About the asymptotic performance of multiple-input/multiple-output linear prediction of subband signals

In the present letter, we investigate the asymptotic performance of causal multiple-input/multiple-output (MIMO) linear prediction of subband signals obtained by means of an orthogonal filterbank. The prediction filters are designed for a minimum mean squared error criterion. It is demonstrated that the geometrical mean of the prediction errors in the subbands is asymptotically equal to the prediction error in the fullband, which is in turn related to the spectral flatness measure of the input process. As a consequence, the maximum performance of a hybrid subband coding system with MIMO prediction is independent of the filter bank used for the subband decomposition.