The performance of DPCM operating on lossy channels with memory

We study the performance of DPCM operating over lossy channels in cell-based networks such as ATM, when statistical multiplexing is used. Since the loss of data is mainly due to traffic congestion, and the minimum unit of loss is a cell which may contain multiple DPCM codewords, the loss process can be highly correlated. The contribution of this work comes in two parts. We first give a formulation which relates the performance of DPCM, in terms of signal-to-reconstruction-noise ratio, to the channel loss conditions. In particular we take into account any correlation that exists in the loss process. We then study in detail mechanisms which give rise to such correlation. Using simple source and multiplexing models, formulae are given which characterise the loss process as a result of statistical multiplexing. As a direct result of this treatment, we know that the maximum achievable performance of DPCM with reoptimised predictors further deteriorates when the loss process is correlated.<>