We consider the problem of stochastic convex optimization under convex
constraints. We analyze the behavior of a natural variance reduced proximal
gradient (VRPG) algorithm for this problem. Our main result is a non-asymptotic
guarantee for VRPG algorithm. Contrary to minimax worst cas