Files

Abstract

We consider multiple description coding for the Gaussian source with K descriptions under the symmetric mean squared error distortion constraints. One of the main contributions is a novel lower bound on the sum rate, derived by generalizing Ozarow’s well-known technique in treating the Gaussian two-description problem, with the new ingredient of expanding the probability space by more than one auxiliary random variables. Two achievability schemes are investigated in details, the first of which is based on successive refinement coding coupled with multilevel diversity coding (SR-MLD), and the second is a generalization of the multi-layer coding scheme proposed by Puri et al.. Moreover, to make the inner and outer bounds compatible, method similar to the enhancement technique in establishing the Gaussian multiple-input multiple-output broadcast channel capacity is used. Comparison between the lower and upper bounds for the Gaussian symmetric-rate individual-description R-D function shows that they are less than a constant away from each other. The bounding constant depends only on the number of descriptions, but not the distortion constraints. Moreover, regardless of the number of descriptions, this gap between the lower bound and the upper bound using the SR-MLD coding scheme is less than 1.5 bits, and for the other case, the gap is less than 1 bit. The results are further extended to general sources, where a similar but looser bound on the gap holds. Generalizing the proof technique for lower bounding the sum rate provides an outer bound for the R-D region. Since the two achievable regions, which have a geometry structure exactly the same as the rate region of the symmetric lossless multilevel diversity coding, are specially tailored for easy comparison with the outer bound, we are able to show the following results. For the case K = 3, the inner and outer bounds can both be represented by ten planes with matching normal directions, and the pairwise difference between them is small. This result is further generalized to K-description case by incorporating the α-resolution method, which yields upper and lower bounds on any bounding plane of the R-D region, and thus provides an approximate characterization of the R-D region.

Details

Actions