20

Decentralized Optimization with Topology-Independent Communication

Main:30 Pages
5 Figures
Bibliography:6 Pages
1 Tables
Abstract

Distributed optimization requires nodes to coordinate, yet full synchronization scales poorly. When nn nodes collaborate through mm pairwise regularizers, standard methods demand O(m)\mathcal{O}(m) communications per iteration. This paper proposes randomized local coordination: each node independently samples one regularizer uniformly and coordinates only with nodes sharing that term. This exploits partial separability, where each regularizer GjG_j depends on a subset Sj{1,,n}S_j \subseteq \{1,\ldots,n\} of nodes. For graph-guided regularizers where Sj=2|S_j|=2, expected communication drops to exactly 2 messages per iteration. This method achieves O~(ε2)\tilde{\mathcal{O}}(\varepsilon^{-2}) iterations for convex objectives and under strong convexity, O(ε1)\mathcal{O}(\varepsilon^{-1}) to an ε\varepsilon-solution and O(log(1/ε))\mathcal{O}(\log(1/\varepsilon)) to a neighborhood. Replacing the proximal map of the sum jGj\sum_j G_j with the proximal map of a single randomly selected regularizer GjG_j preserves convergence while eliminating global coordination. Experiments validate both convergence rates and communication efficiency across synthetic and real-world datasets.

View on arXiv
Comments on this paper