Zero entropy and the obstruction to smooth positive-volume models

Source and verification

This article explains family 152 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.

🏷️ Measurable models and smooth volume

Two probability-preserving systems are measurably isomorphic if there is a measure-preserving bijection between invariant sets of full measure that intertwines their transformations, with both the bijection and its inverse measurable.

A smooth positive-volume model requires a compact finite-dimensional smooth manifold, a $C^{\infty }$ diffeomorphism, and an invariant probability with strictly positive smooth local density. The conjugating map itself need not be continuous.

The measure requirement matters. A smooth map equipped with an arbitrary invariant Borel probability, even one of full support, need not provide a smooth positive-volume model. Nor is a measure merely equivalent to volume necessarily given by a smooth positive density.

🏷️ Entropy is a necessary but proposed insufficient condition

For a finite partition $\mathcal{P}$, let $H_{\mu }(\mathcal{P})$ be the Shannon entropy of its atom probabilities. Kolmogorov–Sinai entropy is

$$\begin{array}{c}{h}_{\mu }(T)\\ =\underset{\mathcal{P}}{\sup }\underset{n\to \infty }{\lim }\frac{1}{n}{H}_{\mu }\left(\underset{j=0}{\overset{n-1}{\bigvee }}{T}^{-j}\mathcal{P}\right).\end{array}$$

Finite entropy is necessary for smooth compact dynamics with smooth volume, because compact derivative bounds control entropy through the differentiable entropy inequality.

Statement in the manuscript

There is an ergodic invertible transformation of a standard nonatomic probability space with entropy zero that has no smooth positive-volume model on any compact finite-dimensional manifold. Smooth boundary is allowed, and orientability is not required. (OpenAI, 2026)

“Standard nonatomic” means measurably equivalent, modulo null sets, to the unit interval with Lebesgue probability. The counterexample is one fixed system, chosen before a hypothetical target dimension or smooth map.

The statement does not rule out a homeomorphism model, a smooth model with singular invariant measure, or every model whose invariant measure is only equivalent to volume.

🏷️ Why zero entropy leaves room for an obstruction

Zero entropy says that, for each fixed finite observation, the information in an orbit word grows sublinearly with word length. It does not prescribe the rate of that growth or the detailed compatibility between distant pieces of a word.

For instance, quantities $\exp (\sqrt{n})$ and $\exp (n/\log (n+2))$ both have zero exponential growth rate, although their scale-dependent complexity differs greatly. The source exploits constraints that ordinary entropy does not record.

Because the prospective conjugacy is only measurable, a metric regularity invariant of one particular presentation would not suffice. The obstruction must survive arbitrary measurable recoding.

🏷️ A switching property of smooth volume

The analytic input is a quantitative way to join an orbit’s future to another orbit’s past. At a dyadic time $n=2^k$, the manifold is partitioned into small cells. Choose a cell according to its volume and then two independent points $x,y$ inside it.

On most retained pairs, the manuscript constructs a point $z$ whose future shadows that of $x$ and whose past shadows that of $y$. Crucially, the output law of $z$ is bounded by twice the original volume. Existence of some shadowing point without control of its distribution would be too weak. (OpenAI, 2026)

The number of cells is at most $\exp (C{B}_{k}n)$, where

$$\sum _k{B}_k^{16}<\infty .$$

This summability limits the information lost by conditioning on a cell at carefully chosen scales.

The proof begins with a square-summable estimate for stationary derivative logs. It does not assume an effective convergence rate in an ergodic theorem. Finite-interval changes of basis solve the linear boundary problem, exterior-power estimates control small minors, and a determinant identity together with injectivity bounds the law of the resulting switch.

🏷️ A symbolic system with incompatible past-future tests

The opposing system is built from finite word families ${\mathcal{W}}_i$ of lengths $m_i$. Higher-level words concatenate lower-level ones and record compatible alignment phases.

Labels satisfy prescribed pairwise compatibility tests at selected power-of-two spacings. At the same time, short lists of labels retain high entropy, and distinct or misaligned words remain separated in symbol distance. The construction balances these requirements while making the final entropy rate zero.

Two independently sampled high-entropy label lists are unlikely to satisfy most compatibility tests. A legal symbolic point, however, must satisfy them. If a smooth model existed, its switching construction would combine a future list from one point with a past list from another into the lists of one legal point.

Measurable conjugacy is handled by approximating symbolic observations on sets of large measure. The bounded output density prevents the switch from concentrating on the exceptional sets where those approximations fail.

🏷️ The scale selection and contradiction

Conditioning on a partition with $N_k$ cells costs at most ${\log }_2{N}_k$ bits. Thus the smooth construction permits a loss controlled by $C{B}_{k}n$, up to conversion of logarithm bases.

The symbolic scales are chosen so that summability of $B_k^{16}$ supplies times where this loss is too small to destroy the independent-list entropy. At those times, the switched lists must simultaneously be typical enough to fail the compatibility tests and legal enough to pass them.

All constants depending on one hypothetical smooth manifold and its finite dimension are absorbed at sufficiently late scales. The symbolic system was fixed beforehand, so the argument is claimed to exclude all finite-dimensional smooth positive-volume models with the same counterexample. (OpenAI, 2026)

The repository’s scope page describes only a finite-entropy obstruction. Its selected statement does not include the paper’s stronger zero-entropy conclusion.

