Chaos with positive area?

The claim, in plain words

The standard map models a rotor kicked once per period: the momentum y gets a kick k sin(2πx), then the angle x advances by the new momentum, both mod 1, so the motion lives on a torus and preserves area. For small kicks smooth invariant curves wrap the torus; as the kicks grow, a chaotic sea spreads. Chaos means nearby points separate exponentially fast: a positive Lyapunov exponent. In our run at k = 5 the typical exponent is 2.756, close to the classical estimate log(πk) = 2.754.

By Pesin’s formula, positive metric entropy for area means exponential separation on a set of positive area, not just somewhere. As the paper puts it, a hyperbolic set may have zero area, and even full Hausdorff dimension does not imply positive area. As it recounts, Duarte exhibited elliptic islands and tangencies; Gorodetski found transitive sets of full dimension for a residual set of large parameters, but residual sets of parameters need not have positive measure; Oliveira proved positive topological entropy and Obata uniqueness of the measure of maximal entropy, neither of which gives entropy for area; perturbed, random or coupled versions change the system. Islands keep coming back: elliptic fixed points appear for kicks just above every positive whole number, and in our run near k = 5 each island covers about 1/6,000 of the torus. Sinai’s conjecture, in the formulation of Berger and Turaev, asks for positive metric entropy on a set of parameters of positive measure.

Result 146 is one OpenAI preprint (45 pages, September 23, 2026). Theorem 1.1 claims a threshold k₀ such that for every k ≥ k₀ the standard map has positive metric entropy with respect to area; equivalently, the largest Lyapunov exponent is positive on a set of positive area. So the good parameters contain a whole tail, and in particular an interval of positive length; the paper “resolves affirmatively Sinai’s positive metric entropy conjecture in the formulation of Berger and Turaev.” Corollary 1.2 gives a positive-area ergodic hyperbolic piece that splits into finitely many cyclic parts, each Bernoulli under the corresponding iterate.

What it doesn’t claim

The threshold k₀ is never computed; the proof is by contradiction. The paper “makes no claim of ergodicity, almost-everywhere positive exponents, or a uniform positive lower bound on entropy,” and says nothing about small or intermediate kicks. It does not call the problem open or famous.

Where it stands

Posted as a preprint

OpenAI, 45 pages, dated September 23, 2026

Peer reviewed

Not yet

Checked in Lean

The theorem and its corollary, in three statements. We have not run the check

Last checked October 19, 2026. The preprint has not been revised. OpenAI published a Lean formalization of the theorem and its corollary, added to the repository in its initial commit on October 6, 2026: three comparator statements (positive entropy, positive Lyapunov exponents, and the Bernoulli component), with metric entropy defined from scratch, and 274 files and 54,864 lines of proof, with no unproved steps that we could find. The files also prove Pesin’s formula for this map, though none of the three checked statements relies on it. We have not run the check. The paper is not among those OSVerify reviewed, and we have seen no expert review.

How the proof goes

1. The stretching game. A tiny displacement is multiplied at each step by a 2×2 matrix whose main entry is 2 + K cos(2πx), with K = 2πk. That entry is usually huge, so displacements usually grow by about K/2 per step on average, but near the two lines where the cosine vanishes the growing direction can be turned into the shrinking one and growth can cancel. The paper tracks the average shortfall from maximal stretching over n steps, a number between 0 and 1.
2. A contradiction across scales. If exponents vanished almost everywhere along parameters going to infinity, the shortfall would climb to 1 at long lengths while tending to 0 at every fixed length, and a doubling bound forces the climb through a critical range of scales.
3. Two engines. Cancellations are exponentially rare: stretching almost maximally for n steps backward and forward yet almost entirely cancelling over the whole stretch has probability decaying exponentially in n, through a local estimate in which a cubic polynomial cannot match two shifted reciprocal-cosine terms. And across a stretch that grows fast, events near its two ends are bounded as if independent, up to a constant factor, with no mixing assumed.
4. Three shapes and a balance. Rescaled limits of the growth geometry are tree-like because the matrices have determinant one; the engines leave three shapes (all slow; slow with one ray of maximal growth; a slow middle between two such rays). A capped bookkeeping of shortfalls telescopes to a positive total, while splitting the limit makes it strictly smaller: a contradiction. Pesin’s formula then turns positive exponents on positive area into positive entropy.

The numbers in the video

• Finite-time top Lyapunov exponents from 600 random points: 2.756 at k = 5, 3.447 at k = 10 and 5.750 at k = 100, against log(πk) = 2.754, 3.447 and 5.750.
• The mean shortfall from maximal stretching for lengths 1 to 256: 0.218 to 0.227 at k = 5, and about 0.109 at k = 100.
• Elliptic fixed points for k between m and √(m² + 4/π²); at k = 5.010 one island has area 1.66 × 10⁻⁴, about 1/6,000 of the torus (islands come in mirror pairs).
• Phase portraits at k = 0.05, 0.3, 1.05 and 5.

Read the source

• Positive Metric Entropy for the Standard Map at Large Parameters (PDF), 45 pages.
• The Lean scope note.
• The collection’s catalog, where this is result 146.

The paper

OpenAI groups related papers into one result; this one has a single paper.

• Positive Metric Entropy for the Standard Map at Large Parameters

Transcript

Spot a mistake?

This episode simplifies on purpose, but it shouldn’t be wrong. If something is, open an issue on GitHub or leave a comment on YouTube. Corrections are noted here with a date.