Half the mean dimension: an exact bridge from dynamics to operator comparison

← All digests

Family 302 · Operator algebras
Importance 5/5 · Provisional editorial rating
Sources consulted 2026-10-25

1. The question

A dynamical system can generate new geometric complexity as its orbit is observed for longer times. Mean dimension measures the number of continuous parameters needed per time step. An associated operator algebra has a different measure of complexity: how much extra tracial size is needed before one positive operator can be fitted inside another.

The Phillips–Toms conjecture predicts a precise bridge between these quantities. Family 302 claims that for every minimal homeomorphism of an infinite compact metrizable space,

$$\operatorname{rc}(C(X)\rtimes_h\mathbb Z) =\frac12\operatorname{mdim}(X,h).$$

The equality includes infinity. One manuscript proves the dynamical statement using an established upper bound; the other supplies a topological compression theorem needed for the new lower bound. The factor one half reflects the relation between complex vector-bundle rank and real covering dimension.

2. Background and essential definitions

A homeomorphism $h:X\to X$ is minimal if every orbit is dense. For infinite compact $X$, minimality rules out periodic points. Metrizability gives a countable, metric framework without requiring finite covering dimension.

For a finite open cover $\mathcal U$, join its translates over $N$ steps to obtain the orbit cover $\mathcal U^{[0,N)}$. Let $D$ be the smallest order of a finite open refinement, where order is one less than the largest number of sets meeting at one point. Then

$$\operatorname{mdim}(X,h)= \sup_{\mathcal U}\lim_{N\to\infty} \frac{D(\mathcal U^{[0,N)})}{N}.$$

It measures average covering complexity along orbits, rather than the dimension of $X$ alone.

The crossed product $C(X)\rtimes_h\mathbb Z$ combines continuous functions on $X$ with a unitary implementing the dynamics. Invariant probability measures yield normalized tracial states on this algebra.

For positive matrix elements $a,b$, Cuntz subequivalence $a\precsim b$ means that $v_j^*bv_j\to a$ in norm after placing the matrices in compatible larger corners. For a normalized trace $\tau$, define

$$d_\tau(a)=\lim_{m\to\infty}(\tau\otimes\operatorname{Tr})(a^{1/m}),$$

with the matrix trace unnormalized. The radius of comparison is the infimum of numbers $r\geq0$ such that

$$d_\tau(a)+r<d_\tau(b)\text{ for all }\tau \quad\Longrightarrow\quad a\precsim b.$$

If no finite $r$ works, the radius is infinity. It records a dimension-like obstruction left over after numerical ranks are compared.

3. Development history and previous partial results

Gromov introduced mean dimension, developed systematically by Lindenstrauss and Weiss. Toms introduced the radius of comparison in C*-algebra classification, where vector bundles show that numerical rank alone need not determine comparison.

Giol–Kerr connected these ideas by constructing minimal systems with positive mean dimension and large comparison radii. Phillips obtained a general upper bound with a larger constant and additive term. Elliott–Niu settled the zero-mean-dimension case through Jiang–Su stability.

Niu then proved the optimal upper bound

$$\operatorname{rc}(C(X)\rtimes_h\mathbb Z) \leq\frac12\operatorname{mdim}(X,h)$$

in a setting including all systems considered here. M2 credits this result; the new problem is the reverse inequality.

Hirshberg–Phillips obtained lower bounds through mean cohomological independence, detecting persistent cup products of translated cohomology classes. These gave strong estimates for particular minimal subshifts. But a general cover may carry dimensional complexity not captured by the required cohomological independence hypothesis.

M1’s topological tools draw on complex cobordism and the Devinatz–Hopkins–Smith nilpotence theorem, with quantitative connective refinements. M2 uses them to turn a comparison assumption directly into a reduction of open-cover complexity, avoiding a separate cohomological independence assumption.

A subsequent October 8 preprint of Elliott and Niu states related equalities for amenable actions with the uniform Rokhlin property and Cuntz comparison of open sets, explicitly using these topological ideas. That is later work with its own hypotheses; the pinned family’s main dynamical theorem concerns integer actions generated by one minimal homeomorphism.

4. What the new manuscripts claim

M2: exact equality and the zero endpoint

For every minimal homeomorphism $h$ of an infinite compact metrizable $X$, M2 claims the equality in Section 1 in $[0,\infty]$. No finite-dimensionality or cohomological independence condition on $X$ is assumed.

It also derives 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 neighborhoods can be chosen whose boundaries have orbit capacity zero: the largest fraction of visits to the boundary over long orbit segments tends to zero. Jiang–Su stability means absorption of $\mathcal Z$; nuclear dimension counts pieces in finite-dimensional positive-map approximations.

The zero-mean-dimension-to-regularity directions use earlier theorems. The new lower bound supplies the converse from operator-algebraic comparison. Positive or infinite mean dimension consequently forces failure of Jiang–Su stability and infinite nuclear dimension in this class.

M1: coherent vanishing on product unions

For a finite pointed CW pair $G\subseteq E$ and a map $v:E\to S^d$, suppose the restriction to $G$ becomes null in complex cobordism. Here a finite CW complex is built from finitely many cells, and complex cobordism is a generalized cohomology theory encoding stable complex-geometric information.

