Nuclear ultrapowers: a finite spectral obstruction

Source and verification

This article explains family 292 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.

🏷️ Preliminaries: nuclear approximation and norm ultrapowers

A $C^{\ast }$-algebra is a norm-closed algebra of bounded operators closed under adjoints. A map between such algebras is completely positive if all its matrix amplifications preserve positivity. A $C^{\ast }$-algebra $B$ is nuclear when its identity can be approximated on finite sets by completely positive contractive maps through finite-dimensional algebras.

Nuclearity also has a tensor-product characterization. The minimal tensor product uses operators on separate Hilbert spaces; the maximal tensor product permits arbitrary commuting representations on one Hilbert space. For nuclear $B$, these two completions of $B\otimes C$ coincide for every $C$. The proof in this family uses precisely this equality.

Fix a free ultrafilter $\omega$ on the natural numbers. An assertion holds on an $\omega$-large set if its set of valid indices belongs to $\omega$. Finite intersections of such sets remain $\omega$-large; arbitrary countable intersections need not. The norm ultrapower is

$$B_{\omega }={\ell }^{\infty }(\mathbb{N},B)/I_{\omega },$$

where $I_{\omega }$ consists of sequences with ${\lim }_{\omega }\Vert b_n\Vert =0$. Thus two bounded sequences represent the same element when their operator-norm difference tends to zero along $\omega$. This is different from a tracial ultrapower, which uses a trace-derived $2$-norm. The source writes $B^{\omega }$ for this norm quotient; we use the subscript to agree with the Jiang–Su stability article.

An embedding is an injective $\ast$-homomorphism. It is unital if it preserves the identity. The Cuntz algebra ${\mathcal{O}}_2$ is generated by two isometries with orthogonal ranges adding to the identity:

$$\begin{array}{c}{s}_i^{\ast }{s}_i=1(i=1,2),\\ s_1{s}_1^{\ast }+s_2{s}_2^{\ast }=1.\end{array}$$

It is nuclear. Kirchberg–Phillips embedding theory places every separable unital exact algebra inside ${\mathcal{O}}_2$; exactness means that minimal tensoring preserves short exact sequences. Kirchberg’s norm-ultrapower question asks whether passing to $({\mathcal{O}}_2{)}_{\omega }$ accommodates every separable unital algebra. This is a norm-approximation problem, distinct from Connes’ tracial embedding problem.

🏷️ The manuscript’s explicit obstruction

For a countable discrete group $G$, its full group algebra $C^{\ast }(G)$ is the universal $C^{\ast }$-algebra generated by unitaries $u_g$ satisfying $u_g{u}_h=u_{gh}$. “Full” means that every unitary representation of $G$ extends to this algebra. The reduced group algebra uses just the regular representation and is a different object.

The manuscript chooses the countable group

$$\begin{array}{c}G=\mathbb{Z}[1/2{]}^3⋊({\mathrm{SL}}_3(\mathbb{Z})\times \mathbb{Z}),\\ (M,k)\cdot v=2^{k}Mv.\end{array}$$

Here $\mathbb{Z}[1/2]$ consists of rational numbers with a power of two as denominator, and ${\mathrm{SL}}_3(\mathbb{Z})$ consists of integer matrices with determinant one. The group records translations together with determinant-one linear changes of coordinates and dyadic dilations.

Nuclear norm-ultrapower obstruction, as stated in the manuscript

The separable unital algebra $A=C^{\ast }(G)$ admits no unital embedding into $B_{\omega }$ for any nonzero unital nuclear $C^{\ast }$-algebra $B$ and any free ultrafilter $\omega$ on $\mathbb{N}$. Neither separability of $B$ nor a choice of $G$ depending on $B$ is required. In particular, $A$ does not embed unitally into $({\mathcal{O}}_2{)}_{\omega }$. (OpenAI, 2026)

The proof combines a spectral finiteness statement inside each nuclear coordinate algebra with a finite chain argument. Merely exhibiting a proper isometry in $A$ would not prove the result: ${\mathcal{O}}_2$ already contains proper isometries.

🏷️ A spectral test for representation types

The useful subgroup is

$$\Gamma ={\mathbb{Z}}^3⋊{\mathrm{SL}}_3(\mathbb{Z}).$$

It has property (T), used here in its spectral-gap form. There is a finite generating set $S$ and a constant $\kappa >0$ such that every unitary representation $\sigma$ without invariant vectors satisfies

$$\sum _{s\in S}\Vert \sigma (s)\xi -\xi {\Vert }^2\ge \kappa \Vert \xi {\Vert }^2.$$

An invariant vector is fixed by every represented group element. The gap therefore separates exact invariance from representations having no invariant part.

