An improved linearization strategy for zero-one quadratic programming problems

Abstract

We present a new linearized model for the zero-one quadratic programming problem, whose size is linear in terms of the number of variables in the original nonlinear problem. Our derivation yields three alternative reformulations, each varying in model size and tightness. We show that our models are at least as tight as the one recently proposed in [7], and examine the theoretical relationship of our models to a standard linearization of the zero-one quadratic programming problem. Finally, we demonstrate the efficacy of solving each of these models on a set of randomly generated test instances.

Non-convex mixed-integer nonlinear programming: A survey

“…Chaovalitwongse et al [97] and Sherali and Smith [98] provide recent, conceptually different O(n) linearization approaches.…”

Optimized cooperative streaming in wireless mesh networks

“…The norm instead prefers solutions that are close to the minimum O I and, among two potential solutions with the same O I , it prefers the one in which the weights of L(F 0 ) are more fairly distributed. This problem is a zero-one quadratic programming problem [29], similar to a quadratic knapsack problem [8]. In this kind of problems one wants to minimize the value of an expression c T x+x T Qx where x = {0, 1} n is an array of binary variables, c 2 R n and Q is a symmetric matrix of size n ⇥ n. The minimization is subject to a constraint of the kind h T x + x T Gx > g where h is an array of size n, G a symmetric matrix of size n ⇥ n and g some real value.…”

The reliable hub-and-spoke design problem: Models and algorithms

“…For example, for a small size problem with 10 nodes, linearization leads to roughly 100,000 new variables and 300, 000 new constraints, which is very difficult for professional solvers to generate solutions in a reasonable time. To address this issue, we adopt a recent linearization strategy [6,36,22] and create the following compact reformulation CptLinear. We mention that this linear formula distinguishes itself by the fact that the number of variables does little change and the number of constraints just linearly increases.…”

“…Results in [6,36,22] show that applying this type of linearization techniques to a 0 − 1 quadratic program will just linearly increase the number of variables and constraints. As…”