Loopy and generalized belief propagation are popular algorithms for approximate inference in markov random fields and Bayesian networks. Fixed points of these algorithms correspond to extrema of the Bethe and →