OctoInformBeta

OpenAI preprint, September 2026 · §3. Shape inequalities for finite step functions

Lemma 3.3: For nondecreasing coefficients and the soft absolute value as terminal datum, the forward data are even and concave, the boundary ratio nonnegative on x≥0x \ge 0

Shape of the forward score and sign of its ratio.

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

Let the order parameter be a nondecreasing step function on a grid of mesh h>0h > 0, with γ=ck\gamma = c_k on [kh,(k+1)h)[kh,(k+1)h). The value function Φγ\Phi_\gamma is computed backward through the grid from the soft absolute value aMa_M, a smooth approximation of |x||x|, at the final time, and forward data run through the grid in the other direction. At each grid time ihih there is a forward datum ℓi\ell_i, a function on ℝ\mathbb{R}, and it defines a forward factor gg; the score of ℓi\ell_i is −∂xlog⁡g-\partial_x \log g.

The Backward and Forward Data

The soft absolute value with parameter MM is the function aMa_M on ℝ\mathbb{R} given by aM(x)=1Mlog⁡(2cosh⁡(Mx)).a_M(x) = \frac{1}{M}\log\bigl(2\cosh(Mx)\bigr).(1)✦ For M>0M > 0 it is smooth, even and convex, with |x|≤aM(x)≤|x|+(log⁡2)/M|x| \le a_M(x) \le |x| + (\log 2)/M.

Given a sequence c=(ck)k≥0c = (c_k)_{k \ge 0} of nonnegative reals, a step h≥0h \ge 0 and a function ff on ℝ\mathbb{R}, the iterated cascade CnkfC^k_n f with nn steps from index kk is the function on ℝ\mathbb{R} defined by recursion on nn: C0kf=f,Cn+1kf=ℒck,hCnk+1f,C^k_0 f = f, \qquad C^k_{n+1} f = \mathcal{L}_{c_k,h}\, C^{k+1}_n f,(2)✦ where ℒc,h\mathcal{L}_{c,h}, the logarithmic heat semigroup with parameter cc over time hh, is the operator on functions on ℝ\mathbb{R} that carries the function at the end of an interval of length hh on which the order parameter equals cc to its value at the start of that interval. The cascade CnkfC^k_n f is the Parisi recursion across nn successive intervals of length hh on which the order parameter equals ck,…,ck+n−1c_k, \dots, c_{k+n-1}, ending at ff.

Given also a number of steps N≥0N \ge 0, the backward datum of ff at index i≥0i \ge 0 is the function Φi\Phi_i on ℝ\mathbb{R} given by Φi=CN−iif(i≤N),Φi=f(i≥N),\Phi_i = C^i_{N-i} f \quad (i \le N), \qquad \Phi_i = f \quad (i \ge N),(3)✦ the value at time ihih of the recursion that ends at ff at time NhNh.

The forward datum of ff at index i≥0i \ge 0 is the function ℓi\ell_i on ℝ\mathbb{R} defined by recursion on ii: ℓ0(x)=−c0Φ0(0),ℓ1(x)=−12log⁡(2πh)−c0Φ0(0)−(c1−c0)Φ1(x),ℓi+1(x)=ℱih,h[ℓi](x)−(ci+1−ci)Φi+1(x)(i≥1),\begin{aligned} \ell_0(x) &= -c_0\,\Phi_0(0),\\ \ell_1(x) &= -\tfrac{1}{2}\log(2\pi h) - c_0\,\Phi_0(0) - (c_1 - c_0)\,\Phi_1(x),\\ \ell_{i+1}(x) &= \mathcal{F}_{ih,h}[\ell_i](x) - (c_{i+1} - c_i)\,\Phi_{i+1}(x) \qquad (i \ge 1), \end{aligned}(4)✦ where ℱih,h\mathcal{F}_{ih,h} is the forward correction over one step of length hh from time ihih.

The Score and the Ratio

For a real number a≠0a \neq 0 and a differentiable function ℓ\ell on ℝ\mathbb{R}, the score of ℓ\ell at time aa is the function sa[ℓ](x)=xa−ℓx(x),s_a[\ell](x) = \frac{x}{a} - \ell_x(x),(5)✦ where g(x)=eℓ(x)−x2/(2a)g(x) = e^{\ell(x) - x^2/(2a)} is the forward factor. The forward score at index i≥1i \ge 1 is si=sih[ℓi]s_i = s_{ih}[\ell_i].

The Wronskian of two differentiable functions rr and ss on ℝ\mathbb{R} is the function W(r,s)=rsx−srx.W(r,s) = r\,s_x - s\,r_x.(6)✦ Where r≠0r \neq 0, W(r,s)=rjW(r,s) = r\,j for the ratio j=r∂x(s/r)j = r\,\partial_x(s/r). The boundary ratio at index i≥1i \ge 1 is the Wronskian W(ri,si)W(r_i, s_i) of ri=ciΦi,xr_i = c_i\,\Phi_{i,x} and the forward score sis_i.✦

The Claim

Theorem 1.

  1. Let MM be a real number, (ck)k≥0(c_k)_{k \ge 0} a nondecreasing sequence of positive reals with ck≤Mc_k \le M for every kk, and h>0h > 0, and let Φi\Phi_i and ℓi\ell_i be the backward and forward data of f=aMf = a_M on NN steps. For all integers N,i≥0N, i \ge 0, the forward datum ℓi\ell_i has the forward shape: ℓi(−x)=ℓi(x),ℓi,xx(x)≤0(x∈ℝ),ℓi,xxx(x)≥0(x≥0).\ell_i(-x) = \ell_i(x), \qquad \ell_{i,xx}(x) \le 0 \qquad (x \in \mathbb{R}), \qquad \ell_{i,xxx}(x) \ge 0 \qquad (x \ge 0).(7)✦

  2. Under the hypotheses of (a), for all integers NN and ii with 0<i≤N0 < i \le N, the boundary ratio satisfies W(ri,si)(x)≥0(x≥0).W(r_i, s_i)(x) \ge 0 \qquad (x \ge 0).(8)✦

  3. Let ℓ\ell and φ\varphi be infinitely differentiable functions on ℝ\mathbb{R}, and a≠0a \neq 0, c≠0c \neq 0 and dd real numbers. Then for every real xx, W(dφx,sa[ℓ−(d−c)φ])(x)=dcW(cφx,sa[ℓ])(x).W\bigl(d\,\varphi_x,\ s_a[\ell - (d-c)\varphi]\bigr)(x) = \frac{d}{c}\, W\bigl(c\,\varphi_x,\ s_a[\ell]\bigr)(x).(9)✦

For i≥1i \ge 1, (7) makes the forward score sis_i odd, with si,x≥1/(ih)s_{i,x} \ge 1/(ih), and concave on x>0x > 0.

What Is Checked

The claims are stated over a nondecreasing sequence of positive coefficients bounded by MM, the soft absolute value aMa_M as terminal datum, the iterated cascade of logarithmic heat semigroups with the backward data Φi\Phi_i and forward data ℓi\ell_i built from it, the score sa[ℓ]s_a[\ell], the Wronskian W(r,s)W(r,s) with the boundary ratio W(ri,si)W(r_i, s_i), and the forward shape.