Trace cones and Razak–Jacelon stabilization

Source and verification

This article explains family 301 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 301 asks how much of a nuclear operator algebra survives tensoring with the Razak–Jacelon algebra. The manuscript claims that the full topological cone of extended traces becomes a complete invariant after this tensoring and stabilization by compact operators. Infinite trace values retain ideal information. The central construction must therefore preserve both finite numerical data and the regions where every relevant value is infinite. (OpenAI, 2026)

🏷️ Preliminaries: stabilization and tracial weights

A $C^{\ast }$-algebra is a norm-closed, adjoint-closed algebra of bounded operators. It may have no identity. It is separable if it has a countable norm-dense subset. It is nuclear if its identity map can be approximated on finite sets by completely positive contractive maps through matrix algebras. Complete positivity means that all matrix amplifications preserve positivity.

Write $\mathcal{K}=\mathcal{K}({\ell }^2(\mathbb{N}))$ for the compact operators. Tensoring with $\mathcal{K}$ is stabilization: it provides arbitrarily large matrix corners. Matrix traces in this direction are unnormalized, so orthogonal copies add their ranks.

The Razak–Jacelon algebra $\mathcal{W}$ is a standard simple nuclear algebra with a unique normalized trace. It is stably projectionless: even its matrix stabilization contains no nonzero projections. Its absorption properties include $\mathcal{W}\otimes \mathcal{W}\cong \mathcal{W}$ and absorption of the universal UHF algebra, an infinite tensor product of matrix algebras. These are imported properties of the tensor factor; the algebras being classified need not be simple.

A closed two-sided ideal $I$ is a closed subspace invariant under multiplication on both sides. It records a region of the algebra on which one can localize operators. Ideals will enter the invariant through extended traces.

An extended lower-semicontinuous tracial weight is an additive, positively homogeneous map $\tau :C_{+}\to [0,\infty ]$ with $\tau (0)=0$ and

$$\tau (x^{\ast }x)=\tau (x{x}^{\ast }).$$

Lower semicontinuity means that $\tau (a)\le {\mathrm{liminf}}_i\tau (a_i)$ whenever positive $a_i$ converge in norm to $a$. Infinite values are allowed, and the finite domain need not be dense. Let $T(C)$ contain all these weights, including zero.

For a positive operator, $(a-\epsilon {)}_{+}$ denotes the continuous functional calculus of $t\mapsto \max (t-\epsilon ,0)$. The canonical topology on $T(C)$ is specified by

$$\begin{array}{c}\underset{i}{\mathrm{limsup}}{\tau }_i((a-\epsilon {)}_{+})\le \tau (a),\\ \tau (a)\le \underset{i}{\mathrm{liminf}}{\tau }_i(a),\end{array}$$

for every $a\ge 0$ and $\epsilon >0$. A topological-cone isomorphism is a homeomorphism preserving addition, zero, and multiplication by positive real scalars. This is stronger data than a bijection of normalized traces.

🏷️ The classification statement

Manuscript claim

Let $A,B$ be separable nuclear complex $C^{\ast }$-algebras. An isomorphism of the topological cones $T(A)$ and $T(B)$ yields an actual isomorphism

$$A\otimes \mathcal{W}\otimes \mathcal{K}\cong B\otimes \mathcal{W}\otimes \mathcal{K}.$$

All tensor products are spatial. No simplicity, unitality, restriction on the ideal space, or universal coefficient theorem hypothesis is imposed. (OpenAI, 2026)

Put $P=A\otimes \mathcal{W}\otimes \mathcal{K}$ and $Q=B\otimes \mathcal{W}\otimes \mathcal{K}$. After transporting the given invariant to a cone map $F:T(Q)\to T(P)$, the manuscript constructs an isomorphism $\Phi :P\to Q$ realizing it:

$$\tau \circ \Phi =F(\tau )(\tau \in T(Q)).$$

The direction reverses because a map of algebras pulls traces back. Both tensor factors belong to the conclusion; it is not a claim that $A$ and $B$ themselves are isomorphic.

