OctoInformBeta

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

Lemma 2.2: Right derivatives at zero and a convex entropy correction with continuous first derivative

The scalar dissipation rate is convex.

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

For a Boolean coordinate, the scalar rate describes dissipation as a function of the entropy deficit from the fair law. The deficit coordinate and the remaining entropy give two parametrizations of the same scalar correction.

The binary entropy deficit ψ(a)\psi(a) of a Boolean law with bias a∈[−1,1]a\in[-1,1] is the entropy LL of a fair bit minus the entropy H(a)H(a) of that law, in natural logarithmic units: L=log⁡2,ψ(a)=L−H(a)=1+a2log⁡(1+a)+1−a2log⁡(1−a).L=\log 2,\qquad \psi(a)=L-H(a) =\frac{1+a}{2}\log(1+a)+\frac{1-a}{2}\log(1-a).(1) At the endpoints, the entropy formula in (1) uses the convention 0log⁡0=00\log 0=0.

The inverse entropy deficit ψ−1:[0,L]→[0,1]\psi^{-1}:[0,L]\to[0,1] is the function that assigns to a deficit II the unique nonnegative bias with that deficit, as specified by (2): ψ(ψ−1(I))=I,ψ−1(I)∈[0,1](0≤I≤L).\psi\bigl(\psi^{-1}(I)\bigr)=I, \qquad \psi^{-1}(I)\in[0,1] \qquad(0\leq I\leq L).(2) The scalar rate F:ℝ→ℝF:\mathbb R\to\mathbb R is the dissipation profile defined by (3), where atanh\operatorname{atanh} denotes the inverse hyperbolic tangent. F(I)=ψ−1(I)atanh⁡(ψ−1(I)).F(I)=\psi^{-1}(I)\operatorname{atanh}\bigl(\psi^{-1}(I)\bigr).(3)✦

The derivative profile DD and curvature profile KK are real functions on (−∞,L)(-\infty,L) defined by (4); in the positive branches, a=ψ−1(I)a=\psi^{-1}(I): D(I)={2,I≤0,1+a(1−a2)atanh⁡a,0<I<L,K(I)={43,I≤0,(1+a2)atanh⁡a−a(1−a2)2(atanh⁡a)3,0<I<L.\begin{aligned} D(I)&= \begin{cases} 2,&I\leq0,\\[2pt] 1+\dfrac{a}{(1-a^2)\operatorname{atanh}a},&0<I<L, \end{cases}\\[6pt] K(I)&= \begin{cases} \dfrac43,&I\leq0,\\[4pt] \dfrac{(1+a^2)\operatorname{atanh}a-a} {(1-a^2)^2(\operatorname{atanh}a)^3},&0<I<L. \end{cases} \end{aligned}(4)

The dissipation excess E:[0,L)→ℝE:[0,L)\to\mathbb R is the function equal to the scalar rate minus twice the deficit, and the zero extension S:(−∞,L)→ℝS:(-\infty,L)\to\mathbb R of EE is the function equal to EE at nonnegative coordinates and zero at negative coordinates, as in (5): E(I)=F(I)−2I(0≤I<L),S(I)={0,I<0,E(I),0≤I<L.E(I)=F(I)-2I\quad(0\leq I<L),\qquad S(I)= \begin{cases} 0,&I<0,\\ E(I),&0\leq I<L. \end{cases}(5) The entropy correction LLean:ℝ→ℝL_{\mathrm{Lean}}:\mathbb R\to\mathbb R and its derivative profile Q:ℝ→ℝQ:\mathbb R\to\mathbb R are the functions of the entropy coordinate hh defined in (6): LLean(h)={E(L−h),0<h<L,0,otherwise,Q(h)={2−D(L−h),0<h<L,0,otherwise.\begin{aligned} L_{\mathrm{Lean}}(h)&= \begin{cases} E(L-h),&0<h<L,\\ 0,&\text{otherwise}, \end{cases}\\[4pt] Q(h)&= \begin{cases} 2-D(L-h),&0<h<L,\\ 0,&\text{otherwise}. \end{cases} \end{aligned}(6)

The Claim

Theorem 1. At the deficit endpoint, the rate and the right derivatives have the values (7), (8) and (9): F(0)=0.F(0)=0.(7)✦ F+′(0)=2.F'_+(0)=2.(8)✦ D+′(0)=43.D'_+(0)=\frac43.(9)✦ For a deficit II strictly between 00 and LL, the derivative identities and curvature inequality are (10), (11) and (12): F′(I)=D(I).F'(I)=D(I).(10)✦ D′(I)=K(I).D'(I)=K(I).(11)✦ K(I)>0.K(I)>0.(12)✦ On the deficit interval [0,L)[0,L), the derivative profile and rate satisfy (13), (14) and (15): D∈C0([0,L)).D\in C^0([0,L)).(13)✦ F is convex on [0,L).F\text{ is convex on }[0,L).(14)✦ D(I)≥2(0≤I<L).D(I)\geq2\qquad(0\leq I<L).(15)✦ For the dissipation excess, (16) and (17) hold on the same interval: E is convex on [0,L).E\text{ is convex on }[0,L).(16)✦ E(I)≥0(0≤I<L).E(I)\geq0\qquad(0\leq I<L).(17)✦ If II and JJ are deficits with 0≤I≤J<L0\leq I\leq J<L, then (18) holds: E(I)≤E(J).E(I)\leq E(J).(18)✦ For positive entropy coordinates, the correction and its derivative profile satisfy (19), (20), (21) and (22): LLean′(h)=Q(h)(h>0).L_{\mathrm{Lean}}'(h)=Q(h)\qquad(h>0).(19)✦ Q∈C0((0,∞)).Q\in C^0((0,\infty)).(20)✦ LLean(h)≥0(h>0).L_{\mathrm{Lean}}(h)\geq0\qquad(h>0).(21)✦ LLean is convex on (0,∞).L_{\mathrm{Lean}}\text{ is convex on }(0,\infty).(22)✦ If h1h_1 and h2h_2 are entropy coordinates with 0<h1≤h20<h_1\leq h_2, then (23) holds: LLean(h1)≥LLean(h2).L_{\mathrm{Lean}}(h_1)\geq L_{\mathrm{Lean}}(h_2).(23)✦

In the paper, the result reads: FF is smooth on [0,L)[0,L), using right derivatives at zero, with F(0)=0F(0)=0, F+′(0)=2F'_+(0)=2, F+″(0)=4/3F''_+(0)=4/3, F″(I)≥4/3F''(I)\geq4/3 and F‴(I)≥0F'''(I)\geq0; the zero extension SS is nonnegative, nondecreasing, convex and C1C^1 on (−∞,L)(-\infty,L), with S(0)=S′(0)=0S(0)=S'(0)=0, and is not C2C^2 at zero.

What Is Checked

The claim is stated over the inverse binary entropy deficit in (2), the scalar rate and derivative profiles in (3) and (4), the excess in (5), and the entropy correction and its derivative profile in (6), with deficit coordinates in [0,L)[0,L) and entropy coordinates in (0,∞)(0,\infty).