Entropy equality and local symmetry in negative curvature

Source and verification

This article explains family 339 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.

🏷️ Preliminaries: geodesic flow and two entropies

Let $(M^n,g)$ be a closed connected Riemannian manifold. Its unit tangent bundle is

$$SM=\{(x,v):x\in M,|v{|}_g=1\}.$$

The geodesic flow ${\varphi }_t$ sends $(x,v)$ to the position and unit velocity reached after following that geodesic for time $t$. Its phase space has dimension $2n-1$.

There are two natural ways to measure its complexity. Topological entropy counts distinguishable orbit segments: if $N(T,\epsilon )$ is the largest number of trajectories that can be pairwise distinguished at resolution $\epsilon$ during time $[0,T]$, then

$$h_{\mathrm{top}}=\underset{\epsilon \downarrow 0}{\lim }\underset{T\to \infty }{\mathrm{limsup}}\frac{\log N(T,\epsilon )}{T}.$$

An invariant probability measure gives a statistical version. For the time-one map, a finite measurable partition records a symbol at each integer time. Its entropy rate is the limiting Shannon entropy of a length-$N$ record divided by $N$. The supremum over finite partitions is the measure entropy $h_{\mu }({\varphi }_1)$.

The variational principle gives $h_{\mu }\le h_{\mathrm{top}}$. A measure attaining equality is a maximal-entropy measure.

The metric supplies normalized Liouville measure $m_L$: choose the base point according to normalized Riemannian volume, then choose a uniformly distributed unit direction there. This probability is invariant under geodesic flow.

Manuscript result

For every closed connected smooth manifold of dimension $n\ge 3$ with strictly negative sectional curvature,

$$\begin{array}{c}{h}_{m_L}({\varphi }_1)=h_{\mathrm{top}}({\varphi }_1)\\ ⟺\\ {\nabla }^g{R}_g=0.\end{array}$$

Thus Liouville measure has maximal entropy exactly when the metric is locally symmetric. No pinching ratio or prescribed symmetric target is assumed. (OpenAI, 2026)

The equation $\nabla R=0$ means that curvature is parallel. In negative curvature, the simply connected symmetric models include real, complex, quaternionic, and Cayley hyperbolic geometry at arbitrary overall scale. Local symmetry therefore does not mean constant sectional curvature.

The prerequisites are Riemannian geometry, elementary ergodic theory, and smooth dynamics. The argument uses a formal classification theorem as a stated interface; its algebraic classification is not developed here.

🏷️ Hyperbolicity and the missing regularity

Strict negative curvature on a compact manifold gives an Anosov splitting

$$T(SM)=E^s\oplus \mathbb{R}X\oplus E^u,$$

where $X$ generates the flow. Vectors in $E^s$ contract exponentially in forward time; vectors in $E^u$ contract exponentially in backward time. Each has dimension $n-1$.

The canonical contact form is

$${\alpha }_{(x,v)}(\xi )=⟨v,d\pi (\xi ){⟩}_g,$$

where $\pi :SM\to M$ is projection. It satisfies $\alpha (X)=1$ and $\ker \alpha =E^s\oplus E^u$. The strong stable and unstable leaves are tangent to $E^s$ and $E^u$; their weak versions also include the flow direction.

Individual leaves are smooth, with bounds for each fixed derivative order on uniform local pieces, called plaques. Their variation across other leaves can be only Hölder continuous. Smoothness along every plaque does not imply ambient smoothness of the distributions.

The main target is to deduce global smoothness of $E^s,E^u$ from entropy equality.

A classical rigidity theorem then recovers a locally symmetric metric, after the flow-to-metric normalization is checked. Applying that theorem at the outset would assume precisely the regularity that needs proof.

🏷️ Expansion, pressure, and a periodic-orbit identity

Jacobi fields describe infinitesimal separation of nearby geodesics. In normal coordinates along a geodesic, the unstable directions are graphs of a positive symmetric Riccati tensor $U$ satisfying

$$\begin{array}{c}XU+U^2+R_v=0,\\ R_v(q)=R(q,v)v.\end{array}$$

Set $J=\mathrm{tr}U$. It is the infinitesimal unstable volume-expansion rate in the source’s graph-volume normalization.

For a potential $f$, pressure is

$$P(f)=\underset{\mu }{\sup }\left(h_{\mu }+\int fd\mu \right),$$

over flow-invariant probabilities. The entropy inequality and the smooth entropy formula give $P(-J)=0$, with Liouville measure an equilibrium state.

