二次问题上的优化性能敏感地依赖于谱的低频部分。对于大型（实际上是无限维的）问题，这部分谱通常可以自然地由 power law 分布表示或近似，从而使得基于梯度的算法对这些问题的迭代求解产生 power law 收敛速率。在本文中，我们提出了一种新的谱条件，为具有 power law 优化轨迹的问题提供了更紧的上界。我们利用该条件为一系列广泛的优化算法——Gradient Descent、Steepest Descent、Heavy Ball 和 Conjugate Gradients——构建了上下界的完整图景，并重点关注学习率和动量的底层调度。特别地，我们展示了如何针对给定的谱形状，以统一的方式获得最优加速方法、其调度以及收敛上界。此外，我们首次为 Steepest Descent 和 Conjugate Gradients 在具有一般指数的 spectral power laws 下的收敛速率提供了紧下界证明。我们的实验表明，所获得的收敛界和加速策略不仅与精确的二次优化问题相关，而且在应用于神经网络训练时也相当准确。
英文摘要：
Performance of optimization on quadratic problems sensitively depends on the low-lying part of the spectrum. For large (effectively infinite-dimensional) problems, this part of the spectrum can often be naturally represented or approximated by power law distributions, resulting in power law convergence rates for iterative solutions of these problems by gradient-based algorithms. In this paper, we propose a new spectral condition providing tighter upper bounds for problems with power law optimization trajectories. We use this condition to build a complete picture of upper and lower bounds for a wide range of optimization algorithms -- Gradient Descent, Steepest Descent, Heavy Ball, and Conjugate Gradients -- with an emphasis on the underlying schedules of learning rate and momentum. In particular, we demonstrate how an optimally accelerated method, its schedule, and convergence upper bound can be obtained in a unified manner for a given shape of the spectrum. Also, we provide first proofs of tight lower bounds for convergence rates of Steepest Descent and Conjugate Gradients under spectral power laws with general exponents. Our experiments show that the obtained convergence bounds and acceleration strategies are not only relevant for exactly quadratic optimization problems, but also fairly accurate when applied to the training of neural networks.
In distributed learning, local SGD (also known as federated averaging) and its simple baseline minibatch SGD are widely studied optimization methods. Most existing analyses of these methods assume independent and unbiased gradient estimates obtained via with-replacement sampling. In contrast, we study shuffling-based variants: minibatch and local Random Reshuffling, which draw stochastic gradients without replacement and are thus closer to practice. For smooth functions satisfying the Polyak-Łojasiewicz condition, we obtain convergence bounds (in the large epoch regime) which show that these s…
For solving finite-sum optimization problems, SGD without replacement sampling is empirically shown to outperform SGD. Denoting by $n$ the number of components in the cost and $K$ the number of epochs of the algorithm , several recent works have shown convergence rates of without-replacement SGD that have better dependency on $n$ and $K$ than the baseline rate of $O(1/(nK))$ for SGD. However, there are two main limitations shared among those works: the rates have extra poly-logarithmic factors on $nK$, and denoting by $κ$ the condition number of the problem, the rates hold after $κ^c\log(nK)$ …
University of Connecticut（康涅狄格大学）；IBM Research（IBM研究院）
We study the convergence of Stochastic Gradient Descent (SGD) for strongly convex objective functions. We prove for all $t$ a lower bound on the expected convergence rate after the $t$-th SGD iteration; the lower bound is over all possible sequences of diminishing step sizes. It implies that recently proposed sequences of step sizes at ICML 2018 and ICML 2019 are {\em universally} close to optimal in that the expected convergence rate after {\em each} iteration is within a factor $32$ of our lower bound. This factor is independent of dimension $d$. We offer a framework for comparing with lower…
Weizmann Institute of Science（魏茨曼科学研究所）；TTI Chicago（芝加哥丰田工业大学）
We provide tight finite-time convergence bounds for gradient descent and stochastic gradient descent on quadratic functions, when the gradients are delayed and reflect iterates from $τ$ rounds ago. First, we show that without stochastic noise, delays strongly affect the attainable optimization error: In fact, the error can be as bad as non-delayed gradient descent ran on only $1/τ$ of the gradients. In sharp contrast, we quantify how stochastic noise makes the effect of delays negligible, improving on previous work which only showed this phenomenon asymptotically or for much smaller delays. Al…
MIT CSAIL（麻省理工学院计算机科学与人工智能实验室）；MIT EECS（麻省理工学院电子工程与计算机科学系）
We study the question of obtaining last-iterate convergence rates for no-regret learning algorithms in multi-player games. We show that the optimistic gradient (OG) algorithm with a constant step-size, which is no-regret, achieves a last-iterate rate of $O(1/\sqrt{T})$ with respect to the gap function in smooth monotone games. This result addresses a question of Mertikopoulos & Zhou (2018), who asked whether extra-gradient approaches (such as OG) can be applied to achieve improved guarantees in the multi-agent learning setting. The proof of our upper bound uses a new technique centered around …
MIT（麻省理工学院）；CUHK（香港中文大学）；MIT-IBM AI Lab（麻省理工学院-IBM人工智能实验室）；IBM Research（IBM研究院）
