Toms–Winter: Strict Comparison Implies Z-Stability

Toms–Winter: Strict Comparison Implies Z-Stability

OpenAI's claim

Manuscript 1 (family 291, 2026-09-23): A separable nuclear C*-algebra that is simple and non-elementary is claimed to absorb Jiang--Su under extended-functional strict comparison, including nonunital and unbounded-trace cases; semigroup divisibility also yields a nonsimple criterion. Manuscript 2 (family 291, 2026-09-23): The paper claims uniform Gamma from real rank zero in the tracial-completion ultrapower for simple unital nuclear stably finite algebras. Comparison gives absorption, also under specified bounded-trace hypotheses in the stably projectionless setting. Manuscript 3 (family 302, 2026-09-25): Whenever an infinite compact metrizable system is minimal under a homeomorphism, its crossed-product comparison radius is claimed to be half its mean dimension, including infinite values; zero forces nuclear dimension at most one.

Mathematical statement of the claimed result

Manuscript 1.

Let $\operatorname{Cu}(A)$ be the ordered Cuntz semigroup of positive elements of $A\otimes\mathcal K$, modulo mutual norm-approximation comparison. For separable simple nuclear non-elementary $A$, OpenAI claims $A\simeq A\otimes_{\min}\mathcal Z$ if, for every $x,y\in\operatorname{Cu}(A)$, inequalities $\varphi(x)\le\varphi(y)$ for all additive order/supremum-preserving functionals $\varphi:\operatorname{Cu}(A)\to[0,\infty]$, strict whenever $0<\varphi(y)<\infty$, imply $x\le y$. It further claims absorption for every nonzero separable nuclear $A$ whose Cuntz semigroup satisfies both \[ (l+1)x\le ly\ \Longrightarrow\ x\le y\quad(l\ge1),\qquad \forall x,\ N\ge1\ \exists u:\ Nu\le x\le(N+1)u. \] In that case a unital $\mathcal Z\to (A_\omega\cap A')/\operatorname{Ann}(A)$ is asserted for every free ultrafilter, where $A_\omega$ is the norm ultrapower and $\operatorname{Ann}(A)=\{z:zA=Az=0\}$. Finally it claims \[ A\simeq A\otimes_{\min}\mathcal Z\quad\Longleftrightarrow\quad \operatorname{Cu}(A)\simeq\operatorname{Cu}(A\otimes_{\min}\mathcal Z) \] for all separable nuclear $A$, allowing an abstract ordered-semigroup isomorphism rather than only the canonical first-factor map. Here $\mathcal Z$ is the Jiang--Su algebra, and non-elementary excludes compact-operator algebras.

Manuscript 2.