🏷️ Infinite values encode ideals

Every ideal $I\subseteq C$ contributes the weight

$${\tau }_I(a)=\{\begin{array}{ll}0,& a\in I_{+},\\ \infty ,& a\notin I_{+}.\end{array}$$

These are exactly the additive idempotents of the cone: $\tau +\tau =\tau$ forces each value to be zero or infinity. Their addition recovers ideal inclusion:

$$I\subseteq J⟺{\tau }_I+{\tau }_J={\tau }_I.$$

For a general weight, let $\mathrm{Fin}(\tau )$ be the closed ideal generated by positive elements of finite trace, and let $\mathrm{Ker}(\tau )$ be its zero ideal. A member of $\mathrm{Fin}(\tau )$ can still have infinite trace; the restriction is densely finite, not necessarily finite everywhere.

The cone topology recovers both ideals through scalar limits:

$$\begin{array}{rl}\underset{r\downarrow 0}{\lim }r\tau & ={\tau }_{\mathrm{Fin}(\tau )},\\ \underset{r\uparrow \infty }{\lim }r\tau & ={\tau }_{\mathrm{Ker}(\tau )}.\end{array}$$

The cut-based topology is essential in the first limit. Pointwise evaluation at an element of infinite trace would keep giving infinity, even if that element lies in the finite ideal.

A trace finite on only one summand

On $C=\mathbb{C}\oplus \mathbb{C}$, define $\tau (a,b)=a$ when $b=0$, and $\tau (a,b)=\infty$ when $b>0$, for $a,b\ge 0$. Its finite ideal is $\mathbb{C}\oplus 0$ and its zero ideal is zero. Scaling toward zero retains the first ideal; scaling toward infinity retains the second. Omitting nondensely finite weights would omit this example.

🏷️ Rank functions and Cuntz comparison

For positive elements in $C\otimes \mathcal{K}$, write $a\precsim b$ when there are operators $v_n$ with $v_n^{\ast }b{v}_n\to a$ in norm. Mutual comparison defines a Cuntz class; these classes form the ordered semigroup $\mathrm{Cu}(C)$ under orthogonal sum.

For a positive contraction, define its trace rank by

$$d_{\tau }(a)=\underset{k\to \infty }{\lim }\tau (a^{1/k}).$$

In matrices this is the trace of the support projection. In a stably projectionless algebra, it remains meaningful without an actual support projection in the algebra.

The rank is determined by its positive cuts, and traces are recovered by spectral integration:

$$\begin{array}{c}{d}_{\tau }(a)=\underset{\epsilon >0}{\sup }{d}_{\tau }((a-\epsilon {)}_{+}),\\ \tau (a)={\int }_0^{\Vert a\Vert }{d}_{\tau }((a-t{)}_{+})dt.\end{array}$$

The manuscript imports trace-realification results: after tensoring with $\mathcal{W}$, the relevant Cuntz semigroup is determined by the trace cone, and its order is pointwise order of rank functions. Thus $F$ induces $G:\mathrm{Cu}(P)\to \mathrm{Cu}(Q)$ with

$$d_{\tau }(G[a])=d_{F(\tau )}(a).$$

The function representation has specific sequential approximation and continuity conditions; it does not identify arbitrary lower-semicontinuous homogeneous functions with Cuntz classes. Realification supplies the invariant-level correspondence. Producing algebra homomorphisms realizing it is the main remaining problem.

🏷️ Local errors that keep their norm

The construction first works in the sequence algebra

$$Q_{\infty }={\ell }^{\infty }(Q)/c_0(Q).$$

Here ${\ell }^{\infty }(Q)$ consists of bounded sequences, and $c_0(Q)$ of norm-null sequences.

A global rank bound can be useless: a full positive element in a stable algebra may have infinite rank at every relevant nonzero trace. A small multiple of infinity is still infinity. The manuscript instead fixes positive controllers inside individual ideals.

The cone correspondence identifies the primitive ideal spaces of $P$ and $Q$; call their common space $X$. Points are kernels of irreducible representations, and open subsets label closed ideals. Write $P(U),Q(U)$ for the ideals corresponding to an open $U\subseteq X$, and $U_a$ for the open support of the ideal generated by $a\in P_{+}$.

