OctoInformBeta

OpenAI preprint, September 2026 · §1. The variational problem and the main theorem

Theorem 1.1: The Stieltjes measure of a minimizer γ\gamma has support [0,1)[0,1), and along every diffusion the second moment of the gradient is tt

Minimizers of the zero-temperature Parisi functional have full support.

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

At zero temperature, the ground state energy of the model of Sherrington and Kirkpatrick, with covariance ξ(t)=t2/2\xi(t) = t^2/2, is the infimum of the Parisi functional 𝒫\mathcal{P} over order parameters γ\gamma on [0,1)[0,1); this is the Parisi formula of Auffinger and Chen. Attached to γ\gamma are the value function Φγ\Phi_\gamma of a stochastic control problem, with gradient uγ=Φγ,xu_\gamma = \Phi_{\gamma,x}, and the diffusion XX with drift γuγ\gamma u_\gamma. The question is the shape of a minimizer: whether its Stieltjes measure μγ\mu_\gamma can leave an interval of [0,1)[0,1) uncharged, so that γ\gamma is constant there.

Order Parameters

An order parameter γ\gamma on the time interval [0,1)[0,1) is a nonnegative, nondecreasing, right-continuous function on [0,1)[0,1) whose extension by zero to the real line is Lebesgue integrable.✦ A Stieltjes measure μγ\mu_\gamma of an order parameter γ\gamma is a measure on [0,1)[0,1) with μγ([0,t])=γ(t)\mu_\gamma([0,t]) = \gamma(t) for every t∈[0,1)t \in [0,1).✦

The Value Function and the Parisi Functional

A Brownian system on a measurable space Ω\Omega is a probability measure ℙ\mathbb{P} on Ω\Omega together with a real process (Bt)t≥0(B_t)_{t \ge 0}, each BtB_t measurable, that is a standard Brownian motion under ℙ\mathbb{P}.✦ Fix one, and write 𝔼\mathbb{E} for the expectation under ℙ\mathbb{P}.

For a time t≥0t \ge 0, the increment filtration after tt is the filtration (ℱst)s≥0(\mathcal{F}^t_s)_{s \ge 0} in which ℱst\mathcal{F}^t_s is 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 (αs)s≥0(\alpha_s)_{s \ge 0} on Ω\Omega, progressively measurable for this filtration, with |αs|≤1|\alpha_s| \le 1 at every time ss and every outcome.✦ The control payoff of a bounded control α\alpha from tt, started at a point x∈ℝx \in \mathbb{R}, is Jt,x(α)=|x+B1−Bt+∫t1γ(s)αs−tds|−12∫t1γ(s)αs−t2ds.J_{t,x}(\alpha) = \Bigl| x + B_1 - B_t + \int_t^1 \gamma(s)\,\alpha_{s-t}\,ds \Bigr| - \frac{1}{2} \int_t^1 \gamma(s)\,\alpha_{s-t}^2\,ds .(1)✦

The value function of γ\gamma at a time tt and a point xx is the supremum of the expected payoff (1) over bounded controls α\alpha from tt: Φγ(t,x)=supα𝔼[Jt,x(α)].\Phi_\gamma(t,x) = \sup_\alpha \mathbb{E}\bigl[J_{t,x}(\alpha)\bigr].(2)✦ The curvature Φγ,xx=∂xuγ\Phi_{\gamma,xx} = \partial_x u_\gamma is the second derivative of Φγ\Phi_\gamma in the position variable.✦

The Parisi functional at an order parameter γ\gamma is 𝒫(γ)=Φγ(0,0)−12∫01tγ(t)dt.\mathcal{P}(\gamma) = \Phi_\gamma(0,0) - \frac{1}{2} \int_0^1 t\,\gamma(t)\,dt .(3)✦ An order parameter γ\gamma is a Parisi minimizer when 𝒫(γ)≤𝒫(η)\mathcal{P}(\gamma) \le \mathcal{P}(\eta) for every order parameter η\eta.✦

The Diffusion

A controlled diffusion of an order parameter γ\gamma is a real process (Xt)t≥0(X_t)_{t \ge 0} on Ω\Omega, progressively measurable for the natural filtration of BB, whose paths are almost surely continuous on [0,1][0,1], start at X0=0X_0 = 0 and satisfy Xt=Bt+∫0tγ(s)uγ(s,Xs)ds(0≤t≤1).X_t = B_t + \int_0^t \gamma(s)\, u_\gamma(s, X_s)\,ds \qquad (0 \le t \le 1).(4)✦ For an order parameter γ\gamma and a real process XX on Ω\Omega, the second moment of the gradient uγu_\gamma along XX is the function qq on [0,1)[0,1) given by q(t)=𝔼[uγ(t,Xt)2]q(t) = \mathbb{E}\bigl[u_\gamma(t,X_t)^2\bigr].✦

The Claim

Theorem 1. Let γ\gamma be an order parameter that is a Parisi minimizer. Then γ\gamma has the full support properties: it has a Stieltjes measure and a controlled diffusion, every Stieltjes measure μγ\mu_\gamma of γ\gamma has support [0,1)[0,1), γ\gamma is strictly increasing on [0,1)[0,1), and every controlled diffusion XX of γ\gamma satisfies, for 0≤t<10 \le t < 1, q(t)=tand𝔼[Φγ,xx(t,Xt)2]=1.q(t) = t \qquad \text{and} \qquad \mathbb{E}\bigl[\Phi_{\gamma,xx}(t,X_t)^2\bigr] = 1 .(5)✦

In the paper, the result reads: if γ\gamma minimizes 𝒫\mathcal{P}, then the support of μγ\mu_\gamma is all of [0,1)[0,1), γ\gamma is strictly increasing, and 𝔼[uγ(t,Xt)2]=t\mathbb{E}[u_\gamma(t,X_t)^2] = t and 𝔼[Φγ,xx(t,Xt)2]=1\mathbb{E}[\Phi_{\gamma,xx}(t,X_t)^2] = 1 for 0≤t<10 \le t < 1.

What Is Checked

The claim (5) is stated over a Brownian system, the order parameters on [0,1)[0,1) and their Stieltjes measures, the Parisi functional (3) built on the value function (2) of bounded controls, and the controlled diffusions (4), along which it reads the gradient uγu_\gamma and the curvature Φγ,xx\Phi_{\gamma,xx}.