Let $A$ be unital, separable, simple, infinite-dimensional, nuclear and stably finite, with nonempty tracial state space $T(A)$. Let $B=\overline A^{T(A)}$ be its completion on the norm-bounded unit ball for $\|a\|_{2,T(A)}=\sup_\tau\tau(a^*a)^{1/2}$, and $B^\omega$ its uniform-tracial ultrapower for a free ultrafilter. If $B^\omega$ has real rank equal to zero when given the quotient $C^*$-norm, OpenAI claims a central projection $p$ satisfying \[ \lambda(px)=\tfrac12\lambda(x)\qquad(x\in B,\ \lambda\in\Lambda_\omega), \] where $\Lambda_\omega$ consists of ultralimits of sequences of designated trace extensions. It also claims every tracial state on $B$ is the unique extension of a trace of $A$, with $|\sigma(b)|\le\|b\|_{2,T(A)}$. For a unital nuclear inclusion $A\subset D$ with $D$ real rank zero and nonempty traces, assume arbitrarily small uniformly tracial full projections in $D$. It claims finite-set approximate central halving projections: given finite $K\subset\mathcal U(A)$ and $\varepsilon>0$, some projection $P\in D$ satisfies $\sup_\tau\|[P,x]\|_{2,\tau}<\varepsilon$ and $\sup_\tau|\tau(Px)-\tau(x)/2|<\varepsilon$ for $x\in K$. The independent unital result takes an infinite-dimensional nuclear complex $C^*$-algebra $A$, assumed simple, separable and unital. For positive $a,b\in A\otimes\mathcal K$, define $d_\tau(a)=\lim_{n\to\infty}(\tau\otimes\operatorname{Tr})(a^{1/n})$, using the unnormalized matrix trace. Assume \[ b\ne0,\quad d_\tau(a)<d_\tau(b)\quad(\tau\in T(A))\quad\Longrightarrow\quad a\precsim b. \] When $T(A)$ is empty, this premise still requires $a\precsim b$ for every positive $a$ and every nonzero positive $b$. OpenAI claims $A\simeq A\otimes\mathcal Z$; when traces exist it also constructs uniform property $\Gamma$. The separate nonunital theorem starts with a nonzero nuclear complex $C^*$-algebra $A$ that is simple and separable, with no nonzero projection in $A\otimes\mathcal K$. Write $T_{\mathrm{lsc}}(A)$ for densely finite lower-semicontinuous traces, and $T_1(A)$ for those bounded traces having norm one. Require every trace in $T_{\mathrm{lsc}}(A)$ to be bounded on the positive unit ball, and require $T_1(A)$ to be nonempty and weak-$*$ compact. The precise comparison premise is \[ \forall a,b\in(A\otimes\mathcal K)_+,\quad b\ne0,\quad \left[d_\tau(a)<1\text{ for every }\tau\in T_{\mathrm{lsc}}(A)\text{ with }d_\tau(b)=1\right] \ \Longrightarrow\ a\precsim b. \] Under these hypotheses it claims uniform property $\Gamma$ and $A\simeq A\otimes\mathcal Z$. Here $a\precsim b$ means $x_n^*bx_n\to a$ in norm for stabilized elements $x_n$, and $\mathcal Z$ is the Jiang--Su algebra.

Manuscript 3.

Let $h$ be a homeomorphism of $X$ whose orbits are all dense, where $X$ is compact, infinite and metrizable. OpenAI claims \[ \operatorname{rc}(C(X)\rtimes_h\mathbb Z)=\tfrac12\operatorname{mdim}(X,h)\quad\text{in }[0,\infty]. \] For positive stabilized elements define $d_\tau(a)=\lim_m\tau(a^{1/m})$. The radius of comparison is the infimum of $r\ge0$ such that $d_\tau(a)+r<d_\tau(b)$ for every normalized trace implies Cuntz subequivalence $a\precsim b$. Mean dimension is the asymptotic covering dimension per iterate, defined from joins of finite open covers. Both finite and infinite mean dimensions are included.

Outline of the argument

Manuscript 1 (family 291, 2026-09-23): Trace transfer into small hereditary supports followed by spectral rounding establishes full Cuntz-semigroup divisibility. Pure-state compression produces norm-visible matrix cones, while localized nuclear approximations produce trace-visible ones. Comparison packs their defects into a central corner, constructing dimension-drop maps whose lifted intertwining relations give Jiang--Su absorption.

Manuscript 2 (family 291, 2026-09-23): Nuclear averaging and scalar block compressions preserve coefficient traces while packing small projections with controlled commutators. A separate comparison construction rounds ranks, forms clipped sums of centered symmetries, and extracts Gaussian trace partitions. In the projectionless case, a combined overlap-and-model cost controls small splittings before reindexing amplifies them to exact central partitions.

Manuscript 3 (family 302, 2026-09-25): Finite-support comparison witnesses are converted to bundle embeddings, which feed the companion's boundary-preserving compression theorem. Packet reconstruction translates that compression into a smaller cover without introducing absent colors. Applying the construction on aligned orbit blocks contradicts any larger cover-growth rate, providing the missing lower bound; established upper bounds and regularity results finish the equality and zero-dimensional consequences.