For an open ideal label $U$, an error $x=[(x_n)]$ belongs to $J(U)$ if, for every $\epsilon >0$, there are a fixed positive $c$ in the corresponding ideal, positive numbers $r_n\to 0$, and a tail on which

$$d_{\tau }((|x_n|-\epsilon {)}_{+})\le r_n{d}_{\tau }(c)$$

for every trace $\tau$. The controller is Cuntz-below a fixed positive cut of an element of that ideal. It is fixed before the coordinate tends to infinity.

This need not imply $\Vert x_n\Vert \to 0$. In an extra UHF factor take nonzero projections $e_n$ of normalized trace $t_n\to 0$. For fixed positive $a$, the operators $a\otimes e_n$ retain norm $\Vert a\Vert$ while their trace ranks are multiplied by $t_n$.

The manuscript proves that $J(U)$ is a closed ideal, invisible to the appropriate regularized traces densely finite on that region. Small stabilizing homomorphisms can therefore absorb errors while retaining enough norm and support to be useful. “Small” here refers to controlled ranks, not operator norm.

🏷️ Regularized traces and preservation of the infinite region

For varying coordinate weights ${\tau }_n$, a free ultrafilter $\omega$, and bounded positive representatives $x_n\ge 0$, the relevant trace on $Q_{\infty }$ is

$$\rho ([(x_n)])=\underset{\epsilon >0}{\sup }\underset{n\to \omega }{\lim }{\tau }_n((x_n-\epsilon {)}_{+}).$$

The cuts make it independent of the chosen representatives. The model identity must hold for every such regularized limit, including weights with a proper or zero finite ideal.

A cone model is a homomorphism from $C_0((0,1])\otimes P$ to $Q_{\infty }$ whose pulled-back weights are products of a finite full-support measure $m$ with $F(\rho {|}_Q)$. The product weight is defined on its densely finite ideal by integration and extended by infinity outside it.

The existence construction has two distinct requirements. Completely positive approximations match first moments and square moments on the finite region. In addition, the chosen controlled positive elementary tests $f\otimes a$ with $f\ne 0$ retain a lower ideal inclusion:

$$Q(U_a)\subseteq {\mathrm{Ideal}}_{Q_{\infty }}(p(f\otimes a)).$$

The second requirement prevents an infinite region from accidentally acquiring finite trace after correction.

Here is the mechanism. Suppose the pulled-back weight were densely finite on an extra open rectangle. Choose a compactly cut-supported tensor $s=f\otimes a$ there. Its image has finite rank. The retained ideal inclusion implies that each positive cut of any constant $c\in Q(U_a)$ is Cuntz-below finitely many copies of a fixed cut of $p(s)$. Consequently every such cut of $c$ has finite rank. This forces $Q(U_a)$ into the original finite ideal, contradicting the purported extra rectangle.

Once the finite ideals coincide, the moments determine the trace inside them, and both sides are infinity outside. Preserving finite moments alone would not establish the full model identity. (OpenAI, 2026)

🏷️ Correction and uniqueness interfaces

The manuscript’s main technical stages construct and compare these models. Rough completely positive maps have multiplicative defects in the local ideals $J(U)$. Small stabilizers put those defects into suitable hereditary targets; Stinespring dilation and ideal-related absorption correct the maps. The construction keeps the fine labels in $(0,1]\times X$, not only their projections to the ideal space $X$. A second stabilizer retains the lower ideal inclusion after correction.

The uniqueness theorem says that two models with identical data have coordinate lifts that become pointwise norm-close after conjugacy by multiplier unitaries. Its proof first aligns tracial data, then absorbs the remaining local errors. Splitting each existing model into two UHF halves supplies a spare model while the other half is replaced. This uniqueness argument is conditional on the models already being given; it does not assume the later existence theorem.

These are substantial manuscript inputs. The complete stabilizer, absorption, measurable-assembly, and correction proofs are outside this post’s selected source audit.

🏷️ An exact positive identity for a corner

A useful explicit part of the last stage avoids subtracting infinite values. Cone models are assembled into a line model. Translation invariance of Lebesgue measure and uniqueness then give a map from the crossed product

$$E\otimes P,E=C_0(\mathbb{R})⋊\mathbb{Z},$$

where the integer action translates by one.

