1: \begin{abstract}

2: An important part of problems in statistical physics and computer

3: science can be expressed as the computation of marginal probabilities

4: over a Markov Random Field. The belief propagation algorithm, which is

5: an exact procedure to compute these marginals when the underlying

6: graph is a tree, has gained its popularity as an efficient way to

7: approximate them in the more general case. In this paper, we focus on

8: an aspect of the algorithm that did not get that much attention in the

9: literature, which is the effect of the normalization of the messages.

10: We show in particular that, for a large class of normalization

11: strategies, it is possible to focus only on belief convergence.

12: Following this, we express the necessary and sufficient conditions for

13: local stability of a fixed point in terms of the graph structure and

14: the beliefs values at the fixed point. We also explicit some connexion

15: between the normalization constants and the underlying Bethe Free

16: Energy.

17: \end{abstract}

18: