Approximating optimization problems over convex functions

TL;DR

This paper introduces a finite difference method using semidefinite programming to approximate solutions to convex optimization problems, achieving convergence with fewer constraints and applicable in higher dimensions.

Contribution

It proposes a novel finite difference approximation with positive semidefinite programs that converges even for non-smooth solutions, in any dimension, with linear constraints.

Findings

Convergence proven under general conditions

Applicable to non-smooth solutions

Works in any dimension with linear constraints

Abstract

Many problems of theoretical and practical interest involve finding an optimum over a family of convex functions. For instance, finding the projection on the convex functions in $H^k(\Omega)$, and optimizing functionals arising from some problems in economics. In the continuous setting and assuming smoothness, the convexity constraints may be given locally by asking the Hessian matrix to be positive semidefinite, but in making discrete approximations two difficulties arise: the continuous solutions may be not smooth, and functions with positive semidefinite discrete Hessian need not be convex in a discrete sense. Previous work has concentrated on non-local descriptions of convexity, making the number of constraints to grow super-linearly with the number of nodes even in dimension 2, and these descriptions are very difficult to extend to higher dimensions. In this paper we propose a finite difference approximation using positive semidefinite programs and discrete Hessians, and prove convergence under very general conditions, even when the continuous solution is not smooth, working on any dimension, and requiring a linear number of constraints in the number of nodes. Using semidefinite programming codes, we show concrete examples of approximations to problems in two and three dimensions.