Lasserre Lower Bounds and Definability of Semidefinite Programming

TL;DR

This paper establishes a logical and complexity-theoretic dichotomy for the Lasserre hierarchy levels needed to solve VCSPs exactly, linking logical definability with semidefinite programming bounds.

Contribution

It proves that if a VCSP isn't solved by its basic LP relaxation, then no sub-linear Lasserre levels can solve it exactly, using logical undefinability techniques.

Findings

Lower bounds are derived via fixed-point logic with counting.

LP relaxations and Lasserre SDP levels are interpretable in VCSPs.

A dichotomy result for Lasserre hierarchy levels in VCSPs and MAXCSPs.

Abstract

For a large class of optimization problems, namely those that can be expressed as finite-valued constraint satisfaction problems (VCSPs), we establish a dichotomy on the number of levels of the Lasserre hierarchy of semi-definite programs (SDPs) that are required to solve the problem exactly. In particular, we show that if a finite-valued constraint problem is not solved exactly by its basic linear programming relaxation, it is also not solved exactly by any sub-linear number of levels of the Lasserre hierarchy. The lower bounds are established through logical undefinability results. We show that the linear programming relaxation of the problem, as well as the SDP corresponding to any fixed level of the Lasserre hierarchy is interpretable in a VCSP instance by means of formulas of fixed-point logic with counting. We also show that the solution of an SDP can be expressed in this logic. Together, these results give a way of translating lower bounds on the number of variables required in counting logic to express a VCSP into lower bounds on the number of levels required in the Lasserre hierarchy to eliminate the integrality gap. As a special case, we obtain the same dichotomy for the class of MAXCSP problems, generalizing some earlier Lasserre lower bound results of Schoenebeck.