This paper investigates the polytope associated with the classical standard linearization technique for the unconstrained optimization of multilinear polynomials in 0-1 variables. A new class of valid inequalities, called 2-links, is introduced to strengthen the LP relaxation of the standard linearization. The addition of the 2-links to the standard linearization inequalities provides a complete description of the convex hull of integer solutions for the case of functions consisting of at most two nonlinear monomials. For the general case, various computational experiments show that the 2-links improve both the standard linearization bound and the computational performance of exact branch & cut methods. The improvements are especially significant for a class of instances inspired from the image restoration problem in computer vision. The magnitude of this effect is rather surprising in that the 2-links are in relatively small number (quadratic in the number of terms of the objective function).

The problem of minimizing a pseudo-Boolean function, that is, a real-valued function of 0-1 variables, arises in many applications. A quadratization is a reformulation of this nonlinear problem into a quadratic one, obtained by introducing a set of auxiliary binary variables. A desirable property for a quadratization is to introduce a small number of auxiliary variables. We present upper and lower bounds on the number of auxiliary variables required to define a quadratization for several classes of specially structured functions, such as functions with many zeros, symmetric, exact k-out-of-n, at least kout-of-n and parity functions, and monomials with a positive coefficient, also called positive monomials. Most of these bounds are logarithmic in the number of original variables, and we prove that they are best possible for several of the classes under consideration. For positive monomials and for some other symmetric functions, a logarithmic bound represents a significant improvement with respect to the best bounds previously published, which are linear in the number of original variables. Moreover, the case of positive monomials is particularly interesting: indeed, when a pseudo-Boolean function is represented by its unique multilinear polynomial expression, a quadratization can be obtained by separately quadratizing its monomials.

No abstract

We investigate the problem of optimizing a multilinear polynomial f in 0-1 variables and characterize instances for which the classical standard linearization procedure guarantees integer optimal solutions. We show that the standard linearization polytope P H is integer exactly when the hypergraph H defined by the higher-degree monomials of f is Berge-acyclic, or equivalently, when the matrix defining P H is balanced. This characterization follows from more general conditions that guarantee integral optimal vertices for a relaxed formulation depending on the sign pattern of the monomials of f .

The Nemhauser-Trotter theorem states that the standard linear programming (LP) formulation for the stable set problem has a remarkable property, also known as (weak) persistency: for every optimal LP solution that assigns integer values to some variables, there exists an optimal integer solution in which these variables retain the same values. While the standard LP is defined by only non-negativity and edge constraints, a variety of stronger LP formulations have been studied and one may wonder whether any of them is persistent as well. We show that any stronger LP formulation that satisfies mild conditions cannot be persistent on all graphs, unless it is always equal to the stable-set polytope.