On robustness of linear prediction based blind identification
Luc Deneire · 1997
Linear prediction based algorithms have been applied to the multi-channel FIR identification problem. In [11], it was shown that oversampled and/or multiple antenna received signals may be modeled as well as low rank MA processes as low rank AR processes. Indeed, taking FIR nature and the singularity of the MA process into account (due to the fact that the number of channels is bigger than the number of sources) leads to a finite order prediction filter (i.e. AR(L ! 1) modeling), which is automatically identified by, e.g., a singular multichannel Levinson algorithm, and can be shown to be robust to AR order overestimation. On the other hand, K.A. Meraim and A. Gorokhov derive other robustness properties based on the equations P(z)H(z) = h(0), where P(z) is the prediction filter H(z) is the channel and h(0) its first coefficient. Although using P(z) of overestimated order, clever use of the previous equations leads to robustness of the estimation of H(z) to channel length overestimation...