Optimal Gradient Sliding and its Application to Optimal Distributed Optimization Under Similarity

We study structured convex optimization problems, with additive objective r := p + q, where r is (µ-strongly) convex, q is L q -smooth and convex, and p is L psmooth, possibly nonconvex. For such a class of problems, we proposed an inexact accelerated gradient sliding method that can skip the gradient computation for one of these components while still achieving optimal complexity of gradient calls of p and q, that is, O( L p /µ) and O( L q /µ), respectively. This result is much sharper than the classic black-box complexity O( (L p + L q )/µ), especially when the difference between L q and L q is large. We then apply the proposed method to solve distributed optimization problems over master-worker architectures, under agents' function similarity, due to statistical data similarity or otherwise. The distributed algorithm achieves for the first time lower complexity bounds on both communication and local gradient calls, with the former having being a longstanding open problem. Finally the method is extended to distributed saddleproblems (under function similarity) by means of solving a class of variational inequalities, achieving lower communication and computation complexity bounds.