1: \begin{abstract}

2: We derive novel conditions that guarantee convergence of the Sum-Product

3: algorithm (also known as Loopy Belief Propagation or simply Belief Propagation)

4: to a unique fixed point, irrespective of the initial messages. The

5: computational complexity of the conditions is polynomial in the number of

6: variables. In contrast with previously existing conditions, our results are

7: directly applicable to arbitrary factor graphs (with discrete variables) and

8: are shown to be valid also in the case of factors containing zeros, under some

9: additional conditions. We compare our bounds with existing ones, numerically

10: and, if possible, analytically. For binary variables with pairwise

11: interactions, we derive sufficient conditions that take into account local

12: evidence (i.e.\ single variable factors) and the type of pair interactions

13: (attractive or repulsive). It is shown empirically that this bound outperforms

14: existing bounds.

15: \end{abstract}

16: