system. Both examples suggest that in the ideal case (perfect scars with harmonic energy gaps, and full overlap with the initial state) the optimal sensing time $t_*$ diverges.
\section{Example: spin-1 DMI model}\label{sec:spin_1}
\subsection{Quantum sensing}\label{sec:spin_1_sensing}
As our first example, we consider a sensing scheme where the goal is the estimation of a magnetic field $\omega$ acting on a system of $N$ interacting spin-1 particles via the Hamiltonian: \begin{equation} \hat{H}_0 = \frac{\omega}{2} \sum_{n=0}^{N-1} \hat{S}_n^z , \end{equation} where $\hat{S}^z = \sum_{m\in\{-1,0,1\}}m\ket{m}\bra{m}$. However, the particles may also interact with each other via the Hamiltonian \begin{equation} \hat{H}_\text{int} = \sum_{n,n'} \lambda_{n,n'} (e^{i\phi} \hat{S}_n^+ \hat{S}_{n'}^{-} + \text{h.c.}) , \label{eq:H_int} \end{equation} where $\hat{S}^\pm = \sum_{m\in\{-1,0,1\}} (1 - \delta_{m, \pm 1}) \ket{m \pm 1}\bra{m}$ are the spin raising and lowering operators, giving a total Hamiltonian $\hat{H} = \hat{H}_0 + \hat{H}_\text{int}$. The spin-spin coupling parameters $\lambda_{n,n'}$ are assumed to be real but are otherwise arbitrary and can, for example, be any real function of the positions of the interacting spins in some $d$-dimensional space. The interaction Hamiltonian can be rewritten as $\hat{H}_\text{int} = \sum_{n,n'}\lambda_{n,n'} [ \cos\phi (\hat{S}_n^x\hat{S}_{n'}^x + \hat{S}_n^y\hat{S}_{n'}^y) + \sin\phi (\hat{S}_n^x\hat{S}_{n'}^y - \hat{S}_n^y\hat{S}_{n'}^x)]$, showing that the phase $\phi$ rotates between an XX-interaction for $\phi \in \{ 0, \pm\pi \}$, and a Dzyaloshinskii-Moriya interaction (DMI) for $\phi = \pm\pi/2$.
After preparing the spins in the initial product state: \begin{eqnarray} \ket{\psi(t=0)} = \ket{\boldsymbol{+}} \equiv \bigotimes_{n=0}^{N-1} \frac{\ket{m_n = +1} + \ket{m_n = -1}}{\sqrt{2}} , \label{eq:psi_plus} \end{eqnarray} we suppose that the system is allowed to evolve by the Hamiltonian $\hat{H}$ for a sensing time $t$, followed by a measurement of the local observable $\hat{\mathcal{O}}_\theta = e^{-i\theta}\sum_n (\hat{S}_n^+)^2 + \text{h.c.}$, where $\theta$ can be tuned for the optimal sensing performance. This prepare-evolve-measure sequence is repeated during a total available time $T$, giving $T/t$ independent repetitions of the measurement. The error in the estimate of $\omega$ is calculated by the propagation-of-error formula given in Eq. \ref{eq:error_formula}. We note that the initial product state in Eq. \ref{eq:psi_plus} is usually relatively easy to prepare in practice. For example, after cooling to the ground state $\otimes_n \ket{m_n = -1}$ of $\hat{H}$ in a strong magnetic field, a collective rotation of all spins can generate the desired state.
\begin{figure}
\includegraphics[width=\columnwidth]{level_spacing}
\caption{Here, we demonstrate the non-integrability of $\hat{H}$ for all values of $\phi$, in a $d=1$ dimensional example with $\lambda_{n,n'} = \lambda/(n-n')^2$ and periodic boundary conditions. The Wigner-Dyson distribution $P(s)$ for the energy spacings $s_i = E_{i+1} - E_i$, and the associated $r$-value of $\langle r \rangle \approx 0.53$ are signatures of non-integrability. All quantities are calculated in the $k = 0$ momentum, and $\sum_n \hat{S}_n^z = 1$ magnetization symmetry sectors. For $\phi \in \{ 0, \pm\pi\}$ we must also restrict to the even reflection parity sector. [Other parameters: $N = 14$, $\omega = 1$, $\lambda = 0.5$.]}
\label{fig:level_statistics}
\end{figure}
As a benchmark, we first calculate the error in the case when there are no interactions between the spins, i.e., when $\lambda_{n,n'} = 0$ for all $n$, $n'$. Then, the time-evolved state is: \begin{eqnarray} \ket{\psi(t)} = \bigotimes_{n=0}^{N-1}\frac{e^{-it\omega/2}\ket{m_n=+1} + e^{it\omega/2}\ket{m_n=-1}}{\sqrt{2}} . \label{eq:psi_t_non_int} \end{eqnarray} We see that the dynamics are periodic, with period $\tau = 2\pi/\omega$, and the spins remain in a product state throughout the evolution. The error, easily calculated by the formula in Eq. \ref{eq:error_formula}, is then equal to the standard quantum limit, given by $\delta\omega = \delta\omega_\text{SQL} \equiv 1/\sqrt{NtT}$.
However, the non-interacting model is a very special case, for which the Hamiltonian $\hat{H}$ is integrable. Generally, for $\hat{H}_\text{int} \neq 0$, $\hat{H}$ is non-integrable. The interactions between the spins might therefore be expected to result in a degradation of the sensing performance. Indeed, when interactions are present the error typically does not decrease monotonically with the sensing time $t$. Rather, it decreases to a minimum value $\delta\omega_* \equiv \min_t \delta\omega$ at an optimal sensing time $t_*$.
\begin{figure}
\includegraphics[width=\columnwidth]{fig1b.pdf}
\caption{(a) The optimised error $\delta\omega_*$, calculated numerically for the $d=1$ example with $\lambda_{n.n'} = \lambda/(n-n')^2$, periodic boundary conditions, and $N=10$. The plots in (b) and (c) are cross sections of (a) (although for $N=12$). (d), (e) show the corresponding optimal sensing times $t_*$. The numerical data (the circles) are closely matched by the equations $\delta\omega_* = 1.09/\sqrt{Nt_* T}$ and $t_* = 0.53/|\lambda \cos\phi|$ (the solid lines). [All plotted for $\omega = 1$, although the error $\delta\omega$ is independent of $\omega$; see Appendix \ref{app:omega_independence}.]}
\label{fig:opt_estimation_error}
\end{figure}
We illustrate these points for an example in $d=1$ dimensions, with an interaction $\lambda_{n,n'} = \lambda/(n-n')^2$ and assuming periodic boundary conditions. We show in Fig. \ref{fig:level_statistics}, by calculating energy level spacing statistics, that this Hamiltonian is non-integrable for any choice of the phase $\phi$. In Fig. \ref{fig:opt_estimation_error} we plot the optimised error $\delta\omega_*$ and the optimal sensing time $t_*$, as a function of the overall coupling strength $\lambda$ and the phase $\phi$. We see that the equations: \begin{equation} \delta\omega_* = \frac{1.09}{\sqrt{Nt_{*}T}} , \quad t_* = \frac{0.53}{|\lambda\cos\phi|} , \label{eq:error_fit} \end{equation} fit the numerical data very well. It is clear that the optimal sensing time is long, and the error is low, when the interaction strength $\lambda$ is small. This is not surprising since it is close to the case of non-interacting spins precessing in the field $\omega$. However, the other notable feature of Fig. \ref{fig:opt_estimation_error} and Eq. \ref{eq:error_fit} is the diverging optimal sensing time $t_* \to \infty$ when $\phi = \pm \pi/2$, even for strong interactions $\lambda$. This may be surprising, considering the non-integrability of the Hamiltonian. We now show that the diverging $t_*$, and the associated low error for $\phi \approx \pm \pi/2$, is due to quantum many-body scarring, and is a feature of the model not only for our $d=1$ dimensional example, but for any choice of the interaction strengths $\lambda_{n,n'}$, in any spatial dimension.
\subsection{Quantum many-body scars in the spin-1 DMI model}\label{sec:spin_1_QMBS}
Before proceeding, it is convenient to introduce the spin-1/2 operators that are obtained by restricting to the $\ket{m_n = \pm 1}$ local basis states of each spin-1 particle. These operators are $\hat{\sigma}_n^\pm \equiv (\hat{S}_n^\pm)^2 = \ket{m_n = \pm 1}\bra{m_n = \mp 1}$, $\hat{\sigma}_n^z = [\hat{\sigma}_n^+, \hat{\sigma}_n^-] = \hat{S}_n^z$, with the associated spin-1/2 collective operators $\hat{J}^\pm \equiv \sum_{n=0}^{N-1}\hat{\sigma}_n^\pm$ and $\hat{J}^z \equiv \frac{1}{2}[\hat{J}^+, \hat{J}^-] = \frac{1}{2}\sum_{n=0}^{N-1}\hat{\sigma}_n^z$. The symmetric Dicke states $\{\ket{\Psi(s)}\}_{s=0}^N$ are defined as simultaneous eigenstates of $\hat{J}^z$ and $\hat{J}^2 \equiv (\hat{J}^x)^2 + (\hat{J}^y)^2 + (\hat{J}^z)^2$, and can be written as \cite{Dic-54}: \begin{equation} \ket{\Psi(s)} \equiv \mathcal{N}_s (\hat{J}^+)^s \ket{\Downarrow} , \quad s \in \{ 0,1, \hdots, N \} , \label{eq:dicke_states} \end{equation} where $\mathcal{N}_s = \frac{1}{s!}\binom{N}{s}^{-1/2}$ is a normalisation factor and $\ket{\Downarrow} \equiv \bigotimes_{n=0}^{N-1}\ket{m_n = -1}$.
It was recently shown \cite{Mar-20b} that the symmetric Dicke states in Eq. \ref{eq:dicke_states} are scar states of the Hamiltonian $\hat{H}$ when $\phi = \pm\pi/2$. In other words, they are ETH-violating eigenstates of $\hat{H}(\phi=\pm\pi/2)$, with a sub-volume law growth of entanglement entropy. We already know that they are eigenstates of the non-interacting part of the Hamiltonian $\hat{H}_0\ket{\Psi(s)} = \omega(s-N/2)\ket{\Psi(s)}$, since this is one of the defining properties of Dicke states. Writing $\hat{h}_{n,n'}(\phi) \equiv e^{i\phi}\hat{S}_n^+\hat{S}_{n'}^- + e^{-i\phi}\hat{S}_n^-\hat{S}_{n'}^+$, in Appendix \ref{app:Dicke_scars} we also show that $\hat{h}_{n,n'}(\pm\pi/2) \ket{\Psi(s)} = 0$ for all $n$, $n'$, implying that $\hat{H}_\text{int}\ket{\Psi(s)} = 0$. Therefore, the Dicke states are eigenstates of the full Hamiltonian $\hat{H} = \hat{H}_0 + \hat{H}_\text{int}$, with the eigenvalue equation: \begin{equation} \hat{H}(\phi=\pm\pi/2) \ket{\Psi(s)} = \omega(s-N/2)\ket{\Psi(s)} . \label{eq:dicke_scar} \end{equation}
Since our initial product state, given in Eq. \ref{eq:psi_plus}, can be rewritten in terms of the Dicke states $\ket{\boldsymbol{+}} = 2^{-N/2} \sum_{s=0}^{N}\binom{N}{s}^{1/2} \ket{\Psi(s)}$, we can use Eq. \ref{eq:dicke_scar} to write the time-evolved state: \begin{eqnarray} \ket{\psi(t)} &=& 2^{-N/2} \sum_{s=0}^{N}\binom{N}{s}^{1/2} e^{-it \omega (s-N/2)}\ket{\Psi(s)} \\ &=& \bigotimes_{n=0}^{N-1}\frac{e^{-it\omega/2}\ket{m_n=+1} + e^{it\omega/2}\ket{m_n=-1}}{\sqrt{2}} , \qquad \label{eq:psi_t_scar} \end{eqnarray} showing that the evolution takes place entirely within the Dicke scar subspace. The product state Eq. \ref{eq:psi_t_scar} is identical to the non-interacting time-evolved state given in Eq. \ref{eq:psi_t_non_int}. The error is therefore the same, $\delta\omega = \delta\omega_\text{SQL} = 1/\sqrt{NtT}$, despite the non-zero interactions and the non-integrability of the Hamiltonian. As the error is always decreasing with time it has no minimum value, i.e., the optimal sensing time diverges $t_* \to \infty$ as indicated by the numerical results in Fig. \ref{fig:opt_estimation_error}.
One might wonder if the emergence of scars (and the associated robustness in the sensing) relies on special symmetries of the spin-1 Hamiltonian at $\phi = \pm\pi/2$. Indeed, a feature of the Hamiltonian at $\phi = \pm\pi/2$ is that it has the ``particle-hole'' symmetry $\{ \hat{\Pi},\hat{H} \} = 0$, where $\hat{\Pi} = \bigotimes_n [\ket{0_n}\bra{0_n} + \ket{1_n}\bra{-1_n} + \ket{-1_n}\bra{1_n}]$. However, as was noted in Refs. \cite{Sch-19, Mar-20b}, this symmetry can be broken by including a term $\hat{H}_D = D\sum_n (\hat{S}_n^z)^2$ in the Hamiltonian, without disturbing the Dicke scar states $\{ \ket{\Psi(s)} \}$. The associated sensing performance therefore remains at the standard quantum limit when $\phi=\pm\pi/2$, despite the addition of $\hat{H}_D$ to the Hamiltonian.
\begin{figure}
\includegraphics[width=\columnwidth]{eigvec_exp_vals_and_ent_entropy.png}
\caption{(a) Expectation values of the observable $\hat{N}_0$, and (b) the half-chain entanglement entropy $S_E$, clearly show the violation of the ETH for the Dicke scar states (the band of isolated states separated from the bulk, with low entropy and $\langle \hat{N}_0 \rangle = 0$). The color scale corresponds to the density of data points, with high-density regions in yellow and low-density regions in black. Parameters: $N = 8$, $\omega = 1.0$, $\Omega = 0.3$, $\eta=D=0$, $\phi = \pi/2$, $\lambda_{n,n'} = c_{n,n'}/(n-n')^2$ where $c_{n,n'}$ is chosen at random from the interval $[0.5,1.0]$, to break translation invariance.}
\label{fig:eigvec_exp_vals_and_ent_entropy}
\end{figure}
The spin-1 Hamiltonian $\hat{H}$ also has a $U(1)$ symmetry associated with the conserved total magnetization $[\hat{H}, \sum_n \hat{S}_n^z] = 0$. However, this symmetry can be broken by adding the local term $\hat{H}_\Omega = \frac{\Omega}{2} \sum_n[ e^{i\eta} (\hat{S}_n^+)^2 + e^{-i\eta} (\hat{S}_n^-)^2]$ to the Hamiltonian. Then the non-interacting part of the Hamiltonian can be written as $\hat{H}_0 + \hat{H}_\Omega = \sqrt{\omega^2 + \Omega^2} \; \hat{U}\hat{H}_0\hat{U}^\dagger$, where $\hat{U} \equiv \exp[\frac{1}{2}e^{-i\eta}\arctan\left(\frac{\Omega}{\omega}\right)\sum_n(\hat{S}_n^-)^2 + \text{h.c.}] $. Despite the broken magnetization symmetry, the full spin-1 Hamiltonian has the rotated set of Dicke scar states $\{\hat{U}\ket{\Psi(s)}\}$ (this is shown in Appendix \ref{app:Dicke_scars}). If we consider a sensing scheme to estimate the Hamiltonian parameter $\sqrt{\omega^2 + \Omega^2}$, with the initial state $\hat{U}\ket{\boldsymbol{+}}$ and the measurement observable $\hat{U}\mathcal{O}_\theta\hat{U}^\dagger$, then the error is again at the standard quantum limit. It appears that the only symmetry that cannot be broken in the Hamiltonian $\hat{H}(\phi=\pm\pi/2)$ without destroying the scars, is the number parity symmetry $[\hat{H},\hat{P}_0]=0$, where $\hat{P}_0 \equiv (-1)^{\hat{N}_0}$ and $\hat{N}_0 \equiv \sum_n \ket{m_n=0}\bra{m_n=0}$.
To verify that the ETH is violated by the Dicke states we consider the expectation values $\bra{E}\hat{N}_0\ket{E}/N$ for eigenstates $\ket{E}$ of the Hamiltonian. For convenience, we suppose that the Hamiltonian has no symmetry except for the conservation of number parity $[\hat{H},\hat{P}_0]=0$. For such an example we expect an infinite temperature thermal eigenstate to have $\bra{E}\hat{N}_0\ket{E}/N \approx 1/3$ (i.e., $1/3$ probability of being in each of the three local basis states $m\in\{-1,0,1\}$). However, the Dicke states all have $\bra{\Psi(s)} \hat{N}_0 \ket{\Psi(s)} = 0$. This ETH-violation is shown numerically for an example in Fig. \ref{fig:eigvec_exp_vals_and_ent_entropy}(a). The half-chain entanglement entropy for each eigenstate in the same example is shown in Fig. \ref{fig:eigvec_exp_vals_and_ent_entropy}(b). We note that it is possible to analytically calculate the entanglement entropy of any Dicke state $\ket{\Psi(s)}$ for any bipartition of the $N$ spins \cite{Mor-18b}. The scar state with the largest entanglement entropy is the $s=N/2$ Dicke state. For a bipartition into two equal-size clusters of spins, its entanglement entropy tends to $S \to \frac{1}{2}\log(N/2)$ as $N\to\infty$. All of the Dicke states therefore have a sub-volume law growth of entanglement entropy \cite{Sch-19}.
\section{Example: spin-1/2 MFI model}\label{sec:MFIM}
\subsection{Quantum sensing}
The robust quantum sensing via many-body scars is not specific to the spin-1 DMI model. In this example we consider a sensing protocol in which the goal is the estimation of the transverse field parameter $\omega$ in a chain of $N$ spin-$1/2$ particles with the mixed-field Ising (MFI) Hamiltonian: \begin{equation} \hat{H}_\text{MFI} = \frac{\omega}{2} \sum_{n=0}^{N-1} \hat{\sigma}_n^x + \frac{\Omega}{2} \sum_{n=0}^{N-1} \hat{\sigma}_n^z + \frac{\lambda}{4} \sum_{n=0}^{N-1} \hat{\sigma}_n^{z}\hat{\sigma}_{n+1}^z . \label{eq:H_MFIM} \end{equation} Here $\hat{\sigma}_n^\mu$, $\mu \in \{x,y,z \}$, are the spin-1/2 Pauli operators at site $n$ on the chain and we assume the periodic boundary conditions $\hat{\sigma}_{n+N}^\mu \equiv \hat{\sigma}_{n}^\mu$.
\begin{figure}
\includegraphics[width=\columnwidth]{fig_MFIMc.png}
\caption{(a) The optimised error $\delta\omega_*$ is represented by both the size and color of the marker. The error is low when $\{ |\Omega|, |\lambda| \} \ll \omega$, since this is close to the ideal case of non-interacting spins. More interestingly, the error is also low when $|\Omega| = |\lambda| \gg \omega$, corresponding to the regime of quantum many-body scarring. (b) As quantum scars emerge the error decreases as $\delta\omega_* \sim |\lambda|^{-1/2}$ for the inital N\'{e}el state $\ket{\mathbb{Z}_2}$ (red lines). This is particularly clear when the scars are enhanced by the perturbation $\eta = -0.1\omega$ to the Hamiltonian (see Eq. \ref{eq:delta_H}). In contrast, outside the regime of QMBS the error increases as $\delta\omega_* \sim |\lambda|^{1/2}$ (black line). For the initially polarised state $\ket{\Downarrow} \equiv \ket{\downarrow\downarrow\downarrow...}$ the error plateaus, but does not decrease as $\lambda$ increases (orange line). (c) As perfect scars emerge the optimal sensing time diverges as $t_* \sim |\lambda|$. [Parameters (unless otherwise stated in legend): $N = 16$; $\ket{\psi(0)} = \ket{\mathbb{Z}_2}$, $\eta = 0$, $\omega = 1$.]}
\label{fig:main}
\end{figure}
After preparing the spins in the initial N\'{e}el state $\ket{\psi(0)} = \ket{\mathbb{Z}_2} \equiv \ket{\uparrow\downarrow\uparrow\downarrow\hdots}$, the spin system evolves by the Hamiltonian $\hat{H}_\text{MFI}$ for a sensing time $t$, followed by the measurement of an observable $\hat{\mathcal{O}} \in \mathcal{M}$, where $\mathcal{M}$ is some set of experimentally accessible observables. The measurement is repeated $T/t$ times, resulting in the estimation error given in Eq. \ref{eq:error_formula}. For the purposes of this example it is sufficient to consider the space of accessible measurement observables $\mathcal{M}$ to be spanned by the basis set $\{ \hat{J}_\text{odd}^\mu , \hat{J}_\text{even}^\mu \}_{\mu\in\{x,y,z\}}$, where $\hat{J}_\text{odd/even}^\mu \equiv \frac{1}{2}\sum_{n \text{ odd/even}} \hat{\sigma}_n^\mu $. A measurement observable is thus of the form $\hat{\mathcal{O}} = \sum_{\mu\in\{x,y,z\}} (c_\text{odd}^\mu \hat{J}_\text{odd}^\mu + c_\text{even}^\mu \hat{J}_\text{even}^\mu)$ for $c_\text{odd/even}^\mu \in \mathbb{R}$. We note that, in practice, such a measurement can be implemented if it is possible to perform arbitrary collective rotations of the odd/even spins separately, just prior to the measurement of a single collective observable, e.g., the total magnetization $\hat{J}^z = \hat{J}_\text{odd}^z + \hat{J}_\text{even}^z$. Such operations do not require single-site addressability of the spins. Also, the initial N\'{e}el state $\ket{\mathbb{Z}_2}$ is often easily prepared, for example by cooling to the ground state of a Hamiltonian with a strong anti-ferromagnetic interaction, or a nearest-neighbour blockade (as in the experiments of Ref. \cite{Ber-17}).
In the case of non-interacting spins without a longitudinal field ($\Omega = \lambda = 0$) the time-evolved state is a product state. The error is easily calculated and is given by the standard quantum limit $\delta\omega_\text{SQL} = 1/\sqrt{NtT}$. For $\lambda \neq 0$ the error no longer decreases monotonically in time, and so we focus our attention on the minimum error $\delta\omega_* = \min_t \delta\omega$. If, in addition, $\Omega \neq 0$ the Hamiltonian is non-integrable, and we resort to numerical simulation to obtain our results. These results are summarised in Fig. \ref{fig:main}(a) where we plot the optimised error $\delta\omega_* = \min_t \delta\omega$ for various values of the longitudinal field $\Omega$ and the Ising coupling $\lambda$. We see that the error $\delta\omega_*$ is small when $\{|\Omega|, |\lambda|\} \ll |\omega|$. This is not surprising, since it is close to the case of non-interacting spins precessing in the transverse field $\omega$. As the interaction strength $\lambda$ increases the approximation to non-interacting spins begins to break down, which we expect to degrade the error. Fig. \ref{fig:main}(a) shows that this is true for the most part, but that something unusual happens for $|\Omega| = |\lambda| \gg |\omega|$, where the error remains small. To show this effect in more detail, in Fig. \ref{fig:main}(b) we plot a cross-section of Fig. \ref{fig:main}(a) along the $\Omega = \lambda$ diagonal line. Along this cross-section we see that the estimation error does not simply increase monotonically as the interaction strength $\lambda$ increases. Rather, there is a range of values for which increasing the interaction strength \emph{improves} the estimation error, before the error eventually plateaus for a sufficiently large interaction strength $\lambda$. Fig. \ref{fig:main}(c) shows that this is associated with an optimal sensing time $t_*$ that is also enhanced by increasing $\lambda$. In contrast, a cross-section along the line $\Omega = 2\lambda$ shows that increasing interactions result in a degraded sensitivity, with the error increasing as $\delta\omega_* \sim |\lambda|^{1/2}$ and the optimal sensing time decreasing as $t_* \sim |\lambda|^{-1}$. This is consistent with the usual expectation that increased interactions between spins leads to faster decoherence.
In the next section we will see that, as with the spin-1 example, the reason for the improved sensor performance is the emergence of quantum many-body scars in the parameter regime $|\Omega| = |\lambda| \gg |\omega|$. Before that however, we consider a perturbation to the mixed-field Ising Hamiltonian that was shown by Choi \emph{et al.} to enhance the many-body scars in that parameter regime \cite{Cho-19}. The perturbation is $\hat{H}=\hat{H}_\text{MFI} + \delta\hat{H}$ where: \begin{equation} \delta\hat{H} \equiv \frac{\eta}{4} \sum_{n=0}^{N-1} \sum_{d=2}^{N/2} c_d (\hat{\sigma}_n^x \hat{\sigma}_{n+d}^z + \hat{\sigma}_n^z \hat{\sigma}_{n+d}^x) , \label{eq:delta_H} \end{equation} with $c_d = (\phi^{d-1} - \phi^{-d+1})^{-2}$ and $\phi = (1 + \sqrt{5})/2$ the golden ratio. Choosing $\eta = -0.1\omega$ we see in Fig. \ref{fig:main}(b) that, with this perturbation, a scaling $\delta\omega_* \sim |\lambda|^{-1/2}$ is maintained for a large range of interaction strengths $\lambda$, giving very low error for large interaction strength. Similarly, for $\lambda \gg \omega$ the sensing time scales as $t_* \sim |\lambda|$, i.e., longer sensing times are achieved with stronger interactions. We now explain that the emergence of quantum many-body scars are responsible for this unusual enhancement in sensitivity with increasing interaction strength.
\subsection{Quantum many-body scars in the MFI model}\label{sec:MFI_scars}
If $\Omega = \lambda$ the mixed-field Ising Hamiltonian can be rewritten as: \begin{equation} \hat{H}_\text{MFI} = \frac{\omega}{2}\sum_{n=0}^{N-1}\hat{\sigma}_n^x + \lambda \sum_{n=0}^{N-1} \ket{\uparrow_n\uparrow_{n+1}}\bra{\uparrow_n\uparrow_{n+1}} , \end{equation} up to an added constant that just shifts all energies by an equal amount. We can see that for $\Omega = \lambda \gg |\omega|$ states $\ket{\hdots\uparrow\uparrow\hdots}$ with two consecutive $\uparrow$-states have a large energy penalty (alternatively, if $\Omega = -\lambda$ states $\ket{\hdots\downarrow\downarrow\hdots}$ with two consecutive $\downarrow$-states have a large energy penalty). Neglecting these states and performing a rotating wave approximation gives the effective ``PXP Hamiltonian'' \cite{Les-11}: \begin{equation} \hat{H}_\text{PXP} = \hat{P}\left( \frac{\omega}{2}\sum_{n=0}^{N-1}\hat{\sigma}_n^x \right)\hat{P} , \end{equation} where $\hat{P} = \prod_{n=0}^{N-1} ( \mathbb{I} - \ket{\uparrow_n\uparrow_{n+1}}\bra{\uparrow_n\uparrow_{n+1}})$ is a projector that forbids any transitions into states with neighbouring spins in the $\uparrow$-state. The PXP Hamiltonian is known to have a set of quantum many-body scars \cite{Tur-18a}. The scars have approximately equal energy gaps, and a large overlap with the initial N\'{e}el state $\ket{\mathbb{Z}_2}$. This results in long-lived revivals of the initial state, and is the origin of the enhanced sensor performance in the parameter regime $\Omega = \lambda \gg \omega$. However, the revivals are not perfect, and they do eventually decay, corresponding to a finite $t_*$ in our numerical simulations of the sensing experiment. It was shown by Choi \emph{et. al} \cite{Cho-19} that the energy gaps between the scar states can be made almost exactly harmonic, and the revivals almost perfect, by adding the perturbation $\hat{P}(\delta\hat{H})\hat{P}$ to the PXP Hamiltonian, with $\delta\hat{H}$ as given in Eq. \ref{eq:delta_H}. If this perturbation leads to perfect revivals of the initial state we can expect the optimal sensing time $t_*$ to diverge in the PXP-limit. This is the origin of the $t_* \sim |\lambda|$ scaling for $\lambda \gg \omega$ in Fig. \ref{fig:main}(c).
Finally, we note that for the non-interacting model $\lambda=\Omega=\eta=0$, we get the same error $\delta\omega = 1/\sqrt{NtT}$, whether we prepare the spins in the initial N\'{e}el state $\ket{\psi(0)}=\ket{\mathbb{Z}_2}$, or if we choose the fully polarised initial state $\ket{\psi(0)} = \ket{\Downarrow} \equiv \ket{\downarrow\downarrow\downarrow\hdots}$. However, in the parameter regime $\lambda=\Omega\gg\omega$ the N\'{e}el state lives in the scarred subspace while the polarized state does not. The polarized state will therefore thermalise and cannot give improving sensor performance with increasing interaction strength. This is shown in the yellow line in Fig. \ref{fig:main}(b).
\section{Robustness and periodic controls}\label{sec:noise}
The two examples presented in section \ref{sec:spin_1} and section \ref{sec:MFIM} show that quantum sensing is very robust to certain strong interactions, as a result of quantum many-body scarring. However, it is natural to ask how stable such sensing is against other perturbations that might appear in any practical realisation of the scheme. This depends partly on the stability of the scars themselves. Despite some work on this topic \cite{Tur-18b, Khe-19, Lin-20, Sur-20, Shi-20}, there is still much unknown.
From the sensing perspective, the situation can be compared to attempts to design a quantum sensor by suppressing all interactions between spins $\hat{H}_\text{int} \approx 0$ (see the discussion in section \ref{sec:general}). In that case, even with very good suppression, in realistic experiments there will always be unwanted small perturbations that degrade the sensor performance. Similarly, although our sensing schemes in section \ref{sec:spin_1} and section \ref{sec:MFIM} are perfectly robust to certain strong interactions, in reality there will always be some degradation compared to the ideal case. For $\hat{H}_\text{int} \approx 0$ it is well known that the application of periodic controls can suppress unwanted interactions \cite{Vio-98}, and that this is compatible with the estimation of alternating signals \cite{Deg-17, Tay-08}. Below, we show that this approach can, in principle, be tailored to quantum sensing in a strongly interacting system with many-body scars.
To illustrate the idea we return to our spin-1 Hamiltonian, this time assuming a one-dimensional nearest neighbour interaction $\hat{H}_\text{int} = \lambda \sum_{n}(e^{i\phi}\hat{S}_n^+\hat{S}_{n+1}^- + \text{h.c.})$. Recall that this can be rewritten as $\hat{H}_\text{int} = \lambda \sum_n [ \cos\phi (\hat{S}_n^x\hat{S}_{n+1}^x + \hat{S}_n^y \hat{S}_{n+1}^y) + \sin\phi (\hat{S}_n^x\hat{S}_{n+1}^y - \hat{S}_n^y \hat{S}_{n+1}^x)]$, separating it into its XX-interaction (the $\cos\phi$ term) and DMI contribution (the $\sin\phi$ term). In Eq. \ref{eq:error_fit} we showed (for a slightly different choice of couplings $\lambda_{n,n'}$) that the optimised error $\delta\omega_* \sim \sqrt{|\lambda\cos\phi|}$ is degraded by the strength of the XX-interaction. This seems to indicate that the sensing is not very robust to such a perturbation.
Now consider the unitary operator $\hat{V}_{\pi} \equiv e^{i\pi\hat{S}_0^x}e^{i\pi\hat{S}_1^y}e^{i\pi\hat{S}_2^x}e^{i\pi\hat{S}_3^y}\hdots$, which implements a $\pi$-rotation of each spin around its $x$-axis (for even spin index $n$) or around its $y$-axis (for odd spin index $n$). The $\pi$-pulse $\hat{V}_{\pi}$ has the property that it commutes with the DMI component of $\hat{H}_\text{int}$, but anticommutes with the damaging XX-component, so that $\hat{V}_{\pi}\hat{H}_\text{int}\hat{V}_{\pi} = \lambda \sum_n [ - \cos\phi (\hat{S}_n^x\hat{S}_{n+1}^x + \hat{S}_n^y \hat{S}_{n+1}^y) + \sin\phi (\hat{S}_n^x\hat{S}_{n+1}^y - \hat{S}_n^y \hat{S}_{n+1}^x)]$. We might therefore expect that periodic application of $\hat{V}_{\pi}$ to the spins will have the effect of suppressing the unwanted XX-interaction, while leaving the DMI, to which the sensing is already robust, relatively unaffected. Note, however, that for $\hat{H}_0 = \frac{\omega}{2}\sum_n\hat{S}_n^z$ we have $\hat{V}_{\pi} \hat{H}_0 \hat{V}_{\pi} = -\hat{H}_0$, so that the signal we wish to measure is also suppressed by the pulse. This is a well known issue in quantum sensing with periodic controls, and can be overcome if we modify our scheme to measure a signal that is alternating in time at the pulse frequency, $\hat{H}_0 \to \hat{H}_0(t) = \frac{\omega}{2}\sin(\pi t/\tau) \sum_n \hat{S}_n^z$, where $\tau$ is the time interval between the periodic $\pi$-pulses. With this modification the reversal of sign of $\hat{H}_0$ on the application of a $\pi$-pulse is compensated by the change of sign in the sinusoid, giving an overall accumulation of signal \cite{Tay-08}.
\begin{figure}
\includegraphics[width=\columnwidth]{error_with_controls.pdf}
\caption{Deviations from $\phi=\pi/2$ (ideal DMI) can degrade the sensitivity in estimating a static parameter $\omega$ (purple line). Periodic $\pi$-pulses can suppress the damaging XX-interaction, extend the optimal sensing time and enhance the sensitivity (in a modified scheme to estimate the amplitude $\omega$ of an alternating signal). Here we plot the error at time $t$ assuming that there have been $m$ periodic $\pi$-pulses $\hat{V}_\pi$ (at the times $t/m, 2t/m,\hdots, t$). [Other parameters: $N=10$, $\lambda = 0.8$, $\phi = 0.4\pi$, $\omega = 1$.]}
\label{fig:error_with_controls}
\end{figure}
In Fig. \ref{fig:error_with_controls} we plot the error in estimating the amplitude $\omega$ of the alternating signal at the sensing time $t$, assuming that the $\pi$-pulse $\hat{V}_\pi$ has been applied $m$ times at periodic intervals $\tau = t/m$ during the preceeding dynamics. We see that the periodic controls suppress the damaging XX-interaction, extend the optimal sensing time and enhance the optimal sensitivity.
Another likely source of noise is inhomogeneity in the local magnetic fields, represented by an added Hamiltonian term $\hat{H}_\Delta = \sum_n \Delta_n \hat{S}_n^z$. Since $\hat{V}_{\pi}\hat{H}_\Delta \hat{V}_{\pi} = -\hat{H}_\Delta$ the $\pi$-pulse is also effective at suppressing such perturbations. In the limit of high-frequency periodic control, $m\to\infty$, the optimal sensing time diverges and the estimation error is $\delta\omega = 2 / \sqrt{\pi^2 NtT}$, as shown in Appendix \ref{app:noise}.
We note that the simple periodic control considered here does not suppress \emph{next}-nearest neighbour XX-interactions or interactions of the form $\hat{H}_Z = \sum_{n,n'}g_{n,n'}\hat{S}_n^z\hat{S}_{n'}^z$, if they are included. Also, pulse errors have not been taken into account in the analysis above, although they may be significant if many pulses $m \gg 1$ are applied. One possible approach to these issues might be to develop a more sophisticated pulse sequence that is robust to more generic noise and to pulse errors \cite{DeL-10}. Another promising approach -- recently demonstated experimentally -- is to employ periodic driving to stabilise the scars and the associated long-lived oscillations by the creation of a robust discrete time-crystal-like phase \cite{Blu-21, Mas-21}. Further work is required to determine if such a robust non-equilibrium phase can be exploited for quantum sensing.
\section{Conclusion}
Quantum many-body scars are special eigenstates of a non-integrable many-body system that, for certain initial states, can prevent or slow down thermalisation. Since this is associated with long coherence times, scars can be exploited for quantum sensing. In this paper we have demonstrated this for two example models: a spin-1 DMI model, and a spin-1/2 MFI model.
Although the two examples appear to be very different, there are some interesting similarities in the structure of their scar subspaces \cite{Sch-19}. Recall that the scar states in the spin-1 DMI model are the Dicke states $\{\ket{\Psi(s)}\}_{s=0}^N$ (defined in Eq. \ref{eq:dicke_states}) with harmonic energy gaps, resulting in SU(2)-spin dynamics in the Dicke subspace. For the spin-1/2 PXP-model, it was shown in Ref. \cite{Tur-18a} that the dynamics in the scar subspace can be though of as approximate SU(2)-spin dynamics, with the scar states playing the role of the Dicke subspace. The perturbation introduced at the end of section \ref{sec:MFIM} improves the approximation, so that the dynamics are almost exactly like those of an SU(2) spin. Whether this SU(2) structure is an essential feature of quantum sensing via many-body scars, or if examples exist without this feature (but still satisfying our criteria in section \ref{sec:general}) is, to the best of our knowledge, an open question.
In both examples we assumed an initial product state of the $N$ probe particles. However, it is well known that for non-interacting systems, entangled initial states such as spin squeezed states can give an enhanced sensitivity compared to separable initial states \cite{Gio-04}. This is also possible for strongly interacting systems with many-body scars, as we show in Appendix \ref{app:squeezing} for the spin-1 DMI example.
Robustness to perturbations is undoubtedly an important topic of further research, if quantum sensors exploiting scars are to become a useful technology. As discussed in section \ref{sec:noise}, periodic controls may offer a route to such robustness.
\begin{acknowledgments}
The author thanks G. Kells for discussions and for helpful comments on the manuscript. This work was funded by Science Foundation Ireland through Career Development Award 15/CDA/3240. The author also wishes to acknowledge the DJEI/DES/SFI/HEA Irish Centre for High-End Computing (ICHEC) for the provision of computational facilities.
\end{acknowledgments}
|
\section{Introduction}
In the construction of a free group (see \cite{von} and \cite[Chapter 6]{DF}), one extends a generating set by a neutral element and inverse elements. The procedure of adding those elements happen entirely by using the set-theoretic operations, i.e. a disjoint union of sets. Therefore it is hard to grasp the meaning of those operations in an algebraic sense.
Thus the natural question is if there exists some approach/point of view from which one could spot the algebraic interpretation of the set-theoretic process of extending the generating set.
The approach presented in the paper uses heaps, a specific variant of universal algebras introduced by R.\ Baer \cite{Bae:ein} and H.\ Pr\"ufer \cite{Pru:the}. Due to both Pr\"ufer \cite{Pru:the} (Abelian case) and R.\ Baer \cite{Bae:ein} (general case), it is known that with every heap one can associate a group and that to every group one can assign a heap. The latter is a functor.
This paper aims to construct a left adjoint to the functor, which assigns a heap to a group, see Theorem~\ref{thm1}. The second goal is to point out what consequences it yields for free groups, see Section~\ref{sect4}. The crucial observations of the main part are Corollary~\ref{cor3} and Corollary~\ref{cor4}, which explains that the free group functor decomposes through the category of heaps.
In the conclusions, we explain that even though in general, the description of a coproduct, which is crucial to define the left adjoint, is not an easy task, in this specific case the structure is transported by the free functor of heaps and therefore is easily described. The last words of the paper briefly explain how the inverses and a neutral element arise through the employment of the free group functor.
\section{Preliminaries}
Following Baer \cite{Bae:ein} and Pr\"ufer \cite{Pru:the} a {\em heap} is a set $H$ together with a ternary operation $[-,-,-]:H\times H \times H\to H$ such that for all $h_1,h_2,h_3,h_4,h_5\in H$ the following holds
\begin{equation}\label{rule:1}
[[h_1,h_2,h_3],h_4,h_5]=[h_1,h_2,[h_3,h_4,h_5]],
\end{equation}
\begin{equation}\label{malcev}
[h_1,h_2,h_2]=h_1=[h_2,h_2,h_1].
\end{equation}
First equation is called an associativity and the second is Mal'cev identities.
If for all $h_1,h_2,h_3\in H$, $[h_1,h_2,h_3]=[h_3,h_2,h_1]$ we say that $H$ is Abelian.
A {\em sub-heap} $S$ of a heap $H$ is a subset of $H$ closed under the ternary operation.
A {\em homomorphism of heaps }is a map between heaps which preserves ternary operation. Observe that a constant map between two heaps is a heap homomorphism since, by Mal'cev identities \eqref{malcev}, a single element $e\in H$ is a sub-heap of $H$.
A sub-heap $S$ is said to be {\em normal} if there exists $e\in S$ such that for all $h \in H$ and $s \in S$ there exists $s' \in S$ such that
$$
[h, e, s] = [s', e, h] \textrm{ or equivalently } [[h,e,s],h,e]=s'
$$
If $S$ is a normal sub-heap then the quotient heap $H/S$ is well defined and canonical map $\mathrm{pi}:H\to H/S$ is a heap epimorphism, see \cite[Proposition 2.10]{Brz:par}.
An important property of heaps is that with every heap $H$ we can associate a group by choosing an element $e\in H$ and defining a binary operation $+_e:=[-,e,-]:H\times H\to H$, we will call the group $(H,+_e)$ {\em a retract of $H$ in $e$} or $e$-{\it retract} and denote by $\mathrm{G}(H;e)$. It is worth to mention that an assignment $\mathrm{G}$ is not a functor, as it is not well-defined on the morphisms. If $\varphi$ is a homomorphism of heaps then $\varphi$ is a homomorphism of appropriate retracts if and only if it preserves neutral elements of the retracts.
In the opposite direction, one can associate with every group $G$ a heap by defining ternary operation, for all $a,b,c\in G$, as $[a,b,c]:=ab^{-1}c$. We call this heap a {\em heap associated with a group G} and denote it by $\mathrm{H}(G)$. In contrast to the assignment $\mathrm{G}$, the assignment $\mathrm{H}:\mathbf{Grp}\to \mathbf{Heap}$, between categories of groups and heaps, is a functor given on morphisms $\varphi:G\to G'$ by $\mathrm{H}(\varphi)=\varphi$. Every group homomorphism is a homomorphism of associated heaps.
By employing both assignments to a group $G$ one gets that a group $\mathrm{G}(\mathrm{H}(G);e)$ is isomorphic to $G$, for all $e\in G$. By applying assignments to a heap $H$, we get that $\mathrm{H}(\mathrm{G}(H;e))=H$, for all $e\in H$. The second link between groups and heaps can be used to show that for all $h_1,h_2,h_3,h_4,h_5\in H,$
\begin{equation}\label{rule:2}
[[h_1,h_2,h_3],h_4,h_5]=[h_1,[h_4,h_3,h_2],h_5]=[h_1,h_2,[h_3,h_4,h_5]],
\end{equation}
see Lemma~2.3 of \cite{Brz:par}.
In Section 3 of \cite{BrzRyb:mod}, one can find the construction of a free heap over a set $X$, $\mathcal{H}(X)$. Since heaps form a variety of algebras a free functor exists, we will denote it by $\mathcal{H}:\mathbf{Set}\to\mathbf{Heap}$ to be coherent with the notation of free heaps. It is a left adjoint to the forgetful functor $\mathrm{U}_{\mathbf{Heap}}:\mathbf{Heap} \to\mathbf{Set}$. In contrast to the case of a free group, in a free heap one does not extend the set of letters by any new letters such as a neutral element or inverse letters.
Since heaps form a variety of algebras small colimits in $\mathbf{Heap}$ exist (see \cite[Theorem 9.4.14]{Ber:inv}), particularly the one in the centre of our attention will be a coproduct.
A {\em coproduct} of two objects $A$ and $B$ in a category $\mathfrak{C}$ is an object $C$ with two morphisms $\iota_A:A\to C$ and $\iota_B:B\to C$, called canonical injections, such that for any object $D$ and morphisms $f:A\to D$ and $g:B\to D$, there exists a unique morphism $\varphi:C\to D$ such that $\varphi\circ \iota_A=f$ and $\varphi\circ \iota_B=g$. In the diagram-like style, the diagram
\begin{equation}\label{sum.diag}
\xymatrix{&& D && \cr A \ar[rr]^{\iota_A}\ar[urr]^f & &C \ar@{-->}[u]_{\varphi} & & B\ar[ll]_{\iota_B}\ar[ull]_g }
\end{equation}
commutes.
In Section 3 of \cite{BrzRyb:mod} one can find the construction of a coproduct in the {\em category of Abelian heaps} $\mathbf{Ah}$, a full subcategory of $\mathbf{Heap}$. The idea is to take two Abelian heaps $A$ and $B$, consider the free heap over theirs disjoint union $\mathcal{H}(A\sqcup B)$ and then divide it by the sub-heap generated by
\begin{equation}\label{eq:1}
[[a,a',a''],[a,a',a'']_{A},e],\quad [[b,b',b''],[b,b',b'']_{B},e],
\end{equation}
for all $a,a',a''\in A$, $b,b',b''\in B$, where
$[---],[---]_{A},[---]_{B}$ are ternary operations in $\mathcal{H}(A\sqcup B)$, $A$ and $B$, respectively.
\section{Main part}
Let us fix some notation. Following \cite{BrzRyb:mod} we will denote the coproduct of two not necessarily Abelian heaps $H$ and $S$ by $H\boxplus S$. It exists by \cite[Theorem 9.4.14]{Ber:inv}, though we do not know exact construction of it for non-Abelian heaps, for Abelian case see \cite[Section 3]{BrzRyb:mod}. The unique filler of the coproduct diagram for morphisms $f$ and $g$, will be called a {\em coproduct map} and will be denoted by $f\boxplus g$.
Our first and main goal is to construct a left adjoint functor to the functor $\mathrm{H}:\mathbf{Grp}\to \mathbf{Heap}$.
A singleton heap is a heap that has only one element, we will denote it by $\{*\}$. For any heap $H$, one can consider a group $\mathrm{Gr}_{*}(H):=\mathrm{G}(H\boxplus\{*\};*)$. The following lemma shows that this group has a very interesting universal property, which will be essential in the construction of the adjoint.
\begin{lemma}\label{heap:lem:univ}
Let $H$ be a heap, S be a group and $f: H\to \mathrm{H}(S)$ be a heap homomorphism. Then there exists a unique group homomorphism $\mathrm{Gr}_{*}(f):\mathrm{Gr}_{*}(H)\to S$ such that $f=\mathrm{H}(\mathrm{Gr}_{*}(f))\circ \iota_{H}$, where $\iota_H$ is a canonical injection into coproduct. In other words, diagram
\begin{equation}\label{graf1}
\xymatrix {H\ar[rr]^-{\iota_{H}} \ar[rrdd]_-f&& \mathrm{H}(\mathrm{Gr}_{*}(H)) \ar@{-->}[dd]^-{\exists !\, \mathrm{H}(\mathrm{Gr}_{*}(f))}
\\
\\
&& \mathrm{H}(S) &}
\end{equation}
commutes, where $\exists !\, \mathrm{H}(\mathrm{Gr}_{*}(f))$ reads ``\it{there exists exactly one \textbf{homomorphism of groups} $\mathrm{Gr}_{*}(f)$}". The pair $(\mathrm{Gr}_{*}(H),\iota_H)$ is a universal arrow, see \cite[Section III.1]{Mac:lane}.
\end{lemma}
\begin{proof}
Observe that by the universal property of coproduct for all groups $S$ and homomorphisms of heaps $f:H\to \mathrm{H}(S)$, $g:\{*\}\to \mathrm{H}(S)$ diagram
\begin{equation}
\xymatrix{&& \mathrm{H}(S)&& \\ H \ar[rr]^{\iota_H}\ar[urr]^{f} & &\mathrm{H}(\mathrm{Gr}_{*}(H)) \ar@{-->}[u]_{\mathrm{H}(\mathrm{Gr}_{*}(f))} & & \{*\}\ar[ll]_{\iota_*}\ar[ull]_{g} }
\end{equation}
commutes. Every homomorphism of groups is a homomorphism of associated heaps. Moreover, a homomorphism of heaps is a homomorphism of retracts if, and only if it maps a neutral element to a neutral element. Hence, $\mathrm{H}(\mathrm{Gr}_{*}(f))$ is a homomorphism of retracts if and only if $g(\iota_*(*))$ is a neutral element of $S$. Observe that $g$ is unique, since $\{*\}$ is a singleton heap. Therefore $\mathrm{H}(\mathrm{Gr}_{*}(f))$ is a unique homomorphism of heaps such that it is also a homomorphism of groups to which heaps were associated. Thus, the preceding diagram commutes.
\end{proof}
Another important observation is that a canonical injection $\iota_H$ has some sort of cancellation property.
\begin{lemma}\label{heap:lem:stepi}
Let $H,L$ be heaps and $f,g:\mathrm{H}(\mathrm{Gr}_{*}(H))\to L$ be homomorphisms of heaps such that $f(\iota_*(*))=g(\iota_*(*))$, then $f\circ \iota_H=g\circ \iota_H$ implies $f=g$.
\end{lemma}
\begin{proof}
Let us consider a homomorphism of heaps $f:\mathrm{H}(\mathrm{Gr}_{*}(H))\to L$. One can easily observe that by the uniqueness of a coproduct map $f=(f\circ\iota_H)\boxplus (f\circ\iota_{*})$. Thus, because $f(\iota_*(*))=g(\iota_{*}(*))$ and $f\circ \iota_H=g\circ \iota_H$, we get that
$$
f=(f\circ\iota_H)\boxplus (f\circ\iota_{*})=(g\circ\iota_H)\boxplus (g\circ\iota_{*})=g.
$$
Therefore, $f(\iota_*(*))=g(\iota_*(*))$ and $(f\circ \iota_H)=(g\circ \iota_H)$ implies $f=g$.
\end{proof}
\begin{corollary}\label{heap:cor:stepi}
Let $e\in L$. If $f,g:\mathrm{Gr}_{*}(H)\to \mathrm{G}(L,+_e)$, are homomorphisms of groups, then $f\circ\iota_H=g\circ\iota_H$ implies $f=g$.
\end{corollary}
\begin{proof}
This follows by Lemma \ref{heap:lem:stepi} since a homomorphism of heaps $\mathrm{H}(f)$ is equal to a homomorphism of groups $f$ as functions.
\end{proof}
The following lemma follows by Theorem 2 (ii) in \cite[Section IV]{Mac:lane}, but for the sake of the unaccustomed reader, we sketch a proof.
Now, we are ready to describe the functor. Let us consider an assignment \linebreak$\mathrm{Gr}:\mathbf{Heap}\to \mathbf{Grp}$ given on a heap $H$ by $H\mapsto \mathrm{Gr}_{*}(H)$. One can easily see that it is a well-defined function. The assignment is given for all homomorphisms of heaps $f:H\to H'$ by $f\mapsto \mathrm{Gr}_{*}(\iota_{H'}\circ f)$. The assignment on morphisms is well-defined since $\iota_H'\circ f$ is a composition of homomorphisms of heaps, so it is a homomorphism of heaps. Therefore by the universal property of $\mathrm{Gr}_{*}$, $\mathrm{Gr}_{*}(\iota_{H'}\circ f)$ is a homomorphism of groups.
\begin{lemma}
The assignment $\mathrm{Gr}:\mathbf{Heap}\to \mathbf{Grp}$ is a functor.
\end{lemma}
\begin{proof}
In the previous discussion, we explained that both assignments are well-defined functions. Thus, we have to show that functor preserves identity and composition.
Obviously $\mathrm{Gr}_{*}(\iota_{H}\circ id_H)=id_{\mathrm{Gr}_{*}(H)}$.
For the composition let us assume that $f:H\to H'$ is a homomorphism of heaps, then $\iota_{H'}\circ f$ is a composition of homomorphisms of heaps, hence $\iota_{H'}\circ f:H\to\mathrm{H}(\mathrm{Gr}(H')))$ is a homomorphism of heaps. If $f:H\to H'$ and $g:H'\to H''$ are homomorphisms of heaps, then
$$
\begin{aligned}
\mathrm{Gr}(g\circ f)\circ \iota_{H}&=\mathrm{Gr}_{*}(\iota_{H''}\circ g\circ f)\circ \iota_{H}=\iota_{H''}\circ g\circ f=\mathrm{Gr}(g)\circ \iota_{H'}\circ f\\ &=\mathrm{Gr}(g)\circ \mathrm{Gr}(f)\circ \iota_{H},
\end{aligned}
$$
where all the equalities follow by Lemma \ref{heap:lem:univ} applied multiple times. Now, since $\mathrm{Gr}(g\circ f)\circ \iota_{H}=\mathrm{Gr}(g)\circ \mathrm{Gr}(f)\circ \iota_{H}$ and $\mathrm{Gr}(g\circ f),\mathrm{Gr}(g), \mathrm{Gr}(f)$ are homomorphisms of groups, applying Corollary \ref{heap:cor:stepi}, one gets that $\mathrm{Gr}(g\circ f)=\mathrm{Gr}(g)\circ \mathrm{Gr}(f)$. Therefore an assignment $\mathrm{Gr}$ preserves composition, hence $\mathrm{Gr}$ is a functor.
\end{proof}
The following theorem confirms that $\mathrm{Gr}$ is a desirable functor and follows by the Theorem 2 (i) in \cite[Section IV]{Mac:lane}, but for the sake of the unaccustomed reader, we sketch a proof.
\begin{theorem}\label{thm1}
A functor $\mathrm{Gr}$ is a left adjoint to the functor $\mathrm{H}$.
\end{theorem}
\begin{proof}
For all heaps $H$ and groups $G$ let us consider functions between sets of morphisms:
$$
\varphi_{H,G}:\mathbf{Grp}(\mathrm{Gr}(H),G)\longrightarrow \mathbf{Heap}(H,\mathrm{H}(G)),\ \
f\longmapsto \mathrm{H}(f)\circ \iota_{H},
$$
$$\varphi_{H,G}^{-1}:\mathbf{Heap}(H,\mathrm{H}(G))\longrightarrow \mathbf{Grp}(\mathrm{Gr}(H),G),\ \ f\mapsto \mathrm{Gr}_{*}(f).$$
To show that $\varphi_{H,G}$ is a bijection let $f\in \mathbf{Heap}(H,\mathrm{H}(G))$ and $g\in \mathbf{Grp}(\mathrm{Gr}(H),G)$, then
$$
\varphi_{H,G}\circ \varphi_{H,G}^{-1}(f)=\varphi_{H,G}(\mathrm{Gr}_{*}(f))=\mathrm{H}(\mathrm{Gr}_{*}(f))\circ\iota_H=f,
$$
where the last equality follows by Lemma~\ref{heap:lem:univ}, and
$$
\varphi_{H,G}^{-1}\circ \varphi_{H,G}(g)=\mathrm{Gr}_{*}(\mathrm{H}(g)\circ\iota_H)=g,
$$
where the last equality follows by the uniqueness of the morphism $\mathrm{Gr}_{*}(\mathrm{H}(f)\circ\iota_H)$.
Hence, $\varphi_{H,G}^{-1}$ is an inverse to $\varphi_{H,G}$. Thus, $\varphi_{H,G}$ is a bijection.
To check naturality conditions, let $G,S$ be groups , $H,L$ be heaps and consider homomorphisms $f:\mathrm{Gr}(H)\longrightarrow G$, $\alpha:L\longrightarrow H$ and $g:G\longrightarrow S$. Then
$$
\varphi_{L,G}(f\circ \mathrm{Gr}(\alpha))=\mathrm{H}(f\circ \mathrm{Gr}(\alpha))\circ\iota_{L}=\mathrm{H}(f)\circ \mathrm{H}(\mathrm{Gr}(\alpha))\circ\iota_{L}=\mathrm{H}(f)\circ\iota_{H}\circ \alpha=\varphi_{H,G}(f)\circ \alpha,
$$
by applying Lemma~\ref{heap:lem:univ} multiple times.
Similarly,
$$
\varphi_{H,S}(g\circ f)=\mathrm{H}(g\circ f)\circ \iota_{H}=\mathrm{H}(g)\circ \varphi_{H,G}(f).
$$
Therefore $\varphi$ is a natural isomorphism and the functor $\mathrm{Gr}$ is a left adjoint to the functor{ }~$\mathrm{H}$.
\end{proof}
\begin{proposition}\label{cor:monad}
The adjunction $\mathrm{Gr} \dashv \mathrm{H}$ is monadic.
\end{proposition}
\begin{proof}
By \cite[3.14. Theorem]{TTT} it is enough to show that $\mathrm{H}$ reflects isomorphisms, $\mathbf{Grp}$ has coequalizers of $\mathrm{H}$-split parallel pairs, and $\mathrm{H}$ preserves those coequalizers.
Let us start with the property of reflecting isomorphisms. Let $g:G\to G'$ be a homomorphism of groups such that $\mathrm{H}(g)$ is an isomorphism of heaps. Then, since $\mathrm{H}(g)=g$ as functions there exists a function which is an inverse to $g$, i.e. $\mathrm{H}(g)^{-1}$. It is obvious that inverse is a group homomorphism. Therefore, $\mathrm{H}$ reflects isomorphisms.
To prove that coequalizers for $\mathrm{H}$-split parallel pairs exist let $f,g:G\to G'$ be an $\mathrm{H}$-split parallel pair, then by definition there exist heap $H$ and heap homomorphisms $$h:\mathrm{H}(G')\to H,\ t:H\to\mathrm{H}(G'),\ s:\mathrm{H}(G')\to \mathrm{H}(G)$$ such that $h\circ \mathrm{H}(f)=h\circ\mathrm{H}(g)$, $s$ and $t$ are sections of $\mathrm{H}(f)$ and $h$, respectively, and $\mathrm{H}(g)\circ s=t \circ h$.
One can check that existence of those homomorphisms imply that a pair $(H,h)$ is a coequalizer of $\mathrm{H}(f)$ and $\mathrm{H}(g)$. Now, let us denote by $e\in G$ and $e'\in G'$ neutral elements of the groups, then we can consider an $h(e')$-retract of $H$, a group $\mathrm{G}(H;h(e'))$.
Observe that all the aforementioned homomorphisms preserve neutral elements of groups. Thus, all of those homomorphisms are homomorphisms of appropriate retracts.
Therefore $(\mathrm{G}(H,e),h)$ is a coequalizer of $\mathrm{H}$-split parallel pair $(f,g)$ and $h$ is surjective. Thus, $\mathrm{H}(\mathrm{G}(H;h(e')))=H$ and $\mathrm{H}$ preserves the coequalizers. Hence, the adjunction is monadic.
\end{proof}
To underline the meaning of the preceding theorem in the context of groups let us consider the following diagram
\begin{large}
\[
\xymatrix @C=30pt{
\mathbf{Grp} \ar@<+0.5ex>[dr]^{\mathrm{H}} \ar@<+0.5ex>[dd]^{\mathrm{U}_{\mathbf{Grp}}} & \\
& \mathbf{Heap}\ , \ar@<+0.5ex>[dl]^{\mathrm{U}_{\mathbf{Heap}}} \ar@<+0.5ex>[ul]^-{\mathrm{Gr}} \\
\mathbf{Set} \ar@<+0.5ex>[uu]^{\mathcal{G}} \ar@<+0.5ex>[ur]^{\mathcal{H}} &
}
\]
\end{large}
where $U_{\mathbf{Grp}}$ is a forgetful functor and $\mathcal{G}$ is its left adjoint, the free functor.
The first observation is that all the opposite arrows are adjoints to each other.
The second observation is that a composition of functors,
$$
\xymatrix{\mathbf{Grp}\ar[rr]^{\mathrm{H}} &&\mathbf{Heap}\ar[rr]^{U_{\mathbf{Heap}}}&&\mathbf{Set}}
$$
is a forgetful functor $\mathrm{U}_{\mathbf{Grp}}$ since for any group $G$, $\mathrm{H}(G)$ and $G$ are equal sets, and every homomorphism of groups $f$ is the same function as $\mathrm{H}(f)$. These two observations leads to the following corollaries.
\begin{corollary}\label{cor3}
A functor $\mathrm{Gr}\circ \mathcal{H}:\mathbf{Set}\to \mathbf{Grp}$ is a free functor, i.e. it is a left adjoint to the functor $U_{\mathbf{Grp}}:\mathbf{Grp}\to \mathbf{Set}$.
\end{corollary}
\begin{proof}
It is easy to show that a composition of two left adjoints is a left adjoint to the composition.
\end{proof}
\begin{corollary}\label{cor4}
For any set $X$, $(\mathrm{Gr}\circ\mathcal{H})(X)\cong\mathcal{G}(X)$.
\end{corollary}
\begin{proof}
Since both functors $\mathrm{Gr}\circ\mathcal{H}$ and $\mathcal{G}$ are left adjoints to the forgetful functor, they are naturally isomorphic, see \cite[Corollary 1, page 85]{Mac:lane}.
\end{proof}
\section{Conclusions}\label{sect4}
To summarise, in the main part we have shown that a free functor from the category of sets to the category of groups is decomposable into two functors, through the category of heaps. The description of the free functor provides a method to construct a free group.
Sadly, in general, it is not an easy task to describe a coproduct of heaps. One intuitively knows it is a quotient of a free heap over the disjoint union. The choice of generators for the normal sub-heap is at least tricky, because one must at the same time deal with the allocations of a ternary operation in elements of the free heap, see the associativity rule \eqref{rule:2}.
Fortunately, since we are interested in a composition of functors $\mathrm{Gr}\circ\mathcal{H}$, we only need to consider the coproduct of two heaps, the singleton heap and a free heap $\mathcal{H}(X)$, for any set $X$. Observe that a heap described on the singleton set is unique up to isomorphism. Thus, we can identify a singleton heap with a free heap $\mathcal{H}(\{*\})$. Now, by definition of $\mathrm{Gr}\circ\mathcal{H}$,
$(\mathrm{Gr}\circ\mathcal{H})(X)=\mathrm{G}(\mathcal{H}(X)\boxplus\mathcal{H}(\{*\});*)$, but $\mathcal{H}$ is a left adjoint functor to the forgetful functor, so it preserves coproducts. Therefore, we have that $\mathcal{H}(X)\boxplus\mathcal{H}(\{*\})\cong\mathcal{H}(X \sqcup \{*\})$. The construction of a free heap is well-known, see for example \cite{BrzRyb:mod}. Hence, we start with taking a set $X$, then consider a disjoint union with $\{*\}$, construct a free heap over $X \sqcup \{*\}$ and take a retract of that heap in $*$. The obtained retract is a free group. Even though one can argue that we still add a disjoint element $*$, in this setup, it has a proper algebraic interpretation in the category of heaps, i.e. taking a coproduct of a free heap with a singleton heap. Another difference is that one does not have to extend the set of generators by inverses, as inverses in the retract are words of the form $[*,w,*]$, for any $w\in \mathcal{H}(X \sqcup \{*\})$.
Hopefully, the reader will find this observation as compelling as I do.
{\bf Acknowledgments.} The author is grateful to Tomasz Brzezi\'nski and Paolo Saracco for all the comments and advice.
|
\section{Case Study: The Mars Polar Lander}
\label{sec:casestudy}
\begin{figure}
\begin{lstlisting}[basicstyle=\ttfamily\small]
//Controller Initialization
prev_td_1 = 0; cur_td_1 = 0; prev_td_2 = 0; cur_td_2 = 0;
health_1 = 1; health_2 = 1;
engine_enabled = 1; event_enabled = 0;
while (engine_enabled == 1) {
//Controller loop start
input radar_alt;$\label{code:readradar}$
prev_td_1 = cur_td_1; input cur_td_1;$\label{code:readtd1}$
prev_td_2 = cur_td_2; input cur_td_2;$\label{code:readtd2}$
state_1 = 0; //Missing, probable cause of crash.$\label{code:statestart}$
state_2 = 0; //Missing, probable cause of crash.$\label{code:mplannii}$
if (prev_td_1 == 1 && cur_td_1 == 1) {
state_1 = 1
};
if (prev_td_2 == 1 && cur_td_2 == 1) {
state_2 = 1
};$\label{code:stateend}$
if ((state_1 == 1 && health_1 == 1) || $\label{code:enginestart}$
(state_2 == 1 && health_2 == 1) &&
event_enabled == 1) {
engine_enabled = 0;
}$\label{code:engineend}$
//Indicator health check
if (radar_alt < 40 && event_enabled == 0) {$\label{code:healthstart}$
if (prev_td_1 == 1 && cur_td_1 == 1) {
health_1 = 0;
};
if (prev_td_2 == 1 && cur_td_2 == 1) {
health_2 = 0;
}
event_enabled = 1;
}$\label{code:healthend}$
}
\end{lstlisting}
\caption{Code of the Mars Polar Lander}
\label{fig:mplcode}
\end{figure}
In this section, we show how belief programming and Epistemic Hoare Logic could be used to implement and verify the control software of the Mars Polar Lander (MPL).
The MPL is a lost space probe, hypothesized to have crashed into the surface of Mars during descent due to a control software error~\citep{mpl}.
We do not claim that belief programming is the first or only technique for preventing the loss of the MPL.
However, the notoriety and subsequent investigation of the MPL's loss has resulted in ample documentation~\citep{mpl} useful for illustrating in detail how belief programming and Epistemic Hoare Logic work.
The code presented in this section is written in BLIMP{}, except that for convenience we define two pieces of syntactic sugar.
The syntax $p_1 \code{=>} p_2$ desugars to $\ncode{!}p_1 \code{||} p_2$
and the syntax $x \code{=} p$ desugars to
$\lcode{if} p \code{\{} x \rcode{= 1}
\code{\} else \{} x \rcode{= 0} \rcode{\}}$.
The code in Figure~\ref{fig:mplcode} is the piece of MPL's software~\citep{mpl} responsible for the final phase of its Martian descent.
The code uses a radar altimeter as well as two touch sensors on the landing legs to monitor its progress along the descent.
Note that this is a simplification from the original software, which used three touch sensors.
The code consists of a state estimator to determine when it reaches the Martian surface, and control code to shut off its engine once it does.
\paragraph{Reading Observations} On Line~\ref{code:readradar} the controller reads the value of the radar altimeter into the variable \lstinline{radar_alt}.
On Lines \ref{code:readtd1} and \ref{code:readtd2} the code reads the values of the touchdown sensors into \lstinline{cur_td_1} and \lstinline{cur_td_2}.
It also stores their previous values in \lstinline{prev_td_1} and \lstinline{prev_td_2}.
\paragraph{State Updates} The block of code from Line~\ref{code:statestart}-\ref{code:stateend} sets the state variables \lstinline{state_1} and \lstinline{state_2} based on the values of the touchdown sensors.
Specifically, if an individual sensor has indicated a \lstinline{1} for two iterations in a row, its state is set to \lstinline{1}.
Otherwise, its state is set to \lstinline{0}.
Notably the annotated lines (Lines \ref{code:statestart} and \ref{code:mplannii}) were missing from the original software, meaning that any two positive sensor readings in a row were sufficient to permanently set the state to \lstinline{1}.
It is hypothesized that this was part of the sequence of events that caused the MPL to crash, and these two lines are the recommended fix~\citep{mpl}.
\paragraph{Engine Shutdown}
The block of code from Line~\ref{code:enginestart}-\ref{code:engineend} determines when to shut down the engine.
If the health check has completed (see below) and at least one of the healthy indicators registers a touchdown, then the program sets \lstinline{engine_enabled} to \lstinline{0}, shutting down the engine.
\paragraph{Health Check}
The block of code from Line~\ref{code:healthstart}-\ref{code:healthend} performs a health check on the touchdown indicators.
The health check assigns into the health variables \lstinline{health_0} and \lstinline{health_1}, and is performed the first time the radar altimeter indicates an altitude lower than 40 meters.
At this point, the lander is off of the ground so both touchdown indicators should read \lstinline{0}.
If an indicator reads \lstinline{1} on both the current and previous time step, it is assumed to be defective, and its health variable is set to \lstinline{0}.
After the health check completes, the program sets the flag \lstinline{event_enabled} to \lstinline{1} to allow the touchdown indicators to shut down the engine.
\subsection{Error Model}
From the above code, we can deduce several sources of error the MPL control software was designed to be robust with respect to:
\begin{itemize}
\item {\it Transient False Positives.} If a touch sensor is momentarily triggered, the software will not immediately assume the lander has contacted the ground.
The software requires two time steps in a row where the sensor was positive before it sets its state variable to true and thereby registers that it has contacted the ground.
The assumption here is that the duration of the transient false positive is no longer than one time step.
\item {\it Permanent False Positives.} If a touch sensor is defective, it may constantly send out an indication that the lander has contacted the ground.
This is detected and corrected for by the indicator health check on Lines~\ref{code:healthstart}-\ref{code:healthend}.
This block of code checks if the touch sensor yields a non-transient positive contact signal when the lander is just under 40 meters above the ground.
If so, the code assumes that sensor is defective and ignores its output for the engine shutoff decision.
The assumption here is that at most one sensor will be defective.
\item {\it Permanent False Negatives.} If a touch sensor is defective, it may constantly send out an indication that the lander has not contacted the ground when in fact it has.
The above code accounts for this by using two sensors and shutting off the engine when either one indicates a touchdown.
The assumption here is that at most one sensor will be defective.
\end{itemize}
\paragraph{Landing Leg Deployment.} Another source of error for the MPL that is not obvious from the code above but that has been well-documented is landing leg deployment.
When the landing legs deploy about 1500 meters above the surface, the process can result in false positives that exceed the one time step assumed for transients~\cite{mpl}.
Without the two annotated lines (Lines \ref{code:statestart} and \ref{code:mplannii}), this would cause the sensor's state variable to be permanently set to \lstinline{1}, causing the engine to shut down immediately after the health check completed.
\begin{figure}
\begin{lstlisting}[basicstyle=\ttfamily\small]
//Model Initialization
prev_err_1 = 0; prev_err_2 = 0; trans_td = 0;
alt = 8000;$\label{code:altinit}$
time_on_ground = 0;
//Permanent errors
perm_1 = choose(. == 0 || . == 1);$\label{code:permstart}$
perm_2 = choose((. == 0 || . == 1) && (perm_1 == 1 => . == 0));
perm_1_v = choose(. == 0 || . == 1);
perm_2_v = choose(. == 0 || . == 1);$\label{code:permend}$
//Controller initialization
...$\label{code:dots1}$
while(engine_enabled == 1) $\label{code:loopcondstart}$
{ (alt > 0 => engine_enabled == 1) && $\label{code:loopinvmpl}$
(trans_td == 0 => time_on_ground < 2) }$\label{code:loopcondend}$
{
//Model start
if (alt == 0 && engine_enabled == 1) {$\label{code:togstart}$
time_on_ground = time_on_ground + 1
};$\label{code:togend}$
alt_rate = choose(0 <= . && . <= 39 && . <= alt);$\label{code:altupdatestart}$
alt = alt - alt_rate;
radar_alt = choose (38 <= . - alt && . - alt <= 38);$\label{code:altupdateend}$
err_1 = choose((prev_err_1 == 1 => . == 0) && (. == 0 || . == 1));$\label{code:errstart}$
prev_err_1 = err_1;
err_2 = choose((prev_err_2 == 1 => . == 0) && (. == 0 || . == 1));
prev_err_2 = err_2;
if (alt == 0 && (err_1 == 1 || err_2 == 1)) { trans_td = 1; };$\label{code:errend}$
leg_err = 1400 <= alt && alt <= 1600;$\label{code:legdep}$
if perm_1 { cur_td_1 = perm_1_v } $\label{code:td1start}$
else if (leg_err == 1 || err_1 == 1) {
cur_td_1 = choose(. == 0 || . == 1)
} else { cur_td_1 = alt == 0; };$\label{code:td1end}$
if perm_2 { cur_td_2 = perm_2_v }$\label{code:td2start}$
else if (leg_err == 1 || err_2 == 1) {
cur_td_2 = choose(. == 0 || . == 1)
} else { cur_td_2 = alt == 0; };$\label{code:td2end}$
//Model end
//Controller loop start
...$\label{code:dots2}$
}
\end{lstlisting}
\caption{Model for verification of the Mars Polar Lander.}
\label{fig:mplmodel}
\end{figure}
\subsubsection{Formalization} The nondeterministic program in Figure~\ref{fig:mplmodel} formalizes these sources of error.
It provides inputs to the control software by setting the variables \lstinline{cur_td_1}, \lstinline{cur_td_2}, and \lstinline{radar_alt}.
It can be composed with the control software by inlining the code in Figure~\ref{fig:mplcode} on Lines \ref{code:dots1} and \ref{code:dots2}.
After composing with the control software, the resulting overall program could then be verified to prove the loop invariant on Lines \ref{code:loopinvmpl} and \ref{code:loopcondend} of Figure~\ref{fig:mplmodel} always holds.
We now describe each piece of the formal model in more detail.
\paragraph{Permanent Errors}
The block of code on Lines~\ref{code:permstart}-\ref{code:permend} of Figure~\ref{fig:mplmodel} models permanent errors, both false positive and false negative.
Each of the variables \lstinline{perm_1} and \lstinline{perm_2} is \lstinline{1} if its sensor has suffered a permanent failure, with the assumption being that neither of these two variables will be \lstinline{1} at the same time.
The variables \lstinline{perm_1_v} and \lstinline{perm_2_v} store the permanent error values, which are the constant values that each of the sensors will read when after suffering a permanent error.
\paragraph{Loop Condition}
The code on Lines~\ref{code:loopcondstart}-\ref{code:loopcondend} of Figure~\ref{fig:mplmodel} specifies the same loop condition as in Figure~\ref{fig:mplcode}, but also adds a loop invariant.
The invariant specifies that when the lander is off the ground, its engine is enabled.
Verifying this property would ensure that the software does not have the bug that caused the MPL to crash.
The invariant also specifies on Line~\ref{code:loopcondend} that the lander should spend less than 2 time steps on the ground with the engine on.
This is another constraint the MPL was designed to satisfy~\citep{mpl}.
However, the model in Figure~\ref{fig:mplmodel} admits transient false negative errors, which can violate this constraint in extreme cases.
Thus, on Line~\ref{code:loopcondend} of Figure~\ref{fig:mplmodel}, we assume this holds in the case where there are no transient false negatives.
\paragraph{Time on Ground}
The code on Lines~\ref{code:togstart}-\ref{code:togend} of Figure~\ref{fig:mplmodel} measures the amount of time the lander has spent on the ground with the engine on and stores it in \lstinline{time_on_ground}.
This is measured purely to evaluate the loop invariant and is not passed to the controller.
\paragraph{Altitude}
The code on Lines~\ref{code:altupdatestart}-\ref{code:altupdateend} of~\ref{fig:mplmodel} specifies how altitude changes and the error model of the radar altimeter.
It stores the rate of altitude change in \lstinline{alt_rate}, the new altitude in \lstinline{alt}, and the altimeter reading in \lstinline{radar_alt}.
We assume that the altitude changes by at most 39 meters and that the altimeter is accurate to within 38 meters.
The original motivation for including touchdown sensors on the MPL was that the radar altimeter is inaccurate below about 40 meters~\citep{mpl}.
This model is designed to conservatively capture this property while still ensuring that the condition on Line~\ref{code:healthstart} of Figure~\ref{fig:mplcode} triggers the indicator health check.
Furthermore, note that on line ~\ref{code:altinit} of Figure~\ref{fig:mplmodel} we have specified that the entry point to the program is when the lander is at an altitude of 8 kilometers.
\paragraph{Transient Errors}
The code on Lines~\ref{code:errstart}-\ref{code:errend} of Figure~\ref{fig:mplmodel} models transient errors.
The variables \lstinline{err_1} and \lstinline{err_2} are set to be \lstinline{1} if a transient error occurred for the first or second touchdown sensor, respectively.
The previous values of these variables are stored in \lstinline{prev_err_1} and \lstinline{prev_err_2}.
This code specifies that a transient error can occur for a sensor if it did not occur at the previous time step.
Furthermore, if a transient occurs after touchdown (i.e.\ a false negative), the code specifies that the variable \lstinline{trans_td} is set to \lstinline{1}.
\paragraph{Landing Leg Deployment}
To account for landing leg deployment errors, the model sets the flag \lstinline{leg_err} whenever the landing gear are deploying.
This occurs at about 1500 meters for the MPL~\cite{mpl}.
We have modeled a deployment window of 100 meters around this nominal value, so that landing leg deployment occurs between 1400 and 1600 meters.
\paragraph{Touchdown Indicators}
The code on Lines~\ref{code:td1start}-\ref{code:td1end} of Figure~\ref{fig:mplmodel} specifies how the first touchdown indication is generated from the various sources of error.
The result is stored in \lstinline{cur_td_1}.
If the indicator suffered a permanent error, then it returns its permanent error value.
Otherwise, during landing leg deployment or a transient error, it may output either \lstinline{0} or \lstinline{1}.
If none of these errors are present, then it indicates \lstinline{1} iff the lander has touched the surface (i.e.\ the altitude is 0 meters).
The code on Lines~\ref{code:td2start}-\ref{code:td2end} of Figure~\ref{fig:mplmodel} specifies how the second touchdown indication is generated and is entirely symmetric.
\begin{figure}
\begin{lstlisting}[basicstyle=\ttfamily\small]
// Model Initialization
...$\label{code:bpinline1}$
engine_enabled = 1;$\label{code:einit}$
while (engine_enabled == 1)$\label{code:mplbploopcond}$
{ $\Box$((alt > 0 => engine_enabled == 1) && $\label{code:mplbploopinv1}$
(trans_td == 0 => time_on_ground < 2)) }$\label{code:mplbploopinv2}$
{
// Model start
...$\label{code:bpinline2}$
// Model end
observe radar_alt;
observe cur_td_1;
observe cur_td_2;
infer $\Box$(alt == 0) {$\label{code:mplinfer}$
engine_enabled = 0
}
}
\end{lstlisting}
\caption{Implementation of the Mars Polar Lander with belief programming}
\label{fig:mplbp}
\end{figure}
\subsection{Belief Programming}
We now show how to use the model in Figure~\ref{fig:mplmodel} to construct a belief program to control the MPL.
We show a fragment of the belief program in Figure~\ref{fig:mplbp}.
This fragment can be completed by inlining the appropriate parts of Figure~\ref{fig:mplmodel} into Lines \ref{code:bpinline1} and \ref{code:bpinline2} of Figure~\ref{fig:mplbp}.
The belief program executes by observing each of the sensor readings generated from the model.
It then determines, on Line~\ref{code:mplinfer}, whether these are sufficient to guarantee the lander is on the ground.
If so, it shuts down the engine.
The belief program also modifies the loop invariant to have a $\Box$ modality, stating that it must be true in every environment (Lines \ref{code:mplbploopinv1} and \ref{code:mplbploopinv2}).
\subsection{Verification}
In this section, we will explain how to verify the loop invariant on Lines~\ref{code:mplbploopinv1}-\ref{code:mplbploopinv2} of Figure~\ref{fig:mplbp}.
We simplify the problem here by only considering the first condition on Line~\ref{code:mplbploopinv1}.
The remaining condition is considered in Appendix~\ref{sec:mplverification}.
\paragraph{Initialization Post-condition}
As specified by the rules in Figure~\ref{fig:logic}, we must show that the initialization code generates a post-condition that satisfies the loop invariant.
This post-condition can be written as $\Box\ncode{(engine\_enabled == 1)} \code{\&\&} \dots$,
where the $\Box$-proposition is generated by the code on Line~\ref{code:einit} of Figure~\ref{fig:mplbp}.
Now, we can apply the fact that
\[
\forall \sigma. \; \sigma \vDash \ncode{(engine\_enabled == 1)} \Rightarrow
\sigma \vDash \ncode{(alt > 0 => engine\_enabled == 1)}
\]
and Theorems \ref{thm:lifting} and \ref{thm:k} to see that
\[
\beta \vDash \Box\ncode{(engine\_enabled == 1)} \code{\&\&} \dots \Rightarrow
\beta \vDash \Box\ncode{(alt > 0 => engine\_enabled == 1)}
\]
\paragraph{Loop Body Post-condition}
According to the rules in Figure~\ref{fig:logic}, we must show that the loop body's post-condition implies the loop invariant.
We can summarize the post-condition as
\begin{align*}
& \Box\ncode{(engine\_enabled\_0 == 1)} \code{\&\&} \\
& \ncode{(}\Box\ncode{(alt == 0)} \code{\&\&} \Box\ncode{(engine\_enabled == 0)}\ncode{)} \code{||}\\
& \ncode{(}\Diamond\ncode{(alt != 0)} \code{\&\&} \Box\ncode{(engine\_enabled == engine\_enabled\_0)}\ncode{)}
\end{align*}
where the first line comes from the loop condition on Line~\ref{code:mplbploopcond} and the second and third line come from the infer statement on Line~\ref{code:mplinfer}.
We assume we have applied the standard rule for strongest-postcondition predicates and taken the disjunction of each branch of the \lstinline{infer}~\citep{floyd}.
We can now show the post-condition implies the invariant. The disjunction gives us two cases. In the first, we can assume that $\Box\ncode{(alt == 0)}$.
Now, we can apply the fact that by vacuous truth,
\[
\forall \sigma. \; \sigma \vDash \ncode{(alt == 0)} \Rightarrow
\sigma \vDash \ncode{(alt > 0 => engine\_enabled == 1)}
\]
and Theorems \ref{thm:lifting} and \ref{thm:k} to see that
\[
\beta \vDash \Box\ncode{(alt == 0)} \Rightarrow
\beta \vDash \Box\ncode{(alt > 0 => engine\_enabled == 1)}
\]
In the second case, we follow a similar logic to the initialization post-condition argument above, with the additional premise that $\Box\ncode{(engine\_enabled == engine\_enabled\_0)}$.
\section{Conclusion}
\label{sec:conclusion}
In this paper, we presented belief programming and Epistemic Hoare Logic.
Belief programming enables developers to write programs that can be directly executed to give state estimators that are derived from environment models. Epistemic Hoare Logic enables developers to reason about belief programs.
We discussed both by reference to the BLIMP{} language, with belief programming described by BLIMP{}'s semantics and Epistemic Hoare Logic operating over BLIMP{} statements.
We determined that belief programming is feasible by evaluating our BLIMP implementation, CBLIMP.
Taken together, this work lays new foundations for soundly reasoning about the behavior of software that executes in partially-observable environments.
\section{Example}
\label{sec:example}
\lstset{xleftmargin=2em,numbers=left}
\begin{wrapfigure}[16]{r}{0.65\textwidth}
\vspace{-2.5mm}
\begin{lstlisting}
cmd = 0
t_max = 10000; t = 0;
while (t < t_max)
{
input obs;$\label{code:readobs}$
// Controller loop start
if (obs < 475) { cmd = 50 } $\label{code:climb}$
else if (obs > 525) { cmd = -50 } $\label{code:descend}$
else { cmd = 0 }
// Controller loop end
t = t + 1
}
\end{lstlisting}
\vspace{-5mm}
\caption{An altitude controller for a UAV.}
\label{fig:dronecode}
\end{wrapfigure}
In this section we show how a developer uses belief programming to implement a controller for a UAV.
The controller's objective is to maintain the UAV's altitude at 500 feet above the ground.
While the UAV can precisely control its altitude, it has to contend with measurement error from its altitude sensing equipment and wind gusts that can blow it off course.
The listing in Figure~\ref{fig:dronecode} shows how a developer can write such a controller in a traditional programming language.
At every time step, up to a maximum of $\ncode{10000}$ steps, the controller receives an altitude observation \lstinline{obs} (Line~\ref{code:readobs}).
We assume the observation comes from a sensor which has some inherent measurement error, so that the value stored in \lstinline{obs} is not precisely the UAV's true altitude.
If the observation is sufficiently low, the controller issues a command to climb by 50 feet (Line~\ref{code:climb}).
Conversely, if the observation is sufficiently high, the controller issues a command to descend (Line~\ref{code:descend}).
The conditions on Lines \ref{code:climb} and \ref{code:descend} form a coarse-grained state estimator that determines if the UAV is too high, too low, or at an acceptable altitude.
We assume the command is stored in \lstinline{cmd}, and that there is an external process that reads \lstinline{cmd} and modifies the UAV's altitude by exactly \lstinline{cmd}.
\begin{wrapfigure}[21]{l}{0.55\textwidth}
\begin{lstlisting}
alt = 500; cmd = 0;
t_max = 10000; t = 0;
while (t < t_max)
{ 450 <= alt && alt <= 550 }$\label{code:loopinv}$
{
alt = choose(alt - 25 <= . && $\label{code:choosealt}$
. <= alt + 25);
obs = choose(alt - 25 <= . && $\label{code:chooseobs}$
. <= alt + 25);
// Controller loop start
...$\label{code:droneinline}$
// Controller loop end
alt = alt + cmd$\label{code:updatealt}$
t = t + 1;
}
\end{lstlisting}
\caption{Environment model for the UAV altitude controller.}
\label{fig:dronemodel}
\end{wrapfigure}
The developer needs to ensure the UAV maintains a consistent altitude for safety reasons.
We will assume that the developer wants to provide this assurance using formal verification, which requires an environment model of how the true and observed altitude are related.
The developer can write such a model as a program that specifies the set of environments the UAV may be in, including the set of values the UAV's true and observed altitude may take on.
The listing in Figure~\ref{fig:dronemodel} shows how a developer can write such a model.
The model is composed with the controller by inlining the annotated lines in Figure~\ref{fig:dronecode} into Line~\ref{code:droneinline} of Figure~\ref{fig:dronemodel}.
Note that this replaces the input \lstinline[breaklines=true]{obs} in Figure~\ref{fig:dronecode} with the value of \lstinline{obs} specified by the model, and we have added a \emph{loop invariant} on Line~\ref{code:loopinv}.
The modeling language includes a nondeterministic assignment operator \lstinline{choose}, which takes a predicate and nondeterministically chooses a value that satisfies that predicate.
The placeholder \lstinline{.} stands for the new value of the variable, so that \lstinline[breaklines=true]{x = choose(x - 1 <= . && . <= x + 1)} picks a new value for \lstinline{x} that is within a distance of \lstinline{1} of its previous value.
At every time step, the model chooses the current altitude \lstinline{alt} from a value within a distance of \lstinline{25} of the previous altitude (Line~\ref{code:choosealt}).
This models a wind gust causing a change of up to 25 feet of altitude per step.
The model then chooses the observation \lstinline{obs} from within a distance of \lstinline{25} of the true altitude (Line~\ref{code:chooseobs}).
This models the altitude instrumentation as having a measurement error of up to 25 feet.
After the controller runs, the model alters the altitude by adding to it the resulting command \lstinline{cmd} (Line~\ref{code:updatealt}).
The condition the developer needs to ensure is given by the loop invariant on Line~\ref{code:loopinv}: \lstinline[breaklines=true]{450 <= alt && alt <= 550}.
This means that the UAV maintains its target altitude of 500 feet within an error margin of 50 feet.
The developer can prove that the composition of the environment model and the controller satisfy this condition using classical verification techniques.
\subsection{Belief Programming}
\begin{figure}
\begin{lstlisting}
alt = 500; cmd = 0; t_max = 10000; t = 0;
while (t < t_max)
{ $\Box$(450 <= alt && alt <= 550) }$\label{code:bploopinv}$
{
alt = choose(alt - 25 <= . && . <= alt + 25);$\label{code:bpchoosei}$
obs = choose(alt - 25 <= . && . <= alt + 25);$\label{code:bpchooseii}$
observe obs;$\label{code:bpobserve}$
infer $\Diamond$(alt < 450) { cmd = 50 } $\label{code:inferstart}$
else infer $\Diamond$(alt > 550) { cmd = -50 }
else { cmd = 0 };$\label{code:inferend}$
alt = alt + cmd;$\label{code:altasgn}$
t = t + 1$\label{code:bodyend}$
}
\end{lstlisting}
\caption{Implementation of the UAV controller using belief programming in BLIMP{}.}
\label{fig:dronebp}
\end{figure}
We will now explain, alternatively, how the developer implements this program using belief programming.
The listing in Figure~\ref{fig:dronebp} shows the code to implement the controller in our belief programming language BLIMP{}.
As this program executes, it maintains a {\em belief state} of the set of possible environments it could be in.
Instead of the conditions over concrete observations in Figure~\ref{fig:dronecode}, the code in Figure~\ref{fig:dronebp} uses conditions over belief states to determine control actions.
To explain how belief programming operators work, we will walk through the execution of the first iteration of the while loop in Figure~\ref{fig:dronebp}.
At the start of the loop, the belief state contains a single environment that has $\ncode{alt} = 500$, $\ncode{cmd} = 0$, $\ncode{t\_max} = 10000$, and $\ncode{t} = 0$.
\paragraph{Choose Statements} The \lstinline{choose} statements on Lines \ref{code:bpchoosei} and \ref{code:bpchooseii} expand the belief state to include all possible environments the nondeterminism could generate.
After the first assignment to \lstinline{alt} on Line~\ref{code:bpchoosei}, the belief state contains all environments such that $\ncode{alt} \in [475, 525]$.
After the assignment to \lstinline{obs} on Line~\ref{code:bpchooseii}, the belief state contains all environments such that $\ncode{obs} \in [450, 550]$, with the additional constraint that the distance between \lstinline{alt} and \lstinline{obs} is less than or equal to $25$.
This means that, for example, the belief state does not contain the environment where $\ncode{alt} = 500$ and $\ncode{obs} = 550$.
\paragraph{Observe Statements} The \lstinline{observe} statement on Line~\ref{code:bpobserve} implicitly receives an input that is the observed altitude \lstinline{obs}.
It updates the belief state to contain only environments that are consistent with that observed altitude.
For example, if the program receives the value $525$, then \lstinline{observe} modifies the belief state
to only contain environments where $\ncode{obs} = 525$ and, correspondingly, where $\ncode{alt} \in [500, 525]$.
\paragraph{Infer Statements}
The \lstinline{infer} statements on Lines~\ref{code:inferstart}-\ref{code:inferend} branch based on belief state conditions.
Conditions use the $\Box$ and $\Diamond$ {\em modal operators} to quantify over environments in the belief state, with $\Box$ meaning ``for all environments in the belief state'' and $\Diamond$ meaning ``there exists an environment in the belief state''.
The condition \lstinline{$\Diamond$(alt < 450)} means that there is an environment in the belief state such that \lstinline{alt} is smaller than $450$.
Similarly \lstinline{$\Diamond$(alt > 550)} means that there is an environment such that \lstinline{alt} is larger than $550$.
Compared to the state estimator on Lines \ref{code:climb} and \ref{code:descend} of Figure~\ref{fig:dronecode}, the infer statements provide more intuition for why the controller was constructed this way.
If one of these conditions is true, then it is possible for the true environment to be outside of the desired range of $500 \pm 50$ feet, and immediate action is needed to correct the situation.
Since, assuming the example observation of $525$, our belief state contains only the environments where $\ncode{alt} \in [500, 525]$, neither of these conditions is true.
Thus, the \lstinline{cmd = 0} branch of the infer statements is executed.
This constrains every environment in the belief state to include \lstinline{cmd = 0}.
Note that under the assumption that the observation localizes $\ncode{alt}$ to within 25 feet, at most one condition of the infer statements will be true in any given belief state.
\paragraph{Assignments} The final line of the loop, Line~\ref{code:altasgn}, updates \lstinline{alt} to be its previous value plus the value of \lstinline{cmd} in every environment in the belief state.
Because the belief state contains the environments such that $\ncode{cmd} = 0$ and $\ncode{alt} \in [500, 525]$, after the update the belief state contains all environments such that $\ncode{alt} \in [500, 525]$.
The loop invariant on Line~\ref{code:bploopinv} states that any environment in the belief state must have a value for $\ncode{alt}$ in the range $[450, 550]$.
Because our belief state constrains $\ncode{alt}$ to the smaller range $[500, 525]$, our belief state satisfies the invariant.
\subsection{Reasoning with Epistemic Hoare Logic}
The previous section describes a single execution of the belief program given concrete observations, but in general developers need to reason about all potential executions given any observations that are allowed under the environment model.
In this section, we show how a developer can use Epistemic Hoare Logic to reason about the potential executions of the belief program in Figure~\ref{fig:dronebp}.
We first present a small example that showcases how Epistemic Hoare Logic can be used to reason about a single statement in the program.
We then explain at a high level how similar reasoning can be used to verify the program maintains its altitude within limits, with the details in Appendix~\ref{sec:exampleverification}.
\subsubsection{Small Example}
\newcommand{s^*}{s^*}
In this section, we consider proving a simple property about the statement on Line~\ref{code:bpchoosei} of Figure~\ref{fig:dronebp} that specifies altitude updates.
We will assume, in accordance with the loop invariant, that the altitude immediately before executing this statement is larger than 450.
We will then show that after executing this statement, the altitude is larger than 425.
This property is expressed by the following judgment in Epistemic Hoare Logic:
\[
\htriplepc{D}{\Box\ncode{(450 <= alt)}}
{\ncode{alt = choose(alt - 25 <= . \&\& . <= alt + 25)}}
{\Box\ncode{(425 <= alt)}}
\]
The context $D$ indicates that the judgement applies when the statement is executed under deterministic control flow.
The pre-condition, $\Box\ncode{(450 <= alt)}$, states that in any environment in the belief state, the variable \texttt{alt} is at least \texttt{450}.
The post-condition, $\Box\ncode{(425 <= alt)}$, states that in any environment in the belief state, the variable \texttt{alt} is at least \texttt{425}.
\paragraph{Epistemic Hoare Logic Rules.} To prove this deduction, we first instantiate the Epistemic Hoare Logic rule for \texttt{choose} statements that we have designed.
We present the general rule in Figure~\ref{fig:logic} of Section~\ref{sec:logic}.
By default, the rule allows for nondeterministic control flow, but we can specialize it to the $D$ context via a subtyping rule.
Instantiating the choose rule and subtyping with our original pre-condition yields the following judgement, using the notation $s^*$ for the statement on Line~\ref{code:bpchoosei}:
\[
\htriplepc{D}{\Box\ncode{(450 <= alt)}}
{s^*}
{\Box\ncode{(450 <= a)} \code{\&\&}
\Box\ncode{(a - 25 <= alt \&\& alt <= a + 25)}}
\]
The rule produces a post-condition by first renaming all instances of the assigned variable \texttt{alt} with a fresh variable \texttt{a} in the pre-condition.
It then conjuncts this with the \texttt{choose} statement's predicate under a $\Box$ with \texttt{alt} replaced with \texttt{a} and the placeholder \lstinline{.} replaced with \texttt{alt}.
\paragraph{Implications.}
To rewrite the post-condition into the desired form, we prove that another predicate is implied by the post-condition and use the rule of consequence (formalized in general in Figure~\ref{fig:logic}) to replace the post-condition with the new predicate.
First, we use the general principle that the $\Box$ operator commutes with \texttt{\&\&}.
We have formalized this principle as Theorem~\ref{thm:box-and} in Appendix~\ref{sec:proptheorems}.
This gives us the following judgment:
\[
\htriplepc{D}{\Box\ncode{(450 <= alt)}}
{s^*}
{\Box\ncode{(450 <= a \&\& a - 25 <= alt \&\& alt <= a + 25)}}
\]
We then use the principle of lifting theorems about environments to theorems about belief states.
This states that if we have an implication over environments, we can wrap the premise and conclusion of this implication with $\Box$, and obtain an implication over belief states.
We have formalized this principle as Theorems \ref{thm:lifting} and \ref{thm:k} in Appendix~\ref{sec:proptheorems}.
In this example, we use the fact that if the environment satisfies \lstinline[breaklines=true]{450 <= a && a - 25 <= alt && alt <= a + 25}, then it also satisfies \lstinline[breaklines=true]{425 <= alt}.
The principle of lifting says that as a result,
if the belief state satisfies \lstinline[breaklines=true]{$\Box$(450 <= a && a - 25 <= alt && alt <= a + 25)}, then it also satisfies \lstinline[breaklines=true]{$\Box$(425 <= alt)}.
Applying the rule of consequence gives us the original judgment we set out to prove:
\[
\htriplepc{D}{\Box\ncode{(450 <= alt)}}
{s^*}
{\Box\ncode{(425 <= alt)}}
\]
\subsubsection{Verification}
The developer would like to ensure that the loop body in Figure~\ref{fig:dronebp} maintains an altitude of 500 feet within a tolerance of 50 feet.
This means that the loop body must preserve its invariant.
This corresponds to the Epistemic Hoare Logic deduction
\[
\htriplepc{D}{\Box\ncode{(450 <= alt \&\& alt <= 550)}}{s}{\Box\ncode{(450 <= alt \&\& alt <= 550)}}
\]
where the pre- and post-conditions are both equal to the loop invariant on Line~\ref{code:bploopinv} of Figure~\ref{fig:dronebp}.
The program $s$ is the loop body on Lines~\ref{code:bpchoosei}-\ref{code:bodyend} of Figure~\ref{fig:dronebp}.
Thus, this deduction states that if the loop body starts in a belief state satisfying the loop invariant, it produces a belief state that also satisfies the loop invariant.
We consider the case of deterministic control flow, as signified with the $D$ context.
This is appropriate because the loop condition, \texttt{t < t\_max}, is consistently either true or false across all environments in the belief state.
The high-level procedure is to assume the loop invariant as the pre-condition for the loop body.
We then use the logic's rules to derive an appropriate post-condition for the loop body, and finally prove that the post condition implies the loop invariant using the reasoning principles in Section~\ref{sec:proptheorems}.
The details of this process are in Appendix~\ref{sec:exampleverification}.
\section{Verification of the UAV Example}
\label{sec:exampleverification}
This section is structured as follows.
Section~\ref{sec:bpverifpost} explains how propositions are generated from belief programs, with an emphasis on how the rules of Epistemic Hoare logic differ from the rules of classical Hoare logic.
Section~\ref{sec:bpverifinv} gives examples of proving an implication with modal propositions.
In each section, we illustrate concepts in the context of verifying the invariant preservation property of the loop body in Figure~\ref{fig:dronebp}.
\subsubsection{Post-condition}
\label{sec:bpverifpost}
\begin{figure}
\begin{lstlisting}
$\exists$ y.
$\Box$(450 <= alt_0 && alt_0 <= 550) && $\label{code:postloopinv}$
$\Box$(alt_0 - 25 <= alt_1 && alt_1 <= x_0 + 25) &&$\label{code:altchoose}$
$\Box$(alt_1 - 25 <= y && y <= alt_1 + 25) &&$\label{code:obschoose}$
$\Box$(obs == y) &&$\label{code:obsprop}$
( ($\Diamond$(alt_1 < 450) && $\Box$(cmd == 50)) ||$\label{code:disstart}$
(!$\Diamond$(alt_1 < 450) && $\Diamond$(alt_1 > 550) && $\Box$(cmd == -50)) ||
(!$\Diamond$(alt_1 < 450) && !$\Diamond$(alt_1 > 550) && $\Box$(cmd == 0))$\label{code:disend}$
) &&
$\Box$(alt == alt_1 + cmd)$\label{code:altprop}$
\end{lstlisting}
\caption{Post-condition of the belief program at the end of the loop.}
\label{fig:dronepost}
\end{figure}
Epistemic Hoare logic is most closely related to strongest-postcondition logics such as~\citet{floyd} that generate a post-condition given a pre-condition and a program.
This means there is a deduction
\[
\htriplepc{D}{\Box\ncode{(450 <= alt \&\& alt <= 550)}}{s}{\prop_\exists'}
\]
where the post-condition $\prop_\exists'$ is generated fairly directly by the program $s$.
Figure~\ref{fig:dronepost} shows such an $\prop_\exists'$, and here we explain how it corresponds to $s$, the loop body from Figure~\ref{fig:dronebp}.
\paragraph{Loop Invariant} The proposition on Line~\ref{code:postloopinv} of Figure~\ref{fig:dronepost} corresponds to the loop invariant on Line~\ref{code:bploopinv} of Figure~\ref{fig:dronebp} that we assumed as our pre-condition.
Because the variable $\ncode{alt}$ is reassigned later in the loop body, the logic renames it to $\ncode{alt\_0}$, a fresh variable that represents the previous value of $\ncode{alt}$ at the start of the loop.
Note that this differs from the standard approach to name conflicts, which uses existential quantification~\citep{floyd}.
This is because Epistemic Hoare logic needs to preserve the original quantification with $\Box$ and/or $\Diamond$ when referring to variables in the environment.
\paragraph{Choose Statements} The choose statement on Line~\ref{code:bpchoosei} of Figure~\ref{fig:dronebp} generates the proposition on Line~\ref{code:altchoose} of Figure~\ref{fig:dronepost}.
This proposition states that the previous altitude $\ncode{alt\_0}$ is within a distance of $25$ of the new altitude $\ncode{alt\_1}$, and is generated from the choose statement's proposition by replacing the placeholder $\code{.}$ with the new altitude.
The updated altitude is renamed to the fresh variable $\ncode{alt\_1}$ because of the reassignment of $\ncode{alt}$ on Line~\ref{code:altasgn} of Figure~\ref{fig:dronebp}.
The choose statement for $\ncode{obs}$ on Line~\ref{code:bpchooseii} of Figure~\ref{fig:dronebp} similarly generates the proposition on Line~\ref{code:obschoose} of Figure~\ref{fig:dronepost}, though as we explain next, the new value of $\ncode{obs}$ is renamed to the existentially quantified variable $\ncode{y}$.
\paragraph{Observations} The observation on Line~\ref{code:bpobserve} of Figure~\ref{fig:dronebp} generates the existentially quantified variable $\ncode{y}$.
This variable stands for the input observed value of $\ncode{obs}$, and thus must satisfy any constraints that $\ncode{obs}$ was originally under.
In this case, that means that $\ncode{y}$ must be within a distance of $25$ of the altitude $\ncode{alt\_1}$.
Furthermore, because $\ncode{obs}$ is being observed, we know that every environment must have it bound to its true value $\ncode{y}$.
Thus, the observation generates the proposition on Line~\ref{code:obsprop} of Figure~\ref{fig:dronepost}.
\paragraph{Infer} The infer statement on Lines~\ref{code:inferstart}-\ref{code:inferend} of Figure~\ref{fig:dronebp} generates the disjunction on Lines~\ref{code:disstart}-\ref{code:disend} of Figure~\ref{fig:dronepost}.
Each term in the disjunction corresponds to one of the branches of the infer statement.
Each term itself is a conjunction of the predicate that causes the branch to be taken and a proposition describing the actions of the branch.
In every case, the predicate is a combination of (possibly negated) $\Diamond$-propositions and the branch specifies that the value of $\ncode{cmd}$ is some constant in every environment.
\paragraph{Assignment}
The assignment on Line~\ref{code:altasgn} of Figure~\ref{fig:dronebp} generates the proposition on Line~\ref{code:altprop} of Figure~\ref{fig:dronepost}. This proposition specifies that in any environment in the belief state, the new altitude $\ncode{alt}$ is equal to the previous altitude $\ncode{alt\_1}$ plus the command $\ncode{cmd}$.
\subsubsection{Modal Implications}
\label{sec:bpverifinv}
In this section, we show how implications of ordinary propositions can be lifted to implications of modal propositions.
We further demonstrate how this can be used to show that the invariant $\prop_\exists'$ presented in Figure~\ref{fig:dronepost} and derived in the Section~\ref{sec:bpverifpost} implies the loop invariant in Figure~\ref{fig:dronebp}.
This completes the proof that the program in Figure~\ref{fig:dronebp} preserves its loop invariant.
The proof that $\prop_\exists'$ implies the loop invariant is structured as follows.
The first step is to treat term of the disjunction on Lines~\ref{code:disstart}-\ref{code:disend} of Figure~\ref{fig:dronepost} as a separate case.
For clarity, we only discuss the first case on Line~\ref{code:disstart} in this section.
In this case, in addition to the post-condition, we can assume the belief state satisfies the proposition $\Diamond\ncode{(alt\_1 < 450)} \code{\&\&} \Box\ncode{(cmd == 50)}$. Our approach will be to a) show that this implies a $\Diamond$-proposition about only $\ncode{y}$ b) that we can convert this $\Diamond$-proposition to a $\Box$-proposition, and c) that the resulting $\Box$-proposition implies the loop invariant.
\paragraph{Implied $\Diamond$-Proposition} The first technique for lifting proposition implications to modal implications is that if for any environment $P \Rightarrow Q$ where $P$ and $Q$ are propositions over environments, then in any belief state $\Diamond P \Rightarrow \Diamond Q$.
In this case, we take $P$ to be the proposition $\ncode{alt\_1 < 450}$ and $Q$ to be the proposition $\ncode{alt\_1 < y - 25 || y <= 475}$.
This means we can assume that \lstinline[breaklines=true]{$\Diamond$(alt_1 < y - 25 || y <= 475)} is true of our belief state.
Furthermore, by the proposition on Line~\ref{code:obschoose}, we know that every environment in the belief state satisfies the negation of $\ncode{alt\_1 < y - 25}$, meaning that if an environment satisfies $Q$ it must be because it satisfies $\ncode{y <= 475}$.
Thus, we can assume that $\Diamond\ncode{(y <= 475)}$.
\paragraph{Conversion to $\Box$-proposition} Note that $\Box$ and $\Diamond$ quantify over environments, and the proposition $\ncode{y <= 475}$ depends only on the quantified variable $\ncode{y}$ and not on any variables in the environment.
This means that if the proposition is true in some environment, it must be true in every environment.
Thus, we can assume $\Box\ncode{(y <= 475)}$
\paragraph{Implied $\Box$-Proposition} Another technique for lifting proposition implications to modal implications is that if for any environment $P \Rightarrow Q$, then for any belief state $\Box P \Rightarrow \Box Q$.
We now apply this to the propositions in Figure~\ref{fig:dronepost} and the assumption $\Box\ncode{(y <= 475)}$.
In this case, we take $P$ to be
\[
\ncode{425 <= alt\_1 \&\& alt\_1 <= 575 \&\& alt\_1 - 25 <= y \&\& y <= alt\_1 + 25 \&\& y <= 475}
\]
and $Q$ to be $\ncode{475 <= alt\_1 + 50 \&\& alt\_1 + 50 <= 550}$. By similar logic, we can see that the proposition $\Box\ncode{(y <= 475)}$ combined with the propositions on Lines ~\ref{code:postloopinv}-\ref{code:obschoose} implies $\Box P$ and that $\Box Q$ combined with the proposition on Line~\ref{code:altasgn} implies the loop invariant.
\section{Future Work}
\subsection{Implementation Efficiency}
There is an open question of how to design an efficient implementation for the belief programming runtime.
CBLIMP directly implements the semantics of Figure~\ref{fig:statementsemantics} using an exhaustive representation of belief states.
This implementation scales poorly with the number of variables in the program, and we discuss here more efficient potential runtime implementation approaches.
\paragraph{Runtime SMT} One approach could symbolically execute BLIMP{} constructs to collect constraints that an SMT solver would use to evaluate modal propositions.
This approach would make use of the enhanced performance of SMT solvers compared to an exhaustive approach, and bears similarities to other languages that deploy solvers at runtime (e.g.~\cite{jeeves, planb}).
\paragraph{Restricted belief states} Known efficient implementations exist when the belief states admitted by the language are more restricted than the full powerset of environments.
Examples of restricted classes of belief states studied in the literature include ellipsoids~\cite{ellipsoids} and polytopes~\cite{polytopes}.
\paragraph{Synthesis} Using the semantics of Figure~\ref{fig:statementsemantics} as a specification, a variety of synthesis tools~\cite{cegis, vsa, dedsynth} exist that could potentially generate more efficient implementations than naive enumeration.
The goal would be to translate \texttt{infer} statements to ordinary \texttt{if} statements, using synthesis to construct a predicate for the \texttt{if} statement that is a function of the observed values and is semantically equivalent to the \texttt{infer} statement's predicate.
\subsection{Logic Automation}
The logic in Figure~\ref{fig:logic} and the theorems in Section~\ref{sec:proptheorems} enable sound, manual reasoning about the behavior of belief programs.
As with many program logics, the gap from manual to automated reasoning is the need for automated techniques for invariant inference and implication checking.
\paragraph{Invariant Inference.} As in traditional program logics, a while loop requires a loop invariant.
Although propositions in the Epistemic Hoare Logic include modalities, classic approaches such as template-based invariant inference may be directly applicable~\citep{houdini} via templates that include modalities.
An additional distinct difference from many traditional program logics is that the rules for if statements require developers to manually determine a suitable post-condition or, in other words, provide an invariant.
Here too template-based techniques may be directly applicable.
In either case, there may also be new opportunities for analysis-based invariant inference techniques that account for modalities.
\paragraph{Discharging Implications.}
A classic approach to discharge implications that appear in the premises of Hoare logic rules is to employ an automatic theorem prover such as Z3~\citep{z3}.
To apply this approach to Epistemic Hoare Logic, one would need to contend with the modalities in propositions.
Recent work holds out the promise of automated reasoning techniques for modal implications via a reduction to SMT~\citep{modalsmt, modalsat}.
\section{Implementation}
\label{sec:impl}
We have demonstrated the feasibility of a belief programming implementation that directly represents the belief state with a set.
Specifically, we wrote the implementation in C using a hash set data structure to represent beliefs.
Called CBLIMP, our implementation is a shallow embedding of BLIMP into C; i.e.\ it is a C library that implements each of the core BLIMP primitives as a C function.
CBLIMP's \texttt{observe} function takes, in addition to the current belief state and the variable to be observed, a parameter that is the observed value of the variable.
CBLIMP's \texttt{infer} function returns a boolean to be used as a branching condition by the C program, and
CBLIMP's \texttt{if} function takes as parameters callbacks for each branch that execute on modified belief states.
All regular C variables in a CBLIMP program can be considered to be deterministic with respect to the belief state (i.e.\ they have the same value in all environments in the belief state).
CBLIMP includes functions that extend BLIMP primitives to simulate the true environment via random sampling.
The simulated \texttt{observe} function presents a different interface: instead of taking the observed value as a parameter, it uses the value from the simulated true environment.
We enforce that environments are finite by augmenting \texttt{choose} statements to include a range that the newly assigned variable must belong to.
For example, as part of the MPL model, Line~\ref{code:errstart} of Figure~\ref{fig:mplmodel} gives the choose statement
\begin{center}
\texttt{err\_1 = choose((prev\_err\_1 == 1 => . == 0) \&\& (. == 0 || . == 1));}
\end{center}
This specifies that \texttt{err\_1} is a boolean value (i.e. it is either 0 or 1) and that whenever \texttt{prev\_err\_1} is 1, \texttt{err\_1} must be 0.
In our implementation, we have alternatively specified this as
\begin{center}
\texttt{err\_1 = choose(0,1, prev\_err\_1 == 1 => . == 0);}
\end{center}
where the arguments \texttt{0} and \texttt{1} of the \texttt{choose} statement are the lower and upper bounds of the range that \texttt{err\_1} must belong to. Note that these statements are equivalent; whereas in the original statement we specified that \texttt{err\_1} is a boolean using the choose statement's proposition, in the new statement we have instead specified this using the range bounds.
\paragraph{Research Question.} We evaluated CBLIMP to answer the question: does the direct implementation achieve practical performance given the latency requirements of the domain?
For the UAV example, a step latency of 1 second is required to match the latency of common GPS receivers~\cite{sparkfungps}.
For the MPL example, a latency of 10ms is required~\cite{mpl}.
\paragraph{Benchmarks.} We used the UAV example with a 100-step time horizon and the MPL example as benchmarks. The code is nearly identical to that in the paper, with the following changes:
\begin{itemize}
\item The MPL benchmark includes an additional intermediate variable that captures the error between the true altitude and the radar altitude.
\item We manually determined bounds for every variable specified by a choose statement to facilitate use of our implementation's augmented choose statements.
\end{itemize}
We also implemented another version of MPL that uses a different \emph{grid size}.
Note that the original version of MPL defines a uniform discretization grid for both true and radar altitude at a resolution of 1 meter.
We modified this to be a non-uniform grid on both true and observed altitude with higher-altitude grid cells exponentially larger than lower-altitude grid cells.
We call this benchmark "MPL-Exp". We have provided its BLIMP{} source code and verified it in Appendix~\ref{sec:mplexp}.
\paragraph{Methodology.} We ran each benchmark 5 times and recorded the mean and standard deviation of the time to execute a single iteration of the main while loop.
We also recorded the maximum time taken by an iteration across all runs.
To construct observations to send as inputs to the belief program, we simulated the true environment alongside the belief program's execution, sampling the new true environment uniformly at random when we encountered a choose statement.
While the latency measurements include both the time to update the belief state and run the simulation, we expect them to be dominated by belief state operations.
\begin{wraptable}{l}{0.5\textwidth}
\caption{Results of performance experiments}
\label{tab:results}
\begin{tabular}{ | l | r | r | }
\hline
Benchmark & Mean +- Std. Dev. & Maximum \\ \hline
UAV & 2.49 +- 0.37 ms & 4.41 ms \\
MPL & 2.45 +- 1.42 s & 11.9 s \\
MPL-Exp & 0.76 +- 0.56 ms & 2.37 ms \\\hline
\end{tabular}
\end{wraptable}
\paragraph{Results.} Table~\ref{tab:results} summarizes the step latency for each benchmark.
We can see that with the UAV benchmark, the direct implementation of belief programming is practical in the sense that the step latency is well under the 1s threshold.
However, with the MPL example, the direct implementation is not practical as-is.
The latency requirement of 10ms is about 1000x faster than the worst-case latency of our implementation.
By contrast, our modified "MPL-Exp" benchmark is practical for the MPL problem because its worst-case latency is considerably below the 10ms threshold.
\paragraph{Threats to validity.}
We ran these benchmarks on a 2017 MacBook Pro with an i7-7920HQ CPU at 3.10GHz and 16 GB of 2133 MHz DDR3.
Both of our benchmarks operate in domains that necessitate embedded computers that are less powerful.
The UAV example enjoys a comfortable margin over the required latency; it is likely that the processors common on larger drones could meet the latency requirements.
The MPL-Exp benchmark would typically be run on a small embedded processor, both for reliability benefits and because such processors are typically the ones hardened against cosmic radiation.
Although we expect such a processor to be slower than a modern laptop computer, with a comfortable 5x slowdown margin until it violates the latency requirement, we speculate that this benchmark could meet the requirement with standard performance engineering techniques.
\section{Introduction}
Computer systems are increasingly deployed in {\em partially observable} environments in which the system cannot exactly determine the environment's true state~\cite{aima, pomdp}.
For example, the software that controls an uncrewed aerial vehicle (UAV) cannot exactly determine the vehicle's true altitude above the ground.
Instead, the vehicle's software receives a measurement from a GPS altimeter that estimates the vehicle's altitude.
This measurement or {\em observation} reveals only partial information about the environment's true state, such as that the UAV's true altitude is within 25 feet of the reported measurement. The primary challenge for a system deployed in such an environment is therefore that it must leverage the partial information provided by an observation to meet its goals, which, in contrast, are typically expressed in terms of the environment's true state.
Consider, for example, a UAV tasked with avoiding a collision with the ground.
Its controller software will include a {\em state estimator} that must infer if it is possible that the vehicle may soon have an altitude of 0 given only estimated altitude measurements.
If the vehicle is indeed at risk, then the controller must take action to ensure the vehicle's altitude stays strictly positive.
However, the discrepancy between the measurements and the vehicle's true altitude introduces the risk that the state estimator's inferences may indicate a strictly positive altitude when the true altitude is in fact~0.
Moreover, the controller's reasoning must soundly work with the state estimator's inferences to intervene whenever the true altitude is dangerously near 0.
In safety-critical domains that desire formal guarantees, such as robotics and vehicle navigation, the best-available approach to formally verify the system is to
1) formally specify an {\em environment model}: the specific relationship between an observation and the system's true state;
2) implement the state estimator and verify the correctness of its inferences in relation to the environment model; and
3) implement the remainder of the controller and verify that the composition of the environment model and controller meets the system's requirements.
\subsection{Belief Programming}
In this paper, we present {\em belief programming}, a programming methodology in which the developer writes a program for the controller that includes a specification of the environment model. From that specification, the program runtime automatically provides a state estimator, eliminating the need to manually implement the state estimator and verify its behavior against the environment~model.
To instantiate the concepts of this programming methodology, we present a new language, BeLief IMP (BLIMP{}), a variant of the pedagogical language IMP~\cite{winskel}. BLIMP{} provides first-class abstractions for environment modeling, for observations, and to interface with the automatically generated state estimator.
\paragraph{Environment Model} The belief programming methodology extends IMP with a nondeterministic choice statement, $\ncode{x = choose(}p\ncode{)}$, that nondeterministically updates the program variable \texttt{x} to a value that satisfies the predicate $p$.
Unlike a traditional {\tt choose} statement~\cite{back}, the value of \texttt{x} is not immediately observable to the controller.
Instead the semantics of \texttt{x} is the set of all possible values that satisfy $p$.
The programming methodology permits such nondeterministic values to be composed with additional computation to produce a jointly nondeterministic and unobserved
set of program variables whose potential values correspond to the partially observable values of the system's physical state.
\paragraph{Observations}
To reveal the true value of an unobserved program variable, the controller must explicitly perform an observation.
Belief programming extends IMP with an observation statement, \texttt{observe y}, that makes the value of an unobserved variable \texttt{y} visible to the program. If, for example, \texttt{y} is a measurement that is derived from another unobservable value \texttt{x}, such as an altitude measurement derived from the UAV's true altitude, then the true value of \texttt{y} reveals partial information about the true value of \texttt{x}.
\paragraph{State Estimation}
The belief programming runtime system dynamically maintains a {\em belief state} that captures the set of all possible values of all unobserved variables.
The belief programming methodology also extends IMP with an inference statement, $\lcode{infer} \prop_\Box$, that computes a boolean inference over the program's belief state
and enables the controller to guide its actions given the validity of a proposition.
For example, if $\Diamond \ncode{(altitude < 1)}$ is true, which states that is possible that the UAV's altitude is less than {\tt 1} foot, then the controller can intervene to avoid a collision.
Together our abstractions for environment modeling, observations, and state estimation enable belief programming to provide runtime capabilities for state estimation that eliminate the need to implement and verify the state estimator itself.
\subsection{Epistemic Hoare Logic}
While the belief programming methodology automates the construction of a sound state estimator, a developer must still verify that the contoller's actions and state estimator soundly work together to meet the system's requirements. To address this problem, we present the {\em Epistemic Hoare Logic} (EHL), a variant of Hoare Logic that supports modal propositions in its assertion logic that model a belief program's dynamically tracked belief state.
EHL includes the modal propositions $\Diamond p$, ``it is possible that $p$ is true", and $\Box p$, ``it is always the case that $p$ is true", that quantify $p$ over the set of all possible values of the program's variables as captured by its belief state.
These propositions, along with EHL's inference rules, enable a developer to represent the state estimator's inferences as propositions in the logic --
e.g.\ $\Diamond \ncode{(altitude < 1)}$ meaning ``it is possible that the true altitude is less than 1 foot" -- and also specify and verify the system's requirements --
e.g.\ $\Box \ncode{(altitude >= 1)}$ meaning ``it is always the case that the true altitude is at least 1 foot."
\subsection{Contributions}
In this paper, we present the following contributions:
\begin{itemize}
\item {\it Belief Programming.} We introduce belief programming, a programming methodology that makes it possible for the program runtime to automatically provide a state estimator given an environment model specification. Specifically, the program runtime tracks the program's belief state: all possible values of the program's unobservable state.
\item {\it Language.} We present the syntax and semantics of Belief IMP (BLIMP{}), a language designed for belief programming. We establish basic properties of BLIMP{} semantics that should be true of any belief programming language. Namely, we show the state estimator that a BLIMP{} program provides soundly and precisely captures the environment's true state.
\item {\it Epistemic Hoare Logic.} We present Epistemic Hoare Logic for verifying properties of BLIMP{} programs. We show that our logic is sound with respect to BLIMP{}'s semantics.
\item {\it Case Study.} We present a case study showing how belief programming can be used to develop a verified implementation of the Mars Polar Lander's flight control software.
The Mars Polar Lander is a lost space probe, hypothesized to have crashed into the surface of Mars during descent due to a control software error~\citep{mpl}.
We present a controller implemented with belief programming, and formally prove using EHL that it does not have the error that caused the MPL crash.
\item {\it Implementation.} We evaluate the feasibility of belief programming by presenting an implementation of BLIMP{} in C called CBLIMP.
Our results show that belief programming is feasible for problems in robotics and vehicle navigation domains.
\end{itemize}
The dual contributions of belief programming and Epistemic Hoare Logic enable developers to more easily program in partially observable environments where correctness is paramount.
Such developers must currently hand-write an environment model and a state estimator, and belief programming enables them to omit the state estimator.
Epistemic Hoare Logic enables developers to reason about the correctness of the resulting belief program, just as Hoare logic allows them to reason about the correctness classical hand-written control software.
\clearpage
\section{Language: Syntax and Semantics of BLIMP{}}
\label{sec:language}
\newcommand{\; \mid \;}{\; \mid \;}
\newcommand{\mid \;}{\mid \;}
\newcommand{\mkproplang}[2]{{%
{#2} \; \mid \; {#1} \code{\&\&} {#1} \; \mid \; {#1} \code{||} {#1} \; \mid \; %
\ncode{!} {#1}
}}
\begin{figure}
\begin{align*}
E & \; \mathrm{\texttt{::=}} \; x \; \mid \; c \; \mid \; E \code{+} E %
\; \mid \; E \code{-} E \; \mid \; E \code{*} E \; \mid \;%
E \code{/} E \; \mid \; y \; \mid \; \ncode{.} \\
P & \; \mathrm{\texttt{::=}} \; \mkproplang{P}{b \; \mid \;%
E \code{<} E \; \mid \; E \code{==} E}\\
\Prop_\Box & \; \mathrm{\texttt{::=}} \; \mkproplang{\Prop_\Box}{\Box P \; \mid \; \Diamond P}\\
\Prop_\exists & \; \mathrm{\texttt{::=}} \; \exists y^*. \; \Prop_\Box
\end{align*}
\caption{Syntax of expressions and propositions.}
\label{fig:propsyntax}
\end{figure}
In this section, we present the belief programming language BLIMP{}. We first present BLIMP{}'s syntax and semantics, and then state and prove properties of the semantics.
We then present an {\em execution model}.
While the semantics describes both the environment modeling and state estimation behavior of BLIMP{} programs, the execution model projects out the state estimation operations.
\subsection{Syntax}
In this section, we present the syntax of BLIMP{}, which gives the constructs of belief programming.
\subsubsection{Expressions and Propositions} Figure~\ref{fig:propsyntax} gives the syntax of expressions and propositions.
\paragraph{Expressions.} We use the notation $E$ to refer to expressions.
An expression may be a variable $x$, a numeric constant $c$, or formed using one of the binary operators \rcode{+}, \rcode{-}, \rcode{*}, \rcode{/}, which have standard interpretations.
An expression may also contain a quantified variable $y$ or the placeholder \code{.}.
Quantified variables only make sense when the variable is bound by an outer quantifier, and the placeholder value only makes sense in the context of an enclosing \texttt{choose} statement.
\paragraph{Propositions.} We use the notation $P$ to refer to propositions.
Propositions may be boolean constants $b \in \{\ncode{true}, \ncode{false}\}$ or comparisons between expressions using the comparison operators $\ncode{<}$ and $\ncode{==}$.
Propositions may also be combined by conjunction, disjunction, and negation through the boolean operators $\ncode{\&\&}$, $\ncode{||}$, and $\ncode{!}$, respectively.
\paragraph{Modal Propositions.} We use the notation $\Prop_\Box$ to refer to modal propositions.
Modal propositions are propositions that are modified using the $\Box$ and $\Diamond$ operators.
They quantify over environments in the belief state, and are used to query the belief state for state estimation.
Modal propositions may be combined by conjunction, disjunction, and negation, using the same syntax as for non-modal propositions.
Note that in contrast to many modal logics, BLIMP{}'s modal operators may only be applied once; hence, propositions such as $\Box \Box P$ are not in the language.
This is a design decision we have made to succinctly capture properties of the domain.
The alternative would be to allow propositions such as $\Box \Box P$, and include a theorem of the form $\Box \Box P \Rightarrow \Box P$.
\paragraph{Existential Propositions.} We use the notation $\Prop_\exists$ to refer to existentially quantified modal propositions.
An existential proposition is a modal proposition prepended with the $\exists$ symbol and a comma-separated list of quantified variables $y^*$ (which could potentially be empty).
Existential propositions form the core propositional language for reasoning with Epistemic Hoare Logic, and appear in BLIMP{} in assertions and loop invariants.
\subsubsection{Statements} Figure~\ref{fig:statementsyntax} presents the syntax of statements. We use the notation $S$ to refer to the set of statements, which is specified by the grammar in Figure~\ref{fig:statementsyntax}. A statement may be an assignment, a choose statement, an assertion, an observation, an if statement, an infer statement, a while loop, a composition of two statements, or a skip statement.
\begin{wrapfigure}[14]{l}{0.45\textwidth}
\begin{minipage}{0.45\textwidth}
\begin{align*}
S & \; \mathrm{\texttt{::=}} \; x \code{=} E \\
& \; \mid \; x \code{= choose(} P \rcode{)}\\
& \; \mid \; \lcode{assert} \Prop_\exists\\
& \; \mid \; \lcode{observe} x\\
& \; \mid \; \lcode{if} P \code{\{} S \code{\} else \{} S \code{\}}\\
& \; \mid \; \lcode{infer} \Prop_\Box \code{\{} S \code{\} else \{} S \code{\}}\\
& \; \mid \; \code{while} P \code{\{} \Prop_\exists \code{\} \{} S \code{\}}\\
& \; \mid \; S \code{;} S\\
& \; \mid \; \ncode{skip}
\end{align*}
\end{minipage}
\caption{Syntax of statements}
\label{fig:statementsyntax}
\end{wrapfigure}
\paragraph{Assignment and Choose.} An assignment statement $x \code{=} E$ and a choose statement $x \code{= choose(} P \rcode{)}$ both assign to the program variable $x$.
The assignment statement does so using an expression while the choose statement does so using a proposition that contains the $\ncode{.}$ placeholder value.
\paragraph{Assertions.} The keyword \texttt{assert} signifies an assertion statement.
An assertion includes an existential proposition specifies a property that must be true at the statement's program point.
\paragraph{Observation.} The keyword \texttt{observe} signifies an observation statement. An observation includes the program variable, $x$, to be observed at the statement's program point.
\paragraph{If and Infer Statements.} The keywords \texttt{if} and \texttt{infer} signify an if statement and an infer statement, respectively.
Both statements select branches based on a condition.
The distinction between an if and an infer statement is that an if statement branches on an ordinary proposition, whereas an infer statement branches on a modal proposition.
If statements facilitate conditional environment models, whereas infer statements facilitate state estimation by branching on belief state queries.
\paragraph{While Loops.} The keyword \texttt{while} signifies a while loop.
Such a loop consists of a body, a condition regarding whether to continue or not, and a loop invariant.
The loop invariant is an existential proposition that must be true at the start and end of each loop, and may be used to express a safety property that must be true at every iteration of a time-step loop.
\paragraph{Composition and Skip.} Statements may be sequentially composed with the $\ncode{;}$ operator.
The \texttt{skip} keyword signifies a skip statement, a null statement that performs no operations.
\section{Epistemic Hoare Logic}
\label{sec:logic}
\newcommand{\; \text{fresh in} \;}{\; \text{fresh in} \;}
\begin{figure}
\begin{mathpar}
\inferrule
{
x_0 \; \text{fresh in} \; \prop_\Box, e\\\\
\prop_\Box' = \prop_\Box[x_0/x]\\
e' = e[x_0/x]
}
{
\htriplepc{N}{\exists \hat{y} . \; \prop_\Box}{x \code{=} e}{
\exists \hat{y}. \; \prop_\Box' \code{\&\&}
\Box \ncode{(} x \code{==} e' \ncode{)}
}
}
\inferrule
{
\htriplepc{N}{\prop_\exists}{s}{\prop_\exists'}
}
{
\htriplepc{D}{\prop_\exists}{s}{\prop_\exists'}
}
\inferrule
{
x_0 \; \text{fresh in} \; \prop_\Box, p\\\\
\prop_\Box' = \prop_\Box[x_0/x]\\
p' = p[x_0/x][x/\ncode{.}]
}
{
\htriplepc{N}{\exists \hat{y}. \; \prop_\Box}
{x \rcode{= choose(} p \ncode{)}}{
\exists \hat{y}. \; \prop_\Box' \code{\&\&}
\Box\ncode{(}p'\ncode{)}
}
}
\inferrule
{ \forall \beta. \; \beta \vDash \prop_\exists \Rightarrow %
\beta \vDash \prop_\exists^a }
{
\htriplepc{N}{\prop_\exists}{
\lcode{assert} \prop_\exists^a
}{\prop_\exists}
}
\inferrule
{
\htriplepc{PC}{\prop_\exists}{s_1}{\prop_\exists'}\\
\htriplepc{PC}{\prop_\exists'}{s_2}{\prop_\exists''}
}
{
\htriplepc{PC}{\prop_\exists}{s_1 \code{;} s_2}{\prop_\exists''}
}
\inferrule { }
{
\htriplepc{N}{\prop_\exists}{\ncode{skip}}{\prop_\exists}
}
\inferrule
{
y_{n+1} \; \text{fresh in} \; \prop_\Box\\
\prop_\Box' = \prop_\Box[y_{n+1}/x]
}
{
\htriplepc{D}{\exists y_0, y_1, \dots, y_n. \; \prop_\Box}
{\lcode{observe} x}{
\exists y_0, y_1, \dots, y_n, y_{n+1}. \;
\prop_\Box' \code{\&\&}
\Box \ncode{(} x \code{==} y_{n+1} \ncode{)}
}
}
\inferrule
{
\htriplepc{N}{\exists \hat{y}. \; \ncode{(}\Box p \ncode{)}
\code{\&\&} \prop_\Box}{s_1}
{\exists \hat{y}'. \; \Box p'}\\
\htriplepc{N}{\exists \hat{y}. \; \ncode{(}\Box \ncode{!}p \ncode{)}
\code{\&\&} \prop_\Box}{s_2}
{\exists \hat{y}'. \; \Box p'}
}
{
\htriplepc{N}{\exists \hat{y}. \; \prop_\Box}
{
\lcode{if} p \code{\{} s_1
\code{\} else \{} s_2 \rcode{\}}
} {
\exists \hat{y}'. \; \Box p'
}
}
\inferrule
{
\forall \beta. \; \beta \vDash \exists \hat{y}. \; \prop_\Box
\Rightarrow \beta \vDash \Box p \code{||}
\Box\ncode{(!} p \ncode{)}\\
\htriplepc{D}{\exists \hat{y}. \; \ncode{(}\Box p \ncode{)}
\code{\&\&} \prop_\Box}{s_1}
{\prop_\exists'}\\\\
\htriplepc{D}{\exists \hat{y}. \; \ncode{(}\Box \ncode{!}p \ncode{)}
\code{\&\&} \prop_\Box}{s_2}
{\prop_\exists'}
}
{
\htriplepc{D}{\exists \hat{y}. \; \prop_\Box}
{
\lcode{if} p \code{\{} s_1
\code{\} else \{} s_2 \rcode{\}}
} {
\prop_\exists'
}
}
\inferrule
{
\htriplepc{PC}{\exists \hat{y}. \; \prop_\Box^i
\code{\&\&} \prop_\Box}{s_1}
{\prop_\exists'}\\
\htriplepc{PC}{\exists \hat{y}. \; \ncode{!(} \prop_\Box^i \lcode{) \&\&}
\prop_\Box}{s_2}
{\prop_\exists'}
}
{
\htriplepc{PC}{\exists \hat{y}. \; \prop_\Box}
{
\lcode{infer} \prop_\Box^i \code{\{} s_1
\code{\} else \{} s_2 \rcode{\}}
} {
\prop_\exists'
}
}
\inferrule {
\htriplepc{N}{\prop_\exists'}{s}{\prop_\exists''}\\
\forall \beta. \; \beta \vDash \prop_\exists \Rightarrow
\beta \vDash \exists \hat{y}. \; \Box p^I\\
\forall \beta. \; \beta \vDash \exists \hat{y}. \;
\Box p^I \code{\&\&} \Box p
\Rightarrow \beta \vDash \prop_\exists' \\
\forall \beta. \; \beta \vDash \prop_\exists'' \Rightarrow
\beta \vDash \exists \hat{y}. \; \Box p^I
}
{
\htriplepc{N}{\prop_\exists}
{\lcode{while} p \code{\{}
\exists \hat{y}. \; \Box p^I \code{\} \{} s \code{\}}}
{ \exists \hat{y}.\; \Box \ncode{(!} p \code{\&\&} p^I \lcode{)} }
}
\inferrule {
\htriplepc{D}{\prop_\exists'}{s}{\prop_\exists''}\\\\
\forall \beta. \; \beta \vDash \exists \hat{y}. \; \prop_\Box^I \Rightarrow
\beta \vDash \Box p \code{||} \Box \ncode{(!} p \ncode{)}\\
\forall \beta. \; \beta \vDash \prop_\exists \Rightarrow
\beta \vDash \exists \hat{y}. \; \prop_\Box^I \\\\
\forall \beta. \; \beta \vDash \exists \hat{y}. \;
\prop_\Box^I \code{\&\&} \Box p
\Rightarrow \beta \vDash \prop_\exists' \\
\forall \beta. \; \beta \vDash \prop_\exists'' \Rightarrow
\beta \vDash \exists \hat{y}. \; \prop_\Box^I
}
{
\htriplepc{D}{\prop_\exists}
{\lcode{while} p \code{\{}
\exists \hat{y}. \; \prop_\Box^I \code{\} \{} s \code{\}}}
{ \exists \hat{y}.\; \Box \ncode{!(} p \lcode{) \&\&} \prop_\Box^I }
}
\inferrule {
\htriplepc{PC}{\prop_\exists'}{s}{\prop_\exists''}\\\\
\forall \beta. \; \beta \vDash \prop_\exists \Rightarrow
\beta \vDash \prop_\exists'\\\\
\forall \beta. \; \beta \vDash \prop_\exists'' \Rightarrow \prop_\exists'''
}
{
\htriplepc{PC}{\prop_\exists}{s}{\prop_\exists'''}
}
\end{mathpar}
\caption{Epistemic Hoare Logic rules. We use the notation $\htriplepc{PC}{\prop_\exists}{s}{\prop_\exists'}$ to mean that in a context $PC$ the statement $s$ maps belief states satisfying $\prop_\exists$ to new belief states satisfying $\prop_\exists'$.}
\label{fig:logic}
\end{figure}
In this section, we present Epistemic Hoare Logic and sketch a proof of its soundness.
Figure~\ref{fig:logic} gives the inference rules for Epistemic Hoare Logic.
Each rule yields a deduction $\htriplepc{PC}{\prop_\exists}{s}{\prop_\exists'}$, where the context $PC$ is drawn from the grammar
$
PC \; \mathrm{\texttt{::=}} \; N \; \mid \; D
$.
The purpose of the context is to ensure observations do not occur under nondeterministic control flow, as that would result in an error according to the semantics.
The deduction $\htriplepc{D}{\prop_\exists}{s}{\prop_\exists'}$ meas that, assuming the statement $s$ is executed under deterministic control flow and terminates, $s$ maps belief states satisfying the pre-condition $\prop_\exists$ to new belief states satisfying the post-condition $\prop_\exists'$.
The deduction $\htriplepc{N}{\prop_\exists}{s}{\prop_\exists'}$ means the same thing, but where the enclosing control flow may be nondeterministic.
\paragraph{Assignment and Choose.} Assignment and choose statements conjunct the pre-condition with a new proposition.
In the case of assignment, the new proposition encodes that the value of the variable is equal to the result of the expression.
In the case of a choose statement, the new proposition encodes that the value of the variable must be consistent with the choose statement's proposition.
In either case the previous value of the variable is encoded with a fresh variable.
\paragraph{Assertions.} An assertion has identical pre- and post-condition propositions.
In order to apply the rule, the developer must show that the pre-condition implies the asserted proposition.
\paragraph{Observations.} An observation adds a new existentially quantified variable $y_{n+1}$ to the variable list from the pre-condition.
The post-condition ensures that the value of the observed variable is deterministic by stating that in every environment it is equal to $y_{n+1}$.
Moreover, $y_{n+1}$ is not completely unrestricted; it must satisfy all properties that the observed variable satisfied in the precondition.
Observations always require that the enclosing control flow is deterministic.
\paragraph{Composition and Skip.} The rules for statement sequencing and skip statements are standard; they are the same as in classical Hoare logic~\citep{floyd, hoare}
\paragraph{If and Infer.} Both if and infer statements require that both branches satisfy the same post-condition.
For if, the pre-condition of each branch includes the statement's condition under the $\Box$ modality.
For infer, the pre-condition of each branch includes the statement's condition, which is itself a modal proposition.
Furthermore, the post-condition of an if statement must use the $\Box$ modality, whereas the post-condition of an infer statement may be any existential proposition.
A more conventional approach would be to have both branches of the if statement imply the same predicate as the post-condition.
For this to be sound, we would need to show that if $\beta_1 \models p_\exists$ and $\beta_2 \models p_\exists$ then $\beta_1 \cup \beta_2 \models p_\exists$, which is in fact false.
However, it is true that if $\beta_1 \models \Box p$ and $\beta_2 \models \Box p$, then $\beta_1 \cup \beta_2 \models \Box p$.
To preserve a deterministic context, developers must ensure that the the if statement's condition has the same value in all environments.
\paragraph{While.}
To verify a while loop with Epistemic Hoare Logic, the developer must prove several properties of the initial pre-condition $p_\exists$, the loop invariant $p_\exists^I$, the loop body's pre-condition $p_\exists'$, and the loop body's post-condition $p_\exists''$.
The developer must show that the pre-condition implies the loop invariant, the invariant implies the body's pre-condition, and the body's post-condition implies the invariant.
Implications such as these are a standard feature of proving properties using Hoare rules, although the overall proof rules are sometimes structured differently.
To preserve a deterministic context, developers must also ensure that the the while loop's condition has the same value in all environments.
\paragraph{Rule of Consequence.} The rule of consequence that any proposition that implies the pre-condition may be substituted for the pre-condition. Conversely, any proposition implied by the post-condition may be substituted for the post-condition.
\subsection{Soundness}
In this section, we formalize and establish the soundness of Epistemic Hoare Logic with respect to the semantics of BLIMP{}.
We start by explaining what supporting lemmas are needed, and then state the main theorem and sketch the proof.
\subsubsection{Substitution}
The main theorem depends on a number of lemmas that relate the substitutions in Figure~\ref{fig:logic} to the environment mapping in Figure~\ref{fig:statementsemantics}.
The following lemma gives the required property for expressions.
It states that if we evaluate an expression under a new environment with a fresh variable $x_0$ that is a rebinding of $x$, we are free to rename $x$ to $x_0$ in the expression without changing the result.
\begin{lemma}[Expression Substitution]
{\ \\}
\vspace{-.35cm}
\begin{center}
If $x_0$ is fresh in $e$, then $\tuple{\sigma, e} \Downarrow c \iff \tuple{\sigma[x_0 \mapsto \sigma(x)], e} \Downarrow c$ and
$\tuple{\sigma, e} \Downarrow c \iff \tuple{\sigma[x_0 \mapsto \sigma(x)], e[x_0/x]} \Downarrow c$
\end{center}
\vspace{-.3cm}
\end{lemma}
\begin{proof}By structural induction on expressions.\end{proof}
By similar means, we establish analogous properties for propositions, modal propositions, and existential propositions.
The full set of lemmas is in Appendices \ref{sec:substproofs}-\ref{sec:constsubstproofs}.
\subsubsection{Observation List Emptiness}
The main theorem depends upon a lemma that states that observations cannot happen under nondeterministic control flow.
We have formalized this property as follows:
\begin{lemma}[Observation List Emptiness]{\ \\}
\vspace{-.35cm}
\begin{center}
If $\htriplepc{N}{\prop_\exists}{s}{\prop_\exists'}$ and $\tuple{\beta, \mu, s} \Downarrow \tuple{C \; | \; o}$, then $o = \ncode{nil}$.
\end{center}
\vspace{-.3cm}
\end{lemma}
\begin{proof}
By structural induction on derivations of $\Downarrow$.
The case of sequencing uses the fact that $\ncode{nil} \concat \code{nil} = \ncode{nil}$
\end{proof}
\subsubsection{Soundness Theorem}
In this section, we establish the soundness of the logic. Our approach is to show partial correctness, meaning that if the program terminates, its final state is described by the post-condition.
The theorem has two parts.
First, we establish the soundness of the logic in the case where control flow may be nondeterministic. The theorem states that if the belief state satisfies the pre-condition, and the semantics produces a configuration, the configuration includes a new belief state that satisfies the post-condition.
Second, we establish the soundness of the logic under deterministic control flow.
The theorem states that if the initial belief state satisfies the pre-condition, and the true environment is in the belief state, then the program executes without error and the new belief state satisfies the post-condition.
\begin{theorem}[Logic Soundness]{\ \\}
\vspace{-.35cm}
\begin{center}
\begin{enumerate}
\item If $\htriplepc{N}{\prop_\exists}{s}{\prop_\exists'}$, $\beta \vDash \prop_\exists$, and $\tuple{\beta, \mu, s} \Downarrow \tuple{C \; | \; o}$, then $C = (\beta', \mu')$ and $\beta' \vDash \prop_\exists'$.
\item If $\htriplepc{D}{\prop_\exists}{s}{\prop_\exists'}$, $\sigma \in \beta$, $\beta \vDash \prop_\exists$, and $\tuple{\beta, \sigma, s} \Downarrow \tuple{C \; | \; o}$, then $C = (\beta', \sigma')$ and $\beta' \vDash \prop_\exists'$.
\end{enumerate}
\end{center}
\vspace{-.2cm}
\label{thm:ehlsound}
\end{theorem}
\begin{proof}
\begin{enumerate}
\item By structural induction on derivations of $\Downarrow$. The cases for assignments and choose statements rely on substitution lemmas.
The cases for deterministic if, infer, and while follow from induction hypotheses, although while loops require destructing the semantic rule.
For non-deterministic if statements, we show that
$\beta_1 \vDash \exists \hat{y}. \; \Box p \wedge
\beta_2 \vDash \exists \hat{y}. \; \Box p \Rightarrow
\beta_1 \cup \beta_2 \vDash \exists \hat{y}. \; \Box p
$, which follows from standard principles of first-order logic.
The details are in Appendix~\ref{sec:ehlsoundproof1}.
\item By structural induction on derivations of $\Downarrow$. The cases except for$\code{observe}$are similar to those above, except that we use Theorem~\ref{thm:bsound} to ensure the premises of the induction hypotheses.
The case for$\code{observe}$relies on the assumption that $\sigma \in \beta$ to instantiate $y_{n+1}$. The details are in Appendix~\ref{sec:ehlsoundproof2}.
\end{enumerate}
\end{proof}
\section{The MPL-Exp Benchmark}
\label{sec:mplexp}
In this section, we describe the MPL-Exp benchmark.
This section is laid out similarly to Section~\ref{sec:casestudy}.
Figure~\ref{fig:mplexpmodel} presents a modified version of Figure~\ref{fig:mplmodel}, and Figure~\ref{fig:mplexpbp} presents a modified version of Figure~\ref{fig:mplbp}.
Section~\ref{sec:mplexpverification} is analogous to \ref{sec:mplverification} and explains how to verify the modified programs using Epistemic Hoare logic.
\begin{figure}
\begin{lstlisting}[basicstyle=\ttfamily\small]
//Model Initialization
prev_err_1 = 0; prev_err_2 = 0; trans_td = 0;
alt = 4;
time_on_ground = 0;
//Permanent errors
perm_1 = choose(. == 0 || . == 1);
perm_2 = choose((. == 0 || . == 1) && (perm_1 == 1 => . == 0));
perm_1_v = choose(. == 0 || . == 1);
perm_2_v = choose(. == 0 || . == 1);
//Controller initialization
...
while(engine_enabled == 1)
{ (alt > -1 => engine_enabled == 1) &&
(trans_td == 0 => time_on_ground < 2) }
{
//Model start
if (alt == -1 && engine_enabled == 1) {
time_on_ground = time_on_ground + 1
};
if (alt >= 3) { alt_rate = choose(. == 0 || . == 1) }
else { alt_rate = choose(. <= prev_alt + 1) };
alt = alt - alt_rate;
if (alt >= 3 ) { radar_alt = choose( alt - 1 <= . && . <= alt + 1) }
else if (alt == 2) { radar_alt = choose ( -1 <= . && . < = 3) }
else { radar_alt = choose(-1 <= . && . <= 2) };
err_1 = choose((prev_err_1 == 1 => . == 0) && (. == 0 || . == 1));
prev_err_1 = err_1;
err_2 = choose((prev_err_2 == 1 => . == 0) && (. == 0 || . == 1));
prev_err_2 = err_2;
if (alt == -1 && (err_1 == 1 || err_2 == 1)) { trans_td = 1; };
leg_err = alt == 4;
if perm_1 { cur_td_1 = perm_1_v }
else if (leg_err == 1 || err_1 == 1) {
cur_td_1 = choose(. == 0 || . == 1)
} else { cur_td_1 = alt == -1; }
else if (leg_err == 1 || err_2 == 1) {
cur_td_2 = choose(. == 0 || . == 1)
} else { cur_td_2 = alt == -1; };
//Model end
//Controller loop start
...
}
\end{lstlisting}
\caption{Model for the MPL-Exp benchmark.}
\label{fig:mplexpmodel}
\end{figure}
\begin{figure}
\begin{lstlisting}[basicstyle=\ttfamily\small]
// Model Initialization
...
engine_enabled = 1;
while (engine_enabled == 1)
{ $\Box$((alt > -1 => engine_enabled == 1) &&
(trans_td == 0 => time_on_ground < 2)) }
{
// Model start
...
// Model end
observe radar_alt;
observe cur_td_1;
observe cur_td_2;
infer $\Box$(alt == -1) {
engine_enabled = 0
}
}
\end{lstlisting}
\caption{Belief program of the MPL-Exp benchmark.}
\label{fig:mplexpbp}
\end{figure}
\paragraph{\bf Altitude Model.} Instead of the uniform 1-meter discretization in Figure~\ref{fig:mplmodel}, the MPL-Exp benchmark uses an exponential discretization. This modified discretization applies to the \texttt{alt} and \texttt{radar\_alt} variables, and is defined as follows:
\begin{itemize}
\item When an altitude variable is 4, the altitude is between 1000 and 10000 meters.
\item When an altitude variable is 3, the altitude is between 100 and 1000 meters.
\item When an altitude variable is 2, the altitude is between 10 and 100 meters.
\item When an altitude variable is 1, the altitude is between 1 and 10 meters.
\item When an altitude variable is 0, the altitude is between 0 and 1 meters.
\item When an altitude variable is -1, the altitude is exactly 0 meters.
\end{itemize}
The model in Figure~\ref{fig:mplexpmodel} is the same as that of Figure~\ref{fig:mplmodel} except that the altitude model has been modified to reflect the exponential discretization.
The update rules are designed to conservatively over-approximate the model in Figure~\ref{fig:mplexpmodel}.
This means that after applying the definition of the discretization, any possible true environment in Figure~\ref{fig:mplmodel} is also possible under Figure~\ref{fig:mplexpmodel}.
The belief program in Figure~\ref{fig:mplexpbp} is the same as that of Figure~\ref{fig:mplbp} except that all references to the \texttt{alt} variable have been modified to use the exponential discretization.
\subsection{Verification}
\label{sec:mplexpverification}
In this section, we explain how to verify the loop invariant of the MPL-Exp example, which describes the same property as that of the MPL example, but with the exponential altitude discretization.
The process proceeds the same as in Section~\ref{sec:mplverification}. We first construct a strengthened invariant, which is the same as the strengthened invariant in Section~\ref{sec:mplverification} but with altitude comparisons to \texttt{0} replaced with comparisons to \texttt{-1}.
Then, using the rules of Epistemic Hoare logic, we construct a formula $\prop_\exists^\mathrm{MPL-Exp}$ by propagating the strengthened invariant through the loop body, and show that $\prop_\exists^\mathrm{MPL-Exp}$ implies the invariant itself.
To show that $\prop_\exists^\mathrm{MPL-Exp}$ implies the strengthened invariant, we follow the same general line of reasoning as in Section~\ref{sec:mplverification}.
All of the proof except the engine disabled soundness condition follows the exact same reasoning as Section~\ref{sec:mplverification} with appropriate changes to take into account the new strengthened invariant and the use of $\prop_\exists^\mathrm{MPL-Exp}$.
Showing engine disabled soundness requires more care because the proof makes use of intermediate propositions that critically depend on the altitude model.
\subsubsection{Engine Disabled Soundness}
To apply an analogous line of reasoning to Section~\ref{sec:engdissound}, we need a value $c^*$ that satisfies the conditions
$
\beta \vDash \Box \code{(y\_radar >} c^* \rcode{)} \Rightarrow
\beta \vDash \Box \rcode{(alt > -1)}
$
and
\lstinline[breaklines=true]{$\beta \vDash \Box$(y_radar <= $c^*$) $\Rightarrow$ $\beta \vDash \Box$(landing_leg_deployment == 0)}.
While we omit the full derivation here, we note that choosing $c^* = \texttt{2}$, applying $\prop_\exists^\mathrm{MPL-Exp}$ and Theorems \ref{thm:lifting} and \ref{thm:k} can prove these implications.
The remaining pieces of the proof are the same as in Section~\ref{sec:engdissound}.
\section{Verifying the MPL with Epistemic Hoare logic}
\label{sec:mplverification}
In this section, we will show a piece of the verification process of the MPL belief program in Figure~\ref{fig:mplbp}.
Specifically, we will show how to verify the belief program's while loop preserves the loop invariant.
Note that some of the variable names in this section are slightly different than the names in Section~\ref{sec:casestudy}.
First, we strengthen the loop invariant to the following proposition.
This strengthened condition contains the additional properties:
\begin{itemize}
\item {\it Altitude and time-on-ground nonnegativity.} The variables \lstinline{altitude} and \lstinline{time_on_ground} never go below \lstinline{0}.
\item {\it Permanent Error Single-upset.} It is never the case that both touchdown sensors suffer permanent failures.
\item {\it Existential Variables.} There are two existentially quantified variables that stand in for the observed touchdown sensor values. This means that if there are no transients on touchdown and the landing legs are not deploying, these sensors will properly indicate a touchdown.
\item {\it Engine Disabled Soundness.} On the second time step after touchdown, if there are no transients on touchdown, the controller will disable the engine.
\end{itemize}
\begin{lstlisting}
$\exists$ y1, y2.
$\Box$(altitude >= 0) &&
$\Box$(time_on_ground >= 0) &&
$\Box$(altitude > 0 => time_on_ground == 0) &&
$\Box$(!(permanent_1 == 1 && permanent_2 == 1)) &&
$\Box$(altitude > 0 => engine_enabled = 1) &&
$\Box$((((y1 != 1 && permanent_1 != 1) ||
(y2 != 1 && permanent_2 != 1)) &&
transient_on_touchdown == 0 &&
landing_leg_deployment == 0) => altitude != 0) ) &&
$\Box$((transient_on_touchdown == 0 && time_on_ground == 1) =>
engine_enabled == 0) &&
$\Box$(transient_on_touchdown == 0 => time_on_ground < 2) &&
\end{lstlisting}
The strengthened invariant can be shown to be correct on the first iteration of the loop applying our logic to the model initialization code.
Our goal here is to show that this invariant is preserved by the belief program's control loop.
Calling this proposition $\prop_\exists^I$, we can propagate it through the loop body using the rules in Figure~\ref{fig:logic}.
For the post-conditions of each if statement, we assume a natural proposition that includes a disjunction representing both branches and wraps the disjunction in a $\Box$ modality.
Similarly, for each infer statement, we include a disjunction that uses the modality of the infer statement's condition.
This yields the following formula, which we call $\prop_\exists^\mathrm{MPL}$. We will not explain the contents of $\prop_\exists^\mathrm{MPL}$ at length, but will refer to it piecemeal in the remainder of the proof.
\begin{lstlisting}[xleftmargin=2em,numbers=left]
$\exists$ y_radar, y_td1, y_td2, y1, y2.
$\Box$(altitude_0 >= 0) &&$\label{code:altznonneg}$
$\Box$(time_on_ground_0 >= 0) &&$\label{code:tognonneg}$
$\Box$(altitude_0 > 0 => time_on_ground_0 == 0) &&$\label{code:alttogz}$
$\Box$(!(permanent_1 == 1 && permanent_2 == 1)) &&$\label{code:perm}$
$\Box$(altitude_0 > 0 => engine_enabled_0 = 1) &&$\label{code:ezenabled}$
$\Box$((((y1 != 1 && permanent_1 != 1) ||$\label{code:yassump}$
(y2 != 1 && permanent_2 != 1)) &&
transient_on_touchdown == 0 &&
landing_leg_deployment_0 == 0) => altitude_0 != 0) ) &&
$\Box$((transient_on_touchdown_0 == 0 && time_on_ground_0 == 1) =>
engine_enabled_0 == 0) &&$\label{code:togezdis}$
$\Box$(transient_on_touchdown_0 == 0 => time_on_ground_0 < 2) &&$\label{code:togbound}$
( $\Box$(altitude_0 == 0 && engine_enabled_ == 1 && $\label{code:togasgn}$
time_on_ground == time_on_ground_0 + 1) ||
(!(altitude_0 == 0 && engine_enabled_0 == 1) &&
time_on_ground == time_on_ground_0)
) &&
$\Box$(0 <= altitude_rate && altitude_rate <= 39 && $\label{code:altrate}$
altitude_rate <= altitude_0) &&$\label{code:altitudenonneg}$
$\Box$(altitude == altitude_0 - altitude_rate) &&$\label{code:altitudeasgn}$
$\Box$((prev_error_1_0 == 1 => error_1 == 0) &&
(error_1 == 0 || error_1 == 1)) &&
$\Box$(prev_error_1 == error_1) &&
$\Box$((prev_error_2_0 == 1 => error_2 == 0) &&
(error_2 == 0 || error_2 == 1)) &&
$\Box$(prev_error_2 == error_2) &&
$\Box$((transient_on_touchdown == 0 && altitude == 0) =>$\label{code:totasgn}$
(error_1 != 1 && error_2 != 1)) &&
( $\Box$(1400 <= altitude && altitude <= 1600 &&
landing_leg_deployment == 1) ||
((1400 > altitude || 1600 < altitude) &&
landing_leg_deployment == 0)
) &&
$\Box$((permanent_1 == 1 && y_td1 == permanent_1_val) ||$\label{code:tdiasgn}$
(permanent_1 != 1 &&
(landing_leg_deployment == 1 || error_1 == 1) &&
y_td1 == 0 || y_td1 == 1
) ||
(permanent_1 != 1 && landing_leg_deployment != 1 &&
error_1 != 1 &&
altitude == 0 => y_td1 == 1 &&
altitude != 0 => y_td1 == 0
)
) &&
$\Box$((permanent_2 == 1 && y_td2 == permanent_2_val) ||$\label{code:tdiiasgn}$
(permanent_2 != 1 &&
(landing_leg_deployment == 1 || error_2 == 1) &&
y_td2 == 0 || y_td2 == 1
) ||
(permanent_1 != 1 && landing_leg_deployment != 1 &&
error_2 != 1 &&
altitude == 0 => y_td1 == 1 &&
altitude != 0 => y_td1 == 0
)
) &&
$\Box$(40 <= y_radar - altitude && y_radar - altitude <= 40) &&$\label{code:radaracc}$
$\Box$(y_radar == radar_altitude) &&$\label{code:yradar}$
$\Box$(y_td1 == current_touchdown_indicator_1) &&
$\Box$(y_td2 == current_touchdown_indicator_2) &&
( ($\Box$(altitude == 0) && $\Box$(engine_enabled == 0)) ||$\label{code:edisabled}$
($\Diamond$(altitude != 0) && $\Box$(engine_enabled == engine_enabled_0))$\label{code:eenabled}$
)
\end{lstlisting}
We will now show that
$\forall \beta. \; \beta \vDash \prop_\exists^\mathrm{MPL}
\Rightarrow \beta \vDash \prop_\exists^I$
We will proceed by handling each conjunction in $\prop_\exists^I$ individually.
\subsubsection{Altitude nonnegativity}
Here, we will show that
$\forall \beta.\; \beta \vDash \prop_\exists^\mathrm{MPL} \Rightarrow
\beta \vDash \Box\ncode{(altitude >= 0)}$
We rely on the fact that, based on standard properties of numerical comparisons and arithmetic operators,
$\forall \sigma. \; \sigma \vDash \code{altitude\_rate <= altitude\_0} \Rightarrow
\sigma \vDash \rcode{altitude\_0 - altitude\_rate >= 0}$
Thus, by applying Theorem~\ref{thm:lifting}, Theorem~\ref{thm:k}, and the assumptions on Lines \ref{code:altitudenonneg} and \ref{code:altitudeasgn}, we have the result.
\subsubsection{Time on ground non-negativity}
Here, we will show that
$\forall \beta.\; \beta \vDash \prop_\exists^\mathrm{MPL} \Rightarrow
\beta \vDash \Box\ncode{(time\_on\_ground >= 0)}$
From Lines \ref{code:tognonneg} and \ref{code:togasgn} of $\prop_\exists^\mathrm{MPL}$, we can deduce that
\begin{lstlisting}[columns=flexible]
$\Box$(time_on_ground_0 >= 0) &&
$\Box$(time_on_ground == time_on_ground_0 || time_on_ground == time_on_ground_0 + 1)
\end{lstlisting}
We will now analyze the two cases of the disjunction separately.
In the first case we use the following implication, which must be true by a simple substitution argument.
\begin{align*}
\forall \sigma.\; \sigma & \vDash \code{(time\_on\_ground\_0 >= 0 \&\& %
time\_on\_ground == time\_on\_ground\_0)} \Rightarrow \\
\sigma & \vDash \code{time\_on\_ground >= 0}
\end{align*}
In the second case, we can use following implication, which follows from the interation of $\code{>=}$ and $\rcode{+}$.
\begin{align*}
\forall \sigma.\; \sigma & \vDash \code{(time\_on\_ground\_0 >= 0 \&\& %
time\_on\_ground == time\_on\_ground\_0 + 1)} \Rightarrow \\
\sigma & \vDash \code{time\_on\_ground >= 0}
\end{align*}
Combining these theorems together and then applying Theorems \ref{thm:lifting} and \ref{thm:k} shows that \lstinline[breaklines=true]{$\Box$(time_on_ground >= 0)}.
\subsubsection{Single upset for permanent errors}
Here, we will show that
\lstinline[breaklines=true]{$\Box$(!(permanent_1 == 1 && permanent_2 == 1))}.
This is directly assumed on Line~\ref{code:perm} of $\prop_\exists^\mathrm{MPL}$.
\subsubsection{Engine enabled soundness}
Here, we will show that
$\forall \beta. \; \beta \vDash \prop_\exists^\mathrm{MPL}
\Rightarrow \beta \vDash \Box \ncode{(altitude > 0 => engine\_enabled == 1)}$.
Notably, the original MPL software did not satisfy this condition and as a result likely shut off its engine too early resulting in a crash.
We case split on the disjunction on Line~\ref{code:edisabled}.
In the first case on Line~\ref{code:edisabled}, the assumption that \lstinline{$\Box$(altitude == 0)} means that any environment in the belief state will vacuously satisfy \lstinline{altitude > 0 => engine_enabled == 1}.
In the second case on Line~\ref{code:edisabled}, we can assume that
\lstinline[breaklines=true]{$\Diamond$(altitude != 0) => $\Box$(engine_enabled == 1)}.
Inlining definitions and using Theorem~\ref{thm:box-diamond}, we can see that
\[
\Box\ncode{(altitude == 0)} \code{||} \Box\ncode{(engine\_enabled == 1)}
\]
Then, applying Theorem~\ref{thm:box-or}, we can deduce that
\[
\Box\ncode{(altitude == 0 || engine\_enabled == 1)}
\]
which is equivalent to
\[
\Box\ncode{(altitude != 0 => engine\_enabled == 1)}
\]
Combining this with the fact that \lstinline[breaklines=true]{$\Box$(altitude >= 0)} Using Theorems \ref{thm:lifting} and \ref{thm:k} proves the result.
\subsubsection{Time on ground vs. altitude}
\label{sec:togalt}
Here, we will show that \lstinline[breaklines=true]{$\Box$(altitude > 0 => time_on_ground == 0)}.
The assumptions on Lines \ref{code:altznonneg}, \ref{code:alttogz}, \ref{code:ezenabled}, \ref{code:togasgn}, and \ref{code:altitudeasgn} mean that any environment in the belief state satisfies,
\begin{lstlisting}[columns=flexible]
altitude_0 > 0 => time_on_ground_0 == 0 &&
altitude_0 > 0 => engine_enabled == 1 &&
altitude_0 == 0 => altitude == 0 &&
(altitude_0 == 0 ||
engine_enabled_0 != 1 ||
altitude_0 > 0 && time_on_ground == time_on_ground_0)
\end{lstlisting}
Thus, any environment in the belief state satisfies \lstinline[breaklines=true]{altitude_0 > 0 => time_on_ground == 0}
\subsubsection{Existential Variables}
In this section, we will show that
\begin{align*}
\beta \vDash \prop_\exists^\mathrm{MPL} \Rightarrow
\beta \vDash \exists & \code{y1, y2. ((y1 || y2) \&\&}\\
& \code{transient\_on\_touchdown == 0 \&\& landing\_leg\_deployment == 0) =>}\\
& \code{altitude == 0}
\end{align*}
The conditions on Lines \ref{code:totasgn}, \ref{code:tdiasgn}, and \ref{code:tdiiasgn} can be rearranged to show that
\begin{align*}
\beta \vDash \prop_\exists^\mathrm{MPL} \Rightarrow
\beta \vDash \exists & \code{y\_td1, y\_td2.} \Box\lcode{(transient\_on\_touchdown == 0 \&\&}\\
& \code{landing\_leg\_deployment == 0 \&\&}\\
&\code{!(permanent\_1 == 1 \&\& permanent\_2 == 1) =>}\\
& \code{((y\_td1 || y\_td2) <=> altitude == 0))}
\end{align*}
Because by the assumption on Line~\ref{code:perm} we have that the belief state satisfies \lstinline[breaklines=true]{$\Box$(!(permanent_1 == 1 && permanent_2 == 1))}, we can $\alpha$-rename \lstinline{y_td1} to \lstinline{y1} and \lstinline{y_td2} to \lstinline{y2} which implies the result.
\subsubsection{Engine Disabled Soundness}
\label{sec:engdissound}
In this section we will show that
\[
\forall \beta. \; \beta \vDash \prop_\exists^\mathrm{MPL} \Rightarrow
\beta \vDash \Box \ncode{(time\_on\_ground == 2 => engine\_enabled == 0)}
\]
We case split on the disjunction on Line~\ref{code:eenabled}.
In the first case, we can assume that the belief state satisfies $\Box\ncode{(engine\_enabled == 0)}$ which ensures the result.
In the second case, we can assume that the belief state satisfies $\Diamond\ncode{altitude != 0}$.
We note the following implication
\begin{align*}
\forall \sigma . \; \sigma \vDash & \code{altitude != 0} \Rightarrow\\
\sigma \vDash & \code{y\_radar > 1000 || } \\
& \code{((y1 != 1 || y\_td1 != 1) \&\& (y2 != 1 || y\_td2 != 1)) ||} \\
& \code{!( (((y1 == 1 \&\& y\_td1 == 1) || (y2 == 1 \&\& y\_td2 == 1)) \&\& y\_radar <= 1000)}\\
& \code{=> altitude == 0)}
\end{align*}
Next, we lift this implication to operate over $\Diamond$.
By taking the contrapositive of Theorem~\ref{thm:k} and applying Theorem~\ref{thm:box-diamond} we can show that
$
(\forall \sigma. \; \sigma \vDash p_1 \Rightarrow \sigma \vDash p_2)
\Rightarrow (\forall \beta. \; \beta \vDash \Diamond p_1 \Rightarrow \beta \vDash \Diamond p_2)
$.
Applying this, we see that
\begin{align*}
\forall \beta . \; \beta \vDash & \code{altitude != 0} \Rightarrow\\
\beta \vDash & \; \Diamond \ncode{(y\_radar > 1000) || } \\
& \; \Diamond \ncode{((y1 != 1 || y\_td1 != 1) \&\& (y2 != 1 || y\_td2 != 1)) ||} \\
& \; \Diamond \ncode{!( (((y1 == 1 \&\& y\_td1 == 1) || (y2 == 1 \&\& y\_td2 == 1)) \&\& y\_radar <= 1000)}\\
& \code{=> altitude == 0)}
\end{align*}
We start by addressing the final term of the disjunction. This is equivalent to saying the belief state satisfies
\begin{align*}
& \ncode{!}\Box \code{((((y1 == 1 \&\& y\_td1 == 1) || (y2 == 1 \&\& y\_td2 == 1)) \&\& y\_radar <= 1000)}\\
& \code{=> altitude == 0)}
\end{align*}
While we omit the full derivation, we note that from $\prop_\exists^{\mathrm{MPL}}$ we can derive that the belief state satisfies the negated proposition:
\begin{align*}
& \Box \code{((((y1 == 1 \&\& y\_td1 == 1) || (y2 == 1 \&\& y\_td2 == 1)) \&\& y\_radar <= 1000)}\\
& \code{=> altitude == 0)}
\end{align*}
Thus, by contradiction, we can disregard this last term of the disjunction.
Because the remaining terms only contain quantified variables, we can apply Theorem~\ref{thm:envindep} to see that
\begin{align*}
\forall \beta . \; \beta \vDash & \rcode{altitude != 0} \Rightarrow\\
\beta \vDash & \; \Box \ncode{(y\_radar > 1000) || } \\
& \; \Box \ncode{((y1 != 1 || y\_td1 != 1) \&\& (y2 != 1 || y\_td2 != 1))} \\
\end{align*}
We now consider each case of the disjunction separately.
In the first case, the approach is to show that for all $\beta$,
\[
\beta \vDash \Box \rcode{(y\_radar > 1000)} \Rightarrow
\beta \vDash \Box \rcode{(altitude > 0)} \Rightarrow
\beta \vDash \Box \ncode{(time\_on\_ground == 0)}
\]
where the first implication uses the assumptions on Lines \ref{code:yradar} and \ref{code:radaracc} of $\prop_\exists^\mathrm{MPL}$ and the second implication uses the result from Section~\ref{sec:togalt}.
In the second case, we will assume that the belief state satisfies \lstinline[breaklines=true]{$\Box$(y_radar <= 1000)}, because otherwise by Theorem~\ref{thm:envindep} we would have
\[
\beta \vDash \Diamond\rcode{(y\_radar > 1000)} \Rightarrow \beta \vDash \Box\rcode{(y\_radar > 1000)}
\]
which by the above argument must ensure the result. Because of this, we can assume without reservation that \lstinline[breaklines=true]{$\Box$(landing_leg_deployment == 0)}.
We will also assume for now that \lstinline[breaklines=true]{$\Box$(permanent_1 != 1)}.
In any environment $\sigma$ in a belief state satisfying this, we have that
\begin{align*}
\sigma \vDash & \code{y\_td1 != 1 \&\& transient\_on\_touchdown == 0} \Rightarrow\\
\sigma \vDash & \code{altitude != 0} \Rightarrow \sigma \vDash \code{time\_on\_ground == 0}
\end{align*}
Where the first implication uses assumptions from Lines \ref{code:totasgn} and \ref{code:tdiasgn} and the second one uses the result from Section~\ref{sec:togalt}. Similarly, we have that
\begin{align*}
\sigma \vDash & \code{y\_t1 != 1 \&\& transient\_on\_touchdown == 0} \Rightarrow\\
\sigma \vDash & \code{altitude\_0 != 0} \Rightarrow%
\sigma \vDash \code{time\_on\_ground\_0 == 0} \Rightarrow \\
\sigma \vDash & \code{time\_on\_ground == 0}
\end{align*}
This means that in any belief state satisfying $\prop_\exists^\mathrm{MPL}$, we will have that
\[
\Box\ncode{(transient\_on\_touchdown == 0 => time\_on\_ground == 0)}
\]
If the assumption \lstinline[breaklines=true]{$\Box$(permanent_1 != 1)} is not true, then by the assumption on Line~\ref{code:perm} we can assume that \lstinline[breaklines=true]{$\Box$(permanent_2 != 1)} and apply a symmetric line of reasoning to \lstinline[breaklines=true]{y_td2} and \lstinline[breaklines=true]{y2}.
\subsubsection{Bounded time-on-ground}
We will now show that
\[
\beta \vDash \prop_\exists^\mathrm{MPL} \Rightarrow \beta \vDash \Box
\ncode{(transient\_on\_touchdown => time\_on\_ground < 2)}
\]
We can summarize the proposition on Line~\ref{code:togasgn} as
\begin{align*}
\forall \beta. \; \beta \vDash & \Box \ncode{%
(engine\_enabled\_0 == 1 => time\_on\_ground == time\_on\_ground\_0 + 1 \&\&}\\
& \rcode{engine\_enabled\_0 != 1 => time\_on\_ground == time\_on\_ground\_0)}
\end{align*}
We can now apply the fact that
\begin{align*}
\forall \sigma. \; \sigma \vDash & \code{transient\_on\_touchdown == 0}\\
& \code{engine\_enabled\_0 == 1 => time\_on\_ground == time\_on\_ground\_0 + 1 \&\&}\\
& \code{engine\_enabled\_0 != 1 => time\_on\_ground == time\_on\_ground\_0 \&\&}\\
& \code{time\_on\_ground\_0 >= 0 \&\&}\\
& \code{time\_on\_ground\_0 < 2 \&\&}\\
& \code{time\_on\_ground\_0 == 1 => engine\_enabled\_0 == 0}\\
\Rightarrow \sigma \vDash & \code{transient\_on\_touchdown => time\_on\_ground < 2}
\end{align*}
can be proven using standard techniques for inequlities and propositional logic. We can lift this to the $\Box$ modality using Theorems \ref{thm:lifting} and \ref{thm:k} and then further apply the assumptions on Lines \ref{code:tognonneg}, \ref{code:togezdis}, and \ref{code:togbound} to show the result.
\section{Proofs}
\label{sec:proofs}
\subsection{Theorem~\ref{thm:bsound}: Belief Soundness}
\label{sec:bsoundproof}
Proceed by structural induction on derivations of $\Downarrow$.
Individual cases are as follows:
\begin{itemize}
\item {\bf Nondeterministic If Statements.} We case split on which branch the true environment takes.
If it is the true branch, then $\sigma \in \beta_T$.
Then, by induction hypothesis $\sigma' \in \beta'_T \Rightarrow \sigma' in \beta'_T \cup \beta'_F$.
The case for the false branch is symmetric.
\item {\bf All other Statements.} This follows directly from the definition of the new belief state and the induction hypotheses.
\end{itemize}
\subsection{Theorem~\ref{thm:bprecise}: Belief Precision}
\label{sec:bpreciseproof}
Proceed by structural induction on derivations of $\Downarrow$.
In each case, we identify the choice of $\sigma_\beta$ given $\sigma_\beta'$ that satisfies the property.
\begin{itemize}
\item {\bf Assignment.} Choose $\sigma_\beta \in \beta$ such that $\tuple{\sigma_\beta, e} \Downarrow c_\beta$ and $\sigma_\beta' = \sigma_\beta[x \mapsto c_\beta]$.
\item {\bf Choose.} Choose $\sigma_\beta \in \beta$ such that
there exists a $c_\beta$ such that
$\sigma_\beta \vDash p[c_\beta/.]$ and
$\sigma_\beta' = \sigma_\beta[x \mapsto c_\beta]$.
\item {\bf Observe.} Choose $\sigma_\beta = \sigma_\beta'$.
\item {\bf Nondeterministic If.} Case split on whether $\sigma_\beta'$ is in $\beta_T'$ or $\beta_F'$.
If it is in $\beta_T'$, choose $\mu_T' = \sigma_\beta$ and apply the inductive hypothesis for $s_1$.
Otherwise, apply the inductive hypothesis for $s_2$.
In either case, the resulting $\sigma_\beta$ must be in $\beta$.
\item {\bf Other Statements.} Apply inductive hypotheses.
\end{itemize}
\subsection{Theorem~\ref{thm:bdeterm}: Belief Determinism}
\label{sec:bdetermproof}
First, we define the following \emph{prefix property}:
\begin{property}[Prefix]
If $\tuple{\beta, \sigma, s} \Downarrow \tuple{\beta' \; | \; o}$
and $\tuple{\beta, \sigma, s} \Downarrow \tuple{\beta' \; | \; o'}$
where $o \neq o'$, then neither is $o$ a prefix of $o'$ nor $o'$ a prefix of $o$.
\end{property}
The prefix property states that if a belief program can generate two different observation lists, the lists have to fundamentally differ and cannot simply be extensions of each other.
Proceed by structural induction on derivations of $\Downarrow$.
We strengthen the inductive hypothesis with the prefix property.
Specific cases are as follows:
\begin{itemize}
\item {\bf Sequencing.} We consider here the rule that produces non-\texttt{nil} observation lists through the concatenate operator.
First, we show the prefix property. If $o_1$ is different between the two executions, then the resulting concatenated list must satisfy the prefix property according to the mechanics of the concatenate operator.
Otherwise, we apply the determinism inductive hypothesis on $s_1$ to show that the intermediate belief state $\beta'$ is the same between executions.
This means we can directly apply the prefix property inductive hypothesis on $s_2$ to show the result.
Now, we show belief determinism.
We can assume the length of all observation lists in this rule are the same across both executions.
Otherwise, $o_1$ in one execution would be a prefix of $o_1$ in the other execution, violating the prefix property.
Due to the mechanics of the list concatenate operation, this means we can assume the values of the lists are the same across both executions.
Thus, we can apply the determinism inductive hypotheses to show the result.
\item {\bf Observe.} The list always has length 1, which ensures the prefix property.
The new belief state is a function of the original belief state and the observed value, which ensures belief determinism.
\item {\bf Other cases.} In the remaining cases, either the observation list is constant and the new belief state is a function of the original one, or the conclusion follows directly from induction hypotheses.
\end{itemize}
\subsection{Logic}
\subsubsection{Substitution Lemmas}
\label{sec:substproofs}
Here, we restate and reprove the substitution lemma from Section~\ref{sec:logic} and state and prove the other substitution lemmas we will need.
\begin{lemma}[Expression Substitution]
If $x_0$ is fresh in $e$, then $\tuple{\sigma, e} \Downarrow c \iff \tuple{\sigma[x_0 \mapsto \sigma(x)], e} \Downarrow c$ and
$\tuple{\sigma, e} \Downarrow c \iff \tuple{\sigma[x_0 \mapsto \sigma(x)], e[x_0/x]} \Downarrow c$
\end{lemma}
\begin{proof}
By structural induction on the language of expressions.
\end{proof}
\begin{lemma}[Proposition Substitution]
If $x_0$ is fresh in $p$, then $\sigma \vDash p \iff \sigma[x_0 \mapsto \sigma(x)] \vDash p$ and
$\sigma \vDash p \iff \sigma[x_0 \mapsto \sigma(x)] \vDash p[x_0/x]$
\end{lemma}
\begin{proof}
By structural induction on the language of propositions. The base case employs the expression substitution lemma above.
\end{proof}
In the following definition, we use the notation $\beta[x_0 \mapsto x]$ to mean a belief state $\beta'$ such that
$\beta' = \{ \sigma[x_0 \mapsto \sigma(x)] \; | \; \sigma \in \beta\}$
\begin{lemma}[Modal Proposition Substitution]
If $x_0$ is fresh in $\prop_\Box$, then $\beta \vDash \prop_\Box \iff \beta[x_0 \mapsto x] \vDash \prop_\Box$ and
$\beta \vDash \prop_\Box \iff \beta[x_0 \mapsto x] \vDash \prop_\Box[x_0/x]$
\end{lemma}
\begin{proof}
By structural induction on the language of propositions.
The base cases employ the proposition substitution lemma above.
\end{proof}
\subsubsection{Agreement}
\label{sec;agreeproofs}
In this section, we show some \emph{agreement lemmas} that are also required to show the main soundness result.
\begin{lemma}[Expression Agreement]
If $\sigma_1$ agrees with $\sigma_2$ on every location except $x_0$ and, $e$ does not depend on $x_0$, then $\tuple{\sigma_1, e} \Downarrow c \iff \tuple{\sigma_2, e} \Downarrow c$.
\end{lemma}
\begin{proof}
By structural induction on the language of expressions.
\end{proof}
\begin{lemma}[Proposition Agreement]
If $\sigma_1$ agrees with $\sigma_2$ on every location except $x_0$ and, $p$ does not depend on $x_0$, then $\sigma_1 \vDash p \iff \sigma_2 \vDash p$.
\end{lemma}
\begin{proof}
By structural induction on the language of propositions.
The base case follows from the expression agreement lemma above.
\end{proof}
\begin{lemma}[Modal Proposition Agreement]
If $\beta_1$ agrees with $\beta_2$ on every location except $x_0$
(i.e.\ $\sigma_1 \in \beta_1$ iff $\sigma_2 \in \beta_2$ and $\sigma_2$ agrees with $\sigma_1$ on every location except $x_0$), then
$\beta_1 \vDash \prop_\Box \iff \beta_2 \vDash \prop_\Box$
\end{lemma}
\begin{proof}
By structural induction on the language of modal propositions.
The base case follows from the proposition agreement lemma above, and we also use the fact that belief agreement is symmetric to prove the only if case.
\end{proof}
\subsubsection{Constant Substitution}
\label{sec:constsubstproofs}
The following is an additional lemma used in the proof of the soundness of Epistemic Hoare logic:
\begin{lemma}[Expression Constant Substitution]
If $\sigma(x) = c_\sigma$, then $\tuple{\sigma, e} \Downarrow c$ iff $\tuple{\sigma, e[c_\sigma/x]} \Downarrow c$.
\end{lemma}
\begin{proof}
By structural induction on the language of expressions.
\end{proof}
\begin{lemma}[Proposition Constant Substitution]
If $\sigma(x) = c_\sigma$, then $\sigma \vDash p$ iff $\sigma \vDash p[c_\sigma/x]$.
\end{lemma}
\begin{proof}
By structural induction on the language of propositions.
For the base cases, we apply the expression constant substitution lemma above.
\end{proof}
\begin{lemma}[Modal Proposition Substitution]
If for all $\sigma \in \beta$, $\sigma(x) = c$, and there exists a $\sigma \in \beta$ such that $\sigma(x) = c$, then $\beta \vDash \prop_\exists \iff \beta \vDash \prop_\exists[x/c]$.
\end{lemma}
\begin{proof}
By structural induction on the language of modal propositions.
For the base cases, we apply the proposition constant substitution lemma above
\end{proof}
\subsubsection{Proof of Theorem~\ref{thm:ehlsound}: Part 1}
\label{sec:ehlsoundproof1}
Proceed by structural induction on derivations of $\Downarrow$.
Specific cases are as follows:
\begin{itemize}
\item {\bf Assignment.} The first part of the post-condition follows from the agreement and substitution lemmas for modal propositions.
The second part follows from expression substitution.
\item {\bf Choose.} The first part of the post-condition follows from the agreement and substitution lemmas.
Then, we can show that since $p[x_0/x]$ does not depend on $x$,
$\sigma \vDash p[x_0/x][./c] \iff
\sigma[x \mapsto c] \vDash p[./x]$.
This can be shown by structural induction, similarly to the proof for the agreement and substitution lemmas.
The second part of the post-condition follows from this fact.
\item {\bf Nondeterministic If.}
By definition, $\beta_T \vDash
\exists \hat{y}. \; \ncode{(}\Box p \ncode{)}
\code{\&\&} \prop_\Box$ and
$\beta_F \vDash
\exists \hat{y}. \; \ncode{(}\Box \ncode{!}p \ncode{)}
\code{\&\&} \prop_\Box$
After applying inductive hypotheses, we can use the fact that $\beta_T \vDash \Box p \wedge \beta_F \vDash \Box p \Rightarrow \beta_T \cup \beta_F \vDash \Box p$ to show the overall result.
\item {\bf Other Cases.} The other cases follow from applying inductive hypotheses and inlining definitions.
In the case of while loops, the premises must be destructed, and the extra premises of the logic rule ensure that the semantics do not evaluate to $\bot$.
\end{itemize}
\subsubsection{Proof of Theorem~\ref{thm:ehlsound}: Part 2}
\label{sec:ehlsoundproof2}
Proceed by structural induction on derivations of $\Downarrow$.
Most cases follow the same reasoning as part 1 above, and we cover here the cases that are unique to part 2.
\begin{itemize}
\item {\bf Observe.} Choose $y_{n+1} = \sigma(x)$.
We show the first part of the post-condition by applying the constant substitution lemma for modal propositions.
The first premise of this lemma follows from the definition of $\beta'$, and the second condition follows from the assumption that $\sigma \in \beta$.
The second part of the post-condition follows directly from the definition of $\beta'$.
\item {\bf Subtyping.} Because this part of the theorem is a strengthened version of the first part, any derivation produced by subtyping must be sound.
\item {\bf Other Cases.} The result follows from applying inductive hypotheses.
\end{itemize}
\section{Theorems About Propositions}
\label{sec:proptheorems}
In addition to the logical rules in Figure~\ref{fig:logic}, developers need to prove the implications in the premises of those rules.
We propose that developers do so by lifting propositional reasoning to the modal operators.
The theorems in this section give a general set of reasoning tools for performing this lifting.
First, we show that modal operator $\Box$ commutes with conjunctions:
\begin{theorem}[$\Box$ commutes with $\rcode{\&\&}$]
$\beta \vDash \Box \ncode{(} p_1 \code{\&\&} p_2 \ncode{)} \iff
\beta \vDash \Box p_1 \code{\&\&} \Box p_2$
\label{thm:box-and}
\end{theorem}
\begin{proof}From first-order logic instantiation.\end{proof}
We can similarly show that modal operator $\Box$ partially commutes with disjunctions, but the implication only applies in one direction
\begin{theorem}[$\Box$ commutes with $\rcode{||}$]
$\beta \vDash \Box p_1 \code{||} \Box p_2 \Rightarrow
\beta \vDash \Box \ncode{(} p_1 \code{||} p_2 \ncode{)}$
\label{thm:box-or}
\end{theorem}
\begin{proof}By case analysis on the premise. Follows from the fact that $\sigma \vDash p_1 \Rightarrow \sigma \vDash p_1 \code{||} p_2$.\end{proof}
The $\Box$ and $\Diamond$ operators are also related by negation, in the same way that $\forall$ and $\exists$ are related in first-order logic.
This means that, while we present all theorems in this section as applying to $\Box$, there are analogous theorems that apply to $\Diamond$.
\begin{theorem}[$\Box$-$\Diamond$ duality]
$\beta \vDash \Box p \iff \beta \vDash \ncode{!} \Diamond \ncode{!} p$
\label{thm:box-diamond}
\end{theorem}
\begin{proof}Follows from definitions and $\forall$-$\exists$ duality.\end{proof}
Next, we show two theorems that enable developers to push implications through modal operators.
The first states that the $\Box$ modality applies to any formulas that are true across all environments.
\begin{theorem}[Knowledge of Theorems]
$\Big( \forall \sigma. \; \sigma \vDash p_1 \Rightarrow \sigma \vDash p_2 \Big)
\Rightarrow \Big( \beta \vDash \Box \ncode{(!} p_1 \code{||} p_2 \ncode{)} \Big)$
\label{thm:lifting}
\end{theorem}
\begin{proof}Follows from $\beta$ being a subset of all possible $\sigma$.\end{proof}
The second theorem states that implications under $\Box$ can be lifted to modal propositions.
\begin{theorem}[Knowledge Instantiation]
$\beta \vDash \Box \ncode{(!} p_1 \code{||} p_2 \ncode{)}
\Rightarrow \Big( \beta \vDash \Box p_1 \Rightarrow \beta \vDash \Box p_2 \Big)$
\label{thm:k}
\end{theorem}
\begin{proof} Follows from first-order instantiation and transitivity of $\Rightarrow$.\end{proof}
Finally, we show that for propositions that do not depend on environment variables (i.e.\ they only depend on quantified variables and constants), truth in one environment implies truth in any.
\begin{theorem}[Environment Independence]{\ \\}
\vspace{-.4cm}
\begin{center}
If $p$ does not contain any $x$, then
$\beta \vDash \exists \hat{y}. \; \prop_\Box \code{\&\&} \Diamond p \Rightarrow
\beta \vDash \exists \hat{y}. \; \prop_\Box \code{\&\&} \Box p$
\end{center}
\vspace{-.3cm}
\label{thm:envindep}
\end{theorem}
\begin{proof}By structural induction, we can show that $\sigma \vDash p$ is independent of $\sigma$. Thus, if $p$ is true, it is true under all environments.\end{proof}
\section{Related Work}
\label{sec:related}
\paragraph{Set-based Uncertainty.}
\cite{signalsets} is a survey paper that gives an overview of how set-based uncertainty is used in the signal processing domain. It explains how programs can over-approximate the true belief state (using e.g. ellipsoids~\cite{schweppe}; more recent work has studied polytopes~\cite{polytopes}), and how the quality of approximation can be measured~\cite{schweppe} to determine if the resulting belief state is too large. It also gives efficient algorithms~\cite{kaczmarz, cimmino} for a restricted set of operations on approximate belief states . By contrast, belief programming reasons about the exact belief state and provides a richer set of operations. However, it cannot achieve the same computational efficiency as an approximation.
\paragraph{Classical Verification}
In Sections~\ref{sec:example} and \ref{sec:casestudy}, we alluded to how the UAV and MPL examples could be verified using classical techniques.
Here, we explain this process in more detail.
The developer would first compose their handwritten environment model (Figures~\ref{fig:dronemodel} and \ref{fig:mplmodel}) with their handwritten state estimator (Figures~\ref{fig:dronecode} and \ref{fig:mplcode}).
The resulting program is in the language IMP~\cite{winskel} with the addition of choose statements that provide nondeterminism.
The developer could obtain a Hoare logic for this language by either extending the logic of IMP~\cite{winskel} to include choose statements, or by rewriting it to a language such as GCL~\cite{gcl} which supports nondeterminism natively and also has a Hoare logic.
Finally, the developer would apply the Hoare logic to the program, which requires discharging verification conditions. Because the proposition language is the standard propositional calculus, we expect there are many techniques in the literature that the developer could apply to this end.
The advantage of classical verification, relative to belief programming, is that developers can expect to draw on a wide variety of verification literature to solve the problem, as the formulation is relatively standard.
The disadvantage is that developers must write the raw code of the state estimator by hand, which is a tedious process.
By contrast, in belief programming, the state estimator consists of \texttt{infer} statements, which are more intuitive to use.
\paragraph{Epistemic and Belief Revision Logics.} The Epistemic Hoare Logic we present in Section~\ref{sec:logic} is similar to dynamic Epistemic logics such as public announcement logic~\citep{pal} and action-based logics~\citep{del}.
Whereas these logics typically use either propositions or abstract action spaces to modify the belief state, our logic uses a belief program to do so.
\paragraph{Synthesis.} To avoid writing state estimators by hand, developers might generate a state estimator by synthesizing it directly from the environment model~\citep{cegis, vsa, dedsynth}.
Such a problem would be challenging due to having hidden state encoded in the belief program that must be explicit in the state estimator.
For example, in the desired handwritten state estimator in Figure~\ref{fig:mplcode}, the program contains the variables \texttt{prev\_td\_1} and \texttt{prev\_td\_2}, which are not related to any of the variables in the model in Figure~\ref{fig:mplmodel}.
We anticipate that existing synthesis tools will have difficulty automatically inferring the existence of such hidden variables.
\paragraph{Dynamic Constraint Solving.} Some existing programming languages~\citep{jeeves,planb} perform constraint solving at runtime using an SMT solver.
Such approaches are necessarily similar to a belief program, which also represents a constrained set of program states at runtime.
Existing systems were designed for other domains, and do not articulate complete set of \texttt{choose}, \texttt{observe}, and \texttt{infer} constructs that BLIMP{} has.
\subsection{Relationship to Probabilistic Programming}
Probabilistic programming languages (PPLs) are also designed to enable developers to reason about uncertainty. We compare BLIMP to PPLs on three core axes: language features, contemporary reasoning techniques, and the practicality of inference (i.e., the implementation of \texttt{infer}) .
\paragraph{Language Features} BLIMP's primary programming constructs are \texttt{choose}, \texttt{observe}, and \texttt{infer} and have probabilistic analogs in probabilistic programming languages. BLIMP's \texttt{choose} has the classic interpretation of non-probabilistic, nondeterministic choice. The analog in probabilistic programming is the probabilistic \texttt{sample} construct that randomly samples a value according to a distribution. BLIMP's \texttt{observe} has a similar semantics to \texttt{observe} constructs in PPLs. BLIMP's \texttt{infer}, which enables the program itself to perform inference, has some support in PPLs as well~\citep{staton,probzelus}.
In general, probabilistic programming provides a more flexible modeling mechanism than BLIMP's set-based uncertainty, enabling a developer to specify distributions for nondeterministic outcomes. For applications for which probabilistic models of outcomes are available, probabilistic programming can be an appropriate choice. However, probabilistic models of outcomes are not available for all applications (e.g., our MPL application) and introducing distributions can complicate reasoning, as we discuss below. BLIMP is an additional design point for applications that do not necessarily benefit from probabilistic modeling.
\paragraph{Reasoning} The verification of probabilistic programs is an active area of research~\cite{passert, probverif}. This work typically provides the ability to express and verify the probability that an assertion is true of the program. In principle, it is possible to verify BLIMP-like modal assertions in such a framework. Specifically, $\Box$ maps to an assertion that a proposition is true with probability 1; $\Diamond$ maps to an assertion that a proposition is true with positive probability.
However, a significant challenge with reasoning about probabilistic programs is that the distribution over states at any given point in a program may not have an analytical characterization. Specifically, the composition of a standard, well-known probability distribution (e.g., a Gaussian distribution) with a computation can result in a distribution that is not well-characterized by a standard, well-known distribution for which standard statistical quantities (such as mean and variance) are easily accessible.
Therefore the full modeling, programming, and reasoning workflow must carefully consider the distributions used in \texttt{sample} statements such that they adhere to the application's uncertainty model and that the resulting distributions in the rest of the program can be precisely reasoned about at an acceptable level of complexity.
BLIMP is, instead, an additional design point for modeling uncertainty that need not rely on an appropriate selection of \texttt{sample} distributions or, more generally, the complexity of techniques for reasoning about probabilistic constructs.
\paragraph{Practicality of Inference.} BLIMP's runtime implementation to support inference -- i.e.\ \texttt{infer} -- tracks all possible environments. The resulting implementation is sound. A probabilistic programming language can take a similar strategy -- i.e., exact inference -- if it desires sound inference.
However, probabilistic programming languages can also leverage approximate inference algorithms for probabilistic programs such as Sequential Monte Carlo methods~\cite{doucet-smc} that need only track a high-probability subset of the possible environments. These methods are approximate in that the probability of an event is estimated with a fidelity that is a function of the size and diversity of the tracked state. Designing algorithms to select this state efficiently is application-specific. Contemporary diagnostics for Monte Carlo methods (e.g.\ \cite{ess, mcmcdiag}) are typically not sound in that they can indicate a good approximation when the true approximation is poor. Establishing bounds on the quality of the resulting approximation is still an active area of research~\cite{issample}.
Therefore, while approximate inference methods can in practice provide empirically good results, reasoning about the soundness of their results is still an active area of research.
\subsection{Semantics}
\label{sec:semantics}
In this section, we illustrate how belief programming works by presenting the semantics of BLIMP{}.
The semantics precisely specify how BLIMP{} manipulates belief states with the goal that the belief states always capture all possible behaviors and only capture realizable behaviors.
\subsubsection{Preliminaries}
\paragraph{Environments.} An environment $\sigma \in \Sigma = (X \rightarrow \mathcal{C})$ is a finite map from program variable names to their numeric values, where each value belongs to a finite subset $\mathcal{C}$ of the integers.
We use the notation $\sigma[x \mapsto c]$ to mean the environment $\sigma$ with the variable $x$ mapped to the value $c$.
An \emph{optional environment} $\mu \in \Sigma \cup \{ \cdot \}$ may be either an environment $\sigma$ or a null value $\cdot$.
The use of null optional environments is described in more detail in Section~\ref{sec:statementsemantics}.
\paragraph{Belief States.} A belief state $\beta \in \mathcal{P}(\Sigma)$ is an element of the powerset of $\Sigma$, i.e.\ it is a set of environments.
The interpretation is that if an environment $\sigma$ is in $\beta$, then the program believes that $\sigma$ is a possibly true environment.
\subsubsection{Expressions}
\begin{figure}
\begin{mathpar}
\inferrule{ }
{\tuple{\sigma, x} \Downarrow \sigma(x)}
\inferrule{ }
{\tuple{\sigma, c} \Downarrow c}
\inferrule{ \tuple{\sigma, e_0} \Downarrow c_0\\\\
\tuple{\sigma, e_1} \Downarrow c_1
}
{
\tuple{\sigma, e_0 \code{+} e_1} \Downarrow c_0 + c_1
}
\inferrule{ \tuple{\sigma, e_0} \Downarrow c_0\\\\
\tuple{\sigma, e_1} \Downarrow c_1
}
{
\tuple{\sigma, e_0 \code{-} e_1} \Downarrow c_0 - c_1
}
\inferrule{ \tuple{\sigma, e_0} \Downarrow c_0\\
\tuple{\sigma, e_1} \Downarrow c_1
}
{
\tuple{\sigma, e_0 \code{*} e_1} \Downarrow c_0 * c_1
}
\inferrule{ \tuple{\sigma, e_0} \Downarrow c_0\\
\tuple{\sigma, e_1} \Downarrow c_1
}
{
\tuple{\sigma, e_0 \code{/} e_1} \Downarrow c_0 / c_1
}
\end{mathpar}
\caption{Semantics of expressions. We use the notation $\tuple{\sigma, e} \Downarrow c$ to mean that the expression $e$ evaluated under the environment $\sigma$ yields the value $c$.}
\label{fig:exprsemantics}
\end{figure}
Figure~\ref{fig:exprsemantics} presents the semantics of expressions.
Our approach is a big-step operational semantics that states what value an expression computes when evaluated under a given environment.
We use the notation $\tuple{\sigma, e} \Downarrow c$ to mean that the expression $e$ evaluated under the environment $\sigma$ yields the value $c$.
The meaning of a variable $x$ is the value of the variable in the input environment $\sigma$.
The meaning of a constant $c$ is the value of the constant.
The arithmetic operators $\ncode{+}$, $\ncode{-}$, $\ncode{*}$, and $\ncode{/}$ have their standard interpretation.
We assume that $\ncode{/}$ denotes integer division.
\subsubsection{Propositions}
\begin{figure}
\begin{align*}
\sigma \vDash b & \iff b = \mathrm{\texttt{true}}\\
\sigma \vDash p_1 \code{\&\&} p_2 & \iff%
\sigma \vDash p_1 \wedge \sigma \vDash p_2\\
\sigma \vDash p_1 \code{||} p_2 & \iff%
\sigma \vDash p_1 \vee \sigma \vDash p_2\\
\sigma \vDash \rcode{!} p & \iff \sigma \not \vDash p\\
\sigma \vDash e_1 \code{<} e_2 & \iff%
\mathrm{\textnormal{there exist}} \; c_1, c_2 \; \mathrm{s.t.} \;%
\tuple{\sigma, e_1} \Downarrow c_1 \wedge%
\tuple{\sigma, e_2} \Downarrow c_2 \wedge%
c_1 < c_2\\
\sigma \vDash e_1 \code{==} e_2 & \iff%
\mathrm{\textnormal{there exist}} \; c_1, c_2. \; \mathrm{s.t.} \;%
\tuple{\sigma, e_1} \Downarrow c_1 \wedge%
\tuple{\sigma, e_2} \Downarrow c_2 \wedge%
c_1 = c_2
\end{align*}
\caption{Semantics of propositions. We use the notation $\sigma \vDash p$ to mean $\sigma$, which must be an environment, satisfies $p$, a proposition.}
\label{fig:propsemantics}
\end{figure}
Figure~\ref{fig:propsemantics} presents the semantics of propositions.
The meaning of a proposition is whether or not, for a given environment, the environment satisfies the proposition (i.e.\ the proposition is true in the environment).
We use the notation $\sigma \vDash p$ to mean that the environment $\sigma$ satisfies the proposition $p$.
The meaning of an equality $\ncode{==}$ or size $\ncode{<}$ comparison is that an environment satisfies the comparison iff evaluating the expressions yields numbers that satisfy the comparison.
Note that the existential quantifiers in the definition are trivial because the expression semantics is a total function of the environment.
The meaning of each of the operators $\ncode{\&\&}$, $\ncode{||}$, and $\ncode{!}$ is its standard interpretation in propositional logic.
\subsubsection{Modal and Existential Propositions}
\begin{figure}[ht]
\begin{subfigure}{0.45\textwidth}
\begin{align*}
\beta \vDash \Box p & \iff%
\mathrm{\textnormal{for every}} \; \sigma \in \beta, \; \sigma \vDash p\\
\beta \vDash \Diamond p & \iff%
\mathrm{\textnormal{there is some}} \; \sigma \in \beta \; \mathrm{s.t.} \; \sigma \vDash p
\end{align*}
\end{subfigure}
\begin{subfigure}{0.45\textwidth}
\begin{align*}
\beta \vDash \exists . \; \prop_\Box & \iff%
\beta \vDash \prop_\Box\\
\beta \vDash \exists y_0, \hat{y}'. \; \prop_\Box & \iff%
\mathrm{\textnormal{there is some}} \; c \; \mathrm{s.t.} \;%
\beta \vDash \exists \hat{y}'. \; \prop_\Box[c/y_0]
\end{align*}
\end{subfigure}
\caption{Semantics of modal and existential propositions. We use the notation $\beta \vDash \prop_\Box$ and $\beta \vDash \prop_\exists$ to mean that the belief state $\beta$ satisfies $\prop_\Box$ or $\prop_\exists$.}
\label{fig:modalpropsemantics}
\end{figure}
Figure~\ref{fig:modalpropsemantics} presents the semantics of modal and existential propositions.
The meaning of a modal or existential proposition is whether or not a given belief state satisfies the proposition (i.e.\ the proposition is true in the belief state).
We use the notation $\beta \vDash \prop_\Box$ and $\beta \vDash \prop_\exists$ to mean that the belief state $\beta$ satisfies $\prop_\Box$ or $\prop_\exists$.
The meaning of $\Box$ is to universally quantify over all environments in the belief state, and the meaning of $\Diamond$ is to existentially quantify over environments in the belief state.
The meaning of the operators $\ncode{\&\&}$, $\ncode{||}$, and $\ncode{!}$ (elided in Figure~\ref{fig:modalpropsemantics}) is their standard propositional logic interpretations, and is the same as Figure~\ref{fig:propsemantics} with $\sigma$ replaced with $\beta$.
The meaning of $\exists$ is its standard meaning in first-order logic.
We use notation $\prop_\Box[c/y_0]$ to mean the proposition $\prop_\Box$ with $c$ substituted for $y_0$, and the notation $\exists. \; \prop_\Box$ to mean an existential proposition with an empty set of quantified variables.
\subsubsection{Statements}
\label{sec:statementsemantics}
\begin{figure}
\begin{mathpar}
\inferrule
{
\beta' = \{\sigma_\beta[x \mapsto c_\beta] \; | \;
\sigma_\beta \in \beta \wedge
\tuple{\sigma_\beta, e} \Downarrow c_\beta
\}
}
{
\tuple{\beta, \cdot, x \code{=} e} \Downarrow
\tuple{\beta', \cdot \; | \; \ncode{nil}}
}
\inferrule
{
\tuple{\beta, \cdot, x \code{=} e} \Downarrow
\tuple{\beta', \cdot \; | \; \ncode{nil}}\\
\tuple{\sigma, e} \Downarrow c
}
{
\tuple{\beta, \sigma, x \code{=} e} \Downarrow
\tuple{\beta', \sigma[x \mapsto c] \; | \; \ncode{nil}}
}
\inferrule
{
\beta' = \{ \sigma_\beta[x \mapsto c_\beta] \; | \;
\sigma_\beta \in \beta \wedge
\sigma_\beta \vDash p[c_\beta/\ncode{.}]
\}
}
{
\tuple{\beta, \cdot, x \rcode{= choose(} p \ncode{)}}
\Downarrow \tuple{\beta', \cdot %
\; | \; \ncode{nil}}
}
\inferrule
{
\tuple{\beta, \cdot, x \rcode{= choose(} p \ncode{)}}
\Downarrow \tuple{\beta', \cdot %
\; | \; \ncode{nil}}\\\\
\sigma \vDash p[c/\ncode{.}]
}
{
\tuple{\beta, \sigma, x \rcode{= choose(} p \ncode{)}}
\Downarrow \tuple{\beta', \sigma[x \mapsto c] %
\; | \; \ncode{nil}}
}
\inferrule
{
\beta \vDash \prop_\Box
}
{
\tuple{\beta, \mu, \lcode{assert} \prop_\Box} \Downarrow
\tuple{\beta, \mu \; | \; \ncode{nil}}
}
\inferrule
{
\beta \not \vDash \prop_\Box
}
{
\tuple{\beta, \mu, \lcode{assert} \prop_\Box} \Downarrow
\tuple{\bot \; | \; \ncode{nil}}
}
\inferrule
{
\beta' = \{ \sigma_\beta \; | \; \sigma_\beta \in \beta
\wedge \sigma_\beta(x) = \sigma(x)\}
}
{
\tuple{\beta, \sigma, \lcode{observe} x} \Downarrow
\tuple{\beta', \sigma \; | \; x \code{:} \sigma(x)}
}
\inferrule
{ }
{
\tuple{\beta, \cdot, \lcode{observe} x} \Downarrow
\tuple{\bot \; | \; \ncode{nil}}
}
\inferrule
{
\beta \vDash \Diamond p \code{\&\&} \Diamond \ncode{(!}p\ncode{)}\\\\
\beta_T = \{ \sigma_\beta \; | \;
\sigma_\beta \in \beta \wedge
\sigma_\beta \vDash p \}\\
\beta_F = \{ \sigma_\beta \; | \;
\sigma_\beta \in \beta \wedge
\sigma_\beta \not \vDash p \}\\\\
\mu_T = \sigma \; \mathrm{if} \; \mu = \sigma \wedge
\sigma \vDash p \; \mathrm{else} \; \cdot \\
\mu_F = \sigma \; \mathrm{if} \; \mu = \sigma \wedge
\sigma \not \vDash p \; \mathrm{else} \; \cdot \\\\
\tuple{\beta_T, \mu_T, s_1} \Downarrow
\tuple{\beta_T', \mu_T' \; | \; \ncode{nil}}\\
\tuple{\beta_F, \mu_F, s_2} \Downarrow
\tuple{\beta_F', \mu_F' \; | \; \ncode{nil}}\\\\
\mu' = {\begin{cases}
\mu_T' & \mu = \sigma \wedge \sigma \vDash p\\
\mu_F' & \mu = \sigma \wedge \sigma \not \vDash p\\
\cdot & \mathrm{else}
\end{cases}}
}
{
\tuple{\beta, \mu, \lcode{if} p \code{\{} s_1
\code{\} else \{} s_2 \rcode{\}}} \Downarrow
\tuple{\beta_T' \cup \beta_F',
\mu' \; | \; \ncode{nil}}
}
\inferrule
{
\beta \vDash \Diamond p \code{\&\&} \Diamond \ncode{(!}p\ncode{)}\\\\
\beta_T = \{ \sigma_\beta \; | \;
\sigma_\beta \in \beta \wedge
\sigma_\beta \vDash p \}\\
\beta_F = \{ \sigma_\beta \; | \;
\sigma_\beta \in \beta \wedge
\sigma_\beta \not \vDash p \}\\\\
\mu_T = \sigma \; \mathrm{if} \; \mu = \sigma \wedge
\sigma \vDash p \; \mathrm{else} \; \cdot \\
\mu_F = \cdot \; \mathrm{if} \; \mu = \sigma \wedge
\sigma \vDash p \; \mathrm{else} \; \sigma \\\\
\tuple{\beta_T, \mu_T, s_1} \Downarrow
\tuple{\beta_T', \mu_T' \; | \; o_T}\\
\tuple{\beta_F, \mu_F, s_2} \Downarrow
\tuple{\beta_F', \mu_F' \; | \; o_F}\\\\
o_T \neq \ncode{nil} \vee o_F \neq \ncode{nil}
}
{
\tuple{\beta, \mu, \lcode{if} p \code{\{} s_1
\code{\} else \{} s_2 \rcode{\}}} \Downarrow
\tuple{\bot \; | \; \ncode{nil}}
}
\inferrule
{
\beta \vDash \Box p\\
\tuple{\beta, \mu, s_1} \Downarrow
\tuple{C_1 \; | \; o}
}
{
\tuple{\beta, \mu, \lcode{if} p \code{\{} s_1
\code{\} else \{} s_2 \rcode{\}}} \Downarrow
\tuple{C_1 \; | \; o}
}
\inferrule
{
\beta \vDash \Box \ncode{(!} p \ncode{)}\\
\tuple{\beta, \mu, s_2} \Downarrow
\tuple{C_2 \; | \; o}
}
{
\tuple{\beta, \mu, \lcode{if} p \code{\{} s_1
\code{\} else \{} s_2 \rcode{\}}} \Downarrow
\tuple{C_2 \; | \; o}
}
\inferrule
{
\beta \vDash \prop_\Box\\
\tuple{\beta, \mu, s_1} \Downarrow \tuple{C_1 \; | \; o}
}
{
\tuple{\beta, \mu, \lcode{infer} \prop_\Box \code{\{}
s_1 \code{\} else \{} s_2 \rcode{\}}}
\Downarrow \tuple{C_1 \; | \; o}
}
\inferrule
{
\beta \not \vDash \prop_\Box\\
\tuple{\beta, \mu, s_2} \Downarrow \tuple{C_2 \; | \; o}
}
{
\tuple{\beta, \mu, \lcode{infer} \prop_\Box \code{\{}
s_1 \code{\} else \{} s_2 \rcode{\}}}
\Downarrow \tuple{C_2 \; | \; o}
}
\inferrule
{
\tuple{\beta, \mu,
\lcode{assert} \prop_\exists^I
\rcode{; if(}p \ncode{) \{} s \code{;}
\lcode{while} p \code{\{} \prop_\exists^I
\code{\} \{} s \rcode{\}}
\lcode{\} else \{ skip \}}
} \Downarrow \tuple{C \; | \; o}
}
{
\tuple{\beta, \mu,
\lcode{while} p \code{\{} \prop_\exists^I
\code{\} \{} s \rcode{\}}} \Downarrow
\tuple{C \; | \; o}
}
\inferrule
{
\tuple{\beta, \mu, s_1} \Downarrow
\tuple{\beta', \mu' \; | \; o_1}\\\\
\tuple{\beta', \mu', s_2} \Downarrow
\tuple{\beta'', \mu'' \; | \; o_2}
}
{
\tuple{\beta, \mu, s_1 \code{;} s_2}
\Downarrow \tuple{\beta'',\mu'' \; | \; o_1 \concat o_2}
}
\inferrule{ }
{
\tuple{\beta, \mu, \ncode{skip}} \Downarrow
\tuple{\beta, \mu \; | \; \ncode{nil}}
}
\end{mathpar}
\caption{Semantics of statements. We use the notation $\tuple{\beta, \mu, s} \Downarrow \tuple{C \; | \; o}$ to mean that the belief state $\beta$ and optional true environment $\mu$ produce the configuration $C$ augmented with the observation list $o$.}
\label{fig:statementsemantics}
\end{figure}
Figure~\ref{fig:statementsemantics} presents the semantics of statements.
We follow a big-step operational approach, where every statement updates the belief state as well as an optional true environment.
While the program execution only updates the belief state, we model programs as simultaneously nondeterministically updating the true environment to provide an easy specification of the set of legal observation inputs.
Namely, a legal observation is one that could have come from the true environment.
A null true environment signifies that, due to nondeterministic control flow, the true environment took a different branch than the belief state.
Observations under a null true environment are illegal.
The nondeterminism of the true environment impacts the belief state through the observations, but for fixed observation inputs, the belief state update is deterministic.
Every statement produces, given an initial belief state and true environment, a {\em configuration}, which is either a new belief state and new optional true environment or an error $\bot$.
We augment the result with an observation list, which documents all observations and is an element of the grammar
\begin{align*}
O \; \mathrm{\texttt{::=}} \; & x : c :: O \; \mid \; \ncode{nil}
\end{align*}
In other words, an observation list is a list of associations of variable names to values.
We use the notation $\tuple{\beta, \mu, s} \Downarrow \tuple{C \; | \; o}$ to mean that the belief state $\beta$ and optional true environment $\mu$ produce the configuration $C$ augmented with the observation list $o$.
\paragraph{Assignment and Choose.} The meaning of either an assignment or choose statement is a new environment, if an original environment was present, and belief state with the statement variable rebound to a new value.
In the case of assignment, the value is given by evaluating the expression.
In the case of a choose statement the value must be consistent with the proposition, meaning that if we replace the placeholder $\ncode{.}$ with the new value, the proposition must hold.
In both cases, the new belief state is obtained by applying this process to each environment in the initial belief state.
A choose statement is nondeterministic with respect to the true environment, but deterministic with respect to the belief state.
\paragraph{Assertions.} The meaning of an assertion, if its predicate is true, is to return the input belief state and environment.
If its predicate is false, the assertion returns an error.
\paragraph{Observations.} An observation $\lcode{observe} x$ does not modify the true environment.
It does modify the belief state to be consistent with the true environment on the observed variable $x$ by only keeping those environments in the initial belief state that have the same value for $x$ as in the true environment.
The semantics also specify that the value of $x$ is in the observation list, and that an error occurs if the true environment is null.
\paragraph{If Statements.}
If the if statement's condition is deterministic (i.e., it is either true in all environments in the belief state or false in all environments), then the execution takes the appropriate branch.
This is specified by semantic rules that require $\beta \vDash \Box p$ or $\beta \vDash \Box \ncode{(!}p\ncode{)}$, where $\beta$ is the initial belief state and $p$ is the if statement's condition.
If the statement's condition is nondeterministic (as specified by requiring $\beta \vDash \Diamond p \code{\&\&} \Diamond \ncode{(!}p\ncode{)}$), the if statement executes both branches, sending as belief-state input to each the set of environments in which the condition has the appropriate value.
It sends the true environment as input to the branch it actually takes and the null environment to the other branch.
The resulting belief state is then the union of environments resulting from either branch, and the resulting true environment is from the branch that the initial true environment actually took.
If an if statement's condition is nondeterministic, then neither of its branches can make observations, or else the result is an error.
This is because it is unclear what interaction with the true environment means within a branch that environment did not necessarily take.
\paragraph{Infer Statements.} The semantics of infer statements are similar to the semantics of if statements in an ordinary language, where infer operates solely on the belief state.
If the belief state satisfies the condition, it evaluates the first branch and otherwise evaluates the second branch.
\paragraph{While Loops.}
The semantics of while loops is defined recursively using if statements.
This is similar to a standard equivalence notion for while-loop programs (see \cite{winskel} Section 2.5).
We additionally include an assertion that requires the loop invariant to be true.
\paragraph{Composition and Skip.} The semantics of statement composition and skips are standard, except that they have been appropriately extended to include the observation list.
Sequencing concatenates observation lists using the $\concat$ operator.
\paragraph{Errors.} We have elided for clarity from Figure~\ref{fig:statementsemantics} the full semantics of how errors propagate.
We assume that errors propagate maximally throughout the program, so if at any point the semantics produce $\bot$, the whole program produces $\bot$.
\subsection{Semantic Properties}
In this section, we establish several properties of the semantics in Figure~\ref{fig:statementsemantics} that we posit any belief programming system should satisfy.
These properties constrain the semantics so that the belief state updates respect the true environment updates.
The first property is that beliefs should be sound.
This means that for each environment in the belief state, and each new environment and new belief state that are reachable according to the semantics, the new environment is in the new belief state.
We formalize this as follows:
\begin{theorem}[Belief Soundness]{\ \\}
\vspace{-.35cm}
\begin{center}
If $\sigma \in \beta$ and $\tuple{\beta, \sigma, s} \Downarrow \tuple{\beta', \sigma' \; | \; o}$, then $\sigma' \in \beta'$.
\end{center}
\vspace{-.3cm}
\label{thm:bsound}
\end{theorem}
\begin{proof}
By structural induction on derivations of $\Downarrow$.
Details are in Appendix~\ref{sec:bsoundproof}.
\end{proof}
This means that it is impossible for the true environment to lie outside of the belief state.
In addition, beliefs should be precise.
This means that for every environment in a new belief state reachable from an initial belief state and the semantics, there is an environment in the initial belief state from which the new environment is reachable.
We formalize this as follows:
\begin{theorem}[Belief Precision]{\ \\}
\vspace{-.35cm}
\begin{center}
If $\tuple{\beta, \mu, s} \Downarrow \tuple{\beta', \mu' \; | \; o}$, then for every
$\sigma_\beta' \in \beta'$, there is some $\sigma_\beta \in \beta$ such that $\tuple{\beta, \sigma_\beta, s} \Downarrow \tuple{\beta', \sigma_\beta' \; | \; o}$
\end{center}
\vspace{-.3cm}
\label{thm:bprecise}
\end{theorem}
\begin{proof}
By structural induction on derivations of $\Downarrow$.
Details are in Appendix~\ref{sec:bpreciseproof}.
\end{proof}
This means that every environment in the belief state could be the true environment.
Finally, beliefs should be deterministic given observations.
This means that the belief state depends on the true environment only through observations.
We formalize this as follows:
\begin{theorem}[Belief Determinism]{\ \\}
\vspace{-.35cm}
\begin{center}
If
$\tuple{\beta, \mu_1, s} \Downarrow \tuple{\beta_1', \mu_1' \; | \; o_1}$ and
$\tuple{\beta, \mu_2, s} \Downarrow \tuple{\beta_2', \mu_2' \; | \; o_2}$ and
$o_1 = o_2$, then $\beta_1' = \beta_2'$.
\label{thm:bdeterm}
\end{center}
\vspace{-.3cm}
\end{theorem}
\begin{proof}
By structural induction on derivations of $\Downarrow$ using a strengthened induction hypothesis.
Details are in Appendix~\ref{sec:bdetermproof}.
\end{proof}
\subsection{Execution Model}
Here, we discuss how belief state updates execute in absence of the true environment.
This models how the belief program executes.
According to Theorem~\ref{thm:bdeterm}, the semantics in Figure~\ref{fig:statementsemantics} depends on the true environment $\mu$ only through the observation list $o$.
By projecting out the operations on belief states, we can use the semantics in Figure~\ref{fig:statementsemantics} to compute the new belief state using only the initial belief state and the sequence of observations.
In other words, if we define the belief execution $\Downarrow_\beta$ as follows
\begin{mathpar}
\inferrule
{
\tuple{\beta,\mu, s} \Downarrow \tuple{\beta', \mu' \; \mid \; o}
}
{
\tuple{\beta, o, s} \Downarrow_\beta \beta'
}
\end{mathpar}
then $\Downarrow_\beta$ is a partial function of $\beta$, $o$, and $s$.
The function $\Downarrow_\beta$ gives the semantics of the belief program's execution on a concrete sequence of observations $o$.
|
\section{Introduction} \label{secc:intro}
Mapping the cosmic history of star-formation is of fundamental importance in our understanding of galaxy formation and evolution not only because it contains the galaxy mass assembly history but also because it represents the footprint of the metal enrichment of the Universe.
A complete, unbiased determination of the star formation rate density (SFRD) requires a multi-wavelength approach
to directly probe the stellar emission of new-born stars
as well as the starlight that has been absorbed (then re-emitted) in dust-enshrouded regions. The former is efficiently measured using rest-frame ultraviolet (UV) observations and the latter using far-infrared (FIR) and (sub-)millimeter surveys that trace the dust re-processed emission from young stars (see \citealt{Madau2014a} for a review).
The current UV census of star formation reaches out to $z\sim11$, close to the formation epoch of the first galaxies (\citealt{Oesch2018a}). Nevertheless, despite some individual detections of dust in galaxies up to $z\sim8$ (\citealt{Watson2015a,Laporte2017a}), studies of the global dust-obscured star formation rate density from FIR/sub-mm surveys are very limited at $z>3$.
The highest redshift estimations of the SFRD from FIR-to-mm surveys, at $z=4.5-5.5$, are uncertain even at the $\approx1.0-1.5\,$dex level (\citealt{Rowan-Robinson2016a,Williams2019a,Loiacono2020a}) and differ by up to two order of magnitudes (\citealt{Michalowski2017a} cf. \citealt{Rowan-Robinson2016a}). This is mostly due to the low number statistics since most of the samples used to derive these measurements range from only one (\citealt{Williams2019a,Loiacono2020a}) to a handful of objects (\citealt{Dudzeviciute2020a,Gruppioni2020a}).
This implies that the total amount of
dust-enshrouded star formation in the earliest epochs of the Universe has remained unknown and, therefore, our current picture of the history of cosmic star formation remains incomplete.
This study aims at estimating the
history of dust-obscured star formation back to $z\sim7$ by combining a backward galaxy evolution model of the dusty star-forming galaxy (DSFG)
population (\citealt{Casey2018b,Casey2018a}) with the state-of-the-art Atacama Large Millimeter/submillimeter Array (ALMA) surveys.
We infer constraints on the prevalence and
characteristics of these galaxies through measurements of galaxy number counts at different
wavelengths. Despite its relative conceptual simplicity, number counts have proven to be a very powerful tool to test and constrain galaxy formation models (e.g. \citealt{Baugh2005a,Hayward2013a}), and consequently, represent an essential measurement in any long-wavelength survey.
The galaxy number counts used in this work include the first arcsecond-resolution interferometric number counts at 2\,mm achieved with ALMA and the most recent estimations at 1.2\,mm and 3\,mm. The former is from our new 2\,mm Mapping Obscuration to Reionization (MORA) survey described in \S\ref{sec:MORA}, the first ALMA large map at this wavelength with a total area of 184\,arcmin$^2$ (an order of magnitude larger than previous interferometric blind surveys). At 3\,mm, the measurements come from the $\sim200\,$arcmin$^2$ ALMA archival program reported in \cite{Zavala2018c}, which we revise in this work after finding three false detections in their sample (see \S\ref{secc:3mm_num_counts}).
These number counts are also complemented with those from the 4.2\,arcmin$^2$ deep ALMA large program ASPECS survey (at 1.2\,mm and 3\,mm; \citealt{Gonzalez-Lopez2019a,Gonzalez-Lopez2020a}), in which 40 continuum sources were detected. All these surveys represent the state-of-the-art observations at long wavelengths, comprising not only the deepest measurements
at those wavebands (beyond those assisted by gravitational amplification), but also the only interferometric surveys at 2 and 3\,mm, which are crucial to constraining the infrared (IR) luminosity function (IRLF) at high-redshift (see discussion by \citealt{Bethermin2015a,Casey2018b,Casey2018a,Zavala2018c}, and \citealt{Wilner1997a} for a pioneering $\sim3\,$mm survey).
The constraints provided by these surveys and by all the far-infrared and sub-millimeter data aggregated over the last two decades, allow us to infer the IR luminosity function and dust-obscured star formation in the last 13 billion years of the Universe, pushing the current redshift limits from $z\sim4-5$ to $z\sim7$. Our constraints are discussed in \S\ref{secc4} and summarized in \S\ref{secc:conclusions}.
Trough this work we assume $H_0=67.7\,\rm km\,s^{-1}\,Mpc^{-1}$ and $\Omega_{\lambda}=0.69$ (\citealt{Planck2016a}), and a Chabrier initial mass function (\citealt{Chabrier2003a}) for SFR estimations.
\section{Mapping Obscuration to Reionization ALMA (MORA) Survey} \label{sec:MORA}
The MORA survey (ALMA project code: 2018.1.00231.S, PI: C. Casey) was originally designed to cover a contiguous area of $\rm230\,$arcmin$^2$ in two different tunings in the 2\,mm band (centered at 147.3\,GHz and 139\,GHz, respectively). The project was only partially observed, covering $\rm184\,$arcmin$^2$ in two separate mosaics of $\rm156\,$arcmin$^2$ and $\rm28\,$arcmin$^2$, respectively.
The largest mosaic, which covers the positions 10 to 20 from the original design and so-called P10-P20, is centered at $\alpha\rm\approx10\,h\,00\,m\,17\,s$, $\delta\approx+02^\circ\,22'\,30''$ and covers an area of $6.6'\times23.6'$.
The smaller one (P03) is centered at $\alpha\rm\approx10\,h\,00\,m\,44\,s$, $\delta\approx+02^\circ\,22'\,30''$ and has $1.2'\times23.6'$ dimensions (see Figure \ref{fig:map}). The deepest portion of the mosaics has an RMS of $\sigma_{\rm 2mm}=60\,\mu\rm Jy\,beam$, with $101\,$arcmin$^2$ covered at or below the proposed map depth of $90\,\mu\rm Jy\,beam$ (see Figures \ref{fig:map} and \ref{fig:area}), with a typical beamsize of $\theta_{\rm FWHM}\approx1.8''\times1.4''$. This project represents the largest ALMA blank-field survey to-date and the only at 2\,mm to search for dusty star-forming galaxies.
\begin{figure*}[t]\centering
\includegraphics[width=0.5\textwidth]{mosiac_map_modified2.eps}
\caption{The MORA 2\,mm SNR maps and the 13 sources detected above $5\sigma$ (orange squares). The noise variations across the mosaics are illustrated by the contours, which range from $60\rm\,\mu Jy$ to $240\rm\,\mu Jy$ in steps of $30\rm\,\mu Jy$ . The typical size of the primary beam response at the frequency of our observations ($\theta_{\rm FWHM}\approx43''$) is represented by the orange circle (note that the synthesized beamsize of $\theta_{\rm FWHM}\sim1.5\arcsec$ is significantly smaller). The $-6\sigma$ peak is also identified with the pale green diamond.
\label{fig:map}}
\end{figure*}
\begin{figure}[t]\hspace{-0.6cm}
\includegraphics[width=0.53\textwidth]{area_per_noise.eps}
\caption{The cumulative
distribution of 1$\sigma$ survey depth and over what solid angle
that RMS depth is achieved. The P03 mosaic alone (covering $28\,\rm arcmin^2$) is represented
in pale green, while the P10--P20 mosaic (covering $156\,\rm arcmin^2$) is shown in blue. The proposed map depth ($90\,\mu$Jy) is illustrated by the orange dashed line. With a total area of $184\,\rm arcmin^2$ (orange line), the MORA program represents the largest ALMA blank-field survey to-date and the only at 2 mm to search for dusty star-forming galaxies.
\label{fig:area}}
\end{figure}
The signal-to-noise ratio (SNR) maps are well modelled by Gaussian statistics within $-4.5\lesssim\rm SNR \lesssim 4.5$ (see Figure \ref{fig:pixels_histogram}). Above $4.5\sigma$ the excess of positive pixels comes from the detection of astronomical sources, as expected. Interestingly, there is also an excess of negative pixels with $\rm SNR<-5$. All these pixels (identified in orange in the figure) come from the same region and are associated with a single noise peak at $-6\sigma$ (plus the effect of beam smearing; see Figure \ref{fig:map}). Nevertheless, assuming Gaussian statistics, the probability of finding a $-6\sigma$ noise peak in our mapped area is $\lesssim0.5\%$. An alternative explanation might be the Sunayev-Zeldovich (SZ) effect (\citealt{Sunyaev1970a}) since its peak decrement is expected close to the observed frequency of our maps ($\sim145\,$GHz). A thorough discussion of the nature of this negative detection will be presented in a future work, after carrying out the required follow-up observations.
Casey et al. (in preparation) will present further details of the survey along with a description of the source catalog and other physical characteristics of the 2\,mm-selected galaxies. Briefly, we first create a noise map using two different approaches. In the first one, we measure the standard deviation of all the pixels with primary beam responses between 0.95 and 1, which give us the minimum noise in our mosaic (or maximum depth); then we multiply
the inverse of the primary beam response map by this minimum noise vale, whose result is adopted as the noise map. In the second approach, we use a 2D boxcar-like function across the primary beam corrected flux mosaic (after applying a sigma clipping procedure to remove the bright sources) to estimate the noise in each pixel, measured as the standard deviation of the pixels within the squared region of the boxcar-like function. Both procedures give us consistent results within $\sim1$\%. Once we have a noise map, we simply divide the primary beam corrected flux map by this noise mosaic in order to obtain a SNR map (which is shown in Figure \ref{fig:map}). This SNR map is then used to find source candidates by searching for pixels above a SNR threshold.
Here, we use the sources detected at $>5\sigma$ (where the false detection rate is expected to be $\lesssim8\%$; see \S\ref{secc:num_counts}) to calculate the first ALMA number counts at 2\,mm, as described below.
\begin{figure}[t]\hspace{-0.2cm}
\includegraphics[width=0.51\textwidth]{noise_pixel_histogram.eps}
\caption{Histogram of pixel values in the SNR maps. The distribution of pixel values within $\rm -4.5\lesssim SNR\lesssim4.5$ is very well fitted by a Gaussian function (pale green solid line), which confirms the Gaussian properties of the map's noise. Above $\sim4.5\sigma$ the distribution starts to diverge due to the presence of positive sources. Our $5\sigma$ threshold to identify robust sources is indicated by the dashed orange line and arrow. All the pixels highlighted in orange, which have values below $-5\sigma$, lie within a region spanning approximately a beamsize. This negative signal could be associated with a single noise peak at $-6\sigma$ or with a detection of the SZ effect, as discussed in \S\ref{sec:MORA}.
\label{fig:pixels_histogram}}
\end{figure}
\subsection{The 2\,mm number counts}\label{secc:num_counts}
The cumulative number counts -- i.e. the number of galaxies above a certain flux density threshold, $S$, per unit area -- can be directly estimated by counting the number of detected galaxies as a function of flux density and making the required corrections for contamination, completeness and flux boosting.
Previous blind ALMA surveys have estimated a minimum, close-to-zero, contamination fraction due to false detections (noise peaks) for SNR thresholds in the range of $\approx4-5$ (e.g. \citealt{Fujimoto2016a,Oteo2016a}). Nevertheless, despite adopting a conservative threshold of $5\sigma$ in our analysis (see Figure \ref{fig:pixels_histogram}), a thorough characterization of the contamination rate is required given the large area covered by the MORA survey and the large number of independent beams in the maps.
The false detection rate of the MORA survey is characterized as follows: first, we create noise maps with the same dimensions as the original maps under the assumption that the noise in the ALMA observations is well represented by Gaussian statistics (as has been shown before for other ALMA data, e.g. \citealt{Franco2018a}, and demonstrated in Figure \ref{fig:pixels_histogram}). Second, the maps are convolved with a beam representative of the average synthesized beam in our observations. Then, to characterize the contamination rate as a function of SNR, the map is scaled such that the final standard deviation of the whole distribution of pixels is equal to one. This is equivalent to a SNR map with no astronomical sources. Finally, we simply quantify the number of peaks above the adopted detection threshold using the same algorithm used to find the real sources. This process is repeated 100 times in order to measure the expected number of false sources and its corresponding uncertainty. At our adopted threshold of $5\sigma$, the expected number of false detections in the MORA catalog is $0.8\pm0.2$. This is in good agreement with the findings in Casey et al. in preparation, who found near-IR counterparts for all the MORA detections with the exception of the lowest significance source in the catalog.
To take into account this false detection rate in the number counts calculation, a statistical approach is adopted in which the contamination fraction is distributed among the 13 sources according to their SNRs with a proportionality given by a normal probability distribution (i.e. estimating the probability that a normal random variable is greater than the sources' SNRs). As expected, this procedure gives a higher probability of false detection to the sources with the lowest SNRs and an (almost) negligible probability to those detected at $\gtrsim5.5\sigma$.
The survey completeness is calculated in a similar fashion in which artificial sources are inserted in the real maps (after removing the bright detections) followed by computing the ratio between the recovered sources and inserted sources. The artificial sources are inserted, one at a time, at random positions around the whole mosaic and then they are recovered with the same source extraction procedure used to build the real source catalog.
The survey completeness can be estimated as a function of input flux density, recovered flux density, detected SNR, local noise RMS, or a combination of these parameters. A source is considered recovered if it is detected within
a synthesized beam from the input position ($\sim1.5''$). In each flux density bin of $25\,\rm\mu Jy$, we repeat this process 10,000 times in order to sample the noise variation across the maps.
The same set of simulations are used to
examine the flux-boosting effects (meaning sources' flux densities systematically biased upwards by noise and the presence of fainter undetected sources). This is done by estimating the average ratio between the sources' intrinsic flux density and the recovered flux density in the simulations. Again, this value depends on the flux density and on the local noise of each source (or similarly, SNR), and therefore, several flux boosting factors can be derived as a function of these parameters. In average,
we find flux boosting factors of around $5\%$ for those source detected just above our SNR threshold. These factors, which decrease with increasing SNR, are however
much lower than the uncertainties on the measured flux densities ($\sim20\%$ for a $5\sigma$ detection).
The number counts can then be directly estimated by counting the number of detected galaxies as a function of flux density and making the appropriate corrections for contamination, completeness and flux boosting. We estimate the contribution of a source
with a deboosted flux density, $S_i$, and measured SNR, $\sigma_i$, to the cumulative number
counts to be:
\begin{equation}
\eta_i(S_i,\sigma_i)=\frac{1-f_{\rm cont}(\sigma_i)}{\zeta(S_i)\,A_{\rm eff}},
\end{equation}
where $f_{\rm cont}(\sigma_i)$ is the estimated fraction of contamination at the measured SNR ($\sigma_i$) of the source, $\zeta$ is the corresponding completeness for a deboosted flux density $S_i$, and $A_{\rm eff}$ is the total area of the MORA maps used for source extraction (i.e. $184\,$sq.\,arcmin$^2$; see Figure \ref{fig:area}).
Finally, the cumulative number counts, $N(>S)$, are estimated by the sum over all sources with a flux density higher than $S$, i.e. $N(>S)=\sum\limits_{i}\eta_i(S_i,\sigma_i)$.
\begin{figure}[t]\hspace{-1.1cm}
\includegraphics[width=0.55\textwidth]{2mm_num_counts.eps}
\caption{The cumulative number counts derived from our main sample (SNR$>5\sigma$) are represented by the blue solid circles. Previous estimation of the number counts using the GISMO camera on the IRAM telescope are represented by the pale green and orange squares (\citealt{Magnelli2019a}) and the yellow stars (\citealt{Staguhn2014a}). Additionally, the predictions from the {\sc shark} galaxy evolution model (\citealt{Lagos2020}) are illustrated by the gray line.
\label{fig:moras_num_counts}}
\end{figure}
\begin{deluxetable*}{cccccccc}[t]
\tablecaption{MORA survey 2\,mm number counts. \label{table:moras_num_counts}}
\tablecolumns{4}
\tablenum{1}
\tablewidth{0pt}
\tablehead{
\multicolumn{4}{c}{Raw number counts} & \multicolumn{4}{c}{Star-forming galaxies number counts} \\
\colhead{$S_{\rm 2mm}$} &
\colhead{N($\rm >S_\nu$)} &
\colhead{$\delta N^-$} &
\colhead{$\delta N^+$ } &
\colhead{$S_{\rm 2mm}$} &
\colhead{N($\rm >S_\nu$)} &
\colhead{$\delta N^-$} &
\colhead{$\delta N^+$ } \\
\colhead{($\rm mJy$)} &
\colhead{(deg$^{-2}$)} &
\colhead{ (deg$^{-2}$)} &
\colhead{ (deg$^{-2}$)} &
\colhead{($\rm mJy$)} &
\colhead{(deg$^{-2}$)} &
\colhead{ (deg$^{-2}$)} &
\colhead{ (deg$^{-2}$)} \\
}
\startdata
0.33 & 440 & 120 & 60 & 0.33 & 380 & 110 & 150 \\
0.36 & 370 & 110 & 140 & 0.36 & 320 & 100 & 130 \\
0.39 & 310 & 90 & 130 & 0.39 & 270 & 80 & 110 \\
0.41 & 260 & 80 & 110 & 0.48 & 220 & 70 & 100 \\
0.48 & 220 & 70 & 100 & 0.52 & 185 & 64 & 91 \\
0.52 & 185 & 64 & 91 & 0.54 & 157 & 59 & 85 \\
0.54 & 157 & 59 & 85 & 0.57 & 131 & 53 & 79 \\
0.57 & 131 & 53 & 79 & 0.61 & 106 & 47 & 72 \\
0.61 & 106 & 47 & 72 & 0.68 & 83 & 41 & 65 \\
0.68 & 83 & 41 & 65 & 0.78 & 60 & 34 & 59 \\
0.78 & 60 & 34 & 59 & 0.88 & 41 & 27 & 53 \\
0.88 & 41 & 27 & 53 & 1.03 & 21 & 17 & 47 \\
1.03 & 21 & 17 & 47 & & & & \\
\enddata
\tablecomments{The star-forming galaxy number counts are estimated after removing source MORA-10, whose flux density falls below our detection threshold after removing the contribution from synchrotron emission.}
\end{deluxetable*}
In order to take into account the uncertainties associated with the correction factors (completeness and false detection) and flux densities in the estimation of the number counts, we perform a Monte Carlo simulation. In each realization, the adopted measured flux density for each source, $S_i$, is extracted from a Gaussian distribution with a standard deviation
equal to the measured error and centered at the observed value. Then, this new flux density is used to estimate a new SNR, $\sigma_i$, by dividing by the local noise. These SNRs are then used to update the contamination fraction and completeness accordingly, whose new values are drawn from Gaussian distributions.
This procedure is repeated 100 times, with the mean values representing the final number counts and the 16th and 85th percentiles their associated confidence interval. Finally, given the relatively small number of sources in our catalog, Poisson uncertainties are added in
quadrature according to \citet{Gehrels1986a}.
The MORA 2\,mm number counts derived in this work are reported in Table \ref{table:moras_num_counts} and are show in Figure \ref{fig:moras_num_counts}.
Given that the 2\,mm flux density of one of our sources, MORA-10, was found to be contaminated by synchrotron emission (at a $\sim40\%$ level; Casey et al. in preparation), we repeat all the process described above after removing this source, aiming at providing the 2\,mm number counts of star-forming galaxies only (note that the source would have been fallen below our detection threshold without the synchrotron emission). This is particularly important since the model used in \S\ref{secc:model} does no take into account non-thermal emission, nevertheless, we highlight that the difference between the two estimations is not significant (see Table \ref{table:moras_num_counts}).
At this wavelength, the only determinations of the number counts reported in the literature beyond our MORA survey measurements come from the GISMO/IRAM surveys reported in \cite{Staguhn2014a} and \cite{Magnelli2019a}, covering $31\,$arcmin$^2$ and $250\,$arcmin$^2$, respectively. The latter are in very good agreement with our estimations while the former lie above by a factor of $\approx2.0-2.5$. This over-estimation is thought to be caused by the uncertainties in the flux deboosting factors in the confusion-limited map of \cite{Staguhn2014a}, as discussed in \cite{Magnelli2019a}, although cosmic variance might be also important given the relatively small mapped area. The predictions from the {\sc shark} semi-analytic model of galaxy formation (\citealt{Lagos2020}) are also plotted in the figure, which are indeed in very good agreement with our estimations.
\section{Revising the 3\,mm number counts}\label{secc:3mm_num_counts}
\cite{Zavala2018c} compiled archival ALMA band 3 observations toward three extragalactic legacy fields: COSMOS, CDF-S, and the UDS, resulting in a total of 135 individual maps adding up a total area of $\approx200\,\rm arcmin^2$.
After masking out the original targets of these observations and any galaxies that were potentially physically associated with them, they derived the first ALMA galaxy number counts at $3\,$mm using a total of 13 sources detected above $5\sigma$.
Our recent ALMA 2mm + 3mm follow-up observations on this sample (ALMA projects: 2018.1.00478.S and 2019.1.00838.S; PI: J. Zavala) show that three of these sources (ALMA-3mm.14, ALMA-3mm.15, and ALMA-3mm.16) are not recovered, meaning they are likely spurious detections. Two of these sources were indeed noted to have a spectral index which might be inconsistent with thermal emission in \cite{Zavala2018c}, nevertheless, without further data at the time, the low SNRs ($5.0-5.2$) prevented a firm conclusion and therefore they were included in the previous number counts estimation.
Thus, we have revised the false detection rate to be higher than the value reported in the original work of \cite{Zavala2018c}, for which only one source was expected to be false. This highlights the complexity of the interferometric data, particularly when using observations with different beamsizes, integration times, and array configurations, as is typical for datasets derived from archival projects.
\begin{deluxetable}{cccc}[h]
\tablecaption{Revised ALMA archival 3\,mm number counts. \label{table:revised_num_counts}}
\tablecolumns{4}
\tablenum{2}
\tablewidth{0pt}
\tablehead{
\colhead{$S_{\rm 3mm}$} &
\colhead{N($\rm >S_\nu$)} &
\colhead{$\delta N^-$} &
\colhead{$\delta N^+$ } \\
\colhead{($\rm \mu Jy$)} &
\colhead{(deg$^{-2}$)} &
\colhead{ (deg$^{-2}$)} &
\colhead{ (deg$^{-2}$)}
}
\startdata 33 & 4440 & 1200 & 1990 \\
52 & 2140 & 640 & 1000 \\
63 & 1420 & 470 & 710 \\
79 & 990 & 360 & 540 \\
98 & 750 & 290 & 450 \\
109 & 550 & 230 & 370 \\
117 & 410 & 190 & 320 \\
127 & 290 & 150 & 280 \\
137 & 170 & 110 & 230 \\
167 & 74 & 55 & 170 \\
\enddata
\end{deluxetable}
Here, we revised the 3\,mm number counts following exactly the same procedure as described in \cite{Zavala2018c} but removing the three spurious sources. Additionally, we updated the flux density of the source ALMA-3mm.03 (a.k.a. ASPECS-3mm.1) since the reported value in \cite{Zavala2018c} seems to be contaminated by a bright CO emission line (\citealt{Gonzalez-Lopez2019a}); we thus adopt the flux density reported in the ASPECS catalog.
The revised 3\,mm number counts are reported in Table \ref{table:revised_num_counts} and shown in Figure \ref{fig:num_counts}. The new values are significantly different at the bright end ($S_{\rm 3mm}>0.2\,$mJy) since two of the false detections were the brightest galaxies in the sample\footnote{Given the inhomogeneous observations in the archival data, a low signal-to-noise ratio does not necessary imply a low flux density since each map has a different noise r.m.s. depth. The two brightest false detections were indeed detected in areas with large r.m.s values.}. At fainter flux densities, the updated number counts are a factor of $\sim1.5\times$ lower, and thus, in better agreement with the ASPECS results (\citealt{Gonzalez-Lopez2019a}), although still a factor of $\sim2\times$ higher. Nevertheless, the two number counts are consistent with each other within the uncertainties. The difference is thus not statistically significant.
\section{Constraining the IR luminosity function and dust-obscured star formation rate density}\label{secc4}
In this section we use a backward evolution model in combination with the state-of-the-art FIR/mm surveys to constrain the IR luminosity function.
The model draws spectral energy distributions (SEDs) from the known breadth and characteristics of dusty galaxies, then works backward to discriminate between different luminosity function scenarios using the galaxy number counts as constraints.
Finally, the dust-obscured star formation rate density is estimated by integrating the best-fit IR luminosity function (see \citealt{Zavala2018c} for a similar analysis).
\subsection{Model description}\label{secc:model}
The adopted backwards evolution model is described in detail in \cite{Casey2018b,Casey2018a}, where the specifics of the model and the assumed values for each parameter can be found. A summary of the salient characteristics and assumptions follows.
The model combines a parameterized evolving galaxy IR luminosity
function with the thermal SED properties of galaxies' dust emission to make predictions for galaxy (sub-)millimeter surveys.
Galaxies' SEDs are described by a modified black-body function of the form $S_\nu\propto(1-e^{-(\nu/\nu_0)^{\beta_{\rm em}}})B_\nu(T_{\rm dust})$ with an extra mid-infrared power law to account for hotter dust emission at $T\gg T_{\rm dust}$ (e.g. \citealt{Casey2012a}). The SED library follows the well-known luminosity-temperature relationship in terms of $L_{\rm IR}-\lambda_{\rm peak}$ and takes into account the observed dispersion in this relation. We highlight that using the peak wavelength rather than dust temperature minimizes the impact of the assumed effective dust opacity in the model ($\tau=1$ at $\lambda_{\rm rest}=100\,\rm\mu m$), which could change the derived dust temperature but has a minor impact on the measured $\lambda_{\rm peak}$. Note that although this relation is not assumed to evolve with redshift (which is supported by the current data at least up to $z\sim4.5$ \citealt{Casey2018a,Dudzeviciute2020a}), if one selects galaxies on the main sequence, a redshift evolution of $\lambda_{\rm peak}$ (proportionally to dust temperature) naturally arises. This evolution, which is driven by the underlying evolution of the main sequence of star-forming galaxies, is in line with recent results from the literature (e.g. \citealt{Bethermin2015a,Schreiber2015a,Magdis2017a}).
As is common practice in the literature (e.g. \citealt{Sanders2003a,Magnelli2011a,Magnelli2013a,Lim2020a}), the assumed IR luminosity function in the model is described by a double power law of the form:
\begin{equation}
\Phi(L,z) =
\begin{cases}
\Phi_\star(z)\left(\frac{L}{L_\star(z)}\right)^{\alpha_{\rm LF}}, & \mbox{if } L<L_\star, \\
\Phi_\star(z)\left(\frac{L}{L_\star(z)} \right)^{\beta_{\rm LF}}, & \mbox{if } L\ge L_\star,
\end{cases}
\end{equation}
where $\alpha_{\rm LF}$ and $\beta_{\rm LF}$, represent the slopes at faint and bright luminosities, respectively, while $L_\star$ and $\Phi_\star$ are the characteristic galaxy luminosity and characteristic number density, two fundamental parameters that are allowed to evolve with redshift.
Following previous works from the literature, the characteristic number density is assumed to evolve as:
\begin{equation}
\Phi_\star \propto
\begin{cases}
(1+z)^{\Psi_1}, & \mbox{if } z<z_{\rm turn}, \\
(1+z)^{\Psi_2}, & \mbox{if } z\ge z_{\rm turn}.
\end{cases}
\end{equation}
Significant observational efforts have provided good constraints on the evolution of $\Phi_\star$ at low redshifts, revealing a flat trend with $\Psi_1$ close to zero and a turnover redshift of $z_{\rm turn}\approx2$, in-line with the peak of the cosmic star formation rate density (CSFRD; \citealt{Magnelli2011a,Gruppioni2013a,Lim2020a}). Although it is clear that at higher redshifts the evolution is significantly steeper, the current estimations are highly uncertain and limited to $z\lesssim4$, with values ranging from $\Psi_2\sim-6$ (or lower; \citealt{Koprowski2017a}) to $-2$ (or even higher; \citealt{Rowan-Robinson2016a}), with a few measurements in between. Indeed, \citet{Zavala2018c} found an intermediate value of $\Psi_2\approx-4.2$ using the initial 3\,mm sample.
In this work, we aim to constrain two of the most important unknowns of the galaxy luminosity function: the faint-end slope, $\alpha_{\rm LF}$, and the redshift evolution of the characteristic number density at high redshift driven by $\Psi_2$ (since $\Phi_\star\propto(1+z)^{\Psi_2}$ for $z\gtrsim2$). We also explore different values of the dust emissivity index, $\beta_{\rm em}$, in order to minimize any possible bias and to test possible degeneracies between the different parameters.
For the rest of the model parameters, we adopt the same values as in \citet{Casey2018a} and \citet{Zavala2018c}. We refer the reader to \citeauthor{Casey2018a} (\citeyear{Casey2018a}, and particularly to its appendix A.1) for a detail discussion about these choices and their impact on the model. In short, the bright-end slope of the luminosity function is fixed to be $\beta_{\rm LF}=-3.0$. This value is relatively well-constrained (at least up to $z\sim4$) with a measured dispersion of $1\sigma\approx0.15$.
The evolution of the characteristic luminosity, $L_\star$, follows $L_\star\propto(1+z)^{\gamma_1}$ for $z<z_{\rm turn}$ and $L_\star\propto(1+z)^{\gamma_2}$ for $z\ge z_{\rm turn}$. While the evolution at lower redshifts ($z<z_{\rm turn}$) is well-characterized (we adopt $\gamma_1=2.8$), the evolution of $L_\star$ beyond $z_{\rm turn}$ is more uncertain. The model adopts $\gamma_2=1.0$, implying that $L_\star$ continues to evolve upward toward higher redshifts, in line with expectations for hierarchical structure formation and cosmic downsizing. Additionally, adopting $\gamma_2=1.0$ makes the $L_\star$ evolution consistent with the $L_\star$ of the quasar luminosity function (\citealt{Hopkins2007a}). Adopting a reversal evolution ($\gamma_2\le0$) under-predicts the number counts at wavelengths longer than $\sim850\rm\,\mu m$, regardless of the adopted number density (for any $\Psi_2<-1.5$), while a more rapid evolution with $\gamma_2\ge1.5$ would imply very bright values for $L_\star$ at $z\sim5$, which might even exceed the luminosity of the brightest DSGFs known to-date. Moreover, in order to match the contribution from LIRGs and ULIRGs to the cosmic SFRD at $z<2.5$ to those reported in the literature, a modest dependence of $z_{\rm turn}$ on $\Psi_2$ of the form $z_{\rm turn}=1.6-0.09\Psi_2$ has to be adopted (see \citealt{Casey2018a} for details).
This model has some caveats that the reader should keep in mind.
The exact values of $L_0$ and $\Phi_0$ are correlated with $\gamma_1$ and $\Psi_1$, respectively\footnote{We refer the reader to equations 8 and 9 from \citet{Casey2018a} for the exact parametrization of the evolution of $L_\star$ and $\Phi_\star$ and the role of $L_0$ and $\Phi_0$.}. As discussed in detailed by \citet{Casey2018a}. The adopted values ($L_0=10^{11.1}\,L_\odot$, $\Phi_o=10^{-3.5}\rm\,Mpc^{-3}\,dex^{-1}$, $\gamma_1=2.8$, and $\Psi_1=0$) were chosen to simultaneously reproduce: {\it(i)} the IRLF at $z\lesssim2$; {\it(ii)} the reported values of $L_\star$ and $\Phi_\star$ from the literature; and {\it(iii)} the relative contribution from LIRGs and ULIRGs to the total SFRD within $0<z<2$. This combination of values are, however, not necessarily unique. Another caveat that the reader should keep in mind is the assumption of a fixed emissivity spectral index at all redshifts, and although several values were explored in the analysis (from $\beta=1.6$ to 2.6, see Figure \ref{fig:triangle_plot}), each realization adopts a single value across all redshifts for all sources. This is not necessarily in disagreement with previous literature, given the current limited data. Although a spread on $\beta$ has been found across different galaxies, these variations have not yet been shown to correlate with redshift or other galaxy characteristics that may impact the model. Indeed, there are clear examples of $\beta\sim1.5$ to $\sim2.0$ in both local galaxies (e.g. \citealt{Remy-Ruyer2013a}) and high-redshift systems (e.g. \citealt{Jin2019a}).
The same is true for the adopted slopes of the luminosity function at both the faint and bright ends ($\alpha_{\rm LF}$ and $\beta_{\rm LF}$, respectively). Although a redshift evolution is possible given that we know similar parameters do evolve for the UV luminosity function (e.g. \citealt{Finkelstein2015a}), the lack of samples of galaxies with spectroscopic redshifts covering a wide range of IR luminosities prevents its confirmation (or rejection). Indeed, non-evolving slopes have been been adopted in several works in the literature (e.g. \citealt{Magnelli2011a,Magnelli2013a,Gruppioni2013a,Lim2020a}). Therefore, in the absence of more information, we choose to fix these quantities. The impact of these assumptions on our results are further discussed below.
\subsection{Fitting methodology and data constraints}
To find the best-fit model parameters,
first, the different evolutionary scenarios of the IRLF are combined with galaxies' SEDs to create mock observations that resemble the real surveys in terms of wavelength, noise depth, and angular resolution, with the CMB heating effects taken into account (following \citealt{daCunha2013a}). Second, the respective number counts are estimated in a similar fashion as in the real observational works.
These simulated number counts are then compared to the measured number counts in a joint analysis that combines multiple observations at different wavelengths simultaneously. Finally, the best-fit model parameters that better reproduce all the observed number counts are derived.
The cumulative number counts are preferred over the differential number counts since the latter suffer from larger uncertainties due to the relative small number of sources in each flux density bin. Though we test that adopting the differential number counts instead do not change significantly our results.
As discussed in \cite{Casey2018b,Casey2018a},
observations at long-wavelengths ($\lambda\gtrsim1\,$mm) are ideal to distinguish between the different scenarios of the model since the majority of constraining data sets at shorter wavelengths only inform about the evolution of the IRLF at $z\lesssim3.0$, where the model parameters are already relatively well understood. Therefore, here we only use number counts at 1.2\,mm, 2\,mm and 3\,mm as constraints, which also reduces significantly the computational cost of the fitting analysis, which would be prohibitive otherwise. We adopt the MORA 2\,mm number counts described above (see \S\ref{secc:num_counts}), the updated 3\,mm number counts reported in this work (see \S\ref{secc:3mm_num_counts}), and the 1.2\,mm and 3\,mm number counts from the ASPECS survey (\citealt{Gonzalez-Lopez2019a,Gonzalez-Lopez2020a}). Altogether, these surveys add a total of $\sim400\,$sq.\,arcmin surveyed area and around 60 detected galaxies. Note that AGN contribution or non-thermal emission (e.g. synchrotron emission) is expected to be negligible at these wavebands and at the flux density ranges explored here\footnote{From all the sources used in this work, we found a non-negligible contamination from synchrotron emission only in one source, namely MORA-10. This source has been removed from our analysis (see \S\ref{secc:num_counts}).} (see discussion in \citealt{Casey2018b,Zavala2018c}).
Once the fiducial number counts have been defined, to derive the best-fit model parameters that better reproduce these number counts we use two different methods: a maximum likelihood approach and a multi-dimensional minimization algorithm.
The maximum likelihood approach is done in a similar fashion as reported by \cite{Zavala2018c}. We create a three-dimensional grid, representing the parameter space formed by
$\alpha_{\rm LF}$, $\Psi_2$, and $\beta_{\rm em}$, with values ranging from $\alpha_{\rm LF}: [-0.6,-0.2]$ in steps of 0.05, $\Psi_2: [-8.8,-2.2]$ in steps of 0.4, and $\beta_{\rm em}: [1.5,2.6]$ in steps of 0.05.
For each combination of the parameters subset (corresponding to different points in the grid), the likelihood function of the measured cumulative number counts relative to the simulated number counts is estimated. Finally,
the best-fit model is assumed to be the one with the highest probability, and the confidence intervals at 68\%, 95\%, and 99.7\% are
obtained by integrating the normalized likelihood distribution.
\begin{figure*}[ht]\hspace{-0.3cm}
\includegraphics[width=1.03\textwidth]{best_fit_number_counts_breakdown.eps}\vspace{0.7cm}
\caption{Cumulative number counts -- galaxies above a given flux density per unit area -- at 1.2\,mm (left panel), 2\,mm (middle panel), and 3\,mm (right panel). The data points used for the fitting analysis are illustrated as blue points
while other measurements from the literature are plotted as green squares. The best-fit number counts from the model, represented by the gray lines, nicely reproduce the number counts at the three different wavelengths simultaneously, including those not used in the fitting procedure as well as number counts at shorter wavelengths (see Appendix \ref{appendix}), spanning more than a decade in wavelength and order of magnitudes in flux density. The best-fit number counts from the model are also broken down into two ranges of luminosity and three redshift bins. Galaxies with $L_{\rm IR}>10^{12}\,L_\odot$ are shown in solid lines while those with $L_{\rm IR}<10^{12}\,L_\odot$ are shown in dotted lines. Similarly, the three different redshift ranges are illustrated by different colors: blue for $z<3$, gold for $3<z<6$, and orange for $z>$6.
\label{fig:num_counts}}
\end{figure*}
The Nelder–Mead optimization method (a.k.a the Amoeba method; \citealt{Nelder1965a,numerical_recipes}) relies on a downhill simplex algorithm to perform a multidimensional minimization of a given parameter, in this case, the square differences ($\chi^2$) between the real and the modeled cumulative number counts. In each realization, we randomly vary the initial starting point of the search, with each search limited to 100 evaluations. During this analysis, we introduce a fourth parameter, $z_{\rm cutoff}$, following \cite{Zavala2018c}, which represents the redshift above which no more dusty galaxies exist (with a range of $z_{\rm cutoff}=5.5-9.0$).
Nevertheless, we find that $z_{\rm cutoff}$ is not well constrained by the data and has a minor impact on the analysis.
As shown in Figure \ref{fig:num_counts}, the best-fit number counts recovered from the model are in good agreement with the measured values, reproducing simultaneously not only the data used in the fitting but also the rest of measurements reported in the literature, as discussed below.
At 1.2\,mm, beyond the ASPECS number counts, the model predictions nicely reproduce the number counts from the 1.6\,deg$^2$ AzTEC surveys reported in \cite{Scott2012a} and those from the ALMA follow-up survey of SCUBA-2 galaxies detected over $\sim1\rm\,deg^2$ reported by \cite{Stach2018a}\footnote{The results from AzTEC observations and other ALMA surveys at 1.1\,mm were scaled by a factor of 0.8 while those from the ALMA/SCUBA-2 survey at $850\,\rm\mu m$ were scaled by a factor of 0.4.}, which probe the brightest flux densities. At fainter flux densities, the model also reproduces the more recent ALMA results from \cite{Dunlop2017a} and \cite{Umehata2017a}. The number counts from the GOODS-ALMA survey (\citealt{Franco2018a}; covering the flux density range of $S_{\rm1.2mm}\approx0.5-2\,$mJy) are, however, lower than our estimations, and those reported by
\citet{Fujimoto2016a} (which probe the faintest flux densities in the figure) lie above our estimations.
Note, however, that the measurements reported by \cite{Fujimoto2016a} come from gravitationally lensed fields, and therefore, these discrepancies might be caused by the uncertainties in the magnification factors and small survey area (see discussion in \citealt{Gonzalez-Lopez2020a}). The discrepancy with the \citet{Franco2018a} results is likely related to the relatively low completeness associated with their high angular resolution observations ($\theta_{\rm FWHM}=0.29''$ and $\theta_{\rm FWHM}=0.60''$ after filtering the data).
At 2\,mm, the only determinations of the number counts reported in the literature beyond our MORA survey measurements come from the GISMO/IRAM surveys reported in \cite{Staguhn2014a} and \cite{Magnelli2019a}, covering $31\,$arcmin$^2$ and $250\,$arcmin$^2$, respectively. The latter are in very good agreement with the model predictions while the former lie above by a factor of $\approx2.0-2.5$. This difference is thought to be caused by the uncertainties in the flux deboosting factors in the confusion-limited map of \cite{Staguhn2014a}, as discussed in \cite{Magnelli2019a}, plus the possible effects of cosmic variance (see \S\ref{secc:num_counts}).
At 3\,mm, the only measurement plotted in Figure \ref{fig:num_counts} which was not used in our analysis, is the brightest bin of the ASPECS survey, which represents only an upper limit. This value is, however, in good agreement with the model predictions.
The success of the model can also be illustrated by its predicting power at other wavelengths.
The model does reproduce the number counts at shorter wavelengths, from $70\,\mu\rm m$ to $850\,\mu\rm m$ spanning over two decades of observations over hundreds of square degrees (see Appendix \ref{appendix}).
\subsection{The IRLF and best-fit parameters}\label{secc:lum_func}
As it is clearly shown in Figure \ref{fig:triangle_plot}, both methods provide consistent results and good constraints on the model parameters governing the IRLF while successfully reproducing the observed number counts (see Figures \ref{fig:num_counts} and \ref{fig:num_counts_short_waves}).
\begin{figure*}[ht]
\includegraphics[width=\textwidth]{triangle_plot_v2.eps}
\caption{Constraints on the IRLF model parameters $\alpha_{\rm LF}$ (the faint-end slope of the luminosity function), $\Psi_2$ (which governs the evolution of $\Phi_\star$ at high redshift given $\Phi_\star\propto(1+z)^{\Psi_2}$), and $\beta_{\rm em}$ (the dust emissivity index). The confidence regions at the 68\%, 95\%, and 99.7\% for each parameter derived using the maximum likelihood approach are represented by the blue contours (from dark blue to light blue).
Additionally, the best-fit values extracted from 100 realizations of the Nelder–Mead multidimensional minimization algorithm (amoeba) are illustrated with the green solid circles. On the top of each column, we show the respective 1D marginalized probability distribution derived from the maximum likelihood approach (solid blue line), and the histogram of best-fit values found by the Nelder–Mead method (green dashed histograms).
The best-fit values derived from both methods and their corresponding $1\sigma$ uncertainties (68\% C.I.) are shown on the top of each row. Consistent results are derived from the two different approaches.
\label{fig:triangle_plot}}
\end{figure*}
The faint-end slope of the infrared luminosity function is found to be flat, with $\alpha_{\rm LF}=-0.42^{+0.02}_{-0.04}$. This is in line with recent studies based on the deepest single-dish observations and interferometric surveys (although those were limited to lower redshifts), with reported values of $\alpha_{\rm LF}=-0.4$ (\citealt{Koprowski2017a}) and $-0.5\pm0.7$ (\citealt{Lim2020a}), and with the observed flattening in the 1.1\,mm luminosity function (\citealt{Popping2020a}).
Although the faint-end slope is fixed for all redshifts in our model, we highlight that an evolution to steeper values with increasing redshift is inconsistent with our data and with the very low number of $z>4$ galaxies detected in the deepest ALMA surveys (\citealt{Dunlop2017a,Hatsukade2018a,Aravena2020a}). Indeed, most of the sources in the ASPECS survey lie at $z<3$ (\citealt{Aravena2020a}), in good agreement with our model predictions (see Figure \ref{fig:num_counts}).
This IR flat slope contrasts with the steep faint-end slope of the UV luminosity function at high redshift, which ranges from $-1.6$ at $z\sim4$ to $-2.0$ at $z\sim7$ (e.g. \citealt{McLure2013a,Finkelstein2015a}).
As a consequence, deep pencil-beam observations at (sub-)mm wavelengths would not significantly increase the number of detected sources, as is commonly the case in the rest-frame UV/optical observations (e.g. \citealt{Ferguson2000a}).
Regarding the evolution of the characteristic number density at high redshift, the best-fit $\Psi_2$ value of $-6.5^{+0.8}_{-1.8}$ implies a steep redshift evolution of the IR number density ($\Phi_\star\propto(1+z)^{-6.5^{+0.8}_{-1.8}}$), which disfavours the dust-rich hypothetical model discussed in \cite{Casey2018a}, for which a value of $\Phi_\star\propto(1+z)^{-2.5}$ was adopted. This indicates that dusty star-forming galaxies are indeed {\it rare} at early epochs. This steep evolution is also in line with the rapid drop-off of the quasar luminosity function at high redshift (\citealt{Hopkins2007a}) and similar to the number density evolution of UV-bright (M$_{\rm UV}=-$21) galaxies ($\propto(1+z)^{-5.9}$; \citealt{Finkelstein2015a}), which might suggest that these galaxies occupy similar dark matter halos.
There is, though, a caveat related to the effective dust optical opacity of our SEDs that the reader should keep in mind. Although, as mentioned before, the SEDs are parameterized in term of $\lambda_{\rm peak}$ rather than dust temperature, incorporating the CMB effects on the heating and detectability of the sources requires an estimation of the dust temperature (\citealt{daCunha2013a}). Assuming an optically thin opacity form would result in lower dust temperatures for the same $\lambda_{\rm peak}$ than those derived from the optically thick model, decreasing the contrast between the dust emission and the CMB. This implies that, in order to match the measured number counts, a higher number of galaxies would be required compared to the optically thick model, i.e. a higher $\Psi_2$ value (see discussion in \citealt{Zavala2018c}). Nevertheless,
this effect is only important if the number counts are dominated by very high-redshift ($z>6$) sources, which our results suggest is likely not the case.
Finally, Figure \ref{fig:triangle_plot} shows that the dust emissivity spectral index is well-constrained to be $\beta_{\rm em}=1.8\pm0.1$, in very good agreement with the values reported in the literature (e.g. \citealt{Dunne2011a,Galliano2018a}; although we note that measurements of this parameter
at high redshifts are scarce). This figure also reveals, unsurprisingly, a mild correlation between
$\beta_{\rm em}$ and $\Psi_2$ since higher values of $\beta_{\rm em}$ imply lower flux densities at long wavelengths (for a given
IR luminosity, redshift, and dust temperature).
We highlight that despite the large range of values explored for the three different parameters and despite the caveats described above, $\alpha_{\rm LF}$ and $\beta_{\rm em}$ are well-constrained, and while the uncertainties on $\Psi_2$ seem large, our constrains rule out values grater than $\Psi_2\sim-4$, which would imply a significantly larger number of DSFGs at high-redshifts.
\subsection{The history of dust-obscured star formation}
The star formation rate density as a function of redshift can now be calculated by integrating the best-fit infrared galaxy luminosity function\footnote{The IR luminosity function is integrated over the interval log$(L/L_\odot):[9,13.8]$. A change of these limits has a minor impact on our results since the majority of the contribution arise from galaxies with luminosities in the range of $L_{\rm IR}=10^{11}-10^{13}\,L_\odot$ (see Figure \ref{fig:sfrd}).} and propagating the associated uncertainties.
An important source of uncertainty is the field-to-field variation due to the large scale structure of the Universe, which is known as cosmic variance. To infer its impact in our estimations, we adopt the results from a model for the dust continuum number counts of galaxies (\citealt{Popping2020a}) that builds upon the {\sc UniverseMachine} model (\citealt{Behroozi2019a}), from which 1000 different light cones of 400\,sq.\,arcmin (approximately the combined area the MORA, ASPECS, and the 3\,mm archival surveys) are used to measure the variance in the number of detected sources. The cosmic variance is then defined following \citet{Moster2011a}:
\begin{equation}
\sigma^2 \equiv \frac{\langle N^2\rangle - \langle N\rangle^2}{\langle N\rangle^2} - \frac{1}{\langle N\rangle},
\end{equation}
where $\langle N^2\rangle$ and $\langle N\rangle$ are the variance and mean number of sources in the light cones.
At the most critical redshifts of our work ($z>3$), the cosmic variance is estimated to be $\approx35\%$. At $z\approx7.0$, the cosmic variance increases significantly and our data-sets suffer a loss of constraining power. Furthermore, an increasing fraction of the star formation activity is expected to be in highly clustered structures (i.e. galaxy proto-clusters; \citealt{Chiang2017a}) that might have been missed in our surveys. Therefore, the model predictions presented in this work are limited to $z\lesssim7.0$.
The inferred dust-obscured star-formation history and its associated uncertainty, including that from cosmic variance, is presented in Figure \ref{fig:sfrd}.
As can be seen, it is dominated by galaxies with $L_{\rm IR}\approx10^{11}-10^{13}$, a population that our surveys are particularly sensitive to (see Figure \ref{fig:lum_limits} in Appendix \ref{appendix2}).
Figure \ref{fig:sfrd} also includes other determinations of the CSFRD from the literature, which are split in two groups: FIR/sub-mm and UV/optical-based measurements. The first set includes the works by \cite{Magnelli2011a,Gruppioni2013a,Magnelli2013a,Casey2012b,Swinbank2014a,Bourne2017a,Koprowski2017a,Liu2018a, Williams2019a,Magnelli2019a,Wang2019b,Lim2020a,Dudzeviciute2020a}, and the second set the results reported in \cite{Wyder2005a,Schiminovich2005a,Robotham2011a,Cucciati2012a,Dahlen2007a,Reddy2009a,Bouwens2012a,Schenker2013a,Finkelstein2015a}. We remind the reader that our model is not a fit to those data points but are shown as a mean for comparison.
\begin{figure*}[ht]
\centering
\includegraphics[width=0.62\textwidth]{sfrd_breakdown_luminosities_v2.eps}
\caption{The inferred dust-obscured star formation history is illustrated by the orange shaded region in the bottom panel. For comparison, we plot independent measurements from the literature based on IR/sub-mm and UV surveys (orange circles and blue squares, respectively) and the average unobscured star-formation derived from rest-frame UV optical surveys (i.e. not corrected for dust attenuation; blue shaded region; \citealt{Finkelstein2015a}). The total inferred SFRD derived in this work is shown in gray. The uncertainties in our estimation include those from the best-fit parameters and cosmic variance. The middle panel represents the fraction of obscured star formation, $\rm SF_{obs}/(SF_{obs}+SF_{unobs})$, and its associated uncertainty (lighter shaded area). The contribution of dust-obscured galaxies, which dominates the cosmic star-formation history through the last $\sim12\,$Gyr, rapidly decreases beyond its maximum, reaching values that are comparable to the unobscured star formation traced by the rest-frame UV/optical surveys by $z\approx4-5$.
The top panel represents the contribution from galaxies with different luminosity ranges to the dust-obscured SFRD, being dominated by ULIRGs (ultra-luminous infrared galaxies; $10^{12}<L_{\rm IR}<10^{13}\,L_\odot$) and LIRGs ($10^{11}<L_{\rm IR}<10^{12}\,L_\odot$).
\label{fig:sfrd}}
\end{figure*}
The dust-obscured component traced by the FIR-to-mm surveys has dominated the cosmic history of star formation for the past $\sim12\,$billion years, with a peak era between $z=2-2.5$ ($\sim10-11\,$Gyr ago) and contributing around $\sim80\%$ of total SFRD (see middle panel in Figure \ref{fig:sfrd}). Beyond this peak redshift, the dust-obscured star formation rapidly decreases, following the strong evolution of the number density of the IR luminosity function (see \S\ref{secc:lum_func}), with values that are comparable to the unobscured star formation traced by the rest-frame UV/optical surveys at $z\sim4$.
At higher redshifts, the dust re-processed star formation rate density becomes less dominant than the unobscured star formation. This is because of the combination of the flat faint-end slope of the IR luminosity function and the steep redshift evolution of its number density as compared to the UV luminosity function. At $z=5$, the dust-obscured star formation represents $35\%^{+10\%}_{-25\%}$ of the total SFRD and decreases to $25\%^{+15\%}_{-20\%}$ at $z=6$.
Given that massive, IR bright galaxies dominate the obscured component (as shown in the top panel of Figure \ref{fig:sfrd}), the observed decline of the dust-obscured star formation likely reflects the dearth of massive galaxies at high redshifts.
\subsubsection{Comparison to other measurements and model predictions}
The SFRD described above is in line with previous results from (sub-)mm surveys, although most of them were limited to $z\lesssim4-5$. For example, \citealt{Dunlop2017a} reported a transition from unobscured-dominated star formation to obscured-dominated at $z\approx4$, in very good agreement with our results. Nevertheless, the small area of their survey ($\sim4.5\,$sq.\,arcmin) prevented them from deriving conclusions beyond this redshift. The results from the larger ALMA surveys presented by \citet{Hatsukade2018a} and \citet{Franco2018a}, covering $26\rm\,arcmin^2$ and $69\rm\,arcmin^2$, respectively, also indicate a minor contribution from DSFGs in the $z\approx4-5$ range (see also \citealt{Yamaguchi2019a}).
More recently, \citet{Dudzeviciute2020a} used ALMA observations to investigate the properties of $\sim700$ DSFGs detected over $\sim1\,\rm deg^2$, and inferred a SFRD in the range of $\rho_{obs}\approx3\times10^{-3}$ at $z\approx4-5$ for galaxies with $S_{850\mu m}>1\,\rm mJy$ (see also \citealt{Koprowski2017a}), in very good agreement with our results (particularly if we look at the contribution from $L_{\rm IR}>10^{12}\,L_\odot$ galaxies; see Figure \ref{fig:sfrd_models}).
The `wedding cake' structure of all these surveys ensure that the contributions from the more abundant faint galaxies and the rare bright sources are accounted for (with the possible exception of the most extreme galaxies with ${\rm log}(L_{\rm IR}/L_\odot)\gtrsim13.5$; see Figure \ref{fig:lum_limits}), suggesting that we are not missing any significant population of galaxies. This is also supported by the fact that the number counts predicted by the model are in good agreement with the deepest ALMA observations achieved to-date and with the large-area single-dish telescope surveys (see Figures \ref{fig:num_counts} and \ref{fig:num_counts_short_waves}).
Other studies based on stacking analysis on (sub-)mm maps and using samples of UV/optically-selected galaxies have also concluded that the fraction of star formation that is obscured by dust decreases at high redshifts (e.g. \citealt{Capak2015a,Bouwens2016a,Fudamoto2020a}). Indeed, \citet{Bouwens2020a} complemented previous results from the literature with their dust-corrected SFRs to estimate the obscured and unobscured components of the cosmic history of star formation, and found that the CSFRD transitions from being primarily unobscured to obscured at $z\sim5$, in relatively good agreement with our estimations\footnote{Note, however, that in \citet{Bouwens2020a}
most of the dust-obscured star formation at $z>4$ is produced by faint (UV-selected) galaxies with $L_{\rm IR}<10^{12}\,L_\odot$. Nevertheless, those estimations were done assuming SED templates whose dust temperatures increase with redshift. If there is no significant evolution on the dust temperature (e.g. \citealt{Dudzeviciute2020a}) then the SFRs derived for these galaxies would be lower by a factor of $\sim2.5$, as discussed by the authors. This would decrease the contribution of faint galaxies and bring our results into better agreement.} from direct mm-selected samples.
Our results are also compatible with estimations from other tracers such as radio observations. For example, using the VLA-COSMOS 3\,GHz radio survey, \cite{Novak2017a} inferred lower limits for the SFRD up to $z\sim5$ by integrating the radio luminosity function after converting the radio luminosities to SFRs. These VLA-COSMOS 3\,GHz radio constraints are consistent with our measurements. The authors also provided a completeness-corrected estimation of this quantity by extrapolating the luminosity function to account for the faintest star-forming galaxies. Their estimates show good consistency with our total SFRD, with the possible exception of their last bin at $z\sim5$. Nevertheless, the large extrapolations involved in this process plus the systematic uncertainties that go into calculating SFRD from radio data introduce very large uncertainties in these measurements. The most recent estimates of SFRD from the VLA-COSMOS 3\,GHz survey presented by \citet{Leslie2020a} are in better agreement with our results.
There are, however, a few other studies that have proposed a different picture, with
the obscured component significantly dominating the CSFRD back to $z\sim5$. \citet{Rowan-Robinson2016a} estimated the obscured star formation rate density using a sample of {\it Herschel}-selected galaxies over $\sim20\,\rm deg^2$, finding values far higher than the UV estimations.
Their constraints, however, cannot rule out the possibility that the UV-based measurements are dominant given the large uncertainties on their estimations, as mentioned by the authors. Additionally, it is possible that their calculations might be contaminated by the effect of gravitationally lensing or by overestimated {\it Herschel} flux densities (in view of their extreme SFRs which extend to $20,000\,M_\odot\,\rm yr^{-1}$).
More recently, \citet{Gruppioni2020a}
derive the dust-obscured SFRD using the serendipitously detected sources in the ALPINE survey, finding values in excess to those derived from UV/optical surveys even at $z\sim5$. While the use of highly-confused {\it Herschel} observations might also overestimate the derived luminosities and SFRs, we think that the discrepancy is mainly due because of the possible clustering of serendipitous detections around the original targets. The clustering is expected since the observations' original targets are massive galaxies (log($M/M_\star$)$\gtrsim10.5$) at $z\approx4-6$. In this case, their measurements would be more representative of an over-dense region in the large-scale structure of the Universe. We note, however, that a significant fraction of their estimates comes from the extrapolation of the IR luminosity function since the
reported total SFRD is a factor of $\sim5$ greater than the SFRD estimated when using only the detected sources. Therefore, it is also possible that the assumptions on the extrapolation of the IRLF are responsible for part of the observed discrepancy.
In Figure \ref{fig:sfrd_models} we compare the SFRD derived in this work with results from galaxy evolution models, including the predictions from the cosmological hydrodynamical IllustrisTNG simulations (\citealt{Pillepich2018a}), the {\sc shark} semi-analytic model (\citealt{Lagos2018a}), and the results from the SIDES simulations (\citealt{Bethermin2017a}). Generally speaking, and taking into account the associated uncertainties in these values, our results are in good agreement with the aforementioned studies, pointing towards a convergence between galaxy evolution models and observations.
Finally, we include the density evolution of luminous ($M_{1450}<-26$) quasars (\citealt{Wang2019c}) in Figure \ref{fig:sfrd_models} (scaled for visualization). The shape of the space density of bright quasar strongly resemble that from the bright DSFGs with $L_{\rm IR}>10^{12}\,L_\odot$ (shaded orange region in the figure), suggesting a connection between these two populations and, therefore, between the onset of star formation and the growth of their massive black holes (e.g. \citealt{Wall2005a}).
\begin{figure}[t]\hspace{-0.6cm}
\includegraphics[width=0.52\textwidth]{sfrd_with_theoretical_predictions.eps}
\caption{The total (obscured $+$ unobscured) SFRD derived in this work (gray region) in comparison to the predictions from galaxy evolution models. The solid blue line represents the predictions from the cosmological hydrodynamical IllustrisTNG simulations (\citealt{Pillepich2018a}),
the dashed champagne line those from the {\sc shark} semi-analytic model (\citealt{Lagos2018a}), and the dash-dotted green line the results from the SIDES simulations (\citealt{Bethermin2017a}). Additionally, we include the density evolution of luminous quasars from \citealt{Wang2019c} (scaled for better visualization), which strongly resemble the shape of the SFRD from bright DSFGs with $L_{\rm IR}>10^{12}\,L_\odot$ illustrated by the orange dashed region.
\label{fig:sfrd_models}}
\end{figure}
\section{Discussion and Conclusions}\label{secc:conclusions}
Exploiting the far-infrared and sub-millimeter data aggregated over the last two decades, and particularly, the state-of-the-art ALMA blind surveys at 1.2\,mm, 2\,mm, and 3\,mm, we have constrained the evolution of the IR luminosity function and dust-obscured star formation in the last 13 billion years, back to $z\sim7$.
This is achieved by combining a model of the dusty star-forming galaxy
population with those long-wavelength observations, inferring constraints on the prevalence and
characteristics of these galaxies through measurements of galaxy number counts at different
wavelengths.
By using a library of SEDs an assuming different evolutionary scenarios for the IR luminosity function, the
model makes predictions for galaxy (sub-)millimeter surveys and then works backward to discriminate between the different luminosity function scenarios through the galaxy number counts. Finally, the dust-obscured star formation rate density is estimated by integrating the IR luminosity function (see also \citealt{Casey2018b,Casey2018a,Zavala2018c}).
The model's constraints used in this work include our 2\,mm galaxy number counts
derived as part of the 2
2\,mm MORA survey, the largest ALMA survey to-date sensitive to detect DSFGs and dust-obscured star formation up to the epoch of reionization and the only one carried out at a 2\,mm wavelength (see also Casey et al. in preparation). Additionally, we use the 3\,mm number counts reported in \S\ref{secc:3mm_num_counts}, which are based on our ALMA follow-up observations on the sample of 3\,mm-selected galaxies reported by \citet{Zavala2018c}. Finally, we also include the number counts from the ASPECS survey (\citealt{Gonzalez-Lopez2019a,Gonzalez-Lopez2020a}), a deep ALMA large program at 1.2\,mm and 3\,mm. Altogether, these surveys add a total of $\sim400\,$sq.\,arcmin. of deep observations with arcsecond-resolution, representing the state-of-the-art blind ALMA surveys to-date.
All these data provide good constraints on the model parameters, and thus, on the
IR luminosity function and its evolution with redshift. Based on our best-fit model, which simultaneously reproduces the far-infrared and sub-millimeter data from single-dish telescopes and interferometric surveys, we constrained the faint-end slope of the infrared luminosity function to be flat, with a value of $\alpha_{\rm LF}=-0.42^{+0.02}_{-0.04}$ (see also \citealt{Koprowski2017a,Lim2020a}). This implies that deep pencil-beam observations at (sub-)mm wavelengths would not significantly increase the number of detected sources, in line with previous results from the literature (e.g. \citealt{Popping2020a}).
The characteristic number density of the luminosity function, $\Phi_\star$, decreases as $\Phi_\star\propto(1+z)^{-6.5^{+0.8}_{-1.8}}$ at $z\gtrsim2$, in a similar fashion as the
quasar luminosity function and the density evolution of UV-bright galaxies (\citealt{Hopkins2007a,Finkelstein2015a}), which might suggest that these galaxies occupy similar dark matter halos.
Our constraints on the dust-obscured star formation indicate that the cosmic history of star formation had a peak at $z\approx2-2.5$, and has been dominated by the dust-obscured component during the last 12 billion years, back to $z\sim4$, when the unobscured and obscured contributions were comparable. Beyond this epoch, the dust re-processed star formation rate density was less dominant than the visible star formation, contributing around $35\%^{+10\%}_{-25\%}$ at $z=5$ and $25\%^{+15\%}_{-20\%}$ at $z=6$. This suggests that the bulk ($\gtrsim80\%$) of the star formation activity in the first billion years of the Universe was not dust enshrouded.
Given the massive nature of DSFGs, this drop-off of the obscured component is in line with the decreasing number of high-mass galaxies with increasing redshift (see also \citealt{Dunlop2017a,Bouwens2020a}).\\
Our picture of the history of the cosmic star formation is consistent with previous results from long-wavelength (FIR-to-mm) surveys (e.g. \citealt{Dunlop2017a,Bourne2017a,Hatsukade2018a,Dudzeviciute2020a}), although most of those were limited to $z\lesssim4$. Hence, our results represent a significant progress on our understanding of the prevalence of DSFGs during the first 1.5 billion years of the Universe, and complement the significant efforts carried out using UV/optically selected galaxies (e.g. \citealt{Bouwens2020a}).
The inferred SFRD is also in broad agreement with the most recent predictions from galaxy evolution models (like IllustrisTNG and {\sc shark}; \citealt{Pillepich2018a,Lagos2020}),
which point towards a convergence between models and observations.
With estimations for both the obscured and unobscured components, the shape of the cosmic history of star formation is now constrained out to the end of the epoch of reionization. This measurement, which preserves the galaxy mass assembly history
of the Universe, provides a benchmark against which to compare galaxy formation models and simulations, and a step forward in our understanding of the dust and metal enrichment of the early Universe.
\acknowledgments
{\small
We thank the anonymous referee for a careful review of our manuscript and his/her valuable suggestions.
ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan), together with NRC (Canada), MOST and
ASIAA (Taiwan), and KASI (Republic of Korea), in
cooperation with the Republic of Chile. The Joint
ALMA Observatory is operated by ESO, AUI/NRAO
and NAOJ. The National Radio Astronomy Observatory
is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. The Dunlap Institute is funded through an endowment established by the David Dunlap family and the University of Toronto. JAZ and CMC acknowledge thank the University of Texas at Austin College of Natural Sciences for support. CMC also thanks the National Science Foundation for support through grants AST-1714528 and AST-1814034 and support from the Research Corporation for Science Advancement from a 2019 Cottrell Scholar Award sponsored by IF/THEN, an initiative of Lyda Hill Philanthropies. M.A. has been supported by the grant “CONICYT + PCI + INSTITUTO MAX PLANCK DE ASTRONOMIA MPG190030” and “CONICYT+PCI+REDES 190194”. KIC acknowledges funding from the European Research Council through the award of the Consolidator Grant ID 681627-BUILDUP. JH acknowledges support of the VIDI research programme with project number 639.042.611, which is (partly) financed by the Netherlands Organisation for Scientific Research (NWO). KK acknowledges support from the Knut and Alice Wallenberg Foundation. GEM acknowledges the Villum Fonden research grant 13160 “Gas to stars, stars to dust: tracing star formation across cosmic time” and the Cosmic Dawn Center of Excellence funded by the Danish National Research Foundation under then grant No. 140. E.T. acknowledges support from FONDECYT Regular 1190818, CONICYT PIA ACT172033 and Basal-CATA AFB170002 grants.
Finally, our deepest gratitude to all the people who continue to provide essential services during the COVID-19 pandemic.\\
{\bf Data and materials availability:}
This paper makes use of the following ALMA data: ADS/JAO. ALMA\#2018.1.00231.S, ADS/JAO.ALMA \#2018.1.00478.S, and ADS/JAO.ALMA\#22019.1.00838.S, archived at https://almascience.nrao.edu/alma-data/archive.
}
|
\section{Introduction}
Results about vertex-disjoint and edge-disjoint cycles have received extensive attention by researchers in graph theory. There are two main lines of research in this area. One line of research provides results that give a sufficient condition for a graph to contain a certain number of vertex-disjoint or edge-disjoint cycles (see, for example, \cite{ Chiba, Corradi, Czygrinow2, Dirac-Erdos, Enomoto, Erdos-Posa, Kierstead1, Kierstead2, Kierstead4, Wang}). Another line of research provides results that classify graphs with no pairs of vertex-disjoint cycles (see, for example, \cite{Dirac} and \cite{Lovaz}). We give some results of the latter type here. Such results are particularly useful in the study of graph structure.
All graphs in this paper contain no loops nor parallel edges unless said otherwise. We use $K_5^-$ to denote the graph obtained by deleting an edge from the complete graph on five vertices. The graphs $K_{3,n}^{\prime}$, $K_{3,n}^{\prime \prime}$, and $K_{3,n}^{\prime \prime \prime}$ are obtained by adding, respectively, one, two, or three edges to a partite class of size three in the graph $K_{3,n}$. We denote by $W_n$ the wheel with $n$ spokes. The following result is independently due to Dirac~\cite{Dirac} and Lov\'{a}sz \cite{Lovaz}:
\begin{theorem}\label{diracandlovasz}
A $3$-connected graph has no pair of vertex-disjoint cycles if and only if it is isomorphic to $W_n$, $K_5$, $K_5^{{-}}$, $K_{3,n}$, $K_{3,n}^{\prime}$, $K_{3,n}^{\prime \prime}$, or $K_{3,n}^{\prime \prime \prime}$ for some integer $n$ exceeding two.
\end{theorem}
We say that a graph $H$ is a \textbf{minor} of a graph $G$ if $H$ is obtained from $G$ by contracting edges, deleting edges and deleting vertices. An \textbf{$H$-minor} of a graph $G$ is a minor of $G$ that is isomorphic to a graph $H$. The \textbf{prism} is the graph obtained from two disjoint triangles by adding a perfect matching connecting the vertices of the different triangles. It follows from Menger's Theorem that Theorem \ref{diracandlovasz} is equivalent to the following result:
\begin{theorem}\label{dirac-prism}
A $3$-connected graph has a prism-minor if and only if it is not isomorphic to $W_n$, $K_5$, $K_5^{{-}}$, $K_{3,n}$, $K_{3,n}^{\prime}$, $K_{3,n}^{\prime \prime}$, or $K_{3,n}^{\prime \prime \prime}$ for some integer $n$ exceeding two.
\end{theorem}
We generalize both Theorems \ref{diracandlovasz} and \ref{dirac-prism}. First we discuss the generalization of Theorem \ref{dirac-prism}. We say that a minor $H$ of a graph $G$ \textbf{uses} an edge $uv\in E(G)$ if $H$ has an edge $xy$ such that each $z\in\{x,y\}$ either is in $\{u,v\}$ or is obtained by the identification of a set of vertices intersecting $\{u,v\}$ when making the contractions to obtain $H$. Some authors call this a minor {\bf rooted} on $uv$. Determining when a class of graphs (resp. matroids) has a minor using a specified edges (resp. elements) is often useful and important in the study of structure of graphs and matroids. For instance, Seymour~\cite{Seymour-Adjacency} established a result on $K_4$ minors rooted on pairs of edges and derived results of disjoint paths on graphs. Results on the existence of $U_{2,4}$-minors rooted on one or two elements on matroids with $U_{2,4}$-minors (Bixby~\cite{Bixby} and Seymour~\cite{Seymour-2round}, respectivelly) are classic results in matroid theory; a property like this is called {\bf roundedness} and has some variations, for instance, $3$-connected graphs graphs with a minor isomorphic to $K_4$, $K_5\del e$ or $K_5$ have a minor rooted on the edge set of each edge-set of a triangle, provided they have these graphs as minors (\cite{Reid} and \cite{Czhou}) and $3$-conneted graphs with $K_{3,3}$-minors containing a degree $3$-vertex have a minor rooted on the edges adjacent to this vertex~\cite{Truemper}. Our first main result is the following generalization of Theorem \ref{dirac-prism}:
\begin{theorem}\label{thm-prism}
Suppose that $G$ is a $3$-connected graph on at least six vertices and $uv$ is an edge of $G$. Then $G$ has no prism-minor using $uv$ if and only if
\begin{enumerate}
\item [(a)] for some $n\ge 3$, $G\cong W_n$, $K_{3,n}$, $K'_{3,n}$, $K''_{3,n}$ or $K'''_{3,n}$ or
\item [(b)] $G$ has a vertex $w$ such that $\{u,v,w\}$ is a vertex-cut of $G$ and each connected component $K$ of $G\del \{u,v,w\}$ is a tree with an unique neighbor of $w$ in respect to $G$.
\end{enumerate}
\end{theorem}
There are variants of Theorem \ref{diracandlovasz}, including \cite{Slilaty}. Motivated by Theorem \ref{diracandlovasz}, we consider a much larger class of graphs. The graphs in this class may contain vertex-disjoint cycles, but they do not contain vertex-disjoint cycles which union contains a specified edge. Equivalently, those are the graph with an edge $e$ with the property that, for each cycle $C$ containing $e$, $G-V(C)$ is a forest. The full and detailed characterization is made in Theorem \ref{main-cor}, the second main result in this paper. The statement of this theorem though requires a considerable amount of new terminologies and for this reason we will only state it in the end of Section \ref{sec-rope bridges}. Although the elaborated statement, Theorem \ref{main-cor} has pratical applications and is of independent interest. Indeed, it is used to prove our first main result, Theorem \ref{thm-prism}. Moreover, we apply it to Theorems \ref{main-strong} and \ref{main-strong-2con} to completely characterize $3$-connected and $2$-connected graphs with no pair of edge-disjoint cycles which union contains a specified edge. A more succinct description of the graphs with no pair of vertex-disjoint cycles which union contains a specified edge is given in Theorem \ref{main}.
The following results completely characterizes the $3$-connected and $2$-connected graphs with no pair of edge-disjoint cycles which union contains a specified edge.
\begin{theorem}\label{main-strong}
If $G$ is a $3$-connected graph with an edge $e=u_1u_2$, then $G$ contains no edge-disjoint cycles using $e$ if and only if $G$ has internally disjoint $(u_1,u_2)$-paths $\alpha=u_1,v_1,\dots,v_n,u_2$ and $\beta=u_1,w_1,\dots,w_n,u_2$ not containing $e$ and there is a family $\mathcal{P}$ of pairwise disjoint pairs of consecutive elements of $\{1,\dots,n\}$ such that
\begin{enumerate}
\item [(a)] $V(G)= V(\alpha)\cup V(\beta)$ and
\item [(b)] $E(G)=\{v_iw_j,v_jw_i:\{i,j\}\in \mathcal P\}\cup\{v_kw_k:k$ is in no member of $\mathcal P\}\cup E(\alpha)\cup E(\beta)\cup \{e\}$.
\end{enumerate}
\end{theorem}
\begin{figure}[h]
\centering
\begin{tikzpicture}
\tikzstyle{node_style} =[shape = circle,fill = black,minimum size = 2pt,inner sep=1pt]
\node[node_style] (u1) at (0.0,1.5) {};
\node[node_style] (u2) at (10,1.5){};
\node[node_style] (x1) at (1,1) {};
\node[node_style] (x2) at (2,1) {};
\node[node_style] (x3) at (3,1) {};
\node[node_style] (x4) at (4,1) {};
\node[node_style] (x5) at (5,1) {};
\node[node_style] (x6) at (6,1) {};
\node[node_style] (x7) at (7,1) {};
\node[node_style] (x8) at (8,1) {};
\node[node_style] (x9) at (9,1) {};
\node[node_style] (y1) at (1,2) {};
\node[node_style] (y2) at (2,2) {};
\node[node_style] (y3) at (3,2) {};
\node[node_style] (y4) at (4,2) {};
\node[node_style] (y5) at (5,2) {};
\node[node_style] (y6) at (6,2) {};
\node[node_style] (y7) at (7,2) {};
\node[node_style] (y8) at (8,2) {};
\node[node_style] (y9) at (9,2) {};
\draw (u1)--(x1)--(x2)--(x3)--(x4)--(x5)--(x6)--(x7)--(x8)--(x9)--(u2);
\draw (u1)--(y1)--(y2)--(y3)--(y4)--(y5)--(y6)--(y7)--(y8)--(y9)--(u2);
\draw (u1)--(0,0)--(10,0)--(u2);
\node () at (1,0.7) {$v_1$};
\node () at (2,0.7) {$v_2$};
\node () at (3,0.7) {$v_3$};
\node () at (4,0.7) {$v_4$};
\node () at (5,0.7) {$v_5$};
\node () at (6,0.7) {$v_6$};
\node () at (7,0.7) {$v_7$};
\node () at (8,0.7) {$v_8$};
\node () at (9,0.7) {$v_9$};
\node () at (1,2.3) {$w_1$};
\node () at (2,2.3) {$w_2$};
\node () at (3,2.3) {$w_3$};
\node () at (4,2.3) {$w_4$};
\node () at (5,2.3) {$w_5$};
\node () at (6,2.3) {$w_6$};
\node () at (7,2.3) {$w_7$};
\node () at (8,2.3) {$w_8$};
\node () at (9,2.3) {$w_9$};
\node () at (5,-0.2) {$e$};
\node () at (-0.1,1.7) {$u_1$};
\node () at (10,1.7) {$u_2$};
\draw (x1)--(y1);
\draw (x2)--(y3); \draw (x3)--(y2);
\draw (x4)--(y4);\draw(x5)--(y5);
\draw (x6)--(y7); \draw (x7)--(y6);
\draw (x8)--(y9); \draw (x9)--(y8);
\end{tikzpicture}
\caption{An example for Theorem \ref{main-strong} with $n=9$ and $\mathcal{P}=\big\{\{2,3\}, \{6,7\}, \{8,9\} \big\}$.}
\end{figure}
The following theorem generalizes Theorem \ref{main-strong} for $2$-connected graphs.
\begin{theorem}\label{main-strong-2con}
Suppose that $G$ is a $2$-connected graph and $e$ is an edge of $G$. Then $G$ has no pair of edge-disjoint cycles whose union contains $e$ if and only if, for some integer $n\ge 1$, $G$ has subgraphs $G_1,\dots, G_n$ and an $(n+1)$-elements set of vertices $U:=\{u_0,\dots,u_n\}$ such that
\begin{enumerate}
\item [(a)] $G=(G_1\cup \cdots\cup G_n)+e$ and $e\notin E(G_i)$ for $i=1,\dots,n$;
\item [(b)] $e=u_0u_n$;
\item [(c)] for each $0\le i<j\le n$, $u_iu_j\notin E(G)$ if $\{i,j\}\neq\{0,n\}$;
\item [(d)] for each $i=1,\dots,n$, $V(G_i)\cap U=\{u_{i-1},u_i\}$;
\item [(e)] for $1\le i<j\le n$, $V(G_i)\cap V(G_j)=\emptyset$ if $j>i+1$ and $V(G_i)\cap V(G_{i+1})=\{u_i\}$; and
\item [(f)] for each $i=1,\dots,n$, one of the following assertions holds:
\begin{enumerate}
\item [(f1)] $G_i$ is connected with $V(G_i)=\{u_{i-1},u_i\}$,
\item [(f2)] $G_i$ is a cycle or
\item [(f3)] $G_i+u_{i-i}u_i$ is a subdivision of a $3$-connected graph with no pair of edge-disjoint cycles whose union contains $u_{i-1}u_i$.
\end{enumerate}
\end{enumerate}
\end{theorem}
Next we state a shorter form of Theorem \ref{main-cor}. The \textbf{hub} of a wheel graph $W_n$ with $n \geq 4$ spokes is the vertex of degree $n$. If $w$ is a vertex of a $3$-connected graph $G$, then a $3$-connected graph $H$ with an edge $e=uv$ is said to be obtained from $G$ by \textbf{splitting} the vertex $w$ by $e$ if the contraction of the edge $e$ from $H$ is the graph $G$ with the vertices $u$ and $v$ of $H$ contracting to form the new vertex $w$ of $G$.
\begin{theorem}\label{main}
Let $G$ be a $3$-connected graph with at least six vertices and $e$ be an edge of $G$. Then $G$ contains no pair of vertex-disjoint cycles whose union contains $e$ if and only if one of the following assertions hold:
\begin{enumerate}
\item [(a)] for some $n\ge 4$, $G$ is obtained from a wheel graph with $n$ spokes by possibly doubling some spokes and splitting the hub by the edge $e$ so that the resulting graph is simple and $3$-connected and both vertices incident to $e$ have degree at least four;
\item [(b)] for $e=uv$, $G$ has a vertex $w$ such that $\{u,v,w\}$ is a vertex-cut of $G$ and each connected component $K$ of $G\del \{u,v,w\}$ is a tree with an unique neighbor of $w$ in respect to $G$; or
\item [(c)] $G$ is a $3$-connected minor of a graph of \textbf{type} (c1), (c2) or (c3) using the edge $e$ as shown in Figures \ref{fig-type1}, \ref{fig-type2}, and \ref{fig-type3}.
\end{enumerate}
\end{theorem}
\begin{center}
\begin{figure}[h]\begin{tikzpicture}\centering[scale=3]
\tikzstyle{node_style2} = [shape = circle,fill = black,minimum size = 0.5pt,inner sep=0.5pt]
\tikzstyle{node_style} =[shape = circle,fill = black,minimum size = 2pt,inner sep=1pt]
\begin{scope
\fill[gray!20,opacity=1] (2,0) -- (0.5,-2) -- (1.25,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (1.75,-2) -- (2.5,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (2.75,-2) -- (3.5,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (4,-1) -- (4,1);
\node[node_style] (v) at (2,0) {}; \node[node_style] (x1) at (4,1) {};\node[node_style] (x2) at (5,1) {};
\node[node_style] (y1) at (4,-1) {};\node[node_style] (y2) at (5,-1) {};
\node[node_style2] () at (0.9,-2) {};
\node[node_style2] () at (1,-2) {};
\node[node_style2] () at (1.1,-2) {};
\node[node_style] (lfl) at (0.5,-2) {};\node[node_style] (lfm) at (0.75,-2) {}; \node[node_style] (lfr) at (1.25,-2) {};
\node[node_style2] () at (1.4,-2){};
\node[node_style2] () at (1.5,-2){};
\node[node_style2] () at (1.6,-2){};
\node[node_style2] () at (2.15,-2){};
\node[node_style2] () at (2.25,-2){};
\node[node_style2] () at (2.35,-2){};
\node[node_style] (mfl) at (1.75,-2) {};\node[node_style] (mfm) at (2,-2) {}; \node[node_style] (mfr) at (2.5,-2) {};
\node[node_style2] () at (3.15,-2){};
\node[node_style2] () at (3.25,-2){};
\node[node_style2] () at (3.35,-2){};
\node[node_style] (rfl) at (2.75,-2) {};\node[node_style] (rfm) at (3,-2) {}; \node[node_style] (rfr) at (3.5,-2) {};
\draw (y2)--(y1)--(v)--(x1)--(x2);
\draw (v)--(lfl)--(lfm)--(v)--(lfr);
\draw (v)--(mfl)--(mfm)--(v)--(mfr);
\draw (v)--(rfl)--(rfm)--(v)--(rfr);
\node[node_style] (1) at (4,-0.6){};
\node[node_style] (2) at (4,-0.2){};
\node[node_style] (3) at (4,0.6){};
\node[node_style2] () at (4,0.1){};
\node[node_style2] () at (4,0.2){};
\node[node_style2] () at (4,0.3){};
\draw (y1)--(1)--(v)--(2)--(1)--(v)--(3)--(x1);
\draw (y2)--(rfr);
\draw (y2)--(3.7,-2.1)--(2.5,-2.1)--(2.5,-2);
\draw (y2)--(3.9,-2.2)--(1.25,-2.2)--(1.25,-2);
\draw (x2)--(5.2,0)--(5.2,-2.4)--(2.75,-2.4)--(rfl);
\draw (x2)--(5.3,0)--(5.3,-2.5)--(1.75,-2.5)--(mfl);
\draw (x2)--(5.4,0)--(5.4,-2.6)--(0.5,-2.6)--(lfl);
\node[node_style] (x3) at (5,1.5){};\node[node_style] (x4) at (4.5,1.5){};
\node[node_style] (y3) at (5,2){};\node[node_style] (y4) at (4.5,2){};
\draw (x3)--(y4);\draw(y3)--(x4);
\node[node_style] (x5) at (4,1.5){};\node[node_style] (x6) at (3.5,1.5){};
\node[node_style] (y5) at (4,2){};\node[node_style] (y6) at (3.5,2){};
\draw (x5)--(y6);\draw(y5)--(x6);
\node[node_style] (x7) at (3,1.5){};\node[node_style] (x8) at (2.5,1.5){};
\node[node_style] (y7) at (3,2){};\node[node_style] (y8) at (2.5,2){};
\draw (x7)--(y8);\draw(y7)--(x8);
\node[node_style] (x9) at (2,1.5){};\node[node_style] (x10) at (1.5,1.5){};
\node[node_style] (y9) at (2,2){};\node[node_style] (y10) at (1.5,2){};
\draw (x9)--(y10);\draw(y9)--(x10);
\node[node_style] (x11) at (1,1.5){};\node[node_style] (x12) at (0.5,1.5){};
\node[node_style] (y11) at (1,2){};\node[node_style] (y12) at (0.5,2){};
\draw (x11)--(y12);\draw(y11)--(x12);
\draw (x2)--(x3)--(0.3,1.5);
\draw (y2)--(5.6,-1)--(5.6,2)--(0.3,2);
\end{scope}
\node (dots) at (0,1.5) {$\dots$};
\node (dots) at (0,2) {$\dots$};
\node (e) at (0,0.2) {$e$};
\draw (-2,0)--(2,0);
\begin{scope}[xscale=-1,yscale=1]
\fill[gray!20,opacity=1] (2,0) -- (0.5,-2) -- (1.25,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (1.75,-2) -- (2.5,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (2.75,-2) -- (3.5,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (4,-1) -- (4,1);
\node[node_style] (v) at (2,0) {};
\node[node_style] (x1) at (4,1) {};\node[node_style] (x2) at (5,1) {};
\node[node_style] (y1) at (4,-1) {};\node[node_style] (y2) at (5,-1) {};
\node[node_style2] () at (0.9,-2) {};
\node[node_style2] () at (1,-2) {};
\node[node_style2] () at (1.1,-2) {};
\node[node_style] (lfl) at (0.5,-2) {};\node[node_style] (lfm) at (0.75,-2) {}; \node[node_style] (lfr) at (1.25,-2) {};
\node[node_style2] () at (1.4,-2){};
\node[node_style2] () at (1.5,-2){};
\node[node_style2] () at (1.6,-2){};
\node[node_style2] () at (2.15,-2){};
\node[node_style2] () at (2.25,-2){};
\node[node_style2] () at (2.35,-2){};
\node[node_style] (mfl) at (1.75,-2) {};\node[node_style] (mfm) at (2,-2) {}; \node[node_style] (mfr) at (2.5,-2) {};
\node[node_style2] () at (3.15,-2){};
\node[node_style2] () at (3.25,-2){};
\node[node_style2] () at (3.35,-2){};
\node[node_style] (rfl) at (2.75,-2) {};\node[node_style] (rfm) at (3,-2) {}; \node[node_style] (rfr) at (3.5,-2) {};
\draw (y2)--(y1)--(v)--(x1)--(x2);
\draw (v)--(lfl)--(lfm)--(v)--(lfr);
\draw (v)--(mfl)--(mfm)--(v)--(mfr);
\draw (v)--(rfl)--(rfm)--(v)--(rfr);
\node[node_style] (1) at (4,-0.6){};
\node[node_style] (2) at (4,-0.2){};
\node[node_style] (3) at (4,0.6){};
\node[node_style2] () at (4,0.1){};
\node[node_style2] () at (4,0.2){};
\node[node_style2] () at (4,0.3){};
\draw (y1)--(1)--(v)--(2)--(1)--(v)--(3)--(x1);
\draw (y2)--(rfr);
\draw (y2)--(3.7,-2.1)--(2.5,-2.1)--(2.5,-2);
\draw (y2)--(3.9,-2.2)--(1.25,-2.2)--(1.25,-2);
\draw (x2)--(5.2,0)--(5.2,-2.4)--(2.75,-2.4)--(rfl);
\draw (x2)--(5.3,0)--(5.3,-2.5)--(1.75,-2.5)--(mfl);
\draw (x2)--(5.4,0)--(5.4,-2.6)--(0.5,-2.6)--(lfl);
\node[node_style] (x3) at (5,1.5){};\node[node_style] (x4) at (4.5,1.5){};
\node[node_style] (y3) at (5,2){};\node[node_style] (y4) at (4.5,2){};
\draw (x3)--(y4);\draw(y3)--(x4);
\node[node_style] (x5) at (4,1.5){};\node[node_style] (x6) at (3.5,1.5){};
\node[node_style] (y5) at (4,2){};\node[node_style] (y6) at (3.5,2){};
\draw (x5)--(y6);\draw(y5)--(x6);
\node[node_style] (x7) at (3,1.5){};\node[node_style] (x8) at (2.5,1.5){};
\node[node_style] (y7) at (3,2){};\node[node_style] (y8) at (2.5,2){};
\draw (x7)--(y8);\draw(y7)--(x8);
\node[node_style] (x9) at (2,1.5){};\node[node_style] (x10) at (1.5,1.5){};
\node[node_style] (y9) at (2,2){};\node[node_style] (y10) at (1.5,2){};
\draw (x9)--(y10);\draw(y9)--(x10);
\node[node_style] (x11) at (1,1.5){};\node[node_style] (x12) at (0.5,1.5){};
\node[node_style] (y11) at (1,2){};\node[node_style] (y12) at (0.5,2){};
\draw (x11)--(y12);\draw(y11)--(x12);
\draw (x2)--(x3)--(0.3,1.5);
\draw (y2)--(5.6,-1)--(5.6,2)--(0.3,2);
\end{scope}
\end{tikzpicture}
\caption{The type (c1) graphs.}
\label{fig-type1}
\end{figure}
\begin{figure}[h]\begin{tikzpicture}\centering
\tikzstyle{node_style2} = [shape = circle,fill = black,minimum size = 0.5pt,inner sep=0.5pt]
\tikzstyle{node_style} =[shape = circle,fill = black,minimum size = 2pt,inner sep=1pt]
\begin{scope
\fill[gray!20,opacity=1] (2,0) -- (0.5,-2) -- (1.25,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (1.75,-2) -- (2.5,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (2.75,-2) -- (3.5,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (4,-1) -- (4.4,-0.2);
\fill[gray!20,opacity=1] (2,0) -- (4,1) -- (4.4, 0.2);
\node[node_style] (v) at (2,0) {};
\node[node_style] (x1) at (4,1) {};\node[node_style] (x2) at (5,1) {};
\node[node_style] (y1) at (4,-1) {};\node[node_style] (y2) at (5,-1) {};
\node[node_style2] () at (0.9,-2) {};
\node[node_style2] () at (1,-2) {};
\node[node_style2] () at (1.1,-2) {};
\node[node_style] (lfl) at (0.5,-2) {};\node[node_style] (lfm) at (0.75,-2) {}; \node[node_style] (lfr) at (1.25,-2) {};
\node[node_style2] () at (1.4,-2){};
\node[node_style2] () at (1.5,-2){};
\node[node_style2] () at (1.6,-2){};
\node[node_style2] () at (2.15,-2){};
\node[node_style2] () at (2.25,-2){};
\node[node_style2] () at (2.35,-2){};
\node[node_style] (mfl) at (1.75,-2) {};\node[node_style] (mfm) at (2,-2) {}; \node[node_style] (mfr) at (2.5,-2) {};
\node[node_style2] () at (3.15,-2){};
\node[node_style2] () at (3.25,-2){};
\node[node_style2] () at (3.35,-2){};
\node[node_style] (rfl) at (2.75,-2) {};\node[node_style] (rfm) at (3,-2) {}; \node[node_style] (rfr) at (3.5,-2) {};
\draw (y2)--(y1)--(v)--(x1)--(x2);
\draw (v)--(lfl)--(lfm)--(v)--(lfr);
\draw (v)--(mfl)--(mfm)--(v)--(mfr);
\draw (v)--(rfl)--(rfm)--(v)--(rfr);
\node[node_style] (lofl) at (4.1,-0.8){};\node[node_style] (lofu) at (4.4,-0.2){}; \draw (y1)--(lofl)--(v)--(lofu);
\node[node_style2] (ldot1) at (4.2,-0.6){};
\node[node_style2] (ldot2) at (4.25,-0.5){};
\node[node_style2] (ldot3) at (4.3,-0.4){};
\node[node_style] (upfu) at (4.1,0.8){};\node[node_style] (upfl) at (4.4, 0.2){}; \draw (x1)--(upfu)--(v)--(upfl);
\node[node_style2] (ldot1) at (4.2,0.6){};
\node[node_style2] (ldot2) at (4.25,0.5){};
\node[node_style2] (ldot3) at (4.3,0.4){};
\draw (lofu)--(x2);
\draw (upfl)--(y2);
\draw (y2)--(rfr);
\draw (y2)--(3.7,-2.1)--(2.5,-2.1)--(2.5,-2);
\draw (y2)--(3.9,-2.2)--(1.25,-2.2)--(1.25,-2);
\draw (x2)--(5.2,0)--(5.2,-2.4)--(2.75,-2.4)--(rfl);
\draw (x2)--(5.3,0)--(5.3,-2.5)--(1.75,-2.5)--(mfl);
\draw (x2)--(5.4,0)--(5.4,-2.6)--(0.5,-2.6)--(lfl);
\node[node_style] (x3) at (5,1.5){};\node[node_style] (x4) at (4.5,1.5){};
\node[node_style] (y3) at (5,2){};\node[node_style] (y4) at (4.5,2){};
\draw (x3)--(y4);\draw(y3)--(x4);
\node[node_style] (x5) at (4,1.5){};\node[node_style] (x6) at (3.5,1.5){};
\node[node_style] (y5) at (4,2){};\node[node_style] (y6) at (3.5,2){};
\draw (x5)--(y6);\draw(y5)--(x6);
\node[node_style] (x7) at (3,1.5){};\node[node_style] (x8) at (2.5,1.5){};
\node[node_style] (y7) at (3,2){};\node[node_style] (y8) at (2.5,2){};
\draw (x7)--(y8);\draw(y7)--(x8);
\node[node_style] (x9) at (2,1.5){};\node[node_style] (x10) at (1.5,1.5){};
\node[node_style] (y9) at (2,2){};\node[node_style] (y10) at (1.5,2){};
\draw (x9)--(y10);\draw(y9)--(x10);
\node[node_style] (x11) at (1,1.5){};\node[node_style] (x12) at (0.5,1.5){};
\node[node_style] (y11) at (1,2){};\node[node_style] (y12) at (0.5,2){};
\draw (x11)--(y12);\draw(y11)--(x12);
\draw (x2)--(x3)--(0.3,1.5);
\draw (y2)--(5.6,-1)--(5.6,2)--(0.3,2);
\end{scope}
\node (dots) at (0,1.5) {$\dots$};
\node (dots) at (0,2) {$\dots$};
\node (e) at (0,0.2) {$e$};
\draw (-2,0)--(2,0);
\begin{scope}[xscale=-1,yscale=1]
\fill[gray!20,opacity=1] (2,0) -- (0.5,-2) -- (1.25,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (1.75,-2) -- (2.5,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (2.75,-2) -- (3.5,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (4,-1) -- (4.4,-0.2);
\fill[gray!20,opacity=1] (2,0) -- (4,1) -- (4.4, 0.2);
\node[node_style] (v) at (2,0) {};
\node[node_style] (x1) at (4,1) {};\node[node_style] (x2) at (5,1) {};
\node[node_style] (y1) at (4,-1) {};\node[node_style] (y2) at (5,-1) {};
\node[node_style2] () at (0.9,-2) {};
\node[node_style2] () at (1,-2) {};
\node[node_style2] () at (1.1,-2) {};
\node[node_style] (lfl) at (0.5,-2) {};\node[node_style] (lfm) at (0.75,-2) {}; \node[node_style] (lfr) at (1.25,-2) {};
\node[node_style2] () at (1.4,-2){};
\node[node_style2] () at (1.5,-2){};
\node[node_style2] () at (1.6,-2){};
\node[node_style2] () at (2.15,-2){};
\node[node_style2] () at (2.25,-2){};
\node[node_style2] () at (2.35,-2){};
\node[node_style] (mfl) at (1.75,-2) {};\node[node_style] (mfm) at (2,-2) {}; \node[node_style] (mfr) at (2.5,-2) {};
\node[node_style2] () at (3.15,-2){};
\node[node_style2] () at (3.25,-2){};
\node[node_style2] () at (3.35,-2){};
\node[node_style] (rfl) at (2.75,-2) {};\node[node_style] (rfm) at (3,-2) {}; \node[node_style] (rfr) at (3.5,-2) {};
\draw (y2)--(y1)--(v)--(x1)--(x2);
\draw (v)--(lfl)--(lfm)--(v)--(lfr);
\draw (v)--(mfl)--(mfm)--(v)--(mfr);
\draw (v)--(rfl)--(rfm)--(v)--(rfr);
\node[node_style] (lofl) at (4.1,-0.8){};\node[node_style] (lofu) at (4.4,-0.2){}; \draw (y1)--(lofl)--(v)--(lofu);
\node[node_style2] (ldot1) at (4.2,-0.6){};
\node[node_style2] (ldot2) at (4.25,-0.5){};
\node[node_style2] (ldot3) at (4.3,-0.4){};
\node[node_style] (upfu) at (4.1,0.8){};\node[node_style] (upfl) at (4.4, 0.2){}; \draw (x1)--(upfu)--(v)--(upfl);
\node[node_style2] (ldot1) at (4.2,0.6){};
\node[node_style2] (ldot2) at (4.25,0.5){};
\node[node_style2] (ldot3) at (4.3,0.4){};
\draw (lofu)--(x2);
\draw (upfl)--(y2);
\draw (y2)--(rfr);
\draw (y2)--(3.7,-2.1)--(2.5,-2.1)--(2.5,-2);
\draw (y2)--(3.9,-2.2)--(1.25,-2.2)--(1.25,-2);
\draw (x2)--(5.2,0)--(5.2,-2.4)--(2.75,-2.4)--(rfl);
\draw (x2)--(5.3,0)--(5.3,-2.5)--(1.75,-2.5)--(mfl);
\draw (x2)--(5.4,0)--(5.4,-2.6)--(0.5,-2.6)--(lfl);
\node[node_style] (x3) at (5,1.5){};\node[node_style] (x4) at (4.5,1.5){};
\node[node_style] (y3) at (5,2){};\node[node_style] (y4) at (4.5,2){};
\draw (x3)--(y4);\draw(y3)--(x4);
\node[node_style] (x5) at (4,1.5){};\node[node_style] (x6) at (3.5,1.5){};
\node[node_style] (y5) at (4,2){};\node[node_style] (y6) at (3.5,2){};
\draw (x5)--(y6);\draw(y5)--(x6);
\node[node_style] (x7) at (3,1.5){};\node[node_style] (x8) at (2.5,1.5){};
\node[node_style] (y7) at (3,2){};\node[node_style] (y8) at (2.5,2){};
\draw (x7)--(y8);\draw(y7)--(x8);
\node[node_style] (x9) at (2,1.5){};\node[node_style] (x10) at (1.5,1.5){};
\node[node_style] (y9) at (2,2){};\node[node_style] (y10) at (1.5,2){};
\draw (x9)--(y10);\draw(y9)--(x10);
\node[node_style] (x11) at (1,1.5){};\node[node_style] (x12) at (0.5,1.5){};
\node[node_style] (y11) at (1,2){};\node[node_style] (y12) at (0.5,2){};
\draw (x11)--(y12);\draw(y11)--(x12);
\draw (x2)--(x3)--(0.3,1.5);
\draw (y2)--(5.6,-1)--(5.6,2)--(0.3,2);
\end{scope}
\end{tikzpicture}
\caption{The type (c2) graphs.}
\label{fig-type2}
\end{figure}
\begin{figure}[h]\begin{tikzpicture}\centering[scale=3]
\tikzstyle{node_style2} = [shape = circle,fill = black,minimum size = 0.5pt,inner sep=0.5pt]
\tikzstyle{node_style} =[shape = circle,fill = black,minimum size = 2pt,inner sep=1pt]
\begin{scope
\fill[gray!20,opacity=1] (2,0) -- (0.5,-2) -- (1.25,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (1.75,-2) -- (2.5,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (2.75,-2) -- (3.5,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (4,-1) -- (4,1);
\node[node_style] (v) at (2,0) {};
\node[node_style] (x1) at (4,1) {};\node[node_style] (x2) at (5,1) {};
\node[node_style] (y1) at (4,-1) {};\node[node_style] (y2) at (5,-1) {};
\node[node_style2] () at (0.9,-2) {};
\node[node_style2] () at (1,-2) {};
\node[node_style2] () at (1.1,-2) {};
\node[node_style] (lfl) at (0.5,-2) {};\node[node_style] (lfm) at (0.75,-2) {}; \node[node_style] (lfr) at (1.25,-2) {};
\node[node_style2] () at (1.4,-2){};
\node[node_style2] () at (1.5,-2){};
\node[node_style2] () at (1.6,-2){};
\node[node_style2] () at (2.15,-2){};
\node[node_style2] () at (2.25,-2){};
\node[node_style2] () at (2.35,-2){};
\node[node_style] (mfl) at (1.75,-2) {};\node[node_style] (mfm) at (2,-2) {}; \node[node_style] (mfr) at (2.5,-2) {};
\node[node_style2] () at (3.15,-2){};
\node[node_style2] () at (3.25,-2){};
\node[node_style2] () at (3.35,-2){};
\node[node_style] (rfl) at (2.75,-2) {};\node[node_style] (rfm) at (3,-2) {}; \node[node_style] (rfr) at (3.5,-2) {};
\draw (y2)--(y1)--(v)--(x1)--(x2);
\draw (v)--(lfl)--(lfm)--(v)--(lfr);
\draw (v)--(mfl)--(mfm)--(v)--(mfr);
\draw (v)--(rfl)--(rfm)--(v)--(rfr);
\node[node_style] (1) at (4,-0.6){};
\node[node_style] (2) at (4,-0.2){};
\node[node_style] (3) at (4,0.6){};
\node[node_style2] () at (4,0.1){};
\node[node_style2] () at (4,0.2){};
\node[node_style2] () at (4,0.3){};
\draw (y1)--(1)--(v)--(2)--(1)--(v)--(3)--(x1);
\draw (y2)--(rfr);
\draw (y2)--(3.7,-2.1)--(2.5,-2.1)--(2.5,-2);
\draw (y2)--(3.9,-2.2)--(1.25,-2.2)--(1.25,-2);
\draw (x2)--(5.2,0)--(5.2,-2.4)--(2.75,-2.4)--(rfl);
\draw (x2)--(5.3,0)--(5.3,-2.5)--(1.75,-2.5)--(mfl);
\draw (x2)--(5.4,0)--(5.4,-2.6)--(0.5,-2.6)--(lfl);
\node[node_style] (x3) at (5,1.5){};\node[node_style] (x4) at (4.5,1.5){};
\node[node_style] (y3) at (5,2){};\node[node_style] (y4) at (4.5,2){};
\draw (x3)--(y4);\draw(y3)--(x4);
\node[node_style] (x5) at (4,1.5){};\node[node_style] (x6) at (3.5,1.5){};
\node[node_style] (y5) at (4,2){};\node[node_style] (y6) at (3.5,2){};
\draw (x5)--(y6);\draw(y5)--(x6);
\node[node_style] (x7) at (3,1.5){};\node[node_style] (x8) at (2.5,1.5){};
\node[node_style] (y7) at (3,2){};\node[node_style] (y8) at (2.5,2){};
\draw (x7)--(y8);\draw(y7)--(x8);
\node[node_style] (x9) at (2,1.5){};\node[node_style] (x10) at (1.5,1.5){};
\node[node_style] (y9) at (2,2){};\node[node_style] (y10) at (1.5,2){};
\draw (x9)--(y10);\draw(y9)--(x10);
\node[node_style] (x11) at (1,1.5){};\node[node_style] (x12) at (0.5,1.5){};
\node[node_style] (y11) at (1,2){};\node[node_style] (y12) at (0.5,2){};
\draw (x11)--(y12);\draw(y11)--(x12);
\draw (x2)--(x3)--(0.3,1.5);
\draw (y2)--(5.6,-1)--(5.6,2)--(0.3,2);
\end{scope}
\node (dots) at (0,1.5) {$\dots$};
\node (dots) at (0,2) {$\dots$};
\node (e) at (0,0.2) {$e$};
\draw (-2,0)--(2,0);
\begin{scope}[xscale=-1,yscale=1]
\fill[gray!20,opacity=1] (2,0) -- (0.5,-2) -- (1.25,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (1.75,-2) -- (2.5,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (2.75,-2) -- (3.5,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (4,-1) -- (4.4,-0.2);
\fill[gray!20,opacity=1] (2,0) -- (4,1) -- (4.4, 0.2);
\node[node_style] (v) at (2,0) {};
\node[node_style] (x1) at (4,1) {};\node[node_style] (x2) at (5,1) {};
\node[node_style] (y1) at (4,-1) {};\node[node_style] (y2) at (5,-1) {};
\node[node_style2] () at (0.9,-2) {};
\node[node_style2] () at (1,-2) {};
\node[node_style2] () at (1.1,-2) {};
\node[node_style] (lfl) at (0.5,-2) {};\node[node_style] (lfm) at (0.75,-2) {}; \node[node_style] (lfr) at (1.25,-2) {};
\node[node_style2] () at (1.4,-2){};
\node[node_style2] () at (1.5,-2){};
\node[node_style2] () at (1.6,-2){};
\node[node_style2] () at (2.15,-2){};
\node[node_style2] () at (2.25,-2){};
\node[node_style2] () at (2.35,-2){};
\node[node_style] (mfl) at (1.75,-2) {};\node[node_style] (mfm) at (2,-2) {}; \node[node_style] (mfr) at (2.5,-2) {};
\node[node_style2] () at (3.15,-2){};
\node[node_style2] () at (3.25,-2){};
\node[node_style2] () at (3.35,-2){};
\node[node_style] (rfl) at (2.75,-2) {};\node[node_style] (rfm) at (3,-2) {}; \node[node_style] (rfr) at (3.5,-2) {};
\draw (y2)--(y1)--(v)--(x1)--(x2);
\draw (v)--(lfl)--(lfm)--(v)--(lfr);
\draw (v)--(mfl)--(mfm)--(v)--(mfr);
\draw (v)--(rfl)--(rfm)--(v)--(rfr);
\node[node_style] (lofl) at (4.1,-0.8){};\node[node_style] (lofu) at (4.4,-0.2){}; \draw (y1)--(lofl)--(v)--(lofu);
\node[node_style2] (ldot1) at (4.2,-0.6){};
\node[node_style2] (ldot2) at (4.25,-0.5){};
\node[node_style2] (ldot3) at (4.3,-0.4){};
\node[node_style] (upfu) at (4.1,0.8){};\node[node_style] (upfl) at (4.4, 0.2){}; \draw (x1)--(upfu)--(v)--(upfl);
\node[node_style2] (ldot1) at (4.2,0.6){};
\node[node_style2] (ldot2) at (4.25,0.5){};
\node[node_style2] (ldot3) at (4.3,0.4){};
\draw (lofu)--(x2);
\draw (upfl)--(y2);
\draw (y2)--(rfr);
\draw (y2)--(3.7,-2.1)--(2.5,-2.1)--(2.5,-2);
\draw (y2)--(3.9,-2.2)--(1.25,-2.2)--(1.25,-2);
\draw (x2)--(5.2,0)--(5.2,-2.4)--(2.75,-2.4)--(rfl);
\draw (x2)--(5.3,0)--(5.3,-2.5)--(1.75,-2.5)--(mfl);
\draw (x2)--(5.4,0)--(5.4,-2.6)--(0.5,-2.6)--(lfl);
\node[node_style] (x3) at (5,1.5){};\node[node_style] (x4) at (4.5,1.5){};
\node[node_style] (y3) at (5,2){};\node[node_style] (y4) at (4.5,2){};
\draw (x3)--(y4);\draw(y3)--(x4);
\node[node_style] (x5) at (4,1.5){};\node[node_style] (x6) at (3.5,1.5){};
\node[node_style] (y5) at (4,2){};\node[node_style] (y6) at (3.5,2){};
\draw (x5)--(y6);\draw(y5)--(x6);
\node[node_style] (x7) at (3,1.5){};\node[node_style] (x8) at (2.5,1.5){};
\node[node_style] (y7) at (3,2){};\node[node_style] (y8) at (2.5,2){};
\draw (x7)--(y8);\draw(y7)--(x8);
\node[node_style] (x9) at (2,1.5){};\node[node_style] (x10) at (1.5,1.5){};
\node[node_style] (y9) at (2,2){};\node[node_style] (y10) at (1.5,2){};
\draw (x9)--(y10);\draw(y9)--(x10);
\node[node_style] (x11) at (1,1.5){};\node[node_style] (x12) at (0.5,1.5){};
\node[node_style] (y11) at (1,2){};\node[node_style] (y12) at (0.5,2){};
\draw (x11)--(y12);\draw(y11)--(x12);
\draw (x2)--(x3)--(0.3,1.5);
\draw (y2)--(5.6,-1)--(5.6,2)--(0.3,2);
\end{scope}
\end{tikzpicture}
\caption{The type (c3) graphs.}
\label{fig-type3}
\end{figure}
\end{center}
A graph as in Theorem~\ref{main} will be called of \label{pg-type}\textbf{type} (a), (b), or (c) according to which case of the theorem conclusion it satisfies. A graph may be of more than one type; for instance, the planar type (a) graphs are also type (c) graphs.
For $4$-connected graphs, we have a strengthening of the previous result. A \textbf{double-wheel} graph is obtained from a cycle $C$ by adding two new vertices, called \textbf{hubs} of the double-wheel, and then adding an edge linking each vertex of $C$ to each one of the hubs. From Theorem \ref{main}, we conclude:
\begin{corollary}\label{4-connected-main}
Let $G$ be a $4$-connected graph with a fixed edge $e$. Then, either $G$ contains a pair of vertex-disjoint cycles using $e$ or $G$ is obtained from a double-wheel by adding the edge $e$ between the two hubs.
\end{corollary}
\begin{proof}Let $G$ be a $4$-connected graph with and edge $e$ such that no pair of vertex-disjoint cycles has the property that $e$ is in one of them. So $G$ is a graph of type (a), (b) or (c) like in Theorme \ref{main}. First suppsoe that $G$ is a type (a) graph, obtained from a wheel as described in Theore \ref{main}. If all spokes are doubed, $G$ is a double wheel, then $G$ is obtained from a double wheel adding $e$ linking the hubs. If some spoke is not doubled, then it is incident to a degree $3$-vertex of $G$ and $G$ is not $3$-connected. As no type (b) graph is $4$-connected, to finish the proof it suffices to prove that type (c) graphs are also not $4$-connected. This follows from the fact that if $C$ is a $k$-vertex cut in a $k$-connected graph $G$ and $X\cont E(G)$, then either the vertices of $G/X$ inherited from $C$ separate $G$ or all but one connected components of $G\del C$ are eliminated when contracting $X$. It is a straighforward verification to use this fact to deduce that the contraction of any set of edges avoiding $e$ in the graphs of Figures \ref{fig-type1}, \ref{fig-type2} or \ref{fig-type3} will result in a non $4$-connected graph.
\end{proof}
One last remark is that the analogous questions for disjoint cycles containing a specified vertex is reduced to Dirac's characterization, as we can see in the next proposition.
\begin{proposition}
Let $G$ be a $3$-connected graph and $v\in V(G)$. If $G$ has no pair of vertex-disjoint cycles using $v$, then $G$ has no pair of vertex-disjoint cycles.
\end{proposition}
\begin{proof}
Suppose that $G$ and $v$ contradict the proposition. Let $C$ and $D$ be vertex-disjoint cycles of $G$. So $v\notin V(C)\cup V(D)$. By Menger's Theorem, there are three $(v,V(C)\cup V(D))$-paths in $G$ meeting only in $v$. We may assume that two of them have an endvertex in $C$. Note that the union of $C$ with those two paths contains a cycle containing $v$ and avoiding $D$, a contradiction.
\end{proof}
The structure of this paper is described next. For this description, we let $e=u_1u_2$ be an edge in a $3$-connected graph $G$ that is not contained in the union of two vertex-disjoint cycles. In Section \ref{sec-lemmas}, we give some terminologies, establish some lemmas, prove that the endvertices of $e$ are in a $3$- or $4$-vertex cut of $G$ and characterize family (a) of Theorem \ref{main}. In Section \ref{sec-one-two}, we introduce, in the remaining cases, a classification for the $4$-vertex-cuts containing $\{u_1,u_2\}$ and characterize the connected components for the removal of a certain type of a $4$-vertex-cut containing $\{u_1,u_2\}$. For the remaining case, in Section \ref{sec-rope bridges}, we define constructively a type of graph called rope-bridge and prove that it is the remaining type of connected component for the removal of a $3$- or $4$-vertex-cut incident to $e$. In the end of Section \ref{sec-rope bridges}, we prove Theorem \ref{main-cor}, which implies Theorem \ref{main}. Section \ref{sec-strong} is dedicated to proving Theorems \ref{main-strong} and \ref{main-strong-2con}. In Section \ref{sec-prism} we prove Theorem \ref{thm-prism}.
\section{Lemmas}\label{sec-lemmas}
We first establish some terminology before giving useful lemmas for the proof of the main results. The paths we consider are simple, we think of paths and cycles both as subgraphs and (cyclic) sequences of vertices. The number of vertices and edges in a graph $G$ are denoted by $|G|$ and $\|G\|$, respectivelly. For vertices $v_1,\dots,v_n$ and a subgraph $H$ of a graph $G$, we say that a \textbf{cycle of the form} $v_1,\dots,v_k,H,v_{k+1},\dots,v_n,v_1$ is a cycle that begins in $v_1$, follows through $v_1,\dots,v_k$, then, through a path of $H$, and, then, returns to $v_1$ through $v_{k+1},\dots,v_n$. We simplify the notations $X\cup\{x\}$ and $X-\{x\}$ by $X\cup x$ and $X-x$ respectively.
Let $e$ be a fixed edge of a graph $G$. We say that $G$ is an \textbf{$e$-Dirac} graph if $G$ contains no vertex-disjoint cycles using $e$, in this case $e$ is a \textbf{Dirac} edge of $G$. If $G$ contains no edge-disjoint cycles using $e$, then $G$ is said to be \textbf{strongly $e$-Dirac}. An elementary observation about the class of $e$-Dirac graphs is that it is closed under minors that use $e$. That is, if $e$ is a Dirac edge of $G$ and $f\in E(G)-e$, then $e$ is also a Dirac edge of $G\del f$ and $G/f$. But the class of strongly $e$-Dirac graphs is closed under deletions but not under contractions of other edges than $e$. For example, when $G$ is the \textbf{prism}, a graph on six vertices obtained by adding a perfect matching between the vertices of two disjoint triangles, and $e$ and $f$ are distinct edges in this perfect matching, then $G$ is strongly $e$-Dirac, but $G/f$ is not.
The proof of the next lemma is straightforward.
\begin{lemma}\label{complete}
Let $e=u_1u_2$ be an edge in a $3$-connected graph $G$ and suppose that $\{u_1,u_2\}$ is contained in no vertex-cut of $G$. Then $G\del\{u_1,u_2\}$ is complete. Moreover, if $G$ is $e$-Dirac, then $|G|\le 5$.
\end{lemma}
Throughout the paper, until the proof of Theorem \ref{main}, we let $G$ be a $3$-connected graph with a Dirac edge $e=u_1u_2$ and a minimum sized set $S$ such that $S^+:=S\cup\{u_1,u_2\}$ is a vertex-cut of $G$. We let $G_1,\dots,G_\kappa$ be the connected components of $G\del S^+$. For a subgraph $H$ of $G$ and $X\cont V(G)-V(H)$ we denote by $H+X$ the graph obtained from $H$ by adding the vertices of $X$ and all edges of $G$ linking the vertices of $X$ with vertices of $V(H)\cup X$.
The minimality of $S$ implies the following elementary result.
\begin{lemma}\label{edges from S}
For each $w\in S$ and each $1 \leq i \leq \kappa$, there is an edge from $w$ to a vertex of $G_i$.
\end{lemma}
\begin{lemma}\label{neighbor}
Each vertex $x\in V(G)-\{u_1,u_2\}$ has at least $|S|$ neighbors in $G\backslash \{u_1,u_2\}$.
\end{lemma}
\begin{proof}Suppose that $x$ has $n\le|S|-1$ neighbors in $G\backslash \{u_1,u_2\}$. Thus $X:=N_G(x)\cup\{x,u_1,u_2\}$ has $n+3\le |S|+2$ elements. As $S^+$ is a vertex-cut, then $|G|\ge |S|+4$. Hence $V(G)-X$ has a vertex $v$. But $N_G(x)\cont X-x$. So, $X-x$ separates $x$ from $v$ and, therefore, $X-x$ is a vertex-cut containing $\{u_1,u_2\}$. But $|X-x|=n+2\le|S|+1<|S^+|$, a contradiction to the minimality of $S$.
\end{proof}
The next lemma has an elementary proof, which we ommit.
\begin{lemma}\label{starbucks}
Let $D$ be a cycle of $G$ such that $e\in E(D)$ and $H$ be a subgraph of $G$ such that $H\del V(D)$ is connected. Suppose that $f$ is an edge of $G$ incident to $x\in V(H)-V(D)$ but such that $f\notin E(H)$. Then, each $(x,V(H)-x)$-path $\gamma$ of $G$ beginning with $f$ intersects $D$. In particular, $\gamma$ has an endvertex in $V(D)$ if $V(D)\cont V(H)$.
\end{lemma}
\begin{lemma}\label{menger}
Let $H$ be a $3$-connected graph, $uv$ be an edge of $H$ and $X$ be a subset of $V(H)$ avoiding $\{u,v\}$ with at least $3$ elements. Then, for some $x\in\{u,v\}$, there are $(X,\{u,v\})$-paths $\alpha$, $\beta$, and $\gamma$ of $G$ such that $V(\alpha)\cap V(\gamma)=V(\beta)\cap V(\gamma)=\emptyset$ and $V(\alpha)\cap V(\beta)=\{x\}$.
\end{lemma}
\begin{proof}
By Menger's Theorem there are $(X,\{u,v\})$-paths $\alpha_1$, $\alpha_2$ and $\alpha_3$ such that $V(\alpha_i)\cap V(\alpha_j)\cont\{u,v\}$ for $1\le i<j\le3$. We may assume that $V(\alpha_i)\cap\{u,v\}=\{v\}$ for $i=1,2,3$. This implies that $d_H(v)\ge 4$. Consider a graph $K$ with minimum degree at least $3$ obtained from $H$ by splitting $v$ into vertices $v_1$ and $v_2$. It is well known and easy to check that $K$ is $3$-connected. By Menger's Theorem, $K$ has three vertex-disjoint $(X,\{u,v_1,v_2\})$-paths. The desired paths are the corresponding paths in $H$.
\end{proof}
\begin{lemma}\label{big-cycle}
Let $u$ be a vertex in a vertex-cut $X$ of a graph $H$. If $u$ has $3$ neighbors in some connected component $K$ of $H\del X$, then $H[V(K)\cup u]$ has a cycle containing $u$ with more than $3$ vertices.
\end{lemma}
\begin{proof}
Let $x,y,z$ be distinct neighbors of $u$ in $K$. As $K$ is connected, there is an $(x,y)$-path and an $(y,z)$-path in $K$. If one of these paths has more than one edge, we are done. Otherwise, both have one edge and $u,x,y,z,u$ is the cycle we seek.
\end{proof}
\begin{lemma}\label{bigone}
If $|S|\ge 3$, $d_G(u_1),d_G(u_2)\ge4$, and, for some $\{i,j\}=\{1,2\}$ and $k\in\{1,\dots,\kappa\}$, there is a pair of edges from $u_i$ to $G_k$, then:
\begin{enumerate}
\item [(a)] $\kappa=2$ and
\item [(b)] $N_G(u_j)\cont V(G_{3-k})$.
\end{enumerate}
\end{lemma}
\begin{proof} We may assume that $k=i=1$. Note that there is a cycle $C$ in $G_1+u_1$ meeting $u_1$. Pick $C$ maximizing $|C|$. By the minimality of $S$, $G\del \{u_1,u_2\}$ is $3$-connected. Using Menger's Theorem if $|V(C)-u_1|\ge 3$ and Lemma \ref{menger} if $|V(C)-u_1|=2$, we conclude that there are $(S,V(C)-u_1)$-paths $\alpha_1$, $\alpha_2$ and $\alpha_3$ in $G\del\{u_1,u_2\}$ such that $V(\alpha_i)\cap V(\alpha_3)=\emptyset$ for $i=1,2$ and $V(\alpha_1)\cap V(\alpha_2)$ is empty if $|V(C)-u_1|\ge3$ and a singleton set contained in $V(C)$ otherwise. Note that the inner vertices of these paths are in $G_1$. For $i=1,2,3$, we let $v_i$ and $a_i$ be the endvertices of $\alpha_i$ in $S$ and $V(C)-u_1$ respectively. We write $C$ as a closed path of the form $u_1,\beta_1,a_1,\beta_{21},a_2,\beta_{23},a_3,\beta_3,u_1$ with $\|\beta_{1}\|,\|\beta_{23}\|,\|\beta_{3}\|\ge 1$ and $\|\beta_{21}\|\ge 1$ if and only if $|C|\ge 4$. In particular, we choose $\alpha_1$, $\alpha_2$ and $\alpha_3$ minimizing $\|\beta_1\|$.
First we prove (a). Suppose for a contradiction that $\kappa\ge 3$. If there is an edge from $u_2$ to $G_1$, then, for distinct $u,v\in S$, $G_1+\{u_1,u_2\}$ has a cycle containing $e$ avoiding a cycle of $G_2\cup G_3+\{u,v\}$, what is a contradiction. Thus, there is no edge from $u_2$ to $G_1$. As $d_G(u_2)\ge 4$, there is an edge linking $u_2$ and $G_2+(S-\{v_2,v_1\})$. Now, we have a cycle of the form
\begin{linenomath*}
\[u_1,u_2,G_2+(S-\{v_2,v_1\}),v_3,\alpha_3,a_3,\beta_3,u_1\]
\end{linenomath*}
containing $e$ and avoiding a cycle of the form \begin{linenomath*}\[a_1,\beta_{21},a_2,\alpha_2,v_2,G_3,v_1,\alpha_1,a_1,\]
\end{linenomath*}
a contradiction. So, $\kappa=2$ and (a) holds.
Now suppose that (b) fails. Choose a counter-example maximizing $|N_G(u_i)\cap V(G_k)|$; this choice is subject to the minimality of $\|\beta_1\|$, which is subject to the maximality of $|C|$. Note that we may still assume that $k=i=1$.
Define $X:=C\cup\alpha_1\cup\alpha_2\cup\alpha_3$. First we check:
\begin{rot}\label{bigone-1}
If $\zeta$ is an $(u_2,V(X)-u_1)$-path of $G_1+(S\cup u_2)$, then either
\begin{enumerate}
\item [(i)] $|C|=3$ and $\zeta$ has an endvertex in $V(\alpha_1)\cup V(\alpha_2)$ or
\item [(ii)] $|C|\ge 4$ and $\zeta$ has an endvertex in $V(\alpha_2)$.
\end{enumerate}
\end{rot}
\begin{rotproof} Let $u_2$ and $z$ be the endvertices of $\zeta$. First we check that $z\in A:=V(\alpha_1)\cup V(\beta_{21})\cup V(\alpha_{2})$. If not, $z$ is a vertex of $X\del A$, which is a connected graph containing $u_1$. This implies that a cycle of $A\cup G_2$ avoids a cycle of the form $u_1,u_2,\zeta,z,X\del A, u_1$, what is a contradiction. So, $z\in A$.
If $|C|=3$, $\|\beta_{21}\|=0$ and $A=V(\alpha_1)\cup V(\alpha_2)$. So, (i) holds.
Now, assume that $|C|\ge 4$. As $z\in A$, analogously, $z\in B:=V(\alpha_3)\cup V(\beta_{23})\cup V(\alpha_{2})$. Therefore, $z\in A\cap B=V(\alpha_2)$. So, (ii) holds.
\end{rotproof}
As (b) fails, $u_2$ has neighbors in $G_1\cup S$ and $G_1+(S\cup u_2)$ has an $(u_2,V(X)-u_1)$-path $\delta$, which we may assume that has an endvertex $d$ of $\alpha_2$ by \ref{bigone-1} ($\alpha_1$ and $\alpha_2$ play similar roles if $|C|=3$). Pick $\delta$ minimizing $|V(\delta)\cap S|$. Note that this choice implies that either $u_2$ has some neighbor in $G_1$ and $V(\delta)\cap S\emptyset$ or $v_2$ has no neighbor in $G_1$ and $V(\delta)\cap S$ has an unique vertex, which is a neighbor of $u_2$ in $S$.
For each $i=1,3$ and each $(u_2,V(X)-u_1)$-path $\zeta$ of $G_1+(S\cup u_2)$ with an endvertex $z$ in $\alpha_2$, we define the following cycle containing $e$:
\begin{linenomath*}
\[C_{\zeta,i}:=u_1,u_2,\zeta,z,\alpha_2,a_2,\beta_{2i},a_i,\beta_i,u_1.\]
\end{linenomath*}
Now we prove:
\begin{rot}\label{bigone-2}
$v_1$ has two neighbors in $V(G_1)$.
\end{rot}
\begin{rotproof}
Suppose the contrary. By the assumption that $\delta$ arrives in $\alpha_2$, $v_1$ avoids $\delta$. If $v_1$ has two neighbors in $V(G_2 + (S-V(\delta))$, then $C_{1,\delta}$ avoids a cycle of $G_2+(S-V(\delta)$, a contradiction. Thus $v_1$ has only one neighbor in $V(G_2 + (S-\int(\delta))$. As $v_1$ has an unique neighbor in $G_1$, it follows by Lemma \ref{neighbor} that $v_1$ has two neighbors in $G_2+S$, and, therefore, a neighbor $v\in V(\delta)\cap S$. As observed before, $v$ is the unique vertex of $S$ in $\delta$, and, as a consequence, $u_2$ has no neighbor in $G_1$ by the minimality of $|V(\delta)\cap S|$. As $d_g(u_2)\ge 4$, $u_2$ has a neighbor $w\in V(G_2)\cup (S-\{v,v_1\})$. Since $\delta$ arrives at $X$ in $\alpha_2$, it follows that $v_3\notin V(\delta)$. Thus the cycle $v_1,v,\delta,d,\alpha_2,a_2,\beta_{21},a_1,\alpha_1,v_1$ avoids a cycle of the form $u_1,u_2,w,G_2,v_3,\alpha_3,a_3,\beta_3,u_1$, a contradiction.
\end{rotproof}
By \ref{bigone-2} $v_1$ has a neighbor $v'_1\in V(G_1)$ such that $v_1v'_1\notin E(\alpha_1)$. Let $\sigma$ be a $(v_1,V(X)\cup\int(\delta))$-path of $G_1+v_i$ beginning with $v_1v'_1$ and let $s$ be the other endvertex of $\sigma$.
Let $H_0$ be the subgraph of $G$ obtained from $G_2\cup X$ by adding the dges that connects $X$ to $G_2$. Note that $H_0\del V(C_{\delta,1})$ is connected. By Lemma \ref{starbucks}, $s\in V(C_{\delta,1})$. If $s\in \int(\beta_1)$, then $\sigma$, $\alpha_2$ and $\alpha_3$ contradict the choice of $\alpha_1$, $\alpha_2$ and $\alpha_3$ minimizing $\|\beta_1\|$. So, $s\notin \int(\beta_1)$ and $\sigma$ avoids $\int(\beta_1)$. As $(H_0\del \int(\beta_1))\del V(C_{\delta,3})$ is connected, it follows from Lemma \ref{starbucks} that $s\in V(C_{\delta,3})$. So,
\begin{linenomath*}
\[s\in V(C_{\delta,1})\cap V(C_{\delta,3}) \cap V(G_1)\cont V(a_2,\alpha_2,d)\cup \int(\delta).\]
\end{linenomath*}
Next we define a path $\tau$ as $\tau:=\sigma$ if $s\in V(\alpha_2)$ and $\tau:=\sigma,s,\delta,d$ otherwise. In the later case by the minimality of $|V(\delta)\cap S|$, once $\delta$ meets $s\in V(G_1)$, $\delta$ does not return to $V(G_2)\cup V(S)$. So, $\tau$ is an $(v_1,V(\alpha_2))$-path of $G_1+v_1$ in all cases. Let $t$ and $v_1$ be the endvertices of $\tau$. We have that $B:=v_1,\tau,t,\alpha_2,a_2,\beta_{21}a_1,\alpha_1,v_1$ is a cycle of $G_1+v_1$.
As $|S|\ge 3$, then by the minimality of $|S|$, $G\del\{u_1,u_2\}$ is $3$-connected. As $u_2$ has degree at least $3$ in $G\del u_1$, $G\del u_1$ is $3$-connected. So $G\del \{u_1,v_1\}$ has two $(u_2,V(X)\cap V(G_1))$-paths $\zeta_1$ and $\zeta_2$ meeting only in $u_2$. Let $H_1:=H_0+S$. Note that $H_1\del V(B)$ is a connected graph containing $u_1$. If $\zeta_k$ meets $H_1\del V(B)$ before meeting $B$, then a cycle of the form $u_1,u_2,\zeta_k,H_1\del V(B),u_1$ avoids $B$, a contradiction. So, $\zeta_k$ does not meet $H_1\del V(B)$ before meeting $B$ and, therefore, $\zeta_1$ and $\zeta_2$ are paths of $G_1+\{u_2,v_1\}$. So one of these paths is in $G_1+u_2$ and is a possibility for the choice of $\delta$. By the minimality of $|V(\delta)\cap S$, $\delta$ is a path of $G_1+u_2$.
Next we check that $|C|\ge4$. Suppose the contrary. By the maximality of $|C|$ and by Lemma \ref{big-cycle}, $u_1$ has at most two neighbors in $G_1$ and, by the maximality of $|N_G(u_i)\cap V(G_k)|$, for each $i,k\in\{1,2\}$, $u_i$ has at most two neighbors in $G_k$. So, since $u_1$ and $u_2$ have degree at least four, it follows that there are $w_1,w_2\in V(G_2)\cup S$ such that $u_1w_1,u_2w_2\in E(G)$. Let $v\in S-\{w_1,w_2\}$. If $v$ has two different neighbors in $G_1$, then a cycle of $G_1+v$ avoids a cycle of the form $u_1,u_2,w_2,G_2,w_1,u_1$, a contradiction. So, by Lemma \ref{neighbor}, $v$ has two different neighbors in $G_2+S$, which has a cycle. But this cycle avoids $C_{\delta,1}$ since $\delta$ is a path of $G_1+u_2$, a contradiction. Hence, $|C|\ge 4$. This implies that $a_1\neq a_2$.
If $\zeta_k$ intersects $\int(\tau)$, then cycles of the forms
\[u_1,u_2,\zeta_k,\tau,v_1,\alpha_1,a_1,\beta_1,u_1\qquad\text{ and}\]
\[v_3,G_2,v_2,\alpha_2,a_2,\beta_{23},a_3,\alpha_3,v_3\] avoid each other, a contradiction. So, $\zeta_k$ does not intersect $\int(\tau)$. By \ref{bigone-1}, the endvertex $z_k$ of $\zeta_k$ in $X$ is in $\alpha_2$. Say that $a_2$, $z_1$, $z_2$ and $v_2$ appear in this order in $\alpha_2$. As $\zeta_k$ has an endvertex in $B$, then $t\in V(z_2,\alpha_2,v_2)$. This implies that $\tau$ avoids $C_{\zeta_1,1}$. Therefore, $C_{\zeta_1,1}$ avoids a cycle of the form $v_1,\tau,t,\alpha_2,v_2,G_2,v_1$, a contradiction.
\end{proof}
\begin{lemma}\label{Sle2}
If $d_G(u_1),d_G(u_2)\ge 4$, then $|S|\le 2$.
\end{lemma}
\begin{proof}\setcounter{rot}{0}
Suppose the contrary.
\begin{rot}\label{Sle2-1}
There is no $i\in\{1,2\}$ and $k\in\{1,\dots,\kappa\}$ such that there are two edges from $u_i$ to $G_k$.
\end{rot}
\begin{rotproof}
Suppose for a contradiction that there are two edges linking $u_i$ to $G_k$. Say $i=k=1$. By Lemma \ref{bigone}, $N_G(u_2)\cont V(G_2)$. But this implies that the hypothesis of Lemma \ref{bigone} also holds for $i=k=2$, then $N_G(u_1)\cont V(G_1)$. For $l=1,2$, as $d_G(u_l)\ge 4$, then, by Lemma \ref{big-cycle}, there is a cycle $C_l$ of $G_l+u_l$ containing $u_l$ with more than $3$-vertices. By the minimality of $|S|$, $G\del \{u_1,u_2\}$ is $3$-connected. So, there are three vertex-disjoint $(C_1-u_1,C_2-u_2)$-paths in $G\del u_1,u_2$. Together with the path $u_1,u_2$, we have four vertex-disjoint $(C_1,C_2)$-paths. Now it is easy to check that there are vertex-disjoint cycles covering these paths. And, therefore, we have two disjoint cycles with one of them containing $e$, a contradiction.
\end{rotproof}
\begin{rot}\label{Sle2-2}
$u_1$ and $u_2$ have no common neighbor in $S$.
\end{rot}
\begin{rotproof}
If $w$ contradicts \ref{Sle2-2}, then $u_1,u_2,w,u_1$ is a cycle containing $e$ that, for some $x,y\in S-w$, avoids a cycle of the form $x,G_1,y,G_2,x$.
\end{rotproof}
\begin{rot}\label{Sle2-3}
$\kappa=2$.
\end{rot}
\begin{rotproof}
Suppose the contrary. First we suppose that $u_1v\in E(G)$ for some $v\in S$. If $N_G(u_2)\cont S\cup u_1$, then, as $d_G(u_2)\ge 4$, by \ref{Sle2-2}, it follows that there is a $3$-subset $\{x_1,x_2,x_3\}$ of $N_G(u_2)-\{u_1,v\}$ and cycles of the form $u_1,u_2,x_1,G_1,v,u_1$ and $x_2,G_2,x_3,G_3,x_2$ avoiding each other, a contradiction. So, we may assume that there is an edge from $u_2$ to $G_1$. But this implies that, for $x,y\in S-v$, cycles of the form $u_1,u_2,G_1,v,u_1$ and $x,G_2,y,G_3,x$ avoid each other, a contradiction. Hence, $u_1$ has no neighbor in $S$ and, analogously, neither has $u_2$.
If both $u_1$ and $u_2$ have neighbors in a common component of $G\del S^+$, say $G_1$, we have, for $x,y\in S$, cycles of the forms $u_1,u_2,G_1,u_1$ and $x,G_2,y,G_3,x$ avoiding each other, a contradiction. Therefore, $u_1$ and $u_2$ have no neighbors in $S$ nor in a same connected component of $G\del S^+$. By \ref{Sle2-1}, $\kappa\ge 4$ and we may assume that $u_i$ has a neighbor in $G_i$ for $i=1,2$. Now, for distinct $x,y,z\in S$, cycles of the form, $u_1,u_2,G_2,z,G_1,u_1$ and $x,G_3,y,G_4,x$ avoid each other, a contradiction.
\end{rotproof}
Let $k\in\{1,2\}$. By \ref{Sle2-1} and \ref{Sle2-3}, $u_k$ has a neighbor $v_k\in S$. Let $x\in S-\{v_1,v_2\}$. If $x$ has two neighbors in $G_1$, then $G_1+x$ has a cycle avoiding a cycle of the form $u_1,u_2,v_2,G_2,v_1,u_1$. Thus, $x$ has only one neighbor in $G_1$ and, analogously, only one neighbor in $G_2$. By Lemma \ref{neighbor}, $x$ has a neighbor $y\in S$. If $y\notin \{v_1,v_2\}$, a cycle of the form $u_1,u_2,v_2,G_1,v_1,u_1$ avoids a cycle of the form $x,y,G_2,x$, a contradiction. So, we may assume that $y=v_1$. As $d_G(u_1)\ge 4$, there is $w\in N_G(u_1)-\{u_2,v_1\}$ and we also may assume that $w\notin V(G_2)$ since $u_1$ has at most one neighbor in $G_2$. Now a cycle of the form $x,y,G_2,x$ avoids a cycle of the form $u_1,u_2,v_2,G_1,w,u_1$, a contradiction.
\end{proof}
\begin{lemma}\label{fan side}
Suppose that $|S|=2$ and, for some connected component $J$ of $G\del S^+$, there are edges from both $u_1$ and $u_2$ to $J$. If $K$ is a connected component other than $J$, then $G[K\cup S]$ is a path with endvertices in $S$, moreover each vertex of $K$ has a neighbor in $\{u_1,u_2\}$.
\end{lemma}
\begin{proof}
Note that $J+\{u_1,u_2\}$ has a cycle containing $e$. Therefore, $G[K\cup S]$ is a tree and so is $K$. This implies that each element of $S$ has an unique neighbor in $K$. Define $S:=\{x,y\}$ and let $v_x$ and $v_y$ be the neighbors of $x$ and $y$ respectively in $K$. Let us check that $v_x$ and $v_y$ are the unique leaves of $K$. Suppose for a contradiction that $l$ is a leave of $K$ different from $v_x$ and $v_y$. This implies that $u_1,u_2,l$ is a cycle of $G$ avoiding a cycle of the form $x,K-l,y,J,x$, a contradiction. So, $v_x$ and $v_y$ are the unique leaves of $K$ and $K$ is an $(v_x,v_y)$-path. Hence, $G[K\cup S]$ is a path with endvertices in $S$. Since each vertex of $G$ has degree at least $3$ and $v_x$ and $v_y$ are the unique neighbors of $x$ and $y$ in $K$ respectively, then each vertex of $K$ has a neighbor in $\{u_1,u_2\}$. This proves the lemma.
\end{proof}
The next Lemma gives a full characterization of the Dirac graphs when $|S|=1$ and has an elementary proof.
\begin{lemma}\label{all type b}
Let $H$ be a $3$-connected graph and suppose that $e:=uv\in E(H)$ is an edge of $H$ and $w\in V(H)$ is such that $\{u,v,w\}$ is a $3$-vertex cut of $H$. Then the following assertions are equivalent:
\begin{enumerate}
\item [(a)] $H$ is $e$-Dirac.
\item [(b)] Each connected component $K$ of $H\del \{u,v,w\}$ is a tree with a special vertex $x_K$ such that $N_H(w)\cap V(K)=\{x_K\}$.
\end{enumerate}
\end{lemma}
\section{One-, two- and no-sided separations}\label{sec-one-two}
Let $X$ be a $4$-vertex-cut of $G$ containing $\{u_1,u_2\}$. We say that $X$ is \textbf{two-sided} if, for each component $K$ of $G\del X$, there is $\{i,j\}=\{1,2\}$ such that there is an edge from $u_i$ to $K$ but no edge from $u_j$ to $K$. We say that $X$ is \textbf{one-sided} if, for each component $K$ of $G\del X$, there are edges from both $u_1$ and $u_2$ to $K$. We say that $X$ is \textbf{no-sided} if $X$ is neither one-sided nor two-sided.
\begin{lemma}\label{ugly}
Suppose that $|S|=2$, $d_G(u_1),d_G(u_2)\ge4$ and all $4$-vertex-cuts of $G$ containing $\{u_1,u_2\}$ are no-sided. Then $G$ is obtained from a wheel by possibly doubling some spokes and splitting the hub by $e$.
\end{lemma}
\begin{proof}
\stepcounter{rotcount}Suppose that the lemma fails. First we check that $G\del X$ has exactly two connected components for each $4$-vertex cut $X$ containing $\{u_1,u_2\}$. Suppose that $X$ contradicts this assertion. As $X$ is no-sided, there is a connected component $H_1$ of $G\del X$ with neighbors of both $u_1$ and $u_2$ and two distinct connected components $H_2,H_3\neq H_1$ of $G\del X$. Then, for $X=\{u_1,u_2,x,y\}$, a cycle in the from $u_1,u_2,H_1,u_1$ avoids a cycle of the form $x,H_2,y,H_3,x$, a contradiction. Thus $G\del X$ has exactly two connected components.
Now, for each $2$-subset $A$ of $V(G)-\{u_1,u_2\}$ such that $A\cup\{u_1,u_2\}$ is a $4$-vertex cut of $G$, $A\cup\{u_1,u_2\}$ is no-sided by hypothesis. The connected component of $G\del A\cup\{u_1,u_2\}$ containg neighbors of both $u_1$ and $u_2$
will be called the {\it $A$-right} component and denoted by $R(A)$, while the component containing neighbors of only one element of $\{u_1,u_2\}$ will be called the {\it $A$-left} component and denoted by $L(A)$. Choose $S$ maximizing the number of vertices in $L(S)$. Say that $u_1$ has a neighbor in $L(S)$ and $u_2$ does not.
Write $S=\{v_1,v_2\}$. By Lemma \ref{fan side}, $G[L(S)\cup\{u_1,v_1,v_2\}]$ is a fan with hub $u_1$ and endvertices $v_1$ and $v_2$. If $|R(S)|=1$, then it is clear that the lemma holds. So, $|R(S)|\ge 2$.
\begin{rot}\label{ugly-1}
For $i=1,2$, $v_i$ has at least two neighbors in $R(S)$. Moreover, $v_1u_2,v_2u_2,v_1v_2\notin E(G)$ and $u_2$ has at least $3$ neighbors in $R(S)$.
\end{rot}
\begin{rotproof}
For the first part, say that $v_1$ has an unique neighbor $x\in R(S)$. Then, for $A:=\{x,v_2\}$, $A\cup\{u_1,u_2\}$ separates $L(S)+ v_1$ from $R(S)\del x$, which is non-empty because $|R(S)|\ge 2$. If $u_2v_1\notin E(G)$, then there are no edges from $u_2$ to $L(S)\cup v_1$ and $L(A)=L(S)+v_1$, contradicting the maximality of $L(S)$. Thus $u_2v_1\in E(G)$. So, both $u_1$ and $u_2$ have neighbors in $L(S)+v_1$. This implies that $R(A)=L(S)+v_1$ and $L(A)=R(S)\del x$. Now, for an unique $k\in\{1,2\}$, $u_k$ has neighbors in $L(A)$. By Lemma \ref{fan side}, $G[L(S)\cup S]$ is a $(v_1,v_2)$-path. By Lemma \ref{fan side} (for $S=A$), $G[L(A)\cup A]=G[R(S)\cup v_2]$ is a $(v_2,x)$-path. Recall that $x$ is the unique neighbor of $v_1$ in $R(S)$. Hence $G\del u_1,u_2$ is a cycle and the Lemma holds in this case, a contradiction. Therefore, $v_i$ has at least two neighbors in $R(S)$ for $i=1,2$.
Now, suppose for a contradiction that $v_2u_2\in E(G)$. Since there are two edges from $v_1$ to $R(S)$, hence $R(S)+v_1$ has a cycle, which must avoid a cycle of the form $u_1,u_2,v_2,L(S),u_1$, a contradiction. So, $v_2u_2\notin E(G)$. Similarly $v_1u_2\notin E(G)$. As $d_G(u_2)\ge 4$, $u_2$ has at least $3$ neighbors in $R(S)$. We already saw that $v_1v_2\notin E(G)$, since $G[L(S)\cup\{v_1,v_2\}]$ is a $(v_1,v_2)$-path. So, \ref{ugly-1} holds.
\end{rotproof}
\begin{rot}\label{ugly-0}
$G\del e$ is $3$-connected.
\end{rot}
\begin{rotproof}
Suppose for a contradiction that $P$ is a $2$-vertex cut of $G\del e$. As $G$ is $3$-connected, $P$ does not meet $\{u_1,u_2\}$ and $u_1$ and $u_2$ are in different connected components $K_1$ and $K_2$ of $G\del P\cup e$ respectively. If $|V(K_i)|=1$, then $d_G(u_i)\le 3$, contradicting our hypothesis, so $|V(K_i)|\ge 2$. This implies that $P\cup\{u_1,u_2\}$ is a $4$-vertex cut of $G$. But, since $G\del P\cup e$ has $u_1$ and $u_2$ in different connected components, it follows that $P\cup\{u_1,u_2\}$ is two-sided, contradicting our hypothesis.
\end{rotproof}
\begin{rot}\label{ugly-2}
$R(S)$ is a tree.
\end{rot}
\begin{rotproof}
By \ref{ugly-0}, there are three pairwise disjoint $(\{v_1,v_2,u_1\},V(C))$-paths $\alpha_1$, $\alpha_2$ and $\alpha_3$ in $G\del e$. As $\{u_1,v_1,v_2\}$ separates $L(S)$ from $C$ in $G\del e$, none of these paths meet $L(S)$. Say that $\{v_1,a_1\}$, $\{v_2,a_2\}$ and $\{u_1,a_3\}$ are the respective pairs of endvertices of $\alpha_1$, $\alpha_2$ and $\alpha_3$ in $C$. Consider also $(u_2,V(C))$-paths $\beta_1$ and $\beta_2$ of $G\del u_1$ intersecting only in $u_2$ and let $b_1$ and $b_2$ be the respective endvertices of $\beta_1$ and $\beta_2$ in $C$.
If for some $j\in\{1,2\}$ and $i\in \{1,2,3\}$, $\beta_j$ meets $\alpha_i$ out of $C$, then $G$ has a \\ $(u_2,\{v_1,v_2,u_1\})$-path $\gamma$ avoiding $C$ and, as a consequence, $C$ avoids a cycle of the form $u_1,u_2,\gamma,L(S)+\{v_1,v_2,u_1\}$ containing $e$, a contradiction. So, $\alpha_i$ and $\beta_j$ do not meet out of $C$.
Let $\delta$ and $\varepsilon$ be the the $(a_1,a_2)$-paths of $C$ meeting and avoiding $a_3$ respectively. If $\beta_j$ has an endvertex in $\int(\delta)$, then, the cycle $u_1,u_2,\beta_j,b_j,\delta,a_3,\alpha_3,u_1$ avoids a cycle of the form $v_1,\alpha_1,a_1,\varepsilon,a_2,v_2,L(S),v_1$, a contradiction. So, $b_1$ and $b_2$ are in $\varepsilon$ and we may assume that $b_1$ is closer to $a_1$ than $a_2$ in $\varepsilon$. See an illustration in Figure \ref{fig-ugly}.
\begin{figure}[h]\label{fig-ugly}
\centering
\begin{tikzpicture}
\tikzstyle{node_style} =[shape = circle,fill = black,minimum size = 2pt,inner sep=1pt]
\node[node_style] (u1) at (-3,1.5) {};\node()at(-3.0,1.3) {$u_1$};
\node[node_style] (v1) at (-2.5,3.0) {};\node()at(-2.5,3.2) {$v_1$};
\node[node_style] (v2) at (-2.5,0.0) {};\node()at(-2.5,-0.2) {$v_2$};
\node[node_style] (u2) at (-2,1.5){};\node()at(-2,1.3) {$u_2$};
\node[node_style] (a1) at (2,3) {};\node()at(2,3.2) {$a_1$};
\node[node_style] (a2) at (2,0) {};\node()at(2,-0.2){$a_2$};
\node[node_style] (a3) at (1.3,1.5) {};\node() at (1.55,1.5) {$a_3$};
\node[node_style] (b1) at (2.7,2.1) {};\node() at (3,2.1) {$b_1$};
\node[node_style] (b2) at (2.7,0.9) {};\node() at (3,0.9) {$b_2$};
\draw plot [smooth, tension=1] coordinates {(a1) (b1) (b2) (a2)};
\draw plot [smooth, tension=2] coordinates {(a1) (a3) (a2)};
\draw (v1) to (a1); \node () at (-0.25,3.15) {$\alpha_1$};
\draw plot [smooth] coordinates {(u1)(-1.5,2)(0,1.5)(a3)}; \node () at (0.5,1.63) {$\alpha_3$};
\draw (v2) to (a2); \node () at (-0.25,-0.17) {$\alpha_2$};
\draw plot [smooth] coordinates{(u2)(0,2.1)(b1)};\node () at (0,2.29) {$\beta_1$};
\draw plot [smooth] coordinates{(u2)(0,0.9)(b2)};\node () at (0,0.68) {$\beta_2$};
\draw (u1)--(u2);
\node () at (2.95,1.5) {$\varepsilon$};
\node () at (2.6,2.6) {$\varepsilon$};
\node () at (2.6,0.4) {$\varepsilon$};
\node () at (1.35,2.6) {$\delta$};
\node () at (1.35,0.4) {$\delta$};
\node (los) at (-4.5,1.5) {};
\draw (v1)--(los);\draw (v2)--(los);\draw (u1)--(los);
\draw (los) [fill=white] circle (0.6cm);
\node () at (los) {$L(S)$};
\end{tikzpicture}
\caption{A illustration of the proof of \ref{ugly-2}}
\end{figure}
Consider the following subgraph of $G$
\begin{linenomath*}\[H:=
(L(S)+\{v_1,v_2,u_1\})\cup C\cup\alpha_1\cup\alpha_2\cup\alpha_3\cup\beta_1\cup\beta_2.\]
\end{linenomath*}
By \ref{ugly-1}, $v_1$ has two different neighbors in $R(S)$ and there is a $(v_1,V(H))$-path $\varphi$ of $R(S)+v_1$ beginning with an edge out of $H$. Let $x$ be the other endvertex of $\varphi$. Next we consider, for $i=1,2$, the following circuit:
\begin{linenomath*}\[D_i:=u_1,u_2,\beta_i,b_i,\varepsilon,a_i,\delta,a_3,\alpha_3,u_1.\]
\end{linenomath*}
Note that $V(D_1)\cap V(D_2)=V(\alpha_3)\cup u_2$ and that $H\del V(D_i)$ is connected for $i=1,2$. By Lemma \ref{starbucks}, $x\in V(\alpha_3)-u_1$. Now, $v_1,\alpha_1,a_1,\delta,a_3,\alpha_3,x,\varphi,v_1$ avoids a cycle of the form $u_1,u_2,\beta_2,\varepsilon,a_2,\alpha_2,v_2,L(S),u_1$, a contradiction.
\end{rotproof}
\begin{rot}\label{ugly-3}
$u_1$ and $u_2$ have no leaf of $R(S)$ as common neighbor.
\end{rot}
\begin{rotproof}
Suppose that $l$ is a leaf contradicting \ref{ugly-3}. By \ref{ugly-1}, $v_1$ and $v_2$ have neighbors in $R(S)-l$. Then, a cycle of the form $v_1,R(S)-l,v_2,L(S),v_1$ avoids the cycle $u_1,u_2,l,u_1$, a contradiction.
\end{rotproof}
\begin{rot}\label{ugly-4}
Let $l$ be a leaf of $R(S)$. Then, $u_2l\notin E(G)$, $u_1l\in E(G)$ and there is an unique index $i\in\{1,2\}$ such that $lv_i\in E(G)$.
\end{rot}
\begin{rotproof}
First let us prove that there is no leaf $l$ of $R(S)$ such that $lv_1,lv_2\in E(G)$. Suppose that $l$ is such a leaf. If $u_1$ has a neighbor in $V(R(S))-l$, then by \ref{ugly-3}, so does $u_2$ and, therefore, cycles of the form $u_1,u_2,R(S)\del l,u_1$ and $v_1,l,v_2,L(S),v_1$ avoid each other, a contradiction. Hence, $l$ is the unique neighbor of $u_1$ in $R(S)$. This implies that $l$ is the unique leaf of $R(S)$ incident to both $v_1$ and $v_2$. Let $l'$ be another leaf of $R(S)$. We may assume that $l'v_1,l'u_2\in E(G)$.
As $l$ is the unique leaf incident to both $v_1$ and $v_2$, it follows from \ref{ugly-1} that $v_2$ has a neighbor in $R(S)\del l,l'$. So, $(R(S)\del l')+v_2$ has a cycle, which avoids a cycle of the form $u_1,u_2,l',v_1,L(S),u_1$, a contradiction. So, $v_1$ and $v_2$ have no leaf as common neighbor in $R(S)$.
By \ref{ugly-1}, this implies that each leaf of $R(S)$ is adjacent to an unique element of $\{u_1,u_2\}$ and an unique element of $\{v_1,v_2\}$. Let us prove that no leaf is adjacent to $u_2$. Suppose that some leaf $l$ is adjacent to $u_2$ and, say, $v_1$. By \ref{ugly-1}, $v_2$ has at least two neighbors in $R(S)\del l$ and cycles of the form $v_2,R(S)-l,v_2$ and $u_1,u_2,l,v_1,L(S),u_1$ avoid each other, a contradiction. This implies \ref{ugly-4}.
\end{rotproof}
\begin{rot}\label{ugly-5}
$R(S)$ has exactly two leaves.
\end{rot}
\begin{rotproof}
Suppose the contrary. Then, $R(S)$ has a vertex $w$ with $d_{R(S)}(w)\ge 3$. Moreover, $R(S)\del w$ has different connected components $K_1$, $K_2$ and $K_3$, each one containing a leaf of $R(S)$. Let $t\in\{1,2,3\}$. By \ref{ugly-4}, $K_t$ has a neighbor of $u_1$ and a neighbor of some $w_t\in\{v_1,v_2\}$. As $u_2$ has a neighbor in $R(S)-w$, we may assume that $u_2$ has a neighbor in $K_1$ and, therefore, $K_1+\{u_1,u_2\}$ has a cycle $C$ containing $e$. Now, there is a cycle of $(L(S)\cup K_2\cup K_3)+\{w,v_1,v_2\}$ avoiding $C$, a contradiction.
\end{rotproof}
We may write $R(S)$ as a path $w_1,\dots, w_n$. By \ref{ugly-1} and \ref{ugly-4} there are $1<a<b<n$ such that $w_a,w_b\in N_G(u_2)$. Also, by \ref{ugly-1} and \ref{ugly-4}, there are edges from $H:=L(S)+\{v_1,v_2\}$ to $w_1$, $w_n$ and a vertex $w_d$ with $1<d<n$. If $d<b$, then a cycle of the form $w_1,\dots,w_d,H,w_1$ avoids $u_1,u_2,w_b,\dots,w_n,u_1$, a contradiction. Thus $d\ge b$. But, symmetrically, we have $d\le a$, a contradiction.
\end{proof}
\section{Rope Bridges}\label{sec-rope bridges}
\renewcommand{\S}{\mathcal{S}}
In this section we describe structures that may appear when $\min\{d_G(u_1),d_G(u_2)\}=3$ or when $|S|=2$ and we may pick $S$ such that $S^+$ is two-sided. Those are the last characterizations that we need to prove Theorem \ref{main}; the theorem is proved in the end of the section as a direct consequence of Theorem \ref{main-cor}.
A graph $R$ with distinct vertices $u,x,y$ is an \textbf{$(u,x,y)$-rope bridge} with \textbf{ropes} $\rho_x$ and $\rho_y$ and \textbf{family of steps} $\S$, if the following assertions hold.
\begin{enumerate}
\item [(RB0)] $d_R(v)\ge 3$ for all $v\in V(R)-\{u,x,y\}$.
\item [(RB1)] $\rho_x$ and $\rho_y$ are paths from $u$ to $x$ and $y$ respectively with $V(\rho_y)\cap V(\rho_x)=\{u\}$.
\item [(RB2)] $\S$ is a family of internally disjoint $(V(\rho_x)-u,V(\rho_y)-u)$-paths (whose members we call \textbf{steps}). We denote by $x_\alpha$ and $y_\alpha$ the extremities of the step $\alpha$ in $\rho_x$ and $\rho_y$ respectively. We say that a step $\alpha$ \textbf{crosses} a step $\beta$ if for some $\{s,t\}=\{x,y\}$, $s_\alpha$ and $s_\beta$ are distinct and appear in this order in $\rho_s$, while $t_\beta$ and $t_\alpha$ are distinct and appear in this order in $\rho_t$.
\item [(RB3)] Each step crosses at most one other step.
\item [(RB4)] Each vertex in $V(R)-V(\rho_y\cup\rho_x)$ is in some step.
\item [(RB5)] Each edge not in a member of $\S\cup\{\rho_x,\rho_y\}$ is incident to $u$.
\item [(RB6)] Let $z\in\{x,y\}$ and $v\in \rho_z$. Suppose that two steps have extremities in $\int(u,\rho_z,v)$. Then, $uv\notin E(R)$ and each step with extremity in $v$ has no inner vertices.
\end{enumerate}
\begin{lemma}\label{smaller rope-bridge}
Suppose that $R$ is an $(u,x,y)$-rope bridge with ropes $\rho_x$ and $\rho_y$ and that $w\in \int(\rho_y)$. Let $w_1$ be the vertex that follows $w$ in $\rho_y$, $\rho:=w_1,\rho_y,y$ and $\sigma_1,\dots,\sigma_k$ be the steps meeting $\rho$. Then, the graph $R'$ obtained from \\ $R\del (\rho\cup\int(\sigma_1),\dots,\int(\sigma_k))$ by suppressing the degree-two vertices in $\int(\rho_x)$ is an $(u,x,w)$-rope bridge.
\end{lemma}
\begin{proof}
Define $W=V(\rho\cup\int(\sigma_1)\cup\cdots\cup\int(\sigma_k))$. For some $F\cont E(\rho_x)$, up to isomorphisms, we have $R'=R\del W/F$. Consider $\rho'_x:=\rho_x/F$ and $\rho_w:=\rho_y\del V(\rho)$ as ropes for $R'$. Also consider as family of steps $\S'=\S-\{\sigma_1,\dots,\sigma_k\}$. Now, its is easy to verify that $R'$ with such ropes and steps, inherits each one of the properties (RB0)-(RB6) from $R$.
\end{proof}
\begin{lemma}\label{rope-bridge criterion}
Let $R$ be a connected graph with vertices $u,x,y$ and paths $\rho_x$ and $\rho_y$ from $u$ to $x$ and $y$ respectively, satisfying $V(\rho_y)\cap V(\rho_x)=\{u\}$. Suppose that $d_R(v)\ge 3$ for all $v\in V(R)-\{u,x,y\}$. Then, the following assertions are equivalent:
\begin{enumerate}
\item [(a)] $R$ is an $(u,x,y)$-rope-bridge with ropes $\rho_x$ and $\rho_y$.
\item [(b)] If $z\in \{x,y\}$ and $C$ is a cycle of $R\del \{u,z\}$, then $R$ has no $(u,z)$-path disjoint from $C$.
\end{enumerate}
\end{lemma}
\begin{proof}
Conditions (RB0) and (RB1) are given in the hypothesis. So, we have to prove that (b) is equivalent to (RB2)-(RB6) for some family of steps $\S$, which we will define ahead. Suppose that (b) holds.
First we prove that each $v\in V(R)-(V(\rho_x)\cup(V\rho_y))$ is in a $(V(\rho_x),V(\rho_y))$-path $\sigma_v$ avoiding $u$. If $v$ is in a cycle of $R\del u$, then, by (b), this cycle must meet $\rho_x$ and $\rho_y$ and this implies the existence of $\sigma_v$. So, we may assume that $v$ is in no cycle of $R\del u$. As $d_{R\del u}(v)\ge 2$, this implies that $R\del u,v$ has different connected components $K_1$ and $K_2$ each one with an unique neighbor of $v$. If $\rho_x$ and $\rho_y$ are each one in a different component of $\{K_1,K_2\}$, then the existence of $\sigma_v$ is straightforward. So, we may assume that $K_1$ avoids $\rho_x$ and $\rho_y$. Let $w_1$ be the unique neighbor of $v$ in $K_1$. As $d_R(w_1)\ge 3$, $K_1$ has more than one vertex. But, for all $w\in V(K_1)-w_1$, $d_{R\del u,v}(w)\ge 2$. This implies that $K_1$ has a cycle, which must avoid $\rho_x$ and $\rho_y$, contradicting (b). This proves the existence of $\sigma_v$.
Now, we define to be the elements of $\S$ the paths $\sigma_v$ with $v\in V(R)-V(\rho_x\cup\rho_y)$ and the paths of length one of the form $x',y'$ with $x'\in \rho_x-u$ and $y'\in \rho_y-u$. We will call {\it steps} the members of $\S$. For each step $\alpha$ we denote by $x_\alpha$ and $y_\alpha$ the endvertices of $\alpha$ in $\rho_x$ and $\rho_y$ respectively.
To establish (RB2) we shall prove that the steps are internally disjoint. Indeed, suppose for a contradiction that a step $\alpha$ intersects a step $\beta$ in an inner vertex $v$. As $\alpha\neq\beta$, we may assume that $x_{\alpha},\alpha,v\neq x_{\beta},\beta,v$. This implies that, $x_{\alpha},\alpha,v,\beta,x_{\beta},\rho_x,x_\alpha$ contains a cycle avoiding $\rho_y$, a contradiction. Thus, the steps are internally disjoint and (RB2) holds.
Suppose that (RB3) does not hold. So, there is a step $\alpha$ crossing different steps $\beta$ and $\gamma$. We may assume that $x_\alpha$, $x_\beta$ and $x_\gamma$ appear in this order in $\rho_x$. Then, the cycle $x_\beta,\rho_x,x_\gamma,\gamma,y_\gamma,\rho_y,y_\beta,\beta,x_\beta$ avoids the path $u,\rho_x,x_\alpha,\alpha,y_\alpha,\rho_y,y$, a contradiction to (b). So, (RB3) holds.
By construction, each vertex of $V(R)-V(\rho_x\cup\rho_y)$ is in a step and we have (RB4).
To prove (RB5) suppose that $f=vw$ is an edge of $R\del u$ not in $\rho_x\cup\rho_y$. If $\{v,w\}$ meet both $\rho_x$ and $\rho_y$ then $f$ is in a step. If $\{v,w\}\subseteq V(\rho_x)$, then there is a cycle in $\rho_x+f$ avoiding $\rho_y$. So, $\{v,w\}\nsubseteq}\newcommand{\cont}{\subseteq V(\rho_x)$ and, analogously, $\{v,w\}\nsubseteq}\newcommand{\cont}{\subseteq V(\rho_y)$. We may assume that $v\notin \rho_x\cup \rho_y$. By (RB4), $v$ is in the interior of a step $\sigma$. If $w$ is in $\sigma$ then either $f\in E(\sigma)$ or $\sigma+f$ has a cycle avoiding one of $\rho_x$ or $\rho _y$. If $w$ is in $\rho_x$ or $\rho_y$, say the former, then $w,v,\sigma,x_\sigma,\rho_x,w$ is a cycle avoiding $\rho_y$. Thus, $w$ is in the interior of a step $\alpha\neq \sigma$. Now $w,v,\sigma,x_\sigma,\rho_x,x_\alpha,\alpha,w$ is a cycle avoiding $\rho_y$, a contradiction. So, (RB5) holds.
Now, we prove (RB6). Let $z\in\{x,y\}$ and $v\in \rho_z$ and suppose that two steps $\alpha$ and $\beta$ have extremities in $\int(u,\rho_z,v)$. Say $z=x$. If $uv\in E(R)$, then the path $u,v,\rho_x,x$ avoids the cycle $C:=x_\alpha,\alpha,y_\alpha,\rho_y,y_\beta,\beta,x_\beta,\rho_x,x_\alpha$, a contradiction. So, $uv\notin E(R)$. If a step $\sigma$ with extremity in $v$ has a inner vertex $w$, then, as $d_R(w)\ge 3$, there is an edge incident to $w$ not in $\sigma$, and by (RB5), this edge is $uw$. Now $u,w,\sigma,v,\rho_x,x$ avoids $C$, a contradiction. So, (RB6) holds and (b) implies (a).
Suppose that $R$ is a graph for which (a) holds but (b) does not hold. Choose $R$ with $|V(R)|$ as small as possible. Consider $z\in \{x,y\}$ such that there is a cycle $C$ of $R\del \{u,z\}$ and a $(u,z)$-path $\gamma$ disjoint from $C$ as short as possible. Say that $z=x$. Let $v$ be the vertex of $\gamma$ such that $v,\gamma,x$ is contained in $\rho_x$ and $v,\gamma,x$ is as long as possible. Let $w$ be the vertex preceding $v$ in $\gamma$. Note that $C$ meets at least two steps, whose endvertices precedes $v$ in $\rho_x$. By (RB6), $w\neq u$ and all steps arriving in $v$ have no internal vertices. If the edge $wv$ is in $\rho_x$, then $w,v,\rho_x,x$ violates the maximality of $v,\rho_x,x$. So, as $w\neq u$, $wv$ is in a step, which may not contain internal vertices. Hence $w\in V(\rho_y)-u$ and $v,w$ is a step. If $w=y$, then $u,\gamma,w$ violates the minimalty of $\gamma$ (for $z=y$), a contradiction. Let $t$ be the vertex following $w$ in $\rho_y$. If $C$ meets $\eta:=t,\rho_y,y$, then, at least two steps contained in $C$ have endvertices in $\eta$, but these steps also have endvertices in $\int(u,\rho_x,v)$ and, therefore, cross $v,w$, contradicting (RB3). So, $C$ does not meet $\eta$. Let $R'$ be obtained from $R$ by deleting $V(\eta)$ and all inner vertices of steps with endpoints in $\eta$ and, then, suppressing the degree-$2$ vertices. By Lemma \ref{smaller rope-bridge}, $R'$ is an $(u,x,w)$-rope bridge with less vertices than $R$. But, the path induced by $u,\gamma,w$ and the cycle induced by $C$ in $R'$ contradict (b). This is a contradiction to the minimality of $|V(R)|$.
\end{proof}
We say that a step in a rope-bridge is \textbf{short} if it has no inner vertices and \textbf{long} otherwise. A vertex $v$ in a rope $rho$ of an $(u,x,y)$-rope bridge is \textbf{clean} if $uv\notin E(G)$ provided $v$ is not the vertex following $u$ in $\rho$ and all steps arriving in $v$ are short. Note that (RB6) says that for the rope $\rho$ containing $v$, $v$ is clean if two steps arrive in $\int(u,\rho,v)$.
The following lemma follows directly from (RB0)-(RB6)
\begin{lemma}\label{constructions}
Let $R$ be an $(u,x,y)$-rope bridge with ropes $\rho_x$ and $\rho_y$. Write, for $z\in\{x,y\}$, $z_0:=u$ and $\rho_z:=z_0,z_1,\dots,z_{n(z)}$.
\begin{enumerate}
\item [(a)] Suppose that, for $m\ge 2$ and $t\ge 1$, $\alpha_1,\dots,\alpha_m$ are the steps arriving in $x_t$. Let $y_{k(i)}:=y_{\alpha_i}$ and suppose $k(1)\le\cdots\le k(m)$. Let $R_1$ be the graph obtained from $R$ by splitting $x_t$ into vertices $v$ and $w$ in such a way that:
\begin{itemize}
\item $x_{t-1}v\in E(R_1)$,
\item either $x_t\neq x$ and $x_{t+1}w\in E(R_1)$ or $x_t=x$ and we consider $w=x$,
\item for some $1\le l<m$, the paths corresponding to $\alpha_1,\dots,\alpha_l$ arrive in $v$ and the ones corresponding to $\alpha_{l+1},\dots,\alpha_m$ arrive in $w$ in $R_1$ and
\item $uw\notin E(R_1)$ and $uv\in E(R_1)$ if and only if $ux_t\in E(R_1)$.
\end{itemize}
Suppose that all steps arriving in $x_{t+1},\dots,x_{n(x)}$ are short. Then, $R_1$ is an $(u,x,y)$-rope bridge if one of the following assertions hold:
\begin{enumerate}
\item [(a1)] $\alpha_{l+1},\dots,\alpha_m$ are short, or
\item [(a2)] $t=l=1$.
\end{enumerate}
Moreover, a similar construction with $x$ and $y$ playing swapped roles also results in an $(u,x,y)$ rope-bridge.
\item [(b)] Suppose that $\alpha$ is a step with endvertices in $x_a$ and $y_b$, $x_{a},\dots,x_{n(x)},y_{b},\dots, y_{n(y)}$ are all clean, $\alpha$ crosses no other step, and no other step has an endvertex in $x_a$ or $y_b$. Let $R_2$ be the graph obtained from $R$ by deleting the edge of $\alpha$, splitting $x_a$ into vertices $v_1$ and $v_2$ and $y_b$ into $w_1$ and $w_2$, then, adding the edges $v_1w_2$ and $v_2w_1$ as steps in such a way that:
\begin{itemize}
\item $x_{a-1}v_1,y_{b-1}w_1\in E(R_2)$,
\item either $x_a\neq x$ and $x_{a+1}v_2\in E(R_2)$ or $x_a=x$ and we consider $v_2=x$ and
\item either $y_b\neq y$ and $y_{b+1}w_2\in E(R_2)$ or $y_b=y$ and we consider $w_2=y$.
\end{itemize}
Then $R_2$ is an $(u,x,y)$-rope bridge.
\end{enumerate}
\end{lemma}
\begin{lemma}\label{rope bridge is a minor}
If $R$ is an $(u,x,y)$-rope bridge then, up to the labels of elements other than $u$, $x$ and $y$, $R$ is a minor of a graph as in Figures \ref{fig-side1} or \ref{fig-side2} (paths that contracts to the ropes are drawn in thick lines).
\end{lemma}
\begin{center}
\begin{figure}[h]
\begin{minipage}{6cm}
\begin{tikzpicture}\centering[scale=3]
\tikzstyle{node_style2} = [shape = circle,fill = black,minimum size = 0.5pt,inner sep=0.5pt]
\tikzstyle{node_style} =[shape = circle,fill = black,minimum size = 2pt,inner sep=1pt]
\begin{scope
\fill[gray!20,opacity=1] (2,0) -- (0.5,-2) -- (1.25,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (1.75,-2) -- (2.5,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (2.75,-2) -- (3.5,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (4,-1) -- (4,1);
\node[node_style] (v) at (2,0) {};\node () at (1.8,0) {$u$};
\node[node_style] (x1) at (4,1) {};\node[node_style] (x2) at (5,1) {};
\node[node_style] (y1) at (4,-1) {};\node[node_style] (y2) at (5,-1) {};
\node[node_style2] () at (0.9,-2) {};
\node[node_style2] () at (1,-2) {};
\node[node_style2] () at (1.1,-2) {};
\node[node_style] (lfl) at (0.5,-2) {};\node[node_style] (lfm) at (0.75,-2) {}; \node[node_style] (lfr) at (1.25,-2) {};
\node[node_style2] () at (1.4,-2){};
\node[node_style2] () at (1.5,-2){};
\node[node_style2] () at (1.6,-2){};
\node[node_style2] () at (2.15,-2){};
\node[node_style2] () at (2.25,-2){};
\node[node_style2] () at (2.35,-2){};
\node[node_style] (mfl) at (1.75,-2) {};\node[node_style] (mfm) at (2,-2) {}; \node[node_style] (mfr) at (2.5,-2) {};
\node[node_style2] () at (3.15,-2){};
\node[node_style2] () at (3.25,-2){};
\node[node_style2] () at (3.35,-2){};
\node[node_style] (rfl) at (2.75,-2) {};\node[node_style] (rfm) at (3,-2) {}; \node[node_style] (rfr) at (3.5,-2) {};
\draw (y2)--(y1)--(v)--(x1)--(x2);
\draw (v)--(lfl)--(lfm)--(v)--(lfr);
\draw (v)--(mfl)--(mfm)--(v)--(mfr);
\draw (v)--(rfl)--(rfm)--(v)--(rfr);
\node[node_style] (1) at (4,-0.6){};
\node[node_style] (2) at (4,-0.2){};
\node[node_style] (3) at (4,0.6){};
\node[node_style2] () at (4,0.1){};
\node[node_style2] () at (4,0.2){};
\node[node_style2] () at (4,0.3){};
\draw (y1)--(1)--(v)--(2)--(1)--(v)--(3)--(x1);
\draw (y2)--(rfr);
\draw (y2)--(3.7,-2.1)--(2.5,-2.1)--(2.5,-2);
\draw (y2)--(3.9,-2.2)--(1.25,-2.2)--(1.25,-2);
\draw (x2)--(5.2,0)--(5.2,-2.4)--(2.75,-2.4)--(rfl);
\draw (x2)--(5.3,0)--(5.3,-2.5)--(1.75,-2.5)--(mfl);
\draw (x2)--(5.4,0)--(5.4,-2.6)--(0.5,-2.6)--(lfl);
\node[node_style] (x3) at (5,1.5){};\node[node_style] (x4) at (4.5,1.5){};
\node[node_style] (y3) at (5,2){};\node[node_style] (y4) at (4.5,2){};
\draw (x3)--(y4);\draw(y3)--(x4);
\node[node_style] (x5) at (4,1.5){};\node[node_style] (x6) at (3.5,1.5){};
\node[node_style] (y5) at (4,2){};\node[node_style] (y6) at (3.5,2){};
\draw (x5)--(y6);\draw(y5)--(x6);
\node[node_style] (x7) at (2.5,1.5){};\node[node_style] (x8) at (2,1.5){};
\node[node_style] (y7) at (2.5,2){};\node[node_style] (y8) at (2,2){};
\draw (x7)--(y8);\draw(y7)--(x8);
\node[node_style](x9)at(1.5,1.5){};\node[node_style](x10)at(1,1.5){};\node()at(0.8,1.5){$x$};
\node[node_style] (y9) at (1.5,2){};\node[node_style] (y10) at (1,2){};\node()at(0.8,2){$y$};
\draw (x9)--(y10);\draw(y9)--(x10);
\draw [line width=1](v)--(x1)--(x2)--(x3)--(3.3,1.5);
\node () at (3.0,1.45) {$\cdots$};
\draw [line width=1] (2.7,1.5)--(x10);
\draw [line width=1] (v)--(y1)--(y2)--(5.6,-1)--(5.6,2)--(3.3,2);
\node () at (3.0,1.95) {$\cdots$};
\draw [line width=1](2.7,2)--(y10);
\end{scope}
\end{tikzpicture}
\caption{}\label{fig-side1}
\end{minipage}
\begin{minipage}{6cm}
\begin{tikzpicture}\centering
\tikzstyle{node_style2} = [shape = circle,fill = black,minimum size = 0.5pt,inner sep=0.5pt]
\tikzstyle{node_style} =[shape = circle,fill = black,minimum size = 2pt,inner sep=1pt]
\begin{scope
\fill[gray!20,opacity=1] (2,0) -- (0.5,-2) -- (1.25,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (1.75,-2) -- (2.5,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (2.75,-2) -- (3.5,-2) -- cycle;
\fill[gray!20,opacity=1] (2,0) -- (4,-1) -- (4.4,-0.2);
\fill[gray!20,opacity=1] (2,0) -- (4,1) -- (4.4, 0.2);
\node[node_style] (v) at (2,0) {};\node () at (1.8,0) {$u$};
\node[node_style] (x1) at (4,1) {};\node[node_style] (x2) at (5,1) {};
\node[node_style] (y1) at (4,-1) {};\node[node_style] (y2) at (5,-1) {};
\node[node_style2] () at (0.9,-2) {};
\node[node_style2] () at (1,-2) {};
\node[node_style2] () at (1.1,-2) {};
\node[node_style] (lfl) at (0.5,-2) {};\node[node_style] (lfm) at (0.75,-2) {}; \node[node_style] (lfr) at (1.25,-2) {};
\node[node_style2] () at (1.4,-2){};
\node[node_style2] () at (1.5,-2){};
\node[node_style2] () at (1.6,-2){};
\node[node_style2] () at (2.15,-2){};
\node[node_style2] () at (2.25,-2){};
\node[node_style2] () at (2.35,-2){};
\node[node_style] (mfl) at (1.75,-2) {};\node[node_style] (mfm) at (2,-2) {}; \node[node_style] (mfr) at (2.5,-2) {};
\node[node_style2] () at (3.15,-2){};
\node[node_style2] () at (3.25,-2){};
\node[node_style2] () at (3.35,-2){};
\node[node_style] (rfl) at (2.75,-2) {};\node[node_style] (rfm) at (3,-2) {}; \node[node_style] (rfr) at (3.5,-2) {};
\draw (y2)--(y1)--(v)--(x1)--(x2);
\draw (v)--(lfl)--(lfm)--(v)--(lfr);
\draw (v)--(mfl)--(mfm)--(v)--(mfr);
\draw (v)--(rfl)--(rfm)--(v)--(rfr);
\node[node_style] (lofl) at (4.1,-0.8){};\node[node_style] (lofu) at (4.4,-0.2){}; \draw (y1)--(lofl)--(v)--(lofu);
\node[node_style2] (ldot1) at (4.2,-0.6){};
\node[node_style2] (ldot2) at (4.25,-0.5){};
\node[node_style2] (ldot3) at (4.3,-0.4){};
\node[node_style] (upfu) at (4.1,0.8){};\node[node_style] (upfl) at (4.4, 0.2){}; \draw (x1)--(upfu)--(v)--(upfl);
\node[node_style2] (ldot1) at (4.2,0.6){};
\node[node_style2] (ldot2) at (4.25,0.5){};
\node[node_style2] (ldot3) at (4.3,0.4){};
\draw (lofu)--(x2);
\draw (upfl)--(y2);
\draw (y2)--(rfr);
\draw (y2)--(3.7,-2.1)--(2.5,-2.1)--(2.5,-2);
\draw (y2)--(3.9,-2.2)--(1.25,-2.2)--(1.25,-2);
\draw (x2)--(5.2,0)--(5.2,-2.4)--(2.75,-2.4)--(rfl);
\draw (x2)--(5.3,0)--(5.3,-2.5)--(1.75,-2.5)--(mfl);
\draw (x2)--(5.4,0)--(5.4,-2.6)--(0.5,-2.6)--(lfl);
\node[node_style] (x3) at (5,1.5){};\node[node_style] (x4) at (4.5,1.5){};
\node[node_style] (y3) at (5,2){};\node[node_style] (y4) at (4.5,2){};
\draw (x3)--(y4);\draw(y3)--(x4);
\node[node_style] (x5) at (4,1.5){};\node[node_style] (x6) at (3.5,1.5){};
\node[node_style] (y5) at (4,2){};\node[node_style] (y6) at (3.5,2){};
\draw (x5)--(y6);\draw(y5)--(x6);
\node[node_style] (x7) at (2.5,1.5){};\node[node_style] (x8) at (2,1.5){};
\node[node_style] (y7) at (2.5,2){};\node[node_style] (y8) at (2,2){};
\draw (x7)--(y8);\draw(y7)--(x8);
\node[node_style](x9)at(1.5,1.5){};\node[node_style](x10)at(1,1.5){};\node()at(0.8,1.5){$x$};
\node[node_style] (y9) at (1.5,2){};\node[node_style] (y10) at (1,2){};\node()at(0.8,2){$y$};
\draw (x9)--(y10);\draw(y9)--(x10);
\draw [line width=1](v)--(x1)--(x2)--(x3)--(3.3,1.5);
\node () at (3.0,1.45) {$\cdots$};
\draw [line width=1] (2.7,1.5)--(x10);
\draw [line width=1] (v)--(y1)--(y2)--(5.6,-1)--(5.6,2)--(3.3,2);
\node () at (3.0,1.95) {$\cdots$};
\draw [line width=1](2.7,2)--(y10);
\end{scope}
\end{tikzpicture}
\caption{}
\label{fig-side2}
\end{minipage}
\end{figure}
\end{center}
\begin{proof}
Suppose that $R'$ is a graph contradicting the lemma. Let $R$ be a graph obtained from $R'$ using the operations of Lemma \ref{constructions} up to the point that they can no longer be performed. If we prove the lemma for $R$, it will also hold for $R'$. If all vertices in the ropes are clean the result is clear. So, we may assume that there is a non-clean vertex. Write $\rho_z=u,z_1,\dots,z_{n(z)}$ for each $z\in \{x,y\}$. Consider the smallest index $c(z)$ such that all vertices $z_{c(z)+1},\dots,z_{n(z)}$ are clean. By (RB6), $1\le c(z) \le 2$.
If $c(z)<t\le n(z)$, there is an unique step arriving in $z_t$ since operation (a) of Lemma \ref{constructions} cannot be performed and (a1) would hold otherwise. In particular, the step arriving in $z_t$ must cross another step since operation (b) of Lemma \ref{constructions} also cannot be performed.
Also, there is an unique step $\alpha_z$ arriving in $z_1$, as we prove next. If $c(z)=2$, it follows from (RB6). If $c(z)=1$, it follows from the fact that item (a1) of Lemma \ref{constructions} does not hold.
If there are no long steps, $x_2$ and $y_2$ are the unique possibly non-clean vetices and the result holds. So, we may assume that there is some long step.
If $c(x)=1$, then $x_1$ is the unique non-clean vertex of $\rho_x$ and, as an unique step arriving at $x_1$, $\alpha_x$ is the unique long step. If $\alpha_y=\alpha_x$, $R$ is a minor of a graph as in Figure \ref{fig-side1}, otherwise it is a minor a graph as in Figure \ref{fig-side2}.
So, we may assume that $c(x)=2$ and, analogously, that $c(y)=2$. If $\alpha_x=\alpha_y$, then $R$ is a minor of a graph as in Figure \ref{fig-side1}. Otherwise, as each step cross at most one other, $\alpha_x$ has an endvertex in $y_2$ and $\alpha_y$ has an endvertex in $x_2$. It follows that $R$ is a minor of a graph as in Figure \ref{fig-side2}.
\end{proof}
\begin{lemma}\label{components are rope-bridges}
Suppose that $S=\{x,y\}$ and let $\{i,j\}=\{1,2\}$. Moreover, suppose
\begin{enumerate}
\item [(i)] $d_G(u_i)\ge 4$ and $S^+$ is a two-sided vertex-cut or
\item [(ii)] $N_G(u_i)=\{u_j,x,y\}$.
\end{enumerate}
For $\{k,l\}=\{1,2\}$, we define an \textbf{$u_k$-component} as a connected component of $G\del S^+$ with a neighbor of $u_k$ and no neighbor of $u_l$. Then, one of the following assertions hold:
\begin{enumerate}
\item [(a)] $N_G(u_j)=\{x,y,u_i\}$,
\item [(b)] There is an unique $u_j$-component $K$ and $K+\{u_j,x,y\}$ is a $(u_j,x,y)$-rope bridge.
\item [(c)] There is at least two $u_j$-components and, for each $u_j$-component $K$, $K+\{u_j,x,y\}$ is a fan with hub $u_j$ and endvertices $x$ and $y$ or a wheel with hub $u_j$ and $xy$ as edge.
\end{enumerate}
\end{lemma}
\begin{proof}
Suppose that (a) fails. Then, there is some $u_j$-component. Note that either there is an $u_i$-component or $N_G(u_i)\cont S^+$. In both cases, for each $z\in\{x,y\}$, there is a $(z,u_i)$-path $\beta_z$ whose internal vertices are all in an $u_i$-component (in case they exist).
First suppose that there is an unique $u_j$-component $K$. By Menger's Theorem, there are three internally disjoint paths from $u_j$ to $u_i$, one of these must contain $x$, namely $\zeta_x$, and other must contain $y$, call it $\zeta_y$. We define, for $z=x,y$, $\rho_z:=u_j,\zeta_z,z$. By the uniqueness of $K$, $\rho_x$ and $\rho_y$ are in $R:=K+\{u_j,x,y\}$. We shall prove that $R$ is an $(u_j,x,y)$-rope-bridge with ropes $\rho_x$ and $\rho_y$. For this we use Lemma \ref{rope-bridge criterion}; if $R$ does not satisfy item (b) of that lemma, there is $z\in\{x,y\}$, a cycle $C$ of $R\del\{u_j,z\}$ and an $(u_j,z)$-path $\alpha$ avoiding $C$. Now $u_i,u_j,\alpha,z,\beta_z,u_i$ is a cycle avoiding $C$, a contradiction. Thus, (b) holds.
Now suppose that there are $u_j$-components $K_1,\dots,K_m$ for some $m\ge 2$. Let us show that each $K_k$ is a tree. Suppose for a contradiction that $D$ is a cycle in, say, $K_1$. Now a cycle of the form $u_i,u_j,K_2,x,\beta_x,u_i$ avoids $D$, a contradiction. Then $K_k$ is a tree for $k=1,\dots,m$.
Let us prove that, for each $k=1,\dots,m$, each element of $\{x,y\}$ has an unique neighbor in $K_k$. Say that $y$ have two different neighbors in $K_1$. Now, $K_1+y$ has a cycle $C$. Note that $C$ avoids a cycle of the form $u_i,u_j,K_2,x,\beta_x,u_i$, a contradiction. So, each element of $\{x,y\}$ has an unique neighbor in $K_k$ for each $k=1,\dots,m$.
If $K_k$ has three different leaves, then two of them have a same neighbor in $\{x,y\}$ a contradiction. Then $K_k$ is a path. Therefore, each leaf of $K_k$ has a different neighbor in $\{x,y\}$ and each vertex of $K_k$ must have $u_j$ as neighbor. This implies (c).
\end{proof}
Recall the definitions of type (a), (b) and (c) graphs from Theorem \ref{main}. It is straighforward to check that all graphs decribed in items (a), (b) and (c) of Theorem \ref{main} have no pair of disjoint cycles whose union contains $e$. Therefore, Theorem \ref{main} is a direct consequence of the following theorem:
\begin{theorem}
\label{main-cor}
Suppose that $G$ is a $3$-connected graph with at least six vertices and an edge $e:=u_1u_2$ such that there is no pair of disjoint cycles of $G$ which union contains the edge $e$. Let $S$ be smallest set such that $S^+:=S\cup\{u_1,u_2\}$ contains a vertex cut of $G$. Then one of the following assertions holds.
\begin{enumerate}
\item [(a)] $|S|=2$, $d_G(u_1),d_G(u_2)\ge 4$, we may not pick $S$ such that $S^+$ is two-sided and $G$ is a type (a) $e$-Dirac graph.
\item [(b)] $|S|=1$ and $G$ is a type (b) $e$-Dirac graph.
\item [(c)] $|S|=2$, $G$ is a type (c) $e$-Dirac graph, we may pick $S$ with the property that for some $\{i,j\}=\{1,2\}$ either $N_G(u_i)=S\cup u_2$ or $d_G(u_i)\ge 4$ and $S^+$ is a $2$-sided vertex cut. Moreover, we call the {\bf $u_j$-components} the connected components of $G-S^+$ with a neighbor of $u_j$ and one of the following assertions holds:
\begin{enumerate}
\item [(c1)] $N_G(u_j)=S\cup u_i$;
\item [(c2)] there is only one $u_j$-component $J$, which has the property that, for $S=\{x,y\}$, $J+\{u_j,x,y\}$ is an $(u_j,x,y)$-rope bridge; or
\item [(c3)] there is more that one $u_j$-component and each $u_j$-conponent $J$ has the property that either $J+(S\cup u_j)$ either is a fan with $u_j$ as hub and with endvertices in $S$ or is a wheel with $u_2$ as hub and contains an edge linking the vertices of $S$.
\end{enumerate}
\end{enumerate}
\end{theorem}
\begin{proof}
Suppose that the theorem fails. If $|S|=1$, (b) follows from Lemma \ref{all type b}. Assume $|S|\ge2$. We will split the proof into two cases as follows.
\smallskip
{\it Case 1: Some vertex in $\{u_1,u_2\}$ has degree $3$ in $G$ or we may pick $S$ such that $S^+$ is a two-sided vertex-cut. }
By Lemma \ref{Sle2}, the hypothesis of Lemma \ref{components are rope-bridges} holds. Consider the terminologies as in that lemma. Let $\{i,j\}=\{1,2\}$ and let $K_1,\dots,K_m$ be the $u_j$-components. To prove that $G$ is a type (c) graph it suffices to prove that there is an isomorphism $\Phi$ between $H:=G[K_1\cup\cdots\cup K_m\cup \{u_j,x,y\}]$ and a minor of a graph like in Figures \ref{fig-side1} or \ref{fig-side2} such that $\Phi$ preserves $x$ and $y$ and $\varphi(u_j)=u$. The result is clear if items (a) or (c) of Lemma \ref{components are rope-bridges} holds. So we may assume that item (b) of that lemma holds. Now, (c) follows from Lemma \ref{rope bridge is a minor}.
\smallskip
{\it Case 2: $d_G(u_1),d_G(u_2)\ge 4$ and we cannot choose $S$ so that $S^+$ is a two-sided vertex cut.}
As $d_G(u_1),d_G(u_2)\ge 4$, by Lemma \ref{Sle2}, $|S|=2$. If all $4$-vertex cuts of $G$ are no-sided, (a) follows from Lemma \ref{ugly}. Assume the contrary. As Case 1 does not hold, we may pick $S=\{x,y\}$ in such a way that $S^+$ is one-sided.
If $\kappa\ge 3$, then cycles of the form $u_1,u_2,G_1,u_1$ and $x,G_2,y,G_3,x$ avoid each other. Thus, $\kappa=2$. If an element of $S$ has two edges to a same component of $G\del S^+$, say $x$ has two edges to $G_1$, then cycles of the form $x,G_1,x$ and $u_1,u_2,G_2,u_1$ avoid each other. Thus, each element of $S$ has exactly one edge to each component of $G\del S^+$.
Let $\{i,j\}=\{1,2\}$. If $G_i$ has a cycle, then it avoids a cycle of the form $u_1,u_2,G_j,u_1$. Hence $G_i$ is a tree. If $l$ is a leaf of $G_i$ with no neighbor in $S$, then $u_1l,u_2l\in E(G)$ and $S$ has neighbors in $G_i\del l$. This implies that a cycle of $G\del{u_i,u_j,l}$ avoids $u_1,u_2,l,u_1$, a contradiction. Thus each leaf of $G_i$ has a neighbor in $\{x,y\}$. As a consequence, if $G_i$ has three different leaves, then two of them have a same neighbor in $S$ and a vertex of $S$ has two different neighbors in $G_i$, a contradiction. Therefore, we may write $G_i$ as a path $w_1,\dots,w_n$ with $N_G(x)\cap V(G_i)=\{w_1\}$ and $N_G(y)\cap V(G_i)=\{w_n\}$. Now, to prove (a), it is left to check that $xy\notin E(G)$. Indeed, suppose the contrary. Then, $G_i+\{x,y\}$ has a cycle $C$. But, as $S^+$ is two sided, $G_2+\{u_1,u_2\}$ has a cycle containing $e$ and avoiding $C$, a contradiction.
\end{proof}
\section{Strongly $e$-Dirac graphs}\label{sec-strong}
In this section, we prove Theorems \ref{main-strong} and \ref{main-strong-2con}. We say that a graph $G$ with an edge $e$ is \textbf{strongly $e$-Dirac} if $G$ has no pair of edge-disjoint cycles whose union contains $e$. Clearly all strongly $e$-Dirac graphs are $e$-Dirac graphs.
\begin{lemma}\label{str type b}
All type (b) $e$-Dirac $3$-connected graphs with more than five vertices are not strongly $e$-Dirac.
\end{lemma}
\begin{proof}
Suppose that $G$ is a graph contradicting the lemma and let $e=uv$. Consider a $3$-vertex-cut $S:=\{u,v,w\}$ of $G$ as in item (b) of Theorem \ref{main}. If $G\del S$ has three distinct connected components $K_1$, $K_2$ and $K_3$, then $G$ has edge-disjoint cycles of the form $u,v,K_1,u$ and $v,K_2,w,K_3,v$, a contradiction. Thus $G\del S$ has exactly two connected components, $K$ and $K'$. As $|G|\ge 6$, we may assume that $|K|\ge 2$. Note that there are two leaves $l_1$ and $l_2$ in $K$ such that $x_K\neq l_1$. As $d_G(l_2)\ge 3$, $l_2$ has a neighbor in $\{u,v\}$, say $v$, while both $u$ and $v$ are adjacent to $l_1$. Consider the $(l_1,l_2)$-path $\gamma$ of $K$. Now, $G$ has $v,l_1,\gamma,l_2,v$ edge-disjoint from a cycle of the form $u,K',v,u$, a contradiction.
\end{proof}
\begin{lemma}\label{str type a}
If $G$ is a $3$-connected type (a) strongly $e$-Dirac graph, then $G$ is isomorphic to the prism or to $K_{3,3}$.
\end{lemma}
\begin{proof}
Suppose that $G$ contradicts the lemma and let $G$ be obtained by splitting the hub of a wheel with rim $R:=x_1,\dots,x_n$. If, in $G$, $x$ is a common neighbor to $u$ and $v$, then $u,v,x,u$ is disjoint from the rim, a contradiction, therefore $u$ and $v$ has no common neighbor in $G$. This implies that $n\ge 4$. Let us prove that $n=4$. Suppose the contrary. Then we may assume that $u$ has three neighbors $y_1$, $y_2$ and $y_3$ in this order in a cyclic ordering of $R$ and $v$ has a neighbor $z$ in $R$, say after $y_3$ and before $y_1$ in this same ordering. Let $\alpha$ be the $(y_1,y_2)$-path of $R$ avoiding $z$ and $\beta$ the $(y_3,z)$-path of $R$ avoiding $\alpha$. Now $u,y_1,\alpha,y_2,u$ is edge-disjoint from $v,u,y_3,\beta,z,v$, a contradiction. So, $n=4$. Now the result is clear.
\end{proof}
We say that an $(u,x,y)$-rope-bridge is \textbf{strong} if all its steps have length one and no pair of steps have a common endvertex.
\begin{lemma}\label{str-char1}
A graph $R$ with a $3$-subset $\{u,x,y\}\cont V(G)$ is a strong $(u,x,y)$-rope-bridge with ropes $\rho_x$ and $\rho_y$ if and only if
$R$ has internally disjoint paths $\rho_x:=v_0,\dots,v_n$ and $\rho_y:=w_0,\dots,w_n$, with $v_n=x$, $w_n=y$ and $v_0=w_0=u$ and there is a family $\mathcal{P}$ of pairwise disjoint pairs of consecutive elements of $\{1,\dots,n\}$ such that
\begin{enumerate}
\item [(a)] $V(R)=V(\rho_x)\cup V(\rho_y)$ and
\item [(b)] $E(R)=\{v_aw_b,v_bw_a:\{a,b\}\in \mathcal P\}\cup\{v_cw_c:c$ is in no member of $\mathcal P\}\cup E(\rho_x)\cup E(\rho_y)$.
\end{enumerate}
\end{lemma}
\begin{proof}
It is clear that a graph satisfying the given conditions is a strong $(u,x,y)$-rope-bridge with ropes $\rho_x$ and $\rho_y$. Let us prove the converse. As no pair of steps shares the same endvertices, then $\rho_x$ and $\rho_y$ have the same number of vertices and we can label $\rho_x=v_0,\dots,v_n$ and $\rho_y=w_0,\dots,w_n$. We let $\sigma_i$ be the step with extremity in $v_i$. Let $\mathcal P$ be the family of pairs $\{i,j\}$ such that $\sigma_i$ crosses $\sigma_j$. By (RB3), the members of $\mathcal P $ are pairwise disjoint. It also follows from (RB3) that each pair in $\mathcal P$ contains consecutive indices. Analogously, for a pair $\{i,j\}\in \mathcal P$, $\sigma_i$ and $\sigma_j$ also have endvertices that are neighbors in $\rho_y$. This implies the lemma.
\end{proof}
\begin{lemma}\label{strong rope-bridge criterion}
Let $R$ be a connected graph with vertices $u,x,y$ and paths $\rho_x$ and $\rho_y$ from $u$ to $x$ and $y$ respectively, satisfying $V(\rho_y)\cap V(\rho_x)=\{u\}$. Suppose that $d_G(v)\ge 3$ for all $v\in V(R)-\{u,x,y\}$. Then the following assertions are equivalent:
\begin{enumerate}
\item [(a)] $R$ is a strong $(u,x,y)$-rope-bridge with ropes $\rho_x$ and $\rho_y$.
\item [(b)] If, for $z\in \{x,y\}$, $C$ is a cycle of $R$, then $R$ has no $(u,z)$-path edge-disjoint from $C$.
\end{enumerate}
\end{lemma}
\begin{proof} It follows from Lemma \ref{str-char1} that (a) implies (b). Suppose (b). This implies item (b) of Lemma \ref{rope-bridge criterion}. So, $R$ is a $(u,x,y)$-rope-bridge with ropes $\rho_x$ and $\rho_y$. Suppose that $R$ is not strong. Then, some step $\sigma$ has an inner vertex $z$ or two steps $\alpha$ and $\beta$ have a common end-vertex. In the former case $\rho_x$ is edge-disjoint from the cycle $u,z,\sigma,y_\sigma,\rho_y,u$, a contradiction. In the later case, we may assume that $x_\alpha=x_\beta$. This implies that $\rho_x$ is edge-disjoint from the cycle $x_\beta,\beta,y_\beta,\rho_y,y_\alpha,\alpha,x_\alpha$, a contradiction again. So, (a) holds.
\end{proof}
\begin{proofof}\emph{Proof of Theorem \ref{main-strong}: } First note that all graphs described in the theorem are strongly $e$-Dirac. Let us prove the converse. Let $G$ be a strongly $e$-Dirac graph. If $G$ is a type (a) or (b) $e$-Dirac graph as in Theorem \ref{main}, then the result follows from Lemmas \ref{str type a} and \ref{str type b} respectively. Assume that $G$ is a type (c) $e$-Dirac graph. If $|G|\le 5$, the result may be verified directly. Thus we may assume that $|G|\ge 6$ and, by Lemma \ref{complete}, the assumptions for the previous results are valid. So, item (c) of Theorem \ref{main-cor} holds and we have the hypothesis of Lemma \ref{components are rope-bridges} holding, consider the terminologies as in that lemma. As $|G|\ge 6$ we may assume the existence of some $u_j$-component.
Let $K_1,\dots, K_m$ be the $u_j$-components of $G$. For each $z\in\{x,y\}$ there is a $(z,u_i)$-path $\alpha_z$ with all internal vertices out of $K_1\cup\cdots\cup K_m\cup S^+$.
If item (b) of Lemma \ref{components are rope-bridges} holds then, $R:=K_1+\{u_j,x,y\}$ is an $(u_j,x,y)$-rope bridge. If for some cycle $C$ of $R$ and $z\in\{x,y\}$, there is an $(u_j,z)$-path $\beta$ edge-disjoint from $C$, then $C$ is edge disjoint from a cycle of the form $u_i,u_j,\beta,z,\alpha_z,u_i$, a contradiction. So, there are no such path and cycle. Thus, by Lemma \ref{strong rope-bridge criterion}, $R$ is a strong $(u_j,x,y)$-rope bridge.
By Lemma \ref{str-char1}, this implies the theorem if all $\{i,j\}=\{1,2\}$ satisfies items (a) or (b) of Lemma \ref{components are rope-bridges}. So, we may assume that item (c) holds for $(i,j)=(1,2)$.
If for some $z\in\{x,y\}$, there is an $(u_2,z)$-path $\gamma$ whose internal vertices are in an $u_j$-component other than $K_1$ and $K_2$, then $u_1,u_2,\gamma,z,\alpha_z,u_1$ is edge-disjoint from a cycle of the form $x,K_1,y,K_2,x$, a contradiction. So, $m=2$ and $u_2x,u_2y\notin E(G)$.
If $u_2$ has two neighbors in a same $u_2$-component, say $K_2$, then $K_2+u_2$ has a cycle which is edge-disjoint from a cycle of the form $u_1,u_2,K_1,x,\alpha_x,u_1$; a contradiction. So, $u_2$ has an unique neighbor in each $u_2$-component. By item (c) of Lemma \ref{components are rope-bridges}, $|K_1|=|K_2|=1$.
If there is no $u_1$-component, either $G\cong K_{3,3}$ and the theorem holds or $xy\in E(G)$ and cycles of the form $x,y,K_1,x$ and $u_1,u_2,k_2,x,u_1$ are edge-disjoint, a contradiction. So, we may assume that $J$ is a $u_1$-component. If there is a second $u_1$-component $J'$, then cycles of the form $u_1,u_2,K_1,x,J,u_1$ and $x,K_2,y,J',x$ are edge-disjoint. Thus, $J$ is the unique $u_1$-component. Hence, item (b) of Lemma \ref{components are rope-bridges} holds for $(i,j)=(2,1)$. If $R$ has a cycle $C$ and aj $(u_1,z)$-path $\gamma$ disjoint from $C$ for some $\in \{x,y\}$, then a cycle of the from $u_1,\gamma,z,K_1,u_2,u_1$ is edge-disjoint from $C$, a contradiction. By Lemma \ref{strong rope-bridge criterion}, $R$ is a strong $(u_1,x,y)$-rope bridge. If $xy\in E(G)$, then cycles of the form $u_1,u_2,K_1,x,J,u_1$ and $x,y,K_2,x$ are edges-disjoint. so $xy\notin E(G)$. Consider labels for the vertices of $R$ like in Lemma \ref{str-char1}. We let $V(K_1):=\{v_{n+1}\}$ and $V(K_2):=\{w_{n+1}\}$. Now it is straightforward to check that $G$ is in the format described in the theorem. \end{proofof}
\begin{proofof}\emph{Proof of Theorem \ref{main-strong-2con}: } Consider a graph $G$ as described in the theorem. Note that all cycles not contained in one of the $G_i$'s are the cycles containing $U$, which are exactly the cycles containing $e$. As each $G_i$ is $u_{i-1}u_i$-Dirac, it follows that $G$ is $e$-Dirac. For the converse, suppose for a contradiction that $G$ is a $2$-connected strongly $e$-Dirac graph not fitting into the description of the theorem.
The $3$-connected strongly $e$-Dirac graphs, described in Theorem \ref{main-strong}, fit into the description in this theorem. So $G$ is not $3$-connected. As the theorem also holds if $|G|\le 3$, then $G$ has a $2$-vertex-cut $\{x,y\}$. This implies that we may write $G$ as the union of two graphs $H$ and $K$ such that $|H|,|K|\ge 3$, $V(H)\cap V(K)=\{x,y\}$, $H+xy$ and $K+xy$ are $2$-connected, $e\in V(H)$ and $xy\notin E(K)$.
Let us check that $H+xy$ is a strongly $e$-Dirac graph. Suppose for a contradiction that $H+xy$ has a pair of edge-disjoint cycles $(D_1,D_2)$ with $e\in D_1$. Then for some $i\in \{1,2\}$, $D_i$ is not a cycle of $G$. So, $xy\in E(D_i)-E(H)$. Let $D$ be a cycle of $K+xy$ containing $xy$. Now $(D_i\cup D)\del xy$ and $D_{3-i}$ are disjoint cycles of $G$ whose union contains $e$, contradicting the fact that $G$ is strongly $e$-Dirac. So $H+xy$ is a $2$-connected strongly $e$-Dirac graph. As $|H|<|G|$, the theorem holds for $H+xy$. Consider, for $H+xy$, graphs $G'_1,\dots,G'_{n'}$ and vertices $u'_0,\dots,u'_{n'}$ as in the theorem, with $e=u'_0u'_{n'}$.
As $K+xy$ is $2$-connected and $xy\notin E(K)$, then either $K$ has a cycle or $K$ is an $xy$-path. In the later case, $G$ is isomorphic to a subdivision of $H$ and as, the theorem holds for $H$, it is straighforward to verify that it also holds for $G$, so $K$ has a cycle $C_K$.
If $H$ has a cycle $C_H$ containing $e$, then $C_H$ and $C_K$ are edge-disjoint cycles of $G$, a contradiction. So $e$ is in no cycle of $H$. This implies that $xy\notin E(H)$ and $\{e,xy\}$ is an edge-cut of $H+xy$ since $H+xy$ is $2$-connected. By the description of $H+xy$ as in the theorem, $\{x,y\}=\{u'_{i-1},u'_i\}$ for some index $i\in \{1,\dots,n'\}$ such that $V(G'_i)=\{u'_{i-1},u'_i\}$. We may assume without loss of generality that $(x,y)=(u'_{i-1},u'_i)$.
Let us check that $K+xy$ is strongly $xy$-Dirac. Suppose for a contradiction that $K+xy$ has a pair of edge-disjoint circuits $(C_1,C_2)$ with $xy\in E(C_1)$. Then for a circuit $C$ of $H+xy$ with $e,xy\in E(C)$, $((C\cup C_1)\del xy,C_2)$ is a pair of edge-disjoint cycles of $G$ whose union contains $e$, a contradiction. So $K+xy$ is a $2$-connected strongly $xy$-Dirac graph. As $|K|<|G|$, we may apply the theorem in $K+xy$ in respect to the edge $xy$. Consider, for $K+xy$, graphs $G''_1,\dots,G''_{n''}$ and vertices $u''_0,\dots,u''_{n''}$ as in the theorem with $(x,y)=(u''_0,u''_{n''})$.
Now the graphs
\[(G_1,\dots,G_n):=(G'_1,\dots,G'_{i-1},G''_1,\dots,G''_{n''},G'_{i+1},\dots,G'_{n'})\] and vertices
\[(u_0,\dots,u_n):=(u'_0,\dots,u'_{i-1},u''_1,\dots,u''_{n''-1},u'_{i},\dots,u'_{n'})\] give a description of $G$ according to the theorem.
\end{proofof}
\section{Prism-Minors}\label{sec-prism}
In this section we prove Theorem \ref{thm-prism}. The theorem follows straightforwardly from Lemmas \ref{prism a}, \ref{prism b} and \ref{prism c}.
If $G$ is a graph with a subgraph $H'$ isomorphic to the subdivision of a graph $H$, we say that and $H$-minor of $H'$ is an \textbf{$H$-topological minor} of $G$. If $G$ has an $H$-topological minor using an edge $e$, then it is clear that $G$ has an $H$-minor using $e$. The converse does not hold in general, but it is easy to verify that it is true provided $G$ and $H$ are $3$-connected and $H$ is cubic, which is the case in our concern: when $H$ is the prism and $G$ is $3$-connected. We will use this fact with no mentions.
Let $G$ be a $3$-connected graph with an edge $e$. By Menger's Theorem, $G$ is not $e$-Dirac if and only if $G$ has a prism-minor $H$ using $e$ as an edge in a triangle of $H$. So, our problem lies within the class of $e$-Dirac graphs. Moreover, the following lemma is valid.
\begin{lemma}\label{prism-menger}
A $3$-connected $e$-Dirac graph has a prism-minor using $e$ if and only if it has vertex-disjoint cycles $C$ and $D$ and three vertex-disjoint $(V(C),V(D))$-paths $\alpha_1$, $\alpha_2$, and $\alpha_3$ such that $e\in E(\alpha_3)$.
\end{lemma}
The next two lemmas proves Theorem \ref{thm-prism} for $e$-Dirac graphs of types (a) and (b).
\begin{lemma}\label{prism a}
If $G$ is a type (a) $e$-Dirac graph, then $G$ has a prism-minor using $e$.
\end{lemma}
\begin{proof}
For some $n\ge 4$, we may assume that $G$ may be obtained from $W_n$ by possibly doubling some spokes and splitting the hub into the edge $e=uv$, where $u$ and $v$ has degree at least four. Say that $|N_G(v)|\ge|N_G(u)|$. Choose $\{a,b\}\cont N_G(u)-v$ minimizing the number of neighbors of $v$ in $\{a,b\}$.
If $n\ge 5$, by the minimality of $|\{a,b\}\cap N_G(v)|$ and as $|N_G(v)|\ge|N_G(u)|\ge 4$, there are at least three neighbors of $v$ out of $\{a,b,u\}$. So $v$ have neighbors $c$ and $d$ with the property that $a$, $b$, $c$ and $d$ appear in this order in some cycle ordering of the cycle $G-\{u,v\}$. This implies that $G$ has a subdivision of the prism containing $e$ and this implies the lemma.
So we may assume that $n=4$. Recall that, by the description of the type (a) graphs, $d_G(u)=d_G(v)\ge4$. Let $a$, $b$, $c$, and $d$ be a cycle ordering of $G-\{u,v\}$ with the property that $b$, $c$, and $d$ are neighbors of $v$. Either $b$ or $d$ is a neighbor of $u$, we may assume it is $b$ as swapping the labels just inverts the cycle ordering. Now $\{u,a,b\}$ and $\{v,c,d\}$ induces triangles in $G$. But $ad$, $bc$ and $e=uv$ are edges of $G$. So $e$ is in a prism-minor of $G$. This finishes the proof.
\end{proof}
\begin{lemma}\label{prism b}
If $G$ is a type (b) $e$-Dirac graph, then $G$ has no prism-minor using $e$.
\end{lemma}
\begin{proof}
For $e=uv$ and some vertex $w$, $A:=\{u,v,w\}$ is a vertex-cut of $G$ and each connected component of $G-A$ is a tree with an unique neighbor of $w$. Suppose for a contradiction that there is a prism-minor of $G$ using $e$. By lemma \ref{prism-menger}, $G$ has vertex disjoint cycles $C$ and $D$ and vertex-disjoint $(V(C),V(D))$-paths $\alpha_1$, $\alpha_2$ and $\alpha_3$ with $e\in E(\alpha_3)$.
As the components of $G-A$ are trees, both $C$ and $D$ meet $A$. Since $w$ has an unique neighbor in each connected component of $G-A$, then it is not possible that $V(C)\cap A=\{w\}$ or $V(D)\cap A=\{w\}$. So, we may assume that $u\in V(C)$ and $v\in V(D)$.
For $i=1,2$, let $c_i$ and $d_i$ be the endvertices of $\alpha_i$ in $C$ and $D$ respectively. Consider the cycle:
\[C'=c_1,C-u,c_2,\alpha_2,d_2,D-v,d_1,\alpha_1,c_1.\]
As the connected components of $G-A$ are trees, $C'$ meets $A$. But $u,v\notin V(C')$. Hence $V(C')\cap A=\{w\}$. So $V(C')-w$ is entirely contained in a connected component of $G-A$, which, therefore, contains the two neighbors of $w$ in $C'$, a contradiction.
\end{proof}
\begin{lemma}\label{prism-creterion}
Suppose that $G$ is a $3$-connected graph, $uv$ is an edge of $G$ and $G-\{u,v\}$ is $2$-connected. Then $G$ has a prism-minor containing $uv$ if and only if $G$ has vertex-disjoint cycles $C$ and $D$ such that $\{u,v\}\cont V(C)\cup V(D)$.
\end{lemma}
\begin{proof}
Suppose that $G$ has such cycles $C$ and $D$. If both $u$ and $v$ are in one of these cycles, say $C$, then we may choose $C$ in such a way that $uv\in E(C)$; by applying Menger's Theorem on $G$ to obtain three vertex-disjoint $(V(C),V(D))$-paths, we get a prism-minor of $G$ using $e$. So, assume that $u\in C$ and $v\in D$. Now we apply Menger's Theorem on $G-\{u,v\}$ to obtain two $(V(C)-u,V(D)-v)$ vertex disjoint-paths that, together with $u,v$, are three vertex-disjoint $(V(C),V(D))$-paths. So, $G$ has a prism-minor using $uv$ in all cases.
Conversely, suppose that $G$ has a prism minor using $uv$. If $G$ is not $uv$-Dirac the result follows from Menger's Theorem. So, we may assume that $G$ is $uv$-Dirac and, by Lemma \ref{prism-menger}, $G$ has vertex-disjoint cycles $C$ and $D$ and vertex-disjoint paths $(V(C),V(D))$-paths $\alpha_1$, $\alpha_2$, and $\alpha_3$ such that $uv\in E(\alpha_3)$. Make the choice of these cycles and paths minimizing $|\alpha_3|$. If $\{u,v\}\cont V(C)\cup V(D)$, we have nothing to prove. So we may assume that $\alpha_3$ has an inner vertex in $\{u,v\}$. Let $x$ and $y$ be the endvertices of $\alpha_3$ in $C$ and $D$ respectively. Let $X:=V(C)\cup V(D)\cup V(\alpha_1)\cup V(\alpha_2)$. As $G$ is $3$-connected, there is a $(V(\alpha_3)-\{x,y\},X)$ path $\beta$ in $G-\{x,y\}$. Let $a$ and $b$ be the endvertices of $\beta$, with $a\in V(\alpha_3)$. We may assume that the edge $uv$ is in the path $a,\alpha,y$. If $b\in V(D)\cup \int(\alpha_i)$ for some $i\in\{1,2\}$, then $D\cup\beta\cup(a,\alpha_3,y)\cup(a,\alpha_i,y)$ has a cycle containing $e$ and avoiding $V(C)$, a contradiction. So $b\in V(C)$. Let $c_1$ and $c_2$ be the respective endvertices of $\alpha_1$ and $\alpha_2$ in $C$. Now $C\cup\beta\cup(x,\alpha_3,a)$ has the cycle $C':=a,\alpha_3,x,C-c_1,b,\beta,a$. For some $\{i,j\}=\{1,2\}$, the cycles $C'$ and $D$ and the vertex-disjoint $(V(C'),V(D))$-paths $(x,C-c_j,c_i,\alpha_i)$, $(b,C-c_i,c_j,\alpha_j)$ and $a,\alpha_3,y$ have the property that $uv\in E(b,\alpha_3,y)$ and $|a,\alpha_3,y|<|\alpha_3|$, contradicting the minimality of $|\alpha_3|$.
\end{proof}
\begin{lemma}\label{prism c}
If $G$ is a type (c) $e$-Dirac graph with at least six vertices, then $G$ has a prism-minor using $e$ or $G\cong W_n$, $K_{3,n}$, $K'_{3,n}$, $K''_{3,n}$ or $K'''_{3,n}$ for some $n\ge 3$.
\end{lemma}
\begin{proof}\setcounter{rot}{0}
Suppose that the lemma fails for $G$ and let $e=u_1u_2$. By Theorem \ref{main-cor}, there are vertices $x$ and $y$ such that $N_G(u_1)=\{u_2,x,y\}$, $N_G(u_2)=\{u_1,x,y\}$, or $\{u_1,u_2,x,y\}$ is a two-sided vertex cut of $G$. Let us use the terminology of that theorem. First we check:
\begin{rot}\label{prism c-0}
$d_G(u_1)=3$ or $d_G(u_2)=3$.
\end{rot}
\begin{rotproof}
Suppose the contrary. For $i=1,2$, let $K^i_1,\dots,K^i_{n_i}$ be the $u_i$-components. As $d_G(u_1),d_g(u_2)\ge 4$, it follows that $n_1,n_2\ge 1$. If, for some $\{z_1,z_2\}=\{x,y\}$ and for each $i=1,2$, there is a cycle $C_i$ of $(K^i_1\cup\cdots\cup K^i_{n_i})+\{z_i,u_i\}$ containing $u_i$, then $C_1$ and $C_2$ contradict Lemma \ref{prism-creterion}. So, we may assume that $H:=(K^1_1\cup\cdots\cup K^1_{n_1})+\{y,u_1\}$ has no cycle containing $u_1$.
If $n_1\ge 2$, then $H$ has a cycle of the form $u_1,K^1_1,y,K^1_2,u_1$, a contradiction. So $n_1=1$.
If $u_1y\in E(G)$, $H$ has a cycle of the form $u_1,K^1_1,y,u_1$, a contradiction again. So $u_1y\notin E(G)$.
As $d_G(u_1)\ge 4$, there are two edges from $u_1$ to $K^1_1$. So $K^1_1+u_1$ has a cycle containing $u_1$, which is a cycle of $H$, another contradiction.
\end{rotproof}
By \ref{prism c-0}, we may assume that $N_G(u_1)=\{u_2,x,y\}$.
\begin{rot}\label{prism c-3}
There is exactly one $u_2$-component.
\end{rot}
\begin{rotproof}
Let $J_1,\dots, J_n$ be the distinct $u_2$-components with $n\ge 2$.
Let us prove first that each one fo these components have exatly one vertex. Say that $J_1$ has more than one vertex. By the description of Theorem \ref{main-cor}, $J_1+\{u_1,x_1,x_2\}$ is either a wheel with $u_2$ as hub or a fan with $u_2$ as hub and $x$ and $y$ as endvertices. In each of these cases, $J_1+u_2$ has a cycle $C$ containing $u_2$. This cycle, together with a cycle of the form $u_1,x,J_2,y,u_1$ contradict Lemma \ref{prism-creterion}. So each $u_2$-component has an unique vertex.
This implies that each vertex of $J_1\cup\cdots\cup J_n\cup\{u_1\}$ has $\{x,y,u_2\}$ as its neighborhood. So, $G\cong K_{3,n+1}$, $K'_{3,n+1}$, $K''_{3,n+1}$ or $K'''_{3,n+1}$, implying the lemma, a contradiction.
\end{rotproof}
Now, by the description of Theorem \ref{main-cor}, $N_G(u_1)=\{x,y,u_2\}$ and $G-u_1$ is an $(u_2,x,y)$-rope bridge. We will denote by $\rho_x$ and $\rho_y$ its ropes. Instead of the ropes, we will argue using the paths $\pi_x:=\rho_x,x,u_1$ and $\pi_y:=\rho_y,y,u_1$ because they have a certain symmetry regarding $u_2$ and $u_1$. Let $\pi_x=u_2,x_1,\dots,x_{m_x},u_2$ and $\pi_y:=u_1,y_1,\dots,y_{m_y},u_1$. For a step $\sigma$ and $k\in\{1,2\}$, we denote by $C_k(\sigma)$ the cycle $u_k,\pi_x,x_\sigma,\sigma,y_\sigma,u_k$.
\begin{rot}\label{prism c-5}
Each pair of steps either cross or have a common endvertex.
\end{rot}
\begin{rotproof}
If the claim fails, there are steps $\alpha$ and $\beta$ such that $u_1$, $z_\alpha$, $z_\beta$ and $u_2$ appear in this order in $\pi_z$ for each $z\in \{x,y\}$. Using Lemma \ref{prism-creterion} for the cycles $C_1(\alpha)$ and $C_2(\beta)$ we conclude that $G$ has a prism minor using $e$, a contradiction.
\end{rotproof}
\begin{rot}\label{prism c-6}
Each step has at most one inner vertex.
\end{rot}
\begin{rotproof}
Suppose that the claim fails and suppose that $\alpha:=v_1,\dots,v_n$ is a step with $n\ge 4$, $v_1\in V(\pi_x)$ and $v_n\in V(\pi_y)$. If $\alpha$ is the unique step, then $G$ is a wheel with $u_2$ as hub and the lemma holds. So, there is another step $\beta$. Now the cycles $u_2,v_1,v_3,u_1$ and $C_1(\beta)$ yield the existence of a prism-minor of $G$ using $e$ by Lemma \ref{prism-creterion}, a contradiction.
\end{rotproof}
\begin{rot}\label{prism c-7}
Let $z\in\{x,y\}$ and let $z_k$ be an inner vertex of $\pi_z$. Suppose that $uz_k\in E(G)-E(\pi_z)$ or a step with an inner vertex contains $z_k$. Then no step has an endvertex in $\{z_{k+1},\dots,z_{m_z}\}$.
\end{rot}
\begin{rotproof}
Suppose that the claim fails. Say that $z=x$. So, there is a step $\beta$ arriving in $x_l$ for some $k<l\le m_x$. If $ux_k\in E(G)-E(\pi_x)$, the cycles $u,\pi_x,z_k,u$ and $C_1(\beta)$ contradict Lemma \ref{prism-creterion} since $G$ has no prism-minor using $e$. So assume that a step $\alpha$ with an inner vertex $w$ has $z_k$ as endvertex. Now $u_2,\pi_x,z_k,\alpha,w,u_2$ and $C_1(\beta)$ contradict Lemma \ref{prism-creterion}.
\end{rotproof}
\begin{rot}\label{prism c-8}
Each pair of steps have a common vertex.
\end{rot}
\begin{rotproof}
By \ref{prism c-6}, it suffices to prove that no pair of steps cross. Suppose that $\alpha$ and $\beta$ are crossing steps. It follows from \ref{prism c-7} that these steps have no inner vertices. So, we may assume that there are indices $1\le a<c \le m_x$ and $1\le b\le d\le m_y$ such that $\alpha=x_a,y_c$ and $\beta=y_b,x_d$.
Let us check that $c=a+1$. Suppose for a contradiction that there is an index $a<k<c$. As each step crosses at most one other step by (RB3), then no step has endvertex in $x_k$. Moreover, by \ref{prism c-7}, $ux_k\notin E(G)$. This implies that the degree of $x_k$ is two, a contradiction. So, $c=a+1$ and, analogously, $d=b+1$.
Let us check now that $a=1$. Suppose that $a>1$. As $d_G(x_1)\ge 3$, there is a step $\gamma$ with $x_1$ as endvertex. As each step crosses at most one other step, $\gamma$ does not cross $\alpha$ nor $\beta$. So, $\gamma$ intersects both $\alpha$ and $\beta$ by \ref{prism c-5}. But, for this to happen, it is necessary that $\gamma$ share a common endvertex in $\pi_y$ with both $\alpha$ and $\beta$, a contradiction. So $a=1$. Analogously, $b=1$.
Let us check that $m_x=2$. Suppose that $m_x\ge 3$. By (RB6) $ux_3\notin E(G)$. So, there is a step $\gamma$ arriving at $x_3$. As argued in the previous paragraph, $\gamma$ share a common vertex in $\pi_y$ with both $\alpha$ and $\beta$, a contradiction. Therefore, $m_x=2$ and, analogously, $m_y=2$.
If there are no other steps than $\alpha$ and $\beta$, then $G\cong K_{3,3}$, so there are other steps. As no other steps cross $\alpha$ nor $\beta$, each other step has endvertices in $\{x_1,y_1\}$ or $\{x_2,y_2\}$.
First suppose that there is a step $\gamma$ with endvertices in $\{x_1,y_1\}$. By \ref{prism c-7}, $\gamma$ may not have inner steps and this establishes the uniqueness of $\gamma$. By \ref{prism c-5}, there is no steps with endevertices in $\{x_2,y_2\}$. So $\alpha$, $\beta$ and $\gamma$ are the unique steps. By (RB6), $u_2x_2,u_2y_2\notin E(G)$ and, therefore, $G\cong K'_{3,3}$, a contradiction.
So, all steps differing from $\alpha$ and $\beta$ have endvertices in $\{x_2,y_2\}$. Note that $N_G(u_1)=N_G(x_1)=N_G(y_1)=\{u_2,x_2,y_2\}$. If $w\in V(G)-\{u_1,u_2,x_1,x_2,y_1,y_2\}$, then $w$ is an inner vertex of a step with endvertices in $\{x_2,y_2\}$. By \ref{prism c-6}, $N_G(w)=\{u_2,x_2,y_2\}$. this implies that each vertex out of $\{u_2,x_2,y_2\}$ have this set as neighborhood and $G\cong K_{3,n}$, $K'_{3,n}$, $K''_{3,n}$ or $K'''_{3,n}$ for some $n\ge 3$; but this implies the lemma.
\end{rotproof}
\begin{rot}\label{prism c-9}
$x_1,y_1$ is the unique step with $x_1$ and $y_1$ as endvertices, but not the unique step.
\end{rot}
\begin{rotproof}
First we prove that there is a step $\beta$ with $x_1$ and $y_1$ as endvertices. Indeed, as $d_G(x_1), d_G(y_1)\ge 3$, there must be a step $\alpha$ containing $x_1$ and a step $\beta$ containing $y_1$. We may assume that $\alpha$ contains a vertex $y_k$ with $k>1$. If the endvertex of $\beta$ in $\pi_x$ is not $x_1$, then $\alpha$ and $\beta$ cross, contradicting \ref{prism c-8}. So there is a step $\beta$ with $x_1$ and $y_1$ as endvertices.
Now let us prove the uniqueness of $\beta$. Suppose that $\gamma$ is a second step with endvertices in $x_1$ and $y_1$. As $G$ has no parallel edges, one of $\gamma$ or $\beta$ has an inner vertex. By \ref{prism c-7} no step meets $Z:=\{x_2,\dots,x_m,y_2,\dots,y_n\}$. By (RB6) there is no edge joining $u_2$ and a vertex of $Z$. Thus the vertices of $Z$ have degree two and $Z$ must be empty. This implies that each vertex $z \in V(G)-\{u_1,u_2,x_1,y_1\}$ is an inner vertex of some step with $x_1$ and $y_1$ as endvertices; by \ref{prism c-6} $N_G(z)=\{u_2,x_1,y_1\}$ and, as $N_G(u_1)=\{u_2,x_1,y_1\}$, this implies that $G\cong K_{3,3}$, $K'_{3,3}$, $K''_{3,3}$ or $K'''_{3,3}$, a contradiction. Thus the uniqueness of $\beta$ as a step with endvertices in $\{x_1,y_1\}$ is established.
Let us prove that $\beta$ is not the unique step. Assume the contrary. This implies that no step has an endvertex in the set $Z$ defined in the last paragraph. Thus $u_2z\in E(G)$ for all $z\in Z$. This implies that $N_G(u_2)=V(G)-u_2$. But the unique edges not incident to $u_2$ are those in $E(\pi_x)-u_2x_1$, $E(\pi_y)-u_2y_1$ or $E(\beta)$. But $u_1,\pi_x,x_1,\beta,y_1,\pi_y,u_1$ induces a cycle in $G$, and, therefore, $G$ is a wheel with $u_2$ as hub, a contradiction.
As there is a second step $\gamma$, by the uniqueness of $\beta$, $\gamma$ arrives at a vertex of $Z$, then by \ref{prism c-7}, $\beta$ has no inner vertex and the claim holds.
\end{rotproof}
Now we finish the proof. By \ref{prism c-9}, there is a step $\alpha\neq x_1,y_1$. By \ref{prism c-8} and by the uniqueness of $x_1y_1$ established in \ref{prism c-9}, we may assume that the endvertices of $\alpha$ are $y_1$ and $x_k$ for some $2\le k\le m_x$. Choose $\alpha$ with $k$ as small as possible.
By \ref{prism c-8}, each step must intersect $\alpha$ and $x_1,y_1$. As $x_1\notin V(\alpha)$, each steps contains $y_1$. Therefore, for $2\le l \le n$, there is no step arriving in $y_l$. Moreover, by (RB6), $u_2y_l\notin E(G)$. So $d_g(y_l)=2$ and such an index $l$ may not exist. So, $m_y=1$.
Let us check that $k=2$. Suppose for a contradiction that there is an index $1<l<k$. By \ref{prism c-7}, $u_2x_l\notin E(G)$ and, therefore there is a step $\beta$ arriving at $x_l$. But $y_1$ must be an endvertex of $\beta$, thus $\beta$ and $l$ contradict the minimality of $\alpha$ and $k$. So, $k=2$
As $|V(G)|\ge 6$, $m_x\ge 3$. Now we check, for each $l=1,\dots m_x$, that $u_2x_l\notin E(G)$ and that each step arriving at $x_l$ has no inner vertices. Indeed, for $l\ge 3$, this follows from (RB6). For $l\le 2$, this follows from \ref{prism c-7}. In particular this implies that $x_ly_1$ is the unique edge of $G$ out of $\pi_x$ incident to $x_l$. Now $G$ is a wheel with $y_1$ as hub. This proves the lemma.
\end{proof}
Theorem \ref{thm-prism} now follows from Lemmas \ref{prism a}, \ref{prism b} and \ref{prism c}.
|
\section{Introduction}
It is shown in \cite[Theorem 3.1]{CKS} that an infinite field of positive characteristic that is definable
in an NTP$_2$ theory has only finitely many Artin-Schreier extensions. NTP$_2$ is a large class of first order
theories (``without the tree property of the second kind'') defined by S.~Shelah generalizing simple and NIP
theories. Algebraic examples of NTP$_2$ structures are given by ultraproducts of $p$-adic fields and certain
valued difference fields. For the precise definition (which we will not need in this paper) and for further
examples, we refer the reader to \cite{CKS}.
In \cite[Proposition~3.2]{CKS} it is shown that the value group of a valued field of characteristic $p>0$
which has only finitely many Artin-Schreier extensions is $p$-divisible. In this note, assuming throughout that
the valuation is nontrivial, we prove a stronger result,
namely, that such a field is dense in its perfect hull. For a valued field of characteristic $p>0$,
this is equivalent to being a deeply ramified field in the sense of \cite{GR}. The density also
implies that the value group is $p$-divisible and the residue field is perfect. By
Theorem~\ref{charst} in Section~\ref{sectdr},
every deeply ramified field of positive characteristic is a semitame field, which we shall define now.
Take a valued field $(K,v)$ of arbitrary characteristic. Its value group will be denoted by $vK$, and
its residue field by $Kv$. Accordingly, the value of an element $a\in K$ will be denoted by $va$, and its residue
by $av$. We say that $(K,v)$ is \bfind{semitame} if either $\chara Kv=0$, or $\chara Kv=p>0$ and the following two
axioms hold:
\sn
{\bf (DRst)} the value group $vK$ is $p$-divisible,
\sn
{\bf (DRvr)} the homomorphism
\[
\cO_{K^c}/p\cO_{K^c} \ni x\mapsto x^p\in \cO_{K^c}/p\cO_{K^c}
\]
is surjective, where $\cO_{K^c}$ denotes the valuation ring of the completion $K^c$ of $(K,v)$.
Note that this condition implies that the residue field $Kv$ is perfect.
\pars
The following result will be proven in Section~\ref{sectcharppf}:
\begin{theorem} \label{MTfinAS}
Take a valued field $(K,v)$ of characteristic $p>0$. If $K$ admits only finitely many distinct Artin-Schreier
extensions, then $(K,v)$ is a semitame field.
\end{theorem}
\begin{corollary}
A nontrivially valued field of positive characteristic that is definable in an NTP$_2$ theory is a semitame field.
\end{corollary}
We will prove Theorem~\ref{MTfinAS} by showing that if $(K,v)$ is not dense in its perfect hull, then it admits
infinitely many distinct Artin-Schreier extensions; see Proposition~\ref{propnotdense}.
However, we will show more than this. We are interested in Galois \bfind{defect extensions} of degree $p$ a prime.
These are immediate Galois extensions $(L|K,v)$ of degree $p$ of valued fields for which $v$ has a unique
extension from $K$ to $L$ (in this case we must have that $p=\chara Kv$). Here, $(L|K,v)$
denotes an extension $L|K$ of fields with $v$ a valuation on $L$ and $K$ endowed with its restriction; the
extension is said to be \bfind{immediate} if the canonical embeddings of $vK$ in $vL$ and of $Kv$ in $Lv$ are
onto. For more details on the defect see Section~\ref{sectfi}, and for further background, see
\cite{BKdrf,Ku1,Ku6}.
In \cite{BKdrf} a classification of Galois defect extensions $\cE=(L|K,v)$ of prime degree $p$ is given as follows.
We show that the set
\[
\Sigma_\sigma\>:=\> \left\{ v\left( \left.\frac{\sigma f-f}{f}\right) \right| \, f\in L^{\times} \right\}
\]
is independent of the choice of a generator $\sigma$ of $\Gal (L|K)$, and we
denote it by $\Sigma_\cE\,$. We say that $\cE$ has \bfind{independent defect} if
\[
\Sigma_{\cE}\>=\> \{\alpha\in vK\mid \alpha >H_\cE\}
\]
for some proper convex subgroup $H_\cE$ of $vK$; otherwise we say that $\cE$ has \bfind{dependent defect}. If
$vK$ is archimedean ordered, or in other words, $(K,v)$ is of \bfind{rank $1$}, then $H_\cE$ can only be equal
to $\{0\}$.
\pars
In Section~\ref{sectcharppf}, we will prove:
\begin{theorem} \label{MTdepAS}
Take a valued field $(K,v)$ of characteristic $p>0$. If it admits an Artin-Schreier extension with dependent
defect, then it admits infinitely many distinct Artin-Schreier extensions with dependent defect.
\end{theorem}
In \cite{Ku6} the classification was originally introduced only for valued fields of positive characteristic.
An Artin-Schreier defect extension $(L|K,v)$ was said to have dependent defect if in a certain way it
depends on immediate purely inseparable extensions of degree $p$ which do not lie in the completion of $(K,v)$.
It was then shown in Section~4.2 of \cite{Ku6} that an Artin-Schreier extension has dependent defect if it is
obtained from such a purely inseparable defect extension by a certain transformation of an inseparable minimal
polynomial which makes it separable. We will describe this transformation in Section~\ref{secttransf} and use it
for the proof of Theorem~\ref{MTdepAS}.
In \cite{BKdrf} we have made the above more precise by showing that {\it every} Artin-Schreier extension with
dependent defect can be obtained by such a transformation. It is also shown that the new definition of the
classification given in \cite{BKdrf} is compatible with the one given in \cite{Ku6}.
\pars
The new definition of the classification became necessary in order to generalize the original definition to the
case of valued fields $(K,v)$ of characteristic $0$ with residue field $Kv$ of positive characteristic $p$
(\bfind{mixed characteristic}), where we cannot rely on nontrivial purely inseparable extensions. We will now
present a partial analogue of Theorem~\ref{MTdepAS} in the mixed characteristic case. To
avoid a number of technical details in the present paper, we will restrict our scope to valued fields $(K,v)$
of rank $1$. We will also assume that $K$ contains a primitive $p$-th root of
unity. We then consider Kummer defect extensions of degree $p$, that is, Galois defect extensions of $(K,v)$
generated by some $\eta\notin K$ such that $\eta^p\in K$. Because such extensions are immediate,
we can show that we can assume $\eta$ to be a \bfind{$1$-unit}, i.e., $v(\eta-1)>0$ (and thus, $v\eta=0$); see
Section~\ref{sectK}.
We need some more preparation. Take an extension $(L|K,v)$ and $a\in L$. We define:
\[
v(a-K)\>:=\>\{v(a-c)\mid c\in K\}\>;
\]
for details on this set, see Section~\ref{sectva-K}.
For $\alpha$ in the divisible hull $\widetilde{vK}$ of the value group $vK$, we will write $v(a-K)<\alpha$
if $v(a-c)<\alpha$ for all $c\in K$, and similarly for ``$\leq$'' in place of ``$<$''. We write
\[
v(a-K)<\!\!|\,\alpha
\]
if $v(a-K)<\alpha$ and $v(a-K)$ is bounded away from $\alpha$ in $\widetilde{vK}$, i.e.,
there is $\beta\in \widetilde{vK}$ such that $v(a-K)\leq\beta<\alpha$.
A Kummer defect extension of degree $p$ generated by a $1$-unit $\eta$ as above always satisfies
\begin{equation} \label{Kde}
v(\eta-K)\><\>\frac{vp}{p-1}\>,
\end{equation}
see Section~\ref{sectK}. In the case of archimedean value groups (where $H_\cE$ cn only be equal to $\{0\}$), the
definition we have given for the extension to have dependent defect is equivalent to the condition that
$v(\eta-K)<\!\!|\, \frac{vp}{p-1}$ (see our discussion at the end of Section~\ref{sectK}).
Note that in \cite{BKdrf} in place of $v(\eta-K)$ a more flexible invariant $\dist(\eta,K)$ is used for our work
with defect extensions. We will give its definition in Section~\ref{sectdist} to enable the reader to relate the results in \cite{BKdrf} and \cite{Ku6} to those in the present paper.
\pars
In view of Theorem~\ref{MTdepAS} it seems reasonable to conjecture that if $(K,v)$ admits a Kummer extension
with dependent defect, then it admits infinitely many distinct Kummer extensions with dependent defect. By
\cite[part 2) of Theorem 1.5]{BKdrf}, this would imply that if $(K,v)$ admits only finitely many Kummer defect
extensions, then it is semitame, in perfect analogy to the equal positive characteristic case. However, so far we
have only been able to prove a weaker result. We will say that a Kummer defect extension as above has
\bfind{super-dependent defect} if
\begin{equation} \label{sd}
v(\eta-K)<\!\!|\, \frac{vp}{p}\>.
\end{equation}
\begin{theorem} \label{MTmixed}
Take a valued field $(K,v)$ of mixed characteristic and rank 1 which contains a primitive $p$-th root of
unity. If it admits a Kummer extension of degree $p$ with super-dependent defect, then it admits infinitely
many distinct Kummer extensions of degree $p$ with super-dependent defect.
\end{theorem}
\n
This leads us to the following
\n
{\bf Open questions:} Does the above theorem also hold with dependent defect in place of super-dependent defect?
What is (possibly) special about a Kummer extension with dependent defect that is not super-dependent?
\sn
We have observed in earlier work already that the threshold $\frac{vp}{p}$ plays a certain role when dealing
with 1-units in mixed characteristic (see
\cite[Corollary 2.11 d)]{Ku8}), but it is not yet sufficiently understood what exactly this role is.
\parm
In the last section of this paper we extend our study of Artin-Schreier extensions of valued fields in order to
fill a gap that occurred in the proof of Lemma 2.31 of \cite{Ku6}, which gives a criterion for such extensions to
have nontrivial defect.
\parm
For general background on valuation theory, we refer the reader to \cite{[En],[EP],[R],[ZS2]}.
\section{Preliminaries} \label{sectprel}
\subsection{Deeply ramified and semitame fields} \label{sectdr}
For a field $K$ of characteristic $p>0$ we denote by $K^{1/p^\infty}$ the perfect hull
of $K$. Further, we set $K^p=\{a^p\mid a\in K\}$ and $K^{1/p}=\{a^{1/p}\mid a\in K\}$; then $K^p$ is a subfield
of $K$, and $K^{1/p}|K$ is a field extension which is trivial if and only if $K$ is perfect.
\pars
In \cite{BKdrf}, the following is proven:
\begin{theorem} \label{charst}
Take a nontrivially valued field $(K,v)$ of characteristic $p>0$. Then the following statements are equivalent:
\n
a) $(K,v)$ is a semitame field,
\n
b) $(K,v)$ is a deeply ramified field,
\n
c) $(K,v)$ satisfies (DRvr),
\n
d) the completion of $(K,v)$ is perfect,
\n
e) $(K,v)$ is dense in its perfect hull, i.e., $K^{1/p^\infty}\subset K^c$,
\n
f) $K^p$ is dense in $(K,v)$.
\end{theorem}
Here, the property ``dense'' refers to the topology induced by the valuation.
\pars
See \cite{BKdrf} for more details and the connection of semitame and deeply ramified fields with the
classification of defect extensions.
\subsection{The set $v(a-K)$} \label{sectva-K}
Take a totally ordered set $(T,<)$. For a nonempty subset $S$ of $T$ and an element $t\in T$ we will write $S<t$ if
$s<t$ for every $s\in S$. A set $S\subseteq T$ is called an \bfind{initial segment} of $T$ if for each
$s\in S$ every $t <s$ also lies in $S$. Similarly, $S\subseteq T$ is called
a \bfind{final segment} of $T$ if for each $s\in S$ every $t >s$ also lies in $S$.
If $(T,<)$ is an ordered abelian group, $n\in\N$, $\alpha\in T$, and $S$ is an initial segment of $T$, then
\[
nS+\alpha\>:=\>\{n\beta+\alpha\mid \beta\in S\}
\]
is an initial segment of $nT$.
\pars
Take a valued field extension $(K(a)|K,v)$. We will now collect various properties of the sets $v(a-K)$ and
$v(a-K)\cap vK$.
\begin{proposition} \label{propva-K}
Take a valued field extension $(K(a)|K,v)$.
\sn
1) If $(K(a)|K,v)$ is immediate, then $v(a-K)$ has no largest element.
\sn
2) If $v(a-K)$ has no largest element, then $v(a-K)\subseteq vK$.
\sn
3) The set $v(a-K)\cap vK$ is an initial segment of $vK$.
\sn
4) The set $v(a-K)\setminus (v(a-K)\cap vK)$ has at most one element.
\sn
5) For every $c,d\in K$,
\[
v(da+c-K) \>=\> v(a-K)+vd\>.
\]
\sn
6) If $a,b$ are elements in some valued field extension of $(K,v)$ such that $v(a-b)>v(a-K)$, then
$v(a-c)=v(b-c)$ for all $c\in K$ and $v(a-K)=v(b-K)$.
\end{proposition}
\begin{proof}
1): This follows from \cite[Theorem 1]{[Ka]}.
\sn
2): Take $c\in K$; we wish to show that $v(a-c)\in vK$. By assumption there is $d\in K$ such that $v(a-d)>v(a-c)$.
Hence $v(a-c)=\min\{v(a-c),v(a-d)\}=v(c-d)\in vK$.
\sn
3): Take $\alpha\in v(a-K)\cap vK$ and $\beta\in vK$ such that $\beta<\alpha$. Choose $b,c\in K$ with $vb=\beta$
and $v(a-c)=\alpha$. Then $c-b\in K$ and $vb=\min\{v(a-c),vb\}=v(a-(c-b))\in v(a-K)$.
\sn
4): Assume that $v(a-c),v(a-d)\in v(a-K)\setminus (v(a-K)\cap vK)$. If they were distinct, say $v(a-d)>v(a-c)$,
then as in the proof of 2) we would obtain that $v(a-c)\in vK$, contradiction.
\sn
5): This follows from the equalities $v(a+c-K)=v(a-K)$ and $v(da-K)=v(a-K)+vd$, which are straightforward to prove.
\sn
6): For all $c\in K$, from $v(a-b)>v(a-c)$ we obtain that $v(a-c)=v(b-c)$ as in the proof of 2); hence $v(a-c)\in
v(b-K)$ and $v(b-c)\in v(a-K)$, which shows that $v(a-K)=v(b-K)$.
\end{proof}
We will mostly work with immediate extensions, in which case we have that $v(a-K)\subseteq vK(a)=vK$ for every element $a\notin K$ in the extension.
However, several of our results hold more generally for extensions that are not necessarily immediate. Some of them, such as the following one, remain true when the set $v(a-K)$ is replaced by $v(a-K)\cap vK$.
\begin{lemma} \label{distinct}
Take a valued field extension $(K(a)|K,v)$. Take $\alpha\in v(a-K)$ and assume that $d\in K$ is such that
\begin{equation} \label{alph+vd}
\alpha+vd\> >\> v(a-K)
\end{equation}
(note that $d$ always exists when $v(a-K)$ is bounded by some element from $vK$). Then the sets
\[
v(a-K)-vd^n\>, \;n\in\N\>,
\]
are pairwise distinct.
\end{lemma}
\begin{proof}
Since $\alpha\in v(a-K)$ and $\alpha+vd>v(a-K)$, we know that $vd>0$. Take any natural numbers
$m<n$. Then
\[
\alpha-mvd\> \geq \> \alpha +vd-nvd \> >\> v(a-K)-nvd\>,
\]
which shows that $\alpha-mvd\notin v(a-K)-nvd=v(a-K)-vd^n$. But $\alpha-mvd\in
v(a-K)-mvd=v(a-K)-vd^m$, so the two sets are distinct.
\end{proof}
\subsection{Distances} \label{sectdist}
Take again a totally ordered set $(T,<)$.
A pair $(\Lambda^L,\Lambda^R)$ of subsets of $T$ is called a \bfind{cut} in $T$ if $\Lambda^L$ is an initial
segment of $T$ and $\Lambda^R=T\setminus \Lambda^L$; it then follows that $\Lambda^R$ is a final segment of $T$.
To compare cuts in $(T,<)$ we will use the lower cut sets comparison. That is, for two cuts
$\Lambda_1=(\Lambda_1^L,\Lambda_1^R) ,\, \Lambda_2=(\Lambda_2^L,\Lambda_2^R)$ in $T$ we will write
$\Lambda_1<\Lambda_2$ if $\Lambda^L_1\varsubsetneq\Lambda^L_2$, and $\Lambda_1\leq \Lambda_2$ if
$\Lambda^L_1 \subseteq \Lambda^L_2$. This defines a linear order on the set of all cuts in $T$.
For a given subset $S$ of $T$ we define $S^+$ to be the cut $(\Lambda^L,\Lambda^R)$ in $T$ such that
$\Lambda^L$ is the least initial segment containing $S$, that is,
\[
S^+\>:=\> (\{t\in T\,|\, \exists\, s\in S:\,t\leq s\}\,,\, \{t\in T\,|\, t>S\} )\>.
\]
Likewise, we denote by $S^-$ the cut $(\Lambda^L,\Lambda^R)$ in $T$ such that $\Lambda^L$ is the largest initial
segment disjoint from $S$, i.e.,
\[
S^-\>:=\> (\{t\in T\,|\, t<S\}\,,\, \{t\in T\,|\, \exists\, s\in S:\,t\geq s\} )\>.
\]
For $s\in T$, we set
\[
s^+\>:=\> \{s\}^+\quad\mbox{ and }\quad s^-\>:=\> \{s\}^-\>.
\]
We note that
\begin{equation} \label{S+<s-}
S^+\>\leq\>s^- \>\Leftrightarrow\> S<s \;\mbox{ and }\;
S^+\><\>s^- \>\Leftrightarrow\> \exists\, t\in T:\, S\leq t<s\>.
\end{equation}
Indeed, $S^+\leq s^-$ means that the smallest initial segment of $T$ containing $S$ is a subset of the initial
segment $\{t\in T\,|\, t<s\}$, and this is equivalent to $S<s$. Likewise, $S^+< s^-$ means that the smallest
initial segment of $T$ containing $S$ is a proper
subset of $\{t\in T\,|\, t<s\}$, and as both are initial segments, this is equivalent to the existence of some
$t\in T$ such that $S\leq t$.
\pars
We embed $T$ in the linearly ordered set of all cuts in $T$ by identifying each $s\in T$ with $s^-$.
\pars
The set $v(a-K)\cap vK$ is an initial segment of $vK$ and thus the lower cut set of a cut in $vK$. However,
in order to be able to compare $v(a-K)$ with $v(a-L)$ when $L|K$ is algebraic, it is more convenient to work
with the cut $v(a-K)$ induces in the divisible hull $\widetilde{vK}$ of $vK$. Indeed,
$\widetilde{vK}$ is equal to the value group of the algebraic closure $K\ac$ of $K$, and if $L|K$ is algebraic,
then $K\ac=L\ac$ and therefore, $\widetilde{vK}=\widetilde{vL}$. Hence we define:
\[
\dist (a,K)\>:=\>(v(a-K)\cap \widetilde{vK})^+\;\; \textrm{ in the divisible hull $\widetilde{vK}$ of $vK\>$.}
\]
We call this cut the \bfind{distance of $z$ from $K$}. The distance replaces the use of suprema in the case of
non-archimedean value groups, which are not contained in $\R$.
\begin{lemma} \label{translate}
Let the situation be as above and assume that $v(a-K)\subseteq \widetilde{vK}$. Take some $\alpha\in
\widetilde{vK}$. Then $\dist (a,K)\leq\alpha^-$ is equivalent to $v(a-K)<\alpha$, and
$\dist (a,K)<\alpha^-$ is equivalent to $v(a-K)<\!\!|\, \alpha$.
\end{lemma}
\begin{proof}
We have:
\begin{eqnarray*}
\dist (a,K)\>\leq\>\alpha^-&\Leftrightarrow& v(a-K)\cap \widetilde{vK}<\alpha \\
&\Leftrightarrow& v(a-K)<\alpha\>,
\end{eqnarray*}
where the first equivalence holds by (\ref{S+<s-}) and the second equivalence holds since $v(a-K)\subseteq \widetilde{vK}$. Similarly,
\begin{eqnarray*}
\dist (a,K)\><\>\alpha^-&\Leftrightarrow& \exists\,\beta\in\widetilde{vK}:\, v(a-K)\cap
\widetilde{vK}\leq\beta<\alpha \\
&\Leftrightarrow& \exists\,\beta\in\widetilde{vK}:\, v(a-K)\leq\beta<\alpha
\>\Leftrightarrow\> v(a-K)<\!\!|\, \alpha\>.
\end{eqnarray*}
\end{proof}
\bn
\subsection{The fundamental equality and the defect} \label{sectfi}
Take an extension $(L|K,v)$ of valued fields such that the extension of $v$ from $K$ to $L$ is unique. Assume that
$\chara Kv=p>0$. Then
\begin{equation} \label{feuniq}
[L:K]\>=\> p^{\nu }\cdot(vL:vK)[Lv:Kv]\>,
\end{equation}
where by the Lemma of Ostrowski $\nu$ is a nonnegative integer (see \cite[Th\'eor\`eme 2, p.~236]{[R]}
or \cite[Corollary to Theorem 25, p.~78]{[ZS2]}). The factor
$d(L|K,v)=p^{\nu }$ is called the \bfind{defect} of the extension $(L|K,v)$.
\bn
\subsection{Criteria for defect extensions} \label{sectde}
\begin{lemma} \label{ueGp1}
Take an extension $(K(a)|K,v)$ of valued fields of degree
$p=\mbox{\rm char}(Kv)$ and such that the extension of $v$ from $K$ to $K(a)$ is unique.
If $v(a-K)$ has no maximal element, then $(K(a)|K,v)$ is immediate with defect~$p$.
\end{lemma}
\begin{proof}
By \cite[part (1) of Theorem 2.21]{Ku6}, $(K(a)|K,v)$ is immediate. As the extension of $v$ from $K$ to $K(a)$
is assumed to be unique, it follows that the defect is equal to the degree of the extension.
\end{proof}
\pars
\begin{lemma} \label{c1}\label{c2}
Assume that $(K(a)|K,v)$ is a normal extension of prime degree $p$.
\sn
1) If $K^h$ is some henselization of $(K,v)$ and $a\notin K^h$, then the extension of $v$ from $K$ to $K(a)$ is
unique.
\sn
2) Assume that $(K,v)$ is of rank 1. If $v(a-K)$ has no maximal element and is bounded in $vK$, then $(K(a)|K,v)$
is immediate and has defect $p$.
\end{lemma}
\begin{proof}
1): Since $K(a)|K$ is normal and of degree $p$, it is linearly disjoint from every other algebraic extension $L$
over $K$ in which it is not contained. This is seen as follows. If $K(a)|K$ is inseparable, then it is purely
inseparable, and so is $L(a)|L$; hence the latter extension can only be of degree $1$ or $p$. If $K(a)|K$ is
separable, then it is Galois with a Galois group of order $p$. Then also $L(a)|L$ is Galois. As restriction embeds
its Galois group in that of $K(a)|K$, again the degree of $L(a)|L$ can only be $1$ or $p$.
Since $a\notin K^h$ by assumption, from what we have just shown we see that $K(a)$ must be linearly disjoint from $K^h$ over $K$, which by
\cite[Lemma~2.1]{BK1} implies our assertion.
\sn
2): Since $(K,v)$ is of rank 1, $K$ lies dense in its henselization $K^h$. By assumption, $v(a-K)$ is bounded in
$vK$, which implies that $a\notin K^h$ and thus by part 1), the extension of $v$ from $K$ to $K(a)$ is unique.
Now our assertion follows from Lemma~\ref{ueGp1}.
\end{proof}
\bn
\subsection{Artin-Schreier extensions of valued fields} \label{sectAS}
Throughout this section, we let $K$ be a field of characteristic $p>0$.
Recall that $L|K$ is called an \bfind{Artin-Schreier
extension} if it is generated by a root of a polynomial of the form $X^p-X-b$ with $b\in K$. Note that every
Artin-Schreier extension of $K$ is a Galois extension of degree $p$, and vice versa.
We call $\vartheta$ an \bfind{Artin-Schreier generator} of an Artin--Schreier extension $L|K$ if $L=K(\vartheta)$
with $\vartheta^p-\vartheta\in K$.
\begin{lemma} \label{allASg}
Assume that $\vartheta$ is an Artin-Schreier generator of an Artin--Schreier extension $L|K$.
Then $\vartheta'$ is another Artin-Schreier generator
of $L|K$ if and only if $\vartheta' =i\vartheta +c$ for some $i\in\F_p$ and $c\in K$.
\end{lemma}
\begin{proof}
If $\vartheta$ and $\vartheta'$ are roots of the same polynomial $X^p-X-b$,
then $\vartheta-\vartheta'$ is a root of $X^p-X$, whose roots are exactly the elements of
$\F_p\,$. Hence, $\vartheta+i$, $i\in\F_p\,$, are all roots of $X^p-X-b$. Pick a nontrivial $\sigma\in
\Gal L|K$. We then have that $\sigma\vartheta-\vartheta=j$ for some $j\in\F_p^\times$.
If $\vartheta'$ is another Artin-Schreier generator of $L|K$ such that $\sigma\vartheta-\vartheta=j=
\sigma\vartheta'-\vartheta'$, then we have $\sigma(\vartheta-\vartheta')
=\vartheta-\vartheta'$. Since $\sigma$ is a generator of $\Gal L|K\isom
\Z/p\Z$, it follows that $\tau(\vartheta-\vartheta')=\vartheta-
\vartheta'$ for all $\tau\in\Gal L|K$, that is, $\vartheta-\vartheta'\in K$.
If $\vartheta'$ is another Artin-Schreier generator of $L|K$ and
$\sigma\vartheta'-\vartheta'=j'\in\F_p^\times$, then there is some $i\in\F_p^\times$
such that $ij=j'$. Since $\sigma i=i$, we then have that $\sigma i\vartheta-i\vartheta=i(\sigma\vartheta-\vartheta)
=ij=j'$. Then by what we have shown before, $\vartheta'=i\vartheta+c$ for some $c\in K$.
