Valid inequalities and global solution algorithm for Quadratically Constrained Quadratic Programs

Abstract: We consider the exact solution of problem $(QP)$ that consists in minimizing a quadratic function subject to quadratic constraints. Starting from the classical convex relaxation that uses the McCormick's envelopes, we introduce 12 inequalities that are derived from the ranges of the variables of $(QP)$. We prove that these general Triangle inequalities cut feasible solutions of the McCormick's envelopes. We then show how we can compute a convex relaxation $(P^*)$ which optimal value equals to the "Shor's plus RLT plus Triangle" semi-definite relaxation of $(QP)$ that includes the new inequalities. We also propose a heuristic for solving this huge semi-definite program that serves as a separation algorithm. We then solve $(QP)$ to global optimality with a branch-and-bound based on $(P^*)$. Moreover, as the new inequalities involved the lower and upper bounds on the original variables of $(QP)$, their use in a branch-and-bound framework accelerates the whole process. We show on the unitbox instances that our method outperforms the compared solvers.