Capacity Regions of Two-Receiver Broadcast Erasure Channels with Feedback and Memory

TL;DR

This paper characterizes the capacity regions of two-receiver broadcast erasure channels with feedback and memory, proposing optimal coding schemes and analyzing their performance under different channel state knowledge scenarios.

Contribution

It provides the first complete capacity region characterization for channels with memory and feedback, including new coding schemes and bounds for both visible and hidden states.

Findings

Capacity region with visible state characterized by a linear program.

Two optimal coding schemes using feedback are proposed and analyzed.

Approximate capacity regions for hidden state channels are provided.

Abstract

The two-receiver broadcast packet erasure channel with feedback and memory is studied. Memory is modeled using a finite-state Markov chain representing a channel state. Two scenarios are considered: (i) when the transmitter has causal knowledge of the channel state (i.e., the state is visible), and (ii) when the channel state is unknown at the transmitter, but observations of it are available at the transmitter through feedback (i.e., the state is hidden). In both scenarios, matching outer and inner bounds on the rates of communication are derived and the capacity region is determined. It is shown that similar results carry over to channels with memory and delayed feedback and memoryless compound channels with feedback. When the state is visible, the capacity region has a single-letter characterization and is in terms of a linear program. Two optimal coding schemes are devised that use feedback to keep track of the sent/received packets via a network of queues: a probabilistic scheme and a deterministic backpressure-like algorithm. The former bases its decisions solely on the past channel state information and the latter follows a max-weight queue-based policy. The performance of the algorithms are analyzed using the frameworks of rate stability in networks of queues, max-flow min-cut duality in networks, and finite-horizon Lyapunov drift analysis. When the state is hidden, the capacity region does not have a single-letter characterization and is, in this sense, uncomputable. Approximations of the capacity region are provided and two optimal coding algorithms are outlined. The first algorithm is a probabilistic coding scheme that bases its decisions on the past L acknowledgments and its achievable rate region approaches the capacity region exponentially fast in L. The second algorithm is a backpressure-like algorithm that performs optimally in the long run.