Conversely, if $\vartheta$ is an Artin-Schreier generator of $L|K$ and
if $i\in\F_p^\times$ and $c\in K$, then $(i\vartheta+c)^p-(i
\vartheta+c)=i(\vartheta^p-\vartheta)+c^p-c\in K$. But $i\vartheta+c$
cannot lie in $K$, so $K(i\vartheta+c)=L$ since $[L:K]$ is a prime. This
shows that also $i\vartheta+c$ is an Artin-Schreier generator of $L|K$.
\end{proof}
\begin{corollary} \label{ASinv}
Let $(L|K,v)$ be an Artin-Schreier extension of valued fields. Then
$v(\vartheta-K)$ is independent of the choice of the Artin-Schreier generator
$\vartheta$ of the extension, so it is an invariant of the extension.
\end{corollary}
\begin{proof}
Take two Artin-Schreier generators $\vartheta,\vartheta'$ of $L|K$. By Lemma~\ref{allASg} we can write
$\vartheta' =i\vartheta +c$ for some $i\in\F_p^\times$ and $c\in K$. Since $vi=0$ for every valuation, it follows
from assertion 5) of Proposition~\ref{propva-K} that $v(\vartheta'-K)=v(\vartheta-K)$.
\end{proof}
\begin{proposition} \label{propve-th}
Assume that $(K(\eta)|K,v)$ is a purely inseparable extension and $(K(\vartheta)|K,v)$ is an
Artin-Schreier extension of valued fields. Then $K(\eta,\vartheta)|K(\vartheta)$ is purely inseparable
and so the extension of $v$ from $K(\vartheta)$ to $K(\eta,\vartheta)$ is unique. If
\begin{equation} \label{ve-th}
v(\eta-\vartheta)\> > \> v(\vartheta-K)\>,
\end{equation}
then the extension of $v$ from $K$ to $K(\vartheta)$ is unique. If in addition, $(K(\eta)|K,v)$ is immediate,
then $(K(\vartheta)|K,v)$ is immediate with defect~$p$.
\end{proposition}
\begin{proof}
Assume that (\ref{ve-th}) holds. Then $\vartheta\notin K^h$ since otherwise it would follow
from Theorem~2 of~\cite{Ku4}
that $K(\eta)|K$ is not purely inseparable, contradicting our assumption. Now our first assertion follows from
part 1) of Lemma~\ref{c1}. By part 6) of Proposition~\ref{propva-K}, (\ref{ve-th}) also implies
that $v(\vartheta-K)=v(\eta-K)$.
\pars
Now assume in addition that $(K(\eta)|K,v)$ is immediate. Then by part 1) of Proposition~\ref{propva-K},
$v(\vartheta-K)=v(\eta-K)$ has no maximal element. Therefore, our second assertion follows from Lemma~\ref{ueGp1}.
\end{proof}
\subsection{Kummer extensions of prime degree of valued fields} \label{sectK}
Take a valued field $(K,v)$ of mixed characteristic, that is, $\chara K=0$ while $\chara Kv=p>0$. We consider
Kummer extensions of degree $p$. Such an extension is generated by an element $\eta$ such that $\eta^p\in K$.
If $(K(\eta)|K,v)$ is immediate, then it can be assumed that $\eta$ and hence also $\eta^p$ is a $1$-unit, i.e.,
$v(\eta-1)>0$. Indeed, since $(K(\eta)|K,v)$ is immediate, we have that $v\eta\in vK(\eta)=vK$,
so there is $c\in K$ such that $vc=-v\eta$. Then $v\eta c=0$, and since $\eta cv\in K(\eta)v=Kv$, there is
$d\in K$ such that $dv=(\eta cv)^{-1}$. Then $v(\eta cd)=0$ and $(\eta cd)v=1$. Hence $\eta cd$ is a $1$-unit.
Furthermore, $K(\eta cd)=K(\eta)$ and $(\eta cd)^p=\eta^pc^pd^p\in K$. Thus we can replace $\eta$ by $\eta cd$
and assume from the start that $\eta$ is a $1$-unit.
\pars
The next proposition follows from \cite[Corollary~3.6 and Proposition 3.7] {BKdrf}:
\begin{proposition} \label{dist_approx}
Take a valued field of mixed characteristic and a Kummer defect extension as detailed above.
Then the distance $\dist(\eta,K)$ does not depend on the choice of the generator $\eta$
of the extension $(K(\eta)|K,v)$ as long as it is a $1$-unit and satisfies $\eta^p\in K$. Moreover,
\begin{equation} \label{disteta1}
0\><\>\dist(\eta,K) \>\leq\> \left(\frac{vp}{p-1}\right)^-\>.
\end{equation}
\end{proposition}
Under the above conditions, the Kummer extension is immediate, hence we have that $v(\eta-K)\subseteq vK$. From the
definition of the distance it then follows that $v(\eta-K)$ is the intersection of the lower cut set of
$\dist(\eta,K)$ with $vK$, hence the distance determines uniquely the set $v(\eta-K)$, showing that this set does
not depend on the choice of the generator $\eta$. Moreover, from Lemma~\ref{translate} we obtain that
inequality (\ref{Kde}) holds.
\parm
In \cite[Proposition 3.7]{BKdrf} we show that a Kummer extension $(K(\eta)|K,v)$, where $\eta$ is a $1$-unit with
$\eta^p\in K$, has dependent defect if and only if
\[
\dist(a,K) \>=\> \frac{vp}{p-1} \,+\, H^-\>,
\]
for some proper convex subgroup $H$ of $\widetilde{vK}$. If $(K,v)$ has rank 1, then $\widetilde{vK}$ is
archimedean ordered and therefore $H=\{0\}$. In this case, the above equation just becomes
\[
\dist(\eta,K)\><\>\left(\frac{vp}{p-1}\right)^-\>.
\]
In view of Lemma~\ref{translate}, this translates to $v(\eta-K)<\!\!|\, \frac{vp}{p-1}$, as mentioned in the
Introduction.
\bn
\subsection{Differences of $p$-th powers in mixed characteristic} \label{sectdpp}
\begin{lemma} \label{lvep-ap}
Take $\eta$ and $a$ to be two elements of some valued field of characteristic 0 with residue field characteristic $p>0$. If
\begin{equation} \label{ve-a}
v(\eta-a)\><\> \frac{vp}{p-1}+v\eta\>,
\end{equation}
then
\begin{equation} \label{vep-ap}
v(\eta^p-a^p) \>=\> pv(\eta-a)\>.
\end{equation}
\end{lemma}
\begin{proof}
Let $\zeta$ be a primitive $p$-th root of unity. We have that
\[
\eta^p-a^p \>=\>\prod_{i=1}^p(\zeta^i\eta-a)\>=\>\prod_{i=1}^p(\zeta^i\eta-\eta+\eta-a)\>.
\]
Using the well known fact that
\[
v(\zeta^i-1)\>=\>\frac{vp}{p-1}
\]
(see e.g.\ the proof of Lemma~2.9 of~\cite{Ku8}) and our assumption, we obtain that
\[
v(\zeta^i-1)+v\eta\>=\> \frac{vp}{p-1}+v\eta \> >\> v(\eta-a)\>.
\]
Hence we have that
\[
v(\zeta^i\eta-a)\>=\> v(\zeta^i\eta-\eta+\eta-a)\>=\>\min\{v(\zeta^i-1)+v\eta, v(\eta-a)\}\>=\>v(\eta-a)\>,
\]
which yields equation (\ref{vep-ap}).
\end{proof}
\bn
\section{The case of valued fields of characteristic $p>0$} \label{sectcharp}
\subsection{A basic transformation} \label{secttransf}
We describe a transformation that was introduced in \cite{Ku6}. Take a valued field $(K,v)$ of characteristic
$p>0$ and a (not necessarily immediate) purely inseparable extension $(K(\eta)|K,v)$ of degree $p$ of valued
fields such that $\eta^p\in K$ and that $v(\eta-K)$ is bounded from above by an element in $vK(\eta)$ (and hence
also by an element in $vK$). We are starting with the minimal polynomial $Y^p-\eta^p$ of $\eta$ over $K$ and turn
it into the separable polynomial
\begin{equation} \label{trans1}
Y^p-d^{p-1}Y-\eta^p\>,
\end{equation}
where $d\ne 0$. Setting $Y=dX$ and then dividing the resulting polynomial by $d^p$, we transform this polynomial
into the Artin-Schreier polynomial
\begin{equation} \label{trans2}
X^p-X-\frac{\eta^p}{d^p}\>.
\end{equation}
Under the condition that $vd$ is large enough, the following lemma describes the behaviour of the set $v(\eta-K)$
when $\eta$ is replaced by roots of these two polynomials.
\begin{lemma} \label{dXeqc}
Take $d\in K$ such that
\begin{equation} \label{vd}
vd^{p-1}\> >\> pv(\eta-K) -v\eta\>.
\end{equation}
Let $\tilde\vartheta_d$ be a root of the polynomial (\ref{trans1}). Then
\[
\vartheta_d\>:=\> \frac{\tilde\vartheta_d}{d}
\]
is a root of (\ref{trans2}) and
$K(\vartheta_d)|K$ is an Artin-Schreier extension with a unique extension of $v$ from $K$ to
$K(\vartheta_d)$. Furthermore, we have that
\begin{equation} \label{dXeqceq1}
v(\tilde\vartheta_d-K)\>=\>v(\eta-K)\>,
\end{equation}
\begin{equation} \label{dXeqceq2}
v\left(\frac{\eta}{d}-\vartheta_d\right)\> >\> v(\vartheta_d-K)\>,
\end{equation}
and
\begin{equation} \label{vth-K}
v(\vartheta_d-K)\>=\> v(\eta-K)-vd\>.
\end{equation}
\end{lemma}
\begin{proof}
Once we prove that
\begin{equation} \label{dXeqceq3}
v(\eta-\tilde\vartheta_d)\> >\> v(\eta-K)\>,
\end{equation}
we obtain equation~(\ref{dXeqceq1}) by part 6) of Proposition~\ref{propva-K}, which in turn implies
equation~(\ref{vth-K}) by part 5) of Proposition~\ref{propva-K}. Further, again using part 5) of
Proposition~\ref{propva-K} again, we obtain
\[
v\left(\frac{\eta}{d}-\vartheta_d\right)\>=\>v(\eta-\tilde\vartheta_d)-vd\> >\> v(\eta-K)-vd\>=\>
v(\tilde\vartheta_d-K)-vd \>=\> v(\vartheta_d-K)\>,
\]
which proves (\ref{dXeqceq2}). Hence we will now prove (\ref{dXeqceq3}). We compute:
\begin{equation} \label{e-th}
(\eta-\tilde\vartheta_d)^p\>=\> \eta^p-\tilde\vartheta_d^p \>=\> \eta^p - \eta^p - d^{p-1}\tilde\vartheta_d \>=\>
- d^{p-1}\tilde\vartheta_d \>.
\end{equation}
\pars
We set $Y=dX$ to obtain that $\vartheta_d$ is a root of the polynomial $d^pX^p-d^{p-1} dX-\eta^p$ and hence also
of the Artin-Schreier polynomial (\ref{trans2}).
\pars
Since $v\eta\in v(\eta-K)$, from (\ref{vd}) we obtain:
\[
(p-1)vd \>=\> vd^{p-1}\> >\> pv\eta - v\eta = (p-1)v\eta\>,
\]
so that
\[
v\frac{\eta^p}{d^p} \>=\> p(v\eta-vd)\><\>0\>.
\]
Hence we have that $v\vartheta_d<0$ and consequently,
$v\vartheta_d^p=pv\vartheta_d<v\vartheta_d$ and
\[
pv\frac{\eta}{d}\>=\>v\frac{\eta^p}{d^p}\>=\>\min\{v\vartheta_d^p,v\vartheta_d\}\>=\>v\vartheta_d^p
\>=\>pv\vartheta_d\>,
\]
which yields that
\[
v\tilde\vartheta_d\>=\> vd+v\vartheta_d \>=\> v\eta\>.
\]
From this together with (\ref{vd}) and (\ref{e-th}) we obtain:
\[
v(\eta-\tilde\vartheta_d)\>=\> \frac{1}{p}v(d^{p-1}\tilde\vartheta_d)\>=\> \frac{1}{p}(vd^{p-1}+v\eta)
\> >\> v(\eta-K)\>,
\]
as desired.
\pars
It remains to prove that $K(\vartheta_d)|K$ is nontrivial (and hence an Artin-Schreier extension),
and that the extension of $v$ from $K$ to $K(\vartheta_d)$ is unique. Since $v(\eta-K)$ is bounded from above in
$vK$ by assumption, the same holds for $v(\eta-K)-vd$ which by (\ref{vth-K})
is equal to $v(\vartheta_d-K)$. This implies in particular that
$\vartheta_d\notin K$, so the extension $K(\vartheta_d)|K$ is nontrivial. Since $\eta$ is purely
inseparable over $K$ by assumption, the same holds for $\eta/d$. Thus our assertion follows from (\ref{dXeqceq2})
together with Proposition~\ref{propve-th}, where we replace $\eta$ by $\eta/d$ and $\vartheta$ by $\vartheta_d\,$.
\end{proof}
\bn
\subsection{Proof of Theorems~\ref{MTfinAS} and~\ref{MTdepAS}} \label{sectcharppf}
Throughout, let $(K,v)$ be a valued field of characteristic $p>0$. First, we prove:
\sn
{\it if $K$ admits only finitely many distinct Artin-Schreier extensions, then $(K,v)$ is a semitame field,}
\sn
which is the assertion of Theorem~\ref{MTfinAS}.
Take a purely inseparable (not necessarily immediate) extension
$K(\eta)|K$ of degree $p$ such that $\eta^p\in K$ and that $v(\eta-K)$ is bounded from above by an element in
$vK$.
\begin{proposition} \label{infASdn}
Assume that $\alpha\in v(\eta-K)$ and $d\in K$ are such that $\alpha+vd> v(\eta-K)$ (that is,
(\ref{alph+vd}) holds for $\eta$ in place of $a$) and $vd^{p-1}>pv(\eta-K) -v\eta$ (that is, (\ref{vd}) holds).
Then with the notation from Lemma~\ref{dXeqc}, the Artin-Schreier extensions
\[
K(\vartheta_{d^n})|K\>, \;n\in\N\>,
\]
are pairwise distinct, and the extensions of $v$ from $K$ to $K(\vartheta_{d^n})$ are unique for all $n\in\N$.
\end{proposition}
\begin{proof}
It follows from Lemma~\ref{dXeqc} that for each $n\in\N$, $K(\vartheta_{d^n})|K$ is an Artin-Schreier extension
and the extension of $v$ from $K$ to all $K(\vartheta_{d^n})$ is unique.
From Lemma~\ref{distinct} we infer that the sets $v(\vartheta_{d^n}-K)=v(\eta-K)-vd^n$ are distinct.
Thus by Corollary~\ref{ASinv}, the extensions $K(\vartheta_{d^n})|K$ are distinct.
\end{proof}
\begin{lemma} \label{pidegp}
If the perfect hull of $K$ does not lie in the completion
of $(K,v)$, then $(K,v)$ admits a purely inseparable extension $K(\eta)|K$ of degree $p$ such that
$v(\eta-K)$ is bounded from above by an element in $vK$.
\end{lemma}
\begin{proof}
Assume that the perfect hull $K^{1/p^{\infty}}$ does not lie in the completion $K^c$ of $(K,v)$, and take an
element
$\tilde\eta\in K^{1/p^{\infty}} \setminus K^c$. We may assume that $\tilde\eta^p\in K^c$ (otherwise, we replace
$\tilde \eta$ by ${\tilde\eta}^{p^{\nu}}$ for a suitable $\nu\geq 1$). Since $\tilde\eta\notin K^c$, we have that
$v(\tilde\eta-K)$ is bounded from above by some $\alpha\in vK$ and $v(\tilde\eta^p-K^p)= pv(\tilde\eta-K)$ is
bounded from above by $p\alpha$. On the other hand, since $\tilde{\eta}^p\in K^c$, there
is some $b\in K$ such that $v(\tilde\eta^p-b)>p\alpha$. We choose $\eta\in K^{1/p}$ such that $\eta^p=b$.
Then $pv(\tilde\eta-\eta)=v(\tilde\eta^p-\eta^p)=v(\tilde\eta^p-b)>p\alpha\geq pv(\tilde\eta-K)$, which yields that
$v(\tilde\eta-\eta)>v(\tilde\eta-K)$. From this it follows by part 6) of Proposition~\ref{propva-K}
that $v(\eta-K)=v(\tilde\eta-K)$, showing that also $v(\eta-K)$ is bounded from above by $\alpha$.
\end{proof}
\pars
Since $v(\eta-K)$ is assumed to be bounded from above by an element in $vK$, there is some $d\in K$ which satisfies
$vd^{p-1}>pv(\eta-K) -v\eta$.
Then the same is true for every $d'\in K$ with $vd'\geq
vd$. Hence $\alpha\in v(\eta-K)$ and $d\in K$ can be chosen such that $\alpha+vd> v(\eta-K)$.
Thus we can use Lemma~\ref{pidegp} together with Proposition~\ref{infASdn} to obtain:
\begin{proposition} \label{propnotdense}
If the perfect hull of $K$ does not lie in the completion
of $(K,v)$, then $K$ admits infinitely many Artin-Schreier extensions.
\end{proposition}
\sn
By the equivalence of assertions a) and e) of Theorem~\ref{charst}, this proposition proves Theorem~\ref{MTfinAS}.
\parm
We will now prove the assertion of Theorem~\ref{MTdepAS}, which states:
\sn
{\it if $(K,v)$ admits an Artin-Schreier extension with dependent
defect, then it admits infinitely many distinct Artin-Schreier extensions with dependent defect.}
\sn
We assume that $(K,v)$ admits an Artin-Schreier extension with dependent defect, which means that it must be
obtained via the transformation described in Section~\ref{secttransf} from a purely inseparable defect extension
$(K(\eta)|K,v)$ of degree $p$. This is only possible if $v(\eta-K)$ is bounded from above by an element in
$vK$. Also, since the extension is of degree $p$ with nontrivial defect, the defect must be $p$ and the extension
must be immediate. Hence by part 1) of Proposition~\ref{propva-K},
$v(\eta-K)$ has no largest element and consequently, the same holds for the sets $v(\eta-K)-vd^n=
v(\vartheta_{d^n}-K)$. Therefore, since the extension of $v$ from $K$ to $K(\vartheta_{d^n})$ is unique by
Proposition~\ref{infASdn}, it follows from Lemma~\ref{ueGp1} that $(K(\vartheta_{d^n})|K,v)$ is immediate with
defect $p$. As this extension is obtained from a purely inseparable defect extension of degree $p$ by the
transformation described in Section~\ref{secttransf}, this defect is dependent by definition. Finally,
Proposition~\ref{infASdn} shows that the extensions $K(\vartheta_{d^n})|K$, $n\in\N$, are distinct. This
completes our proof.
\bn
\section{The case of valued fields of mixed characteristic} \label{sectchar0}
Throughout this section, we assume that $(K,v)$ is a valued field of rank $1$ and of characteristic $0$ with
residue field of characteristic $p>0$. Further, we assume that $K$ contains a primitive $p$-th root of unity,
and that $(K(\eta)|K,v)$ is a Kummer extension, where $\eta$ is a $1$-unit with $\eta^p\in K$.
\subsection{A basic transformation in the mixed characteristic case} \label{secttransf0}
Given $d\in K$, we transform the minimal polynomial $X^p-\eta^p$ of $\eta$ over $K$ into the polynomial
\begin{equation} \label{trans3}
f_{d} (X)\>:=\> X^p+h_d(X)-\eta^p
\end{equation}
with
\[
h_d(X)\>:=\> \sum_{i=1}^{p-1} \binom{p}{i} d^{p-i}X^i\>.
\]
\begin{lemma} \label{btrf0}
Assume that $vd<0$ and that
\begin{equation} \label{ve-K}
v(\eta-K)\><\>\frac{vp+(p-1)vd}{p}\>.
\end{equation}
Take $\tilde\vartheta_d$ to be a root of the polynomial (\ref{trans3}). Then
\[
v(\eta-K)\>=\>v(\tilde\vartheta_d-K)\>.
\]
\end{lemma}
\begin{proof}
We compute the value of the coefficients of $h_d\,$, using our assumption (\ref{ve-K}):
\begin{equation}
v\left(\binom{p}{i} d^{p-i}\right) \>=\> vp+(p-i)vd\>\geq\> vp+(p-1)vd\> >\>pv(\eta-K)\>.
\end{equation}
Since $0=v(\eta-1)\in v(\eta-K)$, this shows that all coefficients of $h_d$ have positive value, while
$v\eta^p=0$; this forces $\tilde\vartheta_d$ to have value $0$ by the ultrametric triangle law. Consequently,
\begin{equation} \label{vtth>}
vh_d(\tilde\vartheta_d) \> >\> pv(\eta-K)\>.
\end{equation}
Suppose that there is $c\in K$ such that $v(\tilde\vartheta_d-\eta)\leq v(\eta-c)$. Combined with our assumption
(\ref{ve-K}), this implies that assumption (\ref{ve-a}) of Lemma~\ref{lvep-ap} holds for $\eta$ and for
$\tilde\vartheta_d$ in place of $a$ since $v\eta=0$. Hence by Lemma~\ref{lvep-ap} we have that
\[
v(\tilde\vartheta_d^p-\eta^p) \>=\> pv(\tilde\vartheta_d-\eta)\>.
\]
Using this together with (\ref{vtth>}) and the fact that
$\tilde\vartheta_d^p=\eta^p+h_d(\tilde\vartheta_d)$ by definition of $\tilde\vartheta_d$, we compute:
\begin{eqnarray*}
v(\tilde\vartheta_d-\eta) &=& \frac 1 p v(\tilde\vartheta_d^p-\eta^p) \>=\> \frac 1 p
v(\eta^p+h_d(\tilde\vartheta_d)-\eta^p)\\
&=& \frac 1 p vh_d(\tilde\vartheta_d)\> >\>v(\eta-K)\>,
\end{eqnarray*}
contradicting our assumption. This shows that $v(\tilde\vartheta_d-\eta)> v(\eta-c)$ for all $c\in K$, whence
$v(\tilde\vartheta_d-\eta)> v(\eta-K)$. Hence by part 6) of Proposition~\ref{propva-K},
$v(\eta-K)=v(\tilde\vartheta_d-K)$.
\end{proof}
\bn
\subsection{Proof of Theorem~\ref{MTmixed}} \label{sectchar0pf}
We let $(K,v)$ be a valued field of mixed characteristic and rank $1$ which contains a primitive $p$-th root of
unity. We have to prove the assertion of Theorem~\ref{MTmixed}, which states:
\sn
{\it If $(K,v)$ admits a Kummer extension of degree $p$ with super-dependent defect, then it admits infinitely
many distinct Kummer extensions of degree $p$ with super-dependent defect.}
\sn
We assume that $(K(\eta)|K,v)$ is a Kummer extension of degree $p$ with super-dependent defect, and that
$\eta$ is a $1$-unit with $\eta^p\in K$. Since $0=v\eta\in v(\eta-K)$ and $v(\eta-K)$ does not contain a
maximal element, it must contain positive elements. Since $(K(\eta)|K,v)$ is a super-dependent defect extension,
we know that (\ref{sd}) holds.
Since by assumption, $(K,v)$ is of rank $1$ and $v(\eta-K)$ does not contain a maximal element but is bounded,
$vK$ must be dense in $\widetilde{vK}$. Hence there is some $\td\in K$ such that $v\td<0$ and
\begin{equation} \label{ve-Kstrong}
v(\eta-K)\><\>\frac{vp}{p}+2v\td\>.
\end{equation}
Then inequality (\ref{ve-K}) of Lemma~\ref{btrf0} is satisfied, because
\[
\frac{vp}{p}+2v\td \><\>\frac{vp+(p-1)v\td}{p}\>.
\]
We obtain from Lemma~\ref{btrf0} that for a root $\tilde\vartheta_{\td}$ of the
polynomial $f_{\td}(X)$ defined in that lemma,
\[
v(\eta-K)\>=\>v(\tilde\vartheta_{\td}-K)\>.
\]
Now we set $X=\td Y$; dividing the resulting polynomial $f_{\td}(\td Y)$ by $\td^p$, we obtain the polynomial
\[
g_d(Y)\>:=\> Y^p + \sum_{i=2}^{p-1} \binom{p}{i} Y^i-\frac{\eta^p}{\td^p}\>.
\]
We observe that
\[
\vartheta_{\td}\>:=\> \frac{\tilde\vartheta_{\td}}{\td}
\]
is a root of $g_{\td}(Y)$ and that by part 5) of Proposition~\ref{propva-K},
\begin{equation}
v(\vartheta_{\td}-K)\>=\> v(\tilde\vartheta_{\td}-K) -v\td\>.
\end{equation}
Now we set $Y=X-1$, which turns the polynomial $g_{\td}(Y)$ into the polynomial
\[
X^p-\left(\frac{\eta^p}{\td^p} +1\right)
\]
with root
\[
\eta_{\td}\>:=\> \vartheta_{\td} +1\>.
\]
Again by part 5) of Proposition~\ref{propva-K}, we have that
\[
v(\eta_{\td}-K)\>=\>v(\vartheta_{\td}-K)\>=\> v(\tilde\vartheta_{\td}-K) -v\td\>=\> v(\eta-K) -v\td
\>\ne\> v(\eta-K)\>,
\]
where the last inequality holds since $v\td\ne 0$ and $v(\eta-K)$ is a bounded subset of an archimedean ordered abelian group. As $vd<0=v\eta$,
\[
\eta_{\td}^p\>=\> \frac{\eta^p}{\td^p} +1
\]
is a $1$-unit.
\pars
Since $vK$ is dense, there are infinitely many $\td'\in K$ with $v\td'<0$ that satisfy (\ref{ve-Kstrong})
in place of $\td$. With the same argument as above, we see that $v(\eta-K) -v\td\ne v(\eta-K) -v\td'$ if
$v\td\ne v\td'$. In this way we obtain infinitely many Kummer extensions $(K(\eta_{\td})|
K,v)$ with distinct sets $v(\eta_{\td}-K)$, which by Proposition~\ref{dist_approx}
shows that these extensions are pairwise distinct.
\pars
Since $(K(\eta)|K,v)$ is a nontrivial immediate extension, part 1) of Proposition~\ref{propva-K}
shows that the set $v(\eta-K)$ has no maximal element, while being bounded by assumption. The same is consequently
true for the sets $v(\eta-K)-v\td=v(\eta_{\td}-K)$. As the rank of $(K,v)$ is assumed to be $1$, part 2) of
Lemma~\ref{c2} shows that $(K(\eta_{\td})|K,v)$ is immediate and has defect $p$. Finally, this defect is
super-dependent since
\[
v(\eta_{\td}-K)\>=\> v(\eta-K)-v\td \><\>\frac{vp}{p}+2v\td-v\td\>=\>\frac{vp}{p}+v\td\>.
\]
\bn
\section{Filling a gap in \cite{Ku6}} \label{sectgap}
In \cite[Lemma 2.31]{Ku6} the following is stated:
\begin{lemma} \label{uniqextv}
Assume that $\chara K=p>0$ and $(K(\vartheta)|K,v)$ is an Artin-Schreier extension with Artin-Schreier generator
$\vartheta$. If $\dist (\vartheta,K)\leq 0^-$ and $v(\vartheta-K)$ has no maximal element, then the
extension of $v$ from $K$ to $K(\vartheta)$ is unique and $(K(\vartheta)|K,v)$ is immediate with defect $p$.
\end{lemma}
\n
In the proof it is written: ``In \cite{Ku4} we show that the assumption that $\dist (\vartheta,K)<0$ implies that
the extension of $v$ from $K$ to $K(\vartheta)$ is unique.'' Since $v(\vartheta-K)$ has no maximal element,
Lemma~\ref{ueGp1} shows that this assertion indeed implies that $(K(\vartheta)|K,v)$ is immediate. In addition,
$\dist (\vartheta,K)<\infty$ implies that $\vartheta\notin K$, so the extension is nontrivial and thus has defect
$p$.
However, the assertion was never proven in \cite{Ku4}. In order to complete the proof of Lemma~\ref{uniqextv}, we
will prove it here. As $v(\vartheta-K)$ has no maximal element, we know from part part 2) of
Proposition~\ref{propva-K} that $v(a-K)\subseteq vK\subseteq \widetilde{vK}$. Hence by Lemma~\ref{translate}
the assumption $\dist (\vartheta,K)\leq 0^-$ is equivalent to $v(\vartheta-K)<0$.
\pars
By part 1) of Lemma~\ref{c1}, it suffices to prove that $\vartheta\notin K^h$. Suppose
otherwise. Then by \cite[Theorem~1]{Ku4}, there are $\alpha\in vK$ and a convex subgroup $H\ne\{0\}$ of $vK$
such that the coset $\alpha+H$ is cofinal in $v(\vartheta-K)$. Since $v(\vartheta-K)<0$, we must have that
$\alpha+H<0$, which yields that $-\alpha>H$ and $-\alpha>0$. As $\alpha\in\alpha+H$, there is some $c\in K$
such that $v(\vartheta-c)\geq\alpha$. Since $\vartheta-c$ is also an Artin-Schreier generator of the extension and
$v(\vartheta-c-K)=v(\vartheta-K)$, we may assume w.l.o.g.\ that $v\vartheta\geq\alpha$.
Let $X^p-X-b$ with $b\in K$ be the minimal polynomial of $\vartheta$ over $K$, and take $d\in K$ with $vd=-\alpha$.
Then by part 5) of Proposition~\ref{propva-K}, $H$ is cofinal in $v(d\vartheta-K)$. Further,
$d\vartheta$ is a root of $d^{-p}X^p-d^{-1}X-b$ and hence also of
\[
X^p-d^{p-1}X-d^pb\>.
\]
Let $\eta$ be the root of $X^p-d^p b$. We wish to show that
\[
v(d\vartheta-K)\>=\>v(\eta-K)\>.
\]
We compute:
\begin{eqnarray*}
pv(d\vartheta-\eta) &=& v((d\vartheta)^p-\eta^p)\>=\>v(d^p\vartheta+d^pb-d^pb)\>=\>vd^p\vartheta\\
&\geq& -p\alpha +\alpha\>\geq\> -\alpha \> >\> H\>.
\end{eqnarray*}
Since $H$ is a convex subgroup of $vK$, this implies that
$v(d\vartheta-\eta) \>\geq\> -\frac 1 p \alpha \> >\> H\>$, and as $H$ is cofinal in $v(d\vartheta-K)$, we obtain:
\[
v(d\vartheta-\eta) \> >\> v(d\vartheta-K)\>.
\]
By construction, $K(\eta)|K$ is purely inseparable, hence by Theorem~2 of~\cite{Ku4},
$d\vartheta$ cannot lie in the henselization $K^h$. We have thus shown that $\vartheta\notin K^h$, as desired.
|
\section{Introduction}
The simplification of a pretrained neural network involves the elimination of neurons that might cause minor effectiveness loss after retraining (\cite{80236,lecun1990optimal}). These methods have been proposed for decades. When applied to the convolutional neural network (CNN) models, they are often referred to as kernel pruning methods. Such methods can speed up, simplify, and improve the explanation of the models~(\cite{he2017channel, huang2018learning, augasta2013pruning}), but their effectiveness is often below the original one~(\cite{luo2017thinet,Jordao:2019, li2016pruning}).
We have observed that kernel pruning methods usually retrain the entire model after each simplification action -- e.g., the elimination of one or a few kernels from a given layer. We call this strategy \emph{complete retraining}. It alters the weights of subsequent layers that have not been involved yet. Given the known problems of retraining deep neural networks, we presume that the accuracy loss in those simplified models is related to complete retraining and the criteria adopted for kernel elimination.
In this letter, we circumvent the problem by presenting kernel pruning methods that operate in a layer-by-layer fashion with effectiveness improvements. The methods include new strategies for (a) kernel elimination and (b) model retraining after each layer simplification. In (a), we present \emph{objective} and \emph{subjective} kernel relevance criteria. In the objective criterion, lower is the negative impact of a kernel in the cross-entropy loss function higher is its priority for removal. In the subjective criterion, the user (a CNN designer) visualizes a 2D projection ~(\cite{maaten2008visualizing}) of the \emph{mean activation maps} that the kernels generate for each class, and eliminates kernels that cannot separate classes in the projection. Each kernel of a given convolutional layer produces one activation map for each input image. The mean activation map of images from the same class indicates the most relevant regions activated by that kernel. We expect that a relevant kernel activates differently for at least one class. Therefore, a kernel remains in the model when at least one class appears separated from the others in the projection of its mean activation maps (Figure~\ref{f.subjective}). This subjective criterion explores the ability of humans to recognize visual patterns and improves the expert's understanding of the model simplification process. In (b), we present a new strategy, named \emph{progressive retraining}, in which the CNN model is retrained from the layer after the simplified one to layer one by backpropagation. Progressive retraining aims to recover the model's generalization capability before the simplification of the next convolutional layer.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=0.9\hsize]{Figs/subjective.pdf}
\end{center}
\caption{Hypothetical 2D projections~(\cite{maaten2008visualizing}) of the mean activation maps of three kernels (A, B, and C) for four color-coded classes. We remove kernel A and keep kernels B and C in the CNN model.}
\label{f.subjective}
\end{figure}
We first discuss related works and emphasize the importance of the above contributions (Section~\ref{sec:related_work}). Section~\ref{sec:methods} describes two new CNN simplification methods that operate in a layer-by-layer fashion based on the proposed kernel relevance criteria and progressive retraining. The experiments first evaluate the advantages of progressive retraining over complete retraining on four challenging image datasets (Section~\ref{sec:exper}). These advantages come with significant efficiency gains over a reference CNN model, Vgg-16~(\cite{vggNet}), and improvements in effectiveness --- a rare phenomenon in kernel pruning methods. We finally show the effectiveness gains of the proposed approaches over two popular kernel pruning methods~(\cite{li2016pruning,hu2016network}) and one approach from the state-of-the-art~(\cite{Jordao:2019}).
\section{Related works}
\label{sec:related_work}
Neural network simplification methods have been investigated for decades.~\cite{80236} estimates the sensitivity of the global error function when removing neurons, and prunes low sensitivity connections.~\cite{549081} propose a similar idea by estimating sensitivity at the output of the layer after the one under simplification.~\cite{POPPI1998187} adopt a special algorithm to measure the impact of removing a neuron in the training error.~\cite{737502} use an extended Kalman filter to measure the importance of a connection weight in a network.~\cite{augasta2013pruning} provide a comparative study of neuron pruning methods on real datasets.
The above methods have been focused on multi-layer perceptron (MLP) models. More recently, after the introduction of CNN models, new kernel pruning methods have appeared. They aim at simplifying
the CNN architecture by eliminating redundant kernels and/or kernels that can be removed with minor negative impact in accuracy. However, it has been difficult for those methods to obtain significant network simplifications with no accuracy loss
~(\cite{luo2017thinet,Jordao:2019, li2016pruning}).
According to~\cite{he2017channel}, recent advances in CNN acceleration can be divided into three categories: (1) optimized implementations~(\cite{vasilache2014fast}), (2) weight quantization~(\cite{10.1007/978-3-319-46493-0_32}), and (3) structured simplification~(\cite{jaderberg2014speeding}). Kernel pruning methods fall into category (3).~\cite{lecun1990optimal}, for instance, identify irrelevant connection weights based on their negligeable impact in accuracy.~\cite{li2016pruning} determine the importance of each kernel using the L1 norm of the weights and then kernels with lower values are pruned. By conducting simplification experiments on Vgg-16, ResNet34, ResNet56, and ResNet101, the authors report small losses in accuracy for reduction factors from $20\%$ to $30\%$ of the original number of kernels.~\cite{luo2017thinet} measure the impact of kernel removal at the output of the next layer to identify possible redundant kernels. Kernels that cause smaller changes in the total activation map of the next layer are selected for removal. They report a slight increase in the top-5 accuracy on the ImageNet dataset, when 50\% of the kernels in the first 10 layers of Vgg-16 are removed, and a slight decrease in the top-5 accuracy, when fully connected layers are replaced by general average pooling, which represents a considerable simplification of Vgg-16. Other kernel relevance criteria may be based on Taylor expansion~(\cite{DBLP:journals/corr/MolchanovTKAK16}) or the percentage of zero activations~(\cite{hu2016network}).
More complex simplification approaches have also been proposed. For instance,~\cite{huang2018learning} propose pruning agents to detect and eliminate unnecessary kernels. This pruning agent is modeled as a second CNN model which takes the kernel weights of the model under simplification as input and output binary decisions about removing or not kernels. The authors report a loss in accuracy of about 3\% after pruning the Vgg-16 and ResNet18 models using the CIFAR-10 dataset.~\cite{he2017channel} employ Lasso regression to determine redundant kernels for pruning and report slight accuracy losses with some simplified CNN models.
More recently,~\cite{Jordao:2019} estimates the importance of kernels for pruning based on the relationship between classes in a latent space, generated by partial least squares, and their variable importance in projection. They first eliminate kernels in the entire network and then retrain the model. They report considerable simplification without penalizing accuracy.
We present kernel pruning methods that simplify the convolutional layers of a CNN model in a layer-by-layer fashion. Differently from all previous approaches, we retrain the model progressively, without affecting the weights of subsequent layers that have not been involved yet. We also introduce a subjective criterion based on the visual analysis of data projections. By that, we would like to call attention for the importance of using visual analytics in the design of deep neural networks. The strategy seems promising to guide human actions in the machine learning loop, such that the resulting models are as simple and effective as possible for a given problem. In this context, there is a lack of interactive methods for the construction of deep neural networks.
Visual analytics have been successfully employed to understand neural networks and explain their decisions~(\cite{morch95,10.1007/978-3-540-24844-6_5,10.1007/978-3-319-10590-1_53,RauberTVCG2017,Hohman:2020}). It has also been used in the design of deep neural network models~(\cite{8019872,GarciaIJCNN19}).~\cite{8019872} present a system, named DeepEyes, to support the design of such models, and~\cite{GarciaIJCNN19} show a first method to build a simplified CNN model in a layer-by-layer fashion. They introduce a kernel selectivity criterion based on the visual analysis of the \emph{activation occurrence maps} for each class. That is, by counting the positive activations for input images from each class, the user can visualize those activation maps as 2D images and choose the kernels with different activation patterns among classes. The method is limited by the number of classes and the 2D dimension of the activation maps (2D input images). In the proposed subjective criterion, we use 2D projections to allow more classes and overcome the limitation about the dimension of the activation maps. However, we are still limited to reasonable numbers of classes, layers, and kernels per layer.
\section{Pruning with progressive retraining}
\label{sec:methods}
This section presents two approaches to simplify CNNs based on kernel relevance criteria followed by progressive retraining (fine-tuning) in a layer by layer fashion (Figure~\ref{f.PPR}). One may conceptually divide a CNN for image classification into three parts: (1) a sequence of convolutional layers, with each layer containing convolution, activation, and optional operations (e.g., pooling, batch normalization) useful for feature extraction; (2) fully connected layers for feature space reduction, which discover the neurons specialized in each class~(\cite{RauberTVCG2017}); and (3) a decision layer for final classification, being (2) and (3) known as a MLP classifier. We have applied the following procedures to simplify pre-trained CNNs based on the analysis of each convolutional layer's output.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=\hsize]{PPR-graphicalabstract.pdf}
\end{center}
\caption{Kernel pruning with progressive retraining using an objective/subjective criterion. Progressive retraining for a simplified layer $l$ starts from the output of layer $l+1$ to layer $1$ by backpropagation.}
\label{f.PPR}
\end{figure}
\subsection{Objective pruning}
The first approach, named \emph{objective Pruning with Progressive Retraining} (oPPR), eliminates kernels of a given convolutional layer $l$ based on an \emph{objective relevance criterion} and fine-tunes the network by backpropagation during $x$ epochs (e.g., $x=40$) from layer $l+1$ to layer $1$, aiming to recover its previous activations at the output of layer $l+1$. This process starts at layer $1$ and proceeds until the simplification of the last convolutional layer $L$, when the fine-tuning involves the weight optimization from the first fully connected layer $L+1$ to layer $1$.
In this approach, the relevance of a kernel $k$ is related to its impact on the neural network's loss function. We use cross-entropy as loss function --- higher is the loss, when removing a kernel $k$ from a layer $l$, more critical the kernel $k$ is for layer $l$. Given that, for a layer $l$, one can remove either a given number of the most irrelevant kernels or the kernels whose relevance is below a given threshold.
Once a set ${\cal M}_l\subset {\cal K}_l$ of kernels have been removed from the original kernel set ${\cal K}_l$ of layer $l$, we apply backprogation to refine the weights from layer $l+1$ to layer $1$. The loss function in this case is the mean square difference of the activation values obtained with and without ${\cal M}_l$ at the output of layer $l+1$. That is, let $I$ be an input image to the network, ${\cal X}$ be the set of training images, $A_{l+1}(I,{\cal K}_l)\in \Re^{n_{l+1}}$ be the output of layer $l+1$, after ReLu and, whenever the case, after pooling, with $n_{l+1}$ activation values, as a result from the input $I$, and $A_{l+1}(I,{\cal K}_l\backslash {\cal M}_l)\in \Re^{n_{l+1}}$ be the same without ${\cal M}_l$. The loss function for progressive retraining is defined as
\begin{eqnarray}
\bar{D}_{l+1} & = & \frac{1}{|{\cal X}|} \sum_{\forall I \in {\cal X}} \|A_{l+1}(I,{\cal K}_l) - A_{l+1}(I,{\cal K}_l\backslash {\cal M}_l)\|_2.
\end{eqnarray}
\subsection{Subjective pruning}
The second approach, named \emph{subjective Pruning with Progressive Retraining} (sPPR), works similarly to oPPR, except for the kernel relevance criterion. In this case, an expert in the design of neural networks visualizes 2D non-linear projections of \emph{mean activation maps} and decides which kernels are the most relevant, according to the expert's subjective judgment of the separation in the projection among the mean activation maps of the classes for each given kernel. One should expect that a relevant kernel can separate at least one class from the others (Figure~\ref{f.sPPR}).
\begin{figure}[hbt!]
\centering
\includegraphics[scale=0.35]{Figs/sPPR-pipeline.pdf}
\caption{sPPR: for each convolutional layer $l$, the expert analyzes the projections of the mean activation maps of the classes from each kernel, and eliminates kernels that form one cluster of points in the projection.}
\label{f.sPPR}
\end{figure}
Let ${\cal X}_j\subset {\cal X}$ be the set of training images from the class $j\in \{1,2,\dots,c\}$. Let $A_{l,k,j}(I)\in \Re^n$ be a map with $n < n_l$ activations at the output of layer $l$ as a consequence of using an input image $I$ from class $j$ and a kernel $k$ up to layer $l$. Note that these are the activations from the subset of the $n_l$ activations, which result from the convolution with kernel $k$ only. One mean activation map $\bar{A}_{l,k,j}$ from kernel $k$ and for each class $j$ can be obtained from the output of layer $l$ as
\begin{eqnarray}
\bar{A}_{l,k,j} & = & \frac{1}{|{\cal X}_j|}\sum_{\forall I\in {\cal X}_j}A_{l,k,j}(I).
\end{eqnarray}
We use the t-SNE algorithm~(\cite{maaten2008visualizing}) to project the mean activation maps $\bar{A}_{l,k,j}$ for all kernels $k$ and classes $j$ at the output of layer $l$ in the 2D space (Figure~\ref{f.kernelsel}a). Nevertheless, the expert analyzes the class projections of each kernel separately. For a reasonable number of kernels (e.g., 512), the user can quickly visualize only $c$ points in 2D per time from each kernel $k$ as an indication of how it discriminates among the images of the $c$ classes (Figures~\ref{f.kernelsel}b-\ref{f.kernelsel}d). If the points are very close to each other, forming one cluster in 2D, we may conclude that the kernel cannot discriminate between the classes. A kernel $k$ remains in layer $l$ when it can separate at least one class from the others. The separation between classes in the 2D projection is a subjective criterion that depends on the expert's visual analysis only.
\begin{figure}[ht!]
\begin{center}
\begin{tabular}{cc}
\fbox{\includegraphics[width=0.39\hsize]{Figs/EggsMeanActivation.png}} &
\fbox{\includegraphics[width=0.39\hsize]{Figs/EggsMeanActivation_kernel18_layer12_cropped.png}} \\
(a) & (b) \\
\fbox{\includegraphics[width=0.39\hsize]{Figs/EggsMeanActivation_kernel220_layer12_cropped.png}} &
\fbox{\includegraphics[width=0.39\hsize]{Figs/EggsMeanActivation_kernel6_layer12_cropped.png}} \\
(c) & (d)
\end{tabular}
\end{center}
\caption{For a layer with $512$ kernels and a dataset with $c=9$ classes. The t-SNE projection of the mean activation maps (a) from all kernels and (b-d) from three distinct kernels. The color indicates the class of images used to create the mean activation maps. The kernels in (b) and (c) are candidates for removal, while the kernel in (d) is a relevant one.}
\label{f.kernelsel}
\end{figure}
After eliminating those irrelevant kernels from layer $l$,
we apply backpropagation to refine the weights from layer $l+1$ to $1$, as described for the $oPPR$ approach. It is also possible to remove redundant kernels among the relevant ones. Redundant kernels have similar weights, so the projection of their weight vectors forms one cluster of points in 2D. Although this may further simplify the CNN, we have evaluated only the elimination of irrelevant kernels.
\subsection{After network simplification}
After simplification, the entire simplified network is retrained with $y$ epochs (e.g., $y=50$). This process starts with random weights in the fully connected layers (MLP classifier)~(\cite{glorot2010understanding}) but only fine-tunes the weights of the convolutional layers, which resulted from progressive retraining. We have observed that the progressive retraining can recover the generalization capability of the CNN better than retraining the entire network after each layer simplification, as recommended in most works in~(\cite{li2016pruning,hu2016network,he2017channel,Jordao:2019}).
\section{Experiments, results, and discussion}
\label{sec:exper}
To evaluate the simplification of CNNs using the $oPPR$ and $sPPR$ approaches, we divided the experiments into two parts. We use the Vgg-16~(\cite{vggNet}) model as reference, pre-trained on the ImageNet dataset~(\cite{imagenet_cvpr09}) and adjusted to the training set of each dataset below:
\begin{itemize}
\item EGG-9: it is a private dataset with images of the eight most common species of helminth eggs in Brazil and one diverse class of fecal impurities.
\item PRO-7: it is a private dataset with images of the six most common species of protozoa cysts in Brazil and one diverse class of fecal impurities.
\item LAR-2: it is a private dataset with images of S. stecoralis (helminth larvae) and fecal impurities.
\item STL-10: it is a well known public dataset with natural images, with considerable variations among objects that define a same class~(\cite{coates2011analysis}).
\end{itemize}
In each dataset of intestinal parasites, EGG-9, PRO-7, and LAR-2, fecal impurities contain examples that are very similar in texture and shape to those of other classes, which makes these datasets very challenging. These datasets belong to a project for the automation of the diagnosis of intestinal parasites~(\cite{SuzukiTBME2013, Suzuki:2013:ISBI, SaitoPR2015, PeixinhoVipImage2015}).
For more statistically reliable results, we divided each dataset into five random splits of training and testing images, respectively, as described in Table~\ref{t.dataset}. To fine-tune Vgg-16 in each dataset, we used 50 epochs, learning rate $1e^{-5}$, and reinitialized the weights of its fully connected layers using the method in~(\cite{glorot2010understanding}).
\begin{table}
\footnotesize
\caption{Number of classes, pixels and bands (resolution), training samples, and testing samples in each dataset. \label{t.dataset}}
\begin{tabular}{|l|r|r|r|r|r|}
\hline
Dataset & Number of & Image & Training & Testing & Total of \\
& Classes & Resolution & samples & samples & samples \\ \hline
EGG-9 & 9 & 200x200x3 & 1265 & 10156 & 11421 \\ \hline
LAR-2 & 2 & 200x200x3 & 478 & 960 & 1438 \\ \hline
PRO-7 & 7 & 200x200x3 & 3733 & 29902 & 33635 \\ \hline
STL-10 & 10 & 96x96x3 & 3900 & 7800 & 11700 \\ \hline
\end{tabular}
\end{table}
We divide the experiments into two parts and discuss the pros and cons of oPPR and sPPR afterwards.
\subsection{The impact of progressive retraining}
First, we evaluate the impact of the progressive retraining after each layer simplification by using variants of oPPR and sPPR, which adopt retraining of the entire network with a single epoch, as in other works~(\cite{li2016pruning,hu2016network,he2017channel,Jordao:2019}). For that, the variants of oPPR and sPPR with complete retraining (fine-tuning) after each layer simplification are named oPCR (\emph{Objective Pruning with Complete Retraining}) and sPCR (\emph{Subjective Pruning with Complete Retraining}), respectively. Table~\ref{t.comb} presents the four approaches compared in this first experimental section.
\begin{table}
\centering
\caption{The proposed approaches, oPPR and sPPR, are first compared with their variants, oPCR and sPCR, that adopt retraining of the entire network after each layer simplification, as recomended in~\cite{he2017channel,Jordao:2019}.}
\label{t.comb}
\begin{tabular}{|l|c|c|}
\hline
Criterion & \multicolumn{2}{c|}{Retraining} \\ \hline
& \multicolumn{1}{l|}{Complete} & \multicolumn{1}{l|}{Progressive} \\ \hline
Objective & oPCR & oPPR \\ \hline
Subjective & sPCR & sPPR \\ \hline
\end{tabular}
\end{table}
In oPPR and sPPR, we use $40$ epochs for progressive retraining and fixed learning rates along the epochs. To fine-tune from layer $l+1$ to layer $1$, the learning rate is fixed as described in Table~\ref{t.lr}, depending on the number of kernels of Vgg-16 at each convolutional layer $l$. Lower learning rates are used as higher is the number of kernels.
After network simplification, the fully convolutional layers (the MLP classifier) are reinitialized with random weights~(\cite{glorot2010understanding}) for retraining and the weights of the simplified convolutional layers are refined. This retraining process uses 50 epochs and learning rate $1e^{-5}$ in all approaches: oPPR, sPPR, oPCR, and sPCR.
\begin{table}
\centering
\caption{Learning rates for progressive retraining.}
\label{t.lr}
\begin{tabular}{cccc}
\hline
\multicolumn{1}{|r|}{Layer} & \multicolumn{1}{l|}{Name} & \multicolumn{1}{r|}{N. of Kernels} & \multicolumn{1}{r|}{Learn. rate} \\ \hline
\multicolumn{1}{|l|}{1} & \multicolumn{1}{l|}{block1\_conv\_1} & \multicolumn{1}{l|}{64} & \multicolumn{1}{l|}{$1e^{-5}$} \\ \hline
\multicolumn{1}{|l|}{2} & \multicolumn{1}{l|}{block1\_conv\_2} & \multicolumn{1}{l|}{64} & \multicolumn{1}{l|}{$1e^{-6}$} \\ \hline
\multicolumn{1}{|l|}{3} & \multicolumn{1}{l|}{block2\_conv1} & \multicolumn{1}{l|}{128} & \multicolumn{1}{l|}{$1e^{-6}$} \\ \hline
\multicolumn{1}{|l|}{4} & \multicolumn{1}{l|}{block2\_conv2} & \multicolumn{1}{l|}{128} & \multicolumn{1}{l|}{$1e^{-7}$} \\ \hline
\multicolumn{1}{|l|}{5} & \multicolumn{1}{l|}{block3\_conv1} & \multicolumn{1}{l|}{256} & \multicolumn{1}{l|}{$1^{e-7}$} \\ \hline
\multicolumn{1}{|l|}{6} & \multicolumn{1}{l|}{block3\_conv2} & \multicolumn{1}{l|}{256} & \multicolumn{1}{l|}{$1^{e-7}$} \\ \hline
\multicolumn{1}{|l|}{7} & \multicolumn{1}{l|}{block3\_conv3} & \multicolumn{1}{l|}{256} & \multicolumn{1}{l|}{$1e^{-8}$} \\ \hline
\multicolumn{1}{|l|}{8} & \multicolumn{1}{l|}{block4\_conv1} & \multicolumn{1}{l|}{512} & \multicolumn{1}{l|}{$1e^{-8}$} \\ \hline
\multicolumn{1}{|l|}{9} & \multicolumn{1}{l|}{block4\_conv2} & \multicolumn{1}{l|}{512} & \multicolumn{1}{l|}{$1e^{-8}$} \\ \hline
\multicolumn{1}{|l|}{10} & \multicolumn{1}{l|}{block4\_conv3} & \multicolumn{1}{l|}{512} & \multicolumn{1}{l|}{$1e^{-8}$} \\ \hline
\multicolumn{1}{|l|}{11} & \multicolumn{1}{l|}{block5\_conv1} & \multicolumn{1}{l|}{512} & \multicolumn{1}{l|}{$1e^{-8}$} \\ \hline
\multicolumn{1}{|l|}{12} & \multicolumn{1}{l|}{block5\_conv2} & \multicolumn{1}{l|}{512} & \multicolumn{1}{l|}{$1e^{-8}$} \\ \hline
\multicolumn{1}{|l|}{13} & \multicolumn{1}{l|}{block5\_conv3} & \multicolumn{1}{l|}{512} & \multicolumn{1}{l|}{$1e^{-8}$} \\ \hline
\end{tabular}
\end{table}
Tables~\ref{t.eggc}-\ref{t.stlc} present the results of comparison among the four approaches on the datasets EGG-9, PRO-7, LAR-2, and STL-10, respectively. One may observe that, except for two cases (oPPR in EGG-9 and sPPR in LAR-2), oPPR and sPPR could simplify Vgg-16 with some effectiveness gain (in accuracy and Cohen kappa) and considerable efficiency gain, as measured by the percentage of reduction in the number of kernels and GFLOPS. For the objective approaches, progressive retraining (oPPR) is consistently better than complete retraining (oPCR). In the subjective approaches, the expert is usually more restrained and eliminates considerably fewer kernels than the objective approaches. Nevertheless, the reduction in the number of kernels and GFLOPS with effectiveness gain is still significant in sPCR and sPPR.
\begin{table}
\scriptsize
\centering
\caption{Impact of the progressive retraining (oPPR and sPPR) after each layer simplification over complete retraining (oPCR and sPCR) and the original network (Vgg-16) on the EGG-9 dataset.}
\label{t.eggc}
\begin{tabular}{|l|l|l|l|l|}
\hline
& Accuracy & Cohen Kappa & GFLOPs & Kernel \\
& & & reduction (\%) & reduction (\%) \\ \hline
Vgg-16 & 97.65 $\pm$ 0.31 & 94.00 $\pm$ 0.79 & 0.00 & 0.00 \\\hline
oPCR & 96.90 $\pm$ 0.24 & 92.02 $\pm$ 1.09 & 74.42 & 50.00 \\\hline
oPPR & 97.62 $\pm$ 0.10 & 93.94 $\pm$ 0.26 & 74.42 & 50.00 \\\hline
sPCR & 97.83 $\pm$ 0.01 & 94.43 $\pm$ 0.05 & 27.33 & 9.42 \\\hline
sPPR & 97.79 $\pm$ 0.20 & 94.34 $\pm$ 0.53 & 54.53 & 21.44\\\hline
\end{tabular}
\end{table}
\begin{table}
\scriptsize
\centering
\caption{Impact of the progressive retraining (oPPR and sPPR) after each layer simplification over complete retraining (oPCR and sPCR) and the original network (Vgg-16) on the PRO-7 dataset.}
\label{t.proc}
\begin{tabular}{|l|l|l|l|l|}
\hline
& Accuracy & Cohen Kappa & GFLOPs & Kernel \\
& & & reduction (\%) & reduction (\%) \\ \hline
Vgg-16 & 96.74 $\pm$ 0.15 & 91.93 $\pm$ 0.36 & 0.00 & 0.00 \\\hline
oPCR & 95.93 $\pm$ 0.48 & 89.91 $\pm$ 1.15 & 74.42 & 50.00 \\\hline
oPPR & 97.11 $\pm$ 0.18 & 92.83 $\pm$ 0.41 & 74.42 & 50.00\\\hline
sPCR & 97.01 $\pm$ 0.37 & 92.61 $\pm$ 0.88 & 29.37 & 10.52 \\\hline
sPPR & 97.05 $\pm$ 0.21 & 92.68 $\pm$ 0.54 & 36.64 & 12.06\\\hline
\end{tabular}
\end{table}
\begin{table}
\scriptsize
\centering
\caption{Impact of the progressive retraining (oPPR and sPPR) after each layer simplification over complete retraining (oPCR and sPCR) and the original network (Vgg-16) on the LAR-2 dataset.}
\label{t.larc}
\begin{tabular}{|l|l|l|l|l|}
\hline
& Accuracy & Cohen Kappa & GFLOPs & Kernel \\
& & & reduction (\%) & reduction (\%) \\ \hline
Vgg-16 & 97.33 $\pm$ 0.81 & 89.65 $\pm$ 3.29 & 0.00 & 0.00\\\hline
oPCR & 96.24 $\pm$ 0.89 & 85.83 $\pm$ 2.70 & 74.42 & 50.00\\\hline
oPPR & 97.49 $\pm$ 0.50 & 90.38 $\pm$ 2.03 & 74.42 & 50.00\\\hline
sPCR & 97.39 $\pm$ 0.27 & 90.01 $\pm$ 1.08 & 38.07 & 16.41 \\\hline
sPPR & 97.22 $\pm$ 0.66 & 89.32 $\pm$ 2.87 & 63.93 & 27.82 \\\hline
\end{tabular}
\end{table}
\begin{table}
\scriptsize
\centering
\caption{Impact of the progressive retraining (oPPR and sPPR) after each layer simplification over complete retraining (oPCR and sPCR) and the original network (Vgg-16) on the STL-10 dataset.}
\label{t.stlc}
\begin{tabular}{|l|l|l|l|l|}
\hline
& Accuracy & Cohen Kappa & GFLOPs & Kernel \\
& & & reduction (\%) & reduction (\%) \\ \hline
Vgg-16 & 77.89 $\pm$ 2.13 & 75.43 $\pm$ 2.37 & 0.00 & 0.00\\\hline
oPCR & 73.68 $\pm$ 0.98 & 70.75 $\pm$ 1.09 & 74.25 & 50.00\\\hline
oPPR & 79.24 $\pm$ 1.74 & 76.93 $\pm$ 1.94 & 74.25 & 50.00 \\\hline
sPCR & 80.95 $\pm$ 2.15 & 78.83 $\pm$ 2.39 & 0.20 & 3.85 \\\hline
sPPR & 79.81 $\pm$ 1.33 & 77.57 $\pm$ 1.48 & 12.61 & 5.97 \\\hline
\end{tabular}
\end{table}
\subsection{Comparison with other methods}
In this section, we compare the proposed approaches (oPPR and sPPR) with other network simplification techniques. These techniques usually differ in two important aspects: (a) the relevance criterion to eliminate kernels and (b) the retraining strategy of the network after each layer simplification. We selected the methods in~\cite{li2016pruning},~\cite{hu2016network}, and~\cite{Jordao:2019} for our comparative analysis. In~\cite{li2016pruning}, the relevance of a kernel is measured by the sum of the absolute values of its weights. In~\cite{hu2016network}, the kernels that generate more outputs with no activation are considered less relevants. In~\cite{Jordao:2019}, the kernels are selected using partial least squares and variable importance in projection. For a fair comparison, the kernels were pruned using iterative pruning with 5 iterations of $10\%$ of pruning ratio, which reduces in $34.39\%$ the number of kernels in the convolutional layers. All baselines adopted their original protocols for kernel pruning followed by complete retraining. The first two network simplification approaches are popular baselines for several works.
In our case, we have fixed at $50\%$ the percentage of reduction in the number of kernels per layer for the objective approach, and repeated the results of the previous experiment for sPPR and the original Vgg-16, as references. For each layer, we eliminate 50\% of irrelevant kernels and apply progressive retraining.
At the end of the simplification process, all simplified models are retrained with $50$ epochs, fixed learning rate $1e^{-5}$, and reinitialized fully connected layers~(\cite{glorot2010understanding}).
Tables~\ref{t.acc50} and ~\ref{t.kappa50} show the results of this experiment on each dataset using accuracy and Cohen kappa as effectiveness measures, respectively. The results indicate that oPPR (and sPPR with less efficiency gain than oPPR) can be considerably more effective than the baselines, especially in Cohen kappa. This result is particularly important, because these datasets are unbalanced and Cohen kappa can penalize errors in small classes.
\begin{table}
\scriptsize
\caption{Mean accuracy with the elimination of $50\%$ of the kernels in the convolutional layers of Vgg-16. The best results are shown in bold.}
\label{t.acc50}
\begin{tabular}{|c|l|l|l|l|}
\hline
& EGG-9 & PRO-7 & LAR-2 & STL-10 \\ \hline
Vgg-16 & 97.65 $\pm$ 0.31 & 96.74 $\pm$ 0.15 & 97.33 $\pm$ 0.81 & 77.89 $\pm$ 2.13 \\ \hline
~\cite{li2016pruning} & 94.48 $\pm$ 0.66 & 93.29 $\pm$ 0.46 & 95.98 $\pm$ 0.66 & 74.54 $\pm$ 2.19 \\ \hline
~\cite{hu2016network} & 96.90 $\pm$ 0.21 & 96.03 $\pm$ 0.34 & 95.6 $\pm$ 0.81 & 73.55 $\pm$ 1.08 \\ \hline
~\cite{Jordao:2019} & 96.71 $\pm$ 0.47 & 94.83 $\pm$ 0.46 & 96.85 $\pm$ 0.30 & 75.89 $\pm$ 1.18 \\ \hline
oPPR & \textbf{97.62 $\pm$ 0.10} & \textbf{97.11 $\pm$ 0.18} & \textbf{97.49 $\pm$ 0.50} & \textbf{79.24 $\pm$ 1.74} \\ \hline
sPPR & \textbf{97.79 $\pm$ 0.20} & \textbf{97.05 $\pm$ 0.21} & \textbf{97.22 $\pm$ 0.66 } & \textbf{79.81 $\pm$ 1.33} \\ \hline
\end{tabular}-
\end{table}
\begin{table}
\scriptsize
\caption{Mean Cohen kappa with the elimination of $50\%$ of the kernels in the convolutional layers of Vgg-16.}
\label{t.kappa50}
\begin{tabular}{|c|l|l|l|l|}
\hline
& EGG-9 & PRO-7 & LAR-2 & STL-10 \\ \hline
Vgg-16 & 94.00 $\pm$ 0.78 & 91.93 $\pm$ 0.36 & 89.65 $\pm$ 3.29 & 75.43 $\pm$ 2.37 \\ \hline
~\cite{li2016pruning} & 85.51 $\pm$ 1.89 & 83.39 $\pm$ 0.83 & 84.30 $\pm$ 2.47 & 71.72 $\pm$ 2.44 \\ \hline
~\cite{hu2016network} & 92.01 $\pm$ 0.61 & 90.19 $\pm$ 0.80 & 83.26 $\pm$ 3.46 & 70.61 $\pm$ 1.20 \\ \hline
~\cite{Jordao:2019} & 91.61 $\pm$ 1.24 & 87.40 $\pm$ 0.98 & 87.84 $\pm$ 1.35 & 73.21 $\pm$ 1.32 \\ \hline
oPPR & \textbf{93.94 $\pm$ 0.26} & \textbf{92.83 $\pm$ 0.41} & \textbf{90.38 $\pm$ 2.03} & \textbf{76.93 $\pm$ 1.94} \\ \hline
sPPR & \textbf{94.34 $\pm$ 0.53} & \textbf{92.68 $\pm$ 0.54} & \textbf{89.32 $\pm$ 2.87} & \textbf{77.57 $\pm$ 1.48} \\ \hline
\end{tabular}
\end{table}
\subsection{Discussion}
The experiments have demonstrated that the proposed progressive retraining can provide higher effectiveness gains than the complete retraining usually adopted in network simplification~(\cite{li2016pruning,hu2016network,he2017channel,Jordao:2019}). When simplifying networks, it is typical for the simplified version to be less effective than the original one, as observed for the methods in~(\cite{li2016pruning,hu2016network,Jordao:2019}) in all datasets. On the other hand, oPPR and sPPR, consistently presented effectiveness gains over Vgg-16 (being oPPR in EGG-9 and sPPR in LAR-2, the exceptions for 50\% of kernel elimination). Given the considerable reduction in the number of kernels and GFLOPS, we may conclude that oPPR and sPPR are relevant contributions to the literature of network simplification.
Although oPPR, with a more aggressive kernel elimination, sometimes presented better performance than sPPR, the expert involvement in deep learning has several advantages. First, the expert can better understand the role of each element of the model and, perhaps, better explain its decisions~(\cite{RauberTVCG2017}). Second, the expert can intervene in the project of the network, as demonstrated by sPPR. Visual analytics techniques, such as t-SNE, play a central role in facilitating communication between machines and humans. They are known as essential tools to improve understanding of the machine learning process. However, we believe they can go beyond and let the user intervene to improve the machine learning process~(\cite{RauberInfoVis2018,BenatoSIBGRAPI2018}). In this sense, sPPR can improve with additional information about objective relevance measures to guide the expert's actions when eliminating kernels. A drawback in sPPR, however, is the limitation to reasonable numbers of classes, layers, and kernels per layer. On the other hand, oPPR completely relies on the objective relevance measure and eliminates a given and fixed percentage of kernels per layer. It seems that, with the expert in the process, some better criterion could be devised to eliminate an adaptive number of irrelevant kernels per layer.
\section{Conclusion}
We have presented two solutions for CNN simplification, named oPPR and sPPR, which explore objective and subjective kernel relevance criteria, respectively, and perform progressive retraining to adjust the model's weights by preserving the weights of subsequent layers not involved in the simplification process. Progressive retraining has shown improvements over complete retraining, usually adopted in kernel pruning methods. The proposed methods have achieved considerable network simplification with effectiveness gains over the original model. They can also be more effective than two popular methods and one approach from the state-of-the-art. We may then conclude that oPPR and sPPR are relevant contributions to the literature of kernel pruning methods.
We are now interested in investigating methods to construct CNNs in a layer-by-layer fashion by exploring objective and subjective kernel relevance criteria. In the context of the diagnosis of intestinal parasites, we intend to create more effective models with considerably simplified architectures.
\section{Acknowledgments}
The authors thank financial support from FAPESP (Proc. 2017/12974-0 and 2014/12236-1) and CNPq (Proc. 303808/2018-7).
\bibliographystyle{model2-names}
|
\section{Effect of Hyperparameters}
\subsection{SysAdmin}
We keep the number of agents fixed to $32$ for the following experiments.
\textbf{Exploration constant $c$}: Keeping number of Monte Carlo iterations fixed at $16000$ and exploration depth fixed at $20$, we varied the exploration constant between $5$ and $40$.
As can be seen from \Cref{fig:hyperparam_fig}, the difference is performance is not significant over the range.
\textbf{Exploration depth $d$}: Keeping the number of Monte Carlo iterations fixed at $16000$ and exploration constant fixed at $20$, we varied the tree search exploration depth.
We found that low exploration can adversely affect FV-MCTS-Var-El, but otherwise does not effect the performance very much.
\textbf{Number of Monte Carlo rollouts $n$}: Keeping the exploration constant fixed at $20$ and exploration depth fixed at $20$, we varied the number of Monte Carlo itertations used by the algorithms.
Low values adversely effected the performance a little.
\section{Algorithm Details}
For completeness we also also describe the FV-MCTS Var-El approach in~\Cref{algo:fvmctsvarel}. Note the difference in structure of tracked statistics.
\begin{algorithm}
\caption{FV-MCTS-Var-El}
\label{algo:fvmctsvarel}
\begin{algorithmic}[1]
\Require{{time limit, exploration depth, exploration constant}}
\Statex
\State $n_e, Q_e$ \Comment{Component statistics}
\State
\Function{FV-MCTS-Var-El}{$s$, depth}
\While{time limit not reached}
\State{\Call{Simulate}{$s$, depth}}
\EndWhile
\State{$\mathbf{a}^* \gets$ \Call{Var-El}{}}
\State{\Return{$\mathbf{a}^*$}}
\EndFunction
\Function{Simulate}{$s$, depth}
\If{$\gamma^{\mathrm{depth}} < \epsilon$}
\State{\Return $0$}
\EndIf
\State{$a\gets$ \Call{Var-El}{exploration constant}}
\State{$s', r \sim T(s,{a}), R(s, a)$}
\State{$q \gets r + \gamma$ \Call{Simulate}{$s'$, depth $- 1$}}
\State{\Call{UpdateStats}{$s, a, q$}}
\EndFunction
\Function{UpdateStats}{$s, a, q$}
\For{every component $e \in \mathcal{G}(s)$}
\State{$n_{e}(s, a_e) \mathrel{{+}{=}} 1$}
\State{$Q_{e}(s, a_e) \mathrel{{+}{=}} \frac{q - Q_{e}(s, a_e)}{n_{e}(s, a_e)} $}
\EndFor
\EndFunction
\end{algorithmic}
\end{algorithm}
\section{On the application to Multi-Agent POMDPs}
Strictly speaking, the prior work of~\citet{Amato2014-io} plans for Multi-Agent POMDPs (rather than MDPs), where the agents receive noisy observations of the true underlying state.
Given that our work is for MMDPs, we briefly comment on how the key challenges for both of us appear in the MMDP formulation itself and on how our approach could be applied directly to an MPOMDP. That is why we chose to use the MMDP for our implementation and evaluations rather than introducing the extraneous confound of a partially observable setting.
First, the prior work focuses almost entirely on the challenge of the exploding action space in a Monte Carlo Tree Search planning algorithm, which is just as pertinent for an MMDP as it is for an MPOMDP. To handle partial observability they use POMCP, a partially observable extension of MCTS (specifically, Upper Confidence Trees) in which particle filters track and update the belief state with new observations~\cite{DBLP:conf/nips/SilverV10}. Our approach could also be used straightforwardly with POMCP and applied to MPOMDPs.
Second, the only segment of the prior work that tries to address the unique challenge of an MPOMDP is a variant of their algorithm that splits joint observation histories into local histories and distributes them over the factors. In MMDPs there is no such thing as an observation history, since the current system state is fully observed; thus there is no equivalent variant to implement for the evaluation. Furthermore, for the MPOMDP case we could again incorporate our approach into their factorized trees variant directly, since our modifications would be to the action selection and exploration steps.
\begin{figure*}
\centering
\begin{subfigure}{\textwidth}
\centering
\begin{subfigure}{0.32\columnwidth}
\centering
\includegraphics[width=\columnwidth]{figures/star_sysadmin_depth_32.pdf}
\end{subfigure}
\begin{subfigure}{0.32\columnwidth}
\centering
\includegraphics[width=\columnwidth]{figures/star_sysadmin_niterations_32.pdf}
\end{subfigure}
\begin{subfigure}{0.32\columnwidth}
\centering
\includegraphics[width=\columnwidth]{figures/star_sysadmin_exploration_constant_32.pdf}
\end{subfigure}
\caption{Star Sysadmin}
\end{subfigure}
\begin{subfigure}{\textwidth}
\centering
\begin{subfigure}{0.32\columnwidth}
\centering
\includegraphics[width=\columnwidth]{figures/bi_sysadmin_depth_32.pdf}
\end{subfigure}
\begin{subfigure}{0.32\columnwidth}
\centering
\includegraphics[width=\columnwidth]{figures/bi_sysadmin_niterations_32.pdf}
\end{subfigure}
\begin{subfigure}{0.32\columnwidth}
\centering
\includegraphics[width=\columnwidth]{figures/bi_sysadmin_exploration_constant_32.pdf}
\end{subfigure}
\caption{Ring Sysadmin}
\end{subfigure}
\caption{Effect of hyperparameters. $32$ agents.}
\label{fig:hyperparam_fig}
\end{figure*}
\section{Conclusion}
\label{sec:conclusion}
We introduced a scalable online planning algorithm for multi-agent MDPs with dynamic coordination graphs. Our approach, FV-MCTS-MP, uses Max-Plus for action coordination, in contrast to the previously introduced FV-MCTS with Variable Elimination. Over the standard SysAdmin and the custom Multi-Drone Delivery domains, we demonstrated that FV-MCTS-MP performs as well as Var-El on static CGs, outperforms it significantly on dynamic CGs, and is far more computationally efficient (enabling online MMDP planning on previously intractable problems).
In the appendix, we discuss how our approach can easily extend to multi-agent POMDPs.
However, we still require a domain expert to pre-define the appropriate coordination graph for the problem. A predetermined CG can be particularly difficult with highly dynamic domains where the CG depends on the state.
More work is required towards learning the dynamic coordination graph itself via interaction with the model.
Similarly, extending ideas from Alpha-Zero~\citep{silver2018general,anthony2017thinking} would be particularly relevant for distilling the coordinated individual actions for the agents into decentralized policies~\citep{phan2019distributed} with FV-MCTS acting as a scalable policy improvement operator \citep{grill2020monte}.
\section{Introduction}
\label{sec:intro}
Coordination is crucial for effective decision-making in cooperative multi-agent systems with a shared objective.
Various real-world problems like formation control~\citep{oh2015survey}, package delivery~\citep{ChoudhurySoloveyETAL2020}, and firefighting~\cite{DBLP:conf/atal/OliehoekSWV08} require a team of autonomous agents to perform a common task.
Such cooperative sequential decision-making problems can be modeled as a multi-agent Markov decision process (MMDP)~\cite{boutilier1996planning}, an extension of the Markov decision process (MDP) \cite{kochenderfer2015decision}.
MMDPs can be reduced to centralized single-agent MDPs with a joint action space for all agents.
Such reductions often make large problems intractable because the action space grows exponentially with the number of agents. Solving independent MDPs for all agents yields suboptimal behavior in problems where reasoning about the effects of joint actions is necessary for better performance~\citep{matignon2012}.
Many previous MMDP approaches have tried to balance these extremes of optimality and efficiency.
In the \emph{offline} setting, these include ad hoc function decomposition approaches, such as Value Decomposition Networks~\citep{sunehag2018value} and QMIX~\citep{rashid2018qmix}, or parameter sharing in decentralized policy optimization~\citep{gupta2017cooperative}.
\citet{Guestrin2002-il} introduced the concept of a coordination graph to reason about joint value estimates from a factored representation, while~\citet{kok2004sparse} used approximations to scale these ideas to larger problems.
Monte Carlo Tree Search (MCTS)~\citep{DBLP:journals/tciaig/BrownePWLCRTPSC12}, a common approach to \emph{online} planning, has been combined with coordination graphs in Factored Value MCTS~\citep{Amato2014-io}.
However, Factored Value MCTS coordinates actions with an exact Variable Elimination (Var-El) step, which negates the anytime nature of MCTS planning.
The key idea of this paper is to \emph{recover the anytime nature of MCTS planning for MMDPs requiring coordination and also scale to larger teams of agents}.
To that end, we propose combining Max-Plus action selection, introduced by \citet{Vlassis2004-da}, with the factored value MCTS of \citet{Amato2014-io}. We do so for many reasons.
Unlike Var-El, which is exact, Max-Plus is an iterative procedure and allows for truly anytime behavior that can trade computation for approximation quality. The representation of Max-Plus is much more efficient than that of Var-El for using
dynamic, i.e., state-dependent, coordination graphs (state-dependent data-structures are a key benefit of online planning for MDPs). Finally, Max-Plus can scale to much larger
and denser coordination graphs than Var-El, and it can be distributed for additional scalability~\cite{DBLP:conf/bnaic/KokV05}.
We present a scalable anytime MMDP planning algorithm called Factored Value MCTS with Max-Plus. On the standard SysAdmin benchmark domain~\citep{Guestrin2002-il}, with static coordination graphs, we demonstrate that our approach performs as well as or better than Factored Value MCTS with Var-El~\cite{Amato2014-io} and is much faster for the same tree search hyperparameters.
We also introduce a new MMDP domain, Multi-Drone Delivery, with dynamic coordination graphs. On the second domain, we show how our approach scales to problem sizes that are entirely intractable for other MCTS variants, while also achieving better performance on smaller problem sizes.
\section{Anytime Factored-Value Monte Carlo Tree Search}
\label{sec:approach}
We now discuss our method for anytime multi-agent MDP planning with coordination graphs, \emph{Factored-Value Monte Carlo Tree Search with Max-Plus}.
To apply the mixture of experts optimization to each node of the search tree, we must define the factored statistics to maintain for each node.
Given a potentially state-dependent undirected \emph{coordination graph} (CG), $\mathcal{G} = \langle \mathcal{V}, \mathcal{E} \rangle$, we factor the CG-induced global payoff at the current state, $\overline{s}$, as follows:
\begin{equation}
\label{eq:maxplus-payoff}
Q(\overline{a}) = \sum_{i \in \mathcal{V}} Q_i(a_i) + \sum_{(i,j) \in \mathcal{E}} Q_{ij}(a_i,a_j).
\end{equation}
Here, $Q_{ij}$ is a local payoff function for agents $i$ and $j$ connected by edge $(i,j)$, while $Q_i$ is an individual utility function for agent $i$ (if applicable to the domain). \emph{All state-dependent quantities in this section's equations are implicitly for the current joint state $\overline{s}$}.
\begin{figure*}[t]
\centering
\includegraphics[width=0.9\textwidth]{figures/mcts_mp}
\caption{Our anytime MMDP planning algorithm, Factored Value MCTS with Max-Plus, computes the best joint action $\overline{a}^{*}$ for the current joint state $\overline{s}$. The tree search uses an Upper Confidence Bound (UCB) exploration bonus during action selection, while the final action coordination does not.}
\label{fig:fv-mcts-mp}
\end{figure*}
Exploiting the duality between computing the maximum a posteriori configuration in a probabilistic graphical model and the optimal joint action in a CG, \citet{Vlassis2004-da} introduced the Max-Plus algorithm for computing the joint action via message passing.
Each node, i.e., agent, iteratively dispatches messages to its neighbours $j \in \Gamma(i)$ in the CG (\Cref{fig:cgfig-maxplus}). A message from agent $i$ is a scalar-valued function of the action space of \emph{receiving agent} $j$, i.e.,
\begin{equation}
\label{eq:maxplus-msg}
\mu_{ij}(a_j) = \max_{a_i} \Big\{Q_i(a_i) + Q_{ij}(a_i,a_j) + \sum_{k \in \Gamma(i)\setminus \{j\}} \mu_{ki}(a_i)\Big\},
\end{equation}
where $\Gamma(i)$ is the set of neighbors of $i$.
Agents exchange messages until convergence or for a maximum number of rounds.
Finally, each agent computes its optimal action individually, i.e.,
\begin{equation}
a_i^{*} = \underset{a_i}{\mathrm{argmax}} \Big\{Q_i(a_i) + \sum_{j \in \Gamma(i)} \mu_{ji}(a_i)\Big\}
\end{equation}
Max-Plus is equivalent to belief propagation in graphical models \citep{DBLP:books/daglib/Pearl89} and its time complexity scales linearly with the CG size (the number of edges);
it is more suitable for real-time systems and more tractable for large numbers of agents than Var-El. %
Similar to Factored Value MCTS with Var-El, our method with Max-Plus (that we illustrate in~\Cref{fig:fv-mcts-mp}) is more efficient than a naive application of MCTS with the joint action space, since it retains fewer statistics and performs efficient action selection. For the rest of this section, we will discuss the key differences from the prior
work of~\citet{Amato2014-io}, which underscore how our approach is more suitable than it for large MMDPs.
\subsection{UCB Exploration with Max-Plus}
\label{sec:approach-exploration}
The key implementation issue for extending MCTS to factored value functions and coordination graphs is
that of action exploration as per the Upper Confidence Bound (UCB) strategy.
In the Var-El case, \citet{Amato2014-io} added the exploration bonus using component-wise statistics during each elimination step in~\Cref{eq:varel-exp}. We cannot apply this strategy with Max-Plus as it does not use components. In contrast, it has two distinct phases of computation. The first is message passing per edge in~\Cref{eq:maxplus-msg}, followed by action selection per node in~\Cref{eq:maxplus-payoff}. We use these two phases to define how our algorithm explores.
\textbf{Edge Exploration:} Analogous to the edge payoff statistics $Q_{ij}$, we keep track of corresponding frequency statistics $N_{a_i,a_j}$ (for pairwise actions).
The natural exploration strategy over edges is to add the bonus to~\Cref{eq:maxplus-msg} as follows:
\begin{multline}
\label{eq:mp-edge-exp}
\mu_{ij}(a_j) = \max_{a_i} \Bigg\{Q_i(a_i) + Q_{ij}(a_i,a_j) + \\ \sum_{k \in \Gamma(i)\setminus \{j\}} \mu_{ki}(a_i) + c\sqrt{\frac{\log(N + 1)}{N_{a_i,a_j}}}\Bigg\}.
\end{multline}
Adding this bonus during the message passing rounds can cause divergence for cyclic graphs with any cycle of length less than the number of rounds.~\Cref{fig:edge_exp} illustrates intuition for this divergent behavior with a simple triangle graph. The bonuses accumulate in successive rounds for messages in either direction along the cycle, making the effective bonus proportional to the total number of rounds (divided by cycle length). Therefore, we only augment each message once \emph{after the final round of message passing}.
\textbf{Node Exploration:} We maintain individual action frequency statistics $N_{a_i}$ and modify \Cref{eq:maxplus-payoff} to add a node exploration bonus during the action selection:
\begin{equation}
\label{eq:mp-node-exp}
a_i^{*} = \underset{a_i}{\mathrm{argmax}} \Bigg\{Q_i(a_i) + \sum_{j \in \Gamma(i)} \mu_{ji}(a_i) +
c\sqrt{\frac{\log(N + 1)}{N_{a_i}}}\Bigg\}
\end{equation}
\noindent
\\
Note that the joint-action payoff $Q(\overline{a})$ can be factorized over the CG nodes and edges as in~\Cref{eq:maxplus-payoff}, but the joint-action exploration bonus $c\sqrt{\frac{\log{N(\overline{s})}}{N(\overline{s},\overline{a})}}$ cannot. Therefore, the node and edge exploration strategies we have defined here are heuristic choices that we make and will evaluate empirically through an ablation, rather than strategies we derive analytically by factorizing the exploration bonus. The same was true even in the exact action selection case of Var-El, where the analysis was restricted to the approximation quality of the factored value function, not the exploration itself. Quoting~\citet{Amato2014-io}, ``it is not possible to demonstrate that the UCB exploration policy is component-wise'', where, as from~\Cref{sec:related-cgs}, the components refer to the cliques in the CG.
\\
We outline our approach in~\Cref{algo:fvmctsmp}
as well as the Max-Plus routine in~\Cref{algo:maxplus}.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{edge_exploration}
\caption{For coordination graphs with cycles, adding an edge exploration bonus to the messages \emph{at every round} can lead to divergent behavior. The exploration bonuses accumulate over rounds from one node to the next (clockwise or anti-clockwise along the cycle). If the number of rounds is greater than the length of a cycle (usually true), the effective relevant exploration bonus for each edge gets compounded each time the messages loop back around. We abuse some notation for convenience, i.e., $N_{ij}$ is a counting function for the pairwise actions of agents $i$ and $j$.}\label{fig:edge_exp}
\vspace{-5pt}
\end{figure}
\algdef{SE}%
[STRUCT]%
{Struct}%
{EndStruct}%
[1]%
{\textbf{struct} \textsc{#1}}%
{}%
\begin{algorithm}[t]
\caption{Factored Value MCTS with Max-Plus}
\label{algo:fvmctsmp}
\begin{algorithmic}[1]
\Require{time limit, depth, exploration constant $c$, state $\overline{s}$}
\Statex
%
\State Initialize $N_i, Q_i$ \Comment{Node statistics}
\State Initialize $N_{ij}, Q_{ij}$ \Comment{Edge statistics}
%
\State
\Function{FV-MCTS-MP}{$\overline{s}$, depth}
\While{time limit not reached}
\State{\Call{Simulate}{$\overline{s}$, depth}}
\EndWhile
\State{$\overline{a}^* \gets$ \Call{MaxPlus}{$0$}} \Comment{No exploration here}
\State{\Return{$\overline{a}^*$}} \Comment{Best joint action}
\EndFunction
\Function{Simulate}{$\overline{s}$, depth}
\If{$\mathrm{depth} = 0$}
\State{\Return $0$}
\EndIf
\State{$\overline{a}\gets$ \Call{MaxPlus}{$c$}}
\State{$\overline{s}', \overline{r} \sim T(\overline{s},{\overline{a}}), R(\overline{s},\overline{a})$} \Comment{Generative model}
\State{$\overline{q} \gets \overline{r} + \gamma \cdot$ \Call{Simulate}{$\overline{s}'$, $\text{depth} - 1$}}
\State{\Call{UpdateStats}{$\overline{s}, \overline{a}, \overline{q}$}}
\EndFunction
\Function{UpdateStats}{$\overline{s}, \overline{a}, \overline{q}$}
\For{every agent $i$}
\State{$N_i(\overline{s}, a_i) \mathrel{{+}{=}} 1$}
\State{$Q_i(\overline{s}, a_i) \mathrel{{+}{=}} \frac{q_i - Q_i(\overline{s}, a_i)}{N_i(\overline{s}, a_i)}$}
\EndFor
\For{every edge $(i, j) \in \mathcal{G}(\overline{s})$}
\State{$N_{ij}(\overline{s}, a_i, a_j) \mathrel{{+}{=}} 1$}
\State{$q_e \gets q_i + q_j$}
\State{$Q_{ij}(\overline{s}, a_i, a_j) \mathrel{{+}{=}} \frac{q_e - Q_{ij}(\overline{s}, a_i, a_j)}{N_{ij}(\overline{s}, a_i, a_j)} $}
\EndFor
\EndFunction
\end{algorithmic}
\end{algorithm}
\algnewcommand{\LineComment}[1]{\State \(\triangleright\) #1}
\begin{algorithm}[t]
\caption{MaxPlus Action Selection}
\label{algo:maxplus}
\begin{algorithmic}[1]
\Require{Coordination Graph $\mathcal{G}(s) = \langle \mathcal{V}, \mathcal{E}\rangle$; state node statistics $N, Q$; max iterations $M$; flags (exploration; normalization)}
\Statex
\Function{MaxPlus}{c}
\For{$t \gets 1$ to $M$}
\State{$\mu_{ij}(a_j) = \mu_{ji}= 0$ for $(i,j) \in \mathcal{E}, a_i \in \mathcal{A}_i, a_j \in \mathcal{A}_j$}
\For{every agent $i$}
\For{all neighbors $j \in \Gamma(i)$}
%
\State{Compute $\mu_{ij}(a_j)$ via \Cref{eq:maxplus-msg}}
\If{message normalization}
\State{$\mu_{ij}(a_j) \mathrel{{-}{=}} \frac{1}{\lvert \mathcal{A}_j \rvert} \sum_{a_j \in \mathcal{A_j}} \mu_{ij}(a_j)$}
\EndIf
\State{send message $\mu_{ij}(a_j)$ to agent $j$}
\If{$\mu_{ij}(a_j)$ close to previous message}
\State{break}
\EndIf
\EndFor
\EndFor
\For{every agent $i$}
\If{edge exploration}
\For{all neighbors $j \in \Gamma(i)$}
%
%
\State{Compute $\mu_{ij}(a_j)$ via \Cref{eq:mp-edge-exp}}
\EndFor
\EndIf
\State{$q_i(a_i) = Q_i(a_i) + \sum_{j\in \Gamma(i)} \mu_{ji}(a_i)$}
\If{node exploration}
\State{$q_i(a_i) \mathrel{{+}{=}} c \sqrt{\frac{\log (N + 1)}{N_i(a_i)}}$} %
\EndIf
\State{$a'_{i} = \arg\max_{a_i} q_i(a_i)$}
\EndFor
%
%
%
%
\If{time limit reached}
\State{break}
\EndIf
\EndFor
\State{\Return{$\overline{a}'$}}
\EndFunction
\end{algorithmic}
\end{algorithm}
\subsection{Other differences from FV-MCTS with Variable Elimination}
\label{sec:approach-differences}
\textbf{Convergence:}
For graphs without cycles, Max-Plus converges to a fixed point in finitely many iterations~\cite{DBLP:books/daglib/Pearl89}.
For cyclic graphs, there are no such guarantees in general~\cite{DBLP:journals/sac/WainwrightJW04}.
However, cyclic message passing can work well in practice~\cite{DBLP:conf/uai/MurphyWJ99}.
\noindent\textbf{Agent Utilities:} The Max-Plus global payoff in~\Cref{eq:maxplus-payoff} includes a utility function $Q_i$ for each individual agent.
The FV-MCTS with Var-El has no such individual utility (unless a node has degree $0$ in the CG).
If such agent utilities were known or learned \emph{independent of the payoffs},
we would naturally use them during action coordination.
However, in FV-MCTS we estimate all statistics from the rewards obtained during tree search with a simulated environment model; the environment model returns precisely one reward vector for each joint state-action pair.
We already account for the simulated rewards in tree search through the $Q_{ij}$ local payoff statistics in~\Cref{eq:varel-q-update}. We do not receive independent per-agent rewards, so
utility statistics would be derived from the same information we use for the payoff statistics.
Our experiments compare the benefit of these derived individual agent (node) utilities, in addition to local edge payoffs. We maintain separate statistics $N_i$ and $Q_i$ for the per-agent frequencies and utilities respectively and estimate them from the joint rewards during tree search; the corresponding updates are $N_i(\overline{s},\overline{a}_i) = N_i(\overline{s},\overline{a}_i) + 1$ and $Q_i(\overline{s}, \overline{a}_i) = Q_i(\overline{s}, \overline{a}_i) + \frac{\overline{r}_i - Q_i(\overline{s}, \overline{a}_i)}{N_{\overline{a}_i}}$ for an agent $i$. The results in~\Cref{sec:results-sysadmin} demonstrate how \emph{including derived node utilities} enables better empirical performance.
\noindent\textbf{Dynamic Coordination Graphs:} Recall that MCTS (and online MDP planning in general) can use computational structures that vary with the current state.
\emph{For FV-MCTS with Var-El, state-dependent or dynamic CGs are not feasible} because
eliminating an agent can change the intermediate CG topology during action coordination (by adding edges).
It is not tractable to maintain statistics for the set of all possible CG components, the size of which is exponential in the number of agents~\cite{zykov1949some}.
On the other hand, Max-Plus only maintains statistics for at most all CG edges, over which the messages are sent.
Therefore, dynamic CGs can be used seamlessly~\citep{yu2020dist}.
\noindent\textbf{Memory Complexity of Statistics}: Factored Value MCTS collects a set of frequency and payoff statistics for each unique joint state encountered during tree search.
Assume an MMDP with the same action set $\mathcal{A}$ for each agent, and a CG with $|\mathcal{V}|$
nodes (agents), $|\mathcal{E}|$ edges, and $C$ local components or cliques. Then, the memory complexity of per-state statistics for Max-Plus is $\mathcal{O}(|\mathcal{V}| |\mathcal{A}| + |\mathcal{E}||\mathcal{A}|^2)$ (the first term only applies if we track per-node utilities).
In contrast, the per-state memory for Var-El statistics is $\mathcal{O}(\sum_{c \in C} |\mathcal{A}|^{|\mathcal{V}|_c} \cdot |\mathcal{V}|_c)$, where $|\mathcal{V}|_c$ is the size of
local component $c$. For a CG that is connected (typically the case), the memory complexity for Var-El is at least $\mathcal{O}(|\mathcal{E}||\mathcal{A}|^2)$, which is the dominant term for Max-Plus, and more generally is exponential in the largest clique. \emph{Therefore, Max-Plus is more memory-efficient than Var-El}. The experiments in~\Cref{sec:results-drone} empirically support this claim by showing how our algorithm can solve problems that cause out-of-memory
issues for the Var-El baseline.
\noindent\textbf{Distributed Implementation:} Unlike with Var-El, \emph{we can execute Max-Plus in a distributed manner} by sending messages in parallel, albeit incurring additional communication complexity.
Such an implementation can allow further scalability with available compute. Note that this is distinct from full decentralization wherein the agent actions can be computed independently. %
\begin{figure*}[t]
\centering
\begin{subfigure}{0.4\textwidth}
\centering
\fbox{\resizebox{0.75\columnwidth}{!}{\input{figures/star_sysadmin.tex}}}
\caption{}
\label{fig:domains-sysadmin}
\end{subfigure}
\begin{subfigure}{0.4\textwidth}
\centering
\fbox{\includegraphics[width=0.95\textwidth]{figures/multi_drone_deliv_pdf}}
\caption{}
\label{fig:domains-multidrone}
\end{subfigure}
\caption{Our two experimental domains: (\subref{fig:domains-sysadmin}) SysAdmin (star topology) and (\subref{fig:domains-multidrone}) Multi-Drone Delivery, where dotted lines illustrate
a subset of the Coordination Graph edges for the current state (for clarity, we omit some edges between drones of the same color, i.e., assigned to the same goal).}
\label{fig:domains}
\end{figure*}
\section{Background and Related Work}
\label{sec:related}
We first review Markov decision processes (MDPs) and their multi-agent formulation. We then describe how coordination graphs can efficiently exploit locality of interactions in multi-agent problems. Finally, we discuss how to use coordination graphs to solve multi-agent MDPs.
\subsection{Multi-Agent Markov Decision Processes}
\label{sec:related-mmdps}
An MDP is defined by the tuple $(\mathcal{S}, \mathcal{A}, {T}, R)$, where $\mathcal{S}$ is the state space, $\mathcal{A}$ is the action space, $T: \mathcal{S} \times \mathcal{A} \times \mathcal{S} \to [0, 1]$
is the transition function, and $R: \mathcal{S} \times \mathcal{A} \to \mathbb{R}$ is the reward function.
The objective for solving an MDP is to obtain a \emph{policy}, $\pi: \mathcal{S} \times \mathcal{A} \to [0,1]$ that specifies a probability distribution over actions for the agent to take from its current state to maximize its \emph{value}, i.e. its expected cumulative reward. An action-value function $Q(s, a)$ defines the expected cumulative reward after taking action $a$ in state $s$ before following the specified policy.
We focus on decision-making settings where multiple agents cooperate to achieve a shared objective~\cite{boutilier1996planning}.
Such problems are multi-agent Markov decision processes (MMDPs), where each agent takes an individual action and the controller policy %
observes the states of all agents.
In principle, we can solve an MMDP as a standard MDP with a joint action space $\mathcal{A} = \prod_{i} \mathcal{A}_i$ ~\citep{DBLP:journals/jair/PynadathT02}.
There exist both offline and online methods for computing such MDP policies~\cite{bertsekas2005dynamic}.
\emph{Offline} methods pre-compute a policy over the entire state space (exactly or approximately) and query the policy during execution. Various exact offline methods exist, but reinforcement learning has emerged as an attractive solution technique due to the complexity of planning in large MMDPs~\citep{DBLP:books/lib/SuttonB98}.
Reinforcement Learning approaches attempt to compute an effective value function $Q(s,a)$
or a policy $\pi(a \mid s)$
through repeated interaction with the environment model.
They still have difficulty with the size of the joint action space, which is exponential in the number of agents. A common strategy is to \emph{decentralize} the policy or value function, such that each only depends on the actions of a single agent~\citep{sunehag2018value,rashid2018qmix,gupta2017cooperative}.
Unfortunately, such decentralization approaches are often suboptimal for coordination and encounter exploration bottlenecks due to uncooperative random actions from the agents~\citep{Bohmer2019-zv}.
\emph{Online} methods use an alternative strategy to deal with the complexity of multi-agent planning; they interleave planning and execution by focusing only on states that are reachable for the current state, while computing the next action to take. Monte Carlo Tree Search (MCTS) is the predominant framework for online planning and has succeeded in a variety of domains~\cite{DBLP:journals/tciaig/BrownePWLCRTPSC12}, including in multi-player contexts~\citep{nijssen2011,zerbel2019multiagent}.
The \emph{anytime} nature of MCTS (search depth and number of simulations) allows us to trade computation time for approximation quality.
However, the exponentially large action space of MMDPs can still be a bottleneck for the naive application of MCTS techniques~\cite{chaslot2008progressive}.
Dec-MCTS tries to work around this bottleneck but does not apply to action-dependent stochastic transitions of an MDP, as it directly chooses the next state~\citep{Best2019-rr}.
\subsection{Coordination Graphs and Variable Elimination}
\label{sec:related-cgs}
Several real-world multi-agent systems demonstrate \emph{locality of interaction}, i.e. the
outcome of an agent's action depends only on the actions of a subset of other agents.
The coordination graph (CG) structure is often used to encode such interactions~\cite{Guestrin2002-il,Guestrin2003}. A CG for a multi-agent system has one node
per agent, and edges connect agents if their payoffs depend on each other.
For now, we assume a stateless or single-shot decision setting (rather than sequential).
The CG structure induces a set of payoff components, \emph{where each component is associated with a clique}, i.e., a subset of agents that are all mutually connected.
For CGs in multi-agent settings, we assume that we can factor the global payoff for a joint action as the sum of local component payoffs, i.e. $Q(\overline{a}) = \sum_{c} Q_c(\overline{a}_c)$, where $\overline{a}$ is the global joint action, and $\overline{a}_c$ is the local component action (the projection of $\overline{a}$ corresponding to component $c$).
Given this factored representation and the local component payoffs%
, we can compute exactly the best joint action, $\mathrm{argmax}_{\overline{a}} \ Q(\overline{a})$,
with the Variable Elimination (Var-El) algorithm originating from the probabilistic inference literature~\cite{Guestrin2003}. Computing the optimal
joint action in a CG is equivalent to obtaining the maximum a posteriori configuration in an undirected probabilistic graphical model~\cite{Vlassis2004-da}.
\begin{figure}
\centering
\begin{subfigure}{0.49\columnwidth}
\centering
\includegraphics[width=0.99\textwidth]{figures/cg_varel_pdf}
\caption{}
\label{fig:cgfig-varel}
\end{subfigure}
\begin{subfigure}{0.49\columnwidth}
\centering
\includegraphics[width=0.99\textwidth]{figures/cg_maxplus_pdf}
\caption{}
\label{fig:cgfig-maxplus}
\end{subfigure}
\caption{A coordination graph for a $4$-agent MMDP with $4$ agents. (\subref{fig:cgfig-varel}) Eliminating agent $2$ in Var-El introduces an edge between nodes (agents) $1$ and $3$ and a
new payoff function $e_2$ (\subref{fig:cgfig-maxplus}) In Max-Plus, agents passes messages along the graph edges; the messages are functions of the actions of the receiving agent, e.g.,
agent $1$ sends $\mu_{12}(a_2)$ to agent $2$.}
\label{fig:cgfig}
\end{figure}
Consider the $4$-agent CG in \Cref{fig:cgfig-varel}. Here, $Q(\overline{a}) = Q_{12}(a_1,a_2) + Q_{23}(a_2,a_3) + Q_{34}(a_3,a_4)$, where $a_i$ is the action variable for agent $i$. In Var-El, we \emph{eliminate}, i.e., maximize over variables one at a time by collecting the local payoffs that depend on them. For instance, if we start with agent $2$, then
\begin{equation}
\max_{a_1,a_3,a_4} Q_{34}(a_3, a_4) + \max_{a_2} \left[Q_{12}(a_1,a_2) + Q_{23}(a_2,a_3) \right]
\end{equation}
is the first elimination. The optimal choice for agent $4$ depends on $a_2$ and $a_3$. The internal max expression is summarized by a new intermediate payoff function $e_2(a_1,a_3) = \max_{a_2} [Q_{12}(a_1,a_2) + Q_{23}(a_2,a_3)]$ and a new edge between $1$ and $3$, after which the algorithm continues with $Q_{34}$ and $e_2$.
After all eliminations, we recover the action for each agent by maximizing the conditional functions in reverse, finally obtaining the optimal joint action. Var-El is exponential in the induced width of the CG, which depends on the elimination order~\cite{DBLP:journals/ai/Dechter99}.
Although most works in the literature assume a domain-dependent static coordination graph structure, some incorporate state-dependent or dynamic CGs~\citep{yu2020dist}, including learning the CG structures~\citep{kok2005utile,li2020deep}.
\subsection{Scalable MMDP Methods with Coordination Graphs}
\label{sec:related-scalable}
In the \emph{offline} context of tabular RL methods, \citet{kok2004sparse} explored action inference with predefined static coordination graphs over factorized value functions;~\citet{Bohmer2019-zv} extended these ideas to the neural network function approximation regime. We focus on \emph{anytime online} planning approaches to solving MMDPs.
\citet{Amato2014-io} provide an online planning solution by combining the idea of coordination graphs and factored values with MCTS.
Although they apply their algorithm to partially observed MDPs, the key ideas are the same for
the fully observed case.
Monte Carlo planning \emph{estimates} quantities by exploring from the current state and gathering relevant statistics through interactions with a simulated generative model of the environment~\cite{DBLP:conf/nips/SilverV10}.
These statistics typically track the average simulated reward obtained for trying an (individual or joint) action, the frequency of action attempts (for Upper Confidence Bound or UCB exploration~\cite{DBLP:conf/ecml/KocsisS06}), and the number of occurences of the (individual or joint) state.
\citet{Amato2014-io} maintain \emph{local component statistics}, i.e. the mean payoff of a local component action $\overline{a}_e$ and the number of times it was attempted in that component (they call this \emph{mixture of experts optimization}, albeit with a simple maximum likelihood estimator expert).
For instance, during tree search from the current joint state $\overline{s} \equiv \{s_i\}$ (where $s_i$ is the state of agent $i)$,
suppose the system simulates a joint action $\overline{a}$ and obtains a reward vector $\overline{r}$.
Then, in a particular CG component $e$ and the corresponding local subset of the joint action $\overline{a}_e$, they augment the local component action frequency statistic $N(\overline{s},\overline{a}_e)$ by $1$ and update the local component payoff statistic $Q_e(\overline{s}, \overline{a}_e)$ as
\begin{equation}
\label{eq:varel-q-update}
Q_e(\overline{s}, \overline{a}_e) \doteq Q_e(\overline{s}, \overline{a}_e) + \frac{\overline{r}_e -Q_e(\overline{s}, \overline{a}_e)}{N(\overline{s},\overline{a}_e)},
\end{equation}
which is a standard running average update.
The UCB exploration step uses the current statistics to select
joint actions, i.e.,
\begin{equation}
\label{eq:varel-exp}
\max_{\overline{a}} \sum_{e} U_e(\overline{s}, \overline{a}_e) = \max_{\overline{a}} \sum_{e} Q_e(\overline{s}, \overline{a}_e) + c \cdot
\sqrt{\frac{\log{N(\overline{s})}}{N(\overline{s},\overline{a}_e)} },
\end{equation}
where $N(\overline{s})$ is the visit frequency for state $\overline{s}$.
Given these local component payoffs, i.e., the $Q_e$ functions, their method computes the best joint action at the next time-step through Variable Elimination over the CG, as in~\Cref{sec:related-cgs}.
\emph{Consequently, it loses the anytime property of MCTS because exact variable elimination cannot be stopped at an intermediate step}. Although \citet{Vlassis2004-da} explored various anytime algorithms for action selection with coordination graphs, they did not investigate their interaction with online planning algorithms like MCTS.
\section{Experiments and Results}
\label{sec:results}
We used cumulative discounted return as the primary metric to evaluate our approach, Factored Value MCTS with Max-Plus (FV-MCTS-MP). Our most relevant baseline is Factored Value MCTS with Variable Elimination (FV-MCTS-Var-El). We also compared against standard MCTS (with no factorization), independent Q-learning (IQL), and a random policy. Besides measuring performance, we examined the effect of different exploration schemes on the performance of FV-MCTS-MP (as we discussed in~\Cref{sec:approach-differences}) and the problem size on MCTS computation time. The appendix provides performance results for FV-MCTS (both variants) with different hyperparameters. Both of our experimental domains represent a range of MMDP problems and underlying coordination graphs (CGs). The source-code for experiments is available at \url{https://sites.google.com/stanford.edu/fvmcts/}. All implementations and simulations are in Julia~\citep{bezanson2017julia}.
We will show qualitatively how our approach recovers the true anytime nature of MCTS by using Max-Plus rather than Var-El. However, \textit{there are many confounds for quantitatively evaluating the anytime property}. Our metric is the average discounted return over the episode, where the Max-Plus routine is called several times; typical anytime evaluation reports improving solution quality with more compute time for a single call to a method. MCTS itself has several parameters that affect the computation-vs-quality tradeoff, such as tree depth, exploration constant, and number of trials. With dynamic CGs as in our multi-drone delivery domain, the same Max-Plus parameters leads to different computation times. Note that our reference for Max-Plus does evaluate its anytime property in a one-shot decision-making domain that does not have any of the above confounds~\cite{Vlassis2004-da}.
\subsection{SysAdmin Domain}
\label{sec:results-sysadmin}
Our first domain is a standard MMDP benchmark: SysAdmin \citep{Guestrin2003}.
Each agent $i$ represents a machine in a network with two state
variables: Status $S_i \in \{\textsc{good}, \textsc{faulty}, \textsc{dead}\}$, and Load $L_i \in \{\textsc{idle}, \allowbreak \textsc{loaded}, \textsc{success}\}$.
A \textsc{dead} machine increases the probability that its neighbor also dies.
The system gets a reward of $1$ if a process terminates successfully,
processes take longer when status is \textsc{faulty}, and a \textsc{dead} machine loses the process.
Each agent must decide whether to reboot its machine, in which case the Status becomes \textsc{good} and any running process is lost.
The discount factor, $\gamma$ used in all the experiments is $0.9$.
All evaluations have been averaged over $40$ runs. Error bars indicate standard deviations.
\Cref{fig:domains-sysadmin} illustrates the star network topology for SysAdmin; we also use the ring topology as well as the ring-of-rings topology.
\noindent
\textbf{Exploration Schemes for FV-MCTS-MP}:
The three knobs affecting exploration in FV-MCTS-MP are per-agent utility, node bonus, and edge bonus. We compared the discounted return of variants that either use or ignore per-agent utilities, and use either or both bonuses.~\Cref{fig:exploration} demonstrates that the combination of \emph{agent utilities and only node exploration (TTF) is enough}; including the edge bonus as well (TTT) does not have much effect. All other schemes are significantly poorer. Therefore, we used the TTF variant of FV-MCTS-MP to compare against the other baselines. Lack of significant difference between some of the exploration strategies has more to do with the small action space of the SysAdmin domain. %
\begin{figure*}[t]
\centering
\begin{subfigure}{0.49\textwidth}
\centering
\includegraphics[width=0.75\columnwidth]{figures/explore_bisysadmin_global_4.pdf}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\centering
\includegraphics[width=0.75\columnwidth]{figures/explore_starsysadmin_global_4.pdf}
\end{subfigure}
\caption{The performance of FV-MCTS-MP varies with different combinations of exploration strategies for the 4-agent Sysadmin on Ring (left) and Star (right) topologies. The True/False (T/F) labels correspond to Agent Utilities, Node Exploration and Edge Exploration in order, e.g. TTF implies agent utilities and only node (but not edge) exploration.}
\label{fig:exploration}
\end{figure*}
\begin{figure*}[t]
\centering
\begin{subfigure}{0.33\textwidth}
\includegraphics[width=0.99\columnwidth]{figures/scaling_agents_ring.pdf}
\end{subfigure}
\begin{subfigure}{0.33\textwidth}
\includegraphics[width=0.99\columnwidth]{figures/scaling_agents_star.pdf}
\end{subfigure}
\begin{subfigure}{0.33\textwidth}
\includegraphics[width=0.99\columnwidth]{figures/scaling_agents_ringofring.pdf}
\end{subfigure}
\caption{On SysAdmin topologies: Ring (left), Star (middle), and Ring-of-Rings (right), FV-MCTS with MaxPlus performs as well as or slightly better than Var-El, while being much more efficient for larger problems as in~\Cref{fig:sysadmin_timing}. \textcolor{red}{NaN} indicates that the algorithm ran out of memory.}
\label{fig:probsize_nodes}
\end{figure*}
\noindent
\textbf{FV-MCTS-MP compared to baselines}:
For all three SysAdmin topologies (and corresponding fixed CGs), we varied the number of machines (agents) and compared the performance of all methods in~\Cref{fig:probsize_nodes}.
With fewer agents, all MCTS methods perform similarly to each other and better than Q-Learning.
However, with more agents, standard MCTS runs out of memory even on our 128GB RAM machine, as expected for large joint action spaces.
Both Factored Value MCTS variants perform comparably on larger problems (on ring-of-rings our Max-Plus variant was better). However, as we discuss subsequently, FV-MCTS-Var-El
is much slower than FV-MCTS-MP, e.g., taking approximately \SI{35}{\second}
versus \SI{16}{\second} for $32$ agents on a single-threaded implementation in the Ring topology.
\emph{Therefore our approach strictly dominates the Var-El baseline on the performance-time tradeoff}.
\noindent\textbf{Effect of Hyperparameters}:
We performed ablation experiments, varying one of exploration constant $c$, tree search exploration depth $d$ and number of Monte Carlo rollouts $n$, while keeping the rest of the hyperparameters constant.
Low values of $n$ adversely effected the performance a little for both FV-MCTS-Var-El and FV-MCTS-MP, while \emph{low values of $c$ vastly degraded the performance of FV-MCTS-Var-El only}.
The difference in performance was not significant over a range of values of $d$.
\Cref{fig:hyperparam_fig} in the appendix, shows results for $32$ agents. Similar results hold for the other domains as well as different numbers of agents.
\noindent\textbf{Computation Time}:
For the same tree search hyperparameters with number of iterations fixed as $16000$, exploration constant as $20$ and tree search depth as $20$, we compared the average time taken for each action for different number of agents in the coordination graphs.
For a fair comparison, we used a single threaded implementation.
As demonstrated in \Cref{fig:sysadmin_timing}, we found FV-MCTS-MP to be consistently faster than FV-MCTS-Var-El.
Although MCTS was faster when there were small number of agents, it ran out of memory as the number of agents increased.
\begin{figure*}[t]
\centering
\begin{subfigure}{0.46\textwidth}
\centering
\includegraphics[width=0.9\columnwidth]{figures/ringtiming.pdf}
\end{subfigure}
\begin{subfigure}{0.46\textwidth}
\centering
\includegraphics[width=0.9\columnwidth]{figures/startiming.pdf}
\end{subfigure}
\caption{Runtime comparisons (lower is better) for the same tree search hyperparameters on SysAdmin with Ring (left) and Star (right) topologies. The NaiveMCTS baseline ran out of memory with more than $8$ agents.}
\label{fig:sysadmin_timing}
\end{figure*}
\subsection{Multi-Drone Delivery Domain}
\label{sec:results-drone}
Besides the SysAdmin domain, previous multi-agent decision-making work has also used the Firefighter~\citep{Amato2014-io} and Traffic Control~\citep{kuyer2008multiagent} domains for benchmarking.
Underneath the differing high-level descriptions, however, \emph{the MMDP details of all three domains are very similar}:
a small state space and binary action space, the degrees of most nodes in the coordination graph are independent of the total number
of agents (except the hub node for Star SysAdmin), and there is no scope in any of them for dynamic CGs.
We introduce and use a truly distinct domain for our second set of experiments.
It simulates \emph{a team of delivery drones navigating a shared operation space to reach their assigned goal regions}.
We are motivated by recent advances in drone delivery technology, from high-level routing to low-level control~\cite{dorling2016vehicle,DBLP:conf/syscon/Lee17};
in particular, drones using ground vehicles as temporary modes of transit to save energy and increase effective travel range~\cite{choudhury2019dynamic,ChoudhurySoloveyETAL2020}.
Our domain models a key component of such drone-transit coordination: multiple drones assigned to board transit vehicles in close
proximity to each other (within the same time window).
\\
\noindent
\textbf{Domain details:}
\Cref{fig:domains-multidrone} illustrates our Multi-Drone Delivery domain; for convenience and consistency with MMDP benchmarks we discretize everything,
but MCTS could accommodate a continuous state space. Each drone starts in a randomly sampled unique cell in a grid (we use larger
grids for more drones in our simulations). There are four circular goal regions, one in each quadrant, that represent a transit vehicle; each
goal region has a radius and maximum capacity of drones it can accommodate, since no two drones can occupy the same grid cell (we also vary
goal radius with grid size). We allocate the drones to the goal regions at random such that at least two drones target every region.
\begin{figure}
\centering
\includegraphics[width=0.99\columnwidth]{figures/uav_scaling.pdf}
\caption{For Multi-Drone Delivery, FV-MCTS-MP vastly outperforms the baselines while effectively using dynamic CGs without any memory issues. \textcolor{red}{NaN} indicates that the algorithm ran out of memory.}
\vspace{-10pt}
\label{fig:drone_scaling}
\end{figure}
Each drone has $10$ actions in total: one for moving to each of the $8$-connected grid neighbors, a \textsc{no-op} action for staying in place,
and a \textsc{board} action that is only valid when the drone is inside its assigned goal region. The MMDP is episodic and terminates only when all drones have reached their goals and executed $\textsc{Board}$ inside them, thus boarding the transit vehicle and receiving a reward of $1000$. Drones also receive an intermediate positive reward if they get closer to their assigned goals. Besides drone movement, the other sources
of negative reward, i.e., cost, are penalties for two or more drones being too close to each other, attempting to enter the same cell (which makes them both stay in place),
and attempting to board in the same goal region at the same time.
Unlike the typical MMDP domains used in prior work, \emph{Multi-Drone Delivery motivates dynamic or state-dependent coordination graphs};
any two drones benefit from coordination only when they are close to each other. Therefore, at the current joint state,
we assign a CG edge between any two drones whose mutual distance is lower than a resolution-dependent threshold
(depicted in \Cref{fig:domains-multidrone}). We also add edges apriori between all drones assigned to
the same goal region, as they need to coordinate while boarding.
For all experiments, we set the discount factor $\gamma$ to $1$, the
goal reaching reward to $1000.0$ units, and the collision penalty to $10.0$ units.~\Cref{tab:multidrone_exp} describes the full set of varying problem resolutions and MCTS hyperparameters.
The average degree of the dynamic coordination graphs ranged from $2.4$ for $8$ agents to $11.8$ for $48$ agents.
We averaged all evaluations over $20$ runs; error bars indicate standard deviations.
\begin{table}
\centering
\resizebox{\columnwidth}{!}{\begin{tabular}{rrrrrr}
\toprule
Agents & XY axis res. & Noise & Expl.\ const. & Expl.\ depth & Iter\\
\midrule
8 & 0.20 & 0.10 & 5 & 10 & 4000 \\
16 & 0.10 & 0.05 & 10 & 10 & 8000 \\
32 & 0.08 & 0.05 & 20 & 10 & 16000 \\
48 & 0.05 & 0.02 & 30 & 10 & 24000 \\
\bottomrule
\end{tabular}}
\caption{Multi-Drone Delivery hyperparameters.}
\label{tab:multidrone_exp}
\vspace{-10pt}
\end{table}
\noindent
\textbf{Performance of FV-MCTS-MP against Baselines}:
As with SysAdmin, we varied the number of drones (agents), discretizing the grid appropriately,
and compared against all baselines (except Random) in~\Cref{fig:drone_scaling}.
We observed that FV-MCTS-Var-El and MCTS quickly ran out of memory, which is expected given the large action space per agent. Even on the problems where Var-El runs, its restriction to static CGs leads to slightly worse performance.
On the other hand, FV-MCTS-MP can solve tasks even with $48$ agents successfully.
Moreover, even on the eight agent problem, FV-MCTS-MP is much faster, taking on average approximately \SI{1}{\second} instead of \SI{40}{\second} for FV-MCTS-Var-El for the same tree search hyperparameters.
\emph{FV-MCTS-MP scales to MMDP problem sizes that FV-MCTS-Var-El cannot even accommodate}.
|
\section{Введение}
Одним из важнейших теоретических результатов в выпуклой оптимизации является разработка ускоренных методов оптимизации \cite{nesterov2018lectures}. На начальном этапе развития этой концепции было предложено множество ускоренных вариантов различных методов, применяющихся к решению многих задач выпуклой оптимизации, однако каждый такой случай требовал отдельного, частного рассмотрения возможности ускорения, ввиду чего предлагаемые конструкции были существенно различны и не позволяли предполагать способ их обобщения. Важным шагом к разработке универсальной схемы ускорения методов оптимизации стали работы, в которых предлагался и исследовался метод, названный каталист, основанный на идее ускоренного проксимального градиентного метода \cite{parikh2014proximal, rockafellar1976monotone} и позволяющий ускорять другие методы оптимизации, используя их для последовательного решения ряда резуляризованных по Моро-Иосиде вспомогательных задач \cite{lin2015universal, lin2017catalyst}. В продолжение этих идей в дальнейшем было предложено множество вариантов применения данного метода и его модификаций \cite{ivanova2019adaptive, kulunchakov2019generic, paquette2017catalyst}. Среди наиболее свежих, на момент написания данной статьи, результатов были также описаны обобщения обсуждаемого подхода на тензорные методы \cite{bubeck2019near, doikov2020contracting, monteiro2013accelerated, gasnikov2020accelerated}. Соответствующее представление ускоренной проксимальной оболочки, если опираться на известные авторам сведения, является наиболее общим из описанных в литературе, и потому в данной работе внимание будет обращено прежде всего именно на методы, предложенные в последней из цитируемых выше работ.
Основной мотив данной работы состоит в том, чтобы описать возможности практического применения универсальных ускоренных проксимальных оболочек для конструирования вычислительно и оракульно эффективных методов оптимизации. Рассмотрим классический покомпонентный метод \cite{bubeck2014convex}, итерация которого для выпуклой функции $f: \mathbb{R}^n \rightarrow \mathbb{R}$ имеет вид:
$$x_{k+1}^i = x_k^i - \eta \nabla_i f(x_k), \quad i \sim \mathcal{U}\{1,...,n\},\;\;\eta > 0.$$
Одним из многих приложений данного метода является оптимизация функционалов, вычисление одной компоненты градиента которых значимо эффективнее, чем вычисление полного вектора градиента; в частности, многие задачи в случае разреженных постановок удовлетворяют данному условию. Однако оракульная сложность данного метода при условии остановки метода при достижении $\varepsilon$-малости невязки по значению функции составляет $\mathcal{O}\left(n \frac{\overline{L} R^2}{\varepsilon}\right)$, где $R^2 = \|x_0 - x_*\|_2^2$, $\overline{L}=\frac{1}{n} \sum_{i=1}^n L_i$~--- среднее констант Липшица компонент градиента, притом эта оценка не является оптимальной для класса выпуклых задач. Рассмотрим теперь ускоренный покомпонентный метод, предложенный Ю.Е.~Нестеровым \cite{nesterov2017efficiency},~--- оракульная сложность данного метода соответствует оптимальной оценке: $\mathcal{O}\left(n\sqrt{\frac{\widetilde{L} R^2}{\varepsilon}}\right)$, где $\sqrt{\widetilde{L}} = \frac{1}{n} \sum_{i=1}^n \sqrt{L_i}$ ~--- среднее квадратных корней из констант Липшица компонент градиента. Вместе с тем, ситуация кардинально меняется в случае рассмотрения алгоритмической сложности метода: пусть даже вычисление одной компоненты градиента имеет сложность $\mathcal{O}(s)$, $s \ll n$, сложность итерации ускоренного покомпонентного метода будет составлять $\mathcal{O}(n)$, в отличие от стандартного метода, сложность итерации которого есть $\mathcal{O}(s)$,~--- содержательно это означает, что степень разреженности задачи при применении ускоренного покомпонентного метода не влияет существенно на сложность алгоритма, и кроме того сложность в таком случае квадратично зависит от размерности задачи: вместе это в некоторой степени обесценивает применение покомпонентного метода в данном случае. Таким образом, интересной задачей является построение ускоренного покомпонентного метода, сложность итерации которого, как и в стандартном варианте метода, составляет $\mathcal{O}(s)$, при сохранении оптимальной оракульной сложности~--- в данной работе это удаётся осуществить благодаря применению универсальной ускоренной проксимальной оболочки ``ускоренный метаалгоритм`` \cite{gasnikov2020accelerated}.
Данная статья состоит из двух основных разделов. В разделе \ref{section1} описываются теоретические результаты о сходимости и алгоритмической сложности покомпонентного метода, ускоренного путём применения оболочки ``ускоренный метаалгоритм``. В разделе \ref{section2} на примере задачи оптимизации функционала вида SoftMax экспериментально проверяется эффективность метода в отношении времени его работы, описываются возможности его вычислительно эффективной имплементации и осуществляется сравнение со стандартными методами.
\section{Теоретические гарантии} \label{section1}
Рассмотрим следующую задачу оптимизации функции $f: \mathbb{R}^n \rightarrow \mathbb{R}$:
$$\min_{x \in \mathbb{R}^n} f(x),$$
при таких условиях:
\begin{enumerate}
\item $f$ дифференцируема на $\mathbb{R}^n$;
\item $f$ выпукла на $\mathbb{R}^n$;
\item $\nabla_i f$ удовлетворяет условию Липшица: $\exists L_{i} \in \mathbb{R}:\;\forall x \in \mathbb{R}^n, u \in \mathbb{R}$
$$\left|\nabla_{i} f\left(x+u e_{i}\right)-\nabla_{i} f(x)\right| \leq L_i|u|,$$
где $e_i$~--- $i$-тый орт базиса, $i \in \{1,...,n\}$;
\item $\nabla f$ удовлетворяет условию Липшица с константой $L$.
\end{enumerate}
Обратимся к содержанию работы \cite{gasnikov2020accelerated}, где предложен общий вариант ``ускоренного метаалгоритма`` решения задач выпуклой оптимизации для композитных функционалов вида $F(x) = f(x) + g(x)$. Для рассматриваемой постановки задачи такая общность не требуется, достаточно применить частный случай описанной схемы при $p=1$, $f \equiv 0$ (используются обозначения соответствующей работы), в котором описанная оболочка принимает вид ускоренного проксимального метода. Псевдокод используемого метода представлен в листинге \ref{am}.
\begin{wrapfigure}{r}{0.51\textwidth}
\begin{minipage}{0.51\textwidth}
\begin{algorithm}[H] \label{am}
\SetAlgoLined
\textbf{Вход:} $H > 0$, $x_0 \in \mathbb{R}^n$\;
\vspace{0.2cm}
$\lambda \leftarrow \nicefrac{1}{2H}$\;
$A_0 \leftarrow 0$; $v_0 \leftarrow x_0$\;
\vspace{0.2cm}
\For{k = 0, ..., $\widetilde{N}-1$}
{
$\displaystyle a_{k+1} \leftarrow \frac{\lambda + \sqrt{\lambda^2 + 4 \lambda A_k}}{2}$\;
$A_{k+1} \leftarrow A_k + a_{k+1}$\;
\vspace{0.2cm}
$\displaystyle \widetilde{x}_k \leftarrow \frac{A_k v_k + a_{k+1} x_k}{A_{k+1}}$\;
\vspace{0.2cm}
Посредством запуска метода $\mathcal{M}$\\
найти с точностью $\varepsilon$ по аргументу\\
решение вспомогательной задачи:\\
\vspace{0.1cm}
$\displaystyle v_{k+1} \in \text{Arg}^{\varepsilon} \min_{y \in \mathbb{R}^n} \left\{ f(y) + \frac{H}{2}\|y - \widetilde{x}_k\|^2_2 \right\}$\;
\vspace{0.5cm}
$\displaystyle x_{k+1} \leftarrow x_k - a_{k+1} \nabla f(v_{k+1})$\;
}
\Return $v_{\widetilde{N}}$\;
\caption{Ускоренный метаалгоритм для метода $\mathcal{M}$ первого порядка}
\end{algorithm}
\vspace{0.5cm}
\begin{algorithm}[H] \label{cdm}
\SetAlgoLined
\textbf{Вход:} $y_0 \in \mathbb{R}^n$\;
\vspace{0.2cm}
$Z \leftarrow \sum_{i=1}^n (H + L_i)$\;
$p_i \leftarrow (H + L_i) / Z,\quad i \in \{1,...,n\}$\;
Дискретное вероятностное\\ распределение $\pi$
с вероятностями $p_i$\;
\vspace{0.2cm}
\For{k = 0, ..., $N-1$}
{
$i \sim \pi\{1,...,n\}$\;
$y_{k+1} \leftarrow y_k$\;
$\displaystyle y_{k+1}^i = y_k^i - \frac{1}{H + L_i} \nabla_i F(y_k)$\;
}
\Return $y_N$\;
\caption{Покомпонентный метод}
\end{algorithm}
\end{minipage}
\vspace{0.5cm}
\end{wrapfigure}
Прежде чем сформулировать какие-либо результаты о сходимости, следует подробнее рассмотреть вопрос о решении вспомогательной задачи~--- её аналитическое решение доступно лишь в редких случаях, и потому необходимо решать её численными методами, а значит неточно. Вспомогательную задачу допустимо решать до выполнения следующего условия останова (\cite{kamzolov2020optimal}, Appendix B):
\begin{equation} \label{monteiro}
\left\|\nabla\left\{F(y_\star) := f(y_\star) + \frac{H}{2}\|y_\star-\widetilde{x}_k\|^2_2\right\}\right\|_2 \leq
\end{equation}
$$\leq \frac{H}{2}\|y_\star-\widetilde{x}_k\|_2.$$
Ввиду того, что $\|\nabla F(y_*)\|_2 = 0$, а также ввиду $(L + H)$-липшицевости $\nabla F$, имеем:
\begin{equation} \label{smooth}
\|\nabla F(y_\star)\|_2 \leq (L + H) \|y_\star - y_*\|_2.
\end{equation}
Выписав неравенство треугольника:
$\|\widetilde{x}_k - y_*\|_2 - \|y_\star - y_*\|_2 \leq \|y_\star - \widetilde{x}_k\|_2$,
и воспользовавшись вместе неравенствами \eqref{monteiro}, \eqref{smooth}, получаем окончательное условие останова:
\begin{equation} \label{crit}
\|y_\star - y_*\|_2 \leq \frac{H}{3H + 2L} \|\widetilde{x}_k - y_*\|_2.
\end{equation}
Содержательно отсюда следует, что необходимая точность решения вспомогательной задачи по аргументу не зависит от требуемой точности решения общей задачи, что позволяет значимо упростить получение дальнейших результатов.
Рассмотрим теперь основной применяемый для решения вспомогательных задач метод: содержание покомпонентного метода \cite{nesterov2012efficiency} (в частном случае $\gamma = 1$) представлено в листинге \ref{cdm}. Для данного метода в случае рассматриваемых вспомогательных задач справедлив результат:
\begin{theorem}{(\cite{bubeck2014convex}, theorem 6.8)} \label{cdm_conv}
Пусть $F$ является $H$-сильно выпуклой относительно $\|\cdot\|_2$. Тогда для последовательности $\{y_k\}_{k=1}^N$, генерируемой покомпонентным методом, выполняется
\begin{equation} \label{f_conv}
\mathbb{E}[F(y_N)] - F(y_*) \leq \left(1 - \frac{1}{\kappa}\right)^N (F(y_0) - F(y_*)),\quad \text{где}\quad\kappa = \frac{H}{Z},\;\;Z = \sum_{i=1}^n (H + L_i).
\end{equation}
\end{theorem}
\noindentИспользуя данный результат, сформулируем утверждение о числе итераций покомпонентного метода, достаточном для выполнения полученного выше условия останова.
\begin{corollary} \label{cor_num}
Матожидание $\mathbb{E}[y_N]$ точки, являющейся результатом работы покомпонентного метода, удовлетворяет условию \eqref{crit} достижения достаточной точности решения вспомогательной задачи ускоренного метаалгоритма в том случае, если для числа итераций метода выполнено
\begin{equation}\label{cor}
N \geq N(\widetilde{\varepsilon}) = \ceil[\Bigg]{\frac{Z}{H} \ln{\left\{\left(1 + \frac{L}{H}\right) \left(3 + \frac{2L}{H}\right)^2\right\}} }, \quad \text{где}\quad\widetilde{\varepsilon} = \frac{H}{2} \left(\frac{H}{3H + 2L}\right)^2 \|y_0 - y_*\|_2^2.
\end{equation}
\end{corollary}
\begin{proof}
Ввиду $(H+L)$-липшицевой гладкости функции $F$ верно: $\displaystyle F(y_0) - F(y_*) \leq \frac{H+L}{2} \|y_0 - y_*\|_2^2$. Используя это неравенство вместе с оценкой \eqref{f_conv}, можем выписать условие достижения заданной точности $\widetilde{\varepsilon}$ по функции: $\displaystyle \frac{H+L}{2} \left(1 - \frac{1}{\kappa}\right)^{N} \|y_0 - y_*\|_2^2 \leq \widetilde{\varepsilon}$. Также верно $1 - 1/\kappa \leq \exp\{-1/\kappa\}$, и, значит:
$$\displaystyle \frac{H+L}{2} \exp\{\kappa / N\} \|y_0 - y_*\|_2^2 \leq \widetilde{\varepsilon}.$$
Логарифмируя и подставляя выражение для $\kappa$, получаем выражение для числа итераций от $\widetilde{\varepsilon}$:
\begin{equation} \label{itt}
N(\widetilde{\varepsilon}) = \ceil[\Bigg]{\frac{Z}{H} \ln {\left\{\frac{(H+L) \|y_0 - y_*\|_2^2}{2 \widetilde{\varepsilon}}\right\}}}.
\end{equation}
Ввиду $H$-сильной выпуклости функции $F$ верно $\displaystyle \|\overline{y}_N - y_*\|_2^2 \leq \frac{2}{H} (F(\overline{y}_N) - F(y_*))$, где $\overline{y}_N = \mathbb{E}[y_N]$. Функция $F$ выпукла, следовательно, по неравенству Йенсена, $F(\overline{y}_N) \leq \mathbb{E}[F(y_N)]$, и отсюда, вместе с \eqref{crit} получаем достаточное условие достижения решения вспомогательной задачи:
$$\mathbb{E}[F(y_N)] - F(y_*) \leq \frac{H}{2} \left(\frac{H}{3H + 2L}\right)^2 \|y_0 - y_*\|_2^2.$$
Подставляя в формулу \eqref{itt} вместо $\widetilde{\varepsilon}$ выражение из правой части данного неравенства, непосредственно приходим к выражению из утверждения.
\end{proof}
Теперь, когда полностью прояснён вопрос о требуемой точности и оракульной сложности решения вспомогательной задачи с помощью покомпонентного метода, можно перейти к результатам о сходимости ускоренного метаалгоритма. Для используемого условия останова \eqref{crit} метода, решающего вспомогательную задачу, справедлив следующий результат о сходимости ускоренного метаалгоритма:
\begin{theorem}{(\cite{gasnikov2020accelerated}, теорема 1)} \label{am_conv}
При $H > 0$, для последовательности $\{v_k\}_{k=1}^{\widetilde{N}}$, генерируемой ускоренным метаалгоритмом, использующим для решения вспомогательной задачи некоторый не стохастический метод, выполняется
\begin{equation} \label{am_th}
f(v_{\widetilde{N}}) - f(x_*) \leq \frac{48}{5} \frac{H \|x_0 - x_*\|_2^2}{\widetilde{N}^2}.
\end{equation}
\end{theorem}
\noindentНа основании последнего утверждения можно сформулировать теорему о сходимости ускоренного метаалгоритма в случае применения стохастического метода и, в частности, покомпонентного градиентного спуска.
\begin{theorem} \label{am_stoch_conv}
При $H > 0$, для некоторого $0 < \delta < 1$, точка $v_{\widetilde{N}}$, являющаяся результатом работы ускоренного метаалгоритма, использующего для решения вспомогательной задачи покомпонентный метод, решающий вспомогательную задачу $N_{\delta}$ итераций, удовлетворяет условию
\begin{equation*}
Pr(f(v_{\widetilde{N}}) - f(x_*) < \varepsilon) \geq 1 - \delta
\end{equation*}
в случае, если
\begin{equation} \label{iters_out_inn}
\widetilde{N} \geq \ceil[\Bigg]{\frac{4 \sqrt{15}}{5} \sqrt{\frac{H \|x_0 - x_*\|_2^2}{\varepsilon}}}, \quad N_{\delta} \geq N\left(\frac{\widetilde{\varepsilon}\delta}{\widetilde{N}}\right) = \ceil[\Bigg]{\frac{Z}{H} \ln { \left\{\frac{\widetilde{N}}{\delta} \left(1 + \frac{L}{H}\right) \left(3 + \frac{2L}{H}\right)^2\right\}} }.
\end{equation}
\end{theorem}
\begin{proof}
В следствии \ref{cor_num} представлена оценка числа итераций, достаточного для выполнения следующего условия на матожидание значения функционала в результирующей точке метода:
$$\mathbb{E}[F(y_{N(\widetilde{\varepsilon})})] - F(y_*) \leq \widetilde{\varepsilon}.$$
Воспользуемся неравенством Маркова и получим формулировку данного условия в терминах оценки вероятности больших отклонений \cite{anikin2015modern}: заведомо выберем допустимое значение вероятности невыполнения поставленного условия, так чтобы $0 < \delta/\widetilde{N} < 1$, где $\widetilde{N}$ выражается из \eqref{am_th}; тогда
$$Pr\left(F\left(y_{N(\widetilde{\varepsilon} \delta/\widetilde{N})}\right) - F(y_*) \geq \widetilde{\varepsilon}\right) \leq \frac{\delta}{\widetilde{N}} \cdot \frac{\mathbb{E}\left[F\left(y_{N(\widetilde{\varepsilon} \delta/\widetilde{N})}\right)\right] - F(y_*)}{\widetilde{\varepsilon} \cdot \delta/\widetilde{N}} = \frac{\delta}{\widetilde{N}}.$$
Поскольку вероятность того, что полученное решение некоторой отдельно взятой вспомогательной задачи не будет удовлетворять поставленному условию, равна $\delta / \widetilde{N}$, значит вероятность того, что за $\widetilde{N}$ итераций ускоренного метаалгоритма условие будет невыполнено хотя бы для одной из задач, есть $\widetilde{N} \cdot \delta / \widetilde{N} = \delta$, откуда и следует доказываемое утверждение.
\end{proof}
Далее, объединяя оценки, приводимые в теореме \ref{am_stoch_conv}, можем получить асимптотическую оценку на общее число итераций покомпонентного метода, достаточное для решения рассматриваемой оптимизационной задачи с некоторой заданной точностью, а также оценку оптимального параметра $H$:
\begin{corollary}
Для того чтобы точка $v_{\widetilde{N}}$, являющаяся результатом работы ускоренного метаалгоритма, удовлетворяла условию
$$Pr(f(v_{\widetilde{N}}) - f(x_*) < \varepsilon) \geq 1 - \delta,$$
достаточно выполнить в сумме
\begin{equation} \label{total_iters}
\hat{N} \geq \widetilde{N} \cdot N_\delta = \mathcal{O}\left(\frac{Z \|x_0 - x_*\|_2}{\sqrt{H}} \cdot \frac{1}{\varepsilon^{1/2}} \log\left\{\frac{1}{\varepsilon^{1/2} \delta}\right\}\right)
\end{equation}
итераций покомпонентного метода для решения вспомогательной задачи. При этом, оптимально значение параметра $H$ регуляризации вспомогательной задачи следует выбирать как $H \simeq \frac{1}{n} \sum_{i=1}^n L_i$ ($\simeq$ обозначает равенство с точностью до малого множителя порядка $\log$).
\end{corollary}
\begin{proof}
Выражение для $\hat{N}$ можно получить путём прямой подстановки одной из оценок, приводимых в \eqref{iters_out_inn}, в другую, и их последующего умножения. Если исключить из рассмотрения малый множитель порядка $\log(L / H)$, константа в оценке будет зависеть от $H$ как:
$$\sqrt{H} \cdot \frac{Z/n}{H} = \sqrt{H} \left(1 + \frac{\frac{1}{n} \sum_{i=1}^n L_i}{H}\right).$$
Минимизируя представленное выражение по $H$, получаем указанный результат.
\end{proof}
Рассмотрим теперь подробнее вопрос об алгоритмической сложности предложенного ускоренного метода покомпонентного градиентного спуска. Очевидным является следующее утверждение:
\begin{proposition} \label{constr}
Алгоритмическая сложность рассматриваемого метода составляет
$$T = \mathcal{O}\left(\widetilde{N} (T_{out} + N_{\delta} T_{inn})\right),$$
где $T_{out}$~--- амортизированная оценка сложности вычислений, производимых на итерации ускоренного метаалгоритма, $T_{inn}$~--- амортизированная оценка сложности итерации покомпонентного метода.
\end{proposition}
\noindentОттолкнувшись от него, сформулируем результат о вычислительной сложности метода:
\begin{theorem}
Пусть сложность вычисления одной компоненты градиентна $f$ составляет $\mathcal{O}(s)$. Тогда алгоритмическая сложность рассматриваемого метода есть
$$T = \mathcal{O}\left(sn \cdot \sqrt{\frac{\overline{L} \|x_0 - x_*\|_2^2}{\varepsilon}} \log\left\{\frac{1}{\varepsilon^{1/2} \delta}\right\}\right),\quad\text{где}\quad \overline{L} = Z / n = \frac{1}{n} \sum_{i=1}^n L_i.$$
\end{theorem}
\begin{proof}
Переформулируем оценку из утверждения \ref{constr} следующим образом:
$$T = \mathcal{O}\left(\hat{N} \cdot T_{iter}\right),$$
где $T_{iter}$~--- амортизированная оценка сложности элементарной итерации ускоренного метаалгоритма, то есть итерации, которая может быть и внутренней итерацией покомпонентного метода, и основной итерацией метаалгоритма. Сложность основной итерации метаалгоритма определяется в первую очередь вычислением полного градиента $f$, а сложность этой процедуры (из условия теоремы) есть $\mathcal{O}(sn)$. В то же время, ввиду $Z = n \overline{L}$, верно также $N_\delta = \widetilde{\mathcal{O}}(n)$, где символ $\widetilde{\mathcal{O}}(\cdot)$ означает то же, что и $\mathcal{O}(\cdot)$, но с возможным присутствием множителей порядка $\log(\cdot)$. Поскольку основная итерация метаалгоритма исполняется через каждые $N_\delta$ элементарных итераций, где $N_\delta$ постоянно, из этого с помощью любого из методов амортизационного анализа тривиально получается амортизированная оценка сложности основной итерации метаалгоритма, составляющая $\mathcal{O}(s)$. Сложность итерации покомпонентного метода (если вместо копирования значений точки выполнять операции ``на месте``, что для данной конструкции вполне допустимо) определяется вычислением одной компоненты градиента, и составляет также $\mathcal{O}(s)$. Отсюда получаем $T_{iter} = \mathcal{O}(s)$. Используя оценку \eqref{total_iters} и подставляя оптимальное значение $H$, получаем приведённую сложность метода.
Заметим также, что сложность метода по памяти при этом составляет $\mathcal{O}(n)$, также как и сложность предварительных вычислений (для покомпонентного метода нет необходимости выполнять их каждый раз заново).
\end{proof}
Сравним полученные для предложенного метода оценки с оценками других методов, которые могут быть использованы для решения задач в описываемой постановке: быстрого градиентного метода (FGM), классического покомпонентного спуска (CDM), ускоренного покомпонентного спуска в варианте Ю.Е.~Нестерова (ACDM) и предложенного в данной работе подхода (Catalyst CDM). Оценки приведены в таблице \ref{table_compare}. Как можно видеть из приведённых асимптотических оценок вычислительной сложности, предложенный метод позволяет достигать эффективной в отношении характера зависимости от размерности задачи $n$ и требуемой точности $\varepsilon$ скорости сходимости, не уступающей другим методам, при некоторой плате за это в виде логарифмического множителя в оценке.
\begin{table}[ht]
\centering
\begin{tabular}{|c|c|c|c|}
\hline
Метод & Ит. сложность & Выч. сложность & Источник \\ \hline
FGM & $ \mathcal{O}\left(sn\right)$ & $\displaystyle \mathcal{O}\left(sn \cdot \frac{1}{\varepsilon^{1/2}} \cdot \sqrt{L}\right)$ & \cite{nesterov2018lectures} \\ \hline
CDM & $ \mathcal{O}\left(s\right)$ & $\displaystyle \mathcal{O}\left(sn \cdot \frac{1}{\varepsilon} \cdot \overline{L}\right)$ & \cite{bubeck2014convex} \\ \hline
ACDM & $ \mathcal{O}\left(n\right)$ & $\displaystyle \mathcal{O}\left(n^2 \cdot \frac{1}{\varepsilon^{1/2}} \cdot \sqrt{\widetilde{L}}\right)$ & \cite{nesterov2017efficiency} \\ \hline
Catalyst CDM & $ \mathcal{O}\left(s\right)$ & $\displaystyle \widetilde{\mathcal{O}}\left(sn \cdot \frac{1}{\varepsilon^{1/2}} \cdot \sqrt{\overline{L}}\right)$ & данная работа \\ \hline
\end{tabular}
\caption{Сравнение эффективности методов}
\label{table_compare}
\end{table}
Заметим, кроме того, что несмотря на существенное сходство оценок, между двумя наиболее эффективными методами в таблице (FGM и Catalyst CDM) существует также различие в константах, характеризующих гладкость функции: $L$~--- в FGM и $\overline{L}$~--- в Catalyst CDM, тем самым поведение рассматриваемого метода для различных задач напрямую зависит от характера их покомпонентной гладкости. В общем случае, нельзя утверждать, что одна из констант асимптотически существенно выгоднее другой, однако в ряде частных случаев возможно выписать оценки значений констант явно, и часто оказывается, что $\overline{L}$ ``меньше`` $L$. Наиболее существенен выигрыш в том случае, если справедливы соотношения $L=\mathcal{O}(n)$, $\overline{L} = \mathcal{O}(1)$~--- тогда в оценке вычислительной сложности предлагаемого метода удаётся редуцировать фактор порядка $\mathcal{O}(\sqrt{n})$ по сравнению с быстрым градиентным методом. В следующем разделе будет рассмотрен пример постановки задачи, в которой данный случай имеет место.
\section{Численные эксперименты} \label{section2}
В данном разделе описывается характер практического поведения метода, на примере следующей оптимизационной задачи для функционала, имеющего вид SoftMax-функции:
\begin{equation}\label{eq:softmax}
\min\limits_{x \in \mathbb{R}^n}~~ \{f(x) = \gamma \ln \left( \sum_{j=1}^m \exp\left(\frac{\left[A x\right]_j}{\gamma}\right) \right) -\langle b,x \rangle\},
\end{equation}
где $b \in \mathbb{R}^n$, $A \in \mathbb{R}^{m \times n}$. Задачи такого вида существенно важны для многих приложений, в частности, они возникают в задачах энтропийно-линейного программирования в качестве двойственной задачи \cite{chernov2016method, gasnikov2016effective}, в том числе в задаче оптимального транспорта, а также исполняют роль сглаженной аппроксимации функции $\max$ (что и дало функционалу название SoftMax) и, соответственно, нормы $\|\cdot\|_{\infty}$, что может быть востребовано в некоторых постановках задачи PageRank или при решении СЛУ. Притом во всех описанных задачах важным частным случаем являются разреженные постановки, в случае которых матрица $A$ разрежена, то есть такова, что среднее число ненулевых элементов в строке $A_j$ не превосходит некоторого $s \ll n$ (будет удобно также предполагать возможность для одной из строк $A_j$ являться полностью неразреженной).
Сформулируем свойства, которыми обладает функция $f$\; \cite{gasnikov2018modern}:\mynobreakpar\vspace{0.2cm}
\begin{enumerate}
\item $f$ дифференцируема;\mynobreakpar
\item $\nabla f$ удовлетворяет условию Липшица с константой $L = \max_{j=1,...,m} \|A_j\|_2^2$;\mynobreakpar
\item $\nabla_i f$ удовлетворяют покомпонентному условию Липшица с константами $L_i = \max_{j=1,...,m} |A_{j i}|$.
\end{enumerate}
Начнём уточнение свойств с первого пункта. Выпишем выражение для $i$-той компоненты градиента данного функционала:
$$\nabla_i f(x) = \frac{\sum_{j=1}^m{A_{j i} \exp\left(\left[A x\right]_j\right)}}{\sum_{j=1}^m{\exp\left(\left[A x\right]_j\right)}}.$$
Как можно видеть, наивное вычисление этого выражения может занимать время, сравнимое с вычислением градиента в целом, что будет значительно влиять на вычислительную сложность, а значит и на время работы метода. Однако в то же время, многие члены в этом выражении могут перевычисляться либо редко, либо покомпонентно, и использоваться при совершении шага метода как члены его дополнительной последовательности, так что сложность итерации будет оставаться эффективной, и применение покомпонентного метода будет оправдано. Для удобства описания используемых вычислительных приёмов, запишем шаг покомпонентного метода в виде:
$$y_{k+1} = y_{k} + \delta e_i,$$
где $\delta$~--- размер шага, умноженный на соответствующую компоненту градиента, $e_i$~--- $i$-тый орт базиса.
\begin{enumerate}
\item Будем хранить набор значений $\left\{\exp\left(\left[A y_k\right]_j\right)\right\}_{j=1}^m$, использующихся для вычисления суммы в числителе. Обновление этих значений после осуществления шага метода имеет сложность $\mathcal{O}(s)$, ввиду того что $A y_{k+1} = A y_k + \delta A_i$, и вместе с тем $A_i$ имеет не более $s$ ненулевых компонент, а значит потребуется вычисление $s$ корректирующих множителей и умножение на них соответствующих значений из набора.
\item Как можно видеть из первого пункта, производить умножение разреженных векторов следует за $\mathcal{O}(s)$, учитывая лишь ненулевые компоненты. В смысле программной реализации это означает необходимость использования для кэшируемых значений и для строк матрицы $A$ разреженного представления, то есть хранение лишь пар индекс-значение ненулевых элементов~--- тогда, очевидно, сложность арифметических операций для таких векторов будет пропорциональна сложности цикла с элементарными арифметическими операциями, число итераций которого равно числу ненулевых элементов (в языке программирования python, например, такой формат хранения реализуется в методе scipy.sparse.csr\_matrix \cite{scipy}).
\item Аналогично, будем хранить значение $\sum_{j=1}^m{\exp\left(\left[A y_k\right]_j\right)}$, являющееся знаменателем. Его обновление осуществляется с той же сложностью, что и обновление последовательности из пункта 1 (путём вычисления суммы ненулевых слагаемых, прибавляемых к каждому значению из набора).
\item Поскольку вычисление указанного выражения требует вычисления значений экспонент, может происходить переполнение типов. Для решения этой проблемы стандартно применяется exp-normalize trick \cite{blanchard2019accurately}. Однако для его применения следует также хранить значение $\max_{j=1,...,m} \left[A y_k\right]_j$. Вместе с тем, нет необходимости знать именно это значение, или, иначе, знать его точно~--- достаточно лишь его приближения, чтобы значения в показателях экспонент были малы, так что перевычислять эту величину можно гораздо реже: например, раз в $m$ итераций метода, в результате чего амортизированная сложность будет равна $\mathcal{O}(s)$.
\end{enumerate}
Итак, в дальнейших рассуждениях можно полагать, что итерация покомпонентного метода при решении соответствующей вспомогательной задачи, имеет амортизированную сложность $\mathcal{O}(s)$.
Далее, рассмотрим подробнее вопрос о величине констант гладкости данного функционала. Можно выписать асимптотические формулы для $L$ и $\overline{L} = \frac{1}{n} \sum_{i=1}^n L_i$:
$$L = \max_{j=1,...,m} \|A_j\|_2^2 = \mathcal{O}(n), \quad \overline{L} = \frac{1}{n} \sum_{i=1}^n \max_{j=1,...,m} |A_{j i}| = \mathcal{O}(1).$$
Используя эти оценки, уточним вычислительную сложность методов FGM и Catalyst CDM в применении к данной задаче:
$$T_{FGM} = \mathcal{O}\left(sn^{3/2} \cdot \frac{1}{\varepsilon^{1/2}}\right), \quad T_{CCDM} = \widetilde{\mathcal{O}}\left(sn \cdot \frac{1}{\varepsilon^{1/2}}\right).$$
Таким образом, в теории, применение метода Catalyst CDM для решения данной задачи позволяет, по сравнению с FGM, редуцировать в асимптотической оценке вычислительной сложности множитель порядка $\mathcal{O}(\sqrt{n})$. Практически, это означает, что предложенный метод разумно применять для задач большой размерности.
Сравним теперь работу предложеннного в статье метода (Catalyst CDM) с рядом альтернативных подходов: градиентным спуском (GM), быстрым градиентным методом (FGM), покомпонентным спуском (CDM) и ускоренным покомпонентным спуском (ACDM), на примере задачи~\eqref{eq:softmax} с искусственно сгенерированной двумя различными способами матрицей $A$. На рис. \ref{fig:umc1} и \ref{fig:umc2} представлены графики сходимости рассматриваемых методов: по оси абсцисс отложено время работы методов в секундах, по оси ординат~--- невязка по функции в логарифмическом масштабе ($f_*$ находится путём поиска соответствующей точки $x_*$ с помощью метода FGM, настроенного на точность, заведомо значительно превосходящую возможную для достижения на выбранном временном промежутке).
\begin{figure}[!ht]
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/results_test.pdf}
\caption{Сходимость методов для задачи SoftMax~\eqref{eq:softmax} с равномерно разреженной случайной матрицей.}
\label{fig:umc1}
\end{minipage}
\hfill
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/results_good.pdf}
\caption{Сходимость методов для задачи SoftMax~\eqref{eq:softmax} с неоднородно разреженной случайной матрицей.}
\label{fig:umc2}
\end{minipage}
\end{figure}
На рис. \ref{fig:umc1} представлен случай, для которого все элементы матрицы $A$ являются н.о.р. случайными величинами из дискретного равномерного распределения $A_{j i} \in \mathcal{U}\{0, 1\}$, при этом число ненулевых элементов составляет $s \approx 0.2 m$, и параметр $\gamma=0.6$ (так же как и во втором случае). В такой постановке предложенный метод демонстрирует более быструю сходимость по сравнению со всеми сравниваемыми методами, за исключением FGM. В то же время, в постановке, отражённой на рис. \ref{fig:umc2}, при которой число ненулевых элементов, по сравнению с первым случаем, увеличено до $s \approx 0.75 m$, а матрица генерируется неравномерно в соответствии с правилом: $0.9 m$ строк с $0.1 n$ ненулевых элементов и $0.1 m$ строк с $0.9 n$ ненулевых элементов, а также одна и строк матрицы является полностью неразреженной, предложенный метод сходится быстрее FGM. Это объясняется тем, что в этом случае $L = n$, тогда как $\overline{L}$ по прежнему остаётся достаточно мало, в результате чего константа в предложенном методе оказывает заметно меньшее влияние на вычислительную сложность, чем в случае FGM. Из результатов эксперимента также можно отметить, что гораздо существеннее степени разреженности задачи на эффективность предложенного метода влияет характер её покомпонентной гладкости.
\newpage
\section{Заключение}
В данной работе предлагается вариант покомпонентного метода, ускоренного с помощью универсальной проксимальной оболочки ``ускоренный метаалгоритм``. Проведённый теоретический анализ предложенного метода позволяет утверждать, что зависимость его вычислительной сложности от размерности задачи и требуемой точности решения не уступает прочим методам, используемым для оптимизации выпуклых липшицево гладких функций, а оценка вычислительной сложности сравнима с оценкой быстрого градиентного метода. При этом, в предложенной схеме сохраняются свойства, характерные для классического покомпонентного метода, в том числе возможность использования свойств покомпонентной гладкости функции. Приведённые численные эксперименты подтверждают практическую эффективность метода, и также подчёркивают особенную релевантность предложенного подхода для часто возникающей в различных приложениях задачи оптимизации функции вида SoftMax.
Данная статья представляет результат работы над проектом, предложенным А.В.~Гасниковым в рамках проектной смены\footnote{Ссылка на сайт проектной смены: \url{https://sochisirius.ru/obuchenie/graduates/smena673/3258}} "Современные методы теории информации, оптимизации и управления"\;Сириус 2-23~августа~2020 г. Авторы выражают благодарность организатору проектной смены А.С.~Ненашеву за создание комфортных условий для работы.
\bibliographystyle{plain}
|
\section{#1}
}
\newcommand{\sect}[1] {
\subsection{#1}
}
\newcommand{\subsect}[1] {
\subsubsection{#1}
}
\newcommand{\subsubsect}[1] {
\paragraph{#1}
}
\section*{Acknowledgments}
We would like to thank Maurizio Nitti for generating the current state-of-the-art results to compare against the newly developed system, and a special thanks to those who appeared in our videos: Iliyan Georgiev, Antoine Milliez, Sarah Pellkofer, Simon Heinzle, Pierre Greisen, and Selina the cat.
\bibliographystyle{eg-alpha}
|
\section{Introduction}
\begin{figure}
\centering
\begin{subfigure}[b]{0.9\linewidth}
\centering
\includegraphics[width=0.8\linewidth]{figures/initial_close.png}
\caption{}
\label{fig:Ng1}
\end{subfigure}
\begin{subfigure}[b]{0.9\linewidth}
\centering
\includegraphics[width=0.8\linewidth]{figures/close_final3.png}
\caption{}
\label{fig:Ng2}
\end{subfigure}
\caption{A robot using our method to autonomously learn to close a microwave. (a) The robot uses a motion planner to generate an initial attempt at closing the microwave door using a kinematic model of the microwave. The resulting plan is unable to fully close the microwave door because of the robot's limited reach. (b) After bootstrapping a motor skill with the trajectory from (a), the robot learns a motor skill that gives the door a push, exploiting its dynamics to fully close the microwave.
}
\label{demopic}
\end{figure}
Robots require motor policies for interacting with objects in their environment. For example, a robot butler may need a motor skill that enables it to open a drawer to fetch utensils for a table, for setting each element of the table, and for pouring wine.
While it is safe to assume that a robot will have an accurate kinematic model of its own body, it is unlikely to have a dynamics model of every object it will ever encounter. This lack of knowledge means that the robot will have to learn how to manipulate the world around it \cite{kroemer2019review}.
Reinforcement learning (RL) provides a framework for robots to acquire motor policies without explicitly modeling the unknown world, but model-free RL methods like policy search \cite{deisenroth2013survey} have high sample-complexity, and often fail to learn a reasonable policy from random initialization. Supervised approaches for policy learning like Learning From Demonstration (LfD) \cite{argall2009survey} can encode human prior knowledge by imitating expert examples, but do not support optimization in new environments. Combining RL with LfD is a powerful method for reducing the sample complexity of policy search, and is often used in practice \cite{levine2013guided, rajeswaran2017learning,Zhu-RSS-18,2018arXiv180510413C}. However, this approach typically requires a human demonstrator for initialization, which fundamentally limits the autonomy, and therefore utility, of a robot which may need to acquire a wide range of motor skills over its operational lifetime. More recently, model-based control techniques (including Model Predictive Control \cite{kahn2017plato, pan2017agile} and LQR \cite{levine2013guided}) have been proposed as exploration methods for policy search; these methods still require human demonstrations or complete dynamic models of both the robot and every object in the scene.
We propose the use of kinematic motion planning to initialize motor skill policies. While previous work have leveraged sample-based motion planners for learning motor skills \cite{tosun2019pixels,jurgenson2019harnessing,jiang2019task}, they only focus on either free-space motions or do not learn a closed-loop controller. To our knowledge, this is the first use of motion planning to provide initial demonstrations for learning closed-loop motor skill policies by leveraging estimated object kinematics.
Motion planning algorithms
generate collision-free behavior and generalize to novel scenarios when the robot has a good kinematic model of itself and the object it aims to manipulate, making it useful for tasks like pick-and-place. We show that given a (potentially approximate, and readily estimated) kinematic description of the environment and the robot, off-the-shelf motion planning algorithms can generate feasible (potentially successful but inefficient) initial trajectories (Figure \ref{fig:Ng1}) to bootstrap an object-manipulation policy that can subsequently be optimized using policy search (Figure \ref{fig:Ng2}). This framework enables the robot to automatically produce its own demonstrations for effectively learning and refining object manipulation policies. Our work enables the robot to realize the benefits of an initial demonstration fully automatically using kinematic planning, requiring no human involvement.
To evaluate our method, we used two different motor policy classes (Dynamic Movement Primitives (DMPs) \cite{ijspeert2003learning} and deep neural networks \cite{JMLR:v17:15-522}). We chose these two different motor policy classes because deep neural networks are extremely expressive in what policies they can represent, but are extremely sample inefficient compared to structured motor primitives like DMPs, and we aim to evaluate how our method performs in both contexts. Using these motor policy classes, we compared bootstrapping with motion planning against learning from scratch in three simulated experiments, and against human demonstrations in real hardware experiments. Human demonstrations provide a baseline for how effective these motor policies can do when bootstrapped with high-quality demonstrations, and learning from scratch provides a baseline for how difficult the task is without any prior information about the task
In the first two experiments---opening a drawer with a dynamic movement primitive representation, and closing a microwave door with a deep neural network policy---we show that motion planning using a kinematic model produces a reasonable initial policy,although suboptimal compared to a supervised human demonstration, that learning adapts to generate efficient, dynamic policies that exploit the dynamics of the object being manipulated.
We also show how our method can bootstrap a motor skill for the challenging dynamic task of learning to hit a ball off a tee, which involves precise and agile movement. In that case, treating the objects in the scene as static and applying kinematic motion planning succeeds in generating a policy that makes contact with the ball, which is sufficient to bootstrap a more dynamic policy that learns to hit the ball several feet. Our method is competitive with human-demonstrated initialization, but requires no human demonstration. It serves as a suitable starting point for learning, and significantly outperforms starting with a random policy. This approach enables robots to efficiently and autonomously learn motor policies for dynamic tasks without human demonstration.
In summary, our contributions are:
\begin{enumerate}
\setlength{\itemsep}{0pt}%
\setlength{\parskip}{0pt}%
\item A fully-autonomous paradigm for policy search, in which an agent first uses goal-directed kinematic planning to devise feasible solution trajectories for itself and objects in a scene.
\item A novel algorithm that autonomously generates a set of initial demonstrations for object manipulation via kinematic planning in object and robot configuration spaces.
\item An empirical evaluation of this algorithm in which we employ model-free policy optimization after bootstrapping, demonstrating that our method's performance is comparable to human expert demonstration and superior to random initialization in hardware and simulation tasks: closing a microwave, opening a drawer, and hitting a ball off a tee.
\end{enumerate}
\section{BACKGROUND}
Our goal is to efficiently and autonomously learn robot motor skill policies. To do so, we develop an approach that uses kinematic motion planning to generate initial trajectories, fits a policy to those trajectories using behavioral cloning, and subsequently optimizes that policy via policy search. We now briefly describe policy search, policy representations, learning from demonstration, and motion planning.
\subsection{Policy Search}
Policy search methods \cite{deisenroth2013survey} are a family of model-free reinforcement learning algorithms that search within a parametric class of policies to maximize reward. Formally, given a Markov Decision Process $\mathcal{M} = \langle \mathcal{S},\mathcal{A},\mathcal{R},\mathcal{T},\gamma \rangle$, the objective of policy search is to maximize the expected return of the policy $\pi_\theta$:
\begin{equation}
\max_\theta \mathop{\large{\mathbb{E}}}_{\mathcal{M},\pi_\theta} \left[ \sum_{t=0}^{T} \gamma^t r_t \right]. \label{eq: rl}
\end{equation}
These approaches can learn motor skills through interaction, and therefore do not require an explicit environment model, and are typically agnostic to the choice of policy class (though their success often depends on the policy class having the right balance of expressiveness and compactness). However, their model-free nature leads to high sample complexity, which often makes them infeasible to apply directly to robot learning problems.
\textbf{Policy representation} describes the class of functions used as the mapping from states to actions.
Our approach is agnostic to policy representation; we demonstrate our approach enables efficient learning using two different common policy representations: Dynamic Movement Primitives and neural network controllers.
Dynamic Movement Primitives \cite{ijspeert2003learning, peters2008natural} are a description of a non-linear second-order differential equation that exhibits attractor dynamics modulated by a learnable forcing function.
DMPs are a popular representation for motor policies because they are parameter-efficient, can express both point and limit cycle attractors, enable real-time computation, and exhibit temporal invariance that does not effect the attractor landscape.
We refer the reader to the work of \citet{ijspeert2003learning} for a more formal introduction to DMPs.
If we have $n$ joints we wish to control, we can model control for each joint independently with $n$ DMPs for each one.
Therefore, the multiple joints of a robot are only coupled through time, which makes this representation very compact.
Neural network controllers have received significant attention in recent years; they are able to learn hierarchical feature representations for approximating functions (in our case, motor skills) operating on high-dimensional input such as robot sensor data. They are more expressive than restricted policy classes such as DMPs and can operate directly on high-dimensional state spaces (e.g. images), yet they typically exhibit higher sample complexity \cite{JMLR:v17:15-522}.
\textbf{Learning from Demonstration} methods
\cite{argall2009survey, ravichandar2019recent} broadly consist of two families of approaches that either mimic (Behavioral Cloning) or generalize (Inverse Reinforcement Learning) the exemplified behavior. Inverse reinforcement learning methods seek to estimate a latent reward signal from a set of demonstrations; we assume a given reward function, and omit a discussion of inverse reinforcement learning methods here.
Behavioral cloning methods
\cite{atkeson1997robot, pastor2009learning, ho2016generative} attempt to directly learn a policy that reproduces the demonstrated policies.
Given a dataset of expert demonstrations $D$, the objective of behavioral cloning is:
$\max_\theta \sum_{(s,a) \in D} \pi_\theta(a | s).$
These methods often result in policies with undesirable behavior in states not observed during demonstrations, though this can addressed with interactive learning \cite{ross2011reduction, nair2017combining, sun2017deeply}. In our approach, the existence of a reward function enables the agent to learn robust behavior in states outside of the initial training distribution. Moreover, our experiments demonstrate our approach's ability to extrapolate beyond suboptimal initial demonstrations.
Many approaches investigate the incorporation of human-provided demonstrations into policy search to drastically reduce sample complexity via a reasonable initial policy and/or the integration of demonstrations in the learning objective \cite{peters2008natural, kang2018policy, 2018arXiv180510413C, rajeswaran2017learning, Zhu-RSS-18, vecerik2017leveraging, levine2013guided}.
\subsection{Motion Planning}
The pose of an articulated rigid body can defined by the state of each of its movable joints. The space of these poses is called the configuration space $\mathcal{C}$ \cite{lozano1981automatic}. Motion planning is the problem of finding a path (sequence of poses) through configuration space such that the articulated object is moved to a desired goal configuration, without encountering a collision.
While there exist many different families of motion planning algorithms, such as geometric, grid-based, and probabilistic road maps \cite{lavalle2006planning}, they all operate in a similar fashion: given a configuration space $\mathcal{C}$ and start and goal joint configurations $q_{0}, q^* \in \mathcal{C}$, return a valid path of joint configurations $\{q_{t}\}_{t=0}^{T}$ between the start and end configurations. We focus on sample-based motion planning approaches.
Probabilistic motion planners provide a principled approach for quickly generating collision-free robot trajectories.
However, online replanning is expensive, and kinematic motion planners are only as effective as their kinematic models are accurate: they generate trajectories directly, and thus cannot be improved through subsequent interaction and learning. Furthermore, kinematic planners produce trajectories that only account for kinematics, not dynamics: they explicitly do not account for forces involved in motion, such as friction, inertial forces, motor torques, etc, which are important for effectively performing contact-rich, dexterous manipulation.
The process of computing the position and orientation $p \in SE(3)$ of a link in a kinematic chain for a given joint variable setting (a point in configuration space) is termed \textit{forward kinematics}. Inversely, computing a configuration to attain a specific end effector pose $p$ is termed \textit{inverse kinematics}. We denote the forward kinematics functions $p=f(q)$.
\section{Bootstrapping Skill Learning with Motion Planning}
Our methodology is inspired by how humans generate reasonable first attempts for accomplishing new motor tasks. When a human wants to learn a motor skill, they do not start by flailing their arms around in a random fashion, nor do they require another person to guide their arms through a demonstration. Instead, they make a rough estimate of how they want an object to move and then try to manipulate it to that goal. For example, before being able to drive stick shift, a human must first learn how to manipulate a gear shifter for their car. Just by looking at the gear shifter, humans can decide (1) what they should grab (the shaft), (2) where they want the shaft to go (positioned in a gear location), and (3) how the shaft should roughly move throughout the action (at the intermediate gear positions). Similarly, a robot that has a good kinematic model of itself, and a reasonable kinematic model of the object it wishes to manipulate, should be able to form a motion plan to achieve the effect it wishes to achieve.
That plan may be inadequate in several ways: its kinematic model may be inaccurate, so the plan does not work; object dynamics (like the weight of a door, or the friction of a joint) may matter, and these are not represented in a kinematic model; and a feasible and collision-free kinematic trajectory may not actually have the desired effect when executed on a robot interacting with a real (and possibly novel) object. But such a solution is a \textit{good start}; we therefore propose to use it to bootstrap motor skill learning.
Our approach, outlined in Figure \ref{fig:overview}, leverages the (partial) knowledge the robot has about its own body and the object it is manipulating to bootstrap motor skills. Our method first assumes access to the configuration space of the robot, denoted as $\mathcal{C}_{R}$, as well as its inverse kinematics function $f^{-1}_R$. This assumption is aligned with the fact that the robot often has an accurate description of its own links and joints and how they are configured during deployment. However, the world is comprised of objects with degrees of freedom that can only be inferred from sensor data. Therefore, our approach only assumes access to estimated kinematics of the object to be manipulated, in the form of configuration space $\mathcal{C}_{O}$ and forward kinematics $f_O$. Recent work has shown that estimating these quantities for novel objects from sensor data in real environments is feasible \cite{benny, li2019category}, though state-of-the-art estimates still include noise.
Finally, our approach assumes that the task goal can be defined in terms of kinematic states of the robot and environment. Examples of such tasks include pick-and-place, articulated object manipulation, and many instances of tool use. (Note that this requirement fails to capture reward functions defined in terms of force, for example exerting a specific amount of force in a target location.) Such a goal, together with object and robot kinematic descriptions, enables us to autonomously generate useful initial trajectories for policy search.
Our approach is outlined in Algorithm \ref{algoloop}, and can broken down into five main steps: 1) collect initial trajectories(s) from a motion planner using estimated object kinematics, 2) fit a policy with these initial trajectories, 3) gather rollouts to sample rewards for the current policy based on the kinematic goal, and 4) update the policy parameters based on the actions and rewards, 5) repeat steps 3-4.
\begin{algorithm}
\caption{Planning for Policy Bootstrapping}
\begin{algorithmic}[1]
\Procedure{PPB}{$C_{R}, f^{-1}_R, C_{O}, f_{O}, q_{O}^*$}
\State $D \leftarrow \varnothing$
\For{$0$ to $N$}
\State $D \leftarrow$ InitialMPDemos($C_{R},f^{-1}_R, C_{O},f_{O}, q_{O}^*$) $\cup$ $D$
\EndFor
\State $\theta \leftarrow$ FitPolicy($D_{0},...,D_{N}$)
\For{$0$ to $E$}
\State $T_{0},..,T_{n} \leftarrow$ Rollout($\pi,\theta,q_{O}^*$)
\State $\theta \leftarrow$ UpdatePolicy($T_{1},..,T_{n},\theta$)
\EndFor
\EndProcedure
\end{algorithmic}
\label{algoloop}
\end{algorithm}
\begin{algorithm}
\caption{Initial Motion Plan Demos }
\begin{algorithmic}[1]
\Procedure{InitialMPDemos}{$C_{R}, f^{-1}_R, C_{O},f_{O}, q_{O}^*$}
\State $T_{O} \leftarrow$ MotionPlanner($C_{O}, q_{O}^*$)
\State $g \leftarrow$ EstimateGrasp($C_{O}$,$f_{O}$)
\State $eepath \leftarrow$ GraspPath($T_{O},C_{O},f_{O}, g$)
\State $T_{R} \leftarrow$ MotionPlanner($C_{R}$, $eepath$, $f^{-1}_R$)
\State return $T_{R}$
\EndProcedure
\end{algorithmic}
\label{mpalgo}
\end{algorithm}
\subsection{Initial Trajectories from Motion Planner}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/iros2020-system-overview.png}
\caption{\textbf{System overview} illustrating our proposed framework for generating demonstrations with a motion planner and subsequently performing policy search. The dashed box contains the steps from Algorithm 2.
}
\label{fig:overview}
\end{figure}
To fit our policy class, we choose to sample $N$ demonstrations using a motion planner. Our approach for collecting initial demonstrations with a kinematic motion planner is outlined in Algorithm \ref{mpalgo}, and proceeds as follows. First, we use a motion planner to find a path through the object's configuration space $C_{O}$ that moves the object from its initial state to a goal state $q_{O}^*$ using an off-the-shelf motion planner. This produces a joint trajectory in object configuration space, $T_{O}$, which transforms the object from its current joint configuration to the desired one.
We then estimate a grasp point on the object to designate the contact point for the robot during manipulation. This can be done by either generating candidate grasps using off-the-shelf grasping algorithms \cite{gualtieri2016high, mahler2017dex} or choosing a part semantically. This produces a local 6D pose, $g$, that represents where the robot should grasp the object during manipulation.
We then use the grasp point $g$, object joint trajectory $T_{O}$, and the object's forward kinematics $f_{O}$ to generate the series of 6D Cartesian poses that the grasp point $g$ will go through as the object proceeds through $T_{O}$. This produces a series of 6D Cartesian poses, $eepath$, which the robot end effector must go through, assuming a fixed grasp pose to the object.
Finally, we solve for a path in robot joint space that achieves the end effector path in Cartesian space using off-the-shelf sample-based motion planners, using the robot's inverse kinematics $f^{-1}_R$ and the sequence of end-effector poses $eepath$.
Note that in our experiments, motion plans were generated offline, rather than recomputed online based on the object's tracked state, but online motion planning is a trivial extension.
\begin{figure*}
\begin{subfigure}[b]{0.32\linewidth}
\centering
\includegraphics[width=\linewidth]{figures/2-29/deep-microwave-relabeled-norm.png}
\caption{Microwave closing (MLP) }
\label{fig:deep-microwave}
\end{subfigure}
\begin{subfigure}[b]{0.32\linewidth}
\centering
\includegraphics[width=\linewidth]{figures/2-29/dmp-drawer-relabeled-norm.png}
\caption{Drawer opening (DMP)}
\label{fig:dmp-drawer}
\end{subfigure}
\begin{subfigure}[b]{0.32\linewidth}
\centering
\includegraphics[width=\linewidth]{figures/2-29/dynamic-final-TEST.png}
\caption{T-ball (DMP) }
\label{fig:dmp-dynamic}
\end{subfigure}
\label{fig:threegraphs}
\caption{\textbf{Simulation Results.} a) Comparison of our method optimized with DAPG against Natural Policy Gradient starting with a random policy in a microwave closing task using Gaussian multi-layer perception policies. b) Comparison of our method against PI$^2$-CMA starting with a random policy in a drawer opening task with DMP policies. c) Our method compared with PI$^2$-CMA with a initially random policy in t-ball with DMP policies. Results are shown as mean and standard error of the normalized returns aggregated across 20 random seeds.}
\label{graphy}
\end{figure*}
\subsection{Fitting a Policy to a Demonstration}
After collecting initial demonstrations from the motion planner, $D$, we can bootstrap our motor policy by initializing the parameters to the policy $\theta$. We can initialize a parameterized motor policy using any behavioral cloning technique; in practice, for DMPs, we use Locally Weighted Regression \cite{schaal1998constructive}, and for neural networks, we maximize the likelihood of the demonstration actions under the policy.
\subsection{Policy Search with Kinematic Rewards}
\label{sec:ps}
To improve the motor policies after bootstrapping, we can perform policy search based on the given (kinematic) reward function. Specifically, we choose a number of epochs $E$ to perform policy search for. For each epoch, we perform an iteration of policy search by executing the policy and collecting rewards based on the goal $q_{O}^*$. We define our reward functions using estimated object states $q_{O}$ and goal states $q_{O}^*$, and add a small action penalty.
\section{EXPERIMENTS}
The aim of our evaluation was to test the hypothesis that motion planning can be used to initialize policies for learning from demonstration without human input. We tested this hypothesis in simulation, against learning from scratch, and on real hardware, against human demonstrations, on three tasks: microwave-closing, drawer-opening, and t-ball. We note that we do not show asymptotic performance because our emphasis is on learning on real hardware from a practical number of iterations. All the components of the motion planning problem - state sampler, goal sampler, distance metrics, etc. - are reused between problems without modification.
\subsection{Simulation Experiments}
\label{sec:experiments}
The aim of our evaluation was to test the hypothesis that motion planning can be used to initialize policies for learning from demonstration without human input. We tested this hypothesis in simulation against learning from scratch, and on real hardware against human demonstrations, on three tasks: microwave-closing, drawer-opening, and t-ball.
\subsection{Simulation Experiments}
\label{simexp}
We used PyBullet \cite{coumans2013bullet} to simulate an environment for our object manipulation experiments.
We used URDFs to instantiate a simulated 7DoF KUKA LBR iiwa7 arm and the objects to be manipulated, which gave us ground-truth knowledge of the robot and object kinematics.
For all our simulated experiments, we compared implementations of our method against starting with a random policy.
For all three tasks, the state was represented as $s_t = [q_R, q_{O}]^T$ where $q_R$ denotes robot configuration and $q_{O}$ denotes object configuration. The action space $A$ was commanded joint velocity for each of the 7 motors.
The reward at each timestep $r_t$ was given as:
\begin{equation}
r_t = -c \ || q_{O}^* - q_{O}||_2^2 - a_t^TRa_t, \label{eq:r_t}
\end{equation}
where $q_{O}$ denotes the object state at time $t$, $q_{O}^*$ denotes desired object state, and $a_t$ denotes the agent's action. We set $c=60$ and R = $I \times 0.001$ for all experiments. As such, maximum reward is achieved when the object is in the desired configuration, and the robot is at rest.
Our first simulated task was to close a microwave door, which consisted of three parts: a base, a door, and a handle.
The pose of the handle was used for the EstimateGrasp method in Algorithm 2. The robot was placed within reaching distance of the handle when the microwave door was in an open position, but was too far to reach the handle in its closed configuration. Thus, the agent was forced to push the door with enough velocity to close it. We used Gaussian policies represented as multi-layer perceptrons with two hidden layers of sizes (32,32) in this experiment. The randomly initialized policy was optimized with natural policy gradient \cite{kakade2002natural}. Ten demonstrations were generated by perturbing the start state and initial kinematic plan with Gaussian noise. The behavior cloning was performed by maximizing likelihood over the demonstration dataset for 10 epochs. Our pretrained policy was optimized using Demo Augmented Policy Gradient \cite{rajeswaran2017learning}, which essentially adds the behavior cloning loss to the natural policy gradient loss, annealing it over time. This ensures that the agent remains close to the demonstrations early in learning, but is free to optimize reward exclusively as learning progresses. Results are shown in Figure \ref{fig:deep-microwave}.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/2-29/real_microwave_curves.png}
\caption{\textbf{Hardware experiment} comparing our initialization scheme with human demonstration. Results are shown as mean and standard error, aggregated across three random seeds. }
\label{fig:real-microwave}
\end{figure}
The second simulated task was to open a drawer. This task required the agent to grasp the drawer's handle and pull the drawer open.
Again, the pose of the object's handle was used for EstimateGrasp method in our algorithm.
In this experiment, we used DMP policies. The weights, goals, and speed parameters of the policies were optimized using PI$^2$-CMA \cite{stulp2012path}. We used 32 basis functions for each of the DMPs. The pretrained policy was initialized using Locally Weighted Regression (LWR) \cite{schaal1998constructive} with a single demonstration. The results of this experiment are shown in Figure \ref{fig:dmp-drawer}.
The third simulated task was to hit a ball off a tee. The ball started at rest on top of the tee. The pose of the ball was used in the EstimateGrasp method. The object state was defined as the object's $y$ position relative to its initial pose. This experiment again used DMP policies initialized with LWR and optimized with PI$^2$-CMA. The results of this experiment are visualized in Figure \ref{fig:dmp-dynamic}.
The results of our simulated tasks can be found in Figure \ref{graphy}. Across all three tasks, we observe that policies initialized with our method dramatically outperform starting learning with a random policy. This confirms our hypothesis that using motion planning to generate demonstrations significantly speeds the acquisition of motor skills in challenging tasks like articulated object manipulation and t-ball.
\begin{figure}
\begin{subfigure}[b]{0.48\linewidth}
\centering
\includegraphics[width=\linewidth]{figures/mp_demo_crop.png}
\caption{Motion Plan Demonstration}
\label{fig:mp-demo}
\end{subfigure}
\begin{subfigure}[b]{0.48\linewidth}
\centering
\includegraphics[width=\linewidth]{figures/mp_learned_crop.png}
\caption{Bootstrapped from (a)}
\label{fig:mp-learned}
\end{subfigure}
\label{fig:ballpic}
\caption{\textbf{Real-world Ball Hitting} Images comparing an autonomous motion plan demonstration generated from our method vs. a bootstrapped motor skill initialized with that demonstration for a real-world robot hitting a ball off a tee as far as possible. Qualitatively, the bootstrapped motor skill outperforms the initial demonstration by learning to take advantage of the latent tasks dynamics. Videos can be found in our supplemental video. (a) A demonstration provided by the motion planner, which moves linearly towards the ball (b) A motor skill bootstrapped by the motion planner demonstration that learns a agile swooping motion to take advantage of the balls dynamics.}
\label{fig:real-tball}
\end{figure}
\subsection{Real-world Experiments}
For all our real-world experiments, we used a 7DoF Jaco arm \cite{campeau2019kinova} to manipulate objects (Figure \ref{demopic}).
We used ROS and MoveIt!\cite{chitta2012moveit} as the interface between the motion planner (RRT* \cite{karaman2011anytime} in our experiments) and robot hardware. For all real-world experiments, we compared implementations of our method against bootstrapping with a human demonstration, which we supplied. We acknowledge this potential bias in expert trajectories, and qualify our decision by only training on human demonstrations that at least accomplished the task. To collect human demonstrations, we had an expert human teleoperate the robot with joystick control to perform the task. For all tasks, the state space, action space, and reward were defined in the same way as in our simulated results (Section \ref{simexp}). Both experiments used DMP policies initialized with LWR \cite{schaal1998constructive} and optimized with PI$^2$-CMA \cite{stulp2012path} with 10 basis functions for each of the DMPs.
Our first real-world task was to close a microwave door, similar to the one described in our simulated domain (Section \ref{simexp}). As in the simulated microwave task, we used the pose of the handle for the EstimateGrasp method in Algorithm \ref{mpalgo}, and also the robot was similarly placed such that it was forced to push the door with enough velocity to close.
We placed an AR tag on the front-face of the microwave to track the microwave's state using a Kinect2.
Results are shown in Figure \ref{fig:real-microwave}.
We observe that the human demonstration is better than the one produced by the motion planner, which we credit to the fact that the motion of the door was heavily influenced by the dynamics of the revolute joint which the motion planner did not account for. Nonetheless, both policies converge to a similar final performance, with our method converging slightly faster.
Note the importance of the policy search phase: the motion planner alone is insufficient for performing the task efficiently.
Our second real-world task was to hit a ball off a tee as far as possible (Figure \ref{fig:real-tball}). Similar to our simulated task, the ball started at rest on top of the tee. The pose of the ball was used in the EstimateGrasp method. The object state was defined as the object's $y$ position relative to its initial pose.
We placed scotchlite-reflective tape on the surface of the ball and conducted our experiments within an OptiTrack motion-capture cage to track the object pose. We observe that when using a motion planner to hit the ball, it moves the bat in a linear motion to make contact, therefore transferring only horizontal motion to the ball. We qualitatively observe that during policy search, the robot learns a dynamic policy that accounts for the dynamics of the ball by applying force under the ball to ``scoop'' the ball upwards and forwards.
\section{RELATED WORK}To our knowledge, our method is the first to
use an object’s estimated kinematics in conjunction with a known robot dynamics model to bootstrap motor policy learning, and we discover and discuss important problems that are only introduced when leveraging policy-learning algorithms, behavioral-cloning, and motion planning algorithms to do so.
In this section, we discuss relevant approaches to motor skill learning.
Recently, Model-Predictive Control (MPC) has been used in the context of imitation learning and reinforcement learning to address the high sample complexity of policy search \cite{kahn2017plato, pan2017agile}. These approaches require a priori object dynamics, or human demonstrations to fit learned models; in constrast, our approach requires only object kinematics, which are much more readily estimated from visual data at runtime \cite{benny, li2019category}. As such, our approaches enables the learning of manipulation skills to be more autonomous than existing MPC-based methods.
\citet{tosun2019pixels} proposed a neural network model for generating trajectories from images, using a motion planner during training to enable the robot to generate a trajectory with a single forward pass at runtime. While this approach uses a motion planner for behavior cloning, it stops short of optimization to improve the resulting policy. In constrast, our method uses object kinematics to produce initial trajectories, while \citet{tosun2019pixels} only use the robot's kinematic model, which is insufficient when the task is to manipulate an object to a specific joint configuration.
While classic robot motor learning papers \cite{atkeson1997robot} leverage the known kinodynamics of the robot, they do not discuss kinematics of external objects or grasp candidates to bootstrap motor policies for object manipulation. We emphasize that we cannot form dynamic plans in the problem setting we are interested in: objects with unknown a priori dynamics.
\citet{kurenkov2019ac} proposed training an initially random RL policy with an ensemble of task-specific, hand-designed heuristics. This improves learning but the initial policy is still random, yielding potentially unsafe behavior on real hardware, and delaying convergence to a satisfying policy. By contrast, we choose to initialize the policy with demonstrations from a kinematic planner, ensuring feasibility, safety, and rapid learning. Moreover, we argue that motion planning is the principled heuristic to use to accelerate learning, as it is capable of expressing manually programmed heuristics like reaching and pulling. Finally, our approach can use the existing estimated object kinematics to provide a principled reward signal for model-free reinforcement learning.
Recently, residual reinforcement learning approaches have been developed which learn a policy superimposed on hand-designed or model-predictive controllers \cite{silver2018residual, johannink2019residual}. Our method is compatible with these approaches, where demonstrations from the motion planner can be used as a base policy on top of which a residual policy can be learned based on kinematic rewards. These methods typically suffer from the same limitations as MPC-based methods mentioned above.
Guided Policy Search (GPS) \cite{levine2013guided} uses LQR to guide policy search into high-reward regions of the state-space. The models employed are fundamentally local approximations, and thus would benefit greatly from a wealth of suboptimal demonstrations from the outset (as made evident by \citet{chebotar2017path}). GPS is one of the state-of-the-art algorithms we expect to be used within our framework as the policy search implementation (Section \ref{sec:ps}). A critical distinction between our work and GPS is the notion of planning trajectories in object configuration spaces and reasoning about grasp candidates to achieve a desired manipulation. This is done using information available apriori, and thus is immediately capable of generating high-value policies, whereas GPS is estimating dynamics models given observed data (obtained either from demonstration or random initialization). In the absence of a human demonstrator, our method would provide far more useful data at the outset of learning than running a naively initialized linear-gaussian controller (as evidenced by our comparisons to random initialization). The ideas proposed in our paper are distinct from those put forth in GPS: we present a method for obtaining demonstrations under certain conditions in the absence of a human.
Most similar to our line of work are those that use sample-based motion planners for improved policy learning. \citet{jurgenson2019harnessing} harness the power of reinforcement learning for neural motion planners by proposing an augmentation of Deep Deterministic Policy Gradient (DDPG) \cite{lillicrap2015continuous} that uses the known robot dynamics to leverage sampling methods like RRT* to reduce variance in the actor update and provide off-policy exploratory behavior for the replay buffer. However, \citet{jurgenson2019harnessing} are only able to address domains where they can assume good estimates of the dynamics model, such as producing free-space motions to avoid obstacles. Our setting, in contrast, focuses on object manipulation, where dynamics are not readily available, but are critical for learning good policies. \citet{jiang2019task} address learning to improve plans produced by a motion planner, but do not bootstrap closed-loop policies. Motion planners aren't expressive enough to leverage the dynamics in object-manipulation tasks, especially in the presence of unknown dynamics, and traditionally are unable to handle perceptual data like RGB images. Our method, on the other hand, enables motion planning to bootstrap policies that are more expressive than the original planner.
\section{CONCLUSION}
\label{sec:conclusion}
We have presented a method that uses kinematic motion planning to bootstrap robot motor policies. By assuming access to a potentially noisy description of the object kinematics, we are able to autonomously generate initial demonstrations that perform as well as human demonstrations, but do not require a human, resulting in a practical method for autonomous motor skill learning.
Our methodology is agnostic to the motion planner, motor policy class, and policy search algorithm, making it a widely applicable paradigm for learning robot motor policies. We demonstrate the power of our methodology by bootstrapping different policy classes with demonstrations from humans and a motion planner, and learn motor policies for three dynamic manipulation tasks: closing a microwave door, opening a drawer, and hitting a ball off a tee. Our framework is the first to enable robots to autonomously bootstrap and improve motor policies with model-free reinforcement learning using only a partially-known kinematic model of the environment.
\section*{Acknowledgement}
This research was supported by NSF CAREER Award 1844960 to Konidaris, and by the ONR under the PERISCOPE MURI Contract N00014-17-1-2699.
Disclosure: George Konidaris is the Chief Roboticist of Realtime Robotics, a robotics company that produces a specialized
motion planning processor.
\bibliographystyle{plainnat}
\section{INTRODUCTION}
\section{RELATED WORK}
Learning dynamic motor primitives using policy gradients \cite{peters2008reinforcement}. Original paper detailing policy gradients \cite{sutton2000policy}. PoWER, extending RL to episodic cases to make big gains on using policy gradients for motor primitives. \cite{kober2009policy}. Trust Region Policy Optimization \cite{sutton2000policy}.
Guided Policy Search \cite{levine2013guided}.
End-to-End Guided Policy Searrch \cite{levine2016end}. Natural Actor Critic \cite{peters2005natural}.
PILCO, model-based policy search \cite{deisenroth2011pilco}. Dynamic motor primitives \cite{ijspeert2003learning,schaal2007dynamics}. Policy search methods for robots \cite{peters2006policy}. Normalized Advantage Actor Critic \cite{gu2016continuous}. Model-based approaches to policy gradients \cite{levine2014learning}. Model-Based polciy search (REPS) \cite{peters2010relative}.
Interactively estiamting the kinematic graph \cite{hausman2015active}. Model-free approach to object manipulation \cite{karayiannidis2013model}.
\section{METHODOLOGY}
\section{EXPERIMENT}
\section{DISCUSSION}
\section{CONCLUSION}
\bibliographystyle{IEEEtran}
|
\section{}
\section{Introduction}
The electronic industry uses a wide palette of metals in various forms. Tiny metallic wires form interconnectors in logic circuits,
thin magnetic films are the media in data storage, mesoscopic layers are found as solders and protective finishings in printed
circuit boards. All these metals undergo corrosion processes, which can lead to degradation and ultimately to failure. In the last
few years the problem has aggravated because of the increased multiplicity of the elements used, the reduced spacing between
the various components, the often unpredictable users' environment and the deterioration of the air quality in region with a high
level of industrial activity. Thus, it is desirable to identify classes of metallic alloys, which are particular resilient to corrosion and,
at the same time, can deliver the functionality desired by the given application.
There are several known mechanisms of corrosion depending on the environmental conditions, such as the mixture of corrosive
agents at play and the humidity level, but a full experimental determination of the dynamics of corrosion is often difficult to achieve.
In general, a corrosive reaction is initiated by the binding of a chemical agent to a metallic surface, followed by the formation of
a new phase, with or without the possible release of new molecules incorporating atoms from the metallic surface (mass loss).
The progression of the corrosive reaction beyond the formation of a few reacted atomic layers is then determined by the diffusion
of the corrosive agent in the metal and the self-diffusion of the metallic ions. At the macroscopic level, such mass transport
mechanisms are further determined by the microstructure of the given sample, for instance through diffusion at grain boundaries.
Given the general complexity of the corrosion process modelling studies must extend across different length and time
scales~\cite{Gunasegaram2014}. These studies usually provide a valid contribution to the understanding of the corrosion dynamics
in a given material. The multiscale approach, however, is not suitable for scanning across materials libraries in the search for the
ideal compound resisting corrosion in a known environment, since the numerical overheads and workflows are computational
prohibitive and often require information from experiments (e.g. the microstructure). Thus, if one wants to determine a simple set
of rules to navigate the large chemical and structural space of metallic alloys, the attention has to focus on one of the steps
encountered in the corrosion path. A natural choice is that of determining the ease with which the first step takes place, namely to
evaluate the reactivity of a given metal to a particular chemical agent.
This is precisely the approach adopted here, where we estimate the reactivity of a vast database of metallic alloys to both O and
S. Oxygen and sulphur are particularly relevant, since for many metallic surfaces oxidation and tarnishing initiate the corrosion
process at the ambient conditions where one typically finds electronic equipment. However, even the simulation of oxidation and
tarnishing is complex and not amenable to a high throughput study. This, in fact, involves determining the full reaction path through
an extensive scan of the potential energy surface or through molecular dynamics. As such here we take a simplified approach
by computing the oxygen and sulphur binding energy to metals and by taking such binding energy as a proxy for reactivity. Clearly
this is a drastic simplification, since sometime a material presents a similar binding energy to O and S but different reactivity, as
in the case of silver~\cite{Saleh2019}.
{Such situations typically occur when the rate-limiting barrier in the reaction path does not correlate well with the
binding energy of the reaction product (see discussion in section~\ref{sec:be-bin-o}), or when the interaction between
the reactants on the surface changes the thermodynamics of the reaction as the coverage increases.}
However, the binding energy still provides a measure of the tendency of O and S to attack
the surface, and it strongly correlates to the stability of the product oxide/sulphide phases (see later). As such it is a valuable
quantity to classify metallic alloys as either weak or robust to corrosion.
Our computational scheme achieves the desired throughput by combining density functional theory (DFT) calculations with an
appropriate {\it descriptor}~\cite{Curtarolo2013}. In particular we use details of the density of state (DOS), namely the shape of
the partial DOS (PDOS) associated to the transition metal $d$-bands, to construct a model that provides an estimate of the binding
energy between a given metal alloy and both O and S. This is based on the notion that the O-metal and S-metal bonds get weaker
as the metal $d$ band is progressively filled~\cite{Norskov1995}. Our strategy then proceeds as follows. We first fit the parameter of the
model to DFT binding energy data for the 4$d$ transition metal series. This is preferred to the 3$d$ one, since it does not present
elemental phases with magnetic order, and to the 5$d$, since the electronic structure can be computed without considering spin-orbit
interaction. In particular, we explore several variations of the model and assess their ability to fit the data. Then, we construct an
automatic workflow to extract the DOS information from the AFLOWLIB.org materials database~\cite{Curtarolo2012}. This involves fitting
the various orbital-resolved PDOS to a semi-elliptical DOS. Finally, we use the descriptor to compute the binding energy for all the
experimentally known binary and ternary transition-metal alloys contained simultaneously in both AFLOWLIB.org and the Inorganic Crystal
Structure Database (ICSD)~\cite{Zagorac2019}.
The paper is organized as follows. In the next section we introduce our methodology, by discussing the various descriptors
considered, their rationale, the computed DFT data and the scheme for extracting data from AFLOWLIB.org. Next we present
the descriptor fitting procedure and evaluate its error in determining the binding energy, before proceeding to show
our results. We first determine the binding energy of O and S to transition-metal binary alloys and then we move to a restricted
number of ternaries. Then we conclude.
\section{Methods}
\label{sec:meth}
\subsection{Rationale for the descriptors}
The main idea beyond the definition of a descriptor is that it should represent a simple relation between a physical observable
and a property of the electronic structure easy to calculate. Once this is established, the descriptor can be used to
scan large databases in the search for particular compounds of interest. In our case an insight on how to construct
a descriptor for the binding energy between O/S and a transition metal alloy can be obtained by looking at Fig.~\ref{fig:1}.
In the figure we report the experimental enthalpy of formation per atom, $-\Delta H_f$, for a wide range of oxides and sulphides across
the 3$d$, 4$d$ and 5$d$ transition-metal series (the data used for constructing Fig.~\ref{fig:1} are listed in the appendix together with
their associated references), where multiple data corresponding to the same transition metal indicate that oxides/sulphides with
different stoichiometry exist for that metal.
\begin{figure}[htb]
\begin{center}
\includegraphics[width=0.45\textwidth]{Fig1.pdf}
\end{center}
\caption{(Color online) Experimental enthalpy of formation per atom for transition-metal oxides and sulphides across the 3$d$ (red
circles), 4$d$ (green squares) and 5$d$ (blue triangles) series. Open symbols are for sulphides and closed symbols for
oxides. Multiple symbols for the same transition metal correspond to oxides/sulphides with different stoichiometry. For
instance, there are five different data points for Ti-O, respectively TiO, Ti$_2$O$_3$, Ti$_3$O$_5$ and TiO$_2$ (both
anatase and rutile). Note that for several transition metals additional stable phases exist, but they are not reported in
the figure because their enthalpy of formation is not available (e.g. for Ti-O there also exist Ti$_6$O, Ti$_3$O, Ti$_2$O,
and Ti$_3$O$_2$).}
\label{fig:1}
\end{figure}
The figure reveals a number of clear trends. Firstly, we note that on average the enthalpies of formation of oxides are
significantly larger than those of sulphides, owning to the fact that the electronegativity of O is larger than that of S.
Secondly, for both oxides and sulphides the absolute value of the enthalpy of formation reduces monotonically (becomes
less negative) across the transition-metal series. The slope of such reduction is significantly more pronounced for oxides
than for sulphides, so that towards the end of the series $-\Delta H_f$ is very similar for these two groups ($-\Delta H_f$
for Ag$_2$O is almost identical to that of Ag$_2$S, $\sim$0.11~eV/fu {- fu = formula unit}). Finally, the
enthalpy of formation increases again beyond the noble metals (Cu, Ag and Au).
Importantly, these trends have not been observed only for the enthalpy of formation but also for the binding energy of transition
metals with monovalent atoms~\cite{Norskov1995,Hammer1995} (e.g. H), with oxygen~\cite{Besenbacher1993} and with a broad
range of molecular adsorbates~\cite{Hammer1996,Rubam1997}. This suggests the formulation of a descriptor, characteristic of
each adsorbate, based solely on the details of the electronic structure of the metal~\cite{Norskov1995}. The crucial observation is
that in typical transition metals the DOS is dominated by a partially filled, extremely broad $s$-$p$ band, and by a relatively narrow
$d$ band. As the atomic number increases the occupation of the $s$-$p$ band changes little, while the $d$ band becomes
progressively more filled, and hence moves to lower energies (with respect to the Fermi energy, $E_\mathrm{F}$). Upon approaching
the surface, the energy level of the adsorbate relevant for the bonding [in the case of O (S) the 2$p$ (3$p$) shell] gets broadened by the
interaction with the $s$-$p$ band. At the same time, it forms a bonding and anti-bonding pair with the $d$ band of the metal, which is here
approximated as a single molecular level. Thus, at the beginning of the transition-metal series one has the situation in which the
bonding level is filled and the anti-bonding one is empty, so that the binding energy is high. However, as the $d$ band fills also the
occupation of anti-bonding level increases, the adsorbate-metal bond weakens and the binding energy reduces.
It is then clear that the energy position of the $d$ band of the transition metal, together with some measure of the strength of the
transition-metal/adsorbate hopping parameter, can define a valuable descriptor for the binding energy. Two classes of such
descriptors are defined in the next section.
\subsection{Definition of the descriptors}
\label{sec:descr}
We model oxygen and sulfur as a single impurity level coupled to a bath of electrons characterising the metal. The level of
description for such bath defines the different models.
\begin{figure}[t]
\centering
\includegraphics[width=8cm, height=3cm]{Fig2}
\caption{(Color online) Level scheme showing the DOS for the Anderson impurity model. The $s$-$p$ band of the metal is wide and
the Fermi level is placed approximately at half filling. In contrast, the $d$ band has a moderate width and it is centered
at $\epsilon_d$. The impurity level, whose width is determined only by the interaction with the metal, is at an energy, $\epsilon_a$.
}
\label{fig:levels}
\end{figure}
The simplest one is often called the Newns-Anderson (NA) model~\cite{PhysRev.178.1123}, and it is defined by the following
Hamiltonian,
\begin{equation}
\label{eq:ham-na}
H_\mathrm{NA} = \epsilon_d d^{\dagger} d + \epsilon_a a^{\dagger} a + V_{ad} (a^{\dagger} d + d^{\dagger} a)\:.
\end{equation}
Here $a^\dagger$ ($a$) and $d^\dagger$ ($d$) are the creation (annihilation) operators, respectively, for the adsorbate and
the metal $d$ band, while $\epsilon_a$ and $\epsilon_d$ are the corresponding energy levels (before binding), and $V_{ad}$
their hybridization (hopping integral). At this level the metal $d$ band is treated as dispersionless and the contribution from the
$s$-$p$ band to the bond is neglected. $H_\mathrm{NA}$ can be easily diagonalised to yield the eigenvalues
$\epsilon_{\pm} = \epsilon_a + \frac 12 (\Delta_{ad} \pm W_{ad})$, where, $\Delta_{ad} = \epsilon_d - \epsilon_a$ and
$W_{ad} = \sqrt{4V^2_{ad}+\Delta^2_{ad}}$. If the fractional occupation of the $d$ band is $f$, then the total energies
before ($V_{ad}=0$) and after ($V_{ad}\ne0$) the coupling are $E_1 = 2(f\epsilon_d + \epsilon_a)$ and $E_2 = 2(1+f)\epsilon_{-}$,
respectively. Thus, the binding energy is given by
\begin{equation}
\label{eq:be-na}
E_\mathrm{b} = E_2 - E_1 = - (1-f)(W_{ad} - \Delta_{ad})\:.
\end{equation}
Finally, one can assume that there is an additional contribution to the binding energy, $E_{sp}$, originating from the
interaction with the $s$-$p$ band. Such contribution, however, is not expected to vary much across the transition-metal
series so that it can be kept constant. The NA model, thus depends on four parameters. Two of them are associated
respectively, to the metal, $\epsilon_d$, and the adsorbate, $\epsilon_a$, alone, one to the interaction between the two,
$V_{ad}$, and one is a constant, $E_{sp}$, specific of each adsorbate.
{Note that when extracting the various parameters from electronic structure theory calculations (see next section),
where the $d$ band has dispersion, the $d$-band energy level, $\epsilon_d$, is replaced by the position of the band
center.}
A more detailed description of the electrons in the metal is provided by the Anderson impurity model,~\cite{PhysRev.124.41}
which is schematically illustrated in Fig.~\ref{fig:levels}. In this case the metal band structure is taken into consideration
through the Hamiltonian,
\begin{align}\nonumber
\label{eq:Ham}
H_\mathrm{A} &= {\epsilon_a a^{\dagger}a}
+ \sum_{{\bf k}}\left({\epsilon^s_{\bf k}s^{\dagger}_{\bf k}s_{\bf k} + \epsilon^d_{\bf k}d^{\dagger}_{\bf k}d_{\bf k}}\right)+\\
&+ \sum_{{\bf k}}\left(V^s_{\bf k}a^{\dagger}s_{\bf k} + V^d_{\bf k}a^{\dagger}d_{\bf k} + \textrm{h.c.}\right)\:,
\end{align}
where now the operators $s^{\dagger}_{\bf k} (s_{\bf k}$) and $d^{\dagger}_{\bf k} (d_{\bf k}$) create (destroy) an electron
with wave-number ${\bf k}$, respectively, in the $s$-$p$ and in the $d$ band. The band energies are $\epsilon^s_{\bf k}$
and $\epsilon^d_{\bf k}$ and the hopping parameters $V^s_{\bf k}$ and $V^d_{\bf k}$.
The model defined by Eq.~(\ref{eq:Ham}) can be solved by constructing the appropriate Green's function, as shown in detail
in Appendix~\ref{appdx-1}. In brief, the `impurity' Green's function can be written as
\begin{equation}
G_{aa}(\omega) = {1 \over \omega -\epsilon_a - \Sigma(\omega)}\:,
\end{equation}
where $\Sigma(\omega)$ is the self energy describing the interaction with the metal. This is given by
\begin{align}
\Sigma(\omega) = \sum_{\bf k}\left( \frac{|V^d_{\bf k}|^2}{\omega -\epsilon^d_{\bf k} + i\eta} + \frac{|V^s_{\bf k}|^2}{\omega -\epsilon^s_{\bf k} + i\eta} \right)\:,
\end{align}
with $\eta\rightarrow0^+$. If we assume that the couplings are independent of ${\bf k}$, namely $V^d_{\bf k}=V_d$ and $V^s_{\bf k}=V_s$,
we can simplify the self energy into $\Sigma(\omega) = \Lambda(\omega) - i\pi\Delta(\omega)$, so that the DOS,
$D_a(\omega) = -\frac 1\pi\text{Im}[G_{aa}(\omega)]$, writes as
\begin{align}\label{DOS-imp}
D_a(\omega) = { \Delta(\omega) \over [\omega - \epsilon_a - \Lambda(\omega)]^2 + \pi^2\Delta(\omega)^2 }\:.
\end{align}
Explicit expressions for the real, $\Lambda(\omega)$, and imaginary part, $\Delta(\omega)$, of the DOS are detailed in Appendix \ref{appdx-1}.
Finally, the binding energy can be obtained by integrating the DOS,
\begin{align}
\label{eq:be-2}
E_\mathrm{b} = \int_{-\infty}^{0}D_a(\omega)d\omega - \epsilon_a\:,
\end{align}
where all the energies are defined from the metal Fermi energy, $E_\mathrm{F}=0$.
As defined, the Anderson impurity model depends on the adsorbate energy level, the metal/adsorbate electronic
coupling and the metal DOS. Here we approximate the metal DOS with a semi-circular model. In particular the
$d$-band DOS, $D_d(\omega)$, is described as a semi-circle with center at $\epsilon_d$ and half bandwidth,
$w_d$, namely as
\begin{equation}
\label{eq:semic-d}
D_d(\omega) = \frac{2}{\pi w_d}\sqrt{1 - \frac{(\omega -\epsilon_d)^2}{w_d^2}}\:.
\end{equation}
In contrast the $s$-$p$ band is taken as having its center at zero and a large bandwidth, $w_s$,
\begin{equation}
\label{eq:semic-s}
D_s(\omega) = \frac{2}{\pi w_s}\sqrt{1 - \frac{\omega^2}{w_s^2}}\:.
\end{equation}
Thus, in addition to $\epsilon_a$ and the relevant hybridisation parameters the Anderson model is uniquely defined by the
center and width of the $d$ band and by the width of the $s$-$p$ one. Furthermore, since we take the approximation that
the $s$-$p$ band remains unchanged across the transition metal series its contribution to the integral of Eq.~(\ref{eq:be-2})
can be simply replaced by a constant, $E_{sp}$, specific for each adsorbate.
In what follows the band parameters, $\epsilon_d$ and $w_d$, will be extracted from DFT calculations with a procedure
described in the next sections, while $\epsilon_a$ and $E_{sp}$ will be considered as fitting parameters. Finally, as far as
$V_{ad}$ is concerned, we will use a well-known strategy~\cite{Hammer2000} of considering the tabulated values extracted
from the LMTO potential functions.~\cite{Andersen1985} These are essentially scaling rules, so that the hybridisation
parameters are all known up-to a general scaling constant, $\beta$, which also will be fitted.
{Note that the same scaling rules are also used for the NA model, which then requires the parameter $\beta$.}
{Note also that the band parameters, which in principle should be computed for each specific
surface, are here replaced with those of the bulk compound. This is an approximation that allows us to perform a
large-scale analysis of the entire transition metal space, but that makes our model insensitive to the surface details.
The validity of such approximation will be discussed in section~\ref{sec:modfit}.}
\begin{table}[th]
\centering
{\renewcommand{\arraystretch}{1.5}
\begin{tabular}{|l|l|l|l|l|l|l|l|}
\hline
Model & \multicolumn{4}{c|}{\color{Green} DFT} & \multicolumn{3}{c|}{\color{Red}FIT} \\
\hline
NA & $\epsilon_d$ & $V_{ad}$ & & & $\epsilon_a$ & $E_{sp}$ & $\beta$\\
\hline
M1 & $\epsilon_d$ & $V_{ad}$ & $w_d$ & & $\epsilon_a$ & $E_{sp}$ & $\beta$\\
\hline
M2 & $\epsilon_d$ & $V_{ad}$ & $w_d$ & $WF$ & $\epsilon_a+WF$ & $E_{sp}$ & $\beta$\\
\hline
\end{tabular}
}
\caption{The different models investigated in this work. Quantities indicated with `DFT' are directly extracted from DFT
calculations for bulk materials, while those in the `FIT' column are taken as fitting parameters. Note that none of the models
is defined by more than three fitting parameters, specific for each adsorbate.}
\label{tab:models}
\end{table}
A summary of the models investigated is presented in Table~\ref{tab:models}, where we separate the quantities that we will
extract from DFT (`DFT' column) from those used as fitting parameters (`FIT' column).
{NA is the original Newns-Anderson model, Eq.~(\ref{eq:be-na}), with $\epsilon_d$ taken as the $d$-band
center. In contrast, M1 and M2 are just numerically defined and essentially implement Eq.~(\ref{eq:be-2}). In M1 the adsorbate
energy level is constant for each adsorbate, while in M2 its position is shifted by the experimental work-function of the metal,
$WF$ (either experimental or extracted from DFT). Note that all the models require only three fitting parameters.}
\subsection{Density functional theory calculations}
\label{sec:DFT}
DFT calculations are performed for the 4$d$ transition metal series (Y to Cd), which is taken as benchmark
for our models and as training dataset for their fit. We use the all-electron FHI-AIMS code\cite{Blum2009}, since
its local-orbital basis set makes it more numerically efficient than plane-wave schemes in computing surfaces. A revised
version of the Perdew-Burke-Ernzerhof (RPBE) exchange-correlation functional~\cite{PhysRevB.59.7413}, extensively
tested for adsorption energies, is considered throughout this work, together with the `light' basis-set FHI-AIMS default.
Tests for the more accurate `tight' basis set have revealed that the average error is minimal compared with that of the
descriptor.
For all the 4$d$ elemental phases we perform two sets of calculations either considering the experimental lattice structure
(see table \ref{tab:4d-1}) or a hypothetical {\it fcc} structure at the RPBE energy minimum. In both cases we construct
4-to-6-layer thick slabs with surfaces along the [100], [110] and [111] directions. The lateral dimensions of the supercell is
such that the surface contains a minimum of 4 atoms, so to minimise the interactions between the periodic images (single
impurity limit). The reciprocal space is sampled with a 12$\times$12$\times$1 grid and the relaxation is converged when
the forces are smaller than $5.0\times10^{-3}$~eV/\AA.
The binding energy is then calculated as
\begin{equation}
E_\mathrm{b} = E_\mathrm{ads+slab} - E_\mathrm{slab} - E_\mathrm{ads}
\end{equation}
where, $E_\mathrm{ads}$ is the DFT energy of the adsorbate alone (oxygen or sulfur)
{in its atomic form},
$E_\mathrm{slab}$ that of
the `relaxed' slab, and $E_\mathrm{ads+slab}$ is the energy of the relaxed slab including the adsorbate at its equilibrium
position. We always relax the top layer of the slab when calculating either $E_\mathrm{slab}$ or $E_\mathrm{ads+slab}$.
In both cases the lower layers are kept fixed to reduce the computational overhead. The orbital-resolved DOS for the
bulk structure and for the various surfaces are always calculated to be used for fitting the models. In that case the
Brillouin zone is sampled with a 144$\times$144$\times$144 and a 144$\times$144$\times$1 grid, respectively.
\begin{table}
\caption{\label{tab:4d-1} Summary table of the elementary phases of the 4$d$ transition-metal series investigated in
this work. For each element we report the atomic number, $Z$, the atomic configuration, the lattice structure of the
thermodynamically stable phase at room temperature, the most stable surface (the one investigated here), the experimental
work function, $WF$ (in eV), the Pauling electronegativity, $EN$, (for O and S this is 3.44 and 2.58 respectively), the
computed energy position of the $d$ band (in eV), $\epsilon_d$, the computed width of the $d$ band, $w_d$ (in eV).
Note that $\epsilon_{d}$ is taken with respect to the Fermi level, which is set to zero.}
\begin{tabular}{lllccclll} \hline\hline
El. & $Z$ & Conf. & Lattice & Surface & $WF$ & $EN$ & $\epsilon_\mathrm{d}$ & $w_\mathrm{d}$ \\ \hline
Y & 39 & 5$s^2$4$d^1$ & hcp & (0001) & 3.1 & 1.22 & ~1.77 & 1.88 \\
Zr & 40 & 5$s^2$4$d^2$ & hcp & (0001) & 4.05 & 1.33 & ~1.12 & 2.14 \\
Nb & 41 & 5$s^2$4$d^3$ & bcc & (100) & 3.95-4.87 & 1.60 & ~0.31 & 2.12 \\
Mo & 42 & 5$s^2$4$d^4$ & bcc & (100) & 4.36-4.95 & 2.16 & -0.18 & 2.19 \\
Tc & 43 & 5$s^2$4$d^5$ & hcp & (0001) & 5.01 & 1.9 & -0.87 & 2.34 \\
Ru & 44 & 5$s^2$4$d^6$ & hcp & (0001) & 4.71 & 2.2 & -1.76 & 2.14 \\
Rh & 45 & 5$s^2$4$d^7$ & fcc & (100) & 4.98 & 2.28 & -1.98 & 1.76 \\
Pd & 46 & 5$s^2$4$d^8$ & fcc & (100) & 5.22-5.6 & 2.2 & -1.96 & 1.41 \\
Ag & 47 & 5$s^2$4$d^9$ & fcc & (100) & 4.64 & 1.93 & -4.3 & 0.90 \\
Cd & 48 & 5$s^2$4$d^{10}$ & hcp & (0001) & 4.08 & 1.69 & -8.95 & 0.45 \\
\hline\hline
\end{tabular}
\end{table}
\subsection{Data extraction from AFLOWLIB.org}
\label{sec:aflow}
As explained before we have carried out novel RPBE-DFT calculations only for the 4$d$ transition-metal
series, which has served as dataset for fitting the model. Once the model has been determined this has
been run over an existing extremely large dataset of electronic structure calculations. In particular we
have extracted data from the AFLOWLIB.org library~\cite{Curtarolo2012}. At present, this contains basic
electronic structure information (DOS, band structure, etc.) for about 3.2 millions compounds, including
about 1,600 binary systems ($\sim$350,000 binary entries) and 30,000 ternary ones (2,400,000 ternary
entries). These have all been computed at the PBE level with the DFT numerical implementation coded in
the VASP package~\cite{PhysRevB.54.11169}, and with extremely standardised convergence criteria.
In particular, a subset of the AFLOWLIB.org data is for experimentally known compounds, namely for
entries of ICSD.\cite{Zagorac2019} Our work investigates that particular subset.
{It must be noted that there may be an apparent inconsistency in constructing the models
by using RPBE electronic structures and applying them to PBE data. However, one has to consider that
RPBE is a variation of PBE designed to improve over the energetics of chemisorption processes. The
two functionals remain relatively similar, and most importantly here, they produce rather close Kohn-Sham
spectra, and hence DOS. The variations in DOS between RPBE and PBE have very little influence on
the determination of the binding energy from our models, and certainly they generate errors much smaller
than that introduced by not considering structural information in our descriptors (see next section).}
The AFLOWLIB.org library is accessible through a web-portal for interactive use, but most importantly
through a RESTful Application Program Interface (API).\cite{TAYLOR2014178} This implements a query
language with a syntax, where comma separated `keywords' (the material properties available) are followed
by a `regular expression' to restrict the range or choice of the keywords. For instance the string
$\textrm{``Egap(1*,*1.6),species(Al,O),catalog(icsd)''}$ will give us a list of compounds containing Al and
O and included in ICSD, whose band gap ranges in the interval [1eV, 1.6eV]. Such queries are submitted
to the server \url{http://aflowlib.duke.edu/search/API/?}. Here we have used the AFLOWLIB.org
RESTful API, together with Python scripting, to search and extract the material information of
transition-metal (i) elemental phases, and both (ii) binary and (iii) ternary alloys. In all cases we have
limited the search to metals only.
We have found that, in general, for a given material prototype AFLOWLIB.org contains multiple entries.
Some correspond to different stable polymorphs, but also there is redundancy for a given lattice, where
multiple entries differ by small variations of the lattice constants. These are typically associated
to independent crystallographic characterisations of the same material, taken under slightly different
experimental conditions (temperature, pressure, etc.) or at different moments in time. Such small variations
typically change little the electronic structure, so that for our purpose they provide no extra information. We
have then removed such `duplicates' by grouping the compounds by lattice symmetry and total DFT energy.
Then, for a given crystal structure we have selected the entry presenting the lowest energy. Such procedure
has returned us 88 elemental phases, 646 binary and about 50 ternary metallic alloys. For these we
have extracted the orbital-resolved DOS, which was then fitted to the semi-circular DOS of
Eq.~(\ref{eq:semic-d}).
The fit proceeds by minimising the mean squared variance between the actual DFT-calculated
DOS, $D_{\text{DFT}}(\omega)$, and the semicircular expression, $D_d(\omega, \epsilon_i, w_i)$, namely by
minimising the following quantity
\begin{equation}
\Delta D=\left(D_{\text{DFT}}(\omega) - \sum_{i=1}^{n}\eta_i D_d(\omega, \epsilon_i, w_i)\right)^2\:.
\end{equation}
Here, $n$ is the number of semicircles used in the fit, while $\eta_i$, $\epsilon_i$, $w_i$ are the weights, centres and
half widths of each semicircle, respectively.
{We have initially used a variable number of semicircles to fit the DOS, but found that a single one
for each atomic orbital was always providing the best result. The fit extends to all species present in a compound and all
orbital channels ($s$, $p$, $d$ and sometime $f$), but only data related to the $d$ band, $\epsilon_d$ and $w_d$,
of all the species are retained when using the model.} An example of such fit is provided in Fig.~\ref{fig:dosfit} for
Pt$_2$Y$_1$.
\begin{figure}[htb]
\begin{center}
\includegraphics[width=8cm]{Fig3.pdf}
\end{center}
\caption{(Color online): Orbital-resolved DOS for Pt$_2$Y$_1$ (ICSD 649861), as extracted from AFLOWLIB.org and its corresponding fits.
In panel (a) we show the DFT-computed total orbital-resolved DOS. Panels (b) and (d), respectively are the element-resolved, orbital-resolved
DOS for Pt and Y. In panels (c) and (e) we show our semi-circular fit to the Eq.~(\ref{eq:semic-d}) semicircular DOS.
}
\label{fig:dosfit}
\end{figure}
\section{Results and Discussion}
\label{sec:result}
\subsection{Fitting the models}
\label{sec:modfit}
\begin{figure}[htb]
\begin{center}
\includegraphics[width=8.5cm]{Fig4.pdf}
\end{center}
\begin{center}
\begin{tabular}{|l|l|l|l|l|l|l|}
\hline
model & $\sigma$: (a) & $\sigma$: (d) & $\sigma$: (b) & $\sigma$: (e) & $\sigma$: (c) & $\sigma$: (f) \\
& O & S & O & S & O & S \\
\hline
NA & 0.66 & 0.40 & 0.66 & 0.40 & 0.51 & 0.67\\
M1 & 0.81 & 0.50 & 0.81 & 0.51 & 0.66 & 0.70\\
M2 & 0.76 & 0.47 & 0.82 & 0.53 & 0.73 & 0.66\\
\hline
\end{tabular}
\end{center}
\caption{(Color online) Comparison between the DFT binding energies and those estimated with the best fit
of the various models. The left column [panels (a)-(c)] is for oxygen, while the right column [panel (d)-(f)] for
sulfur. Circles, squares and diamonds correspond to the NA, M1 and M2 model, respectively. The relative
$\chi^2$ minimum of each model is shown in the legends (see text for details).
The table contains the binding-energy mean deviation, $\sigma = \sqrt{\chi^2(\epsilon_a, \beta, E_{sp})}$.
{The two top panels are for the {\it fcc} structures with the fit done over the DOS of the
specific surfaces; the two middle are again for the {\it fcc} structures, but now we use the bulk DOS for the
fit; the two lower panels are for the actual experimental crystal structures of the compounds.}}
\label{fig:comp-model}
\end{figure}
Each of the three models introduced in the previous sections requires to determine three parameters, $\epsilon_a$, $\beta$
and $E_{sp}$, specific to each adsorbate. These are obtained by fitting the RPBE data for the 4$d$ transition-metal series.
In particular we minimise the sum of the mean squared difference between the DFT binding energies,
$E^\mathrm{DFT}_\mathrm{b}$, and those computed by the models, $E^\mathrm{mod}_\mathrm{b}(\epsilon_a, \beta, E_s)$, namely
\begin{equation}
\chi^2(\epsilon_a, \beta, E_{sp}) = \frac{1}{N_{s}}\sum_{i\in [\mathrm{Y-Cd}]}
[(E^\mathrm{DFT}_\mathrm{b})_i - E^\mathrm{mod}_\mathrm{b}(\epsilon_a, \beta, E_{sp})_i]^2\;.
\end{equation}
where $N_s$ is the total number of surfaces taken over the [Y-Cd] range. For {\it fcc} bulk and surface, $N_s=33$, while for real structures
we had $N_s=28$. The fits for O and S are performed independently, with $\epsilon_d$ taken from the RPBE calculations, $V_{ad}$ from the
scaling laws of Ref.~[\onlinecite{Andersen1985}] and $WF$ from experiments. Finally, the electronic filling of the $d$ band, $f$,
or equivalently the position of the Fermi energy are also taken from the RPBE calculations.
Our best fits are presented in Fig.~\ref{fig:comp-model}, where we show the model-predicted energies against the RPBE values
for both oxygen [panels (a)-(c)] and sulfur [panels (d)-(f)]. The table below the figure reports the mean absolute deviation of the
binding energy, $\sigma = \sqrt{\chi^2(\epsilon_a, \beta, E_{sp})}$. As we go down the rows in the figure, we have three different sets of
fit, which differ for the choice of the DFT-calculated DOS taken to compute the $d$-band centre and bandwidth, and for the
target DFT energies. In the first row [panels (a) and (d)] the DOS is that of the surface atoms of the metals constrained to the {\it fcc}
lattice, and so are the target DFT energies. In the second-row panels [(b) and (e)] the DOS is that of the bulk {\it fcc} lattice, while
the target binding energies remain the same. Finally the last panels [(c) and (f)] use data for the metals in their thermodynamically
stable structure (see Table~\ref{tab:4d-1}).
In general we find that all the models tend to perform better for oxygen than for sulphur, in particular when the actual equilibrium
structures are considered [panels (c) and (f)]. Note that the spread of the DFT RPBE binding energies for O is significantly larger
than that of S (by about 2~eV), reflecting the same trend observed for the enthalpies of formation of oxides and sulphides (see
Fig.~\ref{fig:1}). This means that a similar $\chi^2$ translates in a smaller relative error for O. Interestingly, while in the case of O
our best fit is obtained for the experimental structures, the opposite happens for S, for which the fit for the hypothetical {\it fcc}
lattice is significantly more accurate. In fact, we find that the worst performance is obtained for S and the experimental structures, regardless
of the model used. This large error is associated to a significant scattering in the actual DFT data, in particular towards the beginning
of the series. For instance we find that when going from the most stable (0001) surface of {\it hcp} Y to the same for Zr the binding
energy marginally increases (becomes less negative), as expected from the larger occupation of the $d$ band. However, when moving
to the most stable (100) surface of {\it bcc} Zr, $E_\mathrm{b}$ significantly decreases and in fact it becomes lower than that of
both Y and Zr. Clearly such behaviour cannot be captured by any of the models, since when going from Y to Zr to Nb the position of
$\epsilon_d$ monotonically increases (see Table~\ref{tab:4d-1}). Similar anomalies are found for Ru and Rh, although much
less pronounced.
The much more pronounced spread in binding energies for sulphur can be attributed to its electronegativity, lower than that of
oxygen, and to the associated ability to form compounds involving transition metals over a broad range of stoichiometry. This
is particularly evident towards the beginning of the transition metal series. For instance, while Y forms only one stable oxide and
one sulphide, Y$_2$O$_3$ and Y$_2$S$_3$, so that it takes only the 3+ oxidation state, Zr has a single oxide, ZrO$_2$, but
can form sulphides with five different stoichiometries, ranging from Zr$_3$S$_2$ to ZrS$_3$ (see tables in Appendix \ref{materials_tables}).
Most importantly the enthalpy of formation of these different sulphides varies significantly, from 0.77~/eV/atom for Zr$_3$S$_2$
to 1.99~eV/atom for ZrS$_2$. It is also interesting to note that even when there are stable oxides formed with the same transition
metal but with different stoichiometry, therefore yielding a different oxidation state for the metal, the fluctuation in enthalpy of
formation remains small. For instance one can find NbO (Nb oxidation state 2+) with an enthalpy of formation of 2.17~eV/atom and
Nb$_2$O$_5$ (Nb oxidation state 5+) with an enthalpy of formation of 2.81~eV/atom.
A second important conclusion can be taken by looking at the first two rows of Fig.~\ref{fig:comp-model}, where the same DFT
binding energies computed for the {\it fcc} lattice are modelled by using the band parameters of either the surface atoms [panels
(a) and (d)] or those of the bulk [panels (b) and (e)]. Clearly the two sets of fit present very similar errors, a fact that reflects the
small changes in band parameters when going from the surface to the bulk. Indeed such changes do exist and in fact there
is established evidence for binding energy shifts with $d$-band center shifts.~\cite{GREELEY2005104} However, while the inclusion
of these small corrections improve the fit when a relatively narrow range of metals is considered, they have little impact in our
case, that considers the entire transition metal series. Since our intention is to examine a very broad range of metallic alloys
we can approximate the DOS of the surface with that of the bulk. This allows us to avoid performing surface calculations for the
several hundreds compounds previously selected. {An interesting possibility for improving on such assumption would
be that of constructing a simple descriptor correlating the DOS narrowing at surfaces with the DOS of the bulk.}
{A more quantitative estimate of the accuracy of our model, at least for the elemental phases, can be
obtained by analysing in more detail the distribution of DFT binding energies for the 4$d$ transition-metal series across
the different surfaces of the actual structures and the hypothetical {\it fcc} ones. Such distribution is available in
Fig.~\ref{fig:comp-model}, and it is re-plotted as a function of the atomic number in Fig.~\ref{fig:fig10} in
appendix~\ref{DFTbindenergy}. From the figures we notice that the spread in values is of the order of 1~eV across the series,
with the exception of Tc and both Mo and Nb, but only in the case of S. Clearly, Tc is not a matter of concern, since it is
radioactive and forms a rather limited number of known binaries. Mo and Nb are more problematic and effectively set the
accuracy of our model, which is of the order of $\pm$1~eV.}
Finally, we notice that when comparing the different models we find little difference in accuracy, with the original NA model
performing slightly better than both M1 and M2. {This fact is somehow counterintuitive, since one expects a
better fit for models including more parameters. We attribute such behaviour to the fact that here we apply the models to
a very broad distribution of binding energies, for which the fluctuations of the DFT values are relatively large. Over such range
the accuracy of the model is mainly driven by the $d$-band center, while finer details, such as the bandwidth, appear to have
little impact. Note that in literature there are several examples of model improvements associated to descriptors, which include
more information about the band shape~\cite{Xin2014}. These, however, are related to narrow subsets of compounds, for
which the $d$ band center changes little, and the binding energy is driven by more subtle features of the electronic
structure. For this reason in the remaining of the paper we will consider the NA model only.}
\subsection{Binding energies of binary alloys to O and S}
\label{sec:be-bin-o}
We now discuss the trends in reactivity of transition-metal binary alloys to oxygen and sulfur. Out of the 30 transition metals there
are $30\times29 / 2 = 435$ binary systems, a number that needs to be compared with the 646 binary intermetallic
compounds found at the interception between the AFLOWLIB.org and the ICSD databases. A more detailed view of the
chemical distribution of such 646 compounds can be obtained by looking at Fig.~\ref{fig:ba-bin}, where we graphically plot
the number of stable phases for each of the 435 binary systems. Firstly, we note that there are several binary systems for
which no single compound is found. This does not necessarily means that the two elements are not miscible, but simply that
there is no stable ordered crystalline phase, for which a full crystallographic characterisation is available. This is, for instance,
the case of the Hf-Zr system; the two elements are miscible at any concentration, but the thermodynamically stable phase
is a solid-state solution across the entire composition diagram. A similar situation is found for many binary systems made of
elements belonging to the same group or to adjacent groups, namely along the right-going diagonal of the matrix of Fig.~\ref{fig:ba-bin}.
In contrast, there is a much stronger tendency to form stable intermetallic phases in binary systems comprising
an early ($d^{0}-d^{3}$) and a late ($d^{7}-d^{10}$) transition metal. For instance Ti-Pd is the system presenting
the largest number, namely 12 of stable phases.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.48\textwidth]{Fig5}
\end{center}
\caption{(Colour online): Distribution of stable phases across the entire transition-metal binary system map.
The number of compounds for a given binary system is colour coded.}
\label{fig:ba-bin}
\end{figure}
Next we move to discuss the trend in binding energies to oxygen and sulfur across the binary-system space.
Clearly the binding energy is an object that depends on both the chemical composition and the stoichiometry of
a compound, namely for a binary alloy it is a four dimensional function. Thus for the $A_xB_y$ binary one
has $E_\mathrm{b}(A_xB_y) = f(A, x, B, y)$. We then proceed in the following way. For each binary system
$A$-$B$ we analyse all existing stoichiometry, $A_xB_y$, and compute all the possible binding energies by
running the NA descriptor against the partial DOS of all inequivalent bulk atomic sites. Then, we plot on
a matrix analogous to that of Fig.~\ref{fig:ba-bin} the minimum, $E_\mathrm{b}^\mathrm{min}$,
and maximum, $E_\mathrm{b}^\mathrm{max}$, binding energy found for that system, namely
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.48\textwidth]{Fig6a}
\includegraphics[width=0.48\textwidth]{Fig6b}
\end{center}
\caption{(Colour online) Binding-energy map across the transition-metal binary systems.
The left panel is for oxygen, and right panel for sulfur.
For each given binary the colour of the two semicircles encodes the value of $E_\mathrm{b}^\mathrm{min}$
and $E_\mathrm{b}^\mathrm{max}$ (energies are in eV according to the heat-map scale depicted on the side
of each picture). These are taken across the different stoichiometry and possible binding sites for any given
binary system.}
\label{fig:be-bin-os}
\end{figure*}
\begin{align}\label{Emax}
E_\mathrm{b}^\mathrm{max}(A:B) &= \max_{x,y}\{E_\mathrm{b}(A_xB_y)\}\:,\\
E_\mathrm{b}^\mathrm{min}(A:B) &= \min_{x,y}\{E_\mathrm{b}(A_xB_y)\}\:.
\end{align}
The results of this analysis for both oxygen (left-hand side graph) and sulphur (right-hand side graph) are presented in
Fig.~\ref{fig:be-bin-os}, where $E_\mathrm{b}^\mathrm{min}$ and $E_\mathrm{b}^\mathrm{max}$ occupy the two halves
of a circle and are encoded as a heat map. In the figure red tones indicate a weak binding energy, thus low reactivity, while
the green/blue ones are for strong binding and high reactivity. When the two halves of a particular circle appear approximately
of the same colour there is little variance in binding energy across stoichiometry and binding sites, while a strong contrast means
that for that binary system there are extremely reactive sites (for some stoichiometry) together with weak ones (for the same
compound or for different ones). Note that Fig.~\ref{fig:be-bin-os} just depicts the trend in reactivity, but it cannot be taken as an
absolute measure of the binding energy across the binary space. In fact, it is not guaranteed that either $E_\mathrm{b}^\mathrm{min}$
or $E_\mathrm{b}^\mathrm{max}$ are actually accessible. For instance one may have the situation in which the most reactive site
of a given binary system is associated to an unstable surface, so that it will be hardly available in practice.
Nevertheless, Fig.~\ref{fig:be-bin-os} provides valuable insights into the reactivity to oxygen and sulfur of the transition-metal alloys.
As expected one finds the less reactive compounds among the alloys made of late ($d^{7}-d^{10}$) transition metals,
regardless of the period they belong to. For these the binding energy is small and so is the variance across stoichiometry
and binding sites. The binding energy then becomes increasingly more negative as we move along the diagonal of
the plots towards early ($d^{0}-d^{3}$) transition metals. For this class the variance still remains relatively low owning
to the fact that the position of the center of the $d$ band is similar for elements in near groups. A different situation
is encountered when one moves off the plot diagonal, namely towards alloys combining an early and a late transition
metal. In this case we typically find both strongly and weakly reactive sites. An extreme example is found for the Pt-Ti
system, where an $E_\mathrm{b}^\mathrm{min}$ of -8.62~eV is found for the Ti site of Ti$_3$Pt$_1$, while a
$E_\mathrm{b}^\mathrm{max}$ of -4.53~eV corresponds to Pt in Ti$_1$Pt$_8$. This last group is of particular interest,
in particular if one can find binary alloys with an overall weak reactivity, since the constituent elements do not include only
expensive noble metals. It is also important to note that there is correlation in the variance between $E_\mathrm{b}^\mathrm{min}$
and $E_\mathrm{b}^\mathrm{max}$ and the abundance of compounds in a particular binary system (see Fig.~\ref{fig:ba-bin}).
In fact, we find that more abundant binary systems, which have potentially a larger stoichiometric space (and thus a larger variance
in possible inequivalent binding sites), have a larger colour contrast in Fig.~\ref{fig:be-bin-os}.
Let us now compare the two panels of Fig.~\ref{fig:be-bin-os} for oxygen and sulphur. This is an important exercise, since often in
an ambient-condition corrosion process O and S compete for the same binding site, thus that their relative reactivities may determine
the final products of reaction and the overall reaction rate. In general, a bird-eye view of the data suggests rather similar qualitative
chemical trends for sulphur and oxygen. However, a closer looks reveals a few differences: (i) the overall binding energies for sulphur
are lower than those of oxygen (note that the scales in the two panels of Fig.~\ref{fig:be-bin-os} are different); (ii) the spread, or the
variation over the minima and maxima (basically the colour contrast between the upper and lower semi-circles), for sulphur is typically
larger than for oxygen (e.g. in Hg:Pt).
\subsection{Reactivity of binary alloys to elemental O and S}
\label{sec:be-bin-o}
We are now going to develop a simple criterion for estimating, on a more qualitative ground, the relative reactivity of a given binary
system to S and O. The idea is to compare the predicted reaction rates for O and S absorption and to evaluate these from our
computed binding energies. For simplicity here we take O$_2$ and S$_2$ as the main reactants, so that the reaction of interest
is: $TM+1/2\mathrm{O}_2\rightarrow TM_\mathrm{O}$, where $TM$ indicate the transition metal and $TM_\mathrm{O}$ the
transition metal with one O adsorbed (the same holds for S). The enthalpy of reaction, $\delta E^n$ ($n=$O, S), can then be simply
written as $\delta E^n=E_\mathrm{b}^n+1/2E_\mathrm{atom}^{n_2}$, where $E_\mathrm{atom}^{n_2}$ is the experimental atomization enthalpy for either O$_2$ (5.1~eV) or S$_2$ (4.4~eV), { and $E_\mathrm{b}^n$ is the binding
energy of the $n$ specie}.
Importantly, the enthalpy of reaction is often found to be linearly correlated to the activation
energy, at least in the case of small molecules interacting with late transition metals. These so-called Br\o{}nsted-Evans-Polanyi
relations~\cite{BEP1,BEP2} thus establish a simple connection between a thermodynamical quantity, the enthalpy of reaction, and
a dynamical one, the activation energy, $E_\mathrm{act}^n$. Thus, one has $E_\mathrm{act}^n=\gamma \delta E^n+\xi$, where the
coefficients $\gamma$ and $\xi$ are, in principle, specific of any given reactant. Finally, the reaction rate, $\kappa$, is solely determined
by the activation energy via the usual Arrhenius expression, $\kappa=\nu \mathrm{e}^{-E_\mathrm{act}/k_\mathrm{B}T}$, where $\nu$
is the frequency factor, $T$ the temperature and $k_\mathrm{B}$ the Boltzmann constant.
The crucial point in the discussion is the observation that the scaling coefficient entering the Brønsted-Evans-Polanyi relations,
$\gamma$ and $\xi$, are universal for different classes of molecules and/or bonds~\cite{Norskov2002,Michelides2003}. For instance for
simple diatomic homonuclear molecules (e.g. O$_2$, N$_2$) one finds $\gamma\sim0.95$ and $\xi\sim2$~eV. By assuming that
the same relation is valid also for S$_2$, we can then write an expression for the ratio between the reaction rates of O and S,
namely
\begin{equation}\label{rates}
\frac{\kappa_\mathrm{S}}{\kappa_\mathrm{O}}=\frac{\nu_\mathrm{S}}{\nu_\mathrm{O}}\:\mathrm{exp}\left({-\frac{\gamma(\delta E^\mathrm{S}-\delta E^\mathrm{O})}{k_\mathrm{B}T}}\right)\:.
\end{equation}
If one wants to use Eq.~(\ref{rates}) to determine the relative reaction rate at ambient conditions, then we will write $T=300$~K,
and a further simplification can be made by assuming that the frequency factors for O$_2$ and S$_2$ are similar,
$\nu_\mathrm{S}/\nu_\mathrm{O}\sim1$. Finally, considering that the typical S concentration in the lower atmosphere is about 1~ppm,
one expects similar corrosion to S and O when their reaction rates are in the ratio $\kappa_\mathrm{S}/\kappa_\mathrm{O}\sim10^6$.
This leads to the condition
\begin{equation}\label{rates2}
1=\frac{1}{2}\:\mathrm{exp}\left({-\frac{\gamma(E^\mathrm{S}_\mathrm{b}-E^\mathrm{O}_\mathrm{b})}{0.025~\mathrm{eV}}}\right)\:.
\end{equation}
We can than conclude that a given transition metal alloy will corrode equally to S and O when $E^\mathrm{S}_\mathrm{b}\sim E^\mathrm{O}_\mathrm{b}$,
otherwise the reactivity will be dominated by oxidation.
It is important to note, however, that in the atmosphere S is present mainly in the SO$_2$ and H$_2$S form, and not as S$_2$. Unfortunately
Brønsted-Evans-Polanyi relations are currently unavailable for SO$_2$ and H$_2$S so that a more quantitative analysis of the ambient relative
reactivity of O and S cannot be carried out. Nevertheless, the ratio $\kappa_\mathrm{S}/\kappa_\mathrm{O}$ of Eq.~(\ref{rates}) can serve as a
useful descriptor to analyse the relative reactivity to S and O of a binary system. This analysis is carried out in Fig.~\ref{RateRatio}, where we plot
$\ln(\kappa_\mathrm{S}/\kappa_\mathrm{O})$ over our binary space, and we take $E^\mathrm{max}_\mathrm{b}$ as binding energy.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.48\textwidth]{Fig7}
\end{center}
\caption{(Colour online): Reaction rates ratio $\kappa_\mathrm{S}/\kappa_\mathrm{O}$, Eq.~(\ref{rates}), for the binary systems investigated.
Here $\ln(\kappa_\mathrm{S}/\kappa_\mathrm{O})$ is plotted as a heat map for all the binary compositions by taking as binding energy
$E^\mathrm{max}_\mathrm{b}$.}
\label{RateRatio}
\end{figure}
As expected the $\ln(\kappa_\mathrm{S}/\kappa_\mathrm{O})$ map closely resembles that of the binding energies (see Fig.~\ref{fig:be-bin-os}),
with lower $\kappa_\mathrm{S}/\kappa_\mathrm{O}$ ratios for the late transition metals (particularly in the 5$d$ period), while the reactivity
to O is always largely dominant. From the figure it is clear that the condition $\kappa_\mathrm{S}/\kappa_\mathrm{O}>10^{-6}$, which
makes the ambient corrosion to S stronger than that to O, is met only for a rather limited number of binary systems. In fact, this seems to be
unique to alloys with both atomic species having more than 10 valence electrons (Ni, Pd and Pt). However, it is important to note that this
analysis is based on $E^\mathrm{max}_\mathrm{b}$, so that it is not specific of a particular compound, but simply explore trends in the binary
space. One can than still have binary compounds, where the electronic interplay between the atomic species results in lower binding energies,
and therefore reactivity, with respect to the binary system they belong to. This analysis is carried out next.
\subsection{Reactivity gain for binary compounds}
\label{sec:gain}
Our task here is to identify those binary compounds, in which the binding energy of the different inequivalent sites differs the most from
that of the corresponding single-element phases. In other words we wish to find those binary alloys for which the bond formation between
chemically different atoms alters the most the position of the $d$ band with respect to that of the elemental phase. For a generic $A_xB_y$
binary alloy such property can be captured by the ``elemental energy shift'', a descriptor defined as
\begin{equation}
\Delta E_\mathrm{el}=E^n_\mathrm{b}-E^n_\mathrm{el}\:,
\end{equation}
where $E^n_\mathrm{b}$ is the binding energy for the species $n$ ($n=A, B$) in the binary alloy and $E^n_\mathrm{el}$ is the binding
energy of the elemental phase of $n$. In particular, for any binary compound we compute the maximum and minimum value of $\Delta E_\mathrm{el}$.
Here $\Delta E_\mathrm{el}^\mathrm{max}$ (typically a positive value) corresponds to the particular site whose binding energy has been
increased the most in forming the alloy (the site is less reactive than the same element in its elemental phase), while $\Delta E_\mathrm{el}^\mathrm{min}$
(typically a negative value) is for the site whose binding energy has been reduced the most (the site is more reactive than in its elemental phase).
These two quantities are listed in Table~\ref{Tabgain} for the 24 compounds presenting the largest $\Delta E_\mathrm{el}^\mathrm{max}$ and
$\Delta E_\mathrm{el}^\mathrm{min}$ for both O and S. In the same table we also list the composition-averaged binding energy, $\tilde{E}_\mathrm{b}$,
defined as $\tilde{E}_\mathrm{b}=w_xE^A_\mathrm{b}+w_yE^B_\mathrm{b}$, where $w_x=x/(x+y)$ [$w_y=y/(x+y)$]. This latter energy provides a rough
estimate of the global reactivity of a particular compound.
\begin{table}[th]
\caption{\label{Tabgain} List of the binary compounds possessing binding sites, whose binding energy present large deviations with respect to that of
their corresponding elemental phases. All energies are in eV. In the brackets we report the particular element associated to $\Delta E_\mathrm{el}$.}
\centering
\resizebox{0.495\textwidth}{!}{
\begin{tabular}{c|c|c|c|c|c|c}
\hline \hline
\toprule
{} & \multicolumn{3}{c|}{ {\bf Oxygen}} & \multicolumn{3}{c}{ {\bf Sulphur}} \\ \hline
{\bf Comp} & $\tilde{E}_\mathrm{b}$ & $\Delta E_\mathrm{el}^\mathrm{max}$ & $\Delta E_\mathrm{el}^\mathrm{min}$ & $\tilde{E}_\mathrm{b}$ & $\Delta E_\mathrm{el}^\mathrm{max}$ & $\Delta E_\mathrm{el}^\mathrm{min}$ \\\hline
\midrule
RhY & -5.87 & 4.44 (Y) & -3.18 (Rh) & -4.63 & 2.01 (Y) & -0.83 (Rh) \\
Cu$_{5}$Y & -4.38 & 4.44 (Y) & -4.08 (Cu) & -3.99 & 2.01 (Y) & -1.76 (Cu) \\
Ir$_{2}$Y & -5.14 & 4.43 (Y) & -2.75 (Ir) & -4.32 & 2.00 (Y) & -0.44 (Ir) \\
AuSc & -5.74 & 3.86 (Sc) & -4.08 (Au) & -4.75 & 1.96 (Sc) & -2.10 (Au) \\
Ni$_{5}$Sc & -4.31 & 3.86 (Sc) & -3.63 (Ni) & -3.99 & 1.96 (Sc) & -1.70 (Ni) \\
AuLu & -5.82 & 3.84 (Lu) & -4.25 (Au) & -4.64 & 1.60 (Lu) & -1.88 (Au) \\
Au$_{3}$Lu & -4.72 & 3.84 (Lu) & -4.08 (Au) & -4.13 & 1.60 (Lu) & -1.71 (Au) \\
Co$_{3}$Y & -5.22 & 3.67 (Y) & -3.40 (Co) & -4.61 & -0.37 (Y) & 0.44 (Co) \\
PtZr & -5.82 & 3.58 (Zr) & -3.58 (Pt) & -4.80 & 1.74 (Zr) & -1.61 (Pt) \\
Ni$_{3}$Zr & -4.43 & 3.58 (Zr) & -2.85 (Ni) & -4.04 & 1.74 (Zr) & -1.29 (Ni) \\
IrW & -5.36 & 3.24 (W) & -1.76 (Ir) & -4.83 & 2.02 (W) & -0.87 (Ir) \\
AgHf$_{2}$ & -5.99 & 3.22 (Hf) & -3.44 (Ag) & -4.79 & 1.41 (Hf) & -1.63 (Ag) \\
Au$_{3}$Hf & -4.96 & 3.22 (Hf) & -5.06 (Au) & -4.37 & 1.41 (Hf) & -2.70 (Au) \\
Ni$_{4}$W & -4.44 & 3.16 (W) & -3.33 (Ni) & -4.27 & 1.95 (W) & -2.50 (Ni) \\
IrNb & -5.26 & 3.08 (Nb) & -1.57 (Ir) & -4.63 & 1.59 (Nb) & -0.47 (Ir) \\
Cd$_{3}$Nb & -4.69 & 3.08 (Nb) & -3.94 (Cd) & -4.36 & 1.59 (Nb) & -2.63 (Cd) \\
Ir$_{3}$Y & -5.79 & 3.03 (Y) & -2.55 (Ir) & -5.08 & 0.74 (Y) & -0.32 (Ir) \\
PdTa & -5.47 & 3.02 (Ta) & -3.43 (Pd) & -4.69 & 1.47 (Ta) & -1.88 (Pd) \\
Pd$_{3}$Ta & -4.59 & 3.02 (Ta) & -3.44 (Pd) & -4.33 & 1.47 (Ta) & -2.42 (Pd) \\
Pt$_{3}$Ta & -4.81 & 3.02 (Ta) & -3.76 (Pt) & -4.28 & 1.47 (Ta) & -1.71 (Pt) \\
Rh$_{2}$Ta & -4.89 & 3.02 (Ta) & -2.39 (Rh) & -4.52 & 1.47 (Ta) & -1.41 (Rh) \\
AuMn & -5.19 & 2.99 (Mn) & -2.97 (Au) & -5.00 & 2.59 (Mn) & -2.59 (Au) \\
Ir$_{2}$Lu & -5.55 & 2.95 (Lu) & -2.23 (Ir) & -4.74 & 0.80 (Lu) & -0.12 (Ir) \\
Cu$_{3}$Ti$_{2}$ & -5.04 & 2.89 (Ti) & -3.34 (Cu) & -4.54 & 1.46 (Ti) & -2.10 (Cu) \\
\bottomrule
\hline\hline
\end{tabular}
}
\end{table}
From the table we find, as somehow expected, that compounds formed from elements placed at the different edges of the $d$-metal
period present the largest $\Delta E_\mathrm{el}$. In general one finds that the binding energy of the most reactive elements,
typically Y, Sc, Lu, Zr, Hf, W, Nb and Ta, is drastically reduced (up to 4~eV) with respect to that of the corresponding elemental phase.
At the same time, $E^n_\mathrm{b}$ of the least reactive element increases, often by a relatively similar amount. Such variations
are significantly more pronounced when considering binding to O than to S, mostly because the binding energies to O are larger
and because their dependence on the $d$-band filling factor is more pronounced (see Fig.~\ref{fig:1}).
{Interestingly, we can identify compounds whose composition-averaged binding energy is relatively low,
$\tilde{E}_\mathrm{b}>-4.5$~eV for O, and similar for O and S (within some fraction of eV), and at the same time present inequivalent
sites that bind drastically differently from their element phases. These are mostly Ni-based intermetallics such as
Ni$_5$Sc ($\tilde{E}_\mathrm{b}^\mathrm{O}=$-4.31~eV, $\tilde{E}_\mathrm{b}^\mathrm{S}-\tilde{E}_\mathrm{b}^\mathrm{O}=$0.32~eV),
Ni$_4$W (-4.44~eV, 0.17~eV),
Ni$_3$Zr (-4.43~eV, 0.39~eV),
and also
Pd$_3$Ta (-4.59~eV, 0.36~eV) and
Cu$_{5}$Y (-4.38~eV, 0.39~eV).}
\subsection{Ternary alloys}
\label{sec:be-tern-os}
Finally we turn our attention to the ternary compounds. In this case the set available is significantly smaller than what found for the
binaries, and in fact the same search criterion used before now returns us only 50 ternary phases. This may look a bit surprising,
since ICSD approximately counts about 40,000 binary and 75,000 ternary phases~\cite{Zagorac2019}. However, here we are considering
only compounds made of transition metals. These are then prone to form solid state solutions or highly disordered phases~\cite{Curtarolo2019},
whose structures are typically not part of ICSD. Furthermore, we have only included the compounds that are both in
ICSD and AFLOWLIB.org library~\cite{Curtarolo2012}, namely at the intersection of the `real (ICSD)' (the subset of ICSD reporting
experimentally determined structures) and the `{\it ab initio} (AFLOWlib)' database. In any case, the ternaries considered can be
found from the union of the sets S$_{3d} = $ [Sc-Zn], S$_{4d} = $~[Y-Zr, Mo, Ru, Pd-Cd] and S$_{5d} = $ [Lu-Hf, W-Re, Pt-Hg],
namely they may contain any of the 3$d$ element and a selection of the 4$d$ and 5$d$, with a preference for either early or late
transition metals.
In Fig.~\ref{fig:be-tern} we show the list of these ternary compounds sorted by their composition-averaged binding energy, while
details of the site-dependent binding energy together with the associated elemental energy shifts are provided in Table~\ref{tab:tern-os}
in the appendix. In general, as expected, the ternary phases showing shallower $\tilde{E}_\mathrm{b}$ are those including late
transition metals, often going beyond the noble ones (e.g. Zn and Cd). More interestingly, the subset for which $\tilde{E}_\mathrm{b}$
is approximately the same for O and S are those with an average electronic configuration close to $s^2d^9$, namely that of
Cu, Ag and Au. These, for instance include, Cu$_2$NiZn, CdPt$_2$Zn, AuCuZn$_2$ and AuCuCd$_2$. Among them, Cu$_2$NiZn
appears particularly interesting, since it mimics the electronic structure of a noble metal, without including expensive elements. In
contrast, at the opposite side of the distribution we find alloys with a dominant early transition-metal composition, for which the binding
energy is deep and asymmetric between O and S.
\begin{figure}[htb]
\begin{center}
\includegraphics[width=\columnwidth]{Fig8}
\end{center}
\caption{(Colour online): Composition-averaged binding energy for all the ternary compounds investigated.
The two rectangles below each compound refer to O (left-hand side) and S (right-hand side), respectively,
and include the composition-averaged binding energy. The boxes are then colour coded with the same quantity
fo an easy visualization.}
\label{fig:be-tern}
\end{figure}
\subsection{Model validation for binary and ternary alloys}
\label{sec:be-tern-os}
{In closing this results section we are finally coming back to the question of the accuracy of our model
and the limits of its predictions. We have already remarked (see section~\ref{sec:modfit}) that the
spread in the DFT binding energies across different surfaces for elemental 4$d$ transition-metal compounds
is in the region of 2~eV ($\pm$1~eV). Here we aim at validating such error for binary and ternary alloys.}
{To this goal we have selected four binary compounds and one ternary and computed the DFT binding energies
for O and S for several different surfaces. The compounds in questions are AgZr (ICSD number 605996),
AgZr$_3$ (58392), CuPd (181913), Cu$_3$Pd (103084) and CuHfHg$_2$ (102969). In particular we have
selected two phases from the Ag-Zr binary system, as elemental Ag and Zr provide strongly and weakly coupling
binding sites, respectively; and two compounds from the Cu-Pd system, since it is a low binding-energy one, and hence
interesting for applications. Finally, we have considered CuHfHg$_2$, since it includes elements with a
broad range of binding strengths to O and S. For those we have computed the binding energies at the (100),
(110) and (111) surfaces, and whenever inequivalent, at the (001) one. Note that some of these compounds present
a layered structure, so that different surface terminations are possible. In this case we have computed the binding
energy for all the inequivalent terminations. The calculations then proceed as for the elemental phases, by finding the
minimum energy binding site, and its corresponding, $E_\mathrm{b}$.}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.45\textwidth]{Fig9}
\end{center}
\caption{(Colour online): DFT-computed binding energies of O (a) and S (b) for a number
of selected binary and ternary compounds: AgZr (ICSD \#605996) and AgZr$_3$ (ICSD \#58392); CuPd (ICSD \#181913),
Cu$_3$Pd (ICSD \#103084) and CuHfHg$_2$ (ICSD \#102969). For each compound we report the binding energy
of different surfaces and different absorption sites. Note that for the same surface orientation there may be different
possible terminations. Different symbols correspond to different surfaces (the colour is that of the dominant binding
specie). The colour code is as following: Ag = grey, Cu = green, Hf = magenta, Hg = orange, Pd = blue, Zr = red.}
\label{fig:fig9}
\end{figure}
{Our results are summarised in Fig.~\ref{fig:fig9}, where we report all the computed binding energies
and we colour-code the specific binding sites (the dominant site in the case the adsorbant coordinates with atoms
belonging to different species). Let us start from the Ag-Zr system. In general this presents a bimodal distribution, with
Zr-dominated binding sites showing low binding energies and Ag-dominated ones binding in a much weaker way.
In particular, the Zr sites have binding energies ranging from -10~eV to -8~eV for O and from -7~eV to -5~eV for S,
while the Ag ones are around -4.5~eV for O and -3.5~eV for S. These values have to be compared with those
predicted by our model by using the bulk DOS (see Fig.~\ref{fig:be-bin-os}). For the Zr-Ag system the model returns
a binding energy range of [-7, -5]~eV for O and [-6.5,-4.5]~eV for S. Thus, we find that our model is capable of
predicting the low binding-energy side of our distribution relatively well, while it appears to miss some Zr binding
sites with extremely low binding energy (it predicts the upper side of the Zr distribution). These are associated with
electron-depleted Zr-dominated surfaces, namely having electronic structure rather different from that of the bulk. }
{The situation is significantly better for the Cu-Pd system, where DFT returns a binding energy range
of [-6,-4]~eV for O and [-5,-3.5]eV for S, to be compared with a rather uniform model prediction of -4~eV for both
O and S. Such result is not surprising, since the electronic structure of Pd and Cu are rather similar, so that
large fluctuations at surfaces are not expected. Finally, when looking at our ternary compound we find a
tri-modal distribution of binding energies with those associated to Hf at -9~eV (-6~eV), those to Cu at -4.5~eV
(-3.5~eV) and those to Hg at -3.5~eV (-3~eV) for O (S). For ternary alloys our model was used only to predict the
composition-averaged binding energy, which is -4.7~eV for O and -4.24~eV for S. The composition-averaged
$E_\mathrm{b}$ computed from the DFT values are -5.13~eV for O and -3.88~eV for S, namely in quite good
agreement with the model.}
\section{Conclusion and outlook}
We have here investigated the propensity to oxidation and tarnishing of a large set of binary and ternary intermetallic
compounds. These were selected among existing phases, as reported in the ICSD and AFLOWLIB.org databases. Such
large-scale screening was enable by taking the binding energy to O and S as a proxy to the first stage of the corrosion
process and by the definition of a descriptor. More specifically, we have utilised the well-known $d$-band-position concept and
constructed a descriptor with the associated parameters being fitted to DFT calculations for the 4$d$ transition metal series.
A number of variants of the original Newns-Anderson model were evaluated before choosing one based on an analytical
semi-elliptical density of states. Such descriptor was then put to work against the electronic structures contained in
AFLOWLIB.org, after appropriate fitting, to investigate trends in the binding energy across the compositional space.
In general, we have found the binding to O to be significantly stronger than that to S, a fact that follows closely the behaviour
of the enthalpy of formation of oxides and sulphides. Such difference, however, gets reduced as one moves across the
transition metal period towards the $s^2d^9$ atomic configuration ($s^1d^{10}$ in the solid state), characteristic of Cu,
Ag and Au. A somewhat similar situation is found for binary and ternary compounds. In this case, however, the presence of
chemically and structurally inequivalent sites complicate the analysis, which is better performed by computing for each
compound and binary system the largest and smallest binding energies. This reveals binary systems presenting the
co-existence of strongly and weakly binding sites, composed of an early and a late transition metal. At the same time,
binary phases made of elements belonging to adjacent groups display little variation in binding energy with the particular
binding site.
The thermodynamical information contained in the binding energy can be then converted into a proxy for the reactivity by using
the so-called Br\o{}nsted-Evans-Polanyi relations. These are directly available for O$_2$ and have been extrapolated also
to S$_2$. By using such relations we have established that, at ambient conditions (temperature and relative S/O abundance),
early-stage corrosion to O and S compete only when the binding energies are comparable, otherwise oxidation always
appears to dominate. This first situation takes place only for late, and usually expensive, transition-metal alloys. However, in
the ternary space this seems to be possible also for a handful of alloys presenting an average $s^1d^{10}$ configuration,
but not necessarily including Ag and Au, such as Cu$_2$NiZn.
Overall our work provides a first rough navigation map across the binary and ternary transition metal composition space,
which is useful to categorise materials according to their propensity to oxidation and tarnishing.
{Certainly, our method has several limitations, which need to be overcome in order to establish a high-throughput
quantitative theory of surface reactivity across such vast chemical and structural space. Firstly, we need to improve over our
binding-energy prediction ability. For instance, our descriptor is completely agnostic to the specific surface and absorption site.
One first improvement may be that of running the NA model over the specific surface DOS, an operation that, however, will require
DFT surface calculations for the entire database, a numerically daunting task. In that case one may include additional features of the
surface DOS into the model, which is likely to become more accurate.~\cite{Xin2014} An alternative strategy is to develop models
taking into account the possibly bonding geometry of the different bonding sites. If machine-learning schemes~\cite{Wang2020}
can be constructed, this may limit the number of surface calculations to perform.}
{Secondly, we need to establish a more solid link between the binding energy and the surface chemical activity.
In this case, one has to validate a new set of Br\o{}nsted-Evans-Polanyi relations for O- and S-containing atmospheric
gases,~\cite{Aas2018} such as H$_2$S or SO$_2$. This will involve performing reaction path calculations over a range of
surfaces. The task is relatively straightforward for elemental phases, but becomes much more complex in the case of binary
and ternary alloys. Also in this case a machine-learning strategy generalising or replacing completely the Br\o{}nsted-Evans-Polanyi
approach may be a possible solution. A few examples in such direction exist~\cite{Stocker2020,Lee2020}, but to date the field
remains quite uncharted.}
\section*{ackowledgement}
We thank Corey Oses, Cormac Toher and Stefano Curtarolo for support with the AFLOWLIB.org API.
This work is supported by by Science Foundation Ireland (Amber center 12/RC/2278) and by Nokia Bell Lab.
Computational resources have been provided by the supercomputer facilities at the Trinity Center for High
Performance Computing (TCHPC) and at the Irish Center for High End Computing (ICHEC), projects
(tcphy108c and tcphy120c).
|
\section{Introduction}
Properties of coefficients of generating series \cite{Wi06}, especially
Fourier coefficients of powers of the Dedekind $\eta$-function have been
the focus of research since the times of Euler \cite{Ma72,Se85, Ap90, AE04, On03,HNW18}:
\begin{equation}
\label{eq:1}
\eta \left( \tau \right) ^{r}:=q^{\frac{r}{24}}\prod _{m=1}^{\infty }\left( 1-q^{m}\right) ^{r}=
q^{\frac{r}{24}}\sum _{n=0}^{\infty }a_{n}\left( r \right) q^{n}.
\end{equation}
Here, $q := e^{2\pi i\tau }$, $\func{Im}\left( \tau \right) >0$ and $r \in \mathbb{Z}$.
The coefficients are special values of the D'Arcais polynomials $P_n(x)$ \cite{DA13, Ne55, Co74, We06}.
It has been recently noticed that the growth and vanishing properties of these polynomials
have much in common with properties of other interesting polynomials \cite{HLN19,HN20B}.
These include special orthogonal polynomials as associated Laguerre
polynomials and Chebyshev polynomials of the second kind. Also included are polynomials
attached to reciprocals of the Klein's $j$-invariant and Eisenstein series \cite{HN20A,HN20C}.
In this paper we investigate growth properties and the zero distribution of
polynomials attached to arithmetic functions $g$ and $h$ inspired by Rota \cite{KRY09}.
Let $g$ be normalized and of moderate growth.
Further, let $0<h(n) \leq h(n+1)$. We put $P_0^{g,h}(x)=1$ and
\begin{equation} \label{gh}
P_n^{g,h}(x) := \frac{x}{h(n)} \sum_{k=1}^{n} g(k) \, P_{n-k}^{g,h}(x).
\end{equation}
This definition includes all mentioned examples.
Before providing examples
and
explicit formulas for these polynomials, we give one application for the coefficients of the
Dedekind $\eta $-function. Let $g(n)=\sigma (n):= \sum_{d \mid
n}d$, $h(n)= \func{id}(n)=n$ and
$a_n(r)$ be defined by (\ref{eq:1}), the $n$th coefficient of the $r$th power of the
Dedekind $\eta $-function. Han \cite{Ha10} observed that the Nekrasov--Okounkov hook length formula
\cite{NO06, We06} implies that $a_n(r) \neq 0$ if $ r > n^2 -1$. This improves previous results by Kostant \cite{Ko04}.
In \cite{HN20B} we proved that
\begin{equation} \label{improve}
a_n(r) \neq 0 \text{ holds for } r > \kappa \cdot
(n-1) \text{ where } \kappa =15.
\end{equation}
Numerical investigations show that $\kappa$ has to be larger than $9.55$ (see Table \ref{roots}).
In this paper we prove that (\ref{improve}) is already true for $\kappa = 10.82$.
Since the definition of $P_n^{g,h}(x)$ is quite
abstract, we provide two examples of families of polynomials,
to familiarize the reader with the types of polynomials
we are studying. At first, they appear
to have nothing in common.
Let us start with the Nekrasov--Okounkov hook length formula \cite{NO06}.
Let $\eta(\tau)$ be the Dedekind $\eta$-function.
Let $\lambda$ be a partition of $n$
and let $|\lambda|=n$. By
$\mathcal{H}(\lambda)$ we denote the multiset of hook lengths associated
with $\lambda$
and by $\mathcal{P}$,
the set of all partitions.
The Nekrasov--Okounkov hook length formula (\cite{Ha10}, Theorem 1.2) states that
\begin{equation} \label{ON}
\sum_{n=0}^{\infty} P_n^{\sigma}(z) \, q^n =
\sum_{ \lambda \in \mathcal{P}} q^{|\lambda|} \prod_{ h \in \mathcal{H}(\lambda)} \left( 1 + \frac{z-1}{h^2} \right)
=
q^{\frac{z}{24}} \eta(\tau)^{-z}.
\end{equation}
The identity (\ref{ON}) is valid for all $z \in \mathbb{C}$.
Note that the $P_n^{\sigma}(x)$ are integer-valued polynomials of degree $n$.
From the formula it follows that
$(-1)^n P_n^{\sigma}(x) >0$ for all real $x < -(n^2+1)$.
The second example is of a more artificial nature, discovered recently \cite{HN20A}, when
studying the $q$-expansion of the reciprocals of Klein's $j$-invariant and reciprocals of Eisenstein series \cite{BB05,BK17,HN20C}.
Let $$j(\tau)= \sum_{n=-1}^{\infty} c(n) q^n = q^{-1} + 744 + 196884q+ \ldots$$
denote Klein's $j$-invariant.
Asai, Kaneko, and Ninomiya \cite{AKN97} proved
that the coefficients of the $q$-expansion of
$1/j(\tau)$ are non-vanishing and have strictly alternating signs. This follows from
their result on the zero distribution of the $n$th
Faber polynomials $\varphi_n \left( x\right) $ and
the denominator formula for the monster Lie algebra. The zeros of the Faber polynomials are simple
and lie in the interval $(0,1728)$. They obtained the remarkable identity:
\begin{equation}
\frac{1}{j(\tau)} = \sum_{n=1}^{\infty} \varphi_n'(0) \,
\frac{q^n}{n}.\end{equation}
Let $c^{*}(n):= c(n)/744$. Define the polynomials $Q_{j,n}(x)$ by
\begin{equation}
\sum_{n=0}^{\infty} Q_{j,n}(x) \, q^n :=
\frac{1}{ 1 - x \sum_{n=1}^{\infty} c^{*}(n) \, q^n }.
\end{equation}
We have proved in \cite{HN20A} that $Q_{j,n}(x) = Q_{\gamma_2,n}(x) + 2x Q_{\gamma_2,n}'(x)+ \frac{x^2}{2} Q_{\gamma_2,n}''(x)$, where $Q_{\gamma_2,n}(x)$ are polynomials attached to Weber's cubic root
function $\gamma_2$ of $j$ in a similar way. We have also proved that $Q_{\gamma_2,n}(z)\neq 0$ for all $ \vert z \vert > 82.5$. Hence,
the identity
$$ \frac{\varphi_n'(0)}{n} =
Q_{j,n}(-744) =
\left( Q_{\gamma_2,n}(x) + 2 x Q_{\gamma_2,n}'(x)+ \frac{x^2}{2} Q_{\gamma_2,n}''(x)\right)_{\vert_{x= - 248}}
$$
restates and extends the result of \cite{AKN97}.
Now, let $g(n)$ be a normalized arithmetic function with moderate growth,
such that $\sum_{n=1}^{\infty} \vert g(n) \vert \, T^n$
is analytic at $T=0$. Then the illustrated examples are
special cases of polynomials $P_n^g(x)$ and $Q_n^g(x)$
defined by
\begin{eqnarray}
\sum_{n=0}^{\infty} P_n^g(z) \, q^n & =
& \text{exp} \left( z \sum_{n=1}^{\infty} g(n) \frac{q^{n}}{n}\right), \label{P}
\\
\sum_{n=0}^{\infty} Q_n^g(z) \, q^n & =
& \frac{1}{1- z \sum_{n=1}^{\infty^{\phantom{x}}} g(n) q^n}. \label{Q}
\end{eqnarray}
Note that $P_n^{\func{id}}(x)= x \, L_{n-1}^{(1)}(-x)$ are
associated Laguerre polynomials (see \cite{HLN19}). Letting
$g(n)= \sigma(n)$,
then we recover the polynomials provided by the Nekrasov--Okounkov hook length formula.
The polynomials $Q_n^{\func{id}}(x)$ are related to the Chebyshev polynomials of the second kind \cite{HNT20}.
It is easy to see that $P_n^g(z)$ and $Q_n^g(z)$ are special cases of
polynomials $P_n^{g,h}(x)$ defined by the recursion formula (\ref{gh}).
Here, $P_n^g(x) = P_n^{g,\func{id}}(x)$ and $Q_n^g(x) = P_n^{g,\mathbf{1}}(x)$.
In the next section, we state the main results of this paper.
\section{Statement of main results}
Let $g,h$ be arithmetic functions. Assume that $g$ be normalized
and $ 0 < h(n) \leq h(n+1)$.
It is convenient to extend $h$ by $h(0):=0$.
We start by recalling what is known \cite{HNT20, HN20A, HN20B}.
Assume that $G_1(T):= \sum_{k=1}^{\infty} \vert g(k+1)\vert \, T^k$ has a positive radius $R$ of convergence.
Let $\kappa > 0$ be given, such that $G_1(2/\kappa) \leq \frac{1}{2}$. Let $x \in \mathbb{C}$.
Then we have for all $\vert x \vert > \kappa \,\, h(n-1)$:
\begin{equation}\label{old}
\frac{ \vert x \vert }{2 \, h(n) } \vert P_{n-1}^{g,h}(x) \vert
< \vert P_n^{g,h}(x) \vert < \frac{ 3 \, \vert x \vert }{2 \, h(n) } \left\vert P_{n-1}^{g,h}(x)\right| .
\end{equation}
This implies that $P_n^{g,h}(x)\neq 0$ for all $\vert x \vert > \kappa \,\, h(n-1)$ and
$(-1)^n P_n^{g,h}(x) > 0$ if $x<-\kappa h\left( n-1\right) $.
Let $g(n)= \sigma(n)$. In \cite{HN20B} we proved
that $\kappa=15$ can is an acceptable value.
In the following we state our two main results: Improvement A and Improvement B.
\subsection{Improvement A}
The following result
reproduces our previous result (\ref{old}), if we choose $\varepsilon = \frac{1}{2}$.
\begin{theorem}
\label{schwaecher}Let $0<\varepsilon <1$. Let
$R>0$ be the radius of convergence of
\[
G_{1}\left( T\right) =\sum _{k=1}^{\infty }\left| g\left( k+1\right) \right| T^{k}.
\]
Let $0<T_{\varepsilon }<R$ be such that
$G_{1}\left( T_{\varepsilon }\right) \leq \varepsilon $
and $\kappa=\kappa_{\varepsilon }=\frac{1}{1-\varepsilon }\frac{1}{T_{\varepsilon }}$.
Then
\begin{equation}\label{Th1:formel}
\left| P_{n}^{g,h}\left( x\right) - \frac{x}{h\left( n\right) } P_{n-1}^{g,h}\left( x\right) \right| < \varepsilon \frac{\left| x\right| }{h\left( n\right) }\left| P_{n-1}^{g,h}\left( x\right) \right|,
\end{equation}
if
$\left| x \right| > {\kappa}\,\, h(n-1)$ for all
$n\geq 1$.
\end{theorem}
This result can be reformulated in
the following way, which is more suitable for applications
to growth and non-vanishing properties.
\begin{theorem}\label{versionA2}
Let $0<\varepsilon <1$. Let
$R>0$ be the radius of convergence of
\[
G_{1}\left( T\right) =\sum _{k=1}^{\infty }\left| g\left( k+1\right) \right| T^{k}.
\]
Let $0<T_{\varepsilon }<R$ be such that
$G_{1}\left( T_{\varepsilon }\right) \leq \varepsilon $
and $\kappa=\kappa_{\varepsilon }=\frac{1}{1-\varepsilon }\frac{1}{T_{\varepsilon }}$.
Then
\begin{equation}
\left( 1-\varepsilon \right) \frac{\left| x\right|
}{h\left( n\right) }\left|P_{n-1}^{g,h}\left( x\right) \right| <\left| P_{n}^{g,h}\left( x\right) \right| < \left( 1+\varepsilon \right) \frac{\left| x\right|
}{h\left( n\right) }\left| P_{n-1}^{g,h}\left( x\right) \right|,
\end{equation}
if
$\left| x \right| > {\kappa} \,\, h(n-1)
$ for all
$n\geq 1$.
\end{theorem}
\begin{corollary}
\label{schwaechererkorollar} Let $\kappa$ be chosen as in Theorem
\ref{schwaecher} or as in Theorem \ref{versionA2}. Then
$$P_{n}^{g,h}\left( x\right) \neq 0 \text{ for }
\left| x\right| >\kappa \,\, h(n-1).$$
\end{corollary}
\begin{proof}
This follows from Theorem \ref{versionA2}, since
$\left( 1-\varepsilon \right) \frac{\left| x\right|
}{h\left( n\right) } \neq 0$ and $P_0^{g,h}(x)=1$.
\end{proof}
We note that the smallest possible $\kappa$ is independent of the function $h(n)$.
It is also possible to provide a lower bound for the best possible $\kappa$.
\begin{proposition}
The constant $\kappa_{\varepsilon }$ obtained in
Theorem \ref{schwaecher} has the following lower bound:
\[
\kappa_{\varepsilon }\geq \frac{\left| g\left( 2\right) \right| }{\left( 1-\varepsilon \right) \varepsilon }.
\]
As a lower bound\/ \emph{independent} of $\varepsilon $ we
have $4\left| g\left( 2\right) \right| $.
\end{proposition}
\begin{proof}
If we consider only the first
order term of the power series
\[
G_{1}\left( T\right) =\sum _{k=1}^{\infty }\left| g\left( k+1\right) \right| T^{k},
\]
then for positive $T$ we always have
$G_{1}\left( T\right) =\sum _{k=1}^{\infty }\left| g\left( k+1\right) \right| T^{k}\geq \left| g\left( 2\right) \right| T$.
Thus, $G_{1}\left( T\right) >\varepsilon $ if
$T>\frac{\varepsilon }{\left| g\left( 2\right) \right| }$.
The case
$G_{1}\left( T\right) \leq \varepsilon $ is only possible if
$T\leq \frac{\varepsilon }{\left| g\left( 2\right) \right| }$.
This forces
$T_{\varepsilon }\leq \frac{\varepsilon }{\left| g\left( 2\right) \right| }$.
Applying the last inequality now to
\begin{equation}
\kappa_{\varepsilon }:=\frac{1}{\left( 1-\varepsilon \right) T_{\varepsilon }}
\label{eq:kappaschwaecher}
\end{equation}
Theorem~\ref{schwaecher} shows the lower bound
$\kappa_{\varepsilon }\geq \frac{\left| g\left( 2\right) \right| }{\left( 1-\varepsilon \right) \varepsilon }$
in the
proposition depending on $\varepsilon $. The minimal value of this lower bound is
at $\varepsilon =\frac{1}{2}$ because of the inequality of
arithmetic and geometric means
$\left( 1-\varepsilon \right) \varepsilon \leq \left( \frac{1-\varepsilon +\varepsilon }{2}\right) ^{2}=\frac{1}{4}$.
\end{proof}
\subsection{Improvement B}
\begin{theorem}
\label{main}Let $0<\varepsilon <1$. Let
$R>0$ be the radius of convergence of
\[
G_{2}\left( T \right) =\sum _{k=2}^{\infty }\left| g\left( k+1\right) -g\left( 2\right) g\left( k\right) \right| T^{k}.
\]
Let $0<T_{\varepsilon }<R$ be such that
$G_{2}\left( T_{\varepsilon }\right) \leq \varepsilon$ and
$$ \kappa=\kappa_{\varepsilon }:=\frac{1}{1-\varepsilon }\left( \frac{1}{T_{\varepsilon }}+\left| g\left( 2\right) \right| \right). $$
Then
\begin{equation}\label{B1}
\left| P_{n}^{g,h}\left( x\right) - \frac{x+g\left( 2\right) h\left( n-1\right) }{h\left( n\right) } P_{n-1}^{g,h}\left( x\right) \right| < \varepsilon \frac{\left| x\right| }{h\left( n\right) }\left| P_{n-1}^{g,h}\left( x\right) \right|
\end{equation}
if
$\left| x \right| > {\kappa} \,\, h(n-1)
$ for all $n\geq 1$.
\end{theorem}
\begin{theorem}
\label{variantB} Let $0 <\varepsilon <1$. Let
$R>0$ be the radius of convergence of
\[
G_{2}\left( T \right) =\sum _{k=2}^{\infty }\left| g\left( k+1\right) -g\left( 2\right) g\left( k\right) \right| T^{k}.
\]
Let $0<T_{\varepsilon }<R$ be such that
$G_{2}\left( T_{\varepsilon }\right) \leq \varepsilon$ and
$$ \kappa=\kappa_{\varepsilon }:=\frac{1}{1-\varepsilon }\left( \frac{1}{T_{\varepsilon }}+\left| g\left( 2\right) \right| \right). $$
Then
\begin{eqnarray}
&&\frac{
\left| x+g\left( 2\right) h\left( n-1\right) \right| -\varepsilon \left| x\right|
}{h\left( n\right) }\left|P_{n-1}^{g,h}\left( x\right) \right| \nonumber \\
&<&\left| P_{n}^{g,h}\left( x\right) \right| < \frac{
\left| x+g\left( 2\right) h\left( n-1\right) \right| +\varepsilon \left| x\right|
}{h\left( n\right) }\left| P_{n-1}^{g,h}\left( x\right) \right| \label{B2}
\end{eqnarray}
if
$\left| x \right| > {\kappa} \,\, h (n-1)
$ for all
$n\geq 1$.
\end{theorem}
\begin{corollary}
\label{hauptkorollar} Let $\kappa$ be chosen as in Theorem
\ref{main} or as in Theorem \ref{variantB}. Then
\begin{equation} \label{B3} P_{n}^{g,h}\left( x\right) \neq 0 \text{ for }
\left| x\right| >\kappa \,\, h(n-1).
\end{equation}
\end{corollary}
\begin{proposition}
\label{untereschranke} The constant $\kappa_{\varepsilon }$ obtained in
Theorem \ref{main} has the following lower bound:
$$\kappa_{\varepsilon }\geq \frac{1}{1-\varepsilon }\left( \sqrt{\frac{\left| \left( g\left( 2\right) \right) ^{2}-g\left( 3\right) \right| }{\varepsilon }}+\left| g\left( 2\right) \right| \right).
$$
As a lower bound\/ \emph{independent} of $\varepsilon $ we
have $\frac{3}{2}\sqrt{3\left| \left( g\left( 2\right) \right) ^{2}-g\left( 3\right) \right| }+\left| g\left( 2\right) \right| $.
\end{proposition}
\begin{proof}
If we consider only the second order term of the power series
$G_{2}\left( T\right) =\sum _{k=2}^{\infty }\left| g\left( k+1\right) -g\left( 2\right) g\left( k\right) \right| T^{k}$,
then for positive $T$ we always have
\[
G_{2}\left( T\right) =\sum _{k=2}^{\infty }\left| g\left( k+1\right) -g\left( 2\right) g\left( k\right) \right| T^{k}\geq \left| \left( g\left( 2\right) \right) ^{2}-g\left( 3\right) \right| T^{2}.
\]
Thus, $G_{2}\left( T\right) >\varepsilon $ if
$T>\sqrt{\frac{\varepsilon }{\left| \left( g\left( 2\right) \right) ^{2}-g\left( 3\right) \right| }}$.
The case
$G_{2}\left( T\right) \leq \varepsilon $ is only possible if
$T\leq \sqrt{\frac{\varepsilon }{\left| \left( g\left( 2\right) \right) ^{2}-g\left( 3\right) \right| }}$.
This forces
$T_{\varepsilon }\leq \sqrt{\frac{\varepsilon }{\left| \left( g\left( 2\right) \right) ^{2}-g\left( 3\right) \right| }}$.
Applying the last inequality now to
\begin{equation}
\kappa_{\varepsilon }:=\frac{1}{1-\varepsilon }\left( \frac{1}{T_{\varepsilon }}+\left| g\left( 2\right) \right| \right)
\label{eq:hauptkappa}
\end{equation}
from Theorem~\ref{main} shows the lower bound
$\kappa_{\varepsilon }\geq \frac{1}{1-\varepsilon }\left( \sqrt{\frac{\left| \left( g\left( 2\right) \right) ^{2}-g\left( 3\right) \right| }{\varepsilon }}+\left| g\left( 2\right) \right| \right) $
in the
proposition depending on $\varepsilon $.
It is clear that
\begin{equation}\frac{1}{1-\varepsilon }\left( \sqrt{\frac{\left| \left( g\left( 2\right) \right) ^{2}-g\left( 3\right) \right| }{\varepsilon }}+\left| g\left( 2\right) \right| \right) \geq \frac{1}{1-\varepsilon }\sqrt{\frac{\left| \left( g\left( 2\right) \right) ^{2}-g\left( 3\right) \right| }{\varepsilon }}+\left| g\left( 2\right) \right|
\label{eq:abschaetzung}
\end{equation}
for $0<\varepsilon <1$.
To estimate
$\kappa _{\varepsilon }$
independent of $\varepsilon $
we consider the right hand side of the last inequality as a
function in $\varepsilon $. Thus, we are interested in the
minimal value of this function for $0<\varepsilon <1$.
The inequality of arithmetic and geometric means yields
\begin{eqnarray*}
\left( 1-\varepsilon \right) \varepsilon ^{1/2}&=&2\left( \left( 1-\varepsilon \right) /2\right) ^{1/2}\cdot \left( \left( 1-\varepsilon \right) /2\right) ^{1/2}\cdot \varepsilon \\
&\leq &2\left( \frac{\left( 1-\varepsilon \right) /2+\left( 1-\varepsilon \right) /2+\varepsilon }{3}\right) ^{3/2}=\frac{2}{3\sqrt{3}}.
\end{eqnarray*}
We obtain
$\frac{3}{2}\sqrt{3\left| \left( g\left( 2\right) \right) ^{2}-g\left( 3\right) \right| }+\left| g\left( 2\right) \right| $.
\end{proof}
\subsection{Comparing Improvement A and Improvement B}
Let $0<\varepsilon _{1}<1$ and $T_{\varepsilon _{1}}$ as in
Theorem~\ref{schwaecher}. For all $T\geq 0$ we have that
\begin{eqnarray*}
G_{2}\left( T\right)
&\leq & \sum _{k=2}^{\infty }\left( \left| g\left( k+1\right) \right| +\left| g\left( 2\right) g\left( k\right) \right| \right) T^{k} \\
&=&
\left( 1+\left| g\left( 2\right) \right| T\right) G_{1}\left( T\right) -\left| g\left( 2\right) \right| T.
\end{eqnarray*}
Let
$\varepsilon _{2}$ be such that
\[
\left( 1+\left| g\left( 2\right) \right|
T_{\varepsilon _{1}}\right) G_{1}\left( T_{\varepsilon _{1}}\right) -\left| g\left( 2\right) \right| T_{\varepsilon _{1}}\leq \varepsilon _{2}\leq \left( 1+\left| g\left( 2\right) \right| T_{\varepsilon _{1}}\right) \varepsilon _{1}-\left| g\left( 2\right) \right| T_{\varepsilon _{1}}
<1.
\]
Then
\[
0\leq G_{2}\left( T_{\varepsilon _{1}}\right) \leq \left( 1+\left| g\left( 2\right) \right| T_{\varepsilon _{1}}\right) G_{1}\left( T_{\varepsilon _{1}}\right) -\left| g\left( 2\right) \right| T_{\varepsilon _{1}}
\leq
\varepsilon _{2}.
\]
This shows that we can choose
$T_{\varepsilon _{2}}=T_{\varepsilon _{1}}$.
Let $\kappa _{1,\varepsilon }$ and $\kappa _{2,\varepsilon }$
be the respective constants from Theorems~\ref{schwaecher}
and~\ref{main}. Then
\begin{eqnarray*}
\kappa _{2,\varepsilon _{2}}&=&\frac{1}{1-\varepsilon _{2}}\left( \frac{1}{T_{\varepsilon _{1}}}+\left| g\left( 2\right) \right| \right)
=
\frac{1}{1-\varepsilon _{2}}\left( 1+\left| g\left( 2\right) \right| T_{\varepsilon _{1}}\right) \frac{1}{T_{\varepsilon _{1}}} \\
&\leq
&\frac{1}{1-\left( 1+\left| g\left( 2\right) \right| T_{\varepsilon _{1}}\right) \varepsilon _{1}+\left| g\left( 2\right) \right| T_{\varepsilon _{1}}
}\left( 1+\left| g\left( 2\right) \right| T_{\varepsilon _{1}}\right) \frac{1}{T_{\varepsilon _{1}}} \\
&=&\frac{1}{1-\varepsilon _{1}}\frac{1}{T_{\varepsilon _{1}}}=\kappa _{1,\varepsilon _{1}}.
\end{eqnarray*}
This shows that the minimal value of the
$\kappa _{2,\varepsilon }$ is never larger than the
minimal value of the $\kappa _{1,\varepsilon }$.
\section{Applications}
\subsection{Toy example}
Let us consider the case $g(n)=1$ for all $n \in \mathbb{N}$.
We observe that $G_2(T)=0$
for all $T$. Let $0< \varepsilon < 1$. Then we apply Theorem \ref{variantB}.
For all $\vert x \vert > \frac{1}{1- \varepsilon} h(n-1)$ we obtain
\begin{equation*}
\frac{
\left| x+ h\left( n-1\right) \right| -\varepsilon \left| x\right|
}{h\left( n\right) }\left|P_{n-1}^{\mathbf{1},h}\left( x\right) \right|
<\left| P_{n}^{\mathbf{1},h}\left( x\right) \right| < \frac{
\left| x+ h\left( n-1\right) \right| +\varepsilon \left| x\right|
}{h\left( n\right) }\left| P_{n-1}^{\mathbf{1},h}\left( x\right) \right|.
\end{equation*}
Let $\varepsilon \rightarrow 0$, then for all $\vert x \vert > h(n-1)$:
\begin{equation*}
\frac{
\left| x+ h\left( n-1\right) \right|
}{h\left( n\right) }\left|P_{n-1}^{\mathbf{1},h}\left( x\right) \right|
\leq \left| P_{n}^{\mathbf{1},h}\left( x\right) \right| \leq \frac{
\left| x+ h\left( n-1\right) \right|
}{h\left( n\right) } \left| P_{n-1}^{\mathbf{1},h}\left( x\right) \right|.
\end{equation*}
Then,
$\left\vert P_{n}^{\mathbf{1},h}\left( x\right) \right| = \prod_{k=0}^{n-1} \frac{\vert x + h(k)\vert}{h(k+1)}$ (we define $h(0):=0$).
Since $P_{1}^{\mathbf{1},h}\left( x\right) = x/h(1)$ and
$P_{n}^{\mathbf{1},h}\left( x\right) $ is a polynomial
of degree $n$ with positive leading coefficient, it follows:
\begin{equation}
P_n^{\mathbf{1},h}(x) = \frac{ x (x+h(1))
\cdots (x+h(n-1))}{h(1)
\cdots h(n)}.
\end{equation}
\subsection{Reciprocals of Eisenstein series}
Let $\sigma_{k}(n)= \sum_{d \vert n} d^{k}$ and let $B_{k}$ be the $k$th Bernoulli number. Then
we define Eisenstein series of weight $k$:
\begin{equation}
E_k(\tau):= 1 - \frac{2k}{B_k} \sum_{n=1}^{\infty} \sigma_{k-1}(n) \, q^n \qquad (k=2,4,6,\ldots).
\end{equation}
In \cite{AKN97} it was indicated that the $q$-expansion of the reciprocal of $E_4(\tau)=1 +240 \sum_{n=1}^{\infty} \sigma_3(n) \, q^n$
given by
\begin{equation}
\frac{1}{E_4(\tau)} = \sum_{n=0}^{\infty} \beta_n \, q^n,
\end{equation}
has strictly alternating sign changes: $(-1)^n \beta_n >0$.
Let $\varepsilon_1= \frac{1}{25}
$ and $\varepsilon_2= \frac{1}{982}
$. We can chose $\kappa$ in Theorem 1--4, such that $240 > \kappa$.
(In both cases $T_{\varepsilon }=\frac{87}{20000}$ does the job.
Then $\kappa _{1}=\frac{62500}{261}\approx 239.46$
and $\kappa _{2}=\frac{20408906}{85347}\approx 239.13$.
Note that an approximation of the smallest possible value that
can be obtained by our method is
$\kappa _{2}=\frac{539}{16}\approx 33.7$. This we obtain for
$\varepsilon _{2}=\frac{5}{21}$ and
$T_{\varepsilon _{2}}=\frac{3}{20}$.)
\begin{proof}[Proof of $\kappa _{2}\leq \frac{20408906}{85347}$]
Let $T_{\varepsilon }=\frac{87}{20000}$. Let further $\varepsilon _{1}=\frac{1}{25}$
and $\varepsilon _{2}=\frac{1}{982}$.
We have the well-known estimate
\begin{equation}
\sigma _{3}\left( k\right)
\leq
\left( 1+\int _{1}^{
\infty }t^{-3}\,\mathrm{d}t\right) k^{3}=
3k^{3}/2
.
\label{eq:sigma3}
\end{equation}
Thus,
$\sigma _{3}\left( k\right) \leq 3k^{3}/2\leq 9\binom{k+2}{3}$.
Let $c_{1}\left( k\right) =\sigma _{3}\left( k+1\right) $
for $k\leq 2$ and
$c_{1}\left( k\right) =9\binom{k+3}{3}$ for $k\geq 3$.
Then
$G_{1}\left( T\right) \leq \sum _{k=1}^{\infty }c_{1}\left( k\right) T^{k}=9\frac{1}{\left( 1-T\right) ^{4}}-9-27T-62T^{2}$
and
$$G_{1}\left( \frac{87}{20000}\right) \leq \frac{1248274072444709335238721}{31446822595409952200000000}<\frac{1}{25}.$$
Thus,
$\kappa _{1}\leq \frac{20000}{87}\frac{25}{24}=\frac{62500}{261}\approx 239.46$.
With (\ref{eq:sigma3}) it also follows that
$
\left| 9\sigma _{3}\left( k\right) -\sigma _{3}\left( k+1\right) \right| \leq 15\left( k+1\right) ^{3}\leq 90\binom{k+3}{3}$.
Let $c_{2}\left( k\right) =\left| 9\sigma _{3}\left( k\right) -\sigma _{3}\left( k+1\right) \right| $
for $k\leq 4$ and $c_{2}\left( k\right) =90\binom{k+3}{3}$
for $k\geq 5$. Then
$G_{2}\left( T\right) \leq \sum _{k=2}^{\infty }c_{2}\left( k\right) T^{k}=\frac{90}{\left( 1-T\right) ^{4}}-90-360T-847T^{2}-1621T^{3}-2619T^{4}$
for $T>0$ and $$G_{2}\left( \frac{87}{20000}\right) \leq
\frac{25605878110865247894531439480101}{25157458076327961760000000000000000}<\frac{1}{982}.$$
Thus,
$\kappa _{2}\leq \left( \frac{20000}{87}+9\right)
\frac{982}{981}=\frac{20408906}{85347} \approx 239.13$.
\end{proof}
\ \newline \
Note that $\beta_1= -240$, $\beta_n \in \mathbb{Z}$ and $\beta_1 \mid \beta_n$ for all $n \geq 1$. From
(\ref{old}), Theorem 1--4 and Corollary \ref{schwaechererkorollar} the following properties are obtained.
\begin{eqnarray*}
\frac{1}{2} \vert \beta_1 \beta_{n-1} \vert < & \vert \beta_n \vert & < \frac{3}{2} \vert \beta_1 \beta_{n-1} \vert, \\
\vert \beta_n - \beta_1 \vert \, \left| \beta_{n-1} \right\vert & < & \varepsilon_1 \, \vert \beta_1 \beta_{n-1} \vert,\\
(1 - \varepsilon_1) \vert \beta_1 \beta_{n-1} \vert < & \vert \beta_n \vert & < (1 + \varepsilon_1) \vert \beta_1 \beta_{n-1} \vert, \\
\vert \beta_n -( \beta_1 + 9) \vert & < & \varepsilon_2 \vert \beta_1 \beta_{n-1} \vert,\\
\vert 231 + \varepsilon_2 \beta_1 \vert \, \vert \beta_{n-1} \vert < & \vert \beta_n \vert & <
\vert 231 - \varepsilon_2 \beta_1 \vert \, \vert \beta_{n-1}\vert.
\end{eqnarray*}
Since $\beta_0 =1$ we can deduce that $(-1)^n \beta_n >0$.
In the previous proof we showed that
$G_{2}\left( T_{\varepsilon }\right) <\frac{1}{982}<\frac{1}{250}$
for $T_{\varepsilon }=\frac{87}{20000}$
and $\kappa _{2}<240$.
This leads to the following
\begin{theorem}[\cite{HN20C}]
\label{abschaetzung}Let
$G_{2}\left( T
\right)$ be defined by
$$ \sum _{m=2}^{\infty }\left| \sigma_{3} \left( m+1\right) - 9
\sigma_{3} \left( m\right) \right| T^m $$
with positive radius of convergence $R$. Suppose that there is
$0<T_{\varepsilon }<1$ such that
$G_{2}\left( T_{\varepsilon }\right) \leq \frac{1}{250} $ and
$\kappa_{2 } \leq \frac{250}{249}\left( \frac{1}{T_{\varepsilon }}+ \sigma_{3} \left( 2\right) \right) <\frac{8}{\left| B_{4}\right| }=240$,
then
the absolute value of the
$n$th coefficient $\beta_n$ of $1/E_{4}$
can be estimated by
\begin{equation}
240
\left( \left( 1\pm \frac{1}{250}
\frac{240}{231}
\right)
231
\right) ^{n-1}.
\end{equation}
\end{theorem}
This implies
\begin{equation}
230^{n-1}\leq \frac{\left( -1\right) ^{n}\beta _{n}}{240}\leq 232^{n-1}.
\label{eq:beta}
\end{equation}
The following table displays the first values.
\begin{center}
\begin{minipage}[t]{1.0\textwidth}
{\small
\[
\begin{array}{|r||r|r|r|}
\hline
n & 230^{n-1} & \frac{\beta _{n}}{240} & 232^{n-1} \\ \hline \hline
1 & 1 & -1 & 1 \\ \hline
2 & 230 & 231 & 232 \\ \hline
3 & 52900 & -53308 & 53824 \\ \hline
4 & 12167000 & 12301607 & 12487168 \\ \hline
5 & 2798410000 & -2838775326 & 2897022976 \\ \hline
6 & 643634300000 & 655088819748 & 672109330432 \\ \hline
7 & 148035889000000 & -151171301803544 & 155929364660224 \\ \hline
8 & 34048254470000000 & 34884983226375975 & 36175612601171968 \\ \hline
9 & 7831098528100000000 & -8050218792755033557 & 8392742123471896576 \\ \hline
10 & 1801152661463000000000 & 1857705425589167301906 & 1947116172645480005632 \\
\hline
\end{array}
\]}
\captionsetup{margin={0cm,0cm,0cm,0cm}}
\captionof{table}{Estimation given by (\ref{eq:beta})
}
\end{minipage}
\end{center}
By dividing $\beta _{n}$ by the estimates we obtain
the figures displayed in Table \ref{norm}:
\begin{center}
\begin{minipage}[t]{1.0\textwidth}
\[
\begin{array}{|r||r|r|}
\hline
n & \frac{\beta _{n}}{240\cdot 230^{n-1}} & \frac{\beta _{n}}{240\cdot 232^{n-1}} \\ \hline \hline
0 & -1.00000000 & -1.00000000 \\ \hline
1 & 1.00434783 & 0.99568966 \\ \hline
2 & -1.00771267 & -0.99041320 \\ \hline
3 & 1.01106329 & 0.98513987 \\ \hline
4 & -1.01442438 & -0.97989396 \\ \hline
5 & 1.01779663 & 0.97467598 \\ \hline
6 & -1.02118009 & -0.96948578 \\ \hline
7 & 1.02457479 & 0.96432322 \\ \hline
8 & -1.02798078 & -0.95918815 \\ \hline
9 & 1.03139810 & 0.95408043 \\ \hline
\end{array}
\]
\captionsetup{margin={0cm,0cm,0cm,0cm}}
\captionof{table}{Normalization} \label{norm}
\end{minipage}
\end{center}
\begin{remark}
The value $\varepsilon _{2}=\frac{1}{982}$
improves the inequalities (\ref{eq:beta}) to
\[
230.7648^{n-1}\leq \frac{\left( -1\right) ^{n}\beta _{n}}{240}\leq 231.2353^{n-1}.
\]
The lower bound is quite close to the
optimal value $e^{\pi \sqrt{3}}
=230.764588\ldots $.
\end{remark}
\subsection{Associated Laguerre polynomials and Chebyshev polynomials of the second kind}
We briefly recall the definition of associated Laguerre polynomials $L_n^{(\alpha)}(x)$
and Chebyshev polynomials $U_n(x)$ of the second kind \cite{RS02, Do16}. Both are orthogonal polynomials.
We have
\begin{equation}
L_{n}^{\left( \alpha \right) }\left( x\right) =
\sum _{k=0}^{n}\binom{n+\alpha }{n-k}\frac{(-x)^{k}}{k!} \qquad (\alpha > -1).
\end{equation}
The Chebyshev polynomials are uniquely characterized by
\begin{equation}
U_{n}(\text{cos}(t)) = \frac{\text{sin}((n+1)t)}{\text{sin}(t)} \qquad ( 0 < t < \pi).
\end{equation}
The Chebyshev polynomials are of special interest in the context of applications,
since they are the only classical orthogonal polynomials whose
zeros can be determined in explicit form (see Rahman and Schmeisser \cite{RS02}, Introduction).
Let $g(n)=\func{id}(n)= n$. Then
\begin{eqnarray}
P_n^{\func{id}}(x) & =& \frac{x}{n} L_{n-1}^{(1)}(-x),\\
Q_n^{\func{id}}(x) & =& x \, U_{n-1}\left(\frac{x}{2} + 1 \right).
\label{eq:tschebyscheff}
\end{eqnarray}
The generating series of the Chebyshev polynomial of the second kind is given by
\begin{equation}
\sum_{n=0}^{\infty} U_n(x) \, q^n = \frac{1}{1 - 2x q + q^2}, \qquad \vert x \vert, \vert q \vert <1.
\end{equation}
With this we can
prove equation (\ref{eq:tschebyscheff}). We have
\begin{eqnarray*}
1+xq\sum _{n=0}^{\infty }U_{n}\left( \frac{x}{2}+1\right) q^{n}&=&
1+\frac{xq}{1-\left( 2+x\right) q+q^{2}}=
\frac{1-2q+q^{2}}{1-\left( 2+x\right) q+q^{2}}\\
&=&\frac{1}{1-xq\frac{1}{\left( 1-q\right) ^{2}}}
=
\frac{1}{1-xq\sum _{n=1}^{\infty }nq^{n-1}}\\
&=&\sum _{n=0}^{\infty }Q_{n}\left( x\right) q^{n}
\end{eqnarray*}
using Definition (\ref{Q}).
Note that
$G_{1}\left( T\right) =\sum _{k=1}^{\infty }\left( k+1\right) T^{k}=\frac{1}{\left( 1-T\right) ^{2}}-1$
and
\[
G_{2}\left( T\right) =\sum _{k=2}^{\infty }\left( k-1\right) T^{k}=\frac{T^{2}}{\left( 1-T\right) ^{2}}.
\]
From this
we obtain the
following values:
\begin{center}
\begin{minipage}[t]{0.6\textwidth}
\[
\begin{array}{|c|c|c||c|c|c|}
\hline
\varepsilon _{1}&T_{\varepsilon _{1}}&\kappa _{1}&\varepsilon _{2}&T_{\varepsilon _{2}}&\kappa _{2} \\ \hline
\vphantom{\int }\frac{11}{25}&\frac{1}{6}&\frac{75}{7}&\frac{1
}{4
}&\frac{1
}{3
}&\frac{20
}{3
} \\ \hline
\end{array}
\]
\captionsetup{margin={0cm,0cm,0cm,0cm}}
\captionof{table}{Case $g(n)=n$}
\end{minipage}
\end{center}
If we consider the special case $\varepsilon_1 = 1/2$ in Improvement A, we can chose $T_{\varepsilon _{1}}=2/11$ and finally get $\kappa_1 = 11$.
This
leads to several applications. For example, let $\vert x \vert >\left( 20/3\right) \, n $ then $L_n^{(1)}(x) \neq 0$ and the estimates hold
\begin{equation*}
\left|(\vert x + 2n \vert - 1/4 \vert x \vert )\right| \,\, \vert L_{n-1}^{(1)}(x)\vert < n \vert L_{n}^{(1)}(x)\vert <
|(\vert x + 2n \vert + 1/4 \vert x \vert) | \,\, \vert L_{n-1}^{(1)}(x)\vert.
\end{equation*}
\subsection{Powers of the Dedekind $\eta$-function.}
Let us recall the well-known identity:
\begin{equation}
\prod_{n=1}^{\infty} \left( 1 - q^n \right) = \text{exp} \left( - \sum_{n=1}^{\infty} \sigma(n) \, \frac{q^n}{n} \right) \qquad (z \in \mathbb{C}).
\end{equation}
The $q$-expansion of the $-z$th power of the Euler product defines the
D'Arcais polynomials
\begin{equation}
\sum_{n=0}^{\infty} P_n^{\sigma}(z) \, q^n = \prod_{n=1}^{\infty} \left( 1 - q^n \right)^{-z} \qquad ( z \in \mathbb{C}),
\end{equation}
where $P_0^{\sigma}(x)=1$ and
$P_n^{\sigma}\left( x\right) = \frac{x}{n} \sum_{k=1}^{n} \sigma(k) P_{n-k}^{\sigma}(x)$, as polynomials.
Note that these polynomials evaluated at $-24$ are directly related to the Ramanujan $\tau$-function: $\tau(n)= P_{n-1}^{\sigma}(-24)$,
which gives also a link to the Lehmer conjecture \cite{Le47}.
In the spirit of this paper, let $\varepsilon :=\frac{3}{14}$. Then $T_{\varepsilon }: =\frac{2}{11}$
satisfies the assumptions of Theorem \ref{variantB}. We obtain the
\begin{corollary}\label{top}
Let $\kappa = \frac{119}{11}$. Then $P_n^{\sigma}(z) \neq 0$ for all complex $z$ with $\vert z \vert > \kappa \,\, (n-1)$.
\end{corollary}
We have to show that
$G_2\left( T_{\varepsilon }\right) =
\sum _{k=2}^{\infty }\left| \sigma \left( k+1\right) -3\sigma \left( k\right) \right| T_{\varepsilon }^{k}<\varepsilon $.
For this let
$c\left( k\right) =\left| \sigma \left( k+1\right) -3\sigma \left( k\right) \right| $
for $1\leq k\leq 7$ and
$c\left( k\right) =4\binom{
k+2}{2}
$ for $k\geq 8$. Then
$\left| \sigma \left( k+1\right) -3\sigma \left( k\right) \right| \leq c\left( k\right) $
for all $k\in \mathbb{N}$ since
$$\sigma \left( k\right) \leq \left( 1+\ln \left( k\right) \right) k\leq \left( \frac{k}{4}+\ln \left( 4\right) \right) k\leq \binom{k+1}{2}$$
for $k\geq 4$.
This
implies
$G_2\left( T \right) \leq \sum _{k=2}^{\infty }c\left( k\right) T^{k}$
for $0\leq T \leq 1\leq R$. The upper bound is
now almost, except for the first $8$ terms, a multiple of the second derivative of the geometric
series of $T$. Hence,
\begin{equation*} G_2(T) \leq
\frac{4}{\left( 1-T\right) ^{3}}-4
-12T-
19T^{2}-
35T^{3}-45T^{4}-78T^{5}-84T^{6}-135T^{7}.
\end{equation*}
For $T=T_{\varepsilon }=\frac{2}{11}$ we obtain
\begin{equation*}
G_2 \left( T_{\varepsilon }\right) \leq \frac{3043993780}{14206147659} < \frac{3}{14}=\varepsilon .
\end{equation*}
The claim now follows from Corollary \ref{hauptkorollar}.
\begin{remark}
\begin{itemize}
\item[]
\item[a)]Let $\varepsilon$ and $\kappa$ be as above, and let
$h$ be an arbitrary arithmetic function with $0 < h(n) \leq h(n+1)$.
Then $P_n^{\sigma,h}(x)$ satisfies (\ref{B1}), (\ref{B2}), and (\ref{B3}) obtained by Improvement B.
\item[b)] The value $\varepsilon = \frac{3}{14}$ already leads to
$$\kappa_{\varepsilon} = \frac{119}{11} = 10.\overline{81}.$$ Note only minor further improvements can be achieved.
\item[c)] Corollary \ref{top} improves our previous result \cite{HN20B}, where $\kappa =15$.
\end{itemize}
\end{remark}
\begin{proposition}
Let $\varepsilon = 0.217$ and $T_{\varepsilon}= 0.18289$. Then the assumptions of Theorem \ref{main}
are fulfilled. Furthermore we can take $\kappa=10.815$.
\end{proposition}
\begin{proof}
Let $\varepsilon$ and $T_{\varepsilon}$ be given.
We have to show that
$$G_2\left( T_{\varepsilon }\right) =\sum _{k=2}^{\infty }\left| \sigma \left( k+1\right) -
3\sigma \left( k\right) \right| T_{\varepsilon }^{k}<\varepsilon . $$
Let $c\left( k\right) =\left| \sigma \left( k+1\right) -3\sigma \left( k\right) \right| $
for $1\leq k\leq 11$ and
$c\left( k\right) =4\binom{
k+2}{2}
$ for $k\geq 12$. Then
$\left| \sigma \left( k+1\right) -3\sigma \left( k\right) \right| \leq c\left( k\right) $
for all $k\in \mathbb{N}$ as
\[
\sigma \left( k\right) \leq \left( 1+\ln \left( k\right) \right) k\leq \left( \frac{k}{4}+\ln \left( 4\right) \right) k\leq \binom{k+1}{2}
\]
for $k\geq 4$.
This
implies
$G_2\left( q\right) \leq \sum _{k=2}^{\infty }c\left( k\right) T^{k}$
for $0\leq T \leq 1\leq R$. The upper bound is almost
(except for the first $12$ terms) a multiple
of the second derivative of the geometric
series of $T$. Hence $G_2(T) \leq \sum _{k=2}^{\infty }c\left( k\right) T^{k} \leq$
\begin{eqnarray*}
&&4\sum _{k=0}^{\infty }\binom{k+2}{2} T^{k}-4-12T-19T^{2}-35T^{3}-45T^{4}-78T^{5}-84T^{6}-135T^{7} \\
&&{}-148T^{8}-199T^{9}-222T^{10}-304T^{11} \\
&=&
\frac{4}{\left( 1-T\right) ^{3}}-4
-12T-
19T^{2}-
35T^{3}-45T^{4}-78T^{5}-84T^{6}-135T^{7} \\
&&{}-148T^{8}-199T^{9}-222T^{10}-304T^{11}.
\end{eqnarray*}
For $T=T_{\varepsilon }=0.18289$ we obtain
\begin{equation*}
G_2
\left( T_{\varepsilon }\right)
< 0.216998<\varepsilon .
\end{equation*}
The claim now follows from
Corollary~\ref{hauptkorollar}.
\end{proof}
\begin{minipage}[t]{0.99\textwidth}
\[
\begin{array}{|c||c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
k & 1 &
2 &
3 &
4 &
5 &
6 &
7 &
8 &
9 &
10 &
11 &
12 &
13 &
14 \\ \hline
\left| \sigma \left( k+1\right) -3\sigma \left( k\right) \right| & 0 &
5 &
5 &
15 &
6 &
28 &
9 &
32 &
21 &
42 &
8 &
70 &
18 &
48 \\ \hline
\end{array}
\]
\captionsetup{margin={0cm,0cm,0cm,0cm}}
\captionof{table}{Values of $\left| \sigma \left( k+1\right) -3\sigma \left( k\right) \right| $}
\end{minipage}
\section{Proof of Theorem \ref{schwaecher} and Theorem \ref{versionA2}}
\begin{proof}[Proof of Theorem~\ref{schwaecher}]
The proof will be by induction on $n$.
The case $n=1$ is obvious:
$\left| P_{1}^{g,h}\left( x\right) - \frac{x}{h\left( 1\right) }P_{0}^{g,h}\left( x\right) \right| =0<\varepsilon \frac{\left| x\right| }{h\left( 1\right) }\left| P_{0}^{g,h}\left( x\right) \right| $
for
$\left| x\right| >\kappa \,\, h( 0)$.
Let now $n\geq 2$. Then
\[
P_{n}^{g,h}\left( x\right)
=
\frac{x}{h\left( n\right) }\left( P_{n-1}^{g,h}\left( x\right) +\sum _{k=1}^{n-1}g\left( k+1\right) P_{n-1-k}^{g,h}\left( x\right) \right) .
\]
The basic idea for the induction step is to use the
inequality
\[
\left| P_{n}^{g,h}\left( x\right) -\frac{x}{h\left( n\right) }P_{n-1}^{g,h}\left( x\right) \right|
\leq
\frac{\left| x\right| }{h\left( n\right) }\sum _{k=1}^{n-1}\left| g
\left( k+1\right) \right| \left| P_{n-1-k}^{g,h}\left( x\right) \right| .
\]
We estimate the sum by the following property for $1\leq j\leq n-1$:
\begin{eqnarray*}
\left| P_{j}^{g,h}\left( x\right) \right|
&\geq &
\left| \frac{x}{h\left( j\right) }\right| \left| P_{j-1}^{g,h}\left( x\right) \right| -\left| P_{j}^{g,h}-\frac{x}{h\left( j\right) }P_{j-1}^{g,h}\left( x\right) \right| \\
&> &\left( \frac{\left| x\right| }{h\left( j\right) }-\varepsilon \frac{\left| x\right| }{h\left( j\right) }\right) \left| P_{j-1}^{g,h}\left( x\right) \right| \\
&=&\frac{\left( 1-\varepsilon \right) \left| x\right| }{h\left( j\right) }\left| P_{j-1}^{g,h}\left( x\right) \right| \end{eqnarray*}
for
$\left| x\right| > \kappa \,\, h(n-1)
$.
Thus,
\[
\left| P_{j-1}^{g,h}\left( x\right) \right|
<\frac{h\left( j\right) }{\left( 1-\varepsilon \right) \left| x\right| }\left| P_{j}^{g,h}\left( x\right) \right|.
\]
Further, we have
\begin{eqnarray*}
\left| P_{n-k}^{g,h}\left( x\right) \right|
&<&\left| P_{n-k+1}^{g,h}\left( x\right) \right| \frac{h\left( n-k+1\right)
}{\left( 1-\varepsilon \right) \left| x\right| }<\ldots \\
&<&\left| P_{n-1}^{g,h}\left( x\right) \right| \prod _{j=1}^{k-1}\frac{h\left( n-j\right)
}{\left( 1-\varepsilon \right) \left| x\right| } \\
&\leq &\left| P_{n-1}^{g,h}\left( x\right) \right| \left( \frac{h\left( n-1\right) }{\left( 1-\varepsilon \right) \left| x \right| }\right) ^{k-1}
\end{eqnarray*}
for
$\left| x\right| >\kappa \,\, h (n-1) \geq \kappa \,\, h( n-k)$
for all $2 \leq k \leq n$ by assumption.
Using this, we can now estimate the sum by
\[
\sum _{k=1}^{n-1}\left| g
\left( k+1\right) \right| \left| P_{n-1-k}^{g,h}\left( x\right) \right|
<
\left| P_{n-1}^{g,h}\left( x\right) \right| \sum _{k=2}^{n-1}\left| g
\left( k+1\right) \right| \left( \frac{h\left( n-1\right)
}{\left( 1-\varepsilon \right) \left| x\right| }\right) ^{k}
\]
and we obtain
\begin{eqnarray*}
&&\left| P_{n}^{g,h}\left( x\right) -\frac{x}{h\left( n\right) }P_{n-1}^{g,h}\left( x\right) \right| \\
&<&
\frac{\left| x\right| }{h\left( n\right) }\left| P_{n-1}^{g,h}\left( x\right) \right| \sum _{k=1}^{n-1}\left| g\left( k+1\right) \right| \left( \frac{h\left( n-1\right)
}{\left( 1-\varepsilon \right) \left| x\right| }\right) ^{k}.
\end{eqnarray*}
Estimating the sum using the
assumption from the theorem we obtain
\[
\sum_{k=1}^{n-1}
\left| g(k+1) \right| \left( \frac{h(n-1)}{\left( 1-\varepsilon \right)
\left| x \right|} \right)^{k}
\leq
G_{1}
\left( \frac{h\left( n-1\right)
}{\left( 1-\varepsilon \right) \left| x\right| }\right)
\leq G_{1}\left( T_{\varepsilon }\right) \leq
\varepsilon ,
\]
since
$\left| x\right| >\kappa \,\, h(n-1) =\frac{h\left( n-1\right) }{1-\varepsilon }\frac{1}{T_{\varepsilon }}$
which is equivalent to
$\frac{\left( 1-\varepsilon \right) \left| x\right| }{h\left( n-1\right) }>\frac{1}{T_{\varepsilon }}$
and $G_{1}$ increases on $\left[ 0,R\right) $
as $\left| g\left( k+1\right) \right| \geq 0$ for all $k\in \mathbb{N}$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{versionA2}]
Consider the following upper and lower bounds:
\begin{eqnarray*}
\vert P^{g,h}_{n}\left( x\right) \vert & \leq &
\left|\frac{x}{h\left( n\right) }P^{g,h}_{n-1}\left( x\right) \right|
+ \left| P^{g,h}_{n}\left( x\right) -\frac{x}{h\left( n
\right) }P^{g,h}_{n-1}\left( x\right) \right|, \\
\vert P^{g,h}_{n}\left( x\right) \vert & \geq &
\left|\frac{x}{h\left( n\right) }P^{g,h}_{n-1}\left( x\right) \right|
- \left| P^{g,h}_{n}\left( x\right) -\frac{x}{h\left( n
\right) }P^{g,h}_{n-1}\left( x\right) \right| .
\end{eqnarray*}
Applying (\ref{Th1:formel}) leads to the desired result.
\end{proof}
\section{Proof of Theorem \ref{main} and Theorem \ref{variantB}}
\begin{proof}[Proof of Theorem~\ref{main}]
The proof will be by induction on $n$.
The case $n=1$ is obvious:
$$\left| P_{1}^{g,h}\left( x\right) - \frac{x+g\left( 2\right) h\left( 0\right) }{h\left( 1\right) }P_{0}^{g,h}\left( x\right) \right| =0<\varepsilon \frac{\left| x\right| }{h\left( 1\right) }\left| P_{0}^{g,h}\left( x\right) \right| $$
for
$\left| x\right| >\kappa \,\, h(0) $. Let now $n\geq 2$. Then
\begin{eqnarray*}
&&P_{n}^{g,h}\left( x\right) -g\left( 2\right) \frac{h\left( n-1\right) }{h\left( n\right) }P_{n-1}^{g,h}\left( x\right) \\
&=&\frac{x}{h\left( n\right) }\left( P_{n-1}^{g,h}\left( x\right) +\sum _{k=2}^{n-1}\left( g\left( k+1\right) -g\left( 2\right) g\left( k\right) \right) P_{n-1-k}^{g,h}\left( x\right) \right) .
\end{eqnarray*}
The basic idea for the induction step is to use the
inequality
\begin{eqnarray*}
&&\left| P_{n}^{g,h}\left( x\right) -\frac{x+g\left( 2\right) h\left( n-1\right) }{h\left( n\right) }P_{n-1}^{g,h}\left( x\right) \right| \\
&\leq &\frac{\left| x\right| }{h\left( n\right) }\sum _{k=2}^{n-1}\left| g
\left( k+1\right) -g\left( 2\right) g\left( k\right) \right| \left| P_{n-1-k}^{g,h}\left( x\right) \right| .
\end{eqnarray*}
The sum can be estimated using for $1\leq j\leq n-1$ that
\begin{eqnarray*}
&&\left| P_{j}^{g,h}\left( x\right) \right| \\
&\geq &\left| \frac{x+g\left( 2\right) h\left( j-1\right) }{h\left( j\right) }\right| \left| P_{j-1}^{g,h}\left( x\right) \right| -\left| P_{j}^{g,h}-\frac{x+g\left( 2\right) h\left( j-1\right) }{h\left( j\right) }P_{j-1}^{g,h}\left( x\right) \right| \\
&> &\left( \frac{\left| x\right| }{h\left( j\right) }-\frac{\left| g\left( 2\right) \right| h\left( j-1\right) }{h\left( j\right) }-\varepsilon \frac{\left| x\right| }{h\left( j\right) }\right) \left| P_{j-1}^{g,h}\left( x\right) \right| \\
&=&\frac{\left( 1-\varepsilon \right) \left| x\right| -\left| g\left( 2\right) \right| h\left( j-1\right) }{h\left( j\right) }\left| P_{j-1}^{g,h}\left( x\right) \right| \\
&\geq &\frac{\left( 1-\varepsilon \right) \left| x\right| -\left| g\left( 2\right) \right| h\left( j\right) }{h\left( j\right) }\left| P_{j-1}^{g,h}\left( x\right) \right|
\end{eqnarray*}
for
$\left| x\right| > \kappa \,\, h(n-1)
$.
Note that for
$\left| x\right| >\kappa \,h(n-1) $
we have
\[
\left( 1-\varepsilon \right) \left| x \right| -g\left( 2\right) h\left( j\right) >\left( \frac{1}{T_\varepsilon }+\left| g\left( 2\right) \right| \right) h\left( n-1\right) -g\left( 2\right) h\left( j\right) >0.
\]
Thus,
\[
\left| P_{j-1}^{g,h}\left( x\right) \right|
<\frac{h\left( j\right) }{\left( 1-\varepsilon \right) \left| x\right| -g\left( 2\right) h\left( j\right) }\left| P_{j}^{g,h}\left( x\right) \right|.
\]
We use this inequality and obtain
\begin{eqnarray*}
\left| P_{n-k}^{g,h}\left( x\right) \right|
&<&\left| P_{n-k+1}^{g,h}\left( x\right) \right| \frac{h\left( n-k+1\right)
}{\left( 1-\varepsilon \right) \left| x\right| -\left| g\left( 2\right) \right| h\left( n-k+1\right) }<\ldots \\
&<&\left| P_{n-1}^{g,h}\left( x\right) \right| \prod _{j=1}^{k-1}\frac{h\left( n-j\right)
}{\left( 1-\varepsilon \right) \left| x\right| -\left| g\left( 2\right) \right| h\left( n-j\right) } \\
&\leq &\left| P_{n-1}^{g,h}\left( x\right) \right| \left( \frac{h\left( n-1\right) }{\left( 1-\varepsilon \right) \left| x \right| -\left| g\left( 2\right) \right| h\left( n-1\right) }\right) ^{k-1}
\end{eqnarray*}
for
$\left| x\right| >\kappa \,\, h( n-1) \geq \kappa \,\, h(n-k)
$
for all $2 \leq k \leq n$ by assumption.
Using this, we can now estimate the sum by
\begin{eqnarray*}
&&
\sum _{k=2}^{n-1}\left| g
\left( k+1\right) -g\left( 2\right) g\left( k\right) \right| \left| P_{n-1-k}^{g,h}\left( x\right) \right| \\
&<&
\left| P_{n-1}^{g,h}\left( x\right) \right| \sum _{k=2}^{n-1}\left| g
\left( k+1\right) -g\left( 2\right) g\left( k\right) \right| \left( \frac{h\left( n-1\right)
}{\left( 1-\varepsilon \right) \left| x\right| -\left| g\left( 2\right) \right| h\left( n-1\right) }\right) ^{k}
\end{eqnarray*}
and we obtain
\begin{eqnarray*}
&&\left| P_{n}^{g,h}\left( x\right) -\frac{x+g\left( 2\right) h\left( n-1\right) }{h\left( n\right) }P_{n-1}^{g,h}\left( x\right) \right| \\
&<&
\frac{\left| x\right| }{h\left( n\right) }\left| P_{n-1}^{g,h}\left( x\right) \right| \sum _{k=2}^{n-1}\left| g\left( k+1\right) -g\left( 2\right) g
\left( k\right) \right| \left( \frac{h\left( n-1\right)
}{\left( 1-\varepsilon \right) \left| x\right| -\left| g\left( 2\right) \right| h\left( n-1\right) }\right) ^{k}.
\end{eqnarray*}
Estimating the sum using the
assumption from the theorem we obtain
\begin{eqnarray*}
&&\sum _{k=2}^{n-1}\left| g\left( k+1\right) -g\left( 2\right) g
\left( k\right) \right| \left( \frac{h\left( n-1\right)
}{\left( 1-\varepsilon \right) \left| x\right| -\left| g\left( 2\right) \right| h\left( n-1\right) }\right) ^{k} \\
&\leq &
G_2
\left( \frac{h\left( n-1\right)
}{\left( 1-\varepsilon \right) \left| x\right| -\left| g\left( 2\right) \right| h\left( n-1\right) }\right)
\leq G_2\left( T_{\varepsilon }\right) \leq
\varepsilon
\end{eqnarray*}
since
$\left| x\right| >\kappa \,\, h( n-1) =\frac{\kappa \,\,h(n-1) }{1-\varepsilon }\left( \frac{1}{T_{\varepsilon }}+\left| g\left( 2\right) \right| \right) $
which is equivalent to
$\frac{\left( 1-\varepsilon \right) \left| x\right| }{h\left( n-1\right) }-\left| g\left( 2\right) \right| >\frac{1}{T_{\varepsilon }}$
and $G_2$ is increasing on $\left[ 0,R\right) $
as $\left| g\left( k+1\right) -g\left( 2\right) g\left( k\right) \right| \geq 0$ for all $k\in \mathbb{N}$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{variantB}]
This basically follows from Theorem \ref{main} (see also the proof of Theorem \ref{versionA2}).
\end{proof}
\begin{center}
\begin{minipage}[t]{0.6\textwidth}
\[
\begin{array}{|r|r|}
\hline
n & \min \left\{ \func{Re}\left( x\right) :P_{n}^{\sigma ,\func{id}}\left( x\right) =0\right\} \\ \hline \hline
1 & 0 \\ \hline
2 & -3 \\ \hline
3 & -8 \\ \hline
4 & -14 \\ \hline
5 & -20.61187 \\ \hline
6 & -27.64001 \\ \hline
7 & -34.97153 \\ \hline
8 & -42.53511 \\ \hline
9 & -50.28267 \\ \hline
10 & -58.18014 \\ \hline
50 & -410.63656 \\ \hline
100 & -874.47135 \\ \hline
500 & -4687.67815 \\ \hline
1000 & -9501.75903 \\ \hline
\end{array}
\]
\captionsetup{margin={0cm,0cm,0cm,0cm}}
\captionof{table}{Minimal zeros of $P_{n}^{\sigma ,\func{id}}\left( x\right) $}\label{roots}
\end{minipage}
\end{center}
\begin{Acknowledgement}
To be entered later.
\end{Acknowledgement}
|
\section{Introduction}
In the process of planet formation in a protoplanetary disc, dust grows from individual grains into larger agglomerations.
If planets are to be formed by core accretion, that is through the forming of a self-gravitating solid body as opposed to a gas-dominated self-gravitating collapse, then a chain of mechanisms for growth from the $\mu$m scale to the self-gravitating km scale must be established.
Problematically, between the mm scale and km scales a number of challenges to the growth of solids are present.
As dust agglomerations grow, the speed of their collisions does too.
Given the finite and material dependent bonding forces between the dust monomers, these collisions are eventually become too energetic as the particles grow, and result in breaking the agglomerates back down to smaller fragments, preventing further growth.
If the drag timescale of dust against the gas grows towards the orbital timescale, a second barrier is encountered, because
the timescale for the gas drag on the dust to decay its orbit and drag it into the star reduces.
This introduces a problem, in that even if the dust collisions at this scale are survivable for the dust in question, the time available to have these collisions and grow into large dust is constrained.
Thus, once the dust cannot grow faster than it drifts, it cannot grow to a large enough size for the radial drift to slow down, allowing it to survive.
Furthermore, the scale of the dust collisions velocities is the same order as the inwards dust drift velocity, so this regime is also the least favourable for surviving fragmentation. For a review of these processes, see e.g. \cite{2014prpl.conf..339T}.
To solve these problems the Streaming Instability \citep[SI,][]{2005ApJ...620..459Y} has been a popular proposal.
If dust can be grown both large enough to have long stopping times, and concentrated abundantly enough, the SI can create dust concentrations able to collapse directly into km-scale self-gravitating planetesimals.
On the path to planet formation, these planetesimals can survive dust collisions, and are large enough to be relatively safe from aerodynamic drag decaying their orbits.
Other processes have been proposed for assembling planetesimals from dust.
For example, turbulent clustering of dust, leading directly to gravitational collapse \citep{2001ApJ...546..496C,2008ApJ...687.1432C,2010Icar..208..518C,2010Icar..208..505C,2016MNRAS.456.2383H,2020ApJ...892..120H}, or the concentration of self-gravitating clumps of dust in pressure traps or zonal flows \citep{
1972fpp..conf..211W,2009ApJ...697.1269J,2014ApJ...796...31B,
2018A&A...617A.117R},
vortices \citep{1995A&A...295L...1B,
2008A&A...491L..41L},
and secular gravitational instability \citep{ 1998Icar..133..298S,2011ApJ...731...99Y,2012ApJ...746...35M,2014ApJ...794...55T,2018PASJ...70....3T,2020ApJ...900..182T}
are alternate routes to planetesimal formation.
The difference in the underlying physics between these planetesimal formation scenarios leads to different requirements placed on the dust evolution processes in the molecular cloud, star formation, and protoplanetary discs feeding into them.
Thus, constraining the requirements for planetesimal formation in a given scenario, and comparing to dust evolution models is one way to make progress in understanding which mechanisms succeed and are dominant.
However, the most studied context for a dust-gas streaming scenario in a protoplanetary disc only considers a single size of dust particle, with a single value of the stopping time.
This monodisperse Streaming Instability (mSI) can be viewed as a special case of a more general and physical scenario which accounts for a wide, continuous range of dust sizes, the Polydispserse Streaming Instability (PSI).
Although a number of works have considered multiple dust sizes \citep{2010ApJ...722.1437B,2018A&A...618A..75S} and even approached the continuous limit with a series of multi-species calculations \citep{2019ApJ...878L..30K,2020arXiv200801119Z} the PSI is still largely unexplored.
This paper, the third in the series, employs the tools developed in \citet{paper2} (hereafter \citetalias{paper2}) to study the linear phase of PSI, the fundamentals of which were described in \citet{paper1} (hereafter \citetalias{paper1}).
We focus on extending the range of dust stopping times, and generalizing the dust size distributions to a class of distributions which are plausible outcomes of the dust processing occurring in protoplanetary discs.
These generalisations allow us to draw conclusions about when the linear PSI is conducive to planet formation.
The generalisation of classical mSI (monodispserse Streaming Instability) to PSI allows the consideration of realistic dust size distributions in the protoplanetary disc.
These distributions are non-trivial, as they are the results of the slow processes of dust coagulation-fragmentation and have a dependence on the dust material properties and the disc gas flow
\citep{2011A&A...525A..11B,
2015ApJ...813L..14B, 2016ApJ...818..200E, 2019ApJ...874...26S, 2020SoSyR..54..187K}.
We also consider here in a limited manner the effects of turbulence through a diffusive-viscous model for turbulence, extending the methods used by \citet{2020ApJ...895....4U} and \citet{2020ApJ...891..132C} to a polydisperse context.
\subsection{Review of previous works}
In \citetalias{paper1} we introduced the PSI, and presented a terminal velocity (TV) analysis of the instability.
Our most important finding in this model for well-coupled dust was that for wide dust distributions the instability depends on the size resonance, a resonance related to that found in the Resonant Drag Instability (RDI) theory for the mSI. In RDI theory, fastest mSI growth rates occur when the drift velocity of the dust matches the phase velocity of a gas wave. When considering a size distribution, with a size-dependent drift speed, it may be that a single size in the distribution satisfies an RDI condition, despite the average dust drift being different. We found in \citetalias{paper1} that this size resonance plays a major role in determining PSI growth rates. We finally tested the TV-PSI approximation against the full PSI equations with a direct solution method.
In \citetalias{paper2} we introduced the publically available {\tt psitools} package \citep{psitools} and its advanced numerical methods for solving the PSI linear stability problem.
The direct solver employed in \citetalias{paper1} directly discretizes the entire eigenproblem in size space and solves a matrix form for eigenvalues and eigenvectors.
This direct solver, though faster converging for fast growing modes than the multifluid discretization employed in \citet{2019ApJ...878L..30K} and \citet{2020arXiv200801119Z}, shares the difficulty in attaining results with small relative truncation error and floating point error where the PSI growth rate is small ($\lesssim 10^{-4}\Omega$, where $\Omega$ is the local Keplerian angular velocity).
The root finding algorithm introduced in \citetalias{paper2} uses a scalar formulation of the PSI dispersion relation to find roots in the complex plane.
This allows the use of high accuracy numerical quadrature schemes for integrals over dust size space yielding high accuracy roots of the dispersion relation, and hence pushing the truncation error down to the floating point error limit on the computation of the dispersion relation.
This yields dramatically more accurate growth rates while also dramatically reducing computational cost.
We exploit these tools in this paper to survey the most interesting conditions for PSI growth which could potentially result in the formation of planetesimals.
\citet{2019ApJ...878L..30K} first studied the linear evolution of PSI through multi-fluid SI calculations. We have demonstrated agreement with the results of their calculations in \citetalias{paper1} and \citetalias{paper2}.
Their work found that PSI growth is very slow when the local dust to gas mass ratio is less than unity, and did importantly anticipate that dust evolution may have an important effect on the viability of PSI as a planet formation mechanism.
Recently, \citet{2020arXiv200801119Z} presented numerical calculations of specific cases of multi-fluid SI which display a trend in agreement with our conclusions in \citetalias{paper1} about the role of a critical $\mu$ in enabling growth by the size resonance.
The difference between the size resonance and the resonant drag instability also shows why the PSI behaves in such a different way, with the transition from a monodisperse to a wide dust distribution sometimes increasing, and sometimes decreasing, the instability growth rate.
The sharp boundary of the island of fast growing modes \citet{2020arXiv200801119Z} infer in wavenumber space is rendered in high fidelity by the calculations in this series of papers.
The mapping of the complex valued dispersion relation in \citetalias{paper2} showed how this comes about: in large regions of wavenumber space growing roots flatten towards the real axis asymptotically as the dust distribution is widened.
\subsection{Dust evolution in protoplanetary discs}
Typical dust distributions in the interstellar medium can be roughly described as a power-law in dust size from $\sim0.01\ {\rm \mu m}$ to $\sim0.2\ {\rm \mu m}$ \citep{2011piim.book.....D}.
Dust grains likely grow significantly through coagulation during protostellar collapse, significantly depleting the population of the smallest grains \citep{2020arXiv200704048G}.
Once in the circumstellar disc environment, dust evolution is primarily driven by collisions. The outcome of these collisions depends on the sizes of dust involved, the structure of the aggregates, and the energy of the collision.
For the smallest grains, the dust mutual relative velocities are dominated by Brownian motion, and the gentle collisions result in sticking leading to the growth of aggregates.
Once they have grown larger, the aggregate grains enter a regime where their mutuaal velocities are driven mainly by turbulent motions of the disc gas, which results in faster collisions.
At a critical scale these velocities grow fast enough for the collisions of like-sized grains to result in fragmentation, limiting the growth of the aggregates \citep{2018haex.bookE.136A}.
In the inner parts of a disc, the gas headwind on large aggregates may alone be enough to tear parts off, limiting their growth without particle-particle interactions \citep{2020MNRAS.496.4827R,2020ApJ...898L..13G},
Dust collisions at intermediate scales and unequal-size collisions have further effects, resulting in bouncing, compaction, erosion, or other processes \citep{2010A&A...513A..56G}.
A single dust population does not spend the lifetime of the disc merely interacting locally. Dust of differing sizes and physical properties drifts radially and are lofted vertically to different extents.
Structures in the disc, such as azimuthal pressure bumps, gap and dead zone edges, and vortices likely play a role in the evolution of the dust population, retaining dust which would otherwise drift radially into the star.
This complicated landscape of processes evolves the dust distribution present at any given location of the disc, in a way which is not yet well understood, but must result in the fuel for planet formation.
Understanding the PSI and how the growth rates correspond to dust distributions will reveal which dust evolution scenarios are conducive to planet formation.
Observation of discs can detect the scattered light and/or thermal emission from dust from $0.01\ {\rm\mu m}$ to cm sizes.
The observational signature of dust is a complicated convolution of the physical condition and the radiative transfer properties (such as the opacity, albedo, and polarization) of the entire population of dust.
These radiative transfer properties can have particularly detailed dependence on the physical morphology and mineralogical properties of the dust grains.
This difficulty currently limits what is observationally known about the dust size distribution in discs.
Beyond thermal continuum emission, scattered light observations and details of the system's spectral energy distributions in the optical and near-infrared indicate the presence of dust particles with sizes $\gtrsim 10\ {\rm \mu m}$ high in the disc atmosphere
\citep{2020arXiv200105007A}.
Protoplanetary disc observations show large populations of grains even in old discs, where simple models would suggest the dust should have been depleted by drifting onto the star,
This seeming contradiction indicates a possible role of azimuthal pressure bumps, trapping and slowing dust on its radial journey
\citep{2018ApJ...869L..41A,2018ApJ...869L..46D}.
\subsection{Where might mSI form planetesimals?}
The proposed role of SI-class instabilities in planet formation of producing planetesimals relies on the instability gathering dust overdensities in the disc which exceed the local Roche density
at large scales, and turbulent diffusion at small scales, allowing them to collapse into self-gravitating planetesimals \citep{2007Natur.448.1022J,
2020ApJ...895...91G,2020arXiv200710696K}.
We study in this work only linear instability growth, but the full scenario of planetesimal formation follows from growth to nonlinear amplitudes leading to gravitational collapse of these dust clouds.
To aim our investigation at a useful parameter regime, we have examined existing evidence about the regime where SI driven planetesimal formation is likely to occur, and judge form the current literature what parameters may be optimal.
Simulations producing streaming instability and gravitational collapse for monodispserse particles typically employ particles with stopping times $\sim 1 \ \Omega^{-1}$. Notable examples are
\citet{2007Natur.448.1022J} which used particles with ${\tau_{\rm s}}=0.25$--$1\ \Omega^{-1}$ for the first demonstration of SI driven gravitational fragmentation, and
\citet{2019NatAs...3..808N} which modelled the formation of trans-Neptunian objects employing particles with ${\tau_{\rm s}}=0.3$--$3\ \Omega^{-1}$.
However, due to the difficulty of producing such large particles in coagulation-fragmentation models of dust, another research theme has been an effort to probe the lower limit of particle stopping time required.
This minimum dust stopping time criteria for mSI growth of dust clumps in full simulations is a function of the minimum dust-gas ratio \citep{2015A&A...579A..43C}, but
\citet{2017A&A...606A..80Y} suggested in any case the lower limit of mSI viability for forming dust-rich filaments in stratified simulations is ${\tau_{\rm s}}=10^{-2}$--$10^{-3}\ \Omega^{-1}$.
These latter studies probed stratified nonlinear streaming instability, but did not include gravitational collapse.
In three dimensional stratified simulations, which are able to include particle settling and lofting,
the fundamental parameter is not a local gas/dust volume density ratio, but instead the surface density ratio, and the local dust to gas ratio is a result largely of dust settling to the midplane.
The results of \citet{2017A&A...606A..80Y} show particles with ${\tau_{\rm s}}=10^{-3}\ \Omega^{-1}$ settling to a midplane layer with dust to gas ratio order unity before forming SI filaments.
The collapse to planetesimals successfully from mSI clumps with ${\tau_{\rm s}}=6\times10^{-3}\ \Omega^{-1}$ was shown in \citet{2017ApJ...847L..12S}, and
for a range of disc gas radial pressure gradients with ${\tau_{\rm s}}=5\times10^{-2}\ \Omega^{-1}$ \citep{2019ApJ...883..192A}.
\citet{2020arXiv200801727C} found a tentative requirement of particles with at least ${\tau_{\rm s}}=0.2\ \Omega^{-1}$ for mSI driven fragmentation to occur in 3D simulations including the effect of a background long-lived azimuthal pressure bump, which enhances the concentration of particles.
However, smaller ${\tau_{\rm s}}\simeq 0.02\ \Omega^{-1}$ particles did not drive the formation of fragments, and the dust to gas density ratio only rarely exceeded unity in the simulation volume.\footnote{In \citet{2020arXiv200801727C} the gas density enhancement in the pressure bump does cause the ${\tau_{\rm s}}$ of the particle to vary by location in the domain, but by less than the rough order of magnitude in this discussion.}
Separately, a meta-analysis of previous three dimensional nonlinear simulation results by \citet{2020ApJ...895....4U} suggested that local volumetric gas to dust mass ratios in the midplane exceeding unity, and dust larger than ${\tau_{\rm s}} \gtrsim 0.1\ \Omega^{-1}$ are required to produce mSI driven fragmentation.
\citet{2020ApJ...895....4U} suggests that this ${\tau_{\rm s}} \gtrsim 0.1\ \Omega^{-1}$ criteria comes from the effect of turbulence in the disc gas, beyond the midplane turbulence created purely by a single small particle size streaming through the gas.
Nonlinear SI drives turbulence through gas dust interaction, and ought to occur in a
background with turbulence driven by other instabilities at some level.
\citet{2020ApJ...895....4U} include a model for turbulence in their linear calculations and find only a very small region of viable parameter space for mSI.
Attempting to include the effects of turbulence in a similar formalism,
\citet{2020ApJ...891..132C} suggest a minimum dust stopping time as a function of Shakura-Sunyaev turbulence parameter $\alpha$ of ${\tau_{\rm s}} \gtrsim \alpha^{2/3}\ \Omega^{-1}$ is required for appreciable growth of the linear mSI instability.
For $\alpha < 10^{-2}$ this is a looser stopping time criteria than the one for the formation of fragments from the simulations of \citet{2020arXiv200801727C}.
In summary, the current understanding of nonlinear mSI suggests particle stopping times on the order of ${\tau_{\rm s}} \sim 0.1\ \Omega^{-1}$ are optimal for the production of planetesimals, although under certain assumptions much larger and smaller stopping times may suffice.
This value suggests a centre for the parameter range for the peak particle sizes we will examine in linear PSI and figure-of-merit on which to evaluate the growth rate of PSI against mSI in the context of planetesimal formation.
\subsection{This paper}
In this paper we present circumstances where PSI shows fast growth, and attempt to find regions where PSI is as robust as mSI.
The possible range of parameters is vast, so we present specific cases which, at the current time, seem most relevant and interesting.
In addition, the depth of existing work on streaming and resonant drag instabilities involving the coupled interaction of dust and gas raises myriad questions of the connections between PSI and its generalisations and these previous works.
There is no way for this survey to be exhaustive, so we have made the underlying code publicly available for the community to analyse other scenarios as they arise.
In this paper, Section~\ref{sec:methods} describes the methods used for our calculations,
Section~\ref{sec:dustparam} presents the models for the dust distributions used and the logic behind them,
Section~\ref{sec:laminar} contains results from calculations of the PSI in a laminar background,
Section~\ref{sec:turbulence} presents results for the PSI including a model for the diffusive effects of turbulence,
Section~\ref{sec:discussion} discusses the results in the context of planet formation,
and finally conclusions are drawn in Section~\ref{sec:conclusions}.
\section{Methods}
\label{sec:methods}
In this work we employ the methods of the {\tt psitools} package described in \citetalias{paper2} available through the Zenodo.org archive service \citep{psitools}.
In brief, three tools are used here.
First, a direct discretization of the system of equations for the PSI linear stability problem produces approximate eigenvalues and the associated eigenvectors ({\tt psitools.direct}).
Second, a reduction of the PSI linear stability problem to a single dispersion relation equation allows the application of a complex valued root finder, producing high accuracy eigenvalues ({\tt psitools.psi\_mode}).
Third, a parallel implementation of a grid-refinement method for mapping wavenumber space with the root finder produces maps of the fastest growing PSI mode ({\tt psitools.psi\_grid\_refine}).
Detailed discussion of these methods can be found in \citetalias{paper2}, where also the governing equations in the shearing box are given.
We adopt only $\eta=0.05\ r_0\Omega^2$, where $r_0$ is the fiducial orbital radius of the shearing box, as the value of the radial pressure gradient in the gas disc in this work, although values vary locally around azimuthal pressure bumps, and globally, in a protoplanetary disc. It will also vary in other PSI contexts such as a protolunar disc \citep{2020LPI....51.2976A}.
Smaller pressure gradient values are known to facilitate particle concentration in mSI \citep{2010ApJ...722L.220B}, so situations where these occur may produce differing results.
We calculate growth rate maps covering $K_{x,z}\in [10^{-1},10^3]$ where the wavenumbers are nondimensionalized as ${\bf K} = {\bf k}\eta/\Omega^2$ \citepalias{paper2}. This corresponds to wavelengths of $1.4$ to $1.4\times10^{-4}$ pressure scale heights horizontally and vertically in an MMSN at $1~\mathrm{au}$.
At the short wavelength, high wavenumber end of this scale, a minimal turbulent viscous effect will damp growth \citep[][also see Section~\ref{sec:turbulence}]{2020ApJ...895....4U,2020ApJ...891..132C}.
At low wavenumbers and long wavelengths, the local shearing sheet approximation employed in formulating the linear problem \citepalias{paper1, paper2} has inherent limitations on applicability at scales larger than a pressure scale height when simply considering hydrodynamics \citep{2004A&A...427..855U,2015ApJ...811..121M}.
When in addition considering dust-gas dynamics, the assumption of a constant $\eta$ \citepalias{paper1} may break down at radial scales larger then a pressure scale height, particularly in the presence of an azimuthal pressure bump.
However, for the most part, the fastest growth we find in each map when considering laminar flow is in the high wavenumber (short wavelength) side of the wavenumber space considered.
Finally, we only solve for eigenvalues with growth rates above $2\times10^{-7}\Omega$, due the the technical and computational difficulty of finding roots with these small, physically less important growth rates.
Further discussion of the structure of the PSI dispersion relation and the associated difficulties can be found in \citetalias{paper2}.
\section{Relevant dust parameters}
\label{sec:dustparam}
Notwithstanding the differences between mSI and PSI, we will focus the survey in this paper on conditions approaching those where mSI has previously been proposed to function well, as the dust and gas disc processes leading the system to the point of instability should be the same.
Indeed, the concept behind choosing the particle size in mSI studies has been that this single particle size is representative of, or the dominant one in, the entire dust distribution.
Hence, the stopping time and dust to gas mass ratio regime centres here on ${\tau_{\rm s}} = 10^{-1}\ \Omega^{-1}$ and $\mu \gtrsim 1$.
As established in \citet{2019ApJ...878L..30K}, \citetalias{paper1}, and \citet{2020arXiv200801119Z}, the PSI is likely most
viable as a planetesimal formation mechanism at dust to gas mass ratios $\mu>1$.
Indeed mSI is also most viable as a dense clump-forming mechanism in this high-$\mu$ regime
\citep{2009ApJ...704L..75J,2020arXiv200110000G,2020MNRAS.tmp.2397S,2020arXiv200801727C}.
In this work, we will focus on the linear evolution of PSI in terms of a fixed $\mu$, even when considering the diffusive and viscous effects of turbulence.
Particularly in the polydisperse case, no consistent theory exists
for particle settling outside of the trace dust density ($\mu\ll 1$) regime, so self-consistent study of this regime requires either enhanced theories or nonlinear and vertically extended simulations.
It should however be anticipated that the effect of turbulence, generated by other magnetohydrodynamic or hydrodynamic instabilities, or the PSI itself can be expected to be a crucial mechanism for setting the criteria which determine when planet formation follows from the PSI.
While in the PSI, the convenient variable for parameterizing the size of dust is the stopping time ${\tau_{\rm s}}$, dust coagulation models usually work in terms of dust size $a$ directly. In the Epstein drag regime, ${\tau_{\rm s}} \propto a$ \citepalias[][eq.~35]{paper2}.
This makes it possible to simply parameterize the dust distributions in terms on ${\tau_{\rm s}}$ when convenient, and perform a trivial change of variables to make the gas-dust drag integral in the gas momentum equation \citepalias[][eq.~38]{paper2} to be over the range in stopping time $taus$ instead of particle size $a$. In this section we describe the dust distributions in terms of the size $a$, but in describing our results for PSI growth we will use the stopping time ${\tau_{\rm s}}$.
\subsection{Models for evolved dust size distributions}
\label{sec:dustdists}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{distributions_example.pdf}
\caption{Examples of the MRN, PB, and PBT dust distributions in terms of the size density $\sigma$ as a function of particle radius $a$, for the same total dust volume mass density $10^{-9}\ {\rm[g\ cm^{-3}]}$, and peak particle size $a_{\rm P}=1\ \mathrm{m}$. For the PBT example $a_{\rm BT}=10^{-4}a_{\rm P}$.}
\label{fig:distributions}
\end{figure}
In this paper we consider three dust size distributions as examples of typical results from dust evolution processes.
Solid dust should exist in discs with a wide range of sizes.
The minimum silicate particle size in the interstellar medium is plausibly a single molecule.
For their dust coagulation model
\citet{2011A&A...525A..11B} specify a monomer size of $0.025\ \mu{\rm m}$,
and the inclusion of ISM silicate grains down to at least $0.01\ \mu{\rm m}$ is required to reproduce the optical properties of ISM dust \citep{1984ApJ...285...89D,2011piim.book.....D}.
Collisional dynamics, with dust collision velocities increasing as the particle size grows, may limit the growth of particles when this velocity scale exceeds the fragmentation velocity of aggregates.
If this fragmentation does not limit the size of dust aggregates, particles with stopping times on the order of an orbital period drift inwards through the disc rapidly compared to dust growth timescales \citep{1977MNRAS.180...57W,2012A&A...539A.148B}.
Thus, the upper size of the range under consideration for PSI ought to be on the order of unity.
For these reasons, in this survey we choose a fiducial range of dust stopping times as ${\tau_{\rm s}} \in [10^{-8}, 1]\ \Omega^{-1}$ as being reasonably representative for the disc, and vary the upper size limit.
\subsubsection{MRN distribution}
The particle size distribution in the interstellar medium is commonly modelled in terms of the size density $\sigma$ \citepalias{paper2} as a function of particle radius $a$ as
\begin{align}
\sigma_{\rm MRN}(a) \propto a^{3+\beta}, \ \ \beta=-3.5\ ,
\end{align}
referred to as the MRN distribution \citep{1977ApJ...217..425M,1984ApJ...285...89D}.
Although used as a fit to observations of interstellar dust, this distribution is also mathematically the result of a collisional dust evolution process with pure fragmentation or pure coagulation, and is also a reasonable description of the dust size distribution in a debris disc, and the asteroid belt for this reason \citep{1969JGR....74.2531D,1994Icar..107..117W, 1996Icar..123..450T}.
Hence, this is the basic dust size density distribution that could be expected in the dust inherited form the interstellar medium by a protoplanetary disc.
It has previously been used to study PSI in \citetalias{paper1}, \citetalias{paper2}, and in the works by \citet{2019ApJ...878L..30K} and \citet{2020arXiv200801119Z}.
However, we should expect gas damping of dust velocities in a protoplanetary disc to modify the dust evolution from pure fragmentation.
\subsubsection{Powerlaw-Bump (PB) distribution}
In protoplanetary discs, planet formation can be expected to proceed from dust which has evolved in this context from that found in the interstellar medium.
In the disc, dust relative velocities are smaller then in the ISM, and temperature-density conditions allow for the condensation of further chemical species.
Thus, dust evolution through coagulation-fragmentation allows the formation of larger dust grains.
Additionally, aerodynamic effects can size-sort dust, resulting in altered dust distributions independent of the coagulation-fragmentation evolution.
As this processing has a very large number of parameters, we do not attempt here to produce a full range of possible dust distributions for all regions of protoplanetary discs.
Instead, we consider a simple model derived from the work of \citet{2011A&A...525A..11B} and \citet{2015ApJ...813L..14B} which describes the most apparently common outcomes of dust processing in discs from a fragmentation-coagulation model. We caution that here we only attempt to consider the most common and widely used dust evolution, and in general other special cases not anticipated here might become important.
One of the motivations in making the {\tt psitools} package public is to enable the community to analyse linear PSI with specific dust distributions as they arise in the future.
This recipe can be written as of a power-law in particle number density as a function of radius
\begin{align}
f(a) = a^{\beta} \label{eq:fplain}\, ,
\end{align}
and a Gaussian bump,
\begin{align}
b(a) = 2.0 f(a_{\rm L}) \exp\left( \frac{-\left(a-a_{\rm P}\right)^2}{w}\right)\, ,
\end{align}
with the width
\begin{align}
w = \max\left(
\frac{\min(|a_{\rm R}-a_{\rm P}|,|a_{\rm L}-a_{\rm P}|)}{\sqrt{\ln(2)}}, (0.1)a_{\rm P}
\right)\, ,
\end{align}
and then
\begin{align}
F(a) = \left\{\begin{array}{cll}
f(a) & \text{if} & a \leq a_{\rm L}\\
\max\left(f(a),b(a)\right) & \text{if} & a_{\rm L} \leq a \leq a_{\rm P} \\
b(a) & \text{if} & a_{\rm P} \leq a \leq a_{\rm R}\\
0 &\text{else.}&
\end{array}
\right.
\end{align}
Which then, normalized, gives the dust size density as
\begin{align}
\sigma_{\rm PB}(a) = a^3 F(a)\left/\int_{a_{\rm min}}^{a_{\rm max}} w^3 F(w)\, {\rm d}w\right. . \label{eq:PB}
\end{align}
We refer to this distribution as PB (Powerlaw-Bump).
Physically, \citet{2011A&A...525A..11B} prescribe the Gaussian bump at the top end of the distribution as the result of a local change in size space to the coagulation-fragmentation fluxes due to the prevalence of cratering collisions between small and large dust grains.
For simplicity, we consider here specific values of $a_{\rm L}$ and $a_{\rm R}$ which the \citet{2011A&A...525A..11B} recipe generates for disc conditions typical of a minimum mass solar nebula at 1--10~au (Appendix~\ref{app:bump}):
\begin{align}
a_{\rm L}&= \frac{2}{3} a_{\rm P}\, , \label{eq:aL} \\
a_{\rm R}&= 1.56 a_{\rm P}\, . \label{eq:aR}
\end{align}
In addition we adopt $\beta=-3.5$ for the power-law component as this fiducial power-law index is close to the simplified value adopted in \citet{2015ApJ...813L..14B}, and facilitates the comparison of this density distribution and the MRN distribution.
\subsubsection{Powerlaw-Bump-Tail (PBT) distribution}
Below the critical dust size where Brownian motion dominates over turbulence as a driver of dust collisions, the \citet{2011A&A...525A..11B} model includes a steeper power-law regime of the dust distribution.
To include this steep tail between the dust monomer size $a_0$ and the Brownian motion-turbulence transitions scale $a_{\rm BT}$ we follow the form from \citet{2015ApJ...813L..14B}
and substitute $f(a)$ in the PB distribution with $f_{\rm tail}(a)$ defined as:
\begin{align}
f_{\rm tail}(a) =
\left\{\begin{array}{cll}
a_{\rm BT}^{\beta} \left(\frac{a}{a_{\rm BT}} \right)^{1/2} &\text{if} & a<a_{\rm BT} \\
a^{\beta} &\text{else.} & \
\end{array}
\right. \label{eq:ftail}
\end{align}
We refer to the dust size density distribution Equation~(\ref{eq:PB}) with this tail form Equation~(\ref{eq:ftail}) substituted for Equation~(\ref{eq:fplain}) as the Powerlaw-Bump-Tail (PBT) distribution.
Examples of the three distributions used in this work are shown in Figure~\ref{fig:distributions}. Note all the distributions are what is commonly referred to as `top-heavy' in that by particle mass, the total dust density is dominated by the largest particles. In Figure~\ref{fig:distributions} this is not visually obvious as the size density $\sigma(a)$ is parameterized by particle radius $a$ not particle mass.
\subsection{Model for Gas Turbulence}
\label{sec:turbmodel}
To explore the possible impact of turbulent mixing on the linear growth of PSI, we use a turbulence model to parametrise the effect of turbulent mixing on the gas and dust fractions. Turbulence models usually model subgrid turbulence as a viscous term in the fluid, as turbulence drives momentum diffusion. Analogously, turbulent eddies can move around dust particles suspended in the fluid, driving dust diffusion. Gas turbulence and the consequent dust diffusion can be described consistently using a viscous turbulence model for the gas and a mass diffusion for the dust.
We follow the formulation of \citet{2020ApJ...891..132C}, which uses $\alpha$-viscosity as the viscous term in the gas momentum equation, and a corresponding diffusive term in the dust continuity equation, extending it to our polydisperse dust equations. These terms are an extra turbulent diffusion of momentum, adding to the gas momentum equation \citepalias{paper2}
\begin{align}
\partial_t{\bf{v}_{\rm g}} + ({\bf{v}_{\rm g}}\cdot\nabla){\bf{v}_{\rm g}} =& -\frac{\nabla P}{{\rho_{\rm g}}} + \mathbf{F_{\mathrm{drag}}} +\frac{1}{{\rho_{\rm g}}}\nabla \cdot \mathsf{T}_{\mathrm{visc}}\,
,\label{eq:govsecondturb}
\end{align}
where the final term is the added viscous diffusion term with stress tensor
\begin{align}
\mathsf{T}_{\mathrm{visc}}={\rho_{\rm g}} \nu \left[\nabla {\bf{v}_{\rm g}} + \nabla {\bf{v}_{\rm g}} ^{\dagger} - \frac{2}{3}\mathbb{I}\nabla\cdot{\bf{v}_{\rm g}}\right].
\end{align}
Here $\mathbb{I}$ is the identity tensor.
The turbulent viscosity $\nu$ is then defined following \citet{1973A&A....24..337S}
\begin{align}
\nu=\alpha c^2 \Omega\,,
\end{align}
where $\alpha$ is the dimensionless parameter describing viscosity.
A corresponding turbulent diffusion term is also added to the dust continuity equation \citepalias{paper2}
\begin{align}
\partial_t\sigma + \nabla\cdot(\sigma {\bf{u}} ) =& \nabla\cdot\left[D({\tau_{\rm s}}){\rho_{\rm g}}\nabla\mu\right]\, ,\label{eq:govthirdturb}
\end{align}
where the final term adds dust mass diffusion, with $ D({\tau_{\rm s}})=\delta({\tau_{\rm s}}) c^2 \Omega $ a diffusion coefficient depending on ${\tau_{\rm s}}$. The term $\delta$ relates the turbulent parameter $\alpha$ to the rate of dust diffusion and the dust stopping time ${\tau_{\rm s}}$ \citep{2007Icar..192..588Y},
\begin{align}
\label{eqn:dust-turbulence-coupling}
\delta=\frac{1+{\tau_{\rm s}}\Omega+4\left({\tau_{\rm s}}\Omega\right)^2}{\left(1+({\tau_{\rm s}}\Omega)^2\right)^2}\alpha,
\end{align}
encoding how larger dust particles are less subject to turbulent diffusion than smaller well coupled particles. The dust diffusion coefficient $D\left({\tau_{\rm s}}\right)$ is therefore a function of the grain stopping time ${\tau_{\rm s}}$. Small grains have ${\tau_{\rm s}}\ll 1$, making $\delta({\tau_{\rm s}})\approxeq\alpha$. However, as grains grow, $\delta({\tau_{\rm s}})$ gradually goes to $0$ as the larger grains are less subject to gas driven stirring due to their larger mass to cross-section area ratio.
This model represents only the averaged local effects of turbulence, and does not capture turbulent clumping and other turbulent fluctuations. The model also does not consider the effect of dust on the gas turbulence.
It also assumes only isotropic turbulent diffusion and viscosity effects, whereas at least one possible driver of turbulence, the Vertical Shear Instability \citep{2013MNRAS.435.2610N} is strongly anisotropic, with much greater vertical mixing than horizontal \citep{2017A&A...599L...6S}.
\citet{2020ApJ...891..132C} and \citet{2020ApJ...895....4U} have, in the mSI case, directly connected the strength of turbulence in this model to a dust scale height, and used this to map a dust column density to a midplane dust to gas ratio $\mu$.
However, we do not attempt to make such an identification in this work.
The analytical model used for dust lofting in \citet{2020ApJ...891..132C} and \citet{2020ApJ...895....4U} assumes a trace dust density $\mu\ll 1$, whereas planet formation driven by PSI is most viable at $\mu \gtrsim 1$.
The results we show do not depend on a dust lofting model, as it is unclear whether such models are applicable in the regimes studied.
Instead, in this work we specify the dust to gas mass density ratio and turbulence
strength independently, not invoking a model for the vertical
thickness of the dust distribution.
We similarly do not include the slow viscously driven radial
motion of gas in the midplane which should arise as a consequence of momentum diffusion as this brings complex dependence on disc model and location parameters \citep{1984SvA....28...50U,2017ApJ...837..101P}.
Self-consistent models for dust lofting, including polydisperse dust and valid at $\mu\sim 1$, are thus a topic of great future interest, as the column density of dust is a much more natural input parameter to the PSI initial conditions in a planetesimal forming context than the volume density of dust.
\section{PSI growth rates and eigenmode structures - Laminar}
\label{sec:laminar}
\begin{figure*}
\includegraphics[width=0.9\textwidth]{comparison_MRN_PB_PBT_Mono_mu0p5.pdf}
\caption{Growth rates (in units of $\Omega$) for $\mu=0.5$ truncated at $2\times10^{-7}\ \Omega$ with MRN, PB, PBT, and monodisperse dust distributions. {\sl Grey dashed contour:} Longest growth timescale exceeding dust drift time over a length $0.05r_0$.}
\label{fig:compmu0p5}
\end{figure*}
\begin{figure*}
\includegraphics[width=0.9\textwidth]{comparison_MRN_PB_PBT_Mono_mu3.pdf}
\caption{PSI growth rates (in units of $\Omega$) for $\mu=3$ truncated at $2\times10^{-7}\ \Omega$ with MRN, PB, PBT, and monodisperse dust distributions. {\sl Grey dashed contour:} Longest growth timescale exceeding dust drift time over a length $0.05r_0$.}
\label{fig:compmu3}
\end{figure*}
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{paper3_mu3_MRN_eigenfuncs.pdf}
\caption{Dust components of eigenfunction for the fastest growing mode with $\mu=3$ and ${\tau_{\rm s,peak}}=10^{-2}\ \Omega^{-1}$. {\sl Left:} $K_x=70$ $K_z=1000$, size resonance present. {\sl Right:} $K_x=5$ $K_z=1000$, no size resonance present. {\sl Vertical dashed lines:} Position where the mode phase velocity matches dust radial drift velocity, giving rise to the size resonance.
Each panel for a dust component includes dotted lines for the corresponding scalar component of the gas perturbation from the eigenvector.
Eigenfunctions are normalized to set $\Im({\hat{v}_{{\rm g}x}})=0$ and $|\max({\hat{\sigma}})|=1$. }
\label{fig:mu3eignefuncs}
\end{center}
\end{figure}
\begin{figure*}
\includegraphics[width=0.9\textwidth]{comparison_MRN_PB_PBT_Mono_mu10.pdf}
\caption{PSI growth rates (in units of $\Omega$) for $\mu=10$, truncated at $2\times10^{-7}\ \Omega$ with MRN, PB, PBT, and monodisperse dust distributions. {\sl Grey dashed contour:} Longest growth timescale exceeding dust drift time over a radial distance $0.05r_0$.}
\label{fig:compmu10}
\end{figure*}
\begin{figure}
\begin{center}
\includegraphics[width=0.9\columnwidth]{paper3_eigenfuncs_MRN_mu10_taus1.pdf}
\caption{Dust components of eigenfunction for the fastest growing mode with $\mu=10$, ${\tau_{\rm s,peak}}=10^{0}\ \Omega^{-1}$, $K_x=10$, and $K_z=10$. {\sl Vertical dashed lines:} Position where the mode phase velocity matches dust radial drift velocity, giving rise to the size resonance.
Each panel for a dust component includes dotted lines for the corresponding scalar component of the gas perturbation from the eigenvector.
Eigenfunctions are normalized to set $\Im({\hat{v}_{{\rm g}x}})=0$ and $|\max({\hat{\sigma}})|=1$. }
\label{fig:MRNmu10taus1eigenfuncs}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=1.0\columnwidth]{mrn_vs_pb_eigenfunctions.pdf}
\caption{Dust components of eigenfunction for the fastest growing modes with ${\tau_{\rm s,peak}}=10^{-1}\ \Omega^{-1}$, $\mu=10$. {\sl Left:} MRN dust distribution {\sl Right:} PB dust distribution. {\sl Dashed lines:} Position where the mode phase speed matches $u_{{\rm d},x}-\bar{u}_{{\rm d},x}$, giving rise to the size resonance, in the MRN case
${\tau_{\rm s}} = 1.2\times10^{-2}\ \Omega^{-1}$ and PB case
${\tau_{\rm s}} = 6.0\times10^{-2}\ \Omega^{-1}$.
Eigenfunctions are normalized to set $\Im({\hat{v}_{{\rm g}x}})=0$ and $|\max({\hat{\sigma}})|=1$.}
\label{fig:mrn_vs_pb_eignenfunctions}
\end{center}
\end{figure}
\begin{figure*}
\includegraphics[width=0.9\textwidth]{maximize_alpha.pdf}
\caption{Maximal growth rates (in units of $\Omega^{-1}$) for varying $\alpha$ with MRN, PB, and monodisperse dust distributions at dust to gas ratios $\mu\in {0.5, 3, 10 }$.}
\label{fig:maximize_alpha}
\end{figure*}
\begin{figure*}
\includegraphics[width=\textwidth]{comparison_alpha_MRN_PB_Mono_mu3.pdf}
\caption{Growth rates (in units of $\Omega$) for $\mu=3$ truncated at $2\times10^{-7}\ \Omega$ varying viscous $\alpha$ for MRN, PB, and monodisperse dust distributions with $\mu=3$ and ${\tau_{\rm s,peak}}=10^{-1}\ \Omega^{-1}$, ${\tau_{\rm s,min}}=10^{-8}\ \Omega^{-1}$ (${\tau_{\rm s}}={\tau_{\rm s,peak}}$ for monodisperse).}
\label{fig:alpha_growth_map}
\end{figure*}
The first set of growth rate results we present is a survey of the instability growth rate mapped over wavenumber, for varying dust size density distribution, varying dust to gas mass ratios $\mu$, and varying peak particle size stopping times ${\tau_{\rm s,peak}}$.
Growth rate maps computed with the grid refinement algorithm with 4 runs per point up to a $241^2$ grid \citepalias{paper2} for $\mu=0.5$ are shown in Figure~\ref{fig:compmu0p5}, for $\mu=3.0$ in Figure~\ref{fig:compmu3}, and $\mu=10$ in Figure~\ref{fig:compmu10}.
In each figure, results for the MRN, PB, PBT, and monodisperse dust distributions are shown.
Each column of the figures corresponds to a value for the peak of the size-density distribution ${\tau_{\rm s,peak}}$, which is not in general the peak of the particle mass distribution.
For the MRN distribution ${\tau_{\rm s,peak}}$ is the maximum dust size (and stopping time), for the PB and PBT distributions it is the peak of the Gaussian bump in these functions, and for the monodisperse dust distribution it is the single dust stopping time present. The three columns show results for ${\tau_{\rm s,peak}}$ of $10^{-2}\ \Omega^{-1}$, $10^{-1}\ \Omega^{-1}$, and $1\ \Omega^{-1}$. The MRN, PB, and PBT distributions have fixed minimum ${\tau_{\rm s}}$ of $10^{-8}\ \Omega^{-1}$ and for the PBT distribution the transition between power-law regimes is fixed at ${\tau_{\rm s}}_{\rm, BT}=10^{-4}\ \Omega^{-1}$.
In Figures~\ref{fig:compmu0p5},~\ref{fig:compmu3}, and~\ref{fig:compmu10} we include as a figure of merit for the growth rates a contour drawn at the exponential growth time which equals time for the fastest drifting dust to radially draft a distance of $0.05r_0$, which is on the order of the gas pressure scale height.
This length is also on the order of the largest spatial scales that the local shearing-box approximation used to derive the PSI dispersion relation (\citetalias{paper2}) can describe.
In Figure~\ref{fig:compmu0p5} the $\mu=0.5$, ${\tau_{\rm s}}=10^0\Omega^{-1}$ cases carry the first surprise.
For the pure power-law MRN distribution, no growth above the $2\times10^{-7}\Omega$ threshold is found in the wavenumber space.
However, when the bump is added at the largest dust sizes in the PB distribution, a feature resembling the growth pattern in the monodisperse case both in wavenumber space and growth rates appears.
This ridge in the PB and PBT cases does also have the important feature of a sharp cut-off on the low-$K_x$ side, in contrast to the monodisperse case.
At the smallest $K_x$ in PB and PBT a second island of more slowly growing modes appears.
This is similar to the behaviour seen with MRN dust distributions at $\mu>1$, and more analysis of these modes will be presented when discussing those cases.
For the $\mu=0.5$, ${\tau_{\rm s}}=10^{-1}\Omega^{-1}$ cases in Figure~\ref{fig:compmu0p5}, the suggestion of the size resonance is seen in the fastest growing island of modes, but the peak growth rates for all the polydisperse cases are substantially below the monodisperse case.
The left side of the fastest growing island in the MRN distribution island lies at higher $K_x$ then the ridge of fast growth in the monodisperse case, similar to what was demonstrated in the terminal velocity limit in \citetalias{paper1}.
For the $\mu=0.5$, ${\tau_{\rm s}}=10^{-2}\Omega^{-1}$ cases in Figure~\ref{fig:compmu0p5} there is also a surprising contrast between the behaviour for different dust distributions. For these parameters, the maximal growth of PSI is faster than the equivalent mSI case. Interestingly, this is contrary to the assertion made in \citet{2019ApJ...878L..30K} that the MRN-distributed PSI exhibits maximal growth rates smaller than the corresponding mSI, and agrees with the conclusions of \citet{2020arXiv200801119Z} that such a phenomenon may occur.
This parameter set lies also within the short stopping time regime addressed with the terminal velocity approximation in \citetalias{paper1} and the results are similar. Note though that in \citetalias{paper1} different wavenumber cutoffs were used, so that the island of slow growth is not visible in Figure 4 of \citetalias{paper1}.
In the MRN case the ridge of fast growth is roughly at higher $K_x$ than in the monodisperse case.
Curiously here, the dust abundance bump at largest dust sizes in the PB and PBT distribution inhibits the growth of PSI, despite this modification making the distribution more central and like the monodisperse case.
Instead of producing the faster size resonance growth of the MRN case, or the secular mSI growth of the monodisperse case, a less optimal middle ground is found.
Most of the growth maps in Figures~\ref{fig:compmu0p5}, \ref{fig:compmu3} and~\ref{fig:compmu10} contain two regions of growing modes.
In \citetalias{paper1} is was shown how the size resonance in the terminal velocity limit predicts the onset of growth in a region identified by a cut-off at low $K_x$.
However, a second island of appreciable instability typically appears with a low $K_z$ cutoff, and typically filling all $K_x$ below a cutoff, albeit in some cases a slow decent below the threshold is evident (e.g.~$\mu=0.5$, ${\tau_{\rm s,peak}}=10^{-2}\Omega^{-1}$.
For example, consider the panel for an MRN distribution, $\mu=3$, and ${\tau_{\rm s,peak}}=10^{-1}\ \Omega^{-1}$ in
Figure~\ref{fig:compmu3}.
Eigenfunctions for two fastest growing modes, one from the size density resonance strip, and one from the low-$K_x$ island are shown in Figure~\ref{fig:mu3eignefuncs}.
The eigenfunctions for $K_x=70$, $K_z=1000$ illustrate the role of the size resonance, with the dust size density having a strong response where the equilibrium radial dust drift velocity matches the radial phase velocity of the mode.
However, in the low-$K_x$ island, no size resonance effect is present, with the dust radial drift velocity and mode radial phase velocity never coinciding. Instead, the strongest response from the dust is for the maximum dust size.
This difference between the two sets of modes, one size-resonant and the other being non size-resonant, will appear again in the examination of the effect of turbulent diffusion on the PSI.
In Figure~\ref{fig:compmu3} the $\mu=3$ runs the PSI growth rates are generally faster than those with $\mu=0.5$.
Particularly for the MRN cases, this reflects the faster growth for the high-$\mu$ PSI, with the ${\tau_{\rm s,peak}}=10^{-2}\Omega^{-1}$ MRN case particularly showing how the size-resonance growth in the terminal velocity limit is shifted to higher $K_x$ than the monodisperse case.
This is also apparent in the ${\tau_{\rm s,peak}}=10^{-1}\Omega^{-1}$ MRN case.
However, in the ${\tau_{\rm s,peak}}=10^{0}\Omega^{-1}$ MRN case this ridge of faster size-resonance growth has disappeared. It should be noted at this long stopping time the terminal velocity approximation does not apply \citepalias{paper1}.
In the $\mu=3$ collection of Figure~\ref{fig:compmu3}, the PB dust distribution gives rise to a hybrid map of fast growth with ridges of fast growth at locations slightly shifted form the monodispserse case, but with sharp cutoffs. In particular, in the ${\tau_{\rm s,peak}}=10^{0}\Omega^{-1}$ PB case a fast growing ridge is produced where none existed with the MRN distribution, and at approximately the same location as the fastest growing ridge in the monodisperse case.
Again as in the $\mu=0.5$ case, the PBT and PB distributions exhibit very similar growth maps, and the distribution of the smallest dust again has little impact on the end result.
Finally, in Figure~\ref{fig:compmu10} the $\mu=10$ cases show the behaviour of PSI at the highest dust to gas ratios in this survey.
Notably, at ${\tau_{\rm s,peak}}=10^{0}\ \Omega^{-1}$ here the MRN case contains a ridge of fast growing modes as seen also at lower $\mu$--${\tau_{\rm s,peak}}$ combinations. This suggests the size resonance at longer stopping time requires a higher $\mu$ to produce instability.
Inspecting the eigenfunction shown in Figure~\ref{fig:MRNmu10taus1eigenfuncs} one can see how the size resonance occurs here well below the peak of the dust distribution, at ${\tau_{\rm s}}=2.9\times10^{-2}\ \Omega^{-1}$.
The difference between the growth rate maps for the MRN and PB dust distributions leads to the question of how the presence of the bump at the peak of the PB distribution enables faster growth.
Figure~\ref{fig:mrn_vs_pb_eignenfunctions} shows dust eigenmode structures as a function of dust stopping time for fastest growing modes in the $\mu=10$, ${\tau_{\rm s,peak}}=10^{-1}\ \Omega^{-1}$ maps for MRN and PB dust distributions as calculated with the direct solver.
These show how the additional abundance of the ${\tau_{\rm s}}\sim 10^{-1}\ \Omega^{-1}$ dust in the PB distribution allows significantly faster growth in the PSI.
In the left column for the MRN distribution, the size resonance and the corresponding fast changes in the dust size density are located at ${\tau_{\rm s}}=9.4\times10^{-3}\ \Omega^{-1}$, whereas with the PB distribution in the right column the resonance is shifted up to ${\tau_{\rm s}}=7.4\times10^{-2}\ \Omega^{-1}$.
This also explains why the short stopping time tail of the size density has little effect on the size resonance growth rates -- this tail contains little mass, and does not change the background dust drift velocities, and so has little effect on the location of the size resonance.
Cross checking of the results from the root finding algorithm by the direct solver for select cases shown here can be found in Appendix~\ref{app:psidirectcomp}.
We have also conducted a further study of dust distributions with very long stopping times in Appendix~\ref{app:longstop}.
Such results are mathematically instructive, but the difficulty of growing dust past ${\tau_{\rm s}}\sim 10\ \Omega^{-1}$ in a coagulation scenario and the lack of physical consistency of the model equations in this regime lessens the relevance of these calculations to planetesimal formation.
\section{PSI growth rates - Turbulent}
\label{sec:turbulence}
To investigate the effect of background turbulence in the disc on the PSI, in this section we employ the diffusive turbulence model terms described in section~\ref{sec:turbmodel}.
Figure~\ref{fig:maximize_alpha} displays the maximal growth rates for the PSI in the wavenumber space $K_{x,z}\in[10^{-1}, 10^3]$ for the same dust abundance parameters as employed in section~\ref{sec:laminar} for the monodispse, MRN, an PB distributions.
These curves were obtained by a local optimization in wavenumber space in steps across $\alpha$ to find the maximum growth rate within the space $K_{x,z}\in[10^{-1}, 10^3]$. Starting from the $\alpha=0$ limit with an initial guess derived from the wavenumber space grids in section~\ref{sec:laminar}, the growth rate was maximized at each step with a simplex algorithm search ({\tt scipy.optimize.minimize} with the Nelder-Mead algorithm) using the direct solver with 512 log-spaced points in ${\tau_{\rm s}}$ space as the target function.
The direct solver \citepalias{paper1, paper2} was employed as it gives sufficient accuracy at the maximal growing point due to the second order convergence, but also provides a function which slopes towards that maximum from points in wavenumber space even where there is no physical growth (further discussed in Appendix~\ref{app:psidirectcomp}).
This conveniently prevents the simplex algorithm from failing when those points are tested in intermediate steps while searching for the maximum.
For the cases of most interest for planet formation, with $\mu > 1$, growth rates for either polydisperse distribution in Figure~\ref{fig:maximize_alpha} are uniformly less than those for monodisperse dust.
In this linear analysis PSI does not save the streaming instability from the difficulties posed by the diffusive model of turbulent mixing.
However, it is also at least the case that the PB distribution maintains growth rates almost as large as those for monodisperse dust, so with dust processing able to create such distributions PSI should have similar planet forming potential to mSI even in the presence of turbulence.
Indeed, considering turbulence in a self-consistent manner should result in a stopping time dependent scale height for the dust, with greater settling for large grains than small.
In the $\mu=0.5$ cases shown in Figure~\ref{fig:maximize_alpha} the relation between the $\alpha$-dependency of the PSI and mSI is more complicated than in other scenarios.
We find that at ${\tau_{\rm s,peak}}=10^{0}\Omega^{-1}$ the maximum growth rates scale similarly with $\alpha$ for the PB and monodisperse distributions.
Recall from Figure~\ref{fig:compmu0p5} how in this case at $\alpha=0$ the ridge of fast growth in wavenumber space was very similar for the PB and monodisperse distributions, and see here that it also damps with $\alpha$ in a very similar way.
At ${\tau_{\rm s,peak}}=10^{-1}\Omega^{-1}$ the maximum growth rate for the monodisperse case is much faster in the $\alpha=0$ limit (Figure~\ref{fig:compmu0p5}) than with the MRN or PB distribution, and so the damping with increasing $\alpha$ is also quite different.
Only small islands of fast growth at low $K_z$ are present at $\alpha=0$ in the MRN and PB cases, and $\alpha$ needs to increase to significant level before affecting these.
In the monodisperse case, since the fastest growth at $\alpha=0$ is at high $K_x$ and $K_z$ the growth rates are damped much sooner as $\alpha$ is increased. Nevertheless, the maximum growth rates for the monodispserse case exceed the two PSI cases at all values of $\alpha$.
The surprising case in this $\mu=0.5$ set is the results for ${\tau_{\rm s,peak}}=10^{-2}\Omega^{-1}$. Here, at $\alpha=0$ the maximum growth rate for the MRN and PB distributions was faster than for the monodisperse distribution (Figure~\ref{fig:compmu0p5}).
Since the instability at play is particularly well differentiated between the PSI cases (size resonance) and the monodisperse case (mSI) the effect of increasing $\alpha$ on the two is starkly different.
In the PSI cases, the increase in $\alpha$ gradually reduces the ridge of fast growing modes, from high-$K_z$ down.
For the mSI case, the high-$K_z$ section of the ridge of fastest growth dies away most quickly with increasing $\alpha$, while the low-$K_z$ section is more robust, leading to a decrease in growth in two parts.
We illustrate in detail of how the decrease in growth occurs with $\alpha$ in Figure~\ref{fig:maximize_alpha} for a typical case with $\mu=3$ and ${\tau_{\rm s,peak}} = 10^{-1}\ \Omega^{-1}$ in
Figure~\ref{fig:alpha_growth_map}.
Two basic trends are apparent in this series of wavenumber space growth maps.
First, increasing $\alpha$ results in damping of growth first at higher wavenumbers,
and second, like in the inviscid case, the addition of the large particle enhancement in the PB distribution shifts not just the peak growth rates but the entire growth distribution in wavenumber space towards that of the monodisperse dust.
Second, the set of unstable modes present in the more slowly growing area abutting the $K_z$ axis is damped below the threshold first, before the ridge of fast growing modes.
Recall in Figure~\ref{fig:mu3eignefuncs} that the difference between these two regions of growth in terms of the dust eigenfunctions was illustrated, and it was shown the the size resonance is at play in the PSI cases in the ridge of fast growth but not in the more slowly growing low-$K_x$ modes.
This difference appears to carry over to the viscous case too, with the different classes of modes having a different susceptibility to turbulent viscosity.
This dependence on the type of mode extends to the relation between damping due to turbulent viscosity and damping due to dust diffusion. Damping due to dust diffusion contributes at least in similar magnitude to gas viscosity in all cases we investigated. The relationship between damping due to gas viscosity and dust diffusion is complex overall and depends at least on the dust distribution function and Schmidt number chosen and its dependence on particle size. This warrants further investigation.
Planetesimal formation is however unlikely to be strong affected by this property, as the low $K_x$ modes do not appear to show growth rates $\gtrsim10^{-2}\Omega$ in any case tested in this work.
Considering these results overall, it is apparent that, if this viscous diffusion model captures all of the effects of turbulence on the PSI or mSI, then the maximum $\alpha$ under which any form of SI could drive planetesimal formation in these conditions is a relatively low $\alpha \sim 10^{-5}$.
This suggests that the full nonlinear dynamics of the gas-dust system, beyond that captured here, may well be very important for enabling planet formation.
Stratified nonlinear three dimensional simulations of mSI self-consistently include particle-generated turbulence, which may behave differently than the simplified isotropic diffusion model of turbulence applied here and in similar calculations \citep{2019arXiv190605371U,2020ApJ...891..132C}.
Further development of numerical methods for 3D nonlinear fluid simulations with polydisperse dust will be required to tackle these questions.
\section{Discussion}
\label{sec:discussion}
\subsection{Scenarios for planet formation}
Armed with the results presented here, we are in the position for the first time to speculate on the impact of the real physical distribution of dust sizes on planetesimal formation, and hence planet formation.
It is observationally known that the dust size distribution in protoplanetary discs differs from that in the interstellar medium, so some degree of dust growth and a change to the shape of the dust distribution is known to occur.
To guide intuition, consider the simplest mapping of the canonical scenario for planetesimal formation by mSI to one based on PSI.
If fast linear growth of the instability (on the order of orbits) is a necessary condition for planetesimal formation, and the same underlying history of dust evolution is assumed, then beyond the single, peak size ${\tau_{\rm s,peak}}=10^{-1}\ \Omega^{-1}$ we would expect a PB or PBT-like dust distribution of dust centered on this peak size.
In these conditions, the mSI grows quickly.
For particles with this stopping time, the mSI grows significantly faster for $\mu\gtrsim1$ \citep{2007ApJ...662..627J}.
However, the difference across this boundary is with the PSI is starker, so the requirement for $\mu\gtrsim1$ is stronger in the PSI case.
Producing this high local gas to dust ratio in a protoplanetary disc will require physically separating dust and gas by radial and vertical settling through the disc.
Competing against this dust settling will be turbulent mixing and global gas flows.
Our results suggest that if the dust can concentrate to $\mu\gtrsim1$ then PSI can proceed on timescales only a few times longer than mSI.
However, to form a planetesimal, the PSI must grow to induce gravitational collapse of the concentrated dust cloud.
The conditions needed to produce this may not be the same, and may be more stringent, than those for fast linear growth.
It is possible that the PSI could be nonlinearly seeded by turbulent fluctuations in the dust concentration, violating the assumptions of the linear analysis used here.
Generally, we cannot say whether growing linear PSI modes will result in planetesimal formation via PSI. Nonlinear effects will likely dominate before PSI clumps can become massive enough to collapse under self gravity, so more barriers may need to be overcome in PSI planetesimal formation, and these barriers cannot be studied in the linear regime.
The analogous question for mSI may provide an imperfect guide.
It is difficult to interpret nonlinear mSI simulation results in this context, as although it has been shown that smaller (${\tau_{\rm s,peak}}=10^{-2}\ \Omega^{-1}$) dust is susceptible to mSI growth the resulting concentrations may have a weaker tendency to collapse into planetesimals \citep{2017ApJ...847L..12S,2019ApJ...883..192A,2020arXiv200801727C}.
This could be due to two separate reasons.
If this smaller dust is less prone to settling, then the $\mu$ which can be achieved is limited, and the mSI growth is only fast for this short stopping time dust when $\mu>1$.
There does exist a $\mu<1$ regime where the PSI has a interesting property of supporting much faster growth then the mSI, but this appears to rely on an unevolved pure power-law dust distribution, and the fast growth is particularly susceptible to turbulent diffusion.
Achieving the right conditions for this kind of fast growth then appears to be very difficult.
Thus the most likely conditions for PSI driven planetesimal formation appear to be similar to those already found to produce self-gravitating fragments in nonlinear simulations of mSI.
Achieving the favourable conditions of $\mu\gtrsim1$ and ${\tau_{\rm s,peak}}=10^{-1}\ \Omega^{-1}$ is not a trivial matter however.
For example, to produce this regime at $45\ \mathrm{au}$ in the solar nebula \citet{2019NatAs...3..808N} invoked sub-cm pebbles under the conditions of a late-stage photoevaporating disc where the falling gas density lengthens the stopping time of a fixed size dust grain.
Although this scenario is physically reasonable for a late-formed Kuiper belt object, the Solar System planets would need to form much earlier in a much denser gaseous disc -- particularly the gas giants, which need to acquire a gaseous envelope from the gaseous disc.
In an attempt to elucidate the conditions which appear to be favourable for PSI to drive planetesimal formation, we can appeal to simple disc models and limits for dust growth.
Assuming Epstein drag (typically applicable to disc conditions outside some tens of au) and a vertically isothermal disc structure, we can conveniently relate the dust particle radius to the stopping time as
\begin{align}
\frac{{\tau_{\rm s}}}{\Omega^{-1}} = \frac{a \rho_s}{\Sigma_{\rm g}} \frac{\pi}{2}\, ,
\end{align}
where $\Sigma_{\rm g}$ is the disc gas surface density \citep{2012A&A...539A.148B}.
Combining this with a minimum mass solar nebular disc model as is commonly used as a reference case for planet formation (also Appendix~\ref{app:bump}) yields
\begin{align}
r_{\rm fast} =11 \left(\frac{M_{\rm disc}}{M_{\rm MMSN}} \frac{{\tau_{\rm s}}}{0.1\ \Omega^{-1}}\right)^{2/3}\left(\frac{a}{1\ {\rm cm}}\right)^{-2/3} \ {\rm au}\, ,
\end{align}
where $r_{\rm fast}$ is the radius where the criteria is achieved and $M_{\rm disc}/M_{\rm MMSN}$ is the ratio of the disc mass to a minimum-mass solar nebula (MMSN) \citep{2015A&A...579A..43C}.
Centimetre-scale dust in a MMSN thus has an $r_{\rm fast}=11\ {\rm au}$, and mm-scale dust an $r_{\rm fast}=51\ {\rm au}$.
Achieving $r_{\rm fast}=5\ {\rm au}$ requires dust with radius $\sim3\ {\rm cm}$, and $r_{\rm fast}=1\ {\rm au}$ a dust radius of radius $\sim40\ {\rm cm}$, although in a MMSN these latter two cases are entering into the Stokes drag regime.
If the gas disc is denser than an MMSN model, these radial locations are pushed outwards, and the required dust sizes down.
To estimate the sizes of the largest dust aggregates present in the disc, we can appeal to the model of \citet{2012A&A...539A.148B}, and in a similar MMSN-type disc (Appendix~\ref{app:bump}) obtain for the stopping time of the fragmentation-limited dust aggregates $\tau_{\rm s, frag}$:
\begin{align}
\frac{\tau_{\rm s, frag}}{\Omega^{-1}} = 0.023\left(\frac{u_{\rm f}}{1\ {\rm m/s}}\right)^2 \left(\frac{r}{1\ {\rm au}}\right)^{1/2}\left(\frac{\alpha}{10^{-5}}\right)^{-1}
\end{align}
where $u_{\rm f}$ is the dust aggregate fragmentation velocity, the turbulence parameter $\alpha$ sets the scale of the velocities relative to the sound speed of the vortices driving the collisions.
This fragmentation velocity is the same as appears also in the physics leading to the PB and PBT dust distributions as described in Appendix~\ref{app:bump}.
The appearance of the turbulence parameter $\alpha$, previously also present in the model for the diffusive effects of turbulence (Section~\ref{sec:turbmodel}), is due to the role of turbulent eddies in driving the equal-size dust collisions in \citet{2012A&A...539A.148B}.
A major weakness of this common manner of parameterizing the physics is the likely possibility that the complexity of turbulent phenomena is beyond a single scalar parameter, and such a tight linkage between the diffusive effect and dust collision driving may not be an appropriate model for the underlying physics.
Nevertheless, in this model and at the turbulence level of $\alpha=10^{-5}$ at $1\ {\rm au}$, $\tau_{\rm s, frag}=0.1 \ \Omega^{-1}$ is not achieved with a fragmentation velocity
$u_{\rm f}=1{\rm\ m/s }$, but if it is $10\ {\rm m/s}$ the required stopping time is easily achieved.
However, if we take the diffusive model of turbulence at face value, considering the results of Section~\ref{sec:turbmodel} the value of $\alpha=10^{-5}$ is a significant impediment to the growth of PSI (or mSI).
Instead considering a very low value of $\alpha=10^{-6}$ then $u_{\rm f}=1{\rm\ m/s }$ is sufficient for producing $\tau_{\rm s, frag}=0.1 \ \Omega^{-1}$ at $1\ {\rm au}$ in this model.
If planetesimals can be formed from dust with lower peak stopping times (lower Stokes number) then the role of the dust enhancement at the peak particle size may be lessened. With ${\tau_{\rm s,peak}}=10^{-2}\ \Omega^{-1}$ we found growth at least on the $10^{-2}\ \Omega$ level, in all the laminar cases in this work with the MRN distribution, unlike at longer peak stopping times. This may open a different window of opportunity for planetesimal formation.
This simplified discussion serves to illustrate the difficulty of obtaining the fuel for PSI driven planetesimal formation in the disc, but being simplified it has also omitted many detailed considerations that could still be significant.
Other hydrodynamic dust enrichment and size sorting pathways are thought to occur in protoplanetary discs. Notably, vortices in protoplanetary discs have pressure peaks at their cores which collect dust \citep{1995A&A...295L...1B,2004_johansen}. Larger dust particles collect in the vortex core faster than smaller particles causing a degree of size sorting as well as dust concentration. The lifetimes of these dust loaded vortices has been called into question due to instabilities shown in \citet{Surville_2019} and \citet{DVorticesII_inprep}, though both works agree that, at least in 2D studies, the vortices survive long enough to considerably increase dust concentration in their cores before succumbing to instabilities. Furthermore \citet{DVorticesII_inprep} finds small vortices to be stable for even longer periods, of the order of hundreds of orbits, giving them plenty of time to cause a considerable dust enrichment at their cores.
Turbulent flow in disc gas can be expected to size-sort and concentrate dust, producing both a localized concentration of dust and a lognormal distribution (in stopping time) with the peak located where the particle stopping time equals the Kolmogorov scale eddy turnover time
\citep{2001PhFl...13.2938H}
Depending on the turbulence parameters, this size scale can be appropriate for concentrating chondrule-sized particles
\citep{2001ApJ...546..496C}, but these are typically a few orders of magnitude below the favourable ${\tau_{\rm s}}=0.1\ \Omega^{-1}$ identified for PSI.
As physical evidence that a PB-like dust distribution may occur in some circumstances in a planet-forming disc, one can appeal to the size-sorting of chondrules and other components of chondritic meteorites
\citep{2010M&PS...45.1124T}
Models for the distributions vary, but
chondrule size distributions have relatively narrow peaks with widths often similar or narrower than the PB distribution considered here \citep{2015ChEG...75..419F,2018M&PS...53.1489M}.
Chondritic meteorites, between the chondrules and other inclusions, consist of a fine-grained matrix material.
The grains in this material often have a power-law size distribution, with for example the Allende CV chondrite matrix grains having a top-heavy power law distribution with $\beta\sim -2$
\citep{1977E&PSL..35...25A,1989E&PSL..92..265T}.
That some chondrites have very low fractions of matrix grains and a minimum chondrule size does however suggest an efficient size-sorting process, although one which may be a result of accretion through an asteroid atmosphere and not related to dust concentration in the disc itself
\citep{2015SciA....1E0109J,2019ApJ...887..230M}.
The reproduction of the inclination distribution of binary orbits in the Kuiper belt from mSI simulations is perhaps the strongest argument a the present time for the involvement of SI in the formation of a solar system \citep{2019NatAs...3..808N}.
Simulations in that work tested formation from particle with stopping times of either $0.2\ \Omega^{-1}$ or $2\ \Omega^{-1}$.
Our results suggest that the only modification to their scenario required to produce similar results with the PSI is the dust size evolution or sorting to produce a PB distribution-like (or more severe) abundance of the largest grains in the dust distribution over an MRN power-law distribution.
Testing this will require a new generation of numerical methods to perform nonlinear PSI simulations.
If however, dust evolution models are unable to produce sufficiently large dust with a sufficiently narrow size distribution at sufficient concentration in the disc for fast PSI growth to occur, alternate models of planetesimal formation, such as direct turbulent concentration \citep{2020ApJ...892..120H}, may provide a path to planet formation.
\subsection{Limitations and future directions}
Our study has been limited in various ways.
We have only considered Epstein drag in this work and the preceding two \citepalias{paper1,paper2}.
Particles with $\tau_{\rm s, frag}\sim 0.1 \ \Omega^{-1}$ at $1\ {\rm au}$ in an MMSN ought to lie in the Stokes drag regime.
However, \citet{2020ApJ...891..132C} did not find significant effects in the linear mSI from the inclusion of Stokes drag, and this property may carry over to the PSI.
The aspect of the PB and PBT dust distribution which appears to allow enhanced PSI growth over a pure power-law MRN distribution is the `cratering bump' attributed to the effect on coagulation-fragmentation from the outcome of differing-sized dust collisions included in the \citet{2011A&A...525A..11B} model.
In the denser parts of the disc, aeolian erosion may act to limit the size of weakly bonded aggregates \citep{2020MNRAS.496.4827R,2020ApJ...898L..13G}.
A coagulation-fragmentation model including this effect should be used in the future to determine where this alters the appearance of the cratering bump, and the methods used here employed to test the effect on PSI growth.
Radially varying gas pressure bumps in the disc may slow the radial drift of dust and cause a size-dependent local enhancement in the dust density.
As \citet{2020arXiv200801727C} have proposed that these dust density enhancements due to mild pressure bumps in the disc can promote fragmentation through mSI, the effect of pressure bump and the changes to the dust distribution due to size-dependent traffic-jam effects may have an effect on the local dust size distribution along with the dust to gas ratio, and hence an effect on determining the growth of PSI.
Locally in pressure bump, globally across the radii of a given protoplanetary disc, and in other contexts where PSI may occur, such as lunar formation \citep{2020LPI....51.2976A}, the radial pressure gradient $\eta$ may vary, and in particular be smaller than the value used in this survey.
The effect of this warrants further investigation in both linear and nonlinear studies of the PSI.
We have only considered vertically unstratified models. Linear mSI calculations for the stratified case have only appeared very recently \citep{2020arXiv201112300L}. It will be very interesting to extend these calculations to the polydisperse case.
The results on turbulent PSI in this work are also limited in several ways. First of all, we consider only the gas turbulence to dust stirring relation described in equation \ref{eqn:dust-turbulence-coupling}, which allows easy comparison with previous work done on the mSI. The character of the turbulent PSI is however strongly dependant on the relationship between dust stirring and turbulence, as dust diffusion plays an important part in damping growth in turbulent PSI. In practice what this means is that the way different dust distributions are affected by turbulence ultimately depends very strongly on the relationship between gas turbulence and dust diffusion at different dust sizes. Secondly, we neglect the effects of dust on turbulence. This may have effects on the relationship between dust and turbulence when considering more complete models of dusty turbulence. Overall, further dedicated study of turbulent PSI may yield interesting results if a more self-consistent approach to modelling turbulence is adopted.
The timing and location of planet formation driven by PSI must be dependent on the dust evolution and concentration that leads to fast PSI growth, and so a model of these timescales and locations should encompass a disc-scale dust evolution model.
In addition, size-differential dust concentration by PSI should also be investigated
as it could directly alter the input to dust coagulation-fragmentation processes,
or lead directly to new and different initial conditions for further stages of streaming instability.
Finally, this work and the previous ones in the series have only considered linear stability.
In our linear calculations, we found that the dust which concentrates most strongly is not the largest (i.e.~${\tau_{\rm s}}$ closest to unity) dust as one might naively expect from intuition based on the mSI, but instead the dust near the size-resonance in modes where it dominates.
If this behaviour persists into the nonlinear stage of the instability the expectation for the constituent particle sizes of bodies formed from PSI may not be a simple reflection of the particles which had ${\tau_{\rm s}}$ closest to unity in the formation environment.
Of course, planetesimal formation results from the gravitationally induced collapse of a dust overdensity.
To fully test the ability of the PSI to form planetesimals, nonlinear simulations including self-gravity will be required.
The results of these papers, and the numerical tools released will serve as useful benchmarks for the ability of those simulation codes to capture the PSI.
\section{Conclusions}
\label{sec:conclusions}
We have performed a first survey of the conditions and linear growth rates of the PSI likely to be encountered in the planet formation process.
We confirm that wide MRN top-heavy power-law dust distributions result in damped growth of PSI in the most regimes unstable for the mSI, with an exception occurring for short stopping times and dust to gas ratios of order unity.
The faster growth in these cases is confined to relatively short wavelengths, and so is vulnerable to disruption by other effects.
However, an enhancement of large particles, like that produced by reasonable dust evolution models, allows linear PSI growth nearly as fast the mSI, as long as the dust to gas ratio $\mu\gtrsim 1$.
We have shown that the theory and tools developed in this series of papers is capable of determining if a streaming instability will grow quickly for a given set of dust parameters, and for the first time taking accurately into account an arbitrary dust size distribution.
This provides the basis for solving one of the major unknowns with the proposed role of streaming instabilities in planet formation - the nature of the dust evolution required to allow the instability to proceed.
As streaming instabilities require at least some dust to grow to sizes with stopping times on the order of an orbit for fast growth, the issue of to what extent the largest dust needs to grow, and how dominated by large dust (`top-heavy') the distribution needs to be must be addressed.
Our results enable the outcome of dust evolution processes to be tested to determine if they give rise to fast PSI.
From the examples studied in this work, we are able to conclude that if fast linear growth of the PSI is a prerequisite for planet formation, then one favourable scenario relies both on production of dust of sufficiently large stopping times (${\tau_{\rm s,peak}}=10^{-1}\ \Omega^{-1}$), concentration of dust to dust to gas ratio of at least order unity ($\mu>1$) and a dust evolution to a dust size distribution enhanced above canonical top-heavy power-law distributions inferred from pure fragmentation cascade dust evolution the interstellar medium.
We find that a diffusive model of turbulence has similar impact on the linear PSI as it does to the linear mSI, and so requires similar testing in nonlinear and self-consistent simulations.
These linear results need to be further tested with nonlinear simulations, allowing the examination of the saturation of PSI and the collapse of fragments from the enhanced density dust clouds formed.
Finally, though this survey has been inherently limited, the underlying tools \citep{psitools} have been made available to the community so that similar analysis can be performed for specific predicted dust distributions arising from dust evolution models.
\section*{Acknowledgements}
We thank the anonymous referee for insightful and constructive comments.
We acknowledge useful discussions with Richard Nelson, Chao-Chin Yang, and Mordecai-Mark Mac~Low.
This research was supported by an STFC Consolidated grants awarded to the QMUL Astronomy Unit 2017--2020 ST/P000592/1 and 2020--2023 ST/T000341/1.
We acknowledge that the results of this research have been achieved using the DECI resource Beskow based in Sweden at PDC with support from the PRACE aisbl.
This research utilised Queen Mary's Apocrita HPC facility, supported by QMUL Research-IT \citep{apocrita}.
SJP is supported by a Royal Society URF.
\section*{Data availability}
The software used to perform calculations in this work is publicly archived on Zenodo.org \citep{psitools}.
\bibliographystyle{mnras}
|
\section{Introduction}
\label{intro}
For the past few decades, one of the most active areas of research in observational astronomy has been the identification of sources that contribute to the metagalactic ionizing background. This has been particularly important for building a cohesive picture of reionization --- an important epoch in the history of the Universe in which the first galaxies formed and the bulk of hydrogen in the intergalactic medium (IGM) transitioned from neutral to ionized (e.g., \citealt{bou06,bou12,bou15,ouch09,rob15}). Massive stars and quasars/active galactic nuclei (AGN) both produce ionizing photons, though their relative importance to the global ionizing background appears to evolve over time. Most evidence currently favors a scenario in which dwarf star-forming galaxies (SFGs) are the primary driver of hydrogen reionization \citep{ric00,bou06,font07,font14,rob10,rob15,jap17}, while AGN contributions to the ionizing background are small until $z \sim$ 2 -- 3 \citep{barg03,bolt05,cbt09,cri16,smith20}. However, because these scenarios typically rely on an extrapolation of observed galaxy counts to faint absolute magnitudes unobtainable even in the deepest \textit{HST} imaging, some authors have argued that quasars/AGN could produce a non-negligible or even dominant fraction of UV photons during the era of reionization (e.g., \citealt{font12,mad15}). Of course, arguments for or against this alternate scenario depend critically on the assumed number density of AGN at high redshift, which is still a subject of significant debate (see, e.g., \citealt{gia19,cowie20,graz20} and \citealt{shin20} for some recent analyses).
Another major uncertainty in determining the relative importance of SFGs and AGN to reionization is constraining the escape fraction $f_{esc}$, defined as the fraction of all Lyman continuum (LyC, rest frame $\lambda$ $<$ 912 \AA{}) photons that manage to escape their galaxy of origin to interact with the IGM. Most theoretical and semi-analytical models of reionization require an average $f_{esc}$ for SFGs of about 10\% or greater, if SFGs are to be the primary driver of reionization (e.g., \citealt{bolt07,van12a,feng16,price16,kimm17}; see, however, \citealt{fauch08} and \citealt{matt17}), though at the highest redshifts, $f_{esc}$ remains largely unconstrained by observations. For sources at $z \gtrsim$ 4, the rising density of intervening neutral hydrogen absorption systems leads to much lower transmissivity of the IGM to LyC photons \citep{mad95,song04,in14}. This effectively prohibits direct detections of LyC emission (and hence estimates of $f_{esc}$) along most sightlines at redshifts where such measurements are most needed, though some exceptions have been reported in recent years (e.g., the $z = 4$ source \textit{Ion3} in \citealt{van18} and the $z = 3.8$ source \textit{Ion1} in \citealt{ji20}). Thus, observations focused on analogous objects at slightly lower redshifts are often used to constrain the ionization history of the Universe.
Literature reports of individual or stacked LyC detections suggest that most galaxies have small values of $f_{esc}$, at most $\sim1-3\%$ (e.g., \citealt{lei95,stei01,grim09,cow10,leit13,rut16}), and that the typical escape fraction increases with redshift (e.g., \citealt{mit13,font14,fai16,kai16,jap17}). The latter observation is based on a growing number of LyC detections at $z \sim 2-3$ (e.g., \citealt{van10b,van12b,van18,most15,graz16,shap16,jones18,stei18,laces,riv19,saha20}), though it is now well known that direct LyC searches can be affected by contamination from foreground galaxies (e.g., the projections from \citealt{van10a} or the reexamination of sources first reported in \citealt{shap06} by \citealt{siana15}). Stacking analyses also tend to give a relatively weak average LyC signal at these redshifts (e.g., \citealt{siana10,graz17,rut17,marchi18,nai18,laces}; see, however, \citealt{stei18}).
However, some studies have suggested that particular subgroups of galaxies are more likely to have significant LyC escape. For example, moderately large Ly$\alpha$ equivalent width in emission (e.g., \citealt{mich17,marchi18,stei18,laces}) or multiply-peaked Ly$\alpha$ line profiles \citep{verh17, van20} appear to be signposts of nonzero $f_{esc}$; see, however, the confirmed LyC leaker \textit{Ion1}, which shows Ly$\alpha$ in absorption only \citep{ji20}. A high flux ratio of [OIII]$\lambda\lambda4959,5007$ to [OII]$\lambda\lambda3727,3729$ (O32) has also recently been proposed as an optical marker of LyC escape (e.g., \citealt{nakajima14,izo16a,izo16b,izo18,stei18,laces,tang19}). However, this connection remains tenuous, with several studies reporting that large O32 by itself is insufficient to guarantee significant LyC escape (e.g., \citealt{reddy16b, izo17, rut17, nai18, barr20}).
Legacy fields like the GOODS-North and South \citep{goods04} tend to be especially attractive targets for LyC leaker searches, due to the abundance of deep multiwavelength imaging and thorough spectroscopic coverage. In particular, the advent of the \textit{Hubble} Deep UV (HDUV) Legacy Survey \citep{oesch18} has now made it possible to perform direct photometric searches for LyC leakers at high redshift. Using the HDUV's deep \textit{HST}{}/WFC3 imaging in the F275W and F336W bands, \citet{nai17} identified six candidate LyC sources in the GOODS fields at $z \sim 2$, all with $f_{esc} \gtrsim$ 13\%. However, at the redshifts probed by \citet{nai17}, the Lyman break sits squarely in the middle of the F275W bandpass. Since both ionizing and non-ionizing photons contribute to each object's F275W photometry, a number of modeling assumptions are needed to sift out the contribution of just the LyC to the F275W flux.
However, if one were to push to slightly higher redshifts, say $z \gtrsim 2.4$, as we did in the GOODS-N in \citet{jones18}, then the F275W filter becomes sensitive to LyC photons \textit{only}. In that work, we identified a raw total of six sources with spectroscopic redshifts $z \gtrsim 2.4$ that remained bright in F275W. However, four of these were then shown via optical spectroscopy to be line-of-sight blends of low- and high-redshift galaxies.
In this paper, we turn our attention to the GOODS-S, using the wealth of optical/IR spectroscopy in this field in concert with deep F275W imaging from the HDUV survey, to search for individual candidate LyC-leaking galaxies at $z >$ 2.35. We place constraints
on the mean LyC signal at this redshift using an averaging analysis and on the contributions to the overall ionizing luminosity density from SFGs. In Section \ref{data}, we describe the data we used to select and characterize potential high-redshift LyC leakers, including optical/IR redshift catalogs and spectra, and UV, X-ray, and optical imaging. We present our search for individual candidate LyC leakers in Section \ref{indiv_results} and discuss the properties of our candidate sources, along with evidence for or against contamination by foreground galaxies for each. In Section~\ref{averaging}, we perform an averaging
analysis of all $z = 2.35 - 3.05$ sources with F275W coverage in the GOODS-N, in the GOODS-S, and in the two fields combined, from which we measure their total F275W contributions. In Section~\ref{stacking}, we stack the F275W and F336W images for each GOODS field and for the two
fields combined---first for $z = 2.35 - 3.05$, and then for three redshift bins for the combined images to look for differences that might arise due to ``dilution'' from Lyman absorption along the line-of-sight, which increases with redshift. In Section \ref{flux}, we estimate the metagalactic ionizing background at $2.35<z<3.05$ and compare to the flux required to maintain an ionized IGM at these redshifts.
Finally, we summarize our findings in Section~\ref{summary}.
We assume $\Omega_{M}$ = 0.3, $\Omega_{\Lambda}$ = 0.7, and $H_{0}$ = 70~km~s$^{-1}$~Mpc$^{-1}$ throughout this work. All magnitudes are given in the AB system; magnitude zero-points for the HDUV F275W and F336W observations are given in \citet{oesch18}.
\section{Data}
\label{data}
\subsection{F275W Imaging}
\label{datahduv}
The HDUV survey (GO13872; \citealt{oesch18}) is a 132-orbit WFC3 imaging program centered on the GOODS-North and South fields. Designed to capitalize on existing WFC3/UVIS imaging from the CANDELS \citep{grog11,koek11} and UVUDF \citep{tep13,rafel15} programs, the HDUV survey imaged both of the GOODS fields in the F275W and F336W filters around or within the existing CANDELS and UVUDF footprints.
We focused on the HDUV GOODS-N dataset in \citet{jones18}. Here we focus on the HDUV GOODS-S dataset, which includes one science image per filter and corresponding maps of the per-pixel rms, all drizzled to a 60 mas pixel resolution. The final HDUV images incorporate data from the UVUDF survey, which substantially deepens this region of the field. We do not include data from the region of the CDF-S field observed by the Early Release Science (ERS; \citealt{ers}). This area is almost a magnitude shallower in F275W than the regions that we use, and it contains significantly more artifacts; thus, we are unable to construct a high-significance $<26$ magnitude sample for that region.
When measuring our F275W fluxes (see Section~\ref{uvsample}), we smoothed the F275W and F336W images using the \texttt{scipy.ndimage.gaussian\_filter} function with a standard deviation of 2 to minimize the impact of residual noise on our measurements, though as we will show, these residuals still contribute significantly to our flux uncertainties.
\subsection{Optical/NIR Spectroscopy}
\label{zspec}
Secure spectroscopic redshifts are required for reliable identification of LyC-leaking candidates, as photometric redshift estimates and by-eye inspections are frequently contaminated by galaxies along the line-of-sight (e.g., \citealt{van10a,siana15,jones18}). We therefore narrow our search only to those galaxies with robust spectroscopic redshifts between 2.35 and 3.05. At the lower redshift bound, the Lyman limit sits at 3055.2 \AA, at which point the total throughput of the F275W filter has declined to 2.3\% on the red side. Adopting a slightly stricter redshift minimum of $z = 2.37$, at which point the Lyman limit sits at 3073.4 for a total F275W throughput of 1\%, does not remove any of our candidates, nor does it substantially change the results of our stacking analyses in Section \ref{averaging}.
Similarly, the upper redshift bound relates to the wavelength at which the F336W filter probes only the redshifted LyC. Although we considered F275W-bright candidates with redshifts up to $z=3.55$ in the GOODS-N in \citet{jones18}, in this work we use $z = 3.05$ as the upper redshift bound in both the GOODS-N and GOODS-S subsamples for consistency. In principle, ionizing sources at $z > 3.05$ could also be detected in F336W using the methods described below, though the rapidly increasing opacity of the IGM to LyC photons at $z = 3 - 4$ would make such a search somewhat more difficult (e.g., \citealt{in14}). Since the LyC is redshifted fully into the F275W bandpass at $2.35 < z < 3.05$, the deep and relatively wide F275W coverage provided by the HDUV survey enables us to search for potential sources of ionizing radiation at high redshift.
We consulted spectroscopic catalogs from the literature (\citealt{popesso09,balestra10,cowie12,kurk13,morris15,bacon17,inami17,vandels,barg19,musewide}), along with unpublished spectra that we have obtained with Keck/DEIMOS,
to identify sources in the GOODS-S in the redshift range of interest. We limited ourselves to objects with redshifts marked as high-quality/confidence by the respective authors. For example, we only included spectroscopic redshifts from the MUSE-Deep and MUSE-Wide surveys that are flagged as confidence level 3, which represent unambiguous, multiple-line detections (for MUSE-Deep; \citealt{bacon17} and \citealt{inami17}) or identifications with ``very high certainty" (for MUSE-Wide; \citealt{musewide}).
Our final list of candidates (see Section \ref{uvsample}) has spectroscopic redshifts from the MUSE-Deep program and from the catalogs of \citet{balestra10} and \citet{kurk13}.
\subsection{X-ray Imaging}
\label{xrays}
To identify X-ray counterparts to our F275W sample (see Section \ref{uvsample}) and to weed out probable AGN, we used X-ray data and catalogs from the 7 Ms \textit{Chandra X-ray Observatory} exposure of the \textit{Chandra} Deep Field-South \citep{luo17}, which reaches a limiting flux of $f_{0.5-2 \mathrm{ keV}} \approx 6.4 \times 10^{-18}$~erg~cm$^{-2}$~s$^{-1}$ near the image center. There are 95 F275W sources with X-ray counterparts within a 1\farcs5 search radius. We computed the rest-frame 2 -- 8 keV luminosities, $L_{X}$, of these counterparts from the 0.5 -- 2 keV fluxes with an assumed $\Gamma = 1.8$ and no absorption correction using
\begin{equation}
L_X = 4\pi d^2_L f_{0.5-2 \mathrm{keV}}\bigg(\frac{1+z}{4}\bigg)^{\Gamma-2}\ \mathrm{erg\ s^{-1}.}
\end{equation}
Of the sources with X-ray counterparts, four have $L_X > 10 ^{42}$ erg s$^{-1}$, which classifies them as AGN. One of these is a BALQSO. However, none of our candidate high-redshift LyC-leaking sample (defined below) have X-ray counterparts.
\section{Search for Individual $\lowercase{z} \sim 3$ Candidate LyC Leakers in the GOODS-S}
\label{indiv_results}
\begin{figure}[bht]
\includegraphics[width=3.33in]{z_b.pdf}
\includegraphics[width=3.33in]{GOODS_S_uv_z.pdf}
\caption{\emph{Top:} Spectroscopic redshift vs. \textit{HST}/ACS F435W magnitude for all galaxies in the 43.5~arcmin$^2$ footprint of the HDUV GOODS-S with F850LP $<$ 26, which defines our parent sample.
\emph{Bottom:} Spectroscopic redshift vs. \textit{HST}/WFC3 F275W magnitude for our primary UV sample of 1115 sources in the GOODS-S with F275W magnitudes $<$ 26 and total errors fainter than 27.19, corresponding to a $3\sigma$ detection. Red circles mark sources with an X-ray counterpart in the \textit{Chandra} Deep Field South 7 Ms catalog. The purple shaded region shows the redshift interval where the F275W bandpass straddles the Lyman limit. The four $> 3\sigma$ LyC-leaking candidates lie above this region.
}
\label{uvz}
\end{figure}
\citet{jones18} describes our search for LyC-leaking candidates in the HDUV GOODS-N field. In the GOODS-S, we begin with all F850LP $<$ 26 galaxies that are covered by the 43.5~arcmin$^2$ footprint of the HDUV observations. This area lies entirely within the GOODS-S observations of \citet{goods04} obtained with \textit{HST}{}'s Advanced Camera for Surveys (ACS). At $z \sim$ 2.5, the ACS F850LP filter probes the rest-frame FUV at $\sim$2500 \AA, providing a reasonable selection of star-forming galaxies at these redshifts. In the top panel of Figure~\ref{uvz}, we plot redshift versus F435W magnitude for this area. Spectroscopic identifications are 97\% complete below F435W = 24, 85\% from F435W = 24$-$24.5, and 63\% from F435W = 24.5$-$25.
\subsection{F275W Measurements in the GOODS-S}
\label{uvsample}
Using the Python Source Extraction and Photometry (SEP) library\footnote{https://github.com/kbarbary/sep} \citep{sep}, which is based on the SExtractor program of \citet{sextractor}, we measured F275W fluxes in $2\arcsec$ and $4\arcsec$ diameter apertures at the position of each F850LP $<$ 26 source with UV coverage. We also estimated the local median background at each source position in a $4 - 8\arcsec$ diameter annulus and subtracted this from the aperture fluxes. Comparing background-subtracted small- and large-aperture magnitudes (for extended objects with $21 <$ F275W $< 25$, to avoid saturated or low S/N sources) reveals a median offset of -0.18 mag, which we added back to the $2\arcsec$ magnitudes as an aperture correction.
We initially measured magnitude errors in $2\arcsec$ apertures from the associated rms noise files, with the same -0.18 mag aperture correction applied. However, this resulted in the selection of objects that, though they are detected nominally at moderate ($>3\sigma$) significance in F275W, appeared to be pure noise upon visual inspection.
To estimate the effect of residual noise in the science maps on our flux measurements, we performed the same aperture photometry methods described above at $\sim$10,000 random positions across the HDUV field in both the F275W science and rms images, with the same smoothing applied for consistency with our source flux measurements. If we were to use only the rms image to determine uncertainties, we would underestimate the true noise by a factor of 2. This translates to an offset of -0.376 mag, which we added to our initial magnitude errors.
Our primary UV sample consists of 1115 sources with F275W magnitudes brighter than 26 and total errors fainter than 27.19, corresponding to a $>3\sigma$ detection. In the bottom panel of Figure~\ref{uvz}, we show redshift versus F275W magnitude for this sample. Spectroscopic identifications are nearly fully complete to F275W = 24.5 (with only two sources out of 248 missing or unidentified), 92\% complete from F275W $= 24.5 - 25$, and 57\% complete from F275W $= 25 - 26$. The rapid drop-off in spectroscopic completeness at fainter UV magnitudes is a major motivation for our F275W = 26 cutoff, even though one might expect the number of candidate $z \sim 2.5-3$ LyC leakers to increase at very faint UV magnitudes. We have four LyC-leaking candidates in the redshift range $2.35 < z < 3.05$ from this selection.
Requiring a $>3\sigma$~detection in F275W at the depth of the HDUV imaging will necessarily miss UV-fainter sources that otherwise may merit further consideration as LyC-leaking candidates. We therefore used the following color selection as a secondary way of selecting potential LyC leakers. We required (1) F435W $<$ 25; (2) F606W $-$ F850LP $<$ 1 mag; and (3) F275W signal-to-noise ratio (SNR) $>$ 2. The first condition ensures that this color-selected sample remains fairly spectroscopically complete, while the other conditions select sources with at least marginal F275W detections and relatively flat rest-frame, non-ionizing UV continua. As shown in Figure \ref{color_z}, this selection yields four candidates above our minimum redshift threshold, three of which were already selected directly using our $>3\sigma$ criterion. The remaining source (enclosed in the green circle in Figure \ref{color_z}) is detected at 2.6$\sigma$ in F275W.
In summary, we have identified five candidate high-redshift LyC leakers in the redshift range $2.35 < z < 3.05$ in the GOODS-S, four of which are detected at $>3\sigma$ significance in F275W. With a total of 46 sources with UV coverage and secure redshifts in the redshift range, this leads to a $\sim 10.9\%$ raw success rate in identifying candidate LyC-leaking galaxies. However, after removing two candidates contaminated by foreground galaxies (see Section~\ref{lyccandidates}), this success rate drops to $\sim6.5\%$.
\begin{figure}[t]
\includegraphics[width=3.33in]{uvb_z.pdf}
\caption{F275W $-$ F435W color vs. spectroscopic redshift for sources with relatively flat rest-frame UV continua. Red circles mark sources with an X-ray counterpart in the \textit{Chandra} Deep Field South 7 Ms catalog. Sources with lower limits on F275W $-$ F435W are plotted at their 2$\sigma$ values with blue arrows. The open green circle marks the single $2.35 < z < 3.05$ source detected at $>2\sigma$ significance in F275W that was not already selected with our $>3\sigma$ significance criterion. The purple hatched region shows the redshift interval where the F275W bandpass straddles the Lyman limit.}
\label{color_z}
\end{figure}
\begin{figure*}[ht]
\centering
\begin{tabular}{p{1.8in}p{1.8in}p{1.8in}p{1.8in}}
\includegraphics[width=0.8\linewidth,height=0.8\linewidth]{9890bzh} & \hspace{-1cm}
\includegraphics[width=0.8\linewidth,height=0.8\linewidth]{9890u} &
\includegraphics[width=0.8\linewidth,height=0.8\linewidth]{20231bzh} & \hspace{-1cm}
\includegraphics[width=0.8\linewidth,height=0.8\linewidth]{20231u}
\\
\includegraphics[width=0.8\linewidth,height=0.8\linewidth]{20277bzh} & \hspace{-1cm}
\includegraphics[width=0.8\linewidth,height=0.8\linewidth]{20277u} &
\includegraphics[width=0.8\linewidth,height=0.8\linewidth]{24639bzh} & \hspace{-1cm}
\includegraphics[width=0.8\linewidth,height=0.8\linewidth]{24639u} \\
\hspace{0.49\textwidth}
\includegraphics[width=0.8\linewidth,height=0.8\linewidth]{13629u} &
\includegraphics[width=0.8\linewidth,height=0.8\linewidth]{13629bzh} \\
\end{tabular}
\caption{
Three-color images (\emph{left}; red = F160W, green = F850LP, and blue = F435W) and F275W thumbnails (\emph{right}) of our five candidate LyC leakers. Cyan contours show regions where the F275W per-pixel SNR exceeds 5$\sigma$, and the white dashed circle shows the $2''$ diameter aperture used in measuring the magnitudes. Images are $6''$ on a side; North is up and East is to the left. For each source, the redshift (F275W SNR) is given at the bottom of the left (right) image. Sources 9890 and 20277 were found to be contaminated by low-redshift interlopers.
}
\label{thumbnails}
\end{figure*}
\subsection{Five Candidate LyC Leakers in the GOODS-S}
\label{lyccandidates}
In Table \ref{tab:sources}, we list the basic properties of our five candidate LyC leakers in the GOODS-S. This includes an ID number, identical to those given in Table 3 of \citet{guo13}; decimal coordinates; a spectroscopic redshift from ground-based observations;
the F275W magnitude and error (including the -0.18 mag aperture correction and -0.376 mag noise correction); F435W magnitude; and F606W -- F850LP color.
In Figure \ref{thumbnails}, we show both a three-color thumbnail (left; red = F160W, green = F850LP, blue = F435W) and the F275W thumbnail (right) of each source.
{\setlength{\extrarowheight}{3pt}
\begin{table*}[t]
\normalsize
\centering
\caption{\\\textsc{Summary of Candidate LyC Leakers}}
\begin{tabular}{c c c c c c c}
\hline
\hline
ID$^{a}$ & R.A. & Dec. & $z_{spec}^{b}$
& F275W (Error) & F435W & F606W -- F850LP \\%& Selection$^{c}$\\[3pt]
\hline
\setrow{\bfseries}9890$^{1}$ & 53.096661 & -27.772339 & 2.483 &
25.50 (27.21) & 24.68 & 0.57 \\
20231$^{2}$ & 53.157532 & -27.798981 & 2.678
& 25.67 (27.54) & 25.49 & 0.33 \\
\setrow{\bfseries}20277$^{3}$ & 53.157845 & -27.814756 & 2.573
& 25.93 (27.30) & 24.40 & 0.39 \\
24639$^{3}$ & 53.188301 & -27.829344 & 2.571
& 26.00 (27.22) & 24.58 & 0.21 \\
\hline
13629$^{1}$ & 53.118439 & -27.805323 & 2.627
& 26.18 (27.24) & 24.59 & 0.16 \\
\hline
\end{tabular}
\label{tab:sources}
\tablecomments{$^a$Object IDs are the same as in Table 3 of \citet{guo13}. Sources with ID numbers in bold were later confirmed to be contaminated by foreground galaxies upon inspecting their spectra. \\
$^b$Spectroscopic redshifts from (1) GMASS \citep{kurk13}, (2) MUSE-Deep \citep{bacon17,inami17},\\ or (3) \citealt{balestra10}. \\
}
\end{table*}
}
Although we selected our candidates to have high-quality spectroscopic redshifts from $2.35 < z < 3.05$, the optical spectra for two sources (objects 9890 and 20277 in Table \ref{tab:sources}) show evidence of an intermediate redshift galaxy along the line-of-sight to the high-redshift galaxy. In source 9890, this is betrayed by the presence of an [OIII]$\lambda\lambda4959,5007$/H$\beta$ complex at $z = 0.784$. In source 20277, we see a strong [OII]$\lambda\lambda3727,3729$ doublet at $z = 1.094$. At these intermediate redshifts, the F275W filter probes the rest-frame FUV at $\sim 1300-1500$ \AA, well above the Lyman limit. The UV fluxes for these two sources are almost certainly dominated by the foreground galaxies and hence cannot be used to measure the ionizing flux output at $z \sim 2.5$.
We note that these contaminated galaxies, particularly object 9890, reinforce the need for deep optical and NIR spectroscopy to confirm the nature of any proposed LyC leakers. In the case of object 9890, the photometric redshift without HDUV data ($z_{phot} = 0.698$; \citealt{3dhst}) is more consistent with the spectroscopic redshift of the foreground, low-redshift galaxy. Yet once the HDUV data are included, the photometric redshift better matches the spectroscopic redshift of the background galaxy ($z_{phot} = 2.36$; \citealt{oesch18}), meaning a selection based only on photometric redshifts would falsely label this source as a LyC leaker.
We now briefly discuss our remaining three candidate LyC leakers in the GOODS-S and the foreground contamination rate in our overall candidate sample. First, three distinct components are seen in the three-color thumbnail for source 20231 (upper-right of Figure \ref{thumbnails}): a generally red galaxy with two clumps, both of which are at almost-identical spectroscopic redshifts (R. Bacon, private communication), though it is unclear whether they constitute a single galaxy or an interacting pair. In any case, the F275W flux is not coincident with the blue clump to the southeast, but instead with the relatively red knot of emission near the image center.
Second, in source 24639 (middle-right of Figure \ref{thumbnails}), the F275W emission is offset from the position of the non-ionizing flux from this galaxy by $\sim0\farcs8$, which corresponds to about 6.4 proper kpc at $z = 2.571$. Though regions of LyC escape may not always be fully coincident with the bulk of the stellar emission, the lack of non-ionizing flux at the position of the F275W emission in this source is unusual. We do not remove this object from the list of LyC candidates, but in subsequent sections we note the effects of excluding it from our calculations.
Third, there is a similar off-center clump of F275W emission to the east of source 13629. However, unlike the case for source 24639, object 13629 also shows hints of F275W emission coincident with the position of the non-ionizing flux from this galaxy.
To estimate the expected level of foreground contamination, we ran Monte Carlo simulations to randomize the positions of all galaxies in the HDUV GOODS-S field, regardless of redshift identification or F275W magnitude. This assumes a spatially uniform foreground F275W population with a surface density of around 486,900 deg$^{-2}$ (the number of galaxies within the HDUV GOODS-S footprint divided by its 43.5 arcmin$^{2}$ area), similar to the assumptions in \citet{van10a}. For a sample size of 5, at least one galaxy is contaminated in 27\% of our 500 simulations, with a mean of 0.3 $\pm$ 0.6 contamination events. This is consistent with expectations from a simple binomial probability distribution, in which the probability of any single contamination event is proportional to the source surface density and the aperture size \citep{nestor13}. For the surface density given above and an aperture of radius 1\arcsec, an average of 0.48 $\pm$ 0.66 contamination events are expected in a sample of size 5.
Finally, we must also be concerned about the limitations of our sample selection in that we are only using sources with known spectroscopic redshifts. As we have argued above, it is dangerous to use photometric redshifts, which may be biased against continuum leakers, and where sources with incorrect high redshifts may also result in incorrect high fluxes. However, we have inspected the 185 SFGs that would be placed in the $z$ = 2.35 -- 3.05 redshift range by the \citet{straat16} catalog of photometric redshifts in the field and do not have spectroscopic redshifts. These have only two $3\sigma$ F275W detections and no significant total signal ($-0.25 \pm 0.58~ \mu$Jy).
Although the apparent offsets between the ionizing and non-ionizing radiation in some of these sources (particularly object 24639) are peculiar, we retain all of these objects as candidate LyC leakers. We emphasize that an absence of evidence \textit{for} contamination by foreground galaxies does not equate to positive evidence that such foreground contaminants are absent. For this reason and out of an abundance of caution, we assert that our three remaining sources in the GOODS-S should only be considered LyC-leaking \textit{candidates}.
\section{Mean ionizing emission from an averaging analysis using both GOODS fields}
\label{averaging}
In addition to searches for individual LyC leakers, the depth and breadth of UV coverage offered by the HDUV data enable a robust measurement of the ``typical" ionizing flux output of $z \sim 2.5$ galaxies via averaging.
We select all $2.35 < z < 3.05$ galaxies with high-quality spectra that lie within the F275W footprint of the HDUV survey in both GOODS fields. For each field and band subsample, we applied the same photometric procedure that we used in our search for individual candidate LyC leakers in Section~\ref{uvsample} in order to determine the flux and error for all the sources, including those not detected individually. In Figure~\ref{zm_flux}, we show F275W flux versus F850LP magnitude for our F275W subsamples in the GOODS-S and the GOODS-N.
We next excluded sources with X-ray counterparts and sources whose spectra clearly include emission features from a line-of-sight foreground contaminant. The X-ray non-detection criterion precludes any major contributions from AGNs to our stacks. For the GOODS-S, we use the 7~Ms \textit{Chandra} Deep Field-South catalog (\citealt{luo17}; see Section~\ref{xrays}). For the GOODS-N, we use the 2 Ms \textit{Chandra} Deep Field-North catalog (\citealt{cdfn,xue16}). We also searched for any objects with very nearby ($< 2\arcsec$) projected neighbors that might affect our aperture flux measurements, but we did not find any. Our upper redshift bound is approximately the redshift at which the F336W filter begins to probe only rest-frame wavelengths $<912$~\AA.
We then computed the error-weighted mean flux of all the sources in each subsample and multiplied that value by the number of sources in the subsample. Because the flux errors for our sources are very similar to one another, this is quite similar to computing an unweighted sum. In the GOODS-S, this selection yielded 38 sources. In the GOODS-N, where the areal coverage in F275W is wider than in F336W, this yielded 91 sources with F275W coverage and 69 sources with coverage in both bands. In Table \ref{tab:stacktable}, we summarize the measured properties for the individual and combined GOODS fields.
In the GOODS-S, the overwhelming majority of the total contribution at F275W (0.46 $\mu$Jy of 0.51 $\mu$Jy) comes from the three candidate LyC leakers listed in Table \ref{tab:sources}. If we remove individual candidates from the stack (specifically, source 24639, which appears to have no non-ionizing emission coincident with its F275W detection), the total F275W flux drops to 0.42 $\mu$Jy, of which 0.32 $\mu$Jy is attributable to the remaining two candidates. Removing all three candidates from the stack reduces the error-weighted mean to 0.002 $\pm$ 0.006 $\mu$Jy, giving a total contribution of only 0.070 $\pm$ 0.21 $\mu$Jy.
\begin{figure}[t]
\includegraphics[width=3.33in,trim={0.5cm 1cm 1cm 0},clip]{gs_zm_flux.pdf}
\includegraphics[width=3.33in,trim={0.5cm 1cm 1cm 0},clip]{gn_zm_flux.pdf}
\caption{F275W flux vs. F850LP magnitude for all $2.35 < z < 3.05$ galaxies with high-quality spectra that lie within the F275W footprint of the HDUV survey in the GOODS-S (\textit{top}) and GOODS-N (\textit{bottom}). Sources with X-ray counterparts are marked in green, while $>3\sigma$ detections, both positive and negative, are marked in either red or blue (the latter is used to indicate sources contaminated by foreground galaxies).
Note that GN-UVC-6 is a $2.9\sigma$ source and hence not marked in red in the lower panel.
GN-UVC-1, the GOODS-N quasar shown in the lower panel, is marked in green with an upward pointing arrow, since its measured flux is much higher than the y-axis limit.}
\label{zm_flux}
\end{figure}
{\setlength{\extrarowheight}{3pt}
\begin{table}[tb]
\normalsize
\centering
\caption{\\\textsc{Properties of the Error-Weighted Summed Fluxes}}
\begin{tabular}{c c c c c}
\hline
\hline
Field & F275W & (SNR) & F336W & (SNR)\\
& ($\mu$Jy) & & ($\mu$Jy) &\\
\hline
GOODS-S & 0.51 & (2.3) & 2.83 & (21)\\
GOODS-N & 0.52 & (1.5) & 6.74 & (23)\\
Both & 1.00 & (2.5) & 9.10 & (30)\\
\hline
\end{tabular}
\label{tab:stacktable}
\tablecomments{The area covered in the GOODS-N in F336W is
smaller than that in F275W. \\}
\end{table}
}
In order to check the background, we randomized the positions of the GOODS-S subsample and measured the magnitudes at these positions. We then processed the sample in the same way as the real sample, i.e., by eliminating low redshift sources, X-ray sources, or very nearby bright neighbors. We measured 120 random samples and found a mean background of 0.070 $\pm$ 0.023 $\mu$Jy, which is negligible compared with the measured value in the real sample. 15\% of the random samples had mean backgrounds that exceeded the total contribution of the real sample, which may suggest that we are slightly underestimating the true noise. However, it should be emphasized that we did not carefully curate the random samples in the same way that we did the real sample, and, thus, we may have a higher degree of unrecognized foreground contamination, which could result in an overly large dispersion.
Despite the greater number of sources in the GOODS-N, the total contribution at F275W is similarly small at just 0.52 $\mu$Jy. As with the GOODS-S, a large fraction of the total contribution comes from just a small number of sources; here $\sim40$\% (0.21 $\mu$Jy) is attributable to just two sources. These are GN-UVC-6, a color-selected LyC-leaking candidate at $z = 2.439$ that we discussed in \citet{jones18} (a $2.9\sigma$ detection in F275W), and a $z = 2.98$ galaxy with an F275W magnitude of $\sim$26.21 (an $\sim3\sigma$ detection). The latter object was not discussed in \citet{jones18}, because we considered only high-redshift objects detected at $\geq 4\sigma$ significance in constructing our F275W sample, and the source also did not turn up in our color-selected sample, which probed to lower significance. In any case, this leaves $\sim$ 0.3 $\mu$Jy of F275W unaccounted for, which must come from the remaining 89 sources.
Finally, if we combine the two fields, we calculate a total contribution of 1.00 $\pm$ 0.41 $\mu$Jy at F275W, of which about 0.67 $\mu$Jy is directly attributable to the five LyC-leaking candidates discussed above (3 in the GOODS-S and 2 in the GOODS-N). Removing an individual candidate from this calculation (specifically source 24639, for the reasons mentioned in Section \ref{lyccandidates}) does not substantially change this result: the remaining 128 galaxies combined contribute 0.87 $\pm$ 0.39 $\mu$Jy, of which 0.53 is directly attributable to the remaining candidates. Given the error on the total contribution, both cases are consistent with the remainder of our sample contributing little or nothing to the overall ionizing flux output at $z \sim 2.5$.
\begin{figure}[t]
\centering
\begin{tabular}{c c}
\includegraphics[width=0.45\linewidth]{F275W_Stack.pdf} &
\includegraphics[width=0.45\linewidth]{F336W_Stack.pdf}\\
\includegraphics[width=0.45\linewidth]{F275W_Stack_N.pdf} &
\includegraphics[width=0.45\linewidth]{F336W_Stack_N.pdf}\\
\includegraphics[width=0.45\linewidth]{F275W_Stack_AllGOODS.pdf} &
\includegraphics[width=0.45\linewidth]{F336W_Stack_AllGOODS.pdf}\\
\end{tabular}
\caption{
F275W (\textit{left}) and F336W (\textit{right}) stacks of spectroscopically identified sources between $2.35 < z < 3.05$ with UV coverage from the HDUV survey. Images from top to bottom show stacks from the GOODS-S, the GOODS-N,
and both fields combined. Images are 6\arcsec~on a side and stretched to the same pixel values for each band (except for the F336W image of the combined GOODS fields, which saturates at the color scale used for the individual fields). North is up and East is to the left. The differing number of sources in the GOODS-N stacks (middle row) are due to wider areal coverage of the field in F275W than in F336W.
}
\label{totalstack_fig}
\end{figure}
\section{An Image stacking visualization in both GOODS fields}
\label{stacking}
For visualization purposes, in Figure \ref{totalstack_fig}, we show stacked images in each band for the sample in each GOODS field, as well as for both fields combined. Bright points more than 1\arcsec~away from the image center (for example, the bright spot to the upper right in the first row of Figure \ref{totalstack_fig}) are likely foreground galaxies with significant F275W flux, though these do not affect our measurements at $z > 2.35$. Most of our candidate LyC leakers across both fields lie at $z \sim 2.4 - 2.7$, just above our redshift threshold of 2.35. At higher redshifts, it is increasingly likely that a Lyman limit system along some line-of-sight will wipe out any emergent ionizing flux,
leading to a greater number of non-detections (see Section~\ref{flux}).
Because of this, it is possible that including sources at the higher-redshift end of the range we probe here in the stacks could
``dilute" some underlying average signal from sources at the lower-redshift end.
To test this, we broke apart our stacks for the combined GOODS fields into three redshift bins, each with equal numbers of galaxies, and performed the same analysis. At F275W, we found nondetections in all three redshift bins. This suggests that the ``dilution" does not affect our ability to detect possible LyC-leaking galaxies at the depth of the HDUV F275W data. Meanwhile, the stacked detection in F336W remains fairly strong in all three redshift bins, dropping from SNR = 21.6 in the lowest-redshift bin to SNR = 5.7 in the highest, where most of the band now probes the rest-frame Lyman continuum.
In summary, we are left with (1) a handful of candidate LyC-leaking galaxies that appear to be individually detected at moderate significance (at the $\gtrsim3\sigma$ level), and (2) a general population of galaxies that are essentially invisible at LyC wavelengths, even in stacked images. This has some parallels to the results of the LACES program \citep{laces}, which used deep WFC3/UVIS imaging in F336W of the SSA22 field to search for $z \sim 3.1$ LyC leakers that are also Ly$\alpha$ emitters and Lyman break galaxies. While they successfully detect significant LyC flux in $\sim20\%$ of their targets, the majority of their sources (42 of 54) are faint in F336W (SNR $<$ 4). Stacking these non-detections yielded no net signal, with an upper limit of F336W = 31.8 (3$\sigma$ in a 0.12\arcsec aperture). This led the authors to conclude that detection of LyC emission in their sample is dichotomous, occurring either fairly strongly in individual sources or not at all.
However, we note two important differences in methodology that distinguish this work from \citet{laces}. First, their high success rate in identifying LyC leakers (for comparison, we have a $\sim6.5\%$ success rate in the GOODS-S---that is, three candidates out of 46 total sources with robust redshifts in the $z=2.35-3.05$ range and with UV coverage) can be attributed largely to (1) the depth of their F336W exposures compared to the typical depth of an HDUV pointing ($\sim$30 mag at 3$\sigma$ in a 0\farcs{12} diameter aperture in LACES vs. $\sim$27.5 mag at 5$\sigma$ in a 0\farcs{4} diameter aperture in HDUV), and (2) the higher throughput in F336W compared to F275W, which partially offsets the rising opacity of the IGM to LyC photons at $z > 3$. Second, their \textit{HST}/WFC3 followup was highly targeted, focusing on a set of sources that were already known to have high Ly$\alpha$ equivalent widths and large O32---both of which may correlate with LyC escape \citep{nakajima14,izo16a,izo16b,izo18,mich17,marchi18,stei18}. In contrast, the sample presented here is virtually blind to such selection effects, since we require only that a source be reasonably bright in F275W and that it have a spectroscopic redshift.
\begin{figure}[t]
\includegraphics[width=3.33in,trim={0.5cm 1cm 1cm 0},clip]{transmission_z_flux.pdf}
\caption{
F275W flux vs. redshift for the 129 galaxies with high-quality spectroscopic redshifts and F275W coverage in both GOODS fields.
The quasar in the GOODS-N (GN-UVC-1 from \citealt{jones18})
is marked in green with an upward pointing arrow, since its measured flux is much higher than the y-axis limit. The red and blue
squares are all the $>3\sigma$ detections, positive or negative, with the blue denoting
contamination by foreground galaxies.
Note that GN-UVC-6 is a $2.9\sigma$ source and hence not marked in red.
Thus, there are five $>3\sigma$ LyC-leaking galaxy candidates in the redshift range
$2.35 < z < 3.05$.
The thick black curve shows the transmission over this redshift range (right-hand y-axis scale),
which drops rapidly as one moves to higher redshifts.
The purple curve shows the F275W filter transmission at the LyC edge as a function
of redshift. The y-axis units are arbitrary.
Above $z=2.35$, we will only have LyC photons within the filter.
}
\label{transmission}
\end{figure}
\section{Contributions to the Metagalactic Ionizing Background}
\label{flux}
Determining the absolute escape fraction for each of our candidate LyC leakers is challenging, if not impossible, without reliable knowledge of each source's intrinsic spectral energy distribution (SED) and degree of reddening. Instead, we consider the extent to which our candidate sample contributes to the metagalactic ionizing background. To do so, we combine
the GOODS-N and the GOODS-S data. Excluding the spectroscopically
contaminated sources, we have a total of 129 SFGs together
with 9 AGNs based on the X-ray properties. We have identified six ofthese as candidate LyC leakers. These consist of a single quasar (the LyC-luminous source GN-UVC-1 from \citealt{jones18})
and the five SFGs described in Section~\ref{averaging}.
GN-UVC-1 dominates the flux in the two fields with a F275W flux of 2~$\mu$Jy, roughly twice the contribution of all of the SFGs even when including the candidate LyC leakers ($1.00 \pm 0.41$ $\mu$Jy). The remaining AGNs contribute only an insignificant total of 0.096~$\mu$Jy. Because of the sparseness of the luminous AGNs, the fields are too small to make an accurate estimate of the AGN contributions, and we focus solely on the SFGs.
We converted each SFG's F275W flux density to an ionizing volume emissivity, $\epsilon_{900}$, defined as the luminosity
density at 900 \AA~per unit frequency, divided by the comoving survey volume based on (1) the total survey area in
F275W (131 arcmin$^{2}$) and (2) the redshift bounds $2.35 < z < 3.05$.
We simplified our calculations by assuming that the entirety of each SFG's F275W flux falls at the filter's effective
wavelength of $\sim2714$~\AA, or rest-frame $\sim$750 \AA. We further assumed that the emergent SEDs of these
galaxies are flat in frequency space (i.e., \textit{f$_{\lambda} \propto \lambda^{\beta}$} with $\beta = -2$) both blueward
and redward of 912 \AA, albeit with different amplitudes on each side of the Lyman limit. While a shallower spectral slope
at $\lambda_{rest} < 912$ \AA~may be more plausible, the assumption of a spectrally flat LyC allows us to use our
measured flux densities ``as-is", without the need for scaling to 900 \AA, which would only increase our calculated
values of $\epsilon_{900}$. This assumption thus allows us to remain as conservative as possible when estimating
the ionizing emissivities.
Intergalactic transmission losses are substantial for the higher redshift galaxies seen in the F275W band, and
these losses need to be taken into account in computing the ionizing emissivity. Below the LyC wavelength,
the opacity is dominated by sources with $N(HI)$ column densities near the Lyman limit ($\log N(HI)=17.2$~cm$^{-2}$).
Because the number of absorbing systems is small, there is a wide variation in the mean flux in the F275W filter, after
intergalactic absorption, relative to the galaxy flux at rest-frame 900~\AA\ (e.g., \citealt{in14} and references therein).
The transmission also drops rapidly with increasing galaxy redshift as the intergalactic path length increases.
In order to compute the transmission correction, we ran Monte Carlo simulations,
assuming the number of $N(HI)$ systems per unit $N(HI)$ is a power law with index $-1.7$,
and the number of Lyman limit systems with $\tau > 1$ per unit $dz$ is given by the analytic
form $2.8\times ((1+z)/4.5)^{1.94}$ of \citet{song10}. In Figure~\ref{transmission}, we show
the mean flux in the F275W filter relative to the galaxy flux at rest-frame 900~\AA\
versus redshift, which drops from 0.37 at $z = 2.35$ to 0.065 at $z = 3$. Because of this
rapid drop in the mean transmission, we have chosen to divide the sample into two redshift
intervals in computing the corrected emissivity. We find a logarithmic value for $\epsilon_{900}$ of
$25.0\pm0.3$ erg ~s$^{-1}$~Hz$^{-1}$~Mpc$^{-3}$ in the $z = 2.35$ -- 2.7 interval
and $25.2\pm0.5$ erg ~s$^{-1}$~Hz$^{-1}$~Mpc$^{-3}$ in the $z = 2.7$ -- 3.05 interval.
Removing source 24639 from our calculation in the lower-redshift interval decreases the value of log$(\epsilon_{900})$ by about 0.15 dex.
\begin{figure}[t]
\centering
\includegraphics[trim={7pt 0 8pt 0},clip,width=0.9\linewidth]{EmissivityComparison_AllGOODS}
\caption{Ionizing volume emissivity at $z=2.35-2.7$ and $z=2.7-3.05$ estimated from our 5 LyC-leaking candidates
together with our stacking analysis of the remaining
$z=2.35-3.05$ galaxies in the combined HDUV-GOODS fields (filled gold stars).
Horizontal error bars on our data points indicate the redshift range.
To compare with literature observations of the ionizing emissivity from AGN, we show data points from \citet{cbt09} and \citet{mich17} (red triangles and black squares, respectively) and best-fit models from \citet{cbt09} and \citet{ hm12} (red and blue curves, respectively). For the contribution from SFGs, we show the data point from \citet{stei18} (open green diamond). Cyan circles show the total inferred metagalactic ionizing luminosity density from \citet{beck13}.}
\label{literature}
\end{figure}
The bulk of our ionizing flux ($67\%$) comes from the five SFGs, with the remaining galaxies contributing only a small fraction. This is consistent with the results of \citet{smith18,smith20}, who do not detect significant LyC emission in their own stacking analyses. This bimodality may be caused by favorable lines-of-sight through the ISM of the host galaxy (e.g., \citealt{cenkimm15}) and/or the IGM rather than by the intrinsic galaxy properties, with the small number of candidate LyC leakers corresponding to lines-of-sight with low IGM absorption. Even at $z=2.35$, 14\% of the lines-of-sight have $<10$\% transmission in the F275W filter, and only 30\% of the lines-of-sight have transmission above 50\%, while at $z=2.7$, 48\% of the lines-of-sight have $<10$\% transmission in the F275W filter, and only 5\% of the lines-of-sight have transmission above 50\%. This is consistent with the redshift distribution of the candidate LyC leakers (Figure~\ref{transmission}), four of which lie in the lower-redshift interval ($z=2.35-2.7$).
In Figure \ref{literature}, we summarize the results of our calculations and put them into context using other $z\sim3$ measurements of the ionizing background from the literature. The level of ionizing volume emissivity that we estimate from the SFGs is roughly consistent with other recent estimates, such as that of \cite{stei18} at slightly higher redshifts. It is above the AGN ionizing output measured by \citet{cbt09}, \citet{hm12}, and \citet{mich17} in large samples of quasars and broad-line AGNs at similar redshifts. \citet{beck13} used Ly$\alpha$ forest observations to infer the total ionizing background from $2 < z < 5$. At $z = 3.2$, they estimated a nominal log$(\epsilon_{900}) = 24.9$~erg~s$^{-1}$~Hz$^{-1}$~Mpc$^{-3}$, which is comparable to our estimates of $\epsilon_{900}$ for the SFG ionizing background at these redshifts. However, we emphasize that there may be further contamination in our sample by intervening sources that we have not identified. Hence, the present values should be considered as upper limits.
\section{Summary}
\label{summary}
We have presented a search for candidate LyC leakers at $z \sim 2.5-3$ in the GOODS-S field
using a combination of deep {\em HST}/WFC3 F275W imaging data from the HDUV survey \citep{oesch18}
and extensive optical/NIR spectroscopy. We found four candidate ionizing sources brighter than F275W $= 26$,
plus one additional source just below this cutoff with very blue rest-frame UV colors. However, two of these
five sources are contaminated by foreground lower redshift
galaxies, as seen in their optical spectra, leaving only three possible
candidate LyC leakers.
We performed an averaging analysis of all the sources with spectroscopic redshifts $2.35 < z < 3.05$
in the HDUV areas in both the GOODS-S and GOODS-N and found that the total ionizing output
of galaxies at these redshifts is dominated by just five individual candidate LyC leakers.
These include the three GOODS-S sources described above, and two sources in the GOODS-N
(one of which was presented in \citealt{jones18}).
Allowing for the very substantial effects of intergalactic absorption, we found that the volume ionizing
flux roughly matches that required to ionize the IGM at these redshifts (\citealt{beck13}) and is consistent
with other recent estimates.
\acknowledgements
We thank R. Bacon and the MUSE-Deep team for providing MUSE spectra of some of our LyC-leaking candidates
ahead of publication. Some figures in this work use colormaps from the Python package \texttt{CMasher} \citep{cmasher}.
We gratefully acknowledge support from NSF grant AST-1715145, the William F. Vilas Estate, and the Kellett Mid-Career Award from the University of Wisconsin-Madison Office of the Vice Chancellor for Research and Graduate Education with funding from the Wisconsin Alumni Research Foundation (A. J. B.). Based in part on data obtained at the W. M. Keck Observatory, which is operated as a scientific partnership among the California Institute of Technology, the University of California, and NASA and was made possible by the generous financial support of the W. M. Keck Foundation.
The authors wish to recognize and acknowledge the very significant cultural role and reverence that the summit of Mauna Kea has always had within the indigenous Hawaiian community. We are most fortunate to have the opportunity to conduct observations from this mountain.
\newcommand{A\&A}{A\&A}
|
\section{Introduction}
Understanding city-wide traffic patterns and detecting specific vehicle-related
events in real time has applications ranging from urban planning and
traffic-engineering to law enforcement. In such applications, Natural Language
(NL) descriptions can be one of the most convenient or the only way to interact
with a computer system, \eg taking as input an informal description provided by
a bystander. Thus, it becomes necessary to detect, track and retrieve vehicles,
potentially seen from multiple spatio-temporally disjoint viewpoints, based on
NL queries.
\input{figures/fig-star}
\input{figures/fig-nl-annotations}
Motivated by this demand, we define and address the task of Multi-Target
Multi-Camera (MTMC) by NL Description. The goal of this task is to track a
specific target described by an NL query within a sequence of multi-view video
in the form of a sequence of bounding boxes on each camera view. While it shares
similarities with other Vision+NL tasks like single object tracking (potentially
without distraction) by NL~\cite{feng2020real,feng2019robust,li2017tracking},
NL-based video retrieval~\cite{anne2017localizing,zhang2019man}, and
spatio-temporal localization by
NL~\cite{gavrilyuk2018actor,hu2016segmentation,yamaguchi2017spatio}, the
proposed task requires both temporal and spatial localization of the target from
a multi-view video via an NL query. Up until now, an added challenge in
developing and evaluating algorithms for the proposed task is the lack of
realistic datasets.
Therefore, we introduce CityFlow-NL, an open dataset designed to facilitate
research at the intersection of multi-object tracking, retrieval by NL
specification, and temporal localization of events. Our benchmark is derived
from CityFlow~\cite{tang2019cityflow}, which is itself a public dataset that has
been at the center of several recent workshop-challenges focused on MTMC
tracking and
re-identification~\cite{naphade2017nvidia,naphade20182018,naphade20192019,naphade20204th}.
As a result, the proposed CityFlow-NL dataset exhibits a diversity of real-world
parameters, including traffic density and patterns, viewpoints, ambient
conditions, \etc.
The NL descriptions are provided by at least three crowdsourcing workers, thus
capturing realistic variations and ambiguities that one could expect in such
application domains. Crowdsourcing workers describe the target vehicle using a
carefully designed multi-view annotation platform. The NL descriptions tend to
describe the vehicle color, the vehicle maneuver, the traffic scene, and
relations with other vehicles. Example NL descriptions and vehicle targets are
shown in Fig.~\ref{fig-nl-annotations}.
Although the motivation for building the CityFlow-NL dataset is to tackle the
real-world problem of MTMC by NL, the introduction of the CityFlow-NL also opens
up research possibilities for the retrieval/tracking by NL tasks to more
advanced topics, \eg action recognition by NL and spatio-temporal relation
understanding problems. In this paper, we propose two basic tasks: the vehicle
retrieval by NL description task and the vehicle tracking by NL description
task. We consider these tasks as a foundation for more advanced Vision+NL tasks,
including the MTMC by NL task.
Given a sequence of a \emph{single-view} video and an NL description that
uniquely describes a vehicle target in the video, the goal of the vehicle
tracking task is to predict a spatio-temporal localization in the form of a
sequence of bounding boxes for the target. The vehicle retrieval task is a
simpler form of the tracking task, in which the spatial localizations of all
candidate vehicles are given and the goal is to retrieve the target specified by
the NL query. These tasks are related to object tracking and retrieval systems
and can be evaluated in a similar way, but the NL query setup poses unique
challenges. For example, relationship expressions between targets
(Fig.~\ref{fig-nl-annotations}) and spatial referring expressions
(Fig.~\ref{fig-vtn-results}) have not been investigated in visual trackers.
\section{Related Works}
\label{sec-related-works}
In the past decade, researchers have started to look into exploiting natural
language understanding in computer vision tasks. These models usually
consist of two components: a language model and an appearance model to learn a
new feature space that is shared between both NL and visual
appearance~\cite{johnson2016densecap,vinyals2015show}. More recent object
detection and vision grounding models~\cite{hong2019learning,yang2019fast}
jointly exploit vision and NL using Siamese networks and depth-wise
convolutional neural networks.
The VisualGenome~\cite{krishna2017visual}, Flickr-30K~\cite{young2014image} and
Flickr-30K entities~\cite{plummer2015flickr30k} benchmarks facilitate research
in understanding language grounded in visual contents. These image datasets
provide detailed descriptions of regions in an image but still lack temporal
information that can enable systems to better handle the temporal context and
motion patterns in the NL descriptions.
OTB-99-LANG~\cite{li2017tracking} is the first open dataset that provides NL
descriptions for single object tracking, followed by LaSOT~\cite{fan2019lasot}
which is a large scale single object tracking benchmark annotated with NL
descriptions. These datasets enable researchers to perform spatial localization
of targets based on NL descriptions of the target objects. However, both of
these datasets are annotated with one NL description for the target object for
the entire sequence. The NL descriptions in OTB-99-LANG and LaSOT is limited to
a sentence or a phrase during annotation and can be ambiguous when used for
tracking. These NL descriptions are not able to describe the motion pattern of
the target~\cite{feng2020real}, especially over a longer time interval.
\input{figures/fig-distribution}
On the other hand, temporal localization of targets in videos using NL
descriptions has gained research interest with the introduction of the DiDeMo
dataset~\cite{anne2017localizing}. The DiDeMo dataset provides open domain
videos with NL descriptions for temporal moment. Each NL description refers to a
specific temporal moment and describes the main event in the video. However,
DiDeMo does not provide spatial localization of the target within a frame and
videos collected are intended for event recognition instead of tracking.
The spatial-temporal localization by NL description task is first introduced by
Yamaguchi \etal~\cite{yamaguchi2017spatio}. The ActivityNet dataset is annotated
with NL descriptions to facilitate the training and evaluation of the proposed
task. However, the temporal retrieval in \cite{yamaguchi2017spatio} entails
retrieving the target video clip from a set of video clips. On the contrary, the
goal of the proposed vehicle tracking and retrieval task is to temporally
localize the target object within one sequence of video. Additionally, the
targets in the ActivityNet-NL take up most of the frame and cannot serve as a
tracking benchmark. In this paper, we focus on building a multi-target tracking
with NL description benchmark. The NL descriptions can be used to localize
targets both temporally and spatially in multi-view videos, providing benchmarks
for multiple Vision+NL tasks.
\input{figures/fig-website}
\input{figures/fig-dist-attribute}
\section{CityFlow-NL Benchmark}
In this section, we describe the statistics of the proposed CityFlow-NL
benchmark and how we extend and annotate the CityFlow
Benchmark~\cite{tang2019cityflow}.
\subsection{Overview}
We extend the CityFlow Benchmark with Natural Language (NL) Descriptions for
each target vehicle. The CityFlow-NL makes it possible to build and evaluate
systems that can jointly leverage both language and visual modalities. The
proposed CityFlow-NL consists of 666 targets vehicles in 3,028 (single-view)
tracks from 40 calibrated cameras, and 5,289 unique NL descriptions. The average
number of frames a target vehicle shows in the CityFlow Benchmark is 75.85. The
distribution of the number of frames of target vehicles is shown in
Fig.~\ref{fig-distribution}.
A web based platform, shown in Fig.~\ref{fig-website}, is designed to collect NL
annotations. We provide multi-view video clips with red bounding boxes drawn
around each vehicle target and direct crowd-sourcing workers to give a detailed
NL description of the target vehicle that can be used to uniquely identify the
target vehicle from other vehicles. We collect annotations for each multi-view
track with at least three crowd-sourcing workers from Amazon SageMaker.
We split multi-view tracks from the CityFlow Benchmark by the timestamps of
cameras into multiple multi-view (or single-view) tracks that do not overlap
with each other temporally. For example, if a target shows up in camera 1 and
camera 2 in two temporal segments that do not overlap with each other, we
generate two separate tracks for our annotation process. For each multi-view
track, we clip video segments from all views with bounding boxes drawn around
the target vehicle.
\input{figures/tbl-comparison}
\subsection{Collecting Attributes of Target Vehicle}
Prior to a crowdsourcing worker generating an NL description (in the form of a declarative
sentence of an arbitrary length), the crowdsourcing worker is asked to tag the multi-view
video clip using predefined attributes. The distribution of these attributes are
summarized in Fig.~\ref{fig-dist-attribute}.
Performing such a preliminary task provides two benefits. First, it tends to
focus every crowdsourcing worker's attention on a common set of visual and linguistic
concepts while crowdsourcing worker remain free to create an NL description in any form.
Second, the attributes selected by crowdsourcing workers facilitate automatic verification
of an crowdsourcing worker's engagement with the task and consistency among the various
crowdsourcing workers of the same video clip. Since we direct crowdsourcing workers to
annotate NL descriptions in free form, it introduces uncertainty to the
annotation process and a direct verification of the annotated NL descriptions is
infeasible. Thus, a majority voting mechanism between annotations for the
attributes (color, motion, maneuver) from three different crowdsourcing workers
is used to automate the verification of the annotation and ensures the quality
of the NL descriptions. If no agreement between the three workers is reached,
the annotations are discarded and the target vehicle is re-annotated by another
three workers.
As was the case with the CityFlow dataset~\cite{tang2019cityflow}, the
per-target attributes are not considered part of our CityFlow-NL dataset. One
reason is that the designers of the CityFlow Benchmark made a decision to hide
attributes when releasing their challenge dataset for the AI City
Challenge~\cite{naphade2017nvidia}, in part to prevent challenge participants
from gaining unfair advantage. Another reason is that while attribute-based
tracking received attention in the past~\cite{tripathi2019tracking}, the
motivation for using attributes is not as apparent given strong NL models, and
the requirements of real-world scenarios. Indeed, as shown
in~\cite{feng2020real} NL-based tracking decisively outperforms
attribute-based tracking on a challenging benchmark. The attributes are not
equivalent to the NL descriptions, where additional information including
vehicle maneuver and relations to other targets are essential to a variety of
problems.
\subsection{Collecting NL Descriptions}
NL descriptions of the same target can be very different, as NL descriptions are
subjective and may focus on different aspects of the target. Examples of such
annotations are shown in Fig.~\ref{fig-nl-annotations}. The NL descriptions we
collect tend to describe the vehicle color, the vehicle maneuver, the traffic
scene, and the relation of the vehicle with other vehicles in the scene. We
specifically direct crowdsourcing workers to describe the intrinsic motion for
all maneuvers in the multi-view video clip. Moreover, some conflicting
descriptions are acceptable due to color distortions and view-angle differences
in these multi-view videos. Our majority vote mechanism identifies consensus
among annotators, given the natural variation in their NL descriptions.
\subsection{Comparison to Other NL+Vision Datasets}
We compare the CityFlow-NL to other NL annotated vision datasets and presented
in Fig.~\ref{fig-star} and Tbl.~\ref{tbl-comparison}. The CityFlow-NL is the
first dataset built for the MTMC by NL task.
Comparing to prior NL annotated video datasets, the proposed CityFlow-NL focuses
on building a dataset for multi-object tracking and keeps the unique challenges
for visual tracking as we discussed in Sec.~\ref{sec-related-works}. Each of the
NL descriptions in our dataset describes a specific target for the entire
duration in the multi-view video. Our annotation process avoids potential
observation and annotation biases from crowdsourcing workers for vehicle motion
by presenting multi-view GIFs, using multiple-choice questions for attributes
and a majority vote mechanism implemented during the annotation process.
\section{Vehicle Retrieval by NL Task}
We now define the Vehicle Retrieval by Natural Language Description Task and
present how we use the proposed CityFlow-NL Benchmark for it. We consider this
task as an essential first step towards tackling more advanced Vision+NL tasks
like the MTMC by NL description task.
\subsection{Task Definition}
For the purpose of the retrieval by NL task, we utilize the proposed CityFlow-NL
Benchmark in a \textit{single-view} setup, although the CityFlow-NL could be
used for retrieval tasks with multi-view tracks. For each single-view vehicle
track, we bundle it with a query that consists of three different NL
descriptions for training. During testing, the goal is to retrieve and rank
vehicles tracks based on NL queries.
This variation of the proposed CityFlow-NL contains 2,498 tracks of vehicles
with three unique NL descriptions each. Additionally, 530 unique vehicle tracks
together with 530 query sets (each annotated with three NL descriptions) are
curated for testing.
The proposed NL tracked-object retrieval task offers unique challenges versus
action recognition tasks and content-based image retrieval tasks. In particular,
different from prior content-based image retrieval
systems~\cite{guo2018dialog,hu2016natural,mao2016generation}, retrieval models
for this task need to consider the relation contexts of a vehicle track and
motion within the track as well. While the action recognition by NL description
task~\cite{anne2017localizing} temporally localizes a moment within a video, the
proposed tracked-object retrieval task requires both temporal and spatial
localization within a video.
\subsection{Evaluation Metrics}
The Vehicle Retrieval by NL Description task is evaluated using standard metrics
for retrieval tasks~\cite{manning2008introduction}. We use the Mean Reciprocal
Rank (MRR) as the main evaluation metric. Recall @ 5, Recall @ 10, and Recall @
25 are also evaluated for all models.
For each query in the testing split, we rank all 530 candidate tracks using the
retrieval system and evaluate the retrieval performance based on the
above-mentioned metrics.
\subsection{The Retrieval Baseline Model}
\label{sec-retrieval-baseline}
\input{figures/fig-baseline-model}
\input{figures/fig-baseline-results}
We build the baseline model to measure the similarities between a vehicle track
$\mathcal{T}$ and a NL description $Q$. Our baseline formulation is inspired by
content-based image retrieval systems~\cite{young2014image} and image captioning
models~\cite{johnson2016densecap,vinyals2015show}. The overview of the baseline
model is shown in Fig.~\ref{fig-baseline-model}.
The vehicle track is defined as a sequence of frames in a video clip.
$\mathcal{T} = \left\{F_1,\cdots,F_t\right\}.$ We use the ground truth bounding
box of the target vehicle to crop each frame and reshape the sequence of crops,
denoted as $C$, to the same shape of $(256, 256)$. This sequence of reshaped
frame crops of the target, $\left\{C_1, \cdots, C_t\right\}$, is used as the
visual representation of the vehicle. We build a Siamese embedding model for the
retrieval task, where each crop $C_i$ is embedded by a
ResNet-50~\cite{he2016identity} pretrained on
ImageNet~\cite{russakovsky2015imagenet}, denoted as $E_{C_i}$. We use a
pretrained BERT~\cite{devlin2018bert} to embed the NL description $Q$ into a 256
dimensional vector, denoted as $E_Q$.
The similarity between $C_i$ and $Q$ is defined as:
\begin{equation}
\mathcal{S}\left({C_i}, Q\right) = \exp\left(- \|E_{C_i} - E_Q\|_2\right).
\end{equation}
For the baseline retrieval model, both positive and negative pairs of vehicle
image crop $C_i$ and $Q$ are constructed for training. Negative pairs of $C_i$
and $Q$ are built by randomly selecting NL descriptions that describe other
targets. The baseline retrieval model is trained with a cross entropy loss and
an initial learning rate of 0.001 for 20 epochs on 2 GPUs using a stochastic
gradient descent optimizer.
For the purpose of inference, we measure the similarity between a test track and
a test query as the average of the similarities between all pairs of crops in
the track and NL description in the test query. \ie
\begin{equation}
\mathcal{S}\left(\mathcal{T}, \{Q_1, Q_2, Q_3\}\right) =
\frac{1}{3\cdot t}\sum_{j=1}^{3}\sum_{i=1}^{t}\mathcal{S}(C_i, Q_j).
\end{equation}
\subsection{Retrieval Results and Analysis}
For each test query, we compute this similarity on all test tracks and rank the
tracks for evaluation. The baseline model achieves an MRR of 0.0269, Recall @ 5
of 0.0264, Recall @ 10 of 0.0491, Recall @ 25 of 0.1113.
An example of a failure case from the baseline retrieval model is shown in
Fig.~\ref{fig-baseline-results}. In this specific case, positional referring
expressions (``in the right lane'') is given in the NL description, but the
baseline retrieval model does not consider the position of the target within
the frame. Moreover, it is worth noting that the baseline model does not
consider the motion pattern nor the context of vehicle tracks explicitly and the
retrieval performance could be further improved by measuring the motion within
the frame.
\input{figures/fig-baseline-tracking-model}
\section{Vehicle Tracking by NL Task}
We now describe how we use the proposed CityFlow-NL Benchmark for the vehicle
tracking by NL description task.
\subsection{Problem Definition}
We further investigate the potential of the proposed retrieval model and extend
it to the tracking by NL task. With the same query as in the retrieval task,
together with a sequence of frames $(F_1, \cdots, F_T)$ as inputs, the goal is
to track the specific target the NL query refers to in the form of a sequence of
bounding boxes. For each frame, the tracker should either predict the bounding
box for the target or predict that the target is not present in the frame.
\subsection{Evaluation Metrics}
The vehicle tracking by NL description task is evaluated, following popular
object tracking protocols, using the Area Under Curve (AUC) of the Success Rate
vs. Intersection over Union (IoU) thresholds and Normalized
Precision~\cite{wu2013online}.
The success is defined as 0 for frames where the prediction is ``not present''
while the ground truth is available, or the ground truth is ``not present'' and
a prediction is made. For frames with both a ground truth bounding box and a
prediction, the success is measured by computing the IoU between the ground
truth bounding box and the prediction and comparing this against a predefined
threshold. The AUC of the success rate at different IoU thresholds is reported
as the performance for each tracker~\cite{wu2013online}.
\subsection{The Tracking Baseline Model}
Intuitively, to perform tracking of the target vehicle using NL descriptions, we
use our baseline retrieval model introduced in Sec.~\ref{sec-retrieval-baseline}
for inference on tracks predicted by multi-object trackers. The baseline model
is shown in Fig.~\ref{fig-baseline-tracking-model}.
The tracking baseline model consists of three stages. The first stage of the
model is an object detector which localizes all vehicles in every frame. In the
second stage, we use a generic multi-object tracker to predict vehicle tracks on
the test split of the CityFlow-NL. In the last stage, we use the baseline
retrieval model introduced in Sec.~\ref{sec-retrieval-baseline} to rank the
predicted vehicle tracks for each query. The vehicle track with the highest
average similarity score is returned by the baseline tracker as the prediction.
The baseline vehicle tracking model first localizes and tracks all vehicles in a
video and predicts a retrieval score for each vehicle track in the second pass.
The time complexity of retrieval baselines model is $O(M\times N)$, where $M$ is
the size of the query set and $N$ is the number of tracks predicted by the
multi-object tracker.
In our experiments, we tested our baseline tracker with three different
multi-object trackers: DeepSORT~\cite{wojke2017simple},
MOANA~\cite{tang2019moana}, and TNT~\cite{hsu2019multi} with three different
sets of vehicle detections: MaskRCNN~\cite{he2017mask},
YOLO3~\cite{redmon2018yolov3}, and SSD512~\cite{liu2016ssd}. We use the same
model we trained for the retrieval task for the tracking task. Results are
summarized in Tbl.~\ref{tbl-tracking-results}. Trackers using SSD detections are
outperforming others with a higher recall rate. Theses baseline trackers are
competitive with each other.
\input{figures/fig-vtrm}
\subsection{Vehicle Tracking Network (VTN)}
Instead of the two-stage multi-object tracking and NL retrieval approach, we now
introduce a more efficient and accurate Vehicle Tracking Network (VTN), which is
summarized in Fig.~\ref{fig-vtrm}. Similar to prior works on single object
tracking with NL description~\cite{feng2020real,feng2019robust,li2017tracking},
we design the VTN to be a one-shot tracker by detection that is adapted to
utilize an NL description. The VTN first extracts the visual features of a
frames $F_t$ using a Feature Pyramid Network (FPN)~\cite{lin2017feature}. The
visual features are then used in two separate branches: a Vehicle Presence
Network (VPN) and a Vehicle Localization Network (VLN).
The VPN, shown in Fig.~\ref{fig-vpn}, indicates the target's presence in the
given frame to initiate or terminate a track. The vehicle tracking by NL
description task differs from single object tracking by NL description
task~\cite{li2017tracking} as the presence of the target is not guaranteed for
most of the frames. Thus, a cross-correlation (XCorr)
operation~\cite{bertinetto2016fully} between the NL representation and the
visual features from the four stages of the FPN (``P2'' thru ``P5'') is used to
localize the target temporally. A weighted average trained between the four
presence features is then used to compute the output presence feature, denoted
as $P$. A softmax function between the two elements in $P$ is used to determine
the presence probability $\hat{P}$ of the vehicle target based on the NL query,
\ie
\begin{equation}
\hat{P} = \frac{e^{P_1}}{e^{P_0} + e^{P_1}}.
\end{equation}
We use $P^*$ to denote the ground truth presence of the target, where $P^* = 1$
represents the presence and $P^* = 0$ represents that the target is not present
in the frame. A binary cross entropy loss is used to train the presence branch:
\begin{equation}
\Loss_\text{Presence} = - \hat{P}\cdot \log P^* - (1-\hat{P})\cdot \log (1- P^*).
\end{equation}
\input{figures/fig-vpn}
\input{figures/tbl-tracking-results}
\input{figures/fig-vtn-results}
The VLN is a modified version of the Faster-RCNN~\cite{ren2015faster}, which
proposes regions and localizes the vehicle target based on the NL query. We use
a standard Region Proposal Network (RPN) with anchor sizes of 32, 64, 128, 256,
512 and anchor ratios of 0.5, 1.0 and 2.0 for generating a set of region
proposals. The NL Region of Interest (RoI) Head not only predicts the regression
and classification as in a standard object detection RoI head, it also measures
similarities between visual features for each region proposals and the NL
representation, denoted as $\hat{S}$. The NL similarity layers are trained with
a cross entropy loss that is similar to the classification loss of the RoI loss:
\begin{equation}
\Loss_\text{NL} = \sum_\text{all regions} - \hat{S} \log S^* - (1- \hat{S}) \log (1-S^*).
\end{equation}
where $S^*$ is the ground truth similarity. When curating the ground truth
labels for training the RPN and NL RoI Head, all vehicles in the frame are
marked as positive. However, the target similarity $S^*=1$ if and only if the
region's target regression ground truth bounding box is exactly what the NL
description refers to.
Thus, the loss for the localization branch is:
\begin{equation}
\Loss_\text{Localization} = \Loss_\text{RPN} + \Loss_\text{RoI} + \Loss_\text{NL},
\label{eq-loss-rpn}
\end{equation}
where $\Loss_\text{RPN}$ and $\Loss_\text{RoI}$ is the same as
in~\cite{ren2015faster}. The two branches in the VTN are jointly trained
end-to-end with the above-mentioned losses:
\begin{equation}
\Loss_\text{VTN} = \Loss_\text{Presence} + \Loss_\text{Localization}.
\end{equation}
The VTN is initialized with pretrained weights of a Faster-RCNN with FPN trained
for MSCOCO. We train the RPN and RoI Head with an initial learning rate of 0.001
and the VPN with an initial learning rate of 0.01. The learning rate is decayed
by 0.5 every 2,000 steps with a batch size of 16 frames distributed across 4
GPUs using a stochastic gradient descent optimizer.
When used during inference, we use the VPN with a threshold of 0.5 that is
validated on the training split to indicate if the target is present in a frame.
If the target is found in the frame, the VLN is then used to predict the
similarities between detected vehicle bounding boxes and the NL query. We apply
a sub-window attention~\cite{feng2019robust} to the similarities and pick the
highest scored one as the prediction for the frame.
Results for the vehicle tracking task are presented in
Tbl.~\ref{tbl-tracking-results}. The VTN outperforms baseline trackers and runs
at 20 fps on a single GPU. As the VTN performs the NL retrieval at an earlier
step during the detection phase, the recall of the targets is improved compared
to the baseline trackers, which perform the NL retrieval at the last step after
multi-object tracking. A failure case of the VTN is shown in
Fig.~\ref{fig-vtn-results}. As the VTN is designed as an online tracker by NL
detection and predicts the highest scored bounding box when the VPN predicts
that the target is present, the false positive rate is higher than the baseline
tracking model, especially when contrasting vehicle targets are present.
\section{Conclusion}
This paper has presented CityFlow-NL, the first city-scale multi-target
multi-camera tracking with NL descriptions dataset that provides precise NL
descriptions for multi-view ground truth vehicle tracks. Our NL annotations can
be used to benchmark tasks like vehicle tracking and retrieval by NL, motion
pattern analysis with NL, and MTMC with NL descriptions. We introduce two basic
benchmark tasks with multiple baseline models utilizing the proposed
CityFlow-NL: The Vehicle Retrieval by NL task and The Vehicle Tracking by NL
task, facilitating future research on more advanced Vision+NL tasks and systems.
{\small
\bibliographystyle{ieee}
|
\section{Introduction}\label{sec1}
The search for stable configurations for compact objects play an important role in contemporary Astrophysics. There is a great variety of theoretically proposed objects of the most diverse natures \cite{MM2004}, \cite{jetzer1992},~ \cite{2008PhR...467..117S}, \cite{2005ICRC....3..125M} which continue being studied. Boson stars are ones of the most widely accepted configurations assumed for exploring the possibility of the existence of such compact objects. In the article \cite{cabo2020} a static boson star configuration was found solving the EKG equations for a real scalar field, with spherical symmetry being coupled with matter. These solutions were found numerically by introducing a source being proportional to the energy density of the matter and looking for the value of the scalar field at the origin that would determine a Yukawa potential at large distance, for the scalar field. A main result of that work is to show that the inclusion of the interaction of the scalar field with matter allows for the existence of static solutions of the EKG equations, a property that these equations lack \cite{jetzer1992}.\\
The present work is composed of the following parts, first we present the system of equations that were solved in our previous paper and that determine the boson star solution. In second place, an investigation of the stability criterion is carried out. The relation between matter energy density at the origin and the scalar field also in the origin, which determine the existence of the solutions, is evaluated . Finally, we find the TM vs MED in the origin plot which combined with the weak energy condition, results in two possible ranges for stable solutions of the configuration.