For any tuple of unitaries $\mathbf{z}=(z_s{)}_{s\in S}$, define

$$L(\mathbf{z})=\sum _{s\in S}(z_s-1{)}^{\ast }(z_s-1).$$

This positive “Laplacian” satisfies $0\le L\le R1$, where $R=4|S|$. The tuple need not satisfy the relations of $\Gamma$. If it does come from a representation, the spectrum lies in $\{0\}\cup [\kappa ,R]$.

For a nonzero finite-dimensional irreducible unitary representation $\pi$, let $\bar{\pi }$ be its conjugate representation and set

$$\begin{array}{c}d(\mathbf{z},\pi )\\ =\min \mathrm{spec}L((z_s\otimes \bar{\pi }(s){)}_{s\in S}).\end{array}$$

In a faithful Hilbert-space realization, this is the infimum of the total squared displacement over unit vectors; the infimum need not be attained for an arbitrary tuple. For a genuine representation $\rho$, the value is zero exactly when $\pi$ occurs as a subrepresentation of $\rho$. Indeed, invariant vectors of $\rho \otimes \bar{\pi }$ correspond to maps intertwining $\pi$ with $\rho$; irreducibility makes every nonzero such map injective. Otherwise the spectral gap gives $d\ge \kappa$.

Fix the thresholds

$$b=\frac{{\kappa }^2}{4|S|},0<a<\min \{\kappa ,b\}.$$

The central analytic lemma says that for a fixed unitary tuple $\mathbf{u}$ in a nuclear algebra,

$$D(\mathbf{u})=\{[\pi ]:d(\mathbf{u},\pi )<a\}$$

is finite. Brackets mean unitary equivalence classes. There is no bound on the dimensions of the representations and no assumption that $\mathbf{u}$ satisfies group relations.

🏷️ Nuclearity forces finitely many approximate matches

First consider a representation $\sigma$ on a Hilbert space $K$ with no nonzero finite-dimensional subrepresentation. For arbitrary unitaries $U_s$ on another Hilbert space $E$, the manuscript proves

$$L((U_s\otimes \sigma (s){)}_{s\in S})\ge b1.$$

The arbitrary unitaries in the first tensor factor are removed by taking a reduced density operator. Given a unit vector $\xi \in E\otimes K$, let $T\ge 0$ be its reduced density operator on $K$, with $\mathrm{Tr}T=1$. Put

$${\delta }_s=\Vert (U_s\otimes \sigma (s))\xi -\xi \Vert .$$

Comparison of the two vector states gives

$$\Vert \sigma (s)T\sigma (s{)}^{\ast }-T{\Vert }_1\le 2{\delta }_s.$$

The Powers–Størmer square-root inequality then gives

$$\Vert \sigma (s)T^{1/2}\sigma (s{)}^{\ast }-T^{1/2}{\Vert }_2^2\le 2{\delta }_s.$$

The norms here are trace norm and Hilbert–Schmidt norm. Conjugation by $\sigma$ has no invariant Hilbert–Schmidt operator: the absolute value of a nonzero commuting compact operator would have a nonzero finite-dimensional invariant eigenspace. Applying property (T) to $T^{1/2}$, whose Hilbert–Schmidt norm is one, yields

$$\kappa \le 2\sum _s{\delta }_s\le 2\sqrt{|S|}{\left(\sum _s{\delta }_s^2\right)}^{1/2}.$$

Squaring proves the lower bound $b$. Its constants do not depend on $\dim E$ or $\dim K$.

Now suppose infinitely many inequivalent ${\pi }_j$ satisfy $d(\mathbf{u},{\pi }_j)<a$. Form the auxiliary matrix quotient

$$Q=\frac{{\prod }_j\mathcal{B}({\bar{V}}_{{\pi }_j})}{{⨁}_j\mathcal{B}({\bar{V}}_{{\pi }_j})}.$$

The denominator kills sequences converging to zero in the ordinary norm sense; this quotient has no ultrafilter. The conjugate representations define a representation $q$ into $Q$.

For any fixed finite-dimensional irreducible representation $\tau$, at most one ${\bar{\pi }}_j$ can be equivalent to $\tau$. Outside that exceptional coordinate, the Laplacian for ${\bar{\pi }}_j\otimes \bar{\tau }$ is at least $\kappa$. The exceptional coordinate vanishes in $Q$, so a faithful realization of $q$ has no finite-dimensional subrepresentation. The preceding estimate therefore gives a lower bound $b$ in $B{\otimes }_{\min }Q$.

