Quality of LP-Based Approximations for Highly Combinatorial Problems

Lucian Leahu, Carla P. Gomes

Abstract: We study the quality of LP-based approximation methods for pure combinatorial problems. We found that the quality of the LP-relaxation is a direct function of the underlying constrainedness of the combinatorial problem. More specifically, we identify a novel phase transition phenomenon in the solution integrality of the relaxation. The solution quality of approximation schemes degrades substantially near phase transition boundaries. Our findings are consistent over a range of LP-based approximation schemes. We also provide results on the extent to which LP relaxations can provide a global perspective of the search space and therefore be used as a heuristic to guide a complete solver.