Decentralized encoding of a remote source

For decentralized estimation of a remote source under communication constraints, noisy observations taken by spatially separated sensors are encoded before transmission to a central decision maker. As the observations are not simultaneously available at a given site, the peripheral encoders are to operate disjointly. This problem cannot be cast: as the classical one of encoding a remote source. Given that the peripheral encoders are scalar quantizers, we consider several criteria for partitioning the space of observations and for reproducing the source with minimum distortion. They differ in regard to the knowledge about the source-observations model, for example about the spatial correlation among the observations, and in regard to the complexity of the decoding operations, but all lead to an unifying design approach of cyclical optimization. For a linear Gaussian source-observations model, adopting the quadratic distortion measure, we carry out each design and evaluate the performance of the various schemes in terms of achieved distortion-rate pairs, allowing for some combined parametric variations of the model. We also derive the optimum theoretically achievable performance for the distributed and for the conventional (non-distributed) scheme in terms of distortion-rate bounds. The results demonstrate that distributed encoding-decoding schemes, if properly designed, compare favorably to the non-distributed schemes.