Various applications involve assigning discrete label values to a collection
of objects based on some pairwise noisy data. Due to the discrete---and hence
nonconvex---structure of the problem, computing the optimal assignment
(e.g.~maximum likelihood assignment) becomes intractable at