SOLVING QUADRATIC PROGRAMMING BY CUTTING PLANES

We propose new cutting planes for strengthening the linear relaxations that appear in the solution of nonconvex quadratic problems with linear constraints. By a famous result of Motzkin and Straus, these problems are connected to the clique number of a graph. Our cuts are derived in the context of a spatial branch-and-bound algorithm where linearization variables are introduced to represent products. Their validity is based on the result of Motzkin and Straus, in that it depends on the clique number of certain graphs. For convenience, we derive our cuts for the special case of the so-called standard quadratic programs, where the only (linear) constraint imposes that variables must belong to a simplex. We specifically consider cuts that correspond to carefully constructed complete bipartite graphs. We study the relation between these cuts and the classical ones obtained at the first level of the reformulation-linearization technique. By studying this relation, we derive a new type of valid inequalities that generalizes both types of cuts, and thus leading to a large family of cutting planes to strengthen the linear relaxation. We present extensive computational results using the different cutting planes we propose within the spatial branch and bound implemented by the commercial solver CPLEX. We show that our cuts allow us to obtain a significantly better bound than reformulation-linearization cuts and reduce computing times for global optimality. Finally, we formally establish the scaling necessary for using the proposed cuts to solve nonconvex quadratic problems with linear constraints, e.g., quadratic knapsack problems.