The optimal causal (prediction based) decorrelating scheme is applied to the frameworks
of transform coding, coding of vectorial signals (multichannel audio),
and vectorial DPCM coding. We analyze the effects
of backward adaptation uppon the prediction operations and compare the expressions
of the coding gains under infinite and high resolution assumption.
We generalize the MIMO (Multiple Input/Multiple Output) prediction by
organizing differently the samples in the vectorial signals, which corresponds
to different degrees of non-causality of the intersignals predictors. An extreme
case is the triangular MIMO prediction, for which "causality" becomes
processing the channel in a certain order. The high resolution coding
gain suggests an optimal strategy in the choice of the interband predictors.
For two-dimensionnal vectorial sources (such as stereo signals) we show the superiority
in terms of coding gain of the triangular MIMO predictor over the classical MIMO
prediction. A theorem is established which concerns the optimal ordering
of the signals for the triangular MIMO predictor. When finite prediction orders
are used to perform the intersignal decorrelation, we show that the optimal
positioning of a finite number of taps is fairly straightforward.
