Mean dimension and comparison: boundary compression of orbit covers

Source and verification

This article explains family 302 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.

Overview

Family 302 joins two manuscripts. One claims that a minimal homeomorphism has crossed-product radius of comparison exactly half its mean dimension. The companion supplies a topological compression theorem: under a bundle-rank hypothesis, many coordinates of a cube-valued map can be pushed to their boundaries while every original boundary coordinate remains unchanged. The main proof turns those boundary coordinates into missing colors in an orbit cover. (OpenAI, 2026) (OpenAI, 2026b)

🏷️ Preliminaries: orbit covers and mean dimension

Let $h:X\to X$ be a homeomorphism of a compact metrizable space. It is minimal if every orbit is dense. If $X$ is infinite and $h$ minimal, no point is periodic: a finite orbit would be a proper nonempty closed invariant subset.

For a finite open cover $\mathcal{U}$, its order is one less than the largest number of members having a common point. Let $D(\mathcal{U})$ be the least order of a finite open refinement. The joined orbit cover

$${\mathcal{U}}^{[0,N)}=\underset{j=0}{\overset{N-1}{\bigvee }}{h}^{-j}\mathcal{U}$$

records which cover elements contain each of the first $N$ iterates. Subadditivity gives a finite rate for each fixed cover:

$$\begin{array}{rl}{d}_{\mathcal{U}}& =\underset{N\to \infty }{\lim }\frac{D({\mathcal{U}}^{[0,N)})}{N}\\ & =\underset{N\ge 1}{\inf }\frac{D({\mathcal{U}}^{[0,N)})}{N}.\end{array}$$

The mean dimension is $\mathrm{mdim}(X,h)={\sup }_{\mathcal{U}}{d}_{\mathcal{U}}$. Its supremum may be infinite even though every individual-cover rate is finite. It measures the dimension required per orbit coordinate, rather than ordinary dimension of $X$.

For a finite simplicial complex, a point has barycentric coordinates ${\lambda }_v\ge 0$ summing to one. Its carrier is the simplex formed by the vertices with positive coordinates. A subordinate partition of unity turns a cover into a map to its nerve, a finite simplicial complex whose simplices correspond to intersecting cover members. This connects covering order to the dimension of a target complex.

🏷️ Preliminaries: crossed products and comparison

The crossed product $A=C(X){⋊}_h\mathbb{Z}$ is the operator algebra generated by continuous functions and a unitary implementing the dynamics. Finite sums ${\sum }_{|\ell |\le M}{a}_{\ell }{u}^{\ell }$ are dense. In an orbit representation, functions act diagonally and $u$ shifts orbit coordinates; such a sum has propagation at most $M$.

For positive matrix-valued elements, Cuntz subequivalence $a\precsim b$ means that $v_j^{\ast }b{v}_j\to a$ in norm after placing the elements in compatible matrix corners. It is an approximate analogue of embedding the support of $a$ into that of $b$.

For a normalized tracial state $\tau$, define

$$d_{\tau }(a)=\underset{m\to \infty }{\lim }(\tau \otimes \mathrm{Tr})(a^{1/m}),$$

with the matrix trace unnormalized. A number $r\ge 0$ is a comparison bound when

$$\begin{array}{c}{d}_{\tau }(a)+r<d_{\tau }(b)\text{for every }\tau \\ ⟹a\precsim b.\end{array}$$

The radius of comparison $\mathrm{rc}(A)$ is the infimum of all finite comparison bounds, or infinity if none exist. It measures how much uniform tracial rank surplus guarantees comparison. The infimum need not itself be a comparison bound.

Invariant probability measures yield traces on this crossed product. For a positive matrix function $P\in M_q(C(X))$ and the invariant measure associated with $\tau$,

$$d_{\tau }(P)={\int }_X\mathrm{rank}P(x)d\mu (x).$$

This formula is the bridge from tracial comparison to vector-bundle ranks.