\section{The field equations}
Consider a metric defined by the following squared interval and
coordinates%
\begin{flalign}
&ds^{2}=\mathit{v}(\rho){dx^{o}}^{2}-u(\rho)^{-1}d\rho^{2}-\rho
^{2}(sin^{2}\theta\text{ }d\varphi^{2}+d\theta^{2}),\\
&x^{0}=c\text{ }t\text{, \ \ \ }x^{1}=\rho,\\
&x^{2}=\varphi,\text{ \ }x^{3}\equiv\theta,
\end{flalign}
and an energy-momentum tensor for a real scalar field in interaction with matter:%
\begin{align}
T_{\mu}^{\nu} & =-\frac{\delta_{\mu}^{\nu}}{2}(g^{\alpha\beta}{\Phi
}_{,\alpha}{\Phi}_{,\beta}+m^{2}{\Phi}^{2}+2\text{ }J(\rho)\text{ }%
\Phi)\nonumber\\
& \text{ \ \ \ }+g^{\alpha\nu}{\Phi}_{,\alpha}{\Phi}_{,\mu}+P\text{ }%
\delta_{\mu}^{\nu}+u^{\nu}u_{\mu}(P+e).
\end{align}
After making several coordinate changes as detailed in \cite{cabo2020}, the EKG system can be reached:
\begin{subequations}
\begin{align}
\begin{split}
&{\frac{u^{\prime}(r)}{r}}-{\frac{1-u(r)}{r^{2}}} =-\frac{1}{2}%
(u(r)\phi^{\prime}{(r)}^{2}+\phi(r)^{2} \\
&+2j(r)\phi(r))-e(r),\label{ec1}\\
\end{split}\\
\begin{split}
&\frac{u(r)}{v(r)}\frac{v^{\prime}(r)}{r}-{\frac{1-u(r)}{r^{2}}} =-\frac
{1}{2}(-u(r){\phi}^{\prime}{(r)}^{2}+\phi(r)^{2}\\
&+2j(r)\phi(r))+p(r),\label{ec2}\\
\end{split}\\
\begin{split}
&p^{\prime}(r)+(e(r)+p(r))\frac{v^{\prime}(r)}{2v(r)}-\phi(r)\text{
}j^{\prime}(r) =0,\label{ec3}\\
\end{split}\\
\begin{split}
&j(r)+{\phi}(r)-u(r)\text{ }{\phi}^{\prime\prime}(r)={\phi}^{\prime
}(r){\Large (}\frac{u(r)+1}{r}-r\text{ }(\frac{{\phi}(r)^{2}}{2}\\
&+j(r){\phi}(r)+\frac{e(r)-p(r)}{2}){\Large )},\label{ec4}
\end{split}
\end{align}
\label{straincomponent}
\end{subequations}
\index{non}
where $\phi(r)$ is the scalar field, $j(r)$ is the source and $e(r)$, $p(r)$ the energy density and the matter pressure respectively. Note that we used the Bianchi identity too for completing the system of the equations. So, in general the set of equations is composed as fallow: The two first equations are the Einstein one referent to the two first component of the metric $v(r)$ and $ u(r)$. The other one, \eqref{ec3} is the dynamic equation for the energy, the pressure and the scalar field, and it substitutes the two Einstein
equations that are associated to both angular directions. This relation results from the Bianchi identity. Finally, \eqref{ec4} is the Klein-Gordon equation for the scalar fields.
This system describes the dynamics of the space time for a configuration of scalar field and matter coupled through the source $j(r)$ .\\ Let us fix the boundary conditions at a very small value $\delta = 10^{-6}$ around the origin in the following way:
\begin{align}
u(\delta) & =1,\\
v(\delta) & =1,\\
\phi(\delta) & =\phi_{0},\\
e(\delta) & = e_{0},
\end{align}
and assume a linear relationship between the source and the matter energy density as well as between the pressure and the same matter energy density as
$$j(r) = g ~e(r),$$ $$p(r) = n ~e(r).$$
In the work \cite{cabo2020} the solutions were determined by setting a value of the energy density of the matter at the origin and determining the value of the scalar field also at the origin in such a way that this field takes the form of a Yukawa potential at large radius. In this way a whole family of regular solutions can be found.
\subsection{Stability criterion}
Now we will impose the restrictions placed by the weak energy condition and the fact that the total mass of the system must increase when the value of the energy density is also increased at the origin \cite{shapiro}. These criteria will be examined here for the entire family of solutions, determining those that meet them. These conditions will be easily checked if we start from considering the formula for the total mass
\begin{align}
M(e_{0}) & =\int dr\text{ }4\pi\text{ }r^{2}\text{ }e
_{t}(r),\\
e_{t}(r) & =\frac{1}{2}(u(r)\phi^{\prime}{(r)}^{2}+\phi
(r)^{2}+2j(r)\phi(r))+e(r).\nonumber
\end{align}
\begin{figure}[h!]
\centerline{\includegraphics[width=.8\linewidth]{./GpH/Grafico3.pdf}}
\caption{Relation between matter energy density and the scalar field at the origin~($\phi_{0}$) needful for obtaining plausible solutions. The relation was found fixing $g = 0.9$, ~$n = 0.025$. \label{campoSvseo}}
\end{figure}
As stated in \cite{cabo2020} the value of the initial field at the origin is a function of the matter energy density at the origin. In other words, depending on the density of matter that the object in question has, it will need a specific value of scalar field at the origin in order to show a Yukawa like behaviour of the scalar field in the faraway regions. This relationship was determined using Mathematica software version 11.1 and the results are shown in the table \ref{valores}. The function $\phi_{0}(e_ {0})$ defined by the table is shown in the figure \ref{campoSvseo}.
\begin{center}
\begin{table}[h!]%
\centering
\caption{The scalar field and total mass values corresponding to the fixed matter energy density at the origin.\label{valores}}%
\begin{tabular*}{20pc}{@{\extracolsep\fill}lcc@{\extracolsep\fill}}%
\textbf{$e_{0}$} & \textbf{$\phi_{0}$} & \textbf{$M|_{r=10}$} \\
0.3 & 0.142664736127 & 16.7838 \\
0.5 & 0.215184567562 & 16.9132 \\
0.8 & 0.309367889200 & 17.149\\
0.9 & 0.337735444700 & 17.1977 \\
1.0 & 0.364823000000 & 17.2341\\
1.1 & 0.390743919420 & 17.2535 \\
1.3 & 0.439437983840 & 17.2678 \\
1.5 & 0.484445641130 & 17.2546 \\
2.0 & 0.583890263504 & 17.1616 \\
4.0 & 0.867447339143 & 16.7466 \\
5.0 & 0.969939323350 & 16.6329\\
6.0 & 1.057176196658 & 16.5746\\
6.5 & 1.096426605833 & 16.5623 \\
7.0 & 1.133275616740 & 16.5595 \\
7.5 & 1.168018636795 & 16.5651\\
8.0 & 1.200900067643 & 16.578\\
8.1 & 1.207271535456 & 16.5814 \\
8.15& 1.210432922380 & 16.5832 \\
8.2 & 1.21357829962 & 16.5851 \\
8.3 & 1.219821748053 & 16.589 \\
8.5 & 1.232124212860 & 16.5976 \\
9.0 & 1.261863522592 & 16.6232\\
10.0& 1.317454277700 & 16.6899\\
\end{tabular*}
\end{table}
\end{center}
Note how that, if the MED at the origin increases the scalar field needful to obtain a plausible solution increases too. Doing a fit for the set of points in (table\ref{valores}[1-2]) we obtain that the functional relation between scalar field and the MED can be expressed as
$$\phi_{0}[e_{0}]=1.86882\sqrt[5]{e_{0} + 0.536945}-1.66797.$$
\section{Stability Regions}
Firstly, after solving the \eqref{ec1}-\eqref{ec4} system of equations and fixing the values of the MED at the origin we find the value of the scalar field at the origin, such that the field behaves like a Yukawa potential for large distances.
\begin{figure}[h!]
\centerline{\includegraphics[width=.8\linewidth]{./GpH/MvseoZonas.pdf}}
\caption{ Plot of the total mass vs matter energy density at the origin for the values $g = 0.9$ ,~$n = .025$. The dark zones are the stable ones. Note how in spite of having a positive slope for $e_0$ values over $8.15$ the system is unstable because the total energy density starts to have a negative region.\label{MvsEo}}
\end{figure}
Afterwards, by integrating the total energy from $r = \delta$ to $r = 10$, at which the total energy density is approximately close to zero, we find the total mass vs MED curve as shown in figure \ref{MvsEo}. \\
\begin{figure}[h!]
\centerline{\includegraphics[width=.8\linewidth]{./GpH/ETvsRT.pdf}}
\caption{Behavior of the total energy density with the radial distance from the origin for different values of $e_{0}$ .\label{ETvsRT}}
\end{figure}
\begin{figure}[h!]
\centerline{\includegraphics[width=.8\linewidth]{./GpH/ETvsR.pdf}}
\caption{Behavior of the total energy density with the radial distance from the origin around the value $e_{0}=8.16$.\label{ETvsR}}
\end{figure}
Note that the total mass of the system initially increases with the corresponding increase in the MED at the origin, until it reaches a maximum of $ 17.2678$ at the point $ 1.3 $. From this point it begins to decrease until it reaches a minimum of $ 16.5595 $ in $ e_{0} = 7, $ where it starts to increase again, this time with a lower slope. Then, the slope remains positive for the intervals $ e_{0} <1.3 $ and $ e_{0}> 7 $. Let us analyze the energetic conditions in such ranges. Calculating the total energy density $e_{t}$ for few values of $e_0$, we obtained the plots shown in the figure \ref{ETvsRT}. As it can be observed, for certain values of the energy density of matter at the origin, the total energy begins to have negative values \footnote{For a better illustration, the graphs were only taken for certain values of $ e_{0} $.}. The value of the total energy density at which the curves begin to have a negative region was determined graphically to be $ e_{0} = 8.15 $ in figure \ref{ETvsR}. Also, there are regions where, despite the fact that the derivative of the total mass with respect to the energy is positive, the physical system is not meeting the weak energy condition. Hence, for the system \eqref{ec1}-\eqref{ec4} the stability regions $ e_ {0} <1.3 $ and $ 7 <e_ {0} <8.15 $ result, which are shown as the shaded areas in figure \ref{MvsEo}.
\section{Conclusions}\label{sec5}
We study the stability of the static solutions of the EKG equations for a real scalar field with spherical symmetry and including matter. This analysis was carried out under two simple criteria: The weak energy condition and the need for growth of the mass with the increase in the energy density at the origin \cite{shapiro}. Under these criteria and for a specific configuration described in detail in \cite{cabo2020}, two stability zones were determined, as well as the relationship between the scalar field and the matter energy density at the origin. The analysis will be extended for other values of the pressure and energy relationships as well as for the coupling between the scalar field and matter. In particular, the plans also include studying polytropic relationships.
\section*{Acknowledgments}
The authors acknowledge the partial support of the Office of External Activities of ICTP (OEA), through
the Network on Quantum Mechanics, Particles and Fields (Net-09).
\section{Introduction}\label{sec1}
The search for stable configurations for compact objects play an important role in contemporary Astrophysics. There is a great variety of theoretically proposed objects of the most diverse natures \cite{MM2004}, \cite{jetzer1992},~ \cite{2008PhR...467..117S}, \cite{2005ICRC....3..125M} which continue being studied. Boson stars are ones of the most widely accepted configurations assumed for exploring the possibility of the existence of such compact objects. In the article \cite{cabo2020} a static boson star configuration was found solving the EKG equations for a real scalar field, with spherical symmetry being coupled with matter. These solutions were found numerically by introducing a source being proportional to the energy density of the matter and looking for the value of the scalar field at the origin that would determine a Yukawa potential at large distance, for the scalar field. A main result of that work is to show that the inclusion of the interaction of the scalar field with matter allows for the existence of static solutions of the EKG equations, a property that these equations lack \cite{jetzer1992}.\\
The present work is composed of the following parts, first we present the system of equations that were solved in our previous paper and that determine the boson star solution. In second place, an investigation of the stability criterion is carried out. The relation between matter energy density at the origin and the scalar field also in the origin, which determine the existence of the solutions, is evaluated . Finally, we find the TM vs MED in the origin plot which combined with the weak energy condition, results in two possible ranges for stable solutions of the configuration.
\section{The field equations}
Consider a metric defined by the following squared interval and
coordinates%
\begin{flalign}
&ds^{2}=\mathit{v}(\rho){dx^{o}}^{2}-u(\rho)^{-1}d\rho^{2}-\rho
^{2}(sin^{2}\theta\text{ }d\varphi^{2}+d\theta^{2}),\\
&x^{0}=c\text{ }t\text{, \ \ \ }x^{1}=\rho,\\
&x^{2}=\varphi,\text{ \ }x^{3}\equiv\theta,
\end{flalign}
and an energy-momentum tensor for a real scalar field in interaction with matter:%
\begin{align}
T_{\mu}^{\nu} & =-\frac{\delta_{\mu}^{\nu}}{2}(g^{\alpha\beta}{\Phi
}_{,\alpha}{\Phi}_{,\beta}+m^{2}{\Phi}^{2}+2\text{ }J(\rho)\text{ }%
\Phi)\nonumber\\
& \text{ \ \ \ }+g^{\alpha\nu}{\Phi}_{,\alpha}{\Phi}_{,\mu}+P\text{ }%
\delta_{\mu}^{\nu}+u^{\nu}u_{\mu}(P+e).
\end{align}
After making several coordinate changes as detailed in \cite{cabo2020}, the EKG system can be reached:
\begin{subequations}
\begin{align}
\begin{split}
&{\frac{u^{\prime}(r)}{r}}-{\frac{1-u(r)}{r^{2}}} =-\frac{1}{2}%
(u(r)\phi^{\prime}{(r)}^{2}+\phi(r)^{2} \\
&+2j(r)\phi(r))-e(r),\label{ec1}\\
\end{split}\\
\begin{split}
&\frac{u(r)}{v(r)}\frac{v^{\prime}(r)}{r}-{\frac{1-u(r)}{r^{2}}} =-\frac
{1}{2}(-u(r){\phi}^{\prime}{(r)}^{2}+\phi(r)^{2}\\
&+2j(r)\phi(r))+p(r),\label{ec2}\\
\end{split}\\
\begin{split}
&p^{\prime}(r)+(e(r)+p(r))\frac{v^{\prime}(r)}{2v(r)}-\phi(r)\text{
}j^{\prime}(r) =0,\label{ec3}\\
\end{split}\\
\begin{split}
&j(r)+{\phi}(r)-u(r)\text{ }{\phi}^{\prime\prime}(r)={\phi}^{\prime
}(r){\Large (}\frac{u(r)+1}{r}-r\text{ }(\frac{{\phi}(r)^{2}}{2}\\
&+j(r){\phi}(r)+\frac{e(r)-p(r)}{2}){\Large )},\label{ec4}
\end{split}
\end{align}
\label{straincomponent}
\end{subequations}
\index{non}
where $\phi(r)$ is the scalar field, $j(r)$ is the source and $e(r)$, $p(r)$ the energy density and the matter pressure respectively. Note that we used the Bianchi identity too for completing the system of the equations. So, in general the set of equations is composed as fallow: The two first equations are the Einstein one referent to the two first component of the metric $v(r)$ and $ u(r)$. The other one, \eqref{ec3} is the dynamic equation for the energy, the pressure and the scalar field, and it substitutes the two Einstein
equations that are associated to both angular directions. This relation results from the Bianchi identity. Finally, \eqref{ec4} is the Klein-Gordon equation for the scalar fields.
This system describes the dynamics of the space time for a configuration of scalar field and matter coupled through the source $j(r)$ .\\ Let us fix the boundary conditions at a very small value $\delta = 10^{-6}$ around the origin in the following way:
\begin{align}
u(\delta) & =1,\\
v(\delta) & =1,\\
\phi(\delta) & =\phi_{0},\\
e(\delta) & = e_{0},
\end{align}
and assume a linear relationship between the source and the matter energy density as well as between the pressure and the same matter energy density as
$$j(r) = g ~e(r),$$ $$p(r) = n ~e(r).$$
In the work \cite{cabo2020} the solutions were determined by setting a value of the energy density of the matter at the origin and determining the value of the scalar field also at the origin in such a way that this field takes the form of a Yukawa potential at large radius. In this way a whole family of regular solutions can be found.
\subsection{Stability criterion}
Now we will impose the restrictions placed by the weak energy condition and the fact that the total mass of the system must increase when the value of the energy density is also increased at the origin \cite{shapiro}. These criteria will be examined here for the entire family of solutions, determining those that meet them. These conditions will be easily checked if we start from considering the formula for the total mass
\begin{align}
M(e_{0}) & =\int dr\text{ }4\pi\text{ }r^{2}\text{ }e
_{t}(r),\\
e_{t}(r) & =\frac{1}{2}(u(r)\phi^{\prime}{(r)}^{2}+\phi
(r)^{2}+2j(r)\phi(r))+e(r).\nonumber
\end{align}
\begin{figure}[h!]
\centerline{\includegraphics[width=.8\linewidth]{./GpH/Grafico3.pdf}}
\caption{Relation between matter energy density and the scalar field at the origin~($\phi_{0}$) needful for obtaining plausible solutions. The relation was found fixing $g = 0.9$, ~$n = 0.025$. \label{campoSvseo}}
\end{figure}
As stated in \cite{cabo2020} the value of the initial field at the origin is a function of the matter energy density at the origin. In other words, depending on the density of matter that the object in question has, it will need a specific value of scalar field at the origin in order to show a Yukawa like behaviour of the scalar field in the faraway regions. This relationship was determined using Mathematica software version 11.1 and the results are shown in the table \ref{valores}. The function $\phi_{0}(e_ {0})$ defined by the table is shown in the figure \ref{campoSvseo}.
\begin{center}
\begin{table}[h!]%
\centering
\caption{The scalar field and total mass values corresponding to the fixed matter energy density at the origin.\label{valores}}%
\begin{tabular*}{20pc}{@{\extracolsep\fill}lcc@{\extracolsep\fill}}%
\textbf{$e_{0}$} & \textbf{$\phi_{0}$} & \textbf{$M|_{r=10}$} \\
0.3 & 0.142664736127 & 16.7838 \\
0.5 & 0.215184567562 & 16.9132 \\
0.8 & 0.309367889200 & 17.149\\
0.9 & 0.337735444700 & 17.1977 \\
1.0 & 0.364823000000 & 17.2341\\
1.1 & 0.390743919420 & 17.2535 \\
1.3 & 0.439437983840 & 17.2678 \\
1.5 & 0.484445641130 & 17.2546 \\
2.0 & 0.583890263504 & 17.1616 \\
4.0 & 0.867447339143 & 16.7466 \\
5.0 & 0.969939323350 & 16.6329\\
6.0 & 1.057176196658 & 16.5746\\
6.5 & 1.096426605833 & 16.5623 \\
7.0 & 1.133275616740 & 16.5595 \\
7.5 & 1.168018636795 & 16.5651\\
8.0 & 1.200900067643 & 16.578\\
8.1 & 1.207271535456 & 16.5814 \\
8.15& 1.210432922380 & 16.5832 \\
8.2 & 1.21357829962 & 16.5851 \\
8.3 & 1.219821748053 & 16.589 \\
8.5 & 1.232124212860 & 16.5976 \\
9.0 & 1.261863522592 & 16.6232\\
10.0& 1.317454277700 & 16.6899\\
\end{tabular*}
\end{table}
\end{center}
Note how that, if the MED at the origin increases the scalar field needful to obtain a plausible solution increases too. Doing a fit for the set of points in (table\ref{valores}[1-2]) we obtain that the functional relation between scalar field and the MED can be expressed as
$$\phi_{0}[e_{0}]=1.86882\sqrt[5]{e_{0} + 0.536945}-1.66797.$$
\section{Stability Regions}
Firstly, after solving the \eqref{ec1}-\eqref{ec4} system of equations and fixing the values of the MED at the origin we find the value of the scalar field at the origin, such that the field behaves like a Yukawa potential for large distances.
\begin{figure}[h!]
\centerline{\includegraphics[width=.8\linewidth]{./GpH/MvseoZonas.pdf}}
\caption{ Plot of the total mass vs matter energy density at the origin for the values $g = 0.9$ ,~$n = .025$. The dark zones are the stable ones. Note how in spite of having a positive slope for $e_0$ values over $8.15$ the system is unstable because the total energy density starts to have a negative region.\label{MvsEo}}
\end{figure}
Afterwards, by integrating the total energy from $r = \delta$ to $r = 10$, at which the total energy density is approximately close to zero, we find the total mass vs MED curve as shown in figure \ref{MvsEo}. \\
\begin{figure}[h!]
\centerline{\includegraphics[width=.8\linewidth]{./GpH/ETvsRT.pdf}}
\caption{Behavior of the total energy density with the radial distance from the origin for different values of $e_{0}$ .\label{ETvsRT}}
\end{figure}
\begin{figure}[h!]
\centerline{\includegraphics[width=.8\linewidth]{./GpH/ETvsR.pdf}}
\caption{Behavior of the total energy density with the radial distance from the origin around the value $e_{0}=8.16$.\label{ETvsR}}
\end{figure}
Note that the total mass of the system initially increases with the corresponding increase in the MED at the origin, until it reaches a maximum of $ 17.2678$ at the point $ 1.3 $. From this point it begins to decrease until it reaches a minimum of $ 16.5595 $ in $ e_{0} = 7, $ where it starts to increase again, this time with a lower slope. Then, the slope remains positive for the intervals $ e_{0} <1.3 $ and $ e_{0}> 7 $. Let us analyze the energetic conditions in such ranges. Calculating the total energy density $e_{t}$ for few values of $e_0$, we obtained the plots shown in the figure \ref{ETvsRT}. As it can be observed, for certain values of the energy density of matter at the origin, the total energy begins to have negative values \footnote{For a better illustration, the graphs were only taken for certain values of $ e_{0} $.}. The value of the total energy density at which the curves begin to have a negative region was determined graphically to be $ e_{0} = 8.15 $ in figure \ref{ETvsR}. Also, there are regions where, despite the fact that the derivative of the total mass with respect to the energy is positive, the physical system is not meeting the weak energy condition. Hence, for the system \eqref{ec1}-\eqref{ec4} the stability regions $ e_ {0} <1.3 $ and $ 7 <e_ {0} <8.15 $ result, which are shown as the shaded areas in figure \ref{MvsEo}.
\section{Conclusions}\label{sec5}
We study the stability of the static solutions of the EKG equations for a real scalar field with spherical symmetry and including matter. This analysis was carried out under two simple criteria: The weak energy condition and the need for growth of the mass with the increase in the energy density at the origin \cite{shapiro}. Under these criteria and for a specific configuration described in detail in \cite{cabo2020}, two stability zones were determined, as well as the relationship between the scalar field and the matter energy density at the origin. The analysis will be extended for other values of the pressure and energy relationships as well as for the coupling between the scalar field and matter. In particular, the plans also include studying polytropic relationships.
\section*{Acknowledgments}
The authors acknowledge the partial support of the Office of External Activities of ICTP (OEA), through
the Network on Quantum Mechanics, Particles and Fields (Net-09).
|
\section{{\bf Introduction}}
A {\it hypergraph} ${\mathcal H}$ is an ordered pair ${\mathcal H}=(V,E)$, where $V$ is a set (called the {\it vertex set} of ${\mathcal H}$) and $E$ is a set of some nonempty subsets of $V$ (called the {\it hyperedge set} of ${\mathcal H}$). By a {\it proper vertex coloring} of ${\mathcal H}$, we mean a function
$c:V({\mathcal H}) \rightarrow C$ such that for each $S\in E({\mathcal H})$ with $|S|\geq 2$ we have $|\{c(v) : v\in S\}| \geq 2$. Such a set $C$ is called the set of {\it colors}. By a {\it singleton hyperedge} of ${\mathcal H}$, we mean a hyperedge $e$ with $|e|=1$. Whenever ${\mathcal H}$ has no singleton hyperedges, we define the {\it chromatic number} of ${\mathcal H}$, denoted by $\chi ({\mathcal H})$, as the minimum cardinality of a set $C$ such that a proper vertex coloring $f:V({\mathcal H})\rightarrow C$ exists. If ${\mathcal H}$ has some singleton hyperedges, then we define $\chi ({\mathcal H})=+\infty$.
For any nonnagative integer $n$, let the symbols $[n]$ denote the set $\{1,\ldots,n\}$.
Also, for a positive integer $r\geq 2$,
let $\mathbb{Z}_r=\{\omega,\omega^2,\ldots,\omega^r\}$ be a cyclic group of order $r$ with generator $\omega$.
For a vector $X=(x_1,x_2,\ldots,x_n)\in(\mathbb{Z}_r\cup\{0\})^n\setminus\{\boldsymbol{0}\}$, a subsequence of nonzero terms of $X$ is called {\it alternating} if each two consecutive terms of this subsequence are different. The length of the longest alternating subsequence of $X$ is denoted by ${\rm alt}(X)$. For example, if we set $r=3$, $n=6$, and $X=(\omega^2,0,\omega^2,\omega^1,0,\omega^3)$, then ${\rm alt}(X)=3$.
Also, we define ${\rm alt}(0,\ldots,0)$ to be zero.
For an $X=(x_1,x_2,\ldots,x_n)\in(\mathbb{Z}_r\cup\{0\})^n$ and $\epsilon\in \mathbb{Z}_r$, we define $X^\epsilon\subseteq [n]$ to be the set of all indices $j$ such that $x_j=\epsilon$, i.e.,
$X^\epsilon=\{j:\; x_j=\epsilon\}$.
By abuse of language, we can write $X=\displaystyle{\left(X^\epsilon\right)}_{\epsilon\in\mathbb{Z}_r}$.
Let ${\mathcal H}$ be a hypergraph with $n$ vertices and $\sigma:[n]\longrightarrow V({\mathcal H})$ be an injective mapping.
We define ${\rm alt}_r({\mathcal H},\sigma)$ to be the maximum possible value of $\operatorname{alt}(X)$ for $X=\displaystyle{\left(X^\epsilon\right)}_{\epsilon\in\mathbb{Z}_r}\in (\mathbb{Z}_r\cup\{0\})^n$ such that none of $\sigma(X^\epsilon)$ contains any hyperedge of ${\mathcal H}$.
In other words,
$${\rm alt}_r({\mathcal H},\sigma)=\max\left\{{\rm alt}(X): X\in(\mathbb{Z}_r\cup\{0\})^n \mbox{ and for each } \epsilon\in \mathbb{Z}_r, \mbox{ we have } E({\mathcal H}[\sigma\left(X^\epsilon\right)])=\varnothing \right\}.$$
The {\it alternation number of ${\mathcal H}$,} ${\rm alt}_r({\mathcal H})$, is the minimum possible value for ${\rm alt}_r({\mathcal H},\sigma)$ where the minimum is taken over all
injective mappings $\sigma:[n]\longrightarrow V({\mathcal H})$.
The {\it general Kneser hypergraph} ${\rm KG}^r({\mathcal H})$ is an $r$-uniform hypergraph which has $E({\mathcal H})$ as vertex set and whose hyperedges are formed by $r$ pairwise disjoint hyperedges of ${\mathcal H}$, i.e.,
$$E\left({\rm KG}^r({\mathcal H})\right)=\left\{\{e_1,\ldots,e_r\}:\; e_i\cap e_j=\varnothing \mbox{ for all } i\neq j\right\}.$$
Alishahi and Hajiabolhassan ~\cite{2013arXiv1302.5394A} presented a lower bound for the chromatic number of general Kneser hypergraphs ${\rm KG}^r({\mathcal H})$ in terms of $n$ and ${\rm alt}_r({\mathcal H})$.
\begin{theorem}{\rm (\cite{2013arXiv1302.5394A})}\label{alihaji}
For an integer $r\geq 2$ and a hypergraph ${\mathcal H}$, we have
$$\chi\left({\rm KG}^r({\mathcal H})\right)\geq \left\lceil{ |V({\mathcal H})|-{\rm alt}_r({\mathcal H})\over r-1}\right\rceil.$$
\end{theorem}
Theorem~\ref{alihaji} was applied on various families of graphs to compute their chromatic numbers or investigating their coloring properties;
see~\cite{2015arXiv151006932A,2015arXiv150708456A,2014arXiv1401.0138A,2014arXiv1403.4404A,2014arXiv1407.8035A,HaMe16,Me14}.
For two positive integers $n$ and $r$, the symbol ${[n]\choose r}$ denotes the set of all $r$-subsets of $[n]$.
The {\it complete $r$-uniform hypergraph} $K_n^r$ is a hypergraph with vertex set $[n]$ and the hyperedge set ${[n]\choose r}$.
A set consisting of $m$ pairwise disjoint hyperedges of a hypergraph is called an {\it $m$-matching}. For positive integers $r$ and $t$, the $(r,t)$-{\it hypergraph matching Ramsey number}, $R_r(s_1,s_2,\ldots,s_t)$, is the least integer $n$ such that for any $t$-hyperedge coloring of $K_n^r$, say $c:\;{[n]\choose r}\longrightarrow [t]$, there exists at least one $j\in[t]$ and at least one matching of size $s_j$ such that each hyperedge of this matching receives $j$ as its color.
The hypergraph matching Ramsey number was studied in different languages in the literature, see~\cite{MR857448,cockayne_lorimer_1975,MR0068536,MR514625}, and it was
completely characterized by Alon, Frankl, and Lov\'asz~\cite{MR857448}, as follows.
\begin{theorem} {\rm (\cite{MR857448})} \label{AlonFranklLov}
Let $s_1,s_2,\ldots,s_t$, and $r$ be positive integers with $s_1\leq s_2\leq\cdots\leq s_t$.
Then
$$R_r(s_1,s_2,\ldots,s_t)=1+\sum_{i=1}^ts_i+s_{t}(r-1)-t.$$
\end{theorem}
{\bf Outline of The Paper.}
In the rest, first we establish a development of Theorem \ref{alihaji} in Section \ref{hyperedgecoloring}. Then,
based on the results of Section \ref{hyperedgecoloring}, we resolve Theorem \ref{AlonFranklLov} in Section \ref{Resolve}. Finally, in Section \ref{ZPTucker}, we give a generalization of $\mathbf{\mathbb{Z}_p}$-Tucker Lemma together with an application of this generalized lemma in determining the matching Ramsey number $R_r(s_1,s_2,\ldots,s_t)$.
\section{{\bf Matching Coloring of Hypergraphs}}\label{hyperedgecoloring}
Let ${\mathcal H}$ be a hypergraph with $n$ vertices and $r$ be a positive integer.
A mapping $\tau: \mathbb{N}\longrightarrow \{0,1,\ldots,r-1\}$ is called a {\it color-frequency mapping}.
For a subset $A$ of $\mathbb{N}$, an {\it $(A,\tau)$-matching coloring} of ${\mathcal H}$ is
a mapping $c:E({\mathcal H})\longrightarrow A$ such that for each $a\in A$, the hypergraph induced by the hyperedges receiving $a$ as their color has no matching of size $\tau(a)+1$.
In other words, there is no $a$-monochromatic $(\tau(a)+1)$-matching.
The {\it $\tau$-matching chromatic number} of ${\mathcal H}$, denoted by $\chi_M(\tau, {\mathcal H})$, is the least
possible cardinality of a finite subset $A$ of $\mathbb{N}$ such that ${\mathcal H}$ admits an $(A,\tau)$-matching coloring.
If there exists no such finite subset $A$, then we define $\chi_M(\tau, {\mathcal H})$ to be infinite.
The following theorem provides a sharp lower bound for the $\tau$-matching chromatic number of hypergraphs.
\begin{theorem}\label{mainthmSaeed}
Let $r\geq 2$ be an integer and
$\tau:\mathbb{N}\longrightarrow \{0,1,\ldots,r-1\}$ be a color-frequency mapping.
Then, for any hypergraph ${\mathcal H}$,
we have
$$\chi_M(\tau,{\mathcal H})\geq \min\left(\left\{|A|:\; A\subset \mathbb{N},\ A \mbox{ is finite, and } \sum_{a\in A}\tau(a) \geq |V({\mathcal H})|-{\rm alt}_r({\mathcal H})\right\}\bigcup\left\{+\infty\right\}\right).$$
\end{theorem}
Note that Theorem~\ref{mainthmSaeed} is a generalization of Theorem~\ref{alihaji}. To see this, consider the color-frequency mapping $\tau:\mathbb{N}\longrightarrow \{0,1,\ldots,r-1\}$ such that $\tau(a)=r-1$ for each $a\in \mathbb{N}$.
Now, it is clear that
$\chi_M(\tau, {\mathcal H})=\chi\left({\rm KG}^r({\mathcal H})\right)$
and
$$\min\left\{|A|:\; A\subset \mathbb{N} \mbox{ and } \sum_{a\in A}\tau(a) \geq |V({\mathcal H})|-{\rm alt}_r({\mathcal H})\right\}=\left\lceil{ |V({\mathcal H})|-{\rm alt}_r({\mathcal H})\over r-1}\right\rceil.$$
\noindent{\bf Proof of Theorem~\ref{mainthmSaeed}.}
Let ${\mathcal H}$ be a hypergraph. If there is no finite subset $A\subset \mathbb{N}$ such that ${\mathcal H}$ admits an $(A,\tau)$-matching coloring, then $\chi_M(\tau, {\mathcal H})=+\infty$ and there is nothing to prove.
Let $A$ be a finite subset of $\mathbb{N}$. Our procedure is to show that if ${\mathcal H}$ admits an $(A,\tau)$-matching coloring $c: E({\mathcal H})\longrightarrow A$, then we have
$\displaystyle\sum_{a\in A}\tau(a) \geq |V({\mathcal H})|-{\rm alt}_r({\mathcal H})$.
In this regard, for each $a$ in $A$ with $\tau(a) < r-1$, we add $r-1-\tau(a)$ additional vertices
$x^{(a)}_{1},x^{(a)}_{2},\ldots ,x^{(a)}_{r-1-\tau(a)}$ to ${\mathcal H}$
together with adding all additional singleton hyperedges
$\left\{x^{(a)}_{1}\right\},\left\{x^{(a)}_{2}\right\},\ldots ,\left\{x^{(a)}_{r-1-\tau(a)}\right\}$ to $E({\mathcal H})$
in order to obtain a new hypergraph $\widehat{{\mathcal H}}$.
Also, we extend the hyperedge coloring $c: E({\mathcal H})\longrightarrow A$
to a hyperedge coloring of $\widehat{{\mathcal H}}$, say $\widehat{c}: E\left(\widehat{{\mathcal H}}\right)\longrightarrow A$, in such a way that for each $a$ in $A$ with $\tau(a) < r-1$, all additional singleton hyperedges
$\left\{x^{(a)}_{1}\right\},\left\{x^{(a)}_{2}\right\},\ldots ,\left\{x^{(a)}_{r-1-\tau(a)}\right\}$ are colored by $a$.
On one hand, since $c: E({\mathcal H})\longrightarrow A$ is an $(A,\tau)$-matching coloring
of ${\mathcal H}$, for each $a$ in $A$, the size of each matching in $E({\mathcal H})$
whose hyperedges are colored by $a$ is less than or equal to $\tau (a)$. On the other hand,
there are exactly $r-1-\tau(a)$ hyperedges in $E\left(\widehat{{\mathcal H}}\right)-E({\mathcal H})$
that are colored by $a$ under $\widehat{c}: E\left(\widehat{{\mathcal H}}\right)\longrightarrow A$. We conclude that for each $a$ in $A$, the size of any matching in $E\left(\widehat{{\mathcal H}}\right)$ whose hyperedges are colored by $a$ is at most $r-1$. Therefore,
$\widehat{c}: E\left(\widehat{{\mathcal H}}\right)\longrightarrow A$
is a proper vertex coloring of ${\rm KG}^r\left(\widehat{{\mathcal H}}\right)$. So,
we have
$$|A| \geq \chi\left({\rm KG}^r\left(\widehat{{\mathcal H}}\right)\right)\geq { \left|V\left(\widehat{{\mathcal H}}\right)\right|-{\rm alt}_r\left(\widehat{{\mathcal H}}\right)\over r-1};$$
and therefore,
$$(r-1)|A| \geq \left|V\left(\widehat{{\mathcal H}}\right)\right|-{\rm alt}_r\left(\widehat{{\mathcal H}}\right).$$
Since
$(r-1)|A| = \left( \displaystyle\sum_{a\in A}\tau(a) \right) + \left|V\left(\widehat{{\mathcal H}}\right)\right|-|V({\mathcal H})|$ and ${\rm alt}_r\left(\widehat{{\mathcal H}}\right)={\rm alt}_r({\mathcal H})$,
we have
\begin{center}
$\left( \displaystyle\sum_{a\in A}\tau(a) \right) + \left|V\left(\widehat{{\mathcal H}}\right)\right|-|V({\mathcal H})| \geq \left|V\left(\widehat{{\mathcal H}}\right)\right|-{\rm alt}_r({\mathcal H});$
\end{center}
and therefore,
$\displaystyle\sum_{a\in A}\tau(a) \geq |V({\mathcal H})|-{\rm alt}_r({\mathcal H})$;
as desired.
\hfill$\square$
In the proof of Theorem~\ref{mainthmSaeed}, the procedure is showing that for every finite subset $A$ of $\mathbb{N}$
that an $(A,\tau)$-matching coloring exists, we have
$$ \sum_{a\in A}\tau(a) \geq |V({\mathcal H})|-{\rm alt}_r({\mathcal H}).$$
If ${\mathcal H}$ has at least one hyperedge, then every $(A,\tau)$-matching coloring of ${\mathcal H}$ satisfies $ A\neq\varnothing$. So, we can rewrite Theorem~\ref{mainthmSaeed}
as follows.
\begin{theorem}\label{mainthmemptySaeed}
Let $r\geq 2$ be an integer and
$\tau:\mathbb{N}\longrightarrow \{0,1,\ldots,r-1\}$ be a color-frequency mapping.
Then, for any hypergraph ${\mathcal H}$ with $E({\mathcal H})\neq \varnothing$,
we have
$$\chi_M(\tau,{\mathcal H})\geq \min\left(\left\{|A|:\; \varnothing\neq A\subset \mathbb{N},\ A \mbox{ is finite, and } \sum_{a\in A}\tau(a) \geq |V({\mathcal H})|-{\rm alt}_r({\mathcal H})\right\}\bigcup\left\{+\infty\right\}\right).$$
\end{theorem}
In the rest of this section, we are concerned with showing that the lower bounds in Theorems \ref{mainthmSaeed} and \ref{mainthmemptySaeed} are sharp.
\begin{proposition}\label{upperSaeed}
Let $n$, $k$, and $r$ be positive integers with $r\geq 2$ and
$n\geq k$ and $n-r(k-1)\geq 0$. Also, let $\tau:\mathbb{N}\longrightarrow \{0,1,\ldots,r-1\}$
be a color-frequency mapping such that $r-1\in\tau(\mathbb{N})$.
We have
$$\chi_M\left(\tau,K_n^k\right)\leq \min\left(\left\{|A|:\; \varnothing\neq A\subset \mathbb{N},\ A \mbox{ is finite, and } \sum_{a\in A}\tau(a) \geq n-r(k-1)\right\}\bigcup\left\{+\infty\right\}\right).$$
\end{proposition}
\begin{proof}
If there is no nonempty finite subset $A$ of $\mathbb{N}$ such that $\displaystyle\sum_{a\in A}\tau(a) \geq n-r(k-1)$, then the right-hand side of the inequality is $+\infty$ and the proof
is completed. So, let us suppose that there exists a nonempty finite subset $A:=\{a_1,a_2,\ldots,a_t\}$ of $\mathbb{N}$ with
$\tau(a_{1})\leq\tau(a_{2})\leq\cdots\leq\tau(a_{t})$ and
$\displaystyle\sum_{a\in A}\tau(a) \geq n-r(k-1)$. Our aim is to show that for each such $A$, there exists some subset $B$ of $\mathbb{N}$ such that $|B|=|A|$ and $r-1\in\tau(B)$
and $\displaystyle\sum_{b\in B}\tau(b) \geq n-r(k-1)$ for which $K_n^k$ admits a $(B,\tau)$-matching coloring.
If $\tau(a_{t})=r-1$, then we set $B=A$. Otherwise, if $\tau(a_{t})\neq r-1$, then due to
$r-1\in\tau(\mathbb{N})$, we consider a positive integer $x$ with $\tau(x)=r-1$
and put $B:=(A-\{a_{t}\})\cup \{x\}$. So, in both cases, we have
$|B|=|A|$ and $r-1\in\tau(B)$
and $\displaystyle\sum_{b\in B}\tau(b) \geq \displaystyle\sum_{a\in A}\tau(a) \geq n-r(k-1)$.
For the sake of simplicity of symbols, put $B=\{b_1,b_2,\ldots,b_t\}$ with $\tau(b_{t})=r-1$.
We consider $t$ pairwise disjoint sets $S_1, S_2, \ldots, S_t$ such that
the following three conditions are hold :
\begin{itemize}
\item $\displaystyle\bigcup_{i=1}^{t}S_{i}=[n]$,
\item For each $i$ in $\{1,2,\ldots ,t-1\}$ we have $|S_i|\leq \tau (b_i)$,
\item $ |S_{t}| \leq r(k-1)+\tau (b_t)=r(k-1)+r-1=rk-1$.
\end{itemize}
Define $c:\; E(K_n^k)\longrightarrow B$ such that each $e\in E(K_n^k)={[n]\choose k}$ is mapped to
$c(e):=\min\{i:\; e\cap S_i\neq \varnothing\}$.
The mapping $c$ is a $(B,\tau)$-matching coloring of ${[n]\choose k}$; and therefore, the assertion follows.
\end{proof}
In Proposition \ref{upperSaeed}, if we assume $n\geq rk$ instead of
$n\geq r(k-1)$, we obtain the following proposition.
\begin{proposition}\label{upperemptySaeed}
Let $n$, $k$, and $r$ be positive integers with $r\geq 2$ and
$n\geq rk$. Also, let $\tau:\mathbb{N}\longrightarrow \{0,1,\ldots,r-1\}$
be a color-frequency mapping such that $r-1\in\tau(\mathbb{N})$.
We have
$$\chi_M\left(\tau,K_n^k\right)\leq \min\left(\left\{|A|:\; A\subset \mathbb{N},\ A \mbox{ is finite, and } \sum_{a\in A}\tau(a) \geq n-r(k-1)\right\}\bigcup\left\{+\infty\right\}\right).$$
\end{proposition}
It is worth pointing out that if $n\geq r(k-1)$, then
${\rm alt}_r\left(K_n^k\right)= r(k-1)$. Therefore,
combining Theorem~\ref{mainthmSaeed} and Proposition~\ref{upperemptySaeed}, leads to the
following corollary to show that the lower bound in Theorem~\ref{mainthmSaeed} is sharp.
\begin{corollary}\label{corSaeed}
Let $n$, $k$, and $r$ be positive integers such that $r\geq 2$ and $n\geq rk$.
For any color-frequency mapping $\tau:\mathbb{N}\longrightarrow \{0,1,\ldots,r-1\}$
with $r-1\in\tau(\mathbb{N})$, the $\tau$-matching chromatic number of $K_n^k$, say $\chi_M\left(\tau,K_n^k\right)$, is equal to
$$ \min\left(\left\{|A|:\; A\subset \mathbb{N},\ A \mbox{ is finite, and } \sum_{a\in A}\tau(a) \geq \left|V\left(K_n^k\right)\right|-{\rm alt}_r\left(K_n^k\right)\right\}\bigcup\left\{+\infty\right\}\right).$$
\end{corollary}
Also, combining Theorem~\ref{mainthmemptySaeed} and Proposition~\ref{upperSaeed}, leads to the
following corollary to show that the lower bound in Theorem~\ref{mainthmemptySaeed} is sharp.
\begin{corollary}
Let $n$, $k$, and $r$ be positive integers with
$r\geq 2$ and $n\geq k$ and $n-r(k-1)\geq 0$. Also, let $\tau:\mathbb{N}\longrightarrow \{0,1,\ldots,r-1\}$
be a color-frequency mapping such that $r-1\in\tau(\mathbb{N})$.
Then, $\chi_M\left(\tau,K_n^k\right)$ equals
$$ \min\left(\left\{|A|:\; \varnothing\neq A\subset \mathbb{N},\ A \mbox{ is finite, and } \sum_{a\in A}\tau(a) \geq \left|V\left(K_n^k\right)\right|-{\rm alt}_r\left(K_n^k\right)\right\}\bigcup\left\{+\infty\right\}\right).$$
\end{corollary}
\section{{\bf Resolving Alon-Frankl-Lov\'asz's Theorem}}\label{Resolve}
The aim of this Section is presenting another proof of Theorem \ref{AlonFranklLov}, as follows.
\noindent{\bf Another proof of Theorem~\ref{AlonFranklLov}.}
If $r=1$ or $t=1$ or $s_t=1$, then the assertion follows. Therefore, we may assume that $r\geq 2$ and $t\geq 2$ and $s_t\geq 2$.
Define a color-frequency mapping $\tau:\mathbb{N}\longrightarrow \{0,1,\ldots,s_t-1\}$ such that for each $i\in\mathbb{N}$,
$$\tau(i)=\left\{\begin{array}{ll}
s_i-1 & \mbox { if } i\in[t]\\
0 & \mbox{otherwise.}
\end{array}\right.$$
In view of Theorem~\ref{mainthmSaeed}, if $\displaystyle\sum_{a\in \mathbb{N}}\tau(a)=\displaystyle\sum_{i=1}^t(s_i-1)<n-{\rm alt}_{s_t}(K_n^r)$, then there is no $(A,\tau)$-matching coloring of $K_n^r$ for each finite set $A\subset \mathbb{N}$. This implies that
since ${\rm alt}_{s_t}\left(K_n^r\right)\leq s_t(r-1)$,
for each $n\geq 1+\displaystyle\sum_{i=1}^t(s_i-1)+s_t(r-1)$ and
any $t$-hyperedge coloring
$c:\; {[n]\choose r}\longrightarrow [t]$, there are $j\in[t]$ and $s_j$ pairwise disjoint hyperedges
$e_1,e_2,\ldots,e_{s_j} \in E(K_n^r)$ such that $c(e_1)=\cdots=c(e_{s_j})=j$.
In other words, there is a matching $M$ of size $s_j$ whose hyperedges are colored with the same color $j$. Consequently, we have $$R_r(s_1,s_2,\ldots,s_t)\leq 1+\sum_{i=1}^ts_i+s_t(r-1)-t.$$
So, it suffices to show that
$$R_r(s_1,s_2,\ldots,s_t)\geq1+\sum_{i=1}^ts_i+s_t(r-1)-t.$$
In this regard, we consider $t$ pairwise disjoint sets $S_1, S_2, \ldots, S_t$ such that
the following three conditions are hold :
\begin{itemize}
\item $\displaystyle\bigcup_{i=1}^{t}S_{i}=\left[ \sum_{i=1}^ts_i+s_t(r-1)-t\right] $,
\item For each $i$ in $\{1,2,\ldots ,t-1\}$ we have $|S_i|= s_i-1$,
\item $|S_{t}| = s_t(r-1)+s_t-1=s_tr-1$.
\end{itemize}
For the sake of simplicity, put
$\Lambda:=\displaystyle\sum_{i=1}^ts_i+s_t(r-1)-t.$
Define a {\it bad} $t$-hyperedge coloring $C_{bad}:\; E(K_{\Lambda}^r)\longrightarrow [ t]$ such that each $e\in E(K_{\Lambda}^r)={[{\Lambda}]\choose r}$ is mapped to
$C_{bad}(e):=\min\{i:\; e\cap S_i\neq \varnothing\}$.
Since for each $i$ in $[t]$ there are not any matchings of size $s_{i}$ whose all hyperedges are colored by $i$, we conclude that
$$R_r(s_1,s_2,\ldots,s_t)\geq1+\sum_{i=1}^ts_i+s_t(r-1)-t;$$
which is desired.
\hfill$\square$
\section{{\bf The $\mathbf{\mathbb{Z}}_p$-Tucker Lemma : A Generalization}}\label{ZPTucker}
This section is devoted to present a generalization of ${\mathbb{Z}_p}$-Tucker Lemma.
The ${\mathbb{Z}_p}$-Tucker Lemma is a development of the celebrated Tucker's Lemma (which is equivalent to the famous
Borsuk-Ulam Theorem). There are various important and surprising applications of Tucker's Lemma. For a very detailed discussion in this matter, one can see the interesting reference \cite{MR1988723}.
First, we present the $\mathbb{Z}_p$-Tucker Lemma. In this regard, we should note that for $X_1$ and $X_2$ in $(\mathbb{Z}_p\cup\{0\})^n$,
we write $X_1\subset X_2$ whenever for each $\epsilon\in \mathbb{Z}_p$ we have $X_1^\epsilon\subseteq X_2^\epsilon$.
\begin{lemma}{\rm (\cite{MR2793613,MR1893009}) ($\mathbb{Z}_p$-Tucker Lemma)}
Let $m$ and $n$ be positive integers, $p$ be a prime number, and $\alpha$ be nonnegative integer with $0\leq \alpha \leq m$. Also, let
$$
\begin{array}{rl}
\lambda:\ (\mathbb{Z}_p\cup\{0\})^n\setminus\{\boldsymbol{0}\} & \longrightarrow \mathbb{Z}_p\times[m] \\
X & \longmapsto(\lambda_1(X),\lambda_2(X))
\end{array}
$$
be a mapping that satisfies all of the following three properties simultaneously :
\begin{enumerate}
\item[{\rm 1.}] The mapping $\lambda$ is a $\mathbb{Z}_p$-equivariant mapping, that is, for each $\omega^j\in \mathbb{Z}_p$,
we have $$\lambda_1(\omega^jX)=\omega^j\lambda_1(X) \ {\rm and}\
\lambda_2(\omega^jX)=\lambda_2(X).$$
\item[{\rm 2.}] For all $X_1\subset X_2\in (\mathbb{Z}_p\cup\{0\})^n\setminus\{\boldsymbol{0}\}$,
the condition $\lambda_2(X_1)=\lambda_2(X_2)\leq\alpha$ implies $\lambda_1(X_1)=\lambda_1(X_2)$.
\item[{\rm 3.}] For all $X_1\subset X_2\subset \cdots \subset X_p\in (\mathbb{Z}_p\cup\{0\})^n\setminus\{\boldsymbol{0}\}$,
if $\lambda_2(X_1)=\lambda_2(X_2)=\cdots=\lambda_2(X_p)\geq\alpha+1$,
then $$|\{\lambda_1(X_1),\lambda_1(X_2),\ldots,\lambda_1(X_{p})\}|\leq p-1.$$
\end{enumerate}
Then, we have $$\alpha+(m-\alpha)(p-1) \geq n.$$
\end{lemma}
The next lemma is a generalization of $\mathbb{Z}_p$-Tucker Lemma.
The proof is similar to the one by Meunier~\cite{MR2793613}.
\begin{lemma}\label{Z_ptucker}
Let $p$ be a prime number and let $ \gamma_1,\gamma_2,\ldots,\gamma_{m}\in [p-1]$. Also, let
$$
\begin{array}{rll}
\lambda:(\mathbb{Z}_p\cup\{0\})^n\setminus\{\boldsymbol{0}\} & \longrightarrow & \mathbb{Z}_p\times[m]\\
X& \longmapsto & (\lambda_1(X),\lambda_2(X))
\end{array}$$
be a $\mathbb{Z}_p$-equivariant mapping
such that for each $i\in[m]$ and
for each chain $X_1\subset X_2\subset\cdots\subset X_{l}\in (\mathbb{Z}_p\cup\{0\})^n\setminus\{\boldsymbol{0}\}$, if $\lambda_2(X_1)=\lambda_2(X_2)=\cdots=\lambda_2(X_{l})=i,$ then
$$|\{\lambda_1(X_1),\lambda_1(X_2),\ldots,\lambda_1(X_{l})\}|\leq \gamma_i.$$
Then, we have $$\sum_{i=1}^{m}\gamma_i\geq n.$$
\end{lemma}
\begin{proof}
Clearly,
the mapping $\lambda$ can be considered as a $\mathbb{Z}_p$-equivariant simplicial mapping from
${\rm sd}(\mathbb{Z}^{*n}_p)$ to
$$C=\displaystyle\left(\sigma_{\gamma_1-1}^{p-1}\right)*\cdots*\left(\sigma_{\gamma_m-1}^{p-1}\right).$$
Therefore,
by Dold's Theorem~\cite{MR711043,MR1988723},
the dimension of
$C$ must be strictly larger than the
connectivity of ${\rm sd}\left(\mathbb{Z}^{*n}_p\right)$; which implies
$$\sum_{i=1}^{t}\gamma_i-1> n-2;$$ as desired.
\end{proof}
One can observe that Lemma \ref{Z_ptucker} is a development of $\mathbb{Z}_p$-Tucker Lemma by regarding
$$\gamma_1=\gamma_2=\cdots =\gamma_{\alpha}=1 {\rm ,\ and} \ \gamma_{\alpha +1}=\gamma_{\alpha +2}=\cdots =\gamma_{m}=p-1.$$
The following proposition is a special case of Theorem \ref{mainthmSaeed}; nevertheless, its proof shows an explicit application of Lemma \ref{Z_ptucker}.
\begin{proposition}\label{mainthm}
Let $p$ be a prime number and
$\tau:\mathbb{N}\longrightarrow \{0,1,\ldots,p-1\}$ be a color-frequency mapping.
Then, for any hypergraph ${\mathcal H}$,
we have
$$\chi_M(\tau,{\mathcal H})\geq \min\left(\left\{|A|:\; A\subset \mathbb{N},\ A \mbox{ is finite, and } \sum_{a\in A}\tau(a) \geq |V({\mathcal H})|-{\rm alt}_p({\mathcal H})\right\}\bigcup\left\{+\infty\right\}\right).$$
\end{proposition}
\noindent{\bf Proof of Proposition~\ref{mainthm} by means of Lemma \ref{Z_ptucker}.}
Let $|V({\mathcal H})|=n$ and consider a bijection $\sigma:[n]\longrightarrow V({\mathcal H})$
such that ${\rm alt}_p({\mathcal H},\sigma)={\rm alt}_p({\mathcal H})$.
First note that if there is no finite subset $A\subset \mathbb{N}$ such that ${\mathcal H}$ admits an $(A,\tau)$-matching coloring, then $\chi_M(\tau, {\mathcal H})=+\infty$ and there is nothing to prove.
Therefore, we can assume that $\chi_M(\tau, {\mathcal H})=t$ is finite.
If $t=0$, then ${\mathcal H}$ has not any hyperedges; and therefore,
${\rm alt}_p({\mathcal H})=|V({\mathcal H})|$. Hence, the empty set satisfies
$\sum_{a\in \varnothing}\tau(a) = |V({\mathcal H})|-{\rm alt}_p({\mathcal H})=0$; and we are done.
So, we may assume that $t$ is a positive integer. Let $c: E({\mathcal H})\longrightarrow A$ be an $(A,\tau)$-matching coloring of ${\mathcal H}$
where $A=\{a_1,a_2,\ldots,a_t\}$
and $a_1< a_2<\cdots< a_t$.
To prove the assertion, we need to show that
$n\leq {\rm alt}_p({\mathcal H})+\displaystyle\sum_{a\in A}\tau(a)$.
Let $m={\rm alt}_p({\mathcal H})+t$.
For each $i\in[m]$,
define $$\gamma_i=\left\{\begin{array}{ll}
1 & \mbox{if } i\leq {\rm alt}_p({\mathcal H})\\
\tau(a_{i-{\rm alt}_p({\mathcal H})}) & \mbox{if } i\geq {\rm alt}_p({\mathcal H})+1.
\end{array}\right.$$
Now, we consider a mapping
$$\begin{array}{rll}
\lambda:(\mathbb{Z}_p\cup\{0\})^n\setminus\{\boldsymbol{0}\} & \longrightarrow & \mathbb{Z}_p\times[m]\\
X& \longmapsto & (\lambda_1(X),\lambda_2(X))
\end{array}$$ in order to apply Lemma~\ref{Z_ptucker}.
We endow $2^{[n]}$ with an arbitrary total ordering $\preceq$.
\begin{itemize}
\item If $X\in (\mathbb{Z}_p\cup\{0\})^n\setminus\{\boldsymbol{0}\}$ satisfies ${\rm alt(X)}\leq {\rm alt}_p({\mathcal H},\sigma)$, then set $\lambda_1(X)$ to be the first nonzero component of $X$ and put $\lambda_2(X)={\rm alt}(X)$.
\item If $X\in (\mathbb{Z}_p\cup\{0\})^n\setminus\{\boldsymbol{0}\}$ satisfies ${\rm alt(X)}> {\rm alt}_p({\mathcal H},\sigma)$, then there exists some $\epsilon$ in $\mathbb{Z}_p$
such that $\sigma(X^\epsilon)$ is a superset of at least one hyperedge
$e$ of ${\mathcal H}$, i.e., $e\subseteq \sigma(X^\epsilon)$. We
define $\zeta(X)$ as the maximum positive integer $j$ for which there exists an $\epsilon\in\mathbb{Z}_p$ and some hyperedge $e\in E({\mathcal H})$ such that
$e\subseteq \sigma(X^\epsilon)$ and $c(e)=a_j$. Set $\lambda_2(X)=\zeta(X)+{\rm alt}_p({\mathcal H})$. For determining $\lambda_1(X)$, put $\zeta(X)=j$ and choose $X^{\lambda_1(X)}$ as the maximum $X^\epsilon$ with respect to the ordering $\preceq$ such that
$\sigma(X^\epsilon)$ is a superset of some hyperedge
$e$ with $c(e)=a_{j}$.
\end{itemize}
Since the mapping $\lambda$ satisfies the condition of Lemma~\ref{Z_ptucker}
with respect to prior presented $\gamma_i$'s, we conclude that
$$\begin{array}{lll}
n &\leq & \displaystyle\sum_{i=1}^{m}\gamma_i \\ \\
&=& {\rm alt}_p({\mathcal H})+\displaystyle\sum_{a\in A}\tau(a);
\end{array}$$
as desired.
\hfill$\square$
Using the previous proposition, we can prove the following proposition, which shows an application of Lemma \ref{Z_ptucker} in determining $R_r(s_1,s_2,\ldots,s_t)$ for some special cases.
\begin{proposition}\label{ramsey:main}
Let $s_1,s_2,\ldots,s_t$, $p$, and $r$ be positive integers, where $p$ is a prime number and $s_1\leq s_2\leq\cdots\leq s_t\leq p$.
Then
$$1+\sum_{i=1}^ts_i+s_t(r-1)-t\leq R_r(s_1,s_2,\ldots,s_t)\leq 1+\sum_{i=1}^ts_i+p(r-1)-t.$$
In particular, we have $$R_r(s_1,s_2,\ldots,s_t)=1+\sum_{i=1}^ts_i+s_t(r-1)-t,$$
provided that $s_t$ is a prime number.
\end{proposition}
{\bf Remark.} In a forthcoming paper, the second author extends the familiar results for all variations of colorability defect and equitable colorability defects of hypergraphs and multihypergraphs.
\section*{Acknowledgement}
The authors wish to express their gratitude to Professor Hossein Hajiabolhassan for many stimulating conversations.
\def$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'${$'$}
|
\section{Introduction}
In this paper, we work over ${\mathbb C}$.
For an irreducible variety $Y$, let $\mathbb C(Y)$ be the function field of $Y$.
Let $X$ be a smooth hypersurface of degree $d\geq 4$ in $\mathbb P^{n+1}$, $p$ be a point in $\mathbb P^{n+1}$, and $\pi_p:X\dashrightarrow H$ be a projection with center $p$ where $H$ is a hyperplane not containing $p$.
We have an extension of function fields $\pi^{\ast}:\mathbb C(H)\rightarrow \mathbb C(X)$ such that $[\mathbb C(X):\mathbb C(H)]=d-1$ (resp. $d$) if $p\in X$ (resp. $p\not\in X$).
The structure of this extension does not depend on the choice of $H$ but on the point $p$.
We write $K_p$ instead of $\mathbb C(H)$.
Since $H\cong\mathbb P^n$, $K_p\cong \mathbb C(\mathbb P^n)$ as a field.
Let $Y$ be an irreducible variety $Y$.
Let $K$ be a non-trivial intermediate field between $\mathbb C(Y)$ and $\mathbb C$ such that $K$ is a purely transcendental extension of $\mathbb C$ with the transcendence degree $n$.
The field $K$ is called a maximal rational subfield if there is not a non-trivial intermediate field $L$ between $\mathbb C(Y)$ and $K$ such that $L$ is a purely transcendental extension of $\mathbb C$ with the transcendence degree $n$.
Let $X$ be a smooth hypersurface of degree $d\geq 4$ in $\mathbb P^{n+1}$.
If $n=1$, then the field $K_p$ is a maximal rational subfield of $\mathbb C(X)$ ([\ref{bio:1000}]).
In the case where $n=2$ and $d=4$, if $p$ is not an outer Galois point of $X$,
then the field $K_p$ is a maximal rational subfield.
If $d\geq 5$, then $K_p$ is always a maximal rational subfield.
Please see [\ref{bio:1001},\ref{bio:200}] for details.
\begin{dfn}$([\ref{bio:20},\ref{bio:21},\ref{bio:22}])$.
The point $p\in\mathbb P^{n+1}$ is called a Galois point for $X$ if the extension $\mathbb C(X)/K_p$ is Galois.
Moreover, if $p\in X$ $($resp. $p\not\in X)$, then we call $p$ an inner $($resp. outer$)$ Galois point.
\end{dfn}
Pay attention that if $n=1$ or $p\not\in X$, then $\pi_p$ is a morphism such that $\pi_p:X\rightarrow \mathbb P^n$ is a Galois cover of a variety.
\begin{thm}$([\ref{bio:20},\ref{bio:21},\ref{bio:22}])$.
Let $X$ is a smooth hypersurface of degree $d\geq 4$ in $\mathbb P^{n+1}$, and $p\in\mathbb P^{n+1}$ be a Galois point of $X$.
Then the Galois group of $\mathbb C(X)/K_p$ is induced by a linear automorphism of $X$.
In addition, if $p$ is an inner (resp. outer) Galois point, then the Galois group of $\mathbb C(X)/K_p$ is a cyclic group of $d-1$ (resp. $d$)
\end{thm}
\begin{dfn}
An automorphism $g$ of $X$ is called linear if there is an automorhism $h$ of $\mathbb P^{n+1}$ such that $h(X)=X$ and $h_{|X}=g$.
\end{dfn}
If $X$ is a smooth hypersurface of degree $d\geq 4$ in $\mathbb P^{n+1}$ and $(n,d)\not=(2,4)$,
then the automorphism group Aut$(X)$ of $X$ is a finite subgroup of the group PGL$(n+2,\mathbb{C})={\rm Aut}(\mathbb P^{n+1})$, for instance, see ([\ref{bio:102}]).
\begin{dfn}
Let $p\in\mathbb P^{n+1}$ is a Galois point of $X$.
An automorphism $g$ of $X$ is called an automorphism belonging to the Galois point $p$ if $g$ generates the Galois group of the Galois extension $\mathbb C(X)/K_p$.
\end{dfn}
\begin{dfn}
Let $g$ be a linear automorphism of $X$.
A matrix $A$ is called a representation matrix of $g$ if $g=A$ in ${\rm PGL}(n+2,\mathbb{C})$.
\end{dfn}
A necessary and sufficient condition for a smooth hypersurface $X\subset\mathbb P^{n+1}$ to have Galois points is given by the defining equation of $X$ $([\ref{bio:20},\ref{bio:21},\ref{bio:22}])$.
For the case $n=1$, there is a sufficient condition for a smooth plane $X$ curve to have Galois points by the structure of the automorphism group ${\rm Aut}(X)$ as follows.
\begin{thm}\label{thm:1}$([\ref{bio:1}])$.
Let $X$ be a smooth hypersurface of degree $d\geq 4$ in $\mathbb P^{n+1}$, and $g\in{\rm Aut}(X)$ be a linear automorphism of order $k(d-1)$ $(resp.\ kd)$ for $n,k\geq 1$.
If $n=1$ and $k\geq2$, then $X$ has an inner $(resp.\ outer)$ Galois point $p$, and $g^k$ is an automorphism belonging to the Galois point $p$.
\end{thm}
Smooth curves in $\mathbb P^2$ with Galois points are characterized by other methods as well [\ref{bio:11},\ref{bio:12},\ref{bio:13}].
There are smooth plane curves of degree $d$ with a linear automorphism of order $d-1$ or $d$ acting but without Galois points (see Examples \ref{exa:1} and \ref{exa:2}).
In addition, there is a smooth hypersurface $X$ of degree $d$ in $\mathbb P^4$ with a linear automorphism of order $(d-1)d$ acting but without Galois points (see Example \ref{exa:3}).
Therefore, Theorem \ref{thm:1} does not hold for all $n,k\geq1$.
For $g\in$Aut$(X)$,
we set ${\rm Fix}(g):=\{x\in X\,|\,g(x)=x\ \}$, and
we write the order of $g$ as ord$(g)$.
Recall that if $X$ is a smooth hypersurface and $(n,d)\not=(2,4)$,
then Aut$(X)$ is a subgroup of PGL$(n+2,\mathbb{C})$, i.e. all automorphisms of $X$ are linear.
In this paper, by using Fix$(g)$ and ord$(g)$,
we will study the case $k,n\geq1$ of Theorem \ref{thm:1}.
Our main results are Theorems \ref{thm:2}, \ref{thm:3}, \ref{thm:4}, and \ref{thm:5}.
Theorem \ref{thm:2}, is for $n=k=1$.
\begin{thm}\label{thm:2}
Let $X$ be a smooth plane curve degree $d\geq 4$, and $g$ be a linear automorphism of $X$.\\
$(1)$ If ${\rm ord}(g)=d-1$,
then $\sharp|{\rm Fix}(g)|\not=2$ if and only if $X$ has an inner Galois point $p$,
and $g$ is an automorphism belonging to the Galois point $p$.\\
$(2)$ If ${\rm ord}(g)=d$,
then ${\rm Fix}(g)\not=\emptyset$ if and only if $X$ has an outer Galois point $p$,
and $g$ is an automorphism belonging to the Galois point $p$.
\end{thm}
Theorem \ref{thm:3} is for $k=1$, $n\geq 2$, and an inner Galois point.
\begin{thm}\label{thm:3}
Let $X$ be a smooth hypersurface of degree $d\geq 4$ in $\mathbb P^{n+1}$, and $g\in{\rm Aut}(X)$ be a linear automorphism of order $d-1$.\\
(1) If $n=2$, then ${\rm Fix}(g)$ contains a curve $C'$ which is not a smooth rational curve if and only if
$X$ has an inner Galois point $p$, and $g$ is an automorphism belonging to the Galois point $p$.\\
(2) If $n\geq 3$, then ${\rm Fix}(g)$ has codimension $1$ in $X$ if and only if
$X$ has an inner Galois point $p$, and $g$ is an automorphism belonging to the Galois point $p$.
\end{thm}
Theorem \ref{thm:4} is for $k=1$, $n\geq 2$, and an outer Galois point.
\begin{thm}\label{thm:4}
Let $X$ be a smooth hypersurface of degree $d$ in $\mathbb P^{n+1}$, and $g\in{\rm Aut}(X)$ be a linear automorphism of order $d$.
If $d\geq 2$, then ${\rm Fix}(g)$ has codimension $1$ in $X$ if and only if $X$ has an outer Galois point $p$, and $g$ is an automorphism belonging to the Galois point $p$.
\end{thm}
The following Theorem is for $n,k\geq 2$ and an inner Galois point.
\begin{thm}\label{thm:5}
Let $X$ be a smooth hypersurface of degree $d$ in $\mathbb P^{n+1}$, and $g\in{\rm Aut}(X)$ be a linear automorphism of order $k(d-1)$ for $k\geq 2$.\\
(1) If $n=2$ and $\sharp|{\rm Fix}(g)|\geq5$, then $X$ has an inner Galois point $p$, and $g^k$ is an automorphism belonging to the Galois point $p$.\\
(2) If $n\geq 3$ and ${\rm Fix}(g)$ has codimension $1$ or $2$ in $X$, then $X$ has an inner Galois point $p$, and $g^k$ is an automorphism belonging to the Galois point $p$.
\end{thm}
Theorem \ref{thm:5} does not fold for an outer Galois point (see Example \ref{exa:6}).
For $n=1$, the automorphism groups of curves with Galois points are classified ([\ref{bio:1},\ref{bio:8}]).
There are studies on automorphism groups of plane curves using Galois points ([\ref{bio:1},\ref{bio:10},\ref{bio:11},\ref{bio:12},\ref{bio:16},\ref{bio:17}]).
For the case $n\geq 2$, determining whether $X$ has Galois points from the structure of ${\rm Aut}(X)$ may be an important issue.
\begin{que}\label{q:1}
For $n\geq 1$, is there a group $G_n$ satisfying the following condition ?
The condition: If the automorphism group ${\rm Aut}(X)$ of a smooth hypersurface $X$ of degree $d\geq 4$ in $\mathbb P^{n+1}$ has a subgroup $H$ which is isomorphic to $G$ as a group, then $X$ has a Galois point.
\end{que}
Theorem \ref{thm:1} is an answer to Question \ref{q:1} for the case $n=1$.
However, our main theorems are not answers to Question \ref{q:1}, because they need the fixed points set.
Section 2 is preliminary. We will explain the basic facts of Galois point.
In section 3, we will show Theorems \ref{thm:2}, \ref{thm:3}, \ref{thm:4}, and \ref{thm:5}.
\section{Preliminary}
Let $X$ be a smooth hypersurface of degree $d\geq4$ in $\mathbb P^{n+1}$.
We denote the number of inner (resp. outer) Galois points of $X$ by $\delta(X)$ (resp. $\delta'(X)$).
Here $[s]$ represents the integer part of $s\in\mathbb R$.
\begin{thm}\label{thm:6}$([\ref{bio:20},\ref{bio:21},\ref{bio:22}])$.
Let $X$ be a smooth hypersurface of degree $d\geq 4$ in $\mathbb P^{n+1}$.
The following holds.\\
(1) If $n=1$, then $\delta(X)=0,1$, or $4$, and $\delta'(X)=0,1$, or $3$.
In particular, if $n=1$ and $d\geq 5$, then $\delta(X)=0$ or $1$.\\
(2) If $n\geq2$ and $d=4$, then $\delta(X)\leq 4([\frac{n}{2}]+1)$.
In particular, if $n=2$ and $d=4$, then $\delta(X)=0,1,2,4$, or $8$.\\
(3) If $n\geq2$ and $d\geq5$, then $\delta(X)\leq [\frac{n}{2}]+1$.\\
(4) If $n\geq2$ and $d\geq4$, then $\delta'(X)\leq n+2$.
\end{thm}
The numbers of Galois points of normal hypersurfaces are investigated ([\ref{bio:5},\ref{bio:19}]).
The defining equations for smooth hypersurfaces with a Galois point are determined.
\begin{thm}\label{thm:7}$([\ref{bio:20},\ref{bio:21},\ref{bio:22}])$.
Let $X$ be a smooth hypersurface of degree $d\geq 4$ in $\mathbb P^{n+1}$.
The following holds.\\
(1) $X$ has an inner Galois point $p$ if and only if by replacing the local coordinate system if necessary, $p=[1:0:\cdots:0]$ and $X$ is defined by
\[ X_1X_0^{d-1}+F(X_1,\ldots,X_{n+1})=0.\]
(2) $X$ has an outer Galois point $p$ if and only if
by replacing the local coordinate system if necessary, $p=[1:0:\cdots:0]$ and $X$ is defined by
\[ X_0^d+F(X_1,\ldots,X_{n+1})=0.\]
\end{thm}
The definition equations with many Galois points are also studied (please see [\ref{bio:22}] for more detailed results).
For a positive integer $l$, let $I_l$ be the identity matrix of size $l$, and $e_l$ be a primitive $l$-th root of unity.
Theorem \ref{thm:8} below is a rewrite of Theorem \ref{thm:7} from the viewpoint of a liner automorphism.
\begin{thm}\label{thm:8}
Let $X$ be a smooth hypersurface of degree $d$ in $\mathbb P^{n+1}$, $g\in{\rm Aut}(X)$ be a linear automorphism of order $d-1$ $(resp.\ d)$, and $A$ be a representation matrix of $g$.
There is a Galois point $p$ of $X$ such that $g$ is an automorphism belonging to the Galois point $p$ if and only if
the matrix $A$ is conjugate to a matrix
\[\begin{pmatrix}
a & 0\\
0&bI_{n+1}
\end{pmatrix}\]
such that $\frac{a}{b}=e_{d-1}$ $(resp.\ e_d)$.
In particular, if $A$ is conjugate to the above matrix, then
the Galois point $p$ is the eigenvector corresponding to the eigenvalue $a$.
\end{thm}
From Theorem \ref{thm:8}, we see that the only if parts of Theorems \ref{thm:3} and \ref{thm:4} holds.
From here, we give examples of smooth hypersurfaces of degree $d$ without Galois points which have a linear automorphism such that the order is a multiple of $d-1$ or $d$.
As a corollary of Theorem \ref{thm:8}, we give the following two lemmas.
\begin{lem}\label{lem:1}
Let $X$ be a smooth hypersurface of degree $d\geq4$ in $\mathbb P^{n+1}$, $p\in\mathbb P^{n+1}$, and $g$ be an automorphism belonging to the Galois point $p$.
For any linear automorphism $h$ of $X$, $h(p)$ is also a Galois point of $X$, and $h\circ g\circ h^{-1}$ is an automorphism belonging to the Galois point $h(p)$.
In particular, if $p$ is an inner (resp. outer) Galois point, then $h(p)$ is also an inner (resp. outer) Galois point.
\end{lem}
\begin{proof}
By a linear automorphism $h\circ g\circ h^{-1}$ and Theorem \ref{thm:8}, $h(p)$ is a Galois point of $X$, and
$h\circ g\circ h^{-1}$ is an automorphism belonging to the Galois point $h(p)$.
\end{proof}
\begin{lem}\label{lem:2}
Let $X$ be a smooth hypersurface of degree $d\geq4$ in $\mathbb P^{n+1}$, $p\in\mathbb P^{n+1}$, and $g$ be an automorphism belonging to the Galois point $p$.
For a linear automorphism $k$ of $X$ such that $k(p)=p$, we get that $k\circ g=g\circ k$.
\end{lem}
\begin{proof}
By Lemma \ref{lem:1}, $k\circ g\circ k^{-1}$ is an automorphism belonging to the Galois point $p$.
By Theorem \ref{thm:8}, $k\circ g\circ k^{-1}=g$.
\end{proof}
In Example \ref{exa:1}, we give an example of a smooth plane curve of degree $d$ with a linear automorphism of order $d-1$ but has no Galois points.
Before that, we prepare a lemma.
\begin{lem}\label{lem:3}
Let $A:=(a_{ij})$ be a diagonal $m\times m$ matrix such that $a_{ii}\not=a_{jj}$ for $1\leq i<j \leq m$.
For a $m\times m$ matrix $B:=(b_{ij})$, if $AB=BA$, then $B$ is a diagonal matrix.
\end{lem}
\begin{proof}
We assume that $AB=BA$.
The $(i,j)$-th entry of the matrix $AB$ is $a_{ii}b_{ij}$.
The $(i,j)$-th entry of the matrix $BA$ is $a_{jj}b_{ij}$.
Since $a_{ii}\not=a_{jj}$ for $\leq i<j \leq m$, we get that $b_{ij}=0$ for $\leq i<j \leq m$.
Then the matrix $B$ is a diagonal matrix.
\end{proof}
\begin{exa}\label{exa:1}
Let $d$ be an even number of $6$ or more, and $X$ be a smooth curve in $\mathbb P^2$ defined by
\[X_2^d+X^{d-1}_0X_2+X^{d-1}_1X_2+X_0^{\frac{d}{2}}X_1^{\frac{d}{2}}=0.\]
The curve $X$ has an automorphism $g$ of order $d-1$ such that the following matrix $A$ is a representation matrix of $g$:
\[A:=
\begin{pmatrix}
e_{d-1}^{\frac{d}{2}} &0 &0\\
0 & e_{d-1}^{\frac{d}{2}-1} &0\\
0 & 0 &1
\end{pmatrix}.\]
For $1\leq i<d-1$, we get that $\frac{d}{2}i\not \equiv0$ $({\rm mod}\ d-1)$, $(\frac{d}{2}-1)i\not\equiv0$ $({\rm mod}\ d-1)$, and $\frac{d}{2}i\not\equiv(\frac{d}{2}-1)i$ $({\rm mod}\ d-1)$.
We assume that $X$ has a Galois point $p\in \mathbb P^2$.
By Lemma \ref{lem:1}, $g^j(p)$ is a Galois point for $1\leq j<d-1$.
By Theorem \ref{thm:6}, $\delta(X)\leq 4$ and $\delta'(X)\leq3$.
Since $d\geq 6$, $g^l(p)=p$ for some $1\leq l< d-1$.
Let $h\in{\rm Aut}(X)$ be an automorphism belonging to the Galois point $p$.
Since $g^l(p)=p$, the automorphism $g^l\circ h\circ g^{-l}$ is also an automorphism belonging to the Galois point $p$.
Then $g^l\circ h\circ g^{-l}=h^i$ for some $1\leq i<d-1$.
By Theorem \ref{thm:8}, we can take a representation matrix $B$ of $h$ such that
\[CBC^{-1}=
\begin{pmatrix}
e_k &0 &0\\
0 & 1 &0\\
0 & 0 &1
\end{pmatrix}\]
for some a matrix $C$ where if $p\in X$, then $k=d-1$, and if $p\not \in X$, then $k=d$.
By the equation $g^l\circ h\circ g^{-l}=h^i$, we get that $i=1$, and $A^lBA^{-l}=B$.
Since the diagonal entries of $A^l$ are different from each other, Lemma \ref{lem:3}, and $A^lBA^{-l}=B$, we get that $B$ is a diagonal matrix.
Since $h=B$ is an automorphism belonging to the Galois point $p$, and Theorem \ref{thm:8},
we get that
\[p\in\{[1:0:0],[0:1:0],[0:0:1]\},\]
and the matrix $B$ is one of the following matrices
\[
\begin{pmatrix}
a &0 &0\\
0 & b &0\\
0 & 0 &b
\end{pmatrix},
\begin{pmatrix}
b & 0 &0\\
0 & a &0\\
0 & 0 &b
\end{pmatrix},\ {\rm and}\
\begin{pmatrix}
b &0 &0\\
0 & b &0\\
0 & 0 &a
\end{pmatrix}\]
where if $p\in X$, then $\frac{a}{b}=e_{d-1}$, and if $p\not \in X$, then $\frac{a}{b}=e_d$.
The defining equation of $X$ implies that $h=B$ is not an automorphism of $X$.
This is a contradiction.
Therefore, $X$ does not have Galois points.
\end{exa}
Below is an example of a smooth plane curve of degree $d$ with a linear automorphism $d$ but has no Galois points.
\begin{exa}\label{exa:2}
Let $d_1$ and $d_2$ be integers greater than $4$ such that gcd$(d_1,d_2)=1$.
Let $d:=d_1d_2$, and $X$ be a smooth curve in $\mathbb P^2$ defined by
\[X_0^d+X_1^d+X_2^d+X^{d_1}_0X^{d_2}_1X_2^{d-d_1-d_2}=0.\]
The curve $X$ has an automorphism $g$ of order $d$ such that the following matrix $A$ is a representation matrix of $g$:
\[A:=
\begin{pmatrix}
e_{d_1} &0 &0\\
0 & e_{d_2} &0\\
0 & 0 &1
\end{pmatrix}.\]
For $1\leq i<d$, the diagonal entries of $A^i$ are different from each other.
As like Example \ref{exa:1},
we get that $X$ does not have Galois points.
\end{exa}
We give an example of a smooth surface $X$ of degree $d\geq 4$ in $\mathbb P^3$ such that $X$ has a linear automorphism $g$ of order $(d-1)d$ but has no Galois points.
\begin{exa}\label{exa:3}
Let $d_1\geq 5$ be an odd integer, and $d:=2d_1+1$.
Let $X$ be a smooth surface of degree $d$ in $\mathbb P^3$ defined by
\[X_0^d+X_0^{\frac{d+1}{2}}X_1^{\frac{d-1}{2}}+X_0X_1^{d-1}+X_2^{d-1}X_3+X_2X_3^{d-1}=0.\]
The surface $X$ has an automorphism $g$ of order $(d-1)d$ such that the following matrix $A$ is a representation matrix of $g$
\[
A:=
\begin{pmatrix}
e_{\frac{(d-1)}{2}d}^{1-d}&0&0\\
0&e_{\frac{(d-1)}{2}d}&0\\
0&0&0&1\\
0&0&1&0
\end{pmatrix}.
\]
In addition,
the surface $X$ has an automorphism $h$ of order $(d-2)\frac{(d-1)}{2}d$ such that the following matrix $B$ is a representation matrix of $h$
\[
B:=
\begin{pmatrix}
e_{\frac{(d-1)}{2}d}^{1-d}&0&0\\
0&e_{\frac{(d-1)}{2}d}&0\\
0&0&e_{d-2}&0\\
0&0&0&e_{d-2}^{-1}
\end{pmatrix}.
\]
For $1\leq i<\frac{(d-1)}{2}d$, the diagonal entries of $B^i$ are different from each other.
By Theorem \ref{thm:6}, $\delta(X)\leq 2$ and $\delta'(X)\leq4$.
Since $\frac{(d-1)}{2}d\geq 5$, if $X$ has a Galois point, then there is a Galois point $p$ of $X$ such that $g^l(p)=p$ for some $1\leq l<\frac{(d-1)}{2}d$.
As like Example \ref{exa:1}, this is a contradiction.
Then $X$ does not have Galois points.
\end{exa}
From here, based on [\ref{bio:1}], we explain the orders of automorphisms of smooth plane curves of degree $d\geq 4$.
Let $X$ be a smooth plane curve of degree $d\geq4$, and $g$ be an automorphism of $X$.
By replacing the local coordinate system if necessary, we may assume that $g$ is defined by a diagonal matrix, i.e. $g=
\begin{pmatrix}
\alpha &0 &0\\
0 &\beta &0\\
0 & 0 &\gamma
\end{pmatrix}$.
Let
\[ n(g):=\sharp|{\rm Fix}(g)\cap\{[1:0:0],[0:1:0],[0:0:1]\}|.\]
Since $g$ is defined by a diagonal matrix, $n(g)=X\cap\{[1:0:0],[0:1:0],[0:0:1]\}$.
Then $n(g)=0,1,2$, or $3$.
The following Theorem \ref{thm:20} determines orders of cyclic groups acting on smooth plane curves.
Theorem \ref{thm:2} is shown by Theorems \ref{thm:8} and \ref{thm:20}.
For a smooth hypersurface $X\subset\mathbb P^{n+1}$,
orders of automorphisms of $X$ and the structure of the group Aut$(X)$ are studied for $n\geq 1$ ([\ref{bio:2},\ref{bio:8},\ref{bio:100},\ref{bio:101},\ref{bio:103},\ref{bio:104}]).
Also, as in [\ref{bio:191},\ref{bio:103}], the structures of subgroups of Aut$(X)$ are also investigated based on the way they act on $X$.
In this paper, we examine automorphisms of $X$ that give Galois points.
At the end of this section, we classify abelian groups acting on smooth plane curves (Theorem \ref{thm:21}).
\begin{thm}\label{thm:20}$([\ref{bio:1}])$.
Let $X$ be a smooth curve of degree $d\geq 4$ in $\mathbb P^2$, and $g$ be an automorphism of $X$.
By replacing the local coordinate system if necessary, the order of $g$ and a representation matrix of $g$ are one of Table 1.
\end{thm}
\begin{table}[h]
\caption{Cyclic groups of smooth plane curves of degree $d\geq 4$}
\label{table:data_type}
\centering
\begin{tabular}{|c|c|c|c|}
\hline
No.&$n(g)$&Order $l$ of $g$&Representation matrix of $g$ \\ \hline
1&$0$&$l$ divides $d$&
$\begin{pmatrix}
e_l^s &0 &0\\
0 &e_l^t &0\\
0 & 0 &1
\end{pmatrix}$\\ \hline
2&$1$&$l$ divides $d-1$&
$\begin{pmatrix}
e_l &0 &0\\
0 &1 &0\\
0 & 0 &1
\end{pmatrix}$\\ \hline
3&$1$&$l$ divides $(d-1)d$&
$\begin{pmatrix}
e_l &0 &0\\
0 &e_l^{1-d} &0\\
0 & 0 &1
\end{pmatrix}$\\ \hline
4&$2$&$l$ divides $d-1$&
$\begin{pmatrix}
e_l^s &0 &0\\
0 & e_l^t &0\\
0 & 0 &1
\end{pmatrix}$\\ \hline
5&$2$&$l$ divides $(d-1)^2$&
$\begin{pmatrix}
e_l^{1-d} &0 &0\\
0 & e_l &0\\
0 & 0 &1
\end{pmatrix}$\\ \hline
6&$2$&$l$ divides $(d-2)d$&
$\begin{pmatrix}
e_l &0 &0\\
0 & e_l^{1-d} &0\\
0 & 0 &1
\end{pmatrix}$\\ \hline
7&$3$&$l$ divides $d-1$&
$\begin{pmatrix}
e_l &0 &0\\
0 &1 &0\\
0 & 0 &1
\end{pmatrix}$\\ \hline
8&$3$&$l$ divides $d^2-3d+3$&
$\begin{pmatrix}
e_l &0 &0\\
0 & e_l^{d-1} &0\\
0 & 0 &1
\end{pmatrix}$\\ \hline
\end{tabular}
\end{table}
\begin{thm}\label{thm:21}
Let $X$ be a smooth plane curve of degree $d\geq4$, and $G$ be an abelian subgroup of Aut$(X)$.
If $G$ is not a cyclic group, then $G$ is isomorphic to a subgroup of $\mathbb Z/d\mathbb Z^{\oplus 2}$ as a group.
\end{thm}
\begin{proof}
Since $d\geq4$, $G$ is a finite subgroup of ${\rm PGL}(3,\mathbb C)$.
Let $l:=$max$\{{\rm ord}(k)\,|\,k\in G\}$.
We take an element $g\in G$ such that ord$(g)=l$.
By replacing the local coordinate system if necessary, we may assume that $g$ is defined by a diagonal matrix.
First, we assume that
$g=\begin{pmatrix}
\alpha &0 &0\\
0 &\alpha &0\\
0 & 0 &\beta
\end{pmatrix}$ where $\alpha,\beta\in\mathbb C^{\ast}$.
For simplicity, we may assume that
$\alpha=e_l$ and $\beta=1$.
Let $h$ be an element of $G$ such that $h\not\in \langle g\rangle$, and $A:=(a_{ij})_{1\leq i,j\leq3}$ be a representation matrix of $h$.
Since $g\circ h=h\circ g$, we get that
\[
\begin{pmatrix}
u &0 &0\\
0 &u &0\\
0 & 0 &u
\end{pmatrix}
\begin{pmatrix}
e_l &0 &0\\
0 & e_l &0\\
0 & 0 &1
\end{pmatrix}
\begin{pmatrix}
a_{11} &a_{12} &a_{13}\\
a_{21} &a_{22} &a_{23}\\
a_{31} &a_{32} &a_{33}
\end{pmatrix}
=
\begin{pmatrix}
a_{11} &a_{12} &a_{13}\\
a_{21} &a_{22} &a_{23}\\
a_{31} &a_{32} &a_{33}
\end{pmatrix}
\begin{pmatrix}
e_l &0 &0\\
0 & e_l &0\\
0 & 0 &1
\end{pmatrix},
\]
and hence
$
\begin{pmatrix}
ue_la_{11} &ue_la_{12} &ue_la_{13}\\
ue_la_{21} &ue_la_{22} &ue_la_{23}\\
ua_{31} &ua_{32} &ua_{33}
\end{pmatrix}
=
\begin{pmatrix}
e_la_{11} &e_la_{12} &a_{13}\\
e_la_{21} &e_la_{22} &a_{23}\\
e_la_{31} &e_la_{32} &a_{33}
\end{pmatrix}$
for $u\in \mathbb C^{\ast}$.
If $u\not=1$, then $a_{11}=a_{12}=a_{21}=a_{22}=a_{33}=0$.
Since $A$ is invertible, this is a contradiction.
Therefore, $u=1$.
Then $a_{13}=a_{23}=a_{31}=a_{32}=0$.
This means that
there is an injective homomorphism
\[\vartheta:G\ni k \mapsto C\in {\rm GL}(2,\mathbb C)\ \ {\rm such\ that}\ \ k=\begin{pmatrix}
C &0\\
0 &1
\end{pmatrix}.\]
Since $\vartheta(G)$ is an abelian but not cyclic subgroup of ${\rm GL}(2,\mathbb C)$,
there are two matrices $S_1,S_2\in {\rm GL}(2,\mathbb C)$ such that $\vartheta(G)=\langle S_1\rangle \oplus\langle S_2\rangle $.
In order to show $G\subset \mathbb Z/d\mathbb Z^{\oplus 2}$, we only show that ord$(g')$ is a divisor of $d$ for any $g'\in G$.
Since $G\cong \vartheta(G)=\langle S_1\rangle \oplus\langle S_2\rangle$, by replacing the local coordinate system if necessary, we may assume that $G$ is generated by two diagonal matrices.
We assume that $p:=[1:0:0]\in X$.
Since $G$ is generated by diagonal matrices, we get that $p\in {\rm Fix}(g)$ for any $g\in G$.
Since dim\,$X=1$, and $X$ is smooth, we get that $G$ is a cyclic group.
This contradicts that $G$ is not a cyclic group.
Therefore, we get that $[1:0:0]\not\in X$.
Similarly, we get that $[0:1:0],[0:0:1]\not\in X$.
Since $[1:0:0],[0:1:0],[0:0:1]\not\in X$, $X$ is defined by
\[ aX^d+bY^d+cZ^d+\sum_{i=0}^{d-1}F_{d-i}(Y,Z)X^i=0\]
where $abc\not=0$, $F_{d-i}(Y,Z)$ is a homogeneous polynomial of degree $d-i$ for $0\leq i\leq d-1$, and $F_0(Y,Z)$ has no $Y^d$ and $Z^d$ terms.
Then since $G$ is generated by diagonal matrices, we get that ord$(g')$ is a divisor of $d$ for any $g'\in G$.
Therefore, $G$ is a subgroup of $\mathbb Z/d\mathbb Z^{\oplus 2}$.
\\
Next, we assume that there is not an element $g'\in G$ such that a representation matrix of $g'$ is conjugate to
$
\begin{pmatrix}
\alpha' &0 &0\\
0 &\alpha' &0\\
0 & 0 &\beta'
\end{pmatrix}
$ where $\alpha',\beta'\in\mathbb C^{\ast}$.
Then we may assume that
$g=\begin{pmatrix}
e_l^s &0 &0\\
0 & e_l^t &0\\
0 & 0 &1
\end{pmatrix}$
where $e_l^s\not=e_l^t$, $e_l^s\not=1$, and $e_l^t\not=1$.
Let $h$ be an element of $G$ such that $h\not\in \langle g\rangle$, and $A:=(a_{ij})_{1\leq i,j\leq3}$ be a representation matrix of $h$.
Since $g\circ h=h\circ g$,
$
\begin{pmatrix}
ue_l^sa_{11} &ue_l^sa_{12} &ue_l^sa_{13}\\
ue_l^ta_{21} &ue_l^ta_{22} &ue_l^ta_{23}\\
ua_{31} &ua_{32} &ua_{33}
\end{pmatrix}
=
\begin{pmatrix}
e_l^sa_{11} &e_l^ta_{12} &a_{13}\\
e_l^sa_{21} &e_l^ta_{22} &a_{23}\\
e_l^sa_{31} &e_l^ta_{32} &a_{33}
\end{pmatrix}$
for $u\in \mathbb C^{\ast}$.
If $a_{ii}\not=0$ for some $1\leq i\leq 3$, then $u=1$.
Since $e_l^s\not=e_l^t$, $e_l^s\not=1$, and $e_l^t\not=1$, we get that $a_{ij}=0$ for $i\not=j$, i.e. $A$ is a diagonal matrix.
Since ord$(h)$ divides $l$, and $g$ and $h$ are defined by diagonal matrices,
we get that $\langle g,h\rangle$ contains an automoprhism $k$ such that a representation matrix of $k$ is conjugate to
$
\begin{pmatrix}
\alpha &0 &0\\
0 &\alpha &0\\
0 & 0 &\beta
\end{pmatrix}$ where $\alpha,\beta\in\mathbb C^{\ast}$.
This contradicts the assumption for $G$.
Therefore, $a_{ii}=0$ for any $i=1,2,3$.
Since $A$ is invertible, $a_{12}\not=0$ or $a_{13}\not=0$.
We assume that $a_{12}a_{13}\not=0$.
Then $ue_l^s=e_l^t$ and $ue_l^s=1$, and hence we get that $e_l^t=1$.
This contradicts the assumption that $e_l^t\not=1$.
Therefore, $a_{12}a_{13}=0$ and $(a_{12},a_{13})\not=(0,0)$.
In the same way,
$a_{21}a_{23}=a_{31}a_{32}=0$, $(a_{21},a_{23})\not=(0,0)$, and $(a_{31},a_{32})\not=(0,0)$.
Since $A$ is invertible,
\[
A=
\begin{pmatrix}
0 &a_{12} &0\\
0 &0 &a_{23}\\
a_{31} &0 &0
\end{pmatrix}\ {\rm or}\
\begin{pmatrix}
0 &0 &a_{13}\\
a_{21} &0 &0\\
0 &a_{32} &0
\end{pmatrix}.
\]
If $A$ is the former,
then $ue_l^s=e_l^t$, $ue_l^t=1$, and $u=e_l^s$.
Therefore, we get that ${e_l^s}^3={e_l^t}^3=u^3=1$.
In the same way, for the latter case, we get that ${e_l^s}^3={e_l^t}^3=u^3=1$.
Therefore, we may assume that
$
g=\begin{pmatrix}
e_3^2 &0 &0\\
0 & e_3 &0\\
0 & 0 &1
\end{pmatrix}$, and for an automorphism $k\in G\backslash\langle g\rangle$,
$k$ is defined by a matrix of the form:
\[
\begin{pmatrix}
0 &b_{12} &0\\
0 &0 &b_{23}\\
b_{31} &0 &0
\end{pmatrix}\ {\rm or}\
\begin{pmatrix}
0 &0 &b_{13}\\
b_{21} &0 &0\\
0 &b_{32} &0
\end{pmatrix}.
\]
Note that the square of the former (resp. latter) form of the matrix is of the latter (resp. former) form of the matrix.
From here, we show that $G\cong\mathbb Z/3\mathbb Z^{\oplus 3}$, and the degree $d$ of $X$ is a multiple of $3$.
We assume that there are two automorphisms $h_1,h_2\in G$ such that
\[ h_1=
\begin{pmatrix}
0 &a &0\\
0 &0 &b\\
c & 0 &0
\end{pmatrix},\ \
h_2=
\begin{pmatrix}
0 &a' &0\\
0 &0 &b'\\
c' &0 &0
\end{pmatrix},\]
and $h_1\not\in\langle h_2\rangle$.
Then
\[
h_1^2\circ h_2=\begin{pmatrix}
0 &a &0\\
0 &0 &b\\
c & 0 &0
\end{pmatrix}
\begin{pmatrix}
0 &a &0\\
0 &0 &b\\
c & 0 &0
\end{pmatrix}
\begin{pmatrix}
0 &a' &0\\
0 &0 &b'\\
c' &0 &0
\end{pmatrix}=
\begin{pmatrix}
abc' &0 &0\\
0 &a'bc &0\\
0 & 0 &abc'
\end{pmatrix}.
\]
Since $G$ is abelian, and ord$(h_i)=3$ for $i=1,2$, we get that ord$(h_1^2\circ h_2)=3$.
Since ord$(g)=3$, and the assumption for $G$, we get that $h_1^2\circ h_2\in \langle g\rangle$.
Therefore, $G=\langle g, h\rangle\cong \mathbb Z/3\mathbb Z^{\oplus 3}$ where
\[
g=\begin{pmatrix}
e_3^2 &0 &0\\
0 & e_3 &0\\
0 & 0 &1
\end{pmatrix}\ {\rm and}\
h=
\begin{pmatrix}
0 &a &0\\
0 &0 &b\\
c & 0 &0
\end{pmatrix}.
\]
Since $h([1:0:0])=[0:0:1]$ and $h^2([1:0:0])=[0:1:0]$,
if $\{[1:0:0],[0:1:0],[0:0:1]\}\cap X\not=\emptyset$,
then $\{[1:0:0],[0:1:0],[0:0:1]\}\subset X$, i.e. $n(g)=3$.
By Table 1 and a representation matrix of $g$, we get that $3$ divides $d$.
Then $G$ is a subgroup of $\mathbb Z/d\mathbb Z^{\oplus 2}$.
We assume that $\{[1:0:0],[0:1:0],[0:0:1]\}\cap X=\emptyset$.
By Table 1 and a representation matrix of $g$,
we get that ord$(g)=3$ divides $d^2-3d+3$, and hence $3$ divides $d$.
Therefore, $G$ is a subgroup of $\mathbb Z/d\mathbb Z^{\oplus 2}$.
\end{proof}
\section{Proof of main theorems}
First, we will show Theorem \ref{thm:2} (Theorem \ref{thm:9}).
Theorem \ref{thm:2} is immediately followed by Theorems \ref{thm:8} and \ref{thm:20}.
\begin{thm}\label{thm:9}
Let $X$ be a smooth plane curve degree $d\geq 4$, and $g$ be an automorphism of $X$.\\
$(1)$ If ${\rm ord}(g)=d-1$ and $\sharp|{\rm Fix}(g)|\not=2$, then $X$ has an inner Galois point $p$, and $g$ is an automorphism belonging to the Galois point $p$.\\
$(2)$ If ${\rm ord}(g)=d$ and ${\rm Fix}(g)\not=\emptyset$, then $X$ has an outer Galois point $p$, and $g$ is an automorphism belonging to the Galois point $p$.
\end{thm}
\begin{proof}
Since $d\geq4$, ${\rm Aut}(X)$ is a subgroup of ${\rm PGL}(3,\mathbb C)$.
We will show (1) of this theorem.
Since ${\rm ord}(g)=d-1$, by replacing the local coordinate system if necessary, we may assume that $g$ is defined by a diagonal matrix $A$ such that $A$ is one of no.2, 3, 4, 5, and 7 of Table 1.
By Theorem \ref{thm:8}, if $A$ is one of no.2, 3, 5, and 7 of Table 1, then $X$ has an inner Galois point $p$, and $g$ is an automorphism belonging to the Galois point $p$.
We assume that $A$ is no.4 of Table 1, i.e. $A=
\begin{pmatrix}
e_{d-1}^s &0 &0\\
0 & e_{d-1}^t &0\\
0 & 0 &1
\end{pmatrix}$ where $1\leq s,t<d-1$.
Then $X\cap \{[1:0:0],[0:1:0],[0:0:1]\}\subset{\rm Fix}(g)$ and $\sharp|X\cap \{[1:0:0],[0:1:0],[0:0:1]\}|=2$.
Since $\sharp|{\rm Fix}(g)|\not=2$, $\sharp|{\rm Fix}(g)|\geq3$.
Then we get that $s=t$, $s=1$, or $t=1$.
By Theorem \ref{thm:8}, $X$ has an inner Galois point $p$, and $g$ is an automorphism belonging to the Galois point $p$.
In the same way, we get (2) of this theorem.
\end{proof}
Let $X$ be a smooth hypersurface of degree $d\geq 4$ in $\mathbb P^{n+1}$, $p$ be a point in $\mathbb P^{n+1}$.
Recall that $\pi_p:X\dashrightarrow H$ is a projection with center $p$ where $H$ is a hyperplane not containing $p$.
The following result is obtained for an inner Galois point ([\ref{bio:20}]).
\begin{thm}\label{cro:0}([\ref{bio:20}]).
Let $X$ be a smooth plane curve degree $d\geq 4$, and $\mathbb C(X)$ be the function field of $X$, and $k\subset \mathbb C(X)$ be a subfield such that $k$ is isomorphic to $\mathbb C(\mathbb P^1)$ as a field.
If $\mathbb C(X)/k$ is a Galois extension of degree $d-1$, then $X$ has an inner Galois point $p$, and the Galois extension $\mathbb C(X)/k$ is induced by $\pi_p:X\rightarrow \mathbb P^1$, i.e. $k=\pi_p^{\ast}(\mathbb C(\mathbb P^1))$.
\end{thm}
In the case of the outer Galois point, by Example \ref{ex1}, we see that the same result as in Theorem \ref{cro:0} does not hold.
\begin{exa}\label{ex1}
Let $X$ be a smooth curve of degree $4$ in $\mathbb P^2$ defined by
\[X_0^4+X_1^4+X_2^4=0\]
which is called the Fermat curve of degree $4$.
The $X$ has two automorphism $g_1$ and $g_2$ of order $2$ such that
the followimg matrices $A_1$ and $A_2$ are representation matrices of $g_1$ and $g_2$, respectively
\[
A_1:=
\begin{pmatrix}
-1&0&0\\
0&1&0\\
0&0&1
\end{pmatrix}\ {\rm and}\
A_2:=
\begin{pmatrix}
1&0&0\\
0&-1&0\\
0&0&1
\end{pmatrix}.
\]
Let $G$ be the subgroup of Aut$(X)$ generated by $g_1$ and $g_2$, and $g_3:=g_1\circ g_2\in G$.
Then $G\cong \mathbb Z/2\mathbb Z^{\oplus 2}$, and $\sharp|{\rm Fix}(g_i)|=4$ for $i=1,2,3$.
Let $G_x:=\{g\in G:\,g(x)=x\}$.
For a smooth curve $C$, we write the genus of $C$ as $g(C)$.
By the Riemann-Hurwitz formula,
\[ 2-2g(X)+\sum_{x\in X}(\sharp|G_x|-1)=\sharp|G|(2-2g(X/G)).\]
Since $X$ is a smooth plane curve of degree $4$, we get that $2-2g(X)=4(3-4)=-4$.
Then
\[ 2-2g(X)+\sum_{x\in X}(\sharp|G_x|-1)=-4+12=8.\]
Since $\sharp|G|=4$, and the Riemann-Hurwitz formula, we get that $g(X/G)=0$, and hence $X/G\cong \mathbb P^1$.
Let $p:X\rightarrow X/G$ be the quotient morphism.
Since $G$ is not cyclic group, the Galois extension $\mathbb C(X)/p^{\ast}\mathbb C(\mathbb P^1)$ is not induced by a Galois point of $X$.
\end{exa}
The following theorem shows that similar results hold for an outer Galois point under the assumption of a cyclic extension.
\begin{thm}\label{cro:1}
Let $X$ be a smooth plane curve degree $d\geq 4$, and $\mathbb C(X)$ be the function field of $X$, and $k\subset \mathbb C(X)$ be a subfield such that $k$ is isomorphic to $\mathbb C(\mathbb P^1)$ as a field.
If $\mathbb C(X)/k$ is a cyclic extension of degree $d$, then $X$ has an outer Galois point $p$, and the cyclic extension $\mathbb C(X)/k$ is induced by $\pi_p:X\rightarrow \mathbb P^1$, i.e. $k=\pi_p^{\ast}(\mathbb C(\mathbb P^1))$.
\end{thm}
\begin{proof}
Since $X$ is a smooth curve, there is a cyclic subgroup $G$ of ${\rm Aut}(X)$ such that $X/G\cong \mathbb P^1$, and $k=p^{\ast}\mathbb C(\mathbb P^1)$ where $p:X\rightarrow X/G$ be the quotient morphism.
Since $d\geq4$, $G$ is a subgroup of ${\rm PGL}(3,\mathbb C)$.
Let $g$ be a generator of $G$.
By replacing the local coordinate system if necessary, we assume that there is a diagonal matrix $A$ such that $A$ is a representation matrix of $g$.
Since ord$(g)=d$ and Theorem \ref{thm:9}, we only show that Fix$(g)\not=\emptyset$.
We assume that Fix$(g)=\emptyset$.
By Theorem \ref{thm:8}, that is, by the no.1 of Table 1, we may assume that
$A=
\begin{pmatrix}
e_d^s &0 &0\\
0 & e_d^t &0\\
0 & 0 &1
\end{pmatrix}$.
Since Fix$(g)=\emptyset$,
$X\cap \{[1:0:0],[0:1:0],[0:0:1]\}=\emptyset$.
Then if Fix$(g^i)\not=\emptyset$ for some $1<i<d$, then $\sharp|{\rm Fix}(g^i)|=d$.
By the Riemann-Hurwitz formula and $C/G\cong \mathbb P^1$,
\[ 2-2g(X)+\sum_{x\in X}(\sharp|G_x|-1)=2\sharp|G|=2d\]
Since $X$ is a smooth plane curve of degree $d$, we get that $2-2g(X)=d(3-d)$, and hence
By the matrix $A$, we get that ${\rm Fix}(g^i)\backslash \{[1:0:0],[0:1:0]\}\not=\emptyset$ if and only if $(e_{d-1}^{si}-e_{d-1}^{ti})(e_{d-1}^{si}-1)(e_{d-1}^{ti}-1)=0$ for $1< i<d$.
We define subgroups $G_1$, $G_2$, and $G_3$ of $G$ as follows:
\[ G_1:=\{g\in G\,|\,{\rm a\ representation\ matrix\ of}\ g\ {\rm is}\ \begin{pmatrix}
\alpha &0 &0\\
0 &1 &0\\
0 &0 &1
\end{pmatrix}\ {\rm for\ some}\ \alpha\in\mathbb C^{\ast}
\}.\]
\[ G_2:=\{g\in G\,|\,{\rm a\ representation\ matrix\ of}\ g\ {\rm is}\ \begin{pmatrix}
1 &0 &0\\
0 &\beta &0\\
0 &0 &1
\end{pmatrix}\ {\rm for\ some}\ \beta\in\mathbb C^{\ast}
\}.\]
\[G_3:=\{g\in G\,|\,{\rm a\ representation\ matrix\ of}\ g\ {\rm is}\ \begin{pmatrix}
\gamma &0 &0\\
0 &\gamma &0\\
0 &0 &1
\end{pmatrix}\ {\rm for\ some }\ \gamma\in\mathbb C^{\ast}
\}.\]
We set $a:=\sharp|G_1|$, $b:=\sharp|G_2|$, and $c:=\sharp|G_3|$.
Then $G_i\cap G_j=\{{\rm id}_X\}$ for $1\leq i<j\leq 3$, and ${\rm Fix}(g^i)\not=\emptyset$ if and only if $g^i\in \bigcup_{j=1}^3G_j$ for $1<i<d$.
Then
\[ (d-1)d=\sum_{x\in C}(\sharp|G_x|-1)=d(a+b+c-3).\]
Therefore,
\[ d+2=a+b+c.\]
For simplicity, we may assume that $a\leq b\leq c$.
Since $d+2=a+b+c$, $1<c$.
Since $G_2\cap G_3=\{{\rm id}_X\}$ and $\sharp|G|=d$, we get that $bc|d$.
By the equation $d+2=a+b+c$, we get that $bc+2\leq a+b+c\leq b+2c$, and hence $(b-2)(c-1)\leq 0$.
Since $1<c$, $b\leq 2$.
If $b=2$, then by the equation $bc+2\leq a+b+c$, we get that $a=b=c$.
Since $G_i\cap G_i=\{{\rm id}_X\}$ for $1\leq i<j\leq3$, we get that $\mathbb Z_2^{\oplus 2}\cong \langle G_i,G_j\rangle\subset G$ where $1\leq i<j\leq3$, and
$\langle G_i,G_j\rangle$ is the subgroup of $G$ generated by $G_i$ and $G_j$.
This contradicts that $G$ is a cyclic group.
If $b=1$, then $a=1$ and $c=d$.
This implies that $G=\langle g\rangle =G_3$.
This contradicts that $G=\langle g\rangle$ and Fix$(g)=\emptyset$.
Therefore, Fix$(g)\not=\emptyset$.
By Theorem \ref{thm:9}, $X$ has an outer Galois point $p$, and $g$ is an automorphism belonging to the Galois point $p$.
\end{proof}
From here, we will study $X\subset \mathbb P^{n+1}$ for $n\geq 2$.
First, we give Examples \ref{exa:4} and \ref{exa:5} which imply that Corollary \ref{cro:1} does not hold for $n=2$.
\begin{exa}\label{exa:4}
Let $X$ be a smooth surface of degree $4$ in $\mathbb P^3$ defined by
\[X_0^3X_2+X_1^3X_3+X_2^4+X_3^4=0.\]
The surface $X$ has an automorphism $g$ of order $3$ such that
\[g=
\begin{pmatrix}
e_3&0&0&0\\
0&e_3^2&0&0\\
0&0&1&0\\
0&0&0&1
\end{pmatrix}.\]
Then Fix$(g)$ contains a smooth rational curve.
Since the degree of $X$ is $4$, $X$ is a $K3$ surface.
Since Fix$(g)$ contains a curve, $g$ is a non-symplectic automorphism of order $3$.
Then the quotient space $Y:=X/\langle g \rangle$ is rational.
Let $q:X\rightarrow Y$ be the quotient morphism.
Since $Y$ is rational $k:=q^{\ast}\mathbb C(Y)\cong \mathbb C(\mathbb P^2)$ as a field.
However, by Theorem \ref{thm:8}, there is no a Galois point $p$ of $X$ such that $g$ is an automorphism belonging to the Galois point $p$.
In other words, there is no a Galois point $p$ of $X$ such that $k=\pi_p^{\ast}\mathbb C(\mathbb P^2)$.
Pay attention that $X$ has Galois points, and $\delta(X)=8$ ([\ref{bio:21}]).
\end{exa}
\begin{exa}\label{exa:5}
Let $X$ be a smooth surface in $\mathbb P^3$ defined by
\[X_0^6+X_1^6+X_2^6+X_3^6+X_0^2X_1^3X_2+X_2^3X_3^3=0.\]
The surface $X$ has an automorphism $g$ of order $6$ such that
\[g=
\begin{pmatrix}
-1&0&0&0\\
0&e_3&0&0\\
0&0&1&0\\
0&0&0&1
\end{pmatrix}.\]
Fix$(g^3)=\{X_0=0\}\cap X:=H_1$ and Fix$(g^2)=\{X_1=0\}\cap X:=H_2$ are smooth curves, and Fix$(g)=H_1\cap H_2$.
Then the quotient space $Y:=X/\langle g\rangle$ is smooth.
Let $p:X\rightarrow Y$ be the quotient morphism, and $\mathcal O_X(1):=\mathcal O_{\mathbb P^{3}}(1)$ be the ample line bundle.
By the ramification formula, $K_X=p^{\ast}K_Y+H_1+2H_2$, and hence $p^{\ast}K_Y=K_X-H_1-2H_2$.
Since $K_X=\mathcal O_X(2)$, and $\mathcal O_X(H_i)=\mathcal O_X(1)$ for $i=1,2$,
we get that $p^{\ast}\mathcal O_Y(-K_Y)=\mathcal O_X(1)$ is ample.
Since the morphism $p:X\rightarrow Y$ is finite, $-K_Y$ is ample.
Since $Y$ is a smooth surface, $Y$ is rational, and hence $k:=q^{\ast}\mathbb C(Y)\cong \mathbb C(\mathbb P^2)$ as a field.
However, by Theorem \ref{thm:8}, there is no a Galois point $p$ of $X$ such that $k=\pi_p^{\ast}\mathbb C(\mathbb P^2)$.
\end{exa}
We will show Theorems \ref{thm:3} and \ref{thm:4} (Theorem \ref{thm:10}).
Recall that for a smooth hypersurface $X\subset\mathbb P^{n+1}$ of degree $d\geq 4$, if $(n,d)\not=(2,4)$, then all automorphisms of $X$ are linear.
\begin{thm}\label{thm:10}
Let $X$ be a smooth hypersurface of degree $d\geq 4$ in $\mathbb P^{n+1}$, and $g$ be a linear automorphism of $X$.\\
(1) If $n=2$, ${\rm ord}(g)=d-1$, and ${\rm Fix}(g)$ contains a curve $C'$ which is not a smooth rational curve,
then $X$ has an inner Galois point $p$, and $g$ is an automorphism belonging to the Galois point $p$.\\
(2) If $n\geq 3$, ${\rm ord}(g)=d-1$, and ${\rm Fix}(g)$ has codimension $1$ in $X$, then $X$ has an inner Galois point $p$, and $g$ is an automorphism belonging to the Galois point $p$.\\
(3) If $n\geq 2$, ${\rm ord}(g)=d$, and ${\rm Fix}(g)$ has codimension $1$ in $X$, then $X$ has an outer Galois point $p$, and $g$ is an automorphism belonging to the Galois point $p$.
\end{thm}
\begin{proof}
By replacing the local coordinate system if necessary, we may assume that
\[g=
\begin{pmatrix}
a_{i_1}I_{i_1} & &\\
&\ddots &\\
& &a_{i_m}I_{i_m}
\end{pmatrix}\]
where $I_{i_j}$ is the identity matrix of size $i_j$, $a_{i_j}\in\mathbb C^{\ast}$, $a_{i_j}\not=a_{i_k}$ for $1\leq i_j,i_k\leq m$, and $\sum_{j=1}^mi_j=n+2$.
We assume that Fix$(g)$ contains a hypersurface $H$ in $X$.
Since dim\,$H=n-1$, $i_j\geq n-1$ for some $1\leq j\leq m$.
Then we may assume that
\[g=
\begin{pmatrix}
a &0 \\
0&I_{n+1}
\end{pmatrix}\ {\rm or}\
\begin{pmatrix}
a &0&0 \\
0&b&0\\
0&0&I_{n}
\end{pmatrix}.\]
If $g$ is defined by the former matrix, then by Theorem \ref{thm:8} $X$ has a Galois point $p$, and $g$ is an automorphism belonging to the Galois point $p$.
From here, we will show that if $g$ is defined by the latter matrix, then $n=2$, ${\rm ord}(g)=d-1$, and curves contained in Fix$(g)$ are $\mathbb p^1$.
By the representation matrix of $g$, $H=\{X_0=0\}\cap \{X_1=0\}$.
Let $F(X_0,\ldots ,X_{n+2})$ be the defining equation of $X$.
Since $H=\{X_0=0\}\cap \{X_1=0\}$,
\begin{equation*}
\begin{split}
F(X_0,\ldots ,X_{n+2})=&F_{1,0}(X_2,\ldots,X_{n+2})X_0+F_{0,1}(X_2,\ldots,X_{n+2})X_1\\
&\ \ +\sum_{2\leq i+j\leq d}F_{i,j}(X_2,\ldots,X_{n+2})X^i_0X^j_1.
\end{split}
\end{equation*}
Since $X$ is smooth, $\{F_{1,0}(X_2,\ldots,X_{n+2})=0\}\cap\{F_{0,1}(X_2,\ldots,X_{n+2})=0\}\cap\{X_0=0\}\cap\{X_1=0\}=\emptyset$.
Therefore, $n=2$, curves of Fix$(g)$ are $\mathbb p^1$, $F_{1,0}(X_2,\ldots,X_{n+2})\not=0$, and $F_{0,1}(X_2,\ldots,X_{n+2})\not=0$.
Then $a=b$.
If ${\rm ord}(g)=d$, then $a=b=e_d$.
Then the defining equation of $X$ is as follows.
\[ F(X_0,\ldots ,X_{n+2})=F_{1,0}(X_2,\ldots,X_{n+2})X_0+F_{0,1}(X_2,\ldots,X_{n+2})X_1.\]
Points $[1:0:0:0]$ and $[0:1:0:0]$ are singular points of $X$.
This contradicts that $X$ is smooth.
Therefore, ${\rm ord}(g)=d-1$.
\end{proof}
In the same way, we get Theorem \ref{thm:5} (Theorem \ref{thm:11}).
\begin{thm}\label{thm:11}
Let $X$ be a smooth hypersurface of degree $d$ in $\mathbb P^{n+1}$, $g\in{\rm Aut}(X)$ be a linear automorphism of order $k(d-1)$ for $k\geq 2$.\\
(1) If $n=2$ and $\sharp|{\rm Fix}(g)|\geq5$, then $X$ has an inner Galois point $p$, and $g^k$ is an automorphism belonging to the Galois point $p$.\\
(2) If $n\geq 3$, and the dimension of ${\rm Fix}(g)$ is $n-2$, then $X$ has an inner Galois point $p$, and $g^k$ is an automorphism belonging to the Galois point $p$.
\end{thm}
\begin{proof}
As like the proof of Theorem \ref{thm:10}, we may assume that
\[g=
\begin{pmatrix}
a&0&0 \\
0&b&0\\
0&0&I_{n}
\end{pmatrix}\ {\rm or}\
\begin{pmatrix}
a&0&0&0\\
0&b&0&0\\
0&0&c&0\\
0&0&0&I_{n-1}
\end{pmatrix}\]
where $a$, $b$, $c$, and $1$ are different numbers from each other.
First, we will show that if $g$ is defined by the former matrix, then $X$ has an inner Galois point $p$, and
$g^k$ is an automorphism belonging to the Galois point $p$.
Let $F(X_0,\ldots ,X_{n+2})$ be the defining equation of $X$.
Since dim${\rm Fix}(g)=n-2$,
\[
F(X_0,\ldots ,X_{n+2})=\sum_{1\leq i+j\leq d}F_{i,j}(X_2,\ldots,X_{n+2})X^i_0X^j_1+G(X_2,\ldots,X_{n+2})\]
where $G(X_2,\ldots,X_{n+2})\not=0$.
Let $n(g):=\sharp|\{[1:0:\cdots:0],[0:1:0:\cdots:0]\}\cap X|$.
If $n(g)=0$, then $\sum_{1\leq i+j\leq d}F_{i,j}(X_2,\ldots,X_{n+2})X^i_0X^j_1$ has $X_0^d$ and $X_1^d$ terms.
Since $G(X_2,\ldots,X_{n+2})\not=0$, $a^d=b^d=1$.
This contradicts that ord$(g)>d$.
If $n(g)=1$, then we may assume that $\sum_{1\leq i+j\leq d}F_{i,j}(X_2,\ldots,X_{n+2})X^i_0X^j_1$ has (i) $X_0^d$ and $X_iX_1^{d-1}$ terms, or (ii) $X_0^d$ and $X_0X_1^{d-1}$ terms where $2\leq i\leq n+2$.
The case (i) implies that $a^d=b^{d-1}=1$.
By Theorem \ref{thm:8}, there is an inner Galois point $p$ of $X$, and $g^k$ is an automorphism belonging to the Galois point $p$.
The case (ii) implies that $a^d=ab^{d-1}=1$.
Same as above, $X$ has an inner Galois point $p$, and $g^k$ is an automorphism belonging to the Galois point $p$.
If $n(g)=2$, then we may assume that $\sum_{1\leq i+j\leq d}F_{i,j}(X_2,\ldots,X_{n+2})X^i_0X^j_1$ has (iii) $X_iX_0^d$ and $X_iX_1^{d-1}$ terms, (iv) $X_iX_0^d$ and $X_0X_1^{d-1}$,
or (v) $X_1X_0^{d-1}$ and $X_0X_1^{d-1}$ terms where $2\leq i,j\leq n+2$.
The case (iii) implies that $a^{d-1}=b^{d-1}=1$.
This contradicts that ord$(g)>d-1$.
As like the case $n(g)=1$, if the case is (iv), then
by Theorem \ref{thm:8}, there is an inner Galois point $p$ of $X$, and $g^k$ is an automorphism belonging to the Galois point $p$.
The case (v) implies that $a^{d-1}b=ab^{d-1}=1$.
Then ord$(g)$ divides $(d-2)d$.
This contradicts that ord$(g)=k(d-1)$.
From here, we study the latter case, i.e. $g=
\begin{pmatrix}
a&0&0&0\\
0&b&0&0\\
0&0&c&0\\
0&0&0&I_{n-1}
\end{pmatrix}$.
As like the proof of Theorem \ref{thm:10}, we get that $n\leq3$.
We assume that $n=3$.
Let $F(X_0,\ldots ,X_5)$ be the defining equation of $X$.
Since the dimension of ${\rm Fix}(g)$ is $n-2$,
\begin{equation*}
\begin{split}
F(X_0,\ldots ,X_{n+2})=&\sum_{i=0}^2F_i(X_3,\ldots,X_{n+2})X_i\\
&\ \ +\sum_{2\leq i+j+k\leq d}F_{i,j,k}(X_3,\ldots,X_{n+2})X^i_0X^j_1X_2^k.
\end{split}
\end{equation*}
Since $X$ is smooth, $F_i(X_3,\ldots,X_{n+2})\not=0$ for $i=0,1,2$.
Then $a=b=c$.
This contradicts that ord$(g)=k(d-1)$ for $k\geq2$.
Then $n=2$, and hence
$g=
\begin{pmatrix}
a&0&0&0\\
0&b&0&0\\
0&0&c&0\\
0&0&0&1
\end{pmatrix}$.
Since $a$, $b$, $c$, and $1$ are different numbers from each other,
$\sharp|{\rm Fix}(g)|\leq 4$.
From the above, we get this theorem.
\end{proof}
The following example shows that Theorem \ref{thm:11} does not hold for an outer Galois point.
\begin{exa}\label{exa:6}
Let $d_1\geq 7$ be an odd integer, and $d:=2d_1+1$.
Let $X$ be a smooth hypersurface of degree $d$ in $\mathbb P^4$ defined by
\[X_0^d+X_0^{\frac{d+1}{2}}X_1^{\frac{d-1}{2}}+X_0X_1^{d-1}+
X_2^{d-1}X_4+X_2X_3^{d-1}+X_3X_4^{d-1}=0.\]
The $X$ has an automorphism $g$ of order $\frac{(d-1)}{2}d$ such that
the following matrix $A$ is a representation matrix of $g$:
\[
A:=
\begin{pmatrix}
e_{\frac{(d-1)}{2}d}^{1-d}&0&0&0&0\\
0&e_{\frac{(d-1)}{2}d}&0&0&0\\
0&0&1&0&0\\
0&0&0&1&0\\
0&0&0&0&1
\end{pmatrix}.
\]
Then the dimension of Fix$(g^{d^2-3d+3})$ is $1$.
In addition,
$X$ has an automorphism $h$ such that the following matrix $B$ is a representation matrix of $h$:
\[
B:=
\begin{pmatrix}
e_{\frac{(d-1)}{2}d}^{1-d}&0&0&0&0\\
0&e_{\frac{(d-1)}{2}d}&0&0&0\\
0&0&e_{d^2-3d+3}&0&0\\
0&0&0&e_{d^2-3d+3}^{d-1}&0\\
0&0&0&0&1
\end{pmatrix}.
\]
If $3$ divides $d$, then ord$(h)=\frac{(d-1)}{6}d(d^2-3d+3)$, and if $3$ does not divide $d$, then ord$(h)=\frac{(d-1)}{2}d(d^2-3d+3)$.
For $1\leq i<\frac{d-1}{2}$, the diagonal entries of $B^i$ are different from each other.
By Theorem \ref{thm:6}, $\delta(X)\leq 2$ and $\delta'(X)\leq5$.
Since $\frac{d-1}{2}\geq 7$, if $X$ has a Galois point, then there is a Galois point $p$ of $X$ such that $g^l(p)=p$ for some $1\leq l<\frac{(d-1)}{2}d$.
As like Example \ref{exa:1}, this is a contradiction.
Then $X$ does not have Galois points.
\end{exa}
|
\section{Introduction \label{introduction}}
One of the fundamental postulates of quantum theory is that the Hamiltonian of an isolated system is Hermitian.
This Hermiticity seems to be a compelling postulate because it ensures that the eigenvalues of the Hamiltonian are real. Moreover, a Hermitian Hamiltonian leads to a unitary time evolution and consequently the conservation of probability. However, decoherence effects are ubiquitous in nature because the physical systems are always inevitably affected by the environment. Under this condition, the dynamics of the system is dominated by a Lindbladian generalized master equation. Because accurately solving this master equation is usually a very difficult task, especially in high-dimensional systems, an approximate but fruitful approach \cite{nielsen2010quantum} to describe the dynamics of open quantum systems is directly handling the Schr\"{o}dinger equation $ i\partial_{t} \ket{\psi}=H \ket{\psi}$,
such that the time evolution operator is determined by an effective time-dependent Hamiltonian $ H $ which is not necessarily Hermitian \cite{daley2014quantum,dalibard1992wave}. Non-Hermitian Hamiltonians usually have complex eigenvalue spectra and do not conserve probabilities, and therefore they often only serve as phenomenological descriptions of an open quantum system. Nevertheless, there is a certain class of non-Hermitian Hamiltonians, invariant under a combination of parity ($ \mathcal{P} $) and time-reversal ($ \mathcal{T} $) operations, i.e., $ [\mathcal{PT},H] =0$, whose spectrum can be entirely real as long as they respect $ \mathcal{PT} $ symmetry
\cite{bender2002complex,bender2003must,ahmed2001real,bender2007faster,znojil2001pt}.
Some applications and features of
$ \mathcal{PT} $ symmetry are addressed in Refs. \cite{el2018non,feng2017non,jing2015optomechanically,zhang2015giant,quijandria2018pt,arkhipov2019scully,huybrechts2020validity,zhang2020breaking,arkhipov2020liouvillian,PhysRevA.98.022117,PhysRevA.101.033820}.
In general, $ \mathcal{PT} \!-$symmetric systems exhibit two phases: the unbroken phase in which the entire eigenspectrum is real, and the
broken phase where some or all of the eigenvalues form complex
conjugate pairs. This \textit{phase transition}
occurs at an special point where $ n $ eigenvalues, as
well as their corresponding eigenvectors, coalesce \cite{kato2013perturbation,heiss2012physics,ozdemir2019parity}.
This non-Hermitian degeneracy, also known as exceptional point (EP) of order n (EPn),
has recently emerged as a new way to engineer the response of the open
physical system.
In optics, the abrupt nature
of the phase transitions being encountered around EPs has been demonstrated to lead to many intriguing phenomena, such
as unidirectional invisibility \cite{lin2011unidirectional,peng2014parity}, loss-induced transparency \cite{guo2009observation}, band
merging \cite{zhen2015spawning,makris2008beam}, laser mode selectivity \cite{hodaei2014parity,feng2014single}, topological chirality \cite{doppler2016dynamically,xu2016topological}, new types of thresholdless phonon lasers \cite{jing2014pt,lu2017exceptional}, and even exceptional photon
blockade \cite{huang2020exceptional}.
These important phenomena have been experimentally observed in different platforms based on optomechanics \cite{jing2014pt}, electronics \cite{schindler2011experimental}, metamaterials \cite{kang2013effective}, acoustics \cite{zhu2014p,fleury2015invisible},
and plasmonics \cite{benisty2011implementation}.
Moreover, recently, it has been demonstrated that the bifurcation properties of
second-order non-Hermitian degeneracies can be used as an efficient tool to improve
the sensitivity (frequency shifts) of resonant optical
structures to external perturbations \cite{wiersig2014enhancing}. In addition, it is of particular interest \cite{hodaei2017enhanced} to use even higher-order EPs (greater than second order), which in principle could considerably amplify the effect of perturbations, leading to greater sensitivity beyond what is possible in standard
arrangements \cite{wiersig2020review}.
Recently, anti-$ \mathcal{PT} \!-$symmetric systems where the Hamiltonian is anti-commutative with
the joint $\mathcal{PT}$ operator, i.e., $ \{H,\mathcal{PT}\}=0 $, have attracted much research
interest. Interesting physical phenomena reported in anti-$\mathcal{PT}\!-$symmetric systems include optical
systems with constant refraction \cite{yang2017anti} and optical
materials with balanced positive and negative index \cite{ge2013antisymmetric}. Moreover, several relevant experiments have been realized in diffusive systems \cite{li2019anti},
electrical circuit resonators, and atomic \cite{peng2016anti,chuang2018realization,wang2016optical} or molecular systems \cite{jingwei2020observation}. Additionally, quantum processes such as the observation of EPs \cite{choi2018observation}, symmetry-breaking transition \cite{peng2016anti}, and simulation of anti-$\mathcal{PT}\!-$symmetric Lorentz dynamics \cite{li2019experimental}, which are important phenomena different from Hermitian quantum mechanics, have been addressed in these experiments.
Decoherence control is a key task for practical implementations of nanoscale solid-state quantum information processing, in which the decoherence may be mainly affected by non-Markovian dynamics due to the strong backactions from the environment \cite{breuer2002theory,lidar2019lecture,chen2015using,chen2009detecting,yin2012spin,zhang2012general,
chen2016quantifying,lu2020observing,jahromi2019quantum,jahromi2019multiparameter,PhysRevA.95.052126,jahromi2015precision,NMexp5,LoFrancoNatCom,darrigo2012AOP,
GholipourAnnPhys,mortezapourOpt,lofrancoreview,bellomo2007non,Mortezapour_2017,NM1,NMexp1,
NMexp4,PhysRevA.85.032318,PhysRevA.92.012315,naikoo2019facets,PhysRevA.92.032311,teittinen2018revealing,D_Arrigo_2013,Bellomo_2012}. A fundamental issue is how to accurately define Markovianity in quantum processes \cite{breuer2016colloquium,rivas2014quantum}. In order to present a definition of a Markov process in the quantum regime, it is essential to understand the concept of
Markov process in the classical scenario.
\par
A classical \textit{Markov process}
is defined as a family of random
variables $ \{X(t), t \in I \subset \mathbb{R}\} $, for which the probability that $ X $ takes a value $ x_{n} $ at any arbitrary time $ t_{n} \in I$,
if it took value $ x_{n-1} $ at some previous time
$ t_{n-1} < t_{n} $, can be uniquely determined, and is not influenced
by the possible values of $ X $ at previous times to $ t_{n-1} $. One can
formulate it in terms of conditional probabilities as follows: $ \mathbb{P}(x_{n}, t_{n}|x_{n-1}, t_{n-1};...; x_{0}, t_{0}) = \mathbb{P}(x_{n}, t_{n}|x_{n-1}, t_{n-1})$ for all $\{t_{n} \geq t_{n-1} \geq ... \geq t_{0}\} \subset I$.
Roughly speaking,
its concept is informally encapsulated by the statement that "a
Markov process has no memory of the history of past
values of $ X $".
\par
In order to obtain a similar formulation
in the quantum realm we demand a way to define
$ \mathbb{P}(x_{n}, t_{n}|x_{n-1}, t_{n-1};...; x_{0}, t_{0}) $ for quantum systems. In the classical scenario we can sample a random
variable without affecting its posterior statistics. However, 'sampling'
a quantum system requires measuring process, disturbing the state
of the system and affecting the subsequent outcomes. Therefore,
$ \mathbb{P}(x_{n}, t_{n}|x_{n-1}, t_{n-1};...; x_{0}, t_{0}) $ depend on both the
dynamics and the measurement process. Because in such a case the
Markovian character of a quantum dynamical system would
depend on the the measurement scheme chosen to achieve
$ \mathbb{P}(x_{n}, t_{n}|x_{n-1}, t_{n-1};...; x_{0}, t_{0}) $, a definition
of quantum Markovianity in terms of which is not an easy task.
In fact,
the definition of Markovianity should be independent of what
is required to verify it.
\par
The aforementioned problem can be solved by adopting a different approach
focusing on the study of one-time probabilities $ \mathbb{P}(x,t) $. In \textit{linear} quantum evolutions it may lead to concept of \textit{divisibility} which can be defined without any explicit reference to measurement processes in the quantum
realm \cite{rivas2014quantum}. \textit{Although these probabilities help us to
avoid the difficulties
associated with the measurement disturbance, their efficiency in the scenarios involving some measurement with postselection is controversial.}
One of the most well-known approaches to identify the non-Markovian character of the system dynamics has been proposed by Breuer-Laine-Piilo (BLP), namely the distinguishability of two evolving states of the quantum system \cite{breuer2009measure,laine2010measure}. For two arbitrary states $ \rho_{1} $ and $ \rho_{2} $, this distinguishability is quantified by the trace distance (TD) $D(\rho_{1},\rho_{2})=\frac{1}{2}\text{Tr}|\rho_{1}-\rho_{2}|$, where $ |A|=\sqrt{A^{\dagger}A} $ for some operator $ A $.
To explain the physical origin of this
interpretation we consider two parties, Alice and Bob, and
assume that Alice prepares a quantum system in one of
two states $ \rho_{1} $ or $ \rho_{2} $, with a probability of $ 1/2 $ each. Then, the system is passed into a "black box" where it may be
probed by
Bob in any way allowed by the laws of quantum mechanics. Bob's task is to determine whether the system is in the state
$ \rho_{1} $ or $ \rho_{2} $, by
means of a
\textit{single} quantum measurement.
It has been shown that the maximal success probability which
Bob can achieve through an optimal strategy is directly
related to the trace distance \cite{fuchs1998information,breuer2016colloquium}: $ P_{\text{max}}=1/2\bigg(1+D(\rho_{1},\rho_{2})\bigg). $
Therefore, the trace distance denotes the bias in favor
of a correct state discrimination by Bob, and hence it
can be interpreted as the distinguishability of the quantum states $ \rho_{1} $ and $ \rho_{2} $.
\par
The TD is contractive under completely positive and trace-preserving (CPTP) maps, i.e. $ D(\mathcal{E}_{t}(\rho_{1}),\mathcal{E}_{t}(\rho_{2}))\leq D(\rho_{1},\rho_{2})$, if $\mathcal{E}_{t}$ is a CPTP map.
Therefore, a monotonic decrease in the distinguishability ($ \dot{D} <0 $) indicates unidirectional information flow from the system to the environment. However, an increase in the distinguishability ($ \dot{D} >0 $) signifies backflow of information from the environment to the system, indicating that the time evolution of the system is affected by the history of system-environment interaction. This is one of the most well-known definitions of quantum non-Markovianity in the literature. It should be noted that there are also other ways to define and
detect non-Markovianity or memory effects in quantum mechanics (see \cite{rivas2014quantum} for a review). It should be noted that non-Markovian dynamics is always associated with non-unitary evolution of the system. Contrary to Hermitian systems which can evolve unitarily and non-unitarily, (anti-)$\mathcal{PT}\!-$ dynamics is intrinsically nonunitary both
in the unbroken and broken phases \cite{kawabata2017information}, satisfying one of the necessary conditions to exhibit non-Markovian dynamics.
\par Similar to classical scenario, appearance of a measurement with postselection in the process may call into question the validity of
this non-Markovianity definition, because the time evolution of the system becomes dependent on not only the history of system-environment interaction but also the output of the postselection. \textit{This dependence of the quantum non-Markovianity definition on the measurement
scheme, chosen to achieve the time evolution, is bothering even if we distinguish between memory effects and backflow of information from the environment.} Following this idea, we show that the BLP measure does not have a necessary condition to be a figure of merit for characterizing non-Markovianity, if it is definable in dynamics involving postselection.
To this aim, we investigate the TD contractivity, a necessary condition which should be satisfied by this measure when it is used for defining non-Markovianity.
Recently, the Hilbert-Schmidt speed (HSS), a measure of quantum statistical speed not requiring diagonalization of the system reduced density matrix, has been introduced as a faithful witness of non-Markovianity in Hermitian systems, completely consistent with the BLP witness \cite{jahromi2020witnessing}. Moreover, the HSS has been introduced as an efficient figure of merit
for quantum estimation of phase
encoded into the initial state
of open $ n-$qubit systems \cite{jahromi2021hilbert}. Possibility to enhance quantum sensing near the EPs \cite{wiersig2014enhancing,chen2017exceptional,yu2020experimental,zhang2019quantum} and application of the TD to characterize the criticality in (anti-)$ \mathcal{PT} \!-$symmetric systems \cite{kawabata2017information,xiao2019observation,jingwei2020observation} motivate one to investigate the efficiency of the HSS measure to find these singular points in non-Hermitian systems. Moreover, they motivate us to study the relationship among the HSS, TD, and quantum Fisher information (QFI), playing a central role in quantum estimation theory, in (anti-)$ \mathcal{PT} \!-$symmetric systems.
\par In this paper, we address the aforementioned issue and also \emph{show how the HSS can effectively determine the EPs and reveal the critical behavior of non-Hermitian (anti-)$ \mathcal{PT} \!-$symmetric systems}. Because finding the coalescence of eigenvectors at EPs by numerical
full diagonalization, required when computing the TD, may be a tedious and time-consuming chore, this theoretical development for detecting the EPs could be very useful.
This paper is organized as follows. In Sec.~\ref{HILBERT-SCHMIDT SPEED} we briefly
review the definition of the Hilbert-Schmidt speed. In Sec.~\ref{evolution} the time evolution of
the (anti-)$ \mathcal{PT} \!-$symmetric systems as well as non-contractivity of the trace distance are discussed. In Sec.~\ref{sec:HSSnonHermitianMeasure}
we propose a protocol based on the HSS to characterize phase transitions in (anti-)$ \mathcal{PT} \!-$symmetric systems. The efficiency of this protocol for detecting EPs and criticality in single-qubit systems is discussed in Secs.~\ref{SECTIONEXAMPLES} and \ref{ANTISECTIONEXAMPLES}. Moreover, the sensitivity of this witness is also studied for a high-dimensional $ \mathcal{PT} \!-$symmetric system in Sec.~\ref{qudit}. Finally, Sec.~\ref{conclusion} summarizes the main results and prospects.
\section{HILBERT-SCHMIDT SPEED}\label{HILBERT-SCHMIDT SPEED}
The distance measure, defined as \cite{gessner2018statistical}
\begin{equation}\label{cdis}
[d(p,q)]^{2}=\dfrac{1}{2}\sum\limits_{x}^{}|p_{x}-q_{x}|^{2},
\end{equation}
where $ p = \{p_{x}\}_{x} $ and $ q = \{q_{x}\}_{x} $ are probability distributions, leads to the classical statistical speed
\begin{equation}\label{cspeed}
s[p(\varphi_{0})]=\dfrac{\mathrm{d}}{\mathrm{d}\varphi}\ d(p(\varphi_{0}+\varphi),p(\varphi_{0})).
\end{equation}
Thus, one can define a special kind of quantum statistical speed called the HSS by extending these classical notions to the quantum case.
To this aim, we may consider a given pair of quantum states $ \rho $ and $ \sigma $, and write $ p_{x} = \text{Tr}\{E_{x}\rho\} $ and $ q_{x} = \text{Tr}\{E_{x}\sigma\} $ denoting the measurement
probabilities corresponding to the positive-operator-valued measure (POVM) defined by the $ \{E_{x}\geq 0\} $ which satisfies $\sum\limits_{x}^{} E_{x} = \mathbb{I} $.
Then the associated quantum distance called the Hilbert-Schmidt distance $ \delta_{HS} $ \cite{ozawa2000entanglement} can be achieved by maximizing the classical distance of Eq.~(\ref{cdis}) over all possible choices of POVMs \cite{PhysRevA.69.032106}
\begin{equation}\label{qdistance}
\delta_{HS} (\rho,\sigma)\equiv \max_{\{E_{x}\}}\text{d}(\rho,\sigma)=\sqrt{\frac{1}{2}\text{Tr}[(\rho-\sigma)^{2}]}.
\end{equation}
Consequently, the HSS, the corresponding quantum statistical speed, is obtained by maximizing the classical statistical speed of Eq.~(\ref{cspeed}) over all possible POVMs \cite{paris2009quantum,gessner2018statistical}
\begin{align}\label{HSSS}
HS\!S \big(\rho(\varphi)\big)\equiv HS\!S_{\varphi}
&\equiv \text{S}\big[\rho(\varphi)\big]=\max_{\{E_{x}\}} \text{s}\big[p(\varphi)\big]\nonumber\\
&=\sqrt{\frac{1}{2}\text{Tr}\bigg[\bigg(\dfrac{\mathrm{d}\rho(\varphi)}{\mathrm{d}\varphi}\bigg)^2\bigg]},
\end{align}
which can be easily computed without diagonalizing $ \text{d}\rho(\varphi)/\text{d}\varphi $.
\section{Time evolution of the system governed by a (an) (anti-)$ \mathcal{PT} \!-$symmetric Hamiltonian leading to non-contractivity of TD }\label{evolution}
We directly apply the conventional quantum mechanics on (anti-)$ \mathcal{PT} \!-$symmetric systems to obtain the evolved state.
Accordingly, the dynamics governed by a (an) (anti-)$ \mathcal{PT} \!-$symmetric system with non-Hermitian Hamiltonian $ H_{\mathcal{PT}}^{(anti)} $ is described by \cite{brody2012mixed,kawabata2017information}
\begin{equation}\label{densityR}
\rho(t)=\dfrac{\text{e}^{-iH_{\mathcal{PT}}^{(anti)}}\rho(0)\text{e}^{iH^{(anti)\dagger}_{\mathcal{PT}}}}{\text{Tr}[\text{e}^{-iH_{\mathcal{PT}}^{(anti)}}\rho(0)\text{e}^{iH^{(anti)\dagger}_{\mathcal{PT}}}]},
\end{equation}
where the usual Hilbert-Schmidt inner product is employed. In this situation, the effective dynamics governed by $ H_{\mathcal{PT}}^{(anti)} $ is nonunitary and hence it describes the evolution of an open quantum system \cite{kawabata2017information,ohlsson2020transition,jingwei2020observation}. It originates from the fact that a (an) (anti-)$ \mathcal{PT} \!-$symmetric system cannot be implemented by a closed system. In other words, the lack of Hermiticity of the Hamiltonian, is not observable in closed
systems, in contrast to open systems \cite{brody2016consistency}. Hence,
we do not use a preferentially selected inner product
with which the (anti-)$ \mathcal{PT} \!-$symmetric
Hamiltonian $ H_{\mathcal{PT}}^{(anti)} $ can be
employed to generate the unitary time evolution for the characterization of a closed quantum system.
The metric operators used in this approach for modifying the Hilbert space inner products are not physically observable \cite{brody2016consistency}, as discussed above.
Such physical constraints prohibit experimentalists from modifying the inner product in a laboratory, although it can be used as an effective mathematical tool to nicely formulate the theory of quantum systems whose dynamics is governed by (anti-)$ \mathcal{PT} \!-$symmetric
Hamiltonians (see Ref.~\cite{ju2019non,dogra2021quantum} ).
\par
In more detail, we see that the conventional metric used in Eq.~(\ref{densityR}) leads to results completely consistent with experimental observations \cite{tang2016experimental,naghiloo2019quantum,bian2020quantum,jingwei2020observation,yu2020experimental,wang2020experimental,varma2021simulating,zhang2021observation,dogra2021quantum,wu2019observation} provided that it is applied correctly. In fact, in order to physically implement the non-unitary evolution leading to Eq.~(\ref{densityR}), we can embed the $(anti)\mathcal{PT}\!-$symmetric system into a larger
Hermitian system, realized by adding
ancillary qubits, and perform a measuring process \cite{tang2016experimental,xiao2019observation}. This idea originates from the Naimark
dilation procedure for quantum measurement \cite{naimark1943representation,hayashi2006quantum,kawabata2017information,huang2021solvable}: by including
an ancilla and extending the
Hilbert space, any nonunitary dynamics can be implemented by
a unitary dynamics of the total closed system followed by
quantum measurement acting on the ancilla. When a measurement is performed on
the ancilla and a special definite state is postselected, the evolved state (\ref{densityR}) is realized.
Because of this post-selection occurring in the measuring process, the successful implementation of the non-unitary gate is a probabilistic procedure. This experimental limitation \cite{dogra2021quantum,tang2016experimental,bian2020quantum,jingwei2020observation}, is similar to the situation which occurs in Bell inequality tests \cite{hensen2015loophole,giustina2015significant,shalm2015strong}. Therefore, we can solve the paradoxes \cite{lee2014local,bender2007faster,bender2013pt,ju2019non} associated with violation of \textit{no-go theorems} in $\mathcal{PT}\!-$symmetric theory \cite{ju2019non} using normalized density matrix (\ref{densityR}), without the need for any modification in the Hilbert space (for more details, see Refs.~\cite{tang2016experimental,kawabata2017information}).
A significant property of the extended Hamiltonian discussed above
is that the original system is non-Hermitian if and only
if the characteristic interaction between the original system and the ancilla is nonzero. On the one hand, the information flowed into the environment is actually stored in the entanglement with the ancilla.
On the other hand, because the interaction is global, the
quantum correlation between
the system and the ancilla may oscillate in time \cite{kawabata2017information}. Moreover, the information exchange between the system and this entangled partner hidden in the environment may be one of the
physical origins of the time oscillations of the distance measures
for quantum states of the (anti-)$\mathcal{PT}\!-$symmetric
system. The aforementioned reasons motivates us to define non-Markovianity concept in the (anti-)$\mathcal{PT}\!-$symmetric systems, however, as also described in the Introduction, the dependence of this definition on the postselection process, appearing in all current experimental realizations of (anti-)$\mathcal{PT}\!-$symmetric dynamics \cite{tang2016experimental,wu2019observation,naghiloo2019quantum,jingwei2020observation,bian2020quantum,yu2020experimental,wang2020experimental,dogra2021quantum}, may call into question its validity. Demonstrating the failure of the BLP measure in defining possible non-Markovian effects in (anti-)$\mathcal{PT}\!-$symmetric systems (see Refs. \cite{bian2020quantum,ding2021information}) can support this reasoning.
\par
Adopting the BLP's definition of non-Markovianity, one finds that the \textit{P divisibility} of a linear quantum dynamical map is equivalent to Markovianity of the dynamics \cite{rivas2010entanglement}. In spite of the fact that the (anti-)$ \mathcal{PT}$ dynamics given by Eq.~(\ref{densityR}) is indeed P divisible \cite{brody2012mixed}, the divisibility cannot capture the non-Markovianity of the (anti-)$ \mathcal{PT}$ dynamics because of the \textit{nonlinearity} of quantum operation
$ \mathcal{E} $ \cite{kawabata2017information}. It is emphasized that even in the linear evolutions the divisibility of map has not be used to generally define Markovianity, because the concept of divisibility is limited to processes for which
the inverse of the dynamical map exists \cite{breuer2016colloquium,wissmann2015generalized}, a property which cannot be guaranteed e.g., examples on quantum semi-Markov
processes \cite{vacchini2011markovianity} or the damped Jaynes-Cummings model on
resonance \cite{laine2010measure}, thus making the concept of divisibility
sometimes ill defined.
\par
In order to answer the question of whether the BLP's definition of non-Markovianity can be used in non-Hermitian systems \cite{kawabata2017information,xiao2019observation,jingwei2020observation}, we focus on TD contractivity.
As known, the TD always characterizes the distinguishability between two quantum states. When this property is associated with contractivity, it can be used for the definition of non-Markovianity interpreted as backflow of information from the environment to the system. However, in the absence of contractivity, its oscillation can only be attributed to oscillation of distinguishability, not existence of information backflow from the environment. Hence, to answer the question we should investigate the contractivity of TD in non-Hermitian systems.
\par
Concerning the mutual relations of the quantum operations and trace distance, as referred briefly in Introduction, the following important result is well known \cite{nielsen2010quantum,rastegin2007trace,dajka2011distance,laine2011witness}: if there is no initial correlation between the system and environment and $ \mathcal{E} $ is a trace-preserving quantum operation then
$
D\big( \mathcal{E}(\rho), \mathcal{E}(\sigma)\big) \leq D\big( \rho, \sigma\big)
$
where $ \rho $ and $ \sigma$ denote arbitrary normalized quantum states. This result is usually referred to as contractivity
of the trace distance under the linear trace-preserving quantum operations.
\par
However,
the quantum operation $ \mathcal{E} $, generating the evolved state (\ref{densityR}), is non-trace-preserving \cite{nielsen2010quantum}, since it does not provide a complete description of the processes
occurring in the system. This nondeterministic feature originates from the fact that other measurement outcomes may take place with some probability. Therefore, the contractivity of the TD under the evolution given in Eq. (\ref{densityR}) should be investigated in more detail.
Our numerical calculation, presented in the next sections, shows that \textit{the trace distance may exhibit non-contracticity under (anti-)$ \mathcal{PT}$ dynamics and hence in such systems it loses one of the necessary conditions which should be satisfied by a faithful witness of non-Markovianity.}
\par
It should be noted that contractivity is not a universal feature
but depends on the metric: the dynamics may be
contractive with respect to a given metric and may not be
contractive with respect to other metric measures \cite{dajka2011distance}. In addition,
contractivity of quantum evolution can break down when the system is initially correlated with its environment (for details see \cite{laine2011witness} ).
\section{Detecting criticality THROUGH HSS IN NON-HERMITIAN SYSTEMS}\label{sec:HSSnonHermitianMeasure}
In this section we provide the witness based on the Hilbert-Schmidt speed to faithfully identify the EPs and phase transitions
in non-Hermitian systems.
It is known that in the $ \mathcal{PT}$-symmetric systems the trace distance
oscillates with evolution time when the system symmetry is unbroken while in anti-$ \mathcal{PT}$-symmetric systems, the oscillations of the distinguishability occur if the symmetry is broken \cite{kawabata2017information,xiao2019observation,jingwei2020observation,dogra2021quantum}. Inspired by this fact and the close relationship between HSS and TD described in Ref. \cite{jahromi2020witnessing}, we propose the following easily computable witness to characterize the criticality in (anti-)$ \mathcal{PT}$-symmetric systems:
\textit{For a quantum system with an $n$-dimensional Hilbert space $\mathcal{H}$, let us consider an initial state given by
\begin{equation}\label{initialstate}
|\psi_{0}\rangle=\dfrac{1}{\sqrt{n}}\big(\text{e}^{i\varphi}|\psi_{1}\rangle+\ldots+|\psi_{n}\rangle\big),
\end{equation}
where $\varphi$ is an unknown phase shift and $\{|\psi_{i}\rangle,\ i=1,\ldots,n\}$ represents the computational orthonormal basis. Then, we find that when the dynamics of the HSS, computed with respect to the initial phase $ \varphi $, exhibit an
oscillating pattern, the (anti-)$\mathcal{PT}\!-$symmetric system is in (anti-)$\mathcal{PT}\!-$unbroken (broken) phase.
At EPs or broken (unbroken) phase of the (anti-)$\mathcal{PT}\!-$symmetric system, no oscillation is observed in the HSS dynamics.
Therefore, the EPs can be easily detected by investigating the time evolution of the HSS.}
The sanity check of this protocol
as a faithful witness of EPs and criticality is performed in the
following section.
\par
\section{$\mathcal{PT}$-symmetric two-level system}\label{SECTIONEXAMPLES}
\subsection{Hamiltonian model and computing the witness}
As the first example, we consider
the paradigmatic model of a two-level system described by the Hamiltonian
\begin{equation}\label{HoneQubit}
H_{\mathcal{PT}}=\varepsilon(\sigma_{x}+ia\sigma_{z})=\left(
\begin{array}{cc}
ia\varepsilon&\varepsilon \\
\varepsilon&-ia\varepsilon \\
\end{array}
\right),
\end{equation}
where $ \varepsilon \geqslant 0 $ is an energy scale and $ a\geqslant 0 $ denotes the degree of non-Hermiticity. This model has been previously realized in both classical \cite{ruter2010observation,gao2015observation,liu2016metrology}
and quantum \cite{li2019observation,tang2016experimental,jingwei2020observation} experiments. The eigenenergies are
$ \pm \varepsilon \sqrt{1-a^{2}} $,
and therefore one has an EP2 (exceptional point of order 2) given by $ a=1 $.
\par
The time evolution operator of this system is obtained as \cite{kawabata2017information}
\begin{eqnarray}\label{UoneQubit}
& U_{\mathcal{PT}}=\text{e}^{-iH_{\mathcal{PT}}t}&\nonumber\\
&=\dfrac{1}{\sqrt{1-a^{2}}}\left(
\begin{array}{cc}
\sqrt{1-a^{2}} \text{cos}\theta+a~ \text{sin}\theta & -i ~\text{sin}\theta \\
-i~ \text{sin}\theta & \sqrt{1-a^{2}} \text{cos}\theta-a~ \text{sin}\theta \\
\end{array}
\right),&\nonumber \\
\end{eqnarray}
where $\theta =\sqrt{1-a^{2}}~\varepsilon t $.
\par
In order to check the efficiency of our HSS-based witness, we first compute the normalized evolved state of the system when it is prepared in the initial state $\rho_{0}=\ket{\psi_{0}}\bra{\psi_{0}} $, where $ |\psi_{0}\rangle=(\text{e}^{i\varphi}|0\rangle+|1\rangle)/\sqrt{2}$. We can easily calculate the HSS analytically by inserting the evolved state $ \rho_{t}(\varphi)=U_{\mathcal{PT}}\rho_{0}U_{\mathcal{PT}}^{\dagger}\bigg\{\text{Tr}[U_{\mathcal{PT}}\rho_{0}U_{\mathcal{PT}}^{\dagger}]\bigg\}^{-1} $ into Eq.~(\ref{HSSS}); however, its explicit expression has a cumbersome form and is not reported here.
\subsection{Dynamical behavior of the witness of the quantum criticality}
The qualitative dynamics of the HSS is displayed in Fig.~\ref{HssQfi} for the $ \mathcal{PT} \!-$broken phase ($ a > 1 $) and $ \mathcal{PT} \!-$unbroken phase ($ 0<a<1 $).
In the $ \mathcal{PT} \!-$broken phase, the HSS, as expected, exhibits no oscillations and monotonically decreases with time.
However, in the $ \mathcal{PT} \!-$unbroken phase, it oscillates and eventually returns to its initial value, namely there is a $ T $
\begin{equation}\label{InfRet}
\ \text{such that}~~HS\!S(T)=HS\!S(0),
\end{equation}
where the period is given by $ T=\pi~ \big [1-a^{2}\big]^{-1/2} $. This period $ T $ of the
oscillation, called the recurrence time, increases as the system approaches the EP ($ a=1 $). This period is exactly similar to one achieved theoretically as well as experimentally for the distinguishability oscillations \cite{kawabata2017information,xiao2019observation}. Moreover, checking the experimental data presented in \cite{xiao2019observation}, we find that the EP predicted by the HSS-based witness is quite accurate.
\begin{figure}[t!]
\subfigure{ \includegraphics[width=0.45\textwidth]{HssQfi12.eps}\label{HssQfi12} }
\subfigure{\includegraphics[width=0.45\textwidth]{HssQfi11.eps}\label{HssQfi11}}
\caption{Dynamics of the Hilbert-Schmidt speed $ HS\!S(\rho_{t}(\varphi)) $ (blue
dashed curve) and the Quantum Fisher information $ QFI(t) $ (red solid curve) as a function of the dimensionless
time $ \varepsilon t $ for a two-level system (a) in the $ \mathcal{PT} \!-$broken phase ($ a > 1 $) and (b) in the $ \mathcal{PT} \!-$unbroken phase ($ 0<a<1 $).}
\label{HssQfi}
\end{figure}
\subsubsection*{Behavior at the exceptional point}
As discussed in the Introduction, the exceptional points (EPs) in non-Hermitian Hamiltonians constitute a threshold for the system parameters individuating a phase transition of the system. In the present case of a $ \mathcal{PT} \!-$symmetric system, one has a broken-to-unbroken phase transition by diminishing the value of the non-Hermiticity degree $a$ appearing in the Hamiltonian of Eq.~(\ref{HoneQubit}). Since this phase transition is associated to intriguing physical phenomena, the characterization of the behavior of the non-Hermitian system around EPs is important.
Moreover, because of the importance of the HSS in detecting the EPs, it would be interesting to investigate the behavior of this measure when the system approaches the EP, i.e., when $ a\rightarrow 1 $. In this limit, we find that the HSS is given by
\begin{eqnarray}
HS\!S_{EP}(t)&=&\dfrac{1}{ 4\,{\varepsilon}^{2}{t}^
{2}-4\,{\varepsilon}^{2}{t}^{2}\sin \left( \varphi \right)+2}.
\end{eqnarray}
It is immediate to see that it decreases monotonically with time according to $HS\!S_{EP}(t)\propto t^{-2} $; while only for $ \varphi=\pi/2 $ the HSS gives a constant value equal to $ 1/2 $. These results immediately show that, when the system parameters reach the EP ($ a\rightarrow 1 $), no oscillation in HSS dynamics is observed which arises from the divergence of period $ T $ at this limit.
Overall, we see that the HSS can be employed as an efficient witness to identify phase transitions and detect the EPs in the non-Hermitian system under consideration.
\subsection{Relationship between quantum Fisher information (QFI) and HSS }\label{QFIHSS}
The fundamental question in the theory of quantum estimation is the following: When performing
measurements on the quantum systems affected by some classical parameter $ \varphi $ (which may be the phase encoded into the initial state of the system), how precisely
can $ \varphi $ be estimated? The answer is given by the quantum
Cramer-Rao bound \cite{braunstein1994statistical} indicating that the smallest
resolvable change in $ \varphi $ is $ \delta \varphi =1/\sqrt{F_{\varphi}} $ where $ F_{\varphi} $ denotes the quantum Fisher information (QFI) given by \cite{braunstein1994statistical,liu2019quantum}
\begin{equation}
F_{\varphi}(\rho(t))=2\sum_{i,j} \dfrac{|\langle \phi_{i}|\partial_{\varphi}\rho\left(\varphi \right)|\phi_{j}\rangle|^{2}}{(\lambda_{i}+\lambda_{j})},
\end{equation}
where $ |\phi_{i}\rangle $ and $ \lambda_{i} $ represent, respectively, the eigenvectors and eigenvalues of the density matrix $ \rho\left(t \right) $. It should be noted that recently other expressions for the QFI have been proposed in non-Hermitian systems \cite{li2021cram}, however, for pure states they reduce to the above expression, ignoring a constant coefficient.
\par
Because both QFI and HSS are quantum statistical speeds associated, respectively, with the \textit{Bures} and \textit{Hilbert–Schmidt distances} (for details see \cite{gessner2018statistical}), it is reasonable to explore how they can be
related to each other.
Recently, a strong relationship between the HSS and QFI
has been constructed in the process of phase estimation for $n$-qubit Hermitian systems \cite{jahromi2021hilbert}. It has been found that, when both the HSS and
QFI are computed with respect to the phase parameter encoded into
the initial state of an $n$-qubit system, the zeros of the HSS dynamics are actually equal to those of the QFI dynamics. Likewise, the signs of the time-derivatives of both HSS and QFI exactly coincide.
\par
Now computing the QFI with respect to the phase parameter encoded into the initial state of our one-qubit $ \mathcal{PT} \!-$symmetric system, we obtain the similar result, i.e., both the QFI and HSS exhibit the same qualitative dynamics (see Fig. \ref{HssQfi}). Moreover, in the $ \mathcal{PT} \!-$broken phase ($ a>1 $) illustrated in Fig.~\ref{HssQfi12}, both measures are contractive and hence monotonically decrease with time. However, as observed in Fig.~\ref{HssQfi11},
not only the HSS but also the QFI, contractive under CPTP maps in Hermitian systems, may be non-contractive under the non-trace-preserving evolution in the unbroken phase ($ 0<a<1 $). Moreover, this important relationship between the HSS and QFI shows that the HSS may be introduced as an efficient figure of merit for quantum estimation of phase encoded into the initial state of $n$-qubit $ \mathcal{PT} \!-$symmetric systems. This will be investigated in detail in future studies.
The non-contractivity of the QFI and HSS confirms the fact that quantum information measures and witnesses may exhibit different behaviors for Hermitian and non-Hermitian systems.
It is known that the QFI measures the maximum information about a parameter $ \varphi $, extractable from a given measurement procedure \cite{braunstein1994statistical,liu2019quantum,petz2011introduction,jafarzadeh2020effects,shadehi2020adiabatic}.
In our model, because the parameter to be estimated is encoded into the initial state of the system and the system, initially not correlated with the environment, does not sense it later,
we reasonably expect that the maximum information, achieved in the estimation process, must be extracted from the initial state itself.
Contrary to this intuitive reasoning, it is observed that the non-Hermitian evolution of the system interestingly may enhance the estimation of the initial parameter with time.
\par
We can conclude that the system initially hides some of the encoded information such that it is inaccessible at first and then the non-Hermitian evolution makes this hidden information available for the estimation of the initial phase. In fact, this unusual behavior, leading to non-contractivity of both QFI and HSS, is indistinguishable from the non-Markovian behavior (backflow of information from the environment to the system).
This is why we cannot use the standard Hermitian witnesses based on QFI and HSS to detect the non-Markovianity in this context.
\subsection{Non-contractivity of TD in $\mathcal{PT}\!-$symmetric one-qubit systems}
\begin{figure}[t!]
\includegraphics[width=0.45\textwidth]{noncont1.eps}
\caption{Dynamics of the distinguishability $D $ between two evolved states $ \rho_{1} (t)$ and $ \rho_{2} (t)$ starting from two special pairs of initial states as a function of the dimensionless
time $ \varepsilon t $ for a two-level system in the $ \mathcal{PT} \!-$unbroken phase ($ 0<a<1 $). }
\label{noncont1}
\end{figure}
The non-contractivity of the QFI and HSS motivates us to investigate the TD non-contractivity under non-Hermitian evolution of the system. Our numerical computation shows that the TD may be non-contractive in such systems. Assuming two initial states $ \ket{\psi _{0}^{(1)}}=1/\sqrt{2}\big(\text{e}^{i\varphi} \ket{0}+\ket{1}\big) $ and $ \ket{\psi _{0}^{(2)}}=1/\sqrt{2}\big( \ket{0}-\text{e}^{i\varphi}\ket{1}\big) $, and computing the TD between the two corresponding evolved states, we can observe the non-contractivity of the TD under $ \mathcal{PT}$ dynamics (see Fig. \ref{noncont1}, showing the time evolution of the TD, in which the green solid and orange dashed curves are plotted for ($ a=0.2, \varphi=\pi/3 $ ) and ($ a=0.4, \varphi=\pi/4 $), respectively).
In our model, initially the system is not correlated with the environment. This preparation alongside the fact that the decoherence effects lead to flow of information from the system to the environment, make us
reasonably expect that the distinguishability shows contractivity under the time evolution. However, as discussed in Sec.~\ref{QFIHSS}, the system initially may hide some of the information such that it is inaccessible at first and then the non-Hermitian evolution makes this hidden information available for the system.
This counterintuitive behavior may be one of the reasons that the TD exhibits non-contractive behavior. Moreover, this unusual behavior is indistinguishable from the non-Markovian behavior (i.e., backflow of information from the environment to the system). Therefore, we cannot use the BLP's measure as a faithful witness to detect the non-Markovianity in $\mathcal{PT}\!-$symmetric systems.
\section{ ANTI-$ \mathcal{PT} \!-$SYMMETRIC TWO-LEVEL SYSTEM}\label{ANTISECTIONEXAMPLES}
\subsection{Hamiltonian model and witnesses}
The generalized form of a single-qubit anti-$\mathcal{PT}\!-$symmetric Hamiltonian can be expressed as \cite{bender2002complex}
\begin{equation}\label{HantiPT}
H^{\text{anti}}_{\mathcal{PT}}=\left(
\begin{array}{cc}
\lambda \eta ~\text{e}^{i\vartheta} &i\eta \\
i \eta&-\lambda \eta ~\text{e}^{-i\vartheta} \\
\end{array}
\right),
\end{equation}
where all of the parameters $ \lambda, \vartheta $ and $ \eta$ denote real numbers. It is easy to show that this Hamiltonian satisfies the anti-commutation relation
\begin{equation}
(\mathcal{PT})H^{\text{anti}}_{\mathcal{PT}}(\mathcal{PT})^{-1}=-(H^{\text{anti}}_{\mathcal{PT}})^{T}=-H^{\text{anti}}_{\mathcal{PT}}
\end{equation}
where here operator $\mathcal{P}$ denotes Pauli matrix $ \sigma_{x} $, $\mathcal{T}$ represents the complex
conjugation, and notation $ A^{T} $ means the transpose of matrix $ A $.
The
eigenvalues of Hamiltonian $ H^{\text{anti}}_{PT} $ are given by $ \epsilon_{\pm}=i\lambda \eta~\text{sin}\vartheta \pm \sqrt{\lambda^{2}\eta^{2}\text{cos}^{2}\vartheta-\eta^{2}} $ and the system is denoted in the regime of unbroken anti-$\mathcal{PT}\!-$symmetric phase if $ \lambda^{2}\eta^{2}\text{cos}^{2}\vartheta-\eta^{2}<0 $. For simplicity, as well as comparison with the experimental results presented in \cite{jingwei2020observation} for this kind of system, we consider the scenario in which $ \vartheta=0 $, $ \eta \geq 0 $ being an energy scale and $ \lambda \geq 0 $ denoting the degree of Hermiticity. Therefore, the EP2 is located at $ \lambda=1 $.
We obtain the corresponding time evolution
operator of this system as
\begin{equation}
U^{\text{anti}}_{\mathcal{PT}}=\left(
\begin{array}{cc}
\cos \theta-\frac{i \lambda \sin \theta}{\sqrt{\lambda^2-1}} & \frac{\sin \theta}{\sqrt{\lambda^2-1}} \\
\frac{\sin \theta}{\sqrt{\lambda^2-1}} & \cos \theta+\frac{i \lambda \sin \theta}{\sqrt{\lambda^2-1}} \\
\end{array}
\right),
\end{equation}
where $ \theta=\sqrt{\lambda^2-1}~ \eta t $.
\par
The calculation of the HSS is similar to the approach followed in the previous section. The analytical expressions for this witness is also accessible; however, it does not have compact informative forms, and hence it is not reported here.
\subsection{Dynamical behavior of the witness of quantum criticality}
\begin{figure}[t!]
\subfigure{ \includegraphics[width=0.435\textwidth]{HssQfianti21.eps}\label{HssQfianti21} }
\subfigure{ \includegraphics[width=0.435\textwidth]{HssQfianti22.eps}\label{HssQfianti22} }
\caption{Dynamics of the Hilbert-Schmidt speed $ HS\!S(\rho_{t}(\varphi)) $ (blue
dashed curve) and QFI $ D(t) $ (red solid curve) as a function of the dimensionless
time $ \eta t $ for a two-level system (a) in the anti-$ \mathcal{PT} \!-$unbroken phase ($ \lambda < 1 $) and (b) in the anti-$ \mathcal{PT} \!-$broken phase ($ \lambda >1 $).}
\label{HssQfianti}
\end{figure}
We again see that the HSS works well in detecting the critical behavior of the system. In the anti-$ \mathcal{PT} \!-$unbroken phase ($ \lambda < 1 $), as shown in Fig.~\ref{HssQfianti21} plotted for $ \lambda=0.2 $, the HSS does not oscillate over time. However, its oscillatory behavior for $ \lambda >1 $ exactly predicts that the system is in anti-$ \mathcal{PT} \!-$broken phase (see Fig.~\ref{HssQfianti22} plotted for $ \lambda =1.4 $).
Moreover, we find that the intensity of the HSS oscillations is related to the parameter $ \lambda $, denoting the degree of Hermiticity, such that with an increase in this parameter, its oscillation gradually weakens.
In Ref.~\cite{jingwei2020observation}, the authors
proposed an algorithm for the implementation of
the above generalized anti-$\mathcal{PT}\!-$symmetric evolution with a circuit-based
quantum computing system applying a three-qubit scheme including two ancillary
qubits and one working qubit. The
implementation scheme is based on decomposing the non-Hermitian
Hamiltonian evolution into a linear combination of unitary
operators and realizing the scheme in an enlarged Hilbert
space with ancillary qubits.
The experimental results clearly show that the distinguishability
oscillates with period
\begin{equation}
T=\dfrac{\pi}{\eta \sqrt{\lambda^{2}-1}},
\end{equation}
when the system symmetry is broken. Moreover, the intensity of the distinguishability oscillation is connected to parameter $ \lambda $ such that the transition between the broken and unbroken phases is completely determined by this parameter. Comparing our findings to these experimental results, we see that they are completely consistent. In particular we find that the periods of the HSS and distinguishability oscillations are exactly the same.
This fact hence proves the efficiency of our proposed witness in faithfully detecting the critical behavior of the anti-$ \mathcal{PT} \!-$symmetric systems.
\subsubsection*{Behavior at the exceptional point}
In this case of an anti-$\mathcal{PT}\!-$symmetric system, the exceptional point (EP) individuates a phase transition of the system from the unbroken to the broken phase when increasing the parameter $\lambda$ of the Hamiltonian of Eq.~(\ref{HantiPT}).
In order to complete the analysis, we investigate the behavior of the HSS at the EP $ \lambda=1 $, where it is given by
\begin{eqnarray}
HS\!S_{EP}^{\text{anti}}(t)&=&\frac{1}{2 \left| 2 \eta^2 t^2+2 \eta t \left[ \eta t~ \sin \varphi +\cos \varphi \right] +1 \right| }.
\end{eqnarray}
It is easily seen that the time behavior of the HSS is similar to that we have shown in Fig.~\ref{HssQfianti21} depicting the HSS dynamics in the unbroken phase. This result also confirms that the HSS can be used to detect phase transitions at the EPs in this type of non-Hermitian systems.
\subsection{Relationship between QFI and HSS }
Calculating the QFI with respect to phase parameter
$ \varphi $, we again find that both the QFI and HSS exhibit the same qualitative dynamics in this one-qubit anti-$ \mathcal{PT} \!-$symmetric system (see Fig.~\ref{HssQfianti}).
Interestingly, in both unbroken and broken phases, depicted in Fig.~\ref{HssQfianti}, both measures does not necessarily exhibit contractivity, usually assumed as a necessary property for any faithful witness of non-Markovianity.
\subsection{Non-contractivity of TD in anti-$\mathcal{PT}\!-$symmetric one-qubit systems}
\begin{figure}[t!]
\subfigure{ \includegraphics[width=0.435\textwidth]{antinoncon1.eps}\label{antinoncon1} }
\subfigure{ \includegraphics[width=0.435\textwidth]{antinoncon2.eps}\label{antinoncon2} }
\caption{Dynamics of the distinguishability $D $ between two evolved states $ \rho_{1} (t)$ and $ \rho_{2} (t)$ starting from two special pairs of initial states as a function of the dimensionless
time $ \eta t $ for a two-level system (a) in the anti-$ \mathcal{PT} \!-$unbroken phase ($ \lambda < 1 $) and (b) in the anti-$ \mathcal{PT} \!-$broken phase ($ \lambda >1 $). }
\label{antinoncon}
\end{figure}
Numerically investigating the TD contractivity under the anti-$\mathcal{PT}\!-$symmetric evolution of the system, we find that it is not necessarily contractive in these systems. Starting from two initial states $ \ket{\psi _{0}^{(1)}}=1/\sqrt{2}\big(\text{e}^{i\varphi} \ket{0}+\ket{1}\big) $ and $ \ket{\psi _{0}^{(2)}}=1/\sqrt{2}\big( \ket{0}+\text{e}^{i\varphi}\ket{1}\big) $, and computing the TD between the two corresponding evolved states, we may observe the non-contractivity of the TD under anti-$ \mathcal{PT}$ symmetric dynamics in both unbroken and broken phases (see Fig. \ref{antinoncon1} (\ref{antinoncon2}), showing the time evolution of the TD in the unbroken (broken) phase, where the green solid and orange dashed curves are plotted for $ \bigg(\lambda=0.5 ~(\lambda=1.5), \varphi=3.1 \bigg)$ and $ \bigg(\lambda=0.6 ~(\lambda=1.3), \varphi=2.9 \bigg)$, respectively).
\section{ HIGH-DIMENSIONAL $ \mathcal{PT}\!-$SYMMETRIC SYSTEM}\label{qudit}
\subsection{Hamiltonian model and witnesses}
Now we consider an open, high-dimensional system described by a $4 \times 4$ Hamiltonian \cite{bian2020quantum}
\begin{equation}
H_{\mathcal{PT}}=-JS_{x}+i\gamma S_{z}
\end{equation}
in which $S_{x} $ and $ S_{z} $ denote spin-3/2 representations of the SU(2)
group. In the orthonormal computational basis $ \{\ket{1},\ket{2},\ket{3},\ket{4}\} $, the Hamiltonian can be written in the following matrix form
\begin{equation}\label{HQubit}
H_{\mathcal{PT}}=\dfrac{1}{2}\left(
\begin{array}{cccc}
3i\gamma &-\sqrt{3}J&0 &0 \\
-\sqrt{3}J &i\gamma&-2J &0 \\
0 &-2J &-i\gamma&-\sqrt{3}J \\
0 &0&-\sqrt{3}J &-3i\gamma \\
\end{array}
\right),
\end{equation}
representing a $ \mathcal{PT}\!-$symmetric qudit with $ d=4 $. The eigenvalues of $ H_{PT} $ are
simply given by $ \lambda_{k}=\{-3/2,-1/2,1/2,3/2\}\sqrt{J^{2}-\gamma^{2}}$ $(k=1,2,3,4) $, leading to an EP4 at the $ \mathcal{PT}\!-$breaking
threshold $ \gamma=J $. This Hamiltonian can be easily generalized to an arbitrary dimensional
system and has an EP with the order equal to the dimension of the system \cite{hodaei2017enhanced,graefe2008non,quiroz2019exceptional}.
\begin{figure}[t!]
\centering
\subfigure{ \includegraphics[width=0.48\textwidth]{HSSqudit1.eps}\label{HSSqudit1} }
\subfigure{ \includegraphics[width=0.48\textwidth]{HSSqudit2.eps}\label{HSSqudit2} }
\caption{Dynamics of the Hilbert-Schmidt speed $ HS\!S(\rho_{t}(\varphi)) $ as a function of the dimensionless
time $ \gamma t $ for the qudit system (a) in the $ \mathcal{PT}\!-$broken phase ($ \gamma >J $) and (b) in the $ \mathcal{PT} \!-$unbroken phase ($ \gamma<J $).
}
\label{HSSqudit}
\end{figure}
Let us now assume that the qudit system is $4$-dimensional ($d=4$) and prepared in the pure initial state of Eq.~(\ref{initialstate}): $ \ket{\psi_{0}}=\big(\text{e}^{i\varphi}\ket{1}+\ket{2}+\ket{3}+\ket{4}\big)/\sqrt{4} $.
To obtain the time evolution operator $ U=\text{e}^{-iH_{\mathcal{PT}}t} $, we can expand the initial state in terms of the non-orthogonal eigenvectors $ \ket{\zeta_{k}} $ of the non-Hermitian Hamiltonian $ H_{\mathcal{PT}} $ as $ \ket{\psi_{0}}=\sum_{k}^{}\beta_{k}\ket{\zeta_{k}}$, where the coefficients $\beta_{k} $'s should be determined. To this aim, we first define a square $ d \times d $ matrix $ \Phi $ such that the normalized eigenvectors $ \ket{\zeta_{k}} $ are concatenated as its columns. Therefore, the initial state can be represented as $\ket{\psi_{0}}:= \Phi~ \beta $
where $ \beta$ denotes a column matrix with elements $ \beta_{i}$. Because the columns of $ \Phi $ are linearly independent, it is invertible, and hence we can write $\ket{\psi_{0}}:=\Phi \Phi ^{-1} \ket{\psi_{0}} $, leading to the relation $ \beta=\Phi ^{-1} \ket{\psi_{0}} $. After computing the coefficients $\beta_{k} $'s by the above method, we can easily obtain the evolved state of the system. Although the computation of the HSS (associated with the evolved density matrix $ \rho_{t}(\varphi)=\ket{\psi(t)}\bra{\psi(t)} /\text{Tr}[\ket{\psi(t)}\bra{\psi(t)}]$ where $ \ket{\psi(t)}=\text{e}^{-iH_{\mathcal{PT}}t} \ket{\psi_{0}}= \sum_{k}^{}\text{e}^{-i\lambda_{k}t}\beta_{k}\ket{\zeta_{k}}$, with respect to initial phase $ \varphi $) is straightforward, the explicit analytic expression is extremely complex and is not reported in this paper, whereas the results
are described below.
\subsection{Dynamical behavior of the witness and EP role}
Analyzing the HSS dynamics reveals that the EP4 with $ \gamma = J $ determines the border between $ \mathcal{PT} \!-$symmetric unbrken and broken phases of the qudit (see Fig.~\ref{HSSqudit}). In the $ \mathcal{PT}\!-$broken phase ($ \gamma >J $), the HSS first may increase with time and show a peak. Nevertheless, no oscillation in its dynamics is observed and then it monotonously falls with time (see Fig. \ref{HSSqudit1} plotted for $ J/\gamma=0.9 $). However, in the $ \mathcal{PT}\!-$symmetry unbroken region ($ \gamma < J $), as expected and shown in Fig. \ref{HSSqudit2} plotted for $ J/\gamma=2.2 $, the HSS dynamics exhibits periodic oscillations (see relation (\ref{InfRet})).
\subsection{Non-contractivity of TD in high dimensional $\mathcal{PT}\!-$symmetric systems}
\begin{figure}[t!]
\centering
\subfigure{ \includegraphics[width=0.48\textwidth]{qudit1.eps}\label{qudit1} }
\subfigure{ \includegraphics[width=0.48\textwidth]{qudit2.eps}\label{qudit2} }
\caption{Dynamics of the distinguishability $D $ between two evolved states $ \rho_{1} (t)$ and $ \rho_{2} (t)$ starting from two special pairs of initial states of the qudit system as a function of the dimensionless
time $ \gamma t $ (a) in the $ \mathcal{PT}\!-$broken phase ($ J/\gamma=0.9 $) and (b) in the $ \mathcal{PT} \!-$unbroken phase ($ J/\gamma=2.2 $).
}
\label{qudit12}
\end{figure}
The computation of the trace distance requires diagonalization of $ \rho_{1}(t) -\rho_{2}(t) $ for the pair of optimal initial states $ \rho_{1}(0) $ and $ \rho_{2}(0) $. The optimization process with complexity of computing the eigenvectors of the high dimensional complicated density matrices $ \rho_{i}(t) ~ (i=1,2) $, makes the trace distance too difficult to compute.
We check numerically the contractivity of TD under the $\mathcal{PT}\!-$symmetric evolution of the qudit for a large number of random pairs of initial states. We explicitly show that the TD may exhibit non-contractivity in both unbroken and broken phases. For example, starting from two initial states $ \ket{\psi _{0}^{(1)}}=\big(\text{e}^{i\varphi}\ket{1}+\ket{2}+\ket{3}+\ket{4}\big)/\sqrt{4} $ and $ \ket{\psi _{0}^{(2)}}=\big(\text{e}^{i\theta}\ket{1}+\text{e}^{i\varphi}\ket{2}+\ket{3}+\ket{4}\big)/\sqrt{4} $, and computing the TD between the two corresponding evolved states, one can observe the non-contractivity of the TD under $ \mathcal{PT}$ symmetric dynamics of the qudit in both phases (see Fig. \ref{qudit1} (\ref{qudit2}), representing the dynamics of the the distinguishability in the broken (unbroken) phase, in which the green solid and orange dashed curves are plotted for $\bigg( \theta=\pi/4, \varphi=\pi/2\bigg) $ and $ \bigg(\theta=1.1~(\theta=1.5), \varphi=2.1~(\varphi=0.1) \bigg)$, respectively).
\par
Recently, this interesting four-dimensional system, has been implemented experimentally with single photons and a cascaded interferometric setup \cite{bian2020quantum}. In that work, $ 4\times 4 $ nonunitary evolution operations were realized by six
beam displacers and another one is used for state preparation. Moreover, two different measurements, i.e., the projective measurement and the quantum
state tomography of a four-level system, are carried out at
the output.
We find that our theoretical predictions are completely consistent with the experimental results discussed in \cite{bian2020quantum} in which explicitly reported that the EP4 ($ \gamma = J $) plays the main role in determining the critical behavior of the qudit. Moreover, our theoretical analysis shows that the HSS oscillates with period $ T=2\pi/(\sqrt{J^{2}-\gamma^{2}}) $ in the unbroken phase. This is exactly the period measured experimentally in \cite{bian2020quantum} for oscillations of the dynamics of the quantum information (as quantified by entropy). Because the TD is not contractive in this model, these oscillations cannot be interpreted as evidence
of information backflow from the environment or the signature
of non-Markovianity. They can be attributed to the non-Hermitian evolution of the system making the initial hidden information available, as discussed in Sec. \ref{QFIHSS}.
\section{Conclusions}\label{conclusion}
We have proposed a powerful and easily computable witness, based on the Hilbert-Schmidt speed (HSS), which is a special case of quantum statistical speed, to detect the quantum criticality in systems governed by non-Hermitian (anti-)$\mathcal{PT}\!-$symmetric Hamiltonians. Surprisingly, our theoretical predictions can exactly predict the experimental results.
In addition to its conceptual interest, we remark that the HSS-based witness of criticality does not require the diagonalization of the reduced density matrix of the system. Hence, as discussed in the paper for a four-dimensional qudit, it can be introduced as a faithful witness for characterizing criticality in high-dimensional (anti-)$\mathcal{PT}\!-$symmetric systems in which computation of other measures leads to serious challenges. The role of the HSS-based measure in detecting the exceptional points (EPs) at which breaking of (anti-)$\mathcal{PT}\!-$symmetry occurs has been also analyzed. We have especially illustrated that the system critical behavior appearing at EPs is similar to that of (anti-)$\mathcal{PT}\!-$symmetric system in its broken (unbroken) phase. We stress that our theoretical findings all are in complete agreement with experimental observations. These results thus indicate that the HSS-based witness can be adopted to exactly identify the parameter values where phase transitions in the physical behavior of non-Hermitian systems occur.
As an interesting outlook, the introduced HSS-based witness can be employed to characterize controlled speedup of quantum processes in non-Hermitian systems. Such a line of investigation is suggested by recent advances in the context of shortcuts to adiabaticity (STA), which enable us to control a quantum system evolution with no need of slow driving \cite{guery2019shortcuts,funo2020shortcuts,chen2020shortcuts,alipour2020shortcuts}. Strategies for STA are typically engineered by means of non-Hermitian control Hamiltonians \cite{alipour2020shortcuts,Chen2018,Impens2019}.
On the basis of this argument and seeing the results presented here, one may thus expect that the HSS measure plays a role in optimizing STA via non-Hermitian Hamiltonians. This study will be carried on elsewhere.
Another important feature which should be addressed is the non-contractivity of both trace distance (TD) and HSS in (anti-)$ \mathcal{PT} \!-$symmetric non-Hermitian systems.
Because of this non-contractive behavior, the change in the TD of two arbitrary states can no longer be interpreted as a flow of information between the system and the environment. Therefore, contrary to what happens in Hermitian system \cite{breuer2009measure,jahromi2020witnessing}, both TD and HSS can no longer be used as general measures or definitions of non-Markovianity in (anti-)$ \mathcal{PT} \!-$symmetric systems. Our results show that the non-contractivity of the Hilbert-Schmidt speed in a given model may be a signature of the trace distance non-contractivity and its failure in detecting non-Markovianity.
\par
We explain that the definition of non-Markovianity remains as an important open question in the theory of non-Hermitian quantum systems. In fact, the distinguishability and HSS oscillations in
(anti-)$ \mathcal{PT} \!-$symmetric dynamics may be attributed to the
result of the dynamical overlap between the skew eigenstates in the (anti-)$ \mathcal{PT} \!-$symmetric, an interesting
characteristic absent in conventional open quantum systems. In the unbroken (broken) phase, this overlap is nontrivial and leads to
a beat from the two skew dynamical eigenstates,
such that the beat period equals the oscillation period. However, in the
broken (unbroken) phase, the aforementioned overlap becomes trivial, and in addition,
the amplitudes of the eigenstates monotonically shrink or grow under quantum evolution, leading to a monotonic decrease in the
distinguishability or the HSS dynamics (see \cite{wang2020experimental} for more details).
In Figs.~\ref{HssQfi} and \ref{HssQfianti}, we have compared the behaviors of HSS and quantum Fisher information (QFI) associated with the initial phase $ \varphi $ for one-qubit (anti-)$ \mathcal{PT} \!-$symmetric systems. This investigation shows that the HSS and QFI exhibit the same qualitative dynamics in one-qubit systems. According to the discussion presented in the previous paragraph, we find that by inspecting the QFI dynamics one cannot exactly detect the non-Markovian evolution of the (anti-)PT-symmetric systems. However, as known, the QFI is a faithful witness of non-Markovianity in Hermitian systems \cite{lu2010quantum,fujiwara2001quantum}. Therefore, we conclude that the witnesses proposed for Hermitian systems should be reexamined to check their efficiency in non-Hermitian systems.
Our work thus motivates deeper analyses to clarify the applicability of other witnesses of the non-Markovianity to detect the memory effects in Hermitian and non-Hermitian systems. Our results also pave the way to further studies on HSS applications in detecting the criticality in high-dimensional non-Hermitian multi-qudit systems.
\section*{Acknowledgements}
H.R.J. wishes to acknowledge the financial support of the MSRT of Iran and Jahrom University.
The authors would like to thank Franco Nori and Chia-Yi Ju for fruitful discussions and feedbacks.
|
\section{Introduction}
Explanations matter. In human-human interactions, they provide necessary delineations of reasoning and justification for one's thoughts and actions, and a primary vehicle to transfer knowledge from one person to another~\cite{lombrozo2012explanation}. Explanations play a central role in sense-making, decision-making, coordination, and many other aspects of our personal and social lives~\cite{hoffman2017explaining}. They are becoming increasingly important in human-AI interactions as well. As AI systems are rapidly being employed in high stakes decision-making scenarios in industries such as healthcare~\cite{loftus2020artificial}, finance~\cite{murawski2019mortgage}, college admissions~\cite{pangburn19schools}, hiring~\cite{dattner2019hiring}, and criminal justice~\cite{hao2019jail}, the need for explainability becomes paramount. Explainability is not only sought by users and other stakeholders to understand and develop appropriate trust of AI systems, but also to support discovery of new knowledge and make informed decisions~\cite{liao2020questioning}. To respond to this emerging need for explainability, there has been commendable progress in the field of Explainable AI (XAI), especially around algorithmic approaches to generate representations of how a machine learning (ML) model operates or makes decisions.
Despite the recent growth spurt in the field of XAI, studies examining how people actually interact with AI explanations have found popular XAI techniques to be ineffective~\cite{alqaraawi2020evaluating,poursabzi2018manipulating,zhang2020effect}, potentially risky~\cite{kaur2020interpreting,stumpf2016explanations}, and underused in real-world contexts~\cite{liao2020questioning}. The field has been critiqued for its techno-centric view, where ``inmates [are running] the asylum''~\cite{miller2019explanation}, based on the impression that XAI researchers often develop explanations based on their own intuition rather than the situated needs of their intended audience. Currently, the dominant algorithm-centered XAI approaches make up for only a small fragment of the landscape of explanations as studied in the Social Sciences~\cite{miller2019explanation,lombrozo2012explanation,wang2019designing,mittelstadt2019explaining} and exhibit significant gaps from how explanations are sought and produced by people. Certain techno-centric pitfalls that are deeply embedded in AI and Computer Science, such as Solutionism (always seeking technical solutions) and Formalism (seeking abstract, mathematical solutions)~\cite{selbst2019fairness, green2020algorithmic}, are likely to further widen these gaps.
One way to address the gaps would be to critically reflect on the status quo. Here, the lenses of Agre's Critical Technical Practice (CTP) ~\cite{agre1997toward,agre1997computation} can help. CTP encourages us to question the core epistemic and methodological assumptions in XAI, critically reflect on them to overcome impasses, and generate new questions and hypotheses. By bringing the unconscious aspects of experience to our conscious awareness, critical reflection makes them actionable~\cite{sengers2005reflective,dourish2004action,dourish2004reflective}. Put differently, a CTP-inspired reflective perspective on XAI~\cite{ehsan2020human} will encourage us to ask: by continuing the dominant algorithm-centered paradigm in XAI, what perspectives are we missing? How might we incorporate the marginalized perspectives to embody alternative technology? In this case, a dominant XAI approach can be construed as algorithm-centered that privileges technical transparency and circumscribes the epistemic space of explainable AI around model transparency. An algorithm-centered approach can be effective if explanations and AI systems existed in a vacuum. However, it is not the case that explanations and AI systems are devoid of situated context.
On one hand, explanations (as a construct) are socially situated~\cite{lombrozo2011instrumental, lombrozo2012explanation, wilkenfeld2015inference, miller2019explanation}. Explanation is first and foremost a shared meaning-making process that occurs between an explainer and an explainee. This process is dynamic to the goals and changing beliefs of both parties~\cite{dennett1989intentional,hume2000enquiry,hilton1996mental,heider1958psychology}. For our purposes in this paper, we adopt the broad definition that an explanation is an answer to a \textit{why}-question~\cite{miller2019explanation,dennett1989intentional,lewis1986causal}.
On the other hand, implicit in AI systems are \textit{human-AI assemblages}. Most consequential AI systems are deeply embedded in socio-organizational tapestries in which groups of humans interact with it, going beyond a 1-1 human-AI interaction paradigm. Given this understanding, we might ask: if both AI systems and explanations are socially-situated, then why are we not requiring incorporation of the social aspects when we conceptualize explainability in AI systems? How can one form a holistic understanding of an AI system and make informed decisions if one only focuses on the technical half of a sociotechnical system?
We illustrate the shortcomings of a solely technical view of explainability in the following scenario, which is inspired by incidents described by informants in our study.
\begin{quote}
\textit{You work for a leading cloud software company, responsible for determining product pricing in various markets. Your institution built a new AI-powered tool that provides pricing recommendations based on a wide variety of factors. This tool has been extensively evaluated to assist you on pricing decisions. One day, you are tasked with creating a bid to be the cloud provider for a major financial institution. The AI-powered tool gives you a recommended price. You might think, why should I trust the AI’s recommendation? You examine a variety of technical explanations the system provides: visualizations of the model's decision-making process and descriptions of how the algorithm reached this specific recommendation. Confident at the soundness of the model's recommendation, you create the bid and submit it to the client. You are disheartened to learn that the client rejected your bid and instead accepted the bid from a competitor.}
\end{quote}
Given a highly-accurate machine learning model, along with a full complement of technical explanations, why should the seller's pricing decision not have been successful? It is because the answer to the \textit{why}-question is not limited to the machine explaining itself. It is also in the situational and socio-organizational context, which one can learn from how price recommendations were handled by other sellers. What other factors went into those decisions? Were there regulatory or client-specific (e.g., internal budgetary constraints) issues that were beyond the scope of the model? Did something drastic happen in the operating environment (e.g., a global pandemic) that necessitated a different strategy? In other words, situational context matters and it is with this context the ``why'' questions could be answered effectively and completely.
At a first glance, it may seem that socio-organizational context has nothing to do with explaining an AI system. Therein lies the issue --- where we draw the boundary of our epistemic canvas for XAI matters. If the boundary is traced along the bounds of an algorithm, we risk excluding the human and social factors that significantly impact the way people make sense of a system. Sense-making is not just about opening the closed box of AI, but also about who is around the box, and the sociotechnical factors that govern the use of the AI system and the decision. Thus the ``ability'' in explainability does not lie exclusively in the guts of the AI system~\cite{ehsan2020human}. For the XAI field as a whole, if we restrict our epistemic lenses to solely focus on algorithms, we run the risk of perpetuating the aforementioned gaps, marginalizing the human and sociotechnical factors in XAI design. The lack of incorporation of the socio-organizational context is an epistemic blind spot in XAI. By identifying and critically reflecting on this epistemic blind spot, we can begin to recognize the poverty of algorithm-centered approaches.
In this paper, we address this blind spot and expand the conceptual lens of XAI by reframing explainability beyond algorithmic transparency, focusing our attention to the human and socio-organizational factors around explainability of AI systems. Building upon relevant concepts that promote transparency of social information in human-human interactions, we introduce and explore Social Transparency (ST) in AI systems. Using a scenario-based design, we create a speculative instance of AI-mediated decision-support system and use it to conduct a formative study with 29 AI users and practitioners. Our study explores whether and how proposed constitutive design elements address the epistemic blind spot of XAI -- incorporating socio-organizational contexts into explainability. We also investigate whether and how ST can facilitate AI-mediated decision-making and other user goals. This paper is not a full treatise of how to achieve socially-situated XAI; rather a first step toward that goal by operationalizing the concept in a set of design elements and considering its implications for human-AI interaction. In summary, our contributions are fourfold:
\begin{itemize}
\item We highlight an epistemic blind spot in XAI -- a lack of incorporation of socio-organizational contexts that impact the explainability of AI-mediated decisions -- by using a CTP-inspired reflective approach to XAI.
\item We explore the concept of Social Transparency (ST) in AI systems and develop a scenario-based speculative design that embodies ST, including four categories of design features that reflect \textit{What}, \textit{Why}, \textit{Who}, and \textit{When} information of past user interactions with AI systems.
\item We conduct a formative study and empirically derive a conceptual framework, highlighting three levels of context around AI-mediated decisions that are made visible by ST and their potential effects: technological (AI), decision, and organizational contexts.
\item We share design insights and potential challenges, risks, and tensions of introducing ST into AI systems.
\end{itemize}
\section{Related work}
We begin with a in-depth review of related work in XAI field, further highlighting the danger of the epistemic blind spot. We then discuss a shift in broader AI related work towards sociotechnical perspectives. Lastly, we review work that pushed towards transparency of socio-organizational contexts in human-human interactions.
\subsection{Explainable AI (XAI)}
Although there is no established consensus on the complete set of factors that makes an AI system explainable, XAI work commonly shares the goal of making an AI system's functioning or decisions \textit{easy to understand} by people~\cite{lipton2018mythos,arrieta2020explainable,gilpin2018explaining,miller2019explanation,gunning2017explainable,ras2018explanation,carvalho2019machine, ehsan2019automated}. Recent work also emphasizes that explainability is an audience-dependant instead of a model-inherent property~\cite{arrieta2020explainable,mohseni2018multidisciplinary,arya2019one,miller2019explanation, ehsan2020human}. Explainability is often viewed more broadly than model transparency or intelligibility~\cite{gilpin2018explaining,ras2018explanation,lipton2018mythos}. For example, a growing research area of XAI focuses on techniques to generate \textit{post-hoc} explanations~\cite{ehsan2019automated}. Instead of directly elucidating how a model works internally, post-hoc explanations typically justify an opaque' model's decision by rationalizing the input and output or providing similar examples. Lipton discussed the importance of post-hoc explanations to provide useful information for decision makers, and its similarity with how humans explain~\cite{lipton2018mythos}. At a high level, Gilpin et al.~\cite{gilpin2018explaining} argued that the transparency of model behaviors alone is not enough to satisfy the goal of ``\textit{gain[ing] user trust or produc[ing] insights about the cause of the decisions},'' but rather, explainability requires other capabilities such as providing responses to user questions and the ability to be audited.
Since an explanation is only explanatory if it can be consumed by the recipient, many recognize the importance of taking user-centered approaches to XAI~\cite{miller2019explanation,shneiderman2020human,vaughan20201}, and the indispensable role that the HCI community should play in advancing the field. While XAI has experienced a recent surge in activities, the HCI community has a long history of developing and studying explainable systems, such as explainable recommender systems, context-aware systems, and intelligent agents, as outlined by Abdul et al.~\cite{abdul2018trends}. Moreover, XAI's disconnect with the philosophical and psychological grounds of human explanations has been duly noted~\cite{mittelstadt2019explaining}, as best represented by Miller's call for leveraging insights from the Social Sciences~\cite{miller2019explanation}. Wang et al. reviewed decision-making theories and identified many gaps in XAI output to support the complete cognitive processes of human reasoning~\cite{wang2019designing}. From these lines of work, we highlight a few critical issues that are most relevant to our work.
First, there is a dearth of user studies and a lack of understanding on how people actually perceive and consume AI explanations~\cite{doshi2017towards,vaughan20201}. Only until recently have researchers began to conduct controlled lab studies to rigorously evaluate popular XAI techniques~\cite{buccinca2020proxy,cai2019effects,dodge2019explaining,poursabzi2018manipulating,cheng2019explaining,lai2020chicago, ehsan2019automated}, as well as studies to understand real-world user needs for AI explainability~\cite{liao2020questioning,kaur2020interpreting,hong2020human}. Accumulating evidence shows that XAI techniques are not as effective as assumed. There have been rather mixed results on whether current XAI techniques could appropriately enhance user trust~\cite{cheng2019explaining,poursabzi2018manipulating,yang2020visual} or the intended task performance, whether for decision making~\cite{liao2020questioning,zhang2020effect,buccinca2020proxy}, model evaluation~\cite{cai2019effects,alqaraawi2020evaluating,dodge2019explaining}, or model development~\cite{kaur2020interpreting}. For example, Alqarrawi et al. evaluated the effectiveness of saliency maps~\cite{alqaraawi2020evaluating} -- a popular explanation technique for image classification models -- and found they provided very limited help for evaluating the model. Kauer et al. studied how data scientists use popular model interpretability tools and found them to be frequently misused~\cite{kaur2020interpreting}. Liao et al. interviewed practitioners designing AI systems and reported their struggle with popular XAI techniques due to a lack of actionability for end users. Recent studies also reported detrimental effects of explanations for AI system users including inducing over-trust or over-estimation of model capabilities~\cite{kaur2020interpreting, stumpf2016explanations,smith2020no}, and increasing cognitive workload~\cite{abdul2020cogam,Ghai2020XAL}. Moreover, while XAI is often claimed to be a critical step towards accountable AI, empirical studies have found little evidence that explanations improve a user's perceived accountability or control over AI systems~\cite{rader2018explanations,smith2020no}.
Second, in human reasoning and learning, explanation is both a \textit{product} and a \textit{process}. In particular, it is a \textit{social process}~\cite{miller2019explanation} as part of
a conversation or social interaction. Current technical XAI work typically takes a product-oriented view by generating a representation of a model's internals~\cite{lombrozo2012explanation}. However, explanations are also sought first and foremost as a knowledge transfer process from an explainer to an explainee. A process-oriented view has at least two implications for XAI. First, the primary goal of explanation should be to enable the explainee to gain knowledge or make sense of a situation or event, which may not be limited to a model's internals. Second, as a transfer of knowledge, explanations should be presented relative to the explainee's beliefs or knowledge gaps~\cite{miller2019explanation}. This emphasis on tailoring explanation according to explainee's knowledge gaps has been a focus of prior HCI work on explainable systems~\cite{lim2009and,lim2010toolkit,lim2019these,liao2020questioning}. Recent work has also begun to explore interactive explanations that could address users' follow-up questions as a way to fill individual knowledge gaps~\cite{weld2019challenge,spinner2019explainer}. However, sometimes these knowledge gaps lie outside of the system, which may require providing information that is not related to its internal mechanics~\cite{abdul2018trends}.
Finally, we argue that AI systems are socially situated, but sociotechnical perspectives are mostly absent in current XAI work. One recent study by Hong et al.~\cite{hong2020human} investigated how practitioners view and use XAI tools in organizations using ML models. Their findings suggest that the process of interpreting or making sense of an AI system frequently involves cooperation and mental model comparison between people in different roles, aimed at building trust not only between people and the AI system, but also between people within the organization~\cite{hong2020human}. Our work builds on these observations, as well as prior work on sociotechnical approaches to AI systems which we review below.
\subsection{Sociotechnical approaches to AI}
Our work is broadly motivated by work on sociotechnical approaches to AI. Academia and society at large have begun to recognize the detrimental effect of a techno-centric view on AI~\cite{shneiderman2020human,vaughan20201,sabanovic2010robots}. Since AI systems are socially situated, their development should carefully consider social, organizational, and cultural factors that may govern their usage. Otherwise one may risk deploying an AI system un-integrated into individual and organizational workflows~\cite{makarius2020rising,wolf2019evaluating}, potentially resulting in misuse, mistrust~\cite{yang2016investigating,yang2019unremarkable}, or having profound ethical risks and unintended consequences, especially for marginalized groups~\cite{mohamed2020decolonial,suresh2019framework,sanchez2020does}.
Researchers have proposed ways to make AI systems more human-centered and sensitive to socio-organizational contexts.
Bridging rich veins of work in AI, HCI, and critical theory, such as Critical Technical Practices~\cite{agre1997computation} and Reflective Design~\cite{sengers2005reflective}, Ehsan and Riedl delineate the foundations of a Reflective Human-centered XAI (HCXAI). \textit{Reflective HCXAI} is a sociotechnically informed perspective on XAI that is critically reflective of dominant assumptions and practices of the field~\cite{ehsan2020human}, and sensitive to the values of diverse stakeholders, especially marginalized groups, in its proposal of alternative technology. Zhu et al. proposed Value Sensitive Algorithm Design~\cite{zhu2018value} by engaging stakeholders in the early stages of algorithm creation, to avoid biases in design choices or compromising stakeholder values. Several researchers have leveraged design fictions and speculative scenarios to elicit user values and cultural perspectives for AI system design~\cite{cheon2016integrating,cheon2018futuristic,muller2017exploring}. Šabanovic developed a framework of Mutual-Shaping and Co-production~\cite{sabanovic2010robots} by involving users in the early stages of robot design and engaging in reflexive practices. Jones et al~\cite{jones2013design} proposed a design process for intelligent sociotechnical systems with equal attention to analysis of social concepts in the deployment context and representing such concepts in computational forms.
More fundamentally, using a Science and Technology Studies (STS) lens~\cite{suchman1987plans}, scholars have begun critically reflecting on the underlying assumptions made by AI algorithmic solutions. Mohamed et al.~\cite{mohamed2020decolonial} examined the roles of power embedded in AI algorithms, and suggested applying decolonial approaches to enable AI technologies to center on vulnerable groups that may bear negative consequences of technical innovation. Green and Viljoen~\cite{green2020algorithmic} diagnosed the dominant mode of AI algorithmic reasoning as ``algorithmic formalism'' -- an adherence to prescribed form and rules -- which could lead to harmful outcomes such as reproducing existing social conditions and a technologically-deterministic view of social changes. The authors pointed out that addressing these potential harms requires attending to the internal limits of algorithms and the social concerns that fall beyond the bounds of algorithmic formalism. In the context of fair ML, Selbst et al.\cite{selbst2019fairness} questioned the implications of algorithmic abstraction that are essential to ML. Abstracting away the broader social context can cause AI technical interventions to fall into a number of traps: Framing, Portability, Formalism, Ripple Effect, and Solutionism. The authors suggested to mitigate these problems by extending abstraction boundaries to include social factors rather than purely technical ones. In a similar vein, field work on algorithmic fairness often found that meaningful interventions toward usable and ethical algorithmic systems are non-technical, and that user community derive most value from localized, as opposed to ``scalable'' solutions~\cite{katell2020toward,lee2019webuildai}.
Our work is aligned with and builds on these views obtained through the sociotechnical lens. These perspectives inform our thinking as we expand the boundaries of XAI to include socio-organizational factors, and challenge a formalist perspective that peoples' meaning-making processes could be resolved through algorithmic formalisms. Our work takes an operational step towards sociotechnical XAI systems by expanding the design space with ST.
\subsection{Social transparency and related concepts}
Our work is also informed by prior work that studied social transparency and related concepts in human-human interactions. The concept of making others' activities transparent plays a central role in HCI and Computer-Supported Cooperative Work (CSCW) literature~\cite{suchman1995making,star1999layers}. Erickson and Kellogg proposed the concept of and design principles for Social Translucence, in which ``social cues'' of others' presence and activities are made visible in digital systems, so that people can apply familiar social rules to facilitate effective online communication and collaboration. Gutwin et al.'s seminal work on group awareness~\cite{gutwin2002descriptive} for groupware supporting distributed teams provides an operational design framework. It sets out elements of knowledge that constitute group awareness, including knowledge regarding \textit{Who}, \textit{What}, and \textit{Where} to support awareness related to the present, and \textit{How}, \textit{When}, \textit{Who}, \textit{Where}, and \textit{What} for awareness related to the past. Theses theories have since inspired a bulk of work that created new design features and design spaces for social and collaborative technologies (e.g.~\cite{erickson2003social,mcdonald2012building,gilbert2012designing,gutwin2004group,kim2020enriched}).
Building upon social translucence and awareness, Stuart et al.~\cite{stuart2012social} conceptualized Social Transparency (ST) in networked information exchange. In particular, it extends the visibility of one's direct partner and the effect on their dyadic interactions, to also encompass one's role as an \textit{observer} of others' interactions made visible in the network. Their framework describes three social dimensions made visible to people by ST: identity transparency, content transparency, and interaction transparency. This framework then considers a list of \textit{social inferences} people could make based on these visible dimensions (e.g. perceived similarity and accountability based on identity transparency; activity awareness based on content transparency; norms and social networks based on interaction transparency), and their second order effects for the groups or community. Social transparency theory has been used to design and analyze various social media features and their impact on social learning~\cite{dabbish2012social,nguyen2015perverse}, social facilitation~\cite{huang2013don}, and reputation management~\cite{dabbish2012social}.
The above work focused on how ST -- making others' activities visible -- affects collaboration and cooperative behaviors with other people. Our work also draws on two other important aspects that ST could potentially support for decision-making. One is on knowledge sharing and acquisition. As reviewed by Ackerman et al,~\cite{ackerman2013sharing}, CSCW systems supporting organizational knowledge management fall into two categories: a repository model that externalizes peoples' knowledge as sharable artifacts or objects; and an expertise-sharing model that supports locating the appropriate person to have in-situ access to knowledge. The CSCW community's shift from the former to the latter category represents a shift of emphasis from \textit{explicit} to \textit{tacit} knowledge. Transparency of others' communications could facilitate expertise location through the acquisition of organizational meta-knowledge (e.g., \textit{who knows what} and \textit{who knows whom}), as a type of ``ambient awareness'' coined by Leonardi in the context of enterprise social media~\cite{leonardi2014social,leonardi2015ambient}.
This position is also related to the development of \textit{Transactive Memory Systems (TMS)}~\cite{yoo2001developments,moreland2006transactive,brandon2004transactive,nardi2002networkers} that relies on meta-knowledge to optimize the storage and retrieval of knowledge across different individuals. A sufficiently fluent TMS can evolve to a form of team cognition of or ``collective mind''~\cite{weick1993collective,hutchins1991social} that can lead to better collective performance~\cite{hollingshead2003potential,austin2003transactive}.
Social transparency could also guide or validate peoples' judgment and decision as cognitive heuristics. Cognitive heuristics are a key concept in decision-making~\cite{kahneman1982judgment}, which refers to ``rules of thumb'' people follow to quickly form judgments or find solutions to complex problems. By making visible what other people selected, interacted with, approved or disapproved, ST could invoke many social and group-based heuristics such as bandwagon or endorsement heuristics (following what many others' do), authority or reputation heuristics (following authority), similarity heuristics (following people in similar situations), and social presence heuristics (favoring a social entity over a machine)~\cite{sundar2008main,metzger2010social,metzger2013credibility}. How these ST-rendered heuristics affect peoples' decisions and actions has been studied in a wide range of technologies such as reputation systems~\cite{resnick2000reputation} and social media. In particular, they play a critical role in how people evaluate the trustworthiness, credibility, and agency of technologies~\cite{sundar2008main,metzger2010social,metzger2013credibility}, as well as the sources or organizations behind the technologies~\cite{huang2013don,kramer1999trust}. While these heuristic-based judgments are indispensable for people to navigate complex technological and social environments, they also lead to biases and errors if inappropriately applied~\cite{kahneman1982judgment}, calling for careful study of inferences people make based on ST features and their potential effect.
Our concept of social transparency in AI systems is informed by the aforementioned perspectives, but with several key distinctions: at the center of our work is a desire to support the explainability of AI systems, particularly in AI-mediated decision-making. We are not merely interested in making \textit{others'} activities visible, but more importantly, how \textit{others' interactions with AI} impact the explainability of the system. Within the view of a human-AI assemblage, in which both AI and people have decision-making agency, it is possible to borrow ideas and interpretative lenses from work studying ST in human-human interactions. To study the effects of ST in AI systems, our first-order focus is on users' sense-making of an AI system and their decision-making process, though it may inevitably impact their organizational behaviors as well.
\section{Social Transparency in AI systems: a scenario based design exploration}
\label{4w}
After identifying an epistemic blind spot of XAI, we propose adding Social Transparency (ST) into AI systems--incorporating social-organizational contexts to facilitate explainability of AI’s recommendations. This definition is intentionally left broad, as we follow a broad definition of explainability--ability to answer the why-question. We borrow the term ST from Stuart et al.~\cite{stuart2012social}, and similarly emphasize both making visible of other people in the human-AI assemblage, and other people’s interactions with the “source”, in our case, the AI system. Different from Stuart et al., which proposed the ST concept retrospectively at a time when ST enabling features were pervasive in CSCW systems, we had to consider, prospectively, what kind of features to add to an AI system to make ST possible.
As a formative step, our goal was not to develop a finished treatise of ST in AI systems. Rather, we intended to create an exemplary design of an AI system with ST and use it to conduct formative studies to advance our conceptual development. We opted for a scenario-based design (SBD) method. SBD suspends the needs to define system operations by using narrative descriptions of how a user uses a system to accomplish a task~\cite{rosson2009scenario}. SBD allows interpretive flexibility in a user journey by balancing between roughness and concreteness. SBD is an appropriate choice for our investigation because it is a method oriented for “envisioning future use possibilities” ~\cite{rosson2009scenario}, focusing on people’s needs, evocative, and has been adopted in prior XAI design work~\cite{wolf2019evaluating}.
We started with a range of AI-mediated decision-making scenarios around cybersecurity, hiring (employment), healthcare, and sales, where a user encounters an AI recommendation and seeks answer to a \textit{why}-questions about the recommendation, e.g. “why should I accept or trust the recommendation”. We ran 4 workshops with a total of 21 people from 8 technology companies who are users or stakeholders of relevant AI systems. The scenarios started in a textual form, then we engaged participants in drawing exercises to create visual mock-ups of these scenarios (hereby referred to as visual scenarios), and brainstorming together what kind of information they wanted to see about \textit{other users} of the AI system, and \textit{other users’ interactions with the AI system} if they were the user. When it came to types of design feature that could encode relevant socio-organizational context, people had many suggestions. For instance, suggestions about knowing what happened to other people getting recommendations from the AI systems, who got the recommendations, etc. quickly emerged in the discussions. The ideas converged to what our participants coined as the “4W”—\textit{who} did \textit{what} with the AI system, \textit{when}, and \textit{why} they did what they did— in order to have adequate socio-organizational context around the AI-mediated decisions.
We note an interesting observation that the 4W share similarity with the design elements for group awareness in groupware work~\cite{gutwin2002descriptive}, with the exception of “why”, which is core to explainability. When thinking how to represent the “why”, participants suggested an open ended textual representation to capture the nuances behind a decision. Eventually, we settled on a design of a “commenting” feature (why) together with traces of others’ interactions with the AI system’s recommendations (what), their identities (who) and time of interactions (when). In the rest of the paper, we refer to these constitutive design elements of ST as 4W. Figure~\ref{fig:visual scenario} shows the final visual scenario with the 4W features used in the interview study.
We chose a sales scenario around an AI-mediated price recommendation tool, since it appeared to have a broader reach and accessibility even for workshop participants who did not work in a sales domain. In the study, we intended to interview sellers as targeted users of such an AI system, and also non-sellers to explore the transferability of the ST concept to other AI domains, as we will discuss in detail in the next section.
\begin{figure}[tbh]
\caption{Visual scenario used in the interviews, labeled by blocks to be revealed in the interview in order: (1) Decision information and model explanation: Information of the current sales decision, the AI's recommended price and a ``feature importance'' explanation justifying the model's recommendation, inspired by real-world pricing tools; (2) ST summary: Beginning of ST giving a high-level summary of how many teammates in the past had received the recommendation and how many sold at the recommended price; (3-5): ST blocks with "4W" features containing the historical decision trajectory of three other users.}
\centering
\includegraphics[width=14.95cm]{Probe_picture_v2.png}
\label{fig:visual scenario}
\Description[Figure showing the anatomy of the visual scenario (sales context) we explored in the interviews.]{{Visual scenario used in the interviews, by labeled blocks: (1) Decision information and model explanation: Information of the current sales decision, the AI's recommended price and a ``feature importance'' explanation justifying the model's recommendation, inspired by real-world pricing tools; (2) ST summary: Beginning of ST giving a high-level summary of how many teammates in the past had received the recommendation and how many sold at the recommended price; (3-5): ST blocks with "4W" features containing the historical decision trajectory of from each of the colleagues.}}
\end{figure}
\textit{Design choices in the visual scenario:} We ran 4 pilot studies to finalize the design of the visual scenario in Figure~\ref{fig:visual scenario}, and the procedure to engage participants with the design. We scoped the number of 4W blocks to three to strike a balance between a variety of ST information and avoiding overwhelming the participants, based on what we learned from the pilot studies. Each of the 4W are represented by one or more design features: accepting and rejecting the AI (action [what]), succeeding and failing to make the sale (outcome [what]), one's name, profile picture and organizational role ([who]), a comment on the reasons behind the action ([why]), and a time stamp ([when]). Contents in these components were inspired by the workshop discussions, and showcase a range of socio-organizational contexts relevant to the decision. The pilot runs revealed that presenting the entire visual scenario creates cognitive and visual clutter. Therefore, for the interview, we decided to reveal the five blocks shown in Figure~\ref{fig:visual scenario} one by one, with the interviewer verbally presenting the narrative around each block.
\section{Study Methods}
In this section we share the methodological details of the semi-structured interviews.
\subsection{Recruitment}
As mentioned, we intended to recruit both sellers and non-sellers, who are stakeholders of other AI-mediated decision-making domains. Stakeholders are not limited to end users. We also welcomed different perspectives from designers, data scientists, etc. With this in mind, we recruited participants from six different companies, including a large international technology company where we were able to recruit from multiple lines of products or sales divisions. The recruitment was initiated with an online advertisement posted in company-wide group-chat channels that we considered relevant, followed up by snowball sampling. The advertisement stated two recruiting criteria: First, they needed to have direct experience using or developing or designing an AI system. Second, the AI system should be interacted by multiple users, preferably with multi-user decision-making. We verified that these criteria were met through a series of correspondence (via online messaging) where each participant shared samples of the AI system they intended to discuss.
A total of 29 participants were recruited (17 self-identified as females while the rest as males). The recruitment of sellers turned out to be challenging, given their very limited availability. By using snowball sampling, we were able to recruit 8 sellers. For non-sellers, the snowball sampling resulted in participants clustered in two major domains – healthcare and cybersecurity. We conducted the study in the middle of Covid-19 (a global pandemic in 2020), which added non-trivial burden to the recruitment process and limited our interviews to a remote setting using video conferencing tools. Participants’ ID, role, domains and domain experience is shared in Table~\ref{table:participant details}. To facilitate traceability in the data presented hereafter, we differentiate sellers and non-sellers by appending the participant ID with \textit{-S} for sellers and \textit{-NS} for non-sellers (e.g., 1-S for a seller and 2-NS for a non-seller).
\subsection{Interview procedure}
The semi-structured interviews were conducted online with screen-sharing for the interviewer to present the visual scenario. All interviews were video recorded including the screen activities. The interview had 4 main parts. In the \textit{first} part, after gaining informed consent, we asked participants to share about an AI system that they were currently engaged with, focusing on their or their users’ needs for explainability. We also inquired about the socio-organizational context around the use case, both before and after the AI system was introduced.
The \textit{second} part involved a deep dive into the speculative design with a walk-through of the visual scenario in Figure~\ref{fig:visual scenario}. This is where we explored how incorporation of ST can impact an AI-mediated decision-making scenario, as we revealed the different blocks of the visual scenario in a sequenced manner. Participants were asked to play the role of a salesperson trying to pitch a good price for an Access Management software to Scout Inc. (a client). In the first block revealed, the AI not only recommends a price, but also shows a technical explanation--a set of model features (e.g., cost price, quota goals, etc.) justifying the recommendation. Once the participant showed a good enough understanding on the Decision Information and Model Explanation portion (block 1 in Figure~\ref{fig:visual scenario}), we asked the participant to give a price they would offer and their confidence level (between 1-10, 10 being extremely confident) given what they saw on the screen. Next, we revealed the social transparency portions. First, it was the ST Summary (block 2 in Figure~\ref{fig:visual scenario}) followed by each of the 4W blocks (block 3-5 in Figure~\ref{fig:visual scenario}). We allowed participants to read through the content and guided them through any misunderstandings. They were encouraged to think-aloud during the whole process. Following this, we asked participants to share the top three reactions to the addition of the ST features, either positive or critical. After that, participants were asked to share their final price and confidence level. In addition, we asked them to rank the importance of the 4W (\textit{who}, \textit{what}, \textit{when}, and \textit{why}) for their decision-making process and justify their ranking.
The \textit{third }part was about zooming out from the visual scenario and brainstorming plausible and impactful transfer scenarios of ST in domains our participants resided. At this point, we also gave them a conceptual definition and some vocabulary around ST so that they could brainstorm with us effectively. The goal of this part was to explore the design and conceptual space of ST in domains beyond the sales scenario. For sellers, this meant transferring to their own sales work context, which helped refining our own understanding of the sales scenario. Once participants shared their thoughts on the transferability of ST, they ranked the 4W in the transfer use cases. We wanted to see if there are variations in the rankings as the context switches—an aspect we discuss in the Findings section.
The \textit{fourth} and final part involved discussions around potential unwanted or negative consequences of ST as well as reflective conversations on how incorporation of ST can impact explainability of AI systems.
In summary, in addition to open-ended discussions, our interview collected the following data points from each participant: original and updated price decisions and associated confidence levels, rankings of 4W for both the sales scenario and one's own domain. While our study was not designed to quantitatively evaluate the effect of ST, we will report summary statistics of these data points in the Findings section, which helped guiding our qualitative analysis.
\subsection{Qualitative Analysis of the interviews}
The interviews lasted 58 minutes on average. We analyzed the transcription of roughly 29 hours of interview data using a combination of thematic analysis\cite{braun2006using} and grounded theory~\cite{strauss1994grounded}. Using an open coding scheme, two authors independently went through the videos and transcription to produce in-vivo codes (directly from the data itself). Then we separately performed a thematic analysis, clustering the codes from in-vivo coding to themes. We iteratively discussed and agreed upon the codes and themes, constantly comparing and contrasting the topics each of us found, refining and reducing the variations in each round till consensus was reached. We grouped the codes and themes at the topic level using a combination of mind-mapping and affinity diagramming. Our results section below is organized thematically, with the top-level topics as subsections. When discussing each topic, we highlight codes that add to that topic in \textbf{bold}.
\begin{table*}
\centering
\caption{Participant details}
\label{table:participant details}
\begin{tabular}{llll}
\toprule
\multicolumn{1}{l}{Participant ID} & Role & Domain & \multicolumn{1}{l}{Years of Experience~} \\
\midrule
1-S & Seller & Sales & > 10 \\
2-NS & Designer & Cybersecurity & > 5 \\
3-NS & Designer & Finance and Travel & > 5 \\
4-NS & Consultant & Gov and Non-profit & > 3 \\
5-S & Seller & Sales & > 5 \\
6-NS & Designer & Health- Oncology & > 5 \\
7-NS & Data Scientist & Cybersecurity & > 8 \\
8-S & Seller & Sales & > 3 \\
9-NS & Designer & Health- Radiology & > 5 \\
10-NS & Data Scientist & Cybersecurity & > 10 \\
11-NS & Designer & Health & > 3 \\
12-NS & Designer & Cybersecurity & > 5 \\
13-S & Seller & Data Analytics & > 10 \\
14-NS & Data Scientist & NLP & > 5 \\
15-NS & Designer & Health- Radiology & > 5 \\
16-NS & MD/ Data Scientist & Health- Oncology & > 10 \\
17-NS & Manager & HR & > 5 \\
18-S & Seller & Sales & > 3 \\
19-S & Seller & Sales & > 3 \\
20-S & Seller & Sales & > 10 \\
21-S & Seller & Sales & > 3 \\
22-NS & SOC analyst & Cybersecurity & > 3 \\
23-S & Seller & Sales & > 5 \\
24-NS & SOC analyst & Cybersecurity & > 5 \\
25-NS & SOC analyst & Cybersecurity & > 3 \\
26-NS & SOC analyst & Cybersecurity & > 5 \\
27-NS & SOC Data Scientist & Cybersecurity & > 5 \\
28-NS & SOC Architect & Cybersecurity & > 10 \\
29-NS & SOC analyst & Cybersecurity & > 5 \\
\bottomrule
\end{tabular}
\Description[Table showing participant details]{In this table, we share the participant details. From left to right, the columns are Participant ID, Role, Domain, and Years of Experience}
\end{table*}
\section{Findings}
We begin by sharing how participants' own experience with AI systems demonstrates that technical transparency alone does not meet their explainability needs. They need context beyond the limits of the algorithm. Next, based on how participants reacted to the incorporation of ST in the design scenario, we unpack what context could be made visible by ST and break down the implications at three levels: \textbf{technological (AI)}, \textbf{decision-making}, and \textbf{organizational}, as summarized in Table~\ref{tab:conceptual framework}. We further discuss specific aspects of socio-organizational context that the 4W design features carry and their effects, summarized in Table~\ref{tab:4W}. Based on input from non-seller participants, we also share insights about the potential transferability of ST beyond the sales domain. We end this section with participants’ discussions on the challenges, risks, and tensions of introducing ST into AI systems.
As mentioned, all participants, both sellers and non-sellers, experienced the sales scenario and reflected on its transferability to their own domains. Our analysis revealed substantial alignment between the two groups, possibly due to the accessible nature of our intentional choice of a sales domain and the content in the scenario. With the exception of Section~\ref{transferability}, which focuses on non-sellers' reflection on transferability of ST, we report the results combining the two groups, but mark their IDs differently (-\textit{S} or -\textit{NS}) as shwon in Table~\ref{table:participant details}.
\subsection{Technical Transparency is not enough}
As we began each interview with participants' own experience with AI systems, a core theme that lies at the heart of our findings is the realization that solely relying on technical or algorithmic transparency is not enough to empower complex decision-making. There is a shared understanding that AI algorithms cannot take into account all the contextual factors that matter for a decision: “not everything that you need to actually make the right decision for the client and the company is found in the data” (P25-NS). Participants pointed to the fact that even with an accurate and algorithmically sound recommendation, “there are things [they] never expect a machine to know [such as] clients’ allegiances or internal projects impacting budget behavior” (P1-S). Often, the context of social dynamics that an algorithm is unable to capture is the key: “real life is more than numbers, especially when you think of relationships” (P12-NS). Discussing challenges in interpreting and using AI recommendations in Security Operation Centers (SOC), P29-NS highlighted the need for awareness of others’ activities in the organizational context:
\begin{quote}
Sometimes, even with perfect AI, the most secure thing is to do nothing because you don’t know what the machine doesn't know. There is no centralized process to tell us the context of what's going on elsewhere, what others are doing. One move has ripple effects, you know. So instead of using [the AI’s recommendation], they end up basically doing the most secure thing-- don't touch anything. That’s where the context helps from your colleagues. That’s how actually work really gets done. (P29-NS, a SOC director)
\end{quote}
Moreover, even when provided, technical transparency is not always understandable for end users. While describing how he uses an AI-assisted pricing tool, this seller pointed to how the machine explained itself by sharing a “confidence interval” along with a description of how the AI works, which was meaningless to him:
\begin{quote}
I hate how it just gives me a confidence level and gibberish that the engineers will understand. There is zero context. The only reason I am able to use this tool is [through] guidance from other sellers who gave me the background information on the lead I needed to generate a quote worth their time. (P23-S, senior salesperson using a pricing tool to generate a quote)
\end{quote}
In complex organizational settings, answers to the why-question, i.e. knowledge needed to understand and take informed action for an AI mediated decision, might lie outside the bounds of the machine. As highlighted above, participants repeatedly desired for ``context'' to ``fill in the gaps'' (P27-NS). The ST information in our design scenario is intended to provide such context. After going through the ST portion, 26 out of the 29 participants lowered their sales prices, resulting in a mean final price of $\$73.8$ (SD=$\$15.8$), compared to a mean initial price of $\$110.7$ (SD=$\$57.2$) based only on the AI's recommended price of $\$100$. 24 out of the 29 participants also increased the confidence ratings for their decisions, resulting in a mean final confidence score of $8.3$ out of 10 (SD=$0.9$), compared to a mean initial confidence score of $6.4$ (SD=1.7). These patterns suggest that ST information helped participants to set their price more cautiously and feel more confident about their decisions, by “help[ing] [them] to understand the situation more holistically” (P19-S). This participant succinctly summarized this perspective at the end of her interview:
\begin{quote}
You can’t just get everything in the data that trains the model. The world doesn’t run that way. So why rely just on the machine to make sense of things that are beyond it? To get a holistic sense of the "why" you should or should not trust the AI, you need context. So the context from Social Transparency adds the missing piece to the puzzle of AI explainability”. (P2-NS, a designer of cybersecurity AI systems)
\end{quote}
Now we analyze participants' reaction and reflection from seeing ST features, and unpack the “context” made visible by ST and its effects at three levels: \textbf{technological (AI)}, \textbf{decision-making}, and \textbf{organizational}. For the subsections below each dedicated to a level of context, we begin by summarizing the effects of ST, with codes from the data in bold. These results are summarized in Table~\ref{tab:conceptual framework}.
\begin{table*}
\centering
\caption{Results on the three levels of context made visible by ST and their effects. ``--'' in the last column indicates first-order to second-order effect(s)}
\label{tab:conceptual framework}
\resizebox{\linewidth}{!}{%
\begin{tabular}{>{\hspace{0pt}}p{0.152\linewidth}>{\hspace{0pt}}p{0.3\linewidth}>{\hspace{0pt}}p{0.479\linewidth}}
\toprule
\textbf{Levels} & \textbf{Context made visible} & \textbf{Effects of the visibility} \\
\midrule
Technological (AI) & Trajectory of AI’s past decision outputs and people’s interactions with these outputs & Tracking AI performance --~ Calibrate AI trust~ ~ ~ \par{} Infusing human elements --~ Calibrate AI trust~ \\
\midrule
Decision-making & Local context of past decisions and in-situ access to decision-related (crew) knowledge & Actionable insights -- Improve decisions; Boost decision~ confidence; Support follow-up actions \par{} Social validation~ -- Decision-making resilience;~ ~ ~AI contestability\textit{~} \\
\midrule
Organizational & Organizational meta-knowledge and practices &
Understanding organizational norms and values -- Improve decisions; Set job expectation\par
Fostering auditability and accountability\par Expertise location -- Develop TMS \\
\bottomrule
\end{tabular}
}
\Description[Table showing implications of ST.]{This table shows the results on the three levels of context made visible by ST and their effects. ``--'' in the last column indicates first-order to second-order effect(s)}
\end{table*}
\subsection{Technological (AI) context made visible}
ST makes visible the socially-situated \textbf{technological context}: the trajectory of AI’s past decision outputs as well as people’s interactions with these technological outputs. Such contextual information could help people \textbf{calibrate trust in AI}, not only through \textbf{tracking AI performance}, but also by \textbf{infusing human elements} in AI that could invoke social-based perception and heuristics.
Records of others’ past interactions with the AI system paints a concrete picture of the AI performance, which technical XAI solutions such as performance metrics or model internals would not be able to communicate. Participants felt that the technological context they understood through ST helped them better gauge the AI’s limitations or “actual performance of the AI” (P10-NS). In fact, after going through the sales scenario, many reported on re-calibrating their trust in the AI, which is key to preventing both over-reliance of AI and “AI aversion”~\cite{dietvorst2015algorithm}:
\begin{quote}
Knowing the past context helps me understand that the AI wasn’t perfect. It’s almost like a reality check. The comments helped because real life is more than numbers. I am more confident in myself that I am making the right decision but less trust[ing] the AI. (P12-NS, an XAI designer)
\end{quote}
ST could also affect people’s perception of and trust in the AI system by infusing the much needed human elements of decision-making in the machine. Participants from each of the domains (sales, cybersecurity, healthcare) highlighted that ``there is a human aspect to [their] practice'' (P6-NS), something that ``can never be replaced by a machine'' (P6-NS). Adding these human elements allows one to apply familiar social rules. Many participants commented on a “transitive trust” (P4-NS) from trusting their peers -- ``people are trained to believe [their] peers and trust them'' (P25-NS) -- to trusting the AI system, if others were using the AI systems or accepting the AI’s recommendations. For instance, in the sales domain of the scenario, a transitive trust is often fostered by an organizational hierarchy or job seniority ``as precedence and permission for doing the right thing’’ (P12-NS). Radiologists often want to “know who else used the same logic and for what reason” (P15-NS) when working with AI-powered diagnostic tools. In cybersecurity, “knowing that a senior analyst took a certain route with the recommendation [can be] the difference maker” (P28-NS). Some participants also commented on a positively perceived “humanizing effect” of AI by adding ST, that “[users] would potentially adopt... showing them like [AI is] supporting you not replacing you'' (P6-NS).
The above discussions show that ST could support forming appropriate trust and evaluation of AI through two essential routes, as established in prior work on trust and credibility judgment of technologies~\cite{metzger2010social,metzger2013credibility,sundar2008main}: a central route that is based on a better understanding of the AI system, and a peripheral route by applying social or group-based heuristics such as social endorsement, authority, identity, or social presence heuristics. While the central route tends to be cognitively demanding, the peripheral route is fast and easy, and could be especially impactful to help new users to enhance their trust and adoption of an AI system.
\subsection{Decision-making context made visible}
ST also makes visible the \textbf{decision context} -- the local context of past decisions -- for which many participants described as “in-situ access” to “crew knowledge”
\footnote{The original term used by most participants was "tribal knowledge", which is a term often used in business and management science to refer to unwritten knowledge within a company. We note that, from an Indigenous perspective particularly in North America, the words "tribe" and "tribal" connote both an official status as a recognized Nation, and also a profound sense of identity, often rooted in cultural heritage, a specific ancestral place, and a lived experience of the on-going presence of tribal elders and ancestors (past, present, and future). In our case, participants used the word "tribal" in a non-Indigenous meaning. Being sensitive to potential mis-use of the word, we engaged in critical conversations with potentially affected community members to understand their perspectives. The conversations revealed that it is best to avoid using that word. We went back to the participants who used the word "tribal" and asked if "crew" captures the essence of what they meant by "tribe". All of them agreed that the words were interchangeable. As such, we only present the data using the term "crew knowledge". }.
We will first elaborate on the notion of \textbf{crew knowledge}, then discuss how a combination of decision trajectory, historical context and elements of crew knowledge could (1) lead to \textbf{actionable insights}, which could \textbf{improve decision-making}, \textbf{boost decision confidence} and \textbf{support follow-up actions}; (2) provide social validation that facilitates \textbf{decision-making resilience} and \textbf{contestability of AI}.
The notion of \textit{crew knowledge} emerged during our discussions with many participants regardless of their domains. When asked to elaborate on the concept, participants defined it as “informal knowledge acquired over time through hands-on experience”, knowledge that is not typically “gained through formal means, but knowledge that’s essential to do the job” (P8-S). Crew knowledge is learned "via informal means, mainly through colleague interactions” (P23-S). It can encode “idiosyncrasies like client specific quirks” (P27-NS). Participants referred to their team as their “crew", with a sense of identity and belonging to a community membership. We can think of crew knowledge as informal or tacit knowledge that is acquired over time and locally-situated in a tight-knit community of practice--an aggregated set of “know-hows” of sorts. While ST features may not explicitly encode a complete set of crew knowledge, they provide in-situ access to the vital context of past decisions that carry elements of crew knowledge.
The central position of crew knowledge in participants’ responses demonstrates that ST can act as a vehicle for knowledge sharing and social learning in “one consolidated platform” (P21-S). Participants repeatedly mentioned two types of insights they gained from ST to be particularly actionable for AI-mediated decisions. The first is additional variables important for the decision-making task that are not captured in the AI’s feature space. For example: “I have a lot more variables that I'm aware of to consider, like, the whole pandemic thing...”(P12-NS). These additional variables are often tacit knowledge, idiosyncratic to the decision, or constantly changing, making them impossible to be formalized in an algorithm. ST could support in-situ access to these variables.
Second, ST supports analogical reasoning with similar decisions and their actual outcomes. Participants exhibited a tendency to reason about the similarity and differences between the contexts of the current decision and past decisions made visible by ST. For example: “what did other oncologists do for a patient like that? So, what treatments were chosen for patients like this person?” (P6-NS) or “I see the reasoning why they didn't pay the recommended price the other time... but those were different circumstances and look now, they're were growing customer and we need to push them up closer to the more profitable price” (P1-S).
Gaining actionable insights could ultimately boost decision confidence, as most participants commented on increasing their confidence in the final price. We also observed an interesting bifurcation on how they conceptualize confidence in the AI versus confidence in oneself after being empowered with knowledge about the decision context. This quote encapsulated that perspective well:
\begin{quote}
The system will go by the numbers but I have my “instincts” thanks to my [crew] knowledge. With these comments, you can say I also have my team’s “instincts” to help me. So I am less confident on the AI but more in myself due to the 360 view I have of things-- I have more information than the machine. (P22-NS)
\end{quote}
Moreover, participants commented that learning from the decision context could also support follow-up actions such as interacting with clients or “justifying” (P1-S) the decision to supervisors, as illustrated in the quote below:
\begin{quote}
And I actually learned a lot. I learned from their comments... I feel like this is an education for the next sale. Even [if it is] another customer, I will be more confident...and know what to do with [the AI's recommendation] because I know how to evaluate it. (P12-NS)
\end{quote}
Learning about past decisions from others, especially higher echelons of the organizational hierarchy, also provided social validation. Social validation can reduce the feelings of individual vulnerability in the decision-making process. While going through the sales scenario, participants would often comment how “the director (Jess) offering discounts gives [them] the permission to do the same” (P12-NS). Being able to have a “direct line of sight into the trajectory of how and why decisions were done in the past” (P24-NS) can make one feel empowered, especially if one has to contest the AI. For most participants, their use of AI systems was mandated by their employers. Many a time, the technology got in the way, becoming a “nuisance” (P1-S) they needed to “fight” (P5-S). Contesting the machine often requires time-consuming reporting and manual review, which creates a feeling that one “can’t just say no to the AI” (P25-NS). This participant elaborated on the vulnerability and how social validation could empower one to act:
\begin{quote}
People are afraid—they don’t want to screw up. You look like a dumb*** if you end up in the war room and say you goofed up because you blindly followed the machine. Even if you have at least one other person doing something similar with the AI, you are safe. Just that knowledge is enough to act less scared. [If] your neck is on the line, someone else’s is also on the line. It distributes the risk. (P26-NS)
\end{quote}
\subsection{Organizational context made visible}
Lastly, ST gives visibility to the broader \textbf{organizational context}, including the meta-knowledge about the organization such as who knows what and organizational practices. Different from decision context, which makes visible knowledge localized to the decision, organizational context reflects macro-information about the organization. This differentiation shares similarity with the concepts of content versus interaction transparency in Stuart et al.'s ST in social network ~\cite{stuart2012social}, which emphasizes that transparency of others' interactions enables awareness of ``normative behaviors as well as the social structure within a community''. We observed that such awareness could then: (1) inform an \textbf{understanding of organizational norms and values} that help \textbf{improve decision-making} and \textbf{calibrate people’s overall job expectations}; (2) foster \textbf{accountability and auditability}, and 3) facilitate \textbf{expertise location}, and if done right, over time the \textbf{formation of a Transactive Memory System (TMS)}~\cite{yoo2001developments}. In short, organizational context made visible by ST could foster effective collective actions in the organization and strengthen the human-AI assemblage.
Visibility of others' actions in an organization (what’s done) could translate into an understanding of organizational norms (what’s acceptable) and values (what’s important), which might be otherwise neglected since "norms are often not enshrined in a rule book" (P25-NS). From the comments in the scenario, participants were informed of organizational norms: “the fact that a director offered the discount below cost price means that this is something that's acceptable. I might be able to do” (P12-NS) and values: “seeing Jess [the director in our scenario] give such a steep discount and noting how she did it to retain a customer, tells [us] that relationship matters to this company” (P28-NS). This type of insight is crucial for making informed decisions and setting overall job expectation, especially for new employees to “learn about the culture of the company” (P16-NS). The following participant succinctly summarized this point:
\begin{quote}
The comments...get me a sense of what should be done, what’s expected of me, and what I can also get away with. It tells me what this company values. This helps me understand why certain things are done the way they are, especially if they go against what the AI wanted me to do. This actually explains why I need to do something. (P25-NS).
\end{quote}
The enactment of ST in an AI system shared across an organization enables accountability. Participants felt that if they knew “who did what and why, [then] it provides a nice way to promote accountable actions” (P26-NS). Participants noted that currently there is a level of opaqueness in workers' decision-making processes, making it difficult to uphold accountability, be it during bank audits, sales audits, or standardization on health interventions. ST, according to them, can provide ``peripheral vision'' (P29-NS) that can boost accountability by not only making past decisions traceable, but also socially-situated to better evaluate and attribute responsibilities for, as highlighted by this quote:
\begin{quote}
I think these comments would be extremely important for audits and postmortems after an attack. The traceability is huge. (P26-NS, a senior SOC analyst)
\end{quote}
That being said, there is a potential double-edged-sword nature to traceability and accountability, where people might feel they are being watched or surveilled. The same participant (P26-NS) articulated this concern:
\begin{quote}
You know, there is a dark side to this. If you are part of organizations that love to surveil people, then you are out of luck. That is why organizational culture is so important...
[In our company], we focus on the problem not the person. But you can’t really say this applies [everywhere].
(P26-NS)
\end{quote}
ST also provides awareness of organizational meta-knowledge~\cite{leonardi2015ambient}, such as who does or knows what, and who knows whom. Many participants reacted to the scenario with reaching out to relevant people made visible through ST: such as “who was driving that sales” (P3-NS), or “reach out to Jeff just because it's the most recent and find out what's going on” (P5-S). It shows that ST could potentially solve a pain point for larger, distributed organizations by supporting expertise location.
Beyond expertise sharing, some participants commented that knowing whom to reach out to could facilitate the creation of an “institutional memory” (P28-NS), the passing of “legacy knowledge” (P2-NS), and the ability to “leverage broader resources to lean on” (P8-S). These comments resonate with the core concept of transactive memory systems (TMS)~\cite{moreland2006transactive,brandon2004transactive}, which explains how a group or organization collectively manages the distribution and retrieval of knowledge across different individuals, often through informal networks rather than formal structures~\cite{nardi2002networkers}. TMS could facilitate employee training and benefit new members:
\begin{quote}
You can’t survive without institutional memory… [but] it’s never written anywhere and is always in the grapevines. Even if some of it could be captured like this [with ST], then that’s a game changer... Training newcomers is hard especially when it comes to getting that “instinct” on the proper way to react to the [security] alerts. Imagine how different training would be if everything was there in one place!” (P28-NS)
\end{quote}
A TMS could also facilitate a peer-to-peer support system that gives employees a sense of community:
\begin{quote}
What I really love is the support system you can potentially create over time using ST. This actually reminds of the knowledge repo[sitory] my colleagues and I have set up where we add our nuggets of client specific wisdom which helps others operate better.
As you know, we are a virtual team so having this collective support is crucial. (P27-NS, a SOC data scientist)
\end{quote}
Through tight interactions of the community and repeatedly seeing others' decision processes, a TMS can, over time, enable to formation of a collective mind~\cite{weick1993collective,yoo2001developments}--members of a group form a shared cognitive or decision schema and construct their own actions accordingly. Collective mind is associated with enhanced organizational performance and creativity. Interestingly, one participat speculated on how ST can be construed as “mindware”:
\begin{quote}
This almost reminds me of a mindware in a team, sort of like a group mind. Currently, we tie our [security] incident reports to a slack channel and that acts as a storage of our collective memories...
We even have tagged comments, so when you showed me your thing, it reminded me of that.
(P25-NS)
\end{quote}
\begin{table*}
\centering
\caption{Summary of the design features, supported effect, and rank of the ``4W'' features}
\label{tab:4W}
\resizebox{\linewidth}{!}{%
\begin{tabular}{>{\hspace{0pt}}p{0.08\linewidth}>{\hspace{0pt}}p{0.3\linewidth}>{\hspace{0pt}}p{0.43\linewidth}>{\hspace{0pt}}p{0.1\linewidth}}
\toprule
Category & Design Features & Supported Effect & Overall Rank \\
\midrule
What & Action taken on AI\par{}Decision outcome\par{}Summary statement & Tracking AI performance\par{}Machine contestability (Social validation) & 1st \\
\\
Why & Comments with rationale justifying the decision & Tracking AI performance\par{}Actionable insights\par{}Understanding organizational norms and values\par{}Social validation\par{} & 2nd \\
Who & Name\par{} Organizational role/ job title \par{}Profile picture & Social validation\par{}Transitive trust (Infusing human elements)\par{}Expertise location & 3rd \\ \\
When & Timing of the decision & Temporal relevance (actionable insights) & 4th \\
\bottomrule
\end{tabular}
}
\Description[Summary of the design features, supported effect, and rank of the ``4W'' features]{This table shows how each of the 4W-- what, why, who, and when-- are conceptualized as design features, what effect they have on the implications of ST, and their overall rank}
\end{table*}
\subsection{Design for ST: the 4W}
With the effects of ST at the three levels in mind, now we discuss how participants reacted to specific design features that are intended to reflect ST. As discussed in Section~\ref{4w}, our co-design exercises informed the choices of constitutive elements of ST: \textit{Who} did \textit{What}, \textit{When}, and \textit{Why}, referred as the 4W. The reader might recall that participants were asked to rank and justify the relative importance of 4W twice during the interview. The first ranking was done in the sales scenario. The second was done in discussing the transferability to participants' own domains. By explicitly inquiring about their thoughts and preferences around the 4W, it helped us understand the effects of each of these design features in facilitating ST and the remaining challenges.
Despite domain-dependent variations, we found overall patterns of preference that are informative. In the sales scenario, first, participants wanted to know “what happened?” (mean rank= 1.90). If the outcome was interesting, then they wanted to delve deeper into the “why” (mean rank= 1.97), followed by “who” (mean rank= 3.03) did it and “when” (mean rank= 3.10). The relative order of the 4W remained stable when discussing transferability to participant's individual domains. Table~\ref{tab:4W} summarizes the 4W design features and the types of effect they support based on the codes emerged in the interviews. These codes correspond to the effects of the three levels of context shown in Table~\ref{tab:conceptual framework}.
\subsubsection{What:} In our scenario, the “what” is conveyed by two design features – whether (a) a previous person accepted or rejected the AI’s recommendation and (b) whether the sale was successful or not [the outcome]. There was also a summary feature of What appeared at the top (block 2 in Figure~\ref{fig:visual scenario}). Citing that the outcome is the “consequence of [their] decision” (P16-NS), participants felt that it was a must-have element of ST. Participants referred to the “what” as the “snapshot” of all ST information (P5-S, P8-S, P15-NS, P28-NS) which gave them an overview of the AI’s performance and others’ actions, and guided them to decide “do I want to invest more time and dig through” (P15-NS). As the scenario unfolded in the beginning, seeing the summary What feature often invoked a reaction that one should be cautious from over-relying on the AI’s recommendation, but seek further information to make an informed decision. On a more constructive note, especially when thinking through transfer scenarios in radiology and cybersecurity, participants highlighted the need to present the appropriate level of details so that it does cognitively burden the user-- “the outcome should be a TL;DR. The ‘why’ is there if I am interested” (P29-NS).
\subsubsection{Why:} The “Why” information was communicated in free-form comments left by previous users in our scenario. Participants often referred to the “why” as the “context behind the action” and “to understand the human elements of decision-making” (P17-NS). They felt that the “why” could not only help them understand areas that the technology might be lacking, but also “explain the human and the organization” (P28-NS). “Understanding the rationale behind past decisions allows [one] to make similar decisions… [and] gives you an idea of what you should be doing” (P18-S). Prior rationales can also “give [humans] a justification to reject the machine” (P7-NS) by "know[ing] why someone did something similar” (P28-NS). In short, insights into the why can inform AI performance, provide actionable insights and social validation for the decision, as well as facilitate a better understanding of organizational norms and values. Social validation, in particular, can enable contestability of AI. On a constructive note, participants highlighted the need to process or organize the comments to make them consumable: “not all whys are created equal, [and that there is a] need to ensure things are standardized” (P25-NS). There were concerns that if comments are not quality controlled, they might not serve the purpose of shedding context appropriately. Citing “no one wants a lawsuit on their hands” (P26-NS), participants also suggested the need to be vigilant about compliance and legal requirements to ensure private details (e.g., proprietary information) is not revealed.
\subsubsection{Who: }The “Who” information in our scenario included multiple elements: a previous user’s name, position and a displayed profile picture. Participants engaged with the implications of “who” at multiple levels. For many, ``who'' was the bare minimum that they needed for expertise location – “if [I] knew who to reach out to… [I] could find out the rest of the story” (P5-S). For others, knowing someone’s organizational role or level of experience is more important, because “hierarchy matters” (P16-NS) and one’s experience level influences the “degree of trust we can place on other people’s judgement” (P2-NS). Thus the identity information could affect both social validation for one’s own decision and transitive (dis) trust in AI. On a constructive note, some reflected on how “the collective `who' matter[ed]” (P6-NS) and there needs to be consistency across personnel for them to make sense of the decision. Some participants raised concerns of the profile picture and the name displayed in our scenario. They felt that these features can lead to biases in weighing different ST information. Others welcomed the profile information because it “humanizes” the use of AI (P1-S). Here, the domain of the participant appeared to matter – most salespeople welcomed complete visibility; many of the stakeholders from healthcare and government service domains raised concerns. Such perception differences across domains highlight that we need to pay attention to the values in the community of practice as we design these features.
\subsubsection{When:} The “When” information is expressed by a timestamp. Participants felt that the timing can dictate “if the information is still relevant” (P11-NS), which informs the actionability of context they gain from ST. Knowing the \textit{when} “puts things into perspective” (P16-NS) because it adds “context to the decision and strengthen[s] the why” (P18-S). Timing was particularly useful when participants deliberated on which prior decision they should give more weight to. One comment in the scenario highlighted how Covid-19 (a global pandemic in 2020) influenced the decision-maker’s actions. At the time of the interviews, the world was still going through Covid-19. The “when” “aligned things with a timeline of events and how they transpired” (P21-S).
\subsection{Transferability of ST to other domains}
\label{transferability}
After participants engaged with the sales scenario, we debriefed them on the conceptual idea of adding ST to AI systems. We then asked them to think of transfer scenarios by envisioning how ST might manifest in their own domains or use cases. As Table~\ref{table:participant details} shows, except for 3 people (P4-NS, P11-NS, P17-NS), our participants came from three main domains: sales, cybersecurity, and healthcare (radiology and oncology). Here we give an overview of how participants viewed the potential needs and impact of ST in cybersecurity and healthcare domains.
\subsubsection{Cybersecurity: } Participants working in cybersecurity domain were unanimous in their frustration about the lack of awareness of how their peers make use of AI's recommendations. They saw a rich space where the incorporation of ST could improve their decision-making abilities and provide social validation to foster decision-making resilience. For example, many participants felt that ST would be extremely useful in ticketing systems, where the SOC analyst is tasked with a binary classification deciding if the threat should be escalated or not. Current AI systems have a high rate of false positive alerting a security threat when there is none. This can be stressful for new analysts, as “newcomers [who] always escalates things because they are afraid” (P22-NS). Others pointed out ST can provide insights into organizational practices in the context of compliance regulations. Participants also highlighted ST’s potential to augment a standardized AI with local contexts in different parts of an organization. This would be particularly useful when the AI was trained on a dataset from the Global North but deployed in the Global South:
\begin{quote}
A lot of the companies operate internationally, right? So one of the things we struggle with is working with international clients whose laws are different. On top of that, the system is trained in North American data. Cyber threats mean different things to different people—what’s harmless to me can breach your system. So yeah, if we can do something like this to augment the AI, I think we can catch threats better in a personalized manner to the client. Also, justifying things would be easier because now you have data from both sides [humans and AI]. (P26-NS)
\end{quote}
Most participants highlighted how visibility of the crew knowledge would be instrumental to pass on “client specific legacy knowledge” (P2-NS). In fact, many cybersecurity teams have existing tools to track past decisions “beyond the model details” (P26-NS); for instance, one team manually keep a historical timeline of false positive alerts. This knowledge helps them calibrate their decisions because “no clients like the boy who shouts wolf every single time” (P25-NS). However, none of these aspects were integrated. Some even expressed surprise on the similarity after going through our scenario. This participant commented on how integration of their tracking system could facilitate ST to improve decision-making:
\begin{quote}
It’s not like we don’t have crew knowledge now, you know. But I never really thought about the whole explainability thing from both sides before you showed me this [pointing to the comments in the sales scenario]. Why just have it from the machine? People are black boxes too, you know. Coming at [explainability] from both ends is kind of holistic. I like it. (P29-NS, a SOC analyst)
\end{quote}
Some participants, mainly data scientists, speculated on how one can use the “corpus of social signals” (P10-NS) to feed back into the machine as training data. They wanted to “incorporate the human elements into the machine” (P14-NS) or expert knowledge of “top” analysts back into the AI. They wished the ingestion to not only improve the AI performance, but also to generate socially-situated “holistic explanations”, a point we come back to later in the Discussions section.
\subsubsection{Healthcare (Radiology and Oncology):} Our participants in the healthcare space mainly work in the imaging decision-support domain. Participants felt that ST has promising transfer potential because it would facilitate peer-review and cross-training opportunities. For radiologists and oncologists, participants highlighted that doctors need “explainability especially when their mental models do not match with the AI[s’ recommendation]” (P16-NS). This is where peer feedback and review for similar AI recommendations can be instrumental. One person shared a story of how oncologists rely on “tumor boards” (a meeting made up of specialized doctors to discuss challenging cases). The goal is to decide on the best possible treatment plan for a patient by collaboratively thinking through similar tough cases. This participant equated the tumor board activity to those in the comments, highlighting how the 4W adds a “personal touch” to situate the information amongst “trustworthy peers” (P6-NS).
Participants also valued the context brought in by ST for multi-stakeholder problems, such as deciding on treatment plans for patients going through therapy. ST can help ensure the plans are personalized because doctors can not only see the AI’s recommendation that’s trained on a standard dataset, but can also “consult or reach out to other doctors [who have] treated similar patients and what were all the surrounding contexts that dictated the treatment plan” (P6-NS). According to them, one of the strengths of ST was that the technical and the socio-organizational layers of decision support were integrated in one place, presented side by side, as highlighted in the following quote:
\begin{quote}
You need both [social and technical aspects] integrated. Without integration in one place, context switching just takes a lot of time and no one would use it. Just having these things in one place makes all the difference. It’s funny how we actually IM each other to ask what people did with the AI’s alerts. (P27-NS)
\end{quote}
\subsection{Challenges around ST}
\label{challenge}
Providing ST in AI systems is not without its challenges, risks, and tensions. We discuss four themes that emerged from the interviews on the potential negative consequences of ST. Future work should strive to mitigate these problems.
First and foremost, there is a vital tension between transparency and \textbf{privacy}. Similar issues have been discussed in prior work on social transparency in CSCW~\cite{erickson2003social}. Participants were concerned about making themselves visible to others in the organization, especially with job-critical information such as past performance and competing intellect. Some were also worried that individuals could be coerced into sharing such sensitive information. For example, P6-NS commented, based on her experience working with health professionals, that people may be unwilling to disclose detailed information about their work:
\begin{quote}
I've definitely got on the phone with colleges where they're like, well, you know, not everybody at my practices [are willing to talk about it]... just everybody having access to the outcome probably is not great. Especially if they're not really in a position...and it just becomes like, a point of contention. (P6-NS)
\end{quote}
Some were concerned about revealing personal information. For example, P2-NS reacted by asking: “do I really want this info about me? Who will see it? What can they do with it?” Several suggested to anonymize the \textit{Who} by revealing only general profiles such as position or present the ST information at an aggregated level.
The second tension is around \textbf{biases} that ST could induce on decision-making. The most prominent concern is on group-thinking, by conforming to the group or the majority’ choices. Other biases could also happen by following eminent individuals such as someone in a “senior position” (P17-NS) or “a friend” (P14-NS). As discussed in previous sections, ST could invoke social-based heuristics, which could support both decision-making and judgment of AI. However, biases and cognitive heuristics are inevitably coupled, and should be carefully managed. Users in some domains might be more subject to biases from ST than others. For example, P17-NS were hesitant about introducing ST features into the human resource domain, for example for AI assisted hiring: “issues of bias, cherry picking, and groupthink are much more consequential in HR situation” (P17-NS).
A third challenge is regarding \textbf{information overload and consumption} of ST. While the design scenario listed only 3 comments, participants were concerned about how to effectively consume the information if the number of entries increases, and how to locate the most relevant information in them. There was also a tension in integrating ST in one’s decision-making workflow, which was especially prominent in time-sensitive contexts such as clinical decision-support. Some participants suggested avoiding a constant flow of ST and only provide ST where needed, e.g. “[ST] is not the information that always needs to be up in front of their face, but there should be a way to get back to it, especially when you're building confidence in kind of assistance”(P6-NS). Others suggested providing ST in a structured or processed format such as “summarization” (P17-NS) or “providing some statistics” (P5-S).
Lastly, a tension for the success of ST lies in the \textbf{incentive to contribute}. While there are clear benefits for consumers of ST, it is questionable whether there is enough motivation for people to take the extra effort to contribute, as illustrated by this quote:
\begin{quote}
One thing that we found interesting is oncology... It's a really hard sell to get them to give feedback into a system because they're so time pressed for their workflow...they're giving you work for free. Like, systems should be doing this for them... but the system can break if they don't participate in that loop. (P6-NS)
\end{quote}
This is a classic problem in CSCW systems~\cite{grudin1988cscw}, which may require both lowering the barriers and cost to contribute, and incentivizing contributions with visible and justifiable benefits.
\section{Discussion \& Implications}
Our results identify the potential effects of ST in AI systems, provide design insights to facilitate ST, and point to potential areas of challenges. In this section, we discuss three high-level implications of introducing ST into AI systems: how ST could enable holistic explainability, how ST could strengthen the Human-AI assemblage, and some technical considerations for realizing ST to move towards a socially-situated XAI paradigm.
\subsection{Holistic Explainability through ST}
After participants concluded the scenario walk-through, we debriefed them on the concept of ST and the idea of facilitating explainability of AI-mediated decisions with ST. Despite frequently using AI systems and facing explainability related issues in their daily workflows, many participants were initially surprised by the association between socio-organizational contexts and AI explainability. The surprise was met with a pleasant realization when they reflected on the scenario and how it could transfer to their real-world use cases. Perhaps their reaction is not surprising given that the epistemic canvas of XAI has largely been circumscribed around the bounds of the algorithm. The focus has primarily been on “the AI in X-AI instead of the X [eXplainable], which is a shame because it’s the human who matters” (P25-NS). The following sentiment captures this point:
\begin{quote}
I was taught to think [that] all that mattered [in XAI] was explanations from the model...This is actually the first time I thought of AI explainability from a social perspective, and I am an expert in this space! This goes to show you how much tunnel-visioned we have been. Once you showed me Social transparency…it was clear that organizational signals can definitely help us make sense of the overall system. It’s like we had blinders or something that stopped us from seeing the larger picture. (P27-NS, a SOC data scientist)
\end{quote}
A most common way participants expressed how ST impacted their “ways of answering the why-question” (P28-NS) was how incorporation of the context makes the explainability more “holistic” (P2-NS, P23-S, P6-NS, P27-NS, P29-NS). Acknowledging that “context is king for explanations… [and] there are many ways to answer why” (P27-NS), participants felt that the ST goes “beyond the AI” to provide “peripheral vision” of the organizational context. This, in turn, allows them to answer their why-questions in a holistic manner. For instance, in SOC situations “there is often no single correct answer. There are multiple correct answers” (P2-NS). Since AI systems “don’t produce multitudes of of explanations”, participants acknowledged that “incorporating the social layer into the mix” can expand the ways they view explainability (P28-NS). Moreover, the humanization of the process can also make the decision explainable to non-primary stakeholders in a way that technical transparency alone cannot achieve. For instance, participants felt that having the 4W can make it easier to justify the decisions to clients and regulators.
While in this work our focus is on how a holistic explainability through ST could better support decision-makers, we recognize that there are other types of stakeholders and explanation consumers, as well as other types of AI systems, that could benefit from ST. For example, collecting the 4W in the deployment context could help model developers to investigate how the system performs and why it fails, then incorporate the insights gained about the technological, decision and organizational contexts to improve the model. Auditors or regulatory bodies could also leverage 4W information to better assess the model's performance, biases, safety, etc. by understanding its situated impact. The contributors of ST information are not limited to decision-makers. For example, with automated AI systems where there isn't a human decision-maker involved, its explainability could be facilitated by making visible the social contexts of people who are impacted by the AI systems.
\subsection{Making the Human-AI assemblage concrete}
We highlight that a consequential AI system is often situated in complex socio-organizational contexts, where many people interacting with it. By bringing the human elements of decision-making to the fore-front, ST enables the humans to be explicitly represented, thereby making the Human-AI assemblage concrete. As one participant put it, the socially-situated context can ensure “the human is not forgotten in the mix of things” (P25-NS). In our Findings section, we discussed how organizational meta-knowledge can facilitate formation of Transactive Memory Systems (TMS)~\cite{hollingshead2003potential, moreland2006transactive, wegner1991transactive}, allowing "who knows what" to be explicitly encoded for future retrieval.
Over time, the "heedful interrelations" enabled by TMS and repeatedly seeing others' decision processes with the AI through ST could possibly enable a shared decision schema in the community, leading to the formation of a collective mind~\cite{weick1993collective, yoo2001developments}. This collective mind is one that includes AI as a critical player. With this conceptualization future work could explore the collective actions and evolution of human-AI assemblages.
By prioritizing the view of human-AI assemblage over the AI, adding ST to AI systems calls for critical consideration on what information of the humans and whose information is made visible to whom. Prior work on ST in CSCW systems has warned against developing technologies that make it easy to share information without careful consideration on its longer-term second-order effect on the organization and its members~\cite{stuart2012social}. In Section~\ref{challenge} we identified four potential issues of ST as foreseen by our participants, including privacy, biases, information overload and motivation to contribute, all of which could have profound impact on the functioning of the human-AI assemblage. In general, future work implementing ST in AI systems should take socioechnical approaches to developing solutions that are sensitive to the values of stakeholders and ``localized'' to a human-AI assemblage. For example, regarding the privacy issue, participants were sensitive to how much visibility the rest of the organization has to their shared activities and knowledge. When asked how they might envision the boundaries, participants highlighted that one should “let the individual teams decide because every ‘tribe’ is different” (P28-NS).
\subsection{Towards socially-situated XAI}
Using a scenario-based design (SBD) method, we suspended the needs to define system operations and technical details. Some practical challenges may arise in how to present the 4W to explanation seekers. The first challenge is to handle the quantity of information, especially to fit into the workflow of the users. In addition to utilizing NLP techniques to make the content more consumable, for example by providing memorization or organizing it into facets, it could also be beneficial to give users filtering options, allow them to define ``similarity'' or choose past examples they want to see. Secondly, there needs to be mechanisms in place to validate the quality and applicability of ST information, since ``not all whys are created equal'' (P25-NS). This could be achieved by either applying quality control on the recorded ST information, or through careful design of interfaces to elicit high-quality 4W information from the contributors. Another caveat is that it is common for an AI system to receive model updates or adapt with usage, so its decisions may change over time. In that case, it is necessary to flag the differences of the AI in showing past ST information. Lastly, in certain domains or organizations, it is not advisable or possible to gather all 4W information, sometimes due to the tension with privacy, biases and motivation to contribute, so alternative solutions should be sought, for example by linking past decision trajectories with relevant guidelines or documentation to help users decode the \textit{why} information when it is not directly available.
Several participants, especially those with a data science background, suggested an interesting area for technical innovation--if “the AI can ingest the social data” (P12-NS) to improve both its performance and its explanations. While recent work has started exploring teaching AI with human rationales~\cite{Ghai2020XAL}, ST could enable acquiring such rationales in real-usage contexts. As suggested by what participants learned from the 4W, the decision and organizational contexts made visible by ST could help the AI to learn additional features and localized rules and constraints, then incorporate them into its future decisions. Moreover, a notable area of XAI work focuses on generating human-consumable and domain-specific machine explanations by learning from how humans explain~\cite{ehsan2019automated,hind2019ted}, which could be a fruitful area to explore when combined with the availability of 4W information. That being said, it may be desirable to explicitly separate the technical component (to show how the AI arrives at its decision) and the socio-organizational component (as further support for or caution against AI's decision) in the explanations, as participants had strong opinions to be able to “know how and where to place the trust” (P27-NS).
\section{Limitations \& Future Work}
We view our work as the beginning of a broader cross-disciplinary discourse around what explainability entails in AI systems. With this paper, we have taken a formative step by exploring the concept of Social Transparency (ST) in AI systems, particularly focusing on how incorporation of socio-organizational context can impact explainability of the human-AI assemblage. Given this first step, the insights from our work should be viewed accordingly. We acknowledge the limitations that come with using a scenario-based design~\cite{rosson2009scenario}, including the dependency between the scenario and data. The insights should be interpreted as formative instead of evaluative. We acknowledge that we need to do more work in the future to expand the design space and consider other design elements for ST, further unpack the transferability of our insights, especially where this transfer might be inappropriate. We should also investigate how ST impacts user trust over longitudinal use of ST-infused XAI systems.
Our conception of ST is rooted in and inspired by Phil Agre's notion of Critical Technical Practice~\cite{agre1997computation,agre1997toward} where we identify the dominant assumptions of XAI and critically question the status quo to generate alternative technology that brings previously-marginalized insights into the center. Agre stated that ``at least for the foreseeable future, [a CTP-inspired concept] will require a split identity -- one foot planted in the craft work of design and the other foot planted in the reflexive work of critique.''~\cite{agre1997toward}. As such, ST will, at least for the foreseeable future be a work-in-progress, one that is continuously pushing the boundaries of design and reflexively working on its own blind spots. We have ``planted one foot" in the work of design by identifying a neglected insight--the lack of incorporation of socio-organizational context as a constitutive design element in XAI-- and exploring the design of ST-infused XAI systems. Now, we seek to learn from and with the broader HCI and XAI communities as we ``plant the other foot'' in the self-reflective realm of critique.
\section{Conclusion}
Situating XAI through the lens of a Critical Technical Practice, this work is our attempt to challenge algorithm-centered approaches and the dominant narrative in the field of XAI. Explainability of AI systems inevitably sits at the intersection of technologies and people, both of which are socially-situated. Therefore, an epistemic blind spot in that neglects the ``socio" half of sociotechnical systems would likely render technological solutions ineffective and potentially harmful. This is particularly problematic as AI technologies enter different socio-organizational contexts for consequential decision-making tasks. Our work is both conceptual and practical. Conceptually, we address the epistemic blind spot by introducing and exploring Social Transparency (ST)--the incorporation of socio-organizational context--to enable holistic explainability of AI-mediated decision-making. Practically, we progressively develop the concept and design space of ST through design and empirical research. Specifically, we developed a scenario-based design that embodies the concept of ST in an AI system with four constitutive elements--\textit{who} did \textit{what} with the AI system, \textit{when}, and \textit{why} they did what they did (4W). Using this scenario-based design, we explored the potential effect of ST and design implications through 29 interviews with AI stakeholders. The results refined our conceptual development of ST by discerning three levels of context made visible by ST and their effects: technological, decision, and organizational. Our work also contributes concrete design insights and point to potential challenges of incorporating socio-organizational context into AI systems, with which practitioners and researchers can further explore the design space of ST. By adding formative insights that catalyzes our journey towards a socially-situated XAI paradigm, this work contributes to the discourse of human-centered XAI by expanding the conceptual and design space of XAI.
\begin{acks}
With our deepest gratitude, we acknowledge the time our participants generously invested in this project. We are grateful to members of the Human-Centered AI Lab at Georgia Tech whose continued input refined the conceptualizations presented here.
We are indebted to Werner Geyer, Michael Hind, Stephanie Houde, David Piorkowski, John Richards, and Yunfeng Zhang from IBM Research AI for their generous feedback and time throughout the duration of this project.
Special thanks to Intekhab Hossain and Samir Passi for conversations and feedback throughout the years that have constructively added to the notion of Social Transparency.
This project was partially supported through an internship at IBM Research AI and by the National Science Foundation under Grant No. 1928586.
\end{acks}
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction}
There is an ever-increasing pool of data, particularly unstructured data such as images, text, video, and audio. The vast majority of this data is unlabelled. The process of labelling is time-consuming, labour-intensive, and expensive. Such an environment makes algorithms that can leverage fully unlabelled data particularly useful and important. Such algorithms fall within the realm of unsupervised learning. A particular subset of unsupervised learning is known as Self-Supervised Learning (SSL). SSL is a paradigm in which the data itself provides a supervision signal to the algorithm.
Somewhat related is another core area of research known as transfer learning \cite{transferlearning}. In the context of computer vision, this means being able to pre-train an encoder network offline on a large, varietal dataset, followed by domain-specific fine-tuning on the bespoke task at hand. The state-of-the-art for many transfer learning applications remains dominated by supervised learning techniques \cite{sp3,sp4,sp1,sp2}, in which models are pre-trained on a large labelled dataset.
However, self-supervised learning techniques have more recently come to the fore as potential alternatives that perform similarly on downstream tasks, while requiring no labelled data. Most self-supervised techniques create a supervision signal from the data itself in one of two ways. The one approach are techniques that define a pre-text task beforehand that a neural network is trained to solve, such as inpainting \cite{inpainting} or a jigsaw puzzle \cite{jigsaw}. In this way, the pre-text task is a kind of proxy that, if solved, should produce reasonable representations for downstream visual tasks such as image or video recognition, object detection, or semantic segmentation. The other approach is a class of techniques known as contrastive methods \cite{simclr,moco,simclrv2}. These methods minimise the distance (or maximise the similarity) between the latent representations of two augmented views of the same input image, while simultaneously maximising the distance between negative pairs. In this way, these methods enforce consistency regularisation \cite{fixmatch}, a well-known approach to semi-supervised learning. These contrastive methods often outperform the pre-text task methods and are the current state-of-the-art in self-supervised learning. However, most of these contrastive methods have several drawbacks, such as requiring prohibitively large batch sizes or memory banks, in order to retrieve the negative pairs of samples \cite{simclr,moco}.
The intuition behind our proposed module is that any system tasked with understanding images can benefit from understanding the geometry of the image and the objects within it. An affine transformation is a geometric transformation that preserves parallelism of lines. It can be composed of any sequence of rotation, translation, shearing, and scaling. A homography is a generalisation of this notion to include perspective warping. A homography need not preserve parallelism of lines, however, it ensures lines remain straight. Mathematically, a homography is shown in Equation \ref{eq:homography}. It has 8 degrees of freedom and is applied to a vector in homogenous coordinates. An affine transformation has the same form, but with the added constraint that $\phi_{3,1} = \phi_{3,2} = 0$.
\begin{equation}
\label{eq:homography}
H_\phi = \left[\begin{array}{ccc}
\phi_{1,1} & \phi_{1,2} & \phi_{1,3}\\
\phi_{2,1} & \phi_{2,2} & \phi_{2,3}\\
\phi_{3,1} & \phi_{3,2} & 1\\
\end{array}\right]
\end{equation}
The ability to know how a source image was transformed to get to a target image implicitly means that you have learned something about the geometry of that image. An affine transformation or, more generally, a homography is a natural way to encode this idea. Forcing the network to estimate the parameters of a random homography applied to the source images thereby forces it to learn semantics about the geometry. This geometric information can supplement the signal provided by a contrastive loss, or loss in the latent space.
In this paper, we propose an additional module that can be used in tandem with contrastive self-supervised learning techniques to augment the contrastive objective (the additional module is highlighted in Figure \ref{fig:architecture}). The module is simple, model-agnostic, and can be used to supplement a contrastive algorithm to improve performance and supplement the information learned by the network to converge faster. The module is essentially an additional stream of the network with the objective of regressing the parameters of an affine transformation or homography. In this way, there is a multi-task objective that the network must solve: 1. minimising the original contrastive objective, and 2. learning the parameters of a homography applied to one of the input images from a \emph{vector difference} of their latent representations. We force the latent space to encode the geometric transformation information by learning to regress the parameters of the transformation in an MLP that takes the vector difference of two latent representations of an input, $x$, and its transformed analogue, $x'$. By including the information in this way, the network is \emph{not} invariant to the components of the transformation but is still able to use them as a self-supervised signal for learning. Moreover, this approach serves as a novel hybrid of the pre-text tasks and contrastive learning by enforcing consistency regularisation \cite{fixmatch}.
\begin{figure}[h]
\centering
\includegraphics[scale=0.5]{arch.png}
\caption{Proposed architecture. The highlighted box highlights the additional proposed module tasked with regressing the parameters of an affine transformation or homography.}
\label{fig:architecture}
\end{figure}
Through extensive empirical studies, we show that the additional objective of regressing the transformation parameters serves as a useful supplementary task for self-supervised contrastive learning, and improves performance for all considered datasets in terms of linear evaluation accuracy and convergence speed.
The remainder of the paper is structured as follows. In Section \ref{sec:rw}, we cover the related work in the area of self-supervised learning, going into detail where necessary. In Section \ref{sec:pm} we detail our proposed method. We first introduce a framework and set of notation to make the formalisation of the approach clear. We then delve into the details behind the architecture and choices for the various part of the system. This is followed by a comprehensive set of experiments in Section \ref{sec:exp}, including results of various datasets, as well as an ablative study. Finally, the paper is concluded with some closing remarks in Section \ref{sec:conc}.
\section{Related Work}
\label{sec:rw}
SSL is a popular research area within computer vision. Previous approaches can be broadly classed into two main categories. The first is where pre-text tasks are manually defined, and the goal of the algorithms is to solve these hand-crafted tasks \cite{ssla,patchpred,rot1,colourising1,pretext}. Examples of such methods include inpainting \cite{inpainting}, colourising \cite{colourising1}, jigsaw puzzles \cite{jigsaw}, patch prediction \cite{patchpred}, and geometric image transformations \cite{geotrans} such as using rotation as the pre-text task \cite{rot1,rot2}. Some of these pre-text approaches that deal with geometric image transformations are similar in spirit to our method. \cite{rot1,rot2} are two variants of predicting image rotations as an auxiliary task for unsupervised learning. Perhaps closer to our method is \cite{geotrans}, in which a set of transformations is applied to image patches, and the network is trained in a fully-unsupervised manner to predict surrogate classes defined by a set of transformed image patches by minimising the log loss. Our method, however, investigates a different, particular set of transformations (those that define an affine transformation of general homography), and show this can be used to aid self-supervised performance, using the transformation parameters themselves as targets that need to be regressed (using mean-squared error) by the contrastive algorithm in a multi-task manner. The discrepancy in the network's ability to predict the actual values of the parameters of the affine transformation/homography serves as our additional supervision signal.
A somewhat related approach to our proposed method within the pre-text task domain is proposed by \cite{ssla}. They propose to augment the learning process of a supervised learning algorithm with additional labels constructed using self-supervised labels. These labels are rotation classes and colour permutations. Importantly, they create a loss function which is based on a joint distribution of the original (supervised) labels and the self-supervised (augmented) labels. In this way, the network is not forced to be invariant to the transformations under consideration, since this has been shown to hurt performance \cite{ssla}. Our method is different to this in that we propose a module to be integrated specifically with self-supervised algorithms. Additionally, we regress the transformation parameters in real vector space and do not create classes for the parameters.
The other broad category of SSL is based on contrastive learning \cite{simclr,moco,vfcca}, and this class of techniques represent the current state-of-the-art in self-supervised learning, outperforming the hand-crafted pre-text task methods. These approaches learn representations by contrasting positive pairs of samples from negative pairs of samples in latent space. Such methods typically require that careful attention be paid to the negative samples. Additionally, they have the disadvantage of requiring prohibitively large batch sizes (4096-16000), memory banks, or other mechanisms to retrieve the relevant negative samples.
One popular such method is known as SimCLR \cite{simclr}. SimCLR is a general framework for contrastive learning, and in its vanilla formulation consists of an encoder network parameterised by a CNN (usually a variant of ResNet \cite{resnet}) and an MLP projection head. An input image is sampled, and two distinct views of that same input image are computed using a random augmentation. The augmentation consists of colour jiterring, Gaussian blurring, and random cropping. The two views are sent through the encoder network to produce two latent representations. These latent vectors are then sent through the projection head to produce final latent vectors. It is from these vectors that the loss is computed. In the case of SimCLR, the loss is normalised temperatured cross-entropy (NT-Xent).
A recent approach proposed in \cite{byol} (BYOL) somewhat overcomes the aforementioned disadvantages of requiring negative pairs of samples (which implicitly requires a large batch size). Two separate networks with their own weights are used in tandem to learn the representation. An \emph{online} network (consisting of an encoder, MLP projection head, and MLP prediction network) is trained to predict the representation outputted by a \emph{target} network. During training, the online network parameters are updated using backpropagation of error derivatives computed using a mean-squared error loss. However, the target network parameters are updated using an exponential moving average. In this way, BYOL overcomes collapsed solutions in which every image produces the same representation. We test our module with both SimCLR and BYOL, since these two methods serve as two popular, recent approaches to contrastive SSL.
Some helpful findings for guiding self-supervised research were demonstrated in \cite{sslrevisit}. Core among these are that 1) standard architecture designs that work well in the fully-supervised setting do not necessarily work well in the self-supervised setting, 2) in the self-supervised setting larger CNNs often means higher quality learned representations, and 3) the linear evaluation paradigm for assessing performance may take a long time to converge. Moreover, \cite{usefulviztasks} find that the effectiveness of self-supervised pretraining decreases as the amount of labelled data increases, and that performance on one particular downstream task is not necessarily indicative of performance on other downstream tasks.
\section{Proposed Method}
\label{sec:pm}
We first introduce a mathematical framework for discussing our method. Let $\mathcal{B}_1$ be a set of \emph{base transformations}. A base transformation is a transformation that cannot be decomposed into more basic transformations and is interpreted as per \cite{byol,simclr}. Examples of base transformations include colour jittering, cropping, and horizontal flipping. We define the possible base transformations a-priori, and $|\mathcal{B}_1| < \infty$. Next, we define a new set of base spatial transformations $\mathcal{B}_2$ that correspond to the general affine transformations (i.e. rotation, translation, scaling and shearing) or the full homography (i.e. affine transformations and perspective projection). Further, we impose the following condition:
\begin{equation}
\label{eq:restr}
\mathcal{B}_1 \cap \mathcal{B}_2 = \emptyset
\end{equation}
The reason for this restriction will be apparent later.
A transformation $t_{b, \theta}$ is parameterised by its associated base transformation $b \in \mathcal{B}_1 \cup \mathcal{B}_2$ and transformation parameters $\theta \in \Theta$. Then, the set of all possible transformations for a particular base transformation set $\mathcal{B}$ may be defined as:
\begin{equation}
\mathcal{T}_{i} := \{t_{b, \theta} | b \in \mathcal{B}_i, \theta \in \Theta \}
\end{equation}
Clearly, we may have that $|\mathcal{T}_{i}| = \infty$, since some parameters may take on any value within compact subsets of $\mathbb{R}$. This is important because we want to be able to sample from an infinite sample space during training to ensure the network sees a variety of samples.
We can now define an \emph{augmentation}, which is an \emph{ordered} sequence of $n$ transformations. As such, each unique ordering will necessarily produce a unique augmentation (e.g. flipping and then cropping is different from cropping and then flipping). Formally, an augmentation $a$ is defined as:
\begin{equation}
a(x) = t_{b_n, \theta_n} \circ \cdots \circ t_{b_2, \theta_2} \circ t_{b_1, \theta_1} (x)
\end{equation}
Denote the set of all possible augmentations for a transformation set $\mathcal{T}_{i}$ as $\mathcal{A}_{\mathcal{T}_{i}}$. Under this definition, $\mathcal{A}_{\mathcal{T}_{2}}$ is the set of all possible affine or homographic transformations. Examples of the affine transformations and homographies can be seen in Appendix \ref{apx:dad}.
Now, consider an input image $x$ sampled at random from a dataset of images $X \subset \mathcal{X}$, where $\mathcal{X}$ is the sample space of images. We sample augmentations $a, b \in \mathcal{A}_{\mathcal{T}_{1}}$, and apply them to $x$ to produce augmented views $x_1$ and $x_2$, respectively. We then sample an affine/homographic transformation $c_{\phi} \in \mathcal{A}_{\mathcal{T}_{2}}$ and apply it to $x_1$ to produce $x_1'$. Note that $x_1$ and $x_1'$ are related by a homography. This is a core assumption relied upon by further inductive biases we introduce into our model.
We now describe the proposed architecture as depicted in Figure \ref{fig:architecture}. Let the mapping $f : \mathcal{X} \rightarrow \mathbb{R}^p$ be parameterised by a CNN, and the mappings $g : \mathbb{R}^p \rightarrow \mathbb{R}^k$ and $h : \mathbb{R}^p \rightarrow \mathbb{R}^m$ be parameterised by MLPs, where $p$, $k$, and $m$ are the dimensionality of the encoder latent vector, projection head latent vector, and homography parameter vector, respectively. $f$ and $g$ are the encoder and projection head from the original SimCLR \cite{simclr} and BYOL \cite{byol} formulations, respectively, whereas $h$ is a new MLP tasked with estimating the homography parameters. Note that if we are regressing all parameters of a general affine transformation, then $m = 6$, whereas for a full homography we have $m=8$. For brevity, we have denoted both streams in the architecture to be a network with the same shared weights, although it may be the case that the two streams consist of networks with different weights (as is the case with BYOL).
The loss function for our method contains two terms. First is the original loss function as defined by the original method: NT-Xent for SimCLR and squared $L_2$ for BYOL. We define this first term as $\mathcal{L}_1(z_1, z_2)$, where $z_1 = g(f(x_1))$ and $z_2 = g(f(x_2))$. The second term can be seen as forcing the network to explicitly learn the affine transformation or homography between $x_1$ and $x_1'$. Let the latent representations of $x_1$ and $x_1'$ be $l_1 = f(x_1)$ and $l_1' = f(x_1')$. We send the vector difference $l_1 - l_1'$ through $h$ to produce an estimate of the homography's parameters. We regress to these parameters using mean-squared error: $\mathcal{L}_2(h(l_1 - l_1'), \phi)$, where $\phi$ are the ground truth affine transformation parameters. Thus, the complete loss function is given by:
\begin{equation}
\label{eqn:loss-function}
\mathcal{L}_1(z_1, z_2) + \mathcal{L}_2(h(l_1 - l_1'), \phi)
\end{equation}
The vector difference naturally describes the transformation needed to move from $l_1$ to $l_1'$. With our architecture and learning objective, we force this vector difference transformation vector to encode the homography between $x_1$ and $x_1'$. This interpretation may be seen as natural and intuitive. Hence, the $\mathcal{L}_1$ term enforces invariance to the transformations in $\mathcal{B}_1$ and $\mathcal{L}_2$ enforces non-invariance to the transformations in $\mathcal{B}_2$.
Note that this is still completely self-supervised. Moreover, the restriction imposed in Equation \ref{eq:restr} is necessary because we cannot have any transformations in $c_{\phi}$'s sequence that would destroy the fact that $x_1$ and $x_1'$ are related through a homography. For example, adding a \emph{cropping} transformation would break the homography assumption. One may add transformations that do not break this restriction (e.g. colour jitter), however, we do not explore this here.
We may interpret this extended architecture as solving a multi-task learning objective in which 1) the contrastive loss between differing augmented views of the image must be minimised, and 2) another network must be able to estimate the homography between images, which explicitly forces the latent space to encode this spatial information during training.
\section{Experiments}
\label{sec:exp}
\subsection{Experimental Setup}
This section presents an empirical study comparing the original SimCLR and BYOL techniques on the CIFAR10, CIFAR100 and SVHN benchmark datasets, with and without our proposed module. Our goal is \emph{not} to achieve near state-of-the-art performance on the datasets, but is rather to demonstrate the effectiveness of the proposed additional homography estimation objective under consistent experimental settings. In all cases, the proposed module improves the performance of a linear classifier on the learned representation and improves the learning speed.
The experimental setup for the self-supervised training of SimCLR and BYOL can be found in Table \ref{tbl:expsetup}. The batch size is somewhat lower than the original methods since the original methods focused on performance on ImageNet, which requires a considerably larger batch size to perform well. In some additional experiments, we find performance decreased for our datasets with batch sizes larger than 256 for all methods (original SimCLR and BYOL, as well as our method). Further, we found alternative optimised hyperparameter values (learning rate, optimiser, and weight decay) that worked better than those proposed in the original formulations of SimCLR and BYOL, which can be attributed to similar reasons as the batch size arguments. We use the same type of learning rate decay as the previous methods, and train for the same number of epochs (and warmup epochs) as SimCLR. We use a temperature of 0.5 for the NT-Xent loss and keep all images at their default resolution of $32 \times 32$. Lastly, all reported confidence intervals are the average across 10 trials of the full pipeline trained from scratch (SSL pretraining + linear evaluation).
\begin{table}[!h]
\centering
\caption{Experimental setup.}
\begin{minipage}[t]{0.55\linewidth}
\begin{tabular}[t]{|l|l|l|}
\hline
\multicolumn{3}{|c|}{SSL} \\ \hline
& SimCLR & BYOL \\ \hline
Batch size & 256 & 256 \\ \hline
Optimiser & Adam & SGD \\ \hline
LR & 3e-04 & 0.03 \\ \hline
Momentum & - & 0.9 \\ \hline
Weight decay & 10e-06 & 4e-04 \\ \hline
Epochs (warmup) & 100 (10) & 100 (10) \\ \hline
LR schedule & Cosine decay & Cosine decay \\ \hline
\end{tabular}
\end{minipage}\hfill
\begin{minipage}[t]{0.36\linewidth}
\begin{tabular}[t]{|l|l|l|}
\hline
\multicolumn{3}{|c|}{Linear Evaluation} \\ \hline
& SimCLR & BYOL \\ \hline
Batch size & 64 & 64 \\ \hline
Optimiser & Adam & Adam \\ \hline
LR & 3e-04 & 3e-04 \\ \hline
Epochs & 200 & 200 \\ \hline \hline
\multicolumn{3}{|c|}{Hardware}\\ \hline
\multicolumn{3}{|l|}{V100 16GB GPU}\\ \hline
\end{tabular}
\end{minipage}
\label{tbl:expsetup}
\end{table}
Performance is measured as per the literature, using linear evaluation on the relevant dataset. The experimental setup for linear evaluation can be seen in Table \ref{tbl:expsetup}. We freeze the encoder and only optimise the weights of a final linear layer using cross-entropy.
We parameterise $f$ as ResNet50, while $g$ and $h$ are parameterised as two-layer ReLU MLPs (Figure \ref{fig:architecture}). Further, to ensure consistency with SimCLR, we have that $\mathcal{B}_1 = \{\text{random crop}, \text{random horizontal flip}, \text{colour jitter}, \text{Gaussian blur}, \text{random grayscale}\}$. The output of $h$ is a six-dimensional real vector, where the six components are defined according to the parameters of a general affine transform: 1) rotation angle, 2) vertical translation, 3) horizontal translation, 4) scaling factor, 5) vertical shear angle, and 6) horizontal shear angle. For a homography, the output of $h$ is instead an eight-dimensional vector. For details about the transformations, see Appendix \ref{apx:dad}.
\subsection{Affine and Homography Objective}
From Tables \ref{tbl:simclranalysis} and \ref{tbl:byolanalysis} (`+ H' and `+ A' for homography and affine, respectively) we can see that the estimation of the affine transformation and the homography both assist performance and allow for faster learning. In particular, we note statistically significant improvements across all datasets for both SimCLR and BYOL with the affine objective.\footnote{Using a t-test and significance level of 1\%} We posit that the ability to explicitly estimate the affine transformation or homography between input images in this way allows the encoder to learn complementary information early on in training that is not available from the contrastive supervision signal. The ability to estimate the affine transform or homography means that the network is encoding the geometry of the input images. This explicit geometric information is not directly available from the contrastive signal. Interestingly, the affine objective outperforms the full homography in all cases, even though an affine transformation is a subset of a homography. We perform a sweep of the distortion amount for the homography and find it consistently performs similar to or a little worse than the affine transform (see Appendix \ref{apx:avsh}). When the distortion factor becomes too large, accuracy drops noticeably as the images are too distorted to learn effectively. We note that incorporating our module into a network results in an average 30\% additional training time versus the respective original methods.
\begin{table}[!h]
\centering
\caption{Performance with SimCLR on various datasets (mean $\pm$ 99\% confidence interval).}
\begin{tabular}{|l|l|l|l|}
\hline
& CIFAR10 & CIFAR100 & SVHN \\ \hline
SimCLR & 63.34 $\pm$ 0.0016 & 28.53 $\pm$ 0.0017 & 83.19 $\pm$ 0.0017 \\ \hline
SimCLR + H & 64.04 $\pm$ 0.0029 & 29.10 $\pm$ 0.0025 & 82.37 $\pm$ 0.0024 \\ \hline
SimCLR + A & \textbf{64.71 $\pm$ 0.0023} & \textbf{31.33 $\pm$ 0.0024} & \textbf{83.85 $\pm$ 0.0017} \\ \hline
\end{tabular}
\label{tbl:simclranalysis}
\end{table}
\begin{table}[!h]
\centering
\caption{Performance with BYOL on various datasets (mean $\pm$ 99\% confidence interval).}
\begin{tabular}{|l|l|l|l|}
\hline
& CIFAR10 & CIFAR100 & SVHN \\ \hline
BYOL & 56.56 $\pm$ 0.0026 & 23.25 $\pm$ 0.0026 & 72.34 $\pm$ 0.0047 \\ \hline
BYOL + H & 58.78 $\pm$ 0.0033 & 25.88 $\pm$ 0.0029 & 76.56 $\pm$ 0.0054 \\ \hline
BYOL + A & \textbf{60.19 $\pm$ 0.0016} & \textbf{28.10 $\pm$ 0.0018} & \textbf{78.71 $\pm$ 0.0050} \\ \hline
\end{tabular}
\label{tbl:byolanalysis}
\end{table}
Figure \ref{fig:results-graphs} shows the linear evaluation accuracy trained on the embeddings extracted from the model at each epoch during the SSL training.
We can see that performance and convergence improves with the inclusion of the proposed module. The module and its accompanying additional objective of regressing the affine transform/homography may be seen as a regulariser for the original contrastive objective. This is further evidenced by the shaded regions in the figures, in which the proposed method results in more stable performance.
\begin{figure}
\begin{subfigure}{.5\linewidth}
\centering
\includegraphics[width=0.91\linewidth,clip=true,trim=0.5cm 0.3cm 1.0cm 1.5cm]{simclr.pdf}
\caption{SimCLR results (measured every 10 epochs).}
\label{fig:simclr}
\end{subfigure}%
\begin{subfigure}{.5\linewidth}
\centering
\includegraphics[width=0.91\linewidth,clip=true,trim=0.5cm 0.3cm 1.0cm 1.5cm]{byol.pdf}
\caption{BYOL results (measured every 10 epochs).}
\label{fig:byol}
\end{subfigure}
\caption{Graph of the linear evaluation accuracy at various points during training for both SimCLR and BYOL with and without our proposed module. The shaded region indicates one standard deviation from the mean.}
\label{fig:results-graphs}
\end{figure}
We note that the relative benefit of our proposed module diminishes with longer training time for SimCLR and BYOL. This makes sense as the relative benefit of the module decreases with time as the model learns to estimate the affine transformation or homography more accurately as the epochs progress. We performed additional experiments on SimCLR and BYOL training the model for longer, and note that the proposed module still outperforms or performs similarly to the original methods on all datasets. This is shown in Table \ref{tbl:longtraining}. These results also verify the findings of previous works that find that larger models trained for longer benefits self-supervised architectures \cite{simclr,byol,simclrv2,sslrevisit}.
\begin{table}[!h]
\centering
\caption{Performance comparison of SimCLR and BYOL on various datasets trained for 500 epochs.}
\begin{tabular}{|l|l|l|l|}
\hline
& CIFAR10 & CIFAR100 & SVHN \\ \hline
SimCLR & 76.96 $\pm$ 0.0031 & 42.17 $\pm$ 0.0001 & 88.37 $\pm$ 0.0030 \\ \hline
SimCLR + A & 76.80 $\pm$ 0.0020 & \textbf{42.53 $\pm$ 0.0014} & 88.83 $\pm$ 0.0015 \\ \hline
BYOL & \textbf{78.00 $\pm$ 0.0060} & 38.05 $\pm$ 0.0007 & \textbf{90.78 $\pm$ 0.0067} \\ \hline
BYOL + A & 77.45 $\pm$ 0.0020 & 40.91 $\pm$ 0.0005 & 90.74 $\pm$ 0.0025 \\ \hline
\end{tabular}
\label{tbl:longtraining}
\end{table}
We note that the performance gap between SimCLR and BYOL evident in Tables \ref{tbl:simclranalysis} and \ref{tbl:byolanalysis} in general can be attributed to the fact that in the original works, BYOL was trained for 10x as long as SimCLR, whereas we trained both for the same number of epochs as the original SimCLR work. We posit that BYOL has simply not converged sufficiently, since BYOL eventually outperforms SimCLR (as evidenced by Table \ref{tbl:longtraining}). This is consistent with the findings from the original works.
\subsection{Invariance Is Not Always Desirable}
In order for a function $f$ to be invariant to a transformation $T$, we must have that, for all $x$, $f(x) = f(T x)$. Thus, one way to encourage invariance to $T$ in a neural network $f$ is to add a term to the loss function which minimises:
\begin{equation}
\label{eqn:invarexp}
L(f(x), f(T x))
\end{equation}
for some measure of similarity $L$. If we rewrite our loss function from Equation \ref{eqn:loss-function} in terms of our input image $x$ and augmentations $a, b, c_{\phi}$, we get:
\begin{equation}
\label{eqn:loss-function-x}
\mathcal{L}_1(g(f(a x)), g(f(b x))) + \mathcal{L}_2(h(f(a x) - f(c_{\phi} a x)), \phi)
\end{equation}
The first term of the above loss, corresponding to the SimCLR/BYOL loss, is clearly of the form of Expression \ref{eqn:invarexp}. This means that we are encouraging our representation to be invariant to the transformations within $\mathcal{B}_1$. However, the second term in the loss (i.e. the term corresponding to the affine transformation/homography parameter estimation) is not of the form of Expression \ref{eqn:invarexp}, since we have recast the objective into a parameter prediction task. Thus, we are not encouraging invariance to the transformations within $\mathcal{B}_2$. We provide some empirical evidence for this in Table \ref{tbl:invariance}. When we recast the module to minimise $L_2(f(x_1), f(x_1'))$ ($L_2$ being the mean squared error loss), performance decreases notably on all datasets, with an average relative decrease of over 8\%. This is because, with this loss, we have enforced invariance to the transformations in $\mathcal{B}_2$. In particular, we have encouraged invariance to all the elements of an affine transformation, which proves problematic.
\begin{table}[!h]
\centering
\caption{Performance for SimCLR for transformation invariant and non-transformation invariant representations using the affine objective.}
\begin{tabular}{|l|r|r|r|r|}
\hline
& \multicolumn{1}{l|}{CIFAR10} & \multicolumn{1}{l|}{CIFAR100} & \multicolumn{1}{l|}{SVHN} &
\multicolumn{1}{l|}{`6' vs `9'} \\ \hline
Invariant & 61.43 $\pm$ 0.0053 & 26.13 $\pm$ 0.0074 & 80.32 $\pm$ 0.0047 & 68.64 $\pm$ 0.0085 \\ \hline
Not Invariant & \textbf{64.71 $\pm$ 0.0023} & \textbf{31.33 $\pm$ 0.0024} & \textbf{83.85 $\pm$ 0.0017} & \textbf{72.35 $\pm$ 0.0058} \\ \hline
\end{tabular}
\label{tbl:invariance}
\end{table}
To delve deeper into the effect of transformation invariance on performance, we extract only the `6' and `9' classes of the SVHN dataset as a new dataset and repeat the SSL pre-training and linear evaluation tasks. The goal of this experiment is to observe how performance degrades when the neural network is encouraged to be invariant to certain transformations - including rotation - in a setting where certain invariance (i.e. rotation) is not desirable. Results can also be seen in Table \ref{tbl:invariance}. This further suggests that invariance to certain transformations is not always desirable. Evidence from Table \ref{tbl:invariance} suggests that transformation invariance (for this particular class of transformations) in SSL may not always be desirable, and may, in fact, hurt performance, even when this may not be expected (as in the case with CIFAR10 and CIFAR100, since no classes of these seem as if they should be affected by transformation invariance like the `6' vs `9' case). For more details on the invariance analyses, see Appendix \ref{apx:ia}.
\subsection{Transform Component Analysis}
Table \ref{tbl:tca} shows the performance of the various components of an affine transformation in terms of linear evaluation accuracy on the dataset.\footnote{Due to apparent instability in training BYOL with our module for these low-dimensional outputs (e.g. single real value output for rotation and scale), we temporarily replace MSE with logcosh, which stabilises training in this setting} To compute these results, the output dimensionality of mapping $h$ needs to be changed accordingly. Namely, rotation, translation, scale, and shear have corresponding output dimensionalities $m$ of 1, 2, 1, and 2, respectively. Interestingly, shear alone outperforms the three other transforms on all datasets for both SimCLR and BYOL. We hypothesise that this is because shear corrupts the image the most out of the four transforms, but still in a recognisable way. This forces the networks to learn more complex geometry and information about the object that the other transforms. We leave further investigation of this to future work.
\begin{table}[!h]
\centering
\caption{Performance comparison of the components of an affine transformation for SimCLR and BYOL. Best-performing transformation highlighted in bold for each dataset.}
\resizebox{\linewidth}{!}{
\begin{tabular}{|l|l|l|l|l|l|l|l|}
\hline
& \multicolumn{3}{c|}{SimCLR} & \multicolumn{3}{c|}{BYOL} \\ \hline
& CIFAR10 & CIFAR100 & SVHN & CIFAR10 & CIFAR100 & SVHN \\ \hline
Rotation & 63.57 $\pm$ 0.0031 & 29.40 $\pm$ 0.0035 & 83.46 $\pm$ 0.0036 & 58.66 $\pm$ 0.0020 & 26.46 $\pm$ 0.0042 & 76.34 $\pm$ 0.0073 \\ \hline
Translation & 64.36 $\pm$ 0.0036 & 29.83 $\pm$ 0.0008 & 82.60 $\pm$ 0.0053 & 58.45 $\pm$ 0.0038 & 25.74 $\pm$ 0.0022 & 77.50 $\pm$ 0.0059 \\ \hline
Scale & 64.05 $\pm$ 0.0035 & 30.34 $\pm$ 0.0027 & 82.80 $\pm$ 0.0037 & 59.12 $\pm$ 0.0034 & 25.47 $\pm$ 0.0023 & 76.53 $\pm$ 0.0102 \\ \hline
Shear & \textbf{64.56 $\pm$ 0.0051} & \textbf{31.05 $\pm$ 0.0033} & \textbf{84.25 $\pm$ 0.0020} & \textbf{60.21 $\pm$ 0.0048} & \textbf{26.94 $\pm$ 0.0038} & \textbf{77.69 $\pm$ 0.0059} \\ \hline
\end{tabular}
}
\label{tbl:tca}
\end{table}
\subsection{Additional Ablations}
We perform various additional experiments to motivate the choice of architecture. We experiment with other means of encoding the latent transformation, specifically, concatenation instead of vector difference. However, this results in marginal performance gains of an average of 0.28 percentage points across the 3 datasets. These results do not seem to justify the noticeable additional computational cost from the transformation representation being twice the size. For primarily this reason, we opt to stick with vector difference. Further, we experiment with having the module operate on the output of $g$ instead of $f$. Performance degrades for all datasets (SimCLR): 64.38 for CIFAR10, 29.99 for CIFAR100, and 82.49 for SVHN. Lastly, we perform some preliminary experiments into having two modules: one operating on $x_1$ and the other operating on $x_2$ (instead of just one module as per our original experimental setup). The resulting performance difference is negligible with this setup: CIFAR10 65.28 $\pm$ 0.61, CIFAR100 31.68 $\pm$ 1.04, and SVHN 84.10 $\pm$ 0.23. We posit that this is because if the one module can solve the homography estimation for $x_1$, then a module operating on $x_2$ will have to be able to solve the homography estimation for it, since the same types of random homographies/affine transformations are being applied to both streams.
\section{Conclusion}
\label{sec:conc}
Network size and time of training is a bottleneck in modern SS architectures that can compete on a performance level like supervised alternatives. We have shown that the proposed module that regresses the parameters of an affine transformation or homography as an additional objective assists this training bottleneck with faster convergence and better performance. The architecture of the module does not encourage invariance to the affine or homographic transformation, as invariance has been previously shown to be potentially harmful \cite{ssla}. Rather, the proposed module encourages these transformations to be encoded within the latent space itself by directly estimating the parameters of the transformation. Lastly, we note that the affine transformation performs better in all cases than the full homography, even though the homography is a superset of affine transformations. The experiments suggest that the additional ability of perspective transform in a homography does not yield any tangible benefit over a regular affine transformation in such low-resolution settings.
\bibliographystyle{unsrt}
|
\section{ Introduction}
\label{sec:intro}
Deep learning has experienced tremendous growth in recent years and is considered today of major relevance for computer vision tasks such as object detection \cite{zhao2019object} or image classification \cite{lu2007survey}. The achieved top performance is due to the sophistication of the training algorithms and of the trained models, the growing computing capability of the machines, and the availability of large datasets.
As depicted in Figure~\ref{fig:workflow}, a deep learning model can be interpreted as the concatenation of: a) a \emph{feature extractor}, that generates high-level features from a processed input, and b) a \emph{classifier}, that maps these features to a particular class. Both blocks are generally trained jointly, such that this vision of a deep learning model in two blocks is only useful for interpretation purposes.
Similarly to other classifiers, deep learning methods are susceptible to noise in the training data which can lead to reduced generalization abilities and lack of robustness. Trying to reduce this noise in the raw input domain of data is complicated and could result in overall poorer performance if performed too aggressively.
On the contrary, the latent representation of input samples, provided by the feature extractor, is a better representation for denoising, as it can be viewed as a decomposition of the raw data into more abstract features. For a fixed class, an efficient filtering procedure would therefore nullify individual specificities of labeled samples in that latent space, to improve the representation of the class more generally.
From a signal processing perspective, a sample's own specificities can be seen as high frequencies in an underlying space modeling its class. The framework of Graph Signal Processing (GSP)~\cite{shuman2013emerging} provides the exact tools we need to model such a domain through a graph representation, where vertices correspond to labeled samples of a given class, and features are seen as graph signals. In this work, we aim at constructing one graph for each class, using these graphs to filter not the raw input images of the considered class but instead high-level features typically obtained at a deep layer of a trained deep neural network, corresponding to the output of the previously introduced \emph{feature extractor}
We conduct this analysis to improve the accuracy of the considered classifier. We expect highest gains in the case of few-shot, where only few labeled samples are available for each class. It is worth to mention that in contrast with standard classification, in the few-shot setting the feature vectors that are extracted by a deep neural network are trained on a large dataset (base classes, $\mathcal{D}_{\mathrm{base}}$). These features are then specialized to a smaller and distinct dataset (novel classes, $\mathcal{D}_{\mathrm{novel}}$), generally through \emph{transfer-learning} methods~\cite{torrey2010transfer,rohrbach2013transfer}.
The outline of the paper is as follows. In Section~\ref{sec:methodology} we introduce the proposed methodology. In Section~\ref{sec:theory} we show on simplified settings that the filtering operation asymptotically reduces intra-class variance and preserves the mean.
In Section~\ref{sec:numerical-result} we perform experiments on standardized benchmarks, and show that the proposed methodology can improve the classification accuracy even in the case of state-of-the-art solutions. We consider both few-shot and classical settings. Finally, Section~\ref{sec:conclusion} is a conclusion.
Throughout this article, we use the following notations. Real values and functions are given in italic (e.g., $m$, $h$). Vectors are denoted in bold lowercase (e.g., $\mathbf{x}$), with $i$-th entry as $\mathbf{x}_i$, letters in bold uppercase (e.g., $\mathbf{W}$) denote matrices, with $i$-th row as $\mathbf{W}_{i,:}$, $j$-th column as $\mathbf{W}_{:,j}$, submatrix of $k$ first columns as $\mathbf{W}_{:,:k}$, and entry at intersection of row $i$ and column $j$ as $\mathbf{W}_{i,j}$.
Finally, $|\cdot|$ is the cardinal of a set, and $\mathrm{diag}(\cdot)$ builds a diagonal matrix from a given vector.
\section{Proposed Method}
\label{sec:methodology}
In this work we consider a deep learning framework to address classification problems. Such a framework can be interpreted, as depicted in Figure~\ref{fig:workflow}, as the concatenation of: 1) a feature extractor $\phi$ which is used to map inputs of raw images (matrix $\mathbf{X}\in\mathbb{R}^{n\times l}$, where $n$ is the number of samples and $l$ their dimensions) into high-level easily exploitable features (matrix $\mathbf{F}\in\mathbb{R}^{n\times d}$, where $d$ is the dimension of the feature vectors), and 2) a classifier meant to treat feature vectors as inputs. The purpose of the proposed methodology is to insert an intra-class denoising procedure between these two steps \emph{after the training of the feature extractor}. Once denoised feature vectors have been obtained, only the classifier is retrained for improved accuracy. In the case of few-shot learning, the feature extractor is typically trained on a generic dataset different from the one used to train the classifier~\cite{hu2020leveraging}. Before detailing the proposed denoising solution, let us first introduce some necessary notions and tools.
\begin{figure}
\centering
\begin{tikzpicture}[scale=0.9, thick]
\node[inner sep=0pt](0) at (0,0) {\footnotesize{\begin{tabular}{c}input\\$\mathbf{X}$\end{tabular}}};
\node[inner sep = 0pt, draw, rectangle, minimum height=1.3cm](1) at (1.75,0) {\footnotesize{\begin{tabular}{c}feature\\extractor\\$\phi$\end{tabular}}};
\node[inner sep = 0pt](2) at (4,0) {\footnotesize{\begin{tabular}{c}feature\\tensor\\$\mathbf{F} = \phi (\mathbf{X})$\end{tabular}}};
\node[draw, rectangle, minimum height=1.3cm](3) at (6,0) {\footnotesize{classifier}};
\node(4) at (8,0) {\footnotesize{decision}};
\draw[->,>=stealth']
(0) edge (1)
(1) edge (2)
(2) edge (3)
(3) edge (4);
\end{tikzpicture}
\caption{Architecture of a model based on DNN.}
\label{fig:workflow}
\end{figure}
\subsection{Graph signal processing and graph filters}
The framework of GSP allows to manipulate signals defined on graph structures~\cite{shuman2013emerging, ortega2018graph}. It is then possible to design tools analogous to classical Fourier analysis. We are in particular interested in graph filtering. To introduce these tools, we first need to define graphs. A graph $\mathcal{G} = \langle \mathcal{V}, \mathcal{E}\rangle$ is a tuple made of a countable set $\mathcal{V}$, called vertices, and a set of pairs of vertices $\mathcal{E}$, called edges. Such a graph is usually represented by its binary ($\{0,1\}$) symmetric adjacency matrix $\mathbf{W}$ defined as $\mathbf{W}_{i,j} = 1$ if and only if $(i,j)\in \mathcal{E}$. This matrix can be extended to represent weighted graphs ($\mathbf{W}_{i,j}\in\mathbb{R}^+$).
The diagonal degree matrix $\mathbf{D}$ of $\mathcal{G}$ is defined as:
\begin{equation}
\mathbf{D} = \mathrm{diag}\left( \left[\sum_{j=1}^{|\mathcal{V}|} \mathbf{W}_{i,j}\right]_{i \in \{1, \dots, |\mathcal{V}|\}} \right)\;.
\end{equation}
If all vertices in $\mathcal{G}$ are connected to at least one other vertex, it is possible to define the normalized Laplacian of the graph as the matrix $\mathbf{L}$:
\begin{equation}
\mathbf{L} = \mathbf{I} - \mathbf{D}^{-\frac{1}{2}} \mathbf{W} \mathbf{D}^{-\frac{1}{2}}\;.
\end{equation}
Being symmetric and real-valued, $\mathbf{L}$ can be decomposed into a matrix of orthogonal eigenvectors $\mathbf{U}$, and corresponding nonnegative eigenvalues $\boldsymbol{\lambda}$, with $\lambda_1 \leq \dots \leq \lambda_{|\mathcal{V}|}$, such as $\mathbf{L} = \mathbf{U} \mathrm{diag}(\boldsymbol{\lambda}) \mathbf{U}^\top$.
In the field of GSP, we are interested in manipulating signals over $\mathcal{G}$. A signal is a vector $\mathbf{x}\in \mathbb{R}^{|\mathcal{V}|}$. The product $\hat{\mathbf{x}} = \mathbf{U}^\top \mathbf{x}$ is called the Graph Fourier Transform (GFT) of $\mathbf{x}$. The inverse Graph Fourier Transform (iGFT) is then defined as $\mathbf{x} = \mathbf{U} \hat{\mathbf{x}}$.
By analogy with classical Fourier analysis, values in $\boldsymbol{\lambda}$ are interpreted as frequencies and can thus be used to filter a signal. As such, a filter is usually defined by its spectral response: $h: \lambda \mapsto h(\lambda)$, where $\lambda \in \boldsymbol{\lambda}$. By writing $\mathbf{H} = \mathrm{diag}(h(\boldsymbol{\lambda}))$, the filtered signal $\mathbf{x}^{\mathrm{filter}}$ is then defined as:
\begin{equation}
\mathbf{x}^{\mathrm{filter}} = \mathbf{U}\mathbf{H}\mathbf{U}^\top \mathbf{x}\;.
\end{equation}
To remove high frequencies, a typical low-pass filter $h$ would nullify large values of $\lambda$.
\subsection{Proposed methodology}
\label{sec:methodo}
In this work we propose as stated before to filter out the high frequencies within the feature representation of data to improve classification performance. To achieve this, we first infer graphs from labeled feature vectors within each class and then remove high frequencies using low-pass graph filters.
More precisely, we proceed as follows. Consider all labeled signals in $\mathcal{D}_{\mathrm{base}}$ of class $c$. We denote these signals as $\mathbf{X}$, and the associated features $\mathbf{F} = \phi(\mathbf{X})$ (see Figure \ref{fig:workflow}). We first define a similarity matrix $\mathbf{S}$ between samples in the feature space as follows:
\begin{equation}
\mathbf{S}_{i,j}= s\left(\mathbf{F}_{i,:}, \mathbf{F}_{j,:}\right)\text{~if~} i \neq j \text{~and~} 0 \text{~otherwise}\;
\end{equation}
where $s$ is a similarity measure. In our work, we choose cosine similarity for $s$.
Given a similarity matrix $\mathbf{S}$, we generate the adjacency matrix $\mathbf{W}$ of a graph $\mathcal{G}$ for each class using $k$-nearest neighbors selection: $\mathbf{W}_{i,j} = \mathbf{S}_{i,j}$ if $\mathbf{S}_{i,j}$ is among the $k$ largest entries of $\mathbf{S}_{i,:}$ or $\mathbf{S}_{:,j}$, and 0 otherwise.
The obtained graph is then used to define a low-pass filter $\mathbf{H}$ that we apply on $\mathbf{F}$ as follows:
\begin{equation}
\mathbf{F}^{\mathrm{filter}} = \mathbf{U}\mathbf{H}\mathbf{U}^\top \mathbf{F} \;,
\end{equation}
where $\mathbf{U}$ are the eigenvectors of the Laplacian matrix built from $\mathbf{W}$. Finally, we substitute $\mathbf{F}$ with $\mathbf{F}^{\mathrm{filter}}$ as an input for the classifier in the workflow in Figure \ref{fig:workflow}.
\section{Effect of Low-pass Graph Filters on Centroids}
\label{sec:theory}
We would like to show that graph filtering can have the benefits of both keeping the expectation of centroids of feature vectors invariant while reducing their covariance. As such, graph filtering has the effect of reducing intra-class noise when training the classifier. To this end, we consider a simplified case where data is drawn from a Gaussian model. This is not an undesirable model since many works hypothesize that the features from the same class are aligned with a such a specific distribution~\cite{hu2020leveraging}.
Let $\mathbf{X} \in \mathbb{R}^{m\times l}$ be the subset of labeled samples of class $c$ in $\mathcal{D}_{\mathrm{base}}$, and $\mathbf{F} = \phi(\mathbf{X})$ the associated features ($\mathbf{F} \in \mathbb{R}^{m\times d}$, see Figure \ref{fig:workflow}). In the remainder of this section, we are interested in the distribution of the centroid $\boldsymbol{\gamma} \in \mathbb{R}^{d}$ obtained from features in $\mathbf{F}$. It is defined as:
\begin{equation}
\boldsymbol{\gamma} = \dfrac{1}{m} \sum_{i=1}^{m} \mathbf{F}_{i,:} \;.
\end{equation}
In the following lemma we provide analytical expressions for the mean and the covariance of the filtered centroid, and show their relations with those of the centroid of the original feature vectors.
\begin{lemma}
\label{meancov}
For all $i$, suppose $\mathbf{F}_{i,:}$ are i.i.d such that $\mathbf{F}_{i,:} \sim \mathcal{N}(\boldsymbol{\mu}, \sigma^2 \mathbf{I})$. Denote $\mathbf{L}$ the Laplacian of a graph obtained from $\mathbf{F}$ as described in Section \ref{sec:methodo}, with eigenvectors $\mathbf{U}$ and eigenvalues $\boldsymbol{\lambda}$. Choose $k \in \{1, \dots, m \}$, and define filter $\mathbf{H}$ such that $h(\lambda)=1$ if $\lambda \leq \lambda_k$ and $h(\lambda)=0$ otherwise ($\lambda \in \boldsymbol{\lambda}$). The mean and the covariance of the filtered centroid $\boldsymbol{\gamma}^{\mathrm{filter}}$ are given by:
\begin{equation}
\mathbb{E}\left(\boldsymbol{\gamma}^{\mathrm{filter}}\right) = \dfrac{1}{m} \sum_{j=1}^{k} \left(\mathbf{1}_m^\top\mathbf{U}_{:,j}\right)^2 \mathbb{E}\left(\boldsymbol{\gamma}\right) \;,
\end{equation}
\begin{equation}
\mathrm{Cov}\left(\boldsymbol{\gamma}^{\mathrm{filter}}\right)=\dfrac{1}{m}\left(\sum_{i=1}^{m} \left(\mathbf{U}_{:, :k} \mathbf{U}_{:, :k}^\top\right)_{i,:} \right)^2 \mathrm{Cov}\left(\boldsymbol{\gamma}\right) \;,
\end{equation}
where $\mathbf{1}_m$ is the all-one column vector of dimension $m$.
\end{lemma}
\begin{proof}
See Section \ref{appendix}.
\end{proof}
Notice that the mean and the covariance of the filtered centroid $\boldsymbol{\gamma}^{\mathrm{filter}}$ obtained in Lemma \ref{meancov} are a weighted version of the mean and the covariance of the original centroid $\boldsymbol{\gamma}$, respectively. In the following Corollary \ref{cor:1}, under some conditions on the chosen graph and eigenvectors, we quantify those weights and discuss their effects on the centroids. We show that $\boldsymbol{\gamma}^{\mathrm{filter}}$ exhibits a lower covariance compared to $\boldsymbol{\gamma}$ while keeping the same expectation asymptotically.
\begin{corollary}
\label{cor:1}
Assume that we build a complete graph (i.e., a simple undirected graph in which every pair of distinct vertices is connected by a unique edge) for each class, and under the assumption of Lemma \ref{meancov}, we have:
\begin{equation}
\mathbb{E}(\boldsymbol{\gamma}^{\mathrm{filter}}) = \dfrac{1}{(1-\frac{1}{m})} \mathbb{E}(\boldsymbol{\gamma})\;,
\end{equation}
\begin{equation}
\mathrm{Cov}\left(\boldsymbol{\gamma}^{\mathrm{filter}}\right) = \dfrac{1}{m(1-\frac{1}{m})^2}\mathrm{Cov}(\boldsymbol{\gamma})\;.
\end{equation}
In particular, by letting the number of labeled samples of that class $m \rightarrow \infty$, we get:
\begin{equation}
\mathbb{E}(\boldsymbol{\gamma}^{\mathrm{filter}}) = \mathbb{E}(\boldsymbol{\gamma}) \;,
\end{equation}
\begin{equation}
\mathrm{Cov}(\boldsymbol{\gamma}^{\mathrm{filter}}) =
\left( \mathrm{Cov}(\boldsymbol{\gamma}) \right)\;.
\end{equation}
\end{corollary}
\begin{proof}
See Section \ref{appendix}.
\end{proof}
While this result holds only for the case of i.i.d feature vectors and a specific choice of a graph and a filter, we conjecture that it provides us with interesting insights for other cases as well, as shown experimentally. More specifically, the numerical results in the next section show that the proposed method improves the accuracy on real datasets and competitive classifiers.
\section{Numerical Results}
\label{sec:numerical-result}
In this section we follow the framework described in Figure \ref{fig:workflow} to conduct our numerical analysis. We show that graph filtering improves the performance of different classification methods in the few-shot and standard settings. The filter we use for our experiments is defined as:
\begin{align}
\label{equ:1}
\mathbf{H}_{i,i} & = \begin{cases} 1 & \text{if~} \ i \leq k_1 \;,\\
0.6 & \text{if~} k_1 \leq i \leq k_2 \;,\\
0 & \text{otherwise}\;, \\
\end{cases}
\end{align}
where $(k_1, k_2) \in \{1, \dots, m\}$ are fixed for each scenario. In our experiments we came with empirically chosen values for $k_1$ and $k_2$, as well as for the choice of the value 0.6. They are not necessarily optimal but gave consistent results across datasets. Finding the best design of a graph filter is a direction of research for future work.
\subsection{Few-shot classification}
In the few-shot scenario we first use a pre-trained DNN (backbone) on a bigger dataset of base classes. We then use the backbone to perform transfer learning to the novel classes. For our experiments and following the standard procedure in the field, at each test run $5$ classes are drawn uniformly at random among the novel classes. For each class, $m=5$ labeled samples and $q=15$ unlabeled samples are uniformly drawn at random. We perform 100,000 iterations and report the mean accuracy and 95\% confidence intervals for each test.
\subsubsection{ Feature extractors and Datasets}
We consider two feature extractors. The
first one is a \textbf{Wide Residual Network} denoted \textbf{WRN} \cite{zagoruyko2016wide} and \textbf{Dense Networks} denoted \textbf{DNet121} \cite{wang2019simpleshot}\footnote{\url{https://github.com/yhu01/PT-MAP}.}.
We perform our experiments on four benchmark datasets: \textbf{MiniImageNet} \cite{vinyals2016matching} denoted \textbf{MINet}, \textbf{CUB} \cite{wahcaltech}, \textbf{CIFAR-FS} \cite{bertinetto2018meta} and \textbf{TieredImageNet} \cite{ren2018meta} denoted \textbf{TINet}. These datasets are split into two parts: the base classes that are chosen to train the feature extractor and novel classes.
\subsubsection{Results}
To reduce noise on the labeled features, at each iteration we form $5$ graphs, each of $5$ nodes corresponding to the labeled features in each class. Then, we apply on them the low-pass filter described in (\ref{equ:1}) \cite{ortega2018graph}. In the few-shot scenario we have graphs with 5 nodes and we fix the value of $(k_1, k_2)$ to $(1,4)$. The results in Table \ref{genneral-accuracy} show that the method brings gains on the performance of state-of-the-art classification method \cite{hu2020leveraging} in the case of $5$-shot, for both \textbf{WRN} and \textbf{DNet121} backbones and for all the considered datasets.
\begin{table}[ht]
\small
\centering
\begin{tabular}{|c|c|c|c|}
\hline
\multicolumn{4}{|c|}{ \textbf{PT-MAP} \cite{hu2020leveraging}} \\
\hline
Dataset & Backbone & No Filter \% & With Filter \% \\
\hline
\multirow{2}{*}{ \textbf{MINet}} & \textbf{WRN} & 88.82 $\pm$ 0.013 & $\mathbf{88.90 } \pm \mathbf{0.011}$ \\
& \textbf{DNet121} & $86.82 \pm 0.014$ & $\mathbf{86.87 } \pm \mathbf{0.014}$ \\
\hline
\textbf{CUB} & \textbf{WRN} & 93.99 $\pm$ 0.011 & $\mathbf{94.14 } \pm \mathbf{0.009}$ \\
\hline
\textbf{ CIFAR} & \textbf{WRN} & 90.68 $\pm$ 0.015 & $\mathbf{90.75} \pm \mathbf{0.015}$ \\
\hline
\textbf{TINet} & \textbf{DNet121} & $90.44\pm 0.014$ & $\mathbf{90.62 } \pm \mathbf{0.013}$ \\
\hline
\end{tabular}
\caption{5-shot accuracy of the proposed methodology compared with state-of-the-art method in the literature, for various backbones.
}
\label{genneral-accuracy}
\end{table}
\begin{comment}
\begin{table}[ht]
\small
\centering
\begin{tabular}{|c|c|c|c|}
\hline
& \multicolumn{2}{c|}{ \textbf{MiniImageNet}} \\
\hline
PT+MAP \cite{hu2020leveraging} & DNet121 & $86.82 \pm 0.014$ \% & $\mathbf{86.86} \pm \mathbf{0.014}$ \% \\
\hline
PT+MAP \cite{hu2020leveraging} & WRN & 88.82 $\pm$ 0.013 \% & $\mathbf{88.90 } \pm \mathbf{0.011}$ \% \\
\hline
\hline
& \multicolumn{2}{c|}{ \textbf{CUB}} \\
\hline
PT+MAP \cite{hu2020leveraging} & WRN & 93.99 $\pm$ 0.011 \% & $\mathbf{94.14 } \pm \mathbf{0.009}$ \% \\
\hline
\hline
& \multicolumn{2}{c|}{ \textbf{CIFAR}} \\
\hline
PT+MAP \cite{hu2020leveraging} & WRN & 90.68 $\pm$ 0.015 \% & $\mathbf{90.75} \pm \mathbf{0.015}$ \% \\
\hline
& \multicolumn{2}{c|}{ \textbf{TieredImageNet}} \\
\hline
SimpleShot \cite{wang2019simpleshot} & DNet121 & 86.66 $\pm$ 0.015 \% & $\mathbf{86.81} \pm \mathbf{0.013}$ \%\\
\hline
PT+MAP \cite{hu2020leveraging} & DNet121 & $90.44\pm 0.014$ \% & $\mathbf{90.62 } \pm \mathbf{0.014}$ \% \\
\hline
\end{tabular}
\caption{5-shot accuracy of the proposed method compared with state-of-the-art methods in the literature, for various backbones.
}
\label{genneral-accuracy}
\end{table}
\end{comment}
\begin{table}[ht]
\vspace{-.3cm}
\small
\centering
\begin{tabular}{|l|c|c|c|}
\hline
& \multicolumn{3}{c|}{ \textbf{CIFAR-10}} \\
\hline
Method & \textbf{WRN} & \textbf{ShakeNet} & \textbf{PyramidNet} \\
\hline
original paper & 95.82 \% &97.96 \% & 98.56 \% \\
NCM & 85.81 \% & \textbf{97.97} \% & 98.52 \% \\
1-NN & 95.81 \% & 97.95 \% & 98.54 \% \\
\hline
1-NN+Filter & \textbf{95.92} \% & \textbf{97.97} \% & \textbf{98.61} \% \\
\hline
\end{tabular}
\caption{Accuracy on the \textbf{CIFAR-10} dataset.
}
\label{accuracy-cifar}
\end{table}
The total number of labeled samples $m$ has an impact on the classification performance when using graph filters. To verify that, we evaluate our proposed method while varying $m$. The results of the experiments are presented in Figure \ref{fig:accuracy-n}. We notice that the more we increase $m$ in each class, the better the performance. As a consequence, we observe the same results as those expected after our simplified analysis in Corollary \ref{cor:1}, which experimentally supports our conjecture.
\vspace{0.3cm}
\begin{figure}[h!]
\begin{center}
\begin{tikzpicture}
\begin{axis}[ grid= major ,
width=0.40\textwidth ,
xlabel = {$m$-shot} ,
ylabel = {Accuracy} ,
xmin = 1, xmax = 5,
ymin = 91.45, ymax = 94.40,
height = 5cm,
legend entries={ Without Filter, With Filter},
legend style={at={(1,0)},anchor=south east}]
\addplot coordinates {(1,91.55) (2,93.37) (3,93.77) (4,93.98) (5,93.99) };
\addplot coordinates {(1, 91.55) (2,93.40) (3,93.82) (4,94.05) (5,94.14) };
\end{axis}
\end{tikzpicture}
\end{center}
\vspace{-.7cm}
\caption{ Evolution of the accuracy on CUB (backbone: WRN) as a function of $m$.}
\label{fig:accuracy-n}
\end{figure}
\vspace{-.7cm}
\subsection{Standard classification}
In the standard classification scenario we use the well known \textbf{CIFAR-10} dataset and three pre-trained architectures, \textbf{WRN} \cite{zagoruyko2016wide}, \textbf{ShakeNet} \cite{gastaldi2017shake} and \textbf{PyramidNet} \cite{han2017deep}. The first model is trained with traditional data augmentation techniques (namely random crop and horizontal flip) while the latter two use a stronger learned policy called fast-autoaugment \cite{lim2019fast}. We extract the features, create $k$-nearest neighbor graphs for each class $(k=10)$. We apply the graph filter defined in (\ref{equ:1}) on each graph and generate our filtered features. Here, graphs have 5000 nodes. Parameters $(k_1, k_2)$ are set to (20, 55). Then, we compare the performance of a 1-NN classifier applied to the filtered features with the performance on the original datasets. The obtained results are described in Table \ref{accuracy-cifar}, where 1\% accuracy corresponds to 100 well-classified test images. The 1-NN classifier on the filtered features was able to improve the performance over both the 1-NN classifier and nearest class mean classifier (NCM) applied without denoising, and even beats the performance of untouched full DNN architecture~\cite{mangla2020charting}.
\section{ Conclusion}
\label{sec:conclusion}
We proposed a graph-based method to improve the accuracy of classification methods. The method consists in using techniques from GSP to reduce the noise in the extracted feature vectors that may affect the networks performance. We showed the effectiveness of the method theoretically. In addition, we performed experiments on standardized vision datasets and we obtained gains in two different settings: few-shot classification and standard classification. A possible future work would be to have an automatic way of choosing the best filter parameters or even integrating them as parameters during the learning phase.
\bibliographystyle{IEEEtran}
|
\section*{Appendices}
\section{POS Tag Set}
\label{sec:pos-tagset}
Table ~\ref{Tab:pos-tags} lists all the part-of-speech (POS) tags used in our experiments.
\begin{table}[h!]
\begin{center}
\begin{tabular}{llll}
\toprule
\bf Tag & \bf Description & \bf Tag & \bf Description \\
\hline
ADV & adverb & ADJ & adjective \\
CONJ & conjunction & DET & determiner \\
NOUN & noun & NSUFF & noun suffix \\
NUM & number & PART & particle \\
PUNC & punctuation & PRON & pronoun \\
PREP & preposition & V & verb \\
ABBREV & abbreviation & VSUFF & verb suffix \\
FOREIGN & non-Arabic & FUT\_PART & future particle \\
PROG\_PART & progressive particle & EMOT & Emoticon/Emoji\\
MENTION & twitter mention & HASH & Hashtag \\
URL & URL & ~~~ -- & ~~~ --\\
\bottomrule
\end{tabular}
\end{center}
\caption{\label{Tab:pos-tags} The POS tag set in~\cite{darwish2018multi}.}
\end{table}
\section{Error Analysis}\label{append:error}
\label{app-err}
\begin{table*}[h!]
\footnotesize
\setlength{\extrarowheight}{2pt}
\begin{tabular}{cc|c|c|c|c|}
& \multicolumn{1}{c}{} & \multicolumn{4}{c}{Predicted} \\
& \multicolumn{1}{c}{} & \multicolumn{1}{c}{PER} & \multicolumn{1}{c}{LOC} & \multicolumn{1}{c}{ORG} & \multicolumn{1}{c}{O} \\\cline{3-6}
& PER & 117 & 2 & 2 & 66 \\ \cline{3-6}
Gold & LOC & 11 & 33 & 1 & 39\\\cline{3-6}
& ORG & 5 & 5 & 5 & 57\\\cline{3-6}
& O & 130 & 14 & 15 & 5,940 \\\cline{3-6}
\end{tabular}
\begin{tabular}{cc|c|c|c|c|}
& \multicolumn{1}{c}{} & \multicolumn{4}{c}{Predicted} \\
& \multicolumn{1}{c}{} & \multicolumn{1}{c}{PER} & \multicolumn{1}{c}{LOC} & \multicolumn{1}{c}{ORG} & \multicolumn{1}{c}{O} \\\cline{3-6}
& PER & 120 & 3 & 2 & 62 \\ \cline{3-6}
Gold & LOC & 10 & 34 & 0 & 40\\\cline{3-6}
& ORG & 5 & 6 & 11 & 66\\\cline{3-6}
& O & 54 & 8 & 2 & 6,035 \\\cline{3-6}
\end{tabular}
\caption{NER confusion matrices for fine-tuning (left) and self-training (right) on the development set of the DA NER data.}
\label{tab:conf-mat}
\end{table*}
\begin{table*}[h!]
\centering
\footnotesize
\begin{tabular}{l|c|l|c|l|c}
\toprule
\textbf{no.} & \textbf{Token} &\textbf{Eng.} &\textbf{MSA} & \textbf{Context/Explanation} & \textbf{FT Pred.} \\
\hline
(1) & \<نبي > & we want & \<نريد> & \< نبي نعرف من...> (\textit{we want to know who}) & PER \\
(2) & \<ماكانوا> & wasn't & \<لم يكونوا>
& \<أغلب الي ماكانوا مصدقين>
(\textit{most of those who wasn't believing}) & LOC \\
(3) & \<لوووول > & LOL & \<ضحك> & \< لوووول...> (interjection) & PER \\
(4) & \< عشان > & for & \<لكي > &\<تبي بطاريات عشان تلعب > (\textit{she wants batteries to play}) & LOC \\
(5) & \<دلوقتي > & now & \<الآن > & \<...اقنعوه ينزل دلوقتي> \textit{(convince him to move now}) & PER \\
(6) & \<ايش> & what & \<ماذا>
& \<ايش رأيك>
(\textit{what do you think?}) & PER \\
(7) & \<قادر > & capable & \<قادر> &\<وبقدرة قادر...> (\textit{magically}; idiomatic expression)& PER \\
(8) & \<المشين > & shameful & \<المشين > & \<المشين طنطاوي > (\textit{shameful Tantawy}; Playful for \textit{General Tant.})& PER \\
(9) & \<ايديكوا > & your hands & \<أيديكم > & \<ابوس ايديكوا اقنعوه... > \textit{(I entreat you to convince hi}m) & PER \\
(10) & \<اسالك > & I ask you & \<أسألك> & \< ودي اسالك شنهي > (\textit{I ask you what}) & ORG \\
(11) & \<مين > & who & \<مين>
& \<صوتك مع مين البدوي >
(\textit{who do you vote for, Badawi}) & PER \\
(12) & \<فلوبي ديسك> & floppy disk & \<قرص مرن> & \<ماسك عليه فلوبي ديسك > (\textit{holds a floppy disk against him}) & PER \\
(13) & \<لحبايب> & loved ones & \<الأحباء> & \<تعال علم يف لحبايب > (\textit{come teach your loved ones}) & LOC \\
(14) & \<ماي> & water & \<ماء> & \<جبت لهم ماي> \textit{(brought them water)} & PER \\
(15) & \<ريتويت> & retweet & \<إعادة تغريد> & \<لو قرفان دوس ريتويت> \textit{(if depressed click retweet)} & PER \\
\bottomrule
\end{tabular}
\caption{\textbf{NER task.} Bigger sample false positives mitigated by self-training. These were correctly predicted as the unnamed entity ``O" by the self-trained model.}
\label{tab:fp_analysis_app}
\end{table*}
\begin{table*}[h!]
\footnotesize
\centering
\iffalse
\begin{tabular}{l|c|c|c|c}
\toprule
\textbf{Token(s)} & \textbf{Gold} & \textbf{FT} & \textbf{ST} \\
\hline
\<ياسر > & per & O & per \\
\<رمضان > & per & O & per \\
\< طنطا > & loc & O & loc \\
\<نايف> & per & O & per \\
\<بوالياس> & per & O & per \\
\<فهد> & per & O & per \\
\<ماضي> & per & O & per \\
\<لكمال> & per & O & per \\
\<لكمال> & per & O & per \\
\bottomrule
\end{tabular}
\hspace{1cm}
\fi
\end{table*}
\begin{table*}[h!]
\centering
\footnotesize
\begin{tabular}{l|c|l|c|c|c}
\toprule
\textbf{no.} & \textbf{Token(s)} & \textbf{Context/Explanation} & \textbf{Gold} & \textbf{FT} & \textbf{ST} \\
\hline
(1) & \<بالمبارك> & \<بالمبارك عاد احنا> \textit{(We are still in Mubarak}) & LOC & PER & O \\
(2) & \<محشش> & \<محشش دخل المحاضرة>
\textit{(a drunk entered the lecture)} & O & PER & PER \\
(3) & \<بأمارة> & \<بأمارة ايه وفين>
\textit{(what is the evidence/sign and where?)} & O & LOC & LOC \\
(4) & \<لمستفشي> & \<لمستفشي قصر الدوباره>
\textit{(to Qasr AlDobara Hospital)} & LOC & O & O \\
(5) & \<كنتاكي> & \<عند كنتاكي>
\textit{(by Kentucky [resturant])} & LOC & O & O \\
(6) & \<داون تاون> & \<مشروع داون تاون بطنطا>
\textit{(a down town Tanta project)}&
LOC & O & O \\
(7) & \<يابطل> & \<مبروك يابطل>
\textit{(Congratulations, hero!)} & O & PER & PER \\
(8) & \<الاخوان> & \<نختلف مع الاخوان> \textit{(we disagree with the Muslim brotherhood)} & ORG & O & O \\
(9) & \<قناة العربية> & \<شفت قناة العربية> \textit{(watched Al Arabya Channel)}& ORG & O & O \\
(10) & \<المجلس العسكري> & \<اللي عمله المجلس العسكري> \textit{(what the military council did)} & ORG & O & O \\
\bottomrule
\end{tabular}
\caption{\textbf{NER task.} Sample errors that are not fixed by self-training (shared with the mere fine-tuned model)}
\label{tab:shared-errors}
\end{table*}
\section{Introduction}
\label{intro}
Neural language models~\cite{xu2000can,bengio2003neural} with vectorized word representations~\cite{mikolov2013efficient} are currently core to a very wide variety of NLP tasks. In specific, using representations from transformer-based~\cite{vaswani2017attention} language models~\cite{devlin2018bert,liu2019roberta}, pre-trained on large amounts of unlabeled data and then fine-tuned on labeled task-specific data, has become a popular approach for improving downstream task performance. This pre-training then fine-tuning scheme has been successfully applied to several tasks, including question answering \cite{yang2019enhancing}, social meaning detection~\cite{mageedL2020:aranet}, text classification~\cite{liu2019roberta}, named entity recognition (NER), and part-of-speech (POS) tagging~\cite{tsai2019small,conneau2019unsupervised}. The same setup also works well for cross-lingual learning~\cite{lample2019cross,conneau2019unsupervised}.
Given that it is very expensive to glean labeled resources for all language varieties and dialects, a question arises: ``How can we leverage resource-rich dialects to develop models nuanced to downstream tasks for resource-scarce ones?". In this work, we aim to answer this particular question by
applying self-training to unlabeled target dialect data.
We empirically show that self-training is indeed an effective strategy in \textit{zero-shot} (where no gold dialectal data are included in training set, Section ~\ref{sec:msa-da-zero}) and \textit{few-shot} (where a given number of gold dialectal data points is included in training split, Section~\ref{sec:few-shot}).
Our few-shot experiments reveal that self-training is always a useful strategy that \textit{consistently} improves over mere fine-tuning, even when \textit{all} dialect-specific gold data are used for fine-tuning. In order to understand why this is the case (i.e., why combining self-training with fine-tuning yields better results than mere fine-tuning), we perform an extensive error analysis based on our NER data. We discover that self-training helps the model most (\% = 59.7) with improving false positives. This includes DA tokens whose MSA orthographic counterparts \cite{shaalan2014survey} are either named entities or trigger words that frequently co-occur with named entities in MSA. Interestingly, such out-of-MSA tokens occur in highly dialectal contexts (e.g., interjections and idiomatic expressions employed in interpersonal social media communication) or ones where the social media context in which the language (DA) is employed affords more freedom of speech~\cite{alshehri2020osact4} and a platform for political satire. We present our error analysis in Section~\ref{sec:error}.
\textbf{Context:} Language use in social media tends to diverge from `standard', offline norms~\cite{danet2007multilingual,herring2015computer}. For example, users employ slang, emojis, abbreviations, letter repetitions, and other types of playful practices. This poses a challenge for processing social media data in general. However, there are other challenges specific to Arabic that motivate our work. More specifically, we choose Arabic to apply our approach since it affords a rich context of linguistic variation: In addition to the standard variety, MSA, Arabic also has several spoken dialects~\cite{abdul2018you,bouamor2019madar,mageedetal2020nadi,mageed2020micro}, which differ significantly from the written MSA \cite{zaidan2014arabic} thus offering an excellent context for studying our problem. Arabic dialects differ among themselves and from MSA at various linguistic levels: lexical, phonological, morphological, and syntactic. This makes our case much more challenging than that of standard vs. social media English, for example. For a good zero-shot performance in our case, a model is required to accommodate not only lexical distance between MSA and DA, but also differences in word formation and syntax (related to POS tags, for example) and lexical ambiguity (as the meaning of the same token can vary cross-dialectically). This makes the zero-shot setting even harder, where the performance drops ~20\% F1 points (See section ~\ref{sec:msa-da-zero}).
From a geopolitical perspective, Arabic also has a strategic significance. This is a function of Arabic being the native tongue of $~$ 400 million speakers in 22 countries, spanning across two continents (Africa and Asia)\footnote{https://www.internetworldstats.com/stats19.htm}. In addition, the three dialects of our choice, namely Egyptian (EGY), Gulf (GLF), and Levantine (LEV), are popular dialects that are widely used online. This makes our resulting models highly useful in practical situations at scale. Pragmatically, ability to develop NLP systems on dialectal tasks with no-to-small labeled dialect data immediately eases a serious bottleneck. Arabic dialects differ among themselves and from MSA at all linguistic levels, posing challenges to traditional NLP approaches.
We also note that our method is language-independent, and we hypothesize it can be directly applied to other varieties of Arabic or in other linguistic contexts for other languages and varieties.
\textbf{Tasks:} We apply our methods on two sequence labeling tasks, where we have access to both MSA and DA gold data. In particular, as mentioned above, we perform experiments on POS tagging and NER. Each of these tasks has become an integral part of various other NLP applications, including question answering, aspect-based sentiment analysis, machine translation, and summarization, and hence our developed models should have wide practical use. Again, we note that our approach itself is task-independent. The same approach can thus be applied to other tasks involving DA. We leave testing our approach on other languages, varieties, and tasks for future research.
\newblock
\textbf{Contributions:} Our work offers the following contributions:
\begin{enumerate}
\item We study the problem of MSA-to-DA transfer in the context of sequence labeling and show that when training on MSA data only, a wide performance gap exists between testing on MSA and DA. That is, models fine-tuned on MSA generalize poorly to DA in zero-shot settings.
\item We propose self-training to improve zero- and few-shot MSA-to-DA transfer. Our approach requires little-to-no labeled DA data. We evaluate extensively on 3 different dialects, and show that our method indeed narrows the performance gap between MSA and DA by a margin as wide as \texttildelow 10\% F$_1$ points.
\item We develop state-of-the-art models for the two sequence labeling tasks (NER and POS).
\end{enumerate}
We now introduce our method.
\section{Method}~\label{sec:method}
While the majority of labeled Arabic datasets are in MSA, most daily communication in the Arab world is carried out in DA. In this work, we show that models trained on MSA for NER and POS tagging generalize poorly to dialect inputs when used in zero-shot-settings (i.e., no dialect data used during training). Across the two tasks, we test how self-training would fare as an approach to leverage unlabeled DA data to improve performance on DA. As for self-training, it involves training a model using its own predictions on a set of unlabeled data identical from its original training split. Our proposed self-training procedure is given two sets of examples: a labeled set $L$ and an unlabeled set $U$. To perform zero-shot MSA-to-DA transfer, MSA examples are used as the labeled set, while unlabeled DA examples are the unlabeled set. As shown in Figure ~\ref{Fig:Model}, each iteration of the self-training algorithm consists mainly of three steps. First, a pre-trained language model is fine-tuned on the labeled MSA examples $L$. Second, for every unlabeled DA example $u_i$, we use the model to tag each of its tokens to obtain a set of predictions and confidence scores for each token $p_{u_i} = (l^{(i)}_{1}, c^{(i)}_{1}), (l^{(i)}_{2}, c^{(i)}_{2}), ... (l^{(i)}_{|u_i|}, c^{(i)}_{|u_i|})$, where $(l^{(i)}_{j}, c^{(i)}_{j})$ are the label and confidence score (softmax probability) for the $j$-th token in $u_i$. Third, we employ a selection mechanism to identify examples from $U$ that are going to be added to $L$ for the next iteration.
For a selection mechanism, we experiment with both a thresholding approach and a fixed-size~\cite{dong2019robust} approach. In the thresholding method, a threshold $\tau$ is applied on the minimum confidence per example. That is, we only add an example $u_i$ to $L$ if $\min \limits_{(l^{(i)}_{j}, c^{(i)}_{j}) \in p_{u_i}}{c^{(i)}_j} \geq \tau$. \mam{See Algorithm~\ref{alg}}. The fixed-size approach involves, at each iteration, the selection of the top $S$ examples with respect to the minimum confidence score $\min \limits_{(l^{(i)}_{j}, c^{(i)}_{j}) \in p_{u_i}}{c^{(i)}_j}$ , where $S$ is a hyper-parameter. We experiment with both approaches and report results in Section~\ref{sec:res}.
\begin{figure}[t!]
\hspace{0.6cm}
\includegraphics[width=6.5cm, height=5.5cm]{Self-training3}
\caption{MSA-to-DA Self-training transfer.}
\label{Fig:Model}
\end{figure}
\begin{algorithm}[ht]
\footnotesize
\SetAlgoLined
\textbf{Given} set $L$ of labeled MSA examples, set $U$ of unlabeled DA examples, $\tau$ parameter for probability threshold selection.
\textbf{repeat}
\Indp Fine-tune model $M$ for $K$ epochs on labeled MSA examples $L$;\\
\For{$u_i \in U$}{
Obtain prediction $p_{u_i}$ on unlabeled DA example $u_i$ using model $M$;\\
\uIf{$\min \limits_{(l^{(i)}_{j}, c^{(i)}_{j}) \in p_{u_i}}{c^{(i)}_j} \geq \tau$}{ remove $u_i$ from $U$ and add it to $L$;\\
}
}
\Indm
\textbf{until} stopping criterion satisfied
\caption{\label{alg}MSA-to-DA Self-Training}
\end{algorithm}
For our language model, we use XLM-RoBERTa~\cite{conneau2019unsupervised}, XML-R for short. XLM-R is a cross-lingual model, and we choose it since it is reported to perform better than the multilingual mBERT ~\cite{devlin2018bert}.
XLM-R also uses Common Crawl for training, which is more likely to have dialectal data than the Arabic Wikipedia (used in mBERT), making it more suited to our work. We now introduce our experiments.
\section{Experiments}\label{sec:exps}
We begin our experiments with evaluating the standard fine-tuning performance of XLM-R models on both NER and POS tagging against strong baselines. We then use our best models from this first round to investigate the MSA-to-DA zero-shot transfer, showing a significant performance drop even when using pre-trained XLM-R. Consequently, we employ self-training for both NER and POS tagging in zero- and few-shot settings, showing substantial performance improvements in both cases. We now introduce our datasets
\subsection{Datasets}~\label{subsec:data}
\noindent
\textbf{NER:} For our work on NER, we use 4 datasets: ANERCorp~\cite{Benajiba2007}, \texttildelow 150K tokens; ACE 2003 ~\cite{mitchell2003tides} BNews (BN-2003), \texttildelow 15K tokens; ACE 2003 Newswire (NW-2003), \texttildelow 27K tokens; and Twitter~\cite{darwish2013named}, \texttildelow 81K tokens. Named entity types in all datasets are \textit{location (LOC)}, \textit{organization (ORG)}, and \textit{person (PER)}
\textbf{POS Tagging:} There are a number of Arabic POS tagging datasets, mostly on MSA~\cite{maamouri2004penn} but also on dialects such as EGY~\cite{maamouri2014developing}. To show that the proposed approach is able to work across multiple dialects, we ideally needed data from more than one dialect. Hence, we use the multi-dialectal dataset from~\cite{darwish2018multi}, comprising 350 tweets from each of the 4 varieties MSA, EGY, GLF and LEV. This dataset has 21 POS tags, some of which are suited to social media (since it is derived from Twitter). We show the POS tag set from ~\cite{darwish2018multi} in Table~\ref{Tab:pos-tags} in Appendix ~\ref{sec:pos-tagset}. We now introduce our baselines.
\subsection{Baselines}
For the \textbf{NER task}, we use the following baselines:
\begin{itemize}
\item \textbf{NERA}\textbf{~\cite{abdallah2012integrating}}: A hybrid system of rule-based features and a decision tree classifier.
\item \textbf{WC-BiLSTM}\textbf{~\cite{gridach2016character}}: A character- and a word-level Bi-LSTM with a conditional random fields (CRF) layer.
\item \textbf{WC-CNN}\textbf{~\cite{khalifa2019character}}: A character- and a word-level CNN with a CRF layer.
\item \textbf{mBERT}\textbf{~\cite{devlin2018bert}}: A fine-tuned multilingual BERT-Base-Cased (110M parameters), pre-trained with a masked language modeling objective on the Wikipedia corpus of 104 languages (including Arabic). For fine-tuning, we find that (based on experiments on our development set) a learning rate of $6 \times 10^{-5}$ works best with a dropout of 0.1.
\end{itemize}
In addition, we compare to the published results in~\cite{shaalan2014hybrid}, AraBERT~\cite{antoun2020arabert}, and CAMel~\cite{obeid2020camel} for the ANERCorp dataset. We also compare to the published results in~\cite{khalifa2019character} for the 4 datasets.
For the \textbf{POS tagging task}, we compare to our own implementation of WC-BiLSTM (since there is no published research that uses this method on the task, as far as we know) and run mBERT on our data. We also compare to the CRF results published by~\cite{darwish2018multi}. In addition, for the Gulf dialect, we compare to the BiLSTM with compositional character representation and word representations (CC2W+W) published results in~\cite{alharbi2018part}.
\subsection{Experimental Setup}
Our main models are XLM-RoBERTa base architecture XLM-R\textsubscript{B} $(L=12, H=768, A=12, \textnormal{ 270M params})$ and XLM-RoBERTa large architecture XLM-R\textsubscript{L} $(L=24, H=1024, A=16,\textnormal{ 550M params})$, where $L$ is number of layers, $H$ is the hidden size, $A$ is the number of self-attention heads. For XLM-R experiments, we use Adam optimizer with $1e^{-5}$ learning rate, batch size of 16. We typically fine-tune for 20 epochs, keeping the best model on the development set for testing. We report results on the test split for each dataset, across the two tasks. For all BiLSTM experiments, we use the same hyper-parameters as~\cite{khalifa2019character}.
For the standard fine-tuning experiments, we use the same train/development/test split as in \cite{khalifa2019character} for NER, and the same split provided by \cite{darwish2018multi} for POS tagging.
For all the self-training experiments, we use the dialect subset of the Arabic online news commentary (AOC) dataset~\cite{zaidan2011arabic}, comprising the EGY, GLF, and LEV varieties limiting to equal sizes of 9K examples per dialect (total =27K)~\footnote{We note that our approach could be scaled with an even bigger unlabeled dataset, given the performance gains we report with self-training in this work.}. We use the split from~
\cite{elaraby2018deep} of AOC, removing the dialect labels and just using the comments themselves for our self-training. Each iteration involved fine-tuning the model for $K=5$ epochs. As a stopping criterion, we use early stopping with patience of 10 epochs. Other hyper-parameters are set as listed before. For selecting confident samples, we experiment with a fixed number of top samples $S=[50, 100, 200]$ and selection based on a probability threshold $\tau=[0.80, 0.90, 0.95]$ (softmax values)~\footnote{It is worth noting that our $S$ values are similar to those used in~\cite{dong2019robust}. We also experimented with other values for $\tau$ and $S$, but found them sub-optimal and hence we report performance only for the listed values of these two hyper-parameters here.}. For all evaluations, we use the \textit{seqeval} toolkit.\footnote{\url{https://github.com/chakki-works/seqeval}.}
\section{Results}\label{sec:res}
\subsection{Fine-tuning XLM-R}\label{sec:finetune}
Here, We show the resuts of standard fine-tuning of XLM-R for the two tasks in question. We start by showing the result of fine-tuning XLM-R on the \textbf{named entity task}, on each of the 4 Arabic NER (ANER) datasets listed in Section~\ref{subsec:data}. Table~\ref{Tab:ner} shows the test set macro F$_1$ score on each of the 4 ANER datasets. Clearly, the fine-tuned XLM-R models outperform other baselines on all datasets, except on the NW-2003 where WC-CNN~\cite{khalifa2019character} performs slightly better than XLM-R\textsubscript{L}.
For \textbf{POS Tagging}, Table ~\ref{Tab:pos} shows test set word accuracy of the XLM-R models compared to baselines. Again, XLM-R models (both base and large) outperform all other models. A question arises why XLM-R models outperform both mBERT and AraBERT. As noted before, for XLM-R vs. mBERT, XLM-R was trained on much larger data: CommonCrawl for XLM-R vs. Wikipedia for mBERT. Hence, the \textit{larger dataset} of XLM-R is giving it an advantage over mBERT. For comparison with AraBERT, although the pre-training data for XLM-R and AraBERT may be comparable, even the smaller XLM-R model (XLM-R\textsubscript{B}) has more than twice the number of parameters of the BERT\textsubscript{BASE} architecture on which AraBERT and mBERT are built (270M v. 110M). Hence, XLM-R model \textit{capacity} gives it another advantage. We now report our experiments with zero-shot transfer from MSA to DA.
\begin{table*}[h]
\begin{center}
\footnotesize
\begin{tabular}{lllll}
\toprule
\bf Model & \bf ANERCorp & \bf BN-2003 & \bf NW-2003 & \bf Twitter \\ \hline
NERA \cite{abdallah2012integrating} & 88.77 & ~ ~ -- & ~ ~ -- & ~ ~ -- \\
CAMeL \cite{obeid2020camel} & 85.00 & ~ ~ -- & ~ ~ -- & ~ ~ -- \\
Hybrid \cite{shaalan2014hybrid} & 90.66 & ~ ~ -- & ~ ~ -- & ~ ~ -- \\
WC-BiLSTM \cite{gridach2016character} & 88.56 & 94.92 & 90.32 & 64.93 \\
WC-CNN \cite{khalifa2019character} & 88.77 & 94.12 & \bf 91.20 & 65.34 \\
mBERT (ours) & 85.86 & 89.52 & 87.19 & 58.92 \\
AraBERT \cite{antoun2020arabert} & 84.2 & ~ ~ -- & ~ ~ -- & ~ ~ -- \\
\hline
XLM-R\textsubscript{B} (ours) & 87.75 & 95.35 & 85.25 & 60.39 \\
XLM-R\textsubscript{L} (ours) & \textbf{91.43} & \textbf{97.33} & 91.10 & \textbf{68.91} \\
\bottomrule
\end{tabular}
\end{center}
\caption{\label{Tab:ner} Test set macro F$_1$ scores for NER.}
\end{table*}
\begin{table*}[h]
\begin{center}
\footnotesize
\begin{tabular}{lllll}
\toprule
\bf Model & \bf \bf MSA & \bf EGY & \bf GLF & \bf LEV \\ \hline
BiLSTM (CC2W + W) \cite{alharbi2018part} & ~~-- & ~~-- & 89.7 & ~~-- \\
CRF \cite{darwish2018multi} & 93.6 & 92.9 & 87.8 & 87.9 \\
WC-BiLSTM (ours) & 94.63 & 93.41 & 88.79 & 86.13\\
mBERT (ours) & 90.57 & 92.88 & 87.85 & 72.30 \\
XLM-R\textsubscript{B} (ours) & 96.30 & 94.70 & 92.18 & 89.98\\
XLM-R\textsubscript{L} (ours) & \bf 98.21 & \bf 97.00 & \bf 94.41 & \bf 93.19\\
\bottomrule
\end{tabular}
\end{center}
\caption{\label{Tab:pos} Test set accuracy for POS Tagging.}
\end{table*}
\subsection{MSA-DA Zero-Shot Transfer}\label{sec:msa-da-zero}
We start by the discussion of \textbf{NER experiments}. Since there is no publicly available purely dialectal NER dataset on which we can study MSA-to-DA transfer, we needed to find DA data to evaluate on. We observed that the dataset from ~\cite{darwish2013named} contains both MSA and DA examples (tweets). Hence, we train a binary classifier\footnote{The model we use is XLM-R\textsubscript{B} fine-tuned on the AOC using~\cite{elaraby2018deep} split. We achieve development and test accuracies of 90.3\% and 89.4 \%, respectively, outperforming the best results in \cite{elaraby2018deep}.}
to distinguish DA data from MSA. We then extract examples that are labeled with probability $p > 0.90$ as either DA or MSA. We obtain 2,027 MSA examples (\textit{henceforth}, \texttt{Twitter-MSA}) and 1,695 DA examples (\textit{henceforth}, \texttt{Twitter-DA}), respectively. We split these into development and test sets with 30\% and 70\% ratios. As \textbf{for POS Tagging}, we already have the three previously used DA datasets, namely EGY, GLF and LEV. We use those for the zero-shot setting by omitting their training sets and using only the development and test sets.
We first study how well models trained for NER and POS tagging on MSA data only will generalize to DA inputs during test time. We evaluate this zero-shot performance on both the XLM-R\textsubscript{B} and XLM-R\textsubscript{L} models. \textbf{For NER}, we train on ANERCorp (which is pure MSA) and evaluate on both Twitter-MSA and Twitter-DA. While for POS tagging, we train on the MSA subset~\cite{darwish2018multi} and evaluate on the corresponding test set for each dialect. As shown in Table ~\ref{Tab:zero-shot-all}, for NER, a significant generalization gap of around 20 \% F$_1$ points exists between evaluation on MSA and DA using both models. While for \textbf{POS tagging}, the gap is as large as 18.13 \% accuracy for the LEV dialect with XLM-R\textsubscript{B}. The smallest generalization gap is on the GLF variety, which is perhaps due to the high overlap between GLF and MSA~\cite{alharbi2018part}. In the next section, we evaluate the ability of self-training to close this MSA-DA performance gap.
\begin{table*}[h]
\begin{center}
\footnotesize
\begin{tabular}{c|c|c|c|c|c|c}
\toprule
\textbf{Model}
& \multicolumn{2}{c|}{\textbf{NER}}
& \multicolumn{4}{c}{\textbf{POS}} \\
\hline
& \textbf{MSA} & \textbf{DA} & \textbf{MSA} & \textbf{EGY} & \textbf{GLF} & \textbf{LEV} \\
\hline
XLM-R\textsubscript{B} & 60.42 & 40.07 & 96.30 & 78.38 & 83.72 & 78.17 \\
XLM-R\textsubscript{L} & 68.32 & 47.35 & 98.21 & 82.28 & 85.95 & 81.24 \\
\bottomrule
\end{tabular}
\end{center}
\caption{Zero-shot transfer results on DA For NER (macro F$_1$) and POS Tagging (accuracy). Models are trained on MSA only and evaluated on DA. Datasets used are: Twitter-MSA and Twitter-DA \citep{darwish2013named} for NER, and Multi-dialectal \citep{darwish2018multi} for POS tagging.
\label{Tab:zero-shot-all}
}
\end{table*}
\subsection{Zero-shot Self-Training}\label{sec:msa-da-zero-st}
Here, \textbf{for NER}, similar to Section~\ref{sec:msa-da-zero}, we train on ANERCorp (pure MSA) and evaluate on Twitter-MSA and Twitter-DA. Table ~\ref{Tab:st-ner} shows self-training NER results employing the selection mechanisms listed in Section~\ref{sec:method}, and with different values for $S$ and $\tau$. The best improvement is achieved with the thresholding selection mechanism with a $\tau=0.90$, where we have an F$_1$ gain of 10.03 points. More generally, self-training improves zero-shot performance in all cases albeit with different F$_1$ gains.
It is noteworthy, however, that the much higher-capacity large model deteriorates on MSA if self-trained (dropping from 68.32\% to 67.21\%). This shows the ability of the large model to learn representations very specific to DA when self-trained. It is also interesting to see that the best self-trained base model achieved 50.10\% F$_1$, outperforming the large model before the latter is self-trained (47.35\% in the zero-shot setting). As such, we conclude that \textit{\textbf{a base self-trained model, with less computational capacity, can (and in our case does) improve over a large (not-self-trained) model} that needs significant computation}. The fact that, when self-trained, the large model improves 15.35\% points over the base model in the zero-shot setting (55.42 vs. 40.07) is remarkable.
\begin{table}[h]
\begin{center}
\footnotesize
\begin{tabular}{l|l|l}
\toprule
\bf Model & \bf MSA& \bf DA \\ \hline
XLM-R\textsubscript{B} & 61.88 & 40.07 \\
XLM-R\textsubscript{B}, ST, S=50 & 60.98 & 43.88 \\
XLM-R\textsubscript{B}, ST, S=100 & 61.13 & 42.01 \\
XLM-R\textsubscript{B}, ST, S=200 & 61.46 & 43.49 \\
XLM-R\textsubscript{B}, ST, $\tau=0.80$ & \textbf{63.36} & 46.97 \\
XLM-R\textsubscript{B}, ST, $\tau=0.90$ & 61.02 & \bf 50.10 \\
XLM-R\textsubscript{B}, ST, $\tau=0.95$ & 62.25 & 47.91 \\
\hline
XLM-R\textsubscript{L} & \textbf{68.32} & 47.35 \\
XLM-R\textsubscript{L} + ST, $\tau=0.90$ & 67.21 & \textbf{55.42} \\
\bottomrule
\end{tabular}
\end{center}
\caption{\label{Tab:st-ner} Zero-short self-training (ST) NER results. Models trained on ANERCorp (pure MSA) and evaluated on Twitter-MSA and Twitter-DA we extract from~\cite{darwish2018multi}. Self-training boosts the performance on DA data by 10\% macro F1 points with XLM-R\textsubscript{B} and $\tau=0.90$.}
\end{table}
As \textbf{for POS tagging}, we similarly observe consistent improvements in zero-shot transfer with self-training (Table~\ref{Tab:zero}). The best model achieves accuracy gains of 2.41\% (EGY), 1.41\% (GLF), and 1.74\% (LEV). Again, this demonstrates the utility of self-training pre-trained language models on the POS tagging task even in absence of labeled dialectal POS data (zero-shot).
\begin{table}[h]
\begin{center}
\footnotesize
\begin{tabular}{l|l|l|l|l}
\toprule
\bf Model & \bf MSA & \bf EGY & \bf GLF & \bf LEV \\ \hline
XLM-R\textsubscript{B} & 96.30 & 78.38 & 83.72 & 78.17 \\
XLM-R\textsubscript{B}, ST, S=50 & ~~~ -- & \textbf{80.79} & \textbf{85.13} & \textbf{79.91} \\
XLM-R\textsubscript{B}, ST, S=100 & ~~~ -- & 80.43 & 84.74 & 79.16 \\
XLM-R\textsubscript{B}, ST, S=200 & ~~~ -- & 78.75 & 84.21 & 79.40 \\
XLM-R\textsubscript{B}, ST, $\tau$=0.90 & ~~~ -- & 79.52 & 83.97 & 79.21 \\
XLM-R\textsubscript{B}, ST, $\tau$=0.85& ~~~ -- & 78.97 & 83.53 & 79.06 \\
XLM-R\textsubscript{B}, ST, $\tau$=0.80 & ~~~ -- & 78.88 & 83.72 & 78.50 \\
\hline
XLM-R\textsubscript{L} & 98.21 & 82.28 & 85.95 & 81.24 \\
XLM-R\textsubscript{L}, ST, S=50 & ~~~ -- & \textbf{82.65} & \textbf{87.76} & \textbf{83.70} \\
\bottomrule
\end{tabular}
\end{center}
\caption{\label{Tab:zero} Zero-shot POS tagging transfer accuracy when training on MSA only. \textbf{ST:} self-training.}
\end{table}
\subsection{Few-Shot Self-Training}\label{sec:few-shot}
We also investigate whether self-training would be helpful in scenarios where we have access to some gold-labeled DA data (as is the case with POS tagging). Here, we evaluate the few-shot performance of self-training as increasing amounts of predicted DA data are added to the gold training set. This iteration of experiments focuses exclusively on \textbf{POS tagging}, using a fixed-size $S=100$ of predicted cases for self-training and the XLM-R base model. Figure~\ref{Fig:few-shot} shows how POS tagging test accuracy improves as the percentage of gold DA examples added to the MSA training data increases from 0\% to 100\% on the three dialects (EGY, GLF, and LEV). Comparing these results to those acquired via the standard fine-tuning settings without self-training, we find that self-training does \textit{consistently} improve over fine-tuning. This improvement margin is largest with only 20\% of the gold examples.
\begin{figure*}[t!]
\includegraphics[width=\textwidth, height=3.9cm]{vary-data-sizes.pdf}
\caption{Development accuracy as labeled DA data is added to the training MSA data. \textbf{Left:} Results on EGY. \textbf{Middle: } Results on GLF. \textbf{Right:} Results on LEV. We use fixed-size selection with $S=100$ for self-training models.}
\label{Fig:few-shot}
\end{figure*}
\subsection{Ablation Study}
Here, we conduct an ablation study with the NER task as our playground in order to verify our hypothesis that the performance boost primarily comes from using unlabeled DA data for self-training. By using a MSA dataset with the same size as our unlabeled DA one\footnote{We use a set of MSA tweets from the AOC dataset mentioned before.}, we can compare the performance of the self-trained model in both settings: MSA and DA unlabeled data. We run 3 different self-training experiments using 3 different values for $\tau$ using each type of unlabeled data. Results are shown in table ~\ref{Tab:st-msa}. While we find slight performance boost due to self-training even with MSA unlabeled data, the average F1 score with unlabeled DA is better by 2.67 points, showing that using unlabeled DA data for self-training has helped the model adapt to DA data during testing.
\begin{table}[h]
\begin{center}
\footnotesize
\begin{tabular}{l|l|l}
\toprule
\bf Setting & \bf Unlabeled MSA& \bf Unlabeled DA \\ \hline
ST, $\tau=0.80$ & 43.88 & 44.46 \\
ST, $\tau=0.90$ & 44.69 & 47.83 \\
ST, $\tau=0.95$ & 43.43 & 46.87 \\
\hline
Avg & 43.67 & \textbf{46.34} \\
\bottomrule
\end{tabular}
\end{center}
\caption{\label{Tab:st-msa} Ablation experiment with MSA unlabeled data for zero-shot NER. Development set macro F1 is shown when using both unlabeled MSA and DA data with the same size. Average performance with DA unlabeled data is higher showing the effect of unlabeled DA on model final performance.}
\end{table}
\section{Error analysis}\label{sec:error}
\begin{table}[ht!]
\footnotesize
\centering
\begin{tabular}{c|c|c|c}
\toprule
\textbf{Measure} & \textbf{FT} & \textbf{ST} & \textbf{\% improvement} \\
\hline
True Positives & 155 & 165 & 6.5 \% \\
False Positive & 159 & 64 & \textbf{59.7 \%}\\
False Negative & 162 & 168 & -3.7 \% \\
True Negative & 5,940 & 6,035 & 1.5 \% \\
\bottomrule
\end{tabular}
\caption{Comparison of error categories in percentage between the fine-tuned model (FT) and the model combining fine-tuned+self-trained (ST) model, based on the dialectal part of the dev set of the NER task. }
\label{tab:st-fp}
\end{table}
\begin{table*}[]
\centering
\footnotesize
\begin{tabular}{l|c|l|c|l|c}
\toprule
\textbf{no.} & \textbf{Token} &\textbf{Eng.} &\textbf{MSA} & \textbf{Context/Explanation} & \textbf{FT Pred.} \\
\hline
(1) & \<نبي > & we want & \<نريد> & \< نبي نعرف من...> (\textit{we want to know who}) & PER \\
(2) & \<ماكانوا> & wasn't & \<لم يكونوا>
& \<أغلب الي ماكانوا مصدقين>
(\textit{most of those who wasn't believing}) & LOC \\
(3) & \<لوووول > & LOL & \<ضحك> & \< لوووول...> (interjection) & PER \\
(4) & \< عشان > & for & \<لكي > &\<تبي بطاريات عشان تلعب > (\textit{she wants batteries to play}) & LOC \\
(5) & \<دلوقتي > & now & \<الآن > & \<...اقنعوه ينزل دلوقتي> \textit{(convince him to move now}) & PER \\
(6) & \<ايش> & what & \<ماذا>
& \<ايش رأيك>
(\textit{what do you think?}) & PER \\
(7) & \<قادر > & capable & \<قادر> &\<وبقدرة قادر...> (\textit{magically}; idiomatic expression)& PER \\
\bottomrule
\end{tabular}
\caption{Sample false positives mitigated by self-training. These were correctly predicted as the unnamed entity ``O" by the self-trained model.}
\label{tab:fp_analysis}
\end{table*}
To understand why self-training the pre-trained language model, when combined with fine-tuning, improves over mere fine-tuning, we perform an error analysis. For the error analysis, we focus on the NER task where we observe a huge self-training gain. We use the development set of Twitter-DA (See section ~\ref{sec:msa-da-zero-st}) for the error analysis. We compare predictions of the standard fine-tuned XLM-R\textsubscript{B} model (FT) and the best performing self-training ($\tau=0.9$) model (ST) on the data, and provide the confusion matrices of both models with gold labels in Table ~\ref{tab:conf-mat} (in Appendix ~\ref{app-err}). The error analysis leads to an interesting discovery: The greatest benefit from the ST model comes mostly from reducing \textit{false positives} (see Table~\ref{tab:st-fp}). In other words, self-training helps regularize the model predictions such that tokens misclassified by the original FT model as a named entities are now correctly tagged as \textit{unnamed entity} ``O".
To understand why the ST model improves false positive rate, we manually inspect the cases it correctly identifies that were misclassified by the FT model. We show examples of these cases in Table~\ref{tab:fp_analysis}. As the table shows, the ST model is able to identify dialectal tokens whose equivalent MSA forms can act as trigger words (usually followed by a PER named entity). We refer to this category as \textbf{\textit{false trigger words}}. An example is the word \<نبي > ``prophet" (row 1 in Table~\ref{tab:fp_analysis}). A similar example that falls within this category is in row (2), where the model is confused by the token \<الى> ( ``who" in EGY, but ``to" in MSA and hence the wrong prediction as LOC). A second category of errors is caused by \textbf{\textit{non-standard social media language}}, such as use of letter repetition in interjections (e.g., in row (3) in Table~\ref{tab:fp_analysis}). In these cases, the FT model also assigns the class PER, but the ST model correctly identifies the tag as ``O". A third class of errors arises as a result of \textit{\textbf{out-of-MSA}} vocabulary. For example, the words in rows (4-6) are all out-of-MSA where the FT model, not knowing these, assigns the most frequent named entity label in train (PER). A fourth category of errors occurs as a result of a token that is usually part of a named entity in MSA, that otherwise functions as part of an \textit{\textbf{idiomatic expression}} in DA. Row (7) in Table~\ref{tab:fp_analysis} illustrates this case. Table ~\ref{tab:fp_analysis_app} in Appendix ~\ref{app-err} provides more examples.
We also investigate errors shared by both the FT and ST models (errors which the ST model also could not fix). Some of these errors result from the fact that often times both MSA and DA use the same word for both person and location names. Row (1) in Table ~\ref{tab:shared-errors} (in Appendix~\ref{append:error}) is an example where the word ``Mubarak", name of the ex-Egypt President, is used as LOC. Other errors include \textit{out-of-MSA} tokens mistaken as named entities. An example is in row (3) in Table ~\ref{tab:shared-errors}, where \<بأمارة> ,(``proof" or ``basis" in EGY) is confused for \<بإمارة> (``emirate", which is a location). \textit{False trigger words}, mentioned before, also play a role here. An example is in row (7) where \<يابطل> is confused for PER due to the trigger word \<يا> ``Hey!" that is usually followed by a person name. \textit{\textbf{Spelling mistakes}} cause the third source of errors, as in row (4). We also note that even with self-training, detecting ORG entities is more challenging than PER or LOC. The problem becomes harder when such organizations are not seen in training such as in rows (8) \<الاخوان المسلمين>, (9) \<قناة العربية> and (10) \<المجلس العسكري>, all of which do not occur in the training set (ANERCorp).
\textbf{False negatives.} The ``regularizing" effect caused by self-training we discussed thus far can sometimes produce \textit{false negatives} as shown in Table ~\ref{tab:fn-analysis}. We see a number of named entities that were misclassified by the self-trained model as unnamed ones. As an example, we take the last name \<الجنزوري> which was classified both correctly and incorrectly in different contexts by the self-trained model. Context of correct classification is ``\<هاش تاج لكمال الجنزوري> ", while it is ``\<ماسك على الناس كلها سي دي الا الجنزوري ماسك > \\ \< عليه فلوبي>" for the incorrect classification. First, we note that \<الجنزوري> is not a common name (zero occurrences in the MSA training set). Second, we observe that in the correct case, the word was preceded by the first name \<كمال> which was correctly classified as PER, making it easier for the model to assign PER to the word afterwards as a surname.
\begin{table}[]
\centering
\begin{tabular}{l|c|c|c|c}
\toprule
\textbf{no.} & \textbf{Word} & \textbf{Gold} & \textbf{FT} & \textbf{ST} \\
\hline
(1) & \<الاخوان > & ORG & ORG & O \\
(2) & \<للبرادعي > & PER & PER & O \\
(3) & \< مجدي الجلاد > & PER & PER & O \\
(4) & \<فان ديزل> & PER & PER & O \\
(5) & \<الجنزوري> & PER & PER & O \\
(6) & \<زين يسون> & PER & PER & O \\
\bottomrule
\end{tabular}
\caption{\textbf{NER task.} Sample false negatives produced by self-training.}
\label{tab:fn-analysis}
\end{table}
\section{Related Work}\label{sec:rel}
\textbf{Sequence Labeling.} Recent work on sequence labeling usually involves using a word- or character-level neural network with a CRF layer ~\cite{lample2016neural,ma2016end}. These architectures have also been applied to Arabic sequence tagging~\cite{gridach2016character,alharbi2018part,khalifa2019character,al2020transfer,el2019arabic}, producing better or comparable results to classical rule-based approaches \cite{shaalan2014hybrid}. We refer the interested reader to \cite{shoufan2015natural} and \cite{al2018deep} for surveys on Arabic NLP.
\textbf{Pre-trained Language Models.}
Language models, based on Transformers~\cite{vaswani2017attention}, and pre-trained with the masked language modeling (MLM) objective have seen wide use in various NLP tasks. Examples include BERT~\cite{devlin2018bert}, RoBERTa~\cite{liu2019roberta}, MASS~\cite{song2019mass}, and ELECTRA~\cite{clark2020electra}. While they have been applied to several tasks, including text classification, question answering, named entity recognition~\cite{conneau2019unsupervised}, and POS tagging~\cite{tsai2019small}, a sufficiently large amount of labeled data is required for good performance. Concurrent with our work,~\newcite{mageed2020marbert} released MARBERT, a language model trained on a large amount of dialectal Arabic data. However, the extent to which dialect-specific models such as MARBERT can alleviate lack of labeled data remains untested
\textbf{Cross-lingual Learning.} Cross-lingual learning is of particular importance due to the scarcity of labeled resources in many of the world's languages. The goal is to leverage existing labeled resources in high-resource languages (such as English) to optimize learning for low-resource ones. In our case, we leverage MSA resources for building DA models. With proximity to our work,~\newcite{kim2017cross} trained a POS tagger for different languages using English-resources only using two BiLSTM networks to learn common and language-specific features. \newcite{xie2018neural} made use of bilingual word embeddings with self-attention to learn cross-lingual NER for low-resource languages
Multilingual extensions of LMs have emerged through joint pre-training on multiple languages. Examples include mBERT~\cite{devlin2018bert}, XLM~\cite{lample2019cross} and XLM-RoBERTa~\cite{conneau2019unsupervised}. Such multilingual models have become useful for few-shot and zero-shot cross-lingual settings, where there is little or no access to labeled data in the target language. For instance~\newcite{conneau2019unsupervised} evaluated a cross-lingual version of RoBERTa~\cite{liu2019roberta}, namely XLM-R, on cross-lingual learning across different tasks such as question answering, text classification, and named entity recognition.
\textbf{Self-Training.} Self-Training is a semi-supervised technique to improve learning using unlabeled data. Self-training has been successfully applied to NER~\cite{kozareva2005self}, POS tagging~\cite{wang2007semi}, parsing~\cite{sagae2010self} and text classification~\cite{van2016predicting}. Self-training has also been applied in cross-lingual settings when gold labels are rare in the target language.~\newcite{hajmohammadi2015combination} proposed a combination of active learning and self-training for cross-lingual sentiment classification. ~\newcite{pan2017cross} made use of self-training for named entity tagging and linking across 282 different languages. Lastly,~\newcite{dong2019robust} employed self-training to improve zero-shot cross-lingual classification with mBERT \cite{devlin2018bert}.
\section{Conclusion}\label{sec:conc}
Even though pre-trained language models have improved many NLP tasks, they still need labeled data for fine-tuning. We show how self-training can boost the performance of pre-trained language models in zero- and few-shot settings on various Arabic varieties. We apply our approach to two sequence labeling tasks (NER and POS), establishing new state-of-the-art results on both. Through in-depth error analysis and an ablation study, we uncover why our models work and where they can fail. Our method is \textit{language}- and \textit{task-agnostic}, and we believe it can be applied to other tasks and language settings. We intend to test this claim in future research. Our research also has bearings to ongoing work on language models and self-training, and interactions between these two areas can be the basis of future work. All our models and code are publicly available.
\section*{Acknowledgements}\label{sec:acknow}
MAM gratefully acknowledges support from the Natural Sciences and Engineering Research Council of Canada, the Social Sciences and Humanities Research Council of Canada, Canadian Foundation for Innovation, Compute Canada (\url{www.computecanada.ca}) and UBC ARC-Sockeye (\url{https://doi.org/10.14288/SOCKEYE}).\\
|
\section{Introduction}
The existence of massive, passively evolving galaxies at various cosmic epochs is now well established thanks to observational efforts that span more than two decades, from first discoveries \citep[e.g.][]{Daddi00, daddi05, Cimatti_04, Toft05, toft07, Kriek_06} to the latest spectroscopic confirmations \citep[e.g.][]{toft12,toft17, Gobat12,Whitaker13, Glazebrook17,Schreiber18, Tanaka19,V20,DEugenio20,Stockmann20}. This population of early type galaxies (ETGs), or more generally of quiescent galaxies (QGs){\footnote{in this paper we use the terms ETG and QG interchangeably.}}, experiences low (or negligible) levels of star formation, resides well below the main sequence (MS) of star formation \citep[e.g.][]{Noeske07,daddi07,magdis10, Speagle14,schreiber15_ms} at their corresponding redshift, and is characterised by red colours, indicative of old and evolved stellar populations. With a large fraction of $z \geq$ 2 massive galaxies ($M_{\ast}$\ $> 10^{11}$\,$\rm M_{\odot}$) experiencing a downfall in their star formation activity \citep[e.g.][]{Whitaker_2012,Davidzon17}, the abundance of QGs progressively increases at later times. At the same time, recent studies have attested that a fraction of them are already in place by $z\sim4$ \citep[e.g.][]{Glazebrook17,Schreiber18, Tanaka19,V20}.
While the mere detection of passive and massive galaxies out to $z\sim 3-4$ poses in itself challenges in our theories of galaxy evolution, their existence is perplexing at a more basic level: The very mechanisms responsible for triggering and/or sustaining the cessation of star formation at any redshift remain an unsolved puzzle (for a recent overview, see \citealt{Man18}). Several competing scenarios that attempt to explain quenching have been put forward, promoting mechanisms that prevent gas from cooling (halo quenching, active galactic nuclei and stellar feedback, strangulation, e.g. \citealt{Cattaneo06, Peng15,Henriques15}), as well as gas expulsion via feedback (outflows, e.g. \citealt{Dimatteo05,Hopkins06}) and gas stabilisation (morphological quenching, e.g. \citealt{martig09}). While these plausible quenching mechanisms invoke different processes at different physical scales (from dark matter halos to nuclear activity), it is evident that they all have some common parameters at their core: the gas mass budget, the physical conditions, and the dynamical state of the interstellar medium (ISM). This should not come as a surprise. Since gas and dust are the agents of star formation, it quite naturally follows that they should also be at the heart of its cessation, or in other words, for quenching. Consequently, progress in the field necessitates a careful study of the ISM of QGs, very similar to what has already been achieved for star-forming galaxies (SFGs).
To this end, a large volume of studies have scrutinised the ISM properties of local QGs. By exploring their far-IR (FIR) and millimtetre (mm) dust continuum emission and/or targeting atomic and molecular lines (e.g. low transition CO lines), these studies have revealed low, but not negligible, amounts of dust \citep[e.g.][]{Gomez10, Rowlands12, Smith12, Agius13, Werner14, Boselli14, Lianou16} and gas \citep[e.g.][]{Saintonge11a1, Saintonge11a2, Young11, Cappellari13, Davis14, Boselli14, French15, Alatalo16} in local QGs and overall FIR characteristics that are in contrast to those of local SFGs. In particular, the dust mass fraction ($M_{\rm dust}$/$M_{\ast}$) of local QGs is found to vary between 10$^{-5}$ and 10$^{-4}$, and their molecular gas fraction ($f_{\rm gas}$ = $M_{\rm gas}$/$M_{\ast}$) from 0.3 to 1\%, with both quantities being $\sim 2$ orders of magnitude lower than those of normal galaxies in the local Universe. Similarly, their typical `luminosity-weighted' dust temperatures ($T_{\rm d}$) are $\sim$20\,K \citep[e.g.][]{Smith12}, on average 3-5\,K colder than those of normal SFGs at $z = 0$.
On the other hand, the exploration of the properties of the ISM in QGs beyond the local Universe is still in its infancy, in part due to their strenuous selection (and spectroscopic confirmation), but more importantly due to their faint FIR nature. In spite of these challenges though, some `brave' attempts to measure the $M_{\rm gas}$\ of distant QGs and post-starburst (pSB) galaxies have already been carried out \citep[e.g.][]{Sargent15,Suess17,Rudnick17,Hayashi18,Spilker18,Bezanson19,Williams_20}. These studies of small and inhomogeneous (in terms of selection, and physical properties, e.g. $M_{\ast}$) samples of distant of QGs, have so far provided only loose (if any) constrains in their gas mass budget and its evolution with time, placing their $f_{\rm gas}$\ anywhere between 15\% and $<$3\%. The diverse $f_{\rm gas}$\ measurements emerging from various studies demonstrate the need for a systematic study of the ISM for a carefully and uniformly selected sample of QGs in various redshifts bins.
At the same time, the shape of the FIR spectral energy distributions (SEDs) of distant QGs remains even less explored. Indeed, while reasonable $M_{\rm gas}$\ and $M_{\rm dust}$\ estimates can be achieved with single line (e.g. CO) or single-band dust continuum observations, a full characterisation of their IR SEDs requires multi-wavelength photometry probing both the Wien part and the Rayleigh-Jeans (R-J) tail, a prohibitively expensive task in terms of observing time, even for the limited number of available QGs that benefit from lensing magnification. As such, any progress on this front can only be achieved though stacking \citep[e.g.][G18 hereafter]{Viero13, Man16, Gobat_2018}.
Finally, it is worth recalling that the detailed study of the FIR properties of representative and complete samples of SFGs at various cosmic epochs has advanced our understanding of galaxy evolution through scaling relations and evolutionary tracks that have been used to guide simulations and theoretical models \citep[e.g.][]{Dekel09, Popping14, Narayanan15, Lagos15, Popping17,Dave17,Dave20}. Far-IR studies have revealed that the gas and dust mass fraction of SFGs have sharply declined over time, with SFGs having an ISM mass budget larger by a factor of $\sim$10 at $z \approx 2$ compared to that of SFGs of similar $M_{\ast}$\ in the local Universe \citep[e.g.][]{daddi_2008, daddi_2010, geach11, magdis_2012,magdis17, Tacconi18, Liu19}. At the same time high$-z$ SFGs, despite their higher star formation rates (SFRs), are characterised by shorter gas depletion time scales ($\tau_{\rm dep}$) indicative of a higher star formation efficiency at earlier times \citep[e.g.][]{Tacconi18, Liu19}. Similarly, the $T_{\rm d}$\ of SFGs also increases with redshift \citep[e.g.][]{magdis_2012,magdis17, Magnelli14, schreiber_2018_dust, Cortzen20}, mirroring an evolution in their star formation activity, star formation surface density, and the strength of their radiation field. As a final example, we refer to the construction of representative FIR SEDs of SFGs at various redshifts \citep[e.g.][]{ elbaz11, magdis_2012, schreiber_2018_dust}, which have greatly facilitated models, predictions, and extrapolations.
These successes motivated us to undertake a similar effort, but this time for the population that consists of the ending point of the evolutionary path of SFGs, in other words, for QGs. In this work we aim to characterise the basic FIR properties ($L_{\rm IR}$, $T_{\rm d}$, $M_{\rm dust}$, $M_{\rm gas}$, $f_{\rm gas}$, $\tau_{\rm dep}$), and trace their evolution with time, for a robust and controlled sample of QGs in various redshift bins. For this task we employed a multi-wavelength, mid- to far-IR and millimetre to radio stacking analysis of homogeneously selected and mass complete (log($M_{\ast}$/$M_{\rm\odot}$) $\gtrsim 10.8$) QGs samples in the $z = 0.3 - 1.5$ redshift range, drawn from the COSMOS field. To extend the redshift coverage of our work we further complement our analysis with the $1.5 < z < 2.0$ stacked sample of QGs presented in \citet{Gobat_2018} and which is constructed using the same selection criteria as those employed in our study.
The paper is organised as follows: in Section 2 we present the selection criteria for our samples of QGs in various redshift bins. In Section 3 we present the stacking methodology and the subsequent photometry. In Section 4 we present a detailed analysis of the FIR properties of our stacked ensembles and in Section 5 we explore their evolution with time. In Section 6 we discuss the implications of our findings and provide an outlook for future IR studies of distant QGs with (sub-)millimetre facilities. Finally, in Section 7 we summarise the main findings of our work. Throughout the paper we adopt a \citet{salpeter} initial mass function (IMF) and H$_{0}$ = 70\,km\,s$^{-1}$ Mpc$^{-1}$, $\rm \Omega_{\rm M}$ = 0.3, and $\rm \Omega_{\rm \Lambda}$ = 0.7.
\section{Selection}\label{sec:selection}
\begin{figure*}
\includegraphics[width=0.45\textwidth]{f1.pdf}
\includegraphics[width=0.45\textwidth]{f2.pdf}
\caption{\textbf{Redshift ranges for an efficient use of the \textit{BzK} and
\textit{rJK} colour diagrams}. \textit{Left:} \textit{BzK} diagram for the selection
of $0.3 \leq z < 1.0$ QGs. The yellow, blue, and red contours represent the
loci of stars, SFGs, and QGs at $0.3 \leq z < 1.0$ selected by the $NUVrJ$+$BzK$ selection.
Sources with a $3\sigma$ detection at $24\,\mu$m in \cite{jin_2018} have been excluded. The black lines
represent the original cuts by \citet{daddi_2004} to select
galaxies at $1.4\lesssim z \lesssim 2.5$. The red lines show our
cut at low redshifts (Eq. \ref{eq:bzk_cut}). For reference, the tracks show the
synthetic colours for the
13\,Gyr elliptical (E), early-type spirals
(Sa), and M82-like templates by \citet{polletta_2007}, as marked in the labels. The
colours are computed for $0.3 < z < 1.5$ in step of $\Delta z = 0.3$. No evolutionary effects are included. The black
segments show the reddening vectors at $z = 0.3$, $1$, and
$2$. \textit{Right:} \textit{rJK} diagram for the selection of $1.0\leq z < 1.4$ QGs. The yellow, blue, and red contours represent the
loci of stars, SFGs, and QGs at $1.0\leq z < 1.4$ selected by the $NUVrJ$+$rJK$ selection. The remaining lines are coded as in the left panel.}
\label{fig:polletta07}
\end{figure*}
We selected QGs in the COSMOS field \citep{scoville_2007} based on the latest public photometric catalogue by \cite{laigle_2016}. First, we removed
stars (i.e. \texttt{TYPE=1} flag) and set a magnitude cut at $K_{\rm s} = 24.5$\,mag, the $3\sigma$ depth in $2"$ diameter apertures in the `deep' UltraVista stripes \citep{laigle_2016}.
We further excluded X-ray \textit{Chandra}-detected sources from the latest catalogs by
\cite{civano_2016} and \cite{marchesi_2016}, thus removing active galactic nuclei and strongly star-forming objects. Then, following the approach in G18, we separated red QGs from blue
star-forming sources by combining the rest-frame \textit{NUVrJ} criterion
\citep{arnouts_2007, ilbert_2013} with the observed
\textit{BzK} \citep{daddi_2004} and (a custom) \textit{rJK} diagrams.\\
\textbf{The \textit{BzK} criterion.}
The \textit{BzK} diagram was originally designed to select galaxies at
$1.4\lesssim z\lesssim 2.5$, but it can also be successfully used at lower
redshift \citep{bielby_2012a}. We computed the \textit{B-z},
\textit{z-K} colours using magnitudes in 2'' diameters apertures, following
\cite{mccracken_2010, mccracken_2012}.
We applied a $+0.05$\,mag
correction to the \textit{z-K} colour to better match the stellar
locus from the \cite{lejeune_1997} stellar models reported in the
original selection from \cite{daddi_2004}. Red QGs at $0.3\leq z
< 1.0$ fall in
the portion of the \textit{B-z}, \textit{z-K} plane delimited by:
\begin{equation}
\label{eq:bzk_cut}
\begin{cases}
z-K > 0.3\times(B-z) - 0.5\\
z-K \leq 0.35\times(B-z) + 0.15\\
z-K \leq 0.9\times(B-z)-1.55~ .\\
\end{cases}
\end{equation}
We selected galaxies detected at $5\sigma$ in the $K_{\rm s}$ and $z^+$
bands, and retained sources with lower limits on the $B-z$ colour
compatible with the cuts.\\
\textbf{The \textit{rJK} criterion.}
We adapted the principle of the \textit{BzK} diagram (i.e. two bands
bluer and one band redder than the Balmer/4000\,\AA\ jumps at the
redshift of reference) to classify galaxies in the range $1.0\leq z
<1.4$. We computed the \textit{r-J}, \textit{J-K} colours using
magnitudes measured in 2'' diameter apertures. We selected
red quiescent objects as:
\begin{equation}
\label{eq:rjk_cut}
\begin{cases}
J-K > 0.2\times(r-J) - 0.3\\
J-K < 0.2\times(r-J) + 0.275\\
r-J > 2.0~ .\\
\end{cases}
\end{equation}
We selected sources detected at $5\sigma$ in the $K_{\rm s}$ and $J$
band, and retained objects with lower limits on the $r-J$ colour
compatible with the cuts.\\
We show the observed colour diagrams and the loci occupied by red QGs at $0.3 \leq z \leq 1.4$ in Figure \ref{fig:polletta07}. For guidance, we show
the synthetic colours for a series of templates of different galaxy types from from \cite{polletta_2007} (see
\citealt{bielby_2012a} for a similar exercise with
\citealt{bruzual_2003} stellar population models). We computed the
colours at different redshifts, but without including evolution effects. The locus of 13 Gyr-old QGs falls entirely within the region occupied by $z \lesssim 1$ QGs in the $BzK$ diagram and $z \lesssim 1.4$ QGs in the $rJK$ one. Templates typical of strongly star-forming galaxies lie
entirely outside the quiescent region in both diagrams.
Early-type spirals initially fall in the star-forming
portions of the colour diagrams, later entering the quiescent region.\\
\begin{figure*}
\includegraphics[width=\textwidth]{f11.pdf}
\caption{\textbf{Selection and binning}. \textit{Left:} Grey points showing the final sample of double-colour selected and 24 $\mu$m faint QGs in the redshift vs stellar mass plane. The rose area and the black grid mark the limits of the bins used to group
galaxies for stacking. The number of stacked sources per bin are displayed. The red line indicates the 90\% stellar mass completeness limit for QGs in the `deep' UltraVISTA sample \citep{laigle_2016}, adjusted for the different IMF.
%
\textit{Right:} Blue points showing the position of the stacked massive QGs ($M_\ast>10^{10.8}\, M_\odot$) in the redshift vs sSFR(UV) plane. The error bars on the redshift indicate the width of the bins. Only the statistical uncertainties on sSFR(UV) are shown on the y-axis. The solid red line indicates the $3\sigma$ detection limit for \textit{Spitzer}/MIPS at $24\,\mu$m ($\sigma = 15\,\mu$Jy, \citealt{jin_2018}) converted to SFR using the MS templates from \cite{magdis_2012} and assuming $M_\ast = 10^{11.2}\, M_\odot$, matching the median stellar mass of our sample in each redshift bin. The grey shaded area and the grey line indicate the location of the MS and its 0.3\,dex scatter, computed for a galaxy with $M_\ast = 10^{11.2}\, M_\odot$.}
\label{fig:24um_detections}
\label{fig:binning}
\end{figure*}
The addition of the \textit{BzK} and \textit{rJK} criteria removes more efficiently the red dusty galaxies contaminating the sample of quiescent objects selected with the \textit{NUVrJ} diagram only. Sources without a constraint on the observed colours occupy regions of the \textit{NUVrJ}
diagram with bluer \textit{r-J} and redder \textit{NUV-r} colours,
partially overlapping with the final quiescent sample only below $z<1$. On the other hand, galaxies excluded by the
\textit{BzK} and \textit{rJK} criteria lie close to the line
separating the star-forming population. A 2D Kolmogorov-Smirnov test \citep{fasano_1987} suggests that both the unconstrained and excluded galaxies are drawn from a parent
distribution different from the one of the final quiescent sample.\\
The validity of the double-colour selection is further supported by the lower fraction of \textit{Spitzer}/MIPS
$24\,\mu$m $3\sigma$ detected red galaxies when using
\textit{NUVrJ}+\textit{BzK} and \textit{NUVrJ}+\textit{rJK} colour
selections ($\sim15$\%), rather than a simple \textit{NUVrJ} diagram ($\sim25$\%). For the comparison we use the latest $24\,\mu$m catalogue by \citet{jin_2018}, reaching an rms of $\sim15\,\mu$Jy. Figure
\ref{fig:24um_detections} shows that we can
trace SFRs down to several times below the MS parameterised
as in \citet{schreiber_2015}.
To compute the SFR limit in \citet{jin_2018}, we
rescaled the redshift-dependent, MS templates by \cite{magdis_2012} to match
the observed $24\,\mu$m depth and then integrated over the wavelength
range $8-1000\,\mu$m to derive the total infrared luminosity
($L_{\rm IR}$). We smoothed over the width of the \textit{Spitzer}/MIPS
24\,$\mu$m filter to avoid the strong impact of the PAH
features.
We excluded all the
$24\,\mu$m detected sources in the \citet{jin_2018} catalogue from our sample of QGs. We remark that the sample in the highest redshift bin could still potentially be contaminated by a fraction of galaxies closer to the MS, given the looser constraints on the IR emission. However, most of the contamination in the quiescent sample is due to dusty starbursts (or transitioning galaxies), rather than blue objects on the MS with undetectable IR emission. Indeed, absent IR detections, the median sSFR(UV) of the final stacked samples of massive quiescent objects is constant over the redshift interval under consideration. Therefore, the median location with respect to the MS decreases with redshift, given the steady increase in the normalisation of the former (Figure \ref{fig:binning}).\\
For the purpose of stacking, we further excluded the 1\% of the sample that lay too close to the edges of the IR image mosaics. Finally, we binned our sample by redshift, as shown in Figure \ref{fig:binning}. We split the galaxies selected with the $rJK$ colour at $z = 0.65$, the mid-point of the redshift interval over which this criterion is effective. In this work we focus on a complete sample of massive galaxies at $0.3 < z < 1.5$ (log($M_{\ast}$/$M_{\rm\odot}$) $\geq 10.83)${\footnote{This corresponds to a log($M_{\ast}$/$M_{\rm\odot}$) $\geq 10.6$ mass cut in the original \citet{laigle_2016} catalogue, which is built assuming a Chabrier IMF.}}, matching the median stellar mass of similarly selected objects at $\langle z \rangle = 1.8 $ from G18. This allows us to consistently trace the evolution of the dust content over 10\,Gyr of cosmic evolution. We report the properties of the stacked sample in Table~\ref{tab:properties}.
\section{Source stacking and extraction}\label{sec:method}
Our stacking and source extraction methodology broadly follows that described in G18, namely:
For each redshift bin, we extracted cutouts, centered on each galaxy, from publicly available FIR imaging of the COSMOS field at 24\,$\mu$m \citep{lefloch_2009}, 100\,$\mu$m, and 160\,$\mu$m \citep{lutz_2011}, 250\,$\mu$m, 350\,$\mu$m, and 500\,$\mu$m \citep{oliver_2012}, 850\,$\mu$m \citep{geach_2017}, 1.1\,mm \citep{aretxaga_2011}, 10\,cm \citep{smolcic_2017}, and 20\,cm \citep{schinnerer_2010}. For each band, we then combined the cutouts to create a median image, ignoring masked (e.g. zero-coverage) regions, and estimated the variance of each pixel through bootstrap resampling of the data with half the sample size. We note that these variance cutouts do not show any appreciable structure (i.e. they are flat), suggesting that there is little to no contribution from clustering to the noise \citep[e.g.][]{Bethermin_12}.
As in G18, we model the emission in each median image with a combination of three components: point-like emission from the central QG, an autocorrelation term due to source overlap within the sample, and extended emission from SFGs associated with the central QGs. This includes both true SF satellites of the QGs and the two-halo term arising from clustering with more massive SFGs. For simplicity, we hereafter refer to it as the `satellite halo'. Finally, we also include a flat background term. As the point-source and autocorrelation terms arise from the same galaxy population their amplitudes are tied and we determine their relative scaling beforehand. To model the combined central point-like emission and autocorrelation signal, we first created simulated images of the COSMOS field for each redshift bin, consisting of a blank map populated by model or empirical beams (depending on the data) located at the position of each QG and normalised to the same unitary flux. This latter approximation was used because the scatter of the FIR flux of our QGs is unknown by construction of the sample. Consequently, we do not associate an uncertainty with the shape of the autocorrelation signal. The simulated sources were then extracted using the same procedure as described above. This yields median cutouts that are essentially indistinguishable from point-sources in the case of higher-resolution data (e.g. at 24\,$\mu$m) but deviate substantially from it when the beam is large, due to the blending of sources. In particular, the SPIRE (250\,$\mu$m, 350\,$\mu$m, and 500\,$\mu$m) `point' stacks have therefore extended wing-like emission and higher peak values than a simple beam, while the 850\,$\mu$m and 1.1\,mm point stacks show an opposite effect, with lower peak values, arising from the deep negative rings in their PSFs. This additional step, which accounts for the intrinsic clustering of QGs, is necessary here since we did not eliminate projected pairs during selection, as in G18. Indeed, excluding from the sample all objects that partially overlap in the lowest-resolution maps (SPIRE) would more than halve it\footnote{As a sanity check we have validated that an analysis excluding the pairs yields consistent results albeit with significantly higher uncertainties.}.\\
\begin{figure*}
\centering
\includegraphics[scale=0.8]{f3.pdf}
\caption{\textbf{Observed median IR SEDs of massive QGs}. The magenta circles mark the observed photometry, with arrows indicating 3$\sigma$ upper limits for bands where the measured flux density was negative. We note that the radio points at 10 and 20\,cm were not included in the modelling. The black line and the grey shaded areas indicate the best-fit model from the \cite{draine_2007} templates and the associated uncertainty. The dashed orange line is the best fit adopting a single $T_{\rm d}$\ modified black body. The left, middle, and right panels show the results for the redshift intervals $0.3 < z < 0.65$, $0.65 < z < 1.0$, and $1.0 < z < 1.4$, respectively. Salient properties of the sample are reported in Table \ref{tab:properties}.
}
\label{fig:seds}
\end{figure*}
\begin{table*}
\small
\centering
\caption{Physical properties of the stacked quiescent samples}
\begin{tabular}{cccccccccc}
\hline\hline
Redshift& $N$& $\mathrm{log}(M_\ast)$& SFR(UV)$\dagger$& $\mathrm{log}(L_{\rm IR})$& $T_{\rm d}$& $\langle U \rangle$& $\mathrm{log}(M_{\rm dust})$ &$\mathrm{log}(M^{\rm sol}_{\rm gas})$ & $f^{sol}_{\rm gas}$ \\
& & ($M_{\odot}$)& ($M_{\odot}$ yr$^{-1}$)& ($L_{\odot}$)& (K)& & ($M_{\odot}$)&($M_{\odot}$)& (\%)\\
(1)& (2)& (3)& (4)& (5)& (6)& (7)& (8)& (9)& (10)\\
\hline
$0.30<z<0.65$& $563$& $11.17\pm0.15$& $1.7^{+0.3}_{-0.2}$& $9.40\pm0.05$& $20.0\pm4.0$& $0.6\pm0.3$& $7.51^{+0.20}_{-0.24}$& $9.46^{+0.20}_{-0.24}$& $2.0\pm1.0$\\
$0.65<z<1.00$& $1536$& $11.20\pm0.17$& $1.7^{+0.3}_{-0.2}$& $9.89\pm0.05$& $19.0\pm4.0$& $0.6\pm0.3$& $8.01^{+0.20}_{-0.25}$& $9.97^{+0.20}_{-0.25}$& $5.9\pm3.0$\\
$1.00<z<1.40$& $1394$& $11.16\pm0.15$& $1.9^{+0.4}_{-0.3}$& $10.20\pm0.05$& $22.0\pm3.0$& $1.1\pm0.4$& $8.05^{+0.20}_{-0.24}$& $10.08^{+0.20}_{-0.24}$& $7.3\pm3.0$\\
\hline
\hline
$1.50<z<2.20\ddagger$& $997$& $11.04\pm0.15$& $1.5^{+0.4}_{-0.4}$& $10.35\pm0.13$& $23.5 \pm 2.0$& $1.8 \pm 1.2 $& $8.00^{+0.16}_{-0.19}$&$9.96^{+0.16}_{-0.19}$& $8.4\pm3.0$\\
\hline
\end{tabular}
\tablefoot{Column 1: Redshift bin. Column 2: Number of objects in the redshift bin. Column 3: mean stellar mass. Column 4: mean SFR(UV). Column 5: total infrared luminosity from the SED fitting with \cite{draine_2007} models integrated over $8-1000\,\mu$m. Column 6: `luminosity-weighted' $T_{\rm d}$\ from the SED modelling with a single-temperature modified black body. Column 7: mean intensity of the radiation field from \cite{draine_2007} models. Column 8: dust mass. Column 9: molecular gas fraction assuming a solar gas-to-dust conversion factor of GDR\ = 92. All quantities are computed using a \citet{salpeter} IMF.\\
$\dagger$: Mean and interquartile range.\\
$\ddagger$: Stacked photometry from \citet{Gobat_2018} that for consistency was re-fitted here with the same method and setup as for the rest of the redshift bins.}
\label{tab:properties}
\end{table*}
\begin{table*}
\tiny
\centering
\caption{Stacked photometry of the QG population, as derived through flux density decomposition.}
\begin{tabular}{ccccccccccc}
\hline\hline
Redshift& $24\,\mu m$& $100\,\mu m$ & $160\,\mu m$ & $250\,\mu m$ & $350\,\mu m$& $500\,\mu m$ & $850\,\mu m$ & $1100\,\mu m$ & 3.0\,GHz & 1.4\,GHz \\
&[$\mu Jy$] &[$mJy$]& [$mJy$]& [$mJy$]& [$mJy$]&[$mJy$] & [$mJy$]&[$mJy$]&[$\mu Jy$]&[$\mu Jy$]\\
(1)& (2)& (3)& (4)& (5)& (6)& (7)& (8)& (9)& (10)&(11)\\
\hline
$\langle 0.50 \rangle$& $6.40\pm 1.47$& $0.12\pm0.04$& $0.34\pm0.08$& $0.32\pm0.21$& $0.06\pm0.23$& $0.12\pm0.23$& $0.1\pm0.04$& $<0.225^{\dagger}$& $1.65\pm0.12$& $6.84\pm0.63$\\
$\langle 0.85 \rangle$& $5.82\pm 0.61$& $<0.08^{\dagger}$& $0.08\pm0.04$& $0.47\pm0.22$& $0.80\pm0.21$& $0.64\pm0.19$& $<0.21^{\dagger}$& $0.04\pm0.04$& $1.20\pm0.07$&$2.60\pm0.41$\\
$\langle 1.20 \rangle$& $5.60\pm 0.79$& $0.30\pm0.30$& $0.11\pm0.05$& $0.57\pm0.13$& $0.77\pm0.14$& $0.59\pm0.14$& $0.04\pm0.03$& $0.06\pm0.04$& $1.09\pm0.07$ & $3.53\pm0.43$\\
\hline
\hline
\end{tabular}
\tablefoot{$\dagger$: $3\sigma$ upper limits following the prescription of \citet{Bethermin_20}}
\label{tab:photometry}
\end{table*}
On the other hand, the distribution of FIR emission from satellite halos is estimated using multi-wavelength catalogs available for the COSMOS field \citep{muzzin_2013,laigle_2016}. Satellites were selected in the catalogue as objects that both satisfy the $UVJ$ criterion \citep{williams_2009} for SFGs and have compatible photometric redshifts, namely $z_{\text{L}68}<z_{\text{cen}}<z_{\text{H}68}$, where $z_{\text{cen}}$ is the redshift of the central QG and $z_{*68}$ are the lower and upper 68\% confidence limits to the redshift of satellites, respectively. We estimated the SFR of each satellite candidate by fitting its rest-frame ultraviolet SED with constant star formation history stellar population templates \citep{bruzual_2003}, with and without dust extinction. The difference between these two estimates yields an obscured SFR surface density that, when convolved with the instrumental beam in each band, is assumed to be proportional to the FIR flux distribution from the satellite halo. More details are given in Appendix ~\ref{app:satellites}.\\
To extract the FIR SED of the QGs in each redshift bin, we fitted the cutouts in each band with combinations of these two components plus a flat background term. We let the scaling of the satellite emission vary freely at 24\,$\mu$m, 100\,$\mu$m, and 160\,$\mu$m, where the resolution is high enough to properly decompose the observed signal into point-like and extended components.
Satellite fluxes in these three bands were then fitted with a set of redshift-dependent MS IR templates \citep{magdis_2012} that capture the average value and the scatter of $T_{\rm d}$\ and $\langle U \rangle$ of MS galaxies at a given redshift \citep[e.g.][]{magdis_2012,bethermin_2015}. From these templates (and their scatter) we extrapolated satellite fluxes (and their uncertainties) at $\geq$ 250\,$\mu$m, where both the point-like and halo emission are extended and the large beams preclude a non-degenerate decomposition. For these bands, the scaling of the satellite model was fixed to the extrapolated flux, with only the background and central source being fitted. The satellite fluxes thus removed from the FIR SED correspond to $\mathrm{log}(L_{\rm IR}/L_{\odot}) = 10.40\pm0.09, 10.70\pm0.07, {\rm and}~9.30\pm0.10$~for the high-, intermediate-, and low-redshift bins, respectively. These estimates are consistent with the predicted $L_{\rm IR}$\ infrared from the obscured SFR of the satellites as inferred by the UV/optical modelling ($\mathrm{log}(L_{\rm IR}/L_{\odot}) = 10.47\pm0.13, 10.77\pm0.09, {\rm and}~9.70\pm0.38$). Finally, we note that while these $L_{\rm IR}$ values are comparable to or higher than the $L_{\rm IR}$\ of QGs, the sum of the autocorrelation and the QG component still dominates the $\gtrsim$ 250\,$\mu$m bands due to a combination of lower $T_{\rm d}$\ and large autocorrelation term.
The fits of the various components were done with pixel weights derived from the stacked variance maps described at the beginning of this section. However, these cannot be used to estimate accurate uncertainties on the extracted fluxes as the pixel-to-pixel variance is necessarily correlated in the FIR due to confusion. Therefore, rather than considering the (small) formal errors on scaling of the combined point-source and autocorrelation models, uncertainties on central fluxes were estimated by fitting instrumental beams at random positions in the cutouts, well clear of both the central source and satellite emission. To these we added in quadrature the error of both the satellite fluxes and background, in the corresponding band. We note that the autocorrelation term alone is free of uncertainty by construction, the sample being set and the beam models in $\geq$250\,$\mu$m bands (where it is not negligible) being theoretical and provided without associated errors.
The uncertainties for each decomposition component are presented in Table ~\ref{tab:errors}.
Finally, the measured central QG fluxes were rescaled down to account for contamination due to random clustering in the image using the corrections described in \citet{bethermin_2015}. We note that the background term might not necessarily be flat, due to lensing from the QGs' host dark matter halo. Magnification bias is typically dominated by area increase, so its net result is a deficit in faint source counts with respect to the field. Assuming that the QGs are central to their halo, this should lead to a lower effective background towards the centre of the cutouts, thus slightly underestimated central fluxes. However, its magnitude depends on the shape of the luminosity distribution of background sources, for which we have little to no information. Furthermore, this effect can be mild even for massive galaxy clusters \citep[e.g.][]{Umetsu_2014}. In this case, where the host halos are smaller by a factor of $\sim$10 (G18), we do not expect it to be significant. The final photometry of our stacked QGs is presented in Table \ref{tab:photometry} while the various decomposition components for each redshift bin and each broadband photometry are presented in Appendix~\ref{app:satellites}.
\section{SED modelling}\label{sec:modeling}
To derive the FIR properties of the stacked samples, we employed the dust models of \citet{draine_2007} (DL07), with the standard
parameterisation used in previous studies \citep[e.g.][]{magdis_2012}. Namely, we considered diffuse ISM models with radiation
field intensities ranging from 0.1 $\leq U_{\rm min} \leq $ 50, while fixing the maximum radiation field to $U_{\rm max} = 10^{6}$. We allowed for $
\gamma$ (the fraction of dust exposed to starlight with intensities ranging from $U_{\rm min}$ to $U_{\rm max}$) to vary between
$0 \leq \gamma \leq 0.5$ with a step of 0.01 and adopted Milky Way dust models with $q_{PAH}$ (i.e. the fraction of the dust mass in the form of PAH
grains) from 0.4\% to 4.6\%. To account for the artificial SED broadening due to the redshift distribution of the stacked sources, we constructed an average SED for each DL07 model by co-adding each original DL07 model at each redshift within the bin, weighted by the redshift distribution of stacked sources. The redshift-weighted, `smoothed' models were shifted to the median redshift of each bin and were fit to the data, yielding best-fit $L_{\rm IR}$, $M_{\rm dust}$, and $\langle U \rangle$\ values.
To estimate the uncertainties of the derived parameters we bootstrapped and fitted with same methodology 1000 artificial realisations of the real SEDs, by perturbing the extracted photometry within 1$\sigma$ for the detections (S/N$\geq$3) and within flux $+3\sigma$ for the rest, following the methodology presented in \citet{Bethermin_20}. As uncertainty for each derived parameter from the real data, we adopted the standard deviation of the values of the corresponding parameter inferred from the perturbed SEDs. As a sanity check we also introduced a bootstrapping component in our analysis by randomly excluding one observed data point at each realisation. We also confirm that the data sets of available detections (which include at least one data point at $\lambda_{\rm rest} \geq 220\,\mu$m for all four stacked ensembles, as well as a detection at $24\,\mu$m and at least one detection in the Wien part of the SED), along with the upper limits, ensure against possible systematic effects in the derivation of $M_{\rm dust}$\ and $L_{\rm IR}$\, for all redshift bins (see Appendix~\ref{app:firsimulation}). Finally, to derive `luminosity-weighted' $T_{\rm d}$\ we complemented our analysis by considering a single-temperature optically thin modified blackbody (MBB) models with a fixed effective dust emissivity index of $\beta = 1.8$.
The best fit models to the data (24\,$\mu$m to 1.1\,mm for DL07 and $\lambda_{\rm rest} > $ 50\,$\mu$m to 1.1\,mm for MBB), along with the range of SEDs from the random
realisations, are shown in Figure \ref{fig:seds}. The best fit parameters and their corresponding uncertainties are summarised in Table~\ref{tab:properties}. While the radio data points are not included in the fit, we add to each the best-fit DL07 model a power-law radio slope with a spectral index $\alpha = 0.8$ and a normalisation given by the FIR–radio correlation
\citep[e.g.][]{Delhaize17}, to investigate (and visualise) possible radio excess in our stacked ensembles.
\section{Results}
Using the derived parameters from the SED modelling in the previous section, we attempted to characterise the dust and gas content of QGs and investigate possible evolutionary trends. For the rest of the analysis we will be referring to the $\langle z \rangle = 0.50, 0.85,$ and $1.20$ stacks as low$-z$, mid$-z$, and high$-z$ quiescent sample respectively. We also complemented our study with the stacked ensemble of G18 at $\langle z \rangle =1.75$, which has been constructed and analysed using similar sample selection, stacking, photometry, and SED modelling techniques as those adopted here. Overall, this work covers the average properties of QGs in the redshift range $0.3 < z < 2.0$ with a homogeneous and systematic approach.
\begin{figure*}
\centering
\includegraphics[scale=0.83]{f4.pdf}
\caption{\textbf{Evolution of the dust to stellar mass ratio and of the gas fraction of QGs}. The filled blue circles mark the $M_{\rm dust}$/$M_{\ast}$\ of the stacked QGs in the $0.3 < z < 2.0$ range, including the results from G18. The grey shaded region depicts the trend for MS galaxies from Kokorev et al.\,(2021) in prep. The orange shaded region captures the range of $M_{\rm dust}$/$M_{\ast}$\ for local QGs, as drawn from the literature (see text). The dotted blue line (blue shaded region) corresponds to the best fit (scatter) to the $0.0 < z < 1.0$ $M_{\rm dust}$/$M_{\ast}$\ data of QGs, with a functional form of $M_{\rm dust}$/$M_{\ast}$\ $\propto (1+z)^{5.0}$. The bold, dashed purple line (shaded region) depicts the evolution (scatter) of $M_{\rm dust}$/$M_{\ast}$\ (and of $f_{\rm gas}$) as derived from the progenitor bias analysis of \citet{Gobat_20}, which is discussed in Section 5.2 and Section 6. The conversion of $M_{\rm dust}$/$M_{\ast}$\ to $f_{\rm gas}$\ assumes solar metallicity and GDR\ = 92.}
\label{fig:mdms}
\end{figure*}
\subsection{Basic FIR properties}
The available mid-IR to millimetre photometry that spans from observed 24- to 1100\,$\mu$m allows for a robust determination of the total 8-1000\,$\mu$m infrared
luminosities ($L_{\rm IR}$), dust masses ($M_{\rm dust}$), mean radiation field
(equivalent of dust mass-weighted luminosity or temperature), parametrised as $\langle U
\rangle \propto$ $M_{\rm dust}$/$L_{\rm IR}$, and luminosity-weighted dust
temperatures ($T_{\rm d}$). Our SED fitting yields low infrared
outputs for our stacked ensembles at all redshifts with
$\mathrm{log}(L_{\rm IR}/L_{\odot}) = 9.40 - 10.30$, dust
masses in the range of
$\mathrm{log}(M_{\rm dust}/M_{\odot}) = 7.50 - 8.15$, low mean
radiation field $\langle U \rangle = 0.6-1.2$, and cold $T_{\rm d}$\ $< 23$\,K dust temperatures.
In comparison to previous studies that presented stacked SEDs of QGs at similar redshifts and equivalent $M_{\ast}$\ bins, our analysis yields significantly lower $L_{\rm IR}$\ and $M_{\rm dust}$\ estimates. For example, fitting the photometry presented in \cite{Man16} with same methodology as for our data, we find $\mathrm{log}(L_{\rm IR}/L_{\odot}) = 9.8-11.2$ and $\mathrm{log}(M_{\rm dust}/M_{\odot}) = 8.0-8.6$ for the stacked ensembles of QGs at $z = 0.4 - 1.6$. These values are $0.2-0.9$\,dex larger than the corresponding parameters inferred from our analysis for our sample.
These discrepancies stem from the more conservative sample selection that we adopted in this work (double-colour criterion and exclusion of individually detected 24\,$\mu$m sources). In addition, the adopted stacking and, especially, the treatment and correction for the contribution of blended satellites further affects the comparison with \citet[see also G18]{Man16}. Nevertheless, and in agreement with \citet{Man16}, we do find a radio excess between the observed radio fluxes and those inferred by the infrared-radio correlation of SFGs by \citet{Delhaize17}, indicative of AGN activity and `radio-mode' feedback in massive QGs \citep[e.g][]{Gobat_2018,barisc17}.
In order to test for the possible sensitivity of these recovered parameters to the selection criteria described in Section~\ref{sec:selection}, as well as possible contamination of quiescent samples by dusty star forming objects, we performed a series of tests using sub-samples obtained by either applying an additional 0.1\,mag colour margin or by dividing the low-, mid-, and high-$z$ bins into blue and red halves. Applying the same stacking, source extraction, and modelling procedures, as described in Sections~\ref{sec:method} and \ref{sec:modeling}, to these yields SEDs that are entirely consistent with the full low-, mid-, and high-$z$ samples. This suggests that our analysis is robust with respect to selection effects. More details are given in Appendix~\ref{app:subsamples}.
Finally, we inferred gas masses ($M_{\rm gas}$) estimates by converting the derived $M_{\rm dust}$\ to $M_{\rm gas}$\ through the metallicity-dependent dust to gas mass ratio () equation of \citet{magdis_2012}; $\rm log(GDR) = 10.54 - 0.99\times(12+log(O/H))$. We considered a range of metallicities $Z_{\odot} \leq Z \leq 2Z_{\odot}$ (8.66 $\leq$ 12$+$log(O/H) $\leq$ 9.1 in the \citealt{Pettini04} scale), appropriate for the SFR and $M_{\ast}$\ of our samples \citep[e.g.][]{Mannucci10}. For these metallicities the GDR varies between $35 \leq$ $M_{\rm gas}$/$M_{\rm dust}$\ $\leq 92$. For the rest of our analysis we use the $M_{\rm gas}$\ estimates based on a solar metallicity as benchmark and those for a super-solar metallicity as lower limits. Therefore, throughout the paper, $M_{\rm gas}$\ = $M^{\rm sol}_{\rm gas}$ (and $f_{\rm gas}$\ = $f^{\rm sol}_{\rm gas}$), unless otherwise stated. We note that since the $M_{\rm dust}$/$M_{\rm gas}$ -- $Z$ relation of \citet{magdis_2012} is almost linear (slope of 0.99), we can easily rescale $M_{\rm gas}$\ estimates between solar and an arbitrary metallicity Y by using:
\begin{equation}
\mathrm{log}[M^{Y}_{\rm gas}] = \mathrm{log}[M^{\rm sol}_{\rm gas}] + [8.66 - Y],\,\,\,\, Y = 12 + \mathrm{log(O/H)} .
\end{equation}
We also recall that the GDR -- Z relation yields $M_{\rm gas}$\ estimates that trace the sum of the atomic ($M_{\rm HI}$) and the molecular ($M_{\rm
H_2}$) gas mass (including a $\times$1.36 He contribution). As such, $M_{\rm gas}$\ = $M_{\rm HI}$ + $M_{\rm H_{2}}$.
While the exact state of the cold gas in our QGs cannot be
characterised, the $\langle M_{\rm H2}/M_{\rm HI} \rangle > 1$ ratios
reported in the literature for $z > 0.5$ and for high stellar surface densities \citep[e.g.][]{Blitz06, Bigiel08, Obreschkow09} motivate us to make the simplified assumption that $M_{\rm H_{2}} >> M_{\rm HI}$ and, thus, $M_{\rm H_{2}} \approx$ $M_{\rm gas}$. We note though,
that this estimate is an upper limit to the molecular gas mass of our samples.
\begin{figure*}[!t]
\centering
\includegraphics[scale=1.1]{f5.pdf}
\caption{\textbf{Gas depletion timescales}. \textit{Left}: Gas depletion time scale ($t_{\rm dep}$) as a function of SFR. The blue (grey) circles correspond to the stacked samples of QGs for the case of solar (super
solar) metallicity, assuming GDR = 92 (35). The dashed grey line and the grey shaded region depict the trend for MS galaxies from \citet{Sargent14} with a scatter of $0.23$\,dex. The magenta shaded region marks the area with $\tau_{\rm dep}$\ $<$ 100Myrs, typical of star-bursting galaxies. \textit{Right}: Gas depletion time scale as a function of redshift. The symbols are the same as in the left panel. The dashed grey line and the grey shaded region depict the trend of \citet{Liu19} for MS galaxies at fixed $M_{\ast}$\ = $2\times10^{11}$$\rm M_{\odot}$\ and its corresponding scatter of $0.23$\,dex.}
\label{fig:tdep}
\end{figure*}
\subsection{Evolution of dust and gas mass fractions}
In Figure \ref{fig:mdms} we bring together the $M_{\rm dust}$\ estimates for our stacked ensembles as well as a range of literature $M_{\rm dust}$\ estimates for local QGs of comparable $M_{\ast}$\ \citep{Smith12, Boselli14, Lianou16, Michal19} and explore the evolution of dust mass fraction, $M_{\rm dust}$/$M_{\ast}$, as a function of redshift. Our data indicate that the $M_{\rm dust}$/$M_{\ast}$\ remains roughly constant from $z=2.0$ to $z=1.0$, followed by sharp decline towards the present day, with a functional form of $M_{\rm dust}$/$M_{\ast}$\ $\propto (1+z)^{5.00^{+1.94}_{-1.55}}$ (for $z < 1.0$). As shown in Figure \ref{fig:mdms}, a similar trend characterises the evolution of the dust mass fraction in SFGs (Kokorev et al.\,2020 in prep.), albeit with a shallower slope. Evidently, QGs exhibit systematically lower dust mass fractions with respect to SFGs at all redshifts, indicative of dust consumption, destruction or expulsion during the quenching phase. However, the flat evolution for QGs between $z = 2.0 - 1.0$ and the monotonically decreasing $M_{\rm dust}$/$M_{\ast}$\ for SFGs at all redshift, results in a minimum deviation between the two populations at $z \sim 1.0$.
The evolution of $M_{\rm dust}$/$M_{\ast}$\ we find here also mirrors the evolution of the gas fraction of QGs. Under the assumption of constant solar metallicity for massive QGs at all redshifts, we can directly convert the $M_{\rm dust}$/$M_{\ast}$\ presented in the left y-axis of Figure \ref{fig:mdms} to $f_{\rm gas}$. This approach yields a constant $f_{\rm gas}$\ $\approx 7-8\%$ between $z = 2.0 - 1.0$ that subsequently drops to $f_{\rm gas}$\ $\approx 2\%$ at $z = 0.5$ and $f_{\rm gas}$\ $\leq 1\%$ in the local Universe. These values are in agreement with independent $M_{\rm gas}$\ estimates or upper limits inferred by CO observations of individual QGs and post-starburst galaxies at various redshifts \citep[e.g.][]{French15, Sargent15, Rudnick17, Suess17, Spilker18, Bezanson19}. We note that for a super solar metallicity of $2\times Z_{\odot}$ the $f_{\rm gas}$\ values presented in Figure \ref{fig:mdms} would correspond to $f^{\rm Z}_{\rm gas} \approx 0.4 \times f^{\rm sol}_{\rm gas}$. In Section 6, we will interpret the observed redshift evolution of $M_{\rm dust}$/$M_{\ast}$\ and $f_{\rm gas}$\ of QGs in terms of progenitor bias and ageing of the stellar population.
\subsection{SFR and gas depletion time scales }
The gas masses inferred in the previous section, when coupled with the corresponding SFRs of the stacked ensembles, can be used to define the
star formation efficiency (SFE = SFR/$M_{\rm gas}$) or, inversely, the gas depletion time scale ($\tau_{\rm dep}$\
$=$ 1/SFE). To derive the total SFR in each redshift bin, we
considered the average unobscured component as derived from the optical photometry (SFR$_{\rm
UV}$) and the obscured star formation by converting $L_{\rm IR}$\ to SFR$_{\rm IR}
$ using the \citet{Kennicutt98} scaling relation. We caution the reader that for QGs a
considerable fraction of $L_{\rm IR}$\ could arise from dust heated by old rather than
young stellar populations \citep[e.g.][]{Hayward_14}, as assumed by the Kennicutt SFR-$L_{\rm IR}$\ relation. Therefore,
the estimates of SFR$_{\rm tot}$ = SFR$_{\rm IR}$ + SFR$_{\rm UV}$ should be regarded as upper
limits. To constrain the fraction of $L_{\rm IR}$\ arising from dust heated by old stars, we fit
each galaxy's optical-NIR SED with composite stellar population models derived from
\citet{bruzual_2003} templates. For simplicity, we used delayed exponential star formation
histories, starting at $z_{\rm in}=10$ and with time scales $\tau=100\,\rm Myr-3\,\rm Gyr$, each
truncated 1\,Gyr before observation. Finally, we assumed a \citet{noll_2009} attenuation curve
with variable slope and computed the luminosity difference of the best-fit model before and
after attenuation. We find that this absorbed energy does not make up more than $\sim$30\%
of $L_{\rm IR}$\ at $z\gtrsim1$ but could represent up to 100\% of the infrared luminosity of $z\sim0.5$
QGs. Given the uncertainties and the small IR output indicated by our stacks, we chose to consistently consider SFR$_{\rm tot}$ at all redshift bins.
The emerging (maximum) SFRs are listed in Table~\ref{tab:properties} and they range from $\sim$2.0\,$\rm M_{\odot}\,yr^{-1}$\ in the lowest redshift bin ($\langle z \rangle = 0.5$) to $\sim$6.5\,$\rm M_{\odot}\,yr^{-1}$\ at $\langle z \rangle = 1.8$. These average SFRs place our stacked ensembles $\geq$10$\times$ below the MS at their corresponding
redshifts, confirming their passive nature. At the same time, $\tau_{\rm dep}$, ranges from 1.3 to 3.0\,Gyr for the case of solar metallicity with a mean $\langle$$\tau_{\rm dep}$$\rangle$ = $2.1\pm0.8$\,Gyr and 0.5 to 1.1\,Gyr for a super solar metallicity. These $\tau_{\rm dep}$\ estimates are plotted as a function of SFR and redshift in the two panels of Figure \ref{fig:tdep} along with the corresponding tracks for MS galaxies \citep{Sargent14,Liu19}. We stress again that since we are using the maximum SFR estimates the quoted $\tau_{\rm dep}$\ should be regarded as lower limits for each metallicity. In agreement with \citet{Spilker18}, we find that distant QGs, for their SFRs, have a $\tau_{\rm dep}$\ that is broadly consistent with that of local MS galaxies, but longer than that of MS galaxies at their corresponding redshift. We note that while our data hint upon an increase in $\tau_{\rm dep}$\ from $z=2.0$ to $z=0.8$ followed by a decline towards later cosmic epochs (Figure \ref{fig:tdep}, right), the values are overlapping within their uncertainties yielding a non significant correlation. Possible metallicity effects as well as the uncertainties linked to the conversion of $L_{\rm IR}$\ to SFR, will be investigated in detail in a future study.
\subsection{SED and $T_{\rm d}$\ evolution }
The star formation activity and the dust heating photons per unit ISM mass, both parametrised above in terms of SFR, $M_{\rm dust}$, and SFE, should also be imprinted in the shape of the FIR SED and the characteristic $T_{\rm d}$\ of QGs. Indeed, a large body of literature that has focused on the evolution of $T_{\rm d}$\ as a function of redshift for MS and star-bursting galaxies alike \citep[e.g.][]{Hwang10, magdis_2012,magdis17, Magnelli14,schreiber_2018_dust, Jin19, Cortzen20} has established that the increase in sSFR and SFE at higher redshifts is also followed by a trend of warmer SEDs and increasing $T_{\rm d}$\ at least out to $z\sim4$. At the same time, SEDs and the effective $T_{\rm d}$\ of distant QGs as a function of redshift are far less explored.
In Figure \ref{fig:temp} we plot the $T_{\rm d}$\ of QGs derived from our MBB fits as a function of redshift, along with literature data and trends for MS galaxies. As expected, QGs exhibit colder $T_{\rm d}$\ with respect to star forming galaxies at all redshifts, indicative of a lower star formation activity. Also, at first reading, the $T_{\rm d}$\ estimates for our quiescent ensembles are overlapping within their uncertainties, with an average $\langle T_{\rm d} \rangle = 21.0 \pm 2.0$\,K. Notably, in the redshift range probed by our data, SFGs have witnessed a decrease in their $T_{\rm d}$\ by $\sim8$\,K, indicative of a steep evolution that is not observed for QGs.
However, the hint of small $T_{\rm d}$\ decrease in QGs between $z = 2.0$ and $z = 0.3$ ($\Delta$$T_{\rm d}$\ $\approx
3$\,K), even if not statistically vigorous, provides the grounds for a tempting speculation of a weak
evolution driven by their size evolution. Indeed, a size evolution of $R \propto (1+z)^{-1.48}$
\citep{vdw14} yields a size increase of a factor of $\sim0.5$ between $z = 1.8$ and 0.8. Similarly,
following the analysis of \citet{Chanial07}, $T^{4+\beta}_{\rm d}$ $\propto$ $[\Sigma_{\rm
IR}]^{\gamma}$, where $\gamma=\frac{0.4}{1.4}$ and $\Sigma_{\rm IR} \propto L_{\rm IR}/R^{\rm 2}$.
For fixed $L_{\rm IR}$\ (or SFR), as in the case for our stacked samples in this redshift range, it follows that
$T_{\rm d}(z_1)$/$T_{\rm d}(z_2)$ $
\propto [R(z_2)/R(z_1)]^{2 \times \gamma/(4+\beta)}$. Anchoring this equation at $T_{\rm d}(z=1.8) = 23.5$\,K, as inferred by our analysis and adopting $\beta=1.8,$ provides an evolutionary track that is in broad agreement with our measurements for the $0.3 < z < 2.0$ stacks. We conclude that a weak (if any) evolution of $T_{\rm d}$\ in QGs can be understood in terms of size growth. At the same time an average $\langle T_{\rm d} \rangle = 21.0 \pm 2.0$\,K provides a statistically robust estimate for a characteristic $T_{\rm d}$\ for the ISM of QGs over the last ten billion years.
In addition, the small variance in $T_{\rm d}$\ and $L_{\rm IR}$/$M_{\rm dust}$, suggests that massive QGs at various redshifts share a common FIR SED shape, albeit with a varying normalisation that mirrors the evolution of $L_{\rm IR}$\ and of the ISM mass budget as described in the previous sections. To produce an average SED, we bring the stacked photometry of each redshift bin at rest-frame, normalise it so that all SEDs are anchored at the same reference luminosity of $1.0 \times 10^{10}$\,L$_{\odot}$\ (or SFR $\approx 1.7$\,$\rm M_{\odot}\,yr^{-1}$), and fit with DL07 and MBB models following the methodology presented in Section 4. The observed points (and upper limits) along with the best-fit models are presented Figure \ref{fig:temp}\,(right). The emerging average SED of QGs has $T_{\rm d}$\ $= 20$\,K, and $L_{\rm IR}$/$M_{\rm dust}$\ $\approx 90\,$L$_{\odot}$/$\rm M_{\odot}$, in direct contrast to that of typical $z > 0$ star forming galaxies. As an example, we consider the average SED of $z = 1.0$ MS galaxies from \citet{magdis_2012}, which clearly peaks at shorter $\lambda$ with a $T_{\rm d}$\ $\approx$ 30\,K and has approximately a factor of $\sim10\times$ larger infrared energy output per unit dust mass ($L_{\rm IR}$/$M_{\rm dust}$\ $\approx$ 1000L$_{\odot}$/$\rm M_{\odot}$). Finally, we stress that the radio data are not included in the fit but are presented to demonstrate the radio-excess that has already briefly discussed in a previous section. The average SED of QGs (normalised to $L_{\rm IR}$\ = $1.0 \times 10^{10}$\,L$_{\odot}$\, and $M_{\rm dust}$\ = 1.11 $\times$ 10$^{8}$\,$\rm M_{\odot}$) is publicly available{\footnote{http://www.georgiosmagdis.com/software/}}.
\begin{figure*}
\centering
\includegraphics[scale=1.1]{f6.pdf}
\caption{\textbf{Dust temperatures and average SED}. \textit{Left}: The evolution of $T_{\rm d}$\ in
QGs. The filled blue circles mark the luminosity-weighted $T_{\rm d}$\ for the stacked QGs at $0.3 < z < 2.0$, including the results from G18. Filled grey triangles and squares indicate SFGs on the MS and strong starbursts from \cite{bethermin_2015}. Filled and open grey circles represent the mass-averaged values and the most massive galaxies ($11.0 < \mathrm{log}(M_\ast/M_\odot) < 11.5$, Salpeter IMF) for the star-forming sample of \cite{schreiber_2018_dust}, respectively. The solid and dashed grey lines indicate the best fit models from \cite{schreiber_2018_dust} and \cite{bethermin_2015}, extending the parameterisation of \cite{magdis_2012}. The dotted grey line indicates the expected $T_{\rm d}$\ track for a size evolution of QGs, $R \propto (1+z)^{-1.48}$. The track is normalised to the $T_{\rm d}$\ of the $z = 1.8$ sample. \textit{Right}: Composite and template FIR SED for $z > 0$ QGs. The magenta circles (arrows) correspond to the photometry (3$\sigma$ upper limits) of the stacked ensembles presented in Figure~\ref{fig:seds} normalised to $L_{\rm IR}$\ = 10$^{10}$\,L$_{\odot}$\ (and $M_{\rm dust}$ $= 1.11\times10^{8}$\,$\rm M_{\odot}$). The DL07 and MBB model fits to the SED are shown with grey and red dashed lines. The dotted blue line correspond to the typical SED of $z = 1.0$ MS galaxies from \citet{magdis_2012}, normalised to the same $L_{\rm IR}$. The QG template is publicly available at \href{www.georgiosmagdis.com/software/}{www.georgiosmagdis.com/software/}.}
\label{fig:temp}
\end{figure*}
\section{Discussion}
In the previous sections, we characterised the mass budget and the properties of the ISM of QGs at various redshifts. Here we offer an explanation for the origin and the drivers of the observed trends.
To interpret the evolution of $M_{\rm dust}$\ in our quiescent sample, we first need to consider the expected redshift evolution of a population of QGs. At log($M_{\ast}$/$M_{\rm\odot}$) $\sim 11.0$, their co-moving number density increases significantly between $z \sim 4$ and $z \sim 1$, with little evolution after that \citep[e.g.][]{Muzzin13,Davidzon17}. This implies that an unbiased sample of QGs will contain a large fraction of newly formed objects at any redshift $z \gtrsim 1$ while, few new objects enter the $z < 1$ population. For simplicity, let us neglect dry merging events and equate the formation time of a QG with the quenching time of its precursor. In this case, the average age of the quenching event will vary little between quiescent galaxy samples at $z\gtrsim1$ and increase with cosmic time at $z\lesssim1$. In this context, the apparent non-evolution of the dust fraction between $z\sim2$ and $z \sim1 $, where newly quenched galaxies constitute a large part of the quiescent population, suggests that whatever quenching mechanisms are at play, they leave behind very similar ISM conditions independently of redshift. Conversely, the steep decrease in the dust fraction between $z \sim 1$ and $z = 0$ is consistent with a consumption (or destruction) without replenishment of the dust and gas as few to no new QGs enter the massive population in this redshift range. Combining both regimes yields an apparent evolution very similar to our observed constraints, as shown in Figure~\ref{fig:mdms}. Here we estimated the median gas fraction of the whole population of QGs as a function of redshift by assuming only the QG production rate from stellar mass functions, an initial gas fraction (constant with redshift) of $\sim$10\% immediately following quenching, and a closed-box consumption with a time scale of $\sim$2\,Gyr after that \citep{Gobat_20}. The track derived in this manner reproduces our data well, especially when considering that the latter are likely slightly biased towards dustier objects at lower redshift (see Figure~\ref{fig:24um_detections}), which would flatten the observed trend.
This `progenitor-bias' explanation for the observed evolution of $f_{\rm gas}$\ and $M_{\rm dust}$/$M_{\ast}$\ would also require an anti-correlation between these quantities and the average quenching time, or more simply, the stellar age. Indeed, in a recent study, \citet{Michal19}, presented an exponential decline of the $M_{\rm dust}$/$M_{\ast}$\ with age for a sample of galaxies at a fixed cosmic epoch ($\langle z \rangle = 0.13$) and a range of luminosity-weighted stellar ages between $\rm 9.0 < log(age/yr) < 10$. The authors report a decline of $M_{\rm dust}$/$M_{\ast}$\ with age that follows $e^{\rm -age/\tau}$ and a dust lifetime $\tau \approx$ 2.5\,Gyr. While we do not have any information about the actual stellar age of the galaxies in our stacks, the effect of varying age can be mimicked by the our different redshift bins and under the assumption that the passive galaxies at $z > 0.9$ will, on average, remain passive at later times. Indeed the time interval between $z = 0.9$ and $z = 0.5$ corresponds to $\approx$2.3\,Gyr. According to the formula of \citet{Michal19}, and assuming an average stellar age of $2 - 3$\,Gyr for the QGs at $z = 0.9$, we should expect a drop in log($M_{\rm dust}$/$M_{\ast}$) by $\sim0.4-0.6$\,dex, in very good agreement with our observations. At the same time, as described above, the flat behaviour of the $M_{\rm dust}$/$M_{\ast}$\ (and $f_{\rm gas}$) at $z > 1.0$ could then be explained by the younger, and roughly constant stellar age of the QGs in the higher redshift bins (e.g. log[Age/yr] = 9.0 at $z = 2.0 - 4.0$, \citet{toft17,V20}).
Finally, the rather constant FIR SED shape (and $T_{\rm d}$) of QGs at all redshifts, indicate that $L_{\rm IR}$\ and $M_{\rm dust}$\ roughly scale together irrespective of the time of quenching. If indeed $L_{\rm IR}$\ is a proxy of SFR and $M_{\rm dust}$\ a proxy of $M_{\rm gas}$\ this means that the decrease in $M_{\rm gas}$\ is closely followed by a decrease in SFR. In this case, the universality of the initial ISM conditions, as advocated by the evolution of ISM mass budget (in terms of $f_{\rm gas}$\ and $M_{\rm dust}$/$M_{\ast}$), is followed by a regularity in the internal processes within quenched galaxies.
We should stress that while the progenitor bias offers an appealing empirical explanation for the observed evolution of the ISM properties of QGs, it bares little, if any, information about the underlying physics of the origin and the maintenance of
quenching. For the latter, the radio excess present in our SEDs as well as in previous studies \citep[e.g.][]{Man16, DEugenio20},
point towards `radio-mode' feedback from AGN-driven jets as compelling quenching mechanism that can both prevent gas accretion due to
heating of the halo gas and expel a fraction of the cold gas reservoir \citep[e.g.,][]{croton_2006, Gobat_2018}. This radio-mode
feedback would also need to be stronger at higher redshift to compensate for the higher (putative) gas fractions seen at these epochs. Using
the average 1.4\,GHz excess, with respect to the FIR-radio correlation, and the evolution of the AGN luminosity function \citep{Novak_2018},
we indeed estimate that the AGN duty cycle for our QGs decreases from $>$50\% at $z\sim1.8$ (see G18) to $\sim$15\% in the
low-$z$ redshift bin (see Appendix~\ref{app:dutycyle}).
\subsection{Outlook and ALMA predictions}
From the discussion above, it stems that the $M_{\rm dust}$\ (and subsequently $M_{\rm gas}$\ and $f_{\rm gas}$) of QGs depends both on redshift and on time elapsed since the quenching event. So while our $M_{\rm dust}$\ and $M_{\rm gas}$\ estimates should be valid for the average QGs population at a given redshift, this might not be the case for individual sources due to possible age variations. Clearly, the way forwards necessitates focused ISM studies for large and representative samples of individual QGs at various cosmic epochs. In this regard, the IR SED template presented in Section 5, can facilitate a series of useful predictions regarding the observability of their dust continuum emission.
The main characteristic of the template is the ratio of its native infrared luminosity ($L^{\rm 0}_{\rm IR}$) to its native dust mass ($M^{\rm 0}_{\rm dust}$), which is expressed as $L^{\rm 0}_{\rm IR}$/
$M^{\rm 0}_{\rm dust}= 90$ L$_{\odot}$/$\rm M_{\odot}$. Consequently, for any arbitrary normalisation of the template, it follows that for this template, $L_{\rm IR}$\ = 90$
\times$ $M_{\rm dust}$\ [in L$_{\odot}$]. Similarly, $M_{\rm gas}$, for a fixed metallicity $Z$, scales linearly with $M_{\rm dust}$, with $M_{\rm gas}$\ = GDR($Z$) $\times$ $M_{\rm dust}$\ \citep[e.g.][]{Leroy11, magdis_2012}. Therefore, the template can be rescaled to any arbitrary $M_{\rm dust}$\ that corresponds to a given $M_{\rm gas}$\ (or $f_{\rm gas}$). Indeed,
since $f_{\rm gas}$\ = $M_{\rm gas}$/$M_{\ast}$\ = GDR(Z) $\times$ $M_{\rm dust}$/$M_{\ast}$\, we simply have to
multiply the SED by $N=$ $M_{\rm dust}$/$M^{\rm 0}_{\rm dust}$, which takes the form of:
\begin{equation}
N = \frac{f_{gas} \times M_{\ast}}{GDR(Z)\times M^{\rm 0}_{\rm dust}} .
\end{equation}
Assuming that our SED is representative of massive QGs with a typical log($M_{\ast}$/$M_{\rm\odot}$) = 11.2 at all redshifts and by adopting a universal solar metallicity that corresponds to GDR(Z$_{\odot}$) = 92 \citep[e.g.][]{Leroy11, magdis_2012}, the only free parameter in the SED normalisation is $f_{\rm gas}$. Then we can use our normalised template SED to calculate the emitted flux density at the central wavelength of various ALMA bands at various redshifts for any adopted $f_{\rm gas}$\, after implementing CMB effects as prescribed in \citet{daCunha13}.
\begin{figure*}[!t]
\centering
\includegraphics[scale=0.85]{f7.pdf}
\caption{\textbf{ALMA outlook and systematics}. \textit{Left:} ALMA flux densities in various bands for an evolving $f_{\rm gas}$\ as $(1+z)^{5.0}$ at $0.0 < z < 1.0$ and fixed $f_{\rm gas}$\ = 8\%, at $z > 1.0$. The predicted fluxes are derived using the average DL07 SED presented in Figure \ref{fig:temp}(right) and are corrected for CMB effects. The blue shaded region depicts the range of predicted flux densities in ALMA Band 3 assuming a 50\% uncertainty in the adopted $f_{\rm gas}$. \textit{Middle:} Ratio of $M_{\rm gas}$\ estimates, as inferred for a fixed
monochromatic luminosity at a given rest-frame wavelength from 1) scaling the
average SED of QGs presented in Figure \ref{fig:temp}(right),
$M^{\rm SED}_{\rm gas}$, and 2) the prescription of \citet{Scoville_2017}, $M^{\rm Td=25K\, (mass-weighted)}_{\rm gas}$. Similarly, the right y-axis
depicts the ratio of the monochromatic luminosities
predicted by the two methods for a fixed $M_{\rm gas}$. \textit{Right:} The effect of the adopted $M_{\rm gas}$-method on the recovered evolution of $f_{\rm gas}$. Filled blue circles correspond to the values presented in Figure 4 using the full SED analysis ($f^{\rm SED}_{\rm gas}$). Filled grey squares correspond to the values derived from the \citet{Scoville_2017} prescription, assuming monochromatic observations at $\lambda_{\rm obs} = 850 \mu$m, ($f^{\rm 850}_{\rm gas}$). The empty points (and the colour-shaded region) are an extrapolation to higher redshifts under the assumption of fixed $f^{\rm SED}_{\rm gas}$ = 8\% at $z > 2.0$.}
\label{fig:alma}
\end{figure*}
As an example, we examine the case where $f_{\rm gas}$\ is evolving with redshift as $f_{\rm gas}$\ $\propto(1+z)^{5.0}$ at $z < 1.0$ and then remains flat with $f_{\rm gas}$\ = 8\% at $z \geq 1.0$ (Section 5). The flux density tracks as a function of redshift for various ALMA bands are presented in Figure \ref{fig:alma}(left). We stress that these tracks are specifically tailored around massive (log($M_{\ast}$/$M_{\rm\odot}$) = 11.20) QGs with a solar metallicity. However, the recipe described above is also valid for any arbitrary gas phase metallicity and any $M_{\ast}$\ provided that our template is also representative of less (or more) massive QGs.
We emphasise that the conversion of R-J flux densities ($f_{\rm ALMA}$) to $M_{\rm gas}$, and vice versa, is heavily dependent on the adopted method (and template). Indeed, at fixed $f_{\rm ALMA}$, the commonly adopted monochromatic prescription of \citet{Scoville_2017} yields systematically lower $M_{\rm gas}$\ estimates with respect to those inferred in here. As shown in Figure \ref{fig:alma}(middle) the tension between the two methods becomes more prominent at $\lambda_{\rm rest}$ closer to the peak of the SED and progressively eases moving further into the R-J tail. This is also demonstrated in Figure \ref{fig:alma}(right) where we show how applying the \citet{Scoville_2017} method at $\lambda_{\rm obs}= 850 \mu$m would affect the recovered evolution of $f_{\rm gas}$. We note that while we lack observations at $z > 2.0$ we extrapolate to higher redshifts under the assumption of flat $f_{\rm gas}$\ evolution at $z > 2$ to highlight how discrepancy (for fixed $\lambda_{\rm obs}$) between the two methods increases at higher redshifts.
The tension in the $M_{\rm gas}$\ (and $f_{\rm gas}$) estimates between the two methods, arises predominantly from the assumed $T_{\rm d}$\ of QGs. We recall that the \citet{Scoville_2017} recipe is calibrated on SFGs, assuming a fixed `mass-weighted' $T_{\rm d}$\ $=$ 25\,K. This corresponds to a $\sim$7-10\,K warmer `luminosity-weighted' $T_{\rm d}$\ compared to that of our template. Similarly, as discussed above, our QG template is $\sim$10\,K colder with a $\sim$10 times lower $L_{\rm IR}$/$M_{\rm dust}$\ compared to that of a typical SFG at $z = 1.0$.
It then naturally follows that with respect to the \citet{Scoville_2017} approach (or a $z >$ 0 SFG template) our method predicts lower $f_{\rm ALMA}$ and
subsequently a significantly longer integration time to detect dust emission from a QG of a given $M_{\rm gas}$\ (e.g. $\sim \times6$ for a galaxy at $z = 3$ and $
\lambda_{\rm obs}$ = 1200$\mu$m). This, could explain the scarce dust continuum detections among the limited number of high$-z$ QGs that have
been followed up with ALMA so far.
Another point of concern is selection biases, especially in terms of $M_{\ast}$. \citet{Williams_20}, reports no CO[2-1] detection for a sample of
massive $\langle log(M_{\ast}/M_{\odot}) \rangle \approx 11.7$ (for Salpeter IMF) QGs at $z \sim 1.5$, placing stringent upper limits in their $f_{\rm gas}$ $\leq$ 1.0-3.5\%, that are lower compared to our estimates. However, we note that this sample is $\sim$0.5-0.7dex more massive with respect to the stacked ensembles studied here. If the $f_{\rm gas}$\ of high-$z$ QGs decreases as a function of $M_{\ast}$, similarly to what is observed for SFGs and local ETGs \citep[$f_{\rm gas}$\ $\propto M^{-(0.40-0.55)}_{\ast}$ e.g.][]{Saintonge11a1, magdis_2012}, then rescaling our $f_{\rm gas}$\ estimates accordingly would significantly ease the tension \citep[see also Fig. 6 of][]{Williams_20}. Nevertheless, we caution the reader that 1) the dust method to infer $M_{\rm gas}$, when applied to QGs, could in principle recover a sizeable amount of
HI that is untraceable by the CO lines and 2) DL07 or similar dust emission models could be a poor representation of the dust grain composition of distant
QGs.
Finally, we discuss the impact of the CMB on the predicted flux densities, an effect that was recently observationally confirmed in a sample of `cold' $z > 3$ SFGs \citep{Jin19,Cortzen20}. Our analysis suggests that while $f_{\rm ALMA}$ should remain unaffected up to $z \sim 2.0$ in all ALMA bands, the `dimming' of the dust continuum emission that can be measured against the CMB is not negligable at higher redshifts and lower frequencies. This is demonstrated by the flat evolution of the flux density tracks in Bands 3 $-$ 6, that clearly deviate from the expected rising trend due to the negative $k-$correction (Figure \ref{fig:alma}left). Indeed, our analysis suggests that the recoverable continuum emission drops by a factor of 1.4 at $z$ = 3 - 5 for Band 5 and by a factor of $\sim$1.6-2.0 at $z =$ 4-5 for Band 3 observations. Admittedly, the analysis presented in this work does not extent beyond $z \sim 2$, and therefore the actual IR SED (and $T_{\rm d}$) of QGs at $z > 2$ is still unconstrained. However, if the template SED presented here is also valid at $z > 2.0$, CMB-corrections should be implemented both in the observing strategy and in the interpretation of dust continuum and spectral line observations of $z > 3.0$ QGs in the (sub)-mm.
\subsection{Caveats}
As pointed out in the previous sections, this analysis is not free of caveats which are inherent to the stacking method. First, decomposing the stacked signal into various components is a challenging task. Indeed, `satellite' fluxes in the SPIRE bands, and their corresponding uncertainties, depend on the adopted templates used for extrapolation. While we have considered an uncertainty on the SED of the SF component, our analysis does not account for spatially correlated variations of $T_{\rm dust}$. This would effectively make the satellite emission term frequency-dependent. On the other hand, the average stellar mass of SFGs in the satellite halo varies, with respect to their distance to the QGs, by less than a factor of 2. This suggests that the population of SF satellites is fairly homogeneous.
Furthermore, the combined photometry in the available bands can place constraints in the shape of the SED and in the inferred FIR properties, under the the assumption that the FIR emission of QGs is well-represented by the adopted DL07 models. This is of particular importance, especially since none of our stacked ensembles has a formal detection (S/N$>$3) in the R-J tail ($\lambda_{\rm rest} > 350$$\,\mu$m). With these important considerations in mind we conclude that future infrared and millimetre observations of large and representative samples at $\leq 1-2\arcsec$ resolution, are necessary to capture unequivocally the ISM mass budget and ISM conditions of high-z QGs.
\section{Conclusions}
In this work we presented a robust multi-wavelength stacking analysis of a carefully selected sample of massive QGs from the COSMOS field and in three redshift bins, $\langle z \rangle =$ 0.5, 0.9, 1.2. We further complemented our sample with the stacking
ensemble of Gobat et al.\,(2018), at $\langle z \rangle = 1.8$, constructed following the same selection criteria as those adopted in the current study, as well as with literature data for local QGs. By modelling the stacked photometry, we drew the characteristic FIR SEDs, inferred the ISM mass budget, and explored the evolution of the FIR properties of massive QGs over the last ten billion years. Our results are summarised as follows:
\begin{itemize}
\item The $M_{\rm dust}$/$M_{\ast}$\ ratio rises steeply as a function of redshift up to $z \sim 1.0$ and then remains flat at least out to $ z = 2.0$. The evolution of $f_{\rm gas}$\ follows (by construction) a similar trend, with a normalisation that depends on the assumed gas phase metallicity of the QGs. For solar metallicity, $f_{\rm gas}$\ increases from 2\% to 8\% between $z = 0.5$ and $z = 1.0$. The evolution of $f_{\text{gas}}$ in our QG samples can then be interpreted as a combination of progenitor bias at $z > 1$, a uniformity in ISM conditions among newly quenched galaxies, and a closed-box consumption of the ISM after quenching, with a depletion time of $\sim$2\,Gyr.
\item The gas depletion time scales of massive QGs at all redshifts are comparable to that of local SFGs and systematically longer than that of MS galaxies at their corresponding redshifts.
\item The dust temperature of massive QGs remains roughly constant with $T_{\rm d}$\ = 21$\pm 2$\,K out to $z = 2.0$, with only a weak (if any) evolution towards a marginally higher $T_{\rm d}$\ at higher redshifts, which can be explained as the consequence of the average size evolution of QGs. This motivated us to construct and make publicly available a template IR SED of massive $z > 0$ QGs.
\item Based on our template SED, we provide predictions for the flux densities of the continuum emission in the R-J tail of passive galaxies as a function of redshift, and highlight the need of accounting for CMB effects at $z > 3$. Finally, we argue that $M_{\rm gas}$\ prescriptions calibrated on SFGs could lead to an overestimate of the predicted dust continuum flux density of QGs in the (sub)-mm bands and discuss the implications for ALMA observing strategies.
\end{itemize}
After accounting for progenitor bias, we are ultimately left with the picture of a weak to flat evolution of initial ISM conditions in QGs. That is, our current constraints do not require that the gas fraction of QGs immediately after quenching, nor the depletion time after that, change significantly in the 10\,Gyr between $z = 0$ and $z \sim 2$. This apparent regularity in the internal properties of QGs across most of the history of the Universe mirrors the consistency of star formation physics in MS galaxies throughout cosmic time. We conclude that this uniformity of QGs should extend to earlier $z > 2.0 $ epochs, which so far lack strong constraints in this regard.
\section*{Acknowledgements}
GEM and FV acknowledge the Villum Fonden research grant 13160 “Gas to stars, stars to dust: tracing star formation across cosmic time” and the Cosmic Dawn Center of Excellence funded by the Danish National Research Foundation under then grant No. 140. FV acknowledges support from the Carlsberg Foundation research grant CF18-0388 ``Galaxies: Rise And Death”. S.J. acknowledges financial support from the Spanish Ministry of Science, Innovation and Universities (MICIU) under AYA2017-84061-P, co-financed by FEDER (European Regional Development Funds). K.E.W. wishes to acknowledge funding from the Alfred P. Sloan Foundation.
|
\section{Introduction}
The study of neutron stars using gravitational-wave observations has opened a new window to study how matter behaves at supranuclear densities. The behaviour of neutron-star matter is described by the nuclear equation of state (EoS), which determines the relation between parameters such as the mass and radius of a neutron star. The measurement of the first binary neutron star merger, GW170817, allowed us to constrain the equation of state by measuring the tidal deformability of each neutron star~\citep{LSC_GW170817}. These results have been combined with GW190425~\citep{LSC_GW190425}, and measurements from electromagnetic observations and nuclear theory~\citep[e.g.][]{Raaijmakers_2019,Capano_2020,Dietrich_2020,hernandez_vivanco_2020}.
At sufficiently high densities, strangeness containing matter may appear in the core, e.g., in the form of hyperons or deconfined quark matter, significantly changing the behaviour of the star.
Signatures of strange matter can be inferred during the inspiral and post-merger of a binary neutron star system. \citet{Chatziioannou_2020} show that, during the inspiral, hadron-quark phase transitions can be detected with 50-100 observations assuming a LIGO-Virgo detector network, as long as the phase transition is strong and occurs within the detected population.
If the appearance of strange matter does not occur before or during the inspiral, it can still be triggered after the binary merger. Since the remnant is hot after the merger, the density of the core can reach levels which cannot be reached during the inspiral. These conditions can trigger the appearance of exotic matter such as hyperons or deconfined quarks, which can be measured using gravitational-wave observations~ \citep{Sekiguchi_2011,Radice_2017,Weih_2020}. \citet{Bauswein_2019} show that hadron-quark phase transitions can be detected during the post-merger by comparing the dominant gravitational-wave frequency $f_{\mathrm{peak}}$ with the tidal deformability $\lambda$ measured during the inspiral. If no phase transition to quark matter occurs, the peak frequency depends predominantly on $\lambda$~\citep{Bauswein_2012,Bauswein_2012_b}. If $f_{\mathrm{peak}}$ shifts from the value predicted from the tidal deformability inferred from the inspiral, one may infer the existence of exotic phases.
\citet{Weih_2020} outlines four different outcomes of a post-merger remnant which depend on whether phase transitions to quark matter are triggered after the merger. These scenarios are shown in Fig.~1 of Ref.~\citep{Weih_2020} and correspond to the following outcomes: (1) a phase transition does not occur, (2) a phase transition occurs immediately after the merger, (3) a phase transition is not immediately triggered after the merger, but when it is triggered, the post-merger remnant collapses to a black hole, and (4) a phase transition is not immediately triggered, but when it is triggered, the remnant does not collapse to a black hole and forms a metastable object emitting gravitational waves at higher frequencies than it would without a phase transition.
Scenarios (1), (2) and (3) have been studied in the literature \citep[e.g.][]{Bauswein_2012,Takami_2014,Takami_2015,Kawaguchi_2018,Most_2019}. Scenario (4), referred to as ``delayed phase transition,'' was introduced in \citet{Weih_2020} and takes place a few milliseconds after the merger due to a sudden softening of the equation of state of the core, which causes the core density to overcome a critical phase-transition density. This scenario is particularly interesting because the post-merger gravitational-wave emission may be characterised by two distinct frequencies.
In this paper, we propose a fifth type of post-merger scenario that is triggered due to the cooling of a neutron star. Using finite-entropy realistic equations of state that take into account the effect of $\Lambda$-hyperons~\cite{Banik_2014}, we find that a post-merger remnant can spin up under the right conditions. The neutron star spin up, combined with the softening of the equation of state, shifts the main gravitational-wave emission frequency $f_{\mathrm{peak}}$. When $\Lambda$-hyperons are present in the core of a neutron star, we find that $f_{\mathrm{peak}}$ changes by \unit[$\sim 540$]{Hz}. If this spin-up occurs in nature, it will occur on a timescale larger than the delayed phase transition proposed in \citet{Weih_2020}.
The remainder of this paper is organized as follows. In Sec.~\ref{sec:Method}, we explain why the post-merger remnant is expected to spin up when the entropy drops after the merger. We show the conditions that have to be satisfied in order to observe the post-merger remnant spin up using gravitational-wave observations. In Sec.~\ref{sec:results} we calculate the gravitational-wave emission frequency variation when a delayed appearance of hyperons is triggered. In Sec.~\ref{sec:Discussion}, we discuss our results, and we conclude in Sec.~\ref{sec:Conclusion}.
\section{Neutron-star spin-up}\label{sec:Method}
The radius of a neutron star can significantly change as a function of temperature when a newly born neutron star cools after the merger \citep[e.g.][]{Panda_2010,stone_2019,Nunna_2020}.
\citet{stone_2019} show that when the temperature of a neutron star increases, hyperons appear at lower densities, consequentially changing the radius of a neutron star due to a softening of the equation of state. Figure~\ref{fig:mass_radius} shows the gravitational mass $m_G$ versus radius $R$ profile of the neutron star for several equations of state used in this study. We briefly describe the equations of state here. They were also used by \citet{Nunna_2020}.
\begin{itemize}
\item \textbf{DD2: an equation of state including nucleonic matter.} In this model \cite{Hempel_2010}, nucleonic matter is described by an ensemble of nuclei and interacting nucleons in nuclear statistical equilibrium. While uniform nuclear matter is described by a relativistic mean field model, nuclei are described using nuclear structure calculations based on nuclear Lagrangian density. The transition from the non-uniform phase (nuclei) to uniform nuclear matter is implemented using a thermodynamically consistent description with excluded volume corrections.
\item \textbf{BHB$\Lambda\phi$: an equation of state including $\Lambda$-hyperons.} In this model ~\cite{Banik_2014}, the non-uniform nuclear matter description of the DD2 EoS \cite{Hempel_2010} is used following the standard prescription of minimization of the free energy. As the hyperon-hyperon interaction is mediated via the non-strange vector meson $\phi$, the EoS with $\Lambda$-hyperons is represented by BHB$\Lambda \phi$.
\end{itemize}
The equations of state DD2 and BHB$\Lambda\phi$ are both evaluated at entropy per baryon of \unit[$s=2$]{$k_B$} and \unit[$s=0$]{$k_B$}. From Fig.~\ref{fig:mass_radius}, we see that the radius of a \unit[2]{$M_\odot$} neutron star changes by roughly 12\% when the entropy per baryon decreases from \unit[$s=2$]{$k_B$} to \unit[$s=0$]{$k_B$}. The mass-radius relations are calculated using the numerical library \textsc{Lorene}~\citep{lorene}.
In the remainder of this section, we show that a post-merger remnant can spin up when its temperature decreases after the merger, due to the radius variation shown in Fig.~\ref{fig:mass_radius}. We calculate the increase in the main gravitational-wave emission frequency $f_\mathrm{peak}$ associated with this spin-up.
\subsection{Post-merger remnant spin up} \label{sec:spin-up}
We first show that a post-merger remnant can spin up after the merger. We assume that after the merger, angular momentum is conserved.
With this assumption we are assuming that the spin-up is relatively fast compared to the spin-down timescale.
We revisit this momentarily.
Thus,
\begin{equation}
I_i \Omega_i = I_f \Omega_f.
\end{equation}
Here, $I$ is the moment of inertia, $\Omega$ is the angular velocity and the subscripts $i$ and $f$ refer to the initial and final state of the post-merger remnant. If we additionally assume that mass is conserved during the cool down and that the density of the neutron star is uniform, we find that
\begin{equation} \label{eq:spin-up}
\Omega_f = \left(\frac{R_i}{R_f}\right)^2 \Omega_i.
\end{equation}
Equation (\ref{eq:spin-up}) shows that if the radius of a neutron star changes by $\sim$12\%, as in Fig.~\ref{fig:mass_radius}, its angular velocity increases by $\sim$25\%.
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{mass_radius_LORENE.pdf}
\caption{
Gravitational mass versus radius relations used in this study, corresponding to the nucleonic equation of state DD2, and the BHB$\Lambda\phi$ equation of state containing $\Lambda$-hyperons. This plot shows that the radius of a \unit[2]{$M_\odot$} neutron star decreases by 12\% when the entropy drops from $s=2k_B$ to $s=0k_B$.
}
\label{fig:mass_radius}
\end{figure}
\subsection{Cooling timescale} \label{sec:cooling_timescale}
The post-merger remnant can spin up after the merger due to the radius changing during cooling. However, other torques related to electromagnetic and gravitational-wave emission will spin down the neutron star~\citep[e.g.][]{Paschalidis_2012}.
Here, we explain the necessary conditions under which we expect the spin up to take place.
Let us assume that a post-merger remnant has an initial angular momentum $J$. For simplicity, we assume that gravitational-wave spin down dominates over electromagnetic torques. Under this assumption, the change of angular momentum over time $\dot J$ is given by
\begin{equation}\label{eq:j_dot}
\dot J = \dot \omega I + \omega \dot I.
\end{equation}
The first term, $\dot \omega I$, is associated with spin-down due gravitational-wave emission. The second term, $\omega \dot I$, is associated with spin-up due to the radius changing with the falling temperature of the remnant after the merger.
The post-merger remnant spin-down, caused by $\dot \omega I$ in Eq.~(\ref{eq:j_dot}), occurs during a gravitational-wave timescale $t_\mathrm{gw}$, which can be approximated by~\cite{Paschalidis_2012}
\begin{equation}\label{eq:gravitational_timescale}
t_{\mathrm{gw}} \simeq 200 \left( \frac{\epsilon}{0.5} \right)^{-4} \left( \frac{e}{0.75} \right)^{-2} \left( \frac{R}{20 \mathrm{ km}} \right)^4 \left( \frac{m_G}{2.8 M_\odot} \right)^{-3} \mathrm{ms} .
\end{equation}
Here, $\epsilon$ is the ratio of the star's angular frequency to the break-up angular frequency, $e$ is the ellipticity of the post-merger remnant, $R$ is the radius and $m_G$ is the gravitational mass.
The characteristic timescale for the spin-up, caused by the term $\omega \dot I$ in Eq.~(\ref{eq:j_dot}), is the cooling timescale $t_{\mathrm{cool}}$.
The dominant cooling mechanism is neutrino emission~\cite{Paschalidis_2012}:
\begin{equation}\label{eq:cooling_timescale}
t_{\mathrm{cool}} \simeq 400 \left( \frac{m_G}{2.8M_\odot} \right) \left( \frac{R}{20 \mathrm{ km}} \right)^{-1} \left( \frac{E_\nu}{10 \mathrm{ MeV}} \right)^2 \mathrm{ms} .
\end{equation}
Here, $E_\nu$ is the root mean squared (RMS) neutrino energy. At densities of $\unit[\gtrsim 10^{11}]{g\, cm^{-3}}$ neutrinos are trapped.
Therefore the cooling timescale is predominantly determined by how long it takes for neutrinos to diffuse out of the remnant~\cite{Paschalidis_2012}. The neutrino energy $E_\nu$ (and therefore the cooling timescale) depends on the cooling transfer mechanism which is not well understood.
The condition required to observe the post-merger remnant spin up is given by
\begin{equation}\label{eq:condition}
t_{\mathrm{cool}} < t_{\mathrm{gw}}.
\end{equation}
Equation~(\ref{eq:condition}) implies that if the remnant spin-up takes place with a timescale greater than the gravitational-wave timescale, it is suppressed by the spin-down caused by gravitational-wave emission.
On the other hand, if $t_{\mathrm{cool}} < t_{\mathrm{gw}}$, the remnant spins up faster than it spins-down.
We explore the mass-radius, and the neutrino energy-ellipticity parameter space that satisfy $t_{\mathrm{cool}} - t_{\mathrm{gw}} < 0$ assuming $\epsilon=0.5$.
We consider two values of ellipticity and neutrino energy to determine the values of gravitational mass and radius that would result in a post-merger remnant spin up. The results are shown in Fig.~\ref{fig:fixed_e_and_E}.
We indicate in red the gravitational masses and radii that result in an observable spin-up. The black regions show the gravitational mass and radius values where the spin up is not observable. From Fig.~\ref{fig:fixed_e_and_E}, we see that for neutrino energies in the order of \unit[8]{MeV} and ellipticities in the order of $e=0.5$, the remnant spin-up is observable for masses \unit[$m_G \lesssim 2.3$]{$M_\odot$}. However, for larger neutrino energies, i.e. \unit[$E_\nu =15$]{MeV}, the remnant spin up is not be observable regardless of its mass.
\begin{figure}
\centering
\begin{subfigure}[b]{0.5\textwidth}
\includegraphics[width=\columnwidth]{fixed_e_and_E.pdf}
\caption{}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\includegraphics[width=\columnwidth]{fixed_e_and_E_2.pdf}
\caption{}
\end{subfigure}
\caption[]{(a) Spin up timescales as a function of gravitational mass and radius for a fixed ellipticity $e=0.5$ and neutrino energy \unit[$E_\nu=8$]{MeV}. Only the regions where $t_{\mathrm{cool}}- t_{\mathrm{gw}}<0$, shown in red, correspond to an observable spin up of the post-merger remnant. The blue curve corresponds to the DD2 equation of state considered in Fig.~\ref{fig:mass_radius} at \unit[$s=2$]{$k_B$} rotating with an angular momentum \unit[$J=2.2$]{$GM_\odot^2/c$}. The black dotted line corresponds to the values of masses and radii where $t_{\mathrm{gw}} = t_{\mathrm{cool}}$ (b) Same as (a) but for an ellipticity $e=0.75$ and neutrino energy \unit[$E_\nu=15$]{MeV}.}
\label{fig:fixed_e_and_E}
\end{figure}
In Fig.~\ref{fig:fixed_m_and_r}, we show the values of ellipticity and neutrino energy that would be needed to observe the remnant spin up, assuming the mass of a \unit[$m_G=2.11$]{$M_\odot$} remnant assuming the DD2 equation of state at \unit[$s=2$]{$k_B$} rotating with an angular momentum \unit[$J=2.2$]{$GM_\odot^2/c$}. We see that RMS neutrino energies \unit[$\lesssim 14$]{MeV} result in an observable spin up.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{fixed_m_r.pdf}
\caption[]{Spin up timescales as a function of ellipticity and neutrino energy of a post-merger remnant of mass \unit[$m_G=2.11$]{$M_\odot$}. The remnant follows the DD2 equation of state at entropy \unit{$s=2$}{$k_B$} rotating with an angular momentum \unit[$J=2.2$]{$GM_\odot^2/c$}. Only the regions where $t_{\mathrm{cool}}- t_{\mathrm{gw}}<0$, shown in red, correspond to an observable spin up of the post-merger remnant. The black dotted curve corresponds to the values of ellipticity and neutrino energy where $t_{\mathrm{gw}} = t_{\mathrm{cool}}$.
}
\label{fig:fixed_m_and_r}
\end{figure}
\subsection{Gravitational-wave emission frequency}\label{sec:emission_frequency}
The gravitational-wave emission frequency of a post-merger remnant is proportional to the rotational frequency~\citep{Gaertig_2011,Doneva_2013}. Therefore, if the post-merger remnant spins up, the gravitational-wave emission frequency increases.
Simulations of the evolution of post-merger remnants reveal gravitational-wave spectra with characteristic peaks related to oscillation modes~\cite{Bauswein_2012,Bauswein_2012_b,Bauswein_2016,Paschalidis_2017}. The dominant frequency $f_{\mathrm{peak}}$ (also referred as $f_2$), is related to the co-rotating $l=m=2$ $f$-mode moving at a positive pattern speed in the prograde direction~\cite{Stergioulas_2011,Takami_2015,Doneva_2015}. This mode depends on the mass, radius and rotational frequency of the remnant. Thus if we can accurately measure $f_{\mathrm{peak}}$, we can infer properties of the equation of state.
Gaertig and Kokkotas~\cite{Gaertig_2008,Gaertig_2011} describe relations for the co-rotating $l=|m|=2$ $f$-modes in the Cowling approximation, which assumes that the spacetime remains frozen during the time evolution. \citet{Doneva_2013} expands on the work of Gaertig and Kokkotas using realistic equations of state. The Cowling approximation does not accurately predict the $f$-mode and can include errors up to 20-30\%. Since our aim is to obtain an order of magnitude estimate of $f_\mathrm{peak}$, we approximate $f_\mathrm{peak}$ following Refs.~\cite{Gaertig_2008,Gaertig_2011,Doneva_2013}.
The $l=m=2$ $f$-mode stable branch in the co-rotating frame $\sigma_{\mathrm{corot}}$ is given by \citep{Doneva_2013}
\begin{equation}\label{eq:f_mode}
\frac{\sigma_{\mathrm{corot}}}{\sigma_0} = 1 - 0.235\left(\frac{\Omega}{\Omega_k} \right) -0.358\left(\frac{\Omega}{\Omega_k} \right)^2,
\end{equation}
where $\Omega$ is the neutron star's rotation frequency, $\Omega_k$ is the Keplerian rotation frequency (or mass-shedding limit) and $\sigma_0$ is the $f$-mode of a non-rotating neutron star. An approximate value of the Keplerian frequency can be found in Refs.~\cite{Glendenning_1992,Haensel_1989,Lasota_1996,Friedman_1989,Doneva_2013}. We use the Keplerian approximation presented in Ref.~\cite{Doneva_2013},
\begin{equation}\label{eq:kepler_limit}
\frac{1}{2\pi}\Omega_k \mathrm{[kHz]} = 1.716\sqrt{\frac{\bar{m}_0}{\bar{R}_0^3}} - 0.189.
\end{equation}
Here, $\bar{m}_0 = m_G / 1.4M_\odot$ and $\bar{R}_0 = R / \unit[10]{km}$ is the mass and radius in the non-rotating configuration. Equation (\ref{eq:kepler_limit}) is not a precise estimate. Its true value depends, among other parameters, on the presence of hyperons and the cooling process of the remnant.
Additionally, the $l=2$ $f$-mode of a non-rotating neutron star $\sigma_0$ is given by~\cite{Doneva_2013}
\begin{equation}
\frac{1}{2\pi}\sigma_0 \mathrm{[kHz]} = 1.562 +1.151 \sqrt{\frac{\bar{m}_0}{\bar{R}_0^3}}.
\end{equation}
Finally, the co-rotating frequency $\sigma_{\mathrm{corot}}$ can be transformed to the inertial frame $\sigma_{\mathrm{inertial}}$,
\begin{equation} \label{eq:inertial_frame}
\sigma_{\mathrm{inertial}} = \sigma_{\mathrm{corot}} - m\Omega.
\end{equation}
In the convention used in \citet{Doneva_2013}, the $m<0$ modes are prograde, i.e., the $f$-mode frequency in the inertial frame increases when the angular velocity $\Omega$ increases. We set $m=-2$ in Eq.~(\ref{eq:inertial_frame}).