🏷️ Key steps and proof outlines

Summability selects cheap scales for every proposed model

All logarithms in this scale-selection argument are base two. The symbolic system is fixed before a smooth model is proposed. At level $i$, write $m_i$ for word length, $L_i$ for the number of words, $R_i=(\log L_i)/m_i$, and $K_i=\lfloor \sqrt{\log L_i}\rfloor$. The construction arranges $R_i\to 0$ while ${\sum }_i{K}_i{R}_i^{16}=\infty$. The prospective smooth model supplies a sequence with ${\sum }_k{B}_k^{16}<\infty$.

Test lengths are $n=m_{i}D$, with $D$ a power of two between $2^{K_i/16}$ and $2^{K_i/8}$ up to integer rounding. Their logarithmic indices occupy disjoint sets $J_i$ of at least $K_i/32$ values. If $B_k/R_i$ stayed above some $c>0$ on all late tests, then

$$\sum _k{B}_k^{16}\ge \frac{c^{16}}{32}\sum _i{K}_i{R}_i^{16}=\infty ,$$

a contradiction. Hence along selected tests $B_{\log n}/R_i\to 0$. The cell-partition cost $O(B_{\log n}n)$ is therefore $o((D-1)\log L_i)$, small compared with the entropy in the label lists. This is the scale-selection lemma in “The obstruction to a smooth volume model.” (OpenAI, 2026)

High Shannon entropy controls the compatibility test

Given a smooth cell, the two sampled points are independent, so the corresponding past and future labels have independent laws $p,q$. The symbolic graph test has an associated sign matrix $A$ of operator norm at most $\sqrt{L+1}$, where $L$ is the label count. Its finite character calculation supplies that norm, including the diagonal entries.

High Shannon entropy need not bound the largest atom directly. Split $p$ into heavy atoms with $p(u)>L^{-0.8}$ and the remaining light subprobability. If the heavy mass is ${\theta }_p$, there are at most $L^{0.8}$ heavy atoms, so

$$H(p)\le 1+(1-0.2{\theta }_p)\log L.$$

Average entropy at least $0.99\log L$ therefore makes average heavy mass at most $0.05+5/\log L$. On the light part, $\Vert p^{\mathrm{l}}{\Vert }_2^2\le L^{-0.8}$, and likewise for $q$. The matrix estimate bounds the light correlation by $\sqrt{L+1}{L}^{-0.8}$, which tends to zero. Adding the heavy contributions yields the source’s upper bound $0.62$ for the average compatibility probability. This heavy-light step converts a Shannon-entropy statement into a usable spectral estimate without assuming a uniform pointwise probability bound.

Both sides of the contradiction use the original pair law

Define the passing fraction $Z(x,y)$ from corresponding past labels of $y$ and future labels of $x$ before selecting a switch. The independent upper bound is then ${\mathbb{E}}_{\rho }Z\le 0.62$ under the original cell-weighted pair law $\rho$.

The switch is constructed only on a retained set $G$, but its unnormalized output measure satisfies $z_{\ast }(\rho {|}_G)\le 2\nu$. For a measurable symbolic observation, choose compact cores of its atoms covering all but $\kappa$ of the smooth measure. An original input misses those cores with probability at most $\kappa$; the switched point misses them with retained mass at most $2\kappa$. Shadowing therefore transfers the symbols except on controlled average fractions, without any continuity assumption on the measurable conjugacy.

The word-recognition lemma forces the three alignment phases to agree: a nonzero phase difference would create at least a fixed positive proportion of symbol mismatches in each full block. Distinct-word separation then bounds the number of wrong labels. On retained mass at least $0.95$, the source obtains $Z\ge 0.96$, hence

$${\mathbb{E}}_{\rho }Z\ge 0.95\cdot 0.96=0.912,$$

contradicting the independent upper bound. There is no renormalization by $\rho (G)$ and no re-pairing of entries after observing the switch. Either change could alter the statistic’s law and invalidate the comparison.

Smooth positive density is the analytic interface

The switching theorem’s derivative and determinant estimates provide both shadowing and the output-density bound. Existence of one suitable switched point for each pair would not control how often it lands in the exceptional sets of a measurable coding. The domination by $2\nu$ is what transfers small measure into small retained-pair probability. All constants of a hypothetical finite-dimensional smooth model are fixed before choosing sufficiently late tests; the divergence/summability argument then defeats them using the same symbolic system. The proof makes no corresponding assertion for arbitrary singular invariant measures.

See Also

Simple Lebesgue spectrum constructs a proposed smooth volume model with a special spectral profile. Finite-regularity entropy counterexample gives another setting in which entropy alone fails to reflect a different structural constraint.

🏷️ Manuscripts and verification scope

• A zero-entropy system without a smooth positive-volume model — pinned manuscript and companion source.

The family scope page describes selected formalization coverage. The scope explicitly gives only a finite-entropy obstruction. The manuscript’s stronger zero-entropy conclusion lies outside that comparator statement. 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. A zero-entropy system without a smooth positive-volume model.