A dynamic program for linear sequential coding for Gaussian MAC with noisy feedback
share
In this paper consider a two user multiple access channel with noisy feedback. There are two senders with independent messages who transmit symbols across an additive white Gaussian channel to a receiver, who in turn sends back a symbol which is received by the two senders through two independent noisy Gaussian channels. We consider the case when the feedback is active i.e. the receiver actively encodes the feedback using a linear state process. We pose this as a problem of linear sequential coding at the senders and the receiver to minimize the terminal mean square probability of error at the receiver. This is an instance of decentralized control with no common information at the senders and the receiver. In this paper, we construct two linear controllers at the sender and the receiver. Due to linearity of the policies and the controllers, all the random variables involved are jointly Gaussian. Moreover, the corresponding covariance matrix at the receiver of the estimation process of the senders' messages is a deterministic process, which is a function of the parameters of the controllers and the strategies of the players, and is thus perfectly observed by the senders. Based on this observation, we use deterministic dynamic programming to find the optimal policies and the optimal linear controllers at both the senders and the receiver. The problem with passive feedback can be considered as a special case.
READ FULL TEXT