Filling-in some of the decoding gap between belief propagation (BP) and maximum a posteriori (MAP) for convolutional codes

This paper describes an iterative method of decoding (for convolutional codes) that obtains soft-estimate decoding performances in-between BP decoding and optimal MAP decoding. The idea is to replace an original even-parity function (which is used in BP decoding) with a "reduced" core function and a corresponding multistate trellis. The result is a soft-decoding performance that can be better than the performance for recursive BP decoding. The decoding method does not achieve the optimal MAP performance due to dependencies within the recursive algorithm. The benefit is that the trellis may have far fewer states than the full 2^m-state trellis which is required for MAP decoding