Sharp Analytical Capacity Upper Bounds for Sticky and Related Channels

TL;DR

This paper derives sharp analytical upper bounds on the capacity of binary channels with synchronization errors, such as duplication and geometric channels, advancing understanding beyond previous numerical methods.

Contribution

It provides the first analytical proofs for capacity bounds of these channels, improving upon prior numerical approximations and offering insights into their capacity behavior.

Findings

Upper bounds closely match previous numerical results.

Capacity remains bounded away from 1 for geometric repetition channels with increasing mean.

Analytical methods outperform numerical approximations in understanding channel capacity.

Abstract

We study natural examples of binary channels with synchronization errors. These include the duplication channel, which independently outputs a given bit once or twice, and geometric channels that repeat a given bit according to a geometric rule, with or without the possibility of bit deletion. We apply the general framework of Cheraghchi (STOC 2018) to obtain sharp analytical upper bounds on the capacity of these channels. Previously, upper bounds were known via numerical computations involving the computation of finite approximations of the channels by a computer and then using the obtained numerical results to upper bound the actual capacity. While leading to sharp numerical results, further progress on the full understanding of the channel capacity inherently remains elusive using such methods. Our results can be regarded as a major step towards a complete understanding of the capacity curves. Quantitatively, our upper bounds sharply approach, and in some cases surpass, the bounds that were previously only known by purely numerical methods. Among our results, we notably give a completely analytical proof that, when the number of repetitions per bit is geometric (supported on $\{0,1,2,\dots\}$) with mean growing to infinity, the channel capacity remains substantially bounded away from $1$.