Semi-infinite programming using high-degree polynomial interpolants and semidefinite programming

TL;DR

This paper introduces a novel approach combining rational approximation and polynomial programming to reformulate semi-infinite programs as well-conditioned semidefinite programs, enabling high-degree polynomial solutions without numerical issues.

Contribution

It develops a sum-of-squares interpolant method that maintains good conditioning for high-degree polynomials, facilitating efficient semidefinite programming solutions for semi-infinite constraints.

Findings

Conditioning does not deteriorate with increasing polynomial degree.

Problems with hundreds of degrees can be solved without difficulty.

Memory limits the polynomial degree to about 1000 in practice.

Abstract

In a common formulation of semi-infinite programs, the infinite constraint set is a requirement that a function parametrized by the decision variables is nonnegative over an interval. If this function is sufficiently closely approximable by a polynomial or a rational function, then the semi-infinite program can be reformulated as an equivalent semidefinite program. Solving this semidefinite program is challenging if the polynomials involved are of high degree, due to numerical difficulties and bad scaling arising both from the polynomial approximations and from the fact that the semidefinite programming constraints coming from the sum-of-squares representation of nonnegative polynomials are badly scaled. We combine rational function approximation techniques and polynomial programming to overcome these numerical difficulties, using sum-of-squares interpolants. Specifically, it is shown that the conditioning of the reformulations using sum-of-squares interpolants does not deteriorate with increasing degrees, and problems involving sum-of-squares interpolants of hundreds of degrees can be handled without difficulty. The proposed reformulations are sufficiently well scaled that they can be solved easily with every commonly used semidefinite programming solver, such as SeDuMi, SDPT3, and CSDP. Motivating applications include convex optimization problems with semi-infinite constraints and semidefinite conic inequalities, such as those arising in the optimal design of experiments. Numerical results align with the theoretical predictions; in the problems considered, available memory was the only factor limiting the degrees of polynomials, to approximately 1000.