Predictive coding of correlated sources

A lossy coding scheme is proposed for separate encoding and joint decoding of two correlated sequences. The algorithm simultaneously exploits the correlation between the sequences (using a binning-based quantization scheme) and that between the samples of each sequence (using linear prediction). Under the proposed coding regime, optimal prediction filter design fundamentally deviates from the traditional approach. More specifically, it is, in general, not optimal to employ first-order prediction for noisy observations of a first-order Markov source, even when the noise is negligibly small. Moreover, even if the prediction filter is constrained to be of degree 1, the optimal filter coefficient is different from the correlation coefficient of the Markov source. In the particular example treated in this paper, it is shown that optimal first- and second-order prediction respectively enjoy up to 0.9 dB and 1.15 dB improvement over the traditional approach.