Variable-Length Resolvability for Mixed Sources and its Application to Variable-Length Source Coding

Abstract

In the problem of variable-length $\delta$ -channel resolvability, the channel output is approximated by encoding a variable-length uniform random number under the constraint that the variational distance between the target and approximated distributions should be within a given constant $\delta$ asymptotically. In this paper, we assume that the given channel input is a mixed source whose components may be general sources. To analyze the minimum achievable length rate of the uniform random number, called the $\delta$ -resolvability, we introduce a variant problem of the variable-length $\delta$ -channel resolvability. A general formula for the $\delta$ -resolvability in this variant problem is established for a general channel. When the channel is an identity mapping, it is shown that the $\delta$ -resolvability in the original and variant problems coincide. This relation leads to a direct derivation of a single-letter formula for the $\delta$ -resolvability when the given source is a mixed memoryless source. We extend the result to the second-order case. As a byproduct, we obtain the first-order and second-order formulas for fixed-to-variable length source coding allowing error probability up to $\delta$ .