Gradient Descent Ascent (GDA) methods are the mainstream algorithms for minimax optimization in generative adversarial networks (GANs). Convergence properties of GDA have drawn significant interest in the recent literature. Specifically, for $\min_{\mathbf{x}} \max_{\mathbf{y}} f(\mathbf{x};\mathbf{y})$ where $f$ is strongly-concave in $\mathbf{y}$ and possibly nonconvex in $\mathbf{x}$, (Lin et al., 2020) proved the convergence of GDA with a stepsize ratio $η_{\mathbf{y}}/η_{\mathbf{x}}=Θ(κ^2)$ where $η_{\mathbf{x}}$ and $η_{\mathbf{y}}$ are the stepsizes for $\mathbf{x}$ and $\mathbf{y}$ and…
INRIA（法国国家信息与自动化研究所）；Ecole Normale Supérieure（高等师范学校）；PSL Research University（巴黎文理研究大学）
In the context of statistical supervised learning, the noiseless linear model assumes that there exists a deterministic linear relation $Y = \langle θ_*, X \rangle$ between the random output $Y$ and the random feature vector $Φ(U)$, a potentially non-linear transformation of the inputs $U$. We analyze the convergence of single-pass, fixed step-size stochastic gradient descent on the least-square risk under this model. The convergence of the iterates to the optimum $θ_*$ and the decay of the generalization error follow polynomial convergence rates with exponents that both depend on the regulari…
École Normale Supérieure, Paris（巴黎高等师范学院）；Mila & DIRO, Université de Montréal（蒙特利尔大学）
We consider differentiable games where the goal is to find a Nash equilibrium. The machine learning community has recently started using variants of the gradient method (GD). Prime examples are extragradient (EG), the optimistic gradient method (OG) and consensus optimization (CO), which enjoy linear convergence in cases like bilinear games, where the standard GD fails. The full benefits of theses relatively new methods are not known as there is no unified analysis for both strongly monotone and bilinear games. We provide new analyses of the EG's local and global convergence properties and use…
Byzantine-Robust Distributed SGD: A Unified Analysis and Tight Error Bounds
School of Advanced Interdisciplinary Sciences, University of Chinese Academy of Sciences（中国科学院大学前沿交叉科学学院）；School of Mathematical Sciences, University of Chinese Academy of Sciences（中国科学院大学数学科学学院）；Ministry of Education Key Laboratory of Mathematics and Information Networks, School of Mathematical Sciences, Beijing University of Posts and Telecommunications（北京邮电大学数学科学学院数学与信息网络教育部重点实验室）
The monotone variational inequality is a central problem in mathematical programming that unifies and generalizes many important settings such as smooth convex optimization, two-player zero-sum games, convex-concave saddle point problems, etc. The extragradient algorithm by Korpelevich [1976] and the optimistic gradient descent-ascent algorithm by Popov [1980] are arguably the two most classical and popular methods for solving monotone variational inequalities. Despite their long histories, the following major problem remains open. What is the last-iterate convergence rate of the extragradient…
Tight Convergence Rates for Online Distributed Linear Estimation with Adversarial Measurements
Dept. of Computer Science and Automation, Indian Institute of Science, Bengaluru（印度科学理工学院计算机科学与自动化系）；Department of Computer Science, IMT Atlantique, Plouzané（IMT Atlantique计算机科学系）
Recently, several authors have shown local and global convergence rate results for Douglas-Rachford splitting under strong monotonicity, Lipschitz continuity, and cocoercivity assumptions. Most of these focus on the convex optimization setting. In the more general monotone inclusion setting, Lions and Mercier showed a linear convergence rate bound under the assumption that one of the two operators is strongly monotone and Lipschitz continuous. We show that this bound is not tight, meaning that no problem from the considered class converges exactly with that rate. In this paper, we present tigh…
Douglas-Rachford splitting and the alternating direction method of multipliers (ADMM) can be used to solve convex optimization problems that consist of a sum of two functions. Convergence rate estimates for these algorithms have received much attention lately. In particular, linear convergence rates have been shown by several authors under various assumptions. One such set of assumptions is strong convexity and smoothness of one of the functions in the minimization problem. The authors recently provided a linear convergence rate bound for such problems. In this paper, we show that this rate bo…
Proximal point algorithm has found many applications, and it has been playing fundamental roles in the understanding, design, and analysis of many first-order methods. In this paper, we derive the tight convergence rate in subgradient norm of the proximal point algorithm, which was conjectured by Taylor, Hendrickx and Glineur [SIAM J.~Optim., 27 (2017), pp.~1283--1313]. This sort of convergence results in terms of the residual (sub)gradient norm is particularly interesting when considering dual methods, where the dual residual gradient norm corresponds to the primal distance to feasibility.
