Title: Algorithms for stochastic optimization with expectation constraints

Abstract: This paper considers the problem of minimizing an expectation function over a closed convex set, coupled with functional constraints given in the form of expectation. We present a new stochastic approximation (SA) type algorithm, namely the alternating mirror-descent SA (AMD-SA) algorithm, to minimize these expectation constrained problems, and show that it exhibits the optimal rate of convergence when the objective and constraint functions are convex or strongly convex. We also present a variant of this algorithm to solve a special class of structured nonconvex problems and establish its rate of convergence. It is worth noting that all theses algorithms are primal methods which do not require the estimation of the dual variables. We also provide some promising preliminary numerical results for these algorithms applied to some portfolio optimization and machine learning problems.