Coordinate descent methods beyond smoothness and separability

Abstract

This paper deals with convex nonsmooth optimization problems. We introduce a general smooth approximation framework for the original function and apply random (accelerated) coordinate descent methods for minimizing the corresponding smooth approximations. Our framework covers the most important classes of smoothing techniques from the literature. Based on this general framework for the smooth approximation and using coordinate descent type methods we derive convergence rates in function values for the original objective. Moreover, if the original function satisfies a growth condition, then we prove that the smooth approximations also inherits this condition and consequently the convergence rates are improved in this case. We also present a relative randomized coordinate descent algorithm for solving nonseparable minimization problems with the objective function relative smooth along coordinates w.r.t. a (possibly nonseparable) differentiable function. For this algorithm we also derive convergence rates in the convex case and under the growth condition for the objective.