Positive metric entropy for the standard map at large parameters

Source and verification

This article explains family 146 of the October 6, 2026 OpenAI collection. Research conclusions are attributed to its manuscripts. The review covers statements, introductions, and selected proof mechanisms; it is not an independent verification of the complete proofs.

🏷️ The standard map and its invariant area

On ${\mathbb{T}}^2={\mathbb{R}}^2/{\mathbb{Z}}^2$, the standard sine map in this family is

$$\begin{array}{rl}{x}^{\prime }& =x+y+k\sin (2\pi x),\\ y^{\prime }& =y+k\sin (2\pi x),\end{array}$$

Both coordinates are taken modulo one. Writing $a=2\pi k\cos (2\pi x)$, its derivative is

$$\begin{array}{c}D{f}_k(x,y)=\left(\begin{array}{cc}1+a& 1\\ a& 1\end{array}\right),\\ \det D{f}_k=1.\end{array}$$

The inverse is obtained from $x=x^{\prime }-y^{\prime }$ and then $y=y^{\prime }-k\sin (2\pi x)$. Thus the map is a smooth diffeomorphism preserving normalized Lebesgue area $m$.

The question concerns metric entropy for this particular invariant measure. A chaotic invariant subset of area zero, or positive topological entropy, does not answer it.

🏷️ Entropy and Lyapunov exponents

For a finite measurable partition $\mathcal{P}$, the entropy of the joined observations over $n$ iterates measures how many orbit names are statistically distinguishable. Kolmogorov–Sinai entropy is the supremum of the asymptotic entropy per iterate over finite partitions.

Oseledets’ theorem supplies almost-everywhere exponential derivative-growth rates. Area preservation pairs the two exponents as ${\lambda }_{+}(z)$ and $-{\lambda }_{+}(z)$, with ${\lambda }_{+}\ge 0$. The smooth absolutely continuous version of Pesin’s formula gives

$$h_m(f_k)={\int }_{{\mathbb{T}}^2}{\lambda }_{+}(z)dm(z).$$

Compactness ensures the required derivative integrability at each fixed parameter. Consequently positive metric entropy is equivalent to a positive largest exponent on a set of positive area.

Statement in the manuscript

There exists $k_0>0$ such that $h_m(f_k)>0$ for every $k\ge k_0$. (OpenAI, 2026)

The quantifier is every sufficiently large parameter. The conclusion does not assert full-area hyperbolicity, ergodicity of the entire torus, or a uniform positive lower bound on entropy.

🏷️ Why large derivatives do not settle the problem

Away from the zeros of $\cos (2\pi x)$, one step has a large expanding direction when $k$ is large. Long products can nevertheless lose growth when successive expanding and contracting directions nearly align.

For a simple matrix example, let $A=\mathrm{diag}(L,L^{-1})$ with $L>1$. Both $A$ and $A^{-1}$ have norm $L$, but $A{A}^{-1}=I$. Large norms of factors alone give no lower bound on the norm of their product.

The standard-map proof must therefore control cancellation along actual orbits. Treating successive phases as independent random variables would assume a statistical property that the theorem is meant to establish.

🏷️ Cancellation scarcity and fast bridges

The source writes derivative products in determinant-one transfer coordinates. With a one-step bound $M=2\pi k+4$, it normalizes interval growth by

$$g(i,j)={\log }_M\Vert T_{i,j}\Vert .$$

For the associated time-indexed transfer products, $0\le g(i,j)\le |i-j|$. Products and inverses give the relevant triangle inequalities. The expected shortfall $1-\mathbb{E}g(0,n)/n$ measures the loss relative to maximal normalized growth. (OpenAI, 2026)

The first input bounds the probability that two rapidly growing segments nearly cancel on concatenation. Such cancellation provides a linearized solution small at both ends. Matching expansion scales on both sides yields many local tests. An analytic scarcity estimate controls their joint occurrence, including nearly degenerate reciprocal-cosine configurations.

The second input is a weighted comparison for observations near opposite ends of a rapidly growing bridge. Finite boundary-value problems construct two transverse families of contracting graphs, and slice Jacobians control the relevant changes of variables.

This estimate is restricted to fast bridges. It is not unrestricted mixing or independence across arbitrary orbit segments. Its uniformity over sufficiently long fast bridges is what the multiscale argument needs.

🏷️ A dyadic contradiction

Suppose arbitrarily large parameters had zero exponent almost everywhere. At every fixed length, the shortfall tends to zero as the parameter grows. At a fixed zero-exponent parameter, it tends to one as length grows.

Cancellation scarcity bounds how quickly the shortfall can rise under doubling of length. The manuscript chooses the first range of dyadic scales over which a small increase becomes visible and rescales the resulting distance arrays.

The limiting measures are analyzed away from affine arrays $d(s,t)=w|s-t|$. The bridge comparison and a four-point inequality restrict the remaining shapes to an entirely slow array or configurations with one or two exterior unit-speed rays.

A capped shortfall functional is linear with small slope near zero and equals one on sufficiently slow intervals. Summing the difference between a whole interval and its two halves telescopes to a positive terminal contribution. The source’s classification makes the limiting contribution strictly smaller, yielding a contradiction.

Passing to the limit is itself part of the proof: vague convergence away from affine arrays does not control mass accumulating near them. The cap, bridge estimate, and cancellation truncation handle the different affine speeds separately.

🏷️ The positive-area component

The source further applies the nonuniformly hyperbolic decomposition to the positive-exponent region. It obtains a positive-area invariant ergodic component, partitioned into finitely many cyclic pieces on which an appropriate iterate is Bernoulli. (OpenAI, 2026)

This is a component statement. It preserves the possibility of other invariant regions elsewhere in the torus. The sophisticated research inputs are the cancellation, bridge, and limiting-balance estimates; the determinant and entropy calculations explain exactly what their conclusion implies.

