Zero entropy does not guarantee a smooth model with smooth volume

← All digests

Family 152 · Dynamical systems and ergodic theory
Importance 5/5 · Provisional editorial rating
Sources consulted 2026-10-21

1. The question

An abstract dynamical system can be described by how it rearranges probability, without mentioning coordinates, distances, or derivatives. Can every statistically simple system nevertheless be represented by smooth motion on a finite-dimensional geometric space?

Finite entropy is a necessary condition for many smooth realizations. It limits the average rate at which observing the system reveals new information. Family 152 claims that even zero entropy is insufficient when the geometric model must preserve a strictly positive smooth probability density.

The qualification about the measure is essential. The manuscript does not claim that its system lacks every smooth model with any invariant probability measure. It claims that one fixed abstract system cannot be represented by a smooth diffeomorphism preserving smooth positive volume, in any finite dimension.

2. Background and essential definitions

Measure-preserving systems and changes of coordinates

A probability-preserving transformation $T:X\to X$ rearranges a probability space $(X,\mu)$ while leaving probabilities unchanged. Invertibility means that the past as well as the future is defined. Ergodicity means that an invariant measurable set has probability zero or one.

A standard nonatomic probability space is, after ignoring null sets, equivalent to the unit interval with Lebesgue probability. “Nonatomic” rules out individual points carrying positive probability.

Two systems are measurably conjugate if there is a probability-preserving bijection $F$, defined outside null sets, satisfying

$$F\circ T=S\circ F.$$

Both $F$ and its inverse must be measurable, but neither needs to be continuous. Consequently, visible geometric roughness in a symbolic presentation does not by itself obstruct a smooth model. A measurable change of coordinates can be extremely irregular.

Information growth

Choose a finite partition $\mathcal P$ of $X$, thought of as a finite-resolution measuring device. Recording which cell contains $T^j x$ produces a word of observations. If a finite distribution has probabilities $p_a$, its Shannon entropy in bits is

$$H(p)=-\sum_a p_a\log_2p_a,$$

with $0\log_2 0=0$. The Kolmogorov–Sinai entropy is the greatest asymptotic information rate over all finite partitions:

$$h_\mu(T)=\sup_{\mathcal P}\lim_{n\to\infty} \frac1n H_\mu\!\left(\bigvee_{j=0}^{n-1}T^{-j}\mathcal P\right).$$

The joined partition records all $n$ observations. Zero entropy says their information grows sublinearly for every fixed finite measuring device. It does not mean that observations are constant or that finite stretches lack intricate dependencies.

What counts as smooth volume

The proposed model consists of a compact smooth manifold $M$, a $C^\infty$ diffeomorphism $S$, and an invariant probability $\nu$ that locally has the form

$$d\nu=f(x)\,|dx_1\cdots dx_d|, \qquad f(x)>0,$$

with $f$ smooth. Absolute values allow densities on nonorientable manifolds. A manifold with smooth boundary is also allowed, with density smooth up to the boundary.

A measure with full support merely assigns positive probability to every nonempty open set. It can still be singular. Even a measure equivalent to volume, meaning that it has exactly the same null sets, need not have a smooth positive density. These distinctions separate different realization problems.

3. Development history and previous partial results

Ruelle’s differentiable entropy inequality relates measure entropy to expansion rates of derivatives. On a compact smooth manifold, derivative bounds imply finite entropy. This makes entropy a natural first obstruction to smooth realization.

Positive representation results show that finite entropy is a powerful condition when other requirements are relaxed. Lind and Thouvenot proved that finite-entropy ergodic transformations can be represented by Lebesgue-preserving homeomorphisms of the two-torus. A homeomorphism is continuous with a continuous inverse; differentiability is not required.

They also obtained smooth toral models when one is allowed to choose a suitable invariant Borel probability of full support. Quas and Soo extended the universality results to toral automorphisms with no root-of-unity eigenvalue. Taking a sufficiently high iterate of a hyperbolic toral automorphism provides enough topological entropy to represent any prescribed finite-entropy system with an appropriate invariant measure. That measure need not be smooth volume.

Katok and Thouvenot developed slow entropy invariants to detect complexity below the usual exponential scale. They constructed actions of $\mathbb Z^2$ without Lipschitz realizations on compact spaces of finite box dimension, even allowing arbitrary invariant measures. An action of $\mathbb Z^2$ requires two commuting transformations to be realized together. Its obstruction does not directly answer the one-transformation question here.

Another line of work concerns the difficulty of classifying smooth systems up to measurable isomorphism. Foreman and Weiss, and later work on weakly mixing and mixing zero-entropy diffeomorphisms, established strong anti-classification results. Those theorems describe the complexity of equivalence within classes of smooth systems. They do not by themselves exhibit an abstract system with no smooth positive-volume representative.

The new manuscript addresses that existence question for a single invertible transformation. Its proposed obstruction is more refined than ordinary entropy and is designed to survive an arbitrary measurable conjugacy.

4. What the new manuscripts claim

The manuscript, dated 25 September 2026, claims an ergodic invertible probability-preserving system $(X,\mu,T)$ on a standard nonatomic probability space such that

$$h_\mu(T)=0,$$

but no measurable conjugacy identifies it with any $C^\infty$ diffeomorphism of a compact finite-dimensional manifold preserving a strictly positive smooth probability density.

The same example excludes all finite dimensions. The construction is fixed before a hypothetical dimension or manifold is chosen. Allowing a nonorientable manifold or a smooth boundary does not evade the claim.

