Papers for

statisticians and data analysts

Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.

Entropy concavity conjecture disproved with asymmetric log-concave example

Entropy concavity for log-concave random variables: an asymmetric counterexample

Abstract: The Ball-Nayar-Tkocz entropy concavity conjecture asserts that, if $X,Y$ are independent identically distributed real random variables with a common log-concave density, then the differential entropy of their weighted sum, \[ F(t)=h\bigl(\sqrt{1-t}\,X+\sqrt t\,Y\bigr),\qquad 0\le t\le1, \] is a concave function of the weight parameter $t$. We construct an asymmetric, strictly positive smooth probability density $f$ with mean zero, variance one, and $(\log f)''<-1/2$, for which the corresponding function satisfies $F''(t)>0$ throughout an endpoint neighborhood $0<t<δ$. This disproves the conjecture without an additional symmetry assumption. For a class of Gaussian perturbation densities, we first establish the endpoint expansion \[ F''(t)=-\frac{μ_3J_3}{16\sqrt t}+O(1),\qquad t\downarrow0. \] Here $μ_3=\int x^3f(x)\,dx$ is the third central moment. Writing $ρ=(\log f)'$ for the score function, its third moment is $J_3=\int f(x)ρ(x)^3\,dx$. A Hermite perturbation gives $μ_3>0$ and $J_3<0$, and explicit remainder estimates verify the constructed density and its endpoint curvature. The counterexample does not address the conjecture with an additional symmetry assumption.

Thu 10 SeptInformation Theory
The gist
This paper disproves a mathematical conjecture about how randomness behaves when you mix two similar random values. The original idea suggested that a certain measure of randomness (called entropy) changes in a predictable, curved way when you combine these values. The authors show that if the values are not symmetrical, this predictable behavior can fail. They construct a specific example with an uneven shape where the measure does not curve as expected, challenging previous assumptions.
Open 2609.11418v1