\section{Results}\label{sec:results}
The post-merger remnant spin-up due to cooling described in Sec.~\ref{sec:spin-up} can happen regardless of whether $\Lambda$-hyperons appear in the core. However, we find the peak frequency $f_\mathrm{peak}$, approximated in Sec.~\ref{sec:emission_frequency}, is emitted at a different frequency when hyperons appear during the post-merger.
In Fig.~\ref{fig:rotating_mass_radius}, we present different scenarios that can take place after a binary neutron star merges. The points $A$, $B$, $C$ and $D$ in Fig.~\ref{fig:rotating_mass_radius} correspond to the angular velocities of an \unit[$m_B=2.32$]{$M_\odot$} post-merger remnant assuming an angular momentum \unit[$J=2.2$]{$GM_\odot^2/c$}. Here, $m_B$ is the baryon mass of a post-merger remnant which is conserved during cooling, and \unit[$m_B=2.32$]{$M_\odot$} corresponds to \unit[$m_G=2.11$]{$M_\odot$} assuming the DD2 equation of state at \unit[$s=2$]{$k_B$}. These values are calculated using \textsc{Lorene}~\citep{lorene}.
The post-merger remnant scenarios are described as follows:
\begin{itemize}
\item \textit{No appearance of $\Lambda$-hyperons.} The post-merger remnant cools down from $D \xrightarrow{} B $ from a hot nucleonic neutron star to a cold nucleonic one.
\item \textit{Prompt appearance of $\Lambda$-hyperons.} The post-merger remnant cools down from $D \xrightarrow{} C $; hyperonic matter is present soon after the merger.
\item \textit{Delayed appearance of hyperons caused by density oscillations.} This delayed appearance of hyperons is similar to the hadron-quark phase transition presented in \citet{Weih_2020}, in which the remnant does not undergo a transition immediately after the merger.
In this case, the post-merger remnant cools down from $D \xrightarrow{} B \xrightarrow{} A $ in Fig.~\ref{fig:rotating_mass_radius}, i.e., the remnant cools from a hot nucleonic to a cold nucleonic one, and after a few milliseconds $\Lambda$-hyperons appear. The delayed appearance of hyperons caused by density oscillation is triggered milliseconds after the merger.
\item \textit{Delayed appearance of hyperons caused by cooling.} This is the delayed appearance of hyperons proposed in this paper, which is triggered by the post-merger remnant cooling described in Sec.~\ref{sec:spin-up}. Similar to the delayed appearance of hyperons caused by density oscillations proposed in \citet{Weih_2020}, the post-merger remnant cools down from $D \xrightarrow{} B \xrightarrow{} A $, but occurs on a longer timescale that depends on how fast the post-merger remnant cools.
The transition timescale is significantly longer than a few milliseconds. Since the delayed appearance of hyperons induced by density oscillations takes place milliseconds after the merger \citep[e.g.][]{Weih_2020}, it will be distinguishable from the delayed appearance of hyperons caused by cooling.
\end{itemize}.
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{mass_omega.pdf}
\caption{Baryon mass $m_B$ and rotational frequency $\Omega/2\pi$ sequences of the equations of state BHB$\Lambda\phi$ (with hyperons) and DD2 (pure nucleonic matter). We rotate the equations of state shown in Fig.~\ref{fig:mass_radius} with a constant angular momentum \unit[$J=2.2$]{G$M_\odot^2/c$}. The points $A$, $B$, $C$ and $D$ correspond to the angular rotational frequencies of an \unit[$m_B=2.32$]{$M_\odot$} post-merger remnant. If a delayed appearance of hyperons is triggered, the post-merger remnant cools down from $D \xrightarrow{} B \xrightarrow{} A $, i.e., the remnant first cools from a hot nucleonic EoS to a cold nucleonic EoS, and then hyperons appear in the core. If hyperons do not appear, the post-merger remnant cools down from $D \xrightarrow{} B$.
}
\label{fig:rotating_mass_radius}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{f_peak.pdf}
\caption{
Post-merger peak frequency $f_{\mathrm{peak}}$ as a function of rotation frequency $\Omega/2\pi$ of a \unit[$m_B$=2.32]{$M_\odot$} neutron star rotating with a constant angular momentum \unit[$J=2.2$]{G$M_\odot^2/c$}. If a delayed appearance of of hyperons is triggered by cooling, the post-merger remnant spins up from $D \xrightarrow{} B \xrightarrow{} A $. In this case, $f_\mathrm{peak}$ changes by \unit[$\sim 540$]{Hz}. If hyperons do not appear, the post-merger remnant spins up from $D \xrightarrow{} B $, where $f_\mathrm{peak}$ changes by \unit[$\sim 360$]{Hz}.
}
\label{fig:frequency_shift}
\end{figure}
Using Eqs. (\ref{eq:f_mode}) and (\ref{eq:inertial_frame}), we calculate the $l=m=2$ $f$-mode, which corresponds to the peak frequency of a \unit[$m_B=2.32$]{$M_\odot$} post-merger remnant using the equations of state presented in \citet{Nunna_2020}.
In Fig.~\ref{fig:frequency_shift}, we plot $f_\text{peak}$ versus $\Omega$.
The points $A$, $B$, $C$ and $D$ correspond to the same points shown in Fig.~\ref{fig:rotating_mass_radius}. If hyperons do not appear after the merger, the post-merger remnant cools down from $D \xrightarrow{} B$. In this case, Fig.~\ref{fig:frequency_shift} shows that the peak frequency changes by \unit[$\sim 360$]{Hz}.
In contrast, if a delayed appearance of hyperons is triggered, the post-merger remnant cools down from $D \xrightarrow{} B \xrightarrow{} A $. In this case, Fig.~\ref{fig:frequency_shift} shows that the peak frequency changes by \unit[$\sim 540$]{Hz}. The peak frequency increase will allow us to probe if exotic states of matter are produced after the merger.
\section{Discussion}\label{sec:Discussion}
\citet{Chatziioannou_2017} and \citet{Easter_2020} show that gravitational-wave observations of a post-merger remnant can resolve $f_\mathrm{peak}$ with an accuracy of \unit[20-50]{Hz} at matched-filter signal-to-noise ratio (SNR) 10, depending on the equation of state. Given that the variation of $f_\mathrm{peak}$ in a delayed appearance of hyperons is \unit[$\sim 540$]{Hz}, we expect this frequency variation to be measurable by gravitational-wave detectors. Although measuring a post-merger signal with SNR$\sim10$ is challenging with current detectors, dedicated high-frequency detectors will help make possible such a high SNR~\citep[e.g.][]{Martynov_2019,Ackley_2020}.
For example, on average, an SNR=10 event is expected to be observed every $\sim$$\unit[4-8]{years}$ with a network of Neutron Star Extreme Matter Observatories~\citep{Ackley_2020}.
The evolution of the gravitational-wave frequency caused by cooling will be subtle, depending on precise details of cooling as well as the interplay between angular momentum loss from gravitational waves, electromagnetic radiation, and internal dissipation such as viscosity~\citep[e.g.][]{Alford_2018}. Understanding the evolution of $f_\mathrm{peak}$ will require complex numerical simulations that take into account all of these effects as well as general relativity.
Performing numerical simulations of the delayed appearance of hyperons, which take into account the microphysics such as magnetic fields and neutrino emission is challenging.
Numerical-relativity simulations that take into account phase transitions and appearance of $\Lambda$-hyperons are usually used to study the first \unit[$\sim 20$]{s} after the merger ~\citep[e.g.][]{Sekiguchi_2011_2,Perego_2019,Most_2019,Weih_2020}. For example, \citet{Sekiguchi_2011} simulate transitions from nucleonic matter to hyperons using finite-entropy equations of state taking into account neutrino emission. They find that for \unit[1.35]{$M_\odot$}, equal-mass binaries, the appearance of $\Lambda$-hyperons causes the post-merger remnant to collapse to a black hole \unit[$\sim 12$]{ms} after the merger due to a sudden softening of the core. They find that the appearance of hyperons changes the gravitational-wave characteristic frequency by $\sim 20-30$\%, contrary to the nucleonic case. It remains to be seen if numerical-relativity simulations of lower mass binaries, such as masses corresponding to a total mass of \unit[$m_G \lesssim 2.3$]{$M_\odot$}, agree with the approximations used in this study.
The delayed appearance of hyperons is most readily observable if the cooling timescale $t_\mathrm{cool}$ is less than the gravitational-wave timescale $t_\mathrm{gw}$. The cooling timescale depends on the neutrino energy, which is poorly understood. In order for a post-merger remnant to spin up, we need relatively low neutrino energies, i.e., \unit[$\lesssim 14$]{MeV} assuming a \unit[$m_G=2.11$]{$M_\odot$} remnant. Refs.~\citep{Rosswog_2003,Sekiguchi_2011_2,Richers_2015,sumiyoshi2020properties} suggest that the average neutrino energy can vary between \unit[$\sim 8$]{MeV} and \unit[$\sim 30$]{MeV}. However, the models presented in Refs.~\citep{Rosswog_2003,Sekiguchi_2011_2,Richers_2015} depend sensitively on the neutrino transfer mechanism and do not account for the general theory of relativity, which can significantly alter the results.
The average neutrino energy is similarly not well constrained by observations. A search for neutrinos has been carried out on GW170817~\citep{LSC_GW170817}. Super-Kamiokande found no coincident neutrinos in the range \unit[3.5]{MeV} - \unit[100]{PeV} in two different time frames: \unit[$\pm 500$]{s} around the merger of GW170817~\citep{Abe_2018} and 14 days after the merger. Similarly, a search for high energy neutrinos in the GeV-EeV range was carried out by ANTARES, IceCube, and the Pierre Auger Observatory~\citep{Albert_2017}. They found no coincident neutrinos in the same time frames described in Ref.~\citep{Abe_2018}. However, the non-detection of neutrinos agrees with model predictions of gamma-ray bursts~\citep{Albert_2017}.
In order to observe the delayed appearance of hyperons caused by cooling, the post-merger remnant should be long-lived and have masses \unit[$m_G \lesssim 2.3$]{$M_\odot$}.
The distribution of Galactic binary neutron star is well fit by a Gaussian with mean \unit[1.33]{$M_\odot$} and width \unit[0.09]{$M_\odot$}~\citep{Farrow_2019}. By allowing conservation of rest mass, and taking into account mass loss after the merger, the post-merger mass distribution lies in the range between \unit[2.2-2.5]{$M_\odot$}~\cite{Belczynski_2008,Lasky_2014}.\footnote{The measurement of GW190425~\citep{LSC_GW190425} shows a $\geq 5\sigma$ deviation from the Galactic distribution, although see~\citep{heavy_dns}, in which the authors use a population study to argue that GW190425 is not as different from the Milky Way population as it initially appeared. At any rate, we ignore heavier neutron stars because only low-mass binaries trigger a delayed appearance of hyperons.} In this paper, we focus on an \unit[$m_G= 2.11$]{$M_\odot$}; this limitation is given because the $f$-mode approximation defined in Eq. (\ref{eq:f_mode}) depends on the mass in the non-rotating configuration. Therefore, the maximum value we can use for our estimates is the maximum non-rotating mass (TOV mass) of the BHB$\Lambda\phi$ equation of state at $s=0$, corresponding to \unit[$m_G = 2.11$]{$M_\odot$}.
\section{Conclusion}\label{sec:Conclusion}
We use finite-entropy equations of state, which include the effect of hyperons to study the evolution of a post-merger remnant. We find that under the right circumstances, a post-merger remnant can spin up due to cooling. The spin-up of the remnant may trigger a delayed appearance of hyperons, which can be measured by gravitational-wave observations. Since the gravitational-wave emission frequency of the post-merger is proportional to the angular velocity, we find the main gravitational-wave emission frequency $f_{\mathrm{peak}}$ increases by \unit[$\sim 540$]{Hz} if a delayed appearance of hyperons is triggered. If hyperons do not appear and the remnant spins up, $f_\mathrm{peak}$ changes by \unit[$\sim 360$]{Hz}. This will allow us to test for exotic states of matter during the post-merger while probing protoneutron star cooling times.
The delayed appearance of hyperons is most readily observable when the cooling timescale of the post-merger is less than the gravitational-wave timescale defined by Eqs.~(\ref{eq:gravitational_timescale}) and (\ref{eq:cooling_timescale}) respectively. Assuming a \unit[$m_G=2.11$]{$M_\odot$} post-merger remnant, Eq.(\ref{eq:cooling_timescale}) is satisfied for neutrino RMS energies \unit[$\lesssim 14$]{MeV}.
\section*{Acknowledgments}
This work is supported through Australian Research Council Grant No. CE170100004, No. FT150100281, No. FT160100112, and No. DP180103155. F.H.V. is supported through the Monash Graduate Scholarship (MGS).
|
\section{Linear response property in the thermodynamics limit}
In the limit $L\to \infty$ the spin chain can be seen as a macroscopic environment allowing for a different viewpoint on the dynamics of the qudit. Environments consisting of independent (i.e. noninteracting) degrees of freedom and appropriate scaling of the coupling $\gamma$ fulfil linear response theory, i.e. the influence of the environment on the qudit is fully characterized by the force autocorrelation function of the bath allowing for the description of the influence of the environment in terms of a bath of harmonic oscillators with an effective spectral density.\cite{Makri1999} This idea was used, for example, to describe the dynamics of a two-level system coupled to a spin bath consisting of independent spins\cite{Shao1998, Makri1999} and to a bath of anharmonic vibrational degrees of freedom\cite{Wang2007}. For non-interacting environments this was proven by showing that all but the leading order term in the cumulant expansion of the influence functional vanish in the thermodynamic limit. Despite the fact that the spin chain considered in our work consists of interacting spins, one can show that the linear response property also holds for this model, and thus, the influence of the spin chain on the qudit can be described by a bath of harmonic oscillators with an effective spectral density.
To this end, we follow the derivation given in Ref.\,[\onlinecite{Makri1999}] and extend it to the interacting spin chain considered here. We assume that initially there are no correlations between different spins and the spin chain is in an eigenstate of all $\lbrace\sigma_i^z\rbrace$. In order to evaluate the different terms in the cumulant expansion\cite{Makri1999} the time dependent operator
\begin{align}
\hat{f}(t) &= - \gamma \sum_{i=1}^L \underbrace{{\rm e}^{i H_0 t} \sigma_i^x {\rm e}^{-i H_0 t}}_{\sigma_i^x(t)}
\end{align}
is required, which corresponds to the bath part of the system-bath interaction and describes the force exerted on the system due to its interaction with the environment. Here, $H_0$ is the Hamiltonian of the isolated spin chain. The time evolved operators $\sigma_i^x(t) = \sigma_i^+(t) + \sigma_i^-(t)$ can be calculated analytically yielding
\begin{align}
\sigma_i^+(t) &= {\rm e}^{2ih\xi_it} \frac{1}{2} \bigg[ [\mathds{1}- \sigma_{i-1}^z\sigma_{i+1}^z] + [\mathds{1} + \sigma_{i-1}^z\sigma_{i+1}^z]\cos(4gt) + i [\sigma_{i-1}^z+\sigma_{i+1}^z]\sin(4gt) \bigg] \sigma_i^+ \nonumber \\
&= {\rm e}^{2ih\xi_it} \hat{\varphi}_i(t) \sigma_i^+ \label{eq:tdp-operator},
\end{align}
and $\sigma_i^-(t) = (\sigma_i^+(t))^\dagger$. Here $\sigma_i^+$ and $\sigma_i^-$ are the spin raising and lowering operators, respectively. Using this equation, one can show that
\begin{align}
\big[\sigma_i^x(t), \sigma_j^x(t^\prime) \big]& = 0 ~~~~ \forall~ i,j ~{\rm with} ~|i-j| \geq 2,
\end{align}
This follows from the fact that $\hat{\varphi}_i(t)$ only involves spin operators on sites $i-1$ and $i+1$.
All terms in the cumulant expansion of the influence functional can be expressed\cite{Makri1999} in terms of $N$-time correlation functions defined as
\begin{align}
C^{(N)}(t_1,..., t_N) &= \braket{\hat{f}(t_1)...\hat{f}(t_N)}_0,
\end{align}
where $\braket{...}_0$ denotes the expectation value with respect to the initial state of the environment. Initially, the spin chain is in an eigenstate of all $\lbrace\sigma_i^z\rbrace$, and thus, one finds that
\begin{align}
C^{(1)} = 0,
\end{align}
Consequently, the first term, and by extension all odd order terms in the expansion, vanishes. The second order term can be expressed in terms of the two-point correlation function
\begin{align}
C^{(2)}(t_1, t_2) &= \gamma^2 \sum_i \sum_j \braket{\sigma_i^x(t_1)\sigma_j^x(t_2)}_0.
\end{align}
It is straightforward to check that the expectation value is zero for all $i \neq j$, and thus, the two-time correlation function reduces to
\begin{align}
C^{(2)}(t_1, t_2) &= \gamma^2 \sum_{i=1}^L \braket{\sigma_i^x(t_1)\sigma_i^x(t_2)}_0.
\end{align}
Since $\braket{\sigma_i^x(t_1)\sigma_i^x(t_2)}_0$ is finite but not zero in general, the sum diverges in the thermodynamic limit $L\to \infty$ unless $\gamma\sim \nicefrac{1}{\sqrt{L}}$, which we will assume in the following by setting $\gamma = \gamma_0\sqrt{\nicefrac{L_0}{L}}$. The next non-vanishing term in the cumulant expansion is the fourth term which involves the fourth order correlation function
\begin{align}
C^{(4)}(t_1, t_2, t_3, t_4) &= \gamma^4 \sum_{i,j,k,l} \braket{\sigma_i^x(t_1)\sigma_j^x(t_2)\sigma_k^x(t_3)\sigma_l^x(t_4)}_0.
\end{align}
Since $\hat{\varphi}_i(t)$ does not change the initial state of the spin chain, there can only be up to two different indices in the expectation value. Thus, the four-time correlation function can be written as
\begin{align}
C^{(4)}(t_1, t_2, t_3, t_4) &= \gamma^4 \sum_{i,j} \braket{\sigma_i^x(t_1)\sigma_i^x(t_2)\sigma_j^x(t_3)\sigma_j^x(t_4)}_0 \nonumber \\
&+\gamma^4 \sum\limits_{\substack{i,j \\i \neq j}} \braket{\sigma_i^x(t_1)\sigma_j^x(t_2)\sigma_i^x(t_3)\sigma_j^x(t_4)}_0 \nonumber\\
&+\gamma^4 \sum\limits_{\substack{i,j \\i \neq j}} \braket{\sigma_i^x(t_1)\sigma_j^x(t_2)\sigma_j^x(t_3)\sigma_i^x(t_4)}_0.
\end{align}
In the following we will discuss only the first term on the right hand side. The other two terms can be treated equivalently. The double sum can be decomposed as
\begin{align}
\gamma^4 \sum_{i,j} \braket{\sigma_i^x(t_1)\sigma_i^x(t_2)\sigma_j^x(t_3)\sigma_j^x(t_4)}_0 &= \gamma^4 \sum_{i} \braket{\sigma_i^x(t_1)\sigma_i^x(t_2)\sigma_i^x(t_3)\sigma_i^x(t_4)}_0 \nonumber\\
&+ \gamma^4 \sum_{i} \braket{\sigma_i^x(t_1)\sigma_i^x(t_2)\sigma_{i+1}^x(t_3)\sigma_{i+1}^x(t_4)}_0\nonumber \\
&+ \gamma^4 \sum_{i} \braket{\sigma_{i+1}^x(t_1)\sigma_{i+1}^x(t_2)\sigma_i^x(t_3)\sigma_i^x(t_4)}_0 \nonumber\\
&+ \gamma^4 \sum\limits_{\substack{i,j \\ |i-j|\geq2}}^L \braket{\sigma_i^x(t_1)\sigma_i^x(t_2)\sigma_j^x(t_3)\sigma_j^x(t_4)}_0.
\end{align}
The expectation values in the first three sums on the right hand side are bounded by a constant. If the coupling $\gamma$ is scaled as $\gamma = \gamma_0\sqrt{\nicefrac{L_0}{L}}$, these terms vanish as $L\to \infty$, and thus, we conclude that
\begin{align}
\gamma^4 \sum_{i,j}^L \braket{\sigma_i^x(t_1)\sigma_i^x(t_2)\sigma_j^x(t_3)\sigma_j^x(t_4)}_0 &= \gamma^4 \sum\limits_{\substack{i,j \\ |i-j|\geq2}}^L \braket{\sigma_i^x(t_1)\sigma_i^x(t_2)\sigma_j^x(t_3)\sigma_j^x(t_4)}_0 + \mathcal{O}\bigg(\frac{1}{L}\bigg).
\end{align}
Because the operators in the expectation value act on different Hilbert spaces and the initial state factorizes the expectation values can be factorized as
\begin{align}
\gamma^4 \sum\limits_{\substack{i,j \\ |i-j|\geq2}}^L \braket{\sigma_i^x(t_1)\sigma_i^x(t_2)\sigma_j^x(t_3)\sigma_j^x(t_4)}_0 &= \gamma^4 \sum\limits_{\substack{i,j \\ |i-j|\geq2}}^L \braket{\sigma_i^x(t_1)\sigma_i^x(t_2)}_0 \braket{\sigma_j^x(t_3)\sigma_j^x(t_4)}_0
\end{align}
Adding the terms for $i=j$ and $|i-j|=1$ to the double sum on the right hand side gives an error of $\mathcal{O}(\nicefrac{1}{L})$, and thus, one can write
\begin{align}
\gamma^4 \sum_{i,j} \braket{\sigma_i^x(t_1)\sigma_i^x(t_2)\sigma_j^x(t_3)\sigma_j^x(t_4)}_0 &= \gamma^4 \sum_{i,j} \braket{\sigma_i^x(t_1)\sigma_i^x(t_2)}_0 \braket{\sigma_j^x(t_3)\sigma_j^x(t_4)}_0 + \mathcal{O}\bigg(\frac{1}{L}\bigg) \\
&= C^{(2)}(t_1, t_2) C^{(2)}(t_3, t_4) + \mathcal{O}\bigg(\frac{1}{L}\bigg),
\end{align}
where we have identified the two-time correlation functions. With this, we finally conclude that
\begin{align}
\lim_{L\to \infty} C^{(4)}(t_1, t_2, t_3, t_4) = \lim_{L\to \infty} ~& C^{(2)}(t_1, t_2) C^{(2)}(t_3, t_4)\nonumber \\
&+C^{(2)}(t_1, t_3) C^{(2)}(t_2, t_4) \nonumber \\
&+C^{(2)}(t_1, t_4) C^{(2)}(t_2, t_3).
\end{align}
Using this one finds that the fourth order term in the cumulant expansion in [\onlinecite{Makri1999}] vanishes in the thermodynamic limit. In a similar way one can show that all higher order terms in the expansion vanish, proving that in the thermodynamic limit the influence functional is completely characterized by the force autocorrelation function. Thus, one can construct a bath of harmonic oscillators with an effective spectral density resulting in the same influence functional.
As a last step we show that for the choice of the 'super-Neel' state $\ket{E_{SN}} = \ket{\uparrow\uparrow\downarrow\downarrow...}$ as initial state the force autocorrelation function does not depend on the spin-spin interaction $g$. The force autocorrelation function is defined as
\begin{align}
\braket{\hat{f}(t_1)\hat{f}(t_2)} &= \gamma_L^2 \sum_{i} \braket{E_{SN}|\sigma_i^x(t_1)\sigma_i^x(t_2)|E_{SN}},
\end{align}
where the time-dependent operators are given in equation (\ref{eq:tdp-operator}). For the "super-Neel" state it follows that
\begin{align}
\sigma_{i-1}^z\sigma_{i+1}^z\ket{E_{SN}} &= -\ket{E_{SN}} \nonumber, \\
\big(\sigma_{i-1}^z+\sigma_{i+1}^z\big)\ket{E_{SN}} &= 0,
\end{align}
holds for all $i$ since the spins at site $i-1$ and $i+1$ are always antiparallel. Thus, the action of $\hat{\varphi}_i(t)$ is independent of the index $i$ and gives
\begin{align}
\hat{\varphi}_i(t) \ket{E_{SN}} = \ket{E_{SN}}.
\end{align}
Since the spin-spin interaction enters only via $\hat{\varphi}_i(t)$, we find that the force autocorrelation function, and consequently the parameters of the effective bath harmonic oscillators, are independent of the spin-spin interaction $g$. The corresponding effective spectral density can be calculated from the force autocorrelation\cite{Makri1999} yielding
\begin{align}
J_{\rm eff}(\omega) &= \frac{\pi}{2} \frac{1}{4h} \chi_{[-2h, 2h]}(\omega),
\end{align}
where $\chi_{I}$ is the characteristic function of the interval $I$, i.e. $\chi_{I}(\omega)=1$ if $\omega\in I$ and 0 else. Thus, in the thermodynamics limit a bath of harmonic oscillators with this spectral density gives rise to the same qudit dynamics as the spin chain in the "Super-Neel" state.
\section{Structure of wavefunction}
In addition to modifying the entanglement dynamics at short times, the star-like geometry of this system (depicted in the inset of Figure \ref{fig:mctdh_tree}) should render the concept of locality meaningless. Indeed, from the point of view of operator dynamics, operators for the qudit should immediately spread to $O(L)$ sites after $O(1)$ time\cite{Lucas2019}. Instead we propose to analyze the structure of eigenstates in the Hilbert space of the uncoupled ($\gamma=0$) Hamiltonian. In the spirit of the current understanding of MBL, where eigenstates are weak deformations of the unperturbed system, we parse wavefunctions in the product basis $|s\rangle$ of spins and qudit states -- $|\vect{z}\rangle\otimes|n\rangle$ with $z_i \in \{\uparrow, \downarrow\}, \forall i=1,\ldots,L$ and $n=1,\ldots,d$. We quantify ``deformations'' through a notion of Hilbert space distance as
\begin{align}
\label{eq:hamDis}
\mathcal{D}(|\vect{z},n\rangle, |\vect{z}',n'\rangle) = \max \left( \mathcal{D}_H(\vect{z}, \vect{z}'), |n - n'| \right),
\end{align}
with $\mathcal{D}_H$ being the Hamming distance between the bitstrings $\vect{z}$ and $\vect{z}'$ (here, up-spins are ``1'' and down-spins are ``0''). Intuitively, this measures the minimum number of times the transverse perturbation must be applied to connect two states in the Hilbert space. For eigenstates, the origin ($\equiv |\psi_{\text{ref}}\rangle$) is taken to be the product state with the largest weight while for time evolution, the origin is taken to be the initial state before the quench. Other product states can then be grouped according to their distance $\mathcal{D}$ from $|\psi_{\text{ref}}\rangle$. For each $\mathcal{D}$, we calculate the distribution of expansion coefficients $|\lrangle{s | \psi}|^2$ over disorder realizations and equidistant product states $\{|s\rangle \mid \mathcal{D}(|\psi\rangle, |s\rangle) = x\}$. In the nonergodic phase, these coefficients are suppressed by a factor of $(\gamma/g)^2$ as $\mathcal{D}$ increases (Figure \ref{fig:wavefunction}(a,b)). One can then regard the wavefunction as being exponentially localized in Hilbert space. This can be observed also in non-centrally coupled models of MBL, such as the disordered Ising chain with next-nearest neighbor interactions \cite{Kjall2014}. In contrast, there is no such Hilbert space localization at large enough $\gamma$ in the ergodic phase (inset of Figure \hyperref[fig:wavefunction]{5b}). This is corroborated by the average Hilbert space distance $\lrangle{\mathcal{D}(t)}$, as measured from $|\psi_{\text{ref}}\rangle$. This quantity has been noted by Hauke and Heyl \cite{Hauke2015} to saturate to $L/2$ if the system is (possibly) ergodic, and is consistent with our numerics (Figure \hyperref[fig:wavefunction]{5c}). We additionally find that the experimental accessibility $\lrangle{\mathcal{D}(t)}$ (noted by \cite{Hauke2015}) approximately holds (see top panel of Figure \ref{fig:wavefunction}c), i.e.\ $\lrangle{\mathcal{D}(t)} \geq L (1-q_{EA})/2$ where the inequality is due to the definition of $\mathcal{D}$ we have chosen.
\begin{figure}[H]
\includegraphics[width=0.99\textwidth, trim={0 0 0 0}, clip]{wavefunction.pdf}
\caption{Distributions of wavefunction coefficients in the unperturbed basis $\{ |s\rangle \}$, parsed by the Hilbert space distance $\mathcal{D}$ separating $|s\rangle$ and a reference state $|s_{\text{ref}}\rangle$. The colors of the curves lighten for increasing $\mathcal{D}$, while the dashed vertical lines are guides to the eye for where the expected peak locations should the coefficients decay as $\propto (\gamma/g)^{2\mathcal{D}}$. \textbf{(a)} Distributions in the ten eigenstates closest to the middle of the energy spectrum, choosing $|s_{\text{ref}}\rangle$ to be the unperturbed state with largest weight. \textbf{(b)} Distribution of coefficients for the system at $t=10^{12}$, evolving from $|\psi(0)\rangle = |s_{\text{ref}}\rangle|(d+1)/2\rangle$ with $|s_{\text{ref}}\rangle$ being the super-Neel state. \textbf{(c)} Average Hilbert space distance $\lrangle{D(t)}$ from the super-Neel state (bottom) and in comparison with the spin glass order parameter $q_{EA}$ (top), for $\gamma = 0.1, 0.2, 0.5, 0.7$ and for $L=8\text{(dashed)}, 12\text{(solid)}$.}
\label{fig:wavefunction}
\end{figure}
In the top panel of Figure \ref{fig:wavefunction}c, the periods where $\mathcal{D}/L > (1-q_{EA})/2$ are due to the fast increase and saturation of qudit variance on the timescale of $\sim 1/\gamma$, versus the slower decay of $q_{EA}$. The latter proceeds on slower timescales through a combination of qudit-spin flip transitions and the Ising interaction $g \spin{z}{i}\spin{z}{i+1}$. This interpretation of the spin glass order parameter bring new meaning to the results from ML-MCTDH. At least up to intermediate times, the localization length of the wavefunction in Hilbert space is stable up to $L=96$ when $\gamma \approx 0.106$.
That the system is localized in Hilbert space may be useful in improving the performance of ML-MCTDH in this regime. Currently, ML-MCTDH reduces the size of the full Hilbert space by restricting the dynamics on to a subspace created by uniformly random vectors. For very large systems, these randomly drawn vectors will be heavily weighted towards states farther away from the initial/reference state, since their numbers grow combinatorially quickly.
\section{Trivial limit and convergence of qudit variance}
\label{sec:appA}
The quench setup we examine in this paper -- in which we prepare the full system in an eigenstate of the $\gamma = 0$ Hamiltonian -- results in certain behaviors in the $\gamma\to 0$ limit which we will explore in this section. The reason for the existence of a well-defined limit in the dynamics is due to the off-diagonal coupling being the \textit{only} generator of dynamics in both the spins and the qudit at short times.
We demonstrate this limit by plotting observables rescaled by $\gamma^{-2}$ in Figure \ref{fig:scaledL8}. This scaling at small $\gamma$ converges the dynamics at short times up to $t\leq 10^2$. At small enough coupling, a plateau begins to appear at $t \sim 10^2$. We see hints of this in the MCTDH data for mutual information (Figure \ref{fig:intTime}c), for example, with the establishment of a plateau for $L=48$ and $80$ at times $10 \leq t \lesssim 10^2$. This behavior contrasts with the immediate increase of MI for smaller system sizes at $t\sim 10$.
\begin{figure}[h]
\centerline{\includegraphics[width=0.95\textwidth, trim={0 0 0 0}, clip]{L8scaled.pdf}}
\caption{Dynamics from the ``super-Neel'' state for $L=8$ across different $\gamma$ in the localized phase. Rescaling the \textbf{(a)} spin glass order parameter, \textbf{(b)} qudit variance, and \textbf{(c)} mutual information by the coupling collapses the dynamics at short to intermediate times. The asymptotic behavior in each case should be interpreted as defining a trivial limit analogous to Anderson localization vis-a-vis MBL. }
\label{fig:scaledL8}
\end{figure}
We take advantage of the vanishing coupling, which is scaled to zero with system size, to systematically reconstruct the dynamics using the method of multiple scales. To that end, we solve for the time evolution operator, artificially introducing new `independent' timescales $t$, $t' \equiv \gamma t$, $t'' \equiv \gamma^2 t$, $\ldots$, which allow for control over secular terms growing unboundedly as $\sim t$. Formally, the time development operator is expanded as $U(t) \equiv U_0(t, t', t'', \ldots) + \gamma U_1(t, t', t'', \ldots) + \gamma^2 U_2(t, t', t'', \ldots) + \ldots$ and we shall solve for the full evolution order-by-order. We note that we have made an important assumption that the only timescales of interest are $\sim \gamma^{-n}$. However, we are not interested in describing the dynamics for all $t$ at arbitrary $\gamma$ and $L$, but only up to the $t \sim \gamma^{-1}$ as $\gamma\to 0$.
The propagator evolves according to
\[ \frac{d}{dt} U(t) = -i (H_0 + \Omega \tau^z + \gamma H_1 \tau_x) U(t), \] where we define $\displaystyle \tau^z = \sum_{n=1}^d n|n\rangle\langle n|$, $\displaystyle \tau^+ = \sum_{n=1}^{d-1} |n+ 1\rangle\langle n|$, and $\tau^- = (\tau^+)^\dagger$. From these we construct $\tau^x = \tau^+ + \tau^-$ and $\tau^y = -i \tau^+ + i \tau^-$. With the addition of the new timescales, the time derivative now becomes
\[ \frac{d}{dt} U = \left( \frac{\partial}{\partial t} U_0 \right)
+ \gamma \left( \frac{\partial}{\partial t'} U_0 + \frac{\partial}{\partial t} U_1 \right)
+ \gamma^2 \left( \frac{\partial}{\partial t''} U_0 + \frac{\partial}{\partial t'} U_1 + \frac{\partial}{\partial t} U_2 \right)
+ \ldots \]
At the zeroth order, the equation of motion and its solution are
\begin{align*}
\frac{\partial}{\partial t} U_0 = -i (H_0 + \Omega \tau^z) U_0 && \Longrightarrow && U_0 = e^{-i (H_0 + \Omega \tau^z) t} U^{\text{int}}_0(t', t'', \ldots) \text{ such that } U^{\text{int}}_0(0, 0, \ldots) = 1.
\end{align*}
At first order,
\begin{align}
\frac{\partial}{\partial t'} U_0 + \frac{\partial}{\partial t} U_1 &= -i (H_0 + \Omega \tau^z) U_1 -i H_1 \tau^x U_0 \nonumber \\
\frac{\partial}{\partial t'} U^{\text{int}}_0 + \frac{\partial}{\partial t} U^{\text{int}}_1 &= -i e^{i (H_0 + \Omega \tau^z) t} H_1 \tau^x e^{-i (H_0 + \Omega \tau^z) t} V_0, \label{eq:appA_firstOrder}
\end{align}
where we let $U_1 = e^{-iH_0 t} U^{\text{int}}_1$. Note at this point that the first term on the LHS is independent of $t$. Its contribution to $U^{\text{int}}_1$ would be proportional to $t$ and thus secular. Should there also exist secular terms on the RHS (i.e., independent of $t$), $U^{\text{int}}_0$ should be chosen to offset it. Otherwise, it must be independent of $t'$, i.e.\ $U^{\text{int}}_0 \equiv V_0(t'', \ldots)$. If these secular terms do not exist at all orders, then the multiple scales result would be completely equivalent to the Dyson series in the interaction picture. In our disordered system, we must be careful of secularity and near-secularity. The former, in which two states linked by the perturbation are exactly degenerate, occurs with zero probability since the local field must have a value such that $\Omega = \pm 2(h_i + g(\spin{z}{i-1}+\spin{z}{i+1}))$. More likely is the scenario of near-degeneracies, which at this order can lead to $U^{\text{int}}_1$ growing arbitrarily large after an arbitrarily long time. Such terms make the expansion of $U$ uncontrolled at long times. We can, however, absorb near-secular behavior into $U^{\text{int}}_0$.
Define
\[ A(t') = \sum_{\substack{|a\rangle, |b\rangle \\ |E_a - E_b| < 1}} \lrangle{a | H_1 \tau^x | b} \exp\left(i \frac{\Delta E_{ab}}{\gamma} t'\right) |a\rangle\langle b|. \] To regulate the secular part of (\ref{eq:appA_firstOrder}), we must have $\frac{\partial}{\partial t'} U^{\text{int}}_0 = -i A(t') U^{\text{int}}_0$. This makes \[ U^{\text{int}}_0(t', t'',\ldots) = \exp\left(-i \int\limits_0^{t'} A(\tau) d\tau \right) V_0(t'', \ldots). \] The unknown function $V_0$ will be solved for at higher orders. The argument of the exponential should always be complex, since $A(t')$ is Hermitian. Thus these resonant terms will not cause $U^{\text{int}}_0$ to have unbounded norm, and the perturbative expansion for $U$ remains valid. However, our ability to regulate the secularity in this way should not be taken as a statement on the dynamics being localized. Instead, it implies that more careful consideration of $U^{\text{int}}_0$ is necessary to understand if resonances are able to cause delocalization. At present, all we need is to examine if we can safely neglect $\exp(-i \int^{\gamma t}_0 A)$ if we scale $\gamma \to 0$ with the inverse system size. We can think of $A$ as roughly being the adjacency matrix for states in Hilbert space, where two states are connected by an edge if they are resonant. It is known that the eigenvalues of adjacency matrices are bounded above by the maximum number of edges connecting to a vertex in the graph. A naive upper bound for our system is $2L$, which is the number of states that can be reached by applying the perturbation on to the eigenstates of the unperturbed system. Should the spectrum of $A$ saturate this bound, the matrix exponential will contain time dependence going as $\exp(-i 2 \gamma t L)$ and scaling $\gamma \propto L^{-1/2}$ will not remove the correction factor in $U^{\text{int}}_0$. Regardless, this will not pose a problem in our model for the parameters and initial super-Neel state we have chosen.
Having found the lowest order approximation for $U^{\text{int}}_0$, we are now left with the nonsecular terms. These can be straightforwardly used to solve for $U^{\text{int}}_1$. Let $(H_1 \tau^x)_{\text{reg}}$ be the regular version of the perturbation $H_1 \tau^x$ with resonant matrix elements removed. The equation of motion becomes
\begin{align*}
\frac{\partial}{\partial t}U^{\text{int}}_1 &= -i \, e^{i (H_0 + \Omega \tau^z) t} (H_1 \tau^x)_{\text{reg}} e^{-i (H_0 + \Omega \tau^z) t} \, U^{\text{int}}_0(t', \ldots) \\
\Longrightarrow \qquad U_1 &= -i \, e^{-i (H_0 + \Omega \tau^z) t} \left\{ i V_1(t', t'', \ldots) + \left( \int\limits_0^t e^{i (H_0 + \Omega \tau^z) \tau} (H_1 \tau^x)_{\text{reg}} e^{-i (H_0 + \Omega \tau^z) \tau} d\tau \right) U^{\text{int}}_0(t', \ldots) \right\}
\end{align*}
While we can in principle keep going to higher orders, we will stop here and discuss the dynamics with scaled qudit-spin chain coupling. The results we have obtained so far allow us to accurately describe time evolution up to a timescale $t \sim O(1/\gamma)$. Should we keep decreasing $\gamma$, then all the artificial times $t'$, $t''$, $\ldots$, will tend to zero without affecting the physical time $t$. Using the initial condition for the unknown functions can then give us closed form expressions for $U$. For example, we approximate
\[ U(t) \approx e^{-i (H_0 + \Omega \tau^z) t} + \gamma \, e^{-i (H_0 + \Omega \tau^z) t} \left(-i \int\limits_0^t e^{i (H_0 + \Omega \tau^z) \tau} (H_1 \tau^x)_{\text{reg}} e^{-i (H_0 + \Omega \tau^z) \tau} d\tau \right). \] This is essentially what one finds in usual perturbation theory, except we now have more knowledge of its convergence properties.
We can gain some analytical understanding of the qudit variance and the spin glass order parameter from this approximation to the propagator.
Numerics show that the dynamics of qudit variance $\Delta^2_{\mathcal{Q}} = \lrangle{\tau(t)^2} - \lrangle{\tau(t)}^2$ comes mostly from the first term, as the second term is essentially constant. We calculate $\frac{d}{dt} \Delta^2_{\mathcal{Q}} \approx \lrangle{\acomm{\tau^z}{\frac{d}{dt}\tau^z}}$, where $\frac{d}{dt} \tau^z = \gamma H_1 \tau^y$.
\begin{align*}
\frac{d}{dt} \Delta^2_{\mathcal{Q}} &\approx \gamma \lrangle{\psi | \acomm{\tau^z(t)}{H_1(t) \tau^y(t)} |\psi} \\
&= \gamma \cancelto{0}{\lrangle{\acomm{{\tau^z}^{(0)}}{ H^{(0)}_1 {\tau^y}^{(0)}}} } + \gamma^2 \lrangle{\acomm{{\tau^z}^{(1)}}{ H^{(0)}_1 {\tau^y}^{(0)}}} + \gamma^2 \lrangle{\acomm{{\tau^z}^{(0)}}{ \left( H^{(1)}_1 {\tau^y}^{(0)} + H^{(0)}_1 {\tau^y}^{(1)} \right)}}.
\end{align*}
The first term is zero since the operator is off-diagonal. The last term must also vanish since it is not invariant with respect to redefinition of $\tau^z$, e.g.\ changing its spectrum from $(0, \ldots, d-1)$ to $(1, \ldots, d)$ by adding a constant term to the Hamiltonian. Indeed one can check that it is exactly cancelled by the $- \lrangle{\tau^z}^2$ term we have neglected in our approximation of the qudit variance. A calculation of the remaining term shows
\begin{align}
\label{eq:appA_qdVarRate}
\frac{d}{dt} \Delta^2_{\mathcal{Q}} &\approx 2 \gamma^2 \sum_i \sum_{\pm} \frac{\sin(t \Delta E^{\pm}_i)}{\Delta E^{\pm}_i} \lrangle{\psi^{\pm}_i \middle| \left( \spin{x}{i} \tau^x \right)_{\text{reg}} \middle| \psi},
\end{align}
consistent with usual perturbation theory. Upon disorder averaging, we see that
\begin{align*}
\overline{\Delta^2_{\mathcal{Q}}} &\approx 2 \gamma^2 L \int\limits_0^t d\tau \, \overline{ \sum_{\pm} \frac{\sin(\tau \Delta E^{\pm}_i)}{\Delta E^{\pm}_i} \lrangle{\psi^{\pm}_i \middle| \left( \spin{x}{i} \tau^x \right)_{\text{reg}} \middle| \psi}}.
\end{align*}
Thus the qudit variance -- along with other qudit observables such as the population -- converge to a single curve upon scaling the coupling as $\gamma \propto 1/\sqrt{L}$. Convergence towards this expression should be expected up to time $t \sim O(\gamma^{-1}) \propto O(\sqrt{L})$. For dynamics from the super-Neel state as we study here, the energy difference with spin $i$ flipped is $\Delta E_i = \pm 2 h_i$. There are no resonances for our chosen values of $h_i \in [-1.3, 1.3]$ and $\Omega \approx 3.93$, so we have exactly
\begin{align*}
\Delta^2_{\mathcal{Q}} &\approx 2 \gamma^2 \sum_i \sum_{\pm} \frac{1-\cos\left(t (2 h_i s_i \pm \Omega)\right)}{(2 h_i s_i \pm \Omega)^2} \\
\frac{d}{dt}\overline{\Delta^2_{\mathcal{Q}}} &\approx \gamma^2 L \left( \frac{\operatorname{Si}[t (2h + \Omega)] + \operatorname{Si}[t (2h - \Omega)]}{h} \right),
\end{align*}
where $\operatorname{Si}(t)$ is the sine integral. For fixed $\gamma$ at large enough $L$, this expression will violate the bound on the qudit variance, $(d^2-1)/12$, when all the states of the qudit are equally populated. Thus higher order terms are necessary to prevent this unphysical outcome.
We can similarly look at the spin glass order parameter, and find that for the super-Neel state,
\[ \frac{d}{dt} (1-q) = 4 \gamma^2 \frac{1}{L} \sum_i \sum_{\pm} \frac{\sin(t \Delta E^{\pm}_i)}{\Delta E^{\pm}_i}. \]
This shows the surprising fact that the qudit variance (cf.\ \eqref{eq:appA_qdVarRate}) and spin glass order are linearly related to each other in this limit. This motivates the scaling we take in plotting the results in Figures \ref{fig:shortTime} and \ref{fig:intTime} in the main text.
\section{Ansatz for qudit populations in the thermalized phase}
\label{sec:randMat}
We consider eigenstate thermalization in the sense that
\[ \lrangle{\psi | A |\psi } = Z^{-1} \operatorname{Tr} \left[ e^{-\beta H} A \right], \]
where $\beta$ is the temperature reproducing the same energy $\lrangle{\psi | H |\psi}$. When we time evolve from an initial state sitting at the middle of the many-body spectrum, $|\psi(0)\rangle = |\uparrow \uparrow \downarrow \downarrow \ldots\rangle \big| \frac{d-1}{2}\big\rangle$, we should consider infinite temperature averages, i.e.\
\[ \lrangle{\psi(t) | A |\psi(t) } = d^{-1} \operatorname{Tr}_{qd} \left( 2^{-L} \operatorname{Tr}_S A \right). \] Therefore by setting $A = |n\rangle\langle n|$ for $n=(1-d)/2,\ldots,(d-1)/2$, we should expect a uniform occupation over all $d$ levels of the qudit. We do not observe this in the delocalized phase; instead, the occupations of the qudit seem to saturate to the distributions shown in Fig.\ \ref{fig:occs} (averaging over $\sim 10$ realizations of disorder).
\begin{center}
\begin{figure}[H]
\includegraphics[width=\columnwidth]{occs.pdf}
\caption{Occupations of the qudit at $t\sim 10^2 g^{-1}$. On the right, the occupations have been vertically spaced by $0.05$ for clarity.}
\label{fig:occs}
\end{figure}
\end{center}
On the left, we notice that the occupations appear to be asymmetric about the middle state. This is possibly due to the small number of disorder averages, but may also be affected by finite size effects or the rather special super-Neel initial state. As a proxy for the infinite sample limit, we symmetrize the occupations and note that a one parameter Gaussian ansatz, $p_n \propto \exp \left( -n^2/(2\sigma^2) \right)$, fits well.
We propose a model to reproduce these observations based on the partial diagonalization achieved by invoking Floquet's theorem. By going into the rotating frame through the transformation $\exp \left( - i \Omega \hat{n} t \right)$, the Hamiltonian becomes time-periodic and we can factorize the time evolution in the ``lab frame'' as
\begin{align*}
U(t) &= e^{-i \Omega \hat{n} t} e^{-i K_{\text{rot}}(t)} e^{-i H_{\text{rot}}^{\text{eff}} t} e^{i K_{\text{rot}}(0)}.
\end{align*} We had previously proven that eigenstates can be written in the form
\[ |E\rangle = e^{-i K_{\text{rot}}(0)} |\varepsilon_i\rangle |m\rangle, \qquad E = \varepsilon_i + m\Omega \text{ where } H_{\text{rot}}^{\text{eff}}|\varepsilon_i\rangle |m\rangle = \varepsilon_i |\varepsilon_i\rangle |m\rangle \text{ and } \hat{n}|m\rangle = m|m\rangle \] under the assumptions that (1) the effective Hamiltonian $H_{\text{rot}}^{\text{eff}}$ commutes with $\hat{n}$ and (2) the operator \[ e^{-i \Omega \hat{n} t} e^{-i K_{\text{rot}}(t)} e^{i \Omega \hat{n} t} \] is an analytic function of time $t$. While explicit expressions for $K_{\text{rot}}$ and $H_{\text{rot}}^{\text{eff}}$ can be obtained using the high frequency expansion, the above expressions should hold even when the HFE does not converge, so long as the stated assumptions are satisfied.
By Floquet's theorem, the kick operator $K_{\text{rot}}(t)$ must be time periodic with frequency $\Omega$. Hence it should be representable in a Fourier series in powers of $e^{-i \Omega m t}$. Because in the rotating frame, factors of $e^{\pm i \Omega t}$ are accompanied by the corresponding qudit raising/lowering operator $\tau^{\pm}$, we posit that terms in the Fourier series with $\exp(i \Omega m t)$ should induce transitions between qudit states separated by (signed distance) $m$. The kick operator should then decompose into
\[ K_{\text{rot}}(t) = \sum_{m=1}^{d-1} \sum_{n=-(d-1)/2}^{-m+(d-1)/2} e^{i \Omega m t} |n+m\rangle\langle n| \, B_n^{n+m} + \text{h.c.}, \] where the operators $B^i_j$ act only on the spins. In the HFE, one sees that $B^i_j$ are imaginary and not necessarily Hermitian. We shall assume these properties still hold even when the HFE breaks down.
We shall model the effect of the kick operator on only the qudit states by supposing that matrix elements of $B^i_j$ between two delocalized spin states are random numbers, with possible dependence on $i-j$. In short, we propose the replacement
\[ \operatorname{Tr}_S e^{-i K_{\text{rot}}(0)} \rho e^{i K_{\text{rot}}(0)} \longrightarrow \overline{\exp(-i K) \rho_{qd} \exp(i K)}, \]
where $K$ is a $d\times d$ Hermitian random matrix whose upper triangular part (excluding the diagonal) looks like
\[ \left( K \right)_{mn} = i g \exp \left( - \alpha \left( \frac{m-n}{\gamma/\Omega} \right)^2 \right) R_{mn}, \] for random $R_{mn} \sim \text{Normal}(\mu=0,\sigma^2=1)$ and $g,\alpha > 0$. The factor of $\gamma/\Omega$ was inserted so that $\exp(- i K)$ would be the unit matrix in the decoupled and infinite frequency limits. The average over all realizations of $K$ mimics the nonunitarity of the partial trace over the spins $S$.
For example, we find good fits to the symmetrized occupations for the following values of the parameters, setting $g=1$:
\begin{center}
\begin{tabular}{m{1.5in} m{3.9in}}
\begin{ruledtabular}
\begin{tabular}[b]{@{}ccc@{}}
$L$ & $\gamma$ & $\alpha$ \\ \hline
16 & $0.7/\sqrt{16/12}$ & $1/105$ \\
16 & $0.7$ & $1/83$ \\
20 & $0.7/\sqrt{20/12}$ & $1/140$ \\
24 & $0.7/\sqrt{24/12}$ & $1/175$
\end{tabular}
\end{ruledtabular} &
\includegraphics[scale=0.7,trim={0 0pt 0 0pt}]{fitsymoccs.pdf}
\end{tabular}
\end{center}
We conjecture that the correct form in the limit of large $L$ is
\[ \left( K \right)_{mn} = i \exp \left( - \frac{c}{L} \left( \frac{m-n}{\gamma/\Omega} \right)^2 \right) R_{mn}, \] where $c$ is a positive number of order 1.
\section{Introduction}
The advent of controllable quantum simulation platforms allows for novel explorations of quantum coherent phenomena. Certain such architectures have the advantage of using extra degrees of freedom as a way to easily read out properties of a system \cite{Blais2004}. Examples of such setups include cavity QED with ultracold atoms \cite{Davis2019} and superconducting qubit circuits, the latter of which was recently used to simulate the many-body localized (MBL) phase in a 10 qubit chain with long-range interactions mediated by a central resonator \cite{Xu2018}. Given that such platforms are in their early stages, it is important to explore the interplay of disorder-induced localization and mediated long-range interactions, and how they affect the dynamics of localization in these systems.
If localization exists in these systems, it will naturally be many-body localization since the spins hybridize with the central degree of freedom to give non-trivial interactions. Rigorous results on MBL have already been established in one dimensional systems with short ranged interactions \cite{Imbrie2016}. In such a setting, it is a stable phase of matter, with respect to adding short range perturbations, that can coexist with other types of order \cite{Huse2013, Khemani2016}. While strong disorder enables localization, it cannot prevent thermalization if interactions are long-ranged, decaying slower than $r^{-2D}$, where $D$ is the spatial dimension \cite{Yao2014, Maksymov2020}. Even the MBL phase with short-ranged interactions is fragile. It is destroyed upon coupling to a continuum of bath modes \cite{Nandkishore2014} which, intuitively, can provide arbitrary amounts of energy and allow the system to transition between eigenstates of vastly different character. One sees then that there are two ingredients to this delocalization mechanism: a continuum of energies of large enough bandwidth, and hybridization due to effective infinite-ranged interactions mediated by the non-Markovian bath.
In fact, for a specific type of memoryless bath, nonergodicity does survive. This is the case of Floquet MBL, in which an MBL system is subjected to an external periodic drive with frequency $\Omega$ modeled as a time-dependent Hamiltonian acting on the system \cite{Abanin2016,Ponte2015prl}. The failure of thermalization is due to the the inability of the system to absorb energy in quanta of $\hbar \Omega$, which itself is a consequence of the discreteness of the energy spectrum. The external drive, however, is not inherently dynamical and thus does not capture the backaction present in a fully quantum mechanical system.
In this work, we consider the time evolution of such a system obtained by treating the Floquet drive as a quantum degree of freedom. Specifically, we consider a localized system globally coupled to a $d$-level system (qudit) with finite energy spacing, similar to \cite{Ng2019}. When the qudit is a two-level system, it was shown that localization does not survive at any finite coupling \cite{Ponte2017, Hetterich2018}. But when it is instead a $d>2$ level system, localization was argued to survive under certain conditions \cite{Ng2019}. It is not known, however, what dynamical signatures should be expected in such regimes since the geometry and spin-spin interactions in the system limits the efficiency of usual computational approaches using matrix product operators. We bridge this gap by numerically simulating the non-equilibrium dynamics at much larger system sizes than previously considered. This is done using the multilayer-multiconfigurational time-dependent Hartree (ML-MCTDH) method, which solves the Schr\"odinger equation using the time-dependent variational principle on the manifold of wavefunctions represented by certain tree tensor networks \cite{Wang2003,Manthe2008,MCTDHBook,Oriol2011,Wang2015}.
We furthermore explore the possibility that the additional degree of freedom can provide alternative, nondestructive diagnostics of localization. In experimental settings, the usual observables signaling nonergodic behavior are correlation functions such as the occupation imbalance between odd and even sites of the lattice \cite{Schreiber2015}. More sophisticated setups may attempt to perform tomographic measurements to reconstruct the reduced density matrix for a subsystem and show logarithmic growth of entanglement entropy \cite{Xu2018}, or to measure the energy spectrum of the system in order to retrieve energy level spacing statistics \cite{Roushan2017}. Though these metrics serve as gold standards in characterizing MBL, the latter two methods are difficult to scale with larger systems. In our model, since quantum fluctuations of the spins necessarily involve the qudit, there may be signatures of (de)localization imprinted into the qudit dynamics. Such a possibility has been explored in autocorrelations of qudit observables \cite{Hetterich2018} probing the energy level statistics, as well as dynamics of the occupation number \cite{Sierant2019} by measuring the light intensity output by a single mode cavity. In this work we show that the qudit qualitatively changes the spin chain dynamics, and elucidate the timescale on which this occurs. This provides some insight into the breakdown of localization, and the possible role that non-Floquet physics may play in it.
The structure of this paper is as follows: we will discuss our model and its localization in connection to Floquet MBL; review the essentials of ML-MCTDH, which we then apply to study intermediate time dynamics; present results on thermalizing and nonthermalizing behaviors in dynamical metrics; and discuss what may be expected in experiments, where control over the central coupling may be limited in range.
\section{Model}
\label{sec:model}
We consider a simplified model of many-body localization by coupling a one-dimensional chain of qubits (spins-$1/2$) via global interactions with a central qudit:
\begin{align}
\label{eq:model}
H &= H_0 + \Omega \hat{\tau}^z + \gamma H_1 \left( \hat{\tau}^+ + \text{h.c.} \right), \\
H_0 &= \sum_{i=1}^L h \xi_i \spin{z}{i} + g \spin{z}{i} \spin{z}{i+1}, \qquad H_1 = \sum_{i=1}^L \spin{x}{i},
\nonumber
\end{align}
where $\hat{\tau}^z = \sum_{n=1}^d n |n\rangle\langle n|$, $\hat{\tau}^+ = \sum_{n=1}^{d-1} |n+1\rangle\langle n|$, and the operators $H_0$ and $H_1$ act only on the spin subspace. The states $|n\rangle$ label the states of the central qudit. Here, $h = 1.3$, $g = 1.07$, $\Omega = \pi/0.8$, and $\xi_i$ is a random variable drawn uniformly from $(-1,1)$. When the model with these parameters is mapped on to the corresponding Floquet system (i.e., $d\to\infty$), it shows a localization-delocalization transition at a critical coupling $\gamma_c \lesssim 0.3$ \footnote{This estimate is subject to strong finite size effects}. We restrict our discussion to $\gamma$ either deep in the localized phase ($\gamma < 0.2$) or deep in the ergodic phase ($\gamma\approx 1$). Finally, throughout this paper we restrict ourselves to central qudit size $d=7$, which is large enough to display Floquet-like behavior but small enough that the finite qudit size plays an important role.
The spin part of the Hamiltonian, $H_0$, is a trivial antiferromagnetic Ising chain with longitudinal on-site disorder. The diagonal nature of $H_0$ in the $z$-basis yields trivial localization in the eigenstates. This manifests in eigenstates $|\psi_n\rangle$ as vanishing site-averaged magnetization $L^{-1}\sum_i \lrangle{\psi_n | \spin{z}{i} | \psi_n}$ and maximal value of the spin-glass parameter, $q = L^{-1}\sum_i \lrangle{\psi_n | \spin{z}{i} | \psi_n}^2 = 1$ at high energy densities. Values of $q\approx 1$ suggest that the eigenstates are described mostly by a single pattern of magnetization. Introducing a small coupling to the qudit without longitudinal disorder induces hybridizations that push $q\to 0$. We find that it is necessary to have both qudit coupling and strong disorder to preserve the nonergodicity when probing the system in the middle of the many-body spectrum, where the density of states (DOS) is the greatest.
Several features distinguish our model from those studied previously. While Nandkishore et al.\ \cite{Nandkishore2014} coupled a ``fully MBL'' system to an interacting bath of bosons, the qudit we present here is not bath-like because it does not have a continuous DOS. The model of thermal inclusions studied by Ponte et al.\ \cite{Ponte2017} closely resembles ours, but crucially we place a constant ``magnetic field'' $\Omega \hat{\tau}_z$ on the qudit, thus selecting a preferred direction for the central spin. This greatly impacts the ease with which the qudit fluctuates, which in turn can regulate transitions in the spin states leading to delocalization.
Recent studies have examined how localization can persist in the presence of long-ranged interactions \cite{Sierant2019, Maksymov2020} or with central coupling to a single degree of freedom yielding an effective Hamiltonian with long-ranged interactions \cite{Ponte2017, Hetterich2018}. With the exception of a numerical study \cite{Sierant2019}, these past works have noted that preservation of localization in the thermodynamic limit requires increasing the disorder strength with increasing system size or decreasing the strength of central coupling as $\gamma\to \gamma/L$. Reducing the coupling strength in this way renders the long-ranged part of the effective Hamiltonian for the spin chain subextensive. This is also reflected in the dynamics of the qudit as its transition rate vanishes.
On the other hand, the existence of Floquet MBL affords a different pathway to the coexistence of localization and central coupling. In that context, the persistence of MBL is not due to a vanishing coupling to the external drive, but to a suppression of mixing between different localized eigenstates of the undriven system. This picture suggests that an effective Hamiltonian for only the spin degrees of freedom should show localized behavior. This is indeed the case, as previous work based on the high frequency expansion has shown \cite{Ng2019}. In this limit of $\Omega \to \infty$ the spins are governed by an effective Hamiltonian diagonal in the qudit basis, reproducing the eigenenergies modulo an integer multiple of $\Omega$:
\begin{align}
\label{eq:HFE}
H_{\text{eff}} = H_0 + (H_1)^2 \frac{|d\rangle\langle d| - |1\rangle\langle 1|}{\Omega} + O(\Omega^{-2}).
\end{align}At lowest order in $\Omega^{-1}$, we see that possible delocalization is reserved only for states with $|1\rangle$ or $|d\rangle$, as $(H_1)^2$ induces all-to-all coupling. Increasing $L$ without increasing $d$, as we do in this paper, means that eigenstates occupying $|1\rangle$ will eventually encroach upon the middle of the spectrum and contribute to the quench dynamics we study. This can be seen from the density of states when $H_0$ is dominant as it follows $\rho(E) \propto \exp \left( - \frac{E^2}{(J \sqrt{L})^2} \right)$ for energy scale $J\sim O(1)$, meaning $\rho(E)$ will grow wider with increasing $L$. An energetically dominant $(H_1)^2$ term will both delocalize the eigenstates and deform the Gaussian density of states in the thermodynamic limit.
\begin{figure}[h]
\includegraphics[width=0.95\columnwidth]{phaseDiagram.pdf}
\caption{Schematic phase diagram for the coupled system \eqref{eq:model}, along with the parameters for which we present numerical results from ML-MCTDH. The rough phase boundaries are determined from numerics and analytical arguments. The data are separated into three solid segments -- the strong, intermediate and weak couplings from top to bottom. The angle of the segments comes from fixing the qudit size to $d=7$ and scaling the central coupling $\gamma \propto L^{-1/2}$. For ease of discussion, we group the three coupling regimes as region I (weak and intermediate) and region II (strong).}
\label{fig:phaseDiagram}
\end{figure}
We thus assume $\gamma$ to be small enough such that neither outcome occurs, and ask when this picture will naively break down. In such a limit, we can treat the $H_1^2$ field term in a mean field fashion for each eigenstate:
\[ H_{\text{eff}} \approx H_0 + \frac{\gamma^2 L}{\Omega} + \sum_i \frac{\gamma^2}{\Omega}\lrangle{\sum_{j\neq i} \spin{x}{j}} \spin{x}{i}, \] where the effective field $\lrangle{\sum_{j\neq i} \spin{x}{j}}$ in an eigenstate must be determined self-consistently. For a typical eigenstate, this field should have value $\sim f(\gamma) \sqrt{L}$, where $f(\gamma)$ must vanish when $\gamma=0$. This is the case when $\sum_j \lrangle{\spin{x}{j}}$ is the sum of $L-1$ independent random variables, and the finite $\gamma$ eigenstates are assumed to be perturbatively connected to a corresponding $\gamma=0$ eigenstate. For this model, we take the lowest order approximation $f(\gamma) \approx f_1 \gamma$. With this assumption \footnote{See Supplemental Material at URL for discussion on the cumulant expansion, an ansatz for the steady state of the thermalizing phase, and perturbation theory using multiple scales.}, the effective transverse field on site $i$ will begin to compete with the longitudinal fields in $H_0$ when $\gamma^2 \lrangle{\sum_{j\neq i} \spin{x}{j}} \sim O(g, h_i) \sim O(1)$. For the high energy density eigenstates we are interested in, this effective field will inhibit spin glass ordering and the system should obey the eigenstate thermalization hypothesis. Thus, $\gamma \propto L^{-1/6}$ should serve as a rough separatrix between thermalizing and athermal behaviors. Furthermore, couplings that tend to zero faster than $L^{-1/6}$ will realize a trivial limit, where the localization comes entirely from $H_0$.
The region where this is argument expected to be most significant is denoted in Fig.\ \ref{fig:phaseDiagram} through a color gradient starting around $d/\sqrt{L} \sim 1$.
Note that, in general models where $H_1$ includes operators diagonal in the z-basis, we would have $f(0) \neq 0$; in this case the scaling is replaced by $\gamma \sim L^{-1/4}$.
Besides scaling the coupling to zero, the all-to-all interactions can be avoided by ensuring that eigenstates occupying levels $|d\rangle$ or $|1\rangle$ in the qudit do not participate in the dynamics. For quenches starting from the middle of the many-body spectrum, this condition can be ensured by keeping the qudit size $d$ sufficiently large compared to the typical width of the 1D many-body density of states, $\sqrt{L}$. Dynamics in this limit should closely resemble Floquet physics, since the fluctuations producing effective long-ranged interactions will cancel out after accounting for the processes in which the intermediate qudit state changes by $+1$ or $-1$. Away from this limit, when $d/\sqrt{L} \lesssim O(1)$, the all-to-all interactions are unavoidable. The threshold value of $d/\sqrt{L}$ for delocalization should decrease as the coupling is decreased. These arguments are summarized schematically in Fig.\ \ref{fig:phaseDiagram}.
\begin{figure*}
\begin{minipage}{0.45\textwidth}
\begin{align*}
\ket{\Psi} &= \sum_{j_1=1}^{N_{1}}... \sum_{j_P=1}^{N_{1}} A_{j_1,...,j_P}(t) \prod_{\kappa=1}^P \ket{\varphi_{j_\kappa}^{(\kappa)}(t)}, \\
\ket{\varphi_{j_\kappa}^{(\kappa)}(t)} &= \sum_{i_1=1}^{N_{2}}... \sum_{i_{Q(\kappa)}=1}^{N_{2}} B_{i_1,...,i_{Q(\kappa)}}^{\kappa, j_\kappa}(t) \prod_{q=1}^{Q(\kappa)} \ket{\nu_{i_q}^{(\kappa,q)}(t)}, \\
&..., \nonumber
\end{align*}
\end{minipage}
\begin{minipage}{0.45\textwidth}
\includegraphics[scale=0.6]{tree_ring.pdf}
\end{minipage}
\caption{Expansion of the wave function $\ket{\Psi}$ and the first layer single-particle functions $\ket{\varphi_{j_\kappa}^{(\kappa)}(t)}$ used in the ML-MCTDH approach \textbf{(left)} and a schematic representation of the tree structure of the wave function \textbf{(right)}. The black dots represent single-particle functions (SPFs). The red dot represents the qudit degree of freedom and the blue dots represent the spin degrees of freedom. The binary expansion of the spin wave function is symmetric and, thus, we choose the numbers of SPFs within one layer to be equal. In the example shown, only three spins are grouped together in the lowest layer for better visualization. In the calculation, however, groups of up to 12 spins in the lowest layer are used.}
\label{fig:mctdh_tree}
\end{figure*}
There are two important ways to think of this system and its dynamics: either as a combined many-body systems with localized and delocalized phases, as was done in the previous paragraph, or as a central qudit interacting with an unusual, localized, spin bath. From this latter viewpoint, it will be useful to consider scaling the system bath coupling $\gamma \sim 1/\sqrt{L}$, since that will be shown to achieve a well-defined thermodynamic ($L\to\infty$) limit. This scaled coupling will be used in the majority of our simulations, and is covered in more detail in Section \ref{sec:scaledGamma}. For now, we note that $\gamma \sim L^{-1/2}$ scales to zero faster than the $L^{-1/6}$ that we predict is required for MBL. Therefore, at sufficiently late times, we predict MBL with our scaled coupling.
In this model we use qudits for numerical simplicity due to their finite Hilbert spaces. However, our conclusions can be easily applied also to the case where the central degree of freedom is a single bosonic mode, such as in cavity QED or superconducting circuits. In these setups we expect similar dynamical behaviors when the central coupling is appropriately scaled \cite{Ng2019}.
\section{Numerical method}
\label{sec:numerics}
\begin{figure*}
\centerline{\includegraphics[width=0.99\textwidth, trim={0 0 0 0}, clip]{dynamics_shortTimes.pdf}}
\caption{Dynamics from an initial ``super-Neel'' state in the regimes of weak (top row, $\gamma =0.1/\sqrt{L/12}$) and intermediate coupling (bottom row, $\gamma =0.3/\sqrt{L/12}$). The data is averaged over $O(10^2)$ -- $O(10^3)$ disorder realizations, with the shaded bands indicating deviations of $\pm 1$ standard error of the mean.
\textbf{(left)} Variance of qudit occupations $\Delta^2_Q = \lrangle{(\tau^z)^2} - \lrangle{\tau^z}^2$, \textbf{(center)} deviation from perfect spin glass order $1-q_{EA}$, and \textbf{(right)} entanglement entropy $S_A$ between a contiguous half of the spin chain and its complement. The observables $\Delta_{\mathcal{Q}}^2$ and $1-q_{\text{EA}}$ have been appropriately rescaled \cite{Note2} to show their coincidence (except a factor of 2) at early times for weak coupling. The dynamics are observed to converge to a single curve (black dotted line) and appear to be consistent with the dynamics without the nearest neighbor Ising coupling (dashed lines) as $L\to\infty$.}
\label{fig:shortTime}
\end{figure*}
The non-local interaction induced by the centrally coupled qudit makes the simulation based on matrix product operator techniques like time-evolving block decimation \cite{Paeckel2019} inefficient. And while alternative approaches such as the Floquet-Keldysh DMFT \cite{Lubatsch2019} exist, they are valid only in the well-studied Floquet limit in which the interesting mediated all-to-all couplings are negligible. Thus, we instead employ the Multilayer Multiconfiguration Time-Dependent Hartree (ML-MCTDH) method \cite{Wang2003,Manthe2008,Oriol2011,Wang2015, HeidelbergMCTDH} which has been used to study similar systems in the past, e.g.\ a two-level system coupled to a bath of noninteracting spins \cite{Wang2012}. The ML-MCTDH method generalizes the original MCTDH method \cite{Meyer1990,man92:3199,Beck2000,MCTDHBook,mey12:351} for applications to significantly larger systems. The ML-MCTDH approach represents a rigorous variational basis-set method, which uses a multiconfiguration expansion of the wave function, employing time-dependent basis functions and a hierarchical multilayer representation. Within this framework the wave function is recursively expanded as a superposition of Hartree products as depicted in Fig. \ref{fig:mctdh_tree}. Here, $\ket{\varphi_{j_\kappa}^{(\kappa)}(t)}$, $\ket{\nu_{i_q}^{(\kappa,q)}(t)},\ldots,$ are the so-called ``single-particle functions'' (SPFs) for the first, second, etc.\ layer and the coefficients $A_{j_1,...,j_N}$, $B_{i_1,...,i_{Q(\kappa)}}^{\kappa, j_\kappa}$ are the expansion coefficients of the first, second, etc.\ layer. Despite their name, the SPFs describe multiple degrees of freedom, see Fig.\ \ref{fig:mctdh_tree}. The ML-MCTDH equations of motions for the expansion coefficients and the single-particle functions are obtained by applying the Dirac-Frenkel variational principle \cite{Wang2003, Wang2009}, thus ensuring convergence to the solution of the time-dependent Schrödinger equation upon increasing the number of SPFs. In principle, the recursive multilayer expansion, which corresponds to a hierarchical tensor decomposition in the form of a tensor tree network, can be carried out to an arbitrary number of layers. In practice, the multilayer hierarchy is terminated at a particular level by expanding the single-particle functions in the deepest layer in terms of time-independent basis functions.
In the present application of the ML-MCTDH method, we separate the qudit wave function and the spin chain wave function in the uppermost layer as depicted schematically in Fig.\ \ref{fig:mctdh_tree}. The wave function of the spin chain is then further expanded in a binary tree (i.e.\ $P=Q=2$) up to the lowest layer, which comprises blocks of up to 12 spins. Each of the lowest blocks is expanded in the time-independent local basis of the underlying Hilbert space. Regarding the number of SPFs in the first layer, $N_1$, it can be shown that $N_1>d$ leads to redundant configurations in the expansion \cite{Meyer1990}, and thus, we set $N_1=d$ in all calculations. The required number of SPFs in the other layers of the expansion of the spin-chain wave function was determined by thorough convergence tests and depends on the coupling strength $\gamma$. In general, fewer SPFs are needed for smaller coupling strengths. For L=24, two dynamical layers are employed and the required number $N_2$ of SPFs varies from 30 to 120 SPFs. For L=48, a three layer scheme is used where the number of SPFs in the lowest layer varies from 10 to 30 and in the highest layer from 20 to 60 SPFs. For L=96, four layers are employed with SPFs which vary from 10 to 20 in the lowest and from 35 to 50 in the highest layer.
\section{Results for scaled coupling}
\label{sec:scaledGamma}
We examine the system at infinite temperature by focusing on states in the middle of the many-body spectrum, which have energies close to the midpoint between the maximal and minimal energies of the coupled system, $(E_{\text{max}}$ and $E_{\text{min}})$ respectively. We take $\gamma=0$ for $t<0$ with the spins in a ``super-Neel'' state $|\downarrow \downarrow \uparrow \uparrow \ldots \rangle$ and the qudit occupying its middle state $|(d+1)/2\rangle$. The coupling is switched on instantaneously at $t=0$ to a finite value. The super-Neel state is on average a zero energy eigenstate of $H_0$ and has subextensive energy variance, making it a suitable microcanonical probe. Thus, when the system is thermalizing and shows ensemble equivalence, we expect similar dynamics compared to ones obtained through averaging over random initial product states, mimicking an infinite temperature canonical ensemble.
As there are different dynamical behaviors in our model, we shall organize our discussion around the schematic phase diagram in Fig.\ \ref{fig:phaseDiagram}, similar to the one first introduced in \cite{Ng2019}. In this first section, we will consider scaling the coupling as $\gamma \sim 1/\sqrt{L}$, corresponding to the three solid lines in the phase diagram which, from top to bottom, will be referred to as the strong, intermediate, and weak coupling regimes. The orientation of these cuts comes from the $1/\sqrt{L}$ scaling of $\gamma$. This is natural if we think of the qudit as our main object of interest, as it gives a well-defined thermodynamic limit for the qudit when it is coupled to a non-interacting bath, such as in the spin-boson model \cite{Weiss2012}. This scaling reproduces the Kac prescription \cite{Kac1963} for the all-to-all term in the effective Hamiltonian ensuring also the existence of a thermodynamic limit for the spins. Specifically, we scale $\gamma$ using the following formula:
\begin{equation}
\gamma = \gamma_0 \sqrt{\frac{L_0}{L}}
\label{eq:gamma_scaling}
\end{equation}
where $L_0=12$ throughout for convenience, such that $\gamma=\gamma_0$ at $L=12$. $\gamma_0$ sets the overall strength of the coupling. We will consider three regimes, indicated by the solid lines in Fig.\ \ref{fig:phaseDiagram}: weak coupling ($\gamma_0=0.1$), intermediate coupling ($\gamma_0=0.3$), and strong coupling ($\gamma_0=0.7$).
\begin{figure*}
\centerline{\includegraphics[width=0.99\textwidth, trim={0 0 0 0}, clip]{dynamics_scaled.pdf}}
\caption{Same as in Fig.\ \ref{fig:shortTime}, but with time rescaled by the system size-dependent coupling $\gamma $. Grey dots are independent calculations using the kernel polynomial method, aimed to extend the maximum time from $\Omega t/(2\pi) \sim 6\times 10^2$ to $\Omega t/(2\pi) \sim 1.9 \times 10^3$. Between the weak \textbf{(a,b,c,d)} and intermediate \textbf{(e,f,g,h)} coupling regimes, there is a qualitative shift in the long time behavior of both the qudit- and spin-only observables. This data is suggestive of logarithmic growth in the entanglement entropy becoming the dominant characteristic after $t \sim 1/\gamma ^2$. \textbf{(d,h)} Occupations $p_n$ of the qudit levels, symmetrized around the middle level $|n_{\text{mid}}\rangle = |4\rangle$. While both the weak \textbf{(d)} and intermediate \textbf{(h)} coupling regimes have most of their populations concentrated the initial occupied level, $|4\rangle$, the latter case has a much greater fraction of the total population in the extremes of the qudit's states. The values of $p_n$ for $|n-n_{\text{mid}}|=3$ in the left panel are too small ($\sim O(10^{-4}-10^{-3})$) for the scale.}
\label{fig:intTime}
\end{figure*}
\subsection{Weak and intermediate coupling (region I)}
\label{weakerCoupling}
The first cases we consider are weak and intermediate coupling, which are labeled Region I in Fig.\ \ref{fig:phaseDiagram}. These are both at sufficiently small $\gamma_0$ that we expect MBL for the largest accessible system sizes, but for intermediate coupling ($\gamma_0 = 0.3$) the system will be near the phase transition for small $L$. Three observables -- the qudit variance $\Delta_Q^2$ (Eq. \ref{eq:Delta_Q_def}), the spin glass order parameter $q_\mathrm{EA}$ (Eq. \ref{eq:q_EA_definition}), and the entanglement entropy of the half chain $S_A$ (Eq. \ref{eq:ent_entr_def}) -- are plotted in Figs.\ \ref{fig:shortTime} and \ref{fig:intTime}, which correspond to identical data with different scaling of the time axis. The origin of this scaling will be clarified shortly.
Note first that, by preparing both the qudit and the spins in highly excited states, one would normally expect the system to relax quickly to a featureless ``infinite temperature'' equilibrium. That is, all internal levels of the qudit should be equally occupied, and the spins should be paramagnetic and translationally invariant. This is not true for the disordered system we study, as the numerics demonstrate in Fig.\ \ref{fig:shortTime}: for sufficiently small coupling, the system shows localization in both the qudit and its the surrounding spins. The former is signaled by the variance of the qudit occupations
\begin{equation}
\Delta^2_{\mathcal{Q}} \equiv \lrangle{(\hat{\tau}^z)^2} - \lrangle{\hat{\tau}^z}^2,
\label{eq:Delta_Q_def}
\end{equation}
which saturates to a quantity far below that of the uniform limit, $\Delta^2_{\mathcal{Q}} = (d^2-1)/12 = 4$. Furthermore, the different system sizes exhibit scaling collapse of $\Delta^2_{\mathcal{Q}}$ up to a time scale $t \sim 1/\gamma $. This is a property of the scaled $\gamma$, as it implies that the spin chain acts as a bath for the qudit with a well-defined thermodynamic limit. More specifically, it can be shown that the considered model with scaled coupling $\gamma \propto L^{-1/2}$ fulfills linear response in the thermodynamic limit, meaning that the effect of the spin environment on the qudit is captured by the first two cumulants of the influence functional \cite{Feynman1963,Makri1999,Wang07}. For our model, the first cumulant vanishes and thus the reduced qudit dynamics is determined by the second cumulant, given by the force-force autocorrelation function of the spin chain. This also means that one can construct an effective harmonic bath whose correlation function is the same as that of the spin chain resulting in the same reduced qudit dynamics \cite{Makri1999}. For our model, the effective harmonic bath is characterized by a spectral density which depends in general on the initial state, the random local fields and the spin-spin coupling $g$. For the specific initial state considered here, the spectral density of the effective harmonic bath is equal to the probability distribution of twice the random local fields, and thus is independent of $g$.
Having established scaling collapse of the qudit variance, we now turn our attention to dynamics of the spin chain, starting with the spin glass order parameter
\begin{equation}
q_{\text{EA}}(t) \equiv L^{-1}\sum_i\lrangle{\psi|\spin{z}{i}(t)\spin{z}{i}(0)|\psi}.
\label{eq:q_EA_definition}
\end{equation}
Unlike the qudit variance, the spin glass order parameter displays marked drifts with system size (see insets of Fig.\ \ref{fig:shortTime}(b,e)). The tendency of $q_{\text{EA}}(t) \to 1$ comes from our choice of scaling $\gamma$, since $\gamma$ controls the strength of a local transverse field and thus governs the rate and magnitude of a single spin's precession. On reachable timescales $t\lesssim 10^2$, the largest system size $L=96$ has near perfect memory of the initial state. This behavior is consistent with our claim that the scaling of $\gamma \sim 1/\sqrt{L}$ towards zero with increased system size is sufficiently fast that the system will flow to MBL for arbitrary $\gamma_0$, although proving MBL would require evolution to much later times than we can access.
Though the usefulness of the influence functional approach is restricted to the qudit, we should -- by virtue of the fact that the initial spin dynamics are driven by interactions with the qudit (for initial product states like the super-Neel state we have chosen) -- find that the spin observables are linked to the qudit's. The spin observables should therefore enjoy a similar limiting behavior as $\gamma \propto L^{-1/2} \to 0$. We indeed show this to be the case within first order perturbation theory. In \cite{Note2}, we perform time-dependent perturbation theory using the method of multiple scales. We solve for the time evolution operator perturbatively by introducing new ``independent'' timescales $t$, $t' \equiv \gamma t$, $t'' \equiv \gamma^2 t$, $\ldots$, which allow for control over secular terms growing with $t$. In the thermodynamic limit with scaled coupling, we find that the dynamics of the qudit are described perturbatively to first order up to time $O(1/\gamma )$ (dotted lines in Fig.\ \ref{fig:shortTime}), providing a complementary approach to the linear response solution from the influence functional formalism. The perturbative calculation also demonstrates that spin observables should exhibit similar gradual convergence to a single limit up to timescales $t\sim O(1/\gamma )$. Remarkably, the connection between qudit variance and the spin glass order parameter is even more precise in this limit; they collapse to a single, universal curve in the thermodynamic limit upon scaling as $\Delta_\mathrm{Q}^2/\gamma_0^2$ and $(1-q_\mathrm{EA})L/(2\gamma_0^2)$, as seen in Fig.\ \ref{fig:shortTime}(a,b,d,e). Physically, this comes from the fact that a single perturbative excitation of the qudit through the $\hat{\tau}^+ + \hat{\tau}^-$ component of $H_1$ gives a single spin flip excitation of the spin chain through $\spin{x}{j}$.
Finally, we consider the entanglement entropy
\begin{equation}
S_A = -\mathrm{Tr}\left[\rho_A \log_2 \rho_A\right]
\label{eq:ent_entr_def}
\end{equation}
between a contiguous half of the spins with the rest of the system, which is a defining feature in many body localization. Here $\rho_A$ is the reduced density matrix of half of the spin system, e.g., sites $1$ through $L/2$. As with the previous two quantities, there appears to be a gradual convergence of $S_A$ to a universal curve with increasing $L$, although unlike the other observables, the entanglement depends on the strength of the coupling prefactor $\gamma_0$. By turning off the Ising interaction $g$ (dashed lines in Fig.\ \ref{fig:shortTime}), we see that the dynamics of entanglement at short times $ \lesssim O(1)$ are unchanged -- as predicted from time-dependent perturbation theory -- while growth of entanglement at intermediate times is dependent on this $\spin{z}{i}\spin{z}{i+1}$ interaction.
These observations about the short-time dynamics hold for both weak and intermediate coupling, as seen in Fig.\ \ref{fig:shortTime}. However, we can identify a slower timescale beyond $t \lesssim O(1/\gamma )$ from first order perturbation theory, on which the Ising interactions start to play a role. In Fig.\ \ref{fig:intTime}, the same data is plotted upon rescaling the time by $t/\gamma ^{-2}$. The observables are seen to roughly collapse for both the weak and intermediate couplings and, for intermediate couplings, entanglement in particular shows interesting intermediate time behavior. While the collapse is imperfect, we note a few salient features. First, deep in the localized (weak coupling) regime, the spread of the qudit occupation, the growth of bipartite entanglement entropy, and the decay of the spin-glass order parameter appear to be arrested at long times. It is unclear whether the observables will continue to grow at later times, but our data leaves open the possibility that they saturate and that the asymptotic value may be system-size independent under the chosen scaling. Second, the dynamics of the qudit appear to be correlated with dynamics of the spins, albeit with a slight time delay. Finally, in the intermediate coupling regime, the entanglement entropy continues to grow at late times. For $L=16$, there appears to be a logarithmic growth over three decades in rescaled time (see Fig.\ \ref{fig:intTime}g). The same may be true for the $L\geq 24$, but we have insufficient data to decisively prove slow growth over several decades. As seen in Fig.\ \ref{fig:intTime}h, the period of potentially logarithmic growth coexists with the period of finite occupation in the edge of the qudit spectrum (states $n=1$ and $d$), for which the high-frequency expansion yields all-to-all interactions (see Eq. \ref{eq:HFE}).
It is unclear what drives the logarithmic behavior. When focusing on the bipartite entanglement entropy, two generic mechanisms have been studied in recent years: the slow dephasing from a quench due to interactions between exponentially localized (quasi-local) operators \cite{Znidaric2008,Abanin2019}, and the linearly diverging semiclassical trajectories of the collective spin state \cite{Lerose2020} in long ranged interacting spin systems.
In the former case, it has been found that the slope of the logarithmic growth is independent of the strength of interactions \cite{Kjall2014}. This does not appear to be the case in our numerics, with the larger system sizes $L\geq 24$ ostensibly displaying log growth with a larger prefactor than in the $L=16$ case. Moreover, there does not appear to be any logarithmic trend when the system is deep in the localized phase (see top row of Fig.\ \ref{fig:intTime}). If conserved quasi-local operators do exist in this system, then our results would suggest that their localization lengths are strongly dependent on the coupling $\gamma$.
Another possibility for the appearance of logarithmic growth of $S_A$ could come from the mediated all-to-all interactions predicted in the effective Hamiltonian (Eq. \ref{eq:HFE}). In our qudit system, long ranged interactions begin to play a significant role when the extremal states of the qudit are occupied (see discussion in Sec.\ \ref{sec:model}). It was argued that these mediated interactions are responsible for the localization-delocalization transition upon decreasing $d/\sqrt{L}$, shown in Fig.\ \ref{fig:phaseDiagram}. Consistent with this, we see significantly greater occupation in the extremal qudit states for intermediate couplings -- where logarithmic growth is seen -- compared to weak couplings (see Fig.\ \ref{fig:intTime}(d,h)). It is also clear that the slow growth of $\Delta^2_{\mathcal{Q}}$ for intermediate couplings is due in part to the slow growth in the occupations of the $|1\rangle$ and $|7\rangle$ states.
Regardless of the origin of slow growth, finite occupation at the extremes of the qudit spectrum implies a departure from the Floquet regime. Our finite time numerics are unable to resolve whether this implies delocalization. Should this mechanism give rise to a sharp localization transition, it would possibly be of a different character from the extensively studied MBL transition based on ergodic grains thermalizing nearby insulating regions through short range interactions \cite{Vosk2013, Dumitrescu2017, Goremykina2019, Morningstar2020}.
\subsection{Strong coupling (region II)}
\label{strongCoupling}
\begin{figure}
\centerline{\includegraphics[width=0.99\columnwidth, trim={0 0 0 0}, clip]{dynamics_strong.pdf}}
\caption{Dynamics for strong coupling, $\gamma = 0.7/\sqrt{L/12}$. \textbf{(a)} Qudit variance $\Delta^2_{\mathcal{Q}}$ and \textbf{(b)} bipartite entanglement entropy $S_A$. Results from ML-MCTDH are not included for $\Delta^2_{\mathcal{Q}}$ as they are not converged. For system sizes where the dynamics can be computed exactly (dot-dashed lines), $S_A$ saturates the Page bound $S_A \sim L/2$. The curves from ML-MCTDH (solid lines), corresponding to $L\geq 24$, saturate the bound set by the number of single-particle functions in the second layer, $\log_2{\chi_2}$ (dotted lines).}
\label{fig:strongCoupling}
\end{figure}
In the strong coupling regime (Fig.\ \ref{fig:strongCoupling}), our phase diagram suggests that the system lies deep within the thermalizing phase for our available system sizes due to strong delocalizing interactions between the spins induced by the central qudit. Our data is consistent with this expectation, but we note two effects. First, despite the fact that spins thermalize, we observe that the asymptotic distribution of qudit occupations is nonuniform, similar to the athermal qudit regime found in \cite{Ng2019}. In section \ref{sec:randMat} of \cite{Note2} we introduce a phenomenological picture to explain this based on random matrix theory, suggesting that it is rare for the qudit to make transitions between widely separated states.
Second, we note that due to the ergodic character of the dynamics in the strong-coupling thermalizing regime, the accurate treatment of the dynamics represents a significant challenge for the ML-MCTDH approach and cannot be converged for longer times \cite{Westermann2012}.
This well known limitation of the ML-MCTDH method and other tensor network approaches is due to the following reason. Within the ML-MCTDH approach, the wave function of the system is represented in each layer by sums of Hartree products, the total number of which is determined by the number $N_n$ of SPFs employed in a given layer $n$ for each degree of freedom. For example, in the binary tree depicted in Fig.\ \ref{fig:mctdh_tree}, $N_2$ SPFs are used in the second layer to represent each of the two parts of the spin chain resulting in $(N_2)^2$ Hartree products that represent the spin system in the second layer. As a consequence, the entanglement entropy between the different constituents of the system is bounded by $\log N_2$. However, for ergodic systems, the entanglement entropy is extensive, and thus, starting from an uncorrelated state, the entanglement entropy grows and eventually exceeds the limit of $\log N_2$. This implies that, for longer times, the wave function of the system cannot be represented accurately. The application of the ML-MCTDH formalism in the ergodic phase is thus restricted to short times. Therefore, the results for the qudit variance depicted in Fig.\ \ref{fig:strongCoupling} have been obtained by exact diagonalization and the kernel polynomial method.
Despite this limitation of ML-MCTDH, we are still able to find signatures of thermalization by examining the dependence of the dynamics on $N_2$. In the right panel of Fig.\ \ref{fig:strongCoupling}, we see that the bipartite entanglement entropy is upper bounded by $\log_2 N_2$, corresponding to a maximal entropy state within our variational ansatz. We observe similarly strong dependence of $q_{\text{EA}}$ when increasing $N_2$, which drifts towards zero to indicate paramagnetic behavior in the spin chain. Other observables, such as the populations of the qudit levels cannot be converged, implying that information about the ergodic state is present, but limited.
\section{Results for unscaled coupling}
\label{sec:unscaledGamma}
The dynamics of our system with scaled coupling, $\gamma \sim 1/\sqrt{L}$, is perhaps most interesting because it gives a well-defined thermodynamic limit for the qudit. However, it is also important to understand the dynamics when the coupling is held fixed instead of being scaled by system size, corresponding to the dashed horizontal line in Fig.\ \ref{fig:phaseDiagram}. Fixing the coupling strength may be easier to implement experimentally, for example in cavity QED where the coupling is governed by the position-dependent electric field strength. Doing so, however, means that we can no longer easily separate dynamics occurring on different timescales as in the previous section. Furthermore, we predict that in the thermodynamic limit this will eventually result in thermalization, as the long-range interactions induced by the central qudit will eventually dominate at large enough times and system sizes. We choose the fixed value $\gamma=0.10607$, which precisely matches the intermediate scaled coupling for our largest system size, and therefore lives within the predicted MBL phase for all accessible system sizes.
\begin{figure}[H]
\centerline{\includegraphics[width=0.99\columnwidth, trim={0 0 0 0}, clip]{dynamics_unscaled.pdf}}
\caption{Dynamics with unscaled coupling, $\gamma=0.10607$. On the reachable timescales, the spin glass order parameter $q_{\text{EA}}$ \textbf{(a)} does not show significant system size dependence. However, the bipartite entanglement entropy \textbf{(b)}, in addition to being subextensive -- $S_A \propto L^\alpha$, $0 < \alpha < 1$ -- at short times, already shows qualitatively different behavior at intermediate time for $L \gtrsim 24$. \textbf{(inset)} Mutual information $MI \equiv I(A,B)$ between two contiguous halves $A$ and $B$ of the spin chain (see main text).}
\label{fig:unscaled}
\end{figure}
The dynamics with fixed coupling, shown in Fig.\ \ref{fig:unscaled}, looks similar to the data at intermediate scaled coupling but without as clear a separation of time scales or data collapse. We note that it is harder to detect the sort of logarithmically slow delocalization as seen in $q_{\text{EA}}$ in Fig.\ \ref{fig:intTime}f (cf.\ Fig.\ \ref{fig:unscaled}a). The story is the same with the qudit variance, which is similar to $1-q_{\text{EA}}$ when divided by $L$. However, with the fixed coupling, the entanglement entropy reflects a qualitative change in behavior at large enough system sizes.
We see that at times $\frac{\Omega}{2\pi}t\sim 6\times 10^1$ in the localized phase (Fig.\ \ref{fig:unscaled}), the smaller system sizes $L\leq 16$ establish a subextensive amount of entanglement entropy. Numerics from ML-MCTDH seem to counter this trend, with $S_A$ continuing to grow slowly beyond this timescale. The rate of this growth increases with $L$, which is consistent with it arising from stronger effective all-to-all interactions described by the high-frequency expansion (Eq.\ \ref{eq:HFE}). It also appears to show strong system size dependence at short times, where a subextensive amount of entanglement is established. This should be contrasted with models of MBL without central coupling, in which the short time behavior is system size independent.
As a final note, we point out that the entanglement entropy at fixed $\gamma$ is subextensive, such that $S_A/L$ appears to be trending towards zero with increasing system size. This should be contrasted with mutual information between the two halves of the spin chain,
\begin{equation*}
I(A,B)=S(A) + S(B) - S(A \cup B),
\end{equation*}
which is extensive. In Ref.\ \cite{Ng2019}, mutual information was used as a proxy for entanglement between the two halves of the spin chain, as it nominally removes ``unimportant'' entanglement with the central qudit. However, since entanglement must be subextensive -- indeed, system size independent -- in the MBL phase, our data indicate that entanglement entropy is a better metric than mutual information for capturing this. Our initial expectation was that mutual information would become subextensive at larger system size, but the results obtained with the ML-MCTDH method rule out that possibility.
\section{Conclusions}
In this paper we have studied the dynamical behavior of a qudit coupled to a disordered, interacting bath of up to $L=96$ spins-$1/2$, which altogether can exhibit localization at strong disorder. Using a combination of exact propagation methods and the tensor network-based ML-MCTDH approach, we find evidence of qualitatively different dynamical signatures in local observables such as the spin glass order of the bath and the qudit variance, consistent with a rough phase diagram (Fig.\ \ref{fig:phaseDiagram}). Most notably, we find hints of logarithmically slow decay of localization near the onset of all-to-all interactions in the bath. This behavior was found to occur after timescales $t\sim O(1/\gamma^2)$ where $\gamma$ is the qudit-spin bath coupling.
The behavior of the qudit observed here is, we believe, not specific to this model. Our conclusions should apply equally well to the cases of a cavity photon with rescaled raising/lowering operators $a^\dagger \to (N_0)^{-1/2} a^\dagger$ or central spin-$S$ systems with operators rescaled as $\hat{S} \to (S(S-1))^{-1/2} \hat{S}$. The feature of these systems is that the fundamental commutation relation between the raising and lowering operators vanishes in the limit of large $S$ or large $N_0$. This fact allows for exact cancellation between processes that raise or lower the qudit state. However, this mechanism only serves to protect localization for sufficiently large ``magnetic field'' $\Omega$; it is unclear how these systems interpolate between the $\Omega=0$ limit and the $\Omega > |g|, |h_i|, \ldots$ limit. We note additionally that the limitations of ML-MCTDH for these types of centrally coupled systems with many-body interacting baths in the strong coupling regime requires more clarification. Such clarifications may be necessary to extend the effectiveness of the method into the thermalizing regime on the left side of the phase diagram \ref{fig:phaseDiagram}, which remains numerically inaccessible and thus poorly understood.
\begin{center}
\textbf{Acknowledgements}
\end{center}
This work was performed with support from the National Science Foundation through award number DMR-1945529 (MHK), the Welch Foundation through award number AT-2036-20200401 (MHK), and the German Research Foundation (DFG) through IRTG 2079.
This research used resources of the National Energy Research Scientific Computing Center, a U.S. Department of Energy Office of Science User Facility operated under Contract No. DE-AC02-05CH11231. Furthermore, support by the state of Baden-Württemberg through bwHPC and the DFG through grant no.\ INST 40/575-1 FUGG (JUSTUS 2 cluster) is gratefully acknowledged.
\bibliographystyle{apsrev4-1}
|
\section{Introduction}
A knot $K$ in $S^3$ is said to be {\em periodic} if there exists an integer $p\ge 2$, a diffeomorphism $f:S^3\to S^3$ of order $p$ that preserves the knot $K$, and whose fixed point set Fix$(f)$ is diffeomorphic to $S^1$. In this case we say that $K$ is $p$-periodic, that $p$ is a period of $K$, and we call Fix$(f)$ the {\em axis of $f$}.
See \cite{JabukaNaik} for more background on periodic knots.
In \cite{Edmonds} Edmods famously proved that if $K$ is a $p$-periodic knot of genus $g$, then there exists a Seifert surface $\Sigma$ for $K$ of genus $g$ that is invariant under the diffeomorphism $f$. Said differently, if we define the {\em $p$-periodic} (or {\em equivariant}) {\em 3-genus $g_{3,p}(K)$ of a $p$-periodic knot $K$} as
$$g_{3,p}(K) = \min \{ g\ge 0 \, |\, \text{ $K$ possesses an $f$-invariant Seifert surface of genus $g$} \},$$
then Edmonds' theorem can be seen as saying that $g_3(K) = g_{3,p}(K)$ for every $p$-periodic knot $K$ (with $g_3(K)$ being the Seifert genus of $K$).
The goal of this note is to show that if one considers nonorientable spanning surfaces for periodic knots instead, the analogue of Edmonds' theorem is not true. To state our result, we recall the definition of the {\em nonorientable (nonequivariant) 3-genus $\gamma_3(K)$}, and we define the {\em $p$-periodic} (or {\em equivariant}) {\em nonorientable 3-genus $\gamma_{3,p}(K)$} of a $p$-periodic knot $K$ as
\begin{align*}
\gamma_{3}(K) & = \min \{ b_1(\Sigma) \, |\, \text{ $\Sigma$ is a nonorienatble spanning surface for $K$} \}, \cr
\gamma_{3,p}(K) & = \min \{ b_1(\Sigma) \, |\, \text{ $\Sigma$ is an $f$-invariant nonorienatble spanning surface for $K$} \}.
\end{align*}
We leave it as an easy exercise to show that every periodic knot has an equivariant nonorientable spanning surface, and thus the definition of $\gamma_{3,p}(K)$ is well posed. It is also not hard to show that $\gamma_{3,p}(K) \le 2g_3(K) +p$ for any $p$-periodic knot $K$.
\begin{theorem} \label{main}
Let $K$ be a $p$-periodic knot with $p\ge 3$ and with $\gamma_3(K)\ge 2$. Then $\gamma_{3,p}(K) \ge p$.
\end{theorem}
\begin{proof}
Let $f:S^3\to S^3$ be a diffeomorphism that facilitates the $p$-periodicity of $K$ and let $A=$Fix$(f)$ be its axis. Let further $\Sigma \subset S^3$ be a nonorientable $f$-invariant spanning surface for $K$ and let $\overline \Sigma \subset S^3$ be the quotient of $\Sigma$ by the action of $\mathbb Z_p$ generated by $f$, and note that $\overline \Sigma$ is nonorientable. Then $\Sigma\to \overline \Sigma$ is a $p$-fold cyclic cover, branched along $\lambda \ge 0$ points, with $\lambda$ being the number of points in $\Sigma \cap A$. A straightforward computation of Euler characteristics gives
\begin{equation} \label{EquationEulerCharacteristics}
\chi(\Sigma) = p\cdot \chi (\overline \Sigma) - (p-1)\lambda.
\end{equation}
Write $b_1(\Sigma) = a$ and $b_1(\overline \Sigma) = b$. The assumption $\gamma_3(K) \ge 2$ forces $a\ge 2$, while by definition $b\ge 1$ and $\lambda \ge 0$. Equation \eqref{EquationEulerCharacteristics} then becomes
\begin{equation} \label{EquationOfBAndLambda}
a-1 = p(b-1) +(p-1)\lambda.
\end{equation}
If $b=1$, we obtain $a-1 = (p-1)\lambda$ forcing $\lambda >0$ since $a\ge 2$. This in turn forces the inequality $a-1\ge p-1$ or $a\ge p$. If $b\ge 2$ then \eqref{EquationOfBAndLambda} implies $a-1\ge p$. Thus, in either case we find $a\ge p$ and hence $\gamma_{3,p}(K_p) \ge p$, since $\Sigma$ was an arbitrary equivariant nonorientable spanning surface for $K$.
\end{proof}
\begin{corollary} \label{CorollaryOne}
The difference between the equivariant and nonequivariant nonorientable 3-genera of a periodic knot can become arbitrarily large. Specifically, for every integer $p\ge 3$ there exists a $p$-periodic knot $K_p$ with
$$\gamma_3(K_p) = 2 \qquad \text{ and } \qquad \gamma_{3,p}(K_p) \ge p.$$
\end{corollary}
\begin{proof}
Let $K_p$ be the torus knot $T(4p,2p-1)$. By \cite{Teragaito} (see also \cite{JabukaVanCott}) we obtain $\gamma_3(K_p)=2$ for all $p\ge 3$. The periods of a torus knot $T(a,b)$ are precisely the divisors of $|a|$ and $|b|$, showing that $K_p$ is $p$-periodic. Theorem \ref{main} implies that $\gamma_{3,p}(K_p) \ge p$.
\end{proof}
The inequality $\gamma_{3,p}(K) \ge p$ from Theorem \ref{main} is sharp as seen in the next example.
\begin{example}
Consider the 5-periodic torus knot $K=T(5,3)$. It follows from \cite{Teragaito} that $\gamma_3(K) = 2$ (or use \cite{KnotInfo} where $T(5,3)$ is the knot $10_{124}$), showing that $K$ meets the hypothesis of Theorem \ref{main} and thus $\gamma_{3,5}(K) \ge 5$. An equivariant spanning surface $\Sigma$ for $K$ with $b_1(\Sigma) = 5$ is shown in Figure \ref{FigureFor12-124}, leading to $\gamma_{3,5}(K) = 5$. The values of $a$, $b$, $\lambda$ from the proof of Theorem \ref{main} are 5, 1, 1 respectively, and satisfy equation \eqref{EquationOfBAndLambda}.
\end{example}
\begin{figure}
\includegraphics[width=5cm]{10-124}
\caption{The torus knot $T(5,3)$ shown with an equivariant nonorientable spanning surface $\Sigma$ with $b_1(\Sigma ) = 5$. } \label{FigureFor12-124}
\end{figure}
Another important result of Edmonds' \cite{Edmonds} is the bound $p\le 2g_3(K)+1$ satisfied by any period $p$ of the knot $K$. While it was known prior to Edmonds' work that a knot may only have finitely many periods (cf. Theorem 3 in \cite{Flapan}), the preceding inequality was the first quantitative bound on the number of possible periods of a knot. Corollary \ref{CorollaryOne} shows, as yet another contrast to Edmonds' results, that no upper bound on the periods of a knot can exist by any polynomial function in the nonorientable 3-genus. This conclusion also follows from considering the $p$-periodic alternating torus knots $T(2,p)$ for which $\gamma_3(T(2,p)) = 1 = \gamma_{3,p}(T(2,p))$, with $p\ge 3$ odd.
\bibliographystyle{plain}
|
\section{Introduction}
\label{sec1}
Randomized Controlled Trials (RCT) are considered to be the most reliable form of scientific evidence since it reduces spurious causality and bias. United States Preventive Services Task Force has recognized \emph{"evidence obtained from at least one properly randomized controlled trial with good internal validity"} as the highest quality evidence~\cite{harris2001current}.
It is common in RCT that only one party knows the \emph{opportunities}, how individuals are randomized into test and control groups. The other party only knows the \emph{outcomes} that the treatment was meant to affect (e.g., did an individual take a specific action). Each party wishes to calculate and compare aggregate statistics of how different treatment groups performed, without revealing their input data to the other party.
\begin{figure*}[t]
\centering
\includegraphics[scale=0.45]{steps.png} \\
\caption{Private RCT, high level design}
\label{fig:PLdesign}
\end{figure*}
We design and implement Private RCT; a practical and scalable secure two-party computation system for calculating Randomized Controlled Trial experimental results. Designing a scalable privacy-preserving solution is technically challenging, especially when using one cryptographic primitive in isolation, such as either Garbled Circuit (GC) \cite{saleem2018recent} or Private Set Intersection (PSI) \cite{PrivateMatching}. Our protocol combines a suite of underlying cryptographic primitives such as GC, PSI and Secret Sharing \cite{beimel2011secret}, and incorporates them in a way that is more efficient than using any one secure computation technique by itself. Moreover, we demonstrate how to add distributed computing scaling techniques such as sharding without compromising privacy. Finally, we use Differential Privacy (DP) \cite{Dwork2013} to add noise to the output of the computation and prevent leakage of the input data based on the revealed output.
\subsection{Private Randomized Controlled Trials}
Privacy-preserving computation (secure computation) is a cryptographic method that enables parties to jointly compute a function on each of their secret inputs while preserving the privacy of the inputs. The technology guarantees that the parties will only learn the designated output of the function and they cannot access or derive each other's inputs (also known as input privacy), any intermediate values, or statistical results.
In the context of RCTs, privacy-preserving computation is highly desirable as it enables statistical measurements without giving access to the raw/un-encrypted input databases to external parties. This creates stronger guarantees to enforce the individual expectation of privacy.
Take clinical trials for example. Medical researchers randomly assign subjects into treatment and control groups and measure a few pre-determined health outcomes. While researchers can measure some demographic variables by themselves, much richer but usually unavailable demographic and behavioral data such as living conditions, mental status, social interactions, socio-economic status can be immensely valuable to improve statistical power, cut costs, and extract more credible insights from typically small-scale clinical trials. Unfortunately such auxiliary data is usually held by third parties such as hospitals, physicians, and government agencies. Private RCT solution makes it easier for these independent parties to collaborate in analyzing RCTs.
\subsection{High-level Design}
Private RCT has three main stages. The protocol is designed to make deployment in real world easy by separating the steps of the protocol and address their challenges individually. Figure \ref{fig:PLdesign} shows the overall design.
\begin{enumerate}
\item \emph{Private Identity Matching.}
The first stage is to privately determine the join of users between the two parties. We use two main open source protocols for performing this Private Identity Matching: PID and PS$^3$I \cite{PrivateMatching}. The former computes a full outer join (i.e. union) of the identities using a Decisional Diffie-Hellman (DDH) based construction that is only twice as computationally expensive as standard DDH-based Private Set Intersection (PSI). PS$^3$I computes the inner-join (i.e. intersection) of the identities and generates secret shares of attached data; it uses DDH techniques and additive Homomorphic Encryption (HE). We also have ongoing work to extend both protocols to handle identity matching in cases where each row may have multiple identifiers, and we may wish to output many-to-many matches.
\item \emph{Scalable Secure Computation.} In the second stage, we calculate a pre-defined linear circuit on the output of the previous step. Since we envision different use cases, which differ in their nature of calculation, the computation step should be able to compute any circuit on the joined data. We evaluated multiple general purpose MPC frameworks for this step and concluded to build our solution based on EMP-toolkit. We used sharding mechanism to scale it from handling 2 million rows of input data to 500 million rows. We use an XOR secret sharing mechanism to hide any intermediary result in this step.
\item \emph{Differentially Private RCT Outputs.} The last step is to prepare the output. Since the output is going to be revealed to the both parties, it is important to ensure it does not break the privacy guarantees that the system promises. We need to check for access control, ensure rate limits and finally add DP noise to the output. We calibrate DP noises to ensure two-sided privacy guarantee: participant A gets only a differentially private view of B’s secret input, and vice versa. To provide differentially private confidence intervals of two-sample difference-in-means estimators, we examine several algorithms and compare their tightness, coverage, and computational cost within MPC. We also propose methods to generate DP noises distributively in the presence of semi-honest or malicious adversaries.
\end{enumerate}
\subsection{Our Contributions}
By combining Decisional Diffie-Hellman (DDH) PSI, Secure Multi-Party Computation (MPC), and Differential Privacy (DP), we provide a private yet performant solution to enable two parties to collaborate in RCTs. Starting from the correct underlying cryptographic primitives, we show how to use them, scale to large data sets, and handle the challenges that show themselves in practice. In addition to designing and implementing the Private RCT, we have the following contributions which can be of independent interest to the community:
\begin{enumerate}
\item \emph{Evaluation Framework for MPC.}
We looked into multiple services that enable secure computation to decide on the best framework that meets the requirements of Private RCT. This systematization of knowledge on MPC protocols can be of independent value for researchers and industry cryptographers. We started with a comparison across more than twenty secure computation services --- across secret sharing, garbled circuits, homomorphic encryption-based protocols. We chose ABY, EzPC, EMP-toolkit, Fancy Garbling, Scale-Mamba, Obliv-C and MPyC for a more detailed literature review --- and from them we did performance testing on ABY, EMP-toolkit, and Scale-Mamba. We will discuss the details of this evaluation in section \ref{sec:eval}.
\item \emph{Scaling via Private Sharding.} Handling large volumes of data is important in practice. Unfortunately, most current secure computation platforms cannot handle large enough data sets at the required scale and speed imposed by our application. In this work, we designed and implemented a privacy-preserving sharding technique to make EMP scale to 500M rows. Our protocol does not reveal any intermediary information to the involved parties, i.e., the intermediary result of computation on each shard will remain private. We will discuss the details of this protocol in section \ref{sec:sharding}.
\item \emph{Private Conversion Lift.} To test the efficiency and scalability of our design, we implemented Conversion Lift as an application that uses Private RCT in the backbone. Conversion Lift compares the actions of users in randomized test and control groups to measure the additional business driven by the advertisement. See details of implementation in section~\ref{lift}.
\end{enumerate}
\section{Private RCT Protocol}
In the following sub-sections, we will provide the detailed design for each step of the protocol.
\subsection{Assumptions and Model}
We assume there are two parties involved in the protocol Alice and Bob each have a private input $a$ and $b$ respectively. Each party has their own separate infrastructures that jointly participate in the protocol and there is no trusted third party involved. They jointly wish to compute a function $o=f(a,b)$ that is the outcome of PRC trial over their inputs and receive output $o$.
\emph{Adversarial model.} We assume both parties are semi-honest, in that they will use any information they can learn from the protocol to try and learn the other's data, but they will not deviate from the protocol specification. We assume that both parties want to learn the correct output of the computation and will not seek to poison the output by altering their inputs. Thus, the main thing we are trying to prevent is one party learning the other's inputs. Moreover, we do consider that an adversary may craft their inputs to maximize the information they can learn passively, as discussed in \ref{sec:dp}.
\subsection{ Background on Cryptography Primitives}
\noindent \emph{\textbf{Garbled Circuit (GC)}}
Garbled circuit is a family of secure computation protocol that is more suitable for 2PC but can be extended to MPC (for small numbers). This technology is better for Boolean functions.
In the garbled circuit, the garbler encrypts a Boolean circuit of g = f(a, ·) and sends the encrypted circuit to the evaluator. The evaluator then uses oblivious transfer (OT) to obtain the keys corresponding to the input to decrypt the garbled circuit and evaluate g(b).
OT (the underlying technology of GC) is highly performant compared to secret sharing (on computation cost) and homomorphic encryption (on communication cost) technologies.
\noindent \emph{\textbf{Secret Sharing (SS)}}
In secret sharing, one party (called the dealer) distributes a secret among a group of parties, each of whom is allocated a share of the secret. Each share reveals nothing about the secret to the party possessing it, and the secret can only be reconstructed when a sufficient number of shares are combined together.
In secret sharing based MPC, each party creates secret shares of its input data and shares that among all parties. Then, all parties perform some intermediate computation on the received shares to get a share of the output. In the last stage, all parties will reconstruct the final output together by sending the intermediate outputs to each other.
\noindent \emph{\textbf{Homomorphic Encryption (HE)}}
A fully homomorphic encryption (FHE) scheme performs secure computation over encrypted data without decrypting it.
Current techniques are still slow and can only evaluate small circuits. This restriction is primarily due to noise management techniques (such as bootstrapping) used to deal with a noise term in ciphertexts that increases slightly with homomorphic addition and exponentially with homomorphic multiplication.
Thus, this technique is suitable for special case circuits with more additions and fewer consecutive multiplications.
\subsection{Private Identity Matching}
Before computation can be done for Private RTCs the two data sets must be joined by some identities of the users. There are three matching cases we consider:
\begin{enumerate}
\item \textbf{Single unique identifier.} The simplest situation is that each user has one identifier (e.g. email) which is unique to them in one of the party's input sets. We want to compute on users who have the same identifier in both party's input sets.
\item \textbf{External identifier.} In this situation, there is a common identifier shared between the two parties and each party has a many-to-many relationship between their own user ids and the external ids.
\item \textbf{Many identifiers.} Each user may have multiple identifiers (e.g. email and phone). This case also results in a many-to-many matching between users from both parties.
\end{enumerate}
The paper \cite{PrivateMatching} introduces two protocols for handling the case of a single unique identifier. Both protocols can compose with general-purpose MPC to enable arbitrary computation on the joined data.
The first variant which we call Private-ID (PID), allows the parties to privately compute a set of pseudorandom universal identifiers (UID) corresponding to the records in the union of their sets, where each party additionally learns which UIDs correspond to which items in its set but not if they belong to the intersection or not. This new formulation enables the parties to independently sort their UIDs and the associated records and feed them to any general-purpose MPC that ignores the non-matching records and computes on the matching ones. The Private-ID protocol has the advantage that it only needs the identifiers from the records as input to produce the UIDs and hence for each application, parties can assemble a possibly new set of features/labels per identifier for the downstream computation without re-executing the protocol.
The second protocol from \cite{PrivateMatching} called Private Secret Shared Set Intersection ($\ensuremath{\text{PS}^3\text{I}}$), is a natural extension of PSI where instead of learning the plaintext matched records, parties only learn additive shares of those records which they can feed to any general-purpose MPC to execute the desired computation on. The construction is based on efficiently extending existing DDH-based PSI using any additive homomorphic encryption scheme.
The advantage of $\ensuremath{\text{PS}^3\text{I}}$ over Private-ID is that its output size and hence the complexity of the subsequent MPC is proportional to the size of the intersection which in some cases is much smaller than size of the union of the two original datasets. Its disadvantage, similar to prior work, is that full records and not just the identifiers need to be ready at the time of execution, and it requires a rerun when associated records change for the same identifiers.
We have further generalized these protocols to work in matching scenarios (2) and (3) where the mapping is many-to-many. There are two ways to resolve these many-to-many mappings:
\begin{enumerate}
\item \textbf{Resolve to a single match.} In the matching case where each user has many identifiers, we can define rules of priorities or weights so as to choose the best possible match. Both Private-ID and $\ensuremath{\text{PS}^3\text{I}}$ can be generalized to do this. In cases where the many-to-many mapping results from a shared external identifier, it is generally the case that each occurrence of the identifier is of equal value in defining a match, so the next collecting approach may be better.
\item \textbf{Collect many-to-many matches.} In some applications of Private RCT it is actually best not to resolve to a single match but rather to collect all the matches into a many-to-many matching or many-to-one matching. The $\ensuremath{\text{PS}^3\text{I}}$ protocol generalizes more naturally to do this than Private-ID, and we call this variant Collecting PS3I.
\end{enumerate}
One implication of a many-to-many identity matching is that it affects the RCT validity. If one party partitions users into Test and Control groups, it is possible that one user from the other party's set will match with users from both the Test and Control groups. We call such users contaminated and drop them from the study when running Collecting PS$^3$I on both the test and control groups simultaneously and looking for overlap in the encrypted intersections. When running a many-identifier version of PID, we cannot drop these contaminated users, but we can count how many of them there are and use this to inform our confidence in the results.
\subsection{Private RCT Computation} \label{sec:sharding}
\begin{figure*}[h!]
\centering
\includegraphics[scale=0.4]{MPC-Sharding.png} \\
\caption{ Private RCT design. Garbler and Evaluator have the same overall architecture. }
\label{fig:design}
\end{figure*}
We estimate that a large RCT, as is seen in advertising conversion lift studies, may routinely have up to 500M rows; thus handling large volumes of data is important in practice.
Unfortunately, EMP-toolkit and similar MPC platforms were not able to handle such a large volume of data at once in the required scale and speed imposed by the application.
EMP-toolkit was limited to 2M rows of integer data when we tested on the largest available AWS Fargate option (30GB), although it was able to handle 4M rows on a larger EC2 instance (64GB).
In this section we go over our design for computing a RCT game privately across multiple containers via our privacy-preserving sharding mechanism.
In our protocol, data from the private id match is deterministically partitioned into separate container shards.
Each shard then performs a 2PC with the corresponding shard controlled by the other party.
The intermediary output of each shard remains private, i.e. the garbled values will not be open to the parties.
Instead, each party will learn an XOR share of the value.
Consequently, the aggregation step will happen after the data is reconstructed from the XOR-shares in a garbled circuit and only the final result of the computation will be revealed.
Figure \ref{fig:design} shows the high-level design of the Protocol, as it is replicated on both the Garbler and Evaluator's sides.
Both Garbler and Evaluator each have one Coordinator, multiple Workers and one Aggregator.
The Coordinator partitions the input database into shards in a round robin method, and assigns each shard to a Worker.
Workers will evaluate the sub-circuit on the input shard and send the intermediary result to the Aggregator who evaluates the final output and adds DP noise to it. For each Worker on the Garbler side, there is a corresponding Worker on the Evaluator side that has point-to-point connection with each other and work together to evaluate the garbled circuit. Similarly, the Aggregator on the Garbler side has network connection and computes the aggregation jointly with the Aggregator on the evaluator side.
\subsubsection{Private Partitioning and Aggregation.}
To ensure no party learns the output of computation on one shard, we do not reveal the intermediary output at the end of their game.
Instead, The server will choose a random number as a new input and XOR the result of the computation with that random number in the garbled circuit.
At the end of the game, Evaluator knows the result of computation XORed with that random number and Garbler knows the random number.
Jointly, they can reconstruct the result but individually, they do not have any information about it due to the security of one-time pads.
\subsection{Differential Privacy in MPC} \label{sec:dp}
Deterministic MPC output can leak input information at the individual level. Typical privacy attacks on aggregated statistics include re-identification attacks, database reconstruction, and membership inference \cite{Dwork2017}. To further enhance privacy of the protocol, specifically, to make sure neither party learns any information that they don’t already know about any individual, we use Differential Privacy (DP) to add randomness to the output. See \cite{Dwork2013} for common DP definitions and algorithms. In this work, we use a more recent variant of DP definition called zero-concentrated DP (zCDP) from \cite{Bun2016}.
We aim to protect against attacks to reveal an individual’s membership, treatment status, or any outcome to any party that does not already have the data. Treatment status can itself be highly sensitive (e.g., in a clinical trial), but we also want to emphasize that membership/eligibility in the experiment and the outcome value are equally the focus of protection guaranteed by the DP mechanism.
We address three specific considerations to implement DP for Private RCT below.
\subsubsection{DP Confidence Intervals (CIs)}
Valid statistical inference with RCTs requires valid measures of uncertainty (e.g. Confidence Intervals). Valid CIs consider both the sampling variation and the randomness added by the DP noise. We consider three criteria for CIs. 1) CIs should be differentially private. 2) CIs should have the correct statistical coverage. For example, 95\% CIs should cover the true popular value at least 95\% of the time. 3) CIs should be narrow, given the DP guarantee and the correct coverage.
There are several additional challenges to construct DP confidence intervals in Private RCTs. 1) Most proposed methods for DP confidence intervals aim at estimating unknown parameters from known parametric distributions, e.g. releasing the DP mean estimate and its DP standard error from known distributions \cite{karwa2017finite, Awan2019, du2020differentially, evans2019statistically, ferrando2020general}. In RCTs we estimate the difference-in-means of two samples from unknown distributions, a context considered in \cite{dorazio2015differential}. 2) Nonparametric methods for confidence intervals are typically based on resampling techniques such as bootstrapping, which can be computationally expensive in MPC \cite{brawner2018bootstrap, covington2021unbiased}.
We compared several parametric and non-parametric approaches to construct DP confidence intervals for RCT, and found that a parametric approach based on approximating the distribution of the estimator's sampling distribution strikes a good balance for our use case with large samples. See Algorithm \ref{algo:ad} for detailed description of the algorithm.
\begin{algorithm}
\caption{Differentially Private RCT}
\label{algo:ad}
\begin{flushleft}
\textbf{Input:}
\begin{itemize}
\item $x_T$: user-level outcomes for the test group
\item $x_C$: user-level outcomes for the control group
\item $R$: upper bound of user-level outcomes (lower bound = 0)
\item $\rho_1$: zCDP privacy budget for point estimate
\item $\rho_2$: zCDP privacy budget for standard error
\item $\alpha$: significance level of confidence interval (e.g., 10\%)
\end{itemize}
\textbf{Output: } $\left[\text{DP lift} - w, \text{DP lift} + w\right]$ confidence interval
\end{flushleft}
\begin{algorithmic}[1]
\State Clamp/Winsorize: \[
Y_{i}=\begin{cases}
X_{i} & \text{if }X_{i}\le R \\
R & \text{if }X_{i}>R.
\end{cases}
\]
\State Calculate sample means, variances, and counts: $\bar{y}_T$, $\bar{y}_C$, $s^2_T$, $s^2_C$, $n_T$, $n_C$.
\State $\text{lift} \gets \bar{y}_T - \bar{y}_C$.
\State Standard error of lift: $se_{\text{lift}} \gets \sqrt{s_T^2 / n_T + s_C^2 / n_C}$.
\State Sensitivity of lift: $\Delta_\text{lift} \gets \frac{R}{n_T} + \frac{R}{n_C}$.
\State Sensitivity of the standard error of lift: $\Delta_{se_{\text{lift}}}\gets \sqrt{\frac{N^* - 1}{N^{*3}}}R$, where $N^* = \min \left(n_T, n_C\right)$.
\State Draw scalar random noise $Z_1 \sim \text{Normal}\left(0, \frac{\Delta^2_{\text{lift}}}{2\rho_1}\right)$, $Z_2 \sim \text{Normal}\left(0, \frac{\Delta^2_{se_\text{lift}}}{2\rho_2}\right)$.
\State DP lift $\gets \text{lift} + Z_1$, where $Z_1 \sim \text{Normal}\left(0, \frac{\Delta^2_{\text{lift}}}{2\rho_1}\right)$.
\State DP $se_\text{lift} \gets se_\text{lift} + Z_2$, where $Z_2 \sim \text{Normal}\left(0, \frac{\Delta^2_{se_\text{lift}}}{2\rho_2}\right)$.
\State $w = \sqrt{\left(se_\text{lift} + Z_2\right)^2 + \frac{\Delta^2_{\text{lift}}}{2\rho_1}} \cdot z_{1-\alpha/2}$, where $z_{1-\alpha/2}$ is the critical value of standard normal at $1-\alpha/2$.
\end{algorithmic}
\end{algorithm}
\subsubsection{Two-side Privacy Guarantees}
In Private RCTs, each party’s input data is masked from the other party, thus DP noises need to be calibrated to ensure two-side privacy guarantee: Party A gets only a differentially private view of Party B's secret input, and vice versa \cite{Mcgregor2011}. Recall that different parties hold different parts of the data, for example, party A has the treatment variable and party B holds the outcome variable. The sensitivity of the calculation with respect to the treatment variable is larger than the sensitivity with respect to the outcome variable. It suggests that to protect party A's input data, more noise needs to be added to the final output exposed to party B. In Algorithm \ref{algo:ad}, we simplify the exposition by taking the larger sensitivity of the two parties, thereby showing both party the same results, but in practice parties may see different results due to different scales of the DP noise.
\subsubsection{Distributed DP Noise Generation} \label{sec:babble}
As a first straw man, we consider a simpler semi-honest version of the problem in which only Bob will see the output.
Alice can generate random noise $e_a$ from any desired distribution based on the DP protocol
and adds it as an extra input to MPC which calculates $f(a,b) + e_a$ and sends the sum to Bob where $f(a,b)$ is the intended output of the protocol .
However, a malicious adversary may intentionally choose a huge noise to destroy the other party’s data utility. To address this concern, we propose using the cut and choose method. Protocol asks Alice to generate a vector of $k$ random noises
$\vec{e_a}$ sampled from the desired distribution based on DP protocol and inputs it as an extra input to MPC. Bob will choose an index $i$ and parties jointly calculates $f(a,b)+\vec{e_a}[i]$ in MPC and reveal the sum to Bob. MPC also reveals all the elements in $\vec{e_a}$ other than the $i$th element which we denote as $\vec{e_a[-i]}$ to Bob. Note he can check if ${\vec{e_a}}[- i]$ has correct distribution based on DP-parameters and abort if it is not correct. Thus, our protocol has covert security. The same procedure, with reversed roles, can be done to create Alice’s output. MPC creates Bob’s and Alices’s outputs separately and reveal it only to the corresponding party.
\section{Private Conversion Lift: Implementing a Real World Example}
In an open source deployment of the protocol \cite{facebook2021FBPCFFacebook}, we implemented Private Conversion Lift which is used to measure the incremental (causal) impact of an advertising campaign on consumer conversions, actions such as purchases and website registrations. The party serving ads (publisher) randomly splits eligible users into test or control groups. When there is an opportunity to serve an ad from the campaign, the ad is displayed if that user is in the test group, but not displayed if they are in the control group. In both cases, the ad server logs that the user had the opportunity to see the ad. At the completion of the study, the protocol combines both the opportunity data that the publisher has recorded with outcome data from the advertiser. From these two datasets, the number of conversions in both test and control groups is counted, and the statistical estimates are made to understand incremental outcomes from the ad campaign. More details about Conversion Lift can be found in \cite{conversionlift}.
\vspace{10pt} \noindent \textbf{Identity Matching Step.}
Private Conversion Lift is a two party protocol between an advertiser and a publisher with the following input data. The identity matching we describe here is for a single identifier $x$ (e.g. email) which is unique in each party's input data and uses the PID protocol.
The Publisher and Advertiser have an identifier $x$ for each of their users which they input into the PID protocol. The output of the PID protocol revealed to both parties is a set of pseudorandom identifiers the size of the deduplicated union of the input sets. Each party also receives a mapping from their users into this set of pseudorandom identifiers (which we call the "identity spine" and denote the values as UID's). But neither party learns if a particular one of their users is in the intersection or not. Both parties can then each independently sort the identity spine and align their users to it forming the full-outer join of their sets of users.
\vspace{10pt} \noindent \textbf{Computation Step.} We assume the size of the identity spine output from the PID identity matching is on the order of the 500 million rows that is fed to the Private Conversion Lift calculation step. For each of their users the Publisher has the common identity spine identifier UID, an opportunity timestamp, and a Boolean value to indicate if the user is in the test or control group.
The Advertiser also has the UID row identifier and a set of conversions for each of their users. Each conversion has associated data including a timestamp, value, and value squared that are used in the Conversion Lift calculation.
The calculation is implemented in two games; the Lift game runs within Worker calculates the Lift results on the shard in garbled circuit and outputs the intermediary output in XOR format. The Lift game first compares the conversion's timestamp and opportunity timestamp and adds the conversion value to the total output value if the conversion happened after the opportunity. The Aggregation game runs within Aggregator that receives the intermediary outputs, XORs them pairwise to reconstruct the actual intermediary output in garbled circuit and then runs the aggregation function (addition) on them.
\vspace{10pt} \noindent \textbf{Adding DP noise.} The Aggregation game adds DP noise to the result before revealing it. Specifically, each party follows the distributed DP noise generation protocol in section \ref{sec:babble} and generates DP noises that will be added to the lift estimates and confidence intervals seen by the other party. The DP noises are scaled to the sensitivities to ensure two-sided privacy guarantees, following Algorithm \ref{algo:ad}.
\label{lift}
\subsection{Evaluation Framework}
In this section, we compare the popular open source libraries systematically to choose the most suitable MPC framework for computing RCTs privately at scale. M. Hastings \textit{et~al.}~\cite{8835312} perform a similar evaluation between different MPC libraries for general use cases and we borrowed their docker container as the basis for running the protocols.
We started with more than twenty MPC toolsets and a high level overview of the toolkits helped us initially narrow down to seven frameworks to investigate further; ABY~\cite{demmler2015aby} by Demmler \textit{et~al.}, EzPC~\cite{chandran2017ezpc} by Chandran \textit{et~al.}, EMP-toolkit~\cite{emp-toolkit, wang2017faster, wang2017authenticated} by Wang \textit{et~al.}, Fancy Garbling~\cite{cryptoeprint:2016:969} by Ball \textit{et~al.}, SCALE-MAMBA~\cite{cryptoeprint:2010:514, cryptoeprint:2011:535, cryptoeprint:2011:091, cryptoeprint:2017:214}, Obliv-C~\cite{cryptoeprint:2015:1153} by Zahur and Evans, and MPyC~\cite{mypc} by Schoenmakers. We performed a detailed literature review for these seven MPC frameworks before narrowing down to three contenders. We ran performance tests using a simple RCT example and synthetic data, against the ABY, EMP-toolkit, and SCALE-MAMBA.
\begin{itemize}
\item \textbf{Performance.} We consider the following metrics when performance testing the frameworks and the details of implementation and their full comparison are provided in section~\ref{sec:eval}. Here is a short summary.
\begin{itemize}
\item Data Volume: EMP-toolkit can handle up to one million rows of data per 16GB of memory (4 million rows with 64GB) in the same circuit. ABY can handle up to 100K rows of data with 16GB of memory, but breaks when the circuit size hits $2^{32}-1$ even if more memory is allocated.
\item Time to Compute: In single threaded tests, EMP-toolkit is about 14x faster than SCALE-MAMBA and ABY is about 45x faster than SCALE-MAMBA.
\item Lag Tolerance: Sub-millisecond latency does not dramatically improve performance, but even moderate latency (60ms) can result in a 7-10x slowdown compared to 1ms.
\item Parallelization Speedup: When we extend EMP-toolkit with our sharding design, the new implementation is about 81x faster than single threaded SCALE-MAMBA.
\end{itemize}
\item \textbf{Optimization.} EMP-toolkit identifies and fixes bottlenecks in various building blocks for secure computation that makes it more suitable compared to schemes without any optimization such as Obliv-C.
As an example, EMP-toolkit uses Streaming SIMD Extensions (SSE) to improve the performance of oblivious-transfer extension, and improves the efficiency of the XOR-tree technique to avoid high (non-cryptographic) complexity when applied to large inputs. Theoretically, their optimizations reduce the cost of processing the circuit evaluator’s input by 1000x for 216-bit inputs, and even more for larger inputs. ABY use transition between three types of sharing --- Arithmetic, Boolean, and Yao --- to improve performance. Our Private RCT cannot handle multiple types of sharing and we see this as a potential future direction.
\item \textbf{Ease of Use.} SCALE-MAMBA is remarkably easy to use since the MAMBA language is similar to python. It also benefits from rich documentation and an active mailing list. EMP-toolkit has ample code samples and is an easy language, but does not have a well documented source code.
It was very difficult to implement the lift game in the Fancy-Garbling framework which is a potential blocker for future developments as well. To use ABY, the protocol should be described in the circuit model for computation, which can be difficult.
\item \textbf{Architecture and Extensibility.} For each framework, we look into its architecture to evaluate if it is modular enough to extend in the future.
The EMP-toolkit has a large protocol set that is highly modular since each protocol is usable independently.
The backend cryptography library of EzPC is ABY and as a result ABY is more modular compared to EzPC and its model is easier to follow and extend. EzPC extends ABY with machine learning functionality that is not immediately useful to RCTs.
\item \textbf{Multiparty Support.} It is most common to evaluate RCTs between two parties, but there are realistic scenarios for 3 or more parties contributing data. EMP-toolkit and SCALE-MAMBA work in both two party and multiparty settings. ABY works only in the two party setting. Frigate~\cite{mood2016frigate} , PICCO~\cite{zhang2013picco} and TF-encrypted~\cite{TFEncrypted} cannot handle the two party case, thus we did not look into them deeply.
\item \textbf{Special Dependencies.}
EMP-toolkit and ABY do not require any special hardware. In comparison,
SCALE-MAMBA has a HSM hardware-dependent setup phase for key generation. EzPC requires trusted execution environment such as SGX to be secure against malicious adversaries.
\item \textbf{Completeness.}
EMP-toolkit provides a complete solution to secure computation problem including an efficient compiler and primitive cryptography protocols (circuit garbler and OT). Obliv-C's focus is on improving the efficiency of compiler, but it runs on a very simple and not optimized underlying circuit garbler and OT.
Frigate~\cite{mood2016frigate}, CBMC-GC~\cite{10.1145/2382196.2382278} and FRESCO~\cite{FRESCO} are the only compilers and do not provide a complete secure computation service.
MPyC is suitable for rapid prototyping and teaching but not for production.
\end{itemize}
\section{Performance Evaluation Results} \label{sec:eval}
In this section, we present performance characteristics of a simple lift game in SCALE-MAMBA, ABY and EMP-toolkit, including tests for high data volume and high latency between nodes. We also show the speedup on EMP-toolkit from running multiple processes per container, and multiple containers. We ran all performance tests on AWS EC2 instances with 64 GB of memory and 16 CPUs unless otherwise noted.
\subsection{Performance Testing}
We began by testing the lift game with each party only using a single container. We ran each experiment ten times and provide the average among all of them as the result.
\subsubsection{Performance against input size}
We tested ABY and EMP-toolkit for different data volumes with results shown in Figure \ref{fig:volume}. We discovered that both scaled linearly in time with increased data volume. We were able to scale EMP-toolkit by adding more memory (about 16GB per 1M rows), but we were unable to scale ABY with more memory. ABY failed with an out of memory error between 100K and 110K rows of data, regardless of how much memory was allocated. We ran input size tests in the same AWS Availability Zone.
\begin{figure}[h]
\centering
\includegraphics[scale=0.5]{volume_test.eps}
\caption{ Wall time for ABY and EMP-toolkit at different data volumes.}
\label{fig:volume}
\end{figure}
\subsubsection{Performance against network delay}
We then tested the same game but located the different parties in different AWS zones/regions; results shown in Figure \ref{fig:ping}. For lowest latency, we placed the parties in the same Availability Zone (0.1 ms average latency). We then placed them in different Availability Zones in the same region (1 ms average latency). We also tested different AWS regions (60 ms average latency). We found that all three protocols experienced severe performance degradation with even modest network latency. We ran network delay tests using 10K input rows.
\begin{figure}[h]
\centering
\includegraphics[scale=0.5]{ping_test.eps}
\caption{ Wall time for ABY, EMP-toolkit, and SCALE-MAMBA for different ping times.}
\label{fig:ping}
\end{figure}
\subsection{Performance of the Protocol at Scale}
We ran additional performance tests to measure scalability of conversion lift game in EMP-toolkit.
\subsubsection{Scaling Up --- Local Concurrency}
We tested the effect of running simultaneous games from the same hardware. We used 4M rows of data total in each case --- more simultaneous games means each game has to calculate fewer rows, e.g., with 8 games each game only needs to calculate 500K rows. The running time of each case is shown in Table \ref{scaleuptable}. The number represents amortized time which is calculated as sum of all running time across instances divided by the number of game instances ran.
\begin{table}[h!]
\centering
\footnotesize
\begin{tabular}{| c | c | c | c | c | c |}
\hline
\# of concurrent calculations & 1 & 2 & 4 & 8 & 16 \\
\hline
running time (min) & 74.43 & 40.06 & 21.79 & 14.94 & 12.77 \\
\hline
\end{tabular}
\vspace{5pt}
\caption{Running time against different numbers of concurrent game instances with fixed total input size.}
\label{scaleuptable}
\end{table}
Table \ref{scaleuptable} shows that the running time decreases sub-linearly with the number of concurrent running games, with the most efficient gains coming from a concurrency of 4 games.
\subsubsection{Scaling Out --- Distributed Calculation}
We tested the effect of using many containers simultaneously via our new private sharing mechanism to process a very large RCT with 500M rows of data. We used AWS Fargate containers with 4 CPU and 30GB memory. Each container ran single-threaded. We ran the experiment three times and the average running times with 10 and 50 containers are listed in Table \ref{scaleouttable}. The results show that the running time decreases linearly with the number of simultaneous containers per party, demonstrating negligible coordination overhead.
\begin{table}[h!]
\centering
\footnotesize
\begin{tabular}{| c | c | c |}
\hline
\# of containers on each party & 10 & 50 \\
\hline
running time (min) & 1524 & 252 \\
\hline
\end{tabular}
\vspace{5pt}
\caption{Running time against different numbers of Fargate containers on each party.}
\label{scaleouttable}
\end{table}
\subsubsection{Private Conversion Lift, Results}
We ran a conversion lift study with 500M rows of synthetic data which was sharded into 50 shards of 10M rows.
We let each worker handle one shard (10M rows) and we leveraged one coordinator to handle 50 workers.
In the test run, the publisher called the coordinator that started 50 Fargate containers in an Elastic Container Service (ECS) in a Virtual Private Cloud (VPC) that is connected to the advertiser's VPC through VPC peering. Once the publisher's workers are running, it outputs the IPs of these workers which we shared with the advertiser. The advertiser then called its coordinator to start 50 Fargate containers that run conversion lift along with the publisher's containers. Total running time was 4hr 12min, and the AWS bill of the study is estimated to be \$1600.
\section{Related Work}
\label{related}
The closest line of work to this paper are protocols that compute a function on the intersection of two private sets also known as join and compute protocols. Circuit-based constructions such as~\cite{huang2012private, pinkas2015phasing, pinkas2018efficient, pinkas2019efficient} support arbitrary computation on the intersection by reducing the problem to executing private equality tests using a general-purpose MPC protocol. These constructions are complex, have larger communication costs, but are more generalizable.
Another method to solve the problem is to extend classic PSI wherein two parties compute and reveal the intersection without revealing anything else about the two sets. Custom DDH-style protocols focus on computing the cardinality or linear functions of the intersection~\cite{de2012fast, ion2020deploying}. They are simpler and more communication-efficient but so far have only enabled a limited set of computations on the intersection.
Our design combines the DDH-style and the circuit-based approaches, in that we take advantage of the fast communication-efficient DDH protocol which can scale nicely when composed with a lightweight yet general-purpose downstream circuit based computation.
There is a small but important literature on differentially private statistical inference, which we advance on \cite{dorazio2015differential, karwa2017finite, brawner2018bootstrap, evans2019statistically, ferrando2020general, du2020differentially, covington2021unbiased}. We have described some of this in the body of the paper.
\section{Summary}
We establish that a sharded version of EMP-toolkit is an appropriate choice for building a system able to handle a large advertising effectiveness RCT study. Further research is necessary to develop systems capable of handling the full breadth and volume of studies performed today. MPC is a very active field of research, and we expect to see new libraries emerge, and for future improvements to some existing tool sets. We hope that this research gives context and inspires open source library developers to tackle scaling, hard memory limits, strong coupling between compilers and protocols, and general usability.
\subsection*{Disclosure Statement}
The authors have no conflicts of interest to declare.
\vspace{-0.1cm}
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction}
Tidal interactions are known to drive the late evolution of short-period planetary systems, like Hot-Jupiters orbiting around their host star, and in our solar system the satellites around Jupiter and Saturn \citep[e.g.,][]{O2014,M2019}. In particular, the dissipation of tides in the convective envelope of low-mass host stars and giant planets can modify the spin of the tidally pertubed body, the orbital period and the spin-orbit angle of the perturber \citep[e.g.][]{H1980,FR2006,L2012,BM2016,DM2018}.
Inertial waves, which are driven by tidal forcing and restored by the Coriolis acceleration, are an important source of tidal dissipation in stellar \citep{OL2007,BO2009,BM2016} and planetary convective zones \citep{OL2004}, where the action of turbulent motions on tidal flows is most often modelled as an effective frictional force or a viscous force with an effective viscosity that is much larger than the molecular viscosity \citep[e.g.,][]{Z1966a,Z1977,DB2020}.
For coplanar and circular systems, inertial waves are excited so long as the companion orbits beyond half its corotation radius (the orbit where the host’s rotation frequency is equal to the mean motion).
Low-mass stars from K to F spectral type and giant gaseous planets both harbour a convective envelope surrounding a radiative and a solid (or diluted) core, respectively \citep[e.g.][]{K2012,DC2019}.
In these objects, inertial waves then propagate in a spherical shell and do not form regular normal modes of oscillation as in spherical \aava{and ellipsoidal geometries \citep[respectively]{G1969,B1889}}. In contrast, they can focus on limit cycles also called attractors of characteristics \citep{ML1995} that are confined within the convective envelope \citep[see also][]{RV1997}.
With a non-zero viscosity, attractors take the form of shear layers where the tidal wave’s energy and angular momentum can be deposited by viscous dissipation \citep{RG2001}. Besides, viscous dissipation across shear layers can be more important as viscosity is weaker, as demonstrated notably by \cite{OL2004}, and \cite{AM2015}. In that respect, tidal dissipation of inertial waves can compete with the dissipation of gravito-inertial waves in the radiative core or be greater by several orders of magnitude than the dissipation of equilibrium tidal flows in the convective zone \citep[i.e., the non-wave like fluid's response; see, e.g.,][]{OL2007}.
The dissipation of tidally-forced waves can have a great impact on the orbital and rotational evolution of the system \citep{AL2014,BM2016,GB2018,BR2019}.
Moreover, the dissipation of the stellar dynamical and equilibrium tides varies significantly along the evolution of the star, and is highly
dependent on stellar parameters like the mass, the angular velocity, and the metallicity of stars \citep{M2015,GB2017,BG2017}. This makes desirable the inclusion of all stellar processes on tidal interaction, in particular differential rotation.
The frequency-averaged tidal dissipation is often used to quantify the response of a body subject to tidal perturbations \citep{OL2004,JG2008}. Yet, the dissipation of a tidally-forced inertial wave is strongly correlated with the presence of an attractor at a specific eigenfrequency of the spherical shell \citep[see][]{O2009,RV2010}. Tidal dissipation at a given frequency may then alter differently each orbital and spin elements of the two-body systems as postulated for instance by \cite{L2012} to explain the survival of hot-Jupiters with completely damped spin-orbit angle, and revisited by \cite{DM2018} with an improved treatment of dynamical tides in the convective region. In addition in the context of Jupiter and Saturn moon systems, \cite{FL2016} and \cite{LF2018} also investigated the dependence in frequency of tidal dissipation to explain rapid outward migration of the moons, through resonant locking of tidally-forced internal modes in the giant gaseous planets. This concept could for example explain the high dissipation observed in Saturn as derived from astrometric measurements at the frequency of Rhea \citep{LJ2017}, and at the frequency of Titan \citep{LC2020}.
Furthermore, the fact that all layers in a star or a planet do not rotate at the same speed, i.e. differential rotation, is rarely taken into account in the determination of tidal dissipation. Yet, differential rotation seems ubiquitous in low-mass stars and giant gaseous planets. The Sun's surface is rotating in $\sim25$ days at the equator versus $\sim35$ days near the poles, and a latitude-dependent rotational gradient has also been observed in the Sun's convective envelope thanks to helioseismology \citep{SA1998,TC2003}. Through asteroseismology, latitudinal shears have been found to be comparable to that of the Sun for Sun analogs \citep{BB2019}, and can be even larger for solar-like stars \citep{BB2018}. Essentially, differential rotation in low-mass stars depends on the effective temperature \citep{BCC2005,BJ2017}, and seems to be more important as the convective envelope is thinner.
Solar-like and anti-solar-like (with faster poles and slower equator) rotation profiles are expected for G and K-type stars based on 3D numerical simulations \citep[see in particular][]{BS2017,BS2018}, while cylindrical rotation profile is expected for fast rotators \citep{GW2013}.
Regarding giant gaseous planets in our solar system, the extent of zonal winds, which are visible on their surface as running lengthwise bands, has been recently constrained by the probes Cassini and Juno. They extend to $3000\,\kilo\meter$ depth for Jupiter \citep{KG2017}, while they penetrate down to 9000 km in Saturn \citep{GK2019}. Thus, the outermost molecular convective envelopes \citep{MW2019,DC2019} are the seat of cylindrical differential rotation.
The study of the impact of differential rotation on the propagation and dissipation properties of inertial modes of oscillation began with the work of \cite{BR2013}. They examined the impact of either a shellular (radial) or a cylindrical rotation profile on free inertial waves in an incompressible background, by means of a Wentzel-Kramers-Brillouin-Jeffreys (WKBJ) linear analysis for an inviscid fluid and by solving the linearised hydrodynamics equations for a viscous fluid via a spectral code.
Their linear analysis highlighted major differences compared to the case of solid-body rotation. Two regimes of propagation have been found, in which inertial modes of oscillation can develop along curved paths of characteristic in the entire convective shell (which the authors named D modes), or in a restricted region of the convective shell, encompassed between a turning surface and one of the shell's boundaries (DT modes). Compared to solid-body rotation, the frequency range of propagation of inertial modes is broader. \cite{BR2013} also pointed out strong dissipation of wave energy at corotation resonances where the Doppler-shifted wave frequency vanishes within the fluid. All these new properties have been retrieved by \cite{GB2016}, who in turn examined a conical (latitudinal) rotation profile, which is typical of low-mass (F- to K-type) stars. They also confirmed the existence of unstable inertial modes (i.e., modes with positive growth rate) at corotation resonances, which were found only for shellular rotation in \cite{BR2013}. Tidal forcing of inertial waves with conical rotation has been introduced by \cite{GM2016} within a linear numerical exploration, which also underlined the strong dissipation of inertial waves at corotation resonances, particularly at low viscosities. \cite{FB2014} also studied tidally-forced inertial waves, but through non-linear numerical simulations. Differential rotation was triggered in their simulations by tidal waves depositing energy and angular momentum in an initially uniformly rotating spherical shell. In some cases, they observed hydrodynamical shear instabilities when the Ekman number (the ratio between the viscous and Coriolis accelerations) is sufficiently small.
Understanding how inertial waves interact with corotation resonances is thus a key issue in quantifying tidal dissipation, especially since waves may deeply interact with the background flow at this particular location, which in turn may alter the background flow \citep[as it was proposed first by ][for terrestrial mountain waves]{EP1961}. In binary systems and for late-type stars, \cite{GN1989} have shown that the angular momentum transported by gravity waves and exchanged at corotation can lead to the successive synchronisation of the layers, from the base to the top of the radiative envelope.
More generally, a body of work in various domains from astrophysical disks \citep[e.g.][]{GT1979,BM2008,LB2009,TL2009} to geophysical fluid dynamics \citep[e.g.][]{B1966,YT1984} has tried to understand the properties of wave propagation and dissipation around corotation, and more generally at all special locations in fluids that correspond to singularities in the linear wave propagation equation. We will refer to them as critical levels in the following \citep{M1986}, or to critical layers in the case of a viscous medium. This distinction is analogous to that between shear layers and attractors of characteristics that are kind of singularities for the governing equation of inertial waves in a spherical shell. The aforementioned singularities can act very differently, with either severe absorption at the critical level \citep[like in ][for stratified vertical shear flows]{BB1967}, or no attenuation if the wave propagates in a peculiar direction \citep[for stratified vertical shear flows with rotation and magnetism]{J1967,A1972,G1975}. In other cases, a critical level may even give rise to wave amplification under certain conditions related to the first and second derivatives of the mean flow velocity \citep[for barotropic and stratified shear flows, respectively]{LT1978,LB1985}. These studies have in common the use of an invariant quantity (the Reynolds stress or the wave action for rotating or magnetic flows) as a diagnostic tool to interpret the role of the critical level in terms of energy transmission and to quantify exchanges between the wave and the mean flow \citep{EP1961,B1966}.
In light of these various studies, it is necessary to consider carefully corotation in differentially rotating convective zones.
A local model can notably allow us a detailed understanding of physical processes at critical levels.
While the propagation through a critical level of gravito-inertial waves in stratified shear flows and of Rossby waves in baroclinic and barotropic flows has been largely studied in the past decades, the behaviour of inertial waves in a latitudinal sheared flow with critical levels has been poorly investigated so far \citep[e.g.][for a review]{L1988}. This is why we develop in this work a local Cartesian shearing box model to understand the complex interplay between tidal waves and zonal flows near critical levels. The concept of a shearing box \aava{for tidal flows} has been introduced by \cite{OL2012} to investigate the interactions between \aava{large-scale tidal perturbations} and convective motions. In our model, we focus on latitudinal differential rotation of the mean flow, varying the box orientation to model either cylindrical or conical rotation. The behaviour of free inertial waves in this framework is then examined near critical levels using both analytical and numerical approaches.
This paper is organised as follows.
In Sect. \ref{sec2}, we describe the local shear model with its main assumptions and the system of governing equations.
In Sect. \ref{sec3}, we establish a second-order ordinary differential equation (ODE) for the latitudinal perturbed velocity, and we derive the propagation properties of inertial waves for an inviscid fluid.
This \aava{ODE} is solved near each critical level for both conical and cylindrical rotation profiles, and we interpret energy flux exchanges between the waves and the mean flow. We use in Sect. \ref{secnum} a three-layer numerical model to test our analytical predictions at critical levels. Viscosity is included and non-linear mean flow profiles are also used.
Astrophysical applications are discussed in Sect. \ref{astrodi} with implications for low-mass stars hosting close exoplanets and giant gaseous planets in our solar system. In sect. \ref{conclu}, we summarise the main results of the paper, and discuss some perspectives and caveats.
\section{Local Cartesian model including differential rotation}
\label{sec2}
\begin{figure}
\centering
\includegraphics[width=0.3\textwidth]{figures/box2.pdf}
\caption{Sketch of the local Cartesian box in the convective region of a low-mass star or giant planet. Global spherical coordinates like the depth $r_0$, the inclination of the box $\theta_0$, and the co-latitude $\theta$ of a point of interest $M$ inside the box, are shown to ease the analogy between the spherical and the Cartesian geometries.}
\label{box}
\end{figure}
\subsection{Presentation of the model}
The local model takes the form of an inclined sheared box, centred at a point $C$ of a convective shell, as illustrated in Fig. \ref{box}.
The inclined box model has already been used by \citet{AM2015} to characterise analytically the properties of tidal gravito-inertial waves in the presence of viscous and thermal diffusion in stably stratified or convective regions, and by \citet{AB2017} in layered semi-convective regions in giant planets interiors \citep[see also][for two-dimensional numerical simulations of inertial wave attractors]{JO2014}. The local coordinate system $(x,y,z)$ corresponds to the local azimuthal, latitudinal and radial direction of global spherical coordinates, respectively, as presented in Table \ref{tab_corres}.
The mean flow velocity $\bo U$ is directed along the local azimuthal axis ${\bo e}_x$ (we neglect possible meridional flows) and differential rotation is embodied by a latitudinal shear $\partial_yU$. As the box is tilted by an angle $\theta_0$ relative to the rotation axis, the rotation vector in the local coordinate system is
\begin{equation}
2\bo\Omega_0=(0,2\Omega_0\sin\theta_0,2\Omega_0\cos\theta_0)=2\Omega_0\,(0,\widetilde{f},f),
\end{equation}
where $\Omega_{0}$ is the rotation frequency of the star at the pole, $\widetilde{f}$ and $f$ are the normalised horizontal and vertical Coriolis components, respectively. Note that the inclusion of both these components means that we go beyond the traditional $f$-plane approximation \citep[see also,][]{GZ2008}.
Furthermore, we make several hypotheses to model wave propagation in a latitudinal shear flow.
The buoyancy acceleration is kept in the fluid equations for the background flow. \aava{The effective gravity acceleration $\bo g$ also includes the centrifugal acceleration, the fluid's angular velocity being assumed small compared to the critical angular velocity $\sqrt{GM/R^3}$ where $G$, $M$ and $R$ are the gravitational constant, the mass and the radius of the body, respectively. Thus, the geometry of the body is close to spherical. Furthermore,} the vector $\bo g$ is supposed to be uniform and constant in the whole box. This requires that the typical length of the box $L$ \aava{satisfies} $L\ll H_p$, where $H_p=-p(\mathrm{d}z/\mathrm{d}p)$ is the vertical pressure scale height, with $p$ the pressure. We can assume this because tidally excited waves are expected to have small-scale structures \citep{OL2004,RV2010,AB2017}. Moreover, the dimensions of the box are chosen to be small compared to the depth of the convective envelope so as to remove curvature effects.
\subsection{Mean flow profile}
\label{mean_flow_sect}
In global spherical geometry, the mean flow based on a conical rotation profile $\Omega(\theta)$ is written \citep[e.g. in][]{GB2016}:
\begin{equation}
\bo u=\boldsymbol\Omega\times\bo r=r\sin\theta\Omega(\theta)\bo e_\varphi,
\label{mean_flow_g}
\end{equation}
where $\bo e_\varphi$ is the azimuthal unit vector, $r$ and $\theta$ are the radius and colatitude, respectively. We introduce $\bo u_0=r\sin\theta\Omega_0\bo e_\varphi$ the mean flow at a point $M$ inside the box (see Fig. \ref{box}) without differential rotation, where we remind that $\Omega_0$ is the spin frequency at the pole. We shall also use the shear contrast $\delta\Omega=\Omega(\theta)-\Omega_0$, i.e. the difference between the angular frequency at colatitude $\theta$ and at the pole. The shear contrast is positive for the Sun since the equator rotates faster than the pole, and negative for anti-solar-like rotating stars.
Using the notations of Fig. \ref{box}, the centre $C$ of the box is located at a distance $r_0\sin\theta_0$ from the rotation axis. Accordingly, the \aava{latitudinal coordinate} of the point $M$ in the local frame is
\begin{equation}
y=r\sin(\theta_0-\theta).
\end{equation}
It should be noted that the radial coordinate $r$ of the point $M$ in spherical geometry can be written as $r=r_0+z$. Nevertheless, we neglect vertical displacements in the expression of the local shear, because we are interested in how the (one-dimensional) horizontal shear affects the wave dynamics while a lot of studies on differential rotation in stars have focused on the vertical shear \citep[e.g.][]{MP2004,DM2009,AM2013,MP2018}.
Since $y/r$ and so $\theta_0-\theta$ are small, we have written in Table \ref{tab_corres} the correspondences in terms of mean flows and shears between the two geometries.
\begin{table}
\begin{tabular}{c|cc}
Geometry &Local Cartesian &Global Spherical \\\hline\hline
\rule[-1ex]{0pt}{4ex} Basis &$(\bo e_x,\bo e_y, \bo e_z)$ & $(\bo e_\varphi,-\bo e_\theta, \bo e_r)$ \\
\rule[-1ex]{0pt}{4ex} \begin{tabular}{c} Conical coordinate \end{tabular} & $y/r_0$ & $\theta_0-\theta$\\
\rule[-1ex]{0pt}{4ex} Mean flow & $\boU$ & $\bo u-\bo u_0$ \\
\rule[-1ex]{0pt}{4ex}Conical shear & $\partial_yU$ & $-\partial_\theta(\sin\theta\delta\Omega)$ \\
\end{tabular}
\caption{Correspondence between local and global coordinate systems.}
\label{tab_corres}
\end{table}
As an example, the shear contrast from solid-body rotation used by \cite{GB2016} was:
\begin{equation}
\delta\Omega(\theta)=\Omega_0\chi\sin^2\theta,
\label{law}
\end{equation}
where $\chi$ is the magnitude of the shear between the equator and the pole. Performing a second-order Taylor expansion around a fixed colatitude $\theta_0$, such that $\theta=\theta_0-y/r_0$ and at a specified depth $r_0$ inside the convective region, the local mean flow $U$ can be recast as
\begin{equation}
\begin{aligned}
u-u_0& =r_0\Omega_0\chi\sin^3\theta\\
&\simeq \Omega_0\chi\left[r_0\sin^3\theta_0-3\cos\theta_0\sin^2\theta_0 y\textcolor{white}{\frac{3}{2}}\right.\\
&\hspace{1cm}\left.+\frac{y^2}{r_0}\left(3\cos\theta^2_0\sin\theta_0-\frac{3}{2}\sin^3\theta_0\right)+O(y^3)\right].
\end{aligned}
\label{U_conic}
\end{equation}
We point out that the Taylor expansion
must be pushed further at the pole $\theta_{0}=0$ (and at the pole $\theta_{0}=\pi$ with an opposite sign):
\begin{equation}
u-u_0\simeq-\frac{\Omega_0\chi}{r_0^2}y^3+O(y^4).
\label{cubicpro}
\end{equation}
Accordingly, we can approximate a conical shear as a linear mean flow at the first order when the box is tilted.
We recall that conical shear has been observed in the solar convective zone and is expected in slowly and in moderately rotating solar-like stars \citep[we refer the reader to sect. \ref{appli_lat} for a detailed discussion; see also][]{BG2015,BS2018,BB2018,BB2019}. When the box is at the pole, $y$ becomes the distance from the rotation axis (called hereafter axial distance). Thus, the mean flow mimics a cylindrical differential rotation that can be modelled using a cubic profile in $y$ given Eq. (\ref{cubicpro}). This rotation profile is found in Jupiter and Saturn, as well as in rapidly rotating stars as demonstrated for instance by \cite{GW2013} and \cite{BG2015}.
\subsection{System of equations}
To derive the system of governing equations for tidal waves in the local reference frame, we made several hypotheses.
Stratification terms, which usually drive the propagation of internal gravity waves, have been kept for clarity sake and will be carefully kept or removed after applying the Boussinesq approximation and setting the equations for inertial waves.
Moreover, we assume that the action of turbulence can be modelled as a Rayleigh friction term in the momentum equation with an effective \aava{frictional damping rate} $\sf$. This \aava{simplifies the analytical solution} of the fluid equations compared to the usual modelling of turbulence as an effective viscous force \citep[see in particular][]{O2009}.
The momentum, continuity and thermodynamic equations for tidal waves in a differentially rotating Cartesian framework thus are:
\begin{align}
\df{\bo u} t+2\bo\Omega_0\times\bo u&= -\frac{\bo\nabla p}{\rho}+\bo g-\sf\bo u+\bo f \label{eq:1},\\
\frac{\partial\rho}{\partial t}+\bo\nabla\cdot(\rho\bo u) &=0\label{eq:2},\\
\df\rho t -\frac{1}{c_\mathrm{s}^2}\df p t&=0\label{eq:3},
\end{align}
where $\bo u$, $p$, $\rho$, and $\bo f$ denote the velocity, pressure, density and volumetric tidal forcing, respectively. We have also introduced $c_\mathrm{s}$ the sound speed and \aava{$\df{}{t}=\frac{\partial}{\mathrm{d} t}+\bo u\cdot\bo\nabla$} is the total derivative operator.
All variables are then linearised at first order: zero-order terms correspond to background equilibrium quantities while first-order terms represent the leading perturbation. The local velocity, density and pressure are therefore written:
\begin{equation}
\left\{\begin{aligned}
\bo u&=U(y)\,\bo e_x + \bo u'\\\
\rho&=\rho_0+\epsilon\rho'\\
p&=p_0+\epsilon p'
\end{aligned}\right.,
\end{equation}
where $\bo u=(u,v,w)$ in the local Cartesian basis.
We have introduced the dimensionless parameter $\epsilon$
\begin{equation}
\epsilon=\frac{(2\Omega_0)^2 L}{g},
\end{equation}
where we have used $1/(2\Omega_0)$ a characteristic time scale and $L$ a characteristic length scale of the mean flow. These notations are based on those of \cite{G1975}, and adapted to our model.
In the following, we will work with dimensionless variables using the above scaling, including $2\Omega_{\aava{0}} L$ to scale velocity and $\rho_\mathrm{T} gL$ to scale pressure, with $\rho_\mathrm{T}$ the reference density.
The dimensionless momentum equation of the mean flow is:
\begin{equation}
\epsilon\,\bo n\times\bo{U}=-\frac{\bo\nabla p_0}{\rho_0}-\bo e_z,
\label{eq0}
\end{equation}
with $\bo n$ the unit vector parallel to the rotation axis. Projecting Eq. (\ref{eq0}) into Cartesian coordinates, one can derive:
\begin{equation}
\left\{\begin{aligned}
\partial_xp_0&=0,\\
\partial_yp_0&=-\epsilon\rho_0 fU,\\
\partial_zp_0&=-\rho_0 +\epsilon\rho_0\widetilde{f}U.
\end{aligned}\right.
\end{equation}
At the leading order in $\epsilon$, one can recognise the hydrostatic balance, and at the first-order the geostrophic balance \citep[the set is akin the thermal-wind equilibrium assumption, see e.g.][]{G1975,YT1984}. We underline that tending $\epsilon$ to zero is similar to assuming the Boussinesq approximation. Indeed, all density variations are neglected, except the ones involved in the buoyancy force.
The dimensionless Brunt-Väisälä frequency is
\begin{equation}
N^2=\epsilon^{-1}\left(F\frac{\partial_zp_0}{\rho_0}-\partial_z\ln\rho_0\right),
\end{equation}
where we have introduced the dimensionless number $F=gL/c_\mathrm{s}^2$, which is small when filtering acoustic waves. Consequently, the curl of Eq. (\ref{eq0}) gives
\begin{equation}
\left\{\begin{aligned}
&\partial_x\rho_0=0, \\
&\partial_y\ln\rho_0=\epsilon\left(\widetilde{f}\partial_yU-fUF\right),\\
&\partial_z\ln\rho_0=\epsilon\left(F\widetilde{f} U-N^2\right)-F,
\end{aligned}\right.
\label{rhorel}
\end{equation}
where we neglect the second-order terms in $\epsilon$.
Now, we make several assumptions to treat the propagation of inertial waves. As the convective motions
are essentially adiabatic, the convective zone can be assumed neutrally stratified to a first approximation. Hence, the Brunt-Väisälä frequency $N$ is cancelled out in the third density relationship Eq. (\ref{rhorel}). Moreover, we make the Boussinesq approximation, which means that we \aava{neglect} terms in $\epsilon$ and $F$ in the final set of perturbed equations.
Thus, the dimensionless linearised momentum, continuity, and thermodynamic equations are finally:
\begin{align}
\df{\bo u'} t+v\partial_y\bo U+\bo n\times\bo u'&= -\frac{\bo\nabla p'}{\rho_0}-\frac{\rho'}{\rho_0}\bo e_z-\sf\bo u'+\bo f \label{pert:1},\\
\bo\nabla\cdot\bo u' &=0,\label{pert:2}\\
\df{\rho'}{t}&=-\rho_0v\widetilde{f}\partial_yU\aava{,} \label{rhop}
\end{align}
\aava{where we remind that $v$ is the latitudinal velocity perturbation}.
We emphasise that, although vertical stratification has been filtered in the limit $N$ goes to zero, an horizontal stratification term remains in Eq. (\ref{rhop}). As a result, we consider the inertial waves propagating in the inclined shear box where the mean flow is maintained by the thermal-wind balance.
\subsection{Equilibrium state of the background flow}
\label{aside}
It is noteworthy to discuss the choice of keeping buoyancy forces in the zero-order momentum equation.
Without gravitational forces, the momentum equation for mean dimensional variables is written as a geostrophic balance:
\begin{equation}
2\rho_0\bo\Omega_{\aava{0}}\times\bo U=-\bo\nabla p_0.
\label{stab}
\end{equation}
This balance satisfies the Taylor-Proudman theorem
\citep{R2015}, namely the geostrophic flow is independent of the coordinate parallel to the rotation axis.
When taking the x-axis (the only non-zero) projection of the curl of this equation, one gets the following relationship:
\begin{equation}
(2\bo\Omega_{\aava{0}}\cdot\bo\nabla)(\romU)=0.
\label{relstab}
\end{equation}
Without vertical stratification embodied by the Brunt-Väisälä frequency, nor latitudinal stratification, so for an incompressible fluid, the equilibrium of \aava{a y-dependent} mean flow is not ensured. An alternative to conserve the equilibrium without stratification would be to consider a $z$-dependence of the mean flow.
Such other possibility is not considered in this paper since we are mainly interested in latitudinal mean flow profiles.
Furthermore, in addition to maintaining differential rotation, the latitudinal stratification can allow to construct an invariant that is useful for studying energy transfer at critical levels: the wave action flux. This will further discussed in Sect. \ref{sec3}. Lastly, since $\widetilde{f}=0$ at the poles, the latitudinal stratification term will not appear in the perturbed fluid equations (as we can see from Eq. (\ref{rhop})).
\section{Dynamics of inertial waves at critical levels: analytical predictions}
\label{sec3}
In this section, we investigate analytically the behaviour of inertial waves at critical levels in a non-dissipative fluid at various colatitudes.
For this purpose, we consider perturbations $q$ in the normal mode
\begin{equation}
q(x,y,z,t)=q(y)\exp{i( k_x x+k_zz-\omega t)}+\mathrm{c.c.}
\label{q}
\end{equation}
with $\omega$ the complex inertial frequency, $k_x$ and $k_z$ the real streamwise and vertical wavenumbers, respectively, and c.c. the complex conjugate.
\subsection{\aava{Wave propagation} equation \aava{in the latitudinal direction}}
Using the modal form (\ref{q}) for $\rho$, $p$ and $\bo u$, we solve the set of hydrodynamic equations, Eqs. (\ref{pert:1}) to (\ref{rhop}), for the latitudinal velocity $v$. Considering free inertial waves (i.e. without forcing terms), the set of perturbation equations can be recast into a single second-order ODE for $v$:
\begin{equation}
Av''+Bv'+Cv=0,
\label{Pc}
\end{equation}
where the prime now denotes the derivative according to $y$, and $A$, $B$, and $C$ are the coefficients that can be simplified without friction as follows:
\begin{equation}
\begin{aligned}
A=&\sigma^2-\widetilde{f}^2,\\
B=&-2k_x\widetilde{f}^2\frac{U'}{\sigma}-2ik_z\widetilde{f} f,\\
C=&-k_\perp^2\sigma^2+k_z^2f(f-U')-2ik_xk_z\widetilde{f} f\frac{U'}{\sigma}\\
&-2k_x^2\widetilde{f}^2\frac{U'^2}{\sigma^2}+\frac{k_xU''}{\sigma}\left(\sigma^2-\widetilde{f}^2\right),
\label{coefs}
\end{aligned}
\end{equation}
where $k_{\perp}=\sqrt{k_x^2+k_z^2}$ is the absolute wavenumber in the direction perpendicular to the $y$-direction\aava{, and $\sigma=\omega-k_xU$ is the (dimensionless) Doppler-shifted wave frequency}. We refer the reader to the Appendix \ref{derPc} for the detailed ODE derivation with \aava{friction} and tidal source terms.
Eq. (\ref{Pc}) becomes singular when $A=0$ or $\sigma=0$, and these singular points are called critical levels \citep[see e.g][]{B1966,G1975}. The critical level where the Doppler-shifted frequency equals to zero (i.e. $\sigma=0$) can be met when the mean flow matches the local phase velocity, and is also known as corotation resonance \citep[e.g. in][]{GN1989,GT1979,OL2004}. When the Coriolis acceleration is not taken into account, as to treat internal gravity waves, the corotation resonance is the unique critical level \cite[see e.g.][]{BB1967}.
At colatitudes other than the poles, the critical levels come in three flavours, the corotation $\sigma=0$ and two other critical levels that are defined, in our model, by $\sigma=\pm\widetilde{f}$ (where we remind that $\widetilde{f}$ in the latitudinal component of the rotation vector). These critical levels were similarly reported for vertical shear flows as in the studies of \citet{J1967} and \citet{G1975} for vertical and inclined rotation vectors, respectively. In these work, the Doppler-shifted frequency at critical levels other than the corotation resonance equals to $\pm2\Omega_{v}$ where $\Omega_{v}$ is the vertical component of the rotation vector.
\subsection{Propagation properties}
\label{propa_prop}
\subsubsection{Dispersion relation, group and phase velocities}
The 3D dispersion relation is a fourth-order equation in Doppler-shifted frequency when injecting wave-like solutions in the three directions $x,\ y,\text{ and } z$ in Eq. (\ref{Pc}). In order to understand the main properties of waves at the critical level, we make the short-wavelength approximation as in \cite{BR2013} and \cite{GB2016} in the \aava{meridional plane}. This involves keeping only the second-order derivatives in the $y$ and $z$ directions and it reduces the relation dispersion to a second-order equation when injecting plane wave-like solutions. In the local meridional plane, the \aava{differential} equation reduces to \aava{a Poincaré-like equation}:
\begin{equation}
\left(\sigma^2-\widetilde{f}^2\right)\partial_{y,y}v-2\widetilde{f} f\partial_{y,z}v+\left[\sigma^2+f\left(U'-f\right)\right]\partial_{z,z}v=0,
\end{equation}
\aava{where we recover the Poincaré equation \citep[for the propagation of inertial waves in the inviscid limit,][]{C1922} in the meridional plane when there is no shear ($U'=0$) and at the poles ($\widetilde{f}=0$ and $f=1$).}
Moreover, we set $v\propto\exp{-i(k_z z -k_x x)}$ so as to write the wave dispersion relation for the Doppler-shifted frequency $\sigma$:
\begin{equation}
\sigma^2=\frac{1}{||\bo k||^2}\left[\left(\bo n\cdot\bo k\right)^2-k_z^2fU'\right],
\end{equation}
where $||\bo k||=k_y^2+k_z^2$ is the norm of the wave vector in the \aava{meridional} plane (e.g. for fixed $k_{x}$) like in \citet{BR2013}. Compared to solid-body rotation \citep[see e.g.][]{R2015}, an additional term ($k_z^2fU'$) is present, which accounts for the latitudinal shear. \aava{Assuming that $\sigma^2$ takes positive values} \citep[as in][]{BR2013,GB2016}, we therefore introduce
\begin{equation}
\gamma=\sqrt{(\bo n\cdot\bo k)^2-k_z^2fU'}.
\end{equation}
We can then explicit the phase velocity in the \aava{meridional} plane:
\begin{equation}
\bo v_\phi=\frac{\sigma}{{||\bo k||}^2}{\bo k}=\pm \frac{\gamma}{||\bo k||^3}{\bo k}.
\label{vp}
\end{equation}
In the same way, we can derive the expression for the group velocity in the \aava{meridional} plane:
\begin{equation}
\bo v_\mathrm{g}=\bo\nabla_{\bo k}\sigma=\pm\frac{\gamma}{||\bo k||^3}\left\{-\bo k+\left(\frac{||\bo k||}{\gamma}\right)^2\left[{\bo n}(\bo n\cdot\bo k)-k_zfU' \bo e_z\right]\right\}.
\label{vg}
\end{equation}
Note that without differential rotation, the group velocity reduces to its well-known expression for solid-body rotation \citep[e.g., see][]{R2015}:
\begin{equation}
{\bo v_\mathrm{g}}=\pm \frac{{\bo k}\wedge({\bo n\wedge\bo k})}{||{\bo k}||^3}.
\end{equation}
Moreover, as in solid-body rotation, the group velocity (Eq.~\ref{vg}) and the phase velocity (Eq.~\ref{vp}) lie in perpendicular planes: $\bo v_\mathrm{g}\cdot\bo v_\phi=0$.
When the box is located at the North pole ($\theta_0=0$ in Fig. \ref{box}), by setting $\kappa=1-U'$ we recover
\begin{equation}
\left\{
\begin{aligned}
\sigma^2&=\frac{k_z^2}{||\bo k||^2}\kappa^2\\
\bo v_\phi&=\pm\kappa\frac{k_z\bo k}{||\bo k||^3}\\
\bo v_\mathrm{g}&=\pm\kappa\frac{k_y}{||\bo k||^3}(-k_z\bo e_y+k_y\bo e_z)
\end{aligned}\right.
\end{equation}
as in \citet{LB2009} or \cite{BR2013}, where $\kappa$ can be identified to their epicyclic frequency and $k_y$ corresponds to the cylindrical component of the wavenumber ($k_s$).
\subsubsection{Phase and group velocity at singularities}
We derive in this section the conditions required to meet singularities in terms of wavenumbers and shear, and examine what the implications are for the phase and group velocities.
When the box is inclined, for $\sigma\rightarrow0$ we must have:
\begin{itemize}[label=$-$]
\item$\gamma\rightarrow0$, meaning that $\bm v_\phi\rightarrow\bm0$ while $\ |\bm v_\mathrm{g}\cdot\bm e_y|\rightarrow\infty$ and $|\bm v_\mathrm{g}\cdot\bm e_z|\rightarrow\infty$.
\end{itemize}
\cite{GB2016} found similar results by studying the propagation of free inertial waves in a global frame with conical shear, namely when their parameter $\mathcal{B}$ (which is homogeneous to a frequency and equivalent to our $\gamma$ parameter) goes to zero, the group velocity goes to infinity while the phase velocity cancels out. According to their work, an inertial wave may propagate across the corotation.\\
Now to get $\sigma\rightarrow\pm\widetilde{f}$, we either need:
\begin{itemize}[label=$-$]
\item $|k_y|\rightarrow\infty$ at fixed $k_z$, which implies $\bm v_\phi \rightarrow\bm0$ and $\bm v_\mathrm{g}\rightarrow\bm0$ and means that inertial waves cannot get through the critical level,
\item $|k_z|\rightarrow0$ at fixed $k_y$, which gives $\bm v_\phi\cdot\bm e_z \rightarrow0$ and $\bm v_\mathrm{g}\cdot\bm e_y\rightarrow0$ while $|\bm v_\phi\cdot\bm e_y|\rightarrow\widetilde{f}/k_y$, and $|\bm v_\mathrm{g}\cdot\bm e_z|\rightarrow f/k_y$: the wave may then cross the critical level with some preferential direction.
\end{itemize}
Again, these conditions share some similarities with those observed for corotation in a global spherical geometry. The first aforementioned possibility (first item above) is analogous to the global phase and group velocities tending to zero when $k_s\rightarrow\infty$, with $k_s$ the axial wavenumber in cylindrical coordinates \citep{BR2013,GB2016}. This makes sense since the axial distance is $s=r\sin\theta$, and $y\sim r_0(\theta_0-\theta)$ here. However, the second aforementioned condition (second item above) is slightly different from both these previous works, in that $|k_z|\rightarrow0$ at fixed $k_s$, with $k_z$ the global vertical wavenumber, i.e. along the rotation axis unlike our local vertical wavenumber $k_z$ along the spherical radial coordinate.
We point out that the singularities at $\sigma=\pm\widetilde{f}$ arise in our model because the rotation vector is inclined with respect to the local vertical axis of the box. In the global model of \cite{GB2016}, three conditions for a wave to meet the corotation exist, and these conditions are actually quite similar to the three above conditions for waves in our model to interact either with the corotation $\sigma=0$ or the other critical levels at $\sigma=\pm\widetilde{f}$.
Hence, the local critical levels at $\sigma=\pm\widetilde{f}$ behave partly like the corotation in the global framework, as if we partially broke the degeneracy in the local framework of the origin of the corotation found in the global framework.
When the box is at the North pole, the conditions to meet corotation are similar but lead to different relationships for the phase and group velocities:
\begin{itemize}[label=$-$]
\item$\kappa\rightarrow0$ (i.e. $U'\rightarrow1$) meaning that $\bm v_\phi\rightarrow\bm0$, and $\bm v_\mathrm{g}\rightarrow\bo 0$: the wave is totally absorbed at corotation.
\item $|k_y|\rightarrow\infty$ at fixed $k_z$, which implies $\bm v_\phi \rightarrow\bm0$ and $\bm v_\mathrm{g}\rightarrow\bm0$: same conclusion as in the previous case and analogous case as when the box is tilted,
\item $|k_z|\rightarrow0$ at fixed $k_y$, which gives $\bo v_\phi\rightarrow\bo 0$, $\bm v_\mathrm{g}\cdot\bm e_y\rightarrow0$ while $|\bm v_\mathrm{g}\cdot\bm e_z|\rightarrow \kappa/k_y$: the wave energy does not cross the corotation in the latitudinal direction \citep[equivalent to vertical paths of characteristic in global cylindrical geometry like in ][]{BR2013}.
\end{itemize}
At the north pole\footnotemark\footnotetext{Note that a similar analysis can be undertaken at the South Pole, by using $\kappa=1+U'$.}, we actually have a perfect match with the conditions given by \cite{BR2013} when using a cylindrical rotation profile for the mean flow.
\subsubsection{Energetical aspects}
In this section, we examine the energetic balance associated \aava{with} inertial waves in our inclined shear box model, without assuming the short-wavelength approximation.
This energetic balance does not include potential energy because of the adiabaticity of the convective region but two additional terms appear compared to the solid-body rotation case, coming from the differential rotation.
We denote by $\eta$ and $\zeta$ the displacements along the vertical and latitudinal directions, respectively. Considering that $\omega=k_x c$, where $c$ is the longitudinal phase velocity \citep[e.g. as in][]{BB1967}, we can use the first-order definition
\begin{equation}
\begin{aligned}
v&=\df\zeta t=(U-c)\zeta_x,\\
w&=\df\eta t=(U-c)\eta_x.
\end{aligned}
\end{equation}
It first allows us to express the perturbed density from Eq. (\ref{rhop}) as $\rho=-\rho_0\zeta\widetilde{f}U'$\aava{, where we remind that the symbol $'$ has been dropped out for perturbed quantities}.
Then, by multiplying the momentum equation (\ref{pert:1}) by $\rho_0\bo u$, we can get the energy balance equation:
\begin{equation}
(U-c)\partial_xe_\mathrm{k}+{\bo\nabla\cdot}(p\bo u)=\rho_0\left[-vU' u+\zeta\widetilde{f}U' w-\sf\bo u^2+\bo f\cdot\bo u\right]
\label{balance_e},
\end{equation}
where $e_\mathrm{k}=\rho_0\bm u^2/2$ is the kinetic energy density, and $p{\bm u}$ the so-called acoustic flux.
We now integrate the above energy balance equation over $x$ and $z$, and over one wave period, as the perturbed quantities have a wave-like form in these directions. Further assuming that the box is $\delta$ thick in the $y$-direction, the energetic balance yields:
\begin{equation}
\mathcal{P}_\mathrm{ext}=\mathcal{P_\mathrm{shear}}+D_\mathrm{visc}+\mathcal{P}_\mathrm{f},
\label{power}
\end{equation}
where we have introduced from left to right the power of the external pressures at the boundaries $-\delta/2$ and $\delta/2$ on the perturbed latitudinal flow, the work of the shear, the viscous dissipation, and the forcing power, which read respectively:
\begin{equation}
\begin{aligned}
\mathrm{P}_\mathrm{ext}&=\left[\textcolor{white}{\widetilde{f}}\overline{pv}\textcolor{white}{\widetilde{f}}\right]_{-\delta/2}^{\delta/2},\\
\mathcal{P_\mathrm{shear}}&=-\int_{-\delta/2}^{\delta/2}\romU'\left(\overline{uv}-\widetilde{f}\,\overline{w\zeta}\right)\,\mathrm{d}y,\\
D_\mathrm{visc}&=-\int_{-\delta/2}^{\delta/2}\rho_0\sf\overline{\bo u^2}\,\mathrm{d}y,\\
\mathcal{P}_\mathrm{f}&=\int_{-\delta/2}^{\delta/2}\rho_0\overline{\bo f\cdot\bo u}\,\mathrm{d}y,
\end{aligned}
\end{equation}
where the bar represents the average in the $(z,x)$ plane over one period. Note that the energy density and the acoustic flux in the $x-$ and $z-$directions drop out in Eq. (\ref{power}) when integrating, because of the wave periodicity in those directions. The quantity $\mathcal{P}_\mathrm{shear}$ can also be seen as the power transferred from the mean flow to the perturbation (or conversely) by the Reynolds stress:
\begin{equation}
\tau=-\rho_0(\overline{uv}-\widetilde{f}\,\overline{w\zeta})=-\rho_0\overline{v\left(u+\widetilde{f}\eta\right)},
\label{Re}
\end{equation}
where we have used partial integration and the periodicity of perturbations in the $x$ and $z$ directions.
At the pole, $\widetilde{f}=0$ so we recover the definition of the Reynolds stress in \cite{M1961}, who studied the stability of a 2D stratified $y$-sheared flow, i.e. $\tau=-\rho_0\overline{uv}$. This quantity can also be called the latitudinal flux of horizontal (in the sense $(z,x)$ plane) momentum in reference to the vertical flux of horizontal momentum in stratified $z$-sheared flows. Moreover, we emphasise that the latitudinal flux of energy $\overline{pv}$ is not conserved even in the inviscid free-wave problem. This is due to the presence of the shear, as already stated for example by \cite{EP1961}, who studied stratified vertically sheared flows. They underline that when the mean flow varies with height, the kinetic energy of the mean motion can be converted into wave energy.
Without \aava{friction} and forcing, the $y$-derivative of the latitudinal flux is:
\begin{equation}
\df{}{y}\overline{pv}=-\romU'\overline{v\left(u+\widetilde{f}\eta\right)}.
\label{flux}
\end{equation}
Using the same method as \cite{B1995}, we multiply the $x$-projection of the inviscid force-free momentum equation by $\zeta$:
\begin{equation}
-p_x\zeta=\rho_0(U-c)\zeta u_x+\rho_0(U'-f)v\zeta+\rho_0\widetilde{f} w\zeta.
\end{equation}
By multiplying by $(U-c)$, the latitudinal flux of energy can thus be written as:
\begin{equation}
\overline{pv}=-\rho_0(U-c)\overline{v\left(u+\widetilde{f}\eta\right)}.
\end{equation}
By \aava{differentiating this relationship with respect to $y$}, and by equalising with Eq. (\ref{flux}), one can obtain:
\begin{equation}
-\rho_0(U-c)\frac{\mathrm{d}}{\mathrm{d}y}\left\{\overline{v\left(u+\widetilde{f}\eta\right)}\right\}=0,
\label{rel}
\end{equation}
that is $(U-c)\cfrac{\mathrm{d}\tau}{\mathrm{d} y}=0$ with $\tau$ the Reynolds stress. Eq.~(\ref{rel}) is naturally satisfied at corotation, where $U-c=0$, or if the Reynolds stress is uniform. \cite{BB1967} have shown that the Reynolds stress is discontinuous at a critical level, highlighting exchanges between wave energy and the mean flow.
Compared to the analysis of \cite{B1995} for 3D stratified shear flows, Eq.~(\ref{rel}) is not vectorial, because our base flow is unidirectional.
\subsubsection{Polarisation relations}
For the forthcoming analysis, it is useful to derive expressions of the perturbed projected velocities and the perturbed reduced pressure\footnotemark\footnotetext{The quantity $\Pi$ is actually the enthalpy perturbation but we will use the denomination "reduced pressure" in the following.} $\Pi=p/\rho_0$, namely the polarisation relations (see Appendix \ref{derPi} for more details). In the inviscid free-wave problem, these perturbed quantities can be written in terms of the latitudinal velocity, its derivative, and the shear:
\begin{equation}
\begin{aligned}
\Pi&=\frac{1}{\sigma k_\perp^2}\left\{iAv'+\left[ik_x\frac{U'}{\sigma}A+f(k_z\widetilde{f}-ik_x\sigma)\right]v\right\},\\
w&=\frac{1}{\sigma k_\perp^2}\left\{\left(i\sigma k_z-k_x\widetilde{f}\right)v'+k_x\left[ik_z(U'-f)-k_x\widetilde{f}\frac{U'}{\sigma}\right]v\right\},\\
u&=\frac{1}{\sigma k_\perp^2}\left\{\left(i\sigma k_x+k_z\widetilde{f}\right)v'+k_z\left[ik_z(f-U')+k_x\widetilde{f}\frac{U'}{\sigma}\right]v\right\}.
\end{aligned}
\label{pol}
\end{equation}
Without shear and at $\theta_0=0$, we recover the polarisation relations in the solid-body rotation case \citep[see e.g.][]{R2015}.
\subsubsection{Conservation of the wave action flux}
\label{Co}
While the latitudinal flux of energy is not conserved in the whole domain, there is a conserved quantity, called the wave action flux as introduced
in \cite{G1975}'s paper:
\begin{equation}
\mathcal{A}=\frac{\overline{pv}}{\sigma}=\frac{\rho_0}{2}\Re\left\{\frac{\Pi v^*}{\sigma}\right\},
\label{A}
\end{equation}
which is the latitudinal flux averaged over vertical and longitudinal wavelengths divided by the Doppler-shifted frequency. A general treatment for the derivation of the wave action as a conserved quantity can be similarly found in \cite{AM1978}. The wave action flux is related to the Reynolds stress $\tau$ as $\mathcal{A}=-\tau/k_x$.
By using the expression for the perturbed reduced pressure $\Pi$ derived in the previous section, the wave action flux now reads:
\begin{equation}
\mathcal{A}=\frac{1}{2}\rho_0\Re\left\{\frac{iAv_yv^*+f\widetilde{f} k_z|v|^2}{\sigma^2k_\perp^2} \right\}.
\label{wA}
\end{equation}
Unlike the latitudinal flux of energy, but similarly to the Reynolds stress, this wave action flux is conserved along the latitudinal direction. One can demonstrate that $\mathcal{A}'=0$ in the whole domain except at critical levels, by using the expression for the reduced pressure in Eqs. (\ref{pol}) and the \aava{ODE} (\ref{Pc}). Several works have shown that a properly defined (i.e., conserved) angular momentum transport parameter can be found in $z$-sheared mean flows without rotation \citep{BB1967}, with rotation under the traditional approximation \citep{J1967}, and with rotation under the non-traditional approximation \citep{G1975}.
Verifying the conservation in the whole domain except at critical levels is really important because it brings to the fore energy transfers due to the critical levels. We specify that $\mathcal{A}$ is a measure of wave energy through a surface (in the $(z,x)$ plane) since $\overline{pv}$ is the energy density transported by the group velocity\footnotemark\footnotetext{Note that the velocity of energy density $V_\mathrm{g}$ in the latitudinal direction has been named "group velocity" in the latitudinal direction for obvious physical reasons but it differs from the group velocity defined in Sect. \ref{propa_prop} that depends on latitudinal and vertical wavenumbers unlike $V_\mathrm{g}$.
} $V_\mathrm{g}$ in the latitudinal direction \citep[e.g.][]{BG1968,MB2012}. It should be underlined that the wave action flux has been defined in the inviscid limit and is not conserved when the \aava{friction} is taken into account.
\subsection{Inertial waves at critical levels when the box is tilted}
\label{Fro}
In this section, we analytically investigate waves passing through the various critical levels in the tilted box.
We examine the behaviour of the waves around the corotation $\sigma=0$ and the critical levels $\sigma=\pm\widetilde{f}$ when the box is tilted (for the corotation when the box is at the pole, see Section \ref{polar_case}).
\subsubsection{Critical levels at $\sigma=\pm\widetilde{f}$}
\label{cc_ft}
\begin{figure*}[ht]
\resizebox{\hsize/2}{!}{\includegraphics{figures/colormap_atnu_deg10}}
\resizebox{\hsize/2}{!}{\includegraphics{figures/colormap_atnu_deg80}}
\caption{Transmission rate $T_{\theta_0}$ of a wave passing through any of the critical levels defined by $\sigma=\pm\widetilde{f}$ as a function of the absolute value of the shear Rossby number $|R_\mathrm{o}|=|U'|$ (where $U'$ is scaled by $2\Omega_0$), the ratio of the horizontal wave numbers $\alpha_k=k_z/k_x$, and for a co-latitude of the box $\theta_0=10\degree$ (\textit{left panel}) and $\theta_0=80\degree$ (\textit{right panel}).}
\label{atnu_deg}
\end{figure*}
In this subsection, we treat both singularities $\sigma=\pm\widetilde{f}$ simultaneously. Although Eq. (\ref{Pc}) does not have analytical solutions in general, it is still possible to study the behaviour of an inertial wave close to the critical levels defined by $\sigma=\pm\widetilde{f}$ by approximating the ODE through its first-order Taylor expansion in the vicinity of these singularities, and then by applying the Frobenius method.
We introduce $y_\pm$, the location of the related critical level $\sigma=\pm\widetilde{f}$.
For a linear mean flow profile $U=\Lambda y$, with $\Lambda$ a constant, $y_\pm$ are given by:
\begin{equation}
y_\pm=\frac{\omega\mp\widetilde{f}}{k_x\Lambda}.
\end{equation}
Without any assumption on the mean flow profile, the first-order Taylor expansion of the ODE (\ref{Pc}) near $y_\pm$ is:
\begin{equation}
v''+\frac{\mathcal{B}}{y-y_\pm} v'=0,
\label{PcDL_ft}
\end{equation}
with
\begin{equation}
\begin{aligned}
\mathcal{B}=&1\pm i\frac{k_zf}{k_xU'_{\aava{\pm}}},\\
\end{aligned}
\end{equation}
where the symbol $\pm$
refers to the regular singularities\footnotemark\footnotetext{A singular point $y_0$ of the second-order ODE $v''(y)+p(y)v'(y)+q(y)v(y)=0$ is said regular when the function $p(y)(y-y_0)$ and $q(y)(y-y_0)^2$ are analytical at $y=y_0$.} $y_+$ and $y_-$, respectively\aava{, and $U'_\pm$ is $U'$ evaluated at these singularities.}
The Frobenius method consists in injecting the power function $(y-y_\pm)^\lambda$ in Eq. (\ref{PcDL_ft}), with $\lambda$ a constant to be determined \citep[see e.g.][]{MF1953}. The corresponding indicial equation is then:
\begin{equation}
\lambda(\lambda-1)+\mathcal{B}\lambda=0,
\end{equation}
with solutions:
\begin{equation}
\lambda_\pm=\left\{0,\ \mp i\frac{k_zf}{k_xU_{\aava{\pm}}'}\right\}.
\end{equation}
Therefore, the two independent solutions of Eq. (\ref{PcDL_ft}) can be written as follows:
\begin{equation}
v_{1,\pm}(y)=\sum_{n=0}^{+\infty}a_n(y-y_\pm)^n\ \text{ and }\ v_{2,\pm}(y)=\sum_{n=0}^{+\infty}b_n(y-y_\pm)^{n+\lambda_\pm},
\label{sol}
\end{equation}
where $a_n$ and $b_n$ are complex constants. Both solutions are valid in the vicinity of the critical level around which they are built \aava{in the complex plane}, up to the next singularity if it exists. The coefficients $a_0$ and $b_0$ are unconstrained and depend on boundary conditions, unlike the other factors that can be determined by injecting solutions (\ref{sol}) into the \aava{linearised ODE (\ref{Pc}) around $y_\pm$ at the right order for the desired coefficients.}
Near the critical points $y_\pm$, the total solution $v_\pm$ is well approximated by the lowest orders of $v_{1,\pm}$ and $v_{2,\pm}$:
\begin{equation}
v_\pm(y)\simeq a_0+b_0(y-y_\pm)^{\lambda_\pm}.
\end{equation}
Owing to the existence of a branch point at $y_\pm$ (since $\lambda_\pm$ is complex), reconnecting solutions on either part of the critical levels is not straightforward. This requires both physical and mathematical arguments \citep[see in particular][]{M1961,BB1967,R1998}.
In order to remove degeneracy of the path from positive to negative $y-y_\pm$ (i.e. choose either $\e^{+i\pi}$ or $\e^{-i\pi}$), we make use of a complex inertial frequency $\omega=\omega_\mathrm{R}+i\omega_\mathrm{I}$, assuming the radiation condition $\omega_\mathrm{I}>0$. This condition ensures a non-growing wave toward infinity.
The Taylor expansion of the base flow at first-order in $y-y_\pm$ gives
\begin{equation}
y-y_\pm=\frac{U(y)-U_{\aava{\pm}}}{U'_{\aava{\pm}}},
\end{equation}
and by definition, we have
\begin{equation}
U_{\aava{\pm}}=\frac{\omega\mp\widetilde{f}}{k_x}.
\end{equation}
Consequently, the solution below the critical level is unambiguous in terms of the above solution coefficients, and depends on
\begin{equation}
\sign[\Im\{y-y_\pm\}]=-\sign\left[k_xU'_{\aava{\pm}}\right].
\label{branch}
\end{equation}
In other words, when taking $y-y_\pm$ to decrease from positive to negative values, its complex argument changes continuously from $0$ to $-\sign\left[k_xU'_{\aava{\pm}}\right]\pi$.
Thus, the appropriate path for determining the branch of $(y-y_\pm)^{\lambda_\pm}$ passes under (above) $y_\pm$ as long as $k_xU'_{\aava{\pm}}>0$ ($k_xU'_{\aava{\pm}}<0$) \citep[the same reasoning can be found in ][]{G1975}.
Therefore, the solution on both sides of the critical level $y_\pm$ is:
\begin{equation}
v_{\pm}(y)\simeq\left\{\begin{aligned}
&a_0+b_0|y-y_\pm|^{\mp i k_z f/k_xU'_{\aava{\pm}}}&\text{ for }y>y_\pm\\
&a_0+b_0|y-y_\pm|^{\mp i k_z f/k_xU'_{\aava{\pm}}}\mathrm{exp}(\mp \pi k_z f/|k_xU'_{\aava{\pm}} |)&\text{ for }y<y_\pm
\end{aligned}\right..
\end{equation}
The remaining issue is now to know in which direction the wave is propagating. The second part of the solution can be assimilated to a wave-like solution with the varying latitudinal wavenumber $\mp (k_z f/k_xU'_{\aava{\pm}}) \log|y-y_{\pm}|$.
Moreover, according to Eq. (\ref{wA}), the wave action flux on either side of $y_\pm$ is:
\begin{equation}
\mathcal{A}=
\frac{\rho_0 k_zf}{2\widetilde{f} k_\perp^2}\left\{\begin{aligned}
&|a_0|^2-|b_0|^2&\text{ for }y>y_\pm\\
&|a_0|^2-|b_0|^2\mathrm{exp}(\mp2\pi k_z f/|k_xU'_{\aava{\pm}}|)&\text{ for }y<y_\pm
\end{aligned}\right..
\label{atnu}
\end{equation}
The group velocity gives the direction towards which the energy is transported, recalling that $V_\mathrm{g}\mathcal{E}=\overline{pv}$ with $V_\mathrm{g}$ the group velocity and $\mathcal{E}$ the local energy density. By consequence, $\sign({V_\mathrm{g}})=\sign({\mathcal{A}\sigma})=\mp\sign(fk_z)$ for the solution featuring the coefficient $b_0$. If $k_zf$ is positive, this wave transports energy downward (upward) across the critical level $\sigma=\widetilde{f}$ ($\sigma=-\widetilde{f}$). If $k_zf$ is negative, the wave transports energy upward (downward) across the critical level $\sigma=\widetilde{f}$ ($\sigma=-\widetilde{f}$). In all cases, the action flux of the wave with the
amplitude $|b_0|$ will be transmitted (in the direction given by the sign of $k_zf$ and the critical level $y_+$ or $y_-$) by a factor $T_{\theta_0}$ where
\begin{equation}
T_{\theta_0}=\exp(-2\pi\frac{|k_zf|}{|k_xU'_{\aava{\pm}}|})=\exp(-2\pi\frac{|\alpha_k\cos\theta_0|}{|R_\mathrm{o}|}),
\end{equation}
with $\alpha_k=k_z/k_x$ and $R_\mathrm{o}=U'_{\aava{\pm}}$, after passing through the critical level. Such wave will always be attenuated since $T_{\theta_0}\leq1$.
The transmission factor $T_{\theta_0=10\degree}$ and $T_{\theta_0=80\degree}$ are displayed in Fig. \ref{atnu_deg} in terms of the absolute value of the shear Rossby number $|R_\mathrm{o}|$ and the ratio of wave numbers $\alpha_k$.
The lower the amplitude of the Rossby number and the lower the inclination, the more likely the wave is to be strongly attenuated at any $\alpha_k$. We remind that a low Rossby number refers either to fast rotating stars or to low differential rotation.
At the equator, one should note that \aava{$f=0$ so there is no transmission nor exchange of wave action flux near the critical levels $y_\pm$ in the inviscid limit (see Eq. (\ref{atnu}))}. Results are the same for $\theta_0+k\pi/2$ with $k\in\{0,1,2,3\}$, and for negative Rossby numbers.
However, it has to be emphasised that the cases where the inclination satisfies $\theta_0=k\pi$ with $k\in\{0,1\}$ are not well described by the attenuated factor $T_{\theta_0}$ and require a specific treatment, as discussed in Sect. \ref{polar_case}.
It is important to note that, with fixed parameters $\{k_z, \theta_0,R_\mathrm{o}\}$, the attenuation of the wave action flux is specific to a single direction of wave propagation, i.e. the solution featuring the coefficient $b_0$. The solution of coefficient $a_0$ is not affected by the attenuation. It is the so-called valve effect introduced by \cite{A1972} in the context of hydromagnetic waves in a rotating fluid. It was also evidenced by \cite{G1975}, and further discussed in \cite{G1979} for magneto-gravito-inertial waves in an inviscid and compressible z-sheared fluid.
\subsubsection{Inertial wave crossing corotation}
\label{corot_inclined}
We perform the same analysis as in the previous section to treat the corotation point $y_0$ where $\sigma=0$ (i.e. $U(y_0) \aava{=U_0}=\omega/k_{x}$). \aava{The linearised \aava{ODE}~(\ref{Pc}) near the corotation using the}
Taylor expansion of \aava{$\sigma$ and $U$} \aava{at the lowest orders} is:
\aava{\begin{equation}
\begin{aligned}
&v''+\left(-\frac{2}{y-y_0}+2ik_z\frac{f}{\widetilde{f}}-\frac{U_0''}{U'_0}\right)v'\\&+\left[\frac{2}{(y-y_0)^2}-\frac{2ik_zf}{\widetilde{f}(y-y_0)}+\frac{U_0''}{U_0'(y-y_0)}\right]v=0,
\label{Pcorot}
\end{aligned}
\end{equation}}
where $U_{\aava{0}}'$ and $U_{\aava{0}}''$ are \aava{the first and second derivatives of the mean flow profile $U$} evaluated at the critical level $y_0$.
The singularity at the corotation is a regular singularity and we can use again the Frobenius method. The indicial equation has solutions $\lambda=\{2, 1\}$. Since the difference between the two \aava{values} of the exponent \aava{$\lambda$} is an integer, one expects a second independent solution $v_2$ of Eq. (\ref{Pcorot}) \aava{ that includes a logarithmic part such as \citep[e.g.][]{SH2002}:
\begin{equation}
v_2(y)=\sum_{n=0}^{+\infty}b_n(y-y_0)^{n+1}+L\ln(y-y_0)v_1(y)
\end{equation}
with $v_1(y)=\sum^{+\infty}_{n=0}a_n(y-y_0)^{n+2}$ the first solution, and $a_n$, $b_n$, and $L$ complex coefficients. However, when injecting $v_1+v_2$ in Eq. (\ref{Pcorot}), one finds $L=0$, meaning that a sole polynomial solution in the form
\begin{equation}
v(y)=\sum_{n=0}^{+\infty}c_n(y-y_0)^{n+1}
\label{solln}
\end{equation}
includes all the solutions of Eq. (\ref{Pcorot}), with $c_0=b_0$ and $a_1=b_1+a_0$ determined by boundary conditions, and $c_{n,~n\in\mathds{N}^*\backslash\{1\}}=b_n+a_{n-1}$ determined by recurrence via the expansion of Eq. (\ref{Pc}) around $y_0$. As a result, the wave action flux given by Eq. (\ref{wA}) becomes
\begin{equation}
\mathcal{A}=\frac{\rho_0}{2k_x^2U'^2_0k_\perp^2}\left[\widetilde{f}^2\Im{c_0^*c_1}+f\widetilde{f} k_z|c_0|^2\right]
\label{wAconcor}
\end{equation}
just below and above the corotation, and it is continuous there, similarly as in \citet{G1975}, but here without being restricted to a linear mean flow profile. Hence, no transfer of wave action flux is expected at corotation in the inviscid limit when the box is inclined relative to the rotation axis (i.e. for conical differential rotation), regardless of the mean flow profile. This result also holds true when the box is located at the equator.}
Like in the works of \cite{G1975} and \cite{J1967}, it is tempting to investigate the asymptotic behaviour of a wave when $y\rightarrow\infty$, in order to better constrain the propagation of waves through one or multiple critical levels. Nevertheless, the term $-\sigma^2k_\perp^2$ in the ODE (Eq. (\ref{Pc})), which can not be overlooked like in the aforementioned studies since we do not have vertical stratification, makes the singularity $y=\infty$ an essential (or irregular) singularity and the Frobenius method can not be applied. This term also prevents us to apply an analysis like the WKBJ approximation because even far from critical levels the coefficients $C$ of the ODE (in Eq. (\ref{Pc})) still have a strong dependence on the latitudinal coordinate when the box is tilted.
\subsection{Inertial waves when the box is at the poles}
\label{polar_case}
When the box is located at the North or the South pole, $\widetilde{f}=0$ and the \aava{ODE} (Eq. (\ref{Pc})) is greatly simplified.
For $\theta_0 = 0$, the dimensionless \aava{wave propagation} equation becomes indeed
\begin{equation}
v''+\left[\frac{k_z^2(1-U')}{\sigma^2}+\frac{k_xU''}{\sigma}-k_\perp^2\right]v=0.
\label{Pcpole}
\end{equation}
At the South pole (i.e. $\theta_0=\pi$), the term $1-U'$ in Eq.~(\ref{Pcpole}) is replaced by $1+U'$.
Note that this equation is reminiscent of the differential equation for Rossby waves in the $\beta$-plane, i.e. $2\bo\Omega =(0,\widetilde{f},f)$ and constant $\mathrm{d}f/\mathrm{d}y=\beta$, with $f$ the Coriolis parameter \cite[e.g.][]{M1961,G1975b,GH2020}. However, we can not make a direct comparison at corotation, because the singularity in the equations for Rossby waves and inertial waves is not of the same order.
We have a second-order pole around the corotation while only first-order poles are found in the aforementioned studies. In fact, Eq. (\ref{Pcpole}) is similar to the wave equation in stratified z-sheared flows \citep[e.g.][]{J1968}.
In our polar configuration the y-coordinate is now the axial distance, and it means that the mean flow has a cylindrical profile. Such a rotation profile is expected in giant planets such as Jupiter and Saturn \citep[][respectively]{KG2017,GK2019} as a natural outcome of the Proudman-Taylor theorem for fast-rotating bodies. The propagation and dissipation of inertial modes of oscillations in the presence of critical levels for this kind of mean flow have been investigated by \cite{BR2013} in a spherical shell.
\subsubsection{Analytical solutions with constant shear}
\label{ana_pole}
Analytical solutions of the ODE Eq. (\ref{Pcpole}) are difficult to find for general profiles of the mean flow, e.g. a quadratic mean-flow profile. A linear mean-flow profile, on the other hand, has analytic solutions, that is why we use in this section such a profile, i.e. $U=R_\mathrm{o} y$, with $R_\mathrm{o}$ the shear Rossby number which is taken constant here. Eq. (\ref{Pcpole}) then becomes:
\begin{equation}
v''+\left[\frac{\alpha_k^2(1-R_\mathrm{o})/R_\mathrm{o}^2}{(y-y_0)^2}-k_\perp^2\right]v=0,
\label{pole_cstUp}
\end{equation}
where $y_0=\omega/(k_xR_\mathrm{o})$ and $\alpha_k=k_z/k_x$ the vertical to longitudinal wave number ratio. When the box is located at the South pole, the left-hand term in the bracket is $\alpha_k^2(1+R_\mathrm{o})/R_\mathrm{o}^2$ in the numerator.
This equation takes the form of Whittaker's equation \citep[see][]{AS1972} and solutions can be written in terms of the Whittaker functions $M$:
\begin{equation}
v(y)=AM_{0,\,\mu\,}(\tilde y)+BM_{0,\,-\mu\,}(\tilde y),
\end{equation}
with $\tilde y=2k_\perp(y-y_0)$,
\begin{equation}
\mu=\sqrt{\frac{1}{4}-\frac{\alpha_k^2(1-R_\mathrm{o})}{R_\mathrm{o}^2}},
\end{equation}
and $A$ and $B$ are complex constants given by boundary conditions.
The Whittaker function $M_{0,-\mu}$ allows quite straightforward analytic continuation:
\begin{equation}
M_{0,\,\mu\,}(\e^{-i\pi}\tilde y)=-i\e^{-i\pi\mu}M_{0,\,\mu\,}(\tilde y).
\end{equation}
By consequence, the solution below the critical point $y=y_0$ is:
\begin{equation}
v_\mathrm{W}(y)=-i\varsigma A\e^{- i\varsigma\pi\mu}M_{0,\,\mu\,}(-\tilde y)- i\varsigma B\e^{i\varsigma\pi\mu}M_{0,\,-\mu\,}(-\tilde y).
\end{equation}
Although the Whittaker functions do not feature precisely as wave-like forms, we can already have a good idea of the attenuation factor thanks to analytic continuation as will be shown in the following section.
It is important to point out that $\mu$ can be real or complex depending on the value of
\begin{equation}
\Od_{\theta_0=0,\,\pi}=\alpha_k^2(1\mpR_\mathrm{o})/R_\mathrm{o}^2,
\end{equation}
which we will simply denote by $R$ in the following. This can drastically change the behaviour of a wave passing through the corotation.
A necessary, but not sufficient condition to find an instability is that $R<1/4$ as we will demonstrate in Sect. \ref{inst}. This condition is similar to the Miles-Howard theorem for stratified $z$-sheared flow \citep{MH1964,L1988}. In these studies, the prerequisite for instability is that $\mathrm{Ri}<1/4$ where $\mathrm{Ri}$ is the Richardson number, i.e. the squared ratio of the Brunt-Väisälä frequency and the vertical \citep[or radial \aava{in global spherical geometry},][]{AM2013} shear. In our model, unlike cases where the box is tilted, a WKBJ analysis can be performed for a linear mean flow, mainly provided that $|R|\gg1/4$, in line with the condition of stability derived in the coming sections, and detailed in Appendix \ref{wkbj}.
These various situations regarding the value of $\Od_{\theta_0=0,\,\pi}$ at the North and South poles are illustrated in Fig. \ref{O2}. We stress the particular case where $R_\mathrm{o}=1$ ($R_\mathrm{o}=-1$) at the North (South) pole, and where the differential equation and its solutions take a quite simple form:
\begin{equation}
v''-k_\perp^2v=0,\ \text{ with }\ v\propto\e^{\pm k_\perp y}.
\label{eva}
\end{equation}
Solutions are then fully evanescent for such shears.
One can notice that Eq. (\ref{eva}) is the same far from corotation, for any mean flow.
Finally, it is clear from Fig. \ref{O2} that wave propagation is the same at the North or South pole provided a Rossby number of opposite sign.
As a result, only the equations at the North pole will be treated in the following, and the word "pole" now refers to the North pole.
\begin{figure}
\includegraphics[width=\hsize]{figures/colormap_O2_0pi.pdf}
\caption{Diagram showing how $\Od_{\theta_0=0,\,\pi} \equiv R$ compares to $1/4$ for two positions of the box ($\theta_0=0\text{ and }\pi$), and for a range of Rossby numbers $R_\mathrm{o}$ and of the ratio of the vertical and longitudinal wave numbers $\alpha_k$. The solid and dashed black lines mark where $\Od_{\theta_0=0,\,\pi}=1/4$ at the north and south pole, respectively. The purple domain show where $\Od_{\theta_0=0,\,\pi}<1/4$ and the white region where $\Od_{\theta_0=0,\,\pi}>1/4$, whether the box is at the north or the south pole. In the dark red region, $R_{\theta_0=0}>1/4$ and $R_{\theta_0=\pi}<1/4$ and vice versa in the orange region.
}
\label{O2}
\end{figure}
\subsubsection{Frobenius method at the pole}
\label{Frobpol}
Though analytic solutions are known, it is still useful to determine Frobenius solutions near corotation for two main reasons. First, these solutions can be derived for any mean flow profile near the corotation. Close to corotation, the mean flow is approximated by a Taylor expansion at the first-order $U=U'_{\aava{0}}(y-y_0)$. Secondly, Frobenius solutions may feature wave-like forms, which is helpful for physical interpretation.
Therefore, Eq. (\ref{Pcpole}) can be written near corotation:
\begin{equation}
v''+\frac{\alpha_k^2(1-R_\mathrm{o})/R_\mathrm{o}^2}{(y-y_0)^2}
v=0,
\label{Fropoleeq}
\end{equation}
where $R_\mathrm{o}=U'_{\aava{0}}$ and $R_\mathrm{o}'=U''_{\aava{0}}$.
The indicial equation gives:
\begin{equation}
\lambda=\frac{1}{2}\pm\mu.
\end{equation}
In the two next subsections, we examine both cases where $\mu$ is imaginary or real.
\subsubsection{Theoretical stable regime ($R>1/4$)}
\label{stab_pole}
We address here the case where $R>1/4$.
The same analysis as in Sect. \ref{Fro} can be carried out to determine how a wave behaves upon crossing the corotation.
The solutions of the indicial equation can be recast as
\begin{equation}
\lambda=\frac{1}{2}\pm i|\mu|.
\end{equation}
The first-order solutions to Eq. (\ref{Fropoleeq}) in the vicinity of $y_0$ are :
\begin{equation}
\left\{\begin{aligned}
&a_0(y-y_0)^{\frac{1}{2}+i|\mu|}+b_0(y-y_0)^{\frac{1}{2}-i|\mu|},&y>y_0\\
&-i\varsigma\left[a_0(y_0-y)^{\frac{1}{2}+i|\mu|}e^{\varsigma\pi|\mu|}+b_0(y_0-y)^{\frac{1}{2}-i|\mu|}e^{-\varsigma\pi|\mu|}\right],&y<y_0\end{aligned}\right.
\label{solpol}
\end{equation}
for $a_0,\, b_0 \in \mathbb{C}$. One can recover the same form of Frobenius solutions as in \cite{AM2013}, who examined radially stratified mean flows in spherical geometry.
As $\widetilde{f}=0$, the wave action flux Eq. (\ref{wA}) reduces to
\begin{equation}
\mathcal{A}=-\frac{\rho_0}{2k_\perp^2}\Im{v_yv^*},
\label{wApol}
\end{equation}
that is, injecting the solutions on both sides of the critical level:
\begin{equation}
\mathcal{A}=-\frac{\rho_0}{2k_\perp^2}|\mu|\left\{
\begin{aligned}
&|a_0|^2-|b_0|^2,&y>y_0\\
&-|a_0|^2\e^{2\varsigma\pi|\mu|}+|b_0|^2\e^{-2\varsigma\pi|\mu|},&y<y_0
\end{aligned}\right..
\label{Apol_st}
\end{equation}
This formulation is quite similar to the expression of the Reynolds stress ($\tau$) in vertically stratified mean flows, which can be found in \cite{BB1967} in Cartesian geometry. We recall indeed that $\tau=-k_x\mathcal{A}$. Moreover, given Eq. (\ref{Re}), the Reynolds stress in our model reads $\tau=-\rho_0 \overline{uv}$. Using the polarisation relations for $u$, we recover the wave action flux in Eq. (\ref{wApol}).
\begin{figure}[t!]
\includegraphics[width=\hsize]{figures/colormap_atnu_deg0}
\caption{Transmission rate $T_\mathrm{F}$ of the wave action flux across the corotation when the box is located at the pole for $R>1/4$. It is displayed against the Rossby number $R_\mathrm{o}$ and the ratio of the horizontal wave numbers $\alpha_k=k_z/k_x$. The forbidden region where $R<1/4$ is painted in white, and the dark red cone corresponds to values of $T_\mathrm{F}$ lower than $10^{-9}$.}
\label{atnuO2}
\end{figure}
The pre-factor $i$ in the solutions (\ref{solpol}) below the corotation does not affect the energy flow and simply indicates that the wave undergoes a phase shift of $\pi/2$ through the critical level \citep[see also][]{AM2013}.
Above the critical level, the normalised Doppler-shifted frequency satisfies $\sign(\sigma)=-\varsigma$ as for the corotation in the inclined case. The sign is reversed below the critical level. Thus, the first solution of main amplitude $|a_0|$ carries its latitudinal flux of energy upward (downward) for $\varsigma=+1$ ($\varsigma=-1$), while the second solution transfers its energy in the opposite direction in the various cases.
Therefore, the energy flux of an upward or downward wave is always attenuated by a factor
\begin{equation}
T_\mathrm{F}=\exp\{-2\pi|\mu|\}.
\end{equation}
This attenuation factor is shown in Fig. \ref{atnuO2} versus $R_\mathrm{o}$ and $\alpha_k$. We observe that the wave is largely absorbed at the critical level and thus deposits most, if not all its energy for most couples $(\alpha_k,R_\mathrm{o})$.
\subsubsection{Possible unstable regime ($R<1/4$)}
\label{inst}
We now deal with the case where $R<1/4$, i.e. $\mu$ is real.
Contrary to the situation where $R>1/4$, we can no longer assimilate solutions to wave-like functions. The exponential form of solutions for $R<1/4$ near the critical level reads
\begin{equation}
v(y)=\left\{\begin{aligned}
&a_0(y-y_0)^{\frac{1}{2}+|\mu|}+b_0(y-y_0)^{\frac{1}{2}-|\mu|}, &y>y_0\\
&-i\varsigma\left[\e^{-i\varsigma\pi|\mu|}a_0(y-y_0)^{\frac{1}{2}+|\mu|}\right.\\
&\left.~~~~~~~+\e^{i\varsigma\pi|\mu|}b_0(y-y_0)^{\frac{1}{2}-|\mu|}\right], &y<y_0
\end{aligned}\right.
\label{solpol_inst}
\end{equation}
and makes this region fully evanescent. Furthermore, the associated wave action flux is
\begin{equation}
\mathcal{A}=-\frac{\rho_0}{k_\perp^2}\mu\left\{
\begin{aligned}
&\Im{a_0b_0^*},&y>y_0\\
&-\Im{a_0b_0^*\e^{-2i\varsigma\pi\mu}},&y<y_0
\end{aligned}\right..
\label{Apol_inst}
\end{equation}
\begin{table*}[ht!]
\centering
\begin{tabular}{c|c|| c|c}
\rule[-1ex]{0pt}{5ex} Box & critical level & attenuation & amplification \\\hline\hline
\rule[-1ex]{0pt}{4ex} & $\sigma=\widetilde{f}$
& yes for waves $\downarrow$ & no \\
\rule[-2ex]{0pt}{6ex} \makecell{Inclined box\\$\left(\theta_0\in]0,\pi/2[\right)$} & $ \sigma=-\widetilde{f}$ & yes for waves $\uparrow$ & no \\
\rule[-3ex]{0pt}{6ex} & $\sigma=0$ & no & no \\ \hline
\rule[-1ex]{0pt}{7ex} \makecell{Pole \\ $(\theta_0=0)$}& $\sigma=0$ &
\begin{tabular}{l}
yes if $R<1/4$\\
yes if $\left\{\begin{aligned}R>1/4 \\\varsigma\mathcal{A}^+>0 \end{aligned}\right.$
\end{tabular}
& yes if $\left\{\begin{aligned}R>1/4 \\\varsigma\mathcal{A}^+<0 \end{aligned}\right.$ \\\hline
\rule[-1ex]{0pt}{4ex} \multirow{3}{*}{\makecell{Equator\\($\theta_0=\pi/2$)}} & \aava{$\sigma=\pm1$} & \aava{\Large$\smallsetminus$} & \aava{\Large$\smallsetminus$} \\
\rule[-2ex]{0pt}{6ex} & $\sigma=0$ & no & no \\ \hline
\end{tabular}
\vspace{0.5cm}
\caption{Summary table of analytical results at each critical level when the box is inclined or not with respect to the rotation axis, in the $[0,\pi/2]$ quadrant and for positive wavenumbers. It indicates whether each critical level can cause attenuation or amplification of the (upward $\uparrow$ and downward $\downarrow$) travelling wave \aava{ in the y-direction, which depends notably, at the pole, on $R=\alpha_k(1-R_\mathrm{o})/R_\mathrm{o}^2$, and on} the wave action flux above the critical level $\mathcal{A}^+$. \aava{The symbol $\smallsetminus$ means that no wave action flux is carried across the critical level.}
Moreover, the results are analogous in the other quadrants of the \aava{spherical body (with the direction of the attenuated wave through $\sigma=\pm\widetilde{f}$ varying according to $\sign(k_zf)$).}}
\label{recap_ana}
\end{table*}
Without knowing the direction of the wave or the energy flux since $\sign(V_\mathrm{g})=\varsigma\sign\left(\Im{a_0b_0^*}\right)$, it is difficult to assess the impact of the critical level on wave propagation, i.e. whether it will attenuate waves or on the contrary amplify them.
\cite{LB1985} found a way to investigate the behaviour of internal gravity waves \aava{in the presence of a vertical shear,} passing through a critical level in a regime similar to ours ($\mathrm{Ri}<1/4$) where solutions are fully evanescent.
Their work\aava{, which was carried out in local Cartesian geometry,} has been taken up by \cite{AM2013} \aava{in global} spherical geometry \aava{applied to the radiative zone of solar-like stars and evolved stars}. The method is to determine the reflection and transmission coefficients in a three-zone model. The evanescent region where the Richardson number \aava{satisfies} $\mathrm{Ri}<1/4$ and where the critical level is located (zone II), is sandwiched between two propagating wave layers (zones I and III). Using a linear mean flow profile so as to establish solutions inside zone II, \cite{LB1985} and \cite{AM2013} both used continuity relations of the perturbed vertical (or radial) velocity and its derivatives at the interfaces between zones in order to get the transmission and reflection coefficients. The critical level is located in the middle of zone II of width $\delta$. By consequence, the reflection and transmission coefficients depend, in their works, on the shear and more precisely the Richardson number, and on the width $\delta$.
They both found that, depending on the Richardson number and $\delta$, the reflection and transmission coefficients can be greater than one, meaning that the wave can be over-reflected and/or over-transmitted\aava{, and thus extract energy and angular momentum fluxes from the mean flow, which can lead to potential shear instabilities after successive encounters of the wave with the critical layer}. However, this result is conditioned by the geometry of the model. As shown by \cite{L1988} in his review and references therein, models with one or even two layers with evanescent and eventually a wave-like region, do not allow such phenomena. A first region that allows the wave propagation is mandatory and is combined with a ``sink'' that pulls the wave to cross the critical level. According to \cite{LB1985} and \cite{L1988}, the nature of the sink for wave flux can be either another propagative region or an evanescent region, as in zone II, subject to friction processes. Given this peculiar geometry, instabilities can occur under boundary conditions that allow the wave to return successively to the critical level. Many studies have tried to relate over-reflection and shear instability for a specific wave geometry \cite[see in particular the reviews of][for internal gravity waves and Rossby waves]{L1988,HH2007}.
In the present study, we do not investigate further shear instability by doing, for instance, a temporal analysis to estimate the waves' growth rate \citep[as in][who considered an initial value problem]{LB1985,WA2004}. On the contrary, we give arguments, such as $R<1/4$, of necessary but not sufficient condition to find instabilities. It is important to note that $R$ is constant in the whole domain for a linear mean flow profile, and thus one is stuck with either a propagative (stable) or an evanescent regime. Therefore, finding an adequate geometry to allow over-reflection and over-transmission requires at least that the Rossby number is not the same in the whole domain, by using for instance a non-linear mean flow profile.
Furthermore, in the particular case where $R=0$, (i.e. $R_\mathrm{o}=1$ or $\alpha_{k}=0$ in Eq. (\ref{Pcpole}) when the wave with $k_{z}=0$ propagates in the $(x,y)$-plane), a necessary condition for instability is given by the Inflection Point Theorem \citep{SH2002}. This theorem is particularly used to study barotropic instabilities for Rossby waves \citep[see e.g.][]{LT1978}. In other words, a necessary condition to have unstable modes for $R_\mathrm{o}=1$ is that $U''$ cancels out in the domain of wave propagation. \\
We summarise in Table \ref{recap_ana} the main analytical results of Sects. \ref{Fro} and \ref{polar_case} about wave and wave action flux transmission, either when the box is \aava{tilted} relative to the rotation axis, at the North pole\aava{, or at the equator, in the inviscid limit}. \aava{Note that when ``no'' is given in both attenuation and amplification columns, the wave is fully transmitted across the critical level, regardless of the wavenumbers and of the mean flow profile.}
\section{A three-zone numerical model}
\label{secnum}
In order to test the analytical predictions of the previous section, we have built up a three-zone numerical model to simulate waves passing through critical levels. A similar model has been used, for instance, by \citet{J1967} to explore the behaviour of internal gravity waves passing through critical levels in a fluid with rotation and vertical shear. In our model, we solve the two first-order ODEs satisfied by $v$ and $\Pi$, the combination of which led to the \aava{wave propagation} equation~(\ref{Pc}).
By imposing boundary conditions such that waves satisfy the dispersion relations (see also Appendix \ref{solI_III}), we examine the dynamics of inertial waves propagating in the shear region. Also, whenever possible, we calculate analytically the wave transmission and reflection coefficients as the wave-like solution crosses the shear region.
\begin{figure}[t!]
\centering
\includegraphics[width=0.4\textwidth]{figures/mean_flows}
\caption{Mean flow profiles used in the three-zone numerical model against $y$. In the no-shear regions I ($y<0$) and III ($y > 1$), the mean flow is uniform and set respectively to $U=0$ and $U=\Lambda$. The sheared region II (\textit{gray-shaded}) can have a linear, square, or cubic mean flow profile.}
\label{mean_flow}
\end{figure}
\subsection{Description of the model}
\label{numMod}
The mean flow profile that is used in the three-zone model is illustrated in Fig. \ref{mean_flow}.
The zone with shear (zone II) is surrounded by two no-shear regions, one with no mean flow (zone I), and one with a uniform mean flow (zone III). In the whole domain, the mean flow profile that we adopt is expressed as
\begin{equation}
U(y)=\left\{\begin{aligned}
&0~~~&\mathrm{for}&~y<0: & \textit{zone I},&\\
&\Lambda y^n~~~&\mathrm{for}&~0\leq y\leq1:&\textit{zone II},&\\
&\Lambda~~~&\mathrm{for}&~y>1:& \textit{zone III}.&\\
\end{aligned}\right.
\label{U_pro}
\end{equation}
where $U(y)$ is continuous at each interface, and $n$ is an integer: $n=1$ for a linear shear flow, $n=2$ for a square shear flow or $n=3$ for a cubic shear flow (see also Fig.~\ref{mean_flow}). In zone I, we assume that there is an incident wave that enters the shear zone as well as a wave that is reflected at the interface between zones I and II or in zone II, i.e.:
\begin{equation}
v(y)=A_{I}\exp(ik_{I}y)+A_{R}\exp(ik_{R}y),
\label{IR_waves}
\end{equation}
where $A_{I}$ and $A_{R}$ are the amplitudes, $k_{I}$ and $k_{R}$ are the wavenumbers of the incident and reflected waves, respectively. We further impose as boundary condition in zone III a transmitted wave that propagates towards positive $y$-values:
\begin{equation}
v(y)=A_{T}\exp(ik_{T}y),
\label{T_wave}
\end{equation}
where $A_{T}$ and $k_{T}$ are the amplitude and wavenumber of the transmitted wave, respectively.
We impose $A_{T}=1$ without loss of generality and compute the remaining amplitudes $A_{I}$ and $A_{R}$.
More details on the solutions and the dispersion relations of the waves in zones I and III can be found in Appendix \ref{solI_III}. We have ensured that the transmitted wave carries energy upwards, by deriving the wave action flux in zones I and III (see Appendix \ref{wAI_III}).
We impose the continuity of the latitudinal velocity and reduced pressure at the interfaces (at $y=0$ and $y=1$). By doing so, the wave action flux is continuous at both interfaces. Thus, in the absence of critical points, the wave action flux is conserved in the whole domain, namely $\mathcal{A}_\mathrm{T}=\mathcal{A}_\mathrm{I-R}$ where $\mathcal{A}_{\mathrm{T}}$ is the wave action flux of the transmitted wave and $\mathcal{A}_{\mathrm{I-R}}=\mathcal{A}_{\mathrm{I}}-\mathcal{A}_{\mathrm{R}}$ with $\mathcal{A}_{\mathrm{I}}$ and $\mathcal{A}_{\mathrm{R}}$ for the incident and reflected wave action fluxes, respectively.
To solve the ODE in the three zones and in particular near singularities, we have used MATLAB's solver ode15s, which is suitable for solving stiff differential equations \citep[][]{SR1997}.
To avoid strict singularities at $\sigma^{2}=\tilde{f}^{2}$, we have added a small friction $\sigma_f=\tt{-8}$ in our set of units. Given the boundary conditions, the numerical solver deals with two first-order ODEs for $v$ and $\Pi$, which take the form
\begin{equation}
\left\{
\begin{aligned}
F_{1,v}v+F_{1,p}\Pi=Av',\\
F_{2,v}v+F_{2,p}\Pi=A\Pi',
\end{aligned}\right.
\label{ODEcode}
\end{equation}
where
\begin{equation}
\begin{aligned}
F_{1,v}&=\widetilde{f}\frac{U'}{\sigma}\left(k_x\widetilde{f}+ik_zs\right)-(U'-f)\left(k_xs+ik_z\widetilde{f}\right),\\
F_{1,p}&=-ik_\perp^2s,\\
F_{2,v}&=is\left[s^2+f(U'-f)\right]-i\widetilde{f}^2\left(s+f\frac{U'}{\sigma}\right),\\
F_{2,p}&=ik_zf\widetilde{f}-k_xfs,\\
\end{aligned}
\end{equation}
and where we recall that $s=\sigma+i\sigma_{f}$ is the modified Doppler-shifted frequency due to Rayleigh friction. While $A_{T}=1$ is imposed by the boundary condition, we compute $A_{I}$ and $A_{R}$ by comparing numerical solutions of the system Eq. (\ref{ODEcode}) at $y=0$ with the definition of velocity in zone I (Eq. (\ref{IR_waves})) and its associated reduced pressure (see Eq. (\ref{Pivis})).
\subsection{Numerical exploration at the pole for a constant shear}
\subsubsection{Reflection and transmission coefficients}
\begin{figure*}[ht]
\resizebox{\hsize/2}{!}{\includegraphics[trim=0cm 0cm 0cm 1.4cm,clip]{figures/Ro_omega_T_0p1_0p1}}
\resizebox{\hsize/2}{!}{\includegraphics[trim=0cm 0cm 0cm 1.4cm,clip]{figures/Ro_omega_R_0p1_0p1}}
\caption{Transmission coefficient ($|T|$, left panel) and reflection coefficient ($|R|$, right panel) when the box is at the pole. They are plotted in the regime $R<1/4$ (possibly unstable case, see Sect. \ref{inst}) as a function of the Rossby number $R_\mathrm{o}$ and the inertial frequency $\omega$. The hatched areas do not feature critical points, which correspond to regions where $\omega>k_xR_\mathrm{o}$ in our peculiar geometry (see Appendix \ref{wAI_III} for this particular matter). Vertical and longitudinal wave numbers are fixed: $k_x=0.1$ and $k_z=0.1$. Moreover, the contours that correspond to the coefficients $|R|$ and $|T|$ equal to one are indicated by solid black lines.
Crosses mark the set of parameters used in Fig.~\ref{vA_pole} for the analysis of the behaviour of the velocity in the three-layer model.}
\label{RT_inst}
\end{figure*}
\begin{figure*}[ht]
\resizebox{\hsize/2}{!}{\includegraphics[trim=0cm 0cm 0cm 1.4cm,clip]{figures/Ro_omega_T_0p1_0p1_st}}
\resizebox{\hsize/2}{!}{\includegraphics[trim=0cm 0cm 0cm 1.4cm,clip]{figures/Ro_omega_R_0p1_0p1_st}}
\caption{Same as Fig. \ref{RT_inst}, but for $R>1/4$. When $R_\mathrm{o}<0$ in these panels, hatched areas are also displayed for all $\omega>0$ (see Appendix \ref{wAI_III}).}
\label{RT_st}
\end{figure*}
In most cases, for any inclination of the box and any mean flow profile, there is no analytical solutions in zone II.
Nevertheless, we have shown in Sect. \ref{ana_pole} that, when the box is at the pole and for a linear mean shear flow, solutions can be found in terms of Whittaker functions.
In this section, we will find reflection and transmission coefficients similarly as in \cite{LB1985} and \cite{AM2013}, though there are a few differences. \aava{In particular, our present study differs from the latter by the treatment of inertial waves in convective regions (instead of gravity waves in stably stratified radiative regions in their case) with a latitudinal shear (instead of a vertical/radial shear). Our study, however, uses a local Cartesian model as in \cite{LB1985}}.
Moreover, our boundary conditions are different, as detailed in Sect. \ref{inst}, and the thickness of our shear region is fixed to one in scaled units while \cite{LB1985} and \cite{AM2013} leave the thickness $\delta$ as a control parameter.
We also check the existence of a critical level in the shear zone and the frequency range that delineates the regimes with and without the critical level.
We consider that the perturbed reduced pressure $\Pi$ and velocity $v$ are continuous at the interfaces $y=0$ and $y=1$. In the presence of the critical level $y_0$ in zone II, we have a set of four analytical solutions whose values at $y=0$ and $y=1$ allow us to determine the reflection and transmission coefficients. The solutions to \aava{the wave propagation} equation in zones I, II (below and after the critical level) and III are:
\begin{equation}
v=\left\{\begin{aligned}
&A_\mathrm{I}\e^{ik_\mathrm{I}y}+A_\mathrm{R}\e^{ik_\mathrm{R}y},\ &\text{for }y<0\\
&-i\left[A\e^{-i\pi\mu}M_{0,\,\mu\,}(-\tilde y)+B\e^{i\pi\mu}M_{0,\,-\mu\,}(-\tilde y)\right],\ &\text{for }0<y<y_0\\
&AM_{0,\,\mu\,}(\tilde y)+BM_{0,\,-\mu\,}(\tilde y),\ &\text{for }y_0<y<1\\
&A_\mathrm{T}\e^{ik_\mathrm{T}y},\ &\text{ for }y>1
\end{aligned}\right.
\end{equation}
where $A$ and $B$ are complex coefficients which we will express below. We remind $A_T=1$ in our numerical model. In the shear region (zone II), the reduced pressure perturbation is given by Eqs. (\ref{pol}), which, at the pole, can be recast as
\begin{equation}
\Pi=\frac{i}{k_\perp^2}\left[\sigma v_y+k_x(R_\mathrm{o}-1)v \right]
\label{Pipole},
\end{equation}
with $R_\mathrm{o}=\Lambda$.
In the regions with no shear (zones I and III), $\Pi$ takes a simpler expression with $R_\mathrm{o}=0$, and $\sigma=\omega$ in zone I and $\sigma=\omega-k_xU(1)=\omega-k_xR_\mathrm{o}$ in zone III (noted $\sigma_3$ in the following). Note that while the reduced pressure is kept continuous to conserve the wave action flux across the interfaces, the first derivative of the latitudinal velocity $v'$ is not necessarily continuous at the interfaces.
To find the transmission and reflection coefficients, we solve the system of equations that consist of matching conditions at interfaces as follows:
\begin{my_enumerate}
\item $v~\mathrm{is~continuous~at}~y=0:$\\
$A_\mathrm{I}+A_\mathrm{R}=v_\mathrm{W}(0)$;
\item $\Pi~\mathrm{is~continuous~at}~y=0:$\\
$\left(ik_\mathrm{I}\omega-k_x\right)A_\mathrm{I}+\left(ik_\mathrm{R}\omega-k_x\right)A_\mathrm{R}=\omega v'_\mathrm{W}(0)+k_x(R_\mathrm{o}-1)v_\mathrm{W}(0),$
\item $v~\mathrm{is~continuous~at}~y=1:$\\
$\e^{ik_\mathrm{T}}=v_\mathrm{W}(1),$
\item $\Pi~\mathrm{is~continuous~at}~y=1:$\\
$\left(ik_\mathrm{T}\sigma_3-k_x\right)\e^{ik_\mathrm{T}}=\sigma_3$ $v'_\mathrm{W}(1)+k_x(R_\mathrm{o}-1)v_\mathrm{W}(1),$
\end{my_enumerate}
with the Whittaker functions $v_\mathrm{W}$. At the interfaces below and above the critical level $y_0$ (i.e. $y=0$ and $y=1$), we have
\begin{equation}
\begin{aligned}
v_\mathrm{W}(0)&=-i\left[A\e^{-i\pi\mu}M_{0,\,\mu\,}(2k_\perpy_0)+B\e^{i\pi\mu}M_{0,\,-\mu\,}(2k_\perpy_0)\right],\\
v_\mathrm{W}(1)&=AM_{0,\,\mu\,}\left(2k_\perp(1-y_0)\right)+BM_{0,\,-\mu\,}\left(2k_\perp(1-y_0)\right).
\end{aligned}
\label{whit_01}
\end{equation}
Please note that the first derivative of the Whittaker functions can be computed either numerically or analytically via the relationships in \cite{AS1972}.
The equations in the above continuity relationships 1. to 4. are independent two by two (1. and 2., 3. and 4.), and $A$ and $B$ can be found first:
\begin{equation}
\begin{aligned}
A = &\e^{ik_\mathrm{T}}\frac{\sigma_3(M'_{-\mu,1}-ik_\mathrm{T}M_{-\mu,1})+M_{-\mu,1}k_xR_\mathrm{o}}{\sigma_3(M_{\mu,1}M'_{-\mu,1}-M'_{\mu,1}M_{-\mu,1})},\\
B = &\e^{ik_\mathrm{T}}\frac{\sigma_3(M'_{\mu,1}-ik_\mathrm{T}M_{\mu,1})+M_{\mu,1}k_xR_\mathrm{o}}{\sigma_3(M'_{\mu,1}M_{\mu,1}-M_{\mu,1}M'_{-\mu,1})}.
\end{aligned}
\end{equation}
The amplitude of the incident and reflected waves can be written, in terms of $A$ and $B$, as:
\begin{equation}
\begin{aligned}
A_\mathrm{I}=& \frac{A\e^{-i\mu\pi}X_\mu(k_\mathrm{R})+B\e^{i\mu\pi}X_{-\mu}(k_\mathrm{R})}{\omega(k_\mathrm{R}-k_\mathrm{I})},\\
A_\mathrm{R}=& \frac{A\e^{-i\mu\pi}X_\mu(k_\mathrm{I})+B\e^{i\mu\pi}X_{-\mu}(k_\mathrm{I})}{\omega(k_\mathrm{I}-k_\mathrm{R})},
\end{aligned}
\label{AIR}
\end{equation}
where
\begin{equation}
\begin{aligned}
X_{\mu}(k)=\omega(M'_{\mu,0}-ikM_{\mu,0})+k_xR_\mathrm{o} M_{\mu,0}.\\
\end{aligned}
\end{equation}
The transmission and reflection coefficients are then:
\begin{equation}
|T|=\frac{|A_{\mathrm{T}}|}{|A_{\mathrm{I}}|}=\frac{1}{|A_\mathrm{I}|}\ \text{ and }\ |R|=\frac{|A_\mathrm{R}|}{|A_\mathrm{I}|}.
\end{equation}
We emphasise that these factors depend notably on the location of the critical level and on the inertial frequency, which was not the case in \cite{LB1985} and \cite{AM2013}.
\begin{figure*}[ht!]
\resizebox{\hsize/2}{!}{\includegraphics[trim=0cm 0cm 0cm 1.4cm,clip]{figures/Ro_omega_jump_0p1_0p1_st}}
\resizebox{\hsize/2}{!}{\includegraphics[trim=0cm 0cm 0cm 1.4cm,clip]{figures/Ro_omega_jump_0p1_0p1}}
\caption{Ratio of the wave action flux above and below the critical level against the Rossby number and the inertial frequency, for $R>1/4$ (\textit{left panel}) and $R<1/4$ (\textit{right panel}), when the box is at the pole. Like in Fig. \ref{RT_st}, $k_x=k_z=0.1$, and hatched zones represent areas without critical levels. Again, crosses indicate the set of parameters chosen to analyse the behaviour of the velocity in the three-layer model.}
\label{wA_IRT}
\end{figure*}
We display in Figs. \ref{RT_inst} and \ref{RT_st} the transmission and reflection coefficients as a function of the Rossby number $R_\mathrm{o}$ and the normalised inertial frequency $\omega$ in the regimes $R<1/4$ and $R>1/4$, respectively. We choose $k_{x}=k_{z}=0.1$ in these plots and $R=1/4$ gives $R_\mathrm{o}=-2\pm\sqrt{8}\approx \{-4.8,0.8\}$, which delineates the two regimes as we can see in Fig. \ref{O2}.
Areas that are hatched do not possess the corotation point $\sigma=0$.
In addition, the wavenumber $k_\mathrm{T}$ was chosen with a positive sign in these regions to maintain an upward propagating wave (see Appendices \ref{solI_III} and \ref{wAI_III}).
Areas that are not hatched feature a critical level, according to the table of Appendix \ref{wAI_III}.
In the case where $R<1/4$, over-reflection and over-transmission are both possible (see Fig.~\ref{RT_inst}).
One should notice that for $R_\mathrm{o}>1$, we always have by definition of $R$
\begin{equation}
\frac{R}{(y-y_0)^2}-k_\perp^2<0,
\end{equation}
regardless of $y$, which makes the solutions of Eq. (\ref{pole_cstUp}) tend towards pure exponential functions, i.e. without any imaginary part. Also, we do not see any over-reflection, nor over-transmission in the hatched areas where there is no corotation point.
This highlights the essential role of the critical level in inducing over-reflection or over-transmission of inertial waves crossing the shear region in this regime.
The regime where $R>1/4$ (in Fig. \ref{RT_st}) is more delicate to analyse.
According to our discussion in Sect. \ref{stab_pole}, we expect a strong attenuation of the wave and of the wave action flux as shown in Fig. \ref{atnuO2}.
From this figure and for $\alpha_k=0.1$, the damping is very strong for low positive Rossby numbers.
This tendency is also found for both transmission and reflection coefficients.
Nevertheless, one can also observe an unexpected regime
of over-transmission near $R_\mathrm{o}=0.8$ and low frequency $\omega$.
Still, we must not forget that solutions in this regime, even near the critical level (see Eq.~(\ref{solpol})), are not rigorously equivalent to wave-like functions.
In particular, the amplification term $(y-y_0)^{1/2}$ that can be found at the first-order in the Frobenius solutions becomes more prominent as the thickness of the shear zone is larger.
This is especially true for the transmission coefficient. Assuming that Eq.~(\ref{solpol}) holds throughout zone II and corresponds to upward and downward waves, the transmission coefficient is modulated by $|1-y_0|^{1/2}/|0-y_0|^{1/2}$, the amplitude ratio between the transmitted and incident waves.
This term can be greater than one in the shear region.
In particular, it is always greater than one when $y_0<0$, i.e. no critical level in the regime $R>1/4$ (hatched areas in Fig. \ref{RT_st}).
On the contrary, this ratio is not present for the reflection coefficient since $|R|$ is function of the incident and reflected waves evaluated at $y=0$.
Though this hand-waving explanation does not formally demonstrate the origin of this amplification, it stresses the important role of the shear-region thickness and more generally of the geometry of the model.
In order to clarify whether the amplification is due to the geometry or the critical level, we need to investigate how the wave action flux changes before and after the critical level. The wave action flux is indeed the relevant quantity to investigate energy flux exchanges at a critical level.
\subsubsection{Wave action fluxes below and above the shear region}
Since $v$ and $\Pi$ are continuous at the interfaces between the shear and no-shear regions, the wave action flux is preserved and continuous in all three zones in the absence of \aava{friction} and critical levels. However, it is discontinuous at the corotation point as demonstrated in Sects. \ref{stab_pole} and \ref{inst}. Given the amplitude of the incident and reflected waves (Eq. (\ref{AIR})\aava{)}, we can calculate the ratio of the wave action flux below and after the corotation (see Appendix D.2 for the detailed calculation):
\begin{equation}
\frac{\mathcal{A}_\mathrm{T}}{\mathcal{A}_\mathrm{I-R}}=\pm\frac{\omega^2\sqrt{\sigma_3^2\left[k_z^2f^2+k_\perp^2\left(\widetilde{f}^2-\sigma_3^{2}\right)\right]}}{\sigma_3^2\sqrt{\omega^2\left[k_z^2f^2+k_\perp^2\left(\widetilde{f}^2-\omega^{2}\right)\right]}\left(\left|A_\mathrm{I}\right|^{2}-\left|A_\mathrm{R}\right|^{2}\right)}.
\label{rwA}
\end{equation}
The signs $+$ or $-$ can be chosen in regards to the wave action flux of the transmitted wave that can be positive or negative depending on the presence of the critical level, while the energy flux is always positive in order to have an upward propagating wave in zone III (see Appendix \ref{wAI_III} for a more detailed discussion).
This wave action flux ratio is displayed in Fig. \ref{wA_IRT} in the two regimes $R\lessgtr1/4$. As expected, this ratio is equal to one when no critical level is present (hatched areas).
Unlike in the previous section, the regime where $R>1/4$ has no longer amplification areas, $|\mathcal{A}_\mathrm{T}/\mathcal{A}_\mathrm{I-R}|<1$ everywhere. This supports the idea that the critical level has nothing to do with the amplification phenomenon observed in the left panel of Fig. \ref{RT_st}. As already observed in Fig. \ref{atnuO2}, the damping due to the critical level is strong except for Rossby numbers close to the threshold between the two regimes.
Moreover, $\mathcal{A}_\mathrm{T}/\mathcal{A}_\mathrm{I-R}<0$ means that $\left|A_\mathrm{I}\right|^{2}>\left|A_\mathrm{R}\right|^{2}$ since the minus sign is taken in Eq. (\ref{rwA}). Therefore, no over-reflection due to the critical level is expected in this regime.
The other regime ($R<1/4$, right panel) features areas where the wave is over-reflected for $\left|A_\mathrm{I}\right|^{2}<\left|A_\mathrm{R}\right|^{2}$ (i.e. when $\mathcal{A}_\mathrm{T}/\mathcal{A}_\mathrm{I-R}>0$) and areas where the wave is over-transmitted for $|\mathcal{A}_\mathrm{T}/\mathcal{A}_\mathrm{I-R}|>1$. For the first inequality ($\left|A_\mathrm{I}\right|^{2}<\left|A_\mathrm{R}\right|^{2}$), the threshold between under and over reflection (around $R_\mathrm{o}\approx0.9$) is the same than for the reflection factor (in the right panel of Fig. \ref{RT_inst}). For the second one ($|\mathcal{A}_\mathrm{T}/\mathcal{A}_\mathrm{I-R}|>1$), the comparison with the transmission factor (in the left panel of Fig. \ref{RT_inst}) is more questionable.
Still, these two points suggest that the critical level can induce the over-reflection and over-transmission phenomena in the regime where $R<1/4$.
\subsubsection{Numerical solutions}
\begin{figure*}[ht]
\resizebox{\hsize/3}{!}{\includegraphics[trim=0cm 0cm 0cm 1cm,clip]{figures/v_y_cs_n1_Ro0p3}}
\resizebox{\hsize/3}{!}{\includegraphics[trim=0cm 0cm 0cm 1cm,clip]{figures/v_y_cs_n1_Ro0p8}}
\resizebox{\hsize/3}{!}{\includegraphics[trim=0cm 0cm 0cm 1cm,clip]{figures/v_y_ci_n1_Ro1p8}}
\resizebox{\hsize/3}{!}{\includegraphics[trim=0cm 0cm 0cm 1cm,clip]{figures/a_deg0_y__n1_rf1e-08_Ro0p3}}
\resizebox{\hsize/3}{!}{\includegraphics[trim=0cm 0cm 0cm 1cm,clip]{figures/a_deg0_y__n1_rf1e-08_Ro0p8.pdf}}
\resizebox{\hsize/3}{!}{\includegraphics[trim=0cm 0cm 0cm 1cm,clip]{figures/a_deg0_y__n1_rf1e-08_Ro1p8.pdf}}
\caption{\aava{\textit{Top:}} Real part of the latitudinal velocity $v$ against $y$.
The quantities $v_\mathrm{num}$, $v_\mathrm{I}$, $v_\mathrm{R}$, $v_\mathrm{T}$, $v_\mathrm{F}$ and $v_\mathrm{W}$ are the numerical, incident, reflected, transmitted, first-order Frobenius and Whittaker velocities, respectively.
\aava{\textit{Bottom:}} Wave action flux against $y$. The quantities $\mathcal{A}_\mathrm{num}$, $\mathcal{A}_\mathrm{T}$, $\mathcal{A}_\mathrm{I-R}$, $\mathcal{A}_\mathrm{F}$ are the numerical, transmitted, incident and reflected, Frobenius wave action fluxes, respectively.
For all panels, $\theta_0=0$, the mean flow is linear in the grey-shaded shear regions, and the critical level is marked by dashed lines. The horizontal wave numbers are set to $k_x=k_z=0.1$.
\textit{From left to right:} (i) $\omega=0.02$ and $\aava{R_\mathrm{o}}=0.3$ ($R>1/4$), (ii) $\omega=0.002$ and $\aava{R_\mathrm{o}}=0.8$ ($R>1/4$), and (iii) $\omega=0.09$ and $\aava{R_\mathrm{o}}=1.8$ ($R<1/4$).
}
\label{vA_pole}
\end{figure*}
In the previous sections, we have examined how the shear parameter and the inertial wave frequency impact the reflection and transmission coefficients as well as the wave action flux. We now study particular cases of wave propagation through the critical level for fixed sets of parameters in both regimes $R\lessgtr1/4$.
To do that, we have numerically solved Eqs. (\ref{ODEcode}) for the three-layer model described in Sect. \ref{numMod}, for $\theta_0=0$ and a linear shear flow ($n=1$ in zone II). We have selected three pairs of values for the inertial frequency and the shear, two in the regime $R>1/4$ and one in the regime $R<1/4$. These values are marked by crosses in Figs. \ref{RT_inst}, \ref{RT_st} and \ref{wA_IRT}. In each case, the latitudinal velocity and the wave action flux have been calculated through the three zones successively and plotted in Fig. \ref{vA_pole}. The numerical solution, which has been computed by imposing the boundary condition $A_{\mathrm{T}}=1$ and the continuous interfacial conditions for $v$ and $\Pi$ at $y=\{0,1\}$, is the sum of incident and reflected waves in zone I, and equal to a transmitted wave in zone III. The expressions for the incident, reflected and transmitted waves are given by Eqs. (\ref{IR_waves}) and (\ref{T_wave}) (see also Appendix \ref{solI_III}). In the shear region (zone II) of the upper panel of Fig. \ref{vA_pole}, the Whittaker solution has been added and it matches perfectly with the numerical solution below and above the critical level in each case.
Moreover, Frobenius approximations for the latitudinal velocity (Eqs. (\ref{solpol}) when $R>1/4$, and (\ref{solpol_inst}) when $R<1/4$) and for the wave action flux (Eqs. (\ref{Apol_st}) when $R>1/4$, (\ref{Apol_inst}) otherwise) have also been included. The coefficients $a_0$ and $b_0$ have been determined by matching the numerical solution and its derivative to the Frobenius approximation for the velocity close to the critical level. For both the latitudinal velocity and the wave action flux, this first-order approximation gives satisfactory agreement with numerical solutions, although a slight deviation (regarding velocity) from the numerical solution can be observed as one moves away from the critical level. In particular, it should be mentioned that for the far left panels, $a_0(y-yc)^{1/2+i|\mu|}$ corresponding to an upward wave is sufficient to fit correctly the numerical solution, meaning that the counter-propagating wave in zone I is reflected at $y=0$.
However, for the middle and right-hand side panels, it is not clear whether the first-order Frobenius solutions can be reconnected to the incident and reflected waves at $y=0$.
We now examine attenuation or amplification phenomena in each column of panels in Fig. \ref{vA_pole}, from left to right. In the left panels, for which $R>1/4$, the latitudinal velocity is strongly attenuated at the critical levels, and so is the wave flux action as we can expect from the left panel of Fig. \ref{wA_IRT}. While the transmitted wave is totally absorbed, the reflected wave remains, which is consistent with analytical values of the transmission and reflection coefficients in Fig. \ref{RT_st} (see white crosses). In the middle panels, where we also have $R>1/4$, the wave is over-transmitted but not over-reflected, which is also consistent with Fig. \ref{RT_st} (see black crosses).
In view of the wave action flux, the amplification of the transmitted wave does not seem to be related to the critical level because this quantity is greatly reduced after the critical level (see also the white cross in the left panel of Fig. \ref{wA_IRT}).
The third column of panels now refers to the regime where $R<1/4$. The wave is over-reflected by a factor of $\sim1.5$ and over-transmitted by a factor of $\sim2$ in concordance with the reflection and transmission coefficients plotted in Fig. \ref{RT_inst}. The wave action flux is negative and $|\mathcal{A}_\mathrm{I-R}|<|\mathcal{A}_\mathrm{T}|$ by a factor of $3$ as observed in Fig. \ref{wA_IRT}.
These three case studies reinforce the idea of \cite{BB1967} in the case of stratified z-sheared flows that the wave energy can be lost to the mean flow, or on the contrary that the wave can take energy from the mean flow.
\begin{figure*}[ht]
\resizebox{\hsize/3}{!}{\includegraphics[trim=0cm 0cm 0cm 1cm,clip]{figures/v_y_c+_n1_Ro0p3}}
\resizebox{\hsize/3}{!}{\includegraphics[trim=0cm 0cm 0cm 1cm,clip]{figures/v_y_c0_n1_Ro0p3}}
\resizebox{\hsize/3}{!}{\includegraphics[trim=0cm 0cm 0cm 1cm,clip]{figures/v_y_c-_n1_Ro0p3}}
\resizebox{\hsize/3}{!}{\includegraphics[trim=0cm 0cm 0cm 1cm,clip]{figures/a_deg10_y_+_n1_rf1e-08_Ro0p3}}
\resizebox{\hsize/3}{!}{\includegraphics[trim=0cm 0cm 0cm 1cm,clip]{figures/a_deg10_y_0_n1_rf1e-08_Ro0p3}}
\resizebox{\hsize/3}{!}{\includegraphics[trim=0cm 0cm 0cm 1cm,clip]{figures/a_deg10_y_-0_n1_rf1e-08_Ro0p3}}
\caption{\aava{Same quantities as in Fig. \ref{vA_pole} (when the box is at the pole), but for a box tilted by 10 degrees relative to the pole and for different values of the shear, wavenumbers, and inertial frequencies.} Note that unlike Fig. \ref{vA_pole}, we do not have analytical solutions in the shear region. In all panels, the horizontal wave numbers are set to $k_x=k_z=1$, and the shear is $\aava{R_\mathrm{o}}=0.3$.
\textit{From left to right:} The inertial frequency is set to $\omega=0.31$, $\omega=0.16$, and $\omega=0.1$.
In the third panels, we indicate $y_0$ and $y_-$ by dashed dotted and dashed vertical lines, respectively.
}
\label{vA_10deg}
\end{figure*}
\subsection{Numerical exploration at constant shear when the box is inclined}
We now investigate wave propagation through the different critical levels when the box is inclined with respect to the rotation axis. We still assume that the shear region (zone II) has a linear shear flow profile ($n=1$). In contrast to the polar configuration, we do not have analytical solutions to the ordinary differential equation. Instead of going for an extensive numerical investigation of the parameter space, we will rather focus on the dynamics of inertial waves in our three-layer model as they cross the critical levels $\sigma=\pm\widetilde{f}$ and $\sigma=0$.
Our results are presented in Fig. \ref{vA_10deg} for a box inclination of $10\degree$ relative to the rotation axis, a shear fixed to $R_\mathrm{o}=0.3$ and wave numbers set to $k_x=k_z=1$.
The value of the frequency $\omega$ determines the existence and the nature of the critical level as detailed in the table of Appendix \ref{wAI_III}.
Like in Fig. \ref{vA_pole}, we plot in each column of Fig. \ref{vA_10deg} the latitudinal velocity and the wave action flux.
From left to right, we illustrate our results for the critical levels $\sigma=+\widetilde{f}$, $\sigma=0$, and $\sigma=\{0,-\widetilde{f}\}$ (i.e., there are two critical levels in the rightmost panels).
One can notice that the first-order Frobenius approximation is not in good agreement anymore with the numerical solution in the entire shear region, though it remains a reasonable approximation close enough to a critical level.
Unlike the polar case, the discrepancy far outside the critical levels is due to the linear approximation that the governing ODE takes around critical levels (see Eq. (\ref{PcDL_ft}) and (\ref{Pcorot})).
In the left panels of Fig.~\ref{vA_10deg}, the reflected and transmitted waves are strongly attenuated at the critical level $y=y_+$.
Part of the wave energy is laid down to the mean flow, as corroborated by the drop in the wave action flux.
In the middle panels, we do not see any discontinuity at the corotation $y=y_0$, which is in line with the theoretical analysis in Sect. \ref{corot_inclined} for a constant shear.
However, the wave is over-reflected and over-transmitted, possibly due to the polynomial form of the solutions in the Frobenius series around $y=y_0$.
In the right panels, the wave encounters successively critical levels at $y=y_0$ and at $y=y_-$.
Although the wave going through the shear region is not attenuated at the corotation $y=y_0$, it is completely absorbed at the second critical level $y=y_-$ where the wave action flux drops to zero. This is consistent with the transmission coefficient in the left panel of Fig. \ref{atnu_deg}.
The latitudinal velocities displayed in the top-left and top-right panels of Fig. \ref{vA_10deg} support the concept of a valve effect.
Indeed, according to our analysis in Sect. \ref{cc_ft} with the Frobenius method and given the shear and wave numbers of Fig. \ref{vA_10deg}, the attenuation is strong for a downward wave meeting the critical level $y_+$ (first panel) while the attenuation is strong for an upward wave that meets the critical level $y_-$ (third panel).
Before and after the critical levels $y_-$ and $y_+$, respectively, we observe fast oscillations of shorter and shorter period close to the critical level as already evidenced by \cite{BB1967}.
The analysis to determine how the wave is reflected in the shear zone can hardly be taken any further, because Frobenius solutions are not fully separable into upward and downward waves.
We emphasise that the behaviour of the wave at corotation $y=y_0$ when the box is at the pole stands out clearly different from the case when the box is inclined for a linear mean flow profile.
This is particularly true in terms of the absolute value of the wave action flux that is subject to rise and drop in the polar configuration, whereas it remains conserved when the box is inclined.
In this inclined case, the only way for a wave to be attenuated without \aava{friction} is that it meets critical levels $\sigma=\pm\widetilde{f}$.
Depending on the critical level encountered, an upward or downward wave will not be attenuated in the same way as known by the valve effect.
In addition, we no longer observe amplification due to the critical level but still ``geometric'' amplification (for instance in the middle panels where we can observe over-transmission and over-reflection), that can be explained by the exponential form of the Frobenius series.
\subsection{Numerical results with a non-constant shear}
\label{nonlin}
\begin{table*}[ht!]
\begin{tabular}{ c|c ||c c c | c c c | c c c}
&&\multicolumn{3}{c|}{$R$} & \multicolumn{3}{c|}{$T$} & \multicolumn{3}{c}{$\mathcal{A}_\mathrm{T}/\mathcal{A}_\mathrm{I-R}$}\\
\rule[-1ex]{0pt}{3.5ex}
&\backslashbox{\makecell{case}}{$n=$} & \makecell{$1$\\$\ $} & \makecell{$2$\\$\ $} & \makecell{$3$\\$\ $} & \makecell{$1$\\$\ $} & \makecell{$2$\\$\ $} & \makecell{$3$\\$\ $} & \makecell{$1$\\$\ $} & \makecell{$2$\\$\ $} & \makecell{$3$\\$\ $}\\
\hline\hline
\rule[-1ex]{0pt}{4ex} \multirow{3}{*}{\rotatebox[origin=r]{90}{$\theta_0=0\degree$}} &$R_\mathrm{o}=0.3$ & $1.3\tt{-4}$ & $9.1\tt{-3}$ & $1.5\tt{-2}$ & $1.3\tt{-1}$ & $ 2.1\tt{-2}$ & $7.4\tt{-2}$ & $-3.2\tt{-8}$ & $-1.6\tt{-4}$ & $-1.1\tt{-2}$\\
\rule[-1ex]{0pt}{4ex}& $R_\mathrm{o}=0.8$ & $7.4\tt{-1}$ & $5.7\tt{-3}$ & $1.7\tt{-4}$ & $1.4$ & $9.1\tt{-5}$ & $4.0\tt{-6}$ & $-1.1\tt{-1}$ & $-2.1\tt{-10}$ & $-4.1\tt{-13}$\\
\rule[-1ex]{0pt}{4ex} &$R_\mathrm{o}=1.8$ & $1.5$ & $6.4\tt{-1}$ & $7.0\tt{-1}$ & $1.8$ & $1.1$ & $9.7\tt{-1}$ & $2.9$ & $-1.9$ & $-1.8$ \\\hline
\rule[-1ex]{0pt}{4ex} \multirow{3}{*}{\rotatebox[origin=r]{90}{$\theta_0=10\degree$}} &$y_+$ & $8.2\tt{-2}$ & $2.2\tt{-2}$ & $2.4\tt{-2}$ & $1.0\tt{-1}$ & $8.9\tt{-2}$ & $8.3\tt{-2}$ & $3.5\tt{-1}$ & $2.7\tt{-1}$ & $2.4\tt{-1}$\\
\rule[-1ex]{0pt}{4ex}& $y_0$ & $5.3$ & $4.9$ & $4.7$ & $4.9$ & $4.4$ & $4.3$ & $1.0$ & $1.0$ & $1.0$\\
\rule[-1ex]{0pt}{4ex} &$y_0$ and $y_-$ & $5.0$ & $6.2$ & $6.1$ & $2.3\tt{-4}$ & $4.0\tt{-2}$ & $2.2\tt{-1}$ & $1.1\tt{-9}$ & $2.0\tt{-5}$ & $6.7\tt{-4}$\vspace{0.5cm}
\end{tabular}
\caption{Reflection $R$, transmission $T$ coefficients and ratio of the wave action flux above and below a critical level $\mathcal{A}_\mathrm{T}/\mathcal{A}_\mathrm{I-R}$, for a linear, square and cubic mean flow ($n=1,\ 2, \text{ and }3$, respectively). The six cases presented here are, in order, the three cases examined at the pole in Fig. \ref{vA_pole} and the three cases of a tilted configuration examined in Fig. \ref{vA_10deg}.}
\label{tab_RTA}
\end{table*}
The choice of a linear mean flow profile allows the resolution of the ordinary differential equation at the pole and a simpler implementation of the Frobenius method. Nevertheless, the correspondence between a global and a local mean flows as presented in Sect. \ref{mean_flow_sect} involves higher-order terms than a simple linear dependence.
Therefore, it is important to examine the effect of different mean flow profiles with non-zero $U''$. Now assuming $n > 1$ for the shear flow profile in zone II (see Eq.~\ref{U_pro}), the Frobenius method in the inclined and polar cases still holds provided that
\begin{equation}
\left\{\begin{aligned}
&y_0=\sqrt[n]{\frac{\omega}{k_x\Lambda}},\ \text{ and }\ y_\pm=\sqrt[n]{\frac{\omega\mp\widetilde{f}}{k_x\Lambda}}\\
&R_\mathrm{o}=\Lambda n(y_{0,\pm})^{n-1}.
\end{aligned}\right.
\end{equation}
Moreover, we specify that the critical level in $y<0$ for even values of $n$ is not examined given our numerical set up where the shear region is located in the range $y\in[0,1]$.
The conditions to have critical levels inside the shear region (zone II) are the same as those for a constant shear (see Appendix \ref{wAI_III}).
We show in Table \ref{tab_RTA} numerical values of the reflection and transmission coefficients along with the ratio of the wave action flux below and after the critical level in the six parameter sets illustrated in Figs. \ref{vA_pole} and \ref{vA_10deg} for linear, square and cubic mean flow profiles.
For all cases, $R$, $T$ and $\mathcal{A}_\mathrm{T}/\mathcal{A}_\mathrm{I-R}$ change quite significantly between $n=1$, 2 and 3.
At the pole, the (``geometric'') over-transmission found for $R_\mathrm{o}=0.8$ disappears for $n=2$ and $n=3$, where
the reflected and transmitted waves are strongly attenuated.
Similarly, in the case where $R_\mathrm{o}=1.8$, the over-transmission and over-reflection disappear when $n=3$, whereas $|\mathcal{A}_\mathrm{T}/\mathcal{A}_\mathrm{I-R}|>1$ which entails that the transmitted wave is taken wave action flux from the mean flow just after the critical level.
We find that the Frobenius method does not give consistent results for the case where $R_\mathrm{o}=1.8$ and $n=\{2,3\}$.
When the box is tilted and for the critical level $y_+$, the reflected wave is more attenuated than the transmitted one, which is consistent with stronger attenuation of the counter-propagating wave discussed in the previous section.
At the corotation and when the box is inclined (fifth row of Table \ref{tab_RTA}),
\aava{the wave action flux remains the same in the whole domain for linear or non-linear mean flow profiles, in agreement with the wave action flux derived analytically in Eq. (\ref{wAconcor})}.
We speculate that the over-transmission and over-reflection are due to the polynomial form of the solutions in the shear region. \aava{These wave amplifications may be related to shear instabilities in this particular three-layer configuration, and are probably not linked with the presence of a critical level whose implications on the flow are well diagnosed by the wave action flux}.
In the last case (sixth row), where two critical levels co-exist, the transmitted wave is less attenuated at $y=y_-$ from $n=1$ to $n=3$ but it remains more attenuated than the reflected wave as discussed in the previous section. Again, no jump is found in the wave action flux at the corotation despite analytical predictions.
The same analysis was carried out for different amplitudes of the Rayleigh friction force, up to $\sf=\tt{-2}$.
Of course, the wave action flux is no longer constant, but we observe that all three parameters of Table \ref{tab_RTA} change very little compared to the case where $\sf=\tt{-8}$.
This result is consistent with \cite{AM2013} in the context of gravity waves and vertical shear.
We comment that while a low \aava{friction} is mandatory in the numerical code to solve the ODE at singularities $y_\pm$, it is not the case at the corotation $y=y_0$.
\section{Astrophysical discussion}
\label{astrodi}
\begin{figure}
\includegraphics[width=\hsize]{figures/Dom_age_medn}
\caption{Absolute values of the shear contrast normalised by the rotation at the pole against the age of stars of K and G spectral types. Solid lines feature $0.3<R_\mathrm{of}<0.9$, dashed lines $R_\mathrm{of}\lesssim0.3$ and transparent lines $R_\mathrm{of}>1$.}
\label{Dom}
\end{figure}
\subsection{Latitudinal differential rotation in stars}
\label{appli_lat}
In stars, latitudinal differential rotation is often characterised by the difference in rotation frequency between the equator and the pole, that is the quantity $\Delta\Omega=\Omega_\mathrm{eq}-\Omega_0$, where $\Omega_\mathrm{eq}$ is the rotation frequency at the equator \cite[e.g.,][]{BJ2017}. We will now refer to $\Delta\Omega$ as the shear contrast. Different regimes are distinguished according to the value of $\Delta\Omega$: anti-solar-like rotation for $\Delta\Omega<0$, cylindrical rotation for $|\Delta\Omega|\ll1$, and solar-like rotation rotation for (not-too-low) positive $\Delta\Omega$ (e.g. $\Delta\Omega/\Omega_{0}\simeq0.3$ for the Sun).
Several works based on three-dimensional numerical simulations have explored the range of physical parameters leading to each aforementioned regime in stars and in giant planets \cite[e.g.,][]{GW2013,VS2016,BS2018}
In particular, \cite{BS2017} have derived a criterion based on mixing length theory and calibrated with 3D simulations that determines the rotation profile of a solar-like star. This criterion is based on the fluid Rossby number $R_\mathrm{of}$, defined as
\begin{equation}
R_\mathrm{of}=R_\mathrm{of,\odot}\,\Omega_*^{-0.82}M_*^{1.53},
\label{B1}
\end{equation}
where $R_\mathrm{of,\odot}=0.89$ is the solar fluid Rossby number, and $\Omega_*$ and $M_*$ are the mean rotation and the mass of the star respectively, normalised with their solar value. \cite{BS2017} highlighted the following three regimes:
\begin{itemize}
\item $R_{\mathrm{of}}>1$ for anti-solar-like rotation,
\item $0.3<R_{\mathrm{of}}<0.9$ for solar-like rotation,
\item $R_{\mathrm{of}}\lesssim0.3$ for cylindrical rotation.
\end{itemize}
Furthermore, they introduced the shear contrast $\Delta\Omega_{\mathrm{S}}$ at the colatitude $30\degree$ as
\begin{equation}
\Delta\Omega_\mathrm{S}=\Omega_\mathrm{eq}-\Omega(\theta=30\degree),
\end{equation}
as the rotation frequency is often ill-defined at low colatitudes in 3D numerical simulations in spherical geometry. From their 3D simulations, \citet{BS2017} obtained the following scaling:
\begin{equation}
\Delta\Omega_\mathrm{S}=\Delta\Omega_{\mathrm{S}, \odot}\,M_*^{0.73}\Omega^{0.66},
\label{B2}
\end{equation}
where $\Delta\Omega_{\mathrm{S},\odot}\simeq565\times\tt{-9}\second^{-1}$ is the solar value of $\Delta\Omega_\mathrm{S}$ calculated from \citet{GT2007}. Using the mean flow profile (Eq. (\ref{law})), $\Delta\Omega_\mathrm{S}$ can be related to our shear contrast $\Delta\Omega$ via
\begin{equation}
\frac{\Delta\Omega}{\Omega_0}=\chi=\frac{4}{3}\frac{\Delta\Omega_\mathrm{S}}{\Omega_0}.
\label{deltaOmega}
\end{equation}
We show in Fig. \ref{Dom} the quantity $\Delta\Omega/\Omega_0$, expressed in Eq. (\ref{deltaOmega}), versus the age of solar-like stars for K to G spectral types.
To compute this quantity, we have used grids of the 1D stellar evolution code STAREVOL \cite[see][for details of the code]{AP2019}.
In light of Fig. \ref{Dom}, the stars in the pre-main sequence ($\mathrm{age}\lesssim 100\,\mega\mathrm{yr}$) exhibit cylindrical rotation as they
are fast rotating. During the main sequence, stars mostly feature solar-like rotation, while anti-solar-like rotation is observed at the end of the main sequence from $0.8$ to $1.1\,M_\odot$. According to Fig. \ref{Dom}, a limit on the absolute value of the normalised shear contrast can be set to $|\Delta\Omega|/\Omega_0<0.5$.
However, as already stressed by \cite{BB2018}, the latitudinal shear inferred by asteroseismology can be much larger than predicted by numerical simulations. This can actually be inferred by comparing the shear factors in their work \cite[see Table S3 from supplementary materials in][]{BB2018} with ours given in Fig. \ref{Dom}. Moreover, according to their study, cylindrical and anti-solar differential rotation are hardly unambiguously detectable.
Finally, one should recall that \cite{BS2017}'s scaling laws given in Eqs. (\ref{B1}) and (\ref{B2}) are derived for K and G spectral type stars only.\\
Since we now know values of the shear contrast, we can calculate the ``shear'' Rossby number $R_\mathrm{o}=U'/(2\Omega)$, given by the following relationship:
\begin{equation}
R_\mathrm{o}=-\frac{3}{2}\cos\theta_0\sin^2\theta_0\,\chi,
\end{equation}
which has been derived from Eq. (\ref{U_conic}) by keeping only zero-order terms in $y$.
Taking $\chi\simeq0.3$ as a representative value of the shear contrast for main sequence G and K stars, we find that the Rossby number is maximal when $\theta_0\simeq55\degree$, its maximum value being $R_\mathrm{o}\simeq-0.17$. In particular, $R_\mathrm{o}\simeq-0.013$ for $\theta_0=10\degree$ and $R_\mathrm{o}\simeq-0.076$ for $\theta_0=80\degree$.
These values are useful to interpret wave flux action transmission at critical levels $\sigma=\pm\widetilde{f}$, considering Fig. \ref{atnu_deg} \aava{(we recall that at the corotation the wave action flux is fully transmitted)}. A downward (upward) propagating wave through $\sigma=\widetilde{f}$ ($\sigma=-\widetilde{f}$) is
\begin{itemize}
\item totally absorbed provided that $k_z\gtrsim0.1 k_x$ ($\alpha_k\gtrsim0.1$) at $\theta_0=10\degree$, and that $k_z\gtrsim k_x$ ($\alpha_k\gtrsim1$) at $\theta_0=80\degree$,
\item strongly attenuated for $k_z\sim0.1 k_x$ ($\alpha_k\sim0.1$) at $\theta_0=80\degree$,
\item fully transmitted given that $k_z\ll \tt{-1}k_x$ ($\alpha_k\ll\tt{-1}$) for both inclinations.
\end{itemize}
These results also hold for anti-solar-like differential rotation, since the transmission factor $T_{\theta_0}$ is function of $|R_\mathrm{o}|$. For larger values of $|R_\mathrm{o}|$, i.e. for larger values of the shear contrast, waves are less damped at critical levels $\sigma=\pm\widetilde{f}$ at a given $\alpha_k=k_z/k_x$.
The connection between this ratio of the vertical and azimuthal wavenumbers in the local model and an equivalent ratio of global wavenumbers in the spherical geometry is not straightforward.
A first hint can be to state that $k_z\sim k_r$, where $k_r$ is the wavenumber in the global radial direction, while $k_x\sim m/(r_0\sin\theta_0)$, where $m$ is the azimuthal order of the considered mode of the tidal potential \citep[when $m\neq 0$;][]{ZT1997}. Then, we have $\alpha_k\equiv k_r r_0\sin\theta_0/m \equiv 2\pi\,r_0/\lambda_r\times \sin\theta_0/m$, by introducing $\lambda_r$ the radial wavelength of the tidal wave.
In the case where $r_0>\lambda_r m$, we should thus be in the regime where the tidal wave is attenuated.
\subsection{Cylindrical differential rotation in Jupiter and Saturn}
\begin{figure}
\includegraphics[width=\hsize]{figures/Ro_GiantPlanet.pdf}
\caption{Rossby numbers $R_\mathrm{o}=\frac{r}{2\Omega}\df\Omega r$ for Jupiter and Saturn as a function of the axial distance $r$.}
\label{JuSa}
\end{figure}
\begin{table*}[t!]
\centering
\begin{tabular}{c|cc|cc|c| c}
\rule[-1ex]{0pt}{0ex} & \multicolumn{2}{c|}{ spherical } &\multicolumn{2}{c|}{ equatorial} & \makecell{\\$\alpha_k$} & \makecell{\\Regime}\\
\rule[-1ex]{0pt}{0ex}& $l$ & $m$ & $m$ & $l_\Theta$ & & \\\hline\hline
\rule[-1ex]{0pt}{4ex} asynchronous & 2 & 2 & 2 & \aava{2} & \aava{1} & \aava{total} absorption \\
\rule[-1ex]{0pt}{4ex} inclined &2 & 1 & 1 & 2 & $2$ & total absorption\\
\rule[-1ex]{0pt}{4ex} eccentric & 2 & 0 & 0 & 2 & $\rightarrow\infty$ & total absorption\vspace{0.5cm}
\end{tabular}
\caption{Regime of wave transmission at corotation deduced from Figs. \ref{O2} and \ref{atnuO2} for three orbital states of a satellite around a giant gaseous planet.}
\label{lm}
\end{table*}
Mathis et al. (in prep.) have developed an equatorial model to examine inertial wave properties in the outer convective layers of giant gaseous planets such as Saturn and Jupiter, which are subject to cylindrical differential rotation.
In their model (built in cylindrical coordinates), they derived a Schrödinger-like differential equation for $\Psi=\sqrt{\rho}r^2v_r$ under the anelastic approximation, where $\rho$ is the density, $r$ the axial distance coordinate, and $v_r$ the axial velocity. For free inertial waves, their second-order differential equation is:
\begin{equation}
\dfd{\Psi}{r}+\left(\frac{\kappa_r^2l_\Theta^2}{\hat\sigma^2r^2}-\frac{l_\Theta^2+m^2}{r^2}\right)\Psi=0,
\label{equa}
\end{equation}
where $l_\Theta$ and $m$ denote the equatorial and azimuthal wavenumbers, respectively,
$\hat\sigma=\omega+m\Omega(r)$ is the (linear) Doppler-shifted frequency, and $\kappa_r$ the ``axial'' epicyclic frequency defined as $\kappa_r^2=4\Omega^2+2\Omega r\df{\Omega}{r}$.
For cylindrical differential equation, the corotation resonance $\hat\sigma=0$ results in critical cylinders \citep[see][]{BR2013} characterised by a critical axial distance $r=r_{\mathrm{c}}$. The Taylor expansion of Eq. (\ref{equa}) at first order around $r_{\mathrm{c}}$ gives:
\begin{equation}
\dfd{\Psi}{r}+\left(\frac{ l_\Theta^2(1+R_\mathrm{o})}{m^2R_\mathrm{o}^2(r-r_{\mathrm{c}})^2}-\frac{l_\Theta^2+m^2}{r_{\mathrm{c}}^2}\right)\Psi=0,
\label{equa_exp}
\end{equation}
by setting $R_\mathrm{o}=r_{\mathrm{c}}\left.\df{\Omega}{r}\right|_{r=r_{\mathrm{c}}}/(2\Omega_\mathrm{c})$ the local Rossby number in cylindrical coordinates. By writing $\alpha_k^2=l_\Theta^2/m^2$, Eq. (\ref{equa_exp}) becomes very similar to our \aava{ODE} Eq. (\ref{Pc}) when the box is located at the South pole and for a constant shear (i.e. with $\widetilde{f}=0$, $f=-1$ and $U''=0$).
Note that, when the local shear box model is located at the equator, the latitudinal coordinate $y$ is directed along the (vertical) rotation axis, whereas when the box is at the poles, $y$ is directly the axial distance. That is why a polar configuration of the box best reproduces the ``equatorial'' model of Mathis et al. (in prep.).
Moreover, the convention of a plus sign in the Doppler-shifted frequency explains why Eq. (\ref{equa_exp}) is analogous to our \aava{wave propagation} equation when the box is at the South pole rather than at the North pole.
In Fig. \ref{JuSa}, we show Jupiter's and Saturn's local Rossby number from Mathis et al. (in prep.). Cylindrical differential rotation extends in the outer layer of the convective envelope of both planets, in agreement with Juno and Cassini Grand Finale observations \citep[respectively]{KG2017,GK2019}. According to Fig. \ref{O2}, where we notably plotted $R=\alpha_k^2(1+R_\mathrm{o})/R_\mathrm{o}^2$ at the South pole (dark red and purple areas), we can assess the role of the critical level for wave transmission across the corotation, in terms of the wavenumber ratio $\alpha_k$. In Fig. \ref{JuSa}, the Rossby number \aava{satisfies} $R_\mathrm{o}\lesssim0.27$ for Saturn and $R_\mathrm{o}\lesssim0.07$ for Jupiter. Given this range of values, two regimes can be evidenced for waves and wave action fluxes through the corotation:
\begin{itemize}
\item waves are strongly attenuated for $l_\Theta\gtrsim m$ (see also Fig. \ref{atnuO2} for $R_\mathrm{o}=-|R_\mathrm{o}|$ as the transmission factor in this figure is plotted for the North pole).
\item waves can be over-reflected and over-transmitted for $l_\Theta\ll m$, and can potentially lead to instabilities given specific boundary conditions.
\end{itemize}
To give an idea of the values these wavenumbers can take, we have listed in Table \ref{lm} three typical orbital states, where, in order, the asynchronous, eccentricity and obliquity tides are supposed to be dominant \citep{O2014}.
These states are described by the ``spherical'' quadrupolar components of the dominant terms in the tidal potential, i.e. the degree $l$ and the order $m$ of the spherical harmonics. The analogy with the equatorial model is then made to get $m$, \aava{and $l_\Theta$ is chosen to approximate as best as possible the behaviour of the Legendre polynomial $P_l^m(\cos\Theta)$ around the equator with a simple trigonometric function $\Re\{\exp[i(l_\Theta \Theta+\phi)]\}$, where $\phi$ is the appropriate phase} (Mathis et al. in prep.). \aava{To find the associated wave attenuation at corotation for the three main tides, one can use Figs. \ref{O2} and \ref{atnuO2} for $\alpha_k\geq1$ and $R_\mathrm{o}\lesssim0.27$, and look at the South pole (as Fig. \ref{atnuO2} is plotted at the North pole, one has to take the opposite Rossby number). From Fig. \ref{O2}, we can assess that waves excited by these tides are always in the so-called ``stable'' regime for these ranges of $R_\mathrm{o}$ and $\alpha_k$, which excludes an amplification of these waves. Moreover, from Fig. \ref{atnuO2}, we also observe that waves are completely absorbed at corotation for our given ranges of parameters. By consequence, waves excited by the asynchronous, inclined, or eccentric tides in Jupiter and Saturn are expected to transfer all their wave action flux to the mean flow at corotation}.
\section{Conclusion and perspectives}
\label{conclu}
The present study was motivated by the works of \cite{BR2013} and \cite{GB2016,GM2016}, who showed
that differential rotation can strongly affect the propagation and dissipation properties of (tidal) linear inertial waves. They considered
different rotation profiles typical of stellar and planetary interiors and they pointed out that tidal waves can deeply interact with zonal flows at corotation resonances, leading to intense wave energy dissipation, alongside possible instabilities.
In this paper, we have investigated the transmission of free inertial waves with latitudinal stratification and differential rotation, at the corotation resonance (characterised by a zero Doppler-shifted wave frequency) and more broadly at critical levels (any singularities of the governing second-order wave propagation equation in the inviscid limit).
For this purpose, we have built a new local Cartesian box model with horizontal shear, modelling a small patch of the convective zone of a low-mass star or a giant planet.
By considering the inclination of the local reference frame relative to the rotation axis, we have examined the effect on wave propagation through a critical level of a conical rotation profile at a general colatitude when the box is tilted, or of a cylindrical rotation profile when the box is at the North or South poles.
These rotation profiles are inspired by those observed or expected in the Sun, low-mass stars, and giant gaseous planet in our solar system. Three critical levels can be identified when the box is inclined relative to the rotation axis: the corotation resonance and two others critical levels that arise from the inclination between the gravity and the rotation vectors, and which are defined by a Doppler-shifted frequency equal to plus or minus the latitudinal component of the rotation frequency. When the box is at the poles, critical levels are restricted to the corotation.
In order to diagnose the behaviour of a wave passing through a critical level for both aforementioned rotation profiles, we made use of an invariant called the wave action flux that is independent of the latitudinal coordinate in a non-dissipative fluid flow.
This invariant was used when the ``directional'' flux of angular momentum (here latitudinal) can not be constructed easily from the mean perturbed velocity, as it is the case for example in \cite{LT1978} for Rossby waves in plane-parallel shear flows.
The wave action flux has already been used in vertically-stratified shear flows in the presence of rotation or magnetic fields to interpret the role of critical levels \citep{G1975,AM1978,G1979,M2009,MB2012}.
Using the condition that this invariant is discontinuous at critical levels, we have demonstrated in Sect. \ref{sec3} that waves can be either fully transmitted, damped or even amplified after passing through critical levels as a result of wave action flux exchanges.
These different regimes of wave transmission are found both with conical and cylindrical rotation profiles; they depend on the critical level encountered, on the wave properties (e.g. the propagation direction, wavenumbers) and on the profile of the mean flow. Table \ref{recap_ana} summarises the main analytical results.
We have then confronted our analytical results with a three-layer numerical model that comprises a shear zone in which the critical level is located and two surrounding shear-free zones that allow incident, reflected and transmitted waves.
A difference with the analytical model is the introduction of small dissipative force under the form of a Rayleigh friction \citep[also called frictional force by][]{O2009} to avoid strict singularities.
It does not seem to affect the results, since analytical and numerical results match \aava{quite} well when using a \aava{power-law mean flow profile} and by varying the \aava{friction}.
This conclusion is also shared by the work of \cite{AM2013} who studied corotation resonances for gravity waves propagating in stratified and vertically shear flows.
Based on the analytical results, we have discussed in Sect. \ref{astrodi} possible applications to stellar and planetary interiors. We have estimated the rate of differential rotation in solar-like stars by the shear contrast (the rotation difference between the pole and the equator) and in giant gaseous planets through the local Rossby number (the ratio between the shear and the rotation frequency in cylindrical coordinates). We find that for K and G-type stars along their lifetime, and for Jupiter and Saturn at the present time, a regime where inertial waves are strongly damped is largely preferred in the convective envelope of these objects. Similar conclusions were found by \cite{AM2013} for internal gravity waves through critical levels in the core of solar-like stars.
\\
It is interesting to discuss different regimes of the wave transmission in terms of angular momentum transfer for cases of strong damping and wave amplification.
First, we have to underline that the theoretical analysis presented in Section 3 (using the Frobenius method) does not adequately characterise wave (over-)reflection as similarly observed in the numerical section \ref{secnum}.
What we can access is the wave action flux on either side of the critical level.
The analysis on changes in the wave action flux across the critical level allows us to understand whether energy is deposited to or extracted from the mean flow, in line with the work carried out by \cite{M1961,BB1967,G1975,LB1985}. In the presence of a locally conical differential rotation, we have demonstrated that a valve effect can be found for critical levels other than the corotation, analogous to the results of \cite{A1972} and \cite{G1975} for hydromagnetic and gravito-inertial waves with a vertical shear. For these peculiar critical levels, waves can be attenuated when going one direction, mainly featured by the sign of the rotation components in the box, or fully transmitted when going the other direction.
For cylindrical differential rotation, we have found a criterion analogous to the Miles-Howard theorem for stratified shear flows \citep{MH1964}, which for inertial waves as in this work can be formulated as:
\begin{equation}
R=\frac{k_z}{k_x}\frac{1-R_\mathrm{o}}{R_\mathrm{o}^2}
\left\}\begin{aligned}
&>1/4 \text{: wave attenuation;}\\
&<1/4 \text{: possible wave over- }\left|\begin{tabular}{l} reflection, \\ transmission. \end{tabular}\right.
\end{aligned}\right.
\label{resum}
\end{equation}
The above criterion depends on the shear Rossby number $R_\mathrm{o}=U'/(2\Omega)$ and the vertical ($k_z$) and longitudinal ($k_x$) wavenumbers. \aava{This last point is an important difference from the Miles-Howard criterion, which does not involve wavenumbers. For this reason the analogy between Eq.(\ref{resum}) and the Miles-Howard stability criterion must be taken with care.}
We \aava{also} stress that \aava{Eq. (\ref{resum})} is very different from the Rayleigh's inflection point theorem for Rossby waves, which are a subclass of inertial waves when neglecting the vertical or the radial perturbed velocity \citep{B1966,LT1978}.
The Miles-Howard criterion allows to disentangle between critical levels where strong wave attenuation is expected for $\mathrm{Ri}>1/4$ (where $\mathrm{Ri}$ is the Richardson number), and those where over-reflection and over-transmission can lead to potential shear instabilities for $\mathrm{Ri}<1/4$.
\citet{L1988} warns, however, that over-reflection and over-transmission are a necessary but not sufficient condition for shear instability.
Such amplifications leading to the instability require peculiar conditions in a three-layer model, where the shear zone that features the critical level is surrounded by a region of incoming propagating waves, and a ``sink'' zone to force waves to cross the evanescent shear zone.
Special boundary conditions are necessary for the wave to return successively to the critical level and induce wave amplitude growth. Recent studies \citep[see e.g.][for a review]{CT2012} have revisited instabilities in stratified shear flows by studying multiple counter-propagating waves that can interact to grow in amplitude with time (with conditions as phase-locking). A parallel is drawn between over-reflection mechanisms and interacting counter-propagating waves by \citet{HH2007} to describe baroclinic instabilities for Rossby waves.
Contrary to what our results predict, \citet{BR2013} did not observe any instabilities of inertial waves when using cylindrical differential rotation.
Several reasons can be put forward to explain this discrepancy, such as boundary conditions (as discussed in the previous paragraph), or the values of the shear and horizontal wavenumbers, since they may not be in the regime allowing instabilities according to the criterion Eq. (\ref{resum}), which further needs to be adapted to their global cylindrical geometry.
We stress that when exploring different power-laws for the mean flow profiles, over-reflection was not retrieved for a non-linear mean flow in cylindrical differential rotation. With conical differential rotation, \cite{GB2016,GM2016} did observe instabilities, but only for sufficiently low viscosities, whereas our study showed little dependence on the \aava{friction}, and rather highlights possible over-transmission for non-linear flows.
Lastly, we underline that a temporal analysis on the growth rate of perturbations should be undertaken to unravel instabilities, which has not been performed in this paper but in other separate papers \citep[][]{PPM2020,PPMB2020}.
This ab-initio analytical study is thus a first step to understand how inertial waves interact with a mean flow subject to latitudinal differential rotation at critical levels, in the context of tidal dissipation in differentially rotating stars and planets. Possible feedbacks of the perturbed wave on equilibrium quantities and the mean flow are not taken into account in this study, \aava{nor are non-linearities in the perturbed hydrodynamical wave equations}. Nonetheless, they should be considered in future studies since \cite{BO2010} and \cite{BR2013} suggested important non-linear effects for inertial waves at corotation. Finally, magnetism may also play an important role to dissipate or redistribute angular momentum at critical levels through magnetic stresses \citep[e.g.][]{W2016, W2018,LO2018,AM2019}.
\begin{acknowledgements}
We would like to thank the anonymous referee, as well as A. Barker, for the helpful comments and suggestions regarding our work. A. Astoul, J. Park, and S. Mathis acknowledge funding by the European Research Council through the ERC grant SPIRE 647383. The authors acknowledge the PLATO CNES funding at CEA/IRFU/DAp and IRAP.
This research has made use of NASA's Astrophysics Data System and of the software MATLAB version R2018a.
Finally, we thank Quentin André and Jéremy Ahuir for fruitful discussions on the details of the analytical model.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
As one of the representative examples on inverse problems in partial differential equations, the inverse source problem (ISP) has received a lot of attention in mathematical and engineering communities. Especially, the ISP for wave propagation is an active and important research topic in the field of inverse problems due to the significant applications in diverse scientific areas, such as magnetoencephalography \cite{magnetoencephalography}, photoacoustic tomography \cite{photoacoustic}, ultrasonics imaging \cite{ultrasonics}, antenna design and synthesis \cite{Antennas1,Antennas2}.
The ISP is to determine the unknown source from a knowledge about the solution. It is challenging due to the ill-posedness for lack of uniqueness or stability. This paper is concerned with an ISP for the wave equation, which has been
extensively investigated for the deterministic case. The well-posedness and stability can be found in \cite{BLT-jde10, Cannon+1983, Uniqueness+1981, Liu+Triggiani+2011, Maarten+2016, Uniqueness+Stability+Yamamoto+2003, Uniqueness+Stability+Yamamoto, Uniqueness+Stability+2011} and \cite{Liu+Triggiani+2011, Uniqueness+Stability+Yamamoto+2003, Uniqueness+Stability+Yamamoto, Stability+Imanuvilov+2001, Uniqueness+Stability+2011, Stability+Yamamoto+1995}, respectively. Some of the numerical results may be found in \cite{BaoG+2011, Chen+2020, Hasanov+2016, Nguyen+2019, Sattari+2019} and the references cited therein. We refer to the monograph \cite{theory+Isakov+1990} for a complete account of the general theory on the ISP for the wave equation.
Recently, the field of stochastic inverse problems has been undergoing a rapid development and progressing to an area of intense activity. Stochastic inverse problems refer to inverse problems that involve uncertainties, which play a vital role in mathematical models to handle unpredictability of the environments and incomplete knowledge of the systems and measurements. Compared to deterministic counterparts, stochastic inverse problems have substantially more difficulties
on top of existing hurdles due to randomness and uncertainties. The inverse random source problems for the time-harmonic wave equations have been widely studied. In \cite{Devaney+1979}, it was shown that the correlation of the random source could be determined uniquely by the correlation of the random wave field. Effective computational models were developed in \cite{Bao+Chen+Li+2016, Bao+Chen+Li+2017, Bao+Chow+Li+Zhou+2014, Li+Chen+Li+2017, LiPJ+2011, Li+Yuan+2017, LW2021} for the time-harmonic stochastic acoustic and elastic wave equations, where the goal was to reconstruct the statistical properties of the random source from the boundary measurement of the radiated random wave field at multiple frequencies.
In this paper, we consider an ISP for the following initial-boundary value problem of the stochastic wave equation driven by the fractional Brownian motion (fBm):
\begin{equation}\label{directp}
\begin{cases}
u_{tt}(x,t)-\Delta u(x,t)=F(x,t), &\quad (x,t)\in{D}\times(0,T),\\
u(x,t)=0, & \quad (x,t)\in{\partial D}\times[0,T], \\
u(x,0)=u_t(x,0)=0, &\quad x\in\overline{D},
\end{cases}
\end{equation}
where $D\subset\mathbb R^d$ is a bounded domain with Lipschitz boundary $\partial D$ and the random source is assumed to take the form
\begin{equation}\label{rs}
F(x,t)=f(x)h(t)+g(x)\dot{B}^H(t).
\end{equation}
Here $f(x)$ and $g(x)$ are deterministic functions with compact supports contained in $D$, $h(t)$ is also a deterministic function, $B^H(t)$ is the fBm with the Hurst index $H\in(0,1)$, and $\dot{B}^H(t)$ can be roughly understood as the derivative of $B^H(t)$ with respect to the time $t$. When $H=\frac12$, the fBm reduces to the classical Brownian motion and $\dot{B}^H(t)$ becomes the white noise. Since the source $F(x,t)$ is a random field with low regularity, it is a distribution instead of a function. More precisely, it is shown in \cite{LW2019} that $F(\cdot,t)\in W^{H-1-\epsilon}(\mathbb{R})$ for any $\epsilon>0$. Clearly, we can see that $H-1-\epsilon<0$ as $H\in(0,1)$, and the smaller the Hurst index $H$ is, the lower the regularity of the source has. Given $F$, the direct problem is to determine the wave field $u$; the inverse problem is to recover $f$ and $g$ from the final-time data, i.e., $u(x, T), x\in D$.
So far, there have been considerable work done for the direct problems of the stochastic wave equation with different random sources. In \cite{Orsingher+1982}, the author studied randomly forced vibrations of a string driven by three Gaussian white noises. In \cite{Walsh+1984}, a white noise in both time and space was considered, and the existence and uniqueness of the solution were obtained for the stochastic wave equation. In \cite{Dalang+1998}, the authors examined the conditions on the well-posedness for a two-dimensional stochastic wave equation driven by a space-time Gaussian noise that is white in time but has a non-degenerate spatial covariance. In \cite{Caithamer+2005}, the author gave the upper and lower bounds on both the large and small derivations of several sup norms associated with the solution for a fractional Brownian noise. Some existence and uniqueness results can be found in \cite{Tindel+2007} for the one-dimensional stochastic wave equation driven by a two-parameter fBm. In \cite{
Erraoui+2008}, the existence and uniqueness of solutions were investigated for a class of hyperbolic stochastic partial differential equations driven by a space-time additive fractional Brownian sheet. In \cite{Tang+2020}, the authors showed several solutions of a stochastic wave
equation in the plane with an additive noise which is fractional in time and has a non-degenerate spatial covariance. In
\cite{numerical+1d+2017}, the existence of solutions was obtained and Newton's method was applied for nonlinear stochastic wave equations driven by one-dimensional white noise with respect to time. In \cite{Dengnumerical+2020},
a higher order approximation was proposed to solve the stochastic space fractional wave equation forced by an additive space-time Gaussian noise.
Compared with the direct problems for stochastic wave equations driven by fBm, there are few work for the inverse source problems for the stochastic wave equations driven by fBm. Recently, \cite{LW2019} considered an inverse random source problem for the Helmholtz equation driven by a fractional Gaussian field. The approach was further extended to solve the inverse random source problem for time-harmonic Maxwell's equations driven by a centered complex-valued Gaussian vector field with correlated components \cite{LW2020}. In \cite{Niu+2018} and \cite{FLW2020}, the ISP was studied for the stochastic fractional diffusion equation where the source is assumed to take the form of \eqref{rs} for $H=\frac{1}{2}$ and a general $H\in(0, 1)$, respectively. The goal is to reconstruct $f$ and $|g|$ from the final-time data.
In this work, we consider the ISP for the stochastic wave equation where the source is driven by the fBm. It contains three contributions. First, we show that the direct problem admits a well-defined mild solution which satisfies some stability estimate. Second, the uniqueness and instability are discussed for the inverse problem, which is to use the empirical expectation and correlation of the final-time data $u(x,T)$ to reconstruct $f$ and $|g|$ of the source term $F$. In \cite{FLW2020}, the ISP was studied for the time fractional diffusion equation driven by a fBm, where the parameter of the Caputo fractional derivative $\alpha$ is restricted to $0<\alpha\leq1$. Some of the results presented in \cite{FLW2020} are not valid any more for the hyperbolic equation where $\alpha=2$. In particular, the uniqueness of the inverse problem could not be guaranteed. Therefore it is worthwhile to investigate separately the ISP for the stochastic wave equation. Third, the numerical experiments are given to illustrate
how to obtain $f$ and $|g|$. Since the ISP is ill-posed, the truncation based regularization method is adopted to reconstruct $f$ and $|g|$. The numerical examples show that the method is effective to recover the source functions.
The paper is organized as follows. In Section 2, we introduce some preliminaries for the fBm and the mild solution of the stochastic wave equation. Section 3 is concerned with the well-posedness of the direct problem. Section 4 is devoted to the inverse problem. The uniqueness and instability are discussed. The numerical experiments are presented in Section 5. The paper concludes with some general remarks in Section 6.
\section{Preliminaries}
We begin with a brief introduction to fBms. The details can be found in \cite{Nualart+2006,FLW2020}. Let $(\Omega, \mathcal F, \mathbb{P})$ be a complete probability space, where $\Omega$ is a sample space, $\mathcal F$ is a $\sigma$-algebra on $\Omega$, and $\mathbb{P}$ is a probability measure on the measurable space $(\Omega, \mathcal F)$. For a random variable $X$, denote by $\mathbb{E}(X)$ and $\mathbb{V}(X)=\mathbb{E}(X-\mathbb{E}(X))^2 = \mathbb{E}(X^2)-(\mathbb{E}(X))^2$ the expectation and variance of $X$, respectively. For two random variables $X$ and $Y$, $\text{Cov}(X,Y)=\mathbb{E}[(X-\mathbb{E}(X))(Y-\mathbb{E}(Y))]$ stands for the covariance of $X$ and $Y$. In the sequel, the dependence of random variables on the sample $\omega\in\Omega$ will be omitted unless it is necessary to avoid confusion.
The one-dimensional fBm $B^H(t)$ is a centered Gaussian process, which satisfies $B^H(0)=0$ and is determined by the covariance function
\begin{equation*}
R(t,s)=\mathbb{E}[B^H(t)B^H(s)]=\frac{1}{2}\left(t^{2H}+s^{2H}-|t-s|^{2H}\right)
\end{equation*}
for any $s,t\ge0$. In particular, if $H=\frac12$, $B^H$ turns to be the standard Brownian motion, which is usually denoted by $W$ and has the covariance function $R(t,s)=t\wedge s$.
The increments of fBms satisfy
\begin{equation*}
\mathbb{E}\left[\left(B^H(t)-B^H(s)\right)\left(B^H(s)-B^H(r)\right)\right]=\frac{1}{2}\left[(t-r)^{2H}-(t-s)^{2H}-(s-r)^{2H}\right]
\end{equation*}
and
\begin{equation*}
\mathbb{E}\left[\left(B^H(t)-B^H(s)\right)^2\right]=(t-s)^{2H}
\end{equation*}
for any $0<r<s<t$. It indicates that the increments of $B^H(t)$ in disjoint intervals are linearly dependent except for the case $H=\frac{1}{2}$, and the increments are stationary since their moments depend only on the length of the interval.
Based on the moment estimates and the Kolmogorov continuity criterion, it holds
for any $\epsilon>0$ and $s,t\in[0,T]$ that
\[
|B^H(t)-B^H(s)|\le C|t-s|^{H-\epsilon}
\]
almost surely with constant $C$ depending on $\epsilon$ and $T$. It is clear to note that $H$ represents the regularity of $B^H$ and the trajectories of $B^H$ are $(H-\epsilon)$-H\"older continuous.
The fBm $B^H$ with $H\in(0,1)$ has a Wiener integral representation
\[
B^H(t)=\int_0^tK_H(t,s)dW(s),
\]
where $K_H$ is a square integrable kernel and $W$ is the standard Brownian motion.
For a fixed interval $[0,T]$, denote by $\mathcal{E}$ the space of step functions on $[0,T]$ and by $\mathcal{H}$ the closure of $\mathcal{E}$ with respect to the product
\[
\langle\chi_{[0, t]}, \chi_{[0, s]}\rangle_{\mathcal H}=R(t, s),
\]
where $\chi_{[0, t]}$ and $\chi_{[0, s]}$ are the characteristic functions. For $\psi(t),\phi(t)\in\mathcal{H}$, it follows from the It\^{o} isometry that
(1) if $H\in(0,\frac12)$,
\begin{eqnarray}\label{E012}
&&\mathbb{E}\left[\int_0^t\psi(s)dB^H(s)\int_0^t\phi(s)dB^H(s)\right]\nonumber\\
&=&\int_0^t\bigg\{c_H\Big[\big(\frac{t}{s}\big)^{H-\frac12}(t-s)^{H-\frac12}-(H-\frac12)s^{\frac12-H}\int_s^tu^{H-\frac32}(u-s)^{H-\frac12}du\Big]\psi(s)\nonumber\\
&& +\int_s^t(\psi(u)-\psi(s))c_H\big(\frac{u}{s}\big)^{H-\frac12}(u-s)^{H-\frac32}du\bigg\}\nonumber\\
&&\times \bigg\{c_H\left[\big(\frac{t}{s}\big)^{H-\frac12}(t-s)^{H-\frac12}-(H-\frac12)s^{\frac12-H}\int_s^tu^{H-\frac32}(u-s)^{H-\frac12}du\right]\phi(s)\nonumber\\
&& +\int_s^t(\phi(u)-\phi(s))c_H\big(\frac{u}{s}\big)^{H-\frac12}(u-s)^{H-\frac32}du\bigg\}ds,
\end{eqnarray}
where $c_H=\left(\frac{2H}{(1-2H)\beta(1-2H,H+\frac12)}\right)^{\frac12}$ and $\beta(p,q)=\int_0^1t^{p-1}(1-t)^{q-1}dt$;
(2) if $H=\frac12$,
\begin{align}\label{E12}
\mathbb{E}\left[\int_0^t\psi(s)dB^H(s)\int_0^t\phi(s)dB^H(s)\right]=\mathbb{E}\left[\int_0^t\psi(s)dW(s)\int_0^t\phi(s)dW(s)\right]=\int_0^t\psi(s)\phi(s)ds;
\end{align}
(3) if $H\in(\frac12,1)$,
\begin{align}\label{E121}
\mathbb{E}\left[\int_0^t\psi(s)dB^H(s)\int_0^t\phi(s)dB^H(s)\right]=\alpha_H\int_0^t\int_0^t\psi(r)\phi(u)|r-u|^{2H-2}dudr,
\end{align}
where $\alpha_H=H(2H-1)$.
Since $\dot{B}^H$ is a distribution instead of a classical function, the stochastic wave equation in \eqref{directp} does not hold pointwisely; it should be interpreted as an integral equation and its mild solution is defined as follows.
\begin{definition}
A stochastic process $u$ taking values in $L^2(D)$ is called a mild solution of \eqref{directp} if
\[
u(x,t)=\int_0^t\sin\left((t-\tau)(-\Delta)^{\frac12}\right)(-\Delta)^{-\frac12}f(x)h(\tau)d\tau+\int_0^t\sin\left((t-\tau)(-\Delta)^{\frac12}\right)(-\Delta)^{-\frac12}g(x)dB^H(\tau)
\]
is well-defined almost surely.
\end{definition}
It is known that the operator $-\Delta$ with the homogeneous Dirichlet boundary condition has an eigensystem $\{\lambda_k, \varphi_k\}_{k=1}^{\infty}$, where the eigenvalues satisfy $0<\lambda_1\leq\lambda_2\leq\cdots\leq\lambda_k\leq\cdots$ with $
\lambda_k\to\infty$ as $k\to\infty$, and the eigen-functions $\{\varphi_k\}_{k=1}^{\infty}$ form a complete and orthonormal basis for $L^2({D})$. For any function $v$ in $L^2(D)$, it can be written as
\[
v(x)=\sum_{k=1}^{\infty}v_k\varphi_k(x), \quad v_k=(v,\varphi_k)_{L^2(D)}=\int_D v(x)\varphi_k(x)dx.
\]
Hence, if $u\in L^2(D)$ is a mild solution of \eqref{directp}, we have
\begin{equation}\label{Dsolution}
u(\cdot,t)=\sum_{k=1}^{\infty}u_k(t)\varphi_k,
\end{equation}
where
\begin{eqnarray}\label{Ik}
u_k(t)&=&(u(\cdot,t),\varphi_k)_{L^2(D)}\nonumber\\
&=&f_k\int_0^t\frac{\sin((t-\tau)\sqrt{\lambda_k})}{\sqrt{\lambda_k}}h(\tau)d\tau+g_k\int_0^t\frac{\sin((t-\tau)\sqrt{\lambda_k})}{\sqrt{\lambda_k}}dB^H(\tau)\nonumber\\
&=:&I_{k,1}(t)+I_{k,2}(t).
\end{eqnarray}
Here $f_k=(f,\varphi_k)_{L^2({D})}$, $g_k=(g,\varphi_k)_{L^2({D})}$, and $u_k(t)$ satisfies the stochastic differential equation
\begin{equation}\label{SODE}
\begin{cases}
u_k''(t)+ \lambda_k u_k(t)=f_kh(t)+g_k\dot{B}^H(t), \quad t\in(0,T),\\
u_k(0)=u_k'(0)=0.
\end{cases}
\end{equation}
In particular, if $g=0$, the stochastic differential equation \eqref{SODE} reduces to the deterministic differential
equation
\begin{equation*}
\begin{cases}
u_k''(t)+ \lambda_k u_k(t)=f_kh(t), \quad t\in(0,T),\\
u_k(0)=u_k'(0)=0,
\end{cases}
\end{equation*}
which has the solution given as $I_{k,1}(t)$ in (\ref{Ik}).
\section{The direct problem}
In this section, we discuss the well-posedness of the direct problem. It is only necessary to address the stability since the existence and the uniqueness of the solution has already been considered (cf. \cite{Orsingher+1982,Walsh+1984,Tindel+2007}). We show that the mild solution \eqref{Dsolution} of the initial-boundary value problem \eqref{directp} is well-defined under the following assumptions.
\begin{assumption}\label{assumption}
Let $H\in(0,1)$ and $f,g\in L^2({D})$ with $\|g\|_{L^2(D)}\neq0$. Assume in addition that $h\in L^{\infty}(0,T)$ is a nonnegative function and its support has a positive measure.
\end{assumption}
It is easy to note that the mild solution (\ref{Dsolution}) satisfies
\begin{eqnarray*}
\|u(\cdot,t)\|^2_{L^2({D})} &=&\left\|\sum_{k=1}^{\infty}(I_{k,1}(t)+I_{k,2}(t))\varphi_k(\cdot)
\right\|_{L^2({D})}^2=\sum_{k=1}^{\infty}(I_{k,1}(t)+I_{k,2}(t))^2 \\
&\lesssim& \sum_{k=1}^{\infty} I^2_{k,1}(t)+\sum_{k=1}^{\infty}I_{k,2}^2(t),
\end{eqnarray*}
which gives
\begin{eqnarray}\label{ex}
\mathbb{E}\left[\|u\|^2_{L^2({D}\times[0,T])}\right] & =& \mathbb{E}\left[\int_0^T \|u(\cdot,t)\|^2_{L^2({D})}dt\right]\nonumber\\
&\lesssim&\mathbb{E}\bigg[\int_0^T \bigg(\sum_{k=1}^{\infty}I^2_{k,1}(t)+\sum_{k=1}^{\infty}I_{k,2}^2(t)\bigg)dt\bigg]\nonumber\\
&=&\int_0^T \bigg(\sum_{k=1}^{\infty}I^2_{k,1}(t)\bigg)dt+\mathbb{E}\bigg[\int_0^T \bigg(\sum_{k=1}^{\infty}I_{k,2}^2(t)\bigg)dt\bigg]\nonumber\\
&=&\sum_{k=1}^{\infty}\|I_{k,1}\|^2_{L^2(0,T)}+\int_0^T \bigg(\sum_{k=1}^{\infty}\mathbb{E}[I_{k,2}^2(t)]
\bigg)dt\nonumber\\
&=:&S_1+S_2.
\end{eqnarray}
Hereinafter, $a\lesssim b$ stands for $a\leqslant Cb$, where $C>0$ is a constant and its specific value is not required but should be clear from the context.
Let $G_{k}(t)=\sin(\sqrt{\lambda_k}t)/\sqrt{\lambda_k}$. Then
\begin{eqnarray}\label{Ik11}
\|I_{k,1}\|_{L^2(0,T)}&=&\left(\int_0^T\left|f_k\int_0^tG_k(\tau)h(t-\tau)d\tau\right|^2dt\right)^{\frac12}\notag\\
&\leq&\left(\int_0^T|f_k|^2\|h\|_{L^{\infty}(0,T)}^2\left(\int_0^T|G_k(\tau)|d\tau\right)^2dt\right)^{\frac12}\notag\\
&\leq& T^{\frac12}|f_k|\|G_{k}\|_{L^1(0,T)}\|h\|_{L^\infty(0,T)}.
\end{eqnarray}
A simple calculation gives
\begin{equation}\label{G}
\|G_{k}\|_{L^1(0,T)}=\int_0^T\left|\frac{\sin(\sqrt{\lambda_k}t)}{\sqrt{\lambda_k}}\right|dt\leq\int_0^T tdt=\frac{T^2}{2}.
\end{equation}
Combining (\ref{Ik11}) and (\ref{G}), we obtain
\begin{equation}\label{S1}
S_1=\sum_{k=1}^{\infty}\|I_{k,1}\|^2_{L^2(0,T)}
\leq\frac{T^{5}}{4}\sum_{k=1}^{\infty}|f_k|^2\|h\|^2_{L^\infty(0,T)}
=\frac{T^{5}}{4}\|h\|^2_{L^\infty(0,T)}\|f\|^2_{L^2({D})}.
\end{equation}
Next is to estimate $S_2$. It follows from \eqref{Ik} that
\begin{align}\label{exIk2}
\mathbb{E}[I_{k,2}^2(t)]=g_k^2\mathbb{E}\Big[\Big(\int_0^t\frac{\sin(\sqrt{\lambda_k}(t-\tau))}{\sqrt{\lambda_k}}dB^H(\tau)\Big)^2\Big].
\end{align}
Below, we discuss separately the cases $H=\frac{1}{2}$, $H\in(0,\frac{1}{2})$, and $H\in(\frac{1}{2},1)$ since the covariance operator of $B^H$ takes different forms in these three cases.
For the case $H=\frac{1}{2}$, it follows from It\^o's isometry \eqref{E12} that
\begin{align*}
\mathbb{E}\Big[\Big(\int_0^t\frac{\sin(\sqrt{\lambda_k}(t-\tau))}{\sqrt{\lambda_k}}dB^{\frac{1}{2}}(\tau)\Big)^2\Big]
=\int_0^t\frac{\sin^2(\sqrt{\lambda_k}(t-\tau))}{\lambda_k}d\tau\leq\int_0^t(t-\tau)^2d\tau=\frac{t^3}{3}.
\end{align*}
For the case $H\in(0,\frac{1}{2})$, we have from (\ref{E012}) that
\begin{eqnarray}\label{Ee012}
&&\mathbb{E}\Big[
\Big(\int_0^t\frac{\sin(\sqrt{\lambda_k}(t-\tau))}{\sqrt{\lambda_k}}dB^H(\tau)\Big)^2\Big]\nonumber\\
&=&\int_0^t \bigg\{c_H\Big[\big(\frac{t}{\tau}\big)^{H-\frac12}(t-\tau)^{H-\frac12}-(H-\frac12)\tau^{\frac12-H}\int_\tau^tu^{H-\frac32}(u-\tau)^{H-\frac12}du\Big]\frac{\sin(\sqrt{\lambda_k}(t-\tau))}{\sqrt{\lambda_k}}\nonumber\\
&&+\int_\tau^t\Big(\frac{\sin(\sqrt{\lambda_k}(t-u))}{\sqrt{\lambda_k}}-\frac{\sin(\sqrt{\lambda_k}(t-\tau))}{\sqrt{\lambda_k}}\Big)c_H\big(\frac{u}{\tau}\big)^{H-\frac12}(H-\frac12)(u-\tau)^{H-\frac32}du\bigg\}^2d\tau\nonumber\\
&\lesssim&\int_0^t\Big[\big(\frac{t}{\tau}\big)^{H-\frac12}(t-\tau)^{H-\frac12}\frac{\sin(\sqrt{\lambda_k}(t-\tau))}{\sqrt{\lambda_k}}\Big]^2d\tau\notag\\
&&+\int_0^t\Big[\tau^{\frac12-H}\Big(\int_\tau^tu^{H-\frac32}(u-\tau)^{H-\frac12}du\Big)\frac{\sin(\sqrt{\lambda_k}(t-\tau))}{\sqrt{\lambda_k}}\Big]^2d\tau\nonumber\\
&&+\int_0^t\Big[\int_\tau^t\Big(\frac{\sin(\sqrt{\lambda_k}(t-u))}{\sqrt{\lambda_k}}-\frac{\sin(\sqrt{\lambda_k}(t-\tau))}{\sqrt{\lambda_k}}\Big)\big(\frac{u}{\tau}\big)^{H-\frac12}(u-\tau)^{H-\frac32}du\Big]^2d\tau\nonumber\\
&=:&I_1(t)+I_2(t)+I_3(t).
\end{eqnarray}
A simple calculation gives
\begin{eqnarray}\label{I1}
I_1(t)&=&\int_0^t\big(\frac{t}{\tau}\big)^{2H-1}(t-\tau)^{2H-1}\frac{\sin^2(\sqrt{\lambda_k}(t-\tau))}{\lambda_k}d\tau\nonumber\\
&\leq&\int_0^t(\frac{t}{\tau})^{2H-1}(t-\tau)^{2H-1}(t-\tau)^2d\tau\nonumber\\
&\leq&\int_0^t(t-\tau)^{2H+1}d\tau=\frac{t^{2H+2}}{2H+2}.
\end{eqnarray}
Similarly, we have
\begin{align}\label{I21}
I_2(t)&\lesssim\int_0^t\tau^{1-2H}(t-\tau)^2\Big(\int_\tau^tu^{H-\frac32}(u-\tau)^{H-\frac12}du\Big)^2d\tau.
\end{align}
Using the binomial expansion leads to
\begin{eqnarray*}
\int_\tau^tu^{H-\frac32}(u-\tau)^{H-\frac12}du
&=&\int_\tau^t u^{2H-2}\bigg[\sum_{n=0}^{\infty}
\begin{pmatrix}
H-\frac12\\
n
\end{pmatrix}
\big(-\frac{\tau}{u}\big)^n\bigg]du \\
&=&\sum_{n=0}^{\infty}
\begin{pmatrix}
H-\frac12\\
n
\end{pmatrix}(-1)^n\tau^n\left(\frac{t^{2H-1-n}-\tau^{2H-1-n}}{2H-1-n}\right)\\
&\leq&(t^{2H-1}-\tau^{2H-1})\sum_{n=0}^{\infty}
\begin{pmatrix}
H-\frac12\\
n
\end{pmatrix}\frac{(-1)^n}{2H-1-n}\\
&\lesssim& t^{2H-1}-\tau^{2H-1}.
\end{eqnarray*}
Combing the above estimates, we obtain from (\ref{I21}) that
\begin{eqnarray}\label{I2}
I_2(t)&\lesssim&\int_0^t\tau^{1-2H}(t-\tau)^2\left(t^{2H-1}-\tau^{2H-1}\right)^2d\tau\nonumber\\
&\lesssim&\int_0^t\tau^{1-2H}(t-\tau)^2\left(t^{4H-2}+\tau^{4H-2}\right)d\tau\nonumber\\
&\lesssim&\int_0^t\tau^{2H-1}(t-\tau)^2d\tau\nonumber\\
&\leq&\int_0^t\tau^{2H-1}t^2d\tau
\lesssim t^{2H+2}.
\end{eqnarray}
By the mean value theorem, there holds
\begin{eqnarray}\label{I3}
I_3(t)&\leq&\int_0^t\Big[\int_\tau^t|\tau-u|(\frac{u}{\tau})^{H-\frac12}(u-\tau)^{H-\frac32}du\Big]^2d\tau\nonumber\\
&\leq&\int_0^t\Big[\int_\tau^t(u-\tau)^{H-\frac12}du\Big]^2d\tau \lesssim t^{2H+2}.
\end{eqnarray}
Substituting (\ref{I1}), (\ref{I2}) and (\ref{I3}) into (\ref{Ee012}), we obtain for $H\in(0,\frac12)$ that
\begin{align*}
\mathbb{E}\Big[
\Big(\int_0^t\frac{\sin(\sqrt{\lambda_k}(t-\tau))}{\sqrt{\lambda_k}}dB^H(\tau)\Big)^2\Big]\lesssim t^{2H+2}.
\end{align*}
For the case $H\in(\frac{1}{2},1)$, we have from \eqref{E121} that
\begin{eqnarray*}
&&\mathbb{E}\Big[
\Big(\int_0^t\frac{\sin(\sqrt{\lambda_k}(t-\tau))}{\sqrt{\lambda_k}}dB^H(\tau)\Big)^2\Big]\\
&=&\alpha_H\int_0^t\int_0^t\frac{\sin(\sqrt{\lambda_k}(t-p))}{\sqrt{\lambda_k}}\frac{\sin(\sqrt{\lambda_k}(t-q))}{\sqrt{\lambda_k}}|p-q|^{2H-2}dpdq\\
&\leq&\alpha_H\int_0^t\int_0^t(t-p)(t-q)|p-q|^{2H-2}dpdq\\
&=&2\alpha_H\int_0^t\int_q^tpq(p-q)^{2H-2}dpdq,\\
&=&2\alpha_H\int_0^tq\frac{(t-q)^{2H-1}}{2H-1}\big(t-\frac{t-q}{2H}\big)dq\\
&\lesssim& t\int_0^tq\frac{(t-q)^{2H-1}}{2H-1}dq
\lesssim t^{2H+2}.
\end{eqnarray*}
Combing the above estimates, we obtain for any $H\in (0,1)$ that
\begin{align}\label{eq:EeH01}
\mathbb{E}\Big[\Big(\int_0^t\frac{\sin(\sqrt{\lambda_k}(t-\tau))}{\sqrt{\lambda_k}}dB^H(\tau)\Big)^2\Big]
\lesssim t^{2H+2},
\end{align}
which implies the following stability estimate for the mild solution (\ref{Dsolution}).
\begin{theorem}\label{thereom1}
Let Assumption \ref{assumption} hold. Then the stochastic process $u$
given in (\ref{Dsolution}) satisfies
\begin{equation*}
\mathbb{E}[\|u\|^2_{L^2({D}\times[0,T])}]\lesssim\frac{T^5}{4}\|h\|^{2}_{L^\infty(0,T)}\|f\|^
{2}_{L^2({D})}+\frac{T^{2H+3}}{2H+3}\|g\|^{2}_{L^2({D})}.
\end{equation*}
\end{theorem}
\begin{proof}
The proof follows easily from (\ref{ex}), (\ref{S1}), (\ref{exIk2}), and
(\ref{eq:EeH01}).
\end{proof}
\begin{remark}
The above stability estimate is consistent with \cite[Theorem 3.1]{FLW2020}. The stochastic wave equation (\ref{directp}) can be viewed as the case $\alpha=2$ of the problem (2.1) in \cite{FLW2020}, where
\[
E_{2,2}(z^2)=\frac{\sinh(z)}{z}
\]
and
\[
(t-\tau)^{\alpha-1}E_{\alpha,\alpha}\big(-(t-\tau)^\alpha\lambda_k\big)\big|_{\alpha=2}=(t-\tau)\frac{\sinh\sqrt{-(t-\tau)^2\lambda_k}}{\sqrt{-(t-\tau)^2\lambda_k}}=\frac{\sin(\sqrt{\lambda_k}(t-\tau))}{\sqrt{\lambda_k}}.
\]
\end{remark}
\section{The inverse problem}
In this section, we consider the inverse problem of reconstructing $f$ and $|g|$ from the empirical expectation and correlation of the final-time data $u(x,T)$. More specifically, the data is assumed to be given by
\[
u_k(T):=(u(\cdot,T),\varphi_k(\cdot))_{L^2(D)}.
\]
We discuss the uniqueness and the issue of instability for the inverse problem, separately.
\subsection{Uniqueness}
It follows from (\ref{Dsolution}), (\ref{Ik}) and \eqref{exIk2} that
\begin{equation}\label{Iexp}
\mathbb{E}(u_k(T))=f_k\int_0^Th(\tau)\frac{\sin(\sqrt{\lambda_k}(T-\tau))}{\sqrt{\lambda_k}}d\tau
\end{equation}
and
\begin{align}\label{Icvar}
\text{Cov}(u_k(T),u_l(T))
=g_kg_l
\mathbb{E}\Big[\int_0^T\frac{\sin(\sqrt{\lambda_k}(T-\tau))}{\sqrt{\lambda_k}}dB^H(\tau)\int_0^T\frac{\sin(\sqrt{\lambda_l}(T-\tau))}{\sqrt{\lambda_l}}dB^H(\tau)\Big].
\end{align}
First, we consider the reconstruction of $f$. Let
\begin{align*}
h_k=\int_0^Th(\tau)\sin(\sqrt{\lambda_k}(T-\tau))d\tau=\int_0^Th(T-\tau)\sin(\sqrt{\lambda_k}\tau)d\tau.
\end{align*}
By (\ref{Iexp}), it is impossible to compute $f_k$ if $h_k=0$, which indicates nonuniqueness to reconstruct $f$. It can be seen that $h_k$ may be zero for some $h(t)$ satisfying Assumption \ref{assumption}. Clearly, $\sin(\sqrt{\lambda_k}t)$ is an oscillatory function; the larger the $\lambda_k$ is, the more oscillatory the function $\sin(\sqrt{\lambda_k}t)$ is. Hence the uniqueness cannot be guaranteed to reconstruct $f$ for a general $h$. Below we consider two special cases for the function $h$ where it is sufficient to ensure the uniqueness for the inverse problem.
Case $1$: $h(t)\equiv C>0$ is a constant function and $D=(0,\pi)$. In this case, it is clear that the eigen-system of the Laplacian operator $-\Delta$ is
\[
\lambda_k=k^2, \quad \varphi_k=\sqrt{\frac2\pi}\sin(kx), \quad k=1,2,...
\]
A simple calculation yields
\begin{align*}
h_k~=C\int_0^T\sin(k\tau)d\tau~=\frac Ck(1-\cos(kT)).
\end{align*}
Note that the final time $T$ can be chosen arbitrarily. For example, we can choose $T$ as a rational number so that $h_k>0$ for any $k\in \mathbb{N}$.
Case $2$: $h(t)$ is a continuous and monotonously increasing function and $D=(0,\pi)$. Divide the time interval $(0,T]$ into $n+1$ subintervals as $(t_0,t_1],\dots$, $(t_i,t_{i+1}],\dots$, $(t_n,t_{n+1}]$, where $kt_i=i\pi$, $i=0,1,2,...,n$, and $t_0=0$, $t_{n+1}=T$. Since $h(t)$ is continuous and $\sin(kt)$ is integrable and does not change sign on each subinterval $(t_{i},t_{i+1}]$, by the mean value theorem for the definite integral, there exists $\xi_i\in(t_{i},t_{i+1}]$ such that
\begin{eqnarray*}
h_k&=&\int_0^Th(T-\tau)\sin(k\tau)d\tau=\sum_{i=0}^{n}h(T-\xi_i)\int_{t_i}^{t_i+1}\sin(k\tau)d\tau\nonumber\\
&=&\sum_{i=0}^{n-1}h(T-\xi_i)\frac{2\cdot(-1)^i}{k}+h(T-\xi_n)\Big(\frac{(-1)^n-\cos(kT)}{k}\Big).
\end{eqnarray*}
If $n$ is even, since $h(t)$ is monotonously increasing,
\begin{align*}
h_k=\sum_{i=2m,m=0}^{\frac{n-2}{2}}(h(T-\xi_i)-h(T-\xi_{i+1}))\frac2k+h(T-\xi_n)\Big(\frac{1-\cos(kT)}{k}\Big)>0.
\end{align*}
If $n$ is odd, similarly, we have
\begin{align*}
h_k=\sum_{i=2m,m=0}^{\frac{n-3}{2}}(h(T-\xi_i)-h(T-\xi_{i+1}))\frac2k+h(T-\xi_{n-1})\frac{2}{k}+h(T-\xi_n)\Big(\frac{-1-\cos(kT)}{k}\Big)>0.
\end{align*}
Here, the final time $T$ is chosen as a rational number. If $h(t)$ is strictly monotonously increasing, $T$ can be any real number.
\begin{remark}
Although we only consider the one-dimensional case $D=(0,\pi)$, in fact, if $D$ is a bounded domain in $\mathbb{R}^d, d\in\{1,2,3\}$, then the $j$-th eigenvalue $\lambda_j$ of the homogeneous Dirichlet boundary problem for the Laplacian operator $-\Delta$ in $D$ satisfies
\begin{equation*}
c_0 j^{\frac2d}\leq\lambda_j\leq c_1 j^{\frac2d},
\end{equation*}
where $j\in \mathbb{N}$, and the constants $c_0, c_1$ are independent of the index $j$ \cite{Dengnumerical+2020,Strauss+2008}. Therefore, if we choose $T$ properly, the uniqueness can still be obtained to recover $f$ when the function $h(t)$ belongs to the two cases considered above.
\end{remark}
\begin{remark}
From (\ref{Iexp}), if $f_k\neq0, D=(0,\pi)$, then there holds
\begin{equation}\label{Iexp1}
\frac{k\mathbb{E}(u_k(T))}{f_k}=\int_0^Th(T-\tau)\sin(k\tau)d\tau.
\end{equation}
The right hand side of (\ref{Iexp1}) can be understood as the sine transform of $h(T-\tau)$. Alternatively, (\ref{Iexp1}) can be written as
\begin{equation}\label{Iexp2}
k\mathbb{E}(u_k(T))=\sqrt{\frac2\pi}\int_0^T\int_0^\pi h(T-\tau)f(x)\sin(kx)\sin(k\tau)dxd\tau.
\end{equation}
Therefore, $h(T-t)f(x)$ can be reconstructed by taking the inverse two-dimensional sine transform on both sides of (\ref{Iexp2}) \cite{Bao+Chow+Li+Zhou+2014}.
\end{remark}
Next, we consider the reconstruction of $|g|$. Let
\begin{align}\label{guniqbm}
E_{kl}:=\mathbb{E}\Big[\int_0^T\frac{\sin(\sqrt{\lambda_k}(T-\tau))}{\sqrt{\lambda_k}}dB^H(\tau)\int_0^T\frac{\sin(\sqrt{\lambda_l}(T-\tau))}{\sqrt{\lambda_l}}dB^H(\tau)\Big].
\end{align}
Similarly, since $\sin(\sqrt{\lambda_k}t)$ is an oscillatory function, and the larger the $\lambda_k$ is the more oscillatory the function $\sin(\sqrt{\lambda_k}t)$ is, we can see from (\ref{guniqbm}) that it is difficult to guarantee $E_{kl}\neq0$, particularly for a general $H\in(0, 1)$. Below, we only consider the case $H=\frac{1}{2}$ and $D=(0,\pi) $, where the uniqueness can be obtained to recover $|g|$.
By (\ref{E12}), a straightforward calculation gives
\begin{eqnarray}\label{Ekl12}
E_{kl}&=&\int_0^T\frac{\sin(\sqrt{\lambda_k}(T-\tau))}{\sqrt{\lambda_k}}\frac{\sin(\sqrt{\lambda_l}(T-\tau))}{\sqrt{\lambda_l}}d\tau\nonumber\\
&=&\begin{cases} \frac{1}{2\lambda_k}\left(T-\frac{\sin(2\sqrt{\lambda_k}T)}{2\sqrt{\lambda_k}}\right)>0,\quad & k=l,\\
\frac{1}{\sqrt{\lambda_k\lambda_l}}\bigg(-\frac{\sin((\sqrt{\lambda_k}+\sqrt{\lambda_l})T)}{2(\sqrt{\lambda_k}+\sqrt{\lambda_l})}+\frac{\sin((\sqrt{\lambda_k}-\sqrt{\lambda_l})T)}{2(\sqrt{\lambda_k}-\sqrt{\lambda_l})}\bigg),\quad & k\neq l.
\end{cases}
\end{eqnarray}
When $D=(0,\pi)$, and $k\neq l$, we have
\begin{align*}
E_{kl}=\frac{1}{kl}\bigg(-\frac{\sin((k+l)T)}{2(k+l)}+\frac{\sin((k-l)T)}{2(k-l)}\bigg).
\end{align*}
Without loss of generality, let $k>l$, then it holds that $E_{kl}\neq0$ for any algebraic number $T$. Otherwise, if $E_{kl}=0$, then
\[
\frac{\sin((k-l)T)}{2(k-l)}-\frac{\sin((k+l)T)}{2(k+l)}=0,
\]
which implies
\begin{equation*}
(k+l)e^{i(k-l)T}-(k+l)e^{-i(k-l)T}-(k-l)e^{i(k+l)T}+(k-l)e^{-i(k+l)T}=0.
\end{equation*}
Since $\{i(k-l)T, -i(k-l)T, i(k+l)T, -i(k+l)T\}$ is a set of distinct algebraic numbers, by the Lindemann--Weierstrass theorem (cf. \cite[Theorem 1.4]{ab90}), we have for any non-zero algebraic numbers $c_1, c_2, c_3, c_4$ that
\begin{equation*}
c_1e^{i(k-l)T}+c_2e^{-i(k-l)T}+c_3e^{i(k+l)T}+c_4e^{-i(k+l)T}\neq0,
\end{equation*}
which leads to a contradiction. Therefore $E_{kl}\neq0$ for any $k\neq l\in \mathbb{N}$. Combining with (\ref{Ekl12}), we deduce for any $k,l\in\mathbb{N}$ that $E_{kl}\neq0$ for any algebraic number $T$.
The following results concern the statement of uniqueness for the inverse problem.
\begin{theorem}\label{thereom2}
Let Assumption \ref{assumption} hold.
\begin{itemize}
\item [(1)]If $h(t)$ is monotonously increasing and $T$ is a rational number, or if $h(t)$ is strictly monotonously increasing and $T$ is a real number. Then $f$ can be uniquely determined by the data set
$\{\mathbb{E}(u_k(T)): k \in\mathbb{N}\}$.
\item [(2)]If T is any algebraic number, then the source term $g$ up to sign, i.e. $\pm g$, can be uniquely determined by the data set $\{\text{Cov}(u_k(T),u_l(T)): k,l \in \mathbb{N}\}$.
\end{itemize}
\end{theorem}
\subsection{Instability}
In this subsection, we demonstrate the inverse problem is unstable to recover $f$ and $|g|$. First, it is clear to note that
\begin{equation*}
\left|\int_0^Th(\tau)\frac{\sin(\sqrt{\lambda_k}(T-\tau))}{\sqrt{\lambda_k}}d\tau\right|\leq\frac{1}{\sqrt{\lambda_k}}\int_0^Th(\tau)d\tau\rightarrow0 ~ \text{as } k\rightarrow\infty,
\end{equation*}
which shows that it is unstable to recover $f$ due to (\ref{Iexp}).
Next, we discuss the instability of recovering $g^2_k$, which is equivalent to the instability of recovering $|g|$. We discuss the three different cases $H=\frac{1}{2}, H\in(0,\frac{1}{2})$ and $H\in(\frac{1}{2},1)$, separately.
For the case $H=\frac{1}{2}$, by (\ref{E12}), there holds
\begin{align*}
\mathbb{E}\Big[
\Big(\int_0^T\frac{\sin(\sqrt{\lambda_k}(T-\tau))}{\sqrt{\lambda_k}}dB^{\frac{1}{2}}(\tau)\Big)^2\Big]\leq\frac{T}{\lambda_k}.
\end{align*}
For the case $H\in(0,\frac{1}{2})$, we consider (\ref{Ee012}) with $t=T$ and the estimate of $I_j(T), j=1,2,3.$
For $I_1(T)$, a simple calculation yields
\begin{eqnarray*}
I_1(T)&=&\int_0^T(\frac{T}{\tau})^{2H-1}(T-\tau)^{2H-1}\frac{\sin^2(\sqrt{\lambda_k}(T-\tau))}{\lambda_k}d\tau\nonumber\\
&\leq&\frac{1}{\lambda_k}\int_0^T(\frac{T}{\tau})^{2H-1}(T-\tau)^{2H-1}d\tau
\leq\frac{1}{\lambda_k}\frac{T^{2H}}{2H}.
\end{eqnarray*}
About $I_2(T)$, we have
\begin{eqnarray*}
I_2(T)&\lesssim&\frac{1}{\lambda_k}\int_0^T\tau^{1-2H}\Big(\int_\tau^Tu^{H-\frac32}(u-\tau)^{H-\frac12}du\Big)^2d\tau\nonumber\\
&\lesssim&\frac{1}{\lambda_k}\int_0^T\tau^{1-2H}(T^{2H-1}-\tau^{2H-1})^2d\tau\nonumber\\
&\lesssim&\frac{1}{\lambda_k}\int_0^T\tau^{1-2H}(T^{4H-2}+\tau^{4H-2})d\tau\nonumber\\
&\lesssim&\frac{1}{\lambda_k}\int_0^T\tau^{2H-1}d\tau
\lesssim\frac{1}{\lambda_k}\frac{T^{2H}}{2H}.
\end{eqnarray*}
About $I_3(T)$, there holds
\begin{eqnarray*}
I_3(T)&=&\int_0^T\Big[\int_\tau^T(\frac{\sin(\sqrt{\lambda_k}(T-u))}{\sqrt{\lambda_k}}-\frac{\sin(\sqrt{\lambda_k}(T-\tau))}{\sqrt{\lambda_k}})(\frac{u}{\tau})^{H-\frac12}(u-\tau)^{H-\frac32}du\Big]^2d\tau\\
&\lesssim&\int_0^{t_*}\Big[\int_\tau^{t_*}(\frac{\sin(\sqrt{\lambda_k}(T-u))}{\sqrt{\lambda_k}}-\frac{\sin(\sqrt{\lambda_k}(T-\tau))}{\sqrt{\lambda_k}})(\frac{u}{\tau})^{H-\frac12}(u-\tau)^{H-\frac32}du\Big]^2d\tau\\
&&+\int_0^{t_*}\Big[\int_{t_*}^T(\frac{\sin(\sqrt{\lambda_k}(T-u))}{\sqrt{\lambda_k}}-\frac{\sin(\sqrt{\lambda_k}(T-\tau))}{\sqrt{\lambda_k}})(\frac{u}{\tau})^{H-\frac12}(u-\tau)^{H-\frac32}du\Big]^2d\tau\\
&&+\int_{t_*}^T\Big[\int_\tau^{\tau+t_*}(\frac{\sin(\sqrt{\lambda_k}(T-u))}{\sqrt{\lambda_k}}-\frac{\sin(\sqrt{\lambda_k}(T-\tau))}{\sqrt{\lambda_k}})(\frac{u}{\tau})^{H-\frac12}(u-\tau)^{H-\frac32}du\Big]^2d\tau\\
&& +\int_{t_*}^T\Big[\int_{\tau+t_*}^T(\frac{\sin(\sqrt{\lambda_k}(T-u))}{\sqrt{\lambda_k}}-\frac{\sin(\sqrt{\lambda_k}(T-\tau))}{\sqrt{\lambda_k}})(\frac{u}{\tau})^{H-\frac12}(u-\tau)^{H-\frac32}du\Big]^2d\tau\\
&=:&K_1+K_2+K_3+K_4.
\end{eqnarray*}
For $K_1$ and $K_3$, we have from the differential mean value theorem that
\begin{eqnarray*}
K_1&\lesssim&\int_0^{t_*}\Big[\int_\tau^{t_*}(\frac{u}{\tau})^{H-\frac12}(u-\tau)^{H-\frac12}du\Big]^2d\tau\\
&\lesssim&\int_0^{t_*}\Big[\int_\tau^{t_*}(u-\tau)^{H-\frac12}du\big]^2d\tau\lesssim\int_0^{t_*}(t_*-\tau)^{2H+1}d\tau\lesssim t_*^{2H+2}
\end{eqnarray*}
and
\begin{eqnarray*}
K_3 &\lesssim&\int_{t_*}^T\Big[\int_\tau^{\tau+t_*}(u-\tau)^{H-\frac12}du\Big]^2d\tau\\
&\lesssim&\int_{t_*}^Tt_*^{2H+1}d\tau\lesssim t_*^{2H+1}.
\end{eqnarray*}
For $K_2$ and $K_4$, it follows from straightforward calculations that
\begin{eqnarray*}
K_2&\lesssim&\frac{1}{\lambda_k}\int_0^{t_*}\Big[\int_{t_*}^T(u-\tau)^{H-\frac32}du\Big]^2d\tau\\
&\lesssim&\frac{1}{\lambda_k}\int_0^{t_*}(t_*-\tau)^{2H-1}d\tau\lesssim\frac{1}{\lambda_k}t_*^{2H}
\end{eqnarray*}
and
\begin{eqnarray*}
K_4&\lesssim&\frac{1}{\lambda_k}\int_{t_*}^T\left[\int_{\tau+t_*}^T(u-\tau)^{H-\frac32}du\right]^2d\tau\\
&\lesssim&\frac{1}{\lambda_k}\int_{t_*}^T((T-\tau)^{2H-1}+t_*^{2H-1})d\tau\lesssim\frac{1}{\lambda_k}+\frac{1}{\lambda_k}t_*^{2H-1}.
\end{eqnarray*}
Combining the above estimates and choosing $t_*=\lambda_k^{-1}$, we deduce
\begin{align*}
I_3(T)\lesssim t_*^{2H+2}+\frac{1}{\lambda_k}t_*^{2H}+t_*^{2H+1}+\frac{1}{\lambda_k}+\frac{1}{\lambda_k}t_*^{2H-1}\lesssim\lambda_k^{-2H}.
\end{align*}
For the case $H\in(\frac{1}{2},1)$, we have
\begin{eqnarray*}
&&\mathbb{E}\Big[
\Big(\int_0^T\frac{\sin(\sqrt{\lambda_k}(T-\tau))}{\sqrt{\lambda_k}}dB^H(\tau)\Big)^2\Big]\\
&=&\alpha_H\int_0^T\int_0^T\frac{\sin(\sqrt{\lambda_k}(T-p))}{\sqrt{\lambda_k}}\frac{\sin(\sqrt{\lambda_k}(T-q))}{\sqrt{\lambda_k}}|p-q|^{2H-2}dpdq\\
&\lesssim&\frac{1}{\lambda_k}\int_0^T\int_0^T|p-q|^{2H-2}dpdq
\lesssim\frac{1}{\lambda_k}T^{2H}.
\end{eqnarray*}
Hence, we conclude for any $H\in (0,1)$ that
\begin{align*}
\mathbb{E}\Big[
\Big(\int_0^T\frac{\sin(\sqrt{\lambda_k}(t-\tau))}{\sqrt{\lambda_k}}dB^H(\tau)\Big)^2\Big]\lesssim\lambda_k^{-\gamma},
\end{align*}
where $\gamma=\min\{2H,1\}$. When $k\rightarrow\infty$, $\lambda_k^{-\gamma}\rightarrow0$, which shows that it is unstable to recover $|g|$.
\begin{theorem}\label{theroem3f}
The inverse problem is unstable to recover the source terms $f$ and $\pm g$. Moreover, the following estimates hold:
\begin{equation*
\left|\int_0^Th(\tau)\frac{\sin(\sqrt{\lambda_k}(T-\tau))}{\sqrt{\lambda_k}}d\tau\right|\lesssim\frac{1}{\sqrt{\lambda_k}}
\end{equation*}
and
\begin{equation*
\mathbb{E}\Big[
\Big(\int_0^T\frac{\sin(\sqrt{\lambda_k}(T-\tau))}{\sqrt{\lambda_k}}dB^H(\tau)\Big)^2\Big]\lesssim\lambda_k^{-\gamma},
\end{equation*}
where $\gamma=\min\{2H,1\}$.
\end{theorem}
\section{Numerical experiments}
In this section, we present some numerical experiments for the one-dimensional problem where $D=[0,\pi]$. For some fixed integers $N_t$ and $N_x$, we define the time and space step-sizes $h_t=T/N_t$, $ h_x=\pi/N_x$ and nodes
\[
t_n=nh_t,~n\in\{0,1,2,\ldots,N_t\},\quad x_i=ih_x,~i\in\{0,1,2,\ldots,N_x\}.
\]
For the direct problem, the second order central difference is utilized to generate the synthetic data. Let $u^n_i$ be the numerical approximation to $u(x_i, t_n)$. Then we obtain the following explicit scheme:
\begin{align}\label{ed}
\frac{u_i^{n-1}-2u_i^n+u_i^{n+1}}{h_t^2}-\frac{u_{i-1}^n-2u_i^n+u_{i+1}^n}{h_x^2}=f(x_i)h(t_n)+g(x_i)\frac{B^H(t_{n})-B^H(t_{n-1})}{t_{n}-t_{n-1}}.
\end{align}
We use the ghost point method to handle the discretization of the initial condition $u_t(x,0)$. The virtual point $u_i^{-1}$ is introduced and the first order central difference is adopted, i.e., $\partial_t u|_{t=0}=(u_i^1-u_i^{-1})/(2h_t)$. From (\ref{ed}), when $n=0$, another equation about $u_i^{-1}$ can be obtained. Then we may use the two equations to get rid of $u_i^{-1}$ and obtain $u_i^1$.
For the inverse problem, the coefficients $f_k$ and $g_kg_l$ can be recovered by using (\ref{Iexp}) and (\ref{Icvar}), respectively. Once the coefficients are available, the source functions $f$ and $g^2$ can be expressed as
\[
f=\sum_{k=1}^\infty f_k\varphi_k,\quad g^2=\sum_{k,l=1}^\infty g_kg_l\varphi_k\varphi_l.
\]
Noting that computing $|g|$ and $g^2$ are equivalent, here we consider computing $g^2$. Since the inverse problem is ill-posed, we truncate above series by keeping the first $N$ terms as a regularization.
In the numerical experiments, we choose $N=9$, $N_t=2^{13}, N_x=100$, and $T=1$. The exact functions in \eqref{directp} are chosen as
\begin{equation*}
h(t)=1, \quad f(x)=\sin(3x), \quad g(x)=\exp(-(x-0.5\pi)^2).
\end{equation*}
We compute 1000 sample paths when simulating the covariance of the solution. In addition, the data is polluted by a uniformly distributed noise with level $\delta$.
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{d1hc}\\
\caption{The relative errors of reconstruction for $g^2$ with different $H$ and a fixed $\delta=0.001$.}\label{fig2}
\end{figure}
\begin{table}\label{tabl1}
\centering
\caption{The relative errors of reconstructions for $f$ and $g^2$ with different $\delta$ and a fixed $H=0.9$.
\begin{tabular}{lllllll}
\hline
\hline
$\delta$ & $0.001$ & $0.005$ & $0.01$ & $0.05$ & $0.1$ \\
\hline
$f$ & $0.0260$ & $0.0261$ & $0.0261$ & $0.0286$ & $0.0837$ \\
\hline
$g^2$& $0.0141$ & $0.0147$ & $0.0237$ & $0.0729$ & $0.0683$ \\
\hline
\hline
\end{tabular}
\end{table}
We report the numerical results for different sets of parameters $(H, \delta)$. Figure \ref{fig2} shows the results for the relative errors of $g^2$ with different $H$ and a fixed $\delta=0.001$. For $H=0.9$, the results for the relative errors of $f$ and $g^2$ with different $\delta$ are given in Table 1. Figure \ref{fig3} plots the results of
$\{H=0.9, \delta=0.001\}, \{H=0.2, \delta=0.001\}$, and $\{H=0.9, \delta=0.01\}$. For $H=0.5$ and $\delta=0.1, 0.01, 0.001$, the exact and reconstructed solutions are given in Figure \ref{fig4}. Based on the numerical experiments, it can be observed that the reconstructions would be more accurate if the problem is more regular, i.e., $H$ is larger; if the noise level $\delta$ is smaller, the results would also be better, which exactly implies the ill-posedness of the inverse problem.
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{fh9d1}
\includegraphics[width=0.4\textwidth]{g2h9d1}\\
\includegraphics[width=0.4\textwidth]{fh2d1}
\includegraphics[width=0.4\textwidth]{g2h2d1}\\
\includegraphics[width=0.4\textwidth]{fh9d5}
\includegraphics[width=0.4\textwidth]{g2h9d5}\\
\caption{The exact solution is plotted against the reconstructed solutions with $\{H=0.9,\delta=0.001\}, \{H=0.2,\delta=0.001\}$ and $\{H=0.9,\delta=0.01\}$. (left) $f(x)$; (right) $g^2(x)$.}\label{fig3}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{fH5d01}
\includegraphics[width=0.4\textwidth]{g2H5d01}\\
\includegraphics[width=0.4\textwidth]{fH5d001}
\includegraphics[width=0.4\textwidth]{g2H5d001}\\
\includegraphics[width=0.4\textwidth]{fH5d0001}
\includegraphics[width=0.4\textwidth]{g2H5d0001}\\
\caption{The exact solution is plotted against the reconstructed solutions with $H=0.5$ and $\delta=0.1, 0.01, 0.001$. (left) $f(x)$; (right) $g^2(x)$. }\label{fig4}
\end{figure}
\section{Conclusion}
In this paper, we have studied an inverse random source problem for the wave equation driven by the fBm. We show that the direct problem is well-posed and the inverse problem is ill-posed in the sense that a small deviation of the data may lead to a huge error in the reconstruction. Moreover, for the one-dimensional case, the inverse problem is shown to has a unique solution for the white noise case, i.e., $H=\frac12$. It is unclear if the uniqueness can still hold for the general Hurst index due to the highly oscillatory kernel function for large $k$. We will investigate the uniqueness issue for the ISP of the higher dimensional stochastic wave equation driven by the fBm in the future.
|
\section{Introduction}
\label{sec::intro}
The hydrodynamic behaviour of a spin current has been playing an increasingly prominent role in a variety of physical systems ranging from heavy ion collisions to condensed matter experiments.
In particular, the recent observation of global spin polarization of the $\Lambda$ and $\overline{\Lambda}$ particles in heavy-ion collisions at RHIC \cite{STAR:2017ckg,Adam:2018ivw} and the experimental realization of spin currents induced by vorticity in liquid metals \cite{Takahashi} have aroused strong interest in the subject, calling for a theoretical underpinning of hydrodynamics in the presence of a spin current.
The derivation of a complete and consistent set of constitutive relations in spin hydrodynamics is lacking in the literature. The main goal of this work is to provide the tools for carrying out such an analysis and to use these tools to obtain the constitutive relations for a parity invariant and conformal fluid to subleading order in a derivative expansion.
Recall that hydrodynamics is a universal low energy effective field theory of many body, finite temperature systems. The equations of motion of hydrodynamics consist of local conservation laws (e.g., energy momentum conservation or charge conservation). The dynamical variables are given by a temperature field $T$, a velocity field $u^{\mu}$ (which we normalize such that $u^{\mu}u_{\mu}=-1$ in the relativistic setting we are working in), and chemical potentials associated with other conserved charges present. In the current context angular momentum conservation leads to a (non-)conservation equation for the spin current,
which implies the existence of a spin potential $\mu_{ab}$, i.e. the spin analog of electric chemical potential.
For a system where the only conserved charge is the spin current and energy momentum tensor, and in the absence of anomalies, one finds
\begin{align}
\begin{split}
\label{E:EOMs}
\mathring{\nabla}_{\mu}T^{\mu\nu} & = \frac{1}{2} R^{\rho\sigma\nu\lambda}S_{\rho\lambda\sigma} - T_{\rho\sigma}K^{\nu a b}e^{\rho}{}_{a} e^{\sigma}{}_{b} \\
\mathring{\nabla}_{\lambda}S^{\lambda}{}_{\mu\nu} &= 2 T_{[\mu\nu]} - 2 S^{\lambda}{}_{\rho[\mu}e_{\nu]}{}^{a} e_{\rho}{}^{b} K_{\lambda}{}_{ ab}\,,
\end{split}
\end{align}
where we denote the vielbein by $e^{\mu}_{a}$ and will use it to convert spacetime indices to tangent bundle indices. $A_{[\alpha\beta]} = \frac{1}{2}(A_{\alpha\beta}-A_{\beta\alpha})$, $R^{\alpha}{}_{\beta\gamma\delta}$ is the Riemann tensor and $K_{\mu}{}^{ab}$ is the contorsion tensor, related to the spin connection, $\omega_{\mu}{}^{ab}$ via $\omega_{\mu}{}^{ab} = \mathring{\omega}_{\mu}{}^{ab} + K_{\mu}{}^{ab}$ where $\mathring{\omega}_{\mu}{}^{ab} = e_{\nu}{}^{a} \left(\partial_{\mu}e^{\nu b} + \mathring{\Gamma}^{\nu}{}_{\sigma\mu}e^{\sigma b}\right)$ with $\mathring{\Gamma}^{\alpha}{}_{\beta\gamma} = g^{\alpha\delta}\left(\partial_{(\beta}g_{\gamma)\delta} - \frac{1}{2}\partial_{\delta}g_{\beta\gamma}\right)$. That is, ringed connections and derivatives denote expressions evaluated using the Christoffel connection.
The virtue of using torsion in intermediate stages of the computation is that it allows us to uniquely determine the spin current and stress tensor. Recall that the stress tensor and spin current can be modified via a Belinfante-Rosenfeld transformation \cite{BELINFANTE1940449,rosenfeld1940tenseur}. This transformation, sometimes referred to as a pseudo gauge transformation, leads to an ambiguity in the spin current and energy momentum tensor. Fortunately, in the presence of torsion this ambiguity is removed, much like the ambiguity in the definition of the stress tensor is removed by defining it via its coupling to an external metric. Of course, the torsion and curvature should be set to zero when considering, e.g., heavy ion collisions.
To obtain the explicit form of the equations of motion for the hydrodynamic variables one needs a set of constitutive relations whereby all the conserved charge densities (including the energy momentum tensor) are expressed in terms of $T$, $u^{\mu}$, the relevant chemical potentials and their derivatives. These constitutive relations must satisfy certain criteria which have been shown to be captured by the second law of thermodynamics, at least to leading order in a derivative expansion \cite{Banerjee:2012iz,Jensen:2012jh,Bhattacharyya:2013lha,Glorioso:2016gsa,Jensen:2018hse,Jensen:2018hhx,Haehl:2018uqv}. Often, such constitutive relations are expressed in terms of a truncated expansion in derivatives of the hydrodynamic variables. As we will discuss at length shortly, an unusual feature of hydrodynamics with a spin current is that the spin potential is naturally associated with terms which are first order in derivatives.
In this work we compute the constitutive relations for the stress tensor $T^{\mu\nu}$ and spin current $S^{\mu}{}_{\alpha\beta}$ of a parity invariant conformal theory in $3+1$ dimensions, in a flat, torsionless background geometry, including all terms which contribute to the equations of motion expanded to second order in derivatives. We find
{\small
\begin{widetext}
\begin{align}
\begin{split}
\label{E:constitutive}
T^{(\mu\nu)} =& \epsilon_0 T^4 u^{\mu}u^{\nu} + \frac{1}{3}\epsilon_0 T^4 \Delta^{\mu\nu} - 2\eta_0 T^3 \sigma^{\mu\nu} + T_{BR}^{(\mu\nu)} + \mathcal{O}(\partial^2) \\
T^{-2} T^{[\mu\nu]} = &
\Delta_{\beta}{}^{[\mu}u^{\nu]} \left(
\ell_1 \mathcal{D}_{\alpha}\sigma^{\alpha\beta} + \ell_2 \mathcal{D}_{\alpha} \hat{M}^{\alpha\beta}\right)
+ \ell_3 \Delta^{\rho[\mu}\Delta^{\nu]\sigma} \mathring{\nabla}_{\rho} \hat{m}_{\sigma}
+\ell_4 u^{[\mu}\sigma^{\nu]\rho}\hat{m}_{\rho}
+\ell_5 u^{[\mu}\hat{M}^{\nu]\rho}\hat{m}_{\rho}
+\ell_6 u^{[\mu}{M}^{\nu]\rho}\hat{m}_{\rho} \\
&+\ell_7 \sigma^{[\mu}{}_{\rho} \hat{M}^{\nu]\rho}
+\ell_8 \sigma^{[\mu}{}_{\rho}M^{\nu]\rho}
+\ell_9 \hat{M}^{[\mu}{}_{\rho}M^{\nu]\rho}
-
T^{-2} S^{[\mu\nu]\rho}\left(a_{\rho}-\frac{1}{3}\Theta u_{\rho}\right) +
T^{-2} \left( S^{\rho}{}_{\rho}{}^{[\mu} a^{\nu]}-\frac{1}{3}\Theta S^{\rho}{}_{\rho}{}^{[\mu} u^{\nu]}\right) + T_{BR}^{[\mu\nu]} \\
T^{-2} S^{\mu}{}_{\nu\rho} & = 8 \rho_0 u^{\lambda} M_{\nu\rho} + 2 s_1 u^{\lambda}u_{[\nu}\hat{m}_{\rho]}+2s_2 u^{\lambda}\hat{M}_{\nu\rho} +S_{BR}{}^{\lambda}{}_{\nu\rho}
\end{split}
\end{align}
\end{widetext}
}
with
\begin{align}
\begin{split}
\label{E:BRterms}
T_{BR}^{\mu\nu} =& \frac{1}{2}\mathring{\nabla}{}_{\lambda} \left(S_{BR}{}^{\mu\nu\lambda} + S_{BR}{}^{\nu\mu\lambda}-S_{BR}{}^{\lambda\nu\mu}\right) \\
S_{BR}{}^{\lambda}{}_{\mu\nu} = & 2 T^3 \chi_1 \Delta^{\lambda}{}_{[\mu}u_{\nu]} +2 T^2 \chi_2 M^{\lambda}{}_{[\mu}u_{\nu]}
+ 2 \sigma_1 T^2 \sigma^{\lambda}{}_{[\mu}u_{\nu]} \\
& +2\sigma_2 T^2 \hat{M}^{\lambda}{}_{[\mu}u_{\nu]}
+ 2\sigma_3 T^2 \Delta^{\lambda}{}_{\mu}\hat{m}_{\nu]}
\end{split}
\end{align}
where we have decomposed the spin potential into transverse components,
\begin{equation}
\mu^{ab} = 2 u^{[a}m^{b]} + M^{ab}
\end{equation}
with $m^{a}u_{a}=0$ and $M^{ab}u_{b}=0$ and defined
\begin{align}
\begin{split}
\Theta &= \mathring{\nabla}_{\lambda}u^{\lambda}\,,
\qquad\qquad
a^{\mu}= u^{\alpha}\mathring{\nabla}_{\alpha}u^{\mu}\,,
\\
\Delta^{\mu\nu} & = g^{\mu\nu} + u^{\mu}u^{\nu}\,,
\qquad
\Omega^{\mu\nu} =
\Delta^{\mu\alpha}\Delta^{\nu\beta}\mathring{\nabla}_{[\alpha}u_{\beta]}\,, \\
\sigma^{\mu\nu} & =
\Delta^{\mu\alpha}\Delta^{\nu\beta}\mathring{\nabla}_{(\alpha}u_{\beta)} - \frac{1}{3}\Delta^{\mu\nu}\Theta\,,
\end{split}
\end{align}
which correspond to expansion, acceleration, the transverse projector, vorticity and the shear tensor, respectively. Calligraphic derivatives and hatted quantities are given by
\begin{align}
\begin{split}
\label{E:DandM}
\mathcal{D}_{\alpha}\sigma^{\alpha\beta} &= \mathring{\nabla}{}_{\alpha}\sigma^{\alpha\beta} - 3 a_{\alpha}\sigma^{\alpha\beta}\,, \\
\mathcal{D}_{\alpha}\hat{M}^{\alpha\beta} & = \mathring{\nabla}{}_{\alpha}\hat{M}^{\alpha\beta} - a_{\alpha}\hat{M}^{\alpha\beta}\,, \\
\hat{m}^{\mu} &= m^{\mu}-a^{\mu} \,, \\
\hat{M}^{\mu\nu} &= M^{\mu\nu} + \Omega^{\mu\nu} \,,
\end{split}
\end{align}
and circular brackets denote a symmetrized decomposition of indices, viz., $T^{(\mu\nu)} = \frac{1}{2} \left(T^{\mu\nu}+T^{\nu\mu}\right)$.
Adding terms of the form \eqref{E:BRterms} to the stress tensor and current is usually referred to as a Belinfante-Rosenfeld transformation. Such terms will not modify the equations of motion and are often used to generate a symmetric stress tensor and vanishing spin current from an asymmetric stress tensor and its associated spin current. As we will see shortly, such terms should not be removed. While they do not contribute to the equations of motion, they do contribute to the expectation value of the stress tensor and current. This was first remarked on in \cite{Becattini:2018duy}.
Inserting \eqref{E:constitutive} into \eqref{E:EOMs} one obtains dynamical equations for the hydrodynamic variables $T$, $u^{\mu}$ and $\mu_{ab}$ which can be solved for once supplemented by initial conditions. At late times we expect that the system reaches thermodynamic (or hydrostatic) equilibrium where the temperature, velocity field, and spin chemical potential are fixed in terms of external forces acting on the system. In particular, we find that in equilibrium and in the absence of torsion, the equilibrated spin chemical potential will be proportional to the thermal vorticity, $\Omega^{\mu\nu} - 2 u^{[\mu}a^{\nu]}$ as predicted in \cite{Becattini:2018duy,Becattini:2020riu}.
Spin current hydrodynamics may be relevant for the study of hyperon polarization measurements in heavy ion collisions.
The prediction of global spin polarization in heavy-ion collisions, based on perturbative QCD, was initiated in \cite{Liang:2004ph,Liang:2004xn}. The first attempt to relate spin polarization to hydrostatic vorticity can be found in \cite{Abreu:2007kv} and \cite{Becattini:2007sr} and has been elaborated on in subsequent work \cite{Voloshin:2004ha,Betz:2007kg,Becattini:2013fla,Florkowski:2017dyn,Florkowski:2017ruc,Becattini:2020ngo,Becattini:2020sww}. See also \cite{Florkowski:2018fap,Becattini:2020ngo,Bhadury:2021oat} for reviews.
This thread was continued by studies of entropy production in \cite{Hattori:2019lfp} and a classification of spin current sources and hydrodynamic constitutive relations in \cite{Gallegos:2020otk}. The latter work also pioneered a holographic study of spin transport. Other descriptions of spin hydrodynamics can be found in \cite{Karabali:2014vla,Montenegro:2017rbu,Li:2020eon,Singh:2020rht}. Despite these developments, a fully consistent set of constitutive relations which we derive in this paper did not appear in previous literature.
Before attempting to apply hydrodynamics with spin currents to the study of heavy ion collisions there are two issues that need to be addressed. The first, is whether such a hydrodynamic description is at all relevant for the physical process at hand, and the second is whether the hydrodynamical description of spin currents is complete. In this work we have addressed the second issue but not the first. Nevertheless, in section \ref{sec::heavy-ions} we carry out a simple analysis of perturbed Bjorken flow associated with the constitutive relations \eqref{E:constitutive}. Using some coarse approximations we find a simple one parameter model that nicely fits the experimental results for hyperon polarization. See figure \ref{fig1}.
\section{Spin current hydrostatics}
\label{sec::hydrostatics}
In the presence of time independent sources such as an external metric or gauge field, the fluid is expected to reach a time independent hydrostatic equilibrium configuration whereby Euclidean correlators of the theory decay exponentially. This exponential decay implies that momentum space correlation functions at zero frequency are analytic in the spatial momenta implying that their associated generating function will be a local function of the background fields. Such a generating function was computed explicitly in \cite{Banerjee:2012iz,Jensen:2012jh}. In what follows we use the same technique to study hydrostatically equilibrated spin current dynamics.
The sources which couple to the energy momentum tensor and spin current are the vielbein $e^a{}_{\mu}$ and spin connection $\omega_{\mu}{}^{ab}$,
\begin{equation}
\label{E:Svariation}
\delta S = \int d^4x |e| \left(
T^{\mu}{}_{a} \delta e^{a}{}_{\mu} + \frac{1}{2}S^{\lambda}{}_{ab}\delta\omega_{\lambda}{}^{ab}
\right)\,,
\end{equation}
where the integral is over all space dimensions and a compact Euclidean time direction with parametric length $T_0^{-1}$.
In a hydrostatic setting the sources will be time independent, viz.
\begin{equation}
\label{E:timeindependence}
\pounds_V e^{a}{}_{\mu} = 0\, , \qquad \pounds_V \omega_{\mu}^{ab}=0
\end{equation}
where $V^{\mu}$ points in the time direction and $\pounds_V$ denotes its associated Lie derivative
(the first equality implies that $V^{\mu}$ is a timelike Killing vector).
The generating function for hydrodynamics with a spin current will be given by a local diffeomorphism and Lorentz invariant expression constructed out of the sources $e^a{}_{\mu}$ and $\omega_{\mu}{}^{ab}$, and the time direction $V^{\mu}$.
With some prescience let us denote
\begin{equation}
\label{E:fluidparameters}
T = \frac{T_0}{\sqrt{-V^2}},
\qquad
u^{\mu} = \frac{V^{\mu}}{\sqrt{-V^2}},
\qquad
\mu^{ab} = \frac{\omega_{\mu}{}^{ab} V^{\mu}}{\sqrt{-V^2}}\,.
\end{equation}
These quantities will correspond to the hydrostatic temperature, velocity field and spin potential respectively. To see this we consider the most general generating function which will lead to constitutive relations which contain no derivatives of the parameters \eqref{E:fluidparameters},
\begin{equation}
\label{E:idealS}
\ln Z_{\rm id} = W_{\rm id} = \int d^4x |e| P(T,\,{M}^2,\,m \cdot \tilde{M},\,m^2)
\end{equation}
where $M^2 = 2 M_{\mu\nu}M^{\mu\nu}$ and $m\cdot\tilde{M} = m_{\alpha}M_{\beta\gamma}u_{\delta}\epsilon^{\alpha\beta\gamma\delta}$.
We will refer to a fluid whose constitutive relations are completely determined by \eqref{E:idealS} as an ideal fluid.
The current and stress tensor associated with \eqref{E:idealS} are given by
\begin{subequations}
\label{E:idealC}
\begin{align}
\begin{split}
T_{\rm id}^{\alpha\beta} &= \epsilon u^{\alpha}u^{\beta} + P \Delta^{\alpha\beta} - 2 \left(\frac{\partial P}{\partial m^2}+ 4 \frac{\partial P}{\partial {M}^2} \right)u^{\alpha}M^{\beta\gamma} m_{\gamma} \\
S_{\rm id}^{\lambda}{}_{\alpha\beta} &= u^{\lambda}\rho_{\alpha\beta} \,,
\end{split}
\end{align}
with
\begin{align}
\begin{split}
\label{E:rhodef}
\epsilon =& - P + \frac{\partial P}{\partial T} T + \frac{1}{2} \rho_{ab}\mu^{ab}\,, \\
\rho_{\alpha\beta} =& 8 \frac{\partial P}{\partial {M}^2 } M_{\alpha\beta} + \frac{\partial P}{\partial m \cdot \tilde{M}} \left(4 \tilde{m}_{\alpha\beta} - u_{\alpha}\tilde{M}_{\beta} + \tilde{M}_{\alpha}u_{\beta} \right)
\\
&- 2 \frac{\partial P}{\partial m^2} \left(u_{\alpha} m_{\beta} - m_{\alpha}u_{\beta} \right) \,, \\
\end{split}
\end{align}
\end{subequations}
once we restrict ourselves to a flat, torsionless geometry. Here we defined $\tilde m_{\alpha\beta} = -1/2 \epsilon_{\alpha\beta\gamma\delta} u^\gamma m^\delta$.
There are several lessons to be learnt from \eqref{E:timeindependence} through \eqref{E:idealC}. First, note that the identifications \eqref{E:fluidparameters} yield the expected Gibbs Duhem relations once we identify $P$ with the pressure and $s=\partial P / \partial T$ with the entropy density $s$. As we will show in the next section, the entropy current $J^{\mu} = s u^{\mu}$ is conserved for the ideal fluid, once the equations of motion are satisfied.
Second, since all sources are time independent, we find that \eqref{E:fluidparameters} imply the hydrostatic relations
\begin{align}
\begin{split}
\label{E:hydrostaticidentities}
u^{\mu}K_{\mu}{}^{ab} &= \mu^{ab} + e^{a}{}_{\mu} e^{b}{}_{\nu} \left( \Omega^{\mu\nu} -2 u^{[\mu}a^{\nu]}\right)\,, \\
T \mathring{\nabla}_{\lambda} \frac{\mu^{\rho \sigma}}{T} &= R^{\rho \sigma}{}_{\lambda \alpha} u^\alpha -2 K_{\lambda \alpha}{}^{[\rho} \mu^{\sigma] \alpha} \\
a_{\mu} &= -\frac{\mathring{\nabla}_{\mu}T}{T}\,,
\end{split}
\end{align}
where $R$ is the Riemann tensor in the presence of torsion that is expressed in terms of the torsionless Riemann tensor as
\begin{equation}
R^{\rho \sigma}{}_{\lambda \alpha} = \mathring{R}^{\rho \sigma}{}_{\lambda \alpha} + 2 \mathring{\nabla}_{[\lambda} K_{\alpha]}{}^{\rho \sigma}-2K_\lambda{}^{\kappa [\rho } K_{\alpha \kappa}{}^{\sigma]} \,.
\end{equation}
The first equality in \eqref{E:hydrostaticidentities} has been mentioned in \cite{Becattini:2013fla,Jensen:2013kka} in the absence of torsion. It implies that a non vanishing spin potential must be supported by fluid vorticity or by acceleration (or, alternatively, temperature gradients) in order to maintain thermal equilibrium.
In the absence of torsion, a non flat metric, or other external forces, the fluid will eventually settle down to a thermally equilibrated steady state in which the velocity field and temperature are covariantly constant. The first equality in \eqref{E:hydrostaticidentities} implies that the spin potential must vanish in such an equilibrated state. Therefore, if we wish to construct a gradient expansion around an equilibrium configuration we must count the spin potential as first order in derivatives.
Classifying the spin potential as a first order in derivatives term implies that the timelike components of the torsion tensor, $k^{ab}$, are also first order in derivatives. It remains to classify the transverse components of torsion, $\kappa_{\nu}{}^{ab} = \Delta^{\mu}{}_{\nu}K^{\nu}{}_{ab}$. In what follows we consider $ \kappa_{\mu}{}^{ab} $ as first order in derivatives, but it should be possible to set $\kappa_{\mu}{}^{ab}$ to be zeroth order in derivatives yielding torsio-hydrodynamics, an analog of magneto-hydrodynamics.
Before proceeding with higher order corrections to the ideal fluid, we remark that \eqref{E:hydrostaticidentities} implies that in the absence of torsion, $M^{\mu\nu}+\Omega^{\mu\nu}=0$ and $m^{\mu}-a^{\mu}=0$. Thus, there is an ambiguity in determining the constitutive relations \eqref{E:idealC}. Of course, such an ambiguity will be resolved once non hydrostatic corrections are taken into account. In what follows we will consistently choose $M^{\mu\nu}$ and $m^{\mu}$ as hydrostatic variables in the absence of torsion over $\Omega^{\mu\nu}$ and $a^{\mu}$.
The hydrostatic gradient corrections to the constitutive relations can be obtained by expanding the hydrostatic generating function, $W$, in a derivative expansion. In this work we are interested in the equations of motion expanded to second order in derivatives. Since the antisymmetric components of the stress tensor sources the divergence of the spin current, we must expand the constitutive relations associated with the antisymmetric component of the stress tensor to second order in derivatives and the remaining constitutive relations to first order in derivatives.
Therefore, to compute all possible corrections to the ideal fluid constitutive relations, we must classify all possible first order in derivative scalars which can contribute to the hydrostatic generating function, $W$, and all possible second order in derivative scalars which can contribute to the antisymmetric components of the stress tensor. In order to limit the number of such terms and also simplify future expressions we assume that the system is invariant under parity and also conformally invariant. A full analysis of the constitutive relations which are not restricted by symmetry will be discussed in a future paper.
The Weyl transformation of the spin connection associated with the Christoffel connection, $\mathring{\omega}_{\mu}^{ab}$ can be determined from the Weyl rescaling of the vielbein, $e^{a}{}_{\mu} \to e^{\phi}e^{a}{}_{\mu}$. In what follows we will assume that the spin connection transforms in the same way as $\mathring{\omega}{}_{\mu}{}^{ab}$. Alternately, that the contorsion tensor is inert under Weyl rescalings of the metric. One can argue that if the contorsion tensor transforms non trivially under Weyl rescalings then its transformation properties are such that a vanishing contorsion tensor is conformally equivalent to a non vanishing one \cite{Buchbinder:1985ux,Shapiro:2001rz}.
Using \eqref{E:EOMs}
we find that the change in the stress tensor and spin current due to an infinitesimal Weyl rescaling is given by
\begin{subequations}
\label{E:Crules}
\begin{align}
\begin{split}
\delta T^{\mu\nu} =& -6 \phi T^{\mu\nu} - S^{\mu\nu\rho}\partial_{\rho}\phi + S_{\lambda}{}^{\lambda\mu}\partial^{\nu}\phi \\
\delta S^{\lambda\mu\nu} =& -6 \phi S^{\lambda\mu\nu}
\end{split}
\end{align}
Using \eqref{E:Svariation} we find that tracelesness is replaced by
\begin{equation}
T^{\mu}{}_{\mu} = \mathring{\nabla}_{\mu}S_{\lambda}{}^{\lambda\mu}\,.
\end{equation}
\end{subequations}
We defer an extensive discussion of conformal invariance in the presence of torsion, and the recovery of the canonical transformation laws for the stress tensor in its absence to future work.
It follows that the transverse part of the spin potential, $M^{\mu\nu}$, transforms homogenously under Weyl rescalings while $m^{\alpha}$ does not. Thus, in a conformally invariant theory the pressure $P$ in \eqref{E:idealS} can depend only on $T$ and $M^2$. Counting $M^2$ as second order in derivatives implies that
\begin{equation}
\epsilon = \epsilon_0 T^4\,,
\quad
P= \frac{1}{3}\epsilon_0 T^4\,,
\quad
\frac{\partial P}{\partial M^2} = \rho_0 T^2\,,
\end{equation}
up to second order in derivative corrections.
It is now straightforward, though somewhat tedious to argue that the most general correction to $W$, $W_h$, at the order we are interested in is given by
\begin{equation}
\label{E:W1}
\small
W_h = \int d^4x |e| \left(\chi^{(1)} T^3 \kappa + 2 \chi^{(2)}_1 T^2 \kappa_A^{\mu\nu}M_{\mu\nu} +2 \chi^{(2)}_2 T^2 K^{\mu\nu} M_{\mu\nu} \right)\,,
\end{equation}
where $\chi^{(1)}$ and the $\chi^{(2)}_{j}$'s are numbers, $\kappa = u_a e^{\mu}{}_{b} K_{\mu}{}^{b a}$, ${\kappa}_A^{\alpha\nu} = u_{\beta}\Delta^{\mu[\nu}e^{\alpha]}{}_{a}K_{\mu}^{ab}e^{\beta}{}_{b}$ and $K^{\alpha\beta} = u^{\mu}\Delta^{\alpha}{}_{\gamma} \Delta^{\beta}{}_{\delta} e^{\gamma}{}_c e^{\delta}{}_d K_{\mu}{}^{cd}$.
The stress tensor and spin current derived from \eqref{E:W1} are given by
\begin{widetext}
\begin{align}
\begin{split}
\label{E:hydrostatic1}
T^{(\mu\nu)}_h =& 4 T^3 \chi^{(1)} u^{(\mu}m^{\nu)} + \mathcal{O}(\partial^2) \,, \\
T^{[\mu\nu]}_h = &
T^3 \chi^{(1)}\left( 2 u^{[\mu}m^{\nu]} - M^{\mu\nu} \right)
+4 T^2 \left(\chi^{(2)}_1-\chi^{(2)}_2\right) u^{[\mu}M^{\nu]\alpha}m_{\alpha} +2T^2 \chi^{(2)}_1 \mathring{\nabla}_{\alpha}M^{\alpha [\nu}u^{\mu]}
\\
S_h^{\lambda}{}_{ab} =& 2 T^3 \chi^{(1)} \Delta^{\lambda}{}_{[a}u_{b]} - 4 T^2 \chi^{(2)}_1 M^{\lambda}{}_{[a}u_{b]} + 4 T^2 \chi^{(2)}_{2} u^{\lambda}M_{ab} \,,
\end{split}
\end{align}
\end{widetext}
once we set the torsion to zero.
The contribution of the term associated with $\chi^{(2)}_2$ to the constitutive relations is identical to that of an ideal, conformal fluid at second order in the derivative expansion. At least as far as the antisymmetric part of the stress tensor and the spin current are concerned. Therefore, we may, without loss of generality remove the former by an appropriate shift of the latter.
\section{Spin current hydrodynamics}
\label{sec::hydrodynamics}
The remaining contributions to the stress tensor and current, $T_r^{\mu\nu}$ and $S_r^{\lambda}{}_{ab}$ contain all possible expressions which vanish in equilibrium, are parity invariant, and satisfy \eqref{E:Crules}.
We find
\begin{widetext}
\begin{align}
\begin{split}
\label{E:dissipative}
T^{-3} T^{(\mu\nu)}_r =& -\sigma_6 \sigma^{\mu\nu}
+ \left(\sigma_7 -4 \chi^{(1)} \right) u^{(\mu}\hat{m}^{\nu)}
- \frac{1}{3} \chi^{(1)} \Theta \Delta^{\mu\nu}
- \chi^{(1)} \Theta u^{\mu}u^{\nu} \,, \\
T^{-2} T^{[\mu\nu]}_r = &
T \left(\sigma_7-2\chi^{(1)}\right) u^{[\mu}\hat{m}^{\nu]}
+ T \sigma_8 \hat{M}^{\mu\nu}
+\Delta_{\beta}{}^{[\mu}u^{\nu]} \left(\lambda_1 \mathcal{D}_{\alpha}\sigma^{\alpha\beta}
+ \lambda_2 \mathcal{D}_{\alpha} \hat{M}^{\alpha\beta}\right)
+ \lambda_3 \Delta^{\rho[\mu}\Delta^{\nu]\sigma} \mathring{\nabla}_{\rho} \hat{m}_{\sigma} \\
&
+\lambda_4 u^{[\mu}\sigma^{\nu]\rho}\hat{m}_{\rho}
+\lambda_5 u^{[\mu}\hat{M}^{\nu]\rho}\hat{m}_{\rho}
+(\lambda_6 - 4 \chi^{(2)}_1 +8 \rho_0) u^{[\mu}{M}^{\nu]\rho}\hat{m}_{\rho}
+ \lambda_7 \sigma^{[\mu}{}_{\rho} \hat{M}^{\nu]\rho}
+\lambda_8 \sigma^{[\mu}{}_{\rho}M^{\nu]\rho}
+\lambda_9 \hat{M}^{[\mu}{}_{\rho}M^{\nu]\rho} \\
&+\frac{2}{3} \chi^{(2)}_1 {M}^{\mu\nu}\Theta - T^{-2} S_r^{[\mu\nu]\rho}\left(a_{\rho}-\frac{1}{3}\Theta u_{\rho}\right) + T^{-2} \left( S_r^{\rho}{}_{\rho}{}^{[\mu} a^{\nu]}-\frac{1}{3}\Theta S_r^{\rho}{}_{\rho}{}^{[\mu} u^{\nu]}\right)\\
T^{-2} S_r^{\lambda}{}_{ab} =& 2 \sigma_1 \sigma^{\lambda}{}_{[a}u_{b]}
+2 \sigma_2 \hat{M}^{\lambda}{}_{[a}u_{b]}
+2 \sigma_3 \Delta^{\lambda}{}_{[a}\hat{m}_{b]}
+2 \sigma_4 u^{\lambda}u_{[a}\hat{m}_{b]}
+2 \sigma_5 u^{\lambda} \hat{M}_{ab} \,,
\end{split}
\end{align}
\end{widetext}
A few comments are in order. To help the reader identify the role of the various terms in \eqref{E:dissipative} we have labeled coefficients associated with first order in derivative terms by $\sigma_i$ and coefficients associated with second order in derivative terms by $\lambda_i$. While one often denotes the shear viscosity by $\eta$ we have refrained from doing so for reasons that will become clear shortly. The $\chi^{(1)}$ and $\chi^{(2)}_i$ dependent terms appearing in \eqref{E:dissipative} have been introduced in order to ensure that \eqref{E:Crules} are satisfied out of equilibrium. The same goes for the last two terms on the right hand side of the expression for $T^{[\mu\nu]}$.
Also, we have written \eqref{E:dissipative} in what is usually referred to as the Landau frame where $u^{\mu}$ is an eigenvector of the stress tensor with negative eigenvalue.
Frame transformations offer an additional freedom in redefining the spin potential which we avoid using at this order in the derivative expansion. The hydrostatic stress tensor and spin current, $T_{\rm id}^{\mu\nu} + T_h^{\mu\nu}$ are written in a hydrostatic frame which is more natural from the point of view of the hydrostatic partition function.
To further simplify \eqref{E:dissipative} it is convenient to make the redefinitions
\begin{align}
\notag
\sigma_6 &= 2\eta_0 + \chi^{(1)}\,,&
\quad
\sigma_8 & = \eta_1+\chi^{(1)}\,,
\\
\notag
\lambda_1& = \ell_1+\sigma_1\,, &
\lambda_2&= \ell_2+\sigma_2\,,
\\
\notag
\lambda_3 & = \ell_3 + \sigma_3\,, &
\lambda_4 &= \ell_4-\sigma_3\,,
\\
\label{E:substitutions}
\lambda_5 & = \ell_5 - \sigma_3\,, &
\lambda_6 &= \ell_6 + \sigma_3\,,
\\
\notag
\lambda_7 &= \ell_7-\sigma_1 +\sigma_2\,,&
\lambda_8 & = \ell_8 - 2 \chi^{(2)}_1 + \sigma_1 \,,
\\
\notag
\lambda_9 & = \ell_9 + 2 \chi^{(2)}_1-\sigma_2\,.&&
\end{align}
With these redefinitions, the terms in the spin current associated with $\chi^{(1)}$, $\chi^{(2)}_1$, and $\sigma_i$ with $i=1,\ldots,3$ reduce to Belinfante Rosenfeld terms, discussed in the introduction, and therefore do not affect the equations of motion.
It is also important to note that the first order equations of motion (associated with the zeroth order constitutive relations of the symmetric stress tensor and spin current and first order constitutive relations of the antisymmetric part of the energy momentum tensor) are overdetermined in the sense that the spin potential does not contribute to them. Indeed, in order for these equations to be self consistent it must be the case that $\sigma_7=0$ and $\eta_1=0$.
Combining \eqref{E:idealC}, \eqref{E:hydrostatic1}, \eqref{E:dissipative} and \eqref{E:substitutions}, removing $\chi_2^{(2)}$ following the discussion after \eqref{E:hydrostatic1}, setting $\sigma_7=0$ and $\eta_1=0$, and slightly relabeling coefficients, yields \eqref{E:constitutive}.
We note in passing that the terms associated with $\sigma_4$ and $\sigma_5$ can also be packaged as a Belinfante Rosenfeld term by adding a $\lambda_{10} T^2 u^{\lambda}u^{[\mu}\mathcal{D}_{\lambda}\hat{m}^{\nu]}$ and a $\lambda_{11} T^2 u^{\lambda}\mathcal{D}_{\lambda}\hat{M}^{\mu\nu}$ term to the antisymmetric part of the stress tensor and then redefining $\lambda_{10} = \ell_{10} + \frac{1}{2}\sigma_4$ and $\lambda_{11} = \ell_{11} + \sigma_5$.
(with $u^{\lambda}\mathcal{D}_{\lambda}\hat{m}^{\mu} = u^{\lambda}\mathring{\nabla}_{\lambda} + \frac{1}{3} \theta \hat{m}^{\mu}-m \cdot \hat{m} + \hat{m} \cdot \hat{m} u^{\mu}$ and $u^{\lambda}\mathcal{D}_{\lambda} \hat{M}^{\mu\nu} = u^{\lambda}\mathring{\nabla}_{\lambda}\hat{M}^{\mu\nu} +\frac{1}{3}\theta \hat{M}^{\mu\nu}+2 u^{[\mu}\hat{M}^{\nu]\lambda}m_{\lambda}-2 u^{[\mu}\hat{M}^{\nu]\lambda}\hat{m}_{\lambda})$ The reason these last two terms don't appear in \eqref{E:dissipative} is that we have substituted those expressions with their values under the equations of motion.
The various coefficients multiplying the tensor structures in \eqref{E:constitutive}, e.g., $\eta_0$, are restricted by positivity of entropy production, unitarity of retarded correlation functions or unitarity of the Schwinger-Keldysh generating function \cite{Glorioso:2016gsa,Haehl:2018uqv,Jensen:2018hse,Jensen:2018hhx}. In what follows we will study restrictions on the coefficients in \eqref{E:constitutive} coming from positivity of entropy production. A study of the restrictions on coefficients via other methods is left for future work.
Following \cite{landau2013fluid} we posit the existence of an entropy current $J_S^{\mu}$ satisfying $\mathring{\nabla}_{\mu}J_S^{\mu} \geq 0$ under the equations of motion, such that for an ideal fluid $J_S^{\mu} = s u^{\mu}$ with $s = \partial P/\partial T$. For a non ideal fluid we take $J_S^{\mu} = s u^{\mu} + \mathcal{O}(\partial)$ where $\mathcal{O}(\partial)$ denotes corrections to the entropy current coming from explicit derivative terms appearing in the constitutive relations. Thus, the most general entropy current we may construct, to first order in derivatives is given by
\begin{equation}
\label{E:JS}
J_S^{\mu} = J_{c}^{\mu} + \left( s_1 \Theta u^{\mu} + s_2 a^{\mu} + s_3 m^{\mu} \right)T^2 \,.
\end{equation}
where
\begin{equation}
J_{c}^{\mu} = s u^{\mu} -\frac{u_{\nu}}{T}\left({T}^{\mu\nu}-T_{\rm id}^{\mu\nu}\right) - \frac{1}{2}\frac{\mu^{ab}}{T}\left(S^{\mu}{}_{ab}-S_{\rm id}^{\mu}{}_{ab}\right)
\end{equation}
is referred to as the canonical part of the entropy current. In a conformal theory the $s_i$ are constant.
When expanding the entropy current to first order in derivatives, the divergence of the entropy current is a second order in derivatives scalar. It is useful to classify the latter into two categories. The first are independent second order scalars, these are scalars which can not be written as products of first order scalars. The second includes products of first order scalars. All independent second order scalars appearing in the divergence of the entropy current must vanish on account of the positivity condition. For the same reason all products of first order scalars must arrange themselves into complete squares or vanish.
It is straightforward to show that
\begin{align}
\begin{split}
\label{E:divJcanon}
\mathring{\nabla}_{\mu}J_{c}^{\mu} =& -\mathring{\nabla}_{\mu}\left(\frac{u_{\nu}}{T}\right)\left({T}^{\mu\nu}-T_{\rm id}^{\mu\nu}\right)\\
&-\frac{1}{2}\mathring{\nabla}_{\mu} \left(\frac{\mu^{ab}}{T}\right)\left(S^{\mu}{}_{ab}-S_{\rm id}^{\mu}{}_{ab}\right)
-\frac{\mu^{ab}}{T}T_{ab}\,.
\end{split}
\end{align}
Inserting \eqref{E:JS} into $\mathring{\nabla}_{\mu}J_S^{\mu} \geq 0$
and using \eqref{E:divJcanon} we find
\begin{equation}
\label{E:noncanon}
s_1=-\chi_1\,,
\quad
s_2=\chi_1\,,
\quad
s_3=-\chi_1\,,
\quad
\eta_0 \geq 0\,.
\end{equation}
Let us make the following remarks. Since the spin potential is first order in derivatives the $n-1$th order spin current contributes to the $n$th order entropy current. Thus, the first order entropy current can only constrain the first order energy momentum tensor and zeroth order spin current. In practice, it constrains only $\eta_0$, the shear viscosity.
To determine constraints on the first order terms in the spin current one would need to go to second order in the entropy current. While we have not carried out such an analysis, we note that, at least for spin-less charged fluids, all constraints from the entropy current which imply equality type relations among transport coefficient are already implemented from the partition function. Further, all inequality type constraints appear at leading order in the entropy current \cite{Bhattacharyya:2013lha}.
Another somewhat unusual feature of hydrodynamics with a spin current is that the coefficient of the shear term in $T^{\mu\nu}$ is $-(\eta_0 + \chi_1) T^3 $, c.f, \eqref{E:substitutions}. Nevertheless, it is $\eta_0$ that is constrained to be positive which is perhaps compatible with the fact that $\chi_1$ does not enter into the equations of motion. We have checked that positivity of $\eta_0$ also follows from positivity of the appropriate stress tensor correlator. Note that a computation of two point functions of the stress tensor require knowledge of the expectation value of the stress tensor in the presence of a background metric and spin connection which we have not presented here.
\section{An application to heavy ion collisions}
\label{sec::heavy-ions}
In this short letter we do not presume to carry out a full fledged analysis of heavy ion collision experiments with possible spin currents manifesting during the short collision period. Instead, we consider a perturbed solution to the hydrodynamic equations of motion in the presence of spin with an underlying Bjorken ($SO(1,1)\times ISO(2) \times {Z}_2$) symmetry \cite{Bjorken:1982qr}.
We then attempt to relate the dependence of the spin potential on the initial temperature to the dependence of the average hyperon polarization vector on the beam energy.
Of course, a complete analysis, which we do not carry out in this short letter, should include a proper treatment of initial conditions, a full hydrodynamic simulation, and a comprehensive treatment of hadronization of the quark gluon plasma before reaching the detector.
Consider a collision of two gold ions of radii $R$ initially moving with a relativistic velocity directed along a Cartesian `$z$' coordinate. Let's assume that the fluid formed after the collision has Bjorken symmetry, that is, it is invariant under boosts along the beam direction, translations and rotations along the `$x$' and `$y$' directions, and under $z/t \to -z/t$. Going to a Milne coordinate system, $ds^2 = -d\tau^2 + \tau^2 d\eta^2 + dx^2 + dy^2$ where $\tau = \sqrt{t^2 -z^2}$ and $\eta = \textrm{arctanh}(z/t)$ are proper time and pseudo-rapidity respectively, we find that
\begin{equation}
\label{E:BjorkenT}
u^{\tau} = 1,
\qquad
T= T_0 \left(\frac{\tau_0}{\tau}\right)^{\frac{1}{3}} - \frac{\eta_0}{2 \epsilon_0\tau}\,,
\end{equation}
with all other components of $u^{\mu}$ and $\mu_{ab}$ vanishing, solve the equations of motion. Here $T_0$ is the temperature at the initial time $\tau_0$ when the fluid description is a viable one. Note that the shear viscosity to entropy ratio, $\eta/s$, satisfies $\eta/s=3 \eta_0 / 4 \epsilon_0$.
Since the spin potential vanishes on account of Bjorken symmetry, let's consider linear perturbations of Bjorken flow which break transverse translations and axial rotation, $T \to T+ \int d^2 q \delta T e^{i (q_x x+q_y y)}$, $u^{\mu} \to u^{\mu} + \int d^2 q \delta u^{\mu} e^{i (q_x x+q_y y)} $, and $\mu_{ab} \to \mu_{ab} + \int d^2 q \delta \mu_{ab}e^{i (q_x x+q_y y)}$. To mimic the experiment, we consider a peripheral collision with impact parameter $b$ along the `$x$' axis. Glancing beams are expected to create a non-trivial velocity gradient in the $x$ direction at initial proper time $\tau_0$ at which we assume hydrodynamics becomes applicable. To this end, we consider an initial velocity profile where $\delta u^{\eta}(\tau_0) \propto b q_x$, and other components of the perturbations to the velocity vanish. As a result, we find that $\delta m^{\eta}$, $\delta M^{\eta x} = \delta M^{\eta} i q_x $ and $\delta M^{\eta y} = \delta M^{\eta} i q_y$ are non zero while the temperature perturbations and all other components of the spin potential vanish.
To solve the equations of motion we will use the Floerchinger-Wiedemann (FW) approximation \cite{Florchinger:2011qf}, where
$\frac{3\eta_0}{4\epsilon_0}\frac{1}{T \tau}$
is perturbatively small but
$q^2 \tau^2 \frac{3\eta_0}{4 \epsilon_0}\frac{1}{T\tau} $
(with $q^2 = q_x^2+q_y^2$) is finite. In this approximation, only the leading term for the temperature in \eqref{E:BjorkenT} becomes relevant, the velocity field perturbations take the form
\begin{equation}\label{E:deltau}
\delta u^\eta = i u_0\, b\, q_x\, \tau^{-\frac53} e^{-\frac{9 q^2\eta_0 \tau_0}{16 T_0 \epsilon_0}\left(\frac{\tau}{\tau_0}\right)^\frac43}\, ,
\end{equation}
and $\delta M^{\eta}$ and $\delta m^{\eta}$ are determined algebraically from $\delta u^\eta$ and its derivatives.
Presumably, the stress tensor and spin current will evolve according to hydrodynamic theory from an initial Bjorken time $\tau_0$ to a final time $\tau_f$ when matter hadronizes, $T(\tau_f) = T_f \simeq 150 MeV$. The hadrons yield is then collected by the detector which measures its properties. Converting a hydrodynamics spin current and energy momentum tensor to a Hadron distribution is fraught with difficulty. One often used prescription for doing so works under the assumption that the particle distribution after hadronization follows a thermal distribution with temperature, velocity and chemical potential of the hydrodynamic configuration leading to it \cite{Cooper:1974mv}. Within this framework the polarization vector reads
\begin{equation}
\label{E:Pivec}
\Pi_{\alpha}(p) = -\frac{1}{4}\epsilon_{\alpha\rho\sigma\beta} \frac{p^{\beta}}{m} \frac{\int d\Sigma_{\lambda}p^{\lambda} B \mu^{\rho\sigma}}
{2\int d\Sigma_{\lambda} p^{\lambda} n_F}
\end{equation}
where $\int d\Sigma_{\mu}$ is an integral over the hadronization surface, $d\Sigma_{\mu} = \tau \delta_{\mu}^{\tau} d\eta dx dy$ in Bjorken coordinates, $p^{\mu}$ is the particle momentum, $m$ its mass, $n_F$ is the Fermi Dirac distribution and $B$ is an additional distributional quantity that depends on $u^{\mu}$ and $T$. See, e.g., \cite{Becattini:2013fla,Florkowski:2019qdp} for details.
It is tempting to use our solution to evaluate \eqref{E:Pivec} and compare to data. However, one should keep in mind that our hydrodynamic solution is rather simple minded, involving a linear perturbation on Bjorken symmetry on top of which we used the FW approximation. This perturbation should presumably capture a non vanishing impact parameter. Realistic collisions at mid centrality have an impact parameter of order of the nucleus size and are unlikely to resemble Bjorken flow. They should generate a large enough vorticity for a non trivial spin current to be generated which makes the validity of our linearized approximation somewhat suspect. Still, we have at our disposal an analytic solution to the hydrodynamic equations of motion with spin and it is hard to resist the temptation to compare it with the experimental results using \eqref{E:Pivec}. Hence, throwing caution to the wind, and inserting the perturbed Bjorken solution into \eqref{E:Pivec}, we find
\begin{widetext}
\begin{equation}
\label{E:Pimfull}
\Pi_{\mu}(p)
=
\frac{16 b
e^{-\frac{4 T_f^4 \epsilon_0 x_0^2}{9 T_0^3 \eta_0 \tau_0}}
u_0 \pi^{\frac{3}{2}} T_f^8 x_0 \epsilon_0^{\frac{3}{2}} (\ell_1+\ell_2)\hbox{Erf}\left(\sqrt{\frac{4 T_f^4 \epsilon_0 y_0^2}{9 T_0^3 \eta_0 \tau_0}}\right)}{27 m T_0^{\frac{13}{2}} \eta_0^{\frac{3}{2}} \ell_2 \tau_0^{\frac{13}{6}}}
\times
\begin{pmatrix}
-p^y I^{(1)} \\
0 \\
0 \\
I^{(2)}
\end{pmatrix}
\end{equation}
\end{widetext}
where now $x_0 = R - \frac{b}{2}$, $y_0 = \sqrt{R^2-\frac{b^2}{4}}$ and we have integrated over the range $-x_0<x<x_0$ and $-y_0<y<y_0$ which approximates the area of overlap of the two colliding nuclei.
The expressions for $I^{(n)}$ are given by
\begin{equation}
I^{(n)}(\tau_f) = \frac{\int d\eta~B~(p^{\tau})^n}{2 x_0\times 2 y_0 \times \int d\eta n_F p^{\tau}} \Bigg|_{\tau=\tau_f}\,,
\end{equation}
and $\hbox{Erf}$ denotes the error function.
We are particularly interested in the dependence of the spin polarization on the initial temperature, related to the beam energy. Making the reckless approximation that energy and nucleons are distributed uniformly in the nucleus and that the relation between energy density and temperature is of the form $\epsilon = \epsilon_0 T^4$ as dictated by conformal invariance, we find
\begin{equation}
T_0 = \left(\frac{2 N}{\pi R^2 \epsilon_0 \tau_0}\right)^{\frac{1}{4}}s_{NN}^{\frac{1}{8}}\,,
\end{equation}
where $N$ is the number of nucleons, $\sqrt{s_{NN}}$ is the beam energy per nucleon, and we approximated the volume of the nucleus as $\pi R^2 \tau_0$. It is clear that each of these approximations may be improved and upgraded, but as a preliminary order of magnitude estimate relating our hydrodynamic solution to the polarization vector, they are good enough.
Using, $\epsilon_0 = 12$, $T_f = 150 MeV$, $\eta/s = 1/4\pi$, $\tau_0=1 fm$, $R=7 fm$ and $b=10 fm$ (see \cite{Karsch:2000ps,Kovtun:2004de,Florchinger:2011qf,Karpenko:2016jyx}), we find
\begin{equation}
\frac{4 T_f^4 \epsilon_0 x_0^2}{9 T_0^3 \eta_0 \tau_0}
\simeq
\frac{5.1}{\left(\frac{s_{NN}}{\hbox{GeV}^2}\right)^{\frac{3}{8}}}
\qquad
\sqrt{\frac{4 T_f^4 \epsilon_0 y_0^2}{9 T_0^3 \eta_0 \tau_0}}
\simeq
\frac{5.5}{\left(\frac{s_{NN}}{\hbox{GeV}^2}\right)^{\frac{3}{16}}}\,.
\end{equation}
The overall scaling of $\Pi$ in terms of the energy per nucleon is of the form
\begin{equation}
\label{E:GotPi}
\Pi =
\alpha \frac{\exp\left(-\frac{5.1}{\left(\frac{s_{NN}}{\hbox{GeV}^2}\right)^{\frac{3}{8}}}\right)
\hbox{Erf}\left(\frac{5.5}{\left(\frac{s_{NN}}{\hbox{GeV}^2}\right)^{\frac{3}{16}}}\right)}
{\left(\frac{s_{NN}}{\hbox{GeV}^2}\right)^{\frac{13}{16}}}\,,
\end{equation}
where we have replaced the overall constant in \eqref{E:Pimfull}, which depends on the undetermined initial value for the velocity field perturbations $u_0$, and on the coefficients $\ell_1$ and $\ell_2$, with $\alpha$. To compare to experiment we need to work out $\Pi_{\mu}$ in the center of mass frame of the hyperon. Such a Lorentz transformation will not affect the dependence of $\Pi_{\mu}$ on $s_{NN}$. Therefore, we can attempt to fit \eqref{E:GotPi} to experiment by fitting to a single parameter, $\alpha$. Using the data from \cite{STAR:2017ckg}, we find a surprisingly good fit to $\alpha = 286 \pm 52$. See figure \ref{fig1}.
We emphasize that our phenomenological analysis crucially depends on the new transport coefficients $\ell_1$ and $\ell_2$ in the constitutive relations.
\begin{figure}
\centering
\vspace{0.2in}
\includegraphics[scale=0.92]{AverageFitTodata}
\caption{A comparison of our estimate \eqref{E:GotPi} of the average hyperon polarization (blue) to the STAR measurement \cite{STAR:2017ckg}. Since a magnetic field was not incorporated in our setting, we have compared our estimate to the average value of the polarization of $\Lambda$ and $\overline{\Lambda}$.}
\vspace{0.2in}
\label{fig1}
\end{figure}
\section{Discussion}
\label{sec::discuss}
In this paper, we initiated a fully fledged study of relativistic spin hydrodynamics. The hydrodynamic constitutive relations, relating the spin current and the stress tensor to fluid velocity, temperature, spin potentials and their derivatives, can be found in \eqref{E:constitutive}. For consistency, the constitutive relations for the symmetric part of the stress tensor and spin current were expanded to first order in derivatives while those for the antisymmetric part of the stress tensor were expanded to second order in derivatives.
This mismatch in the derivative expansion is a result of the hydrostatic equilibrium relation \eqref{E:hydrostaticidentities} between torsion, spin potential, vorticity and acceleration which implies that the spin potential and the longitudinal component of the torsion must be first order in derivatives.
In a background with vanishing torsion, such as the one discussed in this work, it is sensible to choose the transverse components of the torsion tensor to be first order, as is the case with the longitudinal components.
However, in certain condensed matter systems such as graphene torsion is used as the long-wavelength description of dislocations and disclinations in the atomic structure \cite{deJuan:2010zz,Mesaros:2009az}. In these systems a sensible choice might be to keep the spatial torsion at zeroth order in derivatives along with temperature and velocity. It would be interesting to develop such a torsio-hydrodynamic theory further, following the route we outlined here. Recent works
\cite{Bandurin2016, Crossno2016,Moll2016,Kumar2017,Guo2017} strongly suggest that graphene, as well as certain clean Dirac and Weyl semimetals, are well described by hydrodynamic theory
which further motivates this study.
Another possible direction is to extend our results to non parity invariant and non-conformal theories. Indeed, Bayesian analysis of heavy-ion data suggest that bulk viscosity may \cite{Bernhard:2016tnd,Everett:2020xug} or may not \cite{Nijs:2020ors,Nijs:2020roc} play an important role in the hydrodynamic description of heavy ion collisions. Finally, for off-central heavy ion collisions, which is the regime where spin hydrodynamics is most relevant, it is desirable to employ a more realistic hydrodynamic configuration in which the rotation symmetry around the beam axis is broken. In this context, extension of the recently introduced a hydrodynamic frame \cite{Bemfica:2017wps,Kovtun:2019hdm} that allows a consistent set of hyperbolic equations, to include spin currents is also desirable.
\section{Acknowledgements}
We thank F. Becattini, K. Fukushima, G. Torrieri and J. Zaanen for useful discussions. DG and UG are partially supported by the Delta-Institute
for Theoretical Physics (D-ITP) funded by the Dutch Ministry of Education, Culture and Science (OCW). In addition, DG is supported in part by CONACyT through the program Fomento, Desarrollo y Vinculacion de Recursos Humanos de Alto Nivel. AY is supported in part by an Israeli Science Foundation excellence center grant 2289/18 and a Binational Science Foundation grant 2016324.
|
\section{First-principles analysis}
We performed first-principles
calculations within the framework of the density functional
theory (DFT) as implemented in the in the
QUANTUM ESPRESSO suite~\cite{Giannozzi_2009,Giannozzi_2017}.
The generalized gradient approximation, as parametrized by the Perdew-Burke-Ernzerhof
generalized gradient approximation (PBE-GGA) functional for the exchange-correlation potential, was used by expanding
the Kohn-Sham wave functions into plane waves up to an
energy cutoff of \(100\,\text{Ry}\approx1.4\,\)keV
and sampling the Brillouin zone on a $4 \times 4 \times 2$ regular mesh~\cite{PhysRevLett.77.3865}.
For \ce{Ba2CuO4} we used the crystal structure parameters available from experiment~\cite{Li12156} (\(a=4.003\,\)\AA, \(c=12.94\)\AA).
Due to the absence of experimental data on the the \ce{LaAlO3}/\ce{LaNiO3} heterostructure (LAO/LNO) proposed in \cite{chaloupka_orbital_2008,hansmann_turning_2009},
we simulated its epitaxial growth by imposing an in-plane lattice constant of \(a=3.905\,\)\AA~and relaxing the out of plane lattice parameter as well as the
atomic positions on a refined $24 \times 24 \times 24$ regular mesh Brillouin zone mesh, resulting in \(c=7.500\,\)\AA.
The extraction of the two-orbital minimal
model used in the presented analysis of superconducting instabilities in
both materials was based on the Wannier functions formalism \cite{Marzari_2012}
and we fixed the Fermi level in our model by enforcing the filling in the
simplified model as opposed to the DFT Fermi level.
\section{Minimal model}
The minimal model to describe the low energy physics of both materials studied here is an extended two-band Hubbard model
on a square lattice. While this becomes apparent from the performed DFT calculations,
one can also arrive at this conclusion by simple chemistry considerations:
Both the LAO/LNO and \ce{Ba2CuO_{3+\delta}} feature a layered perovskite structure.
In case of LNO/LAO nickel oxide planes \ce{Ni^{3+}O2^{2-}} alternate
with an insulating aluminium oxide plane \ce{Al^{3+}O2^{2-}}, allowing
for a two-dimensional description of the low energy electronic degrees
of freedom located on the nickel ions
\cite{hansmann_turning_2009}. For BCO, copper oxide planes
\ce{Cu^{(2+2\delta)+} O2^{2-}}, which are pairwise shifted against
each other, alternate with two insulating barium oxide planes
\ce{Ba_{2}^{2+}O^{2-}}. As such, both systems yield a two-dimensional
description of the low energy electronic degrees of freedom located on
the copper or nickel ions \cite{maier_d_wave_2018,Li12156}.
As it is typical for perovskite structures, the transition metal ion is enclosed by an octahedron of
\ce{O^2-} ions. This results in a crystal field which splits the \(L=2\) d-orbital quintuplet
into a low energy \(t_{2g}\) triplet (composed of the \(d_{xy}\), \(d_{xz}\) and \(d_{yz}\) orbitals)
and an \(e_g\) doublet (composed of the remaining \(d_{x^2-y^2}\) and \(d_{3z^2-r^2}\)) with higher energy.
For both LNO/LAO (\ce{La^{3+}Ni^{3+}O3^{6-}}/\ce{La^{3+}Al^{3+}O3^{6-}}) and \ce{Ba2CuO_{4}} (\ce{Ba2^{4+}Cu^{4+}O4^{8-}})
the transition metal ions nickel and copper have a \(3d^7\) electronic configuration, while the \(3d^6\) configuration
on the aluminum site of LNO/LAO results in a significant excitation gap allowing us to remove it from the low energy description of the system.
In the case of BCO, we additionally simulate the oxygen deficiencies by artifically raising the Fermi level in our model compared to the \ce{Ba2CuO4} calculation, simulating the effect
of the added electrons by removing oxygen from the compound.
The completely filled \(t_{2g}\) multiplett is split of by the octahedral crystal field,
allowing us to restrict our analysis to the \(e_g\) doublet.
Additionally, degeneracy of the \(e_g\) states is lifted by distortions of
the oxygen octahedron. The strength of this Jahn-Teller distortion and the
resulting \(e_g\) splitting \(\epsilon\) is one of the key parameters resolved
in our work. We find our effective low-energy model to be formed by a two band Hubbard model
on a square lattice spanned by the transition metal's \(d_{x^2-y^2}\) (\(o=1\)) and \(d_{3z^2-r^2}\) (\(o=2\)) orbitals:
\begin{align}
\begin{split}
H_0 &=
\sum_i \sum_{o=1}^2 \left( (-1)^o \frac{\epsilon}{2} - \mu \right) c_{o,i}^{\dag} c_{o,i}^{\vphantom{\dag}}
\\&+ \sum_{i} \sum_{j\neq i}
\sum_{o',o=1}^2 t_{o'o,\alpha(i,j)} c_{o',i}^{\dag} c_{o,j}^{\vphantom{\dag}} \label{eq:appH0}
\end{split}
\\&= \sum_{i} \sum_{k}
\sum_{o',o=1}^2 \xi_{o'o}(\vec{k}) c_{o',k}^{\dag} c_{o,k}^{\vphantom{\dag}} \label{eq:appH0k}
\end{align}
with the single-particle bands
\begin{align}
\label{eq:bloch_intra}
\xi_{oo}(\vec{k})&=\left( (-1)^o \frac{\epsilon}{2} - \mu \right) +2t_{oo,1}(\cos k_{x}+\cos k_{y})\nonumber\\
&+4t_{oo,2}\cos k_{x}\cos k_{y}+2t_{oo,3}(\cos 2k_{x}+\cos 2k_{y})\nonumber,\\[10pt]
\xi_{oo'}^{o\neq o'}(\vec{k})&=\,2t_{oo',1}(\cos k_{x}-\cos k_{y})\nonumber\\[5pt]
&+2t_{oo',3}(\cos 2k_{x}-\cos 2k_{y})\nonumber,
\end{align}
where \(c_{o,i}^{\dag}\) creates an electron in the orbital \(o=1,2\) ($d_{x^2-y^2}$, $d_{3z^2-r^2}$) on site
\(\vec{r}_i\).
The index \(\alpha(i,j)\) counts the proximity of different sites \(i\) and \(j\)
and is used to label the corresponding orbital hybridizations \(t_{oo,\alpha}\).
The chemical potential \(\mu\) controls the filling of the system and \(\epsilon\) is the relative onsite energy of the relevant orbitals.
Tab.\,\ref{tab:parameters} lists the tight binding parameters for the minimal model that we obtained by fitting this model to the
results from our DFT calculations.
\begin{table}[!h]
\begin{tabular}{
l
S[table-figures-integer = 1, table-figures-decimal = 4, table-number-alignment = center]
S[table-figures-integer = 1, table-figures-decimal = 4, table-number-alignment = center]
S[table-figures-integer = 1, table-figures-decimal = 4, table-number-alignment = center]
}
\toprule
%
model parameters & {\ce{Ba2CuO_{3+\delta}}}
& {\ce{LaNiO3/LaAlO3}} \\
\midrule
{\(t_0\, / \)\,\si{\eV}} & -0.511
& -0.447 \\
{\(n\, / \)\,electrons} & 2.00
& 1.00 \\
\midrule
{\(\epsilon\, / \,|t_0|\)} & 1.71
& 0.24 \\
{\(\mu\, / \,|t_0|\)} & +0.82
& -1.81 \\
\midrule
\multicolumn{2}{l} {hopping amplitudes \(/\,|t_0| \) } & & \\
\midrule
{\(t_{11,\alpha=1}\)} & -1.00
& -1.00 \\
{\(t_{22,\alpha=1}\)} & -0.44
& -0.48 \\
{\(t_{12,\alpha=1}\)} & +0.63
& +0.62 \\
\midrule
{\(t_{11,\alpha=2}\)} & +0.13
& +0.13 \\
{\(t_{22,\alpha=2}\)} & -0.07
& -0.07 \\
{\(t_{12,\alpha=2}\)} & +0.00
& +0.00 \\
\midrule
{\(t_{11,\alpha=3}\)} & -0.28
& -0.13 \\
{\(t_{22,\alpha=3}\)} & -0.06
& -0.03 \\
{\(t_{12,\alpha=3}\)} & +0.11
& +0.06 \\
\bottomrule
\end{tabular}
\caption{
Parameters for the tight binding Hamiltonian \eqref{eq:appH0} for both LNO/LAO and BCO. \(t_{oo',\alpha}\) denotes the hybridization between two orbitals \(o\) and \(o'\) that are
\(\alpha\)\textsuperscript{th} nearest neighbours (\(\alpha = 1\) corresponds to nearest neigbour, \(\alpha = 2\) to next-nearest neigbour etc.).
}
\label{tab:parameters}
\end{table}
\section{Methodology}
In this section we aim to give a short overview on different diagrammatic schemes to calculate the effective two-particle vertex function, the central object to reveal the superconducting instabilities.
Specifically, we perform Kohn-Luttinger type calculations\,\cite{PhysRevLett.15.524,raghu_superconductivity_2010} in the weak coupling (referred to as weak-coupling renormalization group Kohn-Luttinger analysis (KL) in the following) and through functional renormalization group (FRG)\,\cite{RevModPhys.84.299,platt_functional_2013} as well as random phase approximation (RPA)\,\cite{Altmeyer2016} in the intermediate coupling regime.
Fig.\,\ref{Fig:App:diagram} shows the perturbative expansion of the effective two-particle vertex $\Gamma_{\text{eff}}$ in terms of the bare two-particle vertex $\Gamma_b$. We will show how one can use these diagrams to calculate $\Gamma_{\text{eff}}$ within KL analysis and RPA. Further we give a short derivation of the fRG method, which features the same topological diagrams as the KL analysis, in order to solve a differential equation to determine $\Gamma^{\Lambda_c}$, where $\Lambda_c$ is the critical energy scale of the system.
Clearly all methods employed here include different sets of diagrams, resulting in a significant difference of captured screening effects. Therefore it is necessary to taylor the initial interaction strength $U$ for each approximation scheme individually, in order to obtain comparable results.
A more detailed Discussion is presented in the References \cite{PhysRevB.84.235121,raghu_superconductivity_2010}.
Here we choose \(U = 3.5\,\mbox{eV}\) for the fRG calculations and \(U = 0.8\,\mbox{eV}\) for the RPA calculation and therefore account for the additional screening processes include in the fRG approximation.
\subsection{Functional Renormalization Group}
\textit{Method.}---The here presented summary of the FRG method follows Ref.\,\onlinecite{platt_functional_2013}, which provides a detailed derivation of the flow equation.
First we set up our theory at a high energy scale $\Lambda_{\text{init}}$ with the action
\begin{equation*}
S^{\Lambda_{\text{init}}} = S_0 +S_I ,
\end{equation*}
containing the bare action $S_0$ and an interaction term $S_I$.
In order to connect to lower energy scales we integrate out high-energy modes step by step, adjusting the action accordingly and generating a trajectory between microscopic theory at high energies and an effective low-energy description.
We achieve this by introducing a flow parameter $\Lambda$ interpolating the action from the trivial stating point ($\Lambda = \Lambda_{\text{init}}$) to the fully interacting theory ($\Lambda = 0$).
One can deduce an set of integro-differential equation for all irreducible $2n$-point vertex functions. It turns out that the hierarchy of these so called flow equations does not close, meaning that in order to calculate the flow for an $2n$-point vertex function, one has to know the
vertex function of the next higher order. Consequentially, to numerically solve the integro-differential equation, the flow equation for the four-point vertex function is truncated by neglecting all terms containing higher order vertex functions \cite{RevModPhys.84.299,Salmhofer_fRG}. Also we neglect any corrections
of the self energy $\Sigma$ under the flow. The remaining integro-differential equation
is of first order. The derivative of the irreducible four-point vertex function $\Gamma^{(4),\Lambda}_{k_1',k_2';k_1,k_2}$ with respect to $\Lambda$ turns out to be equal to diagrams which are topologically equivalent to $2a-2c$ in Fig.\,\ref{Fig:App:diagram}, the only difference being that one of the internal Green's functions $G^\Lambda$ gets promoted to a
single scale propagator $S^\Lambda = \partial_\Lambda G^\Lambda|_{\Sigma=\text{const}}$.
\begin{widetext}
\begin{align*}
\frac{d}{d\Lambda}\Gamma^{(4),\Lambda}_{k_1',k_2';k_1,k_2} = \sum_{\substack{k,k' \\ q,q'}}G^\Lambda_{k,k'}S^\Lambda_{q,q'}\Big ( \Gamma^{(4),\Lambda}_{k_1',k_2';k,q} \Gamma^{(4),\Lambda}_{k',q';k_1,k_2}
&- \big [\Gamma^{(4),\Lambda}_{k_1',q';k_1,k} \Gamma^{(4),\Lambda}_{k',k_2';kq,k_2} + (k \leftrightarrow q, k' \leftrightarrow q')\big] \\
&+ \big [\Gamma^{(4),\Lambda}_{k_2',q';k_1,k'} \Gamma^{(4),\Lambda}_{k',k_1';q,k_2} + (k \leftrightarrow q, k' \leftrightarrow q')\big] \Big).
\end{align*}
\end{widetext}
An index $k$ contains the spin $\sigma$, orbital (band) $o$ ($b$), momentum $\boldsymbol k$ and frequency $k_0$ degrees of freedom; the $k$ index on the left corresponds to an in-going particle, while the one on the right to an out-going particle. We further simplify the numerical effort by projecting all momenta to the Fermi surface and compute the flow neglecting all finite frequency contributions.
There are multiple schemes to implement the flow parameter. Wilson's original idea of integrating out momentum modes shell by shell offers one possible implementation \cite{PhysRevA.8.401,POLCHINSKI1984269,WETTERICH199390}, by introducing an energy cutoff depending on $\Lambda$ in the propagators. However this cutoff scheme is
not suited to treat particle-hole fluctuations in an unbiased way\,\cite{platt_functional_2013}. Alternatively, temperature can be used as the flow parameter, as utilized in this work. Doing so not only avoids any issues with the particle hole fluctuations but also offers a very intuitive picture of understanding
the flow as a whole. Integrating out energy modes, descending from high to low energies, can now be interpreted as cooling down our system.
Using the Euler method, we can solve the integro-differential equation. The initial value for the action is chosen to be equivalent to the bare action
\begin{equation}
S^{\Lambda_{\text{init}}} = S_0
\end{equation}
and the initial four-point vertex function at $\Lambda_{\text{init}}$ is set to be the interaction term of the unrenormalized theory, which than flows under the
set up formalism either to an new effective function at $\Lambda=0$ or to a fix point at a given $\Lambda_c$ at which the vertex diverges and the flow breaks down. The later case marks the breakdown of the Fermi surface to a new symmetry broken phase. The order parameter of
the new phase can be analyzed by decomposing the full vertex function into mean field channels. The leading instability is then given by the most diverging of these channels. Analogously to RPA and KL analysis one can further characterize the order parameter according to its transformation behavior under the lattices symmetries, as demonstrated in the main paper.
\textit{Results.}---Fig.\,\ref{fig:Phasediagramm} shows the critical cutoff energy $\Lambda_c$ for the two relevant models.
The superconducting channel of the meanfield decoupled effective vertex $\Gamma^{Sc,\Lambda_c}(k,k')$ at $\Lambda_c$ can further be analysed by solving the eigenvalue equation
\begin{equation}
\Delta_k\lambda= \sum_{k'}\Gamma^{Sc,\Lambda_c}(k,k')\Delta_{k'} \quad \text{with} \quad \Lambda_c \propto T_c
\end{equation}
and identifying the symmetry class of the from factor $\Delta_k$ for the smallest eigenvalue $\lambda$.
\begin{figure}[t!]
\centering
\includegraphics[width=1.\columnwidth]{Phasediagramm_fRG}
\caption{\label{fig:Phasediagramm} Critical cutoff $\Lambda_c$ of fRG-flow and symmetry character of leading and subleading superconducting order parameter for for BCO (unfilled markers) and LNO/LAO (solid markers). The cutoff for the subleading order parameter is estimated by extrapolating the divergence in $\Lambda$ beyond $\Lambda_c$ of the corresponding eigenvalue $\lambda$ . The interaction is given by $U=4J=4J'=3.5\,\mbox{eV}$ while the different values for $V/U$ is achieved by increasing \(V\). The dashed line indicates the generic interaction ratio $U=2V=4J=4J'$.}
\end{figure}
By doing so we can identify two regimes: The first being the case of $V \geq U$, a regime featuring strong repulsive interaction between the pockets of the Fermi surface (FS), which are dominated by opposite orbital character, resulting in an extended $s$-wave ($A_1$ irrep), featuring nodes between the pockets.
The other being $V\ll U$, where intra-pocket scattering makes up for nearly all scattering processes and thus features nodal lines intersecting the FS. In case of \ce{Ba2CuO_{3+\delta}} the leading symmetry class is given by $B_1$ irrep ($d_{x^2-y^2}$-wave), which is the well known $d_{x^2-y^2}$-wave present in most cuprates, to accommodate the large intra orbital nesting $q=(\pi,\pi)$. For \ce{LaNiO3 / LaAlO3} the nature of the incommensurate nesting vectors
would favour nodes between the parallel lines of the outer and inner pocket of the Brillouin zone (BZ). The resulting nodal configuration can be accommodated by a form factor of the symmetry class $B_2$ ($d_{xy}$-wave) and the subleading form factor of the symmetry class $A_2$ ($g$-wave).
In between these two regimes the critical cutoff for \ce{Ba2CuO_{3+\delta}} is nearly constant. By increasing the inter orbital interaction nesting between the parallel lines of the FS of different pockets, connected by transition vectors around $\bs q =(\pi,\pi)$, become relevant, hence the symmetry remains $B_1$. Note that the nesting vectors around $\bs q =(\pi,\pi)$ are yielding favourable results in the gap equation for the $B_1$ and $A_1$ order parameters.
Hence, the transition between these two phases by increasing inter-orbital interactions is quite generic: one just chooses between the nodal configuration which achieves a sign change between the maximal subset of nesting vectors weighted with their interaction strength. This can also be seen in the weak coupling analysis below and in the RPA results from Maier {\it et al.}\,\cite{maier_d_wave_2018}, where the eigenvalues associated with $B_1$ and $A_1$ irreps are nearly degenerate. In fRG these two irreps make up the leading and first subleading symmetry classification of the order parameter for all probed values of $V/U$ with a flow that indicates nearly simultaneous divergences of the respective eigenvalues.
In contrast, for \ce{LaNiO3/LaAlO3} there is no such sequence of phases sharing advantageous nesting options at the vicinity of the phase transition. Its intermediate interaction regime offers no clear favoured nesting option and its unclean orbital makeup yields a uniform interaction between different points of the FS. Hence, it features a dramatically decreased $\Lambda_c$ and a form factor of the symmetry class $E$ ($p$-wave). This decrease is remarkable for two reasons. Firstly, it is counter intuitive in the sense that we add interaction to the system and decrease $T_c${} by up to two orders of magnitude. Secondly, the resulting form factor of symmetry class $E$ has two representations (\textit{i.e.},\ the irrep is two-dimensional), which can be added in such a way that the resulting order parameter leads to a hard gap. The $B_2$ $d$-wave is demoted to subleading irrep in this regime, due to the large additional interaction between the two Fermi pockets, which is not aligned with a nodal configuration provided by a $d_{xy}$ order parameter. For $V\approx 0.7~U$ the $s$-wave becomes the first sublading irrep.
\begin{figure}[t!]
\centering
\includegraphics[width=1.\columnwidth]{Comparison_FS}
\caption{\label{fig:Comparison_FS} Bare susceptibility and corresponding FS for \ce{LaNiO3/LaAlO3} at filling $n=1.1$ (solid lines) and $n=1.0$ (dashed line). The respective critical cutoffs for the generic interaction ratio $U=2V=4J=4J'$ and $U=3.5\,\mbox{eV}$ are given by $\Lambda_c/U = 7.1\cdot 10^{-3}$ ($n=1.1$) and $\Lambda_c/U = 4.2\cdot 10^{-3}$ ($n=1.0$).}
\end{figure}
Overall it seems very reminiscent to frustration effects in magnetism, although this is not related to the geometry of the system, but instead to the relative mixture of different interactions. Following that line of reasoning one can think of \ce{LaNiO3 / LaAlO3} as a {\it frustrated superconductor}.
Electronic structure calculations performed in Ref. \cite{hansmann_turning_2009}, suggested a larger inner FS pocket compared to our results. Increasing the filling of the presented minimal model we find a similar FS (compare Figure \ref{fig:Comparison_FS}). The critical cutoffs fRG $\Lambda_c$ of both fillings are of the same order of magnitude and yield a form factor of symmetry class $E$, thereby demonstrating the independence of the presented many body results from this band structure detail.
\subsection{Random Phase Approximation}
\textit{Method.}---The RPA formulation is based on the idea that selected particle-hole scattering events add up coherently, whereas all other possible scattering channels are suppressed by acquiring random relative phases. Neglecting these terms in the calculation of the effective two-particle vertex results in a summation up to infinite order of pure bubble and ladder diagrams. For the single orbital case both interactions and susceptibilities are scalar. In contrast the multiorbital case the pairing vertex has an additional contraction over the orbital degrees of freedom. This orbital makeup induces additional diagrams with the structure of vertex corrections which are included in the matrix-RPA formulation\,\cite{Altmeyer2016}. A graphic representation of the considered terms are the particle-hole diagrams shown in Fig.\,\ref{Fig:App:diagram} ($1$, $2$b, $2$c, $3$a, ...).
Firstly we define the bare susceptibility as
\begin{align*}
\chi^{0}_{o_1o_2o_3o_4}(\bm{q},\tau)=\frac{1}{N}\sum_{\bm{k}\bm{k}'}\langle T_{\tau} &c^{\dag}_{l_3,\bm{k}+\bm{q},\sigma}(\tau)c_{l_4,\bm{k},\sigma}(\tau) \ \times \\ &c^{\dag}_{l_2,\bm{k}'-\bm{q},\sigma}(0)c_{l_1,\bm{k}',\sigma}(0) \rangle_0 ,
\end{align*}
where $o_i$ are the orbital indices. The bare susceptibility in momentum-frequency space is
\begin{widetext}
\begin{align*}
&\chi^0_{o_1o_2o_3o_4}(\bm{q},i\omega_n)=-\frac{1}{N}\sum_{\mu\nu}\sum_{\bm{k}}a^{o_4}_\mu(\bm{k})a^{o_2*}_{\mu}(\bm{k}) a^{o_1}_\nu(\bm{k}+\bm{q}) a^{o_3*}_{\nu}(\bm{k}+\bm{q})
\frac{n_F(E_{\mu}(\bm{k}))-n_F(E_{\nu}(\bm{k}+\bm{q}))}{i\omega_n+E_{\mu}(\bm{k})-E_{\nu}(\bm{k}+\bm{q})}\ .
\end{align*}
\end{widetext}
where $\mu/\nu$ is the band index. $n_F(\epsilon)$ is the Fermi distribution function, $a^{o_i}_\mu(\bm{k})$ is the $o_i$-th component of the eigenvector for band $\mu$ resulting from the diagonalization of the single-particle Hamiltonian $H_0$, and $E_{\mu}(\bf{k})$ is the eigenvalue of band $\mu$. The intrinsic spin fluctuations are characterized by the susceptibility. The interacting spin susceptibility and charge susceptibility in RPA level are given by,
\begin{align*}
\chi^{RPA}_1(\bm{q})&=[1-\chi_0(\bm{q})U^s]^{-1}\chi_0(\bm{q})\ ,\\[5pt]
\chi^{RPA}_0(\bm{q})&=[1+\chi_0(\bm{q})U^c]^{-1}\chi_0(\bm{q}) .
\end{align*}
Here $U^s$, $U^c$ are the interaction matrices, which are given by
\begin{align*}
\bar{U}^s_{o_1o_2o_3o_4}&=
\begin{cases}
U \phantom{~+2J+} & o_1=o_2=o_3=o_4,\\
V & o_1=o_3\neq o_2=o_4,\\
J & o_1=o_2\neq o_3=o_4,\\
J' & o_1=o_4\neq o_2=o_3,\\
\end{cases}\\
\bar{U}^c_{o_1o_2o_3o_4}&=
\begin{cases}
U & o_1=o_2=o_3=o_4,\\
-V+2J & o_1=o_3\neq o_2=o_4,\\
2\,V-J & o_1=o_2\neq o_3=o_4,\\
J' & o_1=o_4\neq o_2=o_3,\\
\end{cases}
\end{align*}
where we used the notation for the Kanamori interaction parameters introduced in the main text.
The effective interaction obtained in the RPA approximation is given by
\begin{align*} V_{\rm eff}=\sum_{ij,\textbf{k}\textbf{k}'}\Gamma_{ij}(\textbf{k},\textbf{k}')c^{\dag}_{i\textbf{k}\uparrow}c^{\dag}_{i-\textbf{k}\downarrow}c_{j-\textbf{k}'\downarrow}c_{j\textbf{k}'\uparrow}
\end{align*}
where the momenta $\textbf{k}$ and $\textbf{k}'$ are restricted to different FS $C_i$ with $\textbf{k}\in C_i$ and $\textbf{k}'\in C_j$, and $\Gamma_{ij}(\textbf{k},\textbf{k}')$ is the pairing scattering vertex. The pairing vertex can be obtained by projecting the pairing vertex in orbital space onto Fermi surfaces,
\begin{align*}
&\Gamma_{ij}(\textbf{k},\textbf{k}')= \sum_{\substack{o_1,o_2 \\ o_3,o_4}}a^{o_2*}_{v_i}(\textbf{k}) a^{o_3*}_{v_i}(-\textbf{k}) \ \times \\
&\qquad\qquad \text{Re}[\Gamma_{o_1 o_2 o_3 o_4}(\textbf{k},\textbf{k}',\omega=0)] a^{o_1}_{v_j}(\textbf{k}') a^{o_4}_{v_j}(-\textbf{k}').
\end{align*}
The orbital vertex function $\Gamma_{o_1 o_2 o_3 o_4}$ for the singlet channel and triplet channel in the fluctuation exchange formulation\cite{TTakimoto-PRB-69-104504,KKubo-PRB-75-224509,SGraser-NJP-11-025016,AFKemper-NJP-12-073030} are given by
\begin{widetext}
\begin{align*}
\Gamma^S_{o_1 o_2 o_3 o_4}(\textbf{k}, \textbf{k}',\omega)&=[\frac{3}{2}\bar{U}^s \chi^{RPA}_1(\textbf{k}-\textbf{k}',\omega)\bar{U}^s + \frac{1}{2}\bar{U}^s -\frac{1}{2}\bar{U}^c\chi^{RPA}_0(\textbf{k}-\textbf{k}',\omega)\bar{U}^c+\frac{1}{2}\bar{U}^c]_{o_1 o_2 o_3 o4},\\[10pt]
\Gamma^T_{o_1 o_2 o_3 o_4}(\textbf{k}, \textbf{k}',\omega)&=[-\frac{1}{2}\bar{U}^s \chi^{RPA}_1(\textbf{k}-\textbf{k}',\omega)\bar{U}^s+\frac{1}{2}\bar{U}^s -\frac{1}{2}\bar{U}^c\chi^{RPA}_0(\textbf{k}-\textbf{k}',\omega)\bar{U}^c+\frac{1}{2}\bar{U}^c]_{o_1 o_2 o_3 o_4},
\end{align*}
\end{widetext}
where $\bar{U}^{s/c}=U^{s/c}(\bm{k}-\bm{k}')$.
$\chi^{RPA}_0$ describes the charge fluctuation contribution and $\chi^{RPA}_1$ the spin fluctuation contribution. For a given gap function $g(\textbf{k})$, the pairing strength functional is
\begin{align*}
\lambda[g(\textbf{k})]=-\frac{\sum_{ij}\oint_{C_i} \frac{dk_{\|}}{v_F(\textbf{k})} \oint_{C_j} \frac{dk'_{\|}}{v_F(\textbf{k}')} g(\textbf{k})\Gamma_{ij}(\textbf{k},\textbf{k}') g(\textbf{k}')} {4\pi^2\sum_i\oint_{C_i} \frac{dk_{\|}}{v_F(\textbf{k})} [g(\textbf{k})]^2 },
\end{align*}
where $v_F(\textbf{k})=|\triangledown_{\textbf{k}}E_i(\textbf{k})|$ is the Fermi velocity on a given Fermi surface sheet $C_i$. From the stationary condition we find the following eigenvalue problem:
\begin{align*}
-\sum_{j} \oint_{C_j} \frac{dk'_{\|}}{4\pi^2v_F(\textbf{k}')} \Gamma_{ij}(\textbf{k},\textbf{k}') g_{\alpha}(\textbf{k}')=\lambda_{\alpha}g_{\alpha}(\textbf{k}),
\end{align*}
where the interaction $\Gamma_{ij}$ is the symmetric (antisymmetric) part of the full interaction in the singlet (triplet) channel. The leading eigenfunction $g_{\alpha}(\bf{k})$ and eigenvalue $\lambda_{\alpha}$ are obtained from the above equation. The obtained gap function should have the symmetry of one of the irreducible representations for the corresponding point group.
\begin{figure}[t!]
\includegraphics[width=1\columnwidth]{Phasediagram_RPA.pdf}
\caption{\label{fig:PhasediagrammRPA} Leading and subleading eigenvalues $\lambda$ within RPA and symmetry character of the superconducting order parameter of the effective four point RPA vertex $\Gamma$ for LNO/LAO (solid markers) and for BCO (unfilled markers). The interaction is given by $U=4J=4J'=0.8\,\mbox{eV}$ and increasing $V$ and $T=0.02\,$eV. The dashed line indicates the the generic interaction ratio $U=2V=4J=4J'$. For BCO $\Gamma$ diverges for $V/U=1$, signalling the onset of a spin density wave.}
\end{figure}
\textit{Results.}---We show the leading and first subleading eigenvalues calculated within the RPA-matrix formalism for \ce{LaNiO3} and \ce{Ba2CuO_{3+\delta}} in Fig.\,\ref{fig:PhasediagrammRPA}.
The eigenvalues of \ce{Ba2CuO_{3+\delta}} are for the whole parameter regime much larger than \ce{LaNiO3}, due to the overall better nesting. As expected, a large regime of interactions is favoring $d$-wave pairing ($B_1$ irrep), given the large signal in the bare susceptibility. For $V/U > 1$ the extended $s$-wave pairing ($A_1$ irrep), featuring nodes between the Fermi pockets, becomes leading.
In case of \ce{LaNiO3}, there is no clear peak in the bare susceptibility, further the inner Fermi pocket around the $\Gamma$-point is largely inert to superconductivity in a repulsive interaction environment, since it features no density fluctuations, is clean in orbital weight, and is in itself not nested. Due to the density distribution of the outer Fermi sheet around the $M$ point, which is enlarged at the edges of the BZ, the resulting $B_1$ symmetry character is reminiscent of the one band cuprate models. Increasing inter-orbital interactions enlarges the coupling between the parallel lines on the outer Fermi pocket. Hence, additional nodal lines are required, resulting in a $B_2$ or $A_2$ symmetry character. Note that there is, similarly to the results of fRG, a pronounced dip in the eigenvalues of LNO/LAO starting at the pure intra-orbital case and increasing inter-orbital interactions.
Using the eigenvalues from the RPA calculations, we can estimate the transition temperature as $T_c\approx\hbar\omega_Se^{-1/|\lambda_{\rm min}|}$, where $\hbar\omega_S$ is the typical energy scale for spin fluctuations in systems. The T$_c$ of BCO is about 70 K in experiments. To reproduce the experimental value approximately, we take $U=1$ eV, $J/U=0.25$, $\hbar\omega_S=100$ meV and the temperature parameter $T=0.03$ eV in the Fermi distribution function and then get the dominant pairing state with $\lambda_{\rm min}=-0.448$ for BCO, leading to the $T_c$ of the order of 100 K. With the same parameters in LNO/LAO, the dominant pairing eigenvalue is 0.005, much smaller than the case of BCO.
\subsection{Weak coupling Renormalization Group - Kohn-Luttinger method}
\textit{Method.}---Here, we give a short description of the (KL) method; a detailed discussion can be found in various works \cite{raghu_superconductivity_2010,raghu_superconductivity_2011,raghu_effects_2012,cho_band_2013,wolf_unconventional_2018,wolf-20arXiv2004.12624}. Note that one of the most important characteristics of the method is that it becomes asymptotically exact in the limit of infinitesimal coupling, $U\rightarrow 0$.
For the KL analysis, we consider all terms in the perturbative expansion of the effective two-particle-vertex up to second order in the interaction, $U$, which is given by the first four diagrams in Fig.\,\ref{Fig:App:diagram}, \textit{i.e.},\ diagrams 1, 2a, 2b, and 2c.
In contrast to fRG, where higher order diagrams are considered gradually during the flow of the effective interaction parameters, we cut off all diagrams that are of higher than second order in $U$. In RPA, on the contrary, we sum over a geometric series of only ladder and bubble diagrams, \textit{i.e.},\ up to infinite order in $U$.
\begin{figure}[b!]
\includegraphics[width=0.9\columnwidth]{WCRG_vertex_2nd_order.pdf}
\caption{Spinless particle-hole diagrams of the weak coupling expansion of the effective two-particle-vertex in second order in the interaction $U$. Solid lines represent free particle propagators and dashed lines interactions $U$.}
\label{Fig:App:WCRG_diagram}
\end{figure}
The RG flow equation is given by \cite{shankar_renormalization-group_1994,raghu_superconductivity_2010,wolf_unconventional_2018}
\begin{equation}
\label{Eq:App:WCRG_flow}
\frac{\partial \Gamma(k_{2},k_{1})}{\partial \ln(\Omega_0/\Omega)}=-\int_{\rm FS}{\rm d}k_{3}\,\Gamma(k_{2},k_{3})\Gamma(k_{3},k_{1}),
\end{equation}
where $\Gamma$ denotes the effective two-particle-vertex, and $k_{1}$ and $k_{2}$ denote the momenta of the incoming and outgoing pairs of electrons with zero total momentum, respectively. The right hand side is integrated over all Fermi surfaces and we used the short notation $k_{i}=(n_{i},\vec{k}_{i})$ with the band index $n_{i}$. $\Omega_0$ denotes the initial infrared cutoff and $\Omega$ a lowered cutoff. The initial cutoff is chosen in the range
\begin{equation}
\label{eq:cutoff_range}
U^{2}/W\gg\Omega_{0}\gg We^{-1/\rho|U|},
\end{equation}
where $W$ is the bandwidth.
The lower bound ensures that the bare interactions are renormalized due to many-body effects, whereas the upper bound limits all involved modes in the effective interaction to a narrow window around the Fermi surface\,\cite{raghu_superconductivity_2010}.
In the weak coupling limit we can bypass the procedure of the RG flow, \textit{i.e.},\ lowering the cutoff $\Omega$ in consecutive steps, and obtain the final $\Gamma$ directly, where Eq.\,\eqref{eq:cutoff_range} ensures that the final result does not depend on the initial cutoff, $\Omega_{0}$\,\cite{raghu_superconductivity_2010}. All terms which contribute in this manner are the diagrams 2b and 2c in Fig.\,\ref{Fig:App:diagram}, where we use their spinless variants which are given by the four diagrams shown in Fig.\,\ref{Fig:App:WCRG_diagram} (the first order in $U$ diagram just suppresses the plain $s$-wave solution and can thus be neglected).
\begin{figure}[t]
\includegraphics[width=1\columnwidth]{Phasediagram_WC.pdf}
\caption{\label{fig:Phasediagramm_wcRG} KL Results: leading and subleading Eigenvalues $\lambda$ and symmetry character of leading superconducting order of effective four point vertex $\Gamma$ forLNO/LAO (solid markers) and BCO (unfilled markers). The interaction is given by $U=4J=4J'$ and increasing $V$. The dashed line indicates the the generic interaction ratio $U=2V=4J=4J'$.}
\end{figure}
In Eq.\,\eqref{Eq:App:WCRG_flow} $\Gamma$ is scaled such that its eigenvalues $\lambda_i$, independently, fulfil
\begin{equation}
\frac{\partial \lambda_i}{\partial\ln(\Omega_{0}/\Omega)}=-\lambda_i^{2}.
\end{equation}
Hence, finding the leading superconducting instability simplifies to an eigenvalue problem for $\Gamma$. The effective interaction, $V_{\rm eff}=\lambda_{\rm min}/\rho$, of the leading superconducting instability is then obtained from the most negative eigenvalue, $\lambda_{\rm min}$ ($\rho$ is the total density of states at the Fermi level), and the relation to the critical temperature is given by
\begin{equation*}
T_{c}\sim e^{-1/\rho |V_{\rm eff}|}=e^{-1/|\lambda_{\rm min}|}.
\end{equation*}
The corresponding eigenfunction, $\psi_{\rm min}$, yields the formfactor of the order parameter, which can be classified by the irreducible representations of the symmetry group of the crystal lattice (analogous to the fRG as shown in the main paper).
\textit{Results.}---The eigenvalues of the leading superconducting instabilities for each irreducible representation obtained from the KL formalism are shown in Fig.\,\ref{fig:Phasediagramm_wcRG}. Here we plot the eigenvalue $\lambda_{\rm min}$ as a function of the interorbital interaction, $V$, while keeping the other interactions at $U=4J=4J'$.
Firstly, we note that $|\lambda_{\rm min}|$ for Ba$_{2}$CuO$_{3+\delta}$ is larger than for LaNiO$_{3}$ over the whole parameter range. For both systems, an extended $s$-wave solution ($A_1$ irrep) becomes dominant when $V/U$ approaches $1$, in agreement with the fRG and RPA results.
For Ba$_{2}$CuO$_{3+\delta}$, this s-wave ($A_{1}$ irrep) is also the leading instability for an extended range around $U=2V$. For small $V$, there is a transition to a $d_{x^2-y^2}$-wave state ($B_1$ irrep).
Note that for fRG, for which we used the highest bare interaction the regime of the $A_{1}$ irrep in the $V/U$ phase diagram is the smallest compared to the other methods (transition at $V/U\gtrsim 0.9$). In RPA this regime is enlarged (transition at $V/U\approx 0.5 \,...\, 0.6$) and the KL analysis, where U is infinitesimal yields the largest domain with a leading s-wave (transition at $V/U\approx 0.2 \,...\, 0.3$)
For LaNiO$_{3}$ we find a $p$-wave ($E$ irrep) solution for most of the shown parameter range, until the transition towards extended $s$-wave at large values for $V$.
As mentioned in the main paper, due to the asymptotic character, \textit{i.e.},\ the vanishing interacting strength, the KL analysis is extremely sensitive to FS fluctuations; thus differences for the Nickelates compared to RPA and fRG are expected.
|
\section{Bubble Radius and Structural Bias}\label{sec:bubblebias}
We introduce the \emph{(Polarized) Bubble Radius} to quantify how \emph{likely}
users starting their random walk on a vertex $v \in V$ of one color, are to hit
a vertex of the other color in at most $t$ steps.
\begin{definition}\label{def:bubbleradius}
The \emph{(Polarized) Bubble Radius (BR) $\bubble{G}{v}{t}$ of $v$ with
exploration parameter $t$} is
\[
\bubble{G}{v}{t} \doteq \expect{G}{\ttime{t}{v}{\bar{C}_v}} \enspace.
\]
\end{definition}
A random walk starting at a vertex $v$ with high BR is unlikely to hit a vertex
in $\bar{C}_v$ in fewer-than-or-exactly $t$ steps. The following
lemma formalizes this idea on common models for web browsing (random walks with
restarts or with back button~\citep{rwwithbackbutton,rwwithrestart}).
\begin{restatable}{lemma}{lemrestart}\label{lem:restart}
Let $r \in \mathbb{N}$, and consider a user who starts their random walk at
$v \in V$ and may either restart their walk from $v$ or hit the back button up
to $r$ times. Let $\mathcal{T}_v$ be the random variable denoting the number
of steps such user takes to hit a vertex in $\bar{C}_v$. If
$\bubble{G}{v}{t} \geq t (1 - \nicefrac{1}{8r})$, then
$\mathbb{P}\left(\mathcal{T}_v\leq t/2 \right) \leq 1/4$. If instead
$\bubble{G}{v}{t} \leq b$ for some $b > 0$, then $\mathbb{P}\left(\mathcal{T}_v > 4br
\right) \leq 1/4$.
\end{restatable}
Given $t$, it is easy to estimate $\bubble{G}{v}{t}$ for each vertex $v
\in V$ by sampling random walks from $v$. The following result, whose proof uses
the Hoeffding's bound and the union bound, shows the trade-off between the
number of sampled random walks and the accuracy in estimating the BR of $v$.
\begin{restatable}{lemma}{lembubbleapprox}\label{lem:bubbleapprox}
For each $v \in V$, let $w^{(v)}_1,w^{(v)}_2,\dotsc, w^{(v)}_r$ be $r$ random
walks from $v$ and stopped either when they hit a vertex of color
$\ovcolor{v}$ or when they run for $t$ steps, whichever happens first. For
$i=1,\dotsc,r$, let $b^{(v)}_i$ be the length of random walk $w^{(v)}_i$. Let
\[
\bar{B}(v) \doteq \frac{1}{r} \sum_{i=1}^r b^{(v)}_i \enspace.
\]
Let $\varepsilon, \delta \in (0,1)$. If $ r \ge
\frac{t^2}{\varepsilon^2} \ln \frac{2n}{\delta},
$ then
\[
\mathbb{P} \left( \exists v \in V\ \text{s.t.}\ \abs{\bubble{G}{v}{t} -
\bar{B}(v)} > \varepsilon \right) < \delta,
\]
where the probability is over the choice of the random walks.
\end{restatable}
In the rest of the work, we assume for simplicity to have access to the
\emph{exact} BR of every vertex. The above result makes this assumption
reasonable because computing approximations of extremely high quality is
relatively inexpensive.
\paragraph{The structural bias
On the basis of the BR, we define two sets of vertices: \emph{cosmopolitan} and
\emph{parochial}. Given two reals $\goodthres$ and $\badthres$ with
$1\le \goodthres < \badthres \le t$, the set $\good{G}$ of \emph{cosmopolitan}
vertices contains all and only the vertices in $G$ with BR \emph{at most}
$\goodthres$, and the set $\bad{G}$ of \emph{parochial} vertices contains all
and only the vertices in $G$ with BR at least $\badthres$. For ease of
notation, we do not include $\goodthres$ and $\badthres$ in the notation for
$\good{G}$ and $\bad{G}$. In the rest of this work, we assume for simplicity
$\goodthres=2$ and $\badthres=t/2$, but this assumption can be easily removed.
$\good{G}$ and $\bad{G}$ are \emph{disjoint}, but they do not necessarily form a
partitioning of $V$. We will often consider the partitioning of $\bad{G}$ by
color, i.e., the two sets $\badc{R}{G}$ and $\badc{B}{G}$, containing the
parochial vertices of color $R$ or $B$ respectively.
\begin{definition}
The \emph{structural bias} $\bias{G}$ of $G$ is the sum of the BRs of the
parochial nodes of $G$, i.e.,
\begin{equation}\label{eq:bias}
\bias{G} \doteq \sum_{v \in \bad{G}} \bubble{G}{v}{t} \enspace.
\end{equation}
\end{definition}
It is reasonable to consider only the parochial nodes in the definition of
structural bias because they are the ones such that a random walk from them is
very unlikely to hit any vertex of color different than the starting vertex
(see also \cref{lem:restart}).
Our goal in this work is to find a set of edges with extrema of different color
whose addition to $G$ would decrease the structural bias of the network. It is
reasonable to only consider edge with extrema of different color, as they are
always preferable (i.e., will result in a higher decrease of the structural
bias) than edges with monochromatic extrema: the addition of the new edge can
only have positive impact on the parochial vertices of the same color as the
source, and has no impact on the parochial vertices of the other color. If we
could add \emph{any number} of such edges to $G$, it would be easy to bring the
structural bias of $G$ to zero, as there would be no parochial nodes left. This
assumption is not realistic: the number of links that a website editor can add
to a single page and to the whole graph is limited by many factors, such as the
fact that a human-readable page cannot have too many links, and the fact that
the editor can only spend a limited time on this activity. Nevertheless, ideally
one would want to solve the following problem.
\begin{problem}\label{prob:zerobias}
Given a color $C \in \{R, B \}$, find the smallest set $A$ of pairs of
distinct edges $(v, w) \notin E$ with $\vcolor{v} = C$ and $\vcolor{w} \neq C$
such that, for the graph $\Gnew = (V, E \cup A) $ it holds
$\badc{C}{\Gnew}=\emptyset$.
\end{problem}
\begin{restatable}{lemma}{lemreduction}\label{lem:reduction}
\Cref{prob:zerobias} is NP-hard and APX-hard.
\end{restatable}
\section{Reducing the BR with insertions}\label{sec:algo}
Since \cref{prob:zerobias} is hard to even approximate (\cref{lem:reduction}),
we seek to answer a close relative (\cref{prob:kinsertions}). We first introduce
a set of measures to capture the change in the BRs of the (original) parochial
nodes of $G$ after edge insertions. Let $\Gnew$ be obtained from $G$ by
inserting a set $\Sigma$ of directed edges between nodes of different colors,
with each inserted edge $e=(v,w)$ having weight $m_e$ (also denoted as
$m_{vw}$). For a set $U$ of vertices, we define the \emph{gain} of $U$ due to
$\Sigma$ as
\[
\bubblechange{G}{U}{{\Sigma}}{{\{m_{e}\}}_{e \in \Sigma }}{t'} \doteq\
\frac{1}{\card{U}} \sum_{u \in U} \left( \bubble{G}{u}{t'} -
\bubble{\Gnew}{u}{t'} \right) \enspace.
\]
When adding an edge to the graph, we also have to decide its weight. It seems
excessive to assume complete freedom in choosing the weight. We make the
assumption that the weight $m_{vw}$ of an edge $(v,w)$ that we would like to add
is given to us by an oracle which computes $m_{vw}$ only as a function of $v$
and of information \emph{local} to $v$ (e.g., its out-degree) obtained from $G$
and potentially a set of other edges (and their weights) that we want to add
from $v$. The weight $m_{vw}$ is the probability that a random walk arriving at
$v$ will move to $w$ in the next step. When adding $(v,w)$ with weight $m_{vw}$,
the other edges outgoing from $v$ have their weights multiplied by $1-m_{vw}$ to
ensure that the sum of the weights of the edges leaving $v$ is $1$.
The problem we want to solve then is the following.
\begin{problem}\label{prob:kinsertions}
In graph $G$, let $C$ be either blue or red. Find a set
${\Sigma}={\{(v_i, w_i)\}}_{i=1}^k$ of $k$ edges whose source vertices all
have color $C$ and all destination vertices have the other color, that
maximizes $\bubblechange{G}{\pol{C}{G}}{\Sigma}{{\{m_{e}\}}_{e \in
{\Sigma}}}{t}$.
\end{problem}
\algoname\ (\cref{alg:repbublik}) is our algorithm to approximate
\cref{prob:kinsertions}. Before describing it in detail, we give an intuition of
its workings, and present the theoretical results that guided its design.
Specifically, since our objective function is \emph{monotonic and submodular}
(\cref{lem:submodular}), we can greedily choose the edges to be added one by
one. Due to our oracle assumption on the weights, any vertex of color different
than the source can be picked as the target of the added edge, so the problem
essentially \emph{reduces to finding the sources for the edges to be added}.
\Cref{lem:gain} quantifies the gain when picking each source according to a
specific measure depending on the bounded RWCC and on the oracle-given weight
that only depends on the source. In \cref{lem:opt} we show that under mild
conditions this choice is constantly close to an optimal choice. The following
theorem states the approximation qualities of \algoname.
\begin{theorem}\label{thm:main}
Let $\Sigma$ be the output of \algoname\ and $\mathsf{OPT}$ be the optimal
solution to \cref{prob:kinsertions}. Let $\Delta_\Sigma =
\bubblechange{G}{\pol{C}{G}}{\Sigma}{{\{m_e\}}_{e \in \Sigma}}{t}$
Then
\[
\bubblechange{G}{\pol{C}{G}}{\mathsf{OPT}}{{\{m_e\}}_{e \in \mathsf{OPT}}}{t}
\leq\left( 4 \gamma(G)+1\right) \left( 1 + \frac{1}{e} \right)
\Delta_{\Sigma},
\]
where $\gamma(G)$ is the maximum over all $u \in V$ of sum of the
probabilities, for $i=0,\dotsc,t-1$, that a random walk starting at $u$ visits
$u$ at step $i$ without first visiting a vertex in $\bar{C}_u$ (see
also~\eqref{eq:fprob}),
which is a constant for many graphs.
\end{theorem}
We now proceed towards presenting lemmas which together provide a proof for
\cref{thm:main}. \Cref{lem:bubbleapprox,lem:centr} provide bounds of order
$\Theta(nt^2)$ on the runtime of the pre-processing phases of \algoname.
Therefore for small values of $t$, \algoname\ is more efficient than algorithms
that compute hitting times using the Laplacian, which need $\Omega(n^3)$ steps.
For any vertex $v$, and $0 \le i \le t$, let $\visitattime{v}{i}$ be the
probability that a random walk (in $G$) from $v$ visits $v$ at step $i$ before
reaching a vertex in $\bar{C}_v$ (it holds $\visitattime{v}{0}=1$ and
$\visitattime{v}{1}=0$ for every $v$). For any $t' \le t$, let
\begin{equation}\label{eq:fprob}
\mathcal{F}_{t'}(v) = \sum_{i=0}^{t'-1} \visitattime{v}{i}
\enspace.
\end{equation}
The following lemma shows upper and lower bounds to the change in the bubble
radius of a vertex when an new edge from it is added to the graph.
\begin{restatable}{lemma}{lemgainbounds}\label{lem:gainbounds}
Let $v \in \pol{}{G}$, $w \in \bar{C}_v$ and $t'\leq t$. Let $\Gnew$ be
the graph obtained after adding $e=(v,w)$ to $G$, with weight $m_{e}$.
The gain $\bubblechange{G}{v}{e}{m_e}{t'}$ is such that
\[
\left( \bubble{G}{v}{t'}
- 1 \right) m_e \leq \bubblechange{G}{v}{e}{m_e}{t'}\\
\leq \mathcal{F}_{t'}(v) \left( \bubble{G}{v}{t'}
- 1 \right) m_e \enspace.
\]
\end{restatable}
Decreasing the BR of $v$ decreases the BRs of vertices in $C_v$ close to $v$,
and thus the whole network. \Cref{lem:allvertices} quantifies this change.
\begin{restatable}{lemma}{lemallvertices}\label{lem:allvertices}
Let $e = (v,w)$ be the edge with weight $m_e$ added to $G$ to obtain $\Gnew$.
For any other vertex $u \in \pol{C_v}{G}$, it holds
\[
\bubblechange{G}{u}{e}{m_e}{t} =
\sum_{i=1}^{t-2} \left( \bubblechange{G}{v}{e}{m_e}{t - i}
\mathbb{P}\left( \towithin{u}{v}{=i}{G} \right) \right)\enspace.
\]
\end{restatable}
Recall that our greedy choice is to identify a node $v$ that maximizes the gain
$\bubblechange{G}{\pol{}{G}}{(v,w)}{m_{v}}{t}$ where $w$ is any vertex in
$\bar{C}_v$. \Cref{lem:allvertices} suggests that a good candidate $v$ is a
vertex that is likely to be reached by short random walks from many other
vertices in $\pol{C_v}{G}$, a property that is captured by the bounded RWCC
$\centr{t-2}{v}{\pol{C_v}{G}}$ (\cref{sec:prelims:rwcc}).
Now, we first quantify the gain for adding an edge from any vertex with RWCC $c$
(\cref{lem:gain}). Then we show that under mild conditions on the return time of
vertices we get a constant approximation by greedily choosing a vertex with
maximum ${\rm RWCC }\times m_v$ (\cref{lem:opt}).
\begin{restatable}{lemma}{lemgain}\label{lem:gain}
Let $v \in \pol{}{G}$. Let $w \in \bar{C}_v$, and assume to add the edge
$e=(v,w)$ with weight $m_e$.
It holds
\[
\bubblechange{G}{\pol{C_v}{G}}{e}{m_e}{t} \ge \frac{m_e}{2}
\centr{t-2}{v}{\pol{C_v}{G}} \enspace.
\]
\end{restatable}
This lemma suggests that inserting edges from a vertex $v$ with the highest
value of $m_v \centr{t-2}{v}{\pol{C}{G}}$ \emph{may} result in a larger
improvement in the objective function than if we chose a different source. In
the next lemma we compare the effect of choosing such sources to the effect of
an optimal choice.
\begin{restatable}{lemma}{lemopt}\label{lem:opt}
Consider the set $\pol{C}{G}$ where $C$ is either color. Among all vertices in
$\pol{C}{G}$ let $v$ and $\optnode$ be
\begin{align*}
\optnode &= \argmax_{u \in \pol{C}{G}}
\bubblechange{G}{\pol{C}{G}}{e_u}{m_{u}}{t},\\
v &= \argmax_{u \in \pol{C}{G}} m_{u} \centr{t-2}{u}{\pol{C}{G}},
\end{align*}
where $e_u$ is any potentially inserted edge connecting $u$ to
$\bar{C}_u$, and $m_u$ is its weight.\footnote{Our assumption on the oracle
giving the weight ensures that $m_u$ only depends on $u$, not on the target
of $e_u$.}
It holds
\[
\bubblechange{G}{\pol{C}{G}}{e_\optnode}{m_\optnode}{t}
\leq (4 \gamma(G) + 1) \bubblechange{G}{\pol{}{G}}{e_v}{m_v}{t},
\]
where $\gamma(G) = \max_{u \in G} \mathcal{F}_t (u)$.
\end{restatable}
If the probability of getting back to $u$ in less than $t$ steps is less than
$\alpha$ for some constant $\alpha$ then $\gamma(G) \leq \alpha$.
This assumption is realistic since $t$ is usually small and the return time to
$u$ is often much larger than $t$.
\iffalse%
\begin{proof}
By \cref{lem:gainbounds}
we have that
$\bubblechange{G}{u}{(u,w_{u})}{m_{uw_u}}{t}\leq {\mathcal
F}(u)\left(\expect{G}{T^{t'}_u(C)}\right.$
$\left.- (b+1)\right) m_{uw_u}\leq {\mathcal F}(u) (t'- b-1) m_{uw_u} $.
Using \cref{lem:allvertices} we have
$\bubblechange{G}{\pol{C}{G}}{(u,w_u)}{m_{uw_u}}{t}=\frac{1}{\pol{C}{G}}\sum_{x\in \mathcal{P}_{C}}$ $\sum_{i=1}^{t}\bubblechange{G}{u}{(u,w_u)}{m_{uw_u}}{t-i}P^i(x\rightsquigarrow u)$.
Thus,
$\bubblechange{G}{\pol{C}{G}}{(u,w_u)}{m_{uw_u}}{t}\leq \frac{{\mathcal F}(u)m_{uw_u}}{\vert \pol{C}{G}\vert }\sum_{x\in \mathcal{P}_{C}}\sum_{i=1}^{t} (t'-i- b-1) P^i(x\rightsquigarrow u)\leq {\mathcal F}(u)m_{uw_u} {\mathcal I}^t(u;\pol{C}{G})\leq {\mathcal F}(u)m_{v,w_v} {\mathcal I}^t(v;\pol{C}{G})$.
We know by previous lemma that
$m_{vw_v}{\mathcal I}^t(v;{\mathcal P}_c)\leq 2$ $\bubblechange{G}{\pol{C}{G}}{v}{m_{vw_v}}{t}+(b+1)(m_{vw_v})$. Thus,
\noindent
$\bubblechange{G}{\pol{C}{G}}{(u,w_u)}{m_{uw_u}}{t}\leq 2 {\mathcal
F}(u)\left(\bubblechange{G}{\pol{C}{G}}{(v,w_v)}{m_{vw_v}}{t}\right.$
\noindent
$+\left.(b+1)(m_{vw_v})\right).$
\end{proof}
\fi
Finally, we show that the gain function
is monotonic and sub-modular.
\begin{restatable}{lemma}{lemsubmodular}\label{lem:submodular}
Let $C$ be either blue or red and $v, u \in \pol{C}{G}$, and $w_v, w_u \in
\bar{C}_v$, such that $e_v = (v, w_v)$ and $e_v = (u, w_u)$ are not existing
edges. Let $\Sigma = \{e_v, e_u\}$. It holds
\ifextended%
\begin{equation}\label{eq:monotonicity}
\else
\[
\fi
\bubblechange{G}{\pol{C}{G}}{e_v}{m_{e_v}}{t} \leq
\bubblechange{G}{\pol{C}{G}}{\Sigma}{{\{m_e\}}_{e \in \Sigma}}{t},
\ifextended%
\end{equation}
\else
\]
\fi
and
\begin{align}
\bubblechange{G}{\pol{C}{G}}{\Sigma}{{\{m_e\}}_{e \in \Sigma}}{t} \leq &
\bubblechange{G}{\pol{C}{G}}{e_v}{m_{e_v}}{t} \nonumber \\
& + \bubblechange{G}{\pol{C}{G}}{e_u}{m_{e_u}}{t} \enspace.%
\ifextended\label{eq:submod}\else\nonumber\fi
\end{align}
\end{restatable}
We are now ready to prove \cref{thm:main}.
\begin{proof}[Proof of \cref{thm:main}]
\Cref{lem:submodular} shows the monotonicity and submodularity of the
objective function. Thus, a greedy algorithm that picks, iteratively, the $k$
best choices over all parochial vertices of color $C$ as the sources of the
added edges, will result in a $(1 + \nicefrac{1}{e})$-approximation.
\Cref{lem:gain,lem:opt} show that by choosing a vertex $v$ maximizing $m_v
\centr{t-2}{v}{\pol{C}{G}}$ among all parochial vertices of color $C$, we
obtain a vertex such that the gain when adding an edge from this source is a
$4 \gamma(G)+1$-approximation to the greedy choice. Thus, the correctness of our
algorithm is concluded by putting these lemmas together.
\end{proof}
\iffalse%
\begin{proof}
Let $G_v$ be the graph after connecting $v$ to a good vertex, $G_w$ be the graph after connecting $w$ to a good vertex and $G_{vw}$ the graph after both modifications.
Let ${\mathcal E}_{v\geq w}$ be the event of reaching $v$ earlier than $w$ and ${\mathcal E}_{w\geq v}$ be the event of reaching $v$ earlier than $w$. When the underlying graph is ambiguous we mention it in the subscript. We have:
\begin{align}
\bubblechange{G}{u}{\{v,w\}}{\{p_v,p_w\}}{t}=
& \expect{G}{T^t_u(C)}-\expect{G}{T^t_u(C)\vert \neg ({\mathcal E}_{w\geq v}\cup \mathcal{E}_{v\geq w})}\\
&\mathbb{P}(\neg ({\mathcal E}_{w\geq v}\cup {\mathcal E}_{v\geq w}))\\
&-\sum_{i=1}^t (i+\expect{G_{vw}}{T^{t-i}_v(C)}) P^{i}_{G_{vw}}(v)\mathbb{P}({\mathcal E}_{v\geq w})\\
&\ -\sum_{i=1}^t (i+\expect{G_{vw}}{T^{t-i}_w(C)}) P^{i}_{G_{vw}}(w)\mathbb{P}({\mathcal E}_{w\geq v})\\
\\
&=
\expect{G}{T^t_u(C)\vert \mathcal{E}_{w\geq v}\cup {\mathcal E}_{w\geq v}}\mathbb{P}( {\mathcal E}_{w\geq v}\cup \mathcal{E}_{v\geq w})\\
&-\sum_{i=1}^t (i+\expect{G_{vw}}{T^{t-i}_v(C)})P^{i}_{G_{vw}}(v)\mathbb{P}({\mathcal E}_{v\geq w})\\
&\ -\sum_{i=1}^t (i+\expect{G_{vw}}{T^{t-i}_w(C)})P^{i}_{G_{vw}}(w) \mathbb{P}({\mathcal E}_{w\geq v})\\
\\
\leq
&\expect{G}{T^t_u(C)\vert \mathcal{E}_{w\geq v}}\mathbb{P}({\mathcal E}_{w\geq v})+ \expect{G}{T^t_u(C)\vert \mathcal{E}_{w\geq v}}\mathbb{P}( {\mathcal E}_{v\geq w})\\
&-\sum_{i=1}^t (i+\expect{G_{vw}}{T^{t-i}_v(C)})P^{i}_{G_{vw}}(v)\mathbb{P}_{G_{vw}}({\mathcal E}_{v\geq w})\\
&\ -\sum_{i=1}^t (i+\expect{G_{vw}}{T^{t-i}_w(C)})P^{i}_{G_{vw}}(w) \mathbb{P}_{G_{vw}}({\mathcal E}_{w\geq v})\\
\leq
&\left(\expect{G}{T^t_u(C)\vert \mathcal{E}_{w\geq v}}-\sum_{i=1}^t (i+\expect{G_{vw}}{T^{t-i}_v(C)})P^{i}_{G_{vw}}(v)\right)\mathbb{P}_{G_{vw}}(\mathcal{E}_{v\geq w})\\
&
+\left( \expect{G}{T^t_u(C)\vert \mathcal{E}_{w\geq v}} -\sum_{i=1}^t (i+\expect{G_{vw}}{T^{t-i}_w(C)})P^{i}_{G_{vw}}(w)\right) \mathbb{P}_{G_{vw}}({\mathcal E}_{w\geq v})\\
\leq
&\left(\expect{G}{T^t_u(C)\vert \mathcal{E}_{w\geq v}}-\sum_{i=1}^t
(i+\expect{G_{v}}{T^{t-i}_v(C)})P^{i}_{G_{v}}(v)\right)\mathbb{P}_{G_{v}}{\mathcal{E}_{v\geq
w}}\\
&
+\left( \expect{G}{T^t_u(C)\vert \mathcal{E}_{w\geq v}} -\sum_{i=1}^t
(i+\expect{G_{w}}{T^{t-i}_w(C)})P^{i}_{G_{w}}(w)\right)
\mathbb{P}_{G_{w}}(\mathcal{E}_{w\geq v})\\
& \leq
\bubblechange{G}{\pol{C}{G}}{v}{p_v}{t}\\
&+ \bubblechange{G}{\pol{C}{G}}{w}{p_w}{t}.
\end{align}
\end{proof}
\fi
\begin{algorithm}[t]
\algrenewcommand\algorithmicindent{1.0em}
\begin{algorithmic}[1]
\State\textbf{Input}: Graph $G=(V,E)$, desired insertions $k_C$, oracle
$\mathcal{W}_G:V \times 2^{V \times V} \to [0,1]$, $C\in\{R,B\}$.
\State\textbf{Output}: Set $\Sigma_C$ of $k_C$ edges to be inserted, with their weights.
\State$\Sigma_C \gets \emptyset$
\For{$i=1$ \textbf{to} $k_C$}
\State$P \gets$ \texttt{computeParochials}($G \cup \Sigma_C$, $C$)
\State$\mathcal{R} \gets$ \texttt{computeRWCentrality}($P$, $G \cup \Sigma_C$)
\State$v_i \gets {\rm argmax}_{v \in P} \mathcal{R}(v)\times
\mathcal{W}_G(v, \Sigma_C)
\State$u_i\gets $ arbitrary in $ \bar{C}_{v_i}$
\State$\Sigma_C\gets \Sigma_C\cup\{(v_i,u_i)\}$
\EndFor%
\State\textbf{return} $\Sigma_C$
\end{algorithmic}
\caption{\algoname}\label{alg:repbublik}
\end{algorithm}
We can now give the details to \algoname. The algorithm takes as input the graph
$G$, the number $k_C$ of desired edge insertions, the oracle $\mathcal{W}$ that
determines the weights of the new edges, and the set of nodes $C$. It
first creates the empty set $\Sigma_C$ that will store the edges to be added and then enters
a for loop to be repeated for $k_C$ times. At every iteration of the loop, it first
computes the BR of every node in $C$ in the graph (denoted in the pseudocode as $G \cup \Sigma_C$) obtained by adding to $G$ the edges currently in $\Sigma_C$ (with their weights obtained from the oracle $\mathcal{W}_G$) (in practice, the BR is computed using the approximation
algorithm outlined in \cref{lem:bubbleapprox}). Thanks to this computation, the algorithm obtains (line 5) the set $P$ of parochial nodes in this graph (at the first iteration of the loop $P=\badc{C}{G}$). It then obtains the centralities values $\centr{t-2}{v}{P}$ of every node $v \in P$ (in practice, using the approximation algorithm outlined in
\cref{lem:centr}), storing them in a dictionary $\mathcal{R}$ (line 6).
The algorithm then selects the node $v_i \in P$ associated to the
maximum quantity $\mathcal{R}(v_i)\times \mathcal{W}_G(v_i, \Sigma_C)$, and
arbitrarily picks a node $u_i$ of the opposite color of $v_i$ (i.e., of the color other than $C$). The directed edge $(v_i,u_i)$ is added to the set $\Sigma_C$ (lines 7--9). After $k_C$ iterations of the loop,
the algorithm returns $\Sigma_C$, together with the weights obtained from the oracle.
\algoname would require a re-computation of the BRs and of the centralities of all vertices, at every iteration of the loop, which would require to run a very large number of random walks, making it computationally very expensive. We now propose a more practical alternative \algonameplus, at the price of losing the approximation guarantees. \algonameplus\ only computes $\bad{C}{G}$ and $\mathcal{R}$ before entering the for loop, and uses the same values throughout its execution, but trades off the consequences of this choice by adding a penalty factor to the objective function involved in the selection of the source vertices for the edges to be added. Specifically, \algonameplus\ chooses $v_i$ (line 7) by maximizing the quantity $\nicefrac{\mathcal{R}(v) \times \mathcal W_G(v, \Sigma_C)}{\eta_v}$, where $\eta_v$ is a penalty factor equals to one plus the number of edges with source $v$ in $\Sigma_C$ (thus at iteration 1, $\eta_v = 1$ for every node). This penalty factor favours the insertion of edges from nodes that have not yet been altered. Consequently, it indirectly \textit{(1)} handles the possibility that nodes with new edges are no longer parochial, thus we want to avoid to keep adding edges to them; and \textit{(2)} avoids that the new edges are added from a restricted set of nodes, limiting the positive effect of the insertions on $\bubblechange{G}{\badc{C}{G}}{{\Sigma_C}}{{\{m_{e}\}}_{e \in \Sigma }}{t'}$.
\section{Missing proofs}\label{sec:appendix}
We present here the proofs missing from the main body. For convenience, we
repeat the statements of the lemmas.
\lemcentr*
\begin{proof}[Proof of \cref{lem:centr}]
We can write
\[
\centr{t'}{v}{S} = t' - \frac{1}{\card{S}} \sum_{ w \in S }
\expect{G}{\ttime{t'}{w}{v}} \enspace.
\]
We apply Chebyshev's inequality to the r.v.\ $\nicefrac{1}{z} \sum_{i=1}^z
\bar{h}_{w_{i}}$, to bound the deviation from its expectation
\[
\frac{1}{\card{S}} \sum_{ w \in S } \expect{G}{\ttime{t'}{w}{v}} \enspace.
\]
To get an upper bound to the variance of this r.v., we use the fact that the
r.v.'s $\bar{h}_{w_i}$, $i=1,\dotsc,z$, are independent, and, from Popoviciu's
inequality, the fact that each has a variance at most $\nicefrac{t'^2}{4}$, as
$\bar{h}_{w_i} \in [0,t']$.
\end{proof}
The following result is used in the proof of \cref{lem:restart}.
\begin{lemma}[Markov inequality for bounded random variables]\label{lem:markov}
Let $X$ be a random variable satisfying $0 \leq X \leq t$. We have:
\[
\mathbb{P}(X \leq k) \leq \frac{t - \expect{}{X}}{t - k} \enspace.
\]
\end{lemma}
\begin{proof}
It holds
\begin{align*}
\expect{}{X} &= \int_{0}^{k} x p(x) dx +\int_{k}^{t} x p(x) dx \\
&\leq k \left( 1 - \mathbb{P}(X \geq k) \right) + t \mathbb{P}(X \geq k) \enspace.
\end{align*}
Thus,
\[
\mathbb{P}(X \geq k) \geq \frac{\expect{}{X} - k}{t - k},
\]
and
\[
\mathbb{P}(X \leq k) = 1 - \mathbb{P}(X \geq k) \leq 1 - \frac{\expect{}{X} - k}{t - k} =
\frac{t - \expect{}{X}}{t - k} \enspace. \qedhere
\]
\end{proof}
\lemrestart*
\begin{proof}[Proof of \cref{lem:restart}]
Assume first that $\bubble{G}{v}{t} \geq t (1-\nicefrac{1}{8r})$. Consider a
set of $r$ independent random walkers, $w_1, \dotsc, w_r$, each starting from
$v$. We can see the trace of the partial walks taken by our random walker with
restarts as the union of the traces of these walkers. The event $\mathcal{E}'
\doteq \text{``}\mathcal{T}_v \le \nicefrac{t}{2}\text{''}$ is a strict subset
of the event $\mathcal{E}'' \doteq \text{``there is (at least) a walker } w_i$
for which $T_v^t \le \nicefrac{t}{2}$'', as the condition in $\mathcal{E}'$
implies the condition in $\mathcal{E}''$, but not vice versa. Thus,
$\mathbb{P}(\mathcal{E}') < \mathbb{P}(\mathcal{E}'')$. By \cref{lem:markov} we have, for
each walker, that
\[
\mathbb{P}\left(T^t_v\leq \frac{t}{2} \right) \leq
\frac{t - \expect{}{\ttime{t}{v}{S}}}{t - \frac{t}{2}} \le
\frac{\frac{t}{8r}}{\frac{t}{2}} \leq \frac{1}{4r}.
\]
Thus, using the union bound over the $r$ walkers, we get $\Pr({\mathcal
E}'') \le \nicefrac{1}{4}$. Equivalently $\mathbb{P}\left(\mathcal{T}_v\leq \nicefrac{t}{2}
\right) \leq \nicefrac{1}{4}$.
For the case when $\bubble{G}{v}{t} \leq b$,
using Markov inequality we get $\mathbb{P}\left(\mathcal{T}_v > 4br \right) \leq
\nicefrac{1}{4}$.
\end{proof}
\lemreduction*
\begin{proof}[Proof of \cref{lem:reduction}]
We show an approximation-preserving polynomial time reduction from the
minimum set cover problem to \cref{prob:zerobias}. Our reduction does
\emph{not} change the cost of the optimal solution, thus maintaining, in
addition to NP-hardness, the APX-hardness.
