Variable Min-Sum decoding based on generalized mutual information metric

Min Sum algorithm simplifies the non-linear check node operation of Belief Propagation algorithm via linear approximation, which greatly reduces the complexity of realization of decoder but degrades the performance as well. The resulting sub-optimality could be tempered via scaling of LLRs, e.g. fixed optimal scaling applied to Min Sum output resulting in the Normalized Min Sum algorithm, and variable scaling schemes gradually appear in literature. In this paper, we study the variable scaling decoding algorithm, and propose to generate variable scaling sequences via generalized mutual information (GMI) metric. Simulation results on real LDPC codes for different decoding algorithms have shown that our GMI metric performs better than the variable scaling scheme appearing in literature, and meanwhile improves substantially in terms of BER over the conventional Normalized Min Sum algorithm.