Birkhoff Conjecture (Near-Boundary Continuous Invariant Foliation)

Birkhoff Conjecture (Near-Boundary Continuous Invariant Foliation)

OpenAI's claim

Manuscript 1 (family 147, 2026-09-24): A merely continuous invariant foliation of the full grazing phase annulus of a smooth positively curved convex billiard is claimed to create an analytic caustic collar and hence an elliptical table. Manuscript 2 (family 147, 2026-09-24): A smooth positively curved convex billiard table is claimed to be an ellipse whenever its near-boundary region has a continuous full foliation by smooth convex caustics, with no transverse differentiability required.

Mathematical statement of the claimed result

Mathematical claim 1.

Let $\Omega\subset\mathbb R^2$ be bounded and strictly convex, with smooth regular simple closed boundary and positive curvature. In boundary arclength $s\in\mathbb R/L\mathbb Z$ and outgoing angle $\delta\in(0,\pi)$, let $T$ be the billiard reflection map on $M=(\mathbb R/L\mathbb Z)\times(0,\pi)$. Assume a $T$-invariant open annulus $U$ contains $(\mathbb R/L\mathbb Z)\times(0,\eta)$ for some $\eta>0$, and admits a homeomorphism $H:S^1\times(0,1)\to U$ for which each circle leaf is simple, closed, essential and individually $T$-invariant. OpenAI claims \[ \Omega=\{x\in\mathbb R^2:(x-c)^TQ(x-c)<1\} \] for some center $c$ and symmetric positive-definite real matrix $Q$. The assumed phase-space foliation near grazing is essential; circles are included.

Mathematical claim 2.

Let $\Omega$ be a bounded strictly convex planar domain with a smooth regular simple closed boundary of positive curvature. Assume a relatively open collar $A\subset\overline\Omega$ of the boundary and a homeomorphism $H_0:S^1\times[0,1)\to A$ whose level zero is $\partial\Omega$ and whose positive levels are smooth closed convex caustics. A caustic means a curve to which tangency of a billiard segment is preserved by reflection $v\mapsto v-2(v\cdot n)n$, with outward unit normal $n$. OpenAI claims \[ \Omega=\{x:(x-c)^TQ(x-c)<1\},\qquad Q=Q^T>0. \] The statement assumes a continuous collar of caustics and includes circles.

Outline of the argument

Manuscript 1 (family 147, 2026-09-24): Strict twist barriers and area recurrence rule out vertical leaf pairs and rational rotation plateaus, yielding unique periodic graph leaves. Stationary-mesh, Fourier, and buffered-division estimates then promote the continuous phase data to a physical analytic collar before the companion ellipse-rigidity theorem is applied.

Manuscript 2 (family 147, 2026-09-24): Periodic rational caustics give a Fourier action system whose near-identity derivative supports analytic Newton inversion. Values and parameter jets produce a joint glancing conjugacy; finite reflected-chain jet determination then controls complex continuation and rules out a first meromorphic wall, yielding global ellipse rigidity.

Formal verification

Family 147 has no Lean scope page in this snapshot; no formalization of these manuscript results is listed.

This site has not executed the Lean files or checked their dependencies. These descriptions report the scope of the linked declarations, rather than a new verification of the proofs.

Status caveat

OpenAI attributes these results to an internal model. Its release notice cautions that results lacking full formalization can contain mistakes. This site has not independently assessed the proofs; a solved label records the scope of the reported claim.

Sources

• OpenAI, Continuous Phase Foliations Create Analytic Caustic Collars, 2026-09-24; citation key OAI:Continuous-Phase-Foliations-Create-Analytic-Caustic-Collars-September-24-2026. PDF (GitHub); PDF (direct); Manuscript page and citation; Latest version; Statement source: thm:main.
• OpenAI, Rigidity of Smooth Billiards with a Continuous Caustic Collar, 2026-09-24; citation key OAI:Rigidity-of-Smooth-Billiards-with-a-Continuous-Caustic-Collar-September-24-2026. PDF (GitHub); PDF (direct); Manuscript page and citation; Latest version; Statement source: thm:main.
• OpenAI catalogue entry.

Birkhoff Conjecture (Near-Boundary Continuous Invariant Foliation)

Definitions and mathematical statement

Let $\Omega\subset{\mathbb{R}}^2$ be a bounded billiard table with a smooth simple closed boundary of strictly positive curvature and perimeter $L$. Its billiard map $T$ sends one reflection to the next on the phase cylinder $({\mathbb{R}}/L{\mathbb{Z}})\times(0,\pi)$, where $s$ is arclength and $\varphi$ the angle with the positively oriented tangent; reflection obeys equality of incidence and reflection angles. Assume there is an open annular region $U$ adjacent to $\varphi=0$, containing $\{0<\varphi<\eta\}$ for some $\eta>0$, which is continuously foliated by $T$-invariant essential simple closed curves, each winding once around the cylinder. The selected Birkhoff assertion is \[ \text{such a near-boundary invariant foliation exists} \quad\Longrightarrow\quad\partial\Omega\text{ is an ellipse}. \] Circles count as ellipses. Central symmetry, proximity to an ellipse, and a foliation of the whole phase cylinder are not additional hypotheses.

