OctoInformBeta

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

Lemma 2.1: Φγ(t,⋅)\Phi_\gamma(t,\cdot) is even, convex, 11-Lipschitz and close to |x||x|, with xx-derivatives of every order bounded for t≤T<1t \le T < 1 and L1L^1-Lipschitz dependence on γ\gamma

Regularity and stability of the value function.

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 Parisi formula at zero temperature for the Sherrington-Kirkpatrick model, with covariance ξ(t)=t2/2\xi(t) = t^2/2, the ground-state energy is the infimum over order parameters γ\gamma of the Parisi functional 𝒫(γ)=Φγ(0,0)−12∫01tγ(t)dt\mathcal{P}(\gamma) = \Phi_\gamma(0,0) - \frac12\int_0^1 t\gamma(t)\,dt. Here Φγ(t,x)\Phi_\gamma(t,x) is the value of a stochastic control problem in which a Brownian path started from xx at time tt is steered by a bounded drift and rewarded by its distance from the origin at time 11.

The terminal payoff |x||x| has a corner at the origin. The question is how regular Φγ(t,⋅)\Phi_\gamma(t,\cdot) is as a function of the position before the terminal time, and how much Φγ\Phi_\gamma moves when γ\gamma is perturbed in L1L^1.

Order Parameters and the Control Problem

The time interval is the half-open interval [0,1)[0,1).✦ An order parameter is a function γ\gamma from [0,1)[0,1) to ℝ\mathbb{R} that is nonnegative, nondecreasing, right-continuous and integrable.✦ The extension by zero of γ\gamma is the function on ℝ\mathbb{R} equal to γ\gamma on [0,1)[0,1) and to 00 elsewhere; integrals of γ\gamma over intervals of ℝ\mathbb{R} are integrals of this extension.✦ The L1L^1 distance of two order parameters β\beta and γ\gamma is ∥β−γ∥1=∫01|β(s)−γ(s)|ds.\|\beta - \gamma\|_1 = \int_0^1 |\beta(s) - \gamma(s)|\,ds.(1)✦

A Brownian system WW on a measurable space Ω\Omega is a probability measure ℙ\mathbb{P} on Ω\Omega together with a standard Brownian motion B=(Bt)t≥0B = (B_t)_{t \ge 0} under ℙ\mathbb{P}; we write 𝔼\mathbb{E} for the expectation under ℙ\mathbb{P}.✦ For a time t≥0t \ge 0, the increment filtration of WW after tt is the filtration (ℱst)s≥0(\mathcal{F}^t_s)_{s \ge 0} in which ℱst\mathcal{F}^t_s is the σ\sigma-algebra generated by the increments Bt+r−BtB_{t+r} - B_t with 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 ss 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 s≥0s \ge 0 and ω∈Ω\omega \in \Omega.

Fix a Brownian system WW and an order parameter γ\gamma. For a time tt, a position x∈ℝx \in \mathbb{R} and a bounded control α\alpha from time tt, the control payoff is the random variable Jγ(t,x;α)=|x+B1−Bt+∫t1γ(s)αs−tds|−12∫t1γ(s)αs−t2ds.J_\gamma(t,x;\alpha) = \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.(2) The value function of γ\gamma is the function Φγ\Phi_\gamma on ℝ2\mathbb{R}^2 given by Φγ(t,x)=supα𝔼[Jγ(t,x;α)],\Phi_\gamma(t,x) = \sup_\alpha \mathbb{E}\bigl[J_\gamma(t,x;\alpha)\bigr],(3)✦ the supremum running over all bounded controls α\alpha from time tt. At a negative time tt, (2) and (3) are read with BtB_t replaced by B0B_0 and with controls from time 00.

The Claim

Theorem 1. Let WW be a Brownian system, and let β\beta and γ\gamma be order parameters with value functions Φβ\Phi_\beta and Φγ\Phi_\gamma. For every time t∈[0,1]t \in [0,1], the function Φγ(t,⋅)\Phi_\gamma(t,\cdot) is even: Φγ(t,−x)=Φγ(t,x)for all x∈ℝ.\Phi_\gamma(t,-x) = \Phi_\gamma(t,x) \quad \text{for all } x \in \mathbb{R}.(4)✦ For every time t≤1t \le 1, the function Φγ(t,⋅)\Phi_\gamma(t,\cdot) is convex and 11-Lipschitz on ℝ\mathbb{R}: Φγ(t,λx+(1−λ)y)≤λΦγ(t,x)+(1−λ)Φγ(t,y)for all x,y∈ℝ,λ∈[0,1],\Phi_\gamma(t,\lambda x + (1-\lambda)y) \le \lambda\,\Phi_\gamma(t,x) + (1-\lambda)\,\Phi_\gamma(t,y) \quad \text{for all } x, y \in \mathbb{R},\ \lambda \in [0,1],(5)✦ |Φγ(t,x)−Φγ(t,y)|≤|x−y|for all x,y∈ℝ.|\Phi_\gamma(t,x) - \Phi_\gamma(t,y)| \le |x - y| \quad \text{for all } x, y \in \mathbb{R}.(6)✦ For every time t∈[0,1]t \in [0,1] and every x∈ℝx \in \mathbb{R}, |x|≤Φγ(t,x)≤|x|+2(1−t)/π+12∫t1γ(s)ds.|x| \le \Phi_\gamma(t,x) \le |x| + \sqrt{2(1-t)/\pi} + \frac12\int_t^1 \gamma(s)\,ds.(7)✦ For every TT with 0≤T<10 \le T < 1 and every order k≥1k \ge 1, there is a constant C≥0C \ge 0 such that the kk-th derivative of Φγ\Phi_\gamma in xx satisfies |∂xkΦγ(t,x)|≤Cfor all (t,x)∈[0,T]×ℝ.|\partial_x^k \Phi_\gamma(t,x)| \le C \quad \text{for all } (t,x) \in [0,T] \times \mathbb{R}.(8)✦ For every time t∈[0,1]t \in [0,1] and every x∈ℝx \in \mathbb{R}, the value functions of β\beta and γ\gamma satisfy |Φβ(t,x)−Φγ(t,x)|≤32∥β−γ∥1.|\Phi_\beta(t,x) - \Phi_\gamma(t,x)| \le \frac32\,\|\beta - \gamma\|_1.(9)✦

In the paper, the result reads: Φγ\Phi_\gamma is even, convex and 11-Lipschitz in xx, with |x|≤Φγ(t,x)≤|x|+2(1−t)/π+12∫t1γ|x| \le \Phi_\gamma(t,x) \le |x| + \sqrt{2(1-t)/\pi} + \frac12\int_t^1 \gamma; for every k≥1k \ge 1 and T<1T < 1, ∂xkΦγ\partial_x^k \Phi_\gamma is bounded and jointly continuous on [0,T]×ℝ[0,T] \times \mathbb{R}, uniformly over γ\gamma of bounded L1L^1 norm; and if γn→γ\gamma_n \to \gamma in L1(0,1)L^1(0,1), then ∂xkΦγn→∂xkΦγ\partial_x^k \Phi_{\gamma_n} \to \partial_x^k \Phi_\gamma uniformly on [0,T]×ℝ[0,T] \times \mathbb{R} for every k≥1k \ge 1 and T<1T < 1.

What Is Checked

The claim is stated over the value function (3), the supremum of the expected control payoff (2) over bounded controls progressively measurable for the increment filtration of a Brownian system, for order parameters on the time interval [0,1)[0,1) extended by zero, and over the L1L^1 distance (1).