Statistical Physics Methods for Sparse Graph Codes

This thesis deals with the asymptotic analysis of coding systems based on sparse graph codes. The goal of this work is to analyze the decoder performance when transmitting over a general binary-input memoryless symmetricoutput (BMS) channel. We consider the two most fundamental decoders, the optimal maximum a posteriori (MAP) decoder and the sub-optimal belief propagation (BP) decoder. The BP decoder has low-complexity and its performance analysis is, hence, of great interest. The MAP decoder, on the other hand, is computationally expensive. However, the MAP decoder analysis provides fundamental limits on the code performance. As a result, the MAPdecoding analysis is important in designing codes which achieve the ultimate Shannon limit. It would be fair to say that, over the binary erasure channel (BEC), the performance of the MAP and BP decoder has been thoroughly understood. However, much less is known in the case of transmission over general BMS channels. The combinatorial methods used for analyzing the case of BEC do not extend easily to the general case. The main goal of this thesis is to advance the analysis in the case of transmission over general BMS channels. To do this, we use the recent convergence of statistical physics and coding theory. Sparse graph codes can be mapped into appropriate statistical physics spin-glass models. This allows us to use sophisticated methods from rigorous statistical mechanics like the correlation inequalities, interpolation method and cluster expansions for the purpose of our analysis. One of the main results of this thesis is that in some regimes of noise, the BP decoder is optimal for a typical code in an ensemble of codes. This result is a pleasing extension of the same result for the case of BEC. An important consequence of our results is that the heuristic predictions of the replica and cavity methods of spin-glass theory are correct in the realm of sparse graph codes.