Constraint Propagation as Information Maximization