Let $U=\{u_1, u_2,\dots , u_n\}$ be a domain and let $S_1,S_2,\dotsc, S_m$\\
$\subseteq U$ be an instance of the set cover problem. We construct an instance
of \cref{prob:zerobias} as follows. Fix $t \ge 3$. Let $V$ be union of the
following sets: $U$, $S={\{s_i\}}_{i=1}^m$ representing the sets, $T =
\bigcup_{j=1}^m T_j$ where each $T_j$ is a set of $\lceil t/2 \rceil -1 $
distinct vertices, and $\{g\}$. Assume all vertices except $g$ have color red
and $g$ is blue. For each $i\in [n]$ and $j \in [m]$, place an edge from $u_i$
to $s_j$ if and only if $u_i\in S_j$. For each $j\in [m]$, using the vertices
in $T_j$, place a path of length $\lceil t / 2\rceil - 1$ going from $s_j$ to
$g$. For each $1\leq j\leq m$, it holds $\bubble{G}{s_j}{t} = \lceil
\nicefrac{t}{2} \rceil - 1$, and for each $1\leq i\leq n$,
\[
\bubble{G}{u_i}{t} = \frac{1}{\card{\{j\ :\ u_i \in S_j\}}} \sum_{j\
\text{s.t.}\ u_i\in S_j} \bubble{G}{s_j}{t}+1 = \lceil \nicefrac{t}{2}
\rceil \enspace.
\]
Clearly the bubble radius of vertices in $T$ is strictly less than
$\nicefrac{t}{2}$. Thus the parochial vertices are all and only those in $U$.
Assume there is a polynomial-time algorithm for \cref{prob:zerobias}. For any
(optimal) solution $\Sigma \subseteq V \times V$, it holds
$\bubble{G_{\textrm{new}}}{u_i}{t} < \nicefrac{t}{2}$ if and only if $\Sigma$
contains an edge whose source is in $\{u_i\} \cup
\bigcup_{j\ \text{s.t.}\ u_i \in S_j} (\{s_j\} \cup T_j)$, for each $i \in
[n]$. The source vertices of the edges in $\Sigma$ must be distinct, as any
solution containing two edges originating from the same vertex cannot be
optimal. Denote with $Z$ the set of the source vertices of the edges in
$\Sigma$. Consider now the solution $\Sigma'$ obtained by changing (in
polynomial time) $\Sigma$ as follows: \textit{1.} each edge in $\Sigma$ whose
source is in $T_i$ is modified to have source $s_i$, for each $i \in [n]$; and
\textit{2.} each edge in $\Sigma$ whose source is $u \in U$ is changed to
have source $s_j$ where $j$ is such that $u \in S_j$. Clearly $\Sigma'$ is
still an (optimal) solution to \cref{prob:zerobias}. Let $\mathsf{OPT}$ be the
set of source vertices of the edges in $\Sigma'$. Clearly it must be
$\mathsf{OPT} \subseteq S$. We now show that $\Sigma'$ is an (optimal)
solution to \cref{prob:zerobias} if and only if $\mathsf{OPT}$ is such
that $\{ S_j \:\ s_j \in \mathsf{OPT} \}$ is a minimum set cover for the
considered instance. It is evident that $\{ S_j\ :\ s_j \in \mathsf{OPT} \}$
is a set cover, which can be obtained in polynomial time from $\Sigma'$. We
now show that this set cover is minimal. Consider now any set cover $Y
\subseteq \{S_1,\dotsc, S_m\}$, and consider the set of edges $\{(s_i, g)\ :\
S_i \in Y\}$. Adding these edges to $G$ would result in all the vertices in
$U$ to no longer be parochial. This holds in particular for any \emph{minimal}
set cover $Y$, from which we can create an (optimal) solution $\Sigma_Y$ to
\cref{prob:zerobias}. Thus we found a bijection between (optimal)
solutions to \cref{prob:zerobias} and minimal set covers for the
considered instance, and computing one from the other can be done in
polynomial time, showing the NP-hardness of \cref{prob:zerobias}. The
APX-hardness follows because, for any minimum set cover $Y$, the corresponding
optimal solution $\Sigma_Y$ to \cref{prob:zerobias}, built as above, is
such that $\card{\Sigma_Y} = \card{Y}$, thus if we had a constant-factor
polynomial-time approximation algorithm for \cref{prob:zerobias} we
would have an algorithm with the same properties for the minimum set cover
problem.
\end{proof}
\lemgainbounds*
\begin{proof}[Proof of \cref{lem:gainbounds}]
Consider the probability space of all random walks starting from $v$ in $\Gnew$ and $G$. We introduce a coupling between these two probability spaces as follows: consider a walk in $\Gnew$ and couple every step of it to an identical step in $G$.
If a walk in $\Gnew$ never traverses $(v,w)$ then the gain function is zero as it gets coupled to the identical walk in $G$.
Assume that the walk in $\Gnew$ traverses $(v,w)$ at the $i$th step without first visiting a vertex in $\bar{C}_v$. Before traversing $(v,w)$, the two identical walks in $\Gnew$ and $G$ have the same probabilities and the above coupling works. We partition the state space by conditioning on the step $i$ as follows:
Let $\mathcal{E}_i$, $1 \le i \le t'$, be the event that the walk in $\Gnew$ traverses $(v,w)$ at step $i$. Consider all such walks,
at step $i-1$ these walks need one more steps to reach the other color, and they are coupled to walks in $G$ which in expectation need $\bubble{G}{v}{t'-i+1}$ steps to reach $\bar{C}_v$. Thus, assuming $\mathcal{E}_i$, the gain in bubble radius is equal to $\bubble{G}{v}{t'-i+1}-1$.
\iffalse
By extension, let $\mathcal{E}_{t'+1}$ be the event that a walk in $\Gnew$ never traverses $(v,w)$.
We can express
$\bubble{\Gnew}{v}{t'}$ using the law of total expectation as
\begin{align*}
\bubble{G}{v}{t'}-\bubble{\Gnew}{v}{t'} \doteq \expect{G}{\ttime{t'}{v}{\bar{C}_v}}- \expect{\Gnew}{\ttime{t'}{v}{\bar{C}_v}} \\
=\sum_{i=1}^{t'} \left(\expect{G}{\ttime{t'}{v}{\bar{C}_v }\mid \mathcal{E}_i}-\expect{\Gnew}{\ttime{t'}{v}{\bar{C}_v} \mid \mathcal{E}_i}\right) \mathbb{P}({\mathcal E}_i) \enspace.
\end{align*}
It holds $\expect{\Gnew}{\ttime{t'}{v}{\bar{C}_v} \mid \mathcal{E}_{t'+1}} = \expect{G}{\ttime{t'}{v}{\bar{C}_v} \mid \mathcal{E}_{t'+1}}$, where by the conditioning in the expectation over $G$ we mean to consider only the walks in $G$ corresponding to walks in $\Gnew$ for which $\mathcal{E}_{t'+1}$ holds. It also holds
$\expect{G}{\ttime{t'}{v}{\bar{C}_v }\mid \mathcal{E}_i}=i-1+\bubble{G}{v}{t'-i+1} $.
Furthermore, $\expect{\Gnew}{\ttime{t'}{v}{\bar{C}_v} \mid \mathcal{E}_i} = i$,
from the definition of $\mathcal{E}_i$ and the fact that the new edge goes
from $v$ to $w \in \bar{C}_v$, so
\fi
Using the law of total expectation and summing over all $1\leq i\leq t'$, we can write
\begin{equation}\label{eq:4}
\bubblechange{G}{v}{(v,w)}{m_{vw}}{t'} = \sum_{i=1}^{t'} \left(
\bubble{G}{v}{t'-i+1} - 1 \right) \mathbb{P}(\mathcal{E}_i) \enspace.
\end{equation}
The left hand side follows from the fact that $\mathbb{P}({\mathcal E}_1)=m_{vw}$ and that $\bubble{G}{v}{j} \ge 1$ for any $1 \le j \le t'$.
\iffalse
\begin{align*}
&\bubble{\Gnew}{v}{t'} \doteq \expect{\Gnew}{\ttime{t'}{v}{\bar{C}_v}} \\
&\begin{aligned}
\leq \expect{\Gnew}{\ttime{t'}{v}{\bar{C}_v} \mid
\mathcal{E}_1} &\mathbb{P}({\mathcal E}_1) + \\
& \expect{\Gnew}{\ttime{t'}{v}{\bar{C}_v} \mid \neg {\mathcal{E}}_1} (1 -
\mathbb{P}(\mathcal{E}_1))
\end{aligned}\\
&\leq \mathbb{P}(\mathcal{E}_1) + \bubble{G}{v}{t'}
\left( 1 - \mathbb{P}(\mathcal{E}_1) \right)\\
&\leq m_{vw}+\bubble{G}{v}{t'}- \bubble{G}{v}{t'}m_{vw}\enspace.
\end{align*}
Which is equivalent to:
\[
\bubblechange{G}{v}{(v,w)}{m_{vw}}{t'}\geq (\bubble{G}{v}{t'} - 1)m_{vw}\enspace.
\]
\question[from=Matteo,date=1/7]{How is the above sufficient to show the
l.h.s.? The equivalence depends on the assumption that $\mathcal{E}_1$ is
the only event with non-zero probability.
If we assume that only $\mathcal{E}_1$ and $\mathcal{E}_3$ may have non-zero
probability ($\mathcal{E}_2$ always has probability zero, unless we allow
self-loops, and I'm assuming we don't, but if we do, replace
$\mathcal{E}_3$ with $\mathcal{E}_2$), we get
\begin{align*}
\bubble{\Gnew}{v}{t'} \doteq &
\mathbb{P}(\mathcal{E}_1) + \expect{\Gnew}{\ttime{t'}{v}{\bar{C}_v} \mid
\mathcal{E}_3} \mathbb{P}(\mathcal{E}_3) \\
&+ \bubble{G}{v}{t'} (1 - \mathbb{P}(\mathcal{E}_1) - \mathbb{P}(\mathcal{E}_3) ) \\
&= m_{vw} + 3 \mathbb{P}(\mathcal{E}_3)+ \bubble{G}{v}{t'} (1 - m_{vw} -
\mathbb{P}(\mathcal{E}_3))
\end{align*}
which leads to
\begin{align*}
\bubblechange{G}{v}{(v,w)}{m_{vw}}{t'} = &(\bubble{G}{v}{t'} - 1) m_{vw}
\\
&+ ( \bubble{G}{v}{t'} - 3) \mathbb{P}(\mathcal{E}_3) \enspace.
\end{align*}
To get the l.h.s.\ of the thesis under this assumption, we need to argue
that $\bubble{G}{v}{t'} \ge 3$, but why is that true?
What is special about the assumption that only $\mathcal{E}_1$ has non-zero
probability that would allow us to draw the conclusion in general?
}
\fi
The right-hand side is concluded from the fact that $\bubble{G}{v}{t'-i+1}\leq\bubble{G}{v}{t'} $ and that
\[
\sum_{i=1}^{t'} \mathbb{P} (\mathcal{E}_i) = \sum_{i=0}^{t'-1} \visitattime{v}{i} m_{vw}= \mathcal{F}_{t'}(v)m_{vw} \enspace.
\]
\iffalse
and then obviously $\expect{G}{\ttime{t'}{v}{\bar{C}_v}} - i \le
\expect{G}{\ttime{t'}{v}{\bar{C}_v}} - 1$ for $1 \le i \le t'$, which combined
with~\eqref{eq:4} gives the right-hand side of the thesis.
\fi
\end{proof}
\lemallvertices*
\begin{proof}[Proof of \cref{lem:allvertices}]
Using the law of total expectation, for any graph $Z$, it holds
\begin{align*}
\bubble{Z}{u}{t} =& \left( \sum_{i=1}^{t-1} \left( i + \bubble{Z}{v}{t-i}
\right) \mathbb{P}\left( \towithin{u}{v}{=i}{Z} \right) \right) \\
& + \expect{Z}{\ttime{t}{u}{\bar{C}_v} \mid \nottowithin{u}{v}{< t}{Z}}
\mathbb{P}\left( \nottowithin{u}{v}{< t}{Z} \right) \enspace.
\end{align*}
Between $G$ and $\Gnew$, we are only adding an outgoing edge from $v$ and
modifying the weights of the edges outgoing from $v$, so
\begin{align*}
\expect{G}{\ttime{t}{u}{\bar{C}_v} \mid \nottowithin{u}{v}{< t}{G}} =
\expect{\Gnew}{\ttime{t}{u}{\bar{C}_v} \mid \nottowithin{u}{v}{<
t}{\Gnew}},\\
\mathbb{P}\left(\nottowithin{u}{v}{< t}{G} \right) = \mathbb{P}\left(\nottowithin{u}{v}{<
t}{\Gnew} \right),\ \text{and}\ \mathbb{P}\left( \towithin{u}{v}{=i}{G} \right) = \mathbb{P}\left(
\towithin{u}{v}{=i}{\Gnew} \right) \enspace.
\end{align*}
Therefore,
\begin{align*}
\bubblechange{G}{u}{(v,w)}{m_{v}}{t} &\doteq
\bubble{G}{u}{t}-\bubble{G_{\textrm{new}}}{u}{t}\\
&= \sum_{i=1}^{t-1} \left( \bubblechange{G}{v}{(v,w)}{m_{v}}{t-i}\right)
\mathbb{P}\left( \towithin{u}{v}{=i}{G} \right) \\
&= \sum_{i=1}^{t-2} \left( \bubblechange{G}{v}{(v,w)}{m_{v}}{t-i}\right)
\mathbb{P}\left( \towithin{u}{v}{=i}{G} \right)
\enspace.
\end{align*}
The last step follows from the fact that $\bubblechange{G}{v}{(v,w)}{m_v}{1}
=0$ because $\bubble{Z}{u}{1}=1$ for every vertex $u$ of any graph $Z$.
\end{proof}
We need the following technical result in successive proofs.
\begin{lemma}\label{lem:parochiallowert}
If $\bubble{G}{v}{t} \geq \nicefrac{t}{2}$ then, for any
$t' \leq t$, it holds
\[
\frac{t'}{2} \leq \bubble{G}{v}{t'} \leq t' \enspace.
\]
\end{lemma}
\begin{proof}
The rightmost inequality is straightforward from the definition of
$\bubble{G}{v}{t'}$. Expanding the definition of $\bubble{G}{v}{t}$,
it holds, for any $t' < t$,
\begin{align*}
\bubble{G}{v}{t} &= t \mathbb{P} \left( \towithin{v}{\bar{C}_v}{\ge t}{G} \right)
+ \sum_{i=1}^{t-1} i \mathbb{P} \left( \towithin{v}{\bar{C}_v}{= i}{G} \right) \\
&\leq t \mathbb{P} \left( \towithin{v}{\bar{C}_v}{\ge t}{G} \right)
+ t \mathbb{P} \left( \towithin{v}{\bar{C}_v}{t' \le \cdot \le t}{G} \right)
+ \sum_{i=1}^{t'-1} i \mathbb{P} \left( \towithin{v}{\bar{C}_v}{= i}{G} \right) \\
& = t \mathbb{P} \left( \towithin{v}{\bar{C}_v}{\ge t'}{G} \right)
+ \sum_{i=1}^{t'-1} i \mathbb{P} \left( \towithin{v}{\bar{C}_v}{= i}{G} \right)
\enspace.
\end{align*}
Thus, since the l.h.s.~is at least $\nicefrac{t}{2}$,
\[
\frac{t}{2} \leq t \mathbb{P} \left( \towithin{v}{\bar{C}_v}{\ge t'}{G} \right)
+ \sum_{i=1}^{t'-1} i \mathbb{P} \left( \towithin{v}{\bar{C}_v}{= i}{G} \right),
\]
i.e.,
\[
\frac{1}{2} - \frac{1}{t} \sum_{i=1}^{t'-1} i \mathbb{P} \left(
\towithin{v}{\bar{C}_v}{= i}{G} \right)
\leq \mathbb{P} \left( \towithin{v}{\bar{C}_v}{\ge t'}{G} \right)
\enspace.
\]
By expanding $\bubble{G}{v}{t'}$ in a similar way, and plugging in the last
inequality above, we get
\begin{align*}
\bubble{G}{v}{t'} &= t' \mathbb{P} \left( \towithin{v}{\bar{C}_v}{\ge t'}{G} \right)
+ \sum_{i=1}^{t'-1} i \mathbb{P} \left( \towithin{v}{\bar{C}_v}{= i}{G} \right)\\
&\geq t' \left( \frac{1}{2} - \frac{1}{t} \sum_{i=1}^{t'-1} i
\mathbb{P} \left( \towithin{v}{\bar{C}_v}{= i}{G} \right)
\right) + \sum_{i=1}^{t'-1} i \mathbb{P} \left( \towithin{v}{\bar{C}_v}{= i}{G}
\right)\\
& = \frac{t'}{2} + \underbracket{\left( 1 - \frac{t'}{t} \right)
\sum_{i=1}^{t'-1} i \mathbb{P} \left( \towithin{v}{\bar{C}_v}{= i}{G}
\right)}_{\ge 0}\\
&\geq \frac{t'}{2} \enspace. \qedhere
\end{align*}
\end{proof}
\lemgain*
\begin{proof}[Proof of \cref{lem:gain}]
Using \cref{lem:allvertices}, we get
\begin{align}
&\bubblechange{G}{\pol{C_v}{G}}{e}{m_e}{t} = \nonumber\\
&\frac{1}{\card{\pol{C_v}{G}}}
\sum_{u \in \pol{C_v}{G}} \sum_{i=1}^{t-2}
\bubblechange{G}{v}{e}{m_e}{t-i}
\mathbb{P}\left( \towithin{u}{v}{=i}{G} \right) \enspace.\label{eq:gaintech}
\end{align}
It holds from \cref{lem:gainbounds,lem:parochiallowert} that
\[
\bubblechange{G}{v}{e}{m_e}{t'} \geq \left( \frac{t'}{2} - 1 \right)
m_e\ \text{for every}\ 1\leq t' \le t\enspace.
\]
Using this fact, we can continue from~\eqref{eq:gaintech} as follows
\begin{align*}
&\bubblechange{G}{\pol{C_v}{G}}{e}{m_e}{t} \\
&\ge \frac{1}{\card{\pol{C_v}{G}}} \sum_{u \in \pol{C_v}{G}} \sum_{i=1}^{t-2}
\left( \frac{t-i}{2}-1 \right) m_e
\mathbb{P}\left( \towithin{u}{v}{=i}{G} \right)\\
&= \frac{m_e}{2} \underbracket{\frac{1}{\card{\pol{C_v}{G}}} \sum_{u\in \pol{C_v}{G}}
\sum_{i=1}^{t-2} (t-i-2) \mathbb{P}\left( \towithin{u}{v}{=i}{G}
\right)}_{\centr{t-2}{v}{\pol{C_v}{G}}},
\end{align*}
which concludes the proof.
\end{proof}
\lemopt*
\begin{proof}[Proof of \cref{lem:opt}]
It follows from \cref{lem:gainbounds} that, for any $t'$,
\begin{align*}
\bubblechange{G}{\optnode}{e_\optnode}{m_\optnode}{t'}
&\le \mathcal{F}_{t'}(u) \left( \bubble{G}{\optnode}{t'}
- 1\right) m_\optnode \\
&\leq (t'- 1) m_\optnode \mathcal{F}_{t'}(\optnode) \enspace.
\end{align*}
By applying \cref{lem:allvertices} first, and then the above inequality, we get
\begin{align*}
&\bubblechange{G}{\pol{C}{G}}{e_\optnode}{m_\optnode}{t}\\
\le& \frac{1}{\card{\pol{C}{G}}}\sum_{u \in \pol{C}{G}} \sum_{i=1}^{t-2}
\left( \bubblechange{G}{\optnode}{e_\optnode}{m_\optnode}{t-i} \right)
\mathbb{P} \left( \towithin{u}{v}{= i}{G} \right)\\
\le& \frac{1}{\card{\pol{C}{G}}}\sum_{u \in \pol{C}{G}} \sum_{i=1}^{t-3}
\left( \bubblechange{G}{\optnode}{e_\optnode}{m_\optnode}{t-i} \right)
\mathbb{P} \left( \towithin{u}{v}{= i}{G} \right)\\
&\hspace{2cm}+
\bubblechange{G}{\optnode}{e_\optnode}{m_\optnode}{2}
\mathbb{P} \left(\towithin{u}{v}{= t-2}{G}\right)\\
\le& \frac{1}{\card{\pol{C}{G}}}\sum_{u\in\pol{C}{G}} \sum_{i=1}^{t-3}
(t - 1 - i) m_\optnode \mathcal{F}_t(\optnode)
\mathbb{P} \left( \towithin{u}{v}{= i}{G} \right) +1\\
\le& \frac{1}{\card{\pol{C}{G}}}\sum_{u \in \pol{C}{G}} \sum_{i=1}^{t-2}
2 (t - 2 - i) m_\optnode \mathcal{F}_t(\optnode)
\mathbb{P} \left( \towithin{u}{v}{= i}{G} \right)+1 \\
\le&\ 2 m_\optnode
\centr{t-2}{\optnode}{\bar{C}_\optnode} \mathcal{F}_t(\optnode)+1 \le
2 m_v \centr{t-2}{v}{\bar{C}_\optnode} \gamma(G)+1\\
\le&\ ( 4 \gamma(G)+1) \bubblechange{G}{\pol{C}{G}}{e_v}{m_v}{t},
\end{align*}
where the last step follows from \cref{lem:gain}.
\end{proof}
\lemsubmodular*
\begin{proof}[Proof of \cref{lem:submodular}]
Let $G_v$ be the graph after adding only the edge $e_v$, $G_u$ be the graph
after only adding the edge $e_u$, and $G_{vu}$ be the graph after adding both
edges.
We first show the monotonicity of the objective function, i.e.,
that~\eqref{eq:monotonicity} holds. For any $w \in \pol{C}{G}$, it holds
\begin{align*}
\bubblechange{G}{w}{e_v}{m_v}{t} &\doteq \bubble{G}{w}{t} -
\bubble{G_v}{w}{t} \\
&\le \bubble{G}{w}{t} - \bubble{G_{vu}}{w}{t} \\
&\doteq \bubblechange{G}{w}{\{e_v, e_u\}}{\{m_v, m_u\}}{t}
\end{align*}
because $\bubble{G_v}{w}{t} \geq \bubble{G_{vu}}{w}{t}$, as adding an edge
from $u$, which is in $C_w$, to a vertex in $\bar{C}_w$ cannot increase the
bubble radius of $w$. The result generalizes to~\eqref{eq:monotonicity} in a
straightforward way.
We now show the sub-modularity of the objective function, i.e.,
that~\eqref{eq:submod} holds. We start by showing that, for $w \in
\pol{C}{G}$, it holds
\begin{align*}
\bubblechange{G}{w}{\{e_v, e_u\}}{\{m_v, m_u\}}{t} \le &
\bubblechange{G}{w}{e_v}{m_v}{t} \\
&+ \bubblechange{G}{w}{e_u}{m_u}{t} \enspace.
\end{align*}
With an expansion of the definition and a slight rearrangement of the terms,
the above inequality is equivalent to
\[
\underbracket{\bubble{G_v}{w}{t} -
\bubble{G_{vu}}{w}{t}}_{\bubblechange{G_v}{w}{e_u}{m_u}{t}} \le
\underbracket{\bubble{G}{w}{t} -
\bubble{G_{u}}{w}{t}}_{\bubblechange{G}{w}{e_u}{m_u}{t}},
\]
i.e., the gain of adding the same edge (in this case $e_u$) is smaller when
the edge is added to a graph (in this case $G_v$) that has a superset of the
edges (compared to $G$).
Consider all the walks from $w$ that either pass through $v$ or $u$. Among
such walks, let $\mathcal{E}_v$ be the event of seeing $v$ first and
$\mathcal{E}_u$ be the event of seeing $u$ first. If a walk does not pass
through either $v$ or $u$, its probability of hitting the other color is the
same in all graphs, as the graphs differ only in the outgoing edges from these
two nodes and their weights. For the same reason, $\mathbb{P}(\mathcal{E}_v)$ and
$\mathbb{P}(\mathcal{E}_u)$ do not change across the graphs. Thus,
\begin{align*}
&\bubble{G_v}{w}{t} - \bubble{G_{vu}}{w}{t} = \\
&\left( \expect{G_v}{\ttime{t}{w}{\bar{C}_w} \mid \mathcal{E}_v} -
\expect{G_{vu}}{\ttime{t}{w}{\bar{C}_w} \mid \mathcal{E}_v} \right)
\mathbb{P}(\mathcal{E}_v)\\
&+ \left( \expect{G_v}{\ttime{t}{w}{\bar{C}_w} \mid \mathcal{E}_u} -
\expect{G_{vu}}{\ttime{t}{w}{\bar{C}_w} \mid \mathcal{E}_u} \right)
\mathbb{P}(\mathcal{E}_u) \enspace.
\end{align*}
Similarly,
\begin{align*}
&\bubble{G}{w}{t} - \bubble{G_u}{w}{t} = \\
&\left( \expect{G}{\ttime{t}{w}{\bar{C}_w} \mid \mathcal{E}_v} -
\expect{G_u}{\ttime{t}{w}{\bar{C}_w} \mid \mathcal{E}_v} \right)
\mathbb{P}(\mathcal{E}_v)\\
&+ \left( \expect{G}{\ttime{t}{w}{\bar{C}_w} \mid \mathcal{E}_u} -
\expect{G_u}{\ttime{t}{w}{\bar{C}_w} \mid \mathcal{E}_u} \right)
\mathbb{P}(\mathcal{E}_u) \enspace.
\end{align*}
We want to show that it holds
\begin{align}
\expect{G_v}{\ttime{t}{w}{\bar{C}_w \mid \mathcal{E}_v}} -
\expect{G_{vu}}{\ttime{t}{w}{\bar{C}_w \mid \mathcal{E}_v}} \nonumber\\
\le \expect{G}{\ttime{t}{w}{\bar{C}_w \mid \mathcal{E}_v}} -
\expect{G_u}{\ttime{t}{w}{\bar{C}_w \mid \mathcal{E}_v}} \enspace.\label{eq:diffcondonv}
\end{align}
and
\begin{align}
\expect{G_v}{\ttime{t}{w}{\bar{C}_w \mid \mathcal{E}_u}} -
\expect{G_{vu}}{\ttime{t}{w}{\bar{C}_w \mid \mathcal{E}_u}} \nonumber\\
\le \expect{G}{\ttime{t}{w}{\bar{C}_w \mid \mathcal{E}_u}} -
\expect{G_u}{\ttime{t}{w}{\bar{C}_w \mid \mathcal{E}_u}}
\enspace.\label{eq:diffcondonu}
\end{align}
We can write
\begin{align*}
&\expect{G}{\ttime{t}{w}{\bar{C}_w} \mid \mathcal{E}_v} \\
&= \sum_{i=1}^t \left (
i + \expect{G}{\ttime{t-i}{v}{\bar{C}_v}} \right ) \mathbb{P}\left(
\towithin{w}{v}{=i}{G} \mid \mathcal{E}_v \right) \enspace.
\end{align*}
The probability on the right is the same on all graphs. Similar expressions
hold for
\[
\expect{G_v}{\ttime{t}{w}{\bar{C}_w} \mid \mathcal{E}_v},\
\expect{G_u}{\ttime{t}{w}{\bar{C}_w} \mid \mathcal{E}_v}, \
\expect{G_{vu}}{\ttime{t}{w}{\bar{C}_w} \mid \mathcal{E}_v},
\]
and when conditioning on $\mathcal{E}_u$.
To prove~\eqref{eq:diffcondonv} and~\eqref{eq:diffcondonu}, we now show
that, for every $t' \le t$, it holds
\begin{equation}\label{eq:techv}
\expect{G_v}{\ttime{t'}{v}{\bar{C}_v}} -
\expect{G_{vu}}{\ttime{t'}{v}{\bar{C}_v}} \leq
\expect{G}{\ttime{t'}{v}{\bar{C}_v}} -
\expect{G_u}{\ttime{t'}{v}{\bar{C}_v}},
\end{equation}
and
\begin{equation*
\expect{G_v}{\ttime{t'}{u}{\bar{C}_u}} -
\expect{G_{vu}}{\ttime{t'}{u}{\bar{C}_u}} \!\leq\!
\expect{G}{\ttime{t'}{u}{\bar{C}_u}} -
\expect{G_u}{\ttime{t'}{u}{\bar{C}_u}}.
\end{equation*}
We focus on showing~\eqref{eq:techv}, as the same steps, with simple
modifications, can be followed to show the other inquality. For $Z \in \{G,
G_u, G_v, G_{vu}\}$, let $\mathcal{A}_Z$ be the event that a random walk
starting at $v$ reaches $u$ in at most $t$ steps before visiting any vertex in
$\bar{C}_v$, and let $\bar{\mathcal{A}}_Z$ be the complementary event. It holds
\[
\mathbb{P}(\mathcal{A}_{G_v}) = \mathbb{P}(\mathcal{A}_{G_{vu}}) \le \mathbb{P}(\mathcal{A}_{G})
= \mathbb{P}(\mathcal{A}_{G_{u}}),
\]
due to the insertion of $e_v$. It also holds
\[
\expect{G_v}{\ttime{t'}{v}{\bar{C}_v} \mid \bar{\mathcal{A}}_{G_v}} =
\expect{G_{vu}}{\ttime{t'}{v}{\bar{C}_v} \mid \bar{\mathcal{A}}_{G_{vu}}},
\]
and
\[
\expect{G}{\ttime{t'}{v}{\bar{C}_v} \mid \bar{\mathcal{A}}_{G}} =
\expect{G_{u}}{\ttime{t'}{v}{\bar{C}_v} \mid \bar{\mathcal{A}}_{G_{u}}},
\]
Using the law of total expectation (across $\mathcal{A}_Z$ and
$\bar{\mathcal{A}}_Z$) and applying these facts, we can
rewrite~\eqref{eq:techv} as
\begin{align*}
&\left( \expect{G_v}{\ttime{t'}{v}{\bar{C}_v} \mid \mathcal{A}_{G_v}} -
\expect{G_{vu}}{\ttime{t'}{v}{\bar{C}_v} \mid \mathcal{A}_{G_{vu}}} \right)
\mathbb{P}(\mathcal{A}_{G_v})\\
\le &\left( \expect{G}{\ttime{t'}{v}{\bar{C}_v} \mid \mathcal{A}_{G}} -
\expect{G_{u}}{\ttime{t'}{v}{\bar{C}_v} \mid \mathcal{A}_{G_{u}}} \right)
\mathbb{P}(\mathcal{A}_{G}) \enspace.
\end{align*}
The differences between parentheses have the same value, as their
corresponding terms have the same values. The inequality holds because
$\mathbb{P}(\mathcal{A}_{G_v}) \le \mathbb{P}(\mathcal{A}_{G})$ due to the insertion of
$e_v$ in $G$ to obtain $G_v$.
\end{proof}
\begin{figure*}[t]
\begin{center}
\begin{subfigure}[b]{0.22\textwidth}
\includegraphics[width=\textwidth]{CAMERA_mean_graph_guns.pdf}
\includegraphics[width=\textwidth]{CAMERA_bad_nodes_graph_guns.pdf}
\ifextended%
\else
\fi
\caption{Guns
\end{subfigure}%
\begin{subfigure}[b]{0.22\textwidth}
\includegraphics[width=\textwidth]{CAMERA_mean_graph_math_ast.pdf}
\includegraphics[width=\textwidth]{CAMERA_bad_nodes_graph_math_ast.pdf}
\ifextended%
\else
\fi
\caption{MaAs
\end{subfigure}
\begin{subfigure}[b]{0.22\textwidth}
\includegraphics[width=\textwidth]{CAMERA_mean_graph_sociology.pdf}
\includegraphics[width=\textwidth]{CAMERA_bad_nodes_graph_sociology.pdf}
\ifextended%
\else
\fi
\caption{Sociology
\end{subfigure}
\begin{subfigure}[b]{0.22\textwidth}
\includegraphics[width=\textwidth]{CAMERA_mean_graph_politics.pdf}
\includegraphics[width=\textwidth]{CAMERA_bad_nodes_graph_politics.pdf}
\ifextended%
\else
\fi
\caption{Politics
\end{subfigure}
\begin{subfigure}[b]{0.63\textwidth}
\includegraphics[width=\textwidth]{legend.pdf}
\end{subfigure}
\end{center}
\ifextended%
\else
\fi
\caption{The first row shows the $\Delta(G, \Sigma)$ (y-axis) for increasing value of
$k$, reported in terms of $\%\mathcal{L}_G$, the union of possible edges across $\badc{C}{G}$ and $\bar{C}$ for $C \in {R,B}$, (x-axis) for each algorithm. Higher values of
$\Delta$ show more significant reduction of the structural bias. In the second
row, we show the percentage of nodes that are still parochial, $\%\mathcal{P} =
\frac{\card{\pol{}{G}}-\card{\pol{}{\Gnew}}}{\card{\pol{}{G}}}$
after $k$ additions.
\Description{Results of the experiments for different graphs. See the caption
and the text for description of these results.}
\end{figure*}
\section{Conclusion}\label{sec:concl}
We presented \algoname, an algorithm that reduces the structural bias of a graph
by adding $k$ edges. Thanks to the monotonicity and submodularity of the
objective function, \algoname\ is able to return a constant-factor approximation
using a greedy approach based on a task-specific variant of the random walk
closeness centrality. The results of our experimental evaluation show that the
edge insertions suggested by \algoname\ result in a much quicker decrease of the
structural bias than existing methods and reasonable baselines.
The functionality of \algoname \ relies on the existence of an oracle receiving the network and a page in it as input and outputting the transition probabilities of potentially added links to the input page.
We leave the question of designing an algorithm which learns such probabilities from data as future direction of this work.
\section{Experimental evaluation}\label{sec:exper}
The goal of our experimental evaluation is to understand how the addition of the
set $\Sigma = \Sigma_R \cup \Sigma_B$ of $K=k_R+k_B$ edges output by \algonameplus, run separately with $C=R$ and $B$, affects the structural bias of the
network, by computing the gain in the structural bias reduction. In particular, we measure the
gain with $\Delta(G, \Sigma)$, introduced in \cref{sec:algo}, used here with a
simpler notation. We also measure the change
$\card{\bad{G}}-\card{\bad{\Gnew}}$ after adding $\Sigma$.
\textit{Baselines.}
We compare \algonameplus\ to three different baselines (i.e., simplified variants of \algonameplus) and to two existing algorithms, described in the following.
The first baseline, \textit{PureRandom} (PR) selects the source, and the target, nodes of the new edges uniformly at
random from the set $\badc{C}{G}$ and $\bar{C}$, respectively. The second baseline \textit{Random Top-$N$ Central Nodes
($N$-RCN)}, given a parameter $N\in (0,100)$, sorts the nodes in $\badc{C}{G}$
by descending centrality, and picks, uniformly at random, $k_C$ edges
with source in the top-$N$ percent of nodes in $\badc{C}{G}$. The last baseline,
\textit{Random Top-$N$ Weighted Central Nodes ($N$-RWCN)}, differs from $N$-RCN
as the nodes in $\badc{C}{G}$ are sorted in descending order by
$\mathcal{R}(v) \times m_{v,u}$.
We compare \algonameplus\ also to two existing methods, ROV~\cite{garimella2017reducing}, and node2vec~\cite{grover2016node2vec}. The \textit{ROV} algorithm outputs a set of $k$ edges to be added to $G$ to minimize the controversy score (RWC)~\cite{garimella2018quantifying}. The RWC is a metric that characterizes how controversial a topic is by capturing how well separated the two colors are. ROV considers as candidates the edges between the high-degree vertices of each color~\citep[Algorithm 1]{garimella2017reducing}. These edges are sorted by descending impact on the graph controversy score, and the top-$k$ edges are added to the graph. The objective of the comparison between ROV and \algonameplus\ is to verify whether an algorithm developed to minimize the RWC can be used to minimize the structural bias.
\textit{node2vec} is a graph embedding technique that encodes a network in a low-dimensional space retaining characteristics like the nodes' similarity~\citep{grover2016node2vec}. The generation of the embedding is based on random walks. One of the main applications of node2vec is to employ the embedding as the feature space to train link recommendation algorithms. The goal of comparing node2vec to \algonameplus\ is to understand how the predictions of widely-used link recommendation algorithms affect the network's structural bias. In the experiments, we create for each network a 128-dimensional space, then we train a logistic regression (\textit{avg.} AUC 85\%) over these features, and we predict the existence probabilities of edges from $\bad{G}$. We add to the graph the top $k$ edges according to these probabilities.
\textit{Datasets.}
We create graphs obtained from \textit{Wikipedia
, \textit{Amazon}\footnote{\url{https://snap.stanford.edu/data/amazon-meta.html}} and
\textit{PolBlogs}\footnote{\url{http://www-personal.umich.edu/~mejn/netdata/}}. \Cref{tab:info} shows the relevant statistics.
From \textit{Wikipedia} we consider four bi-partitioned subgraphs related to controversial topics: \textit{politics, abortion, guns} and \textit{sociology}~\cite{menghini2020wikipedias}.
Each node in the graph is a page, and is assigned to one color according to Wikipedia's categorization. Directed edges denote links, and are weighted using Wikipedia's clickstream data.\footnote{\url{https://dumps.wikimedia.org/other/clickstream/}}
The \textit{Amazon} dataset contains metadata about \textit{books}~\cite{leskovec2007dynamics}.
Given two book categories, the vertices are all the items in those categories, colored accordingly. There is a directed edge $(u,v)$ if $v$ appears in the list of items similar to $u$. The edge is weighted by $v$'s sales rank.\footnote{Amazon sales rank is a metric of the relationship among products within one category based on their sales performance. It
expresses how well a product is selling relative to other products in the same category.} We built three graphs by considering pairs of the following categories: \textit{Mathematics \& Technology (MaTe)}, \textit{ History of Technology \& Military Science (MiHi)}, and \textit{Mathematics \& Astronomy (MaAs)}.
The \textit{Political Blogs} dataset is a directed network of hyperlinks between weblogs on US politics~\cite{adamic2005political}. Each node represents a blog and is colored according to its political leaning.
Links between blogs were automatically extracted from a crawl of the front page of the blog and represent the edges of the graph. Each edge $(v,u)$ has weight proportional to the out-degree of $v$.
\input{data_stats}
\textit{Setup.}
Given a network, we run \algoname\ and the other algorithms on that network for increasing values of $K$, with $K=1,2,4,6,\dotsc,400$ or $2000$ for larger graphs (\textit{Sociology} and \textit{Politics}). These values of $K$ represent only a small percentage of the set of possible edges to insert and correspond to the total number of edges to add to the graph. Once we set the value of K, accordingly, we allocate $k_B$ and $k_R$ of the $K$ edge insertions to each color proportionally to the sum of the BRs of the parochial vertices in each color.
In particular, we define $Y_C = \sum_{v \in \badc{C}{G}} \bubble{G}{v}{t}$, for $C \in {R,B}$, then $k_B = \left\lceil k \frac{Y_B}{Y_B+Y_R}\right\rceil$ and $k_R = K - k_B$. This allocation strategy is a simple but reasonable heuristic that ensures that more edges are added from nodes whose color is more parochial.
We assign the weight $m_{v,u}=1/(d(v)+1)$ to the added edge $(v,u)$, where $d(v)$ is the out-degree of $v$ before the insertion, and then we re-normalize the weights of the other edges by multiplying each of them by $1-m_{v,u}$. Furthermore, we set $\badthres=5$\ and $\goodthres=2$. Moreover, for the algorithms picking the top-N central nodes $N=10$. To account for variability of the algorithm, we run them 10 times. The variance of the results is low, overall.
The code for our experiments is available from \url{https://github.com/CriMenghini/RePBubLik}.
\textit{Experiment results.}
In \cref{fig:exp:total}, the plots in the first row show how the structural bias
is affected by the insertion of an incrementally larger set of edges, while the
ones on the second row show the reduction in the number of parochial nodes. Each
curve in the plot illustrates the gain by a different algorithm. We can draw the
following observations. (1) \algonameplus\ performs better than the
baselines and the competitors, especially after the insertion of a few edges, as
they obtain much larger gain with fewer insertions, i.e., the average BR of
parochial nodes decreases faster requiring less modifications. (2) N-RCN, N-WRC,
and ROV after a certain point become flat. (3) Overall, \algonameplus\ is the best
algorithm.
(4) The values of \algonameplus and PR converge, at different speed, to the same
value when we add more edges. (5) node2vec, in the best cases, shows little
improvement of the structural bias that, in the remaining cases, stays flat or
even increases. We now explain these behaviours using the plots on the second
row of \cref{fig:exp:total}.
(1) \algonameplus\ chooses edges that directly affect the BR of central nodes and,
with a chain effect, the BR of nodes connected to them. More central are the
nodes we attach the edges to, higher the structural bias drop is. In fact, it
follows, as shown for all the networks, that the addition of even small set of
edges is very effective. Additionally, we observe that the structural bias
reduction corresponds to a significant drop of the number of parochial nodes.
(2) N-RCN, N-WRC, and ROV attach edges only to a subset of $\bad{G}$ and as $k$
increases, so does the probability of adding multiple edges to the same nodes.
These facts imply respectively that, especially on disconnected graphs (see MiHi
in \cref{fig:exp:algo:mihi:red}), the addition of edges may affect few nodes,
and that even the insertion of more edges does not modify the set of nodes on
which the new edges have effect. Thus, the curves of N-RCN, N-WRCN and ROV reach
an early saturation that expresses the scarce impact of subsequent edge
additions. This explanation is confirmed by the percentage of parochial nodes,
which does not decrease after the saturation point. Furthermore, the ROV shows a
stepping behaviour due to it selecting edges between high-degree central nodes
that minimize the RWC without imposing diversity constraints on nodes. And
resulting in many selected edges being attached to the same node. Last, we see
that on \textit{Polblogs} the best algorithms are N-RCN, N-WRCN\@. This surprising
superiority of the random approaches can be explained by the fact that
\textit{Polblogs} is a connected graph, thus edges added to the top-central
nodes potentially affect all the nodes in $\bad{G}$. Thus, even when N-RCN and
N-WRCN add multiple new edges to the same set of nodes, $\Delta$ continues to
increase.
(3) \algonameplus\ shows a consistent behaviour, indeed it increases the gain faster than
other methods, requiring fewer insertions. The penalty factor $\eta$ allows the
algorithm to diversify the set of nodes to which the new edges attach, raising
the chances of lowering the BR of a larger number of parochial nodes, thus
increasing the gain. This feature is important especially on disconnected
graphs, where the vertices in tiny connected components always have lower
centrality compared to those in huge ones. More importantly, we observe that the
size of $\bad{G}$ is often reduced to 0: \algonameplus\ is able to ``heal'' all the bad
vertices, and if we measured the structural bias on the obtained graph it would
be zero.
(4) The variants of \algoname: \algonameplus\ and PR, pick edges from the same
candidate set, thus the more edges they can pick, the more likely they choose
edges with similar effect, thus the average parochial nodes' BR converges. This is the main explanation why the random algorithm performs so well.
(5) Generally, link recommendation algorithms tend to suggest edges between
similar nodes. Node2vec captures this similarity through the nodes'
neighborhood. In this context, graphs partitions have high within- and low
between-density. Nodes in the same partition then lie close in the
embedding space. Edges suggested by node2vec with high probability connect nodes
close to each other in the embedding, which often are in the same partition.
Thus, node2vec has a hard time reducing the structural bias, and in some cases
increases it.
\ifextended%
Plots for \textit{guns}, \textit{sociology}, \textit{politics} and \textit{MaAs}
show similar behaviour and can be found in \cref{sec:appendix}.
\else
Due to space constraint we omit the presentation of the plots for \textit{guns},
\textit{sociology}, \textit{politics} and \textit{MaAs}, which show similar
behaviour. But they can be found in the extend online version~\citemissing
\fi
\section{Introduction}\label{sec:intro}
The World Wide Web often contains thousands or even millions of pages on every
topic, covering the whole spectrum of opinions. Exposure to \emph{diverse
content} is necessary to obtain a complete picture about a topic.
This exposure depends on the hyperlinks connecting the pages to each other. It
can be argued that enabling easier access to diverse content improves society as
it creates a more informed and less polarized general
public~\citep{benhabib1996toward}. Indeed politicians have strongly promoted and
even requested that audiences are exposed to varied content~\citep{LeFebvre17}.
The fact that diverse information is easily \emph{available} does not imply that
\emph{exploring} such diverse information is easy. Rather, echo chambers and
polarization on (social) media and
blogs~\citep{adamic2005political,conover2011political,flaxman2016filter} keep
the user in a \emph{homogeneous bubble}, exposing them only to agreeable
information~\citep{bakshy2015exposure}, and leading to conflicts between users
in different bubbles~\citep{KumarHLJ18,CossardDFMKMPS20}.
A web user can freely click on any hyperlink on the page they are currently
visiting, but the choice of which hyperlinks to include in the page is with the
website owner or editor, who, if not careful, may stop the user from being
exposed to diverse opinions. In other words, the hyperlink topology of a website
may suffer from \emph{structural bias} that traps the user in a bubble of
one-sided content without them knowing~\citep{horta2020youtube,menghini2020wikipedias}.
For example, structural bias on topic-induced networks, such as Wikipedia
topic-induced subgraphs, prevents users from building a well-rounded knowledge
about the topic. On query-/user-induced recommendation networks such as those on
Amazon and YouTube, structural bias hinders the discovery of diversified
content, reducing
serendipity~\citep{mouzhi2010serendipity,kotkov2016chalserendipity,anagnostopoulos2019principal}.
Structural bias thus limits the user's freedom while navigating the Web.
Socially-minded website editors would try to \emph{minimize such structural
bias} by adding appropriate links to pages in the website. As an editor can only
modify a few pages and add only a few links to each edited page, they need to
carefully choose what page $p$ to edit, and what pages to link to from $p$. The
goal of our work is to develop an algorithm that can give editors
recommendations for what links to add in order to reduce the structural bias.
There are key technical challenges that must be solved in order to give
effective link recommendations: \textit{1.} quantify the structural bias of a
page $p$, i.e., how hard it is to reach pages of a different opinion from $p$;
and \textit{2.} decide which of these pages should be linked from $p$. Existing
approaches to link recommendation, such as those based on vertex similarity,
fall short at this task because they are oblivious to the network structural
bias. Their recommendations often \emph{increase} the bias, rather than
decreasing it~\citep{MuscoMT18}, especially for highly controversial topics such
as political blogs (see also our experimental results in \cref{sec:exper}).
Thus, there is a need for a radically different approach to suggest links that
decrease the structural bias.
\paragraph{Contributions.} We study the problem of reducing the structural bias
of a graph by adding edges, and propose an algorithm, \algoname, to suggest such
edges. Our contributions are the following.
\begin{itemize}
\item We consider directed graphs with vertices of two colors, representing a
network of webpages on the same topic, with the two colors identifying the
two opposite opinions on the topic, and edges representing links
between pages. We define the \emph{(Polarized) Bubble Radius} (BR) of a
vertex $p$ as a novel measure to quantify the structural bias of $p$ (see
\cref{def:bubbleradius}), based on a task-specific variant of the hitting
time for random walks, which models the navigation of a user on the web
\citep{rwwithbackbutton,rwwithrestart}. The BR is the expected number of
steps to go from $p$ to a page of different opinion, and can be easily
estimated with a sampling-based approach with probabilistic guarantees
(\cref{lem:bubbleapprox}), which enables us to tackle the first of the key
challenges.
\item We define the \emph{structural bias} of a graph $G$ as the sum of the
BRs of vertices with high BR (Eq.~\ref{eq:bias}). Completely removing the
bias is APX-hard by reduction from set cover (see \cref{lem:reduction}).
We therefore state the \emph{$k$-edge structural bias decrease} problem as
the task of finding the set of $k$ pairs of vertices of different color
such that adding the edge between the vertices in each pair would
\emph{maximally} decrease the structural bias, over all possible sets of $k$
pairs (see \cref{prob:kinsertions,thm:main}). This problem
connects two areas: link recommendation and polarization reduction.
\item We present \algoname, an efficient approximation algorithm for
the $k$-edge structural bias decrease problem, that recommends the addition
of $k$ edges between vertices of different color. Under mild conditions, the
resulting decrease of the structural bias is within a constant factor of the
optimal. Website editors have limited control on the probability that a
newly added edge will be traversed by the users, so our algorithm makes no
assumption or impose any restriction on it, as this probability is
essentially external. At the core of \algoname\ is an analysis of the
submodularity of the objective function (see~\cref{lem:submodular}),
combined with the use of a task-specific variant of random-walk closeness
\citep{rwclosenessWhiteSmyth}, a well-established centrality measure.
\algoname\ requires good estimations of the random walk closeness, so we
also give an approximation algorithm for this quantity (see
\cref{lem:centr}).
\item We evaluate \algoname\ on eight real datasets. We compare it to
baselines and existing methods for edge recommendation either designed with
the goal of reducing the controversy of a
graphs~\citep{garimella2017reducing} or with the more general purpose of
completing the network's link structure~\citep{grover2016node2vec}. Our
algorithm leads to a faster reduction of the average BR (i.e., requiring
fewer edge insertions) than existing contributions.
\end{itemize}
\ifextended%
The proofs of our theoretical results are in \cref{sec:appendix}.
\else
Due to space limitations, many of our proofs are in the appendix of the extended online version~\citemissing.
\fi
\section{Preliminaries}\label{sec:prelims}
Let $G=(V,E)$ be a directed weighted graph with $\card{V} = n$ vertices, such
that no vertex $v \in V$ has only incoming edges and no outgoing edges. $V$ is
partitioned in two disjoint sets $R$ and $B$ (i.e., $R \cap B = \emptyset$ and
$R \cup B = V$), called ``red'' or ``blue'' vertices, respectively. We denote
the color of a vertex $v$ by $\vcolor{v}$ and its opposite color by
$\ovcolor{v}$. The sets of all \emph{other} vertices of the same color as $v$ is
denoted as $C_v$ and the sets of all vertices of color different than $v$ is
denoted as $\bar{C}_v$.
The edge weights are \emph{transition probabilities}, as follows. Let $\trans$
be a $n \times n$ right-stochastic \emph{transition matrix} associated to $G$,
i.e., a matrix such that each entry $m_{i,j}$ is a probability, with $m_{i,j}=0$
if $(i,j) \notin E$, and such that $\sum_{j=1}^n m_{i,j}=1$.
We are interested in random walks on the graph $G$ using the transition matrix
$\trans$. Intuitively, a random walk starting at a vertex $v$ explores the graph
by choosing at each step an outgoing edge from the current vertex, with
probability equal to the weight of such edge, independently from previous
choices. Let $S \subseteq V$ and $v \in V$. Let $T_v(S)$ be the random variable
indicating the first instant when a random walk from $v$ hits (i.e., reaches) any
vertex in $S$. The quantity $\expect{G}{T_v(S)}$ is known as the \emph{hitting
time of $S$ from $v$}, where the expectation is over the space of all random
walks on $G$ starting from $v$, with transition probabilities given by $\trans$.
Variants of random walks, such as random walks with restarts or with back
button, are widespread models for network
exploration~\citep{rwwithbackbutton,rwwithrestart}. It is realistic to assume
that there is an upper bound $t$, which we call the \emph{exploration
factor}, on the length of a walk performed by the users. For example, we can
assume that there is an upper limit on the number of pages that a user will
visit one after the other in a browsing session. The value of the parameter $t$
can be derived, for example, from traces of visits. In most practical cases, $t$
is likely to be bounded by a polylogarithmic quantity in the number of nodes, if
not a constant.
For a random walk starting from $v \in V$, given a set $S \subseteq V$, we
define the random variable $\ttime{t}{v}{S}$ as $\min \{t, T_v(S)\}$.
This variable is more appropriate for measuring the length of browsing sessions,
which have bounded length, than the unbounded length classically used when
discussing random walks.
For a graph $Z$, any vertex $u$, and any set $S$ of vertices, let
$\towithin{u}{S}{\mathsf{cond}}{Z}$, denote the event that a random walk in $Z$
from $u$ hits a vertex in $S$ without first visiting any vertex in $\bar{C}_u$
and while satisfying the condition $\mathsf{cond}$ on the number of
steps needed to hit $S$. For example, $\towithin{u}{S}{< t}{Z}$ is the event
that a random walk in $Z$ from $u$ hits a vertex in $S$ in \emph{less than $t$
steps}, without first visiting any vertex in $\bar{C}_u$. We denote the
complementary event as $\nottowithin{u}{S}{\mathsf{cond}}{Z}$.
\subsection{Random-Walk Closeness Centrality}\label{sec:prelims:rwcc}
We adapt the definition of the standard random-walk closeness
centrality~\citep{rwclosenessWhiteSmyth} to bounded random walks so that the
contribution to the centrality of $v$ by vertices that do not reach $v$ in less
than $t'$ steps (in expectation) is zero, for any $t'$.
\textbf{Random-walk closeness centrality (bounded form).} For a vertex $v \in
V$, and any $t'$, the \emph{$t'$-bounded} Random Walk Closeness Centrality
(RWCC) measure with respect to subset $S \subseteq V$ is
\begin{align*}
\centr{t'}{v}{S} &\doteq \frac{1}{\card{S}} \sum_{w \in S} \left( t' -
\expect{G}{\ttime{t'}{w}{v}} \right) \\
& = \frac{1}{\card{S}} \sum_{w \in S}
\sum_{i=1}^{t'} (t'- i) \mathbb{P}\left( \towithin{w}{v}{=i}{G} \right) \enspace.
\end{align*}
Computing the exact RWCC is expensive. To estimate $\centr{t'}{v}{S}$, we
pick $z$ vertices ${\{w_i\}}_{i=1}^r$ u.a.r.\ from $S$, and run some $\kappa$
random walks to obtain an estimate $\bar{h}_{w_{i}}$ of
$\expect{G}{T^{t'}_{w_i}(v)}$ for each $w_i$. The quantity $\bar{r}(v) \doteq
t' - \nicefrac{1}{z} \sum_{i=1}^z \bar{h}_{w_{i}}$ is a good approximation of
$\centr{t'}{v}{S}$.
\begin{restatable}{lemma}{lemcentr}\label{lem:centr}
Let $z \geq {(\nicefrac{t'}{2\varepsilon})}^2 \delta^{-1}$. Then
\[
\mathbb{P}\left(\card{\bar{r}(v) - \centr{t'}{v}{S}} \geq \varepsilon \right) \leq
\delta \enspace.
\]
\end{restatable}
\section{Related work}\label{sec:related}
Polarization has long been studied in political
science~\citep{Sunstein02,Isenberg86}, and the recent diffusion of (micro-) blog
and social media platforms brought the issue to the attention of the broad
computer science community. Many works focused on showing the existence of
polarization on these
platforms~\citep{MoralesBLB15,adamic2005political,CossardDFMKMPS20,conover2011political,flaxman2016filter},
and on modeling, quantifying, and reducing
polarization~\citep{garimella2018quantifying,garimella2017reducing,MuscoMT18,ChitraM20,MakatosTT17,BeckerCDAG20,GarimellaGPT17,MatakosTG20,AslayMGG18,Akoglu14,GarimellaDFMGM18,MakatosTT17,NelimarkkaLS18,LiaoF14a,LiaoF14b,MunsonLR13},
or the glass ceiling effect~\citep{StoicaC19,StoicaRC18,StoicaHC20}.
The literature is rich, to
the point that times seems ripe for an in-depth survey on the topic. Due to
space limitations, we discuss here only the relationship between our
work and the most relevant algorithmic contributions to polarization
reduction~\citep{garimella2017reducing,ChitraM20,MuscoMT18,AslayMGG18,BeckerCDAG20,GarimellaGPT17,MatakosTG20,menghini2019wikipol,menghini2020wikipedias,StoicaRC18}.
A first important difference of our work with respect to most previous
contributions is that they consider a network of \emph{users}, with edges
representing notions such as friendship or endorsement (e.g.,
retweets)~\citep{garimella2017reducing,ChitraM20,MuscoMT18,AslayMGG18,BeckerCDAG20,GarimellaGPT17,MatakosTG20,StoicaRC18}.
We focus instead on networks of \emph{content}, such as web pages linked to each
other, or products that are connected when similar. This deep difference makes
our contribution quite orthogonal to the ones in these previous works: we focus
on the polarization that is introduced by the topology of the network, rather
than on the polarizing effect of content on users or on the effect of users on
each other. We believe both aspects are important, but the structural bias we
focus on has only been subject to few
studies~\citep{menghini2019wikipol,menghini2020wikipedias}. These works, relying
on the notion of weighted reciprocity, propose a static and dynamic analysis of
structural bias on Wikipedia. The measure of structural bias we use is not
tailored to a specific website.
A second relevant difference from many previous works is that we consider the
``opinion'' of a page (i.e., a vertex) to be fixed, as it depends on its
content, while many previous contributions consider different models of user
opinion dynamics~\citep{MosselT17,DasGM14} to study the evolution of such
opinions as the users are exposed to different content or recommended different
friendships. The problem of recommending changes to the content of a page to
modify the opinion expressed in it is interesting but outside the scope of our
work. Instead, we focus on recommending the addition of links between pages, to
reduce the structural bias.
An interesting line of work studies how to reduce polarization in the content
seen by the users, by adapting information diffusion approaches through better
selection of the seed set for
cascades~\citep{AslayMGG18,BeckerCDAG20,GarimellaGPT17,MatakosTG20,StoicaHC20},
or by directly acting on recommendation systems~\citep{RastegarpahahGC2019}.
These methods can not be adapted to the problem we study, as they do not act on
the graph of content, but on that of users.
The most similar methods to ours are those that act on the structure of the
graph~\citep{ChitraM20,MuscoMT18,garimella2017reducing,StoicaRC18}, although as
we mentioned, they consider a network of users, not of content.
\citet{MuscoMT18} propose a network-design approach: they aim to find the best
set of edges between vertices such that the resulting graph would minimize both
disagreement and polarization. Rather than a ``design-from-scratch'' approach,
which seems mostly of theoretical relevance, we consider instead a practical
incremental approach that suggests modifications to an existing network.
Like us, \citet{garimella2017reducing} consider a graph polarization measure
based on random walks~\citep{garimella2018quantifying}. This measure essentially
quantifies the probability that a user of one opinion is exposed to content from
a user of a different opinion, thanks to a chain of retweets (represented by the
random walks). The measure is based on a variant of personalized PageRank for
sets of users with different opinions. The task requires to recommend new edges,
i.e., retweets, to increase this probability. Our measure of structural bias is
instead defined on the basis of
the (Polarized) Bubble Radius (BR) (\cref{def:bubbleradius}), which is a
vertex-dependent measure that represents the expected number of steps, for a
user starting at the page represented by vertex $v$, to reach, with a random
walk, a vertex with color different from $v$, representing a page expressing a
different opinion. Our measure is appropriate for our task of suggesting new
edges to make it easier for user to reach pages of different opinions. In
\cref{sec:exper} we compare our approach to that of
\citet{garimella2017reducing}.
An important line of work in graph analysis and mining looked at manipulating
the topology to modify different interesting characteristic quantities of
the graph, such as shortest paths and related
measures~\citep{parotsidis2015selecting,papagelis2011suggesting,demaine2010minimizing,perumal2013minimizing},
various forms of
centrality~\citep{parotsidis2016centrality,bergamini2018improving,d2019coverage,wkas2020manipulability,medya2018group,mahmoody2016scalable,angriman2020group},
and more~\citep{arrigo2016edge,arrigo2016updating,chan2014make,tong2012gelling,zeng2012manipulating}.
Despite the fact that we consider a specific centrality to choose the source of
the added edges, these methods cannot be used to solve our task of interest.
Another body of work related to ours are those which estimate graph properties using random walks \cite{Berarw,ChiHad2018rw,peres2018rw,avegragerw}. The studied properties are not defined based on random walks, rather random walks are used as a tool to estimate them. Here based on random walks, we define a new property for networks: \emph{the structural bias}, and we use random walks to estimate it. |
\section{Introduction}\label{sec:intro}
\input{intro}
\section{Related Work}\label{sec:related}
\input{related}
\section{Method}\label{sec:overview}
\input{method}
\section{Implementation Details}\label{sec:implementation}
\input{implementation}
\section{Evaluation and Results}\label{sec:results}
\input{results}
\section{Conclusions}\label{sec:conclusions}
\input{conclusions}
\vspace{-2mm}
\section*{Acknowledgments}
The authors would like to thank Stan Birchfield for the inspiring discussions on the theory of time-to-contact, Gengshan Yang and Junhwa Hur for their kind help in the comparisons with previous works, and anonymous reviewers and ACs for their helpful suggestions.
{\small
\bibliographystyle{ieee_fullname}
\subsection{TTC Data Generation}\label{sec:data_ttc_from_sceneflow}
A dataset providing per-pixel TTC ground truth does not exist.
However, we can use ratio of the depth in different frames to compute the TTC with Equation~\ref{eq:ttc_from_depth_discrete}.
This ratio can be computed from datasets offering per-pixel scene flow $[X(t_0),Y(t_0),Z(t_0)] \rightarrow [X(t_1),Y(t_1),Z(t_1)]$.
To train for TTC estimation we use the Driving dataset from the SceneFlowDatasets~\cite{mayer2016large}.
We also use the KITTI15~\cite{Menze2015kittidataset} dataset to train for TTC estimation.
We split the training dataset of KITTI15 into train and validation split, and show analysis on the validation part of the dataset.
To pre-train our network for the task of binary optical flow, we use the FlyingChairs2~\cite{dosovitskiy2015flownet, iig2018generic} and the FlyingThings3D~\cite{mayer2016large} datasets.
We also use optical flow data available from Driving~\cite{mayer2016large} and KITTI15~\cite{Menze2015kittidataset}, and train our network for both binary optical flow and TTC estimation.
\subsection{Working in the Inverse TTC Domain}\label{sec:inverse_ttc_domain}
We are interested in objects predicted to collide within $\tau_i$ seconds.
However, $\tau \in (-\infty,\infty)$, with positive and negative values indicating objects moving towards and away from the camera, respectively.
That is, some of the pixels whose TTC is smaller than $\tau_i$ will never collide with the camera plane.
In practice, then, we seek to identify the pixels such that
\begin{equation}\label{eq:interesting_ttc_range}
(\tau(x,y) \leq \tau_i) ~~ \cap ~~ (\tau(x,y)>0),
\end{equation}
which introduces an additional complication.
Moreover, the effect of TTC in the image space in not linear.
A simple solution is to work with the ratio-of-depths, the inverse of the TTC domain:
\begin{equation}\label{eq:ratio_of_depths}
\eta = \frac{Z(t_1)}{Z(t_0)} = 1 - \frac{t_1 - t_0}{\tau}.
\end{equation}
Yang and Ramanan termed it motion-in-depth~\cite{yang2020of2ttc}.
Thanks to Equation~\ref{eq:ratio_of_depths}, the effect of TTC is linear in image space and Equation~\ref{eq:interesting_ttc_range} simply reduces to
\begin{equation}\label{eq:eta}
\eta(x,y)\leq \eta_i.
\end{equation}
For binary TTC, then, we simply scale the features of the source image by a factor $\alpha_i = \eta_i$.
Similarly, for continuous estimation, we sample planes uniformly in the inverse TTC domain.
We apply uniformly spaced scale factors to the features of the source image: $\{\alpha_i\}_{i=1:N} = \{\alpha_0 + i\Delta\alpha\}_{i=1:N}$.
The AUC of the resulting segmentation gives us the map $\eta(x,y)$, which we then covert to a TTC map.
\begin{figure*}
\vspace{-3mm}
\centering
\input{figures/results/results_table_new.tex}
\vspace{-2mm}
\caption{\textbf {Comparison on continuous TTC estimation.} We show the predicted TTC and the error map (red and green indicate high and low error, respectively) for all methods. Our method and Yang and Ramanan's method explicitly train for TTC estimation and provide better results than the monocular scene flow method of Hur and Roth. Note the lower error in the TTC estimation for our method.}\label{fig:all_results}
\vspace{-3mm}
\end{figure*}
\subsection{Architecture}\label{sec:architecture}
Figure~\ref{fig:arch} shows our architecture.
The feature extraction module uses spatial pyramid pooling layers as PSMNet~\cite{chang2018pyramid}.
The resulting 32-channel feature maps are at one-third the input image resolution.
We then apply a task-dependent operator,\cP, parametrized by {\color{blue}$\phi$}, to $f_1$ and generate $f_1^\phi$.
For instance, when estimating the horizontal component of binary optical flow,\cP shifts the features by a horizontal shift $u_i$, and for binary TTC, it scales them by $\alpha_i$.
The full list for each task is in the table in Figure~\ref{fig:arch}.
To avoid cropping out features, we zero-pad $f_0$ and $f_1^\phi$ to $1.5\times$ the feature map resolution.
The concatenation of $f_0$ and $f_1^\phi$ is fed to a 2D encoder-decoder network with skip connections.
This module has three heads---one for binary TTC, one for binary horizontal optical flow, and one binary vertical optical flow.
We apply a sigmoid to each output to obtain the respective probability maps.
Intuitively, this network tells us if the features in $f_1^\phi$ are scaling up/down, shifting right/left, or shifting up/down with respect to the corresponding features in $f_0$.
\subsection{Training}\label{sec:training}
We pre-train our network for the binary optical flow task on the FlyingChairs2 dataset and train it further on the FlyingThings3D dataset~\cite{mayer2016large} using a binary cross-entropy (BCE) loss with respect to a thresholded version of the ground truth.
The TTC head is left unsupervised in this stage.
We then fine-tune our network for estimating binary TTC first on the Driving~\cite{mayer2016large} and then on the KITTI15~\cite{Menze2015kittidataset} datasets.
Since both datasets also offer optical flow data, in this second stage we train for both binary optical flow and binary TCC.
In Section~\ref{sec:results} we discuss the impact of this choice on the quality of the results.
We use relative weights of $0.8$ and $0.2$ for binary TTC and binary optical flow tasks respectively.
For training on the KITTI15 dataset, we use the same split as Yang and Ramanan~\cite{yang2020of2ttc}.
For the continuous version, we pre-train the network as before, and
fine-tune it on FlyingThings3D for continuous optical flow.
For each training image pair we uniformly sample horizontal and vertical shifts and stack the maps corresponding to each shift.
We use the resulting volumes to compute continuous optical flow via the AUC operation described in Section~\ref{sec:ttc_continuous}.
We can then continue training the network using a BCE loss on the individual probability maps, and a Smooth-L1 (SL1)~\cite{girshick2015fastrcnn} regression loss on the output of the AUC module.
We use relative weights of $0.1$ and $0.9$ respectively.
To fine-tune our network for continuous TTC estimation, we follow a similar strategy as for the binary segmentation training.
We continue training the network for the task of continuous optical flow and continuous TTC estimation on the Driving and the KITTI15 datasets.
However, this time we form three volumes, two corresponding to optical flow and one to TTC.
As discussed in Section~\ref{sec:inverse_ttc_domain}, we work in the inverse TTC domain and uniformly sample scale factors.
The network trained on the entire KITTI15 dataset is used for scene flow estimation on KITTI15 benchmark images, as explained in Section~\ref{sec:results}.
We refer the reader to our Supplementary for the additional training details.
\subsection{A Review on Time-to-Contact}\label{sec:ttc_overview}
Consider two frames of a static scene captured by a moving camera.
Pixel-level correspondences allow us to compute depth, if the camera information is known.
In the more realistic case of dynamic scenes, however, the problem becomes ill-posed:
the displacement of a pixel is the result of its depth, its velocity, and the camera velocity, all of which cannot be disambiguated without strong priors.
In this case, the concept of time-to-contact (TTC) comes to the rescue.
Given an object $\mathcal{O}$, the TTC $\tau$, \ie, the time at which object $\mathcal{O}$ will (or did) cross the camera plane, can be written as
\begin{equation}\label{eq:ttc_from_depth_ratio}
\tau = - Z_{\mathcal{O}} \bigg/ {\frac{dZ_{\mathcal{O}}}{dt}} = - Z_{\mathcal{O}} / \dot{Z}_{\mathcal{O}},
\end{equation}
where the origin is at the camera, and we assume that the current velocity conditions will continue.
$Z_{\mathcal{O}}$ is the depth of the object from the camera plane, and $\dot{Z}_{\mathcal{O}}$ its relative velocity.
Equation~\ref{eq:ttc_from_depth_ratio} shows the first appealing feature of TTC:
even if the depth and the velocity of the object cannot be estimated independently, $\tau$ can be computed from their ratio.
However, we are interested in computing the TTC from pixel displacements alone.
To do that, we need an additional piece of information: the location of the focus-of-expansion (FOE).
Given two frames of a static scene captured by translating the camera, all the pixels in the image move along lines originating from the FOE, the image of the point towards which the camera is moving.
Under the same assumptions, the FOE coincides with the epipole.
The relationship between the TTC, $\tau$, and the velocity of the pixel is given by
\begin{equation}\label{eq:ttc_from_velocity}
\dot{x} = \frac{x-x_0}{\tau} \quad \text{and} \quad \dot{y} = \frac{y-y_0}{\tau},
\end{equation}
where $(x_0,y_0)$ is the FOE.
Equation~\ref{eq:ttc_from_velocity} can be easily derived by differentiating the projection of a 3D point onto the image plane with respect to time~\cite{horn2007ttc_planar}.
Note that the velocity $(\dot{x}, \dot{y})$ can be computed from the optical flow $(u,v)$, which allows us to write
\begin{equation}\label{eq:ttc_from_flow}
\tau = \frac{x-x_0}{u} \cdot T \quad \text{and} \quad \tau = \frac{y-y_0}{v} \cdot T,
\end{equation}
where $T$ is the time elapsed between the two frames.
Equation~\ref{eq:ttc_from_flow} shows a second compelling reason to use TTC:
there is no need for camera calibration.\footnote{We assume a pin-hole camera model.}
There are also challenges, however.
If a rigid object is dynamic and translates with respect to the scene, its pixels move along lines centered around a different FOE. To estimate the TTC using Equation~\ref{eq:ttc_from_flow}, then, we need to localize the FOE of \emph{each} dynamic object.
Moreover, if an object is deformable, the FOE varies with each pixel, and for general motion, we need to further compensate for rotations.
This is why traditional approaches had to rely on oversimplifying assumptions.
However, we can compute the size of the image of a fronto-parallel, planar, non-deformable object of size $S_{\mathcal{O}}$ at distance $Z_{\mathcal{O}}$ as~\cite{horn2007ttc_planar}
\begin{equation}\label{eq:scale_Z_inverse_relation}
s_{\mathcal{O}} = fS_{\mathcal{O}}/Z_{\mathcal{O}},
\end{equation}
where $f$ is the focal length of the camera.
Since $f$ and $S_{\mathcal{O}}$ are constant, differentiating Equation~\ref{eq:scale_Z_inverse_relation} yields
\begin{equation}\label{eq:Z_to_scale}
Z_{\mathcal{O}}/\dot{Z}_{\mathcal{O}}= -s_{\mathcal{O}}/\dot{s}_{\mathcal{O}},
\end{equation}
which, plugged into Equation~\ref{eq:ttc_from_depth_ratio}, allows us to estimate the TTC using only information about the size of an object's image and its rate of change:
\begin{equation}\label{eq:ttc_from_scale}
\tau = s_{\mathcal{O}}/\dot{s}_{\mathcal{O}}.
\end{equation}
Note that both Equation~\ref{eq:ttc_from_flow}~and~\ref{eq:ttc_from_scale} effectively look at scaling.
However, the latter is independent of the point with respect to which the scaling is performed, whereas for the former, that point is the FOE.
Of course, under more realistic conditions (\eg, not planar or not fronto-parallel objects), Equation~\ref{eq:ttc_from_scale} becomes an approximation, which requires proper handling, as we show later.
\begin{figure}
\vspace{-4mm}
\centering
\input{figures/binary/binary.tex}
\caption{\textbf{Binary TTC maps.} Our method can directly identify the pixels that will contact the camera plane before a given time. From top to bottom, we show results for four TTC values for a case of highway driving and for a camera that rotates.}\label{fig:binary_results}
\vspace{-4mm}
\end{figure}
\subsection{TTC via Multiple Binary Classifications}\label{sec:ttc_as_classification}
In this section, we first provide the intuition motivating our method.
We then describe binary TTC, the core of our method, which detects pixels predicted to collide with the camera plane within a given time.
We further show how an arbitrary number of binary classifications can be combined to estimate, for each pixel, a quantized version of the TTC.
To simplify the description, here we assume positive TTCs, \ie, we focus on objects moving towards the camera rather than away from it.
In Section~\ref{sec:implementation}, we describe a small adaptation of our method that allows us to seamlessly remove this distinction.
\subsubsection{Intuition}\label{sec:ttc_intuition}
Given two images captured at times $t_0$ and $t_1$, Equation~\ref{eq:ttc_from_depth_ratio} can be approximated as
\begin{equation}\label{eq:ttc_from_depth_discrete}
\tau = - Z(t_0) \bigg/ {\bigg(\frac{Z(t_1) - Z(t_0)}{t_1 - t_0}\bigg)} = \frac{t_1 - t_0}{1-\frac{Z(t_1)}{Z(t_0)}}.
\end{equation}
If we assume fronto-parallel and planar objects that do not rotate, we can combine Equations~\ref{eq:ttc_from_depth_discrete}~and~\ref{eq:scale_Z_inverse_relation}, thus expressing the TTC as a function of observations in image space:
\begin{equation}\label{eq:ttc_from_scale_discrete}
\tau = \frac{t_1 - t_0}{1-\frac{s(t_0)}{s(t_1)}} = \frac{t_1 - t_0}{1-\alpha},
\end{equation}
where $s$ is the size of the image of an object or region in the scene.
In other words, with Equation~\ref{eq:ttc_from_scale_discrete}, the TTC can be computed from the scale factor $\alpha$.
This simplifies our task, compared with having to estimate per-pixel FOEs and optical flow.
However, regressing $\alpha$ for each pixel explicitly, which requires defining the size of objects or image regions and tracking its change over time, is not straightforward.
Instead, we consider a pair of images, $I_0$ and $I_1$, and compute $I^{\alpha_i}_1$, a version of $I_1$ scaled by a factor $\alpha_i$.
This scaling factor $\alpha_i$ corresponds to a unique TTC $\tau_i$.
The regions whose size matches between $I_0$ and $I^{\alpha_i}_1$ will collide with the camera plane exactly at $\tau_i$.
Furthermore, we can expect regions that are larger (or smaller) in $I^{\alpha_i}_1$ to collide with the camera before (or after) $\tau_i$, see Figure~\ref{fig:intuition}.
Instead of regressing the scale factor $\alpha$ directly, then, we propose to train a neural network to take such pairs of images and classify regions in $I^{\alpha_i}_1$ as being larger or smaller than the corresponding regions in $I_0$, where the correspondence is learned implicitly.
Our approach is inspired by the work of Badki~\etal~\cite{badki2020Bi3D} for stereo.
They also learn to classify the disparity of a pixel as being larger or smaller than a given disparity, instead of regressing the disparity directly.
\begin{figure}
\vspace{-1mm}
\centering
\includegraphics[width=\columnwidth]{figures/arch/arch.pdf}
\vspace{-6mm}
\caption[]{\textbf{Architecture.} We perform all of our tasks with minor modifications to the same backbone. We preprocess the input images by extracting features and applying a task dependent operation, \cP, to the features of the second image. The input parameter {\color{blue}$\phi$}, the loss, {\color{blue}$\mathcal{L}$}, the ground truth, {\color{blue}GT}, and the output, {\color{blue}$\Psi$} for each task are listed in the table. Note that $\{\cdot\}$ indicates a set, and $\mathbb{1}$ the indicator function.}
\label{fig:arch}
\vspace{-4mm}
\end{figure}
\subsubsection{Binary TTC}\label{sec:ttc_binary}
The inputs to our method are two images and scale factor $\alpha_i$ corresponding to the TTC we want to analyze.
We first extract features $f_0$ and $f_1$ from both images, and scale $f_1$ by $\alpha_i$, to obtain $f_1^{\alpha_i}$.
Rather than classifying the objects as scaling up or down between $f_0$ and $f_1^{\alpha_i}$, we train a lightweight network to directly classify whether each pixel's TTC is larger or smaller than $\tau_i$, using a binary cross-entropy loss.
That is, the network predicts a probability map
\begin{equation}\label{eq:binary_map}
B_{\tau_i}(x,y) = p(\tau(x,y)>\tau_i; f_0, f_1^{\alpha_i}),
\end{equation}
which can be binarized by simple thresholding.
We train directly to predict binary TTC instead of explicitly detecting if the size of the image of objects is getting larger or smaller with respect to $\alpha_i$ for two reasons.
First, ground truth data for the scale factor $\alpha$ is challenging to even define beyond a small neighborhood of pixels, unless objects move rigidly and only translate.
Moreover, as discussed before, Equation~\ref{eq:ttc_from_scale_discrete} is an approximation in the common case of objects that are not planar, and for non-rectilinear motions.
Equation~\ref{eq:ttc_from_depth_discrete}, which establishes the relationship between the change in depth and the TTC, allows us to generate the ground truth data from existing datasets, as we explain in Section~\ref{sec:data_ttc_from_sceneflow}.
A network trained to classify the TTC directly can learn the necessary priors to compensate for the approximations introduced by Equation~\ref{eq:ttc_from_scale_discrete}.
Our core architecture is shown in Figure~\ref{fig:arch}.
This very architecture is also used for all the tasks we describe below, with the caveat that the terms in {\color{blue}blue} differ for each task.
\subsubsection{Quantized TTC}\label{sec:ttc_quantized}
Our core approach naturally extends to estimating a coarsely quantized TTC for each pixel, which may be more useful than binary in certain scenarios.
We note that Equation~\ref{eq:binary_map} is a complementary cumulative distribution function.
Therefore, for two time-to-contact values, $\tau_j > \tau_i$, we can compute
\begin{equation}\label{eq:p_quantized}
p(\tau_i < \tau(x,y) \le \tau_j) = B_{\tau_i}(x,y) - B_{\tau_j}(x,y).
\end{equation}
Consider a set of TTC values $\{\tau_i\}_{i=1:N}$, and assume they are in increasing order.
After computing Equation~\ref{eq:binary_map} for each of the $N$ TTC values, we can estimate the quantization bin in which the TTC of a pixel falls as
\begin{equation}\label{eq:quantized_map}
Q(x,y) = \argmax_i \bigg( p\big(\tau_i < \tau(x,y) \le \tau_{i+1}\big) \bigg).
\end{equation}
The different TTC values can be spaced non-uniformly.
\subsubsection{Continuous and Selective TTC}\label{sec:ttc_continuous}
While our method is specifically designed for binary and quantized TTC estimation, it can also estimate per-pixel, continuous TTC.
In principle, we can approximate continuous values by predicting quantized TCC (Section~\ref{sec:ttc_quantized}) with a larger set of TTC values $\{\tau_i\}_{i=1:N}$.
This, however, fails to exploit the relationship between the binary classifications for different $\tau_i$'s, for a pixel.
We slightly modify the approach while still preserving its binary classification core, as shown in Figure~\ref{fig:arch}.
Specifically, instead of taking consecutive pairs of probability maps and applying Equation~\ref{eq:p_quantized}, we stack them.
\begin{wrapfigure}[4]{r}[0pt]{0.38\columnwidth}
\hspace{-6mm}
\includegraphics[width=0.44\columnwidth]{figures/confidence/confidence.pdf}
\end{wrapfigure}
For a specific pixel $(x_0,y_0)$ we generally see a progression as in the plot on the right.
That is, $B_{\tau_i}(x_0,y_0)$ (the probability that the object will collide \emph{after} $\tau_i$) is consistently high for $\tau_i \ll \tau_*$, and low for $\tau_i \gg \tau_*$, where $\tau_*$ is the correct TTC value.
In the transition region, the network is uncertain, which is why aggregating information across multiple $\tau_i$ is beneficial.
Badki~\etal~\cite{badki2020Bi3D}, who obtain a similar curve for disparity, propose to estimate the transition point, $\tau_*$ in our case, by computing the area under the curve (AUC),
\begin{equation}\label{eq:AUC}
\text{AUC}(x,y) = \sum_i (\tau_{i+1} - \tau_i) \cdot B_{\tau_i}(x,y).
\end{equation}
To understand why Equation~\ref{eq:AUC} yields the desired result, consider the case of a step function, \ie, $\tau_*$ aligns with a quantization boundary: AUC$=\tau_*\cdot 1=\tau_*$.
This relationship holds for the more common case of a smooth transition region:
because the transition is generally symmetric around $\tau_*$, the red and green areas in the plot have similar extent and compensate for each other.
Because the AUC is differentiable, we can use it to fine-tune our network so that combining a set of probability values using the AUC operation yields continuous TTC values.
\subsubsection{A Note on Inference-Time Tradeoff}\label{sec:ttc_discussion}
Binary, quantized, and continuous TTC estimation yield increasingly rich information for navigating an environment.
We note that, when estimating quantized and continuous TTC, the multiple binary classifications can be run in parallel, as they are independent of each other.
Therefore we can compute quantized and continuous TTC at roughly the same frame-rate as for binary quantization---just above $150$ fps.
If multiple binary classifications cannot be run in parallel due to hardware limitations, the cost for computing quantized and continuous TTC grows linearly with the number of levels.
In such cases, one can decide the number of quantization levels dynamically to best leverage the trade-off between accuracy and latency.
For situations where a fast response time is critical, such as highway driving, binary TTC may be sufficient.
For slower navigation (\eg, for robotics or for parking lot driving), our method can be adapted to trade latency for a finer quantization \emph{at inference time}.
\subsection{Dealing with the Lack of Training Data}\label{sec:training_with_of}
As for any learning-based method, having access to a large amount of data is critical to properly train our network.
Unfortunately, there are no datasets with TTC ground truth data and few scene flow datasets that we can use to infer it (Section~\ref{sec:data_ttc_from_sceneflow}).
To increase the training data, we leverage a closely related task: binary optical flow estimation.
We use the same binary approach, but we shift the features (horizontally and vertically) rather than scaling them.
We then classify the direction of the shift (left/right or up/down).
For horizontal shifts, we seek to predict a probability map relative to a given shift $u_i$
\begin{equation}\label{eq:of_u_binary}
B_{u_i}(x,y) = p(u(x,y)>u_i; f_0, f_1^{u_i}),
\end{equation}
where $f_1^{u_i}$ are the features extracted from the second image and shifted by $u_i$.
The equation for the vertical shift $v_i$ is analogous.
Using optical flow to pre-train our network and continuing using it as an auxiliary task yields a stronger inductive bias, as we discuss in Section~\ref{sec:results}.
Although optical flow is an auxiliary task, we show results in Figure~\ref{fig:optical_flow} to allow for a visual evaluation.
\begin{figure}
\vspace{-3mm}
\centering
\input{figures/of/of_figure.tex}
\vspace{-2mm}
\caption{\textbf{Optical flow as an auxiliary task.} We estimate optical flow as an auxiliary task to improve our TTC estimation. Like for continuous TTC, we estimate optical flow via a series of binary classifications. While optical flow is not a goal for us, this figure shows that our prediction is reasonable.}\label{fig:optical_flow}
\vspace{-4mm}
\end{figure}
\subsection{Time-to-Contact}
Time-to-contact (TTC) was studied in psychophysics and psychology, even before it attracted the attention of the computer vision community.
Early work by Lee, for instance, suggested that TTC is sufficient for making decisions about braking, and is likely to be picked up by the driver faster than distance, speed, or acceleration of objects in the scene~\cite{lee1976ttcForBraking}.
From a computational standpoint, TTC is appealing because it only depends on the ratio of depth and velocity, which can be estimated directly from images, even when estimating either one is an ill-posed problem~\cite{horn2009hierarchical}.
Several traditional approaches have been proposed to estimate TTC from the estimated optical flow~\cite{meyer1992estimation, meyer1994time, camus1995calculating}.
Horn~\etal proposed a direct method for TTC estimation that only uses the constant brightness assumptions~\cite{horn2007ttc_planar,horn2009hierarchical}.
While these approaches estimate the TTC for an object that is moving relative to the camera, they require masks for dynamic objects, which limits their practical impact.
We propose a learning-based approach for TTC estimation that handles multiple dynamic objects---without needing any segmentation---and estimates per-pixel TTC.
The elegant, closely related work by Yang and Ramanan estimates optical flow, uses it to compute the scaling factor of objects, and maps it to TTC~\cite{yang2020of2ttc}.
Xu \etal also estimate the scaling of objects, but do so by modeling the change of objects' size explicitly for optical flow estimation~\cite{xu2012scaleinvof}.
However, computing the full optical flow is time-consuming, and estimating TTC from optical flow inherits its limitations.
Instead, following Horn~\etal~\cite{horn2009hierarchical}, we compute TTC directly from the input images, side-stepping optical flow computation altogether.
Moreover, inspired by Badki~\etal~\cite{badki2020Bi3D}, we solve TTC estimation via a series of binary classifications.
Each binary classification can be computed independently of the others at over $150$ fps.
This can be thought of as a temporal geofence, detecting pixels or objects within a given TTC.
For existing methods, including the method by Yang and Ramanan~\cite{yang2020of2ttc}, this can only be achieved by computing the TTC for all the pixels or objects in the scene, and then thresholding it.
\begin{figure*}
\vspace{-5mm}
\centering
\subfloat[Inputs] {\includegraphics[height=80pt, trim={0.04in 0.05in 2.85in 0}, clip]{figures/intuition/intuition_new_with_bg.pdf}}
\subfloat[Scaled inputs]{\includegraphics[height=80pt, trim={1.9in 0.05in 0.9in 0}, clip]{figures/intuition/intuition_new_with_bg.pdf}}
\subfloat[Temporal geofence]{\includegraphics[height=80pt, trim={3.84in 0.05in 0.05in 0}, clip]{figures/intuition/intuition_new_with_bg.pdf}}
\vspace{-2mm}
\caption{\textbf{Intuition.}
Given two images of a dynamic scene, $I_0$ and $I_1$, we define a temporal geofence to detect objects expected to cross the camera plane before a given time $\tau_i$ from the time of capture of $I_0$.
Compare $I_0$ and $I_1$, (a).
The images of the orange and blue cars are smaller in $I_0$ ($\alpha<1$), while the image of the purple car is roughly unchanged.
This allows us to predict that only the first two cars will collide with the camera plane.
We propose to perform this comparison after scaling the source image by a factor $\alpha_i$ (corresponding to TTC value $\tau_i$).
The orange car is still larger in $I_1^{\alpha_i}$ ($\alpha<\alpha_i$), indicating that it will cross the camera plane \emph{before} the specified $\tau_i$.
On the other hand, the blue and purple cars will not.
Rather than regressing the exact scale factor, we classify a pixel as making contact before or after $\tau_i$ by classifying if the objects scale up or down in $I_1^{\alpha_i}$ with respect to $I_0$.
This yields a binary TTC probability map for $\tau_i$, (c).}
\label{fig:intuition}
\vspace{-3mm}
\end{figure*}
\subsection{3D Inference as a Classification Problem}
The idea of posing 3D regression as a classification task has a rich history.
Several learning-based approaches estimate depth via a multi-class classification task~\cite{kendall2017end, chang2018pyramid, zhang2019ga, im2019dpsnet}.
These are more accurate than other learning-based approaches that pose the problem as a regression task~\cite{Menze2015kittidataset,dosovitskiy2015flownet}.
Badki~\etal introduced a method that allows us to control the trade-off between latency and accuracy~\cite{badki2020Bi3D}.
Instead of posing depth as a multi-class classification problem, they solve it via multiple binary classifications.
Each classification provides useful information about the scene at high frame-rates.
Our approach is based on the same intuition.
We are the first learning-based approach to estimate per-pixel TTC directly from the input images, and to pose it as a (binary) classification problem.
\section{Additional Details on Architecture}\label{sec:arch}
\subsection{Binary Estimation}\label{sec:arch_bin}
Our binary segmentation architecture has three main components: a feature extraction network (FeatExtractNet), a segmentation network (SegNet2D), and a refinement network (SegRefineNet).
Our architecture blocks are adopted from Bi3DNet~\cite{badki2020Bi3D}.
We do not use any batch normalization layers in our network.
\paragraph{FeatExtractNet.}
We use the same feature extraction network as used in Bi3DNet~\cite{badki2020Bi3D}.
This feature extraction module is based on the simplified version of the feature extraction network from PSMNet~\cite{chang2018pyramid}.
We normalize each color channel of the input images using the mean and standard deviation of $0.5$ before passing to this network.
The output is a $32$-channel feature map at one third of the input image resolution.
\paragraph{SegNet2D.}
We warp the source image features using a task-dependent operator.
To avoid cropping out the image features during this warping operation, we zero pad the feature maps to $1.5\times$ the feature map resolution.
The reference image feature map and the warped source image feature map are concatenated and fed to the SegNet2D network.
SegNet2D is a 2D encoder-decoder network with skip-connections.
The encoder is comprised of five blocks, each of which has a conv layer that downsamples the features with a stride of 2 followed by another conv layer with a stride of 1.
We use $3\times3$ kernels.
The feature sizes for each of these five blocks are $128$, $256$, $512$, $1024$, and $1024$.
The decoder is comprised of five blocks, each of which has of a deconv layer with $4\times4$ kernels and a stride of 2, followed by a conv layer with $3\times3$ kernels and a stride of 1.
The feature sizes for each of these five blocks are $1024$, $512$, $256$, $128$, and $64$.
We use the LeakyReLU activation function in the network with a slope of $0.1$.
A final conv layer with $3\times3$ kernels, without any activation, is used to generate 3 outputs: one for binary TTC, one for binary horizontal optical flow, and one for binary vertical optical flow.
The final segmentation probability maps can be obtained by cropping out the excess padding, upsampling the outputs to the input image resolution and applying a sigmoid function.
\paragraph{SegRefineNet.}
SegRefineNet is used to refine the segmentation outputs from SegNet2D network using the reference image as a guide.
First, we generate a $16$ channel feature map for the reference image by applying 3 conv layers with $3\times3$ kernels and a stride of 1.
The first two layers use ReLU activation and the final layer does not use any activation.
This feature extraction is done only once and can be used for refining multiple segmentation outputs from the SegNet2D network.
An upsampled segmentation output is concatenated with the reference image features and refined by applying 4 conv layers.
Each conv layer uses $3\times3$ kernels, a feature-size of $8$, and LeakyReLU activation with a slope of 0.1.
The final output of this network is generated by a final conv layer with $3\times3$ kernel and without any activation.
The final binary segmentation probability map can then be generated by applying a sigmoid function.
\subsection{Continuous Estimation}\label{sec:arch_cont}
Given input images we generate the corresponding feature maps using the same FeatExtractNet as in Section~\ref{sec:arch_bin}.
To generate continuous estimation results we uniformly sample the warping parameters in the desired range and use these parameters to warp the features of the source image.
These warped source image features are concatenated with the reference image features to form an input volume for the SegNet2D network.
SegNet2D generates an output volume which upon upsampling to the input image resolution, applying the sigmoid operator followed by AUC operator gives us the continuous estimation map.
This continuous map is further refined using the input image as a guide using a refinement network, ContRefineNet.
This network is based on the disparity refinement network proposed in StereoNet~\cite{khamis2018stereonet}.
\section{Additional Details on Training}\label{sec:train}
\subsection{Binary Estimation}
We pre-train our network for the binary optical flow task on the FlyingChairs2~\cite{dosovitskiy2015flownet, iig2018generic} dataset for 300 epochs and train it further on FlyingThings3D dataset~\cite{mayer2016large} for 400 epochs using a binary cross-entropy (BCE) loss with respect to a thresholded version of the ground truth.
Each batch for training is formed by randomly sampling 16 image pairs from the dataset and then randomly sampling shifts for the two components of the optical flow vector for each image.
Note that our SegNet2D network has three output heads and we leave the head corresponding to binary TTC untouched during this training.
We then fine-tune our network for estimating binary TTC first on the Driving~\cite{mayer2016large} for 500 epochs and then on the KITTI15~\cite{Menze2015kittidataset} datasets for 10k epochs.
Since both datasets also offer optical flow data, in this second stage we train for both binary optical flow and binary TCC.
To do this we form a batch by randomly sampling 16 image pairs from the dataset.
Then for each image pair we form two sets of warped image features: one by randomly sampling shifts for the two components of the optical flow vector and other by randomly sampling scales for training binary TTC.
We select the appropriate segmentation output corresponding to the task and use BCE loss to train our network simultaneously for the binary optical flow and TTC estimation task.
Throughout our training we randomly sample shifts in the range $[-99,99]$, and scales in the range of $[0.5,1.3]$.
We use relative weights of $0.8$ and $0.2$ for binary TTC and binary optical flow task respectively.
For training on the KITTI15 dataset, we split our dataset into training ($160$ examples) and validation ($40$ examples) sets, following the split provided by Yang and Ramanan~\cite{yang2020of2ttc}.
\subsection{Continuous Estimation}
To train our network for the continuous estimation task, we start with the network trained on the FlyingChairs2 and FlyingThings3D datasets for the binary optical flow task and fine-tune it on FlyingThings3D for continuous optical flow for 100 epochs.
Each batch for training is formed by randomly sampling 8 image pairs from the dataset.
For each training image pair we uniformly sample horizontal and vertical shifts and stack the maps corresponding to each shift.
We use the resulting volumes to compute continuous optical flow via the AUC operation which we refine using ContRefineNet.
To fine-tune our network for continuous TTC estimation, we follow a similar strategy as for the binary segmentation training.
We continue training the network for the task of continuous optical flow and continuous TTC estimation on Driving dataset for 100 epochs, followed by KITTI15 datasets for 600 epochs.
We form three volumes, two corresponding for optical flow and one for TTC.
Throughout the training, we use BCE loss on the estimated binary segmentation probability maps, and a SmoothL1 (SL1) regression loss on the output of the AUC module.
We use relative weights of $0.1$ and $0.9$ respectively.
While training on both TTC and OF estimation tasks, we use relative weights of $0.8$ and $0.2$ respectively.
For optical flow we randomly sample a contiguous block of 16 shifts that are divisible by 3 and in the range $[-99, 99]$ for both the components of the optical flow.
We select the shifts to be a factor of 3 since we perform the shifting on the feature maps that are one third the input image resolution.
The AUC operation seamlessly handles the objects with shifts that lie beyond sampled range.
For TTC estimation, we work in the inverse TTC domain and uniformly sample 24 scales in the range $[0.5, 1.3]$.
We perform the training on the cropped images of size $384\times576$.
The cropping is done after the feature warping step.
Again for the KITTI15 dataset, we split the dataset into train and validation sets and use the validation set to compare our method with competing approaches.
The network trained on the entire KITTI15 dataset is used for scene flow estimation on KITTI15 benchmark images as explained in the main paper.
\end{document} |
\section{Four classical number-theoretic results}
By $\Pi$ we denote the set of prime numbers. For a number $x\in\IZ$ by $\Pi_x$ we denote the set of all prime divisors of $x$. Two numbers $x,y\in\IZ$ are {\em coprime} iff $\Pi_x\cap\Pi_y=\emptyset$
In the proof of Theorem~\ref{t:main} we shall exploit the following four known results of Number Theory. The first one is the famous Chinese Remainder Theorem (see. e.g. \cite[3.12]{J}).
\begin{theorem}[Chinese Remainder Theorem]\label{Chinese} If
$b_1,\dots,b_n\in\IZ$ are pairwise coprime numbers, then for any numbers $a_1,\dots,a_n\in\IZ$, the intersection $\bigcap_{i=1}^n(a_i+b_i\IN)$ is infinite.\end{theorem}
The second classical result is not elementary and is due to Dirichlet \cite[S.VI]{Dirichlet}, see also \cite[Ch.7]{Ap}.
\begin{theorem}[Dirichlet]\label{Dirichlet} For any coprime numbers $a,b\in\IN$ the arithmetic progression $a+b\IN$ contains a prime number.
\end{theorem}
The third classical result is a recent theorem of Miha\u\i lescu \cite{Mih} who solved old Catalan's Conjecture \cite{Met}.
\begin{theorem}[Mih\u ailescu]\label{Mihailescu} If $a,b\in \{m^{n+1}:n,m\in \IN\}$, then $|a-b|=1$ if and only if $\{a,b\}=\{2^3,3^2\}$.
\end{theorem}
The fourth classical result we use is due to Karl Zsigmondy \cite{Zsigmondy}, see also \cite[Theorem 3]{Roitman}.
\begin{theorem}[Zsigmondy]\label{Zsigmondy} For integer numbers $a,n\in\IN\setminus\{1\}$ the inclusion $\Pi_{a^n-1}\subseteq\bigcup\limits_{0<k<n}\Pi_{a^k-1}$ holds if and only if one of the following conditions is satisfied:
\begin{enumerate}
\item $n=2$ and $a=2^k-1$ for some $k\in\IN$; then $a^2-1=(a+1)(a-1)=2^k(a-1)$;
\item $n=6$ and $a=2$; then $a^n-1=2^6-1=63=3^2\times 7=(a^2-1)^2\times(a^3-1)$.
\end{enumerate}
\end{theorem}
\section{Superconnected spaces and their superconnecting posets}\label{s:poset}
In this section we discuss superconnected topological spaces and some order structures related to such spaces.
First let us introduce some notation and recall some notions.
For a set $A$ and $n\in\w$ let $[A]^n=\{E\subseteq A:|A|=n\}$ be the family of $n$-element subsets of $A$, and $[A]^{<\w}=\bigcup_{n\in\w}[A]^n$ be the family of all finite subsets of $A$. For a function $f:X\to Y$ and a subset $A\subseteq X$ by $f[A]$ we denote the image $\{f(a):a\in A\}$ of the set $A$ under the function $f$.
For a subset $A$ of a topological space $(X,\tau)$ by $\overline{A}$ we denote the closure of $A$ in $X$. For a point $x\in X$ we denote by $\tau_x:=\{U\in\tau:x\in U\}$ the family of all open neighborhoods of $x$ in $(X,\tau)$.
A {\em poset} is an abbreviation for a partially ordered set.
A family $\F$ of subsets of a set $X$ is called a {\em filter} if
\begin{itemize}
\item $\emptyset\notin\F$;
\item for any $A,B\in\F$ their intersection $A\cap B\in\F$;
\item for any sets $F\subseteq E\subseteq X$ the inclusion $F\in\F$ implies $E\in\F$.
\end{itemize}
A topological space $(X,\tau)$ is called {\em superconnected} if for any $n\in\IN$ and non-empty open sets $U_1,\dots,U_n$ the intersection $\overline{U_1}\cap\dots\cap\overline{U_n}$ is non-empty. This allows us to define the filter
$$\F_\infty=\{B\subseteq X\colon \exists U_1,\dots,U_n\in\tau\setminus\{\emptyset\}\;\;(\overline{U_1}\cap\dots\cap\overline{U_n}\subseteq B)\},$$
called the {\em superconnecting filter} of $X$.
For every finite subset $E$ of $X$ consider the subfilter
$$\F_E:=\{B\subseteq X:\textstyle{\exists (U_x)_{x\in E}\in\prod_{x\in E}\tau_x\;\;(\;\bigcap_{x\in E}\overline{U_x}\subseteq B)}\}$$
of $\F_\infty$. Here we assume that $\F_\emptyset=\{X\}$.
It is clear that for any finite sets $E\subseteq F$ in $X$ we have $\F_E\subseteq \F_F$.
The family $$\mathfrak F=\{\F_E:E\in[X]^{<\w}\}\cup\{\F_\infty\}$$ is endowed with the inclusion partial order and is called the {\em superconnecting poset} of the superconnected space $X$. The filters $\F_\emptyset$ and $\F_\infty$ are the smallest and largest elements of the poset $\mathfrak F$, respectively.
The following obvious lemma shows that the superconnecting poset $\mathfrak F$ is a topological invariant of the superconnected space.
\begin{proposition}\label{p:iso} For any homeomorphism $h$ of any superconnected topological space $X$, the map $$\tilde h:\mathfrak F\to\mathfrak F,\quad \tilde h\colon\F\mapsto \{h[A]\colon A\in\F\},$$ is an order isomorphism of the superconnecting poset $\mathfrak F$.
\end{proposition}
In the following sections we shall study the order properties of the poset $\mathfrak F$ for the Kirch space $(\IZ^\bullet,\tau_K)$ and shall exploit the obtained information in the proof of the topological rigidity of the Kirch space.
\section{Proof of Theorem~\ref{t:main}}
We divide the proof of Theorem~\ref{t:main} into 23 lemmas
\begin{lemma}\label{basic} For any $a,b\in\IZ^\bullet$ the closure $\overline{a+b\IZ}$ of the arithmetic progression $a+b\IZ$ in the Kirch space $(\IZ^\bullet,\tau_K)$ is equal to
$$\IZ^\bullet\cap\bigcap_{p\in\Pi_b}\big(\{0,a\}+p\IZ\big).$$
\end{lemma}
\begin{proof}
First we prove that $\overline{a+b\mathbb Z}\subseteq \{0,a\}+p\IZ$ for every $p\in\Pi_b$.
Take any point $x\in\overline{a+b\IZ}$ and assume that $x\notin p\IZ$. Then $x+p\IZ$ is a neighborhood of $x$ and hence the intersection $(x+p\IZ)\cap(a+b\IZ)$ is not empty. Then there exist $u,v\in\IZ$ such that $x+pu=a+bv$. Consequently, $x-a=bv-pu\in p\IZ$ and $x\in a+p\IZ$.
Next, take any point $x\in\IZ^\bullet\cap\bigcap_{p\in\Pi_b}(\{0,a\}+p\IZ)$. Given any neighborhood $O_x$ of $x$ in $(\IZ^\bullet,\tau_K)$, we should prove that $O_x\cap(a+b\IZ)\ne\emptyset$. By the definition of the Kirch topology there exists a square-free number $d\in\IZ^\bullet$ such that $d,x$ are coprime and $x+d\IZ\subseteq O_x$.
If $\Pi_b\subseteq \Pi_x$, then $b,d$ are coprime and by Chinese Remainder Theorem $\emptyset \ne (x+d\IZ)\cap (a+b\IZ)\subseteq O_x\cap(a+b\IZ)$.
So, we can assume $\Pi_b\setminus\Pi_x\ne\emptyset$.
The choice of $x\in \bigcap_{p\in\Pi_b}(\{0,a\}+p\IZ)$ guarantees that $x\in \bigcap_{p\in\Pi_b\setminus \Pi_x}(a+p\IZ)=a+q\IZ$ where $q=\prod_{p\in \Pi_b\setminus\Pi_x}p$. Since the numbers $x$ and $d$ are coprime and $d$ is square-free, the greatest common divisor of $b$ and $d$ divides the number $q$. Since $x-a\in q\IZ$, the Euclides algorithm yields two numbers $u,v\in\IZ$ such that $x-a=bu-dv$, which implies that $O_x\cap (a+b\IZ)\supseteq (x+d\IZ)\cap(a+b\IZ)\ne\emptyset$.
\end{proof}
Lemma~\ref{basic} implies that the Kirch space $(\IZ^\bullet,\tau_K)$ is superconnected and hence possesses the superconnecting filter
$$\F_\infty=\big\{F\subseteq\IZ^\bullet:\exists U_1,\dots,U_n\in\tau_K\setminus\{\emptyset\}\quad\textstyle\big(\bigcap\limits_{i=1}^n\overline{U_i}\subseteq F\big)\big\}$$and the superconnecting poset
$$\mathfrak F=\{\F_E:E\in[\IZ^\bullet]^{<\w}\}\cup\{\F_\infty\}$$
consisting of the filters $$\F_E=\big\{F\subseteq\IZ^\bullet:\textstyle{\exists (U_x)_{x\in E}\in\prod_{x\in E}\tau_x\;\;\big(\bigcap_{x\in E}\overline{U_x}\subseteq F}\big)\big\}.$$ Here for a point $x\in\IZ^\bullet$ by $\tau_x:=\{U\subseteq\IZ^\bullet:x\in U\}$ we denote the family of open neighborhoods of $x$ in the Kirch topology $\tau_K$.
For a nonempty finite subset $E\subseteq\IZ^\bullet$, let $\Pi_E=\bigcap_{x\in E}\Pi_x$ be the set of common prime divisors of numbers in the set $E$. Also let
$$A_E=\{p\in\Pi:\exists k\in \IN\;\;(E\subset \{0,k\}+p\IZ)\}.$$
Observe that $\Pi_E\subseteq A_E$ and $A_E\ne\emptyset$ because $2\in A_E$.
If $E$ is a singleton, then $A_E=\Pi$;
if $|E|\ge 2$, then $A_E\subseteq \Pi_x\cup\Pi_y\cup\Pi_{x-y}\subseteq\{2,\dots,\max E\}$ for any distinct numbers $x,y\in E$. This inclusion follows from
\begin{lemma}\label{2ae} For any two-element set $E=\{x,y\}\subset \IZ^\bullet$ we have $A_E= \Pi_x\cup\Pi_y\cup\Pi_{x-y}$.
\end{lemma}
\begin{proof}
Each number $p\in\Pi_x$ (resp. $p\in\Pi_y$) belongs to $A_E$ because $\{x,y\}\subset\{0,y\}+p\IZ$ (resp. $\{x,y\}\subset\{0,x\}+p\IZ\}$). Each number $p\in\Pi_{x-y}$ belongs to $A_E$ because $\{x,y\}\subset x+p\IZ\subseteq\{0,x\}+p\IZ$. This proves that $\Pi_x\cup\Pi_y\cup\Pi_{x-y}\subseteq A_E$.
Now take any prime number $p\in A_E$ and assume that $p\notin \Pi_x\cup\Pi_y$. It follows from $\{x,y\}=E\subset\{0,\alpha_E(p)\}+p\IZ$ that $\{x,y\}\subseteq \alpha_E(p)+p\IZ$ and hence $x-y\in p\IZ$ and $p\in\Pi_{x-y}$.
\end{proof}
Let $\alpha_E\colon A_E\to\w$ be the unique function satisfying the following conditions:
\begin{itemize}
\item[\textup{(i)}] $\alpha_E(p)<p$ for all $p\in A_E$;
\item[\textup{(ii)}] $E\subseteq\{0,\alpha_E(p)\}+p\IZ$ for all $p\in A_E$;
\item[\textup{(iii)}] $\alpha_E(2)=1$ and $\alpha_E(p)=0$ for all $p\in \Pi_E\setminus\{2\}$.
\end{itemize}
\begin{lemma}\label{l:realization} Let $A\subset\Pi$ be a finite set containing $2$ and $\alpha:A\to\IN_0$ be a function such that $\alpha(2)=1$ and $\alpha(p)\in\{0,\dots,p-1\}$ for all $p\in A\setminus\{2\}$. Let $x$ be the product of odd prime numbers in the set $A$ and $y$ be any number in the set $\IZ^\bullet\cap\bigcap_{p\in A}(\alpha(p)+p\IZ)$. Then the set $E=\{y,x,2x\}$ has $A_E=A$ and $\alpha_E=\alpha$.
\end{lemma}
\begin{proof} For every prime number $p\in A$ we have $\{x,y\}\subset\{0,y\}+p\IZ$, which implies that $p\in A_E$. Assuming that $A_E\setminus A$ contains some prime number $p$, we conclude that $x\notin p\IZ$ and hence the inclusion $\{y,x,2x\}=E\subset \{0,\alpha_E(p)\}+p\IZ$ implies $\{x,2x\}\subset\alpha_E(p)+p\IZ$ and $x=2x-x\in p\IZ$. This contradiction shows that $A_E=A$. To show that $\alpha_E=\alpha$, take any prime number $p\in A=A_E$.
If $p=2$, then $\alpha(p)=1=\alpha_E(p)$. So, we assume that $p\ne 2$. If $\alpha(p)=0$, then $y\in \alpha(p)+p\IZ=p\IZ$ and hence $p\in \Pi_E$. In this case $\alpha_E(p)=0=\alpha(p)$. If $\alpha(p)\ne 0$, then the number $y\in\alpha(p)+p\IZ$ is not divisible by $p$ and then the inclusions $\{y,x,2x\}\subseteq\{0,\alpha(p)\}+p\IZ$ and $\{y,x,2x\}=E\subset\{0,\alpha_E(p)\}+p\IZ$ imply that $\alpha(p)=\alpha_E(p)$.
\end{proof}
The following lemma yields an arithmetic description of the filters $\F_E$.
\begin{lemma}\label{l:Kirch} For any finite subset $E\subseteq\IZ^\bullet$ with $|E|\ge 2$ we have
$$\F_E=\big\{B\subseteq\IZ^\bullet\colon \exists L\in[\Pi\setminus A_E]^{<\w}\quad\bigcap_{p\in L}p\IZ^\bullet\cap\bigcap_{p\in A_E\setminus\Pi_E}(\{0,\alpha_E(p)\}+p\IZ)\subseteq B\big\}.$$Here we assume that $\bigcap_{p\in\emptyset}p\IZ^\bullet=\IZ^\bullet$.
\end{lemma}
\begin{proof} It suffices to verify two properties:
\begin{enumerate}
\item for any $(U_x)_{x\in E}\in\prod_{x\in E}\tau_x$ there exists a finite set $L\subseteq \Pi\setminus A_E$ such that
$$\bigcap_{p\in L}p\IZ^\bullet\cap\bigcap_{p\in A_E\setminus\Pi_E}(\{0,\alpha_E(p)\}+p\IZ)\subseteq \bigcap_{x\in E}\overline{U_x};$$
\item for any finite set $L\subseteq\Pi\setminus A_E$ there exists a sequence of neighborhoods $(U_x)_{x\in E}\in\prod_{x\in E}\tau_x$ such that
$$\bigcap_{x\in E}\overline{U_x}\subseteq \bigcap_{p\in L}p\IZ^\bullet\cap\bigcap_{p\in A_E\setminus\Pi_E}(\{0,\alpha_E(p)\}+p\IZ).$$
\end{enumerate}
1. Given a sequence of neighborhoods $(U_x)_{x\in E}\in\prod_{x\in E}\tau_x$, for every $x\in E$ find a square-free number $q_x>x$ such that $\Pi_{q_x}\cap \Pi_x=\emptyset$ and $x+q_x\IZ\subseteq U_x$. We claim that the finite set $L=\bigcup_{x\in E}\Pi_{q_x}\setminus A_E$ has the required property. Given any number $z\in \bigcap_{p\in L}p\IZ^\bullet\cap\bigcap_{p\in A_E\setminus\Pi_E}(\{0,\alpha_E(p)\}+p\IZ)$, we should prove that $z\in\overline{U_x}$ for every $x\in E$. By Lemma~\ref{basic},
$$\IZ^\bullet\cap \bigcap_{p\in\Pi_{q_x}}(\{0,x\}+p\IZ)=\overline{ (x+q_x\IZ)}\subseteq\overline{U_x}.$$ So, it suffices to show that $z\in \{0,x\}+p\IZ$ for any $p\in \Pi_{q_x}$. Since the numbers $x$ and $q_x$ are coprime, $p\notin\Pi_x$ and hence $p\notin\Pi_E$. If $p\notin A_E$, then $p\in \Pi_{q_x}\setminus A_E\subseteq L$ and hence $z\in p\IN\subseteq \{0,x\}+p\IZ$. If $p\in A_E$, then $x\in E\subseteq\{0,\alpha_E(p)\}+p\IZ$ and $x\in\alpha_E(p)+p\IZ$ (as $p\notin\Pi_x$). Then $x+p\IZ=\alpha_E(p)+p\IZ$ and
$z\in \{0,\alpha_E(p)\}+p\IZ=\{0,x\}+p\IZ.$
2. Fix any finite set $L\subseteq \Pi\setminus A_E$. For every $x\in E$ consider the neighborhood $U_x=\bigcap_{p\in L\cup A_E\setminus\Pi_x}(x+p\IZ)$ of $x$ in the Kirch topology. By Lemma~\ref{basic},
$$\overline{U_x}=\IZ^\bullet\cap\bigcap_{p\in L\cup A_E\setminus \Pi_x}(\{0,x\}+p\IZ).$$
Given any number $z\in\bigcap_{x\in E} \overline{U_x}$, we should show that $z\in \bigcap_{p\in L}p\IZ^\bullet\cap\bigcap_{p\in A_E\setminus \Pi_E}(\{0,\alpha_E(p)\}+p\IZ)$.
This will follow as soon as we check that $z\in p\IZ^\bullet$ for all $p\in L$ and $z\in\{0,\alpha_E(p)\}+p\IZ$ for all $p\in A_E\setminus \Pi_E$.
Given any $p\in A_E\setminus\Pi_E$, we can find a point $x\in E\setminus p\IZ$ and observe that $x\in E\subseteq\{0,\alpha_E(p)\}+p\IZ$. Then $z\in \overline{U_x}\subseteq\overline{x+p\IZ}\subseteq \{0,x\}+p\IZ=\{0,\alpha_E(p)\}+p\IZ$.
Now take any prime number $p\in L$. Since $L\cap A_E=\emptyset$, we conclude that $E\not\subseteq p\IZ$. So, we can fix a number $x\in E\setminus p\IZ$. Taking into account that $p\notin A_E$, we conclude that $E\not\subseteq \{0,x\}+p\IZ$ and hence there exists a number $y\in E$ such that $p\IZ\ne y+p\IZ\ne x+p\IZ$. Then
$$z\in\overline{U_x}\cap\overline{U_y}\subseteq(\{0,x\}+p\IZ)\cap(\{0,y\}+p\IZ)=p\IZ.$$
\end{proof}
We shall use Lemma~\ref{l:Kirch} for an arithmetic characterization of the partial order of the superconnecting poset $\mathfrak F$ of the Kirch space.
\begin{lemma}\label{l:wo2} For two finite subsets $E,F\subseteq \Pi$ with $\min\{|E|,|F|\}\ge2$ we have $\F_E\subseteq \F_F$ if and only if $$A_F\subseteq A_E,\;\; \Pi_F\setminus\{2\}\subseteq \Pi_E \mbox{ \ and \ }\alpha_E{\restriction}A_F\setminus\Pi_E=\alpha_F{\restriction}A_F\setminus\Pi_E.$$
\end{lemma}
\begin{proof} To prove the ``only if'' part, assume that $\F_E\subseteq\F_F$. By Lemma~\ref{l:Kirch}, the set $$\bigcap_{p\in A_F\setminus A_E}p\IZ^\bullet\cap\bigcap_{p\in A_E\setminus\Pi_E}(\{0,\alpha_E(p)\}+p\IZ)$$ belongs to the filter $\F_E\subseteq\F_F$.
By Lemma~\ref{l:Kirch}, there exists a finite set $L\subset\Pi\setminus A_F$ such that
\begin{equation}\label{eq1}
\bigcap_{p\in L}p\IZ^\bullet\cap\bigcap_{p\in A_F\setminus\Pi_F}(\{0,\alpha_F(p)\}+p\IZ)\subseteq \bigcap_{p\in A_F\setminus A_E}p\IZ^\bullet\cap \bigcap_{p\in A_E\setminus\Pi_E}(\{0,\alpha_E(p)\}+p\IZ).
\end{equation}
This inclusion combined with the Chinese Remainder Theorem~\ref{Chinese} implies $$A_F\setminus A_E\subseteq L\subset \Pi\setminus A_F,\;\;A_E\setminus (\Pi_E\cup\{2\})\subseteq L\cup (A_F\setminus\Pi_F)\mbox{ \ and $\alpha_E(p)=\alpha_F(p)$ for any $p\in (A_F\setminus\Pi_F)\cap (A_E\setminus\Pi_E)$,}$$
and
\begin{equation}\label{eq2}
A_F\subseteq A_E,\;\;\Pi_F\setminus\{2\}\subseteq\Pi_E\mbox{ \ and \ }\alpha_E{\restriction}A_F\setminus\Pi_E=\alpha_F{\restriction}A_F\setminus\Pi_E.
\end{equation}
To prove the ``if'' part, assume that the condition (\ref{eq2}) holds. To prove that $\F_E\subseteq\F_F$, fix any set $\Omega\in\F_E$ and using Lemma~\ref{l:Kirch}, find a finite set $L\subseteq \Pi\setminus A_E$ such that
$$\bigcap_{p\in L}p\IZ^\bullet\cap\bigcap_{p\in A_E\setminus\Pi_E}(\{0,\alpha_E(p)\}+p\IZ)\subseteq\Omega.$$
Consider the finite set $\Lambda=(L\cup A_E)\setminus A_F=L\cup (A_E\setminus A_F)\supseteq L$ and observe that the condition (\ref{eq2}) implies the inclusion
\begin{equation}\label{eq3}\F_F\ni
\bigcap_{p\in\Lambda}p\IZ^\bullet\cap\bigcap_{p\in A_F\setminus\Pi_F}(\{0,\alpha_F(p)\}+p\IZ)\subseteq \bigcap_{p\in L}p\IZ^\bullet\cap\bigcap_{p\in A_E\setminus \Pi_E}(\{0,\alpha_E(p)\}+p\IZ)\subseteq\Omega,
\end{equation}
yielding $\Omega\in\F_F$.
\end{proof}
\begin{lemma}\label{l:wo1} For two nonempty subsets $E,F\subseteq \IN$ with $\min\{|E|,|F|\}=1$ the relation $\F_E\subseteq\F_F$ holds if and only if $|E|=1$ and $E\subseteq F$.
\end{lemma}
\begin{proof} The ``if'' part is trivial. To prove the ``only if'' part, assume that $\F_E\subseteq\F_F$. First we prove that $|E|=1$. Assuming that $|E|>1$ and taking into account that $\min\{|E|,|F|\}=1$, we conclude that $|F|=1$. Choose a prime number $p>\max(E\cup F)$. Since $\bigcap_{y\in E}\overline{y+p\IZ}\in \F_E\subseteq\F_F$, for the unique number $x$ in the set $F$, there exists a square-free number $d$ such that $\Pi_d\cap\Pi_x=\emptyset$ and $\overline{x+dp\IZ}\subseteq\bigcap_{y\in E}\overline{y+p\IZ}$. By Lemma~\ref{basic},
$$x+qp\IZ^\bullet\subseteq \overline{x+dp\IZ}\subseteq\bigcap_{y\in E}\overline{y+p\IZ}=\bigcap_{y\in E}(\{0,y\}+p\IZ)=p\IZ.$$
The latter equality follows from $p>\max E$ and $|E|>1$. Then $x+dp\IZ^\bullet\subseteq p\IZ$ implies $x\in p\IZ$, which contradicts the choice of $p>\max (E\cup F)\ge x$. This contradiction shows that $|E|=1$. Let $z$ be the unique element of the set $E$.
It remains to prove that $z\in F$. To derive a contradiction, assume that $z\notin F$. Take any odd prime number $p>\max(E\cup F)$ and consider the set $\{0,z\}+p\IZ=\overline{z+p\IZ}\in\F_E\subseteq\F_F$. By the definition of the filter $\F_F$, for every $x\in F$ there exists a square-free number $d_x$ such that $\Pi_{d_x}\cap\Pi_x=\emptyset$ and $$\bigcap_{x\in F}\overline{x+d_x\IZ}\subseteq \overline{z+p\IZ}=\{0,z\}+p\IZ.$$ Consider the set $P=\bigcup_{x\in F}\Pi_{d_x}$.
If $p\in \Pi_{d_x}$ for some $x\in F$, we can use the Chinese Remainder Theorem~\ref{Chinese} and find a number $$c\in (x+p\IZ)\cap \bigcap_{q\in P\setminus\{p\}}q\IZ\subseteq \bigcap_{y\in F}\overline{y+d_y\IZ}\subseteq \{0,z\}+p\IZ.$$
Taking into account that $x$ is not divisible by $p$, we conclude that $c\in (x+p\IZ)\cap(z+p\IZ)$ and hence $x-z\in p\IZ$, which contradicts the choice of $p>\max(E\cup F)$. This contradiction shows that $p\notin P$. Since $p\ge 3$, we can find a number $z'\notin\{0,z\}+p\IZ$ and using the Chinese Remainder Theorem~\ref{Chinese}, find a number
$$u\in (z'+p\IZ)\cap\bigcap_{q\in P}q\IZ^\bullet\subseteq \bigcap_{y\in F}\overline{y+d_y\IZ}\subseteq\{0,z\}+p\IZ,$$ which is a desired contradiction showing that $E\subseteq F$.
\end{proof}
As we know, the largest element of the superconnecting poset $\mathfrak F$ is the superconnecting filter $\F_\infty$. This filter can be characterized as follows.
\begin{lemma}\label{filter0} The superconnecting filter $\F_\infty$ of the Kirch space is generated by the base consisting of the sets $q\IN$ for an odd square-free number $q\in\IN$, i.e.
$$\F_\infty=\{B\subseteq \IZ^\bullet\colon \exists q\in (2\IN-1)\setminus\bigcup_{p\in\Pi} p^2\IN\;\;(q\IZ^\bullet\subseteq B)\}.$$
\end{lemma}
\begin{proof} Lemma~\ref{basic} implies that each element $F\in\F_\infty$ contains the set $q\IZ^\bullet$ for some odd square-free number $q$. Conversely, let $q$ be an odd square-free number. Then the sets $U_1=1+q\IZ$ and $U_2=2+q\IZ$ are open in the Kirch topology on $\IZ^\bullet$. By Lemma~\ref{basic} we have
$$\overline{U_1}\cap\overline{U_2}=\mathbb{Z}^\bullet\cap\bigcap_{p\in\Pi_q}(\{0,1\}+p\IZ)\cap(\{0,2\}+p\IZ)=\IZ^\bullet\cap\bigcap_{p\in\Pi_q}p\IZ= q\IZ^\bullet.$$
Hence $q\IZ^\bullet\in\F_\infty$.
\end{proof}
\begin{lemma}\label{l:2max} For a nonempty subset $E\subseteq \IZ^\bullet$ the following conditions are equivalent:
\begin{enumerate}
\item $\F_E=\F_\infty$;
\item $A_E=\{2\}$.
\end{enumerate}
If $|E|=2$, then the conditions \textup{(1), (2)} are equivalent to
\begin{enumerate}
\item[(3)] $E\in\big\{\{2^n,2^{n+1}\},\{-2^n,-2^{n+1}\},\{-2^n,2^n\}:n\in\w\big\}$.
\end{enumerate}
\end{lemma}
\begin{proof}
$(1)\Rightarrow(2)$: Assume $\F_{E}=\F_\infty$. Consider the set $F=\{1,2\}$ and observe that $A_F=\Pi_1\cup\Pi_2\cup\Pi_{2-1}=\{2\}$ and $\Pi_F=\emptyset$. Thus $A_F\subseteq A_E$, $\Pi_F{\setminus}\{2\}\subseteq \Pi_E$ and $\alpha_F{\restriction}A_F{\setminus} \Pi_E=\alpha_E{\restriction} A_F{\setminus} \Pi_E$. Lemma~\ref{l:wo2} implies $\F_{E}\subseteq \F_F$. Since $\F_{E}=\F_\infty$ is the largest element of $\mathfrak{F}$ we get $\F_E=\F_F$. By using again Lemma~\ref{l:wo2} we get $A_E\subseteq A_F$ which implies that $A_E = \{2\}$.
\vskip3pt
$(2)\Rightarrow(1)$:
If $A_E=\{2\}$, then by the Lemma~\ref{l:Kirch}, the filter $\F_{E}$ is generated by the base consisting of the sets $q\IZ^\bullet$ for an odd square-free number $q\in\IZ^\bullet$. Therefore $\F_E=\F_\infty$ by the Lemma~\ref{filter0}.
\smallskip
$(2)\Rightarrow(3)$:
Assume that $|E|=2$ and $A_E=\{2\}$. By Lemma~\ref{2ae}, $E=\{\varepsilon2^a,\delta2^b\}$, where $a,b\in \omega$ and $\varepsilon,\delta\in\{-1,1\}$. Without loss of generality we can assume that $b\leq a$. By Lemma~\ref{2ae}, $\Pi_{\varepsilon2^a-\delta2^b}=\Pi_{ \varepsilon2^b(2^{a-b}-\delta/\varepsilon)}\subseteq\{2\}$. The last inclusion implies that $a-b=1$ and $\delta/\varepsilon=1$ or $a-b=0$ and $\delta/\varepsilon=-1$. In the first case the set $E$ equals $\{2^b,2^{b+1}\}$ or $\{-2^b,-2^{b+1}\}$, in the second case $E=\{2^b,-2^{b}\}$.
$(3)\Rightarrow(2)$: The implication $(3)\Ra(2)$ follows from Lemma~\ref{2ae}.
\end{proof}
In the following lemmas by $\mathfrak F'$ we denote the set of maximal elements of the poset $\mathfrak F\setminus\{\F_\infty\}$.
\begin{lemma}\label{l:max} For a nonempty finite subset $E\subseteq \IZ^\bullet $ the filter $\F_E$ belongs to the family $\mathfrak F'$ if and only if there exists an odd prime number $p\notin\Pi_E$ such that $A_E=\{2,p\}$.
\end{lemma}
\begin{proof} To prove the ``if'' part, assume that $A_E=\{2,p\}$ and $p\notin\Pi_E$ for some odd prime number $p$. By Lemma~\ref{l:2max}, $\F_E\ne\F_\infty$. To show that the filter $\F_E$ is maximal in $\mathfrak F\setminus\{\F_\infty\}$, take any finite set $F\subset \IZ^\bullet $ such that $\F_E\subseteq\F_F\ne\F_\infty$. By Lemmas~\ref{l:wo2} and \ref{l:2max}, $\{2\}\ne A_F\subseteq\{2,p\}$, $\Pi_F\subseteq \Pi_E\cup\{2\}=\{2\}$, and $\alpha_F{\restriction}A_F\setminus \Pi_E=\alpha_E{\restriction}A_F\setminus \Pi_E$. It follows that $A_F=\{2,p\}=A_E$, $\Pi_F\cup\{2\}=\Pi_E\cup\{2\}$ and $\alpha_F=\alpha_E$. Applying Lemma~\ref{l:wo2}, we conclude that $\F_E=\F_F$, which means that the filter $\F_E$ is a maximal element of the poset $\mathcal F\setminus\{\F_\infty\}$.
\smallskip
To prove the ``only if'' part, assume that $\F_E\in\mathfrak F'$. By Lemma~\ref{l:2max}, $A_E\ne\{2\}$ and hence there exists an odd prime number $p\in A_E$.
We claim that $p\notin\Pi_E$. To derive a contradiction, assume that $p\in \Pi_E$ and consider the sets $F=\{p,2p\}$ and $G=\{1,p,2p\}$. By Lemma~\ref{2ae}, $A_F=A_G=\{2,p\}$, $\Pi_F=\{p\}$, and $\Pi_G=\emptyset$. Taking into account that $F\subset G$, $A_F=\{2,p\}\subseteq A_E$, $\Pi_F\setminus\{2\}=\{p\}\subseteq \Pi_E$ and $A_F\setminus\Pi_E\subseteq\{2\}$, we can apply Lemmas~\ref{l:wo2}, \ref{l:2max} and conclude that $\F_E\subseteq \F_F\subseteq\F_G\ne\F_\infty$. The maximality of $\F_E$ implies $\F_E=\F_F=\F_G$. By Lemma~\ref{l:wo2}, the equality $\F_G=\F_F$ implies $p\in\Pi_F\setminus\{2\}\subseteq\Pi_G=\emptyset$, which is a contradiction showing that $p\notin\Pi_E$.
Now consider the set $H=\{\alpha_E(p),p,2p\}$ and observe that $A_H=\{2,p\}$, $\Pi_H=\emptyset$ and $\alpha_H(p)=\alpha_E(p)$. Lemmas~\ref{l:wo2} and \ref{l:2max} guarantee that $\F_E\subseteq\F_H\ne\F_\infty$. By the maximality of $\F_E$, we have $\F_E=\F_H$. Applying Lemma~\ref{l:wo2} once more, we conclude that $A_E=A_H=\{2,p\}$.
\end{proof}
Lemma~\ref{l:max} implies the following description of the set $\mathfrak F'$.
\begin{lemma} $\mathfrak F'=\{\F_{\{a,p,2p\}}:p\in\Pi\setminus\{2\},\;\;a\in\{1,\dots,p-1\}\}$.
\end{lemma}
Let $\mathfrak F''$ be the set of maximal elements of the poset $\mathfrak F\setminus(\mathfrak F'\cup\{\mathcal F_\infty\})$
\begin{lemma}\label{efbis}
For a finite set $E\subset\IZ^\bullet$, the filter $\mathcal F_E$ belongs to the family $\mathfrak F''$ if and only if one of the following conditions holds:
\begin{enumerate}
\item there exists an odd prime number $p$ such that $p\in \Pi_E$ and $A_E=\{2,p\}$;
\item there are two distinct odd prime numbers $p,q$ such that $A_E=\{2,p,q\}$ and $\Pi_E\subseteq\{2\}$.
\end{enumerate}
\end{lemma}
\begin{proof} To prove the ``only if'' part, assume that $\F_E\in\mathfrak{F}''$. By Lemma~\ref{l:2max}, there is an odd prime number $p\in A_E$. If $A_E=\{2,p\}$, then $p\in\Pi_E$ by Lemma~\ref{l:max}, and condition (1) is satisfied. So, we assume that $\{2,p\}\ne A_E$ and find an odd prime number $q\in A_E\setminus\{2,p\}$. By Lemma~\ref{l:realization}, there is a number $x\in\IN$ such that for the set $F=\{x,pq,2pq\}$ we have $A_F=\{2,p,q\}$, $\Pi_F=\emptyset$, $\alpha_F(p)=\alpha_E(p)$ and $\alpha_F(q)=\alpha_E(q)$. Then $\F_E\subseteq \F_F$ by Lemma~\ref{l:wo2}, and $\F_F\in\mathfrak{F}\setminus(\mathfrak{F}'\cup\{\F_\infty\})$ by Lemma~\ref{l:max}. Now the maximality of the filter $\F_E$ implies that $\F_E=\F_F$ and hence $A_E=A_F=\{2,p,q\}$ and $\Pi_E\subseteq \Pi_F\cup\{2\}=\{2\}$, see Lemma~\ref{l:wo2}.
\smallskip
To prove the ``if'' part, we consider two cases. First we assume that $A_E=\{2,p\}$ and $p\in \Pi_E$ for some odd prime number $p$. By Lemmas~\ref{l:2max} and \ref{l:max}, $\F_{E}\in \mathfrak{F}\setminus(\{\F_\infty\}\cup \mathfrak{F}')$. To prove that $\F_E$ is a maximal element of $\mathfrak{F}\setminus(\{\F_\infty\}\cup \mathfrak{F}')$, take any finite set $F\subseteq\IZ^\bullet$ such that $\F_E\subseteq\F_F\in\mathfrak{F}\setminus(\{\F_\infty\}\cup \mathfrak{F}')$. Lemma~\ref{l:wo1} implies that $\min\{|E|,|F|\}\ge 2$ and then by Lemmas~\ref{l:wo2} and \ref{l:max}, we have $A_F=\{2,p\}$, $\Pi_F\setminus\{2\}\subseteq \{p\}$ and
$\alpha_E{\restriction}A_F\setminus\{p\}=\alpha_F{\restriction}A_F\setminus\{p\}$. Now notice that $p \in \Pi_F$ since otherwise $\F_F\in\mathfrak{F}'$ by Lemma~\ref{l:max}. By using again Lemma~\ref{l:wo2} we get $\F_F=\F_E$ which means that $\F_E\in\mathfrak{F}''$.
Now assume that there are two distinct odd prime numbers $p,q$ such that $A_E=\{2,p,q\}$ and $\Pi_E\subseteq\{2\}$. By Lemmas~\ref{l:2max} and \ref{l:max}, $\F_{E}\in \mathfrak{F}\setminus(\{\F_\infty\}\cup \mathfrak{F}')$. To prove that $\F_E$ is a maximal element of $\mathfrak{F}\setminus(\{\F_\infty\}\cup \mathfrak{F}')$, take any finite set $F\subseteq\IZ^\bullet$ such that $\F_E\subseteq \F_F\in\mathfrak{F}\setminus(\{\F_\infty\}\cup \mathfrak{F}')$. Lemma~\ref{l:wo2} implies that $A_F\subseteq \{2,p,q\}$, $\Pi_F\subseteq \{2\}$ and
$\alpha_E{\restriction}A_F\setminus\Pi_E=\alpha_F{\restriction}A_F\setminus\Pi_E$. Taking into account that $\F_F\notin\mathfrak F'\cup\{\F_\infty\}$ and $\Pi_F\subseteq\{2\}$, we can apply Lemmas~\ref{l:max}, \ref{l:2max} and conclude that $A_F=\{2,p,q\}$. We therefore know that $A_F=A_E$, $\Pi_E\cup\{2\}=\Pi_F\cup\{2\}$ and $\alpha_F{\restriction}A_E\setminus\Pi_F=\alpha_E{\restriction}A_E\setminus\Pi_F$. By Lemma~\ref{l:wo2}, $\F_E=\F_F$ and hence $\F_E\in\mathfrak{F}''$.
\end{proof}
\begin{lemma}\label{l:p2p-fix} For any homeomorphism $h$ of the Kirch space and any odd prime number $p$ we have $$\tilde{h}(\F_{\{p,2p\}})=\F_{\{p,2p\}}.$$
\end{lemma}
\begin{proof} By Proposition~\ref{p:iso}, the homeomorphism $h$ induces an order isomorphism $\tilde h$ of the superconnecting poset $\mathfrak F$ on the Kirch space. Then $\tilde h[\mathfrak F']=\mathfrak F'$ and $\tilde h[\mathfrak F'']=\mathfrak F''$.
By Lemmas~\ref{efbis} and \ref{l:realization}, $\mathfrak F''=\mathfrak F''_2\cup\mathfrak F''_3$ where
$$
\begin{aligned}
\mathfrak F''_2&=\{\F_{\{p,2p\}}:p\in\Pi\setminus\{2\}\}\quad\mbox{and}\\
\mathfrak F''_3&=\{\F_{\{x,pq,2pq\}}:p,q\in\Pi\setminus\{3\},\;x\in\{0,\dots,pq-1\}\setminus(p\IZ\cup q\IZ)\}.
\end{aligned}
$$
By Lemmas~\ref{l:wo2} and \ref{l:max}, for every filter $\F_{\{p,2p\}}\in\mathfrak F''_2$ the set
${\uparrow}\F_{\{p,2p\}}=\{\F\in \mathfrak F':\F_{\{p,2p\}}\subset\F_E\}$ coincides with the set $\{\F_{\{a,p,2p\}}:a\in\{1,\dots,p-1\}\}$ and hence has cardinality $p-1$.
On the other hand, for any filter $\F_{\{x,pq,2pq\}}\in\mathfrak F''_3$, the set ${\uparrow}\F_{\{x,pq,2pq\}}=\{\F\in \mathfrak F':\F_{\{x,pq,2pq\}}\subset\F\}$ coincides with the doubleton $\{\F_{\{x,p,2p\}},\F_{\{x,q,2q\}}\}$.
These order properties uniquely determine the filters $\F_{\{p,2p\}}$ for $p\in\Pi\setminus\{3\}$ and ensure that $\tilde h(\F_{\{p,2p\}})=\F_{\{p,2p\}}$ for every $p\in\Pi\setminus\{3\}$.
Next, observe that $\F_{\{3,6\}}$ is a unique element $\F$ of $\mathfrak F''$ such that ${\uparrow}\F\cap\bigcup_{p\in\Pi\setminus\{3\}}{\uparrow}\F_{\{p,2p\}}=\emptyset$. This uniqueness order property of $\F_{\{3,6\}}$ ensures that $\tilde h(\F_{\{3,6\}})=\F_{\{3,6\}}$.
\end{proof}
\begin{lemma}\label{lemmma} Let $E\subseteq \IZ^\bullet$ be a finite subset such that $A_E=\{2,p\}$ for some odd prime number $p\notin \Pi_E$. Then $A_{h[E]}=\{2,p\}$.
\end{lemma}
\begin{proof} By Lemma \ref{l:max}, $\F_E\in\mathfrak{F'}$. Consider the doubleton $\{p,2p\}$ which has $A_{\{p,2p\}}=\{2,p\}$ and $\Pi_{\{p,2p\}}=\{p\}$. By Lemma \ref{l:wo2}, $\mathcal{F}_{\{p,2p\}}\subseteq\F_E$ and by Lemma \ref{l:p2p-fix}, $\mathcal{F}_{\{p,2p\}}=\tilde{h}(\mathcal{F}_{\{p,2p\}})=\mathcal{F}_{\{h(p),h(2p)\}}\subseteq\mathcal{F}_{h[E]}$. By Lemma \ref{l:wo2}, $A_{h[E]}\subseteq A_{\{p,2p\}}=\{2,p\}$. By Lemma 8, $A_{h[E]}=\{2,p\}$.
\end{proof}
\begin{definition} A homeomorphism $h$ of the Kirch space $(\IZ^\bullet,\tau_K)$ is called {\em positive} if $h(1)>0$.
\end{definition}
\begin{lemma}\label{l:2fix} Every positive homeomorphism $h$ of the Kirch space has $h(x)=x$ for any $x\in\{\pm2^n, n\in\omega\}$
\end{lemma}
\begin{proof} Consider the graph $\Gamma_2=(V_2,\mathcal E)$ with set of vertices $V_2=\{\pm2^n:n\in\omega\}$ and set of edges $\mathcal E=\big\{\{2^n,2^{n+1}\},\{-2^n,-2^{n+1}\},\{-2^n,2^n\}:n\in\omega\big\}$.
Observe that $1$ and $-1$ are the unique vertices of $\Gamma_2$ that have order 2. Any other vertices have order 3. This ensures that $h(1)=\pm1$. The positivity of $h$ implies that $h(1)=1$. Then $h(-1)=-1$, $h(2)=2$. Hence $h(\pm2^n)=\pm2^n$ for all $n\in\w$.
\end{proof}
Lemmas~\ref{l:2fix} and \ref{l:p2p-fix} imply
\begin{lemma}\label{l:12fix} For any positive homeomorphism $h$ of the Kirch space and any odd prime number $p$ we have $$\tilde{h}(\F_{\{1,p,2p\}})=\F_{\{1,p,2p\}}\quad\mbox{and}\quad \tilde{h}(\F_{\{2,p,2p\}})=\F_{\{2,p,2p\}}.$$
\end{lemma}
\begin{lemma}\label{ppix} For an integer number $x\in\IZ^\bullet\setminus\{-2,-1,1,2\}$ and an odd prime $p$, the following conditions are equivalent:
\begin{enumerate}
\item $p\in\Pi_x$;
\item $\F_{\{1,x\}}\subseteq \F_{\{1,p,2p\}}$ and $\F_{\{2,x\}}\subseteq \F_{\{2,p,2p\}}$.
\end{enumerate}
\end{lemma}
\begin{proof} If $p\in\Pi_x$, then $A_{\{1,p,2p\}}=\{2,p\}\subseteq A_{\{1,x\}}$, $\Pi_{\{1,x\}}\cup\{2\}=\{2\}=\Pi_{\{1,p,2p\}}\cup\{2\}$ and $\alpha_{\{1,x\}}(p)=1=\alpha_{\{1,p,2p\}}(p)$. By Lemma~\ref{l:wo2}, $\F_{\{1,x\}}\subseteq\F_{\{1,p,2p\}}$. By analogy we can prove that $\F_{\{2,x\}}\subseteq \F_{\{2,p,2p\}}$.
Conversely, assume $\F_{\{1,x\}}\subseteq \F_{\{1,p,2p\}}$ and $\F_{\{2,x\}}\subseteq \F_{\{2,p,2p\}}$. By Lemmas~\ref{l:wo2} and \ref{2ae}, we have
$$\{2,p\}=A_{\{1,p,2p\}}\subseteq A_{\{1,x\}}=\Pi_{x}\cup\Pi_{x-1}
\text{ and }\{2,p\}=A_{\{2,p,2p\}}\subseteq A_{\{2,x\}}=\{2\}\cup\Pi_x\cup\Pi_{x-2}$$
and hence $p\in (\Pi_x\cup\Pi_{x-1})\cap(\Pi_x \cup\Pi_{x-2})\setminus\{2\}\subseteq \Pi_x$.
\end{proof}
Proposition~\ref{p:iso} and Lemmas~\ref{l:2fix}, \ref{l:12fix}, \ref{ppix} imply
\begin{lemma}\label{Pix} For every homeomorphism $h$ of the Kirch space and any number $x\in\IN$ we have $$\Pi_x\cup\{2\}=\Pi_{h(x)}\cup\{2\}.$$
\end{lemma}
For every prime number $p$ consider the set $$V_p=\{\pm2^{n-1}p^m:n,m\in\IN\}$$ of numbers $x\in\IN$ such that $p\in\Pi_x\subseteq \{2,p\}$. Lemmas~\ref{l:2fix} and \ref{Pix} imply that $h[V_p]=V_p$ for every homeomorphism $h$ of the Kirch space.
Consider the graph $\Gamma_p=(V_p,\mathcal E_p)$ on the set $V_p$ with the set of edges $$\mathcal E_p:=\big\{E\in[V_p]^2:A_E=\{2,p\}\big\}.$$
\begin{lemma}\label{l:graph} For every prime number $p$ and every homeomorphism $h$ of the Kirch space, the restriction of $h$ to $V_p$ is an isomorphism of the graph $\Gamma_p$.
\end{lemma}
\begin{proof} Let $E\in \mathcal{E}_p$. Since $p\in \Pi_E$, we can apply Lemma~\ref{efbis} and conclude that $\F_E\in\mathfrak{F}''$. Using fact that $\tilde{h}$ is isomorphism of $\mathfrak{F}$ we get $\F_{h[E]}=\tilde{h}(\F_E)\in\mathfrak{F}''$. Since $h[E]\subseteq h[V_p]=V_p$, we obtain $p\in \Pi_{h[E]}$. Using Lemma~\ref{efbis} once more, we obtain that $A_{h[E]}=\{2,p\}$, which means that $h[E]\in \mathcal{E}_p$. By analogical reasoning we can prove that $h^{-1}[E]\in \mathcal{E}_p$ for every $E\in \mathcal{E}_p$. This means that $h{\restriction}V_p$ is isomorphism of the graph $\Gamma_p$.
\end{proof}
The structure of the graph $\Gamma_p$ depends on properties of the prime number $p$.
A prime number $p$ is called
\begin{itemize}
\item {\em Fermat prime} if $p=2^n+1$ for some $n\in\IN$;
\item {\em Mersenne prime} if $p=2^n-1$ for some $n\in\IN$;
\item {\em Fermat--Mersenne} if $p$ is Fermat prime or Mersenne prime.
\end{itemize}
It is known (and easy to see) that for any Fermat prime number $p=2^n+1$ the exponent $n$ is a power of $2$, and for any Mersenne prime number $p=2^n-1$ the power $n$ is a prime number. It is not known whether there are infinitely many Fermat--Mersenne prime numbers. All known Fermat prime numbers are the numbers $2^{2^n}+1$ for $0\le n\le 4$ (see {\tt oeis.org/A019434} in \cite{OEIS}). At the moment only 51 Mersenne prime numbers are known, see the sequence {\tt oeis.org/A000043} in \cite{OEIS}.
\begin{lemma}\label{struct} Let $p$ be an odd prime number, $p\neq3$.
\begin{enumerate}
\item If $p=3$, then the set $\mathcal E_p$ of the edges of the graph $\Gamma_p$ coincides with the set of doubletons\\
$\{\varepsilon2^{a-1}3^b,\varepsilon2^{a-1}3^{b+1}\}$, $\{\varepsilon2^{a-1}3^b,2\varepsilon^{a-1}3^{b+2}\}$, $\{\varepsilon2^{a-1}3^b,\varepsilon2^{a}3^{b}\}$, $\{\varepsilon2^{a-1}3^b,\varepsilon2^{a+1}3^{b}\}$, $\{\varepsilon2^{a-1}3^{b+1},\varepsilon2^{a+1}3^b\}$,\\
$\{\varepsilon2^{a+1}3^{b},\varepsilon2^{a}3^{b+1}\}$, $\{\varepsilon2^{a+3}3^b,\varepsilon2^a3^{b+2}\},$ $ \{\varepsilon2^{a-1}3^b,-\varepsilon2^{a-1}3^{b+1}\}$, $\{\varepsilon2^{a-1}3^b,-\varepsilon2^{a}3^{b}\}$\\ $\{\varepsilon2^{a-1}3^b,-\varepsilon2^{a+2}3^{b}\}$, $\{\varepsilon2^{a-1}3^b,-\varepsilon2^{a-1}3^b\}$
for some $a,b\in\IN,\ \varepsilon \in\{-1,1\}$.
\item If $p=2^m+1>3$ is Fermat prime, then\newline $\mathcal E_p=\big\{\{\varepsilon2^{a-1}p^b,\varepsilon2^{a-1}p^{b+1}\}, \{\varepsilon2^{a-1}p^b,\varepsilon2^{a}p^b\}, \{\varepsilon2^{a-1}p^b,-\varepsilon2^{a+m-1}p^{b}\},$
$ \{\varepsilon2^{a-1}p^b,-\varepsilon2^{a-1}p^b\}, \{\varepsilon2^{m+a-1}p^{b},\varepsilon2^{a-1}p^{b+1}\}:a,b\in\IN, \ \varepsilon \in\{-1,1\}\big\}$.
\item
If $p=2^m-1>3$ is Mersenne prime, then
$\mathcal E_p=\big\{\{\varepsilon2^ap^b,\varepsilon2^{a-1}p^b\},\{\varepsilon2^{a-1}p^b,\varepsilon2^{m+a-1}p^b\}, $\break$ \{\varepsilon2^{a-1}p^{b+1},\varepsilon2^{m+a-1}p^b\},$ $\{\varepsilon2^{a-1}p^b,-\varepsilon2^{a-1}p^b\}, \{\varepsilon2^{a-1}p^b,-\varepsilon2^{a-1}p^{b+1}\}
:a,b\in\IN, \ \varepsilon \in\{-1,1\}\big\}$.
\item If $p$ is not Fermat--Mersenne, then $\mathcal E_p=\big\{\{\varepsilon2^{a-1}p^b,-\varepsilon2^{a-1}p^b\}, \{\varepsilon2^{a-1}p^b,\varepsilon2^{a}p^b\}:a,b\in\IN, \ \varepsilon \in\{-1,1\}\big\}$.
\end{enumerate}
\end{lemma}
\begin{proof} Proof of Lemma~\ref{struct} in each of cases (1)--(4) will be similar. The edges of graph $\Gamma_p$ are $2$-element subsets of set $V_p$ such that $A_E=\{2,p\}$. Since the vertices of graph $\Gamma_p$ are numbers of the form $\pm2^{n-1}p^m$, where $n,m\in\IN$, we can apply Lemma~\ref{2ae} and conclude that a doubleton $\{x,y\}\subset V_p$ belongs to $\mathcal E_p$ if and only if $\{2,p\}= \Pi_x\cup\Pi_y\cup\Pi_{x-y}$. In subsequent proofs, we will intensively use the Mih\u ailescu Theorem~\ref{Mihailescu} saying that $2^3,3^2$ is a unique pair of consecutive powers.
1. First we consider the case of $p=3$. It is easy to see that the doubletons $\{x,y\}$ written in the statement (1) have $\Pi_x\cup\Pi_y\cup\Pi_{x-y}\subseteq\{2,3\}$, which implies that $\{x,y\}\in\mathcal E_3$. It remains to show that every doubleton $\{x,y\}\in\mathcal E_3$ is of the form indicated in the statement (1). Write $\{x,y\}$ as $\{\varepsilon 2^{a-1}3^b,\delta2^{c-1}3^d\}$ for some $a,b,c,d\in\IN$, $\varepsilon,\delta \in\{-1,1\}$ such that $2^{a-1}3^b\le 2^{c-1}3^d$.
If $a=c$ and $b=d$, then $\varepsilon\ne\delta$ and $\{x,y\}=\{\varepsilon 2^{a-1}3^b,-\varepsilon 2^{a-1}3^b\}$.
If $a=c$, then $b\le d$ and the inclusion $\Pi_{x-y}\subseteq\{2,3\}$ implies that $\Pi_{3^{d-b}-\varepsilon/\delta}\subseteq\{2,3\}$ and hence $3^{d-b}-\varepsilon/\delta$ is a power of $2$. If $\varepsilon/\delta=1$ then by the Mih\u ailescu Theorem~\ref{Mihailescu}, $d-b\in\{1,2\}$, which means that $\{x,y\}$ is equal to $\{\varepsilon2^{a-1}3^b,\varepsilon2^{a-1}3^{b+1}\}$ or $\{\varepsilon2^{a-1}3^b,\varepsilon2^{a-1}3^{b+2}\}$. If $\varepsilon/\delta=-1$ then by the Mih\u ailescu Theorem~\ref{Mihailescu}, $d-b\in\{0,1\}$, which means that $\{x,y\}$ is equal to $\{\varepsilon2^{a-1}3^b,-\varepsilon2^{a-1}3^{b}\}$ or $\{\varepsilon2^{a-1}3^b,-\varepsilon2^{a-1}3^{b+1}\}$.
If $b=d$, then $a\le c$ and the inclusion $\Pi_{x-y}\subseteq\{2,3\}$ implies that $\Pi_{2^{c-a}-\varepsilon/\delta}\subseteq\{2,3\}$ and hence $2^{c-a}-\varepsilon/\delta$ is either 2 or a power of $3$. If $\varepsilon/\delta=1$ then by the Mih\u ailescu Theorem~\ref{Mihailescu}, $c-a\in\{1,2\}$, which means that $\{x,y\}$ is equal to $\{\varepsilon2^{a-1}3^b,\varepsilon2^{a}3^{b}\}$ or $\{\varepsilon2^{a-1}3^b,\varepsilon2^{a+1}3^{b}\}$. If $\varepsilon/\delta=-1$ then by the Mih\u ailescu Theorem~\ref{Mihailescu}, $c-a\in\{0,1,3\}$, which means that $\{x,y\}$ is equal to $\{\varepsilon2^{a-1}3^b,-\varepsilon2^{a-1}3^{b}\}$, $\{\varepsilon2^{a-1}3^b,-\varepsilon2^{a}3^{b}\}$ or $\{\varepsilon2^{a-1}3^b,-\varepsilon2^{a+2}3^{b}\}$.
So, we assume that $a\ne c$ and $b\ne d$. In this case we should consider four subcases.
If $a<c$ and $b<d$, then $\Pi_{x-y}\subseteq\{2,3\}$ implies that each prime divisor of $2^{c-a}3^{d-b}-\varepsilon/\delta$ is equal to $2$ or $3$, which is not possible.
If $a<c$ and $b>d$, then $\Pi_{x-y}\subseteq\{2,3\}$ and $2^{a-1}3^b\le 2^{c-1}3^d$ imply that $2^{c-a}-(\varepsilon/\delta)3^{b-d}=1$ which implies that $\varepsilon=\delta$. Hence $c-a=2$ and $b-d=1$ by the Mih\u ailescu Theorem~\ref{Mihailescu}. In this case $\{x,y\}=\{\varepsilon 2^{a-1}3^{d+1},\varepsilon 2^{a+1}3^d\}$.
If $a>c$ and $b<d$, then $\Pi_{x-y}\subseteq\{2,3\}$ and $2^{a-1}3^b\le 2^{c-1}3^d$ imply that $3^{d-b}-(\varepsilon/\delta)2^{a-c}=1$. This implies that $\varepsilon/\delta=1$ and hence $\langle d-b,a-c\rangle\in\{\langle 1,1\rangle,\langle 2,3\rangle\}$ by the Mih\u ailescu Theorem~\ref{Mihailescu}. In this case $\{x,y\}$ is equal to $\{2^{c+1}3^{b},2^{c}3^{b+1}\}$ or $\{2^{c+3}3^b,2^c3^{b+2}\}$.
The subcase $a>c$ and $b>d$ is forbidden by the inequality $2^{a-1}3^b\le 2^{c-1}3^d$.
\smallskip
\begin{figure}
$$\xymatrix@C50pt{
\vdots\ar@{-}@/_14pt/[dddddddd]&\vdots\ar@{-}@/_14pt/[dddddddd]&\vdots\ar@{-}[dddrr]\ar@{-}@/_14pt/[dddddddd]&\vdots\ar@{-}@/_14pt/[dddddddd]\\
2^3{\cdot}3\ar@{-}@/_14pt/[dddddddd]\ar@{-}[dr]\ar@{-}@/_32pt/[ddddddd]\ar@{-}[dddddddr]\ar@{-}[d]\ar@{-}@/^15pt/[rr]\ar@{-}[r]&2^3{\cdot}3^2\ar@{-}@/_14pt/[dddddddd]\ar@{-}[dr]\ar@{-}@/_32pt/[ddddddd]\ar@{-}[dddddddr]\ar@{-}[r]\ar@{-}[d]\ar@{-}@/^15pt/[rr]&2^3{\cdot}3^3\ar@{-}@/_14pt/[dddddddd]\ar@{-}[dr]\ar@{-}[dddrr]\ar@{-}@/_32pt/[ddddddd]\ar@{-}[dddddddr]\ar@{-}[r]\ar@{-}[d]\ar@{-}@/^15pt/[rr]&2^3{\cdot}3^4\ar@{-}@/_14pt/[dddddddd]\ar@{-}[dr]\ar@{-}@/_32pt/[ddddddd]\ar@{-}[r]\ar@{-}[d]&\dots\\
2^2{\cdot}3\ar@{-}@/_12pt/[dddddd]\ar@{-}[dr]\ar@{-}[ddr]\ar@{-}@/_21pt/[ddddd]\ar@{-}[dddddr]\ar@{-}@/_15pt/[uu]\ar@{-}[d]\ar@{-}@/^15pt/[rr]\ar@{-}[r]&2^2{\cdot}3^2\ar@{-}@/_12pt/[dddddd]\ar@{-}[dr]\ar@{-}@/_21pt/[ddddd]\ar@{-}[dddddr]\ar@{-}[uul]\ar@{-}@/_15pt/[uu]\ar@{-}[r]\ar@{-}[d]\ar@{-}@/^15pt/[rr]&2^2{\cdot}3^3\ar@{-}@/_12pt/[dddddd]\ar@{-}[dr]\ar@{-}@/_21pt/[ddddd]\ar@{-}[dddddr]\ar@{-}[uul]\ar@{-}@/_15pt/[uu]\ar@{-}[r]\ar@{-}[d]\ar@{-}@/^15pt/[rr]&2^2{\cdot}3^4\ar@{-}@/_12pt/[dddddd]\ar@{-}[dr]\ar@{-}@/_21pt/[ddddd]\ar@{-}[uul]\ar@{-}@/_15pt/[uu]\ar@{-}[r]\ar@{-}[d]&\dots\\
2{\cdot}3\ar@{-}@/_10pt/[dddd]\ar@{-}[dr]\ar@{-}@/_18pt/[ddd]\ar@{-}@/^22pt/[dddddd]\ar@{-}[dddr]\ar@{-}@/_15pt/[uu]\ar@{-}[d]\ar@{-}@/^15pt/[rr]\ar@{-}[r]&2{\cdot}3^2\ar@{-}@/_10pt/[dddd]\ar@{-}[dr]\ar@{-}@/_18pt/[ddd]\ar@{-}@/^22pt/[dddddd]\ar@{-}[dddr]\ar@{-}[uul]\ar@{-}@/_15pt/[uu]\ar@{-}[r]\ar@{-}[d]\ar@{-}@/^15pt/[rr]&2{\cdot}3^3\ar@{-}@/_10pt/[dddd]\ar@{-}[dr]\ar@{-}@/_18pt/[ddd]\ar@{-}@/^22pt/[dddddd]\ar@{-}[dddr]\ar@{-}[uuull]\ar@{-}[uul]\ar@{-}@/_15pt/[uu]\ar@{-}[r]\ar@{-}[d]\ar@{-}@/^15pt/[rr]&2{\cdot}3^4\ar@{-}@/_10pt/[dddd]\ar@{-}[dr]\ar@{-}@/_18pt/[ddd]\ar@{-}@/^22pt/[dddddd]\ar@{-}[uuull]\ar@{-}[uul]\ar@{-}@/_15pt/[uu]\ar@{-}[r]\ar@{-}[d]&\dots\\
3\ar@{-}@/_8pt/[dd]\ar@{-}@/^20pt/[dddd]\ar@{-}[rd]\ar@{-}@/_15pt/[uu]\ar@{-}[d]\ar@{-}@/^15pt/[rr]\ar@{-}[r]&3^2\ar@{-}@/_8pt/[dd]\ar@{-}@/^20pt/[dddd]\ar@{-}[rd]\ar@{-}@/_15pt/[uu]\ar@{-}[r]\ar@{-}[d]\ar@{-}@/^15pt/[rr]&3^3\ar@{-}@/_8pt/[dd]\ar@{-}@/^20pt/[dddd]\ar@{-}[rd]\ar@{-}[uul]\ar@{-}[uuull]\ar@{-}@/_15pt/[uu]\ar@{-}[r]\ar@{-}[d]\ar@{-}@/^15pt/[rr]&3^4\ar@{-}@/_8pt/[dd]\ar@{-}@/^20pt/[dddd]\ar@{-}[uul]\ar@{-}[uuull]\ar@{-}@/_15pt/[uu]\ar@{-}[r]\ar@{-}[d]&\dots\ar@{-}[dl]\\
-3\ar@{-}@/_8pt/[uu]\ar@{-}@/_20pt/[uuuu]\ar@{-}[ur]\ar@{-}[d]\ar@{-}@/^15pt/[dd]\ar@{-}@/_15pt/[rr]\ar@{-}[r]&-3^2\ar@{-}@/_8pt/[uu]\ar@{-}@/_20pt/[uuuu]\ar@{-}[ur]\ar@{-}[r]\ar@{-}[d]\ar@{-}@/_15pt/[rr]\ar@{-}[ddl]\ar@{-}@/^15pt/[dd]&-3^3\ar@{-}@/_8pt/[uu]\ar@{-}@/_20pt/[uuuu]\ar@{-}[ur]\ar@{-}[ddl]\ar@{-}[dddll]\ar@{-}@/^15pt/[dd]\ar@{-}[r]\ar@{-}[d]\ar@{-}@/_15pt/[rr]&-3^4\ar@{-}@/_8pt/[uu]\ar@{-}@/_20pt/[uuuu]\ar@{-}[ddl]\ar@{-}@/^15pt/[dd]\ar@{-}[r]\ar@{-}[d]\ar@{-}[dddll]&\dots\ar@{-}[dddll]\ar@{-}[ul]\\
-2{\cdot}3\ar@{-}@/_10pt/[uuuu]\ar@{-}[ur]\ar@{-}@/_22pt/[uuuuuu]\ar@{-}[uuur]\ar@{-}@/^15pt/[dd]\ar@{-}[d]\ar@{-}@/_15pt/[rr]\ar@{-}[r]&-2{\cdot}3^2\ar@{-}@/_10pt/[uuuu]\ar@{-}[ur]\ar@{-}@/_22pt/[uuuuuu]\ar@{-}[uuur]\ar@{-}[ddl]\ar@{-}@/^15pt/[dd]\ar@{-}[r]\ar@{-}[d]\ar@{-}@/_15pt/[rr]&-2{\cdot}3^3\ar@{-}@/_10pt/[uuuu]\ar@{-}[ur]\ar@{-}@/_22pt/[uuuuuu]\ar@{-}[uuur]\ar@{-}[dddll]\ar@{-}[ddl]\ar@{-}@/^15pt/[dd]\ar@{-}[r]\ar@{-}[d]\ar@{-}@/_15pt/[rr]&-2{\cdot}3^4\ar@{-}@/_10pt/[uuuu]\ar@{-}[ur]\ar@{-}@/_22pt/[uuuuuu]\ar@{-}[dddll]\ar@{-}[ddl]\ar@{-}@/^15pt/[dd]\ar@{-}[r]\ar@{-}[d]&\dots\ar@{-}[dddll]\\
-2^2{\cdot}3\ar@{-}@/_12pt/[uuuuuu]\ar@{-}[ur]\ar@{-}[uuuuur]\ar@{-}@/^15pt/[dd]\ar@{-}[d]\ar@{-}@/_15pt/[rr]\ar@{-}[r]&-2^2{\cdot}3^2\ar@{-}@/_12pt/[uuuuuu]\ar@{-}[ur]\ar@{-}[uuuuur]\ar@{-}[ddl]\ar@{-}@/^15pt/[dd]\ar@{-}[r]\ar@{-}[d]\ar@{-}@/_15pt/[rr]&-2^2{\cdot}3^3\ar@{-}@/_12pt/[uuuuuu]\ar@{-}[ur]\ar@{-}[uuuuur]\ar@{-}[ddl]\ar@{-}@/^15pt/[dd]\ar@{-}[r]\ar@{-}[d]\ar@{-}@/_15pt/[rr]&-2^2{\cdot}3^4\ar@{-}@/_12pt/[uuuuuu]\ar@{-}[ur]\ar@{-}[ddl]\ar@{-}@/^15pt/[dd]\ar@{-}[r]\ar@{-}[d]&\dots\\
-2^3{\cdot}3\ar@{-}@/_15pt/[uuuuuuuu]\ar@{-}[ur]\ar@{-}[uuuuuuur]\ar@{-}[d]\ar@{-}@/_15pt/[rr]\ar@{-}[r]&-2^3{\cdot}3^2\ar@{-}@/_15pt/[uuuuuuuu]\ar@{-}[ur]\ar@{-}[uuuuuuur]\ar@{-}[r]\ar@{-}[d]\ar@{-}@/_15pt/[rr]&-2^3\ar@{-}@/_15pt/[uuuuuuuu]\ar@{-}[ur]\ar@{-}[uuuuuuur]{\cdot}3^3\ar@{-}[r]\ar@{-}[d]\ar@{-}@/_15pt/[rr]&-2^3\ar@{-}@/_15pt/[uuuuuuuu]\ar@{-}[ur]{\cdot}3^4\ar@{-}[r]\ar@{-}[d]&\dots\\
\vdots&\vdots&\vdots&\vdots\\
}
$$
\caption{The graph $\Gamma_3$}\label{fig3}
\end{figure}
2. Assume that $p=2^m+1>3$ is a Fermat prime. In this case $m>1$. Since $p>3$, $p$ is not Mersenne prime. It is easy to check that every doubleton $\{x,y\}\in\big\{\{\varepsilon2^{a-1}p^b,\varepsilon2^{a-1}p^{b+1}\}, \{\varepsilon2^{a-1}p^b,\varepsilon2^{a}p^b\}, \{\varepsilon2^{a-1}p^b,-\varepsilon2^{a+m-1}p^{b}\},$\break$ \{\varepsilon2^{a-1}p^b,-\varepsilon2^{a-1}p^b\}$,
$\{\varepsilon2^{m+a-1}p^{b},\varepsilon2^{a-1}p^{b+1}\}:a,b\in\IN, \ \varepsilon \in\{-1,1\}\big\}$ has $A_{\{x,y\}}= \Pi_x\cup\Pi_y\cup\Pi_{x-y}=\{2,p\}$ and hence $\{x,y\}\in\mathcal E_p$.
Now assume that $\{x,y\}\in\mathcal E_p$ is an edge of the graph $\Gamma_p$. Then $\Pi_x\cup\Pi_y\cup\Pi_{x-y}=A_{\{x,y\}}=\{2,p\}$ and $\{x,y\}$ can be written as $\{\varepsilon2^{a-1}p^b,\delta2^{c-1}p^d\}$ for some $a,b,c,d\in\IN$, $\varepsilon,\delta \in\{-1,1\}$ with $2^{a-1}p^b\le 2^{c-1}p^d$.
If $a=c$, $b=d$ and $\varepsilon=-\delta$ then $\Pi_{\varepsilon2^{a-1}p^b-\delta2^{a-1}p^{b}}=\Pi_{\varepsilon2^{a}p^b}\subset\{2,p\}$. In this case $\{x,y\}=\{\varepsilon2^{a-1}p^b,-\varepsilon2^{a-1}p^b\}$.
If $a=c$, then $b\le d$ and the inclusion $\Pi_{x-y}\subseteq\{2,p\}$ implies that $\Pi_{p^{d-b}-\varepsilon/\delta}\subseteq\{2,p\}$ and hence $p^{d-b}-\varepsilon/\delta$ is a power of $2$. By the Mih\u ailescu Theorem~\ref{Mihailescu}, $d-b\in\{0,1\}$. If $d-b=0$, then $\varepsilon=-\delta$ and $\{x,y\}=\{\varepsilon 2^{a-1}p^b,-\varepsilon 2^{a-1}p^b\}$ by the preceding case. So, we assume that $d-b=1$. Since $p$ is not Mersenne prime, we conclude that $\varepsilon=\delta$, and hence $\{x,y\}$ is equal to $\{ \varepsilon2^{a-1}p^b,\varepsilon2^{a-1}p^{b+1}\}$.
If $b=d$, then $a\le c$ and the inclusion $\Pi_{x-y}\subseteq\{2,p\}$ implies that $\Pi_{2^{c-a}-\varepsilon/\delta}\subseteq\{2,p\}$ and hence $2^{c-a}-\varepsilon/\delta$ is a power of $p$. By the Mih\u ailescu Theorem~\ref{Mihailescu}, $2^{c-a}-\varepsilon/\delta\in\{1,p\}=\{1,2^m+1\}$. If $\varepsilon=\delta$ then $c-a=1$, which means that $\{x,y\}$ is equal to $\{\varepsilon2^{a-1}p^b,\varepsilon2^{a}p^{b}\}$. If $\varepsilon=-\delta$ then $c-a=m$ and $\{x,y\}=\{\varepsilon2^{a-1}p^b,-\varepsilon2^{a+m-1}p^{b}\}$.
So, we assume that $a\ne c$ and $b\ne d$. In this case we should consider four subcases.
If $a<c$ and $b<d$, then $\Pi_{x-y}\subseteq\{2,p\}$ implies that each prime divisor of $2^{c-a}p^{d-b}-\varepsilon/\delta$ is equal to $2$ or $p$, which is not possible.
If $a<c$ and $b>d$, then $\Pi_{x-y}\subseteq\{2,p\}$ implies that $2^{c-a}-(\varepsilon/\delta) p^{b-d}=1$ and hence $\varepsilon=\delta$. In this case the Mih\u ailescu Theorem~\ref{Mihailescu} ensures that $b-d=1$ and hence $2^{c-a}=p+1=2^m+2$ which is not possible (as $m>1$).
If $a>c$ and $b<d$, then $\Pi_{x-y}\subseteq\{2,p\}$ implies that $p^{d-b}-(\varepsilon/\delta)2^{a-c}=1$, which implies that $\varepsilon=\delta$. The Mih\u ailescu Theorem~\ref{Mihailescu} implies that $d-b=1$ and hence $2^{a-c}=p-1=2^m$ and $a-c=m$. In this case $\{x,y\}=\{\varepsilon2^{c+m-1}2^b,\varepsilon2^{c-1}p^{b+1}\}$.
The subcase $a>c$ and $b>d$ is forbidden by the inequality $2^{a-1}p^b\le 2^{c-1}p^d$.
\begin{figure}
$$
\xymatrix@C40pt{
\vdots\ar@{-}@/^30pt/[ddddddd]\ar@{-}@/_30pt/[ddddddddd]\ar@{-}[d]\ar@{-}[rdd]&\vdots\ar@{-}@/_30pt/[ddddddddd]\ar@{-}@/^30pt/[ddddddd]\ar@{-}[d]\ar@{-}[rdd]&\ar@{-}@/_30pt/[ddddddddd]\ar@{-}@/^30pt/[ddddddd]\vdots\ar@{-}[d]\ar@{-}[rdd]&\vdots\ar@{-}@/_30pt/[ddddddddd]\ar@{-}@/^30pt/[ddddddd]\ar@{-}[d]\ar@{-}[rdd]&\\
2^3{\cdot}5\ar@{-}[r]\ar@{-}[d]\ar@{-}[rdd]\ar@{-}@/_25pt/[ddddddd]\ar@{-}@/^25pt/[ddddd]&2^3{\cdot}5^2\ar@{-}[r]\ar@{-}[d]\ar@{-}[rdd]\ar@{-}@/^25pt/[ddddd]\ar@{-}@/_25pt/[ddddddd]&2^3{\cdot}5^3\ar@{-}[r]\ar@{-}[d]\ar@{-}@/^25pt/[ddddd]\ar@{-}[rdd]\ar@{-}@/_25pt/[ddddddd]&2^3{\cdot}5^4\ar@{-}[r]\ar@{-}[d]\ar@{-}@/^25pt/[ddddd]\ar@{-}[rdd]\ar@{-}@/_25pt/[ddddddd]&\dots\\
2^2{\cdot}5\ar@{-}@/^12pt/[ddd]\ar@{-}[r]\ar@{-}[d]\ar@{-}[rdd]\ar@{-}@/^30pt/[ddddddd]\ar@{-}@/_20pt/[ddddd]&2^2{\cdot}5^2\ar@{-}@/^12pt/[ddd]\ar@{-}[r]\ar@{-}[d]\ar@{-}[rdd]\ar@{-}@/_20pt/[ddddd]\ar@{-}@/^30pt/[ddddddd]&2^2{\cdot}5^3\ar@{-}@/^12pt/[ddd]\ar@{-}[r]\ar@{-}[d]\ar@{-}[rdd]\ar@{-}@/_20pt/[ddddd]\ar@{-}@/^30pt/[ddddddd]&2^2{\cdot}5^4\ar@{-}@/^12pt/[ddd]\ar@{-}[r]\ar@{-}[d]\ar@{-}[rdd]\ar@{-}@/^30pt/[ddddddd]\ar@{-}@/_20pt/[ddddd]&\dots\\
2{\cdot}5\ar@{-}[r]\ar@{-}[d]\ar@{-}@/^25pt/[ddddd]\ar@{-}@/_15pt/[ddd]&2{\cdot}5^2\ar@{-}[r]\ar@{-}[d]\ar@{-}@/^25pt/[ddddd]\ar@{-}@/_15pt/[ddd]&2{\cdot}5^3\ar@{-}[r]\ar@{-}[d]\ar@{-}@/^25pt/[ddddd]\ar@{-}@/_15pt/[ddd]&2{\cdot}5^4\ar@{-}[r]\ar@{-}[d]\ar@{-}@/^25pt/[ddddd]\ar@{-}@/_15pt/[ddd]&\dots\\
5\ar@{-}[r]\ar@{-}[d]\ar@{-}@/^12pt/[ddd]&5^2\ar@{-}[r]\ar@{-}[d]\ar@{-}@/^12pt/[ddd]&5^3\ar@{-}[r]\ar@{-}[d]\ar@{-}@/^12pt/[ddd]&5^4\ar@{-}[r]\ar@{-}[d]\ar@{-}@/^12pt/[ddd]&\dots\\
-5\ar@{-}[r]\ar@{-}[d]&-5^2\ar@{-}[r]\ar@{-}[d]\ar@{-}[ldd]&-5^3\ar@{-}[r]\ar@{-}[d]\ar@{-}[ldd]&-5^4\ar@{-}[r]\ar@{-}[d]\ar@{-}[ldd]&\dots\ar@{-}[ldd]\\
-2{\cdot}5\ar@{-}[r]\ar@{-}[d]&-2{\cdot}5^2\ar@{-}[r]\ar@{-}[d]\ar@{-}[ldd]&-2{\cdot}5^3\ar@{-}[r]\ar@{-}[d]\ar@{-}[ldd]&-2{\cdot}5^4\ar@{-}[r]\ar@{-}[d]\ar@{-}[ldd]&\dots\ar@{-}[ldd]\\
-2^2{\cdot}5\ar@{-}[r]\ar@{-}[d]&-2^2{\cdot}5^2\ar@{-}[r]\ar@{-}[d]\ar@{-}[ldd]&-2^2{\cdot}5^3\ar@{-}[r]\ar@{-}[d]\ar@{-}[ldd]&-2^2{\cdot}5^4\ar@{-}[r]\ar@{-}[d]\ar@{-}[ldd]&\dots\ar@{-}[ldd]\\
-2^3{\cdot}5\ar@{-}[r]\ar@{-}[d]&-2^3{\cdot}5^2\ar@{-}[r]\ar@{-}[d]&-2^3{\cdot}5^3\ar@{-}[r]\ar@{-}[d]&-2^3{\cdot}5^4\ar@{-}[r]\ar@{-}[d]&\dots\\
\vdots&\vdots&\vdots&\vdots\\
}
$$
\caption{The graph $\Gamma_5$}\label{fig5}
\end{figure}
3. Assume that $p=2^m-1>3$ is a Mersenne prime. In this case $m>2$ and $p$ is not Fermat. It is easy to check that every doubleton $\{x,y\}\in\big\{\{\varepsilon2^ap^b,\varepsilon2^{a-1}p^b\},\{\varepsilon2^{a-1}p^b,\varepsilon2^{m+a-1}p^b\}, \{\varepsilon2^{a-1}p^{b+1},\varepsilon2^{m+a-1}p^b\},$ \break$\{\varepsilon2^{a-1}p^b,-\varepsilon2^{a-1}p^b\}, \{\varepsilon2^{a-1}p^b,-\varepsilon2^{a-1}p^{b+1}\} :a,b\in\IN, \ \varepsilon \in\{-1,1\} \big\}$ has $A_{\{x,y\}}= \Pi_x\cup\Pi_y\cup\Pi_{x-y}=\{2,p\}$ and hence $\{x,y\}\in\mathcal E_p$.
Now assume that $\{x,y\}\in\mathcal E_p$ is an edge of the graph $\Gamma_p$. Then $\Pi_x\cup\Pi_y\cup\Pi_{x-y}=A_{\{x,y\}}=\{2,p\}$ and $\{x,y\}$ can be written as $\{\varepsilon2^{a-1}p^b,\delta2^{c-1}p^d\}$ for some $a,b,c,d\in\IN$, $\varepsilon,\delta \in\{-1,1\}$ with $2^{a-1}p^b\le 2^{c-1}p^d$.
If $a=c$, $b=d$, then $\varepsilon=-\delta$ and $\{x,y\}=\{\varepsilon2^{a-1}p^b,-\varepsilon2^{a-1}p^b\}$.
If $a=c$, then $b\le d$ and the inclusion $\Pi_{x-y}\subseteq\{2,p\}$ implies that $\Pi_{p^{d-b}-\varepsilon/\delta}\subseteq\{2,p\}$ and hence $p^{d-b}-\varepsilon/\delta$ is a power of $2$. By the Mih\u ailescu Theorem~\ref{Mihailescu}, $d-b\in\{0,1\}$. If $d-b=0$, then $\{x,y\}=\{\varepsilon2^{a-1}p^b,-\varepsilon2^{a-1}p^b\}$ by the preceding case. So, we assume that $d-b=1$. If $\varepsilon=\delta$, then $p^{d-b}-\varepsilon/\delta=p-1=2^m-2$ is a power of $2$, which is not true as $m>2$. Therefore $\varepsilon=-\delta$ and $\{x,y\}$ is equal to $\{\varepsilon2^{a-1}p^b,-\varepsilon2^{a-1}p^{b+1}\}$
If $b=d$, then $a\le c$ and the inclusion $\Pi_{x-y}\subseteq\{2,p\}$ implies that $\Pi_{2^{c-a}-\varepsilon/\delta}\subseteq\{2,p\}$ and hence $2^{c-a}-\varepsilon/\delta$ is a power of $p$. By the Mih\u ailescu Theorem~\ref{Mihailescu}, $2^{c-a}-\varepsilon/\delta\in\{1,p\}=\{1,2^m-1\}$, which implies that $\varepsilon=\delta$ and $c-a\in\{1,m\}$. Therefore $\{x,y\}$ is equal to $\{\varepsilon2^{a-1}p^b,\varepsilon2^{a}p^{b}\}$ or $\{\varepsilon2^{a-1}p^b,\varepsilon2^{m+a-1}p^b\}$.
So, we assume that $a\ne c$ and $b\ne d$. By analogy with the case of Fermat primes, we can show that the subcases ($a<c$ and $b<d$) and ($a>c$ and $b>d$) are impossible.
If $a<c$ and $b>d$, then $\Pi_{x-y}\subseteq\{2,p\}$ implies that $2^{c-a}-(\varepsilon/\delta) p^{b-d}=1$, and hence $\varepsilon/\delta=1$. Then the Mih\u ailescu Theorem~\ref{Mihailescu} ensures that $b-d=1$ and hence $2^{c-a}=p+1=2^m$ and $c-a=m$. In this case $\{x,y\}=\{\varepsilon2^{a-1}p^{d+1},\varepsilon2^{a+m-1}p^{d}\}$.
If $a>c$ and $b<d$, then $\Pi_{x-y}\subseteq\{2,p\}$ implies that $p^{d-b}-(\varepsilon/\delta)2^{a-c}=1$ and hence $\varepsilon/\delta=1$. Then Mih\u ailescu Theorem~\ref{Mihailescu} implies that $d-b=1$ and hence $2^{a-c}=p-1=2^m-2$, which is not possible as $m>2$.
\begin{figure}
$$
\xymatrix@C60pt{
\vdots\ar@{-}@/_30pt/[ddddddddd]\ar@{-}[d]\ar@{-}[dddr]\ar@{-}@/^20pt/[ddd]&\ar@{-}@/_30pt/[ddddddddd]\vdots\ar@{-}[d]\ar@{-}@/^20pt/[ddd]\ar@{-}[dddr]&\ar@{-}@/_30pt/[ddddddddd]\vdots\ar@{-}[d]\ar@{-}@/^20pt/[ddd]\ar@{-}[dddr]&\ar@{-}@/_30pt/[ddddddddd]\vdots\ar@{-}[d]\ar@{-}@/^20pt/[ddd]\ar@{-}[dddr]\\
2^3{\cdot}7\ar@{-}[dddddddr]\ar@{-}@/^20pt/[ddd]\ar@{-}[d]\ar@{-}[dddr]\ar@{-}@/_25pt/[ddddddd]&2^3{\cdot}7^2\ar@{-}[dddddddr]\ar@{-}@/^20pt/[ddd]\ar@{-}[d]\ar@{-}[dddr]\ar@{-}@/_25pt/[ddddddd]
&2^3{\cdot}7^3\ar@{-}[dddddddr]\ar@{-}@/^20pt/[ddd]\ar@{-}[d]\ar@{-}[dddr]\ar@{-}@/_25pt/[ddddddd]&2^3{\cdot}7^4\ar@{-}[dddddddr]\ar@{-}@/^20pt/[ddd]\ar@{-}[d]\ar@{-}[dddr]\ar@{-}@/_25pt/[ddddddd]&\dots\ar@{-}\\
2^2{\cdot}7\ar@{-}[dddddr]\ar@{-}@/_20pt/[ddddd]\ar@{-}[d]&\ar@{-}[dddddr]2^2{\cdot}7^2\ar@{-}[dddddr]\ar@{-}@/_20pt/[ddddd]\ar@{-}[d]&2^2{\cdot}7^3\ar@{-}[dddddr]\ar@{-}@/_20pt/[ddddd]\ar@{-}[d]&2^2{\cdot}7^4\ar@{-}[dddddr]\ar@{-}@/_20pt/[ddddd]\ar@{-}[d]&\dots\ar@{-}\\
2{\cdot}7\ar@{-}[dddr]\ar@{-}@/_15pt/[ddd]\ar@{-}[d]&2{\cdot}7^2\ar@{-}[dddr]\ar@{-}@/_15pt/[ddd]\ar@{-}[d]&2{\cdot}7^3\ar@{-}[dddr]\ar@{-}@/_15pt/[ddd]\ar@{-}[d]&2{\cdot}7^4\ar@{-}[dddr]\ar@{-}@/_15pt/[ddd]\ar@{-}[d]&\dots\ar@{-}\\
7\ar@{-}[dr]\ar@{-}[d]&7^2\ar@{-}[dr]\ar@{-}[d]&7^3\ar@{-}[dr]\ar@{-}[d]&7^4\ar@{-}[dr]\ar@{-}[d]&\dots\ar@{-}\\
-7\ar@{-}[ur]\ar@{-}[d]\ar@{-}@/^20pt/[ddd]&-7^2\ar@{-}[ur]\ar@{-}[d]\ar@{-}@/^20pt/[ddd]\ar@{-}[dddl]&-\ar@{-}[ur]7^3\ar@{-}[d]\ar@{-}@/^20pt/[ddd]\ar@{-}[dddl]&-7^4\ar@{-}[ur]\ar@{-}[d]\ar@{-}@/^20pt/[ddd]\ar@{-}[dddl]&\dots\ar@{-}[dddl]\\
-2{\cdot}7\ar@{-}[uuur]\ar@{-}[d]\ar@{-}@/^20pt/[ddd]&-2{\cdot}7^2\ar@{-}[uuur]\ar@{-}[d]\ar@{-}@/^20pt/[ddd]\ar@{-}[dddl]&-2{\cdot}7^3\ar@{-}[uuur]\ar@{-}[d]\ar@{-}@/^20pt/[ddd]\ar@{-}[dddl]&-2{\cdot}7^4\ar@{-}[uuur]\ar@{-}[d]\ar@{-}@/^20pt/[ddd]\ar@{-}[dddl]&\dots\ar@{-}[dddl]\\
-2^2{\cdot}7\ar@{-}[uuuuur]\ar@{-}[d]&-2^2{\cdot}7^2\ar@{-}[uuuuur]\ar@{-}[d]&-2^2{\cdot}7^3\ar@{-}[uuuuur]\ar@{-}[d]&-2^2{\cdot}7^4\ar@{-}[uuuuur]\ar@{-}[d]&\dots\\
-2^3{\cdot}7\ar@{-}[uuuuuuur]\ar@{-}[d]&-2^3{\cdot}7^2\ar@{-}[uuuuuuur]\ar@{-}[d]&-2^3{\cdot}7^3\ar@{-}[uuuuuuur]\ar@{-}[d]&-2^3{\cdot}7^4\ar@{-}[uuuuuuur]\ar@{-}[d]&\dots\\
\vdots&\vdots&\vdots&\vdots&\\
}
$$
\caption{The graph $\Gamma_7$}\label{fig7}
\end{figure}
4. Assume that $p$ is not Fermat--Mersenne. It is easy to check that every doubleton
$$\{x,y\}\in\big\{\{\varepsilon2^{a-1}p^b,-\varepsilon2^{a-1}p^b\}, \{\varepsilon2^{a-1}p^b,\varepsilon2^{a}p^b\}:a,b\in\IN, \ \varepsilon\in\{-1,1\}\big\}$$
has $A_{\{x,y\}}=\Pi_x\cup\Pi_y\cup\Pi_{x-y}=\{2,p\}$ and hence $\{x,y\}\in\mathcal E_p$.
\begin{figure}
$$
\xymatrix@C40pt{
\ar@{-}[d]\ar@{-}@/^40pt/[ddddddddd]\vdots&\ar@{-}[d]\ar@{-}@/^40pt/[ddddddddd]\vdots&\ar@{-}[d]\ar@{-}@/^40pt/[ddddddddd]\vdots&\ar@{-}[d]\ar@{-}@/^40pt/[ddddddddd]\vdots\\
2^3{\cdot}11\ar@{-}[d]\ar@{-}@/^35pt/[ddddddd]&2^2{\cdot}11^3\ar@{-}[d]\ar@{-}@/^35pt/[ddddddd]&2^3{\cdot}11^3\ar@{-}[d]\ar@{-}@/^35pt/[ddddddd]&2^3{\cdot}11^4\ar@{-}[d]\ar@{-}@/^35pt/[ddddddd]&\cdots\\
2^2{\cdot}11\ar@{-}[d]\ar@{-}@/^30pt/[ddddd]&2^2{\cdot}11^2\ar@{-}[d]\ar@{-}@/^30pt/[ddddd]&2^2{\cdot}11^3\ar@{-}[d]\ar@{-}@/^30pt/[ddddd]&2^2{\cdot}11^4\ar@{-}[d]\ar@{-}@/^30pt/[ddddd]&\cdots\\
2{\cdot}11\ar@{-}[d]\ar@{-}@/^25pt/[ddd]&2{\cdot}11^2\ar@{-}[d]\ar@{-}@/^25pt/[ddd]&2{\cdot}11^3\ar@{-}[d]\ar@{-}@/^25pt/[ddd]&2{\cdot}11^4\ar@{-}[d]\ar@{-}@/^25pt/[ddd]&\cdots\\
11\ar@{-}[d]&11^2\ar@{-}[d]&11^3\ar@{-}[d]&11^4\ar@{-}[d]&\cdots\\
-11\ar@{-}[d]&-11^2\ar@{-}[d]&-11^3\ar@{-}[d]&-11^4\ar@{-}[d]&\cdots\\
-2{\cdot}11\ar@{-}[d]&-2{\cdot}11^2\ar@{-}[d]&-2{\cdot}11^3\ar@{-}[d]&-2{\cdot}11^4\ar@{-}[d]&\cdots\\
-2^2{\cdot}11\ar@{-}[d]&-2^2{\cdot}11^2\ar@{-}[d]&-2^2{\cdot}11^3\ar@{-}[d]&-2^2{\cdot}11^4\ar@{-}[d]&\cdots\\
-2^3{\cdot}11\ar@{-}[d]&-2^3{\cdot}11^2\ar@{-}[d]&-2^3{\cdot}11^3\ar@{-}[d]&-2^3{\cdot}11^4\ar@{-}[d]&\cdots\\
\vdots&\vdots&\vdots&\vdots&\\
}
$$
\caption{The graph $\Gamma_{11}$}\label{fig11}
\end{figure}
Now assume that $\{x,y\}\in\mathcal E_p$ is an edge of the graph $\Gamma_p$. Then $\Pi_x\cup\Pi_y\cup\Pi_{x-y}=A_{\{x,y\}}=\{2,p\}$ and $\{x,y\}$ can be written as $\{\varepsilon2^{a-1}p^b,\delta2^{c-1}p^d\}$ for some $a,b,c,d\in\IN$, $\varepsilon,\delta \in\{-1,1\}$ with $2^{a-1}p^b\le 2^{c-1}p^d$.
If $a=c$ and $b=d$, then $\varepsilon\ne\delta$ and $\{x,y\}=\{2^{a-1}p^b,-2^{a-1}p^b\}$.
If $a=c$, then $b\le d$ and the inclusion $\Pi_{x-y}\subseteq\{2,p\}$ implies that $\Pi_{p^{d-b}-\varepsilon/\delta}\subseteq\{2,p\}$ and hence $p^{d-b}-\varepsilon/\delta$ is a power of $2$. By the Mih\u ailescu Theorem~\ref{Mihailescu}, $d-b=1$ and hence $p$ is a Fermat prime or Mersenne prime which is not true.
If $b=d$, then $a\le c$ and the inclusion $\Pi_{x-y}\subseteq\{2,p\}$ implies that $\Pi_{2^{c-a}-\varepsilon/\delta}\subseteq\{2,p\}$ and hence $2^{c-a}-\varepsilon/\delta$ is a power of $p$. By the Mih\u ailescu Theorem~\ref{Mihailescu}, $2^{c-a}-\varepsilon/\delta\in\{1,p\}$. Taking into account that $p$ is not Fermat or Mersenne prime, we conclude that if $\varepsilon=\delta$, $2^{c-a}-1=1$ and hence $c-a=1$. Then $\{x,y\}=\{\varepsilon2^{a-1}p^b,\varepsilon2^{a}p^{b}\}$.
So, we assume that $a\ne c$ and $b\ne d$. By analogy with the case of Fermate primes, we can show that the subcases ($a<c$ and $b<d$) and ($a>c$ and $b>d$) are impossible.
If $a<c$ and $b>d$, then $\Pi_{x-y}\subseteq\{2,p\}$ implies that $2^{c-a}-(\varepsilon/\delta) p^{b-d}=1$. By the Mih\u ailescu Theorem~\ref{Mihailescu} $b-d=1$ and hence $p=2^{c-a}-1$ is a Mersenne prime, which is not true.
If $a>c$ and $b<d$, then $\Pi_{x-y}\subseteq\{2,p\}$ implies that $p^{d-b}-(\varepsilon/\delta)2^{a-c}=1$. By the Mih\u ailescu Theorem~\ref{Mihailescu} $d-b=1$ and hence $p=1+2^{a-c}$ is a Fermat prime, which is not true.
\end{proof}
In Figures~\ref{fig3}, \ref{fig5}, \ref{fig7}, \ref{fig11} we draw the graphs $\Gamma_p$ for $p$ equal to $3,5,7,11$. Observe that $3$ is both Fermat and Mersenne prime, $5$ is Fermat prime, $7$ is Mersenne prime and $11$ is not Fermat--Mersenne.
\begin{lemma}\label{l:graph}
Let $p$ be an odd prime number and $h$ be a positive homeomorphism of the Kirch space.
\begin{enumerate}
\item If $p$ is Fermat-Mersenne, then $h(p)=p$;
\item If $p$ is not Fermat-Mersenne, then $h(p)=p^n$ for some $n\in \IN$.
\end{enumerate}
\end{lemma}
\begin{proof}
1. Lemma~\ref{struct}(1) implies that the degree of $\pm3$ in the graph $\Gamma_3$ is equal to $8$ but the other vertices have degree at least $9$. Hence $h(3)=\pm3$. Assume that $h(3)=-3$. Then by Lemma \ref{l:2fix} and by Lemma \ref{lemmma} $\{2,3\}=A_{\{2,3\}}=A_{h(\{2,3\})}=A_{\{2,-3\}}=\{2,3,5\}$ but this is not true and hence $h(3)=3$.
Assume that $p>3$ is Fermat or Mersenne prime. Lemma~\ref{struct}(2,3) implies that the degree of $\pm p$ in the graph $\Gamma_p$ is $4$ but the other vertices have degree at least $5$. Hence $h(p)=\pm p$. Assume that $h(p)=- p$. By Lemma \ref{l:2fix}, $A_{\{1,p\}}=A_{\{1,h(p)\}}=A_{\{1,-p\}}$, so $\{p\}\cup \Pi_{p-1}=\{p\}\cup\Pi_{p+1}$, according to Lemma \ref{2ae}. This implies that $\Pi_{p-1}=\Pi_{p+1}=\{2\}$. Hence $p$ is both Fermate and Mersenne which is possible iff $p=3$ and this contradicts our assumption. Therefore $h(p)=p$.
2. Let $p$ be an odd prime number, which is not Fermat--Mersenne.
Lemma~\ref{struct}(4) implies that the set $\pm p^\IN=\{\varepsilon p^n:n\in\IN, \varepsilon\in\{-1,1\}\}$ coincides with the set of vertices of order 2 in the graph $\Gamma_p$. Taking into account that $h{\restriction}V_p$ is an isomorphism of the graph $\Gamma_p$, we conclude that $h( p)=\pm p^n$ for some $ n\in \IN$. Assume that $h(p)=-p^n$. Then $h(\{-1,p\})=\{-1,-p^n\}$. By Lemma \ref{lemmma}, $A_{\{-1,p\}}=A_{\{-1,-p^n\}}$, so $\{p\}\cup \Pi_{p+1}=\{p\}\cup\Pi_{p^n-1}$, according to Lemma \ref{2ae}. Since $\{p\}$ does not intersect $\Pi_{p+1}$ and $\Pi_{p^n-1}$ we conclude that $ \Pi_{p+1}=\Pi_{p^n-1}$. Hence we get the inclusion $\Pi_{p-1}\subseteq\Pi_{p^n-1}=\Pi_{p+1}$. If some prime number $d$ divides $p-1$ then the inclusion $\Pi_{p-1}\subseteq\Pi_{p+1}$ implies that $d$ divides ${p+1}$, consequently $d$ divides the difference $(p+1)-(p-1)=2$ and hence $d=2$. As a consequence, $\Pi_{p-1}=\{2\}$ and $p-1=2^m$ for some $m\in \IN$ which contradicts the assumption that $p$ is not Fermat prime. Hence $h( p)= p^n$.
\end{proof}
\begin{lemma}
\label{l:F1x} For any positive homeomorphism $h$ of the Kirch space and any prime number $p$ we have $h(p)=p$.
\end{lemma}
\begin{proof}
If $p=2$, then $h(p)=p$ by Lemma~\ref{l:2fix}. If $p$ is Fermat--Mersenne, then $h(p)=p$ by Lemma~\ref{l:graph}. So, we assume $p$ is not Fermat--Mersenne. By Lemma~\ref{l:graph}, $h(p)=p^n$ for some $n\in\IN$.
By Lemmas~\ref{2ae}, \ref{l:2fix} and \ref{lemmma}, $$\{p\}\cup \Pi_{p-1}=A_{\{1,p\}}=A_{\{1,h(p)\}}=A_{\{1,p^n\}}=\{p\}\cup\Pi_{p^n-1}$$and hence $\Pi_{p^n-1}=\Pi_{p-1}$. Since $p$ is not Mersenne prime, Zsigmondy Theorem~\ref{Zsigmondy} guarantees that $n=1$ and hence $h(p)=p^1=p$
\end{proof
\begin{lemma} The positive homeomorphism group of the Kirch space is trivial.
\end{lemma}
\begin{proof} To derive a contradiction, assume that the Kirch space admits a homeomorphism $h$ such that $h(x)\ne x$ for some number $x$. By the Hausdorff property of the Kirch space and the continuity of $h$, there exists a neighborhood $O_x$ of $x$ in the Kirch topology such that $h[O_x]\cap O_x=\emptyset$. By the definition of the Kirch topology, there exists a square-free number $b$ such that $\Pi_b\cap \Pi_x=\emptyset$ and $x+b\IZ\subseteq O_x$. By the Dirichlet Theorem~\ref{Dirichlet}, the arithmetic progression $x+b\IN\subseteq x+b\IZ$ contains some prime number $p$. Then $h[O_x]\cap O_x=\emptyset$ implies $h(p)\ne p$, which contradicts Lemma \ref{l:F1x}.
\end{proof}
Our final lemma completes the proof of Theorem~\ref{t:main}.
\begin{lemma}Any homeomorphism $h$ of the Kirch space $ \IZ^\bullet$ is equal to $i:\IZ^\bullet\to\IZ^\bullet$, $i:x\mapsto x$ or to $j:\IZ^\bullet\to\IZ^\bullet$$, j:x\mapsto-x$.
\end{lemma}
\begin{proof} If $h$ is positive, then $h=i$ by previous Lemma. If $h$ is not positive then $h(1)<0$ and $j\circ h(1)>0$. Then the homeomorphism $j\circ h$ is positive and equals $i$ by the preceding case. This implies that $$h=i\circ h =(j\circ j)\circ h=j\circ( j \circ h)=j\circ i=j.$$
\end{proof}
\noindent{\bf Acknowledgement.} The author expresses her sincerely thanks to Taras Banakh for his generous help during preparation of this paper.
|
\section{Introduction}
The field of neuromorphic engineering is widely recognized as the realization of Carver Mead's original vision for a new type of electronic hardware engineered to mimic information processing in biological nervous systems \cite{Mead-1990, Mead-2020}. Today, the lowest common denominator of virtually all neuromorphic hardware systems is the co-location of memory and processing units (i.e. non-von Neumann architecture). This minimal neuromorphic feature alone has dramatically improved power efficiency in training various artificial neural network (ANN) models \cite{Zhang-2020}.
A higher level neuromorphic attribute is the ability to learn and while ANN models demonstrate learning in software, learning in hardware is desirable for next-generation stand-alone cognitive devices, especially at the IoT edge \cite{edge}. In hardware, spike-based learning has been successfully implemented in conventional silicon CMOS technology (e.g. \cite{Indiveri-2011,Merolla-2014,Brainscales1,Brainscales2}).
Beyond silicon, nanoelectronic materials with intrinsic neuromorphic properties, including memory and the ability to emulate synaptic connections \cite{Burr-2017,Zhang-2020}, have attracted enormous attention for on-chip learning \cite{Sangwan-2020}. In particular, resistive switching memory (memristive) devices \cite{Waser-Aono-2007,Zidan-2018,Ielmini-Wong-2018,Wang-2020} are leading candidates for efficient neuromorphic computing architectures, with demonstrated neuromorphic learning functionalities such as short-term and long-term potentiation (STP/LTP) and spike-timing dependent plasticity (STDP) \cite{Ohno-2011,Serrano-2013,Serb-2016,Mehonic-2020}.
At the device level, neuromorphic functionalities can be broadly attributed to modification of electronic transport mechanisms by nanoscale geometric confinement, usually across a metal-insulator-metal (MIM) junction \cite{Pershin-DiVentra-2011}. Importantly, synapse-like memristive switching is observed not just in memristors fabricated from conventional bulk materials (e.g. metal oxides), but also in neuromorphic systems self-assembled from nanomaterials using bottom-up techniques \cite{Sangwan-2020}. Here, we focus on self-assembled metallic nanowires because not only do they form memristive switching MIM junctions, but they also form a complex neural-like network topology, with all-in-one connectivity properties such as small-worldness, modularity and recurrent feedback loops \cite{Nirmalraj-2012,Bellew-2014,Avizienis-2012,Demis-2015,Milano-2019,Diaz-Alvarez-2019,Pantone-2018,Loeffler-2020}. The unique neuromorphic topology of self-assembled nanowire networks (NWNs) is responsible for collective functionalities emerging from the interplay between network connectivity and synaptic nonlinear dynamics \cite{Stieg-2012,Sillin-2013,Manning-2018,Kuncic-2018,Diaz-Alvarez-2019,Milano-2020}.
Learning in NWN hardware does not require implementation of an ANN model, as has been demonstrated with associative memory tasks \cite{Diaz-Alvarez-2020,Li-2020} and with temporal information processing tasks using a reservoir computing approach, where the network self-regulates in response to continuous-time input signals and only the readout is trained \cite{Sillin-2013,Fu-2020,Kuncic-2020}.
Varying spatio-temporal input signals (i.e. delivered via different contact electrodes and with time-varying amplitudes) results in the formation of new electrical pathways, analogous to synaptogenetic learning \cite{Zito-2002,Cui-2016}.
Here, we show that NWNs with prior experience of a complex, nonlinear time-series signal can perform better in forecasting the signal than a NWN without prior exposure, thus demonstrating capacity for transfer learning, an important attribute for general intelligence (see \cite{Zhuang-2020} for a recent comprehensive review).
\section{Methods}
\subsection{Modelling network connectivity and memristive junctions}
We performed simulations using a physically motivated model based on polymer-coated Ag nanowires that self--assemble into a complex network \cite{Kuncic-2018,Diaz-Alvarez-2019}. Self--assembly was modelled by distributing individual nanowires on a 2D plane, with uniformly random positions and orientations, and with lengths uniformly sampled from a gamma distribution (mean 100\,$\mu$m, stdev 10\,$\mu$m). The variance in nanowire length is based on experimental observations \cite{Nirmalraj-2012,Bellew-2014,Diaz-Alvarez-2019,Diaz-Alvarez-2020} and increases the probability of forming cross-point junctions between overlapping nanowires. This mimics biological neural networks, in which individual neurons can each make several thousand synaptic connections to neighbouring neurons. In our model of self-assembled nanowire networks (NWNs), a range of nanowire connectivities are possible for a fixed number of nanowires. Importantly, the resulting network structure is more complex than a purely random topology or fully connected network (Fig.~\ref{fig:networks}), with sparseness and recurrence characteristics that are responsible for efficient signal transduction and emergent cognitive function in biological neural networks \cite{Bullmore-Sporns-2009,Lynn-Bassett-2019,Loeffler-2020}. It is also noteworthy that the complex network topology of self-assembled networks differs from the bipartite structure used in ANN models.
\begin{figure}
\centering
\includegraphics[width=1.7in]{NWN}
\includegraphics[width=1.7in]{random}
\caption{Graph representations of 300-node networks: left -- self--assembled nanowire network (2434 edge junctions, average degree 16, small-world propensity 0.67); right -- random network (2400 junctions, average degree 16, small-world propensity 0.29). Nodes in red, edges in black.}
\label{fig:networks}
\end{figure}
Nanowire-nanowire cross-points were modelled as voltage-controlled memristive junctions described by a state-dependent Ohm's law, $I = G (\lambda ) V$, where the conductance $G(\lambda)$ is a function of the state variable $\lambda (t)$ that depends on the past history of voltage input. Physically, $\lambda (t)$ parameterizes the evolution of a conductive filament that forms across the MIM junction above a threshold bias. For polymer-coated Ag nanowires, the polymer is electrically insulating, but ionically conducting, so Ag$^{+}$ cations can migrate across the biased junction \cite{Zhu-2020}. The conducting atomic filament that forms in this way switches the junction from a high-resistance ``off'' state, to a low-resistance ``on'' state, when $\lambda \geq \lambda_{\rm crit}$, where $\lambda_{\rm crit}$ is a threshold. As the polymer thickness ($\approx 2-3\,$nm) is comparable to the Fermi length of Ag ($\approx 0.5\,$nm), resistive switching is modelled as a change in the junction conductance state $G(\lambda)$ by an amount equal to the conductance quanta $G_0 = (13\,{\rm k}\Omega)^{-1}$, consistent with measurements of individual nanowire junctions \cite{Terabe-2005}. The corresponding resistance states are $R_{\rm on} = G_0^{-1}$ and $R_{\rm off} = \zeta R_{\rm on}$, with $\zeta = 10^4$ used in the simulation results presented here. Additionally, the Simmons formula is used to model the low voltage tunneling regime in $G(\lambda)$ when the conductive filament is close to the opposite nanowire \cite{Simmons-1963}. Network conductance is calculated using a modified nodal analysis \cite{Ho-1975} to solve Kirchoff's circuit law equations at each time point.
\subsection{Mackey--Glass time series prediction}
Reservoir computing was implemented on a network with $N=100$ nanowire nodes and 577 memristive junctions. The Mackey--Glass (MG) signal was delivered to one source node as input voltage bias relative to a drain node. MG signals with varying time delays $\tau \geq 17$ were predicted, with $\tau = 17$ corresponding to the onset of chaotic dynamics.
The network state as a function of time $t$ is represented by instantaneous voltage on all the $N$ nodes. The MG signal at a future time $u_{t+\delta t}$ was predicted using a subset of $n=10$ node states weighted by a vector $\mbox{\boldmath$w$}$:
\begin{equation}
u_{t+\delta t} = \mbox{\boldmath$w$} \cdot \mbox{\boldmath$\eta$}_t,
\label{e:predict}
\end{equation}
where $\mbox{\boldmath$\eta$}_t$ is an 11-element vector that includes a 1\,V linear shift element and where $\mbox{\boldmath$w$}$ was determined by least squares regression using all past states of the $n$ nodes and the input (teacher) signal in the time interval $t \in [0,T]$, i.e.
\begin{equation}
[u_{\delta t}, u_{\delta t+1}, ... , u_{T}]
= \mbox{\boldmath$w$} \cdot [\mbox{\boldmath$\eta$}_0, \mbox{\boldmath$\eta$}_1, ... \mbox{\boldmath$\eta$}_{T - \delta t}].
\end{equation}
A history length of $T > \tau$ was used to train the $n$ output weights and the prediction step was set to $\delta t = \tau$.
Accuracy of the prediction task was calculated as
\begin{equation}
\mbox{Accuracy} = 1 - \mbox{RNMSE}
\end{equation}
where RNMSE is the root-normalized mean square error.
Statistical uncertainties were determined by randomly selecting the $n=10$ readout nodes for 100 simulations and averaging the accuracy.
\subsection{Transfer learning}
In conventional reservoir computing, the initial state of the network is homogeneous and for the MG prediction task described above, we set $\mbox{\boldmath$\eta$}_0 =$ \boldmath{$0$}. We modified the task by first exposing the network to a source MG signal with delay parameter $\tau_1$ before training and predicting a second target MG signal with delay parameter $\tau_2$. Delivering the source MG ($\tau_1$) signal for 1.5\,s effectively primed the network to an initial state $\mbox{\boldmath$\eta$}_0^\prime \neq$ \boldmath{$0$} that has memory of previous states associated with the source MG signal. This is analogous to transfer learning methods applied to ANN models, where synaptic weights are trained on a source domain and the knowledge gained is transferred to a different, but related, target domain \cite{Zhuang-2020}. Our case is somewhat different as the network dynamically self-adjusts its own synaptic junction states during the priming period (since in reservoir computing, only the output weights are trained, not the network weights).
We compared the accuracy in predicting the MG target signal to that obtained for a network without prior exposure to the source MG signal during a pre-training period.
\section{Results}
\subsection{Adaptive dynamics}
\begin{figure}[h]
\centering
\includegraphics[width=3.3in]{Gjcn.png}
\includegraphics[width=1.7in]{snapshot_t0.1}
\includegraphics[width=1.7in]{snapshot_t0.4}
\includegraphics[width=1.7in]{snapshot_t0.5}
\includegraphics[width=1.7in]{snapshot_t0.7}
\caption{Top panel -- Individual memristive junction conductances $G_{\rm jn}$ (in units of conductance quanta $G_0$) as a function of time $t$ (in units of total simulation time $T$) for a triangular voltage signal (black) input to a 261-junction NWN. Bottom panel -- snapshots of the network at four sequential time points ($t/T = 0.1, 0.4, 0.5, 0.7$) with colorbar indicating $G_{\rm jn}$. Dark blue junctions denote memristive switches in the off state.}
\label{fig:adaptive}
\end{figure}
Figure~\ref{fig:adaptive} shows the NWN response to a triangular input signal. In the top panel, each colored curve represents the evolution in time of conductance $G_{\rm jn}$ across an individual memristive junction. The network connectivity determines the spatial distribution of voltage at each moment in time. This connectivity influences the voltage-controled memristive dynamics of each junction, resulting in collective switching as $G_{\rm jn}$ continuously adapts. The bottom panel shows this self-regulation of the synaptic junctions in snapshot visualizations of the network at successive time points during evolution. Brightly colored circles evident in the frames at $t = 0.4T$ and $t=0.5T$ represent memristive switches in their on state, with current paths indicated (white). The intrinsic adaptive dynamics of NWNs can in principle be harnessed for information processing. For the parameters used in Fig.~\ref{fig:adaptive} ($f = 0.75\,$Hz. $A = 0.8\,$V), the network exhibits ``edge-of'chaos'' dynamics (e.g. $I-V$ trajectories begin to diverge), which may be optimal for information processing \cite{edge-chaos}.
\subsection{Mackey--Glass time--series prediction}
Fig.~\ref{fig:MG-prediction} plots the time series for training and predicting a MG signal with $\tau = 20$. Network output weights $\mbox{\boldmath$w$}$ are trained using the first 2400 time steps (i.e. $T = 2.4\,$s) after which the signal is predicted 20 steps ahead using eq.~(\ref{e:predict}) (i.e. with $\delta t =0.02\,$s in this case). The target signal is overplotted for comparison. The resulting prediction accuracy is 75\%.
Fig.~\ref{fig:MG-tau} plots the prediction accuracy as a function of $\tau \geq 17$.
Accuracy decreases with $\tau$ because errors amplify exponentially as the MG signal becomes more chaotic.
\begin{figure}[h]
\centering
\includegraphics[width=3.5in]{MG-predict.png}
\caption{Time series of MG source signal during training (gray), followed by target (dashed blue) and predicted (red) signals for $\tau = 20$. The inset shows a zoom-in of part of the prediction period.}
\label{fig:MG-prediction}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=3.0in]{MG-acc}
\caption{Average MG forecasting accuracy as a function of delay parameter $\tau$. Shading indicates standard error.}
\label{fig:MG-tau}
\end{figure}
\subsection{Transfer learning}
Figure~\ref{fig:MG-transfer} plots MG prediction accuracy when the network is primed by a MG signal prior to training. Accuracy is plotted for three different pre-training MG signals ($\tau_1 = 20, 40, 150$) as a function of $\tau_2$ used for MG signal training. For comparison, prediction accuracy without the pre-training (cf. Fig.~\ref{fig:MG-tau}) is also overplotted.
Accuracy improves when the network is first primed with a MG signal. This demonstrates the principle of transfer learning, where knowledge is extracted from a source domain and then leveraged for learning in a related target domain.
\begin{figure}[h]
\centering
\includegraphics[width=3.4in]{MG-transfer}
\caption{Average MG signal prediction accuracy as a function of signal delay parameter $\tau_2$ of the predicted signal and different $\tau_1$ signals used to prime the network before training. Average accuracy without pre-training is shown for comparison (cf. Fig.~\ref{fig:MG-tau}). Shading indicates standard error.}
\label{fig:MG-transfer}
\end{figure}
Learning performance is expected to improve especially when there is insufficient information in the target domain compared to the source domain. In this example, prediction accuracy improves more when the network is primed by a source MG signal that is more chaotic (i.e. more degrees of freedom) than the target MG signal (i.e. $\tau_1 > \tau_2$). This is shown by the blue curve (for $\tau_1 = 150$) in Fig.~\ref{fig:MG-transfer} and by the accuracy difference heatmap in Fig.~\ref{fig:MG-transfer-diff}.
\begin{figure}[h]
\centering
\includegraphics[width=3.4in]{MG-transfer-diff}
\caption{Heatmap showing change in average accuracy in predicting a MG signal with $\tau_2$ when the network is primed using a MG signal with $\tau_1$ relative to prediction without priming.}
\label{fig:MG-transfer-diff}
\end{figure}
Importantly, the target MG signal is predicted without relying on any teacher signal for recall. This suggests learning is achieved by harnessing the network's collective memory of past dynamical states. Priming the network before training improves learning by strengthening the memristive connections in an adaptive way, enabling longer-term memory consolidation.
Prediction accuracy also depends on the instantaneous network state selected for priming. Regardless of the value of $\tau_2$ of the MG signal being predicted, we find accuracy is optimized for a small range of primed network states. This optimal range of states occurs around network activation, coinciding with the formation of a winner-takes-all (WTA) current path (cf. Fig.~\ref{fig:adaptive}, bottom panel). Such WTA gate modules in network circuits are purported to have universal computational power for both digital and analog information processing \cite{Maass-2000}.
\section{Conclusions}
We have demonstrated that the complex interplay between the neural network-like circuitry of nanowire networks and their memristive junctions results in adaptive dynamics, where the network self-regulates to find the optimal signal transduction routes.
We showed how the adaptive dynamics can be harnessed for signal processing using a reservoir computing implementation.
Prediction of the highly nonlinear Mackey--Glass signal was demonstrated well into the strongly chaotic regime. This has not previously been demonstrated with other memristive reservoir computing approaches.
Moreover, we found performance accuracy of this task is improved by transfer learning, where the network is primed by a Mackey--Glass signal before training. Our results show that transfer learning improves performance the most when pre-training with a source signal that is more complex than the target signal to be predicted.
\section*{Acknowledgment}
The authors acknowledge use of the Artemis High Performance Computing resource at the Sydney Informatics Hub, a Core Research Facility of the University of Sydney.
\clearpage
|
\section{INTRODUCTION}
Simultaneous localization and mapping (SLAM) is a key technique in autonomous robots with growing attention from academia and industry. The advent of low-cost sensors accelerated significantly the SLAM development. For example, cameras provided rich angular and color information and enabled the development of visual SLAM such as ORB-SLAM \cite{mur2015orb}. Following that the inertial measurement units (IMU) start to drop price thanks to their massive adoption in smart phone industry, and the utilization in SLAM becomes straightforward to gain additional modality and performance such as in VINS-mono \cite{qin2018vins} and ORB-SLAM3 \cite{campos2020orb}. Among the additional sensors, depth sensors (stereo camera, RGB-D camera) provide direct depth measurement and enables accurate performance in SLAM applications such as ORB-SLAM2 \cite{mur2017orb}. lidar, as the high-end depth sensor, provides long range outdoor capability, accurate measurements and system robustness, and has been widely adopted in more demanding applications such as autonomous driving \cite{urmson2008autonomous}, but also typically come with a hefty price tag. As the autonomous industry progresses, recently many new technology developments have enabled commercialization of low-cost lidars, e.g. Ouster and Livox lidars. Featuring a non-repeating scanning pattern, Livox lidars pose unique advantageous in low-cost lidar-assisted SLAM system. We in this paper present the first Livox lidar assisted visual SLAM system (CamVox) with superior and real-time performance. Our CamVox SLAM built upon the state-of-the-art ORB-SLAM2 by using Livox lidar as the depth sensor, with the following new features:
\begin{enumerate}
\item A preprocessing step fusing lidar and camera input. Careful time synchronization is performed that the distortions in the non-repeating scanned lidar points are corrected by the IMU data and are transformed to camera frames.
\item The accuracy and range of the lidar point cloud is superior compared to other depth cameras, and the proposed SLAM system can perform large-scale mapping with higher efficiency and can robustly operate in an outdoor strong-sunlight environment.
\item We utilized the non-repeating scanning feature of Livox lidar to perform automatic calibrations between the camera and lidar at uncontrolled scenes.
\item CamVox performance is evaluated against some main stream frameworks and exhibits very accurate trajectory results. We also open-sourced the hardware, code and dataset of this work hoping to provide an out-of-the-box CamVox (Camera + Livox lidar) SLAM solution.
\end{enumerate}
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{example.pdf}
\caption{Example of CamVox performance. (a) the trajectory of the robot from CamVox; (b) the dense RGBD map from CamVox; (c) a magnified and rotated part of (b) demonstrating high point cloud quality.}
\label{fig:example}
\end{figure}
Fig. \ref{fig:example} shows an example trajectory and map construction from CamVox. The scale of the map is about 200 meters on the long side. The rest of this paper is structured as the following. Related work is reviewed in Section II. The CamVox system hardware and software framework is described in detail in Section III. The results and evaluation are presented in Section IV. Finally, we conclude and remark the outlook in Section V.
\section{Related work}
Due to the rich angular resolution and informative color information, cameras could provide surprisingly good localization and mapping performance through a simple projection model and bundle adjustment. One of the most famous SLAM systems with monocular camera is ORB-SLAM \cite{mur2015orb}. ORB-SLAM tracks the object by extract the ORB features in the image and use loop-closure detection to optimize the map and pose globally, which is usually fast and stable. However, it cannot accurately recover the real scale factor since an absolute depth scale is unknown to a camera.
To fix the issue that monocular camera doesn’t recover real-world scale, Mur et al. proposed an improvement of ORB-SLAM named ORB-SLAM2 \cite{mur2017orb}, adding the support of stereo camera and RGBD camera for depth estimation. However, there are drawbacks with both of these cameras, especially in estimating the outdoor objects with long depths. The stereo camera requires a long baseline for accurate long-depth estimation, which is usually limited in real world scenarios. Additionally, the calibration between the two cameras is susceptible to mechanical changes and will adversely influence the long-depth estimation accuracy. The RGBD camera is usually susceptible to sun light with a finite range of less than 10 meters typically.
Fusing camera and IMU is another common solution, because camera can partially correct IMU integral drift, calibrate IMU bias while IMU can overcome the scale ambiguity of monocular system. Qin et al. proposed Visual-Inertial Monocular SLAM (VINS) \cite{qin2018vins}, which is an excellent camera and IMU fusion system. Similarly, Campos et al. extended ORB-SLAM2 by fusing camera and IMU measurement and proposed ORB-SLAM3 \cite{campos2020orb}. However, consumer-grade IMU only works well in relatively low precision and suffers from bias, noise and drift while high-end IMU is prohibitively costly.
Lidar on the other hand, provides a direct spatial measurement. Lidar SLAM framework has been developed. One pioneering work is LOAM \cite{zhang2014loam}. Comparing to visual SLAM, it is more robust in a dynamic environment, due to the accurate depth estimation from lidar point cloud. However, due to the sophisticated scanning mechanism, lidar cannot provide as much angular resolution as camera does, so that it is prone to fail in environments with less prominent structures like tunnels or hallways. It also lacks loop-closure detection which make the algorithm focus only on local accuracy without global pose and map optimization. To add loop-closure detection in LOAM, Shan et al. \cite{shan2018lego} proposed an enhanced LOAM algorithm LeGO-LOAM. Comparing to LOAM, LeGO-LOAM improves feature extraction with segmentation and clustering for efficiency improvement, and adds loop-closure detection for long run drift reduction.
Combining lidar and camera in a SLAM framework become an ideal solution. While obtaining the point cloud with accurate depth information, it could make use of the high angular and color information from camera. Zhang et al. proposed VLOAM \cite{zhang2015visual}, which fuse monocular camera and lidar loosely. Similar to LOAM, the estimation is claimed to be accurate enough and no loop-closure is needed. Shin et al. also tried to combine monocular camera and lidar together using direct method rather than feature points to estimate the pose. In addition, it tightly couples the visual data and point clouds, and output the estimated pose. Shao et al. \cite{shao2019stereo} went further fusing the stereo camera, IMU and lidar together. They demonstrated a good performance in outdoor real-scale scenes taking advantage of stereo Visual-Inertial Odometry (VIO) loop closure. VIO and lidar mapping are loosely coupled without further optimization at back-end. It is also limited by its complexity and cost.
Lidar was typically too costly to be useful in practical applications. Fortunately, Livox unveiled a new type of lidars based on prism scanning \cite{liu2020low}. Due to the new scanning method, the cost can be significantly lowered to enable massive adoption. Furthermore, this new scanning method allows non-repeating scanning patterns\footnote{\url{https://www.livoxtech.com/3296f540ecf5458a8829e01cf429798e/assets/horizon/04.mp4}}, ideal for acquiring a relatively high-definition point cloud when accumulated (Fig. \ref{fig:fov}). Even for 100 ms accumulation, the density of Livox Horizon is already as high as 64 lines and continue to increase. This feature can be extremely beneficial in calibrating the lidar and camera, where the traditional multiline lidar lacks the precision for space between the lines. The prism design also features a maximal optical aperture for signal reception and allows long range detection. For example, the Livox Horizon could detect up to 260 m under strong sunlight\footnote{\url{https://www.livoxtech.com/horizon/specs}}. With such superior cost and performance, Lin et al. proposed Livox-LOAM \cite{lin2020loam}, which is an enhancement of LOAM adapted to Livox Mid. Based on this, Livox also released a LOAM framework for Livox Horizon, named livox\_horizon\_loam\footnote{\url{https://github.com/Livox-SDK/livox_horizon_loam}}.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{livox_coverage.pdf}
\caption{Point cloud density of three Livox lidar models compared to traditional lidars as a function of integration time.}
\label{fig:fov}
\end{figure}
Because of the long range detection and high accuracy, extrinsic parameter calibration between the camera and the lidar become a more important consideration. In \cite{levinson2013automatic}, the proposed solutions to the lidar-camera can be clustered in two ways. The first one is whether the calibration process needs a calibration target, while the second is whether the calibration can work without human intervention. During these years, many calibration techniques are based on fixed calibration target or manual effort, like \cite{unnikrishnan2005fast} and \cite{geiger2012automatic}. In \cite{pandey2012automatic}, Pandey et al. use Cramer-Rao-Lower-Bound (CRLB) to prove the existence of calibration parameters and estimate them by calculating the minimum variance unbiased (MVUB) estimator. Iyer et al. proposed a network called CalibNet, a geometrically supervised deep network to estimate the transformation between lidar and camera in \cite{iyer2018calibnet}. No specific scene is required for the above two methods. In addition, Levinson et al. proposed an online calibration method in \cite{levinson2013automatic}, in which they claimed that such a method can calibrate the lidar and camera in real time and it was suitable in any scenes. But so far, the calibration still remains as a challenge task and there is no open-source algorithm for calibrating lidar and camera in uncontrolled scene. Livox lidar’s non-repeating scanning pattern could provided a much easier solution as we will demonstrate.
\section{CamVox framework}
The proposed CamVox is based on ORB-SLAM2 (RGBD model) with separate RGBD input preprocessing and automatic calibration methods at uncontrolled scenes. The framework utilizes lidar-assisted visual keyframes to generate local mapping, and exhibits high robustness thanks to the back-end lightweight pose-graph optimization at various levels of bundle adjustment (BA) and loop closure from ORB-SLAM2.
In the original ORB-SLAM2, keypoints are classified in two categories, close and far, where close points are those with high certainty in depth and can be used for scale, translation and rotation estimations while the far points are only used for rotation estimation and hence less informative. With the dense, long range and accurate points obtained from Livox lidars fused with camera image, many more close points could be assigned than traditional RGBD cameras (due to detection range) or stereo vision cameras (due to limited baseline). As a result, the advantages from both the camera (high angular resolution for ORB detection and tracking) and lidar (long range and accurate depth measurement) can be exploited in a tightly coupled manner.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{platform.pdf}
\caption{(a) The complete robot platform. CamVox hardware is mounted on top of this robot. An additional RGBD camera is mounted for comparison. (b) CamVox hardware close-up including a camera, Livox Horizon, and IMU. Additional GPS/RTK is used for ground truth estimation. (c-d) an example of acquired RGB image, and depth image from lidar point cloud (colored in depth).}
\label{fig:platform}
\end{figure}
\subsection{Hardware and software}
The CamVox hardware includes a MV-CE060-10UC rolling shutter camera, a Livox Horizon lidar and an IMU (Inertial Sense $\mu$INS). Additional GPS-RTK (Inertial Sense $\mu$INS) is used for ground truth estimation. Hard synchronization is performed with all of these sensors by a trigger signal of 10 Hz. The camera output at each trigger signal (10 Hz). The lidar keeps a clock (synced with GPS-RTK) and continuously outputs the scanned point with an accurate timestamp. In the meantime, the IMU outputs at a frequency of 200 Hz synced with the trigger. Data from the GPS-RTK is also recorded for ground truth comparison. An Intel Realsense D435 RGBD camera is mounted for comparison. The whole system mounts on a moving robot platform (Agile X Scout mini). It is noted that the sale price of the Livox Horizon lidar (800 USD) is significantly lower than other similar performance lidars (10k – 80k USD) and this allows building the complete hardware system within a reasonable budget.
The software pipeline runs on several parallel threads as shown in Fig. \ref{fig:overview}. In addition to the major threads from ORB-SLAM2, an additional RGBD input preprocessing thread is added to capture data from synchronized camera and lidar (IMU corrected) and process them into a unified RGBD frame. An automatic calibration thread can be triggered to calibrate the camera and lidar, which happens automatically when the robot is detected not moving or by human interaction. The calibrated result is then evaluated and output for potential parameter update.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{CamVox_flow.pdf}
\caption{CamVox SLAM pipeline. In addition to the ORB-SLAM2 major threads, a RGBD input preprocessing thread is used to convert lidar and camera data to the RGBD format, and an automatic calibration thread can be automatically/manually triggered for camera and lidar extrinsic calibration, which is shown in Section C.}
\label{fig:overview}
\end{figure}
\subsection{Preprocessing}
The preprocessing thread takes the raw points from the lidar, corrected by the IMU and projected into a depth image according to the extrinsic calibration with camera. The RGB image is then combined with the depth image as the output of the RGBD frame, where the two images are formatted with equal size and pixel-wise corresponded as shown in Fig. \ref{fig:platform}c and Fig. \ref{fig:platform}d. Further tracking thread operations such as ORB feature extraction, keypoints generation are then performed based on the output.
Since the lidar continuously scan the environment, each data point is obtained at a slightly different time stamp and needs correction from IMU. This is different from a camera, whereas an image is obtained at almost an instant. Fig. \ref{fig:distortion_correction} shows such an example when the point cloud is distorted while the robot is in a continuous motion. To correct this distortion, the robot motion is interpolated from the IMU pose at each lidar point output time, and transforms the lidar point to the lidar coordinate at the time when trigger signal starts, and that is also the time the camera image is captured. This transformation can be described by Eq. \eqref{eq:tsi} and Eq. \eqref{eq:psi}.
\begin{equation}\label{eq:tsi}
T_{si} = \frac{t_i-t_s}{t_e-t_s}T_{se}
\end{equation}
\begin{equation}\label{eq:psi}
P_i^s=T_{si}P_i^i
\end{equation}
Where $t_s$ and $t_e$ are starting and ending timestamp of a lidar frame respectively. The timestamp of any 3D point in a lidar frame can be represented as $t_i$. According to the robot pose transformation matrix $T_{se}$ between $t_s$ and $t_e$ measured by IMU, $T_{si}$ between $t_s$ and $t_i$ can be calculated in \eqref{eq:tsi}. Finally, all lidar points are converted to the lidar coordinate system at $t_s$, which is shown in \eqref{eq:psi}.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{distortion.pdf}
\caption{Example of motion distortion correction by IMU. (a) the RGB image of a corner scene; (b) the raw point cloud during a scan while the robot is moving; (c) the distortion corrected point cloud after processing with IMU data.}
\label{fig:distortion_correction}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{comparsion.pdf}
\caption{Student dormitory at SUSTech, captured by (a) RGB camera, (b) Realsense D435 and (c) CamVox RGBD output. The blue rectangles show the position of the coconut tree leaf at the three pictures. It is clear that typical RGBD camera are much worse than the CamVox RGBD output in detection distance and resistance to sunlight noise.}
\label{fig:comparsion}
\end{figure}
Furthermore, with the help of long distance Livox lidar, we were able to detect reliably many depth-associated camera feature points beyond 100 meters. In comparison, the Realsense RGBD camera was not able to detect points beyond 10 meters and suffer from sunlight noises. This is clearly demonstrated in Fig. \ref{fig:comparsion}. As a result, in CamVox we can specify the close keypoints to those points with associated depth less than 130 meters. This is far more than the 40 times (ORB-SLAM2) the stereo baseline, which is on the order of 10 cm typically used in commercial stereo cameras.
\subsection{Calibration}
Calibration accuracy is vitally important in CamVox due to the long-range capability of the lidar. A small angular mismatch could result in a large absolute deviation at a large depth. Controlled calibration target such as checkerboards are not always available in the field and misalignment could happen after a random fixation failure or a collision. An automatic calibration method needs to be developed at an uncontrolled scene and update the parameters if a better calibration match is found. Thanks to the non-repeating nature of Livox lidars, as long as we could accumulate a few seconds of scanning points, the depth image could become as high resolution as a camera image (Fig. \ref{fig:correction_process}) and the correspondence to the camera image becomes easy to find. Therefore, we are able to do this calibration at almost all field scenes based on the scene information automatically. The triggering of this automatic calibration is set when the robot is detected to be stationary in order to eliminate the motion blur. We accumulate lidar points for a few seconds while remaining still. Camera image is also captured.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{calibration_process.pdf}
\caption{Calibration thread pipeline with examples. The captured data from camera and lidar (remaining still and accumulate for a few seconds) are processed with initial calibration parameters to form three images (grayscale, reflectivity and depth, the latter two are from lidar). Calibration is performed on the edges detected from these images until a satisfactory set of parameters is obtained.}
\label{fig:correction_process}
\end{figure}
The overall calibration algorithm is structured as in Fig. \ref{fig:correction_process}. The dense point cloud is first projected onto the imaging plane by initial external parameters using both reflectivity and depth values, and contour extractions are then performed to compare with the camera image contour. The cost function is constructed by an improved ICP algorithm, which can be optimized by Ceres \cite{ceres-solver} to get the relatively more accurate external calibration parameters. Cost function from these new parameters is then evaluated against the previous values and a decision is made whether to update the extrinsic calibration parameter at input preprocessing thread.
Suppose the coordinate value of a point in the lidar coordinate system is $X=(x,y,z,1)^T$, the z coordiante value of a point in the camera coordinate system is $Z_{c}$ and the pixel position of the point in 2D image is $Y=(u,v,1)^T$. Given an initial external transform matrix from lidar to camera $T_{lidar}^{cam}$ and camera’s intrinsic parameter $(f_u,f_v,c_u,c_v)$, we can project the 3D point cloud to a 2D image by Eq. \eqref{eq:Prect} and Eq. \eqref{eq:Y}.
\begin{equation}\label{eq:Prect}
P_{rect}=\left(\begin{array}{cccc}
f_{u} & 0 & c_{u} & 0 \\
0 & f_{v} & c_{v} & 0 \\
0 & 0 & 1 & 0
\end{array}\right)
\end{equation}
\begin{equation}\label{eq:Y}
Z_{c}Y = P_{rect}T_{lidar}^{cam}X
\end{equation}
After projecting lidar point cloud to 2D image, we do histogram equalization on all images and extract the edge using Canny edge detector \cite{canny1986computational}. The edges extracted from the depth image and the reflectivity image are combined because they are both from the same lidar but separate information. To extract more reliable edges, the following two kinds of edges are filtered out on those edge images. The first kind is the edge that is less than 200 pixels in length. The second kind of edge are the interior ones that are cluttered together. Finally, some characteristic edges that are present in both camera image and lidar image are obtained and edge matching are performed according to the nearest edge distance.
An initial matching result is shown at lower right of Fig. \ref{fig:correction_process}, where the orange line is the edge of the camera image, the blue line is the edge of the Horizon image, and the red line is the distance between the nearest points. Here we adopted the ICP algorithm \cite{besl1992method} and use K-D tree to speed up the search for the nearest neighbor point. However, sometimes in a wrong match very few points actually participated in the calculation of distance, the value of the cost function is trapped inside this local minimum. In this case, we improve the cost function in ICP by adding a degree of mismatching. The improved cost function is shown in Eq. \eqref{eq:CF}, where n is number of camera edge points that are within distance threshold with lidar edge points, m is the number of nearest neighbors, N is number of all camera points, b is weighing factor. We found a value of 10 for b is a good candidate to start as the default value. Note here that the cost function is an averaged value for each point and thus can be used to compare horizontally at different scenes.
\begin{equation}\label{eq:CF}
CF = \sum_{i=1}^{n} \sum_{i=1}^{m} \frac{\text{Distance}(P_{i}^{cam}, P_{ik}^{lidar})}{n\times m} + b \times \frac{N-n}{N}
\end{equation}
In optimizing this cost function, we adopted coordinate descent algorithms \cite{wright2015coordinate} and iteratively optimized (roll, pitch, yaw) coordinates by Ceres \cite{ceres-solver}. This seems to result in a better convergence.
\section{Results}
In this section we present the evaluation results of CamVox. Specifically, we will first show the results of the automatic calibration. The effect of choosing depth threshold for close keypoints is also evaluated. Finally we evaluate the trajectory of CamVox as compared to some of the main stream SLAM frameworks and give a time analysis.
\subsection{Automatic calibration results}
Our automatic calibration result is shown in Fig. \ref{fig:correction_result}. Shown in Fig. \ref{fig:correction_result}a is an overlay of lidar points onto the RGB image when the sensor and the camera are not calibrated (with a misalignment of more than 2 degrees). The cost function has a value of 7.95. The automatic calibration is triggered and calibrates the result as shown in Fig. \ref{fig:correction_result}(b), where the cost function value is 6.11, while the best possible calibration result done manually is shown in Fig. \ref{fig:correction_result}(c) with value 5.88. The automatic calibration delivers a very close result to the best manual calibration. Additionally, thanks to the image-like calibration scheme, the automatic calibration work robustly on most uncontrolled scenes. Several different scenes are evaluated in Fig. \ref{fig:correction_result}(d-f) and cost values are obtained around or below 6, which we regard as a relatively good value.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{calibration_result.pdf}
\caption{An example of RGB camera and point cloud overlay after calibration. (a) not calibrated. (b) automatically calibrated. (c) the best manual calibration. The automatic calibration algorithms is verified at various scenes, (d) outdoor road with natural trees and grasses, (e) outdoor artificial structures, (f) indoor underexposed structures.}
\label{fig:correction_result}
\end{figure}
\subsection{Evaluation of depth threshold for keypoints}
Because the lidar could detect 260 meters, there are many keypoints in the fused frame that we could characterize as close. These points help significantly in tracking and mapping. From Fig. \ref{fig:evalution}(a-d), by setting the close keypoints depth threshold from 20m to 130m, we see a significant increase in both mapping scale and number of mapped features. In Fig. \ref{fig:evalution}e We evaluated the number of matching points tracked as a function of time in the first 100 frames (10 FPS) after starting CamVox. An increase of feature numbers is observed as more frames is captured (Fig. \ref{fig:evalution}f 0.5 s after start), and the larger threshold obviously tracked more features initially that is helpful in starting a robust localization and mapping. With CamVox, we recommend and set the default value to 130 meters.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{evaluation.pdf}
\caption{Evaluation of close keypoints threshold. (a-d) The reconstructed point cloud map by selecting different values of close keypoints threshold. (e) the number of matching points as a function of frame sequence (10 FPS) from start. (f) the number of matching points as a function of close keypoints threshold, evaluated at the beginning (5th frame) of SLAM process.}
\label{fig:evalution}
\end{figure}
\subsection{Comparison of trajectory}
The comparisons of the trajectories from CamVox, two mainstream SLAM framework and the ground truth are evaluated on our SUSTech dataset shown in Fig. \ref{fig:trajectories} and TABLE \ref{tab:error} using evo tools \cite{grupp2017evo} . Due to the accurate calibration, rich depth-associated visual features and their accurate tracking, CamVox system is very close to ground truth and significantly outperformed the other frameworks such as livox\_horizon\_loam and VINS-mono.
\begin{figure}[h] \centering
\includegraphics[width=\linewidth]{trajectory.pdf}
\caption{Trajectories from livox\_horizon\_loam, VINS-mono and CamVox together with ground truth from SUSTech dataset.}
\label{fig:trajectories}
\end{figure}
\subsection{Timing results}
The timing analysis of the CamVox framework was performed as illustrated in TABLE \ref{tab:time}. The evaluation was performed on a onboard computer system Manifold 2C which has a 4-core Intel Core i7-8550U processor. With such a system, CamVox is able to perform in real-time. The automatic calibration takes about 58s to finish. Because calibration thread only runs occasionally while the robot is in a stationary state, and update of parameter could happen at a later time, such calculation time will not be an issue for real time performance.
\begin{table}[h]
\centering
\caption{Absolute Pose Error (APE) (unit: m)}
\label{tab:error}
\resizebox{\linewidth}{!}{%
\begin{tabular}{@{}cccc@{}}
\toprule
APE & CamVox & VINS-mono & livox\_horizon\_loam \\ \midrule
max & \textbf{3.3} & 27.2 & 9.9 \\
mean & \textbf{1.7} & 6.7 & 6.2 \\
median & \textbf{1.6} & 6.1 & 6.5 \\
min & \textbf{0.2} & 2.8 & 1.8 \\
rmse & \textbf{1.8} & 7.5 & 6.5 \\
sse & \textbf{16066.5} & 50788.6 & 101223.7 \\
std & \textbf{0.7} & 3.5 & 1.9 \\ \bottomrule
\end{tabular}%
}
\end{table}
\begin{table}[h]
\centering
\caption{Time Analysis}
\label{tab:time}
\resizebox{\linewidth}{!}{%
\begin{tabular}{@{}llll@{}}
\toprule
& Framework & CamVox & ORB-SLAM2 \\ \midrule
Setting & Dataset & SUSTech & TUM \\
& Resolution & 1520$\times$568 & 640$\times$480 \\
& Camera FPS & 10 Hz & 30 Hz \\
& ORB Features & 1500 & 1000 \\ \midrule
Thread & Calibration & 58.16 s & / \\
& Tracking & 42.27 ms & 25.58 ms \\
& Mapping & 252.41 ms & 267.33 ms \\
& Loop Closing & 7821.22 ms & 598.70 ms \\ \midrule
RGBD & IMU Corretcion & 0.89 ms & / \\
Preprocessing & Pcd2Depth & 16.35 ms & / \\ \bottomrule
\end{tabular}%
}
\end{table}
\section{Conclusion and perspective}
To summarize, we have proposed CamVox as a new low-cost lidar-assisted visual SLAM framework, aiming to combine the best from both worlds, i.e., the best angular resolution from camera and the best depth and range from lidar. Thanks to the unique working principle of Livox lidar, an automatic calibration algorithm that could perform in uncontrolled scenes is developed. Evaluations of this new framework was carried out in automatic calibration accuracy, depth threshold for close keypoints classification and trajectory comparison. It could also run with real time performance on an onboard computer. We hope that this new framework could be useful for robotic and sensor research and an out-of-the-box low-cost solution for the community.
\addtolength{\textheight}{-12cm}
\section*{ACKNOWLEDGMENT}
The authors thank colleagues at Livox Technology for helpful discussion and support.
|
\section{Introduction}
Many applications of generative modeling, especially in biological systems, are confronted to a limited amount of available data, from which a large number of parameters have to be inferred~\cite{nguyen2017inverse}.
A particularly interesting example is that of
proteins, which belong to the most interesting complex systems in nature and are essential in almost all biological processes. Most of them robustly fold into well-defined three-dimensional structures, which in turn form the basis of their functionality. This triangular sequence-structure-function relationship has, over several decades now, attracted substantial attention in biological physics~\cite{bryngelson1995funnels,dill2012the}.
A fascinating approach to the generative modeling of biological sequences has emerged over the last years \cite{de2013emerging,cocco2018inverse}. In the course of evolution, biological sequences accumulate mutations and become more diverse. We can now easily observe the sequence variability across large families of so-called homologous proteins, {\em i.e.}~proteins of common evolutionary ancestry and of close to equivalent function but in different species or biological pathways~\cite{el2019pfam}. Such homologous proteins may differ by 70-80\% of their amino acids without substantial changes in structure and function. However, their sequence variability is not fully random: a vast majority of mutations is deleterious, reducing protein stability or functionality. They are thus suppressed by natural selection. Only protein variants of similar or even better functionality are maintained.
In this way, the protein's structure and function constrain the viable sequence space that can be explored by evolution. Inverting this argument, the empirically observed variability of homologous sequences contains information about such evolutionary constraints, albeit frequently well hidden.
This idea is at the basis of the concept of data-driven ``sequence landscapes'', {\em i.e.}~classes of models that describe the statistical properties of protein families, assigning high probabilities to functional amino-acid sequences and low probabilities to non-functional ones \cite{levy2017potts,cocco2018inverse}. The log-probability (or minus ``energy'') is thus interpreted as a measure of sequence
fitness, hence the name of sequence (fitness) landscape \cite{de2014empirical}.
Among the best known such models are Potts models (PM), parameterized by local fields and two-site interaction couplings ({\em cf. }~below for details), and
constructed via the Direct Couplings Analysis (DCA) method, which is now firmly established~\cite{levy2017potts,cocco2018inverse}.
The DCA parameters can be obtained via inference or learning procedures ~\cite{balakrishnan2011learning,ekeberg2014fast, ekeberg2013improved, figliuzzi2018pairwise}, and they can be used to extract useful information on molecular structure~\cite{weigt2009identification, morcos2011direct,marks2012protein,ovchinnikov2017protein} and function ~\cite{morcos2014coevolutionary,szurmant2018inter}, on the effects of mutations ~\cite{figliuzzi2015coevolutionary,hopf2017mutation}, and to generate new artificially-designed molecules with specific properties~\cite{tian2018co,russ2020an}.
A concrete implementation of DCA is the following~\cite{figliuzzi2018pairwise}.
Given training data in the form of a Multiple Sequence Alignment (MSA) of $M$ homologous sequences of aligned length $L$, the PM parameters are learned by the so-called Boltzmann machine learning (BML) algorithm \cite{ackley1985learning}. By performing gradient ascent on the log-likelihood of the model given the data, BML determines values of couplings and fields such that the one- and two-site model frequencies match the empirical ones derived from the MSA. A standard pairwise $q$-state PM is thus specified by $q^2 L(L-1)/2$ couplings and $q L$ fields, where, for proteins, $q = 21$ corresponds to the 20 naturally occurring amino acids plus the gap symbol used for insertions or deletions.
Crucially, despite the fact that
modern sequencing techniques are making available a huge amount of biological sequences, and in particular hundreds of millions of protein sequences \cite{uniprot2019}, a serious over-fitting problem is present when PM are used as models of protein families.
In fact, with typical sequence lengths $L\sim 50-500$, the parameters to be inferred are $\sim 10^6-10^8$, which in most cases substantially exceeds
the available information from the MSA.
The resulting over-fitting is manifested in several ways: (i)~many couplings turn out to be rather small and noisy, (ii)~the PM is close to a critical point, {\em i.e.}~it can be very susceptible to small changes in its parameters, and (iii)~different training procedures, {\em e.g.}~with different initial conditions, can lead to significant changes in the sets of parameters without affecting the fitting accuracy, which severely limits the interpretability of the model.
These observations call for a parameter reduction procedure, which aims at identifying a minimal set of couplings needed to accurately describe the training data without overfitting.
Hopfield-Potts models~\cite{shimagaki2019selection} and the more general Restricted Boltzmann Machines (RBM)~\cite{tubiana2019learning} lead to a dimensional reduction of parameter space by learning collective ``patterns'' from sequence data, which in turn can be interpreted as extended sequence motifs and are activated via a limited number of hidden variables. The resulting coupling matrix is low-rank but still dense. A complementary approach aims at sparsifying the network of couplings: $\ell_1$-norm regularization has been used in a number of approximate methods~\cite{jones2012psicov,kamisetty2013assessing}, but cannot be easily used for generative modeling, because the regularization penalizes also non-zero couplings, which in turn assume too small values. Alternatively, a ``color-compression'' scheme~\cite{rizzato2020inference} has been proposed, which groups together sequence symbols with low frequency in specific sites. However, frequent symbols may also be involved into statistically non-supported couplings. Another example is that proposed in \cite{gao2018correlation} where a candidate sparse graph topology is sought by pruning the MSA columns associated with low values of the mutual information. Although this method has to be preferred when $L$ is so large to prevent the standard DCA implementations, it completely loses some information on the target statistical model.
Overall, a statistically principled and efficient approach to construct sparse PM for protein sequence modeling is still lacking.
In this work, we introduce an information-theory based ``decimation'' procedure, which allows for an iterative and controlled removal of irrelevant couplings.
As a result, parameters are removed either if they have no statistical support (as in color compression), or if they have statistical support for being very small.
We show that up to about 90\% of the coupling parameters can be removed without observing any substantial change in the fitting accuracy and in the generative properties of the resulting Sparse Potts Model (SPM).
Although greedy, our pruning scheme does not require to add extra terms in the energy function of the model, at variance with any treatable regularization, like $\ell_1$ or $\ell_2$, and therefore it preserves the generative properties of PM.
Finally, we show that the resulting SPM are not close to criticality, at variance with the original PM learned using standard DCA. Our results thus demonstrate that the observed criticality of PMs inferred from protein sequence data is not an intrinsic feature of the biological systems themselves, {\em cf. }~\cite{mora2011biological}, but results from the over-fitting in the learning procedure.
\section{An information-guided decimation procedure}
With each sequence $S = (s_1,\cdots,s_L)$ of length $L$, in which $s_i$ can take $q$ possible values ($q=21$ for proteins), a PM associates a statistical ``energy'' or Hamiltonian $H(S)$, written as a sum over single-site fields $h_i(s_i)$ and two-site couplings $J_{ij}(s_i,s_j)$:
\begin{equation}} \newcommand{\eeq}{\end{equation}\label{eq:H}
H(S) = -\sum_{1\leq i<j \leq L} J_{ij}(s_i,s_j) - \sum_{1\leq i \leq L} h_i(s_i) \ .
\eeq
The negative of the Hamiltonian can be interpreted as a ``fitness score'' for protein sequence $S$, with an associated Boltzmann probability
$P(S) = \exp\{-H(S)\}/Z$, where $Z=\sum_S \exp\{-H(S)\}$ is the partition function guaranteeing correct normalization of $P$. Hence, the surface defined by $H(S)$ over the space of sequences can be interpreted as a ``fitness landscape'' or -- using a more cautious term -- ``sequence landscape'' for the protein family represented by the training MSA.
We define the ``model density'' $d$ as the number of non-zero couplings $J_{ij}(a,b)\neq 0$ divided by the total number of possible couplings $q^2 L(L-1)/2$. Note that this definition is given element-wise, {\em i.e.}~for each $i,j,a,b$, and not block-wise for entire $q\times q$ matrices $J_{ij}$ coupling two sites $i,j$.
Fields are not decimated and do not contribute to the model density: we consider them an essential ingredient of the model because they encode amino-acid conservation.
A fully connected model, {\em i.e.}~with $d=100\%$, can be trained to arbitrarily high accuracy using standard BML~\cite{figliuzzi2018pairwise}.
Let us define the empirical one-site frequency $f_i(a)$ of observing amino acid $a$ in position $i$ in the MSA, and two-site frequency $f_{ij}(a,b)$ of observing amino acid
$a$ in position $i$ and $b$ in position $j$ in the same sequence of the MSA. BML performs a gradient ascent on the log-likelihood, which gives update equations for the couplings and fields at each learning epoch:
\begin{equation}} \newcommand{\eeq}{\end{equation}\label{eq:PMlearning}
\begin{split}
\delta h_i(a) &= \eta_h [ f_i(a) - p_i(a) ] \ , \\
\delta J_{ij}(a,b) &= \eta_J [ f_{ij}(a,b) - p_{ij}(a,b) ] \ ,
\end{split}\eeq
where the $p_i(a), p_{ij}(a,b)$ are the one- and two-site marginal probabilities of the PM, which are estimated at each iteration of the learning by sampling $P(S)$ via a standard Markov Chain Monte Carlo (MCMC) simulation, and $\eta_h, \eta_J$ are the learning rates for fields and couplings.
These equations are iterated until convergence to a fixed point, at which the model almost perfectly matches the empirical frequencies.
For all the cases we investigate (with one exception, see Appendix~\ref{app:quality}), we can ensure that the MCMC sampling is done in equilibrium and the resulting PM can be
sampled ergodically in relatively short times, see also~\cite{decelle2021equilibrium} for a detailed discussion.
Note that a PM trained in this way also corresponds to the maximum entropy or least constrained model that is compatible with the one- and two-site empirical frequencies ~\cite{lapedes1999correlated,weigt2009identification}.
Our decimation procedure consists in choosing pairs of sites $i<j$ and amino acids $a,b$, and fixing the corresponding coupling permanently to $J_{ij}(a,b)=0$. The coupling is removed from the set of adjustable parameters, and the corresponding two-site frequency $f_{ij}(a,b)$ is no longer explicitly fitted in the subsequent BML epochs.
However, an important property of PM is the so-called ``gauge'' or reparameterization invariance: the transformation
\begin{equation}\begin{split}
J_{ij}(a,b) &\rightarrow} \def\la{\left\langle} \def\ra{\right\rangle J_{ij}(a,b) +{\cal J}_{ij} (a) + {\cal K}_{ij} (b) \ , \\
h_i(a) &\rightarrow} \def\la{\left\langle} \def\ra{\right\rangle h_i(a)-{\cal H}_i -\sum_{j (> i)} {\cal J}_{ij} (a)-\sum_{j (< i)} {\cal K}_{ji} (a) \ ,
\end{split}
\label{eq:gauge}
\end{equation}
leaves the Hamiltonian in Eq.~\eqref{eq:H} and the associated Boltzmann distribution $P(S)$ invariant, for any choice of the $\JJ$, ${\cal K}}\def\UU{{\cal U}} \def\QQ{{\cal Q}} \def\XXX{{\cal X}$ and ${\cal H}$.
Hence, a gauge transformation can transform a zero coupling into a non-zero one and vice versa.
Because the decimation procedure fixes some couplings to zero, it breaks this invariance.
We thus begin our decimation procedure by a ``gauge fixing'' step, which sets to zero $2q-1$ out of all $q^2$ entries of each coupling matrix $J_{ij}$. To do so, we identify, independently for each pair of sites $1\leq i<j\leq L$, the $2q-1$ amino-acid pairs $(a,b)$ of smallest connected correlation $c_{ij}(a,b)=f_{ij}(a,b)-f_i(a)f_j(b)$, and fix the corresponding couplings $J_{ij}(a,b)$ to zero. Only the other $q^2-2q+1 = (q-1)^2$ couplings are updated using the BML, Eq.~\eqref{eq:PMlearning}. This procedure chooses a model of minimal density $d = [(q-1)/q]^2 = 90.7\%$ out of all equivalent PM related by the gauge transformation in Eq.~\eqref{eq:gauge}.
The parameters are initialized using a ``profile model'' fitting only the one-site frequencies $f_i(a)$. This initial model has zero couplings and fields $h_i^{(0)}(a) = \log f_i(a) +{\cal H}_i$, with the constant ${\cal H}_i$ being fixed by $\sum_a h_i^{(0)}(a) =0$ (with a very small pseudo-count added to $f_i(a)$ to avoid infinite fields, see Appendix~\ref{app:training}).
The fitting quality of the learned PM is tested by the Pearson correlation between the empirical $c_{ij}(a,b)$ and their counterparts in the model $P(S)$, the latter being estimated from a large independently and identically distributed MCMC sample. For all protein families considered in this work, this Pearson correlation exceeds 0.95, see Fig.~\ref{fig:pearson} and Appendix~\ref{app:quality}.
Note that the results of our decimation procedure depend on the initialization and gauge fixing described above. We tried a different initialization, either fixing both couplings and fields to zero, or initializing both using pseudo-likelihood maximization (PLM) \cite{ekeberg2013improved}. We found qualitatively similar results, but with slightly worse performance (Appendix~\ref{app:init}).
Any further decimation of couplings changes the model. To measure the impact of removing a given coupling $J_{ij}(a,b)$ from a PM, we determine the symmetric Kullback-Leibler (KL) divergence between the Boltzmann distributions with and without that coupling.
We thus consider a Potts Model with Hamiltonian $H$, and another with Hamiltonian $H'$ in which a given coupling is removed:
\begin{equation}
H'(S)=H(S) + J_{ij}(a,b) \delta_{a,s_i} \delta_{b,s_j} \ .
\end{equation}
We observe that averages over $P'= e^{-H'}/Z'$ can be expressed in terms of averages over $P= e^{-H}/Z$ as
\begin{equation}\begin{split}
\langle O(S) \rangle_{P'}
&= \frac{\sum_S O(S) e^{-H'(S)}}{\sum_S e^{-H'(S)}}\\
& = \frac{\sum_S O(S) e^{-J_{ij}(a,b) \delta_{a,s_i} \delta_{b,s_j}} e^{-H(S)}}{\sum_S e^{-J_{ij}(a,b) \delta_{a,s_i} \delta_{b,s_j}}e^{-H(S)}} \\
& =\frac{\langle O(S) e^{-J_{ij}(a,b) \delta_{a,s_i} \delta_{b,s_j}} \rangle_P}{\langle e^{-J_{ij}(a,b) \delta_{a,s_i} \delta_{b,s_j}} \rangle_P} \ .
\end{split}\end{equation}
Hence, the symmetric Kullback-Leibler divergence of $P$ and $P'$ is
\begin{eqnarray}
D_{ij}^{ab} &=& D_{\rm KL}(P||P') +D_{\rm KL}(P'||P) \nonumber \\
&=& -\sum_S [P(S)-P'(S)] [\log P(S) - \log P'(S)] \nonumber\\
&=&
\langle H'-H \rangle_P - \langle H'-H \rangle_{P'}\nonumber \\
&=&\langle J_{ij}(a,b) \delta_{a,s_i} \delta_{b,s_j} \rangle_P - \langle J_{ij}(a,b) \delta_{a,s_i} \delta_{b,s_j} \rangle_{P'} \label{eq:DKL}\\
&=&\langle J_{ij}(a,b) \delta_{a,s_i} \delta_{b,s_j} \rangle_P \nonumber\\
&& -
\frac{\langle J_{ij}(a,b) \delta_{a,s_i} \delta_{b,s_j} e^{-J_{ij}(a,b) \delta_{a,s_i} \delta_{b,s_j}} \rangle_P}{\langle e^{-J_{ij}(a,b) \delta_{a,s_i} \delta_{b,s_j}} \rangle_P}\nonumber \\
&=& J_{ij}(a,b) p_{ij}(a,b) - \frac{J_{ij}(a,b) p_{ij}(a,b) e^{-J_{ij}(a,b)}}{ p_{ij}(a,b) e^{-J_{ij}(a,b)}+1 - p_{ij}(a,b)}
\ ,\nonumber
\end{eqnarray}
where $p_{ij}(a,b)=\langle \delta_{a,s_i} \delta_{b,s_j} \rangle_P$ is the marginal two-site probability of $P$, which coincides, at convergence of Eq.~\eqref{eq:PMlearning}, with the empirical frequency $f_{ij}(a,b)$.
Note that we could also have equivalently used the non-symmetrized KL divergence (Appendix~\ref{app:score}).
At each decimation step, we now remove the least significant couplings, {\em i.e.}~those with the lowest $D_{ij}^{ab}$. For computational efficiency, this is done for a fixed fraction (in this work we choose $1\%$) of all remaining couplings.
Note that $D_{ij}^{ab}=D(J,p\sim f)$ (dropping the indices for notational simplicity) goes to $0$ either when $f \rightarrow} \def\la{\left\langle} \def\ra{\right\rangle 0$ or $f \rightarrow} \def\la{\left\langle} \def\ra{\right\rangle 1$ at fixed $J$, or when $J \rightarrow} \def\la{\left\langle} \def\ra{\right\rangle 0$ at fixed $f$.
More precisely, we have $D(J,f)\sim Jf(1-e^{-J})$ for $f\rightarrow 0$, $D(J,f)\sim J(f-1)$ for $f\rightarrow 1$, and $D(J,f)\sim J^2f(1-f)$ for $J\rightarrow 0$.
The first and second limits imply that finite couplings can be decimated if the corresponding frequency is close to zero or one,
{\em i.e.}~they have little statistical significance because the corresponding amino acids are almost never observed
(as in color-compression \cite{rizzato2020inference}) or almost always observed. The third limit indicates that small couplings are decimated whatever $f$ is (similar to the procedure proposed in \cite{decelle2014pseudolikelihood} for the inverse Ising problem using PLM).
Numerically, we observe that the percentage of pruned couplings corresponding to each category varies during decimation (Appendix~\ref{app:decimatedcouplings}). After a decimation step, we perform additional BML iterations of Eq.~\eqref{eq:PMlearning} on all undecimated couplings and the fields, to reach convergence again.
In this way, we progressively obtain PMs of reduced density, and we stop the decimation when $d=1\%$.
Note that in order to accurately estimate $D_{ij}^{ab}$, it is important that the PM learning is well converged before each decimation step. We attempted an ``online'' decimation in which couplings are pruned either after a fixed number of iterations of Eq.~\eqref{eq:PMlearning} or for having reached convergence, and found that this provides no advantage (Appendix~\ref{app:onlinelearning}), neither in terms of generative performance ({\em i.e. } the Pearson correlations at $d=1\%$ do not improve),
nor in computational efficiency ({\em i.e. } the computational time required to reach $d=1\%$ is almost unchanged). Other decimation strategies based on $f_{ij}(a,b)$ only (removing statistically unsupported couplings), or on $J_{ij}(a,b)$ only (removing small couplings), or on applying $\ell_1$-norm regularization to select relevant couplings, were found to perform substantially worse than the information-based procedure using Eq.~\eqref{eq:DKL} (Appendix~\ref{app:decimationstrategies}).
\begin{figure}[t]
\begin{center}
\includegraphics[width=\columnwidth]{Fig/Fig1_Pearson_PF76.png}
\caption{ {\bf Fitting and generative quality for PF00076 --} Pearson correlation coefficient between model and data frequencies as a function of the model density.
The one-site frequencies $f_i(a)$ are directly fitted. The two-site connected correlations $c_{ij}(a,b)$ are fully fitted by the densest model, while only a fraction of them are fitted for the sparse models at $d<1$. The three-site connected correlations $c_{ijk}(a,b,c)$ are never fitted. The generative performance of the model is essentially unchanged down to a density of 10\%, and slowly decays for even sparser couplings. However, even down to $d=1\%$, the Pearson coefficients remain at remarkably high values above 95\% for the two-site correlations, and above 84\% for the three-site correlations. }
\label{fig:pearson}
\end{center}
\end{figure}
We have also tested our decimation procedure on synthetic data (see
Appendix~\ref{app:synth}) and found that it is able to correctly identify the
ground-truth sparse model, provided enough data are available.
\section{Results and discussion}
We focus here on a representative protein family, the PF00076 family from the Pfam database~\cite{el2019pfam}, corresponding to a RNA recognition motif (RRM) of about 90 amino acids, known to bind single-stranded RNAs. The MSA provided by Pfam contains $M = 137605$ sequences of aligned length $L = 70$. Results obtained for other families (Appendix~\ref{app:otherproteins}) fully confirm the general conclusions drawn here for the RRM. The features of the protein families used for this work are reported in Appendix~\ref{app:data}.
In Fig.~\ref{fig:pearson} we show, for model densities down to 1\%, the Pearson correlation coefficient between the empirical one-site frequencies $f_i(a)$ obtained from the original MSA, and the model one-site marginal probabilities $p_i(a)$, estimated by MCMC sampling.
Similar curves are also shown for the two-site connected correlations $c_{ij}(a,b)$ and for a selected sub-set (specified in Appendix~\ref{app:quality})
of three-site connected correlations, defined as
\begin{equation}\begin{split}
c_{ijk}(a,b,c) &= f_{ijk}(a,b,c)
-f_{ij}(a,b)f_k(c) \\ -f_{jk}(b,c)f_i(a)
& - f_{ki}(c,a)f_j(b)+2f_i(a)f_j(b)f_k(c) \ , \label{eq:c3}
\end{split}\end{equation}
where $i,j,k$ are the indices of the columns of the MSA (which take value from $1$ to $L$), and
$a,b,c$ run over the amino-acids and the gap symbol (practically, from 1 to $q$).
The one-site frequencies are perfectly reproduced by the model, {\em i.e.}~$f_i(a)=p_i(a)$, as a consequence of the fixed-point condition in Eq.~\eqref{eq:PMlearning}, and the Pearson coefficient thus remains equal to one at all $d$ (up to tiny deviations due to the finite MCMC samples used in BML and in estimating $p_i(a)$).
For the maximal density $d_{\rm max}=[(q-1)/q]^2$ obtained after gauge fixing, the two-site correlations should also be perfectly reproduced because of Eq.~\eqref{eq:PMlearning}.
In practice we only
reach a Pearson coefficient of $\sim0.975$ due to sampling noise (Appendix~\ref{app:sampling}).
On the contrary, for $d<d_{\rm max}$ only a fraction of all two-site frequencies is explicitly fitted by the model via sparse BML.
Nevertheless the two-site Pearson coefficient is essentially independent of $d$, up to a slight reduction when $d<10\%$. Finally, three-site correlations, that are never explicitly fitted by the model (the training process in Eq.~\eqref{eq:PMlearning}
does not include three-site information), are nevertheless very accurately reproduced, with a Pearson coefficient around $0.94$ for all $d>10\%$. Note that the reproduction of unfitted observables is a highly non-trivial test for the generative properties of our models~\cite{figliuzzi2018pairwise}, {\em i.e.}~of the capacity of the model to generate data being statistically close to indistinguishable from the natural sequence data used for model learning. Below density $d=10\%$, the generative quality
of the model for three-site correlations is slightly reduced, remaining nevertheless very high (above 84\% down to $d=1\%$). We discuss the generative property of the sparse models introducing additional metrics in Appendix~\ref{app:similarity}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\columnwidth]{Fig/Fig2_PPV_PF00076.png}
\caption{ {\bf Contact prediction for PF00076 --} Positive predictive values (PPV) for several model densities, {\em i.e.}~the fraction of true positives
among the highest-ranking $k$ pairs $(i,j)$ of sites, when ordered by decreasing $F_{ij}^{\rm APC}$. Even the most sparse model, with only
1.6\% of couplings, shows an excellent performance at contact prediction. The curve for plmDCA, a standard DCA approach for contact prediction, is shown for reference and gives comparable results.}
\label{fig:PPV}
\end{center}
\end{figure}
A second test of model quality is the prediction of structural contacts, which constituted the major application of DCA in the last years. The idea is that pairs of strongly interacting sites in the PM should
correspond to close-by residues in the three-dimensional structure, which display strong coevolution to maintain the proper protein fold and functionality. Using the standard convention for coevolutionary contact prediction, we consider a pair of residues to be in contact if the distance between them is at most 8~$\mathring{\mathrm{A}}$, and we exclude easy-to-predict short-range contacts by considering only pairs with $|i-j|\geq 4$ in our analysis.
The reference (ground-truth) distance was obtained by the package~\cite{edoardo_sarti_andrea_pagnani_2020} that takes the shortest distance between heavy atoms in all protein structures registered in the Protein Data Bank (PDB)~\cite{berman2007worldwide} for the given Pfam family.
We follow the standard procedure for contact prediction, which consists in computing the average-product corrected (APC) Frobenius norms of the coupling matrices
(note that the coupling matrices are transformed into the zero-sum gauge and that the gap states $a,b=q$ are excluded from the sum \cite{feinauer2014improving}),
\begin{equation}} \newcommand{\eeq}{\end{equation}
F_{ij} = \sqrt{ \sum_{a,b=1}^{q-1} J_{ij}(a,b)^2 }\ ;
\ F_{ij}^{\rm APC} = F_{ij} - \frac{\sum_k F_{ik} \sum_k F_{k j}}{\sum_{kl} F_{kl}} \ .
\eeq
In Fig.~\ref{fig:PPV} we show the fraction of true contacts within the first $k$ pairs of sites, ranked in decreasing order of $F_{ij}^{\rm APC}$. We observe that the performance of the model at inferring the structural contacts is only slightly deteriorated even in the sparsest case $d=1.6\%$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.95\columnwidth]{Fig/Fig3_Histograms_dense_sp1276_PF76.pdf}
\caption{ {\bf Coupling distributions --} Distribution of couplings corresponding to true
contacts (top) and to non-contacts (bottom) in the three-dimensional protein fold, for the initial PM with density $d_{\rm max}=91\%$ and a sparser model having density $d=7\%$ associated with a reasonably accurate contact prediction.
}
\label{fig:interpretation}
\end{center}
\end{figure}
In Fig.~\ref{fig:interpretation} we show the probability distributions of couplings $J_{ij}(a,b)$, separately for pairs $i<j$ corresponding to contacts and all the others. We observe that, both for contacts and non-contacts, the decimation affects the shape of the distribution around $J\sim 0$ in a similar way, while the tails are essentially unaffected.
Overall, these
results explain why the performance of the PM for contact prediction using $F_{ij}^{\rm APC}$
is essentially independent of $d$ (Fig.~\ref{fig:PPV}). Unfortunately, the large-$J$ tail of the distributions of
couplings on non-contacting sites does not change upon sparsifying the model, which suggest that our decimation procedure cannot help in devising better contact predictors.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\columnwidth]{Fig/Fig4_Heat_Capacity_PF00076.png}
\caption{ {\bf Criticality --} Heat capacity as a function of temperature for models with different density. The densest models show
a strong peak of specific heat close to the reference scale $T=1$, which is a signature of criticality: the model is extremely sensitive to a small change of couplings, due
to overfitting. On the contrary, sparse models display a much smaller peak, which is also shifted away from $T=1$ towards lower temperatures, indicating a better
robustness of the learning.
}
\label{fig:CV}
\end{center}
\end{figure}
In order to study the criticality of the models, we consider a simple perturbation of the statistical weight,
{\em i.e.}~we rescale the Hamiltonian $H(S)$ by a formal inverse
temperature $\b = 1/T$ and set $P(S)\sim e^{-\b H(S)}$, in such a way that $T=1$ corresponds to the original model trained on data, while measuring the variation of the model entropy $S$.
In Fig.~\ref{fig:CV} we show the heat capacity $C = T \mathrm d S/ \mathrm d T$ of the PM for several sparsities (see Appendix~\ref{app:criticality} for details on the computation of $C$).
Note that a large $C$ indicates a large variation of the model entropy with $T$, or equivalently that the model
statistics changes strongly after a slight change of the parameters. This is indeed
the best definition of criticality in statistical physics, keeping in mind that our models have finite size $L$ and we thus cannot perform a finite-size scaling analysis
to determine if the observed peak in $C$ corresponds to a phase transition in the thermodynamic limit.
In Fig.~\ref{fig:CV} we observe that the denser models display a large peak in $C$ close to $T=1$, which indicates that the models are close to criticality.
Upon sparsifying the model, the peak amplitude is strongly reduced and the peak is also shifted to lower temperatures, {\em i.e.}~further away from the reference scale $T=1$.
These results suggest that the criticality of the dense models comes from over-fitting. Because the dense models have a huge number of parameters, they
are able to fit all the details of the training data. As a consequence, the model becomes very sensitive to noise, and a little change of the parameters changes a lot the model
statistics. On the contrary, sparse models have less parameters and are thus more robust to noise.
Ref.~\cite{melamed2013deep} provides Deep Mutational Scanning (DMS) data for a representative member of the PF00076 family, namely the RRM2 domain of the Poly(A)-Binding Protein (PABP) in the yeast species {\it Saccharomyces cerevisiae}. Using this domain as a reference,
the authors generated a library of 110,745 protein variants, including 1,246 single amino-acid substitutions
and 39,912 double amino-acid substitutions. Each of these variants was experimentally scored for function, by monitoring the growth of mutant yeast and finally, a ``fitness score'' was attributed to each mutant sequence in the experiment~\cite{melamed2013deep}. Within our models, the inferred Hamiltonian $H(S)$ in Eq.~\eqref{eq:H}
is also interpreted as a sequence-fitness score. Hence, a good test of the generative property of our models is to check whether the energy differences $\Delta H = H({\rm mutant})- H({\rm reference})$ between mutant sequences and the PABP reference
correlate well with the experimental fitness variations. Because the mapping between experimental and model fitness may be non-linear, in Fig.~\ref{fig:mut} we show the Spearman's rank correlation between these two variables, both for single and double mutants. In the dense $d=d_{\rm max}$ case, we reproduce the reference values already given in Ref.~\cite{hopf2017mutation}. We also observe
that upon reducing density, once again the model quality is not degraded, down to $d\sim 10\%$. Even for $d=1.6\%$ the model performs quite well, and much better than a profile model, which coincides with the limit $d\rightarrow} \def\la{\left\langle} \def\ra{\right\rangle 0$ of our decimation procedure.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\columnwidth]{Fig/Fig5_Spearm_fitness_PF76_all.pdf}
\caption{ {\bf Single and double mutations --}
Spearman correlation between the experimental fitnesses and the model predictions as a function of the model density,
both for single and double mutants of the PABP, a member of the PF00076 family. The dashed lines show the same correlations for a profile model ({\em i.e.}~$d=0$) as a reference.
}
\label{fig:mut}
\end{center}
\end{figure}
\section{Conclusions}
We introduced a general parameter reduction scheme for Boltzmann Machine Learning, and we applied it to Potts models for protein sequence data, {\em i.e.}~for the learning of highly accurate and generative, but sparse DCA models. Our strategy makes use of a rigorous information-based criterion to select couplings that are iteratively pruned. Intuitively, removed couplings are either statistically unsupported, {\em i.e.}~they correspond to pairs of amino acids that are almost never or almost always observed, similarly
to the color-compression scheme~\cite{rizzato2020inference}, or they are small, {\em i.e.}~they correspond to pairs that are only weakly correlated, or a combination of both. The statistical significance of a coupling is precisely quantified by the symmetric KL divergence between the Boltzmann measure of the Potts model with and without this coupling, which
is exactly computable from the model or the empirical statistics.
While our method is fully general for learning Boltzmann machines from high-dimensional categorical data, here we focused on its application to model protein families via Potts models, in which strong couplings are usually associated with physical contacts in the three-dimensional protein fold. We stress that the aim of this work is not to provide a sparse graph topology underlying the true interaction network, and indeed the pruning is not performed block-wise but at the level of the individual coupling entries, but to provide a general framework of parameters reduction strongly based on information-theoretic assumptions.
We have shown that the model can be decimated down to less than $10\%$ of the original couplings, while losing neither its generative quality,
nor its accuracy in contact prediction.
However, it has to be noted that many couplings not corresponding to structural contacts remain non-zero even in the lowest-density models. The interpretation of such couplings remains unclear. They may result from subtle effects due to the phylogenetic relations between the training sequences \cite{qin2018power,horta2020phylogenetic}, but also from extended functional constraints as those found by Restricted Boltzmann Machines or Hopfield-Potts models~\cite{tubiana2019learning,shimagaki2019selection}. As a result, further work is needed to make DCA-type modeling fully interpretable.
The sparse models resulting from our decimation procedure are also far away from criticality: they do not display the specific-heat peak close to the formal temperature $T=1$ that characterizes the dense models. Hence, we attribute the criticality observed in dense models to over-fitting, and conclude that our decimation procedure makes model learning more robust to finite-sample noise. Finally, the model maintains its performance in predicting the fitness of mutations around a reference sequence, {\em i.e.}~it is capable of predicting the {\it local} shape of the fitness landscape after having been trained on a {\it global} alignment of distantly diverged amino-acid sequences.
Our decimation procedure solves the first two problems mentioned in the introduction: we can eliminate small and noisy couplings, and the resulting model maintains its
fitting and generative qualities, while being statistically more robust. Unfortunately, we were unable to solve the third problem, namely the strong dependence on the initial condition of the training: different initial conditions (zero couplings and fields, profile model, plmDCA) produce fully-connected PMs of equal fitting quality but with slightly different performances in predicting contacts and mutational effects. This difference does not disappear after decimation (Appendix~\ref{app:init}). In other words, our decimation procedure remains sensitive to the initial fully-connected model from which it is started.
The resulting sparse PMs attempt to fit the data by using the minimal number of two-site couplings, {\em i.e}~using coupling matrices that are as sparse as possible. In the context of proteins (or RNA), it is natural to think that the sparse couplings identified by the model are the most relevant to describe the physical two-site correlations
that arise from the need to maintain the three-dimensional folded structure.
This strategy is complementary to collective-feature learning, {\em e.g.}~via RBM or Hopfield-Potts models~\cite{tubiana2019learning,shimagaki2019selection}, in which the number of parameters in the coupling matrix is reduced by
assuming it to be of low-rank. The features learned by these machines are associated with global sequence motifs, related, {\em e.g.}, to protein function or its interactions, but the accuracy of contact prediction is reduced.
An interesting and natural perspective would be to combine these two strategies into a general ``sparse plus low-rank'' scheme, {\em cf. }~\cite{zhang2016improving} for a related idea,
which could lead to an accurate description of protein families in an easily interpretable way, with sparse two-site couplings describing
physical constraints coming from structural contacts, and low-rank couplings describing biological features associated with protein function and its evolutionary history.
To conclude,
we would like to stress once more that our information-based decimation strategy is not specific to the application of Potts models to protein sequence data.
It can directly be used in other applications of inverse statistical physics and Boltzmann machine learning,
as {\em e.g. } in modeling neural or socio-economic data~\cite{nguyen2017inverse}, and may be adapted to more general network reconstruction schemes.
The code for learning and pruning the PMs is available at ~\cite{adabmDCA}.
\acknowledgments
We would like to thank Matteo Bisardi, Simona Cocco, Yaakov Kleeorin, R\'emi Monasson, Rama Ranghanatan, Olivier Rivoire and Jeanne Trinquier for discussions related to this work.
We acknowledge funding by the EU H2020 research and innovation programme MSCA-RISE-2016 (grant 734439 InferNet, to MW), by the Simons Foundation (\#454955, to FZ) and by the Honjo International Scholarship Foundation (PhD grant to KS).
P.B.-C., A.P.M. and K.S. contributed equally to this work. Author contributions:
P.B.-C. and M.W. designed the research;
P.B.-C., A.P.M., K.S., and F.Z. performed the research;
A.P.M. and K.S. analyzed the data;
A.P.M., K.S., M.W., and F.Z. wrote the paper.
|
\section{Introduction}
Scattering amplitudes have been a fertile ground for the study of implications of fundamental principles, manifested as constraints on its analytic properties. It's utility is often based on being expressible in terms of purely on-shell degrees of freedom. In four-dimension massless theory, this is reflected in the use of spinor-helicity formalism (see \cite{1308.1697, 1708.03872} for review). Recently, Arkani-Hamed, Huang and Huang introduced a new on-shell massive spinor helicity formalism \cite{1709.04891}, that manifest the little group covariance of four-dimensional scattering amplitudes. Their supersymmetric extension were further developed by \cite{1902.07204, 1810.04694, 1902.07205}. For any four-point massless tree-amplitude, its residue can be expanded on an orthogonal polynomial basis, with the precise polynomial depending on the helicity of the external and the spin of exchanged states. For external scalars these are just the Legendre polynomials. For external helicity states, the polynomials are proportional to Algebraic Jacobi Function (AJF) $\mathcal{J}^{\alpha,\beta}_{S}\left(-x \right)$, here referred to as spinning polynomials \cite{1709.04891}
\begin{equation}
\boxed{
P_{s}^{\{h_{i}\}}(x)=g(h_{i},s)\mathcal{J}^{\alpha,\beta}_{s}\left(-x \right)}\,,
\end{equation}
where
\begin{align}
\nonumber g(h_{i},s)=&(-1)^{s+h_{1}-h_{2}-h_{3}-h_{4}}\frac{\sqrt{\left( s+\alpha\right) !(s-\alpha)!(s+\beta)!(s-\beta)!}}{(2s)!}\\
\mathcal{J}^{\alpha,\beta}_{s}\left(-x \right) =& \sqrt{\frac{\left(s+\alpha\right) !(s-\alpha)!}{(s+\beta)!(s-\beta)!}}\left( \frac{1+x}{2}\right) ^{\frac{\alpha+\beta}{2}}\left( \frac{1-x}{2}\right) ^{\frac{\alpha-\beta}{2}}\times J^{\alpha+\beta,\alpha-\beta}_{s-\alpha}\left( -x\right).
\end{align}
$x=cos(\theta)$ is scattering angle in center of mass frame thus $-1\leq x\leq1$. $\alpha=h_{1}-h_{2}, \,\beta=h_{3}-h_{4}$ with $h_i$ the helicity of external states. The fact that they are expandable on orthogonal polynomials, is a simple reflection of the uniqueness of the three-point amplitude between one massive and two massless legs in four-dimensions. In this paper, we will show that when the theory enjoys super poincare invariance, the residue of the four-point amplitude is expandable on the super polynomials $\mathcal{P}^{H_{i}}_{\mathcal{N},\mathcal{S}}(x,\eta_{i}^{\mathcal{N}})$, each represent the exchange of a spin-$\mathcal{S}$ massive super multiplet. These are nothing but AJF multiplying Grassmann delta function:
\begin{equation}
\begin{split}
&\boxed{\mathcal{P}^{\{H_{i}\}}_{\mathcal{N},\mathcal{S}}(x,\eta_{i}^{\mathcal{N}})=m^{\mathcal{N}}\delta^{(2\mathcal{N})}(Q^{\dagger}_{}) g(H_{i},\mathcal{S})\mathcal{J}^{H_{1}-H_{2},H_{3}-H_{4}}_{\mathcal{S}}\left(-x \right)
}
\end{split}\, ,
\end{equation}
Instead of using $h_i$ to label helicity in the spinning polynomials, we use capital $H_i$ label the helicity of the super multiplets and $\alpha=H_{1}-H_{2}, \,\beta=H_{3}-H_{4}$ here. Note that since both the super and spinning polynomials are expressed on the same basis of AJF, combined with the fact super multiplet decompose onto a sum of components, implies algebraic relation amongst the AJFs. This is precisely given by the recurrence relation of Jacobi polynomial
\begin{equation}
\begin{split}
J^{a,b}_{n}(-x)=&\frac{(n+a+b+1)}{(2n+a+b+1)}J^{a+1,b}_{n}(-x)-\frac{(n+b)}{(2n+a+b+1)}J^{a+1,b}_{n-1}(-x).
\end{split}
\end{equation}
An explicit example is shown in the fixing external states on specific component field. We use Grassmann derivative to fix external states and the recurrence relation give us:
\begin{equation}
\begin{split}
\frac{\partial}{\partial\eta_{2}}\frac{\partial}{\partial\eta_{4}}\mathcal{P}^{\{H_{i}\}}_{\mathcal{N}=1,\mathcal{S}}(x,\eta_{i})=&m^2P^{\{ H_{1}, H_{2}{-}\frac{1}{2},H_{3},H_{4}{-}\frac{1}{2}\} }_{\mathcal{S}{+}\frac{1}{2}}\left( x\right)\\
&{+}\frac{m^{2}(\mathcal{S}{-}H_{3}{+}H_{4})(\mathcal{S}{-}H_{1}{+}H_{2})}{\left( 2\mathcal{S}\right) \left( 2\mathcal{S}{+}1\right) }P^{ \{H_{1}, H_{2}{-}\frac{1}{2},H_{3},H_{4}{-}\frac{1}{2}\} }_{\mathcal{S}{-}\frac{1}{2}}\left( x\right)\, .
\end{split}
\end{equation}
Said in another way, the algebraic relation amongst the AJFs encodes the possible extension of the Poincare group!
Equipped with the supersymmetric polynomials, we utilize it to constrain the space of EFTs consistent with supersymmetry. As shown in \cite{EFThedron}, unitarity, locality and Poincare invariance allows us to impose bounds on the coefficients of higher dimension operators of the low energy theory. Poincare invaraince is reflected in the fact that the low energy coefficients are bounded by the convex hull of the derivative expansion of the spinning polynomial. By replacing the latter with supersymmetric polynomial, we obtain new bounds which depends on the number of supersymmetry assumed. We show the bounded region of gluon amplitude in MHV helicity configuration which can be generally expanded as:
\begin{equation}
A^\mathcal{N}(1^+,2^-,3^+,4^-)|_{s,t\ll \Lambda_{UV}}=\SB{1 3}^2\AB{2 4}^2(\frac{a}{st}+\frac{b}{s}{+}\frac{b}{t}+\sum_{k,q\geq 0} g_{k,q}s^{k{-}q}t^q)\,.
\end{equation}
The convex hull constrains for $g_{k,q}$ can be found in each $k$ order. We show the constrains for each $g_{k,q}$ in supersymmetry between $\mathcal{N}=0\sim4$. The result in $k=5$ is in fig.\ref{pin}. It is directly to see, the higher $\mathcal{N}$ supersymmetry will give us the smaller allowed region.
\begin{figure}[h!]
\centering
\includegraphics[width=0.4\linewidth]{N5.pdf}
\caption{Cyclic-polytope condition on $(g_{5,1}/g_{5,0},g_{5,2}/g_{5,0})$ with supersymmetry $\mathcal{N}$. We don't show $g_{5,3} , g_{5,4}$ here, it will related to $g_{5,1}, g_{5,2}$ by the crossing symmetry $g_{5,i}=g_{5,5-i}$. The bounded region decrease when the superesymmetry $\mathcal{N}$ growth.}
\label{pin}
\end{figure}\\
This paper is organized as follows. In section \ref{sec:Review of spinning polynomials}, we briefly review the spinning polynomials and algebraic Jacobi polynomials. In section \ref{sec:n=1}, we review on-shell $\mathcal{N}=1$ superspace and fix the three-point amplitude with two massless and one massive state consistent with $\mathcal{N}=1$ super Poincare symmetry. We glue two three-point amplitudes and have the residue of $\mathcal{N}=1$ four massless super amplitude. Following the convention of ~\cite{1709.04891}, we transform the result into center of mass frame with scattering angle $cos(\theta)$ and we have $\mathcal{N}=1$ super polynomial. We decompose super polynomial into spinning polynomials by using recurrence relation of AJF. In section \ref{sec:Extended supersymmetry}, we start from review of $\mathcal{N}=2$ superspace\cite{1902.07204} ,then we derive $\mathcal{N}=2$ super polynomial and the general $\mathcal{N}$ super polynomial. Applying super polynomials on EFT, we show the unitary condition on the low energy coefficients in section \ref{sec:EFT}.
\section{Review of spinning polynomials}
\label{sec:Review of spinning polynomials}
Let's start from review spinning polynomials which are the expansion basis for the residues/discontinuities of massless four-point amplitude. The residues of the four-point are nothing but product of two three-point amplitudes. In spacetime dimensional four, two massless one massive amplitudes are uniquely fixed by the helicities $h_{1}$, $h_{2}$ of legs 1 and 2, and the spin-$s$ of the massive leg.
\begin{equation}
\label{}
\begin{split}
A(1^{h_{1}},2^{h_{2}},\textbf{P}^{(I_{1}I_{2}\ldots I_{2s})}) &=m^{1+h_{1}+h_{2}-s}\langle12\rangle^{-h_{1}-h_{2}-s} \prod_{t=1}^{h_{1}-h_{2}+s} \langle2 \textbf{P} ^{(I_{t}}\rangle \prod_{r=h_{1}-h_{2}+s+1}^{2s}\langle1\textbf{P}^{I_{r})}\rangle
\end{split}
\end{equation}
The symmetrized SU(2) indices forms the irreducible representation of spin-$s$ state. The powers of $m$ is to ensure the amplitude has the canonical mass dimension in 4d, i.e. 1. Gluing two vertices, we have the residue of four-massless amplitude with exchanging a spin-$s$ state.
\begin{equation}
\label{gspinpoly}
\begin{split}
Res_{s=m^2}&[A_{4}(h_{i},s)]= A_{L}(1^{h_{1}},2^{h_{2}},\textbf{P}_{(I_{1}I_{2}\ldots I_{2s})})\cdot A_{R}(3^{h_{3}},4^{h_{4}},-\textbf{P}^{(I_{1}I_{2}\cdots I_{2s})})
\\
&= m^{1+h_{1}+h_{2}-s}
\langle12\rangle^{-h_{1}-h_{2}-s}
\prod_{t=1}^{h_{1}-h_{2}+s}\langle2\textbf{P}_{I_{(t}}\rangle\prod_{r=h_{1}-h_{2}+s+1}^{2s}
\langle 1\textbf{P}_{I_{r)}}\rangle
\\
&\ m^{1+h_{3}+h_{4}-s}
\langle34\rangle^{-h_{3}-h_{4}-s}
\prod_{t=1}^{h_{3}-h_{4}+s}\langle4\vert -\textbf{P}^{I_{(t}}\rangle\prod_{r=h_{3}-h_{4}+s+1}^{2s}
\langle3\vert-\textbf{P}^{I_{r)}}\rangle
\end{split}
\end{equation}
We use convention $\vert{-}\textbf{P}^{I}\rangle=\vert\textbf{P}^{I}\rangle,\,\,\vert{-}\textbf{P}^{I}\rbrack={-}\vert\textbf{P}^{I}\rbrack$ to denote opposite momentum. The tensor contraction of three-point is done by using identity: $\vert \textbf{P}_{I}\rangle_{\alpha}\langle \textbf{P}^{I}\vert^{\beta}=-m_{\textbf{P}}\delta^{\beta}_{\alpha}$. For example, $\left\langle 1\textbf{P}_{I_{r}}\right\rangle $ $\left\langle 3\textbf{P}^{I_{t}}\right\rangle $ can be glued as $\delta^{t}_{r} m_{\textbf{P}}\langle13\rangle$. Then everything we need to do is summing over all contraction in (\ref{gspinpoly}). It is easy to show the combinatorial number in the contraction. Assuming one have contracted spinor $|1\rangle$ and $|3\rangle$ "a" times, then we have the spinor bracket with power
\begin{equation}
\{\langle13\rangle, \langle14\rangle, \langle24\rangle, \langle23\rangle\}=\{a, h_2{-}h_1{+}s{-a}, h_1{-}h_2{+}h_3{-}h_4{+}a, h_{4}{-}h_3{+}s{-}a\}.
\end{equation}
Then we can write down the combinatorial number for each spinor brackets:
\begin{eqnarray}
\nonumber&\left\langle 13\right\rangle^{a}&:
\left(
\begin{array}{cr}
s+h_{2}-h_{1} & \\
a
\end{array}\right) \times\left(
\begin{array}{cr}
s+h_{4}-h_{3} & \\
a
\end{array}\right)\times a!\,
\\
\nonumber&\left\langle 14\right\rangle^{(h_{2}-h_{1}+s-a)}&:
\left(\begin{array}{cr}
h_{3}-h_{4}+s & \\
h_{2}-h_{1}+s-a
\end{array}\right) \times(h_{2}-h_{1}+s-a)!\,
\\
\nonumber&\left\langle24 \right\rangle^{(h_{1}-h_{2}+h_{3}-h_{4}+a)}&:
\left(
\begin{array}{cr}
h_{1}-h_{2}+s & \\
h_{1}-h_{2}+h_{3}-h_{4}+a
\end{array}\right) \times(h_{1}-h_{2}+h_{3}-h_{4}+a)!\,
\\
\label{eq:permutation}&\left\langle23 \right\rangle^{(h_{4}-h_{3}+s-a)}&:(h_{4}-h_{3}+s-a)!\, .
\end{eqnarray}
Summing over "a", we have the contraction result in (\ref{gspinpoly}). It can be
\begin{equation}
\begin{split}
Res_{s=m^2}&[A_{4}(h_{i},s)]=
\\
&\sum_{a}\frac{m^{2-2s+\sum_{i} (h_{i})}}{(2s)!}
\frac{(h_{1}-h_{2}+s)!(h_{2}-h_{1}+s)!(h_{3}-h_{4}+s)!(h_{4}-h_{3}+s)!}{(a)!(s-h_{1}+h_{2}-a)!(s-h_{3}+h_{4}-a)!(h_{1}-h_{2}+h_{3}-h_{4}+a)!}
\\
&\times\left\langle 12\right\rangle^{-h_{1}-h_{2}-s}\left\langle 34\right\rangle^{-h_{3}-h_{4}-s}\langle13\rangle^{a}\langle14\rangle^{s-h_{1}+h_{2}-a}\langle 24\rangle^{h_{1}-h_{2}+h_{3}-h_{4}+a}\langle23\rangle^{h_{4}-h_{3}+s-a} .
\end{split}
\end{equation}
Since the power of spinor brackets are positive integer or zero, the summation is restricted to the region
\begin{equation}
a \geq 0,\quad s-h_{1}+h_{2}-a \geq 0, \quad s-h_{3}+h_{4}-a \geq 0, \quad h_{1}-h_{2}+h_{3}-h_{4}+a \geq 0 .
\end{equation}
It is useful to convert spinor brackets into scattering angle $x=cos\,\theta$ in the center of mass frame. The transformation rule can be found in appendix \ref{sec:spinor-helicity}. As a result, the spinning polynomial is
\begin{equation}
Res_{s=m^{2}}[A_{4}(h_{i},s)]\equiv P_{s}^{\{h_{i}\}}(x)=g(h_{i},s)\mathcal{J}^{\alpha,\beta}_{s}\left(-x \right)
\end{equation}
where
\begin{align}
\nonumber g(h_{i},s)=&(-1)^{s+h_{1}-h_{2}-h_{3}-h_{4}}\frac{\sqrt{\left( s+\alpha\right) !(s-\alpha)!(s+\beta)!(s-\beta)!}}{(2s)!}\\
\nonumber\mathcal{J}^{\alpha,\beta}_{s}\left(-x \right) =& \sqrt{\frac{\left(s+\alpha\right) !(s-\alpha)!}{(s+\beta)!(s-\beta)!}}\left( \frac{1+x}{2}\right) ^{\frac{\alpha+\beta}{2}}\left( \frac{1-x}{2}\right) ^{\frac{\alpha-\beta}{2}}\times J^{\alpha+\beta,\alpha-\beta}_{s-\alpha}\left( -x\right)\\
&-1 \leq x \leq1 .
\end{align}
Here $\alpha, \beta$ define as $\alpha=h_{1}-h_{2}, \beta=h_{3}-h_{4}$.
Note that the spin-$s$ must satisfy $s\geq \mid h_{1}-h_{2}\mid$ and $\{s, h_{1}-h_{2}\}$ must be all integer or half-integer. $\mathcal{J}^{\alpha,\beta}_{s}$ is Algebraic Jacobi Function(AJF) which satisfy orthogonality condition
\begin{equation}
\begin{array}{lll}
\int_{-1}^{1}dx\,\, \mathcal{ J}_s^{\alpha,\beta}(x) \mathcal{ J}_{s'}^{\alpha,\beta}(x) = \int_{-1}^{1}dx\,\, (1-x)^{\alpha}(1+x)^{\beta}J^{\alpha,\beta}_{s}(x)J^{\alpha,\beta}_{s'}(x) &=&
\frac{1}{2s+1}\delta_{s\, s'} .
\end{array}
\end{equation}
The advantage of using AJF as opposed to Jacobi-polynomial is the fact that they satisfy the same orthogonality condition as Legendre polynomials.
Said in anotherway, using AJF is equivalent to taking into account the fact that the residue has a universal prefactor that is determined by the external helicities. This relation can be translated to the spinning polynomials
\begin{equation}
\begin{split}
\begin{array}{lll}\label{eq:normalc}
\int_{-1}^{1}dx\,P_s^{h_{i}}(x)\,\, P_{s'}^{h_{i}}(x) &=&
\frac{\delta_{s\, s'}}{2s+1} \frac{\left( s+\alpha\right) !(s-\alpha)!(s+\beta)!(s-\beta)!}{\left( (2s)!\right)^{2} }.
\end{array}
\end{split}
\end{equation}
\section{$\mathcal{N} = 1$ Supersymmetry}
\label{sec:n=1}
In this section, we give a brief review of on-shell superspace and superfields. This will allow us to fix the 3-point super amplitude, after which one simply follows the previous example of spinning polynomials: gluing three-point to obtain the super polynomials.
\subsection{Review of $\mathcal{N}=1$ super space}
We will begin with $\mathcal{N}=1$, the generalization to extended $\mathcal{N}$ are straight forward. Following the convention of \cite{1902.07204} and \cite{0808.1446}, the SUSY generators can be completely written in terms of on-shell variables by introducing grassmann odd variables $\eta$. For example the massive supercharge are
\begin{equation}
\begin{split}
q _{\dot{a}}=\tilde{\lambda}_{\dot{a}I}\frac{\partial}{\partial \eta_{I}}\,,\quad q^{\dagger} _{ a}=\lambda_{a}^{I}\eta_{I} .
\end{split}
\end{equation}
For massless case, this reduces to
\begin{equation}
\begin{split}
q _{\dot{a}}=\tilde{\lambda}_{\dot{a}}\frac{\partial}{\partial \eta}\,,\quad q^{\dagger} _{ a}=\lambda_{a}\eta .
\end{split}
\end{equation}
\\
Supergenerators satisfy super algebra relations, $\{q_{\dot{a}},q_{a}^{\dagger}\}=\textbf{P}_{a\dot{a}}$. For on-shell SUSY representation, we introduce on-shell superfields whose $\eta$ expansion gives the component degrees of freedom. For example, a massive spin-$S$ $\mathcal{N}{=}1$ multiplet can be written as
\begin{equation}
\label{eq:N1superfield}
\begin{split}
\mathcal{S}^{(I_{1}\ldots I_{2S})}&=\phi^{(I_{1}\ldots I_{2S})}+\eta_{J}\Psi^{J(I_{1}\ldots I_{2S})}-\frac{1}{2}\eta_{J}\eta^{J} \tilde{\phi}^{(I_{1}\ldots I_{2S})}\, ,
\end{split}
\end{equation}
where the $(I_{1}\cdots I_{2S})$ denote $\frac{I_{1}I_{2}\cdots I_{2S}+I_{2}I_{1}\cdots I_{2S}+\cdots}{(2S)!}$.
Each order on the expansion gives different spin states with the same mass. The component field $\Psi^{J(I_{1}\ldots I_{2S})}$ can be separated into irreducible representation by systematically symmetrizing and anti-symmetrizing the free index J with $(I_{1}\cdots I_{2S})$. This leads to $\Psi^{(JI_{1}\ldots I_{2S})}$ and $\Psi_{L}^{\,\, (LI_{2}\ldots I_{2S})}$ representing spin-$S\pm\frac{1}{2}$ states
\begin{equation}
\Psi^{J(I_{1}\ldots I_{2S})}= \Psi^{(JI_{1}\ldots I_{2S})}+\frac{1}{(2S+1)} \sum_{k=1}^{2S}\epsilon^{JI_{k}} \Psi_{L}^{\,\,(LI_{1}\ldots I_{k-1}I_{k+1}\ldots I_{2S})} \, .
\end{equation}
We can derive the component degrees of freedom from (\ref{eq:N1superfield})
\begin{align}
\nonumber\phi^{(I_{1}\ldots I_{2S})}&=\mathcal{S}^{(I_{1}\ldots I_{2S})}\mid_{\eta=0}\\
\nonumber\Psi^{J(I_{1}\ldots I_{2S})}&=\frac{\partial}{\partial \eta_{J}}\mathcal{S}^{(I_{1}\ldots I_{2S})}\mid_{\eta=0}\\
\label{n1de}\tilde{\phi}^{(I_{1}\ldots I_{2S})}&=\frac{1}{2}\frac{\partial}{\partial\eta_{J}}\frac{\partial}{\partial\eta^{J}}\mathcal{S}^{(I_{1}\ldots I_{2S})}\mid_{\eta=0}\, .
\end{align}
For massless multiplet, we instead have
\begin{equation}
\Phi^{H_{i}}=\phi^{H_{i}}+\eta \psi^{H_{i}-\frac{1}{2}}\, ,
\end{equation}
where $\eta$ carries $+\frac{1}{2}$ helicity.\\
\noindent\textbf{$\mathcal{N} = 1$ three-point amplitudes}\\
We now consider the three-point super amplitude for the scattering of $H_{1}$, $H_{2}$ massless and a spin-$S$ massive multiplet, $\mathcal{A}(\Phi_1^{H_1},\Phi_1^{H_1},S_{\textbf{p}}^{(I_{1}\ldots I_{2S})})$. The constraint on $\mathcal{A}$ includes it must have $H_1$, $H_2$ helicity weight, carry $(I_1\ldots I_{2s})$ symmetrised indices, and satisfy the supersymmetry Ward identities
\begin{equation}
Q^{\dagger}_{a}\,\mathcal{A}(\Phi_{1}^{H_{1}},\Phi_{2}^{H_{2}},\mathcal{S}_{\textbf{P}}^{(I_{1}\ldots I_{2S})})=Q_{\dot{a}}\,\mathcal{A}(\Phi_{1}^{H_{1}},\Phi_{2}^{H_{2}},\mathcal{S}_{\textbf{P}}^{(I_{1}\ldots I_{2S})})=0
\end{equation}
where
\begin{equation}
Q^{\dagger}_{ a}=\lambda_{1,a}^{}\eta_{1}+\lambda_{2,a}^{}\eta_{2}+\lambda^{I}_{\textbf{p},a}\eta_{\textbf{P}I}\, .
\end{equation}
Since $Q_{a}^{\dagger}$ is a multiplicative generator, the amplitude must be proportional to to $\delta^{2}(Q)$
\begin{equation}
\mathcal{A}=\delta^{(2)}(Q^{\dagger})f(\lambda,\tilde{\lambda},\eta)\, ,
\end{equation}
where
\begin{equation}
\begin{split}
\delta^{(2)}(Q^{\dagger}) = \left( \langle i\vert\eta_{i}+\langle j\vert\eta_{j}+\langle \textbf{P}^{I}\vert\eta_{PI}\right) \left( \vert i\rangle\eta_{i}+\vert j\rangle\eta_{j}+\vert \textbf{P}_{I}\rangle\eta^{PI}\right)
\\
=\langle1^{}2^{}\rangle\eta_{1^{}}\eta_{2^{}} + \langle1^{}\textbf{P}^{I}\rangle\eta_{1^{}}\eta_{PI} +\langle2^{}\textbf{P}^{I}\rangle\eta_{2^{}}\eta_{PI} +\frac{m}{2}\eta_{PI}\eta_{P}^{I}\, .
\end{split}
\end{equation}
One can then easily check with the momentum conservation
\begin{equation}
Q\delta^{(2)}(Q^{\dagger})=0\, .
\end{equation}
The degree in $\eta$ of $f(\lambda,\tilde{\lambda},\eta)$ for general $\mathcal{N}$ and n-point amplitude is given by
\begin{equation}
[ f ]\leq(n-3)\mathcal{N}\, .
\end{equation}
Since $n=3$, $f(\lambda,\tilde{\lambda},\eta)$ is purely bosonic. Then we have\cite{1902.07204}
\begin{equation}
\label{eq:3pt}
\begin{split}
\mathcal{A}(\Phi_{1}^{H_{1}},\Phi_{2}^{H_{2}},\mathcal{S}_{\textbf{P}}^{(I_{1},\ldots,I_{2S})}) &=\delta^{(2)}(Q^{\dagger})m^{H_{1}+H_{2}-\mathcal{S}}\langle12\rangle^{-H_{1}-H_{2}-\mathcal{S}} \prod_{t=1}^{H_{1}-H_{2}+\mathcal{S}} \langle2\textbf{P}^{(I_{t}}\rangle \prod_{r=H_{1}-H_{2}+\mathcal{S}+1}^{2\mathcal{S}}\langle1\textbf{P}^{I_{r})}\rangle
\end{split}.
\end{equation}
Its' component amplitudes can be extracted by Grassmann derivative. As an example, we show the component amplitude with helicity $H_1{-}\frac{1}{2}$ and $H_2$ on the massless states. There are four states in the massive multiplet and we choose $\Psi^{(I_{1}\ldots I_{2S}J)}, \Theta^{I_{1}\ldots I_{2S}J}$ which both are coming from mixed field $\Psi^{J(I_{1}\ldots I_{2S})}$. Since we are considering mixed field here, we first derive the component amplitude:
\begin{equation}
\label{n1cm}
\begin{split}
A(\psi^{H_{1}-\frac{1}{2}},\phi^{H_{2}},\Psi_{\textbf{P}}^{J(I_{1},\ldots,I_{2S})}) &=\frac{\partial }{\partial\eta_{1}}\frac{\partial }{\partial\eta_{PJ}}\mathcal{A}(\Phi^{H_{1}},\Phi^{H_{2}},\mathcal{S}_{\textbf{P}}^{(I_{1},\ldots,I_{2S})})
\\
&=\langle 1\textbf{P}^J\rangle m^{H_{1}+H_{2}-\mathcal{S}}\langle12\rangle^{-H_{1}-H_{2}-\mathcal{S}} \prod_{t=1}^{H_{1}-H_{2}+\mathcal{S}} \langle2\textbf{P}^{(I_{t}}\rangle \prod_{r=H_{1}-H_{2}+\mathcal{S}+1}^{2\mathcal{S}}\langle1\textbf{P}^{I_{r})}\rangle .
\end{split}
\end{equation}
The mixed field in (\ref{n1cm}) can further decompose onto spin-$\mathcal{S}\pm\frac{1}{2}$ by symmetrizing the SU(2) indices
\begin{equation}
\begin{split}
&A(\psi^{H_{1}-\frac{1}{2}},\phi^{H_{2}},\Psi_{\textbf{P}}^{(I_{1},\ldots,I_{2S}J)})= m^{H_{1}+H_{2}-\mathcal{S}}\langle12\rangle^{-H_{1}-H_{2}-\mathcal{S}} \prod_{t=1}^{H_{1}-H_{2}+\mathcal{S}} \langle2\textbf{P}^{(I_{t}}\rangle \prod_{r=H_{1}-H_{2}+\mathcal{S}+1}^{2\mathcal{S}+1}\langle1\textbf{P}^{I_{r})}\rangle\\
&A(\psi^{H_{1}-\frac{1}{2}},\phi^{H_{2}},\Theta_{\textbf{P}}^{I_{1}\ldots I_{2S}J})= m^{H_{1}+H_{2}-\mathcal{S}}\langle12\rangle^{-H_{1}-H_{2}-\mathcal{S}}
\\
&\times\lbrack\langle 1\textbf{P}^J\rangle\prod_{t=1}^{H_{1}-H_{2}+\mathcal{S}} \langle2\textbf{P}^{(I_{t}}\rangle \prod_{r=H_{1}-H_{2}+\mathcal{S}+1}^{2\mathcal{S}}\langle1\textbf{P}^{I_{r})}\rangle
-\prod_{t=1}^{H_{1}-H_{2}+\mathcal{S}} \langle2\textbf{P}^{(I_{t}}\rangle \prod_{r=H_{1}-H_{2}+\mathcal{S}+1}^{2\mathcal{S}+1}\langle1\textbf{P}^{I_{r})}\rangle\rbrack .
\end{split}
\end{equation}
\subsection{The N=1 super polynomial}
\label{sec:N=1 spinning polynomial}
With the unique three-point super amplitude, we start to construct the basis for 4-point massless super-amplitude by using $\mathcal{N}=1$ supersum
\begin{equation}
\label{eq:Res}
\begin{split}
Res_{s=m^2}&[\mathcal{A}_{4}(H_{i},\mathcal{S})]= \int \frac{1}{2}\epsilon^{IJ} d\eta_{P_{I}} d\eta_{P_{J}}\mathcal{A}_{L}(\Phi^{H_{1}},\Phi^{H_{2}},\mathcal{S}_{\textbf{P}(I_{1}\cdots I_{2S})})\cdot \mathcal{A}_{R}(\Phi^{H_{3}},\Phi^{H_{4}},\overline{\mathcal{S}}_{-\textbf{P}}^{(I_{1}\cdots I_{2S})})
\\
=\int \frac{1}{2}&\epsilon^{IJ} d\eta_{PI} d\eta_{PJ}
\\
\times\delta^{(2)}&(Q_{L}^{\dagger}) m^{H_{1}+H_{2}-\mathcal{S}}
\langle12\rangle^{-H_{1}-H_{2}-\mathcal{S}}
\prod_{t=1}^{H_{1}-H_{2}+\mathcal{S}}\langle2\textbf{P}_{(I_{t}}\rangle\prod_{r=H_{1}-H_{2}+\mathcal{S}+1}^{2\mathcal{S}}
\langle 1\textbf{P}_{I_{r})}\rangle
\\
\times\delta^{(2)}&(Q_{R}^{\dagger}) m^{H_{3}+H_{4}-\mathcal{S}}
\langle34\rangle^{-H_{3}-H_{4}-\mathcal{S}}
\prod_{t=1}^{H_{3}-H_{4}+\mathcal{S}}\langle4\vert-\textbf{P}^{(I_{t}}\rangle\prod_{r=H_{3}-H_{4}+\mathcal{S}+1}^{2\mathcal{S}}
\langle3\vert-\textbf{P}^{I_{r})}\rangle\,,
\end{split}
\end{equation}
where we identify $\vert{-}\textbf{P}\rangle=\vert\textbf{P}\rangle, \vert {-}\textbf{P}\rbrack={-}\vert \textbf{P}\rbrack, \eta^{I}_{{-}P}={-}\eta^{I}_{P}$. The grassmann delta functions are
\begin{align}
\nonumber\delta^{(2)}(Q_{L}^{\dagger}) =& \langle1^{}2^{}\rangle\eta_{1^{}}\eta_{2^{}} + \langle1^{}\textbf{P}^{I}\rangle\eta_{1^{}}\eta_{PI} +\langle2^{}\textbf{P}^{I}\rangle\eta_{2^{}}\eta_{PI} +\frac{m}{2}\eta_{PI}\eta_{P}^{I}\\
\delta^{(2)}(Q_{R}^{\dagger}) =& \langle3^{}4^{}\rangle\eta_{3^{}}\eta_{4^{}} + \langle3^{}\vert-\textbf{P}^{I}\rangle\eta_{3^{}}\eta_{PI} +\langle4^{}\vert-\textbf{P}^{I}\rangle\eta_{4^{}}\eta_{PI} +\frac{m}{2}\eta_{PI}\eta_{P}^{I} .
\end{align}
The integration is super sum which only acts on $\delta^{(2)}(Q_{L}^{\dagger})$ and $\delta^{(2)}(Q_{R}^{\dagger})$. We can reduce Grassmann delta function as
\begin{equation}
\label{eq:eta}
\begin{split}
\int &\frac{1}{2}\epsilon^{AB} d\eta_{PA} d\eta_{PB}\delta^{(2)}(Q_{L}^{\dagger})\delta^{(2)}(Q_{R}^{\dagger})
\\=&m( \langle12\rangle \eta_{1}\eta_{2}+\langle13\rangle \eta_{1}\eta_{3}+\langle14\rangle\eta_{1}\eta_{4}+\langle23\rangle\eta_{2}\eta_{3}+
\langle24\rangle\eta_{2}\eta_{4}+\langle34\rangle\eta_{3}\eta_{4})
\\
=&m\delta^{(2)}(Q^{\dagger}_{4pt})\, .
\end{split}
\end{equation}
As the spinning example (\ref{gspinpoly}), we contract the spinor brackets in (\ref{eq:Res}). The result of the contraction is the same as in (\ref{gspinpoly}) with $h_i\rightarrow H_i$ in (\ref{gspinpoly}). Thus we have:
\begin{equation}
\label{n1sp}
\begin{split}
\boxed{\mathcal{P}^{\{H_{i}\}}_{\mathcal{N}=1,\mathcal{S}}(x,\eta_{i})=m\delta^{(2)}(Q^{\dagger}_{})g_{\left( H_{i},\mathcal{S}\right) }\mathcal{J}^{H_{1}-H_{2},H_{3}-H_{4}}_{\mathcal{S}}\left(-x \right)
}\quad
\end{split},
\end{equation}
where
\begin{equation}
\label{sgp}
g(H_{i},S)=(-1)^{S+H_{1}-H_{2}-H_{3}-H_{4}}\frac{\sqrt{(S+\alpha)!(S-\alpha)!(S+\beta)!(S-\beta)!}}{(2S)!}\, .
\end{equation}
In this paper, we call this result $\mathcal{P}^{\{H_{i}\}}_{\mathcal{N}=1,\mathcal{S}}(x)$: "$\mathcal{N}=1$ super polynomials". For simplicity, we will use $\alpha,\beta$ denote $H_i$, $\alpha=H_{1}-H_{2},\beta=H_{3}-H_{4}$. The variables $(\mathcal{S},\alpha,\beta)$ must be all integer or all half-integer due to the definition of AJF.\\
\noindent\textbf{Expansion on spinning polynomials}\\
Let us show that $\mathcal{P}^{\{H_{i}\}}_{\mathcal{N}=1,\mathcal{S}}(x,\eta_{i})$ can be recast on to the component spinning polynomials with positive coefficients. To facilitate the comparison between super and spinning polynomial, we first fix the external states of the super polynomial to be of specific component. With the fixed external states, super polynomial can be expanded on spinning polynomial
\begin{equation}
\label{eq:n1 general derivative form}
\frac{\partial}{\partial \eta_{j}}\frac{\partial}{\partial \eta_{k}} \mathcal{P}^{\{H_{i}\}}_{\mathcal{N}=1,\mathcal{S}}(x,\eta_{i})=\sum_{s} m^{2}a^{\{h_{i}\}}_{s} P_{s}^{\{h_{i}\}}(x),\quad j\neq k\, .
\end{equation}
We can get the coefficient $a_s$ directly by the orthogonality of spinning polynomial
\begin{equation}
\begin{array}{lll}
\frac{\int_{-1}^{1}dx\,Res_{s=m^2}[ \mathcal{A}_{4}^{(h_{i})}(x)]\,\, P_{s'}^{h_{i}}(x)}{ \frac{( s+\alpha) !(s-\alpha)!(s+\beta)!(s-\beta)!}{( (2s)!)^{2} }} =
\frac{a_{k}}{(2s+1)}
\end{array}\, .
\end{equation}
The same relation can be deduced from the recurrence relation of Jacobi polynomial
\begin{equation}
\label{rcr}
\begin{split}
J^{a,b}_{n}(-x)=\frac{(n+a+b+1)}{(2n+a+b+1)}J^{a+1,b}_{n}(-x)-\frac{(n+b)}{(2n+a+b+1)}J^{a+1,b}_{n-1}(-x)\, .
\end{split}
\end{equation}
With identifying $n=\mathcal{S}{-}\alpha, a=\alpha{+}\beta, b=\alpha{-}\beta$ in (\ref{rcr}), one can also decompose super polynomial. Either way we find
\begin{equation}
\label{n1rc}
\begin{split}
\frac{\partial}{\partial\eta_{2}}\frac{\partial}{\partial\eta_{4}}\mathcal{P}^{\{H_{i}\}}_{\mathcal{N}=1,\mathcal{S}}(x,\eta_{i})=&m^2P^{\{ H_{1}, H_{2}{-}\frac{1}{2},H_{3},H_{4}{-}\frac{1}{2}\} }_{\mathcal{S}{+}\frac{1}{2}}\left( x\right)\\
&{+}\frac{m^{2}(\mathcal{S}{-}H_{3}{+}H_{4})(\mathcal{S}{-}H_{1}{+}H_{2})}{\left( 2\mathcal{S}\right) \left( 2\mathcal{S}{+}1\right) }P^{ \{H_{1}, H_{2}{-}\frac{1}{2},H_{3},H_{4}{-}\frac{1}{2}\} }_{\mathcal{S}{-}\frac{1}{2}}\left( x\right)\, .
\end{split}
\end{equation}
We have similar relations for other external states which we list in Tab \ref{fig:n1}.
The first term in the expansion of (\ref{n1rc}) is related to spinning polynomial which have internal state $\Psi^{(I_{1},\ldots,I_{2S}J)}$. Indeed since
\begin{equation}
\begin{split}
&A_{3}^{\phi^{H_{1}},\psi^{H_{2}-\frac{1}{2}},\Psi^{(I_{1}\ldots I_{2S}J)}}\cdot A_{3}^{\phi^{H_{3}},\psi^{H_{4}-\frac{1}{2}},\tilde{\Psi}_{(I_{1}\ldots I_{2S}J)}}=m^2P^{\{ H_{1}, H_{2}{-}\frac{1}{2},H_{3},H_{4}{-}\frac{1}{2}\} }_{\mathcal{S}{+}\frac{1}{2}}\left( x\right)
\end{split}
\end{equation}
and
\begin{equation}
\begin{split}
A_{3}^{\phi^{H_{1}},\psi^{H_{2}-\frac{1}{2}},\Theta_{\textbf{P}}^{I_{1}\ldots I_{2S}J}}\cdot A_{3}^{\phi^{H_{3}},\psi^{H_{4}-\frac{1}{2}},\Theta_{\textbf{P}I_{1}\ldots I_{2S}J}}=\frac{m^{2}(\mathcal{S}{-}H_{3}{+}H_{4})(\mathcal{S}{-}H_{1}{+}H_{2})}{\left( 2\mathcal{S}\right) \left( 2\mathcal{S}{+}1\right) }P^{ \{H_{1}, H_{2}{-}\frac{1}{2},H_{3},H_{4}{-}\frac{1}{2}\} }_{\mathcal{S}{-}\frac{1}{2}}\left( x\right)\, ,
\end{split}
\end{equation}
eq(\ref{n1rc}) is equivalent to
\begin{equation}
\label{eq:14}
\begin{split}
\frac{\partial}{\partial\eta_{2}}\frac{\partial}{\partial\eta_{4}}\mathcal{P}^{\{H_{i}\}}_{\mathcal{N}=1,\mathcal{S}}(x,\eta_{i})
=&A_{3}^{\phi^{H_{1}},\psi^{H_{2}-\frac{1}{2}},\Psi^{(I_{1}\ldots I_{2S}J)}}\cdot A_{3}^{\phi^{H_{3}},\psi^{H_{4}-\frac{1}{2}},\tilde{\Psi}_{(I_{1}\ldots I_{2S}J)}}\\
&+A_{3}^{\phi^{H_{1}},\psi^{H_{2}-\frac{1}{2}},\Theta_{\textbf{P}}^{I_{1}\ldots I_{2S}J}}\cdot A_{3}^{\phi^{H_{3}},\psi^{H_{4}-\frac{1}{2}},\Theta_{\textbf{P}I_{1}\ldots I_{2S}J}}\, .
\end{split}
\end{equation}
Clearly, super polynomial is the super sum of spinning polynomials. Note that the recurrence relation tell us that the super polynomial with $\{H_i, \mathcal{S}\}$ is expanded with spinning polynomials $\{h_i, s\}$ in the region $\left| s-\mathcal{S}_{}\right|\leq\frac{1}{2} \wedge \left| h_{i}-H_{i}\right|\leq\frac{1}{2}$.
\begin{table}[]
\begin{center}
\begin{tabular}{|l|l|l|l|l|l|l|}
\hline
\multicolumn{4}{|l|}{External states helicity} &$a^{\{h_{i}\}}_{s-\frac{1}{2}}$ & $a^{\{h_{i}\}}_{s}$ & $a^{\{h_{i}\}}_{s+\frac{1}{2}}$ \\ \hline
$H_1{-}\frac{1}{2}$& $H_2{-}\frac{1}{2}$ & $H_3$ & $H_4$ & 0 &1 &0 \\ \hline
$H_1{-}\frac{1}{2}$& $H_2$ & $H_3{-}\frac{1}{2}$ & $H_4$ & $\frac{(\mathcal{S}+H_{3}-H_{4})(\mathcal{S}+H_{1}-H_{2})}{(2\mathcal{S})(2\mathcal{S}+1)}$ & 0 & 1 \\ \hline
\end{tabular}
\caption{\label{fig:n1} The expansion coefficient of super polynomial $\mathcal{P}^{\{H_{i}\}}_{\mathcal{N}=1,\mathcal{S}}$. The difference external states are labeled by helicity $H_i$ and $a^{h_{i}}_{s}$ are the coefficient.}
\end{center}
\end{table}
\section{Extended supersymmetry }
\label{sec:Extended supersymmetry}
Previously we derive $\mathcal{N}=1$ super polynomial which is product of Grassmann delta function and AJF. It is naturally to guess the extended super polynomials are still the product of Grassmann and AJF but Grassmann function should be generalized to $\delta^{(2\mathcal{N})}(Q^{\dagger}_{})$. In this section, we will explicitly show this generalization in $\mathcal{N}=2$ SUSY and list more generalized results in the end.
\subsection{Review of ${N}=2$ super space}
In $\mathcal{N} = 2$ SUSY, we simply consider superalgebra without central charge. The anti-commutation relation of super charges are
\begin{equation}
\left\lbrace q _{\dot{a}A},q^{\dagger B} _{ a}\right\rbrace =\delta^{B}_{A} \textbf{P}_{a\dot{a}}\, ,
\end{equation}
where
\begin{equation}
\begin{split}
q _{\dot{a}A}=\tilde{\lambda}_{\dot{a}I}\frac{\partial}{\partial \eta^{A}_{I}}\,,\quad q^{\dagger A} _{ a}=\lambda_{a}^{I}\eta^{A}_{I}\, .
\end{split}
\end{equation}
We use indices $A,B$ denote SU(2) R-symmetry group and $I,J$ denote little group.
In the massless limit, it can reduce to
\begin{equation}
q _{\dot{a}A}=\tilde{\lambda}_{\dot{a}}\frac{\partial}{\partial \eta^{A}}\,,\quad q^{\dagger A} _{ a}=\lambda_{a}\eta^{A}\, .
\end{equation}
\\
The $\eta$ expansion of on-shell super field in $\mathcal{N}=2$ is
\begin{equation}
\begin{split}
\mathcal{S}^{\left( I_{1}\ldots I_{2S}\right) }=&\phi^{\left( I_{1}\ldots I_{2S}\right) }+\eta^{A}_{I}\Psi^{ I\left( I_{1}\ldots I_{2S}\right)}_{A}-\frac{1}{2} \eta^A_I \eta^B_J (\epsilon^{IJ} \phi^{\left( I_{1}\ldots I_{2S}\right)}_{(AB)} + \epsilon_{AB} W^{(IJ)\left( I_{1}\ldots I_{2S}\right)})
\\
&+\frac{1}{3}\eta_I^B\eta_{JB}\eta^{JA}\tilde{\Psi}_A^{ I\left(I_{1}\ldots I_{2S}\right)}+\eta^{1}_{1}\eta^{2}_{1}\eta^{1}_{2}\eta^{2}_{2}\tilde{\phi}^{\left( I_{1}\ldots I_{2S}\right)} .
\end{split}
\end{equation}
Under massless limit, the super field can be simplified as
\begin{equation}
\Phi^{H_{i}}= \phi^{H_{i}} +\eta^{A}\, \psi_{A}^{H_{i}-\frac{1}{2}}+\frac{1}{2}\eta^{A}\eta_{A}\, w^{H_{i}-1}.
\end{equation}
We can extrapolate the component field by Grassmann derivative
\begin{align}
\nonumber\phi^{\left( I_{1}\ldots I_{2S}\right)}_{(AB)}&=\frac{1}{2}\frac{\partial}{\partial \eta^{JA}} \frac{\partial}{\partial \eta^{B}_{ J}}\mathcal{S}^{(I_{1}\ldots I_{2S})} \mid_{\eta=0}\\
W^{(IJ)\left( I_{1}\ldots I_{2S}\right)}&=\frac{1}{2}\frac{\partial}{\partial \eta^{A}_{I}} \frac{\partial}{\partial \eta_{AJ}}\mathcal{S}^{(I_{1}\ldots I_{2S})} \mid_{\eta=0}\, ,
\end{align}
where derivative of $\phi^{\left( I_{1}\ldots I_{2S}\right) }$, $\Psi^{ I\left( I_{1}\ldots I_{2S}\right)}_{A}$ were list in $\mathcal{N}=1$.
As $\mathcal{N}=1$ SUSY, component fields under $\eta$ expansion are the mixed states. So we start to decompose component fields into irreducible representation. $\phi^{\left( I_{1}\ldots I_{2S}\right) }$ and $\phi^{\left( I_{1}\ldots I_{2S}\right)}_{(AB)}$ are already irreducible and the decomposition of $\Psi^{ I\left( I_{1}\ldots I_{2S}\right)}_{A}$ can be found in $\mathcal{N}=1$. So we only list decomposition of $W^{(IJ)\left( I_{1}\ldots I_{2S}\right)}$ here. Symmetrizing and anti-symmetrizing the little group indices,
\begin{equation}
\begin{split}
W^{(IJ)\left( I_{1}\ldots I_{2S}\right)}&=W^{(IJ I_{1}\ldots I_{2S}) }
\\
&+\frac{(2S+1)}{2(2S)(2S+2)}\sum_{k=1}^{2S}\Bigg\lbrack\epsilon^{JI_{k}}(W^{(LI I_{1}\ldots I_{k-1}I_{k+1}\ldots I_{2S})}_{L}
+W^{(L\vert \,\,\,\vert I I_{1}\ldots I_{k-1}I_{k+1}\ldots I_{2S})}_{\,\,\,\,\,\,\,\,\,L}) +I\leftrightarrow J \Bigg\rbrack
\\
&+\frac{1}{2(2S)(2S+1)}\sum_{k\neq j}^{2S}\left[( \epsilon^{II_{k}}\epsilon^{JI_{j}}W^{\quad(LKI_{1}\ldots I_{k-1}I_{k+1}\ldots I_{j-1}I_{j+1}\ldots I_{2S})}_{LK})+I\leftrightarrow J \right] .
\end{split}
\end{equation}
It is directly to see $W^{(IJ)\left( I_{1}\ldots I_{2S}\right)}$ field contain spin-$S\pm1$ and spin-$S$ states. \\
\noindent\textbf{$\mathcal{N}=2$ three-point amplitudes}\\
Following the same procedure in $\mathcal{N}=1$, the three-point amplitude is uniquely fixed in $\mathcal{N}=2$. We list the result with helicity $H_{1}, H_{2}$ and spin-$S$:
\begin{equation}
\begin{split}
&\mathcal{A}(\Phi^{H_{1}},\Phi^{H_{2}},\mathcal{S}_{\textbf{P}}^{\left( I,\ldots,I_{2S}\right) })
\\=&\delta^{(4)}(Q^{\dagger})m^{H_{1}+H_{2}-\mathcal{S}}\langle12\rangle^{-H_{1}-H_{2}-\mathcal{S}} \prod_{t=1}^{H_{1}-H_{2}+\mathcal{S}} \langle1\textbf{P}^{( I_{t}}\rangle \prod_{r=H_{1}-H_{2}+\mathcal{S}+1}^{2\mathcal{S}}\langle2\textbf{P}^{I_{r}) }\rangle ,
\end{split}
\end{equation}
where
\begin{equation}
\begin{split}
\delta^{(4)}(Q^{\dagger}) = \prod^{\mathcal{N}=2}_{A=1}\left( \langle1^{}2^{}\rangle\eta^{A}_{1^{}}\eta^{A}_{2^{}} + \langle1^{}\textbf{P}^{I}\rangle\eta^{A}_{1^{}}\eta^{A}_{PI} +\langle2^{}\textbf{P}^{I}\rangle\eta^{A}_{2^{}}\eta^{A}_{PI} +\frac{m}{2}\eta^{A}_{PI}\eta_{P}^{AI}\right) .
\end{split}
\end{equation}
\subsection{The $\mathcal{N}=2$ super polynomial}
\label{sec:n2 spinning polynomials}
Having uniquely three-point result in $\mathcal{N}=2$, we start to glue them up. We use super sum to glue them just as what we have done in Sec.\ref{sec:N=1 spinning polynomial}. The difference in $\mathcal{N}=2$ is the super sum should also sum over all the R-symmetry indices
\begin{equation}
\label{eq:Res2}
\begin{split}
Res_{s=m^2}&[\mathcal{A}^{\mathcal{N}=2}(H_{i},\mathcal{S})]= \int \prod_{A=1}^{2}\left( \frac{1}{2} d\eta^{A}_{PI} d\eta_{P}^{AI}\right) \mathcal{A}_{L}(\Phi^{H_{1}},\Phi^{H_{2}},\mathcal{S}_{\textbf{P}(I_{1}\cdots I_{2S})})\cdot \mathcal{A}_{R}(\Phi^{H_{3}},\Phi^{H_{4}},\overline{\mathcal{S}}_{-\textbf{P}}^{(I_{1}\cdots I_{2S})})
\\
=&\int \frac{1}{2^{2}} d\eta^{A}_{PI} d\eta_{P}^{AI}\\
\times\delta^{(4)}&(Q_{L}^{\dagger}) m^{H_{1}+H_{2}-\mathcal{S}}
\langle12\rangle^{-H_{1}-H_{2}-\mathcal{S}}
\prod_{t=1}^{H_{1}-H_{2}+\mathcal{S}}\langle2\textbf{P}_{(I_{t}}\rangle\prod_{r=H_{1}-H_{2}+\mathcal{S}+1}^{2\mathcal{S}}
\langle 1\textbf{P}_{I_{r})}\rangle
\\
\times\delta^{(4)}&(Q_{R}^{\dagger}) m^{H_{3}+H_{4}-\mathcal{S}}
\langle34\rangle^{-H_{3}-H_{4}-\mathcal{S}}
\prod_{t=1}^{H_{3}-H_{4}+\mathcal{S}}\langle4\vert-\textbf{P}^{(I_{t}}\rangle\prod_{r=H_{3}-H_{4}+\mathcal{S}+1}^{2\mathcal{S}}
\langle3\vert-\textbf{P}^{I_{r})}\rangle .
\end{split}
\end{equation}
The $\eta$ integration only hit on Grassmann delta function and the integral result is
\begin{equation}
\label{eq:supersumn2}
\begin{split}
&\int \frac{1}{2^{2}}\prod_{A=1}^{2} \left( d\eta^{A}_{PI} d\eta_{P}^{AI} \right) \delta^{(4)}(Q_{L}^{\dagger})\delta^{(4)}(Q_{R}^{\dagger})
\\
&=m^{2}\prod_{A=1}^{2}( \langle12\rangle \eta^{A}_{1}\eta^{A}_{2}+\langle13\rangle \eta^{A}_{1}\eta^{A}_{3}+\langle14\rangle\eta^{A}_{1}\eta^{A}_{4}+\langle23\rangle\eta^{A}_{2}\eta^{A}_{3}+
\langle24\rangle\eta^{A}_{2}\eta^{A}_{4}+\langle34\rangle\eta^{A}_{3}\eta^{A}_{4})
\\
&=m^{2}\delta^{(4)}(Q^{\dagger}) .
\end{split}
\end{equation}
As the same procedure in spinning polynomial (\ref{gspinpoly}), we contract spinor brackets and have
\begin{equation}
\label{eq:4pt2}
\begin{split}
&\boxed{\mathcal{P}^{\{H_{i}\}}_{\mathcal{N}=2,\mathcal{S}}(x,\eta_{i}^{\mathcal{N}})=m^{2}\delta^{(4)}(Q^{\dagger}) g(H_{i},\mathcal{S})\mathcal{J}^{H_{1}-H_{2},H_{3}-H_{4}}_{\mathcal{S}}\left(-x \right)
}
\end{split}
\end{equation}
where $g(H_{i},\mathcal{S})$ is defined in (\ref{sgp}). As we expected, $\mathcal{N}=2$ super polynomial is still product of Grassmann delta function with AJF.\\
\noindent\textbf{Expansion on spinning polynomial}\\
Using the recurrence relation help us to decompose super polynomial into spinning polynomials in $\mathcal{N}=1$. Here we will use the recurrence relation again and show the decomposition in $\mathcal{N}=2$. The only difference between $\mathcal{N}=1$ and $\mathcal{N}=2$ is we need to repeat the recurrence relation twice in $\mathcal{N}=2$ and $P^{\{H_{i}\}}_{\mathcal{N}=2,\mathcal{S}}$ will be decomposed into four elements. The recurrence relation for Jacobi polynomial is:
\begin{equation}
\label{eq:1324recurrence relationJ}
\begin{split}
J^{a,b}_{n}(x)=& \frac{(n+a+b+1)}{(2n+a+b+1)}J^{a,b+1}_{n}(x)+\frac{(n+a)}{(2n+a+b+1)}J^{a,b+1}_{n-1}(x)\\
=&\frac{(n+a+b+1)}{(2n+a+b+1)}\frac{1}{(n+\frac{a}{2}+\frac{b}{2}+1)(1+x)}[(n+1)J^{a,b}_{n+1}(x)+(n+b+1)J^{a,b}_{n}(x)]\\
&+\frac{(n+a)}{(2n+a+b+1)}\frac{1}{(n+\frac{a}{2}+\frac{b}{2})(1+x)}[(n)J^{a,b}_{n}(x)+(n+b)J^{a,b}_{n-1}(x)]\, ,
\end{split}
\end{equation}
Applying (\ref{eq:1324recurrence relationJ}) into AJF, we can simply decompose super polynomial. As an example, we consider specific component amplitude here
\begin{equation}
\frac{\partial}{\partial\eta^{A}_{1}}\frac{\partial}{\partial\eta^{B}_{2}}\frac{\partial}{\partial\eta^{A}_{3}}\frac{\partial}{\partial\eta^{B}_{4}}\mathcal{P}_{\mathcal{N}=2,\mathcal{S}}^{\{H_{i}\}}(x,\eta^{C}_{i})\, .
\end{equation}
Identifying $n=\mathcal{S}-\alpha, a=\alpha{+}\beta, b=\alpha{-}\beta$ in recurrence relation, we decompose the component amplitude into
\begin{equation}
\label{N2ex}
\begin{split}
&\frac{\partial}{\partial\eta^{A}_{1}}\frac{\partial}{\partial\eta^{B}_{2}}\frac{\partial}{\partial\eta^{A}_{3}}\frac{\partial}{\partial\eta^{B}_{4}}\mathcal{P}_{\mathcal{N}=2,\mathcal{S}}^{\{H_{i}\}}(x,\eta^{C}_{i})=
\\
&m^{4}\{ P^{ \{H_{i}-\frac{1}{2}\} }_{\mathcal{S}+1}\left( x\right)+\Big(\frac{\left(\mathcal{S}+\beta \right) \left(\mathcal{S}-\alpha \right)}{\left( 2\mathcal{S}\right) \left( 2\mathcal{S}+1\right) }+\frac{\left(\mathcal{S}-\beta+1 \right) \left(\mathcal{S}+\alpha+1 \right)}{\left( 2\mathcal{S}+1\right) \left( 2\mathcal{S}+2\right) }\Big)P^{ \{H_{i}-\frac{1}{2}\} }_{\mathcal{S}}\left( x\right)
\\
&+\frac{\left(\mathcal{S}+\beta \right)\left(\mathcal{S}-\beta \right) \left(\mathcal{S}+\alpha \right)\left(\mathcal{S}-\alpha \right)P^{ \{H_{i}-\frac{1}{2}\} }_{\mathcal{S}-1}\left( x\right)}{\left( 2\mathcal{S}-1\right)\left( 2\mathcal{S}\right)^2 \left( 2\mathcal{S}+1\right) }\}\, .
\end{split}
\end{equation}
Similar results for other external states are in Tab.\ref{fig:n2} where
\begin{equation}
\label{N2cc}
\begin{split}
\mathfrak{a}&= \frac{(\mathcal{S}+\alpha)(\mathcal{S}+\alpha-1)(\mathcal{S}-\beta)(\mathcal{S}-\beta-1)}{(2\mathcal{S}-1)(2\mathcal{S})^{2}(2\mathcal{S}+1)}
\\
\mathfrak{b}&=\frac{(\mathcal{S}+\alpha)(\mathcal{S}-\beta)}{(2\mathcal{S})(2\mathcal{S}+1)}+\frac{(\mathcal{S}+\alpha)(\mathcal{S}-\beta)}{(2\mathcal{S}+1)(2\mathcal{S}+2)}
\\
\mathfrak{c}&=\frac{\left(\mathcal{S}+\beta \right)\left(\mathcal{S}-\beta \right) \left(\mathcal{S}+\alpha \right)!}{\left(\mathcal{S}+\alpha-2 \right)!\left( 2\mathcal{S}-1\right)\left( 2\mathcal{S}\right)^2 \left( 2\mathcal{S}+1\right) }
\\
\mathfrak{d}&=\frac{\left(\mathcal{S}+\beta \right) \left(\mathcal{S}+\alpha \right)}{\left( 2\mathcal{S}\right) \left( 2\mathcal{S}+1\right) }
+\frac{\left(\mathcal{S}-\beta+1 \right) \left(\mathcal{S}+\alpha \right)}{\left( 2\mathcal{S}+1\right) \left( 2\mathcal{S}+2\right) }
\\
\mathfrak{e}&=\frac{(\mathcal{S}+\alpha)(\mathcal{S}-\beta)}{(2\mathcal{S})(2\mathcal{S}+1)}
\\
\mathfrak{f}&=\frac{(\mathcal{S}+\alpha)(\mathcal{S}-\alpha)(\mathcal{S}+\beta)(\mathcal{S}-\beta)}{(2\mathcal{S}-1)(2\mathcal{S})^{2}(2\mathcal{S}+1)}
\\
\mathfrak{g}&=\frac{(\mathcal{S}-\alpha)(\mathcal{S}+\beta)}{(2\mathcal{S})(2\mathcal{S}+1)}+\frac{(\mathcal{S}+\alpha+1)(\mathcal{S}-\beta+1)}{(2\mathcal{S}+1)(2\mathcal{S}+2)} .
\end{split}
\end{equation}
\begin{table}[]
\begin{tabular}{|l|l|l|l|l|l|l|l|l|}
\hline
\multicolumn{4}{|c|}{External states helicity} & $a^{\{h_{i}\}}_{s-1}$ &$a^{\{h_{i}\}}_{s-\frac{1}{2}}$ &$a^{\{h_{i}\}}_{s}$ & $a^{\{h_{i}\}}_{s+\frac{1}{2}}$ & $a^{\{h_{i}\}}_{s+1}$ \\ \hline
$(H_1{-}1)_{AB}$& $(H_2{-}1)_{AB}$ & $H_3$ & $H_4$ & 0 & 0 & $\epsilon^{BA}\epsilon^{BA}$ & 0 &0 \\ \hline
$(H_1{-}1)_{AB}$& $H_2$ & $H_3$ & $(H_4{-}1)_{AB}$ & $\mathfrak{a}\epsilon^{BA}\epsilon^{BA}$ & 0 &$\mathfrak{b}\epsilon^{BA}\epsilon^{BA}$ & 0 & $\epsilon^{BA}\epsilon^{BA}$ \\ \hline
$(H_1{-}1)_{AB}$ & $H_2$ & $(H_3{-}\frac{1}{2})_{A}$ & $(H_4{-}\frac{1}{2})_{B}$ & $\mathfrak{c}\epsilon^{BA}$ & 0 & $\mathfrak{d}\epsilon^{BA}$ & 0 &$\epsilon^{BA}$ \\ \hline
$(H_1{-}1)_{AB}$ & $(H_2{-}\frac{1}{2})_{A}$ & $H_3$ & $(H_4{-}\frac{1}{2})_{B}$ & 0 &$\mathfrak{e}\epsilon^{BA} $ &0 &$\epsilon^{BA}$ & 0 \\ \hline
$(H_1{-}\frac{1}{2})_{A}$ & $(H_2{-}\frac{1}{2})_{A}$ & $(H_3{-}\frac{1}{2})_{B}$& $(H_4{-}\frac{1}{2})_{B}$ & 0 &0 &1 &0 &0 \\ \hline
$(H_1{-}\frac{1}{2})_{A}$ & $(H_2{-}\frac{1}{2})_{B}$ & $(H_3{-}\frac{1}{2})_{A}$& $(H_4{-}\frac{1}{2})_{B}$ & $\mathfrak{f}$ & 0 & $\mathfrak{g}$ &0 &1 \\ \hline
\end{tabular}
\caption{\label{fig:n2} Super polynomial $\mathcal{P}^{\{H_{i}\}}_{\mathcal{N}=2,\mathcal{S}}(x,\eta_{i}^{C})$ decomposed into spinning polynomials by recurrence relation. We use helicity $H_i$ to label each component states and the indices A,B are R indices in $\mathcal{N}=2$. Number $\mathfrak{a}, \mathfrak{b}, \mathfrak{c}, ......$ in the coefficient $a^{\{h_{i}\}}_{s}$ list in eq(\ref{N2cc}). }
\end{table}
The decomposition of (\ref{N2ex}) are precisely correspond to sum over spinning polynomials with internal states: $\phi_{\textbf{P}(CD)}^{(I_{1},\ldots,I_{2S})}$, $\mathcal{W}_{S}$ and $\mathcal{W}_{S\pm1}$ where
\begin{align}
\nonumber\mathcal{W}_{S+1}=&W_{\textbf{P}}^{(IJI_{1},\ldots,I_{2S})}\\
\nonumber\mathcal{W}_{S}=&\frac{(2S+1)}{2(2S)(2S+2)}\sum_{k=1}^{2S}\left[\epsilon^{JI_{k}}(W^{\,(LI I_{1}\ldots I_{k-1}I_{k+1}\ldots I_{2S})}_{L}+W^{(L\vert \,\,\,\vert I I_{1}\ldots I_{k-1}I_{k+1}\ldots I_{2S})}_{\,\,\,\,\,\,\,\,\,L}) +I\leftrightarrow J\right]\\
\nonumber\mathcal{W}_{S-1}=&\frac{1}{2(2S)(2S+1)}\sum_{k \neq j}^{2S}\left[( \epsilon^{II_{k}}\epsilon^{JI_{j}}W^{\quad(LKI_{1}\ldots I_{k-1}I_{k+1}\ldots I_{j-1}I_{j+1}\ldots I_{2S})}_{LK})+I\leftrightarrow J\right] .
\end{align}
Eq(\ref{N2ex}) is equivalent to
\begin{equation}
\label{eq:N=2,13A24B split to RtripletRsinglet}
\begin{split}
&\frac{\partial}{\partial\eta^{A}_{1}}\frac{\partial}{\partial\eta^{B}_{2}}\frac{\partial}{\partial\eta^{A}_{3}}\frac{\partial}{\partial\eta^{B}_{4}}\mathcal{P}_{\mathcal{N}=2,\mathcal{S}}^{\{H_{i}\}}(x,\eta^{C}_{i})
\\
=&A_{3}^{\psi^{H_{1}-\frac{1}{2}}_{A},\psi^{H_{2}-\frac{1}{2}}_{B},\mathcal{W}_{S+1}}\cdot A_{3}^{\psi^{H_{1}-\frac{1}{2}}_{A},\psi^{H_{2}-\frac{1}{2}}_{B},\mathcal{W}_{S+1}}+A_{3}^{\psi^{H_{1}-\frac{1}{2}}_{A},\psi^{H_{2}-\frac{1}{2}}_{B},\mathcal{W}_{S}}\cdot A_{3}^{\psi^{H_{1}-\frac{1}{2}}_{A},\psi^{H_{2}-\frac{1}{2}}_{B},\mathcal{W}_{S}}
\\
&+A_{3}^{\psi^{H_{1}-\frac{1}{2}}_{A},\psi^{H_{2}-\frac{1}{2}}_{B},\mathcal{W}_{S-1}}\cdot A_{3}^{\psi^{H_{1}-\frac{1}{2}}_{A},\psi^{H_{2}-\frac{1}{2}}_{B},\mathcal{W}_{S-1}}+A_{3}^{\psi^{H_{1}-\frac{1}{2}}_{A},\psi^{H_{2}-\frac{1}{2}}_{B},\phi_{\textbf{P}(CD)}^{(I_{1},\ldots,I_{2S})} }\cdot A_{3}^{\psi^{H_{3}-\frac{1}{2}}_{A},\psi^{H_{4}-\frac{1}{2}}_{B},\phi^{(CD)}_{\textbf{P}(I_{1},\ldots,I_{2S})}}\, .
\end{split}
\end{equation}
\noindent\textbf{Generalize super polynomial}\\
We have super polynomial in $\mathcal{N}=1$, $\mathcal{N}=2$ SUSY. These results can be easily generalize to extended supersymmetry by replace the Grassmann delta function. For $\mathcal{N}$ -extended supersymmetry, $m\delta^{(4)}(Q^{\dagger}_{4pts})$ is replaced with $m^{\mathcal{N}}\delta^{(2\mathcal{N})}(Q^{\dagger}_{4pts})$ in the super polynomial. The $\mathcal{N}$ super polynomials are
\begin{equation}
\label{eq:4ptn}
\begin{split}
&\boxed{\mathcal{P}^{\{H_{i}\}}_{\mathcal{N},\mathcal{S}}(x,\eta_{i}^{\mathcal{N}})=m^{\mathcal{N}}\delta^{(2\mathcal{N})}(Q^{\dagger}) g(H_{i},\mathcal{S})\mathcal{J}^{H_{1}-H_{2},H_{3}-H_{4}}_{\mathcal{S}}\left(-x \right)
}.
\end{split}
\end{equation}
\section{The supersymmetric EFThedron }\label{sec:EFT}
In this section, we will consider difference R-symmetry SUSY theory in low energy limit. The low energy means our kinematics $s,t$ are much smaller than UV mass scale $\Lambda_{UV}$. We further reduce super amplitude to component amplitude and show how R-symmetry bound EFT couplings. Consider gluon MHV amplitude $\mathcal{A}^{N}$ reduce from super-amplitude:
\begin{equation}
\mathcal{A}^{\mathcal{N}}(1^+,2^-,3^+,4^-)=\SB{1 3}^2\AB{2 4}^2\mathcal{A}^{\mathcal{N}}(s,t).
\end{equation}
We factor out common factor of spinor $\SB{1 3}^2\AB{2 4}^2$ in each $\mathcal{A}^{\mathcal{N}}(1^+,2^-,3^+,4^-)$ amplitudes and focus on improvement of $\mathcal{A}^{\mathcal{N}}(s,t)$ under SUSY increase. The low energy expansion of $\mathcal{A}^{\mathcal{N}}$ is
\begin{equation}
\mathcal{A}^{\mathcal{N}}(s,t)|_{s,t\ll \Lambda_{UV}}=(\frac{a}{st}+\frac{b}{s}{+}\frac{b}{t}+\sum_{k,q\geq 0} g_{k,q}s^{k{-}q}t^q)\,.
\end{equation}
Here we assume there is crossing symmetry $g_{k,q}=g_{k,k-q}$ in amplitude.
$g_{k,q}$ correspond to the effective couplings under low energy limit which can be extrapolated by contour integral around zero in the complex $s$-plane
\begin{equation}
g_{k,q}=\frac{1}{q!}\frac{d^q}{dt^q}\left(\frac{i}{2p}\oint\frac{ds}{s^n} \mathcal{A}(s,t)\right)\bigg|_{t=0}\,.
\end{equation}
We can deform the integral and catch the poles and discontinuities on real s-axis from
\begin{equation}
\label{jocobiexpand}
\mathcal{A}^{\mathcal{N}}(1^+,2^-,3^+,4^-)|_{s\rightarrow m^2}=\sum_{\ell} c_\ell\frac{\AB{2 4}^{\mathcal{N}} \mathcal{P}_{\ell}^{\alpha,\beta}(\cos\theta)}{s-m^2},\quad \cos\theta=1{+}\frac{2t}{m^2}\, ,
\end{equation}
where $\alpha,\beta$ correspond to super-field helicity weight and for MHV amplitude $\mathcal{M}(\bar{\Phi}^{+1},\Phi^{\mathcal{N}/2-1},\bar{\Phi}^{+1},\Phi^{\mathcal{N}/2-1})$ we consider is
\begin{equation}
\alpha=H_1-H_2=2-\frac{\mathcal{N}}{2},\quad\beta=H_3-H_4=2-\frac{\mathcal{N}}{2}.
\end{equation}
We expand (\ref{jocobiexpand}) in Taylor series and factor out common spinor factor. The coefficients of expanding super polynomial are $v_{\ell,q}^{\alpha,\beta}$
\begin{equation}
\AB{2 4}^{\mathcal{N}} \mathcal{P}_{\ell}^{\alpha,\beta}(1{+}\frac{2t}{m^2})=\SB{1 3}^2\AB{2 4}^2\sum_{\ell}v_{\ell,q}^{\alpha,\beta}t^{q}.
\end{equation}
At fix power of $k$, we have
\begin{equation}
\vec{g}_k=\sum \left(\begin{array}{c} g_{k,0} \\ g_{k,1} \\ \vdots \\ g_{k,n}\end{array}\right)\in\sum_{a} c_a \vec{\mathcal{P}}_{\ell_a}^{\alpha,\beta},\quad
\vec{\mathcal{P}}_{\ell}^{\alpha,\beta}=\left(\begin{array}{c} v_{\ell, 0}^{\alpha,\beta} \\ v_{\ell, 1}^{\alpha,\beta} \\ \vdots \\ v_{\ell, n}^{\alpha,\beta}\end{array}\right)\,,
\end{equation}
where $a$ labels the spectrum of the UV states and $c_a>0$. This imply vector $\vec{g}_k$ lie inside the convex hull of the super polynomial vector $\vec{\mathcal{P}}_{\ell}^{\alpha,\beta}$. Due to positivity properties of $\vec{\mathcal{P}}_{\ell}^{\alpha,\beta}$, its convex hull is a cyclic polytope. The boundary of cyclic polytope is constructed by adjacent pair of $\vec{\mathcal{P}}_{\ell}^{\alpha,\beta}$, inside the convex hull have
\begin{equation}
\langle\vec{g}_k, \vec{\mathcal{P}}_i, \vec{\mathcal{P}}_{i+1}...\vec{\mathcal{P}}_j, \vec{\mathcal{P}}_{j+1}\rangle>0.
\end{equation}
We use cyclic polytope constrain bound vector $\vec{g}_k$ in each order $k$ and comparing the difference results between $\mathcal{N}=0\sim4$.
\subsection{Expansion}
To expansion $k=2$, we use cyclic-polytope condition constrain $\vec{g}_{2}=(1,g_{2,1}/g_{2,0},g_{2,2}/g_{2,0})$ in projective space. Under the crossing symmetry, $\vec{g}_{2}$ is only one variable vector $\vec{g}_{2}=(1,g_{2,1}/g_{2,0},1)$. We consider super polynomials in each $\mathcal{N}$, and list the most constrained condition from cyclic-polytope for $g_{2,1}/g_{2,0}$:
\begin{eqnarray}
\nonumber \mathcal{N}=0 :&\quad \langle\vec{g}_2,\vec{v}_3,\vec{v}_4\rangle\geq0\quad\rightarrow\quad &0\leq g_{2,1}/g_{2,0} \leq44/7\\
\nonumber \mathcal{N}=1 :&\quad \langle\vec{g}_2,\vec{v}_2,\vec{v}_3\rangle\geq0\quad\rightarrow\quad&0\leq g_{2,1}/g_{2,0} \leq16/3\\
\nonumber \mathcal{N}=2 :&\quad \langle\vec{g}_2,\vec{v}_2,\vec{v}_3\rangle\geq0\quad\rightarrow\quad&0\leq g_{2,1}/g_{2,0} \leq22/5\\
\nonumber \mathcal{N}=3 :&\quad \langle\vec{g}_2,\vec{v}_1,\vec{v}_2\rangle\geq0\quad\rightarrow\quad&0\leq g_{2,1}/g_{2,0} \leq7/2\\
\mathcal{N}=4 :&\quad \langle\vec{g}_2,\vec{v}_1,\vec{v}_2\rangle\geq0\quad\rightarrow\quad&0\leq g_{2,1}/g_{2,0} \leq8/3.
\end{eqnarray}
We find that with higher order $\mathcal{N}$, $g_{2,1}/g_{2,0}$ is bounded in tighter region. The result also show in fig.\ref{N2}.
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{N2.pdf}
\caption{Cyclic-polytope condition on $g_{2,1}/g_{2,0}$.}
\label{N2}
\end{subfigure}
\begin{subfigure}[b]{0.45\linewidth}
\includegraphics[width=\linewidth]{N3.pdf}
\caption{Cyclic-polytope condition on $g_{3,1}/g_{3,0}$.}
\label{N3}
\end{subfigure}
\caption{Cyclic-polytope condition. $\mathcal{N}$ label R-symmetry group}
\label{N45}
\end{figure}
The same procedure can go to $k=3$ expansion. In $k=3$, $\vec{g}_3=(1,g_{3,1}/g_{3,0},g_{3,2}/g_{3,0},g_{3,3}/g_{3,0})$ still be an one variable vector $(1,g_{3,1}/g_{3,0},g_{3,1}/g_{3,0},1)$. One can read out cyclic-polytope constrain and have upper bound for $\{\mathcal{N}{=}0,\mathcal{N}{=}1,\mathcal{N}{=}2,\mathcal{N}{=}3,\mathcal{N}{=}4\}=\{42/5,15/2,20/3,6,6\}$ as show in fig.\ref{N3}. The region become smaller with increasing $\mathcal{N}$. In $k=3$, $\mathcal{N}=3$ and $\mathcal{N}=4$ share same boundary. It is not necessary for them have same boundary, one can find difference boundaries in others $k$.
In $k=4$ and $k=5$, there are both two independent variables. We show the constrain in fig.\ref{N45}.
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{0.3\linewidth}
\includegraphics[width=\linewidth]{N4.pdf}
\caption{Cyclic-polytope condition on $(g_{4,1}/g_{4,0},g_{4,2}/g_{4,0})$}
\end{subfigure}
\begin{subfigure}[b]{0.3\linewidth}
\includegraphics[width=\linewidth]{N5.pdf}
\caption{Cyclic-polytope condition on $(g_{5,1}/g_{5,0},g_{5,2}/g_{5,0})$}
\end{subfigure}
\caption{Cyclic-polytope condition. $\mathcal{N}$ label R-symmetry group.}
\label{N45}
\end{figure}
The region become smaller is reflected the fact, the decomposition of $\mathcal{N}$ super polynomial is positive sum of $\mathcal{N}{-}1$ polynomials. It can be easily show from recurrence relation
\begin{equation}
\begin{split}
\frac{\partial}{\partial\eta^{A}_{2}}\frac{\partial}{\partial\eta^{A}_{4}}\mathcal{P}^{\mathcal{N}}_{\{H_{i}\},S}\vert_{\eta^{A}=0}=m^{2}\mathcal{P}^{\mathcal{N}-1}_{\{h_{i}\},S+\frac{1}{2}}+\frac{m^{2}(S-H_{3}+H_{4})(S-H_{1}+H_{2})}{\left( 2S\right) \left( 2S+1\right) }\mathcal{P}^{\mathcal{N}-1}_{\{h_{i}\},S-\frac{1}{2}}\, .
\end{split}
\end{equation}
RHS have positive coefficients when we identify $H_1=H_3, H_2=H_4$ which means $\mathcal{N}$ super polynomial is living inside of $\mathcal{N}{-}1$ convex hull. As we know the boundary of $\mathcal{N}$ convex hull are living inside of $\mathcal{N}{-}1$, the $\mathcal{N}$ space must smaller than $\mathcal{N}{-}1$.
\section{Conclusion}
In this paper, we proposed an super Poincare invariant expansion basis for the residue and discontinuity of four-point amplitudes in four dimensions. These super spinning polynomials are given in terms of algebraic Jacobi functions, whose recurrence relations reflect the decomposition of the SUSY multiplet in terms of components. These polynomials were identified through gluing of three-point super amplitudes constructed out of massive and massless on-shell super space.
As an application, we use dispersion relations to derive bounds on EFT coefficients. By expanding the imaginary part of the amplitude on these super spinning polynomials, we derive bounds that reflect the underlying supersymmetric UV completion.
\section{Acknowledgements}
We would like to thank Yu-tin Huang for enlightening discussions. J-Y Liu and Z-M You are supported by MoST Grant No. 106-2628-M-002-012-MY3.
\label{sec:Conclusion}
|
\section{Conclusion}
We have introduced the first learning-based approach to estimating the 6D pose of a self-occluded object immersed in a scene. Our method is intuitive, relying on an instance segmentation module followed by a pose estimation one. As evidenced by our empirical results, this strategy is effective, outperforming the state-of-the-art traditional and learning-based methods, which so far have mostly focused on registering partial views of the same object or scene. We believe that the problem we have tackled here is central in many practical applications. In particular, in the future, we will aim to deploy our approach for LiDAR-based registration in the context of non-cooperative rendezvous in space.
\section{Experiment}
\input{fig/mAP_noback}
\input{fig/mAP_withback}
We first compare our approach to the state-of-the-art methods on the task of estimating the pose of a self-occluded object without a scene background, and then turn to the more challenging scenario where the object is immersed in context.
Finally we analyze the influence of the different components of our approach.
\subsection{Self-Occluded Object Registration without Background}
\label{no_background}
\paragraph{Dataset.} For this experiment, we use the auto-aligned ModelNet40 dataset~\cite{Wu15,Sedaghat16}. This dataset contains mesh models for $40$ object categories. The point clouds are normalized in the range $[-1, 1]$ on each axis. As in~\cite{Wang19f}, we split the data into $9,843$ training and $2,468$ testing mesh models.
We use the mesh models to render depth maps given camera viewpoints and intrinsic parameters. Specifically, we treat the full point clouds corresponding to the meshes as source sets $\mathcal{X}$ and the resulting depth maps as target sets $\mathcal{Y}$. We use the look-at method~\footnote{https://www.scratchapixel.com/lessons/mathematics-physics-for-computer-graphics/lookat-function} to place the camera, and set the distance between the camera and the center of the object to be $0.65$. We then randomly sample the elevation in $[15^{\circ}, 75^{\circ}]$ and azimuth in $[0^{\circ}, 89^{\circ}]$. To deal with the large number of points, following~\cite{Choy20,Yang20}, we first voxelize the point clouds with a voxel size of $0.05 \times (\frac{\sqrt2}{2})$, and then randomly sample the desired number of points from the voxelized point clouds.
\vspace{-.4cm}
\paragraph{Evaluation metrics.} We report the rotation error and translation error between the predictions $\hat{R},\hat{\textbf{t}}$ and the ground truth $R_{gt},\textbf{t}_{gt}$. These errors are computed as
\begin{equation}
\begin{aligned}
&E_{rot}(\hat{R},R_{gt}) = arccos\frac{trace(\hat{R}^{\top}R_{gt}) - 1)}{2} \;,\\
&E_{trans}(\hat{\textbf{t}},\textbf{t}_{gt}) = \norm{\hat{\textbf{t}} - \textbf{t}_{gt}}^{2}_{2} \;.
\end{aligned}
\end{equation}
We summarize the results in terms of mean average precision (mAP) of the estimated relative pose under varying accuracy thresholds, as in ~\cite{Yi18}. For the rotation, we use the three thresholds $[5^{\circ}, 10^{\circ}, 15^{\circ}]$, and for the translation, we set the thresholds to be $[1\times 10^{-3}, 5\times 10^{-3}, 1\times 10^{-2}]$.
\vspace{-0.4cm}
\paragraph{Implementation details.}
We implement our pose estimation network in Pytorch~\cite{Paszke17} and train it from scratch. We use the Adam optimizer~\cite{Kingma15} with a learning rate of $10^{-3}$ and mini-batches of size $20$, and train the network for $40,000$ iterations. We set the number of points for $\mathcal{X}$ and $\mathcal{Y}$ to be 1024 and 768, respectively, encoding the fact that $\mathcal{Y}$ only contains a visible portion of $\mathcal{X}$. We use the same parameters for both Ours\_SoftMax and Ours\_OT. For the OT layer, we use $k = 50$ iterations and set $\lambda = 0.5$. Training was performed on one NVIDIA RTX8000 GPU.
To build the ground-truth assignment matrix $\mathcal{M}$ for Ours\_SoftMax and Ours\_OT, we transform $\mathcal{X}$ using the ground-truth transformation $\mathcal{T}$, giving us $\tilde{\mathcal{X}}$. We then compute the pairwise Euclidean distance matrix between $\tilde{\mathcal{X}}$ and $\mathcal{Y}$, which we threshold to $0.05$ to obtain a correspondence matrix $\mathcal{M} \in \{0, 1\}$. For Ours\_OT, we augment $\mathcal{M}$ with an extra row and column acting as outlier bins to obtain $\mathcal{\bar{M}}$. The points without any correspondence are treated as outliers, and the corresponding positions in $\mathcal{\bar{M}}$ are set to one. This strategy does not guarantee a bipartite matching, which we address using a forward-backward check.
\vspace{-.5cm}
\paragraph{Registration results.}
We compare our approach to PRNet~\cite{Wang19f}, DCP~\cite{Wang19e}, ICP~\cite{Besl92a}, FPFH-RANSAC~\cite{Rusu09}, FGR~\cite{Zhou16} and TEASER++~\cite{Yang19,Yang20}. Note that, despite our best effort, we were unable to successfully train DGR~\cite{Choy20} on our dataset. We believe this to be due to the different nature of the task addressed by DGR, i.e., scene to scene registration. This is further evidenced by the fact that, in~\cite{Choy20}, the DGR authors mentioned that they were unable to train DCP on the 3DMatch dataset~\cite{Zeng17b}, whereas we underwent not problems with DCP in our scenario. Similarly, we were unable to train RPMNet~\cite{Yew20} on our data, encountering the instabilities reported by the RPMNet authors, with training crashing after 10 epochs because of an error during SVD.
\input{fig/syn_qualitative}
For the PRNet, we use the pretrained partial-to-partial model provided by the authors, which we fine-tune on our dataset with a learning rate of $0.0001$ for $10$ epochs, where we observed convergence. For DCP, we use the pretrained DCP-v2 model (clean version), and similarly fine-tune it on our dataset with a learning rate of $0.001$ for $30$ epochs corresponding to convergence.
For ICP, FPFH-RANSAC and FGR, we use the Open3D~\cite{Zhou18} implementations. For TEASER++ we use the official implementation
The results are summarized in Fig.~\ref{mAP_noback}. Our methods outperforms all the baselines in all settings, with our two variants yielding similar results.
TEASER++ achieves the best result among the baselines for a threshold of $5^{\circ}$, but is outperformed by
DCP for $10^{\circ}$ and $15^{\circ}$, which justifies our choice of DCP as our base network. We believe this to illustrate a downside of exploiting handcrafted features, such as FPFH, as TEASER++ does; while such features are reliable for some objects, they do not generalize well to all of them. By contrast, deep networks can learn to leverage different sources of information for different objects. We attribute the relatively poor performance of PRNet to the fact that it was designed specifically for object-to-object registration.
\subsection{Self-Occluded Object Pose Estimation with Background}
\label{with_background}
Let us now turn to the more challenging scenario where the object is immersed in a scene.
\vspace{-0.5cm}
\paragraph{Dataset.}
To this end, we generate a dataset using the BlenderProc~\cite{Denninger19} photo-realistic renderer discussed in Section~\ref{sec:training}.
Specifically, we create cubic scenes of length $4$ with random texture images from the CC0textures dataset~\footnote{https://cc0textures.com}. In each scene, we place an object from the ModelNet40 dataset~\cite{Wu15} with a randomly sampled elevation and azimuth. We then render depth maps from given cameras, together with a mask of the object's visible portion, and the object pose and bounding box. We follow the same procedure to generate the training and test data, using the same object splits and the same sampling range for the look-at camera model as in Section~\ref{no_background}.
The main difference between this experiment and the previous one is that the algorithms must identify the object in the scene.
The whole scene contains $640 \times 480$ points, which after voxelization still leaves between $15,000$ and $300,000$ points. Among these, the object itself covers only $1,000$ to $9,000$ points, depending on the structure of the mesh model.
This corresponds to an inlier rate ranging from $0.5\%$ to $23.5\%$, with $79.41\%$ of the samples having an inlier rate less than $5\%$.
\vspace{-0.6cm}
\paragraph{Implementation details.}
\input{fig/real_scene}
For our instance segmentation network, we use a pre-trained Resnet101 backbone~\cite{He16}.
Since the input to our model is not an image, we replicate the depth map three times to match the number of input channels. As output, we keep only one category, which is used to classify each pixel as belonging to the object or not. We train the resulting network for $100,000$ iterations with mini-batches of size $16$, and using SGD with a learning rate of $10^{-3}$.
We then train our pose estimation module in the same manner as in Section~\ref{no_background}, but discarding the points that do not belong to the object using the ground-truth mask. This, of course, can only be achieved during training, and leads to a mismatch between training and testing time, since, at testing time, we cannot expect our instance segmentation module to always give perfect results. To handle this, at test time, we erode the boundary of the predicted masks with a $3\times 3$ kernel. Furthermore, we remove all the points that lie outside a sphere centered on the object and of radius $1.2$, accounting for the fact that the object coordinates were normalized in the range $[-1, 1]$.
\vspace{-0.4cm}
\paragraph{Registration results.}
We compare our approach to the same baselines as in Section~\ref{no_background}, using the same hyper-parameter values to train them. The only exception to this is for TEASER++, for which we limit the number of points in the target point set to a maximum of $50,000$ because we observed too large target point sets
to produce a segmentation fault in the code.
Note that, prior to our work, no method was designed to explicitly estimate the 6D pose of a self-occulded object observed in a scene. Therefore, the baselines directly tackle the 3D registration task from the raw depth maps. By contrast, our segmentation module outputs a mask that facilitates the task of the registration one.
The result are summarized in Fig.~\ref{mAP_withback}, with a few examples in Fig.~\ref{fig:syn_qualitative}. Note that our methods outperform the baselines by a large margin. Since the DCP and PRNet directly act on the raw depth maps, the point-cloud downsampling process, required to make inference tractable, leaves only very few points belonging to the object, thus leading to the failure of these methods. By contrast, TEASER++, which was designed to handle such large amounts of outliers, yields reasonable results, nevertheless clearly outperformed by our learning-based strategy.
\input{fig/mAP_real}
\subsection{Dealing with Real Data}
To evidence that our approach generalizes to real depth maps, we use the TUD-L dataset~\cite{Hodan18}, which contains training and testing image sequences that show three moving objects under eight lighting conditions. We retrain our complete model with OT layer from scratch on this dataset, keeping the same hyper-parameter values as for the previous experiments.
To compare our method to the baselines, we select 2000 images from each testing sequence. We restrict our baseline evaluation to the methods that gave reasonable results in the experiment of Section~\ref{no_background}, i.e., TEASER++ and FGR. As can be shown in Fig.~\ref{mAP_real}, our method also outperforms the baselines by a large margin in this scenario.
In Fig.~\ref{real_scene}, we show a few qualitative results obtained with our framework, evidencing the accuracy of our predictions.
\input{table/inference_time}
\subsection{Method Analysis}
\paragraph{Inference Time.}
In Table~\ref{tab:inference_time}, we compare the inference time of our registration method to that of the baselines for different point set sizes. For this comparison to be fair, we ran all methods on a desktop computer with an Intel(R) Core(TM) i7-7700K CPU @ 4.20GHz, an Nvidia GTX 2080 Ti GPU, and 32GB memory.
Theoretically, the inference time for Ours\_SoftMax time is the same as DCP, thus we omit it. Ours\_OT yields run-times comparable to FGR, while being slightly faster than PRNet. Note that the run-times reported in Table~\ref{tab:inference_time} for our approach do not include the instance segmentation module, which takes 0.0285s to process an image of size $480\times 640$.
\vspace{-.3cm}
\paragraph{Effectiveness of the Optimal Transport Layer.} Note that the best results with our approach were obtained without the optimal transport layer. Here, we illustrate that this layer is nonetheless important in scenarios where the training data is noisy. To this end, we replace our rendering on-the-fly procedure with BlenderProc during training so as to generate noisy training observations. To be specific, we use the ground-truth mask to obtain the depth of the object and transform it to the target point cloud, and then jitter the rotation in the range $[0, 45^{\circ}]$ and the translation in the range $[-1, 1]$ on each axis. Only Our\_OT could be trained on the resulting dataset, thanks to its outlier bins. We provide the corresponding results as Ours\_OT\_BlenderProc in Fig.~\ref{mAP_withback}.
\section{Introduction}
\input{fig/teaser}
3D registration aims to determine the rigid transformation, i.e., 3D rotation and 3D translation, between two 3D point sets. The traditional approach to this problem is the Iterative Closest Point (ICP) algorithm~\cite{Besl92}. In its vanilla version, this iterative algorithm easily gets trapped in poor local optima. While globally-optimal solutions~\cite{Yang15,Zhou16} have been proposed to remedy this, they
lack robustness to noise, thus greatly reducing their practical applicability.
Recently, learning-based methods~\cite{Aoki19,Wang19e} were shown to outperform the previous traditional, optimization-based techniques.
While these initial methods were designed to work under the unrealistic assumption that both point sets are fully observed, even at test time, follow-up works focused on the more practical partial-to-partial registration scenario~\cite{Wang19f,Yew20}.
Nevertheless, these techniques still assume the observations to depict a point-cloud in full 3D.
In practice, however, many sensors, such as depth cameras and LiDARs, only provide 2.5D measurements. Furthermore, when focusing on objects, as we do, and not entire scenes, as in, e.g.,~\cite{Choy20}, one needs to handle the measurements coming from the object's context, i.e., the scene itself, which are irrelevant to estimate the object 6D pose but are nonetheless captured by the sensor.
Only few methods have attempted to estimate the 6D pose of an object in a scene from depth data~\cite{Vidal18,Yang19,Yang20}.
To this end, these methods extract handcrafted representations of the 3D points, such as Point Pair Features (PPF)~\cite{Vidal18} or Fast Point Feature Histograms (FPFH)~\cite{Yang19,Yang20}
and handle the presence of contextual information by casting the problem as an outlier removal task.
As such, they do not truly focus on the 2.5D nature of the measurements, and, for example, the effectiveness of the most recent one, TEASER++~\cite{Yang20},
was only demonstrated in a restricted scenario, where the reference object point-cloud was in one-to-one correspondence with that observed in the scene context.
Furthermore, these methods all rely on handcrafted features, and, to the best of our knowledge, 6D pose estimation of a self-occluded object in context, as illustrated by Fig.~\ref{fig:teaser}, has never been tackled in a deep learning fashion.
In this paper, we address this by introducing a deep learning framework for 3D registration of self-occluded objects in context. Our approach consists of an instance segmentation module to separate the object from the surrounding scene, followed by a 6D pose estimation module able to handle the mismatch between the complete 3D model and the observed, self-occluded point-set. Our method
works in a one-shot fashion; at test time, we estimate the object pose by a single forward pass through the network, without the need for an expensive iterative procedure. This makes our approach amenable to practical applications relying on 2.5D sensors and requiring real-time performance, such as autonomous navigation and robotics manipulation.
Our contributions can be summarized as follows:
\vspace{-1.5mm}
\begin{itemize}
\itemsep-0.2em
\item We tackle, for the first time, the challenging problem of estimating the 6D pose of a self-occluded object in context in a learning-based manner.
\item We develop a one-shot framework to address this task, consisting of an instance segmentation and a pose estimation module.
\item We design a memory- and computation-effective training method based on on-the-fly rendering, which prevents the need for expensive offline rendering and storing of the generated data, as required by standard procedures, such as BlenderProc~\cite{Denninger19}, while giving us access to virtually infinite amounts of training data.
\end{itemize}
\vspace{-1mm}
We demonstrate the effectiveness of our approach using the ModelNet40 benchmark, from which we generate depth maps to match our application scenario. Our approach significantly outperforms the state-of-the-art traditional and leaning-based registration techniques~\cite{Wang19e,Wang19f}, including our main competitor TEASER++~\cite{Yang19,Yang20}. We will make our code publicly available.
\section{Methodology}
Let us now introduce our approach to estimating the 6D pose of an object, represented with a reference point cloud in full 3D, from 2.5D observations of this object in a scene. To address this challenging scenario, we develop the deep learning framework depicted by Fig.~\ref{fig:pipeline}. It relies on two main modules, an instance segmentation network and a pose estimation network, which we discuss in detail below.
\subsection{Instance Segmentation}
To tackle the realistic scenario where the object is immersed in a scene, we first process the input depth map with an instance segmentation module. Specifically, we build on the YOLACT~\cite{Bolya19} framework that has proven highly effective for image-based instance segmentation. To exploit this framework in our scenario, we replicate the depth map $I_{depth}$ three times, so as to obtain 3 channels, as an RGB image, which allows us to benefit from the pre-training of the model on an image dataset. Furthermore, instead of predicting multiple object categories, we train our model to discriminate between object and background, which we found to be better suited to handle depth that does not carry appearance information. Note that we nonetheless train our model using multiple object categories. Finally, after non-maximum suppression, we only keep the bounding box with the highest score, which we observed to work well in practice, as will be shown by our experiments, and as evidenced qualitatively in Fig.~\ref{synthetic_scene}.
Ultimately, our instance segmentation module produces a mask image, expressed as
\begin{equation}
I_{mask} = \phi_{seg}(I_{depth})\;.
\end{equation}
We then use $I_{mask}$ to isolate the target point cloud, which we input to the pose estimation network described below.
\input{fig/pose_net}
\input{fig/synthetic_scene}
\vspace{-.2cm}
\subsection{Pose Estimation}
Let us now turn to the task of estimating the 6D pose of the object of interest. To this end, let $\mathcal{X} \in \mathbb{R}^{M \times 3}$ and $\mathcal{Y} \in \mathbb{R}^{N \times 3}$ be two sets of 3D points sampled from the same object surface. We typically refer to $\mathcal{X}$ as the source point set and to $\mathcal{Y}$ as the target point set. We obtain the source point set $\mathcal{X}$ by uniform sampling from the mesh model, and target one $\mathcal{Y}$ from $I_{depth}$ and $I_{mask}$, assuming known camera intrinsic parameters. Registration then aims to find a rigid transformation $\mathcal{T}$ that aligns $\mathcal{X}$ to $\mathcal{Y}$. In this work, we focus on the case where the object is self-occluded. To be specific, we assume that each point in $\mathcal{Y}$ has a corresponding point in $\mathcal{X}$, but not the opposite.
We build our pose estimation module on DCP~\cite{Wang19e}, extending it to handle the partial correspondence scenario. As illustrated in Fig.~\ref{pose_net}, we rely on the DCP-v2 design, which consists of a DGCNN~\cite{Wang18b} followed by a Transformer~\cite{Vaswani17}.
Specifically, the DGCNN takes a point set as input, constructs a k-NN graph from it, and then extracts point-wise features via standard convolutions on this graph, encoding diverse levels of context by max-pooling the local features and concatenating the resulting representations to the point-wise ones.
Let $\theta^x$, resp. $\theta^y$, be the final feature matrix, i.e., one $P$-dimensional feature vector per 3D point, for $\mathcal{X}$, resp. $\mathcal{Y}$. The transformer then learns a function $\phi : \mathbb{R}^{M\times P} \times \mathbb{R}^{N\times P} \rightarrow \mathbb{R}^{M\times P}$, that combines the information of the two point sets. Ultimately, this produces descriptor matrices $\mathbf{f}^x$, resp. $\mathbf{f}^y$, for $\mathcal{X}$, resp. $\mathcal{Y}$, written as
\begin{equation}
\mathbf{f}^x = \theta^x + \phi(\theta^x, \theta^y), \;\;\;
\mathbf{f}^y = \theta^y + \phi(\theta^y, \theta^x)\;.
\end{equation}
Given these matrices, we then form a score map $\mathcal{S} \in \mathbb{R}^{M\times N}$ by computing the similarity between each source-target pair of descriptors. That is, we compute the $(i,j)$-th element of $\mathcal{S}$ as
\begin{equation}
\mathcal{S}_{i,j} = <\mathbf{f}^x_i, \mathbf{f}^y_j>, \;\;\forall (i, j) \in [1,M]\times [1,N]\;,
\end{equation}
where $<\cdot, \cdot>$ is the inner product, and $\mathbf{f}^x_i, \mathbf{f}^y_j \in \mathbb{R}^P$.
In DCP~\cite{Wang19e}, this score map is passed through a row-wise softmax so as to obtain correspondences. These correspondences are then processed within the network via an SVD layer to solve the Procrustes problem, and the resulting rigid transformation is compared to the ground-truth one with a mean squared error (MSE) loss. Here, when $I_{mask}$ can be assumed to be clean, we keep the softmax because each point in $\mathcal{Y}$ should be visible in $\mathcal{X}$. However, we remove the SVD layer, which, if not carefully designed, can make the training process unstable~\cite{Dang18,Dang20,Wang19g,Yew20}. Instead, we impose direct supervision on the output of the softmax. To this end, let $\mathcal{M} \in \{0, 1\}^{M \times N}$ be the matrix of ground-truth correspondences, with a $1$ indicating a correspondence between a pair of points. Such correspondences can be estimated using the ground-truth transformation matrix, as discussed in more detail in Section~\ref{no_background}. We then express our loss function as the negative log-likelihood
\begin{equation}
\mathcal{L}(\mathcal{P}, \mathcal{M}) = \frac{- \sum\limits_{i =1}^{M}\sum\limits_{j = 1}^{N} (\log \mathcal{P}_{i,j})\mathcal{M}_{i, j}}{\sum\limits_{i = 1}^{M}\sum\limits_{j = 1}^{N} \mathcal{M}_{i, j}}\;,
\label{eq:nll}
\end{equation}
where $\mathcal{P}$ is the output of the softmax, and where the denominator normalizes the loss value so that different training samples containing different number of correspondences have the same influence in the overall empirical risk.
\vspace{-.4cm}
\paragraph{Dealing with noisy target points.}
\input{table/algorithm}
Although the softmax works well when the points in $\mathcal{Y}$ all have a corresponding point in $\mathcal{X}$, in practice, $\mathcal{Y}$ may include noise, due, for example, to an imperfect $I_{mask}$ prediction.
To address this, we replace the softmax with an optimal transport layer~\cite{Yew20,Sarlin19}
including an outlier bin to handle the noise.
Specifically, we extend the score matrix $\mathcal{S}$ by one row and one column to form an augmented score matrix $\mathcal{\bar{S}}$. The values at the newly-created positions in $\mathcal{\bar{S}}$ are set to
\begin{equation}
\mathcal{\bar{S}}_{i, N + 1} = \mathcal{\bar{S}}_{M + 1, j} = \mathcal{\bar{S}}_{M + 1, N + 1} = \alpha,
\end{equation}
$\forall i \in [1, M], \;\forall j \in [1, N]$, where $\alpha \in \mathbb{R}$ is a fixed parameter, which set to be $0.01$ in practice. The values at the other indices directly come from $\mathcal{S}$. Given the augmented score map $\mathcal{\bar{S}}$, we aim to find a partial assignment $\mathcal{\bar{P}} \in \mathbb{R}^{(M+1)\times (N+1)}$, defining correspondences between the two point sets, extended with the outlier bins. Let $\mathbf{U} (\mathbf{a}, \mathbf{b})$ be the set of probability matrices defined as
\begin{equation}
\left\{\mathcal{\bar{P}} \in \mathbb{R}_{+}^{(M+1)\times (N+1)}: \mathcal{\bar{P}} \mathbbm{1}_{M+1} = \mathbf{a} \: \text{and}\: \mathcal{\bar{P}}^{\top} \mathbbm{1}_{N+1} = \mathbf{b} \right\}\;,
\end{equation}
where $\mathbf{a} = [\mathbbm{1}_{M}^{\top}, N]^{\top}$, and $\mathbf{b} = [\mathbbm{1}_{N}^{\top}, M]^{\top}$, with $\mathbbm{1}_{M} = [1, 1, ..., 1]^{\top}\in \mathbb{R}^{M}$. Then, from the optimal transport theory~\cite{Peyre19,Cuturi13}, an assignment probability matrix $\mathcal{\bar{P}}$ can be obtained by solving
\begin{equation}
\min_{\mathcal{\bar{P}} \in \mathbf{U} (\mathbf{a}, \mathbf{b})} \langle \mathcal{\bar{S}}, \mathcal{\bar{P}} \rangle - \lambda E(\mathcal{\bar{P}})\;,
\label{entropy_func}
\end{equation}
where $\langle\cdot,\cdot\rangle$ is the Frobenius dot product and $E(\cdot)$ is an entropy regularization term defined as $E(\mathcal{\bar{P}}) = - \displaystyle\sum_{i, j} \mathcal{\bar{P}}_{i, j} \left(\log(\mathcal{\bar{P}}_{i, j}) - 1\right)$. In practice, this optimization problem can be solved by using the log-domain Sinkhorn algorithm, which we summarized in Algorithm~\ref{alg:sinkhorn}, where the matrix operator $logsumexp(\mathcal{A}) = log(exp(\mathcal{A}_{1,1}) + ... + exp(\mathcal{A}_{i,j}) + ... + exp(\mathcal{A}_{M,N}))$. Since all operations performed by this algorithm are differentiable, the training errors can be backpropagated to the rest of the network. We then still use the negative log-likelihood of Eq.~\ref{eq:nll} as loss function, but replacing $\mathcal{M}$ with $\mathcal{\bar{M}}$, which contains an extra row and column acting as outlier bins.
\subsection{Training via On-the-fly Rendering}
\label{sec:training}
To train our network, we follow a two-stage procedure, starting with the instance segmentation module and following with the pose estimation one.
Unfortunately, there exist virtually no real dataset providing high-quality depth data for 6D object pose estimation in context.
For instance, the datasets in the BOP challenge~\cite{Hodan18} mostly provide relatively low-quality depth maps, with very sparse depth measurements for the object. Such low-quality depth observations, however, are not representative of the resolutions that modern sensors can now achieve~\cite{Lu20}.
To better reflect the progress of these sensors, in this work, we rely on BlenderProc~\cite{Denninger19} to generate high-quality synthetic data, as will be discussed in more detail in Section~\ref{with_background}. However, even such synthetic data remains imperfect at the object boundaries, mixing the object depth with the background one, as shown in Fig.~\ref{fig:syn_qualitative}. Such outliers at the boundaries affect the training process, and we therefore use BlenderProc to train our instance segmentation module but not our pose estimation one. Instead, and to simultaneously benefit from having access to virtually infinite amounts of training data, we follow an on-the-fly rendering strategy. Specifically, we use Pytorch3D~\cite{ravi20} to generate scenes that contain individual objects without background, thus fulfilling the pose estimation network's assumption and precluding the presence of outliers at the object boundaries. The use of Pytorch3D allows us to interface the rendering process with the network training, thus preventing the need to generate and store a fixed set of training samples in a pre-processing stage.
\begin{comment}
\paragraph{Training the instance segmentation module.} To train our instance segmentation module, we use the BlenderProc~\cite{Denninger19} photo-realistic renderer. Specifically, we create cubic scenes with texture images from the dataset CC0textures~\footnote{https://cc0textures.com}, within which we place an object mesh model from the ModelNet40 dataset~\cite{Wu15} with a randomly sampled elevation and azimuth. We then render depth map from given cameras, together with a mask of the object's visible portion, and the object pose and bounding box. The test data used in our experiments to evaluate all algorithms was also generated in this manner.
\paragraph{On-the-fly rendering to train the pose estimation module.}
We do not use BlenderProc to train our pose estimation module because the training data generated in this manner contains outliers. To be specific, the resulting masks are inaccurate, and tend to incorporate background information at the objects' boundaries, as shown in Fig.~\ref{fig:syn_qualitative} \zd{the depth at the boundary of the object mixed up with the background's depth.}
To address this problem, and simultaneously benefit from having access to virtually infinite amounts of training data, we follow an on-the-fly rendering strategy.
Specifically, we use Pytorch3D~\cite{ravi20} to generate scenes that contain only individual objects, without background, thus fulfilling the pose estimation network's assumption. We then render depth maps of these objects using camera and object pose parameters sampled from the same range as with BlenderProc.
\end{comment}
\section{Related Work}
\textbf{Traditional point cloud registration.}
ICP is the best-known algorithm for solving the point cloud registration problem. It comprises two steps: One whose goal is to find the closest target point for each source point to generate 3D-3D correspondences, and the other that computes the rigid transformation from these correspondences by solving a least-square problem. These two steps are repeated until a termination condition is satisfied. Several variants, such as Generalized-ICP~\cite{Segal09} and Sparse ICP~\cite{Bouaziz13}, have been proposed to improve robustness to noise and mismatches, and we refer the reader to~\cite{Pomerleau15,Rusinkiewicz01} for a complete review of ICP-based strategies. The main drawback of these methods is their requirement for a reasonable initialization to converge to a good solution. Only relatively recently has this weakness been addressed by the globally-optimal registration method Go-ICP~\cite{Yang15}. In essence, this approach follows a branch-and-bound strategy to search the entire 3D motion space $SE(3)$. Motivated by this, other approaches to finding a global solution have been proposed, via, e.g., Riemannian optimization~\cite{Rosen19}, convex relaxation~\cite{Maron16}, and mixed-integer programming~\cite{Jzatt20}. While globally optimal, these methods all come at a much higher computational cost than vanilla ICP. This was, to some degree, addressed by the Fast Global Registration (FGR) algorithm~\cite{Zhou16}, which leverages a local refinement strategy to speed up computation. While effective, FGR still suffers from the presence of noise and outliers in the point sets, particularly because, as vanilla ICP, it simply relies on 3D point-to-point distance to establish correspondences. In principle, this can be addressed by designing point descriptors that can be more robustly matched. Over the years, several works have tackled this task, in both a non learning-based~\cite{Johnson99,Rusu08,Rusu09} and learning-based~\cite{Zeng17b,Khoury17} fashion. Nowadays, however, these approaches are outperformed by end-to-end learning frameworks, which directly take the point sets as input.
\textbf{End-to-end learning with point sets.}
A key requirement to enable end-to-end learning-based registration was the design of deep networks acting on unstructured sets. Deep sets~\cite{Zaheer17} and PointNet~\cite{Qi17} constitute the pioneering works in this direction. They use shared multilayer perceptrons to extract high-dimensional features from the input point coordinates, and exploit a symmetric function to aggregate these features. This idea was then extended in PointNet++~\cite{Qi17a} via a modified sampling strategy to robustify the network to point clouds of varying density, in DGCNN~\cite{Wang18b} by building a graph over the point cloud, in PointCNN~\cite{Li18c} by learning a transformation of the data so as to be able to process it with standard convolutional layers, and in PCNN~\cite{Atzmon18} via an additional extension operator before applying the convolutions. While the above-mentioned works focused on other tasks than us, such as point cloud classification or segmentation, end-to-end learning for registration has recently attracted a growing attention. In particular, PointNetLK~\cite{Aoki19} combines the PointNet backbone with the traditional, iterative Lucas-Kanade (LK) algorithm~\cite{Lucas81} so as to form an end-to-end registration network; DCP~\cite{Wang19e} exploits DGCNN backbones followed by Transformers~\cite{Vaswani17} to establish 3D-3D correspondences, which are then passed through an SVD layer to obtain the final rigid transformation. While effective, PointNetLK and DCP cannot tackle the partial-to-partial registration scenario. That is, they assume that both point sets are fully observed, during both training and test time. This was addressed by PRNet~\cite{Wang19f} via a deep network designed to extract keypoints from each input set and match these keypoints. This network is then applied in an iterative manner, so as to increasingly refine the resulting transformation.
Similarly, RPM-Net~\cite{Yew20} builds on DCP, replacing its softmax layer with an optimal transport ones so as to handle outliers, and as PRNet, relies on an iterative strategy to refine the computed transformation.
In any event, the methods discussed above were designed to handle point-clouds in full 3D, and were thus not demonstrated for registration from 2.5D measurements. By contrast, DGR~\cite{Choy20}, which
uses a deep network to reject the outliers from an input set of correspondences, was shown to be applicable to depth data. Nevertheless, this was achieved in the context of registering two partial views of a scene, whereas we focus on estimating the 6D pose of an object captured in a scene.
To the best of our knowledge, only few methods have been proposed to address this challenging scenario. In particular,~\cite{Vidal18} relies on generating pose hypotheses via feature matching, followed by a RANSAC-inspired method to choose the candidate pose with the largest number of support matches.
Similarly, TEASER~\cite{Yang19} and its improved version TEASER++~\cite{Yang20} take putative correspondences obtained via feature matching as input and remove the outlier ones by an adaptive voting scheme.
By relying on handcrafted features designed for 3D point-clouds, these methods do not explicitly address the case of 2.5D measurements. This is what we address in this work, and, motivated by the great progress of deep networks to address the classical 3D registration scenario, introduce the first deep learning framework capable of estimating the 6D pose of a 3D object observed with a 2.5D sensor in a scene.
\section{Supplymentary}
In \Cref{tudl_1,tudl_2,tudl_3}, we provide additional qualitative results on the three test sequences of the TUD-L dataset~\cite{Hodan18}.
Furthermore, we provide qualitative results for every class in ModelNet40~\cite{Wu15,Sedaghat16} in \Cref{blenderproc_1,blenderproc_2,blenderproc_3,blenderproc_4,blenderproc_5,blenderproc_6,blenderproc_7,blenderproc_8}.
\clearpage
\clearpage
\clearpage
\clearpage
\clearpage
\clearpage
|
\section{Introduction}
Despite the fast development of neural architecture search (NAS) \cite{zoph2016neural} to aid network design in vision tasks like classification~\cite{tan2019efficientnet}, object detection~\cite{ghiasi2019fpn}, and segmentation~\cite{liu2019auto}, there has been an urging demand for faster searching algorithms. Early methods based on massive evaluation of candidate models \cite{zoph2016neural,tan2018mnasnet,howard2019searching} require unaffordable costs (typically 2K GPU days). In the light of weight-sharing mechanism introduced in SMASH \cite{brock2017smash}, a variety of low-cost approaches have emerged \cite{bender2018understanding,pham2018efficient,liu2018darts}. Differentiable architecture search \cite{liu2018darts} has taken the dominance with a myriad of follow-up works \cite{wu2018fbnet, cai2018proxylessnas,xie2018snas,dong2019searching,chen2019progressive,zela2019understanding}. In this paper, we investigate the single-path based branch \cite{dong2019searching} for its efficiency that only requires a single path in memory.
Unlike many DARTS variants that have to perform all candidate operations,
single-path based methods~\cite{stamoulis2019single,dong2019searching,xie2018snas}, also termed as memory-efficient NAS\footnote{In this paper, we interchangeably use the term `single-path' and `memory-efficient' for describing such kinds of NAS methods.}, are developed to sample and activate one operation from each edge. Specifically, the Gumbel-Softmax reparameterization tricks~\cite{jang2016categorical,maddison2016concrete} are employed to achieve differentiable and memory-efficient search \cite{wu2018fbnet,dong2019searching,xie2018snas}.
In this paper, we discover that existing single-path based searching methods yet suffer a severe instability problem, upon which we propose a novel algorithm to disentangle the search for optimal topology (i.e. edge connections) from operation choices. Specifically, in addition to parameters $\bm{\alpha}$ representing the importance of operations, we involve topology parameters $\bm{\beta}$ to represent the relative importance of edges. A single-path architecture is induced by independently sampling candidate connections and operations based on $\bm{\beta}$ and $\bm{\alpha}$ respectively.
Moreover, to robustify the training of operation weights and architecture parameters, we propose a gradient accumulation strategy during the bi-level optimization. In a nutshell, our contributions are:
\textbf{1) Deep dive into the collapse in single-path based NAS.}
We discover that single-path based methods also suffer from the instability issue. Similar to the performance collapse problem in DARTS that has been discussed in recent works~\cite{zela2020understanding,chen2020stabilizing,xie2018snas,liang2019darts+}, the architectures searched by single-path based methods can also be dominated by parameterless operations, especially skip connections. This motivates us to scheme techniques to improve its stability.
\textbf{2) Consistent search and evaluation by disentangling topology search from operation selection.}
Independent topological distribution is designed to achieve consistent architecture between the search and inference stage, which separates the search for optimal topology from operations. To our best knowledge, this is the first work to achieve consistent search and evaluation for single-path based differentiable NAS, though there are counterpart techniques like edge normalization \cite{Xu2020PC-DARTS} developed in the full-path scenario.
\textbf{3) Robustifying bi-level optimization via gradient accumulation.} Compared with the full-path methods whereby all the operations are involved for training, the instability issue by the random sampling of operations in the single-path methods can be more pronounced. We propose to accumulate the gradients over sampling iterations to acquire more stable gradient estimation, for both architecture parameters and network weights. Our work is one of the first attempts to address the stability issue during the bi-level optimization of single-path based NAS.
\textbf{4) Strong performance while maintaining low memory cost.}
Our memory-efficient approach achieves state-of-the-art on various search spaces and datasets across \textbf{15} benchmarks.
We will provide the source code and hyperparameter settings to make our work reproducible.
\section{Related Work}
\textbf{Differentiable Neural Architecture Search.}
Similar to \cite{zoph2017learning,pham2018efficient} that uses a directed acyclic graph to represent a cell, DARTS~\cite{liu2018darts} constructs a cell-based supernet and further introduces architecture parameters to represent the importance of each operation. It formulates the searching process as a bi-level optimization problem \cite{anandalingam1992hierarchical}. DARTS proposes two types (first-order and second-order) of approximation that alternatively update network weights and architecture parameters with stochastic gradient descent. However, since the supernet subsumes all connections and operations within the search space, DARTS risks exhausting the memory when searching for a larger network. In this regard, DARTS searches for a smaller network as a proxy and then increases the depth to have a large model in the evaluation stage. A possible attempt to resolve this issue is done by progressively pruning the operations and gradually increasing the depths \cite{chen2019progressive}, which is still an indirect approach and requires strong regularization tricks. Apart from that, recent works~\cite{zela2020understanding,chen2020stabilizing} also point out an instability phenomenon of DARTS. These issues significantly restrict its application.
\textbf{Memory-efficient NAS.}
To reduce the memory cost, several prior works revise the forward procedure of a supernet. PC-DARTS~\cite{Xu2020PC-DARTS} makes use of partial connections instead of the full-fledged supernet. MergeNAS~\cite{wang2020mergenas} proposes to merge all parametric operations into one convolution, a similar super-kernel strategy is also used in \cite{stamoulis2019single}.
ProxylessNAS~\cite{cai2018proxylessnas} samples two paths instead of all paths during the search process, which enables proxyless searching on large datasets. Single-path based supernets~\cite{guo2019single,chu2019fairnas} sample only a single path at each iteration, which are two-stage methods that require additional searching to choose the final models. GDAS \cite{dong2019searching} sample a subgraph of the DAG at each iteration, which is by far the most efficient. However, we observe that a severe instability issue occurs when applying a single-path based differentiable method, which previously has been neglected.
\textbf{Performance collapse of DARTS.} The collapse issue is one of the most critical problems in differentiable architecture search \cite{liang2019darts+,Xu2020PC-DARTS,chu2019fair,chen2020stabilizing,zela2020understanding}. It has been shown that DARTS~\cite{liu2018darts} prefers to choose the parameterless operations, leading to its performance collapse \cite{liang2019darts+,chu2019fair}. Recent works \cite{zela2020understanding,chen2020stabilizing} utilize the eigenvalue of the Hessian matrix as an indicator of collapse and design various techniques to regularize high eigenvalues. Instead, \cite{liang2019darts+,chen2019progressive} directly constricts the number of skip connections to 2 to step aside the collapse. Nonetheless, the previous methods are specifically designed for the full-path training scheme. Whereas the collapse problem in single-path mode has been rarely studied. Our method addresses this issue with a simplistic sampling technique, which preserves the memory-friendly feature. We believe our effort is a promising direction towards efficient and stable NAS.
\section{Methodology}
We first introduce the preliminary of differentiable single-path methods \cite{dong2019searching,dong2019one} that exploit Gumbel reparameterization, and next we disclose a new collapse. We dive into its cause, propose our method, and discuss its strengths.
\subsection{Review on Differentiable Single-Path Methods}
DARTS~\cite{liu2018darts} optimizes a supernet stacked by normal cells and reduction cells. A cell is represented by a DAG with $N$ nodes $\{x_i\}_{i=1}^{N}$ where each node $x_i$ denotes latent representation. Edge $e_{i,j}$ from node $x_i$ to $x_j$ integrates all the candidate operations $\mathcal{O}$, and the output $\bar{o}_{i, j}(x_i)$ is:
\begin{equation} \label{darts}
\bar{o}_{i,j}(x_i) = \sum_{o\in\mathcal{O}} \frac{\exp(\alpha^{o}_{i,j})} {\sum_{o'\in\mathcal{O}} \exp(\alpha^{o'}_{i,j})} o(x_{i}),
\end{equation}
where $\alpha^{o}_{i,j}$ is the architecture weight for operation $o(x_{i})$, as the operation importance. Each intermediate node is connected by all of its predecessors: $x_{j}=\sum_{i<j} \bar{o}_{i,j}(x_{i})$.
Different from the compound operation in Eq. \ref{darts}, single-path methods construct a memory-efficient structure. For edge $e_{i,j}$, a one-hot random vector $z_{i,j} \in \mathbb{R}^{|\mathcal{O}|}$ is sampled. Based on the sampled $z_{i,j}$, only one candidate operation is selected during the forward pass:
\begin{equation} \label{memory-efficient}
\bar{o}_{i,j}(x_i) = \sum_{o\in\mathcal{O}} z_{i,j}^{o}\cdot o(x_{i}),
\end{equation}
where $z_{i,j}^{o}$ equals to 1 if the operation $o$ is sampled otherwise 0.
To distinguish the most important operations, DARTS uses a bi-level optimization by alternately updating the operation parameters $\bm{\theta}$ and architecture weights $\bm{\alpha}$. Since the weights of all operations are involved, DARTS is memory-consuming. This is addressed by differentiable single-path methods where one out of all operations is sampled and activated at each iteration. \cite{dong2019searching,wu2018fbnet,xie2018snas} utilize Gumbel-Softmax reparameterization to make the sampling procedure differentiable, and adopt a bi-level optimization,
\begin{equation}
\begin{split}
\bm{\alpha} &\leftarrow \bm{\alpha} - \nabla_{\bm{\alpha}} L_{val}(\bm{\theta}_z, \bm{\alpha}), \\
\bm{\theta}_{z'} &\leftarrow \bm{\theta}_{z'} - \nabla_{\bm{\theta}_{z'}} L_{train}(\bm{\theta}_{z'}, \bm{\alpha}),
\end{split}
\end{equation}
where $z,z' \sim p(z | \bm{\alpha})$ indicates the activated architectures sampled under a distribution determined by the architecture weights $\bm{\alpha}$, and $\bm{\theta}_z$ indicates the selected operation weights in the architecture $z$.
The main difference between GDAS \cite{dong2019searching} and SNAS \cite{xie2018snas} is that the former provides a memory-efficient version by activating paths in a one-hot way. Note that the normal cell of the best model found in ~\cite{dong2019searching} has 4 skip connections. And for GDAS (FRC), the normal cell contains 5 skip connections. These observations motivate us to dive into whether there is performance collapse like DARTS found in recent works \cite{zela2020understanding,chu2019fair,chen2020stabilizing,zhou2020NAS}. Therefore, we rerun the released code \cite{dong2019searching} to search several times, where the found normal cells are composed of almost all parameterless operations such as skip connection and max pooling, shown in Fig.~\ref{fig:gdas-fail-geno}. It resembles the collapse problem in DARTS, so we also call this phenomenon \textbf{performance collapse}. However, the cause of it in a single-path based method differs from that in DARTS.
On one hand, for the fourth intermediate node, its output is generated by integrating the output of five paths, which brings more difficulties to optimize and thus slows down the convergence of searching loss.
Once skip connection is selected during the sampling process, it is beneficial to faster loss descent, which in turn increases $\alpha_{skip}$ and makes it eventually dominate the architecture parameters.
On the other hand, unlike the sampling scheme that each edge owns one operation in the search phase, the final model is inferred by only preserving two input edges for each node. Such a difference in network topology can mislead the searching objective.
In general, we address this collapse by focusing on two issues: inconsistency between the searching and inferring stage, and instability of the bi-level optimization.
\subsection{Root Analysis for Single-Path NAS Collapse}
\label{subsec:issue}
\textbf{Inconsistency between the searching and inferring stage.}
For one-shot NAS, the structure inconsistency between supernet and final network is mainly reflected in both operation-level and topology-level.
The operation-level inconsistency has been alleviated in recent single-path based methods~\cite{dong2019searching,xie2018snas} by sampling one operation on each edge at each iteration in the search phase.
However, the topology-level inconsistency issue is neglected. Specifically, nodes in the supernet connect with all its predecessors, while nodes in the final network only have two in-degrees.
\begin{figure}[tb!]
\centering
\includegraphics[width=0.45\columnwidth]{img/gdas_fail_normal.pdf}
\includegraphics[width=0.45\columnwidth]{img/gdas_fail_2_normal.pdf}
\caption{Two failure examples of GDAS \cite{dong2019searching} in the DARTS search space produced in our experiment by running the authors' code, where normal cells are full of parameterless operations.}
\label{fig:gdas-fail-geno}
\end{figure}
In this paper, we propose to eliminate the search-evaluation inconsistency at both operation and topology level by disentangling the search for topology and operations. Simultaneously, our proposed approach is even more memory-efficient due to less storage and computation cost.
\textbf{Instability of the stochastic bi-level optimization.}
Instability issues for single-bath based methods can result from its stochastic nature that involves sub-sampling. Specifically, only a subset of operations is sampled at each iteration, resulting in insufficient training of weights $\bm{\theta}$ and $\bm{\alpha}$. Additionally, the final architecture is inferred by the architecture parameters $\bm{\alpha}$, while the sampling scheme involves high variance for the gradient of $\bm{\alpha}$, affecting the convergence of searching. Therefore, it is critical to stabilize such a challenging optimization procedure.
To achieve more accurate estimation and reduce the approximation variance, we propose accumulating and averaging the gradients of architecture parameters.
\subsection{Disentanglement of the Search for Topology} \label{subsec:disentanglement}
To disentangle the search for topology and operations on each edge, we use an indicator $\bm{B}_{i,j}\in\{0,1\}$ to denote whether edge $e_{i,j}$ is selected, and $\bm{A}_{i,j}^o\in\{0,1\}$ to denote whether operation $o$ is selected given edge $e_{i,j}$. The probability of selecting connection $z_{i,j}^o$ can be formulated as:
\begin{equation}
\label{eq4}
p(z_{i, j}^o = 1) = p(\bm{B}_{i, j} = 1) p(\bm{A}_{i, j}^o = 1).
\end{equation}
Therefore, sampling architecture $z$ with $M$ connections can be decomposed into two parts: independently sample $M$ candidate edges and their operations.
\begin{comment}
To disentangle the search for the topology and operations on each edge, we regard the indicator $z_{i, j}^o$ as the joint product of two independent random indicators $\bm{B}_{i, j}$ and $\bm{A}_{i,j}^o$.
\begin{equation}
\bm{z}_{i, j}^o = \bm{B}_{i, j} \bm{A}_{i, j}^o.
\end{equation}
$\bm{B}_{i, j}$ indicates whether the edge $e_{i, j}$ between node $x_i$ and $x_j$ is sampled while $\bm{A}_{i, j}^o$ indicates whether the operator $o$ is sampled among all operators on the edge $e_{i, j}$. $A$ and $B$ are mutually independent, thus
\begin{equation}
\label{eq4}
p(z_{i, j}^o = 1) = p(\bm{B}_{i, j} = 1) p(\bm{A}_{i, j}^o = 1).
\end{equation}
\end{comment}
\textbf{Sampling for operations.}
To sample operation on each edge, we follow Eq. \ref{darts} and use Gumbel-Max~\cite{dong2019searching,xie2018snas}:
\begin{align}
\label{eq:prob_alpha}
\tilde{\bm{\alpha}}_{i,j}^o &= \frac{\exp(\bm{\alpha}^{o}_{i,j})} {\sum_{o'\in\mathcal{O}} \exp(\bm{\alpha}^{o'}_{i,j})}, \\
\bm{A}_{i, j} = \mathrm{one}\underline{\hbox to 0.2cm{}} \mathrm{hot} &\left[\arg\max_{o\in\mathcal{O}}(\log {\tilde{\bm{\alpha}}_{i,j}^{o}} + \bm{g}_{i,j}^{o})\right] \in \mathbb{R}^{|\mathcal{O}|},
\end{align}
where $\bm{g}_{i,j}^o$ are i.i.d. samples drawn from Gumbel (0,1) distribution~\footnote{$\bm{g}_{i,j}^o=-\log(-\log(\bm{u}_{i,j}^o))$, $\bm{u}_{i,j}^o$ is drawn from uniform distribution}. $\bm{A}_{i,j}$ is the one-hot vector indicating which operation is selected on $e_{i,j}$. As $\tilde{\bm{\alpha}}_{i, j}^o$ is the probability of the operator being chosen among all the operators on edge $e_{i, j}$, we have:
\begin{gather}
p(z_{i, j}^o) = \tilde{\bm{\alpha}}_{i, j}^o = p(\bm{A}_{i,j}^o = 1).
\end{gather}
To make search objective differentiable to operation parameter $\bm{\alpha}$, we relax the discrete distribution to a continuous one by Gumbel-Softmax:
\begin{gather}\label{eq:gumbelsoftmax}
\tilde{\bm{A}}_{i,j}^{o} = \frac{\exp\left[(\log\tilde{\bm{\alpha}}_{i,j}^{o}+\bm{g}_{i,j}^{o})/\tau\right]} {\sum_{o'=1}^{|\mathcal{O}|} \exp\left[(\log\tilde{\bm{\alpha}}_{i,j}^{o'}+\bm{g}_{i,j}^{o'})/\tau\right]}, \nonumber \\
\bm{A}_{i, j} = \mathrm{one}\underline{\hbox to 0.2cm{}} \mathrm{hot} \left[\arg\max_{o\in\mathcal{O}}{\tilde{\bm{A}}}_{i,j}^{o}\right],
\end{gather}
in which $\tau$ is the temperature that gradually decreases during the searching stage.
\textbf{Sampling for edges.}
As analyzed in Sec.~\ref{subsec:issue}, topology inconsistency issue occurs in the prior single-path based methods~\cite{dong2019searching,xie2018snas}, where all 14 edges in a cell are selected in the search stage but only 8 edges are retained in the inferring stage. To eliminate such an inconsistency issue, we propose an end-to-end method, referred to as ROME, to directly search the target model, which means sampling 8 edges per cell during the search session as well.
Each intermediate node should connect with exact two predecessors, satisfying DARTS's constraints on the cell topology:
\begin{gather}
\label{eq:constraints_topology}
\sum_{i < j} \bm{B}_{i, j} = 2, \quad \forall j.
\end{gather}
Specifically, we introduce two versions of ROME based on different edge sampling ideas: ROME-v1 and ROME-v2.
\paragraph{1) ROME-v1.}
Suppose node $x_j$ has $j$ possible predecessors to connect to, there exists $\frac{j \times (j - 1)}{2}$ types of edge selection.
We introduce $\bm{I}_j^{(i,k)}, i<k<j$ to indicate if node $x_j$ is connected to $x_i$ and $x_k$. When $\bm{I}_j^{(i,k)}=1$, we have:
\begin{equation}
\bm{B}_{m,j}=\left\{
\begin{aligned}
1, &\quad m=i \ \text{or} \ m=k \\
0, &\quad otherwise
\end{aligned}.
\right.
\end{equation}
We then set a trainable variable $\bm{\beta}_j^{(i,k)}$ to denote the intensity of each edge selection case for node $x_j$.
\begin{gather}
p\left(\bm{I}^{(i, k)}_j = 1\right) = \frac{\exp(\bm{\beta}^{(i, k)}_j)} {\sum_{i'< k'<j} \exp(\bm{\beta}^{(i', k')}_j)} \triangleq \tilde{\bm{\beta}}_j^{(i,j)}.
\end{gather}
Gumbel-Max reparameterization is also applied, where $\bm{g}^{(i,k)}_j$ are i.i.d. samples from Gumbel (0,1) distribution:
\begin{gather} \label{eq:I}
\bm{I}_j = \mathrm{one}\underline{\hbox to 0.2cm{}} \mathrm{hot} \left[\arg\max_{i<k<j}(\log {\tilde{\bm{\beta}}^{(i, k)}_j} + \bm{g}^{(i, k)}_j)\right]
\in \mathbb{R}^{\frac{j \times (j - 1)}{2}}.
\end{gather}
Taking the whole cell into consideration, if edge $e_{i, j}$ is chosen, there must exist another chosen edge $e_{k, j}$, thus leading to
\begin{equation}
\label{eq:v1_sample}
\bm{B}_{i,j} = \sum_{k<i} \bm{I}_j^{(k,i)} + \sum_{k>i} \bm{I}_j^{(i,k)}.
\end{equation}
\begin{comment}
\begin{align}\label{eq:v1_sample}
\bm{B}_{i, j} & = \bm{I}_j^{(1, i)} + \bm{I}_{2, i} + \cdots + \bm{I}_{i - 1, i} \notag
\\ & + \bm{I}_{i, i + 1} + \cdots + \bm{I}_{i, j - 2} + \bm{I}_{i, j - 1}.
\end{align}
\end{comment}
Intending to retain the gradient information, Gumbel-Softmax reparameterization is employed again:
\begin{gather} \label{eq:I_gumbelsoftmax}
\tilde{\bm{I}}_j^{(i,k)} = \frac{\exp\left\{\left[\log {\tilde{\bm{\beta}}^{(i, k)}_j} + \bm{g}^{(i, k)}_j\right]/\tau\right\}}{\sum_{i'<k'<j} \exp\left\{\left[\log {\tilde{\bm{\beta}}^{(i', k')}_j} + \bm{g}^{(i', k')}_j\right]/\tau\right\}},\nonumber \\
\bm{I}_j = \mathrm{one}\underline{\hbox to 0.2cm{}} \mathrm{hot} \left[\arg\max_{i<k<j}\tilde{\bm{I}}^{(i, k)}_j\right].
\end{gather}
\begin{comment}
\begin{gather} \label{eq:I_gumbelsoftmax}
\tilde{\bm{I}}_j^{(i,k)} = \frac{\exp\left(\frac{\log {\tilde{\bm{\beta}}^{(i, k)}_j} + \bm{g}^{(i, k)}_j}{\tau}\right)}{\sum_{i'<k'<j} \exp\left(\frac{\log {\tilde{\bm{\beta}}^{(i', k')}_j} + \bm{g}^{(i', k')}_j}{\tau}\right)}
\end{gather}
\end{comment}
In ROME-v1, we utilize 40 variables to represent the topology of normal and reduction cells because 4 intermediate nodes in a cell have 2, 3, 4, 5 predecessors respectively. By applying Gumbel-Softmax on operation and topology levels, we construct the target model efficiently instead of adding considerable extra parameters to learn.
\paragraph{2) ROME-v2.} Different from ROME-v1 that utilizes $\sum_{i=2}^{N+2}\binom{i}{2}$ variables to represent the topology of a cell with $N$ intermediate nodes and two input nodes, ROME-v2 introduces the probability of selecting edge $e_{i,j}$ as $p(e_{i,j})$, which can be calculated as follows:
\begin{equation}
p(e_{i,j}) = \frac{\exp(\bm{\beta}_{i,j})}{\sum_{k<j}\exp(\bm{\beta}_{k,j})} \triangleq \tilde{\bm{\beta}}_{i,j},
\end{equation}
where $\bm{\beta}_{k,j}$ is the topology parameter. To satisfy the constraints on the cell topology (Eq.~\ref{eq:constraints_topology}), ROME-v2 extends Gumbel-Max reparameterization to a Gumbel-Top2 trick:
\begin{equation}\label{eq:pq_gumbelsoftmax}
\tilde{\bm{B}}_{i,j} = \frac{\exp\left((\log\tilde{\bm{\beta}}_{i,j}+\bm{g}_{i,j})/\tau\right)} {\sum_{i'<j} \exp\left((\log\tilde{\bm{\beta}}_{i',j}+\bm{g}_{i',j})/\tau\right)},
\end{equation}
\begin{equation}\label{eq:pq}
\bm{B}_{i,j}=\left\{
\begin{aligned}
1, &\quad i \in \arg\mathop{\mathrm{top2}}\limits_{i'<j} (\tilde{\bm{B}}_{i',j}) \\
0, &\quad otherwise
\end{aligned}.
\right.
\end{equation}
ROME-v2 provides an efficient way to sample edges, and only 28 variables are used to represent the topology of normal and reduction cell with 4 intermediate nodes.
\paragraph{Theoretical analysis on Gumbel-Top2.} We show that our Gumbel-Top2 technique is equivalent to sampling two different edges without replacement with probability simplex $p_i$. We sketch our proof as follows and the rigorous full version can be found in supplemental materials.
Let $p_i$ be the probability of $e_i$ being selected by single choice among $n$ edges. Without loss of generality, we suppose that $e_1$ is chosen.
i) Consider sampling two edges without replacement:
The probability of $e_i$ being sampled is $p_1 + \sum_{i=2}^n p_i \frac{p_1}{1 - p_i}$.
ii) Consider the Gumbel-Top2 scheme:
$\epsilon_i$ is sampled uniformly from $[0, 1]$ and the largest two $\log p_i - \log (- \log \epsilon_i)$s are chosen. The cases that $p_1$ is chosen can be divided into two disjoint parts.
\textbf{A)} $\log p_1 - \log (-\log \epsilon_1)$ is the largest, and the probability is $p_1$ due the our knowledge of Gumbel-Max scheme.
\textbf{B)} $\log p_1 - \log (-\log \epsilon_1)$ is the second largest only next to $\log p_i - \log (-\log \epsilon_i)$. We directly get its probability:
\begin{gather*}
\int_{0}^1 \epsilon_1^{p_2 / p_1} \epsilon_1^{p_3 / p_1} \cdots \ (1 - \epsilon_1^{p_i / p_1}) \cdots \epsilon_1^{p_n / p_1} d\epsilon_1\\
= \int_{0}^1 (1 - \epsilon_1^{p_i / p_1})\epsilon_1^{\frac{1 - p_i}{p_1} - 1} d\epsilon_1 \\
= \frac{p_1}{1 - p_i} - p_1
= p_i \frac{p_1}{1 - p_i}.
\end{gather*} Therefore, the probability is $p_1 + \sum_{i=2}^n p_i \frac{p_1}{1 - p_i}$, which means the two schemes are equivalent.
\begin{comment}
ROME-v2 denotes the probability of edge $e_{i,j}$ being sampled as $\beta_{i,j}$ and selects the predecessor edges with the top two probabilities for each node. Specifically, we formulate the procedure of sampling two edges as follows:
\begin{equation}\label{eq:pq}
\begin{split}
\sum_{i < j} \bm{B}_{i, j} &= \bm{I}_{p} + \bm{I}_{q} \\
p,q &= \arg \mathrm{top2}_{i<j}~(\log {\beta_{i, j}} + g_{i, j}).
\end{split}
\end{equation}
where $\bm{I}_{p}$ is a one-hot vector with 1 at the $p^{th}$ position. We can theoretically (or empirically) substantiate that our Gumbel-Top2 trick is equivalent to sampling twice without replacement according to the probability distribution $\bm{\beta}$ in \ref{proof}.
\end{comment}
\subsection{Gradient Summation in Bi-level Optimization}
The single-path based GDAS~\cite{dong2019searching} borrows DARTS's bi-level optimization. It samples one architecture and applies one-step gradient descent to update operation parameters and architecture weights alternately. Fig.~\ref{fig:gdas-fail-geno} shows this procedure can incur instability.
In our analysis, the sampling variance can mislead the architecture parameters. Moreover, insufficient training for weights of parametric operations should be considered since only the selected operations contribute to the network output, and other operations have no gradient. Consequently, we propose a gradient accumulation strategy to robustify the bi-level optimization.
\begin{algorithm}[tb!]
\caption{ROME (with two versions v1 and v2).}
\label{alg:romeNAS}
\textbf{Input}: iteration count $T$; number of sampling $K$; initialized network weights $\bm{\theta}$; and architecture parameters $\bm{\alpha}$, $\bm{\beta}$;\\
\textbf{Output}: optimal architecture $z^*$;
\begin{algorithmic}[1]
\FOR {$t = 1 \to T$}
\STATE Sample two batches of data samples $D_s$ and $D_t$; \\
\FOR {$k = 1 \to K$}
\STATE Operation sampling by Eq.~\ref{eq:gumbelsoftmax};
\STATE Topology sampling for search-evaluation consistency by Eq.~\ref{eq:v1_sample} (for v1) or Eq.~\ref{eq:pq} (for v2);\\
\STATE Get sampled architecture $z_k$;\\
\ENDFOR
\STATE Gradient accumulation and update $\bm{\alpha}, \bm{\beta}$ by Eq.~\ref{eq:update_alpha}; \\
\FOR {$k = 1 \to K$}
\STATE Operation sampling by Eq.~\ref{eq:gumbelsoftmax};
\STATE Topology sampling for search-evaluation consistency by Eq.~\ref{eq:v1_sample} (for V1) or Eq.~\ref{eq:pq} (for v2);\\
\STATE Get sampled architecture $z'_k$;\\
\ENDFOR
\STATE Gradient accumulation and update $\bm{\theta}$ by Eq.~\ref{eq:update_theta};
\ENDFOR
\STATE \textbf{return}: $z^*=\arg\max\limits_z p(z;\bm{\alpha}, \bm{\beta})$
\end{algorithmic}
\end{algorithm}
As shown in Sec.~\ref{subsec:disentanglement}, we utilize architecture weights $\bm{\alpha}$ and $\bm{\beta}$ to construct a parametric distribution. We select a candidate architecture by independently sampling edges and operations. Suppose there are $M$ edges $\{e_1,...,e_M\}$ and the corresponding operations $\{o_1,...,o_M\}$ make up an architecture $z=\{(e_1,o_1),...,(e_M,o_M)\}$, and thus the probability of $z$ being selected is given by:
\begin{equation}
p(z;\bm{\alpha}, \bm{\beta}) = \prod_{i=1}^M p(e_i;\bm{\beta})\times p(o_i|e_i;\bm{\alpha}).
\end{equation}
NAS can be then modeled as finding optimal $\bm{\alpha}$ and $\bm{\beta}$ to minimize the expectation of validation loss of the architectures as Eq.~\ref{eq:opt_operation}, where $\bm{\theta}_z^*$ denotes the optimal operation parameters for the sampled architecture $z$.
\begin{equation}
\begin{split}
\min_{\bm{\alpha}, \bm{\beta}} \quad & \mathbb{E}_{z\sim p(z;\bm{\alpha}, \bm{\beta})}\left[L_{val}(\bm{\theta}_z^*, z) \right]
\\
\mathrm{s.t.} \quad & \bm{\theta}_z^* = \arg\min_{\bm{\theta}} L_{train}(\bm{\theta}, z) \label{eq:opt_operation}
\end{split}
\end{equation}
Rather than GDAS~\cite{dong2019searching} that approximates the expectation by sampling once, we propose to sample for $K$ times and utilize the average loss to approximate the expectation. Suppose that the gradient of $\alpha$ is a random variable with a variance of $\sigma^2$, sampling $K$ times and averaging generate an unbiased estimator with a variance of $\frac{\sigma^2}{K}$. To make the average loss differentiable to architecture weights, we refer to DARTS~\cite{liu2018darts} and use current $\bm{\theta}_z$ to approximate $\bm{\theta}_z^*$.
Accordingly, we propose to train operation parameters by accumulating the gradients of operation weights under multiple candidate architectures. Not only gradient accumulation is able to alleviate the issue of insufficient training for operation parameters, but also it makes the approximation of $\bm{\theta}_z^*$ much more stable. Consequently, we can alternately update operation parameters $\bm{\theta}$ and architecture
parameters $\bm{\alpha}$ (similar for $\bm{\beta}$) as:
\begin{align}
\bm{\alpha} &\leftarrow \bm{\alpha} - \frac{1}{K}\sum_{k=1}^K \nabla_{\bm{\alpha}} L_{val}(\bm{\theta}, z_k) \label{eq:update_alpha},\\
\bm{\theta} &\leftarrow \bm{\theta} - \sum_{k=1}^K \nabla_{\bm{\theta}} L_{train}(\bm{\theta}, z'_k) \label{eq:update_theta},
\end{align}
where $z_k,z'_k, k=1,...,K$ denote the sampled architectures. The overall method is shown in Alg.~\ref{alg:romeNAS}
\section{Experiments}
\label{sec:experiments}
\subsection{Protocols}
\textbf{S0 (DARTS's search space).} DARTS's search space is comprised of a stack of duplicate normal cells and reduction cells. Each cell is represented by a DAG of 4 intermediate nodes. There are candidate operations \{maxpool, avgpool, skip\_connect, sep\_conv 3$\times$3 and 5$\times$5, dil\_conv 3$\times$3 and 5$\times$5\}, between each two nodes.
We search and evaluate on both CIFAR-10~\cite{krizhevsky2009learning} and ImageNet ~\cite{deng2009imagenet}.
\textbf{S1-S4 (RobustDARTS's reduced search space).} We also conduct experiments on the reduced search spaces introduced by RobustDARTS~\cite{zela2020understanding}.
S1 is a pre-optimized search space and the two candidate operations on each edge is different, see \cite{zela2020understanding} for details. In other three search spaces, the candidate operations on each edge are the same. Specifically, S2 has \{3 $\times$ 3 SepConv, SkipConnect\}, S3 has \{3 $\times$ 3 SepConv, SkipConnect, Zero (None)\}, and S4 has \{3 $\times$ 3 SepConv, Noise\}.
To demonstrate the robustness of our method, we search and evaluate under the above four search spaces on CIFAR-10, CIFAR-100~\cite{krizhevsky2009learning}, and SVHN~\cite{netzer2011reading}.
\textbf{Search settings.}
Similar to DARTS, the supernet consists of 8 cells with 16 initial channels
We alternately update operation parameters and architecture weights for $50$ epochs and set the sampling number $K=7$ by default. For operation parameters, we use the SGD optimizer with 0.9 momentum and set the base learning rate $0.05$; For architecture weights, we adopt Adam optimizer and set the base learning rate $3\times 10^{-4}$.
\textbf{Evaluation settings.}
We use standard evaluation settings as DARTS \cite{liu2018darts}. Specially, we train the inferred model for 600 epochs using SGD with a batch size of 96.
For S0 search space, networks are constructed by stacking 20 cells with 36 initial channels, and are trained with the same strategy and data processing tricks as the previous works~\cite{chen2019progressive,liu2018darts}.
For S1-S4 search spaces, we strictly follow the settings in RobustDARTS \cite{zela2020understanding} for fair comparison. We also increase the number of cells and initial channels to 20 and 36 as in SDARTS \cite{chen2020stabilizing} for further evaluation.
\subsection{Ablation Study}
Our method disentangles the search for topology and utilizes gradient accumulation strategy to resolve the search-evaluation inconsistency issue and alleviate the instability issue for single-path based methods. To show the efficacy of each component, we conduct an ablation study under S0 search space on CIFAR-10.
\begin{table}[tb!]
\begin{center}
\setlength{\tabcolsep}{2pt}
\small
\begin{tabular}{|c|c|c|c|}
\hline
Disentanglement & GA for $\bm{\theta}$ & Avg for $\bm{\alpha}$ & Acc ($\%$) \\
\hline\hline
$\times$ & $\times$ & $\times$ & 96.52$\pm$0.07 \\
\checkmark & $\times$ & $\times$ & 97.12$\pm$0.06 \\
\checkmark & $\times$ & \checkmark & 97.22$\pm$0.07 \\
\checkmark & \checkmark & $\times$ & 97.34$\pm$0.07 \\
\checkmark & \checkmark & \checkmark & \textbf{97.42$\pm$0.07} \\
\hline
\end{tabular}
\end{center}
\vspace{-10pt}
\caption{Component analysis of our method as searched and evaluated on CIFAR-10. We conduct three parallel tests on each setting and report the mean and standard deviation of top-1 accuracy. }
\label{tab:componet_analysis}
\end{table}
\textbf{Component analysis for the instability issue.}
To evaluate how each component contributes to the instability issue, we utilize the disentanglement and gradient accumulation strategy separately. Results are shown in Table~\ref{tab:componet_analysis}, where `GA for $\bm{\theta}$' indicates applying gradient accumulation when training the network weights, `Avg for $\bm{\alpha}$' indicates sampling candidate architectures for $K=7$ times and utilize the average loss to approximate the expectation.
We observe that disentanglement can significantly improve the searching performance, indicating that the inconsistency issue is the principal reason for performance collapse. On one hand, inconsistent topology between search and evaluation results in inconsistent search objective; On the other hand, training weights of 14 operations (without disentanglement) is much more difficult than training 8 operations (with disentanglement), making the convergence slower. Moreover, we observe that gradient accumulation strategy on $\bm{\theta}$ and $\bm{\alpha}$ can both improve the search performance.
\begin{table}[tb!]
\begin{center}
\small
\begin{tabular}{|c|c|c|}
\hline
\# Samples ($K$) & Acc (\%) & \# Params (M) \\
\hline\hline
1 & 97.12$\pm$0.06 & 3.34 \\
4 & 97.28$\pm$0.07 & 3.57 \\
7 & 97.42$\pm$0.07 & 3.73 \\
10 & 97.46$\pm$0.12 & 4.06 \\
\hline
\end{tabular}
\end{center}
\vspace{-10pt}
\caption{Sensitivity study of sampling number $K$. We do three parallel tests on each setting and report mean and standard deviation of top-1 accuracy. Models are searched and evaluated on CIFAR-10. The disentanglement technique is used in search phase.}
\label{tab:sensitivity_K}
\end{table}
\textbf{Sensitivity to the sampling number $K$.}
Table~\ref{tab:sensitivity_K} compares accuracy by setting $K$ as $1, 4, 7, 10$. The average performance increases monotonically with $K$ but saturates when $K$=7. Therefore, we use $K$=7 by default.
\textbf{Memory analysis.} Table~\ref{tab:mem_com} compares GPU memory cost under S0 search space on CIFAR-10. ROME has the lowest memory cost and this advantage can be signified given more candidate operations in search space.
\begin{table}[tb!]
\begin{center}
\small
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
\multirow{2}{*}{\textbf{Method}} & \multirow{2}{*}{DARTS} & \multirow{2}{*}{GDAS} & \multicolumn{2}{c|}{PC-DARTS} & \multirow{2}{*}{ROME}\\
\cline{4-5}
&&&K=4&K=2&\\
\hline\hline
\textbf{Memory} (G) & 9.4 & 3.1 & 3.7 &5.7& 2.3 \\
\hline
\end{tabular}
\end{center}
\vspace{-10pt}
\caption{Memory cost comparison. This evaluation is measured based on a batch size of 64, 16 initial channels, and 8 layers.}
\label{tab:mem_com}
\end{table}
\begin{table}[tb!]
\small
\centering
\setlength{\tabcolsep}{2pt}
\begin{tabular}{|l*{4}{|c}|c|}
\hline
\textbf{Models} & \textbf{Params} & \textbf{FLOPs} & \textbf{Error} & \textbf{Cost}& SP\\
& \scriptsize{(M)} & \scriptsize{(M)} & \scriptsize{(\%)} & \tiny{GPU Days} & \\
\hline\hline
NASNet-A \cite{zoph2017learning} & 3.3 & 608$^\dagger$ & 2.65 & 2000&\checkmark\\
ENAS \cite{pham2018efficient} & 4.6 & 626$^\dagger$ & 2.89 & 0.5 &\checkmark\\
DARTS-V1 \cite{liu2018darts} & 3.3 & 528$^\dagger$ & 3.00 &0.4&$\times$\\
SNAS \cite{xie2018snas} & 2.8 & 422$^\dagger$ & 2.85 & 1.5&$\times$\\
P-DARTS \cite{chen2019progressive} & 3.4 & 532$^\dagger$ & 2.50 & 0.2&$\times$\\
PC-DARTS \cite{Xu2020PC-DARTS} & 3.6 & 558$^\dagger$ & 2.57 & 0.1&$\times$\\
GDAS \cite{dong2019searching} & 3.4 & 519$^\dagger$ & 2.93 & 0.2&\checkmark\\
ROME-v1-a &4.5 & 683 & 2.53& 0.3&\checkmark\\
\textbf{ROME-v2-a} & 3.6& 591 &2.48&0.3&\checkmark\\
\hline
DARTS-V1 \cite{Yu2020Evaluating} &-&-&3.38$\pm$0.23&0.4&$\times$\\
P-DARTS \cite{chen2019progressive}$^\ddagger$ & 3.3$\pm$0.2 & 540$\pm$34 & 2.81$\pm$0.14 & 0.3 &$\times$\\
PC-DARTS \cite{Xu2020PC-DARTS} $^\ddagger$ & 3.6$\pm$0.5 & 592$\pm$90 & 2.89$\pm$0.22 & 0.1 &$\times$\\
R-DARTS \cite{zela2020understanding} & - & - & 2.95$\pm$0.21 & 1.6&$\times$\\
SDARTS-ADV \cite{chen2020stabilizing} & 3.3 & - & 2.61$\pm$0.02 & 1.3&$\times$\\
ROME-v1 & 4.0$\pm$0.6 & 670$\pm$21 & 2.63$\pm$0.09 & 0.3&\checkmark\\
\textbf{ROME-v2 } & 3.7$\pm$0.2 & 595$\pm$28&2.58$\pm$0.07&0.3&\checkmark\\
\hline
\end{tabular}
\caption{Performance comparisons on CIFAR-10. The first block reports the accuracy of the best model. The second block reports the averaged performance using 4 independent searches as recommended by \cite{zela2020understanding,chen2020stabilizing,Yu2020Evaluating}. $^\ddagger$: reproduced result using their released code since they didn't report the average performance. $^\dagger$: FLOPs is calculated by their released architecture. SP: single-path based method, which is memory-efficient.
}
\label{tab:cifar10-s0}
\end{table}
\begin{comment}
\begin{figure}[tb!]
\centering
\includegraphics[width=0.38\columnwidth]{img/ROME_cifar10_best_normal.pdf}
\includegraphics[width=0.58\columnwidth]{img/ROME_cifar10_best_reduction.pdf}
\caption{Best normal and reduction by ROME-v2 on CIFAR-10.}
\label{fig:rome-v2-cifar10-best-geno}
\end{figure}
\end{comment}
\subsection{Search Results Analysis}
\textbf{Result on CIFAR-10.}
We conduct 4 parallel tests by searching with different random seeds and then evaluating on CIFAR-10. The average and standard deviation of the performance of our method are reported in Table~\ref{tab:cifar10-s0}, showing that ROME achieves state-of-the-art performance with only 0.3 GPU-days search cost\footnote{GDAS cost 0.2 GPU-days to search 250 epochs. We search 50 epochs with $K=7$, being equivalent to search by GDAS for 350 epochs.}. Specifically, our best model achieves $97.52\%$ accuracy with $3.6$M parameters
Moreover, it's recommended to report average results across independent runs of searching~\cite{zela2020understanding,chen2020stabilizing,Yu2020Evaluating,chu2019fair}. ROME-v2 achieves an average of 2.58$\pm$0.07\% error rate, which slightly outperforms up-to-date SOTAs such as SDARTS-ADV~\cite{chen2020stabilizing}. However, ours takes 3$\times$ fewer time than it.
\begin{comment}
\begin{figure}[tb!]
\centering
\includegraphics[width=0.38\columnwidth]{img/ROME_V1_CIFAR10_s0_best-normal.pdf}
\includegraphics[width=0.38\columnwidth]{img/ROME_V1_CIFAR10_s0_best-reduction.pdf}
\caption{Best normal and reduction by ROME-v1 on CIFAR-10.}
\label{fig:rome-v1-cifar10-best-geno}
\end{figure}
\end{comment}
\textbf{Result on CIFAR-100.} We search on CIFAR-100 as \cite{liu2018darts} and show the results in Table~\ref{tab:comparison-cifar100} (supplementary).
Our method also outperforms other SOTAs with a clear margin.
\begin{table}[tb!]
\small
\centering
\setlength{\tabcolsep}{2pt}
\begin{tabular}{|l|c|c|c|c|c|c|}
\hline
\textbf{Models} & \textbf{FLOPs} & \textbf{Params} & \textbf{Top-1} & \textbf{Cost} & Way\\
& \scriptsize{(M)} & \scriptsize{(M)} & \scriptsize{(\%)} & \scriptsize{(GPU days)} & \\
\hline\hline
AmoebaNet-A \cite{real2019regularized} & 555 & 5.1 & 74.5& 3150&TF \\
NASNet-A \cite{zoph2017learning} & 564 & 5.3 &74.0 & 2000&TF\\
PNAS \cite{liu2018progressive} & 588 & 5.1 & 74.2 &225&TF\\
DARTS \cite{liu2018darts} & 574 & 4.7 & 73.3 & 0.4&TF\\
MdeNAS \cite{zheng2019multinomial} & - & 6.1 & 74.5& 0.16&TF\\
PC-DARTS \cite{Xu2020PC-DARTS} & 586 & 5.3 & 74.9 & 0.1&TF\\
FairDARTS-B \cite{chu2019fair}& 541 & 4.8 &75.1 & 0.4&TF\\
SNAS \cite{xie2018snas} &522&4.3&72.7 & 1.5&TF\\
GDAS \cite{dong2019searching} &581&5.3&74.0 & 0.2&TF \\
ROME-b (ours) &576&5.2&75.3&0.3&TF\\
\hline
PC-DARTS \cite{Xu2020PC-DARTS} & 597 & 5.3 & 75.8 & 3.8&DS\\
FairNAS-C \cite{chu2019fairnas} &321 & 4.4 & 74.7 &12&DS \\
SPOS \cite{guo2019single} &323&3.5&74.4& 12&DS\\
ProxylessNAS GPU \cite{cai2018proxylessnas} & 465 & 7.1 & 75.1 & 8.3&DS \\
FBNet-C \cite{wu2018fbnet} & 375 & 5.5 & 74.9 & 9&DS \\
GDAS$^*$ \cite{dong2019searching}&405&3.6&72.5&0.8&DS\\
ROME-c (ours) & 556 & 5.1 & 75.5 & 0.5&DS\\
\hline
\end{tabular}
\caption{Results on ImageNet. TF: transferred to ImageNet; DS: direct search on ImageNet.
Our method can support direct search on large scale dataset for its memory efficiency. FBNet, FairNAS, SPOS and ProxylessNAS use a different search space as \cite{tan2018mnasnet}, which can achieve higher accuracy with fewer FLOPs. $^*$: obtained by our experiments based on the code publicly released by the authors.}
\label{tab:comparison-imagenet}
\end{table}
\begin{figure}
\subfigure[GDAS]{
\includegraphics[width=0.58\columnwidth]{img/GDAS_IMAGENET_S0_best-normal}
}
\subfigure[ROME]{
\includegraphics[width=0.36\columnwidth]{img/ROME_IMAGENET_S0_best-normal}
}
\caption{The architecture of normal cells searched by GDAS and ROME on ImageNet under S0 search space. Network searched by GDAS is dominated by skip connection and only obains $72.5\%$ accuracy on ImageNet, while our method is much more stable and achieves $75.5\%$ accuracy.}
\label{fig:compare-s0-imagenet}
\end{figure}
\begin{table*}[tb!]
\center
\small
\setlength{\tabcolsep}{2pt}
\begin{tabular}{|c*{7}{c}H*{2}{c}|*{2}{c}H*{2}{c}|}
\hline
\multicolumn{2}{|c}{\multirow{2}*{\textbf{Benchmark}}} & \multirow{2}*{\textbf{DARTS}$^\dagger$} & \multicolumn{2}{c}{\textbf{R-DARTS}$^\dagger$} & \multicolumn{2}{c}{\textbf{DARTS}$^\dagger$} & \multirow{2}*{\textbf{SDARTS-RS}$^\dagger$} & & \multicolumn{2}{c|}{\textbf{ROME}$^{\dagger}$} & \multirow{2}*{\textbf{PC-DARTS}$^\ddagger$} & \multirow{2}*{\textbf{SDARTS-RS}$^\ddagger$} & & \multicolumn{2}{c|}{\textbf{ROME}$^\ddagger$} \\
& & & DP & L2 & ES & ADA & & ADV & v1 & v2 & & & ADV & v1 & v2 \\
\hline\hline
\multirow{4}*{C10}& S1 & 3.84 & 3.11 & 2.78 & 3.01 & 3.10 & 2.78 & 2.73 & \textbf{2.68} & \textbf{\underline{2.62}} & 3.11 & 2.78 & 2.73 & \textbf{2.68} & \textbf{\underline{2.62}} \\
~ & S2 & 4.85 & 3.48 & 3.31 & 3.26 & 3.35 & 3.33$^\star$ & 3.41$^\star$ & \textbf{3.24} & \textbf{\underline{2.95}} & 3.02 & \textbf{2.75} & \textbf{2.65} & 2.79 & \textbf{\underline{2.62}} \\
~ & S3 & 3.34 & 2.93 & \textbf{\underline{2.51}} & 2.74 & 2.59 & \textbf{2.53} & \textbf{\underline{2.49}} & 2.65 & 2.61 & \textbf{\underline{2.51}} & \textbf{2.53} & \textbf{\underline{2.49}} & 2.65 & 2.61 \\
~ & S4 & 7.20 & 3.58 & 3.56 & 3.71 & 4.84 & 4.84$^\star$ & 4.28$^\star$ & \textbf{\underline{3.21}} & \textbf{3.31} & 3.02 & \textbf{2.93} & 2.87 & 3.61 & \textbf{2.68} \\
\hline
\multirow{4}*{C100} & S1 & 29.46 & 25.93 & 24.25 & 28.37 & 24.03 & 23.51 & \textbf{22.33} & 22.34 & \textbf{\underline{22.04}} & 18.87 & \textbf{\underline{17.02}} & \textbf{\underline{16.88}} & 17.27 & \textbf{17.24} \\
~ & S2 & 26.05 & 22.30 & 22.24 & 23.25 & 23.52 & 22.28 & \textbf{\underline{20.56}} & \textbf{\underline{21.95}} & \textbf{22.12} & 18.23 & 17.56 & 17.24 & \textbf{17.09} & \textbf{\underline{17.06}} \\
~ & S3 & 28.90 & 22.36 & 23.99 & 23.73 & 23.37 & \textbf{\underline{21.09}} & \textbf{\underline{21.08}} & 22.56 & \textbf{22.11} & 18.05 & 17.73 & 17.12 & \textbf{16.95} & \textbf{\underline{16.94}} \\
~ & S4 & 22.85 & 22.18 & 21.94 & \textbf{21.26} & 23.20 & 21.46 & 21.25 & 21.33 & \textbf{\underline{20.44}} & 17.16 & 17.17 & \textbf{\underline{15.46}} & \textbf{\underline{15.99}} & \textbf{16.18} \\
\hline
\multirow{4}*{SVHN} & S1 & 4.58 & 2.55 & 4.79 & 2.72 & 2.53 & 2.35 & \textbf{2.29} & \textbf{2.33} & \textbf{\underline{2.27}} & 2.28 & 2.26 & 2.16 & \textbf{\underline{2.07}} & \textbf{2.14} \\
&S2 & 3.53 & 2.52 & 2.51 & 2.60 & 2.54 & 2.39 & \textbf{2.35} & \textbf{2.39} & \textbf{\underline{2.30}} & 2.39 & 2.37 & 2.07 & \textbf{2.14} & \textbf{\underline{2.07}} \\
&S3 & 3.51 & 2.49 & \textbf{2.48} & 2.50 & 2.50 & \textbf{\underline{2.36}} & 2.40 & 2.58 & 2.51 & 2.27 & 2.21 & 2.05 & \textbf{2.14} & \textbf{\underline{2.07}}\\
&S4 & 3.05 & 2.61 & 2.50 & 2.51 & 2.46 & 2.46 & \textbf{2.42} & \textbf{2.43} & \textbf{\underline{2.34}} & 2.37 & 2.35 & 1.98 & \textbf{2.00} & \textbf{\underline{1.99}} \\
\hline
\end{tabular}
\caption{Comparison in RobustDARTS \cite{zela2020understanding} reduced search spaces and 3 datasets. We report the \textbf{lowest error rate} of 3 found architectures. $^\dagger$: using the settings of \cite{zela2020understanding} where CIFAR-100 and SVHN models have 8 layers and 16 initial channels, CIFAR-10 models have 20 layers and 36 initial channels except that S2 and S4 have 16 initial channels. $^\star$: retrained with 16 initial channels. $^\ddagger$: using training settings where all models have 20 layers and 36 initial channels in \cite{chen2020stabilizing}. The best is underlined and in bold face, the second best is in bold.}
\label{tab:comparison-rdarts-s1-4-best}
\end{table*}
\begin{table}[tb!]
\smallskip
\centering
\setlength{\tabcolsep}{1pt}
\footnotesize
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
\multicolumn{2}{|c|}{\scriptsize{\textbf{Benchmark}}} & \scriptsize{\textbf{DARTS}} & \scriptsize{\textbf{DARTS-ES}} & \scriptsize{\textbf{DARTS-ADA}} & \scriptsize{\textbf{ROME-v1}} & \scriptsize{\textbf{ROME-v2}} \\
\hline\hline
\multirow{4}*{C10} & S1 & 4.66$\pm$0.71 & 3.05$\pm$0.07 & 3.03$\pm$0.08 & \textbf{2.93$\pm$0.09} & \textbf{\underline{2.66$\pm$0.06}} \\
~ & S2 & 4.42$\pm$0.40 & 3.41$\pm$0.14 & 3.59$\pm$0.31 & \textbf{3.34$\pm$0.12} & \textbf{\underline{3.14$\pm$0.14}} \\
~ & S3 & 4.12$\pm$0.85 & 3.71$\pm$1.14 & 2.99$\pm$0.34 & \textbf{2.72$\pm$0.09} & \textbf{\underline{2.61$\pm$0.00}} \\
~ & S4 & 6.95$\pm$0.18 & 4.17$\pm$0.21 & 3.89$\pm$0.67 & \textbf{\underline{3.21$\pm$0.00}} & \textbf{3.44$\pm$0.12} \\
\hline
\multirow{4}*{C100} & S1 & 29.93$\pm$0.41 & 28.90$\pm$0.81 & 24.94$\pm$0.81 & \textbf{\underline{22.65$\pm$0.45}} & \textbf{22.71$\pm$0.71} \\
~ & S2 & 28.75$\pm$0.92 & 24.68$\pm$1.43 & 26.88$\pm$1.11 & \textbf{23.14$\pm$0.98} & \textbf{\underline{22.91$\pm$0.75}} \\
~ & S3 & 29.01$\pm$0.24 & 26.99$\pm$1.79 & 24.55$\pm$0.63 & \textbf{23.03$\pm$0.66} & \textbf{\underline{22.43$\pm$0.36}} \\
~ & S4 & 24.77$\pm$1.51 & 23.90$\pm$2.01 & 23.66$\pm$0.90 & \textbf{21.33$\pm$0.00} & \textbf{\underline{20.95$\pm$0.45}} \\
\hline
\multirow{4}*{SVHN}& S1 & 9.88$\pm$5.50 & 2.80$\pm$0.09 & 2.59$\pm$0.07 & \textbf{2.37$\pm$0.04} & \textbf{\underline{2.34$\pm$0.06}} \\
~ & S2 & 3.69$\pm$0.12 & 2.68$\pm$0.18 & 2.79$\pm$0.22 & \textbf{2.49$\pm$0.14} & \textbf{\underline{2.41$\pm$0.07}} \\
~ & S3 & 4.00$\pm$1.01 & 2.78$\pm$0.29 & \textbf{2.58$\pm$0.07} & 2.61$\pm$0.03 & \textbf{\underline{2.56$\pm$0.03}} \\
~ & S4 & 2.90$\pm$0.02 & 2.55$\pm$0.15 & 2.52$\pm$0.06 & \textbf{2.43$\pm$0.00} & \textbf{\underline{2.34$\pm$0.00}} \\
\hline
\end{tabular}
\vspace{-5pt}
\caption{Comparison in 4 search spaces and 3 datasets. We report the mean and variance about \textbf{lowest error rate} of 3 found architectures. We use the model settings as in \cite{zela2020understanding}. In case the variance is zero, it means that we found three identical networks.}
\label{tab:comparison-rdarts-s1-4-avg}
\end{table}
\textbf{Result on ImageNet.} As a common practice \cite{liu2018darts,chen2019progressive,li2019sgas,chu2019fair}, we transfer the architecture searched on CIFAR-10 to ImageNet. Following the same setting as \cite{li2019sgas,chu2019fair}, we train models for 250 epochs with a batch size of 1024. Specifically, we utilize the SGD optimizer with 0.9 momentum. The learning rate is initialized as 0.5 and decayed by the cosine strategy. We also utilize cutout and 0.4 weighted auxiliary head loss. The result is shown in Table~\ref{tab:comparison-imagenet}, our model ROME-b achieves 75.3$\% $ top-1 accuracy.
Moreover, ROME allows direct and proxyless search on ImageNet due to its low memory cost.
We randomly sampling 10\% and 10\% images respectively from each class to train operation parameters and architecture weights. We construct a supernet by stacking 8 cells with 16 initial channels and search for 30 epochs with $K=3$. The search time can be reduced to 0.4 GPU days on a single Tesla v100. The searched architecture (ROME-c in Tabel~\ref{tab:comparison-imagenet}) is evaluated with the same training setting as above. As shown in Tabel~\ref{tab:comparison-imagenet}, ROME-c achieves 75.5$\%$ top-1 accuracy on ImageNet validation set
To make fair comparisons, we also search for 90 epochs by GDAS~\cite{dong2019searching} under the same settings. However, the network is dominated by skip connection and only achieves $72.5\%$ top-1 accuracy.
The structure of normal cells searched by GDAS and ROME are shown in Fig.~\ref{fig:compare-s0-imagenet}. The structure of reduction cells are shown in the supplementary ( Fig.~\ref{fig:gdas-s0-imagenet} and Fig.~\ref{fig:rome-s0-imagenet}).
\subsection{Robustness Evaluation on 12 Benchmarks}
We further evaluate the performance and generalization of our method on recent 4 harder search spaces across 3 datasets from \cite{zela2020understanding}. For each benchmark, we perform four independent searches and train the searched 4 models to report their performance by mean and standard variance. This process is recommended by \cite{Yang2020NAS,chen2020stabilizing,zela2020understanding} to fairly compare different NAS methods. The result is shown in Table~\ref{tab:comparison-rdarts-s1-4-avg}. Our methods robustly outperform RobustDARTS with a clear margin across the 12 benchmarks with 3$\times$ fewer search cost. The best cells found by ROME v1 and v2 are shown
in Sec. \ref{supp:fig-geno} in the supplementary material.
\subsection{Discussions}
\textbf{Memory cost comparison with PC-DARTS.} PC-DARTS utilizes partial channels during the search stage to reduce GPU memory cost, in which the partial ratio is controlled by a hyperparameter $K$. $K$ requires careful calibration for different tasks. Moreover, it still needs a partial whole supernet to perform the search. Given a mixed micro operator with N candidate operations, the memory cost is $\mathcal{O}(KN)$. In contrast, ours doesn't require calibrating such an extra hyperparameter and enjoys a cost of $\mathcal{O}(1)$.
\textbf{Discussion on collapse behavior across popular NAS benchmarks.} We argue that excluding an important operation for search space can cause illusive conclusions. Specifically, NAS-Bench-1Shot1 \cite{zela2020nas} suggests that Gumbel-based NAS is quite robust. However, this observation is laying on the basis that popular skip connections are not included in the search space \cite{pmlr-v97-ying19a}. After adding skip connection into the choices, we perform the GDAS search using their released code \footnote{https://github.com/automl/nasbench-1shot1}. The best model found is full of skip connections, which again supports our discovery of collapse issue in single-path based NAS, see Fig.~\ref{fig:1shot1-gdas-fail} and more
in Fig.~\ref{fig:1shot1-gdas-fail-rest} in the supplemental material.
Instead, while performing ROME-v2 in such modified search spaces, we don't suffer the same issue, see
Fig.~\ref{fig:1shot1-rome-best} in the supplemental material.
\begin{figure}[tb!]
\centering
\subfigure[GDAS \cite{dong2019searching}]{
\includegraphics[width=0.45\columnwidth]{img/GDAS_SKIP_S2_short.pdf}
}
\subfigure[ROME]{
\includegraphics[width=0.49\columnwidth]{img/ROME_SKIP_S2_short.pdf}
}
\caption{GDAS fails on NAS-Bench-1Shot1~\cite{zela2020nas} on CIFAR-10 when adding skip connection to the second search space. Notice that nodes with no out-degrees have no contribution to the output.
}
\label{fig:1shot1-gdas-fail}
\end{figure}
\section{Conclusion}In this paper, we highlight a new performance collapse of single-path NAS paradigm based on Gumbel reparameterization, and attribute it to inconsistent searching for topology and stochastic nature of sampling candidate operations.
Therefore, we propose topology disentanglement for consistent search and gradient accumulation mechanism to reduce the sampling variance.
Unlike the previous single-path based works that suffer from instability issue, our method robustly achieves state-of-the-art results across 15 recent popular benchmarks.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Conclusion}
We considered the problem of unsupervised video representation learning, and introduced
Hierarchically Decoupled Spatial-Temporal Contrast (HDC). By decomposing the target into subtasks emphasizing different features and performing learning in a hierarchical manner, HDC is able to capture both rich spatial and temporal semantics at multiple scales. Extensive experiments of action recognition and nearest neighbor retrieval on UCF101 and HMDB51 using 3 different backbones suggest the potential of manipulating augmentations as regularization and demonstrate the state-of-the-art performance of HDC.
\section{Experiments}
We follow a common protocol~\cite{misra2016shuffle} to evaluate the
effectiveness of our HDC by using the learned representations as
initialization and fine-tuning on the downstream task of action
recognition on UCF101~\cite{soomro2012ucf101} and
HMDB51~\cite{kuehne2011hmdb}.
UCF101~\cite{soomro2012ucf101} consists of 13,320 videos and 101 classes
of human action. It has three train/test splits with a split ratio of about 7:3.
HMDB51~\cite{kuehne2011hmdb} is another widely-used action
recognition dataset containing 6,766 videos and 51
classes. It also has three splits with a similar split
ratio as UCF101. In our ablation studies, if not explicitly mentioned, we report Top-1 accuracy on only UCF101 split 1.
When comparing our method with other state-of-the-art, results
averaged on three splits of UCF101 and HMDB51 are reported.
In fine-tuning, we use the same network (C3D, 3D-ResNet18, or R(2+1)D-10) as we did
in self-supervised learning. A dropout layer~\cite{srivastava2014dropout} of ratio 0.5 is added after global average pooling, followed by a single fully-connected layer and softmax activation for classification. The instance
normalization layers are kept as they are. Blocks 1-5 are initialized with the learned weights by self-supervised
training on Kinetics-400. The last layer is randomly initialized.
During fine-tuning, we use stochastic gradient descent with learning
rate $0.01$, momentum $0.9$, decay $0.0001$, and $L2$ regularizer
$5e^{-5}$. Each mini-batch contains 32 clips, each
with 16 continuous frames randomly cropped to $128\times128\times3$.
Augmentations including random spatial cropping, scale
jittering, and horizontal flipping are applied,
resulting in an input of shape $32\times16\times112\times112\times3$.
During testing, each video is divided into non-overlapping 16-frame
clips. A center crop and four corner crops are taken for each
clip~\cite{wang2015towards}. The class score for each video is
obtained by averaging over all crops and clips.
\subsection{Ablation Studies}
\label{sec:ablation}
We conduct ablation experiments to analyze our design choices.
\begin{table}
\vspace{-12pt}
{\footnotesize{\textsf{
\begin{center}
\scalebox{1.0}{
\begin{tabular}{ccccc}
\toprule
SC & TC & $\alpha_k$ & $\beta_k$ & Top-1 Acc\\
\midrule
5 & - & 1 & - & 61.3\\
- & 5 & - & 1 & 62.7 \\
5 & 5 & 1 & 1 & 65.5 \\
4, 5 & 4, 5 & 1, 1 & 1, 1 & 66.9 \\
4, 5 & 4, 5 & 0.5, 1 & 0.5, 1 & 67.8 \\
3, 4, 5 & 3, 4, 5 & 1, 1, 1 & 1, 1, 1 & 66.8 \\
\midrule
\multicolumn{2}{c}{Training from scratch} & - & - & 43.5 \\
\midrule
3, 4, 5 & 3, 4, 5 & 0.25, 0.5, 1 & 0.25, 0.5, 1 & \textbf{69.0} \\
\bottomrule
\end{tabular}}
\end{center}
}}}
\vspace{-12pt}
\caption{Ablation study of decoupled contrast and hierarchical contrast. We use 3D-ResNet18 as the backbone, pretrain on Kinetics-400 and report top-1 accuracy on UCF101 split 1. SC and TC indicate the
scale at which we perform Spatial Contrast (SC) or Temporal
Contrast (TC) self-supervised learning -- i.e., the index of the
block we take features from. $\alpha_k$ and $\beta_k$
are defined in Eq.~\ref{eq:12} (listed in order
of scale).}
\vspace{-12pt}
\label{tab:ablation}
\end{table}
\xhdr{Decoupled Contrast.} We first evaluate the
effectiveness of decomposing the video representation learning task
into subtasks of Spatial and Temporal Contrast and performing joint
learning with them in Tab.~\ref{tab:ablation}. The first three rows
present the results with only Spatial Contrast learning, with only
Temporal Contrast learning, and with both Spatial Contrast and
Temporal Contrast learning. Hierarchical Contrast
learning is not involved and the learning is based on the features of
the last (5th) block. We see that, first, Temporal Contrast achieves slightly better performance than Spatial Contrast, perhaps because temporal semantics are more important in video learning. Then, by
incorporating both Spatial and Temporal Contrast, performance improves
significantly. This indicates that Spatial Contrast and Temporal
Contrast learn complementary features, and validates our hypothesis that
applying different augmentations as a way of regularization guides
the network to learn different features. Furthermore, the nearest neighbor retrieval results in Fig.~\ref{fig:pre} and Fig.~\ref{fig:retri} show that Spatial Contrast and Temporal Contrast focus on spatial and temporal features respectively.
We note that Spatial Contrast in and of itself can be viewed as directly
migrating the basic contrastive learning model from the image to video
domain, while Temporal Contrast learning adapts to this new domain by
further applying temporal augmentations to create the sample
variant. Thus the results suggest that
traditional contrastive learning models should be further adapted to
learn a good embedding space for video, and we present
one possible solution.
\xhdr{Hierarchical Contrast.}
Comparing rows 3 with 4 or 6
of Tab.~\ref{tab:ablation}, we observe that
Hierarchical Contrast learning at more scales yields better performance. This suggests that instance-level
invariance widely exists for mid-level and high-level features from
previous layers, and can be used to capture multi-scale
semantics. However, rows 4 and 6 show that simply adding more scales does not consistently bring improvement. We argue that this is because instance-level
invariance may be weaker for mid-level features, and adding more scales while giving them the same weight of significance will distract the network and harm the learning. We discuss this
below.
\xhdr{Significance of Instance-level Invariance at Different Scales.}
Zeiler et al.~\cite{zeiler2014visualizing} showed
that early layers of neural networks learn low-level features
which change a lot due to augmentations. As we perform
Hierarchical Contrast learning at more scales, those mid-level
features may not share the same level of invariance against
augmentations as the last layer's features. We use $\alpha_k$ and $\beta_k$ to model the significance of instance-level invariance at different
scales. We conducted experiments with different values of $\alpha_k$ and
$\beta_k$, and the results (row 4 vs row 5, row 6 vs the last row
in Tab.~\ref{tab:ablation}) show that smaller weights for lower
levels of the hierarchy yields better performance than assigning 1's. This
suggests that significance decreases at lower layers, which
is consistent with~\cite{zeiler2014visualizing}.
We did not exhaustively tune $\alpha_k$ and $\beta_k$, so
better performance is likely possible through tuning.
\begin{table}
\vspace{-12pt}
{\footnotesize{\textsf{
\begin{center}
\scalebox{0.85}{
\begin{tabular}{ccccc}
\toprule
color jittering & channel replication & flipping & scale jittering & Top-1 Acc\\
\midrule
\xmark & & & & 65.4\\
& \xmark & & & 61.0 \\
\xmark & \xmark & & & 49.4 \\
& & \xmark & & 68.0 \\
& & & \xmark & 64.7 \\
\midrule
\multicolumn{4}{c}{Our full model (HDC)} & \textbf{69.0} \\
\bottomrule
\end{tabular}}
\end{center}
}}}
\vspace{-12pt}
\caption{Ablation study of different spatial augmentations. We use our full HDC model with 3D-ResNet18 as the backbone, pretrain on Kinetics-400, and show top-1 accuracy on UCF101 split 1. An \xmark~indicates the transformation is not used during pretraining to generate augmented variants.}
\vspace{-12pt}
\label{tab:augab}
\end{table}
\xhdr{Spatial augmentation ablations.} Tab.~\ref{tab:augab} shows
ablation results of using different spatial augmentations. We find
that channel replication is crucial for the model to learn good
features, perhaps because it is a non-linear projection of RGB channels
which can effectively prevent the network from learning trivial
solutions based on color distribution~\cite{lee2017unsupervised}. As
another way to prevent such trivial solutions, color jittering
uses a linear function and thus is less effective. However, when
neither color jittering nor channel replication is used, the accuracy
drops greatly, indicating the network may suffer from the trivial
solution.
We note that, as shown in~\cite{chen2020simple}, there are other
spatial augmentations which can further improve the performance of
contrastive learning. However, the purpose of this paper is not to
exhaustively explore effects of different augmentations. By applying a
set of simple augmentations, we show the generalizability of our HDC.
\xhdr{Spatial augmentations in Temporal Contrast.} We verify the
importance of applying spatial augmentations after temporal random
cropping in Temporal Contrast. We use our full model of HDC with C3D
as the backbone, pretrain on Kinetics-400, and report average accuracy
on 3 splits of UCF101 and HMDB51. It achieves $72.3\%$ on UCF101 and
$39.3\%$ on HMDB51 when spatial augmentations are used in Temporal
Contrast versus only $68.9\%$ on UCF101 and $38.0\%$ on HMDB51
when spatial augmentations are not used. This supports our design
choice of having spatial augmentations in Temporal Contrast to obtain
variants whose spatial context varies as much as possible.
\setlength{\tabcolsep}{2pt}
\begin{table}
\centering
\vspace{-12pt}
{\footnotesize{\textsf{
\begin{center}
\scalebox{0.9}{
\begin{tabular}{lccccc}
\toprule
\multicolumn{2}{l}{Self-supervised Learning} & & \multicolumn{2}{c}{Accuracy ($\%$)} \\
\cmidrule{1-2} \cmidrule{4-5}
Method & Architecture & & HMDB51 & UCF101\\
\midrule
Random Initialization~\cite{misra2016shuffle} & CaffeNet && 13.3 & 38.6 \\
Shuffle \& Learn~\cite{misra2016shuffle} & CaffeNet && 18.1 & 50.2 \\
B{\"u}chler et al.~\cite{buchler2018improving} & CaffeNet && 25.0 & 58.6 \\
\midrule
Random Initialization~\cite{lee2017unsupervised} & VGG-M-2048 && 18.3 & 51.1 \\
OPN~\cite{lee2017unsupervised} & VGG-M-2048 && 23.8 & 59.8 \\
\midrule
Random Initialization~\cite{xu2019self} & R3D-18 && 21.5 & 54.4 \\
Clip Order~\cite{xu2019self} & R3D-18 && 29.5 & 64.9 &\\
VCP~\cite{luo2020video} & R3D-18 && 31.5 & 66.0 \\
\midrule
Random Initialization~\cite{cyclecontrast} & R3D-18+1 && 19.4 & 44.7 \\
CCL~\cite{cyclecontrast} & R3D-18+1 && 37.8 & 69.4 \\
\midrule
Random Initialization~\cite{han2019video} & 2D3D-ResNet18 && 17.1 & 46.5 \\
DPC~\cite{han2019video} & 2D3D-ResNet18 && 34.5 & 68.2 \\
\midrule
Random Initialization~\cite{benaim2020speednet} & I3D && \textbf{29.6} & 47.9 \\
SpeedNet~\cite{benaim2020speednet} & I3D && \textbf{43.7} & 66.7 \\
\midrule
Random Initialization~\cite{3dresnet} & 3D-ResNet18 && 17.1 (19.1) & 42.4 (43.8) \\
3D-RotNet~\cite{rotnet} & 3D-ResNet18 && 33.7 & 62.9 \\
3D-ST-Puzzle~\cite{kim2019self} & 3D-ResNet18 && 33.7 & 65.8 \\
\midrule
\textbf{Our method (HDC\textdagger)} & \textbf{3D-ResNet18} && \textbf{38.4} & \textbf{69.8} \\
\textbf{Our method (HDC)} & \textbf{3D-ResNet18} && \textbf{38.1} & \textbf{68.5} \\
\midrule
Random Initialization~\cite{xu2019self} & C3D && 23.2 (22.5) & 61.6 (51.4) \\
Motion \& Appearance~\cite{wang2019self} & C3D && 33.4 & 61.2 \\
Clip Order~\cite{xu2019self} & C3D && 28.4 & 65.6 \\
VCP~\cite{luo2020video} & C3D && 32.5 & 68.5 \\
Cho et al.~\cite{cho2020self} & C3D && 34.3 & 70.4 \\
\midrule
\textbf{Our method (HDC\textdagger)} & \textbf{C3D} && \textbf{39.5} & \textbf{73.1} \\
\textbf{Our method (HDC)} & \textbf{C3D} && \textbf{39.3} & \textbf{72.3} \\
\midrule
Random Initialization~\cite{xu2019self} & R(2+1)D-10 && 22.0 & 56.2 \\
Clip Order~\cite{xu2019self} & R(2+1)D-10 && 30.9 & 72.4 \\
VCP~\cite{luo2020video} & R(2+1)D-10 && 32.2 & 66.3 \\
Cho et al.~\cite{cho2020self} & R(2+1)D-10 && 36.8 & 74.8 \\
PacePrediction~\cite{pacepredict} & R(2+1)D-10 && 36.6 & \textbf{77.1} \\
\midrule
\textbf{Our method (HDC\textdagger)} & \textbf{R(2+1)D-10} && \textbf{40.0} & \textbf{76.8} \\
\textbf{Our method (HDC)} & \textbf{R(2+1)D-10} && \textbf{39.8} & \textbf{76.2} \\
\bottomrule
\end{tabular}}
\end{center}
}}}
\vspace{-6pt}
\caption{Top-1 accuracy averaged on 3 splits of
UCF101 and HMDB51. Parentheses show accuracy obtained with
our own implementation. HDC with~\textdagger~is pretrained by decreasing learning rate more slowly.}
\label{tab:compa}
\vspace{-12pt}
\end{table}
\setlength{\tabcolsep}{4pt}
\begin{table}
\centering
\vspace{-12pt}
{\footnotesize{\textsf{
\begin{center}
\scalebox{1.0}{
\begin{tabular}{lccccc}
\toprule
Methods & Top1 & Top5 & Top10 & Top20 & Top50\\
\midrule
OPN~\cite{lee2017unsupervised} & 19.9 & 28.7 & 34.0 & 40.6 & 51.6 \\
B{\"u}chler et al.~\cite{buchler2018improving} & 25.7 & 36.2 & 42.2 & 49.2 & 59.5 \\
\midrule
Random Initialized C3D & 16.7 & 27.5 & 33.7 & 41.4 & 53.0 \\
Clip Order (C3D)~\cite{xu2019self} & 12.5 & 29.0 & 39.0 & 50.6 & 66.9 \\
VCP (C3D)~\cite{luo2020video} & 17.3 & 31.5 & 42.0 & 52.6 & 67.7 \\
\midrule
CCL (R3D-18+1)~\cite{cyclecontrast} & 22.0 & 39.1 & 44.6 & 56.3 & 70.8 \\
\midrule
\textbf{Our HDC (C3D)} & \textbf{33.9} & \textbf{49.6} & \textbf{55.7} & \textbf{61.6} & \textbf{69.9} \\
\bottomrule
\end{tabular}}
\end{center}
}}}
\vspace{-12pt}
\caption{Nearest neighbor retrieval results on UCF101.}
\label{tab:nnrucf}
\end{table}
\begin{table}
\vspace{-12pt}
{\footnotesize{\textsf{
\begin{center}
\scalebox{1.0}{
\begin{tabular}{lccccc}
\toprule
Methods & Top1 & Top5 & Top10 & Top20 & Top50\\
\midrule
Random Initialized C3D & 7.4 & 20.5 & 31.9 & 44.5 & 66.3 \\
Clip Order (C3D)~\cite{xu2019self} & 7.4 & 22.6 & 34.4 & 48.5 & 70.1 \\
VCP (C3D)~\cite{luo2020video} & 7.8 & 23.8 & 35.3 & \textbf{49.3} & \textbf{71.6} \\
\midrule
\textbf{Our HDC (C3D)} & \textbf{14.6} & \textbf{28.8} & \textbf{36.1} & \textbf{44.8} & \textbf{57.9} \\
\bottomrule
\end{tabular}}
\end{center}
}}}
\vspace{-12pt}
\caption{Nearest neighbor retrieval results on HMDB51.}
\label{tab:nnrhmdb}
\vspace{-12pt}
\end{table}
\begin{figure*}[t]
\begin{center}
\includegraphics[width=1.0\textwidth]{figures/nnr_cvpr.pdf}
\end{center}
\vspace{-18pt}
\caption{Sample results of nearest neighbor retrieval on UCF101 split 1, showing the top-1 retrieved clip from our full model (\textit{HDC}), and Spatial Contrast (\textit{SC}), Temporal
Contrast (\textit{TC}), and Joint Spatial-Temporal Contrast
(\textit{SC + TC}) models. Below each clip is its
action category, with green underlined text indicating correct retrievals and red text indicating incorrect. Our proposed HDC achieved better results because of the ability to capture both spatial and temporal features at multiple scales.}
\label{fig:retri}
\vspace{-12pt}
\end{figure*}
\subsection{Comparison with the State-of-the-art}
For fairness, we compare with methods under similar settings. There are other methods that achieve great success. But they adopt much more advanced backbones~\cite{cvrl, yao2020seco, diba2019dynamonet}, require extra preprocessing~\cite{timearrow, Han20, han2020self}, or needs additional information to prepare the input~\cite{korbar2018cooperative,alwassel2019self, sun2019learning, miech2020end}.
We report top-1 accuracy averaged over 3 splits of UCF101 and HMDB51
in Tab.~\ref{tab:compa}. HDC achieves better or comparable performance on both datasets. We note that: 1) PacePrediction~\cite{pacepredict} shows slightly better results on UCF101, perhaps because its best performance is achieved by further adapting contrastive laerning in addition to pace prediction; 2) SpeedNet~\cite{benaim2020speednet} benefits from the backbone ($29.6\%$ when trained from scratch) and thus has better performance on HMDB51.
By comparing HDC variants using
different backbones, we observe that a more advanced backbone always
leads to better performance, suggesting that HDC will be able to further benefit from future advances in network architectures. We also find that decreasing the learning rate more slowly is beneficial, perhaps because it allows better optimization.
\subsection{Nearest Neighbor Retrieval}
\label{sec:retrieval}
We follow~\cite{xu2019self} and perform nearest neighbor retrieval experiments.
As shown in Tables~\ref{tab:nnrucf} and~\ref{tab:nnrhmdb}, our
method significantly outperforms other methods on both UCF101 and
HMDB51. This implies that we learn better
features, and explains the good performance on downstream tasks.
As shown in Fig.~\ref{fig:pre} and Fig.~\ref{fig:retri}, qualitative
results further support our idea of manipulating augmentations to
guide the network to learn different features and the benefit of
hierarchical learning. For example, in the first row of the right
column in Fig.~\ref{fig:retri}, HDC succeeds in retrieving a clip of
the correct action while the other variants fail: SC focuses more on
spatial information of people outdoors, TC pays more attention to the
actions of people walking, and ST+TC fails perhaps because it does not
learn features at different hierarchies, although it successfully
captured both spatial and temporal information of multiple moving
people.
\section{Introduction}
\label{sec:intro}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.48\textwidth]{figures/intro_cvpr.pdf}
\end{center}
\vspace{-16pt}
\caption{Sample results of nearest neighbor retrieval on UCF101 split 1 demonstrate the effectiveness of our approach for decoupling spatial-temporal feature learning. We compare the top-1 retrieved clip from our full model (\textit{HDC}) with the clip retrieved by models performing only one subtask of HDC (Spatial Contrast (\textit{SC}) or Temporal Contrast (\textit{TC})). Below each clip is its corresponding action category, with green underlined text indicating a correct retrieval and red text indicating an incorrect retrieval. During retrieval, SC tends to focus more on spatial similarity with the query clip, while TC focuses more on temporal similarity. For example, in the first row, SC's retrieval is perhaps based on spatial similarity of people indoors, while TC's retrieval may be due to temporal similarity of crawling. Since our proposed HDC solves SC and TC hierarchically and simultaneously, it is able to capture both spatial and temporal semantics and made the correct retrieval.}
\label{fig:pre}
\vspace{-12pt}
\end{figure}
As a solution to the growing need for large-scale labeled data for
training complex neural network
models~\cite{alexnet,resnet,inceptionv1,c3d,i3d}, unsupervised
representation learning aims to learn good feature embeddings from
data without annotations. Using the learned representations as
initialization, downstream tasks only need to be fine-tuned on a
relatively small labeled dataset in order to yield reasonable
performance. Much recent progress in unsupervised image representation
learning~\cite{wu2018unsupervised,ye2019unsupervised,bachman2019learning,hjelm2018learning,oord2018representation,henaff2019data,he2019momentum,misra2019self,tian2019contrastive}
is driven by contrastive learning. While they solve the same pretext task of
instance-level variant matching, these methods differ in how they
obtain variant embeddings of the same instance, e.g., using
augmentations~\cite{wu2018unsupervised,ye2019unsupervised,bachman2019learning,hjelm2018learning},
future representations~\cite{oord2018representation}, or momentum
features~\cite{he2019momentum}. By optimizing a contrastive
loss~\cite{hadsell2006dimensionality}, they are essentially maximizing
intra-instance embedding similarity and minimizing inter-instance
embedding similarity, which leads to a spread-out feature
space~\cite{zhang2017learning}.
Compared with the success of contrastive learning for images, most
state-of-the-art unsupervised video representation learning work
relies on context-based proxy tasks such as time arrow
classification~\cite{timearrow} or clip order
prediction~\cite{xu2019self}. While the results of some recent
methods~\cite{sun2019learning, han2019video, cyclecontrast, Han20}
suggest the potential of contrastive learning for videos, they work
in a ``one for all'' manner by solving a single contrastive learning
task with the activation maps from the final layer, expecting the model to
capture all the features through the learning procedure. As a result,
the learned model may lack a general understanding of spatial and
temporal semantics, and instead just memorize spatial-temporal combinations.
In this paper, we argue that good video representations should be able
to capture spatial and temporal features in a more general form at
multiple scales, and thus it is helpful to decompose the overall
goal of learning spatial-temporal features into hierarchical subtasks,
respectively emphasizing spatial and temporal features.
To this end, we present Hierarchically Decoupled Spatial-Temporal Contrast (HDC),
in which we decouple the learning objective into separate
subtasks of Spatial Contrast and Temporal
Contrast and perform the learning hierarchically.
Neural networks are notorious for learning shortcuts to
``cheat''~\cite{doersch2015unsupervised,timearrow}: if
there is an easy way to solve the problem, the network will hardly
try to find a more complex solution. We make use of this property for
the decoupling and direct the network to learn different features by
manipulating augmentations. In particular, in the Spatial Contrast
subtask, we deliberately provide a shortcut by creating augmented
variants with only spatial augmentations (random spatial cropping,
color jittering, etc.). As the timestamps of the query clip and its
augmented copy are the same, it is possible to solve the matching task
based merely on consistency of spatial semantics, and thus the network
will try to ``cheat'' by focusing more on spatial features. In
Temporal Contrast learning, we randomly select a new clip from the
video of the query clip (random temporal cropping) before applying
spatial augmentations in order to obtain a variant whose spatial
semantics are as different from the query clip as possible. Since
spatial context may vary dramatically after applying the temporal and
spatial transformations, the model is prevented from ``cheating''
through spatial similarity and encouraged to rely more on similarity
of temporal semantics to solve the pretext task. Fig.~\ref{fig:pre}
shows nearest neighbor retrieval results demonstrating the
effectiveness of our approach for the decoupling, and suggests the potential of directing the network to learn desired features through manipulating augmentations as regularization.
In order to capture multi-scale features, we further perform Spatial
Contrast and Temporal Contrast learning hierarchically by optimizing
towards a compound loss. During hierarchical learning, we model the
significance of instance-wise consistency in a given layer with
different weights, because features from different layers do not share
the same level of invariance against
augmentations~\cite{zeiler2014visualizing}.
\newenvironment{itemize*}%
{\begin{itemize}[leftmargin=*]%
\setlength{\itemsep}{0pt}%
\setlength{\parskip}{0pt}}%
{\end{itemize}}
In summary, our contributions are as follows:
\begin{itemize*}
\item We demonstrate the effectiveness of spatial-temporal feature learning decoupling and hierarchical learning in the context of unsupervised learning for the first time.
\item We show that the network can be guided to learn desired features in contrastive learning by manipulating augmentations as regularization, and introduce a new way for the network to separately capture spatial and temporal semantics in self-supervised video representation learning.
\item We propose an approach to improve hierarchical contrastive learning by modeling the divergent levels of invariance in different layers as different loss weights.
\item By optimizing a novel compound loss, our Hierarchically Decoupled Spatial-Temporal Contrast (HDC) outperforms
other unsupervised methods and sets a new
state-of-the-art on downstream tasks of action recognition on
UCF101 and HMDB51.
