Successive Lagrangian relaxation algorithm for nonconvex quadratic optimization

Abstract

Optimization problems whose objective function and constraints are quadratic polynomials are called quadratically constrained quadratic programs (QCQPs). QCQPs are NP-hard in general and are important in optimization theory and practice. There have been many studies on solving QCQPs approximately. Among them, a semidefinite program (SDP) relaxation is a well-known convex relaxation method. In recent years, many researchers have tried to find better relaxed solutions by adding linear constraints as valid inequalities. On the other hand, the SDP relaxation requires a long computation time, and it has high space complexity for large-scale problems in practice; therefore, the SDP relaxation may not be useful for such problems. In this paper, we propose a new convex relaxation method that is weaker but faster than SDP relaxation methods. The proposed method transforms a QCQP into a Lagrangian dual optimization problem and successively solves subproblems while updating the Lagrange multipliers. The subproblem in our method is a QCQP with only one constraint for which we propose an efficient algorithm. Numerical experiments confirm that our method can quickly find a relaxed solution with an appropriate termination condition.