Max-log-MAP decoding with reduced memory complexity

Given an M-state (recursive) convolutional encoder and information sequence of length n, the space complexity of unoptimized Bahl-Cocke-Jelinek-Raviv (BCJR) decoder is considered to be O(nm). However, if BCJR's forward alpha coefficients are continuously recomputed instead of stored in memory, it can be shown that the space complexity will drop to O(m). In this paper we start from these observations and present a technique for memory reduction in the Max-Log-MAP algorithm. We test our design on a rate-1/2 1025-bit-long Turbo Code and show considerable memory saving.