OctoInformBeta

OpenAI preprint, September 2026 · §5. Uniform control of returning geodesics

Lemma 5.3 [Turning duality, with a direct-estimate replacement]: Direct Cauchy estimates with a boundary penalty and scalar turning coefficients

Endpoint control with a return duration bound.

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

A field along a returning geodesic has a transverse component zz and a scalar component ww. For return duration LL, the endpoint estimate controls the transverse displacement X=z(L)−z(0)X=z(L)-z(0) by the cometric coordinate energy of the field. Here ss denotes geodesic time, vv the height velocity, and pp the conserved transverse momentum; ee is a fixed transverse covector.

Transverse vectors belong to ℝ2\mathbb R^2, and transverse matrices belong to ℝ2×2\mathbb R^{2\times2}. For a transverse matrix AA and transverse vectors u,r∈ℝ2u,r\in\mathbb R^2, the matrix bilinear pairing B(A;⋅,⋅)B(A;\cdot,\cdot) is the bilinear form on ℝ2\mathbb R^2 defined by (1): B(A;u,r)=u⊤Ar.B(A;u,r)=u^\top Ar.(1)✦ Let P(s)P(s) be the transverse cometric and Q(s)=P(s)−1Q(s)=P(s)^{-1} its metric matrix. For a transverse matrix path QQ, transverse vector paths kk and z′z', and real functions hh, ww and w′w', the completed-square energy density ε\varepsilon is the real function in (2), whose integral is the cometric coordinate energy EE: ε(s)=B(Q(s);z′(s)−w(s)k(s),z′(s)−w(s)k(s))+w′(s)2+h(s)w(s)2,E=∫0Lε(s)ds.\begin{aligned} \varepsilon(s) &=B\bigl(Q(s);z'(s)-w(s)k(s),z'(s)-w(s)k(s)\bigr) +w'(s)^2+h(s)w(s)^2,\\ E&=\int_0^L\varepsilon(s)\,ds. \end{aligned}(2)✦

The turning constants a0a_0 and β\beta are the real numbers defined in (3): a0=exp⁡(−3/2),β=1+exp⁡(−9/26)2.a_0=\exp(-3/2),\qquad \beta=\frac{1+\exp(-9/26)}{2}.(3) For real coefficients Ce,CmC_e,C_m, the smallness condition on the neck parameter δ\delta consists of the inequalities in (4): 0<δ≤1,δ≤12Cm,δ≤1−β2Cm+Cea0.0<\delta\le1,\qquad \delta\le\frac{1}{2C_m},\qquad \delta\le\frac{1-\beta}{2C_m+C_ea_0}.(4)✦

The Claim

Theorem 1. Let L,K≥0L,K\ge0. Suppose that P,Q,k,h,v,z′,w,w′P,Q,k,h,v,z',w,w' are continuous on [0,L][0,L] and that z,v,wz,v,w have derivatives z′,v′,w′z',v',w' at every point of this interval. Assume throughout the interval that PP and QQ are symmetric, QP=1QP=1, and QQ is positive semidefinite. Suppose also that h≥0h\ge0, (e⊤k)2≤Kh(e^\top k)^2\le Kh, p⊤k=−2v′p^\top k=-2v', and w(0)=w(L)=0w(0)=w(L)=0. The direct endpoint Cauchy estimate (5) holds for every real α\alpha: |(p−αe)⊤X|2≤(∫0L[B(P(s);p−αe,p−αe)+4v(s)2+α2K]ds)E.\bigl|(p-\alpha e)^\top X\bigr|^2 \le \left(\int_0^L \bigl[B\bigl(P(s);p-\alpha e,p-\alpha e\bigr) +4v(s)^2+\alpha^2K\bigr]\,ds\right)E.(5)✦

Let c,E≥0c,E\ge0, μ>0\mu>0, and f,y,α∈ℝf,y,\alpha\in\mathbb R, and suppose that f2≤cEf^2\le cE. The scalar boundary Cauchy inequality is (6): (f+αy)2≤(c+α2μ−1)(E+μy2).(f+\alpha y)^2 \le\bigl(c+\alpha^2\mu^{-1}\bigr)\bigl(E+\mu y^2\bigr).(6)✦

For real coefficients spp,spes_{pp},s_{pe} and a positive real denominator DD, the quadratic minimum bound (7) holds for every real α\alpha: spp−spe2D≤spp−2αspe+α2D.s_{pp}-\frac{s_{pe}^2}{D} \le s_{pp}-2\alpha s_{pe}+\alpha^2D.(7)✦

Let Ce,Cm>0C_e,C_m>0, let δ\delta satisfy (4), and let L,v0,μ>0L,v_0,\mu>0. Suppose that the real scalars p1,κ,spp,spe,seep_1,\kappa,s_{pp},s_{pe},s_{ee} satisfy (8): κ=δa0,μ=κv04,v02≤κp12,4v0≤βκp12L,spp≤L(1+v02),0≤see≤CeL,|spe−Lp1|≤CmδL|p1|.\begin{aligned} \kappa&=\delta a_0,& \mu&=\frac{\kappa v_0}{4},\\ v_0^2&\le\kappa p_1^2,& 4v_0&\le\beta\kappa p_1^2L,\\ s_{pp}&\le L(1+v_0^2),& 0\le s_{ee}&\le C_eL,\\ \bigl|s_{pe}-Lp_1\bigr|&\le C_m\delta L|p_1|. \end{aligned}(8) There is a real α\alpha satisfying the returning quadratic bound (9): spp−2αspe+α2(see+μ−1)≤L.s_{pp}-2\alpha s_{pe}+\alpha^2\bigl(s_{ee}+\mu^{-1}\bigr)\le L.(9)✦

In the paper, the result reads: with SS the Hilbert dual form, every constant covector qq and every gauged field satisfy |q⊤X|2≤S(q,q)E|q^\top X|^2\le S(q,q)E, and for every real α\alpha, |p⊤X|2≤(S(p−αe,p−αe)+α2/μ)(E+μ(e⊤X)2)|p^\top X|^2\le\bigl(S(p-\alpha e,p-\alpha e)+\alpha^2/\mu\bigr) \bigl(E+\mu(e^\top X)^2\bigr), whose smallest first factor is S(p,p)−S(p,e)2/(S(e,e)+μ−1)S(p,p)-S(p,e)^2/(S(e,e)+\mu^{-1}), attained at α=S(p,e)/(S(e,e)+μ−1)\alpha=S(p,e)/(S(e,e)+\mu^{-1}).

What Is Checked

The endpoint estimate (5) is stated over continuous real transverse matrix paths and differentiable fields with zero scalar endpoints, while the scalar inequalities (6), (7) and (9) are stated over real numbers subject to their respective hypotheses.