Given a graphical model, one of the most useful queries is to find the most likely configuration of its variables. This task, known as the maximum a posteriori (MAP) problem, can be solved efficiently via message passing techniques when the graph is a tree, but is NPhard for general graphs. Jebara (2009) shows that the MAP problem can be converted into the stable set problem, which can be solved in polynomial time for a broad class of graphs known as perfect graphs via a linear programming

more »

... ation technique. This is a result of great theoretical interest. However, the article additionally claims that max-product linear programming (MPLP) message passing techniques of Globerson and Jaakkola (2007) are also guaranteed to solve these problems exactly and efficiently. We investigate this claim, show that it does not hold, and repair it with alternative message passing algorithms.