CCF- Unified View of Multiterminal Source Coding

University Of Illinois At Urbana-Champaign, Urbana IL

Investigators

Abstract

In a variety of applications, such as weather monitoring, smart buildings, and manufacturing systems, a network of sensors, each with a communication constraint, is remotely deployed with the goal of making collective inference. However, the governing principle, as well as the performance limit, of such sensor (information source) networks is not well understood. Consequently, most practical systems are designed conservatively, allowing significant room for future improvement and cost reduction. In this backdrop, the investigators identify certain fundamental principles underlying network (multiterminal) source coding, and study performance bounds in the spirit of Shannon. This research entails several benefits: (1) attractive economic returns on research investment due to design benchmarking; (2) unified conceptual understanding of a large class of network coding problems; (3) development of academic materials and courses for training engineers in design and maintenance of distributed systems. The investigators give a canonical theory characterizing the asymptotic performance bound for a broad class of multiterminal source coding problems. In particular, they employ a unifying distortion-independent coding strategy, which manifests distortion only upon suitable post-processing. Importantly, the ubiquitous notion of single-letter description is connected to a certain graphical representation; existence of such description is guaranteed if the corresponding graph admits no cycles. As a consequence, the investigators find a single-letter solution to the hitherto-open single-helper problem. On the contrary, graphs with cycles, seen, for example, in the famous Berger-Tung problem, likely admit no single-letter solution. Accordingly, the Berger-Tung bound is conjectured to be loose, and work towards a conclusive answer is in progress. However, in absence of a single-letter description, the canonical theory still provides a general road map towards a computable description, and tractable computational algorithms.