Optimal convergence rates of totally asynchronous optimization

TL;DR

This paper establishes explicit, order-optimal convergence rates for asynchronous optimization algorithms like PIAG and Async-BCD under total asynchrony, revealing how delay growth impacts convergence times.

Contribution

It provides the first explicit convergence rates for these algorithms under total asynchrony, extending beyond bounded delay assumptions.

Findings

Derived order-optimal convergence rates under total asynchrony

Showed how delay growth affects convergence times

Validated theoretical results with numerical example

Abstract

Asynchronous optimization algorithms are at the core of modern machine learning and resource allocation systems. However, most convergence results consider bounded information delays and several important algorithms lack guarantees when they operate under total asynchrony. In this paper, we derive explicit convergence rates for the proximal incremental aggregated gradient (PIAG) and the asynchronous block-coordinate descent (Async-BCD) methods under a specific model of total asynchrony, and show that the derived rates are order-optimal. The convergence bounds provide an insightful understanding of how the growth rate of the delays deteriorates the convergence times of the algorithms. Our theoretical findings are demonstrated by a numerical example.