🏷️ The main equality

Main manuscript claim

For every minimal homeomorphism of an infinite compact metrizable space,

$$\begin{array}{c}\mathrm{rc}(C(X){⋊}_h\mathbb{Z})\\ =\frac{1}{2}\mathrm{mdim}(X,h)\end{array}$$

in the extended nonnegative reals. No finite-dimensionality assumption on $X$ is made. (OpenAI, 2026b)

The manuscript imports Niu’s upper bound with the factor one half. Its new lower-bound argument proves $d_{\mathcal{U}}\le 2r$ for every finite cover and every finite comparison bound $r$. Taking the supremum over covers and infimum over comparison bounds gives the reverse inequality, including infinity.

The proof is by contradiction. Starting from $d_{\mathcal{U}}>2r$, comparison supplies a bundle embedding. The topological companion converts it into boundary compression. Reconstruction then gives an orbit-cover rate strictly below its defining infimum.

🏷️ Cube slots and the topological companion

A slot is a whole vector in $[-1,1{]}^t$. For $n$ slots, set

$$C(t,n)=([-1,1{]}^t{)}^n.$$

A boundary slot belongs to $\partial [-1,1{]}^t$; at least one scalar coordinate has magnitude one. Preserving it means fixing the entire vector.

Sparse parameters $z\in ([-2,2{]}^t{)}^n$ have at most $\lfloor an\rfloor$ nonzero slots. Denote their compact union of coordinate cubes by $Z_{an}(t,n)$.

Fix an integer $p\ge 1$ and put $t=2p+1$. The companion constructs a small embedded torus $L\cong ({\mathbb{T}}^2{)}^p\subset (-1/32,1/32{)}^t$ with a framed normal line. It carries a rank-$p$ complex vector bundle $H$, a sum of one degree-one line bundle from each two-torus factor, embedded in a trivial rank-$2p$ bundle.

For a continuous $f:Y\to C(t,n)$, the torus locus is

$$\begin{array}{c}\mathcal{T}=\{(y,z):z\in Z_{an}(t,n),\\ f(y)-z\in L^n\}.\end{array}$$

Its slot bundle is the direct sum of the $n$ pulled-back copies of $H$, of rank $pn$.

Torus compression, as stated in the companion

Let $0<a<1/16$ and $Y$ be compact metrizable. If the slot bundle has one continuous embedding into a trivial bundle of rank $K<2pn$ over the entire torus locus, then for arbitrarily large integers $u$ there is a continuous map $G:Y^u\to C(t,un)$ with at least $\lceil (a/8)un\rceil$ boundary slots at every point. Every slot that was already on the boundary in the product map $f^{\times u}$ is fixed exactly. (OpenAI, 2026a)

The embedding must remain continuous when the support of the sparse parameter changes. Separate embeddings on each fixed-support piece do not meet the hypothesis. The locus may be empty; if it is nonempty, the hypothesis already forces $K\ge pn$.

The strict upper inequality matters. At $n=1$ and $K=2p$, the identity cube map satisfies the canonical bundle embedding. A compression with a boundary slot everywhere would map a product cube into its total boundary while fixing original boundary slots. Its boundary restriction is homotopic there to the identity, hence has degree one, but extension over the cube forces degree zero.

🏷️ Cobordism, coherent products, and padding

The topological input uses complex cobordism, a generalized cohomology theory denoted $MU$. A stable class is a homotopy class after allowing suspension; an $MU$ class records its image under the unit from stable cohomotopy to complex cobordism.

A relative point test for $f$ collapses the region where $f(y)-z$ stays away from the origin and records the remaining difference map into a sphere. Retaining this relative support allows later perturbations to be small.

The bundle hypothesis forces the test to vanish in $MU$. List the $MU$ first Chern classes of its $r=pn$ line summands as $x_i$. Each comes from a two-torus, and the connective theory has $M{U}^4(T^2)=0$, so $x_i^2=0$. A complementary bundle of rank $K-r<r$ has total $MU$ Chern polynomial

$$c_v^{MU}(F)=\prod _{i=1}^r(1-x_{i}v).$$

Its coefficient of degree $r$ must vanish by rank, forcing ${\prod }_i{x}_i=0$ after pullback. A framed Thom class on a neighborhood of the torus converts this into vanishing of the relative point test. Absolute vanishing alone would omit the support needed for avoidance.

The companion’s filtered-product theorem then supplies a single stable nullhomotopy on a union of products where a fixed positive proportion of factors are constrained. It uses coherent lifts and quantitative nilpotence; separate nullhomotopies on individual product pieces would not ensure agreement on intersections. These deep compatibility and nilpotence arguments are treated as manuscript inputs here.

Stable avoidance also requires a dimension range. The padding construction addresses this without assuming a dimension bound on $Y$: at each stage it keeps only the previous output, whose dimension is bounded by its ambient cube, together with one new batch of inputs. A fixed initial padding makes every stage fit the avoidance range. Boundary restoration fixes old boundary vectors, so previous gains persist.

For arbitrary compact metrizable $Y$, a finite polyhedral model retains every coordinate of $f$ exactly and adds finitely many coordinates needed for the bundle embedding. This exact factorization preserves the original-boundary clause on pullback; an arbitrary close approximation of $f$ would not suffice.

🏷️ How comparison constructs the bundle embedding

Let a long orbit block have length $k$, and let its nerve dimension be $m=D({\mathcal{U}}^{[0,k)})$. After barycentric subdivision, color the nerve vertices with $m+1$ colors, distinct on each simplex. Group colors in packets of $2p+2$; the number of full packets is

$$g=\lfloor \frac{m+1}{2p+2}\rfloor .$$

Each packet supplies a slot map. The factor one half enters through $p/(2p+2)\to 1/2$: $p$ complex line-bundle ranks correspond to a packet of $2p+2$ real cover colors.

Open dynamical towers allocate length-$k$ blocks on a base $B$ with invariant measures close to $1/k$. Choose a target rank $R$ with $(R-pg)/k>r$. Positive matrix functions $P_0,P_1$ have ranks at most $pg$ and exactly $R$ on the relevant base, giving

$$d_{\tau }(P_0)+r<d_{\tau }(P_1).$$

Comparison produces $P_0\precsim P_1$. Approximate one witness by a finite Fourier sum $T_0$ so that

$$\Vert T_0^{\ast }{P}_1{T}_0-P_0\Vert <1/2.$$

In an orbit window, the resulting finite matrix $W$ is injective on the whole sum of selected genuine projection ranges:

$$\Vert W\xi {\Vert }^2\ge \frac{1}{2}\Vert \xi {\Vert }^2.$$

Checking each fiber separately would not establish this simultaneous injection.

Continuity across alignment changes and sparse supports requires additional work. Two fixed linear maps compress the output coordinates and the exceptional slot coordinates. A generic linear map can be injective on every coordinate subspace supported on at most a fixed number of indices: there are only finitely many forbidden rank conditions. It need not be injective on the sum of all those possible subspaces.

The two resulting maps together embed the full slot bundle. Tower, sparse-parameter, and finite-propagation losses fit inside a strict budget giving target rank less than $2pn$. The finite propagation range is fixed before the orbit window grows. (OpenAI, 2026b)

🏷️ Exact reconstruction of the cover

A packet consists of $t+1$ color weights $w_j$ of total mass $M_j$. When $M_j>0$, the proportions $w_j/M_j$ lie in a simplex; identify that simplex with the slot cube by a boundary-preserving homeomorphism.

Division fails at zero mass. Choose $\theta >0$ small and let $b$ be the simplex barycenter. Keep $q_j=w_j/M_j$ when $M_j\ge \theta$; for $M_j\le \theta$, use

$$q_j=b+(w_j-M_{j}b)/\theta .$$

After compression, multiply the new simplex vector by $\eta (M_j)M_j$, where $\eta$ is zero below $\theta$ and one above $2\theta$. Thus any colors introduced by regularization occur only in packets subsequently discarded. Leftover colors retain their weights, and the total retained mass satisfies $Z\ge 1-2g\theta >0$.

