Graph Matching: Relax at Your Own Risk

We present theoretical results addressing the ubiquitous problem of graph matching. Two correlated random Bernoulli graphs have, via their correlation, a natural underlying correspondence between their two vertex sets. We prove, in this very broad random-graph model, and under very general conditions, that a popular convex relaxation of graph matching---solved exactly---will almost always fail to provide the underlying correspondence and that, under mild conditions, a nonconvex relaxation of graph matching---solved exactly---will almost always produce the underlying correspondence. Experimental results illuminate these theoretical findings and provide hints on how to mitigate the risky use of popular convex relaxations in graph matching.