OctoInformBeta

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

Lemma 3.2: Nondecreasing coefficients 0<ck≤M0<c_k\le M and datum M−1log⁡(2cosh⁡Mx)M^{-1}\log(2\cosh Mx) make the drift odd, increasing, concave on x≥0x\ge 0, with nonpositive, nondecreasing relative backward change

Drifts of increasing soft absolute value cascades have backward shape.

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

On an interval where the order parameter γ\gamma, a function on [0,1)[0,1), takes a constant value γ=c\gamma = c, the Parisi equation for the value function Φ\Phi, with covariance ξ(t)=t2/2\xi(t) = t^2/2, reads ∂tΦ+12Φxx+c2Φx2=0,\partial_t\Phi + \tfrac12\,\Phi_{xx} + \tfrac{c}{2}\,\Phi_x^2 = 0,(1) and the drift r=cΦx=γuγr = c\,\Phi_x = \gamma u_\gamma of the diffusion XX, given by dX=γ(t)uγ(t,X)dt+dBdX = \gamma(t)\,u_\gamma(t,X)\,dt + dB with X0=0X_0 = 0 for a Brownian motion BB and uγ=Φxu_\gamma = \Phi_x, solves the viscous Burgers equation ∂tr+12rxx+rrx=0.\partial_t r + \tfrac12\, r_{xx} + r\, r_x = 0.(2)

Fix a real number MM. A step function γ\gamma on a grid of common step length hh takes values c0≤c1≤⋯c_0 \le c_1 \le \cdots in (0,M](0, M], and the terminal datum is the soft absolute value fMf_M of (3), a smooth version of |x||x| at scale 1/M1/M. The question is the shape of the profile x↦r(t,x)x \mapsto r(t,x) at a fixed time, and how fast it moves, relative to its size, as time runs backward.

Cascades and Finite Values

The soft absolute value with parameter MM is the function fMf_M on ℝ\mathbb{R} given by fM(x)=1Mlog⁡(2cosh⁡(Mx)).f_M(x) = \frac1M \log\bigl(2\cosh(Mx)\bigr).(3)✦

For s≥0s \ge 0, a coefficient cc and a function gg on ℝ\mathbb{R}, the step operator TcsT^s_c sends gg to the function TcsgT^s_c g giving, by the Cole-Hopf formula, the value at the start of a step of length ss where γ=c\gamma = c of the solution of (1) with value gg at the end of the step, with BB the Brownian motion, Tcsg(x)=1clog⁡𝔼exp⁡(cg(x+Bs)).T^s_c\, g(x) = \frac1c \log \mathbb{E}\, \exp\bigl(c\, g(x + B_s)\bigr).(4)

Fix a step length h≥0h \ge 0 and coefficients c0,c1,…≥0c_0, c_1, \ldots \ge 0. The iterated cascade of a function ff on ℝ\mathbb{R} is the family of functions fknf^n_k, for n,k≥0n, k \ge 0, given by fk0=f,fkn+1=Tckhfk+1n.f^0_k = f, \qquad f^{n+1}_k = T^h_{c_k}\, f^n_{k+1}.(5)✦

The finite value of ff is the family of functions ΦkN(t,x)\Phi^N_k(t,x), for N,k≥0N, k \ge 0, given by Φk0(t,x)=f(x)\Phi^0_k(t,x) = f(x) and ΦkN+1(t,⋅)={Tckh−tfk+1N,t≤h,Φk+1N(t−h,⋅),t>h.\Phi^{N+1}_k(t,\cdot) = \begin{cases} T^{h-t}_{c_k}\, f^N_{k+1}, & t \le h,\\ \Phi^N_{k+1}(t-h,\cdot), & t > h. \end{cases}(6)✦ The function ΦkN\Phi^N_k is the value function Φγ\Phi_\gamma of the step function γ\gamma equal to ck+jc_{k+j} on [jh,(j+1)h)[jh,(j+1)h) for 0≤j<N0 \le j < N, with terminal datum ff at time NhNh.

Backward Shape

The Burgers rate of a function r(t,x)r(t,x) is the function ρr(t,x)=12rxx(t,x)+r(t,x)rx(t,x);\rho_r(t,x) = \tfrac12\, r_{xx}(t,x) + r(t,x)\, r_x(t,x);(7)✦ along a solution of (2) it equals −∂tr-\partial_t r, the rate at which rr changes as time runs backward. The backward Wronskian of rr is the function Jr(t,x)=r(t,x)ρr,x(t,x)−rx(t,x)ρr(t,x).J_r(t,x) = r(t,x)\,\rho_{r,x}(t,x) - r_x(t,x)\,\rho_r(t,x).(8)✦

A function rr on ℝ\mathbb{R} has backward shape if it is odd, its derivative is positive everywhere, and on the half-line x≥0x \ge 0 its second derivative, its Burgers rate and its backward Wronskian, taken without the variable tt, satisfy r″(x)≤0,ρr(x)≤0,Jr(x)≥0(x≥0).r''(x) \le 0, \qquad \rho_r(x) \le 0, \qquad J_r(x) \ge 0 \qquad (x \ge 0).(9)✦

For rr of backward shape, the relative backward change ρr/r\rho_r/r is nonpositive and nondecreasing on x>0x > 0.

The Claim

Theorem 1. Let MM be a real number, let (ck)k≥0(c_k)_{k \ge 0} be a nondecreasing sequence with 0<ck≤M0 < c_k \le M for every kk, and let h≥0h \ge 0; let fknf^n_k and ΦkN\Phi^N_k be the iterated cascade and the finite value of the soft absolute value fMf_M for these coefficients and step length hh. Then for all N,i≥0N, i \ge 0 and every t≤ht \le h the drift x↦ciΦi,xN+1(t,x)x \longmapsto c_i\, \Phi^{N+1}_{i,x}(t,x)(10)✦ has backward shape, and for all N,i≥0N, i \ge 0 so does the drift x↦ci(fiN)′(x).x \longmapsto c_i\, (f^N_i)'(x).(11)✦

What Is Checked

The claim is stated over the soft absolute value fMf_M, its iterated cascade and its finite value for nondecreasing coefficients 0<ck≤M0 < c_k \le M on a grid of step h≥0h \ge 0, and over the backward shape, built from the Burgers rate and the backward Wronskian.