Belief propagation for general graphical models with loops

Abstract: Belief Propagation (BP) decoders for quantum error correcting codes are not always precise. There is a growing interest in the application of tensor networks to quantum error correction in general and, in particular, in degenerate quantum maximum likelihood decoding and the tensor network decoder. We develop a unified view to make the generalized BP proposal by Kirkley et. al explicit on arbitrary graphical models. We derive BP schemes and provide inference equations for BP on loopy tensor networks and, more generally, loopy graphical models. In doing so we introduce a tree-equivalent approach which allows us to relate the tensor network BlockBP to a generalized BP for loopy networks. Moreover, we show that the tensor network message passing approach relies essentially on the same approximation as the method by Kirkley. This allows us to make tensor network message passing available for degenerate quantum maximum likelihood decoding. Our method and results are key to obtaining guidelines regarding how the exchange between complexity and decoding accuracy works between BP and tensor network decoders. Finally, we discuss how the tree-equivalent method and the method by Kirkley can justify why message scheduling improves the performance of BP.