Converse Theorems for the DMC with Mismatched Decoding

TL;DR

This paper establishes converse theorems for mismatched decoding over discrete memoryless channels, showing the equivalence of various decoding strategies and providing upper bounds on mismatch capacity using probabilistic and information-theoretic tools.

Contribution

It proves the equality of mismatch capacity with constant margin decoding and introduces new upper bounds for sequences of codebooks with empirical distributions close to a fixed distribution.

Findings

Mismatch capacity equals the capacity with a constant margin decoder.

Supremum rates with threshold decoding are upper bounded by the mismatch capacity.

Fast decay of error probability implies rate bounds by the mismatch capacity.

Abstract

The problem of mismatched decoding with an additive metric $q$ for a discrete memoryless channel $W$ is addressed. The "product-space" improvement of the random coding lower bound on the mismatch capacity, $C_q^{(\infty)}(W)$, was introduced by Csisz\'ar and Narayan. We study two kinds of decoders. The {\it $\delta$-margin mismatched decoder} outputs a message whose metric with the channel output exceeds that of all the other codewords by at least $\delta$. The {\it $\tau$-threshold decoder} outputs a single message whose metric with the channel output exceeds a threshold $\tau$. Both decoders declare an error if they fail to find a message that meets the requirement. It is assumed that $q$ is bounded. It is proved that $C_q^{(\infty)}(W)$ is equal to the mismatch capacity with a constant margin decoder. We next consider sequences of $P$-constant composition codebooks, whose empirical distribution of the codewords are at least $o(n^{-1/2})$ close in the $L_1$ distance sense to $P$. Using the Central Limit Theorem, it is shown that for such sequences of codebooks the supremum of achievable rates with constant threshold decoding is upper bounded by the supremum of the achievable rates with a constant margin decoder, and therefore also by $C_q^{(\infty)}(W)$. Further, a soft converse is proved stating that if the average probability of error of a sequence of codebooks converges to zero sufficiently fast, the rate of the code sequence is upper bounded by $C_q^{(\infty)}(W)$. In particular, if $q$ is a bounded rational metric, and the average probability of error converges to zero faster than $O(n^{-1})$, then $R\leq C_q^{(\infty)}(W)$. Finally, a max-min multi-letter upper bound on the mismatch capacity that bears some resemblance to $C_q^{(\infty)}(W)$ is presented.