\end{itemize*}
\section{Our Method}
To explore the potential of self-supervised video representation
learning merely from RGB clips, we formulate a novel pretext task of
Hierarchically Decoupled Spatial-Temporal Contrast (HDC) in which
we maximize intra-instance representation similarity and minimize
inter-instance representation similarity spatially (Fig.~\ref{fig:stc}), temporally (Fig.~\ref{fig:stc}), and
hierarchically (Fig.~\ref{fig:histc}).
\subsection{Decoupled Contrast}
\label{sec:stc}
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.85\textwidth]{figures/intro-s.pdf}
\end{center}
\vspace{-12pt}
\caption{A simple example of how our Decoupled Contrast
works between two sample videos. In Spatial Contrast (SC), two sets of spatial transformations, $\varphi_{o,i}$ and $\varphi_{s,i}$, are sampled from families of candidate spatial transformations $\Phi_o$ and $\Phi_s$, and applied to the same clip $x_i$ to obtain variants $u_{o,i}$ and $u_{s,i}$ for SC learning. In Temporal Contrast (TC), we first extract another clip $x'_i$ from the same video as $x_i$ by random temporal cropping $\Gamma(\cdot)$, and then apply $\varphi_{t,i}$ sampled from families of candidate spatial transformations $\Phi_t$ to obtain a variant $u_{t,i}$, which will be used with $u_{o,i}$ for TC learning.}
\label{fig:stc}
\vspace{-12pt}
\end{figure*}
Motivated by the observation in supervised learning that factoring 3D filters into separate spatial and temporal components yields significant gains~\cite{tran2018closer}, we propose to decouple the overall objective of unsupervised spatial-temporal feature learning into separate subtasks and provide regularizations to guide them to emphasize spatial and temporal
features, respectively (Fig.~\ref{fig:stc}).
\textbf{Spatial Contrast} is designed to focus on learning spatial
representations. Neural networks are notorious for learning shortcuts
to ``cheat''~\cite{doersch2015unsupervised,timearrow}. While previous
work tries to avoid all
cheating~\cite{doersch2015unsupervised,timearrow,lee2017unsupervised,kim2019self},
we intentionally make use of this property. In Spatial Contrast, the
network is provided with a shortcut by augmenting clips with only
spatial transformations. This allows the model to ``cheat'' by using
only spatial features to capture intra-instance similarity and
inter-instance difference. In other words, the network will not bother
with temporal features because capturing
spatial similarity is already enough for solving the matching task.
In particular, as shown in Fig.~\ref{fig:stc}, given a batch of clips $X=[x_1,x_2,...,x_B]$, each
from a different video, we augment each clip $i$ with
augmentations $\varphi_{o,i}$ and $\varphi_{s,i}$ sampled from
families of candidate spatial transformations $\Phi_o$ and $\Phi_s$
to obtain corresponding variants $u_{o,i}= \varphi_{o,i}(x_i),$
$\varphi_{o,i}(\cdot)\sim\Phi_o$ and $u_{s,i} = \varphi_{s,i}(x_i),$
$\varphi_{s,i}(\cdot)\sim\Phi_s$, which are only spatially-augmented.
Our primary goal is to train an encoder so that the similarity between
feature embeddings of $u_{o,i}$ and $u_{s,j}$ is maximized when $i=j$,
and minimized otherwise. Let $f(\cdot)$ denote the encoder. The
embeddings of $u_{o,i}$ and $u_{s,j}$ are $v_{o,i} = f(u_{o,i})$ and
$v_{s,j} = f(u_{s,j})$, where $v_{o,i}$ and $v_{s,j}$ are both vector
embeddings, e.g., obtained through global pooling.
By measuring similarity with cosine distance
$sim(v_{o,i},v_{s,j}) = (v_{o,i}\cdot
v_{s,j})/(|v_{o,i}|\cdot|v_{s,j}|)$, our goal can be achieved by
optimizing a contrastive loss called
InfoNCE~\cite{oord2018representation},
\begin{equation}
L_{s} = -\sum_{i=1}^{B}\log\frac{\exp(sim(v_{o,i},v_{s,i})/\tau)}{\sum_{j=1}^{B} \exp(sim(v_{o,i},v_{s,j})/\tau)},
\label{eq:6}
\end{equation}
\noindent
where $\tau$ is a temperature controlling the concentration
of the feature embedding distribution~\cite{hinton2015distilling} (usually around $0.1$).
Eq.~\ref{eq:6} can be viewed as a cross entropy loss
between a pseudo prediction $S \in R^{B \times B}$ and a pseudo label
$\Bar{S} \in R^{B \times B}$. $\Bar{S}$ is an identity matrix and $S$
is a similarity matrix where $S_{ij}$ measures the similarity of one
clip variant $u_{o,i}$ to another variant $u_{s,j}$ in the feature
space,
\begin{equation}
S_{ij} = \frac{\exp(sim(v_{o,i},v_{s,j})/\tau)}{\sum_{m=1}^{B} \exp(sim(v_{o,i},v_{s,m})/\tau)}.
\label{eq:7}
\end{equation}
In this spirit, the Spatial Contrast subtask is a self-supervised
multi-way classification problem in which we want to match one clip
variant to another when they come from the same video and distinguish
it from variants coming from different videos.
\textbf{Temporal Contrast} emphasizes learning temporal
representations. To do this, we need to prevent the network from
cheating through spatial feature similarity: we want variants whose
spatial context varies dramatically from the original clip, but whose
temporal context remains nearly the same and thus is essential for
capturing instance-level invariance. To do this, we add temporal
augmentations of random temporal cropping before applying spatial
transformations to produce another temporally-augmented variant. Note
that spatial augmentations are crucial here to further alter spatial
context in order to prevent cheating, and we show its importance
in Section~\ref{sec:ablation}.
Specifically, to produce a temporally-augmented variant
$u_{t,j}$ for each clip $x_j$, random temporal cropping
$\Gamma(\cdot)$ is first applied, followed by spatial augmentations
$\varphi_{t,j}$ sampled from another family of candidate spatial
transformations $\Phi_t$, after which we have $u_{t,j} =
\varphi_{t,j}(\Gamma(x_j)), ~\varphi_{t,j}(\cdot)\sim\Phi_t$.
Temporal Contrast is modeled between $u_{o,i}$ and $u_{t,j}$
using a similar technique as in Spatial Contrast. Then
we minimize,
\begin{equation}
L_{t} = -\sum_{i=1}^{B}\log\frac{\exp(sim(v_{o,i},v_{t,i})/\tau)}{\sum_{j=1}^{B} \exp(sim(v_{o,i},v_{t,j})/\tau)}.
\label{eq:9}
\end{equation}
As with Spatial Contrast learning, the
Temporal Contrast subtask is also a self-supervised multi-way
classification problem of matching clip variants of the same video.
\subsection{Hierarchical Contrast}
\label{sec:histc}
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.85\textwidth]{figures/model.pdf}
\end{center}
\vspace{-12pt}
\caption{An example illustrating Hierarchical Contrast. Here we only show learning among variants from the same video and thus the network is maximizing similarity between all corresponding pairs. Otherwise if variants are from different videos, the network will minimize similarity.
Feature maps at multiple scales are 2D global average pooled along the
spatial dimension (\textit{S-Avg Pooling}) in Hierarchical Spatial Contrast
learning, and 3D global average pooled along the temporal and spatial dimensions (\textit{G-Avg Pooling}) in Hierarchical Temporal Contrast learning. The channel dimension is omitted for clarity.}
\label{fig:histc}
\vspace{-12pt}
\end{figure*}
Inspired by the effectiveness of multi-scale features in supervised learning~\cite{zhou2018temporal, hussein2019timeception}, we introduce hierarchical learning to unsupervised learning by conducting Decoupled
Contrast learning hierarchically. As illustrated in
Fig.~\ref{fig:histc}, feature maps from different layers or blocks of
the encoder $f(\cdot)$ are collected and pooled to produce multi-scale
vector embeddings. Two pooling strategies are applied to fully
leverage instance-level consistency among different layers: (1)
Temporal Contrast uses 3D global average pooling along both
temporal and spatial dimensions, and (2) Spatial Contrast performs
2D global average pooling only along the spatial dimension. Therefore,
for each scale, we obtain one vector representation for each clip
variant in Hierarchical Temporal Contrast learning, but may get more
than one vector for each variant in Hierarchical
Spatial Contrast learning, depending on the length of input clip and
the temporal downsample factor of the encoder for a certain layer. Our
motivation is that because the timestamps in Spatial Contrast are
the same for two clip variants from the same video, the
spatial instance-level invariance should exist not only for the whole
clip, but also for corresponding sub-clips.
As shown in Fig.~\ref{fig:histc}, for Hierarchical Spatial Contrast learning at scale $k$,
we apply 2D global average pooling and
obtain vector embeddings of $v^{k,n}_{o,i}$, $n=1,2,...,N$ for clip
variant $u_{o,i}$ and vector embeddings of $v^{k,n}_{s,j}$,
$n=1,2,...,N$ for variant $u_{s,j}$, where $N$ is the size
of the temporal dimension of the feature maps. The loss
at scale $k$ becomes,
\begin{equation}
L^k_{s} = -\frac{1}{N}\sum_{n=1}^{N}\sum_{i=1}^{B}\log\frac{\exp(sim(v^{k,n}_{o,i},v^{k,n}_{s,i})/\tau)}{\sum_{j=1}^{B} \exp(sim(v^{k,n}_{o,i},v^{k,n}_{s,j})/\tau)},
\label{eq:10}
\end{equation}
\noindent
while the loss for
Hierarchical Temporal Contrast remains
almost the same for scale $k$,
\begin{equation}
L^k_{t} = -\sum_{i=1}^{B}\log\frac{\exp(sim(v^k_{o,i},v^k_{t,i})/\tau)}{\sum_{j=1}^{B} \exp(sim(v^k_{o,i},v^k_{t,j})/\tau)},
\label{eq:11}
\end{equation}
where $v^k_{o,i}$ are $v^k_{t,i}$ are obtained from 3D global average pooling.
\subsection{Self-supervised Learning with HDC}
We consider 3 different backbones, C3D~\cite{c3d}, 3D-ResNet18~\cite{resnet,3dresnet} and R(2+1)D-10~\cite{tran2018closer}, as our
encoder $f(\cdot)$ for fair comparison with existing methods. All
fully-connected layers are removed (last 3 layers of C3D, and last
layer of 3D-ResNet18 and R(2+1)D-10). Batch normalization layers~\cite{batchnorm} in
3D-ResNet18 and R(2+1)D-10 are replaced with instance normalization
layers~\cite{ulyanov2016instance} to prevent cheating via batch
statistics~\cite{henaff2019data,he2019momentum}; for C3D, an instance
normalization layer is inserted after each convolution layer. For
multi-scale learning, we consider feature maps from blocks
$3, 4, 5$ (last 3 blocks). Instead of directly performing the matching pretext task
with the feature vectors from pooling, we project each vector to a
lower-dimension space with linear projections,
following~\cite{wu2018unsupervised}. In particular, there are
projections $g^k_o(\cdot)$, $g^k_s(\cdot)$, and $g^k_t(\cdot)$ for
projecting $v^k_{o,i}$, $v^k_{s,i}$, and $v^k_{t,i}$, respectively, at
each scale $k$, each implemented as a single fully-connected layer
with linear activation projecting a vector into 128-d.
(See~\cite{chen2020simple} for more experiments and discussion about
the importance of such linear projections.) The families of spatial
augmentations $\Phi_o$, $\Phi_s$, and $\Phi_t$ contain the same set of
transformations of random spatial cropping, scale jittering,
horizontal flipping, color jittering, and channel
replication~\cite{lee2017unsupervised}. We use $\tau=0.07$ for computing the contrastive loss.
We perform self-supervised training on
Kinetics-400~\cite{kay2017kinetics}, which has 400 action classes
and over 400 videos per class. We resize frames,
preserving aspect ratio, so that the shorter
side is 128 pixels. Each mini-batch contains 128 clips from 128 videos
and each clip consists of 16 randomly-cropped continuous frames of
shape $128 \times 128 \times 3$. After spatial and temporal
augmentations, three sets of clip variants are obtained, each of shape
$128 \times 16 \times 112 \times 112 \times 3$. Before being fed to
the encoder, these clips are rescaled following~\cite{i3d} so
pixel values are between -1 and 1.
Our model is implemented with Tensorflow~\cite{tensorflow} and
Keras~\cite{keras}. We use stochastic gradient descent with learning
rate $0.1$, momentum $0.9$, decay $0.0001$, and $L2$ regularizer
$5e^{-5}$.
HDC is trained as a whole towards minimizing a novel compound contrastive loss, namely HD-NCE,
\begin{equation}
L = \sum_k (\alpha_k \cdot L^k_s + \beta_k \cdot L^k_t),
\label{eq:12}
\end{equation}
where $L^k_s$ and $L^k_t$ are in Eqs.~\ref{eq:10}
and~\ref{eq:11}, and $k=3,4,5$ as we use features from blocks $3, 4, 5$ of
the encoder.
\xhdr{Significance of instance-wise
consistency.} $\alpha_k$ and $\beta_k$ represent
the significance of instance-wise
consistency in each layer, essentially weighting how much the corresponding subtask will contribute to our main goal. We simply tested (1)
$\alpha_3=\beta_3=\alpha_4=\beta_4=\alpha_5=\beta_5=1.0$, and (2)
$\alpha_3=\beta_3=0.25$, $\alpha_4=\beta_4=0.5$ and
$\alpha_5=\beta_5=1.0$, and the second set worked better (see
Section~\ref{sec:ablation}).
\subsection{Difference from Recent Work}
\label{sec:diff}
Some work~\cite{yang2020hierarchical, yang2020vthcl} also explores hierarchical
contrastive learning in videos. However, while HDC matches variants obtained with carefully designed augmentations, ~\cite{yang2020vthcl} brings closer features from the slow and fast streams of a Slowfast network~\cite{feichtenhofer2019slowfast}, and ~\cite{yang2020hierarchical} predicts future motion patterns. Moreover, our work introduces spatial-temporal learning decoupling, which is another novel component.
Another similar method is CVRL~\cite{cvrl}, which is similar in spirit to the variant of our model trained with only Temporal Contrast incorporated in the last layer. Our HDC decouples the learning into separate subtasks emphasizing spatial and temporal features and further performs the learning hierarchically. As later experiments (Sec.~\ref{sec:ablation} and Tab.~\ref{tab:ablation}) show, HDC achieves significant improvement over the CVRL-alike variant (the second row in Tab.~\ref{tab:ablation}).
\section{Related Work}
\newcommand{\xhdr}[1]{\vspace{6pt}\noindent{\textbf{#1}}}
\xhdr{Unsupervised Video Representation Learning}
was originally
based on input
reconstruction~\cite{vincent2008extracting,ranzato2007unsupervised,lee2007efficient,le2011learning,hinton2006reducing,hinton2006fast},
while more recent methods derive
implicit pseudo-labels from the
unlabeled data to use as self-supervision signals for the
corresponding pretext task~\cite{raina2007self,wang2015unsupervised}.
For example, several models use
chronological order of video frames to define proxy tasks
such as frame order prediction~\cite{lee2017unsupervised} or
verification~\cite{misra2016shuffle,fernando2017self}, clip
order prediction~\cite{xu2019self}, and time arrow
classification~\cite{timearrow}. Other pretext tasks, such as
spatial-temporal jigsaw~\cite{kim2019self}, future
prediction~\cite{vondrick2016generating,srivastava2015unsupervised,han2019video,mathieu2015deep,vondrick2016anticipating,lotter2016deep},
temporal correspondence
estimation~\cite{wang2019learning,jayaraman2016slow,wang2015unsupervised,isola2015learning},
audio-video clustering~\cite{alwassel2019self}, video
colorization~\cite{vondrick2018tracking}, motion and appearance
statistics prediction~\cite{wang2019self}, and loss distillation across multiple tasks~\cite{piergiovanni2020evolving} have also been explored.
\xhdr{Contrastive Learning}~\cite{hadsell2006dimensionality} is very
effective for unsupervised representation learning for
images~\cite{wu2018unsupervised,ye2019unsupervised,bachman2019learning,hjelm2018learning,oord2018representation,henaff2019data,he2019momentum,misra2019self,tian2019contrastive}.
These methods try to learn a feature space in which variants of the
same sample are close together while variants from different samples
are far apart, and mainly differ in how they create the variants. For
example, Oord~\etal~\cite{oord2018representation} predict the future
in the latent space as a variant of the real embeddings of the future.
In the video domain, Han~\etal~\cite{han2019video} predict dense
feature maps of future clips and match them with corresponding real
embeddings from other distractions. This idea is further extended to a
memory-augmented version~\cite{Han20} for
improvement. Sun~\etal~\cite{sun2019learning} proposed to use
bidirectional transformers for multi-modal contrastive learning from
text and videos. A cycle-contrastive loss inspired by
CycleGAN~\cite{CycleGAN2017} is presented by
Kong~\etal~\cite{cyclecontrast} to use the relationship
between videos and frames. Recasens \etal~\cite{recasens2021broaden} and Wang \etal~\cite{wang2021long} both explore a new pretext task of matching short-term view to long-term view of the same video. Tschannen \etal~\cite{tschannen2019self}
also use video-induced invariance to formulate a pretext task
based on contrastive learning. Despite its motivation, interestingly,
it aims to learn image representations instead of video embeddings.
There is more discussion in Sec.~\ref{sec:diff}.
|
\section{Introduction}
The structure of link homology theories is closely related to the representation theory of Lie algebras, and to categorification of quantum groups. This paper initiates the study of lifting these relations to the level of stable homotopy types. In the process of doing so, we provide the first example (to our knowledge) of constructing a stable homotopy refinement via framed flow categories of a homological invariant involving signs and cancellations.
We work in the setting of the $\sl_2$ homology theory $Kh_{\mathbb{A}}(L)$ for links $L$ in the thickened annulus, sometimes referred to as sutured annular Khovanov homology. Following constructions in \cite{APS, BN2, Ro}, this triply graded theory may be obtained from the usual Khovanov chain complex \cite{Kh1} of a link diagram $D$ by taking the annular degree zero part of the differential. It was shown in \cite{BPW} that given a tangle $T$, the homology $Kh_{\mathbb{A}}(\widehat T)$ of its annular closure may also be obtained as the Hochschild homology of the complex of bimodules over the Chen-Khovanov algebra \cite{CK} associated to $T$.
Motivated in part by results of Auroux, Grigsby and Wehrli \cite{AGW} and by the work of Lauda \cite{Lauda} on categorified $\sl_2$, Grigsby, Licata and Wehrli showed \cite{GLW} that the annular Khovanov homology $Kh_{\mathbb{A}}(L)$ carries an action of $\sl_2$. Analogous actions of $\sl_n$ on annular Khovanov-Rozansky homology were defined in \cite{QR}. In fact, the $\sl_2$ action is constructed on the chain complex level, and it is natural with respect to annular link cobordisms. Here a trivial simple closed curve in the annulus is assigned the trivial representation of $\sl_2$, and essential curves are assigned the fundamental representation or its dual; \cite{GLW} extended the action to $Kh_{\mathbb{A}}(L)$ for any annular link $L$ and showed that the action is a link invariant; see Section \ref{sec:sl2 action} for a detailed discussion.
Our results concern the stable homotopy refinement of Khovanov homology, introduced by Lipshitz and Sarkar \cite{LS}. This construction, building on the work of Cohen, Jones and Segal \cite{CJS}, associates to a diagram of an oriented link a framed flow category. The resulting suspension spectrum is a stable homotopy invariant of the link $L$; its cohomology is isomorphic to the Khovanov homology of $L$. An alternative, more combinatorial construction using the Burnside category was given by Lawson, Lipshitz and Sarkar in \cite{LLS}.
The stable homotopy refinement $\X_\mathbb{A}(L)$ of an annular link $L$ may be constructed using an analogous method, cf. \cite[Section 4.3]{AKW}. Alternatively, lifting the homology isomorphism with the Hochschild homology of Chen-Khovanov bimodules, $\X_\mathbb{A}(L)$ may be defined as the topological Hochschild homology of the tangle invariants introduced by Lawson-Lipshitz-Sarkar in \cite{LLSCK}.
These new links invariants valued in spectra can be studied using tools of stable homotopy theory. For example, they admit an action of the Steenrod algebra, and they are known to be stronger invariants than the underlying link homology, cf. \cite{Seed}. It is an interesting question what features of link homology admit a lift to the level of spectra. We address this question, giving a new structure on the annular Khovanov spectrum and thus providing a new invariant of annular links and cobordisms. To this end, our main result is the following. Let $J$ denote one of the standard generators $E, F, H$ of $\sl_2$ (see Section \ref{sec:sl2 action}).
\begin{theorem}\label{main thm} Given an annular diagram $D$ of a link $L$ in the thickened annulus, there exists a map of spectra $\mathcal{J}:\X_\mathbb{A}(D) \rightarrow \X_\mathbb{A}(D)$, whose induced map $\mathcal{J}^*$ on cohomology is equal to the action of $J$ on $Kh_{\mathbb{A}}(L)$. The homotopy class of $\mathcal{J}$ is an invariant of the link $L$.
\end{theorem}
As a consequence, we have the following statement about the $\sl_2$ action on annular Khovanov homology.
\begin{corollary}
The action of $\sl_2$ on the annular homology $Kh_{\mathbb{A}}(L)$ of a link $L$ commutes with the Steenrod cohomology operations.
\end{corollary}
We carry out the construction in the setting of framed flow categories, recalled in Section \ref{sec:ffc review}. (An analogous construction likely can be done using a suitably defined ``framed Burnside category''; compare the discussion in \cite[Section 2.5]{LSsurvey}.) The overall strategy is to define a stable homotopy refinement of the cone of the chain map $J$, see Section \ref{sec:constructing a map in general}.
An important problem that immediately presents itself is the existence of {\em cancellations}. That is, within the cube-shaped chain complex ${\rm Cone}(J)$, there exist pairs of consecutive edges so that a single generator is sent to a pair of generators, which then cancel by the second edge differential. This does not happen in the usual Khovanov chain complex, since the coefficients of the differential with respect to the canonical generators are either $0$ or $1$. There are signs but still no cancellations in the setting of odd Khovanov homology, whose stable homotopy refinement was given in \cite{SSS}.
A specific example of cancellations in ${\rm Cone}(J)$ is discussed in Section \ref{sec:cubic illustration}. Its manifestation in the context of framed flow categories is the presence of a pair of oppositely framed points, which cobound a framed interval in the $1$-dimensional moduli space. In general this leads to moduli spaces of dimensions $2$ and higher with non-trivial topology, see Sections \ref{sec:cubic illustration}, \ref{sec:m-1 to m} and Figure \ref{fig:Hexagon times I}.
The construction of the Khovanov homotopy type \cite{LS} and other recent results in this setting involved framed flow categories whose $n$-dimensional moduli spaces are trivial covers of permutohedra, so topologically they are (disjoint unions of) $n$-dimensional balls. In general, if one attempts to build a framed flow category and the moduli spaces have non-trivial topology in some dimension, the existence of higher-dimensional moduli spaces is not assured. Indeed, there are highly restrictive compatibility conditions on how the moduli spaces of various dimensions fit together in a framed flow category (see Section \ref{sec:ffc review}), and there are choices involved in their construction. The union of moduli spaces up to a dimension $n-1$ are required to form the boundary of the moduli spaces of dimension $n$, forming manifolds with corners. Thus extending the construction to dimension $n$ amounts to finding a collection of null-cobordisms, a problem which a priori might not have a solution.
A key point in our definition of a framed flow category for ${\rm Cone}(J)$ in the proof of Theorem \ref{main thm} is that the moduli spaces may be constructed as (trivial covers of) codimension-1 submanifolds of some canonical spaces. The bulk of the paper, Sections \ref{0 dim sec} - \ref{sec:Path moduli spaces}, deals with setting up the base cases: $0$- and $1$-dimensional moduli spaces. The inductive step, using the Pontryagin-Thom construction and propagating the codimension-1 condition, is given in Section \ref{sec:m-1 to m}.
Another important feature in our work, as compared to prior constructions, is the necessity of performing a global combinatorial analysis of configurations of curves and surgery arcs in a link diagram. Recall that the structure of the 1-dimensional moduli spaces used in the construction of the Khovanov homotopy type is analyzed by hand only for 2- and 3-dimensional faces of the Khovanov cube \cite[Sections 5.4,5.5]{LS}. These analyses are local in the sense that no more than three surgery arcs need to be considered. The triviality of the permutohedral covers then ensures that all higher dimensional moduli spaces can be built. In contrast, the more intricate structure of moduli spaces in our setting requires a detailed analysis of configurations of an arbitrary number of surgery arcs; see Section \ref{sec:Path moduli spaces}.
The construction of the map $\mathcal{J}$ in Theorem \ref{main thm} is stated as Corollary \ref{cor:ffcj refines Cone J}; the main technical result underlying its definition is Theorem \ref{thm:ffcJ}. The invariance of the homotopy class of $\mathcal{J}$ with respect to various choices made in the construction, as well as invariance under Reidemeister moves, is established in Section \ref{sec:invariance}.
Recall that the $\sl_2$ action on the annular complex is natural with respect to link cobordisms. However, link cobordism maps on spectra are only defined \cite{LS2} for link cobordisms which are presented as a sequence of Morse moves and Reidemeister moves. These maps are not known to be well-defined in general, although they are conjectured to be so.
\begin{theorem}
\label{thm:J commutes with cobs}
The map $\mathcal{J}$ constructed in Theorem \ref{main thm} is natural with respect to link cobordisms presented as sequences of Morse moves and Reidemeister moves.
\end{theorem}
The results of this paper concern the action of generators of $\sl_2$. The problem of lifting the relations of $\sl_2$ is outside the scope of our present work; we hope to address it in a future paper. Another interesting problem concerns quantum annular homology $Kh_{\mathbb{A}_\mathfrak{q}}(L)$. This theory was defined by Beliakova, Putyra and Wehrli in \cite{BPW}. Extending the work of \cite{GLW}, they showed that $Kh_{\mathbb{A}_\mathfrak{q}}(L)$ admits an action of $U_{\mathfrak{q}}(\sl_2)$. Our earlier work \cite{AKW} defined a stable homotopy refinement of quantum annular homology, which takes the form of a $\mathbb{Z}/r\mathbb{Z}$-equivariant spectrum ${\mathcal X}^r_{\mathbb{A}_\mathfrak{q}}(L)$, where $r\geq 2$. Conjecture 1.4 in \cite{AKW} states that the action of $U_{\mathfrak{q}}(\sl_2)$ on $Kh_{\mathbb{A}_\mathfrak{q}}(L)$ can be lifted to an action on ${\mathcal X}^r_{\mathbb{A}_\mathfrak{q}}(L)$. The methods of the present paper do not immediately extend to the setting of ${\mathcal X}^r_{\mathbb{A}_\mathfrak{q}}(L)$ which was constructed using an equivariant version of the Burnside category. Nevertheless, our results are consistent with that conjecture, and a reformulation in terms of a ``framed Burnside category", briefly discussed above, is likely to lead to a lift of the action of the generators of $U_{\mathfrak{q}}(\sl_2)$.
{\bf Acknowledgements.} We are grateful to Sucharit Sarkar for many very helpful discussions on framed flow categories and the Burnside category. MW is grateful to Matt Stoffregen for several helpful conversations.
RA was supported by NSF RTG Grant DMS-1839968 and the Jefferson Scholars Foundation.
VK was supported in part by the Miller Institute for Basic Research in Science at UC Berkeley, Simons Foundation fellowship 608604, and NSF Grant DMS-1612159. MW was supported in part by NSF FRG Grant DMS-1563615.
\section{Background and conventions}
\subsection{Khovanov homology}\label{sec:Khovanov homology}
We begin with a brief overview of the Bar-Natan category $\mathcal{BN}$ and the construction of the formal chain complex $[[D]]$ assigned to a link diagram $D$; for a complete treatment see \cite{BN1}.
Recall the (dotted) Bar-Natan category $\mathcal{BN}=\mathcal{BN}(\mathbb{R}^2)$. Let $I :=[0,1]$ denote the unit interval. Objects of $\mathcal{BN}$ are formal direct sums of formally graded collections of simple closed curves in the plane $\mathbb{R}^2$. Morphisms are matrices whose entries are formal $\mathbb{Z}$-linear combinations of dotted cobordisms properly embedded in $\mathbb{R}^2 \times I$. Cobordisms are considered up to isotopy relative to the boundary, and they are subject to the local relations shown in Figure \ref{fig:BN relations}.
\begin{figure
\centering
\begin{subfigure}[b]{.3\textwidth}
\begin{center}
\includegraphics{BN_rel1.pdf}
\end{center}
\caption{Sphere}\label{fig:sphere}
\end{subfigure}%
\begin{subfigure}[b]{.4\textwidth}
\begin{center}
\includegraphics{BN_rel2.pdf}
\end{center}
\caption{Neck-cutting}\label{fig:neck-cutting}
\end{subfigure}\\
\vskip1em
\begin{subfigure}[b]{.3\textwidth}
\begin{center}
\includegraphics{BN_rel3.pdf}
\end{center}
\caption{Dotted sphere}\label{fig:dotted sphere}
\end{subfigure}%
\begin{subfigure}[b]{.35\textwidth}
\begin{center}
\includegraphics{BN_rel4.pdf}
\end{center}
\caption{Two dots}\label{fig:two dot}
\end{subfigure}
\caption{Bar-Natan relations}\label{fig:BN relations}
\end{figure}
Let $A= \mathbb{Z}[X]/(X^2)$ denote the Frobenius algebra with comultiplication, unit, and counit described below.
\begin{align*}
\Delta \: A & \to A \o A & \eta \: \mathbb{Z} & \to A & \varepsilon \: A & \to \mathbb{Z} \\
1 & \mapsto X \o 1 + 1\o X & 1 &\mapsto 1 & 1 & \mapsto 0\\
X & \mapsto X\o X & & & X & \mapsto 1
\end{align*}
This is the Frobenius algebra underlying $\sl_2$ link homology \cite{Kh1}. Define the quantum grading $\operatorname{qdeg}$ on $A$ by setting
\begin{equation}\label{eq:1 and X gradings}
\operatorname{qdeg}(X) = -1 \hskip2em \operatorname{qdeg}(1) = 1.
\end{equation}
Let $\mathbb{Z} {\rm -}\operatorname{gmod}$ denote the category of $\mathbb{Z}$-graded abelian groups and graded maps (of any degree) between them.
The Frobenius algebra $A$ defines a $(1+1)$-dimensional TQFT, which descends to a functor
\begin{equation}\label{eq:tqft F}
\mathcal{F}_{Kh} \colon \mathcal{BN} \to \mathbb{Z}{\rm -}\operatorname{gmod}.
\end{equation}
We will refer to $\mathcal{F}_{Kh}$ as the \emph{Khovanov TQFT}. The construction is reviewed below.
On an object $\mathscr{C}$ which is a collection of $n$ disjoint circles in the plane, $\mathcal{F}_{Kh}(\mathscr{C}) = A^{\o n}$, generated by elements of the form
\begin{equation}
\label{eq:generators expanded as labels on each circle}
y=y_1\o \cdots \o y_n\in A^{\o n}
\end{equation}
where each $y_i \in \{1, X\}$; this corresponds to a label of each circle in $\mathscr{C}$ by $1$ or $X$. We refer to such generators as \emph{Khovanov generators}.
Morphisms in $\mathcal{BN}$ are generated by so-called \emph{elementary cobordisms} which contain only a single non-degenerate critical point with respect to the height function $\mathbb{R}^2\times I \rightarrow I$. Index 0 (cup) and 2 (cap) elementary cobordisms are assigned the unit and co-unit maps of $A$ respectively; index 1 (saddle) elementary cobordisms are assigned multiplication or comultiplication maps according to whether the cobordism merged two circles into one, or split one circle into two. The presence of a dot on a cobordism corresponds to multiplication by $X\in A$.
Now let $D$ be a planar diagram for an oriented link $L\subset S^3$. We recall the construction of the formal Bar-Natan chain complex $[[D]]$ from \cite{BN2}. First form the \textit{cube of resolutions}. Label the crossings of the diagram by $1,\ldots, n$. Each crossing may be resolved in one of two ways, called the \textit{0-smoothing} and \textit{1-smoothing}, as in \eqref{fig:0 and 1 smoothings}.
\begin{equation}\label{fig:0 and 1 smoothings}
\begin{aligned}
\includestandalone{two_smoothings}
\end{aligned}
\end{equation}
For each $u = (u_1,\ldots, u_n)\in \{0,1\}^n$, perform the $u_i$-smoothing at the $i$-th crossing. This results in a collection of disjoint simple closed curves in the plane, which we denote $D_u$. Identifying $\{0,1\}^n$ with the vertices of an $n$-dimensional cube, we decorate each vertex $u$ by its corresponding smoothing $D_u$.
Let $u=(u_1,\ldots, u_n)$ and $v=(v_1,\ldots, v_n)$ be vertices which differ only in the $i$-th entry, where $u_i = 0$ and $v_i=1$. The diagrams $D_u$ and $D_v$ are identical outside of a small disk around the $i$-th crossing. There is a cobordism from $D_u$ to $D_v$, which is the obvious saddle near the $i$-th crossing and the identity (product cobordism) elsewhere. We will call this the \emph{saddle cobordism} from $D_u$ to $D_v$, and denote it by $d_{u,v}$. Decorate each edge of the $n$-dimensional cube by the corresponding saddle cobordism. The result is a commutative cube in the category $\mathcal{BN}$. There is an assignment $s_{u,v} \in \{0,1\}$ for each edge so that multiplying the edge map $d_{u,v}$ by $(-1)^{s_{u,v}}$ results in an anti-commutative cube (see \cite[Section 2.7]{BN2}, also \cite[Definition 4.5]{LS}).
For $u=(u_1,\ldots, u_n) \in \{0,1\}^n$, set $\vert u \vert = \sum_i u_i$. Now, form the chain complex $[[D]]$ by setting
\[
[[D]]^i = \bigoplus_{\vert u \vert = i+ n_-} D_{u}\{i + n_+ -n_-\}
\]
where $n_-$, $n_+$ are the number of negative and positive crossings in $D$, and $\{-\}$ denotes the formal grading shift in $\mathcal{BN}$. The differential is defined on each summand by the edge map $(-1)^{s_{u,v}}d_{u,v}$. That $[[D]]$ is a chain complex follows from anti-commutativity of the cube.
\begin{theorem}\emph{(\cite[Theorem 1]{BN2})}
If diagrams $D$ and $D'$ are related by a Reidemeister move, then $[[D]]$ and $[[D']]$ are chain homotopy equivalent.
\end{theorem}
Let
\[
CKh(D) := \mathcal{F}_{Kh}([[D]])
\]
denote the chain complex obtained by applying the Khovanov TQFT to $[[D]]$; it is precisely the invariant defined in \cite[Section 7]{Kh1}.
\subsection{Annular Khovanov homology}\label{sec:annular Khovanov homology}
Asaeda-Przytycki-Sikora \cite{APS} introduced a homology theory for links in $I$-bundles over surfaces. The case of the thickened annulus is known as annular Khovanov homology, sometimes called \emph{sutured} annular Khovanov homology. This section reviews its construction, following \cite{Ro, GLW}.
Let $\mathbb{A} = S^1\times I$ denote the annulus. An \textit{annular link} is a link in the thickened annulus $\mathbb{A}\times I$, and its diagram is obtained by projecting onto the first factor $\mathbb{A}$. Embed $\mathbb{A}$ standardly in $\mathbb{R}^2$ via
\[
\mathbb{A} = \{ x\in \mathbb{R}^2 \mid 1\leq \lr{x} \leq 2\},
\]
so that an annular link diagram and all of its smoothings are drawn in the punctured plane $\mathbb{R}^2\setminus (0,0)$. We represent the annulus by simply indicating the deleted point using the symbol $\times$. See Figure \ref{fig:annular link} for an example of an annular link diagram drawn using this convention. We distinguish two types of circles in $\mathbb{A}$: \emph{trivial} circles, which are contractible in $\mathbb{A}$, and \emph{essential} ones, which are not contractible.
\begin{figure}
\centering
\includegraphics{QAH49.pdf}
\caption{An annular link diagram}\label{fig:annular link}
\end{figure}
Let $\mathcal{BN}(\mathbb{A})$ denote the Bar-Natan category of the annulus. Its objects are formal direct sums of formally $\mathbb{Z}^2$-graded collections of simple closed curves in $\mathbb{A}$. Morphisms are matrices with entries formal $\mathbb{Z}$-linear combinations of dotted cobordisms properly embedded in $\mathbb{A} \times I$, modulo isotopy relative to the boundary, and subject to the local Bar-Natan relations shown in Figure \ref{fig:BN relations}.
Let $\mathbb{Z} {\rm -}\operatorname{ggmod}$ denote the category of $\mathbb{Z}^2$-graded abelian groups and graded maps (of any bidegree) between them. We now describe the annular TQFT
\[
\Fa \colon \mathcal{BN}(\mathbb{A}) \to \mathbb{Z}{\rm -}\operatorname{ggmod},
\]
following the construction in \cite{Ro}.
Let $\mathscr{C} \subset \mathbb{A}$ be a collection of $n$ trivial and $m$ essential circles. Viewing $\mathbb{A}$ as a subspace of $\mathbb{R}^2$, apply the Khovanov TQFT $\mathcal{F}_{Kh}$ from Section \ref{sec:Khovanov homology},
\[
\mathcal{F}_{Kh}(\mathscr{C}) = A^{\o n} \o A^{\o m}.
\]
Introduce a second grading, called the \emph{annular} grading and denoted $\operatorname{adeg}$, on $\mathcal{F}_{Kh}(\mathscr{C})$ as follows. Every tensor factor corresponding to a trivial circle is concentrated in annular degree zero. For a factor $A$ associated to an essential circle, introduce the notation
\[
v_- = X,\hskip2em v_+ = 1
\]
to denote a basis for this copy of $A$, and define $\operatorname{adeg}$ by setting
\[
\operatorname{adeg}(v_-) = -1 \hskip2em \operatorname{adeg}(v_+) = 1.
\]
The underlying abelian group of $\Fa(\mathscr{C})$ is defined to be $\mathcal{F}_{Kh}(\mathscr{C})$, with the bigrading given by $(\operatorname{qdeg}, \operatorname{adeg})$. To define $\Fa$ on a cobordism $S\subset \mathbb{A}\times I$, first view $S$ as a surface in $\mathbb{R}^2 \times I$ and consider the map $\mathcal{F}_{Kh}(S)$ assigned to $S$ by the Khovanov TQFT. It was observed in \cite[Section 2]{Ro} that $\mathcal{F}_{Kh}(S)$ is of the form
\begin{equation}\label{eq:non alpha cob split}
\mathcal{F}_{Kh}(S) = \mathcal{F}_{Kh}(S)_0 + \mathcal{F}_{Kh}(S)_-
\end{equation}
where $\mathcal{F}_{Kh}(S)_0$ preserves $\operatorname{adeg}$ and $\mathcal{F}_{Kh}(S)_-$ lowers $\operatorname{adeg}$. Set
\[
\Fa(S) := \mathcal{F}_{Kh}(S)_0
\]
to be the part of $\mathcal{F}_{Kh}(S)$ that preserves $\operatorname{adeg}$. It follows from \eqref{eq:non alpha cob split} that $\Fa$ is functorial with respect to composition of cobordisms, and it clearly factors through the relations in Figure \ref{fig:BN relations}. We will refer to $\Fa$ as the \emph{annular TQFT}. Note that if $S$ carries $d$ dots, then $\Fa(S)$ is a map of $(\operatorname{qdeg}, \operatorname{adeg})$ bidegree
\[
(\chi(S) - 2d, 0).
\]
We distinguish the bigraded modules assigned to trivial and essential circles by writing
\[
V = \Fa(C)
\]
if $C$ is an essential circle, with basis written as $\{v_-, v_+\}$. The notation $A$ is reserved for the module assigned to a trivial circle, with basis $\{1, X\}$. Then if $\mathscr{C} \subset \mathbb{A}$ consists of $n$ trivial and $m$ essential circles, the module assigned to $\mathscr{C}$ by $\Fa$ is written as
\[
\Fa(\mathscr{C}) = A^{\o n} \o V^{\o m}.
\]
Cobordisms between trivial circles are assigned the same map by $\mathcal{F}_{Kh}$ and $\Fa$. We record here the maps assigned to the four elementary saddle cobordisms involving at least one essential circle, shown in Figure \ref{fig:el saddles}, using the $V$ notation.
\noindent\begin{minipage}{.5\linewidth}
\begin{align}\label{eq:formula1}
\begin{aligned}
&V\o A \rar{\hyperref[fig:type1]{\operatorname{(I)}}} V \\
&v_- \o 1 \mapsto v_- \\
&v_+\o 1 \mapsto v_+ \\
&v_- \o X \mapsto 0 \\
&v_+ \o X \mapsto 0
\end{aligned}
\end{align}
\end{minipage}%
\begin{minipage}{.35\linewidth}
\begin{align}\label{eq:formula2}
\begin{aligned}
&V \o V \rar{\hyperref[fig:type2]{\operatorname{(II)}}} A \\
&v_- \o v_- \mapsto 0 \\
&v_+ \o v_- \mapsto X \\
&v_- \o v_+ \mapsto X \\
&v_+ \o v_+ \mapsto 0
\end{aligned}
\end{align}
\end{minipage}
\vskip1em
\noindent\begin{minipage}{.49\linewidth}
\begin{align}\label{eq:formula3}
\begin{aligned}
&V \rar{\hyperref[fig:type3]{\operatorname{(III)}}} V \o A \\
&v_- \mapsto v_- \o X \\
&v_+ \mapsto v_+ \o X
\end{aligned}
\end{align}
\end{minipage}
\begin{minipage}{.4\linewidth}
\begin{align}\label{eq:formula4}
\begin{aligned}
&A \rar{\hyperref[fig:type4]{\operatorname{(IV)}}} V\o V \\
&1 \mapsto v_+ \o v_- + v_- \o v_+ \\
&X \mapsto 0
\end{aligned}
\end{align}
\end{minipage}
\vskip1em
\begin{figure}
\begin{subfigure}[b]{.2\textwidth}
\begin{center}
\includegraphics{AD_type1.pdf}
\end{center}
\caption{Type (I)}\label{fig:type1}
\end{subfigure}
\begin{subfigure}[b]{.2\textwidth}
\begin{center}
\includegraphics{AD_type2b.pdf}
\end{center}
\caption{Type (II)}\label{fig:type2}
\end{subfigure}
\begin{subfigure}[b]{.2\textwidth}
\begin{center}
\includegraphics{AD_type3.pdf}
\end{center}
\caption{Type (III)}\label{fig:type3}
\end{subfigure}
\begin{subfigure}[b]{.2\textwidth}
\begin{center}
\includegraphics{AD_type4b.pdf}
\end{center}
\caption{Type (IV)}\label{fig:type4}
\end{subfigure}
\caption{Saddles involving essential circles. The vertical red arc is the central axis of $\mathbb{A} \times I \subset \mathbb{R}^2\times I$.}\label{fig:el saddles}
\end{figure}
Finally, let $D$ be a diagram for an oriented annular link $L$. The construction of $[[D]]$ as described in Section \ref{sec:Khovanov homology} is completely local, and crossings are away from the puncture $\times$. Thus we may view $[[D]]$ as a chain complex over $\mathcal{BN}(\mathbb{A})$, with the grading shifts $\{-\}$ in $\mathcal{BN}$ rewritten as bigrading shifts $\{-,0\}$ in $\mathcal{BN}(\mathbb{A})$. Annular link diagrams representing isotopic annular links are related by Reidemeister moves away from the puncture. Therefore the chain homotopy class of $[[D]]$, viewed as a chain complex over $\mathcal{BN}(\mathbb{A})$, is an invariant of $L$. Applying the annular TQFT $\Fa$, we obtain the annular Khovanov chain complex
\begin{equation}\label{eq:annular chain complex}
CKh_\mathbb{A}(D) := \Fa([[D]]).
\end{equation}
It is an invariant of $L$ up to chain homotopy equivalence.
\subsection{Arc diagrams}
\label{sec:Arc diagrams}
In our resolution diagrams throughout the Khovanov cube for a link diagram $D$, we will use additional decorations to record whether the $0$-smoothing or $1$-smoothing was used at a crossing. These form an enhancement of the \emph{resolution configurations} from \cite[Section 2]{LS}, and they will be crucial to our combinatorial analysis in Section \ref{sec:Path moduli spaces}.
\begin{definition}
\label{def:arc diagram}
Let $D$ be an $n$-crossing link diagram and let $u\in \{0,1\}^n$ be a vertex. The \emph{arc diagram $D_u$} is obtained from the resolution $D_u$ as follows. Join the two strands at each smoothing by one of two types of arcs, a solid red arc, called a \emph{future arc}, if the $0$-resolution was used, or a dashed blue arc, called a \emph{past arc}, if the $1$-resolution was used.
\end{definition}
The local picture is shown in Figure \ref{fig:smoothings with arcs}. To explain the terminology, a $0$-smoothing means that the saddle corresponding to that crossing is yet to be performed, while a $1$-smoothing means that the associated saddle has already been performed. For an arc $a$ in an arc diagram, we can obtain a new arc diagram by performing surgery along $a$ and then inserting the dual surgery arc to $a$ and switching its type, as shown in Figure \ref{fig:surgery along an arc}. We will refer to this process as performing surgery along $a$, with the insertion of the dual arc of opposite type implied. An edge $D_u \to D_v$ in the cube of resolutions of $D$ thus corresponds to performing surgery along a future arc in $D_u$. More generally, given any pair of vertices $u,v\in\{0,1\}^n$, the arc diagram $D_v$ can be obtained from $D_u$ by a sequence of such surgeries along past and future arcs in $D_u$.
\begin{figure}
\hspace*{\fill}
\begin{subfigure}[t]{.4\linewidth}
\begin{center}
\includestandalone{two_smoothings_with_arcs}
\end{center}
\caption{Future ($0$-smoothing) and past ($1$-smoothing) arcs.}
\label{fig:smoothings with arcs}
\end{subfigure}
\hfill
\begin{subfigure}[t]{.4\linewidth}
\begin{center}
\includestandalone{surgery_along_an_arc}
\end{center}
\caption{Surgery along a future arc $a$ gives a new arc diagram with the dual to $a$ drawn as a past arc.}
\label{fig:surgery along an arc}
\end{subfigure}
\hspace*{\fill}
\caption{}
\end{figure}
In what follows, we will depict annular arc diagrams by drawing planar tangles in the rectangle, with the understanding that the annular arc diagram is obtained by connecting the left endpoints to the right. Thus horizontal intervals are segments of distinct essential circles. See Figure \ref{fig:convention for arc diagrams} for an example. We will omit the dashed platforms in the future.
\begin{figure}
\begin{subfigure}[b]{.4\textwidth}
\begin{center}
\includestandalone{arc_diagram_convention_2}
\end{center}
\caption{}\label{fig:arc diagram}
\end{subfigure}
\begin{subfigure}[b]{.4\textwidth}
\begin{center}
\includestandalone{arc_diagram_convention_1}
\end{center}
\caption{}\label{fig:cut open arc diagram}
\end{subfigure}
\caption{A rectangular arc diagram and its annular closure}\label{fig:convention for arc diagrams}
\end{figure}
An arc diagram is \emph{connected} if it is connected as a subspace of the plane. Similarly, we will refer to the connected components of arc diagrams. Note that, since surgery along an arc preserves connectedness, there is a natural correspondence between the connected components of $D_u$ and the connected components of $D_v$ for any two vertices $u,v\in\{0,1\}^n$.
Now let $u=(0,\cdots,0)$ be the starting vertex of the cube, and suppose we have a partition of $D_u$ into disjoint closed (not necessarily connected) components
\[D_u = \mathscr{C}_1 \sqcup \cdots \sqcup \mathscr{C}_k.\]
In a similar fashion, this partition determines such a partition of the arc diagram $D_v$ at every vertex $v\in\{0,1\}^n$. Abusing notation slightly, we will use the same notation for these closed components at every vertex:
\[D_v = \mathscr{C}_1 \sqcup \cdots \sqcup \mathscr{C}_k.\]
Then for any Khovanov generator $x\in\mathcal{F}_{Kh}(D_v)$, we will write
\begin{equation}
\label{eq:restrictions of generators}
x=x|_{\mathscr{C}_1} \o \cdots \o x|_{\mathscr{C}_k} \in \mathcal{F}_{Kh}(D_v)
\end{equation}
where $x \vert_{\mathscr{C}_i}$ denotes the restriction of the labels in $x$ to the circles in $\mathscr{C}_i\subset D_v$. Note that this is again a slight abuse of notation, since the use of the symbol $\otimes$ in Equation \eqref{eq:restrictions of generators} is, strictly speaking, incompatible with the use of the same symbol in Equation \eqref{eq:generators expanded as labels on each circle}, but this should not cause any confusion in our arguments.
In many places throughout the paper, we will be concerned with certain \emph{subcubes} within the large cube $\{0,1\}^n$. A $k$-dimensional subcube, denoted $\subcube{k}{v}{u}$, is determined by a pair of vertices $u,v$ such that $u_i\leq v_i$ for all $i$, with precisely $k$ coordinates $i$ where $u_i=0$ and $v_i=1$. When focusing on such a subcube, we will draw the arc diagram $D_u$ by including only the $k$ future arcs corresponding to these $k$ coordinates, and likewise for any intermediate vertices $u\leq w \leq v$ within the subcube. See Figure \ref{fig:subcube arc diagram} for an example.
\begin{figure}
\hspace*{\fill}
\begin{subfigure}[b]{.4\textwidth}
\begin{center}
\includestandalone{subcube_arc_diagram1}
\end{center}
\caption{The arc diagram at the starting vertex $(0,0,0,0)$}\label{fig:subcube arc diagram1}
\end{subfigure}
\hfill
\begin{subfigure}[b]{.4\textwidth}
\begin{center}
\includestandalone{subcube_arc_diagram2}
\end{center}
\caption{An arc diagram for the subcube $(1,0,0,0) \leq_2 (1,0,1,1)$}\label{fig:subcube arc diagram2}
\end{subfigure}
\hspace*{\fill}
\caption{}\label{fig:subcube arc diagram}
\end{figure}
\subsection{Action of $\sl_2$ on annular Khovanov homology}\label{sec:sl2 action}
It is shown in \cite{GLW} that the annular TQFT $\Fa$ can be viewed as taking values in the category $\operatorname{gRep}(\sl_2)$ of graded $\sl_2$ representations. We first recall some basic definitions. Let $\sl_2$ be the $\mathbb{Z}$-span of the $2 \times 2$ matrices
\[
E =
\begin{pmatrix}
0 & 1 \\
0 & 0
\end{pmatrix},\,
F =
\begin{pmatrix}
0 & 0 \\
1 & 0
\end{pmatrix},\,
H =
\begin{pmatrix}
1 & 0 \\
0 & -1
\end{pmatrix}.
\]
The usual commutator bracket,
\[
[x,y] = xy-yx,
\]
makes $\sl_2$ into a Lie algebra, with the bracket given on generators by
\begin{equation}\label{eq:sl2 relations}
[E,F] = H, \; [H,E] = 2E, \; [H,F] = -2F.
\end{equation}
\begin{remark}
The Lie algebra $\sl_2$ in \cite{GLW} is defined over $\mathbb{C}$. For the purpose of this paper we continue working over $\mathbb{Z}$, since all desired results already hold integrally.
\end{remark}
Recall that if $M$ is an $\sl_2$-module, then its linear dual
\[
M^* = \operatorname{Hom}_\mathbb{Z}(M, \mathbb{Z})
\]
is again an $\sl_2$-module via
\[
(xf)(m) := -f(xm)
\]
for $x\in \sl_2$, $f\in M^*$, and $m\in M$. Likewise, if $N$ is another $\sl_2$-module, then the tensor product $M\o_\mathbb{Z} N$ inherits an $\sl_2$ action by setting
\[
x(m\o n) := xm \o n + m \o xn
\]
for $x\in \sl_2$, $m\in M$, and $n\in N$.
Let $V_1 = \mathbb{Z} \oplus \mathbb{Z}$ denote the fundamental representation of $\sl_2$ with standard basis vectors denoted $\boldsymbol{v}_0= (0,1), \boldsymbol{v}_1 = (1,0)$, and let $V_1^*$ denote its dual, with dual basis vectors $\boldsymbol{v}_0^*, \boldsymbol{v}_1^* \in V_1^*$. The action of $\sl_2$ on $V_1$ and $V_1^*$ is summarized in \eqref{eq:sl2 action}.
\begin{align}\label{eq:sl2 action}
\begin{aligned}
E \boldsymbol{v}_0 &= \boldsymbol{v}_1 & F \boldsymbol{v}_0 &= 0 & H \boldsymbol{v}_0 &= - \boldsymbol{v}_0\\
E \boldsymbol{v}_1 &= 0 & F \boldsymbol{v}_1 &= \boldsymbol{v}_0 & H \boldsymbol{v}_1 &= \boldsymbol{v}_1\\
& & & & & \\
E \boldsymbol{v}_0^* &= 0 & F \boldsymbol{v}_0^* &= -\boldsymbol{v}_1^* & H \boldsymbol{v}_0^* &= \boldsymbol{v}_0^*\\
E \boldsymbol{v}_1^* &= -\boldsymbol{v}_0^* & F \boldsymbol{v}_1^* &= 0 & H \boldsymbol{v}_1^* &= - \boldsymbol{v}_1^*
\end{aligned}
\end{align}
Let $\mathscr{C} \subset \mathbb{A}$ be a collection of $n$ trivial and $m$ essential circles. There is a natural ordering on the essential circles, starting from the innermost circle (closest to the puncture $\times$). Make
\[
\Fa(\mathscr{C}) = A^{\o n} \o V^{\o m}
\]
into an $\sl_2$-module as follows. Every factor $A$ corresponding to a trivial circle is assigned the trivial (zero) two-dimensional representation of $\sl_2$. For a factor $V$ corresponding to the $i$-th essential circle in $\mathscr{C}$, identify $V$ with $V_1$ if $i$ is odd via the $\mathbb{Z}$-linear isomorphism \eqref{eq:odd id} and with $V_1^*$ if $i$ is even via \eqref{eq:even id}.
\begin{align}
V_1 & \to V,\, \boldsymbol{v}_0 \mapsto v_-,\, \boldsymbol{v}_1 \mapsto v_+, \label{eq:odd id} \\
V_1^* & \to V,\, \boldsymbol{v}_0^* \mapsto v_+,\, \boldsymbol{v}_1^* \mapsto v_-. \label{eq:even id}
\end{align}
Thus trivial circles are always assigned the trivial two-dimensional representation, and essential circles are assigned $V_1$ and $V_1^*$ in an alternating manner, with the convention that the innermost essential circle is assigned $V_1$.
\begin{example}\label{ex:action on two essential circles}
The action of $E$ and $F$ on
two essential circles is recorded below.
\begin{align*}
E(v_- \o v_-) &= v_+ \o v_- - v_- \o v_+ & E(v_+\o v_-) &= - v_+ \o v_+ \\
E(v_- \o v_+) &= v_+ \o v_+ & E(v_+ \o v_+) &= 0\\
& & & & & & \\
F(v_- \o v_-) &= 0 & F(v_+\o v_-) &= v_- \o v_- \\
F(v_- \o v_+) &= -v_- \o v_- & F(v_+ \o v_+) &= v_- \o v_+ - v_+ \o v_-
\end{align*}
\end{example}
\begin{remark}
The $\sl_2$ representations $V_1$ and $V_1^*$ are isomorphic via
\[
\boldsymbol{v}_0 \mapsto -\boldsymbol{v}_1^*, \hskip1em \boldsymbol{v}_1 \mapsto \boldsymbol{v}_0^*.
\]
However, the map $V_1 \to V_1^*$ obtained by composing \eqref{eq:odd id} with the inverse of \eqref{eq:even id} is not $\sl_2$-linear; indeed, the identifications \eqref{eq:odd id} and \eqref{eq:even id} equip $V$ with two actions of $\sl_2$ which differ by a sign.
\end{remark}
By \cite[Lemma 4]{GLW}, maps assigned to cobordisms in $\mathbb{A} \times I$ commute with the action of $\sl_2$. Therefore the annular TQFT $\Fa$ is upgraded to a functor
\[
\Fa \colon \mathcal{BN}(\mathbb{A}) \to \operatorname{gRep}(\sl_2)
\]
landing in the category of $\mathbb{Z}$-graded $\sl_2$ representations. The annular grading corresponds to the weight space decomposition, and the quantum grading corresponds to the external grading. It follows that $\sl_2$ acts on the annular chain complex \eqref{eq:annular chain complex} and its homology.
Let $\mathscr{C} \subset \mathbb{A}$ be a collection of $n$ trivial and $m$ essential circles. We will always assume that the essential circles are ordered from innermost to outermost, so that the $i$-th tensor factor of $V$ in
\[
\Fa(\mathscr{C}) = A^{\o n} \o V^{\o m}
\]
corresponds to the $i$-th essential circle. Any standard generator in $V^{\o m}$ can be written as $v_{s_1} \o \cdots \o v_{s_m}$, with each $s_i \in \{\pm \}$. Thus any standard generator $x\in \Fa(\mathscr{C})$ is of the form
\begin{equation}\label{eq:standard generator}
x =y \o v_{s_1} \o \cdots \o v_{s_m}
\end{equation}
where $y\in A^{\o n}$ is a standard generator for the trivial circles in $\mathscr{C}$.
By construction, $\sl_2$ weights correspond to the annular grading, so that for a standard generator $x\in \Fa(\mathscr{C})$, $H$ acts as multiplication by $\operatorname{adeg}(x)$,
\begin{equation}\label{eq:H map}
Hx = \operatorname{adeg}(x) x.
\end{equation}
Although the formula \eqref{eq:H map} is straightforward, we can refine it by writing
\begin{equation}\label{eq:H map refined}
Hx = \sum_{i=1}^m s_i y \o v_{s_1} \o \cdots \o v_{s_m}
\end{equation}
using the notation \eqref{eq:standard generator}. The action of $E$ and $F$ can similarly be written as
\begin{align}\label{eq:E and F explicit}
\begin{aligned}
Ex &= \sum_{i=1}^m y \o v_{s_1} \o \cdots\o E v_{s_i} \o \cdots \o v_{s_m},
\\ Fx & = \sum_{i=1}^m y \o v_{s_1} \o \cdots\o F v_{s_i} \o \cdots \o v_{s_m}.
\end{aligned}
\end{align}
In the formulas \eqref{eq:E and F explicit}, $E v_{s_i}$ and $F v_{s_i}$ are given by
\[
E v_{s_i} =
\begin{cases}
(-1)^{i+1} v_+ & \text{ if } s_i = - \\
0 & \text{ if } s_i = +
\end{cases}
\hskip2em
F v_{s_i} =
\begin{cases}
(-1)^{i+1} v_- & \text{ if } s_i = + \\
0 & \text{ if } s_i = -
\end{cases}
\]
Letting $C$ denote the $i$-th essential circle in $\mathscr{C}$, will say that $E$ \emph{acts on $C$} to mean that we apply $E$ on the $i$-th factor of $V$ in $\Fa(\mathscr{C})$,
\[
x \mapsto y \o v_{s_1} \o \cdots\o E v_{s_i} \o \cdots \o v_{s_m},
\]
and likewise for $F$. If $J$ denotes one of $E$, $F$, or $H$, we will also say the \emph{$J$ map} to mean the endomorphism of the module assigned to a collection of circles
\[
J\: \Fa(\mathscr{C}) \to \Fa(\mathscr{C}), \; x\mapsto Jx,
\]
or on the annular chain complex
\[
J \: CKh_\mathbb{A}(D) \to CKh_\mathbb{A}(D),
\]
which is built out of the $J$ map on each smoothing.
The following is a combination of \cite[Lemma 2]{GLW} and \cite[Lemma 5]{GLW}.
\begin{lemma}\label{lem:E and F duality}
Let $D$ be an annular link diagram. For each smoothing $D_u$, let $\Theta_u \: \Fa(D_u) \to \Fa(D_u)$ be the involution given by $v_{\pm} \mapsto v_{\mp}$ on essential circles and the identity map on the trivial circles. The maps $\Theta_u$ assemble into an isomorphism
\[
\Theta \: CKh_\mathbb{A}(D) \rar{\sim} CKh_\mathbb{A}(D)
\]
of chain complexes of abelian groups, which fits into the commutative diagram
\[
\begin{tikzcd}
CKh_\mathbb{A}(D) \ar[d, "E"'] \ar[r, "\Theta"] & CKh_\mathbb{A}(D) \ar[d, "F"] \\
CKh_\mathbb{A}(D) \ar[r, "\Theta"] & CKh_\mathbb{A}(D).
\end{tikzcd}
\]
\end{lemma}
Therefore the action of $F$ is related to $E$ simply by $F = \Theta E \Theta$. Moreover, $\Theta$ sends generators to generators bijectively, so in our combinatorial analysis we can focus on one of $E$ or $F$. Note however that $\Theta$ does not preserve $(\operatorname{qdeg}, \operatorname{adeg})$-bidegree. A modified quantum grading is used in \cite{GLW}, there denoted as $j'$. In our notation, $j' = \operatorname{qdeg} - \operatorname{adeg}$.
\section{Khovanov spectra}
\subsection{Using framed flow categories to construct Khovanov spectra}
\label{sec:ffc review}
There is a general method due to Cohen-Jones-Segal \cite{CJS} for constructing a spectrum $\X$ with cellular cochain complex $C^*(\X)$ isomorphic to a given cochain complex $C^*$. The essential idea is to build a cell complex whose cells correspond to the generators of $C^*$ and whose attaching maps correspond to the differential of $C^*$. Of course, the various compositions of attaching maps must be homotopically coherent in order for this to make sense. The data needed to guarantee this coherence is organized in a \emph{framed flow category}, which we will review here. Most of this material is based upon the corresponding material in \cite{LS,LLS} where the authors successfully constructed such a category (and thus a spectrum) lifting the Khovanov chain complex of a link in $S^3$.
\begin{definition}
\label{def:flow cat}
(\cite[Definition 3.12]{LS})
A \emph{flow category} $\ffc$ is a category consisting of the following data.
\begin{itemize}
\item A finite set of \emph{objects}, also denoted $\ffc$, together with an indexing map $h:\ffc\rightarrow\mathbb{Z}$. In analogy with Morse theory, an object $x\in\ffc$ is thought of as a critical point of index $h(x)$ for some Morse function.
\item For each $x\neq y\in\ffc$, a \emph{morphism space} (usually called a \emph{moduli space}) $\ModSp(y,x)$ which is a $(h(y)-h(x)-1)$-dimensional manifold with corners (see \cite[Section 3.1]{LS})); negative dimensional manifolds are all empty. In analogy with Morse theory, $\ModSp(y,x)$ is often thought of as the moduli space of flow lines from $y$ down to $x$. Furthermore, the codimension-$m$ boundary $\de_{[m]}\ModSp(y,x)$ of any such moduli space can be identified with the disjoint union
\[ \de_{[m]}\ModSp(y,x) \cong \coprod_{(z_1,\cdots,z_m)\in\ffc^m} \ModSp(y,z_m) \times \ModSp(z_m,z_{m-1}) \times \cdots \times \ModSp(z_1,x).\]
In analogy with Morse theory, $\de_{[m]}\ModSp(y,x)$ is thought of as the $m$-times broken flow lines from $y$ down to $x$.
\end{itemize}
\end{definition}
\begin{remark}
In \cite{LS,LLS}, the authors use a notion of an \emph{$\langle n \rangle$-manifold}, which is a manifold with corners $X$ whose boundary $\de X$ is decomposed in a particular manner into $n$ pieces, $\de X = \de_1 X \cup \cdots \cup \de_n X$. We use square brackets in $\de_{[m]} X$ to denote the codimension $m$ boundary of a manifold with corners in order to distinguish from the earlier notation.
\end{remark}
\begin{definition}
\label{def:framed flow cat}
(\cite[Definition 3.20]{LS})
A \emph{framed flow category} is a flow category $\ffc$ whose moduli spaces come equipped with \emph{framed neat embeddings} into cornered Euclidean space (see \cite[Definition 3.7]{LLS}) in a manner consistent with the boundary identifications indicated in Definition \ref{def:flow cat}.
\end{definition}
\begin{definition}
\label{def:totalization and refinement}
A framed flow category $\ffc$ determines a cochain complex $\operatorname{Tot}^*(\ffc)$, called the \emph{totalization complex} of $\ffc$,
defined as follows.
\begin{itemize}
\item The generators of $\operatorname{Tot}^*(\ffc)$ are in bijection with the objects of $\ffc$.
\item For any $x\in\ffc$, $h(x)$ gives the homological degree of the corresponding generator $x\in \operatorname{Tot}^*(\ffc)$.
\item For any $x\in\operatorname{Tot}^*(\ffc)$, we have
\[dx=\sum_{\{y\in\ffc \mid h(y)=h(x)+1\}} \vert \ModSp(y,x) \vert y\]
where $|\ModSp(y,x)|$ is the signed count of the framed points comprising the 0-dimensional manifold $\ModSp(y,x)$.
\end{itemize}
In analogy with Morse theory, $\operatorname{Tot}^*(\ffc)$ can be viewed as the Morse cochain complex associated to the corresponding Morse function. Then we say $\ffc$ \emph{refines} a complex $C^*$ if $\operatorname{Tot}^*(\ffc)\cong C^*$.
\end{definition}
\begin{remark}
Note that the totalization notation in \cite{LLS} is assigned to Burnside functors, while the term \emph{associated cochain complex} is used in \cite{LS} to describe the complex we are calling the totalization in Definition \ref{def:totalization and refinement}
\end{remark}
\begin{theorem}\emph{(\cite[Lemmas 3.24 -- 3.26]{LS})}
\label{thm:framed flow cat gives spectrum}
A framed flow category $\ffc$ gives rise to a cell complex $|\ffc|$ whose cellular cochain complex is isomorphic to $\operatorname{Tot}^*(\ffc)$ up to an overall homological shift. The stable homotopy type of $|\ffc|$ is an invariant of the given framed flow category $\ffc$.
\end{theorem}
The definitions above are considered in much greater detail in \cite{LS, LLS}, together with the process by which one can build the cell complex $|\ffc|$. In short, each object $x\in\ffc$ corresponds to a cell of dimension $h(x)+k$ for some fixed offset $k$. The boundary of such a cell is made up of cornered Euclidean spaces, and the framed embeddings of the moduli spaces into such cornered Euclidean spaces provide homotopically coherent attaching instructions.
We will not need to delve into any greater detail here, because (following \cite{LS, LLS}) our framed flow category will actually be built upon an already existing one.
\begin{definition}
\label{def:cube flow cat}
For any fixed $n$, the \emph{cube flow category} of dimension $n$ is $\ffc_\mathcal{Q}$ whose objects are the vertices $u$ of an $n$-dimensional cube, viewed as elements of $\{0,1\}^n$, and whose non-empty moduli spaces $\cMod$ are permutohedra. The grading $h\: \ffc \to \mathbb{Z}$ is given by $h(u) = \sum_i u_i$. For more precise details, see \cite[Definition 4.1]{LS} where $\ffc_\mathcal{Q}$ is also described via a specific Morse function on $\mathbb{R}^n$, or \cite[Section 3.4]{LLS} for a combinatorial description.
\end{definition}
\begin{remark}
In \cite{LS}, the notation $\mathscr{C}$ is used for flow categories. We choose to use $\ffc$ in order to reserve $\mathscr{C}$ for collections of circles in a diagram.
\end{remark}
\begin{proposition}\emph{(\cite[Section 4]{LS}, \cite[Section 3.6]{LLS})}
\label{prop:cube flow is framed}
The cube flow category $\ffc_\mathcal{Q}$ is a framed flow category. In particular, its moduli spaces can be neatly embedded (with framing) into cornered Euclidean spaces in a coherent fashion.
\end{proposition}
The proof of Proposition \ref{prop:cube flow is framed} does involve a choice of \emph{sign assignment} for the edges of the cube which correspond to a choice of signs making a cube-shaped chain complex anti-commute so that $d^2=0$ (see \cite[Proposition 4.12]{LS}). However, different choices of sign assignments give rise to stably equivalent spectra, in a lift of the corresponding argument for chain complexes \cite[Proposition 6.1]{LS}.
In this framework, \cite{LS} show how to refine the Khovanov chain complex of a link $L\subset S^3$ with diagram $D$ into a flow category $\ffc_{Kh}(D)$ that ``trivially covers" $\ffc_\mathcal{Q}$ in the following sense.
\begin{definition}
\label{def:cubical flow cat}
(\cite[Definition 3.21]{LLS})
A \emph{cubical flow category} is a flow category $\ffc$ which comes equipped with a graded functor $f:\ffc\rightarrow\ffc_\mathcal{Q}$ to some fixed cube category such that for all $x,y\in\ffc$, the map $\ModSp_\ffc (y,x) \rightarrow \ModSp_\mathcal{Q} (f(y),f(x))$ is a trivial covering map.
\end{definition}
\begin{proposition}
\label{prop:cubical flow is framed}
Any cubical flow category $\ffc$ can be upgraded to a framed flow category.
\end{proposition}
\begin{proof}
The covering maps $\ModSp_\ffc \rightarrow \ModSp_\mathcal{Q}$ can be composed with the embeddings of $\ModSp_\mathcal{Q}$ into cornered Euclidean space, and all of the framings are then inherited. These framings also provide a consistent manner in which to separate the components of the trivial covers. In \cite{LLS} such embeddings are called \emph{cubical neat embeddings}; see Section 3.6 in that reference for more details.
\end{proof}
\begin{remark}
Alternatively, one may follow the arguments in Sections 3.1 and 3.2 of \cite{LS} to upgrade a cubical flow category into a framed flow category. Our compositions of covering maps and embeddings give rise to \emph{neat immersions} with coherent framings in the language employed in that reference, which can be perturbed to give \emph{neat embeddings}. See also Section 3.4.1 in that reference.
\end{remark}
\begin{theorem}
\label{thm:Khovanov flow covers cube flow}
(\cite{LS,LLS})
Given a link $L\subset S^3$ with diagram $D$, there is a cubical flow category $\ffc_{Kh}(D)$ which refines the Khovanov complex $CKh^*(D)$. The resulting suspension spectrum $\X_{Kh}(D)= \Sigma^{\infty} |\ffc_{Kh}(D)|$ is an invariant of the link $L$ up to stable homotopy equivalence, allowing the notation $\X_{Kh}(L)$ for the spectrum.
\end{theorem}
One crucial feature of the Khovanov chain complex used in the proof of Theorem \ref{thm:Khovanov flow covers cube flow} is that, before sign assignment, all signs in the differential are positive. Thus the commutation of any square face in the Khovanov cube corresponds to a \emph{bijection} of 0-dimensional moduli spaces coming from the edges of the square. Such bijections correspond to trivial covers of the 1-dimensional permutohedron (which is an interval), allowing $\ffc_{Kh}$ to satisfy the requirements of Definition \ref{def:cubical flow cat} for the 1-dimensional moduli spaces. In our situation however, this feature will not be present.
\begin{remark}
\label{rmk:Odd Khovanov as signed cover}
It should be noted that the odd Khovanov differential used to define the odd Khovanov complex $CKh_o^*(D)$ in \cite{ORS} does include signs within the cube before the sign assignment. In \cite{SSS}, the authors used the language of signed correspondences in a suitable Burnside category to resolve this issue. In terms of our language here, $\ffc_{Kh^o}$ could be thought of as a ``signed" cubical flow category, where the various covering maps carry signs which indicate the need to reverse framing on certain components of the given cover. Despite this need for signs, it is shown in \cite{SSS} that the square faces in the odd Khovanov cube still correspond to (signed) bijections, which give (signed) trivial covers of intervals as needed.
\end{remark}
The rest of the arguments in \cite{LS,LLS} used to construct $\ffc_{Kh}$ are less immediately relevant, so we simply summarize. The moduli spaces are built inductively. The 1-dimensional moduli spaces in $\ffc_{Kh}$ corresponding to square faces involve a choice of bijection between the compositions of $0$-dimensional moduli spaces coming from the two ways to traverse the edges in a square. For all but one case, these two compositions are either both empty or both consist of one point, so the bijection is determined. The remaining nontrivial case is known as the \emph{ladybug configuration} \cite[Figure 5.1]{LS}, where the two $0$-dimensional moduli spaces each have two points. In this situation, the choice is the \emph{ladybug matching} (\cite[Section 5.4]{LS}). The 2-dimensional $\ModSp_{Kh}$ are then bounded by some cover of $\de_{[1]}\ModSp_\mathcal{Q}$ (a hexagon, topologically a circle $S^1$). The hexagon relation (\cite[Section 5.5]{LS}) is the check that this cover is indeed the trivial cover, so that the 2-dimensional $\ModSp_{Kh}$ can be chosen to be a trivial cover of $\ModSp_\mathcal{Q}$. The higher dimensional $\de\ModSp_\mathcal{Q}$ are all topologically spheres of higher dimension, and so have only trivial covers, allowing the construction of all higher dimensional $\ModSp_{Kh}$ from here.
If we instead let $L\subset \mathbb{A} \times I$ be an annular link with diagram $D$, then the methods in \cite{LS, LLS} extend in a straightforward way to build a stable homotopy refinement of annular Khovanov homology. A formulation using the language of Burnside functors is given in \cite[Section 4.3]{AKW}. The correspondence between Burnside functors and cubical flow categories \cite[Section 4.3]{LLS} yields a flow category $\ffc_\mathbb{A}(D)$ refining $CKh_\mathbb{A}(D)$. The associated suspension spectrum
\begin{equation}
\label{eq:annular spectrum}
\X_\mathbb{A}(D) = \Sigma^\infty|\ffc_\mathbb{A}(D)|
\end{equation}
is an invariant of $L$ up to stable homotopy equivalence. An alternative construction can be found in \cite{LLSCK}.
\begin{remark}
In \cite{LS}, there is a global choice between two types of ladybug matchings, the so-called \emph{left pair} and \emph{right pair}. By \cite[Proposition 6.5]{LS}, the two spectra associated to these choices are stably equivalent. To the authors' knowledge, it is not known if the annular spectrum is independent of the choice of ladybug matching; in particular the argument in \cite[Proposition 6.5]{LS} does not hold for links in $\mathbb{A} \times I$. For the remainder of this paper we fix one of the two choices.
\end{remark}
\subsection{Constructing a map between Khovanov spectra}
\label{sec:constructing a map in general}
Given an annular link diagram $D$, each of the generators $E, F$, and $H \in \sl_2$ yield chain maps $CKh_\mathbb{A}(D) \to CKh_\mathbb{A}(D)$. The goal of this paper is to lift these endomorphisms to maps of spectra. This section discusses our general strategy for lifting maps of cochain complexes to maps of spectra.
\begin{definition}
\label{def:grading shift ffc}
Let $\ffc$ be a flow category with grading $h \: \ffc \to \mathbb{Z}$. The \emph{suspension of $\ffc$}, denoted $\Sigma \ffc$, is the flow category whose underlying category is identical to $\ffc$ but whose grading $\Sigma h$ is shifted up by one,
\[
\Sigma h (x) = h(x) +1.
\]
\end{definition}
It is straightforward to verify that if $\ffc$ is a framed flow category, then $\Sigma \ffc$ inherits a natural framing, and moreover there is a homeomorphism
\[
\vert \Sigma \ffc \vert \cong \Sigma \vert \ffc \vert.
\]
Consider a map $A^*\xrightarrow{f} B^*$ of cochain complexes, where $A^*$ and $B^*$ are refined by framed flow categories $\ffc(A)$ and $\ffc(B)$ respectively. In order to lift $f$ to a map of cell complexes $|\ffc(B)|\xrightarrow{F}|\ffc(A)|$ satisfying $F^* = f$ on cohomology (up to a shift), the strategy is to refine the complex $\mathrm{Cone}~f$ into a framed flow category $\ffc(\mathrm{Cone}~f)$. The key point is that, by construction, $\ffc(\mathrm{Cone}~f)$ will contain complementary downward and upward closed subcategories (\cite[Section 3.4]{LS}) corresponding to $\ffc(A)$ and $\Sigma \ffc(B)$ respectively. The downward closed subcategory yields a subcomplex $\vert \ffc(A) \vert \subset \vert \ffc(\mathrm{Cone}~f) \vert$ whose quotient complex is canonically identified with $\vert \Sigma \ffc(B) \vert = \Sigma\vert \ffc(B)\vert$. The Puppe sequence yields
\[
\vert \ffc(A)\vert \hookrightarrow \vert \ffc(\mathrm{Cone}~f)\vert \twoheadrightarrow \vert \ffc(\mathrm{Cone}~f)\vert / \vert \ffc(A)\vert = \Sigma \vert \ffc(B) \vert \xrightarrow{\star} \Sigma \vert \ffc(A) \vert \to \cdots
\]
where $\star$ is the Puppe map. After shifting degrees, the Puppe map is the desired $F$.
Now we note that, if $A^*=B^*=CKh^*(D)$ for a link diagram $D$ having $n$ crossings and $f$ is assembled from maps $\Fa(D_u) \to \Fa(D_u)$ on each smoothing, then the complex $\mathrm{Cone}~f$ takes the shape of an $(n+1)$-dimensional cube $\mathcal{Q}$. If in addition the map $f$ contains no signs, then one can seek to build a cubical flow category $\ffc_{Kh}(\mathrm{Cone}~f)$ covering $\ffc_{\mathcal{Q}}$ as before. Assuming this can be done, the resulting Puppe map $F$ induces a map of spectra $\X(D)\xrightarrow{\Sigma^\infty F} \X(D)$ which lifts $f$ as desired. All of this is illustrated by the exact triangle for a crossing (\cite[Theorem 2]{LS}) which is also used in \cite[Section 3.3]{LS2} to define the map induced by a saddle cobordism between links.
In our situation, $D$ is an annular link and $f$ is one of $E, F$, or $H$. Moreover, as illustrated in Section \ref{sec:cubic illustration}, these maps contain signs leading to cancellations, which prevents $\ffc(\mathrm{Cone}~f)$ from being made into a cubical flow category. Nevertheless, we will refine $\mathrm{Cone}~f$ and lift $E, F$, and $H$ by using the above procedure.
\begin{remark}
\label{rmk:old Burnside version}
All of the constructions thus far can be converted into the language of functors from cube categories to Burnside categories. This is the framework introduced in \cite{LLS} to investigate disjoint unions and connect sums of links, and it provides a convenient combinatorial approach to constructing stable homotopy types, as in \cite{SSS} for the odd Khovanov spectrum. In this language, chain maps are lifted as natural transformations of Burnside functors. In a previous paper on annular Khovanov spectra \cite{AKW}, we used (an equivariant version of) this approach to refine the quantum annular Khovanov homology constructed in \cite{BPW} and the action coming from the generators $K,K^{-1} \in U_q(\sl_2)$ on quantum annular Khovanov homology. However, as we shall see, this viewpoint is not sufficient to refine the action of the remaining generators of $U_q(\sl_2).$ See \cite[Section 8]{AKW} for a related discussion.
\end{remark}
\section{New features: cancellations and topology of moduli spaces}
\label{sec:cubic illustration}
This section considers a particular example to illustrate new complexities that appear in the analysis of the $\sl_2$ action. These features motivate the introduction of new techniques in follow-up sections.
Consider the $2$-crossing annular knot diagram and the associated surgery arc diagram (see Section \ref{sec:Arc diagrams}) in Figure \ref{fig:example}. The cube of resolutions for this link diagram (to be more precise, a square since there are two crossings) is given in Figure \ref{fig:example square}; the coordinate directions are labeled in accordance with the numbering of the surgery arcs in Figure \ref{fig: example arc diagram}.
\begin{figure}[ht]
\hspace*{\fill}
\begin{subfigure}[b]{.4\textwidth}
\begin{center}
\includestandalone{cubic_knot_diagram}
\end{center}
\caption{An annular link diagram} \label{fig:example diagram}
\end{subfigure}
\hfill
\begin{subfigure}[b]{.4\textwidth}
\begin{center}
\includestandalone{cubic_surgery_diagram}
\end{center}
\caption{The corresponding arc diagram} \label{fig: example arc diagram}
\end{subfigure}
\hspace*{\fill}
\caption{}\label{fig:example}
\end{figure}
\begin{figure}
\begin{center}{\includestandalone{cubic_cube_of_resolutions}}
\end{center}
\caption{The cube of resolutions for the link in Figure \ref{fig:example diagram}}\label{fig:example square}
\end{figure}
The cube-shaped chain complex ${\rm Cone}(E)$ is shown in Figure \ref{fig:example cube}. This cube consists of two copies (pictured horizontally) of the cube of resolutions from Figure \ref{fig:example square}; the action of $E$ takes place along the vertical axis. Each diagram in the cube is endowed with past/future arcs as in Definition \ref{def:arc diagram}.
\begin{figure}
\begin{center}{\includestandalone{cubic_cube_of_resolutions_prime}}
\end{center}
\caption{The cube-shaped complex associated to $E$}\label{fig:example cube}
\end{figure}
Consider a particular generator, $1\otimes v_-$, at the upper left vertex of the cube. We use the computations from Section \ref{sec:sl2 action} to arrive at the following diagram which summarizes the cube edge maps (we suppress the tensor product notation for the $v$ factors, for brevity).
\begin{equation} \label{cubic cube}
\begin{tikzcd}[row sep=scriptsize,column sep=scriptsize]
& v_+v_-v_-+ v_-v_+v_- \arrow[rr, "1"] \arrow[dd, "E", pos=0.2] & & X\otimes v_- \arrow[dd, "E"] \\
1\otimes v_- \arrow[rr, crossing over, "1", pos=0.7] \arrow[dd, "E"] \arrow[ur, "2"] & & v_- \arrow[ur, "2"] \\
& v_+v_-v_++v_-v_+v_+ \arrow[rr, "1", pos=0.8] & & X \otimes v_+ \\
1\otimes v_+ \arrow[rr, "1"] \arrow[ur, "2"] & & v_+ \arrow[ur, "2"] \arrow[uu, crossing over, leftarrow, "E"', pos=0.8]\\
\end{tikzcd}
\end{equation}
A crucial feature in this example is {\em cancellation}. It is not seen in the result of the algebraic calculation in the cubical diagram above, and we will explain it now in more detail. The generator $1\otimes v_-$ is sent by the edge map labeled $2$ to $v_+v_-v_-+ v_-v_+v_-$. The value of the $E$ map on these two summands is given as follows:
\begin{equation} \label{1st eq} E(v_+v_-v_-)= v_+v_-v_+ -v_+v_+v_-
\end{equation}
\begin{equation} \label{2nd eq} E(v_-v_+v_-)= v_-v_+v_+ + v_+v_+v_-
\end{equation}
Considering the composition $E\circ 2$, the terms $-v_+v_+v_-$, $v_+v_+v_-$ in equations (\ref{1st eq}), (\ref{2nd eq}) cancel; this is an instance of cancellation that is the root of the complexity of this problem as we explain next.
\begin{figure}[ht]
\begin{center}{\includestandalone{cubic_2_morphism}}
\end{center}
\caption{}\label{fig:example 2 morphism}
\end{figure}
A relevant part of the moduli spaces is illustrated in Figure \ref{fig:example 2 morphism}. The vertical dashed lines are labeled by compositions of edges of length three in the cube, and the corresponding $1$-morphisms are indicated by thickened dots. For example, the effect of the edge path $E21$ on the generator $1\otimes v_-$ is
$$ 1\otimes v_-\overset{1}\longmapsto v_-\overset{2}\longmapsto X\otimes v_-\overset{E}\longmapsto X\otimes v_+$$
The resulting single flow line is represented by a dot above the label {$E21$} in Figure \ref{fig:example 2 morphism}. There is a single flow line associated to the edge path $E12$ as well, since the summand $v_+v_-v_-$ is sent to zero by the edge map labeled 1. The interesting case in this example is the edge path $1E2$: there are three flow lines
$$ 1\otimes v_-\longmapsto v_+v_-v_- \longmapsto v_+v_-v_+\longmapsto X\otimes v_+$$
$$ 1\otimes v_-\longmapsto v_+v_-v_-\longmapsto -v_+v_+v_-\longmapsto - X\otimes v_+$$
$$ 1\otimes v_-\longmapsto v_-v_+v_-\longmapsto v_+v_+v_-\longmapsto X\otimes v_+$$
represented by the three dots above the label $1E2$. (Note that $v_-v_+v_+$ is sent to zero by the map $1$, so this term does not contribute a flow line.) The rest of the edge paths give rise to a single flow line.
The vertical strips between dashed lines in Figure \ref{fig:example 2 morphism} represent $2$-morphisms, or in other words $1$-dimensional moduli spaces corresponding to square faces in the cube ${\rm Cone}(E)$.
Restricting to the square face in the back of the cube (\ref{cubic cube}), the flow line
\[v_-v_+v_-\longmapsto X\otimes v_-\longmapsto X\otimes v_+\]
along the edge path $E1$ corresponds to the flow line
\[v_-v_+v_-\longmapsto v_+v_+v_-\longmapsto X\otimes v_+\]
along $1E$. This explains the matching of the flow lines in the $2$-morphism $E12-1E2$, i.e. the interval that goes across that vertical strip in Figure \ref{fig:example 2 morphism}.
On the other hand, $v_+v_-v_-$ is sent to zero by the map $1$, and thus it does not contribute to a flow line along $E1$. The same generator $v_+v_-v_-$ is sent by the map $E$ to a pair of generators in equation (\ref{1st eq}) which then cancel under the map $1$. Thus the $2$-morphism $E12-1E2$ has to ``match'' the empty set at $E12$ to two canceling flow lines at $1E2$. It is natural to represent this pair of flow lines by oppositely framed points in the interval labeled by $1E2$, and to connected them by a framed ``turnback'' in the strip $E12-1E2$.
A similar analysis along the square face on the left of the cube forces the matching of the flow lines under the $2$-morphism $1E2-12E$, creating the ``cubic'' shape of the $1$-dimensional moduli space indicated in Figure \ref{fig:example 2 morphism}.
This example shows that bijections across square faces, used in \cite{LS} in the setting of Khovanov homology, are insufficient in the presence of cancellations. On the other hand, $2$-dimensional faces of the cube force the matching of flow lines associated to edge paths by framed $1$-dimensional moduli spaces. As indicated in the introduction, in general building higher-dimensional moduli spaces is a rather non-trivial problem. Our solution in this paper is based on representing the moduli spaces as codimension-$1$ submanifolds, as illustrated in Figure \ref{fig:example 2 morphism}. A subtlety in this approach is that such a codimension-1 position of the flow lines ($0$-dimensional moduli spaces) in the interval has to be prescribed for all edge paths in a given cube-shaped chain complex ${\rm Cone}(E)$. Indeed, these flow lines for the edge paths of arbitrary lengths are a part of the inductive construction of higher-dimensional moduli spaces, based on Definition \ref{def:flow cat}. Moreover, given any two generators at two arbitrary vertices $u,v$ in the cube, the positions in the interval for all edge paths connecting $u,v$ in the cube have to be consistent, in the sense that they can be connected by the prescribed framed $1$-dimensional moduli spaces in codimension-1 without intersections. The analysis that enables a solution to this problem is the content of
Sections \ref{0 dim sec} - \ref{sec:Path moduli spaces}.
The construction of higher-dimensional moduli spaces in Section \ref{sec:m-1 to m}
proceeds by induction. For example, in the basic case considered above, the $2$-dimensional moduli space has to be constructed for the two chosen generators $1\otimes v_-$, $X\otimes v_+$ in the cube (\ref{cubic cube}). Identifying the leftmost and rightmost vertical intervals in Figure \ref{fig:example 2 morphism} gives a boundary condition for the $2$-dimensional moduli space in a hexagon times an interval, see Figure \ref{fig:Hexagon times I}. In general, the higher-dimensional inductive step amounts to finding a null-cobordism in a permutohedron times an interval $I$, as explained in Section \ref{sec:m-1 to m}.
\section{Constructing the framed flow category: The set-up and the base case} \label{0 dim sec}
We begin by fixing an $n$-crossing annular link diagram $D$, and we let $J$ denote one of the maps $E$, $F$, or $H$. As discussed in Section \ref{sec:constructing a map in general}, the chain map $J$ on the annular Khovanov complex $CKh_\mathbb{A}(D)$ gives rise to $\mathrm{Cone}~J$, a complex in the shape of an $(n+1)$-dimensional cube $\mathcal{Q}$. Our goal is to build a framed flow category $\ffcJ(D)$ (we will often omit the diagram $D$ from the notation) that refines $\mathrm{Cone}~J$ with the help of $\ffc_\mathcal{Q}$, the framed flow category for the cube $\mathcal{Q}$.
\subsection{Definitions and the overall strategy}
\label{sec:Definitions and overall strategy}
To begin with we set some notation and basic definitions. The vertices of the cube $\mathcal{Q}$ will use labels $u,v,w$, with \[u=(u',i)\in\{0,1\}^n \times \{0,1\}.\]
Here we are viewing the first $n$ coordinates $u'$ as indicating where in $CKh_\mathbb{A}(D)$ the corresponding vertex lies, while the last coordinate $i$ indicates whether we are in the domain or range of the $J$ map.
\begin{definition}
\label{def:resolution diagram}
Given a vertex $u=(u',i)\in\mathcal{Q}$, the \emph{resolution diagram} $D_u$ is the resolution of our link diagram $D$ at the vertex $u'$ in the $CKh_\mathbb{A}(D)$. We further define $\KG{u}$ to be the set of Khovanov generators for $D_u$; see Equation \eqref{eq:generators expanded as labels on each circle}.
\end{definition}
\begin{definition}
An \emph{ $m$-dimensional sub-cube } is a choice of two vertices
\[\scube\]
such that $u$ and $v$ match in all but $m$ indices where $u$ has entry zero and $v$ has entry one. A $1$-dimensional sub-cube is also called an \emph{edge}. We say that a subcube $\scube$ \emph{involves $J$} if $u=(u',0)$ and $v=(v',1)$ (so that some edges corresponding to the $J$-map are included).
\end{definition}
\begin{definition}
For any sub-cube $\scube$, the \emph{cube moduli space} $\cMod$ is by definition the $(m-1)$-dimensional permutohedron (see Definition \ref{def:cube flow cat}). This is a manifold with corners (for example, if $m=3$ this is a hexagon) that is topologically equivalent to $D^{m-1}$, whose codimension $i$ boundary (for all $i=1,\dots,m-1)$ is identified with $i$-fold ``composition'' moduli spaces
\[\de_{[i]} \cMod \cong \coprod \cubeMod{v}{w_i} \times \cdots \cubeMod{w_1}{u},\]
where the disjoint union is taken over all sequences of vertices $u\leq w_1 \leq \cdots \leq w_i \leq v$. For more details, see \cite[Definition 4.1]{LS} and \cite[Section 3.4]{LLS}.
\end{definition}
\begin{remark}
\label{rmk:direction reversal}
Note the reversal of direction. For vertices $u$ and $v$ with $u\leq_1 v$, there is an edge from $u$ to $v$ in the cube of resolutions. However, moduli spaces are directed from $v$ to $u$.
\end{remark}
As described in Section \ref{sec:ffc review}, the original construction \cite{LS} used trivial coverings of the moduli spaces in the cube flow category (see Definition \ref{def:cubical flow cat}). The example of Section \ref{sec:cubic illustration} shows that this framework alone is insufficient for building our desired $\ffcJ$. Instead, we will need to keep track of signs of the $J$ map with the help of framings. First we give ourselves an extra dimension to allow for meaningful framings to occur.
\begin{definition}
\label{def:cube ambient space}
For any sub-cube $\scube$, the \emph{cube ambient space} $\AmbS$ is defined depending on whether or not the $J$ map was involved in $\scube$.
\begin{equation}\label{eq:amb space}
\AmbS := \begin{cases}
\cMod & \text{if $J$ is not involved}\\
\cMod \times (0,1) & \text{if $J$ is involved}
\end{cases}
\end{equation}
In either case, $\AmbS$ is a manifold with corners whose boundary structure is determined by that of $\cMod$. In our figures throughout the paper, we will draw the $(0,1)$-direction horizontally.
\end{definition}
Next we introduce the types of maps that we will be using in place of the trivial covering maps utilized in \cite{LS} via Definition \ref{def:cubical flow cat}.
\begin{definition}
\label{def:thick embedding}
A \emph{thick embedding} is a framed map $f=e\circ t$ where $t$ is a trivial covering and $e$ is an embedding. We will denote thick embeddings with the symbol
$\thickemb$. If $t$ is the trivial identity cover (i.e. the identity map), we call $f$ a \emph{thin embedding} and denote with the usual symbol $\hookrightarrow$. Thus a thin embedding is just a framed embedding.
\end{definition}
\begin{remark}
Note that according to the definition, different sheets in a thick embedding may have different framings. In codimension $1$, such framings are either identical or are reversed. As we show later in the paper, the former case occurs for $J=E, F$ and the latter case occurs for $J=H$.
\end{remark}
With all of this notation in place, we can state our main construction as a theorem.
\begin{theorem}
\label{thm:ffcJ}
Fix an $n$-crossing annular link diagram $D$ and an $\sl_2$ generator $J=E,F,$ or $H$ giving rise to a chain map on annular Khovanov complexes with mapping cone $\mathrm{Cone}~J$. Let $\mathcal{Q}$ denote the $(n+1)$-dimensional cube, with corresponding framed flow category $\ffc_\mathcal{Q}$. Then there exists a flow category $\ffc_J=\ffc_J(D)$ together with a ``thick embedding functor" $f:\ffcJ\rightarrow \ffc_\mathcal{Q}$ satisfying the following properties.
\begin{enumerate}
\item The objects of $\ffc_J$ are in bijection with the Khovanov generators throughout the complex $\mathrm{Cone}~J$:
\[\mathrm{Ob}(\ffcJ):=\coprod_{u\in\mathcal{Q}} \KG{u},\]
with grading given by homological degree in $\mathrm{Cone}~J$, and graded functor $f:\mathrm{Ob}(\ffcJ)\rightarrow\mathrm{Ob}(\ffc_\mathcal{Q})$ being the obvious covering map of sets
\[f(x)= u \quad \Leftrightarrow \quad x\in \KG{u}.\]
\item For any generators $x\in\KG{u},y\in\KG{v}$ such that $u\not\leq v$, the moduli space $\MspJxy$ is the empty set.
\item
\label{item:ffcj boundary definitions}
Given a sub-cube $\scube$ and a pair of generators $x\in \KG{u}, y\in \KG{v}$, the moduli space $\MspJxy$ is an $(m-1)$-dimensional manifold with corners, whose codimension $i$ boundary is identified with $i$-fold compositions
\[
\de_{[i]}\MspJxy \cong \coprod \MspJ{y}{z_i} \times \cdots \times \MspJ{z_1}{x}
\]
where the disjoint union is taken over all tuples of generators $(z_1,\dots,z_i)\in \KG{w_1}\times \cdots \times \KG{w_i}$ for sequences of vertices $u\leq w_1 \leq \cdots \leq w_i \leq v$ in the subcube $\scube$.
\item
\label{item:ffcj thick embedding condition}
Each moduli space $\MspJxy$ of item \eqref{item:ffcj boundary definitions} comes equipped with a thick embedding
\[f:\MspJxy\thickemb \AmbS\]
such that, for any tuple $(z_1,\dots,z_i)\in \KG{w_1}\times \cdots \times \KG{w_i}$ giving rise to a component of $\de_{[i]}\MspJxy$, the following diagram of thick embeddings (using the product framing as necessary) commutes
\begin{equation}
\begin{aligned}
\label{eq:respecting boundary structure}
\begin{tikzpicture}[x=1.5in,y=-.5in]
\node(A) at (0,0) {$\MspJxy$};
\node(B) at (0,1) {$\de_{[i]} \MspJxy$};
\node(C) at (0,2) {$\MspJ{y}{z_i} \times \cdots \times \MspJ{z_1}{x}$};
\node(D) at (2,0) {$\AmbS$};
\node(E) at (2,1) {$\de_{[i]} \AmbS$};
\node(F) at (2,2) {$\AmbSp{v}{w_i}\times\cdots\times \AmbSp{w_1}{u}$};
\xthickembtikz{A}{D}{f}
\thickembtikz{B}{E}
\xthickembtikz{C}{F}{f\times \cdots \times f}
\draw[right hook->] (C) -- (B);
\draw[right hook->] (B) -- (A);
\draw[right hook->] (F) -- (E);
\draw[right hook->] (E) -- (D);
\end{tikzpicture}
\end{aligned}
\end{equation}
\item
\label{item:differential condition}
In the case $m=1$, the $0$-dimensional moduli spaces correspond to the components of the differential in $\mathrm{Cone}~J$.
\end{enumerate}
\end{theorem}
Item \eqref{item:ffcj boundary definitions} ensures that the moduli space boundaries of $\ffcJ$ are arranged in the same way as those of $\ffc_\mathcal{Q}$, while item \eqref{item:ffcj thick embedding condition} ensures that this arrangement is respected by the thick embeddings as well. Meanwhile item \eqref{item:differential condition} refers to the edges of the cube, where the moduli spaces are all thickly embedded points in either $D^0$ or $D^0\times(0,1)$ corresponding to either the annular Khovanov differential in $CKh_\mathbb{A}(D)$ or the $J$ map, respectively.
Recall the annular Khovanov spectrum $\X_\mathbb{A}(D)$ from Equation \eqref{eq:annular spectrum} and the surrounding discussion.
\begin{corollary}
\label{cor:ffcj refines Cone J}
The flow category $\ffcJ(D)$ defined above can be upgraded to a framed flow category which refines the complex $\mathrm{Cone}~J$, giving rise to a map of spectra $\mathcal{J}:\X_\mathbb{A}(D) \rightarrow \X_\mathbb{A}(D)$ whose induced map $\mathcal{J}^*$ on cohomology is equal to the action of $J$.
\end{corollary}
\begin{proof}
The cube flow category $\ffc_\mathcal{Q}$ is a framed flow category, so each $\cMod$ comes equipped with a framed neat embedding $\cMod\hookrightarrow \cEuc$ where $\cEuc$ denotes a suitable cornered Euclidean space (see \cite[Section 3.6]{LLS}). As in Definition \ref{def:cube ambient space}, we set
\begin{equation}\label{eq:Euc space}
\EucS := \begin{cases}
\cEuc & \text{if $J$ is not involved}\\
\cEuc \times \mathbb{R} & \text{if $J$ is involved}
\end{cases}
\end{equation}
We then have embeddings $\AmbS\hookrightarrow\EucS$, where $(0,1)$ embeds into $\mathbb{R}$ in the standard way when $J$ is involved. We then compose with our thick embeddings provided by Item \eqref{item:ffcj thick embedding condition} to give
\[\MspJxy \thickemb \AmbS \hookrightarrow \EucS.\]
Just as in Proposition \ref{prop:cubical flow is framed} for the cubical flow categories of \cite{LLS}, the framing inherited by this composition provides us with a consistent way to separate the various components of each trivial cover, giving us framed neat embeddings $\MspJxy \hookrightarrow \EucS$ which respect the necessary boundary conditions as required (note that crossing with $\mathbb{R}$ in Equation \eqref{eq:Euc space} does not affect the cornered structure of $\cEuc$). Item \eqref{item:differential condition} then ensures our resulting framed flow category $\ffcJ(D)$ has refined $\mathrm{Cone}~J$. This gives rise to the desired map as described in Section \ref{sec:constructing a map in general}.
\end{proof}
The rest of this section together with Sections \ref{sec:m=2}, \ref{sec:Path moduli spaces}, and \ref{sec:m-1 to m}
are devoted to proving Theorem \ref{thm:ffcJ}. Items \eqref{item:ffcj boundary definitions} and \eqref{item:ffcj thick embedding condition} of the theorem are the most crucial aspects of the construction, and the various thickly embedded moduli spaces will be built inductively. If the sub-cube $u\leq_m v$ does not involve the $J$ map, then it corresponds to a sub-cube in the cube of resolutions of $D$. In this case the thick embedding $f:\MspJxy \thickemb \AmbS=\cMod$ will be a trivial cover built exactly as in \cite{LS} (based on the annular differential rather than the traditional Khovanov differential). If $u\leq_m v$ involves the $J$ map, then $\MspJxy$ will be thickly embedded as a codimension $1$ submanifold of $\AmbS$. We handle the base case below.
\subsection{0-dimensional moduli spaces for edges}
\label{sec:0-dim moduli spaces}
Fix $u\leq_1 v$, $x\in \KG{u}$, and $y\in \KG{v}$. We begin by defining the $0$-dimensional moduli space $\MspJxy$.
Suppose first that $u\leq_1 v$ is not the $J$ edge. Then it corresponds to an edge $u\to v$ in the cube of resolutions of $D$. Let
\[
\psi_{u,v} \: \Fa(D_u) \to \Fa(D_v)
\]
denote the map assigned to the saddle cobordism $D_u \to D_v$, and write
\[
\psi_{u,v}(x) = \sum_{z\in \KG{v}} k_z z
\]
where each $k_z \in \{0,1\}$. Say that $y$ \emph{appears in $\psi_{u,v}(x)$} if the above coefficient $k_y$ is nonzero. If $y$ appears in $\psi_{u,v}(x)$, then $\MspJxy$ is defined to consist of a single element, which we denote
\[
x\to y.
\]
If $y$ does not appear in $\psi_{u,v}(x)$, set $\MspJxy = \varnothing$.
Now suppose that $u \leq_1 v$ is the $J$ map. Let $s$ denote the number of essential circles in $D_u = D_v$. We distinguish two cases. First, if $J$ is one of $E$ or $F$, again write
\[
Jx = \sum_{z\in \KG{v}} \l_z z
\]
where now each $\l_z\in \{-1,0,1\}$, and say that $y$ appears in $Jx$ if $\l_y\neq 0$. If $y$ appears in $Jx$, then it is obtained from $x$ by the action of $J$ on an essential circle in $D_u$. Let $1\leq i \leq s$ denote the position of this essential circle, and define $\MspJxy$ to consist of a single element, which we denote
\[
x \xrightarrow{i} y.
\]
If $y$ does not appear in $Jx$, then set $\MspJxy = \varnothing$.
Finally, if $J=H$, set $\ModSp_H(y,x)=\varnothing$ unless the labels ($v_\pm,1,X$) on each circle in $x$ and $y$ match, in which case $\ModSp_H(y,x)$ contains a point for every essential circle in $D_u=D_v$. These $s$ points in $\ModSp_H(y,x)$ will be denoted
\[
x \xrightarrow{i} y
\]
for $1\leq i \leq s$.
We now describe the thick embeddings, which again depend on the type of edge $u\leq_1 v$. If $u\leq_1 v$ is not the $J$ map, then $\AmbS$ is a point, so the map is unique.
Suppose now that $u\leq_1 v$ corresponds to the $J$ map. Let $s$ denote the number of essential circles in $D_u = D_v$. Recall that $\AmbS = \cMod \times (0,1) \cong (0,1)$. We define our thick embedding $\MspJxy\xthickemb{f}\AmbS$ by setting\footnote{The important data here is the ordering of points in the interval; a specific formula is given just to fix the convention.}
\begin{equation}
\label{eq:m=1 J-embedding}
f(x\xrightarrow{i}y):=\frac{i}{s+1}\in(0,1),
\end{equation}
with framing determined as follows. If $J$ is either $E$ or $F$, then the framing is given by the sign of $(-1)^{i+1}$. If $J=H$, then the framing is $\pm 1$ if $x$ (and thus $y$) is labeled by $v_{\pm }$ on the $i$-th essential circle in $D_u$.
Note that all of these are actually thin embeddings when the subcube is $1$-dimensional. It follows from the construction of the spaces and thin embeddings that item \eqref{item:differential condition} of Theorem \ref{thm:ffcJ} is clearly satisfied.
\begin{remark}
Recall from \eqref{eq:H map} that $Hx = \operatorname{adeg}(x) y$, where the circle labels in $y$ match those in $x$. It may seem natural to define the $0$-dimensional moduli space $\ModSp_H(y,x)$ to contain $\vert\operatorname{adeg}(x) \vert$ points, and define the thick embedding to be constant, with framing given by the sign of $\operatorname{adeg}(x)$. In the above definition of $\ModSp_H(y,x)$ we are motivated by the refined formula \eqref{eq:H map refined} in which $H$ acts on each essential circle.
\end{remark}
\begin{example}
\label{ex:0-dim ex}
We illustrate the $0$-dimensional moduli spaces for an edge corresponding to various $J$ maps when $D_u = D_v$ consists of two essential circles. The thin horizontal segment is the interval $\AmbS \cong (0,1)$, the dot $\bullet$ is the image of $\MspJxy$, and the arrows indicate the framing.
Moduli spaces $\ModSp_E(y,x)$ are depicted below.
\begin{equation*}
\includestandalone{moduli_space_ex_E}
\end{equation*}
For $J=F$ the situation is similar; see also Lemma \ref{lem:E and F duality} and the discussion surrounding it.
Moduli spaces $\ModSp_H(y,x)$ are shown below. There are four generators of $\Fa(D_u)$, hence four nonempty moduli spaces, each containing two elements of the form $x \rar{i} y$ for $i=1, 2$. The numbers above the dots $\bullet$ in the following diagram indicate the image of the point $x \rar{i} y$.
\begin{equation*}
\includestandalone{moduli_space_ex_H}
\end{equation*}
The moduli spaces for pairs $(x,y)$ which are not depicted above are all empty.
\end{example}
\section{Constructing the 1-dimensional moduli spaces for square faces}
\label{sec:m=2}
We now define the $1$-dimensional moduli spaces and their thick embeddings, with an eye towards the inductive step in Section \ref{sec:m-1 to m}. Although the combinatorial analysis in this section can be simplified by the more general considerations in Section \ref{sec:Path moduli spaces}, we present a ``hands-on'' approach to illustrate the notions in Section \ref{sec:Definitions and overall strategy}.
Fix $u\leq_2 v$, $x\in \KG{u}$, and $y\in \KG{v}$. The moduli space $\MspJxy$ will be a disjoint union of closed intervals, so we need to specify which points in $\de \MspJxy$ bound an interval in $\MspJxy$. Item \eqref{item:ffcj boundary definitions} of Theorem \ref{thm:ffcJ} requires that these boundary points come from products corresponding to composition along the edges of the subcube. Letting $w_1, w_2$ denote the two vertices with $u\leq_1 w_1, w_2 \leq_1 v$, we introduce the notations
\begin{align}
\begin{split}
\pathPts{w_1}{\MspJxy} & := \coprod_{z\in \KG{w_1}} \ModSp_J(y,z) \times \ModSp_J(z,x),\\
\pathPts{w_2}{\MspJxy} & := \coprod_{z\in \KG{w_2}} \ModSp_J(y,z) \times \ModSp_J(z,x)
\end{split}
\label{eq:m=2 path moduli spaces}
\end{align}
for these two composition moduli spaces - one for the edge-path passing through $w_1$, and the other for the edge-path passing through $w_2$. Thus we have
\[\de\MspJxy = \pathPts{w_1}{\MspJxy} \,\coprod\, \pathPts{w_2}{\MspJxy}.\]
We write elements of each $\pathPts{w_i}{\MspJxy}$ as
\begin{equation}
\label{eq:elements in composition moduli spaces}
(x\rightarrow z \rightarrow y):= (z\rightarrow y , x\rightarrow z)\in \ModSp_J(y,z) \times \ModSp_J(z,x)
\end{equation}
for various $z\in\KG{w_i}$. If one of the edges of $\scube$ corresponds to the $J$ map, we decorate the corresponding arrow in $(x\rightarrow z \rightarrow y)$ according to which essential circle was acted upon as in Section \ref{sec:0-dim moduli spaces}. Each composition moduli space comes equipped with a thick product embedding
\begin{equation}
\label{eq:m=2 boundary embeddings}
\pathPts{w_i}{\MspJxy} \thickemb \AmbSp{v}{w_i}\times \AmbSp{w_i}{u}\cong
\begin{cases}
D^0 & \text{if $J$ is not involved}\\
(0,1) & \text{if $J$ is involved}
\end{cases}
\end{equation}
induced by the definitions in Section \ref{sec:0-dim moduli spaces}. Our goal now is to construct a moduli space with thick embedding
\[\MspJxy \thickemb \AmbS \cong
\begin{cases}
D^1 & \text{if $J$ is not involved}\\
D^1\times (0,1) & \text{if $J$ is involved}
\end{cases}
\]
which `fills' the thick embeddings \eqref{eq:m=2 boundary embeddings} in the sense that Equation \eqref{eq:respecting boundary structure} is satisfied. Naturally, we split into cases depending on whether or not $J$ was involved.
\subsection{Square faces involving only the annular Khovanov differential}
\label{sec:m=2 J not involved}
A subcube $u\leq_2 v$ that does not involve the $J$ edge corresponds to a square face in the cube of resolutions of $D$, and in this case we proceed exactly as in \cite{LS}. From \eqref{eq:m=2 path moduli spaces} and the analysis in \cite{LS}, we see that
\[ \size{\pathPts{w_1}{\MspJxy}} = \size{\pathPts{w_2}{\MspJxy}} =: N \in \{0,1,2\}\]
where the number of points $N$ depends on the combinatorics of the square face $D_u\rightarrow D_v$.
Following \cite{LS}, we define
\[
\ModSp_J(y,x) := \coprod_S \cMod
\]
where $S$ is a set of $N$ elements. If $N\leq 1$, then there is no ambiguity in the above definition. If $N=2$, then the ladybug matching \cite[Section 5.5]{LS} is used to determine the pairing of the two pairs of boundary points in $\de \ModSp_J(x,y)$. To define the thick embedding, recall that
\[
\AmbS = \cMod
\]
in this case, so we let
\[
\ModSp_J(y,x) \xthickemb{\coprod \operatorname{id}} \AmbS
\]
be a disjoint union of the identity map. In particular, the above thick embedding is a trivial cover requiring no framing data, and this is exactly how the cubical flow category $\ffc_{Kh}$ is constructed in \cite{LS} (the framing data in \cite{LS} is entirely inherited by the neat embedding of the cube moduli spaces $\cMod\hookrightarrow \cEuc$; see Section \ref{sec:ffc review} and Definition 5.5 there).
\subsection{Square faces involving $E$ or $F$}
\label{sec:m=2 J=E/F involved}
Now we suppose that the subcube $u\leq_2 v$ involves the $J=E$ map; the case involving $J=F$ is nearly identical (see Lemma \ref{lem:E and F duality}). Without loss of generality, let the edges $u \to w_1$ and $w_2\to v$ be saddle maps $S$, while $u\to w_2$ and $w_1\to v$ are the $E$ map. In particular, we have equality of resolutions $D_u = D_{w_2}$, $D_{w_1} = D_v$, and the saddles $D_u \to D_{w_1}$ and $D_{w_2} \to D_v$ are identical.
\begin{equation*}
\label{eq:square with J}
\begin{tikzcd}
D_u \ar[d, "E"'] \ar[r, "S"] & D_{w_1} \ar[d, "E"] \\
D_{w_2} \ar[r, "S"] & D_v
\end{tikzcd}
\end{equation*}
We focus first on connected arc diagrams $D_u$, and we begin by considering those for which the saddle $S$ does not change the number of essential circles. There are four of these, as shown in Figure \ref{fig:saddles not changing number of ess circles}.
\begin{figure}
\begin{subfigure}[b]{.25\textwidth}
\begin{center}
\includestandalone{arc_diagram_multiplying_trivials}
\end{center}
\caption{}\label{fig:multiplying trivial circles}
\end{subfigure
\begin{subfigure}[b]{.25\textwidth}
\begin{center}
\includestandalone{arc_diagram_splitting_trivial_into_trivials}
\end{center}
\caption{}\label{fig:splitting trivial into trivials}
\end{subfigure
\begin{subfigure}[b]{.25\textwidth}
\begin{center}
\includestandalone{arc_diagram_multiplying_trivial_and_essential}
\end{center}
\caption{}\label{fig:multiplying trivial and essential}
\end{subfigure
\begin{subfigure}[b]{.25\textwidth}
\begin{center}
\includestandalone{arc_diagram_splitting_trivial_from_essential}
\end{center}
\caption{}\label{fig:splitting trivial from essential}
\end{subfigure}
\caption{Connected arc diagrams where the number of essential circles is unchanged}\label{fig:saddles not changing number of ess circles}
\end{figure}
For Figures \ref{fig:multiplying trivial circles} and \ref{fig:splitting trivial into trivials}, the $E$ map is the zero map, and all of the moduli spaces are empty.
In the case of Figure \ref{fig:multiplying trivial and essential}, we see that all of the moduli spaces are empty unless $x=1 \o v_-$ and $y=v_+$, in which case we have single point composition moduli spaces
\[
\pathPts{w_1}{\MspExy} = \{(1\o v_- \rightarrow v_- \xrightarrow{1} v_+)\},
\]
\[
\pathPts{w_2}{\MspExy} = \{(1 \o v_- \xrightarrow{1} 1 \o v_+ \rightarrow v_+)\}
\]
as illustrated in Figure \ref{fig:mult ess and triv gens}.
\begin{figure}
\hspace*{\fill}
\begin{subfigure}{.4\textwidth}
\begin{center}
\includestandalone{square_face_bijection_ex}
\end{center}
\caption{}\label{fig:mult ess and triv gens}
\end{subfigure}
\hfill
\begin{subfigure}{.2\textwidth}
\begin{center}
\includestandalone{framed_embedding_bijection_ex}
\end{center}
\caption{}\label{fig:mult ess and triv modsp}
\end{subfigure}
\hspace*{\fill}
\caption{The relevant generators in the subcube of resolutions for the case of Figure \ref{fig:multiplying trivial and essential}, together with the corresponding moduli space embedded and framed into $D^1\times (0,1)$.}
\end{figure}
Figure \ref{fig:splitting trivial from essential} is dual to Figure \ref{fig:multiplying trivial and essential}, so the analysis is similar. We have empty moduli spaces unless $x=v_-$ and $y=X\o v_+$, in which case we have
\begin{align*}
\pathPts{w_1}{\MspExy} & = \{(v_-\rightarrow X \o v_- \xrightarrow{1} X \o v_+)\},\\
\pathPts{w_2}{\MspExy} & = \{(v_- \xrightarrow{1} v_+ \rightarrow X \o v_+)\}
\end{align*}
In all of these cases where $S$ does not change the number of essential circles, we see that there is a bijection of one-point composition moduli spaces that commutes with the thin embeddings
\begin{equation}
\begin{aligned}
\label{eq:1dim bij for trivial S}
\begin{tikzpicture}[x=1in,y=-.5in]
\node(A) at (0,0) {$\pathPts{w_1}{\MspExy}$};
\node(B) at (0,2) {$\pathPts{w_2}{\MspExy}$};
\node(C) at (2,0) {$\AmbSp{v}{w_1}\times\AmbSp{w_1}{u}$};
\node(D) at (2,2) {$\AmbSp{v}{w_2}\times\AmbSp{w_2}{u}$};
\node(E) at (2,1) {$(0,1)$};
\draw[->] (A)-- node[left]{$\cong$} (B);
\draw[right hook->] (A) -- (C);
\draw[right hook->] (B) -- (D);
\path (C)-- node[rotate=-90]{$\cong$} (E);
\path (E)-- node[rotate=-90]{$\cong$} (D);
\end{tikzpicture}.
\end{aligned}
\end{equation}
This commuting bijection allows us to define $\MspExy$ as a single closed interval $D^1$ which is thinly embedded as an identity
\[\MspExy:=D^1\xhookrightarrow{\operatorname{id} \times\{p_1\}} D^1\times (0,1)=\AmbS\]
with the constant framing. See Figure \ref{fig:mult ess and triv modsp} for a depiction of the case where $S$ is the saddle in Figure \ref{fig:multiplying trivial and essential}.
Next, we consider the case where $S$ changes the number of essential circles. For this there are two possible connected arc diagrams, illustrated in Figure \ref{fig:saddles changing number of ess circles}.
\begin{figure}
\begin{subfigure}[t]{.4\textwidth}
\begin{center}
\includestandalone{arc_diagram_splitting_trivial_into_essentials}
\end{center}
\caption{}\label{fig:splitting trivial into essentials}
\end{subfigure
\begin{subfigure}[t]{.4\textwidth}
\begin{center}
\includestandalone{arc_diagram_multiplying_essentials}
\end{center}
\caption{}\label{fig:multiplying essentials}
\end{subfigure
\caption{Connected arc diagrams where the number of essential circles is changed.}\label{fig:saddles changing number of ess circles}
\end{figure}
In Figure \ref{fig:splitting trivial into essentials}, we see that all moduli spaces are empty unless $x=1$ and $y=v_+\otimes v_+$, in which case we have $\pathPts{w_1}{\MspExy}$ containing two points while $\pathPts{w_2}{\MspExy}$ is empty (see Figure \ref{fig:comult triv gens}).
\[\pathPts{w_1}{\MspExy} = \{(1\rightarrow v_-\otimes v_+ \xrightarrow{1} v_+\otimes v_+), (1\rightarrow v_+\otimes v_- \xrightarrow{2} v_+\otimes v_+)\},\]
\[\pathPts{w_2}{\MspExy} = \varnothing.\]
\begin{figure}
\hspace*{\fill}
\begin{subfigure}{.4\textwidth}
\begin{center}
\includestandalone[scale=.8]{square_face_turnback_ex}
\end{center}
\caption{}\label{fig:comult triv gens}
\end{subfigure}
\hfill
\begin{subfigure}{.2\textwidth}
\begin{center}
\includestandalone{framed_embedding_turnback_ex}
\end{center}
\caption{}\label{fig:comult triv modsp}
\end{subfigure}
\hspace*{\fill}
\caption{The relevant generators in the subcube of resolutions for the case of Figure \ref{fig:splitting trivial into essentials}, together with the corresponding moduli space embedded and framed into $\AmbS \cong D^1\times (0,1)$.}
\end{figure}
Following the embedding instructions from Section \ref{sec:0-dim moduli spaces}, we see that our two points in $\pathPts{w_1}{\MspExy}$ are embedded in $\AmbSp{v}{w_1}\times\AmbSp{w_1}{u}\cong (0,1)$ with framing shown below.
\begin{equation*}
\begin{aligned}
\includestandalone{framed_embedding_turnback_boundary_ex}
\end{aligned}
\end{equation*}
We then define $\MspExy$ in this setting to be a single interval which is thinly embedded into $\AmbS\cong D^1\times (0,1)$ as a framed turnback as indicated in Figure \ref{fig:comult triv modsp}.
The last connected case is dual to the previous one, where the moduli spaces are empty unless $x=v_-\otimes v_-$ and $y=X$, and the roles of $\pathPts{w_1}{\MspExy}$ and $\pathPts{w_2}{\MspExy}$ are reversed. The details here are left to the reader.
Finally, we consider the situation when the diagram $D_u$ is disconnected. We let $\Cs\subset D_u$ denote the connected component of the saddle arc $S$ within $D_u$, while $\Cs'$ denotes the complement of $\Cs$. As in Section \ref{sec:Arc diagrams}, these components determine connected components in the resolutions at every vertex, which we also denote by $\Cs$ and $\Cs'$. Recall from Section \ref{sec:Arc diagrams} the notion of restriction of generators. This allows us to write generators as $x=x|_\Cs \otimes x|_{\Cs'}$ and $y=y|_\Cs\otimes y|_{\Cs'}$, and we use this breakdown to aid in our analysis.
\begin{lemma}
\label{lem:m=2 conn comp and annular deg}
Suppose some $\pathPts{w_i}{\MspExy}$ is nonempty. Then precisely one of the two following statements are true.
\begin{enumerate}
\item \label{it:m=2 E acts on C}
$\operatorname{adeg}(y|_\Cs)=\operatorname{adeg}(x|_\Cs)+2$ and $\operatorname{adeg}(y|_{\Cs'}) = \operatorname{adeg}(x|_{\Cs'})$. This corresponds to $E$ acting on an essential circle in $\Cs$.
\item \label{it:m=2 E acts on C'}
$\operatorname{adeg}(y|_\Cs)=\operatorname{adeg}(x|_\Cs)$ and $\operatorname{adeg}(y|_{\Cs'}) = \operatorname{adeg}(x|_{\Cs'})+2$. This corresponds to $E$ acting on an essential circle in $\Cs'$.
\end{enumerate}
When $J=F$, the value $2$ above is replaced by $-2$.
\end{lemma}
\begin{proof}
Saddle maps preserve annular degree, but the $E$ map increases annular degree by two. Similarly, the $F$ map decreases annular degree by two.
\end{proof}
If $\pathPts{w_1}{\MspExy}=\pathPts{w_2}{\MspExy}=\varnothing$, we define $\MspExy:=\varnothing$ as well. Otherwise, we use Lemma \ref{lem:m=2 conn comp and annular deg} to consider the two cases.
In the case \eqref{it:m=2 E acts on C} where $E$ was acting upon the component $\Cs$ containing $S$, we must have the labels of $y|_{\Cs'}$ matching the labels of $x|_{\Cs'}$, in which case these components are irrelevant for the analysis of the moduli spaces and we may define
\[\MspExy:=\MspE{y|_\Cs}{x|_\Cs}\]
as defined and embedded earlier. This case is illustrated for a specfic example in Figure \ref{fig:framed embedding turnback ex disconnected}.
Meanwhile, in the case \eqref{it:m=2 E acts on C'} where $E$ was acting on some essential circle in $\Cs'$ disconnected from the saddle $S$, we see that the generator at $w_1$ must be $y|_\CC \o x|_{\CC'}$ while the generator at $w_2$ must be the `opposite' $x|_\CC \o y|_{\CC'}$. This allows us to show that the maps $S$ and $E$ (and their corresponding edge moduli spaces) commute in the following sense:
\begin{equation}
\label{eq:edges commute}
\begin{aligned}
\pathPts{w_1}{\MspExy} &=
\MspE{y|_\CC \o y|_{\CC'}} {y|_\CC \o x|_{\CC'}} \times \MspE{y|_\CC \o x|_{\CC'}} {x|_\CC \o x|_{\CC'}}\\
& \cong \MspE{y|_{\CC'}}{x|_{\CC'}} \times \MspE{y|_\CC}{x|_\CC}\\
& \cong \MspE{y|_\CC}{x|_\CC} \times \MspE{y|_{\CC'}}{x|_{\CC'}}\\
& \cong \MspE{y|_\CC \o y|_{\CC'}} {x|_\CC \o y|_{\CC'}} \times \MspE{x|_\CC \o y|_{\CC'}} {x|_\CC \o x|_{\CC'}}\\
& = \pathPts{w_2}{\MspExy}.
\end{aligned}
\end{equation}
The framed embedding data for such moduli spaces is determined by which circle is being acted on by $E$, and this is maintained throughout all of the equivalences above. Therefore we again have a (one point) bijection $\pathPts{w_1}{\MspExy}\cong\pathPts{w_2}{\MspExy}$ that commutes with the thin embeddings as in Equation \eqref{eq:1dim bij for trivial S}, allowing us to once again define $\MspExy$ as
\[\MspExy:=D^1\xhookrightarrow{I\times\{p_1\}} D^1\times (0,1)=\AmbS.\]
This case is illustrated for a specific example in Figure \ref{fig:framed embedding bijection ex disconnected}.
\begin{figure}
\begin{subfigure}{\textwidth}
\begin{center}
\includestandalone[scale=.8]{square_face_turnback_ex_disconnectedv2}
\end{center}
\caption{The generators and differentials within the subcube $\subcube{2}{v}{u}$. The numbers on maps indicate which essential circle is being acted on by $E$. Certain tensor products are written with subscripts to avoid clutter.}
\end{subfigure}\\
\hspace*{\fill}
\begin{subfigure}{.4\textwidth}
\begin{center}
\includestandalone{framed_embedding_turnback_ex_disconnected}
\end{center}
\caption{The moduli space for $x=1\o v_-$ and $y=v_+ \o v_+ \o v_-$, embedded and framed in $\AmbS$.}
\label{fig:framed embedding turnback ex disconnected}
\end{subfigure}
\hfill
\begin{subfigure}{.4\textwidth}
\begin{center}
\includestandalone{framed_embedding_bijection_ex_disconnected_1}
\end{center}
\caption{The moduli space for $x=1\o v_-$ and either $y=v_-\o v_+\o v_+$ or $y=v_+\o v_-\o v_+$, embedded and framed in $\AmbS$.}
\label{fig:framed embedding bijection ex disconnected}
\end{subfigure}
\hspace*{\fill}
\caption{Some examples of $\MspExy\hookrightarrow\AmbS\cong D^1 \times(0,1)$ for a subcube $\subcube{2}{v}{u}$ involving $E$ for a disconnected arc diagram. For the fixed $x=1\o v_-$, different values of $y$ correspond to different cases of Lemma \ref{lem:m=2 conn comp and annular deg}. }
\end{figure}
\subsection{Square faces involving $H$}
\label{sec:m=2 J=H involved}
Finally, we turn to the case where $J=H$. As before, we distinguish between two cases.
Suppose first that the saddle $S$ does not change the number of essential circles in $D_u$. Let $s$ denote the number of essential circles in $D_u$ (equivalently, in any of $D_u, D_{w_1}, D_{w_2}$, or $D_v$). If $s=0$, then $\de_{w_1} \MspHxy = \de_{w_2} \MspHxy = \varnothing$, and we define $\MspHxy = \varnothing$. Otherwise, there are unique generators in $z \in \KG{w_1}$ and $z'\in \KG{w_2}$ such that
\begin{align*}
\de_{w_1}\MspHxy &= \MspH{y}{z} \times \MspH{z}{x} \\ \de_{w_2} \MspHxy &= \MspH{y}{z'} \times \MspH{z'}{x}.
\end{align*}
In fact, $z=y$ and $z'=x$ as generators of $D_u=D_{w_2}$ and $D_v=D_{w_1}$, respectively. The edge moduli spaces $\de_{w_1}\MspHxy$ and $\de_{w_2}\MspHxy$ each have $s$ elements, thinly embedded in $(0,1)$ in order according to Equation \eqref{eq:m=1 J-embedding}. From the formulas \eqref{eq:formula1} and \eqref{eq:formula3}, we see that the essential circles in $D_u = D_{w_2}$ and $D_v=D_{w_1}$ have the same labels in $x=z'$ and $y=z$, respectively, so the $i$-th point in $\de_{w_1}\MspHxy$ has the same framing as the $i$-th point in $\de_{w_2}\MspHxy$. Thus we have a bijection of $s$-point composition moduli spaces that commutes with the thin embeddings precisely as in Equation \eqref{eq:1dim bij for trivial S}. We therefore define $\MspHxy$ to be a disjoint union of $s$ intervals, embedded in $\AmbS$
as a framed identity cobordism from $\de_{w_1}\MspHxy$ to $\de_{w_2}\MspHxy$ connecting the $i$-th point of $\de_{w_1}\MspHxy $ to the $i$-th point of $\de_{w_2} \MspHxy$.
Suppose now that $S$ changes the number of circles. Suppose that $D_v=D_{w_1}$ has $s$ circles and $D_u = D_{w_2}$ has $s+2$, so that $S$ is the Type II saddle from Figure \ref{fig:type2}. If both $\de_{w_1}\MspHxy = \de_{w_2}\MspHxy = \varnothing$, then define $\MspHxy = \varnothing$. Otherwise, there are generators $z\in \KG{w_1}$, $z' \in \KG{w_2}$ such that
\begin{align*}
\de_{w_1}\MspHxy &= \MspH{y}{z} \times \MspH{z}{x} \\ \de_{w_2} \MspHxy &= \MspH{y}{z'} \times \MspH{z'}{x}.
\end{align*}
The edge moduli spaces $\MspH{z}{x}$ and $\MspH{y}{z'}$ each have one element, while $\MspH{y}{z}$ and $\MspH{z'}{x}$ have $s$ and $s+2$ elements, respectively, corresponding to the number of essential circles in $D_{w_1}$ and $D_{w_2}$. Note that the two essential circles merged by $S$ are consecutive in $D_{w_2}$, and their labels in $z'$ are different. Define $\MspHxy$ to consist of $s+1$ intervals, with $s$ of them thinly embedded in $\AmbS$ matching the $s$ points in $\MspH{y}{z}$ with the $s$ points in $\MspH{z'}{x}$ that correspond to essential circles not participating in the saddle. The remaining interval in $\MspHxy$ is thinly embedded as a turnback with boundary the two remaining unmatched points in $\de_{w_2} \MspHxy$ corresponding to the essential circles created by $S$. See Figure \ref{fig:m=2 framed embedding for H ex} for an example. Finally, the case when $S$ increases the number of essential circles by two is dual to the above discussion.
\begin{figure}
\hspace*{\fill}
\begin{subfigure}{.55\textwidth}
\begin{center}
\includestandalone[scale=.8]{square_face_turnback_ex_disconnected_H}
\end{center}
\caption{The generators and differentials within the subcube $\subcube{2}{v}{u}$.}
\end{subfigure}
\hfill
\begin{subfigure}{.3\textwidth}
\begin{center}
\includestandalone{framed_embedding_turnback_ex_disconnected_H}
\end{center}
\caption{The moduli space $\MspHxy$ embedded and framed in $\AmbS$.}
\end{subfigure}
\hspace*{\fill}
\caption{An example of $\MspHxy\hookrightarrow\AmbS\cong D^1 \times(0,1)$ for a subcube $\subcube{2}{v}{u}$ involving an $H$ edge and a saddle $S$ which decreases the number of essential circles by two. The two merged circles must be consecutive, allowing a framed turnback to `cancel' the corresponding points in $\de_{w_2}\MspHxy$.}\label{fig:m=2 framed embedding for H ex}
\end{figure}
\section{Path moduli spaces}
\label{sec:Path moduli spaces}
This section contains a detailed combinatorial analysis that will be crucial for completing our inductive construction in Section \ref{sec:m-1 to m}. For an $m$-dimensional subcube $u\leq_m v$ and generators $x\in \KG{u}, y\in \KG{v}$, Item \eqref{item:ffcj boundary definitions} of Theorem \ref{thm:ffcJ} states that the codimension $m-1$ boundary of $\MspJxy$ is built out of products of the moduli spaces associated to a sequence of edges from $u$ to $v$. Such a sequence of edges is a \emph{path} from $u$ to $v$, and each edge moduli space, along with its thick embedding, was defined in Section \ref{sec:0-dim moduli spaces}. We analyze these \emph{path moduli spaces} and their thick embeddings, obtained as the product of the thick embeddings for each edge moduli space. While each edge moduli space is thinly embedded, taking the disjoint union over intermediate generators as in Theorem \ref{thm:ffcJ} \eqref{item:ffcj boundary definitions} may produce a thick embedding.
We focus on connected arc diagrams in this section. The key results are Proposition \ref{prop:E/F conn is thin} and Proposition \ref{prop:conn H}. In Proposition \ref{prop:E/F conn is thin} we show that for $J=E, F$ all path moduli spaces involving $J$ are in fact thinly embedded. For $J=H$, path moduli spaces are thinly embedded except for one type of arc diagram, for which an explicit description is provided, see Proposition \ref{prop:conn H}. The outline is as follows. In Section \ref{sec:basic defs and lemmas} we introduce path moduli spaces, their thick embeddings, and present a helpful lemma which bounds the number of points in some path moduli spaces. Section \ref{sec:equiv path moduli spaces} establishes two key lemmas. First, Lemma \ref{lem:biject to all ess} allows us to assume that a path has the special property that the arc diagram just before the $J$ edge (the so-called \emph{$J$ resolution}) consists entirely of essential circles. Second, in Lemma \ref{lem:forbidden configs} we give restrictions on the type of arc configurations that may appear in such a diagram.
To establish thinness we focus on the $J$-resolution arc diagram. While the two cases $J=E,F$ and $J=H$ require different arguments, the overall strategy is the following. Connectivity and the formulas \eqref{eq:formula2} and \eqref{eq:formula4} imply that picking an essential circle for $J$ to act on and a label $v_\pm$ on that circle determines labels on all the other circles. Lemma \ref{lem:genus considerations} ensures thinness ``before and after'' the $J$ edge. With further analysis, depending on $J=E, F$ or $J=H$, we arrive at our desired results. Note that connectivity of the arc diagram is critical: disconnected arc diagrams may yield non-thinly embedded path moduli spaces. See also the discussion at the beginning of Section \ref{sec:m-1 to m J=E disconn}.
\subsection{Basic definitions and lemmas}
\label{sec:basic defs and lemmas}
\begin{definition}
\label{def:path}
A \emph{path} $P$ through a subcube $\scube$ is a sequence of edges, denoted as
\[P = (u\rightarrow w_1 \rightarrow \cdots \rightarrow w_{m-1} \rightarrow v).\]
Fixing generators $x\in\KG{u},y\in\KG{v}$, any such path determines a \emph{path moduli space}
\[\pathP := \coprod \MspJ{y}{z_{m-1}} \times \MspJ{z_{m-1}}{z_{m-2}} \times \cdots \times \MspJ{z_1}{x},\]
where the disjoint union is taken over all sequences $(z_1,\ldots,z_{m-1})$ with $z_i\in\KG{w_i}$ for each $1\leq i \leq m-1$. As in Section \ref{sec:m=2}, points in $\pathP$ will be denoted as
\[(x\rightarrow z_1 \rightarrow\cdots \rightarrow z_{m-1} \rightarrow y)\in\pathP,\]
with a label on an arrow corresponding to the $J$-edge indicating which essential circle is being acted on by $J$ if necessary (see Equation \eqref{path eq} for an example of such a label).
\end{definition}
Recall from Definition \ref{def:flow cat} that we use the notation $\de_{[i]}\ModSp$ to indicate the codimension-$i$ portion of the boundary of a moduli space. In order to satisfy the demands of Theorem \ref{thm:ffcJ}, we define
\begin{equation}\label{disjoint union paths}
\de_{[m-1]}\MspJxy \cong \coprod_{\text{paths $P$}} \pathP
\end{equation}
where the disjoint union is taken over paths $P$ from $u$ to $v$; to make the notation less cumbersome we omit reference to $u, v$.
Furthermore, every $\pathP$ inherits a thick embedding
\begin{equation}
\label{eq:product cube ambient space}
\pathP \xthickemb{\varphi_P} \mathcal{E}(P) := \AmbSp{v}{w_{m-1}} \times \cdots \times \AmbSp{w_1}{u}
\end{equation}
via the disjoint union of products of the thin embeddings
\[
\MspJ{z_{m-i}}{z_{m-i-1}} \hookrightarrow \AmbSp{w_{m-i}}{w_{m-i-1}}
\]
defined in Section \ref{sec:0-dim moduli spaces}. If the subcube $\scube$ does not involve $J$, then $\pAmb$ is a point and the thick embedding is constant. Otherwise, $J$ is involved and there is a natural identification $\pAmb \cong (0,1)$. Then the thick embedding \eqref{eq:product cube ambient space} sends a point
\begin{equation} \label{path eq}
(x \to z_1 \to \cdots \to z_i \rar{j} z_{i+1} \to \cdots \to z_{m-1} \to y) \in \pathPts{P}{\MspJxy},
\end{equation}
to the fixed $j$-th point in $(0,1)$ with framing determined by the sign $(-1)^{j+1}$ if $J=E, F$ and otherwise given by the label of the $j$-th circle in $z_i$ if $J=H$, as described in Section \ref{sec:0-dim moduli spaces}.
We may also consider \emph{sub-paths} of a path $P$. If $w_k$ is any vertex along a path
\[P=(u\rightarrow w_1\rightarrow \cdots \rightarrow w_k \rightarrow \cdots \rightarrow v),\]
we may define $P_{\leq w_k} :=(u\rightarrow\cdots \rightarrow w_k)$ and $P_{\geq w_k} :=(w_k\rightarrow \cdots \rightarrow v)$.
This then allows us to split the path moduli space as a product
\[\pathP = \coprod_{z\in \KG{w_k}} \pathPts{P_{\geq w_k}}{\MspJ{y}{z}} \times \pathPts{P_{\leq w_k}}{\MspJ{z}{x}}\]
embedded as a product with the product framing.
\begin{definition}
For a fixed path $P$ through a sub-cube $\scube$ that involves $J$, we single out the vertices along $P$ directly before and after the $J$ map has been applied:
\begin{itemize}[label={$\bullet$}]
\item $e_0=e_0(P)$ denotes the source vertex of the $J$ edge in $P$, and
\item $e_1=e_1(P)$ denotes the target vertex of the $J$ edge in $P$.
\end{itemize}
The arc diagram $D_e:=D_{e_0}=D_{e_1}$ being acted on by $J$ will be referred to as the \emph{$J$-resolution of $P$}.\footnote{To relate this notation to equation (\ref{path eq}), $e_0=w_i$, $e_1=w_{i+1}$.}
\end{definition}
Note that the $J$-resolution of a path $P$ depends on the path, although we do not include $P$ in the notation to avoid clutter.
Now if $u\leq_m v$ does not involve the $J$ map, then any path $P$ from $u$ to $v$ represents a path in the cube of resolutions of $D$. For each such path there is an associated cobordism $S \: D_u \to D_v$.
\begin{lemma}\label{lem:genus considerations}
Let $P$ be a path from $u$ to $v$ which does not involve the $J$ map, and let $S \: D_u \to D_v$ denote its corresponding surface cobordism. Let $g$ denote the maximal genus among all components of $S$.
\begin{enumerate}
\item If $g=0$, then for every $x\in \KG{u}$ and $y\in \KG{v}$, the moduli space $\pathPts{P}{\MspJxy}$ contains at most one element.
\item If $g \geq 1$ and one of $D_u$ or $D_v$ consists only of essential circles, then $\Fa(S) = 0$. In particular in this case, for every $x\in \KG{u}$ and $y\in \KG{v}$, the moduli space $\pathPts{P}{\MspJxy}$ is empty.
\end{enumerate}
\end{lemma}
\begin{proof}
For (1), first view $S$ as a cobordism in $\mathbb{R}^2\times I$ using the inclusion $\mathbb{A} \times I \hookrightarrow \mathbb{R}^2\times I$. Since $\Fa$ is the $\operatorname{adeg}$-preserving part of $\mathcal{F}_{Kh}$, the cardinality of $\pathPts{P}\MspJxy$ is bounded above by the coefficient of $y$ in $\mathcal{F}_{Kh}(S)(x)$. Represent the generators $x$ and $y$ using (dotted) cup cobordisms, denoted $\Sigma_x$ and $\Sigma_y$ respectively. Let $\Sigma_y^*$ denote the dual cap cobordism to $\Sigma_y$, so that the closed surface
\[
\Sigma_y^* \circ \Sigma_y
\]
evaluates to $1$ under $\mathcal{F}_{Kh}$. The coefficient of $y$ in $\mathcal{F}_{Kh}(S)(x)$ is then the value of $\mathcal{F}_{Kh}$ on the closed surface
\[
S':= \Sigma_y^* \circ S \circ \Sigma_x.
\]
Since $S$ has genus zero, every component of $S'$ is a (possibly dotted) sphere. Then $\mathcal{F}_{Kh}(S') \leq 1$, which completes the proof of statement (1).
We now address (2). Let $\overline S\subset S$ be a component of positive genus. It suffices to show $\Fa(\overline S)=0$. By neck-cutting, Figure \ref{fig:neck-cutting}, we can write $\overline S$ as a sum
\[
S_1 + \cdots + S_k,
\]
where each component of each $S_i$ is either an annulus with essential boundary or a disk, possibly carrying dots. We may assume that each component carries at most one dot. By assumption, each $S_i$ contains an annulus. We will show that each $S_i$ has a dotted annulus, which implies $\Fa(\overline S) = 0$ since $X$ acts by zero on essential circles, see formula \eqref{eq:formula1}.
Fix $1\leq i \leq k$. Recall that the Bar-Natan relations are homogeneous, so $\deg(\overline S) = \deg(S_i)$. Let $n$ denote the number of (possibly dotted) disks in $S_i$. If none of the annuli in $S_i$ are dotted, then
\[
\deg(S_i) \geq -n.
\]
On other other hand, considering the connected surface $\overline{S}$,
\[
\deg(\overline S) = \chi(\overline S) < -n,
\]
since $g(\overline S)\geq 1$ and the number of boundary components of $\overline S$ is strictly greater than $n$. Therefore $S_i$ contains a dotted annulus, which completes the proof.
\end{proof}
\subsection{Equivalent path moduli spaces and a non-emptiness condition}
\label{sec:equiv path moduli spaces}
In this section we analyze the manner in which two paths $P,P'$ can give rise to equivalent path moduli spaces $\pathPts{P}{\MspJxy}\cong\pathPts{P'}{\MspJxy}$. This analysis will allow us to choose certain preferred paths through our subcubes. It will also give rise to an important non-emptiness condition on path moduli spaces.
\begin{definition}
Let $P$ and $P'$ be two paths in a subcube $u\leq_m v$. Write
\begin{align*}
P &= \left( u = w_0 \to w_1 \to \cdots \to w_{m} = v \right) \\
P' &= \left( u = w_0' \to w_1' \to \cdots \to w_{m}' = v \right).
\end{align*}
We say that $P$ and $P'$ are \emph{related by a square face} if for some $1 \leq j \leq m-1$, we have $w_j\neq w_j'$ while $w_i = w_i'$ for all indices $i\neq j$. We say the square face involves $J$ if one of the edges in the $2$-dimensional subcube $w_{j-1} \leq_2 w_{j+1}$ is the $J$ map. In this case, two of the opposite edges are $J$, and the remaining two opposite edges are a saddle.
\end{definition}
\begin{lemma}\label{lem:square face bijections}
Let $P$ and $P'$ be two paths in a subcube $u\leq_m v$ that are related by a square face $w_{j-1} \leq_2 w_{j+1}$ as above, and let $x\in \KG{u}$, $y\in \KG{v}$. If either
\begin{enumerate}
\item the square face does not involve the $J$ map, or
\item the square face involves the $J$ map, and the number of essential circles in the resolutions appearing in $w_{j-1} \leq_2 w_{j+1}$ is the same,
\end{enumerate}
then there is a bijection
\[
\pathP \cong \pathPts{P'}{\MspJxy}
\]
such that the diagram \eqref{eq:path bijection} commutes.
\begin{equation}
\label{eq:path bijection}
\begin{tikzpicture}[x=1.2in,y=-.8in,baseline=(current bounding box.center)]
\node(A) at (0,0){$\pathP$};
\node(B) at (0,1){$\pathPts{P'}{\MspJxy}$};
\node(C) at (1,0){$\mathcal{E}(P)$};
\node(D) at (1,1){$\mathcal{E}(P')$};
\thickembtikz{A}{C}
\thickembtikz{B}{D}
\draw[->] (A) -- node[left]{$\cong$} (B);
\draw[->] (C) -- node[left]{$\cong$} (D);
\end{tikzpicture}
\end{equation}
\end{lemma}
\begin{proof}
The desired bijection is assembled from the bijections in Section \ref{sec:m=2} on the factors of $\pathP$ and $\pathPts{P'}{\MspJxy}$ involving the square face and the identity on all other factors. Commutativity of the diagram is immediate for item (1), since $\mathcal{E}(P) = D^0 = \mathcal{E}(P')$ in this case. For item (2) it follows from commutativity of the diagram \eqref{eq:1dim bij for trivial S} and the discussion below Lemma \ref{lem:m=2 conn comp and annular deg}.
\end{proof}
In the case that the number of essential circles in the $J$-resolution has changed, we instead see turnbacks as in Figure \ref{fig:comult triv modsp}, but this fact does not concern us at the moment.
\begin{corollary}
\label{cor:fixed J edge}
Let $P$ and $P'$ be paths from $u$ to $v$ such that the $J$-resolution of $P$ and $P'$ are at the same vertex, $e_0(P) = e_0(P')$. Then there is a bijection
\[
\pathP \cong \pathPts{P'}{\MspJxy}
\]
such that the diagram \eqref{eq:path bijection} commutes.
\end{corollary}
\begin{proof}
The paths $P$ and $P'$ can be related by a sequence of square faces that do not involve the $J$ map. The claim follows from Lemma \ref{lem:square face bijections}.
\end{proof}
Now we consider how the various bijections of Lemma \ref{lem:square face bijections} can be used to replace a given path through a subcube with one that is easier to analyze.
\begin{lemma}\label{lem:biject to all ess}
Consider a path $P$ through the sub-cube $\scube$, specifying a $J$-resolution arc diagram $D_e$, and fix generators $x\in\KG{u},y\in\KG{v}$. Then there exists a path $P'$ through $\scube$, with $J$-resolution $D_{e'}$, satisfying the following properties.
\begin{itemize}
\item Each connected component of $D_{e'}$ contains the same number of essential circles as the corresponding component of $D_e$.
\item Any connected component of $D_{e'}$ which contains at least one essential circle also contains no trivial circles.
\item There is a bijection $\pathPts{P}{\MspJxy} \cong \pathPts{P'}{\MspJxy}$ such that the diagram \eqref{eq:path bijection} commutes.
\end{itemize}
\end{lemma}
\begin{proof}
We work one connected component $\mathscr{C}$ of $D_e$ at a time. If $\mathscr{C}$ contains only essential circles or only trivial circles, then we are done. Otherwise, $\mathscr{C}$ must contain a trivial circle that is joined by an arc (either past or future) to an essential circle. This arc specifies a square face in the subcube $\scube$, and the swap across this square face yields a path $P''$ in which the corresponding component $\mathscr{C}''$ in the $J$-resolution of $P''$ contains one fewer trivial circle and the same number of essential circles as $\mathscr{C}$. By Lemma \ref{lem:square face bijections} (2), the moduli spaces for $P$ and $P''$ are in bijection with diagram \eqref{eq:path bijection} commuting. We continue this process until there are no trivial circles left in any of the connected components with essential circles.
\end{proof}
Lemma \ref{lem:biject to all ess} ensures that it is enough to understand $\pathPts{P}{\MspJxy}$ for paths $P$ having the $J$-resolution $D_e$ consisting of only essential circles, possibly together with a disconnected component consisting of trivial circles. However, in the following sections, we will focus largely on connected arc diagrams; the disconnected cases will follow from the connected ones in a manner very similar to the arguments of Section \ref{sec:m=2 J=E/F involved}.
We end this section with an important consequence of Corollary \ref{cor:fixed J edge}. Recall from Section \ref{sec:Arc diagrams} that a subcube $\scube$ gives rise to its own arc diagrams at each vertex including only those arcs that correspond to saddle maps within the subcube. Note further that, for any path $P$, the arc diagrams $D_{e_0}$ and $D_{e_1}$ directly before and after the $J$ map are the same. The following lemma provides an important restriction on the types of arc diagrams that can appear for non-empty path moduli spaces.
\begin{lemma}\label{lem:forbidden configs}
Let $P$ be a path through $u\leq_m v$ and let $w$ be a vertex in $P$. Suppose that any of the configurations in Figure \ref{fig:forbidden configs}, as well as their vertical and horizontal reflections, appear for some pair of arcs in the arc diagram $D_w$. Then $\pathPts{P}\MspJxy = \varnothing$ for every $x\in \KG{u}, y\in \KG{v}$.
\begin{figure}
\begin{subfigure}[b]{.3\textwidth}
\begin{center}
\includestandalone{forbidden1}
\end{center}
\caption{}\label{fig:forbidden1}
\end{subfigure}
\begin{subfigure}[b]{.3\textwidth}
\begin{center}
\includestandalone{forbidden3}
\end{center}
\caption{}\label{fig:forbidden3}
\end{subfigure}
\begin{subfigure}[b]{.3\textwidth}
\begin{center}
\includestandalone{forbidden5}
\end{center}
\caption{}\label{fig:forbidden5}
\end{subfigure} \\
\vskip1em
\begin{subfigure}[b]{.3\textwidth}
\begin{center}
\includestandalone{forbidden2}
\end{center}
\caption{}\label{fig:forbidden2}
\end{subfigure}
\begin{subfigure}[b]{.3\textwidth}
\begin{center}
\includestandalone{forbidden4}
\end{center}
\caption{}\label{fig:forbidden4}
\end{subfigure}
\begin{subfigure}[b]{.3\textwidth}
\begin{center}
\includestandalone{forbidden6}
\end{center}
\caption{}\label{fig:forbidden6}
\end{subfigure}
\caption{Forbidden configurations; only the two relevant arcs are pictured, but there could be an arbitrary number of additional past and future arcs anywhere within such a diagram.}\label{fig:forbidden configs}
\end{figure}
\end{lemma}
\begin{proof}
Surgery arcs can be performed in any order by Corollary \ref{cor:fixed J edge}. We assume that the depicted pair of arcs correspond to consecutive edges in $P$. Direct calculations verify that they all yield the zero map.
\end{proof}
\subsection{The $E$ and $F$ maps for connected arc diagrams}
\label{sec:The $E$ and $F$ maps for connected arc diagrams}
In this section we analyze the path moduli spaces $\pathPts{P}\MspJxy$ for connected arc diagrams when $J=E,F$ and show that the thick embeddings \eqref{eq:product cube ambient space} in such cases are in fact thin. (We will see in Section \ref{sec:m-1 to m J=E disconn} how genuinely thick embeddings appear when the arc diagrams are disconnected.) The analysis is stated only for the $E$ map, but the $F$ map is nearly identical, see Lemma \ref{lem:E and F duality}; in particular all results here hold with $F$ and $E$ interchanged, though the proofs may need slight modifications.
Fix an annular link diagram $D$ with $n$-crossings, so that together with the $E$ map we have an $(n+1)$-dimensional cube $\mathcal{Q}$. We also fix a sub-cube $u\leq_m v$. For the remainder of this section, we assume that the sub-cube $u\leq_m v$ contains the $E$ edge and that the arc diagram $D_u$ is connected. Our first lemma rules out a very specific arrangement of arcs for such a diagram.
\begin{lemma}\label{lem:forbidden E configuration}
Suppose the arc diagram $D_u$ is connected. Let $P$ be a path through $u\leq_m v$ such that the $E$-resolution $D_e$ contains only essential circles. If the configuration shown in \eqref{eq:forbidden E configuration} appears for some pair of arcs in the arc diagram $D_{e_0}$, where the two horizontal intervals are segments of distinct essential circles, then $\pathPts{P}\MspExy = \varnothing$ for every $x\in \KG{u}, y\in \KG{v}$.
\begin{equation}
\label{eq:forbidden E configuration}
\begin{aligned}
\includestandalone{forbidden_E_configuration}
\end{aligned}
\end{equation}
\end{lemma}
\begin{proof}
We consider two cases. If there are no other essential circles in $D_e$, then by Corollary \ref{cor:fixed J edge} we may assume that the past arc occurs directly before the $E$ map, and the future arc occurs directly after. It is straightforward to verify that the configuration results in empty moduli spaces. If there are other essential circles, then since the arc diagram $D_e$ is connected, at least one of the two depicted circles must be connected to a third circle. Then one of Figure \ref{fig:forbidden5} or Figure \ref{fig:forbidden6} appears in the arc diagram, so we are done by earlier considerations.
\end{proof}
We now present the main technical result of this section which concerns the thinness of path moduli spaces for connected arc diagrams.
\begin{proposition}\label{prop:E/F conn is thin}
Suppose $P$ is a path through a subcube $\scube$ having connected arc diagram $D_u$. Then for any fixed generators $x\in\KG{u},y\in\KG{v},$ the thick embedding $\varphi_P$ from \eqref{eq:product cube ambient space} is in fact a thin embedding
\[\pathPts{P}{\MspExy}\xhookrightarrow{\varphi_P} (0,1)\cong \AmbS.\]
\end{proposition}
\begin{proof}
If the path moduli space $\pathPts{P}\MspExy$ is empty, then the statement is trivially true. Otherwise, by Lemma \ref{lem:biject to all ess}, we can assume that our $E$-resolution consists entirely of essential circles. We fix two points $z,z'\in \pathPts{P} \MspExy$ and seek out to prove that, if $\varphi_P(z)=\varphi_P(z'),$ then in fact we must have had $z=z'$.
We write $z,z'$ as
\begin{align*}
z &= (x \to z_1 \to \cdots \to z_i \rar{j} z_{i+1} \to \cdots \to z_{m-1} \to y)\\
z' &= (x \to z'_1 \to \cdots \to z'_i \rar{j'} z'_{i+1} \to \cdots \to z_{m-1}' \to y),
\end{align*}
where $z_i, z_i'\in \KG{e_0}$ are the generators that $E$ acts on. Then by definition of $\varphi_P$, we must have $j=j'$. Letting $C$ denote the $j$-th essential circle in $D_e$, this implies that the label on $C$ must be $v_-$ in $z_i$ and $v_+$ in $z_{i+1}$. Based on the assumption that the arc diagram is connected and using Lemmas \ref{lem:forbidden configs}, \ref{lem:forbidden E configuration}, one observes that the curves are connected by past/future arcs in an alternating manner as illustrated in the following figure (also see Figure (\ref{eq:alternating arcs})).
The labels discussed above then determine the labels on the circles adjacent to $C$ as shown below.
\begin{equation*}
\includestandalone{propagation}
\end{equation*}
By propagating this argument we see that the labels on all essential circles in $D_e$ are determined by $j$, and thus for $j=j'$ we must have $z_i=z'_i$ and $z_{i+1}=z'_{i+1}$ as well.
From this we can conclude that both $z$ and $z'$ belong to the same composition moduli space
\[z,z'\in \pathPts{P_{\geq e}}{\MspE{y}{z_{i+1}}} \times \MspE{z_{i+1}}{z_i} \times \pathPts{P_{\leq e}}{\MspE{z_i}{x}}.\]
Then since $D_e$ contains only essential circles, Lemma \ref{lem:genus considerations} ensures that both $\pathPts{P_{\leq e}}{\MspE{z_i}{x}}$ and $\pathPts{P_{\geq e}}{\MspE{y}{z_{i+1}}}$ are at most single point moduli spaces, while $\MspE{z_{i+1}}{z_i}$ is an edge moduli space and so contains at most a single point as indicated in Section \ref{sec:0-dim moduli spaces}. Thus we have $z=z'$ as desired, and the embedding $\varphi_P$ must in fact be thin.
\end{proof}
Although Proposition \ref{prop:E/F conn is thin} suffices to continue our construction, we can actually prove a stronger statement about our path moduli spaces which makes them easier to understand in concrete examples. We begin with a special case.
\begin{lemma}\label{lem:E/F conn is one point per circle preliminary}
Suppose $P$ is a path through a subcube $u\leq_m v$ with (connected) $E$-resolution arc diagram $D_e$ consisting entirely of $s$ essential circles, with precisely one arc connecting each consecutive pair. Fix $x\in \KG{u}, y\in \KG{v}$ such that the moduli space $\pathPts{P}{\MspExy}$ is nonempty. Then for each $1\leq j \leq s$, there is at least one point of the form
\[ (z^j):= (x\rightarrow\cdots\rightarrow z_i \xrightarrow{j} z_{i+1} \rightarrow \cdots \rightarrow y)\in\pathPts{P}{\MspExy}.\]
\end{lemma}
\begin{proof}
Let $S_{\leq e} \: D_u \to D_{e}$, $S_{\geq e} \: D_e\to D_v$ denote the cobordisms obtained by composing the saddle maps in $P$ before the $E$ map and after the $E$ map, respectively.
Fixing $1\leq j \leq s$, we first note that the $j$-th essential circle in $D_e$ must be labeled by $v_-$ in $z_i$. Using the argument in Proposition \ref{prop:E/F conn is thin}, there is a unique choice of generators $z_i,z_{i+1}$ satisfying this condition. To construct $(z^j)$, we will show that $\mathcal{F}_\mathbb{A}(S_{\geq e})(z_{i+1})$ is a generator of $\Fa(D_v)$, which we denote $y'$. Then we show there exists a unique $x'\in \KG{u}$ such that $z_i$ appears in $\mathcal{F}_\mathbb{A}(S_{\leq e})(x')$. Finally we will show that $x'=x$ and $y'=y$, completing the proof.
By Lemma \ref{lem:forbidden configs}, Figures \ref{fig:forbidden5} and \ref{fig:forbidden6}, the arc diagram for $D_e$ consists of $s$ essential circles connected by past and future arcs in an alternating manner,
\begin{equation}\label{eq:alternating arcs}
\begin{aligned}
\includestandalone{alternating_arcs}
\end{aligned}
\end{equation}
Let us first show that $z_{i+1}$ is not killed by applying the future arcs in $D_e$. Note that each circle in $D_e$ is incident to at most one future arc. Pick circles $C_1, C_2$ in $D_e$ that are connected by a future arc $a$. Surgery along $a$ results in a trivial circle $C'$, and by construction of $z_{i+1}$ the labels on $C_1$ and $C_2$ are different. Therefore the map assigned to $a$ is nonzero on $z_{i+1}$, and moreover $C'$ is labeled by $X$ in the image of $z_{i+1}$. Repeating this for all future arcs, we see that $y':=\mathcal{F}_\mathbb{A}(S_{\geq e})(z_{i+1})$ is a generator of $D_v$. Each trivial circle in $D_v$ is labeled $X$ in $y'$. Essential circles in $D_v$ are those with no future arc incident on them in $D_e$, and they have the same label as in $z_{i+1}$.
Let us now define $x'\in \KG{u}$ such that $z_i$ appears in $\mathcal{F}_\mathbb{A}(S_{\leq e})(x')$. As before, each circle in $D_e$ is incident to at most one past arc. If $a'$ is a past surgery arc joining the essential circles $C_1'$ and $C_2'$ in $D_e$, then undoing $a'$ we see a trivial circle $T$ splitting into the two essential circles $C_1' \cup C_2'$,
\begin{equation*}
\begin{tikzpicture}[yscale=.75]
\coordinate (p1) at (0,0);
\coordinate (p2) at (.5,0);
\coordinate (p3) at (0,1);
\coordinate (p4) at (.5,1);
\coordinate (p1') at (2,0);
\coordinate (p2') at (2.5,0);
\coordinate (p3') at (2,1);
\coordinate (p4') at (2.5,1);
\draw[line width=.4mm] (p1) -- (p2);
\draw[line width=.4mm] (p3) -- (p4);
\draw[line width=.4mm] (p2) arc[start angle=-90, end angle=90, radius=.5];
\draw[line width=.4mm] (p1') -- (p2');
\draw[line width=.4mm] (p3') -- (p4');
\draw[line width=.4mm] (p3') arc[start angle=90, end angle=270, radius=.5];
\coordinate (l1) at (4.5,0);
\coordinate (r1) at (7,0);
\coordinate (l2) at (4.5,1);
\coordinate (r2) at (7,1);
\draw[line width=.4mm] (l1) -- (r1);
\draw[line width=.4mm] (l2) -- (r2);
\node[draw=none] at (3.5,.5) {$\longrightarrow$};
\coordinate (l1') at (2.25,2);
\coordinate (r1') at (4.75,2);
\coordinate (l2') at (2.25,3);
\coordinate (r2') at (4.75,3);
\coordinate (m1) at (3.5,2);
\coordinate (m2) at (3.5,3);
\draw[line width=.4mm] (l1') -- (r1');
\draw[line width=.4mm] (l2') -- (r2');
\draw[line width=1mm, dashed, blue] (m1) -- (m2);
\node[draw=none, anchor = north] at (m1) {$a'$};
\node[draw=none, anchor = west] at (r1) {$C_2'$};
\node[draw=none, anchor = west] at (r2) {$C_1'$};
\node[draw=none, anchor = east] at (p1) {$T$};
\end{tikzpicture}
\end{equation*}
We must label $T$ by the generator $1$. By construction, $C_1'$ and $C_2'$ have opposite labels in $z_i$. The saddle map corresponding to $a'$ sends $1\in \Fa(T)$ to a sum of two generators in $\Fa(C_1' \cup C_2')$, one of which matches the labels on $C_1'\cup C_2'$ in $z_i$. Repeat this for all past arcs in $D_e$ to obtain the desired generator $x'\in \KG{u}$. Note that trivial circles in $D_u$ are labeled by $1$ in $x'$. Essential circles in $D_u$ correspond to essential circles in $D_e$ with no past arc incident on them, and they have the same label in $x'$ and $z_i$.
It remains to show that $x'=x$ and $y'=y$. By the above analysis, $x'$ and $x$ are labeled $1$ on all trivial circles in $D_u$, and likewise $y'$ and $y$ are labeled $X$ on all trivial circles in $D_v$. We will argue that labels on essential circles in $D_u$ and $D_v$ are determined by the arc diagram $D_e$.
Let $r$ and $t$ denote the number of essential circles in $D_u$ and $D_v$, respectively. Since the arc diagram $D_e$ is connected, Lemma \ref{lem:forbidden configs}, Figures \ref{fig:forbidden5} and \ref{fig:forbidden6} imply that all but the first and last essential circles in $D_e$ are connected to adjacent essential circles by both a past and future arc. It follows that the pair $(r,t)$ is one of $(2,0)$, $(1,1)$, or $(0,2)$, which can be read off from the type of arcs incident on the first and last circle in $D_e$,
\begin{equation*}
\includestandalone{first_and_last_circles}
\end{equation*}
Finally, note that if $x'' \in \KG{u}$, $y''\in \KG{v}$ such that $\pathPts{P}{\MspE{y''}{x''}} \neq \varnothing$, then $\operatorname{adeg}(y'') = \operatorname{adeg}(x'') + 2$. For each of the three possibilities for $(r,t)$, we see that the labels on essential circles in $D_u$ and $D_v$ are always uniquely determined whenever the moduli space is nonempty. Therefore $x=x'$ and $y=y'$.
\end{proof}
With Lemma \ref{lem:E/F conn is one point per circle preliminary} in place, we can state and prove the following proposition which characterizes all non-empty path moduli spaces in any subcube for a connected diagram which involves the $E$ map.
\begin{proposition}\label{prop:E/F conn is one point per circle}
Suppose $P$ is a path through a subcube $\scube$ having connected arc diagram $D_u$, such that the $E$-resolution $D_e$ of $P$ contains $s$ essential circles. Then for any fixed generators $x\in\KG{u},y\in\KG{v}$, if the path moduli space $\pathPts{P}{\MspExy}$ is non-empty, then the thick embedding $\varphi_P$ from \eqref{eq:product cube ambient space} is a thin embedding mapping $\pathPts{P}{\MspExy}$ bijectively onto $\{ \frac{1}{s+1}, \ldots, \frac{s}{s+1} \} \subset (0,1)$.
\end{proposition}
\begin{proof}
Thinness of the embedding $\pathPts{P}{\MspExy}\xhookrightarrow{\varphi_P} (0,1)$ was established in Proposition \ref{prop:E/F conn is thin}. By Lemma \ref{lem:biject to all ess}, we can assume that our $E$-resolution consists entirely of $s$ essential circles. It thus remains to show that each essential circle in $D_e$ contributes at least one point to $\pathPts{P}{\MspExy}$.
Fix $1\leq j \leq s$. By Corollary \ref{cor:fixed J edge}, the order in which the surgery arcs in the sub-cubes $u\leq e_0$ and $e_1 \leq v$ are performed is irrelevant, and we use this implicitly.
Pick future surgery arcs $a_1, \ldots, a_f$ and past surgery arcs $b_1, \ldots, b_p$ in $D_e$, such that consecutive circles in $D_e$ are connected by exactly one of the chosen arcs. We assume that in the path $P$, the $f$ edges directly after the $E$ edge correspond to performing surgery on $a_1, \ldots, a_f$, and the $p$ edges directly before the $E$ edge correspond to surgery on the dual arcs to $b_1, \ldots, b_p$. In other words, $P$ is of the form
\[
u \to \cdots \to u_0 \overset{b_1}{\to} u_1 \overset{b_2}{\to} \cdots \overset{b_p}{\to} e_0 \overset{E}{\to} e_1 \overset{a_1}{\to} v_1 \overset{a_2}{\to} \cdots \overset{a_f}{\to} v_f \to \cdots \to v.
\]
Consider the sub-cube $u_0\leq v_f$ and the path $P^2$ through this sub-cube, consisting of all edges in $P$ from $u_0$ to $v_f$. Let $D_{u_0}'$ denote the diagram obtained from $D_{u_0}$ by including only arcs in $P^2$. The $E$-resolution of $P^2$ is connected by construction, so $D_{u_0}'$ is connected as well. Let $P^1$ and $P^3$ denote the sub-paths of $P$ consisting of edges from $u$ to $u_0$ and $v_f$ to $v$, respectively.
\[
\lefteqn{\overbrace{\phantom{u \to \cdots \to u_0}}^{P^1}}u \to \cdots \to
\underbrace{u_0 \overset{b_1}{\to} u_1 \overset{b_2}{\to} \cdots \overset{b_p}{\to} e_0 \overset{E}{\to} e_1 \overset{a_1}{\to} v_1 \overset{a_2}{\to} \cdots \overset{a_f}{\to}
\lefteqn{\overbrace{\phantom{ v_f \to \cdots \to v}}^{P^3}}v_f}_{P^2} \to \cdots \to v
\]
Since $\pathPts{P}{\MspExy}$ is nonempty, there exist generators $x'\in \KG{u_0}$, $y'\in \KG{v_f}$ such that the moduli spaces
\[
\pathPts{P^1}\MspE{x'}{x},\; \pathPts{P^2}{\MspE{y'}{x'}},\; \pathPts{P^3} \MspE{y}{y'}
\]
are all nonempty. By Lemma \ref{lem:E/F conn is one point per circle preliminary} applied to $D_{u_0}'$ and $P^2$, we can find $z^2\in \pathPts{P^2}{\MspE{y'}{x'}}$ of the form
\[z^2 = (x' \rightarrow \cdots \rightarrow z_i \xrightarrow{j} z_{i+1} \rightarrow \cdots \rightarrow y')\in \pathPts{P^2}{\MspE{y'}{x'}}.\]
Then choosing any $z^1 \in \pathPts{P^1}\MspE{x'}{x}$ and $z^3 \in \pathPts{P^3} \MspE{y}{y'}$ gives rise to at least one point in the composition moduli space
\[z=( z^3,z^2,z^1 ) \in \pathPts{P^3}{\MspE{y}{y'}} \times \pathPts{P^2}{\MspE{y'}{x'}} \times \pathPts{P^1}{\MspE{x'}{x}} \subset \pathPts{P}{\MspExy}\]
which $\varphi_P$ maps to the $j$-th point in $(0,1)$ as desired.
\end{proof}
\subsection{The $H$ map for connected arc diagrams}
As in the previous section, we continue to fix a subcube $u\leq_m v$ involving $J=H$ such that the initial arc diagram $D_u$ is connected (and hence all intermediate diagrams $D_w$ for $u\leq w \leq v$ are connected as well). The main technical result is a characterization of the path moduli spaces in $u\leq_m v$, Proposition \ref{prop:conn H}. We begin with several lemmas.
\begin{lemma}
\label{lem:conn is thin H}
Let $P$ be a path through $u\leq_m v$ such that the $H$-resolution $D_e$ consists entirely of $s$ essential circles. Fix generators $x\in \KG{u}$, $y\in \KG{v}$ such that $\pathPts{P}{\ModSp_H(y,x)}$ is nonempty. If every pair of consecutive circles in $D_e$ is joined by exactly one type of arc (either past or future), then the thick embedding $\varphi_P$ from \eqref{eq:product cube ambient space} is in fact thin.
\end{lemma}
\begin{proof}
Let $z,z'\in \pathPts{P}{\MspHxy}$ with
\begin{align*}
z &= (x\rightarrow\cdots\rightarrow z_i \xrightarrow{j} z_{i+1} \rightarrow \cdots \to y),\\
z' &= (x\rightarrow\cdots\rightarrow z_i' \xrightarrow{j} z_{i+1}' \rightarrow \cdots \to y),
\end{align*}
for some $1\leq j \leq s$. Let $C$ denote the $j$-th circle in $D_e$. We will first show that $z_i = z_i'$ (and thus $z_{i+1} = z_{i+1}'$ as well).
Choosing a label of $v_{\pm}$ on $C$ uniquely determines the generator $z_i$ (and thus $z_{i+1}$), as in the proof of Proposition \ref{prop:E/F conn is thin}. Therefore, in order to have $z_i\neq z_i'$, we must have the labels on $C$ in $z_i$ and $z_i'$ be opposites. It follows that $z_i$ is obtained from $z_i'$ by swapping the labels on all circles in $D_e$, and likewise for $z_{i+1}$ and $z'_{i+1}$.
On the other hand, Lemma \ref{lem:forbidden configs} (Figures \ref{fig:forbidden5} and \ref{fig:forbidden6}) implies that the circles in $D_e$ are connected by past and future arcs in an alternating manner as in \eqref{eq:alternating arcs},
where there may be more than one past or future arc between any two adjacent circles. Since the arc diagram $D_e$ is connected, all but the first and last essential circles in $D_e$ are connected to adjacent essential circles by both a past and future arc. Therefore at least one of $D_u$ or $D_v$ has an essential circle. Moreover, the labels on the essential circle(s) in $x$ (resp. $y$) are the same as the labels on the corresponding essential circle(s) in $z_i$ (resp. $z_{i+1}$). Since $x$ and $y$ are fixed, at least one of these prevents the circles in $D_e$ from having opposite labels for either $z_i$ and $z_i'$, or $z_{i+1}$ and $z_{i+1}'$.
Thus we must have that $z_i=z_i'$ and $z_{i+1}=z'_{i+1}$, and the lemma follows by the same argument as in the proof of Proposition \ref{prop:E/F conn is thin}.
\end{proof}
Lemma \ref{lem:conn is thin H} does not handle all cases where $D_e$ consists of only essential circles. After ruling out the forbidden configurations Figure \ref{fig:forbidden5} and Figure \ref{fig:forbidden6} from Lemma \ref{lem:forbidden configs}, it remains to analyze the situation where $D_e$ consists of exactly two essential circles with both a past and future arc connecting them, as in \eqref{eq:forbidden E configuration}. The following lemma gives an explicit description of such a path moduli space and its thick embedding.
\begin{lemma}
\label{lem:conn is thick H outlier}
Let $P$ be a path through $u\leq_m v$ such that the $H$-resolution $D_e$ consists entirely of two essential circles, with at least one past and one future arc connecting them. Fix generators $x\in \KG{u}$, $y\in \KG{v}$ such that $\pathPts{P}{\ModSp_H(y,x)}$ is nonempty. Then $\pathPts{P}{\MspHxy}$ consists of four points, and the thick embedding from \eqref{eq:product cube ambient space} is a two-to-one map onto the points $\left\{\frac 13,\frac 23\right\} \subset (0,1) \cong \AmbS$. Moreover, any two points in $\pathPts{P}{\MspHxy}$ that map to the same point in $\AmbS$ are oppositely framed.
\end{lemma}
\begin{proof}
Let $a_p$ and $a_f$ denote the saddles corresponding to one past and one future surgery arc in $D_e$, respectively. By Corollary \ref{cor:fixed J edge}, we may assume that $a_p$ and $a_f$ are performed directly before and after the $H$ edge, respectively. Let $P'$ denote the 3-edge subpath of $P$ corresponding to $a_p$, $H$, and $a_f$, and note that the diagrams at the start and end of this path both consist of a single trivial circle. Let $P_{\leq p}$ (resp. $P_{\geq f}$) denote the subpath consisting of the edges before $a_p$ (resp. after $a_f$).
It is a straightforward calculation to check that, in order for $\pathPts{P'}{\MspH{z_f}{z_p}}$ to be non-empty, the generator $z_p$ (resp. $z_f$) must consist of a label $1$ (resp. $X$) on the trivial circle at the start (resp. end) of the path $P'$. In this unique case one further computes that $\pathPts{P'}{\MspH{z_f}{z_p}}$ consists of four points thickly embedding into $(0,1)$ as in the statement of the lemma.
Then for the full path moduli space we have
\[\pathPts{P}{\MspHxy} \cong \pathPts{P_{\geq f}}{\MspH{y}{z_f}} \times \pathPts{P'}{\MspH{z_f}{z_p}} \times
\pathPts{P_{\leq p}}{\MspH{z_p}{x}}.\]
The cobordism corresponding to the path $P_{\leq p}$ (resp. $P_{\geq f}$) must have genus zero, otherwise Lemma \ref{lem:genus considerations} would cause the path moduli space to be empty after composing with $a_p$ (resp. $a_f$) and connecting to $D_e$ which consists of all essential circles. Thus both $\pathPts{P_{\leq p}}{\MspH{z_p}{x}}$ and $\pathPts{P_{\geq f}}{\MspH{y}{z_f}}$ consist of a single trivially embedded point and we are done.
\end{proof}
\begin{remark}
The two past and future arcs in Lemma \ref{lem:conn is thick H outlier} form a ladybug configuration (\cite[Figure 5.1]{LS}). Lemma \ref{lem:forbidden E configuration} disallows this configuration for $J=E, F$, and it is the only case among connected arc diagrams for which the thick embedding \eqref{eq:product cube ambient space} is not thin. The configuration in Lemma \ref{lem:forbidden configs}, Figure \ref{fig:forbidden1} is also a ladybug, which is disallowed for all $J$.
\end{remark}
The following important lemma concerns the planar topology of an arc diagram $D_e$ as in the hypothesis of Lemma \ref{lem:conn is thick H outlier}.
\begin{lemma}
\label{lem:ray}
Let $P$ be a path through $u\leq_m v$ such that the $H$-resolution $D_e$ consists entirely of two essential circles, with at least one past and one future arc connecting them. Let $x\in \KG{u}$, $y\in \KG{v}$ be generators such that $\pathPts{P}{\ModSp_H(y,x)}$ is nonempty. Then there exists a ray from the puncture $\times$ to the point at infinity which is disjoint from all the arcs in $D_e$ and which intersects the circles in $D_e$ in exactly two points.
\end{lemma}
\begin{proof}
Let $a_p$ and $a_f$ denote two past and future arcs, respectively, which join the two essential circles in $D_e$. We assume that $a_p$ and $a_f$ are adjacent, in the sense that they can be connected by an interval whose interior is disjoint from the arc diagram $D_e$. We will show that there exists such a ray passing between them, as in Figure \ref{fig:cutting H configuration}. Indeed, the only possible obstruction to such a ray would be an arc, or a sequence of interlacing arcs, along one of the essential circles with endpoints `outside' of the arcs $a_p,a_f$ (see Figure \ref{fig:forbidden H configuration} with single arc case and Figure \ref{fig:forbidden linking configuration} for the interlacing arc case). By considering the forbidden configurations of Figures \ref{fig:forbidden1} and \ref{fig:forbidden2}, we see that the arcs on one side of the essential circle must be either all past or all future. In either case, we will arrive at another forbidden configuration as in Figure \ref{fig:forbidden3} or Figure \ref{fig:forbidden4}. Thus no such arrangement may exist and our ray can be drawn.
\begin{figure}
\hspace*{\fill}
\begin{subfigure}[t]{.2\textwidth}
\begin{center}
\includestandalone{cutting_H_configuration}
\end{center}
\caption{The desired ray}\label{fig:cutting H configuration}
\end{subfigure}
\hfill
\begin{subfigure}[t]{.2\textwidth}
\begin{center}
\includestandalone{forbidden_configurations_for_H}
\end{center}
\caption{One obstructing arc}\label{fig:forbidden H configuration}
\end{subfigure}
\hfill
\begin{subfigure}[t]{.5\textwidth}
\begin{center}
\includestandalone{forbidden_linking_configuration}
\end{center}
\caption{An obstructing sequence of interlaced arcs}\label{fig:forbidden linking configuration}
\end{subfigure}
\hspace*{\fill}
\caption{}\label{fig:ray obstruction}
\end{figure}
\end{proof}
We are now ready for the main result in this section.
\begin{proposition}
\label{prop:conn H}
Consider a subcube $u\leq_m v$ such that the arc diagram $D_u$ is connected, and fix generators $x\in \KG{u}, y\in \KG{v}$. Then one of the following two holds.
\begin{enumerate}
\item
\label{item:H path is thin}
For any path $P$ from $u$ to $v$ such that $\pathPts{P}\MspHxy \neq \varnothing$, the thick embedding from \eqref{eq:product cube ambient space} is in fact thin.
\item
\label{item:H path is thick}
For any path $P$ from $u$ to $v$ such that $\pathPts{P}\MspHxy \neq \varnothing$, the moduli space $\pathPts{P}{\MspHxy}$ consists of four points, and the thick embedding from \eqref{eq:product cube ambient space} is a two-to-one map onto the points $\left\{\frac 13,\frac 23\right\} \subset (0,1) \cong \AmbS$. Moreover, any two points in $\pathPts{P}{\MspHxy}$ that map to the same point in $\AmbS$ are oppositely framed.
\end{enumerate}
\end{proposition}
\begin{proof}
If all paths yield empty moduli spaces then there is nothing to show. Take a path $P$ from $u$ to $v$ with $\pathPts{P}\MspHxy \neq \varnothing$. By Lemma \ref{lem:biject to all ess}, we may assume that the $H$-resolution $D_e$ of $P$ consists entirely of essential circles. We proceed by considering two cases for this fixed $P$.
Suppose first that $D_e$ contains a pair of essential circles which are joined by both a past and future arc. We will argue that case (2) of the lemma holds. By Lemma \ref{lem:forbidden configs}, Figures \ref{fig:forbidden5} and \ref{fig:forbidden6}, we know that $D_e$ contains exactly two essential circles. It was shown in Lemma \ref{lem:conn is thick H outlier} that the conclusion of case (2) holds for $\pathPts{P} \MspHxy$, so it remains to show that the same holds for every path from $u$ to $v$. To that end, let $P'$ be another path from $u$ to $v$ with nonempty moduli space, and denote its $H$-resolution by $D_{e'}$. As usual, it suffices to assume that $D_{e'}$ contains only essential circles. The arc diagram $D_{e'}$ can be obtained from $D_e$ by performing a sequence of surgeries along the arcs in $D_e$. Then Lemma \ref{lem:ray} and the assumption that $\pathPts{P'}\MspHxy \neq \varnothing$ guarantees that $D_{e'}$ contains two essential circles. Moreover, the ray constructed in the proof of Lemma \ref{lem:ray}, along with the forbidden configurations of Lemma \ref{fig:forbidden configs}, Figures \ref{fig:forbidden1} and \ref{fig:forbidden2}, imply that the two circles in $D_{e'}$ are joined by both a past and a future arc. Thus (2) holds by Lemma \ref{lem:conn is thick H outlier}.
Now suppose that every pair of consecutive circles in $D_e$ are joined by exactly one type (either past or future) of arc. The above argument guarantees that the $H$-resolution of every path with nonempty moduli space also has the property that each pair of consecutive circles are joined by exactly one type of arc. Then we conclude that case (1) holds by Lemma \ref{lem:conn is thin H}.
\end{proof}
Proposition \ref{prop:conn H} suffices to continue our construction. However, as in Section \ref{sec:The $E$ and $F$ maps for connected arc diagrams}, we can completely describe the thickly embedded path moduli spaces for $J=H$. In Case \eqref{item:H path is thick} above, the moduli spaces are already described completely. Meanwhile, for Case \eqref{item:H path is thin}, we have the following proposition; compare with Proposition \ref{prop:E/F conn is one point per circle}.
\begin{proposition}
\label{prop:H conn bijection}
Let $x\in \KG{u}$, $y\in \KG{v}$ such that we are in Case \eqref{item:H path is thin} of Proposition \ref{prop:conn H}. Then for any path $P$ with $\de_P\MspHxy\neq\varnothing$, the thin embedding $\varphi_P$ from \eqref{eq:product cube ambient space} surjects onto $\left\{\frac{1}{s+1}, \ldots, \frac{s}{s+1}\right\}$ where $s$ denotes the number of essential circles in $D_e$.
\end{proposition}
\begin{proof}
Pick an element
\[
(x\rightarrow\cdots\rightarrow z_i \xrightarrow{j} z_{i+1} \rightarrow \cdots \to y) \in \pathPts{P}\MspHxy
\]
for some $1\leq j \leq s$. Since $H$ does not change the labels on essential circles, there must be an element of the form
\[
(x\rightarrow\cdots\rightarrow z_i \xrightarrow{k} z_{i+1} \rightarrow \cdots \to y) \in \pathPts{P}\MspHxy
\]
for every $1\leq k \leq s$.
\end{proof}
\section{Constructing the higher dimensional moduli spaces (the inductive step)}
\label{sec:m-1 to m}
In order to complete the proof of Theorem \ref{thm:ffcJ}, we induct on the dimension of the subcube being considered. That is to say, we fix $m>2$ and inductively assume that, for all subcubes $\subcube{i}{b}{a}$ with $i\leq m-1$ and generators $r\in\KG{a},s\in\KG{b}$, we have constructed moduli spaces $\MspJ{s}{r}$ together with thick embeddings
\[\MspJ{s}{r}\thickemb \AmbSp{b}{a}\]
satisfying all relevant boundary conditions. For any $\subcube{i}{b}{a}$ that does not involve $J$, this thick embedding is just
\[\MspJ{s}{r} \cong \coprod \cubeMod{b}{a} \xthickemb{\coprod \operatorname{id}} \cubeMod{b}{a} = \AmbSp{b}{a}.\]
Meanwhile, for any $\subcube{i}{b}{a}$ that does involve $J$, the thick embedding is of codimension $1$ due to the extra factor of $(0,1)$ in the definition of $\AmbSp{b}{a}$.
Now we fix an $m$-dimensional subcube $\scube$ and seek to build new moduli spaces $\MspJxy$ thickly embedding into $\AmbS$ for various $x\in\KG{u},y\in\KG{v}$. In the case that $\scube$ does not involve $J$, none of the subfaces of $\scube$ involve $J$ either and the entire construction follows from \cite{LS}. In this case, the ``trivial cover" language of \cite{LS} gives a thickened identity embedding
\[\MspJxy \cong \coprod \cMod \xthickemb{\coprod \operatorname{id}} \cMod = \AmbS\]
and we are done. Otherwise, the subcube $\scube$ involves $J$ and we split into cases.
In each case, the essential strategy is the same. The boundary of $\MspJxy$ is made up of pieces of various codimensional pieces $\de_{[i]}\MspJxy$, each of which consists of products of lower dimensional moduli spaces, which we inductively assume that we have already constructed and thickly embedded into the relevant ambient spaces. When we assemble these pieces together, we have a thick embedding of $\de\MspJxy$ into $\de\AmbS$. Completing the inductive step is then equivalent to proving that this thick embedding can be `thickly filled' by a moduli space embedding $\MspJxy\thickemb\AmbS$ that satisfies Equation \eqref{eq:respecting boundary structure}.
\subsection{Subcubes involving $E$ or $F$ with connected arc diagrams}
\label{sec:m-1 to m J=E conn}
In the case that $J=E$ (the case $J=F$ is again dual to this case and left to the reader), we begin by considering subcubes $\scube$ whose corresponding arc diagram $D_u$ is connected. We fix generators $x\in\KG{u},y\in\KG{v}$ and consider the codimension $(m-1)$ (i.e. dimension $0$) part of the boundary
\[\de_{[m-1]}\MspExy = \coprod_{\text{paths $P$}} \pathPts{P}{\MspExy}.
\]
If all of these path moduli spaces are empty, then we define $\MspExy:=\varnothing$. Otherwise, Proposition \ref{prop:E/F conn is thin} shows that any non-empty $\pathPts{P}{\MspExy}$ is \emph{thinly} embedded into $\pAmb\cong (0,1)$.
Now consider $\de_{[m-2]}\MspExy$, which Theorem \ref{thm:ffcJ} demands is of the form
\[
\de_{[m-2]} \MspExy = \bigcup \left( \coprod_{\substack{r\in\KG{a} \\s\in\KG{b}}} \pathPts{P^b}{\MspE{y}{s}} \times \MspE{s}{r} \times \pathPts{P^a}{\MspE{r}{x}} \right),
\]
where the outer union is taken over
\begin{itemize}
\item all square sub-faces $\subcube{2}{b}{a}$ in the subcube $\scube$, and
\item all sub-paths $P^a$ from $u$ to $a$ and $P^b$ from $b$ to $v$.
\end{itemize}
We will denote the parenthetical term by $\sqMspExy$. Then our (inductively defined) lower dimensional moduli space embeddings determine a thick product embedding
\begin{equation}\label{eq:E/F square face embedding}
\sqMspExy \thickemb \pathAmb{P^b} \times \AmbSp{b}{a} \times \pathAmb{P^a} \cong [0,1]\times(0,1)
\end{equation}
with boundary conditions based on the embeddings of the two path moduli spaces corresponding to composing $P^a$ and $P^b$ with the two paths around the edges of the square face $\subcube{2}{b}{a}$.
\begin{equation}
\begin{aligned}
\begin{tikzpicture}[commutative diagrams/every diagram, row sep = 1.4ex]
\matrix[matrix of math nodes, name=m, commutative diagrams/every cell]
{ & & & c & & & \\
u & \cdots & a & & b &\cdots & v\\
& & & c' & & & \\
};
\path[commutative diagrams/.cd,every arrow,every label]
(m-2-3) edge (m-1-4)
(m-2-3) edge (m-3-4)
(m-1-4) edge (m-2-5)
(m-3-4) edge (m-2-5)
(m-2-1) edge (m-2-2) (m-2-2) edge (m-2-3) (m-2-5) edge (m-2-6) (m-2-6) edge (m-2-7)
;
\draw [
thick,
decoration={
brace,
mirror,
raise=0.5cm
},
decorate
] (-5.1,.9) -- (-1.5,.9)
node [pos=0.5,anchor=north,yshift=-0.55cm] {$P^a$};
\draw [
thick,
decoration={
brace,
mirror,
raise=0.5cm
},
decorate
] (1.5,.9) -- (5.1,.9)
node [pos=0.5,anchor=north,yshift=-0.55cm] {$P^b$};
\end{tikzpicture}
\end{aligned}
\end{equation}
The key point is that, because the path moduli spaces on either boundary of $\sqMspExy$ are \emph{thinly} embedded for our connected $D_u$, we can conclude that our composition moduli space embeddings \eqref{eq:E/F square face embedding} are thin as well. Thus we see that all of $\de_{[m-2]}\MspExy$ is thinly embedded, and we move on to consider $\de_{[m-3]}\MspExy$. This is again built from products of lower dimensional moduli spaces that we inductively assume have been constructed and thickly embedded in such a way that their boundaries are built out of parts of the thinly embedded $\de_{[m-2]}\MspExy$ - and thus $\de_{[m-3]}\MspExy$ must actually have been thinly embedded also.
This reasoning continues all the way until we consider the thinly embedded $\de_{[1]}\MspExy=\de\MspExy$. This is an $(m-2)$-dimensional manifold with corners $\de\MspExy$ thinly embedded into the $(m-1)$-dimensional ambient space boundary $\de\AmbS\cong S^{m-2}\times (0,1)$. We may view this as a codimension-1 framed embedding $\de\MspExy\hookrightarrow S^{m-1}$ that misses the poles. Using the Pontryagin-Thom construction, this framed submanifold corresponds to a map $S^{m-1}\longrightarrow S^1$. Since $m>2$, the map is null-homotopic so it extends to a map $D^m\longrightarrow S^1$. Therefore $\de\MspExy\hookrightarrow S^{m-1}$ bounds a framed embedding of an $(m-1)$-dimensional manifold into the interior of $D^m$, which is equivalent to the interior of $D^{m-1}\times (0,1)\cong\AmbS$. It is worth noting that this is the point in the proof that uses (and propagates) the crucial codimension $1$ inductive assumption. This is our construction of $\MspExy$ for connected $D_u$. See Figure \ref{fig:Hexagon times I}.
\begin{figure}[ht]
\includegraphics[height=5cm]{HexagonI5.pdf}
\caption{Examples of 2-dimensional $\MspExy\thickemb\AmbS\cong\ModSp_\mathcal{Q}(v,u)\times(0,1)$, where $\ModSp_\mathcal{Q}(v,u)$ is the 2-dimensional permutohedron (a hexagon, topologically $D^2$). The Pontryagin-Thom construction ensures that our framed embedded $\de\MspExy$ (drawn in red) can be filled as shown, regardless of the presence of turnbacks (framing data is omitted to avoid clutter).}
\label{fig:Hexagon times I}
\end{figure}
\subsection{Subcubes involving $E$ or $F$ with disconnected arc diagrams}
\label{sec:m-1 to m J=E disconn}
Next we consider the situation when $J=E$ and the diagram $D_u$ is disconnected. Unlike in Section \ref{sec:m-1 to m J=E conn}, path moduli spaces for subcubes with disconnected arc diagrams need not be thinly embedded. For example, adding a disjoint, contractible ladybug configuration (\cite[Figure 5.1]{LS}) has the effect of turning a nonempty thinly embedded path moduli space into one where the thick embedding is not injective. We begin with a simple lemma generalizing Lemma \ref{lem:m=2 conn comp and annular deg} from the $m=2$ construction.
\begin{lemma}\label{lem:general E acts on single conn comp}
Suppose there is some path $P$ for which $\pathPts{P}{\MspExy}\neq\varnothing$. Then there is a single connected component $\mathscr{C}$ of the arc diagram $D_u$ such that
\[ \operatorname{adeg}(y|_\mathscr{C})=\operatorname{adeg}(x|_\mathscr{C})+2,\quad \operatorname{adeg}(y|_{\mathscr{C}'})=\operatorname{adeg}(x|_{\mathscr{C}'}),\]
where $\mathscr{C}'$ is the complement of $\mathscr{C}$. We say that $E$ is acting on the component $\mathscr{C}$. When $J=F$, the value $2$ above is replaced by $-2$.
\end{lemma}
\begin{proof}
As in the proof of Lemma \ref{lem:m=2 conn comp and annular deg}, saddles maps preserve annular degree, but the $E$ map increases annular degree by two.
\end{proof}
As before, we consider
\[\MspExy=\MspE{y|_\mathscr{C} \otimes y|_{\mathscr{C}'}}{x|_\mathscr{C} \otimes x|_{\mathscr{C}'}}.\]
We let $\MspECCxy$ denote the moduli space that would be built for the diagram $\mathscr{C}$ for generators $x|_\mathscr{C},y|_\mathscr{C}$ if $\mathscr{C}'$ were not present. We also include the $E$ edge in the cube used to construct $\MspECCxy$, since $E$ is acting on $\mathscr{C}$. Because $\mathscr{C}$ is connected, we can build $\MspECCxy$ using the techniques of Sections \ref{sec:0-dim moduli spaces}, \ref{sec:m=2 J=E/F involved}, and \ref{sec:m-1 to m J=E conn}. Furthermore, if $\mathscr{C}'$ contains no saddles, then $y|_{\mathscr{C}'}=x|_{\mathscr{C}'}$ and we can define
\[\MspExy\cong\MspECCxy\]
as thickly embedded moduli spaces, and so we are done.
Therefore we may assume that $\mathscr{C}'$ contains at least one saddle. In this case we let $\MspCCpxy$ denote the moduli space that would be built for the diagram $\mathscr{C}'$ for generators $x|_{\mathscr{C}'},y|_{\mathscr{C}'}$ if $\mathscr{C}$ were not present (not including the $E$ edge, since $E$ is not acting on $\mathscr{C}'$). Note that, because $E$ is not involved, we can build $\MspCCpxy$ as a trivial cover of a suitable permutohedron using the techniques in \cite{LS}.
Now consider that, in the large subcube $\scube$, there are many $k$-dimensional subfaces $\subcube{k}{b}{a}$ whose edges consist of all of the saddles in $\mathscr{C}'$. Indeed for any path $P$ starting from $u$ which consists of edges coming only from saddles in $\mathscr{C}$ (or the $E$ edge), the ending vertex of $P$ can be taken as the starting vertex $a$ of such a subface.
\begin{lemma}\label{lem:J=E subcube for the complement}
Let $k>0$ denote the number of saddles in $\mathscr{C}'$ and let $\subcube{k}{b}{a}$ denote \emph{any} subface of the subcube $\scube$ consisting of only edges corresponding to these $k$ saddles. Then for any choice of generator $z$ on $\mathscr{C}\subset D_a$, we have a commuting diagram
\begin{equation}
\begin{aligned}
\label{eq:J=E framed bij for the complement}
\begin{tikzpicture}[x=1in,y=-.5in]
\node(A) at (0,0) {$\MspE{z\otimes (y|_{\mathscr{C}'})}{z\otimes (x|_{\mathscr{C}'})}$};
\node(B) at (0,2) {$\coprod \cubeMod{b}{a}$};
\node(C) at (2,0) {$\AmbSp{b}{a}$};
\node(D) at (2,2) {$\cubeMod{b}{a}$};
\node(E) at (0,1) {$\MspCCpxy$};
\path (A)-- node[rotate=-90]{$\cong$} (E);
\path (E)-- node[rotate=-90]{$\cong$} (B);
\path (C)-- node[rotate=-90]{$=$} (D);
\thickembtikz{A}{C}
\xthickembtikz{B}{D}{\coprod \operatorname{id}}
\end{tikzpicture}
\end{aligned}
\end{equation}
\end{lemma}
\begin{proof}
Because the arbitrary generator $z$ is kept constant throughout such a subcube $\subcube{k}{b}{a}$, the definitions make this clear.
\end{proof}
Implicit in Lemma \ref{lem:J=E subcube for the complement} is the fact that the diagram \eqref{eq:J=E framed bij for the complement} commutes for any portion of the boundaries of the moduli spaces in question. Furthermore, this trivial cover remains consistent across all choices of $\subcube{k}{b}{a}$ and $z$.
The idea now is to assign a \emph{label} to each component of this trivial cover \[\MspCCpxy\xthickemb{\coprod \operatorname{id}} \cubeMod{b}{a}.\] Then because $\mathscr{C}$ is disjoint from $\mathscr{C}'$, any such label will be maintained across the entire subcube $\scube$ and we are free to build the large moduli space $\MspExy$ one label at a time.
For each fixed label, we can regard $\MspCCpxy$ as a single copy of $\cubeMod{b}{a}$ embedding via the identity map. Thus in any product used to build a composition moduli space, any factors coming from subfaces in $\mathscr{C}'$ contribute only trivial identity embeddings. In this way we are free to build $\MspExy$ in the same manner as in the previous sections, constructing the boundary one codimension at a time. At each stage of this process, products with the trivial identity embeddings maintain thinness and we eventually arrive at a thinly embedded $\de\MspExy$ of codimension one which can be filled using the Pontryagin-Thom construction as before. Since this procedure is identical for each fixed label of component in $\MspCCpxy$, we conclude that our eventual thick embedding
\[\MspExy \thickemb \AmbS\]
contains a covering map of the same order as the trivial cover $\MspCCpxy\xthickemb{\coprod \operatorname{id}} \cubeMod{b}{a}$ and we are done.
\subsection{Subcubes involving $H$ with connected arc diagrams}
\label{sec:m-1 to m J=H conn}
In the case that $J=H$ and the subcube $\scube$ has corresponding arc diagram $D_u$ that is connected, Proposition \ref{prop:conn H} splits us into two further cases. In case (1) of the proposition, we again have that all of our path moduli spaces are thin and the construction of $\MspHxy$ proceeds in precisely the same manner as in Section \ref{sec:m-1 to m J=E conn} when $J=E$.
Therefore we focus on case (2) of Proposition \ref{prop:conn H}, where for fixed $x\in\KG{u},y\in\KG{y}$, every non-empty path moduli space $\pathPts{P}\MspHxy$ consists of four points thickly embedded in two-to-one fashion onto $\left\{\frac 13 , \frac 23 \right\}\subset (0,1)\cong \AmbS$. The proposition also shows that the two points embedded at any single $\frac{i}{3}$ are oppositely framed. Thus for any path $P$ we can split $\pathPts{P}\MspHxy$ into two disjoint sets
\[\pathPts{P}\MspHxy=\pathPts{P}^+\MspHxy\sqcup\pathPts{P}^-\MspHxy,\]
where each $\pathPts{P}^\pm\MspHxy$ consists of two points, one embedded at $\frac{1}{3}$ with framing $\pm 1$, and the other embedded at $\frac{2}{3}$ with framing $\mp 1$. In particular, each $\pathPts{P}^\pm\MspHxy$ is \emph{thinly} embedded into $\AmbS$.
The main idea then is to construct $\MspHxy$ in two `signed' pieces, which we denote by $\ModSp_H^\pm(y,x)$. Each of these is built and thinly embedded into $\AmbS$ precisely as in Section \ref{sec:m-1 to m J=E conn} starting from the thinly embedded
\[\de_{[m-1]}\ModSp_H^\pm(y,x):=\coprod_{\text{paths $P$}} \pathPts{P}^\pm\MspHxy,\]
and ending with a codimension one Pontryagin-Thom argument. Note that the computations in Section \ref{sec:m=2 J=H involved} that are used to build $\de_{[m-2]}\MspHxy$ respect this decomposition into signed pieces, since any turnback described there must connect two oppositely framed points at $\frac 13$ and $\frac 23$. The processes for building $\ModSp_H^+(y,x)$ and $\ModSp_H^-(y,x)$ are thus identical except that all framings are reversed, and we can combine them into a two-to-one thick embedding
\[\MspHxy=\ModSp_H^+(y,x)\sqcup\ModSp_H^-(y,x)\thickemb\AmbS.\]
\begin{remark}
The process above can be compared to Section \ref{sec:m-1 to m J=E disconn}, where identical thin constructions could be done one label at a time leading to thickly embedded moduli spaces. The difference here is that our labels $+$ and $-$ give rise to oppositely framed moduli spaces throughout.
\end{remark}
\subsection{Subcubes involving $H$ with disconnected arc diagrams}
\label{sec:m-1 to m J=H disconn}
Finally we consider subcubes $\scube$ involving the $J=H$ edge, where the corresponding arc diagram $D_u$ is disconnected. The construction in this case involves a mixture of the ideas in Sections \ref{sec:m-1 to m J=E disconn} and \ref{sec:m-1 to m J=H conn}, which we will refer to freely without reiterating any details. As always, we begin by fixing $x\in\KG{u}$ and $y\in\KG{v}$ with some path $P$ through $\scube$ having non-empty $\pathPts{P}\MspHxy$.
Suppose the arc diagram $D_u$ has $k$ homologically essential connected components, which we denote by $\mathscr{C}_1,\dots,\mathscr{C}_k$, ordered by their nesting in the annulus.
The first idea is to build $\MspHxy$ one homologically essential connected component at a time, in the spirit of Section \ref{sec:m-1 to m J=H conn}. To this end, note that each path moduli space $\pathPts{P}\MspHxy$ can be decomposed as
\[\pathPts{P}\MspHxy = \pathPts{P}^{\mathscr{C}_1}\MspHxy \sqcup \cdots \sqcup \pathPts{P}^{\mathscr{C}_k}\MspHxy,\]
where each $\pathPts{P}^{\mathscr{C}_i}\MspHxy$ consists of only the points corresponding to $H$ acting on essential circles in $\mathscr{C}_i$. This decomposition is clearly respected throughout the entire cube, allowing us to decompose our desired moduli space $\MspHxy$ as
\[\MspHxy = \ModSp_H^{\mathscr{C}_1}(y,x) \sqcup \cdots \sqcup \ModSp_H^{\mathscr{C}_k}(y,x).\]
Then fixing a connected component $\mathscr{C}_i$, we seek to build and thickly embed the space $\ModSp_H^{\mathscr{C}_i}(y,x)$ starting from its codimension $(m-1)$ boundary
\[\de_{[m-1]}\ModSp_H^{\mathscr{C}_i}(y,x) :=
\coprod_{\text{paths $P$}} \pathPts{P}^{\mathscr{C}_i}\MspHxy\]
as usual.
In order to do this, we appeal to the arguments in Section \ref{sec:m-1 to m J=E disconn} letting $\mathscr{C}_i$ take the role of $\mathscr{C}$ as notated there. If the component $\mathscr{C}_i$ corresponds to case (1) of Proposition \ref{prop:conn H}, then this construction proceeds precisely as before and we produce a trivial cover according to the usual Lipshitz-Sarkar construction for generators on $\mathscr{C}_i'$, the complement of $\mathscr{C}_i$. If the component $\mathscr{C}_i$ corresponds to case (2) of Proposition \ref{prop:conn H}, then the procedures of Section \ref{sec:m-1 to m J=E disconn} are done one signed label at a time, leading to a trivial cover with twice the order of the covering coming from Lipshitz-Sarkar's construction for generators on $\mathscr{C}_i'$.
In any case, each moduli space $\ModSp_H^{\mathscr{C}_i}(y,x)$ can be built and thickly embedded into a copy of $\AmbS$ which we write as
\[\ModSp_H^{\mathscr{C}_i}(y,x)\thickemb\AmbS\cong \ModSp_\mathcal{Q} \times (i-1,i).\] We then `stack' all of these thick embeddings together as
\[\MspHxy = \coprod_{i=1}^k \ModSp_H^{\mathscr{C}_i}(y,x) \thickemb \coprod_{i=1}^k \ModSp_\mathcal{Q} \times (i-1,i) \hookrightarrow \ModSp_\mathcal{Q} \times (0,k) \cong \AmbS\]
to complete our construction. See Figure \ref{fig:Hexagon times I stacked}.
\begin{figure}
\includegraphics[height=5cm]{HexagonI6.pdf}
\caption{A possible embedding for $\MspHxy$ for a disconnected diagram. The two components are constructed and embedded individually for $H$ acting on different connected components of the arc diagram $D_u$. They are then horizontally `stacked' together as shown.}
\label{fig:Hexagon times I stacked}
\end{figure}
\section{Invariance and naturality with respect to link cobordisms}
\label{sec:invariance}
In order to complete the proof of Theorem \ref{main thm}, we must show that our construction of the map $\mathcal{J}:\X_\mathbb{A}(D)\rightarrow\X_\mathbb{A}(D)$, or equivalently our spectrum $\X_J(D)$ lifting $\mathrm{Cone}(J)$, is invariant under the choices involved. The majority of these choices are considered in \cite{LS} where invariance of the similarly constructed $\X_{Kh}(D)$ is shown, and the proofs there apply to our construction in the same manner. However, in Section \ref{sec:m-1 to m} we made an additional choice of null-cobordism while appealing to the Pontryagin-Thom construction. We show below that $\X_J(D)$ is invariant under this choice as well.
Finally, we will also show that $\mathcal{J}$ commutes, up to homotopy, with the maps on spectra assigned to link cobordisms. Proposition \ref{prop:J commutes with cobs} below will concern elementary link cobordisms, including Reidemeister moves, and will thus conclude the proofs of both Theorem \ref{main thm} and Theorem \ref{thm:J commutes with cobs}.
\subsection{Invariance under choices of cobordisms}
The inductive step in the construction of higher-dimensional moduli spaces (see Section \ref{sec:m-1 to m}) relied on choosing a codimension-1 framed null-cobordism for the boundary. We will now show that different choices are in turn related by a framed codimension-1 cobordism. (Even though we work with codimension-1 orientable submanifolds where the framing can always be assumed to exist, we choose to explicitly specify the framing since it is central to the construction in the context of framed flow categories.)
\begin{lemma} \label{lem:cob choices linked by higher cob}
Suppose $\MspJxy,\ModSp'_J(y,x)$ are two moduli spaces constructed as in Section \ref{sec:m-1 to m} using some choices of null-cobordisms. Then there exists a codimension-1 cobordism, that is a thick embedding
\[\overline{\ModSp}_J(y,x)\thickemb \AmbS\times[0,1]\]
whose boundary image in $\AmbS\times\{0\}$ (respectively $\AmbS\times\{1\}$) recovers the thick embedding $\MspJxy\thickemb\AmbS$ (respectively $\ModSp'_J(y,x)\thickemb\AmbS$).
\end{lemma}
\begin{proof}
The construction of $\overline{\ModSp}_J(y,x)$ is done inductively in parallel with that of the moduli spaces. There is no choice involved in the construction of $0$- and $1$-dimensional moduli spaces, so in these cases $\overline{\mathcal{M}}$ is defined to be the product $\mathcal M \times I$. For the inductive step we focus first on the case when $J=E$ (or $F$) for connected diagrams as in Section \ref{sec:m-1 to m J=E conn}.
Recall that we begin with the $(m-2)$-manifold with corners $\de\MspExy$ (for $m>2$) thinly embedded into $\de\AmbS$, viewed as $S^{m-1}$ minus the poles. Consider $\de\MspExy\subset S^{m-1}\times\{0\}$, $\de\mathcal{M}'_E(x,y)\subset S^{m-1}\times\{1\}$. Just as in the arguments of Section \ref{sec:m-1 to m J=E conn}, the inductive assumption of this lemma provides the existence of a manifold with corners (denote it $\partial \overline{\mathcal{M}}_E(x,y)$) thinly embedded as a framed cobordism for these boundaries in $S^{m-1}\times[0,1]$. Assembling this data: \[\MspExy\cup \partial \overline{\mathcal{M}}_E(x,y)\cup \mathcal{M}'_E(x,y),\]
we have a codimension-1 framed submanifold of $\partial(D^m\times I)$. The corresponding map $\partial(D^m\times I)\longrightarrow S^1$ extends to $D^m\times I$, giving a desired cobordism in $\AmbS\times[0,1]$ between $\MspExy, \mathcal{M}'_E(x,y)$. Note that it extends the cobordism $\partial \overline{\mathcal{M}}_E(x,y)$ in $\partial \AmbS\times[0,1]$ that was given by the inductive assumption.
The extension to the other cases, where the embeddings are indeed thick, is done in direct analogy to the constructions in Sections \ref{sec:m-1 to m J=E disconn}, \ref{sec:m-1 to m J=H conn}, and \ref{sec:m-1 to m J=H disconn}, where all components of the trivial cover can be considered one at a time.
\end{proof}
Next we will use Lemma \ref{lem:cob choices linked by higher cob} to show that stable homotopy types constructed using different choices of moduli spaces are homotopy equivalent.
\begin{proposition}
\label{prop:invariance under cobs}
Fix an annular link diagram $D$, and let $\X_J(D),\X_J'(D)$ denote two stable homotopy types built via Theorem \ref{thm:ffcJ} using different choices of null-cobordisms to construct the moduli spaces as described above. Then $\X_J(D)\simeq\X_J'(D)$.
\end{proposition}
\begin{proof}
We work in the setting of \cite[Section 3.3]{LS}. An analogue of Lemma 3.16 in that reference applied to our construction in Lemma \ref{lem:cob choices linked by higher cob}, together with the ability to `separate sheets of the cover' as in the proof of Proposition \ref{prop:cubical flow is framed}, gives a collection of embeddings $\overline{\mathcal M}_J(y,x)\subset \mathbb{E}_d[{\rm gr}(y):{\rm gr}(x)]\times[0,1]$ which extend the neat embeddings $\iota_{y,x}, \iota'_{y,x}$ of ${\mathcal M}_J(y,x), {\mathcal M}'_J(y,x)$ into $\mathbb{E}_d[{\rm gr}(y):{\rm gr}(x)]\times\{ 0\}$, respectively $\times \{ 1\}$. Informally, this may be thought of as a $1$-parameter family of neat embeddings of flow categories, with singularities at times corresponding to critical points of the cobordisms $\overline{\mathcal M}$.
Consider the construction of the cell complex corresponding to a framed flow category \cite[Definition 3.24]{LS}. To each object $y$ in the flow category, one associates a cell $\mathcal{C}(y)$. To define the attaching map of $\mathcal{C}(y)$ to a lower dimensional cell $\mathcal{C}(x)$, one uses the neat embedding of the moduli space $\ModSp(y,x)$ to identify a subset $\mathcal{C}_x(y)\cong\ModSp(y,x)\times \mathcal{C}(x)\subset \de\mathcal{C}(y)$. The map is then defined on $\mathcal{C}_x(y)$ to be the projection to $\mathcal{C}(x)$, while sending all of $\de {\mathcal C}(y)\smallsetminus \cup_x {\mathcal C}_x(y)$ to the basepoint.
In our setting, the neat embedding of $\overline{\ModSp}_J(y,x)$ into $\mathbb{E}_d[{\rm gr}(y):{\rm gr}(x)]\times[0,1]$ provides an identification of $\overline{\ModSp}_J(y,x)\times\mathcal{C}(x)\subset(\de\mathcal{C}(y))\times[0,1]$, and the corresponding projection $\overline{\ModSp}_J(y,x)\times\mathcal{C}(x) \longrightarrow \mathcal{C}(x)$ gives the desired homotopy between attaching maps corresponding to the neat embeddings of $\ModSp_J(y,x)$ and $\ModSp'_J(y,x)$. The constructions of the thickly embedded $\overline{\ModSp}_J(y,x)$ in Lemma \ref{lem:cob choices linked by higher cob} give rise to homotopies of this sort which are coherent across all attaching maps of all cells, showing that our resulting spectra $\X_J(D)$ and $\X'_J(D)$ are stably homotopy equivalent.
\end{proof}
\subsection{Link cobordisms and Reidemeister invariance}
Our goal in this section is to show that the $\mathcal{J}$ map constructed in Corollary \ref{cor:ffcj refines Cone J} commutes up to homotopy with the maps assigned to link cobordisms in \cite[Section 3]{LS2}. This will include the case of link isotopies, indicating that $\mathcal{J}$ commutes with the stable equivalences assigned to Reidemeister moves in \cite[Section 6]{LS}. We begin with some general notions about certain subcategories of flow categories that are used to construct the desired maps.
Let $\ffc$ be a framed flow category. Recall from \cite[Definition 3.29]{LS} the notion of a downwards (resp. upwards) closed subcategory $\mathcal{Z} \subset \ffc$, which is a certain full subcategory whose geometric realization $\lr{\mathcal{Z}}$ is naturally a subcomplex (resp. quotient complex) of $\lr{\ffc}$. Note that any full subcategory is specified by its objects. Thus if $\mathcal{Z} \subset \ffc$ is downward closed, then it has a complementary upward closed subcategory $\mathcal{Z}'$ whose objects are $\mathrm{Ob}(\ffc)\setminus \mathrm{Ob}(\mathcal{Z})$. Moreover, upon geometric realization, $\lr{\mathcal{Z}}$ is a subcomplex of $\lr{\ffc}$ with quotient complex $\lr{\mathcal{Z}'}$, allowing us to write the cofibration sequence
\begin{equation}
\label{eq:complementary cofib seq}
\lr{\mathcal{Z}}\hookrightarrow \lr{\ffc} \twoheadrightarrow \lr{\mathcal{Z}'}.
\end{equation}
We will say a full subcategory $\mathcal{Z}\subset \ffc$ is \emph{closed} if it is either upwards or downwards closed.
We now specialize to the situation at hand. Let $D$ be an annular link diagram, and let $\ffcJ(D)$ be the flow category refining $\mathrm{Cone}~J$ (see Theorem \ref{thm:ffcJ}). By definition, the objects of $\ffcJ(D)$ consist of two copies of the objects of $\ffc_\mathbb{A}(D)$,
\[
\mathrm{Ob}(\ffcJ(D)) = \mathrm{Ob}(\ffc_\mathbb{A}(D))_0 \coprod \mathrm{Ob}(\ffc_\mathbb{A}(D))_1,
\]
where we use the subscript $0$ and $1$ to indicate that some objects are before the $J$ map (subscript $0$), and some are after (subscript $1$). The gradings of objects in $\mathrm{Ob}(\ffc_\mathbb{A}(D))_1$ are shifted up by $1$ from the gradings in $\mathrm{Ob}(\ffc_\mathbb{A}(D))_0 $.
\begin{definition}
For a full subcategory $\mathcal{Z} \subset \ffc_\mathbb{A}(D)$, let $\overline{\mathcal{Z}}$ denote the full subcategory of $\ffcJ(D)$ whose objects are two copies of the objects of $\mathcal{Z}$,
\[
\mathrm{Ob}(\overline{\mathcal{Z}}) = \mathrm{Ob}({\mathcal{Z}})_0 \coprod \mathrm{Ob}({\mathcal{Z}})_1,
\]
where $\mathrm{Ob}({\mathcal{Z}})_i \subset \mathrm{Ob}(\ffc_\mathbb{A}(D))_i$. We call $\overline{\mathcal{Z}}$ the \emph{double} of $\mathcal{Z}$, and we let $\mathcal{Z}_0,\mathcal{Z}_1$ denote the two full subcategories of $\overline{\mathcal{Z}}\subset\ffcJ(D)$ defined by
\[
\mathrm{Ob}(\mathcal{Z}_i) = \mathrm{Ob}(\mathcal{Z})_i.
\]
\end{definition}
As with $\ffcJ(D)$, we think of $\mathcal{Z}_0$ and $\mathcal{Z}_1$ as two parallel copies of $\mathcal{Z}$ within the double $\overline{\mathcal{Z}}$, separated by the $J$ edge direction. In this way, $\mathcal{Z}_0$ and $\mathcal{Z}_1$ are complementary downward and upward closed subcategories of $\overline{\mathcal{Z}}$, so that upon geometric realization, there is an inclusion and projection
\[
\lr{\mathcal{Z}_0} \hookrightarrow \lr{\overline{\mathcal{Z}}} \twoheadrightarrow \lr{\mathcal{Z}_1}.
\]
Combining this with the canonical\footnote{The subcategories $\mathcal{Z}_0, \mathcal{Z}_1$ do not involve the $J$ edge, so they are built in exactly the same way as $\mathcal{Z}$.} identifications $\mathcal{Z}_0=\mathcal{Z}, \mathcal{Z}_1 = \Sigma\mathcal{Z}$ (the $\Sigma(-)$ denotes a grading shift as in Definition \ref{def:grading shift ffc}) allows us to write, for any full subcategory $\mathcal{Z}\subset\ffc_\mathbb{A}(D)$, the cofibration sequence
\begin{equation}
\label{eq:doubles cofib seq}
\lr{\mathcal{Z}} \hookrightarrow \lr{\overline{\mathcal{Z}}} \twoheadrightarrow \Sigma\lr{\mathcal{Z}}.
\end{equation}
Meanwhile, if $\mathcal{Z}'$ denotes the complementary full subcategory of $\mathcal{Z}$ in $\ffc_\mathbb{A}(D)$, then the complement of the double is the double of the complement, $\overline{\mathcal{Z}}'=\overline{\mathcal{Z}'}$.
However, if $\mathcal{Z}$ is closed in $\ffc_\mathbb{A}(D)$, then its double $\overline{\mathcal{Z}}$ need not be closed in $\ffcJ(D)$, since the $J$ map may take certain generators corresponding to objects of $\mathcal{Z}$ into $\mathcal{Z}'$, or vice versa.
\begin{lemma}
\label{lem:combined cofib seqs}
Suppose that $\mathcal{Z}$ is a closed subcategory of $\ffc_\mathbb{A}(D)$ whose double $\overline{\mathcal{Z}}$ is also a closed subcategory of $\ffcJ(D)$. Then the cofibration sequences \eqref{eq:complementary cofib seq} for complementary pairs $\mathcal{Z},\mathcal{Z}'$ and $\overline{\mathcal{Z}},\overline{\mathcal{Z}}'$ commute with the cofibration sequences \eqref{eq:doubles cofib seq} for complementary copies within a double, as illustrated below for the case that $\mathcal{Z}$ is downward closed.
\begin{equation}
\label{eq:commuting cofib seqs}
\begin{tikzpicture}[x=1in,y=-.6in,baseline=(current bounding box.center)]
\node(AA) at (0,0) {$\lr{\mathcal{Z}}$};
\node(AB) at (1,0) {$\lr{\overline{\mathcal{Z}}}$};
\node(AC) at (2,0) {$\Sigma\lr{\mathcal{Z}}$};
\node(BA) at (0,1) {$\lr{\ffc_\mathbb{A}(D)}$};
\node(BB) at (1,1) {$\lr{\ffcJ(D)}$};
\node(BC) at (2,1) {$\Sigma\lr{\ffc_\mathbb{A}(D)}$};
\node(CA) at (0,2) {$\lr{\mathcal{Z}'}$};
\node(CB) at (1,2) {$\lr{\overline{\mathcal{Z}}'}$};
\node(CC) at (2,2) {$\Sigma\lr{\mathcal{Z}'}$};
\draw[right hook->] (AA)--(AB);
\draw[right hook->] (BA)--(BB);
\draw[right hook->] (CA)--(CB);
\draw[right hook->] (AA)--(BA);
\draw[right hook->] (AB)--(BB);
\draw[right hook->] (AC)--(BC);
\draw[->>] (AB)--(AC);
\draw[->>] (BB)--(BC);
\draw[->>] (CB)--(CC);
\draw[->>] (BA)--(CA);
\draw[->>] (BB)--(CB);
\draw[->>] (BC)--(CC);
\end{tikzpicture}
\end{equation}
\end{lemma}
\begin{proof}
The inclusion maps, the quotient maps, and the canonical identifications involved are all defined by identifying objects of the categories with cells in the geometric realizations, so the definitions make the commutation clear.
\end{proof}
Note that, whenever we do have a closed subcategory $\mathcal{Z}\subset\ffc_\mathbb{A}(D)$ whose double $\overline{\mathcal{Z}}$ remains closed in $\ffcJ(D)$ as in Lemma \ref{lem:combined cofib seqs}, each row in \eqref{eq:commuting cofib seqs} gives rise to a `horizontal' Puppe map which can be identified as the $\mathcal{J}$ map restricted to the setting of its row. Similarly, each column in \eqref{eq:commuting cofib seqs} also gives rise to a `vertical' Puppe map.
An annular link cobordism can be decomposed into a sequence of so-called \emph{elementary cobordisms} corresponding to either Morse moves or Reidemeister moves. To prove Theorem \ref{thm:J commutes with cobs}, it suffices to show that the map $\mathcal{J}$ commutes up to homotopy with all maps assigned to such elementary cobordisms.
\begin{proposition}
\label{prop:J commutes with cobs}
Let $\phi:\X_\mathbb{A}(D)\rightarrow\X_\mathbb{A}(D')$ be the map on annular Khovanov spectra assigned to an elementary annular link cobordism from $D'$ to $D$ as in \cite[Section 3]{LS2}. Then the map $\mathcal{J}$ commutes with $\phi$ up to homotopy:
\[
\begin{tikzcd}
\X_\mathbb{A}(D) \ar[r, "\mathcal{J}"] \ar[d, "\phi"'] & \X_\mathbb{A}(D) \ar[d, "\phi"] \\
\X_\mathbb{A}(D') \ar[r, "\mathcal{J}"] & \X_\mathbb{A}(D')
\end{tikzcd}
\]
In particular, if $D,D'$ are two diagrams for the same annular link, then the stable equivalence $\X_\mathbb{A}(D) \cong \X_\mathbb{A}(D')$ commutes with $\mathcal{J}$ up to homotopy.
\end{proposition}
\begin{proof}
Each elementary cobordism map arises from identifying a closed subcategory $\mathcal{Z}$ in the framed flow category $\ffc_\mathbb{A}(\td{D})$ for a suitable diagram $\td{D}$, leading to a cofibration sequence of the form \eqref{eq:complementary cofib seq} (or in the case of Reidemeister II and III, a finite sequence of such closed subcategories). We claim that each such closed subcategory $\mathcal{Z}$ used in this way gives rise to a closed double $\overline{\mathcal{Z}}\subset \ffcJ(\td{D})$. Indeed the saddle map arises from a diagram $\td{D}$ having a crossing at the place of the saddle, so that $\mathcal{Z}$ and its complement $\mathcal{Z}'$ correspond to 0- and 1-resolutions of this crossing, which clearly remain closed upon doubling. Meanwhile, all of the other elementary cobordisms (cups, caps, and Reidemeister moves) identify closed subcategories $\mathcal{Z}$ by fixing labels on \emph{trivial} circles in their corresponding $\td{D}$. Since the $J$ map does not affect the labels on trivial circles, any such closed subcategory remains closed upon doubling.
Thus any closed subcategory $\mathcal{Z}$ used to define an elementary cobordism map $\phi$ gives rise to a commuting diagram of cofibrations as in \eqref{eq:commuting cofib seqs}. In all such cases, the map being built will be identified with one of the vertical maps in the first and/or last column (or perhaps the corresponding `vertical' Puppe map), while the $\mathcal{J}$ maps will be identified with the `horizontal' Puppe maps, and these two maps will commute due to the naturality of the Puppe construction.
\end{proof}
\begin{remark}
Conjecturally the maps $\phi$ assigned to link cobordisms in \cite[Section 3]{LS2} are well-defined and do not depend (up to homotopy) on the decomposition into elementary cobordisms. If this is the case, the statement of Theorem \ref{thm:J commutes with cobs} can be similarly strengthened as well.
\end{remark}
\bibliographystyle{alpha}
|
\section{Introduction} \label{secc::introduction}
In the last years, our capacity to collect data has increased at an unprecedented rate \cite{atzori2010, manyika2011}. Today, billions of sensors, transducers or videocameras get data from physical systems, while the advent of the Internet of Things will hugely increase the amount and variety of such data \cite{atzori2010}. This ability is progressively moving us from the parametric regression approach used in model-based Science to non-parametric regression methods, typical of Artificial Intelligence and, in particular, of Artificial Neural Networks (ANN)\cite{maggiora1992computational}, a technique that has had an impressive success in many fields \cite{stulp2015many, hoffmann2019benchmarking, Hill2006, Aneshensel2013, Raghupathi2014, Lopez-Moreno2014, Krizhevsky2012, purwins2019deep}. This new paradigm may be interpreted as a certain return from the well-established hybrid inductive-deductive approach, to the original pure inductive method that only uses direct regressions between input and output, without imposing any additional hypothesis on the mathematical structure of the regression model \cite{minto2018}. However, this move back is being helped by the huge amount and variety of the data available, the power of data preprocessing techniques, especially for non-structured data, the increasing computer power with dedicated capacities \cite{LeCun2019} and a new generation of software tools that simplify and optimize the construction of the networks, the training process, and, finally, the validation of the NN-based model, such as TensorFlow and Keras \cite{gulli2017deep,geron2019hands}, Theano \cite{bergstra2011theano,bastien2012theano} or Pytorch \cite{paszke2017automatic,paszke2019pytorch}.
This latter approach still faces, however, some important drawbacks in its application to physical sciences. For example, scientific data are biased by centuries of knowledge\cite{berry2011, gould1981, kitchin2014big}. Also, spurious correlations between them may be unnoticed by AI methods \cite{kitchin2014, kitchin2014big}, so a blind algorithm without any additional information may lead to wrong predictions. Also, many times, we lack of sufficient data, in terms of quantity, variety and quality, to extract the characteristic features in problems that often involve a large number of variables that interact in a complex manner. As a consequence, we can expect poor extrapolation capacity of such models or, alternatively, overfitting behavior \cite{Xue2019}. Finally, a physically-based model is also useful to get new information by interpreting its structure, parameters, and mathematical properties, being this the reason for the important efforts made to whitening the black-box way of working of current machine-learning predictive algorithms \cite{xu2019explainable}. All these characteristics, with especial emphasis on the latter, have delayed in Physics and Engineering the success that data-driven applications have achieved in other domains.
Despite these current limitations, data-based models, helped by artificial intelligence techniques, have started to gain more and more relevance in predictive Physics. The term coined for this new paradigm is Physically-Guided Data Science (PGDS) \cite{ayensa2019unsupervised, karpatne2017theory, raissi2017physics, raissi2019physics, li2019combination}. A straightforward application of these techniques is dynamic data-driven systems (DDS) \cite{darema2004dynamic, peherstorfer2015dynamic, kirchdoerfer2016data, kutz2016dynamic, ayensa2018new}. The DDS idea is improving the predictive capability of physical models by adding information from the experimental data. As it is well-known, in any physical system, two types of state variables may be identified: i) observable (measurable) ones, that can be obtained directly from physical sensors such as position, temperature or forces; ii) internal non-observable (not directly measurable) variables, that integrate locally other observable magnitudes and depend on the particular internal structure of the system, that is condensed and therefore lost in this integration process. This homogenization procedure is required in any "averaged" theory, established at scales higher than the one of quantum mechanics and first principles. In general, these internal state variables (e.g. stresses, plastic strains, damage, etc.) depend upon the whole time-history of the system, and collect the internal changes in the microstructure. Since they are not directly measurable, the only way of relating them with the former observable ones is by means of physical experiments, sufficiently simplified to allow assuming a certain internal state of the system (e.g. uniform distribution of stresses in the central section of a sample under uniaxial tension). The results obtained are extrapolated to more general conditions by additional assumptions (internal state models). In the DDS approach, the experimental results are used directly, without appealing to that latter extrapolation procedure, thus avoiding using any state model with its inherent assumptions and associated errors. Of course, the counterpart is the need of big amounts of experimental data under sufficiently varied conditions. This is costly and does not avoid the need for simplified experimental designs and the implicit assumptions to approximate in them the values of the internal state variables, since they are not directly measurable.
An opposite perspective is integrating physical knowledge into data science models to inform and improve the data prediction capability of neural networks, that is, to constrain the prediction domain of the standard data model coming from pure data-treatment to fulfill some physical constraints. This idea has been applied to different examples in \cite{karpatne2017physics, raissi2017physics, raissi2019physics, lu2019deepxde, long2019pde, bar2019unsupervised, haghighat2020deep}, while a classification of research topics was enumerated in \cite{karpatne2017theory}. However, in all these works, the physical information was introduced directly as relations between the input and output layers. Only in \cite{raissi2019physics, lu2019deepxde} a first attempt was made to provide the network with some explanatory capacity by adding as output some of the parameters associated with the internal state model. A particular idea in this framework is using the physical universal laws to inform the neural network in such a way that it is then possible to associate some neuron values to internal state variables (observable or not). This idea, named as Physically-Guided Neural Networks with Internal Variables (PGNNIV), was introduced in a previous paper \cite{ayensajimenez2020identification}. In such method, the equations of evolution (physical principles) are treated as constraints between neuron values in the NN, while the network adopts a particular topology imposed by the Physics. Moreover, the internal state equations, that represent the averaged behavior of the internal structure of the system, are directly derived from the NN outcome. This latter characteristic endows the methodology with explanatory capacity. The beauty and power of this idea is that only observable data are used while the internal variables may be predicted by the NN itself, thus relaxing as much as we want the internal state model.
In particular, in continuum Physics (deformable solids and fluid mechanics, electromagnetism, energy and mass transport problems, etc.) the universal physical principles (NN constraints) are written in terms of partial differential equations that are currently computationally solved by means of numerical methods (finite elements, boundary elements, finite volumes, meshless methods and a long etcetera) \cite{ruas2016, larsson2009}. These use a previous discretization step in space, driving to an algebraic, in general non-linear, system, that is then solved by means of standard matrix manipulation. Also, for time-dependent (evolution) problems, another discretization step in time is required to transform the time-continuum problem into a discrete one (alternatively automatic differentiation may be used \cite{raissi2019physics}). This includes the selection of a suitable time integrator (i.e. Euler, Multi-step, Runge-Kutta, among many others) \cite{larsson2009}. In the PGNNIV approach, solving this problem is straightforward by working with the discretized version of the problem, assigning an internal neuron to each nodal variable.
There are several reasons that supports the interest of using PGNNIV in continuum Physics: on one hand, the multiple problems of interest, in many disciplines, that are expressed in this framework. Secondly, the increasing use of images and videos as the main method for data provision in engineering and healthcare, for example, which is equivalent to having a continuous distribution of data in space, previously discretized (pixels in 2D images and voxels in 3D ones), and also in time (time frames in a video). In third place, and following this trend, in the last years, there has been a tremendous effort in treating images and in developing AI tools to make predictions from such images (e.g. convolutional neural networks - CNN-) \cite{yoo2015deep, rawat2017deep, mccann2017convolutional, anwar2018medical, pujari2018}. Finally, the mathematical operators acting on the image can be extended to consider the standard differential operators in continuum Physics, thus allowing PGNNIV to leverage all current possibilities in image treatment to predict the evolution of a physical system (predictive capacity of PGNNIV) as well as to extract information on its structure (explanatory capacity of PGNNIV). Therefore, the similarity in the mathematical language of image treatment and image-based NN predictions (convolutional filters) and of discretized continuous physical problems (discretized differential operators) is so high that makes it plausible to think of a common framework for both, which would allow to take profit of the tools available in both sides to improve the other.
The objective of this work is, therefore, to extend the PGNNIV methodology to continuum problems, showing its predictive capacity to get the input-output relation in a physical system from a sufficient set of data, as well as its unraveling (explanatory) ability to extract knowledge on the system internal structure. This is always performed considering the constraints imposed by Physics, and using only observable (measurable) variables in the training set (here related to continuous distributions of data values in a spatial and/or time domain). Also, the similarities between PDEs and CNN are highlighted and described in detail.
The structure of the paper is as follows: first, in section \ref{sec::problem}, we introduce the problem to solve and the main objectives. Then, we present in section \ref{sec::methods} the general methodology to study continuum physics under the PGNNIV framework, that is, we reformulate the mathematical foundations of continuum physics in the domain of artificial neural networks. Both the fields and the operators are recast in a standard ANN language, as it is TensorFlow. Next, we present in section \ref{sec::validation} several validation examples where the methodology is fully illustrated and its performance is analyzed when dealing with heterogeneous and nonlinear problems. The predictive and explanatory capacity of the methodology is here revealed. Later, in section \ref{sec::numerical_experiments}, we perform several numerical experiments to demonstrate the performance and the main features of the methodology and how it depends on the training dataset size and noise level and the deep learning structure. Finally, we finish the paper with a discussion and the main conclusions of the work.
\section{Problem statement.} \label{sec::problem}
\paragraph{The physical problem (PP).} Let us consider a certain physical problem defined by a set of (possibly nonlinear) partial differential equations which can usually be split into two main groups, universal physical laws and constitutive equations:
\begin{subequations}\label{eq::eq_fundamental}
\begin{align}
\mathcal{F}[\bs{u},\bs{v}] &= \bs{f} \label{eq::eq_fundamental1}\\
\mathcal{H}[\bs{u},\bs{v}] &= \bs{0} \label{eq::eq_fundamental2}
\end{align}
\end{subequations}
Eq. (\ref{eq::eq_fundamental}) must be completed with appropriate initial and/or boundary conditions to make the problem well-posed:
\begin{equation} \label{eq::eq_fundamental_bc}
\mathcal{G}[\bs{u},\bs{v},\bs{g}] = 0
\end{equation}
where $\bs{u}$ and $\bs{v}$ are unknown tensor fields and $\bs{f}$ and $\bs{g}$ are other tensor fields assumed to be known.
The (also known) functionals $\mathcal{F}$ and $\mathcal{G}$ represent the whole set of universal laws associated with the problem in hands, as well as particular geometrical and environmental constraints, whereas $\mathcal{H}$ represents all (possibly unknown) internal state or constitutive relations of the problem. Splitting of the problem in these two sets of partial differential equation systems also drives to the distinction between two kinds of fields: the essential measurable fields, $\bs{u}$, and the internal state fields, $\bs{v}$, that are particular to each continuum based field theory. In many contexts, the underlying theory is formulated such that Eq. (\ref{eq::eq_fundamental2}) may be expressed in the form $\bs{v} = \mathcal{H}(\bs{u})$. For instance, in solid mechanics, $\mathcal{F}$ encodes mass, momentum and energy conservation equations, while $\mathcal{H}$ expresses the material-dependent constitutive relations referred to stresses and strains (or displacements). In this context, $\bs{u}$ is the displacement field and $\bs{v}$ represents stresses and any other internal variable associated with the constitutive framework established (plastic strains, plastic multipliers, stress rates...). To fix ideas, and in the case of linear elasticity theory, we have:
\begin{subequations}\label{eq::eq_example}
\begin{align}
\bs \nabla \cdot \bs{\sigma} &= \rho\bs{b} \\
\bs{\varepsilon} &= \frac{1}{2}\left(\bs \nabla \otimes \bs{u} + \bs{u} \otimes \bs \nabla\right) \\
\bs{\sigma} - \bs{C}:\bs{\varepsilon}&= \bs{0}
\end{align}
\end{subequations}
with boundary conditions:
\begin{subequations}\label{eq::eq_example_bc}
\begin{align}
\bs{u} &= \bar{\bs{u}}. \quad \mathrm{in} \quad \Gamma_D \\
\bs{\sigma} \cdot \boldsymbol{n} &= \bar{\bs{t}}, \quad \mathrm{in} \quad \Gamma_N
\end{align}
\end{subequations}
The two first equations in (\ref{eq::eq_example}) are the equilibrium and kinematics expressions (the former derived from universal laws: the variation of linear and angular momenta, while the latter is a mere definition of the strains in terms of the displacements). On the contrary, the third one is the constitutive relation, which is postulated as a linear relationship between stress and strain. This example may be enriched when considering finite strains and displacements, and nonlinear/inelastic materials\cite{bonet2016nonlinear}, keeping the structure of Eq. (\ref{eq::eq_fundamental}). The general mixed boundary constraints given by (\ref{eq::eq_example_bc}) may be considered as special cases of more general functional operators. For instance $\mathcal{G}[\bs{u}]$ may be defined as $\mathcal{G}[u](\bs{x}) = \int_\Omega \left((\bs{u}(\bs{y})-\bar{\bs{u}}(\bs{y})\right)\delta(\bs{x}-\bs{y}) \, d\bs{y}, \forall \bs{x} \in \Gamma_D$ and $\mathcal{G}[u](\bs{x}) = \bs{0}, \forall \bs{x} \notin \Gamma_D$. For this particular problem, $\bs{u}$ is the essential measurable field and $\bs{\sigma}$ is an internal field, which identifies with $\bs{v}$ in Eq. (\ref{eq::eq_fundamental}), and arising from classical field theory. Finally, $\bs{b}$ corresponds to the stimulus $\bs{f}$, and $\bar{\bs{u}}$ and $\bar{\bs{t}}$ are the known values of the respective Dirichlet and Neumann prescribed boundary conditions identified with $\bs{g}$.
To solve numerically the physical problem (\ref{eq::eq_fundamental}) with boundary conditions (\ref{eq::eq_fundamental_bc}), it has to be previously discretized both in space and time by means of one of the many discretization available techniques available \cite{larsson2009}, such as the Finite Difference Method \cite{langtangen1999computational}, Finite Element Method \cite{zienkiewicz1971finite}, or other spectral techniques \cite{boyd2001chebyshev}. Once discretized, the physical problems writes:
\begin{subequations}\label{eq::eq_fudamental_d}
\begin{align}
\bs{F}(\bs{u},\bs{v}) &= \bs{f} \label{eq::eq_fundamental_d1}\\
\bs{H}(\bs{u},\bs{v}) &= \bs{0} \label{eq::eq_fundamental_d2}
\end{align}
\end{subequations}
with boundary conditions:
\begin{equation} \label{eq::eq_fudamental_bc_d}
\bs{G}(\bs{u},\bs{v},\bs{g}) = 0
\end{equation}
where now $\bs{u}$, $\bs{v}$ are unknown vectors of dimension $n$ (number of degrees of freedom of the problem) containing all nodal values, $\bs{f}$ and $\bs{g}$ are known vectors of dimension $n$ and $n'$ (number of prescribed degrees of freedom at the boundary), respectively, and $\bs{F}$, $\bs{G}$ and $\bs{H}$ are corresponding array-valued functions. As commented before, often, we replace Eq. (\ref{eq::eq_fundamental_d2}) by $\bs{v} = \bs{H}(\bs{u})$. Note that in this discretized version of the problem, the continuous position label $\bs{x}$ is replaced by a discrete index $j$. Similarly, the fields $\bs{u}$, $\bs{v}$, $\bs{f}$ and $\bs{g}$ are replaced by their discrete counterparts, $\bs{u}=(u_1,\ldots,u_n)$, $\bs{v}=(u_1,\ldots,u_n)$, $\bs{f}=(f_1,\ldots,f_n)$ and $\bs{g}=(g_1,\ldots,g_{n'})$.
\paragraph{The Data Science problem (DSP).} Alternatively to the physical problem, we can think in a \emph{Data-Science Problem} consisting in a collection of input-output data $\mathcal{D} = \{\bs{x}^i,\bs{y}^i\}_{i=1,\ldots,N}$ and whose goal is to learn the underlying implicit relationship $\bs{y} = \bs{F}(\bs{x})$, or, in other words, to build an estimate $\tilde{\bs{y}}$ from $\bs{x}$. To face this problem, Deep Learning (DL) applied to regression problems arises as one possibility, so it is possible to set-up a DL model relating $\bs{x}$ and $\bs{y}$ that may be expressed as $\tilde{\bs{y}} = \mathsf{Y}(\bs{x})$ (note that the use of sans serif notation indicates that there is a DL model relating these two variables).
Once the error $\bs{e} = \tilde{\bs{y}} - \bs{y}$ is defined, the construction of the model $\mathsf{Y}$ is performed by solving a minimization problem. We usually define a cost function related to the norm of the error for the whole learning dataset $\mathcal{D}$, for instance $\mathrm{CF} = \sum_{i=1}^N\|\bs{e}^i\|^2$, where $\bs{e}^i = \mathsf{Y}(\bs{x}^i) - \bs{y}^i$.
\section{Methodology} \label{sec::methods}
\subsection{Coupling physics and data science problems}
\paragraph{General recipe.} In order to link both the Physics-based and Data Science problems we build a physically-based
neural network with internal variables (PGNNIV), following the three steps described below. Additional details on the
PGNNIV methodology may be found in \cite{ayensajimenez2020identification}:
\begin{enumerate}
\item \textbf{Selection of appropriate input and output variables:} Since we consider that they come from any type of data-capturing sensor or device, we require that these variables have to be measurable, so they are a subset of variables $\bs{u}$, $\bs{f}$ and $\bs{g}$ of the problem. Of course, the output variables are always the variables that the engineer or scientist want to predict. For notation purposes, we write $\bs{x} = \bs{I}(\bs{u},\bs{f},\bs{g})$ and $\bs{y}=\bs{O}(\bs{u},\bs{f},\bs{g})$.
\item \textbf{Physical constraints:} The physics of the problem given by Eq. (\ref{eq::eq_fundamental_d1}) is now supplied to the NN as constraints on some prescribed layers (PILs). However, this equation includes the internal variables $\bs{v}$, so for the problem to make sense, we have to include Eq. (\ref{eq::eq_fundamental_d2}) as an element of the network. For example, by defining $\bs{u} = \bs{H}(\bs{v})$ or $\bs{v} = \bs{H}(\bs{u})$.
\item \textbf{Model relaxation:} Since the interest of this methodology is both \emph{to predict} new values of the variable $\bs{y}$ and \emph{to unravel} the constitutive model $\bs{H}$, this latter is generally only partially known. That is, we may know partly its functional structure, or some of the associated parameters. Therefore, the model is replaced by a subnetwork, for instance, $\bs{v} = \mathsf{H}(\bs{u})$. Some guidelines to the set-up of this $\mathsf{H}$ are given hereafter.
\end{enumerate}
It is important to note that these three steps have to be balanced and consistent: each model relaxation has to be complemented with the definition of supplementary output variables and/or the addition of physical constraints in order to learn both the output variables and the constitutive model. Classical simulation (by means of the NN) is recovered as a limit case when the model is not relaxed at all (the constitutive relation is assumed to be known) and no learning on its structural parameters is required. In this case, the state model is ``exactly'' imposed as an internal constraint. On the other end, the standard Data Science appears as another particular case of this methodology, when no physical information is introduced to the network, and the constitutive model is considered totally relaxed.
\paragraph{PGNNIV formulation.} Now, all ingredients of the PGNNIV approach have been already set-up: we have a predictive input-output neural network, that we call the Reduced Order Modelling (ROM) network, with appropriate physical constraints acting on some prescribed internal layers (PILs). Renaming all the constraints associated with the known physics of the problem as $\bs{R}$, that is, the whole set of relations in $\bs{F}$ and $\bs{G}$ and, if desired, part of the relations in $\bs{H}$ or any other knowledge on the system, it is possible to write the PGNNIV problem as:
\begin{equation} \label{eq::optimization}
\begin{aligned}
\bs{y} &= \mathsf{Y}(\bs x); \;\;\; \bs{v} = \mathsf{H}(\bs{u}) \\
& \text{s. t.} \; \; \; \; \bs{x} = \bs{I}(\bs u, \bs f, \bs g) \\
& \quad \quad \; \; \bs{y} = \bs{O}(\bs u, \bs f, \bs g) \\
& \quad \quad \; \; \bs{R}(\bs u, \bs v, \bs f, \bs g) = \bs{0}
\end{aligned}
\end{equation}
or, equivalently, following the standard approach in NN, a cost function is minimized, including now the physical constraints as penalty terms. This is equivalent to consider a physically augmented neural network where the new output variables are identically equal to zero and are related to the internal neuron layers through the predefined relations. If we define $e = \norm{\bs{e}}$ and $\pi_j = \norm{\bs{R}_j}$, this leads to the minimization of the cost function:
\begin{equation} \label{eq::optimization_penalty1}
\mathrm{CF} = \frac{1}{N}\sum_{i=1}^N \left((e^i)^2 + \sum_{j=1}^rc_j (\pi^i_j)^2\right)
\end{equation}
where the upper index $i$ indicates that the values of $e$ and $\pi_j$ are associated with the sample $i$ of the dataset $\mathcal{D}$, and $c_j$ are penalty coefficients. Using a standard data science notation, we can write:
\begin{equation} \label{eq::optimization_penalty2}
\mathrm{CF} = \mathrm{MSE}(e) + \sum_{j=1}^rc_j \mathrm{MSE}(\pi_j)
\end{equation}
Note that the coefficients $c_j$, $j=1,\ldots,r$, are numerical parameters of the optimization procedure, so they are new metaparameters of the neural network. Eq (\ref{eq::optimization_penalty2}) may be written in a more compact form as:
\begin{equation} \label{eq::optimization_penalty3}
\mathrm{CF} = \sum_{j=0}^rc_j \mathrm{MSE}(\pi_j)
\end{equation}
just by adding a penalty coefficient to the loss term $c_0$ and defining $\pi_0 = e$.
\subsection{Data and field description} \label{secc::data_fields}
In continuum physical problems, a time-dependent tensor field is a point-dependent magnitude $\bs{f} = \bs{f}(\bs{x},t)$ indexed therefore by the point coordinate $\bs{x}$ and the time $t$. In many problems in continuum Physics, we deal with tensor fields that, once discretized, are represented by arrays of appropriate dimension. For example, the time dependent ($l$-covariant, $m$-contravariant) tensor field $\bs{f}$ is represented by the multi-indexed array $\bs{F}$, where $F^{i_1,\ldots,i_l,k_1,k_2,k_3,k_4}_{j_1,\ldots,j_m} = F^{i_1,\ldots,i_l}_{j_1,\ldots,j_m}(x_{k_1},y_{k_2},z_{k_3},t_{k_4})$. Note that we have used the contravariant indexes for referring the spatial coordinates and time (voxels and time frames when referring to a particular image or video).
The TensorFlow framework \cite{abadi2016tensorflow} is particularly suitable for working with data associated with a physical discretized field. Indeed, a tensor field $\bs{f}$ is represented in TensorFlow notation by a multiarray tensor $\mathtt{f}$. When that tensor field varies among samples of a given dataset, the tensor rank is expanded to take this into consideration. For instance, the value of a two dimensional discretized displacement field at a given time $t_k = k \Delta t$, $\bs{f}(\bs{x},t_k)$, for a given sampled value $i$, is represented by $\mathtt{f}[i,\cdot,\cdot,k]$, that is a 2nd-rank tensor, where the symbol ``$\cdot$'' represents the two spatial indexes as a whole.
One main feature of TensorFlow is that it allows working with both data-independent and data-dependent tensors. Data-independent tensors are unalterable over the learning dataset, whereas data-dependent tensors depend on the considered sampled. For instance, in continuum mechanics, the displacement field $\bs{u}$ and the stress field $\bs{\sigma}$ are dependent on the boundary conditions, $\bs{g}$, that may change with the input of each problem. On the contrary, the elastic tensor, even if we consider the material as heterogeneous, so that such tensor is spatially-dependent, is constant for any possible input stimulus (internal forces and boundary conditions). Depending on the selection of the input and output variables and the specific conditions of the problem, the different fields involved are either constant or variable fields. For instance, for the linear elastic problem given by Eqs. (\ref{eq::eq_example}) with boundary conditions (\ref{eq::eq_example_bc}) we have at least two possibilities:
\begin{itemize}
\item The boundary conditions are fixed and we want to learn the displacement field $\bs{u}$ from the external forces $\bs{b}$. In that case, $\bar{\bs{u}}$, $\bs{n}$, $\bar{\bs{t}}$ and $\bs{C}$ are data-independent tensor fields whereas $\bs{\varepsilon}$, $\bs{\sigma}$ and $\bs{u}$ are data-dependent fields.
\item The boundary conditions (displacements and/or normal tractions at the boundaries) are taken as input and we want to learn the displacement field $\bs{u}$ when the external forces are known and fixed. In that case, $\bs{b}$, $\bs{n}$ and $\bs{C}$ are data-independent and $\bar{\bs{u}}$, $\bar{\bs{t}}$, $\bs{\varepsilon}$, $\bs{\sigma}$ and $\bs{u}$ are data-dependent fields.
\end{itemize}
\subsection{Operator description}
\subsubsection{General considerations and notations}
The operator $\mathcal{F}$, on a vector field $\bs{u}$, acts, after discretization, as a vectorial function $\bs{F}$. This is directly reframed in the TensorFlow language by defining a function relating two multiarray tensors. Briefly, a functional relationship $\bs{u} = \mathcal{F}(\bs{v})$ is first discretized into a $n$-variables function $\bs{v} = \bs{F}(\bs{u})$, which in turns is expressed in TensorFlow as a tensor relationship $\mathtt{v} = F(\mathtt{u})$. Nonetheless, there are two fundamental observation that has to be mentioned:
\begin{enumerate}
\item First, most of the operators acting in the formulation of the continuum physics are either (i) functions acting over the field values or (ii) linear functional operators. Even more, almost all linear operators involved in the formulation of the continua are local operators. TensorFlow is a framework that seems to have been developed for that purpose, as both cases of operators may be seen as convolution filters.
The first case, that is an operator such that $v_i = f(u_i)$, may be expanded using convolution filters in the TensorFlow framework into a multilayer perceptron. For example, if $\bs{v}$ is a 2D vector field, then, $v^{i_1,k_1,k_2}$ is a rank 3 tensor:
\begin{equation} \label{eq::operator1}
\underbrace{\mathtt{v_0}}_{[2,n_x,n_y]} \underbrace{\longrightarrow}_{[2,m_1,n_x,n_y]} \underbrace{\mathtt{v_1}}_{[m_1,n_x,n_y]} \underbrace{\longrightarrow}_{[m_1,m_2,n_x,n_y]} \cdots \underbrace{\longrightarrow}_{[m_{k-1},2,n_x,n_y]} \underbrace{\mathtt{v}_k}_{[2,n_x,n_y]} \\
\end{equation}
where $\mathtt{v_0}=\mathtt{u}$ and $\mathtt{v_k}=\mathtt{v}$. $k$ is the number of hidden layers and $m_1,\ldots,m_k$ are the number of neurons at layer $k$. The universal approximation theorem \cite{cybenko1989approximations,hornik1991approximation,lu2017expressive,hanin2017universal} guarantees that every regular enough function $f$ may be approximated by multilayer perceptrons so this approximation makes sense.
The second case, that is local linear operators, may be reframed in the TensorFlow framework using convolution filters of a given size. If $\bs{v} = \bs{F}(\bs{u})$ is a local linear operator relating a rank $k$ mutiarray and a rank $k'$ multiarray, both representing 2D spatial fields:
\begin{equation} \label{eq::operator2}
\underbrace{\mathtt{u}}_{[\underbrace{2,\ldots,2}_k,n_x,n_y]} \underbrace{\longrightarrow}_{[\underbrace{2,\ldots,2}_k,\underbrace{2,\ldots,2}_{k'},n_x,n_y]} \underbrace{\mathtt{v}}_{[\underbrace{2,\ldots,2}_{k'},n_x',n_y']}
\end{equation}
It is important to note that, as the considered operator is local, the spatial field $\bs{v}$ is undefined at some values close to the boundaries, so $n_x' \neq n_x$ and $n_y' \neq n_y$.
To summarize, all common operators in physical problems may be framed in terms of artificial network structures using convolutional filters or multilayer perceptrons including proper activation functions, in order to take into account possible nonlinearities.
\item Among the data-independent tensors, we distinguish between constant (non-trainable) and variable (trainable) tensors. Once we have fixed the physical problem and decided which is the input-output relation that has to be learned, the selection of the role of each operator tensor is natural: when a tensor is involved in a known operator, such as the ones related with $\bs{F}$ and $\bs{G}$ functions, it is a TensorFlow constant tensor and is denoted with a star. One particular example is the tensor associated with the derivation operator, $\mathtt{D}^*$. If the tensor is associated with an unknown relationship, such as the ROM network $\mathsf{Y}$ or the model network $\mathsf{H}$, the tensor is a variable tensor. An example is the (possibly heterogeneous) elastic tensor $\mathtt{C}$.
\end{enumerate}
\subsubsection{Brief taxonomy of operators}
To illustrate the introduced concepts above, we particularize the general ideas above to a brief taxonomy of different operators found in continuum physical problems. This discussion is not intended to be exhaustive and complete, but showing on the suitability of the PGNNIV formulation to handle a very wide range of operators.
\paragraph{Common linear differential operators.}
With the presented framework, all differential operators can be cast as pre-defined filters acting on field tensors. A (discretized) differential operator is a function $\bs{D}$ transforming one multiarray tensor into another. For linear differential operators, $\mathcal{D}$ are linear. Consequently, they are encoded as known constant tensors $\mathtt{D}^*$. To fix ideas, let us consider the equilibrium equation in solid mechanics (infinitesimal theory):
\begin{equation} \label{eq::equilibrium_intrinsic}
\boldsymbol \nabla \cdot \sigma = \rho \boldsymbol{b}
\end{equation}
Eq. (\ref{eq::equilibrium_intrinsic}) involves covariant derivative, as $\mathrm{div}$ is a linear differential operator defined for a two contravariant tensor. If $\boldsymbol{f}=\mathrm{div}(\boldsymbol{X})$, $f^i=X^{ij}_{|j}$, where $|$ represents the covariant derivative. In a coordinate representation, the covariant derivative is expressed for the considered tensor as $X^{ij}_{|j}= X^{ij}_{,j} +\Gamma^i_{jk}X^{kj} + \Gamma^j_{jk}X^{ik}$ where $\Gamma^k_{ij}$ are the Christoffel symbols that, for the Levi-Civita connection, are defined in terms of the metric tensor $\bs{\mathfrak{g}}$, and satisfies the following linear equation $\mathfrak{g}_{kl}\Gamma^k_{ij} = \frac{1}{2}\left(\mathfrak{g}_{jl,i}+ \mathfrak{g}_{li,j}- \mathfrak{g}_{ij,l}\right)$. Now, as $\bs{\mathfrak{g}}$ encodes the geometry of the problem, the only ingredient to reframe Eq. (\ref{eq::equilibrium_intrinsic}) to the multiarray framework is to select a discretization of the common, one dimensional, derivative operator $\partial_i$ as a tensor operator. For instance, let us consider a two-dimensional problem. $\partial_1$, may be defined using first-order finite difference approximation as:
\begin{equation} \label{eq::partial_i}
\left[\partial_1 F\right]^{\bullet,kl}_\bullet = \frac{1}{l} F^{\bullet,k+1\, l}_\bullet -\frac{1}{l} F^{\bullet,k\, l}_\bullet
\end{equation}
so the tensorial expression of $\partial_1$ is $\left[\partial_i\right]_{mn}^{rs}= -1/l$ if $r=m,s=n$, $\left[\partial_i\right]_{mn}^{rs}= 1/l$ if $r=m+1,s=n$ and $\left[\partial_i\right]_{mn}^{rs}= 0$ otherwise. Note that $\partial_i$ is a convolutional filter, given in a planar 2D representation by the kernel stencil (moving from left to right and from bottom to top):
\begin{equation}
D_1 = \begin{bmatrix}
0 & 0 & 0 \\
0 & -1/l & 1/l \\
0 & 0 & 0
\end{bmatrix}
\end{equation}
Analogously, the tensorial representation of $\partial_2$ is $\left[\partial_i\right]_{mn}^{rs}= -1/l$ if $r=m,s=n$, $\left[\partial_i\right]_{mn}^{rs}= 1/l$ if $r=m,s=n+1$ and $\left[\partial_i\right]_{mn}^{rs}= 0$ otherwise and its planar representation is:
\begin{equation}
D_2 = \begin{bmatrix}
0 & 1/l & 0 \\
0 & -1/l & 0 \\
0 & 0 & 0
\end{bmatrix}
\end{equation}
If we go back to the equilibrium equation, the tensor $\bs \sigma$ associated with the stress field is represented as $\sigma^{ij,kl}$ and the tensor $\bs{b}$ associated with the external forces field per unit mass is represented as $b^{i,kl}$ so the divergence operator may be expressed as a tensor $\mathtt{D}^*$ so that:
\begin{equation} \label{eq::div_components}
\rho b^{i,rs} = D^{rs*}_{j,kl}\sigma^{ij,kl}
\end{equation}
In particular, let us consider that there is no curvature, $\Gamma^k_{ij}=0$. The non zero components of the tensor are $D^{\bullet,11}_{1,11}=D^{\bullet,11}_{2,11}=-1/l$ and $D^{\bullet,11}_{1,12}=D^{\bullet,11}_{2,12}=1/l$ for a two-by-two grid.
Recall that all differential operators may be reframed as convolutional filters in the spatial slots. This has important consequences from a practical point of view:
\begin{itemize}
\item Operator tensors are sparse in the discretization dimensions (that are those of greater dimensionality). This allows sparse-based algebra and storage, resulting in high performance computations and less demanding requirements.
\item Differential operators may be easily built and used in standard neural network software codes and tools, such as TensorFlow, although some care must be taken in indexing.
\end{itemize}
In summary, all differential operators involved in the fundamental balance equations in Continuum Physics (universal laws) may be encapsulated in this tensor framework, provided we have established two main ingredients: the space geometry ($\bs{\mathfrak{g}}$) and a given discretization rule for differentiation ($\delta$).
\paragraph{Constitutive models.}
Constitutive models (or internal state equations) define the internal state (in general, non-measurable) variables of the problem in terms of the essential (measurable) ones \cite{ayensajimenez2020identification}. They can be written in a general case as:
\begin{equation} \label{eq::const_1}
\bs{v}= \mathcal{H}(\bs{u})
\end{equation}
where $\bs{v}$ is the set of internal variable fields and $\bs{u}$ is the set of essential variable fields (for instance, stresses and displacements in continuum mechanics, macroscopic -$\boldsymbol{D},\boldsymbol{H}$- and microscopic -$\boldsymbol{E},\boldsymbol{B}$- intensity of electromagnetic fields in electromagnetism...) and $\mathcal{H}$ must be interpreted as a functional (e.g. $\bs{\sigma} = \bs{H}(\bs{u}) = \frac{1}{2}\bs{C}:\left(\bs{\nabla} \otimes \bs{u} + \bs{u} \otimes \bs{\nabla}\right)$ in linear elasticity).
Once discretized, Eq. (\ref{eq::const_1}) is expressed as:
\begin{equation} \label{eq::const_2}
\bs{v} = \bs{H}(\bs{u})
\end{equation}
where now $\bs{v}$ and $\bs{u}$ are the tensor fields associated to the nodal field values and $\bs{H}$ is a (in general nonlinear) mapping between tensors.
All symmetries and simplifications that may be assumed in the constitutive equation relating the two internal variable fields may be transcribed to the structure of the function $\bs{H}$:
\begin{itemize}
\item The linearity of the functional $\mathcal{H}$ is translated directly into the linearity of the function $\bs{H}$. Using the neural network language, this is equivalent to no internal layers between neurons associated with the tensors $\mathtt{u}$ and $\mathtt{v}$, associated with the fields considered. Different levels of complexity and non-linearity of $\mathcal{H}$ (and therefore $\bs{H}$) may be handled with appropriate topologies of the deep neural network relating $\bs{v}$ and $\bs{u}$.
\item The structure of the function $\bs{H}$ is further exploited using the deep neural network topology, involving different levels of sparsity (see Fig. \ref{fig::net}):
\begin{itemize}
\item Local constitutive laws, such that that the operator $\mathcal{H}$ is local, that is, the value of $\bs{v}$ depends on the values of $\bs{u}(\bs{x})$ in a neighbourhood of $\bs{x}$. This means that $\mathcal{H}(\bs{u}) = \bs{H}(D^1[\bs{u}](\bs{x}),\ldots,D^m[\bs{u}](\bs{x}))$ where $D^k$ is a differential operator of order $k$. $m<\infty$ is called the order of the locality.
When $m=0$ we speak about order-zero local or pointwise constitutive laws. In that case $\bs{v}(\bs{x}) = \mathcal{H}(\bs{u}(\bs{x}))$, or, using an embedding notation, $\frac{\partial H^{\bullet,I}_{\bullet}}{\partial u^{\bullet,J}_{\bullet}}=\delta^I_J$. This entails \emph{block-diagonal} tensor structures, in the slots associated with the spatial discretization: $\mathtt{v}[:,i,j,k,l]=H(u[:,i,j,k,l])$. In the deep learning framework, these kinds of relationships are associated with partitioned networks, as it is illustrated in Fig. \ref{fig::net_1}.
When $m>0$, the tensors are sparse but not necessarily \emph{block-diagonal}. In the deep learning framework, these operators are associated with convolutional filters, as shown in Fig. \ref{fig::net_2}.
\item Non-local constitutive laws are models so that $\mathcal{H}$ is not a local operator, that is, the value of $\bs{v}$ depends on the values of $\bs{v}(\boldsymbol{x})$ on the whole spatial domain. In the language of differential operators, $\mathcal{H}(\bs{u}) = \bs{H}(D^1[\bs{u}](\boldsymbol x),\ldots,D^m[\bs{u}](\boldsymbol x),\ldots)$. There are many ways of defining non-local functionals (see for instance \cite{ros2015nonlocal} and included references for a motivation and examples in elliptic operators). In that case, the tensors are dense and so it is the topology of the neural network associated with the model, as illustrated in Fig. \ref{fig::net_3}.
\end{itemize}
Obviously, these different situations may be modulated in several hierarchical levels in the network. For instance, in linear elasticity, $\bs \sigma = \frac{1}{2}\bs{C}:\left(\nabla \otimes \bs{u} + \bs{u} \otimes \bs \nabla\right)$, so the material is local with respect to $\bs{u}$. Therefore, an accurate deep learning network for working with this model is obtained by combining the previous ideas, as shown in Fig. \ref{fig::net_4}.
\item The difference between homogeneous and heterogeneous constitutive relations may be exploited also. For local models (block-diagonal or sparse tensors), the different blocks or filters may be the same or may be dependent on the spatial considered point. For the former, the indexes referring to the spatial part (denoted using capital letters) of the tensor are spurious and therefore may be omitted. For instance, in the linear elastic problem, in general $\sigma^{ij,I} = C^{ijkl}(I,J)\upvarepsilon^{J}_{kl}$. The fact that elasticity assumes a relationship $\bs{\sigma}(\bs{x}) = \bs{H}(\bs \varepsilon(\boldsymbol x))$ (point-wise) implies that $C^{ijkl}(I,J)=C^{ijkl}\delta^I_J$ where $\bs{C}(I)$ is the common elasticity tensor. For homogeneous materials, $\bs{C}(I) = \bs{C}$.
\item The tensor relations may be adapted for the exploitation of further symmetries of the constitutive equation. This includes:
\begin{itemize}
\item Relations derived from the principe of objectivity, that is, reference frame independence.
\item Constraints related to the physical or geometrical foundations of the model (e.g. major and minor symmetries of the elastic tensor, associated with thermodynamics, angular momentum conservation, and compatibility constraints).
\item Additional constraints related to special symmetries of the constitutive model, that is, orthotropy, isotropy...
\end{itemize}
All these symmetries may be enforced by adding constraints to the PGNNIV (that is, in an implicit way) or by assuming a given topology for the deep neural network (explicit way). Indeed, if $\bs{v} = \bs{H}(\bs{u})$, the existence of a given symmetry is equivalent, in the Noether sense, to the action of a given group of transformations, so that $\bs{H}(\bs{u}) = \bs{H}(\bs{A}(\bs{u}))$ where $\bs{A} \in \mathcal{A}$, a group of transformations. Therefore, we can look for a finite set of transformations $\bs{A}_k$ in a way such that $\bs{H}(\bs{u}) = \bs{H}(\bs{A}(\bs{u})), A \in \mathcal{A} \Leftrightarrow \bs{A}_k(\bs{H}(\bs{u})) = \bs{H}(\bs{u})$ or to \emph{a priory} set up a topology for the deep neural network so that it is invariant under the action of all $\bs{A}$.
\item Finally, the classical framework of parametric fitting is a very particular case, in which some of the internal layers are related to the others by means of a parametric explicit expression. In that case, $\bs{v} = \bs{H}(\bs{u};\bs{\lambda})$ where the function $\bs{H}$ is explicitly imposed, and the functional relationship depends on the value of unknown parameters $\bs{\lambda}$, that are variable TensorFlow scalars obtained, in general, during the training process.
\end{itemize}
\begin{figure}
\begin{subfigure}[b]{.48\textwidth}
\centering
\begin{tikzpicture}[x=2cm, y=1cm, >=stealth]
\foreach \m [count=\y] in {1,2,missing,3}
\node [every neuron/.try, neuron \m/.try] (input-\m) at (0,4.5-\y*1.5) {};
\foreach \m [count=\y] in {1,2,3,4,5,6}
\node [every neuron/.try, neuron \m/.try,fill=white] (hidden-\m) at (1,4-\y*0.5) {};
\foreach \m [count=\y] in {7,8,9}
\node [every neuron/.try, neuron \m/.try,fill=white] (hidden-\m) at (1,-0.5-\y*0.5) {};
\foreach \m [count=\y] in {1,2,missing,3}
\node [every neuron/.try, neuron \m/.try,fill=white] (output-\m) at (2,4.5-\y*1.5) {};
\node [] at (input-1) {$u_1$};
\node [] at (input-2) {$u_2$};
\node [] at (input-3) {$u_n$};
\node [] at (output-1) {$v_1$};
\node [] at (output-2) {$v_2$};
\node [] at (output-3) {$v_n$};
\foreach \i in {1,2,3}
\draw [->] (input-1) -- (hidden-\i);
\foreach \i in {1,2,3}
\draw [->] (hidden-\i) -- (output-1);
\foreach \i in {4,5,6}
\draw [->] (input-2) -- (hidden-\i);
\foreach \i in {4,5,6}
\draw [->] (hidden-\i) -- (output-2);
\foreach \i in {7,8,9}
\draw [->] (input-3) -- (hidden-\i);
\foreach \i in {7,8,9}
\draw [->] (hidden-\i) -- (output-3);
\end{tikzpicture}
\caption{Pointwise constitutive relation.}
\label{fig::net_1}
\end{subfigure}
\begin{subfigure}[b]{.48\textwidth}
\centering
\begin{tikzpicture}[x=2cm, y=1cm, >=stealth]
\foreach \m [count=\y] in {1,2,missing,3}
\node [every neuron/.try, neuron \m/.try] (input-\m) at (0,4.5-\y*1.5) {};
\foreach \m [count=\y] in {1,2,missing,3}
\node [every neuron/.try, neuron \m/.try,fill=white] (output-\m) at (2,4.5-\y*1.5) {};
\node [] at (input-1) {$u_1$};
\node [] at (input-2) {$u_2$};
\node [] at (input-3) {$u_n$};
\node [] at (output-1) {$v_1$};
\node [] at (output-2) {$v_2$};
\node [] at (output-3) {$v_n$};
\draw [->] (input-1) -- (output-1);
\draw [->] (input-2) -- (output-1);
\draw [->] (input-2) -- (output-2);
\draw [->] (input-missing) -- (output-2);
\draw [->] (input-3) -- (output-3);
\node [] at (1,-2) {};
\end{tikzpicture}
\caption{Local constitutive relation.}
\label{fig::net_2}
\end{subfigure} \vspace{0.4cm} \\
\begin{subfigure}[b]{.48\textwidth}
\centering
\begin{tikzpicture}[x=2cm, y=1cm, >=stealth]
\foreach \m [count=\y] in {1,2,missing,3}
\node [every neuron/.try, neuron \m/.try] (input-\m) at (0,4.5-\y*1.5) {};
\foreach \m [count=\y] in {1,2,missing,3}
\node [every neuron/.try, neuron \m/.try,fill=white] (output-\m) at (2,4.5-\y*1.5) {};
\node [] at (input-1) {$u_1$};
\node [] at (input-2) {$u_2$};
\node [] at (input-3) {$u_n$};
\node [] at (output-1) {$v_1$};
\node [] at (output-2) {$v_2$};
\node [] at (output-3) {$v_n$};
\foreach \i in {1,missing,2,3}
\foreach \j in {1,2,missing,3}
\draw [->] (input-\i) -- (output-\j);
\draw [->] (input-3) -- (output-missing);
\node [] at (1,-2) {};
\end{tikzpicture}
\caption{Global constitutive relation.}
\label{fig::net_3}
\end{subfigure}
\begin{subfigure}[b]{.48\textwidth}
\centering
\begin{tikzpicture}[x=2cm, y=1cm, >=stealth]
\foreach \m [count=\y] in {1,2,missing,3}
\node [every neuron/.try, neuron \m/.try] (input-\m) at (0,4.5-\y*1.5) {};
\foreach \m [count=\y] in {1,2,missing,3}
\node [every neuron/.try, neuron \m/.try,fill=white] (hidden-\m) at (1,4.5-\y*1.5) {};
\foreach \m [count=\y] in {1,2,missing,3}
\node [every neuron/.try, neuron \m/.try,fill=white] (output-\m) at (2,4.5-\y*1.5) {};
\node [] at (input-1) {$u_1$};
\node [] at (input-2) {$u_2$};
\node [] at (input-3) {$u_n$};
\node [] at (output-1) {$v_1$};
\node [] at (output-2) {$v_2$};
\node [] at (output-3) {$v_n$};
\draw [->] (input-1) -- (hidden-1);
\draw [->] (input-2) -- (hidden-1);
\draw [->] (input-2) -- (hidden-2);
\draw [->] (input-missing) -- (hidden-2);
\draw [->] (input-3) -- (hidden-3);
\draw [->] (hidden-1) -- (output-1);
\draw [->] (hidden-2) -- (output-2);
\draw [->] (hidden-3) -- (output-3);
\node [] at (1,-2) {};
\end{tikzpicture}
\caption{Hierarchical combination.}
\label{fig::net_4}
\end{subfigure}
\caption{\textbf{Illustration of the different structures associated with constitutive equations.} The different schemes are illustrative for one-dimensional problems. Note that when convolutional layers are applied (local operators), values at the boundaries may not be conveniently described.}
\label{fig::net}
\end{figure}
Sometimes it is useful to introduce other operators to enforce higher-order discretizations, or to adapt the problem to other numerical methods. This is the case of special filters for meshless approaches such as SPH \cite{gingold1977smoothed}, DEM \cite{nayroles1992generalizing} or NEM \cite{sukumar1998natural, chinesta2014} among a crowd. Also, it is easy to adapt this framework to integral formulations as in the FEM. Indeed, Finite Element integrals may be expressed in terms of the nodal values, being this relationship dependent on the shape function and the chosen numerical integrator, but otherwise fixed for a given degree of approximation. For instance, a moving averaging filter applied to the nodes recovers the framework of linear shape functions for a given element. It is also possible to increase the order of the differential operators. This relies on the fact that a higher-order differential operator can be expressed as the subsequent application of lower-order ones, enriching the differentiation scheme. For instance, if $\Delta^+_h$ is the forward difference operator ($\Delta^+_h f^i = \frac{1}{h}(f^{i+1}-f^{i})$), of order $h$, $\frac{1}{h}(\Delta_h^+ - \frac{1}{2}(\Delta_h^+)^2)$ is a forward difference operator of order $h^2$. Finally, another useful possibility is that of stabilization filters in time-dependent problems to ensure the fulfillment of well-known stability criteria \cite{fischer2001filter }or the filters that are designed for obtaining and high fidelity time integrations, such as Crank-Nicolson integration or Runge-Kutta integrators \cite{raissi2019physics}.
The last application of filters are probes. Probes are measurable or quantifiable values related to the different fields by a known function. The most common probe is the value of a tensor field evaluated at a point or a region. Other common probes are measurements defined in the Data-Driven context \cite{ayensa2019unsupervised}, such as surface stress forces over a plane or strains along a direction, at a certain point. These values may be expressed in terms of data-dependent or data-independent tensor quantities (contractions with other vector/tensors, trace, etc.). As a particular case, we may consider integral quantities such as the flow of a tensorial field over a surface (mass, momentum or energy flux), related to conserved quantities in field theories (Noether charges).
\clearpage
\section{Examples of validation} \label{sec::validation}
\subsection{Problem description} \label{sec::problem_description}
The next example illustrates the use of PGNNIV in continuum physical problems after discretization. Let us suppose the following partial differential equation corresponding to a diffusion problem:
\begin{equation}
\label{eq::diffusion}
\boldsymbol \nabla \cdot \left( \boldsymbol{k} \boldsymbol \nabla u\right) = f
\end{equation}
where $u$ is the solution field and $f$ is the source term. This problem is ubiquitous in physics and engineering. Indeed, Eq. (\ref{eq::diffusion}) is used for instance in stationary $u$ in heat transfer conduction problems with $u$ the temperature, also in steady-state, water seepage in soil mechanics, or in electrostatics, among others. Eq. (\ref{eq::diffusion}) is the combination of two different laws:
\begin{itemize}
\item \textbf{A fundamental functional principle} as it is energy conservation (heat transfer), mass conservation (diffusion) or Gauss law (electrostatics), that states as $\boldsymbol \nabla \cdot \boldsymbol{q} = f$, where $\boldsymbol{q}$ is the flow vector (heat flow, mass flow or electric displacement field) and $f$ is the source term (heat source, mass source or electric charge density).
\item \textbf{A constitutive functional equation} as it is the Fourier law (heat transfer), Fick's law (diffusion) or Dielectric behavior (electrostatics), relating the flux variable $\boldsymbol{q}$ that plays the role of internal state field (non-measurable if no additional assumption is made, e.g. uniform distribution of the transported magnitude through a certain area), with the essential field $u$. Commonly, this relationship is formulated in tensor form as $\boldsymbol{q} = -\boldsymbol{K} \boldsymbol{\nabla} \boldsymbol{u}$, where $\boldsymbol{K}$ is the (thermal) conductivity tensor, the diffusion tensor or the dielectric permittivity tensor, respectively, depending on the particular physical problem considered. For general nonlinear problems, the tensor $\boldsymbol{K}$ may be dependent (in a functional sense) on the field $u$ as well as on the point $\boldsymbol{x}$ as any other field. It is common however to particularize this equation for linear ($\boldsymbol{u}$-independent), homogeneous ($\boldsymbol{x}$-independent) and isotropic simplifications.
\end{itemize}
With these assumptions, Eq. (\ref{eq::diffusion}) may be splitted in:
\begin{subequations}
\label{eq::diffusion_PGNN}
\begin{alignat}{2}
\boldsymbol \nabla \cdot \boldsymbol{q} &= f &&\quad \text{(fundamental principle),} \label{eq::diffusion_PGNN_1} \\
\boldsymbol{q} &= -\boldsymbol{k} \boldsymbol \nabla u &&\quad \text{(constitutive equation),} \label{eq::diffusion_PGNN_2}
\end{alignat}
\end{subequations}
together with appropriate boundary conditions.
Using the framework described above, Eq. (\ref{eq::diffusion_PGNN_1}) is the universal law of the problem, and Eq. (\ref{eq::diffusion_PGNN_2}) is the internal state equation. The only difference is that here, the fundamental principle and the state equation are expressed in functional form. But this subtlety is bypassed by using any common discretization technique (Finite Element Method, Finite Differences Method,...) so the values of $u$, $\boldsymbol{q}$ and $\boldsymbol{k}$ are replaced by the corresponding interpolating (nodal) values or by the approximation parameters, depending on the particular approach. Note that, if $L$ is the mesh-size and $n$ is the number of nodes, in the limit case when $n \rightarrow \infty$ and $L \rightarrow 0$, $\boldsymbol{u}$ and $u$ are indistinguishable, and so it is for the tensorial fields $\boldsymbol{q}$ and $\bs{k}$.
For instance, for one-dimensional problems, using forward first-order finite differences, a discretized version of Eq. (\ref{eq::diffusion_PGNN}) is:
\begin{subequations}
\label{eq::diffusion_PGNN_disc}
\begin{alignat}{2}
\frac{q_{i+1} - q_i}{L} &= f_{i+1}, &&\quad i=0,\ldots,n. \label{eq::diffusion_PGNN_disc_1} \\
q_i &= -k_i \frac{u_{i+1} - u_i}{L}, &&\quad i=0,\ldots,n. \label{eq::diffusion_PGNN_disc_2}
\end{alignat}
\end{subequations}
where $L$ is an appropriate mesh size, $u_i$ are the field variables, $q_i$ are the internal state variables and the functional relationship $q = \mathcal{H}(u)$ is now in the form of an algebraic equation $\bs{q} = \bs{H}(\bs{u})$, where we have defined $\bs{q}= (q_0,\ldots,q_n)$, $\bs{u}= (u_0,\ldots,u_n)$ and $\bs{f}= (f_1,\ldots,f_{n})$. Of course, when solving Eq. (\ref{eq::diffusion_PGNN}), or its corresponding discrete version Eq. (\ref{eq::diffusion_PGNN_disc}), proper boundary conditions, parametrized in terms of a set of variables $\bs{g}$, must be supplied. If the problem is now formulated within the PGNNIV framework, and for many cases, these latter boundary values, together with $\bs{f}$, are the natural inputs of the problem, being $\bs{u}$ the output ones. We refer to this problem as the \emph{prediction problem}. However, in other cases, we are rather interested in characterizing a given material from its response to different stimuli. In that case $\bs{g}$, $\bs{f}$ and $\bs{u}$ are the input variables while $\bs{k}$ is the output one in what we denote as the \emph{characterization problem}.
We focus on this work in the prediction of the fields $u$, $q$ and $k$ given a value of the boundary conditions $\bs{g}$ and of the stimuli $\bs{f}$. This will be possible by using the approach stated in Section \ref{sec::methods} such that the nodal values of $\bs{u}$ will be learned from a sufficiently big and varied data set of input-output values, but
constrained by the two following equations:
\begin{subequations}
\label{eq::constraints}
\begin{alignat}{2}
\frac{q_{i+1} - q_i}{L} &= f_{i+1}, &&\quad i=0,\ldots,n. \label{eq::constraints_1} \\
q_i &= -k_i(u_1,\ldots,u_n) \frac{u_{i+1} - u_i}{L}, &&\quad i=0,\ldots,n. \label{eq::constraints_2} \\
u_0 &= g_1, \quad u_n = g_2 \label{eq::constraints_3}
\end{alignat}
\end{subequations}
Eqs. (\ref{eq::constraints}) are formulated so that the internal state variables verify the fundamental principle of flow conservation given by Eq. (\ref{eq::diffusion_PGNN_disc_1}) and the specific boundary conditions. At this level, the main problem relies on the form of the function $k_i = k_i(u_1,\ldots,u_n)$ that is a multiple input - multiple output relationship that will be learned using Deep Learning regression techniques. Further assumptions can be made about the functional form of this relationship, that may be translated to the structure of the deep subnetwork associated to the constitutive equation. For example:
\begin{itemize}
\item Assuming a local relationship of order $m$ between $q$ and $u$, that is, $k_i=F\left((\Delta^+)^m (u_i),(\Delta^+)^{m-1}(u_i),\ldots,(\Delta^+)^0 (u_i)\right)$ where $\Delta^+$ is the forward difference operator and $(\Delta^+)^0 (u_i) = u_i$. Furthermore, it is possible to extend the methodology for non-local operators \cite{ciaurri2018nonlocal}, with the inconvenience of the numerical and computational complexity. In particular, a nonlinear relationship may be reduced to a separable form involving the field $u$, the gradient of the field $u$, and higher-order derivatives, or even non-local operators. Thus, $q_i=\Pi_{j=1}^kF_j\left((\Delta^+)^{r_j}(u_i)^{s_j}\right)$ where now $F_j$ are the functions to be learned. Although this functional form seems arbitrary to some extent, it is ubiquitous in mathematical \cite{osserman2013survey}, physical \cite{frank2005nonlinear,ishizuka2008integral}, engineering \cite{barenblatt1989theory,caffarelli2010nonlinear} and financial \cite{barles1998option,ankudinova2008numerical} problems as it involves a huge range of problems. As a very particular but common case, a pointwise relationship is expressed as $k_i = F_0\left((\Delta^+)^0(u_i)^1\right)$.
\item Assuming a linear (possibly heterogeneous) relationship between the gradient of $u$ and the flow $q$. In that case, $q_i = k_i \Delta^+ \mathrm{u}_i$, that is, $k_i$ are constants. The homogeneous case is a particular one, provided that $k_i = k, i=1,\ldots,n$.
\end{itemize}
Finally the function $k_i = k_i(u_1,\ldots,u_n)$ may be parametrized using model parameters $\bs{\lambda}$ with physical meaning. This approach recovers the classical parametric fitting when a given constitutive model structure is assumed, and is either useful for prediction problems and for model selection or validation \cite{ayensajimenez2020identification}.
For illustrative purposes let us consider the problem:
\begin{equation}
\label{eq::p1}
\frac{d}{dx} \left(k\frac{du}{dx} \right) = 0
\end{equation}
with boundary conditions
\begin{subequations}
\begin{align}
\label{eq::p1_b}
u(x=0) &= g_1 \\
u(x=1) &= g_2
\end{align}
\end{subequations}
To facilitate the discussion, we shall analyze separately the effect of including heterogeneity and nonlinearities, since each problem has its own particularities, even if both problems may be simultaneously studied in one stroke.
\subsection{Homogeneous and heterogeneous problem}
\subsubsection{Network construction}
In this case, we seek for solutions to the problem Eq. (\ref{eq::p1}) and boundary conditions (\ref{eq::p1_b}) when considering $k = k(x)$. Let us consider for such purposes two PGNNIVs that will be defined hereafter. To evaluate the network performance, we will consider the two cases $k(x)=1$ (homogeneous problem, P1) and $k(x) = x+1$ (heterogeneous problem, P2). The analytical solutions to these two problems are respectively $u(x) = (g_2-g_1)x+b_1$ and $u(x) = \frac{g_2-g_1}{\ln 2} \ln(x+1)+g_1$. The solution in terms of all the fields involved is summarized in Table \ref{table::solution_Homo_vs_Hetero}.
\begin{table}[htbp]
\centering
\begin{tabular}{lccc}
\textbf{Fields} & $u(x)$ & $q(x)$ & $k(x)$ \\ \hline
\textbf{P1} & $(g_2-g_1)x+g_1$ & $g_2-g_1$ & $1$ \\ \hline
\textbf{P2} & $u(x) = \frac{g_2-g_1}{\ln 2} \ln(x+1)+g_1$ & $\frac{g_2-g_1}{\ln 2}$ & $x+1$ \\ \hline
\end{tabular}
\caption{\textbf{Analytical solutions associated with the two considered problems.} The solution $u$, flow $q$ and diffusivity $k$ fields are shown.}
\label{table::solution_Homo_vs_Hetero}
\end{table}
$N$ profiles of $u(x)$ were synthetically generated for different values of the boundary conditions, independently and uniformly sampled on $[0;1]^2$, that is $g_1,g_2 \sim \mathcal{U}[0;1]$ and independents. Together with $g_1$ and $g_2$, $q_1 = q(x=0)$ and $q_2 = q(x=1)$ were considered as input variables to ensure that the PGNNIV is associated with a well-posed problem: we need at least one value of the flow, the solution is unique up to an additive constant value. The values of the field $u$ were particularized at $n=10$ points, $x_i = (i-1)/(n-1), i=1,\ldots,n$ so the output variables correspond therefore to the nodal values $u_i = u(x_i), i=1,\ldots,10$.
\paragraph{Construction of the Reduced Order Model NN.} Let us denote the tensor input as $\mathtt{x}$ (shape $[N,4]$), such that $\mathtt{x}[j,1]=g^j_1$, $\mathtt{x}[j,2]=g^j_2$, $\mathtt{x}[j,3]=q^j_1$ and $\mathtt{x}[j,4]=q^j_2$ while the tensor output is denoted as $\mathtt{u}$ (shape $[N,n]$) such that $\mathtt{u}[i,j]=u^j_i$. To predict the values $u_i$ we use standard neural network regression techniques. In this work, we consider multilayer perceptrons as the fundamental tool for building ROM models, although more sophisticated neural networks may be suitable for other problems. In particular, we used a 4-layer network with two hidden layers, each with 15 neurons, such that $\mathtt{y}=\mathtt{Y}[\mathtt{x}]$ with $\mathtt{Y}$ the nonlinear operator that identifies the input-output relation in the neural network. Up to this point, there is no novelty compared to a traditional ANN approach. With all these notations, the prediction error is
\begin{equation}
\mathtt{e} = \mathtt{y} -\mathtt{u}
\end{equation}
\paragraph{Construction of the continuum PGNNIV.} We establish now a tensor operator $\mathtt{D}^*$ associated with the first-order forward differential operator $\Delta^+$ and we define $\mathtt{dy}=\mathtt{D}^*[\mathtt{y}]$ (shape $[N,n-1]$). Now we set a variable tensor, $\mathtt{K}$, of shape $[n-1,n-1]$. The structure of the mathematical equation implies that the tensor $\mathtt{K}$ is diagonal. Here we try two possibilities, resulting in two different PGNNIV: a scalar-tensor (N1) and a general diagonal tensor (N2). The first one is associated with a generic homogeneous problem and the second one to a heterogeneous one. Now we define $\mathtt{q} = -\mathtt{dy}\cdot \mathtt{K}$ (shape $[N,n-1]$) and we concatenate with the two boundary flow values obtaining $\tilde{\mathtt{q}}$ (shape $[N,n+1]$). Finally, we define $\mathtt{f} = \mathtt{D}^*[\tilde{\mathtt{q}}]$, $\bar{\mathtt{u}} = (\mathtt{y}[\cdot,0] - g_1, \mathtt{y}[\cdot,n] - g_2)$ and $\bar{\mathtt{q}} = (\mathtt{q}[\cdot,0] - q_1, \mathtt{q}[\cdot,n-1] - q_2)$. Consequently, the penalties involved in the problem are
\begin{subequations}
\begin{align}
\mathtt{\pi}_1 &= \mathtt{f} \\
\mathtt{\pi}_2 &= \bar{\mathtt{u}} \\
\mathtt{\pi}_3 &= \bar{\mathtt{q}}
\end{align}
\end{subequations}
\paragraph{Cost function, learning algorithm and metaparameters.} As explained in \cite{ayensajimenez2020identification}, a common PGNNIV may be interpreted as a standard NN where the output space is augmented by including new variables (associated with the added constraints) whose exact value is identically zero. Therefore, as in all ANN problems, we have to specify a cost function and a learning algorithm and its associated metaparameters, but also we have to specify the weights associated with the PILs related to the constraints. In the formal minimization problem to be solved, these weights are the penalty coefficients that become, therefore, new metaparameters associated with the NN. In this work, we consider the mean squared error (MSE) as the cost function, both for the error and penalty terms, and we select the ADAMS optimizer \cite{kingma2014adam}. Note that when referring to the MSE of a tensor, we understand the MSE of the sum of the squares of all its components. The differences in the tensor sizes and physical nature (i.e. units) is the reason for the capital role of the selection of the penalty weights. The resulting cost function (CF) of the optimization procedure is, therefore:
\begin{equation}
\mathrm{CF} = c_0\mathrm{MSE}(\mathtt{e})+c_1\mathrm{MSE}(\mathtt{\pi}_1)+c_2\mathrm{MSE}(\mathtt{\pi}_2)+c_3\mathrm{MSE}(\mathtt{\pi}_3)
\end{equation}
The metaparameters of the PGNNIV are summarized in Table \ref{table::metaparam}.
\begin{table}[htbp]
\centering
\begin{tabular}{lcc}
\textbf{Parameter} & & \textbf{Value} \\ \hline
Learning rate & & $0.0003$ \\ \hline
Error coefficients & & $10^7$ \\ \hline
& Flow conservation & $10^2$ \\
Penalty coefficients & Essential boundary conditions & $10^3$ \\
& Natural boundary contitions & $10^3$ \\ \hline
\end{tabular}
\caption{\textbf{PGNNIV metaparameters.}}
\label{table::metaparam}
\end{table}
\subsubsection{Network convergence}
To evaluate the performance of the continuum-based PGNNIV we generated $N=10^4$ samples of input-output values for both P1 and P2 problems. We used $80\%$ of the generated values as training data and $20\%$ as test data. At each iteration along the optimization process, the PGNNIV was fed with the whole training data. The process was stopped after $10^5$ iterations. The convergence of both neural networks is shown in Fig. \ref{fig::HH_convergence}, demonstrating that a low value of the cost function was obtained in both cases for the homogeneous problem. Nonetheless, for the homogeneous problem, the homogeneous network showed a faster convergence as it includes less learning parameters. However, only the heterogeneous network reached similar low values for the heterogeneous problem.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.8\linewidth]{HH_conv.png}
\caption{\textbf{Learning curve for the different networks and datasets.} CF values are smoothed using a constant filter of bandwidth $w=1000$ to avoid sharp oscillations in the logarithmic scale.}
\label{fig::HH_convergence}
\end{figure}
The inability of the homogeneous network to reach low error predictions for the heterogeneous problem is explained by the impossibility of obtaining arbitrarily small values for the penalties associated with flow conservation for such a limited model. This is well illustrated in Fig. \ref{fig::HH_loss}, wich shows the value of the different penalty terms during the optimization process. The penalty terms associated with the flow are unable to reach sufficiently low values because they are incompatible with the assumption of a homogeneous material.
\begin{figure}
\centering
\begin{subfigure}{.8\textwidth}
\includegraphics[width=\linewidth]{HH_Loss_1.png}
\caption{Homogeneous problem.}
\label{fig::HH_Loss_1}
\end{subfigure}\\
\begin{subfigure}{.8\textwidth}
\includegraphics[width=\linewidth]{HH_Loss_2.png}
\caption{Heterogeneous problem.}
\label{fig::HH_Loss_2}
\end{subfigure}
\caption{\textbf{Evolution of the loss term along the optimization process.} The different loss terms are normalized by its maximal value for a simpler and more consistent comparison. Curves associated with $\mathtt{e}$ and $\mathtt{\pi_2}$ overlap as they are both associated with displacement fields.}
\label{fig::HH_loss}
\end{figure}
Finally, it is interesting to compare the convergence of the network in terms of the model network parameters. This is illustrated in Fig. \ref{fig::HH_weights}. As for the homogeneous problem, all model network parameters converge to the theoretical value ($k=1$), both for the homogeneous and heterogeneous network, albeit the homogeneous one converges faster. However, for the heterogeneous problem, each model parameter of the heterogeneous network model converges to a value satisfying the nodal constitutive relationship (that is, the nodal value $k_i$), while the homogeneous network does whatever possible to reduce the CF, that is, the only model parameter converges to an intermediate value of the diffusivity $k$, so it never achieves the same predictive power as the heterogeneous network.
\begin{figure}
\centering
\begin{subfigure}{.8\textwidth}
\includegraphics[width=\linewidth]{HH_Weights_1.png}
\caption{Homogeneous problem.}
\label{fig::HH_Weights_1}
\end{subfigure} \\
\begin{subfigure}{.8\textwidth}
\includegraphics[width=\linewidth]{HH_Weights_2.png}
\caption{Heterogeneous problem.}
\label{fig::HH_Weights_2}
\end{subfigure}
\caption{\textbf{Network model parameter convergence.} The value of the parameters of the tensor $\mathtt{K}$ are plotted until $M=2\times 10^4$ iterations, where the model network has converged.}
\label{fig::HH_weights}
\end{figure}
\subsubsection{Predictive capacity}
Once the network has converged, we can predict the values of the field $u$ by simply interpolating the obtained values for the nodes, $u_i$, predicted by the network. Note that this prediction for the whole space where the input variables have been sampled (in our case, $b_1,b_2\in [0;1]$), has a minimal cost (the one of a single evaluation), as in any other neural network once trained, since it does not require the inversion of any system of equations neither any iteration procedure. Moreover, the values of the fields $q$ and $k$ are obtained as a byproduct of the network without any post-process beyond the nodal interpolation. As a simple illustration, Fig. \ref{fig::HH_illustration} shows the neural network prediction of the essential and derivative fields for the heterogeneous problem and one particular set of boundary conditions ($g_1 = 0.31$, and $g_2 = 0.79$).
\begin{figure}
\centering
\begin{subfigure}{.7\textwidth}
\includegraphics[width=\linewidth]{HH_pred_u.png}
\caption{Field $u$.}
\label{fig::HH_pred_u}
\end{subfigure}\\
\begin{subfigure}{.7\textwidth}
\includegraphics[width=\linewidth]{HH_pred_q.png}
\caption{Field $q$.}
\label{fig::HH_pred_q}
\end{subfigure} \\
\begin{subfigure}{.7\textwidth}
\includegraphics[width=\linewidth]{HH_pred_k.png}
\caption{Field $k$.}
\label{fig::HH_pred_k}
\end{subfigure}
\caption{\textbf{PGNNIV prediction of the output fields.} Both neural networks are able to predict the output field $u$ for the heterogeneous problem but the heterogeneous network outperforms the homogeneous one when predicting fields $q$ and $k$.}
\label{fig::HH_illustration}
\end{figure}
Even if we show the field associated with one single value of the inputs, the performance is general for the whole coverage of the input values. Indeed, the statistics of the normalized $L^2$ errors corresponding to the prediction of the different fields are shown in Table \ref{table::HH_table_L2_1} and \ref{table::HH_table_L2_2}. This error is computed by using the estimate:
\begin{equation}
E^2_r[f] = \frac{\int_0^1(\tilde{f}(x)-f(x))^2\, dx}{\int_0^1 f(x)^2\, dx}
\end{equation}
with $\tilde{f}$ the predicted value of the field and $f$ the true value given by the analytical solution (see Table \ref{table::solution_Homo_vs_Hetero}).
\begin{table}[htbp]
\centering
\begin{tabular}{|l|ccccc|ccccc|}
\hline
& & & $E^2_r[u]$ & & & & & $E^2_r[q]$ & & \\ \hline
& min & $Q_1$ & $Q_2$ & $Q_3$ & max & min & $Q_1$ & $Q_2$ & $Q_3$ & max \\\hline
N1 & $1 \cdot 10^{-4}$ & $1 \cdot 10^{-4}$ & $1 \cdot 10^{-4}$ & $2 \cdot 10^{-4}$ & $4 \cdot 10^{-1}$ & $2 \cdot 10^{-4}$ & $6 \cdot 10^{-4}$ & $1 \cdot 10^{-3}$ & $2 \cdot 10^{-3}$ & $1$ \\
N2 & $5 \cdot 10^{-5}$ & $1 \cdot 10^{-4}$ & $1 \cdot 10^{-4}$ & $1 \cdot 10^{-4}$ & $4 \cdot 10^{-1}$ & $5 \cdot 10^{-4}$ & $6 \cdot 10^{-4}$ & $7 \cdot 10^{-4}$ & $1 \cdot 10^{-3}$ & $1$ \\ \hline
\end{tabular}
\caption{\textbf{Statistics of the error when predicting the output fields for the homogeneous problem.}}
\label{table::HH_table_L2_1}
\end{table}
\begin{table}[htbp]
\centering
\begin{tabular}{|l|ccccc|ccccc|}
\hline
& & & $E^2_r[u]$ & & & & & $E^2_r[q]$ & & \\ \hline
& min & $Q_1$ & $Q_2$ & $Q_3$ & max & min & $Q_1$ & $Q_2$ & $Q_3$ & max \\ \hline
N1 & $7 \cdot 10^{-6}$ & $7 \cdot 10^{-4}$ & $1 \cdot 10^{-3}$ & $1 \cdot 10^{-3}$ & $4 \cdot 10^{-1}$ & $2 \cdot 10^{-1}$ & $2 \cdot 10^{-1}$ & $2 \cdot 10^{-1}$ & $2 \cdot 10^{-1}$ & $1$ \\
N2 & $1 \cdot 10^{-4}$ & $3 \cdot 10^{-4}$ & $6 \cdot 10^{-4}$ & $1 \cdot 10^{-3}$ & $4 \cdot 10^{-1}$ & $1 \cdot 10^{-3}$ & $2 \cdot 10^{-3}$ & $2 \cdot 10^{-3}$ & $3 \cdot 10^{-3}$ & $1$ \\ \hline
\end{tabular}
\caption{\textbf{Statistics of the error when predicting the output fields for the heterogeneous problem.}}
\label{table::HH_table_L2_2}
\end{table}
\begin{table}[htbp]
\centering
\begin{tabular}{|lc|c|}
\hline
& & $E^2_r[k]$ \\ \hline
Homogeneous problem & N1 & $7 \cdot 10^{-6}$ \\
& N2 & $5 \cdot 10^{-4}$\\ \hline
Heterogeneous problem & N1 & $2 \cdot 10^{-1}$ \\
& N2 & $7 \cdot 10^{-3}$\\ \hline
\end{tabular}
\caption{\textbf{Statistics of the error when predicting the diffusivity field for both problems.}}
\label{table::HH_table_L2_3}
\end{table}
As the output field $k(x)$ does not depend on the value of the boundary conditions, all quantile indicators collapse to a single error value, as shown in Table \ref{table::HH_table_L2_3}.
Note that except for some very particular predictions, the error remains small if the PGNNIV is able to learn the constitutive relation (less than $1\%$ error for more than $75\%$ of the predictions). Only the homogeneous network fails when estimating the values of the field $q$, which is the one associated with the constitutive model. Consequently, the PGNNIV is not capable either of learning accurately the value of $k$ (error of the order $20\%$).
In Figs. \ref{fig::HH_errors_b_1} and \ref{fig::HH_errors_b_2} the error is depicted as a function of the boundary conditions for both neural networks and problems. As it may be seen, the error remains always small, independently of the values of $g_1$ and $g_2$, when estimating the field $u$ for both problems and neural networks. The heterogeneous PGNNIV is additionally able to accurately estimate the field $q$, except for values close to the line $g_1=g_2$, when of course $q(x)=0$ and therefore $E^2_r[q] \rightarrow \infty$. Apart from this singular case, the error is generally higher when getting closer to the boundaries of the dataset coverage ($g_1=0,1$, and/or $g_2=0,1$), which is expected due to the self-learning nature of the method presented.
\begin{figure}
\centering
\begin{subfigure}{.8\textwidth}
\includegraphics[width=\linewidth]{HH_errors_b_1a.png}
\caption{Homogeneous network.}
\label{fig::HH_errors_b_1a}
\end{subfigure} \\
\begin{subfigure}{.8\textwidth}
\includegraphics[width=\linewidth]{HH_errors_b_1b.png}
\caption{Heterogeneous network.}
\label{fig::HH_errors_b_1b}
\end{subfigure}
\caption{\textbf{Errors in the homogeneous problem for both neural networks.} Both neural networks are able to predict accurately the fields $u$ and $q$ (less than $1\%$ error) except for the case $g_1=g_2$ when $q(x)=0$.}
\label{fig::HH_errors_b_1}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}{.8\textwidth}
\includegraphics[width=\linewidth]{HH_errors_b_2a.png}
\caption{Homogeneous network.}
\label{fig::HH_errors_b_2a}
\end{subfigure} \\
\begin{subfigure}{.8\textwidth}
\includegraphics[width=\linewidth]{HH_errors_b_2b.png}
\caption{Heterogeneous network.}
\label{fig::HH_Weights_2b}
\end{subfigure}
\caption{\textbf{Errors in the heterogeneous problem for both neural networks.} $u$ is accurately predicted by the two neural networks, while only the heterogeneous neural network is able to predict accurately the field $q$ (less than $1\%$ error, except for the case $g_1=g_2$ when $q(x)=0$).}
\label{fig::HH_errors_b_2}
\end{figure}
\clearpage
\subsection{Linear and nonlinear problem}
\subsubsection{Network construction}
With the aim of getting now a nonlinear version of the problem (\ref{eq::p1}) and boundary conditions (\ref{eq::p1_b}), we state $k = k(u)$. Let us consider now three PGNNIVs that are described in what follows. To evaluate the network performance, we shall consider three cases $k(u)=1$ (constant diffusivity, P1) and $k(u) = u$ (linear diffusivity, P2) and $k(u)=\exp(u)$ (exponential diffusivity, P3). The analytical solutions to these three problems are, respectively, $u(x) = (g_2-g_1)x+g_1$, $u(x) = \sqrt{(g^2_2 - g^2_1)x+b_1^2}$ and $u(x) = \ln\left((\exp(g_2)- \exp(g_1))x+\exp(g_1)\right)$ as well as the corresponding fields are summarized again in Table \ref{table::solution_NN}.
\begin{table}[htbp]
\centering
\begin{tabular}{lccc}
\textbf{Fields} & $u(x)$ & $q(x)$ & $k(x)$ \\ \hline
\textbf{P1} & $(g_2-g_1)x+g_1$ & $g_2-g_1$ & $1$ \\ \hline
\textbf{P3} & $u(x) = \sqrt{(g^2_2 - g^2_1)x+g_1^2}$ & $\frac{g^2_2-g^2_1}{2}$ & $\sqrt{(g^2_2 - g^2_1)x+g_1^2}$ \\ \hline
\textbf{P2} & $u(x) = \ln\left((\exp(g_2)- \exp(g_1))x+\exp(g_1)\right)$ & $\exp(g_2)- \exp(g_1)$ & $(\exp(g_2)- \exp(g_1))x+ \exp(g_2)$ \\ \hline
\end{tabular}
\caption{\textbf{Analytical solutions associated with the three considered problems.} The solution $u$, flow $q$ and diffusivity field $k$ are shown.}
\label{table::solution_NN}
\end{table}
The input and output values that feed the neural network are generated analogously to the previous example.
\paragraph{Construction of the Reduced Order Model NN.} One of the advantages of this methodology is that the ROM neural network only depends on the nature of the input and output variables. Therefore, the ROM-NN for this nonlinear problem is exactly the same as in the previous example and so it is for the prediction error. The nature of the hidden state equation only affects the physical constraints associated with the PILs layers.
\paragraph{Construction of the continuum PGNNIV.} It is at this level where a supplementary effort has to be made. From the tensor $\texttt{y}$, we obtain the derivative field again using the tensorial operator $\mathtt{D}^*$, $\mathtt{dy}=\mathtt{D}^*[\mathtt{y}]$ (shape $[N,n-1]$). However, we define additionally a new tensor $\mathtt{k}$ representing the diffusivity associated with each element. Note that this field has a shape of $[N,n-1]$ as it is defined on the elements rather than on the nodes. When linking the diffusivity with the value of the field, this has to be done at the element level, so we define an element field $\mathtt{um}=\mathtt{M}^*[\mathtt{u}]$ (shape $[N,n-1]$) obtained by averaging the nodal values of the field associated with the considered element. Note that, as stated in the Methodology section, this operator is formulated in terms of a tensor operator, so a convolutional one-dimensional filter of size 2 with constant kernel $1/2$ was used. The next step is to define a neural network model relating the tensors $\mathtt{um}$ and $\mathtt{k}$. Note that the point-wise character of this (unknown) relationship, that is, the fact that $\mathtt{k}[N,j] = f(\mathtt{um}[N,j])$, is easily formulated in the deep learning TensorFlow framework by defining a convolutional neural network that expands in higher dimensional spaces the content of each neuron associated with each element.
For illustrative purposes we try two possibilities, resulting in two different PGNNIV: a 2-layer CNN (no hidden layers) and a 3-layer CNN. For the first, no activation functions were used so the CNN is able to reproduce only linear relationships. For the second one, we used sigmoid activation functions in the hidden layer. In all layers, we consider bias terms before and after applying the activation function. Another possibility, which was explored alternatively, was to prescribe a parametric relation between the two tensors $\mathtt{um}$ and $\mathtt{k}$. Here we illustrate this possibility by prescribing $k(u) = \alpha + \beta u^\gamma$ where $\alpha, \beta$ and $\gamma$ are model parameters.
As a summary the tensorial flow at the model network is, for the three proposed approaches:
\begin{align}
&\underbrace{\mathtt{um}}_{[N,n-1]} \underbrace{\rightarrow}_{[1,n-1,n-1]} \underbrace{\mathtt{k}}_{[N,n-1]} \nonumber \\
&\underbrace{\mathtt{um}}_{[N,n-1]} \underbrace{\rightarrow}_{[1,m,n-1,n-1]} \underbrace{\mathtt{h}}_{[N,m,n-1]} \underbrace{\rightarrow}_{[m,1,n-1,n-1]} \underbrace{\mathtt{k}}_{[N,n-1]} \nonumber \\
&\underbrace{\mathtt{um}}_{[N,n-1]} \underbrace{\rightarrow}_{[1,n-1,n-1]} \underbrace{\mathtt{k}}_{[N,n-1]} \nonumber
\end{align}
where we indicate under each tensor or operation, the shape of the tensor or filter. The different penalty terms related to the PGNNIV are the same used in the previous example.
\paragraph{Cost function, learning algorithm and metaparameters.} All algorithms and metaparameters used in these examples are the same as in the preceding one, with the exception that there is an extra metaparameter related to the network topology, which is the size of the convolutional filter in the second network, that is, $m$. For the following results obtained, we set a value of $m=5$.
\subsubsection{Network convergence}
The convergence in terms of the CF and the different penalties presents the same trend discussed before when analyzing Fig. \ref{fig::HH_convergence}. However, it is interesting here to show the evolution of the model parameters during the optimization problem, which is informative about the nature of the model $k = k(u)$ (Fig. \ref{fig::NN_weights}).
First, it is easy to observe that the parameters of both the 2L-CNN model and the parametric CNN are easily interpreted in physical terms. Indeed, for the problem $k(u) = 1$, the weight of the 2L-CNN converges to $0$ and the bias converges to $1$ because $k(u) = 1\cdot u + 0$. This also happens in the parametric network, where $\alpha \rightarrow 1$ and $\beta \rightarrow 0$ although there is a third spurious parameter, $\gamma$ that remains undetermined and therefore its convergence is not guaranteed. The parameters associated with the 3L-CNN, althought not so easily interpretable, do converge, which is indicative of the fact that the CNN network is able to find an optimal solution, despite the solution got accuracy or not.
For the linear diffusivity case, the interpretation is similar: the weight of the 2L-CNN converges to $1$ and the bias converges to $0$ while for the parametric network, $\alpha \rightarrow 0.06$, $\beta \rightarrow 0.97$ and $\gamma \rightarrow 1.2$. Note that even if the underlying model is learned well enough, the numerical error intrinsic to the network induces another error in the parameter estimation, which darkens the linear $k-u$ relationship. This may be dramatic when extrapolating, being this a well-known drawback when using complex parametric models without paying attention to overfitting. The 3L-CNN did not totally converged after $M=10^5$ iterations even if, as we will see later, yields good enough results. This is possibly due to an excess of network parameters (excess of neurons) and/or to an insufficiently good model learning approach that could be improved with longer runs or using different optimization algorithms (from the mathematical point of view, the problem is not bounded or the search algorithm has not reached a local minimum).
Finally, in the exponential diffusivity problem, the 2L-CNN network and the parametric network reach convergence close to $M=3\times 10^4$ iterations while for the parametric network we get $\alpha = 1.06$, $\beta = 1.60$ and $\gamma=1.47$. This may be interpreted as the optimal least squares solution for the parametric problem, when using the CF considered. This solution may be also obtained by using another optimization approach different from backpropagation (for instance, the Levenberg-Marquardt algorithm\cite{levenberg1944method} commonly used in parametric fitting). This latter family of algorithms is usually resource-intensive and complex to use in large scale problems. The behavior of the 3L-CNN performance is, in that case, similar to the one for the linear diffusivity problem.
\begin{figure}
\centering
\begin{subfigure}{.7\textwidth}
\includegraphics[width=\linewidth]{NN_weights_1.png}
\caption{Constant diffusivity.}
\label{fig::NN_Weights_1}
\end{subfigure} \\
\begin{subfigure}{.7\textwidth}
\includegraphics[width=\linewidth]{NN_weights_2.png}
\caption{Linear diffusivity.}
\label{fig::NN_Weights_2}
\end{subfigure} \\
\begin{subfigure}{.7\textwidth}
\includegraphics[width=\linewidth]{NN_weights_3.png}
\caption{Exponential diffusivity.}
\label{fig::NN_Weights_3}
\end{subfigure}
\caption{\textbf{Network model parameter convergence.} The values of all the network parameters are displayed, but only those represented in continuous line have a direct interpretation in terms of parameters in the relation $k = k(u)$ .}
\label{fig::NN_weights}
\end{figure}
\subsubsection{Predictive capacity}
As in the previous example (homogeneous vs heterogeneous), the performance of the presented methodology in predicting the value of the different fields is now evaluated. There is, however, a difference with respect to that example that is inherent to nonlinear problems: the value of the field $k(x)$ depends now on the boundary conditions, as $k = k(u(x))$ and $u$ depends on the boundary conditions. Fig. \ref{fig::NN_illustration} shows the neural network prediction of all the fields involved in the problem in the hardest case analyzed, that is, exponential diffusivity, for one particular set of boundary conditions ($g_1 = 0.31$ and $g_2 = 0.79$).
\begin{figure}[htbp]
\centering
\begin{subfigure}{.7\textwidth}
\includegraphics[width=\linewidth]{NN_pred_u.png}
\caption{Field $u$.}
\label{fig::NN_pred_u}
\end{subfigure}\\
\begin{subfigure}{.7\textwidth}
\includegraphics[width=\linewidth]{NN_pred_q.png}
\caption{Field $q$.}
\label{fig::NN_pred_q}
\end{subfigure} \\
\begin{subfigure}{.7\textwidth}
\includegraphics[width=\linewidth]{NN_pred_k.png}
\caption{Field $k$.}
\label{fig::NN_pred_k}
\end{subfigure}
\caption{\textbf{PGNNIV prediction of the output fields.} The three neural networks are able to predict the output field $u$ for the exponential diffusivity problem but the predictive capacity varies from one network to another for $q$ and $k$.}
\label{fig::NN_illustration}
\end{figure}
More important than a particular prediction for one set of boundary conditions are the statistics of the errors for the boundary conditions varying in the whole learning space. These statistics are shown in Tables \ref{table::NN_table_L2_1}, \ref{table::NN_table_L2_2} and \ref{table::NN_table_L2_3} for the fields $u$ and $q$ and the three problems, respectively, while the one of the diffusion field $k$ is presented in Table \ref{table::NN_table_L2_4}.
\begin{table}[htbp]
\centering
\begin{tabular}{|l|ccccc|ccccc|}
\hline
& & & $E^2_r[u]$& & & & & $E^2_r[q]$ & & \\ \hline
& min & $Q_1$ & $Q_2$ & $Q_3$ & max & min & $Q_1$ & $Q_2$ & $Q_3$ & max \\ \hline
2L-CNN & $4\cdot 10^{-7}$ & $1\cdot 10^{-6}$ & $2\cdot 10^{-6}$ & $4\cdot 10^{-6}$ & $1\cdot 10^{-1}$ & $3\cdot 10^{-5}$ & $3\cdot 10^{-5}$ & $5\cdot 10^{-5}$ & $7\cdot 10^{-5}$ & $3\cdot 10^{-1}$ \\
3L-CNN & $9\cdot 10^{-7}$ & $5\cdot 10^{-6}$ & $9\cdot 10^{-6}$ & $2\cdot 10^{-5}$ & $2\cdot 10^{-1}$ & $4\cdot 10^{-5}$ & $8\cdot 10^{-5}$ & $1\cdot 10^{-4}$ & $2\cdot 10^{-4}$ & $4\cdot 10^{-1}$ \\
Parametric & $3\cdot 10^{-5}$ & $3\cdot 10^{-5}$ & $4\cdot 10^{-5}$ & $4\cdot 10^{-5}$ & $1\cdot 10^{-1}$ & $1\cdot 10^{-4}$ & $2\cdot 10^{-4}$ & $4\cdot 10^{-4}$ & $8\cdot 10^{-4}$ & $4\cdot 10^{-1}$ \\ \hline
\end{tabular}
\caption{\textbf{Statistics of the error when predicting the output fields for the problem with constant diffusivity}}
\label{table::NN_table_L2_1}
\end{table}
\begin{table}[htbp]
\centering
\begin{tabular}{|l|ccccc|ccccc|}
\hline
& & & $E^2_r[u]$& & & & & $E^2_r[q]$ & & \\ \hline
& min & $Q_1$ & $Q_2$ & $Q_3$ & max & min & $Q_1$ & $Q_2$ & $Q_3$ & max \\ \hline
2L-CNN & $3\cdot 10^{-4}$ & $3\cdot 10^{-3}$ & $5\cdot 10^{-3}$ & $8\cdot 10^{-3}$ & $3\cdot 10^{-1}$ & $5\cdot 10^{-3}$ & $2\cdot 10^{-2}$ & $3\cdot 10^{-2}$ & $4\cdot 10^{-2}$ & $2$ \\
3L-CNN & $2\cdot 10^{-4}$ & $3\cdot 10^{-3}$ & $5\cdot 10^{-3}$ & $8\cdot 10^{-3}$ & $3\cdot 10^{-1}$ & $4\cdot 10^{-3}$ & $2\cdot 10^{-3}$ & $3\cdot 10^{-3}$ & $5\cdot 10^{-2}$ & $2\cdot 10^{1}$ \\
Parametric & $3\cdot 10^{-4}$ & $2\cdot 10^{-3}$ & $4\cdot 10^{-3}$ & $6\cdot 10^{-3}$ & $9\cdot 10^{-2}$ & $2\cdot 10^{-3}$ & $1\cdot 10^{-2}$ & $2\cdot 10^{-2}$ & $3\cdot 10^{-2}$ & $2$ \\ \hline
\end{tabular}
\caption{\textbf{Statistics of the error when predicting the output fields for the problem with linear diffusivity}}
\label{table::NN_table_L2_2}
\end{table}
\begin{table}[htbp]
\centering
\begin{tabular}{|l|ccccc|ccccc|}
\hline
& & & $E^2_r[u]$& & & & & $E^2_r[q]$ & & \\ \hline
& min & $Q_1$ & $Q_2$ & $Q_3$ & max & min & $Q_1$ & $Q_2$ & $Q_3$ & max \\ \hline
2L-CNN & $1\cdot 10^{-4}$ & $2\cdot 10^{-3}$ & $3\cdot 10^{-3}$ & $5\cdot 10^{-3}$ & $1\cdot 10^{-1}$ & $3\cdot 10^{-3}$ & $2\cdot 10^{-2}$ & $3\cdot 10^{-2}$ & $4\cdot 10^{-2}$ & $7\cdot 10^{-1}$ \\
3L-CNN & $1\cdot 10^{-4}$ & $2\cdot 10^{-3}$ & $3\cdot 10^{-3}$ & $5\cdot 10^{-3}$ & $2\cdot 10^{-1}$ & $4\cdot 10^{-3}$ & $2\cdot 10^{-2}$ & $3\cdot 10^{-2}$ & $5\cdot 10^{-2}$ & $7\cdot 10^{-1}$ \\
Parametric & $1\cdot 10^{-4}$ & $2\cdot 10^{-3}$ & $4\cdot 10^{-3}$ & $5\cdot 10^{-3}$ & $2\cdot 10^{-1}$ & $2\cdot 10^{-3}$ & $9\cdot 10^{-3}$ & $2\cdot 10^{-2}$ & $3\cdot 10^{-2}$ & $ 8\times 10^{-1}$ \\ \hline
\end{tabular}
\caption{\textbf{Statistics of the error when predicting the output fields for the problem with exponential diffusivity}}
\label{table::NN_table_L2_3}
\end{table}
\begin{table}[htbp]
\centering
\begin{tabular}{|lc|ccccc|}
\hline
& & & & $E^2_r[k]$ & & \\
& & min & $Q_1$ & $Q_2$ & $Q_3$ & max \\ \hline
Constant diffusivity & 2L-CNN & $3 \cdot 10^{-4}$ & $3 \cdot 10^{-4}$ & $3 \cdot 10^{-4}$ & $4 \cdot 10^{-4}$ & $4\cdot 10^{-4}$ \\
& 3L-CNN & $7 \cdot 10^{-7}$ & $4 \cdot 10^{-5}$ & $5 \cdot 10^{-5}$ & $5 \cdot 10^{-5}$ & $2\cdot 10^{-3}$ \\
& Parametric & $2 \cdot 10^{-7}$ & $1 \cdot 10^{-5}$ & $2 \cdot 10^{-5}$ & $2 \cdot 10^{-5}$ & $2 \cdot 10^{-5}$ \\ \hline
Linear diffusivity & 2L-CNN & $4 \cdot 10^{-3}$ & $7 \cdot 10^{-3}$ & $1 \cdot 10^{-2}$ & $2 \cdot 10^{-2}$ & $\infty$ \\
& 3L-CNN & $1 \cdot 10^{-3}$ & $9 \cdot 10^{-3}$ & $1 \cdot 10^{-2}$ & $2 \cdot 10^{-2}$ & $\infty$ \\
& Parametric & $1 \cdot 10^{-3}$ & $6 \cdot 10^{-3}$ & $1 \cdot 10^{-2}$ & $2 \cdot 10^{-2}$ & $\infty$ \\ \hline
Exponential diffusivity & 2L-CNN & $6 \cdot 10^{-4}$ & $2 \cdot 10^{-2}$ & $3 \cdot 10^{-2}$ & $3 \cdot 10^{-2}$ & $8 \cdot 10^{-2}$ \\
& 3L-CNN & $6 \cdot 10^{-4}$ & $1 \cdot 10^{-2}$ & $2 \cdot 10^{-2}$ & $5 \cdot 10^{-2}$ & $2 \times 10^{-1}$ \\
& Parametric & $6 \cdot 10^{-4}$ & $4 \cdot 10^{-3}$ & $7 \cdot 10^{-3}$ & $1 \cdot 10^{-2}$ & $7 \cdot 10^{-2}$ \\ \hline
\end{tabular}
\caption{\textbf{Statistics of the error when predicting the diffusivity field for the three problems.}}
\label{table::NN_table_L2_4}
\end{table}
In addition to the error associated to the fields $u$ and $q$, it is possible to evaluate the error of the field $k$ for the different boundary conditions. Fig. \ref{fig::NN_k} shows the relative $L_2$ error for the three tested networks and the three datasets.
From the figures and tables presented, we can draw several important observations. For the first problem, the three PGNNIV have a good accuracy, though the 2L-PGNNIV, whose learning power is specific for linear models, provides the best results in terms of errors. Besides, as the problem is linear, the error has a linear shape when visualized in terms of the boundary conditions. The 3L-PGNNIV and the parametric PGNNIV are nonlinear models and so it is the error. For the second problem, the three PGNNIV estimate accurately the value of the field $k$ except for some values close to the boundaries of the sampling space. The three networks are useful therefore as model learners. Finally, for the third problem, we observe that, although the error is low in general for the three models, the 3L-CNN and the parametric network achieves smaller values. That is because the first network has less predictive power than the second for a general class of functions, while, for that case, a parametric model of the form $k(u) = \alpha + \beta u^\gamma$ is able to describe the underlying physics accurately enough. Note that this, however, is particular to the problem in hands, that is, $k(u)=\exp(u)$. For all cases, the highest errors appear always close to the boundaries of the learning domain. If we compare only the second and third PGNNIV, the third one reaches better predictions close to the boundaries. This is a consequence of the multi-parametric nature of neural networks when used in regression problems: certain overfitting is unavoidable to some extent that is glimpsed especially close to the boundary values.
\begin{figure}[htbp]
\centering
\begin{subfigure}{.82\textwidth}
\includegraphics[width=\linewidth]{NN_k_D1.png}
\caption{Constant diffusivity.}
\label{fig::NN_k_D1}
\end{subfigure}\\
\begin{subfigure}{.82\textwidth}
\includegraphics[width=\linewidth]{NN_k_D2.png}
\caption{Linear diffusivity.}
\label{fig::NN_k_D2}
\end{subfigure} \\
\begin{subfigure}{.82\textwidth}
\includegraphics[width=\linewidth]{NN_k_D3.png}
\caption{Exponential diffusivity.}
\label{fig::NN_k_D3}
\end{subfigure}
\caption{\textbf{PGNNIV error in predicting the field $\bs{k}$.} All PGNNIV are able to estimate the value of the field $k$ for the constant and linear diffusivity problems. Although the 2L-CNN network yields small enough errors for the exponential case, the 3L-CNN and the parametric networks provide more accurate results.}
\label{fig::NN_k}
\end{figure}
\subsubsection{Unraveling capacity}
As explained in Section \ref{sec::problem_description}, PGNNIVs have both predictive and unraveling capacity. This has been explored in the precedent example by reproducing the field $k = k(x)$. This field is a direct output of the problem, but when expressed in terms of the variable $x$ may be seen as an explanation (identification) of the heterogeneous constitutive model. For nonlinear problems, however, the problem becomes richer. The interest here is to learn the model $k = k(u)$. One way is to sample the two output fields $k(x)$ and $u(x)$ for each $x$ and all boundary conditions. This leads however to a point cloud due to the noise and discretization errors. But there is one more elegant alternative that consists on exporting the network related to the model. Its convolutional nature makes it independent of the considered point $x$ (that is, element independent). Fig. \ref{fig::NN_model} shows the model predictions for the three datasets and the three tested PGNNIV. It is important to emphasize once again that the parametric network has good learning capacity since $1.06+1.60u^{1.47}$ is a good enough approximation of $\exp(u)$, although this characteristic is specific to the problem in hands.
\begin{figure}
\centering
\begin{subfigure}{.7\textwidth}
\includegraphics[width=\linewidth]{NN_model_1.png}
\caption{Constant diffusivity.}
\label{fig::NN_model_1}
\end{subfigure}\\
\begin{subfigure}{.7\textwidth}
\includegraphics[width=\linewidth]{NN_model_2.png}
\caption{Linear diffusivity.}
\label{fig::NN_model_2}
\end{subfigure} \\
\begin{subfigure}{.7\textwidth}
\includegraphics[width=\linewidth]{NN_model_3.png}
\caption{Exponential diffusivity.}
\label{fig::NN_model_3}
\end{subfigure}
\caption{\textbf{Unraveling capacity of the PGNNIV.} The three presented PGNNIV try to explain the constitutive model $k = k(u)$. Dataset coverage leads to sampled values of $\mathtt{um}$ in $[0.13,0.87]$ which explains the differences obtained beyond those values.}
\label{fig::NN_model}
\end{figure}
\clearpage
\section{Numerical experiments} \label{sec::numerical_experiments}
Next, we evaluate the performance of the methodology presented according to different parameters, inherent to the presented methodology: the dataset size, the error in the training dataset and the size of the hidden layers. The following discussion corresponds to the more complex exponential diffusivity problem using the 3L-CNN network.
\subsection{Dataset size}
Fig. \ref{fig::NE_Dataset_error} shows the $E^2_r$ errors when varying the dataset size, both considering the predictive and unraveling capacity of the network. The main conclusion is that the dataset size has an important impact on accuracy and precision, but not much on the model learning capacity. This result is expected since for each sample of the dataset, the model learning is performed at the nodal level, so the learning capacity is amplified as a consequence of the discretization. This is even more evident when analyzing the spatial error defined for a spatial field $f$ as:
\begin{equation}
\varepsilon_r[f](x) =\left|\frac{\tilde{f}(x)-f(x)}{f(x)}\right|
\end{equation}
This is illustrated in Fig. \ref{fig::NE_Dataset_space}: even if for all fields, a larger dataset implies better estimations, the field $k$ is the less data demanding to be learned.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.8\linewidth]{NE_Dataset_error.png}
\caption{\textbf{Impact of the dataset size on the predictive and unraveling capacity of the network.} The relative $E_r^2$ is shown for different dataset sizes. As the prediction of the fields depends on the boundary conditions, the error is shown with its associated standard error bar.}
\label{fig::NE_Dataset_error}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=\linewidth]{NE_Dataset_space.png}
\caption{\textbf{Impact of the dataset size on the spatial field prediction.} As the different fields involveds depend on boundary conditions, the median of the relative error is shown together with a $95\%$ confident interval.}
\label{fig::NE_Dataset_space}
\end{figure}
\subsection{Noise impact}
To analyze the impact of the data noise on the results we added a white noise to the training data. The noise level is evaluated by introducing a white noise, proportional to the standard deviation of spatial profile. That is, $u^n_i \sim \mathcal{N}(u_i,s)$ where $u_i$ is the noise-free nodal values, $u^n_i$ is their noisy counterparts, and $s = p \sigma$, being $\sigma$ the standard deviation of $\{u_i\}_{i=1,\ldots,n}$. First, we analyze the noise impact on the network convergence. Fig. \ref{fig::NE_Noise_conv} shows the CF evolution as well as the different penalty terms during the optimization process. As the noise affects directly the output field $u$, the differences in the convergence of the CF function are associated mainly to the prediction error term associated with $u$, already in the 1000-th iteration. However, once the output field noise has been filtered by the PGNNIV, there is another error source associated with the numerical discretization, which plays an important role in later stages of convergence, around $M=2 \times 10^4$.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.8\linewidth]{NE_Noise_conv.png}
\caption{\textbf{Impact of the noise level on the network convergence.} The CF plotted at the left panel is the squared sum of the terms shown at the right panel.}
\label{fig::NE_Noise_conv}
\end{figure}
With respect to the impact of the noise in the accuracy and precision of the network, we analyze now the impact of the noise in the $E^2_r$ errors, both for predictive and unraveling capacity. The results are shown in Fig. \ref{fig::NE_Noise_error}. The interpretation is similar to the one for the dataset impact. However, there is one subtlety: the fields $k$ and $q$ are associated with the derivatives of the field $u$ so even if the filtering capacity of the network is remarkable for the prediction of all fields, as it is common when using PGNNIV methodology \cite{ayensajimenez2020identification}, the error in the prediction is higher for the fields involving a derivative of the discretized function, as the error is amplified by numerical discretization. Fig. \ref{fig::NE_Dataset_space} shows the spatial errors, illustrating this fact more clearly.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.8\linewidth]{NE_Noise_error.png}
\caption{\textbf{Impact of the noise level on the predictive and unraveling capacity of the network.} The relative $E_r^2$ is shown for different dataset sizes. As the prediction of the fields depends on the boundary conditions, the error is shown with its associated standard error bar.}
\label{fig::NE_Noise_error}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=\linewidth]{NE_Noise_space.png}
\caption{\textbf{Impact of the noise level on the spatial field prediction.} As the different fields involved depend on boundary conditions, the median of the relative error together with a $95\%$ confidence interval is shown.}
\label{fig::NE_Noise_space}
\end{figure}
\subsection{Learning space size}
Finally, we evaluate the effect of varying the learning space associated with the model. The first comment corresponds to the convergence of the network, which is shown in Fig. \ref{fig::NE_Neurons_conv}: the larger the learning space (i.e. the higher is the number of neurons in the network), the higher the converge cost.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.8\linewidth]{NE_Neurons_conv.png}
\caption{\textbf{Impact of the learning space size on the network convergence.} The upper pannel shows the value of the CF and the lower pannel the value of the weights at each iteration. CF values are smoothed using a constant filter of bandwidth $w=1000$ for avoiding sharp oscillations in the visualization in the logarithmic scale. Note that a sharp change in the value of the parameters is associated with a significative change in the value of CF.}
\label{fig::NE_Neurons_conv}
\end{figure}
The predictive and unraveling capacity of the network is shown in Fig. \ref{fig::NE_Neurons_error}. As it is common in the neural network framework, the learning power of the model increases with the model parameters (number of neurons) until it reaches a stagnation point, beyond which the accuracy does not improve. It is important however to remark that an augmentation of the learning space is always related with an average accuracy improvement but not necessarily with a precision improvement for all specific problems. Indeed, if the learning space is large, the model risks being overfitted, resulting in poor predictions for some special cases. This may be seen when comparing the error bar for the different predictions. To illustrate this fact, Fig. \ref{fig::NE_Neurons_space} plots the spatial distributions of the errors of fields $q$ and $k$ (the ones involved in the model learning). The results are in agreement with our conclusion: lower values of median estimations and larger confident intervals.
\begin{figure}
\centering
\begin{subfigure}{.6\textwidth}
\includegraphics[width=\linewidth]{NE_Neurons_error_1.png}
\caption{Predictive capacity.}
\label{fig::NE_Neurons_error_1}
\end{subfigure}\\
\begin{subfigure}{.6\textwidth}
\includegraphics[width=\linewidth]{NE_Neurons_error_2.png}
\caption{Unraveling capacity.}
\label{fig::NE_Neurons_error_2}
\end{subfigure} \\
\caption{\textbf{Predictive and unraveling capacity of the PGNNIV.} The unraveling capacity is evaluated using the $L^2$ error instead of the relative $L^2$ error since it is not reported in the same plot.}
\label{fig::NE_Neurons_error}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}{.6\textwidth}
\includegraphics[width=\linewidth]{NE_Neurons_space_2.png}
\caption{Field $q$.}
\label{fig::NE_Neurons_space_2}
\end{subfigure}\\
\begin{subfigure}{.6\textwidth}
\includegraphics[width=\linewidth]{NE_Neurons_space_3.png}
\caption{Field $k$.}
\label{fig::NE_Neurons_space_3}
\end{subfigure} \\
\caption{\textbf{Spatial distribution of the error for $q$ and $k$ fields and 2L-CNN and 3L-CNN PGNNIV.} As the different fields involved depend on boundary conditions, the median of the relative error is shown together with a $95\%$ confidence interval.}
\label{fig::NE_Neurons_space}
\end{figure}
In summary, an increase of the number of neurons is a good starting point for reaching good enough results, but it is a rude strategy when the goal is to fit the model precisely and avoiding overfitting. For these purposes, better and more suited strategies are required \cite{hou2017convnets}.
\clearpage
\section{Discussion}
The presented framework introduces a new and singular way of combining of physical knowledge and the power of the most recent data science techniques to solve problems in continuum physics. Although the presented illustrative problems are simple and academic, they highlight all the ingredients and the main features of the methodology. We summarize the following achievements:
\begin{itemize}
\item The capacity of dealing with arbitrary complex models, equations and structures. We have considered nonlinearities and different degrees of material knowledge such as spatial symmetry. Also, locality has been implicitly considered when establishing a local relation between the flow and the essential field.
\item Flexibility to add some or the whole available physical knowledge to the network. Several degrees of knowledge have been tested, involving symmetries, the mathematical character of the constitutive operators and their explicit parametric dependence.
\item The two-stages character of the methodology. In the first step, the predictive and unraveling capacities of the methods are clearly revealed. This process is computationally expensive. However, nowadays, there are a lot of resources to perform this task, such as scalability and parallelization tools, cloud and distributed computing and adapted hardware technologies such as graphical processor units (GPU) and, tensorial processor units (TPUs) and Field Programmable Gate Arrays (FPGAs). The second step is a pure evaluation of the output for the desired input. Primary and derived fields and parameters are obtained in evaluation cost. Only, the post process is related to the interpolation of the nodal values to get the different fields in the whole domain.. As an example, Fig. \ref{fig::NE_Neurons_space} was generated using a $100 \times 100$ grid, so it includes the resolution of $10^4$ nonlinear PDEs. It was however generated in less than one second using a personal laptop as it involves only $10^4$ network evaluations for each model.
\end{itemize}
Nonetheless, there is an important limitation that arises from the present study. It is crucial to correcly choose the best strategy when selecting what is known and what is not in the problem to solve. If something is known, the best strategy is to include it in the PGNNIV, explicitly if possible, or implicitly if not. In the limit case, parametric models are the best ones and less expensive to train. So they are the ones that provide the best results, when correctly assumed. Indeed, centuries dedicated to establish models cannot be wrong. This may drive to think that in that limit case nothing new has been presented, but this is simplistic since the model learning and the predictive capacity are acquired in one stroke and once for all and can be continuously improved by new sets of data input (Dynamic Data Driven Applications Systems, DDDAS). Then, any prediction may be performed later via offline calculation in one single evaluation. This is extremely advantageous for optimization, inverse problems or stochastic computations based on Montecarlo strategies, among others. In particular, we may establish the whole constitutive model, which is equivalent to build a response surface using NN procedures or to derive a Reduced Order Model using backpropagation, with the particularity that the computational cost is reduced due to the constraints with respect to usual NN procedures \cite{ayensajimenez2020identification}.
Finally, two last important remarks should be done about the limitations of the presented methodology:
\begin{itemize}
\item The first one is the obvious fact that the methodology is based on the availability of enough data. Data quantity but mainly data quality is required: large amounts of data are not enough, but they have also to be well distributed and uncorrelated for the model to be accurately learned. This is more complicated to be thought, since there is not a simple way to guarantee that internal variables have a pertinent coverage (see for instance Fig. \ref{fig::NE_Neurons_error}) when the constitutive relation is not \emph{a priori} known.
\item A rough increase of the discretization (that is, to augment the number of nodes or to reduce the mesh size) does not necessarily improve the prediction or unraveling capacity. This is a fundamental difference between this methodology and the usual simulation approach (that is merely predictive). Convergence with the mesh is a very critical aspect, as shown in Fig. \ref{fig::Discretizationv}. Here we point out the need of new research results in this line, related to the mathematical structure of a broad range of problems. In a certain sense, we still suffer from the lack of mathematical results, playing the role that, for instance, convergence theorems play in Finite Element analysis.
\end{itemize}
\begin{figure}[htbp!]
\centering
\includegraphics[width=1.0\linewidth]{Discretization.png}
\caption{\textbf{Effect of the discretization in the predictive and unraveling results.} Note that a mesh refinement is not necessarily accompanied by an improvement in the results.}
\label{fig::Discretizationv}
\end{figure}
As future work we point in two directions:
\begin{itemize}
\item The methodology is new and lacks of sound theoretical results that justify its application for general classes of problems. Indeed, the performance of the trained PGNNIV may be evaluated by analyzing the value of the cost function for the test data, or, even better, the value of each of the loss terms, as each one is associated with one physical aspect of the problem in hands (e.g. mass, momentum or energy conservation, etc.). Although the predictive capacity is directly assessable by evaluating the corresponding loss term of the cost function, it is not the case for the unraveling capacity: if input and output data are not rich enough or the physical added content is insufficient, the network may have good predictive capacity while misleading the constitutive state model. Consider, for example, the case of an elastic clamped beam subjected to axial load. If the displacement is measured only at the loaded node, many possible stiffness ($K(x) = E(x)A(x)/L$) distributions along the beam are possible resulting in the same displacement associated with a given load. The problem is solved: (i) by monitoring the whole displacement field or (ii) by explicitly imposing stiffness homogeneity. This fundamental problem, which we call the slack problem, requires further research, as now its treatment is based on the modeler's previous knowledge and intuition about the problem.
\item As the methodology is based on a TensorFlow reframing of mathematical problems, it is especially well suited for easy and structured mesh discretizations. An octree-based formulation is one of the common strategies to face complex geometries, but this would, in principle, greatly increase the input and output size of the problem. These problems also aggravates when going to higher dimensions in what is called the curse of dimensionality \cite{poggio2017and}. It is therefore important to think of strategies to bypass this difficulty.
\end{itemize}
\section{Conclusions}
In the present work, we introduce a general framework for the analysis of problems in continuum physics from a data science perspective that incorporates the fundamental physical laws in the computations. We have taken advantage of the characteristics of a very recent idea, the so-called Physically-Guided Neural Networks with Internal Variables, for a very general formulation of a broad class of problems whose physical content is expressed by means of a system of partial differential equations. The key point is a cunning splitting of the equations of the problem into two sets, what is known and what is not, and therefore you want it to be learned. The result is a method that offers both predictive and unraveling capacity: it is possible to predict in a single evaluation the state of the system for any prescribed value of the input variable, and to learn about the mathematical structure of the constitutive state equation of the problem in hands. Indeed, good results have been demonstrated for some paradigmatic cases in science and engineering related to constitutive modeling: local character, heterogeneity and nonlinearity.
Several numerical experiments have been carried out to illustrate some of the characteristics of the method, in particular those related to the dataset size and noise, and the size of the model networks. The results show the trends and features that are common to machine learning techniques, in particular, the importance of the dataset, the filtering capacity and the problem of overfitting. Nevertheless, the numerical experiments and examples shown in this work should be reinforced by further research exploring the mathematical structure of the PGNNIV problem, depending on the particular selection of the topology for $\mathsf{Y}$ (ROM) and $\mathsf{H}$ (constitutive state equation) subnetworks, as well as the prescription of the known functions $\bs{I}$, $\bs{O}$ and $\bs{R}$ in Eq. (\ref{eq::optimization}). In any case, PGNNIV are very promising for both predicting and unraveling problems formulated in terms of partial differential equations, ubiquitous in sciences and engineering.
\section*{Acknowledgements}
The authors gratefully acknowledge the financial support from the Spanish Ministry of Economy and Competitiveness (MINECO) and FEDER, UE through the project PGC2018-097257-B-C31, the Spanish Ministry of Science and Innovation through the project PID2019-106099RB-C44/AEI/10.13039/501100011033, the Government of Aragon (DGA) through the grant T24\_17R and the Centro de Investigacion Biomedica en Red en Bioingenieria, Biomateriales y Nanomedicina (CIBER-BBN). CIBER-BBN is Financed by the Instituto de Salud Carlos III with assistance from the European Regional Development Fund.
\bibliographystyle{unsrt}
|
\section{Introduction}\label{sec1: intro}
The {\em (large) Schr\"oder numbers }
$$
S_n=\sum_{i=0}^n\frac{1}{i+1}\binom{2i}{i}\binom{n+i}{n-i}
$$
are one of the most fundamental integer sequences in mathematics. They arise in many classical combinatorial enumeration problems~\cite{kit,shapi2,stan}. In pattern avoidance, one of the most important classes of permutations, known as {\em separable permutations}, are counted by the Schr\"oder numbers~\cite{shapi,wes,sk}.
There are two distinct ways to define the separable permutations. One is in terms of pattern avoiding permutations. Let $\mathfrak{S}_n$ be the set of permutations of $[n]:=\{1,2,\ldots,n\}$. A permutation $\pi\in\mathfrak{S}_n$ is said to {\em avoid} a permutation (or a pattern) $\sigma\in\mathfrak{S}_k$, $k\leq n$, if there exists no subsequence of $\pi$ that is order isomorphic to $\sigma$. Separable permutations are permutations that avoid both the patterns $2413$ and $3142$.
Another description of separable permutations is via two elementary operations, called direct sum and skew sum of permutations.
The \emph{direct sum} $\pi\oplus\sigma$ and the \emph{skew sum} $\pi\ominus\sigma$, of $\pi\in\mathfrak{S}_k$ and $\sigma\in\mathfrak{S}_l$, are permutations in $\mathfrak{S}_{k+l}$ defined respectively as
$$
(\pi\oplus\sigma)_i=
\begin{cases}
\pi_i, &\text{for $1\leq i\leq k$};\\
\sigma_{i-k}+k, &\text{for $k+1\leq i\leq k+l$}.
\end{cases}
$$
and
$$
(\pi\ominus\sigma)_i=
\begin{cases}
\pi_i+l, &\text{for $1\leq i\leq k$};\\
\sigma_{i-k}, &\text{for $k+1\leq i\leq k+l$}.
\end{cases}
$$
For instance, we have $123\oplus 21=12354$ and $123\ominus 21=34521$. The following characterization of separable permutations is folkloric (see~\cite[p.~57--58]{kit}) in pattern avoidance.
\begin{proposition}\label{desides}
A permutation is separable if and only if it can be built from the permutation $1$ by applying the operations $\oplus$ and $\ominus$ repeatedly.
\end{proposition}
This characterization induces a natural bijection~\cite{shapi} between separable permutations and di-sk trees, where a {\em di-sk tree} is a rooted binary tree whose nodes are labeled by $\oplus$ or $\ominus$, and no node has the same label as its right child. The trees considered in this paper will all be di-sk trees. The main objective of this paper is to construct a combinatorial bijection on di-sk trees that proves a symmetric quintuple equidistribution on separable permutations involving two Comtet statistics \cite{flw}, the number of components and the length of the initial ascending run.
For a permutation $\pi=\pi_1\pi_2\cdots\pi_n\in\mathfrak{S}_n$, define six statistics
\begin{align*}
\mathrm{LMAX}(\pi) &:=\{\pi_i:\pi_j<\pi_i,\: \forall 1\le j < i\};\\
\mathrm{LMIN}(\pi) &:=\{\pi_i:\pi_j>\pi_i,\: \forall 1\le j < i\};\\
\mathrm{DES}(\pi)&:=\{i\in[n-1]:\pi_i<\pi_{i+1}\};\\
\mathrm{DESB}(\pi)&:=\{\pi_{i+1}: i\in\mathrm{DES}(\pi)\};\\
\mathsf{iar}(\pi)&:=\min(\mathrm{DES}(\pi)\cup\{n\});\\
\mathsf{comp}(\pi)&:=|\{i: \forall j\leq i,\: \pi_j\leq i\}|;
\end{align*}
called {\em the set of values of left-to-right maxima}, {\em the set of values of left-to-right minima}, {\em the set of positions of descents}, {\em the set of descent bottoms}, {\em the length of initial ascending run} and {\em the number of components} of $\pi$, respectively. For a (finite) collection of patterns $P$, we write $\mathfrak{S}_n(P)$ for the set of all permutations in $\mathfrak{S}_n$ that avoid simultaneously every pattern contained in $P$.
\begin{theorem}\label{thm:sep:sym}
There exists an involution $\Phi$ on $\mathfrak{S}_n(2413,3142)$ that preserves the triple of set-valued statistics $(\mathrm{LMAX},\mathrm{LMIN},\mathrm{DESB})$ but exchanges the pair $(\mathsf{comp},\mathsf{iar})$. Moreover, $\Phi$ restricts to an involution on $\mathfrak{S}_n(312)$.
\end{theorem}
The inspiration of Theorem~\ref{thm:sep:sym} stems from the work of Comtet~\cite[Ex.~VI.14]{com} and several recent results. The two statistics $\mathrm{LMAX}/\mathrm{LMIN}$ and $\mathrm{DESB}$ are respectively the set-valued extensions of the classical Stirling and Eulerian statistics, since the number of left-to-right maxima/minima over $\mathfrak{S}_n$ gives the {\em Stirling numbers of the first kind} and the descent polynomial on $\mathfrak{S}_n$ is the {\em$n$-th Eulerian polynomial} (see~\cite{bon,FS,pet}). Note that $\mathsf{comp}(\pi)$ equals the maximum number of components in an expression of $\pi$ as a direct sum of permutations~\cite{adin}. The statistic $\mathsf{comp}$ dates back at least to Comtet~\cite[Ex.~VI.14]{com} and following \cite{flw}, any statistic equidistributed with $\mathsf{comp}$ over a class of restricted permutations will be called a {\em Comtet statistic} over such class. The statistic $\mathsf{iar}$ was considered by Claesson and Kitaev in~\cite{CK}, but under the different notation $\mathsf{lir}$. It was known that
\begin{equation}\label{eq:cat}
|\{\pi\in\mathfrak{S}_n(321): \mathsf{iar}(\pi)=k\}|=C_{n,n-k}= |\{\pi\in\mathfrak{S}_n(321): \mathsf{comp}(\pi)=k\}|,
\end{equation}
where $\{C_{n,k}=\frac{n-k}{n}\binom{n-1+k}{k}\}_{0\le k\le n-1}$ forms the {\em Catalan triangle} (see \cite[A009766]{oeis}). Thus, $\mathsf{iar}$ is a Comtet statistic over $321$-avoiding permutations.
Recently, Adin, Bagno, and Roichman~\cite{adin} proved a generalization of~\eqref{eq:cat}, which is equivalent to the equidistribution of $(\mathrm{LMAX},\mathsf{iar})$ and $(\mathrm{LMAX},\mathsf{comp})$ on $321$-avoiding permutations. This result was shown to imply that the class of $321$-avoiding permutations with a prescribed number of components is {\em Schur positive}. Rubey~\cite{rub} later found a symmetric generalization of the Adin--Bagno--Roichman equidistribution via constructing an involution on Dyck paths and using Krattenthaler's bijection~\cite{kra} from Dyck paths to $321$-avoiding permutations. His symmetric equidistribution was shown~\cite{flw} to be equivalent to the equidistribution of $(\mathrm{LMAX},\mathsf{iar},\mathsf{comp})$ and $(\mathrm{LMAX},\mathsf{comp},\mathsf{iar})$ on $321$-avoiding permutations, up to some elementary transformations on permutations. Since the classical {\em Knuth--Richards bijection} (see~\cite{CK}) between $\mathfrak{S}_n(321)$ and $\mathfrak{S}_n(312)$ preserves the triple $(\mathrm{LMAX},\mathsf{iar},\mathsf{comp})$ and the fact that $\mathrm{DESB}(\pi)\cup\mathrm{LMAX}(\pi)=[n]$ for $\pi\in\mathfrak{S}_n(312)$, the two quadruples
$$
(\mathrm{LMAX},\mathrm{DESB},\mathsf{iar},\mathsf{comp})\quad \text{and}\quad (\mathrm{LMAX},\mathrm{DESB},\mathsf{comp},\mathsf{iar})
$$
have the same distribution over $\mathfrak{S}_n(312)$.
On the other hand, Claesson, Kitaev, and Steingr\'imsson~\cite[Thm~2.2.48]{kit} constructed a bijection between separable permutations of length $n+1$ with $k+1$ components and {\em Schr\"oder paths} of length $2n$ with $k$ horizontals on the $x$-axis. Combining this bijection with the recent work in~\cite[Thm~3.2]{fw} justifies $\mathsf{iar}$ being a Comtet statistic on separable permutations.
The above results lead us to find Theorem~\ref{thm:sep:sym}. See also~\cite{flw} for other interesting consequences of Theorem~\ref{thm:sep:sym}. For other studies of pattern avoiding permutations that emphasize bijective maps, the reader is referred to~\cite{CK,EP,LK,SS}.
The rest of this paper is mainly devoted to the proof of Theorem~\ref{thm:sep:sym}. In the next section, we transform the involved statistics from separable permutations to di-sk trees, and then in section~\ref{sec:constr of phi}, we construct a combinatorial bijection on di-sk trees to build the involution $\Phi$ for Theorem~\ref{thm:sep:sym}. In section~\ref{sec:tree-tra}, we discuss some further results from the perspective of tree traversal, and derive several new Comtet statistics over di-sk trees. Finally, we conclude our paper by posing several questions for further investigation.
\section{From separable permutations to di-sk trees}
The set of all di-sk trees with $n-1$ nodes is denoted as $\mathfrak{DT}_n$. For each $T\in\mathfrak{DT}_n$, we use the \emph{inorder} (traversal) to compare nodes on $T$: starting with the root node, we recursively traverse the left subtree to the parent then to the right subtree if any (see the first tree in Fig.~\ref{8traversal}). We call the first (by inorder) node of $T$ the \emph{inorder root} (abbreviated as {\em iroot} in the sequel) of $T$.
We will apply a natural bijection $\eta:\mathfrak{S}_n(2413,3142)\rightarrow \mathfrak{DT}_n$ found by Shapiro and Stephens~\cite{shapi}. The recursive description of $\eta$ recalled below is from~\cite{flz}. Let $\mathrm{id}_1=1$ be the unique permutation of length $1$ and we set $\eta(\mathrm{id}_1)=\emptyset$. For $\pi=\pi_1\ldots\pi_n\in\mathfrak{S}_n(2413,3142)$ with $n\geq2$, find the greatest index $i\in[n-1]$ such that either
\begin{equation}\label{eq:greind}
\min\{\pi_1,\ldots,\pi_i\}>\max\{\pi_{i+1},\ldots,\pi_n\}\quad\text{or}\quad\max\{\pi_1,\ldots,\pi_i\}<\min\{\pi_{i+1},\ldots,\pi_n\}.
\end{equation}
In view of Proposition~\ref{desides}, such an index $i$ exists and is unique. We distinguish two cases:
\begin{itemize}
\item If the first inequality in \eqref{eq:greind} holds, then $\pi=\omega\ominus\rho$ with $\omega=(\pi_1+i-n)\cdots(\pi_i+i-n)\in\mathfrak{S}_i(2413,3142)$ and $\rho=\pi_{i+1}\cdots\pi_{n}\in\mathfrak{S}_{n-i}(2413,3142)$. Define $\eta(\pi)=(\eta(\omega),\ominus,\eta(\rho))$, the tree with the left subtree $\eta(\omega)$ and the right subtree $\eta(\rho)$ attached to the root $\ominus$.
\item Otherwise, $\pi=\omega\oplus\rho$, where $\omega=\pi_1\cdots\pi_i\in\mathfrak{S}_i(2413,3142)$ and $\rho=(\pi_{i+1}-i)\cdots(\pi_{n}-i)\in\mathfrak{S}_{n-i}(2413,3142)$. Then define $\eta(\pi)$ to be the tree with the left subtree $\eta(\omega)$ and the right subtree $\eta(\rho)$ attached to the root $\oplus$.
\end{itemize}
See Fig.~\ref{di-sk} for an example of $\eta$ for $\pi=5\,2\,3\,4\,1\,9\,11\,10\,6\,8\,7\in\mathfrak{S}_{11}(2413,3142)$. One of the important features of $\eta$ was proved in~\cite{flz}.
\begin{figure}
\begin{tikzpicture}[scale=0.3]
\draw[-] (17,10) to (19,12);
\draw[-] (20,12) to (22,10);
\draw[-] (22,9) to (20,7);
\draw[-] (23,9) to (25,7);
\draw[-] (20,6) to (22,4);
\draw[-] (26,6) to (28,4);
\draw[-] (16,9) to (14,7);
\node at (13.5,6.5) {$\ominus$};
\textcolor{red}{\draw[-] (13,6) to (12,5);}
\textcolor{blue}{\node at (11.5,4.5) {$5$};}
\draw[-] (14,6) to (16,4);
\node at (16.5,3.5) {$\oplus$};
\textcolor{red}{\draw[-] (17,3) to (18,2);}
\textcolor{blue}{\node at (18.5,1.5) {$4$};}
\draw[-] (16,3) to (14,1);
\node at (13.5,0.5) {$\oplus$};
\textcolor{red}{\draw[-] (13,0) to (12,-1);}
\textcolor{blue}{\node at (11.5,-1.5) {$2$};}
\textcolor{red}{\draw[-] (14,0) to (15,-1);}
\textcolor{blue}{\node at (15.5,-1.5) {$3$};}
\node at (16.5,9.5) {$\ominus$};
\textcolor{red}{\draw[-] (17,9) to (18,8);}
\textcolor{blue}{\node at (18.5,7.5) {$1$};}
\node at (19.5,12.5) {$\oplus$};
\node at (22.5,9.5) {$\ominus$};
\node at (19.5,6.5) {$\oplus$};
\textcolor{red}{\draw[-] (19,6) to (18,5);}
\textcolor{blue}{\node at (17.5,4.5) {$9$};}
\node at (25.5,6.5) {$\oplus$};
\textcolor{red}{\draw[-] (25,6) to (24,5);}
\textcolor{blue}{\node at (23.5,4.5) {$6$};}
\node at (22.5,3.5) {$\ominus$};
\textcolor{red}{\draw[-] (22,3) to (21,2);}
\textcolor{blue}{\node at (20.5,1.5) {$11$};}
\textcolor{red}{\draw[-] (23,3) to (24,2);}
\textcolor{blue}{\node at (24.5,1.5) {$10$};}
\node at (28.5,3.5) {$\ominus$};
\textcolor{red}{\draw[-] (29,3) to (30,2);}
\textcolor{blue}{\node at (30.5,1.5) {$7$};}
\textcolor{red}{\draw[-] (28,3) to (27,2);}
\textcolor{blue}{\node at (26.5,1.5) {$8$};}
\end{tikzpicture}
\caption{The bijection $\eta:\mathfrak{S}_n(2413,3142)\rightarrow \mathfrak{DT}_{n}$.\label{di-sk}}
\end{figure}
\begin{lemma}[Theorem~2.3 in~\cite{flz}]
\label{lem:sep1}
The mapping
$\eta: \mathfrak{S}_n(2413,3142)\rightarrow\mathfrak{DT}_{n}$ is a bijection such that
\begin{align}\label{minu:des}
i\in \mathrm{DES}(\pi) \,\,\Longleftrightarrow \text{ the $i$th node (by inorder) of $\eta(\pi)$ is $\ominus$}
\end{align}
for each $\pi\in\mathfrak{S}_n(2413,3142)$.
\end{lemma}
For each tree $T\in\mathfrak{DT}_n$, let $\mathsf{iop}(T)$ be the number of {\em initial $\oplus$-nodes} (by inorder) in $T$.
It follows from~\eqref{minu:des} that
\begin{equation}\label{eq:iop}
\mathsf{iar}(\pi)-1=\mathsf{iop}(\eta(\pi))
\end{equation}
for any $\pi\in\mathfrak{S}_n(2413,3142)$.
But what is $\mathsf{comp}(\pi)$ corresponding to in the di-sk tree $\eta(\pi)$? Let us consider the {\em spine} of $T$, i.e., the path from the root of $T$ to the iroot of $T$. Let $\mathsf{top}(T)$ be the number of top consecutive $\oplus$-nodes in the spine of $T$. For instance, the spine of $T$ in Fig.~\ref{di-sk} is $\oplus-\ominus-\ominus$ (from the top) and so $\mathsf{top}(T)=1$.
\textcolor{red}{
}
\begin{lemma}\label{lem:sep2}
For each permutation $\pi\in\mathfrak{S}_n(2413,3142)$, we have
$$\mathsf{comp}(\pi)-1=\mathsf{top}(\eta(\pi)).
$$
\end{lemma}
\begin{proof}
Recall that $\mathsf{comp}(\pi)-1$ equals the cardinality of the set $\{k\in [n-1]: \forall j\leq k,\: \pi_j\leq k\}$.
We need to consider two cases:
\begin{itemize}
\item If $\{k\in [n-1]: \forall j\leq k,\: \pi_j\leq k\}=\emptyset$, then the root of $\eta(\pi)$ is a $\ominus$-node and so $\mathsf{top}(\eta(\pi))=0=\mathsf{comp}(\pi)-1$.
\item Otherwise, let $l$ be the greatest integer in $\{k\in [n-1]: \forall j\leq k,\: \pi_j\leq k\}$. Clearly, $l$ is the greatest index smaller than $n$ such that~\eqref{eq:greind} holds. Thus, by the construction of $\eta$ we have $\eta(\pi)=(\eta(\omega),\oplus,\eta(\rho))$ assuming that $\pi=\omega\oplus\rho$ with $\omega=\pi_1\cdots\pi_l$. It then follows by induction on $n$ that
$$
\mathsf{comp}(\pi)-1=\mathsf{comp}(\omega)=1+\mathsf{top}(\omega)=\mathsf{top}(\pi).
$$
\end{itemize}
In either case, the assertion is true.
\end{proof}
\begin{remark}
The statistic $\mathsf{top}$ on di-sk trees was previously considered by Corteel, Martinez, Savage and Weselcouch in~\cite[Corollary~5]{cor}, where they constructed a bijection from $021$-avoiding inversion sequences of length $n$ with $k$ initial zeros to $\{T\in\mathfrak{DT}_n: \mathsf{top}(T)=k-1\}$.
On the other hand, Kim and Lin~\cite{LK} built a bijection from $021$-avoiding inversion sequences to $(2413,4213)$-avoiding permutations which transforms positions of ascents to positions of descents. Combining these two bijections gives
\begin{equation}\label{eq:2413-4213}
|\{T\in\mathfrak{DT}_n: \mathsf{top}(T)=k-1\}|=|\{\pi\in\mathfrak{S}_n(2413,4213): \mathsf{iar}(\pi)=k\}|.
\end{equation}
It would be interesting to construct a natural bijection (probably in similar flavor as $\eta$) between $\mathfrak{DT}_n$ and $\mathfrak{S}_n(2413,4213)$ that proves~\eqref{eq:2413-4213}.
See also section~\ref{sec:tree-tra} for another interpretation of $\mathsf{top}$ in terms of tree traversal.
\end{remark}
\section{The construction of \texorpdfstring{$\Phi$}{Phi}}\label{sec:constr of phi}
This section is devoted to the construction of $\Phi$. We begin with an elementary operation on di-sk tree that will be used frequently during our construction of $\Phi$.
Let $T$ be a di-sk tree and $v$ be an $\ominus$-node of $T$. We introduce the di-sk tree $\mathcal{L}(v,T)$ whenever there exists an $\oplus$-node $w$ ($w$ is not a right child) in the following two situations:
\begin{itemize}
\item $v$ is the left child of $w$.
\item $v$ is the right child of an $\oplus$-node, denoted $w'$, whose parent is $w$.
\end{itemize}
In either case, define $\mathcal{L}(v,T)$ to be the di-sk tree obtained from $T$ by cutting the $\oplus$-node $w$, together with its right subtree (if any), and inserting it as the left child of $v$. The original left child of $v$ (if any) becomes the left child of $w$, while the original right parent of $w$ (if any) becomes the right parent of $v$ (resp.~$w'$) for the first case (resp.~the second case), keeping the remaining nodes and edges of $T$ unchanged. See Fig.~\ref{left-opera} for the illustration of $\mathcal{L}(v,T)$ in the above two cases. Since the edges we have inserted/deleted in the construction of $\mathcal{L}(v,T)$ are all left edges, we see that $\mathcal{L}(v,T)$ is still a di-sk tree. Moreover, both cases are seen to be invertible and if $T':=\mathcal{L}(v,T)$, we will denote the inverse map as $\mathcal{L}^{-1}(v,T')=T$.
\begin{figure}
\begin{tikzpicture}[scale=0.3]
\draw[-,dashed] (7.7,12.7) to (10,15);\draw[-] (5.5,10.5) to (7.3,12.3);
\draw[-] (7.5,12.5) to (9.1,9.9);\draw[-] (7.5,12.5) to (11.1,11.9);\draw[-] (9.1,9.9) to (11.1,11.9);
\node at (9.3,11.4) {$C$};
\node at (5,10) {$\oplus$};\node at (4,10.5) {$w$};\draw[-] (4.5,9.5) to (3,8);
\draw[-] (5.4,9.6) to (7,7);\draw[-] (5.4,9.6) to (9,9);\draw[-] (7,7) to (9,9);
\node at (7.2,8.5) {$A$};
\node at (2.5,7.5) {$\ominus$};\node at (1.5,8) {$v$};\draw[-,dashed] (-0.2,4.8) to (-3,2);\draw[-] (0.2,5.2) to (2,7);
\draw[-] (0,5) to (1.6,2.4);\draw[-] (0,5) to (3.6,4.4);\draw[-] (1.6,2.4) to (3.6,4.4);
\node at (1.8,3.9) {$D$};
\draw[-] (2.9,7.1) to (4.5,4.5);\draw[-] (2.9,7.1) to (6.5,6.5);\draw[-] (4.5,4.5) to (6.5,6.5);
\node at (4.7,6) {$B$};
\node at (1.5,1.5) {$T$};\node at (11.7,1.5) {$\mathcal{L}(v,T)$};
\node at (11,10) {$\longrightarrow$};
\draw[-,dashed] (17.7,12.7) to (20,15);\draw[-] (15.5,10.5) to (17.3,12.3);
\draw[-] (17.5,12.5) to (19.1,9.9);\draw[-] (17.5,12.5) to (21.1,11.9);\draw[-] (19.1,9.9) to (21.1,11.9);
\node at (19.3,11.4) {$C$};
\node at (15,10) {$\ominus$};\node at (14,10.5) {$v$};\draw[-] (14.5,9.5) to (13,8);
\draw[-] (15.4,9.6) to (17,7);\draw[-] (15.4,9.6) to (19,9);\draw[-] (17,7) to (19,9);
\node at (17.2,8.5) {$B$};
\node at (12.5,7.5) {$\oplus$};\node at (11.5,8) {$w$};\draw[-,dashed] (9.8,4.8) to (7,2);\draw[-] (10.2,5.2) to (12,7);
\draw[-] (10,5) to (11.6,2.4);\draw[-] (10,5) to (13.6,4.4);\draw[-] (11.6,2.4) to (13.6,4.4);
\node at (11.8,3.9) {$D$};
\draw[-] (12.9,7.1) to (14.5,4.5);\draw[-] (12.9,7.1) to (16.5,6.5);\draw[-] (14.5,4.5) to (16.5,6.5);
\node at (14.7,6) {$A$};
\draw[-,dashed] (29.8,12.8) to (32,15);\draw[-] (27.5,10.5) to (29.3,12.3);
\draw[-] (29.5,12.5) to (31.1,9.9);\draw[-] (29.5,12.5) to (33.1,11.9);\draw[-] (31.1,9.9)to (33.1,11.9);
\node at (31.3,11.4) {$C$};
\node at (27,10) {$\oplus$};\node at (26,10.5) {$w$};\draw[-] (26.5,9.5) to (25,8);
\draw[-] (27.4,9.6) to (29,7);\draw[-] (27.4,9.6) to (31,9);\draw[-] (29,7) to (31,9);
\node at (29.2,8.5) {$A$};
\node at (24.5,7.5) {$\oplus$};\node at (23.5,8) {$w'$};\draw[-,dashed] (24,7) to (21,4);
\draw[-] (25,7) to (26.5,5.5);\node at (27,5) {$\ominus$};\draw[-,dashed] (24.3,2.3) to (21.5,-0.5);
\node at (26,5) {$v$};
\draw[-] (26.5,4.5) to (24.7,2.7);
\draw[-] (24.5,2.5) to (26.1,-0.1);\draw[-] (24.5,2.5) to (28.1,1.9);\draw[-] (26.1,-0.1) to (28.1,1.9);
\node at (26.3,1.4) {$D$};
\draw[-] (27.4,4.6) to (29,2);\draw[-] (27.4,4.6) to (31,4);\draw[-] (29,2) to (31,4);
\node at (29.2,3.5) {$B$};
\node at (26,-1) {$T$};
\node at (33.5,10) {$\longrightarrow$};
\draw[-,dashed] (40.8,12.8) to (43,15);\draw[-] (38.5,10.5) to (40.3,12.3);
\draw[-] (40.5,12.5) to (42.1,9.9);\draw[-] (40.5,12.5) to (44.1,11.9);\draw[-] (42.1,9.9)to (44.1,11.9);
\node at (42.3,11.4) {$C$};
\node at (38,10) {$\oplus$};\node at (37,10.5) {$w'$};\draw[-,dashed] (37.5,9.5) to (34.5,6.5);
\draw[-] (38.5,9.5) to (40,8);\node at (40.5,7.5) {$\ominus$};\node at (39.3,7.5) {$v$};
\draw[-] (40.9,7.1) to (42.5,4.5);\draw[-] (40.9,7.1) to (44.5,6.5);\draw[-] (42.5,4.5) to (44.5,6.5);
\node at (42.7,6) {$B$};
\draw[-] (37.5,4.5) to (35.7,2.7);
\draw[-] (35.5,2.5) to (37.1,-0.1);\draw[-] (35.5,2.5) to (39.1,1.9);\draw[-] (37.1,-0.1) to (39.1,1.9);
\node at (37.3,1.4) {$D$};
\draw[-] (40,7) to (38.5,5.5);\node at (38,5) {$\oplus$};\draw[-,dashed] (35.5,2.5) to (32.5,-0.5);
\node at (36.7,5) {$w$};
\draw[-] (38.4,4.6) to (40,2);\draw[-] (38.4,4.6) to (42,4);\draw[-] (40,2) to (42,4);
\node at (40.2,3.5) {$A$};
\node at (37,-1) {$\mathcal{L}(v,T)$};
\end{tikzpicture}
\caption{Two cases to obtain the di-sk tree $\mathcal{L}(v,T)$ from $T$.\label{left-opera}}
\end{figure}
The reason to introduce the transformation $\mathcal{L}$ lies in the following lemma.
\begin{lemma}\label{lem:L}
Let $\pi=\eta^{-1}(T)$ and $\pi'=\eta^{-1}(\mathcal{L}(v,T))$ for a di-sk tree $T$.
Then,
\begin{align*}
\mathrm{DESB}(\pi)&=\mathrm{DESB}(\pi'),\\
\mathrm{LMAX}(\pi)&=\mathrm{LMAX}(\pi')\quad{and}\\
\mathrm{LMIN}(\pi)&=\mathrm{LMIN}(\pi').
\end{align*}
\end{lemma}
\begin{proof}
We need to describe the inverse $\eta^{-1}: \mathfrak{DT}_{n} \rightarrow \mathfrak{S}_n(2413,3142)$ of $\eta$. For a given di-sk tree $
T\in\mathfrak{DT}_n$, let us add some edges to $T$ so that the out-degree of each node is exactly two (see the red edges in Fig.~\ref{di-sk}). There are $n$ new edges to be added to $T$, thus creating $n$ new leaves. The next step of $\eta^{-1}$ is to assign the integers in $[n]$ to these $n$ new leaves so that for each $\oplus$-node (resp.~$\ominus$-node) of $T$, the integers assigned to the leaves belonging to the left subtree of this node are all smaller (resp.~greater) than those assigned to leaves belonging to the right subtree. Such an assignment is unique and the permutation $\eta^{-1}(T)$ can be derived from reading these $n$ integers by the inorder of this augmented tree (see Fig.~\ref{di-sk}).
From the above description of $\eta^{-1}$, we see that the transformation $\mathcal{L}$ preserves the assignment of the augmented tree, namely, if a new leaf of $T$ has been assigned an integer $k$, then the corresponding new leaf (i.e., the leaf of $\mathcal{L}(v,T)$ added to the corresponding node under $\mathcal{
L}$) receives the same integer $k$. Notice that an integer is a descent bottom of $\pi=\eta^{-1}(T)$ if and only if it is assigned to a leaf appearing immediately after an $\ominus$-node by the inorder of the augmented tree of $T$. Thus, $\mathrm{DESB}(\pi)=\mathrm{DESB}(\pi')$ holds. To see that $\mathrm{LMAX}(\pi)=\mathrm{LMAX}(\pi')$, we divide letters assigned to the augmented tree of $T=\eta(\pi)$ (see Fig.~\ref{left-opera}) into three subclasses:
\begin{enumerate}
\item[I \;] letters assigned to the right subtree $A$;
\item[II ] letters assigned to the right subtree $B$;
\item[III] all the remaining letters.
\end{enumerate}
We observe that: 1) all letters in class I are greater than all letters in class II; 2) none of the letters in class II is a left-to-right maximum. Now $\pi'$ is obtained from $\pi$ by swapping class I letters with class II letters, therefore the status of being a left-to-right maximum or not remains the same for each letter in classes I, II, and III. So $\mathrm{LMAX}(\pi)=\mathrm{LMAX}(\pi')$ as desired. The proof of $\mathrm{LMIN}(\pi)=\mathrm{LMIN}(\pi')$ is similar by noting that none of the letters in class I is a left-to-right minimum.
\end{proof}
Let $\widetilde\mathfrak{DT}_n$ be the set of trees $T\in\mathfrak{DT}_n$ such that the spine of $T$ has at least one $\ominus$-node. The next result contains the main ingredient for our construction of $\Phi$.
\begin{theorem}\label{thm:di-sk}
Let $\widetilde\mathfrak{DT}_n^{(k,l)}:=\{T\in\widetilde\mathfrak{DT}_n: \mathsf{top}(T)=k, \mathsf{iop}(T)=l\}$. If $k\geq1$, then there exists a bijection $\phi:\widetilde\mathfrak{DT}_n^{(k,l)}\rightarrow \widetilde\mathfrak{DT}_n^{(k-1,l+1)}$ satisfying
\begin{equation}\label{desb:lmax}
(\mathrm{DESB},\mathrm{LMAX},\mathrm{LMIN})\, \eta^{-1}(T)= (\mathrm{DESB},\mathrm{LMAX},\mathrm{LMIN})\, \eta^{-1}(\phi(T))
\end{equation}
for any $T\in\widetilde\mathfrak{DT}_n^{(k,l)}$.
\end{theorem}
\begin{figure}
\begin{tikzpicture}[scale=0.4]
\node at (10,30) {$\bullet$};\draw[-,dashed] (10,30) to (11,29);
\draw[-] (10,30) to (9,29);
\node at (9,29) {$\bullet$};\draw[-,dashed] (9,29) to (10,28);
\draw[-] (9,29) to (8,28);
\node at (8,28) {$\bullet$};\draw[-] (8,28) to (8.5,26.5);\draw[-] (8.5,26.5) to (9.5,27.5);\draw[-] (8,28) to (9.5,27.5);
\node at (8.7,27.3) {$\alpha$};
\draw[-] (8,28) to (7,27);\node at (6.85,26.85) {$\circ$};\node at (6.2,27.2) {$v_1$};\draw[-,dashed] (6.95,26.74) to (7.8,25.85);
\draw[-] (6.7,26.75) to (5.7,25.75);\node at (5.7,25.75) {$\bullet$};\draw[-,dashed] (5.7,25.75) to (6.7,24.75);
\draw[-] (5.7,25.75) to (4.7,24.75);
\node at (4.7,24.75) {$\bullet$};\draw[-] (4.7,24.75) to (5.2,23.25);\draw[-] (5.2,23.25) to (6.2,24.25);
\draw[-] (6.2,24.25) to (4.7,24.75);
\node at (5.4,24) {$\beta$};
\draw[-] (4.7,24.75) to (3.7,23.75);\node at (3.55,23.6) {$\circ$};\node at (2.9,24) {$v_2$};\draw[-,dashed] (3.7,23.5) to (4.6,22.6);
\draw[-] (3.4,23.5) to (2.4,22.5);\node at (2.4,22.5) {$\bullet$};
\draw[-] (2.4,22.5) to (2.9,21);\draw[-] (2.9,21) to (3.9,22);\draw[-] (2.4,22.5) to (3.9,22);
\node at (3,21.8) {$\gamma$};
\draw[-] (2.4,22.5) to (1.2,21.3);\node at (1.2,21.3) {$\bullet$};
\draw[-] (1.2,21.3) to (0.2,20.3);\node at (0.2,20.3) {$\bullet$};
\draw[-] (0.2,20.3) to (-0.8,19.3);\node at (-0.8,19.3) {$\bullet$};
\draw[-] (-0.8,19.3) to (-1.8,18.3);\node at (-1.8,18.3) {$\bullet$};
\draw[-] (1.2,21.3) to (2.2,20.3);\node at (2.35,20.15) {$\circ$};\node at (1.7,20.2) {$v_3$};\draw[-,dashed] (2.46,20.04) to (3.45,19.05);
\draw[-] (2.23,20.03) to (1.2,19);\node at (1.32,19.12) {$\bullet$};\draw[-,dashed] (1.32,19.12) to (2.32,18.12);
\draw[-] (1.32,19.12) to (0.2,18);\node at (0.2,18) {$\bullet$};
\draw[-] (0.2,18) to (0.7,16.5);\draw[-] (0.7,16.5) to (1.7,17.5);\draw[-] (0.2,18) to (1.7,17.5);
\node at (0.9,17.2) {$\delta$};
\node at (5.9,18) {$\longrightarrow$};\node at (5.9,19) {Step 1};
\node at (16.5,18) {$\longrightarrow$};\node at (16.5,19) {Step 2};
\draw[-] (0.2,18) to (-0.6,17.2);\node at (-0.74,17.06) {$\circ$};\node at (-1.4,17.4) {$v_4$};\draw[-,dashed] (-0.62,16.94) to (0.36,15.96);
\draw[-] (-0.86,16.94) to (-1.86,15.94);\node at (-1.86,15.94) {$\bullet$};
\draw[-] (-1.86,15.94) to (-0.86,14.94);\node at (-0.72,14.8) {$\circ$};\node at (-1.4,14.9) {$v_5$};\draw[-,dashed] (-0.6,14.68) to (0.28,13.8);
\draw[-] (-0.84,14.68) to (-1.84,13.68);\node at (-1.84,13.68) {$\bullet$};\draw[-,dashed] (-1.84,13.68) to (-0.84,12.68);
\draw[-] (-1.84,13.68) to (-2.84,12.68);\node at (-2.84,12.68) {$\bullet$};\node at (-3.4,12.8) {$w$};
\draw[-] (-2.84,12.68) to (-3.84,11.68);\node at (-3.98,11.54) {$\circ$};\node at (-4.5,11.8) {$v$};\draw[-] (-4.1,11.42) to (-4.98,10.54);\node at (-4.98,10.54) {$\bullet$};\draw[-,dashed] (-3.86,11.42) to (-2.98,10.54);
\draw[-] (-2.84,12.68) to (-1.34,11.18);\node at (-1.2,11.04) {$\circ$};
\draw[-] (-1.32,10.92) to (-2.2,10.04);
\draw[-] (-2.2,10.04) to (-3.2,9.04);\node at (-3.34,8.9) {$\circ$};\draw[-] (-3.46,8.78) to (-4.46,7.78);
\textcolor{red}{
\node at (-2.2,10.04) {$\bullet$};
\draw[-] (-2.2,10.04) to (-1.2,9.04);\node at (-1.06,8.9) {$\circ$};\draw[-] (-0.94,8.78) to (-0.06,7.78);\node at (-0.06,7.78) {$\bullet$};
\draw[-] (-1.18,8.78) to (-2.18,7.78);\node at (-2.32,7.64) {$\circ$};
}
\textcolor{blue}{
\node at (-4.46,7.78){$\bullet$};\draw[-] (-4.46,7.78) to (-3.46,6.78);\node at (-3.32,6.64) {$\circ$};
\draw[-] (-3.44,6.52) to (-4.44,5.52);\node at (-4.44,5.52){$\bullet$};
\draw[-] (-4.46,7.78) to (-5.46,6.78);\node at (-5.46,6.78){$\bullet$};
}
\textcolor{magenta}{
\draw[-,dashed] (-3,13.5) to (-4,12.5);
\draw[-,dashed] (-3,13.5) to (1.8,8.7);\draw[-,dashed] (1.8,8.7) to (-3.2,3.7);
\draw[-,dashed] (-6.2,6.7) to (-3.2,3.7);\draw[-,dashed] (-6.2,6.7) to (-2.2,10.7);
\draw[-,dashed] (-4,12.5) to (-2.2,10.7);
}
\node at (-3,2.5){$T$};\node at (7,2.5){$T'$};\node at (17,2.5){$\phi(T)$};\node at (6.5,4.5){\textcolor{magenta}{$B$}};\node at (-3,5){$\textcolor{magenta}{B}$};
\node at (20,30) {$\bullet$};\draw[-,dashed] (20,30) to (21,29);
\draw[-] (20,30) to (19,29);
\node at (19,29) {$\bullet$};\draw[-,dashed] (19,29) to (20,28);
\draw[-] (19,29) to (18,28);\node at (17.86,27.86) {$\circ$};\node at (17.2,28) {$v_1$};\draw[-,dashed] (17.98,27.74) to (18.86,26.86);
\draw[-] (17.74,27.74) to (16.86,26.86);
\node at (16.86,26.86){$\bullet$};\draw[-] (16.86,26.86) to (17.36,25.36);\draw[-] (17.36,25.36) to (18.36,26.36);
\draw[-] (16.86,26.86) to (18.36,26.36);
\node at (17.56,26.16) {$\alpha$};
\draw[-] (16.86,26.86) to (15.86,25.86);\node at (15.86,25.86){$\bullet$};
\draw[-] (15.86,25.86) to (14.86,24.86);\draw[-,dashed] (15.86,25.86) to(16.86,24.86);
\node at (14.72,24.72) {$\circ$};\node at (14,25) {$v_2$};\draw[-,dashed] (14.84,24.6) to(15.74,23.7) ;
\draw[-] (14.6,24.6) to (13.72,23.72);\node at (13.72,23.72){$\bullet$};
\draw[-] (13.72,23.72) to (14.22,22.22);\draw[-] (14.22,22.22) to (15.22,23.22);\draw[-] (13.72,23.72) to (15.22,23.22);
\node at (14.42,22.9) {$\beta$};
\draw[-] (13.72,23.72) to (12.5,22.5);\node at (12.5,22.5){$\bullet$};
\draw[-] (12.5,22.5) to (11.5,21.5);\node at (11.5,21.5){$\bullet$};
\draw[-] (11.5,21.5) to (10.5,20.5);\node at (10.5,20.5){$\bullet$};
\draw[-] (10.5,20.5) to (9.5,19.5);\node at (9.5,19.5){$\bullet$};
\draw[-] (12.5,22.5) to (13.5,21.5);\node at (13.64,21.36) {$\circ$};\node at (13,21.4) {$v_3$};\draw[-,dashed] (13.78,21.22) to(14.64,20.36);
\draw[-] (13.52,21.24) to (12.64,20.36);\node at (12.64,20.36){$\bullet$};
\draw[-] (12.64,20.36) to (13.14,18.86);\draw[-] (13.14,18.86) to (14.14,19.86);\draw[-] (12.64,20.36) to (14.14,19.86);
\node at (13.24,19.66) {$\gamma$};
\draw[-] (12.64,20.36) to (11.64,19.36);\node at (11.64,19.36){$\bullet$};\draw[-,dashed] (11.64,19.36)to(12.64,18.36);
\draw[-] (11.64,19.36) to (10.64,18.36);\node at (10.5,18.22) {$\circ$};\node at (9.9,18.6) {$v_4$};\draw[-,dashed] (10.62,18.12)to(11.5,17.22);
\draw[-] (10.38,18.1) to (9.5,17.22);\node at (9.5,17.22){$\bullet$};
\draw[-] (9.5,17.22) to (10.5,16.22);\node at (10.64,16.08) {$\circ$};\node at (9.98,16.12) {$v_5$};\draw[-,dashed] (10.78,15.94)to(11.64,15.08);
\draw[-] (10.52,15.96) to (9.64,15.08);\node at (9.64,15.08){$\bullet$};
\draw[-] (9.64,15.08) to (10.14,13.58);\draw[-] (10.14,13.58) to (11.14,14.58);\draw[-] (9.64,15.08)to (11.14,14.58);
\node at (10.34,14.28) {$\delta$};
\draw[-] (9.64,15.08) to (8.64,14.08);\node at (8.64,14.08){$\bullet$};
\draw[-,dashed] (8.64,14.08) to (9.64,13.08);
\draw[-] (8.64,14.08) to (7.64,13.08);\node at (7.5,12.94) {$\circ$};\node at (7,13.2) {$v$};
\draw[-,dashed] (7.62,12.82) to (8.5,11.94);
\draw[-] (7.38,12.82) to (6.5,11.94);\node at (6.5,11.94){$\bullet$};\node at (5.9,12.2) {$w$};
\draw[-] (6.5,11.94) to (5.5,10.94);\node at (5.5,10.94){$\bullet$};
\draw[-] (6.5,11.94) to (8,10.54);\node at (8.14,10.4) {$\circ$};\node at (8.65,10) {$v_1'$};
\draw[-] (8.02,10.28) to (7.14,9.4);
\draw[-] (7.14,9.4) to (6.14,8.4);\node at (6,8.26) {$\circ$};\node at (6.6,8) {$v_2'$};\draw[-] (5.88,8.14) to (4.88,7.14);
\textcolor{red}{
\node at (7.14,9.4) {$\bullet$};
\draw[-] (7.14,9.4) to (8.14,8.4);\node at (8.28,8.26) {$\circ$};\draw[-] (8.4,8.14) to (9.28,7.14);\node at (9.28,7.14) {$\bullet$};
\draw[-] (8.16,8.14) to (7.16,7.14);\node at (7.02,7) {$\circ$};
}
\textcolor{blue}{
\node at (4.88,7.14){$\bullet$};\draw[-] (4.88,7.14) to (5.88,6.14);\node at (6.02,6) {$\circ$};
\draw[-] (5.9,5.88) to (4.9,4.88);\node at (4.9,4.88){$\bullet$};
\draw[-] (4.88,7.14) to (3.88,6.14);\node at (3.88,6.14){$\bullet$};
}
\textcolor{magenta}{
\draw[-,dashed] (6.34,12.86) to (5.34,11.86);
\draw[-,dashed] (6.34,12.86) to (11.14,8.06);\draw[-,dashed] (11.14,8.06) to (6.14,3.06);
\draw[-,dashed] (3.14,6.06) to (6.14,3.06);\draw[-,dashed] (3.14,6.06) to (7.14,10.06);
\draw[-,dashed] (5.34,11.86) to (7.14,10.06);
}
\node at (30,30) {$\bullet$};\draw[-,dashed] (30,30) to (31,29);
\draw[-] (30,30) to (29,29);
\node at (29,29) {$\bullet$};\draw[-,dashed] (29,29) to (30,28);
\draw[-] (29,29) to (28,28);\node at (27.86,27.86) {$\circ$};\draw[-,dashed] (27.98,27.74) to (28.86,26.86);
\draw[-] (27.74,27.74) to (26.86,26.86);
\node at (26.86,26.86){$\bullet$};\draw[-] (26.86,26.86) to (27.36,25.36);\draw[-] (27.36,25.36) to (28.36,26.36);
\draw[-] (26.86,26.86) to (28.36,26.36);
\node at (27.56,26.16) {$\alpha$};
\draw[-] (26.86,26.86) to (25.86,25.86);\node at (25.86,25.86){$\bullet$};
\draw[-] (25.86,25.86) to (24.86,24.86);\draw[-,dashed] (25.86,25.86) to(26.86,24.86);
\node at (24.72,24.72) {$\circ$};\draw[-,dashed] (24.84,24.6) to(25.74,23.7) ;
\draw[-] (24.6,24.6) to (23.72,23.72);\node at (23.72,23.72){$\bullet$};
\draw[-] (23.72,23.72) to (24.22,22.22);\draw[-] (24.22,22.22) to (25.22,23.22);\draw[-] (23.72,23.72) to (25.22,23.22);
\node at (24.42,22.9) {$\beta$};
\draw[-] (23.72,23.72) to (22.5,22.5);\node at (22.5,22.5){$\bullet$};
\draw[-] (22.5,22.5) to (21.5,21.5);\node at (21.5,21.5){$\bullet$};
\draw[-] (21.5,21.5) to (20.5,20.5);\node at (20.5,20.5){$\bullet$};
\draw[-] (20.5,20.5) to (19.5,19.5);\node at (19.5,19.5){$\bullet$};
\draw[-] (22.5,22.5) to (23.5,21.5);\node at (23.64,21.36) {$\circ$};\draw[-,dashed] (23.78,21.22) to(24.64,20.36);
\draw[-] (23.52,21.24) to (22.64,20.36);\node at (22.64,20.36){$\bullet$};
\draw[-] (22.64,20.36) to (23.14,18.86);\draw[-] (23.14,18.86) to (24.14,19.86);\draw[-] (22.64,20.36) to (24.14,19.86);
\node at (23.24,19.66) {$\gamma$};
\draw[-] (22.64,20.36) to (21.64,19.36);\node at (21.64,19.36){$\bullet$};\draw[-,dashed] (21.64,19.36)to(22.64,18.36);
\draw[-] (21.64,19.36) to (20.64,18.36);\node at (20.5,18.22) {$\circ$};\draw[-,dashed] (20.62,18.12)to(21.5,17.22);
\draw[-] (20.38,18.1) to (19.5,17.22);\node at (19.5,17.22){$\bullet$};
\draw[-] (19.5,17.22) to (20.5,16.22);\node at (20.64,16.08) {$\circ$};\draw[-,dashed] (20.78,15.94)to(21.64,15.08);
\draw[-] (20.52,15.96) to (19.64,15.08);\node at (19.64,15.08){$\bullet$};
\draw[-] (19.64,15.08) to (20.14,13.58);\draw[-] (20.14,13.58) to (21.14,14.58);\draw[-] (19.64,15.08)to (21.14,14.58);
\node at (20.34,14.28) {$\delta$};
\draw[-] (19.64,15.08) to (18.64,14.08);\node at (18.64,14.08){$\bullet$};
\draw[-,dashed] (18.64,14.08) to (19.64,13.08);
\draw[-] (18.64,14.08) to (17.64,13.08);\node at (17.5,12.94) {$\circ$};\node at (17,13.2) {$v$};
\draw[-,dashed] (17.62,12.82) to (18.5,11.94);
\draw[-] (17.38,12.82) to (16.5,11.94);\draw[-] (16.5,11.94) to (14.8,10.24);\node at (14.2,10.45) {$w$};\node at (14.8,10.24){$\bullet$};
\textcolor{red}{
\node at (16.5,11.94){$\bullet$};\draw[-] (16.5,11.94) to (17.5,10.94);\node at (17.64,10.8) {$\circ$};
\draw[-] (17.52,10.68) to (16.74,9.9);\node at (16.6,9.76) {$\circ$};\draw[-] (17.76,10.68) to (18.64,9.8);\node at (18.64,9.8){$\bullet$};
}
\draw[-] (14.8,10.24) to (13.8,9.24);\node at (13.8,9.24){$\bullet$};
\draw[-] (14.8,10.24) to (16.3,8.74);\node at (16.44,8.6){$\circ$};\node at (17.1,8.3) {$v_1'$};
\draw[-] (16.32,8.48) to (15.44,7.6);
\draw[-] (14.44,6.6) to (13.44,5.6);\node at (13.3,5.46){$\circ$};\node at (13.8,5.1) {$v_2'$};
\textcolor{blue}{
\node at (15.44,7.6){$\bullet$};\draw[-] (15.44,7.6) to (14.44,6.6);\node at (14.44,6.6){$\bullet$};
\draw[-] (15.44,7.6) to (16.44,6.6);\node at (16.58,6.46){$\circ$};
\draw[-] (16.46,6.34) to (15.58,5.46);\node at(15.58,5.46){$\bullet$};
}
\textcolor{magenta}{
\draw[-,dashed] (14,10) to (16.5,12.5);\draw[-,dashed] (16.5,12.5) to (20,9);\draw[-,dashed] (20,9) to (14.5,3.5);
\draw[-,dashed] (12.5,5.5) to (14.5,3.5);\draw[-,dashed] (12.5,5.5) to (15.5,8.5);\draw[-,dashed] (14,10) to (15.5,8.5);
}
\end{tikzpicture}
\caption{An illustration of $\phi$, where each $\oplus$-node (resp.~$\ominus$-node) in di-sk trees is replaced by a solid (resp.~hollow) circle, for simplicity.\label{bij:disk}}
\end{figure}
\begin{proof}For a fixed di-sk tree $T\in\widetilde\mathfrak{DT}_n^{(k,l)}$ with $k\geq1$,
the construction of $\phi(T)$ can be performed in the following two steps. In the first step, we do the ``{\bf swing down}'' on $T$ (see Step 1 of Fig.~\ref{bij:disk}), i.e.,
\begin{itemize}
\item Find the topmost $\ominus$-node, say $v_1$, on the spine of $T$ (since $T\in\widetilde\mathfrak{DT}_n$, such a $v_1$ always exists). Find the first (by inorder) $\ominus$-node, say $v$ (possibly $v=v_1$), of $T$.
\item In tree $T$, there is a unique path $P$ from $v_1$ to $v$. Let $v_1,v_2,\ldots,v_{k-1},v_k=v$ be all the $\ominus$-nodes on the path $P$ in the order we visit them when walking from $v_1$ to $v$. Note that by our choice of $v$, the path $P$ cannot have two consecutive right edges, making all of the $\ominus$-nodes $v_1,v_2,\ldots,v_{k-1},v_k=v$ on $P$ eligible for applying the transformation $\mathcal{L}$.
\item Define the di-sk tree $T'$ by
$$
T'=\mathcal{L}(v_k,\mathcal{L}(v_{k-1},\ldots,\mathcal{L}(v_2,\mathcal{L}(v_1,T))\cdots)).
$$
\end{itemize}
For the second step, we do the ``{\bf backward shift}'' on $T'$ to obtain $\phi(T)$ (see the shift inside the dotted box in Step 2 of Fig.~\ref{bij:disk}), i.e.,
\begin{itemize}
\item Let $w$ be the left child of $v$ in $T'$. Then $w$ must be an $\oplus$-node according to the construction of the first step above.
\item Let $B$ (possibly empty) be the right subtree of $w$. If $B$ is empty, then set $\phi(T)=T'$. Otherwise, the root of $B$ is an $\ominus$-node according to the definition of di-sk trees, as $w$ is an $\oplus$-node. Let $P'$ be the spine of $B$.
\item Let $v_1',v_2',\ldots,v_{\ell}'$ be all the $\ominus$-nodes in the path $P'$ from the top to the bottom. Then $v_1'$ is the topmost node of $P'$, which is also the root of $B$. For $1\leq i\leq \ell-1$, suppose the number of $\oplus$-nodes on $P'$ between $v_i'$ and $v_{i+1}'$ is $c_i$. Suppose the number of $\oplus$-nodes on $P'$ below $v_{\ell}'$ is $c_{\ell}$. For instance, for the middle tree $T'$ in Fig.~\ref{bij:disk}, we have $c_1=1$ and $c_2=2$.
\item Introduce $\mathcal{L}^{-k}(v',T'):=\mathcal{L}^{-1}(v',\mathcal{L}^{-k+1}(v',T'))$ recursively for $k\geq2$. Define the di-sk tree $\phi(T)$ by
$$
\phi(T)=\mathcal{L}^{-c_{\ell}}(v_{\ell}',\mathcal{L}^{-c_{\ell-1}}(v_{\ell-1}',\ldots,\mathcal{L}^{-c_2}(v_2',\mathcal{L}^{-c_1}(v_1',T'))\cdots)).
$$
\end{itemize}
By the above construction, we see that the node immediately after (by inorder) $w$ of $\phi(T)$ is an $\ominus$-node and therefore $\phi(T)$ is a di-sk tree in $\widetilde\mathfrak{DT}_n^{(k-1,l+1)}$. We aim to show that $\phi$ is a bijection by defining $\phi^{-1}$ explicitly.
For a di-sk tree $\tilde T\in\widetilde\mathfrak{DT}_n^{(k-1,l+1)}$ with $k\geq1$, $l\ge 0$, we use two steps to obtain $\phi^{-1}(\tilde T)$ from $\tilde T$ as follows. In the first step, we do the ``{\bf forward shift}'' on $\tilde T$ to obtain a di-sk tree $\tilde T^*$, i.e.,
\begin{itemize}
\item Find the $\oplus$-node, say $w$, immediately before the first $\ominus$-node (by inorder), such a $w$ always exists since $\mathsf{iop}(\tilde T)=l+1\ge 1$. Trace the unique path from $w$ back to the root of $\tilde T$ to locate the first $\ominus$-node, say $v$. Such a $v$ always exists since $\tilde T\in\widetilde\mathfrak{DT}_n$. Let $c_1$ be the number of nodes (necessarily $\oplus$-nodes) in the path from $v$ to $w$, with $v$ and $w$ excluded.
\item Let $\tilde B$ (possibly empty) be the right subtree of $w$ in $\tilde T$. If $\tilde B$ is empty, then $v$ must be the first $\ominus$-node by inorder. In this case, $c_1=0$ and we set $\tilde T^*=\tilde T$. Otherwise, $\tilde B$ is not empty and let us consider the the spine $\tilde P^*$ of $\tilde B$.
\item Let $v_1',v_2',\ldots,v_{\ell}'$ be all the $\ominus$-nodes in the path $\tilde P^*$ from the top to the bottom. Then $v_1'$ and $v_{\ell}'$ (possibly coincide) are the root and the tail of the path $\tilde P^*$, respectively. For $2\leq i\leq \ell$, let $c_i$ be the number of $\oplus$-nodes between $v_{i-1}'$ and $v_i'$ in the path $\tilde P^*$.
\item Introduce $\mathcal{L}^{k}(v',\tilde T):=\mathcal{L}(v',\mathcal{L}^{k-1}(v',\tilde T))$ recursively for $k\geq2$. Define the di-sk tree $\tilde T^*$ by
$$
\tilde T^*=\mathcal{L}^{c_{\ell}}(v_{\ell}',\mathcal{L}^{c_{\ell-1}}(v_{\ell-1}',\ldots,\mathcal{L}^{c_2}(v_2',\mathcal{L}^{c_1}(v_1',\tilde T))\cdots)).
$$
\end{itemize}
We see that $\tilde T^*\in\widetilde\mathfrak{DT}_n$ and the ``{\bf forward shift}'' is clearly inverse to the ``{\bf backward shift}''. For the second step, we do the ``{\bf swing up}'' on $\tilde T^*$ to obatin $\phi^{-1}(\tilde T)$:
\begin{itemize}
\item From the construction of the ``{\bf forward shift}'', the $\ominus$-node $v$ becomes the parent of $w$ in $\tilde T^*$. Let $v_1$ be the topmost $\ominus$-node in the spine of $\tilde T^*$. Such a $v_1$ always exists since $\tilde T^*\in\widetilde\mathfrak{DT}_n$.
\item In the tree $\tilde T^*$, there is a unique path $\tilde P$ from $v_1$ to $v$, which contains no consecutive right edges. Let $v_1,v_2,\ldots,v_{k-1},v_k=v$ be all the $\ominus$-nodes on $\tilde P$ in the order we visit them when walking from $v_1$ to $v$.
\item Define the di-sk tree $\phi^{-1}(\tilde T)$ by
$$
\phi^{-1}(\tilde T)=\mathcal{L}^{-1}(v_1,\mathcal{L}^{-1}(v_{2},\ldots,\mathcal{L}^{-1}(v_{k-1},\mathcal{L}^{-1}(v_k,\tilde T^*))\cdots)).
$$
\end{itemize}
Since the ``{\bf swing up}'' is inverse to the ``{\bf swing down}'', the mapping $\phi^{-1}$ is indeed the inverse of $\phi$. As every step of the bijection $\phi$ involves only the elementary transformation $\mathcal{L}$, Lemma~\ref{lem:L} guarantees that the desired property~\eqref{desb:lmax} holds. This ends the proof of the theorem.
\end{proof}
\begin{example}
As an example of $\phi$, the di-sk tree $T\in\widetilde\mathfrak{DT}_{11}^{(1,0)}$ in Fig.~\ref{di-sk} and its image $\phi(T)\in\widetilde\mathfrak{DT}_{11}^{(0,1)}$ under $\phi$ are drawn in Fig.~\ref{exam:phi}. One can check that $\eta^{-1}(T)=\pi=5\,2\,3\,4\,1\,9\,11\,10\,6\,8\,7$, $\eta^{-1}(\phi(T))=\pi'=5\,9\,6\,8\,7\,11\,10\,2\,3\,4\,1$,
$$
(\mathrm{DESB},\mathrm{LMAX},\mathrm{LMIN})\, \pi=(\{1,2,6,7,10\},\{5,9,11\},\{5,2,1\})=(\mathrm{DESB},\mathrm{LMAX},\mathrm{LMIN})\, \pi',
$$
$(\mathsf{comp},\mathsf{iar})\,\pi=(2,1)$ and $(\mathsf{comp},\mathsf{iar})\,\pi'=(1,2)$.
\begin{figure}
\begin{tikzpicture}[scale=0.2]
\draw[-] (17,8) to (19,10);
\draw[-] (20,10) to (22,8);
\draw[-] (22,7) to (20,5);
\draw[-] (23,7) to (25,5);
\draw[-] (20,4) to (22,2);
\draw[-] (26,4) to (28,2);
\draw[-] (16,7) to (14,5);
\node at (13.5,4.5) {$\ominus$};
\draw[-] (14,4) to (16,2);
\node at (16.5,1.5) {$\oplus$};
\draw[-] (16,1) to (14,-1);
\node at (13.5,-1.5) {$\oplus$};
\node at (16.5,7.5) {$\ominus$};
\node at (19.5,10.5) {$\oplus$};
\node at (22.5,7.5) {$\ominus$};
\node at (19.5,4.5) {$\oplus$};
\node at (25.5,4.5) {$\oplus$};
\node at (22.5,1.5) {$\ominus$};
\node at (28.5,1.5) {$\ominus$};
\node at (31,6.5) {$\longrightarrow$};\node at (31,8) {$\phi$};
\node at (44.5,10.5) {$\ominus$};
\draw[-] (44,10) to (42,8);\node at (41.5,7.5) {$\ominus$};
\draw[-] (44,4) to (42.5,2.5);\node at (42,2) {$\oplus$};
\draw[-] (42,7) to (44,5);\node at (44.5,4.5) {$\oplus$};
\draw[-] (41,7) to (38,4);\node at (37.5,3.5) {$\oplus$};
\draw[-] (37,3) to (35,1);\node at (34.5,0.5) {$\oplus$};
\draw[-] (38,3) to (40,1);\node at (40.5,0.5) {$\ominus$};
\draw[-] (35,0) to (37,-2);\node at (37.5,-2.5) {$\ominus$};
\draw[-] (38,-3) to (40,-5);\node at (40.5,-5.5) {$\oplus$};
\draw[-] (41,-6) to (43,-8);\node at (43.5,-8.5) {$\ominus$};
\end{tikzpicture}
\caption{An example of the bijection $\phi$.\label{exam:phi}}
\end{figure}
\end{example}
Now we are ready to prove our main result, Theorem~\ref{thm:sep:sym}.
\begin{proof}[{\bf Proof of Theorem~\ref{thm:sep:sym}}] The involution $\Phi$ is defined recursively using $\phi$. Set $\Phi(\mathrm{id}_1)=\mathrm{id}_1$. For each $\pi=\pi_1\pi_2\cdots\pi_n\in\mathfrak{S}_n(2413,3142)$ and $n\geq2$, we need to distinguish two cases as follows.
\begin{itemize}
\item If $\eta(\pi)\in\mathfrak{DT}_n\setminus \widetilde \mathfrak{DT}_n$, i.e., the spine of $\eta(\pi)$ is composed of $\oplus$-nodes only, hence $\pi=1\oplus\sigma$, where $\sigma=(\pi_2-1)(\pi_3-1)\cdots(\pi_n-1)\in\mathfrak{S}_{n-1}(2413,3142)$. Define $\Phi(\pi)=1\oplus\Phi(\sigma)$.
\item Otherwise, we have $\eta(\pi)\in\widetilde \mathfrak{DT}_n^{(k,l)}$ for some $k$ and $l$. Let us define
$$\Phi(\pi)=\eta^{-1}\circ\phi^{k-l}\circ\eta(\pi).$$
\end{itemize}
For instance, if $\pi=2\,4\,5\,9\,6\,8\,7\,11\,10\,3\,1\in\mathfrak{S}_{11}(2413,3142)$, then $\eta(\pi)\in\widetilde \mathfrak{DT}_{11}^{(0,3)}$ and so
$$
\Phi(\pi)=\eta^{-1}\circ\phi^{-3}\circ\eta(\pi)=2\,1\,4\,3\,5\,9\,11\,10\,6\,8\,7.
$$
It follows from Lemmas~\ref{lem:sep1},~\ref{lem:sep2} and Theorem~\ref{thm:di-sk} that $\Phi$ is an involution that preserves the triple of set-valued statistics $(\mathrm{LMAX},\mathrm{LMIN},\mathrm{DESB})$ but exchanges the pair $(\mathsf{comp},\mathsf{iar})$, which completes the proof of the first statement of Theorem~\ref{thm:sep:sym}.
For the second statement, we observe that $\pi\in\mathfrak{S}_n(312)$ if and only if each $\ominus$-node of $\eta(\pi)$ has no right child. It is clear that the elementary transformation $\mathcal{L}$, which deletes/inserts left edges only, preserves this kind of property. Consequently, $\Phi$ indeed restricts to an involution on $\mathfrak{S}_n(312)$.
\end{proof}
\begin{remark}
Since $\Phi$ restricts to an involution on ${\mathfrak{S}_{n}(312)}$, we get immediately that the two quintuples
$(\mathrm{LMAX},\mathrm{LMIN},\mathrm{DESB},\mathsf{comp},\mathsf{iar})$ and $(\mathrm{LMAX},\mathrm{LMIN},\mathrm{DESB},\mathsf{iar},\mathsf{comp})$ have the same distribution over $\mathfrak{S}_n(312)$. However, as $\mathrm{LMIN}(\pi)=\{1,2,\ldots,\pi_1\}$ for any $\pi\in\mathfrak{S}_n(312)$ and $\pi_1=\min(\mathrm{LMAX}(\pi))$, this result is equivalent to Theorem 1.1~(ii) of~\cite{flw}.
\end{remark}
\section{tree traversal}\label{sec:tree-tra}
In retrospect, we note that for each di-sk tree $T$, the statistic $\mathsf{top}(T)$ also equals the number of initial $\oplus$-nodes in $T$ when we use the {\em preorder} (i.e., recursively traversing the parent to the left subtree then to the right subtree) instead of the inorder to traverse $T$. Motivated by this new perspective, we investigate in this section the distributions of the number of initial $\oplus$-nodes with respect to the following eight types of tree traversal (see Tab.~\ref{eight-tree-tra}). The first six of which are usually called {\em depth first traversals}, while the last two are called {\em breadth first traversals}.
\begin{table}[bp]\caption{Eight types of tree traversal for a di-sk tree $T$ and their associated statistics for the number of initial $\oplus$-nodes}\label{eight-tree-tra}
\centering
\begin{tabular}{|c||c|c|c|c|}
\hline
Name & inorder & right inorder & preorder & right preorder \\
\hline
Rule & Left--Root--Right & Right--Root--Left & Root--Left--Right & Root--Right--Left \\
\hline
Stat & $\mathsf{iop}(T)$ & $\mathsf{riop}(T)$ & $\mathsf{top}(T)$ & $\mathsf{rtop}(T)$ \\
\hline
Name & postorder & right postorder & level order & right level order \\
\hline
Rule & Left--Right--Root & Right--Left--Root & Left--Right--Next level & Right--Left--Next level \\
\hline
Stat & $\mathsf{pop}(T)$ & $\mathsf{rpop}(T)$ & $\mathsf{lop}(T)$ & $\mathsf{rlop}(T)$ \\
\hline
\end{tabular}
\end{table}
\begin{figure}
\begin{tikzpicture}[scale=0.3]
\node at (2.5,19.5) {$\bullet$};
\node at (4.5,21.5) {$\circ$};
\node at (6.5,23.5) {$\bullet$};
\node at (2.5,15.5) {$\bullet$};
\node at (4.5,17.5) {$\circ$};
\node at (6.5,19.5) {$\circ$};
\node at (8.5,21.5) {$\circ$};
\node at (8.5,17.5) {$\bullet$};
\node at (10.5,19.5) {$\bullet$};
\draw[-] (2.7,19.7) to (4.33,21.33);
\draw[-] (4.7,21.7) to (6.33,23.33);
\draw[-] (2.7,15.7) to (4.33,17.33);
\draw[-] (6.7,19.7) to (8.33,21.33);
\draw[-] (2.67,19.33) to (4.3,17.7);
\draw[-] (6.67,23.33) to (8.3,21.7);
\draw[-] (6.67,19.33) to (8.3,17.7);
\draw[-] (8.67,21.33) to (10.3,19.7);
\node at (2,20.5) {$1$};
\node at (2,16.5) {$2$};
\node at (4.5,18.7) {$3$};
\node at (4,22.5) {$4$};
\node at (6,24.5) {$5$};
\node at (6.5,20.7) {$6$};
\node at (9,18.5) {$7$};
\node at (9,22.5) {$8$};
\node at (11,20.5) {$9$};
\node at (6.5,14) {$\mathsf{iop}(T)=2$};
\node at (14.5,19.5) {$\bullet$};
\node at (16.5,21.5) {$\circ$};
\node at (18.5,23.5) {$\bullet$};
\node at (14.5,15.5) {$\bullet$};
\node at (16.5,17.5) {$\circ$};
\node at (18.5,19.5) {$\circ$};
\node at (20.5,21.5) {$\circ$};
\node at (20.5,17.5) {$\bullet$};
\node at (22.5,19.5) {$\bullet$};
\draw[-] (14.7,19.7) to (16.33,21.33);
\draw[-] (16.7,21.7) to (18.33,23.33);
\draw[-] (14.7,15.7) to (16.33,17.33);
\draw[-] (18.7,19.7) to (20.33,21.33);
\draw[-] (14.67,19.33) to (16.3,17.7);
\draw[-] (18.67,23.33) to (20.3,21.7);
\draw[-] (18.67,19.33) to (20.3,17.7);
\draw[-] (20.67,21.33) to (22.3,19.7);
\node at (14,20.5) {$9$};
\node at (14,16.5) {$8$};
\node at (16.5,18.7) {$7$};
\node at (16,22.5) {$6$};
\node at (18,24.5) {$5$};
\node at (18.5,20.7) {$4$};
\node at (21,18.5) {$3$};
\node at (21,22.5) {$2$};
\node at (23,20.5) {$1$};
\node at (18.5,14) {$\mathsf{riop}(T)=1$};
\node at (26.5,19.5) {$\bullet$};
\node at (28.5,21.5) {$\circ$};
\node at (30.5,23.5) {$\bullet$};
\node at (26.5,15.5) {$\bullet$};
\node at (28.5,17.5) {$\circ$};
\node at (30.5,19.5) {$\circ$};
\node at (32.5,21.5) {$\circ$};
\node at (32.5,17.5) {$\bullet$};
\node at (34.5,19.5) {$\bullet$};
\draw[-] (26.7,19.7) to (28.33,21.33);
\draw[-] (28.7,21.7) to (30.33,23.33);
\draw[-] (26.7,15.7) to (28.33,17.33);
\draw[-] (30.7,19.7) to (32.33,21.33);
\draw[-] (26.67,19.33) to (28.3,17.7);
\draw[-] (30.67,23.33) to (32.3,21.7);
\draw[-] (30.67,19.33) to (32.3,17.7);
\draw[-] (32.67,21.33) to (34.3,19.7);
\node at (26,20.5) {$3$};
\node at (26,16.5) {$5$};
\node at (28.5,18.7) {$4$};
\node at (28,22.5) {$2$};
\node at (30,24.5) {$1$};
\node at (30.5,20.7) {$7$};
\node at (33,18.5) {$8$};
\node at (33,22.5) {$6$};
\node at (35,20.5) {$9$};
\node at (30.5,14) {$\mathsf{top}(T)=1$};
\node at (38.5,19.5) {$\bullet$};
\node at (40.5,21.5) {$\circ$};
\node at (42.5,23.5) {$\bullet$};
\node at (38.5,15.5) {$\bullet$};
\node at (40.5,17.5) {$\circ$};
\node at (42.5,19.5) {$\circ$};
\node at (44.5,21.5) {$\circ$};
\node at (44.5,17.5) {$\bullet$};
\node at (46.5,19.5) {$\bullet$};
\draw[-] (38.7,19.7) to (40.33,21.33);
\draw[-] (40.7,21.7) to (42.33,23.33);
\draw[-] (38.7,15.7) to (40.33,17.33);
\draw[-] (42.7,19.7) to (44.33,21.33);
\draw[-] (38.67,19.33) to (40.3,17.7);
\draw[-] (42.67,23.33) to (44.3,21.7);
\draw[-] (42.67,19.33) to (44.3,17.7);
\draw[-] (44.67,21.33) to (46.3,19.7);
\node at (38,20.5) {$7$};
\node at (38,16.5) {$9$};
\node at (40.5,18.7) {$8$};
\node at (40,22.5) {$6$};
\node at (42,24.5) {$1$};
\node at (42.5,20.7) {$4$};
\node at (45,18.5) {$5$};
\node at (45,22.5) {$2$};
\node at (47,20.5) {$3$};
\node at (42.5,14) {$\mathsf{rtop}(T)=1$};
\node at (2.5,6.5) {$\bullet$};
\node at (4.5,8.5) {$\circ$};
\node at (6.5,10.5) {$\bullet$};
\node at (2.5,2.5) {$\bullet$};
\node at (4.5,4.5) {$\circ$};
\node at (6.5,6.5) {$\circ$};
\node at (8.5,8.5) {$\circ$};
\node at (8.5,4.5) {$\bullet$};
\node at (10.5,6.5) {$\bullet$};
\draw[-] (2.7,6.7) to (4.33,8.33);
\draw[-] (4.7,8.7) to (6.33,10.33);
\draw[-] (2.7,2.7) to (4.33,4.33);
\draw[-] (6.7,6.7) to (8.33,8.33);
\draw[-] (2.67,6.33) to (4.3,4.7);
\draw[-] (6.67,10.33) to (8.3,8.7);
\draw[-] (6.67,6.33) to (8.3,4.7);
\draw[-] (8.67,8.33) to (10.3,6.7);
\node at (2,7.5) {$3$};
\node at (2,3.5) {$1$};
\node at (4.5,5.7) {$2$};
\node at (4,9.5) {$4$};
\node at (6,11.5) {$9$};
\node at (6.5,7.7) {$6$};
\node at (9,5.5) {$5$};
\node at (9,9.5) {$8$};
\node at (11,7.5) {$7$};
\node at (6.5,1) {$\mathsf{pop}(T)=1$};
\node at (14.5,6.5) {$\bullet$};
\node at (16.5,8.5) {$\circ$};
\node at (18.5,10.5) {$\bullet$};
\node at (14.5,2.5) {$\bullet$};
\node at (16.5,4.5) {$\circ$};
\node at (18.5,6.5) {$\circ$};
\node at (20.5,8.5) {$\circ$};
\node at (20.5,4.5) {$\bullet$};
\node at (22.5,6.5) {$\bullet$};
\draw[-] (14.7,6.7) to (16.33,8.33);
\draw[-] (16.7,8.7) to (18.33,10.33);
\draw[-] (14.7,2.7) to (16.33,4.33);
\draw[-] (18.7,6.7) to (20.33,8.33);
\draw[-] (14.67,6.33) to (16.3,4.7);
\draw[-] (18.67,10.33) to (20.3,8.7);
\draw[-] (18.67,6.33) to (20.3,4.7);
\draw[-] (20.67,8.33) to (22.3,6.7);
\node at (14,7.5) {$7$};
\node at (14,3.5) {$5$};
\node at (16.5,5.7) {$6$};
\node at (16,9.5) {$8$};
\node at (18,11.5) {$9$};
\node at (18.5,7.7) {$3$};
\node at (21,5.5) {$2$};
\node at (21,9.5) {$4$};
\node at (23,7.5) {$1$};
\node at (18.5,1) {$\mathsf{rpop}(T)=2$};
\node at (26.5,6.5) {$\bullet$};
\node at (28.5,8.5) {$\circ$};
\node at (30.5,10.5) {$\bullet$};
\node at (26.5,2.5) {$\bullet$};
\node at (28.5,4.5) {$\circ$};
\node at (30.5,6.5) {$\circ$};
\node at (32.5,8.5) {$\circ$};
\node at (32.5,4.5) {$\bullet$};
\node at (34.5,6.5) {$\bullet$};
\draw[-] (26.7,6.7) to (28.33,8.33);
\draw[-] (28.7,8.7) to (30.33,10.33);
\draw[-] (26.7,2.7) to (28.33,4.33);
\draw[-] (30.7,6.7) to (32.33,8.33);
\draw[-] (26.67,6.33) to (28.3,4.7);
\draw[-] (30.67,10.33) to (32.3,8.7);
\draw[-] (30.67,6.33) to (32.3,4.7);
\draw[-] (32.67,8.33) to (34.3,6.7);
\node at (26,7.5) {$4$};
\node at (26,3.5) {$9$};
\node at (28.5,5.7) {$7$};
\node at (28,9.5) {$2$};
\node at (30,11.5) {$1$};
\node at (30.5,7.7) {$5$};
\node at (33,5.5) {$8$};
\node at (33,9.5) {$3$};
\node at (35,7.5) {$6$};
\node at (30.5,1) {$\mathsf{lop}(T)=1$};
\node at (38.5,6.5) {$\bullet$};
\node at (40.5,8.5) {$\circ$};
\node at (42.5,10.5) {$\bullet$};
\node at (38.5,2.5) {$\bullet$};
\node at (40.5,4.5) {$\circ$};
\node at (42.5,6.5) {$\circ$};
\node at (44.5,8.5) {$\circ$};
\node at (44.5,4.5) {$\bullet$};
\node at (46.5,6.5) {$\bullet$};
\draw[-] (38.7,6.7) to (40.33,8.33);
\draw[-] (40.7,8.7) to (42.33,10.33);
\draw[-] (38.7,2.7) to (40.33,4.33);
\draw[-] (42.7,6.7) to (44.33,8.33);
\draw[-] (38.67,6.33) to (40.3,4.7);
\draw[-] (42.67,10.33) to (44.3,8.7);
\draw[-] (42.67,6.33) to (44.3,4.7);
\draw[-] (44.67,8.33) to (46.3,6.7);
\node at (38,7.5) {$6$};
\node at (38,3.5) {$9$};
\node at (40.5,5.7) {$8$};
\node at (40,9.5) {$3$};
\node at (42,11.5) {$1$};
\node at (42.5,7.7) {$5$};
\node at (45,5.5) {$7$};
\node at (45,9.5) {$2$};
\node at (47,7.5) {$4$};
\node at (42.5,1) {$\mathsf{rlop}(T)=1$};
\end{tikzpicture}
\caption{Eight types of traversal for a di-sk tree $T$ and its associated statistics}
\label{8traversal}
\end{figure}
We use $\mathsf{st}_1\sim_{S}\mathsf{st}_2$ to indicate that the two statistics $\mathsf{st}_1$ and $\mathsf{st}_2$ are equidistributed over the set $S$. For instance, Theorem~\ref{thm:sep:sym} implies in particular that $\mathsf{iar}\sim_{\mathfrak{S}_n(2413,3142)}\mathsf{comp}$, or equivalently upon applying the mapping $\eta$, we have $\mathsf{iop}\sim_{\mathfrak{DT}_n}\mathsf{top}$. We simply write $\mathsf{st}_1\sim\mathsf{st}_2$ when the set $S$ is clear from the context.
Our first result in this section classifies the corresponding eight statistics in Tab.~\ref{eight-tree-tra} into three equidistribution classes.
\begin{theorem}\label{thm:transver}
Over the set of di-sk trees $\mathfrak{DT}_n$, the distributions of the eight statistics in Tab.~\ref{eight-tree-tra} group into three classes. Namely, we have two classes
\begin{align}
&\mathsf{rtop}=\mathsf{rlop}, \text{ and}\label{two equal}\\
&\mathsf{riop}\sim\mathsf{iop}\sim\mathsf{top}\sim\mathsf{pop}\sim\mathsf{rpop}, \label{five equal}
\end{align}
where $\mathsf{st}_1=\mathsf{st}_2$ means that $\mathsf{st}_1(T)=\mathsf{st}_2(T)$ for each $T\in\mathfrak{DT}_n$. The distribution of $\mathsf{lop}$ over $\mathfrak{DT}_n$ is different from the other seven statistics for $n\ge 5$.
\end{theorem}
\begin{proof}
Firstly, it can be quickly checked that the eight statistics have the same distribution among the two di-sk trees in $\mathfrak{DT}_2$ and the six di-sk trees in $\mathfrak{DT}_3$. We next show that $\mathsf{rtop}=\mathsf{rlop}$ by induction on $n$. Take any di-sk tree $T\in\mathfrak{DT}_n$, if the root of $T$ is an $\ominus$-node, then $\mathsf{rtop}(T)=\mathsf{rlop}(T)=0$. Otherwise we can write $T=(T_1,\oplus,T_2)$, with $T_1$ and $T_2$ being the left subtree and the right subtree of $T$, respectively. Now if $T_2$ is nonempty, then its root must be an $\ominus$-node since $T$ is a di-sk tree, thus we have $\mathsf{rtop}(T)=\mathsf{rlop}(T)=1$. If $T_2$ is empty, then we have
$\mathsf{rtop}(T)=1+\mathsf{rtop}(T_1)=1+\mathsf{rlop}(T_1)=\mathsf{rlop}(T)$ as well. We have finished the proof of \eqref{two equal}.
Next, we prove $\mathsf{riop}\sim\mathsf{iop}$ via a bijection. Recall the elementary operation {\em reverse-complement} on permutations:
\begin{align*}
\pi=\pi_1\cdots\pi_n &\mapsto \mathrm{rc}(\pi):=(n+1-\pi_n)\cdots (n+1-\pi_1).
\end{align*}
It then suffices to note that for each $T\in\mathfrak{DT}_n$,
\begin{align*}
\mathsf{riop}(T)=\mathsf{iop}(\eta\circ\mathrm{rc}\circ\eta^{-1}(T)),
\end{align*}
and the fact that $\mathfrak{S}_n(2413,3142)$ is closed under the reverse-complement map.
Since we have already proved the symmetry of $(\mathsf{iop},\mathsf{top})$ in Theorem~\ref{thm:di-sk}, it remains to show the two equidistributions $\mathsf{top}\sim\mathsf{rpop}$ and $\mathsf{pop}\sim\mathsf{rpop}$
to finish the proof of~\eqref{five equal}. The first one follows from the symmetry of $(\mathsf{rpop},\mathsf{top})$ that will be proved in Theorem~\ref{thm:rpop} using generating functions, while the second one is proved
via a recursively constructed bijection $\theta$ that we define next. Indeed, for each $n\ge 1$, we construct a bijection $\theta:\mathfrak{DT}_n\rightarrow\mathfrak{DT}_n$ satisfying
\begin{align}\label{bij:pop=rpop}
(\mathsf{pop},\mathsf{rpop}) T=(\mathsf{rpop},\mathsf{pop}) \theta(T),
\end{align}
for every $T\in\mathfrak{DT}_n$. Let $\emptyset$ be the empty tree and we set $\theta(\emptyset)=\emptyset$. For a di-sk tree $T \in \mathfrak{DT}_n$ with $n \geq 2$, suppose $T=(T_1,r,T_2)$, where $r=\oplus$ or $\ominus$ is the root of $T$, while $T_1$ (resp.~$T_2$) is the left (resp.~right) subtree of $T$. We consider the following three cases.
\begin{itemize}
\item If $T_2=\emptyset$, then we define $\theta(T)=(\theta(T_1),r,\emptyset)$.
\item If $T_2\neq\emptyset$ and $T_1=\emptyset$, then we let $\theta(T)$ be the unique di-sk tree in $\mathfrak{DT}_n$ having no left subtree and $\theta(T_2)$ as its right subtree.
\item Otherwise $T_1\neq\emptyset$ and $T_2\neq\emptyset$, then we let $\theta(T)$ be the unique di-sk tree in $\mathfrak{DT}_n$ having $\theta(T_2)$ and $\theta(T_1)$ as its left and right subtrees, respectively.
\end{itemize}
In all three cases, we can verify that~\eqref{bij:pop=rpop} holds true assuming it has been proved for trees with fewer nodes. It should also be clear how to invert the map $\theta$. The proof is now completed.
\end{proof}
For a di-sk tree $T$, let $\mathsf{omi}(T)$ be the number of $\ominus$-nodes in $T$.
\begin{theorem}\label{thm:rpop}
For $n\geq1$, the two triples $(\mathsf{omi},\mathsf{rpop},\mathsf{top})$ and $(\mathsf{omi},\mathsf{top},\mathsf{rpop})$ have the same distribution over $\mathfrak{DT}_n$.
In particular, the pair $(\mathsf{rpop},\mathsf{top})$ is symmetric over $\mathfrak{DT}_n$.
\end{theorem}
\begin{proof}
We will prove the equidistribution by using generating functions.
Let us consider the generating function
$$
S(x,y)=S(t,x,y;z):=1+\sum_{n\ge 1}z^n\sum_{T\in\mathfrak{DT}_{n+1}}t^{\mathsf{omi}(T)}x^{\mathsf{rpop}(T)}y^{\mathsf{top}(T)}=1+S^{\oplus}(x,y)+S^{\ominus}(x,y),
$$
where $S^{\oplus}(x,y)$ (resp.~$S^{\ominus}(x,y)$) is the generating function for nonempty di-sk trees whose root is an $\oplus$-node (resp.~$\ominus$-node). For the sake of simplicity, we set $S=S(1,1)$. By the work in~\cite{flz}, the function $S$ satisfies the algebraic equation
\begin{equation}\label{S:des}
S=tz^2S^3+tz^2S^2+(1+t)zS+1.
\end{equation}
On the other hand, by Theorem~\ref{thm:di-sk} the two pairs $(\mathsf{omi},\mathsf{top})$ and $(\mathsf{omi},\mathsf{iop})$ are equidistributed over di-sk trees. Thus, the pair $(\mathsf{omi},\mathsf{top})$ over di-sk trees has the same distribution as the pair $(\mathsf{des},\mathsf{iar}-1)$ over separable permutations and it follows from~\cite[Eq.~(5.4)]{flw} that
\begin{equation}\label{S:top}
S(1,y)=\frac{S}{1+(1-y)zS}.
\end{equation}
Let $T=(T_1,r,T_2)$ be a nonempty di-sk tree, with $r=\oplus$ or $\ominus$ being the label of its root, $T_1$ and $T_2$ being its left subtree and right subtree, respectively. We need to consider two cases:
\begin{itemize}
\item {\bf Case~1: $r=\oplus$}. We further distinguish two cases.
\begin{enumerate}
\item $T_2=\emptyset$.
\begin{itemize}
\item[a)] $T_1$ is a di-sk tree (possibly empty) without any $\ominus$-node. This case contributes to $S^{\oplus}(x,y)$ the enumerator
$$\frac{zxy}{1-zxy}.$$
\item[b)] Otherwise, $T_1$ is a di-sk tree with at least one $\ominus$-node. This case contributes to $S^{\oplus}(x,y)$ the enumerator
$$zy\biggl(S(x,y)-\frac{1}{1-zxy}\biggr).$$
\end{itemize}
\item $T_2\neq\emptyset$. This case contributes to $S^{\oplus}(x,y)$ the enumerator
$$zyS(1,y)S^{\ominus}(x,1).$$
\end{enumerate}
\item {\bf Case~2: $r=\ominus$}. We further distinguish two cases.
\begin{enumerate}
\item $T_2$ is a di-sk tree (possibly empty) without any $\ominus$-node. This case contributes to $S^{\ominus}(x,y)$ the enumerator
$$
\frac{tzS(x,1)}{1-zx}.
$$
\item $T_2$ is a di-sk tree with at least one $\ominus$-node. This case contributes to $S^{\ominus}(x,y)$ the enumerator
$$tzS\biggl(S^{\oplus}(x,1)-\frac{zx}{1-zx}\biggr).$$
\end{enumerate}
\end{itemize}
Combining all the above cases results in a system of functional equations
\begin{equation}\label{sys:rpop}
\begin{cases}
\,\, S(x,y)=1+S^{\oplus}(x,y)+S^{\ominus}(x,y),\\
\,\, S^{\oplus}(x,y)=\frac{zxy}{1-zxy}+zy\biggl(S(x,y)-\frac{1}{1-zxy}\biggr)+zyS(1,y)S^{\ominus}(x,1),\\
\,\, S^{\ominus}(x,y)=\frac{tzS(x,1)}{1-zx}+tzS\biggl(S^{\oplus}(x,1)-\frac{zx}{1-zx}\biggr).
\end{cases}
\end{equation}
Solving this system of equations for $y=1$ (using Maple) yields
\begin{align}
S(x,1)&=-(tz^2S^2+tz^2S+z-1)/A,\label{S:rpop}\\
S^{\oplus}(x,1)&=(txz^2(zx-1)S^2+tz(1-zx)^2S+x(1-z)(1-zx)+tz(1-x))/B,\label{Sop:rpop}\\
S^{\ominus}(x,1)&=(z-1)tz/B\label{Som:rpop},
\end{align}
where
\begin{align*}
A&:=(txz^3-tz^2)S^2+(txz^3-2tz^2)S+xz^2-tz-xz-z+1,\\
B&:=(1-zx)(tz^2(1-zx)S^2+tz^2(2-zx)S+tz+zx+z-xz^2-1).
\end{align*}
It follows from~\eqref{S:rpop} and~\eqref{S:top} that
\begin{align*}
S(x,1)-S(1,x)=\frac{(z-1)(tz^2S^3+tz^2S^2+(1+t)zS+1-S)}{A(xzS-zS-1)}=0,
\end{align*}
in view of~\eqref{S:des}. Equivalently,
\begin{equation}\label{sym:S}
S(x,1)=S(1,x)=\frac{S}{1+(1-x)zS}.
\end{equation}
Substituting~\eqref{sym:S},~\eqref{Sop:rpop} and~\eqref{Som:rpop} into~\eqref{sys:rpop} and solving this system of equations (using Maple) gives an expression for $S(x,y)$, which is a rational function in $S,t,z,x$ and $y$ (too complicated to be reported here). It turns out that there is a factor
$$tz^2S^3+tz^2S^2+(1+t)zS+1-S,$$ which is zero in view of~\eqref{S:des}, in the numerator of the difference $S(x,y)-S(y,x)$. This proves $S(x,y)=S(y,x)$, as desired.
\end{proof}
In Theorems~\ref{thm:di-sk},~\ref{thm:transver} and~\ref{thm:rpop}, we have already proved three symmetries over di-sk trees:
\begin{align*}
&(\mathsf{omi},\mathsf{top},\mathsf{iop})\sim(\mathsf{omi},\mathsf{iop},\mathsf{top}),\\
&(\mathsf{pop},\mathsf{rpop})\sim(\mathsf{rpop},\mathsf{pop}),\\
& (\mathsf{omi},\mathsf{top},\mathsf{rpop})\sim(\mathsf{omi},\mathsf{rpop},\mathsf{top}).
\end{align*}
Recall that our proof of $\mathsf{riop}\sim\mathsf{iop}$ is essentially the application of the reverse-complement operation composed with the bijection $\eta$. Note that this composed map preserves the statistics $\mathsf{top}$ and $\mathsf{omi}$. This means we get the following symmetry over di-sk trees for free:
\begin{align*}
&(\mathsf{omi},\mathsf{top},\mathsf{riop})\sim(\mathsf{omi},\mathsf{riop},\mathsf{top}).
\end{align*}
Among the remaining six pairs taken from the equidistributed quintuple $$\mathsf{riop}\sim\mathsf{iop}\sim\mathsf{top}\sim\mathsf{pop}\sim\mathsf{rpop},$$ our calculations suggest the following three more symmetric pairs.
\begin{conj}\label{conj:sym}
Over di-sk trees, the following three symmetries hold:
\begin{align*}
&(\mathsf{omi},\mathsf{riop},\mathsf{rpop})\sim(\mathsf{omi},\mathsf{rpop},\mathsf{riop}),\\
&(\mathsf{omi},\mathsf{iop},\mathsf{rpop})\sim(\mathsf{omi},\mathsf{rpop},\mathsf{iop}),\\
& (\mathsf{riop},\mathsf{pop})\sim(\mathsf{pop},\mathsf{riop}).
\end{align*}
\end{conj}
Although Conjecture~\ref{conj:sym} may be proved by computing their generating functions similarly as the proof of Theorem~\ref{thm:rpop}, combinatorial involution proofs are preferred.
\subsection{Initial $\ominus$-nodes in di-sk trees}
In the previous section, we have considered eight different types of tree traversal and their associated statistics for the number of initial $\oplus$-nodes in di-sk trees. One could also consider the statistics of the number of initial $\ominus$-nodes with respect to the eight types of tree traversal; see Tab.~\ref{eight-tree-omi} for the notations of these eight associated statistics. Since $\oplus$-nodes and $\ominus$-nodes are symmetry in di-sk trees, for a fixed type of traversal, the statistic of the number of initial $\oplus$-nodes is equidistributed with the statistic of the number of initial $\ominus$-nodes. For instance, we have $\mathsf{iop}\sim\mathsf{iom}$ over $\mathfrak{DT}_n$. Thus, by Theorem~\ref{thm:transver} we have five new Comtet statistics over di-sk trees
$$
\mathsf{riom}\sim\mathsf{iom}\sim\mathsf{tom}\sim\mathsf{pom}\sim\mathsf{rpom}.
$$
It would be interesting to investigate systematically the joint distribution for one of the five Comtet statistics above and one in~\eqref{five equal}. As one example, in the rest of this paper, we aim to prove combinatorially that the Comtet pair $(\mathsf{top},\mathsf{iom})$ is symmetric over di-sk trees.
\begin{table}[bp]\caption{Eight types of tree traversal for a di-sk tree $T$ and their associated statistics for the number of initial $\ominus$-nodes}\label{eight-tree-omi}
\centering
\begin{tabular}{|c||c|c|c|c|}
\hline
Name & inorder & right inorder & preorder & right preorder \\
\hline
Rule & Left--Root--Right & Right--Root--Left & Root--Left--Right & Root--Right--Left \\
\hline
Stat & $\mathsf{iom}(T)$ & $\mathsf{riom}(T)$ & $\mathsf{tom}(T)$ & $\mathsf{rtom}(T)$ \\
\hline
Name & postorder & right postorder & level order & right level order \\
\hline
Rule & Left--Right--Root & Right--Left--Root & Left--Right--Next level & Right--Left--Next level \\
\hline
Stat & $\mathsf{pom}(T)$ & $\mathsf{rpom}(T)$ & $\mathsf{lom}(T)$ & $\mathsf{rlom}(T)$ \\
\hline
\end{tabular}
\end{table}
For $n\geq1$ and $0\leq k,l\leq n-1$, let us consider the set
\begin{equation}\label{omi-top}
\mathfrak{DT}_n^{(k,l)}:=\{T\in\mathfrak{DT}_n: \mathsf{top}(T)=k, \mathsf{iom}(T)=l\}
\end{equation}
and denote its cardinality by $s_n^{(k,l)}$. Interestingly, it turns out that the $n\times n$ matrix $M_n^{\mathsf{top},\mathsf{iom}}$, with entry $s_n^{(i-1,j-1)}$ in row $i$ and column $j$, is upper anti-triangular and Hankel. The first values of $M_n^{\mathsf{top},\mathsf{iom}}$ are
$$
\begin{bmatrix}\label{tri:Sch2}
0& 1 \\
1 & 0
\end{bmatrix},
\begin{bmatrix}\label{tri:Sch3}
1 & 1 & 1 \\
1 & 1 & 0\\
1 & 0 & 0
\end{bmatrix},
\begin{bmatrix}\label{tri:Sch4}
4 & 4 & 2&1 \\
4 & 2 & 1&0\\
2 & 1 & 0&0\\
1&0&0&0
\end{bmatrix},
\begin{bmatrix}\label{tri:Sch5}
17 & 16 & 8&3 &1\\
16 & 8 & 3&1&0\\
8 & 3& 1&0&0\\
3&1&0&0&0\\
1&0&0&0&0
\end{bmatrix},
\begin{bmatrix}\label{tri:Sch}
76 & 69 & 34&13 &4&1\\
69 & 34 & 13&4&1&0\\
34 & 13& 4&1&0&0\\
13&4&1&0&0&0\\
4&1&0&0&0&0\\
1&0&0&0&0&0
\end{bmatrix}.
$$
The upper anti-triangular property of $M_n^{\mathsf{top},\mathsf{iom}}$ follows from the simple fact that $\mathsf{top}(T)+\mathsf{iom}(T)\leq n-1$ for each $T\in\mathfrak{DT}_n$, while the Hankel property of $M_n^{\mathsf{top},\mathsf{iom}}$ is a consequence of the following result.
\begin{theorem}\label{thm:bij2}
Let $\mathfrak{DT}_n^{(k,l)}$ be defined in~\eqref{omi-top}. If $k\geq1$, then there exists a bijection $\psi:\mathfrak{DT}_n^{(k,l)}\rightarrow \mathfrak{DT}_n^{(k-1,l+1)}$. Consequently, the pair $(\mathsf{comp},\mathsf{idr})$ of double Comtet statistics is symmetric over $\mathfrak{S}_n(2413,3142)$, where $\mathsf{idr}(\pi)$ denotes the length of the initial descending run of a permutation $\pi$.
\end{theorem}
In order to facilitate our construction of $\psi$, we define three basic operations, called {\em conjugation, insertion}, and {\em extraction}, for di-sk trees.
\begin{Def}[Conjugation]
Given a di-sk tree $T\in\mathfrak{DT}_n$, we reverse the labels of all of its nodes, i.e., $\oplus$-nodes become $\ominus$-nodes and $\ominus$-nodes become $\oplus$-nodes. This yields a new di-sk tree in $\mathfrak{DT}_n$ that we call the {\em conjugate} of $T$. Denote by $\widebar{T}$ the conjugate of $T$.
\end{Def}
\begin{figure}
\begin{tikzpicture}[scale=0.3]
\node at (9.5,8.5) {$\circ$};
\node at (7.5,6.5) {$\circ$};
\node at (5.5,4.5) {$\circ$};
\node at (3.5,2.5) {$\circ$};
\draw[-,dashed] (1,0) to (3.33,2.33);
\draw[-] (3.7,2.7) to (5.33,4.33);
\draw[-,dashed] (5.7,4.7) to (7.33,6.33);
\draw[-] (7.7,6.7) to (9.33,8.33);
\draw[-,dashed] (9.7,8.7) to (12,11);
\draw[-,dashed] (3.67,2.33) to (6,0);
\draw[-,dashed] (5.67,4.33) to (8,2);
\draw[-,dashed] (7.67,6.33) to (10,4);
\draw[-,dashed] (9.67,8.33) to (12,6);
\draw[-,dashed] (7,11) to (9.3,8.7);
\node at (2.75,3.25) {$a$};
\node at (4.75,5.25) {$c$};
\node at (6.75,7.25) {$d$};
\node at (9.5,10) {$b$};
\node at (14.5,5.5) {$\stackrel{\text{extract}}{\longrightarrow}$};
\node at (20.5,6.5) {$\circ$};
\node at (18.5,4.5) {$\circ$};
\draw[-,dashed] (18.7,4.7) to (20.33,6.33);
\draw[-,dashed] (18.67,4.33) to (21,2);
\draw[-,dashed] (20.67,6.33) to (23,4);
\node at (17.75,5.25) {$c$};
\node at (19.75,7.25) {$d$};
\node at (26.5,5.5) {$\Longrightarrow$};
\node at (33.5,6.5) {$\circ$};
\node at (31.5,4.5) {$\circ$};
\draw[-,dashed] (29,2) to (31.33,4.33);
\draw[-] (31.7,4.7) to (33.33,6.33);
\draw[-,dashed] (33.7,6.7) to (36,9);
\draw[-,dashed] (31.67,4.33) to (34,2);
\draw[-,dashed] (33.67,6.33) to (36,4);
\draw[-,dashed] (31,9) to (33.3,6.7);
\node at (30.75,5.25) {$a$};
\node at (33.5,8) {$b$};
\node at (6,-1.3) {$T_3$};
\node at (19.75,-1.3) {$T_2$};
\node at (32,-1.3) {$T_1=T_3\backslash T_2$};
\node at (7.5,20.5) {$\circ$};
\node at (5.5,18.5) {$\circ$};
\draw[-,dashed] (3,16) to (5.33,18.33);
\draw[-] (5.7,18.7) to (7.33,20.33);
\draw[-,dashed] (7.7,20.7) to (10,23);
\draw[-,dashed] (5.67,18.33) to (8,16);
\draw[-,dashed] (7.67,20.33) to (10,18);
\draw[-,dashed] (5,23) to (7.3,20.7);
\node at (4.75,19.25) {$a$};
\node at (7.5,22) {$b$};
\node at (6,13) {$T_1$};
\node at (19.75,13) {$T_2$};
\node at (33,13) {$T_3=T_1/T_2(a,b)$};
\node at (14.5,19.5) {$\stackrel{\text{insert}}{\longleftarrow}$};
\node at (20.5,20.5) {$\circ$};
\node at (18.5,18.5) {$\circ$};
\draw[-,dashed] (18.7,18.7) to (20.33,20.33);
\draw[-,dashed] (18.67,18.33) to (21,16);
\draw[-,dashed] (20.67,20.33) to (23,18);
\node at (17.75,19.25) {$c$};
\node at (19.75,21.25) {$d$};
\node at (26.5,19.5) {$\Longrightarrow$};
\node at (35.5,22.5) {$\circ$};
\node at (33.5,20.5) {$\circ$};
\node at (31.5,18.5) {$\circ$};
\node at (29.5,16.5) {$\circ$};
\draw[-,dashed] (27,14) to (29.33,16.33);
\draw[-] (29.7,16.7) to (31.33,18.33);
\draw[-,dashed] (31.7,18.7) to (33.33,20.33);
\draw[-] (33.7,20.7) to (35.33,22.33);
\draw[-,dashed] (35.7,22.7) to (38,25);
\draw[-,dashed] (29.67,16.33) to (32,14);
\draw[-,dashed] (31.67,18.33) to (34,16);
\draw[-,dashed] (33.67,20.33) to (36,18);
\draw[-,dashed] (35.67,22.33) to (38,20);
\draw[-,dashed] (33,25) to (35.3,22.7);
\node at (28.75,17.25) {$a$};
\node at (30.75,19.25) {$c$};
\node at (32.75,21.25) {$d$};
\node at (35.5,24) {$b$};
\end{tikzpicture}
\caption{Insertion and extraction.}
\label{fig:ins-ext}
\end{figure}
\begin{Def}[Insertion and extraction]
Given two trees $T_1,T_2$, suppose $a,b$ are two nodes of $T_1$ such that $a$ is the left child of $b$, and suppose $c$ (resp.~$d$) is the iroot (resp.~root) of $T_2$. We derive a new tree, say $T_3$, by first deleting the edge between $a$ and $b$, then attaching the subtree rooted at $a$ to the left of $c$, and attaching the subtree rooted at $d$ (now contains $a$) to the left of $b$. This operation is illustrated in Fig.~\ref{fig:ins-ext} (wherein $\circ$ does not indicate the label) and is called the {\em insertion of $T_2$ into $T_1$ at $a,b$}. We denote $T_3=T_1/T_2(a,b)$. Conversely, suppose $T_3$ and $T_2$ (with root $d$ and iroot $c$) are two trees such that $T_2$ can be embedded in $T_3$ satisfying
\begin{itemize}
\item[i.] $d$ is the left child of certain node $b\in T_3$;
\item[ii.] if we denote the left child of $c$ in $T_3$ as $a$, then the two edges $ca$ and $bd$ are the only edges connected to $T_2$ but not contained in $T_2$.
\end{itemize}
We derive a new tree, say $T_1$, from $T_3$ by first deleting the edges $ca$ and $bd$, then connecting $a$ and $b$ with a left edge. This operation is also illustrated in Fig.~\ref{fig:ins-ext} and is called the {\em extraction of $T_2$ from $T_3$}. We denote $T_1=T_3\backslash T_2$.
\end{Def}
\begin{remark}
To make the above definition applicable in more situations, we allow either $a$ or $b$ to be the empty node $\emptyset$. We explain here the meaning of such special cases for the insertion, while the extraction should be understood similarly. The meaning of $T_1/T_2(\emptyset,b)$ should be clear. For the case of $b=\emptyset$, $a$ must then be the root of $T_1$, and $T_1/T_2(a,\emptyset)$ is the tree rooted at $d$, with $a$ attached to $c$ by a left edge.
\end{remark}
Now we are ready for the proof of Theorem~\ref{thm:bij2}.
\begin{figure}
\begin{tikzpicture}[scale=0.4]
\draw[-] (2.15,3.15) to (2.9,3.9);
\draw[-] (3.15,4.15) to (4.9,5.9);
\draw[-] (5.15,6.15) to (6,7);
\draw[-] (8.15,7.15) to (8.9,7.9);
\draw[-] (8.15,9.15) to (8.9,9.9);
\draw[-] (9.15,10.15) to (9.9,10.9);
\draw[-] (10.15,11.15) to (11,12);
\draw[-] (11.9,12.9) to (11.15,12.15);
\draw[-] (3.15,6.15) to (3.9,6.9);
\draw[-] (4.15,7.15) to (4.9,7.9);
\draw[-] (5.15,8.15) to (8.9,11.9);
\draw[-] (9.15,12.15) to (10.9,13.9);
\draw[-] (6.85,6.15) to (5.1,7.9);
\draw[-] (10,7) to (9.1,7.9);
\draw[-] (8,9) to (8.85,8.15);
\draw[-] (10,9) to (9.1,9.9);
\draw[-] (10.13,10.87) to (10.85,10.15);
\draw[-] (12.1,12.9) to (12.85,12.15);
\draw[-] (11,14) to (11.85,13.15);
\node at (2,3) {$\circ$};
\node at (3,4) {$\circ$};
\node at (3.5,3.5) {$u_1$};
\node at (4,5) {$\bullet$};
\node at (4.5,4.5) {$v_1$};
\node at (5,6) {$\circ$};
\node at (6,7) {$\bullet$};
\node at (7,6) {$\circ$};
\node at (8,7) {$\bullet$};
\node at (9,8) {$\circ$};
\node at (10,7) {$\bullet$};
\node at (8,9) {$\bullet$};
\node at (9,10) {$\bullet$};
\node at (10.15,8.85) {$\circ$};
\node at (10,11) {$\circ$};
\node at (10.8,11) {$v_2$};
\node at (11,10) {$\bullet$};
\node at (11,12) {$\bullet$};
\node at (12,13) {$\circ$};
\node at (13,12) {$\bullet$};
\node at (11,14) {$\bullet$};
\node at (10.6,14.3) {$v$};
\node at (9,12) {$\circ$};
\node at (7,10) {$\bullet$};
\node at (5,8) {$\circ$};
\node at (4,7) {$\circ$};
\node at (3,6) {$\circ$};
\textcolor{red}{\draw[-,dashed] (11,15) to (10,14);}
\textcolor{red}{\draw[-,dashed] (10,14) to (11,13);}
\textcolor{red}{\draw[-,dashed] (11,13) to (9.2,11.2);}
\textcolor{red}{\draw[-,dashed] (9.2,11.2) to (11.2,9.2);}
\textcolor{red}{\draw[-,dashed] (11.2,9.2) to (14,12);}
\textcolor{red}{\draw[-,dashed] (11,15) to (14,12);}
\textcolor{blue}{\draw[-,dashed] (8.8,10.8) to (11.8,7.8);}
\textcolor{blue}{\draw[-,dashed] (8.8,10.8) to (7,9);}
\textcolor{blue}{\draw[-,dashed] (7,9) to (8,8);}
\textcolor{blue}{\draw[-,dashed] (8,8) to (7,7);}
\textcolor{blue}{\draw[-,dashed] (7,7) to (9,5);}
\textcolor{blue}{\draw[-,dashed] (9,5) to (11.8,7.8);}
\textcolor{red}{\draw[-,dashed] (16.8,5.8) to (19.8,2.8);}
\textcolor{red}{\draw[-,dashed] (16.8,5.8) to (16,5);}
\textcolor{red}{\draw[-,dashed] (16,5) to (17,4);}
\textcolor{red}{\draw[-,dashed] (17,4) to (15,2);}
\textcolor{red}{\draw[-,dashed] (15,2) to (17,0);}
\textcolor{red}{\draw[-,dashed] (17,0) to (19.8,2.8);}
\textcolor{blue}{\draw[-,dashed] (20,9) to (18,7);}
\textcolor{blue}{\draw[-,dashed] (18,7) to (19,6);}
\textcolor{blue}{\draw[-,dashed] (19,6) to (18.2,5.2);}
\textcolor{blue}{\draw[-,dashed] (18.2,5.2) to (20.2,3.2);}
\textcolor{blue}{\draw[-,dashed] (20.2,3.2) to (23,6);}
\textcolor{blue}{\draw[-,dashed] (20,9) to (23,6);}
\node at (14.5,8.7) {$\longrightarrow$};
\node at (14.5,9.4) {$\psi$};
\node at (14.5,8) {$\longleftarrow$};
\node at (14.5,7.3) {$\psi^{-1}$};
\draw[-] (16,2) to (16.9,2.9);
\draw[-] (17.15,3.15) to (18,4);
\draw[-] (19.15,5.15) to (20,6);
\draw[-] (15.15,3.15) to (15.9,3.9);
\draw[-] (16.15,4.15) to (16.9,4.9);
\draw[-] (17.15,5.15) to (18.9,6.9);
\draw[-] (19.15,7.15) to (19.9,7.9);
\draw[-] (20.15,8.15) to (21.9,9.9);
\draw[-] (22.15,10.15) to (23,11);
\draw[-] (20.15,10.15) to (20.9,10.9);
\draw[-] (21.15,11.15) to (21.9,11.9);
\draw[-] (22.15,12.15) to (23.9,13.9);
\draw[-] (16.85,1.15) to (16,2);
\draw[-] (18.85,3.15) to (17.1,4.9);
\draw[-] (20.85,5.15) to (19.1,6.9);
\draw[-] (21,7) to (20.1,7.9);
\draw[-] (23.85,10.15) to (22.1,11.9);
\node at (24,14) {$\circ$};
\node at (23,13) {$\bullet$};
\node at (22,12) {$\circ$};
\node at (21,11) {$\circ$};
\node at (20,10) {$\circ$};
\node at (23,11) {$\bullet$};
\node at (22,10) {$\circ$};
\node at (21,9) {$\bullet$};
\node at (21.7,9) {$v_1$};
\node at (20,8) {$\circ$};
\node at (19,7) {$\circ$};
\node at (17,5) {$\circ$};
\node at (16.5,5.2) {$v$};
\node at (16,4) {$\circ$};
\node at (15.3,4.3) {$u_1$};
\node at (15,3) {$\circ$};
\node at (24,10) {$\circ$};
\node at (21.15,6.85) {$\bullet$};
\node at (20,6) {$\bullet$};
\node at (19,5) {$\circ$};
\node at (18,4) {$\bullet$};
\node at (17,3) {$\circ$};
\node at (16,2) {$\bullet$};
\node at (16.8,2) {$v_2$};
\node at (21,5) {$\circ$};
\node at (19,3) {$\circ$};
\node at (17,1) {$\circ$};
\draw[-] (30,7) to (30.9,7.9);
\draw[-] (29.15,8.15) to (30,9);
\draw[-] (30.85,8.15) to (30,9);
\node at (30,9) {$\bullet$};\node at (30,9.6) {$v_1$};
\node at (29,8) {$\circ$};
\node at (31,8) {$\circ$};\node at (31.5,7.5) {$v_2$};
\node at (30,7) {$\bullet$};
\node at (33,8) {$\leftrightarrow$};
\draw[-] (34.15,7.15) to (34.9,7.9);
\draw[-] (35.15,8.15) to (36,9);
\draw[-] (36,7) to (35.12,7.88);
\node at (36.1,9.1) {$\circ$};
\node at (35,8) {$\circ$};\node at (34.8,8.7) {$v_1$};
\node at (34,7) {$\circ$};
\node at (36,7) {$\bullet$};\node at (36.7,6.8) {$v_2$};
\end{tikzpicture}
\caption{Two examples of $\psi$: a general one and a small one \label{bij:psi}}
\end{figure}
\begin{proof}[{\bf Proof of Theorem~\ref{thm:bij2}}]
For each tree $T\in\mathfrak{DT}_n^{(k,l)}$, since $k\geq1$, we can write $T=(T_1,v,T_2)$ with $v=\oplus$. We perform the following {\em cut-and-paste} procedure to get $\psi(T)$:
\begin{itemize}
\item[Step 1] Let $v_1$ be the first $\oplus$-node (by inorder) of $T$ and $u_1$ be its left child (if $v_1$ has no left child then set $u_1=\emptyset$). Now if $T_2=\emptyset$, set $T^*=(\emptyset,\ominus,\emptyset)$ and jump to Step 3.
\item[Step 2] Otherwise $T_2\neq\emptyset$ and the root of $T_2$ must be an $\ominus$-node since $T$ is a di-sk tree. Now denote the lowest $\ominus$-node on the spine of $T_2$ as $v_2$, and denote the left subtree (possibly empty) of $v_2$ as $T_3$. We set $T^*$ to be the {\bf conjugate} of the following tree
$$((T\backslash T_1)\backslash T_3)/T_3(v,\emptyset)$$
\item[Step 3] If $v_1$ coincides with $v$ (see a small example on right side of Fig.~\ref{bij:disk}), then take $\psi(T)=T_1/T^*(u_1,\emptyset)$. Otherwise, we take $\psi(T)=T_1/T^*(u_1,v_1)$.
\end{itemize}
It should be clear from our construction that $\psi(T)\in\mathfrak{DT}_n^{(k-1,l+1)}$ indeed. See Fig.~\ref{bij:disk} (on the left side) for an illustration of $\psi$.
It remains to show that $\psi$ is invertible. We will construct its inverse $\psi^{-1}$ explicitly. Given a di-sk tree $T\in\mathfrak{DT}_n^{(k-1,l+1)}$, we perform the following {\em inverse cut-and-paste} procedure to get $\psi^{-1}(T)$:
\begin{itemize}
\item[Step 1]
Since $T\in\mathfrak{DT}_n^{(k-1,l+1)}$, we have $\mathsf{iom}(T)=l+1$, so we can find the $(l+1)$-th (by inorder) node, which is an $\ominus$-node and denoted $v$. Let $v_2$ be the $(l+2)$-th node (by inorder), which is an $\oplus$-node. In the special case of $l+1=n-1$, we set $v_2=\emptyset$. If $v_2$ is not a descendent of $v$ (including the case $v_2=\emptyset$), set $T^*=(\emptyset,v,\emptyset)$, $\widetilde{T}=\widebar{T^*}$, and jump to Step 3.
\item[Step 2]
Otherwise $v_2$ must be a descendent of $v$. Now starting with $v$, we find the maximal chain of consecutive $\ominus$-nodes connected by left edges: $w_1=v, w_2,\ldots,w_m$, such that $w_m$ is the root of $T$ or $w_m$ has the right parent which is an $\oplus$-node. (Note that $w_m$ cannot have an $\oplus$-node as its left parent, since this will contradict with the fact that $v$ is the $(l+1)$-th initial $\ominus$-node). Let $T^*$ (resp.~$T_1$) be the tree having $w_m$ and $v$ (resp.~$w_2$) as its root and iroot, respectively. In the special case of $m=1$, simply take $T_1=\emptyset$. Furthermore, we let $\widetilde{T}$ be the {\bf conjugate} of the following tree
$$(T^*\backslash T_1)/T_1(\emptyset,v_2).$$
\item[Step 3]
Let $w$ be the root of $T$. If $w$ coincides with $v$, then take $\psi^{-1}(T)=(T\backslash T^*)/\widetilde{T}(u_1,\emptyset)$, where $u_1$ is the left child of $v$ in $T$. Otherwise, we take
$$\psi^{-1}(T)=(T\backslash T^*)/\widetilde{T}(w,\emptyset).$$
\end{itemize}
It is routine to check that $\psi$ and $\psi^{-1}$ are inverse to each other and so $\psi$ is indeed a bijection.
By Lemmas~\ref{lem:sep1} and~\ref{lem:sep2}, the pair $(\mathsf{comp},\mathsf{idr})$ of Comtet statistics over $\mathfrak{S}_n(2413,4213)$ is equidistributed with the pair $(\mathsf{top},\mathsf{iom})$ over $\mathfrak{DT}_n$. Since $\psi:\mathfrak{DT}_n^{(k,l)}\rightarrow \mathfrak{DT}_n^{(k-1,l+1)}$ is a bijection, the pair $(\mathsf{top},\mathsf{iom})$ is symmetry over $\mathfrak{DT}_n$ and so does the pair $(\mathsf{comp},\mathsf{idr})$ over $\mathfrak{S}_n(2413,4213)$.
\end{proof}
\section{Concluding remarks
\label{sec:fin}
The main achievement of this paper is the construction of a combinatorial bijection on di-sk trees that proves the equidistribution of two quintuples $(\mathrm{LMAX},\mathrm{LMIN},\mathrm{DESB},\mathsf{iar},\mathsf{comp})$ and $(\mathrm{LMAX},\mathrm{LMIN},\mathrm{DESB},\mathsf{comp},\mathsf{iar})$ over separable permutations.
At this point, we would like to pose several open problems.
\begin{?}
Our proof of the symmetry of $(\mathsf{rpop},\mathsf{top})$ in Theorem~\ref{thm:rpop} is purely algebraic, can one find a direct bijective proof (probably in the same spirit as $\psi$ constructed in Theorem~\ref{thm:bij2})? Could the three symmetries in Conjecture~\ref{conj:sym} be proved bijectively?
\end{?}
The three Comtet statistics $\mathsf{riop}$, $\mathsf{iop}$ and $\mathsf{top}$ in Theorem~\ref{thm:transver} have interpretations in terms of natural statistics over separable permutations under the bijection $\eta$. This makes us wonder whether there are natural interpretations of $\mathsf{pop}$ and $\mathsf{rpop}$ in terms of separable permutations.
\begin{?}
Define explicitly two statistics, say $\mathsf{st}$ and $\mathsf{st}'$, for every separable permutation $\pi$, such that
$$\mathsf{st}(\pi)-1 = \mathsf{pop}(\eta(\pi)),\quad \text{and} \quad \mathsf{st}'(\pi)-1 = \mathsf{rpop}(\eta(\pi)).$$
In view of \eqref{eq:iop} and \eqref{five equal}, such two statistics are (new?) Comtet statistics over separable permutations. Similar question can be asked for the statistics $\mathsf{pom}$ and $\mathsf{rpom}$.
\end{?}
\begin{?}
Sitting at the heart of our proof of Theorem~\ref{thm:sep:sym} is di-sk tree, we need the mapping $\eta$ to transform the results back and forth between separable permutations and di-sk trees. This is reminiscent of Rubey's proof \cite{rub} of the equidistribution of $(\mathrm{LMAX},\mathsf{iar})$ and $(\mathrm{LMAX},\mathsf{comp})$ on $321$-avoiding permutations using Dyck paths, where Krattenthaler's bijection plays the role of $\eta$. So one may ask, is there a way to bypass the use of di-sk tree and prove Theorem~\ref{thm:sep:sym} directly on permutations? This has been done in our previous work \cite{flw} for the case of $321$-avoiding permutations.
\end{?}
Many classical permutation statistics, such as {\em Eulerian} statistics, {\em Mahonian} statistics or {\em Stirling} statistics, have been extensively investigated in the literature (see the excellent book exposition~\cite{kit} of Kitaev) not only on permutations avoiding ordinary patterns, but also on permutations avoiding consecutive patterns or the more general vincular patterns. It would be interesting to explore systematically the distributions of the two Comtet statistics, $\mathsf{iar}$ and $\mathsf{comp}$, on permutations avoiding vincular patterns.
\section*{Acknowledgement
We thank Martin Rubey for sharing with us his observation that the set-valued statistic $\mathrm{LMIN}$ could be added to Theorem~1.2.
The second author was supported
by the National Science Foundation of China grants 11871247 and the project of Qilu Young Scholars of Shandong University.
|
\section{Introduction}
\label{sec:intro}
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=0.85\linewidth]{images/OBS_SOFA_connected.png}
\vspace{-0.25cm}
\caption{\footnotesize The first and final observations of septic patients in MIMIC-III, colored by SOFA score, visualized via Principal Component Analysis. Blue lines connect observations of patients who recovered, while red lines signify those that did not. Notably, these raw observations of severe health are not directly separable.}
\label{fig:intro_pca}
\end{center}
\vspace{-.35cm}
\end{figure}
Many problems in healthcare are a form of sequential decision making, e.g., clinical staff making decisions about the best ``next step'' in care \citep{ghassemi2019practical}.
Solving these problems is similar to finding an optimal decision making policy, requiring estimation and optimization of the cumulative effects of decisions over time~\citep{sox2007medical}.
Recently, reinforcement learning (RL) has been proposed as a promising approach for finding an optimal policy for such processes from data~\citep{gottesman2019guidelines}.
However, the development of a successful policy rests on the ability to derive informative states from observations. In healthcare these observations are noisy, irregular, and may not convey the entirety of a patient's condition \citep{obermeyer2016predicting}. While there are many proposed state construction approaches to handle these challenges~\citep{li2019optimizing,chang2019dynamic,peng2018improving,prasad2017reinforcement,raghu2017deep,RaghuKomorowskiCeliEtAl2017}, few consider the sequential nature of observations, choosing instead to isolate the features from a single time step to construct the state.
Critical care is one specific setting where sequential data is crucial for predictive modelling. Raw physiological observations may not be clearly separable with respect to patient acuity or outcome, complicating downstream prediction and treatment models~\citep{ibrahim2020classifying} (see Figure~\ref{fig:intro_pca}). Complex model architectures have shown improved performance on such tasks, due in part to their improved ability to generate high-quality representations~\citep{choi2016multi,sadati2018representation,weng2019representation}. Yet within healthcare, the design and learning of patient representations for RL is an open problem~\citep{yu2019reinforcement}.
In this work, we provide a controlled investigation of sequentially encoded state representations for use within RL applied to healthcare. We focus on the problem of treating septic patients~\citep{liu2020reinforcement}, using a patient cohort defined by~\citet{KomorowskiCeliBadawiEtAl2018} from the MIMIC-III dataset~\citep{mimicweb}.
We compare seven encoding architectures, and evaluate representations learned from sequential patient observations through three experiments.
First, we examine the effect of representation dimension when training models to predict \underline{\bf s}ubsequent physiological \underline{\bf o}bservations (SO) through autoencoding~\citep{baldi2012autoencoders} prior physiological observations. We position this as an auxiliary task to the development of a treatment policy~\citep{jaderberg2017reinforcement}.
Second, we investigate the impact of including contextual information as well as regularization when training these models. Context is added by augmenting the physiological observations with 5 demographic features. When regularizing model training, the learned representations are regularized to correlate with three clinical patient acuity scores -- OASIS \citep{jones2009sequential}, SAPS II \citep{le1993new} and SOFA \citep{johnson2013new}. We then qualitatively evaluate representations to determine their correlation with these scores, and embed them into a lower dimensional visualization to demonstrate their separability in contrast to the raw data.
Finally, we learn treatment policies from the encoded patient state representations using a state of the art off-policy RL algorithm, the discretized form of Batch Constrained Q-learning (dBCQ)~\citep{fujimoto2019benchmarking}. Policies are evaluated using weighted importance sampling~\citep{mahmood2014weighted}.
To our knowledge, we present the first rigorous empirical evaluation of learned patient state representations that facilitate policy learning.
A summary of our contributions are:
\begin{itemize}[leftmargin=*]
\item We show that, keeping all other hyperparameters constant, increasing the latent dimensionality could reduce prediction accuracy, indicating that high capacity representations are not always most informative.
\item We find that including demographic context when learning the state representation generally improves the performance of predicting SO.
\item We demonstrate that sequentially formed state representations can facilitate effective policy learning in batch settings. In particular, we find that representations learned through the recent Neural CDE~\citep{kidger2020neural} facilitate an especially effective policy.
\end{itemize}
\section{Background and related work}
\label{sec:background}
State representation learning has a long history within RL as a primary means of making complex control tasks computationally tractable~\citep{sutton1999between}.
Recent research has also separated feature extraction from policy learning~\citep{raffin2019decoupling}, where the goal is to isolate relevant features of the recorded observations in the representation, and provide more salient information to the policy learning algorithm.
Problems modeled as POMDPs often require a state representation to be specified, typically deriving from prior observations and actions~\citep{kaelbling1998planning}.
Past work in state construction has ranged from concatenation of a finite number of consecutive observations~\citep{mnih2013playing} to using the final layer of a recurrent neural network (RNN) to collectively embed a sequence of inputs~\citep{hausknecht2015deep}.
Most prior work in the context of RL and healthcare has constructed states from unprocessed observations, framing the problem as a fully observable MDP~\footnote{For reference, Table~\ref{tab:background} summarizes these approaches, found in Appendix~\ref{sec:apdx_prior_table}}. This approach naively abstracts the true nature of the data generating process which is inherently partially observable. Missingness as well as an incomplete understanding of biological and physiological processes contribute to the partial nature of healthcare observations. There is a growing set of RL literature in this applied space that accounts for partial observability explicitly. The literature specific to sepsis treatment~\citep{tsoukalas2015data,li2018actor,peng2018improving, li2019optimizing,lu2020deep} often learns state representations by utilizing recurrent methods, encoding sequentially observed features of the patient's condition into a hidden state.
To date, none of these works provide any analysis or justification of specific state representation choices. In this paper we address this empirical gap by rigorously evaluating multiple recurrent state representation learning approaches for use in healthcare. With this study we hope to provide a foundation for further research into representation learning for sequential decision problems within healthcare.
\section{Data}
\label{sec:dataset}
We consider the treatment of septic patients using data from the Medical Information Mart for Intensive Care (MIMIC-III) dataset (v1.4) \cite{johnson2016mimic}.
We follow~\citet{KomorowskiCeliBadawiEtAl2018} to extract and preprocess\footnote{Code available at \url{https://github.com/matthieukomorowski/AI_Clinician}} relevant vital and lab measurements to build a cohort of 19,418 patients among which there is an observed mortality rate just above 9\% (determined by death within 48h of the final observation).
To evaluate the formation of sequential representations of a patient's condition, we focus on patient vital signs and lab measurements that change over time, whether in response to selected treatments or as a consequence of their acute condition. This creates a dataset of $33$ features $\mathcal{O}$ with a discrete categorical action space with $25$ possible choices of combination between fluid and vasopressor amounts. We also experiment with including 5 additional demographic features $\mathfrak{D}$.
We include a list of features in Table~\ref{tab:features} with additional details included in Section~\ref{sec:apdx_data} of the Appendix.
\section{Methods}
\label{sec:methods}
\begin{figure*}[ht]
\vspace{-0.5cm}
\hspace{-.75cm}
\includegraphics[width=1.15\textwidth]{images/ml4h_comb_arch.png}
\caption{\footnotesize The architectures used to construct state representations via predicting future observations. {\bf a)} basic RNN autoencoder {\bf b)} Approximate Information State~\citep{AIS} {\bf c)} Neural CDE~\citep{kidger2020neural} {\bf d)} Decoupled Dynamics Module~\citep{zhang2018decoupling} {\bf e)} Deep Signature Transform~\citep{bonnier2019deep} {\bf f)} ODE-RNN~\citep{rubanova2019latent} {\bf g)} a non-recurrent Autoencoder. See Table~\ref{tab:arch_overview} in Appendix, Section~\ref{sec:apdx_arch} for a summary.}
\label{fig:all_archs}
\vspace{-0.5cm}
\end{figure*}
In this section, we provide a general overview of state representation learning via autoencoding architectures. We focus on the context of our first experimental analysis, where representations are used to predict the subsequent observation (SO).
\subsection{General overview}
With a batch of observed patient trajectories---comprised of transitions between subsequent observations $O_{t}$ and $O_{t+1}$ following treatment action $A_t$---we seek to learn an encoding function $\psi: \mathcal{H}_{t,t-1}\rightarrow\hatS_t$ as well as a decoding prediction function $\phi:\hatS_t\timesA_t\rightarrow\hatO_{t+1}$. Here the history $\mathcal{H}_{t, t-1}$ contains all observations $O_{0:t}$ and actions $A_{0:t-1}$ preceding the target observation $O_{t+1}$. Together, the encoding and decoding functions form a prediction $\hatO_{t+1}$ using the learned state representation $\hatS_t$.
That is, {\small $\hatO_{t+1} = \phi\left(\psi\left(\mathcal{H}_{t,t-1}\right), A_t\right) = \phi(\hatS_t, A_t)$.}
To facilitate sequentially stable predictions for the state representation $\hatS_t$ we choose encoding functions $\psi$ with a recurrent structure. Thus, $\hatS_t$ implicitly embeds the history $\mathcal{H}_{t,t-1}$, which has been shown to improve sepsis treatment policies~\citep{li2019optimizing}.
We jointly train the encoding function $\psi$ and decoding function $\phi$ via a loss function $\mathcal{L}(O_{t+1}, \hatO_{t+1})$, which in general computes the mean squared error between the predicted and true SO, as specified by the particular encoding approach.
\subsection{Information encoding models}
\label{sec:models}
We target six recurrent modeling approaches, largely motivated by their development to learn dynamics models:
\begin{itemize}[leftmargin=*,noitemsep]
\item Basic RNN Autoencoder (RNN)~\citep{chung2014empirical}
\item Approximate Information State (AIS,~\cite{AIS})
\item Neural Controlled Differential Equations (CDE,~\cite{kidger2020neural})
\item Decoupled Dynamics Module (DDM,~\cite{zhang2018decoupling})
\item Deep Signature Transforms (DST,~\cite{bonnier2019deep})
\item And the ODE-RNN (ODE,~\cite{rubanova2019latent})
\end{itemize}
These approaches are depicted in Figure~\ref{fig:all_archs}. We also compare these approaches to a simple non-recurrent Autoencoder (AE). A comparative overview of the features that differentiate each approach as well as specific details about how each are trained are presented in Sec.~\ref{sec:apdx_arch} of the Appendix.
The unifying feature among these approaches is the development of a latent representation space $\hatS$ that encodes information about the patient observations made over time. The formation of $\hatS$ is meant to develop informative representations to facilitate better downstream policy learning, by implicitly accounting for the history $\mathcal{H}_{t, t-1}$. That is, we seek to develop a strategy to select treatments based on the encoded history via the learned state representation: $A_t \sim \pi(\hatS_t|\mathcal{H}_{t,t-1})$. Through the remainder of this work, we evaluate the characteristics of the representations embedded in $\hatS$.
\noindent{\bf Model training:} We separate the data into a 70/15/15 train/validation/test split using stratified sampling. This maintains the same proportions of each terminal outcome (survival or mortality), and ensures that no patients are repeated across splits. All models were trained for the same number of epochs, using a variety of learning rates and 5 random initializations. The final settings for each model architecture are provided in the Appendix, Section~\ref{sec:apdx_arch}.
\subsection{Augmenting the learning process}
\label{sec:augmentation}
In hopes of ensuring that the intermediate state representations $\hatS_t$ retain clinically relevant features, we investigate augmenting the training of the representation space $S$ through a combination of two options: \begin{inparaenum} \item[(1)] Include the demographic context features $\mathfrak{D}$ (e.g. age, gender, etc.) as input to the encoder function $\psi$. When training with this option the history $\mathcal{H}_{t,t-1}$ contains observations $O_i^+ = [O_i, \mathfrak{D}_i]$. \item[(2)] Regularize the loss function by the Pearson correlation between the state representation and a set of acuity scores derived from the patient observations. \end{inparaenum} We utilize three independent acuity scores --- SOFA, SAPS II and OASIS --- through a linear combination of the correlation coefficients to subtract from the loss. The complete objective function when using this form of regularization is then,
\begin{equation*}
Loss = \mathcal{L}(O_{t+1}, \hatO_{t+1}) -\lambda~\rho(\hatS_t)
\end{equation*}
where {\small $\lambda~\rho(\hatS_t) = \lambda_1~\rho^{\text{SOFA}}(\hat{S_t}) + \lambda_2~\rho^{\text{SAPS II}}(\hatS_t) + \lambda_3~\rho^{\text{OASIS}}(\hatS_t)$}. We choose the hyperparameter $\lambda$ so that the final prediction loss of the regularized model is not inordinately larger than its unregularized counterpart. Additionally, we set $\lambda_1=\lambda_2=\lambda_3$ for simplicity in this paper yet these hyperparameters could be chosen independently of one another.
\subsection{Policy development}
\label{sec:policymethods}
We train policies on each of the learned state representations outlined in Section~\ref{sec:models}. As we do not have the ability to generate more data through an exploration of novel treatment strategies, we develop a policy using offline, batch reinforcement learning. In this setting, it is critical that the estimated value function not extrapolate to actions that are absent from the provided data~\citep{gottesman2019guidelines}. To avoid this extrapolation error \citet{fujimoto2019off} developed an algorithm that truncates any Q-function estimate corresponding to actions that fall outside the support of the dataset. This algorithm, Batch Constrained Q-Learning (BCQ), originally designed for continuous control problems was later adapted for use in discrete action settings~\citep{fujimoto2019benchmarking}.
We use this discretized BCQ algorithm to learn treatment policies from state representations $\hatS$. We train the policies using the encoded training subset of our data and validate the performance with the testing subset using weighted importance sampling (WIS), following~\cite{li2019optimizing}. The WIS return for each policy throughout training is computed by: $R^{\mathrm{WIS}} = \frac{\sum_n^N w_n R_n}{\sum_n^N w_n}$, where the $w_n$ are the per-trajectory $\mathrm{IS}$ weights and $R_n$ is the observed outcome of the trajectory. All further details regarding policy training and intermediate results are provided in Section~\ref{sec:apdx_policy_training}.
\section{Empirical Study}
\label{sec:empirical_study}
We evaluate the representations $\hat S_t$ learned from patient data following the three experiments outlined at the conclusion of Section~\ref{sec:intro}. All analyses and results reported through the remainder of this section are provided using only the test set of the patient cohort. All code used to extract and preprocess the data, train and evaluate the encoding models as well as the policies can be found at \url{https://github.com/MLforHealth/rl_representations}.
\subsection{Representation dimension in SO prediction}
\label{sec:prediction}
We evaluate the accuracy of predicting the SO $O_{t+1}$ from $O_t$ and $A_t$. Our primary investigation considers the effect of varying the dimension $\hatd_s$ of the learned state representation $\hatS_t$ from the set {\small $\hat d_s\in\{4,8,16,32,64,128,256\}$}. Other than varying the latent dimension in each run, we keep all other model and optimization hyperparameters constant. This experiment evaluates the information capacity needed in the state representation $\hatS_t$ to adequately predict the SO.
\begin{figure}[ht]
\hspace{-0.2cm}
\includegraphics[width=1.05\linewidth]{images/next_step_comb.png}
\caption{\footnotesize Mean squared error for SO prediction as a result of varying $\hatd_s$, comparing various training settings. Error bars are twice the std. dev. of each model over 5 random seeds. We note that augmenting the input to the encoding function $\psi$ with demographic context generally improves prediction performance. See Table~\ref{tab:next_step} for the best performing settings.}
\label{fig:next_step_results}
\vspace{-0.5cm}
\end{figure}
Results from these models, learned through the described training settings, are presented in Figure~\ref{fig:next_step_results} and Table~\ref{tab:next_step}. We see that the prediction performance of these models saturates as the dimension increases beyond 64, with the test loss increasing with larger representations. Aside from DST, the best performing settings of all other approaches converge between a loss of $0.46$ and $0.48$. This indicates that the highest capacity representations may not be the most informative for this prediction task.
\subsection{Augmenting learning in SO prediction}
We evaluate the two proposed training augmentations (see Sec.~\ref{sec:augmentation}) --- adding demographic features $\mathfrak{D}$ during training, and regularizing $\hatS$ to be correlated with SOFA, SAPS II and OASIS --- via the accuracy of predicting the SO $O_{t+1}$ from $O_t$ and $A_t$.
