Philip A. Chou
Shannon's source/channel coding theorem states that a source can be communicated with distortion D(R) over a channel with capacity C whenever R is less than C (where both are expressed in information bits per second). Furthermore, this can be achieved using a block code separated into a source code producing R1>R information bits per second and a block channel code consuming R1<C information bits per second. Unfortunately, in order to achieve this for R arbitrarily close to C, the length of the block code, and hence its complexity and delay, must be arbitrarily large. Thus the theorem is not strictly applicable when the complexity or delay is bounded as is the case for any real-time communication system. Of course, separation of source code and channel coding has been a viable technique for many years. Nevertheless we expect gains in performance, sometimes quite large, when the assumptions of the theorem are violated, e.g., when the souce or channel is time-varying with a time constant that is large relative to the delay bound; when the source or channel is unknown; when there are multiple receivers with different channels; when there are multiple channels (e.g., with different QoS) to each receiver; when feedback is available; when partial communication between multiple receivers is available; when available power must be shared between compression and transmission; and so forth, all of which are realities in today's networked, mobile world. These are the conditions under which joint source/channel coding can achieve the greatest gains.