EXTRA is a popular method for dencentralized distributed optimization and has broad applications. This paper revisits EXTRA. First, we give a sharp complexity analysis for EXTRA with the improved O((μL+1−σ2(W)1)logϵ(1−σ2(W))1) communication and computation complexities for μ-strongly convex and L-smooth problems, where σ2(W) is the second largest singular value of the weight matrix W. When the strong convexity is absent, we prove the O((ϵL+1−σ2(W)1)log1−σ2(W)1) complexities. Then, we use the Catalyst framework to accelerate EXTRA and obtain the O(μ(1−σ2(W))Llogμ(1−σ2(W))Llogϵ1) communication and computation complexities for strongly convex and smooth problems and the O(ϵ(1−σ2(W))Llogϵ(1−σ2(W))1) complexities for non-strongly convex ones. Our communication complexities of the accelerated EXTRA are only worse by the factors of (logμ(1−σ2(W))L) and (logϵ(1−σ2(W))1) from the lower complexity bounds for strongly convex and non-strongly convex problems, respectively.
Cite
@article{arxiv.2002.10110,
title = {Revisiting EXTRA for Smooth Distributed Optimization},
author = {Huan Li and Zhouchen Lin},
journal= {arXiv preprint arXiv:2002.10110},
year = {2020}
}