OctoInformBeta

OpenAI preprint, September 2026 · §2. The scalar rate and its curvature

Lemma 2.4: An upper bound in the open entropy interval and a separate value at zero

Entropy weights control scalar rate curvature.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

Binary information on the discrete cube gives a scalar rate parametrized by the entropy deficit of a binary bias. A bias a∈[0,1]a\in[0,1] describes a sign with probabilities (1+a)/2(1+a)/2 and (1−a)/2(1-a)/2.

The binary entropy deficit ψ\psi is the function of the bias in (1), with natural logarithms and the usual entropy convention 0log⁡0=00\log 0=0: L=log⁡2,ψ(a)=1+a2log⁡(1+a)+1−a2log⁡(1−a).L=\log 2,\qquad \psi(a)=\frac{1+a}{2}\log(1+a)+\frac{1-a}{2}\log(1-a).(1) The inverse entropy deficit ψ−1\psi^{-1} is the inverse of ψ\psi restricted to [0,1][0,1]: for s∈[0,L]s\in[0,L], its value is the unique bias in that interval with entropy deficit ss.✦

The scalar rate FF is the function on [0,L)[0,L) defined parametrically by (2): F(ψ(a))=aatanh⁡(a),0≤a<1.F(\psi(a))=a\operatorname{atanh}(a),\qquad 0\le a<1.(2) The profile curvature F″F'' is a real function of the entropy-deficit argument ss, equal to 4/34/3 for s≤0s\le0 and to the expression in (3) for s>0s>0, where as=ψ−1(s)a_s=\psi^{-1}(s): F″(s)=(1+as2)atanh⁡(as)−as(1−as2)2(atanh⁡(as))3,s>0.F''(s)= \frac{(1+a_s^2)\operatorname{atanh}(a_s)-a_s} {(1-a_s^2)^2\bigl(\operatorname{atanh}(a_s)\bigr)^3}, \qquad s>0.(3)

The complementary entropy coordinate hh associated with a deficit ss is the real number specified by h=L−sh=L-s. The correction curvature CC is the function of the real entropy coordinate hh defined by (4): C(h)={F″(L−h),h<L,0,h≥L.C(h)= \begin{cases} F''(L-h), & h<L,\\ 0, & h\ge L. \end{cases}(4)✦

The Claim

Theorem 1. For an entropy coordinate hh with 0<h<L0<h<L, h2F″(L−h)≤23.h^2 F''(L-h)\le\frac23.(5)✦ For a deficit ss with 0<s<L0<s<L, (L−s)2F″(s)≤23.(L-s)^2 F''(s)\le\frac23.(6)✦ The endpoint value is F″(0)=43.F''(0)=\frac43.(7)✦ For every positive entropy coordinate hh, h2C(h)≤23.h^2 C(h)\le\frac23.(8)✦

In the paper, the result reads: for 0<h≤L0<h\le L, put P(h)=h2F″(L−h)P(h)=h^2F''(L-h); then 12≤P(h)≤M=4L2/3\frac12\le P(h)\le M=4L^2/3, with P(L)=MP(L)=M.

What Is Checked

The statements (5), (6) and (7) concern the profile curvature defined from the inverse binary entropy deficit on the open physical interval and at zero, while (8) concerns the piecewise correction curvature on the positive entropy axis.