Delayed supermartingale convergence lemmas for stochastic approximation with Nesterov momentum

TL;DR

This paper develops delayed supermartingale convergence lemmas to establish almost sure convergence of stochastic approximation methods with Nesterov momentum, addressing delayed information challenges in acceleration.

Contribution

It introduces a new framework based on delayed supermartingale lemmas that proves almost sure convergence for various stochastic optimization algorithms with Nesterov acceleration.

Findings

Framework applies to stochastic subgradient and proximal methods

Methods with Nesterov acceleration achieve almost sure convergence

Numerical experiments confirm theoretical results

Abstract

This paper focus on the convergence of stochastic approximation with Nesterov momentum. Nesterov acceleration has proven effective in machine learning for its ability to reduce computational complexity. The issue of delayed information in the acceleration term remains a challenge to achieving the almost sure convergence. Based on the delayed supermatingale convergence lemmas, we give a series of framework for almost sure convergence. Our framework applies to several widely-used random iterative methods, such as stochastic subgradient methods, the proximal Robbins-Monro method for general stochastic optimization, and the proximal stochastic subgradient method for composite optimization. Through the applications of our framework, these methods with Nesterov acceleration achieve almost sure convergence. And three groups of numerical experiments is to check out theoretical results.