Polymatroids and mean-risk minimization in discrete optimization