M1 claims that sufficiently large smash powers of $v$ are stably null on the whole union where at least a fixed positive proportion of factors lie in $G$. A smash product is a pointed product with all basepoint-coordinate subsets collapsed; stable nullity means nullity after passing to stable homotopy. The conclusion gives one compatible nullhomotopy on the union, not unrelated nullhomotopies on its pieces.

M1: boundary-preserving cube compression

Let a continuous map from any compact metrizable space have $n$ cube-valued slots of dimension $2p+1$. A specified torus locus over sparse translations carries a complex bundle of rank $pn$. If that bundle has one continuous embedding over the entire parameter locus into a trivial bundle of rank $K<2pn$, M1 obtains compression on arbitrarily large Cartesian powers.

For a sparsity parameter $0<a<1/16$, the output places at least an $a/8$ fraction of slots on their boundaries. Every slot already on the boundary is preserved as its entire vector, including coordinates that were interior within that slot. There is no covering-dimension bound on the input space.

The strict inequality $K<2pn$ is essential to this general criterion: M1 gives a cube example at equality where the required compression is impossible. The embedding must vary continuously as the sparse parameter’s support changes.

5. Why the advance matters

Importance: 5/5. If correct, the equality identifies an exact numerical operator-algebra invariant from topological dynamics across all minimal homeomorphisms in the stated class. It also makes zero mean dimension precisely the boundary of the crossed product’s Jiang–Su and finite-dimensional approximation regularity.

The new lower-bound mechanism is important beyond improving a constant. It detects arbitrary orbit-cover complexity, including infinite-dimensional spaces where a selected cohomological invariant may miss information.

M1 supplies a reusable topological principle: local or restricted vanishing must remain compatible across product intersections before it can support a global compression. Its boundary-preservation requirement is what lets the compression improve a cover without destroying its original refinement information.

6. The main idea

Assume comparison is too easy

Fix a finite comparison bound $r$ and an open cover whose mean cover rate exceeds $2r$. M2 represents a long orbit block by a finite simplicial complex. After subdivision, vertices receive colors distinct on each simplex.

Groups of $2p+2$ colors become packets whose proportions define cube-valued slots of dimension $2p+1$. Zero colors record absent members of the original cover. A useful compression must preserve these absences, or the resulting map may no longer describe a refinement.

Turn a comparison witness into a bundle embedding

Each slot carries a small torus and a rank-$p$ bundle assembled from degree-one line bundles. Niu’s open tower theorem aligns most orbit blocks with a small uniform remainder.

A finite-propagation Cuntz comparison witness gives a bundle embedding on the aligned pieces. Additional linear maps handle sparse support changes and imperfect alignment. M2 obtains the single continuous embedding over the full parameter locus required by M1, with target rank strictly below twice the source rank.

Compress coherently and reconstruct a smaller cover

M1 first converts cobordism vanishing into stable vanishing on a constrained product union. Compatible local duality then turns this into avoidance of a sparse arrangement. A padding construction removes the dimension restriction that would otherwise obstruct conversion from stable data to an actual map.

Radial normalization produces boundary slots while restoring original boundary vectors exactly. M2 interprets each new boundary slot as at least one missing color. Small packets are discarded carefully, and boundary preservation prevents any originally absent color from reappearing.

The reconstructed map lands in a complex of smaller dimension while still refining the full orbit cover. That contradicts the chosen cover rate. The ratio of $p$ complex lines to $2p+2$ colors tends to one half, producing the exact constant after controlling the other losses.

Include infinite mean dimension

The argument bounds every finite-cover rate by $2r$. Taking the supremum over covers gives mean dimension at most $2r$. If mean dimension is infinite, no finite comparison bound can exist. Combining this lower bound with Niu’s established upper bound completes the stated equality.

7. Scope, limitations, and open questions

The main theorem concerns minimal homeomorphisms of infinite compact metrizable spaces. It does not by itself settle the conjecture for every action of every amenable group, and the later cited extension retains additional dynamical assumptions.

The topological compression theorem requires a continuous bundle monomorphism over the full sparse-parameter locus, not separate embeddings at individual parameters. It gives arbitrarily large suitable powers, rather than specifying compression at every power or a universal small power.

The equality uses normalized algebra traces with unnormalized matrix traces. Changing rank normalization changes the comparison convention and can obscure the factor one half. The point at infinity is part of the theorem, not an extrapolation from finite examples.

M2 depends on M1’s compression theorem and on Niu’s upper-bound and tower results. The regularity corollary also invokes published simple-algebra theorems. These dependencies are part of the claimed result’s structure, rather than independent verifications performed for this digest.

8. Sources and further reading

• Hirshberg and Phillips, Radius of comparison and mean cohomological independence dimension, for earlier dynamical lower bounds.
• Elliott and Niu, Mean dimension and radius of comparison, for the later extension under its stated hypotheses.
• M2 gives the precise Niu upper-bound and tower inputs; M1 states the compatible product-vanishing and strict-rank compression theorems used in the new lower bound.

Source manuscripts

• M1. OpenAI, Filtered products and boundary-preserving compression in complex cobordism (September 25, 2026).
• M2. OpenAI, Radius of comparison equals half the mean dimension (September 25, 2026).

Sources consulted on 25 October 2026 at repository revision 1581d4c. Download the manuscript citations.