OctoInformBeta

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

§2 (setting): inverse entropy, rate and surplus: The inverse entropy deficit, the rate surplus, and the correction in the entropy variable

Entropy coordinates for the scalar rate.

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

A binary variable can be parametrized by its bias or by its entropy deficit. The scalar rate uses the nonnegative bias, and the inverse entropy deficit expresses that rate in information coordinates.

Write H(v)H(v) for the entropy of a random sign with mean vv, measured in natural logarithms. The constant LL and the binary entropy deficit ψ\psi are given by (1), with 0log⁡0=00\log 0=0: L=log⁡2,ψ(v)=L−H(v)=1+v2log⁡(1+v)+1−v2log⁡(1−v),−1≤v≤1.L=\log 2,\qquad \psi(v)=L-H(v) =\frac{1+v}{2}\log(1+v)+\frac{1-v}{2}\log(1-v), \qquad -1\leq v\leq 1.(1) The endpoint convention for the real-valued formulas here is atanh⁡(1)=0\operatorname{atanh}(1)=0.

The Claim

Definition 1. The inverse entropy deficit ψ−1:[0,L]→[0,1]\psi^{-1}:[0,L]\to[0,1] is the inverse of the restriction of ψ\psi to [0,1][0,1], as in (2): ψ−1=(ψ|[0,1])−1.\psi^{-1}=\bigl(\psi|_{[0,1]}\bigr)^{-1}.(2)✦ The scalar rate F:[0,L]→ℝF:[0,L]\to\mathbb R is the function defined by (3): F(s)=ψ−1(s)atanh⁡(ψ−1(s)).F(s)=\psi^{-1}(s)\operatorname{atanh}\bigl(\psi^{-1}(s)\bigr).(3)✦ The rate surplus G:[0,L]→ℝG:[0,L]\to\mathbb R is the excess of FF over twice the entropy deficit, given by (4): G(s)=F(s)−2s.G(s)=F(s)-2s.(4)✦ The entropy correction LLean:ℝ→ℝL_{\mathrm{Lean}}:\mathbb R\to\mathbb R is the zero extension of the surplus in the entropy variable, defined by (5): LLean(h)={G(L−h),0<h<L,0,otherwise.L_{\mathrm{Lean}}(h)= \begin{cases} G(L-h),&0<h<L,\\ 0,&\text{otherwise}. \end{cases}(5)✦

Theorem 1. For an entropy deficit s∈[0,L]s\in[0,L], the inverse identity is (6): ψ(ψ−1(s))=s.\psi\bigl(\psi^{-1}(s)\bigr)=s.(6)✦ For a nonnegative bias u∈[0,1]u\in[0,1], the reverse identity is (7): ψ−1(ψ(u))=u.\psi^{-1}\bigl(\psi(u)\bigr)=u.(7)✦ The inverse has the continuity property (8): ψ−1 is continuous on [0,L].\psi^{-1}\text{ is continuous on }[0,L].(8)✦ For a nonnegative bias u∈[0,1]u\in[0,1], the rate has the parametrization (9): F(ψ(u))=uatanh⁡(u).F\bigl(\psi(u)\bigr)=u\operatorname{atanh}(u).(9)✦ For positive entropy hh, the correction has the clipped form (10): LLean(h)=G(max⁡{L−h,0}).L_{\mathrm{Lean}}(h)=G\bigl(\max\{L-h,0\}\bigr).(10)✦

What Is Checked

The statements concern the real binary entropy deficit restricted to nonnegative biases in [0,1][0,1], its inverse on [0,L][0,L], the associated rate and surplus, and the zero extension of the entropy correction, with the clipped formula restricted to h>0h>0.