For low-density parity-check (LDPC) codes widely used in NAND flash memories, the bit-error rate performance is closely tied to the number of bits per message used by the message-passing decoder. This paper describes a technique to generate message-passing decoding mapping functions for LDPC codes using 3 and 4 bits per message. These maps are not derived from belief-propagation decoding or one of its approximations, instead, the maps are based on a channel quantizer that maximizes mutual information. More precisely, the construction technique is a systematic method, which uses an optimal quantizer at each step of density evolution to generate message-passing decoding mappings. Numerical results show, for high-rate codes suitable for flash memories, that 4 bits per message and a few iterations (10-20 iterations) are sufficient to approach full belief-propagation decoding, less than 5-7 bits per message typically needed. The construction technique is flexible, since it can generate maps for arbitrary number of bits per message, and can be applied to arbitrary memoryless channels.

A recent result has shown connections between statistical learning theory and channel quantization. In this paper, we present a practical application of this result to the implementation of LDPC decoders. In particular, we describe a technique for designing the message-passing decoder mappings (or lookup tables) based on the ideas of channel quantization. This technique is not derived from sum-product algorithm or any other LDPC decoding algorithm. Instead, the proposed algorithm is based on an optimal quantizer in the sense of maximization of mutual information, which is inserted in the density evolution algorithm to generate the lookup tables. This algorithm has low complexity since it only employs 3-bit messages and lookup tables, which can be easily implemented in hardware. Two quantized versions of the min-sum decoding algorithm are used for comparison. Simulation results for a binary-input AWGN channel show 0.3 dB and 1.2 dB gains versus the two quantized min-sum algorithms. On the binary symmetric channel also a gain is seen.