If an original color is absent in a retained packet, its original simplex point is on the boundary. Exact boundary preservation fixes that point, so the missing color cannot reappear. Normalizing by $Z$ therefore yields a nerve point whose positive vertex coordinates were already positive before reconstruction.

Each compressed boundary packet removes at least one of its colors. If $s$ packets succeed, the new carrier dimension is at most $m-s$. This is measured against the common bound $m$; it is not necessarily a decrease of $s$ from the original carrier dimension.

🏷️ From pointwise savings to one smaller target

Boundary compression takes place on a large Cartesian power. Pull it back along actual consecutive orbit-window starts. Most macro-intervals have a continuous alignment label; on them, reconstruct the internal blocks and retain two unchanged edge blocks so the entire interval remains covered.

Where alignment weakens, combine the aligned description with an unchanged description by a topological join. A join interpolates continuously between its two branches and adds one dimension when both weights are positive. Its extra cost occurs only on a uniformly small fraction $b_0$ of macro-intervals.

For macro-length $k{H}_0$, the manuscript’s rate budget is

$$\begin{array}{rl}k{H}_0{E}_k& =(H_0+1)m\\ & +b_0(H_{0}m+1)\\ & -(a/8-b_0)(H_0-1)g.\end{array}$$

The first term pays for the reconstructed and edge blocks. The second pays for exceptional joins. The last records retained successful slots.

Starting from $d_{\mathcal{U}}>2r$, choose $p$, the small losses, and $H_0$, then fix $k$ so that both the strict bundle-rank inequality and $E_k<d_{\mathcal{U}}$ hold. The orbit-window length is fixed before invoking arbitrarily large product powers.

All image points then lie in the low-dimensional cells of one finite target complex. Their union with faces is an actual subcomplex. Its open-star cover pulls back to a refinement of the full orbit cover, because reconstruction preserved every original positive-coordinate subordination relation.

For a suitable total window length $L$, this gives

$$\frac{D({\mathcal{U}}^{[0,L)})}{L}\le E_k<d_{\mathcal{U}},$$

contradicting the infimum defining $d_{\mathcal{U}}$. A varying pointwise carrier bound without a common finite complex would not complete this step.

🏷️ The zero endpoint and scope

The main manuscript also records, for these same minimal infinite systems, equivalence of zero mean dimension, the small boundary property, Jiang–Su stability of the crossed product, finite nuclear dimension, and nuclear dimension at most one. The small boundary property means that sufficiently small neighborhoods can have boundaries visited with zero uniform asymptotic orbit frequency.

Jiang–Su stability is absorption of the standard algebra $\mathcal{Z}$. Nuclear dimension measures finite-dimensional completely positive approximations with a bounded number of orthogonality-preserving return pieces. These regularity implications use established external theorems; the new equality supplies the direction from comparison at zero back to zero mean dimension.

If mean dimension is infinite, the individual-cover inequality excludes every finite comparison bound. No exchange of infinite quantities or attainment of the comparison infimum is required.

🏷️ Key steps and proof outlines

Padding pays for dimension only once per new batch

The companion’s boundary-compression proof must handle a parameter space $P$ whose dimension may greatly exceed the number of cube coordinates. Taking larger Cartesian powers alone does not improve that ratio. Instead, choose a batch of $\ell$ inputs with $b=n\ell$ output slots and a guaranteed gain $H=\lfloor ab/4\rfloor$. Add $B$ initial padding batches, with

$$tb(B+1)\ge \max \{20,2\ell \dim P\}.$$

At stage $j$, retain only the previous output, a point in the cube region with at least $(j-1)H$ boundary slots. The next problem has domain

$$W_j=Q_{(j-1)H}(t,b(B+j-1))\times P^{\ell },$$

where $Q_s(t,N)$ denotes the region with at least $s$ boundary slots. Its dimension is bounded by the old ambient cube dimension plus $\ell \dim P$, rather than the dimension of all past inputs. (OpenAI, 2026a)

The relative nullhomotopy also has an explicit compatibility argument. A sparse parameter with at most $jH-1$ nonzero slots must either use at most $(j-1)H-1$ old slots or at most $H-1$ new slots. In the first case, one old boundary slot receives zero parameter, so the old difference map is at distance at least one from zero and is collapsed to the basepoint. In the second, the new batch’s point test is stably null. On the intersection both descriptions agree at the basepoint. The closed pasting lemma therefore produces one relative map on the union, with a single nullhomotopy; separate piecewise vanishings would not suffice.

