Abstract:The convergence analysis of optimization algorithms using continuous-time dynamical systems has received much attention in recent years. In this paper, we investigate applications of these systems to analyze the convergence of linearized proximal ADMM algorithms for nonconvex composite optimization, whose objective function is the sum of a continuously differentiable function and a composition of a possibly nonconvex function with a linear operator. We first derive a first-order differential inclusion for the linearized proximal ADMM algorithm, LP-ADMM. Both the global convergence and the convergence rates of the generated trajectory are established with the use of Kurdyka-Łojasiewicz (KL) property. Then, a stochastic variant, LP-SADMM, is delved into an investigation for finite-sum nonconvex composite problems. Under mild conditions, we obtain the stochastic differential equation corresponding to LP-SADMM, and demonstrate the almost sure global convergence of the generated trajectory by leveraging the KL property. Based on the almost sure convergence of trajectory, we construct a stochastic process that converges almost surely to an approximate critical point of objective function, and derive the expected convergence rates associated with this stochastic process. Moreover, we propose an accelerated LP-SADMM that incorporates Nesterov's acceleration technique. The continuous-time dynamical system of this algorithm is modeled as a second-order stochastic differential equation. Within the context of KL property, we explore the related almost sure convergence and expected convergence rates.

Local Linear Convergence of an ADMM-Type Splitting Framework for Equality Constrained Optimization

Linear Convergence of the Alternating Direction Method of Multipliers for a Class of Convex Optimization Problems

A Bregman-Style Improved ADMM and its Linearized Version in the Nonconvex Setting: Convergence and Rate Analyses

Linearized ADMM for Nonconvex Nonsmooth Optimization with Convergence Analysis

Convergence Analysis of Generalized ADMM with Majorization for Linearly Constrained Composite Convex Optimization

Sublinear and Linear Convergence of Modified ADMM for Distributed Nonconvex Optimization

A Bregman-style Partially Symmetric Alternating Direction Method of Multipliers for Nonconvex Multi-block Optimization

An over-relaxed ADMM for separable convex programming and its applications to statistical learning

Understanding the ADMM Algorithm via High-Resolution Differential Equations

Stochastic linearized generalized alternating direction method of multipliers: Expected convergence rates and large deviation properties

Differentiable Linearized ADMM

Dynamical convergence analysis for nonconvex linearized proximal ADMM algorithms

A Convergence Analysis of Nonlinearly Constrained ADMM in Deep Learning

A Bi-level Globalization Strategy for Non-convex Consensus ADMM and ALADIN

Bregman Proximal Linearized ADMM for Minimizing Separable Sums Coupled by a Difference of Functions

On the Convergence Analysis of the Alternating Direction Method of Multipliers with Three Blocks

Adaptive Relaxed ADMM: Convergence Theory and Practical Implementation

ADMM for Nonsmooth Composite Optimization under Orthogonality Constraints

Partition-Insensitive Parallel ADMM Algorithm for High-dimensional Linear Models

Global Complexity Bound of a Proximal ADMM for Linearly Constrained Nonseparable Nonconvex Composite Programming