The theorem does not exclude a smooth model with a singular invariant measure. Nor does it exclude all models whose invariant measure is merely equivalent to volume through a nonsmooth density. These are explicit limits of the stated result.

The proof is organized around two ingredients: a quantitative operation available in smooth positive-volume systems that joins one orbit’s future to another’s past, and a symbolic system whose long-range information constraints resist that operation.

5. Why the advance matters

Provisional importance: 5/5. If correct, the result shows that entropy alone cannot characterize smooth positive-volume realizability, even at its smallest possible value.

This distinguishes two ideas that are easily conflated. A system may generate no positive asymptotic information per step and still impose complicated restrictions on which pasts can coexist with which futures. Smooth geometry with a regular invariant density would impose additional compatibility conditions on those restrictions.

The uniformity over dimension strengthens the result. A family of examples, each requiring a larger manifold than the previous one, would leave open whether every individual system eventually fits somewhere. One example excluding every finite dimension closes that escape route for the stated class of models.

6. The main idea

A geometric operation with statistical control

For a hypothetical smooth model, consider a long time $n=2^k$. Partition the manifold into small cells. First choose a cell with probability equal to its volume, then independently choose two points $x,y$ inside it.

The manuscript’s smooth switching theorem claims that, for most such pairs, one can find a point $z$ whose future through time $n$ stays close to that of $x$, and whose past back to time $-n$ stays close to that of $y$. This resembles local product constructions in hyperbolic dynamics, but the argument does not assume that the whole system is uniformly hyperbolic.

Existence of a suitable $z$ is not enough. A construction could concentrate all outputs on a statistically exceptional set. The manuscript therefore controls the output distribution: on retained pairs, its unnormalized law is at most twice $\nu$. In particular, a small-probability exceptional set cannot absorb many switched outputs.

The number of cells has the form

$$\#\mathcal D_k\le \exp(CB_k2^k), \qquad \sum_k B_k^{16}<\infty.$$

The constant $C$ and the positive sequence $B_k$ may depend on the chosen smooth model. Summability is stronger than simply saying that the cost per time tends to zero; it forces sufficiently cheap scales that the symbolic construction can exploit.

The analytic work starts with stationary derivative-growth estimates on dyadic intervals. Finite-time changes of basis separate suitable directions, and a boundary-value construction joins the orbit halves. Determinant estimates and injectivity control the volume distortion. The smooth density assumption enters this quantitative distributional control.

Build legal words with incompatible independent halves

The opposing system is constructed from longer and longer finite words. A word at one level is a concatenation of words from the previous level, and additional alignment information records where those blocks begin.

Labels for blocks are nonzero vectors over $\mathbb F_2$, the field with two elements. A fixed alternating bilinear form specifies relations between labels at selected power-of-two separations. Informally, certain distant blocks must be orthogonal under that form.

The construction preserves high information content in selected short lists of labels, even while its overall entropy rate tends to zero as block lengths grow. It also makes distinct words, and words read with incorrect alignment, differ in a definite proportion of symbols. This recognizability prevents an apparent orbit match from silently changing block boundaries.

Two independent label lists with high entropy are unlikely to satisfy most of the imposed pair relations. But the past and future of a single legal symbolic point are required to satisfy them. This creates the mismatch that a smooth switch would expose.

Transfer through a merely measurable conjugacy

A measurable conjugacy need not send nearby manifold points to nearby symbolic sequences. The argument instead approximates a finite symbolic observation on a large-measure set where spatial control is useful. The bounded output law ensures that switched points do not disproportionately land outside this good set.

At selected scales, the small cell-partition cost leaves the two input lists with too much independent information. Yet shadowing and word recognition force most entries to agree with the past and future lists of one legal output point. The manuscript obtains incompatible bounds for the probability that the required label relations hold.

The symbolic scales are chosen first. After any candidate finite-dimensional smooth model is proposed, the summability estimate supplies scales defeating its particular constants. This order is what makes one example work against every finite dimension.

7. Scope, limitations, and open questions

This is a claim about a smooth map together with a smooth positive invariant density. Removing the density requirement changes the problem, and known realization results already provide smooth models with more general invariant measures.

The theorem also should not be read as ruling out nonsmooth densities equivalent to volume. The manuscript identifies its output-law estimate as the point where the regularity of the density matters; it makes no corresponding assertion for all equivalent measures.

Zero entropy is used in the measure-theoretic sense. It says nothing by itself about the topological entropy of a hypothetical ambient map with a different measure. Nor does the theorem assert mixing, a finite symbolic presentation, or an explicit minimal differentiability threshold for the obstruction.

A natural remaining direction is to identify additional invariants that distinguish realizable zero-entropy systems from nonrealizable ones, and to understand how the answer changes as regularity of the map or its invariant density is weakened.

8. Sources and further reading

Lind and Thouvenot’s torus realization theorem and Quas and Soo’s ergodic universality theorem explain the positive results with different regularity or measure requirements.

Katok and Thouvenot’s slow entropy paper treats commuting transformations and motivates looking below exponential information growth. Ruelle’s entropy inequality gives the classical differentiable obstruction.

Foreman and Weiss’s anti-classification theorem concerns the distinct question of classifying smooth models. The smooth switching statement and the single-transformation counterexample summarized here are the claims of the pinned manuscript below.

Source manuscripts

• M1. OpenAI, A zero-entropy system without a smooth positive-volume model (September 25, 2026).

Sources consulted on 21 October 2026 at repository revision 1581d4c. Download the manuscript citations.