Windowed Erasure Codes

Abstract

The design of erasure correcting codes and their decoding algorithms is now at the point where capacity achieving codes are available with decoding algorithms that have complexity that is linear in the number of information symbols. One aspect of these codes is that the overhead (number of coded symbols beyond the number of information symbols required to achieve decoding completion with high probability) is linear in k. This work considers a new class of random codes which have the following advantages: (i) the overhead is constant (in the range of 5 to 10) (ii) the probability of completing decoding for such an overhead is essentially one (iii) the codes are effective for a number of information symbols as low as a few tens. The price for these properties is that the decoding complexity is greater, on the order of k 3/2. However, for the lower values of k where these codes are of particular interest, this increase in complexity might be outweighed by other significant advantages. The parity check matrices of these codes are chosen at random as windowed matrices i.e. for each column an initial starting position of a window of length w is chosen and the succeeding w positions are chosen at random by zero or one. It can be shown that it is necessary that w = O(k1/2) for the probabilistic matrix rank properties to behave as a non-windowed random matrix. The sufficiency of the condition has so far been established by extensive simulation, although other arguments strongly support this conclusion