Decoding of expander codes at rates close to capacity

The decoding error probability of codes is studied as a function of their block length. It is shown that the existence of codes with a polynomially small decoding error probability implies the existence of codes with an exponentially small decoding error probability. Specifically, it is assumed that there exists a family of codes of length N and rate R = (1 - epsilon)C (C is a capacity of a binary-symmetric channel), whose decoding probability decreases inverse polynomially in N. It is shown that if the decoding probability decreases sufficiently fast, but still only inverse polynomially fast in N, then there exists another such family of codes whose decoding error probability decreases exponentially fast in N. Moreover, if the decoding time complexity of the assumed family of codes is polynomial in N and 1/epsilon, then the decoding time complexity of the presented family is linear in N and polynomial in 1/epsilon. These codes are compared to the recently presented codes of Barg and Zemor, "Error Exponents of Expander Codes," IEEE Transactions on Information Theory, 2002, and "Concatenated Codes: Serial and Parallel," 11 IEEE Transactions on Information Theory, 2005. It is shown that the latter families cannot be tuned to have exponentially decaying (in N) error probability, and at the same time to have decoding time complexity linear in N and polynomial in 1/epsilon.