Min-Sum Belief Propagation not working on a chain model with equal unary potentials

27 Views Asked by At

Given is a chain factor graph as presented in the image below with the following properties:

  • Each node can take values 0 or 1
  • All unary potentials are equal (e.g. $U(a) = 0$ for every node)
  • All pairwise potentials are defined as $P(a,b) = \lvert a - b\rvert$

I want to use the Min-Sum Belief Propagation algorithm to find MAP solution to the problem.

Chain factor graph

The obvious answer is that there exist two correct solutions:

  1. All variables have value 0
  2. All variables have value 1

However, the Min-Sum Belief Propagation algorithm, as the final output, gives equal beliefs for 0 and 1 at each label. As the messages are initialized to 0, they remain unchanged after the message passing and updating iterations. That and the equal unary potentials imply equal beliefs. I cannot see how that information can be used to obtain correct results.

Viterbi algorithm solves the problem easily and correctly. AFAIK, Min-Sum Belief Propagation should always find optimal solution when applied on acyclic graphs. But in this concrete example, it doesn't.

This is minimal repro example.

Does the Min-Sum Belief Propagation algorithm has a mode of failure or am I misunderstanding something?