When augmenting the input to the encoding function $\psi$ with demographic context $\mathfrak{D}$ the prediction performance is generally improved (see the dashed curves in Figure~\ref{fig:next_step_results}). In contrast, the performance slightly degrades when the learned representations are regularized to be correlated with acuity scores (see the dotted and dot-dash lines in Figure~\ref{fig:next_step_results}), except for the DDM and DST models where there is a noticeable negative effect on model performance.
\begin{table}[ht]
\centering
\vspace{-0.15cm}
\caption{\footnotesize Optimal model settings for each approach when predicting the SO. Models are trained with observations $\mathcal{O}$ and can be augmented with demographic context $\mathcal{D}$ or by the correlation regularization $\mathcal{C}$.}
\label{tab:next_step}
\vspace{-0.25cm}
{\small
\resizebox{\linewidth}{!}{
\begin{tabular}{cccc}
{\bf Approach} & {\bf Best MSE} & {\bf $\hat d_s$} & {\bf Training Setting}\\ \midrule\midrule
AE & 0.4804$\pm$0.001 & 64 & w/ $\mathcal{O} + \mathfrak{D}$ \\
\midrule
AIS & 0.4679$\pm$ 0.004 & 64 &$\mathcal{O} + \mathfrak{D}$ \\
CDE & 0.4887$\pm$ 0.019 & 32 & $\mathcal{O} + \mathfrak{D}$ \\
DDM & 0.4654$\pm$ 0.002 & 128 & $\mathcal{O} + \mathfrak{D}$ \\
DST & 0.5863$\pm$ 0.013 & 64 & $\mathcal{O}$ \\
ODE & 0.4698$\pm$ 0.003 & 32 & $\mathcal{O} + \mathfrak{D}$ \\
RNN & 0.4658$\pm$ 0.002 & 64 & $\mathcal{O} + \mathfrak{D}$ \\
\bottomrule
\end{tabular}
}
}
\end{table}
\subsection{Qualitative analysis}
The following analyses investigate the qualitative impact that the separate training strategies have on learned representations.
\noindent{\bf Representation-to-acuity score correlation} We first evaluate the average correlation coefficient between the representations and derived acuity scores (see Section~\ref{sec:apdx_acuity} for more information). This is to demonstrate the capacity of the representations $\hatS_t$ to maintain clinically relevant information. We perform this analysis with and without the correlation regularization described in the previous subsection. The intention of this regularization is that the more positively correlated the representation is to the acuity scores, the more clinically informative the learned representation is. This was designed in hopes to improve SO prediction and policy learning yet there was no demonstrated advantage in doing so as shown in Figure~\ref{fig:next_step_results} and Figure~\ref{fig:apdx_policy_training}.
\begin{figure}[ht]
\centering
\includegraphics[width=\linewidth]{images/comb_corrHtmp_mainBody.png}
\caption{\footnotesize The average Pearson correlation coefficient between the state representations from each encoding approach and acuity scores. Shown here are the average coefficients when regularizing the learning process and demographic features are omitted (left) or are included (right) as input. For SAPS II and OASIS, the inclusion of demographic features when constructing the state representations results in higher correlation.}
\label{fig:corr_heatmap}
\end{figure}
We show the average correlation coefficients of the learned state representations with acuity scores in Figure~\ref{fig:corr_heatmap} for the two training settings where regularization is included (with and without demographic context). Unregularized representations fail to encode information that is correlated with the acuity scores (see Figure~\ref{fig:apdx_corrHtmp} in the Appendix). Between the two settings, representations are better correlated with the acuity scores when a patient's demographic context is included. This suggests that clinical acuity scores are strongly entangled with demographics features. Further investigation into the effects of this entanglement, including questions of fairness, is outside the scope of this study and is therefore a suggested element of future work.
\begin{figure*}[ht!]
\centering
\includegraphics[width=\linewidth]{images/comb_pca.png}
\caption{\footnotesize Representations of patient health, learned through a non-recurrent autoencoder (AE), Approximate Information State (AIS), a Neural CDE (CDE) and an ODE-RNN (ODE) (left-to-right, all other approaches are included in the Appendix, Section~\ref{sec:apdx_pca}) for two training settings. We show the first and final observations made of septic patients in the MIMIC-III dataset, colored by the SOFA score. Blue lines represent the trajectory of patients who recovered, while red lines connect observations of those those that did not.}
\label{fig:comb_pca}
\vspace{-0.5cm}
\end{figure*}
\noindent{\bf Visualizing learned state representations} Next, we use principal component analysis (PCA) to project the learned representations into a lower dimensional space. PCA embeddings are fit using the encoded representations for the entire test set but only the first and final representation from a patient trajectory are vizualized. To aid in connecting these two points, we have drawn a line between them colored by the patient outcome, survival (blue) vs. death (red).
As shown in Figure~\ref{fig:intro_pca}, PCA projections of raw observations are not separable. Separability is desirable because a representation that separates patients who are most at risk of death could be used to more easily facilitate prediction models.
In Figure~\ref{fig:comb_pca} we show PCA projections for AE, AIS, CDE and ODE in two training settings (remaining approaches and training settings in the Appendix, Section~\ref{sec:apdx_pca}). We focus on the role of including demographics without acuity regularization (top), and when it is included (bottom). With exception of AIS, regularization provides better separation between the patients that survive their sepsis infection and those that do not. Additionally, the regularization compresses the feature space of some encoding approaches. In combination with findings in Section~\ref{sec:prediction}, this compression suggests that the information prioritized via acuity regularization does not contribute to an improved state representations despite improved separability in representation space. Further analysis of the information content stored in the representations as a consequence of being regularized to correlate with acuity scores is a subject of future work.
\subsection{Policy training and evaluation}
\label{sec:policy_training}
We investigate the quality of treatment policies learned from the state representations via the approaches outlined in Section~\ref{sec:methods}, following the procedure outlined in \ref{sec:policymethods}. We train policies using discretized BCQ, and evaluate with weighted importance sampling (WIS).
In Figure~\ref{fig:policy_comparison} we present the best performing policies learned from state representations. For each approach, excepting ODE, the top policies were learned from representations trained with the demographic context included as input to the encoding function $\psi$. The best ODE policy was developed from representations learned from the observations alone (see Figure~\ref{fig:apdx_policy_training} in the Appendix).
\begin{figure}[ht]
\centering
\includegraphics[width=\linewidth]{images/ml4h_policy_learning_comb.png}
\caption{\footnotesize WIS evaluation of policies trained from the representations encoded by the architectures outlined in Section~\ref{sec:methods}. Policies are trained from an experience replay buffer comprised of the training batch of patient trajectories for 200k iterations, evaluating the trained policy every 500 iterations. Results presented here are averaged over 5 random seeds, the shaded region measures a single standard deviation across seeds.}
\label{fig:policy_comparison}
\vspace{-0.35cm}
\end{figure}
Among the various approaches, policies learned from representations encoded by the Neural CDE (CDE) far outperform the others. Simpler recurrent based architectures such as AIS and RNN also obtain higher performance than the non-recurrent autoencoding baseline (AE). These results contribute toward the validation of our empirical hypothesis, that recurrent architectures provide better state representations in sequential partially observed settings. However, the AE based policy still learns a better policy than those based from far more complex methods (DST, DDM and ODE) signifying that the representations from these methods did not adequately encode sufficient information to learn a policy from in the batch setting, possibly due to dataset limitations.
\section{Discussion}
\label{sec:discussion}
In this paper we have empirically evaluated seven information encoding approaches to develop sequential state representations of patient health, useful for learning effective treatment policies. We performed several experiments to determine characteristics useful for training state representations from noisy patient data that is inherently partially observed. To support the formation of informative representations we designed a supervised task where the representation implicitly encodes a history $\mathcal{H}_{t,t-1}$ of previous observations and actions to predict the next SO. This auxiliary task allowed us to investigate several properties of the representation space $\hatS$ based on decisions of how to execute the training.
In Section~\ref{sec:prediction} we showed that higher dimensional representations reduce prediction accuracy, indicating the high capacity representations are not the most informative. In tandem we demonstrated that the inclusion of demographic context improves the learned state representations. This was verified (see Section~\ref{sec:apdx_policy_training}) when learning treatment policies. The best performing policies for each information encoding approaches presented in this paper were trained from representations learned with demographic context.
{\bf Future work} In future work we intend to explore the use of multi-task learning~\citep{mcdermott2020comprehensive,lin2019adaptive} to jointly train the representation space. Additionally, we plan to investigate methods that incorporate indicators of feature missingness and other underlying contextual variables~\citep{agor2019value,fleming2019missingness,sharafoddini2019new,che2018recurrent,lipton2016directly}. We intend to study the effect these approaches have on the representation space, including a quantification of any performance reductions that may arise through use of demographic information encoding bias in the representations~\citep{chen2018my}.
The class of Neural Differential Equation methods~\citep{chen2018neural,rubanova2019latent,kidger2020neural} were developed to account for irregular time series with missing values and have demonstrated high performance in prediction tasks when provided feature sets with varying rates of missingness. Following the analyses performed in this paper, the Neural CDE appears promising for constructing state representations in the midst of the missingness and other irregularities inherent in healthcare data.
The conceptual separation between representation learning and policy learning in this paper was motivated by prior literature on state representation learning~\citep{raffin2019decoupling}. This choice allowed us to focus on the formation and analysis of the representation space $\hatS$. Another reason for this design choice was to enable straightforward use of current state of the art batch RL training algorithms. However, this decoupling is not necessary for developing off-policy sepsis treatment policies as discussed and demonstrated by~\cite{li2020optimizing}. Another line of future work utilizing the findings of this paper is to similarly develop an end-to-end policy development approach that combines the objectives of the auxiliary tasks and RL algorithm, explicitly accounting for the state representation space as it encodes features of the expected outcome via the RL objective.
Additionally, it is necessary to more fully evaluate and interpret what the learned state representations encode and whether clinically relevant relationships are preserved~\citep{bai2018interpretable}. It will be beneficial for the future use of these state representations to determine whether they embed trends in the data following the improving (or degrading) health of the patient beside only encoding features relevant for inferring the SO.
{\bf Conclusion} Such investigations and state representation learning will provide mechanisms by which we can better understand the cumulative effects of prescribed actions, chosen by following observed or learned policies.
State representations and learned value functions used in this manner can enable the identification of reliable treatment policies, developed following a learning process that acknowledges the sequential and partial nature of the observations that are made.
This paper recommends possible ways of thinking of representation learning as a form of auxiliary task within policy development. Among the various research directions that are natural extensions from this work, we affirm the necessity of thoughtfully designing the representation learning process to honor the partial and sequential nature of the data generating process. These opportunities for learning optimal state representations for RL in healthcare offer an exciting new area of research that we anticipate being fruitful for establishing future advances in clinically relevant sequential decision making problems.
\acks{\small We thank our many colleagues who contributed to thoughtful discussions and provided timely advice to improve this work. We specifically appreciate the feedback provided by Nathan Ng, Vinith Suriyakumar and Karsten Roth.
This research was supported in part by Microsoft Research, a CIFAR AI Chair at the Vector Institute, a Canada Research Council Chair, and an NSERC Discovery Grant.
Resources used in preparing this research were provided, in part, by the Province of Ontario, the Government of Canada through CIFAR, and companies sponsoring the Vector Institute \url{www.vectorinstitute.ai/\#partners}.}
\section{Introduction}
\label{sec:intro}
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=0.85\linewidth]{images/OBS_SOFA_connected.png}
\vspace{-0.25cm}
\caption{\footnotesize The first and final observations of septic patients in MIMIC-III, colored by SOFA score, visualized via Principal Component Analysis. Blue lines connect observations of patients who recovered, while red lines signify those that did not. Notably, these raw observations of severe health are not directly separable.}
\label{fig:intro_pca}
\end{center}
\vspace{-.35cm}
\end{figure}
Many problems in healthcare are a form of sequential decision making, e.g., clinical staff making decisions about the best ``next step'' in care \citep{ghassemi2019practical}.
Solving these problems is similar to finding an optimal decision making policy, requiring estimation and optimization of the cumulative effects of decisions over time~\citep{sox2007medical}.
Recently, reinforcement learning (RL) has been proposed as a promising approach for finding an optimal policy for such processes from data~\citep{gottesman2019guidelines}.
However, the development of a successful policy rests on the ability to derive informative states from observations. In healthcare these observations are noisy, irregular, and may not convey the entirety of a patient's condition \citep{obermeyer2016predicting}. While there are many proposed state construction approaches to handle these challenges~\citep{li2019optimizing,chang2019dynamic,peng2018improving,prasad2017reinforcement,raghu2017deep,RaghuKomorowskiCeliEtAl2017}, few consider the sequential nature of observations, choosing instead to isolate the features from a single time step to construct the state.
Critical care is one specific setting where sequential data is crucial for predictive modelling. Raw physiological observations may not be clearly separable with respect to patient acuity or outcome, complicating downstream prediction and treatment models~\citep{ibrahim2020classifying} (see Figure~\ref{fig:intro_pca}). Complex model architectures have shown improved performance on such tasks, due in part to their improved ability to generate high-quality representations~\citep{choi2016multi,sadati2018representation,weng2019representation}. Yet within healthcare, the design and learning of patient representations for RL is an open problem~\citep{yu2019reinforcement}.
In this work, we provide a controlled investigation of sequentially encoded state representations for use within RL applied to healthcare. We focus on the problem of treating septic patients~\citep{liu2020reinforcement}, using a patient cohort defined by~\citet{KomorowskiCeliBadawiEtAl2018} from the MIMIC-III dataset~\citep{mimicweb}.
We compare seven encoding architectures, and evaluate representations learned from sequential patient observations through three experiments.
First, we examine the effect of representation dimension when training models to predict \underline{\bf s}ubsequent physiological \underline{\bf o}bservations (SO) through autoencoding~\citep{baldi2012autoencoders} prior physiological observations. We position this as an auxiliary task to the development of a treatment policy~\citep{jaderberg2017reinforcement}.
Second, we investigate the impact of including contextual information as well as regularization when training these models. Context is added by augmenting the physiological observations with 5 demographic features. When regularizing model training, the learned representations are regularized to correlate with three clinical patient acuity scores -- OASIS \citep{jones2009sequential}, SAPS II \citep{le1993new} and SOFA \citep{johnson2013new}. We then qualitatively evaluate representations to determine their correlation with these scores, and embed them into a lower dimensional visualization to demonstrate their separability in contrast to the raw data.
Finally, we learn treatment policies from the encoded patient state representations using a state of the art off-policy RL algorithm, the discretized form of Batch Constrained Q-learning (dBCQ)~\citep{fujimoto2019benchmarking}. Policies are evaluated using weighted importance sampling~\citep{mahmood2014weighted}.
To our knowledge, we present the first rigorous empirical evaluation of learned patient state representations that facilitate policy learning.
A summary of our contributions are:
\begin{itemize}[leftmargin=*]
\item We show that, keeping all other hyperparameters constant, increasing the latent dimensionality could reduce prediction accuracy, indicating that high capacity representations are not always most informative.
\item We find that including demographic context when learning the state representation generally improves the performance of predicting SO.
\item We demonstrate that sequentially formed state representations can facilitate effective policy learning in batch settings. In particular, we find that representations learned through the recent Neural CDE~\citep{kidger2020neural} facilitate an especially effective policy.
\end{itemize}
\section{Background and related work}
\label{sec:background}
State representation learning has a long history within RL as a primary means of making complex control tasks computationally tractable~\citep{sutton1999between}.
Recent research has also separated feature extraction from policy learning~\citep{raffin2019decoupling}, where the goal is to isolate relevant features of the recorded observations in the representation, and provide more salient information to the policy learning algorithm.
Problems modeled as POMDPs often require a state representation to be specified, typically deriving from prior observations and actions~\citep{kaelbling1998planning}.
Past work in state construction has ranged from concatenation of a finite number of consecutive observations~\citep{mnih2013playing} to using the final layer of a recurrent neural network (RNN) to collectively embed a sequence of inputs~\citep{hausknecht2015deep}.
Most prior work in the context of RL and healthcare has constructed states from unprocessed observations, framing the problem as a fully observable MDP~\footnote{For reference, Table~\ref{tab:background} summarizes these approaches, found in Appendix~\ref{sec:apdx_prior_table}}. This approach naively abstracts the true nature of the data generating process which is inherently partially observable. Missingness as well as an incomplete understanding of biological and physiological processes contribute to the partial nature of healthcare observations. There is a growing set of RL literature in this applied space that accounts for partial observability explicitly. The literature specific to sepsis treatment~\citep{tsoukalas2015data,li2018actor,peng2018improving, li2019optimizing,lu2020deep} often learns state representations by utilizing recurrent methods, encoding sequentially observed features of the patient's condition into a hidden state.
To date, none of these works provide any analysis or justification of specific state representation choices. In this paper we address this empirical gap by rigorously evaluating multiple recurrent state representation learning approaches for use in healthcare. With this study we hope to provide a foundation for further research into representation learning for sequential decision problems within healthcare.
\section{Data}
\label{sec:dataset}
We consider the treatment of septic patients using data from the Medical Information Mart for Intensive Care (MIMIC-III) dataset (v1.4) \cite{johnson2016mimic}.
We follow~\citet{KomorowskiCeliBadawiEtAl2018} to extract and preprocess\footnote{Code available at \url{https://github.com/matthieukomorowski/AI_Clinician}} relevant vital and lab measurements to build a cohort of 19,418 patients among which there is an observed mortality rate just above 9\% (determined by death within 48h of the final observation).
To evaluate the formation of sequential representations of a patient's condition, we focus on patient vital signs and lab measurements that change over time, whether in response to selected treatments or as a consequence of their acute condition. This creates a dataset of $33$ features $\mathcal{O}$ with a discrete categorical action space with $25$ possible choices of combination between fluid and vasopressor amounts. We also experiment with including 5 additional demographic features $\mathfrak{D}$.
We include a list of features in Table~\ref{tab:features} with additional details included in Section~\ref{sec:apdx_data} of the Appendix.
\section{Methods}
\label{sec:methods}
\begin{figure*}[ht]
\vspace{-0.5cm}
\hspace{-.75cm}
\includegraphics[width=1.15\textwidth]{images/ml4h_comb_arch.png}
\caption{\footnotesize The architectures used to construct state representations via predicting future observations. {\bf a)} basic RNN autoencoder {\bf b)} Approximate Information State~\citep{AIS} {\bf c)} Neural CDE~\citep{kidger2020neural} {\bf d)} Decoupled Dynamics Module~\citep{zhang2018decoupling} {\bf e)} Deep Signature Transform~\citep{bonnier2019deep} {\bf f)} ODE-RNN~\citep{rubanova2019latent} {\bf g)} a non-recurrent Autoencoder. See Table~\ref{tab:arch_overview} in Appendix, Section~\ref{sec:apdx_arch} for a summary.}
\label{fig:all_archs}
\vspace{-0.5cm}
\end{figure*}
In this section, we provide a general overview of state representation learning via autoencoding architectures. We focus on the context of our first experimental analysis, where representations are used to predict the subsequent observation (SO).
\subsection{General overview}
With a batch of observed patient trajectories---comprised of transitions between subsequent observations $O_{t}$ and $O_{t+1}$ following treatment action $A_t$---we seek to learn an encoding function $\psi: \mathcal{H}_{t,t-1}\rightarrow\hatS_t$ as well as a decoding prediction function $\phi:\hatS_t\timesA_t\rightarrow\hatO_{t+1}$. Here the history $\mathcal{H}_{t, t-1}$ contains all observations $O_{0:t}$ and actions $A_{0:t-1}$ preceding the target observation $O_{t+1}$. Together, the encoding and decoding functions form a prediction $\hatO_{t+1}$ using the learned state representation $\hatS_t$.
That is, {\small $\hatO_{t+1} = \phi\left(\psi\left(\mathcal{H}_{t,t-1}\right), A_t\right) = \phi(\hatS_t, A_t)$.}
To facilitate sequentially stable predictions for the state representation $\hatS_t$ we choose encoding functions $\psi$ with a recurrent structure. Thus, $\hatS_t$ implicitly embeds the history $\mathcal{H}_{t,t-1}$, which has been shown to improve sepsis treatment policies~\citep{li2019optimizing}.
We jointly train the encoding function $\psi$ and decoding function $\phi$ via a loss function $\mathcal{L}(O_{t+1}, \hatO_{t+1})$, which in general computes the mean squared error between the predicted and true SO, as specified by the particular encoding approach.
\subsection{Information encoding models}
\label{sec:models}
We target six recurrent modeling approaches, largely motivated by their development to learn dynamics models:
\begin{itemize}[leftmargin=*,noitemsep]
\item Basic RNN Autoencoder (RNN)~\citep{chung2014empirical}
\item Approximate Information State (AIS,~\cite{AIS})
\item Neural Controlled Differential Equations (CDE,~\cite{kidger2020neural})
\item Decoupled Dynamics Module (DDM,~\cite{zhang2018decoupling})
\item Deep Signature Transforms (DST,~\cite{bonnier2019deep})
\item And the ODE-RNN (ODE,~\cite{rubanova2019latent})
\end{itemize}
These approaches are depicted in Figure~\ref{fig:all_archs}. We also compare these approaches to a simple non-recurrent Autoencoder (AE). A comparative overview of the features that differentiate each approach as well as specific details about how each are trained are presented in Sec.~\ref{sec:apdx_arch} of the Appendix.
The unifying feature among these approaches is the development of a latent representation space $\hatS$ that encodes information about the patient observations made over time. The formation of $\hatS$ is meant to develop informative representations to facilitate better downstream policy learning, by implicitly accounting for the history $\mathcal{H}_{t, t-1}$. That is, we seek to develop a strategy to select treatments based on the encoded history via the learned state representation: $A_t \sim \pi(\hatS_t|\mathcal{H}_{t,t-1})$. Through the remainder of this work, we evaluate the characteristics of the representations embedded in $\hatS$.
\noindent{\bf Model training:} We separate the data into a 70/15/15 train/validation/test split using stratified sampling. This maintains the same proportions of each terminal outcome (survival or mortality), and ensures that no patients are repeated across splits. All models were trained for the same number of epochs, using a variety of learning rates and 5 random initializations. The final settings for each model architecture are provided in the Appendix, Section~\ref{sec:apdx_arch}.
\subsection{Augmenting the learning process}
\label{sec:augmentation}
In hopes of ensuring that the intermediate state representations $\hatS_t$ retain clinically relevant features, we investigate augmenting the training of the representation space $S$ through a combination of two options: \begin{inparaenum} \item[(1)] Include the demographic context features $\mathfrak{D}$ (e.g. age, gender, etc.) as input to the encoder function $\psi$. When training with this option the history $\mathcal{H}_{t,t-1}$ contains observations $O_i^+ = [O_i, \mathfrak{D}_i]$. \item[(2)] Regularize the loss function by the Pearson correlation between the state representation and a set of acuity scores derived from the patient observations. \end{inparaenum} We utilize three independent acuity scores --- SOFA, SAPS II and OASIS --- through a linear combination of the correlation coefficients to subtract from the loss. The complete objective function when using this form of regularization is then,
\begin{equation*}
Loss = \mathcal{L}(O_{t+1}, \hatO_{t+1}) -\lambda~\rho(\hatS_t)
\end{equation*}
where {\small $\lambda~\rho(\hatS_t) = \lambda_1~\rho^{\text{SOFA}}(\hat{S_t}) + \lambda_2~\rho^{\text{SAPS II}}(\hatS_t) + \lambda_3~\rho^{\text{OASIS}}(\hatS_t)$}. We choose the hyperparameter $\lambda$ so that the final prediction loss of the regularized model is not inordinately larger than its unregularized counterpart. Additionally, we set $\lambda_1=\lambda_2=\lambda_3$ for simplicity in this paper yet these hyperparameters could be chosen independently of one another.
\subsection{Policy development}
\label{sec:policymethods}
We train policies on each of the learned state representations outlined in Section~\ref{sec:models}. As we do not have the ability to generate more data through an exploration of novel treatment strategies, we develop a policy using offline, batch reinforcement learning. In this setting, it is critical that the estimated value function not extrapolate to actions that are absent from the provided data~\citep{gottesman2019guidelines}. To avoid this extrapolation error \citet{fujimoto2019off} developed an algorithm that truncates any Q-function estimate corresponding to actions that fall outside the support of the dataset. This algorithm, Batch Constrained Q-Learning (BCQ), originally designed for continuous control problems was later adapted for use in discrete action settings~\citep{fujimoto2019benchmarking}.
We use this discretized BCQ algorithm to learn treatment policies from state representations $\hatS$. We train the policies using the encoded training subset of our data and validate the performance with the testing subset using weighted importance sampling (WIS), following~\cite{li2019optimizing}. The WIS return for each policy throughout training is computed by: $R^{\mathrm{WIS}} = \frac{\sum_n^N w_n R_n}{\sum_n^N w_n}$, where the $w_n$ are the per-trajectory $\mathrm{IS}$ weights and $R_n$ is the observed outcome of the trajectory. All further details regarding policy training and intermediate results are provided in Section~\ref{sec:apdx_policy_training}.
\section{Empirical Study}
\label{sec:empirical_study}
We evaluate the representations $\hat S_t$ learned from patient data following the three experiments outlined at the conclusion of Section~\ref{sec:intro}. All analyses and results reported through the remainder of this section are provided using only the test set of the patient cohort. All code used to extract and preprocess the data, train and evaluate the encoding models as well as the policies can be found at \url{https://github.com/MLforHealth/rl_representations}.
\subsection{Representation dimension in SO prediction}
\label{sec:prediction}
We evaluate the accuracy of predicting the SO $O_{t+1}$ from $O_t$ and $A_t$. Our primary investigation considers the effect of varying the dimension $\hatd_s$ of the learned state representation $\hatS_t$ from the set {\small $\hat d_s\in\{4,8,16,32,64,128,256\}$}. Other than varying the latent dimension in each run, we keep all other model and optimization hyperparameters constant. This experiment evaluates the information capacity needed in the state representation $\hatS_t$ to adequately predict the SO.
\begin{figure}[ht]
\hspace{-0.2cm}
\includegraphics[width=1.05\linewidth]{images/next_step_comb.png}
\caption{\footnotesize Mean squared error for SO prediction as a result of varying $\hatd_s$, comparing various training settings. Error bars are twice the std. dev. of each model over 5 random seeds. We note that augmenting the input to the encoding function $\psi$ with demographic context generally improves prediction performance. See Table~\ref{tab:next_step} for the best performing settings.}
\label{fig:next_step_results}
\vspace{-0.5cm}
\end{figure}
Results from these models, learned through the described training settings, are presented in Figure~\ref{fig:next_step_results} and Table~\ref{tab:next_step}. We see that the prediction performance of these models saturates as the dimension increases beyond 64, with the test loss increasing with larger representations. Aside from DST, the best performing settings of all other approaches converge between a loss of $0.46$ and $0.48$. This indicates that the highest capacity representations may not be the most informative for this prediction task.
\subsection{Augmenting learning in SO prediction}
We evaluate the two proposed training augmentations (see Sec.~\ref{sec:augmentation}) --- adding demographic features $\mathfrak{D}$ during training, and regularizing $\hatS$ to be correlated with SOFA, SAPS II and OASIS --- via the accuracy of predicting the SO $O_{t+1}$ from $O_t$ and $A_t$.
When augmenting the input to the encoding function $\psi$ with demographic context $\mathfrak{D}$ the prediction performance is generally improved (see the dashed curves in Figure~\ref{fig:next_step_results}). In contrast, the performance slightly degrades when the learned representations are regularized to be correlated with acuity scores (see the dotted and dot-dash lines in Figure~\ref{fig:next_step_results}), except for the DDM and DST models where there is a noticeable negative effect on model performance.
\begin{table}[ht]
\centering
\vspace{-0.15cm}
\caption{\footnotesize Optimal model settings for each approach when predicting the SO. Models are trained with observations $\mathcal{O}$ and can be augmented with demographic context $\mathcal{D}$ or by the correlation regularization $\mathcal{C}$.}
\label{tab:next_step}
\vspace{-0.25cm}
{\small
\resizebox{\linewidth}{!}{
\begin{tabular}{cccc}
{\bf Approach} & {\bf Best MSE} & {\bf $\hat d_s$} & {\bf Training Setting}\\ \midrule\midrule
AE & 0.4804$\pm$0.001 & 64 & w/ $\mathcal{O} + \mathfrak{D}$ \\
\midrule
AIS & 0.4679$\pm$ 0.004 & 64 &$\mathcal{O} + \mathfrak{D}$ \\
CDE & 0.4887$\pm$ 0.019 & 32 & $\mathcal{O} + \mathfrak{D}$ \\
DDM & 0.4654$\pm$ 0.002 & 128 & $\mathcal{O} + \mathfrak{D}$ \\
DST & 0.5863$\pm$ 0.013 & 64 & $\mathcal{O}$ \\
ODE & 0.4698$\pm$ 0.003 & 32 & $\mathcal{O} + \mathfrak{D}$ \\
RNN & 0.4658$\pm$ 0.002 & 64 & $\mathcal{O} + \mathfrak{D}$ \\
\bottomrule
\end{tabular}
}
}
\end{table}
\subsection{Qualitative analysis}
The following analyses investigate the qualitative impact that the separate training strategies have on learned representations.
\noindent{\bf Representation-to-acuity score correlation} We first evaluate the average correlation coefficient between the representations and derived acuity scores (see Section~\ref{sec:apdx_acuity} for more information). This is to demonstrate the capacity of the representations $\hatS_t$ to maintain clinically relevant information. We perform this analysis with and without the correlation regularization described in the previous subsection. The intention of this regularization is that the more positively correlated the representation is to the acuity scores, the more clinically informative the learned representation is. This was designed in hopes to improve SO prediction and policy learning yet there was no demonstrated advantage in doing so as shown in Figure~\ref{fig:next_step_results} and Figure~\ref{fig:apdx_policy_training}.
\begin{figure}[ht]
\centering
\includegraphics[width=\linewidth]{images/comb_corrHtmp_mainBody.png}
\caption{\footnotesize The average Pearson correlation coefficient between the state representations from each encoding approach and acuity scores. Shown here are the average coefficients when regularizing the learning process and demographic features are omitted (left) or are included (right) as input. For SAPS II and OASIS, the inclusion of demographic features when constructing the state representations results in higher correlation.}
\label{fig:corr_heatmap}
\end{figure}
We show the average correlation coefficients of the learned state representations with acuity scores in Figure~\ref{fig:corr_heatmap} for the two training settings where regularization is included (with and without demographic context). Unregularized representations fail to encode information that is correlated with the acuity scores (see Figure~\ref{fig:apdx_corrHtmp} in the Appendix). Between the two settings, representations are better correlated with the acuity scores when a patient's demographic context is included. This suggests that clinical acuity scores are strongly entangled with demographics features. Further investigation into the effects of this entanglement, including questions of fairness, is outside the scope of this study and is therefore a suggested element of future work.
\begin{figure*}[ht!]
\centering
\includegraphics[width=\linewidth]{images/comb_pca.png}
\caption{\footnotesize Representations of patient health, learned through a non-recurrent autoencoder (AE), Approximate Information State (AIS), a Neural CDE (CDE) and an ODE-RNN (ODE) (left-to-right, all other approaches are included in the Appendix, Section~\ref{sec:apdx_pca}) for two training settings. We show the first and final observations made of septic patients in the MIMIC-III dataset, colored by the SOFA score. Blue lines represent the trajectory of patients who recovered, while red lines connect observations of those those that did not.}
\label{fig:comb_pca}
\vspace{-0.5cm}
\end{figure*}
\noindent{\bf Visualizing learned state representations} Next, we use principal component analysis (PCA) to project the learned representations into a lower dimensional space. PCA embeddings are fit using the encoded representations for the entire test set but only the first and final representation from a patient trajectory are vizualized. To aid in connecting these two points, we have drawn a line between them colored by the patient outcome, survival (blue) vs. death (red).
As shown in Figure~\ref{fig:intro_pca}, PCA projections of raw observations are not separable. Separability is desirable because a representation that separates patients who are most at risk of death could be used to more easily facilitate prediction models.
In Figure~\ref{fig:comb_pca} we show PCA projections for AE, AIS, CDE and ODE in two training settings (remaining approaches and training settings in the Appendix, Section~\ref{sec:apdx_pca}). We focus on the role of including demographics without acuity regularization (top), and when it is included (bottom). With exception of AIS, regularization provides better separation between the patients that survive their sepsis infection and those that do not. Additionally, the regularization compresses the feature space of some encoding approaches. In combination with findings in Section~\ref{sec:prediction}, this compression suggests that the information prioritized via acuity regularization does not contribute to an improved state representations despite improved separability in representation space. Further analysis of the information content stored in the representations as a consequence of being regularized to correlate with acuity scores is a subject of future work.
\subsection{Policy training and evaluation}
\label{sec:policy_training}
We investigate the quality of treatment policies learned from the state representations via the approaches outlined in Section~\ref{sec:methods}, following the procedure outlined in \ref{sec:policymethods}. We train policies using discretized BCQ, and evaluate with weighted importance sampling (WIS).
In Figure~\ref{fig:policy_comparison} we present the best performing policies learned from state representations. For each approach, excepting ODE, the top policies were learned from representations trained with the demographic context included as input to the encoding function $\psi$. The best ODE policy was developed from representations learned from the observations alone (see Figure~\ref{fig:apdx_policy_training} in the Appendix).
\begin{figure}[ht]
\centering
\includegraphics[width=\linewidth]{images/ml4h_policy_learning_comb.png}
\caption{\footnotesize WIS evaluation of policies trained from the representations encoded by the architectures outlined in Section~\ref{sec:methods}. Policies are trained from an experience replay buffer comprised of the training batch of patient trajectories for 200k iterations, evaluating the trained policy every 500 iterations. Results presented here are averaged over 5 random seeds, the shaded region measures a single standard deviation across seeds.}
\label{fig:policy_comparison}
\vspace{-0.35cm}
\end{figure}
Among the various approaches, policies learned from representations encoded by the Neural CDE (CDE) far outperform the others. Simpler recurrent based architectures such as AIS and RNN also obtain higher performance than the non-recurrent autoencoding baseline (AE). These results contribute toward the validation of our empirical hypothesis, that recurrent architectures provide better state representations in sequential partially observed settings. However, the AE based policy still learns a better policy than those based from far more complex methods (DST, DDM and ODE) signifying that the representations from these methods did not adequately encode sufficient information to learn a policy from in the batch setting, possibly due to dataset limitations.
\section{Discussion}
\label{sec:discussion}
In this paper we have empirically evaluated seven information encoding approaches to develop sequential state representations of patient health, useful for learning effective treatment policies. We performed several experiments to determine characteristics useful for training state representations from noisy patient data that is inherently partially observed. To support the formation of informative representations we designed a supervised task where the representation implicitly encodes a history $\mathcal{H}_{t,t-1}$ of previous observations and actions to predict the next SO. This auxiliary task allowed us to investigate several properties of the representation space $\hatS$ based on decisions of how to execute the training.
In Section~\ref{sec:prediction} we showed that higher dimensional representations reduce prediction accuracy, indicating the high capacity representations are not the most informative. In tandem we demonstrated that the inclusion of demographic context improves the learned state representations. This was verified (see Section~\ref{sec:apdx_policy_training}) when learning treatment policies. The best performing policies for each information encoding approaches presented in this paper were trained from representations learned with demographic context.
{\bf Future work} In future work we intend to explore the use of multi-task learning~\citep{mcdermott2020comprehensive,lin2019adaptive} to jointly train the representation space. Additionally, we plan to investigate methods that incorporate indicators of feature missingness and other underlying contextual variables~\citep{agor2019value,fleming2019missingness,sharafoddini2019new,che2018recurrent,lipton2016directly}. We intend to study the effect these approaches have on the representation space, including a quantification of any performance reductions that may arise through use of demographic information encoding bias in the representations~\citep{chen2018my}.
The class of Neural Differential Equation methods~\citep{chen2018neural,rubanova2019latent,kidger2020neural} were developed to account for irregular time series with missing values and have demonstrated high performance in prediction tasks when provided feature sets with varying rates of missingness. Following the analyses performed in this paper, the Neural CDE appears promising for constructing state representations in the midst of the missingness and other irregularities inherent in healthcare data.
The conceptual separation between representation learning and policy learning in this paper was motivated by prior literature on state representation learning~\citep{raffin2019decoupling}. This choice allowed us to focus on the formation and analysis of the representation space $\hatS$. Another reason for this design choice was to enable straightforward use of current state of the art batch RL training algorithms. However, this decoupling is not necessary for developing off-policy sepsis treatment policies as discussed and demonstrated by~\cite{li2020optimizing}. Another line of future work utilizing the findings of this paper is to similarly develop an end-to-end policy development approach that combines the objectives of the auxiliary tasks and RL algorithm, explicitly accounting for the state representation space as it encodes features of the expected outcome via the RL objective.
Additionally, it is necessary to more fully evaluate and interpret what the learned state representations encode and whether clinically relevant relationships are preserved~\citep{bai2018interpretable}. It will be beneficial for the future use of these state representations to determine whether they embed trends in the data following the improving (or degrading) health of the patient beside only encoding features relevant for inferring the SO.
{\bf Conclusion} Such investigations and state representation learning will provide mechanisms by which we can better understand the cumulative effects of prescribed actions, chosen by following observed or learned policies.
State representations and learned value functions used in this manner can enable the identification of reliable treatment policies, developed following a learning process that acknowledges the sequential and partial nature of the observations that are made.
This paper recommends possible ways of thinking of representation learning as a form of auxiliary task within policy development. Among the various research directions that are natural extensions from this work, we affirm the necessity of thoughtfully designing the representation learning process to honor the partial and sequential nature of the data generating process. These opportunities for learning optimal state representations for RL in healthcare offer an exciting new area of research that we anticipate being fruitful for establishing future advances in clinically relevant sequential decision making problems.
\acks{\small We thank our many colleagues who contributed to thoughtful discussions and provided timely advice to improve this work. We specifically appreciate the feedback provided by Nathan Ng, Vinith Suriyakumar and Karsten Roth.
This research was supported in part by Microsoft Research, a CIFAR AI Chair at the Vector Institute, a Canada Research Council Chair, and an NSERC Discovery Grant.
Resources used in preparing this research were provided, in part, by the Province of Ontario, the Government of Canada through CIFAR, and companies sponsoring the Vector Institute \url{www.vectorinstitute.ai/\#partners}.}
|
\section{Introduction}
Restricted by the imaging technology, it remains incomplete parts and noise in legacy grayscale photos. It is highly challenging to restore them due to the great information loss of real world. Also, there is high demand for high-quality and colorful legacy photos. Recently, as deep learning techniques have been demonstrated to successfully applied to many low-level computer vision tasks, the legacy photo enhancement becomes possible. In this paper, we would first discover the representation of blind noise from legacy images as a prior, and then perform image editing based on the estimated noise prior.
Editing legacy photos is highly challenging since there are multiple degradation types in legacy photos. Firstly, there exist noises with unknown distribution and intensity. The noises may be caused by many reasons such as sensor noise, camera distortion, jpeg compression, preservation technology, etc. However, most of current denoisers \cite{zhang2017beyond, tai2017memnet, zhang2018ffdnet, liu2018multi, gu2019self, liu2020densely} are trained with specific noise models such as Gaussian and Poisson distribution. Directly applying those denoisers to legacy photos cannot well enhance the images \cite{abdelhamed2020ntire}. Secondly, it remains flaws or cracks in legacy photos, which are not global noise but regional artifacts. Moreover, the levels of the artifacts are different for distinct pixels, which are hard to estimate. Finally, the grayscale legacy photos lack of color. Thus, the colorization process is significant to attach vivid colors to them. In conclusion, the pipeline of legacy photo editing can be categorized into three parts: denoising, inpainting, and colorization.
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth, height=6.1cm]{img/figure1-2.pdf}
\caption{The edited real legacy photo samples by proposed method (chosen from LP dataset, captured around 1950). The first, second and third row denote the real legacy photos, image enhancement results by \cite{gu2019self, yu2019free, zhang2016colorful} sequentially and the proposed pipeline, respectively.}
\label{teaser}
\end{figure*}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth, height=8cm]{img/figure2-2.pdf}
\caption{The details of edited real legacy photo by 3 samples. The left part includes the input patches. The right part includes denoised result, final result by the proposed system, denoised result by \cite{gu2019self}, and previous pipeline \cite{gu2019self, yu2019free, zhang2016colorful}, respectively.}
\label{teaser}
\end{figure}
To address the issues, we propose a system to implement the pipeline sequentially. Firstly for denoising, the noise distribution of legacy photos is always unknown. However, the current denoisers pre-define a fixed noise model. It is not practical to directly apply the denoisers to process legacy photos with blind noise. If denoised images are not clean enough, the following inpainting and colorization will also be affected. Moreover, there are no pairs of degraded and clean target legacy photos (i.e. legacy photos are normally noisy). There may be three approaches to address the issue such as estimating noise model \cite{abdelhamed2019noise, wei2020physics}, unsupervised training \cite{lehtinen2018noise2noise, batson2019noise2self} and learning blind noise distribution \cite{chen2018image, zamir2020cycleisp}. Since the camera settings are unknown (i.e. the ISP of old cameras is extremely hard to acquire), and the unsupervised training methods also assume a noise distribution, we alternatively propose to learn the noise prior on unpaired legacy photos and clean images by a NEGAN.
Based on the CycleGAN framework \cite{goodfellow2014generative, zhu2017unpaired}, we proposed the NEGAN to estimate the blind noise model. Firstly, we notice that the noisy regions normally include more high-frequency components than common regions; whereas flatten (or noise-free) areas comprise the low-frequency components. Thus, we utilize discrete wavelet transform (DWT) to extract the high-frequency components of generated images and real noisy photos, which are used for computing the domain adversarial loss. Secondly, we randomly select patches rather than resizing whole images, in order to maintain the low-level statistics. In addition, we collect a legacy photo dataset (LP dataset), which contains more than 25000 old photos with different levels of noises.
If the NEGAN is well trained, we can obtain training pairs by manually degrading the clean images from a large-scale dataset, such as ImageNet \cite{russakovsky2015imagenet}. The degraded images have similar statistics with legacy photos. Thus, the following inpainting and colorization processes are based on paired data. The solutions can be briefly categorized into reference-based and automatic. To improve the quality of generated images, we propose to perform the image editing by an IEGAN, i.e. reference-based inpainting and colorization. For the inpainting, it is hard to annotate the real cracks on each legacy photo. Thus, we alternatively collect some templates for modelling the cracks. By multiplying the cracks and clean images, we can obtain the masked images. While for colorization, we use the color scribbles as additional input guidance to enhance colorization reality. The adversarial losses used in IEGAN aim to improve perceptually quality of generated images.
We evaluate the proposed pipeline on ImageNet \cite{russakovsky2015imagenet} validation set. Compared with previous pipelines (i.e. denoising with AWGN, inpainting, and colorization networks), the proposed system achieves the best perceptual performance. Also, we visualize some samples in LP dataset in Figure \ref{teaser} (resolution 1760$\times$1760) and details in Figure \ref{teaser} (resolution 256$\times$512). Since the images in LP dataset generally have little cracks, we manually add the masks to real legacy photos for better visualization.
The main contributions of this paper are as follows:
1) We propose a novel NEGAN for estimating blind noise of legacy photos using unpaired data;
2) We create a new legacy photo dataset (LP dataset) including different types of degradation of real legacy photos for learning noise prior;
3) We propose an IEGAN that jointly performs denoising, inpainting and colorization in a user-guided way based on the noise prior estimated by NEGAN.
\section{Related Work}
\textbf{Image Denoising.} Image denoising is a fundamental problem in low-level vision. Recently, researches have shown that deep learning technologies outperform traditional methods such as bilateral filtering, BM3D \cite{dabov2007image}, non-local algorithm \cite{buades2005non}. Mao et al. \cite{mao2016image} designed U-Net shaped network to perform image denoising, which was improved by DnCNN \cite{zhang2017beyond} using residual learning and MemNet \cite{tai2017memnet} using long memory. Using a tunable noise level map as the input, FFDNet \cite{zhang2018ffdnet} handled a wide range of noise levels and removed spatially variant noise. Considering both Gaussian-Poisson Model and in-camera processing pipeline, CBDNet \cite{guo2019toward} further improved the blind denoising ability by embedding a noise estimation network. To further improve the network architecture, MWCNN \cite{liu2018multi} utilized DWT to avoid down-sampling information loss. SGN \cite{gu2019self} greatly decreased the memory consumption and runtime, while it was further improved by DSWN \cite{liu2020densely} using residual path and reconstruction path.
\textbf{Image Inpainting.} The image inpainting denotes the process of filling cracks of images. Normally, the masks of corresponding masked images are known. Pathak et al. \cite{pathak2016context} firstly adopted a conditional GAN \cite{goodfellow2014generative} for context completion. It was enhanced by jointly utilizing global and local discriminators by Iizuka et al. \cite{iizuka2017globally} to strengthen sharpness for filled regions. Liu et al. \cite{liu2018image} introduced a partial convolution with automatically updated status to deal with irregular input masks. It was improved by gated convolution \cite{yu2019free}. It is the combination of vanilla convolution and gate state, which generalizes the partial convolution by a learnable dynamic feature selection mechanism. The EdgeConnect \cite{nazeri2019edgeConnect} proposed an edge generator and image completion network to minimize blurry effect. Xiong et al. \cite{xiong2019foreground} further enhanced it for foreground-aware image inpainting.
\textbf{Image Colorization.} The existing colorization methods can be briefly categorized into three classes: scribble-based \cite{levin2004colorization, xu2013sparse, chen2012manifold, zhang2017real}, example-based \cite{ironi2005colorization, reinhard2001color, welsh2002transferring, he2018deep, iizuka2019deepremaster}, and fully-automatic \cite{cheng2015deep, zhang2016colorful, iizuka2016let, deshpande2017learning}. The former two kinds of approaches are user-guided that learn a mapping function to propagate user hint to the grayscale image. Since grayscale images only include the edge information, the results are highly relevant to the reasonability of human hints. On the other hand, fully-automatic algorithms directly solve an end-to-end objective from grayscale images to corresponding color embeddings. Normally, these approaches are trained on a very large dataset, which is essential for the system to exploit necessary information from the large-scale database without any human intervention.
\textbf{Generative Adversarial Network for Image Enhancement.} The image enhancement is a general idea to improve the image quality. It is addressed by a list of sub-tasks including demosaicking \cite{zhao2019saliency, chen2018learning}, deblurring \cite{kupyn2018deblurgan, kupyn2019deblurgan}, super-resolution \cite{wang2018esrgan, zhang2019ranksrgan, Xuehui_2020_ACCV}, etc. The performance of image enhancement has been greatly improved through the data-driven deep learning approaches. Generative adversarial network (GAN), developed by Goodfellow et al. \cite{goodfellow2014generative}, defines a minmax game between generator and discriminator. The goal of generator is producing convincing samples which fool discriminator, so as to distinguish generated samples from ground truth. The first well-known general GAN-based image enhancer is Pix2Pix \cite{isola2017image} that translates the images from two different domains. It was improved by Wang et al. \cite{wang2018high} for processing high-resolution images and Zhu et al. \cite{zhu2017unpaired} for multimodal generation.
\section{Methodology}
\subsection{Problem Formulation}
Suppose the clean images are from the domain $Z$ and legacy noisy images are in domain $N$. The target is to process the legacy photo $n \in N$ and obtain colorful clean image $z \in Z$. The images in both domain $Z$ and $N$ are totally different in terms of noise, content, and color.
However, the spatial pixels of clean image $z$ and legacy photo $n$ are not aligned. To constitute valid training pairs, we propose to decolorize $z$ and add pseudo noise to clean image $x \in X$ to obtain image $\hat{x} \in N$, which exists similar low-level characteristics of $n \in N$. We utilize a neural network $G$ to simulate the blind noise for clean image $x$. This transformation process can be formulated as:
\begin{equation}
\hat{x} = G ( x ).
\label{pf1}
\end{equation}
In order to recover colorful clean image $z$ from the artificial degraded image $\hat{x}$, we summarize the process as denoising, inpainting, and colorization, respectively. Similarly, they are implemented by neural networks due to the highly non-linear process. To simulate random mask of legacy photos, we use several binary mask samples $m$ to process $\hat{x}$ and obtain masked degraded image $\ddot{x} = \hat{x} \times m$. In addition to input image, we provide the masks and color scribbles to edit final colorized image and obtain sound perceptual quality. It can be represented as:
\begin{equation}
z = col ( inp ( den(\ddot{x}) ), s ),
\label{pf2}
\end{equation}
where $col(*)$, $inp(*)$ and $den(*)$ represent the colorization, inpainting, and denoising operations, respectively. The $s$ is the color
scribble provided by user. In practice, we combine $den(*)$ and $inp(*)$ into one architecture $C$ to accelerate inference, while $col(*)$ is implemented by network $R$.
\begin{figure*}[t]
\centering
\includegraphics[height=6.7cm, width=\linewidth]{img/figure3.pdf}
\caption{Illustration of training pipeline of proposed method. It contains four convolutional neural networks $G$, $F$, $C$ and $R$. The left part (NEGAN) represents the process that learns noise prior. The right part (IEGAN) shows image editing procedure including the joint denoising, inpainting and scribble-based colorization.}
\label{network}
\end{figure*}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{img/figure4.pdf}
\caption{Illustration of testing stage of proposed system. The user-provided mask and color scribble map assist the system to produce photorealistic colorizations from legacy photos.}
\label{three}
\end{figure}
\subsection{Training and Testing Pipeline}
Figure \ref{network} shows the training of the proposed pipeline. Specifically, the left and right part of Figure \ref{network} correspond to representations of equation \ref{pf1} and \ref{pf2}, respectively. They are concluded by two architectures, i.e. NEGAN and IEGAN, where NEGAN comprises the sub-network $F$ and $G$, IEGAN comprises the sub-network $C$ and $R$. The NEGAN is trained firstly. Then, it is used to degrade clean images, which are for IEGAN training. To stabilize training of IEGAN sub-networks, we propose to enforce a loss on $C$ directly. Figure \ref{three} shows the testing process, where only $C$, $R$ with additional mask and color scribble are adopted to edit legacy photos. The network architectures and the training details will be presented in following paragraphs.
\subsection{Noise Estimation GAN}
The Noise Estimation GAN (NEGAN) including $G$ and $F$ aims to implement equation \ref{pf1}, i.e. NEGAN translates the images $x \in X$ to noisy image domain $N$. Since there is no evident noise model and paired training data in our application, the unique characteristics of noise become significant. A complete image is composed of low-frequency and high-frequency parts and we notice the noise occupies most high-frequency components of images. Based on the observation, we propose to utilize a clean image $x \in X$ and keep the original low-frequency part $x_L$. Then, we replace its high-frequency component $x_H$ with statistics of noisy image from domain $N$ to implement the translation. Therefore, we propose a Noise Estimation GAN (NEGAN) based on unpaired images to learn the implicit noise distribution, which is called noise prior in following text.
To divide the low-frequency and high-frequency parts, we need to map the images into frequency domain. The common way is to utilize low-pass and high-pass filters, e.g. Gaussian filter and its inverse. It can be defined as:
\begin{equation}
x = x_L + x_H = w \ast x + (\delta - w) \ast x,
\label{pf3}
\end{equation}
where $x$, $w$, $\delta$ represent clean image, low-pass filter, and impulse function, respectively, while ($\delta - w$) is viewed as high-pass filter since it is the reverse of filter $w$. The ``$\ast$'' is convolution operator. But kernel $w$ is often set artificially, which cannot well separate different frequencies. To improve the functionality of the kernel, we introduce discrete wavelet transform (DWT) for frequency division and inverse discrete wavelet transform (IDWT) for image construction. Suppose two components $x_L$ and $x_H$ of input image $x$ are derived from DWT, the whole learning losses for training NEGAN can be represented as:
\begin{equation}
\begin{aligned}
L_{low}(G, F) &= \mathbb{E}[||G(x)_L - x_L||_1] \\ &+ \mathbb{E}[||F(G(x))_L - x_L||_1],
\end{aligned}
\label{negan1}
\end{equation}
\begin{equation}
\begin{aligned}
L_{high}(G, D_N, X, N) &= \mathbb{E}_{x \thicksim X}[||( D_N( G(x)_H ) )^2||] \\ &+ \mathbb{E}_{n \thicksim N}[||( D_N(n_H) - 1 )^2||],
\end{aligned}
\label{negan2}
\end{equation}
\begin{equation}
\begin{aligned}
L_{cycle}(G, F) &= \mathbb{E}_{x \thicksim X}[|| F(G(x)) - x ||_1] \\ &+ \mathbb{E}_{n \thicksim N}[|| G(F(n)) - n ||_1],
\end{aligned}
\label{negan3}
\end{equation}
\begin{equation}
\begin{aligned}
L_{NEGAN} &= \lambda_{low} L_{low}(G, F) + \lambda_{cycle} L_{cycle}(G, F) \\
& + L_{high}(G, D_N, X, N) + L_{high}(F, D_X, X, N),
\end{aligned}
\label{negan4}
\end{equation}
where $G$, $F$, $D_X$, $D_N$ denote generator from domain $X$ to $N$, generator from domain $N$ to $X$, and their corresponding discriminators, respectively. The $x$ and $n$ are random samples from both domains. The $L_{high}$ utilizes the LSGAN loss term \cite{mao2017least}. The $L_{high}$ only matches the low-frequency part of images, which is different from CycleGAN. Also, the discriminators distinguish between fake and real noisy images by matching only the high-frequency part.
\subsection{Image Editing GAN}
The second step of proposed method is to recover a high-quality image from the pseudo noisy image by an Image Editing GAN (IEGAN). The inference of IEGAN is divided into two sub-networks: inpainting network ($C$) and colorization network ($R$). The $C$ generates a complete grayscale image and the $R$ colorizes the output of $C$. As shown in Figure \ref{network}, the proposed IEGAN framework receives pseudo noisy grayscale image with additional mask and color map guidances.
We utilize L1 loss for both sub-networks $C$ and $R$. The losses for them share same representations. It is defined as:
\begin{equation}
L_1 = \mathbb{E}[||t_1 - t_2||_1] ,
\label{iegan1}
\end{equation}
where the two variables $t_1$ and $t_2$ equal to $\hat{y}$ and $x$ for $C$, meanwhile they equal to $\hat{z}$ and $z$ for $R$. The input $\ddot{x} = \hat{x} \odot m$ is a masked grayscale image with an additional Gaussian noise added. The outputs $\hat{y} = C(\ddot{x}, m)$ and $\hat{z} = R(y, s)$. The definitions can be found in Figure \ref{network}.
To boost perceptual quality of generated images, we adopt perceptual loss \cite{johnson2016perceptual}, which is defined as:
\begin{equation}
L_{percep} = \mathbb{E}[||\phi_l (t_1) - \phi_l (t_2)||_1] ,
\label{iegan2}
\end{equation}
where $\phi_l (*)$ represents the features of the $l$-th layer of the pre-trained CNN. In our experiment, we use the $conv_{4\_3}$ layer of VGG-16 \cite{simonyan2014very} network, which is pre-trained on ImageNet \cite{russakovsky2015imagenet} dataset.
Instead of traditional GAN training method \cite{goodfellow2014generative}, we utilize the PatchGAN \cite{isola2017image} with LSGAN critic \cite{mao2017least} to minimize the Pearson $\chi^2$ divergence between the generated samples and ground truth. It is defined as:
\begin{equation}
L_{G} = \cfrac{1}{2} \mathbb{E}[(D(t_1) - 1)^2],
\label{iegan3}
\end{equation}
\begin{equation}
L_{D} = \cfrac{1}{2} \mathbb{E}[(D(t_2) - 1)^2] + \cfrac{1}{2} \mathbb{E}[(D(t_1))^2].
\label{iegan4}
\end{equation}
The total loss functions of IEGAN can be defined as:
\begin{equation}
L_{IEGAN} = L_{1C} + L_{1R} + \lambda_{percep} L_{percepR} + \lambda_{G} L_{GR},
\label{iegan5}
\end{equation}
where inpainting network $C$ only adopts L1 loss term $L_{1C}$. The colorization network $R$ utilizes all three loss terms $L_{1R}$, $L_{percepR}$, and $L_{GR}$. The definitions of the loss terms can also be found in Figure \ref{network}.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{img/figure5.pdf}
\caption{Illustration of mask templates used in this paper.}
\label{mask}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{img/figure6-2.pdf}
\caption{Illustration of image editing results. The input masked images are obtained by multiplication operation of grayscale images and masks. Different columns represent different samples edited by methods in experiment. They are randomly selected from validation set.}
\label{compare_fig}
\end{figure*}
\section{Experiment}
\subsection{Implementation Details}
\textbf{Dataset.} We use LP dataset to include enough modes of noisy image domain $N$, for NEGAN training. There are over 25000 grayscale legacy photos with different resolutions in the dataset. Also, we choose ImageNet \cite{russakovsky2015imagenet} (1.3 million images) for clean image domain $X$. It contains 1000 categories, which is general and robust for learning the mapping. At training, we randomly select unpaired sample $n \in N$ and $x \in X$. The images are randomly cropped to 256$\times$256 local patches and normalized to range of [0, 1]. Moreover, the binary mask samples $m$ is randomly cropped from templates, as shown in Figure \ref{mask}.
\textbf{Network Architecture.} For NEGAN architecture, the generators adopt 8 residual blocks \cite{he2016deep} as transformer with residual connection between input and output. There are no downsampling and upsampling operations since they may affect the low-level details. The discriminators adopt $16 \times 16$ PatchGAN architecture and all layers are spectral normalized \cite{miyato2018spectral}. The pre-trained NEGAN produces corresponding degraded images from input while maintains the low-frequency parts. For IEGAN architecture, the generator $C$ and $R$ adopt U-Net structure \cite{ronneberger2015u}. The convolutional layer of $C$ is replaced by gated convolution \cite{yu2019free} to learn adaptive inpainting. The discriminator $D_C$ and $D_R$ adopt convolution part of a VGG-16 architecture while the final output is one channel. The networks are instanced normalized \cite{ulyanov2016instance}. Each layer is LeakyReLU activated \cite{maas2013rectifier}.
\textbf{Optimization.} At first stage, the parameters of all networks are initialized using Xavier method \cite{glorot2010understanding} and the learning rate is initialized as $1 \times 10^{-4}$. The NEGAN and two sub-networks of IEGAN are trained independently for 20 epochs. At second stage, all networks are optimized jointly. The learning rate is fixed to $5 \times 10^{-5}$ while the system is trained for another 20 epochs. The learning rate is fixed in both stages. We use Adam optimizer \cite{kingma2014adam} with $\beta_1$ = 0.5, $\beta_2$ = 0.999 and batch size of 32. Moreover, we randomly select 0 - 30 color scribbles as hint for network $R$. The hyperparameters $\lambda_{percep}, \lambda_{G}$ equal to 10 and 0.1, respectively. We implement our system with PyTorch framework and train it on 4 NVIDIA Titan Xp GPUs. It takes approximately 2 weeks to complete the whole training process.
\begin{table}[t]
\caption{Comparison results of the proposed pipeline and other 6 state-of-the-art pipelines. The grayscale images (clean) are obtained from ground truth colorful images. In ``Reference'' item, the ``mask'' and ``color'' denote the additional mask and color scribble input. Also, the \cite{zhang2017beyond,yu2019free,zhang2016colorful} represents using \cite{zhang2017beyond}, \cite{yu2019free}, \cite{zhang2016colorful} for inference sequentially.}
\begin{center}
\begin{tabular}{lccc}
\hline
Method & Reference & PSNR & SSIM \\
\hline
\hline
Grayscale (clean) & / & 23.24 & 0.9443 \\
\hline
\cite{zhang2017beyond,yu2019free,zhang2016colorful} & mask & 21.26 & 0.8865 \\
\hline
\cite{gu2019self,yu2019free,zhang2016colorful} & mask & 21.18 & 0.8865 \\
\hline
\cite{zhang2017beyond,yu2019free,zhang2017real}v1 & mask & 23.62 & 0.9059 \\
\hline
\cite{zhang2017beyond,yu2019free,zhang2017real}v2 & mask, color & 27.51 & 0.9233 \\
\hline
\cite{gu2019self,yu2019free,zhang2017real}v1 & mask & 23.50 & 0.9024 \\
\hline
\cite{gu2019self,yu2019free,zhang2017real}v2 & mask, color & 27.34 & 0.9194 \\
\hline
Proposed & mask, color & \textbf{28.02} & \textbf{0.9408} \\
\hline
\end{tabular}
\end{center}
\label{comparison}
\end{table}
\subsection{Validation on Image Editing Quality}
In this section, we quantitatively evaluate the image enhancement quality of the proposed system. Since there is no ground truth for legacy photos, we alternatively adopt the ImageNet validation 50000 images. We convert the images to grayscale and rescale them to 256$\times$256. Each validation image is added a pseudo mask and an additive Gaussian noise with standard deviation of 0.05 to simulate a legacy image, which is similar to training process. At inference stage, only IEGAN is used since the noise prior modelled by NEGAN is implied in $C$ at training. We utilize different combinations of denoisers \cite{zhang2017beyond, gu2019self}, inpainting network \cite{yu2019free}, and colorization networks \cite{zhang2016colorful,zhang2017real} as pipelines and there are overall 6 combinations. All aforementioned algorithms are trained on ImageNet training dataset. Specifically, the denoisers are trained on the same noise level (i.e. AWGN) as validation data, whereas IEGAN is trained on blind noise learned from noise prior. The method \cite{zhang2017real} is a scribble-based colorization algorithm while \cite{zhang2016colorful} is fully-automatic colorization method. Color scribbles are used in both IEGAN and method \cite{zhang2017real}; therefore all the approaches in experiment adopt reference information. There are 30 scribbles used for IEGAN and for \cite{zhang2017real} at test.
The comparison results are summarized in Table \ref{comparison} and illustrated in Figure \ref{compare_fig}. The methods using color scribbles achieve the PSNR higher than 27 and obviously outperform others since they have precise color prior. Note that, the proposed method achieves the highest PSNR and SSIM since the two parts of IEGAN are trained collaboratively. Therefore, the colorized images are more natural and realistic than other methods.
\begin{table}[t]
\caption{Comparison results for ablation study.}
\begin{center}
\begin{tabular}{lccc}
\hline
Methods & Setting & PSNR & SSIM \\
\hline
\hline
w/o DWT-based losses & 1) & 27.91 & 0.9397 \\
\hline
w/o perceptual loss & 2) & 27.78 & 0.9369 \\
\hline
w/o GAN loss & 2) & 27.83 & 0.9396 \\
\hline
w/o both losses & 2) & 26.18 & 0.9334 \\
\hline
10 color scribbles & 3) & 26.78 & 0.9354 \\
\hline
20 color scribbles & 3) & 27.59 & 0.9390 \\
\hline
w/o joint training & 4) & 27.45 & 0.9358 \\
\hline
Proposed & / & \textbf{28.02} & \textbf{0.9408} \\
\hline
\end{tabular}
\end{center}
\label{ablation}
\end{table}
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{img/figure7-2.pdf}
\caption{Comparison of legacy photo enhancement results of the proposed and previous \cite{gu2019self, yu2019free} pipeline. The left part and right part include the full resolution legacy photos and local patches, respectively. The colorful rectangles denote the locations of selected patches.}
\label{denoised}
\end{figure*}
\subsection{Ablation Study}
In order to demonstrate the effectiveness of NEGAN and IEGAN losses, we set up 4 ablation study settings. We use 50000 ImageNet validation data for validation. All images are added unknown noise by pre-trained NEGAN to simulate legacy photos. The settings are show as:
1) Drop the DWT-based loss terms that NEGAN noise prior learner retrogrades to a CycleGAN \cite{zhu2017unpaired};
2) Drop the perceptual loss or GAN loss or both loss terms of IEGAN to compare their effectiveness, while the NEGAN remains unchanged;
3) Decrease the number of color scribbles to 20 or 10;
4) Train two sub-networks of IEGAN framework separately in order to evaluate joint training scheme.
As shown in Table \ref{ablation}, the full system reaches the best performance on PSNR and SSIM \cite{wang2004image}. If the DWT-based loss terms are dropped, the system is hard to handle the ``real noise'' generated by the NEGAN. Also, each loss term or joint training contributes to better performance. In conclusion, all components of proposed method and significant.
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{img/figure8.pdf}
\caption{Comparison of the proposed and previous \cite{gu2019self, yu2019free, zhang2017real} pipeline on real legacy photos. The rectangles denote the highlighted areas.}
\label{legacy}
\end{figure*}
\subsection{Validation on Legacy Photo Enhancement}
In this section, we assess the denoising and inpainting ability of the proposed system, i.e. network $C$ of IEGAN. The state-of-the-art denoising and inpainting methods \cite{gu2019self, yu2019free} are used for comparison. For the denoising, the results of proposed approach are more sharper than \cite{gu2019self, yu2019free}. For instance, the eyebrows, cheeks and beard generated by the proposed method are more clear, as shown in the second patch in Figure \ref{denoised}. For inpainting, the patches produced by the proposed method are also realistic. For instance, the color of filled regions are closer to clothes, as shown in the first patch. Also, the patch of proposed model in third row is much more smoother than \cite{gu2019self, yu2019free}. Since the NEGAN better estimates the noise model, the generated results are cleaner and sharper. Moreover, the inpainted regions are more plausible due to better denoising ability.
\subsection{Validation on Legacy Photo Colorization}
In this section, we assess the editing quality of the proposed system on real legacy photos. We utilize the state-of-the-art pipelines, i.e. \cite{gu2019self, yu2019free, zhang2017real} for comparison and 15 color scribbles are adopted for both methods, as shown in Figure \ref{legacy}. The samples are selected from the proposed LP dataset. The proposed method produces more plausible colors than compared method since the it utilizes joint training scheme for image denoising, inpaiting and colorization. Moreover, the proposed method learns noise prior well, thus it produces high-quality images.
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{img/figure9.pdf}
\caption{Illustration of some failure image editing cases of proposed method, including color bleeding, artifacts and inconsistent colors.}
\label{failure}
\end{figure*}
\section{Failure Cases}
Our system can predict relatively reasonable colorizations in many cases; however, there are still some common failure cases, shown in Figure \ref{failure}. For left part (left 3 rows in the figure), there exists slight color bleeding effect when given “not reasonable color scribbles”. Since the degradation degree is shifted for many legacy photos, the output images of center part still contain artifacts. Finally, as color scribbles provided by users are not reasonable enough, the results are also not very plausible, as right part shows. We will enhance the design of the proposed framework and. Moreover, we will add semantic information into our framework to guide inpainting and colorization in the future.
\section{Conclusion}
In this paper, we present a novel framework for editing legacy photos in an end-to-end manner. Since the legacy photographs are captured by old cameras, they are corrupted with undesirable noise, artifacts and saved in grayscale format. The noise is often blind, thus it is difficult to use a specific distribution for modelling. Thus, we propose a NEGAN to simulate noise prior learned from real legacy photos based on unpaired data training. We enforce the NEGAN to focus more on noisy parts (i.e. high-frequency components) of images by introducing DWT-based loss functions. Moreover, we collect a large-scale legacy photo dataset (LP dataset) including more than 25000 real photographs in different scenes for training NEGAN. Moreover, to remove the artifacts and colorize legacy photos, we propose an IEGAN that performs joint denoising, inpainting and scribble-based colorization sequentially, based on estimated noise prior. At test phase, users can edit the legacy photo by providing masks and color scribbles. Experimental results show that the proposed framework has better performance than the state-of-the-art pipelines.
\section*{Acknowledgement}
The authors would like to thank the colleagues of City University of Hong Kong for their support for collecting legacy photos. The authors would also like to thank the anonymous reviewers and conference chairs for their kind suggestions.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}
\label{sec:introduction}
Spin-transfer torque magnetic random access memory (STT-MRAM) is considered as one of the most promising non-volatile memory technologies, since it features high density, nearly unlimited endurance, and negligible leakage power \cite{Fong2016}.
Thus, many foundries worldwide have been investing heavily in its commercialization.
For example, SK hynix demonstrated in 2016 \cite{Chung2016} a 4\,Gb STT-MRAM prototype targeting the replacement of DRAM and flash memories.
Samsung and Intel also presented their STT-MRAM solutions in 2018 \cite{SongYJ2018} and 2019 \cite{Wei2019a}, respectively.
In STT-MRAMs, data is stored in magnetic tunnel junction (MTJ) devices which contain multiple ferromagnetic layers. Each of them generates a stray field, which has a significant impact on the device's performance. It has been shown that the stray field increases as the MTJ dimension shrinks \cite{Han2015, Wu2020survey}, which makes \textit{intra-cell magnetic coupling} a critical constraint for STT-MRAM designs at advanced technology nodes. Furthermore, to compete with DRAM and flash memories, high-density STT-MRAM arrays are required. It was reported that the STT-MRAM array pitch can be made as small as 1.5$\times$ the MTJ diameter at sub-\SI{20}{\nano\meter} nodes, using advanced nano-patterning techniques \cite{Nguyen2018}. As the pitch decreases, MTJ devices are pushed closer to each other. This makes \textit{inter-cell magnetic coupling} between neighboring cells become increasingly evident, which may lead to write errors \cite{chappert2010emergence}. Therefore, it is crucial to analyze magnetic coupling and evaluate its impacts quantitatively when it comes to high-density STT-MRAMs at advanced technology nodes.
There is limited work published on magnetic coupling considering both intra- and inter-cell effects in dense STT-MRAM arrays based on designs that were manufactured.
In \cite{Jiancheng2015}, Huang et al. observed experimentally that the stray field results in a distinct difference in the thermal stability factor $\Delta$ of the two binary states of MTJ devices.
Wang et al. \cite{Wang2012} reported that the stray field has a non-uniform distribution over the cross-section of the MTJ device, and it results in a significant variation in the switching time based on micromagnetic simulations. In \cite{Golonzka2018}, Golonzka et al. even observed that some of Intel's devices were locked to one certain state due to a strong stray field. However, these papers did not take into consideration the stray fields from neighboring cells. Augustine et al. \cite{Augustine2010} analyzed the stray field from four adjacent cells for in-plane MTJ devices and concluded that inter-cell stray fields can cause up to 80\% increase in switching time. Yoon et al. \cite{Yoon2018} explored the effect of inter-cell magnetic coupling on the MTJ's properties in a compact memory array. Overall, the related prior work has the following shortcomings: 1) not based on real-world MTJ stack designs with perpendicular magnetic anisotropy, dual MgO, and synthetic anti-ferromagnetic (SAF) pinned layer; 2) intra- and inter-cell magnetic coupling are not considered simultaneously; 3) a lack of magnetic characterization data.
In this paper, we address these issues by characterizing and subsequently modeling intra-cell magnetic coupling for isolated MTJ devices fabricated at IMEC. This model is then used to calculate stray fields at the victim cell located in the center of a representative 3$\times$3 memory array. The contributions of this paper are as follows.
\vspace*{-2pt}
\begin{itemize}
\item Magnetic characterization results of MTJ devices with various sizes ranging from \SI{35}{\nano\meter} to \SI{175}{\nano\meter}.
\item An analytical intra-cell magnetic coupling model, which is calibrated and validated by silicon data.
\item Extrapolating the above model to study inter-cell magnetic coupling on a cell array with varying pitches.
\item Introduction of the inter-cell magnetic coupling factor $\Psi$ to indicate the coupling strength.
\item Evaluation of the impact of magnetic coupling on the MTJ's write characteristics and retention time.
\end{itemize}
The rest of this paper is organized as follows.
Section~\ref{sec:background} provides a background on the MTJ device technology and the magnetic coupling mechanism.
Section~\ref{sec:intracell_mc_characterization} presents magnetic characterization data.
Section~\ref{sec: MC_modeling} elaborates the modeling methodology of intra-cell and inter-cell magnetic coupling.
Section~\ref{sec:MC_impact} evaluates the impact of magnetic coupling on the MTJ's performance.
Section~\ref{sec:conclusion} concludes this paper.
\section{Background}
\label{sec:background}
\subsection{MTJ Device Technology}
{\em Magnetic tunnel junction} (MTJ) devices are the data-storing elements in STT-MRAMs.
Each MTJ device stores one-bit data in the form of binary magnetic configurations \cite{Wu2019TETC}. Fig.~\ref{fig:STTMRAM_basics}a shows the MTJ stack which essentially consists of four layers: FL/TB/RL/HL.
The {\em hard layer} (HL) is composed of [Co/Pt]\textsubscript{{x}}, which is used to pin the magnetization in the upper {\em reference layer} (RL). The RL is generally built up with a Co/spacer/CoFeB multilayer, which is anti-ferromagnetically coupled to the HL. These two layers form a {\em synthetic anti-ferromagnetic} (SAF) structure, providing a strong fixed reference magnetization in the RL. The {\em tunnel barrier} (TB) layer is made of dielectric MgO, typically $\sim$\SI{1}{\nano\meter}. The {\em resistance-area} (RA) product is commonly used to evaluate the TB resistivity, as it depends on the TB thickness but not the device size.
The CoFeB-based {\em free layer} (FL) is the data-storing layer where the magnetization can be switched by a spin-polarized current.
To work properly as memory devices, MTJs need to provide read and write mechanisms, which are realized by the {\em tunneling magneto-resistance} (TMR) effect and the {\em spin-transfer-torque} (STT) effect \cite{ Khvalkovskiy2013d}. Due to the TMR effect, the MTJ's resistance is low ($R_{\mathrm{P}}$) when the magnetization in the FL is parallel to that in the RL; the resistance is high ($R_{\mathrm{AP}}$) when in anti-parallel state.
If the write current magnitude (with sufficiently long pulse width) is larger than the {\em critical switching current} ($I_{\mathrm{c}}$), the magnetization in the FL can switch to the opposite direction. It is a fundamental parameter to characterize the switching capability by current. The STT-induced switching behavior also depends on the current direction, as shown in Fig.~\ref{fig:STTMRAM_basics}a. $I_{\mathrm{c}}$(AP$\rightarrow$P) can be significantly different from $I_{\mathrm{c}}$(P$\rightarrow$AP) due to the bias dependence of STT efficiency and external field disturbance \cite{Khvalkovskiy2013d}. In addition, the {\em average switching time} ($t_{\mathrm{w}}$) \cite{Wu2018} is another critical parameter, which is inversely correlated with the write current. In other words, the higher the write current over $I_{\mathrm{c}}$, the less the time required for the magnetization in FL to flip. In practice, $t_{\mathrm{w}}$(AP$\rightarrow$P) can also differ from $t_{\mathrm{w}}$(P$\rightarrow$AP) depending on the write current magnitude and duration.
In addition, enough retention time is required for STT-MRAMs depending on the target application. Storage applications require $>$10\,years typically, while cache applications only necessitate ms-scale retention time \cite{Jog2012a}. an STT-MRAM retention fault occurs when the magnetization in the FL of the MTJ flips spontaneously due to thermal fluctuation. Thus, the STT-MRAM retention time is generally characterized by the {\em thermal stability factor} ($\Delta$) \cite{Khvalkovskiy2013d}. The higher the $\Delta$, the longer the retention time.
\begin{figure}[!t]
\centering
\includegraphics[width=0.5 \textwidth ]{Background}
\caption{ (a) MTJ stack and the intra-cell stray fields from the RL and HL, (b) 3$\times$3 MTJ array and the inter-cell stray fields from neighboring cells, (c) SEM image of the 0T0R wafer floor plan, and (d) SEM image of MTJ array.}
\label{fig:STTMRAM_basics}
\end{figure}
\subsection{Magnetic Coupling Mechanism }
To obtain high TMR and strong interfacial perpendicular magnetic anisotropy (iPMA), our MTJ devices were annealed at \SI{375}{\celsius} for \SI{30}{\minute} in a vacuum chamber under the perpendicular (out-of-plane) magnetic field of \SI{20}{\kilo Oe}. Once the ferromagnetic layers (i.e., FL, RL, and HL) in the MTJ stack are magnetized, each of them inevitably generates a stray field in the space. Fig.~\ref{fig:STTMRAM_basics}a illustrates the intra-cell stray field ($\bm{H_{\mathrm{s\_intra}}}$) perceived at the FL, generated by the RL and HL together; its in-plane component ($H_{\mathrm{s\_intra}}^{x-y}$) is marginal \cite{Wang2012}, while its out-of-plane component ($H_{\mathrm{s\_intra}}^{z}$) at the FL has a significant influence on the energy barrier ($E_{\mathrm{b}}$) between the P and AP states \cite{Fong2016}. For example, if $H_{\mathrm{s\_intra}}^{z}$ has the same direction as the magnetization in the FL in AP state, it leads to an increase in $E_{\mathrm{b}}$(AP$\rightarrow$P) and a decrease in $E_{\mathrm{b}}$(P$\rightarrow$AP). The deviation of energy barriers along the two switching directions has a significant impact on the retention and the STT-switching characteristics of MTJ devices, as reported in \cite{Wang2012,Jiancheng2015,Golonzka2018}.
In the extreme case where $H_{\mathrm{s\_intra}}^{z}$ exceeds the FL {\em coercivity} ($H_{\mathrm{c}}$), defined as the reverse field needed to drive the magnetization of a ferromagnet to zero, the bistable states will disappear \cite{Han2015}.
Furthermore, as the density of STT-MRAMs increases, the spacing between neighboring MTJ devices becomes narrower (i.e., smaller pitch). This makes stray fields from neighboring cells not negligible any more \cite{Augustine2010,Fong2016}.
Fig.~\ref{fig:STTMRAM_basics}b shows a 3$\times$3 MTJ array, where the eight cells C0-C7 (aggressors) surrounding cell C8 (victim) in the center inevitably generate an inter-cell stray field ($\bm{H_{\mathrm{s\_inter}}}$) acting on the victim cell. Fig.~\ref{fig:STTMRAM_basics}c and Fig.~\ref{fig:STTMRAM_basics}d show the scanning electron microscope (SEM) images of our 0T1R wafer floor plan and MTJ array, respectively.
\section{ Intra-Cell Magnetic Coupling Characterization}
\label{sec:intracell_mc_characterization}
$H_{\mathrm{s\_intra}}^{z}$ can be extracted from R-H hysteresis loops. Fig.~\ref{fig:Hs_characterization}a shows a measured R-H hysteresis loop for a representative MTJ device with the HL/RL configuration shown in Fig.~\ref{fig:STTMRAM_basics}a.
During the measurement, an external field was applied perpendicularly to the device under test. It was ramped up from \SI{0}{Oe} to \SI{3}{kOe}, then it went backwards to \SI{-3}{kOe} and finished at \SI{0}{Oe}. In total, we measured 1000 field points, each of which was followed by a read operation to read out the device resistance with a voltage of \SI{20}{\milli\volt}. It can be seen that the MTJ device switches from AP state (high resistance) to P state (low resistance) when the field reaches at $H_{\mathrm{sw\_p}}$, and it switches back to AP state at a negative field $H_{\mathrm{sw\_n}}$. The device coercivity can be obtained by $H_{\mathrm{c}}=(H_{\mathrm{sw\_p}}-H_{\mathrm{sw\_n}})/2$. Due to the existence of stray fields at the FL, the loop is always offset to the positive side for the device configuration in Fig.~\ref{fig:STTMRAM_basics}a. The offset field $H_{\mathrm{offset}}$ is equal to $(H_{\mathrm{sw\_p}}+H_{\mathrm{sw\_n}})/2$, as shown in the figure. Since $H_{\mathrm{offset}}$ is essentially equivalent to the extra external field applied to cancel out $H_{\mathrm{s\_intra}}$, the relation of these two parameters is $H_{\mathrm{s\_intra}}=-H_{\mathrm{offset}}$. Given the fact that the {\em resistance-area product} ($\mathit{RA}$) does not change with the device size, the {\em electrical Critical Diameter} (eCD) of each device can be derived by \cite{JWu2018}:
$eCD=\sqrt{\frac{4}{\pi}\cdot \frac{RA}{R_{\mathrm{P}}}},$
where $\mathit{RA}$$=$$4.5$\,\si[inter-unit-product = \ensuremath{\cdot}]{\ohm\um\squared}\xspace (measured at blanket stage) for this wafer, and $R_{\mathrm{P}}$ can be extracted from the R-H loop (i.e., the lower horizontal line in Fig.\ref{fig:Hs_characterization}a). The calculated eCD=\SI{55}{\nano\meter} for the device shown in Fig.\ref{fig:Hs_characterization}a.
In this way, we can obtain $H_{\mathrm{s\_intra}}^{z}$ and eCD for MTJ devices with different sizes on the same wafer. The measurement results are shown in Fig. \ref{fig:Hs_characterization}b. The error bars indicate the device-to-device variation in the the measured values due to process variations and the intrinsic switching stochasticity.
It can be seen that the smaller the device size (i.e., smaller eCD), the higher $H_{\mathrm{s\_intra}}^{z}$; the trend even tends to grow exponentially for eCD$<$\SI{100}{\nano\meter}. The solid curve in the figure represents simulation results which will be explained in the next section.
\begin{figure}[!t]
\centering
\includegraphics[width=0.5 \textwidth ]{Hs_intra_characterization}
\caption{(a) Measured R-H hysteresis loop, (b) device size dependence of $H_{\mathrm{s\_intra}}^z$: measured vs. simulated.}
\label{fig:Hs_characterization}
\end{figure}
\section{Modeling of Magnetic Coupling}
\label{sec: MC_modeling}
To analyze the impact of magnetic coupling on the MTJ's performance, we need to develop an analytical model for both inter-cell and intra-cell magnetic coupling. In this section, we first model and calibrate $\bm{H_{\mathrm{s\_intra}}}$ for isolated MTJ devices, based on the measurement data as presented in the previous section. Thereafter, we extrapolate this model to derive $\bm{H_{\mathrm{s\_inter}}}$ for an memory array with various pitches.
\subsection{Intra-Cell Magnetic Coupling }
\label{subsec:Intra_MC_modeling}
Under the assumption that each ferromagnetic layer (i.e., FL, RL, and HL) in MTJ devices is uniformly magnetized, the produced field is identical to the field that would be produced by the bound current \cite{griffiths2013}. Fig. \ref{MC_modeling_results1}a depicts a thin ferromagnet with tiny current loops representing dipoles. All internal currents cancel each other while there is no adjacent loop at the edge to do the canceling. As a result, the net effect is a macroscopic current $I_{\mathrm{b}}$ (referred to as bound current) flowing around the boundary. The magnetic moment of this ferromagnet can be expressed as $\bm{m} =\bm{M_s}\cdot A \cdot t $ \cite{griffiths2013},
where $\bm{M_s}$ is the saturation magnetization, $A$ is the cross-sectional area, and $t$ is the thickness of this ferromagnet. Considering the bound current $I_{\mathrm{b}}$, $\bm{m}$ can also be written as $I_{\mathrm{b}}\cdot A \cdot \bm{\hat{n}}$
where $\bm{\hat{n}}$ is the unit vector along the direction of $\bm{M_s}$ \cite{griffiths2013}. Therefore, one can easily derive $I_{\mathrm{b}}=M_s t $. For each ferromagnet in the MTJ stack, the $M_st $ product is measured at blanket film level by vibrating sample magnetometry (VSM) measurements.
\begin{figure*}[!h]
\centering
\includegraphics[width=1 \textwidth ]{MC_modeling_results1}
\caption{(a) Bound current, (b) Biot-Savart law, (c) intra-cell stray field $\bm{H_{\mathrm{s\_intra}}}$ from the HL and RL for an MTJ with eCD=\SI{55}{\nano\meter}, and (d) the out-of-plane component $H_{\mathrm{s\_intra}}^z$ distribution over the cross-section of the FL, with respect to various eCDs.}
\label{MC_modeling_results1}
\end{figure*}
With the derived bound current $I_{\mathrm{b}}$ for each ferromagnet in the MTJ stack, the generated stray field in the space can be modeled as the field of a current loop with current $I_{\mathrm{b}}$, as shown in Fig.~\ref{MC_modeling_results1}b. In this way, the stray field at any point P($x_{\mathrm{p}}$,$y_{\mathrm{p}}$,$z_{\mathrm{p}}$) in the space can be calculated by the Biot-Savart law \cite{griffiths2013}:
\begin{equation}
\bm{H}(\bm{r})=\frac{\mu_{0}}{4\pi}\oint \frac{I_{\mathrm{b}}\bm{dl}\times \bm{r}}{|\bm{r}|^3}, \label{BS_law}
\end{equation}
where $\bm{dl}$ is an infinitesimal length of the current loop, $\bm{r}$ the vector distance from $\bm{dl}$ to the point P, and $\mu_{0}$ the vacuum permeability. To calculate the above integral in a discrete form, we can divide the current loop into a large number of small segments, thereafter sum up the fields of all segments at point P as an approximation of $\bm{H}(\bm{r})$.
Assume the current loop is cut into $N$ segments. For the $k^{th}$ segment $\bm{dl_k}$ ($k\in[0,N-1]$), we derive:
\begin{align*}
\bm{dl_k}&=(x_{k+1}-x_{k}, y_{k+1}-y_{k}, z_{k+1}-z_{k}), \\
\bm{r_k}&=(x_{p}-x_{k}, y_{p}-y_{k}, z_{p}-z_{k}).
\end{align*}
Therefore, $\bm{dl_k}\times \bm{r_k}=(S_k^x,S_k^y,S_k^z)$, where
\begin{align*}
S_k^x&=(y_{k+1}-y_{k})\cdot(z_{p}-z_{k})-(z_{k+1}-z_{k})\cdot(y_{p}-y_{k}), \\
S_k^y&=(z_{k+1}-z_{k})\cdot(x_p-x_k)+(x_{k+1}-x_k)\cdot(z_p-z_k), \\
S_k^z&=(x_{k+1}-x_k)\cdot(y_p-y_k)-(y_{k+1}-y_k)\cdot(x_p-x_k).
\end{align*}
The field generated by the tiny segment $\bm{dl_k}$ is
$$ \bm{dH_k}=(dH_k^x,dH_k^y,dH_k^z)=\frac{\mu_{0}}{4\pi} \cdot \frac{I_{\mathrm{b}}}{|\bm{r_k}|^3}\cdot(S_k^x,S_k^y,S_k^z).$$
By summing up the fields of all $N$ segments, we derive the overall field at the spot P generated by the entire current loop:
$$\bm{H}=\sum\limits_{k=0}^{N-1}\bm{dH_k} =(\sum\limits_{k=0}^{N-1}dH_k^x, \sum\limits_{k=0}^{N-1}dH_k^y, \sum\limits_{k=0}^{N-1}dH_k^z) $$
In this way, we can calculate the intra-cell stray field from the HL ($\bm{H_{\mathrm{s\_HL}}}$) and intra-cell stray field from the RL ($\bm{H_{\mathrm{s\_RL}}}$), respectively. The overall intra-cell stray field is the vector sum of these two fields (i.e., $\bm{H_{\mathrm{s\_intra}}}=\bm{H_{\mathrm{s\_HL}}}+\bm{H_{\mathrm{s\_RL}}}$), which is visualized in Fig.~\ref{MC_modeling_results1}c for an MTJ device with eCD=\SI{55}{\nano\meter}. Fig. \ref{MC_modeling_results1}d shows the distribution of the z-component $H_{\mathrm{s\_intra}}^{z}$ (i.e., the out-of-plane component) over the horizontal cross-section of the FL. It can be seen that $H_{\mathrm{s\_intra}}^{z}$ is not uniformly distributed at the FL; its magnitude is smaller at the edge than at the center. We took the values at the center (i.e., at radial position=\SI{0}{\nano\meter}) and calibrated them with the measured data. Fig.~\ref{fig:Hs_characterization}b presents the simulation results of $H_{\mathrm{s\_intra}}^{z}$ vs. eCD, which match the silicon data.
\subsection{Inter-Cell Magnetic Coupling }
\label{subsec:Inter_MC_modeling}
To study the inter-cell magnetic coupling effect, we extrapolate the intra-cell magnetic coupling model from a single MTJ device to a 3$\times$3 MTJ array in Cartesian Coordinates. The nine devices are named C0 to C8, as illustrated in Fig.~\ref{fig:STTMRAM_basics}b. Cell C8 in the center is considered as the victim whereas the four direct neighbors (C0-C3) and four diagonal neighbors (C4-C7) are aggressor cells. In this way, the inter-cell magnetic coupling effect is translated to the impact of net stray field from the eight neighboring cells (denoted as $\bm{H_{\mathrm{s\_inter}}}$) on the FL of the victim C8. $\bm{H_{\mathrm{s\_inter}}}$ can be calculated by:
$$
\bm{H_{\mathrm{s\_inter}}}=\sum_{i=0}^{7} (\bm{H_{\mathrm{s\_HL}}}(\mathrm{C}i)+\bm{H_{\mathrm{s\_RL}}}(\mathrm{C}i)+\bm{H_{\mathrm{s\_FL}}}(\mathrm{C}i)).$$
Since the HL and RL are both fixed layers after the fabrication of MTJ devices, $\bm{H_{\mathrm{s\_HL}}}$ and $\bm{H_{\mathrm{s\_RL}}}$ are fixed, given an eCD and a pitch node. However, the direction of $\bm{H_{\mathrm{s\_FL}}}$ changes dynamically depending on the data stored in the MTJ device though its magnitude remains the same. As a result, $\bm{H_{\mathrm{s\_inter}}}$ depends on the {\em neighborhood pattern} in the eight neighboring cells (i.e., C0-C7), which we denote as NP\textsubscript{8}. In the binary form, NP\textsubscript{8} can be expressed as:
$[d_0,...,d_7]_2$,
where $d_i\in\{0,1\}$ represents the data stored in C$i$. In addition, NP\textsubscript{8} can also be transformed to the decimal form: $[n]_{10}$, where $n\in[0,255]$.
Fig. \ref{MC_modeling_results2}a shows the resultant $H_{\mathrm{s\_inter}}^z$ values at the FL of victim C8 as a function of the number of 1s in direct neighbors C0-C3 (marked in yellow) and the number of 1s in diagonal neighbors C4-C7 (marked in skyblue). Since C0-C3 are in symmetric positions and C4-C7 are also in symmetric positions, there are 25 distinct combinations as shown in the figure.
For this example, we set eCD=\SI{55}{\nano\meter} and pitch=\SI{90}{\nano\meter} (design spec. from the SK hynix high-density STT-MRAM design in \cite{Chung2016}).
It can be seen that $H_{\mathrm{s\_inter}}^z$ reaches its lowest point (\SI{-16}{Oe}) when C0-C7 are all in 0 (P) state (i.e., NP\textsubscript{8}=0). In this case, the magnetization in the FL of every aggressor cell is in parallel with that of the RL; together, they generate a stray field which is stronger enough to compensate the stray field from the HL. As the bit number of 1s increases, $H_{\mathrm{s\_inter}}^z$ increases; it increases in a step of \SI{15}{Oe} with the number of 1s in direct neighbors and in a step of \SI{5}{Oe} with the number of 1s in the diagonal neighbors. When C0-C7 are all in 1 (AP) state (i.e., NP\textsubscript{8}=255), $H_{\mathrm{s\_inter}}^z$ reaches the peak (\SI{64}{Oe}). Therefore, the maximum variation in $H_{\mathrm{s\_inter}}^z$ among the 256 neighborhood patterns is \SI{80}{Oe} in this case. If the value is too large compared to the device coercivity ($H_{\mathrm c}$=\SI{2.2}{kOe} for the measured devices in this paper), it may result in a significant variation in the device performance. To quantitatively evaluate the inter-cell magnetic coupling strength, we defined { \em inter-cell magnetic coupling factor} $\Psi$ as the ratio of the maximum variation in $H_{\mathrm{s\_inter}}^z$ to $H_{\mathrm c}$.
$\Psi$ will be used as an indicator of inter-cell magnetic coupling strength in the rest of this paper.
The $\Psi$ value varies with device size and array pitch, as shown in Fig. \ref{MC_modeling_results2}b. In our simulations, we set the minimum pitch to 1.5$\times$eCD according to \cite{Nguyen2018} for high-density STT-MRAMs and the maximum pitch to \SI{200}{\nano\meter}, which is adopted by both Samsung and Intel \cite{Song2016,Wei2019a}. It can be seen that $\Psi\approx0$\% at pitch=\SI{200}{\nano\meter} for all three device sizes, indicating the inter-cell magnetic coupling is negligible due to the far distance between devices. As the pitch decreases, $\Psi$ increases gradually until reaching a threshold point after which it goes up exponentially. For our devices, $\Psi$=2\% (marked with the dashed line) can be considered as the threshold point, where the array density is maximized with negligible inter-cell magnetic coupling. For a device with eCD=\SI{35}{\nano\meter}, this corresponds to pitch=$\sim$\SI{80}{\nano\meter}.
\begin{figure*}[!h]
\centering
\includegraphics[width=1 \textwidth ]{MC_modeling_results2}
\caption{ (a) $H_{\mathrm{s\_inter}}^{z}$ at the FL of victim C8 under various combinations of the number of 1s in direct neighbors and diagonal neighbors, (b) $\Psi$ vs. pitch with respect to three MTJ sizes, and (c) $I_{\mathrm{c}}$ vs. pitch under the circumstance of different stray fields. }
\label{MC_modeling_results2}
\end{figure*}
\section{Evaluation of Magnetic Coupling Impacts}
\label{sec:MC_impact}
In this section, we evaluate the impact of magnetic coupling on the critical switching current $I_{\mathrm{c}}$ and the average switching time $t_{\mathrm{w}}$, using the proposed model. Thereafter, we investigate the impact on the thermal stability factor $\Delta$ in a similar way. Due to the space limitation, we will only present the simulation results for devices with eCD=\SI{35}{\nano\meter}.
\subsection{Impact on the Critical Switching Current}
\label{subsec:Ic}
Under the influence of stray field, $I_{\mathrm{c}}$ can be expressed as follows \cite{Khvalkovskiy2013d}:
\vspace{-5pt}
\begin{equation}
I_{\mathrm{c}} (H_{\mathrm{stray}}^z) =\frac{1}{\eta } \frac{2\alpha e}{\hbar} M_{\mathrm{s}} \cdot V \cdot H_{\mathrm{k}}\cdot(1\pm\frac{H_{\mathrm{stray}}^z}{H_{\mathrm k}}) , \label{eq:Ic} \\
\end{equation}
where $\eta$ is the STT efficiency, $\alpha$ the magnetic damping constant, $e$ the elementary charge, $\hbar$ the reduced Planck constant, $M_{\mathrm s}$ the saturation magnetization, $V$ the volume of the FL, $H_{\mathrm k}$ the magnetic anisotropy field.
The sign in the parentheses is `$+$' for $I_{\mathrm{c}}$(P$\rightarrow$AP) and `$-$' for $I_{\mathrm{c}}$(AP$\rightarrow$P), given the definition of coordinates in this paper.
In Equation~(\ref{eq:Ic}), $H_{\mathrm{stray}}^{z}=H_{\mathrm{s\_intra}}^{z}+H_{\mathrm{s\_inter}}^{z}$ can be calculated with our proposed magnetic coupling model taking into account both intra-cell and inter-cell stray fields, while $H_{\mathrm{k}}$ needs to be extracted from measurement data. the other parameters in the equation are measured at blanket stage before etch.
Since the switching points (i.e., $H_{\mathrm{sw\_p}}$ and $H_{\mathrm{sw\_n}}$ in Fig.~\ref{fig:Hs_characterization}a) are intrinsically stochastic, we measured the R-H loop of the same device for 1000 cycles to obtain a statistical result of the switching probability at varying fields. With the technique proposed in \cite{Thomas2014}, we are able to extract $H_{\mathrm{k}}$ and $\Delta_0$ by performing curve fitting. $\Delta_0$ is the intrinsic thermal stability factor without any stray field at the FL; it will be used in the next subsection. By doing this for a large number of devices, we obtained $\Delta_0=45.5$ and $H_{\mathrm{k}}=\SI{4646.8}{Oe}$ (both in median) for devices with eCD=\SI{35}{\nano\meter}.
Fig. \ref{MC_modeling_results2}c shows the critical switching current $I_{\mathrm{c}}$ for C8 (for both P$\rightarrow$AP switching and AP$\rightarrow$P switching) at different pitches with respect to various stray fields.
For isolated devices without any stray field (i.e., ideal case, $H_{\mathrm{stray}}^{z}=0$), the intrinsic $I_{\mathrm{c}}$ for the two switching directions is supposed to show no difference; $I_{\mathrm{c}}=\SI{57.2}{\micro\ampere}$. When taking into account the intra-cell stray field (i.e., $H_{\mathrm{stray}}^{z}=H_{\mathrm{s\_intra}}^{z}$), a static shift in $I_{\mathrm{c}}$ is introduced, making
$I_{\mathrm{c}}$(AP$\rightarrow$P)=\SI{61.7}{\micro\ampere} (i.e., 7\% above the intrinsic $I_{\mathrm{c}}$) and $I_{\mathrm{c}}$(P$\rightarrow$AP)=\SI{52.8}{\micro\ampere} (i.e., 7\% below). When considering both intra-cell and inter-cell stray fields (i.e., $H_{\mathrm{stray}}^{z}=H_{\mathrm{s\_intra}}^{z}$+$H_{\mathrm{s\_inter}}^{z}$) for different neighborhood patterns NP\textsubscript{8}, the impact on $I_{\mathrm{c}}$ shows a clear dependence on the array pitch. $I_{\mathrm{c}}$(AP$\rightarrow$P) becomes larger at smaller pitches when NP\textsubscript{8}=0, while it shows an opposite trend when NP\textsubscript{8}=255. This indicates that the variation in $I_{\mathrm{c}}$(AP$\rightarrow$P) between different neighborhood patterns increases as the pitch goes down.
It can be seen that at pitch$\approx$\SI{80}{\nano\meter} (corresponding to $\Psi=2\%$), the variation is marginal. Similar observations can be seen on the P$\rightarrow$AP switching direction.
\subsection{Impact on the Average Switching Time}
\label{subsec:tw}
The average switching time $t_{\mathrm w}$ in the presence of $H_{\mathrm{stray}}^{z}$ in the precessional regime (namely, switched by the STT-effect) can be estimated using Sun's model as follows \cite{Wu2019a}:
\begin{align}
t_{\mathrm w}(H_{\mathrm{stray}}^{z}) &= (\frac {2} {C+\ln(\frac {\pi ^2\Delta }{4})} \cdot \frac{\mu_{\mathrm{B}}P}{e m(1+P^2)}\cdot I_{\mathrm{m}})^{-1}, \\
I_{\mathrm{m}}&=\frac{V_{\mathrm{p}}}{R(V_{\mathrm{p}})}- I_{\mathrm{c}}(H_{\mathrm{stray}}^{z}) . \label{eq:tw_prec}
\end{align}
Here, $C$$\approx$$0.577$ is Euler's constant, $\mu_{\mathrm{B}}$ the Bohr magneton, $P$ the spin polarization, $e$ the elementary
charge, and $m$ the FL magnetization. $V_{\mathrm{p}}$ is the voltage applied on the MTJ device to switch its state. $R({V_{\mathrm{p}}})$ is the resistance of the MTJ device as a function of the applied voltage $V_{\mathrm{p}}$; it shows a non-linear dependence on $V_{\mathrm{p}}$ \cite{Wu2019a}.
Fig. \ref{fig:tw_vs_Vmtj}a-c shows the voltage dependence of the average switching time from AP state to P state ($t_{\mathrm{w}}$(AP$\rightarrow$P)) for MTJs with eCD=\SI{35}{\nano\meter} at pitch=3$\times$eCD, 2$\times$eCD, and 1.5$\times$eCD. Due to the space limitation, the simulation results of $t_{\mathrm{w}}$(P$\rightarrow$AP) are excluded. It can be seen that $t_{\mathrm{w}}$(AP$\rightarrow$P) becomes larger for MTJ devices in the presence of $H_{\mathrm{stray}}^z$ (solid lines), comparing to devices without any stray field (dashed lines). It is worth noting that the larger the voltage, the smaller the impact of the stray field on $t_{\mathrm{w}}$(AP$\rightarrow$P). However, an increase in the switching voltage $V_{\mathrm{p}}$ also results in more power consumption and a higher vulnerability to breakdown. In addition, when the pitch goes from 3$\times$eCD (Fig.~\ref{fig:tw_vs_Vmtj}a) to 2$\times$eCD (Fig.~\ref{fig:tw_vs_Vmtj}b), the inter-cell magnetic coupling factor $\Psi$ increases from 1\% to 2\% and the change in $t_{\mathrm{w}}$(AP$\rightarrow$P) is negligible. However, when the pitch goes down to 1.5$\times$eCD (Fig. \ref{fig:tw_vs_Vmtj}c), $\Psi$ increases to 7\% and the variation in $t_{\mathrm{w}}$(AP$\rightarrow$P) between different NPs (i.e., $H_{\mathrm{s\_inter}}^{z}$) becomes very visible. For example, at a voltage of \SI{0.72}{\volt}, $t_{\mathrm{w}}$(AP$\rightarrow$P) under NP\textsubscript{8}=0 is $\sim$\SI{4}{\nano\second} slower than NP\textsubscript{8}=255, as shown in Fig. \ref{fig:tw_vs_Vmtj}c. This indicates that a larger write margin (e.g., a longer pulse) is required to avoid write failure in the worst-case (i.e., NP\textsubscript{8}=0).
\begin{figure*}[!t]
\centering
\includegraphics[width=1 \textwidth ]{tw_vs_Vmtj}
\caption{Impact of magnetic coupling on the voltage dependence of $t_{\mathrm{w}}$(AP$\rightarrow$P) with eCD=\SI{35}{\nano\meter} at various pitches: (a) 3$\times$eCD, (b) 2$\times$eCD, and (c) 1.5$\times$eCD.}
\label{fig:tw_vs_Vmtj}
\end{figure*}
\begin{figure}[!t]
\centering
\includegraphics[width=0.5 \textwidth ]{Delta_vs_T}
\caption{Impact of magnetic coupling on $\Delta$ with eCD=\SI{35}{\nano\meter} at: (a) pitch= 2$\times$eCD and (b) worst-case $\Delta$ for pitch=3$\times$eCD, 2$\times$eCD, and 1.5$\times$eCD. }
\label{fig:Delta_vs_T}
\end{figure}
\subsection{Impact on the Thermal Stability Factor}
\label{subsec:Delta}
The intrinsic thermal stability factor $\Delta_0$ (without any stray field at the FL) of the MTJ device is given by \cite{Khvalkovskiy2013d}:
$\Delta_0 =\frac{H_{\mathrm{k}}M_{\mathrm{s}}V}{2k_BT}$,
where $k_{\mathrm{B}}$ is the Boltzmann constant and $T$ is the absolute temperature. However, in the presence of stray fields, the thermal stability factor in AP state deviates from that in P state, i.e., $\Delta_{\mathrm{AP}}$$\neq$$\Delta_{\mathrm{P}}$. The $\Delta$ value in the presence of $H_{\mathrm{stray}}^z$ is given by \cite{Khvalkovskiy2013d}:
\begin{equation}
\Delta(H_{\mathrm{stray}}^z) =\Delta_0(1\pm\frac{H_{\mathrm{stray}}^z}{H_{\mathrm{k}}})^2, \label{eq:Delta_Hstray}
\end{equation}
where the sign in the parentheses is `$+$' for $\Delta_{\mathrm{P}}$ and `$-$' for $\Delta_{\mathrm{AP}}$ for the devices considered in this paper. $H_{\mathrm{stray}}^z$ can be calculated with our proposed magnetic coupling model, while $H_{\mathrm{k}}$ and $\Delta_0$ are extracted from measurement data.
Fig. \ref{fig:Delta_vs_T}a shows the thermal stability factor $\Delta$ at varying temperature for eCD=\SI{35}{\nano\meter} and pitch=2$\times$eCD, corresponding to $\Psi=2\%$. It can be seen that the intra-cell stray field $H_{\mathrm{s\_intra}}^{z}$ introduces a static shift in $\Delta_{\mathrm{AP}}$ and $\Delta_{\mathrm{P}}$; $\Delta_{\mathrm{AP}}$ is $\sim$30\% smaller than $\Delta_{\mathrm{P}}$ comparing the dash-dotted line to the dotted one. The solid lines represent the thermal stability factors considering both intra-cell and inter-cell magnetic coupling. It can be seen that the MTJ device has the smallest $\Delta$ (highest vulnerability to a retention fault) when the victim cell is in P state and all neighboring cells are also in P state (i.e., NP\textsubscript{8}=0). Fig. \ref{fig:Delta_vs_T}b compares the worst-case $\Delta$, i.e., $\Delta_{\mathrm{P}}$(NP\textsubscript{8}=0), at pitch=3$\times$eCD, 2$\times$eCD, and 1.5$\times$eCD.
One can observe that $\Delta_{\mathrm{P}}$(NP\textsubscript{8}=0) shows a marginal degradation when the array pitch goes down to 1.5$\times$eCD, in comparison to pitch=2$\times$eCD.
\section{ Conclusion}
\label{sec:conclusion}
Magnetic coupling including both intra- and inter-cell can be a critical design constraint for high-density STT-MRAM designs as the MTJ device scales down. Intra-cell magnetic coupling needs to be minimized by further device stack innovation, as it leads to a significant bifurcation in the switching characteristics and retention for AP$\rightarrow$P and P$\rightarrow$AP switching directions. The inter-cell magnetic coupling depends on the device size as well as the array pitch. When the pitch reaches at $\sim$2 times of device diameter (corresponding to $\Psi=2\%$), the array density is maximized with negligible impact on the device performance. However, more aggressive exploration in array density (e.g., pitch=1.5$\times$eCD) shows significant variations in average switching time, especially at low switching voltages, caused by inter-cell magnetic coupling. Moreover, we also observed a marginal degradation of retention due to the increased inter-cell magnetic coupling.
\section*{Acknowledgment}
We would like to acknowledge and thank the efforts of Sebastien Couet, Farrukh Yasin and Davide Crotti in fabricating the STT-MRAM devices and Nico Jossart for providing the SEM images. This work was performed as part of imec's industrial affiliation program on STT-MRAM devices.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\def\url#1{}
\bibliographystyle{IEEEtran}
|
\section{Introduction}
One of the best way for data protection is to hide real information in fake information. The English term "disinformation" is defined as "false information with the intention to deceive opinion" with slightly negative meaning. However, the idea of disinformation can be borrowed into data science world for private data protection. As more and more data are becoming digitized, efficiency of data transfer, replication and usage has increased significantly in recent years. In addition, due to the development of system like federated data networks, data sharing among large number of institutions and devices becomes possible \cite{jordon2018pate,liu2019two,mukherjee2019protecting,shao2019privacy}. These advances bring convenience to our society. Nevertheless, they also raise big concerns on data security and privacy, especially in sensitive fields like healthcare and personal finance. Despite efforts in data encryption, secure computation and other methods trying to protect data, there is always a chance that data can be leaked due to either technological or human reasons. In addition to technological level protection of data, human level protection is often overlooked \cite{jensen2013challenges,lee2009ethical,joly2016data}. In this study, we aim at developing a technology that provides an additional layer of protection of private data
by mapping the original information onto a fake domain that looks realistic but unidentifiable to human. By doing this, even if the data is leaked, the malicious attackers will not be able to know whether the data they obtained are fake, and, therefore, unable to retrieve the original real information, which is similar to the concept of disinformation.
\par
One of the biggest promises of deep learning is its ability to discover rich representation and approximate complex mapping functions \cite{goodfellow2014generative,bengio2009learning}. Recent advances in generative models such as generative adversarial networks (GANs) utilized these properties of deep learning and significantly simplified some previously difficult tasks such as image-to-image translation and images based text generation \cite{zhu2017unpaired,mirza2014conditional}. Derivatives and modifications of GAN based models have been used and achieved outstanding performance in many fields \cite{odena2017conditional,makhzani2015adversarial,radford2015unsupervised}. These tools can also be used to protect data privacy.
\par
GANs and its derivatives have achieved impressive result in data generation, style transferring and many other fields \cite{goodfellow2014generative,zhu2017unpaired}. The ability of GAN related models to generate data that are indistinguishable from the real data comes from the idea of adversarial loss. Cycle consistency is a concept originated from machine translation, where a phrase translated from one language to another, after translated back, should be identical to the original phrase. Cycle consistency has been widely applied in machine learning especially in computer vision related tasks \cite{godard2017unsupervised,zhou2016learning}. One recent success in combing GANs and cycle consistency was in image style transferring by CycleGAN \cite{zhu2017unpaired}. In this study, we combined GANs and consistency loss to develop a method named FakeSafe to map the private information onto a fake message that looks indistinguishable from the real messages. The fake message can be either from the same domain of the original private information, or from a completely different domain. FakeSafe can be used during data transfer, data storage, data usage or other scenarios in combination of traditional encryption and security technologies. Using toy data sets as well as real world clinical data set, we conducted a proof-of-concept experiments to explore how well FakeSafe can help protect private information at human level and the quality of reconstructed data from fake domain.
\section{FakeSafe}
\subsection{Motivation and formulation}
The purpose of FakeSafe method is to map the original private information onto a fake but realistically looking message. The method consists of two parts: 1) a function $F$ that maps a private message $X$ into a fake message $X^{fake}$ , i.e. $X^{fake}=F(X)$. $X^{fake}$ can be from the same domain as $X$, such as a human face image, or a completely different domain. 2) a reconstruction function $R$ that maps the fake message back to original message. $F$ and $R$ are specific to each data set.
\subsection{System}
In our system of interest, we are assuming there is a sender of private information, a targeted receiver of information. The data transfer/storage infrastructure is not 100\% safe and the malicious attackers might try to steal the private information. Only the data sender has access to function $F$ to map the private data to fake domain. Only targeted receiver has access to function $R$ to retrieve the original data. Even if the attacker obtains $X^{fake}$, without additional information, it is impossible for him or her to know that the data is fake, because $X^{fake}$ looks realistic. If $X$ and $X^{fake}$ are from the same data domain, e.g. mapping a set of human faces into another set of human faces, even if the attacker knows before hand what the data is supposed to look like, it is difficult to notice if the data has been mapped by FakeSafe method into fake messages.
\begin{figure}
\centering
{\setlength{\fboxsep}{0pt}%
\setlength{\fboxrule}{0pt}%
\fbox{\includegraphics[width=8.25cm, height=13.2cm]{Demo1.png}}
%
}%
\label{Demo}
\caption{Human level data protection by mapping original information into fake data domain (disinformation). The original information can be recovered from fake messaged using a trained reconstruction function}
\end{figure}
\begin{figure*}
\centering
{\setlength{\fboxsep}{0pt}%
\setlength{\fboxrule}{0pt}%
\fbox{\includegraphics[width=12cm, height=8cm]{Picture1.jpg}}
%
}%
\label{Mechanism}
\caption{FakeSafe method uses a Generative Adversarial Network (GAN) with cycle consistency to map original private data to fake data (A) and reconstruct original information from the fake message (B) }
\end{figure*}
\subsection{Generative adversarial networks(GAN) with cycle consistency loss}
GAN was used as the function $F$ to map private information to fake message in this study due its good performance in generating fake data sets that look realistic to human. Generative model $F(X)$ is trained against discriminator $D$ to make the outputs of $X^{fake}=F(X)$ look indistinguishable from the samples used to train $D$. We name this type of $X^{fake}$ messages as FakeSafe messages. $D$ and $F$ were trained in an alternating manner. The objective loss function for training generator and discriminator is:
\begin{multline}
L(x,x^{fake},F,D)=\\E[log(D(X^{fake}))]+E[log(1-D(F(X)))]
\end{multline}
$F$ generates data points that look indistinguishable from real data in fake message domain. Least loss was used to train the GAN due to reported stability \cite{zhu2017unpaired}. Therefore, when training GAN, we train $F$ to minimize $E [1-D(F(X))]^2$ and train $D$ to minimize $E [D(F(X))^2]$.
After the other party receives the FakeSafe message, it will be recovered using a trained model $R$ such that $R(F(X))\approx X$. To enable the ability of $R$ to retrieve the original message from $X^{fake}$, cycle-consistency was used to make reconstructed data $R(X^{Fake})$ matching the original data $X$. The loss function is
\[L(F,R,X)=E[||R(F(X))-X||]\]
For reconstruction errors, we used absolute loss. For simplicity, fully connected neural network with leaky ReLU was used in both generator and discriminator models.
\subsection{Model implementation}
As this is a proof-of-concept study, 1) for the image-image generator model, we used a simple 3 layer fully connected neural network with 256, 512 and 1024 units. 2) For the text-image generator model, we used a 4 layer fully connected neural network with 64, 256, 512 and 1024 units. Leaky ReLU was used as the activation functions for hidden layers and batch normalization was applied in both image-image and text-image generator model. 3) For the image-text generator model, we used a 4 layer fully connected neural network with 128, 256, 512 and 1024 units. Leaky ReLU was also used as the activation function, and a dropout with rate 0.2 was introduced to avoid over-fitting. The adam optimization with learning rate of 0.0002 was used in the above 3 cases.
\section{Experiments and results}
In order to understand whether our FakeSafe method works in protecting information transfer, we conducted three types of proof-of-concept experiments. First, we encoded information into fake messages from the same data domain, using MNIST and MNIST fashion as example. Second, we encoded information into fake messages from a different domain, such as MNIST digitals to MNIST fashion. Last, we explored the possibility of multi-step FakeSafe encoding of information. The reconstructing accuracy decreased as we increased the number of steps to encode information.
In addition, we tested its potential values in real world application using a face video frame from a clinical settings using an open source data set.
\begin{figure}
\centering
{\setlength{\fboxsep}{0pt}%
\setlength{\fboxrule}{0pt}%
\fbox{\includegraphics[width=8cm, height=4cm]{SameDomain.jpg}}
%
}%
\label{SameDomain}
\caption{Use FakeSafe method to map private information into fake data from the same domain }
\end{figure}
\subsection{Date set}
To conduct proof-of-concept experiments, four data sets were used in this study:
1) MNIST hand written digits data set, 2) MNIST fashion data set, 3) an English text data set from Tatoeba 4) the UNBC-McMaster Shoulder Pain Expression Archive Data set \cite{lucey2011painful}. UNBC-McMaster Shoulder Pain is a real world data set from clinical setting and consists of human face video frames from different individuals with shoulder pain. We use this data set as an example of real use case of FakeSafe in medical setting.
\subsection{FakeSafe mapping onto the same data domain}
One potential application of FakeSafe is to map private information on other same data domain but different data points. We conducted four experiments: MNIST$\rightarrow$F$\rightarrow$MNIST$\rightarrow$R$\rightarrow$MNIST, Fashion$\rightarrow$F$\rightarrow$Fashion$\rightarrow$R$\rightarrow$Fashion, Face$\rightarrow$F$\rightarrow$Face$\rightarrow$R$\rightarrow$Face and Word$\rightarrow$F$\rightarrow$Word$\rightarrow$R$\rightarrow$Word.
\par
When conducting experiments on MNIST, Models $F,D,R$ were all trained using the training sets with images of 10 hand-written digits. Therefore, $X^{fake}=F(X)$ can be any possible number from the training set and might not have to be the same digits as $X$. As shown in figure \ref{SameDomain} , the recovered images $R(F(X))$ have the same labels as the original images $X$, while the FakeSafe images $F(X)$ are different.
\par
In a similar manner, when conducting experiments on MNIST fashion data set which contain objects from 10 different categories, such as "shoe" or "dress", $R(F(X))$ have the same labels as the original message $X$ and could differ from labels of $X^{Fake}=F(X)$.
\par
When conducting experiments on human face images, original data $X$ is a human face image which was mapped to another human image $X^{Fake}=F(X)$ that could be from the same person or a different person.
When conducting experiments on English words, original data $X$ is a 50-dimension word embeddings which was mapped to another 300-dimension word embeddings $X^{Fake}=F(X)$ that could be from the same word or a different word.
\begin{figure}
\centering
{\setlength{\fboxsep}{0pt}%
\setlength{\fboxrule}{0pt}%
\fbox{\includegraphics[width=8cm, height=8cm]{Cross-domain.jpg}}
%
}%
\label{CrossDomain}
\caption{Use FakeSafe method to map private information into fake data in a different domain. (A)}
\end{figure}
\par
In order to evaluate quality of the reconstructed message$R(F(X))$, two metrics were used. First, reconstruction errors between $R(F(X))$ and $X$ were calculated as mean squared errors. Second, in order to know whether the reconstructed messages $R(F(X))$ still look like from the same class or individual as $X$ to human, we trained a classifier $C$ on $X$ in the training set to classify their labels, i.e. the digits, fashion category or individual ID, and apply $C$ onto reconstructed data $R(F(X))$. The accuracy, F1 score, precision and recall of $C(R(F(X)))$ were compared with the original labels of $X$.
\begin{table*}
\caption{Performance of FakeSafe mapping onto the same domain or different domain}
\label{Performance}
\resizebox{17.5cm}{!}{%
\begin{tabular}{|c|c|c|c|c|c|c|c|c|}
\hline
\textbf{Experiment} &
\textbf{Original message} &
\textbf{FakeSafe message} &
\textbf{Reconstruction error} &
\textbf{Precision} &
\textbf{Recall} &
\textbf{F1 Score} &
\textbf{Type} &
\textbf{Remark} \\
\hline
\textbf{Face-\textgreater{}F-\textgreater{}Face-\textgreater{}R-\textgreater{}Face} & Face image & Face image & 1.06 & 0.95 & 0.92 & 0.92 & Same domain&- \\ \hline
\textbf{MNIST-\textgreater{}F-\textgreater{}MNIST-\textgreater{}R-\textgreater{}MNIST} & MNIST digits & MNIST digits & 1.62 & 0.9 & 0.8 & 0.81 & Same domain&- \\ \hline
\textbf{Fashion-\textgreater{}F-\textgreater{}Fashion-\textgreater{}R-\textgreater{}Fashion} & Fashion image & Fashion image & 1.2 & 0.71 & 0.72 & 0.7 & Same domain &- \\ \hline
\textbf{Word-\textgreater{}F-\textgreater{}Word-\textgreater{}R-\textgreater{}Word} & English words (50d embeddings) & English words (300d embeddings) & NA & 0.65 & 0.68 & 0.66 & Same domain & All 202 words \\ \hline
\textbf{Face-\textgreater{}F-\textgreater{}MNIST-\textgreater{}R-\textgreater{}Face} & Face image & MNIST digits & 0.1 & 1 & 1 & 1 & Cross domain &- \\ \hline
\textbf{Fashion-\textgreater{}F-\textgreater{}MNIST-\textgreater{}R-\textgreater{}Fashion} & Fashion image & MNIST digits & 0.79 & 0.85 & 0.76 & 0.79 & Cross domain &- \\ \hline
\textbf{{MNIST-\textgreater{}F-\textgreater{}Fashion-\textgreater{}R-\textgreater{}MNIST}} & MNIST digits & Fashion image & 1.32 & 0.91 & 0.88 & 0.88& Cross domain &- \\ \hline
\textbf{Word-\textgreater{}F-\textgreater{}Fashion-\textgreater{}R-\textgreater{}Word} & English words (tokens) & Fashion image & NA & 0.8 & 0.84 & 0.81 & Cross domain & Top 100 frequent words \\ \hline
\textbf{Word-\textgreater{}F-\textgreater{}Fashion-\textgreater{}R-\textgreater{}Word} & English words (50d embeddings) & Fashion image & NA & 0.96 & 0.96 & 0.96 & Cross domain & Top 100 frequent words \\ \hline
\textbf{Word-\textgreater{}F-\textgreater{}Fashion-\textgreater{}R-\textgreater{}Word} & English words (50d embeddings) & Fashion image & NA & 0.65 & 0.68 & 0.66 & Cross domain & All 202 words \\ \hline
\end{tabular}%
}
\end{table*}
\par
The MNIST$\rightarrow$F$\rightarrow$MNIST$\rightarrow$R$\rightarrow$MNIST FakeSafe experiment achieved a reconstruction error of 1.62, classifier precision of 0.90, recall of 0.80 and F1 score of 0.81. In the Fashion$\rightarrow$F$\rightarrow$Fashion$\rightarrow$R$\rightarrow$Fashion experiment (Table \ref{Performance} ), FakeSafe achieved a reconstruction error of 1.2, classifier precision of 0.71, recall of 0.72 and F1 score of 0.70. The real world human face image data set, FakeSafe method achieved a reconstruction error of 1.06, precision of 0.95, recall of 0.92 and F1 score of 0.92.
\subsection{FakeSafe mapping onto a different data domain}
Hide privacy information onto fake messages of the same type can help protect the information by misleading the malicious attackers. However, sometimes it is better not to expose the original information domain at all. Therefore, we conducted experiments to FakeSafe map information into the message in a different domain. We conducted 4 experiments, MNIST$\rightarrow$F$\rightarrow$Fashion$\rightarrow$R$\rightarrow$MNIST, Fashion$\rightarrow$F$\rightarrow$MNIST$\rightarrow$R$\rightarrow$Fashion, Face$\rightarrow$F$\rightarrow$MNIST$\rightarrow$R$\rightarrow$Face and Word$\rightarrow$F$\rightarrow$Fashion$\rightarrow$R$\rightarrow$Word. The performances are comparable to FakeSafe mapping onto the same data domain (Table \ref{Performance}).
\par
Specifically, for the experiment Word$\rightarrow$F$\rightarrow$Fashion$\rightarrow$R$\rightarrow$Word, we have tried two different approaches to map the original messages. In the first approach, we will tokenize the original messages, which are the English words, and then map the tokens to MNIST fashion images using FakeSafe. During the decoding process, we will map the MNIST fashion images back to tokens, which will be eventually converted back to English words. In the second approach, we will first convert the words to word embeddings with 50 dimensions, using GloVe Word Embeddings, and then map the word embeddings to MNIST fashion images. During the decoding process, we will use FakeSafe to map the MNIST fashion images back to word embeddings, and then decode back to the original words by finding the word with the smallest cosine similarity with the decoded word embeddings. It is noteworthy that, the second approach, which uses word embeddings as the original messages, is proved to achieve better performance than the first approach, which only uses word tokens as the original messages.
\begin{figure*}
\centering
{\setlength{\fboxsep}{0pt}%
\setlength{\fboxrule}{0pt}%
\fbox{\includegraphics[width=16cm, height=4cm]{2-Step.jpg}}
%
}%
\label{Two-step}
\caption{Multi-step FakeSafe mapping. The reconstruction errors increase with depth}
\end{figure*}
\begin{table}
\caption{Performance of multi-step FakeSafe mapping)}
\label{Multi-step-Performance}
\resizebox{17.5cm}{!}{%
\begin{tabular}{|c|c|c|c|c|c|}
\hline
\textbf{Experiment} &
\textbf{Reconstruction error} &
\textbf{Precision} &
\textbf{Recall} &
\textbf{F1 Score} &
\textbf{Type} \\ \hline
\textbf{mnist-\textgreater{}mnist-\textgreater{}fashion-\textgreater{}mnist-\textgreater{}mnist} &
9.4 &
0.75 &
0.64 &
0.63 &
two-step \\ \hline
\textbf{fashion-\textgreater{}mnist-\textgreater{}mnist-\textgreater{}mnist-\textgreater{}fashion} &
1.1 &
0.81 &
0.72 &
0.73 &
two-step \\ \hline
\textbf{face-\textgreater{}fashion-\textgreater{}mnist-\textgreater{}fashion-\textgreater{}face} &
37.1 &
0.66 &
0.56 &
0.59 &
two-step \\ \hline
\textbf{mnist-\textgreater{}mnist-\textgreater{}fashion-\textgreater{}fashion-\textgreater{}fashion-\textgreater{}mnist-\textgreater{}mnist} &
17.3 &
0.48 &
0.44 &
0.39 &
three-step \\ \hline
\textbf{face-\textgreater{}mnist-\textgreater{}fashion-\textgreater{}fashion-\textgreater{}fashion-\textgreater{}mnist-\textgreater{}face} &
88.8 &
0.29 &
0.36 &
0.29 &
three-step \\ \hline
\end{tabular}
}
\end{table}
\section{Deeper FakeSafe mapping}
To guarantee the safety of sensitive data, one may ask why not map the original private multiple times using a cascade of different $F$ functions, so that even if the attacker knows the message is fake, he or she will not know how many steps the messages were mapped. In order to explore the feasibility of deeper FakeSafe mapping, we conducted a series of experiments of 2-step and 3-step FakeSafe using MNIST, fashion and face images (Figure \ref{Two-step} and Table \ref{Multi-step-Performance}). Our results suggest that even it is possible to conduct multi-step FakeSafe mapping, the reconstruction error increased and classification accuracy decreased dramatically.
\clearpage
\section{Conclusion}
In this article, we propose a method, named FakeSafe, to provide human-level private data protection by mapping each data point into a fake message that looks realistic to human. We utilized GANs with cycle-consistency to build a function to map the original data to fake message and another function to map the fake message back to the original data. Both functions are data set specific and can be easily adjusted for the other data sets. FakeSafe method gives users flexibility to map private data onto different data domains depending on use cases. In addition, FakeSafe can be easily used in combination with traditional data protection technologies but focus on human-level protection which takes human factors in data security and privacy into consideration.
\bibliographystyle{aaai} |
\section*{Abstract}
\justifying
We report a straightforward method to control main spatio-temporal couplings in a CPA laser chain system using a specially designed chromatic doublet in a divergent beam configuration. The centering of the doublet allows for the control of the spatial chirp of the CPA laser chain, while its longitudinal position in the divergent beam enables the control of the amount of longitudinal chromatism in a wide dynamic range. The performance of this technique is evaluated by measuring main spatio-temporal couplings with a simple method, based on an ultrafast pulse shaper, which allows for a selection of narrow windows of the spectrum.
\section{Introduction}
Chirped pulse amplification (CPA) laser chains are being more and more used as they provide applications in a wide range of domains, thanks to their ultrashort and intense pulses\cite{Nolte:16}. The beam properties in the spatial domain can be controlled thanks to deformable mirrors or spatial light modulators \cite{Weiner:00}, while acousto-optic programmable dispersive filters allows for a temporal and spectral shaping of ultrashort beam\cite{Verluise:00}. However, the control of spatio-temporal couplings (STCs), described in \cite{Akturk:10}, still represents a challenge in the shaping of ultrashort pulses.
Indeed, high energy CPA laser systems require large beam expanders, which can introduce couplings, better known as pulse front curvature in the near field and longitudinal chromatism in the far field, deteriorating the spatio-temporal focusing performance\cite{Bor:89}. A precise and dynamic adjustment of the STCs would also pave the way for new advances. Indeed, numerical calculations show that the combination between longitudinal chromatism and temporal chirp allows for the control of the velocity of ultrashort pulses under vacuum\cite{Sainte-Marie:17}, \cite{Froula:18}, therefore offering perspectives for laser-driven acceleration of ions\cite{Macchi:13}, or to avoid velocity mismatch effects in experiments involving the propagation of multiple pulses of different frequencies in a dispersive medium\cite{Depresseux:15}. Adjustment of longitudinal chromatism, along with spatial-phase shaping, also opens up prospects for phase-locked laser wakefield acceleration of relativistic electrons\cite{Caizergues:20}, \cite{Froula:20}.
Longitudinal chromatism due to lenses can be entirely avoided by replacing them by reflective optics, like spherical or parabolic mirrors. Yet, the former often introduce spherical aberration and even astigmatism, when the latter are expensive and more difficult to implement. Several compensation schemes have been proposed using reflective, refractive, or diffractive optical components\cite{Madjidi-Zolbanine:79}, \cite{Piestun:01}, \cite{Bahk:14}, \cite{Neauport:07}. All these methods have their own advantages and drawbacks. However, they have in common the inability to adjust continuously the longitudinal chromatism.
A dynamic chromatic aberration pre-compensation scheme for ultrashort laser chains has been suggested\cite{Cui:19}. It consists in a system, requiring vacuum to avoid air breakdown, composed of two lenses and a spherical mirror, each being on a translation stage. Although being efficient and having a good dynamic, this system remains expensive, complex and bulky.
Another STC is inherent to ultrashort laser chains: pulse compressors require the introduction of massive amounts of angular dispersion, i.e. angular variation of the average wavelength in the beam\cite{Gu:04}, which is equivalent to pulse front tilt in the Fourier space. Focusing a beam containing angular dispersion/pulse front tilt generates spatial chirp, i.e. variation of the focus transverse position with respect to the wavelength. Although supposed to be completely removed with a perfect alignment, small misalignments as well as optics imperfections, are likely to cause non-desired angular dispersion, once again affecting the spatio-temporal focusing performance of the beam. On the other hand, some applications, like femtosecond pulse shaping\cite{Weiner:88} or control of the emission angle of laser-plasma accelerated electrons\cite{Popp:10}, require to separate spatially the different wavelengths deliberately and in a controlled manner. This can easily be done in the dispersion direction of the compressor gratings by changing their parallelism. More generally, some solutions have been suggested, like placing a lens one focal length away from a grating\cite{Weiner:88}. But this proposition is only suitable to introduce a precise amount of spatial chirp, depending on the properties of the chosen grating.
In this letter, we report a straightforward method to control STCs using a chromatic doublet, inserted in a divergent beam configuration. This technique is of high interest since it allows to adjust both spatial chirp thanks to the centering of the doublet, and longitudinal chromatism in a wide dynamic range. The doublet design can easily be adapted to all types of applications, from full correction of the STCs to a precise control of their amount.
\section{Principle and design of the doublet}
The control of the STCs (longitudinal chromatism and spatial chirp) is ensured by the properties of the doublet, and its longitudinal and transverse position in the beam. The principle on which this technique is based is as follows, and illustrated on figure~\ref{fig:Principle}.
\begin{figure}[htbp]
\centering
\includegraphics[width=12cm]{Principle.pdf}
\captionsetup{justification=justified}
\caption{Principle of the experiment. Rays plots for : a) a doublet designed to compensate exactly the longitudinal chromatism of the beam. b) and c) a doublet designed to introduce longitudinal chromatism for the same beam divergence but for a different longitudinal position of the doublet in this beam. d) a chromatic doublet off-center with respect to the beam. In d), the input beam is collimated to ease comprehension. A paraxial lens is added in each case to visualize the impact of the doublet.}
\label{fig:Principle}
\end{figure}
For decades, doublets have been used to reduce chromatic aberrations. They consist of two lenses, one convergent and one divergent, made of two different glasses which disperse wavelengths differently. Their properties are chosen to correct exactly the longitudinal chromatism. Therefore, there is no spreading of the focusing point along the beam propagation direction. Such an achromat is displayed on figure~\ref{fig:Principle} a) for a divergent beam. The represented doublet has an infinite focal length.
On the contrary, doublets can be used to introduce a certain amount of longitudinal chromatism, characterized by the spectral focal shift (SFS), i.e. the distance between the respective focus of the considered wavelengths. The amount of longitudinal chromatism cannot be adjusted for a given doublet used in a collinear beam. However, this is not the case in a divergent beam as it can be seen on figure~\ref{fig:Principle} b) and c). They represent ray plots for an input beam with the same divergence, but for a different longitudinal position of the doublet in this beam. The focal planes of the different wavelengths are much closer to each other in b) than in c). Thus, using a doublet in a divergent beam allows for the adjustment of longitudinal chromatism.
As shown on figure~\ref{fig:Principle} d), a decentering of the doublet induces angular dispersion and therefore a variation of the focus transverse position with respect to the wavelength. The more off-center the beam, the more the wavelengths are separated from each other in the transverse plane. Therefore, varying on the transverse (vertical and lateral) positions of the doublet, the amount of spatial chirp can be adjusted within the plane perpendicular to the propagation axis, and not only along one axis.
The doublet used for this experiment was specially designed for our laser chain using the software Zemax. The doublet is composed of a convergent lens in BK7 and of a divergent lens in SF5. The doublet is vacuum-spaced. The B integral of the doublet is estimated to be 3 mrad. It has an infinite focal length at 800 nm and is suited to a beam with a divergence of $\theta$ = 1.75°. More details concerning its design are given in the supplementary materials. It is conceived to introduce a variation of a few fs/cm$^{2}$ when translating the doublet longitudinally over a range of 40 cm. The doublet is optimized to minimize other aberrations. More details about the residual aberrations are provided in the supplementary materials. Moreover, moving the doublet by 40 cm along the optical axis only introduce a defocus of a few hundred of microns. Thus, the adjustment of the longitudinal chromatism can be performed without modifying the position of the focal plane.
The experiment was carried out using a 5-stages amplification Ti:Sa laser chain delivering 30fs-duration pulses with an energy up to 5J at a repetition rate of 1Hz. The spectrum of the laser is centered around 800 nm, with a spectral width of 150 nm at 1/$e^{2}$. The doublet is inserted in a divergent beam configuration, before the last lens of the last afocal system of the chain. The doublet is placed on a 40 cm translation mount in the longitudinal direction. Its axial position will therefore be expressed by the distance d between itself and the last lens of the last afocal system. The doublet is also motorized in vertical and lateral directions, to ease its centering and thus the adjustment of the spatial chirp. The measurement of longitudinal chromatism has been performed with a 60 mm beam diameter at the output of the last afocal system, while the measurement of spatial chirp has been performed with a 30 mm beam diameter.
\section{Measurement and control of longitudinal chromatism}
\begin{figure}[htbp]
\centering
\includegraphics[width=\linewidth]{Setup.pdf}
\captionsetup{justification=justified}
\caption{Setup of the experiment. An ultrafast pulse shaper, installed upstream the amplification stages, selects a narrow window of the initial spectrum. The longitudinal focal plane, for a given spectral component, is located precisely with a CCD camera scanning through the focus of a lens of 2 m focal length. The doublet is placed on a 40 cm translation mount in the longitudinal direction and is also motorized in vertical and lateral directions. The measurements are performed for different distances between the doublet and the last lens of the afocal system. The whole laser chain is not represented, except its last afocal system.}
\label{fig:Setup}
\end{figure}
\begin{figure*}[ht]
\centering
\includegraphics[width=\linewidth]{Chromatism.pdf}
\captionsetup{justification=justified}
\caption{Peak intensity of the beam with respect to the longitudinal position of the CCD camera for 775 nm in blue, 800 nm in green, 825 nm in red, a) for d = 12 cm, b) for d = 30 cm, c) for d = 50 cm.}
\label{fig:Scan}
\end{figure*}
The performance of our dynamic correction using the specially designed doublet was evaluated thanks to the following method, performed at reduced power to prevent damaging the amplifiers. An acousto-optic programmable dispersive filter (FastLite Dazzler ultrafast pulse shaper) installed upstream the amplification stages, is used to select a narrow window of the spectrum of width $\Delta\lambda$ = 10 nm. In practice, the incident beam propagates through a birefringent crystal, mounted on a piezoelectric transducer which creates a longitudinal traveling acoustic wave. The diffraction efficiency for each phase-matched optical spectral component is adjusted by modulating the amplitude of the traveling acoustic wave. By modulating its frequency, the longitudinal position where the diffraction occurs in the crystal for each spectral component can be tuned. Thus, controlling the spectral phase and the amplitude of the incident beam, certain components can be attenuated, resulting in the selection of a narrow window of the spectrum. The beam is then focused by a lens of 2 m focal length. With a CCD camera scanning through the Rayleigh zone of this lens, the longitudinal position of the focal plane for a given spectral component can be precisely located. For this measurement, three narrow windows, respectively centered on 775, 800 and 825 nm, were used. The setup of the experiment is shown on figure~\ref{fig:Setup}. The whole laser chain is not represented, except its last afocal system. The measurements are performed for different values of the distance d between the doublet and the last lens of the last afocal system.
The results of these measurements are displayed on figure~\ref{fig:Scan} for several values of d : a) 12 cm, b) 30 cm, c) 50 cm. It shows the peak intensity of the beam with respect to the longitudinal position of the CCD camera for 775 nm in blue, 800 nm in green and 825 nm in red. The measurement points are represented by squares, with error bars calculated as the standard error determined by repeatability over 5 shots. These points are fitted by Gaussians, represented in solid lines.
The performance of the doublet is evaluated through measurement of the shift of the focal plane for different wavelengths, called spectral focal shift (SFS). The obtained values are listed in the table~\ref{Table:Values}. The uncertainties are evaluated considering the standard error of the mean value of each fit.
Therefore, our specially designed doublet allows for the control of the longitudinal chromatism in a wide dynamic range. In our configuration, a 40 cm range on the position of the doublet induces a linear variation of the SFS between 775 and 825 nm from 1.81 mm to 5.33 mm, which corresponds to a variation of pulse front curvature from $\alpha$ = 1.26 fs/cm$^{2}$ to 3.70 fs/cm$^{2}$. This range can be settled with a different design of the doublet, and thus be adjusted to all kind of applications, from a complete correction of the longitudinal chromatism for a chromatic laser chain to a precise control of the amount of longitudinal chromatism, wether in positive or in negative.
\begin{table}
\begin{center}
\begin{tabular}{|c||c|c|c|}
\hline
& SFS 775-800 nm & SFS 775-825 nm \\
\hline
d = 12 cm & (1.12 $\pm$ 0.18) mm & (1.81 $\pm$ 0.19) mm \\
d = 30 cm & (1.80 $\pm$ 0.19) mm & (3.76 $\pm$ 0.25) mm \\
d = 50 cm & (2.70 $\pm$ 0.11) mm & (5.33 $\pm$ 0.10) mm \\
\hline
\end{tabular}
\end{center}
\captionsetup{justification=justified}
\caption{Measured values of the spectral focal shift between 775 and 800 nm, and of the spectral focal shift between 775 and 825 nm for different values of d.}
\label{Table:Values}
\end{table}
\section{Spatial chirp control}
\begin{figure}[htbp]
\centering
\includegraphics[width=7cm]{ChirpAllSpectrum.pdf}
\captionsetup{justification=justified}
\caption{Peak intensity at focus with respect to the transverse position of the doublet. Purple stands for the horizontal direction, orange for the vertical direction. At the top, spatial profiles of the beam for different vertical doublet positions.}
\label{fig:ChirpAllSpectrum}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=6cm]{Chirp2wavelengths.pdf}
\captionsetup{justification=justified}
\caption{Spatial profiles of two-colour laser beam pulse at the focus: a) for a perfect centering of the doublet, b) for a perfect vertical centering and a lateral decentering of 13 mm, c) for a perfect lateral centering and a vertical decentering of 15 mm, and d) for a vertical decentering of 7 mm and a lateral decentering of 13 mm. The beam radius at 1/e² on a) is of the order of 200 $\mu$m.}
\label{fig:Chirp2wavelengths}
\end{figure}
The use of this doublet also enables the correction of another STC. To visualize the impact of the centering of the doublet on the spatial chirp, we measure the peak intensity of the beam at focus with respect to the transverse position of the doublet. It is represented on figure~\ref{fig:ChirpAllSpectrum}, in purple for the lateral direction and in orange for the vertical direction. This figure enhances a maximum peak intensity for a perfect centering of the doublet, marked by position 0 in both directions. For each direction, there is a clear decrease of intensity when the doublet is not perfectly centered. Indeed, as it has been shown on figure~\ref{fig:Principle}, a decentering of the doublet introduces spatial chirp which results in a separation of the different frequency components at focus in the transverse directions. This can easily be visualized at the top of figure~\ref{fig:ChirpAllSpectrum}, showing spatial profiles of the beam at focus for different vertical doublet positions. For a perfect centering, the profile is very intense and circular, while it stretches in the direction of the decentering due to the separation of wavelengths.
Another way to display the spatial chirp consists in using again the ultrafast pulse shaper to adjust the spectrum so that it remains only two peaks, in this case at 775 and 825 nm respectively, each with a width $\Delta\lambda$ = 10 nm. A similar method is used in \cite{Borzsonyi:12} to measure spatial dispersion by filtering spectrally the beam to create well separated peaks in the spectrum. Figure~\ref{fig:Chirp2wavelengths} shows spatial profiles of the beam at the focus for different decentering of the doublet : a) was obtained for a perfect centering of the doublet, b) for a perfect vertical centering and a lateral decentering of 13 mm, c) for a perfect lateral centering and a vertical decentering of 15 mm, and d) a vertical decentering of 7 mm and a lateral decentering of 13 mm. The maximum decentering shown here, 15 mm in the vertical direction, causes an spatial chirp of 3.2 $\mu$rad/nm. For comparison, the spatial chirp required in \cite{Popp:10} was 1$\mu$rad/nm to deflect the electron beam by 4 mrad. Such a decentering of the doublet introduces small aberrations : simulations performed with Zemax showed that a decentering of 13 mm in lateral and 7 mm in vertical introduces 0.05 $\lambda$ of 45° astigmatism. As this value does not vary with the wavelength, it can easily be corrected with a deformable mirror.
Varying the vertical and lateral position of the doublet, the position of one wavelength relatively to the other can easily be selected, and thus the amount of spatial chirp can be adjusted. This should be of high interest for pulse shaping experiments as the accuracy of pulse shaping depends on the degree of spatial chirp at the focal plane.
\section{Summary}
In summary, we have reported a straightforward method to evaluate the longitudinal chromatism of a CPA laser chain, using an ultrafast pulse shaper to locate precisely the focal plane of different spectral components. This technique is easy-to-implement since it only requires an acousto-optic dispersive filter, available in most laser facilities, and a focusing optic. This process allowed us to assess the performance of a chromatic doublet regarding the control of STCs. The longitudinal position of the doublet, inserted in a divergent beam configuration, enables the regulation of the amount of longitudinal chromatism in a wide dynamic range. This technique can easily be adapted to most laser chains by adjusting the design of the doublet. Furthermore, the centering of the doublet allows for the control of the degree of spatial chirp at focus. It opens up prospects for numerous applications requiring the adjustment of STCs.
\section*{Acknowledgments}
This work has received funding from the European Union’s Horizon 2020 Research and Innovation programme under Grant No. 730871 (project ARIES).
\bigskip
|
\section{Introduction}
Evaporation-induced self-assembly (EISA) methods are useful in applications such as inkjet printing, production of photonic crystals, application of transparent conductive coatings, development of biosensors, and others~\cite{Kolegov2020}. Modeling the formation of colloidal particle precipitation structures is important. It allows us to understand the main mechanisms of such processes, identify key parameters, and study how to influence the system to obtain the required patterns. For example, the model~\cite{ZhaoM2019} describes the self-assembly of particles on the free surface of a droplet. The capillary interaction of particles and their transport by surface flow resulting from evaporation is considered in this paper. The authors~\cite{ZhaoM2019} performed calculations for spherical and triangle droplets using the lattice Boltzmann method (LBM). Numerical results showed the dependence of the sediment on the curvature of the free surface. Clusters of particles formed at the liquid-vapor interface appear on the substrate after complete evaporation of the liquid. Another mechanism for the formation of particle clusters as a result of their capillary attraction is their self-assembly on the substrate where the thickness of the liquid layer is less than the particle size~\cite{Wouters2020}. The calculations were performed using the modification of LBM. The feature of the algorithm is the ability to simulate the dynamics of soft particles~\cite{Wouters2020}. The mass transfer of hard particles in a drying film was studied in~\cite{Chun2020}. The LBM-based model takes into account the hydrodynamic interaction of particles and their diffusion. The authors~\cite{Chun2020} performed calculations for different values of the Peclet number associated with different evaporation rates. At high values of the Peclet number, a layer of particles is formed near the free surface of the film due to the rapid movement of the two-phase boundary as a result of intense evaporation. A uniform distribution of particles is observed at a small value of the Peclet number. It is explained by diffusion transfer. The periodic predominance of capillary flow or solutal Marangoni flow sometimes leads to the formation of concentric precipitation of colloidal particles~\cite{Seo2020}. The authors~\cite{Seo2020} used the lattice model and the Monte Carlo method to study the effect of the surfactant concentration on this process. The magnetic interaction of particles in a liquid and the formation of chains was studied in~\cite{Darras2017}. The model is based on the Soft Sphere Discrete Element Method. At each time step, Newton's equations were solved to predict the motion of particles. Calculations were performed for small viscosity values~\cite{Darras2017}. The model~\cite{Yang2020} takes into account the diffusion and sedimentation of particles. These particles have no volume, but they have mass. The model is based on the phase field method and the Monte Carlo method. Two shapes of the droplet surface were considered: a spherical segment and an asymmetric shape from the experiment~\cite{Deegan1997}. In both cases, a large number of deposited particles per unit area near the periphery is shown (the coffee stain effect). The model predicts a significant accumulation of particles in areas of low curvature for an asymmetric droplet shape. In the experiment~\cite{Deegan1997}, the particles accumulated mainly in areas of high curvature. The reason for this is the capillary flow, which is not taken into account in the model~\cite{Yang2020}. The formation of particle chains (clusters) inside the annular sediment during the evaporation of a colloidal droplet on a substrate was described in~\cite{Kolegov2019}. The model takes into account diffusion, advection, and capillary attraction of particles. Capillary interaction of particles occurs near the boundary (fixation radius $R_f$) where the thickness of the liquid layer is comparable to the particle size. This process was mimicked as follows. In one time step, the particle was shifted to the nearest neighboring particle in the small neighborhood $R_n$ at the presence of the last one. Calculations were performed for only one parameter value $R_n$. In the current work, we consider the set of parameter $R_n$ values and determine the value which leads the average cluster size is similar to the experiment~\cite{Park2006}. Also, here we analyze the modification of the algorithm associated with multiple attempts to shift those moving particles whose location has not changed in a time step as a result of a collision with a neighboring particle.
\section{Methods}
\subsection{Problem statement and model description}
Here we consider an axisymmetric sessile droplet on a substrate in the mode of a pinned three-phase boundary. Colloidal particles are suspended in the liquid. They are subject to transfer by flow, diffusion, and capillary attraction. At the initial time, the contact angle $\theta_0=\pi/18$. The contact radius of the droplet with the substrate $R=$ 60~$\mu$m. The radius of spherical particles $r_p=$ 0.35~$\mu$m. As the liquid dries, these particles precipitate. The deposit is a monolayer of particles since the droplet is thin (its height $h\ll R$). Time of full droplet evaporation $t_\mathrm{max}=$ 10~s. A more detailed description of the problem statement is given in~\cite{Kolegov2019}.
The mathematical model~\cite{Kolegov2019} is semi-discrete. Each particle is considered separately, but the liquid is described with a continuum approach. It is assumed that the particle velocity is equal to the fluid flow velocity, which is calculated by the formula
$$
\bar v_r (r,t)=
\frac{R}{4 r/R (t_\mathrm{max}-t)}\left[ \frac{1}{\sqrt{1-(r/R)^2}} - \left( 1-(r/R)^2 \right) \right].
$$
It was derived from the mass conservation law (references and formula derivation are given in~\cite{Kolegov2019}). The model considers the radial flow averaged over the drop height. Here $r$ and $t$ is the radial coordinate of the particle center and the current time of the process, respectively.
Coordinate of the particle at the next time step as a result of displacement by the flow
\begin{equation*}
r_{\tau+1} =
\begin{cases}
r_\tau + \bar v_r \delta t, \; r_{\tau+1}\eqslantless R_f \\
R_f,\text{ otherwise,}
\end{cases}
\end{equation*}
where $\delta t$ is the time step and $\tau$ is the time step number.
Capillary forces act on the particles in the region of the fixation boundary $R_f (t) = \sqrt{R^2 - 4 r_p R/ \theta (t)}$. The time dependence of the contact angle is written as
$\theta(t) = \theta_0 \left( 1 - t/ t_\mathrm{max} \right)$. We mimic the capillary attraction of particles as follows. If there are particles in the vicinity of the current particle that have precipitated or are also subject to capillary attraction, then it moves close to the nearest neighboring particle. By neighborhood, we mean a circle with a radius $R_n$. Some particles precipitate if $r > R_f+r_p$. In this case, the local thickness of the liquid layer is less than the particle size, $h < 2r_p$.
The diffusion motion of particles is modeled by the Monte Carlo method. A random offset angle $\alpha \in [-\pi; \pi)$ is generated at each time step $\tau$. The new position of the particle is calculated using the formulas $x_{\tau +1}= x_\tau + \cos (\alpha)\sqrt{2
D \delta t}$ and $y_{\tau +1}= y_\tau + \sin (\alpha)\sqrt{2
D \delta t}$, where $x$ and $y$ are the Cartesian coordinates of the particle center. The diffusion coefficient is calculated using the Einstein formula $D = kT/(6\pi \eta r_p)$, where $k$ is the Boltzmann constant. The values of the temperature $T$ and the viscosity $\eta$ of the liquid are taken for water under normal room conditions ($D\approx 6\cdot 10^{-13}$ m$^2$/s). The value of the time step $\delta t =$ 0.1~ms was selected based on a series of computational experiments in such a way that the Einstein relation for the mean square displacement $\langle l^2 \rangle = 2D t_{max}$ was fulfilled. Here $$\langle l^2 \rangle =\frac{1}{N_p} \sum_{i=1}^{N_p}\left( \left( x_i(t_\mathrm{max}) - x_i(0) \right)^2 + \left( y_i(t_\mathrm{max}) - y_i(0) \right )^2 \right)$$
and $N_p$ is the number of particles.
\subsection{Algorithm}
For each particle, we store coordinates and status in memory. Let's denote movable particles with green and black colors. Here green particles are subject to convective and diffusive transfer. We do not use a separate status for particles diffusing near $R_f$ to simplify the algorithm~\cite{Kolegov2019}. Capillary forces act on black particles. Unmovable particles that have precipitated are marked in red. The particle status changes according to the rules in Figure~\ref{fig:algorithmDescription} (right).
\begin{figure}[H]
\begin{minipage}{0.3\textwidth}
\includegraphics[width=0.99\textwidth]{schemeToTheAlgorithm}
\end{minipage}\hspace{0.025\textwidth}
\begin{minipage}{0.3\textwidth}
\includegraphics[width=0.99\textwidth]{particleStatus}
\end{minipage}
\begin{minipage}{0.35\textwidth}
\caption{\label{fig:algorithmDescription} Scheme to the algorithm description (left) and rules for changing the particle status (right).}
\end{minipage}
\end{figure}
For more details about the algorithm, see Figure~\ref{fig:algorithmDescription} (left) and pseudocode~\ref{alg:particleDynamicsModified}. At the beginning of the process, all particles are green by default. If a green particle touches the fixation radius, it is repainted black. The green particle cannot go beyond the fixation radius, unlike the black one. After all, the movement in that area is mainly due to capillary forces. But due to the movement of the fixation radius itself, the green particle may be behind it (after that, it will be repainted red). A black particle becomes red after it is beyond the moving boundary $R_f$ as a result of the motion of the last one or after it shifts due to capillary forces.
\begin{algorithm}[H
\caption{Particle dynamics algorithm.}
\label{alg:particleDynamicsModified}
\begin{algorithmic}[1]
\State Problem parameters definition: $R_n$, $r_p$, $R$, $N_p$, and
$t_\mathrm{max}$.
\State Generation random coordinates of the particles
$x_i$ and $y_i$ ($i \in [1;N_p]$).
\State By default, all particles are marked green.
\For {$\tau \leftarrow 1, t_\mathrm{max}/\delta t$}
\State calculate $R_f$.
\For {$i \leftarrow 1, N_p$}
\State changing the particle status if necessary.
\EndFor
\While {(not all particles are displaced) $\And$ (there is a shift of at least one particle)}
\State shuffle the array of particle numbers.
\For {$i \leftarrow 1, N_p$}
\State red particles are skipped.
\If {(green particle)}
\State calculate the new particle coordinates due to diffusion.
\If {(no collision)}
\State move the particle.
\EndIf
\State Calculate the new particle coordinates due to advection.
\If {(no collision)}
\State move the particle.
\EndIf
\EndIf
\If {(the particle is black) $\And$ (there is a black or red particle within $R_n$ )}
\State Calculate the new particle coordinates due to the capillary shift.
\If {(no collision)}
\State move the particle close to the nearest one.
\EndIf
\EndIf
\EndFor
\EndWhile
\State Write the particle coordinates and colors to a file for the current time step.
\EndFor
\end{algorithmic}
\end{algorithm}
At each time step, we calculate the new value of $R_f$. Then we go through all the non-red particles and check the conditions for changing their status. Next, we shuffle the numbers of particles so that their subsequent search is random. After that, we repeatedly iterate over the moving particles that have not yet moved at the current time step and try to move each of them in turn. Failure to shift can be caused by a collision (neighbor particle overlapping) or the exit of a green particle beyond the $R_f$ boundary as a result of a diffusive motion. Repeated iterations are completed if all the moving particles were displaced at the current time step or there were no displacements at all during a repeat of the loop. To check that there is no collision of a particle with any other, it is advisable to consider the distance only to the nearest neighbors. To do this, we have to store additionally and periodically update information about subdomains containing particles (numbers of particles included in the subdomain).
\subsection{Determination of cluster sizes}
To determine the value of the average cluster size, an experimental picture was taken from~\cite{Park2006} for further processing. The data we are interested in was extracted as follows. The image was imported into the bitmap editor GIMP. A transparent layer was applied on top of it, which we worked with later. The brush size was set to match the size of the particles in the image~\cite{Park2006}. All particles not included in the annular sediment were drawn on the transparent layer manually on top of the experimental image. At the same time, we tried to choose different colors for each cluster for the convenience of further calculation.
\begin{algorithm}[H]
\caption{Algorithm for finding clusters.}
\label{alg:clustering}
\begin{algorithmic}[1]
\State Read the coordinates of the particles $x_i$, $y_i$ and their cluster numbers $c_i \leftarrow 0$ ($i \in [1;N_p]$).
\For {$i \leftarrow 1, N_p$}
\State $r\leftarrow \sqrt{x_i^2 + y_i^2}$
\If {$r>r_*$}
\State remove this particle from the array.
\EndIf
\EndFor
\State Cluster counter $C_n \leftarrow 1$.
\For {$i \leftarrow 1, N_p$}
\For {$j \leftarrow 1, N_p$}
\State Calculate the distance $s_{ij}$ between particles $i$ and $j$.
\If {$s_{ij}=d_p$ $\And$ $i \neq j$}
\If{$c_i= 0$ $\And$ $c_j= 0$ }
\State $c_i\leftarrow c_j\leftarrow C_n$
\State $C_n \leftarrow C_n +1$
\EndIf
\If{$c_i \neq 0$ $\And$ $c_j= 0$}
\State $c_j \leftarrow c_i$
\EndIf
\If {$c_i= 0$ $\And$ $c_j \neq 0$}
\State $c_i \leftarrow c_j$
\EndIf
\EndIf
\EndFor
\EndFor
\While{there are changes to cluster numbers}
\For {$i \leftarrow 1, N_p$}
\For {$j \leftarrow 1, N_p$}
\State skip the zero cluster particles.
\State $s_{ij}= \sqrt{(x_i-x_j)^2 + (y_i-y_j)^2}$
\If {$s_{ij}=d_p$ $\And$ $i \neq j$}
\If {particles $i$ and $j$ are not from one cluster}
\State We assign a value $\max(c_i,c_j)$ to the particle with $\min(c_i,c_j)$.
\EndIf
\EndIf
\EndFor
\EndFor
\EndWhile
\end{algorithmic}
\end{algorithm}
Sometimes it was difficult to visually determine whether a particular particle belongs to a particular cluster or not. In the process, we were guided by the following rule. If a particle is located close to another particle in the cluster, or the distance between these particles is much smaller than the particle size, then we assume that the current particle also belongs to this cluster. As a result, the layer with the original image was removed. Based on the received image, we manually calculated the average cluster size.
Processing of numerical simulation results was performed using the algorithm~\ref{alg:clustering} for finding clusters.
To determine the contact of two particles, the condition $s = d_p\pm \varepsilon d_p$ was used, where $s$ is the distance between two particles, the diameter of a spherical particle $d_p=2r_p$, and $\varepsilon$ is the constant ($\varepsilon=$ 0.01 is used here). This nuance was also taken into account when detecting a collision of two particles ($s < d_p- \varepsilon d_p$). The calculation of the average cluster size based on the results of ten repeated tests was performed using a script written in Python.
\section{Results and discussions}
A series of calculations were performed for different values of $R_n$ in the range from $3r_p$ to $22r_p$. Then the numerical result was approximated using the least-squares method (Figure~\ref{fig:averageNc_vs_Rn}). The plot in Figure~\ref{fig:averageNc_vs_Rn} resembles a sigmoid. The value of the average cluster size determined by us based on experimental data~\cite{Park2006} is $\langle N_c \rangle \approx 8.2$ (average number of particles in the chain). The model predicts this value for $R_n \approx 7.9 r_p$.
\begin{figure}[h]
\begin{minipage}[c]{0.7\textwidth}
\centering
\includegraphics[width=0.99\textwidth]{averageNc_vs_Rn}\\
\end{minipage}
\begin{minipage}[c]{0.3\textwidth}
\caption{\label{fig:averageNc_vs_Rn}Dependence of the average cluster size on the parameter $R_n$ (the standard deviation was used to estimate the error).}
\end{minipage}
\end{figure}
For comparison, the Figure~\ref{fig:experimentAndSimulationClusters} shows the final sediment structures based on experimental data~\cite{Park2006} and our simulation results. Only the particles located inside the annular sediment are shown. The subset of particles belonging to the ring was subtracted from their set using the condition $r>r_*$. The inner radius of the ring $r_*\approx 0.75R$ was determined based on calculation data~\cite{Kolegov2019} ($r_*$ corresponds to the place of a high gradient of the number of particles per unit area). The internal structure of such rings was discussed in detail earlier~\cite{Kolegov2019}.
Here we observe that for the given parameters, the model predicts the absence of single particles (Figure~\ref{fig:experimentAndSimulationClusters}). The visualization of experimental data shows a very small number of such particles. In both cases, there are clusters of different sizes (from two to several dozen particles in a cluster). Some clusters are similar in shape. It should be noted that based on the results of modeling, tree-shaped clusters are expected to appear. Their root element is located closer to the periphery. These ``trees'' grow towards the center of the area. It is explained by the nature of movement and the shape of the fixation boundary $R_f$. Such branching structures are not observed in the experiment. This distinctive feature of the numerical results is most likely related to the approximate description of the domain geometry~\cite{Kolegov2019}. After all, the shape of the drop at the final stage of evaporation is more like a film, in which local ruptures can occur~\cite{Kolegov2020}. Thus, the action of capillary forces occurs in a larger area. The model boundary $R_f$ and its neighborhood are only a rough approximation of such a region. Besides, the mimic of the capillary interaction of particles is used here. In a good way, it is necessary to solve the n-body problem numerically. Only in this problem, instead of gravity, there are capillary forces.
\begin{figure}[h]
\begin{minipage}[c]{0.35\textwidth}
\includegraphics[width=0.99\textwidth]{experimentClusters}
\end{minipage
\begin{minipage}[c]{0.35\textwidth}
\includegraphics[width=0.99\textwidth]{simulationClusters}
\end{minipage}
\begin{minipage}[c]{0.25\textwidth}
\caption{\label{fig:experimentAndSimulationClusters} Visualisation of clusters from experiment~\cite{Park2006} (left) and simulation results for $R_n = 8 r_p$ (right).}
\end{minipage}
\end{figure}
Analysis of the algorithm~\ref{alg:particleDynamicsModified} showed a weak nonlinearity of the dependence of the simulation time on the number of particles $N_p$ (Figure~\ref{fig:NpVsTime}). Also, it was found that the average (over time steps) rate of movable particle displacement failures for the modified algorithm (with multiple attempts to particle displacement) is 1--2\% according to the results of calculations. For comparison, we tested the original algorithm (with a single attempt to shift the particles) and got a failure rate of 3-4\%. Five tests were run simultaneously for each type of calculation (the error is less than the marker size on the Figure~\ref{fig:NpVsTime}). Thus, for a sufficiently small value of the time step, the error associated with the failure of the particle shift is small. Repeated attempts to shift particles can slow down the program and not give a significant difference in the result.
\begin{figure}[h]
\begin{minipage}[c]{0.7\textwidth}
\centering
\includegraphics[width=0.99\textwidth]{NpVsTime}\\
\end{minipage}
\begin{minipage}[c]{0.3\textwidth}
\caption{\label{fig:NpVsTime}Dependence of the simulation time on the particle number (the program was written in Python, Intel (R) Core (TM) i9-9900K CPU 3.6 GHz was used).}
\end{minipage}
\end{figure}
\section{Conclusion}
Numerical results~\cite{Kolegov2019} are qualitatively agreed with the experiment~\cite{Park2006}. As the droplet dries, chains of particles (clusters) are formed inside the annular sediment. In the current work, we have processed experimental data~\cite{Park2006}, found out the average cluster size, and defined the value of a model parameter $R_n$ to predict a quantitatively agreed value $\langle N_c \rangle$. But it is also worth mentioning the qualitative difference between the simulation and experimental results, related to the shape of some clusters. The simple model under consideration often predicts tree-shaped clusters, which is not typical for the experiment~\cite{Park2006}. Further study is needed in this direction to better understand the processes that are taking place. Some questions can be answered by using the high-speed shooting of the formation of such clusters. Additional experimental data are also needed, for example, on the shape of the free surface of the liquid at the time stage when these clusters begin to appear. Further creation of more complex and accurate models that can predict the internal structure of such precipitation is important.
\section*{Acknowledgment}
This work is supported by the grant 18-71-10061 from the Russian Science Foundation. We thank Svetlana Kolegova for her help in processing experimental data and Yuri Tarasevich for useful comments.
\section*{References}
|
\subsection{Appendix 1}
\subsection{Definition of WoodNet CNN in PyTorch Code}
\label{app:woodnet}
\begin{code}
\captionof{listing}{Code that defines the neural network used}
\label{code:woodnet}
\inputminted[breaklines]{Python}{./Code/woodnet.m}
\end{code}
\newpage
\subsection{Definition of BadNet in PyTorch Code}
\label{app:badnet}
\begin{code}
\captionof{listing}{Code that defines BadNet, a simple fully connected network used for comparison}
\label{code:badnet}
\inputminted[breaklines]{Python}{./Code/badnet.m}
\end{code}
\newpage
\subsection{Example Output from the Training Loop after Three Epochs}
\label{app:loop-output}
\begin{center}
\lstset{%
caption=Example of the output produced by the training loop,
basicstyle=\ttfamily\footnotesize\bfseries,
frame=tb,
label=lst:loop-output
}
\begin{lstlisting}
Training only last layer of SqueezeNet
Epoch 0/24
----------
Epoch: 0 (train): 10
train Loss: 0.1488 Acc: 0.9476
Epoch: 0 (val): 10
val Loss: 0.0499 Acc: 0.9851
Epoch 1/24
----------
Epoch: 1 (train): 10
train Loss: 0.0826 Acc: 0.9714
Epoch: 1 (val): 10
val Loss: 0.0353 Acc: 0.9900
Epoch 2/24
----------
Epoch: 2 (train): 10
train Loss: 0.0704 Acc: 0.9760
Epoch: 2 (val): 10
val Loss: 0.0301 Acc: 0.9910
Epoch 3/24
----------
Epoch: 3 (train): 10
train Loss: 0.0666 Acc: 0.9766
Epoch: 3 (val): 1
\end{lstlisting}
\end{center}
\newpage
\subsection{Loss and Accuracy for WoodNet and Squeeznet for Four Epochs}
\begin{table}[H]
\centering
\begin{tabular}{c|l|ll|ll}
\multicolumn{1}{l|}{} & & \multicolumn{2}{c|}{\textbf{WoodNet}} & \multicolumn{2}{c}{\textbf{SqueezeNet}} \\
\multicolumn{1}{l|}{\textbf{Epoch}} & \textbf{Phase} & \multicolumn{1}{c}{\textbf{Loss}} & \multicolumn{1}{c|}{\textbf{Accuracy}} & \multicolumn{1}{c}{\textbf{Loss}} & \multicolumn{1}{c}{\textbf{Accuracy}} \\ \hline
\multirow{2}{*}{\textbf{1}} & Train & 1.1977 & 0.4230 & 0.1737 & 0.9446 \\
& Val & 0.5233 & 0.7943 & 0.0709 & 0.9805 \\ \hline
\multirow{2}{*}{\textbf{2}} & Train & 0.2519 & 0.9061 & 0.0863 & 0.9738 \\
& Val & 0.0815 & 0.9739 & 0.0521 & 0.9849 \\ \hline
\multirow{2}{*}{\textbf{3}} & Train & 0.0626 & 0.9786 & 0.0745 & 0.9765 \\
& Val & 0.0340 & 0.9884 & 0.0424 & 0.9887 \\ \hline
\multirow{2}{*}{\textbf{4}} & Train & 0.0286 & 0.9900 & 0.0656 & 0.9793 \\
& Val & 0.0249 & 0.9920 & 0.0396 & 0.9894
\end{tabular}
\caption[Loss and accuracy for WoodNet and SqueezeNet for the first four epochs.]{Loss and accuracy for WoodNet and SqueezeNet for the first four epochs of training and validation on the face-cropped dataset.}
\label{tab:training-table}
\end{table}
\section{Introduction}
\label{sec:introduction}
This paper addresses the supervised classification problem of \emph{facial recognition} using \emph{convolutional neural networks} (CNNs). That is, given an image, classify the image into one of a finite set of classes, each representing a person. This problem is addressed with the intention of applying it to entry access control, an approach that could provide fast, easy, and secure, key-less access to e.g. an apartment. This is an interesting problem for a number of reasons. For one, the prevalence of biometric recognition is growing, highlighting the relevance of this paper. For instance, facial recognition systems are now commonplace in virtually all smartphones. Furthermore, it is said that there is nothing new under the sun – except the extensive dataset used in this project. Applying CNNs for image classification is not exactly new, however, this project's novelty stems from a meticulously constructed dataset, a network made from scratch, and the comparison with a pre-trained network. It is thus a novel application of facial recognition which, if successful, may yield very tangible and meaningful results (to the authors at least). This was a critical factor for the authors in choosing this project.
To classify faces, a model that accepts images and outputs a prediction is created. The data used for this task is color images of people and other motifs. The model's goal is to classify the images into one of the four classes: $\{Kjartan, Lars, Morgan, Other\}$, where the three former classes represent the faces of the authors of the paper. Such a model may function as the discriminator in an automated face-based admission control system, where three people have approved access while the ``Other'' category represent trespassers - i.e. everyone else. One application of such a system could be controlling access of the door to an apartment the three authors share.
The data used for training has been gathered and processed by the authors. Collection starts by filming faces in different settings with varied conditions (e.g lighting). The videos are converted into separate images, and each image is processed and augmented to increase the amount of data. This way, the model may avoid discriminating classes based on spurious correlations with faces, like background, clothing, lighting etc. To gather data for each class, both the authors and other people are filmed in the differing settings. The steps taken to preprocess the data is a substantial part of the workload, and is described in further detail in \autoref{sec:data}.
In order to classify faces, a \emph{deep learning} model, specifically a CNN called \emph{WoodNet} is designed by the authors. CNNs are appropriate for facial recognition as they are well suited for analyzing grid-like data such as images through their distinctive convolution operation\footnote{The convolution operation is elaborated upon in detail in the method paper previously written by the authors.}. This enables CNNs to exploit several invariant properties of the data such as \emph{translational invariance} and \emph{spatial locality}, allowing for identification of patterns in images, such as features of a face. \emph{Parameter sharing} makes CNNs small compared to a comparable dense neural network. In order to train the network, \emph{Stochastic Gradient Descent} (SGD) with \emph{backpropagation} is utilized, a well-documented method for optimization of the network. See \autoref{sec:methods} for details.
In addition to implementing a CNN, the machine learning method \emph{transfer learning} is also applied. Numerous publicly available neural networks, trained with large amount of images (e.g ImageNet with 10 million images over 10 000 categories) exist. Transfer learning is utilized by modifying the final layer of such a large and well-trained net. This allows for rapid retraining while leveraging the accurate feature extraction of the trained model. In this paper, a condensed version of AlexNet \citep{NIPS2012_c399862d}, called SqueezeNet \citep{iandola2016squeezenet} is chosen. See \autoref{sec:related-work} and \autoref{sec:methods} for further details.
Lastly, the authors would like to emphasize that working on this project has meant iteratively experimenting on countless different models, data processing approaches and hours upon hours of training without seeing many results. Since this paper is part of the NTNU course ``TDT4173 - Machine Learning'', the different attempts performed are supplied in the \href{https://github.com/Neuralwood-Net/woodnet/tree/main/notebooks/archive}{\color{blue}\texttt{notebooks/archive}} folder in this project's GitHub repository to document the learning process.
\section{Related Work}
\label{sec:related-work}
This section contextualizes this paper's project within the field of deep learning, with emphasis on CNNs and facial recognition. Presentations of related work is divided into two groups: first, outline of important research for deep learning generally, and second, an overview of crucial contributions to the problem of facial recognition. As \autoref{fig:fr-overview} shows, over the last three decades, the state-of-the-art facial recognition algorithms' accuracy have improved in correlation with an increasing prevalence of deep learning approaches and CNNs.
\emph{Convolutional neural networks} \\
The breakthrough application of CNNs on the MNIST dataset in 1998 \citep{LeCun1998GradientbasedLA} marked the beginning of a revolution in using neural networks for complex machine learning, in particular image recognition, classification, detection, and segmentation. Ever since \emph{AlexNet} \citep{NIPS2012_c399862d} won the \emph{ImageNet Large Scale Visual Recognition Challenge} (ILSVRC) competition in 2012, CNNs and deep learning have proven to be the most proficient methods for image classification. Another method used in this paper, which is important for the evolution of the field of deep learning, is SGD \citep{MIT-6-036}, which computes the gradient of the parameters using only a random sample of training examples. \emph{Adam} \citep{Adam} is a slightly more complex and adaptive variant that often leads to more efficient training, and is discussed in \autoref{sec:methods}.
\emph{Facial recognition} \\ AlexNet also contributed to the field of facial recognition specifically. It was used as an underlying architecture for many of the first deep learning facial recognition algorithms implemented in the years around 2013 \citep{DBLP:journals/corr/abs-1804-06655}. In this paper, a smaller\footnote{In terms of number of weights.} version of AlexNet, \emph{SqueezeNet} \citep{iandola2016squeezenet} was used as the base for the transfer learning approach and proved to achieve the higher accuracy faster (see \autoref{sec:Transfer-Learning}).
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{Images/RelatedWork/Overview_of_facial_recogition.png}
\caption[Historical overview of facial recognition approaches.]{Historical overview of facial recognition approaches.}
\source{\cite{DBLP:journals/corr/abs-1804-06655}}
\label{fig:fr-overview}
\end{figure}
One highly relevant facial recognition algorithm in relation to this paper, is the \emph{FaceNet} model, developed by Google researchers \citep{DBLP:journals/corr/SchroffKP15}. It still is one of the most accurate facial recognition algorithms\footnote{\href{http://vis-www.cs.umass.edu/lfw/results.html}{\color{blue}http://vis-www.cs.umass.edu/lfw/results.html}}. FaceNet uses a large private database of faces, and complex loss function to train a GoogLeNet \citep{DBLP:journals/corr/SzegedyLJSRAEVR14} to extract face embeddings. It achieves a staggering 99.63\% accuracy on the \emph{Labeled Faces in the Wild} (LFW) database which consists of 13,000 labeled pictures \citep{LFWTech}. The paper also shows that the sizes of the datasets continue to matter for large sample sizes. Using tens of millions of images resulted in a clear boost of accuracy as compared to ``only'' millions of images. Over this interval the relative reduction in error is 60\%. The data collected for this project consists of approximately 150 000 images (see \autoref{sec:data}). Experiences from FaceNet might indicate that this project could have benefited from gathering a significantly larger dataset.
\section{Data}
\label{sec:data}
To classify the authors' faces, a dataset with 156 241 color images of size 224 by 224 pixels have been produced. There are $156\,240 / 4 = 39\,060$ images for each of the four classes $\{Lars, Morgan, \\ Kjartan, Other\}$. The images are split into a training, validation, and test set in parts of 70\%, 15\%, and 15\% or 109 368, 23 436, and 23 436 images, respectively. This split is chosen because it provides a reasonable trade-off between sufficient test and validation data and having enough data to train on. In testing the models developed during the project, two versions of the dataset was produced for comparison: one that uses face detection to crop out faces, and one that crops the image around the center, making it square.
As the raw data is high resolution videos, a number of processing steps are required to create a dataset ready to fit a model with. This section outlines the collection and transformations of the data used. See \autoref{fig:pipeline} for a visual overview. The raw images and the finished datasets can be found under ``additional resources'' below the abstract.
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{Images/Data/processing_pipeline.pdf}
\caption[The image processing pipeline used to generate training data.]{Pipeline for generating training data. An alternative pipeline without face detection was also used, where images were cropped to square around the center (see \autoref{subsec:compression}). The numbers below the images give the dimensions at each step. N is any size and depends on the size of the face in any given image.}
\label{fig:pipeline}
\end{figure}
\subsection{Raw Data Collection}
Data collection was performed by recording videos of each of the authors' and their friends' faces repeatedly, with a variety of backgrounds and light settings. Video recordings were used to capture high volumes of images efficiently. The videos were converted to separate images, each with one person in focus. While recording, the captured face was centered, while keeping nothing else in the frame constant. This was done by capturing videos with different backgrounds, surrounding colors, lighting, clothes etc. This way, data that induces learning of facial features specifically should be generated. The variation should prevent the neural networks from learning to classify individuals by, for instance, a dazzling sweater. More formally, given two images of the same class, the ambition is for their intersection to exclusively be some structure the computer can recognize as a single individual's face.
In addition to images of friends of the authors, the \emph{Other} class also contains pictures without people in them in an effort to make the networks learn to predict this class when there is no person in the frame. In an entry access control scenario, this ensures that the system does not wrongfully activate and allow access when there is no one in front of the camera.
As the data collection resulted in a varied number of images per class, the dataset was balanced before further processing was conducted. Let $n$ denote the number of images in the class containing the fewest images. In order to balance the dataset, a random sample of $n$ images were extracted for each of the three other classes. This resulted in 7812 images per class before augmentation. Balancing prevents bias towards heavier classes and prevents reduced predictive performance regarding the underrepresented classes.
\subsection{Compression}
\label{subsec:compression}
For efficient storage, and more importantly, rapid training, the images were compressed to a smaller size. This is a trade-off between information and performance, as one has to balance information with data size and network size. An image size of 224x224 was chosen after iterative experimentation, where several different sizes were tested. For images of size 64x64, there was clearly a substantial loss of information, and even the human eye struggled with classification. At 128x128, the human eye could easily distinguish between classes, but training the model to do so proved challenging as low accuracies were achieved. At 224x224 pixels the images seem detailed enough for efficient feature extraction. At the other extreme, images of size 1080x1080 (length of short end of original images) would increase use of space and complexity by a factor of $\frac{1080^2}{224^2}\approx 23$. Lastly, since ImageNet images are 224x224, that further reinforces the assumption that this size is a good trade-off between information and size. The preprocessing of the data was done in three parts. These are detailed below, accompanied by reflections around and justifications behind each processing step. The code used to perform this processing can be found in \href{https://github.com/Neuralwood-Net/woodnet/blob/main/notebooks/data-extraction.ipynb}{\color{blue}\texttt{notebooks/data-extraction.ipynb}}.
\begin{enumerate}
\item \textit{Face detection.} To remove background and focus on the faces, faces were detected using a third-party pre-trained CNN, an implementation of the \emph{Multi-Task Cascaded CNN} (MTCNN) as described in \cite{mtcnn}. See \autoref{fig:pipeline} for an example of face detection applied to the collected data. The intent is to crop the image around the detected face (see step 2). This should improve the model's learning rate as it allows the neural network to train only on what is important: faces. The problem of detecting a face is a vastly different task to that of recognizing (classifying) one, and therefore using the pre-trained network to bound the face in each image seems suitable. As \cite{DBLP:journals/corr/abs-1804-06655} shows, separating the task of extracting a face from an image from the task of matching a face with a name is a common practice for even the most proficient facial recognition models. Furthermore, feeding WoodNet two considerably similar photos yielded two different predictions, where only one was correct, see \autoref{fig:morgan-zoom}. This illustrates how different the task of face detection and recognition is, and that cropping faces seems to be conducive for learning.
\item \textit{Crop image to square.} After detecting a face in the image, the image is cropped around the bounding box yielded from the face detection CNN to produce a square picture of an individual's face. While obtaining the benefits of removing background, cropping the image helps reduce file sizes. For the dataset without face detection, this is the first step of the compression processing, and cropping is done by simply removing pixels from the edges of the image to make it square.
\item \textit{Resizing.} As a final step, all images were resized to the size of 224 by 224 RGB pixels. In most cases, this resize is a downsize, as the input image will only be smaller in the case where the face detection algorithm detected a face far away. The resizing is primarily done to save resources, reduce training time and interface with SqueezeNet, although most likely at the cost of losing some valuable information in the data.
\end{enumerate}
\begin{figure}
\centering
\includegraphics[width=.4\textwidth]{Images/Data/morgan_twice.eps}
\caption[Demonstration of separating facial detection and recognition.]{Demonstration of separating facial detection and recognition. WoodNet does not recognize Morgan to the left, and is 99.98\% certain it is the \emph{Other} class (i.e. less than 0.02\% certain that it \textit{is} Morgan). However, on the right, WoodNet is 97.95\% certain it is him.}
\label{fig:morgan-zoom}
\end{figure}
\subsection{Augmentation}
The last step in the preprocessing pipeline is image augmentation. After collecting the data and compressing it, there are 7 812 224x224 RGB images. As the images stem from videos, many of the individual images are similar. One can imagine a neural network overfitting to images that appear quite similar. Furthermore, it would be problematic if the dataset consisted of a small number of original images. To mitigate both of these issues, each image is randomly augmented in order to grow both the size and diversity of the dataset. Each original image is augmented 19 times, yielding 20 images per original image. Every augmentation include the same types of manipulations, however they are applied in random order and with random parameters within their respective ranges. The following transformations were used: Rotation of $\pm 5^{\circ}$, scaling the image by $-5\%$ to $+10\% $ by cropping and padding, additive Gaussian noise using a kernel with $0<\sigma\leq1$, changing the brightness from $-10$ up to $+10$ of the original value, and finally translating the image in $x$ and $y$ direction by $-10\%$ to $+10\%$ per axis. The transformations and their parameters were chosen on the basis that they sufficiently altered the picture without making it unrecognizable for the human eye.
The right side of \autoref{fig:pipeline} shows an example image that has gone trough these transformations, creating 19 augmentations. By generating 19 new augmentations and including the original image, the dataset size is increased to $7\,812\cdot 20 = 156\,240$ images. The choice of 19 augmentations was made as a result of a qualitative evaluation of the trade-off between a larger dataset and more similar images. Naturally, by producing more augmentations per image the size of the data could be considerably increased, but at the risk of new images being a regurgitation of old ones.
\section{Methods}
\label{sec:methods}
The high level approach used in this paper is a deep convolutional neural network, where loss is measured by \emph{cross-entropy loss}, optimized with the stochastic gradient descent variant Adam, and trained on a dataset generated by the authors (see \autoref{sec:data}). Transfer learning by adapting a fully-trained network to this specific problem is also tested. This section elaborates on these methods and their rationale. In \href{https://github.com/Neuralwood-Net/woodnet/blob/main/notebooks/training-and-plotting.ipynb}{\color{blue}\texttt{notebooks/training-and-plotting.ipynb}} one can see and run the code used for defining the networks, training and plotting.
\subsection{Convolutional Neural Network Architecture}
As discussed in \autoref{sec:related-work}, CNNs have for a decade been dominant in most machine learning applications involving data with some aspect of temporal or spatial locality. As the problem for this paper is an image classification problem, the CNN is a natural choice.
To ease the implementation effort, the project relies on the deep learning framework \emph{PyTorch} (version 1.7.0+cu101)\footnote{\href{https://pytorch.org/}{\color{blue}https://pytorch.org/}}. This framework is chosen for several reasons. First, from the authors' research, it appears to be the most beginner friendly and pythonic, i.e. it feels most like programming pure python. Second, it is used by companies like Tesla to power the neural nets in their self-driving cars\footnote{\href{https://www.youtube.com/watch?v=oBklltKXtDE}{\color{blue}PyTorch at Tesla - Andrej Karpathy, Tesla}}, and should therefore be sufficient for this undertaking as well.
Through iterative experimentation, a network architecture prioritizing understanding, simplicity, and learning was designed. This was prioritized over creating the absolute best-performing network achievable. Certainly, when taking \autoref{sec:related-work} and for instance FaceNet into consideration, the fact that the neural network was trained using thousands of hours on Google CPU clusters \citep{DBLP:journals/corr/SchroffKP15}, with the company's private database of faces in the hundreds of millions, makes a goal of matching their accuracy seem unreasonable for this project. This resulted in the implementation of a fairly simple architecture. Inspiration for possible architectures has been drawn from distinguished architectures such as FaceNet \citep{DBLP:journals/corr/SchroffKP15}, VGG \citep{simonyan2015deep}, AlexNet \citep{NIPS2012_c399862d}, SqueezeNet \citep{iandola2016squeezenet}, etc. Simple architectures that solve the MNIST challenge were also researched. The end result is an architecture in between the two extremes. In short, the development started with a very simple network, see e.g. the code for ``BadNet''\footnote{Tongue in cheek name for a network that performs a lot worse than the final architecture of WoodNet.} in \autoref{code:badnet} in \autoref{app:badnet} of the appendix. This architecture did not prove sufficient for efficient learning. Iteratively, the net was extended with more layers while monitoring the performance on validation data until a sufficiently deep network was found in WoodNet. Kernel size and number of dropout layers (see \autoref{sec:activation-function}) were also varied. As an example, WoodNet with dropout after every layer was much less apt for learning than the version with dropout only after the second-to-last layer. The PyTorch code defining the final network, WoodNet, is given in \autoref{code:woodnet} in \autoref{app:woodnet} of the appendix, and a visualization of the architecture can be seen in \autoref{fig:architecture}.
WoodNet consists of a feature extraction phase and a classification phase. The feature extraction phase is centered around triplets of the two-dimensional convolution layer, \emph{max pooling} \citep{Riesenhuber1999HierarchicalMO}, and a \emph{Rectified Linear Unit} (ReLU) non-linear activation \citep{ramachandran2017searching}. There are four sets of these triplets. The convolution layer is the distinctive element of a CNN \citep{Goodfellow-et-al-2016}. The max pooling cuts the height and width of the image in half for every convolution unit, but will, in theory, output more and more refined features. In short, an image of 224 by 224 pixels by 3 channels is fed in, and iteratively the model attempts to extract larger and larger features and ultimately outputs them in the form of a $7\times 7\times 64$ tensor. The max pool layer is chosen over e.g. average pooling because by allowing only the maximum value through the filter, it will emphasize the strongest signal which can help in propagating features through the network.
The output tensor from the feature extraction phase is then passed to the classification layer. This consists of three fully connected layers. The first layer takes in a flattened tensor of length $7\cdot 7\cdot 64 = 3136$, representing features if working as intended. The input layer consists of 2048 neurons, and is connected to the hidden classification layer that has 1024 neurons, before it is passed to the final layer that has four neurons, i.e. one for each class in the dataset. Every fully connected layer except the last is followed by a ReLU non-linearity and a dropout layer \citep{labach2019survey}.
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{Images/Methods/architecture.png}
\caption[Visualization of the WoodNet architecture.]{Visualization of the architecture of WoodNet. The first 5 boxes from the left represent convolution units with given heights, widths, number of channels and filter sizes. The final four boxes to the right represent fully connected layers resembling arrays.}
\label{fig:architecture}
\end{figure}
\subsection{Choice of Activation Function and Regularization}
\label{sec:activation-function}
To ensure that the neural network is actually a function approximator that can approximate non-linear functions, the network requires non-linear activation functions \citep{MIT-6-036}, i.e. functions that take the output of one layer and performs a non-linear transformation to it. The ReLU is such a function and is defined as $\textnormal{ReLU}(x) = \max(0, x)$ \citep{ramachandran2017searching}. This activation function is chosen as it is simple and easy to understand and its derivative is trivially calculated, which speeds up training. Furthermore, it has proven effective in practice and is utilized by all the successful architectures mentioned previously. The final layer does not have an activation function. The \emph{Softmax} function could have been used, but was not necessary due to the choice of loss function (see \autoref{subsec:loss-func-optimization}). The Softmax function is however used to normalize the logits the network outputs to be able to interpret the model outputs as probabilities, when appropriate.
In an effort to make the network generalize and not be overly fitted to the training data, the network incorporates \emph{dropout} layers. Dropout prevents overfitting by preventing co-adaption of the network parameters on the data \citep{hinton2012improving}. This is achieved by randomly turning off connections between neurons during training. This way, it is theorized, that the remaining neurons are forced to generalize to the data, and over time the ``signal'' will emerge out of the noise. In the network, dropout is implemented in between the two last fully connected layers. Too much dropout can hinder the network from learning anything at all, as the authors experienced during experimentation.
\subsection{Loss Function and Optimization Criterion}
\label{subsec:loss-func-optimization}
The basic, and most widely used approach to training neural networks is stochastic gradient descent (SGD) \citep{MIT-6-036}. This involves calculating gradients of weights in the network, to figure out how they can be nudged towards an optimum. This requires a way to quantify how well or poorly the network performs at the classification task, i.e. a \emph{loss function}. The loss function solves this by providing a quantitative measure of how bad a certain prediction is. When this function is differentiated with respect to the parameters in the model, with the help of the chain rule for derivatives, one can use the resulting gradient to decide how the parameters in the network can be tuned to most efficiently decrease the loss, and thus increase the accuracy. This paper uses the \emph{Cross Entropy Loss} or \emph{Negative Log Likelihood Loss}. Cross entropy loss is useful for classification problems, generalizes over an arbitrary number of classes, and is easily differentiated \citep{Goodfellow-et-al-2016}. As this task is a classification task, cross entropy loss suits the problem well. Given $N$ input samples, $M$ classes, binary indicator $y_{i,c}$ stating whether observation $i$ is classified as class $c$ and $p_{i,c}$, the predicted probability of observation $i$ being of class $c$, the multi-class cross entropy loss function can be written as
\begin{align*}
H_p(q)=-\frac{1}{N}\sum_{i=1}^{N}\sum_{c=1}^{M} y_{i,c} log(p_{i,c}). \quad \text{\citep{MIT-6-036}}
\end{align*}
As mentioned, the network will be trained using SGD. However, there are variations of the classic algorithm that provide some advantages. In normal SGD, the step size used to adjust the weights are shared between all weights, and is normally decayed at some fixed rate. In this project, Adam is used instead, which utilizes adaptive estimates of the \emph{moment}, or change in loss to change in parameters, to more accurately set appropriate learning rates for efficient learning \citep{kingma2017adam}. Furthermore, the authors argue that Adam is computationally efficient, requires little memory, and is well suited for problems requiring a large number of parameters or data. Lastly, they argue that it outperforms most other stochastic optimization for neural networks available. It was even designated as the default algorithm to use by Andrej Karpathy at Stanford University in the course \textit{CS231n: Convolutional Neural Networks for Visual Recognition\footnote{\href{http://cs231n.stanford.edu/}{\color{blue}http://cs231n.stanford.edu/}}}. It was thus a natural choice for this project.
\subsection{Training Loop and Data}
The model trains on the dataset the authors created, as described in \autoref{sec:data}. Prior to the model ever seeing data, it is divided into three distinct set of images: training, validation, and test. The model is then trained on the training data and validated on the validation data after each epoch of training. An epoch is one full iteration through the training data. That means that the model will see the training data multiple times and will therefore be prone to over-fit it, and not generalize to the embedded features in them. That is why the model is tested on the validation set after every epoch. The accuracy on both the training data and the validation data is expected to increase in tandem for a number of epochs, while the model is learning. However, once the model starts overfitting, the accuracy on the training data can be seen to increase while the accuracy on the validation data decreases as a consequence of the model ``memorizing'' the images in the training data, and not learning the generalized features in the images. That way, one can keep track of the best model as measured by accuracy on the validation data, and save those parameters.
The held-out test data is meant as a last validation of model generalization. It will only be used once the model and hyperparameters are decided upon. This is a measure to prevent overfitting to the validation data, which might happen if one experiments with a large array of different networks, hyperparameters, and configurations.
\subsection{Transfer Learning for Better Performance with Less Data}
\label{sec:Transfer-Learning}
Despite efforts in generating a large dataset specific for the task as discussed in \autoref{sec:data}, the dataset is in all likelihood too small and not diverse enough to enable a model to fully tease out specific features that are important for face recognition when trained from scratch. Therefore, the method of transfer learning is applied and contrasted with the results from learning from nothing. The network used for transfer learning is SqueezeNet \citep{iandola2016squeezenet}. This network is chosen because it shows similar performance to the well-known AlexNet on ImageNet, while having only one 50th the number of parameters and taking up 510 times less space. This suits this use-case because the network can be downloaded and fine-tuned quickly. The model is fully trained\footnote{\emph{``fully trained''} means that the model achieves state of the art performance on the given classification task, i.e. winning accuracy for ILSVRC 2012 of 57.5\% in this specific case.} on ImageNet with 1000 classes. To adapt it to the problem in this paper, the final classification layer will be replaced to predict for four classes instead. Then, the model is trained on the novel data. Fine-tuning the parameters in the final layer only will be experimented with. In theory, a model fully trained on ImageNet has learned how to effectively extract features from images, and will be able to effectively learn how to map those features to novel classes \citep{inductive-transfer}.
\subsection{Experiments and Evaluation}
The most important evaluation metric to be used to measure the performance of the models is the accuracy, i.e. percentage of correctly classified images. Furthermore, the loss value is recorded, as this is an indication of how ``bad'' the models are performing. In addition, precision and recall are interesting metrics given the context of access control. Considering a prediction of one of $\{Kjartan, Lars, Morgan\}$ as a positive prediction, and predicting $\{Other\}$ as a negative, precision and recall may be calculated. False positives represent allowing unauthorized access and the consequences of granting access for the wrong person are larger than not granting access for the right person. Thus, a high precision is desirable. Furthermore, a confusion matrix will be used to expand on this analysis, examining whether some classes are misclassified more often than others, and whether some classes are more often predicted as the result of a misclassification.
Another interesting experiment is to test the models with completely novel data, e.g. separate images of the authors from entirely different settings. An example of this is running live inference using an image stream from a webcam as input. Performance on these images will be a qualitative indicator of the models' ability to generalize. The dataset the networks are trained on is in all likelihood not sufficiently large or varied enough to enable the models to fully learn the features that define the faces of the authors. Therefore, it is to be expected that the accuracy on completely novel images is lower than that from the training, validation, and test set.
\section{Results}
\label{sec:results}
This section presents the main results of training WoodNet and SqueezeNet using the two different datasets (images cropped to faces and cropped around center). Discussion and reflections regarding the results are also included. Experiments run to settle on the final dataset is discussed in \autoref{sec:data} and experiments deriving the final CNN architecture of WoodNet is discussed in \autoref{sec:methods}.
\subsection{Presentation of Results}
\label{sec:presentation-of-results}
A summary of the key statistical measures of the two networks for each test dataset is presented in \autoref{tab:key-metrics}. The test dataset was never tested on the networks before the final models were settled on for the higher validity. Overall, WoodNet achieves a marginally lower loss and higher accuracy than SqueezeNet. Notably, the face-cropped dataset achieves a higher loss and lower recall for both models, but in turn achieves a higher accuracy and precision.
\autoref{fig:training-plot} and \autoref{tab:training-table} in the appendix shows the accuracy and loss from training WoodNet and SqueezeNet on the face-cropped dataset. As a full epoch of training is completed before a full epoch of validation, validation always start at a better level than training. An observation to highlight is that both networks seem to learn fairly quickly, dropping loss to zero and accuracy close to 100\% after two epochs. Furthermore, SqueezeNet seems to learn significantly faster than WoodNet in terms of number of images seen. A final observation is that none of the graphs exert any signs of overfitting, which is normally visible through diverging graphs for training and validation.
\begin{figure}
\centering
\includegraphics[width=0.9\textwidth]{Images/results/training_cropped.png}
\caption[Training of WoodNet and SqueezeNet.]{Training of WoodNet and SqueezeNet. Each row shows the training of a given network. Each column shows the accuracy and loss respectively, both during training and validation, over the course of the first three epochs of training. The black, vertical lines marks the change of epoch. The $x$-axis is the cumulative number of images trained on.}
\label{fig:training-plot}
\end{figure}
\begin{table}
\small
\centering
\begin{tabular}{l|ll|ll}
& \multicolumn{2}{c|}{\textbf{WoodNet}} & \multicolumn{2}{c}{\textbf{SqueezeNet}} \\
\textbf{Metric} & \multicolumn{1}{c}{\textbf{Generic crop}} & \multicolumn{1}{c|}{\textbf{Face-crop}} & \multicolumn{1}{c}{\textbf{Generic crop}} & \multicolumn{1}{c}{\textbf{Face-crop}} \\ \hline
Loss & 0.00895 & 0.00985 & 0.02685 & 0.03501 \\
Accuracy & 0.99710 & 0.99780 & 0.99200 & 0.99020 \\
Precision & 0.99903 & 0.99960 & 0.99624 & 0.99685 \\
Recall & 0.99875 & 0.99852 & 0.99749 & 0.99305
\end{tabular}
\caption[Summary of important metrics for the two networks.]{Summary of important metrics for the two networks and the two sets of training data, calculated from the held-out test set that no network was tested on prior to settling on final models.}
\label{tab:key-metrics}
\end{table}
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{Images/results/all_confusion_matrices.jpeg}
\caption[The confusion matrices for each of the tested models.]{The confusion matrices for WoodNet, SqueezeNet and BadNet when tested on the generically cropped test dataset. The $x$-axis represents what the model predicts, and the $y$-axis represents the true label of the image. BadNet is a small, fully connected neural network.}
\label{fig:confusion_matrices}
\end{figure}
Confusion matrices give an overview of how well a network predicts each class. A confusion matrix for each of the tested models can be seen in \autoref{fig:confusion_matrices}. For both SueezeNet and WoodNet, \textit{Kjartan} is most easily recognized, i.e. least likely to be misclassified. In fact this happened only 19 and 13 times for each network respectively. Furthermore, one can note that for both networks, \textit{Lars} and \textit{Kjartan} is most often misinterpreted to be \textit{Morgan}, with an average of 13 times for the two classes. \textit{Morgan} is most often wrongly predicted to be \textit{Other} by SqueezeNet and \textit{Lars} by WoodNet. Lastly, a small fully connected neural network named BadNet is included to show how a simpler network performs on the data. Note how \textit{Morgan} is its most likely predicted class for all classes except for \textit{Other}.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{Images/results/demo.png}
\caption[Sample screenshot from a live demonstration.]{Sample screenshot of live inference using the SqueezeNet architecture with face detection. Each bounding box represents a detected face, with corresponding predicted label and certainty.}
\label{fig:demo}
\end{figure}
Finally, a live demonstration was developed, running inference in real time using a web camera feed as input. See \autoref{fig:demo} for a sample screenshot. A video including the demonstration is accompanied with this paper and can be referred to for further examination of the performance of the network in a deployed scenario. Qualitatively, the authors were recognized well, however inaccuracies often occurred.
\subsection{Interpretations and Discussion}
As mentioned, \autoref{fig:training-plot} clearly shows that SqueezeNet learns much faster than WoodNet. This seems logical since SqueezeNet is already trained to achieve high accuracy on ImageNet, which is arguably a harder classification task. In theory, that is because SqueezeNet already knows how to effectively extract features from images. By training the last layer (incidentally called ``classifier''), it is only taught how to decide upon a class based on the features it has extracted. In the case of WoodNet, in addition to learning how do this classification, it must also learn how to extract features. Naturally this requires more exposure to images and more resources. Therefore, it seems reasonable to conclude that, as long as it is feasible, one should start with a trained model and adapt it instead of training from scratch.
The confusion matrices in \autoref{fig:confusion_matrices} contain interesting findings. Based on the observations made in \autoref{sec:presentation-of-results}, it seems the two networks do not contain the symmetric property of relations, $R(a,b)$ where $a$ and $b$ are classes and $R$ is ``most similar to''. For SqueezeNet, this is evident since \textit{Lars} is most often mistaken for \textit{Morgan} and \textit{Morgan} is most regularly misinterpreted as \textit{Other}. As such, we have $R(Lars, Morgan) \centernot\iff R(Morgan, Lars)$. Therefore, the way SqueezeNet recognizes a face seems to work differently than what humans are used to, assuming that a ``most similar to'' relation is symmetric for humans.
Despite zealous effort, over-fitting seems to be riddling both models. Accuracies on the training, validation, and test set, as seen in \autoref{tab:training-table}, \autoref{tab:key-metrics}, and \autoref{fig:training-plot}, are all very close to 100\%. Even though the authors naturally are happy with the produced model, it must be conceded that this result is most likely too good to be true. Inaccuracies during the live demonstration further indicate this. In the authors' best guess, the cause of over-fit is lack of data, both in terms of variety and numbers. Despite a large number of images, the method of producing these images might be conducive to over-fitting in two ways. The first is that images were extracted from a video recording 30 frames per second, producing similar frames despite constant movement. Furthermore, even though augmentations create different images, they are in many ways still just different views of the same underlying information, and might not be a perfect substitute to actual, unique images. However, as made evident by the video demonstration, the model really does seem to have learned some deeper features of the authors' faces. It is in no way perfect, but it is clear that it performs significantly better than chance, and the authors are more often than not correctly identified.
\section{Conclusion}
\label{sec:conclusion}
This paper has presented attempted solutions to the problem of facial recognition using convolutional neural networks as a discriminator in entry access control systems. Two proposed solutions in the form of neural network architectures were proposed and compared: building a neural network from scratch and training it using only self-collected data (WoodNet), and leveraging transfer learning by utilizing a network pre-trained for feature extraction of faces and adapting this to our self-collected data (SqueezeNet). This resulted in two models very capable of recognizing the faces of the authors, albeit at the cost of signs of overfitting to the specific data. Both models achieve high accuracy, approaching 100\%, however WoodNet reaches such a level considerably slower during training compared to SqueezeNet as a result of having to learn from scratch.
Considering time constraints and the ambitions of the project, the results seem satisfactory. In particular, the pipeline of data collection, processing, and training the models worked well. However, the quality of the outcome could have been improved through a number of actions. Firstly, as the models seem to be somewhat overfitted to the data, a larger, more diverse dataset could increase generalizability. The methodology for collecting data could have been more rigorous by ensuring less similar frames and using more different settings regarding lighting, background, colors etc. Additionally, collecting more data is always beneficial.
Results kept in mind, deploying the proposed solutions in a entry access control system could present some challenges. It is not entirely unlikely that the models classify strangers as one of the authors, effectively granting unauthorized access. Furthermore, there is little stopping an intruder from displaying an image of one of the authors to gain access. Thus, for such a system to work, further research into improving the models and implementing the access control system may be necessary. One solution could be to combine a camera-based system with other biometric approaches such as fingerprint scanning or voice recognition, or require several successful classifications before granting entry.
Given access to more resources like data, compute, and time, the authors would investigate the possibility of improving the models through use of a larger, more diverse dataset and a deeper network that may be able to extract more detailed features. This would both counteract overfitting as well as improve accuracy of inference. In addition, the evaluation method of \emph{cross-validation} would be leveraged, as it utilizes the data better and provides a valuable metric for the performance of the models, but requires considerable resources. Looking into alternative loss functions that penalize false positives harder than false negatives could also be an effective way to make the system more resilient to intruders.
|
\section{Convolutional Autoencoder Configurations} \label{conv_ae_appendix}
\begin{table}[H]
\centering
\begin{tabular}{c|c|c|c}
\textbf{Encoder} & & & \\ \hline \hline
\textbf{Convolutional Layers} & Kernels & Filter Size & Stride \\ \hline
Convolutional Layer 1 & 4 & $5\times 5$ & $2\times 2$ \\
Batchnormalization 1 & & & \\
Convolutional Layer 2 & 8 & $5\times 5$ & $2\times 2$ \\
Batchnormalization 2 & & & \\
Convolutional Layer 3 & 16 & $5\times 5$ & $2\times 2$ \\
Batchnormalization 3 & & & \\
Convolutional Layer 4 & 32 & $5\times 5$ & $2\times 2$ \\
Batchnormalization 4 & & & \\
Flatten & & & \\ \hline
\textbf{Dense Layers} & Neurons & & \\ \hline
Dense Layer 1 & $N_l$ & & \\ \hline
& & & \\ \hline
\textbf{Decoder} & & & \\ \hline \hline
\textbf{Dense Layers} & Neurons & Size & \\ \hline
Dense Layer 1 & 512 & & \\
Reshape & & $4\times 4\times 32$ & \\ \hline
\textbf{Convolutional Layers} & Kernels & Filter Size & Stride \\ \hline
Transposed Convolutional Layer 1 & 16 & $5\times 5$ & $2\times 2$ \\
Batchnormalization 1 & & & \\
Transposed Convolutional Layer 2 & 8 & $5\times 5$ & $2\times 2$ \\
Batchnormalization 2 & & & \\
Transposed Convolutional Layer 3 & 4 & $5\times 5$ & $2\times 2$ \\
Batchnormalization 3 & & & \\
Transposed Convolutional Layer 5 & 1 & $5\times 5$ & $2\times 2$ \\ \hline
\end{tabular}
\caption{Convolutional autoencoder configuration for the heat equation anf advection equation.}
\label{tab:con_ae_configuration_heat_adv}
\end{table}
\begin{table}[H]
\centering
\begin{tabular}{c|c|c|c}
\textbf{Encoder} & & & \\ \hline \hline
\textbf{Convolutional Layers} & Kernels & Filter Size & Stride \\ \hline
Convolutional Layer 1 & 4 & $5\times 5$ & $2\times 2$ \\
Batchnormalization 1 & & & \\
Convolutional Layer 2 & 8 & $5\times 5$ & $2\times 2$ \\
Batchnormalization 2 & & & \\
Convolutional Layer 3 & 16 & $5\times 5$ & $2\times 2$ \\
Batchnormalization 3 & & & \\
Convolutional Layer 4 & 32 & $5\times 5$ & $2\times 2$ \\
Batchnormalization 4 & & & \\
Flatten & & & \\ \hline
\textbf{Dense Layers} & Neurons & & \\ \hline
Dense Layer 1 & $64$ & & \\
Dense Layer 2 & $32$ & & \\
Dense Layer 3 & $N_l$ & & \\ \hline
& & & \\ \hline
\textbf{Decoder} & & & \\ \hline \hline
\textbf{Dense Layers} & Neurons & Size & \\ \hline
Dense Layer 1 & 32 & & \\
Dense Layer 2 & 64 & & \\
Dense Layer 3 & 1568 & & \\
Reshape & & $7\times 7\times 32$ & \\ \hline
\textbf{Convolutional Layers} & Kernels & Filter Size & Stride \\ \hline
Transposed Convolutional Layer 1 & 16 & $5\times 5$ & $2\times 2$ \\
Batchnormalization 1 & & & \\
Transposed Convolutional Layer 2 & 8 & $5\times 5$ & $2\times 2$ \\
Batchnormalization 2 & & & \\
Transposed Convolutional Layer 3 & 4 & $5\times 5$ & $2\times 2$ \\
Batchnormalization 3 & & & \\
Transposed Convolutional Layer 5 & 2 & $5\times 5$ & $2\times 2$ \\ \hline
\end{tabular}
\caption{Convolutional autoencoder configuration for the lid driven cavity problem.}
\label{tab:con_ae_configuration_ldc}
\end{table}
\section{Time Evolution Neural Network Configuration} \label{evolution_NN_appendix}
For the CCNN memory encoding the layers are organized as shown in Figure \ref{fig:CCNN_types}. For the LSTM we work network architectures of 3 layers 32 neurons in each LSTM layer.
Furthermore, before the LSTM or CCNN layers a each previous state is passed through a dense layer with 16 neurons. In TensorFlow 2.0 this type of layer is denoted \texttt{TimeDistributed}.
For the parameter encoding the neural network is a simple 3 layer deep network with 16 neurons in each layer.
For the final prediction we utilize a 3 layer deep NN with 32 neurons in each layer.
\begin{comment}
\section{Results} \label{appendix:appendix_results}
\subsection{Heat Equation}
\begin{figure}[H]
\centering
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/heat/heat_sol_t0}
\caption{$t=0$}
\label{fig:heat_sol_t0}
\end{subfigure}%
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/heat/heat_sol_t75}
\caption{$t=1$}
\label{fig:heat_sol_t75}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/heat/heat_sol_t15}
\caption{$t=10$}
\label{fig:heat_sol_t15}
\end{subfigure}\\
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/heat/heat_sol_pred_t0}
\caption{$t=0$}
\label{fig:heat_sol_pred_t0}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/heat/heat_sol_pred_t75}
\caption{$t=1$}
\label{fig:heat_sol_pred_t75}
\end{subfigure}%
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/heat/heat_sol_pred_t15}
\caption{$t=10$}
\label{fig:heat_sol_pred_t15}
\end{subfigure} \\
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/heat/heat_error_t0}
\caption{$t=0$}
\label{fig:heat_error_t0_appendix}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/heat/heat_error_t75}
\caption{$t=1$}
\label{fig:heat_error_t75_appendix}
\end{subfigure}%
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/heat/heat_error_t15}
\caption{$t=10$}
\label{fig:heat_error_t15_appendix}
\end{subfigure}
\caption{Solution to the heat equation diffusion rates, $\mu=(1.09044445, 0.35502073, 0.12671445, 1.01588356 )$. First row shows the high-fidelity solution, the middle row shows the neural network prediction (CCNN), and the lower column shows the pointwise absolute error.}
\label{fig:heat_sols}
\end{figure}
\subsection{Linear Advection Equation}
\begin{figure}[H]
\centering
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/linear_advection/linear_adv_sol_t0}
\caption{$t=0$}
\label{fig:linear_adv_sol_t0_appendix}
\end{subfigure}%
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/linear_advection/linear_adv_sol_t75}
\caption{$t=7.5$}
\label{fig:linear_adv_sol_t75_appendix}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/linear_advection/linear_adv_sol_t15}
\caption{$t=15$}
\label{fig:linear_adv_sol_t15_appendix}
\end{subfigure}\\
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/linear_advection/linear_adv_sol_pred_t0}
\caption{$t=0$}
\label{fig:linear_adv_sol_pred_t0_appendix}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/linear_advection/linear_adv_sol_pred_t75}
\caption{$t=7.5$}
\label{fig:linear_adv_sol_pred_t75_appendix}
\end{subfigure}%
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/linear_advection/linear_adv_sol_pred_t15}
\caption{$t=15$}
\label{fig:linear_adv_sol_pred_t15_appendix}
\end{subfigure} \\
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/linear_advection/linear_adv_error_t0}
\caption{$t=0$}
\label{fig:sub2}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/linear_advection/linear_adv_error_t75}
\caption{$t=7.5$}
\label{fig:linear_adv_error_t75_appendix}
\end{subfigure}%
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/linear_advection/linear_adv_error_t15}
\caption{$t=15$}
\label{fig:linear_adv_error_t15_appendix}
\end{subfigure}
\caption{Solution to the linear advection equation with velocity, $\mu_1=1.41605949$, and initial angle, $\mu_2=2.87439799$. First row shows the high-fidelity solution, the middle row shows the neural network prediction (CCNN), and the lower column shows the pointwise absolute error.}
\label{fig:linear_advection_sols_appendix}
\end{figure}
\subsection{Lid Driven Cavity}
\begin{figure}
\centering
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/ldc_vel_magnitude_t1.pdf}
\caption{$t=1$}
\label{fig:ldc_vel_magnitude_t1}
\end{subfigure}%
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/ldc_vel_magnitude_t5.pdf}
\caption{$t=3$}
\label{fig:ldc_vel_magnitude_t5}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/ldc_vel_magnitude_t10.pdf}
\caption{$t=10$}
\label{fig:ldc_vel_magnitude_t10}
\end{subfigure}\\
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/ldc_vel_magnitude_t1_pred.pdf}
\caption{$t=1$}
\label{fig:ldc_vel_magnitude_t1_pred}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/ldc_vel_magnitude_t5_pred.pdf}
\caption{$t=3$}
\label{fig:ldc_vel_magnitude_t5_pred}
\end{subfigure}%
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/ldc_vel_magnitude_t10_pred.pdf}
\caption{$t=10$}
\label{fig:ldc_vel_magnitude_t10_pred}
\end{subfigure} \\
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/ldc_vel_magnitude_t1_error.pdf}
\caption{$t=1$}
\label{fig:ldc_vel_magnitude_t1_error_appendix}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/ldc_vel_magnitude_t5_error.pdf}
\caption{$t=3$}
\label{fig:ldc_vel_magnitude_t5_error_appendix}
\end{subfigure}%
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/ldc_vel_magnitude_t10_error.pdf}
\caption{$t=10$}
\label{fig:ldc_vel_magnitude_t10_error_appendix}
\end{subfigure}
\caption{Velocity magnitude for the lid driven cavity problem with $Re = 190$. First row shows the high-fidelity solution, the middle row shows the neural network prediction (CCNN), and the lower column shows the pointwise absolute error.}
\label{fig:ldc_sols_and_error}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/CCNN_memory_test.pdf}
\caption{CCNN.}
\label{fig:ldc_CCNN_memory_test}
\end{subfigure}%
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/LSTM_memory_test.pdf}
\caption{LSTM.}
\label{fig:ldc_LSTM_memory_test}
\end{subfigure}
\caption{Comparison of CCNN and LSTM in relative error for each time step in high-fidelity space for the lid driven cavity problem for varying memory, $\xi$. The error for each time step is a computed average over 15 test cases with the standard error. }
\label{fig:test}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/CCNN_train_test.pdf}
\caption{CCNN.}
\label{fig:CCNN_train_test}
\end{subfigure}%
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/LSTM_train_test.pdf}
\caption{LSTM.}
\label{fig:LSTM_train_test}
\end{subfigure}
\caption{Comparison of CCNN and LSTM in relative error for each time step in high-fidelity space for the lid driven cavity problem for number of training samples, $N_{train}$.}
\label{fig:ldc_train_test}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/CCNN_residual_test.pdf}
\caption{CCNN.}
\label{fig:ldc_CCNN_residual_test}
\end{subfigure}%
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/LSTM_residual_test.pdf}
\caption{LSTM.}
\label{fig:ldc_LSTM_residual_test}
\end{subfigure}
\caption{Comparison of computing the next step directly the residual for the CCNN and LSTM. The figures show relative error for each time step in high-fidelity space for the lid driven cavity problem.}
\label{fig:ldc_residual_test}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/CCNN_beta_test_reg3.pdf}
\caption{$\beta_1 = 1e-3$.}
\label{fig:ldc_CCNN_beta_test_reg3}
\end{subfigure}%
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/CCNN_beta_test_reg6.pdf}
\caption{$\beta_1 = 1e-6$.}
\label{fig:ldc_CCNN_beta_test_reg6}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/CCNN_beta_test_reg6.pdf}
\caption{$\beta_1 = 1e-9$.}
\label{fig:ldc_CCNN_beta_test_reg9}
\end{subfigure}
\caption{Impact of the two regularization terms, weight decay, $\beta_1$, and Jacobian, $\beta_2$ for the lid driven cavity problem. The average relative error in high-fidelity space over 15 test trajectories for each time step is shown. Each figure shows the error for a constant $\beta_1$ and varying $\beta_2$. }
\label{fig:ldc_beta_test}
\end{figure}
\end{comment}
\end{document}
\section{Approximating Parameterized Time Evolution using Neural Networks}\label{Approximating Parameterized Time Evolution using Neural Networks}
In the previous section we presented the general framework for nonlinear dimensionality reduction and showcased how convolutional autoencoders fit into this framework. In this section we aim to explain how neural networks will be utilized for approximating the dynamics in the latent space. For a brief review of the relevant types of neural networks, see \ref{appendix:neural_network}. Neural networks have already shown to be able to approximate dynamical systems \cite{xu2019multi, fresca2020comprehensive, erichson2019physics}. For this reason, it makes sense to test out various architectures for this exact purpose.
We aim to approximate the dynamics in the latent space non-intrusively by a function, $\Psi \approx F_{l,\delta t}$:
\begin{align}
u^{n+1} = \Psi(u^{n}),
\end{align}
where $\Psi$ is a neural network. The approximated latent states will be denoted, $\tilde{u}^n_l(\mu)$, to distinguish from the encoded high-fidelity state, $u_l^n(\mu)=\Phi_{enc}(u_h^n(\mu))$. Thereby, we aim to achieve:
\begin{align}
\begin{split}
&\left\{ \tilde{u}_l^{0}(\mu_1), \ldots, \tilde{u}_l^{N_t}(\mu_1) , \ldots , \tilde{u}_l^{0}(\mu_{N_{train}}) , \ldots, \tilde{u}_l^{N_t}(\mu_{N_{train}}) \right\} \\
\approx &\left\{ u_l^{0}(\mu_1), \ldots, u_l^{N_t}(\mu_1) , \ldots , u_l^{0}(\mu_{N_{train}}) , \ldots, u_l^{N_t}(\mu_{N_{train}}) \right\}
\end{split}
\end{align}
\subsubsection*{Taking Longer Steps}
Using high-fidelity methods for time-stepping often includes some restrictions on the step size in order for the scheme to be stable. An example is the Courant–Friedrichs–Lewy (CFL) condition for advection-dominated problems \cite{leveque2007finite}. With our strategy, where we aim to learn a neural network representation of the time evolution map, there is no immediate connection between step size and stability. Therefore, in order to speed up online computations, the neural network can be trained to learn to take steps of size $s\delta t$. Hence, $\Psi \approx F_{l,s\delta t}$.
In the offline phase the high-fidelity trajectories are still computed with step size $\delta t$, to ensure stability, but only every $s$'th step is used for training the NN:
\begin{align}
\left\{ u_h^{0}(\mu),u_h^{1}(\mu),u_h^{2}(\mu) \ldots, u_h^{N_t}(\mu) \right\} \mapsto \left\{ u_h^{0}(\mu), u_h^{s}(\mu), u_h^{2s}(\mu) \ldots, u_h^{N_t}(\mu) \right\}.
\end{align}
In general, two states one time-step apart, say $u_h^{n}(\mu)$ and $u_h^{n+1}(\mu)$, are highly correlated. In practice the means that we gain very little extra information by using both in the training of the NN. Therefore, it makes sense only use every $s$'th step to save memory and speed up the training. Hoeever, the number $s$ must be chosen according to various factors, like the requested detail of the dynamics in the online phase. It should further be kept in mind that larger $s$ results in a more complicated map to learn, and thus complicates the training.
For simplicity we will use the notation $\left\{ u_h^{0}(\mu),u_h^{1}(\mu),u_h^{2}(\mu) \ldots, u_h^{N_t}(\mu) \right\}$ when referring to the trajectory used for training the neural network.
\subsubsection*{Approximating the State vs. Residual}
At first glance, it makes sense to train a neural network to approximate $u_l^{n+1}$ directly given $u^{n}$. However, it is shown in \cite{pawar2019deep} and \cite{gin2019deep} that learning the residual instead of the next state often improves the accuracy. Hence, we consider the case
\begin{align}
u_l^{n+1} = \Psi(u^{n}) = u_l^{n} + R(u_l^{n}),
\end{align}
where $R$ is being approximated by a neural network. This practically makes $\Psi$ what is often referred as a residual neural network.
\subsubsection*{Incorporating Memory}
In \cite{pawar2019deep} and \cite{gonzalez2018deep} the potential benefits of not only using the present state but also incorporating several previous time steps for the future predictions were shown. Therefore, we now consider
\begin{align}
u_l^{n+1} =\Psi(u_l^{n},u^{n-1},\ldots,u_l^{n-\xi}) = u_l^{n} + R(u_l^{n},u_l^{n-1},\ldots,u_l^{n-\xi}),
\end{align}
where $\xi$ is the number of previous states included as input into the residual computation by the NN. The idea of incorporating several previous timesteps can loosely be compared to linear multistep methods where the order of approximation can be increased by using several previous steps \cite{leveque2007finite}. In contrast to linear multistep methods, NNs incorporate the previous time steps in a nonlinear fashion.
We consider two different types of networks in this paper: LSTM, and the CCNN. The two types of neural networks take varying computational time to train, have varying numbers of parameters, and vary in regards to how they interpret memory. In the appendix, there is a short description of the two types. Regarding the CCNNs, there are different ways to include memory. In this paper, we have chosen to include memory in shape of adding more layers, see Figure \ref{fig:CCNN_types} for examples for $\xi=8$, $\xi=6$, $\xi=4$, and $\xi=2$.
\begin{figure}
\centering
\includegraphics[width=.9\linewidth]{figures/CCNN_types.eps}
\caption{CCNN network architectures for varying memory.}
\label{fig:CCNN_types}
\end{figure}
\subsubsection*{Parameterized Dynamics}
We aim to simulate parameterized trajectories of the latent dynamics. Hence, we need to incorporate the parameters as input to the residual computation. For now, we consider constant parameters, but note that it should be possible to incorporate time-dependent parameters. For this reason, the parameters do not need to be part of the memory aware section of the network. We propose a parallel architecture consisting of two branches combining into one: One branch interpreting the last $\xi$ states and one branch processing the parameters. The two branches then connect and provide one final prediction for the residual. Having a single neural network incorporating the previous states and the parameters enables simultaneous training of the two branches. See Figure \ref{fig:network_parallel} for an illustration of the network structure. This ensures that the learned latent features from both branches are optimal with respect to predicting the next state. This is in contrast to what is done in \cite{xu2019multi} where the memory and the parameters are incorporated into two completely separate networks.
\begin{figure}
\centering
\includegraphics[width=0.75\linewidth]{figures/full_NN_model.eps}
\caption{Illustration of the parallel neural network structure. The "Encoding previous timesteps" part is visualized using the CCNN, but it should be noted that an LSTM network (or any suitable time series encoder) could be put in its place. }
\label{fig:network_parallel}
\end{figure}
For the parameter branch we simply make use of a dense FFNN. There is no immediate reason to believe that more complicated architectures are necessary, since we are not dealing with time-dependent nor high-dimensional or continuously spatially varying input. Note, however, that there is no reason to believe that this methodology will not work if the FF network in the parameter branch is replaced with a memory aware network in more advanced settings.
The training of the full time-evolution network is done by minimizing the loss function
\begin{align}
L(u_l,\mu;\theta) = \frac{1}{N_{train}N_{t}}\sum_{i=1}^{N_{train}}\sum_{n=\xi}^{N_T} \norm{u^{n+1}(\mu_i)-\Psi(u_l^{n}(\mu_i),\ldots,u_l^{n-\xi}(\mu_i),\mu_i;\theta)}^2_2,
\end{align}
with respect to the NN parameters $\theta$.
\subsubsection*{Imposing Stability and Generalization}
It is well-known that NNs do not necessarily generalize well beyond the training data without some kind of regularization. Combining that with the general risk of having instability in discrete dynamical systems makes it crucial to address these problems during the training.
The arguably most common technique is to add $L^1$- or $L^2$-regularization to the loss function. Furthermore, specifically for dynamical systems, it has been shown in \cite{erichson2019physics} and \cite{pan2018long} that regularizing the eigenvalues of the Jacobian of the dynamics with respect to the state variable, $D_u \Psi$, does improve long term predictions. In short this is related to linear and Lyapunov stability analysis of dynamical systems, which deals with the analysis of when a system is unstable. Hence, we propose adding the term $\norm{D_u \Psi}_2$, which is the matrix 2-norm, i.e. the spectral radius of the Jacobian of $\Psi$, to the loss function. In practice, by utilizing the relation
\begin{align}
\norm{D_u \Psi}_2 \leq \norm{D_u \Psi}_F,
\end{align}
we instead add the computationally much cheaper Frobenius norm.
It can empirically be shown that the long term predictions are significantly better if the network takes several steps at a time instead of a single one. Hence, we modify the output of the NN to
\begin{align}
R\left(u_l^{n}(\mu),\ldots,u_l^{n-\xi}(\mu),\mu;\theta\right) = \left[R_1, R_2, \ldots, R_\zeta\right]^T,
\end{align}
which gives future predictions,
\begin{align}
\begin{bmatrix}
u_l^{n+1}(\mu) \\ \vdots\\ u_l^{n+\zeta}(\mu)
\end{bmatrix}
= \begin{bmatrix}
\Psi_1\left(u_l^{n}(\mu),\ldots,u_l^{n-\xi}(\mu),\mu;\theta\right) \\ \vdots\\ \Psi_{\zeta}\left(u_l^{n}(\mu),\ldots,u_l^{n-\xi}(\mu),\mu;\theta\right)
\end{bmatrix} =
u_l^{n}(\mu) +
\begin{bmatrix}
R_1\left(u_l^{n}(\mu),\ldots,u_l^{n-\xi}(\mu),\mu;\theta\right) \\ \vdots\\ R_{\zeta}\left(u^{n}(\mu),\ldots,u_l^{n-\xi}(\mu),\mu;\theta\right)
\end{bmatrix}
\end{align}
Empirically we see that this modification keeps the prediction from exploding for longer time and it reduces spurious oscillations.
The final loss function for the dynamics NN is given by:
\begin{align}
\begin{split}
L(u,\mu;\theta) &=\frac{1}{N_{train}N_{t}} \sum_{i=1}^{N_{train}}\sum_{n=\xi}^{N_T} \norm{\sum_{k=1}^\zeta \left[u_l^{n+k}(\mu_i)- \Psi_k\left(u_l^{n}(\mu_i),\ldots,u_l^{n-\xi}(\mu_i),\mu_i;\theta\right)\right]}_2^2\\
&+ \underbrace{\beta_1 \norm{\theta}_2^2}_{\text{Weight decay}} + \underbrace{\beta_2 \norm{D_u R}_F}_{\text{Jacobian regularization}},
\end{split}
\end{align}
\subsection{The Complete Scheme}
Putting it all together, we have a scheme subdivided into two parts that are trained independently: The CAE, and the time evolution. The whole process is divided into an online phase and an offline phase.
In the offline phase the CAE is trained on a series of high-fidelity snapshots in order to identify a nonlinear reduced trial manifold. Then the CAE is used to reduce the high-fidelity snapshots to the latent space. The latent space trajectories are then used to train the time evolution NN. The training of the two neural networks is visualized in Figure \ref{fig:offline_stage} and outlined in Algorithm \ref{algo:offline_train}. Note that in steps 3 and 5, where the autoencoder and the time evolution network, respectively, are being trained, the considerations mentioned in Appendix \ref{appendix:neural_network} have to be included. Here, we refer to early-stopping, multiple-initialization, choice of optimizer, etc. In Algorithm \ref{algo:offline_tuning} an algorithm to automatically choose the latent dimension, number of training trajectories, memory, and future steps per iteration is presented. Note that this is a very simple approach to tune the network. More advanced methods such as Bayesian optimization or reinforcement learning could be utilized here. Furthermore, it is worth noting that we can, assuming no time constraints, generate as many training samples as necessary.
In the online phase the first $\xi$ time steps of the state, computed with a high-fidelity scheme for a given parameter realization $\mu$, are projected onto the latent space using the encoder part of the CAE. From there, the time evolution NN computes the parameterized latent space trajectories iteratively. The latent space trajectories are then transformed to the high-fidelity space using the decoder of the CAE. The online stage is visualized in Figure \ref{fig:online_stage} and described in pseudo code in Algorithm \ref{algo:online_train}.
\begin{figure}
\centering
\begin{subfigure}{1.\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/offline_stage.eps}
\caption{Illustration of the offline stage.}
\label{fig:offline_stage}
\end{subfigure} \\
\begin{subfigure}{1.\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/online_stage.eps}
\caption{Illustration of the online stage.}
\label{fig:online_stage}
\end{subfigure}
\caption{}
\label{fig:stages}
\end{figure}
\begin{algorithm}
\caption{Offline Stage - Training}\label{algo:offline_train}
\KwIn{$N_l$, $\zeta$, $\xi$, $N_{train}$.}
\KwOut{Trained CAE, $\Phi$, and Time Evolution Network, $R$, and test error, E.}
Sample $N_{train}$ parameter samples from the parameter space. \\
Generate high-fidelity trajectories,
$$\left\{ u_h^{0}(\mu_1), \ldots, u_h^{N_t}(\mu_1) , \ldots , u_h^{0}(\mu_{N_{train}}) , \ldots, u_h^{N_t}(\mu_{N_{train}}) \right\}.$$ \\
Train CAE, $\Phi=\Phi_{dec}\circ\Phi_{enc}$, with latent space dimension $N_l$, by minimizing
\begin{align}
\arg \min_{\theta} \frac{1}{N_{train}N_t}\sum_{i=1}^{N_{train}} \sum_{n=0}^{N_t} \left|\left| u_n(\mu_i) - \Phi(u_n(\mu_i);\theta) \right|\right|^2.
\end{align} \\
Encode high-fidelity trajectories to get latent state space trajectories
$$\left\{ \Phi_{enc}\left(u_h^{0}(\mu_1)\right), \ldots, \Phi_{enc}\left(u_h^{N_t}(\mu_1)\right) , \ldots , \Phi_{enc}\left(u_h^{0}(\mu_{N_{train}})\right) , \ldots, \Phi_{enc}\left(u_h^{N_t}(\mu_{N_{train}})\right) \right\}.$$ \\
Train time evolution network, $R$, to take the last $\zeta$ states and output the residuals for the next $\xi$ states, by minimizing
\begin{align*}
\frac{1}{N_{train}N_{t}} \sum_{i=1}^{N_{train}}\sum_{n=\xi}^{N_T} \norm{\sum_{k=1}^\zeta \left[u_l^{n+k}(\mu_i)- \Psi_k\left(u_l^{n}(\mu_i),\ldots,u_l^{n-\xi}(\mu_i),\mu_i;\theta\right)\right]}_2^2
+ \beta_1 \norm{\theta}_2^2+ \beta_2 \norm{D_u R}_F,
\end{align*} \\
Estimate error on a test set.
\end{algorithm}
\begin{algorithm}
\caption{Offline Stage - Tuning} \label{algo:offline_tuning}
\KwIn{Desired test error, $E^*$}
\KwOut{Optimal latent dimension, $N_l$, $\zeta$, $\xi$, and $N_{train}$.}
Initialize $E = \infty$, $N_l=1$, $\zeta=0$, $\xi=1$, $N_{train}$. \\
\While{$E^*<E$}
{
Train $\Phi$ and $R$ and compute test error, $E$, using Algorithm \ref{algo:offline_train}. \\
Update $N_l$, $\zeta$, $\xi$, $N_{train}$ and according to some update rule.
}
\end{algorithm}
\begin{algorithm}
\caption{Online Stage} \label{algo:online_train}
\KwIn{$\Phi_{dec}$, $R$, $\mu$, $u^0_h(\mu),\ldots, u^{\xi}_h(\mu)$}
\KwOut{Approximated trajectory in high-fidelity space.}
Encode the initial $\xi$ high-fidelity states,
\begin{align*}
(u^0_l(\mu),\ldots, u^{\xi}_l(\mu)) = \left(\Phi_{enc}\left(u^0_h(\mu)\right), \ldots, \Phi_{enc}\left(u^\xi_h(\mu)\right) \right)
\end{align*} \\
Compute approximated latent trajectory by iterating,
\begin{align*}
\left(\tilde{u}^{n}_l(\mu),\ldots,\tilde{u}^{n+\zeta}_l(\mu)\right) = \tilde{u}^{n}_l(\mu) + R\left(\tilde{u}^{n}(\mu),\ldots,\tilde{u}^{n-\xi}(\mu),\mu;\theta\right),
\end{align*}
until desired end time has been reached. \\
Decode approximated latent space trajectories to high-fidelity space:
\begin{align*}
\left\{ \tilde{u}_h^{0}(\mu), \ldots, \tilde{u}_h^{N_t}(\mu) \right\} =
\left\{ \Phi_{dec}\left(\tilde{u}_l^{0}(\mu)\right), \ldots, \Phi_{dec}\left(\tilde{u}_l^{N_t}(\mu)\right) \right\}.
\end{align*}
\end{algorithm}
\end{document}
\section{Artificial Neural Networks} \label{appendix:neural_network}
\subsection{Feedforward Neural Networks}
The arguably most common ANN architecture is the feedforward neural network (FNN). An FNN can be considered a function, $G:\mathbb{R}^{N_i}\rightarrow \mathbb{R}^{N_o}$, consisting of a series of affine transformations, $T_i$, followed by an element-wise (nonlinear) activation function, $\sigma_i$:
\begin{align} \label{FNN}
G(x;\theta) = \sigma_L\circ T_L\circ \ldots \circ \sigma_1 \circ T_1 (x).
\end{align}
The combination of an afine transformation followed by the activation is called a neuron. The afine transformation can be written as $T_i(x)=W_ix+b_i$, where $W\in \mathbb{R}^{M_i\times M_{i-1}}$ and $b\in\mathbb{R}^{M_i}$. We call $W_i$ the weight matrix, $b_i$ the bias vector, and $M_i$ the number of neurons in layer $i$, and $L$ the number of layers. \eqref{FNN} is conveniently visualized as a network of neurons. We will refer to the set of parameters as $\theta = \left\{W_1,b_1, \ldots,W_L,b_L\right\}$.
\begin{figure}
\centering
\includegraphics[width=.35\textwidth]{figures/NN_dense.eps}
\caption{Visualization of a feedforward densely connected neural network.}
\label{fig:NN_dense}
\end{figure}
In supervised learning one tries to approximate a function by an ANN, typically done by minimizing the empirical risk w.r.t. the parameters $\theta$:
\begin{align} \label{minimization}
\theta^* = \text{arg}\min_\theta \mathbb{E}_{(x,y)\sim \mathcal{P}_{\text{data}}} \left[ \mathcal{L}(G(x;\theta),y) \right],
\end{align}
where $\mathcal{P}_{\text{data}}$ is the distribution generating the data and $\mathcal{L}$ is a chosen loss function measuring the discrepancy between the predicted output and the target. For regression type problems the mean squared error (MSE) is the most common choice. However, especially for physics-informed machine learning the physics is often incorporated in the loss function in shape of extra term \cite{raissi2017physics, erichson2019physics}.
Computing \eqref{minimization}, i.e. training the ANN, is mostly done using stochastic gradient descent (SGD) or a variants such as ADAM \cite{kingma2014adam}.
\subsection{Convolutional Neural Networks}
Convolutional neural networks (CNNs) gained attention due their great performance in image recognition. The general idea is to utilize local properties of the data instead of only considering global properties. This is done by having local connections and shared weights in the neural networks. These properties are not only great for detecting patterns in data but it also makes it possible to do computations on very high-dimensional data.
A convolutional layer is effectively a feature map where each unit in the layer is connected to a local patch of the previous layer through a filter bank and an activation function. A feature map at layer $l$ is a tensor, $H^l\in\mathbb{R}^{N_{chan}^l\times N_1^l \times N_2^l}$, where $H^l_{i,j,k}$ is a unit at channel $i$, row $j$, and column $k$. The filter bank at layer $l$ is a 4-dimensional tensor, $F^l\in \mathbb{R}^{N_{filter}^l\times N_{chan}^{l-1}\times k_1 \times k_2}$, where $F^l_{i,j,m,n}$ connects a unit in channel $i$ of the output and channel $j$ of the input with $m$ and $n$ being the offset of rows and columns respectively. $N_{filter}^l$ denotes the number of filters in the feature bank in layer $l$ and $k_1$ and $k_2$ denotes the kernel size. The convolution operation between a feature map and a filter bank is given by
\begin{align}
H^l_{i,j,k} = \sigma_l\left(\sum_{r=1}^{N_{chan}^{l-1}}\sum_{m=1}^{N_1^{l-1}}\sum_{n=1}^{N_2^{l-1}}
H^{l-1}_{r,(j-1)s+m,(k-1)s+n} F_{i,r,m,k}^{l} + B^l_{i,j,k} \right),
\end{align}
where $B^l_{i,j,k}$ is a bias term and $\sigma_l$ is an activation function applied element-wise. $s$ denotes the stride and effectively downsamples the feature map between layers. The filters, $F_{i,r,m,k}^{l}$, and biases, $B^l_{i,j,k}$, are the learnable parameters while the kernel sizes, $k_1, k_2$, the stride, $s$, and the number of filters, $N_{filter}^l$, are chosen. Often these are subject case specific objectives or hyperparameter optimization.
\subsection{Causal Convolutional Neural Networks}
As the name suggests, causal convolutional neural networks (CCNNS) are related to convolutional neural networks. CCNNs are sometimes referred to as temporal convolutional neural networks, but in this paper we use the term CCNN.
CCNNs are used for encoding time series data with the purpose of forecasting or classification. The general idea is to use 1-dimensional convolutions on time series data. In the multivariate case the multiple dimensions are interpreted as channels. The term causal refers to the fact that the filter banks are only convolved with the current and previous time steps, thus establishing a causal relationship between the past the future.
\subsection{Recurrent Neural Networks and Long Short-Term Memory}
A recurrent neural network (RNN) is an alternative to CCNNs for interpreting time series data. The general idea is to process and retain information from previous time steps in an efficient manner. In this paper, we solely focus on a specific RNN called long short-term memory (LSTM) \cite{hochreiter1997long}. For an input consisting of several previous time steps, $x^n$, an LSTM layer consists of four components \cite{gonzalez2018deep}: An input gate:
\begin{align}
i^{n+1} = \sigma\left(W_i x^{n} + b_i \right),
\end{align}
a forget gate:
\begin{align}
f^{n+1} = \sigma\left(W_f x^{n} + b_f \right),
\end{align}
an output gate:
\begin{align}
o^{n+1} = \sigma\left(W_o x^{n} + b_o \right),
\end{align}
and a cell state
\begin{align}
c^{n+1} = i \odot c^{n} + i^n \odot \tanh\left(W_c x^{n} + b_c \right).
\end{align}
The prediction is then given by
\begin{align}
x^{n+1} = o^n \odot \tanh\left(c^n\right).
\end{align}
$W_i,b_i,W_f,b_f,W_o,b_o,W_c,b_c$ are the trainable weight matrices and bias vectors, and $\odot$ is the Hadamard product. Ideally, the input gate identifies what information to be passed to the from the cell state, the forget what to be dropped, and the output gate decides what to be passed to the final prediction.
\end{document}
\section{Conclusion} \label{Conclusion}
We presented a novel deep learning approach to non-intrusive reduced order modeling for parameterized time-dependent PDEs using CAEs for dimensionality reduction and CCNNs and LSTMs combined with FFNNs for time evolution. This approach was demonstrated on various test cases and was shown to perform well in the online phase, showcasing the potential of using deep learning based ROMs for different physical phenomena.
Regarding dimensionality reduction, a discussion and comparison of linear and nonlinear methods was presented with POD and CAEs as the focus points. The discussion focused on why a nonlinear approach has the potential to outperform a linear approach.
For time stepping, the general idea was to encode the previous states and the parameters separately in parallel and then combine the encoded data to make a final prediction using a FFNN. The two encoding NNs, as well as the final prediction NN, constitute a single network, meaning everything is trained simultaneously. This ensures that both the memory and parameters are encoded in relation to one another. Furthermore, various methods to ensure generalization, stability, and precision were discussed and tested.
In all the test cases errors were found to be between $10^{-4}$ and $10^{-2}$ for all time steps. These are in many cases acceptable errors considering the significant speed-ups. Furthermore, we saw that the model performs well on both rather high-dimensional trial manifolds (the heat equation), purely advection dominated problems (the advection equation) and nonlinear multiple vector field computations (the lid driven cavity) and it delivered significant online speed-ups without sacrificing accuracy.
In summary, the contributions in this work include a nonlinear dimensionality reduction scheme using convolutional autoencoders, a novel parallel neural network architecture for parameterized time-stepping using CCNNs and LSTMs, and a discussion on different approaches to achieve stability and generalization for neural network-based time-stepping.
In the future the methodology will be tested on more advanced PDE problems. By advanced problems, we are both referring to increasing nonlinearity, higher dimensions, and multi-query problems such as uncertainty quantification, model predictive control, and data assimilation.
Besides considering other use cases one could work on improving the NN architecture and training by, e.g. incorporating the physics in the training \cite{erichson2019physics, raissi2017physics}, and use Bayesian optimization \cite{archetti2019bayesian} or reinforcement learning \cite{sugiyama2015statistical} to ensure effective snapshot generation. Furthermore, with the amount of hyperparameters ($\xi$, $\beta_1$, $\beta_2$, number of layers and neurons, etc.) the task of hyperparameter tuning is not trivial and could be solved more effectively using modern approaches.
\section*{Acknowledgement}
This work is supported by the Dutch National Science Foundation NWO under the grant number 629.002.213.
which is a cooperative projects with IISC Bangalore and Shell research as project partners.
The authors furthermore acknowledge fruitful discussions with Dr. B. Sanderse.
\begin{comment}
In general, this work should be considered a proof a concept. There is still much work to be done on the framework. Firstly, it is necessary to test how difficult problems can be handled. How large Reynolds numbers can be considered before the model does not compute sufficient details? How high-dimensional can the trial manifold be? These and many other questions are worth asking.
Another issue worth considering is the training. For the heat equation we used up to 2500 training trajectories, for the advection equation up to 120, and for the lid driven cavity up to 80. It is not difficult to imagine problems where more trajectories are needed or cases where even generating training trajectories is so expensive that it quickly becomes infeasible. To accommodate these issues one could look into new areas such as active learning \cite{settles2009active}, reinforcement learning \cite{sugiyama2015statistical}, Bayesian optimization \cite{archetti2019bayesian}, or adaptive sampling. All of these approaches aim to derive optimal sampling strategies which could minimize the number of trajectories necessary.
A much more fundamental issue is the lack of theoretical understanding of neural networks. Here we are especially referring to lack of interpretability, unknown uncertainties in the model, no clear strategy for hyperparameter tuning, etc. While this is a major issue, it is worth noting that much work is currently being done on this topic \cite{kutyniok2019theoretical}.
\end{comment}
\end{document}
\section{Dimensionality Reduction} \label{Dimensionality Reduction}
The fundamental idea of dimensionality reduction is that the minimal number of variables necessary to represent the state, also called the intrinsic dimension, of the dynamical system is low compared to the dimension of the high-fidelity model. However, identifying an optimal low-dimensional representation is, in general, not a trivial task. In this section we will give a brief overview of linear dimensionality reduction, particularly, the well-known proper orthogonal decomposition (POD). Then, from the linear outset, we will describe the more general case of nonlinear dimensionality reduction.
In general, for both linear and nonlinear dimensionality reduction, we assume that a state, $u_h^n(\mu)\in\mathbb{R}^{N_h}$, can be approximated,
\begin{align} \label{general_dim_reduction}
u_h^n(\mu) \approx \Phi(u_h^n) = \Phi_{dec} \circ \Phi_{enc} (u_h^n(\mu)),
\end{align}
where $\Phi_{enc} (u_h)\in \mathbb{R}^{N_l}$ with $N_l \ll N_h$. $\Phi_{enc}$ is referred to as the \textit{encoder} and $\Phi_{dec}$ the \textit{decoder}. The encoder transforms the high-dimensional input to a \textit{latent space} of low dimension and the decoder transforms the latent variable back to the high-fidelity space. The latent space is often denoted the \textit{reduced trial manifold}. The state at time step $n$ in the latent space is denoted $u_l^n(\mu) = \Phi_{enc}(u_h^n(\mu))$, and will be referred to as the latent state.
Ideally, $\Phi$ reconstructs the input perfectly for any given parameters and time step. However, that is, in general, not possible. The precision of the reconstruction is heavily dependent on the dimension of the latent space, as this determines the amount of compression applied. One computes $\Phi$ by choosing a latent dimension, $N_l$, and then solving the minimization problem,
\begin{align} \label{reduction_minimization}
\Phi^* = \text{arg}\min_{\Phi} \sqrt{\int_{\mu\in P} \left[\sum_{n=0}^{N_t}\norm{u_h^n(\mu) - \Phi (u_h^n(\mu))}^2_2 \right] d\mu},
\end{align}
where $\norm{\cdot}_2$ denotes the $l^2$-norm. Theoretically, the reconstruction error should decrease when $N_l$ is increased until the intrinsic dimension of the problem is reached. From thereon, increasing the dimension of the latent space should have very little effect on the reconstruction error.
There are many ways of solving \eqref{reduction_minimization} \cite{quarteroni2015reduced}. In this paper we focus on a data-driven approach, sometimes referred to as the \textit{method of snapshots}. A snapshot is a high-fidelity solution for a given parameter realization at a certain time. The idea of this approach is to make $N_{train}$ samples from the parameter space and then compute a series of $N_{T}+1$ snapshots, i.e. trajectories, per parameter sample,
\begin{align} \label{M_train}
M_{N_{train},h,\delta t} = \left\{ u_h^{0}(\mu_1), \ldots, u_h^{N_t}(\mu_1) ,u_h^{0}(\mu_2), \ldots, u_h^{N_t}(\mu_2), \ldots , u_h^{0}(\mu_{N_{train}}) , \ldots, u_h^{N_t}(\mu_{N_{train}}) \right\}.
\end{align}
Then \eqref{reduction_minimization} is rewritten into an empirical minimization problem:
\begin{align}\label{reduction_minimization_empirical}
\Phi^* = \text{arg}\min_{\Phi} \sqrt{\sum_{i=1}^{N_{train}} \sum_{n=0}^{N_t} \left|\left| u_h^{n}(\mu_i) - \Phi \left(u_h^{n}(\mu_i)\right) \right|\right|^2_2}.
\end{align}
The idea is that sampling a finite number of discrete trajectories a sufficient number of times yields a good enough representation of the time-discrete high-fidelity solution manifold. It should be noted that computing \eqref{M_train} is potentially very expensive and even infeasible in some cases.
When a reduction scheme is computed, one can then compute the parameterized trajectories in the latent space, by
\begin{align} \label{time_discrete_problem_reduced}
u^{n+1}_l(\mu) = F_{l,\delta t}(u^{n}_l;\mu), \quad u^{0}_l(\mu) = \Phi_{enc}\left(u^{0}_h(\mu)\right),
\end{align}
from which the trajectories in the high-fidelity space can be recovered by $u^{n}_h(\mu) = \Phi_{dec}\left(u^{n}_l(\mu)\right)$. $F_{l,\delta t}$ can be derived in many ways and much time and effort have been put into deriving optimal latent dynamics.
\subsection{Linear Dimensionality Reduction}
In linear dimensionality reduction the strategy is to find a reduced linear trial manifold of low dimension. Since the sought manifold is linear it can be written as the column space, $\text{Col}(V)$ of some matrix, $V\in\mathbb{R}^{N_h\times N_l}$. The column space is the space spanned by the columns of the matrix $V$. From the orthogonal projection theorem, it can be shown that the optimal projection onto a latent linear space is given by
\begin{align} \label{PCA}
u_h \approx VV^T u_h.
\end{align}
Hence, this is a special case of \eqref{general_dim_reduction} where
\begin{align}
\Phi = VV^T, \quad \Phi_{enc} = V^T, \quad \Phi_{dec} = V.
\end{align}
This simplification reduces \eqref{reduction_minimization_empirical} to
\begin{align}\label{reduction_minimization_empirical_linear}
V^* = \text{arg}\min_{V} \sqrt{\sum_{i=1}^{N_{train}} \sum_{n=0}^{N_t} \left|\left| u_h^{n}(\mu_i) - VV^T u_h^{n}(\mu_i) \right|\right|^2_2},
\end{align}
often accompanied by the constraint that the columns of $V$ are orthogonal, $V^T V=0$. It can be shown that \eqref{reduction_minimization_empirical_linear} has an exact solution \cite{quarteroni2015reduced}. By collecting the snapshots in a \textit{snapshot matrix},
\begin{align}
S = \left[ u_h^{0}(\mu_1) \: | \: \ldots \: | \: u_h^{N_t}(\mu_1) \: | \: \ldots \: | \: u_h^{0}(\mu_{N_{train}}) \:| \: \ldots \: | \: u_h^{N_t}(\mu_{N_{train}}) \right],
\end{align}
one can show that the optimal $V\in\mathbb{R}^{N_h\times N_l}$ is given by the first $N_l$ \textit{left singular vectors}. The left singular vectors are computed through the singular value decomposition (SVD),
\begin{align}
S = U \Sigma Z^T,
\end{align}
where $U$ is a matrix whose columns are the left singular vectors, $Z$ is a matrix whose columns are the right singular vectors, and $\Sigma$ is a diagonal matrix with the singular values on the diagonal. $V$ is then chosen to be the first $N_l$ columns of $U$. This method of obtaining $V$ is the \textit{proper orthogonal decomposition} (POD) \cite{quarteroni2015reduced}, also denoted \textit{principal component analysis} (PCA) \cite{friedman2001elements}.
To obtain $F_{l,\delta t}$ a Petrov Galerkin projection is often performed, which yields
\begin{align}
F_{l,\delta t}(u^{n}_l;\mu) = W^T F_{h,\delta t}(Vu^{n}_l;\mu).
\end{align}
When $W=V$ it is denoted the Galerkin projection. This approach is \textit{intrusive} which means that direct access to the model, $F_{h,\delta t}$, is required. Furthermore, in the online phase a transformation between the latent space and the high-fidelity space must be performed in each time step in order to be able to evaluate $F_{h,\delta t}(Vu^{n}_l;\mu)$, which slows down the computations. Various methods to circumvent that problem, such the discrete empirical interpolation methods \cite{quarteroni2015reduced}, already exists. However, in recent years there are many studies exploring approximating $F_{l,\delta t}$ with neural networks \cite{pawar2019deep, mucke2019reduced}.
While there are many advantages of a linear reduction scheme, such as the explicit solution to \eqref{reduction_minimization_empirical_linear}, there are, indeed, disadvantages as well. A significant problem is the restriction to a linear trial manifold. The optimal trial manifold, i.e. the trial manifold of the intrinsic dimension, is rarely linear. Especially for advection-dominated and nonlinear problems, it has been shown that a linear reduced approximation does not necessarily lead to significant speed-ups.
\subsection{Nonlinear Dimensionality Reduction}
The extension from linear to nonlinear dimensionality reduction comes naturally and addresses several of the drawbacks of linear dimensionality reduction. The fundamental difference is that we remove the constraint that the latent space has to be a linear manifold. Due to this generalization, we cannot write the projection operator as the matrix product $VV^T$ anymore, but instead we must use the general form in \eqref{general_dim_reduction}, where $\Phi_{enc}$ and $\Phi_{dec}$ can be any type of nonlinear functions. This gives rise to a major difference in solving \eqref{reduction_minimization_empirical}, since no general exact solution exists and therefore \eqref{reduction_minimization_empirical} will be solved numerically.
Even though extra approximation steps have to be introduced in the nonlinear case, the potential gains will, in some cases, outweigh this hurdle. This is due to the fact that with a nonlinear reduction scheme it is theoretically possible to reduce the high-fidelity space down to its intrinsic dimension, $N_P + 1$. However, this relies on the choice of $\Phi$ and the minimization scheme.
A common method for nonlinear dimensionality reduction from the machine learning communities is, among others, kernel PCA. Here, the nonlinear manifold is embedded into a linear space, often of higher dimension, using a predefined nonlinear mapping, $\psi:\mathbb{R}^{N_h}\rightarrow \mathbb{R}^{N_k}$, $N_k>N_h$. From thereon, a linear PCA is performed on the high-dimensional linear data. In order to speed up computations the so-called kernel trick is typically invoked. Utilizing that the nonlinear embedding induces a kernel, $K = k(\psi(x),\psi(y)) = \psi(x)^T\psi(y)$, one can compute the low-dimensional basis without explicitly transforming the data and perform PCA in the high-dimensional space. For more details see \cite{vidal2005generalized}.
This approach works well in many cases but suffers from one crucial downside: Choosing the nonlinear mapping, $\psi$, or the kernel, $K$, is far from trivial. There exist no clear guiding principles that work across several cases.
\subsubsection{Autoencoders}
To overcome the problems of other nonlinear dimensionality reduction methods, such as kernel PCA and DEIM, we present Autoencoders (AEs). AEs are a type of NN. For a brief introduction to NNs and the terminology used in this paper, see \ref{appendix:neural_network}. In the context of dimensionality reduction one can interpret an AE as a kernel PCA where the kernel is learned during the training process. Thus, one circumvents the problem of choosing a suitable kernel. Note that this interpretation is merely presented in order to give an intuition of AEs in context of other methods. To further explain the connections between AEs and PCA it is worth noting that a single hidden layer AE with linear activation functions is equivalent to PCA. A single hidden layered AE without bias terms can be written as
\begin{align} \label{one_layer_AE}
\Phi(u_h^n(\mu);\theta) = T_2\circ T_1(x) = W_2 W_1 u_h^n(\mu),
\end{align}
where $\theta=\left\{W_1,W_2\right\}$, $T_1:\mathbb{R}^{N_h}\rightarrow \mathbb{R}^{N_l}$, and $T_2:\mathbb{R}^{N_l}\rightarrow \mathbb{R}^{N_h}$ are linear maps and $W_1$ and $W_2$ are matrices. Typically, the mean squared error is chosen as the loss function for AEs, which gives the following minimization problem for the single hidden layer AE:
\begin{align}\label{AE_single_layer_loss}
\arg \min_{W_1,W_2} \frac{1}{N_{train}N_t}\sum_{i=1}^{N_{train}} \sum_{n=0}^{N_t} \left|\left| u_h^n(\mu_i) - W_2 W_1 u_h^n(\mu_i) \right|\right|^2.
\end{align}
Hence, training a single layer AE is equivalent to solving the PCA minimization problem, eq. \eqref{reduction_minimization_empirical_linear}, without the orthogonality constraint. Conclusively, PCA, eq. \eqref{PCA}, can be considered a special case of an AE.
By dividing the AE into the encoder and decoder parts and allowing an arbitrary number of layers and nonlinear activation functions, it is easier to understand the similarities to linear dimensionality reduction and why AEs have the potential to perform significantly better. Consider the encoder part with a linear activation in the final layer,
\begin{align}
\Phi_{enc}(u_h^n(\mu);\theta) =T_L\circ \underbrace{\sigma_{L-1} \circ T_{L-1} \ldots \circ \sigma_1 \circ T_1 (u_h^n(\mu))}_{\psi_{enc}(u_h^n(\mu);\theta)} = W_L\psi_{enc}(u_h^n(\mu);\theta) = z,
\end{align}
where $\psi_{enc}(x;\theta)=(\psi_{enc}^1(x;\theta),\ldots, \psi_{enc}^{N_e}(x;\theta))\in \mathbb{R}^{N_e}$ and $W_L\in\mathbb{R}^{N_e \times N_l}$. For convenience we ignore bias terms. We see that this corresponds to a nonlinear embedding onto $\mathbb{R}^{N_e}$ and then a projection onto the space spanned by the vectors ${\psi_{enc}^1, \ldots, \psi_{enc}^{N_e}}$. This is quite similar to the idea behind kernel PCA. The difference is that in the AE framework we adjust the nonlinear embedding in the training instead of defining it beforehand. The decoder part is similarly written as
\begin{align}
\Phi_{dec}(z;\theta) =T_L\circ \underbrace{\sigma_{L-1} \circ T_{L-1} \ldots \circ \sigma_1 \circ T_1 (z)}_{\psi_{dec}(z)} = W_L\psi_{dec}(z;\theta) = \tilde{u}_h^n(\mu),
\end{align}
where $W_L\in\mathbb{R}^{N_d\times N_h}$.
Note that this is merely a brief discussion of the topic of AEs aiming to give an intuitive understanding. For more details see \cite{yu2019understanding}.
\subsubsection{Convolutional Autoencoders}
Convolutional autoencoders (CAEs) are a special type of AEs utilizing convolutional layers instead of dense layers. A brief introduction to convolutional neural networks (CNNs) can be found in \ref{appendix:neural_network}. It can be shown that dense and convolutional neural networks are equivalent regarding approximation rates \cite{petersen2020equivalence}, which means that theoretical approximation results for dense NNs translate directly to CNNs. For practical purposes, however, convolutional layers are often to be preferred due to especially the following two properties:
\begin{itemize}
\item \textit{Shared weights}, which in practice makes the affine transformations very sparse and enables location invariant feature detection.
\item \textit{Local connections}, which utilizes that spatial nodes close to each other are highly correlated.
\end{itemize}
An additional advantage is that it is straightforward to handle multiple spatially distributed states. These occur in coupled PDEs such as the Navier Stokes equations where one is dealing with both the $x-$, $y-$ and $z-$components of the velocity field as well as the pressure field. In the framework of CAEs these can all be included by interpreting them as different channels. This enables the possibility of including multiple spatial states without increasing the number of weights in the neural network significantly. The connection between PDEs and CNNs has already been made, see e.g. \cite{ruthotto2019deep}.
In Figure \ref{fig:conv_AE} one sees an illustration of a CAE. The encoding consists of a series of convolutional layers with an increasing number of filters and decreasing dimension, effectively down sampling the number of degrees of freedom, followed by dense layers. Similarly, the decoding consists of series of dense layers followed by a series of deconvolutional layers with a decreasing number of filters and increasing dimension, effectively up sampling. The down sampling is often performed by utilizing pooling layers or strides larger than one.
It is worth noting that computing the decoder, $\Phi_{dec}$, of a CAE in the training phase is effectively solving an inverse problem. Inverse problems are, in general, ill-posed and therefore require some sort of regularization. $L^2$-regularization, often referred to as weight decay, is frequently used, and results in the following minimization problem to solve:
\begin{align}\label{CAE_loss}
\arg \min_{\theta} \frac{1}{N_{train}N_t}\sum_{i=1}^{N_{train}} \sum_{n=0}^{N_t} \left|\left| u_h^n(\mu_i) - \Phi(u_h^n(\mu_i);\theta) \right|\right|^2 + \alpha \norm{\theta}_2^2,
\end{align}
where $\alpha$ is a hyperparameter to be tuned. Besides ensuring well-posedness the term also ensures generalization.
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{figures/conv_AE.eps}
\caption{Illustration of a convolutional autoencoder}
\label{fig:conv_AE}
\end{figure}
\end{document}
\section{Introduction}
Simulations based on first-principles models often form an essential element for understanding, designing, and optimizing problems in, for example, physics, engineering, chemistry, and economics. However, with an increasing complexity of the mathematical models under consideration, it is not always possible to achieve the desired fidelity of such simulations in a satisfactory time frame. This is especially the case when dealing with multi-query and/or real time problems as encountered in uncertainty quantification and model predictive control, where the computational model is typically parameterized.
There are several approaches to reduce the computation time bottleneck. The arguably most common ones include high-performance computations \cite{hager2010introduction}, high-order discretizations \cite{kopriva2009implementing}, iterative and/or multigrid methods \cite{saad2003iterative, trottenberg2000multigrid}, and reduced order modeling (ROM) \cite{quarteroni2015reduced}. High-performance computing may be costly; the improvements due to high-order discretization strongly depend on the smoothness of solutions at hand, and iterative methods are highly dependent on being able to identify suitable preconditioners. Furthermore, these approaches may suffer from the curse of dimensionality. ROM, a relatively recent research area, is an interesting alternative to the other approaches.
The ROM solution process is generally divided into two distinct stages \cite{quarteroni2015reduced}: A so-called ''offline stage'', in which the reduced model is derived, and an ''online stage'', where the reduced model is utilized and solved. Popular choices for the two stages are the proper orthogonal decomposition (POD) model definition, combined with a (Galerkin) projection procedure in the online stage \cite{quarteroni2015reduced, hesthaven2016certified}. Whereas this combination has shown important successes, it has also been shown that the POD and projection approaches perform worse in certain settings, such as for advection-dominated or nonlinear problems. Furthermore, projection-based methods are intrusive, as they require access to the underlying high-fidelity model. Nowadays, it is a reasonable assumption that an industrial model is not directly accessable, and therefore non-intrusive approaches, i.e. approaches that are only based on a series of snapshots of solutions, are increasingly interesting alternatives.
Machine learning has recently gained the attention from the scientific computing community due to great successes of artificial intelligence in various settings. Specifically Artificial Neural Networks (ANNs), often simply denoted neural networks (NNs), have shown remarkable results in tasks such as image analysis and speech recognition. Much of the success has been boosted further by the availability of open source software frameworks, such as PyTorch \cite{paszke2017automatic} and Tensorflow \cite{abadi2016tensorflow}, which has made implementation and training possible without expert knowledge and the availability of computation accelerating hardware, such as GPUs, has made training of very large models feasible. These recent advances have accelerated research in especially deep learning, i.e. multilayered NNs, which was not possible few years ago, resulting in many NN architectures specialized in certain tasks, such as time series forecasting and dimensionality reduction, that are able to interpret data with unprecedented accuracy.
Mathematically, NNs have many interesting properties, like universal function approximation \cite{cybenko1989approximation} and pattern recognition that require deep architectures. For these reasons, NNs have gained traction within the mathematics, numerical analysis, and engineering communities either as a replacement or as a supplement to conventional function approximation methods. For an overview of articles, prospects, and future challenges see e.g. \cite{lee2018basic, baker2019workshop, brunton2019machine}. In this paper, we will combine ROM and machine learning in both the offline and the online stages to showcase the potential of using these technologies on conventional problems from scientific computing.
Important work has already been done on the topic of NN-based ROM, which was typically based on Proper Orthogonal Decomposition (POD) for dimensionality reduction and feedforward neural networks (FFNNs) as a parametric map or as a time-stepping scheme \cite{mucke2019reduced,pawar2019deep,pan2018long}. For NN-based dimensionality reduction, see the work in \cite{lee2020model}, where convolutional autoencedoers (CAEs) are used while the time-stepping is done intrusively using multistep methods on a reduced model derived from a Galerkin projection procedure. In \cite{gonzalez2018deep}, autoencoders are utilized for dimensionality reduction and long-short term memory (LSTM) neural networks are used for non-parameterized time-stepping.
Closest to our work are the methodologies described in \cite{fresca2020comprehensive} and \cite{xu2019multi}. In both papers autoencoders are used for dimensionality reduction, and neural networks are used for parameterized time advancement in the latent space. The training of the different networks is done seperately in both papers however. In \cite{fresca2020comprehensive} the dynamics network takes in the parameters and a desired time for the state to be predicted. Hence, the ROM cannot be given an arbitrary initial state and advance in time from there. In \cite{xu2019multi} the time-stepping is further divided into two networks: a causal convolutional neural network (CCNN) is used for encoding the previous states, and a dense FFNN is used to predict the next state based on the encoded dynamics and parameters. Thus, the memory-aware and the parameter neural networks are trained independently from each other.
In our work, we present a non-intrusive framework, based on deep learning, for computing parameterized spatio-temporal dynamics. The resulting reduced order model is divided into two distinct stages: Firstly, a dimensionality reduction stage based on convolutional autoencoders (CAEs), and secondly a memory-aware NN stage for parameterized time stepping. This methodology utilizes the effectiveness of CAEs as nonlinear dimensionality reduction techniques for spatially distributed data. To discuss the advantages of using CAEs, we make a comparison to the widely used linear counterpart, POD. Specifically we show that POD is a special case of an autoencoder. Furthermore, we present a flexible neural network structure for time-stepping that takes into consideration previous states as well as parameters. The framework is quite general and allows for various types of neural network architectures, hence allowing the researcher to use state-of-the-art techniques that fit the problem at hand. We present and compare two modern time series forecasting architectures, Long Short-Term Memory (LSTM) networks \cite{hochreiter1997long}, and Causal Convolutional Neural Networks (CCNNs) \cite{oord2016wavenet}. Furthermore, we present and discuss a series of approaches to ensure stability and generalization of the time-stepping netowork.
To the best of our knowledge, there is no other work on deep learning-based ROM that is non-intrusive, uses CAEs for dimensionality reduction, has memory-aware and parameterized time-stepping, compares modern time series encoding architectures, and discusses practical approaches to ensure stability and generalization. The result is a flexible offline-online scheme that works for various physical phenomena and can easily be modified according to the specific problem at hand. This makes the presented approach suitable for multi-query problems as they occur in model predictive control, uncertainty quantification, and solving inverse problems.
The structure of the present paper is as follows. In Section \ref{Parameterized Time-Dependent PDEs} we present parameterized time-dependent PDEs. In Section \ref{Dimensionality Reduction} we discuss dimensionality reduction. Furthermore, we discuss how convolutional autoencoders are used for nonlinear dimensionality reduction. In Section \ref{Approximating Parameterized Time Evolution using Neural Networks} we present the parameterized memory-aware time-stepping neural network. In Section \ref{Results} we showcase the performance on three test problems: the linear heat equation, a linear advection equation, and the lid driven cavity incompressible Navier-Stokes problem.
\end{document}
\section{Parameterized Time-Dependent PDEs}\label{Parameterized Time-Dependent PDEs}
The model under consideration is of the form
\begin{align}\label{partial_diff}
\partial_t u(t,x;\mu) = F(t,x,u;\mu), \quad u(0,x;\mu) = u_0(x;\mu),
\end{align}
where $F$ is a (nonlinear) differential operator, $u:\mathbb{R}\times\mathbb{R}^{d}\times\mathbb{R}^{N_p}\rightarrow\mathbb{R}$ or $u:\mathbb{R}\times\mathbb{R}^{d}\times\mathbb{R}^{N_p}\rightarrow\mathbb{R}^d$, $t\in [0,T]$, and $x\in\mathbb{R}^{d}$. Equation \eqref{partial_diff} is a very general parameterized PDE. $\mu$ is to be considered a vector of parameters on which the solution depends. These parameters could be diffusion rate, Reynolds number, parameterize an initial or boundary condition, etc. For technical reasons, the parameter space $\mathcal{P}$ is chosen to be a compact subspace of $\mathbb{R}^{N_P}$ \cite{quarteroni2015reduced}.
Spatially discretizing \eqref{partial_diff}, using finite elements, finite volumes, finite differences \cite{quarteroni2009numerical}, gives the following finite-dimensional dynamical system,
\begin{align} \label{general_problem}
\partial_t u_h(t,\mu) = F_h(t,u_h(t,\mu);\mu), \quad u_h(0,\mu) = u^0_h(\mu),
\end{align}
$h$ defines the granularity of the discretization, i.e. grid size, number of elements, etc. We will not go into detail regarding these discretizations and it should be assumed that the discretized system is stable and converges to the exact solution when granularity is refined. $u_h(t,\mu)\in \mathbb{R}^{N_h}$ will be referred to as the high-fidelity or full-order solution.
The manifold of high-fidelity solutions, parameterized by time and the parameters, is called the spatial discrete solution manifold,
\begin{align}
M_{h} = \left\{ u_h(t,\mu) \: | \: \mu\in \mathcal{P}, \: t\in [0,T] \right\} \subset \mathbb{R}^{N_h},
\end{align}
Our goal is to approximate this manifold.
Defining a time discretization, $\{t_0,t_1,\ldots,t_{N_t}\}$, $t_n=n\delta t$, and using a time stepping scheme gives us the time discrete approximation of \eqref{general_problem}:
\begin{align} \label{time_discrete_problem}
u^{n+1}_h(\mu) = F_{h,\delta t}(u^{n}_h;\mu),
\end{align}
where $u^n(\mu)=u(t_n,\mu)$. We will refer to $u^{n}_h(\mu)$ as the state at time step $n$. Note that the discrete time evolution map is not necessarily restricted to only depend on the last state, but can take in several previous states, as is done in e.g. multistep methods, or it could depend on the current state as in implicit methods. We can now define the time-discrete high-fidelity solution manifold:
\begin{align}
M_{h,\delta t} = \left\{ u^n_h(\mu) \: | \: \mu\in \mathcal{P}, \: n=0,\ldots N_t \right\} \subset \mathbb{R}^{N_h}.
\end{align}
The subscripts $h$ and $\delta t$ refer to the chosen spatial and time discretizations respectively. $M_{h,\delta t}$ can be seen as the set of discrete state trajectories parameterized by the set of parameters.
In general, $N_h$ will be very large, which makes advancing the state with \eqref{time_discrete_problem} for many time steps time consuming. This is especially the case when dealing with high-dimensional domains and multiphysics problems. It is indeed a problem when dealing with multi-query problems such as uncertainty quantification and data assimilation or when real-time solutions are of importance as in real-time control settings.
\end{document}
\section{Results} \label{Results}
The aim of this section is to showcase how well our frameworks perform for various parameterized PDE problems. Furthermore, we show how the various approaches, regularizations, and parameters affect the performance.
To assess the performance measure the error on $N_{test}$ test trajectories for parameters values, $\left\{\mu_1,\ldots,\mu_{N_{test}}\right\}$, that the NNs have not seen in the training phase. we measure the mean relative error (MRE) at every time step and take the mean over multiple runs of the test cases:
\begin{align}
\text{MRE}(u^n_h(\mu_i),\tilde{u}^n_h(\mu_i)) = \frac{1}{N_{test}} \sum_{i=1}^{N_{test}}
\frac{\norm{u^n_h(\mu_i) - \tilde{u}^n_h(\mu_i)}_2^2}{\norm{u^n_h(\mu_i)}_2^2},
\end{align}
where
\begin{align}
\norm{u^n_h(\mu_i)}_2^2 = (u_h^n(\mu_i))^T u_h^n(\mu_i).
\end{align}
\begin{comment}
and
\begin{align}
u_h(\mu_i) = \left\{ u_h^{0}(\mu_i), \ldots, u_h^{N_t}(\mu_i) \right\},
\end{align}
are the high-fidelity trajectories, and
\begin{align}
\tilde{u}_h(\mu_i) &= \left\{ \tilde{u}_h^{0}(\mu_i), \ldots, \tilde{u}_h^{N_t}(\mu_i) \right\},
\end{align}
are the trajectories computed using the neural networks in the latent space and decoded back to the high-fidelity space.
\end{comment}
Besides showing the MRE, we also show the standard error:
\begin{align}
\text{Standard Error} = \frac{\sigma}{\sqrt{{N_{test}}}},
\end{align}
where $\sigma$ is the variance of MRE. With this measure we can assess if the trained NN performs similarly on all the test data. I.e. we empirically show robustness and generalization.
For all problems we compare the reconstruction error of the CAE with the reconstruction error of a POD approach. Note that regarding POD the reconstruction error is the same as the projection error. The measure for the POD error is the MRE as for the CAE.
\subsection*{Neural Network Setup}
All neural networks are implemented in Tensorflow 2.0 \cite{abadi2016tensorflow} in Python. The training is performed in the Google Colab framework on NVIDIA Tesla P100 GPUs. \\
The neural network architecture configurations can be found in \ref{conv_ae_appendix} and \ref{evolution_NN_appendix}. \\
\\
\textbf{Remark.} We only present results on dimensionality reduction using convolutional autoencoders and compare them to POD. It should be noted that dense autoencoders were also tested and showed significantly worse results. \\
\\
\textbf{Remark.} As mentioned earlier, we only consider FFNNs, LSTMs, and CCNNs for the time -stepping. We also studied other achitectures, such as neural ODEs \cite{chen2018neural}, gated recurrent units (GRUs), and simple recursive neural networks. However, we chose to not include those results. Neural ODEs performed significantly worse and the training took much longer time. GRUs performed similarly to LSTMs and simple recursive neural networks performed slightly worse.
\subsection{Heat Equation}
In this section we consider results for the linear heat equation parameterized by a space-dependent diffusion rate, $\mu(x)$, on the domain $\Omega=[0,1]^2$:
\begin{subequations}
\begin{alignat}{2}
\partial_t u(\mu) - \mu(x)\Delta u(\mu) &= f, &\quad &\text{in} \quad \Omega, \\
u(\mu) &= u_d, & &\text{on} \quad \Gamma_d, \\
\textbf{n} \cdot \nabla u(\mu) &= u_n, & &\text{in} \quad \Gamma_n, \\
u(\mu) &= 0, & &\text{for} \quad t=0.
\end{alignat}
\end{subequations}
We consider a unit square domain divided into four equally-sized square subdomains, $\Omega_1,\Omega_2,\Omega_3,\Omega_4$ with a specific diffusion rate, $\mu_1,\mu_2,\mu_3,\mu_4$ on each subdomain, where $\mu=(\mu_1,\mu_2,\mu_3,\mu_4)\in [0.1,1.5]^4$. Hence, our parameter space is four-dimensional. The boundary conditions are given by $u_d=0$ on $\Gamma_d = \Omega\cap \{y=1\}$, $u_n=0$ on $\Omega\cap\{x=0, x=1\}$, and $u_n=1$ on $\Omega\cap \{y=0\}$ See Figure \ref{fig:heat_setup} for a visualization of the setup.
The high-fidelity snapshots are computed on a $100\times 100$ grid using second-order Lagrange finite elements, which gives third order convergence. The high-fidelity problem has 40401 degrees of freedom. For the implementation we used the FEniCS library in Python \cite{logg2012automated}. The time-stepping is done using the Crank-Nicolson scheme with 100 time steps of size 0.1, meaning the time horizon spans from $t=0$ to $t=10$.
In Figure \ref{fig:POD_AE_comparison} the convergence of the reconstruction (projection) error for the CAE and the POD is compared. It is clear the CAE achieves higher precision with much lower latent dimension. Indeed, we see that the convergence stagnates around a latent dimension, $N_l=4$, which is actually lower than the intrinsic dimension of the parameterized solution manifold, $\mathcal{M}_{h,\delta t}$. At $N_l=4$ the mean relative error is below $10^{-4}$. To achieve the same accuracy using POD one needs at least a 10-dimensional latent space. The remaining results regarding the heat equation are computed with $N_l=4$.
\begin{figure}
\centering
\begin{subfigure}{.4\textwidth}
\centering
\includegraphics[width=1.\textwidth]{figures/heat/heat_setup.eps}
\caption{Heat equation setup.}
\label{fig:heat_setup}
\end{subfigure}
\begin{subfigure}{.4\textwidth}
\centering
\includegraphics[width=1.\textwidth]{figures/heat/POD_AE_comparison.eps}
\caption{Comparison of convergence of the time averaged MRE of the reconstruction using CAE and POD for the heat equation. }
\label{fig:POD_AE_comparison}
\end{subfigure}
\caption{}
\end{figure}
In Figure \ref{fig:heat_CCNN_memory_test} and \ref{fig:heat_LSTM_memory_test} we see how the MRE evolves in time for varying memory in the time-stepping NN for the CCNN and LSTM, respectively. The CCNN architecture changes as visualized in Figure \ref{fig:CCNN_types}. As expected, for both the CCNN and the LSTM the error decreases with increasing memory. Furthermore, there is not a big difference between the CCNN and LSTM, but the CCNN seems to perform slightly better. From hereon, all results regarding the heat equation are computed with a memory of $\xi=8$.
\begin{figure}
\centering
\begin{subfigure}{.4\textwidth}
\centering
\includegraphics[width=.8\linewidth]{figures/heat/CCNN_memory_test.pdf}
\caption{CCNN.}
\label{fig:heat_CCNN_memory_test}
\end{subfigure}%
\begin{subfigure}{.4\textwidth}
\centering
\includegraphics[width=.8\linewidth]{figures/heat/LSTM_memory_test.pdf}
\caption{LSTM.}
\label{fig:heat_LSTM_memory_test}
\end{subfigure}
\caption{Comparison of CCNN and LSTM in relative error for each time step in high-fidelity space for the heat equation for varying memory, $\xi$. The error for each time step is a computed average over 15 test cases with the standard error. }
\label{fig:heat_memory_test}
\end{figure}
One of the proposed improvements of the time-stepping in Section \ref{Approximating Parameterized Time Evolution using Neural Networks} was to approximate the residual instead of the next state directly. In Figure \ref{fig:heat_residual_test} we see that for the CCNN there is close to no difference, with the state approximation performing slightly better. The same results are apparent for the LSTM in Figure \ref{fig:heat_LSTM_residual_test}, but with the residual case being the best choice.
\begin{figure}
\centering
\begin{subfigure}{.4\textwidth}
\centering
\includegraphics[width=.8\linewidth]{figures/heat/CCNN_residual_test.pdf}
\caption{CCNN.}
\label{fig:heat_CCNN_residual_test}
\end{subfigure}%
\begin{subfigure}{.4\textwidth}
\centering
\includegraphics[width=.8\linewidth]{figures/heat/LSTM_residual_test.pdf}
\caption{LSTM.}
\label{fig:heat_LSTM_residual_test}
\end{subfigure}
\caption{Comparison of computing the next step directly and the residual for the CCNN and LSTM. The figures show relative error for each time step in high-fidelity space for the heat equation.}
\label{fig:heat_residual_test}
\end{figure}
In Figure \ref{fig:heat_CCNN_train_test} and \ref{fig:heat_LSTM_train_test} it is showcased how the number of training trajectories affects the performance for the CCNN and LSTM respectively. As expected, the relative error decreases with increasing number of training trajectories. One further sees that the LSTM requires fewer training trajectories to achieve higher precision in the short term than the CCNN. On the other hand, the MRE seems to stabilize for more time steps in contrast to the LSTM, where the MRE seems to increase. This suggests that the CCNN is the preferable choice for long term predictions with limited training data for this problem. It is worth noting that the relatively large number of training trajectories is due to the fact that we are dealing with a 4-dimensional parameter space, which requires quite a few samples to explorer the full parameter space.
\begin{figure}
\centering
\begin{subfigure}{.4\textwidth}
\centering
\includegraphics[width=.8\linewidth]{figures/heat/CCNN_train_test.pdf}
\caption{CCNN.}
\label{fig:heat_CCNN_train_test}
\end{subfigure}%
\begin{subfigure}{.4\textwidth}
\centering
\includegraphics[width=.8\linewidth]{figures/heat/LSTM_train_test.pdf}
\caption{LSTM.}
\label{fig:heat_LSTM_train_test}
\end{subfigure}
\caption{Comparison of CCNN and LSTM in relative error for each time step in high-fidelity space for the heat equation for number of training samples, $N_{train}$.}
\label{fig:heat_train_test}
\end{figure}
In Section \ref{Approximating Parameterized Time Evolution using Neural Networks} we discussed two regularization terms, weight decay, $\beta_1$, and Jacobian, $\beta_2$, with the aim of promoting generalization and stability. The effect of those terms is shown in Figure \ref{fig:heat_beta_test} for the CCNN. It is, somewhat surprisingly, apparent that the size of the weight decay regularization term has very little effect on the performance, while it is clear that the Jacobian regularization has a large impact. In general, the performance improves with smaller amount of regularization suggesting that stability is not an issue. This could be because the dynamics in this diffusion type problem rather quickly go to a steady state.
\begin{figure}
\centering
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/heat/CCNN_beta_test_reg3.pdf}
\caption{$\beta_1 = 10^{-3}$.}
\label{fig:heat_CCNN_beta_test_reg3}
\end{subfigure}%
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/heat/CCNN_beta_test_reg6.pdf}
\caption{$\beta_1 = 10^{-6}$.}
\label{fig:heat_CCNN_beta_test_reg6}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/heat/CCNN_beta_test_reg9.pdf}
\caption{$\beta_1 = 10^{-9}$.}
\label{fig:heat_CCNN_beta_test_reg9}
\end{subfigure}
\caption{Impact of the two regularization terms, weight decay, $\beta_1$, and Jacobian, $\beta_2$ for the heat equation. The average relative error in high-fidelity space over 15 test trajectories for each time step is shown. Each figure shows the error for a constant $\beta_1$ and varying $\beta_2$. }
\label{fig:heat_beta_test}
\end{figure}
In Figure \ref{fig:reduced_heat_sols} we see the high-fidelity trajectories reduced to the latent space as well as the NN time-stepping approximation for a specific test case. It is clear that in the latent space the dynamics reach a steady state from around time step 20. Furthermore, the approximated trajectories are very close to the reduced high-fidelity trajectories. In figure \ref{fig:heat_time_error_CCNN} a comparison of the error in the latent space, $\text{MRE}(u_l(\mu),\tilde{u}_l(\mu))$, and in the high-fidelity space, $\text{MRE}(u_h(\mu),\tilde{u}_h(\mu))$, is made. Throughout all time steps the error in the latent space is lower than in the high-fidelity space. This increase in error comes from the decoding step of the CAE.
\begin{figure}
\centering
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=.8\linewidth]{figures/heat/reduced_heat_sols.pdf}
\caption{Latent space trajectories for the heat equation.}
\label{fig:reduced_heat_sols}
\end{subfigure}%
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=.8\linewidth]{figures/heat/heat_time_error.pdf}
\caption{Mean relative error in the latent space and high-fidelity space for the heat equation.}
\label{fig:heat_time_error_CCNN}
\end{subfigure}
\caption{$\mu =(0.1902, 0.4583 , 1.2648, 0.7116)$.}
\label{fig:heat_latent_sol_and_error}
\end{figure}
For time steps $t=0$, $t=1$, and $t=10$, the pointwise absolute error is shown for a specific test case in Figure \ref{fig:heat_error}.
\begin{figure}
\centering
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/heat/heat_error_t0}
\caption{$t=0$}
\label{fig:heat_error_t0}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/heat/heat_error_t75}
\caption{$t=1$}
\label{fig:heat_error_t1}
\end{subfigure}%
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/heat/heat_error_t15}
\caption{$t=10$}
\label{fig:heat_error_t10}
\end{subfigure}
\caption{Pointwise absolute error between the high-fidelity solution and the neural network prediction (CCNN) for the heat equation with diffusion rates, $\mu=(1.0904, 0.3550, 0.1267, 1.0159 )$ .}
\label{fig:heat_error}
\end{figure}
\subsection{Linear Advection Equation }
We consider a linear advection equation on the domain $\Omega=[0,1]^2$:
\begin{subequations}
\begin{alignat}{2}
\partial_t u(\mu) + b\cdot \nabla u(\mu) &= 0, & \quad &\text{in} \quad \Omega, \\
u(\mu) &= 0 & & \text{on} \quad \Gamma,
\end{alignat}
where $\Gamma=\partial\Omega$,
\begin{align}
b = \mu_1 \begin{pmatrix}
- y - \frac{1}{2} \\ x-\frac{1}{2}
\end{pmatrix},
\end{align}
\end{subequations}
with initial condition
\begin{align}
u_0(\mu) = \exp\left(\frac{1}{2}\left[\frac{(x-x_{0})^2}{0.005} + \frac{(y-y_{0})^2}{0.005}\right]\right),
\end{align}
where
\begin{align}
\begin{pmatrix} x_0 \\ y_0 \end{pmatrix} =
\frac{1}{4}\begin{pmatrix} \cos(\mu_2) \\ \sin(\mu_2) \end{pmatrix}+ \frac{1}{2}.
\end{align}
This problem models a Gaussian curve being advected with velocity $\mu_1$ in a circle with origin at $[\frac{1}{2},\frac{1}{2}]$ and radius $\frac{1}{4}$, starting at the position given by the angle $\mu_2$. This problem is parameterized by two parameters, $\mu=(\mu_1,\mu_2)\in [0.5,1.5]\times [0,2\pi]$. The first parameter, the velocity, is directly affecting the dynamics, while the other, $\mu_2$, is only dictating the initial placement of the Gaussian curve. Hence, we are dealing with a 2-dimensional parameter space, while the dynamics are only parameterized by a single parameter.
The high-fidelity snapshots are computed on a $60\times 60$ grid using the discontinuous Galerkin method with linear Lagrange elements, resulting in a second-order convergence scheme that suits advection dominated problems well. The high-fidelity model consists of 21600 degrees of freedom. For the implementation we used the FEniCS library in Python \cite{logg2012automated}. The time-stepping is done using the Crank-Nicolson scheme with time steps of size 0.0075 for 2000 steps. resulting in a time interval, $t\in[0,15]$. The training of the neural networks is done using every 4th time step, $s=4$, meaning the model is trained to take steps of size 0.03.
The same tests as for the heat equation have been performed on the advection equation. The NN configuration that performs the best uses a memory of $\xi=6$, $\beta_1=10^{-9}$, $beta_2=10^{-6}$, and computes the residual rather than the state directly. Furthermore, the training has been performed with 120 training trajectories.
In Figure \ref{fig:advection_POD_vs_AE}, we see a significant improvement by using the CAE compared to the POD approach. Using a latent dimension of 2, which is also the intrinsic dimension of the solution manifold, the CAE reconstructs the high-fidelity solution with an MRE between $10^{-3}$ and $10^{-4}$. To achieve the same accuracy using the POD method, one needs a latent dimension of at least 17. This supports the previous claim that POD does, in general, not perform well on advection dominated problems.
Figure \ref{fig:reduced_adv_sols} shows the encoded high-fidelity trajectories as well as the latent space trajectories computed by using the NN. As expected, we see periodic behavior and close to no discrepancy between the encoded high-fidelity trajectory and the NN computed latent space trajectory.
Comparing the errors in the latent space with the errors in the high-fidelity space in Figure \ref{fig:adv_time_error_CCNN}, we observe that the latent space errors are in general an order of magnitude smaller. However, the errors in the latent space show much more spurious oscillations, suggesting that the CAE is not sensitive to small perturbations in the latent space.
By looking at the pointwise error between the high-fidelity and the NN solutions in Figure \ref{fig:linear_advection_sols} it is clear that the NN approximation introduces a small phase error. This error is, however, not large and could possibly be corrected in a post-processing step.
\begin{figure}
\centering
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\textwidth]{figures/linear_advection/advection_POD_vs_AE.eps}
\caption{}
\label{fig:advection_POD_vs_AE}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/linear_advection/reduced_adv_sols.pdf}
\caption{}
\label{fig:reduced_adv_sols}
\end{subfigure}%
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/linear_advection/adv_time_error.pdf}
\caption{}
\label{fig:adv_time_error_CCNN}
\end{subfigure}
\caption{(a) Comparison of convergence of the time averaged MRE of the reconstruction using CAE and POD for the heat equation. (b) Latent space trajectories with advection velocity $\mu_1=1.4161$, and initial angle, $\mu_2=2.8744$ and (c) average test errors computed for the linear advection equation. The predictions are computed using the CCNN dynamic network. }
\label{fig:adv_reduced_sols}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/linear_advection/linear_adv_error_t0}
\caption{$t=0$}
\label{fig:linear_adv_error_t0}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/linear_advection/linear_adv_error_t75}
\caption{$t=7.5$}
\label{fig:linear_adv_error_t75}
\end{subfigure}%
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/linear_advection/linear_adv_error_t15}
\caption{$t=15$}
\label{fig:linear_adv_error_t15}
\end{subfigure}
\caption{Pointwise absolute error between the high-fidelity solution and the neural network prediction (CCNN) for the linear advection equation with velocity, $\mu_1=1.41605949$, and initial angle, $\mu_2=2.87439799$.}
\label{fig:linear_advection_sols}
\end{figure}
\begin{comment}
In Figure \ref{fig:adv_memory_test} we see the CCNN and the LSTM performs on the advection problem for varying memory. In both cases having memory of $\xi = 6$ gives the best results and a memory of $\xi = 2$ the worst. Increasing the memory to $\xi=8$ could lead to overfitting or provide unnecessary information which results in a worse model compared to a memory of $\xi=6$. Hence, from hereon, we will only show results with $\xi=6$ for the advection equation.
\begin{figure}
\centering
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=.8\linewidth]{figures/linear_advection/CCNN_memory_test.pdf}
\caption{CCNN.}
\label{fig:adv_CCNN_memory_test}
\end{subfigure}%
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=.8\linewidth]{figures/linear_advection/LSTM_memory_test.pdf}
\caption{LSTM.}
\label{fig:adv_LSTM_memory_test}
\end{subfigure}
\caption{Comparison of CCNN and LSTM in relative error for each time step in high-fidelity space for the linear advection equation for varying memory, $\xi$. The error for each time step is a computed average over 15 test cases with the standard error. }
\label{fig:adv_memory_test}
\end{figure}
When it comes to assessing the affect of varying number of training trajectories, we see a quite surprising result for the CCNN. In Figure \ref{fig:adv_CCNN_train_test} we see that having larger training sets does not necessarily lead to better models. However, for the LSTM this is the case (see Figure \ref{fig:adv_LSTM_train_test}).
\begin{figure}
\centering
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=.8\linewidth]{figures/linear_advection/CCNN_train_test.pdf}
\caption{CCNN.}
\label{fig:adv_CCNN_train_test}
\end{subfigure}%
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=.8\linewidth]{figures/linear_advection/LSTM_train_test.pdf}
\caption{LSTM.}
\label{fig:adv_LSTM_train_test}
\end{subfigure}
\caption{Comparison of CCNN and LSTM in relative error for each time step in high-fidelity space for the linear advection equation for number of training samples, $N_{train}$.}
\label{fig:adv_train_test}
\end{figure}
For the heat equation it did not have much of an effect whether the model predicts the next state directly or the residual. For the advection equation, however, this has a very large effect. For both the CCNN, Figure \ref{fig:adv_CCNN_residual_test}, and the LSTM, Figure \ref{fig:adv_LSTM_residual_test}, computing the residual leads to much better outcome. The reason for this can probably be found in the behaviour of the solution the advection equation. Here, we are dealing a fixed shape moving around. This means that there the next state is not fundamentally different from the previous one.
\begin{figure}
\centering
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=.8\linewidth]{figures/linear_advection/CCNN_residual_test.pdf}
\caption{CCNN.}
\label{fig:adv_CCNN_residual_test}
\end{subfigure}%
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=.8\linewidth]{figures/linear_advection/LSTM_residual_test.pdf}
\caption{LSTM.}
\label{fig:adv_LSTM_residual_test}
\end{subfigure}
\caption{Comparison of computing the next step directly the residual for the CCNN and LSTM. The figures show relative error for each time step in high-fidelity space for the linear advection equation.}
\label{fig:adv_residual}
\end{figure}
As for the heat equation, the weight decay, $\beta_1$, does not have a significant effect on the performance of the CCNN. However, there is no clear choice in Jacobian regularization parameter, as $\beta_2=10^{-6}$, $\beta_2=10^{-8}$, and $\beta_2=10^{-10}$ result in quite similar outcomes, meaning that the network is not so sensitive to changes on this matter. In general, that is a good sign, since it points to the fact that it is not prone to overfitting.
\begin{figure}
\centering
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/linear_advection/CCNN_beta_test_reg=3.pdf}
\caption{$\beta_1 = 10^{-3}$.}
\label{fig:adv_CCNN_beta_test_reg=3}
\end{subfigure}%
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/linear_advection/CCNN_beta_test_reg=6.pdf}
\caption{$\beta_1 = 10^{-6}$.}
\label{fig:dv_CCNN_beta_test_reg=6}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/linear_advection/CCNN_beta_test_reg=9.pdf}
\caption{$\beta_1 = 10^{-9}$.}
\label{fig:linear_advection/adv_CCNN_beta_test_reg=9}
\end{subfigure}
\caption{Impact of the two regularization terms, weight decay, $\beta_1$, and Jacobian, $\beta_2$. The average relative error in high-fidelity space over 15 test trajectories for each time step is shown. Each figure shows the error for a constant $\beta_1$ and varying $\beta_2$. }
\label{fig:adv_beta_test}
\end{figure}
In Figure \ref{fig:reduced_adv_sols} we see the latent space trajectories predicted by the CCNN plotted with the encoded high-fidelity trajectories. As expected, we see a periodic behaviour. Furthermore, it is very difficult to distinguish the predicted trajectories from the encoded high-fidelity trajectories. However, we do see a slight error in the phase, which is also apparent in Figure \ref{fig:linear_adv_error_t0}, \ref{fig:linear_adv_error_t75}, and \ref{fig:linear_adv_error_t15}, where we see that the error evolves in to two peaks suggesting a difference in phase. For plots of the full solutions computed by the NN see Figure \ref{fig:linear_advection_sols_appendix} in the appendix. Looking at the prediction error in the latent space and the prediction in the high-fidelity space in Figure \ref{fig:adv_time_error_CCNN}, we see that it is much more oscillatory in the latent space. This suggests that decoder smooths the solution, meaning that small variations in the latent space does not lead to significant different outcomes when decoded.
\begin{figure}
\centering
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\textwidth]{figures/linear_advection/advection_POD_vs_AE.eps}
\caption{}
\label{fig:advection_POD_vs_AE}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=.9\linewidth]{figures/linear_advection/reduced_adv_sols.pdf}
\caption{}
\label{fig:reduced_adv_sols}
\end{subfigure}%
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/linear_advection/adv_time_error.pdf}
\caption{}
\label{fig:adv_time_error_CCNN}
\end{subfigure}
\caption{(a) CAE and POD convergence as well as (b) latent space trajecoties and (c) errors computed for the linear advection equation with advection velocity $\mu_1=1.41605949$, and initial angle, $\mu_2=2.87439799$. The predictions are computed using the CCNN dynamic network. }
\label{fig:adv_reduced_sols}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/linear_advection/linear_adv_error_t0}
\caption{$t=0$}
\label{fig:linear_adv_error_t0}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/linear_advection/linear_adv_error_t75}
\caption{$t=7.5$}
\label{fig:linear_adv_error_t75}
\end{subfigure}%
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/linear_advection/linear_adv_error_t15}
\caption{$t=15$}
\label{fig:linear_adv_error_t15}
\end{subfigure}
\caption{Pointwise absolute error between the high-fidelity solution and the neural network prediction (CCNN) for the linear advection equation with velocity, $\mu_1=1.4161$, and initial angle, $\mu_2=2.8744$.}
\label{fig:linear_advection_sols}
\end{figure}
\end{comment}
\subsection{2D Nonlinear Equation - Lid Driven Cavity}
In this section we showcase results for the lid driven cavity incompressible Navier-Stokes problem, parameterized by the Reynolds number:
\begin{subequations}
\begin{alignat}{3}
\partial_t u(Re) + (u(Re)\cdot \nabla)u(Re) - \nabla p(Re) &= \frac{1}{Re} \Delta u(Re), &\text{in} \quad \Omega, \\
\nabla \cdot u(Re) &= 0, &\text{in} \quad \Omega, \\
u(Re) &= 0, &\text{on} \quad \Gamma, \\
(u_x(Re),u_y(Re)) &= (1,0) &\text{for} \quad y=0,
\end{alignat}
\end{subequations}
where $\Omega = (0,1)\times (0,1)$ and $\Gamma$ is the boundary except for the one with $y=0$, and the initial condition is $u(Re)=0$ and $p(Re)$. The goal is to compute the velocity field, $u(Re) = (u_x(Re),u_y(Re))$ and the pressure field, $p(Re)$ in the full time domain and for all Reynolds numbers $Re$ in an interval. Thus, $\mu=Re$. We consider $Re\in[100,300]$. The high-fidelity snapshots are computed using a second-order finite volume scheme on a $100\times 100$ staggered non-uniform grid, ensuring higer resolution near the boundaries. This gives $3\cdot100\cdot100=30000$ degrees of freedom - the $x$-component and the $y$-component of the velocity field, and the pressure field in each point \cite{sanderse2018energy}. The time-stepping is done with a fourth-order Runge-Kutta time-stepping scheme \cite{sanderse2012accuracy}.
The training data consists of 80 parameterized trajectories, each computed with 4000 time steps with $t\in [0,13]$. However, we train the latent time-stepping network using only every fourth time step, $s=4$, conclusively enabling the network to take longer time steps.
The convolutional autoencoder is trained to encode the velocity field as well as the pressure field. Hence, the features are a three channel matrix input, consisting of the velocity in the $x$-direction, the $y$-direction, and the pressure field.
For the lid driven cavity problem the same tests as for the heat equation have been made. However, the plots showing the performance for varying parameters have been omitted. To summarize, we see that the best performance is achieved with a CCNN with a memory of $\xi=8$, $\beta_1=10^{-6}$, $\beta_2=10^{-10}$, a training set of 80 trajectories, and by approximating the residual.
Since we are only dealing with varying one parameter, the Reynolds number, the intrinsic dimension of the solution manifold is 2. In Figure \ref{fig:pod_vs_AE_ldc} we see that the error reaches reaches an order of magnitude at $10^{-3}$ when using a latent space dimension of exactly that. Increasing the dimension only leads to modest improvements. To achieve the same precision using POD one needs a latent space of dimension 6. Hence, we do get a significant improvement.
In Figure \ref{fig:reduced_ldc_sols} we see that the predicted latent trajectories are following the trend well. Furthermore, in Figure \ref{fig:ldc_time_error} one sees that when the trajectories are decoded the error remains around the same order of magnitude as in the latent space, suggesting that small errors in the latent space are not inherited into the high-fidelity space.
\begin{figure}
\centering
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1\textwidth]{figures/ldc/AE_vs_POD.eps}
\caption{}
\label{fig:pod_vs_AE_ldc}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1\linewidth]{figures/ldc/reduced_ldc_sols.pdf}
\caption{}
\label{fig:reduced_ldc_sols}
\end{subfigure}%
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1\linewidth]{figures/ldc/ldc_time_error.pdf}
\caption{}
\label{fig:ldc_time_error}
\end{subfigure}
\caption{(a) CAE and POD convergence as well as (b) trajectories and (c) errors for the lid driven cavity problem computed with Reynolds number, $Re=287$. The predictions are computed using the CCNN time-stepping network.}
\label{fig:ldc_error_time}
\end{figure}
When looking at common measures for the lid driven cavity, the $x$-velocity and the $y$-velocity in the along $y=0.5$ and $x=0.5$ respectively, in Figure \ref{fig:ldc_horizontal_vertical} one sees that the NN approximates the high-fidelity solution throughout the entire time interval. Even at the boundary layer we see a that the NN captures the dynamics. This is further established in Figure \ref{fig:ldc_horizontal_vertical} where one sees that the errors remain quite small during the whole timespan.
\begin{figure}
\centering
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=.8\linewidth]{figures/ldc/horizontal.eps}
\caption{Horizontal velocity.}
\label{fig:horizontal}
\end{subfigure}%
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=.8\linewidth]{figures/ldc/vertical.eps}
\caption{Vertical velocity.}
\label{fig:vertical}
\end{subfigure}
\caption{Velocity in the horizontal and vertical direction and $0\leq x\leq1$, $y=0.5$ and $x=0.5$, $0\leq y\leq1$ respectively for $Re=287$.}
\label{fig:ldc_horizontal_vertical}
\end{figure}
Lastly, considering the pointwise error of the velocity magnitude, Figure \ref{fig:ldc_vel_magnitude_t1_error}-\ref{fig:ldc_vel_magnitude_t10_error}, and the pointwise error of the pressure, Figure \ref{fig:ldc_pres_magnitude_t1_error}-\ref{fig:ldc_pres_magnitude_t10_error}
\begin{figure}
\centering
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/ldc_vel_magnitude_t1_error.pdf}
\caption{$t=1$}
\label{fig:ldc_vel_magnitude_t1_error}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/ldc_vel_magnitude_t5_error.pdf}
\caption{$t=3$}
\label{fig:ldc_vel_magnitude_t5_error}
\end{subfigure}%
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/ldc_vel_magnitude_t10_error.pdf}
\caption{$t=13$}
\label{fig:ldc_vel_magnitude_t10_error}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/ldc_pres_magnitude_t1_error.pdf}
\caption{$t=1$}
\label{fig:ldc_pres_magnitude_t1_error}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/ldc_pres_magnitude_t5_error.pdf}
\caption{$t=3$}
\label{fig:ldc_pres_magnitude_t5_error}
\end{subfigure}%
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/ldc/ldc_pres_magnitude_t10_error.pdf}
\caption{$t=13$}
\label{fig:ldc_pres_magnitude_t10_error}
\end{subfigure}
\caption{Pointwise absolute error of the velocity magnitude (top row), $\sqrt{u_x^2+u_y^2}$, and the pressure field (bottom row) between the high-fidelity solution and the neural network prediction (CCNN) for the lid driven cavity problem with $Re = 287$.}
\label{fig:ldc_error}
\end{figure}
\subsection{Computation Time and Accuracy}
Above, we have showed and discussed performance regarding relative error for our three test cases. The results for the three cases, using the CCNN and the LSTM, are summarized in Table \ref{tab:precision}, where the time averaged error is shown. We clearly see that the differences between the LSTM and CCNN are subtle.
As mentioned in the introduction, the aim is to be able to compute solutions fast in the online stage. In Table \ref{tab:online_time} the high-fidelity as well as the NN online time is shown. In the online stage there has not been used any form of parallelization. Therefore, it should be noted that significant speed ups for both the high-fidelity and the NN approaches could be achieved with a greater effort on this matter. The NN online time and the high-fidelity computation time is computed on an Intel Xeon 2.30GHz CPU in the CPU case and a Tesla p100 in the GPU case for the NNs. For the CCNN there is no significant difference between the online time when using a CPU compared to a GPU, while for the LSTM, the computation time approximately doubles when using a CPU.
In general, it is clear that the NN is significantly faster in the online stage. Especially, for the advection equation, we see a massive speed up. This is due to the fact that purely advection dominated problems requires relatively small time steps to avoid spurious oscillations in the solution.
In table \ref{tab:offline_time} the offline time is shown, divided into NN training time and the time it took to generate the training trajectories. In cases where the training trajectories come from collected data the simulation step is unnecessary, and hence the training time alone is the relevant number. For the training we used a Tesla p100 GPU in Google Colab. Compared to the online stage it makes a massive difference to use a GPU instead of a CPU due to the heavy computations associated with backpropagation. We have chosen to only show the GPU training time. It is clear that the most time consuming part is generating the training trajectories. However, it should be noted that a significant amount of time has been spent on hyperparameter tuning which is not documented here.
\begin{table}[]
\centering
\begin{tabular}{|c|c|c|}
\hline
Test Problem & CCNN & LSTM \\ \hline
Heat Equation & $4.41 \cdot 10^{-4}$ & $3.55\cdot 10^{-4}$
\\
Advection Equation & $1.54 \cdot 10^{-3}$ & $5.02\cdot 10^{-3}$
\\
Lid Driven Cavity & $1.56\cdot 10^{-3}$ & $1.49\cdot 10^{-3}$ \\ \hline
\end{tabular}
\caption{Time averaged MRE for the three test problems using the CCNN and LSTM.}
\label{tab:precision}
\end{table}
\begin{table}[]
\centering
\begin{tabular}{ c c||c|c||c|c| }
\cline{2-6}
\multicolumn{1}{c||}{} & \multicolumn{1}{c||}{High-Fidelity} & \multicolumn{2}{c||}{CCNN} & \multicolumn{2}{c|}{LSTM} \\
\hline
\multicolumn{1}{|c||}{Test Problem}&\multicolumn{1}{c||}{CPU} & CPU & GPU & CPU & GPU\\
\hline
\multicolumn{1}{|c||}{Heat Equation} & 42.62 & 0.58 & 0.48 & 1.21 & 0.49 \\
\multicolumn{1}{|c||}{Linear Advection Equation} & 532.22 & 2.08 & 2.03 & 4.63 & 2.39\\
\multicolumn{1}{|c||}{Lid Driven Cavity} & 94.83 & 5.15 & 4.54 & 11.18 & 4.52 \\
\hline
\end{tabular}
\caption{Online computationn. We remind the reader that the following schemes are used for the high-fidelity computations: i) Third-order Galerkin finite element method on a $100\times 100$ grid and a Crank-Nicolson time-stepping scheme for the heat equation, ii) second-order discontinuous Galerkin finite element method on a $60\times 60$ grid and a Crank-Nicolson time-stepping scheme for the advection equation, and iii) a second-order finite volume scheme on a $100\times 100$ non-uniform staggered grid and a fourth-order explicit Runge-Kutta time-stepping scheme for the lid driven cavity problem. }
\label{tab:online_time}
\end{table}
\begin{table}[]
\centering
\begin{tabular}{|c|c|c|c|} \hline
Test Problem & CCNN & LSTM & Generation of Trajectories \\ \hline
Heat Equation & 1178.89 & 1273.63 & 11485.50 (1450 trajectories) \\
Advection Equation & 396.67 & 622.28 & 50763.13 (120 trajectories) \\
Lid Driven Cavity & 798.00 & 761.59 & 7121.4 (80 trajectories) \\ \hline
\end{tabular}
\caption{Offline computation time, i.e. NN training time, in seconds for the CCNN and LSTM using GPUs. Furthermore we show the time it took to generate the training trajectories. Note that the generation of training trajectories is not necessary in cases where the data already exists. }
\label{tab:offline_time}
\end{table}
\end{document}
|
\section{Introduction}
Our setting is $\mathbb{R}^n$ endowed with the measure $\gamma_{-1}$ whose density with respect to the Lebesgue measure is $\pi ^{n/2}e^{|x|^2}$, $x\in \mathbb{R}^n$. The measure $\gamma_{-1}$ is called the inverse Gaussian measure. The study of harmonic analysis operators in $(\mathbb{R}^n,\gamma _{-1})$ was began by Salogni (\cite{Sa}). The principal motivation for the Salogni's studies was the connection with the Gaussian setting. However, as Bruno and Sj\"ogren (\cite{BrSj}) pointed out, $(\mathbb{R}^n,\gamma _{-1})$ can be seen as a model of a variety of settings where a theory of singular integrals has not been developed. Also, the natural Laplacian on $(\mathbb{R}^n,\gamma _{-1})$, that we will denote by $\mathcal{A}$, can be interpreted as a restriction of the Laplace-Beltrami operator associated with a warped-product manifold whose Ricci tensor is unbounded from below. A complete exposition of the theory of this kind of manifolds can be found in \cite{Chen}.
The aim of this paper is to study $L^p(\mathbb{R}^n,\gamma _{-1})$-boundedness properties of higher order Riesz transforms in the inverse Gaussian setting. Also, we characterize the UMD Banach spaces by using these Riesz transforms.
We consider the second order differential operator $\mathcal{A}_0$ defined by
$$
\mathcal{A}_0f(x)=-\frac{1}{2}\Delta f(x)-x\cdot \nabla f(x),\quad x\in \mathbb{R}^n,
$$
where $f\in C_c^\infty(\mathbb{R}^n)$, the space of the smooth functions with compact support in $\mathbb{R}^n$. Here, $\Delta$ and $\nabla$ denote the usual Euclidean Laplacian and gradient, respectively.
$\mathcal{A}_0$ is essentially selfadjoint in $L^2(\mathbb{R}^n,\gamma _{-1})$. $\mathcal{A}$ denotes the closure of $\mathcal{A}_0$ in $L^2(\mathbb{R}^n,\gamma _{-1})$.
For every $k=(k_1,...,k_n)\in \mathbb{N}^n$ by $H_k$ we represent the $k$-th Hermite polynomial given by $H_k(x)=\prod_{i=1}^nH_{k_i}(x_i)$, $x=(x_1,...,x_n)\in \mathbb{R}^n$, where, for every $m\in \mathbb{N}$,
$$
H_m(z)=(-1)^me^{z^2}\frac{d^m}{dz^m}e^{-z^2},\quad z\in \mathbb{R}.
$$
We have that, for every $k=(k_1,...,k_n)\in \mathbb{N}^n$,
$$
\mathcal{A}\widetilde{H}_k=(|k|+n)\widetilde{H}_k,
$$
where $|k|=k_1+k_2+...+k_n$ and $\widetilde{H}_k(x)=e^{-|x|^2}H_k(x)$, $x\in \mathbb{R}^n$. The spectrum of $\mathcal{A}$ in $L^2(\mathbb{R}^n,\gamma _{-1})$ is the discrete set $ \{n+m\}_{m\in \mathbb{N}}$.
The operator $-\mathcal{A}$ generates a diffusion semigroup (in the Stein sense \cite{StLP}) $\{T_t^\mathcal{A}\}_{t>0}$ in $(\mathbb{R}^n,\gamma _{-1})$ where, for every $t>0$, we have that
$$
T_t^\mathcal{A}(f)(x)=\int_{\mathbb{R}^n}T_t^\mathcal{A}(x,y)f(y)dy,\quad x\in \mathbb{R}^n,
$$
for every $f\in L^p(\mathbb{R}^n,\gamma _{-1})$, $1\leq p<\infty$, and being
$$
T_t^\mathcal{A}(x,y)=\frac{e^{-nt}}{\pi ^{n/2}(1-e^{-2t})^{n/2}}\exp\left(-\frac{|x-e^{-t}y|^2}{1-e^{-2t}}\right),\quad x,y\in \mathbb{R}^n,\;t>0.
$$
The maximal operator $T_*^\mathcal{A}$ defined by
$$
T_*^\mathcal{A}f=\sup_{t>0}|T_t^\mathcal{A}f|,
$$
was studied by Salogni (\cite{Sa}). She proved that $T_*^\mathcal{A}$ is bounded from $L^1(\mathbb{R}^n,\gamma _{-1})$ into $L^{1,\infty}(\mathbb{R}^n,\gamma _{-1})$. From the general results in \cite{StLP} it can be deduced that $T_*^\mathcal{A}$ is bounded from $L^p(\mathbb{R}^n,\gamma _{-1})$ into itself, for every $1<p<\infty$. Recently, Betancor, Castro and de Le\'on-Contreras \cite{BCdL} have characterized the K\"oethe function spaces with the Hardy-Littlewood property by using the maximal operators
$$
T_{*,k}^\mathcal{A}f=\sup_{t>0}|t^k\partial _t^kT_tf|,\quad k\in \mathbb{N}.
$$
In \cite{Sa} $L^p(\mathbb{R}^n,\gamma _{-1})$-boundedness properties with $1<p<\infty$ for some spectral multipliers associated with the operator $\mathcal{A}$ were proved. The imaginary power $\mathcal{A}^{i\gamma}$, $\gamma \in \mathbb{R}\setminus\{0\}$, of $\mathcal{A}$ is a special case of the multipliers studied in \cite{Sa}. Bruno (\cite{Br}) established endpoints results for $\mathcal{A}^{i\gamma}$, $\gamma \in \mathbb{R}\setminus\{0\}$, proving that $\mathcal{A}^{i\gamma }$ is bounded from $L^1(\mathbb{R}^n,\gamma _{-1})$ into $L^{1,\infty}(\mathbb{R}^n,\gamma _{-1})$. Also, he showed that, for $\lambda \geq 1$, the shifted first order Riesz transform $\nabla (\mathcal{A}+\lambda I)^{-1/2}$ is bounded from $L^1(\mathbb{R}^n,\gamma _{-1})$ into $L^{1,\infty}(\mathbb{R}^n,\gamma _{-1})$. These operators are studied on new Hardy type $H^1$-spaces.
Higher order Riesz transforms associated with the operator $\mathcal{A}$ were studied by Bruno and Sj\"ogren \cite{BrSj}. For every $\alpha =(\alpha _1,...,\alpha _n)\in \mathbb{N}^n\setminus\{0\}$ the $\alpha$-th Riesz transform is defined by $R_\alpha =\partial ^\alpha \mathcal{A}^{-|\alpha |/2}$, where $\partial ^\alpha =\frac{\partial ^{|\alpha|}}{\partial x_1^{\alpha _1}...\partial x_n^{\alpha _n}}$ and $|\alpha|=\alpha _1+...+\alpha _n$. In \cite[Theorem 1.1]{BrSj} it was established that $R_\alpha$ is bounded from $L^1(\mathbb{R}^n,\gamma _{-1})$ into $L^{1,\infty}(\mathbb{R}^n,\gamma _{-1})$ if and only if $|\alpha|\leq 2$.
In \cite[Remark 2.6]{Br} Bruno proved that, for every $\alpha \in \mathbb{N}^n$ with $|\alpha |=1$, $R_\alpha$ is bounded from $L^p(\mathbb{R}^n,\gamma _{-1})$ into itself, for every $1<p<\infty$. In \cite{BrSj} Bruno and Sj\"ogren say that they do not know whether $R_\alpha$ is bounded from $L^p(\mathbb{R}^n,\gamma _{-1})$ into itself for every $1<p<\infty$ and $\alpha \in \mathbb{N}^n$, $|\alpha |>1$, though they expect so. In our first results we prove that, as they expected, $R_\alpha$ is bounded from $L^p(\mathbb{R}^n,\gamma _{-1})$ into itself when $1<p<\infty$ and $\alpha \in \mathbb{N}^n\setminus\{0\}$. We also obtain a representation of $R_\alpha$ as a principal value singular integral. In order to prove our result we need to use some properties of the negative power $\mathcal{A}^{-\beta}$, $\beta >0$, of $\mathcal{A}$. In Section 2 we analyze $\mathcal{A}^{-\beta}$, $\beta >0$. We obtain that, for every $\beta >0$, the operator $\mathcal{A}^{-\beta}$ is bounded from $L^1(\mathbb{R}^n,\gamma _{-1})$ into $L^{1,\infty}(\mathbb{R}^n,\gamma _{-1})$. This result contrasts with the one in \cite[Proposition 6.2]{GCMST2} where it is proved that $\mathcal{L}^{-\beta}$, $\beta >0$, is not bounded from $L^1(\mathbb{R}^n,\gamma _1)$ into $L^{1,\infty}(\mathbb{R}^n,\gamma _1)$, where $\mathcal{L}$ represents the Ornstein-Uhlenbeck operator and $\gamma _1$ denotes the Gaussian measure ($d\gamma _1(x)=\pi ^{-n/2}e^{-|x|^2}dx$) on $\mathbb{R}^n$.
\begin{thm}\label{Th1.1}
Let $\alpha =(\alpha _1,...,\alpha _n)\in \mathbb{N}^n\setminus \{0\}$. For every $f\in C_c^\infty (\mathbb{R}^n)$, the derivative $\partial ^{\alpha}\mathcal{A}^{-|\alpha|/2}(f)(x)$ exists for almost all $x\in \mathbb{R}^n$ and there exists $c_\alpha \in \mathbb{R}$ such that
$$
\partial_x ^{\alpha}\mathcal{A}^{-|\alpha|/2}(f)(x)=\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon}R_\alpha (x,y)f(y)dy+c_\alpha f(x),\quad \mbox{for almost all }x\in \mathbb{R}^n,
$$
where $c_\alpha =0$ if $\alpha _i$ is odd for some $i=1,...,n$.
Furthermore, when $n=1$ and $\alpha$ is even, the last integral is actually absolutely convergent for every $x\in \mathbb{R}$ and in this case no principal value is needed.
Here
$$
R_\alpha (x,y)=\frac{1}{\Gamma (\frac{|\alpha|}{2})}\int_0^\infty \partial _x^{\alpha}T_t^\mathcal{A}(x,y)t^{\frac{|\alpha |}{2}-1}dt,\quad x,y\in \mathbb{R}^n,\;x\not=y.
$$
\end{thm}
Let $\alpha \in \mathbb{N}^n$. Since, for every $\ell \in \mathbb{N}$,
\begin{equation}\label{derivH}
\frac{d}{dz}\widetilde{H}_\ell (z)=-\widetilde{H}_{\ell +1}(z),\quad z\in \mathbb{R},
\end{equation}
we have that, for every $k\in \mathbb{N}^n$,
$$
\partial ^{\alpha}_x\mathcal{A}^{-|\alpha|/2}(\widetilde{H}_k)(x)=\frac{(-1)^{|\alpha|}}{(|k|+n)^{\frac{|\alpha|}{2}}}\widetilde{H}_{k+\alpha}(x),\quad x\in \mathbb{R}^n.
$$
Let $f\in L^2(\mathbb{R}^n,\gamma _{-1})$. We can write $f=\sum_{k\in \mathbb{N}^n}c_k(f)\widetilde{H}_k$ in $L^2(\mathbb{R}^n,\gamma _{-1})$, where, for every $k\in \mathbb{N}^n$,
$$
c_k(f)=\frac{\pi ^{\frac{n}{2}}}{\|\widetilde{H}_k\|_{L^2(\mathbb{R}^n,\gamma _{-1})}^2}\int_{\mathbb{R}^n}f(y)\widetilde{H}_k(y)e^{|y|^2}dy.
$$
We define
$$
R_\alpha f=(-1)^{|\alpha|}\sum_{k\in \mathbb{N}^n}\frac{c_k(f)}{(|k|+n)^{\frac{|\alpha|}{2}}}\widetilde{H}_{k+\alpha}.
$$
For every $k=(k_1,...,k_n)\in \mathbb{N}^n$,
\begin{equation}\label{norma2}
\|\widetilde{H}_{k+\alpha}\|_{L^2(\mathbb{R}^n,\gamma _{-1})}^2=\int_{\mathbb{R}^n}(H_{k+\alpha }(y))^2e^{-|y|^2}dy=\pi ^n 2^{|\alpha|+|k|}\prod_{i=1}^n\Gamma (k_i+\alpha _i+1).
\end{equation}
Then
\begin{align*}
\|R_\alpha f\|_{L^2(\mathbb{R}^n,\gamma _{-1})}^2&=\sum_{k\in \mathbb{N}^n}\frac{|c_k(f)|^2\|\widetilde{H}_{k+\alpha}\|_{L^2(\mathbb{R}^n, \gamma _{-1})}^2}{(|k|+n)^{|\alpha |}}\\
&=\sum_{k\in \mathbb{N}^n}(c_k(f)\|\widetilde{H}_{k}\|_{L^2(\mathbb{R}^n, \gamma _{-1})})^2\frac{\|\widetilde{H}_{k+\alpha}\|_{L^2(\mathbb{R}^n, \gamma _{-1})}^2}{(|k|+n)^{|\alpha |}\|\widetilde{H}_{k}\|_{L^2(\mathbb{R}^n, \gamma _{-1})}^2}\\
&\leq C\sum_{k\in \mathbb{N}^n}(c_k(f)\|\widetilde{H}_{k}\|_{L^2(\mathbb{R}^n, \gamma _{-1})})^2=C\|f\|_{L^2(\mathbb{R}^n,\gamma_{-1})}^2.
\end{align*}
Hence $R_\alpha $ is bounded from $L^2(\mathbb{R}^n,\gamma_{-1})$ into itself.
If $f\in C_c^\infty (\mathbb{R}^n)$, then $R_\alpha (f)(x)=\partial ^{\alpha}_x\mathcal{A}^{-|\alpha|/2}(f)(x)$, $x\in \mathbb{R}^n$.
\begin{thm}\label{Th1.2}
Let $\alpha =(\alpha _1,...,\alpha _n)\in \mathbb{N}^n\setminus\{0\}$ and $1<p<\infty$. The Riesz transform $R_\alpha$ can be extended from $L^2(\mathbb{R}^n,\gamma_{-1})\cap L^p(\mathbb{R}^n,\gamma_{-1})$ to $L^p(\mathbb{R}^n,\gamma_{-1})$ as a bounded operator from $L^p(\mathbb{R}^n,\gamma_{-1})$ into itself. By denoting again $R_\alpha$ to this extension we have that, there exists $c_\alpha \in \mathbb{R}$ such that, for every $f\in L^p(\mathbb{R}^n,\gamma_{-1})$,
$$
R_\alpha (f)(x)=\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon}R_\alpha (x,y)f(y)dy+c_\alpha f(x),\quad \mbox{for almost all }x\in \mathbb{R}^n,
$$
where $c_\alpha =0$ if $\alpha_i$ is odd for some $i=1,...,n$.
When $n=1$ and $\alpha\in \mathbb{N}$ is even the integral defining $R_\alpha$ is absolutely convergent.
\end{thm}
As it was mentioned, Bruno and Sj\"ogren (\cite[Theorem 1.1]{BrSj}) proved that $R_\alpha $ is bounded from $L^1(\mathbb{R}^n,\gamma_{-1})$ into $L^{1,\infty}(\mathbb{R}^n,\gamma_{-1})$ if and only if $1\leq |\alpha|\leq 2$. This property also holds in the Gaussian setting (see \cite{FS} and \cite{GCMST1}). Aimar, Forzani and Scotto (\cite{AFS}) introduced Riesz type operators $\mathfrak{R}_\alpha$, $\alpha \in \mathbb{N}^n\setminus\{0\}$, related to the Ornstein-Uhlenbeck operator. $\mathfrak{R}_\alpha$ is bounded from $L^1(\mathbb{R}^n,\gamma_1)$ into $L^{1,\infty}(\mathbb{R}^n,\gamma_1)$ for every $\alpha \in \mathbb{N}^n\setminus\{0\}$. Motivated by the results in \cite{AFS} we define Riesz transform in the inverse Gaussian setting whose behavior in $L^1(\mathbb{R}^n,\gamma_{-1})$ is different from the one for $R_\alpha$.
We can write $\mathcal{A}_0=\sum_{i=1}^n\delta _i\partial_{x_i}$, where, for every $i=1,...,n$, $\delta _i=-\frac{1}{2}e^{-x_i^2}\partial _{x_i}e^{x_i^2}$. We consider the operator $\mathcal{\bar{A}}_0=\sum_{i=1}^n\partial_{x_i}\delta_i$. We have that $\mathcal{\bar{A}}_0=-n+\mathcal{A}_0$. Let $\mathcal{\bar A}$ the closure of $\mathcal{\bar A}_0$ in $L^2(\mathbb{R}^n,\gamma_{-1})$. For every $k\in \mathbb{N}^n$, $\mathcal{\bar A}\widetilde{H}_k=|k|\widetilde{H}_k$, and the spectrum of $\mathcal{\bar A}$ in $L^2(\mathbb{R}^n,\gamma_{-1})$ is the set of nonnegative integers. For every $\ell , m\in \mathbb{N}$, we have that $\delta_u^m\widetilde{H}_\ell (u)=(-1)^m\frac{\Gamma (\ell +1)}{\Gamma (\ell -m +1)}\widetilde{H}_{\ell -m}(u)$, $u\in \mathbb{R}$. Here, we understand $\widetilde{H}_\ell=0$ when $\ell <0$. Then, for every $\alpha =(\alpha _1,...,\alpha _n)\in \mathbb{N}^n$, by denoting $\delta ^\alpha=\prod_{i=1}^n\delta _i^{\alpha_i}$, we get
$$
\delta ^\alpha \widetilde{H}_k=(-1)^{|\alpha|}\prod_{i=1}^n\frac{\Gamma (k_i+1)}{\Gamma (k_i-\alpha _i+1)}\widetilde{H}_{k-\alpha},\quad k=(k_1,...,k_n)\in \mathbb{N}^n,\;k_r\geq \alpha _r, \;r=1,...,n.
$$
If $\alpha =(\alpha _1,...,\alpha_n)\in \mathbb{N}^n\setminus\{0\}$ and $k=(k_1,...,k_n)\in \mathbb{N}^n\setminus\{0\}$, with $k_r\geq \alpha _r$, $r=1,...,n$,
$$
\delta ^\alpha \mathcal{\bar A}^{-|\alpha|/2}(\widetilde{H}_k)=\frac{(-1)^{|\alpha|}}{|k|^{\frac{|\alpha|}{2}}}\prod_{i=1}^n\frac{\Gamma (k_i+1)}{\Gamma (k_i-\alpha _i+1)}\widetilde{H}_{k-\alpha}.
$$
In other case, $\delta^\alpha \mathcal{\bar A}^{-|\alpha|/2}(\widetilde{H}_k)=0$ (see Section \ref{barA} for details).
Let $\alpha =(\alpha _1,...,\alpha_n)\in \mathbb{N}^n\setminus\{0\}$. We define the Riesz transform $\bar{R}_\alpha$ on $L^2(\mathbb{R}^n,\gamma_{-1})$ as follows
$$
\overline{R}_\alpha (f)=(-1)^{|\alpha|}\sum_{\substack{k=(k_1,...,k_n)\in \mathbb{N}^n \\k_r\geq \alpha_r, r=1,...,n}}\frac{1}{|k|^{\frac{|\alpha|}{2}}}\prod_{i=1}^n \frac{\Gamma (k_i+1)}{\Gamma (k_i-\alpha _i+1)}c_k(f)\widetilde{H}_{k-\alpha}.
$$
Thus, $\overline{R}_\alpha$ is bounded from $L^2(\mathbb{R}^n,\gamma_{-1})$ into itself. If $f\in C_c^\infty (\mathbb{R}^n)$ and $c_0(f)=0$ then $\overline{R}_\alpha f=\delta ^\alpha \mathcal{\bar A}^{-|\alpha|/2}f$.
\begin{thm}\label{Th1.3}
Let $\alpha =(\alpha _1,...,\alpha_n)\in \mathbb{N}^n\setminus\{0\}$. The Riesz transform $\overline{R}_\alpha$ can be extended from $L^2(\mathbb{R}^n,\gamma_{-1})\cap L^p(\mathbb{R}^n,\gamma_{-1})$ to $L^p(\mathbb{R}^n,\gamma_{-1})$ as a bounded operator from
(i) $L^p(\mathbb{R}^n,\gamma_{-1})$ into itself, for every $1<p<\infty$.
(ii) $L^1(\mathbb{R}^n,\gamma_{-1})$ into $L^{1,\infty}(\mathbb{R}^n,\gamma_{-1})$, provided that $n=1$ or $|\alpha|>n$, when $n>1$.
By denoting again $\overline{R}_\alpha$ to the extension we have that, for every $f\in L^p(\mathbb{R}^n,\gamma_{-1})$, $1\leq p<\infty$,
$$
\overline{R}_\alpha (f)(x)=\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon}\overline{R}_\alpha (x,y)f(y)dy+c_\alpha f(x),\quad \mbox{for almost all }x\in \mathbb{R}^n,
$$
where $c_\alpha=0$ when $\alpha _i$ is odd for some $i=1,...,n$.
Here
$$
\overline{R}_\alpha (x,y)=\frac{1}{\Gamma (\frac{|\alpha|}{2})}\int_0^\infty \delta_x^\alpha T_t^\mathcal{\bar{A}}(x,y)t^{\frac{|\alpha|}{2}-1}dt,\quad x,y\in \mathbb{R}^n,\;x\not=y.
$$
\end{thm}
Let $X$ be a Banach space. Suppose that $\{M_r\}_{r=1}^m$ is a $X$-valued martingale. The sequence $\{d_r=M_r-M_{r-1}\}_{r=1}^m$, where $M_0$ is understood as 0, is called the martingale difference associated with $\{M_r\}_{r=1}^m$. We say that $\{d_r\}_{r=1}^m$ is a $L^p$-martingale difference sequence when it is the difference sequence associated with a $L^p$-martingale. If $1<p<\infty$, $X$ is said to be a $UMD_p$-space when there exists $\beta >0$ such that for all $X$-valued $L^p$-martingale difference sequence $\{d _r\}_{r=1}^m$ and for all $(\varepsilon _r)_{r=1}^m\in \{-1,1\}^{m}$,
$$
\mathbb{E}\Big\|\sum_{r=1}^m\varepsilon _rd_r\Big\|^p\leq \beta \mathbb{E}\Big\|\sum_{r=1}^md_r\Big\|^p.
$$
$UMD$ is an abbreviation of unconditional martingale difference. If $X$ is $UMD_p$ for some $1<p<\infty$, then $X$ is $UMD_p$ for every $1<p<\infty$. This fact justifies to call $UMD$ to the property without any reference to $p$. Burkholder \cite{Bu} and Bourgain \cite{Bou} proved that the $UMD$ property of $X$ is necessary and sufficient for the boundedness of the Hilbert transform in $L^p(\mathbb{R},X)$, $1<p<\infty$. The $UMD$ property is a central notion in the development of the harmonic analysis when the functions are taking values in infinite dimensional spaces. $UMD$ Banach spaces have been characterized by using other singular integrals that can be seen as Riesz transforms associated to orthogonal systems (see \cite{AT}, \cite{BFMT}, \cite{BFRT} and \cite{HTV}, for instance). In the following result we characterize the Banach spaces with the $UMD$ property by using Riesz transforms in the inverse Gaussian setting. For every $i=1,...,n$, we define $e^i=(e_1^i,...,e_n^i)$ where $e_j^i=0$, $i\not =j$, and $e_i^i=1$.
\begin{thm}\label{Th1.4}
Let $X$ be a Banach space. The following assertions are equivalent.
(i) $X$ is $UMD$.
(ii) For every $i=1,...,n$, $R_{e^i}$ can be extended from $L^p(\mathbb{R}^n,\gamma _{-1})\otimes X$ to $L^p(\mathbb{R}^n,\gamma _{-1},X)$ as a bounded operator from $L^p(\mathbb{R}^n,\gamma _{-1},X)$ into itself, for every $1<p<\infty$.
(iii) For every $i=1,...,n$, $R_{e^i}$ can be extended from $L^p(\mathbb{R}^n,\gamma _{-1})\otimes X$ to $L^p(\mathbb{R}^n,\gamma _{-1},X)$ as a bounded operator from $L^p(\mathbb{R}^n,\gamma _{-1}, X)$ into itself, for some $1<p<\infty$.
(iv) For every $i=1,...,n$, $R_{e^i}$ can be extended from $L^1(\mathbb{R}^n,\gamma _{-1})\otimes X$ to $L^1(\mathbb{R}^n,\gamma _{-1},X)$ as a bounded operator from $L^1(\mathbb{R}^n,\gamma _{-1}, X)$ into $L^{1,\infty}(\mathbb{R}^n,\gamma _{-1}, X)$.
Also the equivalences hold when in the properties (ii), (iii) and (iv) we replace $R_{e^i}$ by the maximal operator $R_{e^i}^*$ defined by
$$
R_{e^i}^*(f)(x)=\sup_{\varepsilon >0}\Big\|\int_{|x-y|>\varepsilon}R_{e^i}(x,y)f(y)dy\Big\|,\quad x\in \mathbb{R}^n\mbox{ and }i=1,...,n.
$$
\end{thm}
\begin{thm}\label{Th1.5}
Let $X$ be a Banach space. The following assertions are equivalent.
(i) $X$ is $UMD$.
(ii) For every $1\leq p<\infty$ there exists, for each $i=1,...,n$, the limit
$$
\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon }R_{e^i}(x,y)f(y)dy,\quad \mbox{ for almost all }x\in \mathbb{R}^n,
$$
for every $f\in L^p(\mathbb{R}^n,\gamma _{-1},X)$.
(iii) For some $1\leq p<\infty$, there exists, for every $i=1,...,n$, the limit
$$
\lim_{\varepsilon\rightarrow 0^+}\int_{|x-y|>\varepsilon}R_{e^i}(x,y)f(y)dy,\quad \mbox{ for almost all }x\in \mathbb{R}^n,
$$
for each $f\in L^p(\mathbb{R}^n,\gamma _{-1},X)$.
(iv) For every $1\leq p<\infty$, $f\in L^p(\mathbb{R}^n,\gamma _{-1},X)$ and $i=1,...,n$, $R_{e^i}^*(f)(x)<\infty$, for almost all $x\in \mathbb{R}^n$.
(v) For some $1\leq p<\infty$ and for every $f\in L^p(\mathbb{R}^n,\gamma _{-1},X)$ and $i=1,...,n$, $R_{e^i}^*(f)(x)<\infty$, for almost all $x\in \mathbb{R}^n$.
\end{thm}
The properties stated in Theorems \ref{Th1.4} and \ref{Th1.5} can also be established when we replace $R$-Riesz transforms by $\overline{R}$-Riesz transforms.
Next aim is to state characterizations of the Banach spaces with the $UMD$ property by using imaginary powers $\mathcal{A}^{i\gamma}$, $\gamma \in \mathbb{R}\setminus\{0\}$, of $\mathcal{A}$. Salogni (\cite[Theorem 3.4.3]{Sa}) proved that, for every $1<p<\infty$ and $\gamma \in \mathbb{R}\setminus\{0\}$,
$$
\|\mathcal{A}^{i\gamma }\|_{L^p(\mathbb{R}^n,\gamma_{-1})\rightarrow L^p(\mathbb{R}^n,\gamma _{-1})}\sim e^{\phi^*_p|\gamma|},
$$
as $\gamma \rightarrow \infty$, when $\phi _p=\arcsin |\frac{2}{p}-1|$. Actually, $\mathcal{A}^{i\gamma}$ is a Laplace transform type multiplier associated with $\mathcal{A}$ defined by the function $\phi _\gamma (t)=\frac{t^{-i\gamma}}{\Gamma (1-i\gamma)}$, $t>0$, for every $\gamma \in \mathbb{R}\setminus\{0\}$. Then, since $\{T_t^\mathcal{A}\}_{t>0}$ is a Stein diffusion semigroup, the $L^p(\mathbb{R}^n,\gamma_{-1})$-boundedness of $\mathcal{A}^{i\gamma}$, $\gamma \in \mathbb{R}\setminus\{0\}$, follows from the general results established in \cite[Chapter III]{StLP}. Recently, Bruno (\cite[Theorem 4.1]{Br}) proved that $\mathcal{A}^{i\gamma}$, $\gamma \in \mathbb{R}\setminus\{0\}$, is bounded from $L^1(\mathbb{R}^n,\gamma_{-1})$ into $L^{1,\infty}(\mathbb{R}^n,\gamma_{-1})$.
Let $\gamma \in \mathbb{R}\setminus\{0\}$. We have that
$$
\mathcal{A}^{i\gamma}f=\sum_{k\in \mathbb{N}^n}(|k|+m)^{i\gamma}c_k(f)\widetilde{H}_k,\quad f\in L^2(\mathbb{R}^n,\gamma_{-1}).
$$
It is immediate to see that $\mathcal{A}^{i\gamma}$ is bounded from $L^2(\mathbb{R}^n,\gamma_{-1})$ into itself. For every $f\in L^2(\mathbb{R}^n,\gamma_{-1})\otimes X$ we define $\mathcal{A}^{i\gamma}$ in the obvious way when $X$ is a Banach space. In order to the operator $\mathcal{A}^{i\gamma}$ is bounded from $L^2(\mathbb{R}^n,\gamma_{-1})\otimes X$ into itself as subspace of $L^2(\mathbb{R}^n,\gamma_{-1},X)$ we need to impose some additional property to the Banach space $X$. For instance, if $X$ is isomorphic to a Hilbert space then $\mathcal{A}^{i\gamma}$ can be extended from $L^2(\mathbb{R}^n,\gamma_{-1})\otimes X$ to $L^2(\mathbb{R}^n,\gamma_{-1},X)$ as a bounded operator from $L^2(\mathbb{R}^n,\gamma_{-1},X)$ into itself. We are going to characterize the $UMD$ Banach spaces as those Banach spaces for which $\mathcal{A}^{i\gamma}$ can be extended from $(L^2(\mathbb{R}^n,\gamma_{-1})\cap L^p
(\mathbb{R}^n,\gamma_{-1}))\otimes X$ to $L^p(\mathbb{R}^n,\gamma_{-1},X)$ as a bounded operator from $L^p(\mathbb{R}^n,\gamma_{-1},X)$ into itself, when $1<p<\infty$, and from $L^1(\mathbb{R}^n,\gamma_{-1},X)$ into $L^{1,\infty}(\mathbb{R}^n,\gamma_{-1},X)$. Our result is motivated by the one in \cite[p. 402]{G} where $UMD$ Banach spaces are characterized by the $L^p(\mathbb{R},dx)$-boundedness properties of the imaginary power $(-\frac{d^2}{dx^2})^{i\gamma}$, $\gamma \in \mathbb{R}\setminus\{0\}$, of $-\frac{d^2}{dx^2}$. Guerre-Delabri\'ere's result was extended to higher dimensions by considering imaginary powers of the Laplacian in \cite[Proposition 1]{BCFR}. In \cite{BCCR} this kind of characterization for $UMD$ Banach spaces is obtained in Hermite and Laguerre settings. As far as we know this property has not been proved for the Ornstein-Uhlenbeck operator in the Gaussian framework.
\begin{thm}\label{Th1.6}
Let $\gamma \in \mathbb{R}\setminus\{0\}$. For every $f\in L^2(\mathbb{R}^n,\gamma_{-1})$, we have that
$$
\mathcal{A}^{i\gamma}(f)(x)=\lim_{\varepsilon \rightarrow 0^+}\Big(\int_{|x-y|>\varepsilon }K_\gamma ^\mathcal{A}(x,y)f(y)dy+\alpha (\varepsilon )f(x)\Big),\quad \mbox{ for almost all }x\in \mathbb{R}^n,
$$
where
$$
K_\gamma ^\mathcal{A}(x,y)=-\int_0^\infty \phi _\gamma (t)\partial_tT_t^\mathcal{A}(x,y)dt,\quad x,y\in \mathbb{R}^n,\;x\not=y,
$$
and
$$
\alpha (\varepsilon)=\frac{1}{\Gamma (\frac{n}{2})}\int_0^\infty \phi_\gamma \big(\frac{\varepsilon ^2}{4u}\big)e^{-u}u^{\frac{n}{2}-1}du,\quad \varepsilon \in (0,\infty),
$$
being $\phi_\gamma(t)=\frac{t^{-i\gamma }}{\Gamma (1-i\gamma )}$, $t\in (0,\infty)$.
Let $X$ be a Banach space. The following assertions are equivalent.
(i) $X$ is $UMD$.
(ii) For every $1<p<\infty$, $\mathcal{A}^{i\gamma}$ can be extended from $(L^2(\mathbb{R}^n,\gamma_{-1})\cap L^p(\mathbb{R}^n,\gamma_{-1}))\otimes X$ to $L^p(\mathbb{R}^n,\gamma_{-1},X)$ as a bounded operator from $L^p(\mathbb{R}^n,\gamma_{-1},X)$ into itself.
(iii) For some $1<p<\infty$, $\mathcal{A}^{i\gamma}$ can be extended from $(L^2(\mathbb{R}^n,\gamma_{-1})\cap L^p(\mathbb{R}^n,\gamma_{-1}))\otimes X$ to $L^p(\mathbb{R}^n,\gamma_{-1},X)$ as a bounded operator from $L^p(\mathbb{R}^n,\gamma_{-1},X)$ into itself.
(iv) $\mathcal{A}^{i\gamma}$ can be extended from $(L^2(\mathbb{R}^n,\gamma_{-1})\cap L^1(\mathbb{R}^n,\gamma_{-1}))\otimes X$ to $L^1(\mathbb{R}^n,\gamma_{-1},X)$ as a bounded operator from $L^1(\mathbb{R}^n,\gamma_{-1},X)$ into $L^{1,\infty}(\mathbb{R}^n,\gamma_{-1},X)$.
We define the maximal operator $\mathcal{A}^{i\gamma}_*$ by
$$
\mathcal{A}_*^{i\gamma}(f)(x)=\sup_{\varepsilon >0}\left\|\int_{|x-y|>\varepsilon}K_\gamma ^\mathcal{A}(x,y)f(y)dy\right\|,\quad f\in \bigcup_{p\geq 1}L^p(\mathbb{R}^n,\gamma _{-1},X).
$$
The following assertions are equivalent to (i).
(v) For every $1<p<\infty$, $\mathcal{A}_*^{i\gamma}$ is bounded from $L^p(\mathbb{R}^n,\gamma _{-1},X)$ into itself.
(vi) For some $1<p<\infty$, $\mathcal{A}_*^{i\gamma}$ is bounded from $L^p(\mathbb{R}^n,\gamma _{-1},X)$ into itself.
(vii) $\mathcal{A}_*^{i\gamma}$ is bounded from $L^1(\mathbb{R}^n,\gamma _{-1},X)$ into $L^{1,\infty}(\mathbb{R}^n,\gamma _{-1},X)$.
(viii) For every $1\leq p<\infty$ and every $f\in L^p(\mathbb{R}^n,\gamma _{-1},X)$ there exists the limit
$$
\lim_{\varepsilon \rightarrow 0^+}\left(\int_{|x-y|>\varepsilon}K_\gamma ^\mathcal{A}(x,y)f(y)dy+\alpha (\varepsilon)f(x)\right),\quad \mbox{ for almost all }x\in \mathbb{R}^n.
$$
(ix) For some $1\leq p<\infty$ and every $f\in L^p(\mathbb{R}^n,\gamma _{-1},X)$ there exists the limit
$$
\lim_{\varepsilon \rightarrow 0^+}\left(\int_{|x-y|>\varepsilon}K_\gamma ^\mathcal{A}(x,y)f(y)dy+\alpha (\varepsilon)f(x)\right),\quad \mbox{ for almost all }x\in \mathbb{R}^n.
$$
(x) For every $1\leq p<\infty$ and every $f\in L^p(\mathbb{R}^n,\gamma _{-1},X)$, $\mathcal{A}_*^{i\gamma}(f)(x)<\infty$, for almost all $x\in \mathbb{R}^n$.
(xi) For some $1\leq p<\infty$ and every $f\in L^p(\mathbb{R}^n,\gamma _{-1},X)$, $\mathcal{A}_*^{i\gamma}(f)(x)<\infty$, for almost all $x\in \mathbb{R}^n$.
\end{thm}
This paper is organized as follows. In Section 2 we study the negative power $\mathcal{A}^{-\beta}$, $\beta >0$, of $\mathcal{A}$. Higher order Riesz transforms in the inverse Gauss setting are considered in Section 3 where we prove Theorems \ref{Th1.1} and \ref{Th1.2}. Theorem \ref{Th1.3} is established in Section 4 and Theorems \ref{Th1.4} and \ref{Th1.5} in Section 5. Finally, Section 6 is devoted to show the proof of Theorem \ref{Th1.6}.
Throughout this paper $C$ and $c$ denote positive constants that can change in each occurrence.
\section{Negative powers of $\mathcal{A}$.}
In this section we prove $L^p(\mathbb{R}^n,\gamma _{-1})$-boundedness properties of the negative powers $\mathcal{A}^{-\beta}$, $\beta >0$, of $\mathcal{A}$. These properties are different than the ones of the negative powers of the Ornstein-Uhlenbeck operator $\mathcal{L}=-\frac{1}{2}\Delta +x\nabla$. We prove that, for every $\beta >0$, $\mathcal{A}^{-\beta}$ defines a bounded operator from $L^1(\mathbb{R}^n,\gamma _{-1})$ into $L^{1,\infty}(\mathbb{R}^n,\gamma _{-1})$. However, in \cite[Proposition 6.2]{GCMST2} it was proved that if $\beta >0$, $\mathcal{L}^{-\beta}$ is not bounded from $L^1(\mathbb{R}^n,\gamma _1)$ into $L^{1,\infty}(\mathbb{R}^n,\gamma _1)$.
Let $\beta >0$. We define
\begin{equation}\label{Abeta}
\mathcal{A}^{-\beta}(f)=\sum_{k\in \mathbb{N}^n}\frac{c_k(f)}{(|k|+n)^\beta}\widetilde{H}_k,\quad f\in L^2(\mathbb{R}^n,\gamma _{-1}).
\end{equation}
$\mathcal{A}^{-\beta}$ is bounded from $L^2(\mathbb{R}^n,\gamma _{-1})$ into itself. Moreover, when $\beta >1$ the series also converges pointwisely in $\mathbb{R}^n$. Indeed, let $f\in L^2(\mathbb{R}^n,\gamma _{-1})$. We have that
$$
|c_k(f)|\leq \pi ^{\frac{n}{2}}\|f\|_{L^2(\mathbb{R}^n,\gamma _{-1})}\|\widetilde{H}_k\|_{L^2(\mathbb{R}^n,\gamma _{-1})}^{-1},\quad k\in \mathbb{N}^n.
$$
Also, for every $k\in \mathbb{N}^n$, (see \eqref{norma2}),
\begin{equation}\label{2.1}
\|\widetilde{H}_k\|_{L^2(\mathbb{R}^n,\gamma _{-1})}=\pi ^{\frac{n}{2}}2^{\frac{|k|}{2}}\Big(\prod_{j=1}^n\Gamma (k_j+1)\Big)^{\frac{1}{2}},
\end{equation}
and, according to \cite[p. 324]{San},
\begin{equation}\label{2.2}
|H_j(z)|\leq 2\sqrt{\Gamma (j+1)}2^{\frac{j}{2}}e^{\frac{z^2}{2}},\quad z\in \mathbb{R}\mbox{ and }j\in \mathbb{N}.
\end{equation}
Then,
$$
|c_k(f)\widetilde{H}_k(x)|\leq 2^ne^{-\frac{|x|^2}{2}}\|f\|_{L^2(\mathbb{R}^n,\gamma_{-1})},\quad x\in \mathbb{R}^n,\;k\in \mathbb{N}^n,
$$
and if $\beta >1$ it follows that
$$
\sum_{k\in \mathbb{N}^n}\frac{|c_k(f)|}{(|k|+n)^\beta}|\widetilde{H}_k(x)|\leq Ce^{-\frac{|x|^2}{2}}\|f\|_{L^2(\mathbb{R}^n,\gamma _{-1})},\quad x\in \mathbb{R}^n.
$$
The series in \eqref{Abeta} converges pointwise absolutely for each $\beta >0$ when $f\in C_c^\infty (\mathbb{R}^n)$. Indeed, let $f\in C_c^\infty (\mathbb{R}^n)$. Partial integration allows us to see that, for every $r\in \mathbb{N}$ there exists $C=C(f,r)>0$ such that
\begin{equation}\label{2.3}
|c_k(f)|\leq \frac{C}{(|k|+n)^r\|\widetilde{H}_k\|_{L^2(\mathbb{R}^n,\gamma_{-1})}},\quad k\in \mathbb{N}^n.
\end{equation}
Then,
$$
\sum_{k\in \mathbb{N}^n}\frac{|c_k(f)|}{(|k|+n)^\beta}|\widetilde{H}_k(x)|\leq Ce^{-\frac{|x|^2}{2}}\sum_{k\in \mathbb{N}^n}\frac{1}{(|k|+n)^{\beta +r}}<\infty .
$$
We also consider the operator defined, for every $f\in L^2(\mathbb{R}^n,\gamma_{-1})$, by
$$
S_\beta(f)=\frac{1}{\Gamma (\beta)}\int_0^\infty T_t^\mathcal{A}(f)t^{\beta -1}dt,
$$
where the integral is understood in the $L^2(\mathbb{R}^n,\gamma _{-1})$-Bochner sense.
For each $t>0$ we can write
$$
T_t^\mathcal{A} (f)=\sum_{k\in \mathbb{N}^n}e^{-(|k|+n)t}c_k(f)\widetilde{H}_k,\quad f\in L^2(\mathbb{R}^n,\gamma_{-1}).
$$
We obtain
\begin{align*}
\|T_t^\mathcal{A}(f)\|_{L^2(\mathbb{R}^n,\gamma_{-1})}^2&=\sum_{k\in \mathbb{N}^n}e^{-2(|k|+n)t}|c_k(f)|^2\|\widetilde{H}_k\|_{L^2(\mathbb{R}^n,\gamma_{-1})}^2
\leq e^{-2nt}\sum_{k\in \mathbb{N}^n}|c_k(f)|^2\|\widetilde{H}_k\|_{L^2(\mathbb{R}^n,\gamma_{-1})}^2\\
&=e^{-2nt}\|f\|_{L^2(\mathbb{R}^n,\gamma_{-1})}^2,\quad f\in L^2(\mathbb{R}^n, \gamma_{-1}),\;t>0.
\end{align*}
Then,
\begin{align*}
\|S_\beta(f)\|_{L^2(\mathbb{R}^n,\gamma_{-1})}&\leq \frac{1}{\Gamma(\beta)}\int_0^\infty \|T_t^\mathcal{A}(f)\|_{L^2(\mathbb{R}^n,\gamma_{-1})}t^{\beta -1}dt\leq \frac{\|f\|_{L^2(\mathbb{R}^n,\gamma_{-1})}}{\Gamma(\beta)}\int_0^\infty e^{-nt}t^{\beta -1}dt\\
&\leq \frac{\|f\|_{L^2(\mathbb{R}^n,\gamma_{-1})}}{n^\beta},\quad f\in L^2(\mathbb{R}^n,\gamma_{-1}).
\end{align*}
Suppose now that $f\in C_c^\infty(\mathbb{R}^n)$. By using \eqref{2.3} we obtain that there exists $C=C(f)>0$ such that
$$
|T_t^\mathcal{A}(f)(x)|\leq Ce^{-\frac{|x|^2}{2}}e^{-nt},\quad x\in \mathbb{R}^n,\;t>0.
$$
We can write
\begin{align*}
S_\beta(f)(x)&=\frac{1}{\Gamma (\beta)}\sum_{k\in \mathbb{N}^n}c_k(f)\widetilde{H}_k(x)\int_0^\infty e^{-(|k|+n)t}t^{\beta -1}dt\\
&=\sum_{k\in \mathbb{N}^n}\frac{c_k(f)}{(|k|+n)^\beta}\widetilde{H}_k(x)=\mathcal{A}^{-\beta}(f)(x),\quad x\in \mathbb{R}^n.
\end{align*}
Since $C_c^\infty (\mathbb{R}^n)$ is dense in $L^2(\mathbb{R}^n,\gamma_{-1})$, $\mathcal{A}^{-\beta}(f)=S_\beta(f)$, $f\in L^2(\mathbb{R}^n,\gamma_{-1})$.
According to \cite[Theorem 2.5]{Br} we have that, for every $f\in C_c^\infty (\mathbb{R}^n)$
$$
\mathcal{A}^{-\beta}(f)(x)=\int_{\mathbb{R}^n}M_\beta (x,y)f(y)dy,
$$
for all $x$ outside the support of $f$, where
$$
M_\beta (x,y)=\frac{1}{\Gamma (\beta)}\int_0^\infty T_t^\mathcal{A}(x,y)t^{\beta -1}dt,\quad x,y\in \mathbb{R}^n,\;x\not =y.
$$
\begin{prop}\label{Prop2.1}
Let $\beta >0$. The operator $\mathcal{A}^{-\beta}$ can be extended from $L^p(\mathbb{R}^n, \gamma_{-1})\cap L^2(\mathbb{R}^n, \gamma_{-1})$ to $L^p(\mathbb{R}^n, \gamma_{-1})$ as a bounded operator from $L^p(\mathbb{R}^n, \gamma_{-1})$ into itself, when $1<p<\infty$, and from $L^1(\mathbb{R}^n, \gamma_{-1})$ into $L^{1,\infty}(\mathbb{R}^n, \gamma_{-1})$.
\end{prop}
\begin{proof}
We use the method consisting in decomposing the operator $\mathcal{A}^{-\beta}$ in two parts called local and global parts. This procedure of decomposition was employed by Muckenhoupt (\cite{Mu1} and \cite{Mu2}) in the Gaussian setting.
From now on we consider the function $m$ given by $m(x)=\min\{1,\frac{1}{|x|^2}\}$, $x\in \mathbb{R}^n\setminus\{0\}$, and $m(0)=1$ and the region $N$ defined by
$$
N=\left\{(x,y)\in \mathbb{R}^n\times \mathbb{R}^n:|x-y|\leq n\sqrt{m(x)}\right\}.
$$
We decompose $\mathcal{A}^{-\beta}$ as follows
$$
\mathcal{A}^{-\beta}=\mathcal{A}_{\rm loc}^{-\beta}+\mathcal{A}_{\rm glob}^{-\beta},
$$
where $\mathcal{A}_{\rm loc}^{-\beta}(f)(x)=\mathcal{A}^{-\beta}(f\chi _N(x,\cdot))(x)$, $x\in \mathbb{R}^n$.
According to \cite[Lemma 3.3.1]{Sa} we get
$$
M_\beta (x,y)\leq C\left(\int_0^{m(x)}\frac{e^{-\frac{c|x-y|^2}{t}}}{t^{\frac{n}{2}}}t^{\beta -1}dt+\int_{m(x)}^\infty \frac{e^{-nt}}{(1-e^{-2t})^{\frac{n}{2}}}t^{\beta -1}dt\right),\quad (x,y)\in N.
$$
By choosing $0<\varepsilon <\min\{2\beta,n\}$ we obtain
\begin{align*}
M_\beta(x,y)&\leq C\left(\int_0^{m(x)}\frac{e^{-\frac{c|x-y|^2}{t}}}{t^{\frac{n-\varepsilon}{2}}}t^{\beta -\frac{\varepsilon}{2}-1}dt+\int_{m(x)}^\infty \frac{dt}{t^{\frac{n-\varepsilon}{2}+1}}\right)\\
&\leq C\left(\frac{1}{|x-y|^{n-\varepsilon}}\int_0^1t^{\beta -\frac{\varepsilon}{2}-1}dt+m(x)^{-\frac{n-\varepsilon}{2}}\right)\leq \frac{C}{|x-y|^{n-\varepsilon}},\quad (x,y)\in N.
\end{align*}
In the last inequality we have taken into account that $\sqrt{m(x)}\sim \frac{1}{1+|x|}$, $x\in \mathbb{R}^n$, and that $(1+|x|)|x-y|\leq C$, provided that $(x,y)\in N$.
We have that
\begin{align*}
\sup_{x\in \mathbb{R}^n}\int_{\mathbb{R}^n}M_\beta(x,y)\chi_N(x,y)dy&\leq C\sup_{x\in \mathbb{R}^n}\int_{|x-y|\leq n\sqrt{m(x)}}\frac{dy}{|x-y|^{n-\varepsilon}}\\
&\leq C\sup_{x\in \mathbb{R}^n}\int_0^{n\sqrt{m(x)}}r^{\varepsilon -1}dr\leq C\sup_{x\in \mathbb{R}^n}m(x)^{\frac{\varepsilon}{2}}<\infty.
\end{align*}
Also, since $\sqrt{m(x)}\sim \frac{1}{1+|x|}\sim\frac{1}{1+|y|}\sim \sqrt{m(y)}$, $(x,y)\in N$,
$$
\sup_{y\in \mathbb{R}^n}\int_{\mathbb{R}^n}M_\beta(x,y)\chi_N(x,y)dx<\infty.
$$
Hence, the operator $M_{\beta ,{\rm loc}}$ defined by
$$
M_{\beta ,{\rm loc}}(f)(x)=\int_{\mathbb{R}^n}M_\beta(x,y)\chi_N(x,y)f(y)dy,\quad x\in \mathbb{R}^n,
$$
is bounded from $L^p(\mathbb{R}^n,dx)$ into itself, for every $1\leq p<\infty$. Since $M_{\beta ,{\rm loc}}$ is a local operator, by using \cite[Proposition 3.2.5]{Sa} we deduce that $M_{\beta, {\rm loc}}$ is bounded from $L^p(\mathbb{R}^n,\gamma_{-1})$ into itself, for every $1\leq p<\infty$.
Suppose that $f,g\in C_c^\infty (\mathbb{R}^n)$. As above we take $0<\varepsilon <\min\{2\beta, n\}$ and we get
\begin{align*}
\int_{\mathbb{R}^n}|f(x)|\int_{\mathbb{R}^n}\int_0^\infty t^{\beta -1}T_t^\mathcal{A}(x,y)\chi _N(x,y)|g(y)|dtdydx\leq C\int_{\mathbb{R}^n}|f(x)|\int_{\mathbb{R}^n}|g(y)|\frac{\chi _N(x,y)}{|x-y|^{n-\varepsilon}}dydx<\infty .
\end{align*}
Then,
$$
\int_{\mathbb{R}^n}f(x)\mathcal{A}_{\rm loc}^{-\beta}(g)(x)dx=\int_{\mathbb{R}^n}f(x)M_{\beta ,{\rm loc}}(g)(x)dx.
$$
We deduce that $\mathcal{A}_{\rm loc}^{-\beta}(g)(x)=M_{\beta,{\rm loc}}(g)(x)$, for almost all $x\in \mathbb{R}^n$. Since $\mathcal{A}_{\rm loc}^{-\beta}$ and $M_{\beta,{\rm loc}}$ are bounded from $L^2(\mathbb{R}^n,\gamma_{-1})$ into itself, $\mathcal{A}_{\rm loc}^{-\beta}(f)=M_{\beta,{\rm loc}}(f)$, $f\in L^2(\mathbb{R}^n,\gamma_{-1})$. It follows that, for every $1\leq p<\infty$, $\mathcal{A}_{\rm loc}^{-\beta}$ can be extended from $L^2(\mathbb{R}^n,\gamma_{-1})\cap L^p(\mathbb{R}^n,\gamma_{-1})$ to $L^p(\mathbb{R}^n,\gamma_{-1})$ as a bounded operator from $L^p(\mathbb{R}^n,\gamma_{-1})$ into itself.
We now study the operator $M_{\beta,{\rm glob}}$ defined by
$$
M_{\beta,{\rm glob}}(f)(x)=\int_{\mathbb{R}^n}M_\beta (x,y)\chi_{N^c}(x,y)f(y)dy,\quad x\in \mathbb{R}^n.
$$
By making the change of variables $s=1-e^{-2t}$, $t\in (0,\infty)$, and taking into account that $|x-ry|^2=|y-rx|^2+(1-r^2)(|x|^2-|y|^2)$, $x,y\in \mathbb{R}^n$, $r\in \mathbb{R}$, we obtain
$$
M_\beta (x,y)=\frac{\pi ^{-\frac{n}{2}}}{2^\beta\Gamma (\beta)}e^{|y|^2-|x|^2}\int_0^1\frac{e^{-\frac{|y-x\sqrt{1-s}|^2}{s}}}{s^{\frac{n}{2}}}(1-s)^{\frac{n}{2}-1}(-\log (1-s))^{\beta -1}ds,\quad x,y\in \mathbb{R}^n.
$$
We now use some notations that were introduced in \cite{PeSo} and proceed as in the proof of \cite[Proposition 2.2]{PeSo}. For every $x,y\in \mathbb{R}^n$ we define
\begin{equation}\label{uPeSo}
a=|x|^2+|y|^2,\quad b=2\langle x,y\rangle,\quad s_0=2\frac{\sqrt{a^2-b^2}}{a+\sqrt{a^2-b^2}}\quad \mbox{and}\quad u(s)=\frac{|y-x\sqrt{1-s}|^2}{s}, \;\;s\in (0,1).
\end{equation}
Assume that $(x,y)\not \in N$. Suppose first that $b\leq 0$. In this case, $u(s)\geq \frac{a}{s}-|x|^2$, $s\in (0,1)$. Furthermore, $s\leq C(-\log (1-s))\leq C(1-s)^{-1/(4\beta)}$, $s\in (0,1)$. Then,
$$
M_\beta (x,y)\leq Ce^{|y|^2}\int_0^1\frac{e^{-\frac{a}{s}}}{s^{\frac{n}{2}+1}(1-s)^{\frac{3}{4}}}ds.
$$
By making the change of variable $r=\frac{1}{s}-1$, $s\in (0,1)$, and taking into account that $a\geq \frac{1}{2}$ we obtain
$$
M_\beta (x,y)\leq Ce^{-|x|^2}\int_0^\infty \frac{e^{-\frac{r}{2}}(1+r)^{\frac{n}{2}-\frac{1}{4}}}{r^{\frac{3}{4}}}dr\leq Ce^{-|x|^2}.
$$
Suppose now $b>0$. We write $u_0=u(s_0)=\frac{|y|^2-|x|^2}{2}+\frac{|x+y||x-y|}{2}$. By using that
$$
\sup_{s\in (0,1)}\frac{e^{-u(s)}}{s^{\frac{n}{2}}}\sim \frac{e^{-u_0}}{s_0^{n/2}},
$$
we have that
$$
M_\beta (x,y)\leq Ce^{|y|^2-|x|^2}\sup_{s\in (0,1)}\frac{e^{-u(s)}}{s^{\frac{n}{2}}}\int_0^1(1-s)^{\frac{n}{2}-1}(-\log (1-s))^{\beta -1}ds\leq Ce^{|y|^2-|x|^2}\frac{e^{-u_0}}{s_0^{n/2}}.
$$
Since $s_0\sim \frac{\sqrt{a^2-b^2}}{a}$ we conclude that, when $(x,y)\not \in N$,
\begin{equation}\label{Mbeta}
M_\beta (x,y)\leq C\left\{
\begin{array}{ll}
e^{-|x|^2},&\mbox{ if }\langle x,y\rangle \leq 0,\\
\displaystyle \Big(\frac{|x+y|}{|x-y|}\Big)^{\frac{n}{2}}\exp\Big(\frac{|y|^2-|x|^2}{2}-\frac{|x-y||x+y|}{2}\Big),&\mbox{ if }\langle x,y\rangle >0.
\end{array}
\right.
\end{equation}
Let $1<q<\infty$. Since $||y|^2-|x|^2|\leq |x+y||x-y|$, $x,y\in \mathbb{R}^n$, and $|x-y||x+y|\geq n$, when $(x,y)\in N^c$, as in \cite[p. 501]{Pe} we obtain
\begin{align*}
\int_{\mathbb{R}^n}e^{\frac{|x|^2}{q}-\frac{|y|^2}{q}}M_\beta (x,y)\chi_{N^c}(x,y)dy&\leq C\left(\int_{\mathbb{R}^n}e^{-|x|^2(1-\frac{1}{q})}e^{-\frac{|y|^2}{q}}dy\right.\\
&\quad +\left.\int_{\mathbb{R}^n}|x+y|^ne^{-|x+y||x-y|(\frac{1}{2}-|\frac{1}{q}-\frac{1}{2}|)}dy\right)\leq C,\quad x\in \mathbb{R}^n.
\end{align*}
Also, we have that
$$
\sup_{y\in \mathbb{R}^n}\int_{\mathbb{R}^n}e^{\frac{|x|^2}{q}-\frac{|y|^2}{q}}M_\beta (x,y)\chi_{N^c}(x,y)dx<\infty .
$$
We deduce that $M_{\beta ,{\rm glob}}$ is bounded from $L^p(\mathbb{R}^n,\gamma_{-1})$ into itself, for every $1<p<\infty$.
Next, we are going to see that $M_{\beta ,{\rm glob}}$ is bounded from $L^1(\mathbb{R}^n,\gamma_{-1})$ into $L^{1,\infty}(\mathbb{R}^n,\gamma_{-1})$. We decompose $M_\beta (x,y)$, $x,y\in \mathbb{R}^n$, as follows
$$
M_\beta (x,y)=\frac{\pi ^{-\frac{n}{2}}}{2^\beta \Gamma (\beta)}e^{|y|^2-|x|^2}\left(\int_0^{\frac{1}{2}}+\int_{\frac{1}{2}}^1\right)\frac{e^{-\frac{|y-x\sqrt{1-s}|^2}{s}}}{s^{\frac{n}{2}}}(1-s)^{\frac{n}{2}-1}(-\log (1-s))^{\beta -1}ds=I_1(x,y)+I_2(x,y).
$$
For every $x,y\in \mathbb{R}^n\setminus\{0\}$ we denote by $\theta (x,y)\in [0,\pi ]$ the angle between $x$ and $y$ (we understand $\theta (x,y)=0$, when $n=1$). By using \cite[Lemma 3.3.3]{Sa} we get that, for every $(x,y)\in N^c$, $x,y\not=0$,
\begin{align*}
I_1(x,y)&\leq Ce^{|y|^2-|x|^2}\sup_{s\in (0,1)}\frac{e^{-\frac{|y-x\sqrt{1-s}|^2}{s}}(1-s)^{\frac{n}{2}}}{s^{\frac{n}{2}}}\int_0^{\frac{1}{2}}\frac{(-\log (1-s))^{\beta -1}}{1-s}ds\\
&\leq Ce^{|y|^2-|x|^2}\sup_{r\in (0,1)}\frac{e^{-\frac{|(1+r)y-(1-r)x|^2}{r}}(1-r)^n}{r^{\frac{n}{2}}}\leq Ce^{|y|^2-|x|^2}\min\{(1+|x|)^n,(|x|\sin \theta (x,y))^{-n}\}.
\end{align*}
On the other hand, by proceeding as in \cite[Proposition 5.1]{BrSj} we estimate $I_2(x,y)$, $(x,y)\in N^c$. We first observe that
\begin{align*}
I_2(x,y)&\leq Ce^{|y|^2-|x|^2}\int_{\frac{1}{2}}^1e^{-|y-x\sqrt{1-s}|^2}(1-s)^{\frac{n}{2}-1}(-\log (1-s))^{\beta -1}ds\quad x,y\in \mathbb{R}^n.
\end{align*}
If $|y|\geq 2|x|$ then $|y-x\sqrt{1-s}|\geq 3|y|/4$, $s\in (1/2,1)$, and it follows that
$$
I_2(x,y)\leq Ce^{-c|y|^2}e^{|y|^2-|x|^2}\int_{\frac{1}{2}}^1(1-s)^{\frac{n}{2}-1}(-\log (1-s))^{\beta -1}ds\leq C\frac{e^{|y|^2-|x|^2}}{|y|^{n-1}}\leq C\frac{e^{|y|^2-|x|^2}}{|x|^{n-1}},\quad x,y\not=0.
$$
If $x,y\in \mathbb{R}^n\setminus\{0\}$ we define $r_0=\frac{|y|}{|x|}\cos \theta (x,y)$ and we write $y=y_x+y_\bot$, where $y_x$ is parallel to $x$ and $y_\bot$ is orthogonal to $y$.
By making the change of variables $r=\sqrt{1-s}$ and since $|y-rx|^2=|r-r_0|^2|x|^2+|y_\bot|^2$, it follows that, when if $|y|\leq 2|x|$,
\begin{align*}
I_2(x,y)&\leq Ce^{-c|y_\bot|^2}e^{|y|^2-|x|^2}\int_0^{\frac{1}{2}}(|r-r_0|^{n-1}+|r_0|^{n-1})(-\log r)^{\beta -1}e^{-c|r-r_0|^2|x|^2}dr\\
&\leq Ce^{-c|y_\bot|^2}e^{|y|^2-|x|^2}\left(\frac{1}{|x|^{n-1}}+\Big(\frac{|y|}{|x|}\Big)^{n-1}\right)\int_0^{\frac{1}{2}}(-\log r)^{\beta -1}dr\\
&\leq Ce^{-c|y_\bot|^2}e^{|y|^2-|x|^2}\left (\frac{1}{|x|^{n-1}}+|x|\Big(\frac{|y|}{|x|}\Big)^{n-1}\right),\quad (x,y)\in N^c.
\end{align*}
In the last inequality we have used that $|x|\geq C$ when $(x,y)\in N^c$ and $2|x|\geq |y|$.
By combining the above estimates we obtain
\begin{align}\label{2.4}
M_\beta (x,y)&\leq Ce^{|y|^2-|x|^2}\Big(\min\big\{(1+|x|)^n,(|x|\sin \theta (x,y))^{-n}\big\}+\frac{1}{|x|^{n-1}}\\ \nonumber
&\quad+e^{-c|y_\bot|^2}|x|\Big(\frac{|y|}{|x|}\Big)^{n-1}\chi_{\{|y|\leq 2|x|\}}(x,y)
\Big),\quad (x,y)\in N^c.
\end{align}
According to \cite[Lemma 3.3.4]{Sa} and \cite[Lemmas 4.2 and 4.3]{BrSj} we can see that the operator $M_{\beta ,{\rm glob}}$ is bounded from $L^1(\mathbb{R}^n,\gamma_{-1})$ into $L^{1,\infty}(\mathbb{R}^n,\gamma_{-1})$.
The estimation \eqref{2.4} allows us to prove that, for every $g\in C_c^\infty (\mathbb{R}^n)$, $\mathcal{A}_{\rm glob}^{-\beta}(f)(x)=M_{\beta, {\rm glob}}(f)(x)$, for almost all $x\in \mathbb{R}^n$. Then, since $\mathcal{A}_{\rm glob}^{-\beta}$ and $M_{\beta, {\rm glob}}$ are bounded from $L^2(\mathbb{R}^n,\gamma_{-1})$ into itself, $\mathcal{A}_{\rm glob}^{-\beta}(f)=M_{\beta, {\rm glob}}(f)$, $f\in L^2(\mathbb{R}^n,\gamma_{-1})$. Hence, $\mathcal{A}_{\rm glob}^{-\beta}$ can be extended from $L^2(\mathbb{R}^n,\gamma_{-1})\cap L^p(\mathbb{R}^n,\gamma_{-1})$ to $L^p(\mathbb{R}^n,\gamma_{-1})$ as a bounded operator from $L^p(\mathbb{R}^n,\gamma_{-1})$ into itself, when $1<p<\infty$, and from $L^1(\mathbb{R}^n,\gamma_{-1})$ into $L^{1,\infty}(\mathbb{R}^n,\gamma_{-1})$.
Thus the proof is finished.
\end{proof}
\section{Higher order Riesz transforms associated with the operator $\mathcal{A}$}
In this section we prove Theorems \ref{Th1.1} and \ref{Th1.2} concerning to the higher order Riesz transforms in the inverse Gaussian setting.
\subsection{Proof of Theorem \ref{Th1.1}.} Let $f\in C_c^\infty (\mathbb{R}^n)$. If $\ell\in \mathbb{N}^n$ we have that
$$
\partial_x ^\ell T_t^\mathcal{A}(x,y)=(-1)^{|\ell |}\frac{e^{-nt}e^{-\frac{|x-e^{-t}y|^2}{1-e^{-2t}}}}{\pi^{\frac{n}{2}}(1-e^{-2t})^{\frac{n+|\ell|}{2}}}H_\ell \Big(\frac{x-e^{-t}y}{\sqrt{1-e^{-2t}}}\Big),\quad x,y\in \mathbb{R}^n,\;t>0,
$$
and then
\begin{equation}\label{derivT}
|\partial _x^\ell T_t^\mathcal{A}(x,y)|\leq C\frac{e^{-nt}e^{-c\frac{|x-e^{-t}y|^2}{1-e^{-2t}}}}{(1-e^{-2t})^{\frac{n+|\ell|}{2}}},\quad x,y\in \mathbb{R}^n,\;t>0.
\end{equation}
Suppose that $k\in \mathbb{N}$ and $\ell \in \mathbb{N}^n$ such that $|\ell|<k$. Then,
\begin{equation}\label{3.1}
\int_{\mathbb{R}^n}|f(y)|\int_0^\infty \big|\partial_x^\ell T_t^\mathcal{A}(x,y)\big|t^{\frac{k}{2}-1}dtdy<\infty,\quad x\in \mathbb{R}^n.
\end{equation}
Indeed, by considering the function $m$ defined in Proposition \ref{Prop2.1} and using \eqref{derivT} and \cite[Lemma 3.3.1]{Sa} we obtain, for $\varepsilon \in (0,1)$,
\begin{align*}
\int_0^\infty \big|\partial ^\ell _xT_t^\mathcal{A}(x,y)\big|t^{\frac{k}{2}-1}dt&\leq C\left(\int_0^{m(x)}\frac{e^{-c\frac{|x-y|^2}{t}}}{t^{\frac{n+|\ell|-k}{2}+1}}dt+\int_{m(x)}^\infty \frac{e^{-nt}}{(1-e^{-2t})^{\frac{n+|\ell|}{2}}}t^{\frac{k}{2}-1}dt\right)\\
&\hspace{-4cm}\leq C\left(\frac{1}{|x-y|^{n-\varepsilon}}\int_0^1t^{\frac{k-|\ell|-\varepsilon}{2}-1}dt+\int_{m(x)}^\infty \frac{dt}{t^{\frac{n+|\ell|-k}{2}+1}}\right)\leq C\left(\frac{1}{|x-y|^{n-\varepsilon}}+\frac{1}{m(x)^{\frac{n+|\ell|-k}{2}}}\right)\\
&\hspace{-4cm}\leq C\left(\frac{1}{|x-y|^{n-\varepsilon}}+\frac{1}{m(x)^{\frac{n-\varepsilon}{2}}}\right)\leq \frac{C}{|x-y|^{n-\varepsilon}},\quad (x,y)\in N.
\end{align*}
On the other hand, by reading the proof of \cite[Lemma 3.3.3]{Sa} we deduce that
\begin{align}\label{3.2}
\int_0^\infty \big|\partial_x^\ell T_t^\mathcal{A}(x,y)\big|t^{\frac{k}{2}-1}dt&\leq Ce^{|y|^2-|x|^2}\sup_{t\in (0,\infty)}\frac{e^{-c\frac{|y-e^{-t}x|^2}{1-e^{-2t}}}}{(1-e^{-2t})^{\frac{n+|\ell|}{2}}}\int_0^\infty e^{-nt}t^{\frac{k}{2}-1}dt\nonumber\\
&\leq Ce^{|y|^2-|x|^2}(1+|x|)^{n+|\ell|},\quad (x,y)\in N^c.
\end{align}
Note that this estimation also holds when $|\ell|=k$.
Since $f\in C_c^\infty (\mathbb{R}^n)$, \eqref{3.1} holds. Hence, according to \cite[Lemma 4.2]{BFRT}, we have that, when $k\in \mathbb{N}$, $\ell \in \mathbb{N}^n$ and $|\ell|<k$,
\begin{equation}\label{derivA}
\partial_x^\ell\mathcal{A}^{-k/2}(f)(x)=\frac{1}{\Gamma (\frac{k}{2})}\int_{\mathbb{R}^n}f(y)\int_0^\infty \partial _x^\ell T_t^\mathcal{A}(x,y)t^{\frac{k}{2}-1}dtdy,\quad \mbox{ for almost all }x\in \mathbb{R}^n.
\end{equation}
We assume $\alpha =(\alpha _1,...,\alpha _n)\in \mathbb{N}^n\setminus\{0\}$, so we can suposse without loss of generality that $\alpha _1\geq 1$. Let us take $\ell =(\alpha _1-1,\alpha_2,...,\alpha _n)$. According to \eqref{derivA} we can write
$$
\partial_x^\alpha \mathcal{A}^{-|\alpha|/2}(f)(x)=\partial_{x_1}\left(\frac{1}{\Gamma (\frac{|\alpha|}{2})}\int_{\mathbb{R}^n}f(y)\int_0^\infty \partial _x^\ell T_t^\mathcal{A}(x,y)t^{\frac{|\alpha|}{2}-1}dtdy\right),\quad \mbox{for almost }x\in \mathbb{R}^n.
$$
Assume first $n>1$ and write
$$
\int_{\mathbb{R}^n}f(y)\int_0^\infty \partial _x^\ell T_t^\mathcal{A}(x,y)t^{\frac{|\alpha|}{2}-1}dtdy=F(x)+G(x),\quad x\in \mathbb{R}^n,
$$
where
$$
F(x)=\int_{\mathbb{R}^n}f(y)\int_0^\infty \partial _x^\ell [T_t^\mathcal{A}(x,y)-W_t(x-y)]t^{\frac{|\alpha|}{2}-1}dtdy,\quad x\in \mathbb{R}^n,
$$
and
$$
G(x)=\int_{\mathbb{R}^n}f(y)\int_0^\infty \partial _x^\ell W_t(x-y)t^{\frac{|\alpha|}{2}-1}dtdy,\quad x\in \mathbb{R}^n.
$$
Here $W_t$, $t>0$, denotes the classical heat kernel $W_t(z)=\frac{e^{-|z|^2/(2t)}}{(2\pi t)^{n/2}}$, $z\in \mathbb{R}^n$.
Next we show that
\begin{equation}\label{derivF}
\partial_{x_1}F(x)=\int_{\mathbb{R}^n}f(y)\int_0^\infty \partial _x^\alpha [T_t^\mathcal{A}(x,y)-W_t(x-y)]t^{\frac{|\alpha|}{2}-1}dtdy,\quad \mbox{for almost }x\in \mathbb{R}^n.
\end{equation}
By taking into account \eqref{derivT} we get
\begin{align}\label{dif1}
\int_{m(x)}^\infty \big|\partial _x^\alpha [T_t^\mathcal{A}(x,y)-W_t(x-y)]\big|t^{\frac{|\alpha|}{2}-1}dt&\leq C\int_{m(x)}^\infty \left(\frac{e^{-nt}t^{\frac{|\alpha|}{2}-1}}{(1-e^{-2t})^{\frac{n+|\alpha|}{2}}}+\frac{e^{-c\frac{|x-y|^2}{t}}}{t^{\frac{n}{2}+1}}\right)dt\\\nonumber
&\leq C\int_{m(x)}^\infty \frac{dt}{t^{\frac{n}{2}+1}}\leq \frac{C}{m(x)^{\frac{n}{2}}}\leq C(1+|x|)^n,\quad x,y\in \mathbb{R}^n.
\end{align}
Also by using \eqref{3.2} we obtain,
\begin{align}\label{dif2}
\int_0^{m(x)} \left|\partial _x^\alpha [T_t^\mathcal{A}(x,y)-W_t(x-y)]\right|t^{\frac{|\alpha|}{2}-1}dt&\leq C\left(e^{|y|^2-|x|^2}(1+|x|)^{n+|\alpha|}+\int_0^{m(x)} \frac{e^{-c\frac{|x-y|^2}{t}}}{t^{\frac{n}{2}+1}}dt\right)\\\nonumber
&\hspace{-5cm}\leq C\left(e^{|y|^2-|x|^2}(1+|x|)^{n+|\alpha|}+\frac{1}{|x-y|^{n+|\alpha|}}\right)\leq Ce^{|y|^2}(1+|x|)^{n+|\alpha|},\quad (x,y)\in N^c.
\end{align}
Now we are going to estimate
$$
I(x,y)=\int_0^{m(x)}\ \left|\partial _x^\alpha [T_t^\mathcal{A}(x,y)-W_t(x-y)]\right|t^{\frac{|\alpha|}{2}-1}dt,\quad (x,y)\in N.
$$
We can write
\begin{align}\label{decomposition}
\partial _x^\alpha [T_t^\mathcal{A}(x,y)-W_t(x-y)]&=\frac{(-1)^{|\alpha|}}{\pi ^{\frac{n}{2}}}\left(\frac{e^{-nt}}{(1-e^{-2t})^{\frac{n+|\alpha|}{2}}}\widetilde{H}_\alpha\left(\frac{x-e^{-t}y}{\sqrt{1-e^{-2t}}}\right)-\frac{1}{(2t)^{\frac{n+|\alpha|}{2}}}\widetilde{H}_\alpha\left(\frac{x-y}{\sqrt{2t}}\right)\right)\nonumber\\
&\hspace{-4cm}=\frac{(-1)^{|\alpha|}}{\pi ^{\frac{n}{2}}}\left\{\frac{e^{-nt}-1}{(1-e^{-2t})^{\frac{n+|\alpha|}{2}}}\widetilde{H}_\alpha\left(\frac{x-e^{-t}y}{\sqrt{1-e^{-2t}}}\right)+\left(\frac{1}{(1-e^{-2t})^{\frac{n+|\alpha|}{2}}}-\frac{1}{(2t)^{\frac{n+|\alpha|}{2}}}\right)\widetilde{H}_\alpha\left(\frac{x-e^{-t}y}{\sqrt{1-e^{-2t}}}\right)\right.\nonumber\\
&\hspace{-4cm}\quad \left.+\frac{1}{(2t)^{\frac{n+|\alpha|}{2}}}\left(\widetilde{H}_\alpha\left(\frac{x-e^{-t}y}{\sqrt{1-e^{-2t}}}\right)-\widetilde{H}_\alpha \left(\frac{x-y}{\sqrt{2t}}\right)\right)\right\}=\sum_{j=1}^3 I_j(t,x,y),\quad x,y\in \mathbb{R}^n.
\end{align}
Then
$$
I(x,y)\leq \sum_{j=1}^3\int_0^{m(x)}|I_j(t,x,y)|t^{\frac{|\alpha|}{2}-1}dt=\sum_{j=1}^3I_j(x,y),\quad x,y\in \mathbb{R}^n.
$$
By proceeding as in the proof of \cite[Lemma 3.3.1]{Sa} we can see that
\begin{equation}\label{3.3}
\frac{e^{-c\frac{|x-e^{-t}y|^2}{1-e^{-2t}}}}{(1-e^{-2t})^{\frac{n}{2}}}\leq C\frac{e^{-c\frac{|x-y|^2}{t}}}{t^{\frac{n}{2}}},\quad (x,y)\in N,\;t\in (0,1).
\end{equation}
Then, by taking into account that
\begin{equation}\label{differences}
|e^{-nt}-1|\leq Ct\quad \mbox{ and }\quad \left|\frac{1}{(1-e^{-2t})^{\frac{n+|\alpha|}{2}}}-\frac{1}{(2t)^{\frac{n+|\alpha|}{2}}}\right|\leq \frac{C}{t^{\frac{n+|\alpha|}{2}-1}},\quad t\in (0,1),
\end{equation}
it follows by using \eqref{3.3} that
$$
I_1(x,y)+I_2(x,y)\leq C\int_0^{m(x)}\frac{e^{-c\frac{|x-y|^2}{t}}}{t^{\frac{n}{2}}}dt\leq \frac{C}{|x-y|^{n-\frac{1}{2}}}\int_0^1\frac{dt}{t^{\frac{1}{4}}}\leq \frac{C}{|x-y|^{n-\frac{1}{2}}},\quad (x,y)\in N.
$$
Let us analyze the term $I_3(x,y)$, $(x,y)\in N$. For every $z=(z_1,...,z_n)$, $w=(w_1,...,w_n)\in \mathbb{R}^n$ we write
\begin{align*}
\widetilde{H}_\alpha (z)-\widetilde{H}_\alpha (w)&=\prod_{j=1}^n \widetilde{H}_{\alpha_j}(z_j)-\prod_{j=1}^n\widetilde{H}_{\alpha _j}(w_j)\\
&=\sum_{k=1}^n\left(\prod_{j=1}^{k-1} \widetilde{H}_{\alpha_j}(w_j)\Big( \widetilde{H}_{\alpha_k}(z_k)- \widetilde{H}_{\alpha_k}(w_k)\Big)\prod_{j=k+1}^n \widetilde{H}_{\alpha_j}(z_j)\right).
\end{align*}
Let $(x,y)\in N$, $t\in (0,1)$, and consider $z=\frac{x-e^{-t}y}{\sqrt{1-e^{-2t}}}$ and $w=\frac{x-y}{\sqrt{2t}}$. By taking into account that $(1+|x|+|y|)|x-y|\leq C$ it follows that $e^{-c|z|^2}\leq Ce^{-c|w|^2}$ and then, for each $k=1,...,n$,
$$
\prod_{j=1}^{k-1} |\widetilde{H}_{\alpha_j}(w_j)|\prod_{j=k+1}^n |\widetilde{H}_{\alpha_j}(z_j)|\leq C\exp\Big(-c\sum_{j=1, j\not=k}^n\frac{|x_j-y_j|^2}{t}\Big),
$$
and, by considering also \eqref{derivH} and using the mean value theorem we get
\begin{align*}
\Big|\widetilde{H}_{\alpha_k}(z_k)- \widetilde{H}_{\alpha_k}(w_k)\Big|&\leq Ce^{-c\frac{|x_k-y_k|^2}{t}}|z_k-w_k|\leq Ce^{-c\frac{|x_k-y_k|^2}{t}}\left(\frac{(1-e^{-t})|y_k|}{\sqrt{1-e^{-2t}}}+|x_k-y_k|\Big|\frac{1}{\sqrt{1-e^{-2t}}}-\frac{1}{\sqrt{2t}}\Big|\right)\\
&\hspace{-2cm}\leq Ce^{-c\frac{|x_k-y_k|^2}{t}}\sqrt{t}(|y_k|+|x_k-y_k|)\leq Ce^{-c\frac{|x_k-y_k|^2}{t}}\sqrt{t}(|x_k|+|x_k-y_k|)\leq Ce^{-c\frac{|x_k-y_k|^2}{t}}\sqrt{t}(1+|x|).
\end{align*}
Then,
\begin{equation}\label{Halpha}
\left|\widetilde{H}_\alpha \left(\frac{x-e^{-t}y}{\sqrt{1-e^{-2t}}}\right)-\widetilde{H}_\alpha \left(\frac{x-y}{2t}\right)\right|\leq Ce^{-c\frac{|x-y|^2}{t}}\sqrt{t}(1+|x|),\quad (x,y)\in N,\;t\in (0,1),
\end{equation}
and thus,
\begin{align*}
I_3(x,y)&\leq C(1+|x|)\int_0^{m(x)}\frac{e^{-c\frac{|x-y|^2}{t}}}{t^{\frac{n+1}{2}}}dt\leq C\frac{1+|x|}{|x-y|^{n-\frac{1}{2}}}\int_0^{m(x)}t^{-\frac{3}{4}}dt\\
&=C\frac{(1+|x|)m(x)^{\frac{1}{4}}}{|x-y|^{n-\frac{1}{2}}}\leq C\frac{\sqrt{1+|x|}}{|x-y|^{n-\frac{1}{2}}},\quad (x,y)\in N.
\end{align*}
We deduce that
\begin{equation}\label{acotI}
I(x,y)\leq C\frac{\sqrt{1+|x|}}{|x-y|^{n-\frac{1}{2}}},\quad (x,y)\in N.
\end{equation}
This estimation, jointly with \eqref{dif1} and \eqref{dif2}, leads to
\begin{align*}
\int_{\mathbb{R}^n}|f(y)|\int_0^\infty \left|\partial _x^\alpha [T_t^\mathcal{A}(x,y)-W_t(x-y)]\right|t^{\frac{|\alpha|}{2}-1}dtdy&\nonumber\\
&\hspace{-6cm}\leq C\int_{\mathbb{R}^n}|f(y)|\Big(e^{|y|^2}(1+|x|)^{n+|\alpha|}+\frac{\sqrt{1+|x|}}{|x-y|^{n-\frac{1}{2}}}\Big)dy \leq C(1+|x|)^{n+|\alpha|}<\infty,\quad x\in \mathbb{R}^n,
\end{align*}
where we have used that $f$ has compact support. According to \cite[Lemma 4.2]{BFRT} \eqref{derivF} is then established.
We are going to evaluate $\partial _{x_1}G(x)$. We write $G(x)=\int_{\mathbb{R}^n}f(x-y)\Phi(y)dy$, $x\in \mathbb{R}^n$, where
$$
\Phi (z)=\int_0^\infty \partial _z^\ell (W_t(z))t^{\frac{|\alpha|}{2}-1}dt,\quad z\in \mathbb{R}^n.
$$
Since $n>1$ we have that
$$
|\Phi (z)|\leq C\int_0^\infty \frac{e^{-c\frac{|z|^2}{t}}}{t^{\frac{n+1}{2}}}dt\leq \frac{C}{|z|^{n-1}},\quad z\in \mathbb{R}^n\setminus\{0\}.
$$
According to \cite[Lemma 4.2]{BFRT} we can derivate under the integral sign obtaining
$$
\partial_{x_1}G(x)=\int_{\mathbb{R}^n}\partial _{x_1}(f(x-y))\Phi (y)dy=-\int_{\mathbb{R}^n}\partial_{y_1}(f(x-y))\Phi (y)dy,\quad \mbox{ for almost all }x\in \mathbb{R}^n,
$$
where the last integral is absolutely convergent.
For every $z=(z_1,...,z_n)\in \mathbb{R}^n$ we define $\overline{z}=(z_2,...,z_n)$. Partial integration leads to
\begin{align*}
\int_{\mathbb{R}^n}\partial_{y_1}(f(x-y))\Phi (y)dy&=\lim_{\varepsilon \rightarrow 0^+}\int_{|y|>\varepsilon}\partial_{y_1}(f(x-y))\Phi (y)dy\\
&=\lim_{\varepsilon \rightarrow 0^+}\left(\int_{|\overline{y}|<\varepsilon}\int_{-\infty}^{-\sqrt{\varepsilon ^2-|\overline{y}|^2}}+\int_{|\overline{y}|<\varepsilon}\int_{\sqrt{\varepsilon ^2-|\overline{y}|^2}}^{+\infty}+\int_{|\overline{y}|>\varepsilon}\int_\mathbb{R}\right)\partial_{y_1}(f(x-y))\Phi (y)dy_1d\overline{y}\\
&=\lim_{\varepsilon \rightarrow 0^+}\left(-\int_{|y|>\varepsilon}f(x-y)\partial _{y_1}\Phi (y)dy+\int_{|\overline{y}|<\varepsilon}f(x-y)\Phi (y)\Big]_{y_1=-\infty}^{y_1=-\sqrt{\varepsilon ^2-|\overline{y}|^2}}d\overline{y}\right.\\
&\quad +\left.\int_{|\overline{y}|<\varepsilon}f(x-y)\Phi (y)\Big]_{y_1=\sqrt{\varepsilon ^2-|\overline{y}|^2}}^{y_1=+\infty}d\overline{y}\right)\\
&=\lim_{\varepsilon \rightarrow 0^+}\left(-\int_{|\overline{y}|>\varepsilon}f(x-y)\partial_{y_1}\Phi (y)dy+J_\varepsilon (x)\right),\quad x\in \mathbb{R}^n,
\end{align*}
where
\begin{align*}
J_\varepsilon (x)&=\int_{|\overline{y}|<\varepsilon}f(x_1+\sqrt{\varepsilon ^2-|\overline{y}|^2},\overline{x}-\overline{y})\Phi (-\sqrt{\varepsilon ^2-|\overline{y}|^2},\overline{y})d\overline{y}\\
&\quad -\int_{|\overline{y}|<\varepsilon}f(x_1-\sqrt{\varepsilon ^2-|\overline{y}|^2},\overline{x}-\overline{y})\Phi (\sqrt{\varepsilon ^2-|\overline{y}|^2},\overline{y})d\overline{y},\quad x\in \mathbb{R}^n.
\end{align*}
Let us estimate $\lim_{\varepsilon \rightarrow 0^+}J_\varepsilon (x)$, $x\in \mathbb{R}^n$. We recall that $\Phi$ can be written as follows
$$
\Phi (z)=\frac{(-1)^{|\alpha|-1}}{2^{\frac{|\ell|}{2}}(2\pi)^{\frac{n}{2}}}\int_0^\infty \widetilde{H}_\ell \Big(\frac{z}{\sqrt{2t}}\Big)\frac{dt}{t^{\frac{n+1}{2}}},\quad z\in \mathbb{R}^n.
$$
Suppose now that $\alpha _1$ is odd. Then,
$$
\Phi (-\sqrt{\varepsilon ^2-|\overline{y}|^2},\overline{y})=\Phi (\sqrt{\varepsilon ^2-|\overline{y}|^2},\overline{y}),\quad \overline{y}\in \mathbb{R}^{n-1},\;|\overline{y}|<\varepsilon.
$$
We have that, for every $x\in \mathbb{R}^n$ and $\varepsilon>0$,
\begin{align*}
J_\varepsilon (x)&=\int_{|\overline{y}|<\varepsilon}\big(f(x_1+\sqrt{\varepsilon ^2-|\overline{y}|^2},\overline{x}-\overline{y})-f(x_1-\sqrt{\varepsilon ^2-|\overline{y}|^2},\overline{x}-\overline{y})\big)\Phi (\sqrt{\varepsilon ^2-|\overline{y}|^2},\overline{y})d\overline{y}\\
&=\varepsilon ^{n-1}\int_{|\overline{z}|<1}\big(f(x_1+\varepsilon\sqrt{1-|\overline{z}|^2},\overline{x}-\varepsilon \overline{z})-f(x_1-\varepsilon\sqrt{1-|\overline{z}|^2},\overline{x}-\varepsilon\overline{z})\big)\Phi (\varepsilon\sqrt{1-|\overline{z}|^2},\varepsilon\overline{z})d\overline{z}.
\end{align*}
On the other hand, by performing the change of variable $s=\frac{\varepsilon ^2}{2t}$, we get
\begin{align*}
\Phi (\varepsilon\sqrt{1-|\overline{z}|^2},\varepsilon\overline{z})&=\frac{(-1)^{|\alpha|-1}}{2^{\frac{|\ell|}{2}}(2\pi)^{\frac{n}{2}}}\int_0^\infty \widetilde{H}_{\alpha _1-1}\Big(\frac{\varepsilon\sqrt{1-|\overline{z}|^2}}{\sqrt{2t}}\Big)\prod_{i=2}^n\widetilde{H}_{\alpha _i}\Big(\frac{\varepsilon z_i}{\sqrt{2t}}\Big)\frac{dt}{t^{\frac{n+1}{2}}}\\
&=\frac{(-1)^{|\alpha|-1}\varepsilon^{1-n}}{2^{\frac{|\alpha|}{2}}\pi^{\frac{n}{2}}}\int_0^\infty \widetilde{H}_{\alpha _1-1}(\sqrt{s(1-|\overline{z}|^2)})\prod_{i=2}^n\widetilde{H}_{\ell _i}(z_i\sqrt{s})s^{\frac{n-3}{2}}ds,\quad \varepsilon >0,\;\overline{z}\in \mathbb{R}^{n-1},\;|z|<1.
\end{align*}
It follows that
\begin{align*}
J_\varepsilon (x)&=\frac{(-1)^{|\alpha|-1}}{2^{\frac{|\alpha|}{2}}\pi^{\frac{n}{2}}}\int_{|\overline{z}|<1}\big(f(x_1+\varepsilon\sqrt{1-|\overline{z}|^2},\overline{x}-\varepsilon \overline{z})-f(x_1-\varepsilon\sqrt{1-|\overline{z}|^2},\overline{x}-\varepsilon \overline{z})\big)\\
&\quad \times\int_0^\infty H_{\alpha _1-1}(\sqrt{s(1-|\overline{z}|^2)})\prod_{i=2}^nH_{\ell _i}(z_i\sqrt{s})e^{-s}s^{\frac{n-3}{2}}ds d\overline{z},\quad x\in \mathbb{R}^n,\;\varepsilon >0.
\end{align*}
Then, by using the dominated convergence theorem we obtain
$$
\lim_{\varepsilon \rightarrow 0^+}J_\varepsilon (x)=0,\quad x\in \mathbb{R}^n.
$$
Suppose now that $\alpha _1$ is even. Then,
$$
\Phi (-\varepsilon\sqrt{1-|\overline{y}|^2},\overline{y})=-\Phi (\varepsilon\sqrt{1-|\overline{y}|^2},\overline{y}),\quad \varepsilon >0,\;\overline{y}\in \mathbb{R}^{n-1},\;|\overline{y}|<\varepsilon,
$$
and proceeding as above we get
\begin{align*}
J_\varepsilon (x)&=\frac{(-1)^{|\alpha|}}{2^{\frac{|\alpha|}{2}}\pi^{\frac{n}{2}}}\int_{|\overline{z}|<1}\big(f(x_1+\varepsilon\sqrt{1-|\overline{z}|^2},\overline{x}-\varepsilon \overline{z})+f(x_1-\varepsilon\sqrt{1-|\overline{z}|^2},\overline{x}-\varepsilon \overline{z})\big)\\
&\quad \times \int_0^\infty H_{\alpha _1-1}(\sqrt{s(1-|\overline{z}|^2)})\prod_{i=2}^nH_{\alpha _i}(z_i\sqrt{s})e^{-s}s^{\frac{n-3}{2}}dsd\overline{z},\quad x\in \mathbb{R}^n,\;\varepsilon >0.
\end{align*}
It follows that
$$
\lim_{\varepsilon \rightarrow 0^+}J_\varepsilon (x)=\frac{(-1)^{|\alpha|}}{2^{\frac{|\alpha|}{2}-1}\pi^{\frac{n}{2}}}f(x)\int_{|\overline{z}|<1}\int_0^\infty H_{\alpha _1-1}(\sqrt{s(1-|z|^2)})\prod_{i=2}^{n-1}H_{\alpha _i}(z_i\sqrt{s})e^{-s}s^{\frac{n-3}{2}}dsd\overline{z}=-c_\alpha f(x),\quad x\in \mathbb{R}^n.
$$
We note that if $\alpha _i$ is odd for some $i=2,...,n$, then $c_\alpha =0$. Thus, we conclude that
\begin{align}\label{derivG}
\partial _{x_1}G(x)&=\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon}f(y)\partial _{x_1}\Phi (x-y)dy+c_\alpha f(x)\nonumber\\
&\quad =\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon}f(y)\int_0^\infty \partial _x^\alpha W_t(x-y)t^{\frac{|\alpha|}{2}-1}dtdy+c_\alpha f(x),\quad \mbox{ for almost all }x\in \mathbb{R}^n,
\end{align}
where $c_\alpha=0$ when $\alpha_i$ is odd for some $i=1,...,n$.
We have obtained
\begin{align*}
\partial_x^{\alpha}\mathcal{A}^{-|\alpha|/2}(f)(x)&=\frac{1}{\Gamma (\frac{|\alpha|}{2})}\partial_{x_1}(F(x)+G(x))\\
&=\frac{1}{\Gamma (\frac{|\alpha|}{2})}\left(\int_{\mathbb{R}^n}f(y)\int_0^\infty \partial _x^\alpha [T_t^\mathcal{A}(x,y)-W_t(x-y)]t^{\frac{|\alpha|}{2}-1}dtdy\right.\\
&\quad \left.+\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon}f(y)\int_0^\infty \partial _x^\alpha W_t(x-y)t^{\frac{|\alpha|}{2}-1}dtdy+c_\alpha f(x)\right)\\
&=\frac{1}{\Gamma (\frac{|\alpha|}{2})}\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon }f(y)\int_0^\infty \partial _x^{\alpha }T_t^\mathcal{A}(x,y))t^{\frac{|\alpha|}{2}-1}dtdy+c_\alpha f(x),\quad \mbox{ for almost all }x\in \mathbb{R}^n.
\end{align*}
We now deal with the case $n=1$. We have $\alpha \in \mathbb{N}$, $\alpha \geq 1$. According to \eqref{derivA} we can write
$$
\frac{d^\alpha}{dx^\alpha}\mathcal{A}^{-\alpha /2}(f)(x)=\frac{1}{\Gamma (\frac{\alpha}{2})}\frac{d}{dx}\int_{\mathbb{R}^n}f(y)\int_0^\infty \partial_x^{\alpha -1}T_t^\mathcal{A}(x,y)t^{\frac{\alpha}{2}-1}dtdy=\frac{1}{\Gamma (\frac{\alpha}{2})}\frac{d}{dx}(\overline{F}(x)+\overline{G}(x)),\quad x\in \mathbb{R},
$$
where
$$
\overline{F}(x)=\int_\mathbb{R}f(y)\int_0^\infty\Big[\partial _x^{\alpha -1}T_t^\mathcal{A}(x,y)-\Big((\frac{d^{\alpha -1}}{dx^{\alpha -1}}W_t)(x-y)- (\frac{d^{\alpha -1}}{dx^{\alpha -1}}W_t)(0)\chi_{(1,\infty)}(t)\Big)\Big]t^{\frac{\alpha}{2}-1}dtdy,\quad x\in \mathbb{R},
$$
and
$$
\overline{G}(x)=\int_\mathbb{R}f(y)\int_0^\infty \Big((\frac{d^{\alpha -1}}{dx^{\alpha -1}}W_t)(x-y)- (\frac{d^{\alpha -1}}{dx^{\alpha -1}}W_t)(0)\chi_{(1,\infty)}(t)\Big)t^{\frac{\alpha}{2}-1}dtdy,\quad x\in \mathbb{R}.
$$
Note that if $\alpha $ is even, then $(\frac{d^{\alpha -1}}{dx^{\alpha -1}}W_t)(0)=0$.
By proceeding as in the case of $n>1$ and taking into account that
\begin{align*}
\left|(\frac{d^{\alpha -1}}{dx^{\alpha -1}}W_t)(x-y)- (\frac{d^{\alpha -1}}{dx^{\alpha -1}}W_t)(0)\right|\leq \frac{C}{t^{\frac{\alpha}{2}}}\Big|\widetilde{H}_{\alpha -1}\big(\frac{x-y}{\sqrt{2t}}\big)-\widetilde{H}_{\alpha -1}(0)\Big|\leq C\frac{|x-y|}{t^{\frac{\alpha +1}{2}}},\quad x,y\in \mathbb{R},\;t>1,
\end{align*}
we can see that the integral defining $\overline{F}(x)$ is absolutely convergent for every $x\in \mathbb{R}$ and
$$
\frac{d}{dx}\overline{F}(x)=\int_\mathbb{R}f(y)\int_0^\infty \partial _x^\alpha [T_t^\mathcal{A}(x,y)-W_t(x-y)]t^{\frac{\alpha}{2}-1}dtdy,
$$
being also this integral absolutely convergent for almost $x\in \mathbb{R}$.
On the other hand, by considering
\begin{align*}
\overline{\Phi }(y)&=\int_0^\infty \Big((\frac{d^{\alpha -1}}{dx^{\alpha -1}}W_t)(x-y)- (\frac{d^{\alpha -1}}{dx^{\alpha -1}}W_t)(0)\chi_{(1,\infty)}(t)\Big)t^{\frac{\alpha}{2}-1}dt\\
&=\frac{(-1)^{\alpha -1}}{2^{\frac{\alpha }{2}}\sqrt{\pi}}\int_0^\infty \Big(\widetilde{H}_{\alpha-1}\Big(\frac{y}{\sqrt{2t}}\Big)-\widetilde{H}_{\alpha -1}(0)\chi_{(1,\infty )}(t)\Big)\frac{dt}{t},\quad y\in \mathbb{R}\setminus\{0\},
\end{align*}
we can write
$$
\overline{G}(x)=\int_\mathbb{R}f(x-y)\overline{\Phi}(y)dy,\quad x\in \mathbb{R}^n,
$$
and according to \cite[Lemma 4.2]{BFRT} we get
$$
\frac{d}{dx}\overline{G}(x)=-\int_\mathbb{R}\partial_y (f(x-y))\overline{\Phi}(y)dy\\
=-\lim_{\varepsilon \rightarrow 0^+}\int_{|y|>\varepsilon}\partial_y(f(x-y))\overline{\Phi}(y)dy,\quad \mbox{ for almost all }x\in \mathbb{R}.
$$
By partial integration we obtain
$$
\int_{|y|>\varepsilon}\partial_y(f(x-y))\overline{\Phi }(y)dy=-\int_{|y|>\varepsilon}f(x-y)\overline{\Phi}'(y)dy+f(x+\varepsilon )\overline{\Phi}(-\varepsilon)-f(x-\varepsilon)\overline{\Phi}(\varepsilon),\quad \varepsilon >0,\;x\in \mathbb{R}^n.
$$
Then
$$
\frac{d}{dx}\overline{G}(x)=\lim_{\varepsilon \rightarrow 0^+}\left(\int_{|y|>\varepsilon}f(x-y)\overline{\Phi}'(y)dy+\overline{J}_\varepsilon(x)\right),\quad \mbox{ for almost all }x\in \mathbb{R},
$$
where
$$
\overline{J}_\varepsilon(x)=f(x-\varepsilon)\overline{\Phi}(\varepsilon)-f(x+\varepsilon )\overline{\Phi}(-\varepsilon),\quad \varepsilon >0,\;x\in \mathbb{R}.
$$
By taking into account \eqref{derivH} and that $\widetilde{H}_\alpha (z)\leq Ce^{-cz^2}$, $z\in \mathbb{R}$, it follows that
$$
\overline{\Phi}'(y)=\frac{(-1)^{\alpha }}{2^{\frac{\alpha +1}{2}}\sqrt{\pi}}\int_0^\infty \widetilde{H}_\alpha \Big(\frac{y}{\sqrt{2t}}\Big)t^{-\frac{3}{2}}dt=\int_0^\infty (\frac{d^\alpha}{dx^\alpha}W_t)(y)t^{\frac{\alpha}{2}-1}dt,\quad y\in \mathbb{R}\setminus\{0\}.
$$
On the other hand we have that
\begin{align*}
\overline{\Phi}(y)&=\frac{(-1)^{\alpha-1}}{2^{\frac{\alpha}{2}}\sqrt{\pi}}\left(\int_0^1\widetilde{H}_{\alpha-1}\Big(\frac{y}{\sqrt{2t}}\Big)\frac{dt}{t}+\int_1^\infty \Big(\widetilde{H}_{\alpha -1}\Big(\frac{y}{\sqrt{2t}}\Big)-\widetilde{H}_{\alpha-1}(0)\Big)\frac{dt}{t}\right)\\
&=\frac{(-1)^{\alpha-1}}{2^{\frac{\alpha }{2}}\sqrt{\pi}}\left(\int_{\frac{y^2}{2}}^\infty \widetilde{H}_{\alpha -1}(\sqrt{s})\frac{ds}{s}+\int_0^{\frac{y^2}{2}}(\widetilde{H}_{\alpha-1}(\sqrt{s})-\widetilde{H}_{\alpha-1}(0))\frac{ds}{s}\right),\quad y\in \mathbb{R}\setminus\{0\}.
\end{align*}
Let $x\in \mathbb{R}$. Since $f\in C_c^\infty (\mathbb{R}^n)$ we get
$$
|f(x+\varepsilon)-f(x-\varepsilon)|\leq C\varepsilon,\quad \varepsilon >0.
$$
If $\alpha$ is odd, then $\Phi$ is even and, for every $\varepsilon \in (0,1)$,
\begin{align*}
|\overline{J}_\varepsilon(x)|&=|(f(x+\varepsilon)-f(x-\varepsilon))\overline{\Phi}(\varepsilon)|\leq C\varepsilon \left(\int_{\frac{\varepsilon ^2}{2}}^\infty |\widetilde{H}_{\alpha-1}(\sqrt{s})|\frac{ds}{s}+\int_0^{\frac{\varepsilon ^2}{2}}|\widetilde{H}_{\alpha -1}(\sqrt{s})-\widetilde{H}_{\alpha-1}(0)|\frac{ds}{s}\right)\\
&\leq C\varepsilon \left(\int_1^\infty e^{-s}ds+\int_{\frac{\varepsilon ^2}{2}}^1 \frac{ds}{s}+\int_0^{\frac{\varepsilon ^2}{2}}\sqrt{s}\frac{ds}{s}\right)\leq C\varepsilon \big(1+|\log\varepsilon|+\varepsilon \big).
\end{align*}
Hence, if $\alpha$ is odd, $\lim_{\varepsilon \rightarrow 0^+}\overline{J}_\varepsilon(x)=0$.
If $\alpha$ is even, then $\overline{\Phi}$ is odd, $H_{\alpha-1}(0)=0$, and
$$
\overline{\Phi }(\varepsilon)=-\frac{1}{2^{\frac{\alpha}{2}}\sqrt{\pi}}\int_0^\infty \widetilde{H}_{\alpha-1}(\sqrt{s})\frac{ds}{s},\quad \varepsilon >0.
$$
We obtain
$$
\lim_{\varepsilon \rightarrow 0^+}\overline{J}_\varepsilon(x)=-\frac{f(x)}{2^{\frac{\alpha}{2}-1}\sqrt{\pi}}\int_0^\infty \widetilde{H}_{\alpha-1}(\sqrt{s})\frac{ds}{s},
$$
provided that $\alpha$ is even.
We conclude that, for certain $c_\alpha \in \mathbb{R}$
$$
\frac{d}{dx}G(x)=\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon}f(y)\int_0^\infty(\frac{d^\alpha}{dx^\alpha}W_t)(x-y)t^{\frac{\alpha}{2}-1}dtdy+c_\alpha f(x),\quad \mbox{for almost all }x\in \mathbb{R}.
$$
Then, we get
$$
\frac{d^\alpha}{dx^\alpha}\mathcal{A}^{-\alpha/2}(f)(x)=\frac{1}{\Gamma (\frac{\alpha}{2})}\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon }f(y)\int_0^\infty \partial_x^\alpha T_t^\mathcal{A}(x,y)t^{\frac{\alpha}{2}-1}dtdy+c_\alpha f(x),\quad \mbox{for almost all }x\in \mathbb{R},
$$
where $c_\alpha=0$ when $\alpha$ is odd.
Finally we observe that when $\alpha$ is even
\begin{align*}
\int_0^\infty (\frac{d^\alpha}{dz^\alpha}W_t)(z)t^{\frac{\alpha}{2}-1}dt&=\frac{1}{2^{\frac{\alpha +1}{2}}\sqrt{\pi}}\int_0^\infty e^{-\frac{z^2}{2t}}H_\alpha \Big(\frac{|z|}{\sqrt{2t}}\Big)t^{-\frac{3}{2}}dt=\frac{1}{2^{\frac{\alpha }{2}-1}\sqrt{\pi}}\frac{1}{|z|}\int_0^\infty e^{-s^2}H_\alpha (s)ds=0.
\end{align*}
By taking into account the arguments developed in this proof we can see that
$$
\int_{\mathbb{R}}\left|f(y)\int_0^\infty \partial_x^\alpha T_t^\mathcal{A}(x,y)t^{\frac{\alpha}{2}-1}dt\right|dy=\int_{\mathbb{R}}\left|f(y)\int_0^\infty \partial_x^\alpha \big(T_t^\mathcal{A}(x,y)-W_t(x-y)\big)t^{\frac{\alpha}{2}-1}dt\right|dy<\infty,
$$
for every $x\in \mathbb{R}$. Then,
$$
\frac{d^\alpha}{dx^\alpha}\mathcal{A}^{-\alpha/2}(f)(x)=\frac{1}{\Gamma (\frac{\alpha}{2})}\int_{\mathbb{R}}f(y)\int_0^\infty \partial_x^\alpha T_t^\mathcal{A}(x,y)t^{\frac{\alpha}{2}-1}dtdy+c_\alpha f(x),\quad \mbox{ for almost all }x\in \mathbb{R}.
$$
The proof of Theorem \ref{Th1.1} is completed.
\subsection{Proof of Theorem \ref{Th1.2}.}
For every $f\in C_c^\infty (\mathbb{R}^n)$ we have that
$$
\mathcal{A}^{-|\alpha|/2}(f)(x)=\sum_{k\in \mathbb{N}^n}\frac{c_k(f)}{(|k|+n)^{\frac{|\alpha|}{2}}}\widetilde{H}_k(x),\quad x\in \mathbb{R}^n,
$$
and according to \eqref{2.1}, \eqref{2.2} and \eqref{2.3} the last series is pointwise absolutely convergent, it defines a smooth function on $\mathbb{R}^n$ and
$$
\partial _x^\alpha\mathcal{A}^{-|\alpha|/2}(f)(x)=\sum_{k\in \mathbb{N}^n}\frac{c_k(f)}{(|k|+n)^{\frac{|\alpha|}{2}}}\partial _x^\alpha\widetilde{H}_k(x)=(-1)^{|\alpha|}\sum_{k\in \mathbb{R}^n}\frac{c_k(f)}{(|k|+n)^{\frac{|\alpha|}{2}}}\widetilde{H}_{k+\alpha}(x),\quad x\in \mathbb{R}^n.
$$
Then, according to Theorem \ref{Th1.1}, for each $f\in C_c^\infty (\mathbb{R}^n)$,
\begin{equation}\label{limC_c}
R_\alpha (f)(x)=\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon}R_\alpha (x,y)f(y)dy+c_\alpha f(x),\quad \mbox{for almost all }x\in \mathbb{R}^n.
\end{equation}
Here $c_\alpha =0$ when $\alpha _i$ is odd for some $i=1,...,n$.
To establish our result it is sufficient to show that for every $f \in L^p(\mathbb{R}^n,\gamma_{-1})$ the limit
$$
\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon}R_\alpha (x,y)f(y)dy
$$
exists for almost $x\in \mathbb{R}^n$ and the operator $L_\alpha $ defined by
\begin{equation}\label{Lalpha}
L_\alpha (f)(x)=\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon}R_\alpha (x,y)f(y)dy+c_\alpha f(x)
\end{equation}
is bounded from $L^p(\mathbb{R}^n,\gamma_{-1})$ into itself. Thus, $L_\alpha$ is the unique extension of $R_\alpha$ from $L^2(\mathbb{R}^n,\gamma_{-1})\cap L^p(\mathbb{R}^n,\gamma_{-1})$ to $L^p(\mathbb{R}^n,\gamma_{-1})$ as a bounded operator from $L^p(\mathbb{R}^n,\gamma_{-1})$ into itself.
For every $\beta >0$ we define the set
$$
N_\beta =\Big\{(x,y)\in \mathbb{R}^n\times\mathbb{R}^n:|x-y|\leq \beta n\min\big\{1,\frac{1}{|x|}\big\}\Big\}.
$$
Observe that $N_1=N$ and that if $\beta >0$, $a \in (0,1)$ and $(x,y)\in N_\beta^c$, then
$$
|a x-a y|\geq a \beta n\min\Big\{1,\frac{1}{|x|}\Big\}\geq a^2 \beta n\min\Big\{1,\frac{1}{|ax|}\Big\},
$$
that is, $(a x,a y)\in N_{a^2 \beta}^c$. In particular we have that if $a\in (0,1)$ then $(a x,a y)\in N^c$ provided that $(x,y)\in N_{1/a^2}^c$.
Let $\beta >0$. We consider the operators $R_{\alpha,{\rm loc}}$ and $R_{\alpha,{\rm glob}}$ defined on $C_c^\infty(\mathbb{R}^n)$ by
\begin{equation}\label{RlocRglob}
R_{\alpha,{\rm loc}}(f)(x)=R_\alpha (f\chi_{N_\beta}(x,\cdot))(x),\quad R_{\alpha,{\rm glob}}(f)(x)=R_\alpha (f\chi_{N_\beta ^c}(x,\cdot))(x),\quad x\in \mathbb{R}^n.
\end{equation}
We recall that
$$
R_\alpha (x,y)=\frac{(-1)^{|\alpha|}}{\pi ^{\frac{n}{2}}\Gamma (\frac{|\alpha |}{2})}\int_0^\infty \frac{e^{-nt}}{(1-e^{-2t})^{\frac{n+|\alpha|}{2}}}\widetilde{H}_\ell \Big(\frac{x-e^{-t}y}{\sqrt{1-e^{-2t}}}\Big)t^{\frac{|\alpha|}{2}-1}dt,\quad x,y\in \mathbb{R}^n,\;x\not=y.
$$
Then,
\begin{equation}\label{acotRalphab}
|R_\alpha (x,y)|\leq C\int_0^\infty \frac{e^{-nt}e^{-\eta \frac{|x-e^{-t}y|^2}{1-e^{-2t}}}}{(1-e^{-2t})^{\frac{n+|\alpha|}{2}}}t^{\frac{|\alpha|}{2}-1}dt,\quad x,y\in \mathbb{R}^n,\;x\not=y,
\end{equation}
for every $\eta \in (0,1)$. From now on we consider $1/p<\eta <1$ and $\beta =1/\eta$. Since $|x-ry|^2=|y-rx|^2+(1-r^2)(|x|^2-|y|^2)$, $x,y\in \mathbb{R}^n$, $r\in \mathbb{R}$, by making the change of variables $s=1-e^{-2t}$ in the last integral we get
$$
|R_\alpha (x,y)|\leq Ce^{\eta (|y|^2-|x|^2)}\int_0^1\frac{e^{-\eta \frac{|y-x\sqrt{1-s}|^2}{s}}}{s^{\frac{n+|\alpha|}{2}}}(1-s)^{\frac{n}{2}-1}(-\log (1-s))^{\frac{|\alpha|}{2}-1}ds,\quad x,y\in \mathbb{R}^n,\;x\not=y.
$$
Assume first that $(x,y)\in N_\beta^c$. Then, $(\sqrt{\eta}x,\sqrt{\eta}y)\in N^c$. By proceeding as in the proof of \eqref{Mbeta} it follows that, when $\langle x,y\rangle\leq 0$,
$$
|R_\alpha (x,y)|\leq Ce^{-\eta |x|^2}\int_0^\infty \frac{e^{-\frac{r}{2}}(1+r)^{\frac{n+|\alpha|}{2}-\frac{1}{4}}}{r^{\frac{3}{4}}}dr\leq Ce^{-\eta|x|^2}.
$$
Suppose now that $\langle x,y\rangle>0$. Again, as in the estimation in \eqref{Mbeta}, since $(\sqrt{\eta}x,\sqrt{\eta}y)\in N^c$ we obtain
$$
\int_{\frac{1}{2}}^1\frac{e^{-\eta \frac{|y-x\sqrt{1-s}|^2}{s}}}{s^{\frac{n+|\alpha|}{2}}}(1-s)^{\frac{n}{2}-1}(-\log (1-s))^{\frac{|\alpha|}{2}-1}ds\leq C\Big(\frac{|x+y|}{|x-y|}\Big)^{\frac{n}{2}}\exp\Big(\frac{\eta }{2}\big(y|^2-|x|^2-|x+y||x-y|\big)\Big).
$$
On the other hand, proceeding as in \cite[p. 862]{PeSo} and considering the notation in \eqref{uPeSo} and \cite[Lemma 2.3]{PeSo} we have
\begin{align*}
\int_0^{\frac{1}{2}}\frac{e^{-\eta \frac{|y-x\sqrt{1-s}|^2}{s}}}{s^{\frac{n+|\alpha|}{2}}}(1-s)^{\frac{n}{2}-1}(-\log (1-s))^{\frac{|\alpha|}{2}-1}ds&
\leq C\sup_{s\in (0,1)}\Big(\frac{e^{-\eta u(s)}}{s^{\frac{n}{2}}}\Big)^{1-\frac{1}{n}}\int_0^{\frac{1}{2}}\frac{e^{-\eta \frac{u(s)}{n}}}{\sqrt{s}}\frac{(-\log (1-s))^{\frac{|\alpha|}{2}-1}}{s^{\frac{|\alpha|}{2}}}ds\\
&\hspace{-5cm}\leq C\Big(\frac{e^{-\eta u_0}}{s_0^{n/2}}\Big)^{1-\frac{1}{n}}\int_0^{\frac{1}{2}}\frac{e^{-\frac{u(s)}{n}}}{s^{\frac{3}{2}}}ds\leq C\Big(\frac{e^{-\eta u_0}}{s_0^{n/2}}\Big)^{1-\frac{1}{n}}\int_0^1\frac{e^{-\frac{u(s)}{n}}}{s^{\frac{3}{2}}\sqrt{1-s}}ds\leq C\Big(\frac{e^{-\eta u_0}}{s_0^{n/2}}\Big)^{1-\frac{1}{n}}\frac{e^{-\eta \frac{u_0}{n}}}{\sqrt{s_0}}\\
&\hspace{-5cm}\leq C\frac{e^{-\eta u_0}}{s_0^{n/2}}\leq C\Big(\frac{|x+y|}{|x-y|}\Big)^{\frac{n}{2}}\exp\Big(\frac{\eta }{2}\big(y|^2-|x|^2-|x+y||x-y|\big)\Big).
\end{align*}
From the above estimates we conclude that, when $(x,y)\in N_\beta ^c$,
\begin{equation}\label{acotRalpha}
|R_\alpha (x,y)|\leq C\left\{
\begin{array}{ll}
e^{-\eta |x|^2},&\langle x,y\rangle \leq 0,\\
\displaystyle \Big(\frac{|x+y|}{|x-y|}\Big)^{\frac{n}{2}}\exp\Big(\eta\Big(\frac{|y|^2-|x|^2}{2}-\frac{|x-y||x+y|}{2}\Big)\Big),& \langle x,y\rangle >0.
\end{array}
\right.
\end{equation}
On the other hand, we consider the kernel
\begin{equation}\label{Rclassical}
\mathbb{R}_\alpha (x,y)=\frac{1}{\Gamma(\frac{|\alpha|}{2})}\int_0^\infty\partial _x^\alpha W_t(x-y)t^{\frac{|\alpha|}{2}-1}dt,\quad x,y\in \mathbb{R}^n,\;x\not=y.
\end{equation}
Let us show that
\begin{equation}\label{acotdif}
|R_\alpha (x,y)-\mathbb{R}_\alpha (x,y)|\leq C\frac{\sqrt{1+|x|}}{|x-y|^{n- \frac{1}{2}}},\quad (x,y)\in N_\beta , \;x\not=y .
\end{equation}
For every $x\in \mathbb{R}^n$, we have that
\begin{align*}
R_\alpha (x,y)-\mathbb{R}_\alpha (x,y)&=\frac{1}{\Gamma(\frac{|\alpha|}{2})}\left(\int_0^{m(x)}+\int_{m(x)}^\infty\right) \partial _x^\alpha (T_t^\mathcal{A}
(x,y)-W_t(x-y))t^{\frac{|\alpha|}{2}-1}dt\\
&=I(x,y)+J(x,y),\quad x,y\in \mathbb{R}^n,\;x\not=y.
\end{align*}
The same proof of \eqref{acotI} allows us to obtain that
$$
|I(x,y)|\leq C\frac{\sqrt{1+|x|}}{|x-y|^{n-\frac{1}{2}}},\quad (x,y)\in N_\beta,\;x\not=y.
$$
Also, we get
\begin{align*}
|J(x,y)|&\leq C\left(\int_{m(x)}^\infty \frac{e^{-nt}}{(1-e^{-2t})^{\frac{n+|\alpha|}{2}}}\Big|\widetilde{H}_\alpha \Big(\frac{x-ye^{-t}}{\sqrt{1-e^{-2t}}}\Big)\Big|t^{\frac{|\alpha|}{2}-1}dt+\int_{m(x)}^\infty \Big|\widetilde{H}_\alpha \Big(\frac{x-y}{\sqrt{2t}}\Big)\Big|\frac{dt}{t^{\frac{n}{2}+1}}\right)\\
&\leq C\left(\int_{m(x)}^\infty e^{-nt}\frac{e^{-c\frac{|x-e^{-t}y|^2}{1-e^{-2t}}}}{(1-e^{-2t})^{\frac{n+|\alpha|}{2}}}t^{\frac{|\alpha|}{2}-1}dt+\int_{m(x)}^\infty \frac{e^{-c\frac{|x-y|^2}{2t}}}{t^{\frac{n}{2}+1}}dt\right)\leq C\int_{m(x)}^\infty \frac{dt}{t^{\frac{n}{2}+1}}\\
&= \frac{C}{m(x)^\frac{n}{2}}\leq C(1+|x|)^n\leq C\frac{\sqrt{1+|x|}}{|x-y|^{n- \frac{1}{2}}},\quad (x,y)\in N_\beta ,\; x\not=y,
\end{align*}
and thus, \eqref{acotdif} is established.
From \eqref{acotRalpha} we obtain that $R_{\alpha,{\rm glob}}$, can be extended to $L^p(\mathbb{R}^n,\gamma_{-1})$ as a bounded operator from $L^p(\mathbb{R}^n,\gamma_{-1})$ into itself, and the extension is given by \eqref{RlocRglob}. Indeed, when $(x,y)\in N_\beta ^c$, we have that $|x-y||x+y|\geq C$. By taking into account also that $||y|^2-|x|^2|\leq |x+y||x-y|$ we get
\begin{align*}
\int_{\mathbb{R}^n}|R_\alpha (x,y)|e^{\frac{|x|^2-|y|^2}{p}}\chi _{N_\beta ^c}(x,y)dy&\\
&\hspace{-4cm}\leq C\left(\int_{\mathbb{R}^n}e^{-(\eta -\frac{1}{p})|x|^2-\frac{|y|^2}{p}}dy+\int_{\mathbb{R}^n}|x+y|^n\exp\Big(\Big(\frac{\eta}{2}-\frac{1}{p}\Big)(|y|^2-|x|^2)-\frac{\eta }{2}|x-y||x+y|\Big)dy\right)\\
&\hspace{-4cm}\leq C\left(e^{-(\eta -\frac{1}{p})|x|^2}\int_{\mathbb{R}^n}e^{-\frac{|y|^2}{p}}dy+\int_{\mathbb{R}^n}|x+y|^n\exp\Big(-|x-y||x+y|\big(\frac{\eta}{2}-\big|\frac{1}{p}-\frac{\eta}{2}\big|\big)\Big)dy\right),\quad x\in \mathbb{R}^n.
\end{align*}
Since $\eta >1/p$ it follows that
$$
\sup_{x\in \mathbb{R}^n}\int_{\mathbb{R}^n}|R_\alpha (x,y)|e^{\frac{|x|^2-|y|^2}{p}}\chi_{N_\beta ^c}(x,y)dy<\infty,
$$
and in a similar way
$$
\sup_{y\in \mathbb{R}^n}\int_{\mathbb{R}^n}|R_\alpha (x,y)|e^{\frac{|x|^2-|y|^2}{p}}\chi_{N_\beta ^c}(x,y)dx<\infty,
$$
from which we deduce that $R_{\alpha,{\rm glob}}$ is bounded on $L^p(\mathbb{R}^n,\gamma_{-1})$.
On the other hand, by using \eqref{acotdif} and that $\sqrt{m(x)}\sim \frac{1}{1+|x|}$, $x\in \mathbb{R}^n$, we have that
\begin{align*}
\int_{\mathbb{R}^n}|R_\alpha (x,y)-\mathbb{R}_\alpha (x,y)|\chi_{N_\beta }(x,y)dy&\leq C\sqrt{1+|x|}\int_{\mathbb{R}^n}\frac{1}{|x-y|^{n-\frac{1}{2}}}\chi_{N_\beta }(x,y)dy\\
&\hspace{-3cm}=C\sqrt{1+|x|}\int_0^{\beta n\sqrt{m(x)}}r^{-\frac{1}{2}}dr=C\sqrt{(1+|x|)\sqrt{m(x)}}\leq C,\quad x\in \mathbb{R}^n.
\end{align*}
Also, since $m(x)\sim m(y)$, $(x,y)\in N_\beta$, we get
$$
\sup_{y\in \mathbb{R}^n}\int_{\mathbb{R}^n}|R_\alpha (x,y)-\mathbb{R}_\alpha (x,y)|\chi_{N_\beta }(x,y)dx<\infty.
$$
Hence the operator $S_\alpha$ defined by
$$
S_\alpha(f)(x)=\int_{\mathbb{R}^n}(R_\alpha (x,y)-\mathbb{R}_\alpha (x,y))\chi_{N_\beta}(x,y)f(y)dy,\quad x\in \mathbb{R}^n,
$$
is a bounded operator from $L^p(\mathbb{R}^n,dx)$ into itself. Since $S_\alpha$ is a local operator, by \cite[Proposition 3.2.5]{Sa} $S_\alpha$ is bounded from $L^p(\mathbb{R}^n, \gamma_{-1})$ into itself.
We now observe that the kernel $\mathbb{R}_\alpha$ is a standard Calder\'on-Zygmund kernel. Indeed, we get
$$
|\mathbb{R}_\alpha (x,y)|\leq C\int_0^\infty \Big|H_\alpha \big(\frac{x-y}{\sqrt{2t}}\big)\Big|\frac{e^{-\frac{|x-y|^2}{2t}}}{t^{\frac{n}{2}+1}}dt\leq C\int_0^\infty \frac{e^{-c\frac{|x-y|^2}{t}}}{t^{\frac{n}{2}+1}}dt\leq \frac{C}{|x-y |^n},\quad x,y\in \mathbb{R}^n,\;x\not =y.
$$
Let $i=1,...,n$ and denote $\alpha^i=(\alpha _1,...,\alpha_i+1,...,\alpha _n)$. We have that
$$
\partial_{x_i}\mathbb{R}_\alpha (x,y)=\frac{(-1)^{|\alpha|+1}}{(2\pi )^{\frac{n}{2}}2^{\frac{|\alpha|+1}{2}}\Gamma (\frac{|\alpha|}{2})}\int_0^\infty H_{\alpha ^i}\Big(\frac{x-y}{\sqrt{2t}}\Big)\frac{e^{-\frac{|x-y|^2}{2t}}}{t^{\frac{n+3}{2}}}dt,\quad x,y\in \mathbb{R}^n,\;x\not=y.
$$
Then,
$$
\Big|\partial _{x_i}\mathbb{R}_\alpha (x,y)\Big|\leq \frac{C}{|x-y |^{n+1}},\quad x,y\in \mathbb{R}^n,\;x\not =y.
$$
The Euclidean $\alpha$-order Riesz transform $\mathbb{R}_\alpha$ is bounded from $L^q(\mathbb{R}^n,dx)$ into itself, for every $1<q<\infty$. According to \cite[Proposition 3.2.5]{Sa} the operator $\mathbb{R}_{\alpha ,{\rm loc}}$ defined by
$$
\mathbb{R}_{\alpha ,{\rm loc}}(f)(x)=\mathbb{R}_{\alpha }(f\chi _{N_\beta }(x,\cdot))(x)
$$
is bounded from $L^q(\mathbb{R}^n,\gamma_{-1})$ into itself, for every $1<q<\infty$.
We can write $R_{\alpha,{\rm loc}}=S_\alpha (f)+\mathbb{R}_{\alpha ,{\rm loc}}(f)$ on $C_c^\infty(\mathbb{R}^n)$. Then, $R_{\alpha,{\rm loc}}$ can be extended from $C_c^\infty (\mathbb{R}^n)$ to $L^p(\mathbb{R}^n,\gamma_{-1})$ as a bounded operator from $L^p(\mathbb{R}^n,\gamma_{-1})$ into itself.
Since $R_\alpha=R_{\alpha,{\rm loc}}+R_{\alpha,{\rm glob}}$ we conclude that $R_\alpha$ can be extended from $C_c^\infty (\mathbb{R}^n)$ to $L^p(\mathbb{R}^n,\gamma_{-1})$ as a bounded operator from $L^p(\mathbb{R}^n,\gamma_{-1})$ into itself.
Let us consider the maximal operator
$$
R_{\alpha,*}(f)(x)=\sup_{\varepsilon >0}\left|\int_{|x-y|>\varepsilon }R_\alpha (x,y)f(y)dy\right|,\quad f\in L^p(\mathbb{R}^n,\gamma_{-1}),\;1<p<\infty.
$$
Let $f\in L^p(\mathbb{R}^n,\gamma_{-1})$, $1<p<\infty$. For every $\varepsilon >0$, by using the above estimates we can write
\begin{align*}
\left|\int_{|x-y|>\varepsilon }R_\alpha (x,y)f(y)dy\right|& \leq \int_{|x-y|>\varepsilon}|R_\alpha (x,y)-\mathbb{R}_\alpha (x,y)|\chi _{N_\beta }(x,y)|f(y)|dy\\
&\hspace{-3cm} +\int_{|x-y|>\varepsilon}|R_\alpha (x,y)|\chi _{N_\beta ^c}(x,y)|f(y)|dy+\int_{|x-y|>\varepsilon}|\mathbb{R}_\alpha (x,y)|\chi _{N_\beta }(x,y)|f(y)|dy<\infty ,\quad x\in \mathbb{R}^n.
\end{align*}
We also have that
\begin{align*}
R_{\alpha ,*}(f)(x)&\leq C\left(\int_{\mathbb{R}^n}\frac{\sqrt{1+|x|}}{|x-y|^{n-\frac{1}{2}}}\chi_{N_\beta }(x,y)|f(y)|dy+\int_{\mathbb{R}^n}|R_\alpha (x,y)|\chi _{N_\beta ^c}(x,y)|f(y)|dy\right.\\
&\left.\quad +\sup_{\varepsilon >0}\left|\int_{|x-y|>\varepsilon}\mathbb{R}_\alpha (x,y)\chi _{N_\beta }(x,y)f(y)dy\right|\right),\quad x\in \mathbb{R}^n.
\end{align*}
Since the maximal operator
$$
\mathbb{R}_{\alpha ,*}(f)(x)=\sup_{\varepsilon >0}\left|\int_{|x-y|>\varepsilon}\mathbb{R}_\alpha (x,y)f(y)dy\right|
$$
is bounded from $L^p(\mathbb{R}^n,dx)$ into itself, by using a vector valued version of \cite[Proposition 3.2.5]{Sa} (see \cite[Proposition 2.3]{HTV}), we deduce that the local maximal operator
$$
\mathbb{R}_{\alpha,{\rm loc }, *}(f)(x)=\sup_{\varepsilon >0}\left|\int_{|x-y|>\varepsilon}\mathbb{R}_\alpha (x,y)\chi _{N_\beta}(x,y)f(y)dy\right|
$$
is bounded from $L^p(\mathbb{R}^n,\gamma_{-1})$ into itself.
By using the same arguments as above we conclude that $R_{\alpha ,*}$ is bounded from $L^p(\mathbb{R}^n,\gamma_{-1})$ into itself.
From \eqref{limC_c} and since $C_c^\infty ( \mathbb{R}^n)$ is dense in $L^p(\mathbb{R}^n,\gamma_{-1})$ and $\mathbb{R}_{\alpha ,*}$ is bounded in $L^p(\mathbb{R}^n,\gamma_{-1})$, by using a standard procedure we can conclude
that the limit
$$
\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon}R_\alpha (x,y)f(y)dy
$$
exists for almost $x\in \mathbb{R}^n$ and $L_\alpha$ (defined in \eqref{Lalpha}) is a bounded operator from $L^p(\mathbb{R}^n,\gamma_{-1})$ into itself.
\begin{rem}The $L^p(\mathbb{R}^n,\gamma_{-1})$-boundedness of the local part $R_{\alpha,{\rm loc}}$ of $R_\alpha$ can be proved also by using Calder\'on-Zygmund theory. We have preferred to do it by comparing $R_{\alpha,{\rm loc}}$ with the classical local Riesz transform $\mathbb{R}_{\alpha,{\rm loc}}$ because in this way we can know how the singularity of $R_{\alpha,{\rm loc}}$ is. Furthermore, these comparative results will be useful in the proof of Theorem \ref{Th1.4} and \ref{Th1.5}.
\end{rem}
\section{Riesz transform associated with the operator $\bar{\mathcal{A}}$}\label{barA}
We define $L_0^2(\mathbb{R}^n,\gamma _{-1})$ as the space that consists of all those $f\in L ^2(\mathbb{R}^n,\gamma _{-1})$ such that $c_0(f)=\int_{\mathbb{R}^n}f(x)dx=0$. Let $\beta >0$. For every $f\in L_0^2(\mathbb{R}^n,\gamma _{-1})$, $\bar{\mathcal{A}}^{-\beta}f$ is defined by
$$
\bar{\mathcal{A}}^{-\beta }(f)=\sum_{k\in \mathbb{N}^n\setminus\{0\}}\frac{c_k(f)}{|k|^\beta}\widetilde{H}_k.
$$
We have that
$$
\|\bar{\mathcal{A}}^{-\beta }(f)\|_{L^2(\mathbb{R}^n,\gamma_{-1})}^2=\sum_{k\in \mathbb{N}^n\setminus\{0\}}\frac{|c_k(f)|^2\|\widetilde{H}_k\|_{L^2(\mathbb{R}^n,\gamma_{-1})}^2}{|k|^{2\beta}}\leq \|f\|^2_{L^2(\mathbb{R}^n,\gamma_{-1})},\quad f \in L_0^2(\mathbb{R}^n,\gamma_{-1}).
$$
We introduce the operator $\overline{S}_\beta $ defined by
$$
\overline{S}_\beta (f)=\frac{1}{\Gamma (\beta)}\int_0^\infty(T_t^{\bar{\mathcal{A}}}(f)-c_0(f)e^{|\cdot|^2}) t^{\beta -1}dt,\quad f\in L^2(\mathbb{R}^n,\gamma_{-1}).
$$
Let $f\in L^2(\mathbb{R}^n,\gamma_{-1})$. We have that
$$
T_t^{\bar{\mathcal{A}}}(f)-c_0(f)e^{-|\cdot|^2}=\sum_{k\in \mathbb{N}^n}e^{-|k|t}c_k(f)\widetilde{H}_k-e^{|\cdot|^2}c_0(f)=\sum_{k\in \mathbb{N}^n\setminus\{0\}}e^{-|k|t}c_k(f)\widetilde{H}_k,\quad t>0.
$$
Then,
$$
\|T_t^{\bar{\mathcal{A}}}(f)-c_0(f)e^{-|\cdot|^2}\|_{L^2(\mathbb{R}^n,\gamma_{-1})}\leq e^{-t}\|f\|_{L^2(\mathbb{R}^n,\gamma_{-1})},\quad t>0.
$$
Hence, the integral defining $\overline{S}_\beta (f)$ converges in the $L^2(\mathbb{R}^n,\gamma_{-1})$-Bochner sense.
Let $f\in C_c^\infty(\mathbb{R}^n)$. According to \eqref{2.1}, \eqref{2.2} and \eqref{2.3} we get
$$
\sum_{k\in \mathbb{N}^n\setminus\{0\}}e^{-|k|t}|c_k(f)||\widetilde{H}_k(x)|\leq Ce^{-t}e^{-\frac{|x|^2}{2}}\sum_{k\in \mathbb{N}^n\setminus\{0\}}\frac{1}{|k|^2}\leq Ce^{-t}e^{-\frac{|x|^2}{2}},\quad t>0,\;x\in \mathbb{R}^n.
$$
It follows that that the series that defines $T_t^{\bar{\mathcal{A}}}(f)(x)-c_0(f)e^{-|x|^2}$ converges pointwisely and absolutely and
$$
\overline{S}_\beta(f)(x)=\frac{1}{\Gamma (\beta)}\sum_{k\in \mathbb{N}^n\setminus\{0\}}c_k(f)\widetilde{H}_k(x)\int_0^\infty e^{-|k|t}t^{\beta -1}dt=\sum_{k\in \mathbb{N}^n\setminus\{0\}}\frac{c_k(f)}{|k|^\beta}\widetilde{H}_k(x)=\bar{\mathcal{A}}^{-\beta}(f_0)(x),\quad x\in \mathbb{R}^n,
$$
where $f_0(x)=f(x)-c_0(f)e^{-|x|^2}$, $x\in \mathbb{R}^n$.
On the other hand, since $\supp f$ is compact we have that
\begin{align*}
\int_{m(x)}^\infty\int_{\mathbb{R}^n}\big|T_t^{\bar{\mathcal{A}}}(x,y)-e^{-|x|^2}\big||f(y)|dy t^{\beta -1}dt&\\
&\hspace{-4cm} \leq C\int_{m(x)}^\infty \left(\int_{\mathbb{R}^n}e^{-\frac{|x-e^{-t}y|^2}{1-e^{-2t}}}\left|(1-e^{-2t})^{-\frac{n}{2}}-1\right||f(y)|dy\right.\\
&\hspace{-4cm}\quad +\int_{\mathbb{R}^n}\left|e^{-\frac{|x-e^{-t}y|^2}{1-e^{-2t}}}-e^{-|x-e^{-t}y|^2}\right||f(y)|dy\left. +\int_{\mathbb{R}^n}\left|e^{-|x-e^{-t}y|^2}-\-e^{-|x|^2}\right||f(y)|dy\right)t^{\beta -1} dt\\
&\hspace{-4cm}\leq C\int_0^\infty e^{-t}t^{\beta -1}dt,\quad x\in \mathbb{R}^n,
\end{align*}
and, taking $0<\varepsilon <\min\{2\beta ,n\}$,
\begin{align*}
\int_0^{m(x)}t^{\beta -1}\int_{\mathbb{R}^n}\big|T_t^{\bar{\mathcal{A}}}(x,y)-e^{-|x|^2}\big||f(y)|dydt &\leq C\left(\int_0^{m(x)}t^{\beta -1}\int_{\mathbb{R}^n}\Big(e^{c|x|}\frac{e^{-c\frac{|x-y|^2}{t}}}{t^{\frac{n}{2}}}+e^{-|x|^2})|f(y)|dydt\right)\\
&\hspace{-2cm}\leq C\left(e^{c|x|}\int_{\mathbb{R}^n}\frac{|f(y)|}{|x-y|^{n-\varepsilon}}dy+m(x)^\beta e^{-|x|^2}\int_{\mathbb{R}^n}|f(y)|dy\right)\\
&\hspace{-2cm}\leq C\Big(e^{c|x|}+m(x)^\beta e^{-|x|^2}\Big),\quad x\in\mathbb{R}^n.
\end{align*}
Then, we obtain
$$
\overline{S}_\beta (f)(x)=\int_{\mathbb{R}^n}\overline{K}_\beta (x,y)f(y)dy,\quad x\in \mathbb{R}^n,
$$
where
$$
\overline{K}_\beta (x,y)=\frac{1}{\Gamma (\beta)}\int_0^\infty (T_t^{\bar{A}}(x,y)-e^{-|x|^2})t^{\beta -1}dt,\quad x,y\in \mathbb{R}^n,
$$
and
$$
T_t^{\bar{\mathcal{A}}}(x,y)=e^{nt}T_t^\mathcal{A}(x,y)=\frac{e^{|y|^2-|x|^2}}{\pi ^{\frac{n}{2}}(1-e^{-2t})^{\frac{n}{2}}}\exp\left(-\frac{|y-e^{-t}x|^2}{1-e^{-2t}}\right), \quad x,y\in \mathbb{R}^n, \;t>0.
$$
By denoting $\Pi _0$ the projection from $L^2(\mathbb{R}^n,\gamma_{-1})$ to $L_0^2(\mathbb{R}^n,\gamma_{-1})$ we have proved that, for every $f\in C_c^\infty (\mathbb{R}^n)$,
$$
\bar{\mathcal{A}}^{-\beta}\Pi _0(f)(x)=\int_{\mathbb{R}^n}\overline{K}_\beta (x,y)f(y)dy,\quad x\in \mathbb{R}^n.
$$
Let $f\in C_c^\infty (\mathbb{R}^n)$. Next we show that
\begin{equation}\label{deltaA}
\delta_x^\alpha \bar{\mathcal{A}}^{-|\alpha|/2}\Pi _0(f)(x)=\lim_{\varepsilon \rightarrow 0^+}\frac{1}{\Gamma (\frac{|\alpha|}{2})}\int_{|x-y|>\varepsilon }f(y)\int_0^\infty \delta_x^\alpha T_t^{\bar{A}}(x,y)t^{\frac{|\alpha|}{2}-1}dtdy+c_\alpha f(x),
\end{equation}
for almost all $x\in \mathbb{R}^n$. Here $c_\alpha\in \mathbb{R}$ and, when $\alpha _i$ is odd for some $i=1,...,n$, $c_\alpha =0$.
For every $\ell \in \mathbb{N}^n$,
\begin{align*}
\delta_x^\ell T_t^{\bar{\mathcal{A}}}(x,y)&=\frac{(-1)^{|\ell|}e^{-|x|^2}}{2^{|\ell|}}\partial ^\ell _x\big(e^{|x|^2}T_t^{\bar{\mathcal{A}}}(x,y)\big)\\
&=\frac{(-1)^{|\ell|}e^{|y|^2-|x|^2}}{2^{|\ell|}\pi ^{\frac{n}{2}}(1-e^{-2t})^{\frac{n+|\ell|}{2}}}e^{-|\ell|t}\widetilde{H}_\ell \left(\frac{y-e^{-t}x}{\sqrt{1-e^{-2t}}}\right),\quad x,y\in \mathbb{R}^n, \;t>0,
\end{align*}
Then, for each $\ell \in \mathbb{N}^n$,
$$
|\delta_x^\ell T_t^{\bar{A}}(x,y)|\leq Ce^{|y|^2-|x|^2}\frac{e^{-|\ell|t}e^{-c\frac{|y-e^{-t}x|^2}{1-e^{-2t}}}}{(1-e^{-2t})^{\frac{n+|\ell|}{2}}},\quad x,y\in \mathbb{R}^n,\;t>0.
$$
Since $\delta_x^\ell (e^{-|x|^2})=0$, $x\in \mathbb{R}^n$, when $\ell \in \mathbb{N}^n\setminus\{0\}$, by proceeding as in the proof of \eqref{derivA} we get that for every $\ell \in \mathbb{N}^n\setminus\{0\}$ and $k\in \mathbb{N}$ being $|\ell|<k$,
$$
\delta_x^\ell \bar{\mathcal{A}}^{-k/2}\Pi _0(f)(x)=\frac{1}{\Gamma (\frac{k}{2})}\int_{\mathbb{R}^n}f(y)\int_0^\infty \delta_x^\ell T_t^{\bar{A}}(x,y)t^{\frac{k}{2}-1}dtdy,\quad\mbox{ for almost all }x\in \mathbb{R}^n.
$$
Without loss of generality we can assume that $\alpha _1\geq 1$ and consider $\ell =(\alpha _1-1,\alpha_2,...,\alpha_n)$. When $\ell \in \mathbb{N}^n\setminus\{0\}$, we can proceed as in the proof of Theorem \ref{Th1.1}. For $n>1$ we write
\begin{align}\label{decomp100n}
\delta_x^\alpha \bar{\mathcal{A}}^{-|\alpha|/2}\Pi _0(f)(x)&=\frac{1}{\Gamma (\frac{|\alpha|}{2})}\delta_{x_1}\left(\int_{\mathbb{R}^n}f(y)\int_0^\infty \delta_x^\ell \big(T_t^{\bar{\mathcal{A}}}(x,y)-e^{|y|^2-|x|^2}W_t(y-x)\big)t^{\frac{|\alpha|}{2}-1}dtdy\right.\nonumber\\
&\quad +\left. \int_{\mathbb{R}^n}f(y)\int_0^\infty \delta_x^\ell (e^{|y|^2-|x|^2}W_t(y-x))t^{\frac{|\alpha|}{2}-1}dtdy\right)=\frac{1}{\Gamma (\frac{|\alpha|}{2})}\delta_{x_1}\big(F(x)+G(x)\big),\quad x\in \mathbb{R}^n.
\end{align}
We observe that, for every $r\in \mathbb{N}^n$,
$$
\delta_x^r (e^{|y|^2-|x|^2}W_t(y-x))=\frac{(-1)^{|r|}}{2^{|r|}}e^{|y|^2-|x|^2}\partial_x^r(W_t(y-x))=\frac{(-1)^{|r|}}{2^{|r|}\pi ^{\frac{n}{2}}}\frac{e^{|y|^2-|x|^2}}{(2t)^{\frac{n+|r|}{2}}}\widetilde{H}_r \Big(\frac{y-x}{\sqrt{2t}}\Big),\quad x,y\in \mathbb{R}^n,\;t>0.
$$
By considering the decomposition
\begin{align}\label{decomp}
\delta_x^\alpha \big(T_t^{\bar{\mathcal{A}}}(x,y)-e^{|y|^2-|x|^2}W_t(y-x)\big)&=\frac{(-1)^{|\alpha|}}{2^{|\alpha|}\pi ^{\frac{n}{2}}}e^{|y|^2-|x|^2}\left(\frac{e^{-|\alpha|t}}{(1-e^{-2t})^{\frac{n+|\alpha|}{2}}}\widetilde{H}_\ell \left(\frac{y-e^{-t}x}{\sqrt{1-e^{-2t}}}\right)-\frac{1}{(2t)^{\frac{n+|\alpha|}{2}}}\widetilde{H}_\alpha \Big(\frac{y-x}{\sqrt{2t}}\Big)\right)\nonumber\\
&\hspace{-4cm}=\frac{(-1)^{|\alpha|}}{2^{|\alpha|}\pi ^{\frac{n}{2}}}e^{|y|^2-|x|^2}\left\{\frac{e^{-|\alpha|t}-1}{(1-e^{-2t})^{\frac{n+|\alpha|}{2}}}\widetilde{H}_\alpha\Big(\frac{y-e^{-t}x}{\sqrt{1-e^{-2t}}}\Big)+\Big(\frac{1}{(1-e^{-2t})^{\frac{n+|\alpha|}{2}}}-\frac{1}{(2t)^{\frac{n+|\alpha|}{2}}}\Big)\widetilde{H}_\alpha\left(\frac{y-e^{-t}x}{\sqrt{1-e^{-2t}}}\right)\right.\\
&\hspace{-4cm}\quad \left.+\frac{1}{(2t)^{\frac{n+|\alpha|}{2}}}\left(\widetilde{H}_\alpha\left(\frac{y-e^{-t}x}{\sqrt{1-e^{-2t}}}\right)-\widetilde{H}_\alpha \left(\frac{y-x}{\sqrt{2t}}\right)\right)\right\},\quad x,y\in \mathbb{R}^n.\nonumber
\end{align}
we can argue as in the proof of \eqref{derivF} to obtain that
$$
\delta_{x_1}F(x)=\int_{\mathbb{R}^n}f(y)\int_0^\infty \delta_x^\alpha \big(T_t^{\bar{\mathcal{A}}}(x,y)-e^{|y|^2-|x|^2}W_t(y-x)\big)t^{\frac{|\alpha|}{2}-1}dtdy, \quad \mbox{ for almost all }x\in \mathbb{R}^n.
$$
On the other hand, to deal with $\delta_{x_1}G(x)$ we consider $g(x)=f(x)e^{|x|^2}$ and $\Psi (x)=\int_0^\infty \partial_x^\ell (W_t)(-x)t^{\frac{|\alpha|}{2}-1}dt$, $x\in \mathbb{R}^n$, and write
$$
G(x)=\frac{e^{-|x|^2}}{2^{|\ell|}}\int_{\mathbb{R}^n}g(x-y)\Psi (y)dy,\quad x\in \mathbb{R}^n.
$$
We proceed as in the proof of \eqref{derivG} to get
$$
\delta_{x_1}G(x)=\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon}f(y)\int_0^\infty \delta _x^\alpha (e^{|y|^2-|x|^2}W_t(x-y))t^{\frac{|\alpha |}{2}-1}dtdy+c_\alpha f(x),\quad \mbox{ for almost all }x\in \mathbb{R}^n,
$$
where $c_\alpha \in \mathbb{R}$ and $c_\alpha=0$ if $\alpha _i$ is odd for some $i=1,...,n$. Thus \eqref{deltaA} is established when $n>1$.
If $n=1$ we can also follow the proof of Theorem \ref{Th1.1} by using the decomposition
\begin{equation}\label{decomp1001}
\delta_x^\alpha \bar{\mathcal{A}}^{-\alpha/2}\Pi _0(f)(x)=\frac{1}{\Gamma (\frac{|\alpha|}{2})}\delta_x (\overline{F}(x)+\overline{G}(x)),\quad x\in \mathbb{R}^n,
\end{equation}
where
$$
\overline{F}(x)=\int_{\mathbb{R}^n}f(y)\int_0^\infty \delta_x^\ell T_t^{\bar{\mathcal{A}}}(x,y)-\delta _x^\ell \big(e^{|y|^2-|x|^2}[W_t(y-x)-\frac{1}{2^\ell}\frac{d^\ell}{dx^\ell}W_t (0)\chi_{(1,\infty)}(t)]\big)t^{\frac{|\alpha|}{2}-1}dtdy,
,\quad x\in \mathbb{R}^n,
$$
and
$$
\overline{G}(x)=\int_{\mathbb{R}^n}f(y)\int_0^\infty \delta_x^\ell \big(e^{|y|^2-|x|^2}[W_t(y-x)-\frac{1}{2^\ell}\frac{d^\ell}{dx^\ell}W_t (0)\chi_{(1,\infty)}(t)]\big)t^{\frac{|\alpha|}{2}-1}dtdy,\quad x\in \mathbb{R}^n.
$$
When $\ell =0$, that is, $\alpha =(1,0,...,0)$, we can replace $F$ in \eqref{decomp100n} and $\overline{F}$ in \eqref{decomp1001} by
$$
F(x)=\int_{\mathbb{R}^n}f(y)\int_0^\infty \big(T_t^{\bar{\mathcal{A}}}(x,y)-e^{-|x|^2}-e^{|y|^2-|x|^2}W_t(y-x)\big)\frac{dtdy}{\sqrt{t}},\quad x\in \mathbb{R}^n,
$$
and
$$
\overline{F}(x)=\int_{\mathbb{R}^n}f(y)\int_0^\infty \Big(T_t^{\bar{\mathcal{A}}}(x,y)-e^{-|x|^2}-\big(e^{|y|^2-|x|^2}[W_t(y-x)-\frac{1}{2^\ell}\frac{d^\ell}{dx^\ell}W_t (0)\chi_{(1,\infty)}(t)]\Big)\frac{dtdy}{\sqrt{t}},\quad x\in \mathbb{R}^n,
$$
respectively, and proceed as above to obtain \eqref{deltaA}.
According to \eqref{2.1}, \eqref{2.2} and \eqref{2.3} we get
$$
\overline{R}_\alpha f(x)=\delta ^\alpha \mathcal{A}^{-|\alpha|/2}\Pi_0f(x),\quad x\in \mathbb{R}^n,
$$
and then,
$$
\overline{R}_\alpha f(x)=\lim_{\varepsilon \rightarrow 0^+}\frac{1}{\Gamma (\frac{|\alpha|}{2})}\int_{|x-y|>\varepsilon }f(y)\int_0^\infty \delta_x^\alpha T_t^{\bar{A}}(x,y)t^{\frac{|\alpha|}{2}-1}dtdy+c_\alpha f(x),\quad \mbox{ for almost all }x\in \mathbb{R}^n,
$$
with $c_\alpha=0$, when $\alpha_i$ is odd for some $i=1,...,n$.
We are going to show the $L^p(\mathbb{R}^n,\gamma_{-1})$-boundedness properties of $\overline{R}_\alpha $. We recall that
$$
\overline{R}_\alpha (x,y)=\frac{(-1)^{|\alpha|}}{2^{|\alpha|}\pi ^{\frac{n}{2}}\Gamma(\frac{|\alpha|}{2})}e^{|y|^2-|x|^2}
\int_0^\infty \frac{e^{-|\alpha|t}}{(1-e^{-2t})^{\frac{n+|\alpha|}{2}}}\widetilde{H}_\alpha\Big(\frac{y-e^{-t}x}{\sqrt{1-e^{-2t}}}\Big)t^{\frac{|\alpha|}{2}-1}dt,\quad x,y\in \mathbb{R}^n,\quad x\not=y.
$$
Consider first $1<p<\infty$ and choose $1-\frac{1}{p}<\eta <1$. By making the change of variables $s=1-e^{-2t}$, $t\in (0,\infty)$, we obtain
$$
|\overline{R}_\alpha (x,y)|\leq Ce^{|y|^2-|x|^2}\int_0^1\frac{e^{-\eta\frac{|y-x\sqrt{1-s}|^2}{s}}}{s^{\frac{n+|\alpha|}{2}}}(1-s)^{\frac{|\alpha|}{2}-1}(-\log(1-s))^{\frac{|\alpha|}{2}-1}ds,\quad x,y\in \mathbb{R}^n,\;x\not=y.
$$
Let $\beta=\eta^{-1}$ and consider the local and global operators defined on $C_c^\infty (\mathbb{R}^n)$ by
$$
\overline{R}_{\alpha ,{\rm loc}}(f)(x)=\overline{R}_\alpha (f\chi _{N_\beta}(x, \cdot ))(x),\quad \mbox{ and }\quad \overline{R}_{\alpha ,{\rm glob}}(f)(x)=\overline{R}_\alpha (f\chi _{N_\beta^c}(x,\cdot ))(x),\quad x\in \mathbb{R}^n.
$$
By proceeding as in the proof of \eqref{acotRalpha} it follows that, for each $(x,y)\in N_\beta ^c$,
$$
\overline{R}_\alpha (x,y)|\leq C\left\{\begin{array}{ll}
\displaystyle e^{(1-\eta)|y|^2-|x|^2}, & \langle x,y\rangle \leq 0, \\[0,5cm]
\Big(\frac{|x+y|}{|x-y|}\Big)^n\exp\big((1-\frac{\eta}{2})(|y|^2-|x|^2)-\frac{\eta }{2}|x+y||x-y|\big), & \langle x,y\rangle >0.
\end{array}
\right.
$$
We have that
\begin{align*}
\int_{\mathbb{R}^n}e^{-\frac{|y|^2}{p}+\frac{|x|^2}{p}}|\overline{R}_\alpha (x,y)|\chi _{N_\beta ^c}(x,y)dy&\leq C\left(\int_{\mathbb{R}^n}e^{|y|^2(1-\eta -\frac{1}{p})}e^{(\frac{1}{p}-1)|x|^2}dy\right.\\
&\left.+\int_{\mathbb{R}^n}|x+y|^n\exp \Big(-|x+y||x-y|\big(\frac{\eta }{2}-|1-\frac{1}{p}-\frac{\eta}{2}|\big)\Big)dy\right),\quad x\in \mathbb{R}^n.
\end{align*}
Then, since $\eta >1-\frac{1}{p}$,
$$
\sup_{x\in \mathbb{R}^n}\int_{\mathbb{R}^n}e^{-\frac{|y|^2}{p}+\frac{|x|^2}{p}}|\overline{R}_\alpha (x,y)|\chi _{N_\beta ^c}(x,y)dy<\infty.
$$
Also we get
$$
\sup_{y\in \mathbb{R}^n}\int_{\mathbb{R}^n}e^{-\frac{|y|^2}{p}+\frac{|x|^2}{p}}|\overline{R}_\alpha (x,y)|\chi _{N_\beta ^c}(x,y)dx<\infty.
$$
We conclude that $\overline{R}_{\alpha ,{\rm glob}}$ is bounded from $L^p(\mathbb{R}^n,\gamma_{-1})$ into itself.
We are going to study the operator $\overline{R}_{\alpha, {\rm loc}}$. We write
\begin{align*}
\overline{R}_\alpha (x,y)&=\frac{1}{\Gamma (\frac{|\alpha|}{2})}\left(\int_0^\infty \delta _x^\alpha [T_t^\mathcal{\bar{A}}(x,y)-e^{|y|^2-|x|^2}W_t(y-x)]t^{\frac{|\alpha |}{2}-1}dt+\int_0^\infty \delta _x^\alpha[e^{|y|^2-|x|^2}W_t(y-x)]t^{\frac{|\alpha |}{2}-1}dt\right)\\
&=I(x,y)+J(x,y),\quad x,y\in \mathbb{R}^n, \;x\not=y.
\end{align*}
By taking into account that
$$
|\delta _x^\alpha [T_t^\mathcal{\bar{A}}(x,y)-e^{|y|^2-|x|^2}W_t(y-x)]|\leq Ce^{|y|^2-|x|^2}\left(\frac{e^{-|\alpha|t}}{(1-e^{-2t})^{\frac{n+|\alpha|}{2}}}+\frac{e^{-c\frac{|x-y|^2}{t}}}{t^{\frac{n+|\alpha|}{2}}}\right),\quad x,y\in \mathbb{R}^n,\;t>0,
$$
we get (see \eqref{dif1})
$$
\int_{m(x)}^\infty |\delta _x^\alpha [T_t^\mathcal{\bar{A}}(x,y)-e^{|y|^2-|x|^2}W_t(y-x)]|t^{\frac{|\alpha|}{2}-1}dt\leq Ce^{|y|^2-|x|^2}(1+|x|)^n,\quad x,y\in \mathbb{R}^n.
$$
Also, from \eqref{decomp} and proceeding as in the proof of \eqref{acotI} we can see that
$$
\int_0^{m(x)}|\delta _x^\alpha [T_t^\mathcal{\bar{A}}(x,y)-e^{|y|^2-|x|^2}W_t(y-x)]|t^{\frac{|\alpha|}{2}-1}dt\leq Ce^{|y|^2-|x|^2}\frac{\sqrt{1+|x|}}{|x-y|^{n-\frac{1}{2}}},\quad (x,y)\in N_\beta.
$$
Since $||y|^2-|x|^2|\leq C$ when $(x,y)\in N_\beta$, we obtain that
\begin{equation}\label{acotIb}
|I(x,y)|\leq C\frac{\sqrt{1+|x|}}{|x-y|^{n-\frac{1}{2}}},\quad (x,y)\in N_\beta,\;x\not=y.
\end{equation}
On the other hand, we observe that
$$
J(x,y)=\frac{e^{|y|^2-|x|^2}}{2^{|\alpha|}\Gamma(\frac{|\alpha|}{2})}\int_0^\infty \partial _y^\alpha (W_t(y-x))t^{\frac{|\alpha|}{2}-1}dt=2^{-|\alpha|}e^{|y|^2-|x|^2}\mathbb{R}_\alpha (y,x),\quad x,y\in \mathbb{R}^n,\;x\not=y,
$$
where $\mathbb{R}_\alpha (\cdot,\cdot)$ is the classical kernel considered in \eqref{Rclassical}. The result can be established by proceeding as in the proof of Theorem \ref{Th1.2} by taking into account that the Euclidean Riesz transform $\mathbb{R}_\alpha$ is bounded from $L^p(\mathbb{R}^n,dx)$ into itself.
To deal with the case $p=1$ we consider the local and global operators $\overline{R}_{\alpha,{\rm loc}}$ and $\overline{R}_{\alpha ,{\rm glob}}$ defined above with $N$ instead of $N_\beta$. Since the classical Riesz transform $\mathbb{R}_\alpha$ is bounded from $L^1(\mathbb{R}^n,dx)$ into $L^{1,\infty}(\mathbb{R}^n,dx)$ we can use \eqref{acotIb} and argue as in the proof of Theorem 1.2 to obtain that $\overline{R}_{\alpha ,{\rm loc}}$ defines a bounded operator from $L^1(\mathbb{R}^n, \gamma_{-1})$ into $L^{1,\infty }(\mathbb{R}^n,\gamma_{-1})$.
In order to prove that $\overline{R}_{\alpha ,{\rm glob}}$ defines a bounded operator from $L^1(\mathbb{R}^n,\gamma_{-1})$ into itself we make the change of variables $r=e^{-t}$, $t\in (0, \infty )$, and write $\overline{R}_\alpha (x,y)=\overline{R}_{\alpha, 1}(x,y)+\overline{R}_{\alpha ,2}(x,y)$, where
$$
\overline{R}_{\alpha ,1}(x,y)=\frac{(-1)^{|\alpha|}}{2^{|\alpha|}\pi ^{\frac{n}{2}}\Gamma(\frac{|\alpha|}{2})}e^{|y|^2-|x|^2}\int_0^{\frac{1}{2}}\frac{r^{|\alpha|-1}}{(1-r^2)^{\frac{n+|\alpha|}{2}}}\widetilde{H}_\alpha \Big(\frac{y-rx}{\sqrt{1-r^2}}\Big)(-\log r)^{\frac{|\alpha|}{2}-1}dr,\quad x,y\in \mathbb{R}^n,\;x\not=y.
$$
Suppose that $n=1$ or $|\alpha |\geq n+1$ when $n>1$. By using \cite[Lemma 3.3.3]{Sa} it follows that
\begin{align*}
|\overline{R}_{\alpha ,1}(x,y)|&\leq Ce^{|y|^2-|x|^2}\int_0^{1/2}r^{|\alpha|-1}\frac{e^{-c\frac{|y-rx|^2}{1-r^2}}}{(1-r^2)^{\frac{n}{2}}}(-\log r)^{\frac{|\alpha|}{2}-1}dr\\
&\leq Ce^{|y|^2-|x|^2}\sup_{r\in (0,1)}\frac{r^n e^{-c\frac{|y-rx|^2}{1-r^2}}}{(1-r^2)^{\frac{n}{2}}}\int_0^1(-\log r)^{\frac{|\alpha|}{2}-1}dr\\
&\leq Ce^{|y|^2-|x|^2}\min\{(1+|x|)^n,(|x|\sin \theta (x,y))^{-n}\}, \quad (x,y)\in N^c.
\end{align*}
On the other hand, by proceeding as in the estimation of $K_2^0(x,y)$ in \cite[proof of Proposition 5.1]{BrSj} we obtain, for every $(x,y)\in N^c$,
$$
|\overline{R}_{\alpha ,2}(x,y)|\leq Ce^{|y|^2-|x|^2}\int_{1/2}^1\frac{e^{-c\frac{|y-rx|^2}{1-r^2}}}{(1-r^2)^{\frac{n+2}{2}}}dr\leq Ce^{|y|^2-|x|^2}(|x|^{-n}+\min\{(1+|x|)^n,(|x|\sin \theta (x,y))^{-n}\}).
$$
From \cite[Lemma 3.3.4]{Sa} and \cite[Lemma 4.2]{BrSj} we deduce that $\overline{R}_{\alpha ,{\rm glob}}$ defines a bounded operator from $L^1(\mathbb{R}^n, \gamma_{-1})$ into $L^{1,\infty }(\mathbb{R}^n,\gamma_{-1})$.
Thus, we conclude that $R_\alpha$ can be extended to $L^1(\mathbb{R}^n, \gamma_{-1})$ as a bounded operator from $L^1(\mathbb{R}^n, \gamma_{-1})$ into $L^{1,\infty }(\mathbb{R}^n,\gamma_{-1})$.
\section{UMD spaces and Riesz transforms in the inverse Gaussian setting}
\subsection{Proof of Theorem \ref{Th1.4}.}
For every $i=1,...,n$, by $\mathbb{R}_{e^i}$ we denote the $i$-th Euclidean Riesz transform defined, for every $f\in L^p(\mathbb{R}^n,dx)$, $1<p<\infty$, by
$$
\mathbb{R}_{e^i}(f)(x)=\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon} \mathbb{R}_{e^i}(x-y)f(y)dy,\quad \mbox{fol almost all }x\in \mathbb{R}^n,
$$
where
$$
\mathbb{R}_{e^i}(z)=\frac{1}{\sqrt{\pi}}\int_0^\infty \partial_{x_i}W_t(z)\frac{dt}{\sqrt{t}},\quad z\in \mathbb{R}^n, \;z\not=0.
$$
Observe that
$$
\mathbb{R}_{e^i}(z)=-\frac{\sqrt{2}\Gamma (\frac{n+1}{2})}{\pi ^{\frac{n+1}{2}}}\frac{z_i}{|z|^{n+1}},\quad z\in \mathbb{R}^n,\;z\not=0,\;i=1,...,n.
$$
Let $X$ be a Banach space. For every $i=1,...,n$, we define $\mathbb{R}_{e^i}$ on $L^p(\mathbb{R}^n,dx)\otimes X$, $1\leq p<\infty$, in the obvious way.
The UMD-property for $X$ can be characterized by using $\mathbb{R}_{e^i}$, $i=1,...,n$. The properties stated in Theorems \ref{Th1.4} and \ref{Th1.5} hold when $R_{e^i}$ is replaced by $\mathbb{R}_{e^i}$, $i=1,...,n$. The estimations established in the proofs of Theorems \ref{Th1.1} and \ref{Th1.2} allow us to pass from $\mathbb{R}_{e^i}$ to $R_{e^i}$, $i=1,...,n$.
Let $i=1,...,n$ and $1<p<\infty$. We are going to see that the following two assertions are equivalent:
\begin{enumerate}
\item[(i)] $R_{e^i}$ can be extended from $(L^2(\mathbb{R}^n,\gamma_{-1})\bigcap L^p(\mathbb{R}^n,\gamma_{-1}))\otimes X$ to $L^p(\mathbb{R}^n,\gamma_{-1}, X)$ as a bounded operator from $L^p(\mathbb{R}^n,\gamma_{-1}, X)$ into itself.
\item [(ii)] $\mathbb{R}_{e^i}$ can be extended from $(L^2(\mathbb{R}^n,dx)\bigcap L^p(\mathbb{R}^n,dx)\otimes X$ to $L^p(\mathbb{R}^n,dx, X)$ as a bounded operator from $L^p(\mathbb{R}^n,dx, X)$ into itself.
\end{enumerate}
We choose $\frac{1}{p}<\eta <1$ and consider the global and local operators as in the previous sections according to the region $N_\beta$, with $\beta =\eta ^{-1}$.
Suppose that (ii) holds. We can write
$
R_{e^i}=(R_{e^i,{\rm loc}}-\mathbb{R}_{e^i,{\rm loc}})+\mathbb{R}_{e^i,{\rm loc}}+R_{e^i,{\rm glob}}.
$
Since $\mathbb{R}_{e^i}$ is a Calder\'on-Zygmund operator, by using a vectorial version of \cite[Proposition 3.2.5]{Sa} (see \cite[Proposition 2.3]{HTV}) we deduce that $\mathbb{R}_{e^i,{\rm loc}}$ can be extended from $(L^2(\mathbb{R}^n,\gamma_{-1})\bigcap L^p(\mathbb{R}^n,\gamma_{-1}))\otimes X$ to $L^p(\mathbb{R}^n,\gamma_{-1}, X)$ as a bounded operator from $L^p(\mathbb{R}^n,\gamma_{-1}, X)$ into itself.
According to \eqref{acotRalpha} and \eqref{acotdif} we have that
\begin{equation}\label{5.1}
|R_{e^i}(x,y)-\mathbb{R}_{e^i}(x-y)|\leq L_i(x,y),\;\;(x,y)\in N_\beta,\quad \mbox{ and }\quad |R_{e^i}(x,y)|\leq M_i(x,y),\;\;(x,y)\in N_\beta ^c,
\end{equation}
and the integral operators
$$
L_i(f)(x)=\int_{\mathbb{R}^n}L_i(x,y)\chi _{N_\beta}(x,y)f(y)dy,\quad \mbox{ and }\quad M_i(f)(x)=\int_{\mathbb{R}^n}M_i(x,y)\chi _{N_\beta ^c}(x,y)f(y)dy
$$
are bounded from $L^p(\mathbb{R}^n,\gamma_{-1})$ into itself. Then, $L_i$ and $M_i$ define bounded operators from $L^p(\mathbb{R}^n,\gamma_{-1},X)$ into itself, and the same property holds for the operators $R_{e^i,{\rm loc}}-\mathbb{R}_{e^i,{\rm loc}}$ and $R_{e^i,{\rm glob}}$. We conclude that (i) holds.
Suppose now that (i) holds. By \eqref{acotRalphab} we get
\begin{align*}
|R_{e^i}(x,y)|&\leq C\int_0^\infty e^{-nt}\frac{e^{-c\frac{|x-e^{-t}y|^2}{1-e^{-2t}}}}{(1-e^{-2t})^{\frac{n+1}{2}}}\frac{dt}{\sqrt{t}}\leq C\left(\int_0^{m(x)}\frac{e^{-c\frac{|x-y|^2}{t}}}{t^{\frac{n}{2}+1}}dt+\int_{m(x)}^\infty \frac{dt}{t^{\frac{n}{2}+1}}\right)\\
&\leq C\left(\frac{1}{|x-y |^n}+\frac{1}{m(x)^{\frac{n}{2}}}\right)\leq \frac{C}{|x-y|^n},\quad (x,y)\in N_\beta ,\;x\not=y.
\end{align*}
In a similar way we get, for each $k=1,...,n$,
$$
|\partial _{x_k}R_{e^i}(x,y)|\leq C\int_0^\infty e^{-nt}\frac{e^{-c\frac{|x-e^{-t}y|^2}{1-e^{-2t}}}}{(1-e^{-2t})^{\frac{n}{2}+1}}\frac{dt}{\sqrt{t}}\leq \frac{C}{|x-y|^{n+1}},\quad (x,y)\in N_\beta,\;x\not=y .
$$
Then, according to a vector valued version of \cite[Propositions 3.2.5 and 3.2.7]{Sa} (see \cite[Propositions 2.3 and 2.4]{HTV}) we deduce that $R_{e^i,{\rm loc}}$ defines a bounded operator from $L^p(\mathbb{R}^n,dx, X)$ into itself. Also, $R_{e^i,{\rm loc}}-\mathbb{R}_{e^i,{\rm loc}}$ defines a bounded operator from $L^p(\mathbb{R}^n,dx, X)$ into itself. We conclude that $\mathbb{R}_{e^i,{\rm loc}}$ defines a bounded operator from $L^p(\mathbb{R}^n,dx, X)$ into itself. Since $\mathbb{R}_{e^i}$ is dilatation invariant, by proceeding as in the proof of \cite[Theorem 1.10, $(ii)\Rightarrow (i)$]{HTV} it follows that $\mathbb{R}_{e^i}$ can be extended from $(L^2(\mathbb{R}^n,dx)\bigcap L^p(\mathbb{R}^n,dx)\otimes X$ to $L^p(\mathbb{R}^n,dx, X)$ as a bounded operator from $L^p(\mathbb{R}^n,dx, X)$ into itself.
