Inequalities in completely convex stochastic programming

A two-stage stochastic programming problem in which the random variable enters in a convex manner is called completely convex. For such problems we give a sequence of inequalities and equalities showing the equivalence of optimality over plans and optimality of a two-stage procedure related to dynamic programming and giving upper bounds on the expected value of perfect information. Our assumptions are the weakest possible to guarantee the results in the completely convex case and supersede previous related results which have received erroneous proofs or have been established under highly restrictive conditions. In the course of our argument we exhibit a new measurable selection theorem and a rather general form of Jensen's inequality. We also present a multistage generalization of our central theorem.