Assume entropy equality, with common value $h$. Liouville measure is then also an equilibrium state for the constant potential $-h$, whose pressure is zero.

The two descriptions give Gibbs estimates for the same orbit neighborhoods. Around a periodic orbit of period $\ell$, follow $k$ repetitions. Comparing the upper and lower estimates yields a constant $C$, independent of $k$, such that

$$\left|k{\int }_0^{\ell }(J({\varphi }_{t}v)-h)dt\right|\le C.$$

Dividing by $k$ proves that the integral vanishes on every periodic orbit. The Livšic theorem converts these periodic identities into

$$J=h+XF$$

for a Hölder transfer function $F$ differentiable along the flow. Thus deviations of unstable expansion from $h$ are a flow derivative.

On one unstable plaque, backward contraction gives the more informative formula

$$\begin{array}{rl}& F(y)-F(x)\\ & ={\int }_{-\infty }^0(J({\varphi }_{t}y)-J({\varphi }_{t}x))dt.\end{array}$$

Positive leaf derivatives of the integrand decay exponentially. Differentiating under the integral proves smoothness of $F$ along weak unstable plaques. It still does not prove transverse differentiability.

🏷️ Transported conormal volumes

Put $m=n-1$. The transfer function produces horizontal $m$-forms $u,w$ such that $u$ annihilates $E^u$, $w$ annihilates $E^s$, and

$$\begin{array}{c}{\varphi }_t^{\ast }u=e^{-ht}u,{\varphi }_t^{\ast }w=e^{ht}w,\\ u\wedge w=c_{\ast }(d\alpha {)}^m,c_{\ast }\ne 0.\end{array}$$

These are volume data on complementary directions. They are initially continuous in the ambient space, with smooth finite jets along the appropriate weak leaves.

The transport law follows from Jacobi’s equation. If a Jacobi vector is represented by $(q,q^{\prime })$, then

$$\frac{d}{dt}(q^{\prime }-Uq)=-U(q^{\prime }-Uq).$$

The determinant therefore changes at rate $-J$. Multiplication by $e^F$ changes that rate to $-h$. Reversing the direction of the geodesic supplies the opposite transport law.

The wedge ratio with $(d\alpha {)}^m$ is continuous and flow invariant. Ergodicity makes it constant almost everywhere, and full support of Liouville measure makes it constant everywhere. The proof carefully separates such a continuity extension from an unsupported assertion that a measure-one identity already holds at every point.

🏷️ Finite jets and the formal alternative

A finite jet is a finite list of Taylor coefficients. The manuscript records jets of actual plaques and their conormal densities in finite-dimensional spaces on which the dynamics contracts in the appropriate direction.

Polynomial normal forms organize these records using the Lyapunov expansion rates. Relations satisfied by actual plaque jets produce coherent families of possible conormal data. Coherence means that along an actual leaf, the proposed coefficients differentiate as the center moves, as genuine derivatives would.

The resulting formal symmetry analysis yields an alternative. In one case it supplies coherent formal fields with values $E^s,E^u$. In the other it supplies nonzero stable and unstable directions joined by a formal flat connection.

This classification is a substantial source input. A formal field is only an order-by-order collection of coefficients; it is not yet a smooth field defined in a neighborhood. The remainder of the proof must realize enough of these coefficients as actual geometry, with estimates that survive differentiation.

🏷️ Excluding the mixed-rectangle alternative

The second alternative is converted into a complete proper ray $a(t)$ in one strong stable leaf, together with an actual path $z(t)$ satisfying

$$z(t)\in W^u(a(t))\cap W^s(z(0)).$$

Constructing these mixed stable–unstable rectangles uses finite Taylor truncations, transport by the dynamics, and derivative estimates. Formal flatness alone would not give the displayed incidences.

On the universal cover, a geodesic has a past and a future endpoint at infinity. A Busemann function records signed distance relative to a ray toward one such endpoint. A strong stable leaf fixes the future endpoint and its Busemann level; a strong unstable leaf fixes the corresponding past data.

Let $\eta$ be the future endpoint of the stable ray and ${\eta }^{\prime }\ne \eta$ the future endpoint of $z(0)$. Properness of $a(t)$ forces its past endpoint ${\zeta }_t$ to approach $\eta$. With the future level fixed, the past Busemann value of $a(t)$ then tends to $-\infty$.