Formal verification

For family 291, the scope page documents the following selections. The selection gives uniform Gamma from real rank zero of the designated tracial ultrapower. The documented comparison scope also treats Jiang--Su absorption and Gamma under strict comparison, with additional trace and target-rank conditions in the projectionless case. Manuscript 1 is not among the papers selected by that page, so the family link does not formalize its main conclusion. It does not establish formal coverage of the complete problem clause. The documented statement covers the main claim of manuscript 2. Its selected declaration also covers the clause used for this problem. Comparator files: UniformGamma.lean. Declarations: OAI.ComparatorModel.CurrentMain.main.

Family 302 has no Lean scope page in this snapshot; no formalization of these manuscript results is listed.

This site has not executed the Lean files or checked their dependencies. These descriptions report the scope of the linked declarations, rather than a new verification of the proofs.

Status caveat

OpenAI attributes these results to an internal model. Its release notice cautions that results lacking full formalization can contain mistakes. This site has not independently assessed the proofs; a solved label records the scope of the reported claim.

Sources

• OpenAI, Cuntz comparison and Jiang–Su absorption, 2026-09-23; citation key OAI:Cuntz-comparison-and-Jiang-Su-absorption-September-23-2026. PDF (GitHub); PDF (direct); Manuscript page and citation; Latest version; Statement source: ext:thm:main; Statement source: thm:condition-C; Statement source: cor:canonical; Statement source: cor:abstract-cu.
• OpenAI, Tracial projection methods and uniform property $\Gamma$, 2026-09-23; citation key OAI:Tracial-projection-methods-and-uniform-property-Gamma-September-23-2026. PDF (GitHub); PDF (direct); Manuscript page and citation; Latest version; Statement source: r:thm:main; Statement source: r:cor:completion-traces; Statement source: r:thm:finite; Statement source: u:thm:main; Statement source: n:thm:main.
• Family 291: Lean scope.
• Comparator UniformGamma.json.
• Comparator UniformGamma.lean.
• OpenAI, Radius of comparison equals half the mean dimension, 2026-09-25; citation key OAI:Radius-of-comparison-equals-half-the-mean-dimension-September-25-2026. PDF (GitHub); PDF (direct); Manuscript page and citation; Latest version; Statement source: thm:main.
• OpenAI catalogue entry.

Toms–Winter: Strict Comparison Implies Z-Stability

Definitions and mathematical statement

Let $A$ be a separable simple nuclear C*-algebra, not isomorphic to an algebra of compact operators; no unit is required. Its Cuntz semigroup $\operatorname{Cu}(A)$ consists of positive elements of $A\otimes\mathcal K$ modulo mutual Cuntz subequivalence, where $a\precsim b$ means $x_j^*bx_j\to a$ for some sequence. In the selected strict-comparison convention, $x\le y$ whenever every additive order-preserving functional $\varphi:\operatorname{Cu}(A)\to[0,\infty]$, preserving zero and increasing sequential suprema, satisfies $\varphi(x)\le\varphi(y)$, with strict inequality whenever $0<\varphi(y)<\infty$. The conjecture asserts \[ A\text{ has strict comparison}\quad\Longrightarrow\quad A\cong A\otimes_{\min}\mathcal Z, \] where $\mathcal Z$ is the Jiang--Su algebra. Separability means a countable norm-dense subset; simplicity means no nontrivial closed two-sided ideals; nuclearity is the completely positive matrix-factorization approximation property. No UCT, finite nuclear dimension, or extra trace-space hypothesis is appended.

Formulation and concept references: [S1].

Short English statement

Does strict comparison force every separable simple nuclear non-elementary C*-algebra to absorb the Jiang--Su algebra, including nonunital cases?

Research attempt

The remaining implication and its quantifiers.

The exact comparison convention in the question is the one printed after Conjecture 5.1 of [S1]. Its functionals can take the value $+\infty$; the strict inequality is required only at finite, strictly positive values of the comparison target. Replacing this by a condition on a single normalized trace would change the problem, especially in the nonunital setting.

