OctoInformBeta

OpenAI preprint, September 2026 · §2. Integrable coefficients and the controlled diffusion

Lemma 2.2: The equation dX=γ(t)uγ(t,X)dt+dBdX = \gamma(t)\,u_\gamma(t,X)\,dt + dB, X0=0X_0 = 0, has a solution on [0,1][0,1], along which q(0)=0q(0) = 0 and the formula for q′q' holds on [0,1)[0,1)

The controlled diffusion is unique and q′=𝔼Φγ,xx2q' = \mathbb{E}\,\Phi_{\gamma,xx}^2..

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 γ\gamma be an order parameter of the zero-temperature Sherrington-Kirkpatrick model and Φγ\Phi_\gamma the value function of the stochastic control problem behind its Parisi formula. The gradient uγ=Φγ,xu_\gamma = \Phi_{\gamma,x} is the feedback of the diffusion dX=γ(t)uγ(t,X)dt+dBdX = \gamma(t)\,u_\gamma(t,X)\,dt + dB driven by a Brownian motion BB and started from X0=0X_0 = 0, and the second moment q(t)q(t) of uγ(t,Xt)u_\gamma(t,X_t) is the function from which the contact potential G(s)=∫s1(q(t)−t)dtG(s) = \int_s^1 (q(t) - t)\,dt is built. The questions are whether this diffusion exists and is unique, and how qq starts and changes in tt.

The Control Problem

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 standard Brownian motion B=(Bt)t≥0B = (B_t)_{t \ge 0} under ℙ\mathbb{P}, each BtB_t a random variable on Ω\Omega.✦ Expectations 𝔼\mathbb{E} are taken under ℙ\mathbb{P}.

An order parameter is a function γ\gamma on the time interval [0,1)[0,1) that is nonnegative, nondecreasing and right-continuous, and whose extension by zero, the function on ℝ\mathbb{R} equal to γ\gamma on [0,1)[0,1) and to 00 elsewhere, is Lebesgue integrable.✦ Below, γ\gamma also denotes this extension.

For a time t≥0t \ge 0, the increment filtration (ℱst)s≥0(\mathcal{F}^t_s)_{s \ge 0} of WW after tt is the natural filtration of the increments of BB after tt: ℱst\mathcal{F}^t_s is the σ\sigma-algebra generated by Bt+r−BtB_{t+r} - B_t, 0≤r≤s0 \le r \le s.✦ A bounded control from time tt is a real process α=(αs)s≥0\alpha = (\alpha_s)_{s \ge 0} on Ω\Omega, indexed by the time elapsed since tt, that is progressively measurable for (ℱst)s≥0(\mathcal{F}^t_s)_{s \ge 0} and satisfies |αs(ω)|≤1|\alpha_s(\omega)| \le 1 for all ss and ω\omega.✦

The value function of γ\gamma is the function Φγ\Phi_\gamma of a time tt and a position xx given by the supremum, over bounded controls α\alpha from time tt, of the expectation of the control payoff, the random variable in brackets: 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)✦ a negative time being read as 00 wherever BB or α\alpha is evaluated. The 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✦, and the curvature Φγ,xx(t,x)\Phi_{\gamma,xx}(t,x) is the derivative of uγ(t,⋅)u_\gamma(t,\cdot) at xx.✦

The Diffusion

A controlled diffusion of γ\gamma driven by WW is a real process YY on Ω\Omega, progressively measurable for the natural filtration of BB, such that for ℙ\mathbb{P}-almost every ω\omega the path t↦Yt(ω)t \mapsto Y_t(\omega) is continuous on [0,1][0,1], starts at Y0(ω)=0Y_0(\omega) = 0 and satisfies Yt(ω)=Bt(ω)+∫0tγ(s)uγ(s,Ys(ω))ds,0≤t≤1.Y_t(\omega) = B_t(\omega) + \int_0^t \gamma(s)\,u_\gamma\bigl(s, Y_s(\omega)\bigr)\,ds, \qquad 0 \le t \le 1.(2)✦ The set of controlled diffusions of γ\gamma driven by WW is written 𝒟γ(W)\mathcal{D}_\gamma(W).

A bounded Lipschitz drift is a jointly measurable function bb of a time tt and a position xx, with a bound M≥0M \ge 0 and a constant L≥0L \ge 0, 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 and yy.✦ The Picard iterates of a bounded Lipschitz drift bb along a real process ZZ on Ω\Omega are the processes CnC^n with C0=0C^0 = 0 and Ctn+1=∫0tb(s,Zs+Csn)ds,C^{n+1}_t = \int_0^t b(s, Z_s + C^n_s)\,ds,(3)✦ The drift correction of bb along ZZ is the process Ct=lim⁡nCtnC_t = \lim_n C^n_t, the limit taken at each time and sample point✦, and the drift equation solution of bb driven by ZZ is the process Z+CZ + C, the Picard solution of Vt=Zt+∫0tb(s,Vs)dsV_t = Z_t + \int_0^t b(s, V_s)\,ds.✦