The same arguments allow us to prove that the following assertions are equivalent.
\begin{enumerate}
\item[(iii)] $R_{e^i}$ can be extended from $(L^1(\mathbb{R}^n,\gamma_{-1})\bigcap L^2(\mathbb{R}^n,\gamma_{-1}))\otimes X$ to $L^1(\mathbb{R}^n,\gamma_{-1}, X)$ as a bounded operator from $L^1(\mathbb{R}^n,\gamma_{-1}, X)$ into $L^{1,\infty}(\mathbb{R}^n,\gamma_{-1}, X)$.
\item[(iv)] $\mathbb{R}_{e^i}$ can be extended from $(L^1(\mathbb{R}^n,dx)\bigcap L^2(\mathbb{R}^n,dx))\otimes X$ to $L^1(\mathbb{R}^n,dx, X)$ as a bounded operator from $L^1(\mathbb{R}^n,dx, X)$ into $L^{1,\infty}(\mathbb{R}^n,dx, X)$.
\end{enumerate}
Furthermore, in a similar way we can see that (i)$\iff$(ii) and (iii)$\iff$(iv) when $R_{e^i}$ and $\mathbb{R}_{e^i}$ are replaced by $R_{e^i,*}$ and $\mathbb{R}_{e^i,*}$, respectively.
The proof of Theorem \ref{Th1.4} is thus finished.
\subsection{Proof of Theorem \ref{Th1.5}.} Let $1<p<\infty$ and $i=1,...,n$. We are going to see that the following two assertions are equivalent.
(a) For every $f\in L^p(\mathbb{R}^n,\gamma_{-1}, X)$, there exists
$$
\lim_{\varepsilon \rightarrow 0^+} \int_{|x-y|>\varepsilon}R_{e^i}(x,y)f(y)dy,\quad \mbox{for almost all }x\in \mathbb{R}^n.
$$
(b) For every $f\in L^p(\mathbb{R}^n,dx, X)$, there exists
$$
\lim_{\varepsilon \rightarrow 0^+} \int_{|x-y|>\varepsilon}\mathbb{R}_{e^i}(x-y)f(y)dy,\quad \mbox{for almost all }x\in \mathbb{R}^n.
$$
We consider again $\frac{1}{p}<\eta<1$ and $\beta=\eta ^{-1}$. Suppose that $(a)$ is true. Let $f\in L^p(\mathbb{R}^n,dx, X)$. We can write
\begin{align*}
\int_{|x-y|>\varepsilon}\mathbb{R}_{e^i}(x-y)f(y)dy&=\int_{|x-y|>\varepsilon}(\mathbb{R}_{e^i}(x-y)-R_{e^i}(x-y))\chi _{N_\beta}(x,y)f(y)dy+\int_{|x-y|>\varepsilon}R_{e^i}(x,y)\chi _{N_\beta}(x,y)f(y)dy\\
&\quad +\int_{|x-y|>\varepsilon}\mathbb{R}_{e^i}(x-y)\chi _{N_\beta ^c}(x,y)f(y)dy,\quad x\in \mathbb{R}^n,\;\varepsilon >0.
\end{align*}
Since \eqref{5.1} holds and the operator $L_i$ is bounded from $L^p(\mathbb{R}^n,dx, X)$ into itself, there exists the limit
$$
\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon}(R_{e^i}(x,y))-\mathbb{R}_{e^i}(x-y)\chi _{N_\beta}f(y)dy,\quad \mbox{ for almost all }x\in \mathbb{R}^n.
$$
On the other hand, we get
\begin{align*}
\int_{\mathbb{R}^n}|\mathbb{R}_{e^i}(x-y)|\chi _{N_\beta ^c}(x,y)\|f(y)\|dy&\leq C\int_{\mathbb{R}^n}\frac{1}{|x-y|^n}\chi _{N_\beta ^c}(x,y)\|f(y)\|dy\\
&\hspace{-3cm}\leq C \left(\int_{|x-y|>\beta n\sqrt{m(x)}}\frac{dy}{|x-y|^{np'}}\right)^{1/p'}\|f\|_{L^p(\mathbb{R}^n, dx, X)}\\
&\hspace{-3cm}\leq C\left(\int_{\beta n\sqrt{m(x)}}^\infty \frac{dr}{r^{n(p'-1)+1}}\right)^{1/p'}\|f\|_{L^p(\mathbb{R}^n, dx, X)}\leq \frac{C}{m(x)^{n/(2p)}}\|f\|_{L^p(\mathbb{R}^n, dx, X)},\quad x\in \mathbb{R}^n.
\end{align*}
Then, there exists the limit
$$
\lim_{\varepsilon \rightarrow 0^+}\int_{\mathbb{R}^n}\mathbb{R}_{e^i}(x-y)\chi _{N_\beta ^c}(x,y)f(y)dy,\quad x\in \mathbb{R}^n.
$$
Suppose that $g\in L^p(\mathbb{R}^n, \gamma_{-1},X)$. It was seen in the proof of Theorem \ref{Th1.2} that
$$
\int_{\mathbb{R}^n}|R_{e^i}(x,y)|\chi _{N_\beta ^c}(x,y)\|g(y)\|dy\in L^p(\mathbb{R}^n,\gamma_{-1}).
$$
Then,
$$
\lim_{\varepsilon \rightarrow 0^+}\int_{\mathbb{R}^n}R_{e^i}(x,y)\chi _{N_\beta ^c}(x,y)g(y)dy=\int_{\mathbb{R}^n}R_{e^i}(x,y)\chi _{N_\beta^c}(x,y)g(y)dy,\quad \mbox{ for almost }x\in \mathbb{R}^n.
$$
Since (a) holds, there also exists the limit
$$
\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon}R_{e^i}(x,y)\chi _{N_\beta}(x,y)g(y)dy,\quad \mbox{ for almost all }x\in\mathbb{R}^n.
$$
Let $k\in \mathbb{N}$. We have that $|y|\leq \beta n+k$ provided that $|x|\leq k$ and $|x-y|\leq \beta n\min\{1,|x|^{-1}\}$. Then, for every $\varepsilon >0$,
$$
\int_{|x-y|>\varepsilon}R_{e^i}(x,y)\chi _{N_\beta}(x,y)f(y)dy=\int_{|x-y|>\varepsilon}R_{e^i}(x,y)\chi _{N_\beta}(x,y)\chi _{B(0,\beta n+k)}(y)f(y)dy,\quad |x|\leq k.
$$
Since $f\in L^p(\mathbb{R}^n,dx,X)$, $\chi_{B(0,\beta n+k)}f\in L^p(\mathbb{R}^n, \gamma_{-1},X)$ and then there exists
$$
\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon}R_{e^i}(x,y)\chi _{N_\beta}(x,y)f(y)dy,\quad \mbox{ for almost all }x\in B(0,k).
$$
Hence, we get that there exists
$$
\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon}R_{e^i}(x,y)\chi _{N_\beta}(x,y)f(y)dy,\quad \mbox{ for almost all }x\in\mathbb{R}^n.
$$
We conclude that there exists
$$
\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon}\mathbb{R}_{e^i}(x-y)f(y)dy,\quad \mbox{ for almost all }x\in\mathbb{R}^n.
$$
In a similar we can see that (a) holds provided that (b) is true. Note that $L^p(\mathbb{R}^n,\gamma_{-1},X)\subset L^p(\mathbb{R}^n,dx,X)$.
As it was proved in \cite{BrSj} the operator $S_i$ defined by
$$
S_i(f)(x)=\int_{\mathbb{R}^n}|R_{e^i}(x,y)|\chi _{N_\beta ^c}(x,y)f(y)dy,\quad x\in \mathbb{R}^n,
$$
is bounded from $L^1(\mathbb{R}^n, \gamma_{-1})$ into $L^{1,\infty}(\mathbb{R}^n, \gamma_{-1})$. Then, for every $f\in L^1(\mathbb{R}^n, \gamma_{-1})$ there exists
$$
\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon}R_{e^i}(x,y)\chi _{N_\beta}(x,y)f(y)dy=\int_{\mathbb{R}^n}R_{e^i}(x,y)\chi _{N_\beta}(x,y)f(y)dy,\quad \mbox{ for almost all }x\in\mathbb{R}^n.
$$
The same arguments allow us to prove that (a) $\iff$ (b) when $p=1$.
By using a $n$-dimensional version of \cite[Theorem D]{TZ} we deduce that the properties $(i)$, $(ii)$ and $(iii)$ in Theorem \ref{Th1.5} are equivalent.
By proceeding as above we can see that the properties in (a) and (b) continue being equivalent when we replace the principal values by the maximal operators $R_{e^i,*}$ and $\mathbb{R}_{e^i,*}$ in (a) and (b), respectively. Then, the property (b) is equivalent to the property UMD for $X$ (see the comments before the proof of \cite[Theorem 1.10, p. 19]{HTV}).
Thus the proof of Theorem \ref{Th1.5} is finished.
\section{UMD spaces and the imaginary powers of $\mathcal{A}$}
In this section we prove Theorem \ref{Th1.6}.
According to \cite[(11)]{BCFR} we have that, for every $f\in C_c^\infty (\mathbb{R}^n)$,
$$
\Big(-\frac{\Delta }{2}\Big)^{i\gamma}f(x)=\lim_{\varepsilon \rightarrow 0^+}\Big(\alpha (\varepsilon )f(x)+\int_{|x-y|>\varepsilon }K_\gamma (x-y)f(y)dy\Big),\quad \mbox{ for a.e. }x\in \mathbb{R}^n,
$$
where
$$
K_\gamma (z)=-\int_0^\infty \phi_\gamma (t)\partial _tW_t(z)dt,\quad z\in \mathbb{R}^n\setminus\{0\},
$$
$$
\alpha (\varepsilon )=\frac{1}{\Gamma (\frac{n}{2})}\int_0^\infty \phi_\gamma \big(\frac{\varepsilon ^2}{4u}\big)e^{-u}u^{\frac{n}{2}-1}du,\quad \varepsilon \in (0,\infty ),
$$
and
$$
\phi _\gamma (t)=\frac{t^{-i\gamma}}{\Gamma (1-i\gamma)},\quad t\in (0,\infty ).
$$
Note that the limit $\lim_{t\rightarrow 0^+}\phi_\gamma (t)$ does not exist.
Let $f,g\in C_c^\infty (\mathbb{R}^n)$. By proceeding as in \cite[p. 213]{BCFR} we can see that
\begin{align*}
\langle \mathcal{A}^{i\gamma}f,g\rangle_{L^2(\mathbb{R}^n,\gamma_{-1})}&=\int_0^\infty \phi _\gamma (t)\Big(-\frac{d}{dt}\Big)\int_{\mathbb{R}^n}\int_{\mathbb{R}^n}T_t^\mathcal{A}(x,y)f(y)dy\overline{g(x)}\gamma_{-1}(x)dxdt\\
&=\int_0^\infty \phi _\gamma (t)\Big(-\frac{d}{dt}\Big)\int_{\mathbb{R}^n}\int_{\mathbb{R}^n}W_t(x-y)f(y)dy\overline{g(x)}\gamma_{-1}(x)dxdt\\
&\quad +\int_0^\infty \phi _\gamma (t)\Big(-\frac{d}{dt}\Big)\int_{\mathbb{R}^n}\int_{\mathbb{R}^n}(T_t^\mathcal{A}(x,y)-W_t(x-y))f(y)dy\overline{g(x)}\gamma_{-1}(x)dxdt.
\end{align*}
We have that
\begin{align*}
\partial_tT_t^\mathcal{A}(x,y)&=\Big(-n-2e^{-t}\sum_{i=1}^ny_i(x_i-e^{-t}y_i)+\frac{2e^{-2t}|x-e^{-t}y|^2}{1-e^{-2t}}\Big)\frac{T_t^\mathcal{A}(x,y)}{1-e^{-2t}}\\
&=\Big(-n-2e^{-t}(\langle x,y\rangle-e^{-t}|y|^2)+\frac{2e^{-2t}|x-e^{-t}y|^2}{1-e^{-2t}}\Big)\frac{T_t^\mathcal{A}(x,y)}{1-e^{-2t}},\quad x,y\in \mathbb{R}^n,\;t>0,
\end{align*}
and writing $2e^{-t}\langle x,y\rangle=|x|^2 +e^{-2t}|y|^2-|x-e^{-t}y|^2$, we get
\begin{equation}\label{derivadaT}
\partial_tT_t^\mathcal{A}(x,y)=\Big(-n+e^{-2t}(|y|^2-|x|^2)- (1-e^{-2t})|x|^2+\frac{1+e^{-2t}}{1-e^{-2t}}|x-e^{-t}y|^2)\Big)\frac{T_t^\mathcal{A}(x,y)}{1-e^{-2t}},\quad x,y\in \mathbb{R}^n,\;t>0.
\end{equation}
Also, we have that
$$
\partial_tW_t(x-y)=\Big(-n+\frac{|x-y|^2}{t}\Big)\frac{W_t(x-y)}{2t},\quad x,y\in \mathbb{R}^n,\;t>0.
$$
We can write, for every $x,y\in \mathbb{R}^n$ and $t>0$,
\begin{align}\label{difbas}
\partial_tT_t^\mathcal{A}(x,y)-\partial_tW_t(x-y)&=\Big(-n+\frac{|x-y|^2}{t}\Big)\frac{T_t^\mathcal{A}(x,y)-W_t(x-y)}{2t}\nonumber \\
&\hspace{-3cm}+\left[-n\Big(\frac{1}{1-e^{-2t}}-\frac{1}{2t}\Big)+\frac{e^{-2t}}{1-e^{-2t}}(|y|^2-|x|^2)-|x|^2-\frac{|x-e^{-t}y|^2}{1-e^{-2t}}+2\Big(\frac{|x-e^{-t}y|^2}{(1-e^{-2t})^2}-\frac{|x-y|^2}{(2t)^2}\Big)\right]T_t^\mathcal{A}(x,y).
\end{align}
The derivative under the integral sign is justified and we get
\begin{align*}
\langle \mathcal{A}^{i\gamma}f,g\rangle_{L^2(\mathbb{R}^n,\gamma_{-1})}&=-\int_0^\infty \phi_\gamma (t)\int_{\mathbb{R}^n}\int_{\mathbb{R}^n}\partial _tW_t(x-y)f(y)dy\overline{g(x)}\gamma_{-1}(x)dxdt\\
&\quad -\int_0^\infty \phi_\gamma (t)\int_{\mathbb{R}^n}\int_{\mathbb{R}^n}\partial_t(T_t^\mathcal{A}(x,y)-W_t(x-y))f(y)dy\overline{g(x)}\gamma_{-1}(x)dxdt.
\end{align*}
To ensure the change on the order of integration we are going to see that
\begin{equation}\label{6.1}
J(x)=\int_0^\infty \int_{\mathbb{R}^n}\left|\partial_t(T_t^\mathcal{A}(x,y)-W_t(x-y))\right||f(y)|dydt<\infty, \quad x\in \mathbb{R}^n,
\end{equation}
and
\begin{equation}\label{6.2}
\int_0^\infty \int_{\mathbb{R}^n}\int_{\mathbb{R}^n}\left|\partial _t(T_t^\mathcal{A}(x,y)-W_t(x-y))\right||f(y)||g(x)|\gamma_{-1}(x)dydxdt<\infty .
\end{equation}
We observe that, since $g\in C_c^\infty (\mathbb{R}^n)$, it is sufficient to see that $J(x)\leq h(x)$, $x\in \mathbb{R}^n$, for certain continuous function $h$.
We consider the decomposition
$$
J(x)=\left(\int_0^{m(x)}+\int_{m(x)}^\infty\right)\int_{\mathbb{R}^n}\left|\partial_t(T_t^\mathcal{A}(x,y)-W_t(x-y))\right||f(y)|dydt=J_1(x)+J_2(x),\quad x\in \mathbb{R}^n.
$$
Since $f\in C_c^\infty (\mathbb{R}^n)$ we have that
$$
J_2(x)\leq C(1+|x|)^2\int_{m(x)}^\infty \int_{{\rm supp }f}\frac{1}{t^{\frac{n}{2}+1}}dtdy\leq C\frac{(1+|x|)^2}{m(x)^{\frac{n}{2}}},\quad x\in \mathbb{R}^n.
$$
We now estimate $J_1(x)$, $x\in \mathbb{R}^n$. We take into account that, if $\supp f\subset B(0,r_f)$, with $r_f>0$, then
\begin{align*}
|x-e^{-t}y|^2&\geq |x-y|^2+(1-e^{-t})^2|y|^2-2|x-y||y|(1-e^{-t})\nonumber\\
&\geq |x-y|^2-d(x)(1-e^{-t}), \quad x\in \mathbb{R}^n,\;y\in {\rm supp }f,\; t>0,
\end{align*}
where $d(x)=2r_f(r_f+|x|)$, $x\in \mathbb{R}^n$. By considering the decomposition \eqref{decomposition} for $\alpha =0$, the estimations \eqref{differences} and the proof of \eqref{Halpha} we can see that
$$
|T_t^\mathcal{A}(x,y)-W_t(x-y)|\leq Ce^{d(x)}e^{-c\frac{|x-y|^2}{t}}\Big(\frac{1}{t^{\frac{n}{2}-1}}+(|y|+|x-y|)^n\Big),\quad x\in \mathbb{R}^n,\;\;y\in {\rm supp }f,\;t\in (0,1).
$$
Then, according to \eqref{difbas} we have that
\begin{align*}
|\partial_t(T_t^\mathcal{A}(x,y)-W_t(x-y))|&\leq Ce^{d(x)}e^{-c\frac{|x-y|^2}{t}}\left(\frac{1+|x|^2}{t^{\frac{n}{2}}}+\frac{(|y|+|x-y|)^n}{t}+\frac{|x-y||y+x|}{t^{\frac{n}{2}+1}}+\frac{|y|+ |x-y|}{t^{\frac{n}{2}}}\right)\\
&\leq Ce^{d(x)}e^{-c\frac{|x-y|^2}{t}}\left(\frac{1+|x|^2}{t^{\frac{n}{2}}}+\frac{(1+|x|)^n}{t}+\frac{1+|x|}{t^{\frac{n+1}{2}}}\right),\quad x\in \mathbb{R}^n,\;\;y\in {\rm supp }f,\;t\in (0,1).
\end{align*}
Since $m(x)\sim (1+|x|)^{-2}$, $x\in \mathbb{R}^n$, we get when $t\in (0,m(x))$,
$$
|\partial_t(T_t^\mathcal{A}(x,y)-W_t(x-y))|\leq Ce^{d(x)}e^{-c\frac{|x-y|^2}{t}}\left(\frac{1+|x|}{t^{\frac{n+1}{2}}}+\frac{(1+|x|)^n}{t}\right),\quad x\in \mathbb{R}^n,\;\;y\in {\rm supp }f.
$$
We deduce that
\begin{align*}
J_1(x)&\leq Ce^{d(x)}\int_0^{m(x)}\int_{\supp f}\left(\frac{1+|x|}{t^{\frac{3}{4}}}+(1+|x|)^nt^{\frac{n}{2}-\frac{5}{4}}\right)\frac{dy}{|x-y|^{n-\frac{1}{2}}}\\
&\leq Ce^{d(x)}\Big((1+|x|)m(x)^{\frac{1}{4}}+(1+|x|)^nm(x)^{\frac{n}{2}-\frac{1}{4}}\Big)\leq Ce^{d(x)}\sqrt{1+|x|},\quad x\in \mathbb{R}^n.
\end{align*}
Then, $|J(x)|\leq C((1+|x|)^{n+2}+e^{d(x)}\sqrt{1+|x|})$, $x\in \mathbb{R}^n$ and \eqref{6.1} and \eqref{6.2} are established.
Note that the estimations that we have just proved are depending on $f$.
By interchanging the order of integration we get
$$
\langle \mathcal{A}^{i\gamma}f,g\rangle _{L^2(\mathbb{R}^n,\gamma_{-1})}=\langle \big(-\frac{\Delta}{2}\big)^{i\gamma}f,g\rangle_{L^2(\mathbb{R}^n,\gamma_{-1})}+\int_{\mathbb{R}^n}\int_{\mathbb{R}^n}f(y)\int_0^\infty \phi _\gamma (t) \Big(\partial_tT_t^\mathcal{A}(x,y)-\partial_tW_t(x-y)\Big)dtdyg(x)\gamma_{-1}(x)dx.
$$
It follows that
\begin{align*}
\mathcal{A}^{i\gamma}(f)(x)&=\big(-\frac{\Delta}{2}\big)^{i\gamma}f(x)-\int_{\mathbb{R}^n}f(y)\int_0^\infty \phi _\gamma (t) \Big(\partial_tT_t^\mathcal{A}(x,y)-\partial_tW_t(x-y)\Big)dtdy\\
&=\big(-\frac{\Delta}{2}\big)^{i\gamma}f(x)-\lim_{\varepsilon \rightarrow 0^+}\int_{|x-y|>\varepsilon}f(y)\int_0^\infty \phi _\gamma (t) \Big(\partial _tT_t^\mathcal{A}(x,y)-\partial_tW_t(x-y)\Big)dtdy,
\end{align*}
for almost all $x\in \mathbb{R}^n$.
We conclude that
$$
\mathcal{A}^{i\gamma}(f)(x)=\lim_{\varepsilon \rightarrow 0^+}\Big(\int_{|x-y|>\varepsilon}K_\gamma ^\mathcal{A}(x,y)f(y)dy+\alpha (\varepsilon )f(x)\Big),\quad\mbox{for almost all }x\in \mathbb{R}^n,
$$
where
$$
K_\gamma ^\mathcal{A}(x,y)=-\int_0^\infty \phi _\gamma (t)\partial_tT_t^\mathcal{A}(x,y)dt,\quad x,y\in \mathbb{R}^n,\;t>0.
$$
Salogni (\cite[Theorem 3.4.3]{Sa}) proved that $\mathcal{A}^{i\gamma}$ is bounded from $L^p(\mathbb{R}^n,\gamma_{-1})$ into itself, for every $1<p<\infty$, and Bruno (\cite[Theorem 4.1, (i)]{Br}) established that $\mathcal{A}^{i\gamma}$ is bounded from $L^1(\mathbb{R}^n,\gamma_{-1})$ into $L^{1,\infty}(\mathbb{R}^n,\gamma_{-1})$. In order to extend $\mathcal{A}^{i\gamma}$ to functions taking values in a Banach space we need to prove these results in a different way by making a comparation between $\mathcal{A}^{i\gamma}$ and $(-\frac{\Delta}{2})^{i\gamma}$.
Let $\beta >0$. We define the local and global part as follows
$$
\mathcal{A}^{i\gamma}_{\rm loc}(f)(x)=\lim_{\varepsilon \rightarrow 0^+}\Big(\alpha (\varepsilon )f(x)+\int_{|x-y|>\varepsilon}K_\gamma ^\mathcal{A}(x,y)\chi_{N_\beta}(x,y)f(y)dy\Big),\quad x\in \mathbb{R}^n,
$$
and
$$
\mathcal{A}^{i\gamma}_{\rm glob}(f)(x)=\int_{|x-y|>\varepsilon}K_\gamma ^\mathcal{A}(x,y)\chi_{N_\beta^c}(x,y)f(y)dy,\quad x\in \mathbb{R}^n.
$$
The operators $(-\frac{\Delta}{2})_{\rm loc}^{i\gamma}$ and $(-\frac{\Delta}{2})_{\rm glob}^{i\gamma}$ are defined in analogous way.
We are going to see that
\begin{equation}\label{6.4}
|K_\gamma ^\mathcal{A}(x,y)-K_\gamma (x-y)|\leq C\frac{\sqrt{1+|x|}}{|x-y|^{n-\frac{1}{2}}},\quad (x,y)\in N_\beta.
\end{equation}
We can write,
\begin{align*}
|K_\gamma ^\mathcal{A}(x,y)-K_\gamma (x-y)|&\leq C\left(\int_0^{m(x)}|\partial_t(T_t^\mathcal{A}(x,y)-W_t(x-y))|dt\right.\\
&\left.+\int_{m(x)}^\infty|\partial_tT_t^\mathcal{A}(x,y)|dt+\int_{m(x)}^\infty|\partial_tW_t(x-y)|dt\right)=\sum_{i=1}^3I_i(x,y),\quad x,y\in \mathbb{R}^n.
\end{align*}
First we observe that
$$
I_3(x,y)\leq C\int_{m(x)}^\infty \frac{e^{-c\frac{|x-y|^2}{t}}}{t^{\frac{n}{2}+1}}dt\leq \frac{C}{m(x)^{\frac{n}{2}}}\leq C(1+|x|)^n\leq C\frac{\sqrt{1+|x|}}{|x-y|^{n-\frac{1}{2}}}, \quad (x,y)\in N_\beta.
$$
By using \eqref{derivadaT} and since $||y|^2-|x|^2|\leq |x-y||x+y|\leq C$ when $(x,y)\in N_\beta$, we have that
$$
I_2(x,y)\leq C(1+|x|^2)\int_{m(x)}^\infty \frac{e^{-\frac{n}{2}t}}{t^{\frac{n}{2}+1}}dt\leq C\frac{(1+|x|^2)}{m(x)^{\frac{n}{2}+1}}\leq C(1+|x|)^n\leq C\frac{\sqrt{1+|x|}}{|x-y|^{n-\frac{1}{2}}}, \quad (x,y)\in N_\beta.
$$
Finally, from \eqref{difbas}, proceeding as above for the estimation of $J_1$, but now by taking into account that $(x,y)\in N_\beta$, we obtain that
\begin{align*}
I_1(x,y)&\leq C\int_0^{m(x)}e^{-c\frac{|x-y|^2}{t}}\left(\frac{1+|x|}{t^{\frac{n+1}{2}}}+\frac{(1+|x|)^n}{t}\right)dt\leq C\frac{(1+|x|)m(x)^{\frac{1}{4}}+(1+|x|)^nm(x)^{\frac{n}{2}-\frac{1}{4}}}{|x-y|^{n-\frac{1}{2}}}\\
&\leq C\frac{\sqrt{1+|x|}}{|x-y|^{n-\frac{1}{2}}},\quad (x,y)\in N_\beta.
\end{align*}
Thus, \eqref{6.4} is proved.
As it was established in Section 3.2,
$$
\sup_{x\in \mathbb{R}^n}\int_{\mathbb{R}^n}\frac{\sqrt{1+|x|}}{|x-y|^{n-\frac{1}{2}}}\chi_{N_\beta}(x,y)dy<\infty\quad \mbox{ and }\quad \sup_{y\in \mathbb{R}^n}\int_{\mathbb{R}^n}\frac{\sqrt{1+|x|}}{|x-y|^{n-\frac{1}{2}}}\chi_{N_\beta}(x,y)dx<\infty.
$$
Then, the operator $L_\beta$ defined by
$$
L_\beta (f)(x)=\int_{\mathbb{R}^n}|K_\gamma ^\mathcal{A}(x,y)-K_\gamma (x,y)|\chi_{N_\beta}(x,y)f(y)dy,\quad x\in \mathbb{R}^n,
$$
is bounded from $L^p(\mathbb{R}^n,dx)$ into itself for every $1\leq p<\infty$. From \cite[Proposition 3.2.5]{Sa} it follows that $L_\beta$ is bounded from $L^p(\mathbb{R}^n,\gamma_{-1})$ into itself, for every $1\leq p<\infty$.
Note that
$$
|\mathcal{A}^{i\gamma}_{\rm loc}(f)-\big(-\frac{\Delta}{2}\big)^{i\gamma}_{\rm loc}(f)|\leq CL_\beta (|f|).
$$
We now study the global operator $\mathcal{A}^{i\gamma}_{\rm glob}$. We recall that $\mathcal{A}^{i\gamma}_{\rm glob}$ is the integral operator defined by
$$
\mathcal{A}^{i\gamma}_{\rm glob}(f)(x)=-\int_{\mathbb{R}^n}\int_0^\infty \phi _\gamma (t)\partial_tT_t^\mathcal{A}(x,y)dt\chi_{N_\beta ^c}(x,y)f(y)dy,\quad x\in \mathbb{R}^n.
$$
We have, by making $r=e^{-t}$, $t\in (0,\infty)$, that
\begin{align*}
\int_0^\infty |\phi _\gamma (t)|\partial_tT_t^\mathcal{A}(x,y)|dt&\leq C\int_0^1 \Big|\partial_r\Big[r^n\frac{e^{-\frac{|x-ry|^2}{1-r^2}}}{(1-r^2)^{\frac{n}{2}}}\Big]\Big|dr\\
&\hspace{-2.5cm}\leq C\int_0^1\left|\frac{n(1-r^2) +2r(1-r^2)\sum_{i=1}^ny_i(x_i-ry_i)-2r^2|x-ry|^2}{(1-r^2)^2}\right|r^{n-1}\frac{e^{-\frac{|x-ry|^2}{1-r^2}}}{(1-r^2)^{\frac{n}{2}}}dr\\
&\hspace{-2.5cm}=C\int_0^1\left|\frac{(1-r^2)(n+|x|^2-r^2|y|^2)-(1+r^2)|x-ry|^2}{(1-r^2)^2}\right|r^{n-1}\frac{e^{-\frac{|x-ry|^2}{1-r^2}}}{(1-r^2)^{\frac{n}{2}}}dr.
\end{align*}
For every $x,y\in \mathbb{R}^n$ there exists a polynomial $P_{x,y}$ with degree 4 and a positive function $Q_{x,y}$ such that
$$
\partial_r\Big[r^n\frac{e^{-\frac{|x-ry|^2}{1-r^2}}}{(1-r^2)^{\frac{n}{2}}}\Big]=P_{x,y}(r)Q_{x,y}(r),\quad r\in (0,1).
$$
Then, for every $x,y\in \mathbb{R}^n$, the function
$$
\partial_r\left[r^n\frac{e^{-\frac{|x-ry|^2}{1-r^2}}}{(1-r^2)^{\frac{n}{2}}}\right]
$$
changes the sign at most four times in $(0,1)$. We deduce that
$$
\int_0^\infty |\phi_\gamma (t)|\big|\partial _tT_t^\mathcal{A}(x,y)\big|dt\leq C\int_0^1\Big|\partial_ r\Big[r^n\frac{e^{-\frac{|x-ry|^2}{1-r^2}}}{(1-r^2)^{\frac{n}{2}}}\Big]\Big|dr\leq C\sup_{r\in (0,1)}r^n\frac{e^{-\frac{|x-ry|^2}{1-r^2}}}{(1-r^2)^{\frac{n}{2}}},\quad x,y\in \mathbb{R}^n.
$$
If $(u,v)\in N_\sigma $, with $\sigma >0$, we have that
$$
\frac{1}{(1+\sigma n)(1+|v|)}\leq \frac{1}{1+|u|}\leq \frac{1+\sigma n}{1+|v|}.
$$
Also, $\frac{1}{2}\min\{1,|x|^{-1}\}\leq (1+|x|)^{-1}\leq \min\{1,|x|^{-1}\}$, $x\in \mathbb{R}^n$. Then, if $(y,x)\in N$, then $(x,y)\in N_{2(1+n)}$.
We take $\beta=2(1+n)$. Since $(y,x)\not\in N$ when $(x,y)\not \in N_\beta$, according to \cite[Proposition 2.1]{MPS} we get, for $(x,y)\not\in N_\beta$,
$$
\sup_{r\in (0,1)}\frac{e^{-\frac{|x-ry|^2}{1-r^2}}}{(1-r^2)^{\frac{n}{2}}}\leq C\left\{\begin{array}{ll}
e^{-|x|^2},&\langle x,y\rangle \leq 0,\\[0.3cm]
\displaystyle \Big(\frac{|x+y|}{|x-y|}\Big)^{\frac{n}{2}}\exp\Big(\frac{|y|^2-|x|^2}{2}-\frac{|x-y||x+y|}{2}\Big),&\langle x,y\rangle >0.
\end{array}
\right.
$$
By proceeding as in the proof of Theorem \ref{Th1.2} we can see that, for every $1<p<\infty$,
$$
\sup_{x\in \mathbb{R}^n}\int_{\mathbb{R}^n}e^{\frac{|x|^2-|y|^2}{p}}\chi _{N_\beta ^c}(x,y)\int_0^\infty |\phi _\gamma (t)|\Big|\partial_tT_t^\mathcal{A}(x,y)\Big|dtdy<\infty,
$$
and
$$
\sup_{y\in \mathbb{R}^n}\int_{\mathbb{R}^n}e^{\frac{|x|^2-|y|^2}{p}}\chi _{N_\beta ^c}(x,y)\int_0^\infty |\phi _\gamma (t)|\Big|\partial_tT_t^\mathcal{A}(x,y)\Big|dtdy<\infty .
$$
Hence the operator $\mathcal{L}_\beta$ defined by
$$
\mathcal{L}_\beta (f)(x)=\int_{\mathbb{R}^n}f(y)\chi _{N_\beta ^c}(x,y)\int_0^\infty |\phi _\gamma (t)|\Big|\partial_tT_t^\mathcal{A}(x,y)\Big|dtdy,\quad x\in \mathbb{R}^n,
$$
is bounded from $L^p(\mathbb{R}^n,\gamma_{-1})$ into itself, for every $1<p<\infty$.
On the other hand, according to \cite[Lemma 3.3.3]{Sa}
$$
\sup_{t>0}T_t^\mathcal{A}(x,y)\leq Ce^{-|x|^2}\min\{(1+|x|)^n, (|x|\sin \theta (x,y))^{-n}\},\quad (x,y)\in N_\beta ^c,\;x,y\not=0.
$$
We recall that $\theta (x,y)\in [0,\pi)$ represents the angle between the nonzero vectors $x$ and $y$ when $n>1$ and $\theta (x,y)=0$, $x,y\in \mathbb{R}^n$, when $n=1$. By \cite[Lemma 3.3.4]{Sa} the operator $\mathcal{L}_\beta $ is bounded from $L^1(\mathbb{R}^n,\gamma _{-1})$ into $L^{1,\infty}(\mathbb{R}^n,\gamma_{-1})$.
Note that $|\mathcal{A}_{\rm glob}^{i\gamma}(f)|\leq \mathcal{L}_\beta(|f|)$.
By taking into account that $(-\frac{\Delta}{2})^{i\gamma}$ is a Calder\'on-Zygmund operator, the arguments developed in the proofs of Theorems \ref{Th1.2} and \ref{Th1.3} allow us to finish the proof of this theorem.
\bibliographystyle{acm}
|
\section{Introduction}
\setcounter{equation}{0}
The study of the coupling free flow
and porous media flow is of considerable interest and
has attracted a lot of attentions in recent years due to its wide applications in geosciences (e.g., karst aquifers, hyporheic flow, contaminant transport), health sciences (e.g., blood flow) and industrial processes, see \cite{DQ09} and the references therein. In this paper, we investigate a Navier--Stokes--Darcy--Boussinesq system that models thermal convection in an incompressible viscous fluid overlying a saturated porous medium (see, e.g., \cite{MMW19}). The convection phenomenon under consideration is much more complicated than that in a single fluid (cf. \cite{FMT1987} for the free-flow and \cite{F86,LT99} for fluids in a porous medium), since more physical parameters will affect the heat transport process. Linear and nonlinear stability analysis, properties of bifurcation and dynamic transition for the coupled system, with the Navier--Stokes equations and Darcy's equation governing the free-flow and the porous regions, have been provided in \cite{MMW19,HWW20} under suitable interface conditions. Here, our aim is to perform a first-step analysis on the well-posedness of the associated initial boundary value problem, proving the existence of global weak solutions and their uniqueness property.
Assume that the fluid is confined in a bounded connected domain $\Omega \subset \mathbb{R}^d$ ($d=2,3$) with $C^{2,1}$ boundary $\partial \Omega$. The unit outer normal vector on $\partial \Omega$ is denoted by $\mathbf{n}=\mathbf{n}(x)$. The domain $\Omega$ is partitioned into two non-overlapping regions such that $\overline{\Omega}=\overline{\Omega}_f\cup \overline{\Omega}_m$ and $\Omega_f \cap \Omega_m= \emptyset$, where $\Omega_f$ and $\Omega_m$ represent the free-flow region and the porous matrix region, respectively. We denote by $\partial \Omega_f$ and $\partial \Omega_m$ the boundaries of the free-flow and the matrix part, with $\widehat{\mathbf{n}}_f$, $\widehat{\mathbf{n}}_m$ being the corresponding unit outer normals on them. Both $\partial \Omega_f$ and $\partial \Omega_m$ are assumed to be Lipschitz continuous. The interface between the two parts (i.e., $\partial \Omega_f\cap \partial \Omega_m$) is denoted by $\Gamma_{i}$. On the free-flow/matrix interface $\Gamma_{i}$, $\mathbf{n}_{i}$ stands for the unit normal on $\Gamma_{i}$ pointing from the free-flow region to the matrix, and $\{\boldsymbol{\tau}_j\}$ $(j=1,...,d-1)$ stands for a local orthonormal basis for the tangent plane to $\Gamma_{i}$. Let $\Gamma_f=\partial \Omega_f\backslash \Gamma_{i}$ and $\Gamma_m=\partial \Omega_m\backslash \Gamma_{i}$ with $\mathbf{n}_f, \mathbf{n}_m$ being the unit outer normals to
$\Gamma_{f}$ and $\Gamma_{m}$. We assume that $\Gamma_m$ and $\Gamma_{i}$ have positive measure (i.e., $|\Gamma_m|>0$, $|\Gamma_{i}|>0$) but allow $\Gamma_f=\emptyset$, namely, $\Omega_f$ can be enclosed completely by $\Omega_m$. When $d=3$, we assume that the surfaces $\Gamma_f$, $\Gamma_m$ and $\Gamma_{i}$ have Lipschitz continuous boundaries.
In the sequel, the subscript $f$ (or $m$) indicates that the variables are for the free-flow part (or the matrix part). We denote by $\mathbf{u}$ the mean velocity of the (incompressible) fluid and $\theta$ the (relative) temperature of the fluid. The following convention will be assumed throughout the paper
$$
\mathbf{u}|_{\Omega_f}=\mathbf{u}_f, \ \ \ \ \mathbf{u}|_{\Omega_m}=\mathbf{u}_m, \ \ \ \ \theta|_{\Omega_f}=\theta_f, \ \ \theta|_{\Omega_m}=\theta_m.
$$
\textbf{Governing PDE system}. We shall consider the following Navier--Stokes--Darcy--Boussinesq system (in a nondimensionallized form, see \cite{MMW19})
\begin{align}
&\partial_t \mathbf{u}_f+(\mathbf{u}_f\cdot\nabla )\mathbf{u}_f=\nabla\cdot\mathbb{T}(\mathbf{u}_f,P_f) + \theta_f\mathbf{k}, &\mbox{in}\ \Omega_f\times(0,T),
\label{uf1}\\
&\nabla\cdot\mathbf{u}_f=0, &\mbox{in}\ \Omega_f\times(0,T),
\label{uf2}\\
&\partial_t\theta_f+(\mathbf{u}_f\cdot\nabla)\theta_f={\rm div}(\lambda_f(\theta_f)\nabla\theta_f), &\mbox{in}\ \Omega_f\times(0,T),
\label{tf1}\\
&\varpi\partial_t\mathbf{u}_m+\nu(\theta_m)\mathbb{K}^{-1}\mathbf{u}_m
=-\nabla P_m + \theta_m\mathbf{k},
&\mbox{in}\ \Omega_m\times(0,T),
\label{um1}\\
&\nabla\cdot\mathbf{u}_m=0, &\mbox{in}\ \Omega_m\times(0,T), \label{um2}\\
&\partial_t\theta_m+(\mathbf{u}_m\cdot\nabla)\theta_m={\rm div}(\lambda_m(\theta_m)\nabla\theta_m), &\mbox{in}\ \Omega_m\times(0,T),
\label{tm1}
\end{align}
where $T\in (0,+\infty)$ is an arbitrary final time and $\mathbf{k}$ is the upward pointing unit vector. In the free-flow region, the motion of the incompressible fluid is characterized by the Navier--Stokes equations \eqref{uf1}--\eqref{uf2} with Boussinesq approximation (buoyancy force), coupled with the advection-diffusion equation \eqref{tf1} for the temperature. While for the fluid in porous medium, we employ the Darcy system \eqref{um1}--\eqref{um2} (valid under the small porosity assumption that is generally applicable to geophysical systems) with the advection-diffusion equation \eqref{tm1}. The Cauchy stress tensor $\mathbb{T}$ in equation \eqref{uf1} is given by
\begin{align}
\mathbb{T}(\mathbf{u}_f,p_f) = 2\nu(\theta_f)\mathbb{D}(\mathbf{u}_f)-P_f\mathbb{I},
\label{CT}
\end{align}
where $\mathbb{D}(\mathbf{u}_f)=\frac{1}{2}(\nabla \mathbf{u}_f + \nabla^T \mathbf{u}_f)$ is the symmetric rate of deformation tensor and $\mathbb{I}$ denotes the $d\times d$ identity matrix. The scalar functions $P_f$ and $P_m$ stand for the pressures in the free-flow and matrix regions. The fluid viscosity is denoted by $\nu$. The thermal diffusivity coefficients may differ in the free-flow and matrix regions and are denoted by $\lambda_f$, $\lambda_m$, respectively. The viscosity and thermal diffusivity are allowed to be functions that may depend on the temperature $\theta$, which are physically important in the study of non-isothermal fluids (see, e.g., \cite{LB96}). In equation \eqref{um1}, $\mathbb{K}$ is a $d\times d$ matrix standing for the permeability of the porous medium, which is usually assumed to be a bounded, symmetric and uniformly positive definite matrix but could be heterogeneous \cite{Bear}. The parameter $\varpi$ in \eqref{um1} is a nonnegative constant related to the so-called Darcy--Prandtl number.
Since the Darcy--Prandtl number for porous medium flows is heuristically small in the regime with a small Darcy number \cite{Jo76}, the term $\varpi\partial_t\mathbf{u}_m$ is often neglected in the literature (see e.g., \cite{LT99,Va}), while in some other works, this time derivative term is kept primarily for the benefit of the energy analysis, and it would allow more accurate description of temporal transitions \cite{MMW19}. In this paper, we shall treat both cases with or without this time derivative term. For the sake of simplicity, physical coefficients that are not important for our subsequent mathematical analysis are set to be one (for instance, the fluid density and those in the Boussinesq approximation related to the gravitational acceleration and thermal expansion coefficient etc).
Next, we describe the initial, boundary as well as interface conditions of the coupled system \eqref{uf1}--\eqref{CT}.
\textbf{Initial conditions}. The system \eqref{uf1}--\eqref{CT} is subject to the initial conditions
\begin{align}
& \mathbf{u}_f|_{t=0}=\mathbf{u}_{0f}(x), \qquad \text{in}\ \Omega_f,\quad \text{and}\qquad \mathbf{u}_m|_{t=0}=\mathbf{u}_{0m}(x), &\text{in}\ \Omega_m,\label{ini1}\\
& \theta|_{t=0}=\theta_{0}(x), &\text{in}\ \Omega. \label{ini2}
\end{align}
In particular, when $\varpi=0$, we do not need to specify the initial velocity $\mathbf{u}_{0m}$ as it can be recovered from $\theta_{0m}$ by solving the Darcy equation (see e.g., \cite{LT99}). \medskip
\textbf{Boundary conditions on $\Gamma_f$ and $\Gamma_m$}. Since we are mainly interested in the coupling on the interface $\Gamma_i$ between sub-domains, we impose the following standard boundary conditions on the outer boundaries $\Gamma_f$ and $\Gamma_m$:
\begin{align}
& \mathbf{u}_f=\mathbf{0}, &\text{on}\ \Gamma_f\times(0,T),
\label{IBC0}\\
& \mathbf{u}_m\cdot\mathbf{n}_m=0, &\text{on}\ \Gamma_m\times(0,T),
\label{IBC1}\\
& \theta_f=0, &\text{on}\ \Gamma_f\times(0,T),
\label{IBC2}\\
& \theta_m=0, &\text{on}\ \Gamma_m\times(0,T).
\label{IBC3}
\end{align}
\textbf{Interface conditions on $\Gamma_{i}$}. Now on $\Gamma_i$, we assume that the system \eqref{uf1}--\eqref{CT} are coupled through the following set of interface conditions:
\begin{align}
&\mathbf{u}_f\cdot\mathbf{n}_{i} = \mathbf{u}_m\cdot\mathbf{n}_{i},
&\mbox{on}\ \Gamma_{i}\times(0,T),
\label{IBCi5}\\
&-\mathbf{n}_{i}\cdot(\mathbb{T}(\mathbf{u}_f,P_f){\mathbf{n}_{i}})+\frac12|\mathbf{u}_f|^2=P_m,
&\mbox{on}\ \Gamma_{i}\times(0,T),
\label{IBCi6}\\
&-\boldsymbol{\tau}_j\cdot(\mathbb{T}(\mathbf{u}_f,P_f){\mathbf{n}_{i}})
=\frac{\alpha\nu(\theta_m)}{\sqrt{\text{trace}(\mathbb{K})}}\boldsymbol{\tau}_j\cdot\mathbf{u}_f,\quad j=1,..,d-1,
&\mbox{on}\ \Gamma_{i}\times(0,T),
\label{IBCi7}\\
&\theta_f=\theta_m,&\mbox{on}\ \Gamma_{i}\times(0,T),
\label{IBCi1}\\
& \lambda_f(\theta_f)\frac{\partial \theta_f}{\partial \mathbf{n}_{i}}=\lambda_m(\theta_m)\frac{\partial \theta_m}{\partial \mathbf{n}_{i}},
&\mbox{on}\ \Gamma_{i}\times(0,T).
\label{IBCi2}
\end{align}
The condition \eqref{IBCi5} indicates the continuity in normal velocity that guarantees the conservation of mass, i.e., the exchange of fluid
between the two sub-domains is conservative. The condition \eqref{IBCi6}
represents the balance of the forces normal to the interface taking into consideration of the so-called dynamic pressure $\frac12|\mathbf{u}_f|^2$. With this quadratic term, condition \eqref{IBCi6} is known as the Lions interface condition in the literature (see e.g., \cite{CR09,DQ09,GR09}). This specific choice gives rise to a dissipative energy law that facilitates the analysis significantly \cite{CR08}. On the other hand, in the absence of this dynamic pressure term, the kinetic energy could increase without external forcing, which is physically unrealistic. Recently in \cite{MMW19}, the authors showed that the nonlinear dynamic pressure term is somewhat small, thus the difference between solutions produced with the Lions interface condition and its linear counterpart is heuristically small as well. More precisely, with a formal asymptotic argument, they showed that the order of the dynamic pressure term is $\mathcal{O}(\mathrm{Da})$ in the small Darcy number limit as $\mathrm{Da}\to 0$ and it begins to affect solutions to the perturbed systems at the scale
$\mathcal{O}(\mathrm{Da}^2)$ (see \cite[Appendix]{MMW19}).
Next, the interface condition \eqref{IBCi7} is
the so-called Beavers--Joseph--Saffman--Jones (BJSJ) condition (cf. \cite{Jo73, Sa71}), where $\alpha> 0$ is an empirical constant usually determined by the domain geometry and the porous material. We note that the BJSJ condition is a simplified variant of the well-known Beavers--Joseph (BJ) condition (see \cite{B-J}) that addresses how the porous medium affects the conduit flow at the interface:
$$-\boldsymbol{\tau}_j\cdot(2\nu\mathbb
D(\mathbf{u}_f))\mathbf{n}_{i}
=\frac{\alpha_{BJ}\nu}{\sqrt{\text{trace}(\mathbb{K})}}\boldsymbol{\tau}_j\cdot(\mathbf{u}_f-\mathbf{u}_m), \ \ \mbox{on}\ \Gamma_{i}, \ j=1,...,d-1.
$$
This empirical condition essentially says that the
tangential component of the normal stress that the free-flow incurs along the interface is proportional to the jump in the tangential velocity over the interface. To get the BJSJ condition, the term $-\boldsymbol{\tau}_j\cdot \mathbf{u}_m$ on the right-hand
side is simply dropped from the corresponding BJ condition
(as long as the Darcy number is small). Mathematically rigorous justification of this simplification under appropriate assumptions can be found in \cite{J-M}. At last, the interface conditions \eqref{IBCi1}--\eqref{IBCi2} involving $\theta$ are simply the continuity conditions for the temperature functions and their conormal derivatives across the interface (sometimes also referred to as transmission conditions).
The coupling system \eqref{uf1}--\eqref{CT} contains the Navier--Stokes--Darcy system for an incompressible viscous flow as a subsystem, which has been extensively studied in the literature. We just mention a few results related to the mathematics analysis and refer readers to the references cited therein.
For the simplified case of coupled (stationary) Stokes and Darcy equations, analysis of weak solutions has been done in \cite{CGHW10,LSY03,DMQ01}. We also refer to \cite{HWW} for the existence of global weak solutions and weak-strong uniqueness of a Cahn--Hilliard--Stokes--Darcy system for incompressible two-phase fluids. For the coupling of Navier--Stokes and Darcy equations, the stationary case has been studied in \cite{BDQ10,ChR09,DQ09, GR09}, and in \cite{CR12}, a Navier--Stokes/Darcy--transport system modelling the contamination of ground water was analyzed. On the other hand, the time-dependent problem have been investigated in \cite{CR08,CR09,CVR13} with different choices of interface conditions. In those works, existence and uniqueness of weak solutions are obtained under certain specific conditions, for instance, the small data assumption.
We also note that our system \eqref{uf1}--\eqref{CT} contains the Boussinesq system either for the free-flow \cite{FMT1987,HL05,LPZ11,LT16,LB96,LB99,SZ13,Temam1997,WZ11} or for the flow in a porous medium \cite{F86,LT99}, which have been studied in a vast literature under various settings. In our current case, we have to deal with new difficulties due to the complicated coupling of flows governed by different physical processes, the complex geometry of domain, and in particular, the coupled nonlinear interface conditions.
The aim of this paper is two-fold. First, we prove the existence of global weak solutions to problem \eqref{uf1}--\eqref{IBCi2} with variable fluid viscosity, thermal diffusivity and a nonnegative Darcy--Prandtl number $\varpi$ in both two and three dimensions (see Theorem \ref{thmEx}). Here we choose to work with the Lions and BJSJ interface conditions, which lead to a dissipative energy law for the solutions so that no smallness assumptions on the initial data are necessary (cf. \cite{CR08,CVR13}). The proof is based on a semi-implicit discretization scheme with a Brinkman type regularization that can be solved by the Leray--Schauder principle (cf. \cite{ADG13,HWW} for related applications to some hydrodynamic systems for two-phase flows). Using the energy inequality, we derive uniform estimates of approximate solutions and then obtain the existence of global weak solutions to the original system by a compactness argument.
Second, we prove the weak-strong uniqueness property of problem \eqref{uf1}--\eqref{IBCi2} (see Theorem \ref{thmuniq}), namely, a weak solution coincides with a strong solution emanating from the same initial data as long as the latter exists. Uniqueness of weak solutions to problem \eqref{uf1}--\eqref{IBCi2} is not trivial even when the spatial dimension is two due to the nonlinear interface condition of Lions type. Besides the Navier--Stokes equations, additional difficulties come from the thermal advection terms, low regularity of the solution $\mathbf{u}_m$ to Darcy's equation and variable viscosity/thermal diffusivity coefficients. We remark that our approach can be further applied to study the Cahn--Hilliard--Navier--Stokes--Darcy--Boussinesq system for thermal convection of two-phase flows in a fluid layer overlying a porous medium (see e.g., \cite{CHWZ20}).
Finally, due to the decomposed domain setting in the B\'enard convection problem \eqref{uf1}--\eqref{IBCi2} and in particular, the complicated coupling via interface conditions, existence of strong or classical solutions (especially with higher-order spatial regularity) remains a challenging open problem, which is quite different from the case in a simple domain (cf. \cite{FMT1987,HL05,LPZ11,LB99,LT99,SZ13} and the references therein). We refer to \cite{LL19} for an attempt on the Navier--Stokes--Darcy system for isothermal incompressible flows in a two-dimensional strip domain.
The rest of this paper is organized as follows.
In Section 2, we introduce the function spaces, present the definition of weak solutions and state the main results. Section 3 is devoted to the proof for the existence of global weak solutions to problem \eqref{uf1}--\eqref{IBCi2}. In Section 4, we prove the weak-strong uniqueness property by the energy method.
\section{Main Results}
\setcounter{equation}{0}
\subsection{Preliminaries}
Let $d=2,3$ be the spatial dimension. For arbitrary vectors $\mathbf{a},\mathbf{b}\in \mathbb{R}^d$, we denote $\mathbf{a}\otimes \mathbf{b}=(a_jb_k)^d_{j,k=1}$ and $\mathbf{a}\cdot \mathbf{b}=\sum_{j=1}^d a_jb_j$.
Let $X$ be a Banach space with its dual denoted by $X'$. We denote by
$\langle u, v\rangle \equiv \langle u, v\rangle_{X'
,X} $ the duality product for $u \in X'$, $v\in X$. The inner product on a Hilbert space $H$ is denoted by $(\cdot, \cdot)_H$.
We use $L^q(\Omega)$, $1 \leq q \leq +\infty$ to denote the usual Lebesgue space on $\Omega$ and $\|\cdot\|_{L^q(\Omega)}$ for its norm. Similarly, $W^{m,q}(\Omega)$, $m \in \mathbb{N}$, $1 \leq q \leq + \infty$, denote the usual
Sobolev spaces with norm $\|\cdot \|_{W^{m,p}(\Omega)}$, and for $q=2$, we simply denote $W^{m,2}(\Omega)$ by $H^m(\Omega)$. The fractional order Sobolev spaces $H^s(\Omega)$ ($s\in \mathbb{R}$) are defined as in \cite[Section 4.2.1]{Tri}. If $I$ is an interval of $\mathbb{R}^+$, we use $L^p(I;X)$, $1 \leq p \leq +\infty$, to denote the function space that consists of $p$-integrable
functions with values in $X$. Moreover, $C_w(I;X)$ denotes the topological space of all bounded and weakly continuous functions from $I$ to $X$, while $W^{1,p}(I;X)$ with $1\leq q<+\infty$ stands for the space of all functions $u$ such that $u, \frac{du}{dt}\in L^p(I;X)$, where $\frac{du}{dt}$ denotes the distributional time derivative of $u$. Bold characters are used to denote vector or matrix valued spaces.
Let $\Omega$ be the domain with decomposition that has been described in the Introduction. For our problem, we introduce the following spaces
\begin{eqnarray}
\mathbf{H}({\rm div}; \Omega_k) &:=&\{\mathbf{w}\in \mathbf{L}^2(\Omega_k)~|~\nabla \cdot \mathbf{w}\in L^2(\Omega_k)\},\nonumber \\
\widetilde{\mathbf{H}}_{k,0}&:=&\{\mathbf{w}\in \mathbf{H}({\rm div}; \Omega_k)~|~\mathbf{w}\cdot \mathbf{n}_k=0 \ \text{on}\ \Gamma_{k}\}, \nonumber \\
\widetilde{\mathbf{H}}_{k,\mathrm{div}}
&:=&\{\mathbf{w}\in\widetilde{\mathbf{H}}_{k,0}~|~\nabla \cdot\mathbf{w}=0\},\nonumber \\
\mathbf{H}_{k,0}&:=&\{\mathbf{w}\in \mathbf{H}^1(\Omega_k)~|~\mathbf{w}=\mathbf{0}\text{ on
}\Gamma_{k}\}, \nonumber \\
\mathbf{H}_{k,\text{div}}&:=&\{
\mathbf{w}\in\mathbf{H}_{k,0}~|~\nabla \cdot\mathbf{w}=0\},\nonumber\\
\widehat{\mathbf{H}}_{k,0}&:=&\{\mathbf{w}\in \mathbf{H}^1(\Omega_k)~|~\mathbf{w}=\mathbf{0}\text{ on
}\partial\Omega_{k}\}, \nonumber \\
\widehat{\mathbf{H}}_{k,\text{div}}&:=&\{
\mathbf{w}\in\widehat{\mathbf{H}}_{k,0}~|~\nabla \cdot\mathbf{w}=0\},\nonumber
\end{eqnarray}
with index $k\in \{f,m\}$.
For simplicity, we denote $(\cdot, \cdot)_f$, $(\cdot, \cdot)_m$ the inner products on the spaces $L^2(\Omega_f)$, $L^2(\Omega_m)$, respectively (also for the corresponding vector or matrix valued spaces). The inner product on $L^2(\Omega)$ is simply denoted by $(\cdot, \cdot)$. For any function $u\in L^2(\Omega)$ with $u_m:=u|_{\Omega_m}$ and $u_f:=u|_{\Omega_f}$, it holds
$$
(u,v)=(u_m,v_m)_m+(u_f,v_f)_f, \quad \|u\|_{L^2(\Omega)}^2=\|u_m\|_{L^2(\Omega_m)}^2+\|u_f\|_{L^2(\Omega_f)}^2.
$$
On the interface $\Gamma_{i}$, we consider the fractional Sobolev spaces
$H^{\frac12}_{00}(\Gamma_{i})$ and $H^\frac12(\Gamma_{i})$ for a (Lipschitz) surface $\Gamma_{i}$ when $d=3$ or a curve when $d=2$, with the following equivalent norms (see \cite[Chapter 1, Section 11]{L-M}, or \cite{Gri85}):
\begin{eqnarray}
&& \|u\|_{H^\frac12(\Gamma_{i})}^2=\int_{\Gamma_{i}} |u|^2 dS +\int_{\Gamma_{i}}\!\int_{\Gamma_{i}}\frac{|u(x)-u(y)|^2}{|x-y|^d} dxdy,\nonumber\\
&& \|u\|_{H_{00}^\frac12(\Gamma_{i})}^2=\|u\|_{H^\frac12(\Gamma_{i})}^2
+\int_{\Gamma_{i}} \frac{|u(x)|^2}{\rho(x, \partial \Gamma_{i})} dx,\nonumber
\end{eqnarray}
where $\rho(x, \partial \Gamma_{i})$ denotes the distance from $x$ to $\partial \Gamma_{i}$. We note that these norms are not equivalent except when $\Gamma_{i}$ is a closed surface or curve. Besides, if $\Gamma_{i}$ is a subset of $\partial \Omega_f$ with positive measure, then $H^{\frac12}_{00}(\Gamma_{i})$ is a trace space of functions of $H^1(\Omega_f)$ that vanish on $\Gamma_{f}$ (see \cite{CVR13}). Similarly in the vectorial case, we have $\mathbf{H}^{\frac12}_{00}(\Gamma_{i})=\mathbf{H}_{f,0}|_{\Gamma_{i}}$. $H^{\frac12}_{00}(\Gamma_{i})$ is a non-closed subspace of
$H^{\frac12}(\Gamma_{i})$ and it has a continuous zero extension to
$H^{\frac12}(\partial\Omega_{f})$.
Moreover, we have the following continuous embedding result (see \cite{CGHW10}): $H^{\frac12}_{00}(\Gamma_{i})\subsetneqq H^{\frac12}(\Gamma_{i}) \subsetneqq H^{-\frac12}(\Gamma_{i}) \subsetneqq(H^{\frac12}_{00}(\Gamma_{i}))'$.
Let $H^{-\frac12}(\partial\Omega_f)|_{\Gamma_{i}}$ be defined in the
following way: for all $f\in H^{-\frac12}(\partial\Omega_f)|_{\Gamma_{i}}$ and $ g\in
H^{\frac12}_{00}(\Gamma_{i})$,
$\langle f,g\rangle_{H^{-\frac12}(\partial\Omega_f)|_{\Gamma_{i}},\,
H^{\frac12}_{00}(\Gamma_{i})}:=\langle f,\widetilde g\rangle_{H^{-\frac12}(\partial\Omega_f),\,
H^{\frac12}(\partial\Omega_f)}$ with $\widetilde g$ being the zero
extension of $g$ to $\partial\Omega_f$. Then we note that
$H^{-\frac12}(\partial\Omega_f)|_{\Gamma_{i}} \subset(H^{\frac12}_{00}(\Gamma_{i}))'$ but
$H^{-\frac12}(\partial\Omega_f)|_{\Gamma_{i}} \nsubseteq H^{-\frac12}(\Gamma_{i})$.
For any vector $\mathbf{u} \in \mathbf{H}({\rm div}; \Omega_f)$, its normal component $\mathbf{u} \cdot \mathbf{n}_{i}$ is well defined in $(H^{\frac12}_{00}(\Gamma_{i}))^\prime$, and for all $q \in H^1(\Omega_f)$ such that $q =0$ on $\Gamma_{f}$, we have
\begin{align}
(\nabla \cdot \mathbf{u}, q)_f=-(\mathbf{u}, \nabla q)_f+\langle \mathbf{u}\cdot \mathbf{n}_{i}, q\rangle_{(H^{\frac12}_{00}(\Gamma_{i}))^\prime,\, H^{\frac12}_{00}(\Gamma_{i})}.\nonumber
\end{align}
Similar results hold also on the sub-domain $\Omega_m$.
In our decomposed domain setting, the boundary $\Gamma_f=\emptyset$ is allowed, i.e., $\Omega_f$ can be enclosed completely by the matrix part $\Omega_m$. Since the classical Korn's inequality (see, e.g., \cite{Ho}) may not apply when $\Gamma_f=\emptyset$, in order to overcome this difficulty, we introduce the space
\begin{align}
\mathbf{Z}=\big\{\mathbf{u}\ |\ \mathbf{u}_f=\mathbf{u}|_{\Omega_f}\in \mathbf{H}_{f, \mathrm{div}},\ \mathbf{u}_m=\mathbf{u}|_{\Omega_m}\in \widetilde{\mathbf{H}}_{m,\mathrm{div}}, \ \mathbf{u}_f\cdot\mathbf{n}_{i}=\mathbf{u}_m\cdot\mathbf{n}_{i} \ \text{on}\ \Gamma_{i}\big\},
\label{Z}
\end{align}
whose natural norm can be given by $\|\mathbf{u}_f\|_{\mathbf{H}^1(\Omega_f)}+
\|\mathbf{u}_m\|_{\mathbf{L}^2(\Omega_m)}$.
In view of \cite[Lemma 3.9]{HWW}, we have the following result:
\begin{lemma}
\label{equinorml}
The norm defined by
\begin{align}\label{equinorm}
\|\mathbf{u}\|_{\mathbf{Z}}^2
:=\|\mathbb{D}(\mathbf{u}_f)\|_{\mathbf{L}^2(\Omega_f)}^2
+\sum_{j=1}^{d-1}\|\mathbf{u}_f \cdot \mathbf{\boldsymbol{\tau}}_j\|_{L^2(\Gamma_{i})}^2+
\|\mathbf{u}_m\|_{\mathbf{L}^2(\Omega_m)}^2
\end{align}
is an equivalent norm on $\mathbf{Z}$. There exists a constant $C$ independent of $\mathbf{u}$ such that
$$
\|\mathbf{u}_f\|_{\mathbf{H}^1(\Omega_f)}^2+
\|\mathbf{u}_m\|_{\mathbf{L}^2(\Omega_m)}^2\leq C\|\mathbf{u}\|_{\mathbf{Z}}^2,\qquad \forall\, \mathbf{u}\in \mathbf{Z}.
$$
\end{lemma}
\subsection{Main results}\label{WM}
We make the following assumptions on the viscosity $\nu$, thermal diffusivity $\lambda_f$, $\lambda_m$ as well as the permeability matrix $\mathbb{K}$.
\begin{itemize}
\item[(A1)] $\nu\in C^1(\mathbb{R})$, $\underline{\nu} \leq \nu(s)\leq\bar{\nu} $ and $|\nu'(s)|\leq \tilde{\nu}$ for $s\in \mathbb{R}$, where $\bar{\nu}$, $\underline{\nu}$ and $\tilde{\nu}$ are positive constants.
\item[(A2)] ${\lambda}_j\in C^1 (\mathbb{R})$, $\underline{\lambda} \leq {\lambda}_j(s)\leq \bar \lambda $ and $|{\lambda}_j'(s)|\leq \tilde{\lambda}$ for $s\in \mathbb{R}$, where $\bar{\lambda}$, $\underline{\lambda}$ and $\tilde{\lambda}$ are positive constants, $j\in \{f,m\}$.
\item[(A3)] The permeability matrix $\mathbb{K}$ is isotropic, bounded from above and below, namely, $\mathbb{K}=\kappa(x)\mathbb{I}$ with $\mathbb{I}$ being the $d\times d$ identity matrix and $\kappa(x)\in L^\infty (\Omega)$ such that there exist $\bar{\kappa}>\underline{\kappa}>0$, $\underline{\kappa}\leq \kappa(x)\leq \bar{\kappa}$ a.e. in $\Omega$.
\end{itemize}
Next, we introduce the notion of weak solutions.
\begin{definition}\label{defweak}
Suppose that $d=2,3$ and $T>0$ is arbitrary.
\textbf{Case 1}: $\varpi>0$. Consider the initial data $\mathbf{u}_{0f}(x)\in \widetilde{\mathbf{H}}_{f,\mathrm{div}}$, $\mathbf{u}_{0m}(x)\in \widetilde{\mathbf{H}}_{m,\mathrm{div}}$ with
$\mathbf{u}_{0f}\cdot\mathbf{n}_{i} = \mathbf{u}_{0m}\cdot\mathbf{n}_{i}$
on $\Gamma_{i}$, and $\theta_{0}\in L^2(\Omega)$. The triple $(\mathbf{u}_f, \mathbf{u}_m, \theta)$ satisfying
\begin{eqnarray}
&& \mathbf{u}_f\in C_w([0,T]; \widetilde{\mathbf{H}}_{f,\mathrm{div}})\cap L^2(0, T; \mathbf{H}_{f,\mathrm{div}})\cap W^{1,\frac43}(0,T; (\mathbf{H}_{f,\mathrm{div}})'), \label{regubc}\\
&& \mathbf{u}_m\in C_w([0, T]; \widetilde{\mathbf{H}}_{m,\mathrm{div}}) \cap L^2(0, T; \widetilde{\mathbf{H}}_{m,\mathrm{div}})\cap H^1(0,T; (\widetilde{\mathbf{H}}_{m,\mathrm{div}})'), \label{regubm}\\
&& \theta\in C_w([0,T]; L^2(\Omega))\cap L^2(0,T; H^1_0(\Omega))\cap H^1(0,T; (W^{1,3}_0(\Omega))'),
\end{eqnarray}
is called a weak solution to problem \eqref{uf1}--\eqref{IBCi2}, if the following conditions are fulfilled:
(1) For any $\mathbf{v}_f\in C^1_0((0,T); \mathbf{H}_{f,\mathrm{div}})$, $\mathbf{v}_m\in C^1_0((0,T); \widetilde{\mathbf{H}}_{m,\mathrm{div}})$ with $\mathbf{v}_f\cdot \mathbf{n}_i=\mathbf{v}_m\cdot \mathbf{n}_i$ on $\Gamma_{i}$, it holds
\begin{eqnarray}
&&-\int_0^T(\mathbf{u}_f,\partial_t\mathbf{v}_f)_f dt
-\varpi\int_0^T(\mathbf{u}_m,\partial_t\mathbf{v}_m)_m dt\nonumber \\
&&
+\int_0^T \big(\mathrm{div}(\mathbf{u}_f\otimes \mathbf{u}_f), \mathbf{v}_f\big)_fdt
+2\int_0^T\big(\nu(\theta_f)\mathbb{D}(\mathbf{u}_f),\mathbb{D}(\mathbf{v}_f)\big)_fdt
\nonumber \\
&& + \int_0^T \left(\nu(\theta_m)\mathbb{K}^{-1}\mathbf{u}_m,\mathbf{v}_m)\right)_m dt
\nonumber \\
&&+\sum_{j=1}^{d-1} \int_0^T\!\int_{\Gamma_{i}}\frac{\alpha \nu(\theta_m)}{\sqrt{{\rm trace} (\mathbb{K})}} (\mathbf{u}_f\cdot\boldsymbol{\tau}_j)(\mathbf{v}_f\cdot\boldsymbol{\tau}_j)dSdt\nonumber \\
&&- \int_0^T \!\int_{\Gamma_{i}} \frac12|\mathbf{u}_f|^2 (\mathbf{v}_f\cdot \mathbf{n}_{i}) dSdt \nonumber \\
&=& \int_0^T (\theta_f \mathbf{k}, \mathbf{v}_f )_f dt +
\int_0^T (\theta_m \mathbf{k}, \mathbf{v}_m )_m dt.\label{weak1}
\end{eqnarray}
(2) For any $\phi\in C_0^1((0,T); W_0^{1,3}(\Omega))$, it holds
\begin{eqnarray}
&&-\int_0^T(\theta,\partial_t\phi)dt+\int_0^T(\lambda(\theta) \nabla \theta,\nabla\phi)dt=\int_0^T(\theta \mathbf{u}, \nabla \phi)dt.
\label{weak3}
\end{eqnarray}
(3) $\mathbf{u}_f|_{t=0}=\mathbf{u}_{0f}(x)$, $\mathbf{u}_m|_{t=0}=\mathbf{u}_{0m}(x)$, $\theta|_{t=0}=\theta_{0}(x)$.\medskip
\textbf{Case 2}: $\varpi=0$. Consider the initial data $\mathbf{u}_{0f}(x)\in \widetilde{\mathbf{H}}_{f,\mathrm{div}}$, $\theta_{0}\in L^2(\Omega)$. The triple $(\mathbf{u}_f, \mathbf{u}_m, \theta)$ satisfying
\begin{eqnarray}
&& \mathbf{u}_f\in C_w([0,T]; \widetilde{\mathbf{H}}_{f,\mathrm{div}})\cap L^2(0, T; \mathbf{H}_{f,\mathrm{div}})\cap W^{1,\frac43}(0,T; (\mathbf{H}_{f,\mathrm{div}})'),\label{regubc0}\\
&& \mathbf{u}_m\in L^2(0, T; \widetilde{\mathbf{H}}_{m,\mathrm{div}}),\label{regubm0}\\
&& \theta\in C_w([0,T]; L^2(\Omega))\cap L^2(0,T; H^1_0(\Omega))\cap W^{1,\frac87}(0,T; (W^{1,4}_0(\Omega))'),
\end{eqnarray}
is called a weak solution to problem \eqref{uf1}--\eqref{IBCi2}, if the following conditions are satisfied:
(1) For any $\mathbf{v}_f\in C^1_0((0,T); \mathbf{H}_{f,\mathrm{div}})$, $\mathbf{v}_m\in C^1_0((0,T); \widetilde{\mathbf{H}}_{m,\mathrm{div}})$ with $\mathbf{v}_f\cdot \mathbf{n}_i=\mathbf{v}_m\cdot \mathbf{n}_i$ on $\Gamma_{i}$, it holds
\begin{eqnarray}
&&-\int_0^T(\mathbf{u}_f,\partial_t\mathbf{v}_f)_f dt
+\int_0^T \big(\mathrm{div}(\mathbf{u}_f\otimes \mathbf{u}_f), \mathbf{v}_f\big)_fdt\nonumber \\
&&
+2\int_0^T\big(\nu(\theta_f)\mathbb{D}(\mathbf{u}_f),\mathbb{D}(\mathbf{v}_f)\big)_fdt
+ \int_0^T \left(\nu(\theta_m)\mathbb{K}^{-1}\mathbf{u}_m,\mathbf{v}_m)\right)_m dt
\nonumber \\
&&+\sum_{j=1}^{d-1} \int_0^T\!\int_{\Gamma_{i}}\frac{\alpha \nu(\theta_m)}{\sqrt{{\rm trace} (\mathbb{K})}} (\mathbf{u}_f\cdot\boldsymbol{\tau}_j)(\mathbf{v}_f\cdot\boldsymbol{\tau}_j)dSdt\nonumber \\
&&- \int_0^T\! \int_{\Gamma_{i}} \frac12|\mathbf{u}_f|^2 (\mathbf{v}_f\cdot \mathbf{n}_{i}) dSdt \nonumber \\
&=& \int_0^T (\theta_f \mathbf{k}, \mathbf{v}_f )_f dt+
\int_0^T (\theta_m \mathbf{k}, \mathbf{v}_m )_m dt.\label{weak10}
\end{eqnarray}
(2) For any $\phi\in C_0^1((0,T); W_0^{1,4}(\Omega))$, it holds
\begin{eqnarray}
&&-\int_0^T(\theta,\partial_t\phi)dt+\int_0^T(\lambda(\theta) \nabla \theta,\nabla\phi)dt=\int_0^T(\theta \mathbf{u}, \nabla \phi)dt.
\label{weak30}
\end{eqnarray}
(3) $\mathbf{u}_f|_{t=0}=\mathbf{u}_{0f}(x)$, $\theta|_{t=0}=\theta_{0}(x)$.
\end{definition}
%
\begin{remark}
We note that the interface/boundary conditions \eqref{IBC0}--\eqref{IBCi2} are enforced as consequences of the weak formulation stated in Definition \ref{defweak}. The equivalence for smooth (or strong) solutions between the weak formulation and the classical form can be verified in a straightforward way using integration by parts. We may refer to \cite{CR08} for detailed computations about the two dimensional Navier--Stokes--Darcy system, and we mention \cite{Zha17} for the weak formulation of second order parabolic transmission problems. Here in \eqref{weak3}, \eqref{weak30} and below, we always use the convention $$\lambda(\theta)|_{\Omega_f}=\lambda_f(\theta_f),\quad \lambda(\theta)|_{\Omega_m}=\lambda_m(\theta_m)$$ for the thermal diffusivity.
\end{remark}
\smallskip
Now we are in a position to state the main results of this paper.
\begin{theorem}[Existence of global weak solutions]\label{thmEx}
Suppose that $d=2,3$, $T>0$ being arbitrary and the assumptions (A1)--(A3) are satisfied.
\begin{itemize}
\item[(i)] If $\varpi>0$, for any $\mathbf{u}_{0f}\in \widetilde{\mathbf{H}}_{f,\mathrm{div}}$, $\mathbf{u}_{0m}\in\widetilde{\mathbf{H}}_{m,\mathrm{div}}$ with $\mathbf{u}_{0f}\cdot\mathbf{n}_{i} = \mathbf{u}_{0m}\cdot\mathbf{n}_{i}$
on $\Gamma_{i}$, and $\theta_0 \in L^2(\Omega)$, problem \eqref{uf1}--\eqref{IBCi2} admits at least one global weak solution $(\mathbf{u}_f, \mathbf{u}_m, \theta)$ on $[0,T]$.
\item[(ii)] If $\varpi=0$, for any $\mathbf{u}_{0f}\in \widetilde{\mathbf{H}}_{f,\mathrm{div}}$ and $\theta_0 \in L^2(\Omega)$, problem \eqref{uf1}--\eqref{IBCi2} admits at least one global weak solution $(\mathbf{u}_f, \mathbf{u}_m, \theta)$ on $[0,T]$.
\end{itemize}
\end{theorem}
\begin{remark}
The pressure terms $P_f, P_m$ will be understood in the distributional sense and can be constructed, for instance, as in \cite[Chapter 3, Section 3.5]{Temam1977}. Besides, when $\varpi=0$, the pressure $P_m$ in the matrix part can be also regarded as a weak solution of the Neumann problem
\begin{equation*}
\begin{cases}
-\Delta P_m=\mathrm{div}(\nu(\theta_m)\mathbb{K}^{-1}\mathbf{u}_m- \theta_m\mathbf{k}),\qquad\quad\ \ \text{in}\ \Omega_m,\\
\partial_{\widehat{\mathbf{n}}_m}P_m=(\nu(\theta_m)\mathbb{K}^{-1}\mathbf{u}_m- \theta_m\mathbf{k})\cdot \widehat{\mathbf{n}}_m, \qquad \ \text{on}\ \partial\Omega_m.
\end{cases}
\end{equation*}
For any function $h\in H^1(\Omega)$, it holds
\begin{align*}
& \big\langle \mathrm{div}(\nu(\theta_m)\mathbb{K}^{-1}\mathbf{u}_m- \theta_m\mathbf{k}),\,h\big\rangle_{(H^1(\Omega_m))',H^1(\Omega_m)}\\
&\quad \leq C(\|\mathbf{u}_m\|_{\mathbf{L}^2(\Omega_m)}+\|\theta_m\|_{L^2(\Omega_m)})\|h\|_{H^1(\Omega_m)}.
\end{align*}
On the other hand, at least for the simple case when $\nu$, $\kappa$ are positive constants, one can verify that $(\nu \mathbb{K}^{-1}\mathbf{u}_m- \theta_m\mathbf{k})\cdot \widehat{\mathbf{n}}_m \in L^2(0,T;H^{-\frac12}(\partial \Omega_m))$. Then we can obtain $P_m\in L^2(0,T; H^1(\Omega_m))$.
\end{remark}
Our second result concerns the uniqueness property of solutions to problem \eqref{uf1}--\eqref{IBCi2}. More precisely, we deduce the following weak-strong uniqueness result in both two and three dimensions:
\begin{theorem}[Weak-strong uniqueness]\label{thmuniq}
Suppose that $d=2,3$, $\varpi\geq 0$ and the assumptions (A1)--(A3) are satisfied.
Let $(\mathbf{u}_f,\mathbf{u}_m,\theta)$, $(\bar{\mathbf{u}}_f,\bar{\mathbf{u}}_m,\bar{\theta})$ be two solutions to problem \eqref{uf1}--\eqref{IBCi2} on a certain time interval $[0,T]$, both emanating from the same initial data $\mathbf{u}_{0f}\in \widetilde{\mathbf{H}}_{f,\mathrm{div}}$, $\varpi \mathbf{u}_{0m}\in\widetilde{\mathbf{H}}_{m,\mathrm{div}}$ with $\mathbf{u}_{0f}\cdot\mathbf{n}_{i} = \mathbf{u}_{0m}\cdot\mathbf{n}_{i}$
on $\Gamma_{i}$ (if $\varpi>0$) and $\theta_0 \in L^2(\Omega)$. In particular, we assume that
$(\mathbf{u}_f,\mathbf{u}_m,\theta)$ is a global weak solution obtained in Theorem \ref{thmEx} and $(\bar{\mathbf{u}}_f,\bar{\mathbf{u}}_m,\bar{\theta})$ is a strong solution with the following additional regularity
\begin{equation}
\bar{\mathbf{u}}_f\in L^4(0,T;\mathbf{W}^{1,6}(\Omega_f)),\ \bar{\mathbf{u}}_m\in L^4(0,T;\mathbf{L}^6(\Omega_m)),\ \bar{\theta}\in L^8(0,T;W^{1,4}(\Omega)).
\label{areg}
\end{equation}
Then it holds $$(\mathbf{u}_f,\mathbf{u}_m,\theta)=(\bar{\mathbf{u}}_f,\bar{\mathbf{u}}_m,\bar{\theta}),\quad \text{on}\ [0,T].$$
\end{theorem}
\begin{remark}
The additional regularity conditions \eqref{areg} can be weakened for some special cases. For instance, if the fluid viscosity $\nu$ is a positive constant, the condition $\bar{\mathbf{u}}_f\in L^4(0,T;\mathbf{W}^{1,6}(\Omega_f))$ can be replaced by the classical condition $\bar{\mathbf{u}}_f\in L^4(0,T;\mathbf{H}^{1}(\Omega_f))$ (see \cite{Be95}), and the condition $\bar{\mathbf{u}}_m\in L^4(0,T;\mathbf{L}^6(\Omega_m))$ can simply be dropped. Besides, in the two dimensional case, if we assume that $\nu, \lambda_f, \lambda_m$ are all positive constants and $\theta_0\in L^\infty(\Omega)$, then one can easily check that the conclusion on uniqueness holds under just one additional regularity condition $\bar{\mathbf{u}}_f\in L^\eta(0,T;\mathbf{H}^{1}(\Omega_f))$, for any $\eta>2$. Different from the case for a single homogeneous incompressible fluid in a simple domain (i.e., the classical uniqueness result for weak solutions of the Navier--Stokes equations in 2D), this additional requirement with $\eta>2$ is essentially due to the nonlinear Lions interface condition \eqref{IBCi6}.
\end{remark}
\section{Existence of Global Weak Solutions} \setcounter{equation}{0}
In this section, we prove Theorem \ref{thmEx} on the existence of global weak solutions to problem \eqref{uf1}--\eqref{IBCi2}. The proof will be given in the three dimensional case and the two dimensional case can be easily treated with minor modifications.
First, we recall an important feature of problem \eqref{uf1}--\eqref{IBCi2}, that is, it obeys a basic energy law which can lead to certain nonlinear stability of the system under suitable assumptions (see e.g., \cite{MMW19}). Denote the total energy of the coupled system by
\begin{equation}
\mathcal{E}_\sigma(t)=\int_{\Omega_f}\frac{1}{2}|\mathbf{u}_f|^2 dx
+ \int_{\Omega_m} \frac{\varpi}{2}|\mathbf{u}_m|^2 dx
+ \int_{\Omega}\frac{\sigma}{2}|\theta|^2dx,
\label{totenergy}
\end{equation}
for $\varpi\geq 0$ and some $\sigma>0$, where $\sigma$ is an arbitrary positive constant and it will be chosen in a suitable way below. By a similar calculation like in \cite[Section 4]{MMW19}, we have the following formal result:
\begin{lemma}[Basic energy law]\label{BEL}
Let $(\mathbf{u}_m, \mathbf{u}_f, \theta)$ be a smooth solution to the initial boundary value problem \eqref{uf1}--\eqref{IBCi2} on $[0,T]$. Then $(\mathbf{u}_m, \mathbf{u}_f, \theta)$ satisfies the following energy equality:
\begin{equation}
\frac{d}{dt} \mathcal{E}_\sigma (t)+\mathcal{D}_\sigma (t)= \mathcal{R}(t),\quad \forall\, t\in (0,T),
\label{EnergyLaw}
\end{equation}
where $\mathcal{E}_\sigma(t)$ is given by \eqref{totenergy} and
\begin{eqnarray}
\mathcal{D}_\sigma (t)
&=& \int_{\Omega_f}2\nu(\theta_f)|\mathbb{D}(\mathbf{u}_f)|^2dx
+\int_{\Omega_m} \nu(\theta_m)\mathbb{K}^{-1}|\mathbf{u}_m|^2dx
\nonumber \\
&&
+ \sum_{j=1}^{d-1}\int_{\Gamma_{i}} \frac{\alpha\nu(\theta_m)}{\sqrt{{\rm trace}(\mathbb{K})}}
|\mathbf{u}_f\cdot\boldsymbol{\tau}_j|^2 dS
+ \sigma \int_{\Omega}\lambda(\theta)|\nabla\theta|^2dx,\label{D}\\
\mathcal{R}(t)&=&\int_{\Omega_f} (\mathbf{u}_f\cdot \mathbf{k})\theta_f dx+\int_{\Omega_m} (\mathbf{u}_m\cdot \mathbf{k})\theta_m dx.
\label{R}
\end{eqnarray}
\end{lemma}
Inspired by \cite{ADG13,HWW}, below we apply a semi-discretization approach to prove Theorem \ref{thmEx}. First, we introduce a discrete in time, continuous in space numerical scheme for a regularized system with an approximation of Brinkman's type in the Darcy equation. The existence of weak solutions to the regularized discrete problem is then proved by using the Leray--Schauder principle. After that, we construct approximate solutions and derive uniform estimates using a discrete version of the basic energy law. Finally, by a two-step compactness argument we show the convergence of approximate solutions to a global weak solution of the original problem \eqref{uf1}--\eqref{IBCi2}.
\subsection{Time discretization of a regularized problem}
Let $\xi\in (0,1)$ be an arbitrary but fixed constant. We consider the following weak formulation of a regularized problem for the original one \eqref{uf1}--\eqref{IBCi2}:
\begin{eqnarray}
&&-\int_0^T(\mathbf{u}_f,\partial_t\mathbf{v}_f)_f dt
-\varpi\int_0^T(\mathbf{u}_m,\partial_t\mathbf{v}_m)_m dt\nonumber \\
&&
+\int_0^T \big(\mathrm{div}(\mathbf{u}_f\otimes \mathbf{u}_f), \mathbf{v}_f\big)_fdt
+2\int_0^T\big(\nu(\theta_f)\mathbb{D}(\mathbf{u}_f),\mathbb{D}(\mathbf{v}_f)\big)_fdt
\nonumber \\
&& + \int_0^T \left(\nu(\theta_m)\mathbb{K}^{-1}\mathbf{u}_m,\mathbf{v}_m)\right)_m dt +\xi \int_0^T \left(\nabla \mathbf{u}_m, \nabla \mathbf{v}_m\right)_m dt
\nonumber \\
&&+\sum_{j=1}^{d-1} \int_0^T\!\int_{\Gamma_{i}}\frac{\alpha \nu(\theta_m)}{\sqrt{{\rm trace} (\mathbb{K})}} (\mathbf{u}_f\cdot\boldsymbol{\tau}_j)(\mathbf{v}_f\cdot\boldsymbol{\tau}_j)dSdt\nonumber \\
&&- \int_0^T\! \int_{\Gamma_{i}} \frac12|\mathbf{u}_f|^2 (\mathbf{v}_f\cdot \mathbf{n}_{i}) dSdt \nonumber \\
&=& \int_0^T (\theta_f \mathbf{k}, \mathbf{v}_f )_f dt +
\int_0^T (\theta_m \mathbf{k}, \mathbf{v}_m )_m dt,
\label{weak1r}
\end{eqnarray}
for any $\mathbf{v}_f\in C^1_0((0,T); \mathbf{H}_{f,\mathrm{div}})$, $\mathbf{v}_m\in C^1_0((0,T); \mathbf{H}_{m,\mathrm{div}})$ with $\mathbf{v}_f\cdot \mathbf{n}_i=\mathbf{v}_m\cdot \mathbf{n}_i$ on $\Gamma_i$, and
\begin{eqnarray}
&&-\int_0^T(\theta,\partial_t\phi)dt+\int_0^T(\lambda(\theta) \nabla \theta,\nabla\phi)dt=-\int_0^T( \mathbf{u}\cdot \nabla \theta, \phi)dt,
\label{weak3r}
\end{eqnarray}
for any $\phi\in C_0^1((0,T); H_0^1(\Omega))$. Besides, the following initial conditions are satisfied:
\begin{align}
\mathbf{u}_f|_{t=0}=\mathbf{u}_{0f}(x),\quad \varpi\mathbf{u}_m|_{t=0}=\varpi\mathbf{u}_{0m}(x),\quad \theta|_{t=0}=\theta_{0}(x).
\label{weakrini}
\end{align}
We introduce a semi-implicit time discretization scheme for the regularized problem \eqref{weak1r}--\eqref{weakrini}.
For arbitrary but fixed $T>0$ and a positive integer $N\in \mathbb{N}$, we denote by $\delta=\Delta t=\frac{T}{N}$ the size of time step. Given a triple $(\mathbf{u}_f^k, \mathbf{u}_m^k, \theta^k)$, $k=0,1,2,...,N-1$, our aim is to determine $(\mathbf{u}_f, \mathbf{u}_m, \theta)=(\mathbf{u}_f^{k+1}, \mathbf{u}_m^{k+1}, \theta^{k+1})$ as a solution of the following nonlinear elliptic system
\begin{eqnarray}
&&
\left(\frac{\mathbf{u}_f^{k+1}-\mathbf{u}_f^k}{\delta} ,\mathbf{v}_f\right)_f
+\varpi \left(\frac{\mathbf{u}_m^{k+1}-\mathbf{u}_m^k}{\delta} ,\mathbf{v}_m\right)_m\nonumber \\
&&+\left(\mathrm{div}(\mathbf{u}_f^{k+1}\otimes \mathbf{u}_f^{k+1}), \mathbf{v}_f\right)_f
+2\left(\nu(\theta_f^k)\mathbb{D}(\mathbf{u}_f^{k+1}),\mathbb{D}(\mathbf{v}_f)\right)_f\nonumber \\
&&+\left(\nu(\theta_m^k)\mathbb{K}^{-1}\mathbf{u}_m^{k+1},\mathbf{v}_m)\right)_m +\xi \left(\nabla \mathbf{u}_m^{k+1}, \mathbf{v}_m\right)_m \nonumber \\
&& +\sum_{j=1}^{d-1} \int_{\Gamma_{i}}\frac{\alpha\nu(\theta^k_m)}{\sqrt{{\rm trace} (\mathbb{K})}} (\mathbf{u}_f^{k+1}\cdot\boldsymbol{\tau}_i)(\mathbf{v}_f\cdot\boldsymbol{\tau}_i)dS\nonumber \\
&& - \int_{\Gamma_{i}} \frac12|\mathbf{u}_f^{k+1}|^2 (\mathbf{v}_f\cdot \mathbf{n}_{i}) dS \nonumber \\ \nonumber \\
&=& \big(\theta^{k+1}_f \mathbf{k}, \mathbf{v}_f \big)_f
+ \big(\theta^{k+1}_m \mathbf{k}, \mathbf{v}_m \big)_m,
\label{app1}
\end{eqnarray}
\begin{equation}
\left(\frac{\theta^{k+1}-\theta^k}{\delta},\phi\right)
+
\big(\lambda(\theta^k)\nabla \theta^{k+1},\nabla\phi\big)
=-\big(\mathbf{u}^{k+1}\cdot \nabla \theta^{k+1}, \phi\big),
\label{app2a}
\end{equation}
for any $\mathbf{v}_f\in \mathbf{H}_{f,{\rm div}}$, $\mathbf{v}_m\in \mathbf{H}_{m,{\rm div}}$ with $\mathbf{v}_f\cdot \mathbf{n}_i=\mathbf{v}_m\cdot \mathbf{n}_i$ on $\Gamma_i$ and $\phi\in H^1_0(\Omega)$. When $\varpi=0$, we simply take $\mathbf{u}_m^0=\mathbf{0}$. In the above weak formulation for $\theta^{k+1}$, we implicitly use the modified interface condition $$\lambda_f(\theta_f^k)\frac{\partial \theta_f^{k+1}}{\partial \mathbf{n}_{i}}=\lambda_m(\theta_m^k)\frac{\partial \theta_m^{k+1}}{\partial \mathbf{n}_{i}},\quad \text{on}\ \Gamma_i.$$
In the remaining part of this subsection, we will omit the superscript $k+1$ for $\mathbf{u}_f^{k+1}$, $\mathbf{u}_m^{k+1}$, $\theta^{k+1}$ for the sake of simplicity.
The next lemma shows that the solution to problem \eqref{app1}--\eqref{app2a}, if exists, satisfies an discrete energy inequality.
\begin{lemma}[Discrete energy inequality] \label{DEE}
Suppose that $k=0,1,...,N-1$, $\mathbf{u}_f^k\in \widetilde{\mathbf{H}}_{f,{\rm div}}$, $\mathbf{u}_m^k\in \widetilde{\mathbf{H}}_{m,{\rm div}}$, $\theta^k\in H^1_0(\Omega)$. Let
$(\mathbf{u}_f, \mathbf{u}_m, \theta)\in \mathbf{H}_{f,{\rm div}} \times \mathbf{H}_{m,{\rm div}} \times H^1_0(\Omega)$ with $\mathbf{u}_f\cdot \mathbf{n}_i=\mathbf{u}_m\cdot \mathbf{n}_i$ on $\Gamma_i$ be a solution to the discrete problem \eqref{app1}--\eqref{app2a}. Then
the following energy inequality holds
\begin{eqnarray}\label{DisEnLaw}
&&
\mathcal{E}_\sigma (\mathbf{u}_f,\mathbf{u}_m,\theta)
+ \delta \left(\nu(\theta_f^k)\mathbb{D}(\mathbf{u}_f), \mathbb{D}(\mathbf{u}_f)\right)_f
+\frac12\delta \left(\nu(\theta_m^k)\mathbb{K}^{-1}\mathbf{u}_m, \mathbf{u}_m\right)_m
\nonumber \\
&& +\frac12 \delta \sum_{j=1}^{d-1}\int_{\Gamma_{i}} \frac{\alpha\nu(\theta^k_m) }{\sqrt{{\rm trace} (\mathbb{K})}} |\mathbf{u}_f\cdot\boldsymbol{\tau}_j|^2 dS
+\frac12 \delta \sigma \int_\Omega \lambda(\theta^k)|\nabla\theta|^2 dx
\nonumber \\
&&+\frac{1}{2}\big(\mathbf{u}_f-\mathbf{u}_f^k ,\mathbf{u}_f-\mathbf{u}_f^k\big)_f +\frac{\varpi}{2}\big(\mathbf{u}_m-\mathbf{u}_m^k,\mathbf{u}_m-\mathbf{u}_m^k\big)_m\nonumber \\
&& +\frac{1}{2}\sigma\big(\theta-\theta^k,\theta-\theta^k \big)+ \delta \xi \left(\nabla \mathbf{u}_m,\nabla \mathbf{u}_m\right)_m\nonumber \\
&\le& \mathcal{E}_\sigma \big(\mathbf{u}_f^k,\mathbf{u}_m^k,\theta^k\big),
\label{disBEL}
\end{eqnarray}
where $\mathcal{E}_\sigma$ is defined as in \eqref{totenergy}
with a sufficiently large constant $\sigma$ that is independent of $(\mathbf{u}_f,\mathbf{u}_m,\theta)$ and $(\mathbf{u}_f^k,\mathbf{u}_m^k,\theta^k)$.
\end{lemma}
\begin{proof}
Taking $\mathbf{v}_f=\mathbf{u}_f$, $\mathbf{v}_m=\mathbf{u}_m$ in \eqref{app1}, using the elementary identity
\begin{align}
a\cdot(a-b)=\frac{1}{2}\left(|a|^2-|b|^2+|a-b|^2\right), \quad \forall\, a, b\in \mathbb{R} \ \ \text{or}\ \ \mathbb{R}^d,\label{eek1}
\end{align}
we have
\begin{eqnarray}
&&\frac{1}{2\delta}\left(\mathbf{u}_f,\mathbf{u}_f\right)_f+ \frac{1}{2\delta}\big(\mathbf{u}_f-\mathbf{u}_f^k ,\mathbf{u}_f-\mathbf{u}_f^k\big)_f
+
\frac{\varpi}{2\delta}\left(\mathbf{u}_m,\mathbf{u}_m\right)_m\nonumber \\
&&+ \frac{\varpi}{2\delta}\big(\mathbf{u}_m-\mathbf{u}_m^k ,\mathbf{u}_m-\mathbf{u}_m^k\big)_m+2\left(\nu(\theta_f^k)\mathbb{D}(\mathbf{u}_f),\mathbb{D}(\mathbf{u}_f)\right)_f\nonumber \\
&&+\left( \nu(\theta_m^k)\mathbb{K}^{-1}\mathbf{u}_m, \mathbf{u}_m\right)_m
+\sum_{j=1}^{d-1} \int_{\Gamma_{i}}\frac{\alpha \nu(\theta^k_m)}{\sqrt{{\rm trace} (\mathbb{K})}} |\mathbf{u}_f\cdot\boldsymbol{\tau}_j|^2dS \nonumber \\
&& + \xi\left(\nabla \mathbf{u}_m,\nabla \mathbf{u}_m\right)_m\nonumber \\
&=&\frac{1}{2\delta}\big(\mathbf{u}_f^k,\mathbf{u}_f^k\big)_f
+\frac{\varpi}{2\delta}\big(\mathbf{u}_m^k,\mathbf{u}_m^k\big)_m +(\theta_f \mathbf{k}, \mathbf{u}_f )_f
+ (\theta_m \mathbf{k}, \mathbf{u}_m )_m.
\label{eapp1}
\end{eqnarray}
Next, taking the test function $\phi=\theta$ in \eqref{app2a}, using the boundary and interface conditions, after integration by parts, we get
\begin{align}
& \frac{1}{2\delta}\left(\theta,\theta \right)
+\frac{1}{2\delta}\big(\theta-\theta^k,\theta-\theta^k \big) + \int_\Omega \lambda(\theta^k)|\nabla\theta|^2 dx
=\frac{1}{2\delta}\big(\theta^k,\theta^k \big).
\label{eapp2}
\end{align}
Using Lemma \ref{equinorml}, the H\"{o}lder, Young and Poincar\'e inequalities, we estimate the last two terms on the right-hand side of \eqref{eapp1} as follows:
\begin{align}
(\theta_f \mathbf{k}, \mathbf{u}_f )_f + (\theta_m \mathbf{k}, \mathbf{u}_m )_m
& \leq \|\theta_f\|_{L^2(\Omega_f)}\|\mathbf{u}_f \|_{\mathbf{L}^2(\Omega_f)}+\|\theta_m\|_{L^2(\Omega_m)}\|\mathbf{u}_m \|_{\mathbf{L}^2(\Omega_m)}\nonumber \\
& \leq C\|\theta_f\|_{L^2(\Omega_f)}\|\mathbf{u} \|_{\mathbf{Z}}+ \|\theta_m\|_{L^2(\Omega_f)}\|\mathbf{u} \|_{\mathbf{Z}}\nonumber \\
&\leq \epsilon \|\mathbf{u} \|_{\mathbf{Z}}^2
+C\epsilon^{-1}\|\theta_f\|_{L^2(\Omega_f)}^2+\epsilon^{-1} \|\theta_m\|_{L^2(\Omega_m)}^2\nonumber \\
&\leq \epsilon \|\mathbf{u} \|_{\mathbf{Z}}^2+ C\epsilon^{-1}\|\theta\|_{L^2(\Omega)}^2\nonumber \\
&\leq \epsilon \|\mathbf{u} \|_{\mathbf{Z}}^2+ C\epsilon^{-1}\|\nabla \theta\|_{\mathbf{L}^2(\Omega)}^2,\nonumber
\end{align}
for any $\epsilon>0$.
In view of assumptions (A1)--(A3), we can take $\epsilon$ to be sufficiently small,
for instance,
\begin{align}
\epsilon= \frac{1}{4}\min\left\{\underline{\nu},\, \underline{\nu}\overline{\kappa}^{-1},\,
\alpha\underline{\nu}\overline{\kappa}^{-\frac12}
\right\}.\label{epsi1}
\end{align}
Then multiplying \eqref{eapp2} by a sufficiently large constant $\sigma$ that depends on $\epsilon$ and adding the resultant with \eqref{eapp1}, we obtain the discrete energy inequality \eqref{DisEnLaw}.
\end{proof}
To prove the existence of solutions of the discrete problem \eqref{app1}--\eqref{app2a}, we shall adapt a fixed point argument involving the Leray--Schauder principle (cf. \cite{ADG13} for a diffuse interface model for the two-phase flow with unmatched densities and \cite{HWW} for the Cahn--Hilliard--Stokes--Darcy system for the two-phase flow in decomposed domains). For this purpose, it will be convenient to reformulate the problem \eqref{app1}--\eqref{app2a} (again dropping the superscript $k+1$ for simplicity) as follows
\begin{eqnarray}
&&
2\left(\nu(\theta_f^k)\mathbb{D}(\mathbf{u}_f),\mathbb{D}(\mathbf{v}_f)\right)_f
+\left(\nu(\theta_m^k)\mathbb{K}^{-1}\mathbf{u}_m,\mathbf{v}_m)\right)_m
+\xi \left(\nabla \mathbf{u}_m,\nabla \mathbf{v}_m\right)_m\nonumber \\
&& +\sum_{j=1}^{d-1} \int_{\Gamma_{i}}\frac{\alpha \nu(\theta^k_m)}{\sqrt{{\rm trace} (\mathbb{K})}}(\mathbf{u}_f\cdot\boldsymbol{\tau}_i)(\mathbf{v}_f\cdot\boldsymbol{\tau}_i)dS\nonumber \\
&=&-\left(\frac{\mathbf{u}_f-\mathbf{u}_f^k}{\delta} ,\mathbf{v}_f\right)_f
-\varpi \left(\frac{\mathbf{u}_m-\mathbf{u}_m^k}{\delta} ,\mathbf{v}_m\right)_m-\big(\mathrm{div}(\mathbf{u}_f\otimes \mathbf{u}_f), \mathbf{v}_f\big)_f \nonumber \\
&&
+ \int_{\Gamma_{i}} \frac12|\mathbf{u}_f|^2 (\mathbf{v}_f\cdot \mathbf{n}_{i}) dS +(\theta_f \mathbf{k}, \mathbf{v}_f )_f
+ (\theta_m \mathbf{k}, \mathbf{v}_m )_m,
\label{wapp1}
\end{eqnarray}
\begin{equation}
\big(\lambda(\theta^k)\nabla \theta,\nabla\phi\big)
=-\left(\frac{\theta-\theta^k}{\delta},\phi\right)
- ( \mathbf{u} \cdot\nabla \theta, \phi).
\label{wapp2}
\end{equation}
Define the function spaces
\begin{align}
& \mathbf{V} =\big\{(\mathbf{u}_f,\mathbf{u}_m)\ |\ \mathbf{u}_f\in \mathbf{H}_{f, \mathrm{div}},\ \mathbf{u}_m\in \mathbf{H}_{m,\mathrm{div}}, \ \mathbf{u}_f\cdot\mathbf{n}_{i}=\mathbf{u}_m\cdot\mathbf{n}_{i} \ \text{on}\ \Gamma_{i}\big\},
\label{V} \\
& \mathbf{X}=\mathbf{V}\times H_0^1(\Omega),\quad \mathbf{Y}=\mathbf{V}'\times H^{-1}(\Omega).
\label{XY}
\end{align}
First, we introduce the operator $\mathcal{L}_k: \mathbf{V}\to \mathbf{V}'$ given by
\begin{eqnarray}
&&\big \langle \mathcal{L}_k(\mathbf{u}_f, \mathbf{u}_m), (\mathbf{v}_f, \mathbf{v}_m)\big \rangle_{\mathbf{V}', \mathbf{V}}\nonumber \\
&=&2\left(\nu(\theta_f^k)\mathbb{D}(\mathbf{u}_f),\mathbb{D}(\mathbf{v}_f)\right)_f
+\left(\nu(\theta_m^k)\mathbb{K}^{-1}\mathbf{u}_m,\mathbf{v}_m)\right)_m
+\xi \left(\nabla \mathbf{u}_m,\nabla \mathbf{v}_m\right)_m\nonumber \\
&& +\sum_{j=1}^{d-1} \int_{\Gamma_{i}}\frac{\alpha\nu(\theta^k_m)}{\sqrt{{\rm trace} (\mathbb{K})}}(\mathbf{u}_f\cdot\boldsymbol{\tau}_i)(\mathbf{v}_f\cdot\boldsymbol{\tau}_i)dS,
\label{Lk}
\end{eqnarray}
for any $(\mathbf{u}_f, \mathbf{u}_m)$, $(\mathbf{v}_f, \mathbf{v}_m) \in \mathbf{V}$.
Using the assumptions (A1), (A3) and Lemma \ref{equinorml}, it is straightforward to verify that $\mathcal{L}_k$ is a strictly monotone, bounded
and coercive operator on $\mathbf{V}$. Hence, it easily follows from the Lax--Milgram theorem that
\begin{lemma}\label{LLk}
Let the assumptions (A1) and (A3) be satisfied. For any given function $\theta^k\in H^1_0(\Omega)$, the operator $\mathcal{L}_k: \mathbf{V}\to \mathbf{V}'$ is invertible and its inverse $\mathcal{L}_k^{-1}: \mathbf{V}' \to \mathbf{V}$ is continuous.
\end{lemma}
Next, we consider the operator induced by the left-hand side of \eqref{wapp2}. Define the operator
$\mathrm{div}_D: \mathbf{L}^2(\Omega)\to H^{-1}(\Omega)$ by
$ \big \langle \mathrm{div}_D \mathbf{v}, \phi\big\rangle_{H^{-1}(\Omega), H_0^1(\Omega)}=-(\mathbf{v}, \nabla \phi)$ for any $\phi \in H_0^1(\Omega)$. Then for $\lambda\in L^\infty(\Omega)$ such that $\lambda(x)\geq \underline{\lambda}>0$ almost everywhere in $\Omega$, we introduce the operator
$\mathrm{div}_D(\lambda(x) \nabla \cdot): H_0^1(\Omega)\to H^{-1}(\Omega)$ given by
$$ \big \langle \mathrm{div}_D( \lambda(x) \nabla \theta), \phi\big\rangle_{H^{-1}(\Omega), H_0^1(\Omega)}=-(\lambda(x) \nabla \theta, \nabla \phi), \quad \forall\, \phi \in H_0^1(\Omega).$$
Again, one can check that the operator $\mathrm{div}_D(\lambda(x) \nabla \cdot)$ is an isomorphism by an easy application of the Lax--Milgram theorem. Then we have
\begin{lemma}\label{LDk}
Let the assumption (A2) be satisfied.
For any given $\theta^k\in H^1_0(\Omega)$, the operator
\begin{equation}
\mathcal{M}_k:=-\mathrm{div}_D\big(\lambda(\theta^k) \nabla \cdot\big): H_0^1(\Omega)\to H^{-1}(\Omega)\label{Dk}
\end{equation}
is invertible and its inverse $\mathcal{M}_k^{-1}: H^{-1}(\Omega) \to H_0^1(\Omega)$ is continuous.
\end{lemma}
Concerning the terms on the right-hand side of problem \eqref{wapp1}--\eqref{wapp2}, we consider the following operator $\mathcal{J}_k: \mathbf{X}\to \mathbf{V}'$:
%
\begin{align}
& \big\langle \mathcal{J}_k(\mathbf{w}), (\mathbf{v}_f, \mathbf{v}_m) \big\rangle_{\mathbf{V}', \mathbf{V}}\nonumber \\
&\quad = -\left(\frac{\mathbf{u}_f-\mathbf{u}_f^k}{\delta} ,\mathbf{v}_f\right)_f
-\varpi \left(\frac{\mathbf{u}_m-\mathbf{u}_m^k}{\delta} ,\mathbf{v}_m\right)_m\nonumber \\
&\qquad -\left(\mathrm{div}(\mathbf{u}_f\otimes \mathbf{u}_f), \mathbf{v}_f\right)_f
+ \int_{\Gamma_{i}} \frac12|\mathbf{u}_f|^2 (\mathbf{v}_f\cdot \mathbf{n}_{i}) dS \nonumber \\
&\qquad + (\theta_f \mathbf{k}, \mathbf{v}_f )_f
+ (\theta_m \mathbf{k}, \mathbf{v}_m )_m, \qquad \forall \, (\mathbf{v}_f, \mathbf{v}_m) \in \mathbf{V}, \label{Jk}
\end{align}
and the operator $\mathcal{K}_k: \mathbf{X}\to H^{-1}(\Omega)$ given by
\begin{align}
& \big \langle \mathcal{K}_k(\mathbf{w}),\phi\big \rangle_{H^{-1}(\Omega),H^1_0(\Omega)} = -\left(\frac{\theta-\theta^k}{\delta}, \phi\right) - (\mathbf{u}\cdot \nabla \theta , \phi),\quad \forall\, \phi\in H^1_0(\Omega),
\label{KK}
\end{align}
where we denote $\mathbf{w}=(\mathbf{u}_f, \mathbf{u}_m, \theta)$.
Let $\mathbf{u}_f^k\in \widetilde{\mathbf{H}}_{f,\mathrm{div}}$, $\mathbf{u}_m^k\in \widetilde{\mathbf{H}}_{m,\mathrm{div}}$ and $\theta^k\in H^1_0(\Omega)$ be given.
Using the above formulations \eqref{Lk}--\eqref{KK}, we now introduce the nonlinear operators $\mathcal{T}_k$, $\mathcal{G}_k: \mathbf{X}\to \mathbf{Y}$ such that
\begin{equation}
\mathcal{T}_k (\mathbf{w}) =
\left(
\begin{array}{c}
\mathcal{L}_k(\mathbf{u}_f,\mathbf{u}_m)\\
\mathcal{M}_k (\theta)\\
\end{array}
\right), \label{Tk}
\end{equation}
and
\begin{equation}
\mathcal{G}_k (\mathbf{w}) =
\left(
\begin{array}{c}
\mathcal{J}_k(\mathbf{w}) \\
\mathcal{K}_k(\mathbf{w})
\end{array}
\right), \label{Gk} %
\end{equation}
where $\mathbf{w}=(\mathbf{u}_f, \mathbf{u}_m, \theta)$.
As a consequence, denoting the solution to problem \eqref{app1}--\eqref{app2a} by $\mathbf{w}=(\mathbf{u}_f, \mathbf{u}_m, \theta)$, we can write \eqref{app1}--\eqref{app2a} into the following abstract form:
\begin{align}
\mathcal{T}_k(\mathbf{w})=\mathcal{G}_k(\mathbf{w}). \label{Abs}
\end{align}
\noindent
Indeed, from the above discussions, we have
\begin{proposition}\label{Abequi}
The triple $(\mathbf{u}_f, \mathbf{u}_m, \theta)\in \mathbf{X}$ is a solution of problem \eqref{app1}--\eqref{app2a} if and only if
$\mathbf{w}=(\mathbf{u}_f, \mathbf{u}_m, \theta) \in \mathbf{X}$ satisfies the equation
$\mathcal{T}_k(\mathbf{w})=\mathcal{G}_k(\mathbf{w})$.
\end{proposition}
We now proceed to show that the abstract equation \eqref{Abs} admits at least one solution $\mathbf{w} \in \mathbf{X}$. Recalling the definition of $\mathcal{T}_k$ and Lemmas \ref{LLk}--\ref{LDk}, we can conclude that
\begin{lemma}\label{LTG}
Let the assumptions (A1)--(A3) be satisfied. For any given function $\theta^k\in H^1_0(\Omega)$,
$\mathcal{T}_k: \mathbf{X}\to\mathbf{Y}$ is an invertible mapping and its inverse $\mathcal{T}^{-1}_k: \mathbf{Y}\to \mathbf{X}$ is continuous.
\end{lemma}
Next, concerning the operator $\mathcal{G}_k$, we introduce the space
$$
\widetilde{\mathbf{Y}}= \left(\big(\mathbf{H}^\frac{3}{4}(\Omega_f)\big)'
\times \mathbf{L}^2(\Omega_m)\right)\times \big(H^\frac12(\Omega)\big)',
$$
then we have
\begin{lemma}\label{LGk}
The operator $\mathcal{G}_k: \mathbf{X}\to\widetilde{\mathbf{Y}}$ is continuous and it maps bounded sets into bounded sets.
Moreover, the mapping $\mathcal{G}_k: \mathbf{X}\to\mathbf{Y}$ is compact.
\end{lemma}
\begin{proof}
For all $\mathbf{w}=(\mathbf{u}_f, \mathbf{u}_m, \theta)\in \mathbf{X}$, using the Sobolev embedding theorems $(d=2, 3)$, we can show that
$
\mathcal{G}_k (\mathbf{w})\in \widetilde{\mathbf{Y}}$.
Indeed, the estimates for the linear terms are obvious, thus we only need to estimate those terms that are nonlinear:
\begin{align}
\|\mathrm{div}(\mathbf{u}_f\otimes \mathbf{u}_f)\|
_{\big(\mathbf{H}^\frac{3}{4}(\Omega_f)\big)'}
&\leq C\|\mathrm{div}(\mathbf{u}_f\otimes \mathbf{u}_f)\|_{\mathbf{L}^\frac43(\Omega_f)}\nonumber \\
&\leq C\|\mathrm{div}(\mathbf{u}_f\otimes \mathbf{u}_f)\|_{\mathbf{L}^\frac32(\Omega_f)}\nonumber \\
&\leq C \|\nabla \mathbf{u}_f\|_{\mathbf{L}^2(\Omega_f)}\|\mathbf{u}_f\|_{\mathbf{L}^6(\Omega_f)}
\nonumber \\
&\leq C\|\mathbf{u}_f\|_{\mathbf{H}^1(\Omega_f)}^2,\nonumber
\end{align}
\begin{align}
\sup_{\|\mathbf{v}_f\|_{\mathbf{H}^{\frac34}(\Omega_f)}\leq 1}\left|\int_{\Gamma_{i}} |\mathbf{u}_f|^2 (\mathbf{v}_f\cdot \mathbf{n}_{i}) dS\right|
&\leq \big\||\mathbf{u}_f|^2\big\|_{\big(\mathbf{H}^{\frac14}(\Gamma_i)\big)'}
\|\mathbf{v}_f\|_{\mathbf{H}^{\frac14}(\Gamma_i)}\nonumber \\
&\leq C \big\||\mathbf{u}_f|^2\big\|_{\mathbf{L}^\frac{8}{5}(\Gamma_i)}
\|\mathbf{v}_f\|_{\mathbf{H}^{\frac34}(\Omega_f)}\nonumber \\
&\leq C\|\mathbf{u}_f\|^2_{\mathbf{L}^\frac{16}{5}(\Gamma_i)}
\leq C\|\mathbf{u}_f\|^2_{\mathbf{L}^4(\Gamma_i)}\nonumber \\
&\leq C\|\mathbf{u}_f\|^2_{\mathbf{H}^\frac{1}{2}(\Gamma_i)}
\leq C\|\mathbf{u}_f\|_{\mathbf{H}^1(\Omega_f)}^2,\nonumber
\end{align}
\begin{align}
\|\mathbf{u}\cdot\nabla \theta\|_{\big(H^\frac12(\Omega)\big)'}&\leq
\|\mathbf{u}\cdot\nabla \theta\|_{L^\frac32(\Omega)}\nonumber \\
&\leq \|\mathbf{u}\|_{\mathbf{L}^6(\Omega)}\|\nabla \theta\|_{\mathbf{L}^2(\Omega)}\nonumber \\
&\leq C\|(\mathbf{u}_f,\mathbf{u}_m)\|_{\mathbf{V}}\|\theta\|_{H^1_0(\Omega)}.\nonumber
\end{align}
The second conclusion on compactness of $\mathcal{G}_k$ easily follows from the fact $\widetilde{\mathbf{Y}}\hookrightarrow\hookrightarrow \mathbf{Y}$.
\end{proof}
Since the operator $\mathcal{T}_k: \mathbf{X}\to\mathbf{Y}$ is invertible, for any $\mathbf{w}\in \mathbf{X}$ we introduce $\mathbf{q}= \mathcal{T}_k (\mathbf{w})$ and then
the abstract equation \eqref{Abs} can be rewritten into an equivalent form such that $\mathbf{q}=(\mathcal{G}_k\circ\mathcal{T}_k^{-1})(\mathbf{q})$.
Thanks to Lemmas \ref{LTG} and \ref{LGk},
we see that the mapping $$\mathcal{N}_k\overset{\text{def}}{=}\mathcal{G}_k\circ\mathcal{T}_k^{-1}:\, \mathbf{Y}\to \mathbf{Y}$$ is indeed a compact operator, because
$\mathcal{T}_k^{-1}$ is continuous and $\mathcal{G}_k$ is compact.
Now the original problem can be reduced to find a fixed point $\mathbf{q}$ of the operator $\mathcal{N}_k$ in $\mathbf{Y}$, that is,
\begin{equation}
\mathbf{q}=\mathcal{N}_k(\mathbf{q}).\label{IaNK}
\end{equation}
Existence of such a fixed point follows from an application of the Leray--Shauder principle. More precisely, we have
\begin{lemma}\label{degree}
Assume that assumptions (A1)--(A3) are satisfied. For any $\mathbf{u}_f^k\in \widetilde{\mathbf{H}}_{f,\mathrm{div}}$, $\mathbf{u}_m^k\in \widetilde{\mathbf{H}}_{m,\mathrm{div}}$ and $\theta^k\in H^1_0(\Omega)$, the abstract equation \eqref{IaNK} admits a solution $\mathbf{q}\in \widetilde{\mathbf{Y}}\subset \mathbf{Y}$.
\end{lemma}
\begin{proof}
According to the abstract result \cite[Theorem 6.A]{Zei92}, it remains to show that there exists a constant $R>0$ such that if $\mathbf{q}\in \mathbf{Y}$ and $0\leq s\leq 1$ satisfying $\mathbf{q}=s\mathcal{N}_k(\mathbf{q})$, then $\|\mathbf{q}\|_{\mathbf{Y}}\leq R$.
We consider $\mathbf{q}\in \mathbf{Y}$ and $0\leq s\leq 1$ satisfying $\mathbf{q}=s\mathcal{N}_k(\mathbf{q})$. Denote $\mathbf{w} = \mathcal{T}_k ^{-1}(\mathbf{q})$. Then we have
$\mathcal{T}_k(\mathbf{w})=s\mathcal{G}_k(\mathbf{w})$ that is equivalent to the weak formulation
\begin{eqnarray}
&&
2\left(\nu(\theta_f^k)\mathbb{D}(\mathbf{u}_f),\mathbb{D}(\mathbf{v}_f)\right)_f
+\left(\nu(\theta_m^k)\mathbb{K}^{-1}\mathbf{u}_m,\mathbf{v}_m)\right)_m
\nonumber \\
&& +\xi (\nabla \mathbf{u}_m,\nabla \mathbf{v}_m)_m +\sum_{j=1}^{d-1} \int_{\Gamma_{i}}\frac{\alpha \nu(\theta^k_m)}{\sqrt{{\rm trace} (\mathbb{K})}}(\mathbf{u}_f\cdot\boldsymbol{\tau}_i)(\mathbf{v}_f\cdot\boldsymbol{\tau}_i)dS\nonumber \\
&=&-s\left(\frac{\mathbf{u}_f-\mathbf{u}_f^k}{\delta} ,\mathbf{v}_f\right)_f
-s\varpi \left(\frac{\mathbf{u}_m-\mathbf{u}_m^k}{\delta} ,\mathbf{v}_m\right)_m\nonumber \\
&&-s\left(\mathrm{div}(\mathbf{u}_f\otimes \mathbf{u}_f), \mathbf{v}_f\right)_f
+ s\int_{\Gamma_{i}} \frac12|\mathbf{u}_f|^2 (\mathbf{v}_f\cdot \mathbf{n}_{i}) dS \nonumber \\
&& +s(\theta_f \mathbf{k}, \mathbf{v}_f )_f
+ s(\theta_m \mathbf{k}, \mathbf{v}_m )_m,
\label{wapp1c}
\end{eqnarray}
and
\begin{equation}
\big(\lambda(\theta^k)\nabla \theta,\nabla\phi\big)
=-s\left(\frac{\theta-\theta^k}{\delta},\phi\right) - s(\mathbf{u}\cdot \nabla \theta ,\phi).
\label{wapp2c}
\end{equation}
Similar to the derivation of the discrete energy inequality in Lemma \ref{DEE}, we can derive
\begin{eqnarray}
&&\frac{s}{2\delta}\left(\mathbf{u}_f,\mathbf{u}_f\right)_f+ \frac{s}{2\delta}\big(\mathbf{u}_f-\mathbf{u}_f^k ,\mathbf{u}_f-\mathbf{u}_f^k\big)_f
+
\frac{\varpi s}{2\delta}\left(\mathbf{u}_m,\mathbf{u}_m\right)_m\nonumber \\
&&+ \frac{\varpi s}{2\delta}\big(\mathbf{u}_m-\mathbf{u}_m^k ,\mathbf{u}_m-\mathbf{u}_m^k\big)_m+2\left(\nu(\theta_f^k)\mathbb{D}(\mathbf{u}_f),\mathbb{D}(\mathbf{u}_f)\right)_f\nonumber \\
&&+\big( \nu(\theta_m^k)\mathbb{K}^{-1}\mathbf{u}_m, \mathbf{u}_m\big)_m +\xi (\nabla \mathbf{u}_m,\nabla \mathbf{u}_m)_m
\nonumber \\
&& +\sum_{j=1}^{d-1} \int_{\Gamma_{i}}\frac{\alpha \nu(\theta^k_m)}{\sqrt{{\rm trace} (\mathbb{K})}} |\mathbf{u}_f\cdot\boldsymbol{\tau}_j|^2dS \nonumber \\
&=&\frac{s}{2\delta}\big(\mathbf{u}_f^k,\mathbf{u}_f^k\big)_f
+\frac{\varpi s}{2\delta}\big(\mathbf{u}_m^k,\mathbf{u}_m^k\big)_m + s(\theta_f \mathbf{k}, \mathbf{u}_f )_f + s(\theta_m \mathbf{k}, \mathbf{u}_m )_m,
\label{eapp1b}
\end{eqnarray}
and
\begin{equation}
\frac{s}{2\delta}\left(\theta,\theta \right)
+\frac{s}{2\delta}\left(\theta-\theta^k,\theta-\theta^k \right) + \int_\Omega \lambda(\theta^k)|\nabla\theta|^2 dx
=\frac{s}{2\delta}\left(\theta^k,\theta^k \right).
\label{eapp2b}
\end{equation}
Similar to the previous argument for \eqref{DisEnLaw}, using the fact $s\in [0,1]$, we can derive the following discrete energy inequality with respect to $s$:
\begin{eqnarray}\label{DisEnLawb}
&&
s\mathcal{E}_\sigma (\mathbf{u}_f,\mathbf{u}_m,\theta)
+\delta \underline{\nu}\left(\mathbb{D}(\mathbf{u}_f), \mathbb{D}(\mathbf{u}_f)\right)_f
+\frac12\delta\underline{\nu}\bar{\kappa}^{-1} \left(\mathbf{u}_m, \mathbf{u}_m\right)_m
\nonumber \\
&& +\frac12 \delta \alpha \underline{\nu}\bar{\kappa}^{-\frac12}\sum_{j=1}^{d-1}\int_{\Gamma_{i}} |\mathbf{u}_f\cdot\boldsymbol{\tau}_j|^2 dS
+\delta\xi (\nabla \mathbf{u}_m,\nabla \mathbf{u}_m)_m
+\frac12 \delta \sigma \underline{\lambda} \int_\Omega |\nabla\theta|^2 dx
\nonumber \\
&\le& s\mathcal{E}_\sigma(\mathbf{u}_f^k,\mathbf{u}_m^k,\theta^k)\leq \mathcal{E}_\sigma(\mathbf{u}_f^k,\mathbf{u}_m^k,\theta^k),
\end{eqnarray}
with a sufficiently large constant $\sigma$ that is independent of $s$.
As a consequence, we have (recall also Lemma \ref{equinorml})
$$
\|\mathbf{w}\|_{\mathbf{X}}\leq C_k,\quad \forall\, s\in [0,1],
$$
where the constant $C_k$ may depend on $\delta$, $\xi$ and $\sigma$, but is independent of $s$.
In order to derive an estimate for $\mathbf{q}=\mathcal{T}_k(\mathbf{w})\in \mathbf{Y}$, we recall the relation $\mathbf{q} = s\mathcal{G}_k(\mathbf{w})$ and Lemma \ref{LGk} that $\mathcal{G}_k:\,\mathbf{X}\to \widetilde{\mathbf{Y}}$ maps bounded sets in $\mathbf{X}$ into bounded (compact) sets in $\mathbf{Y}$. Thus, we get $\mathbf{q}\in \widetilde{\mathbf{Y}}$ and
$$\|\mathbf{q}\|_{\mathbf{Y}} = \|s\mathcal{G}_k(\mathbf{w})\|_{\mathbf{Y}}\leq C(\|\mathbf{w}\|_{\mathbf{X}}+1)\leq R,$$
where the constant $R$ depends on $C_k$,
but is independent of $s$.
The proof of Lemma \ref{degree} is complete.
\end{proof}
As a consequence, we can conclude the existence of a weak solution to the time discrete problem \eqref{app1}--\eqref{app2a} from Lemmas \ref{DEE}, \ref{degree} and Proposition \ref{Abequi}, that is
\begin{lemma}\label{DEEl}
Assume that assumptions (A1)--(A3) are satisfied.
For every $\mathbf{u}_f^k\in \widetilde{\mathbf{H}}_{f,\mathrm{div}}$, $\mathbf{u}_m^k\in \widetilde{\mathbf{H}}_{m,\mathrm{div}}$ and $\theta^k\in H^1_0(\Omega)$, there exists a weak solution $(\mathbf{u}_f, \mathbf{u}_m, \theta)$ to the discrete problem \eqref{app1}--\eqref{app2a} such that
$$
(\mathbf{u}_f,\mathbf{u}_m) \in \mathbf{V}, \quad \theta \in H_0^1(\Omega).
$$
Moreover, the solution $(\mathbf{u}_f, \mathbf{u}_m, \theta)$ satisfies the discrete energy inequality \eqref{DisEnLaw}.
\end{lemma}
\subsection{Construction of approximate solutions}
Once we have proved the existence of weak solutions to the time-discrete problem \eqref{app1}--\eqref{app2a}, we are able to construct approximate solutions to the regularized time-continuous system \eqref{weak1r}--\eqref{weak3r}. Recall that $\delta= \frac{T}{N}$, where $T>0$ and $N$ is an positive integer. We set $$t_k=k \delta, \quad k=0, 1,\cdots, N.$$
Let $(\mathbf{u}_f^{k+1}, \mathbf{u}_m^{k+1}, \theta^{k+1})$ ($k=0, 1,\cdots, N-1$) be chosen successively as a solution of the discrete problem \eqref{app1}--\eqref{app2a} with $(\mathbf{u}_f^{k}, \mathbf{u}_m^k, \theta^{k})$ being the ``initial value" (see Lemma \ref{DEEl}). In particular, we set $(\mathbf{u}_f^{0}, \mathbf{u}_m^{0}, \theta^{0})=(\mathbf{u}_{0f}, \mathbf{u}_{0m}, \theta_0)$ with the choice $\mathbf{u}_m^0=\mathbf{0}$ when $\varpi=0$.
Then for $k=0,1, \cdots, N-1$, we define the approximate solutions as follows
\begin{align*}
&\theta^\delta :=\frac{t_{k+1}-t}{\delta}\theta^{k}+\frac{t-t_k}{\delta}\theta^{k+1}, &\text{for}\ t \in [t_k, t_{k+1}], \\
&\mathbf{u}^\delta_f:= \frac{t_{k+1}-t}{\delta}\mathbf{u}_f^{k}+\frac{t-t_k}{\delta}\mathbf{u}_f^{k+1}, &\text{for}\ t \in [t_k, t_{k+1}],\\
&\mathbf{u}^\delta_m:= \frac{t_{k+1}-t}{\delta}\mathbf{u}_m^{k}+\frac{t-t_k}{\delta}\mathbf{u}_m^{k+1}, &\text{for}\ t \in [t_k, t_{k+1}],\\
& \widehat{\mathbf{u}}_f^\delta: =\mathbf{u}_f^{k+1},&\text{for}\ t \in (t_k, t_{k+1}],\\
& \widehat{\mathbf{u}}_m^\delta: =\mathbf{u}_m^{k+1},&\text{for}\ t \in (t_k, t_{k+1}],\\
& \widehat{\mathbf{u}}^\delta|_{\Omega_f}= \widehat{\mathbf{u}}_f^\delta, \quad \widehat{\mathbf{u}}^\delta|_{\Omega_m}=\widehat{\mathbf{u}}^\delta_m,
& \text{for}\ t \in (t_k, t_{k+1}],\\
&\widehat{\theta}^\delta := \theta^{k+1}, & \text{for}\ t \in (t_k, t_{k+1}], \\
&\widetilde{\theta}^\delta := \theta^{k}, & \text{for}\ t \in [t_k, t_{k+1}).
\end{align*}
\begin{remark}
It follows from the above definitions that $\theta^\delta$, $\mathbf{u}^\delta_f$, $\mathbf{u}^\delta_m$ are continuous piecewise linear functions in time, while
$\widehat{\mathbf{u}}_f^\delta$, $\widehat{\mathbf{u}}_m^\delta$, $\widehat{\theta}^\delta$ are piecewise constant (in time) functions being right continuous at the nodes
$\{t_{k+1}\}$ and $\widetilde{\theta}^\delta$ is left continuous at the nodes $\{t_k\}$.
\end{remark}
Using the above definition of approximate solutions, we can derive from the discrete problem \eqref{app1}--\eqref{app2a} that the following identities hold:
\begin{eqnarray}
&&\int_0^T\big(\partial_t\mathbf{u}_f^\delta,\mathbf{v}_f\big)_f dt
+\varpi\int_0^T\big(\partial_t \mathbf{u}_m^\delta,\mathbf{v}_m\big)_m dt\nonumber \\
&&
+\int_0^T \big(\mathrm{div}(\widehat{\mathbf{u}}_f^\delta\otimes \widehat{\mathbf{u}}_f^\delta), \mathbf{v}_f\big)_fdt
+2\int_0^T\left(\nu(\widetilde{\theta}_f^\delta)\mathbb{D}(\widehat{\mathbf{u}}_f^\delta),
\mathbb{D}(\mathbf{v}_f)\right)_fdt
\nonumber \\
&& + \int_0^T \left(\nu(\widetilde{\theta}_m^\delta)\mathbb{K}^{-1}\widehat{\mathbf{u}}_m^\delta,\mathbf{v}_m)\right)_m dt + \xi \int_0^T \left(\nabla \widehat{\mathbf{u}}_m^\delta,\nabla \mathbf{v}_m\right)_m dt
\nonumber \\
&&+\sum_{j=1}^{d-1}\int_0^T\!\int_{\Gamma_{i}}
\frac{\alpha \nu(\widetilde{\theta}_m^\delta)}{\sqrt{{\rm trace} (\mathbb{K})}}
(\widehat{\mathbf{u}}_f^\delta \cdot\boldsymbol{\tau}_j)(\mathbf{v}_f\cdot\boldsymbol{\tau}_j)dSdt\nonumber \\
&&- \int_0^T\! \int_{\Gamma_{i}} \frac12|\widehat{\mathbf{u}}_f^\delta|^2 (\mathbf{v}_f\cdot \mathbf{n}_{i}) dSdt \nonumber \\
&=& \int_0^T \big(\widehat{\theta}_f^\delta \mathbf{k}, \mathbf{v}_f \big)_f dt +
\int_0^T \big(\widehat{\theta}_m^\delta \mathbf{k}, \mathbf{v}_m \big)_m dt,\label{weak1ra}
\end{eqnarray}
for any $\mathbf{v}_f\in C^1_0((0,T); \mathbf{H}_{f,\mathrm{div}})$, $\mathbf{v}_m\in C^1_0((0,T); \mathbf{H}_{m,\mathrm{div}})$ with $\mathbf{v}_f\cdot \mathbf{n}_i=\mathbf{v}_m\cdot \mathbf{n}_i$ on $\Gamma_i$, and
\begin{align}
&\int_0^T\big(\partial_t \theta^\delta,\phi\big)dt+\int_0^T\big(\lambda(\widetilde{\theta}^\delta) \nabla \widehat{\theta}^\delta,\nabla\phi\big)dt =\int_0^T\big(\widehat{\mathbf{u}}^\delta \widehat{\theta}^\delta ,\nabla \phi\big)dt
\label{weak3ra}
\end{align}
for any $\phi\in C_0^1((0,T); H^1(\Omega))$.
Besides, in analogy to the estimates for \eqref{eapp1} and \eqref{eapp2}, we can obtain the energy inequalities for $t\in [0,T]$:
\begin{eqnarray}
&&\frac{1}{2}\|\widehat{\mathbf{u}}^\delta_f(t)\|_{\mathbf{L}^2(\Omega_f)}^2
+
\frac{\varpi}{2}\|\widehat{\mathbf{u}}^\delta_m(t)\|_{\mathbf{L}^2(\Omega_m)}^2
+2\int_0^t\left(\nu(\widetilde{\theta}^\delta_f)\mathbb{D}(\widehat{\mathbf{u}}^\delta_f),\mathbb{D}(\widehat{\mathbf{u}}^\delta_f)\right)_f d\tau\nonumber \\
&&+\int_0^t \left( \nu(\widetilde{\theta}^\delta_m)\mathbb{K}^{-1}\widehat{\mathbf{u}}^\delta_m, \widehat{\mathbf{u}}^\delta_m\right)_m d\tau
+\sum_{j=1}^{d-1}\int_0^t\! \int_{\Gamma_{i}}\frac{\alpha \nu(\widetilde{\theta}^\delta_m)}{\sqrt{{\rm trace} (\mathbb{K})}} |\widehat{\mathbf{u}}^\delta_f\cdot\boldsymbol{\tau}_j|^2 dS d\tau \nonumber \\
&& + \xi\int_0^t\left(\nabla \widehat{\mathbf{u}}^\delta_m,\nabla \widehat{\mathbf{u}}^\delta_m\right)_m d\tau\nonumber \\
&\leq &\frac{1}{2}\|\mathbf{u}_{0f}\|_{\mathbf{L}^2(\Omega_f)}^2
+\frac{\varpi}{2}\|\mathbf{u}_{0m}\|_{\mathbf{L}^2(\Omega_m)}^2 + \int_0^t \big(\widehat{\theta}^\delta_f \mathbf{k}, \widehat{\mathbf{u}}^\delta_f \big)_f + \big(\widehat{\theta}^\delta_m \mathbf{k}, \widehat{\mathbf{u}}^\delta_m \big)_m d\tau,
\label{eapp1c}
\end{eqnarray}
and
\begin{align}
& \frac{1}{2}\|\widehat{\theta}^\delta(t)\|_{L^2(\Omega)}^2
+\int_0^t \int_\Omega \lambda(\widetilde{\theta}^\delta)|\nabla\widehat{\theta}^\delta|^2 dx d\tau
\leq \frac{1}{2}\|\theta_0\|_{L^2(\Omega)}^2.
\label{eapp2c}
\end{align}
Let $\mathcal{E}_\sigma^\delta(t)$ be the piecewise linear interpolation of the discrete energy $\mathcal{E}_\sigma(\mathbf{u}_f^{k},\mathbf{u}_m^{k}, \theta^k)$ (with the same choice for the constant $\sigma$ as in Lemma \ref{DEE}) such that
\begin{equation} \mathcal{E}_\sigma^\delta(t)=\frac{t_{k+1}-t}{\delta}\mathcal{E}_\sigma(\mathbf{u}_f^{k},\mathbf{u}_m^{k}, \theta^k)+\frac{t-t_k}{\delta}\mathcal{E}_\sigma(\mathbf{u}_f^{k+1},\mathbf{u}_m^{k+1}, \theta^{k+1}), \quad \text{for}\ t \in [t_k, t_{k+1}], \nonumber
\end{equation}
and $\mathcal{D}_\sigma^\delta(t)$ be the approximate energy dissipation
\begin{eqnarray}
\mathcal{D}_\sigma^\delta(t)&=& 2 \left(\nu(\theta_f^k)\mathbb{D}(\mathbf{u}_f^{k+1}), \mathbb{D}(\mathbf{u}_f^{k+1})\right)_f+ \left( \nu(\theta_m^k)\mathbb{K}^{-1}\mathbf{u}_m^{k+1}, \mathbf{u}_m^{k+1}\right)_m\nonumber \\
&& +\xi(\nabla \mathbf{u}_m^{k+1}, \nabla \mathbf{u}_m^{k+1})_m + \sigma \int_\Omega \lambda(\theta^k)|\nabla\theta^{k+1}|^2 dx\nonumber \\
&& +\sum_{j=1}^{d-1}\int_{\Gamma_{i}} \frac{\alpha \nu(\theta^k_m)}{\sqrt{{\rm trace}(\mathbb{K})}}|\mathbf{u}_f^{k+1}\cdot\boldsymbol{\tau}_j|^2 dS, \quad \text{for}\ t \in (t_k, t_{k+1}).
\nonumber
\end{eqnarray}
We see from the discrete energy estimate \eqref{DisEnLaw} that for $k=0,1, \cdots, N-1$, it holds
\begin{align}
\frac{d}{dt}\mathcal{E}_\sigma^\delta(t)
&=\frac{1}{\delta}\left[\mathcal{E}_\sigma(\mathbf{u}_f^{k+1}, \mathbf{u}_m^{k+1}, \theta^{k+1})-\mathcal{E}_\sigma(\mathbf{u}_f^{k}, \mathbf{u}_m^{k}, \theta^k)\right]\nonumber \\
&\leq - \frac12\mathcal{D}_\sigma^\delta(t),\quad \text{for}\ t \in (t_k, t_{k+1}).\label{bela}
\end{align}
In particular, we have for all $t\in [0,T]$,
\begin{align}
\mathcal{E}_\sigma(\widehat{\mathbf{u}}_f^{\delta}(t), \widehat{\mathbf{u}}_m^{\delta}(t), \widehat{\theta}^{\delta}(t))
+ \frac12\int_0^t \mathcal{D}_\sigma^\delta(\tau) d\tau\leq \mathcal{E}_\sigma(\mathbf{u}_{0f},\mathbf{u}_{0m}, \theta_0).\label{ee1}
\end{align}
\subsection{Proof of Theorem \ref{thmEx}}
We now proceed to finish the proof of Theorem \ref{thmEx}. First, we prove the conclusion for the case $\varpi>0$, and then we point out necessary modifications for the case $\varpi=0$.
\subsubsection{Case $\varpi>0$}
\textbf{Step 1. Passage to the limit $\delta\to 0$}.
First, we pass to the limit as $\delta\to 0$ while keeping the regularizing parameter $\xi>0$ fixed. From the energy inequality \eqref{ee1}, we can derive estimates on the approximate solutions that are uniform in $\delta$:
\begin{align}
&\|\widehat{\mathbf{u}}^\delta_f\|_{L^\infty(0,T;\mathbf{L}^2(\Omega_f))}
+\varpi^\frac12\|\widehat{\mathbf{u}}^\delta_m\|_{L^\infty(0,T;\mathbf{L}^2(\Omega_m))}
+\|\widehat{\theta}^\delta\|_{L^\infty(0,T;L^2(\Omega))} \leq C,\label{con1} \\
&\|\mathbb{D}(\widehat{\mathbf{u}}^\delta_f)\|_{L^2(0,T;\mathbf{L}^2(\Omega_f))}
+\sum_{j=1}^{d-1}\|\widehat{\mathbf{u}}_f^\delta\cdot\boldsymbol{\tau}_j\|_{L^2(0,T; L^2(\Gamma_{i}))}\leq C, \\
&\|\widehat{\mathbf{u}}_m^\delta\|_{L^2(0,T;\mathbf{L}^2(\Omega_m))}
+\xi^\frac12\|\nabla \widehat{\mathbf{u}}_m^\delta\|_{L^2(0,T;\mathbf{L}^2(\Omega_m))}
\leq C, \label{conm1}\\
& \|\nabla \widehat{\theta}^\delta\|_{L^2(0,T;\mathbf{L}^2(\Omega))} \leq C,\label{mues1}
\end{align}
where the constant $C$ depends on $\mathcal{E}_\sigma(\mathbf{u}_{0f},\mathbf{u}_{0m},\theta_0)$
and $\Omega$, but is independent of the parameters $\delta$ and $\xi$.
From the uniform estimates \eqref{con1}--\eqref{mues1} and Lemma \ref{equinorml}, we deduce that there exists a convergent subsequence $\{(\widehat{\mathbf{u}}_f^\delta, \widehat{\mathbf{u}}_m^\delta, \widehat{\theta}^\delta)\}$ (still denoted by the same symbols for simplicity) as $\delta \to 0$ (or equivalently $N\to +\infty$) such that
\begin{equation}
\begin{cases}
\widehat{\mathbf{u}}_f^\delta \rightarrow \mathbf{u}_f^\xi & \text{ weakly star in } L^\infty(0,T; \mathbf{L}^2(\Omega_f)),\\
&\text{ weakly in } L^2(0,T; \mathbf{H}^1(\Omega_f)),\\
\widehat{\mathbf{u}}_m^\delta \rightarrow \mathbf{u}_m ^\xi &\text{ weakly star in } L^\infty(0,T; \mathbf{L}^2(\Omega_m)),\\
&\text{ weakly in } L^2(0,T; \mathbf{H}^1(\Omega_m)),\\
\widehat{\theta}^\delta \rightarrow \theta^\xi & \text{ weakly star in } L^\infty(0,T; L^2(\Omega)), \\
& \text{ weakly in } L^2(0,T; H^1_0(\Omega)),
\end{cases}
\label{conwmu}
\end{equation}
for certain limit functions $(\mathbf{u}_f^\xi, \mathbf{u}_m^\xi, \theta^\xi)$ satisfying
\begin{align*}
& \mathbf{u}_f^\xi \in L^\infty(0, T; \widetilde{\mathbf{H}}_{f,\mathrm{div}})\cap L^2(0, T; \mathbf{H}_{f,\mathrm{div}}), \\
& \mathbf{u}_m^\xi \in L^\infty(0, T; \widetilde{\mathbf{H}}_{m,\mathrm{div}})\cap L^2(0,T; \mathbf{H}_{m,\mathrm{div}}),\\
& \theta^\xi \in L^\infty(0, T; L^2(\Omega))\cap L^2(0, T; H^1_0(\Omega)),
\end{align*}
with $\mathbf{u}_f^\xi\cdot\mathbf{n}_{i}=\mathbf{u}_m^\xi\cdot\mathbf{n}_{i}$ on $\Gamma_i$.
In order to pass to the limit in those nonlinear terms, we need to obtain some information on the strong convergence of $\widehat{\theta}^\delta$ (up to a subsequence). It follows from equation \eqref{weak3ra}, the Gagliardo--Nirenberg inequality and the Sobolev embedding theorem that
\begin{eqnarray}
&& \|\partial_t \theta^\delta\|^{\frac{4}{3}}_{L^{\frac{4}{3}}(0,T;H^{-1}(\Omega))}\nonumber \\
& \leq& C\int_0^T \left(\|\nabla \widehat{\theta}^\delta\|_{\mathbf{L}^2(\Omega)}^\frac{4}{3}
+ \|\widehat{\theta}^\delta\|_{L^3(\Omega)}^\frac{4}{3}
\| \widehat{\mathbf{u}}^\delta\|_{\mathbf{L}^6(\Omega)}^\frac{4}{3}
\right) dt\nonumber \\
&\leq& C\int_0^T \|\nabla \widehat{\theta} ^\delta\|_{\mathbf{L}^2(\Omega)}^{\frac{4}{3}}dt
+ C \sup_{0\leq t\leq T} \|\widehat{\theta}^\delta\|_{L^2(\Omega)}^\frac{2}{3}
\int_0^T \|\widehat{\theta}^\delta\|_{H^1(\Omega)}^\frac{2}{3}\| \widehat{\mathbf{u}}^\delta\|_{\mathbf{H}^1(\Omega)}^\frac{4}{3} dt\nonumber \\
&\leq& C\int_0^T \Big(\|\nabla \widehat{\theta}^\delta\|_{\mathbf{L}^2(\Omega)}^2+1\Big)dt+ C \sup_{0\leq t\leq T} \|\widehat{\theta}^\delta\|_{L^2(\Omega)}^\frac{2}{3} \int_0^T \Big(\|\widehat{\theta}^\delta\|_{H^1(\Omega)}^2+\| \widehat{\mathbf{u}}^\delta\|_{\mathbf{H}^1(\Omega)}^2\Big) dt\nonumber \\
&\leq& C_T,\label{vpt1}
\end{eqnarray}
where the constant $C_T$ is independent of $\delta$, but may depend on $\xi$.
As a result, it follows that $\partial_t \theta^\delta \in L^{\frac43}(0,T; H^{-1}(\Omega))$ is bounded and $
\partial_t \theta^\delta \to \partial_t \theta^\xi $ in the sense of distribution. On the other hand, by the definition of $\theta^\delta$, it satisfies similar uniform estimates like those for $\widehat{\theta}^\delta$. Hence, applying Simon's compactness lemma (see e.g., \cite{Si85}), we deduce that there exists $$\theta^*\in L^2(0,T; H^{1-\beta}(\Omega))\cap C([0,T]; H^{-\beta}(\Omega)),$$
for some $\beta \in (0,\frac12)$ such that up to a subsequence,
$$ \theta^\delta \to \theta^* \quad \text{strongly in}\ L^2(0,T; H^{1-\beta}(\Omega))\cap C([0,T]; H^{-\beta}(\Omega))\quad
\text{as}\ \delta\to 0.
$$
Due to the uniqueness of limit (for the same convergent subsequence), we have $\theta^*=\theta^\xi$. Besides, by \eqref{conwmu} and \eqref{vpt1}, we also have $\theta^\xi\in C_w([0,T]; L^2(\Omega))$. Thus, concerning the initial datum, since by definition $\theta^\delta|_{t=0}=\theta_0$, we infer that
$$\theta^\xi|_{t=0}= \theta_0.$$
Next, since
$$\|\widehat{\theta}^\delta-\theta^\delta\|_{H^{-1}(\Omega)}
=\left\|(t_{k+1}-t)\frac{(\theta^{k+1}-\theta^k)}{\delta}\right\|_{H^{-1}(\Omega)}
\leq \delta \|\partial_t \theta^\delta\|_{H^{-1}(\Omega)}, \quad t\in (t_k, t_{k+1}],$$
for $k=0,1,...,N-1$, we infer from \eqref{vpt1} that
\begin{equation}
\int_0^T \|\widehat{\theta}^\delta-\theta^\delta\|_{H^{-1}(\Omega)} ^\frac43 dt\leq \delta^\frac43 \int_0^T \|\partial_t \theta^\delta\|_{H^{-1}(\Omega)}^\frac43 dt\to 0 \quad \text{as}\ \delta \to 0,
\end{equation}
which implies
\begin{align*}
\widehat{\theta}^\delta-\theta^\delta \to 0 \quad \text{strongly in}\ \ L^\frac43(0,T; H^{-1}(\Omega))\ \ \text{ as } \delta \rightarrow 0.
\end{align*}
Similarly, one can show that $\|\widetilde{\theta}^\delta -\theta^\delta\|_{L^\frac43(0,T; H^{-1}(\Omega))} \rightarrow 0$ as $\delta \rightarrow 0$. Therefore, the sequences $\{\theta^\delta\}$, $\{\widehat{\theta}^\delta\}$ and $\{\widetilde{\theta}^\delta\}$, if convergent, should converge to the same limit $\theta^\xi$. Besides, it follows from the above strong convergence, the uniform bounds \eqref{con1}, \eqref{mues1} and an interpolation argument that as $\delta \rightarrow 0$,
\begin{equation}
\widehat{\theta}^\delta, \widetilde{\theta}^\delta \to \theta^\xi \quad \text{strongly in}\ \ L^\frac{32}{17}(0,T; H^\frac34(\Omega))\cap L^q(0,T; L^2(\Omega)).\label{vpt}
\end{equation}
for any $q\in[2,+\infty)$. The strong convergence results also imply the pointwise almost everywhere convergence of $\theta^\delta$, $\widehat{\theta}^\delta$, $\widetilde{\theta}^\delta$ in $\Omega\times (0,T)$ (again up to a subsequence).
Next, using equation \eqref{weak1ra} and taking $\mathbf{v}_f\in L^{4} \big(0,T; \widehat{\mathbf{H}}_{f,\mathrm{div}})$, $\mathbf{v}_m=\mathbf{0}$, we can deduce that
\begin{eqnarray}
&& \|\partial_t \mathbf{u}_f^\delta\|^{\frac{4}{3}}_{L^{\frac{4}{3}} \big(0,T;(\widehat{\mathbf{H}}_{f,\mathrm{div}})^\prime\big)}\nonumber \\
&\leq&
C\int_0^T
\left(\|\widehat{\mathbf{u}}_f^\delta\|_{\mathbf{L}^3(\Omega_f)}^\frac43
\|\nabla \widehat{\mathbf{u}}_f^\delta\|_{\mathbf{L}^2(\Omega_f)}^\frac43+
\| \widehat{\mathbf{u}}_f^\delta\|_{\mathbf{H}^1(\Omega_f)}^\frac43
+\|\widehat{\theta}_f^\delta\|_{L^2(\Omega_f)}^\frac43
\right) dt\nonumber \\
&\leq&
C\int_0^T
\left(\|\widehat{\mathbf{u}}_f^\delta\|_{\mathbf{H}^1(\Omega_f)}^2
\|\widehat{\mathbf{u}}_f^\delta\|_{\mathbf{L}^2(\Omega_f)}^\frac{2}{3} + \| \widehat{\mathbf{u}}_f^\delta\|_{\mathbf{H}^1(\Omega_f)}^2
+ \|\widehat{\theta}^\delta_f\|_{L^2(\Omega_f)}^2 +1
\right) dt\nonumber \\
&\leq& C_T.\label{ut}
\end{eqnarray}
In a similar manner, taking test functions $\mathbf{v}_f=\mathbf{0}$ and $\mathbf{v}_m\in L^2(0,T;\widehat{\mathbf{H}}_{m,\mathrm{div}})$, we have
\begin{eqnarray}
&& \varpi^2\|\partial_t \mathbf{u}_m^\delta\|^2_{L^2(0,T;(\widehat{\mathbf{H}}_{m,\mathrm{div}})^\prime)}\nonumber \\
&\leq&
C\int_0^T
\left( \xi^2 \| \nabla \widehat{\mathbf{u}}_m^\delta\|_{\mathbf{L}^2(\Omega_m)}^2
+\|\widehat{\mathbf{u}}_m^\delta\|_{\mathbf{L}^2(\Omega_m)}^2
+\|\widehat{\theta}^\delta_m\|_{L^2(\Omega_m)}^2\right) dt\nonumber \\
&\leq& C_T.\label{utm}
\end{eqnarray}
The constant $C_T$ in \eqref{ut} and \eqref{utm} is independent of $\delta$ and $\xi$ (recalling that $\xi\in (0,1)$). If the nonlinear term on the interface are involved, we notice that for any $\mathbf{v}_f\in L^{4} \big(0,T;\mathbf{H}_{f,\mathrm{div}})$, it holds
\begin{eqnarray}
&& \sup_{\|\mathbf{v}_f\|_{L^4(0,T; \mathbf{H}^{1}(\Omega_f))}\leq 1} \left|\int_0^T\!\int_{\Gamma_{i}} |\widehat{\mathbf{u}}_f^\delta|^2 (\mathbf{v}_f\cdot \mathbf{n}_{i}) dSdt\right|\nonumber \\
&\leq & \int_0^T \big\||\widehat{\mathbf{u}}_f^\delta|^2\big\|_{\big(H_{00}^{\frac12}(\Gamma_i)\big)'}
\|\mathbf{v}_f\|_{\mathbf{H}_{00}^{\frac12}(\Gamma_i)}dt\nonumber \\
&\leq & C \int_0^T \big\||\widehat{\mathbf{u}}_f^\delta|^2\big\|_{L^\frac{4}{3}(\Gamma_i)}
\|\mathbf{v}_f\|_{\mathbf{H}^{1}(\Omega_f)} dt\nonumber \\
&\leq & C \left(\int_0^T \|\widehat{\mathbf{u}}_f^\delta\|^\frac83_{\mathbf{L}^\frac{8}{3}(\Gamma_i)} dt\right)^\frac34 \left(\int_0^T \|\mathbf{v}_f\|_{\mathbf{H}^{1}(\Omega_f)}^4 dt\right)^\frac14\nonumber \\
&\leq & C\left( \int_0^T \|\widehat{\mathbf{u}}_f^\delta\|^\frac83_{\mathbf{H}^\frac{1}{4}(\Gamma_i)} dt\right)^\frac34\nonumber \\
&\leq & C\left(\int_0^T \|\widehat{\mathbf{u}}_f^\delta\|^\frac83_{\mathbf{H}^\frac{3}{4}(\Omega_f)} dt\right)^\frac34\nonumber \\
&\leq & C\left(\int_0^T\|\widehat{\mathbf{u}}_f^\delta\|_{\mathbf{H}^1(\Omega_f)}^2 dt\right)^\frac34
\sup_{0\leq t\leq T}\|\widehat{\mathbf{u}}_f^\delta(t)\|_{\mathbf{L}^2(\Omega_f)}^\frac{1}{2}\nonumber \\
&\leq & C.\nonumber
\end{eqnarray}
Then in equation \eqref{weak1ra}, taking test functions $\mathbf{v}_f\in L^{4} (0,T; \mathbf{H}_{f,\mathrm{div}})$, $\mathbf{v}_m\in L^2(0,T;\mathbf{H}_{m,\mathrm{div}})$ with $\mathbf{v}_f\cdot\mathbf{n}_i=\mathbf{v}_m\cdot \mathbf{n}_i$ on $\Gamma_i$, we infer from the above estimates that
\begin{eqnarray}
\left|\int_0^T\!\int_{\Omega_f}\partial_t \mathbf{u}_f^\delta \cdot \mathbf{v}_f dxdt + \int_0^T\!\int_{\Omega_m}\partial_t \mathbf{u}_m^\delta \cdot \mathbf{v}_m dxdt \right| \leq C_T,\label{utd}
\end{eqnarray}
which also implies $\partial_t \mathbf{u}_f^\delta\in L^\frac43(0,T; (\mathbf{H}_{f,\mathrm{div}})')$ and $\partial_t \mathbf{u}_m^\delta \in L^2(0,T;(\mathbf{H}_{m,\mathrm{div}})')$.
From the estimates \eqref{ut}, \eqref{utm} on time derivatives, we can conclude the weak continuity property that $\mathbf{u}_f^\xi\in C_w([0,T]; \widetilde{\mathbf{H}}_{f,\mathrm{div}})$, $\mathbf{u}_m^\xi\in C_w([0,T]; \widetilde{\mathbf{H}}_{m,\mathrm{div}})$ and thus the initial conditions $\mathbf{u}_f^\xi|_{t=0}= \mathbf{u}_{0f}$, $\mathbf{u}_m^\xi|_{t=0}= \mathbf{u}_{0m}$ are fulfilled.
Besides, parallel to the arguments for $\theta^\delta$, $\widehat{\theta}^\delta$, we obtain the strong convergence as $\delta \rightarrow 0$ (up to a subsequence) such that
\begin{align}
\mathbf{u}_f^\delta \to \mathbf{u}_f^\xi, \quad &\text{strongly in}\ \ L^2(0,T; \mathbf{H}^{1-\beta}(\Omega_f))\cap C([0,T]; \mathbf{H}^{-\beta}(\Omega_f)), \label{vvf0}\\
\widehat{\mathbf{u}}_f^\delta-\mathbf{u}_f^\delta \to \mathbf{0}, \quad &\text{strongly in}\ \ L^\frac{4}{3}(0,T; (\widehat{\mathbf{H}}_{f,\mathrm{div}})^\prime), \label{vvf1} \\
\widehat{\mathbf{u}}_f^\delta \to \mathbf{u}_f^\xi, \quad &\text{strongly in}\ \ L^\frac{32}{17}(0,T; \mathbf{H}^\frac34(\Omega_f))\cap L^q(0,T; \mathbf{L}^2(\Omega_f)), \label{vvf2}\\
\mathbf{u}_m^\delta \to \mathbf{u}_m^\xi, \quad &\text{strongly in}\ \ L^2(0,T; \mathbf{H}^{1-\beta}(\Omega_m))\cap C([0,T]; \mathbf{H}^{-\beta}(\Omega_m)), \label{vvm0}\\
\widehat{\mathbf{u}}_m^\delta-\mathbf{u}_m^\delta \to \mathbf{0}, \quad &\text{strongly in}\ \ L^2(0,T; (\widehat{\mathbf{H}}_{m,\mathrm{div}})^\prime), \label{vvm1} \\
\widehat{\mathbf{u}}_m^\delta\to \mathbf{u}_m^\xi, \quad &\text{strongly in}\ \ L^\frac{32}{17}(0,T; \mathbf{H}^\frac34(\Omega_m)) \cap L^q(0,T; \mathbf{L}^2(\Omega_m)), \label{vvm2}
\end{align}
for some $\beta\in (0,\frac12)$ and any $q\geq 2$. Hence, we can further deduce the strong convergence of nonlinear terms
\begin{eqnarray}
& &\int_0^T \|\widehat{\mathbf{u}}_f^\delta\otimes \widehat{\mathbf{u}}_f^\delta -
\mathbf{u}_f^\xi\otimes \mathbf{u}_f^\xi\|_{\mathbf{L}^{\frac43}(\Omega_f)}^\frac{32}{17} dt\nonumber \\
&\leq &\int_0^T \|\widehat{\mathbf{u}}_f^\delta-\mathbf{u}_f^\xi\|_{\mathbf{L}^{4}(\Omega_f)}^\frac{32}{17}
\|\widehat{\mathbf{u}}_f^\delta+\mathbf{u}_f^\xi\|_{\mathbf{L}^{2}(\Omega_f)}^\frac{32}{17} dt\nonumber \\
& \leq & \sup_{0\leq t\leq T}\|\widehat{\mathbf{u}}_f^\delta(t)+\mathbf{u}_f^\xi(t)\|_{\mathbf{L}^{2}(\Omega_f)}
^\frac{32}{17}\int_0^T \|\widehat{\mathbf{u}}_f^\delta-\mathbf{u}_f^\xi\|_{\mathbf{H}^{\frac34}(\Omega_f)}^\frac{32}{17} dt\nonumber \\
& \to & 0\quad \text{as}\ \delta\to 0,\label{covua}
\end{eqnarray}
and
\begin{eqnarray}
& &\int_0^T
\big\||\widehat{\mathbf{u}}_f^\delta|^2-|\mathbf{u}_f^\xi|^2\big\|_{L^\frac{4}{3}(\Gamma_i)}^\frac{32}{29}dt\nonumber \\
& \leq & \int_0^T \|\widehat{\mathbf{u}}_f^\delta-\mathbf{u}_f^\xi\|_{\mathbf{L}^\frac{8}{3}(\Gamma_i)}^\frac{32}{29}
\|\widehat{\mathbf{u}}_f^\delta+\mathbf{u}_f^\xi\|_{\mathbf{L}^\frac{8}{3}(\Gamma_i)}^\frac{32}{29}dt\nonumber \\
& \leq & C\int_0^T \|\widehat{\mathbf{u}}_f^\delta-\mathbf{u}_f^\xi\|_{\mathbf{H}^\frac{1}{4}(\Gamma_i)}^\frac{32}{29}
\|\widehat{\mathbf{u}}_f^\delta+\mathbf{u}_f^\xi\|_{\mathbf{H}^\frac{1}{4}(\Gamma_i)}^\frac{32}{29}dt\nonumber \\
& \leq & C\int_0^T \|\widehat{\mathbf{u}}_f^\delta-\mathbf{u}_f^\xi\|_{\mathbf{H}^\frac{3}{4}(\Omega_f)}^\frac{32}{29}
\|\widehat{\mathbf{u}}_f^\delta+\mathbf{u}_f^\xi\|_{\mathbf{H}^1(\Omega_f)}^\frac{24}{29}
\|\widehat{\mathbf{u}}_f^\delta+\mathbf{u}_f^\xi\|_{\mathbf{L}^2(\Omega_f)}^\frac{8}{29} dt\nonumber \\
& \leq & C\left(\int_0^T \|\widehat{\mathbf{u}}_f^\delta-\mathbf{u}_f^\xi\|_{\mathbf{H}^\frac{3}{4}(\Omega_f)} ^\frac{32}{17}\right)^\frac{17}{29}
\left(\int_0^T \|\widehat{\mathbf{u}}_f^\delta+\mathbf{u}_f^\xi\|_{\mathbf{H}^1(\Omega_f)}^2dt\right)^\frac{12}{29}
\nonumber \\
&& \quad \times \sup_{0\leq t\leq T}\|\widehat{\mathbf{u}}_f^\delta(t)+\mathbf{u}_f^\xi(t)\|_{\mathbf{L}^2(\Omega_f)}^\frac{8}{29}\nonumber \\
& \to & 0 \quad \text{as}\ \delta\to 0.
\label{covub}
\end{eqnarray}
Based on the a.e. and strong convergence of $\widetilde{\theta}^\delta$, the assumptions (A1)--(A2) and the Sobolev embedding theorem, we see that
\begin{align*}
\nu(\widetilde{\theta}^\delta)\to \nu(\theta^\xi), \quad &\text{strongly in}\ \ L^2(0,T; L^4(\Omega)),\\
\nu(\widetilde{\theta}_m^\delta)\to \nu(\theta^\xi_m), \quad &\text{strongly in}\ \ L^2(0,T; L^\frac{8}{3}(\Gamma_i)),\\
\lambda(\widetilde{\theta}^\delta)\to \lambda(\theta^\xi),\quad & \text{strongly in}\ \ L^2(0,T; L^4(\Omega)).
\end{align*}
Then we deduce that
\begin{eqnarray*}
&&\left|\int_0^T\left(\nu(\widetilde{\theta}_f^\delta)\mathbb{D}(\widehat{\mathbf{u}}_f^\delta),
\mathbb{D}(\mathbf{v}_f)\right)_fdt
- \int_0^T\left(\nu(\theta_f^\xi)\mathbb{D}(\mathbf{u}_f^\xi),
\mathbb{D}(\mathbf{v}_f)\right)_fdt\right|
\\
&=&\left|\int_0^T\left((\nu(\widetilde{\theta}_f^\delta)-\nu(\theta_f^\xi))\mathbb{D}(\widehat{\mathbf{u}}_f^\delta),
\mathbb{D}(\mathbf{v}_f)\right)_fdt\right|\\
&&
+\left|\int_0^T\left(\nu(\theta_f^\xi)(\mathbb{D}(\widehat{\mathbf{u}}_f^\delta)-\mathbb{D}(\mathbf{u}_f^\xi)),
\mathbb{D}(\mathbf{v}_f)\right)_fdt\right|\\
&\leq& \sup_{0\leq t\leq T}\|\mathbb{D}(\mathbf{v}_f(t))\|_{\mathbf{L}^\infty(\Omega_f)}
\left(\int_0^T \|\nu(\widetilde{\theta}_f^\delta)-\nu(\theta_f^\xi)\|_{L^2(\Omega_f)}^2dt\right)^\frac12
\left(\int_0^T \|\mathbb{D}(\widehat{\mathbf{u}}_f^\delta)\|_{\mathbf{L}^2(\Omega_f)}^2dt\right)^\frac12\\
&& +\left|\int_0^T\left((\mathbb{D}(\widehat{\mathbf{u}}_f^\delta)-\mathbb{D}(\mathbf{u}_f^\xi)),
\nu(\theta_f^\xi)\mathbb{D}(\mathbf{v}_f)\right)_fdt\right|\\
&\to& 0\quad \text{as}\ \delta\to 0,
\end{eqnarray*}
for any $\mathbf{v}_f\in C([0,T]; \mathbf{H}_{f,\mathrm{div}}\cap \mathbf{W}^{1,\infty}(\Omega_f))$. In a similar manner, we get
\begin{align*}
\left| \int_0^T \left(\nu(\widetilde{\theta}_m^\delta)\mathbb{K}^{-1}\widehat{\mathbf{u}}_m^\delta,\mathbf{v}_m)\right)_m dt- \int_0^T \left(\nu(\theta_m^\xi)\mathbb{K}^{-1}\mathbf{u}_m^\xi,\mathbf{v}_m)\right)_m dt \right|
\to 0 \quad \text{as}\ \delta\to 0,
\end{align*}
for any $\mathbf{v}_m\in C([0,T]; \mathbf{H}_{m,\mathrm{div}}\cap \mathbf{L}^{\infty}(\Omega_m))$;
\begin{eqnarray*}
& & \left|\sum_{j=1}^{d-1}\int_0^T\!\int_{\Gamma_{i}}
\frac{\alpha\nu(\widetilde{\theta}_m^\delta)}{\sqrt{{\rm trace} (\mathbb{K})}}
(\widehat{\mathbf{u}}_f^\delta \cdot\boldsymbol{\tau}_j)(\mathbf{v}_f\cdot\boldsymbol{\tau}_j)dSdt\right.\\
& &\quad \left. -\sum_{j=1}^{d-1}\int_0^T\!\int_{\Gamma_{i}}\frac{\alpha\nu(\theta_m^\xi)}{\sqrt{{\rm trace} (\mathbb{K})}}
(\mathbf{u}_f^\xi \cdot\boldsymbol{\tau}_j)(\mathbf{v}_f\cdot\boldsymbol{\tau}_j)dSdt\right|\to 0 \quad \text{as}\ \delta\to 0,
\end{eqnarray*}
for any $\mathbf{v}_f\in C([0,T]; \mathbf{H}_{f,\mathrm{div}}\cap \mathbf{H}^2(\Omega_f))$; and
\begin{align*}
& \left| \int_0^T\big(\lambda(\widetilde{\theta}^\delta) \nabla \widehat{\theta}^\delta,\nabla\phi\big)dt
-
\int_0^T\big(\lambda(\theta^\xi) \nabla \theta^\xi,\nabla\phi\big)dt\right|
\to 0\quad \text{as}\ \delta\to 0,\\
& \left|\int_0^T \big(\widehat{\mathbf{u}}^\delta \widehat{\theta}^\delta, \nabla\phi\big)dt- \int_0^T \big(\mathbf{u}^\xi \theta^\xi, \nabla \phi\big)dt\right| \to 0\quad \text{as}\ \delta\to 0,
\end{align*}
for any $\phi \in C([0,T]; H^1_0(\Omega)\cap W^{1,\infty}(\Omega))$.
Using the above convergence results, we are able to pass to the limit as $\delta\to 0$ (up to a subsequence) in \eqref{weak1ra}--\eqref{weak3ra} to show that the triple $(\mathbf{u}^\xi_f, \mathbf{u}^\xi_m, \theta^\xi)$ is indeed a weak solution to the regularized system \eqref{weak1r}--\eqref{weak3r} on $[0,T]$. \medskip
\textbf{Step 2. Passage to the limit $\xi\to 0$}.
Next, we pass to the limit as $\xi\to 0$ in the weak form \eqref{weak1r}--\eqref{weak3r}. To this end, we show that $(\mathbf{u}_f^\xi, \mathbf{u}_m^\xi, \theta^\xi)$ fulfills some energy estimates uniform in $\xi$.
It follows from the strong convergence results \eqref{vpt}, \eqref{vvf2} and \eqref{vvm2} that as $\delta \to 0$, for almost all $t\in (0, T)$, we have (up to a subsequence),
\begin{eqnarray}
&& \widehat{\mathbf{u}}_f^\delta(t)\to \mathbf{u}_f^\xi(t), \quad \ \, \text{strongly in}\ \ \mathbf{L}^2(\Omega_f),\nonumber \\
&& \widehat{\mathbf{u}}_m^\delta(t)\to \mathbf{u}_m^\xi(t), \quad \text{strongly in}\ \ \mathbf{L}^2(\Omega_m),\nonumber \\
&& \widehat{\theta}^\delta(t) \to \theta^\xi(t),\ \quad\ \ \text{strongly in}\ \ L^2(\Omega),\nonumber
\end{eqnarray}
which imply that
\begin{equation}
\mathcal{E}_\sigma(\widehat{\mathbf{u}}_f^{\delta}(t), \widehat{\mathbf{u}}_m^{\delta}(t), \widehat{\theta}^{\delta}(t))\to \mathcal{E}_\sigma(\mathbf{u}_f^\xi(t), \mathbf{u}_m^\xi(t), \theta^\xi(t)),\quad \text{for a.a.} \ t\in (0, T).\nonumber
\end{equation}
By the estimate \eqref{ee1}, we have
\begin{align}
\mathcal{E}_\sigma(\mathbf{u}_f^\xi(t), \mathbf{u}_m^\xi(t), \theta^\xi(t))
&\leq \mathcal{E}_\sigma(\mathbf{u}_{0f},\mathbf{u}_{0m}, \theta_0),\label{ee1a}
\end{align}
for a.a. $t\in (0,T)$, and thanks to the lower semi-continuity of norms, we get
\begin{align}
\int_0^T \mathcal{D}_\sigma^\xi(\tau) d\tau\leq C \mathcal{E}_\sigma(\mathbf{u}_{0f},\mathbf{u}_{0m}, \theta_0),\label{ee2a}
\end{align}
where
\begin{align}
\mathcal{D}_\sigma^\xi &=2 \big(\mathbb{D}(\mathbf{u}_f^{\xi}), \mathbb{D}(\mathbf{u}_f^{\xi})\big)_f+ \big(\mathbf{u}_m^{\xi}, \mathbf{u}_m^{\xi}\big)_m +\xi\big(\nabla \mathbf{u}_m^{\xi}, \nabla \mathbf{u}_m^{\xi}\big)_m \nonumber \\
&\quad +\sum_{j=1}^{d-1}\int_{\Gamma_{i}}|\mathbf{u}_f^{\xi}\cdot\boldsymbol{\tau}_j|^2 dS +\sigma\int_\Omega |\nabla\theta^{\xi}|^2dx.\label{Dxi}
\end{align}
Hence, it follows from \eqref{ee1a}--\eqref{Dxi} that
\begin{equation}
\begin{cases}
\mathbf{u}_f^\xi \rightarrow \mathbf{u}_f & \text{ weakly star in } L^\infty(0,T; \mathbf{L}^2(\Omega_f)),\\
&\text{ weakly in } L^2(0,T; \mathbf{H}^1(\Omega_f)),\\
\mathbf{u}_m^\xi \rightarrow \mathbf{u}_m &\text{ weakly star in } L^\infty(0,T; \mathbf{L}^2(\Omega_m)),\\
\xi \nabla \mathbf{u}_m^\xi \rightarrow \mathbf{0} &\text{ strongly in } L^2(0,T; \mathbf{L}^2(\Omega_m)),\\
\theta^\xi \rightarrow \theta & \text{ weakly star in } L^\infty(0,T; L^2(\Omega)), \\
& \text{ weakly in } L^2(0,T; H^1_0(\Omega)),
\end{cases}
\label{conwmua}
\end{equation}
for certain functions $(\mathbf{u}_f, \mathbf{u}_m, \theta)$ satisfying
\begin{align*}
& \mathbf{u}_f \in L^\infty(0, T; \widetilde{\mathbf{H}}_{f,\mathrm{div}})\cap L^2(0, T; \mathbf{H}_{f,\mathrm{div}}), \\
& \mathbf{u}_m \in L^\infty(0, T; \widetilde{\mathbf{H}}_{m,\mathrm{div}}),\\
& \theta \in L^\infty(0, T; L^2(\Omega))\cap L^2(0, T; H^1_0(\Omega)),
\end{align*}
and $\mathbf{u}_f\cdot \mathbf{n}_i=\mathbf{u}_m\cdot \mathbf{n}_i$ on $\Gamma_i$. Then by similar arguments like for \eqref{vpt1}, \eqref{ut} and \eqref{utm}, we can deduce that for any $\xi\in(0,1)$, the following estimates hold:
\begin{eqnarray}
&& \|\partial_t \theta^\xi\|^{2}_{L^{2}(0,T;(W^{1,3}_0(\Omega))')}\nonumber \\
& \leq& C\int_0^T \left(\|\nabla \theta^\xi\|_{\mathbf{L}^2(\Omega)}^2
+ \|\theta^\xi\|_{L^6(\Omega)}^2
\|\mathbf{u}^\xi\|_{\mathbf{L}^2(\Omega)}^2
\right) dt\nonumber \\
&\leq& C\int_0^T \|\nabla \theta ^\xi\|_{\mathbf{L}^2(\Omega)}^{2}dt
+ C \sup_{0\leq t\leq T} \|\mathbf{u}^\xi(t)\|_{\mathbf{L}^2(\Omega)}^2
\int_0^T \|\theta^\xi\|_{H^1(\Omega)}^2 dt\nonumber \\
&\leq& C_T,\label{vpt1a}
\end{eqnarray}
\begin{eqnarray}
&& \|\partial_t \mathbf{u}_f^\xi\|^{\frac{4}{3}}
_{L^{\frac{4}{3}}(0,T;(\widehat{\mathbf{H}}_{f,\mathrm{div}})^\prime)}\nonumber \\
&\leq&
C\int_0^T
\left(\|\mathbf{u}_f^\xi\|_{\mathbf{L}^3(\Omega_f)}^\frac43
\|\nabla \mathbf{u}_f^\xi\|_{\mathbf{L}^2(\Omega_f)}^\frac43+
\| \mathbf{u}_f^\xi\|_{\mathbf{H}^1(\Omega_f)}^\frac43
\|\theta^\xi\|_{L^2(\Omega)}^\frac43
\right) dt\nonumber \\
&\leq& C_T,\label{uta}
\end{eqnarray}
\begin{eqnarray}
&& \varpi^2\|\partial_t \mathbf{u}_m^\xi\|^2_{L^2(0,T;(\widehat{\mathbf{H}}_{m,\mathrm{div}})^\prime)}
\nonumber \\
&\leq&
C\int_0^T
\left[ \xi \big(\xi \| \nabla \widehat{\mathbf{u}}_m^\xi\|_{\mathbf{L}^2(\Omega_m)}^2\big)
+\|\mathbf{u}_m^\xi\|_{\mathbf{L}^2(\Omega_m)}^2
+\|\theta^\xi\|_{L^2(\Omega)}^2\right] dt\nonumber \\
&\leq& C_T,\label{utma}
\end{eqnarray}
where the constant $C_T$ in the above estimates is independent of $\xi$. Besides, similar to \eqref{utd}, we can see that $\partial_t \mathbf{u}_f^\xi\in L^\frac43(0,T; (\mathbf{H}_{f,\mathrm{div}})')$ and $\partial_t \mathbf{u}_m^\xi \in L^2(0,T;(\mathbf{H}_{m,\mathrm{div}})')$ for all $\xi \in(0,1)$.
The estimates \eqref{vpt1a}--\eqref{utma} on time derivatives
imply the weak continuity in time such that $\theta\in C_w([0,T]; H^1_0(\Omega))$, $\mathbf{u}_f\in C_w([0,T]; \widetilde{\mathbf{H}}_{f,\mathrm{div}})$, $\mathbf{u}_m\in C_w([0,T]; \widetilde{\mathbf{H}}_{m,\mathrm{div}})$ and thus $\theta|_{t=0}= \theta_0$, $\mathbf{u}_f|_{t=0}= \mathbf{u}_{0f}$, $\mathbf{u}_m|_{t=0}= \mathbf{u}_{0m}$. We also infer the strong convergence as $\xi \rightarrow 0$ (up to a subsequence) such that
\begin{align}
\theta^\xi \to \theta, \quad &\text{strongly in}\ L^2(0,T; H^{1-\beta}(\Omega))\cap C([0,T]; H^{-\beta}(\Omega)),\label{vvt2a}\\
\mathbf{u}_f^\xi \to \mathbf{u}_f, \quad &\text{strongly in}\ \ L^2(0,T; \mathbf{H}^{1-\beta}(\Omega_f))\cap C([0,T]; \mathbf{H}^{-\beta}(\Omega_f)), \label{vvf2a}\\
\mathbf{u}_m^\xi \to \mathbf{u}_m, \quad &\text{strongly in}\ \ C([0,T]; \mathbf{H}^{-\beta}(\Omega_m)), \label{vvm2a}
\end{align}
for some $\beta\in (0,\frac12)$. Hence, following exactly the same argument as for \eqref{covua}, \eqref{covub}, we can further deduce the strong convergence of nonlinear terms
\begin{align}
& \int_0^T \|\mathbf{u}_f^\xi\otimes \mathbf{u}_f^\xi -
\mathbf{u}_f\otimes \mathbf{u}_f\|_{\mathbf{L}^{\frac43}(\Omega_f)}^2 dt
\to 0\quad \text{as}\ \xi\to 0,\nonumber
\end{align}
\begin{align}
& \int_0^T
\big\||\mathbf{u}_f^\xi|^2-|\mathbf{u}_f|^2\big\|_{L^\frac{4}{3}(\Gamma_i)}^\frac87dt
\to 0 \quad \text{as}\ \xi\to 0.\nonumber
\end{align}
Besides, using assumptions (A1)--(A2) and \eqref{vvt2a}, we get
\begin{align*}
\nu(\theta^\xi)\to \nu(\theta), \quad &\text{strongly in}\ \ L^2(0,T; L^4(\Omega)),\\
\nu(\theta^\xi_m)\to \nu(\theta_m), \quad &\text{strongly in}\ \ L^2(0,T; L^\frac{8}{3}(\Gamma_i)),\\
\lambda(\theta^\xi)\to \lambda(\theta),\quad & \text{strongly in}\ \ L^2(0,T; L^4(\Omega)),
\end{align*}
which together with \eqref{vvt2a}--\eqref{vvm2a} yield
\begin{align*}
&\left|\int_0^T\left(\nu(\theta_f^\xi)\mathbb{D}(\mathbf{u}_f^\xi),
\mathbb{D}(\mathbf{v}_f)\right)_fdt
- \int_0^T\left(\nu(\theta_f)\mathbb{D}(\mathbf{u}_f),
\mathbb{D}(\mathbf{v}_f)\right)_fdt\right|
\to 0\quad \text{as}\ \xi\to 0,
\end{align*}
for any $\mathbf{v}_f\in C([0,T]; \mathbf{H}_{f,\mathrm{div}}\cap \mathbf{W}^{1,\infty}(\Omega_f))$;
\begin{align*}
&\left|\sum_{j=1}^{d-1}\int_0^T\!\int_{\Gamma_{i}}
\frac{\alpha\nu(\theta_m^\xi)}{\sqrt{{\rm trace} (\mathbb{K})}}
(\mathbf{u}_f^\xi \cdot\boldsymbol{\tau}_j)(\mathbf{v}_f\cdot\boldsymbol{\tau}_j)dSdt\right.\\
&\quad \left.-\sum_{j=1}^{d-1}\int_0^T\!\int_{\Gamma_{i}}
\frac{\alpha\nu(\theta_m)}{\sqrt{{\rm trace} (\mathbb{K})}}
(\mathbf{u}_f \cdot\boldsymbol{\tau}_j)(\mathbf{v}_f\cdot\boldsymbol{\tau}_j)dSdt\right|\ \to 0\quad \text{as}\ \xi\to 0,
\end{align*}
for any $\mathbf{v}_f\in C([0,T]; \mathbf{H}_{f,\mathrm{div}}\cap \mathbf{H}^2(\Omega_f))$;
\begin{align*}
&\left|\int_0^T \left(\nu(\theta_m^\xi)\mathbb{K}^{-1}\mathbf{u}_m^\xi,\mathbf{v}_m)\right)_m dt- \int_0^T \left(\nu(\theta_m)\mathbb{K}^{-1}\mathbf{u}_m,\mathbf{v}_m)\right)_m dt\right|\\
&\leq \left|\int_0^T \left((\nu(\theta_m^\xi)-\nu(\theta_m))\mathbb{K}^{-1}\mathbf{u}_m^\xi,\mathbf{v}_m)\right)_m dt\right|+\left| \int_0^T \left(\nu(\theta_m)\mathbb{K}^{-1}(\mathbf{u}_m^\xi-\mathbf{u}_m),\mathbf{v}_m)\right)_m dt\right|\nonumber \\
&\to 0\quad \text{as}\ \xi\to 0,
\end{align*}
for any $\mathbf{v}_m\in C([0,T]; \mathbf{H}_{m,\mathrm{div}}\cap \mathbf{L}^\infty(\Omega_m))$;
and
\begin{align*}
&\left|\int_0^T\left(\lambda(\theta^\xi) \nabla \theta^\xi,\nabla\phi\right)dt
-
\int_0^T\left(\lambda(\theta) \nabla \theta,\nabla\phi\right)dt\right|\to 0\quad \text{as}\ \xi\to 0,\\
& \left|\int_0^T \big(\mathbf{u}^\xi \theta^\xi, \nabla\phi\big)dt- \int_0^T \big(\mathbf{u} \theta, \nabla \phi\big)dt\right| \to 0\quad \text{as}\ \xi\to 0,
\end{align*}
for any $\phi \in C([0,T]; H^1_0(\Omega)\cap W^{1,\infty}(\Omega))$.
Then we are able to pass to the limit as $\xi\to 0$ (up to a subsequence) in the weak formulation \eqref{weak1r}--\eqref{weak3r} to show that the limit triple $(\mathbf{u}_f, \mathbf{u}_m, \theta)$ is indeed a weak solution to the original system \eqref{weak1}--\eqref{weak3} on $[0,T]$, keeping in mind that by integration by parts, it holds
\begin{eqnarray}
&&\int_0^T\left(\partial_t \mathbf{u}_f ,\mathbf{v}_f\right)_f dt =- \int_0^T\left( \mathbf{u}_f ,\partial_t \mathbf{v}_f\right)_f dt,\nonumber \\
&&\int_0^T\left(\partial_t \mathbf{u}_m ,\mathbf{v}_m\right)_m dt =- \int_0^T\left( \mathbf{u}_m ,\partial_t \mathbf{v}_m\right)_m dt,\nonumber \\
&& \int_0^T \left(\partial_t \theta,\phi\right)dt =-\int_0^T \left( \theta,\partial_t \phi\right)dt,\nonumber
\end{eqnarray}
for sufficiently regular test functions that have compact support in $(0,T)$.
On the other hand, we see that for any $(\mathbf{v}_f,\mathbf{v}_m)\in L^{4} \big(0,T;\mathbf{H}_{f,\mathrm{div}}) \times L^2(0,T; \widetilde{\mathbf{H}}_{m,\mathrm{div}})$ with $\mathbf{v}_f\cdot \mathbf{n}_i=\mathbf{v}_m\cdot \mathbf{n}_i$ on $\Gamma_{i}$, it holds (cf. \eqref{ut}--\eqref{utd})
\begin{eqnarray}
&& \left|\int_0^T(\partial_t\mathbf{u}_f,\mathbf{v}_f)_f dt
+\varpi\int_0^T(\partial_t \mathbf{u}_m,\mathbf{v}_m)_m dt\right|\nonumber \\
&\leq &
C\left[\int_0^T
\left( \|\mathbf{u}_f\|_{\mathbf{H}^1(\Omega_f)}^2
\|\mathbf{u}_f\|_{\mathbf{L}^2(\Omega_f)}^\frac{2}{3} + \| \mathbf{u}_f\|_{\mathbf{H}^1(\Omega_f)}^2
+ \|\theta_f\|_{L^2(\Omega_f)}^2 +1
\right) dt\right]^\frac34\nonumber \\
&&\times \left(\int_0^T\|\mathbf{v}_f\|_{\mathbf{H}^1(\Omega)}^4dt\right)^\frac14 \nonumber \\
&& + C\left[\int_0^T\left(\|\mathbf{u}_m\|_{\mathbf{L}^2(\Omega_m)}^2
+\|\theta_m\|_{L^2(\Omega_m)}^2\right)dt\right]^\frac12 \left(\int_0^T\|\mathbf{v}_m\|_{\mathbf{L}^2(\Omega)}^2dt\right)^\frac12\nonumber \\
&\leq& C,\nonumber
\end{eqnarray}
which implies $\partial_t\mathbf{u}_f \in L^{\frac43}(0,T; (\mathbf{H}_{f,\mathrm{div}})')$ and $ \partial_t \mathbf{u}_m\in L^2(0,T;(\widetilde{\mathbf{H}}_{m,\mathrm{div}})')$.
Finally, in view of the energy inequalities \eqref{eapp1c}, \eqref{eapp2c} and using the above convergence results we can further conclude that for almost all $t\in [0,T]$, the global weak solution satisfies
\begin{eqnarray}
&&\frac{1}{2}\|\mathbf{u}_f(t)\|_{\mathbf{L}^2(\Omega_f)}^2
+
\frac{\varpi}{2}\|\mathbf{u}_m(t)\|_{\mathbf{L}^2(\Omega_m)}^2
+2\int_0^t\big(\nu(\theta_f)\mathbb{D}(\mathbf{u}_f),\mathbb{D}(\mathbf{u}_f)\big)_f d\tau\nonumber \\
&&+\int_0^t \left( \nu(\theta_m)\mathbb{K}^{-1}\mathbf{u}_m, \mathbf{u}_m\right)_m d\tau
+\sum_{j=1}^{d-1}\int_0^t\! \int_{\Gamma_{i}}\frac{\alpha \nu(\theta_m)}{\sqrt{{\rm trace} (\mathbb{K})}} |\mathbf{u}_f\cdot\boldsymbol{\tau}_j|^2 dS d\tau \nonumber \\
&\leq &\frac{1}{2}\|\mathbf{u}_{0f}\|_{\mathbf{L}^2(\Omega_f)}^2
+\frac{\varpi}{2}\|\mathbf{u}_{0m}\|_{\mathbf{L}^2(\Omega_m)}^2 + \int_0^t (\theta_f \mathbf{k}, \mathbf{u}_f )_f + (\theta_m \mathbf{k}, \mathbf{u}_m )_m d\tau,
\label{eapp1d}
\end{eqnarray}
and
\begin{align}
& \frac{1}{2}\|\theta(t)\|_{L^2(\Omega)}^2
+\int_0^t\! \int_\Omega \lambda(\theta)|\nabla\theta|^2 dx d\tau
\leq \frac{1}{2}\|\theta_0\|_{L^2(\Omega)}^2.
\label{eapp2d}
\end{align}
\subsubsection{Case $\varpi=0$}
For the case of $\varpi=0$, we note that only some weaker estimates are available for the velocity $\mathbf{u}_m$ in the matrix part. Keeping this in mind, below we point out necessary modifications for the proof of global weak solutions.
First, for the discrete system \eqref{weak1r}--\eqref{weak3r}, comparing with \eqref{con1} and \eqref{conm1}, we now only have the estimate for
$\|\widehat{\mathbf{u}}_m^\delta\|_{L^2(0,T;\mathbf{H}^1(\Omega_m))}$, which implies
$$
\widehat{\mathbf{u}}_m^\delta \rightarrow \mathbf{u}_m ^\xi \text{ weakly in } L^2(0,T; \mathbf{H}^1(\Omega_m))\quad \text{as}\ \delta\to 0
$$
for some limit function $\mathbf{u}_m^\xi \in L^2(0,T; \mathbf{H}_{m,\mathrm{div}})$. The sequential strong convergence of $\widehat{\mathbf{u}}_m^\delta$ is no longer available. Taking $\delta \to 0$ (while keeping $\xi>0$ fixed), we still have the uniform estimates for the limit triple $(\mathbf{u}_f^\xi, \mathbf{u}_m^\xi,\theta^\xi)$ that is a weak solution to the regularized system \eqref{weak1r}--\eqref{weak3r} with $\varpi=0$:
\begin{align}
\mathcal{E}_\sigma(\mathbf{u}_f^\xi(t), \theta^\xi(t))
&\leq \mathcal{E}_\sigma(\mathbf{u}_{0f}, \theta_0),\label{ee1b}
\end{align}
for a.a. $t\in (0,T)$ and
\begin{align}
\int_0^T \mathcal{D}_\sigma^\xi(\tau) d\tau\leq C \mathcal{E}_\sigma(\mathbf{u}_{0f}, \theta_0),\label{ee2b}
\end{align}
where $ \mathcal{D}_\sigma^\xi$ is given by \eqref{Dxi}.
Then in \eqref{conwmua}, we now only have as $\xi\to 0$ (up to a subsequence)
\begin{align*}
&\mathbf{u}_m^\xi \rightarrow \mathbf{u}_m \quad\ \text{ weakly in } L^2(0,T; \mathbf{L}^2(\Omega_m)),\\
&\xi \nabla \mathbf{u}_m^\xi \rightarrow \mathbf{0}\quad \text{ strongly in } L^2(0,T; \mathbf{L}^2(\Omega_m)),
\end{align*}
for some $\mathbf{u}_m \in L^2(0, T; \widetilde{\mathbf{H}}_{m,\mathrm{div}})$. Next, we see that under the current regularity of $\mathbf{u}_m^\xi$, it holds
\begin{eqnarray}
&& \|\partial_t \theta^\xi\|^{\frac87}_{L^{\frac87}(0,T;(W^{1,4}_0(\Omega))')}\nonumber \\
& \leq& C\int_0^T \left(\|\nabla \theta^\xi\|_{\mathbf{L}^2(\Omega)}^\frac87
+ \|\theta^\xi\|_{L^4(\Omega)}^\frac87
\|\mathbf{u}^\xi\|_{\mathbf{L}^2(\Omega)}^\frac87
\right) dt\nonumber \\
&\leq& C\int_0^T \left(\|\nabla \theta ^\xi\|_{\mathbf{L}^2(\Omega)}^{2}+1\right)dt\nonumber \\
&& + C \sup_{0\leq t\leq T} \|\theta^\xi(t)\|_{L^2(\Omega)}^\frac{2}{7}
\int_0^T \|\mathbf{u}^\xi\|_{\mathbf{L}^2(\Omega)}^\frac87 \|\theta^\xi\|_{H^1(\Omega)}^\frac{6}{7} dt\nonumber \\
&\leq& C\int_0^T \left(\|\nabla \theta ^\xi\|_{\mathbf{L}^2(\Omega)}^{2} +1\right)dt\nonumber \\
&& + C \sup_{0\leq t\leq T} \|\theta^\xi(t)\|_{L^2(\Omega)}^\frac{2}{7}
\left(\int_0^T \|\mathbf{u}^\xi\|_{\mathbf{L}^2(\Omega)}^2dt\right)^{\frac47} \left(\int_0^T\|\theta^\xi\|_{H^1(\Omega)}^2dt\right)^\frac{3}{7} \nonumber \\
&\leq& C_T,\label{vpt1b}
\end{eqnarray}
where $C_T$ is independent of $\xi$.
Keeping these modifications in mind, we can pass to the limit as $\xi\to 0$ and conclude the existence of a weak solution to system \eqref{weak1}--\eqref{weak3} on $[0,T]$ by a similar argument for the case $\varpi >0$. Moreover, the energy inequalities \eqref{eapp1d}--\eqref{eapp2d} still hold (now with $\varpi=0$).
The proof of Theorem \ref{thmEx} is complete.
\hfill$\square$
\section{Weak-Strong Uniqueness}\setcounter{equation}{0}
In this section, we prove Theorem \ref{thmuniq} on the weak-strong uniqueness of solutions to problem \eqref{uf1}--\eqref{IBCi2}.
Let $(\mathbf{u}_f, \mathbf{u}_m, \theta)$ be a weak solution to problem \eqref{uf1}--\eqref{IBCi2}. Then from the previous section, it satisfies the energy inequalities \eqref{eapp1d}--\eqref{eapp2d} for $\varpi\geq 0$. On the other hand, the regular solution $(\bar{\mathbf{u}}_f, \bar{\mathbf{u}}_m, \bar{\theta})$ that assumed to exist on $[0,T]$ is allowed to be used as a test function in its weak formulation (i.e., \eqref{weak1}--\eqref{weak3}). By a direct computation, we obtain the following energy equalities for $(\bar{\mathbf{u}}_f, \bar{\mathbf{u}}_m, \bar{\theta})$:
\begin{eqnarray}
&&\frac{1}{2}\|\bar{\mathbf{u}}_f(t)\|_{\mathbf{L}^2(\Omega_f)}^2
+
\frac{\varpi}{2}\|\bar{\mathbf{u}}_m(t)\|_{\mathbf{L}^2(\Omega_m)}^2
+2\int_0^t\left(\nu(\bar{\theta}_f)\mathbb{D}(\bar{\mathbf{u}}_f),\mathbb{D}(\bar{\mathbf{u}}_f)\right)_f d\tau\nonumber \\
&&+\int_0^t \left( \nu(\bar{\theta}_m)\mathbb{K}^{-1}\bar{\mathbf{u}}_m, \bar{\mathbf{u}}_m\right)_m d\tau
+\sum_{j=1}^{d-1} \int_0^t\!\int_{\Gamma_{i}}\frac{\alpha \nu(\bar{\theta}_m)}{\sqrt{{\rm trace} (\mathbb{K})}} |\bar{\mathbf{u}}_f\cdot\boldsymbol{\tau}_j|^2 dS d\tau \nonumber \\
&=&\frac{1}{2}\|\mathbf{u}_{0f}\|_{\mathbf{L}^2(\Omega_f)}^2
+\frac{\varpi}{2}\|\mathbf{u}_{0m}\|_{\mathbf{L}^2(\Omega_m)}^2 +\int_0^t (\bar{\theta}_f \mathbf{k}, \bar{\mathbf{u}}_f )_f + (\bar{\theta}_m \mathbf{k}, \bar{\mathbf{u}}_m )_m d\tau,
\label{eapp1e}
\end{eqnarray}
and
\begin{align}
& \frac{1}{2}\|\bar{\theta}(t)\|_{L^2(\Omega)}^2
+\int_0^t\! \int_\Omega \lambda(\bar{\theta})|\nabla\bar{\theta}|^2 dx d\tau
= \frac{1}{2}\|\theta_0\|_{L^2(\Omega)}^2.
\label{eapp2e}
\end{align}
Next, in the weak formulation \eqref{weak1} for $(\mathbf{u}_f, \mathbf{u}_m)$, we take the test function $(\mathbf{v}_f,\mathbf{v}_m)=(-\bar{\mathbf{u}}_f,-\bar{\mathbf{u}}_m)$ and perform integration by parts to get
\begin{eqnarray}
&&
-\big(\mathbf{u}_f(t),\bar{\mathbf{u}(t)}_f\big)_f -\varpi\big(\mathbf{u}_m(t),\bar{\mathbf{u}}_m(t)\big)_m
-2\int_0^t\big(\nu(\theta_f)\mathbb{D}(\mathbf{u}_f),\mathbb{D}(\bar{\mathbf{u}}_f)\big)_fd\tau
\nonumber \\
&& - \int_0^t \left(\nu(\theta_m)\mathbb{K}^{-1}\mathbf{u}_m,\bar{\mathbf{u}}_m\right)_m d\tau - \sum_{j=1}^{d-1} \int_0^t\!\int_{\Gamma_{i}}\frac{\alpha \nu(\theta_m)}{\sqrt{{\rm trace} (\mathbb{K})}} (\mathbf{u}_f\cdot\boldsymbol{\tau}_j)(\bar{\mathbf{u}}_f\cdot\boldsymbol{\tau}_j)dSd\tau\nonumber \\
&=& -\|\mathbf{u}_{0f}\|_{\mathbf{L}^2(\Omega_f)}^2
-\int_0^t(\mathbf{u}_f,\partial_t\bar{\mathbf{u}}_f)_f d\tau
-\varpi\|\mathbf{u}_{0m}\|_{\mathbf{L}^2(\Omega_m)}^2 - \varpi\int_0^t(\mathbf{u}_m,\partial_t\bar{\mathbf{u}}_m)_m d\tau\nonumber \\
&& +\int_0^t (\mathrm{div}(\mathbf{u}_f\otimes \mathbf{u}_f), \bar{\mathbf{u}}_f)_fd\tau - \int_0^t\! \int_{\Gamma_{i}} \frac12|\mathbf{u}_f|^2 (\bar{\mathbf{u}}_f\cdot \mathbf{n}_{i}) dSd\tau
\nonumber \\
&& -
\int_0^t (\theta_f \mathbf{k}, \bar{\mathbf{u}}_f )_f d\tau -
\int_0^t (\theta_m \mathbf{k}, \bar{\mathbf{u}}_m )_m d\tau\nonumber \\
&=& -\|\mathbf{u}_{0f}\|_{\mathbf{L}^2(\Omega_f)}^2
-\varpi\|\mathbf{u}_{0m}\|_{\mathbf{L}^2(\Omega_m)}^2
+\int_0^t (\mathrm{div}(\mathbf{u}_f\otimes \mathbf{u}_f), \bar{\mathbf{u}}_f)_fd\tau
\nonumber \\
&& + \int_0^t (\mathbf{u}_f, \mathrm{div}(\bar{\mathbf{u}}_f\otimes \bar{\mathbf{u}}_f))_fd\tau +2\int_0^t(\nu(\bar{\theta}_f)\mathbb{D}(\mathbf{u}_f),\mathbb{D}(\bar{\mathbf{u}}_f))_fd\tau
\nonumber \\
&&- \int_0^t\! \int_{\Gamma_{i}} \frac12|\mathbf{u}_f|^2 (\bar{\mathbf{u}}_f\cdot \mathbf{n}_{i}) dSd\tau - \int_0^t\! \int_{\Gamma_{i}} \frac12|\bar{\mathbf{u}}_f|^2 (\mathbf{u}_f\cdot \mathbf{n}_{i}) dSd\tau
\nonumber \\
&& + \sum_{j=1}^{d-1} \int_0^t\!\int_{\Gamma_{i}}\frac{\alpha \nu(\bar{\theta}_m)}{\sqrt{{\rm trace} (\mathbb{K})}} (\mathbf{u}_f\cdot\boldsymbol{\tau}_j)(\bar{\mathbf{u}}_f\cdot\boldsymbol{\tau}_j)dSd\tau+ \int_0^t \left(\nu(\bar{\theta}_m)\mathbb{K}^{-1}\mathbf{u}_m,\bar{\mathbf{u}}_m)\right)_m d\tau\nonumber \\
&& -
\int_0^t (\theta_f \mathbf{k}, \bar{\mathbf{u}}_f )_f d\tau
-
\int_0^t (\bar{\theta}_f \mathbf{k}, \mathbf{u}_f )_f d\tau
-
\int_0^t (\theta_m \mathbf{k}, \bar{\mathbf{u}}_m )_m d\tau\nonumber \\
&& -
\int_0^t (\bar{\theta}_m \mathbf{k}, \mathbf{u}_m )_m d\tau.
\label{weak1z}
\end{eqnarray}
In the weak formulation \eqref{weak3} for $\theta$, we take the test function $\phi=-\bar{\theta}$ and obtain
\begin{eqnarray}
&&-(\theta(t), \bar{\theta}(t))
-\int_0^t(\lambda(\theta)\nabla \theta,\nabla\bar{\theta})d\tau \nonumber \\
&=&-\|\theta_0\|_{L^2(\Omega)}^2
-\int_0^t(\theta,\partial_t\bar{\theta})d\tau
-\int_0^t(\theta \mathbf{u}, \nabla \bar{\theta})d\tau \nonumber \\
&=&-\|\theta_0\|_{L^2(\Omega)}^2 +\int_0^t(\lambda(\bar{\theta})\nabla \theta,\nabla\bar{\theta})d\tau
-\int_0^t(\theta \mathbf{u}, \nabla \bar{\theta})d\tau -\int_0^t(\bar{\theta} \bar{\mathbf{u}}, \nabla \theta)d\tau.
\label{weak3z}
\end{eqnarray}
Summing up the relations \eqref{eapp1d}, \eqref{eapp1e} and \eqref{weak1z}, we obtain
\begin{eqnarray}
&& \frac12\|\mathbf{u}_f(t)-\bar{\mathbf{u}}_f(t)\|_{\mathbf{L}^2(\Omega_f)}^2+
\frac{\varpi}{2}\|\mathbf{u}_m(t)-\bar{\mathbf{u}}_m(t)\|_{\mathbf{L}^2(\Omega_m)}^2
\nonumber \\
&& +2\int_0^t\big(\nu(\theta_f)\mathbb{D}(\mathbf{u}_f-\bar{\mathbf{u}}_f),\mathbb{D}(\mathbf{u}_f-\bar{\mathbf{u}}_f)\big)_f d\tau\nonumber \\
&& +\int_0^t \big( \nu(\theta_m)\mathbb{K}^{-1}(\mathbf{u}_m-\bar{\mathbf{u}}_m), (\mathbf{u}_m-\bar{\mathbf{u}}_m)\big)_m d\tau\nonumber \\
&& +\sum_{j=1}^{d-1} \int_0^t\!\int_{\Gamma_{i}}\frac{\alpha \nu(\theta_m)}{\sqrt{{\rm trace} (\mathbb{K})}} |(\mathbf{u}_f-\bar{\mathbf{u}}_f)\cdot\boldsymbol{\tau}_j|^2 dS d\tau \nonumber \\
&\leq& I_1+I_2+I_3+I_4+I_5+I_6, \label{adiff1}
\end{eqnarray}
where
\begin{eqnarray}
I_1 &=& \int_0^t \big(\mathrm{div}(\mathbf{u}_f\otimes \mathbf{u}_f), \bar{\mathbf{u}}_f\big)_fd\tau + \int_0^t \big(\mathbf{u}_f, \mathrm{div}(\bar{\mathbf{u}}_f\otimes \bar{\mathbf{u}}_f)\big)_fd\tau,\nonumber \\
I_2&=& - \int_0^t\! \int_{\Gamma_{i}} \frac12|\mathbf{u}_f|^2 (\bar{\mathbf{u}}_f\cdot \mathbf{n}_{i}) dSd\tau - \int_0^t\! \int_{\Gamma_{i}} \frac12|\bar{\mathbf{u}}_f|^2 (\mathbf{u}_f\cdot \mathbf{n}_{i}) dSd\tau,\nonumber \\
I_3 &=& \int_0^t \big((\theta_f-\bar{\theta}_f) \mathbf{k}, \mathbf{u}_f-\bar{\mathbf{u}}_f \big)_fd\tau +\int_0^t \big(\theta_m-\bar{\theta}_m) \mathbf{k}, \mathbf{u}_m- \bar{\mathbf{u}}_m \big)_m d\tau,\nonumber \\
I_4 &=& -2\int_0^t\Big((\nu(\theta_f)-\nu(\bar{\theta}_f)) (\mathbb{D}(\mathbf{u}_f)-\mathbb{D}(\bar{\mathbf{u}}_f)), \mathbb{D}(\bar{\mathbf{u}}_f)\Big)_fd\tau,\nonumber \\
I_5 &=& - \int_0^t \Big((\nu(\theta_m)-\nu(\bar{\theta}_m)) \mathbb{K}^{-1}(\mathbf{u}_m-\bar{\mathbf{u}}_m),\bar{\mathbf{u}}_m\Big)_m d\tau,\nonumber \\
I_6 &=& - \sum_{j=1}^{d-1} \int_0^t\!\int_{\Gamma_{i}}\frac{\alpha (\nu(\theta_m)- \nu(\bar{\theta}_m)) }{\sqrt{{\rm trace} (\mathbb{K})}} \big((\mathbf{u}_f-\bar{\mathbf{u}}_f)\cdot\boldsymbol{\tau}_j\big) (\bar{\mathbf{u}}_f\cdot\boldsymbol{\tau}_j)dSd\tau.\nonumber
\end{eqnarray}
Using the facts $\mathrm{div}\mathbf{u}_f=\mathrm{div}\bar{\mathbf{u}}_f=0$ and integration by parts, we have
\begin{eqnarray}
&& \big(\mathrm{div}(\mathbf{u}_f\otimes \mathbf{u}_f), \bar{\mathbf{u}}_f\big)_f + \big(\mathbf{u}_f, \mathrm{div}(\bar{\mathbf{u}}_f\otimes \bar{\mathbf{u}}_f)\big)_f\nonumber \\
&=& \int_{\Omega_f} (\mathbf{u}_f \cdot \nabla \mathbf{u}_f)\cdot \bar{\mathbf{u}}_f dx +\left[- \int_{\Omega_f} (\bar{\mathbf{u}}_f\otimes \bar{\mathbf{u}}_f):\nabla \mathbf{u}_f dx +\int_{\Gamma_i}(\bar{\mathbf{u}}_f\otimes \bar{\mathbf{u}}_f)\mathbf{n}_i\cdot \mathbf{u}_f dS \right]\nonumber \\
&& +\left[ - \int_{\Omega_f} \big[ (\mathbf{u}_f-\bar{\mathbf{u}}_f)\cdot \nabla \bar{\mathbf{u}}_f\big]\cdot \bar{\mathbf{u}}_f dx+\frac12\int_{\Gamma_i}\big[(\mathbf{u}_f-\bar{\mathbf{u}}_f)\cdot \mathbf{n}_i\big]|\bar{\mathbf{u}}_f|^2dS\right]\nonumber \\
&=& \int_{\Omega_f} \big[(\mathbf{u}_f-\bar{\mathbf{u}}_f)\cdot \nabla (\mathbf{u}_f-\bar{\mathbf{u}}_f) \big]\cdot \bar{\mathbf{u}}_f dx +\int_{\Gamma_i}(\bar{\mathbf{u}}_f\cdot \mathbf{u}_f) (\bar{\mathbf{u}}_f\cdot \mathbf{n}_i) dS\nonumber \\
&& +\frac12\int_{\Gamma_i}\big[(\mathbf{u}_f-\bar{\mathbf{u}}_f)\cdot \mathbf{n}_i\big]|\bar{\mathbf{u}}_f|^2dS.\nonumber
\end{eqnarray}
Then we can deduce that
\begin{eqnarray}
I_1+I_2 &=& \int_0^t\!\int_{\Omega_f} \big[(\mathbf{u}_f-\bar{\mathbf{u}}_f)\cdot \nabla (\mathbf{u}_f-\bar{\mathbf{u}}_f)\big]\cdot \bar{\mathbf{u}}_f dxd\tau \nonumber \\
&& - \frac12 \int_0^t\!\int_{\Gamma_i} |\mathbf{u}_f-\bar{\mathbf{u}}_f|^2(\bar{\mathbf{u}}_f\cdot\mathbf{n}_i) dSd\tau\nonumber \\
&\leq & \int_0^t \|\mathbf{u}_f-\bar{\mathbf{u}}_f\|_{\mathbf{L}^3(\Omega_f)} \|\bar{\mathbf{u}}_f\|_{\mathbf{L}^6(\Omega_f)} \|\nabla (\mathbf{u}_f-\bar{\mathbf{u}}_f)\|_{\mathbf{L}^2(\Omega_f)} d\tau\nonumber \\
&& +\frac12\int_0^t\| \mathbf{u}_f-\bar{\mathbf{u}}_f\|_{\mathbf{L}^\frac{8}{3}(\Gamma_i)}^2\|\bar{\mathbf{u}}_f\|_{\mathbf{L}^4(\Gamma_i)}d\tau \nonumber \\
&\leq & \frac{\epsilon}{2}\int_0^t\|\mathbf{u}_f-\bar{\mathbf{u}}_f\|_{\mathbf{H}^1(\Omega_f)}^2d\tau +\frac{C}{\epsilon}\int_0^t\|\bar{\mathbf{u}}_f\|_{\mathbf{H}^1(\Omega_f)}^4 \|\mathbf{u}_f-\bar{\mathbf{u}}_f\|_{\mathbf{L}^2(\Omega_f)}^2d\tau \nonumber \\
&& +C \int_0^t\| \mathbf{u}_f-\bar{\mathbf{u}}_f\|_{\mathbf{H}^\frac{1}{4}(\Gamma_i)}^2 \|\bar{\mathbf{u}}_f\|_{\mathbf{H}^\frac12 (\Gamma_i)}d\tau \nonumber \\
&\leq & \frac{\epsilon}{2}\int_0^t\|\mathbf{u}_f-\bar{\mathbf{u}}_f\|_{\mathbf{H}^1(\Omega_f)}^2d\tau +\frac{C}{\epsilon}\int_0^t\|\bar{\mathbf{u}}_f\|_{\mathbf{H}^1(\Omega_f)}^4 \|\mathbf{u}_f-\bar{\mathbf{u}}_f\|_{\mathbf{L}^2(\Omega_f)}^2d\tau \nonumber \\
&& + C \int_0^t\| \mathbf{u}_f-\bar{\mathbf{u}}_f\|_{\mathbf{H}^\frac{3}{4}(\Omega_f)}^2\|\bar{\mathbf{u}}_f\|_{\mathbf{H}^1(\Omega_f)}d\tau\nonumber \\
&\leq& \epsilon \int_0^t\|\mathbf{u}_f-\bar{\mathbf{u}}_f\|_{\mathbf{H}^1(\Omega_f)}^2 d\tau + \frac{C}{\epsilon} \int_0^t \|\bar{\mathbf{u}}_f\|_{\mathbf{H}^1(\Omega_f)}^4 \|\mathbf{u}_f-\bar{\mathbf{u}}_f\|_{\mathbf{L}^2(\Omega_f)}^2d\tau,
\label{I1I2}
\end{eqnarray}
where $\epsilon>0$ is a small constant to be chosen later.
Concerning $I_3$, by Lemma \ref{equinorml}, we have
\begin{eqnarray}
I_3&\leq& \int_0^t \|\theta_f-\bar{\theta}_f\|_{L^2(\Omega_f)}\| \mathbf{u}_f-\bar{\mathbf{u}}_f\|_{\mathbf{L}^2(\Omega_f)}+ \|\theta_m-\bar{\theta}_m\|_{L^2(\Omega_m)}\| \mathbf{u}_m- \bar{\mathbf{u}}_m \|_{\mathbf{L}^2(\Omega_m)} d\tau\nonumber \\
&\leq & \frac{\epsilon}{2}\int_0^t \left(
\| \mathbf{u}_f-\bar{\mathbf{u}}_f\|_{\mathbf{L}^2(\Omega_f)} ^2 +
\| \mathbf{u}_m- \bar{\mathbf{u}}_m \|_{\mathbf{L}^2(\Omega_m)}^2 \right) d\tau \nonumber \\
&& +\frac{1}{2\epsilon} \int_0^t \left( \|\theta_f-\bar{\theta}_f\|_{L^2(\Omega_f)}^2+ \|\theta_m-\bar{\theta}_m\|_{L^2(\Omega_m)}^2\right) d\tau\nonumber \\
&\leq& C\epsilon \int_0^t
\| \mathbf{u}_f-\bar{\mathbf{u}}_f\|_{\mathbf{Z}} ^2d\tau +\frac{1}{2\epsilon} \int_0^t \|\theta-\bar{\theta}\|_{L^2(\Omega)}^2 d\tau.
\label{I3}
\end{eqnarray}
The terms $I_4,I_5,I_6$ are due to the temperature dependent viscosity and can be estimated as follows:
\begin{eqnarray}
I_4&\leq & C\int_0^t \|\nu(\theta_f)-\nu(\bar{\theta}_f)\|_{L^3(\Omega_f)} \|\mathbb{D}(\mathbf{u}_f)-\mathbb{D}(\bar{\mathbf{u}}_f)\|_{\mathbf{L}^2(\Omega_f)} \|\mathbb{D}(\bar{\mathbf{u}}_f)\|_{\mathbf{L}^6(\Omega_f)}d\tau \nonumber \\
&\leq & C\int_0^t \|\theta_f-\bar{\theta}_f\|_{L^2(\Omega_f)}^\frac12 \|\nabla (\theta_f-\bar{\theta}_f)\|_{\mathbf{L}^2(\Omega_f)} ^\frac12 \|\mathbb{D}(\mathbf{u}_f)-\mathbb{D}(\bar{\mathbf{u}}_f)\|_{\mathbf{L}^2(\Omega_f)} \|\mathbb{D}(\bar{\mathbf{u}}_f)\|_{\mathbf{L}^6(\Omega_f)}d\tau\nonumber \\
&\leq & \epsilon \int_0^t \|\mathbb{D}(\mathbf{u}_f)-\mathbb{D}(\bar{\mathbf{u}}_f)\|_{\mathbf{L}^2(\Omega_f)}^2 d\tau + \epsilon \int_0^t \|\nabla (\theta_f-\bar{\theta}_f)\|_{\mathbf{L}^2(\Omega_f)}^2 d\tau\nonumber \\
&& + \frac{C}{\epsilon^3} \int_0^t \|\mathbb{D}(\bar{\mathbf{u}}_f)\|_{\mathbf{L}^6(\Omega_f)}^4 \|\theta_f-\bar{\theta}_f\|_{L^2(\Omega_f)}^2 d\tau.
\label{I4}
\end{eqnarray}
In a similar manner, we get
\begin{eqnarray}
I_5 &\leq & C\int_0^t \|\nu(\theta_m)-\nu(\bar{\theta}_m)\|_{L^3(\Omega_m)} \|\mathbf{u}_m-\bar{\mathbf{u}}_m\|_{\mathbf{L}^2(\Omega_m)} \|\bar{\mathbf{u}}_m\|_{\mathbf{L}^6(\Omega)}d\tau \nonumber \\
&\leq & \epsilon\int_0^t \|\mathbf{u}_m-\bar{\mathbf{u}}_m\|_{\mathbf{L}^2(\Omega_m)}^2d\tau + \epsilon \int_0^t \|\nabla (\theta_m-\bar{\theta}_m)\|_{\mathbf{L}^2(\Omega_m)}^2 d\tau\nonumber \\
&& + \frac{C}{\epsilon^3} \int_0^t \|\bar{\mathbf{u}}_m\|_{\mathbf{L}^6(\Omega_m)}^4 \|\theta_m-\bar{\theta}_m\|_{L^2(\Omega_m)}^2 d\tau.\label{I5}
\end{eqnarray}
By the trace theorem, the Gagliardo--Nirenberg inequality and Young's inequality, we deduce that
\begin{eqnarray}
I_6 &\leq & C\int_0^t \|\nu(\theta_m)- \nu(\bar{\theta}_m)\|_{L^\frac{8}{3}(\Gamma_i)} \|\mathbf{u}_f-\bar{\mathbf{u}}_f\|_{\mathbf{L}^4(\Gamma_i)} \|\bar{\mathbf{u}}_f\|_{\mathbf{L}^\frac{8}{3}(\Gamma_i)} d\tau\nonumber \\
&\leq& C \int_0^t \|\theta_m-\bar{\theta}_m\|_{H^\frac{3}{4}(\Omega_m)} \|\mathbf{u}_f-\bar{\mathbf{u}}_f\|_{\mathbf{H}^1(\Omega_f)} \|\bar{\mathbf{u}}_f\|_{\mathbf{H}^\frac{3}{4}(\Omega_f)}d\tau \nonumber \\
&\leq & \epsilon \int_0^t \|\mathbf{u}_f-\bar{\mathbf{u}}_f\|_{\mathbf{H}^1(\Omega_f)}^2 d\tau + \epsilon\int_0^t \|\theta_m-\bar{\theta}_m\|_{H^1(\Omega_m)}^2d\tau \nonumber \\
&& +\frac{C}{\epsilon^7}\sup_{0\leq t\leq T}\|\bar{\mathbf{u}}_f\|_{\mathbf{L}^2(\Omega_f)}^5 \int_0^t \|\bar{\mathbf{u}}_f\|_{\mathbf{W}^{1,6}(\Omega_f)}^3 \|\theta_m-\bar{\theta}_m\|_{L^2(\Omega_m)}^2 d\tau.\label{I6}
\end{eqnarray}
Next, we sum the relations \eqref{eapp2d}, \eqref{eapp2e} and \eqref{weak3z} to get
\begin{eqnarray}
&& \frac{1}{2}\|\theta(t)-\bar{\theta}(t)\|_{L^2(\Omega)}^2
+\int_0^t\! \int_\Omega \lambda(\theta)|\nabla(\theta-\bar{\theta})|^2 dx d\tau
\leq I_7+I_8,\label{adiff2}
\end{eqnarray}
where
\begin{eqnarray}
I_7&=& -\int_0^t \int_\Omega (\lambda(\theta)-\lambda(\bar{\theta}))\nabla (\theta-\bar{\theta})\cdot \nabla \bar{\theta}dx d\tau,\nonumber \\
I_8&=& -\int_0^t(\theta \mathbf{u}, \nabla \bar{\theta})d\tau -\int_0^t(\bar{\theta} \bar{\mathbf{u}}, \nabla \theta)d\tau.\nonumber
\end{eqnarray}
It follows from assumption (A2) that
\begin{eqnarray}
I_7&\leq& C\int_0^t\|\theta -\bar{\theta}\|_{L^4(\Omega)}\|\nabla (\theta-\bar{\theta})\|_{\mathbf{L}^2(\Omega)}\| \nabla \bar{\theta}\|_{\mathbf{L}^4(\Omega)}d\tau \nonumber \\
&\leq & C\int_0^t\|\theta -\bar{\theta}\|_{L^2(\Omega)}^\frac14\|\nabla (\theta-\bar{\theta})\|_{\mathbf{L}^2(\Omega)}^\frac{7}{4} \| \nabla \bar{\theta}\|_{\mathbf{L}^4(\Omega)}d\tau \nonumber \\
&\leq & \epsilon \int_0^t \|\nabla (\theta-\bar{\theta})\|_{\mathbf{L}^2(\Omega)}^2 d\tau + \frac{C}{\epsilon^7}\int_0^t \| \nabla \bar{\theta}\|_{\mathbf{L}^4(\Omega)}^8\|\theta -\bar{\theta}\|_{L^2(\Omega)}^2 d\tau. \label{I7}
\end{eqnarray}
Then using the incompressibility condition, integration by parts, the Gagliardo--Nirenberg inequality and Lemma \ref{equinorml}, we obtain
\begin{eqnarray}
I_8&=& -\int_0^t (\theta(\mathbf{u}-\bar{\mathbf{u}}),\nabla \bar{\theta})d\tau - \int_0^t\big[(\bar{\theta} \bar{\mathbf{u}},\nabla \theta)+(\theta\bar{\mathbf{u}},\nabla \bar{\theta})\big]d\tau\nonumber \\
&=& -\int_0^t \big((\theta-\bar{\theta})(\mathbf{u}-\bar{\mathbf{u}}),\nabla \bar{\theta}\big)d\tau
-\int_0^t \big( (\theta-\bar{\theta})\bar{\mathbf{u}},\nabla(\theta-\bar{\theta})\big)d\tau \nonumber \\
&=& -\int_0^t \big((\theta-\bar{\theta})(\mathbf{u}-\bar{\mathbf{u}}),\nabla \bar{\theta}\big)d\tau\nonumber \\
&\leq& \int_0^t \| \mathbf{u}- \bar{\mathbf{u}} \|_{\mathbf{L}^2(\Omega)} \|\theta -\bar{\theta}\|_{L^4(\Omega)}\|\nabla \bar{\theta}\|_{\mathbf{L}^4(\Omega)}d\tau \nonumber \\
&\leq & \epsilon \int_0^t \| \mathbf{u}- \bar{\mathbf{u}} \|_{\mathbf{Z}}^2 d\tau + \epsilon \int_0^t \|\nabla (\theta-\bar{\theta})\|_{\mathbf{L}^2(\Omega)}^2 d\tau\nonumber \\
&& + \frac{C}{\epsilon^7}\int_0^t \|\nabla \bar{\theta}\|_{\mathbf{L}^4(\Omega)}^8\|\theta-\bar{\theta}\|_{L^2(\Omega)}^2 d\tau.\label{I8}
\end{eqnarray}
Collecting the above estimates for $I_1,...,I_8$ and taking the constant $\epsilon$ to be sufficiently small (cf. assumptions (A1)--(A3)), we deduce from \eqref{adiff1} and \eqref{adiff2} that
\begin{eqnarray}
&& \|\mathbf{u}_f(t)-\bar{\mathbf{u}}_f(t)\|_{\mathbf{L}^2(\Omega_f)}^2+
\varpi\|\mathbf{u}_m(t)-\bar{\mathbf{u}}_m(t)\|_{\mathbf{L}^2(\Omega_m)}^2
+\|\theta(t)-\bar{\theta}(t)\|_{L^2(\Omega)}^2
\nonumber \\
&& +2\int_0^t\underline{\nu}\|\mathbb{D}(\mathbf{u}_f-\bar{\mathbf{u}}_f)\|_{\mathbf{L}^2(\Omega_f)}^2 d\tau
+\int_0^t \left( \underline{\nu}\bar{\kappa}^{-1}(\mathbf{u}_m-\bar{\mathbf{u}}_m), (\mathbf{u}_m-\bar{\mathbf{u}}_m)\right)_m d\tau\nonumber \\
&& +\sum_{j=1}^{d-1} \int_0^t\!\int_{\Gamma_{i}}\alpha \underline{\nu}\bar{\kappa}^{-\frac12} |(\mathbf{u}_f-\bar{\mathbf{u}}_f)\cdot\boldsymbol{\tau}_j|^2 dS d\tau
+\int_0^t \underline{\lambda}\|\nabla(\theta-\bar{\theta})\|_{\mathbf{L}^2(\Omega)}^2 d\tau
\nonumber \\
&\leq & C\int_0^t h(\tau)\left( \|\mathbf{u}_f-\bar{\mathbf{u}}_f\|_{\mathbf{L}^2(\Omega_f)}^2+ \|\theta-\bar{\theta}\|_{L^2(\Omega)}^2\right) d\tau,\nonumber
\end{eqnarray}
where
$$
h(t)= \|\bar{\mathbf{u}}_f(t)\|_{\mathbf{W}^{1,6}(\Omega_f)}^4 + \|\bar{\mathbf{u}}_m(t)\|_{\mathbf{L}^6(\Omega_m)}^4 + \|\nabla \bar{\theta}(t)\|_{\mathbf{L}^4(\Omega)}^8+1.
$$
By the additional regularity assumptions on $(\bar{\mathbf{u}}_f,\bar{\mathbf{u}}_m,\bar{\theta})$, i.e.,
$$\bar{\mathbf{u}}_f\in L^4(0,T;\mathbf{W}^{1,6}(\Omega_f)),\quad \bar{\mathbf{u}}_m\in L^4(0,T;\mathbf{L}^6(\Omega_m)),\quad \bar{\theta}\in L^8(0,T;W^{1,4}(\Omega)),$$
we see that $h(t)\in L^1(0,T)$. This enables us to apply Gronwall's lemma and Lemma \ref{equinorml} to conclude that
$$
\|\mathbf{u}_f(t)-\bar{\mathbf{u}}_f(t)\|_{\mathbf{L}^2(\Omega_f)}^2+
\|\theta(t)-\bar{\theta}(t)\|_{L^2(\Omega)}^2 +\int_0^t \|\mathbf{u}(\tau)-\bar{\mathbf{u}}(\tau)\|_{\mathbf{Z}}^2 d\tau =0,
$$
for a.a. $t\in [0,T]$. As a consequence, we obtain the weak-strong uniqueness result for problem \eqref{uf1}--\eqref{IBCi2} with $\varpi\geq 0$.
The proof of Theorem \ref{thmuniq} is complete. \hfill $\square$
\section*{Acknowledgments}
The first author was partially supported by NNSFC 11871159 and Guangdong Provincial Key Laboratory for Computational Science and Material Design
2019B030301001. The second author was partially supported by NNSFC 12071084 and the Shanghai Center for Mathematical Sciences.
|
\section{Introduction}
Video, patres conscripti, in me omnium vestrum ora atque oculos esse
conversos, video vos non solunn de vestro ac rei publicae, verum
etiam, si id depulsum sit, de meo periculo esse sollicitos. Est mihi
iucunda in malis et grata in dolore vestra erga me voluntas, sed eam,
per deos inmortales, deponite atque obliti salutis meae de vobis ac
de vestris liberis cogitate. Mihi si haec condicio consulatus data
est, ut omnis acerbitates, onunis dolores cruciatusque perferrem,
feram non solum fortiter, verum etiam lubenter, dum modo meis
laboribus vobis populoque Romano dignitas salusque pariatur.
\begin{figure}
\begin{center}
\includegraphics[height=4cm]{jcaesar.eps}
\caption{Gaius Julius Caesar, 100--44 B.C.}
\label{fig1}
\end{center}
\end{figure}
\subsection{A subsection}
Marcus Tullius Cicero, 106--43 B.C. was a Roman statesman, orator,
and philosopher. A major figure in the last years of the Republic,
he is best known for his orations against Catiline\footnote{
This footnote should be very brief.}
and for his mastery of Latin prose \cite{Heritage:92}. He was a
contemporary of Julius Caesar (Fig.~\ref{fig1}).
\section{The argument}
Some words might be appropriate describing equation~(\ref{e1}), if
we had but time and space enough.
\begin{equation} \label{e1}
{{\partial F}\over {\partial t}} =
D{{\partial^2 F}\over {\partial x^2}}.
\end{equation}
See \cite{Abl:56}, \cite{AbTaRu:54}, \cite{Keo:58} and
\cite{Pow:85}.
This equation goes far beyond the celebrated theorem ascribed to the great
Pythagoras by his followers.
\begin{thm}
The square of the length of the hypotenuse of a right triangle equals the sum of the squares
of the lengths of the other two sides.
\end{thm}
\begin{prop}
The square of the length of the hypotenuse of a right triangle equals the sum of the squares
of the lengths of the other two sides.
\end{prop}
\section{Epilogue}
A word or two to conclude, and this even includes some inline
maths: $R(x,t)\sim t^{-\beta}g(x/t^\alpha)\exp(-|x|/t^\alpha)$.
\begin{ack}
Partially supported by the Roman Senate.
\end{ack}
\bibliographystyle{plain}
\section{Introduction}
\label{sec:intro}
There has been a growing interest in applications where data about individuals are not accessible, instead aggregate population-level observations in the form of counts of the individuals are available~\cite{SheDie11,LuoXuZhe16}. For various reasons including measurement fidelity, privacy preservation, cost of data collection, and scalability, data is often collected as aggregates. For example, in human ensemble flow analysis, individual trajectories may not be readily accessible due to privacy concerns, but the number of individuals in a certain geographical area can typically be counted by cell phone carriers. More examples include voter turnout based on demography from census data~\cite{Kin13} and bird migration analysis~\cite{SunSheKum15}. One fundamental part in modeling such aggregate data is the training phase for estimating the model parameters. Learning the underlying individual model from aggregate observations is a challenging task since the full trajectory of each individual is not accessible.
We are interested in learning hidden Markov models (HMMs) using aggregate data.
The HMM is a popular graphical model used in various scenarios involving unobservable (hidden) data sequences arising in ecology, social dynamics, and emergence of an epidemic~\cite{RabJua86,CapMouRyd06,DonPenHel12,SinHaaZha20}. Due to their ability to address the nonstationarity in observed data sequences, HMMs are capable of modeling a rich class of problems. In aggregate HMM settings, a large set of homogeneous individuals transit from one state to another according to the underlying HMM and at each time-step, corresponding aggregated observations are recorded. For example, in epidemiology, one can model spread of an infectious disease such as COVID-19 over time in a geographical area using the population level aggregate data generated by an HMM. In this work, we consider the problem of estimating the parameters of a time-homogeneous hidden Markov model, i.e., transition and observation probabilities, from noisy aggregate data.
A traditional method for learning HMM is the Baum-Welch algorithm~\cite{BauEag67,BauPetSou70}, which is a special case of the expectation-maximization (EM) algorithm~\cite{DemLaiRub77}. For the given observations sampled from a model consisting of latent variables (variables that are not observable) with unknown parameters, the EM algorithm aims to find the maximum likelihood estimates of the model parameters. In its first step (E-step), the EM algorithm estimates a function of the expected values of the latent variables and subsequently in the second step (M-step), it finds the maximum likelihood parameter estimates. For the case of learning HMM parameters, inference algorithms such as belief propagation (BP) algorithm \cite{Pea88} is utilized in the E-step of the EM algorithm. The Baum-Welch algorithm for estimating an HMM uses the forward-backward inference algorithm, one type of BP algorithms, in the E-step to complete the data. Unfortunately, traditional HMM learning methods such as Baum-Welch algorithm~\cite{BauPetSou70} can not be applied to aggregate setting. Learning the individual model from such population-level observations becomes challenging since great amount of information about individuals is lost due to data aggregation and observation noise.
Recently, the learning and inference problems in aggregate settings have been formalized under the collective graphical model (CGM) framework~\cite{SheDie11}. Within the CGM framework, for learning the parameters of the individual model, several aggregate inference methods such as non-linear belief propagation (NLBP)~\cite{SunSheKum15} and Bethe-RDA~\cite{VilBelSheMcc15} algorithms has been utilized in the E-step of the EM algorithm aiming to maximize the complete data likelihood. Both of the inference algorithms work on an explicit observation model.
In addition, since NLBP does not exhibit convergence guarantee, it does not lead to stable learning methods.
The primary contribution of our work is a novel algorithm for estimating the HMM parameters with theoretical guarantees from noisy aggregate observations. We utilize a modified EM algorithm for the learning task, where the E-step of the algorithm is solved using recently proposed aggregate inference method, the Sinkhorn belief propagation (SBP) algorithm~\cite{SinHaaZha20}. We show that our algorithm exhibits a convergence guarantee. Moreover, our algorithm naturally reduces to the standard Baum-Welch algorithm when the observed data is based on a single individual. We further extend our algorithm to learn the model parameters with \textit{continuous} observation noise model. We evaluate the performance of our algorithm on a variety of scenarios including bird migration and human ensemble flow on real-world dataset.
\textbf{Related Work:} Estimating Markov chains from aggregate data, also referred to as \textit{macro} data in earlier works, has a long history. It was first studied in \cite{LeeJudZel70} where the transition matrices were estimated based on maximum likelihood method. In \cite{Sun75,Mac77,KalLawVol83}, the modeling of a single Markov chain was studied by maximizing the aggregate posterior. More recent learning methods from aggregate data include~\cite{LuoXuZhe16,PasFuBou12}. After the CGM framework proposed by \cite{SheDie11}, there have been a few works on learning the underlying individual model from aggregate data. The non-linear belief propagation algorithm~\cite{SunSheKum15}, a message passing type algorithm for approximate inference in CGMs, has been utilized in EM for the task of learning a Markov chain. Another existing aggregate inference algorithm utilized in the E-step of the EM algorithm is Bethe-RDA~\cite{VilBelSheMcc15} which exhibits convergence guarantees.
Finally, \cite{BerShe16} proposed a method of moments estimator for learning a Markov chain within the CGM framework. Other works along this line include estimating spatio-temporal population flow~\cite{IwaShi19} and
recurrent estimation of HMM~\cite{LyuGriDun19} from aggregate data, learning stochastic behaviour of aggregate data~\cite{MaLiuShu20}, and estimating group behavior from ensemble observations \cite{Zen19}.
The rest of the paper is organized as follows. In Section~\ref{sec:background}, we briefly discuss related background including probabilistic graphical models, collective graphical models, and the Sinkhorn belief propagation algorithm. We present our main results and algorithms in Section~\ref{sec:main_results} for discrete observations. The counterpart with continuous observations is developed in Section \ref{sec:hmm_cts}. This is followed by experimental results in Section~\ref{sec:experiments} and a concluding remark in Section~\ref{sec:conclusion}.
\section{Background}
\label{sec:background}
Our algorithmic framework for learning HMM is based on a modified version of the EM algorithm that utilizes SBP for inference over a graphical model. In this section, we present related background including probabilistic graphical models, their extension to aggregate settings, and the SBP inference algorithm.
\subsection{Probabilistic Graphical Models}
\label{subsec:pgm}
A probabilistic graphical model (PGM)~\cite{WaiJor08} represents the dependencies between a collection of random variables using a graph. Let the set of $J$ random variables be $X_1,X_2,\ldots,X_J$, where each variable takes one of the $d$ possible discrete states from ${\mathcal X}$ ($|{\mathcal X}| = d$). Consider a graph $G = (V,E)$ with the set of nodes of the graph $V$ representing random variables and the set of edges $E$ representing dependencies between the variables. Then, the joint probability of the distribution of random variables can be written as
\begin{equation}\label{eq:individual}
p({\bf x}) := p(x_1,x_2,\ldots,x_J) = \frac{1}{Z} \prod_{(i,j) \in E} \psi_{ij}(x_i,x_j),
\end{equation}
where $x_1,x_2,\ldots, x_J$ are realizations of the corresponding random variables, $\psi_{ij}$ are edge potentials, and $Z$ is a normalization constant. The hidden Markov model is a special PGM.
There are two fundamental problems in PGMs: learning and inference. The learning problem in PGMs is concerned with estimating the parameters and structure of the underlying graphical model using observation data sampled from the model. The inference problem aims to infer the statistics of the node variables given the model parameters. The inference algorithm is a key component of learning algorithms such as EM algorithm \cite{NeaHin98}.
{\bf {Belief Propagation}:}
One of the most effective inference algorithms in PGMs is belief propagation (BP)~\cite{Pea88}, which estimates the marginal distribution of each node based on the messages from the neighbors of the node.
Let $m_{i\rightarrow j} (x_j)$ be the message from variable node $i$ to variable node $j$, then
\begin{equation}\label{eq:BP}
m_{i \rightarrow j} (x_j) \propto \sum_{x_i} \psi_{ij}(x_i,x_j)
\prod_{k\in N(i)\backslash j} m_{k \rightarrow i}(x_i),
\end{equation}
where $N(i)$ is the set of neighboring nodes of $i$, and $N(i)\backslash j$ denotes the set of neighbors of $i$ except for $j$. The message in the above equation represents belief of node $i$ on the marginal of node $j$. The BP algorithm iteratively updates the messages in \eqref{eq:BP} over the graph.
The BP algorithm is guaranteed to converge globally when the underlying graph is a tree~\cite{YedFreWei01} and one can recover the true marginals exactly upon convergence as $\hat{p}_i(x_i) \propto \prod_{k \in N(i)} m_{k \rightarrow i} (x_i)$.
\textbf{Expectation-Maximization:}
The learning problem in PGMs is to estimate parameters from a dataset of $M$ independent, identically distributed (i.i.d.) training samples generated from the model. When every variable in a PGM is observable, one can utilize maximum-likelihood estimation technique for the estimation of model parameters given observed data. Suppose we are given an i.i.d. training dataset consisting of $M$ number of complete observations ${\bf x}^{(1)}, {\bf x}^{(2)},\ldots,{\bf x}^{(m)},\ldots,{\bf x}^{(M)}$, where each observation ${\bf x}^{(m)} = \{ x^{(m)}_1, x^{(m)}_2,\ldots,x^{(m)}_J \}$ contains a value assignment to all the $J$ variables in the model. Then, the maximum-likelihood method estimates the model parameters $\Psi$ (set of all the edge potentials) by maximizing the log-likelihood $L(\Psi;{\bf x}^{(1)}, {\bf x}^{(2)},\ldots,{\bf x}^{(M)}) = \sum_{m=1}^M\mathrm{log}~ p({\bf x}^{(m)} | \Psi)$. However, in practice due to economic or feasibility reasons, some variables are not observable (known as hidden or latent variables) and observations are made corresponding to only a subset of variables.
Expectation-maximization (EM)~\cite{Bis06} is a general technique for maximum-likelihood estimation of the model parameters in presence of hidden (unobservable) variables. The EM algorithm is an iterative method that involves two steps in each iteration: E-step and M-step. In the E-step, the values associated with the hidden variables are estimated to make the data complete and then, in M-step, the parameters of the underlying model are optimized based on the complete data likelihood.
Assume that only a subset of variables $\Gamma \subset V$ are observable in the given graphical model. The training dataset consists of i.i.d. observations ${\mathbf o}^{(1)}, {\mathbf o}^{(2)},\ldots, {\mathbf o}^{(M)} $, where each observation ${\mathbf o}^{(m)}= \{o^{(m)}_1, o^{(m)}_2,\ldots, o^{(m)}_{|\Gamma|} \}$ contains a value assignment to all the $|\Gamma |$ number of variables. Let the complete data log likelihood be $L(\Psi;{\bf X},{\mathbf o})$, where ${\bf X} = \{X_1,X_2,\ldots,X_{|V| - |\Gamma|} \}$ denote the set of hidden variables , ${\mathbf o} =\{{\mathbf o}^{(1)}, {\mathbf o}^{(2)},\ldots, {\mathbf o}^{(M)} \}$ represents the observed data and $\Psi$ be the unknown set of parameters to be learned. The E-step of the EM algorithm computes an auxiliary function which is the expected value of $L(\Psi;{\bf X},{\mathbf o})$ given the observed data ${\mathbf o}$ and the current estimate of parameters $\Psi^{\mathrm{old}}$:
\begin{equation}\label{eq:e_step}
Q(\Psi,\Psi^{\mathrm{old}}) = \sum_{m=1}^M \mathbb{E}_{{\bf X} | {\mathbf o}^{(m)}, \Psi^{\mathrm{old}}} \left[ L(\Psi;{\bf X},{\mathbf o}^{(m)})\right].
\end{equation}
Subsequently, in the M-step the parameters are estimated as
\begin{equation}\label{eq:m_step}
\Psi^{\mathrm{new}} = \underset{\Psi}{\mathrm{argmax}}~ Q(\Psi,\Psi^{\mathrm{old}}).
\end{equation}
The above two steps are repeated iteratively until convergence. It has been proven that the EM algorithm is guaranteed to converge at least to a local maximum. In case of estimating the parameters of a PGM, the BP algorithm is used in the E-step of the algorithm.
\subsection{Collective Graphical Models}
\label{subsec:cgm}
Collective graphical model (CGM)~\cite{SheDie11} is a framework for learning and inference from noisy aggregate data derived from a graphical model describing the behavior of individuals.
The aggregate data is generated from $M$ independent individuals following a certain individual graphical model as in \eqref{eq:individual}. Denote the state of the $m^{th}$ individual at node $i$ as random variable $X^{(m)}_i$. Let the sample vectors be ${\bf x}^{(1)},...,{\bf x}^{(M)}$ drawn from the individual probability model representing the individuals in a population, where each entry of the vector ${\bf x}^{(m)}$ corresponding to node $i$ is $x^{(m)}_i$ that takes one of the $d$ possible states. Let ${\mathbf n}_i \in \mathbb{R}^d$ be the aggregate \textit{node} distribution with entries $n_i(x_i)= \sum_{m=1}^M \mathbb{I}[X^{(m)}_i= x_i]$ that count the number of individuals in each state with $\mathbb{I}[\cdot]$ being the indicator function. Moreover, let ${\mathbf n}_{ij}$ be the aggregate edge distributions with entries $n_{ij}(x_i,x_j) = \sum_{m=1}^M \mathbb{I}[X_i^{(m)}= x_i, X_j^{(m)}= x_j]$. The vectors ${\mathbf n}_1,\ldots,{\mathbf n}_J$ constitute the aggregate data and the aggregate edge distributions ${\mathbf n}_{ij}$ represent sufficient statistics of the individual model~\cite{SheDie11}. Let ${\mathbf n}=\{{\mathbf n}_i,{\mathbf n}_{ij}\}$ denote the the aggregate node distributions ${\mathbf n}_i$ together with the aggregate edge distributions ${\mathbf n}_{ij}$. In its original form, the CGM explicitly models the observation noise as a conditional distribution $p({\mathbf y}|{\mathbf n})$ with ${\mathbf y}$ being the aggregate noisy observations.
The learning problem in CGMs is concerned with estimating the individual model parameters of the underlying graph~\footnote{We assume that the structure of the underlying graph is known.} from noisy aggregate observations ${\mathbf y}$. For learning the parameters of the model, the EM algorithm is utilized that consists of two operations: Expectation-step (computes the complete data log-likelihood) and Maximization-step (maximizes the complete data log-likelihood). The E-step of an EM algorithm for learning the individual model from aggregate data requires inferring the aggregate node distributions ${\mathbf n}$.
The goal of inference in CGMs is to estimate ${\mathbf n}$ from the aggregate noisy observations through the conditional distribution $p({\mathbf n} | {\mathbf y}) \propto p({\mathbf n})p({\mathbf y}|{\mathbf n})$, where $p({\mathbf n})$ is known as the CGM distribution~\cite{SheDie11} which is derived from the individual model~\eqref{eq:individual}. For the tree structured graph, the CGM distribution is given by
\begin{align}
p({\mathbf n}| \Psi) &= \left( M! \frac{\prod_{i\in V} \prod_{x_i}((n_i(x_i)!)^{(d_i - 1)} }{\prod_{(i,j) \in E} \prod_{x_i,x_j} n_{ij}(x_i,x_j)!} \right) \nonumber \\
& \quad ~ \left( \frac{1}{Z^M}\!\! \prod_{(i,j) \in E} \prod_{x_i,x_j}\! \psi_{ij}(x_i,x_j)^{n_{ij}(x_i,x_j)} \right),
\end{align}
where $\Psi = \{\psi_{ij}(x_i,x_j) \}$ is the set of parameters of the graphical model and $d_i$ is the number of neighbors of node $i$ in the underlying graph $G$. The first term in the above equation is a count of the number of different ordered samples contributing to the sufficient statistics ${\mathbf n}$; it does not depend on the parameters $\Psi$. The second term is the joint probability of the entire population.
Moreover, the support of the CGM distribution $p({\mathbf n})$ is such that each entry of ${\mathbf n}$ is an integer and satisfies the following constraints
\begin{equation}\label{eq:nconstraints}
\begin{aligned}
&\sum_{x_i}n_i(x_i) = M,\qquad\qquad ~\forall i \in V \\
&n_i(x_i) = \sum_{x_j} n_{ij} (x_i,x_j), \quad \forall i \in V, ~j\in N(i).
\end{aligned}
\end{equation}
The exact inference of ${\mathbf n}$ based on $p({\mathbf n}|{\mathbf y})$ is computationally intractable for large populations as the computational complexity increases very quickly with increase in the population size $M$ \cite{SheSunKumDie13}.
It was first discovered in \cite{SheSunKumDie13} that,
$-\ln p({\mathbf n}|{\mathbf y})$ can be approximated by (up to a constant addition and multiplication) the CGM free energy
\begin{equation}\label{eq:cgm_free_energy}
F_{\rm CGM}({\mathbf n}) = U_{\rm CGM}({\mathbf n}) - H_{\rm CGM}({\mathbf n}),
\end{equation}
where
\begin{align*}
U_{\rm CGM}({\mathbf n}) & = - \sum_{(i,j) \in E} \sum_{x_i,x_j} n_{ij}(x_i,x_j) ~\ln~\psi_{ij}(x_i,x_j) \\
& \quad \quad - \ln~p({\mathbf y}|{\mathbf n}),
\end{align*}
and
\begin{equation*}
\begin{aligned}
H_{\rm CGM}({\mathbf n}) = - \sum_{(i,j) \in E} \sum_{x_i,x_j} n_{ij}(x_i,x_j) \ln~n_{ij}(x_i,x_j) \\ + \sum_{i \in V} (d_i - 1) \sum_{x_i} n_i(x_i) ~\ln~n_{i}(x_i).
\end{aligned}
\end{equation*}
After relaxing the integer constraints on $n_i(x_i), n_{ij}(x_i,x_j)$ and under the assumption that the observation model $p({\mathbf y}\mid{\mathbf n})$ is log-concave, the resulting problem of minimizing $F_{\rm CGM}$ is a convex optimization problem. This is the approximate MAP \cite{SheSunKumDie13} inference problem in the CGM framework. Note that the problem size of minimizing $F_{\rm CGM}$ is independent of the population size $M$.
Some of the earlier algorithms for aggregate inference problem in CGMs include non-linear belief propagation (NLBP)~\cite{SunSheKum15} and Bethe-RDA~\cite{VilBelSheMcc15}. These algorithms are based on an explicit noise model that utilize standard BP on a modified set of model parameters. In contrast, recently proposed SBP algorithm~\cite{SinHaaZha20} consider the observation noise within the graph and exhibits convergence guarantees.
Next, we discuss the SBP algorithm in detail that we use to solve the E-step of the EM algorithm for the purpose of learning the underlying model parameters.
\subsection{Sinkhorn Belief Propagation}
\label{subsec:sbp}
Sinkhorn belief propagation~\cite{SinHaaZha20} is a recently proposed algorithm to solve the inference problem in CGMs efficiently based on Sinkhorn algorithm for multi-marginal optimal transport~\cite{Nen16,Pas12,BenCarCut15}. As compared to the other existing methods including NLBP and Bethe-RDA for aggregate inference under the CGM framework that explicitly model the observation noise, the observation noise model employed in SBP is incorporated within the underlying graph, ensuring that it reduces to standard BP in case of fixed delta observations. Moreover, SBP enjoys convergence guarantees and is faster than its counterparts.
\begin{figure}[h]
\centering
\includegraphics[scale=0.8]{figures/noise_model_general.pdf}
\caption{Aggregate observation model in SBP (shaded nodes represent aggregate noisy observations).}
\label{fig:noise_model_general}
\end{figure}
The generative model in the SBP algorithm is such that the noise model is incorporated withing the underlying graph. This can be explained via the graph depicted in Figure~\ref{fig:noise_model_general}, where the aggregate observations are made corresponding to variables $X_4$ and $X_5$, whereas rest of the variables are unobservable. More generally, let $G=(V,E)$ be the underlying CGM with joint aggregate distribution ${\mathbf n}$ and let $\Gamma\subset\{1,\dots,J\}$ be the set of indices representing observed variables such that ${\mathbf n}_i= {\mathbf y}_i$ for a given set of observations ${\mathbf y}_i$, for $i\in\Gamma$~\footnote{See \cite{SinHaaZha20} for more details on the observation model used in SBP.}.
Here, without loss of generality, it can be assumed that the observations are recorded corresponding to a subset of leaf nodes of the underlying graph $G$~\cite{SinHaaZha20}.
The goal of the aggregate inference algorithm is to find such aggregate distribution ${\mathbf n}$ that maximizes the posterior $\mathrm{log}~p({\mathbf n} \mid {\mathbf y}) \propto \mathrm{log}~p({\mathbf n},{\mathbf y})$ given model parameters $\Psi$ (set of edge potentials). The SBP algorithm employs the approximation of the log-likelihood of the CGM distribution $- \mathrm{log}~p({\mathbf n},{\mathbf y} \mid \Psi)$ by $F_{\mathrm{Bethe}}({\mathbf n},{\mathbf y} \mid \Psi)$ and solves the following
\begin{subequations}\label{eq:MOT_bethe}
\begin{eqnarray}
\min_{\{{\mathbf n}_{ij},{\mathbf n}_i \}} &&F_{\rm Bethe}({\mathbf n}, {\mathbf y} \mid \Psi) \label{eq:MOT_bethe_a}
\\
\text{s.t.} && n_i(x_i) = y_i(x_i),~ \forall i \in \Gamma \label{eq:MOT_bethe_b} \\
&& \sum_{x_j} n_{ij}(x_i,x_j) = n_i(x_i), \forall (i,j) \in E \label{eq:MOT_bethe_c} \\
&& \sum_{x_i} n_i(x_i)= 1, ~ \forall i \in V,\label{eq:MOT_bethe_d}
\end{eqnarray}
\end{subequations}
where $F_{\rm Bethe}({\mathbf n})$ is the the Bethe free energy~\cite{YedFreWei05} given by
\begin{equation}\label{eq:Bethe_energy_MOT}
\begin{aligned}
F_{\rm Bethe}({\mathbf n} ,{\mathbf y} \mid \Psi) = \sum_{i,j} \sum_{x_i,x_j} n_{ij}(x_i,x_j) \ln \frac{n_{ij}(x_i,x_j)}{\psi_{ij}(x_i,x_j)} - \\ \sum_{i=1} (d_i - 1) \sum_{x_i} n_i(x_i) \ln n_i(x_i).
\end{aligned}
\end{equation}
Moreover, \eqref{eq:MOT_bethe_b} corresponds to the aggregate observation constraints, \eqref{eq:MOT_bethe_c} are consistency constraints, and \eqref{eq:MOT_bethe_d} are the normalization constraints. All the steps of the SBP algorithm are listed in Algorithm~\ref{alg:sbp}. Note that in the step (ii) of the Algorithm~\ref{alg:sbp}, $i_{\rm next}$ is next node followed by $i$ in $\bar\Gamma$ with $\bar\Gamma$ representing a sequence of nodes such that each element in $\Gamma$ appear in $\bar \Gamma$ infinitely often. The SBP algorithm is guaranteed to converge when the underlying graph is a tree \cite{SinHaaZha20}.
\begin{algorithm}[tb]
\caption{Sinkhorn Belief Propagation (SBP)}
\label{alg:sbp}
\begin{algorithmic}
\STATE Initialize the messages $m_{i\rightarrow j} (x_j)$ to $1$, $\forall~ i,j \in E$
\WHILE{not converged}
\FOR{$i \in \bar\Gamma$}
\STATE i) Update $m_{i\rightarrow j} (x_j)$ as
\begin{align}\label{eq:is_bp_MOT_a}
m_{i\rightarrow j} (x_j) \propto & \sum_{x_i} \psi_{ij}(x_i,x_j) \frac{y_i(x_i)}{m_{j \rightarrow i} (x_i)},~j\in N(i),
\end{align}
\STATE ii) Update all the messages on the path from $i$ to $i_{\rm next}$
\begin{align}\label{eq:is_bp_MOT_b}
m_{j\rightarrow k} (x_k) \propto & \sum_{x_j} \psi_{jk}(x_j,x_k) \prod_{l \in N(j)\backslash k} m_{l \rightarrow j}(x_j),
\end{align}
\ENDFOR
\ENDWHILE
\STATE Compute the required aggregate distributions as
\begin{align*}
{n}_i (x_i) \propto & \prod_{k \in N(i)} m_{k \rightarrow i}(x_i), ~\forall i \notin \Gamma \\
{n}_{ij} (x_i,x_j) \propto & ~~\psi_{ij} (x_i,x_j) \!\!\!\! \prod_{k \in N(i) \backslash j} \!\!\!\! m_{k \rightarrow i} (x_i) \!\!\!\! \prod_{l \in N(j) \backslash i} \!\!\!\! m_{l \rightarrow j} (x_j)
\end{align*}
\end{algorithmic}
\end{algorithm}
One can interpret the expressions \eqref{eq:is_bp_MOT_a} and \eqref{eq:is_bp_MOT_b} in Algorithm~\ref{alg:sbp} as messages between nodes, similar to the standard BP algorithm. The messages \eqref{eq:is_bp_MOT_a} can be understood as a scaling step, which guarantees that the constraints \eqref{eq:MOT_bethe_b} remain satisfied.
\begin{rem}
It is worth noting the connection between collective and standard filtering problems.
In case of a single individual case ($M=1$), the aggregate distributions are concentrated at the fixed observations and subsequently, the SBP algorithm for aggregate inference coincides with the standard BP. We refer the reader to \cite{HaaSinZha2020} for more details.
\end{rem}
\section{Learning discrete HMMs}
\label{sec:main_results}
In this section, we first propose a learning algorithm called approximate EM algorithm with collective data in general graphs. We then specialize it to learning the parameters of a discrete hidden Markov model (HMM) from aggregate data.
\subsection{Aggregate Learning based on Approximate EM Algorithm}
\label{subsec:mode_maximization}
The EM algorithm~\cite{Mur12} increases the log likelihood of the observed data through maximizing \textit{expected complete data log likelihood} iteratively. However, in the aggregate setting, it becomes intractable \cite{SheSunKumDie13} to calculate expected complete data log likelihood $Q(\Psi,\Psi^{(\ell-1)})={\mathbb E}_{p({\mathbf n}|{\mathbf y}, \Psi^{(\ell-1)})} [\textrm{log}~ p({\mathbf n},{\mathbf y}| \Psi)]$ at iteration $\ell$ precisely. Fortunately, as pointed out in \cite{SheSunKumDie13}, the conditional distribution $p({\mathbf n}|{\mathbf y}, \Psi^{(\ell-1)})$ concentrates on the minimizer ${\mathbf n}^*$ of the Bethe free energy in \eqref{eq:MOT_bethe}, i.e., ${\mathbf n}^* = \underset{{\mathbf n}}{\mathrm{argmin}}~ F_{\mathrm{Bethe}}({\mathbf n}, {\mathbf y}\mid \Psi^{(\ell-1)})$. This can be explained by the large deviation theory. We refer the reader to \cite{HasRinChe19,SinHaaZha20} for more justifications on this approximation.
Subsequently, the auxiliary function is approximated as
\begin{equation} \label{eq:mode_log_likelihood}
Q(\Psi,\Psi^{(\ell-1)}) \approx - F_{\mathrm{Bethe}}({\mathbf n}^*,{\mathbf y} \mid \Psi ),
\end{equation}
with ${\mathbf n}^* = \underset{{\mathbf n}}{\mathrm{argmin}}~ F_{\mathrm{Bethe}}({\mathbf n}, {\mathbf y}\mid \Psi^{(\ell-1)})$. The approximation error vanishes as the population size $M$ increases.
Based on \eqref{eq:mode_log_likelihood}, we propose approximate EM algorithm for parameter learning in CGM as listed in Algorithm~\ref{alg:mm_cgm}.
\begin{algorithm}[ht]
\caption{Approximate EM algorithm for parameter learning in CGM}
\label{alg:mm_cgm}
\begin{algorithmic}
\STATE Initialize model parameters $\Psi^{(0)}$ arbitrarily
\FOR{$\ell = 1,2,\ldots$}
\STATE Approximate Expectation-step: Obtain ${\mathbf n}^*$ that maximizes $-F_{\mathrm{Bethe}}({\mathbf n}^*,{\mathbf y} \mid \Psi^{(\ell-1)})$ with the SBP algorithm
\STATE Maximization-step: Compute $\Psi = \mathrm{arg}~\underset{\Psi}{\mathrm{max}}~ - F_{\mathrm{Bethe}}({\mathbf n}^*,{\mathbf y} \mid \Psi )$
\STATE Update parameters $\Psi^{(\ell)} \leftarrow \Psi$
\ENDFOR
\end{algorithmic}
\end{algorithm}
The convergence and effectiveness of Algorithm \ref{alg:mm_cgm} are characterized by Theorem \ref{thm:mean-ll} below.
\begin{thm}
\label{thm:mean-ll}
The Approximate Expectation-Maximization algorithm (Algorithm \eqref{alg:mm_cgm}) converges. Moreover, the data likelihood approximately increases after each iterations.
\end{thm}
\begin{proof}
See Appendix~\ref{appendix:proof_algo2}.
\end{proof}
The approximate EM algorithm given by Algorithm~\ref{alg:mm_cgm} works on a general graph. The maximization-step may requires solving another optimization problem. Next we turn our attention to learning discrete HMMs from aggregate data, where the maximization-step has a closed-form solution.
\subsection{Aggregate Hidden Markov Models}
\label{subsec:collective_hmm}
\begin{figure}[h]
\centering
\includegraphics[scale=1]{figures/hmm_graphical_model.pdf}
\caption{A length $T$ HMM.}
\label{fig:hmm_graphical_model}
\end{figure}
An HMM is a Markov chain where the variables are not directly observable, but corresponding noisy variables are observed. Denote the unobserved variables as $X_1,X_2,\ldots$ and observed variables as $O_1,O_2,\ldots$. Assume that each hidden variable takes one of the discrete values from a finite set $\mathcal{X}_h$ and each observation variable from a finite set $\mathcal{X}_o$, where in general $|\mathcal{X}_h| \neq |\mathcal{X}_o|$. A time-homogenous HMM is parameterized by the initial distribution $\pi(X_1)$, the state transition probabilities $p(X_{t+1} \mid X_t)$, and the observation probabilities $p(O_{t} \mid X_{t})$ for each time step $t = 1,2,\ldots$. The graphical representation of a length $T$ HMM is shown in Figure~\ref{fig:hmm_graphical_model}, where $V = \{X_1,X_2,\ldots, X_T,O_1,O_2,\ldots,O_T\}$ with $\Gamma = \{O_1,O_2,\ldots,O_T\}$.
The joint distribution of an (individual) HMM with length $T$ factorizes as
\begin{align} \label{eq:HMM_distribution}
p({\bf x}, {\mathbf o}) &:= p(x_1,x_2,\ldots,x_T,o_1,o_2,\ldots,o_T ) \nonumber \\
&= \pi(x_1)~ \prod_{t=1}^{T-1} ~p(x_{t+1} \mid x_{t}) ~ \prod_{t=1}^{T}~p(o_t \mid x_t),
\end{align}
where ${\bf x} = \{x_1,x_2,\ldots,x_T \}$ and ${\mathbf o}= \{o_1,o_2,\ldots,o_T \}$ denote particular assignments to the hidden and observation variables, respectively.
Note that the model given by \eqref{eq:HMM_distribution} is a directed graphical model but it can be equivalently converted to the undirected graphical model \eqref{eq:individual} by considering the transition and observation probabilities as edge potentials $\psi_{ij}$ in \eqref{eq:individual}. Denote the set of parameters to be learned as $\Psi = \{ \pi(x_1), p(x_{t+1} \mid x_{t}), p(o_{t} \mid x_{t})\}$. The aggregate data constitute ${\mathbf n}_1, {\mathbf n}_2, \ldots, {\mathbf n}_{J}$ ($J = 2T$) and let the aggregate observation be ${\mathbf y}_i, \forall i \in \Gamma$. For collective HMM, the aggregate observation model for length of $T=3$ is depicted in Figure~\ref{fig:noise_model}, wherein aggregate data is generated based on $M$ number of individual trajectories~\footnote{Note that in our aggregate observation model, the variables corresponding to the observations are within the graphical model as compared to explicit observation model considered in original CGM framework~\cite{SheDie11}.}.
\begin{figure}[h]
\centering
\includegraphics[scale=0.8]{figures/noise_model_hmm.pdf}
\caption{Aggregate HMM observation model (shaded nodes represent aggregate noisy observations).}
\label{fig:noise_model}
\end{figure}
Let $X_t^{(m)}$ be the (unobservable) state of $m^{th}$ individual at time $t$ and $O_t^{(m)}$ be the observable state of $m^{th}$ individual at time $t$. The observations are made in the form of
$y_t(o_t) = \sum_{m=1}^M \mathbb{I}[O_t^{(m)}= o_t] = n_t^o(o_t)$. Given these aggregate observations, the SBP algorithm estimates the latent distributions $n_{t,t+1}(x_t,x_{t+1}) = \sum_{m=1}^M \mathbb{I}[X_t^{(m)}= x_t, X_{t+1}^{(m)}= x_{t+1}]$, $n_{t,t}(x_t,o_{t}) = \sum_{m=1}^M \mathbb{I}[X_t^{(m)}= x_t, O_t^{(m)}= o_t]$, and $n_t(x_t) = \sum_{m=1}^M \mathbb{I}[X_t^{(m)}= x_t]$. In case of collective HMM, the SBP algorithm is nothing but the generalization of traditional forward-backward algorithm~\cite{KolFri09} to aggregate settings, known as collective forward-backward algorithm~\cite{SinHaaZha20}, which we present next.
{\bf Collective Forward-Backward Algorithm:}
The collective forward-backward algorithm (CFB) is a special case of the general SBP algorithm (Algorithm~\ref{alg:sbp}) when the underlying graph is characterized by an HMM. Moreover, for the case of single observation trajectory ($M=1$), it reduces to standard forward-backward algorithm.
\begin{figure}[h]
\centering
\includegraphics[scale=0.8]{figures/hmm_message.pdf}
\caption{Messages for inference in collective HMM.}
\label{fig:hmm_message}
\end{figure}
The messages in collective HMM are shown in Figure~\ref{fig:hmm_message} with $\alpha_t(x_t)$ being the messages in the forward direction and $\beta_t(x_t)$ are the messages in the backward direction. Moreover, $\gamma_t(x_t)$ denote the messages from observation node to hidden node and $\xi_t(o_t)$ are the messages from hidden node to observation node. These messages are characterized by
\begin{subequations}\label{eq:forward_backward}
\begin{eqnarray}
\alpha_t(x_t) \propto \sum_{x_{t-1}} p(x_t|x_{t-1}) \alpha_{t-1} (x_{t-1}) \gamma_{t-1}(x_{t-1}) \label{eq:forward_backward1} \\
\beta_t(x_t) \propto \sum_{x_{t+1}} p(x_{t+1}|x_{t}) \beta_{t+1} (x_{t+1}) \gamma_{t+1}(x_{t+1}) \label{eq:forward_backward2} \\
\gamma_t(x_t) \propto \sum_{o_{t}} p(o_{t}|x_{t}) \frac{y_t(o_t)}{\xi_t(o_t)} \label{eq:forward_backward3} \\
\xi_t(o_t) \propto \sum_{x_{t}} p(o_{t}|x_{t}) \alpha_{t} (x_{t}) \beta_{t}(x_{t}) \label{eq:forward_backward4}
\end{eqnarray}
\end{subequations}
with boundary conditions
\begin{equation*}
\alpha_1(x_1) = \pi(x_1) \quad \text{and} \quad \beta_T(x_T) = 1.
\end{equation*}
The sequence of update steps are listed in Algorithm~\ref{alg:collective_forward_backward}.
\begin{algorithm}[h]
\caption{Collective Forward-backward algorithm}
\label{alg:collective_forward_backward}
\begin{algorithmic}
\STATE Initialize all the messages $\alpha_t(x_t), \beta_t(x_t), \gamma_t(x_t), \xi_t(o_t)$
\WHILE{not converged}
\STATE \textbf{Forward pass:}
\FOR{$t = 2,3,\ldots,T$}
\STATE i) Update $\gamma_{t-1}(x_{t-1})$
\STATE ii) Update $\alpha_t(x_t), \xi_t(o_t)$
\ENDFOR
\STATE \textbf{Backward pass:}
\FOR{$t = T-1,\ldots,1$}
\STATE i) Update $\gamma_{t+1}(x_{t+1})$
\STATE ii) Update $\beta_t(x_t), \xi_t(o_t)$
\ENDFOR
\ENDWHILE
\STATE Estimate required marginals as
\begin{subequations}\label{eq:marginals_hmm}
\begin{eqnarray}
&& n_t(x_t) \propto \alpha_t(x_t) \beta_t(x_t) \gamma_t(x_t), \label{eq:marginals_hmm_a} \\
&& n_{t,t+1}(x_t,x_{t+1}) \propto p(x_{t+1} | x_t) \alpha_t(x_t) \gamma_t(x_t) \nonumber \\ && \qquad \qquad \qquad \qquad \qquad \qquad \beta_t(x_{t+1}) \gamma_t(x_{t+1}) \label{eq:marginals_hmm_b} \\
&& n_{t,t}(x_t,o_t) \propto \frac{p(o_t | x_t) \alpha_t(x_t) \beta_t(x_t)}{\xi_t(o_t)}
\label{eq:marginals_hmm_c}
\end{eqnarray}
\end{subequations}
\end{algorithmic}
\end{algorithm}
\subsection{Learning Aggregate HMMs}
\label{subsec:learning_HMM}
Now we address the problem of parameter learning in an approximate way via EM algorithm with aggregate observation noise model described above.
Theorem \ref{thm:mean-ll} is applicable to general graph with aggregate observations. The practical implementation requires efficient E-step and M-step. In this subsection, we show there exists an efficient algorithm for aggregate HMM. We propose to implement the E-step via the SBP algorithm and derive a practical method to address the maximization-step characterized via Proposition \ref{prop:Maximization_HMM}.
\begin{prop}
\label{prop:Maximization_HMM}
The Maximization-step updates in learning aggregate HMM are given by
\begin{subequations}\label{eq:parameters}
\begin{eqnarray}
&& \pi(x_1) = n_1(x_1), \label{eq:parameters_a} \\
&& p(x_{t+1} \mid x_t) = \frac{\sum_{t=1}^{T-1} n_{t,t+1}(x_t,x_{t+1})}{\sum_{t=1}^{T-1} n_{t}(x_t)}, \label{eq:parameters_b} \\
&& p(o_{t} \mid x_t) = \frac{\sum_{t=1}^{T} n_{t,t}(x_t,o_{t})}{\sum_{t=1}^{T} n_{t}(x_t)}. \label{eq:parameters_c}
\end{eqnarray}
\end{subequations}
\end{prop}
\begin{proof}
See Appendix~\ref{appendix:proof_thm3}.
\end{proof}
\begin{rem}
If parts of the parameters are known, then we only need to update the other parameters in the maximization step. For instance, if the emission probability are known, then only the steps \eqref{eq:parameters_a}-\eqref{eq:parameters_b} are needed.
\end{rem}
Based on Proposition \ref{prop:Maximization_HMM} along with the Approximate EM algorithm (Algorithm~\ref{alg:mm_cgm}), we list the steps to learn the individual HMM parameters in Algorithm~\ref{alg:em_hmm_discrete}. Moreover, Algorithm~\ref{alg:em_hmm_discrete} converges because of Theorem~\ref{thm:mean-ll}.
\begin{algorithm}[ht]
\caption{Learning aggregate HMMs}
\label{alg:em_hmm_discrete}
\begin{algorithmic}
\STATE \textbf{Require:} Aggregate observations ${\mathbf y}_1,{\mathbf y}_2,\ldots, {\mathbf y}_T$
\STATE Initialize $\pi(x_1)$, $p(x_{t+1} \mid x_t)$, and $p(o_{t} \mid x_t)$
\REPEAT
\STATE Compute the hidden counts $n_{t,t+1}(x_t,x_{t+1}),n_t(x_t),n_{t,t}(x_t,o_{t})$ using CFB
\STATE $
\pi(x_1) = n_1(x_1)
$
\STATE $
p(x_{t+1} \mid x_t) = \frac{\sum_{t=1}^{T-1} n_{t,t+1}(x_t,x_{t+1})}{\sum_{t=1}^{T-1} n_{t}(x_t)}
$
\STATE $
p(o_{t} \mid x_t) = \frac{\sum_{t=1}^{T} n_{t,t}(x_t,o_{t})}{\sum_{t=1}^{T} n_{t}(x_t)}
$
\UNTIL{convergence}
\end{algorithmic}
\end{algorithm}
Algorithm~\ref{alg:em_hmm_discrete} is for a single sequence of aggregate data generated from a certain number of samples.
Learning from multiple sequences was initially explored in the Baum-Welch algorithm \cite{RabJua89} to improve learning process. Sharing with same merits, Algorithm~\ref{alg:em_hmm_discrete_ensemble} extends to an ensemble of $K$ number of aggregate observation sequences generated from the same HMM model. Here $K$ number of sequences are generated by aggregating trajectories of $N$ individuals such that each sequence results from $M$ individual trajectories, i.e., $K = N/M$.
\begin{algorithm}[ht]
\caption{Learning HMM From an Ensemble of Aggregate Observations}
\label{alg:em_hmm_discrete_ensemble}
\begin{algorithmic}
\STATE \textbf{Require:} Ensemble of aggregate observations ${\mathbf y}_1^k,{\mathbf y}_2^k,\ldots, {\mathbf y}_T^k$, for $k=1,2,\ldots,K$
\STATE Initialize $\pi(x_1)$, $p(x_{t+1} \mid x_t)$, and $p(o_{t} \mid x_t)$
\REPEAT
\STATE Compute unobserved distributions $n_{t,t+1}^k(x_t,x_{t+1}),~n_t^k(x_t),~n_{t,t}^k(x_t,o_{t})$ for $k=1,2,\ldots,K$ using CFB
\STATE $\pi(x_1) = \frac{1}{K}\sum_{k=1}^K n_1^k(x_1)$
\STATE $
p(x_{t+1} \mid x_t) = \frac{\sum_{k=1}^{K} \sum_{t=1}^{T-1} n_{t,t+1}^k(x_t,x_{t+1})}{\sum_{k=1}^{K} \sum_{t=1}^{T-1} n_{t}^k(x_t)}
$
\STATE $
p(o_{t} \mid x_t) = \frac{\sum_{k=1}^{K} \sum_{t=1}^{T} n_{t,t}^k(x_t,o_{t})}{\sum_{k=1}^{K} \sum_{t=1}^{T} n_{t}^k(x_t)}
$
\UNTIL{convergence}
\end{algorithmic}
\end{algorithm}
\begin{prop}\label{prop:connection_Baum_Welch}
Algorithms~\ref{alg:em_hmm_discrete} and \ref{alg:em_hmm_discrete_ensemble} reduce to the Baum-Welch algorithm when observations are from populations of size $M=1$.
\end{prop}
\begin{proof}
See Appendix~\ref{appendix:proof_prop_connection_Baum_Welch}.
\end{proof}
Next, we extend our learning framework to HMMs with continuous emission densities, more specifically, to HMMs with Gaussian emission probabilities.
\section{Learning aggregate HMMs with continuous observations}
\label{sec:hmm_cts}
Now we turn our attention to the problem of estimating the parameters of continuous HMMs from a set of continuous observation trajectories. A continuous HMM is similar to the discrete HMM except for the continuous emission densities, i.e., instead of taking values from a finite number of discrete symbols, the observations are allowed to take values from continuous $s$-dimensional observation space $\mathbb{R}^s$. In standard HMM case, the continuous observation model has been studied in~\cite{Jua85,JuaLevSon86}. In this section, we extend our learning algorithm to continuous emission densities in aggregate observation settings. We have a total of $M$ trajectories of continuous observations over a $T$ length HMM. The observations constitute $\{ o_1^{(m)}, o_2^{(m)}, \ldots , o_T^{(m)}\},~\forall m=1,2,\ldots,M$ with $o_t^{(m)} \in \mathbb{R}^s$ being the continuous observation of $m^{th}$ individual at time $t$.
Although all the observation trajectories are recorded, the assignments to the corresponding state is not known. Now that the observation space is continuous, we represent the assignments of the sample observations as $n_{t}^{(m)}(x_t)$, which is the probability that observation of the $m^{th}$ individual at time $t$, $o^{(m)}_t$, has been generated by hidden state $x_t$.
Recently, the inference problem in aggregate HMMs with continuous emission densities has been studied in~\cite{ZhaSinChe20}. It was shown that the required marginals can be estimated as (Corollary 2, \cite{ZhaSinChe20})
\begin{subequations}\label{eq:marginals_hmm_cts}
\begin{eqnarray}
&& n_t(x_t) \propto \alpha_t(x_t) \beta_t(x_t) \gamma_t(x_t), \label{eq:marginals_hmm_a_cts} \\
&& n_{t,t+1}(x_t,x_{t+1}) \propto p(x_{t+1} | x_t) \alpha_t(x_t) \gamma_t(x_t) \nonumber \\
&& \qquad \qquad \qquad \qquad \qquad \beta_t(x_{t+1}) \gamma_t(x_{t+1}) \label{eq:marginals_hmm_b_cts} \\
&& n_{t}^{(m)}(x_t)\propto \frac{p(o_t^{(m)} | x_t) \alpha_t(x_t) \beta_t(x_t)}{\xi_t(m)},
\label{eq:marginals_hmm_c_cts}
\end{eqnarray}
\end{subequations}
$\forall t =1,2,\ldots,T$, where $\alpha_t(x_t), \beta_t(x_t),$ and $\gamma_t(x_t)$ are the messages in aggregate HMMs as depicted in Figure~\ref{fig:hmm_message} that are the fixed points of the following updates
\begin{subequations}\label{eq:forward_backward_cts}
\begin{eqnarray}
\alpha_t(x_t) = \sum_{x_{t-1}} p(x_t|x_{t-1}) \alpha_{t-1} (x_{t-1}) \gamma_{t-1}(x_{t-1}), \label{eq:forward_backward_cts1} \\
\beta_t(x_t) = \sum_{x_{t+1}} p(x_{t+1}|x_{t}) \beta_{t+1} (x_{t+1}) \gamma_{t+1}(x_{t+1}), \label{eq:forward_backward_cts2} \\
\gamma_t(x_t) = \frac{1}{M} \sum_{m=1}^{M} \frac{p(o_{t}^{(m)}|x_{t})} {\xi_t(m)}, \label{eq:forward_backward_cts3} \\
\xi_t(m) = \sum_{x_{t}} p(o_{t}^{(m)}|x_{t}) \alpha_{t} (x_{t}) \beta_{t}(x_{t}), \label{eq:forward_backward_cts4}
\end{eqnarray}
\end{subequations}
with
\begin{equation*}
\alpha_1(x_1) = \pi(x_1) \quad \text{and} \quad \beta_T(x_T) = 1.
\end{equation*}
The inference estimates given by above are applicable to any general continuous emission density. Now, we derive the formulas for parameter estimation of the underlying continuous observation HMM with Gaussian emission density.
Assuming the Gaussian noise model for emission density, it takes the form
\begin{equation}
p(o_t|x_t) = \mathcal{N}(o_t; \mu(x_t), \Sigma(x_t)),
\end{equation}
i.e., each (discrete) hidden state corresponds to a single Gaussian density parameterized by mean $\mu(x_t)$ and variance $\Sigma(x_t)$. In such a model, an observation $o_t^{(m)}$ corresponding to the $m^{th}$ individual at time $t$ is nothing but a sample from one of the Gaussian densities.
\begin{algorithm}[t]
\caption{Learning aggregate Gaussian-HMMs}
\label{alg:em_hmm_cts}
\begin{algorithmic}
\STATE \textbf{Require:} Continuous observations $o_1^{(m)}, o_2^{(m)}, \ldots , o_T^{(m)},~\forall m=1,2,\ldots,M$
\STATE Initialize $\pi(x_1)$, $p(x_{t+1} \mid x_t)$, $\mu(x_{t})$, and $\Sigma(x_t)$
\REPEAT
\STATE Compute the hidden marginals $n_{t,t+1}(x_t,x_{t+1}),n_t(x_t),n_{t,t}(x_t,o_{t}^{(m)})$ using CFB
\STATE $
\pi(x_1) = n_1(x_1)
$
\STATE $
p(x_{t+1} \mid x_t) = \frac{\sum_{t=1}^{T-1} n_{t,t+1}(x_t,x_{t+1})}{\sum_{t=1}^{T-1} n_{t}(x_t)}
$
\STATE $
\mu(x_t) = \frac{\sum_{t=1}^{T} \sum_{m=1}^M n_{t}^{(m)}(x_t) ~ o_t^{(m)}} { \sum_{t=1}^{T} n_{t}(x_t) }
$
\STATE $
\Sigma(x_t) = \frac{\sum_{t=1}^{T} \sum_{m=1}^M n_{t}^{(m)}(x_t) \left( o_t^{(m)} - \mu_t \right) \left( o_t^{(m)} - \mu_t \right)' }{ \sum_{t=1}^{T} n_{t}(x_t) }
$
\UNTIL{convergence}
\end{algorithmic}
\end{algorithm}
\begin{figure*}
\centering
\begin{subfigure}{.32\textwidth}
\centering
\includegraphics[scale=0.32]{figures/node3.png}
\caption{$d=3$}
\label{fig:synthetic_d3}
\end{subfigure}
\begin{subfigure}{.32\textwidth}
\centering
\includegraphics[scale=0.32]{figures/node5.png}
\caption{$d=5$}
\label{fig:synthetic_d5}
\end{subfigure}
\begin{subfigure}{.32\textwidth}
\centering
\includegraphics[scale=0.32]{figures/node10.png}
\caption{$d=10$}
\label{fig:synthetic_d10}
\end{subfigure}
\caption{The learning curves of HMMs with discrete observation models. Curves in different color depict the results with different $M$. All three experiments share the same values of $T=5, N=5000$. The figures show how $\Delta NLL$ evolves as the number of iterations increases, for $d=3,~d=5$ and $d=10$ respectively. The shaded region represents standard deviation of $\Delta NLL$ over 10 random seeds.}
\label{fig:synthetic}
\end{figure*}
\begin{figure*}
\centering
\begin{subfigure}{.32\textwidth}
\centering
\includegraphics[scale=0.32]{figures/test_go_20.png}
\caption{$d=20$}
\label{fig:synthetic_go_d3}
\end{subfigure}
\begin{subfigure}{.32\textwidth}
\centering
\includegraphics[scale=0.32]{figures/test_go_50.png}
\caption{$d=50$}
\label{fig:synthetic_go_d5}
\end{subfigure}
\begin{subfigure}{.32\textwidth}
\centering
\includegraphics[scale=0.32]{figures/test_go_100.png}
\caption{$d=100$}
\label{fig:synthetic_go_d100}
\end{subfigure}
\caption{The learning curves of various HMMs with Gaussian observation models. Curves in different color depict the results with different $M$. All three experiments are HMMs with $T=5$ and $N=5000$. The figures show how $\Delta NLL$ evolves for $d=20,~d=50$ and $d=100$. The shaded region represents standard deviation of $\Delta NLL$ over 10 random seeds.}
\label{fig:synthetic_go}
\end{figure*}
Note that with this aggregate Gaussian-HMM settings, the estimation updates of the initial distribution and transition probabilities remain same as described in Algorithm~\ref{alg:em_hmm_discrete}. For learning the model parameters, the required marginals in the E-step of the algorithm are computed using Equation~\eqref{eq:marginals_hmm_cts} and updates in the M-step are characterized via the following.
\begin{prop}
\label{prop:Maximization_HMM_cts}
The Maximization-step updates in HMM with continuous GMM observation model take the form
\begin{subequations}\label{eq:parameters_cts}
\begin{eqnarray}
&& \pi(x_1) = n_1(x_1), \label{eq:parameters_cts_a} \\
&& p(x_{t+1} \mid x_t) = \frac{\sum_{t=1}^{T-1} n_{t,t+1}(x_t,x_{t+1})}{\sum_{t=1}^{T-1} n_{t}(x_t)}, \label{eq:parameters_cts_b} \\
&& \mu(x_t) = \frac{\sum_{t=1}^{T} \sum_{m=1}^M n_{t}^{(m)}(x_t)~ o_t^{(m)}} { \sum_{t=1}^{T} n_{t}(x_t) }, \label{eq:parameters_cts_c} \\
&& \Sigma(x_t) = \frac{\sum_{t=1}^{T} \sum_{m=1}^M n_{t}^{(m)}(x_t) \left( o_t^{(m)} - \mu_t \right) \left( o_t^{(m)} - \mu_t \right)' }{ \sum_{t=1}^{T} n_{t}(x_t) }, \nonumber \\ \label{eq:parameters_cts_d}
\end{eqnarray}
\end{subequations}
where prime denotes transpose operation.
\end{prop}
\begin{proof}
See Appendix~\ref{appendix:proof_prop_cts}.
\end{proof}
Based on Proposition~\ref{prop:Maximization_HMM_cts}, the parameters of a Gaussian-HMM are estimated using Algorithm~\ref{alg:em_hmm_cts}.
\begin{rem}
Convergence of Algorithm~\ref{alg:em_hmm_cts} follows from Theorem~\ref{thm:mean-ll}.
\end{rem}
\begin{rem}
Similar to Algorithm~\ref{alg:em_hmm_discrete_ensemble}, one can extend the learning of aggregate Gaussian-HMM to the case of an ensemble of continuous observation sequences.
\end{rem}
\begin{figure*}
\centering
\begin{subfigure}{.32\textwidth}
\centering
\includegraphics[scale=0.32]{figures/test_go_length_1.png}
\caption{$M=1$}
\label{fig:synthetic_go_T_M1}
\end{subfigure}
\begin{subfigure}{.32\textwidth}
\centering
\includegraphics[scale=0.32]{figures/test_go_length_10.png}
\caption{$M=10$}
\label{fig:synthetic_go_T_M10}
\end{subfigure}
\begin{subfigure}{.32\textwidth}
\centering
\includegraphics[scale=0.32]{figures/test_go_length_100.png}
\caption{$M=100$}
\label{fig:synthetic_go_T_M100}
\end{subfigure}
\caption{ Effect of the HMM length on the learning performance. Curves with different color correspond to different $T$ values. All three experiments are HMMs with $d=50$. With larger $T$, our aggregate learning algorithm achieve lower negative log likelihood per dimension. It also shows that aggregate learning shows the similar performance as EM algorithm~(M=1).}
\label{fig:synthetic_go_T}
\end{figure*}
\begin{figure*}
\centering
\begin{subfigure}{.32\textwidth}
\centering
\includegraphics[scale=0.32]{figures/test_go_traj_1.png}
\caption{$M=1$}
\label{fig:synthetic_go_traj_M1}
\end{subfigure}
\begin{subfigure}{.32\textwidth}
\centering
\includegraphics[scale=0.32]{figures/test_go_traj_10.png}
\caption{$M=10$}
\label{fig:synthetic_go_traj_M10}
\end{subfigure}
\begin{subfigure}{.32\textwidth}
\centering
\includegraphics[scale=0.32]{figures/test_go_traj_100.png}
\caption{$M=100$}
\label{fig:synthetic_go_traj_M100}
\end{subfigure}
\caption{Performance of aggregate learning with various data sizes. Curves with different color depict the learning curves with different data sizes $N$. The insufficient data causes overfitting to the training data. Our algorithm shows better performance with more samples available. All three experiments are HMMs with $d=50,~ T=5$.}
\label{fig:synthetic_go_traj}
\end{figure*}
\section{Experiments}
\label{sec:experiments}
To validate the efficacy of the proposed aggregate learning algorithms, we perform multiple sets of experiments on synthetic as well as real-world datasets.
\subsection{Learning HMMs with synthetic data}
We begin with synthetic data. We perform multiple sets of experiments for performance comparison of fitted time-invariant HMM models with discrete as well as continuous observations. The initial state probability is sampled from the uniform distribution over the probability simplex.
To produce the transition matrix, we firstly randomly permute rows of noised Identity matrix $\mathcal{I} + 0.05 \times \sqrt{d} \times \exp(Uniform[-1,1])$. We scale rows of the permuted matrix so that the resulting matrix is a valid conditional distribution.
To evaluate the performance on HMM with discrete observation, the emission matrix is generated in a similar way as transition matrix, but with a different random seed.
In case of HMMs with continuous observations, we consider the Gaussian emission model. For each state, the corresponding Gaussian distribution is parameterized by a random mean and variance. The mean is sampled from $Uniform[-5d,5d]$ and variance is from $Uniform[1,5]$. In continuous observation setting, the algorithm is required to estimate the initial distribution, the transition matrix and the means of Gaussian emission densities.
We generate $N$ individual trajectories from an HMM parameterized with $\Psi^*$. Each aggregate sequence consists of collective observations of $M$ independent trajectories of length $T$. After learning the parameters by $[\frac{N}{M}]$ sequences, we generate another $N$ individual trajectories for testing purpose.
We use the negative log likelihood ($NLL$) as a metric for evaluating performance of our learning algorithm. The difference of $NLL$ between the learned model $\Psi$ and ground truth $\Psi^*$ is
\begin{align}
\Delta NLL(\Psi) &= NLL(\Psi) - NLL(\Psi^*).
\end{align}
The model with learned parameters is evaluated on test datasets. Figure \ref{fig:synthetic} shows the performance of our algorithm for different values of state dimension $d$ and population size $M$ on HMMs with discrete observations. Curves in the same figure show learning performance with different values of $M$ but fixed $d,~T$, and $N$.
Figure \ref{fig:synthetic_go} demonstrates the performance of our algorithm for Gaussian observation models. It shows that our algorithm can effectively learn the generative models. Larger aggregate size corresponds to lower convergence rate, which is intuitive;
experiments with larger aggregate size lose more information about individuals. We also observe that in both discrete as well as continuous observation models, the performance of our aggregate learning algorithm degrades as model dimension increases.
To further demonstrate the scalability of our algorithm, we conduct experiments with various HMM lengths and sample sizes as depicted in Figure \ref{fig:synthetic_go_T} and Figure \ref{fig:synthetic_go_traj}, respectively.
In Figure \ref{fig:synthetic_go_T}, the curve in different color depicts the learning performance with various HMM lengths.
We observe the larger $T$ leads to better performance. This is because larger $T$ is associated with more training data.
Figure \ref{fig:synthetic_go_traj} demonstrates that the overfitting problem can be eased with more data available in aggregate learning.
\begin{figure}
\centering
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[scale=0.45]{figures/Tokyo_04.png}
\caption{$2:00$}
\label{fig:tokyo2_heat}
\end{subfigure}
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[scale=0.45]{figures/Tokyo_28.png}
\caption{$14:00$}
\label{fig:tokyo14_heat}
\end{subfigure}
\caption{ Heatmap observation of population around Tokyo city. The whole area is divided into $14\times16$ blocks.
With more people stay in a block, color inside the block becomes deeper. The underlying green curves represent main roads around the city.}
\label{fig:city_heat}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[scale=0.45]{figures/Estimated_04.png}
\end{subfigure}
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[scale=0.45]{figures/Tokyo_gt_04.png}
\end{subfigure}
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[scale=0.45]{figures/Estimated_16.png}
\end{subfigure}
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[scale=0.45]{figures/Tokyo_gt_16.png}
\end{subfigure}
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[scale=0.45]{figures/Estimated_28.png}
\end{subfigure}
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[scale=0.45]{figures/Tokyo_gt_28.png}
\end{subfigure}
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[scale=0.45]{figures/Estimated_40.png}
\caption{Estimation}
\end{subfigure}
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[scale=0.45]{figures/Tokyo_gt_40.png}
\caption{Ground Truth}
\end{subfigure}
\caption{Comparison between estimation based on our algorithm and ground truth movement. The four rows show the comparison at 2:00, 8:00, 14:00 and 20:00.
The red arrow depicts that flow between two block exceeds a threshold, 35.}
\label{fig:city}
\end{figure}
\subsection{Estimating Spatio-Temporal Population Flow}
\label{subsec:flow}
Now we evaluate our aggregate learning algorithm in estimating the population movement around Japanese city Tokyo. The dataset\footnote{Data Source: SNS-based People Flow Data, Nightley, Inc., Shibasaki \& Sekimoto Laboratory, the University of Tokyo, Micro Geo Data Forum, People Flow project, and Center for Spatial Information Science at the University of Tokyo,http://nightley.jp/archives/1954} consists of anonymous individual trajectories. We discretize the whole city area into $14\times16$ blocks with each block representing a $15 km \times 15 km$ area.
We interpolate individual trajectories every 30 minutes. A total of 6,432, 9,166, 6,822, 10,134, 6,646, 10,338 trajectories are collected respectively on July 1, July 7, October 7, October 13, December 16 and December 29 in the year 2013. We assume that the observation model suffer from a small Gaussian Noise. Moreover, with a small chance, a point in the center block can be categorized to eight neighbouring blocks incorrectly, which account for sensor inaccuracy. In Figure~\ref{fig:city_heat}, we show the aggregate observation at timestamps 2:00 and 14:00 generated from the noisy observation model.
Our task is to estimate the population flow at timestamps $2:00,~8:00,~14:00$, and $20:00$. At each timestamp, we take the observation generated in one and an half hour and construct a time-invariant HMM graph with $M=20$ to infer the movement of the population. Figure \ref{fig:city} presents the comparison between our estimation and ground truth movement at the four timestamps, from which we see that our algorithm successfully recovers the underlying movement of population with noisy aggregate observations.
\section{Conclusion}
\label{sec:conclusion}
In this paper, we proposed an algorithm for learning the parameters of a time-homogeneous HMMs from aggregate data. Our algorithm is based on a modified version of the EM algorithm, wherein we utilized the Sinkhorn belief propagation algorithm to infer the unobservable states. In contrast to the existing state-of-the-art algorithms that explicitly consider the aggregate observation noise, our algorithm employs the aggregate observation noise within the graphical model and due to which it is consistent with the standard Baum-Welch algorithm when aggregate data consists of only a single individual. Moreover, our algorithm enjoys convergence guarantees. We further extended our algorithm to incorporate continuous observations and presented estimates for Gaussian observation model. In this work, we have assumed that the HMMs are time-homogeneous, which restricts the modeling capability of the data. We plan to explore learning of time-varying HMMs in our future research.
\bibliographystyle{plain}
|
\section{Introduction}
Potential maximal cliques (PMCs) have been used to obtain large families of exact exponential time and FPT algorithms for graph problems~\cite{DBLP:journals/algorithmica/FominLMT18,DBLP:journals/siamcomp/FominTV15,korhonen_ipec,DBLP:journals/algorithmica/LiedloffMT19}.
These algorithms enumerate all PMCs of the input graph, so their time complexity is lower bounded by the number of PMCs.
In~\cite{DBLP:journals/algorithmica/FominLMT18} it was shown that graphs have $O^*(4^k)$\footnote{The $O^*(\cdot)$ notation suppresses factors polynomial in the number of vertices.} PMCs, where $k$ is the size of minimum vertex cover.
Algorithms with time complexity $O^*(4^k)$ were obtained via a corresponding PMC enumeration algorithm.
In the conclusion of~\cite{DBLP:journals/algorithmica/FominLMT18}, the authors remark that to them, the bound $O^*(4^k)$ for PMCs does ``not seem to be tight''.
In this paper we give an upper bound of $4^k + n$ for the number of PMCs.
We show that our bound is tight up to constant factors by giving a family of graphs with $\Omega(4^k)$ PMCs and $O(k^2)$ vertices.
Note that also the term $n$ is necessary, as $n$-vertex star graphs have $k=1$ and $n-1$ PMCs.
\section{Preliminaries}
The notation $S(n, k)$ denotes the Stirling numbers of the second kind, i.e., the number of ways to partition an $n$-element set into $k$ non-empty parts.
We consider graphs that are undirected and simple.
We denote the vertices of a graph $G$ with $V(G)$ and edges with $E(G)$.
The neighbors of a vertex $v$ are $N(v)$ and the neighborhood of a vertex set $X$ is $N(X) = \cup_{v \in X} N(v)$.
The closed neighborhood is denoted by $N[v] = N(v) \cup \{v\}$.
The set of vertex sets of connected components of $G$ is denoted by $\mathcal{C}(G)$.
For a vertex subset $X \subseteq V(G)$, $G[X]$ denotes the subgraph of $G$ induced by $X$ and $G \setminus X$ the subgraph of $G$ induced by $V(G) \setminus X$.
A full component of a vertex set $X \subseteq V(G)$ is a component $C \in \mathcal{C}(G \setminus X)$ such that $N(C) = X$.
A vertex subset is a PMC if it is a maximal clique in some minimal triangulation~\cite{DBLP:journals/siamcomp/BouchitteT01}.
Our proofs are based on the following characterization of PMCs.
\begin{proposition}[\cite{DBLP:journals/siamcomp/BouchitteT01}]
\label{pro:pmc}
Let $G$ be a graph.
A set $\Omega \subseteq V(G)$ is a PMC of $G$ if and only if
\begin{enumerate}
\item for each $C \in \mathcal{C}(G \setminus \Omega)$, $N(C) \subsetneq \Omega$, i.e., $\Omega$ has no full components, and
\item for each pair of distinct vertices $u,v \in \Omega$, either $(u, v) \in E(G)$ or there is a component $C \in \mathcal{C}(G \setminus \Omega)$ with $\{u,v\} \subseteq N(C)$.
\end{enumerate}
\end{proposition}
We refer to the first condition of Proposition~\ref{pro:pmc} as the \emph{no full component condition} and to the second condition as the \emph{cliquish condition}.
We call a PMC $\Omega$ of $G$ free if each of its vertices is adjacent to a vertex in $V(G) \setminus \Omega$.
\begin{lemma}
\label{lem:nonfree_ct}
If a PMC $\Omega$ is not free, then $\Omega = N[v]$ for some vertex $v$.
\end{lemma}
\begin{proof}
Let $v$ be a vertex of $\Omega$ that is not adjacent to any vertex in $V(G) \setminus \Omega$.
We have $N[v] \subseteq \Omega$ by definition and $N[v] = \Omega$ follows from the cliquish condition.
\end{proof}
In particular, the number of non-free PMCs is at most $n$, the number of vertices.
\section{Upper Bound}
Let $G$ be a graph and $V_k \subseteq V(G)$ a vertex cover of $G$ of size $k$.
We first reduce the task of upper bounding the number of PMCs of $G$ to this task in an induced supergraph $M(G, V_k)$ of $G$.
Then we show that the PMCs of $M(G, V_k)$ have a relatively simple structure that we can handle with some case analysis.
The following lemma allows us to reduce to an induced supergraph.
\begin{lemma}[\cite{DBLP:journals/tcs/BouchitteT02}]
\label{lem:ind_sub}
Let $G$ be a graph and $X \subseteq V(G)$.
The number of PMCs of $G[X]$ is at most the number of PMCs of $G$.
\end{lemma}
Now we add $2^k-1$ vertices to $G$ to get $M(G, V_k)$.
In particular, for each non-empty subset $X \subseteq V_k$ we add a vertex $M_X$ with $N(M_X) = X$.
The vertices $V_k$ form a vertex cover of $M(G, V_k)$, and we refer to them as the inner vertices of $M(G, V_k)$.
We refer to other vertices of $M(G, V_k)$ as the outer vertices.
Now we start to expose the structure of PMCs of $M(G, V_k)$.
\begin{lemma}
Let $\Omega$ be a PMC of $M(G, V_k)$.
The vertices $V_k$ intersect at most three distinct connected components of $M(G, V_k) \setminus \Omega$.
\end{lemma}
\begin{proof}
Suppose that there are four disjoint subsets $P_1, P_2, P_3, P_4$ of $V_k$, each a subset of a distinct connected component of $M(G, V_k) \setminus \Omega$.
The vertices $M_{P_1 \cup P_2}$ and $M_{P_3 \cup P_4}$ are both in $\Omega$, because otherwise $P_1$ would be connected to $P_2$ and $P_3$ to $P_4$.
However, the pair of vertices $M_{P_1 \cup P_2}$, $M_{P_3 \cup P_4}$ violates the cliquish condition because the neighborhoods of them are in different connected components of $M(G, V_k) \setminus \Omega$, which is a contradiction.
\end{proof}
Now, given a PMC $\Omega$, the inner vertices $V_k$ can be partitioned into at most four parts: the intersection $V_k \cap \Omega$, and at most three connected subsets.
We denote this partition by $P(\Omega) = \{V_k \cap \Omega, P_1, P_2, P_3\}$, or by subsets of it if some parts are missing.
Our goal is to show that given this partition, we can (almost) determine for each vertex of $M(G, V_k)$ whether it is in $\Omega$ or not.
First, fixing $V_k \cap \Omega$ determines which of the inner vertices are in $\Omega$.
Second, if the neighborhood of an outer vertex intersects multiple of the components $P_i$, then it must be in the PMC because otherwise these would not be disjoint components.
Third, we have the following lemma.
\begin{lemma}
\label{lem:subsv}
Let $\Omega$ be a PMC of $M(G, V_k)$ and $P_i \subseteq V_k \setminus \Omega$ a part of $P(\Omega)$.
If $v$ is an outer vertex of $M(G, V_k)$ with $N(v) \subseteq P_i$, then $v \notin \Omega$.
\end{lemma}
\begin{proof}
Suppose that $v \in \Omega$.
All neighbors of $v$ are in a same component of $M(G, V_k) \setminus \Omega$.
Therefore, if $\Omega$ would satisfy the cliquish condition, then $P_i$ would be a full component, which is a contradiction.
\end{proof}
By Lemma~\ref{lem:subsv}, $P(\Omega)$ determines the outer vertices $v$ with $N(v) \subseteq P_i$ for $i \in \{1,2,3\}$.
Now, to determine outer vertices $v$ with $N(v) \subseteq V_k \cap \Omega$ we assume that the PMC $\Omega$ is free.
It follows that these vertices must not be in the PMC because otherwise the PMC would subsume their neighborhood, and therefore be non-free.
What is left undetermined are outer vertices whose neighborhood intersects $V_k \cap \Omega$ and a single part $P_i$.
To finish the argument, we consider different cases based on what parts there are in $P(\Omega)$.
If $P(\Omega) = \{P_1, P_2, P_3\}$ we are already ready, and the number of such PMCs is $S(k, 3)$.
There are no PMCs with $P(\Omega) = \{P_1, P_2\}$ or $P(\Omega) = \{P_1\}$ because they would have a full component.
Also, there are no free PMCs $\Omega$ with $P(\Omega) = \{V_k \cap \Omega\}$ because the vertex $M_{V_k}$ would be forced in $\Omega$ making it non-free.
The remaining cases are that we have a non-empty part $V_k \cap \Omega$, and at least one but at most three other parts $P_i$.
\begin{lemma}
\label{lem:p1}
For a fixed partition $\{V_k \cap \Omega, P_1\}$ of $V_k$, there are at most $k$ free PMCs $\Omega$ of $M(G, V_k)$ such that $P(\Omega) = \{V_k \cap \Omega, P_1\}$.
\end{lemma}
\begin{proof}
Let $C_1$ be the component of $M(G, V_k) \setminus \Omega$ with $P_1 \subseteq C_1$.
To satisfy the no full component condition, there must at least one vertex $v \in V_k \cap \Omega$ that is not in $N(C_1)$.
Next we prove that fixing such a vertex $v$ uniquely determines $\Omega$.
Each outer vertex $u$ whose neighborhood intersects both $P_1$ and $v$ must be in $\Omega$.
Each outer vertex $u$ whose neighborhood intersects $P_1$ and $V_k \cap \Omega$ but not $v$ must not be in $\Omega$ because otherwise it would violate the cliquish condition with $v$.
\end{proof}
By Lemma~\ref{lem:p1}, there are at most $k 2^k$ free PMCs $\Omega$ that induce a partition of type $P(\Omega) = \{V_k \cap \Omega, P_1\}$.
\begin{lemma}
\label{lem:p2}
For a fixed partition $\{V_k \cap \Omega, P_1, P_2\}$ of $V_k$, there are at most $2k$ free PMCs $\Omega$ of $M(G, V_k)$ such that $P(\Omega) = \{V_k \cap \Omega, P_1, P_2\}$.
\end{lemma}
\begin{proof}
Recall that only the outer vertices whose neighborhoods intersect $V_k \cap \Omega$ and exactly one of $P_1$ and $P_2$ are undetermined.
Consider two such vertices, one whose neighborhood intersects $P_1$ and one whose neighborhood intersects $P_2$.
They both cannot be in $\Omega$ because they would violate the cliquish condition.
Now we have two options, to add only neighbors of $P_1$ in $\Omega$ or only neighbors of $P_2$.
Let's assume that we choose $P_1$ and multiply the count by two.
Now we are left with the exact same situation as in Lemma~\ref{lem:p1}, and by the same arguments we get at most $k$ PMCs.
\end{proof}
By Lemma~\ref{lem:p2}, there are at most $6k \cdot S(k, 3)$ free PMCs $\Omega$ that induce a partition of type $P(\Omega) = \{V_k \cap \Omega, P_1, P_2\}$.
\begin{lemma}
\label{lem:p3}
For a fixed partition $\{V_k \cap \Omega, P_1, P_2, P_3\}$ of $V_k$, there is at most one free PMC $\Omega$ of $M(G, V_k)$ such that $P(\Omega) = \{V_k \cap \Omega, P_1, P_2, P_3\}$.
\end{lemma}
\begin{proof}
Let $v$ be an outer vertex whose neighborhood intersects $V_k \cap \Omega$ and $P_i$.
The vertex $v$ cannot be in $\Omega$, because it would violate the cliquish condition with $M_{P_j \cup P_l}$ where $i,j,l$ are distinct.
\end{proof}
By Lemma~\ref{lem:p3}, there are at most $4 \cdot S(k, 4)$ free PMCs $\Omega$ that induce a partition of type $P(\Omega) = \{V_k \cap \Omega, P_1, P_2, P_3\}$.
\begin{theorem}
A graph with vertex cover of size $k$ has at most $4^k + n$ PMCs.
\end{theorem}
\begin{proof}
Let $G$ be a graph with vertex cover $V_k$ of size $k$.
By Lemma~\ref{lem:ind_sub} it suffices to bound the number of PMCs of $M(G, V_k)$.
The number of vertices of $M(G, V_k)$ is $n + 2^k-1$, so the number of non-free PMCs of it is at most $n + 2^k - 1$ by Lemma~\ref{lem:nonfree_ct}.
By Lemmas~\ref{lem:p1}, \ref{lem:p2}, and \ref{lem:p3}, the number of free PMCs of $M(G, V_k)$ is at most $S(k, 3) + k 2^k + 6k \cdot S(k, 3) + 4 \cdot S(k, 4)$.
By computing the values for small $k$ and observing that $4 \cdot S(k, 4) \le 4^k / 6$ we can verify that this sums up to at most $4^k + n$ in total.
\end{proof}
\section{Lower Bound}
We consider graph $G_k$, whose vertex set is $[k] \cup \binom{[k]}{2}$ consisting of the set $[k]$ of positive integers up to $k$ and the set $\binom{[k]}{2}$ of unordered pairs of positive integers up to $k$.
The edge set of $G_k$ is defined by connecting each vertex $(i,j) \in \binom{[k]}{2}$ to vertices $i$ and $j$ in $[k]$.
Therefore both $[k]$ and $\binom{[k]}{2}$ are independent sets and $[k]$ is a vertex cover of $G_k$.
We first show that $G_k$ has at least $S(k, 3) = \Omega(3^k)$ free PMCs $\Omega$ with $\Omega \cap [k] = \emptyset$.
Then we extend this bound by induction to show that for each $i \le k$, $G_k$ has at least $S(k-i, 3)$ free PMCs $\Omega$ with $\Omega \cap [k] = [i]$.
By symmetry and binomial theorem, this implies the lower bound $\Omega(4^k)$.
\begin{lemma}
\label{lem:base}
The graph $G_k$ has at least $S(k, 3)$ free PMCs $\Omega$ with $\Omega \cap [k] = \emptyset$.
\end{lemma}
\begin{proof}
Consider a tripartition $P = \{P_1,P_2,P_3\}$ of $[k]$.
Construct a PMC $\Omega$ such that $(i, j) \in \Omega$ if and only if $i$ and $j$ are in different parts of $P$.
Note that from such $\Omega$ we can uniquely recover $P$ because vertices in $\binom{[k]}{2} \setminus \Omega$ are grouped into three components.
Now it suffices to prove that $\Omega$ is indeed a PMC, and it follows that the number of such PMCs is at least the number of tripartitions, i.e., $S(k, 3)$.
The graph $G_k \setminus \Omega$ has exactly tree connected components, each consisting of vertices $P_i \cup \binom{P_i}{2}$ for $i \in \{1,2,3\}$.
None of the components are full, because for the component $P_i \cup \binom{P_i}{2}$ each vertex of form $v \in P_j \times P_l$ (with $i,j,l$ distinct) is in $\Omega$ but not in the neighborhood of $P_i \cup \binom{P_i}{2}$.
Each vertex $v \in \Omega$ has exactly two adjacent components because it is of form $v \in P_i \times P_j$ (with $i \neq j$).
Therefore $\Omega$ is free, and any pair of vertices in $\Omega$ has a common adjacent component so $\Omega$ satisfies the cliquish condition.
\end{proof}
Let $\Pi_f(G_k, i)$ denote the set of free PMCs $\Omega$ of $G_k$ with $\Omega \cap [k] = [i]$.
By Lemma~\ref{lem:base} we have $|\Pi_f(G_k, 0)| \ge S(k, 3)$.
\begin{lemma}
\label{lem:ind}
It holds that $|\Pi_f(G_k, i)| \ge |\Pi_f(G_{k-1}, i-1)|$.
\end{lemma}
\begin{proof}
Let $\Omega'$ be a free PMC of $G_{k-1}$ with $\Omega' \cap [k-1] = [i-1]$.
We claim that $\Omega = \Omega' \cup \{k\}$ is a free PMC of $G_k$.
The inequality follows from this claim by permuting the indices appropriately.
First, note that $G_k \setminus \Omega$ can be obtained from $G_{k-1} \setminus \Omega'$ by adding the vertices $(j, k)$ for $j<k$.
These vertices have degree one in $G_k \setminus \Omega$, so their addition can not create a full component, so $\Omega$ does not have a full component in $G_k$ because $\Omega'$ does not have a full component in $G_{k-1}$.
For the cliquish condition, the vertex $k$ is connected to each vertex in $[i-1]$ via the new vertices $(j,k)$.
Other vertices in $\Omega'$ are of form $(a,b)$, but either $a \notin \Omega'$ or $b \notin \Omega'$ because $\Omega'$ is free, so $k$ can reach them via either $(a, k)$ or $(b,k)$.
Finally, $\Omega$ is free because $(1, k) \notin \Omega$.
\end{proof}
\begin{theorem}
The number of PMCs of $G_k$ is $\Omega(4^k)$.
\end{theorem}
\begin{proof}
By combining Lemmas~\ref{lem:base} and~\ref{lem:ind}, we get that $|\Pi_f(G_k, i)| \ge S(k-i, 3)$.
By symmetry, the total number of PMCs of $G_k$ is at least $$\sum_{i=0}^k \binom{k}{i} |\Pi_f(G_k, i)| \ge \sum_{i=0}^k \binom{k}{i} S(k-i, 3) \ge 3^{-3} \sum_{i=0}^{k-3} \binom{k}{i} 3^{k-i} = \Omega(4^k).$$
\end{proof}
\section{Conclusion}
We showed that a graph with $n$ vertices and vertex cover of size $k$ has at most $4^k + n$ PMCs.
Furthermore, this bound is tight up to constant factors.
Our result refines the upper bound $O^*(4^k)$ given in~\cite{DBLP:journals/algorithmica/FominLMT18} and closes the open problem of giving a lower bound.
We hope that in addition to showing limitations of the PMC approach, our proofs could yield positive insights on PMCs.
Our proofs show that in the worst case, the vertex cover is an independent set, and that intuitively, the more edges there are within the vertex cover, the less PMCs there are.
This seems to generalize the graph class of split graphs, which have a clique vertex cover and at most $n$ PMCs.
\bibliographystyle{plain}
|
\section{Introduction}
The simple exclusion process is a fundamental interacting particle system obtained by adding a site-exclusion interaction to multiple random walks \cite{KL}. We assume here that particles lie on a random locally finite subset of ${\ensuremath{\mathbb R}} ^d$ (simple point process) and allow the jump probability rates to be random as well, but symmetric (i.e. they do not depend on the orientation of the jump).
We require that the law of the environment is stationary and ergodic w.r.t. the action of a group ${\ensuremath{\mathbb G}} $ of ${\ensuremath{\mathbb R}} ^d$-translations, ${\ensuremath{\mathbb G}} $ being the full group of translations or a subgroup isomorphic to ${\ensuremath{\mathbb Z}} ^d$.
Under weak second moment assumptions on the jump rates and a percolation assumption
assuring the existence of the process,
we then prove for almost all environments that the
simple exclusion process admits a hydrodynamic limit (HL) in path space with hydrodynamic equation $\partial_t \rho = \nabla\cdot ( D \nabla \rho)$, $D$ being the non random effective homogenized matrix associated to a single random walk ($D$ can also be degenerate).
The above result (stated in Theorem \ref{teo1} in Section \ref{HL_EP}) covers a very large class of
simple exclusion processes in symmetric random environments, e.g. those obtained by adding a site-exclusion interaction to random walks on $\mathbb{Z}^d$ and on general crystal lattices with random (possibly arbitrarily long) conductances, to random walks performing a Mott variable range hopping, to simple random walks on Delaunay triangulations or on supercritical percolation clusters
(in Section \ref{sec_esa} we discuss some examples).
In Section \ref{cima} we provide a brief presentation of
our class of models and main results, without insisting on technicalities (faced in the subsequent sections). We discuss below how the present work relates with the existing literature, the strategy we have followed and the most original aspects of our contribution.
Given a realization of the environment the resulting simple exclusion process is non--gradient.
The usual derivation of the HL for non-gradient interacting particle systems based on the method introduced by Varadhan and further developed by Quastel (cf.
\cite{KL,Q1,V}) is very technical. It becomes even harder in the disordered case (cf. \cite{FMar,Q2}). On the other hand, for disordered simple exclusion processes with symmetric jump rates one can try to avoid the non-gradient machinery by exploiting
some \emph{duality property} between the particle system and the single random walk
and some \emph{averaging property} of the single random walk.
This was first realized by K.~Nagy in \cite{N} for the simple exclusion process on ${\ensuremath{\mathbb Z}} $ with symmetric random jump rates. Nagy's analysis had two main ingredients: a representation of the exclusion process in terms of compensated Poisson processes and the Markov semigroup of the random walk
(see \cite[Eq.~(12), (13)]{N} and a quenched CLT for the random walk uniformly in the starting point (see \cite[Theorem~1]{N}). Nagy's representation (coming from duality) has been further generalize in \cite{F0,F1} and in \cite{F1} we showed that Nagy's second ingredient can be replaced but a suitable homogenization result of the $L^2$-Markov semigroup of the random walk. The advantage comes from the fact that homogenization requires much weaker assumptions than quenched CLT's (moreover, it is also more natural from a physical viewpoint: the light bulb turns on because of the motion of many electrons and not of a single one).
One advantage of the approach based on Nagy's representation and homogenization is that one can prove the HL without proving the uniqueness of the weak solution of the Cauchy problem associated to the hydrodynamic limit. On the other hand, one gets the HL at a fixed macroscopic time (in the form usually stated e.g. in \cite{KL}) but not in path space.
To gain the HL in path space, one has to prove the tightness of the empirical measure. This has been
achieved in \cite{GJ} by developing the method of corrected empirical measure (initially introduced in \cite{JL}). This method again relies on duality and on homogenization property of the resolvent of the random walk.
Once proved the tightness one can proceed in two ways. If a uniqueness result for the Cauchy problem is available, one can try to push further the analysis of the corrected empirical measure and characterize all limit points of the empirical measures as in \cite{GJ}. Otherwise, one can try to extend Nagy's representation
and use homogenization (or some averaging, in general) to get the HL for a fixed time, avoiding results of uniqueness. This has revealed useful e.g. for the subdiffusive system considered in \cite{FJL}, where a quenched CLT for varying and converging initial
points was used instead of homogenization.
Of course, the above strategies have been developed in specific contexts and not in full generality. The applications to other models require some work, already in the choice of the right functional spaces and topologies.
In our proof we used the corrected empirical measure and homogenization to prove tightness. To proceed we have presented the two independent routes: by proving uniqueness for the Cauchy problem in weak form we characterize the limit points of the empirical measure continuing to work with the corrected one; alternatively we prove in Appendix \ref{sec_passetto} Nagy's representation in our context and use homogenization to get the HL at a fixed time.
We comment now how our result differs from the previous contributions concerning the diffusive HL of simple exclusion processes in symmetric environments. The main novelty is the huge
class of models for which the HL has been proved. In particular, (i) we go beyond the lattice (${\ensuremath{\mathbb Z}} ^d$ or toroidal) structure and deal with a very broad range of random environments including geometrically amorphous ones, (ii) our assumptions on the jump rates are minimal and given by 2nd moment assumptions, (iii) we remove ellipticity conditions on the jump rates and treat also the case of degenerate effective homogenized matrix $D$, (iv) the jump range can be unbounded.
Concerning Item (i) we point out that
to gain such a generality we have used the theory of ${\ensuremath{\mathbb G}} $--stationary random measures where ${\ensuremath{\mathbb G}} ={\ensuremath{\mathbb R}} ^d$, ${\ensuremath{\mathbb Z}} ^d$ (cf. \cite{Ge,Km}).
This also allows to describe the ergodicity properties of the environment in terms the Palm distribution. To achieve the HL in great generality we needed the same for the homogenization results. This part, which has also an independent interest, has been presented in the companion work \cite{Fhom3}, where our homogenization analysis is based on 2-scale convergence. Although \cite{Fhom3} has been preliminary to the present one, here we have kept the presentation self-contained.
For completeness, we point out that Theorem \ref{teo1}
includes also as very special cases the HL in \cite{F1}, \cite{N} and \cite{RSS} (for the part concerning non-dynamical random environments in \cite{RSS}).
We recall that in \cite{RSS} the authors prove the HL for the random conductance model on ${\ensuremath{\mathbb Z}} ^d$ with possibly time-dependent random conductances in a given interval $[a,b]$, with $0<a<b<+\infty$.
Finally, we point out that this work takes origin from our previous unpublished notes \cite{Fhom1}.
\smallskip
{\bf Outline of the paper}. In Section \ref{cima} we give a non-technical presentation of setting and results. In Section \ref{flauto} we present more precisely our setting and basic assumptions for the single random walk. In Section \ref{HL_EP} we state our HL (see Theorem \ref{teo1}).
In Section \ref{sec_esa} we discuss some examples. In Section \ref{figlio_stress} we recall the homogenization results from \cite{Fhom3} used in the proof of Theorem \ref{teo1}. In Section \ref{sec_GC} we present the graphical construction of the simple exclusion process and analyze its Markov semigroup. In Section \ref{dualino} we collect some results concerning duality.
In Section \ref{sec_mammina} we recall some properties of the space $\ensuremath{\mathcal M}$ of Radon measures on ${\ensuremath{\mathbb R}} ^d$ and of the Skohorod space $D([0,T],\ensuremath{\mathcal M})$ and show the uniqueness of the weak solution of the Cauchy problem. In Section \ref{tipicone} we study the family of typical environments, for which
the HL will be proved.
In Section \ref{ida} we prove Theorem \ref{teo1}. Appendix \ref{app_localino} concerns the proof of Proposition \ref{prop_SEP}. In Appendix \ref{sec_passetto} we give an independent proof of the HL for fixed times by proving Nagy's representation in our context and by using homogenization.
\section{Overview}\label{cima}
In this section we give a brief presentation of our context and results postponing a detailed discussion to Sections \ref{flauto} and \ref{HL_EP}.
Not surprisingly, this story starts with a probability space $(\O, \ensuremath{\mathcal F}, \ensuremath{\mathcal P})$. Here the other characters: the group ${\ensuremath{\mathbb G}} $ acting on the probability space and acting by translations on ${\ensuremath{\mathbb R}} ^d$, a simple point process and a family of jump probability rates.
The group ${\ensuremath{\mathbb G}} $ can be ${\ensuremath{\mathbb R}} ^d$ or ${\ensuremath{\mathbb Z}} ^d$ (the former endowed with the Euclidean distance, the latter with the discrete topology). ${\ensuremath{\mathbb G}} $ is a measurable space endowed with the Borel $\s$--algebra and it acts on $(\O, \ensuremath{\mathcal F}, \ensuremath{\mathcal P})$
by
a family of maps $(\theta_g)_{g\in {\ensuremath{\mathbb G}} }$, with $\theta_g :\O\to \O$, such that
\begin{equation}\label{eq_azione}
\begin{cases}
\theta_0=\mathds{1},\\
\theta _g \circ \theta _{g'}= \theta_{g+g'} \text{ for all } g,g'\in {\ensuremath{\mathbb G}} ,\\
\text{the map ${\ensuremath{\mathbb G}} \times \O \ni (g,\o) \mapsto \theta _g \o \in \O$ is measurable.}
\end{cases}
\end{equation}
The group ${\ensuremath{\mathbb G}} $ acts also on the space ${\ensuremath{\mathbb R}} ^d$ by translations. We denote its action by $(\t_g)_{g\in {\ensuremath{\mathbb G}} }$, where $\t_g:{\ensuremath{\mathbb R}} ^d\to {\ensuremath{\mathbb R}} ^d$ is given by
\begin{equation}\label{trasferta}
\t_g x=x+ g_1v_1 + \cdots +g_d v_d\;, \qquad g=(g_1,\dots, g_d) \in {\ensuremath{\mathbb G}} \,,
\end{equation}
for a fixed basis $v_1, \dots, v_d$ of ${\ensuremath{\mathbb R}} ^d$.
For many applications, $\t_g x= x+g$. When dealing with processes on general lattices (as e.g. the triangular or hexagonal lattice on ${\ensuremath{\mathbb R}} ^2$), the general form \eqref{trasferta} is more suited (see Section \ref{sec_esa}).
We assume to have a simple point process on ${\ensuremath{\mathbb R}} ^d$ defined on our probability space. In particular, to each $\o \in \O$ we associate a locally finite subset $\hat \o \subset {\ensuremath{\mathbb R}} ^d$ by a measurable map $\O \ni \o \to \hat \o \in \ensuremath{\mathcal N} $. Above, $\ensuremath{\mathcal N}$ is the measurable space of locally finite subsets of ${\ensuremath{\mathbb R}} ^d$ with $\s$--algebra generated by the sets $\{ |\hat \o \cap A|=n\}$, where $A\subset {\ensuremath{\mathbb R}} ^d$ is Borel and $n \in {\ensuremath{\mathbb N}} $ (cf. \cite{DV}). As discussed in \cite{DV} one can introduce a metric $d$ on $\ensuremath{\mathcal N}$ such that the above $\s$--algebra equals the Borel $\s$--algebra.
Finally, we fix a measurable function
\begin{equation}\label{ciccino} c: \O\times {\ensuremath{\mathbb R}} ^d \times {\ensuremath{\mathbb R}} ^d \ni (\o, x, y)\mapsto c_{x,y}(\o) \in [0,+\infty)\,,\end{equation}
symmetric in $x,y$: $c_{x,y}(\o)= c_{y,x}(\o)$.
As it will be clear below, only the value of $ c_{x,y}(\o)$ with $x\not =y$ in $ \hat \o$ will be relevant. Hence,
without loss of generality, we take
\begin{equation}\label{cicciobello}
c_{x,x}(\o)\equiv 0 \text{ and } c_{x,y}(\o)\equiv 0 \text{ if }\{x,y\}\not \subset \hat \o\,.
\end{equation}
All the above objects are related by ${\ensuremath{\mathbb G}} $-invariance. As detailed in Section \ref{flauto}, we assume that $\ensuremath{\mathcal P}$ is stationary and ergodic for the action $(\theta_g)_{g\in {\ensuremath{\mathbb G}} }$. We recall that stationarity means that $\ensuremath{\mathcal P} \circ\theta_g^{-1}=\ensuremath{\mathcal P}$ for all $g\in {\ensuremath{\mathbb G}} $, while ergodicity means that $\ensuremath{\mathcal P}(A)=1$ for all translation invariant sets\footnote{Strictly speaking, one should call $A$ ${\ensuremath{\mathbb G}} $-invariant, but we can identity ${\ensuremath{\mathbb G}} $ with a subset of Euclidean translations by \eqref{trasferta}.} $A \in \ensuremath{\mathcal B}(\O)$ (i.e. such that
$\theta_g A= A$ for all $g\in {\ensuremath{\mathbb G}} $).
We also assume that, for $\ensuremath{\mathcal P}$--a.a. $\o\in \O$ and for all $g\in {\ensuremath{\mathbb G}} $, it holds
\begin{align}
&\widehat{\theta_g\o} =\t_{-g} ( \hat{\o}) \,,\label{kiwi1}\\
& c_{x,y} (\theta_g\o)= c_{\t_g x, \t_g y} (\o) \qquad \forall x,y \in {\ensuremath{\mathbb R}} ^d \,.
\end{align}
The minus sign in \eqref{kiwi1} could appear ugly, but indeed if one identifies $\hat \o$ with the counting measure $\mu_\o (A):= \sharp ( \hat \o \cap A)$, one would restate \eqref{kiwi1} as $\mu_{\theta_g\o} (A)= \mu_\o ( \t_g A)$ for all $A\subset {\ensuremath{\mathbb R}} ^d$ Borel.
Given the environment $\o$, we will introduce by the standard graphical construction the simple exclusion process on $\hat\o$ with probability
rate $c_{x,y}(\o)$ for a jump between $x$ and $y$ when the exclusion constraint is satisfied. As discussed in Section \ref{sec_GC} this simple exclusion process is a Feller process whose Markov semigroup on $C(\{0,1\}^{\hat \o})$ has infinitesimal generator $\ensuremath{\mathcal L}_\o$
acting on local functions as
\begin{equation}\label{mammaE}
\ensuremath{\mathcal L}_{\o } f(\eta) = \sum_{x\in \hat \o} \sum_{y\in \hat \o}c_{x,y}(\o) \eta(x) \bigl( 1- \eta(y)\bigr) \left[ f( \eta ^{x,y})- f(\eta)\right]\,,\;\; \eta \in \{0,1\}^{\hat \o}\,.
\end{equation}
Above and in what follows, $\{0,1\}^{\hat \o}$ is endowed with the product topology and $C(\{0,1\}^{\hat \o})$
denotes the space of continuous functions on $\{0,1\}^{\hat \o}$ endowed with the uniform topology.
We recall that a function $f$ on $\{0,1\}^{\hat \o}$ is called local if $f(\eta)$ depends on $\eta$ only through $\eta(x)$ with $x$ varying among a finite set.
The configuration $\eta^{x,y}$ is obtained from $\eta$ by exchanging the occupation variables at $x$ and $y$, i.e.
\begin{equation}\label{furia}
\eta^{x,y}(z)=\begin{cases}
\eta(y) & \text{ if } z=x\,,\\
\eta(x) & \text{ if } z=y\,,\\
\eta(z) & \text{ otherwise}\,.
\end{cases}
\end{equation}
The generator $\ensuremath{\mathcal L}_\o$ given in \eqref{mammaE} can be thought of as
an exchange operator:
\begin{equation}\label{mahmood}
\ensuremath{\mathcal L}_\o f (\eta) = \sum _{\{x,y\}\subset \hat \o} c_{x,y}(\o) \bigl[ f(\eta^{x,y})-f(\eta) \bigr]\,.
\end{equation}
When the starting configuration is given by a single particle, the dynamics reduces to a random walk in random environment, denoted by
$X_t^\o$. In Sections \ref{flauto} and \ref{HL_EP}
we will fix basic assumptions assuring the existence of the above processes for all times for $\ensuremath{\mathcal P}$--a.a. $\o$.
\smallskip
We can now present the content of our
Theorem \ref{teo1} (see Section \ref{HL_EP}), in which we show that, under suitable weak assumptions, for $\ensuremath{\mathcal P}$--a.a. environments $\o$ the above simple exclusion process admits a hydrodynamic limit under diffusive rescaling. More precisely, for $\ensuremath{\mathcal P}$-a.a. $\o$ the following holds.
Fix an initial macroscopic profile given by a Borel function $\rho_0 :{\ensuremath{\mathbb R}} ^d \to [0,1]$.
Suppose that for any $\e>0$ the simple exclusion process starts with an initial distribution $\mathfrak{m}_\e$ such that
\begin{equation*}
\lim_{\e\downarrow 0} \mathfrak{m}_\e\Big( \Big| \e^d \sum_{x \in \hat \o} \varphi (\e x) \eta(x) -\int _{{\ensuremath{\mathbb R}} ^d} \varphi(x) \rho_0(x) dx \Big|>\e\Big)=0 \qquad \forall \varphi \in C_c({\ensuremath{\mathbb R}} ^d)\,.
\end{equation*}
Call ${\ensuremath{\mathbb P}} _{ \o,\mathfrak{m}_\e }^\e$ the law of the exclusion process on $ \hat \o$ with initial distribution $\mathfrak{m}_\e$ and generator $\e^{-2} \ensuremath{\mathcal L}_ \o$. Call $m\in(0,+\infty)$ the mean density of the point process $\hat \o$. Then for all $T>0$ one has
\begin{equation*}\lim_{\e\downarrow 0}
{\ensuremath{\mathbb P}} ^\e _{ \o,\mathfrak{m}_\e } \Big(\sup_{0\leq t\leq T} \Big| \e^d \sum_{x \in \hat \o} \varphi (\e x) \eta_t( x) - \int _{{\ensuremath{\mathbb R}} ^d} \varphi(x) \rho(x,t) dx\Big| >\d
\Big)=0 \; \;\forall \varphi \in C_c({\ensuremath{\mathbb R}} ^d)\,,
\end{equation*}
where $\rho:{\ensuremath{\mathbb R}} ^d\times[0,\infty)\to {\ensuremath{\mathbb R}} $ is given by $\rho(x ,t ): =P_t \rho_0(x) $, $(P_t)_{t\geq 0}$ being the Markov semigroup on bounded measurable functions of the Brownian motion with diffusion matrix $2D$. Above $D$ is the so called \emph{effective homogenized matrix}.
We recall that, by ergodicity (cf.~Proposition \ref{prop_ergodico}), the mean density $m$ (defined in \eqref{puffo1}, \eqref{puffo2} and \eqref{zazzera}) satisfies $m=\lim _{\ell \uparrow \infty} \sharp ( \hat \o \cap [-\ell,\ell]^d)/(2\ell)^d $ $\ensuremath{\mathcal P}$--a.s.
The matrix $D$ is a $d\times d$ symmetric non-negative matrix, admitting a variational characterization in terms of the Palm distribution $\ensuremath{\mathcal P}_0$ associated to $\ensuremath{\mathcal P}$ (cf.~Definition \ref{def_D}). $D$ is related to the homogenization properties of the diffusively rescaled random walk $\e X^\o _{ \e^{-2} t} $ on $\e \hat \o$ as discussed in \cite{Fhom3}.
Some of these properties are collected in Proposition \ref{replay}.
\section{Basic assumptions and homogenization}\label{flauto}
In this section we describe our setting and our basic assumptions for the single random walk $X_t^\o$ (hence the site-exclusion interaction does not appear here). The context is the same of \cite{Fhom3} with the simplification that the jump rates are symmetric,
hence the counting measure on $\hat \o$ is reversible for $X_t^\o$.
We first fix
some basic notation.
We denote by $e_1, \dots, e_d$ the canonical basis of ${\ensuremath{\mathbb R}} ^d$, by $\ell (A)$ the Lebesgue measure of the Borel set $A\subset {\ensuremath{\mathbb R}} ^d$, by $a \cdot b$ the standard scalar product of $a,b\in {\ensuremath{\mathbb R}} ^d$.
Given a topological space $W$, without further mention, $W$ will be thought of as a measurable space endowed with the $\s$--algebra $\ensuremath{\mathcal B}(W) $ of its Borel subsets.
$\ensuremath{\mathcal N}$ is the space of locally finite subset $\{x_i\}$ of ${\ensuremath{\mathbb R}} ^d$. $\ensuremath{\mathcal N}$ is endowed with a metric such that the Borel $\s$--algebra $\ensuremath{\mathcal B}(\ensuremath{\mathcal N})$ is generated by the sets $\{ |\hat \o \cap A|=n\}$, where $A\in \ensuremath{\mathcal B}( {\ensuremath{\mathbb R}} ^d)$ and $n \in {\ensuremath{\mathbb N}} $ (cf. \cite{DV}).
\smallskip
Recall that ${\ensuremath{\mathbb G}} $ acts on the probability space $(\O, \ensuremath{\mathcal F}, \ensuremath{\mathcal P})$ by $(\theta_g)_{g\in {\ensuremath{\mathbb G}} }$ (see \eqref{eq_azione}) and that $\ensuremath{\mathcal P}$ is assumed to be stationary and ergodic for this action. Moreover, ${\ensuremath{\mathbb G}} $ acts
on ${\ensuremath{\mathbb R}} ^d$ by $(\t_g)_{g\in {\ensuremath{\mathbb G}} }$, where (cf. \eqref{trasferta})
\begin{equation}\label{trasferta2}
\t_g x= x+Vg\,, \qquad V:= [ v_1|v_2|\cdots|v_d]\,.
\end{equation}
Above, $V$ is the matrix with columns given by the basis vectors $v_1,v_2, \dots, v_d$, fixed once and for all.
\smallskip
We set
\begin{equation}\label{simplesso}
\D:= \{t_1 v_1 + \cdots + t_d v_d \,:\, (t_1,\cdots, t_d) \in [0,1)^d\}\,.
\end{equation}
Given $x\in {\ensuremath{\mathbb R}} ^d$, the \emph{${\ensuremath{\mathbb G}} $--orbit} of $x$ is defined as the set $\{\t_g x\,:\, g\in {\ensuremath{\mathbb G}} \}$.
If ${\ensuremath{\mathbb G}} ={\ensuremath{\mathbb R}} ^d$, then the orbit of the origin of ${\ensuremath{\mathbb R}} ^d$ equals ${\ensuremath{\mathbb R}} ^d$ and in this case we set
\begin{equation}\label{attimino}
g(x):=g \text{ if } x= \t_g 0=Vg\,.
\end{equation}
When $V={\ensuremath{\mathbb I}} $ (as in many applications), we have $\t_g x=x+g$ and therefore $g(x)=x$.
If ${\ensuremath{\mathbb G}} ={\ensuremath{\mathbb Z}} ^d$, $\D$ is a set of orbit representatives
for the action $(\t_g)_{g\in {\ensuremath{\mathbb G}} }$. We introduce the functions $\b: {\ensuremath{\mathbb R}} ^d \to \D$ and $g: {\ensuremath{\mathbb R}} ^d \to {\ensuremath{\mathbb G}} $ as follows:
\begin{equation}\label{attimo}
x= \t_g \bar x \text{ and } \bar x \in \D \; \Longrightarrow \; \b(x):=\bar x \,, \; g(x) := g\,.
\end{equation}
Hence, given $x\in {\ensuremath{\mathbb R}} ^d$, $\bar x$ denotes the unique element of $\D$ such that $x$ and $\bar x$ are in the same ${\ensuremath{\mathbb G}} $--orbit, and $g(x)$ denotes the unique element in ${\ensuremath{\mathbb G}} $ such that $x= \t_{g(x)} \bar x$.
\subsection{An example with ${\ensuremath{\mathbb G}} ={\ensuremath{\mathbb Z}} ^d$ and $V\not ={\ensuremath{\mathbb I}} $}\label{virgilio} Although we will discuss several examples in Section \ref{sec_esa}, our mathematical objects for ${\ensuremath{\mathbb G}} ={\ensuremath{\mathbb Z}} ^d$ and $V\not ={\ensuremath{\mathbb I}} $ could appear very abstract at a first sight.
To have in mind something concrete to which refer below, we present an example related to the random walk and the simple exclusion process on the infinite cluster of the supercritical site Bernoulli percolation on the hexagonal lattice (see Section \ref{sec_esa} for a further discussion).
Consider the hexagonal lattice graph $\ensuremath{\mathcal L}=(\ensuremath{\mathcal V},\ensuremath{\mathcal E})$ in ${\ensuremath{\mathbb R}} ^2$, partially drawn in Figure \ref{apetta}.
$\ensuremath{\mathcal V}$ and $\ensuremath{\mathcal E}$ denote respectively the vertex set and the edge set.
The vectors $v_1$, $v_2$ in Figure \ref{apetta} form a fundamental basis for the hexagonal lattice.
\begin{figure}[!ht]
\begin{center}
\centering
\mbox{\hbox{
\includegraphics[width=0.3\textwidth]{fig_esa_lattice.pdf}}}
\end{center}
\caption{Hexagonal lattice, fundamental cell $\D$, basis $\{v_1,v_2\}$}\label{apetta}
\end{figure}
We take $\O:=\{0,1\}^\ensuremath{\mathcal V}$ endowed with the product topology and with the Bernoulli product probability measure $\ensuremath{\mathcal P}$ with supercritical parameter $p$. We set ${\ensuremath{\mathbb G}} :={\ensuremath{\mathbb Z}} ^2$.
The action $(\theta_g )_{g\in {\ensuremath{\mathbb Z}} ^2}$ is given by $\theta _{(g_1,g_2)} \o = ( \o_{x-g_1 v_1 -g_2 v_2} ) _{x\in \ensuremath{\mathcal V}}$ if $\o=(\o_x) _{x\in \ensuremath{\mathcal V}}$ (note that $v_1,v_2$ are 2d vectors and not coordinates, while $(g_1,g_2)\in {\ensuremath{\mathbb Z}} ^2$). Trivially, $\ensuremath{\mathcal P}$ is stationary and ergodic for this action.
The action of ${\ensuremath{\mathbb Z}} ^2$ on ${\ensuremath{\mathbb R}} ^2$ is given by the translations $\t _{(g_1,g_2)}x:= x+g_1 v_1 +g_2 v_2$. Note that $V=[v_1|v_2]$.
The cell $\D$ in \eqref{simplesso} is here the fundamental cell of the lattice $\ensuremath{\mathcal L}$ given by the parallelogram
with ticked border in Figure \ref{apetta} (one has to remove the upper and right edges). Then the map $\b : {\ensuremath{\mathbb R}} ^2 \to \D$ in \eqref{attimo} is the map $\b(x): =\bar x $ where $\bar x $ is the unique element of $\D$ such that $x =\bar x \text{ mod } {\ensuremath{\mathbb Z}} v_1+{\ensuremath{\mathbb Z}} v_2$. Moreover, the map $g:{\ensuremath{\mathbb R}} ^2 \to {\ensuremath{\mathbb Z}} ^2$ in \eqref{attimo} assigns to $x $ the only element $g=(g_1,g_2) \in {\ensuremath{\mathbb Z}} ^2$ such that $x\in \t_g \D= \D+ g_1 v_1 +g_2 v_2$.
We now describe the simple point process $\hat \o$. As $p$ is supercritical, for $\ensuremath{\mathcal P}$--a.a. $\o$ the set $\{x\in \ensuremath{\mathcal V}\,:\, \o_x=1\}$ has a unique infinite connected component $\ensuremath{\mathcal C}(\o)$ inside the lattice $\ensuremath{\mathcal L}$. We set $\hat \o:= \ensuremath{\mathcal C}(\o)$. To extend this definition to all $\o$, we set $\ensuremath{\mathcal C}(\o):=\emptyset$ if $\o$ has not a unique infinite connected component.
\subsection{Palm distribution}\label{subsec_palm} We recall that we have a simple point process on ${\ensuremath{\mathbb R}} ^d$ defined on our probability space $(\O, \ensuremath{\mathcal F},\ensuremath{\mathcal P})$. This means that to each $\o \in \O$ we associate a locally finite subset $\hat \o \subset {\ensuremath{\mathbb R}} ^d$ by a measurable map $\O \ni \o \to \hat \o \in \ensuremath{\mathcal N} $.
We now recall the definition of Palm distribution $\ensuremath{\mathcal P}_0$ associated to our simple point process
by distinguishing between two main cases and a special subcase.
For a more detailed discussion we refer to \cite{Fhom3} and references therein (we point out here \cite{Ge,GL}). We remark that our treatment reduces to the one in \cite{DV} when ${\ensuremath{\mathbb G}} ={\ensuremath{\mathbb R}} ^d$, $\O=\ensuremath{\mathcal N}$, $\hat \o=\o$, $V={\ensuremath{\mathbb I}} $ (i.e. $\t_g x=x+g)$ and $\theta_g \o := \t_{-g} \o=\o -g$.
We will write ${\ensuremath{\mathbb E}} [\cdot]$ and ${\ensuremath{\mathbb E}} _0[\cdot]$ for the expectation w.r.t. $\ensuremath{\mathcal P}$ and $\ensuremath{\mathcal P}_0$, respectively\footnote{With some abuse, when $f$ has a complex form, we will write ${\ensuremath{\mathbb E}} [f(\o)]$ instead of ${\ensuremath{\mathbb E}} [f]$, and similarly ${\ensuremath{\mathbb E}} _0[f(\o)]$ instead of ${\ensuremath{\mathbb E}} _0[f]$.}.
\smallskip
$\bullet$
{\sl Case ${\ensuremath{\mathbb G}} ={\ensuremath{\mathbb R}} ^d$}. The intensity of the simple point process $\hat\o$ is defined as
\begin{equation}\label{puffo1}
m:= {\ensuremath{\mathbb E}} \left[ \sharp \left( \hat \o \cap [0,1)^d \right)\right]\,.\end{equation}
We will assume that $m\in (0,+\infty)$.
By the ${\ensuremath{\mathbb G}} $-stationarity of $\ensuremath{\mathcal P}$ we have $m \ell(B)={\ensuremath{\mathbb E}} \left[ \sharp \left(\hat \o \cap B\right)\right]$ for any $B\in \ensuremath{\mathcal B} ({\ensuremath{\mathbb R}} ^d)$.
Then the Palm distribution $\ensuremath{\mathcal P}_0$ is the probability measure on $(\O,\ensuremath{\mathcal F})$ such that, for any $U\in \ensuremath{\mathcal B}({\ensuremath{\mathbb R}} ^d)$ with $0<\ell (U)<\infty$ ($\ell(U)$ is the Lebesgue measure of $U$),
\begin{equation}\label{palm_classica}
\ensuremath{\mathcal P}_0(A):=\frac{1}{m \ell(U) }\int _\O d\ensuremath{\mathcal P}(\o) \sum _{x\in \hat \o \cap U} \mathds{1}_A(\theta_{g(x)} \o)\,, \qquad \forall A\in \ensuremath{\mathcal F}
\,.\end{equation}
One can check that $\ensuremath{\mathcal P}_0$ has support inside the set $ \O_0:=\{\o\in \O\,:\, 0\in \hat \o\}$.
\smallskip
$\bullet$ {\sl Case ${\ensuremath{\mathbb G}} ={\ensuremath{\mathbb Z}} ^d$}.
The intensity of the simple point process $\hat \o$ is defined as
\begin{equation}\label{puffo2}
m:= {\ensuremath{\mathbb E}} \left[ \sharp \left( \hat \o \cap \D \right)\right]/ \ell(\D) \,.
\end{equation}
By the ${\ensuremath{\mathbb G}} $-stationarity of $\ensuremath{\mathcal P}$, $m \ell(B)={\ensuremath{\mathbb E}} \left[ \hat \o\left( B \right)\right]$ for any $B\in \ensuremath{\mathcal B}({\ensuremath{\mathbb R}} ^d)$ which is an overlap of translated cells $\t_g \D$ with $g\in {\ensuremath{\mathbb G}} $. We will assume that $m\in (0,+\infty)$.
Then the Palm distribution $\ensuremath{\mathcal P}_0$ is the probability measure on $\left(\O\times \D,\ensuremath{\mathcal F}\otimes \ensuremath{\mathcal B}(\D)\right)$ such that
\begin{equation}\label{Palm_Z}
\ensuremath{\mathcal P}_0(A):=\frac{1}{m\, \ell(\D)}\int _\O d\ensuremath{\mathcal P}(\o) \sum _{x\in \hat \o \cap \D} \mathds{1}_A( \o,x )\,, \qquad \forall A\in \ensuremath{\mathcal F}\otimes \ensuremath{\mathcal B}(\D)
\,.\end{equation}
$\ensuremath{\mathcal P}_0$ has support inside $\O_0:=\{(\o, x)\in \O\times \D\,:\,x\in \hat \o\}$.
Note that in the Example of Subsection \ref{virgilio}, the set $\hat \o \cap \D$ equals $\{0,a\}\cap \ensuremath{\mathcal C}(\o)$, $a$ being as in Figure \ref{apetta}. Moreover, $\O_0= \{ (\o, 0)\,:\, \o\in \O\,, \;0 \in \ensuremath{\mathcal C}(\o)\}
\cup \{ (\o, a)\,:\, \o\in \O\,, \;a \in \ensuremath{\mathcal C}(\o)\} $.
\smallskip
$\bullet$ {\sl Special discrete case: ${\ensuremath{\mathbb G}} ={\ensuremath{\mathbb Z}} ^d$, $V={\ensuremath{\mathbb I}} $ and $\hat \o \subset {\ensuremath{\mathbb Z}} ^d$ $\forall \o\in \O$} (see \eqref{trasferta2}). This is a subcase of the previous one and in what follows we will call it simply \emph{special discrete case}. Due to its relevance in discrete probability, we discuss it apart pointing out some simplifications. As $V={\ensuremath{\mathbb I}} $ we have $\D=[0,1)^d$. In particular (see the case ${\ensuremath{\mathbb G}} ={\ensuremath{\mathbb Z}} ^d$) $\ensuremath{\mathcal P}_0$ is concentrated on $\{ \o \in \O:0\in \hat \o \}\times \{0\}$. Hence we can think of $\ensuremath{\mathcal P}_0$ simply as a probability measure
concentrated on the set $\O_0:=\{ \o \in \O:0\in \hat \o\}$.
Formulas \eqref{puffo2} and \eqref{Palm_Z}
then read
\begin{equation}\label{zazzera}
m:= \ensuremath{\mathcal P}(0\in \hat \o) \,, \qquad
\ensuremath{\mathcal P}_0(A):= \ensuremath{\mathcal P}\left( A\,|\, 0 \in \hat \o \right) \qquad \forall A\in \ensuremath{\mathcal F}
\,.
\end{equation}
In what follows, when treating the special discrete case, we will use the above identifications without explicit mention.
\subsection{Basic assumptions} Recall that
the jump probability rates are given by the measurable function $c_{x,y}(\o)$ in
\eqref{ciccino}, which is symmetric in $x,y$ (i.e. $c_{x,y}(\o)= c_{y,x}(\o)$) and recall our convention \eqref{cicciobello}. We also define
\begin{equation} \label{clacson}
c_x(\o):= \sum _{y \in \hat \o } c_{x,y} (\o) \qquad \forall x \in \hat \o\,.
\end{equation}
We define the functions
$\l_k:\O_0 \to [0,+\infty]$ (for $k=0,2$) as follows:\\
\begin{equation}\label{altino15}
\begin{split}
&
\begin{cases}
\l_k(\o):=\sum_{x\in \hat \o} c_{0,x}(\o)|x|^k\\
\O_0=\{\o\in \O\,:\, 0\in \hat \o\}
\end{cases}
\Large{\substack{ \text{\;\;Case ${\ensuremath{\mathbb G}} ={\ensuremath{\mathbb R}} ^d$ and}\\\text{\;\;\;\;\;\;special discrete case}}\,,}
\\
& \begin{cases}
\l_k(\o,a):=
\sum_{x\in \hat \o} c_{a,x}(\o)|x-a|^k \\
\O_0=\{(\o, x)\in \O\times \D\,:\,x\in \hat \o \}
\end{cases}
\text{\;\;Case ${\ensuremath{\mathbb G}} ={\ensuremath{\mathbb Z}} ^d$\,.}
\end{split}
\end{equation}
For ${\ensuremath{\mathbb G}} ={\ensuremath{\mathbb R}} ^d$ and in the special discrete case, $\l_0(\o)=c_0(\o)$ for all $\o\in\O_0$.
We collect below all our assumptions leading to homogenization of the massive Poisson equation of the diffusively rescaled random walk (some of them have already been mentioned
in Section \ref{cima}). We will not recall here the above homogenization results obtained in \cite{Fhom3}, as not necessary. On the other hand, we will collect some of their consequences in Proposition \ref{replay} in Section \ref{figlio_stress}, since used in the proof of Theorem \ref{teo1}.
\smallskip
\noindent
{\bf Assumptions for homogenization}:
\emph{
\begin{itemize}
\item[(A1)] $\ensuremath{\mathcal P}$ is stationary and ergodic w.r.t. the action $(\theta_g)_{g\in {\ensuremath{\mathbb G}} }$ of the group ${\ensuremath{\mathbb G}} $;
\item[(A2)] the $\ensuremath{\mathcal P}$--intensity $m$ of the random measure $\hat \o $
is finite and positive (cf. \eqref{puffo1}, \eqref{puffo2} and \eqref{zazzera});
\item[(A3)] the $\o$'s in $\O$ such that $ \theta_g\o\not = \theta _{g'} \o $ for all $ g\not =g'$ in ${\ensuremath{\mathbb G}} $ form a measurable set of $\ensuremath{\mathcal P}$--probability $1$;
\item[(A4)] the $\o$'s in $\O$ such that,
for all $g\in {\ensuremath{\mathbb G}} $ and $x,y \in {\ensuremath{\mathbb R}} ^d$,
\begin{align}
&\widehat{\theta_g\o} =\t_{-g}( \hat \o ) \,,\label{base}\\
& c_{x,y} (\theta_g\o)= c_{\t_g x, \t_g y} (\o) \,,\label{montagna}
\end{align}
form a measurable set of $\ensuremath{\mathcal P}$--probability $1$;
\item[(A5)] for $\ensuremath{\mathcal P}$--a.a. $\o\in \O$, for all $x,y\in {\ensuremath{\mathbb R}} ^d$ it holds \begin{equation}\label{pietra}
c_{x,y}(\o) = c_{y,x}(\o)\,;
\end{equation}
\item[(A6)] for $\ensuremath{\mathcal P}$--a.a. $\o\in \O$, given any $x,y \in \hat \o$ there exists a path $x=x_0$, $x_1$,$ \dots, x_{n-1}, x_n =y$ such that $x_i \in \hat \o$ and $c_{x_i, x_{i+1}}(\o) >0$ for all $i=0,1, \dots, n-1$;
\item[(A7)] $\l_0 , \l_2 \in L^1(\ensuremath{\mathcal P}_0)$;
\item[(A8)] $L^2(\ensuremath{\mathcal P}_0)$ is separable.
\end{itemize}
}
\smallskip
The above assumptions implies that $\ensuremath{\mathcal P}$--a.s. the random walk $X_t^\o$ on $\hat \o$ introduced in Section \ref{cima} is well defined for all times $t\geq 0$
(recall that a set $A\subset \O$ is called translation invariant if $\theta_g A=A$ for all $g\in {\ensuremath{\mathbb G}} $):
\begin{Lemma}\label{lemma_no_TNT}\cite[Lemma 3.5]{Fhom3} There exists a translation invariant measurable set $\ensuremath{\mathcal A} \subset\O$ with $\ensuremath{\mathcal P}(\ensuremath{\mathcal A})=1$ such that, for all $\o \in \ensuremath{\mathcal A}$, (i) $c_x(\o)\in (0,+\infty)$ for all $x \in \hat \o$ (cf. \eqref{clacson}), (ii) the continuous--time Markov chain on $\hat \o$ starting at any $x_0\in \hat \o$, with waiting time parameter $c_x(\o)$ at $x\in \hat \o$ and with probability $c_{x,y}(\o)/c_x(\o) $ for a jump from $x$ to $y$, is non-explosive.
\end{Lemma}
In Section \ref{HL_EP} we will make an additional assumption (called Assumption (SEP)) assuring that the simple exclusion process introduced via the universal graphical construction is well defined for all times (see \eqref{mammaE} for its generator on local functions). Hence, by thinking the random walk $X_t^\o$ as a simple exclusion process with only one particle, also Assumption (SEP) guarantees the well-definedness of $X_t^\o$.
We now report some other comments on the above assumptions (A1),...,(A8) taken from \cite[Section~2.4]{Fhom3} (where more details are provided).
By Zero-Infinity Dichotomy (see \cite[Proposition~12.1.VI]{DV2}) and Assumptions (A1) and (A2), for $\ensuremath{\mathcal P}$--a.a.~$\o$ the set $\hat \o$ is infinite.
(A3) is a rather fictitious assumption as one can add some randomness by enlarging $\O$ to assure (A3). The assumption of measurability in (A3) and (A4) is always satisfied for ${\ensuremath{\mathbb G}} ={\ensuremath{\mathbb Z}} ^d$ by \eqref{cicciobello} (as discussed in \cite[Section~2.4]{Fhom3}, one can even weaken this requirement).
Considering the random walk $X_t^\o$, (A5) and (A6) correspond $\ensuremath{\mathcal P}$--a.s. to reversibility of the counting measure and to irreducibility. Finally, we point out that, by \cite[Theorem~4.13]{Br}, (A8) is fulfilled if $(\O_0,\ensuremath{\mathcal F}_0,\ensuremath{\mathcal P}_0)$ is a separable measure space where $\ensuremath{\mathcal F}_0:=\{A\cap \O_0\,:\, A\in \ensuremath{\mathcal F}\}$ (i.e. there is a countable family $\ensuremath{\mathcal G}\subset \ensuremath{\mathcal F}_0$ such that the $\s$--algebra $\ensuremath{\mathcal F}_0$ is generated by $\ensuremath{\mathcal G}$). For example, if $\O_0$ is a separable metric space and
$\ensuremath{\mathcal F}_0= \ensuremath{\mathcal B}(\O_0)$ (which is valid if $\O$ is a separable metric space and
$\ensuremath{\mathcal F}= \ensuremath{\mathcal B}(\O)$) then (cf. \cite[p.~98]{Br}) $(\O_0,\ensuremath{\mathcal F}_0,\ensuremath{\mathcal P}_0)$ is a separable measure space and (A8) is valid.
\medskip
We now explain why the Palm distribution $\ensuremath{\mathcal P}_0$ will play a crucial role in the hydrodynamic limit of the simple point process. $\ensuremath{\mathcal P}_0$ is indeed the natural object to express the ergodicity property of the environment when dealing with observables keeping track also of the local microscopic details of the environment. This is formalized by the following result which will be frequently used below (cf. \cite[App.~B]{Fhom1}, \cite[Prop.~3.1]{Fhom3} and recall that ${\ensuremath{\mathbb E}} _0$ denotes the expectation w.r.t. $\ensuremath{\mathcal P}_0$):
\begin{Proposition} \label{prop_ergodico} Let $f: \O_0\to {\ensuremath{\mathbb R}} $ be a measurable function with $\|f\|_{L^1(\ensuremath{\mathcal P}_0)}<\infty$. Then there exists a translation invariant measurable subset $\ensuremath{\mathcal A}[f]\subset \O$ such that $\ensuremath{\mathcal P}(\ensuremath{\mathcal A}[f])=1$ and such that, for any $\o\in \ensuremath{\mathcal A}[f]$ and any $\varphi \in C_c ({\ensuremath{\mathbb R}} ^d)$, it holds
\begin{equation}\label{limitone}
\lim_{\e\downarrow 0} \int d \mu_\o^\e (x) \varphi (x ) f(\theta_{g( x/\e)} \o )=
\int dx\,m\varphi (x) \cdot {\ensuremath{\mathbb E}} _0[f]\,,
\end{equation}
where $\mu^\e_\o := \sum_{x\in \hat \o} \e^d \d_{\e x}$.
\end{Proposition}
We can now also introduce the effective homogenized matrix $D$, defined in terms of the Palm distribution:
\begin{Definition}\label{def_D} We define the \emph{effective homogenized matrix} $D$ as
the unique $d\times d$ symmetric matrix such that:
\smallskip
$\bullet$ {\bf Case ${\ensuremath{\mathbb G}} ={\ensuremath{\mathbb R}} ^d$ and special discrete case}
\begin{equation}\label{def_D_R}
a \cdot Da =\inf _{ f\in L^\infty(\ensuremath{\mathcal P}_0) } \frac{1}{2}\int _{\O_0} d\ensuremath{\mathcal P}_0(\o)\sum_{x\in \hat \o} c_{0,x}(\o) \left
(a\cdot x - \nabla f (\o, x)
\right)^2\,,
\end{equation}
for any $a\in {\ensuremath{\mathbb R}} ^d$,
where $\nabla f (\o, x) := f(\theta_{g(x)} \o) - f(\o)$.
\smallskip
$\bullet$ {\bf Case ${\ensuremath{\mathbb G}} ={\ensuremath{\mathbb Z}} ^d$ }
\begin{align}
& a \cdot Da= \label{def_D_Z} \\
& \inf _{ f\in L^\infty(\ensuremath{\mathcal P}_0) } \frac{1}{2} \int_{\O\times \D}d\ensuremath{\mathcal P}_0(\o,x) \sum_{y \in \hat \o} c_{x,y}(\o) \left
(a\cdot (y-x) -
\nabla f (\o, x,y-x)
\right)^2\,,\nonumber
\end{align}
for any $a\in {\ensuremath{\mathbb R}} ^d$,
where $\nabla f (\o, x,y-x) := f(\theta_{g(y)} \o, \beta(y) ) - f(\o,x)$ (recall \eqref{attimo}).
\end{Definition}
Finally we introduce an additional assumption assuring a weak form of convergence for the $L^2$-Markov semigroup and the $L^2$-resolvent associated to the random walk $X_t^\o$ as discussed in Section \ref{figlio_stress} (recall definition \eqref{simplesso} of $\D$).
\smallskip
\noindent
{\bf Additional assumption for semigroup and resolvent convergence}:
\emph{
\begin{itemize}
\item[(A9)]
At least one of the following conditions is satisfied:
\begin{itemize}
\item[(i)] for $\ensuremath{\mathcal P}$--a.a.~$\o$
$\exists C(\o)>0$ such that
\begin{equation}\label{crostino}
\sharp( \hat \o \cap \t_k \D) \leq C(\o) \text{ for all } k\in {\ensuremath{\mathbb Z}} ^d\,;
\end{equation}
\item[(ii)]
at cost to enlarge the probability space $\O$ one can define random variables $(N_k) _{k\in {\ensuremath{\mathbb Z}} ^d}$ with $ \sharp( \hat \o \cap \t_k \D) \leq N_k$ and such that, for some $C_0\geq 0$, it holds
\begin{align}
& \sup_{k \in {\ensuremath{\mathbb Z}} ^d} {\ensuremath{\mathbb E}} [N_k]<+\infty\,,\qquad \sup_{k \in {\ensuremath{\mathbb Z}} ^d}{\ensuremath{\mathbb E}} [N_k^2] <+\infty\,,\label{zuppetta}
\\
& |\text{Cov}\,(N_k, N_{k'})| \leq
C_0 |k-k'| ^{-1} \qquad \forall k\not = k'\text{ in }{\ensuremath{\mathbb Z}} ^d \,.\label{intinto}
\end{align}
\end{itemize}
\end{itemize}
}
\begin{Remark} If one set $N_k:= \sharp( \hat \o \cap \t_k \D) $ for $k\in {\ensuremath{\mathbb Z}} ^d$, then to check Condition (ii) in (A9) it is enough to check that ${\ensuremath{\mathbb E}} [N_0^2]<+\infty$ and \eqref{intinto} (due to (A1) and (A2)). As discussed in \cite[Remark~4.2]{Fhom3},
when ${\ensuremath{\mathbb G}} ={\ensuremath{\mathbb R}} ^d$, by defining $N_k:= \sharp( \hat \o \cap \t_k \D)$ one can replace the cells $\{\t_k \D\}_{k \in {\ensuremath{\mathbb Z}} ^d}$ by the cells of any lattice partition of ${\ensuremath{\mathbb R}} ^d$.
\end{Remark}
\section{Hydrodynamic limit }\label{HL_EP}
Given $\o \in \O$ we consider the simple exclusion process on $\hat\o$ with particle exchange probability rate $c_{x,y}(\o)$.
To have a well defined process for all times $t\geq 0$, $\ensuremath{\mathcal P}$--a.s., we will use in Section \ref{sec_GC} Harris' percolation argument \cite{Du}. To this aim, we define
\begin{equation}\label{pitagora}
\ensuremath{\mathcal E}_\o :=\{ \,\{x,y\}\,:\, x,y \in \hat \o\,, \; x\not=y\,\}\,.
\end{equation}
Then,
given $\o$, we associate to each unordered pair $\{x,y\}\in \ensuremath{\mathcal E}_\o $ a Poisson process $( N_{x,y}(t))_{t\geq 0}$ with intensity $c_{x,y}(\o)$, such that the $N_{x,y}(\cdot)$'s are independent processes when varying the pair $\{x,y\}$.
The random object $( N_{x,y}(\cdot ) )_{\{x,y\}\in \ensuremath{\mathcal E}_\o}$ takes value in the product space $ D({\ensuremath{\mathbb R}} _+, {\ensuremath{\mathbb N}} )^{\ensuremath{\mathcal E}_\o}$, $ D({\ensuremath{\mathbb R}} _+, {\ensuremath{\mathbb N}} )$ being endowed with the standard Skohorod topology. In the rest, we will denote by $\ensuremath{\mathcal K}= ( \ensuremath{\mathcal K}_{x,y}(\cdot ) )_{\{x,y\}\in \ensuremath{\mathcal E}_\o}$ a generic element of $ D({\ensuremath{\mathbb R}} _+, {\ensuremath{\mathbb N}} )^{\ensuremath{\mathcal E}_\o}$.
Moreover, we denote by ${\ensuremath{\mathbb P}} _\o$ the law on $ D({\ensuremath{\mathbb R}} _+, {\ensuremath{\mathbb N}} )^{\ensuremath{\mathcal E}_\o}$
of $( N_{x,y}(\cdot ) )_{\{x,y\}\in \ensuremath{\mathcal E}_\o}$.
\medskip
In this section we add the following assumption (we call it ``SEP" for ``simple exclusion process" as the assumption is introduced to assure the existence of the simple exclusion process):
\smallskip
\noindent {\bf Assumption (SEP)}. \emph{For $\ensuremath{\mathcal P}$--a.a. $\o$ there exists $t_0=t_0(\o)>0$ such that
for ${\ensuremath{\mathbb P}} _\o$--a.a. $\ensuremath{\mathcal K}$
the undirected graph $\ensuremath{\mathcal G} _{t_0}(\o,\ensuremath{\mathcal K})$ with vertex set $\hat \o$ and edges
\[ \{ \{x,y\}\in \ensuremath{\mathcal E}_\o \,:\, \ensuremath{\mathcal K}_{x,y}(t_0) \geq 1\}\] has only connected components with finite cardinality.}
\smallskip
In Section \ref{sec_GC} we discuss the universal graphical construction of the exclusion process on $\hat \o$ under Assumption (SEP). For $\ensuremath{\mathcal P}$--a.a. $\o$ the resulting process is a Feller process and the infinitesimal generator $\ensuremath{\mathcal L}_\o$
acts on local functions as in \eqref{mammaE} and \eqref{mahmood} (see Proposition \ref{prop_SEP}).
We denote by $\ensuremath{\mathcal M}$ the space of Radon measures on ${\ensuremath{\mathbb R}} ^d$ endowed with the
vague topology and we denote by $D([0,T], \ensuremath{\mathcal M})$ the
Skohorod space of c\`adl\`ag paths from $[0,T]$ to $\ensuremath{\mathcal M}$ endowed with the Skohorod metric (see Section \ref{sec_mammina} for details).
For each $\e>0$ we consider the map
\[ \{0,1\}^{\hat \o}\ni \eta \mapsto \pi^\e_\o [\eta]:=\e^d \sum_{x\in \hat \o} \eta(x) \d_{\e x} \in \ensuremath{\mathcal M}\,.\]
Above $\pi^\e_\o [\eta]$ is the so called \emph{empirical measure} associate to $\eta$.
%
Given a path $\eta_\cdot= (\eta_s )_{0 \leq s \leq T}$ and given $t\in [0,T]$, we define
$\pi^\e_{\o,t} [ \eta_\cdot]:= \pi^\e_\o [ \eta _t ]$.
In what follows, given $\e>0$ and a probability measure $\mathfrak{m} $ on $\{0,1\}^{ \hat \o}$,
we denote by ${\ensuremath{\mathbb P}} ^\e_{\o, \mathfrak{m} }$ the law of the diffusively rescaled exclusion process on $ \hat \o$ with generator $\e^{-2} \ensuremath{\mathcal L}_\o$ and initial distribution $\mathfrak{m} $.
Note that the time $T$ is understood and does not appear in the notation.
We denote by $(B_t)_{t\geq 0}$ the Brownian motion on ${\ensuremath{\mathbb R}} ^d$ with diffusion matrix given by $2D$, $D$ being the effective homogenized matrix (see Definition \ref{def_D}). As $D$ is symmetric we
can fix an orthonormal basis
$\mathfrak{e}_1$,...,$ \mathfrak{e}_d$ of eigenvectors of $D$, such that $\mathfrak{e}_1$,...,$ \mathfrak{e}_{d_*}$ have positive eigenvalues, while the other basis vectors have zero eigenvalue.
Then the Brownian motion $(B_t)_{t\geq 0}$ is not degenerate when projected on $\text{span}(\mathfrak{e}_1, \dots, \mathfrak{e}_{d_*})$, while no motion is present along
$\text{span}(\mathfrak{e}_{d_*+1}, \dots, \mathfrak{e}_{d})$. Given a bounded function $f:{\ensuremath{\mathbb R}} ^d\to {\ensuremath{\mathbb R}} $ we set $P_t f(x):=E\left[ f(x+B_t)\right]$.
\begin{TheoremA}\label{teo1}
Suppose that Assumptions (A1),\dots,(A9) and Assumption (SEP) are satisfied.
Then there exists a translation invariant measurable set $\O_{\rm typ}\subset \O$ with $\ensuremath{\mathcal P}(\O_{\rm typ})=1$, such that
for any $\o \in \O_{\rm typ}$ the simple exclusion process
is well defined for any initial distribution and exhibits the following hydrodynamic behavior.
Let $\rho_0: {\ensuremath{\mathbb R}} ^d \to [0,1]$ be a measurable function and let $\rho:{\ensuremath{\mathbb R}} ^d\times[0,\infty)\to [0,1]$ be the function $\rho(x,t):= P_t \rho_0 (x)$.
Let $\{\mathfrak{m}_\e\}_{\e >0}$ be an $\e$--parametrized family of probability measures on $\{0,1\}^{\hat{\o}}$ such that the random empirical measure $\pi_\o ^\e[\eta] $ in $\ensuremath{\mathcal M}$, with $\eta$ sampled according to $\mathfrak{m}_\e$, converges in probability to $\rho_0(x)dx$ inside $\ensuremath{\mathcal M}$. In other words, we suppose that,
for all $\d>0$ and $\varphi\in C_c({\ensuremath{\mathbb R}} ^d)$, it holds
\begin{equation}\label{marzolino}
\lim_{\e\downarrow 0} \mathfrak{m}_\e\Big( \Big| \e^d \sum_{x \in \hat \o} \varphi (\e x) \eta(x) -\int _{{\ensuremath{\mathbb R}} ^d} \varphi(x) \rho_0(x) dx \Big|>\d\Big)=0\,.
\end{equation}
Then:
\begin{itemize}
\item[(i)] For all $T>0$, as $\e\downarrow 0$ the random path $( \pi^\e_{\o,t} [ \eta_\cdot])_{0\leq t \leq T} $ in $D ( [0,T], \ensuremath{\mathcal M})$, with
$\eta_\cdot=( \eta_t )_{0\leq t \leq T} $ sampled according to ${\ensuremath{\mathbb P}} ^\e_{\o, \mathfrak{m}_\e }$, converges in probability to the deterministic path $(\rho(x,t) dx )_{0\leq t \leq T}$.
\item[(ii)] For all $T>0$, $\varphi \in C_c({\ensuremath{\mathbb R}} ^d)$ and $ \d>0$, it holds
\begin{equation}\label{pasqualino}
\lim_{\e\downarrow 0} {\ensuremath{\mathbb P}} ^\e _{ \o,\mathfrak{m}_\e } \Big(\sup_{0\leq t \leq T}\Big| \e^d \sum_{x \in \hat \o} \varphi (\e x) \eta_t (x) - \int _{{\ensuremath{\mathbb R}} ^d} \varphi(x) \rho(x,t) dx\Big| >\d
\Big)=0\,.
\end{equation}
\end{itemize}
\end{TheoremA}
The proof of the above theorem is given in Section \ref{ida} (Section \ref{silenzioso} can be replaced by Appendix \ref{sec_passetto}, the two approaches are alternative). The function $\rho(x,t)= P_t \rho_0 (x)$ is the unique weak solution of the Cauchy system
\begin{equation}\label{debolezza}
\begin{cases}
\partial_t \rho = \nabla \cdot ( D \nabla \rho) \text{ for } t>0\,,\\
\rho(0,\cdot) = \rho_0 \,,
\end{cases}
\end{equation}
in the sense specified by Lemma \ref{timau} in Section \ref{silenzioso}.
\section{Some applications}\label{sec_esa}
There are plenty of examples to which Theorem \ref{teo1} can be applied. We discuss here three main classes.
We also stress that one can apply Theorem \ref{teo1} to the same context of \cite{F1}, replacing there the uniform upper bound on conductances by integrability (see also Section \ref{nonno} and \cite[Section~5.3]{Fhom3}).
\subsection{Nearest-neighbor random conductance model on ${\ensuremath{\mathbb Z}} ^d$, $d\geq 1$}\label{nonno}
We take ${\ensuremath{\mathbb G}} :={\ensuremath{\mathbb Z}} ^d$ acting on ${\ensuremath{\mathbb R}} ^d$ by standard translations, i.e. $\t_g x=x+ g$. Let ${\ensuremath{\mathbb E}} ^d$ be the set of unoriented edges of ${\ensuremath{\mathbb Z}} ^d$ and endow $\O:=(0,+\infty)^{{\ensuremath{\mathbb E}} ^d}$ with the product topology. Given $\o\in \O$, we write $\o_{x,y}$ for the component of $\o$ associated to the edge $\{x,y\}\in {\ensuremath{\mathbb E}} ^d$.
The action $(\theta _x)_{x\in {\ensuremath{\mathbb Z}} ^d}$ is the standard one: $(\theta_x \o) _{a,b}:= \o _{a+x,b+x}$.
We set $\hat \o := {\ensuremath{\mathbb Z}} ^d$, hence the exclusion process lives on ${\ensuremath{\mathbb Z}} ^d$. We define $c_{x,y}(\o):=\o_{x,y}$ if $\{x,y\}\in {\ensuremath{\mathbb E}} ^d$ and $c_{x,y}(\o):=0$ otherwise.
It is simple to check that
Assumptions (A1),..., (A9) are satisfied whenever $\ensuremath{\mathcal P}$ is stationary and ergodic, $\ensuremath{\mathcal P}$ satisfies (A3) (which is a fictitious assumption) and ${\ensuremath{\mathbb E}} [\o_{x,y}]<+\infty$ for all $\{x,y\}\in {\ensuremath{\mathbb E}} ^d$.
Below, given $k>0$, we say that the random conductances $\o_{x,y}$ are $k$--dependent if, given $A,B\subset {\ensuremath{\mathbb Z}} ^d$ with Euclidean distance between $A$ and $B$ larger than $k$, the random fields
\[ \left( \o_{x,y}\,:\,\{x,y\} \in {\ensuremath{\mathbb E}} ^d\,,\; x,y\in A\right) \text{ and }\left( \o_{x,y}\,:\,\{x,y\} \in {\ensuremath{\mathbb E}} ^d\,,\; x,y\in B\right)
\] are independent (see \cite[page 178]{G} for a similar definition).
\begin{Proposition} Assumption (SEP) is satisfied if at least one of the following conditions is satisfied:
\begin{itemize}
\item[(i)] $\ensuremath{\mathcal P}$--a.s. there exists a constant $C(\o)$ such that $\o_{x,y}\leq C(\o)$ for all $\{x,y\}\in {\ensuremath{\mathbb E}} ^d$;
\item[(ii)] under $\ensuremath{\mathcal P}$ the random conductances $\o_{x,y}$ are independent;
\item[(iii)] under $\ensuremath{\mathcal P}$ the random conductances $\o_{x,y}$ are $k$--dependent with $k>0$.
\end{itemize}
\end{Proposition}
We note that, by ergodicity, in Item (i) one could just restrict to a non-random upper bound $C$.
Item (ii) is a special case of Item (iii).
\begin{proof} We start with Item (i). As
${\ensuremath{\mathbb P}} _\o ( \ensuremath{\mathcal K}_{x,y}(t_0)>0)= 1-e^{ - \o_{x,y} t_0}$, it is enough to take $t_0$ small to have $1-e^{ - C (\o) t_0}< p_c$, $p_c>0$ being the critical probability for the Bernoulli bond percolation on ${\ensuremath{\mathbb Z}} ^d$.
Let us consider Items (ii) and (iii). We present an argument valid for all $d\geq 1$ (but for $d=1$ one can give easily a more direct proof). By ${\ensuremath{\mathbb Z}} ^d$--stationarity the distribution of $\o_{x,y}$ depends only on the axis parallel to the edge $\{x,y\}$. To simplify the notation we suppose that the conductances are identically distributed with common distribution $\nu$ (otherwise one has just to deal with a finite family of distributions $\nu_1,\nu_2,\dots, \nu_d$ in the stochastic domination below).
We observe that, for any $C_0>0$, the graph $\ensuremath{\mathcal G}_{t_0} (\o,\ensuremath{\mathcal K})$ described in Assumption (SEP) is contained in the graph $\ensuremath{\mathcal G}'_{t_0}(\o,\ensuremath{\mathcal K})$ with edges $\{x,y\}\in {\ensuremath{\mathbb E}} ^d$
such that
\[ \o_{x,y} > C_0 \qquad \text{ or } \qquad
\begin{cases}
\o_{x,y} \leq C_0\,,\\
\ensuremath{\mathcal K}_{x,y}(t_0)>0\,.
\end{cases}
\]
Given $e\in {\ensuremath{\mathbb E}} ^d$ we set $Y_e(\o,\ensuremath{\mathcal K}):=1 $ if $e$ is present in $\ensuremath{\mathcal G}'_{t_0}(\o,\ensuremath{\mathcal K})$, otherwise we set $Y_e=0$. We define $\a(C_0):= \nu \left( (C_0 ,+\infty) \right) $. Then, under $ {\ensuremath{\mathbb P}} := \int d\ensuremath{\mathcal P}(\o) {\ensuremath{\mathbb P}} _\o$, the random field $Y=(Y_e)_{e\in {\ensuremath{\mathbb E}} ^d}$ is stationary, satisfies
${\ensuremath{\mathbb P}} ( Y_e=1)\leq \a(C_0) + (1-\a(C_0))( 1-e^{ - C _0 t_0})$ and is given by independent r.v.'s under (ii) and by $k$--dependent r.v.'s under (iii). Hence, fixed $p_*\in(0,p_c)$, we can first choose $C_0$ large and afterwards $t_0$ small to have ${\ensuremath{\mathbb P}} ( Y_e=1)\leq p_*$. In particular, in case (ii) we conclude that ${\ensuremath{\mathbb P}} $--a.s. $Y$ does not percolate. Similarly to \cite[Theorem (7.65)]{G} (invert the role between $0$ and $1$ there), taking $p_*\in(0,p_c)$ and $\d=\d(p_*)$ small enough, we get that the random field $Y$ is stochastically dominated by a subcritical Bernoulli bond percolation of parameter $p_*$ and therefore ${\ensuremath{\mathbb P}} $--a.s. $Y$ does not percolate. Hence, in both cases (ii) and (iii), by suitably choosing $C_0,t_0$, the graph $\ensuremath{\mathcal G}'_{t_0}(\o,\ensuremath{\mathcal K})$ has only connected components with finite
cardinality ${\ensuremath{\mathbb P}} $ a.s. (i.e. for $\ensuremath{\mathcal P}$--a.a. $\o$ and for ${\ensuremath{\mathbb P}} _\o$--a.a. $\ensuremath{\mathcal K}$). The same then must hold for $\ensuremath{\mathcal G}_{t_0}(\o,\ensuremath{\mathcal K})\subset \ensuremath{\mathcal G}'_{t_0}(\o,\ensuremath{\mathcal K})$.
\end{proof}
\subsection{Nearest--neighbor random conductance models on a generic crystal lattice} We consider a generic crystal lattice $\ensuremath{\mathcal L}=(\ensuremath{\mathcal V},\ensuremath{\mathcal E})$ in ${\ensuremath{\mathbb R}} ^d$, $d\geq 1$, as follows.
We fix a basis $v_1,\dots, v_d$ of ${\ensuremath{\mathbb R}} ^d$, write $V$ for the matrix with columns $v_1,\dots ,v_d$ and write $\D$ for the $d$--dimensional cell \eqref{simplesso}.
Given $g\in {\ensuremath{\mathbb G}} :={\ensuremath{\mathbb Z}} ^d$, we denote by $\t_g$ the translation \eqref{trasferta2}, i.e. $\t_g x = x+ V g $.
We fix a finite set $\ensuremath{\mathcal A} \subset \D$.
Then the vertex set $\ensuremath{\mathcal V}$ of the crystal lattice is given by $\sqcup_{g\in {\ensuremath{\mathbb G}} } ( \t_g \ensuremath{\mathcal A})$.
The edge set $\ensuremath{\mathcal E}$ has to be a family of unoriented pairs of vertexes $\{x,y\}$ with $x\not =y$ in $\ensuremath{\mathcal V}$, such that $\t_g \ensuremath{\mathcal E}= \ensuremath{\mathcal E}$ for all $g\in {\ensuremath{\mathbb G}} $. In particular, the crystal lattice $\ensuremath{\mathcal L}=(\ensuremath{\mathcal V},\ensuremath{\mathcal E})$ is left invariant by the action $ (\t_g )_{g\in {\ensuremath{\mathbb G}} }$ on ${\ensuremath{\mathbb R}} ^d$.
As example consider the hexagonal lattice $\ensuremath{\mathcal L}=(\ensuremath{\mathcal V},\ensuremath{\mathcal E})$ in ${\ensuremath{\mathbb R}} ^2$
(cf. Section \ref{virgilio}).
Then $\ensuremath{\mathcal A}=\{0, a\}$ (see Figure \ref{apetta}).
We take $\O:= (0,+\infty)^{\ensuremath{\mathcal E}}$ endowed with the product topology and set $\o_{x,y}:=\o_{\{x,y\}}$.
The action of $(\theta_g)_{g\in {\ensuremath{\mathbb G}} }$ on $\O$ is given by
$
\theta _g \o: =(\o_{x-Vg,y-Vg }\,:\, \{x,y\}\in \ensuremath{\mathcal E}) $ if $ \o=(\o_{x,y}\,:\,\{x,y\}\in \ensuremath{\mathcal E})$.
For any $\o \in \O$, we set $\hat \o := \ensuremath{\mathcal V}$, hence our simple exclusion process lives on $\ensuremath{\mathcal V}$.
The set $\O_0$ introduced after \eqref{Palm_Z} equals $\O\times \ensuremath{\mathcal A}$ and, by \eqref{puffo2}, $m \ell (\D)= |\ensuremath{\mathcal A}| $. Hence (see \eqref{Palm_Z})
$\ensuremath{\mathcal P}_0(d\o,dx)= \ensuremath{\mathcal P}(d\o) \otimes {\rm Av}_{u\in\ensuremath{\mathcal A} } \d_u(dx) $, where ${\rm Av}$ denotes the arithmetic average and $\d_u$ is the Dirac measure at $u$.
We set $c_{x,y}(\o):= \o_{x,y}$ if $\{x,y\}\in \ensuremath{\mathcal E}$ and $c_{x,y}(\o):=0$ otherwise. If $\ensuremath{\mathcal P}$ satisfies (A1), (A2), (A3) and the crystal lattice is connected, then all assumptions (A1),\dots,(A9) are satisfied if $\sum _{y\in \ensuremath{\mathcal V}}\sum_{u \in \ensuremath{\mathcal A}} {\ensuremath{\mathbb E}} [ \o_{u,y}]|y-u|^2<+\infty$.
It the crystal lattice is locally finite (i.e. vertexes have finite degree), then the above moment bound equal the bound ${\ensuremath{\mathbb E}} [\o_{x,y}]<+\infty$ for $\{x,y\}\in \ensuremath{\mathcal E}$ (by ${\ensuremath{\mathbb G}} $--stationarity and local finiteness, we have just a finite family of bounds).
For locally finite crystal lattices, by reasoning as done for the lattice ${\ensuremath{\mathbb Z}} ^d$, we get that Assumption (SEP) is satisfied if
the conductances $\o_{x,y}$
are uniformly bounded or if the conductances $\o_{x,y}$ are independent or $k$--dependent under $\ensuremath{\mathcal P}$.
\subsection{Mott variable range hopping with site-exclusion interaction}\label{ex_mott} Mott random walk (see e.g. \cite{FM,FSS}) is a mean-field model for Mott variable range hopping in amorphous solids \cite{POF}.
The following exclusion process is obtained by adding a site-exclusion interaction to multiple Mott random walkers.
We take ${\ensuremath{\mathbb G}} := {\ensuremath{\mathbb R}} ^d$ ($d\geq 1$) acting on ${\ensuremath{\mathbb R}} ^d$ by standard translations ($\t_g x=x+ g$). $\O$ is given by the space of marked counting measures with marks in ${\ensuremath{\mathbb R}} $ \cite{DV}, hence $(\O,\ensuremath{\mathcal F},\ensuremath{\mathcal P})$ describes a marked simple point process \cite{DV}. By identifying $\o$ with its support, we have $\o=\{(x_i,E_i)\}$ where $E_i\in {\ensuremath{\mathbb R}} $ and the set $\{x_i\}$ is locally finite. The action $\theta_x$ on $\O$ is given by $\theta _x \o:=\{ (x_i-x, E_i)\}$ if $\o=\{(x_i, E_i)\}$.
Given the environment $\o=\{(x_i,E_i)\}$, Mott random walk
has state space $\hat \o =\{x_i\}$ and jump rates
\begin{equation}\label{vento}
c_{x_i,x_j}(\o):= \exp\bigl\{ -|x_i-x_j| - |E_{x_i}|-|E_{x_j}|-|E_{x_i}-E_{x_j}|\bigr\}\,\qquad x_i\not =x_j
\end{equation}
(above, without loss of generality, we have fixed to absolute constants the inverse temperature and the localization length).
Suppose that $\ensuremath{\mathcal P}$ satisfies (A1),(A2) and (A3). Then $\ensuremath{\mathcal P}_0$ is simply the standard Palm distribution associated to the marked simple point process with law $\ensuremath{\mathcal P}$ \cite{DV}. Assumptions (A4), (A5), (A6) are automatically satisfied. As the above space $\O$ is Polish (see \cite{DV}) and $\O_0=\{\o\,:\, 0 \in \hat \o\} $ is a Borel subset of $\O$, $\O_0$ is separable and therefore (A8) is satisfied.
As proven in \cite[Section 5]{Fhom3},
(A7) is equivalent to the bound ${\ensuremath{\mathbb E}} \bigl[ |\hat \o \cap [0,1]^d | ^{2} \bigr]<+\infty$.
Assumption (A9) is verified in numerous examples of marked simple point processes, including the Poisson point process (PPP) with intensity $m\in (0,+\infty)$. Assumption (SEP) is of percolation nature. We show its validity for PPP's. Moreover, since one can consider as well other jump rates $c_{x,y}(\o)$ for a random walk on a marked simple point process, we state our percolation result in a more general form.
\begin{Proposition}\label{igro1} Suppose that under $\ensuremath{\mathcal P}$ the random set $\{x_i\}$ is a PPP with intensity $m\in (0,+\infty)$. Take
jump rates $c_{x,y}(\o)$ satisfying
\eqref{montagna} in (A4) and (A5). Suppose that, for $\ensuremath{\mathcal P}$--a.a. $\o$,
$c_{x,y}(\o) \leq g(|x-y|)$ for any $x,y \in \hat \o$, where $g(r)$ is a fixed bounded function such that
the map $ x \mapsto g(|x|) $ belongs to $L^1({\ensuremath{\mathbb R}} ^d, dx)$ (for example take $g(r)=e^{-r}$ for \eqref{vento}).
Then
Assumption (SEP) is satisfied.
\end{Proposition}
\begin{proof}
Note that ${\ensuremath{\mathbb P}} _\o ( \ensuremath{\mathcal K}_{x,y}(t) \geq 1)= 1- e^{-c_{x,y}(\o)t} \leq 1- \exp\{-g(|x-y|) t\}\leq C_1 g(|x-y|) t$ for some fixed $C_1>0$ if we take $t\leq 1$ (since $g$ is bounded). We restrict to $t$ small enough such that $C_1\|g\|_\infty t <1$ and $t\leq 1$.
Consider the random connection model \cite{MR} on a PPP with intensity $m$ where
an edge between $x\not =y$ is created with probability $C_1 g(|x-y|) t $.
Due to the independence of the Poisson processes $N_{x,y}(\cdot)'s$ given $\o$, one can couple the above random connection model with the field $(\o, \ensuremath{\mathcal K})$ with law ${\ensuremath{\mathbb P}} :=\int d \ensuremath{\mathcal P} (\o){\ensuremath{\mathbb P}} _\o$ in a way that the graph in the random connection model contains the graph $\ensuremath{\mathcal G} _{t}(\o,\ensuremath{\mathcal K})$.
We choose $t=t_0$ small enough to have $ m C_1 t_0 \int _{{\ensuremath{\mathbb R}} ^d} dx g(|x|) <1$.
The above bound and the branching process argument in the proof of \cite[Theorem~6.1]{MR} (cf. (6.3) there) imply that a.s. the random connection model has only connected components with finite cardinality. Hence the same must hold for $\ensuremath{\mathcal G} _{t_0}(\o,\ensuremath{\mathcal K})$
\end{proof}
\subsection{Simple exclusion processes on infinite clusters}
For completeness we give an example associated to the random geometric structure introduced in Section \ref{virgilio}. Recall that there $\ensuremath{\mathcal L}=(\ensuremath{\mathcal V},\ensuremath{\mathcal E})$ is the hexagonal lattice, $\O=\{0,1\}^{\ensuremath{\mathcal V}}$, $\ensuremath{\mathcal P}$ is a Bernoulli site percolation, $\hat \o = \ensuremath{\mathcal C}(\o)$ is the unique infinite percolation cluster inside $\ensuremath{\mathcal L}$ $\ensuremath{\mathcal P}$--a.s. We consider the simple exclusion process on $\ensuremath{\mathcal C}(\o)$ with $c_{x,y}(\o)=1$ if $x,y \in \ensuremath{\mathcal C}(\o)$ and $\{x,y\}\in \ensuremath{\mathcal E}$. The it is trivial to check that Assumptions (A1),$\dots$,(A9) and (SEP) are all satisfied.
\section{Random walk semigroup and resolvent convergence by homogenizaton}\label{figlio_stress}
In this section we recall the main results from \cite{Fhom3} which will be used in the proof of Theorem \ref{teo1}.
As in Proposition \ref{prop_ergodico} we introduce the atomic measure
\begin{equation}\label{franci}
\mu_\o^\e := \sum _{ x\in \hat \o} \e^d \d_{\e x}\,.\end{equation}
We also introduce the set (recall \eqref{clacson})
\begin{equation}\label{alba_chiara}
\begin{split}
\O_1:=\{\o\in \O\,:\, c_x(\o) <+\infty \; \forall x \in \hat \o\,,\;\;
c_{x,y}(\o)=c_{y,x}(\o) \;\forall x,y \in \hat \o\}\,.
\end{split}
\end{equation}
As explained in \cite[Section~3.3]{Fhom3}, the set $\O_1$ is translation invariant and satisfies $\ensuremath{\mathcal P}(\O_1)=1$.
Let us fix $\o \in \O_1$.
We call $\ensuremath{\mathcal C}(\e \hat \o)$ the space of local functions $f : \e \hat \o \to {\ensuremath{\mathbb R}} $ (here local means that $f$ has finite support).
We define
\[
\ensuremath{\mathcal D}_\o^\e:=\bigl\{ f \in L^2(\mu_\o^\e)\,:\,
\sum_{x\in\hat \o} \sum _{y\in \hat \o} c_{x,y}(\o) ( f(\e y) -f (\e x) )^2 <+\infty\bigr\}
\] and introduce the bilinear form
\[
\ensuremath{\mathcal E}_\o^\e (f,g) :=\frac{\e^{d-2} }{2} \sum _{x\in \hat \o} \sum _{y \in \hat \o } c_{x,y}(\o)
\bigl (f(\e y) - f (\e x ) \bigr)\bigl ( g(\e y) - g(\e x) \bigr)
\]
with domain $\ensuremath{\mathcal D}_\o ^\e$.
Since $\o \in \O_1$ it holds $C (\e \hat \o) \subset \ensuremath{\mathcal D} _\o ^\e$, as explained in \cite[Section~3.3]{Fhom3}.
We call $\ensuremath{\mathcal D}_{\o,*}^\e $ the closure of $C (\e \hat \o)$ w.r.t. the norm $\|f\|_{L^2(\mu^\e_\o)} + \ensuremath{\mathcal E}_\o^\e (f,f)^{1/2}$.
Then, as stated in \cite[Example~1.2.5]{FOT}, the bilinear form $ \ensuremath{\mathcal E}_\o^\e $ restricted to $\ensuremath{\mathcal D}_{\o,*}^\e $
is a regular Dirichlet form. In particular, there exists a unique nonpositive self-adjoint operator ${\ensuremath{\mathbb L}} ^\e_\o$ in $L^2(\mu^\e_\o)$ such that $\ensuremath{\mathcal D}_{\o,*}^\e $ equals the domain of $\sqrt{-{{\ensuremath{\mathbb L}} }^\e_\o}$ and $ \ensuremath{\mathcal E}_\o^\e (f,f) = \|\sqrt{-{\ensuremath{\mathbb L}} ^\e_\o} f\|^2_{L^2(\mu^\e_\o)}$ for any $f\in \ensuremath{\mathcal D}_{\o,*}^\e $ (see
\cite[Theorem~1.3.1]{FOT}). Due to \cite[Lemma~1.3.2 and Exercise~4.4.1]{FOT}, ${\ensuremath{\mathbb L}} ^\e_\o$ is the infinitesimal generator of
the strongly continuous Markov semigroup $(P^\e _{\o ,t} )_{t \geq 0}$ on $L^2(\mu^\e _\o)$ associated to the random walk $(\e X^\o _{ \e^{-2} t} )_{t\geq 0}$ on $\e \hat \o$
defined in terms of holding times and jump probabilities (see Lemma \ref{lemma_no_TNT}).
Hence, $P^\e _{\o ,t} f(x) = E_x\bigl[ f(\e X^\o _{ \e^{-2} t}) \bigr]$ for $f\in L^2(\mu^\e _\o)$ and $x\in \e\hat \o$, $E_x$ denoting the expectation when the random walk starts at $x$.
For completeness, although not used below, we report that (using that $\o \in \O_1$) one can check that
$\ensuremath{\mathcal C} (\e \hat \o)\subset \ensuremath{\mathcal D}( {\ensuremath{\mathbb L}} ^\e_\o)$ and that
${\ensuremath{\mathbb L}} ^\e_\o f(\e x) = \sum_{y \in \hat \o} c_{x,y}(\o) \left( f(\e y)-f(\e x) \right)$ for all $x \in \hat \o,\; \forall f\in \ensuremath{\mathcal C}(\e \hat \o)$
(the series in the r.h.s. is well defined being absolutely convergent).
We recall that we write $( P_t )_{t \geq 0} $ for the Markov semigroup associated to the Brownian motion $(B_t)_{t\geq 0}$ on ${\ensuremath{\mathbb R}} ^d$ with diffusion matrix $2 D$ given in Definition \ref{def_D} (strictly speaking it would be natural here to refer to the semigroup on $L^2(mdx)$ but $P_t$ will be applied below to bounded functions, hence one can keep the same definition of $P_t$ as for Theorem \ref{teo1}).
Recall that we write $(P^\e _{\o ,t} )_{t \geq 0}$ for the $L^2(\mu^\e _\o)$--Markov semigroup associated to the random walk $(\e X^\o _{ \e^{-2} t} )_{t\geq 0}$ on $\e \hat \o$. In particular, $P^\e _{\o ,t}=e^{t {\ensuremath{\mathbb L}} ^\e _\o }$. Given $\l>0$ we write $R^\e _{\o ,\l}: L^2(\mu^\e _\o)\to L^2(\mu^\e _\o) $ for the resolvent associated to the random walk $\e X^\o _{ \e^{-2} t} $, i.e. $R^\e _{\o ,\l}:= (\l -{\ensuremath{\mathbb L}} ^\e _\o )^{-1} =\int_0^\infty e^{- \l s} P^\e _{\o ,s} ds $. We write $ R_\l : L^2(m dx) \to L^2(m dx)$ for the resolvent associated to the above Brownian motion on ${\ensuremath{\mathbb R}} ^d$ with diffusion matrix $2 D$.
\begin{Proposition}\label{replay}\cite[Theorem 2]{Fhom3} Let Assumptions (A1),...,(A9) be satisfied.
Then there exists a translation invariant measurable
set $\O_\sharp\subset \O $ with $\ensuremath{\mathcal P}( \O_\sharp)=1$ such that for any $ \o \in \O_\sharp $, any $f \in C_c({\ensuremath{\mathbb R}} ^d)$, $\l>0$, $t\geq 0$ it holds:
\begin{align}
& \lim_{\e \downarrow 0} \int \bigl | P^\e_{\o,t} f(x) - P_t f (x) \bigr| d \mu^\e_\o (x)=0\,.\label{marvel2}\\
& \lim_{\e \downarrow 0} \int \bigl | R^\e_{\o,\l} f(x) -R_\l f (x) \bigr| d \mu^\e_\o (x)=0\,.\label{flavia}
\end{align}
\end{Proposition}
\begin{Remark} As stated in \cite[Remark~4.2]{Fhom3} Assumption (A9) is used in \cite{Fhom3} only to prove for $\ensuremath{\mathcal P}$--a.a. $\o$ that
\begin{equation} \label{claudio2}
\lim _{\ell \uparrow \infty} \varlimsup_{\e \downarrow 0} \int d \mu^\e _\o (x)\psi(|x|) \mathds{1}_{\{ |x| \geq \ell\}}=0\,,
\qquad \psi(r):=1/(1+ r^{d+1})
\,. \end{equation}
Hence, in Theorem \ref{teo1} one could replace (A9) by any other condition leading to the above property.
\end{Remark} For later use, we also point out that
the $\o$'s satisfying \eqref{claudio2} form a translation invariant measurable set.
\section{Graphical construction of the simple exclusion process}\label{sec_GC}
Let $t_0=t_0(\o)$ be as in Assumption (SEP). Recall definition \eqref{pitagora} of $\ensuremath{\mathcal E}_\o$.
\begin{Definition}[Property $(P_r)$] \label{def_pr} Given $r\in {\ensuremath{\mathbb N}} $ we say that the pair $(\o, \ensuremath{\mathcal K})\in \O\times D({\ensuremath{\mathbb R}} _+,{\ensuremath{\mathbb N}} )^{\ensuremath{\mathcal E}_\o}$ satisfies property $(P_r)$ if
the
undirected graph $\ensuremath{\mathcal G} ^r_{t_0}(\o,\ensuremath{\mathcal K})$ with vertex set $\hat \o$ and edge set $\{ \{x,y\}
\in \ensuremath{\mathcal E}_\o
\,:\, \ensuremath{\mathcal K}_{x,y}((r+1) t_0)> \ensuremath{\mathcal K}_{x,y} (r t_0) \}$ has only connected components with finite
cardinality.
\end{Definition}
\begin{Definition}[Set $\tilde \O$] \label{omesso} The set $\tilde \O$ is given by the elements $\o\in \O$ such that $c_x (\o)<+\infty \;\forall x \in \hat \o$ and such that the properties in Assumptions (A4) and (A5) are fulfilled (namely, \eqref{base}, \eqref{montagna}, \eqref{pietra} hold for all $x,y,g$).
\end{Definition}
As already pointed out in Section \ref{figlio_stress}, $\O_1$ is a translation invariant set and $\ensuremath{\mathcal P}(\O_1)=1$. It is trivial to check that the same holds for $\tilde \O\subset \O_1$.
\begin{Definition}[Sets ${\ensuremath{\mathbb K}} _\o$, $\O_*$] \label{def_omega_*}
Fixed $\o\in \O$, ${\ensuremath{\mathbb K}} _\o$ is the set given by the elements $\ensuremath{\mathcal K}\in D({\ensuremath{\mathbb R}} _+,{\ensuremath{\mathbb N}} )^{\ensuremath{\mathcal E}_\o}$ such that
\begin{itemize}
\item[(i)] $(\o,\ensuremath{\mathcal K})$ satisfies property $(P_r)$ for all $r\in {\ensuremath{\mathbb N}} $;
\item[(ii)] the jump time sets $\{ t>0\,: \ensuremath{\mathcal K}_{x,y}(t-)\not = \ensuremath{\mathcal K}_{x,y} (t) \}$ are disjoint as $\{x,y\}$ varies among $\ensuremath{\mathcal E}_\o$;
\item[(iii)] $\ensuremath{\mathcal K}_x(t):=\sum_{y: \{x,y\}\in \ensuremath{\mathcal E}_\o} \ensuremath{\mathcal K}_{x,y}(t) <+\infty$ for all $x\in \hat \o$ and $t\geq 0$.
\end{itemize}
We define $\O_*$ as the set of $\o \in \tilde \O$ such that ${\ensuremath{\mathbb P}} _\o( {\ensuremath{\mathbb K}} _\o)=1$.\end{Definition}
Since $\ensuremath{\mathcal P}(\tilde \O)=1$ and by the loss of memory of the Poisson point process, we have that $\ensuremath{\mathcal P}(\O_*)=1$. It is simple to check that $\O_*$ is translation invariant.
Also for later use, we now recall
the graphical construction of the simple exclusion process. To this aim it is convenient to think the simple exclusion process as an exchange process.
Let us fix $\o \in \O_*$ and $\ensuremath{\mathcal K}\in {\ensuremath{\mathbb K}} _\o$.
Given a particle configuration $\xi \in \{0,1\}^{\hat \o}$ we now define a deterministic trajectory $(\eta^\xi_t[\ensuremath{\mathcal K}])_{t \geq 0 }$ in $D({\ensuremath{\mathbb R}} _+,\{0,1\}^{\hat \o} )$ and starting at $\xi$ by an iterative procedure. We set
$\eta^\xi_0[ \ensuremath{\mathcal K}]:=\xi$.
Suppose that the deterministic trajectory has been defined up to time $r t_0$, $r\in {\ensuremath{\mathbb N}} $ (note that for $r=0$ this follows from our definition of $\eta^\xi_0[ \ensuremath{\mathcal K}]$).
As $\ensuremath{\mathcal K} \in {\ensuremath{\mathbb K}} _\o$ all connected components of $\ensuremath{\mathcal G}^r_{t_0}(\o,\ensuremath{\mathcal K})$ have finite cardinality.
Let $\ensuremath{\mathcal C}$ be
such a connected component and let
\begin{multline*}
\{s_1<s_2< \cdots <s_k\} =\\
\bigl \{s \,: \ensuremath{\mathcal K}_{x,y}(s) = \ensuremath{\mathcal K}_{x,y}(s-)+1\,, \;\{x,y\} \text{ bond in } \ensuremath{\mathcal C}, \; r t_0 <s \leq (r+1) t_0\bigr\}\,.
\end{multline*}
As $\ensuremath{\mathcal K}\in {\ensuremath{\mathbb K}} _\o$, the l.h.s. is indeed a finite set. The local evolution $\eta^\xi _t[ \ensuremath{\mathcal K}](z) $ with $z \in \ensuremath{\mathcal C}$ and $r t_0 < t \leq (r+1) t_0$ is described as follows.
Start with $\eta^\xi_{rt_0}[ \ensuremath{\mathcal K}]$ as configuration at time $r t_0$ in $\ensuremath{\mathcal C}$. At time $s_1$ exchange the values between $\eta(x)$ and $\eta(y)$ if $\ensuremath{\mathcal K}_{x,y}(s_1)= \ensuremath{\mathcal K} _{x,y}(s_1-)+1$ and $\{x,y\}$ is an edge in $\ensuremath{\mathcal C}$ (there is exactly one such edge as $\ensuremath{\mathcal K} \in {\ensuremath{\mathbb K}} _\o$). Repeat the same operation orderly for times $s_2,s_3, \dots , s_k$. Then move to another connected component of $\ensuremath{\mathcal G}_{t_0}^r(\o,\ensuremath{\mathcal K})$ and repeat the above construction and so on. As
the connected components are disjoint, the resulting path does not
depend on the order by which we choose the connected components in the above algorithm.
This procedure defines $ \eta^\xi _t[\ensuremath{\mathcal K}]_{ r t_0< t \leq (r+1) t_0}$.
Starting with $r=0$ and progressively increasing $r$ by $1$ we get the trajectory $ \eta^\xi_t[\ensuremath{\mathcal K}]_{t\geq 0}$.
\smallskip
We recall that $C(\{0,1\}^{\hat \o})$ is the space of continuous functions on $\{0,1\}^{\hat \o}$.
Given $\o\in \O_*$ we consider
the probability space $( {\ensuremath{\mathbb K}} _\o, {\ensuremath{\mathbb P}} _\o)$, and write ${\ensuremath{\mathbb E}} _\o$ for the associated expectation. We set
\[
S(t) f (\xi) := {\ensuremath{\mathbb E}} _\o [ f( \eta^\xi _t [\ensuremath{\mathcal K}])]\,, \qquad t\geq 0\,,\; f \in C(\{0,1\}^{\hat \o})\,.
\]
\begin{Proposition}\label{prop_SEP} Take $\o \in \O_*$ and
fix $\xi \in \{0,1\}^{\hat \o}$. Then the random trajectory $\bigl( \eta^\xi_t [\ensuremath{\mathcal K}]\bigr)_{t\geq 0}$ with $\ensuremath{\mathcal K}$ sampled in the probability space $( {\ensuremath{\mathbb K}} _\o, {\ensuremath{\mathbb P}} _\o)$ belongs to the Skohorod space $D( {\ensuremath{\mathbb R}} _+, \{0,1\}^{\hat \o})$ and it starts at $\xi$. It describes a Feller process, called simple exclusion process. In particular, $( S(t) )_{t\geq 0}$ is a Markov semigroup on $C(\{0,1\}^{\hat \o})$. Moreover, the domain of its infinitesimal generator $\ensuremath{\mathcal L}_\o$ contains the family of local functions and for any local function $f$ the function $\ensuremath{\mathcal L}_\o f$ is given by the right hand sides of \eqref{mammaE} and \eqref{mahmood}, which are absolutely convergent series in $\ensuremath{\mathcal C}(\{0,1\}^{\hat \o})$.
\end{Proposition}
The above proposition can be derived by the standard arguments used for the graphical construction of the SEP usually presented under the assumption of finite range jumps (see e.g. \cite[Section 2.1]{timo}).
The only exception is given by the
derivation of the identities \eqref{mammaE} and \eqref{mahmood} for local functions, due to possible unbounded jump range.
We refer to Appendix \ref{app_localino} for the proof of \eqref{mammaE} and \eqref{mahmood}.
\section{Duality}\label{dualino}
In order to prove the tightness of the empirical measure by means of the corrected empirical one, we need to deal with non local functions on $\{0,1\}^{\hat \o}$. In what follows we collect the extended results concerning $\ensuremath{\mathcal L}_\o$ and Dynkin martingales, which will be used in Section \ref{sec_teso}. Recall \eqref{clacson}.
In all this section we restrict to $\o\in \O_*$ (cf.~Definition \ref{def_omega_*}).
\begin{Definition}\label{birillino}
Given a function $u : \e \hat \o \to {\ensuremath{\mathbb R}} $ such that $\sum _{x\in \hat \o} c_x (\o) |u(\e x) |<+\infty$, we define
$\tilde {\ensuremath{\mathbb L}} ^\e_\o u(x) :=\e^{-2} \sum _{y\in \hat\o} c_{x,y} (\o) ( u(\e y) - u(\e x) )$.\end{Definition}
By symmetry of the jump rates we have
\begin{equation}\label{fischio0}
\sum _{x\in \hat \o} \sum _{y\in \hat \o} c_{x,y} (\o) (| u(\e x)| + |u(\e y) |) =2 \sum _{x\in \hat \o} c_x(\o) | u(\e x) |\,.
\end{equation}
Hence, if $\sum _{x\in \hat \o } c_x (\o) |u(\e x) |<+\infty$, the series defining $\tilde {\ensuremath{\mathbb L}} ^\e_\o u(x)$ is
absolutely convergent for any $x\in \hat \o$.
\medskip
In what follows, to simplify the notation, we write $\pi^\e_\o(u)$, $\pi^\e_{\o,t}(u)$
for the integral of $u$ w.r.t. $\pi^\e_\o[\eta]$, $\pi^\e_{\o,t}[\eta]$, respectively. Recall that ${\ensuremath{\mathbb L}} ^\e_\o$ is the Markov generator in $L^2(\mu^\e_\o)$ of the random walk $(\e X_{\e^{-2} t}^\o)_{t\geq 0}$ (see Section \ref{figlio_stress}).
We now state two lemmas which will be crucial when dealing with the corrected empirical measure. We postpone their proofs to the end of the section.
\begin{Lemma}[Duality] \label{ringo}
Suppose that $u : \e \hat \o \to {\ensuremath{\mathbb R}} $ satisfies
\begin{equation} \label{norma}
\sum_{x\in \hat \o } |u(\e x)| <+\infty \text{ and }
\sum_{x\in \hat \o } c_x (\o) |u(\e x)| <+\infty\,.
\end{equation}
Then $ \pi^{\e} _{\o}(u) =\e^d\sum_{x\in \hat \o } u(\e x) \eta(x)$ is an absolutely convergent series in $ C(\{0,1\}^{\hat \o})$. It belongs to the domain $\ensuremath{\mathcal D}(\ensuremath{\mathcal L}_\o) \subset C(\{0,1\}^{\hat \o})$ of $\ensuremath{\mathcal L}_\o$ and
\begin{equation}\label{airone25} \ensuremath{\mathcal L} _\o \left( \pi^\e_\o (u)\right)=
\e^{d+2} \sum _{x\in \hat \o} \eta(x) \tilde {\ensuremath{\mathbb L}} ^\e_\o u (\e x) \,,
\end{equation}
the r.h.s. of \eqref{airone25} being an absolutely convergent series in $C(\{0,1\}^{\hat \o})$.
If, in addition to \eqref{norma}, it holds $u \in \ensuremath{\mathcal D}( {\ensuremath{\mathbb L}} ^\e_\o )\subset L^2(\mu^\e_\o)$, then ${\ensuremath{\mathbb L}} ^\e_\o u = \tilde {\ensuremath{\mathbb L}} ^\e_\o u $ and in particular we have the duality relation
\begin{equation}\label{airone26} \ensuremath{\mathcal L} _\o \left( \pi^\e_\o (u)\right)=
\e^{d+2} \sum _{x\in \hat \o} \eta(x) {\ensuremath{\mathbb L}} ^\e_\o u (\e x) \,.
\end{equation}
\end{Lemma}
Let $u:\e \hat \o \to {\ensuremath{\mathbb R}} $ be a function satisfying \eqref{norma}. As, by Lemma \ref{ringo}, $\pi^{\e} _{\o}(u) \in \ensuremath{\mathcal D}(\ensuremath{\mathcal L}_\o)$, we can introduce on the Skohorod space $D\bigl({\ensuremath{\mathbb R}} _+, \{0,1\}^{\hat \o}\bigr)$ the Dynkin martingale $(\ensuremath{\mathcal M}^\e_{\o,t})_{t\geq 0}$ given by (see e.g. \cite[Appendix~1]{KL} or \cite[Section 3.2]{timo})
\begin{equation} \label{tremo}
\ensuremath{\mathcal M}^\e_{\o,t} : = \pi^{\e} _{\o,t}(u) - \pi^{\e} _{\o,0}(u) -\e^{-2} \int _0 ^t \ensuremath{\mathcal L}_\o \left( \pi^{\e} _{\o}(u) \right)(\eta_s) ds\,.
\end{equation}
$(\ensuremath{\mathcal M}^\e_{\o,t})_{t\geq 0}$ is square integrable martingale w.r.t. the filtered probability space $\left( D\bigl({\ensuremath{\mathbb R}} _+, \{0,1\}^{\hat \o}\bigr), {\ensuremath{\mathbb P}} ^\e_{\o,\mathfrak{n}_\e }, (\ensuremath{\mathcal F}_t)_{t\geq 0}\right)$, $\mathfrak{n}_\e$ being an arbitrary initial distribution and $\ensuremath{\mathcal F}_t$ being the $\s$--field generated $\{\eta_s:0\leq s\leq t\}$. Square integrability follows from the property that $\| \ensuremath{\mathcal M}^\e_{\o,t} \|_\infty<+\infty$ as the same holds for all addenda in the r.h.s. of \eqref{tremo} (see Lemma \ref{ringo}).
\begin{Lemma}\label{star}
Let $u:\e \hat \o \to {\ensuremath{\mathbb R}} $ be a function satisfying \eqref{norma}.
Suppose in addition that $ \sum_{x\in \hat \o} c_x(\o) u(\e x )^2 <+\infty$.
Then the sharp bracket process of $\ensuremath{\mathcal M}^\e_{\o,t}$ is given by
$\langle \ensuremath{\mathcal M}^\e_{\o}\rangle _t= \int _0 ^t B^\e_{\o} (\eta_s)ds $, where
\begin{equation}\label{compenso_IV}
B^\e_{\o}(\eta) =
\e^{2d-2} \sum _{x\in \hat \o} \sum _{y \in \hat \o} c_{x,y}(\o)\bigl[u(\e x)-u(\e y)\bigr]^2
\eta (x) \bigl (1- \eta (y) \bigr)\,.
\end{equation}
\end{Lemma}
Note that the bound $\sum_{x\in \hat \o} c_x(\o) u(\e x )^2 <+\infty
$ implies that the r.h.s. of \eqref{compenso_IV} is an absolutely convergent series of functions in $C(\{0,1\}^{\hat \o})$. For later use, we recall that $\langle \ensuremath{\mathcal M}^\e_{\o}\rangle _t$ can be characterized as
the unique predictable increasing process such that $(\ensuremath{\mathcal M}^\e_{\o,t})^2- \langle \ensuremath{\mathcal M}^\e_{\o}\rangle _t$ is a martingale \cite[Theorem~8.24]{Kl}.
\begin{Remark} In the proof of Theorem \ref{teo1} (see Section \ref{sec_teso}) we will apply the above Lemmas \ref{ringo} and \ref{star} just to functions $u$ of the form $R^\e_{\o,\l } \psi$ for suitable functions $\psi \in C_c ({\ensuremath{\mathbb R}} ^d)$.
\end{Remark}
\begin{proof}[Proof of Lemma \ref{ringo}]
As $\sum_{x\in \hat \o } |u(\e x) | <+\infty$, it is simple to check that the series defining
$\pi^{\e} _{\o}(u)$ is indeed an absolutely convergent series of continuous functions w.r.t. the uniform norm. The same holds for the series corresponding to the r.h.s. of \eqref{airone25}. Indeed, by \eqref{fischio0}, $ \sum _{x\in \hat \o} \| \eta(x) \tilde {\ensuremath{\mathbb L}} ^\e_\o u (\e x) \|_\infty \leq 2 \e^{-2} \sum_{x\in \hat \o } c_x(\o) |u(\e x) | <+\infty$.
When the function $u$ is local, also the map $\eta\mapsto \pi^\e_\o (u)$ is local.
By locality and Proposition \ref{prop_SEP}, this map belongs to $\ensuremath{\mathcal D}(\ensuremath{\mathcal L}_\o) $. In the case of local $u$, \eqref{airone25} follows from easy computations by \eqref{mahmood}.
We now treat the general case.
Given $n\in {\ensuremath{\mathbb N}} $, we define $u_n(\e x ):= u(\e x ) \mathds( |\e x | \leq n) $. As observed above,
$\pi^{\e} _{\o}(u_n)$ is a local function on $\{0,1\}^{\hat \o}$ belonging to $\ensuremath{\mathcal D}(\ensuremath{\mathcal L}_\o)$ and \eqref{airone25} holds with $u_n$ instead of $u$.
We claim that
\begin{align}
& \lim_{n \to \infty} \| \pi^{\e} _{\o}(u_n) - \pi^{\e} _{\o}(u) \|_\infty=0\,, \label{fischio1}\\
& \lim_{n \to \infty}\| \sum _{x\in \hat \o} \eta(x) \tilde {\ensuremath{\mathbb L}} ^\e_\o u_n (\e x) - \sum _{x\in \hat \o} \eta(x) \tilde {\ensuremath{\mathbb L}} ^\e_\o u (\e x) \|_\infty =0\,.\label{fischio2}
\end{align}
As $\ensuremath{\mathcal L}_\o$ is a closed operator being a Markov generator, \eqref{airone25} with $u_n$ instead of $u$,
\eqref{fischio1} and \eqref{fischio2} imply that $ \pi^{\e} _{\o}(u)\in \ensuremath{\mathcal D}(\ensuremath{\mathcal L}_\o)$ and that \eqref{airone25} holds.
To prove \eqref{fischio1} and \eqref{fischio2} it is enough to bound the uniform norms appearing there by, respectively, $\e^d\sum_{x\in \hat \o: |\e x|>n}| u(\e x) |$ and $2 \e^{-2}\sum_{x\in \hat \o: |\e x|>n} c_x(\o) |u(\e x) |$ and use \eqref{norma}. This concludes the proof of \eqref{airone25}.
It remains to show that ${\ensuremath{\mathbb L}} ^\e_\o u = \tilde {\ensuremath{\mathbb L}} ^\e_\o u $ if $u \in \ensuremath{\mathcal D}( {\ensuremath{\mathbb L}} ^\e_\o )\subset L^2(\mu^\e_\o)$ in addition to \eqref{norma}. Given a function $f\in C(\{0,1\}^{\hat \o})$ we write $S(t) f (\eta ):={\ensuremath{\mathbb E}} _\eta [ f(\eta_t) ]$ for the Markov semigroup associated to the simple exclusion process (without any time rescaling). Then \eqref{airone25} can be read as
\begin{equation}\label{bibita}
\lim _{t\downarrow 0}\sup_{\eta \in \{0,1\}^{\hat \o} }\Big| \frac{\left(S(t) \pi^{\e} _{\o}(u)\right)(\eta)- \pi^{\e} _{\o}[\eta](u)}{t}- \e^{d+2} \sum _{x\in \hat \o} \eta(x) \tilde {\ensuremath{\mathbb L}} ^\e_\o u (\e x)\Big|=0\,.
\end{equation}
Given $x_0 \in \hat \o$ we take $\eta$ corresponding to a single particle located at $x_0$. Then
$\left(S(t) \pi^{\e} _{\o}(u)\right)(\eta)= \e^d E_{x_0} [ u ( \e X^{\o}_t) ]$ and
\eqref{bibita} implies that
$\frac{d}{dt} E_{x_0} [ u ( \e X^{\o}_{\e^{-2}t}) ]_{|t=0}= \tilde {\ensuremath{\mathbb L}} ^\e_\o u (\e x_0)$.
On the other hand, we know that $u \in \ensuremath{\mathcal D}({\ensuremath{\mathbb L}} ^\e_\o)$. Hence
\begin{equation}
\lim _{t\downarrow 0}\sum_{x\in \hat \o} \Big| \frac{ E_{x} [ u ( \e X^{\o}_{\e^{-2} t})] - u(x)}{t}- {\ensuremath{\mathbb L}} ^\e_\o u (\e x)\Big|^2 =0,
\end{equation}
which implies that $\frac{d}{dt} E_{x_0} [ u ( \e X^{\o}_{\e^{-2}t}) ]_{|t=0}= {\ensuremath{\mathbb L}} ^\e_\o u (\e x_0)$. Then it must be ${\ensuremath{\mathbb L}} ^\e_\o u (\e x_0)=\tilde {\ensuremath{\mathbb L}} ^\e_\o u (\e x_0)$.
\end{proof}
\begin{proof}[Proof of Lemma \ref{star}]
For $u$ local both $\pi^\e_\o (u)$ and its square belong to $\ensuremath{\mathcal D}(\ensuremath{\mathcal L}_\o)$ being local functions of $\eta$. Then the statement in the lemma can be checked by simple computations due to \eqref{mahmood}, Lemma \ref{ringo} and \cite[Lemma~5.1, App.~1]{KL} (equivalently, \cite[Exercise~3.1 and Lemma~8.3]{timo}). For the computation of the sharp bracket process we just comment that, by using the symmetry of $c_{x,y}(\o)$, one easily gets
\[\ensuremath{\mathcal L} _\o ( \pi^\e_\o (u)^2 )-
2 \pi^\e_\o (u) \ensuremath{\mathcal L} _\o ( \pi^\e_\o (u) ) = \e^{2d} \sum _{x\in \hat \o} \sum _{y \in \hat \o} c_{x,y}(\o) [ u(\e x)- u (\e y)]^2
\eta(x) (1- \eta(y)).
\]
We now move to the general case. For simplicity of notation we write $\ensuremath{\mathcal M}_t$, $B(\eta)$ instead of $\ensuremath{\mathcal M}^\e_{\o,t}$, $B^\e_\o(\eta)$. Similarly, we define $\ensuremath{\mathcal M}_{n,t} $ and $B_n(\eta)$ as in \eqref{tremo} and \eqref{compenso_IV} with $u$ replaced by $u_n$, $u_n(\e x):= u(\e x) \mathds{1}(|\e x| \leq n)$. Note that $\lim _{n \to +\infty}\sup_{t\in [0,T]} \| \ensuremath{\mathcal M}_t - \ensuremath{\mathcal M} _{n,t}\|_\infty =0$ for any $T>0$ (see \eqref{airone25}, \eqref{fischio1} and \eqref{fischio2}).
Hence, by the characterization of the sharp bracket process recalled after Lemma \ref{star} and by our results for the local case (applied to $u_n$), to get \eqref{compenso_IV} it is enough to show that $\lim_{n\to \infty} \|B_n(\eta)-B(\eta) \|_\infty =0$. To this aim it is enough to show that $\sum _{x\in \hat \o} \sum _{y \in \hat \o} c_{x,y}(\o)\bigl[ u_n(\e x)- u_n (\e y)\bigr]^2$ converges, as $n\to \infty$, to the analogous expression with $u$ instead of $u_n$. This follows from the dominated convergence theorem, by dominating $\bigl[ u_n(\e x)- u_n (\e y)\bigr]^2$ with $2 u(\e x)^2 + 2 u(\e y)^2$ and by using that $\sum _{x\in \hat \o} \sum _{y \in \hat \o} c_{x,y}(\o)\bigl[ u(\e x)^2 + u(\e y)^2\bigr]=2 \sum_{x\in \hat \o} c_x(\o) u(\e x)^2<+\infty $.
\end{proof}
\section{Space $\ensuremath{\mathcal M}$ of Radon measures and Skohorod space $D([0,T], \ensuremath{\mathcal M})$}\label{sec_mammina}
Given a measure $\mu$ on ${\ensuremath{\mathbb R}} ^d$ and a real function $G$ on ${\ensuremath{\mathbb R}} ^d$, we will denote by $\mu(G)$ the integral $\int d\mu(x) G(x)$.
We denote by $\ensuremath{\mathcal M}$ the space of Radon measures on ${\ensuremath{\mathbb R}} ^d$, i.e. locally bounded Borel measures on ${\ensuremath{\mathbb R}} ^d$.
$\ensuremath{\mathcal M}$ is endowed with the vague topology, for which $\mu_n \to \mu$ if and only if $\mu_n(f) \to \mu(f)$ for all $f\in C_c({\ensuremath{\mathbb R}} ^d)$. This topology can be defined through a metric, that we now recall also for later use (for more details, see e.g.~\cite[Appendix~A.10]{timo}). To this aim we set $B_r:=\{x\in {\ensuremath{\mathbb R}} ^d: |x| \leq r\}$.
For each $\ell\in {\ensuremath{\mathbb N}} $ we choose a sequence of functions $(\varphi_{\ell, n})_{n \geq 0}$ such that\footnote{Some of our requirements will be used to prove the hydrodynamic behavior for $\ensuremath{\mathcal P}$--a.a. $\o$ and are not strictly necessary to define the metric on $\ensuremath{\mathcal M}$. }
\begin{itemize}
\item[(i)] $\varphi_{\ell, n}
\in C^\infty_c ({\ensuremath{\mathbb R}} ^d)$ and $\varphi_{\ell, n}$ is supported on $B_{\ell+1}$;
\item[(ii)] the family
$(\varphi_{\ell, n})_{n \geq 0}$ contains a function with values in $[0,1]$, equal to $1$ on $B_\ell$ and equal to $0$ outside $B_{\ell+1}$;
\item[(iii)] for each $\d>0$ and $\varphi \in C^\infty_c({\ensuremath{\mathbb R}} ^d)$ with support in $B_\ell$ there exists $n\geq 0$ such that $\| \varphi_{\ell, n} - \varphi\|_\infty \leq \d$ and $\sup_{i,k=1}^d\| \partial^2_{x_i, x_k} \varphi_{\ell, n} -\partial^2_{x_i,x_k} \varphi\|_\infty \leq \d$.
\end{itemize}
For the existence of such a set of functions $\varphi_{\ell, n} $ we refer \cite[Appendix~A]{timo} and discuss only Item (iii) which is in part new. To deal with Item (iii) we use an extended version of the classical Weierstrass approximation theorem (see \cite[Theorem~1.6.2]{Nara}) implying that, for any compact set $K$, the family $\ensuremath{\mathcal P}$ of polynomial functions with rational coefficients is dense in $C^2({\ensuremath{\mathbb R}} ^d)$ w.r.t. to the semi-norm $\| f\|_K := \|f\|_{L^\infty(K)}+\sum_{i=1}^d \| \partial_{x_i} f\|_{L^\infty(K)}+ \sum_{i,j=1}^d \|\partial^2_{x_i,x_j} f\|_{L^\infty (K)}$. For each $\ell$ fix a function $g_\ell$ as in Item (ii). Given $\varphi \in C^\infty_c({\ensuremath{\mathbb R}} ^d)$ as in Item (iii), by applying Leibniz rule to $\varphi-g_\ell f= g_\ell (\varphi-f)$, one easily gets that
$\| \varphi - g_\ell f\|_{B_{\ell+1}}\leq C(d) \| g_\ell \|_{B_{\ell+1}} \| \varphi - f\|_{B_{\ell+1}}$. Hence, to fulfill Item (iii), it is enough to include into $\{\varphi_{\ell,n}\}$ the countable family of functions $\{ g_\ell f\, : \, f\in \ensuremath{\mathcal P}\}$.
\begin{Definition}\label{tuono20}
By a relabeling, we write $( \varphi_j)_{j\in {\ensuremath{\mathbb N}} } $ for the family $ ( \varphi_{\ell, n})_{ \ell, n \in {\ensuremath{\mathbb N}} }$.
\end{Definition}
On $\ensuremath{\mathcal M}$ we define the metric $d_\ensuremath{\mathcal M}$ as
$d_\ensuremath{\mathcal M} (\mu, \nu):= \sum_{j=0}^\infty 2^{-j} \left( 1\wedge | \mu(\varphi_j) -\nu(\varphi_j) | \right)$.
It can be proved that $(\ensuremath{\mathcal M},d_\ensuremath{\mathcal M})$ is a Polish space and that the topology induced by the metric $d_\ensuremath{\mathcal M}$ coincides with the vague topology (see e.g.~\cite[Appendix~A.10]{timo}, \cite{DV}).
We write $D([0,T],\ensuremath{\mathcal M})$ for the Skohorod space of $\ensuremath{\mathcal M}$--valued c\`adl\`ag paths $(\mu_t)_{0\leq t \leq T}$. We recall (cf. \cite[Section 4.1]{KL}) that $D([0,T],\ensuremath{\mathcal M})$ is a Polish space endowed with the metric
\begin{equation}
d\left( \mu_\cdot, \nu_\cdot\right)
:= \inf _{\l \in \L} \max \Big\{ \|\l\|, \sup_{0 \leq t \leq T} d_\ensuremath{\mathcal M} \bigl( \mu_t,\nu_t \bigr) \Big\}\,,
\end{equation}
where $\L$ is the set of strictly increasing continuous functions $\l:[0,T]\to [0,T]$ with $\l(0)=0$, $\l(T)=T$,
and
$\|\l\|:=\sup_{s\not = t} \big |\ln[(\l(t)-\l(s))/(t-s)] \big|$. As a subset $A\subset \ensuremath{\mathcal M}$ is relatively compact if and only if $\sup\{\mu(K) \,:\, \mu \in A\}<+\infty$ for any compact set $K\subset {\ensuremath{\mathbb R}} ^d$ (cf. \cite[App.~A]{timo}),
by the same arguments used in the proof of \cite[Prop.~1.7, Chapter~4]{KL} one gets the following:
\begin{Lemma}\label{proiettore} Given an index set $\ensuremath{\mathcal A}$, a family of probability measures $\{ Q^\a\}_{\a \in \ensuremath{\mathcal A}}$ on $D([0,T],\ensuremath{\mathcal M})$ is relatively compact (w.r.t. to weak convergence) if and only if for any $j\in {\ensuremath{\mathbb N}} $ the family of probability measures $\{ Q^\a\circ \Phi_j ^{-1} \}_{\a \in \ensuremath{\mathcal A}}$ on $D([0,T],{\ensuremath{\mathbb R}} )$ is relatively compact, where
\begin{equation}
\Phi_j: D([0,T],\ensuremath{\mathcal M})\ni (\mu_t )_{0\leq t\leq T} \mapsto (\mu_t(\varphi_j) )_{0\leq t\leq T}\in D([0,T],{\ensuremath{\mathbb R}} )\,.
\end{equation}
\end{Lemma}
Recall that $B_r:=\{x\in {\ensuremath{\mathbb R}} ^d\,:\,|x|\leq r\}$.
The following fact can be obtained by suitably modifying and afterwards extend the proof of \cite[Theorem~A.28 in Appendix~A]{timo}. W.r.t. the version in \cite{timo}, we have removed the assumption of non-degenerate diffusion matrix and we have modified the mass bounds.
\begin{Lemma}\label{timau} Let $v_0: {\ensuremath{\mathbb R}} ^d\to {\ensuremath{\mathbb R}} $ be Borel and bounded. Let $\a: [0,T]\to \ensuremath{\mathcal M}$ be a map such that
\begin{itemize}
\item[(i)] $\a$ is continuous when $\ensuremath{\mathcal M}$ is endowed with the vague topology;
\item[(ii)] $\a_0 (dx)=v_0(x) dx$;
\item[(iii)] for all $\varphi \in C_c^\infty ({\ensuremath{\mathbb R}} ^d)$ and $t \in [0,T]$ it holds
\begin{equation}
\int _{{\ensuremath{\mathbb R}} ^d} \varphi(x) \a_t( dx)= \int _{{\ensuremath{\mathbb R}} ^d} \varphi(x) \a_0(dx) +\int_0^t ds \int_{{\ensuremath{\mathbb R}} ^d} \nabla \cdot ( D \nabla \varphi)(x) \a_s(dx) \,;
\end{equation}
\item[(iv)] for some constants $C,r_0,\g>0$ it holds $ \a_s( B_r) \leq C r^\g$ for all $s\in [0,T]$ and all $r\geq r_0$.
\end{itemize}
Then $\a_t(dx)=P_t v_0(x) dx$ for all $t\in [0,T]$.
\end{Lemma}
\begin{proof} We distinguish two cases according to the non-degeneracy of $D$.
$\bullet$ \emph{Case $D$ non-degenerate}. The proof is the same of \cite[Theorem~A.28]{timo} apart of modifying \cite[Eq.~(A.40)]{timo}. To this aim, as in \cite{timo}, let $f\in C_c^\infty({\ensuremath{\mathbb R}} ^d)$ be a nonnegative, symmetric function with $\int_{{\ensuremath{\mathbb R}} ^d} f(x)dx=1$. Set $f^\epsilon(x):=\epsilon^{-d} f(x/\epsilon)$ and $v^\epsilon (x,t):=\int_{{\ensuremath{\mathbb R}} ^d} f^\epsilon(x-y)\a_t(dy)$. Then \cite[Eq.~(A.40)]{timo} has to be replaced
by the bound (for $0<\epsilon\leq 1$)
\[ |v^\epsilon (x,t) |\leq \|f^\epsilon\|_\infty \a_t (x+ B_1) \leq \|f^\epsilon \|_\infty \a_t( B_{r_0+1+|x|}) \leq C \|f \|_\infty \epsilon^{-d}
(r_0+1+|x|)^\g\,,
\]
which holds uniformly in $t\in [0,T]$ due to Item (iv).
The above bound is enough to apply \cite[Theorem~A.30]{timo} (which is a byproduct of \cite[Theorems~1 and 7, Section~2.3]{evans}). Then one can proceed and conclude as in \cite{timo}.
$\bullet$ \emph{Case $D$ degenerate}.
Without loss, at cost of a linear change of coordinates, we can assume that $D$ is diagonal with strictly positive eigenvalues on $e_1, e_2, \dots, e_{d_*}$, and zero eigenvalue on $e_{d_*+1}, \dots, e_d$ ($e_1,\dots, e_d$ being the canonical basis).
By writing $p_t(\cdot,\cdot )$ for the probability transition kernel of the Brownian motion on ${\ensuremath{\mathbb R}} ^{d_*}$ with non--degenerate diffusion matrix $2 \tilde D:=(2 D_{i,j})_{1\leq i,j \leq d_*}$, it holds
\begin{equation}\label{pesciolini}
P_t v_0(x',x'') = \int _{{\ensuremath{\mathbb R}} ^{d_*}} p_t ( x', z') v_0( z', x'' ) dz'\,\qquad (x', x'')\in {\ensuremath{\mathbb R}} ^{d_*} \times {\ensuremath{\mathbb R}} ^{d-d_*}= {\ensuremath{\mathbb R}} ^d\,.
\end{equation}
Given $\psi \in C_c ^\infty({\ensuremath{\mathbb R}} ^{d-d_*})$ with $\psi\geq 0$, we define $\tilde \a_t (dx')$ as the measure on ${\ensuremath{\mathbb R}} ^{d_*}$ given by
$ \tilde \a_t (B):= \int _{{\ensuremath{\mathbb R}} ^d} \mathds{1}_{B}(x') \psi(x'') \a_t(d x', d x'')$ for all Borel $ B\subset {\ensuremath{\mathbb R}} ^{d_*}$.
Above, and in what follows, $x'\in {\ensuremath{\mathbb R}} ^{d_*}$ and $x''\in {\ensuremath{\mathbb R}} ^{d-d_*}$.
Then $\tilde \a _t\in \ensuremath{\mathcal M}({\ensuremath{\mathbb R}} ^{d_*})$, where $\ensuremath{\mathcal M}({\ensuremath{\mathbb R}} ^{d_*})$ is defined as $\ensuremath{\mathcal M}$ but with ${\ensuremath{\mathbb R}} ^{d_*}$ instead of ${\ensuremath{\mathbb R}} ^d$. Due to Item (i) the path $\tilde \a:[0,T]\to \ensuremath{\mathcal M}({\ensuremath{\mathbb R}} ^{d_*})$ is continuous. Due to Item (ii) we have $\tilde \a_0 ( dx') =\tilde v_0(x')d x'$ where $\tilde v_0(x')=\int_{{\ensuremath{\mathbb R}} ^{d-d_*}} v_0 (x',x'') \psi (x'') dx''$.
Moreover, taking $\varphi(x',x''):=\tilde \varphi(x')\psi(x'')$ in Item (iii) with $\tilde \varphi\in C^\infty_c({\ensuremath{\mathbb R}} ^{d_*})$, we get that $\int _{{\ensuremath{\mathbb R}} ^{d_*}} \tilde\varphi(x') \tilde \a_t( dx')= \int _{{\ensuremath{\mathbb R}} ^{d_*}} \tilde \varphi(x') \tilde \a_0(dx') +\int_0^t ds \int_{{\ensuremath{\mathbb R}} ^{d_*}} \nabla \cdot (\tilde D \nabla \tilde \varphi)(x') \tilde \a_s(dx') $.
We set $\tilde B_r:=\{x'\in {\ensuremath{\mathbb R}} ^{d_*}\,:\, |x'|\leq r\}$ and
let $r_\psi$ be the minimal radius such that $\psi$ has support in the ball of ${\ensuremath{\mathbb R}} ^{d-d_*}$ centered at the origin with radius $r_\psi$. Then,
due to Item (iv), it holds
$\tilde \a_s ( \tilde B_r) = \int _{{\ensuremath{\mathbb R}} ^d} \mathds{1}_{\tilde B_r}(x') \psi(x'') \a_s(d x', d x'') \leq
C 2^\g r^\g
$
if $r \geq \tilde r_0:= \max\{r_0,r_\psi\}$. Hence, we have checked that the path $\tilde \a$ satisfies the same conditions appearing in Lemma \ref{timau}, restated for ${\ensuremath{\mathbb R}} ^{d_*}$ with $D$ replaced by $\tilde D$. By the non-degenerate case we conclude that
$\tilde \a_t (d x')=[ \int_{{\ensuremath{\mathbb R}} ^{d_*}} dz' p_t (x', z') \tilde v_0(z')] dx'$. Hence, for all $\tilde \varphi \in C_c({\ensuremath{\mathbb R}} ^d)$ and $\psi \in C_c ^\infty({\ensuremath{\mathbb R}} ^{d-d_*})$ with $\psi \geq 0$, we have
\begin{equation}
\begin{split}
& \int _{{\ensuremath{\mathbb R}} ^{d}} \tilde \varphi(x') \psi (x'') \a_t (dx',dx'') =\int _{{\ensuremath{\mathbb R}} ^{d_*}} dx' \tilde \varphi(x') \int_{{\ensuremath{\mathbb R}} ^{d_*}} dz' p_t (x', z') \tilde v_0(z')\\
& =\int _{{\ensuremath{\mathbb R}} ^{d_*}} dx' \int_{{\ensuremath{\mathbb R}} ^{d-d_*}} d x'' \tilde \varphi(x') \psi(x'') \int_{{\ensuremath{\mathbb R}} ^{d_*}} dz'p_t (x', z') v_0(z',x'')\\
&= \int _{{\ensuremath{\mathbb R}} ^{d_*}} dx' \int_{{\ensuremath{\mathbb R}} ^{d-d_*}} d x'' \tilde \varphi(x') \psi(x'') P_t v_0 (x',x'')\,.
\end{split}
\end{equation}
By additivity and density we then get that $\a_t(dx) = P_t v_0(x) dx $.
\end{proof}
\section{Set $\O_{\rm typ}$ of typical environments}\label{tipicone}
In this section we describe the set $\O_{\rm typ}$ of typical environments $\o$ for which the properties stated in Theorem \ref{teo1} will hold.
We denote by $p^\e_{\o,t}(\cdot , \cdot)$ the transition probability kernel of $( \e X^\o _{\e^{-2} t})_{t\geq 0}$. Recall Definition \ref{tuono20}.
\begin{Definition}[Set $\hat \O$] \label{acqua74} We define $\hat \O$ as the family of $\o\in \O$ such that
\begin{equation}\label{rinato}
\sum _{x\in \hat \o} \sum _{y \in \hat \o:\e y \in B_r} c_x(\o) \int_0^\infty e^{- t} p^\e_{\o,t}(\e x, \e y) dt<+\infty
\end{equation}
for all $\e, r\in (0,+\infty)$, where $B_r:=\{x\in {\ensuremath{\mathbb R}} ^d\,:\, |x|\leq r\}$.
\end{Definition}
\begin{Definition}\label{buddha}(Set $\O_{\rm typ}$)
The set $\O_{\rm typ} $ is given by the environments $\o \in \ensuremath{\mathcal A}[1]\cap \O_\sharp \cap \O_*
\cap \hat \O $ satisfying
\eqref{claudio2} (see respectively Proposition~\ref{prop_ergodico}, Proposition~\ref{replay},
Definition~\ref{def_omega_*} and Definition~\ref{acqua74}).
\end{Definition}
\begin{Remark}\label{cipcip}
Due to Proposition \ref{prop_ergodico}, for any $\o\in \O_{\rm typ}$ we have $\lim_{\e \downarrow 0} \mu^\e_\o(\varphi)= \int_{{\ensuremath{\mathbb R}} ^d} \varphi(x) m dx $ for all $\varphi\in C_c({\ensuremath{\mathbb R}} ^d)$.
\end{Remark}
Given $G\in \{\varphi_j\}$ and $\l>0$, we set $G^{(\l)} := \l G- \nabla\cdot D \nabla G $. Moreover, we denote by $G_{\o, \l}^\e\in \ensuremath{\mathcal D}({\ensuremath{\mathbb L}} ^\e_\o) $ the unique solution of $\l G_{\o,\l}^{\e}-{\ensuremath{\mathbb L}} ^\e_\o G_{\o,\l}^{\e}= G^{(\l)}$ in ${\ensuremath{\mathbb L}} ^2(\mu_\o^\e)$. As $G_{\o, \l}^\e = \int_0^\infty e^{- \l t} P^\e_{\o,t} G ^{(\l)} dt$ (for the notation see Section \ref{figlio_stress}), we have the integral representation
\begin{equation} \label{orbite}G_{\o, \l}^\e (\e x)= \sum _{y \in \hat \o} \int_0^\infty e^{- \l t} p^\e_{\o,t}(\e x, \e y) G^{(\l)}(\e y)dt\,,\qquad \forall x \in \hat\o \,.
\end{equation}
\begin{Remark}\label{origano} If $\o\in \O_{\rm typ}$, then for all $G\in \{\varphi_j\}$, $\e>0$ and $\l=1$ it holds
\begin{equation}\label{eriksen} \sum _{x\in \hat \o} c_x(\o) |G_{\o, \l}^\e(\e x)|<+\infty \text{ and }\sum _{x\in \hat \o} c_x(\o) G_{\o, \l}^\e(\e x)^2<+\infty\,.
\end{equation}
Indeed, by \eqref{orbite} we get that $\| G_{\o, \l}^\e \|_\infty \leq \|G^{(\l)}\|_\infty$. Hence, one has just to check the first bound in \eqref{eriksen}, which follows from \eqref{rinato} and \eqref{orbite} as $G^{(\l)}\in C^\infty_c({\ensuremath{\mathbb R}} ^d)$.
\end{Remark}
\begin{Proposition}\label{udine} $\O_{\rm typ}$ is measurable, translation invariant and $\ensuremath{\mathcal P}(\O_{\rm typ})=1$.
\end{Proposition}
\begin{proof}The sets
$\ensuremath{\mathcal A}[1]$, $ \O_\sharp$, $\O_*$
are translation invariant measurable sets of $\ensuremath{\mathcal P}$--probability one as stated
in Proposition \ref{prop_ergodico}, in Proposition \ref{replay}
and
after Definition \ref{def_omega_*}, respectively.
The same holds for the set of $\o$'s satisfying \eqref{claudio2} as stated after Proposition \ref{replay}.
To conclude it is enough to
show that $\hat \O\cap \O_*$ is a translation invariant measurable set with $\ensuremath{\mathcal P}(\hat \O\cap \O_*)=1$.
For what concerns measurability, it is enough to show that $\hat \O$ is measurable. Trivially, in Definition \ref{acqua74} one can restrict to $r\in {\ensuremath{\mathbb Q}} \cap (0,+\infty)$.
It is also simple to check that one can restrict also to $\e\in {\ensuremath{\mathbb Q}} \cap (0,+\infty)$ by using that, given $0<\e_*<\e$ with $\e_*\in {\ensuremath{\mathbb Q}} $, it holds
\begin{multline}
\int_0^\infty e^{- t} p^\e_{\o,t}(\e x, \e y) dt=
\int_0^\infty e^{- t} p^1_{\o,\e^{-2}t}( x, y) dt
= \e^2 \int_0^\infty e^{- \e^2 s} p^1_{\o,s}( x, y) ds \\
\leq \e^2 \int_0^\infty e^{- \e_* ^2 s} p^1_{\o,s}( x, y) ds
= (\e/\e_* )^2 \int_0^\infty e^{- t} p^{\e_*}_{\o,t}(\e_* x, \e_* y) dt \,.
\end{multline}
Since in in Definition \ref{acqua74} one can restrict to $r,\e \in {\ensuremath{\mathbb Q}} \cap (0,+\infty)$ (hence to a countable set of parameters), we conclude that $\hat\O$ is measurable.
To prove that $\ensuremath{\mathcal P}(\hat \O\cap \O_*)=1$, it is enough to prove that, given $r,\e\in {\ensuremath{\mathbb Q}} \cap (0,+\infty) $, it holds $H(\o)<+\infty$ for $\ensuremath{\mathcal P}$--almost all $\o\in \O_*$ (recall that $\ensuremath{\mathcal P}(\O_*)=1$), where $H$ denotes the l.h.s. of \eqref{rinato}.
We treat the case ${\ensuremath{\mathbb G}} ={\ensuremath{\mathbb R}} ^d$ (the case ${\ensuremath{\mathbb G}} ={\ensuremath{\mathbb Z}} ^d$ can be reduced to the present one by the transformation described in \cite[Section~6]{Fhom3}). We also assume that $V={\ensuremath{\mathbb I}} $ in \eqref{trasferta2}, w.l.o.g. at cost to apply an affine transformation. This implies that $\t_g x= x+g$ and $g(x)=x$
(see \eqref{trasferta2} and \eqref{attimino}).
Given $\o\in \O_*$ and given $x, y \in \hat \o$ (see Definitions \ref{omesso} and \ref{def_omega_*}), we then have $p^\e_{\o,t}(\e x, \e y)=p^\e_{\o,t}(\e y, \e x) = p^\e_{\theta_y\o,t}(0, \e (x-y))$, $c_x(\o)=c_{x-y}(\theta_y\o)$ and $\widehat{\theta_y\o}=\t_{-y}\hat \o=\hat \o-y$.
We set $F(\xi):=\sum_{z\in \hat \xi} c_z(\xi) \int_0^\infty e^{- t}p^\e_{\xi,t} (0, \e z) dt$ for $\xi\in \O_0=\{\xi \in \O\,:\, 0\in \hat\xi \}$.
By the above observations, given $\o\in \O_*$ we get
\begin{equation}
H(\o) = \sum _{y \in \hat \o:\e y \in B_r} \sum _{z\in \widehat{\theta _y \o}} c_z(\theta_y\o) \int_0^\infty e^{- t} p^\e_{\theta_y\o,t}(0, \e z) dt= \sum _{y \in \hat \o:\e y \in B_r}F(\theta_y\o)\,.
\end{equation}
Hence, to prove that $\ensuremath{\mathcal P}(\hat\O\cap \O_*)=1$, we just need to show that the last expression is finite $\ensuremath{\mathcal P}$--a.s. To this aim we apply Campbell's identity (see \cite[Appendix~A]{Fhom3}): for any
nonnegative measurable function $f$ on $ {\ensuremath{\mathbb R}} ^d\times \O_0$ it holds
\begin{equation}\label{campanello}
\int_{{\ensuremath{\mathbb R}} ^d}dx \int _{\O_0} d\ensuremath{\mathcal P}_0 ( \o) f(x, \o) =\frac{1}{m } \int _{\O}d\ensuremath{\mathcal P}(\o) \sum_{x\in \hat \o} f(x , \theta_x \o)
\end{equation}
(we recall that $\ensuremath{\mathcal P}_0$ denotes the Palm distribution associated to $\ensuremath{\mathcal P}$).
Taking $f(x,\o ):= \mathds{1}_{B_r} (\e x) F(\o)$ we get
$\e^{-d} \ell (B_r) {\ensuremath{\mathbb E}} _0[ F] = m^{-1} {\ensuremath{\mathbb E}} [ \sum _{x\in \hat \o: \e x \in B_r} F (\theta _x \o)]$, where
${\ensuremath{\mathbb E}} _0,{\ensuremath{\mathbb E}} $ denote the expectation w.r.t.~$\ensuremath{\mathcal P}_0$, $\ensuremath{\mathcal P}$ respectively and $\ell(B_r)$ denotes the Lebesgue measure of the ball $B_r$.
Hence, to conclude it is enough to show that ${\ensuremath{\mathbb E}} _0[F]<+\infty$.
As can be easily deduced from the proof of \cite[Lemma 3.5]{Fhom3}, the Palm distribution $\ensuremath{\mathcal P}_0$ is a reversible and ergodic (w.r.t. time shifts) distribution for the environment viewed from the random walk, i.e. for the process $(\theta_{X_t^\o }\o) _{t\geq 0}$. Indeed, in the proof of \cite[Lemma 3.5]{Fhom3} we considered the jump chain associated to
the environment viewed from the random walk and proved that ${\ensuremath{\mathbb E}} [c_0]^{-1} c_0(\o) d \ensuremath{\mathcal P}_0(\o) $ is reversible and ergodic for the jump chain.
As accelerating
time does not change the class of reversible and ergodic distributions, we get that $\ensuremath{\mathcal P}_0$ is a reversible and ergodic distribution also
for the process $(\bar{\o}_t)_{t\geq 0}$, $\bar{\o}_t:=\theta_{ X^\o _{\e^{-2} t}} \o $.
On the other hand, by (A7), ${\ensuremath{\mathbb E}} _0 [ c_0]={\ensuremath{\mathbb E}} _0[\l_0]<+\infty$. Hence, by the $L^1$--Birkhoff ergodic theorem, we get
$\lim_{t\to +\infty} \frac{1}{t} \int_0^t c_0( \bar{\o}_s)ds= {\ensuremath{\mathbb E}} _0[ c_0] $ in $L^1({\ensuremath{\mathbb P}} _{\ensuremath{\mathcal P}_0})$, where ${\ensuremath{\mathbb P}} _{\ensuremath{\mathcal P}_0}$ is the law of the random path $(\bar{\o}_t)_{t\geq 0}$ when the starting configuration $\o$ is sampled with distribution $\ensuremath{\mathcal P}_0$. As the above limit implies the limit of expectations and $c_0(\theta_z\o)=c_z(\o)$, we have
\begin{equation}\label{piero88}
\lim_{t\to +\infty} \frac{1}{t} \sum_{z\in \hat \o}\int \ensuremath{\mathcal P}_0(d\o) \int_0^t p^\e_{\o,s} (0, \e z) c_z(\o) ds = {\ensuremath{\mathbb E}} _0 [ c_0]\,.
\end{equation}
Now note that, for some positive constant $C$, it holds
\begin{equation}
\begin{split}
F(\o)
\leq C \sum_{n=0}^\infty \frac{e^{- n/2 }}{n+1} \sum_{z\in \hat \o} \int _{n}^{n+1} p^\e_{\o,t} (0, \e z) c_z(\o) dt\,.
\end{split}
\end{equation}
Hence, setting $ a_n:= \frac{1}{n+1} \int \ensuremath{\mathcal P}_0(d \o) \sum_{z\in \hat \o} \int _{0}^{n+1} p^\e_{\o,t} (0, \e z) c_z(\o) dt$, we get ${\ensuremath{\mathbb E}} _0[ F] \leq C \sum_{n=0}^\infty e^{- n/2 } a_n$.
By \eqref{piero88} we have $\lim_{n\to \infty} a_n = {\ensuremath{\mathbb E}} _0 [ c_0]<+\infty$, hence the series $ \sum_{n=0}^\infty e^{- n/2 } a_n$ is finite, thus implying that ${\ensuremath{\mathbb E}} _0[ F] <+\infty$. This concludes the proof that $\ensuremath{\mathcal P}(\hat \O\cap \O_*)=1$
We now show
that $\hat \O\cap \O_*$ is
translation invariant (always restricting to ${\ensuremath{\mathbb G}} ={\ensuremath{\mathbb R}} ^d$ and $V={\ensuremath{\mathbb I}} $). Take $ \e, r>0$, $g\in {\ensuremath{\mathbb G}} $ and $\o \in \hat \O\cap \O_* $.
Then we have
\begin{equation*}
\begin{split}
H(\theta_g \o) &= \sum_{x\in \widehat{\theta_g \o}} \sum _{y \in \widehat{\theta_g \o}: \e y \in B_r } c_x(\theta_g \o)
\int_0^\infty e^{- t}
p^\e_{\theta_g\o,t}(\e x, \e y)dt
\\
& \leq \sum_{a\in \hat \o}\sum _{b \in \hat \o: \e b \in B_{r+ \e|g|}} c_a(\o) \int_0^\infty e^{- t} p^\e_{\o,t}(\e a, \e b)dt <+\infty\,.
\end{split}
\end{equation*}
This proves that $\theta_g (\hat \O\cap \O_*)\subset \hat \O$ for all $g\in {\ensuremath{\mathbb G}} $.
Using that $\O_*$ is translation invariant, it is then trivial to conclude that
$\theta_g (\hat \O\cap \O_*)\subset \hat \O \cap \O_*$ for all $g \in {\ensuremath{\mathbb G}} $, which implies the translation invariance of $\hat \O\cap \O_*$.
\end{proof}
\section{Proof of Theorem \ref{teo1}}\label{ida}
In Proposition \ref{prop_SEP} we discussed the existence
of the simple exclusion process for $\o\in \O_*\supset \O_{\rm typ}$. In Proposition \ref{udine} we
showed that $\O_{\rm typ}$ is a translation invariant measurable set of $\ensuremath{\mathcal P}$-probability one.
To get the hydrodynamic behavior we will proceed as follows. We fix $\o\in \O_{\rm typ}$. We consider the random path $(\pi^\e_{\o,t} [ \eta_\cdot])_{0\leq t\leq T}$ with $\eta_\cdot$ sampled according to ${\ensuremath{\mathbb P}} ^\e_{ \o,\mathfrak{m}_\e }$.
We call $Q^\e$ its law, which is a probability measure on $D([0,T], \ensuremath{\mathcal M})$. Note that, to simplify the notation, $\o$ is understood in $Q^\e$.
We call $Q^\infty $ the law of the deterministic path $\bigl( \rho(x,t) dx \bigr)_{0\leq t\leq T}$ in
$D([0,T], \ensuremath{\mathcal M})$ where $\rho(x,t)= P_t \rho_0(x)$. To get Theorem \ref{teo1} it is enough to prove that, for $\o\in \O_{\rm typ}$, $Q^\e$ weakly converges to $Q^\infty$. Indeed, this implies the convergence in probability of the random path $(\pi^\e_{\o,t} [ \eta_\cdot])_{0\leq t\leq T}$ towards $\bigl( \rho(x,t) dx \bigr)_{0\leq t\leq T}$.
As $\bigl( \rho(x,t) dx \bigr)_{0\leq t\leq T}\in C([0,T], \ensuremath{\mathcal M}) $,
the above convergence in probability implies \eqref{pasqualino} (cf. \cite[page 124]{B}).
By adapting the method of the corrected empirical process of \cite{GJ} to the $L^2$-context and the unbounded domain ${\ensuremath{\mathbb R}} ^d$, we prove the tightness of $\{Q^\e\}$ in Section \ref{sec_teso} (since $D([0,T], \ensuremath{\mathcal M})$ is a Polish space, tightness is here equivalent to relative compactness, cf. \cite[Theorems~5.1~and~5.2]{B}).
The homogenization result used in this part is given by \eqref{flavia}.
After getting tightness, one can proceed in two ways.
A first route is to show that all limit points of $\{Q^\e\}$ must equal $Q^\infty$ since concentrated on continuous $\ensuremath{\mathcal M}$--valued paths solving in a weak sense the hydrodynamic equation with initial value $\rho_0(x) dx$ and satisfying suitable mass bounds on balls. Then one can invoke the uniqueness result for these weak solutions given by Lemma \ref{timau}.
This is the route followed in Section \ref{silenzioso} in the same spirit of \cite{GJ}. Again, the homogenization result used here is given by \eqref{flavia}.
We now describe the second route. Due to tightness and by \cite[Theorem 13.1]{B}, to prove that $Q^\e \Rightarrow Q^\infty$ it is enough to show the finite dimensional distribution convergence and, by a union bound, the convergence for the distribution at a fixed time: i.e. that
for any $t\geq 0$, $\d>0$ and $\varphi \in C_c({\ensuremath{\mathbb R}} ^d)$ it holds
\begin{equation}\label{pasqualino_fisso}
\lim_{\e\downarrow 0} {\ensuremath{\mathbb P}} ^\e _{ \o,\mathfrak{m}_\e } \Big(\Big| \e^d \sum_{x \in \hat \o} \varphi (\e x) \eta_t( x) - \int _{{\ensuremath{\mathbb R}} ^d} \varphi(x) \rho(x,t) dx\Big| >\d
\Big)=0\,.
\end{equation}
This can be obtained by a completely autonomous analysis with two main ingredients: an extension to our context of Nagy's representation of the simple exclusion process (based on duality with the random walk) and the homogenization limit \eqref{marvel2}. Note that here one does not need Lemma \ref{timau}.
We have discussed this second route in Appendix \ref{sec_passetto}. %
\subsection{Relative compactness of the empirical measure} \label{sec_teso}
To simplify the notation, we fix once and for all a sequence $\{\e_n\}$ of positive numbers with $\e_n \downarrow 0$. In what follows all limits $\e\downarrow 0$ have to be thought along the above sequence $\{\e_n\}$.
By Lemma \ref{proiettore}, to prove that the family $\{Q^\e\}$ is relatively compact as $\e\downarrow 0$
it is enough to prove that, given $G\in \{\varphi_j\}$ (cf. Definition~\ref{tuono20}), the $\e$--parameterized laws of the random paths $(\pi^\e_{\o,t}(G))_{0\leq t\leq T}$ form a relatively compact family of probability measures on $D([0,T], {\ensuremath{\mathbb R}} )$ as $\e\downarrow 0$. Note that we have dropped from the notation the dependence on the path $\eta_\cdot $.
By \cite[Theorem~1.3]{KL} and Aldous' criterion given in \cite[Proposition~1.6]{KL}, it is enough to prove that
\begin{itemize}
\item[(i)] for every $t\in [0,T]$ and every $\b>0$, there exists $\ell>0$ such that $\varlimsup_{\e \downarrow 0} {\ensuremath{\mathbb P}} _{\o,\mathfrak{m}_\e } ^\e (| \pi ^\e_{\o,t}(G) |> \ell) \leq \b$;
\item[(ii)] calling $\mathfrak{I}_T$ the family of stopping times bounded by $T$ w.r.t. to the filtration $(\ensuremath{\mathcal F}_t)_{t\geq 0}$, with $\ensuremath{\mathcal F}_t:=\s \{ \eta_s: 0 \leq s \leq t\}$, for any $\b>0$ it holds
\begin{equation}
\lim_{\g \downarrow 0} \varlimsup_{\e \downarrow 0}\sup_{
\substack{\t \in \mathfrak{I}_T\\ \theta\leq \g}}{\ensuremath{\mathbb P}} ^\e _{\o,\mathfrak{m}_\e } \left( \left| \pi ^\e_{\o,\t}(G)-\pi ^\e_{\o,(\t+\theta)\wedge T }(G)\right | >\b\right)=0\,.
\end{equation}
\end{itemize}
Item (i) gives no problem. Indeed, $ |\pi ^\e_{\o,t}(G) |\leq \mu_\o^\e (|G|) \to \int dx m |G(x)| dx$ as
$\o \in \O_{\rm typ} $ (cf.~Remark \ref{cipcip}). Item (ii) is more delicate and can be treated by the corrected empirical measure. To this aim we fix $\l>0$
(let us take $\l=1$ as in Remark \ref{origano}) and define $G^{(\l)}\in C^\infty_c({\ensuremath{\mathbb R}} ^d) $ as
\begin{equation}\label{glambda} G^{(\l)} := \l G- \nabla\cdot D \nabla G \,,\end{equation}
where $D$ is the effective homogenized matrix (see Definition \ref{def_D}).
As in Section \ref{tipicone},
we define $G_{\o,\l}^{\e}$ as the unique element of $\ensuremath{\mathcal D}({\ensuremath{\mathbb L}} ^\e_\o)\subset L^2(\mu ^\e_\o)$ such that
\begin{equation}\label{carnevale}
\l G_{\o,\l}^{\e}-{\ensuremath{\mathbb L}} ^\e_\o G_{\o,\l}^{\e}= G^{(\l)} \qquad \text{ in }L^2(\mu ^\e_\o)\,.
\end{equation}
By using the resolvent operators $R_\l$ and $R^\e_{\o,\l}$ defined in Section \ref{figlio_stress}, we can restate the above definitions as
\begin{equation}\label{fuochino}
G=R_\l G^{(\l)}\,, \qquad G_{\o,\l}^{\e}= R_{\o,\l}^{\e}G^{(\l)}\,.
\end{equation}
We point out some standard bounds which we will be useful below. By
taking the scalar product with $G_{\o,\l}^{\e}$ in the massive Poisson equation
\eqref{carnevale} and by applying Schwarz inequality, we get that
\begin{align}
& \l \| G_{\o,\l}^{\e}\|_{L^2(\mu^\e_\o)}
\leq \| G^{(\l)} \|_{L^2(\mu^\e_\o)}\,, \label{aperitivo1}\\
& \langle G_{\o,\l}^{\e}, -{\ensuremath{\mathbb L}} ^\e_\o G_{\o,\l}^{\e} \rangle _{L^2(\mu^\e_\o)} \leq \langle G_{\o,\l}^{\e},
G^{(\l)} \rangle _{L^2(\mu^\e_\o)}\leq \l^{-1}\|G^{(\l)} \| ^2 _{L^2(\mu^\e_\o)}
\,.\label{aperitivo2}
\end{align}
We also note that
\begin{equation}\label{aperitivo3}
\langle G_{\o,\l}^{\e}, -{\ensuremath{\mathbb L}} ^\e_\o G_{\o,\l}^{\e} \rangle _{L^2(\mu^\e_\o)} =
\frac{\e^{d-2} }{2} \sum _{x\in \hat \o} \sum _{y \in \hat \o} c_{x,y}(\o)\bigl[ G_{\o,\l}^{\e} (\e x)- G_{\o,\l}^{\e} (\e y)\bigr]^2\,.
\end{equation}
To justify \eqref{aperitivo3} we proceed as follows. For any $f\in \ensuremath{\mathcal D}( {\ensuremath{\mathbb L}} ^\e_\o)\subset \ensuremath{\mathcal D}( \sqrt{- {\ensuremath{\mathbb L}} ^\e_\o})$ we have $ \langle f, - {\ensuremath{\mathbb L}} ^\e_\o f \rangle _{L^2(\mu^\e_\o)}= \|\sqrt{-{\ensuremath{\mathbb L}} ^\e_\o} f\|^2_{L^2(\mu^\e_\o)}=
\ensuremath{\mathcal E}_\o^\e (f,f) $, the last identity being discussed in Section \ref{figlio_stress}. By taking $f=G_{\o,\l}^{\e}$, we then get \eqref{aperitivo3}.
We now use our homogenization result for the resolvent convergence. Indeed, $\O_{\rm typ}\subset \O_\sharp$ and \eqref{fuochino} and \eqref{flavia} imply that \begin{equation}\label{ratto_gatto}
\lim _{\e \downarrow 0} \e^d \sum_{x\in \hat \o} | G(\e x) - G_{\o,\l}^{\e} (\e x)| =0 \,.
\end{equation}
As $ | \pi^{\e} _{\o,t}(G) -\pi^{\e} _{\o,t}(G_{\o,\l}^{\e}) |\leq \e^d \sum_{x\in \hat \o} | G(\e x) - G_{\o,\l}^{\e} (\e x)|$, we get
\begin{equation} \label{violoncello}
\lim _{\e \downarrow 0} {\ensuremath{\mathbb P}} ^\e_{\o, \mathfrak{m}_\e} \Big( \sup_{0\leq t\leq T} \bigl| \pi^{\e} _{\o,t}(G) -\pi^{\e} _{\o,t}(G_{\o,\l}^{\e}) \bigr| >\d
\Big) =0\,.
\end{equation}
By \eqref{violoncello}, to get Item (ii), it is enough to prove the same result with $G$ replaced by $G_{\o,\l}^{\e}$, i.e. that for any $\b>0$ it holds
\begin{equation}\label{il_sole!}
\lim_{\g \downarrow 0} \varlimsup_{\e \downarrow 0}\sup_{
\substack{\t \in \mathfrak{I}_T\\ \theta\leq \g}}{\ensuremath{\mathbb P}} ^\e_{\o,\mathfrak{m}_\e } \left( \left| \pi ^\e_{\o,\t}(G_{\o,\l}^{\e})-\pi ^\e_{\o,(\t+\theta)\wedge T }(G_{\o,\l}^{\e})\right | >\b\right)=0\,.
\end{equation}
We have now to deal with the Dynkin martingale associated to $\pi^{\e} _{\o}(G_{\o,\l}^{\e})$. We will use below Lemmas \ref{ringo} and \ref{star}. Let us check their hypotheses. Trivially, $G_{\o,\l}^{\e}\in \ensuremath{\mathcal D}({\ensuremath{\mathbb L}} ^\e_\o)$ (by definition).
We claim that
\[G_{\o, \l}^\e \in L^1 (\mu^\e_\o)\,,\;\; \sum _{x\in \hat \o} c_x(\o) |G_{\o, \l}^\e(\e x)|<+\infty\,,\;\; \sum _{x\in \hat \o} c_x(\o) G_{\o, \l}^\e(\e x)^2 <+\infty\,.
\]
The last two bounds follow from Remark \ref{origano}. To derive the first one we observe that, by the integral representation \eqref{orbite} and the symmetry of $p^\e_{\o,t}(\cdot, \cdot)$, it holds $\| G_{\o, \l}^\e \|_{L^1(\mu^\e_\o)} \leq \|G^{(\l)}\|_{L^1(\mu^\e_\o)} /\l<+\infty$.
By applying Lemma \ref{ringo}, we get that $\pi^{\e} _{\o}(G_{\o,\l}^{\e}) $ corresponds to an absolutely convergent series in $C(\{0,1\}^{\hat \o})$ and, as function of $\eta$, belongs to the domain of $\ensuremath{\mathcal L}_\o$.
This observation allows us to introduce the Dynkin martingale
\begin{equation} \label{sirenetta}
M^\e_{\o,t} : = \pi^{\e} _{\o,t}(G_{\o,\l}^{\e}) - \pi^{\e} _{\o,0}(G_{\o,\l}^{\e})
-\e^{-2} \int _0 ^t \ensuremath{\mathcal L}_\o \left( \pi^{\e} _{\o}(G_{\o,\l}^{\e} ) \right)(\eta_s) ds\,.
\end{equation}
By \eqref{airone26} and \eqref{carnevale}, we can rewrite $M^\e_{\o,t}$ as
\begin{equation} \label{tremo_bis}
M^\e_{\o,t}= \pi^{\e} _{\o,t}(G_{\o,\l}^{\e}) - \pi^{\e} _{\o,0}(G_{\o,\l}^{\e}) -
\e^{d} \sum _{x\in \hat \o}\int_0^t \eta_s(x) \left( \l G_{\o,\l}^{\e} - G^{(\l)}\right ) (\e x)ds\,.
\end{equation}
We can now prove \eqref{il_sole!}.
Due to \eqref{tremo_bis} it is enough to prove that
\begin{align}
& \lim_{\g \downarrow 0} \varlimsup_{\e \downarrow 0}\sup_{
\substack{\t \in \mathfrak{I}_T\\ \theta\leq \g}}{\ensuremath{\mathbb P}} ^\e_{\o,\mathfrak{m}_\e } \Big(
\e^{d} \theta \sum _{x\in \hat \o} \left| \l G_{\o,\l}^{\e} - G^{(\l)}\right | (\e x)
>\b/2\Big)=0\,,\label{il_sole1}\\
& \lim_{\g \downarrow 0} \varlimsup_{\e \downarrow 0}\sup_{
\substack{\t \in \mathfrak{I}_T\\ \theta\leq \g}}{\ensuremath{\mathbb P}} ^\e_{\o,\mathfrak{m}_\e } \Big( \Big |M^\e_{\o,(\tau+\theta)\wedge T}- M^\e_{\o,\t}
\Big | >\b/2\Big)=0\,.
\label{il_sole2}
\end{align}
$\bullet$ \emph{Proof of \eqref{il_sole1}}.
The inequality inside \eqref{il_sole1} is indeed deterministic.
As $\theta\leq \g \downarrow 0$, to prove \eqref{il_sole1} it is enough to prove that
\begin{equation}\label{freddo}
\varlimsup_{\e \downarrow 0} \e^{d} \sum _{x\in \hat \o} \left| \l G_{\o,\l}^{\e} - G^{(\l)}\right | (\e x)<+\infty\,.
\end{equation}
We have already observed that $\| G_{\o, \l}^\e \|_{L^1(\mu^\e_\o)} \leq \|G^{(\l)}\|_{L^1(\mu^\e_\o)} /\l$. Then, to get \eqref{freddo} it is enough to apply Remark \ref{cipcip}.
\medskip
$\bullet$ \emph{Proof of \eqref{il_sole2}}. We write ${\ensuremath{\mathbb E}} ^\e_{\o,\mathfrak{m}_\e }$ for the expectation w.r.t. ${\ensuremath{\mathbb P}} ^\e_{\o,\mathfrak{m}_\e}$.
We bound the probability in \eqref{il_sole2} by $(2/\b)^{2} {\ensuremath{\mathbb E}} ^\e_{\o,\mathfrak{m}_\e } \left[ (M^\e_{\o,(\t+\theta)\wedge T} - M^\e_{\o,\t})^2
\right]$. Using that $\t$ is a stopping time and
the form of the sharp bracket process in Lemma \ref{star}, we get (cf. \eqref{aperitivo2} and \eqref{aperitivo3})
\begin{multline}\label{minecraft}
{\ensuremath{\mathbb E}} ^\e_{\o,\mathfrak{m}_\e } \left[ (M^\e_{\o,(\t+\theta)\wedge T} - M^\e_{\o,\t})^2
\right] \leq \theta \e^{2d-2} \sum _{x\in \hat \o} \sum _{y \in \hat \o} c_{x,y}(\o)\bigl[ G_{\o,\l}^{\e} (\e x)- G_{\o,\l}^{\e} (\e y)\bigr]^2\\
= 2 \theta \e^d \langle G_{\o,\l}^{\e} , -{\ensuremath{\mathbb L}} ^\e_\o G_{\o,\l}^{\e} \rangle _{L^2(\mu^\e_\o)} \leq 2\theta \e^d \l^{-1} \| G^{(\l)}\|^2_{L^2(\mu^\e_\o)} \,.
\end{multline}
As $\o \in \O_{\rm typ}$ (see Remark \ref{cipcip}), as $\e\downarrow 0$ we have $\| G^{(\l)}\|^2_{L^2(\mu^\e_\o)}\to C_0:= \int dx \,m G^{(\l)}(x) ^2 $.
In conclusion we have proved that the probability in \eqref{il_sole2} is bounded from above by
$ (2/\b)^{2} 2 \theta \e^d \l^{-1} ( C_0+o(1)) $ as $\e\downarrow 0$. This implies \eqref{il_sole2}.
\subsection{Characterization of the limit points}\label{silenzioso} Recall that $\o\in \O_{\rm typ}$ is fixed.
Let $Q$ be any limit point $\{Q^\e\}$ as $\e \downarrow 0$. We claim that $Q$ is concentrated on paths $\a\in D([0,T],\ensuremath{\mathcal M})$ satisfying the conditions of Lemma \ref{timau} with $v_0=\rho_0$. Then, by applying Lemma \ref{timau},
we can conclude that $Q=\d _{(\rho (x,t) dx)_{0 \leq t \leq T}}$, thus completing the proof of Theorem \ref{teo1}.
Let us prove our claim. Item (ii) in Lemma \ref{timau} follows from condition \eqref{marzolino}. We move to Item (iv).
We recall that, for any integer $\ell\geq 0$, there is some $[0,1]$--valued function $\varphi_{j_0}\in \{\varphi_j\}$ equal to $1$ on $B_\ell$ and equal to zero outside $B_{\ell+1}$.
Then, by Remark \ref{cipcip}, we have for all $t\in [0,T]$ and for a suitable constant $C(d)$ depending only on the dimension $d$ that
\begin{equation}\label{pico85}
\pi_{\o,t}^\e (\varphi_{j_0}) \leq \mu^\e_\o(\varphi_{j_0}) \stackrel{\e \downarrow 0}{\to} m\int \varphi_{j_0}(x) dx \leq C(d) m \ell^d \,.
\end{equation}
Setting
$H:=\{ \a \,:\, \sup_{0\leq t\leq T} \a_t (\varphi_{j_0}) \leq 2 C(d) m \ell^d \}$, we get $\lim_{\e \downarrow 0} {\ensuremath{\mathbb P}} ^\e _{ \o,\mathfrak{m}_\e }( \pi_{\o,\cdot}^\e \in H )=1$. As $H$ is closed in
$D([0,T],\ensuremath{\mathcal M}) $, we conclude that $Q(H)=1$ by varying $\ell $ in ${\ensuremath{\mathbb N}} $. This implies Item (iv) in Lemma \ref{timau} with $\g=d$.
We move to Item (iii).
By Doob's inequality and reasoning as in \eqref{minecraft} we get
\begin{equation} \label{liuto}
{\ensuremath{\mathbb P}} ^\e_{\o, \mathfrak{m}_\e}( \sup _{t\in [0,T]} |M^\e_{\o,t} | \geq \d) \leq \d^{-2} {\ensuremath{\mathbb E}} ^\e_{\o, \mathfrak{m}_\e}( (M^\e_{\o,T} )^2) \leq 2 \d^{-2} T \e^d \l^{-1} \| G^{(\l)}\|^2_{L^2(\mu^\e_\o)} \stackrel{\e\da0}{\to} 0 \,.
\end{equation}
By
\eqref{glambda} and \eqref{ratto_gatto} (the latter is due to \eqref{flavia} in Proposition \ref{replay}), we have
\begin{equation}\label{viola}
\sup_{0\leq t\leq T} \Big|\int_0^t \pi^\e_{\o,s} \left( \l G_{\o,\l}^{\e} - G^{(\l)}- \nabla \cdot D \nabla G \right) ds
\Big|
\leq
T \e^{d} \sum _{x\in \hat \o} | \l G_{\o,\l}^{\e} - \l G | (\e x) \stackrel{\e \downarrow 0}{\to} 0\,.
\end{equation}
At this point, by combining \eqref{violoncello}, \eqref{tremo_bis}, \eqref{liuto} and \eqref{viola} we get that
\begin{equation}\label{razzo1}
{\ensuremath{\mathbb P}} ^\e _{ \o,\mathfrak{m}_\e } \Big(
\sup_{0\leq t\leq T} |
\pi^{\e} _{\o,t}(G) - \pi^{\e} _{\o,0}(G) -\int_0^t \pi^{\e} _{\o,s}(
\nabla \cdot D \nabla G ) ds |\leq \d)=1\,.
\end{equation}
As a consequence, given $G\in\{\varphi_j\}$, $Q$--a.s. it holds $ \a_t(G)-\a_0(G)-\int_0^t \a_s ( \nabla \cdot D \nabla G ) ds =0$ for all $0\leq t\leq T$ (adapt the proof of \cite[Lemma 8.7]{timo} to show that $\{ \a:
\sup_{0\leq t\leq T} |
\a_t(G) - \a_0(G) -\int_0^t \a_s (
\nabla \cdot D \nabla G ) ds |\leq \d \}$ is closed in $ D([0,T],\ensuremath{\mathcal M})$). By the construction of $\{\varphi_j\}$ in Section \ref{sec_mammina},
given a generic $\varphi \in C_c^\infty ({\ensuremath{\mathbb R}} ^d)$ with support in some $B_\ell$, we know that for each $\d>0$
there exists $G\in\{\varphi_j\}$ with support in $B_{\ell+1}$ such that $\| G - \varphi\|_\infty \leq \d$ and $\sup_{1\leq i,k\leq d}\| \partial^2_{x_i, x_k} G -\partial^2_{x_i,x_k} \varphi\|_\infty \leq \d$. Hence both $ \sup_{0 \leq t \leq T}\bigl| \a_t (G) -\a_t(\varphi) \bigr| $ and $ \sup_{0 \leq t \leq T}\bigl| \a_t ( \nabla \cdot D \nabla G ) -\a_t( \nabla \cdot D \nabla \varphi) \bigr| $ can be bounded by $ C\d \sup_{0\leq t \leq T} \a_t(B_{\ell+1}) $, where $C=C(D)$. Due to Item (iv) (already proved) and by density,
we conclude that $Q$--a.s. it holds $ \a_t(\varphi)-\a_0(\varphi)-\int_0^t \a_s ( \nabla \cdot D \nabla \varphi ) ds =0$ for all $0\leq t\leq T$ and all $\varphi \in C^\infty_c({\ensuremath{\mathbb R}} ^d)$. Hence Item (iii) in Lemma \ref{timau} is verified.
We now check Item (i) in Lemma \ref{timau}.
By Remark \ref{cipcip}, given $G\in \{\varphi_j\}$
we get
\begin{equation}\label{razzo2}
\sup _{\substack{0\leq s\leq t\leq T\\ |t-s| < \b}} \Big| \int_s^t \pi^{\e} _{\o,u}(
\nabla \cdot D \nabla G ) du \Big|\leq \b \e^{d} \sum _{x\in \hat \o} |\nabla \cdot D \nabla G (\e x) | \stackrel{\e \downarrow 0}{\to} C(G)\b\,.
\end{equation}
We set $\tilde H:=\{ \a\,:\, |\a_t(G)-\a_s(G)|\leq 2 C(G) \b \text{ for all } 0\leq s\leq t\leq T \text{ with }|t-s|< \b\}$.
By combining \eqref{razzo1} and \eqref{razzo2} we get $\lim_{\e \downarrow 0}{\ensuremath{\mathbb P}} ^\e _{ \o,\mathfrak{m}_\e }( \pi_{\o,\cdot}^\e \in \tilde H)=1$. As $\tilde H$ is closed in $D([0,T],\ensuremath{\mathcal M})$, we conclude that $Q(\tilde H)=1$. By varying $G$ among $\{\varphi_j\}$ and by taking $\b \downarrow 0$ along a sequence, we get that
$Q(C([0,T],\ensuremath{\mathcal M}) )=1$.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.