Writing $D_j=tb(B+j)$, the excluded sparse set has codimension $q_j=D_j-t(jH-1)>7D_j/8$. Padding ensures $\dim W_j<3D_j/2$ and $D_j\ge 20$, hence

$$2q_j-5>\frac{7}{4}{D}_j-5\ge \frac{3}{2}{D}_j>\dim W_j.$$

This is the stable-avoidance range. Boundary restoration then gives $jH$ boundary slots while fixing all old ones. Since $ab\ge 16$, one has $H/b\ge 3a/16$; for $j\ge 2B$, the retained fraction is at least $(2/3)(3a/16)=a/8$. The fixed padding loss disappears as more batches are added.

Reconstruct vertex weights without introducing new vertices

The main manuscript’s “Cover reconstruction” turns boundary slots into a lower-dimensional target. A subtle point is assigning modified color weights back to vertices when a color’s original weight approaches zero. Let $c_l$ be the old total weight of color $l$, ${\hat{c}}_l$ its modified weight, $Z={\sum }_l{\hat{c}}_l>0$, and ${\lambda }_v$ the old weight of a vertex of color $l$. Define

$${\tilde{\lambda }}_v=\{\begin{array}{ll}\frac{{\lambda }_v}{c_l}\frac{{\hat{c}}_l}{Z},& c_l>0,\\ 0,& c_l=0.\end{array}$$

The first ratio lies in $[0,1]$. Exact boundary preservation proves $c_l=0\Rightarrow {\hat{c}}_l=0$, so the second ratio tends to zero at a vanishing color. Thus the displayed piecewise formula is continuous, even though its first factor alone need not have a limit. The new weights have support contained in the original carrier and still sum to one. (OpenAI, 2026b)

Each successful packet has at least one missing color after compression. Since packets are disjoint, $s$ successes leave at most $m+1-s$ positive colors, hence carrier dimension at most $m-s$. This is a bound relative to the global nerve dimension $m$, not a claim of deleting $s$ vertices from every original carrier.

Finally, subordination survives because no new vertex appears. The reconstructed images lie in low-dimensional cells of one fixed finite complex; adjoining their faces gives an actual subcomplex. Its open-star cover therefore pulls back to a refinement of the orbit cover. Together with the manuscript’s rank and alignment budgets, this yields an orbit-cover rate smaller than its defining infimum. The topological vanishing, stable-avoidance and comparison interfaces retain their stated hypotheses; these calculations explain how their outputs compose into the contradiction.

🏷️ Links and reading scope

• Family 291 discusses nuclear approximation and absorption; the present family links those regularity properties to a dynamical dimension.
• Family 301 defines Cuntz ranks as an invariant; here the uniform rank gap measures comparison rather than classification.
• The collection guide places the pair in the operator-analysis sequence.

The selected reading covers both introductions, the main compression interface, cover reconstruction, lower-bound assembly and regularity argument, selected dynamical comparison passages, and the companion’s padding, torus criterion, point-test and boundary-restoration passages. The full coherent-product and local-duality proofs and all external inputs were not independently audited.

🏷️ Manuscripts and verification scope

• Filtered products and boundary-preserving compression in complex cobordism — pinned manuscript and companion source.
• Radius of comparison equals half the mean dimension — pinned manuscript and companion source.

No family scope page was present at lean/docs/302.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 2026a. Filtered products and boundary-preserving compression in complex cobordism.

🐻  OpenAI 2026b. Radius of comparison equals half the mean dimension.