Formulation and concept references: [S1].

Short English statement

If all trajectories sufficiently close to grazing a smooth strictly convex billiard table form a continuous invariant foliation, must the table be an ellipse?

Research attempt

The regularity and extent of the foliation matter. The requested foliation is continuous and occupies a whole annulus near grazing. Theorem 1.1 of [S1] assumes central symmetry and an invariant curve of four-periodic orbits, together with a foliation of the region between that curve and grazing. Its discussion explicitly asks about smaller regions and relaxing symmetry. The stated target contains neither its symmetry hypothesis nor its distinguished rotation-$1/4$ curve. In particular, an annulus $0<\varphi<\eta$ can be too thin to contain such a curve. We do not infer those extra properties from continuity of the given foliation.

This attempt verifies the full near-boundary foliation for ellipses by a direct first integral, proves rigidity under the stronger assumption that the angle itself is invariant, and computes why the leading grazing asymptotics cannot force an arbitrary strictly convex curve to be an ellipse.

The circle and the preserved two-form. For a circle of radius $R$, elementary chord geometry gives \[ T(s,\varphi)=(s+2R\varphi,\varphi)\pmod{2\pi R}. \] Thus every horizontal circle is invariant, and its rotation number is $\varphi/\pi$. This checks the angle convention: near $\varphi=0$ the next reflection is near the current one in the positive boundary direction.

For a general arclength parametrization $r(s)$, put $\mathcal L(s,s')=|r(s')-r(s)|$. Differentiating a chord's length gives \[ d\mathcal L=-\cos\varphi\,ds+\cos\varphi'\,ds'. \] Taking an exterior derivative shows that the billiard map preserves $\sin\varphi\,ds\wedge d\varphi$. The mixed second derivative is \[ \partial_s\partial_{s'}\mathcal L =\frac{\sin\varphi\sin\varphi'}{\mathcal L}>0. \] For a proof of the last formula, differentiate the unit chord direction: only the component of the endpoint tangent perpendicular to the chord survives; its length is $\sin\varphi'$, while projection against the initial tangent supplies $\sin\varphi$. Strict convexity gives the positive sign in the stated orientation. These identities establish the twist geometry but do not establish integrability.

A direct invariant for every ellipse. Write the ellipse as $q^TD^{-1}q=1$, where $D=\operatorname{diag}(a^2,b^2)$ and $a,b>0$. At a boundary point $q$, let $v$ be the outgoing inward-pointing unit velocity. Define \[ A=q^TD^{-1}v,\qquad I(q,v)=A^2. \] Along the straight flight to $q'=q+tv$, the boundary equation gives \[ 0=2tA+t^2v^TD^{-1}v, \qquad t=-\frac{2A}{v^TD^{-1}v}>0. \] Consequently $q'^TD^{-1}v=A+t\,v^TD^{-1}v=-A$. At the second boundary point an outward normal vector is $n'=D^{-1}q'$. Specular reflection changes the velocity to \[ v'=v-2\frac{v\cdot n'}{|n'|^2}n', \] and hence $q'^TD^{-1}v'=v'\cdot n'=-v\cdot n'=A$. Therefore $I(q',v')=I(q,v)$ exactly. No near-circular assumption or perturbation expansion was used.

In the arclength-angle coordinates, the outgoing inward normal component is $-\sin\varphi$, so \[ I(s,\varphi)=|D^{-1}r(s)|^2\sin^2\varphi. \] Let $n(s)=|D^{-1}r(s)|$, a positive smooth periodic function. For every sufficiently small $c>0$, the branch of the level set near grazing is the smooth essential circle \[ \varphi_c(s)=\arcsin\frac{\sqrt c}{n(s)}. \] These disjoint graphs vary smoothly with $c$, and fill an open annulus adjacent to $\varphi=0$. Uniform near-grazing continuity of the billiard map ensures that for small $c$ the image stays on the branch near zero, rather than the other branch near $\pi$; invariance of $I$ then makes each graph invariant. The same holds for the inverse map. Compactness supplies an $\eta>0$ such that all $0<\varphi<\eta$ lie in this annulus. This verifies precisely the hypothesis for ellipses, including noncircular ellipses.

