An SDP-Based Approach for Computing the Stability Number of a Graph

TL;DR

This paper introduces new relaxations and a branch-and-bound algorithm to efficiently compute the stability number of a graph, improving runtime while maintaining tight bounds compared to previous SDP-based methods.

Contribution

The paper presents two novel relaxations for faster stability number bounds and integrates them into a B&B algorithm, enhancing efficiency over existing SDP approaches.

Findings

New relaxations reduce computation time without losing bound quality.

The B&B algorithm with new relaxations explores fewer nodes.

Existing bounds by Gaar and Rendl significantly cut down search space.

Abstract

Finding the stability number of a graph, i.e., the maximum number of vertices of which no two are adjacent, is a well known NP-hard combinatorial optimization problem. Since this problem has several applications in real life, there is need to find efficient algorithms to solve this problem. Recently, Gaar and Rendl enhanced semidefinite programming approaches to tighten the upper bound given by the Lov\'asz theta function. This is done by carefully selecting some so-called exact subgraph constraints (ESC) and adding them to the semidefinite program of computing the Lov\'asz theta function. First, we provide two new relaxations that allow to compute the bounds faster without substantial loss of the quality of the bounds. One of these two relaxations is based on including violated facets of the polytope representing the ESCs, the other one adds separating hyperplanes for that polytope. Furthermore, we implement a branch and bound (B&B) algorithm using these tightened relaxations in our bounding routine. We compare the efficiency of our B&B algorithm using the different upper bounds. It turns out that already the bounds of Gaar and Rendl drastically reduce the number of nodes to be explored in the B&B tree as compared to the Lov\'asz theta bound. However, this comes with a high computational cost. Our new relaxations improve the run time of the overall B&B algorithm, while keeping the number of nodes in the B&B tree small.