Beyond Fixed Local Updates: An Adaptive Decentralized Quasi-Newton Method Free of Stepsize Degradation
read the original abstract
This paper proposes a novel Adaptive Decentralized Quasi-Newton (AdaDQN) method for solving smooth nonconvex optimization problems over undirected networks. While existing decentralized algorithms with multiple local updates suffer from a pessimistic theoretical bottleneck, where the convergence stepsize must be inversely proportional to the number of local updates, our work overcomes this limitation. We revisit local update methods through a Majorization-Minimization (MM) lens and establish a Robust Inexact Algorithm (RIA) framework. Inspired by this framework, AdaDQN integrates a safeguarded consensus-aware termination criterion to dynamically balance local computational gains against consensus error, complemented by a scalable memoryless BFGS update and an event-triggered communication protocol to significantly reduce communication overhead. We establish the global convergence of AdaDQN to a first-order stationary point. Crucially, we prove that the convergence stepsize bound is decoupled from the reciprocal of maximum number of local updates. Numerical experiments demonstrate that AdaDQN achieves a superior computation-communication tradeoff, outperforming state-of-the-art decentralized methods across various performance metrics.
This paper has not been read by Pith yet.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.