The inspected Theorem 5.6 of [S1] proves the implication in the unital tracial case with the additional hypothesis of uniform property $\Gamma$. That additional hypothesis is not part of this catalogue entry. The argument below gives explicit comparison calculations, constructs uniform central trace divisions in an infinite tensor-tail subcase, and identifies the missing passage from order information to approximately central structure. It does not assume finite nuclear dimension, UCT or a regular trace boundary in order to declare the target solved.

What Cuntz comparison means in the elementary test case.

For positive compact operators $a,b$ on a separable infinite Hilbert space, Cuntz subequivalence is exactly support-rank comparison: \[ a\precsim b\quad\Longleftrightarrow\quad \operatorname{rank}a\leq\operatorname{rank}b, \tag{380.1} \] where infinite rank is allowed. Here rank means the dimension of the closure of the range, not algebraic rank on a dense subspace.

To prove sufficiency, truncate the spectral decomposition of $a$ to finitely many eigenvalues. If the truncation has rank $r$, choose $r$ positive eigenvectors of $b$, map the corresponding eigenvectors of $a$ to them and rescale by square roots of the eigenvalue ratios. This finite-rank operator $x$ satisfies $x^*bx=a_{\rm trunc}$. The spectral truncations converge to $a$ in norm, proving the subequivalence. The operators $x$ may have large norm; the definition does not require a uniform norm bound.

For necessity, if $b$ has finite rank $r$, then every $x^*bx$ has rank at most $r$. A positive operator with at least $r+1$ positive eigenvalues cannot be a norm limit of such operators: on the span of those eigenvectors its quadratic form is bounded below by their positive minimum, whereas every rank-$r$ operator has a kernel vector in that $(r+1)$-dimensional subspace. Thus rank cannot increase in the limit. If $b$ has infinite rank, the necessary inequality is automatic. This proves $380.1$.

Since $M_d\otimes\mathcal K\cong\mathcal K$, it follows that $\operatorname{Cu}(M_d)=\{0,1,2,\ldots,\infty\}$ with the usual order and addition. Every finite rank $r$ is a sum of $r$ rank-one classes, and increasing ranks have supremum infinity.

The non-elementary exclusion prevents an actual counterexample.

The matrix algebra $M_d$ satisfies the printed strict-comparison condition. To check the implication, use the rank functional $\varphi(r)=r$. If $y$ is finite and positive, the premise forces $x<y$; if $y=0$, it forces $x=0$; if $y=\infty$, every $x$ is already at most $y$. Thus the required order conclusion holds in every case in which the functional premise holds.

Nevertheless $M_d$ does not absorb the infinite-dimensional Jiang--Su algebra: $M_d\otimes\mathcal Z$ is infinite-dimensional, whereas $M_d$ is finite-dimensional. The same stabilization calculation explains the elementary compact-operator case excluded by the statement. These are not counterexamples under the selected hypotheses; they show that deleting the non-elementary exclusion would make the implication false.

The test also warns against equating comparison with exact divisibility. In $\operatorname{Cu}(M_d)$ the rank-one class cannot be twice a nonzero class. Even the weaker request $kz\leq1\leq(k+1)z$ has no solution for $k\geq2$: the left inequality forces $z=0$, contradicting the right. Order detected by functionals does not automatically create intermediate semigroup elements. Non-elementarity is necessary to exclude this obstruction, but merely noting its necessity is not a construction of the missing elements in a general algebra.

Strictly ordered projections versus central comparison.

In the unital stably finite simple nuclear setting, normalized trace functionals give dimension functions $d_\tau(a)=\lim_{m\to\infty}\tau(a^{1/m})$; for a projection $p$, this equals $\tau(p)$. The correspondence with the relevant Cuntz functionals is part of the framework recorded in [S1]. Strict comparison then turns uniform strict dimension inequalities into Cuntz subequivalence. Its output, however, is only a sequence of witnesses $x_j$ with $x_j^*bx_j\to a$.

