Asymmetric Gaussian multiple descriptions and asymmetric multilevel diversity coding

We consider asymmetric multiple description (MD) source coding for Gaussian source under mean squared error distortion constraints, and focus on the three description problem. Inner and outer bounds for the rate region are derived, both of which can be represented as the intersection of ten half spaces with matching normal directions. Moreover, the gap between the inner and outer bounds is shown to be small. The inner bound relies on the rate region characterization of a lossless asymmetric multilevel diversity (MLD) coding problem treated in our earlier work, which is a natural generalization of the symmetric MLD coding problem previously considered by Roche et al. Different from symmetric MLD coding, superposition coding is not sufficient in the asymmetric case, and ideas akin to network coding need to be used strategically. Equipped with this finding, and motivated by the connection between symmetric MD and symmetric MLD coding, in this work we consider asymmetric MD as a lossy version of the asymmetric MLD coding, which requires coding beyond simple superposition. An outer bound is also derived, which bears a geometric structure particularly suitable for comparison with the inner bound. Combining the inner and outer bounds provides an approximate characterization of the rate region for the asymmetric Gaussian three description problem.