But $z(t)$ has endpoints $({\zeta }_t,{\eta }^{\prime })$ and a fixed future Busemann level. Their limiting pair $(\eta ,{\eta }^{\prime })$ consists of distinct points, so these vectors stay in a compact region and their past Busemann values stay bounded. Membership in the same strong unstable leaf forces the past values of $z(t)$ and $a(t)$ to agree. This contradiction excludes the alternative.

The properness and actual-leaf conclusions are essential. A merely formal rectangle, or a stable ray that does not escape compact sets, would not support the endpoint argument.

🏷️ Recovering smoothness from finite approximations

In the remaining alternative, choose one typical strong stable leaf on the universal cover. Its points label past endpoints other than its fixed future endpoint. These labels define a continuous map $q$ on a corresponding open subset of the unit tangent bundle, constant along actual weak unstable leaves.

The formal construction first supplies coherent Taylor coefficients along the chosen axis. Backward iteration brings compact subsets toward it exponentially. Pulling back a finite Taylor polynomial of order $R$ gives smooth approximants $p_{N,R}$ with

$$\begin{array}{rl}\Vert p_{N,R}-q{\Vert }_{C^0}& \le C{e}^{-[b(R+1)-\epsilon ]N},\\ \Vert p_{N,R}{\Vert }_{C^j}& \le C{e}^{(B_j+\epsilon )N}.\end{array}$$

The key is that the growth exponent $B_j$ is independent of $R$. Constants may depend on $R$, the compact set, and the derivative order.

Fix a desired derivative order $j$, and choose $r>j$. Interior interpolation applied to consecutive approximants combines their rapidly decaying $C^0$ difference with their controlled $C^r$ growth. Choosing one finite $R$ sufficiently large makes the resulting $C^j$ differences summable.

The limit is the already identified continuous label $q$. Since this works for every $j$, $q$ is smooth. A single Taylor cutoff need not work for all derivative orders, and no infinite formal series has been evaluated.

Smooth labels alone might have critical points. The source also proves two-sided volume bounds for weak-unstable holonomy between smooth transversals:

$$\begin{array}{c}c\le \frac{\mathrm{Vol}(h(E))}{\mathrm{Vol}(E)}\le C,\\ \mathrm{Vol}(E)>0.\end{array}$$

If the derivative of a label map on an $m$-dimensional transversal had deficient rank, the image of a radius-$r$ ball would lie in a box with $m-1$ widths of order $r$ and one width $o(r)$. Its volume would be $o(r^m)$, contradicting the lower bound. Hence the label is a submersion.

Its kernel is the actual weak unstable distribution:

$$\begin{array}{c}\ker dq=E^u\oplus \mathbb{R}X,\\ E^u=(\ker dq)\cap \ker \alpha .\end{array}$$

A second typical axis covers the omitted endpoint. Reversing geodesics gives smoothness of $E^s$ as well. Because these are the actual invariant bundles, the smooth structures descend from the universal cover.

🏷️ From a smooth flow model to the metric

The classical smooth-foliation rigidity input gives a symmetric geodesic-flow model, initially allowing finite covers and a closed-one-form time change.

A geodesic and its reverse have the same period. The integral of a closed one-form changes sign under reversal. Comparing the two period formulas eliminates its periods; an orbit shift removes the resulting exact time correction. The remaining time change is a constant.

For a conjugacy satisfying ${\Psi }_{\ast }{X}_g=a{X}_0$, the contact forms and volumes obey

$$\begin{array}{c}{\Psi }^{\ast }{\alpha }_0=a{\alpha }_g,\\ \mathrm{Vol}(g_0)=a^n\mathrm{Vol}(g),\\ h_{\mathrm{top}}(g)=a{h}_{\mathrm{top}}(g_0).\end{array}$$

Rescale $g$ to $\hat{g}=a^{2}g$. Its volume and volume entropy match those of the symmetric model. The equality case of minimal entropy rigidity, with the source’s stated correction for the Cayley case, then recovers the metric up to homothety. Local symmetry descends through finite covers.

This final step uses a symmetric target only after the smooth splitting has produced it. It does not assume at the start that $M$ lies in a prescribed symmetric homotopy class.

