A bias removal technique for the prediction-based blind adaptive multichannel deconvolution

The problem of identifying/equalizing a digital communication channel based on its temporally or spatially oversampled output has recently gained much attention (multichannel deconvolution). In particular, blind identification methods were proposed relying on the linear prediction of the received signals, making these methods well suited to an adaptive implementation. However, in practical situations with noise corrupted data, the estimated prediction coefficients are biased, causing serious impairment in the channel estimation. In this contribution we propose a low cost algorithm for the adaptive computation of the unbiased prediction coefficients, that does not require the the knowledge of the noise variance. The technique is based on the minimization of a constrained prediction criterion, which moreover provides an estimate of the noise level. We concentrate on the blind multichannel deconvolution context, but this bias removal technique may also be used in other kinds of linear prediction-based problems.