On the other hand, constant copies of $B$ and coordinate copies of $Q$ give commuting homomorphisms into

$$\frac{{\prod }_j(B\otimes \mathcal{B}({\bar{V}}_{{\pi }_j}))}{{⨁}_j(B\otimes \mathcal{B}({\bar{V}}_{{\pi }_j}))}.$$

The maximal tensor product maps to this quotient by its universal property. Nuclearity identifies the maximal tensor product with the minimal one, allowing the spectral lower bound to pass to the quotient. Applying the continuous function $t\mapsto (b-t{)}_{+}$ shows that

$$(b-d(\mathbf{u},{\pi }_j){)}_{+}⟶0.$$

This contradicts $d(\mathbf{u},{\pi }_j)<a<b$. No identification of the minimal tensor product itself with the coordinate quotient is assumed. That distinction is exactly where an unjustified tensor-product interchange would otherwise enter the argument. (OpenAI, 2026)

🏷️ Uniformity across changing matrix dimensions

Suppose a unital embedding $C^{\ast }(G)\to B_{\omega }$ exists. Lift the group unitaries to unitaries $x_n(g)\in B$. Polar correction gives genuine unitary lifts because both unitary defects tend to zero. Their multiplication satisfies

$$\Vert x_n(g)x_n(h)-x_n(gh)\Vert {⟶}_{\omega }0$$

for every fixed pair $g,h$. Individual coordinates need not be representations. This is why the finiteness lemma was formulated for arbitrary tuples.

Let ${\mathbf{x}}_n=(x_n(s){)}_{s\in S}$ and define the finite sets $D_n=D({\mathbf{x}}_n)$. Their cardinalities and constituent dimensions may grow with $n$.

For every sequence of finite-dimensional representations ${\pi }_n$, the tuples $x_n(g)\otimes {\bar{\pi }}_n(g)$ give a genuine representation in the norm ultraproduct of the varying algebras $B\otimes \mathcal{B}({\bar{V}}_{{\pi }_n})$. Its Laplacian has the property-(T) gap. Since $0\le P_n\le R1$ for the coordinate Laplacians,

$$\begin{array}{c}\min \mathrm{spec}[(P_n)]\\ =R-\Vert R1-[(P_n)]\Vert \\ =\underset{\omega }{\lim }\min \mathrm{spec}{P}_n.\end{array}$$

Consequently,

$$\underset{\omega }{\lim }d({\mathbf{x}}_n,{\pi }_n)\in \{0\}\cup [\kappa ,R].$$

If $[{\pi }_n]\in D_n$ on an $\omega$-large set, this limit must be zero. This statement permits an adversarial representation choice changing at every coordinate. Checking one fixed representation at a time would not provide the uniformity needed below.

🏷️ Finite-index transport and the predecessor relation

The dyadic dilation $t\in G$ implements the injective map

$$\theta (v,M)=(2v,M)$$

on $\Gamma$. Its image $\Lambda$ has index eight, with coset representatives the translations in $\{0,1{\}}^3$.

Choose an irreducible representation ${\pi }_{\ast }$ factoring through reduction modulo two and nontrivial on translations. Such a representation occurs in the regular representation of the finite quotient. For every $j\ge 1$, the pullback ${\pi }_{\ast }\circ {\theta }^j$ is trivial on translations, so it is inequivalent to ${\pi }_{\ast }$.

Inducing ${\pi }_{\ast }$ from $\Gamma$ to $G$ gives a representation whose restriction contains ${\pi }_{\ast }$. Universality of the full group algebra makes the corresponding Laplacian have spectral minimum zero. Injectivity of the proposed embedding preserves this spectrum, so $[{\pi }_{\ast }]\in D_n$ on an $\omega$-large set.

Write $\pi \prec \rho$ when $\pi$ is a subrepresentation of $\rho$. The analytic transport argument proves the simultaneous assertion

$$\begin{array}{c}\text{for every }[\pi ]\in D_n,\\ \text{there is }[{\pi }^{\prime }]\in D_n\text{ with }\pi \prec {\pi }^{\prime }\circ \theta \end{array}$$

on one $\omega$-large set of coordinates.

To establish this quantifier order, suppose it fails and choose a bad ${\pi }_n$ at each offending coordinate. The varying-dimension gap makes its matching error tend to zero. Conjugation by $x_n(t)$ transports almost invariant vectors to $\Lambda$, for the representation ${\alpha }_n={\pi }_n\circ {\theta }^{-1}$.

