OctoInformBeta

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

Lemma 2.3: The first two derivatives of q′q' on an interval where γ=c\gamma = c, as expectations along XX of polynomials in the spatial derivatives of Φγ\Phi_\gamma

Where γ\gamma is constant, q′q' is twice differentiable.

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 with value function Φγ(t,x)\Phi_\gamma(t,x) and value gradient uγ=Φγ,xu_\gamma = \Phi_{\gamma,x}, and let XX be the diffusion dX=γ(t)uγ(t,X)dt+dBdX = \gamma(t)\,u_\gamma(t,X)\,dt + dB with X0=0X_0 = 0, driven by the Brownian motion BB. Write q(t)=𝔼[uγ(t,Xt)2]q(t) = \mathbb{E}[u_\gamma(t,X_t)^2]; its derivative q′q' is the second moment of the curvature Φγ,xx\Phi_{\gamma,xx} along XX. The question is what q′q' does on an open interval where γ\gamma is constant, equal to cc: whether it can be differentiated twice there, and how its derivatives read in terms of Φγ,xx\Phi_{\gamma,xx}, Φγ,xxx\Phi_{\gamma,xxx} and Φγ,xxxx\Phi_{\gamma,xxxx} along XX.

The Value Function and the Diffusion

The time interval is [0,1)[0,1).✦ The extension by zero of a function ff on [0,1)[0,1) is the function on ℝ\mathbb{R} that agrees with ff on [0,1)[0,1) and vanishes elsewhere.✦ An order parameter γ\gamma is a nonnegative, nondecreasing, right-continuous function on the time interval whose extension by zero is Lebesgue integrable; off [0,1)[0,1), γ(s)\gamma(s) stands for the value of this extension.✦

A Brownian system WW is a probability measure ℙ\mathbb{P} on a measurable space Ω\Omega 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; 𝔼\mathbb{E} is the expectation under ℙ\mathbb{P}.✦

A bounded control from a time tt is a real process α=(αs)s≥0\alpha = (\alpha_s)_{s \ge 0}, indexed by the time elapsed since tt, with |αs|≤1|\alpha_s| \le 1, that is progressively measurable for the increment filtration of WW after tt, whose σ\sigma-algebra at ss is generated by the increments Bt+r−BtB_{t+r} - B_t for 0≤r≤s0 \le r \le s.✦ The value function Φγ\Phi_\gamma is the supremum over these controls of the expected control payoff: 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 value curvature Φγ,xx\Phi_{\gamma,xx} of Φγ\Phi_\gamma is the derivative in xx of the value gradient Φγ,x\Phi_{\gamma,x}.✦ The value rate of Φγ\Phi_\gamma is the right-hand side of the Parisi equation ∂tΦγ=−12(Φγ,xx+γΦγ,x2)\partial_t\Phi_\gamma = -\frac12\bigl(\Phi_{\gamma,xx} + \gamma\,\Phi_{\gamma,x}^2\bigr).✦

The driving Brownian motion of WW is the progressive version of BB started at zero, the process B̃t−B̃0\tilde B_t - \tilde B_0, where B̃t\tilde B_t is the limit as n→∞n \to \infty of BB at the left mesh time ⌊(n+1)t⌋/(n+1)\lfloor (n+1)t \rfloor/(n+1); it is written BB again, and the diffusions are driven by it.✦

A bounded Lipschitz drift is a jointly measurable function f(t,x)f(t,x) with |f(t,x)|≤M|f(t,x)| \le M and |f(t,x)−f(t,y)|≤L|x−y||f(t,x) - f(t,y)| \le L\,|x-y| for constants M,L≥0M, L \ge 0.✦ The solution of a bounded Lipschitz drift ff is the Picard solution of Zt=Bt+∫0tf(s,Zs)dsZ_t = B_t + \int_0^t f(s,Z_s)\,ds: the process Z=B+YZ = B + Y, where the drift correction YY is the limit of the Picard iterates Y0=0Y^0 = 0 and Ytn+1=∫0tf(s,Bs+Ysn)dsY^{n+1}_t = \int_0^t f(s, B_s + Y^n_s)\,ds✦.

Fix a horizon TT in [0,1)[0,1) and let t‾=max⁡(0,min⁡(T,t))\bar t = \max(0,\min(T,t)) be the clamp of a real tt to the strip [0,T][0,T].✦ The clamped drift is the clamped coefficient γ(t‾)\gamma(\bar t) times the clamped gradient uγ(t‾,x)u_\gamma(\bar t,x), the drift γ(t)uγ(t,x)\gamma(t)\,u_\gamma(t,x) of XX for 0≤t≤T0 \le t \le T.✦ The strip drift is the clamped drift, a bounded Lipschitz drift with bound γ(T)\gamma(T); its solution YTY^T is the diffusion with drift frozen after TT.✦

For the horizon times Tn=1−1/(n+2)T_n = 1 - 1/(n+2), increasing from 1/21/2 toward 11, the localized diffusion is Xn=YTnX^n = Y^{T_n}✦ and the raw diffusion Xˆt\hat X_t is the limit of XtnX^n_t as n→∞n \to \infty✦. The raw feedback FtF_t is the gradient cut at time 11 along Xˆ\hat X, equal to uγ(t,Xˆt)u_\gamma(t,\hat X_t) for t<1t < 1 and to 00 for t≥1t \ge 1✦, and the diffusion is Xt=Bt+∫0tγ(s)Fsds,t≥0.X_t = B_t + \int_0^t \gamma(s)\,F_s\,ds, \qquad t \ge 0.(2)✦

The jet of order kk along the diffusion XX at a time tt is the random variable ∂xkΦγ(t,Xt)\partial_x^k\Phi_\gamma(t,X_t).✦ The second moment of curvature of XX is the function mm of time m(t)=𝔼[Φγ,xx(t,Xt)2].m(t) = \mathbb{E}\bigl[\Phi_{\gamma,xx}(t,X_t)^2\bigr].(3)✦

The Strip Terms

A bounded smooth function is a C∞C^\infty function on ℝ\mathbb{R} with every derivative bounded✦, and a bounded smooth family is a jointly measurable function F(t,x)F(t,x) whose sections F(t,⋅)F(t,\cdot) are bounded smooth functions with each ∂xkF\partial_x^k F bounded uniformly in tt and xx✦. A time field with horizon TT is a pair of bounded smooth families, a function ff and a rate dd, such that ff is jointly continuous, d(s,x)→d(t,x)d(s,x) \to d(t,x) as s↓ts \downarrow t, and for 0≤t<T0 \le t < T the right derivative of s↦f(s,x)s \mapsto f(s,x) at tt is d(t,x)d(t,x).✦ Sums and scalar multiples of time fields are taken in function and rate alike, and the product of fields (f1,d1)(f_1,d_1) and (f2,d2)(f_2,d_2) has function f1f2f_1 f_2 and rate d1f2+f1d2d_1 f_2 + f_1 d_2✦. The expected value of a time field (f,d)(f,d) along a bounded Lipschitz drift is the function t↦𝔼f(t,Zt)t \mapsto \mathbb{E}\,f(t,Z_t), where ZZ is the solution of that drift.✦

For an order nn, the clamped jet field is the time field with horizon TT whose function is the clamped gradient jet ∂xnuγ(t‾,x)=∂xn+1Φγ(t‾,x)\partial_x^n u_\gamma(\bar t,x) = \partial_x^{n+1}\Phi_\gamma(\bar t,x) and whose rate is ∂xn+1\partial_x^{n+1} of the clamped value rate, the value rate at (t‾,x)(\bar t,x).✦ The product of the field of order 11 with itself is the squared curvature field, with function Φγ,xx(t‾,x)2\Phi_{\gamma,xx}(\bar t,x)^2.✦ For a real cc, sums, scalar multiples and products of the clamped jet fields of orders 11, 22, 33 and the squared curvature field give the second and third curvature fields, with functions gc(t‾,x)g_c(\bar t,x) and hc(t‾,x)h_c(\bar t,x), where gc(t,x)=(Φγ,xxx2−2cΦγ,xx3)(t,x),g_c(t,x) = \bigl(\Phi_{\gamma,xxx}^2 - 2c\,\Phi_{\gamma,xx}^3\bigr)(t,x),(4)✦ hc(t,x)=(Φγ,xxxx2−12cΦγ,xxΦγ,xxx2+6c2Φγ,xx4)(t,x).h_c(t,x) = \bigl(\Phi_{\gamma,xxxx}^2 - 12c\,\Phi_{\gamma,xx}\Phi_{\gamma,xxx}^2 + 6c^2\,\Phi_{\gamma,xx}^4\bigr)(t,x).(5)✦

The strip second term and the strip third term are the expected values of these two fields along the strip drift: ΛT,c(t)=𝔼gc(t‾,YtT),ΘT,c(t)=𝔼hc(t‾,YtT).\Lambda_{T,c}(t) = \mathbb{E}\,g_c(\bar t, Y^T_t), \qquad \Theta_{T,c}(t) = \mathbb{E}\,h_c(\bar t, Y^T_t).(6)

The Claim

Theorem 1. Let WW be a Brownian system, γ\gamma an order parameter and TT a time in [0,1)[0,1). Let aa, bb and cc be real numbers with 0≤a0 \le a and b≤Tb \le T such that γ(s)=c\gamma(s) = c for every ss with a<s<ba < s < b. Then at every tt in (a,b)(a,b) the second moment of curvature mm is differentiable, with m′(t)=ΛT,c(t),m'(t) = \Lambda_{T,c}(t),(7)✦ and its derivative m′m' is differentiable at tt, with m″(t)=ΘT,c(t).m''(t) = \Theta_{T,c}(t).(8)✦ For every real number cc and every time t≤Tt \le T, the strip terms are expectations of polynomials in the jets along XX: ΛT,c(t)=𝔼gc(t,Xt),\Lambda_{T,c}(t) = \mathbb{E}\,g_c(t,X_t),(9)✦ ΘT,c(t)=𝔼hc(t,Xt).\Theta_{T,c}(t) = \mathbb{E}\,h_c(t,X_t).(10)✦

In the paper, the result reads: if γ=c\gamma = c on an open interval I⊂(0,1)I \subset (0,1) and c>0c > 0, then within II, (q′)′=c−2𝔼[w2−2v3](q')' = c^{-2}\,\mathbb{E}[w^2 - 2v^3] and (q′)″=c−2𝔼[z2−12vw2+6v4](q')'' = c^{-2}\,\mathbb{E}[z^2 - 12vw^2 + 6v^4], with v=cΦγ,xxv = c\,\Phi_{\gamma,xx}, w=cΦγ,xxxw = c\,\Phi_{\gamma,xxx} and z=cΦγ,xxxxz = c\,\Phi_{\gamma,xxxx} along XX, and if c=0c = 0, q′q' is strictly increasing on II.

What Is Checked

The claim is stated over the second moment of curvature mm of the diffusion XX, built from the value function Φγ\Phi_\gamma of an order parameter γ\gamma on a Brownian system WW; over the jets of Φγ\Phi_\gamma along XX; and over the strip terms ΛT,c\Lambda_{T,c} and ΘT,c\Theta_{T,c}, the expected values of the second and third curvature fields along the diffusion YTY^T with drift frozen after a horizon TT in [0,1)[0,1), under the hypothesis that γ=c\gamma = c on an interval (a,b)(a,b) with 0≤a0 \le a and b≤Tb \le T.

The library also proves two facts about these constructions: every horizon TnT_n lies in [0,1)[0,1)✦, and every spatial derivative ∂xnuγ(t‾,x)\partial_x^n u_\gamma(\bar t,x) of the clamped gradient is jointly continuous in (t,x)(t,x)✦.