OctoInformBeta

OpenAI preprint, September 2026 · §6. Full support and regularity

Lemma 6.1: Their limit is 12∫01(β−γ)(q−t)dt\frac12\int_0^1(\beta-\gamma)(q-t)\,dt for steps 1/(n+1)1/(n+1) toward β\beta, and this integral is nonnegative when γ\gamma minimizes 𝒫\mathcal P

Difference quotients of the Parisi functional along chords converge.

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

The zero-temperature Parisi formula expresses the ground-state energy of the Sherrington-Kirkpatrick model as the minimum of the Parisi functional 𝒫\mathcal P over order parameters γ\gamma, where 𝒫(γ)\mathcal P(\gamma) is computed from the value function Φγ\Phi_\gamma of a stochastic control problem. Given an order parameter γ\gamma and another one, β\beta, the question is how 𝒫\mathcal P changes to first order when γ\gamma moves a small step toward β\beta, and what this says when γ\gamma is a minimizer.

Order Parameters and the Parisi Functional

The extension by zero of a real function ff on the time interval [0,1)[0,1) is the function on ℝ\mathbb R equal to f(t)f(t) for 0≤t<10\le t<1 and to 00 elsewhere.✦ An order parameter is a real function γ\gamma on [0,1)[0,1) that is nonnegative, nondecreasing and right-continuous, and whose extension by zero is Lebesgue integrable.✦ Wherever an order parameter is evaluated outside [0,1)[0,1), its extension by zero is meant.

For order parameters γ\gamma, β\beta and a weight ε∈[0,1]\varepsilon\in[0,1], the chord (1−ε)γ+εβ(1-\varepsilon)\gamma+\varepsilon\beta is the order parameter whose value at t∈[0,1)t\in[0,1) is (1−ε)γ(t)+εβ(t)(1-\varepsilon)\gamma(t)+\varepsilon\beta(t).✦ The variation step of index n∈ℕn\in\mathbb N is εn=1/(n+1)\varepsilon_n=1/(n+1)✦, and the chord perturbation γn\gamma_n is the chord (1−εn)γ+εnβ(1-\varepsilon_n)\gamma+\varepsilon_n\beta✦.

A Brownian system WW on a measurable space Ω\Omega is a probability measure ℙ\mathbb P on Ω\Omega, its law, together with a family B=(Bt)t≥0B=(B_t)_{t\ge0} of random variables on Ω\Omega that is a standard Brownian motion under ℙ\mathbb P.✦ Expectation under ℙ\mathbb P is written 𝔼\mathbb E. The increment filtration of WW after a time t≥0t\ge0 is the filtration (ℱst)s≥0(\mathcal F^t_s)_{s\ge0} in which ℱst\mathcal F^t_s is generated by the increments Bt+r−BtB_{t+r}-B_t, 0≤r≤s0\le r\le s.✦ A bounded control α=(αs)s≥0\alpha=(\alpha_s)_{s\ge0} from time tt is a real process, indexed by the time elapsed since tt, that is progressively measurable for this filtration and satisfies |αs(ω)|≤1|\alpha_s(\omega)|\le1 for all ss and ω\omega.✦

The value function of an order parameter γ\gamma on a Brownian system WW is the function Φγ\Phi_\gamma of a time tt and a position xx given by Φγ(t,x)=supα𝔼[|x+B1−Bt+∫t1γ(s)αs−tds|−12∫t1γ(s)αs−t2ds],\Phi_\gamma(t,x)=\sup_\alpha\mathbb E\Bigl[\Bigl|x+B_1-B_t+\int_t^1\gamma(s)\,\alpha_{s-t}\,ds\Bigr| -\frac12\int_t^1\gamma(s)\,\alpha_{s-t}^2\,ds\Bigr],(1)✦ the supremum running over bounded controls α\alpha from time tt; the quantity in the brackets is the control payoff of α\alpha. The value gradient uγ(t,x)=Φγ,x(t,x)u_\gamma(t,x)=\Phi_{\gamma,x}(t,x) is the derivative of Φγ(t,⋅)\Phi_\gamma(t,\cdot) at xx.✦

The Parisi functional on WW assigns to an order parameter γ\gamma the number 𝒫(γ)=Φγ(0,0)−12∫01tγ(t)dt.\mathcal P(\gamma)=\Phi_\gamma(0,0)-\frac12\int_0^1t\,\gamma(t)\,dt.(2)✦ An order parameter γ\gamma is a Parisi minimizer when 𝒫(γ)≤𝒫(γ′)\mathcal P(\gamma)\le\mathcal P(\gamma') for every order parameter γ′\gamma'.

The Diffusion and the Function qq

The left mesh time of level nn of a time t≥0t\ge0 is τn(t)=⌊(n+1)t⌋/(n+1)\tau_n(t)=\lfloor(n+1)t\rfloor/(n+1), the last point of the grid of mesh 1/(n+1)1/(n+1) not after tt.✦ The progressive version of BB is the process B̃t=lim⁡n→∞Bτn(t)\tilde B_t=\lim_{n\to\infty}B_{\tau_n(t)}, the limit taken at each sample point.✦ The driving Brownian motion of WW is the process B‾t=B̃t−B̃0\bar B_t=\tilde B_t-\tilde B_0.✦

