Distributed Optimization Based on Gradient-tracking Revisited: Enhancing Convergence Rate via Surrogation

Abstract

We study distributed multiagent optimization over (directed, time-varying) graphs. We consider the minimization of $F+G$ subject to convex constraints, where $F$ is the smooth strongly convex sum of the agent's losses and $G$ is a nonsmooth convex function. We build on the SONATA algorithm: the algorithm employs the use of surrogate objective functions in the agents' subproblems (going thus beyond linearization, such as proximal-gradient) coupled with a perturbed (push-sum) consensus mechanism that aims to track locally the gradient of $F$. SONATA achieves precision $\epsilon>0$ on the objective value in $\mathcal{O}(\kappa_g \log(1/\epsilon))$ gradient computations at each node and $\tilde{\mathcal{O}}\big(\kappa_g (1-\rho)^{-1/2} \log(1/\epsilon)\big)$ communication steps, where $\kappa_g$ is the condition number of $F$ and $\rho$ characterizes the connectivity of the network. This is the first linear rate result for distributed composite optimization; it also improves on existing (non-accelerated) schemes just minimizing $F$, whose rate depends on much larger quantities than $\kappa_g$ (e.g., the worst-case condition number among the agents). When considering in particular empirical risk minimization problems with statistically similar data across the agents, SONATA employing high-order surrogates achieves precision $\epsilon>0$ in $\mathcal{O}\big((\beta/\mu) \log(1/\epsilon)\big)$ iterations and $\tilde{\mathcal{O}}\big((\beta/\mu) (1-\rho)^{-1/2} \log(1/\epsilon)\big)$ communication steps, where $\beta$ measures the degree of similarity of the agents' losses and $\mu$ is the strong convexity constant of $F$. Therefore, when $\beta/\mu < \kappa_g$, the use of high-order surrogates yields provably faster rates than what achievable by first-order models; this is without exchanging any Hessian matrix over the network.