There is no approximate commutation requirement on those $x_j$. The issue is visible in $M_3$: the rank-one projection $e_{11}$ is subequivalent to $e_{22}+e_{33}$, witnessed by $e_{21}$ in the convention $x^*bx=a$. Yet a scalar witness cannot do this, since $c^*(e_{22}+e_{33})c$ is supported orthogonally to $e_{11}$. Its distance from $e_{11}$ is at least one. A witness approximately commuting with all of $M_3$ is close to a scalar, so it cannot make this error tend to zero.

For the last assertion, average over the finite clock-and-shift unitaries $U^rV^s$, $0\leq r,s<3$. Their conjugation average of an element $x$ is $\operatorname{tr}_3(x)1$. If all the corresponding commutators are at most $\varepsilon$, this average differs from $x$ by at most $\varepsilon$. Thus the obstruction is quantitative. The example is elementary and excluded from the final conjecture, but it disproves a purely formal inference from comparison witnesses to central comparison witnesses.

An explicit source of central trace divisions.

Let $U=\bigotimes_{j=1}^\infty M_2$ be the infinite tensor product with its canonical units, and let $B$ be any unital $C^*$-algebra. Set $A=B\otimes U$. We prove the following approximation property directly. For a finite set $\mathcal F$ in $A$, an integer $n\geq2$ and $\varepsilon>0$, there are orthogonal projections $e_1,\ldots,e_n$ in $A$, summing to one, such that \[ \|[e_i,a]\|<\varepsilon, \qquad \left|\tau(ae_i)-\frac1n\tau(a)\right|<\varepsilon \tag{380.2} \] for every $a\in\mathcal F$, every tracial state $\tau$ on $A$ and every $i$.

Approximate $\mathcal F$ in norm within $\delta$ by elements $a_0$ of $B\otimes\bigotimes_{j=1}^mM_2$. Take a later block of $k$ tensor factors, isomorphic to $M_D$ with $D=2^k$. Partition its standard basis into $n$ pieces whose sizes $r_i$ differ by at most one. The diagonal projections onto these pieces satisfy \[ \sum_ie_i=1,\qquad e_ie_j=0\ (i\ne j),\qquad \left|\frac{r_i}D-\frac1n\right|\leq\frac1D. \] They commute exactly with every $a_0$.

For any tracial state on an algebra containing a commuting matrix factor, its restriction to the first factor times that matrix factor is the first-factor trace tensored with normalized matrix trace. Indeed off-diagonal matrix units have trace zero after conjugation by diagonal unitaries, and all diagonal matrix units have equal trace after permutation conjugation. Applying this to $a_0e_i$ gives $\tau(a_0e_i)=(r_i/D)\tau(a_0)$, uniformly over all traces. Consequently \[ \|[e_i,a]\|\leq2\delta,\qquad \left|\tau(ae_i)-\frac1n\tau(a)\right| \leq2\delta+\frac{\|a_0\|}{D}. \tag{380.3} \] Choose $\delta$ small and then $k$ large. This proves $380.2$. The rounding of ranks is handled explicitly; powers of two need not be divisible by the prescribed integer $n$.

Passing this construction to a central sequence algebra.

If $A$ is separable, choose an increasing sequence of finite sets whose union is norm dense. Apply $380.2$ with errors tending to zero to produce sequences $e_i^{(j)}$. In the quotient $A_\infty=\ell^\infty(A)/c_0(A)$ their classes commute with every constant element of $A$, are orthogonal projections, and sum to the unit. Along any ultrafilter and any sequence of tracial states, the trace-splitting equalities pass to the limit: \[ \tau_\omega(ae_i)=\frac1n\tau_\omega(a) \quad(a\in A). \tag{380.4} \] The construction also descends to the tracial ultrapower, since norm-null sequences are uniformly trace-two-norm null. Thus it supplies, in this tensor-tail setting, the uniform property $\Gamma$ used in [S1]. The trace estimate is simultaneous over all traces; separate constructions for each individual trace would not have implied $380.4$.

