Definability of semidefinite programming and lasserre lower bounds for CSPs

Abstract

We show that the ellipsoid method for solving semi-definite programs (SDPs) can be expressed in fixed-point logic with counting (FPC). This generalizes an earlier result that the optimal value of a linear program can be expressed in this logic.As an application, we establish lower bounds on the number of levels of the Lasserre hierarchy required to solve many optimization problems, namely those that can be expressed as finite-valued constraint satisfaction problems (VCSPs). In particular, we establish a dichotomy on the number of levels of the Lasserre hierarchy that are required to solve the problem exactly. 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 SDP corresponding to any fixed level of the Lasserre hierarchy is interpretable in a VCSP instance by means of FPC formulas. Our definability result of the ellipsoid method then implies that the solution of this 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 earlier Lasserre lower bound results by Schoenebeck [17]. Recently, and independently of the work reported here, a similar linear lower bound in the Lasserre hierarchy for general-valued CSPs has also been announced by Thapper and Zivný [20], using different techniques.