Title:Instance-specific linear relaxations of semidefinite optimization problems

Abstract:We present a simple yet flexible systematic way to significantly improve linear relaxations of semidefinite optimization programs (SDPs) using instance-specific information. We get inspiration for our approach by studying the celebrated SDP relaxation for max cut (GW) due to Poljak and Rendl [1995] and analyzed by Goemans and Williamson [1995]. Using the instance at hand, we provide a pair of closely related compact linear programs that sandwich the optimal value of the GW semidefinite program. The instance specific information allows us to trivially avoid hardness of approximation results of semidefinite programs using linear ones, such as Braun et al. [2015] and Kothari et al. [2021].
We give sufficient conditions that guarantee that the optimal value of both our programs match the optimal value of the semidefinite relaxation. We use these conditions to prove that our two bounds give the same value as the GW SDP for max cut on distance-regular graphs, and in particular, strongly regular graphs such as cycles and the Petersen graph. Further, we show that the approximation of the upper bounding problem is strictly better than the eigenvalue bounds of Mohar and Poljak [1990], Alon and Sudakov [2000] for max cut.
To the best of our knowledge, no previously proposed linear program proposed for max cut has this guarantee. We extensively test our methodology on synthetic and real graphs, and show how our ideas perform in practice. Even though our methodology is inspired by the SDP relaxation for max cut, we show experimentally that our ideas can be applied successfully to obtain good solutions to other computationally hard SDPs such as sparse PCA and the Lovasz Theta number.