A bounded Lipschitz drift is a jointly measurable function bb of a time tt and a position xx, with a bound M≥0M\ge0 and a Lipschitz constant L≥0L\ge0, such that |b(t,x)|≤M|b(t,x)|\le M and |b(t,x)−b(t,y)|≤L|x−y||b(t,x)-b(t,y)|\le L|x-y| for all tt, xx, yy.✦ Along a real process ZZ, the correction iterates of bb are Y0=0Y^0=0 and Ytk+1=∫0tb(s,Zs+Ysk)dsY^{k+1}_t=\int_0^tb(s,Z_s+Y^k_s)\,ds✦, and its drift correction YY is their limit as k→∞k\to\infty✦. The drift equation solution of bb along ZZ is the Picard solution Z+YZ+Y of Vt=Zt+∫0tb(s,Vs)dsV_t=Z_t+\int_0^tb(s,V_s)\,ds.✦

For a horizon T∈[0,1)T\in[0,1), write [t]T=max⁡(0,min⁡(T,t))[t]_T=\max(0,\min(T,t)) for the clamp of a time tt to the strip [0,T][0,T].✦ The clamped drift is the product γ([t]T)uγ([t]T,x)\gamma([t]_T)\,u_\gamma([t]_T,x) of the clamped coefficient and the clamped gradient; for 0≤t≤T0\le t\le T it is the drift γ(t)uγ(t,x)\gamma(t)\,u_\gamma(t,x).✦ The strip drift with horizon TT is the bounded Lipschitz drift given by the clamped drift.✦

The horizons Tn=1−1/(n+2)T_n=1-1/(n+2), n∈ℕn\in\mathbb N, increase from T0=1/2T_0=1/2 toward 11.✦ The localized diffusion XnX^n is the drift equation solution along B‾\bar B for the strip drift with horizon TnT_n, so that Xtn=B‾t+∫0tγ(min⁡(s,Tn))uγ(min⁡(s,Tn),Xsn)dsX^n_t=\bar B_t+\int_0^t\gamma(\min(s,T_n))\,u_\gamma(\min(s,T_n),X^n_s)\,ds.✦ The raw diffusion limit is Xˆt=lim⁡n→∞Xtn\hat X_t=\lim_{n\to\infty}X^n_t.✦ The gradient cut at time 11 is the function of a time tt and a position xx equal to uγ(t,x)u_\gamma(t,x) for t<1t<1 and to 00 for t≥1t\ge1, and the raw feedback FtF_t is the value of the gradient cut at (t,Xˆt)(t,\hat X_t).

The diffusion of γ\gamma is the process Xt=B‾t+∫0tγ(s)Fsds,t≥0,X_t=\bar B_t+\int_0^t\gamma(s)\,F_s\,ds,\qquad t\ge0,(3)✦ the integrated form of dXt=γ(t)uγ(t,Xt)dt+dB‾tdX_t=\gamma(t)\,u_\gamma(t,X_t)\,dt+d\bar B_t, X0=0X_0=0, with the gradient read along Xˆ\hat X. The second moment of the gradient along XX is the function qq of a time tt given by q(t)=𝔼[uγ(t,Xt)2].q(t)=\mathbb E\bigl[u_\gamma(t,X_t)^2\bigr].(4)✦

The Claim

Theorem 1. Let WW be a Brownian system on Ω\Omega, let γ\gamma be an order parameter, and let qq be the second moment of the gradient along the diffusion of γ\gamma. For every order parameter β\beta, the chord perturbations γn\gamma_n from γ\gamma toward β\beta satisfy limn→∞𝒫(γn)−𝒫(γ)εn=12∫01(β(t)−γ(t))(q(t)−t)dt.\lim_{n\to\infty}\frac{\mathcal P(\gamma_n)-\mathcal P(\gamma)}{\varepsilon_n} =\frac12\int_0^1\bigl(\beta(t)-\gamma(t)\bigr)\bigl(q(t)-t\bigr)\,dt.(5)✦ If γ\gamma is a Parisi minimizer, then every order parameter β\beta satisfies ∫01(β(t)−γ(t))(q(t)−t)dt≥0.\int_0^1\bigl(\beta(t)-\gamma(t)\bigr)\bigl(q(t)-t\bigr)\,dt\ge0.(6)✦

In the paper, the result reads: if η∈L1(0,1)\eta\in L^1(0,1) and, for some ε0>0\varepsilon_0>0, γ+εη\gamma+\varepsilon\eta is an order parameter for 0≤ε≤ε00\le\varepsilon\le\varepsilon_0, then the one-sided derivative of ε↦𝒫(γ+εη)\varepsilon\mapsto\mathcal P(\gamma+\varepsilon\eta) at ε=0+\varepsilon=0^+ is 12∫01η(t)(q(t)−t)dt\frac12\int_0^1\eta(t)\,(q(t)-t)\,dt.

What Is Checked

The limit (5) and the inequality (6) are stated over a Brownian system WW and order parameters γ\gamma and β\beta, through the Parisi functional (2) built on the value function (1), the chord perturbations γn\gamma_n with steps εn=1/(n+1)\varepsilon_n=1/(n+1), the condition that γ\gamma be a Parisi minimizer, and the function qq of (4) along the diffusion (3).

About these constructions the library proves that each variation step εn\varepsilon_n lies in [0,1][0,1]✦ and each horizon TnT_n in [0,1)[0,1)✦, and that for a horizon T∈[0,1)T\in[0,1) the clamped drift is jointly measurable, bounded in absolute value by γ(T)\gamma(T) and Lipschitz in the position.