An Analysis Tool for Push-Sum Based Distributed Optimization

TL;DR

This paper introduces an analysis tool for push-sum algorithms in distributed optimization, improving convergence rate proofs and enabling new algorithm variants over directed graphs.

Contribution

It establishes explicit probability sequences and quadratic Lyapunov functions for push-sum algorithms, enhancing convergence analysis and proposing a flexible heterogeneous algorithm.

Findings

Subgradient-push converges at O(1/√t) for convex functions.

Stochastic gradient-push converges at O(1/t) for strongly convex functions.

The proposed methods achieve optimal convergence rates matching single-agent algorithms.

Abstract

The push-sum algorithm is probably the most important distributed averaging approach over directed graphs, which has been applied to various problems including distributed optimization. This paper establishes the explicit absolute probability sequence for the push-sum algorithm, and based on which, constructs quadratic Lyapunov functions for push-sum based distributed optimization algorithms. As illustrative examples, the proposed novel analysis tool can improve the convergence rates of the subgradient-push and stochastic gradient-push, two important algorithms for distributed convex optimization over unbalanced directed graphs. Specifically, the paper proves that the subgradient-push algorithm converges at a rate of $O(1/\sqrt{t})$ for general convex functions and stochastic gradient-push algorithm converges at a rate of $O(1/t)$ for strongly convex functions, over time-varying unbalanced directed graphs. Both rates are respectively the same as the state-of-the-art rates of their single-agent counterparts and thus optimal, which closes the theoretical gap between the centralized and push-sum based (sub)gradient methods. The paper further proposes a heterogeneous push-sum based subgradient algorithm in which each agent can arbitrarily switch between subgradient-push and push-subgradient. The heterogeneous algorithm thus subsumes both subgradient-push and push-subgradient as special cases, and still converges to an optimal point at an optimal rate. The proposed tool can also be extended to analyze distributed weighted averaging.