A.1 Nonconvexstochasticoptimization Wegiveproofsofthetheoremsinsection3. From Assumption 2,forthemini-batch gradientfSk(xk) = 1nk