OctoInformBeta

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

A gap contradicts contact: For q(a)=aq(a)=a and q(b)=bq(b)=b with a<ba<b, a strictly convex q′q' forces q′(a)>1q'(a)>1 or q′(b)>1q'(b)>1, and q″>0q''>0 forces q′(b)>1q'(b)>1

Conditions on q′q' that exclude a gap between contact points.

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, a function on [0,1)[0,1) with Stieltjes measure μγ\mu_\gamma, μγ([0,t])=γ(t)\mu_\gamma([0,t])=\gamma(t). Let Φγ\Phi_\gamma be its value function, uγ=Φγ,xu_\gamma=\Phi_{\gamma,x}, and XX the diffusion dXt=γ(t)uγ(t,Xt)dt+dBtdX_t=\gamma(t)\,u_\gamma(t,X_t)\,dt+dB_t with X0=0X_0=0, driven by a Brownian motion BB. Set q(t)=𝔼[uγ(t,Xt)2]q(t)=\mathbb{E}[u_\gamma(t,X_t)^2]. A contact point of γ\gamma is a time t∈[0,1)t\in[0,1) with q(t)=tq(t)=t, where the graph of qq meets the diagonal. A gap of γ\gamma is an interval (a,b)(a,b) with a<ba<b to which μγ\mu_\gamma gives no mass; on [a,b)[a,b), γ\gamma equals a constant cc.

The question is whether a gap can have contact at both ends, q(a)=aq(a)=a and q(b)=bq(b)=b. The answer turns on the shape of q′q' on [a,b][a,b] and on how its values at aa and bb compare with 11, the slope of the diagonal. There are two cases: for c>0c>0, q‴>0q'''>0 on (a,b)(a,b), so that q′q' is strictly convex; for c=0c=0, q′q' is strictly increasing. The library states the answer for an arbitrary real function qq on ℝ\mathbb{R} in place of t↦𝔼[uγ(t,Xt)2]t\mapsto\mathbb{E}[u_\gamma(t,X_t)^2].

The Claim

Theorem 1. Let a<ba<b be real numbers, and let qq and q′q' be real functions on ℝ\mathbb{R} such that qq is continuous on [a,b][a,b], qq has derivative q′(t)q'(t) at every t∈(a,b)t\in(a,b), and q(a)=aq(a)=a, q(b)=bq(b)=b.

(i) If q′q' is strictly convex on [a,b][a,b], then max⁡{q′(a),q′(b)}>1.\max\{q'(a),q'(b)\}>1.(1)✦

(ii) If q′q' is continuous on [a,b][a,b] and q‴(t)>0q'''(t)>0 for every t∈(a,b)t\in(a,b), then max⁡{q′(a),q′(b)}>1.\max\{q'(a),q'(b)\}>1.(2)✦

(iii) If q′q' is continuous on [a,b][a,b] and q″(t)>0q''(t)>0 for every t∈(a,b)t\in(a,b), then q′(b)>1.q'(b)>1.(3)✦

Part (iii) states the case c=0c=0 as q″>0q''>0 on (a,b)(a,b).

What Is Checked

The statements (1), (2) and (3) are stated over two real functions qq and q′q' on ℝ\mathbb{R} and two reals a<ba<b, with qq continuous on [a,b][a,b], q′(t)q'(t) the derivative of qq at every t∈(a,b)t\in(a,b), q(a)=aq(a)=a and q(b)=bq(b)=b; the first assumes q′q' strictly convex on [a,b][a,b], the other two assume q′q' continuous on [a,b][a,b], with q‴>0q'''>0 and q″>0q''>0 on (a,b)(a,b) respectively.