We present a novel way of generating Lyapunov functions for proving linear convergence rates of first-order optimization methods. Our approach provably obtains the fastest linear convergence rate that can be verified by a quadratic Lyapunov function (with given states), and only relies on solving a small-sized semidefinite program. Our approach combines the advantages of performance estimation problems (PEP, due to Drori & Teboulle (2014)) and integral quadratic constraints (IQC, due to Lessard et al. (2016)), and relies on convex interpolation (due to Taylor et al. (2017c;b)).
We perform the first tight convergence analysis of the gradient method with varying step sizes when applied to smooth hypoconvex (weakly convex) functions. Hypoconvex functions are smooth nonconvex functions whose curvature is bounded and assumed to belong to the interval $[μ, L]$, with $μ<0$. Our convergence rates improve and extend the existing analysis for smooth nonconvex functions with $L$-Lipschitz gradient (which corresponds to the case $μ=-L$), and smoothly interpolates between that class and the class of smooth convex functions. We obtain our results using the performance estimation f…
Although originally designed and analyzed for convex problems, the alternating direction method of multipliers (ADMM) and its close relatives, Douglas-Rachford splitting (DRS) and Peaceman-Rachford splitting (PRS), have been observed to perform remarkably well when applied to certain classes of structured nonconvex optimization problems. However, partial global convergence results in the nonconvex setting have only recently emerged. In this paper we show how the Douglas-Rachford envelope (DRE), introduced in 2014, can be employed to unify and considerably simplify the theory for devising globa…
The present paper is devoted to estimating the speed of convergence towards consensus for a general class of discrete-time multi-agent systems. In the systems considered here, both the topology of the interconnection graph and the weight of the arcs are allowed to vary as a function of time. Under the hypothesis that some spanning tree structure is preserved along time, and that some nonzero minimal weight of the information transfer along this tree is guaranteed, an estimate of the contraction rate is given. The latter is expressed explicitly as the spectral radius of some matrix depending up…
Peppy: An AI-Assisted Workflow for Tight Convergence Analysis of Optimization Algorithms
Rice University（莱斯大学）；Johns Hopkins University（约翰斯·霍普金斯大学）；Google Inc.（谷歌公司）；Data Science and AI Institute, Johns Hopkins University（约翰斯·霍普金斯大学数据科学与人工智能研究所）
The present paper considers leveraging network topology information to improve the convergence rate of ADMM for decentralized optimization, where networked nodes work collaboratively to minimize the objective. Such problems can be solved efficiently using ADMM via decomposing the objective into easier subproblems. Properly exploiting network topology can significantly improve the algorithm performance. Hybrid ADMM explores the direction of exploiting node information by taking into account node centrality but fails to utilize edge information. This paper fills the gap by incorporating both nod…
The Pennsylvania State University（宾夕法尼亚州立大学）；Massachusetts Institute of Technology（麻省理工学院）
Despite the established convergence theory of Optimistic Gradient Descent Ascent (OGDA) and Extragradient (EG) methods for the convex-concave minimax problems, little is known about the theoretical guarantees of these methods in nonconvex settings. To bridge this gap, for the first time, this paper establishes the convergence of OGDA and EG methods under the nonconvex-strongly-concave (NC-SC) and nonconvex-concave (NC-C) settings by providing a unified analysis through the lens of single-call extra-gradient methods. We further establish lower bounds on the convergence of GDA/OGDA/EG, shedding …
In this paper, we provide a simple convergence analysis of proximal gradient algorithm with Bregman distance, which provides a tighter bound than existing result. In particular, for the problem of minimizing a class of convex objective functions, we show that proximal gradient algorithm with Bregman distance can be viewed as proximal point algorithm that incorporates another Bregman distance. Consequently, the convergence result of the proximal gradient algorithm with Bregman distance follows directly from that of the proximal point algorithm with Bregman distance, and this leads to a simpler …
Tight Stability, Convergence, and Robustness Bounds for Predictive Coding Networks
University of South Florida（南佛罗里达大学）；VERSES AI Research Lab（VERSES AI 研究院）；Vienna University of Technology（维也纳技术大学）；Rochester Institute of Technology（罗切斯特理工学院）
The stability of stochastic Model Predictive Control (MPC) subject to additive disturbances is often demonstrated in the literature by constructing Lyapunov-like inequalities that guarantee closed-loop performance bounds and boundedness of the state, but convergence to a terminal control law is typically not shown. In this work we use results on general state space Markov chains to find conditions that guarantee convergence of disturbed nonlinear systems to terminal modes of operation, so that they converge in probability to a priori known terminal linear feedback laws and achieve time-average…
This paper proposes tight semidefinite relaxations for polynomial optimization. The optimality conditions are investigated. We show that generally Lagrange multipliers can be expressed as polynomial functions in decision variables over the set of critical points. The polynomial expressions can be determined by linear equations. Based on these expressions, new Lasserre type semidefinite relaxations are constructed for solving polynomial optimization. We show that the hierarchy of new relaxations has finite convergence, or equivalently, the new relaxations are tight for a finite relaxation order…
