Thinning, photonic beamsplitting, and a general discrete entropy power inequality

TL;DR

This paper develops an axiomatic framework for a natural discrete-variable entropy power inequality (EPI), drawing parallels with the continuous case and exploring its implications in quantum channel capacity.

Contribution

It introduces a discrete-variable EPI based on geometric distributions and a discrete scaled addition, linking it to quantum information theory and prior work on ultra-log-concave distributions.

Findings

Proposes a discrete EPI with geometric distribution as maximum entropy.

Defines entropy power as the mean of a geometric random variable.

Shows the discrete EPI holds when entropy power is redefined as e^{H(X)}.

Abstract

Many partially-successful attempts have been made to find the most natural discrete-variable version of Shannon's entropy power inequality (EPI). We develop an axiomatic framework from which we deduce the natural form of a discrete-variable EPI and an associated entropic monotonicity in a discrete-variable central limit theorem. In this discrete EPI, the geometric distribution, which has the maximum entropy among all discrete distributions with a given mean, assumes a role analogous to the Gaussian distribution in Shannon's EPI. The entropy power of $X$ is defined as the mean of a geometric random variable with entropy $H(X)$. The crux of our construction is a discrete-variable version of Lieb's scaled addition $X \boxplus_\eta Y$ of two discrete random variables $X$ and $Y$ with $\eta \in (0, 1)$. We discuss the relationship of our discrete EPI with recent work of Yu and Johnson who developed an EPI for a restricted class of random variables that have ultra-log-concave (ULC) distributions. Even though we leave open the proof of the aforesaid natural form of the discrete EPI, we show that this discrete EPI holds true for variables with arbitrary discrete distributions when the entropy power is redefined as $e^{H(X)}$ in analogy with the continuous version. Finally, we show that our conjectured discrete EPI is a special case of the yet-unproven Entropy Photon-number Inequality (EPnI), which assumes a role analogous to Shannon's EPI in capacity proofs for Gaussian bosonic (quantum) channels.