Distributed Gradient Tracking for Unbalanced Optimization With Different Constraint Sets
Songsong Cheng, Shu Liang, Yuan Fan, Yiguang Hong
Abstract
Gradient tracking methods have become popular for distributed optimization in recent years, partially because they achieve linear convergence using only a constant step-size for strongly convex optimization. In this article, we construct a counterexample on constrained optimization to show that direct extension of gradient tracking by using projections cannot guarantee the correctness. Then, we propose projected gradient tracking algorithms with diminishing step-sizes rather than a constant one for distributed strongly convex optimization with different constraint sets and unbalanced graphs. Our basic algorithm can achieve <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$O(\ln T/{T})$</tex-math></inline-formula> convergence rate. Moreover, we design an epoch iteration scheme and improve the convergence rate as <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$O(1/{T})$</tex-math></inline-formula> .