Induce ${\alpha }_n$ back to $\Gamma$. Its Hilbert space has eight copies of the original representation space. If ${\eta }_n$ is the transported vector, the induced test vector has components

$$({\zeta }_n{)}_i=8^{-1/2}(x_n(r_i)\otimes 1){\eta }_n.$$

Its norm is one. Multiplying a coset representative by a generator uses only one of finitely many fixed identities $s{r}_i=r_{j}h$, with $h\in \Lambda$. The lift defects in these identities tend to zero, independently of $\dim {\pi }_n$. Thus the induced representation has matching error tending to zero.

Its Laplacian is a finite direct sum over irreducible constituents, so one constituent ${\pi }_n^{\prime }$ belongs to $D_n$. Every constituent has ${\pi }_n\prec {\pi }_n^{\prime }\circ \theta$: projection onto the identity coset is a nonzero $\Lambda$-intertwiner, hence surjective onto the irreducible ${\alpha }_n$, and its adjoint gives the required inclusion. The projection is nonzero because translating a vector with a nonzero $i$th coordinate by $r_i^{-1}$ moves that coordinate to the identity coset. This contradicts the bad choice. The proof never takes a countable intersection of $\omega$-large sets.

🏷️ A finite chain forces the contradiction

Fix a coordinate at which $D_n$ contains ${\pi }_{\ast }$ and every member has a predecessor. Starting at ${\tau }_0={\pi }_{\ast }$, choose a chain

$${\tau }_i\prec {\tau }_{i+1}\circ \theta .$$

Because $D_n$ is finite, some class repeats: ${\tau }_l\simeq {\tau }_{l+p}$ for $p\ge 1$. Composing inclusions around that segment gives

$${\tau }_l\prec {\tau }_l\circ {\theta }^p.$$

The dimensions are equal, so this inclusion is an equivalence.

It remains to transfer the periodicity back to the starting class. Every pullback ${\tau }_l\circ {\theta }^j$ is irreducible. Otherwise a proper invariant subspace would remain invariant under a deeper pullback ${\theta }^{kp}$ with $kp\ge j$, contradicting its equivalence to the irreducible ${\tau }_l$. The initial chain gives ${\pi }_{\ast }\prec {\tau }_l\circ {\theta }^l$, so irreducibility makes this an equivalence as well. Therefore

$${\pi }_{\ast }\simeq {\pi }_{\ast }\circ {\theta }^p,$$

contradicting the translation behavior modulo two.

The repetition alone would only locate a cycle somewhere in the chain. Equal dimensions and irreducibility are what propagate periodicity to ${\pi }_{\ast }$ and complete the argument.

🏷️ Key steps and proof outlines

Remove arbitrary first-factor unitaries with a density operator

Let $\sigma$ be a representation with no finite-dimensional invariant subspace, and let $(U_s{)}_{s\in S}$ be any unitary tuple, with no relations required. For a unit vector $\xi \in E\otimes K$, let $T$ be its reduced density operator on $K$: $T\ge 0$, $\mathrm{Tr}(T)=1$. If ${\xi }_s=(U_s\otimes \sigma (s))\xi$ and ${\delta }_s=\Vert {\xi }_s-\xi \Vert$, the reduced density changes to $T_s=\sigma (s)T\sigma (s{)}^{\ast }$, independently of $U_s$. Comparing vector states gives $\Vert T_s-T{\Vert }_1\le 2{\delta }_s$.

The square-root trace inequality then gives

$$\Vert \sigma (s)T^{1/2}\sigma (s{)}^{\ast }-T^{1/2}{\Vert }_2^2\le 2{\delta }_s.$$

Conjugation on Hilbert–Schmidt operators has no invariant vector: any nonzero invariant compact operator would have a nonzero finite-dimensional invariant eigenspace. Applying the representation spectral gap $\kappa$ to the unit vector $T^{1/2}$ and then Cauchy–Schwarz yields

$$\kappa \le 2\sum _s{\delta }_s\le 2\sqrt{|S|}{\left(\sum _s{\delta }_s^2\right)}^{1/2}.$$

Thus the tensor Laplacian is at least $b={\kappa }^2/(4|S|)$ for every first-factor tuple. This is the quantitative input that nuclearity transfers to its tensor-product comparison. (OpenAI, 2026)

Let matrix sizes vary inside the quotient, rather than fixing them afterward

Lift a hypothetical embedding into a norm ultrapower to unitary coordinate representatives $x_n(g)$. Their relations become exact only in the quotient. Suppose ${\pi }_n$ is a sequence of finite-dimensional representations whose dimensions may grow. Form the norm ultraproduct of the varying algebras $B\otimes \mathcal{B}(\overline{V_{{\pi }_n}})$.

