Sending a Bivariate Gaussian Source over a Gaussian MAC with Unidirectional Conferencing Encoders
Abstract: A memoryless bivariate Gaussian source is observed separately by two transmitters; Transmitter 1 observes the first source component and Transmitter 2 observes the second source component. The bivariate source is to be transmitted over an average-power constrained Gaussian two-to-one multiple-access channel. In the considered model, prior to each transmission block, Transmitter 1 is allowed to communicate with Transmitter 2 via a unidirectional noise-free bit-pipe of given capacity. Both transmitters then cooperate in describing the source components to the central receiver which wishes to reconstruct each source component subject to a desired expected squared-error distortion. We seek to characterize the pairs of expected squared-error distortions that are simultaneously achievable on the two source components.
The main results of this work are sufficient and necessary conditions for the achievability of a distortion pair expressed as a function of the channel SNR and of the source correlation. In the high-SNR regime, and when the capacity of the conference channel is unlimited, these necessary and sufficient conditions are shown to agree. In addition, we evaluate the precise high-SNR asymptotics of the optimal distortion pairs when the capacity of the conference channel is unlimited.
* The research presented in the seminar was carried out towards the M.Sc. degree, with the supervision of Dr. Shraga Bross