A Semidefinite Hierarchy for Disjointly Constrained Multilinear Programming

TL;DR

This paper introduces a semidefinite hierarchy based on sum-of-squares polynomials to solve disjointly constrained multilinear programming problems, proving finite convergence under generic conditions.

Contribution

It develops a novel hierarchy of semidefinite relaxations that converges finitely to the multilinear program's optimal value, with applications to game theory and polytope containment.

Findings

Hierarchy converges finitely in generic cases

Nash equilibria can be computed finitely using sum-of-squares

Polytope containment problem decidable finitely under geometric conditions

Abstract

Disjointly constrained multilinear programming concerns the problem of maximizing a multilinear function on the product of finitely many disjoint polyhedra. While maximizing a linear function on a polytope (linear programming) is known to be solvable in polynomial time, even bilinear programming is NP-hard. Based on a reformulation of the problem in terms of sum-of-squares polynomials, we study a hierarchy of semidefinite relaxations to the problem. It follows from the general theory that the sequence of optimal values converges asymptotically to the optimal value of the multilinear program. We show that the semidefinite hierarchy converges generically in finitely many steps to the optimal value of the multilinear problem. We outline two applications of the main result. For nondegenerate bimatrix games, a Nash equilibrium can be computed by the sum of squares approach in finitely many steps. Under an additional geometric condition, the NP-complete containment problem for projections of $\mathcal{H}$-polytopes can be decided in finitely many steps.