Conservation laws for coding

This work deals with coding systems based on sparse graph codes. The key issue we address is the relationship between iterative (in particular belief propagation) and maximum a posteriori decoding. We show that between the two there is a fundamental connection, which is reminiscent of the Maxwell construction in thermodynamics. The main objects we consider are EXIT-like functions. EXIT functions were originally introduced as handy tools for the design of iterative coding systems. It gradually became clear that EXIT functions possess several fundamental properties. Many of these properties, however, apply only to the erasure case. This motivates us to introduce GEXIT functions that coincide with EXIT functions over the erasure channel. In many aspects, GEXIT functions over general memoryless output-symmetric channels play the same role as EXIT functions do over the erasure channel. In particular, GEXIT functions are characterized by the general area theorem. As a first consequence, we demonstrate that in order for the rate of an ensemble of codes to approach the capacity under belief propagation decoding, the GEXIT functions of the component codes have to be matched perfectly. This statement was previously known as the matching condition for the erasure case. We then use these GEXIT functions to show that in the limit of large blocklengths a fundamental connection appears between belief propagation and maximum a posteriori decoding. A decoding algorithm, which we call Maxwell decoder, provides an operational interpretation of this relationship for the erasure case. Both the algorithm and the analysis of the decoder are the translation of the Maxwell construction from statistical mechanics to the context of probabilistic decoding. We take the first steps to extend this construction to general memoryless output-symmetric channels. More exactly, a general upper bound on the maximum a posteriori threshold for sparse graph codes is given. It is conjectured that the fundamental connection between belief propagation and maximum a posteriori decoding carries over to the general case.