Exact penalties for maximum-edge bicliques in bipartite graphs

Abstract

Given a bipartite graph $G$, we denote by $g(G)$ the largest product $|A||B|$ over pairs of vertex sets $A$ and $B$, one on each side, with no edge between them. These pairs are the bicliques of the bipartite complement of $G$; hence, computing $g(G)$ is the maximum-edge biclique problem, which is NP-hard, as shown by Peeters in 2003. Gillis and Glineur characterized it by a continuous quadratic program, in which a large enough penalty on the edges of $G$ makes the global maximizers correspond to the maximum pairs and the local maximizers to the maximal ones. We propose a new threshold for the penalty, which depends only on $g(G)$. For any penalty at or above it, the global maximizers correspond exactly to the maximum pairs, and every local maximizer with positive value corresponds to a maximal pair. This threshold is never larger than theirs and is sharp up to an additive $1/2$. Spectral and semidefinite bounds give computable penalties; in experiments, a simple local ascent succeeds far more often with the semidefinite one than with the earlier thresholds, although on our graphs that bound already reveals $g(G)$. The penalty formulation, whose optimal value is $\sqrt{g(G)}/2$, is also a copositive formulation of $g(G)$. Strengthening a semidefinite bound of Laurent, Polak and Vargas with the completely positive cone gives exactly $\sqrt{g(G)}/2$, while the doubly nonnegative strengthening is already inexact on six vertices. Finally, Pólya's linear programming approximations of the copositive cone become finite at an explicit level and converge to $\sqrt{g(G)}/2$ without ever being exact.