The computation also shows why imposing constant $\varphi$ would change the problem: for a noncircular ellipse the invariant angle varies with $s$ through $n(s)$. A proposed proof that all invariant leaves must be horizontal would incorrectly exclude these intended examples.

Grazing expansion and a stronger rigidity subcase. Let $\kappa(s)>0$ be curvature, with inward unit normal $N(s)$ and positive unit tangent $T(s)$. For a short positive arclength increment $\tau$, Frenet's equations give \[ r(s+\tau)-r(s) =\left(\tau-\frac{\kappa^2\tau^3}{6}\right)T +\left(\frac{\kappa\tau^2}{2} +\frac{\kappa'\tau^3}{6}\right)N+O(\tau^4). \] Taking the angle of this chord with $T(s)$ yields \[ \varphi=\frac{\kappa\tau}{2}+\frac{\kappa'\tau^2}{6} +O(\tau^3). \] The tangent turns by $\kappa\tau+\kappa'\tau^2/2+O(\tau^3)$ along the arc. The incoming chord makes angle $\varphi'$ with the new tangent; reflection preserves that magnitude. Subtraction gives \[ \varphi'=\frac{\kappa\tau}{2}+\frac{\kappa'\tau^2}{3} +O(\tau^3), \] and consequently \[ s'-s=\frac{2\varphi}{\kappa(s)}+O(\varphi^2),\qquad \varphi'-\varphi= \frac{2\kappa'(s)}{3\kappa(s)^2}\varphi^2+O(\varphi^3). \] The errors are uniform in $s$ because the boundary is smooth and compact and curvature is bounded below.

Suppose the stronger condition $\varphi' =\varphi$ held for every sufficiently small positive $\varphi$ at every $s$. Divide the last identity by $\varphi^2$ and let $\varphi\downarrow0$. It follows that $\kappa'(s)=0$. A closed plane curve with positive constant curvature is a circle: integrate $T'(s)=\kappa N(s)$ to obtain a uniformly rotating tangent and then integrate $r'=T$. Thus exact conservation of the ordinary angle near grazing implies the circular subcase. The supplied foliation does not assert this stronger angle conservation, as the ellipse example has just demonstrated.

Why the first asymptotic invariant gives no classification. Set $J(s,\varphi)=\kappa(s)^{-1/3}\varphi$. The preceding two expansions show \[ J(s',\varphi')-J(s,\varphi)=O(\varphi^3) \] for every smooth strictly convex boundary. Indeed the contribution from the angle increment is $2\kappa'\varphi^2/(3\kappa^{7/3})$, while the change in $\kappa^{-1/3}$ times $\varphi$ contributes its negative. The cancellation is exact at order two. Thus an almost conserved grazing coordinate is generic at this order; identifying it cannot prove the exact foliation or rigidity. Summing a small one-step error over an unbounded number of reflections also requires a separate uniform argument.

Even the presence of many invariant caustics does not automatically fill an annulus. A collection parametrized by a Cantor set may have positive measure and accumulate at the boundary while leaving open gaps. The problem excludes such gaps by requiring a continuous foliation of an entire annular region. One cannot replace that hypothesis by density, positive measure, or a sequence of invariant curves tending to grazing.

Rational leaves and the regularity trap. An orientation-preserving circle homeomorphism with rational rotation number has a periodic orbit, but need not have every point periodic. For example, the time-one map of a smooth vector field on the circle with isolated zeros has fixed points and nonconstant orbits between them; its rotation number is zero. Passing to a suitable finite cyclic configuration gives the same distinction for other rational rotation numbers. An invariant smooth positive density on the leaf would instead conjugate the map, through its cumulative distribution, to a rigid rotation, making rational leaves periodic. The source uses an absolutely continuous invariant measure in its $C^1$-foliation discussion. A merely continuous ambient foliation does not justify differentiating its leaf parameter or automatically constructing that density. The smooth area form on the two-dimensional annulus alone is not a specified smooth conditional density on every individual leaf.

Verification and outcome. The diagnostic checks straight flight, reflection, and preservation of $I$ numerically for specified noncircular ellipses and checks the cancellation coefficients in the grazing expansion with exact rational arithmetic. It does not infer an invariant foliation from a finite orbit sample. The attempt proves the ellipse implication in the easy direction, an exact stronger circle-rigidity subcase, and the universal leading-order cancellation. The missing step is to extract a global geometric constraint from the exact continuous foliation on an arbitrarily thin annulus without adding symmetry, regularity, or a distinguished four-periodic leaf.

Sources

• S1 Bialy and Mironov, The Birkhoff–Poritsky conjecture for centrally-symmetric billiard tables, introduction and Section 2; continuous invariant-curve near-boundary generalization. https://arxiv.org/pdf/2008.03566. Formulation source.