If this $A$ is simple, separable, nuclear and unital, has a nonempty tracial state space, and has strict comparison, the inspected Theorem 5.6 of [S1] now implies $\mathcal Z$-stability. This is a complete conditional subcase obtained by combining an explicit verification of the additional hypothesis with a named external theorem. The deep passage from uniform central divisions and comparison to absorption is attributed to that theorem, not reproved by tensor bookkeeping.

Why individual traces do not give a uniform construction.

Suppose for every trace $\tau$ one found elements splitting its values approximately. Those elements may depend on $\tau$. The property needed above has the order \[ \text{for each finite set and tolerance, choose elements that work for every trace.} \] This is different from choosing elements after fixing the trace. Even compactness of a trace space only produces finitely many locally valid choices; one still has to combine them inside the algebra while preserving orthogonality, positivity and approximate centrality. A scalar partition of unity on the trace space is not automatically a central partition inside $A$.

Likewise, strict inequalities between dimension functions do not by themselves give a positive numerical margin uniform over all traces unless the relevant functions and compactness hypotheses justify it. Dimension functions of positive elements need not be continuous affine functions in the same way that traces of a fixed element are. Passing from pointwise comparison to a central construction can therefore lose two kinds of uniformity, both across traces and across the prescribed finite set.

The infinite tensor tail avoids these difficulties for a concrete reason: the same matrix block commutes with every approximating head element, and every trace has exactly the same normalized matrix-factor formula. A general nuclear approximation by completely positive maps into matrices is not such a commuting matrix subalgebra inside $A$.

Nuclear approximation is not a multiplicative splitting.

Nuclearity provides finite-dimensional completely positive factorizations approximating finite sets in norm. It does not say that the matrix-factor maps are multiplicative or that their images commute with the original finite set. Taking matrix projections and applying a general completely positive map need not preserve projections or orthogonality. For instance the normalized trace map $M_2\to\mathbb C$ sends two orthogonal rank-one projections to the same nonzero scalar $1/2$. Thus the tail construction cannot be copied by replacing its actual matrix block with a generic nuclear factorization.

One must obtain appropriately orthogonal order-zero or central sequence structure with uniform tracial size. Theorem 5.6 encodes a route once uniform property $\Gamma$ is present. Deriving that structure from only the selected comparison hypothesis is the unproved step in this attempted approach.

The nonunital scope is not eliminated by adjoining a unit.

If $A$ is nonunital and simple, its unitization contains $A$ as a proper ideal and has scalar quotient. It is generally not simple. Applying a theorem for simple unital algebras to the unitization therefore fails its hypotheses. Replacing $A$ by a unital full corner also requires a suitable nonzero projection; stably projectionless algebras do not supply one. Neither operation is a formal way of reducing the entire printed nonunital statement to the unital tracial theorem used above.

The functional convention in the question is designed to retain these cases. A proof restricted to normalized tracial states must state its restriction explicitly, rather than treating an empty or noncompact normalized trace space as harmless.

Diagnostics and outcome.

The exact checks verify the finite-rank comparison witnesses, clock-and-shift averaging, rank partitions for tail blocks and the error bound from rounding. They do not compute an infinite Cuntz semigroup or certify $\mathcal Z$-absorption for arbitrary input algebras.

We have explained why the elementary exclusion is necessary, proved an explicit central division construction for an infinite tensor-tail family, and used the source theorem only with its additional unital tracial and uniform-$\Gamma$ hypotheses. The first unresolved step for the catalogue is constructing the requisite uniform approximately central structure from strict comparison alone, including the separate nonunital difficulties. The general implication remains unresolved by this attempt.

Sources

• S1 Castillejos, Evington, Tikuisis and White, Uniform Property Gamma, Conjecture 5.1 and following comparison definition; Schafhauser, Tikuisis and White, Nuclear C*-algebras: 99 problems, Problem XVIII. https://arxiv.org/pdf/1912.04207. Formulation source.