Capacity, source/channel coding and separation: A fresh perspective
In this talk we revisit tradition notions of capacity, end-to-end distortion, and source/channel code separation optimality for general channels that are not necessarily stationary, ergodic, or information-stable. We introduce new definitions of capacity and end-to-end distortion for such channels that incorporate notions of outage and expected performance, and show how these new definitions impact separation optimality and the source/channel code interface. We then present new source and channel code designs based on these ideas, and demonstrate their performance gains relative to existing techniques. Finally, we discuss broadening the notion of separation optimality beyond Shannon's view that a single rate serves as the source/channel code interface.