Graph-based Encoders and their Achievable Rates for Channels with Feedback

This paper investigates graph-based encoders for the unifilar finite-state channel (FSC) with feedback. A recent paper introduced the Q-graph as a tool for the recursive quantization of channel outputs on a directed graph. The Q- graph approach yielded single-letter lower and upper bounds on the feedback capacity of unifilar FSCs, termed here Q-LB and Q-UB, respectively. The current paper provides two computable optimization problems for the Q-LB and the Q-UB. The first, for the Q-LB, aims to find the graph-based encoder with the highest achievable rate. Specifically, for a structured cooperation between the encoder and the decoder, that is given by a particular Q-graph, the optimization problem maximizes the Q-LB over all input distributions. The resultant graph-based encoder from the optimization problem has a corresponding posterior matching scheme that achieves the Q-LB. The second optimization problem provides a formulation of the Q-UB as a convex optimization problem. Numerical results of the Q-LB and the Q-UB are presented for the Ising channel and a simplified version of a fading channel. The numerical results are then translated into analytical expressions for graph-based encoders and their achievable rates.