The classes of $x_n(g)\otimes \overline{{\pi }_n}(g)$ give an exact representation, because multiplication errors are unchanged by tensoring with unitaries. Its Laplacian therefore has the genuine gap. For the positive coordinate Laplacians $P_n\le R1$, the norm formula gives

$$\min \mathrm{spec}[(P_n)]=R-\Vert R1-[(P_n)]\Vert =\underset{\omega }{\lim }\min \mathrm{spec}(P_n).$$

Hence those spectral minima have ultralimit zero or at least $\kappa$. This proves uniformity across changing representations without interchanging an ultralimit with an unbounded family of fixed-size estimates.

Obtain predecessors for every member at one coordinate

Nuclear finiteness says the set $D_n$ of irreducible approximate matches at coordinate $n$ is finite. The needed statement is that, on one ultrafilter-large set, every member has a predecessor in that same $D_n$ under the relation $\pi \prec {\pi }^{\prime }\circ \theta$.

Argue by contradiction: if that fails on a large set, choose a violating ${\pi }_n$ at every such coordinate. The varying-size gap makes its matching error tend to zero. Transport almost invariant vectors using the fixed implementing group element, then induce from the fixed finite-index subgroup. The induction has eight coordinates; for a fixed generating set it uses only $8|S|$ fixed multiplication tests. Its errors are therefore independent of $\dim {\pi }_n$.

The induced Laplacian is a direct sum over irreducible constituents. Its small spectral minimum selects a constituent ${\pi }_n^{\prime }$ in $D_n$. Projection to the identity coset is a nonzero intertwiner onto the original irreducible subgroup representation; its adjoint embeds that representation into every chosen constituent’s restriction. Thus ${\pi }_n\prec {\pi }_n^{\prime }\circ \theta$, contradicting the selected failure. This diagonal choice is what proves the simultaneous quantifier statement.

Finiteness forces periodicity all the way back to the initial representation

Follow predecessors in one finite set: ${\tau }_i\prec {\tau }_{i+1}\circ \theta$. Eventually ${\tau }_l\simeq {\tau }_{l+p}$. Composing inclusions around that segment gives ${\tau }_l\prec {\tau }_l\circ {\theta }^p$; equal dimensions make it an equivalence.

Every pullback ${\tau }_l\circ {\theta }^j$ is irreducible. Otherwise a proper invariant subspace would remain invariant after further pullback to a multiple $kp\ge j$, contradicting ${\tau }_l\circ {\theta }^{kp}\simeq {\tau }_l$. The initial segment then gives ${\tau }_0\simeq {\tau }_l\circ {\theta }^l$, so periodicity also holds for ${\tau }_0$.

The distinguished initial representation is not periodic: it detects translations modulo two, whereas every positive iterate of $\theta$ makes those translations invisible. This finite contradiction completes the analytic argument. The nuclear-finiteness and ultrapower-contradiction sections establish the interfaces; the group construction supplies this final nonperiodicity test.

🏷️ Consequences and verification boundaries

The manuscript also deduces that embeddability into $({\mathcal{O}}_2{)}_{\omega }$ can fail to survive a crossed product by a single automorphism. Such a crossed product adjoins a unitary implementing that automorphism. This consequence uses an imported equivalence: the algebra and automorphism witnessing it are existential, not explicitly identified by the dyadic group construction. The source’s further logical consequences are outside the requested subject coverage.

Jiang–Su stability uses nuclear approximation to construct central structure; here equality of tensor norms instead forces a finite set of approximate representation types. The similarity article provides another use of estimates uniform in auxiliary matrix dimensions.

All principal source sections were read for this exposition. The review follows the proof’s tensor-product passage, spectral estimates, varying-dimension quantifiers and finite-cycle argument. Property (T) of the affine subgroup and the structural embedding and consequence theorems remain imported results; this is not an independent certification of the complete research claim.

🏷️ Manuscripts and verification scope

• An explicit obstruction to nuclear norm-ultrapower embeddings — pinned manuscript and companion source.

The family scope page describes selected formalization coverage. The documented selection is the explicit separable unital full group algebra and its obstruction to unital embeddings into the norm ultrapower of every nonzero unital nuclear algebra, for every free ultrafilter on the natural numbers. This includes the Cuntz algebra O₂. The page does not list the additional existential or crossed-product consequences. 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. An explicit obstruction to nuclear norm-ultrapower embeddings.