Slepian-Wolf codes for parallel sources: design and error-exponent analysis∗

The distributed source coding problem, introduced by Slepian and Wolf, is finding application in diverse areas ranging from sensor networks to robust videocompression to compression of encrypted data. Such applications have led to a flurry of recent work in practical code constructions. However, most of these constructions are designed for independent identically distributed random source models, which fail to capture the high degree of correlation inherent in real-world sources like image and video signals. Motivated by its success in the modeling of real-world sources, in this paper we invoke a parallel source model. The traditional way to deal with parallel source models has been through the so-called waterfilling prescription of using multiple shorter codes, each matched to one of the source components. Our main contribution is the proposal of a single long-block-length distributed source code that takes into account the parallel nature of the source. We first present an information-theoretic analysis of the gains made possible by this approach by using the well-developed theory of error exponents. More significantly, our study exposes a new code design problem which we describe for an LDPC framework. We show simulation results to validate our design.