On performance analysis and implementation issues of iterative decoding for graph based codes

There is no doubt that long random-like code has the potential to achieve good performance because of its excellent distance spectrum. However, these codes remain useless in practical applications due to the lack of decoders rendering good performance at an acceptable complexity. The invention of turbo code marks a milestone progress in channel coding theory in that it achieves near Shannon limit performance by using an elegant iterative decoding algorithm. This great success stimulated intensive research oil long compound codes sharing the same decoding mechanism. Among these long codes are low-density parity-check (LDPC) code and product code, which render brilliant performance. In this work, iterative decoding algorithms for LDPC code and product code are studied in the context of belief propagation. A large part of this work concerns LDPC code. First the concept of iterative decoding capacity is established in the context of density evolution. Two simulation-based methods approximating decoding capacity are applied to LDPC code. Their effectiveness is evaluated. A suboptimal iterative decoder, Max-Log-MAP algorithm, is also investigated. It has been intensively studied in turbo code but seems to be neglected in LDPC code. The specific density evolution procedure for Max-Log-MAP decoding is developed. The performance of LDPC code with infinite block length is well-predicted using density evolution procedure. Two implementation issues on iterative decoding of LDPC code are studied. One is the design of a quantized decoder. The other is the influence of mismatched signal-to-noise ratio (SNR) level on decoding performance. The theoretical capacities of the quantized LDPC decoder, under Log-MAP and Max-Log-MAP algorithms, are derived through discretized density evolution. It is indicated that the key point in designing a quantized decoder is to pick a proper dynamic range. Quantization loss in terms of bit error rate (BER) performance could be kept remarkably low, provided that the dynamic range is chosen wisely. The decoding capacity under fixed SNR offset is obtained. The robustness of LDPC code with practical length is evaluated through simulations. It is found that the amount of SNR offset that can be tolerated depends on the code length. The remaining part of this dissertation deals with iterative decoding of product code. Two issues on iterative decoding of\u27 product code are investigated. One is, \u27improving BER performance by mitigating cycle effects. The other is, parallel decoding structure, which is conceptually better than serial decoding and yields lower decoding latency