Exponential lower bounds on fixed-size psd rank and semidefinite extension complexity

TL;DR

This paper establishes exponential lower bounds on the size of semidefinite programming formulations for certain polytopes, extending previous LP bounds to SDPs with fixed-size positive semidefinite blocks, highlighting limitations of practical SDP relaxations.

Contribution

It proves exponential lower bounds for SDP formulations over fixed-size psd cones, generalizing known LP bounds and addressing practical SDP relaxation limitations.

Findings

SDP formulations over fixed-size psd cones require exponential size

Results include LP bounds as a special case when block size is 1

No small SDP formulations exist for the cut polytope with fixed block size d=2

Abstract

There has been a lot of interest recently in proving lower bounds on the size of linear programs needed to represent a given polytope P. In a breakthrough paper Fiorini et al. [Proceedings of 44th ACM Symposium on Theory of Computing 2012, pages 95-106] showed that any linear programming formulation of maximum-cut must have exponential size. A natural question to ask is whether one can prove such strong lower bounds for semidefinite programming formulations. In this paper we take a step towards this goal and we prove strong lower bounds for a certain class of SDP formulations, namely SDPs over the Cartesian product of fixed-size positive semidefinite cones. In practice this corresponds to semidefinite programs with a block-diagonal structure and where blocks have constant size d. We show that any such extended formulation of the cut polytope must have exponential size (when d is fixed). The result of Fiorini et al. for LP formulations is obtained as a special case when d=1. For blocks of size d=2 the result rules out any small formulations using second-order cone programming. Our study of SDP lifts over Cartesian product of fixed-size positive semidefinite cones is motivated mainly from practical considerations where it is well known that such SDPs can be solved more efficiently than general SDPs. The proof of our lower bound relies on new results about the sparsity pattern of certain matrices with small psd rank, combined with an induction argument inspired from the recent paper by Kaibel and Weltge [arXiv:1307.3543] on the LP extension complexity of the correlation polytope.