For a horizon T∈[0,1)T \in [0,1), the clamp of a time tt to the strip [0,T][0,T] is t‾=max⁡(0,min⁡(T,t))\bar t = \max(0, \min(T,t))✦, and the clamped drift is bT(t,x)=γ(t‾)uγ(t‾,x)b_T(t,x) = \gamma(\bar t)\,u_\gamma(\bar t, x).✦ For 0≤t≤T0 \le t \le T it is the drift γ(t)uγ(t,x)\gamma(t)\,u_\gamma(t,x) of (2). The strip drift at TT is the drift bTb_T, which is a bounded Lipschitz drift.✦ The horizons are Tn=1−1/(n+2)T_n = 1 - 1/(n+2) for natural numbers nn, increasing from T0=1/2T_0 = 1/2 toward 11.✦

The progressive version of BB is the process B̃t=lim⁡nB⌊(n+1)t⌋/(n+1)\tilde B_t = \lim_n B_{\lfloor (n+1)t \rfloor/(n+1)}, the limit of BB at the last points not after tt of the grids of mesh 1/(n+1)1/(n+1).✦ The driving Brownian motion of WW is β=B̃−B̃0\beta = \tilde B - \tilde B_0.✦

The localized diffusion XnX^n is the drift equation solution driven by β\beta for the strip drift at TnT_n, the Picard solution of Xtn=βt+∫0tγ(min⁡(s,Tn))uγ(min⁡(s,Tn),Xsn)ds.X^n_t = \beta_t + \int_0^t \gamma(\min(s,T_n))\,u_\gamma\bigl(\min(s,T_n), X^n_s\bigr)\,ds.(4)✦ The raw diffusion limit is Xˆt=lim⁡nXtn\hat X_t = \lim_n X^n_t, at each time and sample point✦, and the raw feedback is Ft=uγ(t,Xˆt)F_t = u_\gamma(t, \hat X_t) for t<1t < 1 and Ft=0F_t = 0 for t≥1t \ge 1. The diffusion process of γ\gamma driven by WW is Xt=βt+∫0tγ(s)Fsds,t≥0.X_t = \beta_t + \int_0^t \gamma(s)\,F_s\,ds, \qquad t \ge 0.(5)✦

The second moment of the gradient along a real process YY is the function t↦𝔼[uγ(t,Yt)2]t \mapsto \mathbb{E}[u_\gamma(t,Y_t)^2]✦; along the diffusion process XX it is q(t)=𝔼[uγ(t,Xt)2],q(t) = \mathbb{E}\bigl[u_\gamma(t,X_t)^2\bigr],(6) and the second moment of the curvature along XX is t↦𝔼[Φγ,xx(t,Xt)2]t \mapsto \mathbb{E}[\Phi_{\gamma,xx}(t,X_t)^2].✦

The Claim

Theorem 1. Let WW be a Brownian system with law ℙ\mathbb{P} and Brownian motion BB, let γ\gamma be an order parameter, let XX be the diffusion process of γ\gamma driven by WW, and let qq be the second moment of the gradient along XX. Then X∈𝒟γ(W),X \in \mathcal{D}_\gamma(W),(7)✦ and for every Y∈𝒟γ(W)Y \in \mathcal{D}_\gamma(W) and ℙ\mathbb{P}-almost every ω\omega, Yt(ω)=Xt(ω)for all t∈[0,1].Y_t(\omega) = X_t(\omega) \qquad \text{for all } t \in [0,1].(8)✦ Moreover q(0)=0,q(0) = 0,(9)✦ and qq is differentiable on [0,1)[0,1), from the right at 00, with q′(t)=𝔼[Φγ,xx(t,Xt)2],0≤t<1.q'(t) = \mathbb{E}\bigl[\Phi_{\gamma,xx}(t,X_t)^2\bigr], \qquad 0 \le t < 1.(10)✦

In the paper, the result reads: the diffusion has a unique strong solution on [0,1][0,1]; on every strip [0,T][0,T] with T<1T < 1, uγ(t,Xt)=∫0tΦγ,xx(s,Xs)dBsu_\gamma(t,X_t) = \int_0^t \Phi_{\gamma,xx}(s,X_s)\,dB_s; in particular qq is C1C^1 on [0,1)[0,1) with q(0)=0q(0) = 0 and q′(t)=𝔼[Φγ,xx(t,Xt)2]q'(t) = \mathbb{E}[\Phi_{\gamma,xx}(t,X_t)^2].

What Is Checked

The theorem is stated over a Brownian system WW and an order parameter γ\gamma: the value function Φγ\Phi_\gamma of (1) with its gradient and curvature, the controlled diffusions of (2), the diffusion process (5) built from the strip drifts at the horizons TnT_n by Picard iteration and a limit in nn, and the second moments of the gradient and of the curvature along it.

The library also shows that every horizon TnT_n lies in [0,1)[0,1).✦