Finite-Length Scaling of Turbo-Like Code Ensembles on the Binary Erasure Channel

A possibility of estimating the finite-length performance of sparse-graph code ensembles gives two opportunities: to compare different codes of the same length in a context very close to real, practical applications and to perform the parameter optimization for a given code length [2]. We need a finite-length approximation that is valid for any code ensemble. The scaling approach seems to be a tool, general enough to provide such an approximation. However, the analytical derivation of parameters of the scaling approximation has been successful only for LDPC codes [1]; despite several attempts [25], [20], no such result was proposed for other code ensembles. In this paper, we focus on the finite-length performance of turbo-like codes, by applying the scaling approach to this case. In particular, by assuming the transmission over the binary erasure channel, we conjecture the scaling law and derive its scaling parameter. As examples, we present the performance estimation for Repeat-Accumulate codes [11], parallel turbo codes [8] and TLDPC codes [5], in all cases matching well the numerical results.