🏷️ Key steps and proof outlines

Near cancellation forces a solution small at both endpoints

Let $T_{0,-n}$ and $T_{0,n}$ be determinant-one transfer products from a common center, with norms $D_{-},D_{+}$ and most-expanded unit input directions $a_{-},a_{+}$. Suppose both norms are at least $M^{(1-\delta )n}$ but the full product across the interval has norm at most $M^{\delta n}$.

Take a unit vector $b$ perpendicular to $a_{-}$. Determinant one means the singular values are reciprocal, so $\Vert T_{0,-n}b\Vert =D_{-}^{-1}$. On the other side, the component of $b$ in the expanding direction gives $\Vert T_{0,n}b\Vert \ge D_{+}|a_{-}\wedge a_{+}|$. Comparing through the full product yields

$$|a_{-}\wedge a_{+}|\le M^{-(2-3\delta )n}.$$

Resolving the same vector into the other product’s singular directions then gives $\Vert T_{0,n}b\Vert \le D_{+}^{-1}+D_{+}|a_{-}\wedge a_{+}|\le 2M^{-(1-3\delta )n}$. After normalizing one central coordinate, the scalar recurrence solution has value one near the center and is exponentially small at both distant endpoints. This is the starting point of “Nearly full-rate cancellations.” (OpenAI, 2026)

A conserved Wronskian creates many local tests

Reindex each outward orbit path at its normalized center, so $u_0=1$, and write the linearized recurrence as $u_{j+1}=v_j{u}_j-u_{j-1}$. The companion solution with $t_0=0,t_1=1$ satisfies

$$u_{j-1}{t}_j-u_j{t}_{j-1}=1.$$

The identity follows by substituting the two recurrences, and its value at $j=1$ is one. Thus a small $u$-pair at the endpoint forces a large $t$-pair there. The normalized logarithms of successive pair norms change by at most one per step, because each transfer and its inverse have norm at most $M$.

A path whose total expansion is nearly its maximal length must therefore have many indices with nearly maximal local expansion and simultaneous contraction of $u$. The source makes this counting quantitative: with a fixed rate $c=99/100$, the number of failed “good index” conditions is at most $A(\delta n+1)/(1-c)$. Matched scales on the two sides produce many compatible analytic tests in the single orbit parameter. The local scarcity theorem, including its near-degenerate cases, bounds their simultaneous occurrence and yields the exponential cancellation probability $C{e}^{-\gamma n}$. The Wronskian and counting argument explain the reduction to that analytic theorem; they do not replace its proof.

Four-point comparison controls the shapes of growth arrays

Fix a base time $i$, let $D_j=\Vert T_{i,j}\Vert$, and define

$$R_{jp}=\frac{\Vert T_{j,p}\Vert }{D_j{D}_p},Q_{jp}=\max \{|a_j\wedge a_p|,D_j^{-2},D_p^{-2}\}.$$

Singular-direction decomposition gives $Q_{jp}\le R_{jp}\le 3Q_{jp}$. The wedge term measures relative angle; the reciprocal-norm terms retain the possibility that one side has little growth. The triangle inequality for angles then gives $R_{jp}\le 6\max (R_{j\ell },R_{\ell p})$. This is a deterministic inequality for each orbit, not a probabilistic independence statement.

Taking normalized logarithms and rescaling time converts it into a four-point constraint on limiting growth arrays. The fast-bridge comparison supplies the additional probabilistic restriction needed for their classification. It applies only where the bridge is fast, so the limiting argument must treat slow intervals separately.

Nonnegative defects permit dyadic telescoping

For an interval $I=[s,t]$ in a rescaled array $d$, define its shortfall $h(I)=1-d(s,t)/(t-s)$. For its two halves, put

$$J(I)=h(I)-\frac{h(I_l)+h(I_r)}{2},V(I)=(h(I_l)-h(I_r){)}^2.$$

The triangle inequality gives $J\ge 0$, and $V\ge 0$ directly. The first records loss on concatenation; the second detects unequal speeds that concatenation loss alone might miss. Their vanishing across all dyadic intervals identifies the affine arrays $d(s,t)=w|s-t|$.

At selected initial and terminal scales, expected shortfalls are respectively negligible compared with $ϵ$ and equal to $ϵ$. Summing the dyadic $J$ increments and dividing by $ϵ$ gives a nonzero terminal contribution. Because the resulting measures can have unbounded total mass near affine arrays, ordinary compactness does not justify passing this identity to the limit. The source uses the local defects, cancellation truncation, and a capped shortfall functional to control those concentrations. Its classified nonaffine limits contribute strictly less than the telescoping identity requires. This yields the contradiction to zero exponent at arbitrarily large parameters, after which Pesin’s formula gives positive metric entropy for area.

See Also

A C1 entropy counterexample distinguishes topological entropy from homological growth. Simple Lebesgue spectrum describes a proposed smooth mixing system with zero metric entropy and zero exponents.

🏷️ Manuscripts and verification scope

• Positive Metric Entropy for the Standard Map at Large Parameters — pinned manuscript and companion source.

The family scope page describes selected formalization coverage. The scope includes the positive parameter tail, positive Lyapunov exponents on positive area, and the positive-area hyperbolic component with its finite Bernoulli decomposition. It does not assert ergodicity of the entire torus. No Lean build, axiom audit, or manuscript-to-formal-statement equivalence check was performed for these notes.

OpenAI collection provides prerequisites and the reading sequence; coverage register records completed and pending families.

🐻  OpenAI 2026. Positive Metric Entropy for the Standard Map at Large Parameters.