🏷️ The converse and the meaning of a strict gap

For a locally symmetric negatively curved metric, the curvature operator has the same negative eigenvalue list at every unit tangent vector and is parallel along geodesics. The unstable Riccati solution is

$$U(v)=\sqrt{-R_v}.$$

Consequently $J=\mathrm{tr}U$ is constant. The pressure identity $P(-J)=0$ and the smooth entropy formula immediately give equality of the two entropies.

For a metric that is not locally symmetric, the theorem gives a strict entropy gap. Its unique maximal-entropy probability is then distinct from Liouville measure. Both are ergodic, so they are mutually singular. This is stronger than saying their densities differ.

The statement is qualitative: no explicit positive lower bound for the gap is claimed in terms of a distance from symmetric metrics. It also concerns strictly negative curvature on a closed manifold, not arbitrary nonpositive curvature or noncompact finite-volume spaces.

🏷️ Key steps and proof outlines

The periodic identity removes fluctuations only along the flow

Comparing Gibbs estimates on repeated periodic orbits is stronger than comparing their average entropies once. The comparison constant stays fixed while the orbit is repeated $k$ times, so division by $k$ forces the integral of $J-h$ to vanish exactly on every periodic orbit. Livšic then produces a transfer function. The backward integral on an unstable plaque differentiates because each positive leaf derivative decays exponentially under backward flow.

This gives smoothness along the appropriate leaves, but no ambient derivative estimate. The formal symmetry and actualization parts of the manuscript are needed precisely to cross this gap. Their algebraic classification is a stated input here, consistent with the family’s focus on differential geometry and dynamics. (OpenAI, 2026)

A finite Taylor cutoff can prove arbitrary smoothness

Fix a compact set inside the domain of the actual endpoint label $q$. Backward iteration brings that set exponentially close to a typical axis. Truncate the coherent axis Taylor data at degree $R$ and pull the resulting polynomial back to obtain a genuinely smooth function $p_{n,R}$.

To estimate its error, compare with the actual plaque through the pulled-back point. Coherence says that the discrepancy has zero Taylor coefficients through degree $R$ along that plaque. Taylor’s remainder therefore gives

$$\Vert p_{n,R}-q{\Vert }_{C^0}\le C{e}^{-a_{R}n},a_R=b(R+1)-\epsilon .$$

The chain rule gives $\Vert p_{n,R}{\Vert }_{C^r}\le C{e}^{(B_r+\epsilon )n}$. The exponential rate $B_r$ is independent of $R$: differentiating a degree-$R$ polynomial changes constants, while the exponential losses arise from at most $r$ derivatives of the dynamics.

Interior interpolation of successive approximants now gives a decay exponent

$$a_R(1-j/r)-(B_r+\epsilon )j/r$$

for the $C^j$ difference. Fix $r>j$ and choose one finite $R$ large enough to make this positive. The differences are summable in $C^j$, and their limit is the already identified continuous label $q$. Repeat for each $j$. This proves smoothness without evaluating an infinite formal series or requiring one cutoff to serve all derivative orders. (OpenAI, 2026)

Holonomy volume rules out critical label points

The source passes bounded-Jacobian estimates from strong to weak unstable holonomy by thickening transversals in the flow direction. In flow-box coordinates the lifted map has form $(a,t)\mapsto (h(a),t+\tau (a))$. Fubini shows that the image of $E\times I$ has product volume $|I|\mathrm{Vol}(h(E))$, even though the time shift is only continuous. Bounded smooth coordinate densities therefore transfer the two-sided volume estimate to $h$.

If the now smooth label had deficient rank on a transversal, differentiability would squeeze the image of a radius-$r$ ball into a box with one width $o(r)$ and all others $O(r)$. Its volume would be $o(r^{n-1})$, contradicting the holonomy lower bound. Thus the label is a submersion and its kernel is the actual weak unstable distribution. Intersecting with the contact distribution recovers the strong bundle. Reversing geodesics gives the stable bundle, supplying the precise regularity hypothesis of the final rigidity theorem.

See Also

Cartan–Hadamard isoperimetry studies geometric inequalities on simply connected nonpositively curved spaces. Here negative curvature supplies hyperbolic dynamics, while the decisive hypothesis is equality of two entropy invariants on a compact quotient.

🏷️ Manuscripts and verification scope

• Entropy equality and local symmetry in negative curvature — pinned manuscript and companion source.

No family scope page was present at lean/docs/339.md in the pinned snapshot. This does not establish that all related formalization code is absent. 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. Entropy equality and local symmetry in negative curvature.