1: \begin{abstract}

2: 

3: Max-product ``belief propagation"  is an iterative, local,

4: message-passing algorithm for finding the maximum a posteriori (MAP)

5: assignment of a discrete probability distribution specified by a

6: graphical model. Despite the spectacular success of the algorithm in

7: many application areas such as iterative decoding, computer vision

8: and combinatorial optimization which involve graphs with many

9: cycles, theoretical results about both correctness and convergence

10: of the algorithm are known in few cases \cite{WeissFreeman,

11: wainwright04, YFW00-22,Urbanke}.

12: 

13: In this paper we consider the problem of finding the Maximum

14: Weight Matching (MWM) in a weighted complete bipartite graph.  We

15: define a probability distribution on the bipartite graph whose MAP

16: assignment corresponds to the MWM.  We use the max-product

17: algorithm for finding the MAP of this distribution or

18: equivalently, the MWM on the bipartite graph.  Even though the

19: underlying bipartite graph has many short cycles, we find that

20: surprisingly, the max-product algorithm always converges to the

21: correct MAP assignment as long as the MAP assignment is unique. We

22: provide a bound on the number of iterations required by the

23: algorithm and evaluate the computational cost of the algorithm.

24: We find that for a graph of size $n$, the computational cost

25: of the algorithm scales as $O(n^3)$, which is the same as the

26: computational cost of the best known algorithm. Finally, we

27: establish the precise relation between the max-product algorithm

28: and the celebrated {\em auction} algorithm proposed by Bertsekas.

29: This suggests possible connections between dual algorithm and max-product

30: algorithm for discrete optimization problems.

31: 

32: \end{abstract}

33: