OctoInformBeta

OpenAI preprint, September 2026 · §5. Strict curvature on constant intervals

Proposition 5.1: m″>0m''>0 where γ=c>0\gamma=c>0 and m′>0m'>0 where γ=0\gamma=0, for m(t)=𝔼Φγ,xx(t,Xt)2m(t)=\mathbb{E}\,\Phi_{\gamma,xx}(t,X_t)^2, with a quartic lower bound on the strip third moment

Constant γ\gamma makes the curvature moment strictly convex or increasing.

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

In the zero-temperature Parisi formula for the Sherrington-Kirkpatrick model, with covariance ξ(t)=t2/2\xi(t) = t^2/2, an order parameter γ\gamma on [0,1)[0,1) determines the value function Φγ(t,x)\Phi_\gamma(t,x) of a stochastic control problem and a diffusion XX with drift γ(t)Φγ,x(t,Xt)\gamma(t)\,\Phi_{\gamma,x}(t,X_t). The question is how the second moment of the curvature Φγ,xx(t,Xt)\Phi_{\gamma,xx}(t,X_t) moves with tt on a stretch of times where γ\gamma equals a constant cc: whether it is strictly convex there when c>0c > 0, and strictly increasing when c=0c = 0. For c>0c > 0 one also wants a quantitative lower bound in terms of the fourth moment 𝔼v4\mathbb{E}\,v^4 of v=cΦγ,xxv = c\,\Phi_{\gamma,xx}.

The Value Function

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

A Brownian system WW on a measurable space Ω\Omega is a probability measure ℙ\mathbb{P} on Ω\Omega together with a real process B=(Bt)t≥0B = (B_t)_{t \ge 0}, measurable at each time, that is a standard Brownian motion under ℙ\mathbb{P}; 𝔼\mathbb{E} is the expectation under ℙ\mathbb{P}.✦

A bounded control α\alpha from time tt is a real process (αs)s≥0(\alpha_s)_{s \ge 0} that is progressively measurable for the increment filtration after tt, the natural filtration of s↦Bt+s−Bts \mapsto B_{t+s} - B_t, and satisfies |αs|≤1|\alpha_s| \le 1 at every time and outcome.✦

The value function Φγ\Phi_\gamma of γ\gamma is the function of time tt and position xx whose value is the supremum, over bounded controls α\alpha from tt, of the expected control payoff, the absolute value of the terminal point minus the running cost: Φγ(t,x)=supα𝔼[|x+B1−Bt+∫t1γ(s)αs−tds|−12∫t1γ(s)αs−t2ds].\Phi_\gamma(t,x) = \sup_{\alpha}\,\mathbb{E}\Big[\Big|x + B_1 - B_t + \int_t^1 \gamma(s)\,\alpha_{s-t}\,ds\Big| - \frac12\int_t^1 \gamma(s)\,\alpha_{s-t}^2\,ds\Big].(1)✦

The value gradient is uγ=Φγ,xu_\gamma = \Phi_{\gamma,x}, the derivative of Φγ(t,⋅)\Phi_\gamma(t,\cdot) in the position✦, and the value curvature is Φγ,xx\Phi_{\gamma,xx}, the derivative of uγ(t,⋅)u_\gamma(t,\cdot).✦ The value rate of γ\gamma is the function of (t,x)(t,x) given by −12Φγ,xx(t,x)−12γ(t)Φγ,x(t,x)2-\tfrac12\Phi_{\gamma,xx}(t,x) - \tfrac12\gamma(t)\,\Phi_{\gamma,x}(t,x)^2, the right side of the Parisi equation ∂tΦγ=−12(Φγ,xx+γΦγ,x2)\partial_t\Phi_\gamma = -\tfrac12\big(\Phi_{\gamma,xx} + \gamma\,\Phi_{\gamma,x}^2\big).✦

The Diffusion

A bounded Lipschitz drift is a jointly measurable function b(t,x)b(t,x) for which there are constants M,L≥0M, L \ge 0 with |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.✦ Given such a drift bb and a process ZZ, the solution of the drift equation Xt=Zt+∫0tb(s,Xs)dsX_t = Z_t + \int_0^t b(s,X_s)\,ds is the process Z+YZ + Y, where the drift correction Y=lim⁡nYnY = \lim_n Y^n is the limit of the correction iterates Y0=0Y^0 = 0, Ytn+1=∫0tb(s,Zs+Ysn)dsY^{n+1}_t = \int_0^t b(s, Z_s + Y^n_s)\,ds.

The driving Brownian motion of WW is the process Bt∘=B‾t−B‾0B^\circ_t = \bar B_t - \bar B_0, where B‾t=lim⁡n→∞B⌊(n+1)t⌋/(n+1)\bar B_t = \lim_{n\to\infty} B_{\lfloor (n+1)t\rfloor/(n+1)} is the progressive version of BB, the limit of BB along the left mesh times ⌊(n+1)t⌋/(n+1)\lfloor (n+1)t\rfloor/(n+1).✦

For a horizon TT, the clamp to the strip [0,T][0,T] sends a real tt to t‾=max⁡(0,min⁡(T,t))\bar t = \max(0,\min(T,t)).✦ For T∈[0,1)T \in [0,1), the strip drift of horizon TT is the bounded Lipschitz drift (t,x)↦γ(t‾)uγ(t‾,x)(t,x) \mapsto \gamma(\bar t)\,u_\gamma(\bar t,x) with bound γ(T)\gamma(T); the strip diffusion XTX^T is the solution of its drift equation driven by B∘B^\circ, XtT=Bt∘+∫0tγ(s‾)uγ(s‾,XsT)ds.X^T_t = B^\circ_t + \int_0^t \gamma(\bar s)\,u_\gamma(\bar s, X^T_s)\,ds.(2)✦

For a natural number nn, the localized diffusion XnX^n is the strip diffusion XTnX^{T_n} of the horizon Tn=1−1/(n+2)T_n = 1 - 1/(n+2), and the raw diffusion limit is Xˆt=lim⁡n→∞Xtn\hat X_t = \lim_{n\to\infty} X^n_t.✦ The diffusion XX is the process Xt=Bt∘+∫0tγ(s)uγ(s,Xˆs)ds,X_t = B^\circ_t + \int_0^t \gamma(s)\,u_\gamma(s,\hat X_s)\,ds,(3)✦ in which the raw feedback uγ(s,Xˆs)u_\gamma(s,\hat X_s) is the gradient cut at time 11, read along Xˆ\hat X, and is 00 for s≥1s \ge 1.

For t∈[0,1)t \in [0,1), the jet of order kk along the diffusion is the random variable ∂xkΦγ(t,Xt)\partial_x^k\Phi_\gamma(t,X_t).✦ The second moment of curvature, the curvature moment mm, is the function of time m(t)=𝔼Φγ,xx(t,Xt)2.m(t) = \mathbb{E}\,\Phi_{\gamma,xx}(t,X_t)^2.(4)✦

The Strip Third Moment

A bounded smooth function is a C∞C^\infty function gg on ℝ\mathbb{R} all of whose derivatives g(n)g^{(n)} are bounded.✦ A bounded smooth family is a jointly measurable function F(t,x)F(t,x) such that each section F(t,⋅)F(t,\cdot) is a bounded smooth function and each ∂xnF\partial_x^n F is bounded uniformly in (t,x)(t,x).✦ 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).✦ The sum and the product of two time fields (f,d)(f,d) and (f′,d′)(f',d') with horizon TT are the time fields (f+f′,d+d′)(f + f', d + d') and (ff′,df′+fd′)(ff', df' + fd'), and the scaled time field c(f,d)c(f,d) is (cf,cd)(cf, cd).✦

For 0≤T<10 \le T < 1 and an order nn, the clamped jet field of order nn 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 squared curvature field is the product of the clamped jet field of order 11 with itself.✦ For a constant cc, the third curvature field is the time field with horizon TT formed from the clamped jet fields of orders 11, 22, 33 and the squared curvature field by sums, products and scalar multiples, with function Kc(t,x)=(Φγ,xxxx2−12cΦγ,xxΦγ,xxx2+6c2Φγ,xx4)(t‾,x).K_c(t,x) = \big(\Phi_{\gamma,xxxx}^2 - 12c\,\Phi_{\gamma,xx}\Phi_{\gamma,xxx}^2 + 6c^2\,\Phi_{\gamma,xx}^4\big)(\bar t,x).(5)✦

For a time field with function ff and a bounded Lipschitz drift bb, the expected field value is the function t↦𝔼f(t,Xtb)t \mapsto \mathbb{E}\,f(t,X^b_t), where XbX^b is the solution of the drift equation of bb driven by B∘B^\circ.✦ The strip third moment with horizon T∈[0,1)T \in [0,1) and constant cc is the expected field value of the third curvature field along the strip drift, ΘT,c(t)=𝔼Kc(t,XtT).\Theta_{T,c}(t) = \mathbb{E}\,K_c(t,X^T_t).(6)✦ With v=cΦγ,xxv = c\,\Phi_{\gamma,xx}, w=cΦγ,xxxw = c\,\Phi_{\gamma,xxx} and z=cΦγ,xxxxz = c\,\Phi_{\gamma,xxxx}, one has c2Kc=z2−12vw2+6v4c^2K_c = z^2 - 12vw^2 + 6v^4.

The Claim

Theorem 1. Let WW be a Brownian system on Ω\Omega, γ\gamma an order parameter, and mm the second moment of curvature (4).

(i) Let 0≤a<b<10 \le a < b < 1 and c>0c > 0. If γ(s)=c\gamma(s) = c for every s∈(a,b)s \in (a,b), then m″(t)>0for every t∈(a,b).m''(t) > 0 \qquad \text{for every } t \in (a,b).(7)✦

(ii) Let 0≤a<b<10 \le a < b < 1. If γ(s)=0\gamma(s) = 0 for every s∈(a,b)s \in (a,b), then m′(t)>0for every t∈(a,b).m'(t) > 0 \qquad \text{for every } t \in (a,b).(8)✦

(iii) Let t,T∈[0,1)t, T \in [0,1) with 0<t≤T0 < t \le T, suppose γ\gamma is continuous at tt, and put c=γ(t)c = \gamma(t). Then 11000c4𝔼Φγ,xx(t,Xt)4≤c2ΘT,c(t).\frac{1}{1000}\,c^4\,\mathbb{E}\,\Phi_{\gamma,xx}(t,X_t)^4 \le c^2\,\Theta_{T,c}(t).(9)✦

In the paper, the curvature moment mm is written q′q', and the result reads: where γ=c>0\gamma = c > 0 on an open interval, q′q' is strictly convex there, with q‴≥𝔼[v4]/(1000c2)>0q''' \ge \mathbb{E}[v^4]/(1000c^2) > 0, and where γ=0\gamma = 0, q′q' is strictly increasing.

What Is Checked

The claim is stated over an order parameter γ\gamma and a Brownian system WW: the curvature moment (4) of the value function (1) along the diffusion (3), and the strip third moment (6) along the strip diffusion (2).

The library also shows that each horizon Tn=1−1/(n+2)T_n = 1 - 1/(n+2) lies in [0,1)[0,1)✦, and that for 0≤T<10 \le T < 1 every spatial derivative ∂xkuγ(t‾,x)\partial_x^k u_\gamma(\bar t,x) of the clamped gradient is jointly continuous in (t,x)(t,x).✦