Abstract: Aimed at the loopy belief propagation's unknown convergence and too many iterations on Bayesian networks with cycles, an algorithm named recursive belief propagation (RBP) is proposed. Different from the classical belief propagation, a special order (well-order) is followed by RBP. Passing messages according to this order is equivalent to breaking the cycles sequentially, so the messages cannot circulate and the fixed point is arrived at in one trial. Furthermore, this fixed point is the same as the result of the loopy belief propagation or the minimal point of Bethe free energy. At last, we prove the good performance of the proposed algorithm by experiments.