Choose $\zeta \in C_c(\mathbb{R}{)}_{+}$ with

$$\sum _{n\in \mathbb{Z}}\zeta (t-n{)}^2=1,{\int }_{\mathbb{R}}{\zeta }^2=1.$$

The manuscript constructs a projection $p\in E$ from the finite Fourier coefficients $p_n(t)=\zeta (t)\zeta (t-n)$. The displayed partition identity proves $p^2=p$ by convolution.

Choose finitely many compactly supported continuous real functions $q_i$, each supported in an interval of diameter less than one, with ${\sum }_i{q}_i^2=1$ on $\mathrm{supp}\zeta$. For positive $a$, set $y_i=q_{i}p\otimes a^{1/2}$. Then

$$\begin{array}{rl}\sum _i{y}_i^{\ast }{y}_i& =p\otimes a,\\ \sum _i{y}_i{y}_i^{\ast }& ={\zeta }^2\otimes a.\end{array}$$

Every extended trace has the same value on the two sums. A product trace with Lebesgue measure therefore gives exactly its point value on $p\otimes a$, including infinity. Restricting the crossed-product model to this corner yields a point model $P\to Q_{\infty }$.

🏷️ From sequence models to an actual isomorphism

Nuclearity gives completely positive contractive lifts ${\phi }_n:P\to Q$ of the point model. They need not converge.

Uniqueness for every pair of subsequences implies that any two sufficiently late lifts are close on a prescribed finite set after unitary conjugacy. Otherwise, two sequences of bad pairs would define models contradicting uniqueness.

Choose a subsequence and cumulative multiplier conjugacies so that the resulting maps ${\psi }_k$ satisfy

$$\Vert {\psi }_{k+1}(a)-{\psi }_k(a)\Vert <2^{-k}$$

on increasing finite sets with dense union. Summability gives a point-norm limit $\varphi :P\to Q$. Vanishing multiplicative defects make it a homomorphism.

One cannot pass an unbounded weight through this norm limit by ordinary continuity. Instead, the corrected sequence and constant limit define the same element of $Q_{\infty }$. Regularized traces and unitary invariance recover the prescribed trace identity; positive powers and cuts recover $G$.

Repeat in the reverse direction. Uniqueness makes both compositions approximately the identity. An approximate intertwining with summable errors, including the images of earlier finite tests, produces mutually inverse limiting homomorphisms. Final cut comparisons and spectral integration show that the resulting isomorphism still realizes the original cone map.

🏷️ Key steps and proof outlines

Normalize the corner using only positive identities

The point-model stage needs a corner whose trace mass is exactly one for every extended weight, including weights taking value infinity. The source’s “An explicit trace-normalized corner” constructs it directly. Start with $\eta \in C_c(\mathbb{R}{)}_{+}$ strictly positive on $[0,1]$, and set

$$H(t)=\sum _{n\in \mathbb{Z}}\eta (t-n{)}^2,\zeta (t)=H(t{)}^{-1/2}\eta (t).$$

The sum is locally finite, positive and periodic. Thus ${\sum }_n\zeta (t-n{)}^2=1$, and integration over one period gives $\int {\zeta }^2=1$. In the translation crossed product, the coefficients $p_n(t)=\zeta (t)\zeta (t-n)$ define a finite Fourier sum $p$. Its convolution square has coefficient

$$(p^2{)}_k(t)=\zeta (t)\zeta (t-k)\sum _n\zeta (t-n{)}^2=p_k(t).$$

The involution has the same coefficients, so $p$ is a projection. (OpenAI, 2026)

Choose the functions $q_i$ described above with supports of diameter less than one. In the $n$th Fourier coefficient of ${\sum }_i{q}_{i}p{q}_i$, every term contains $q_i(t)q_i(t-n)$. For $n\ne 0$ this is zero by the support diameter. At $n=0$, the coefficient is ${\zeta }^2{\sum }_i{q}_i^2={\zeta }^2$. Hence the positive partition identity is exact. Applying the trace to the two finite sums ${\sum }_i{y}_i^{\ast }{y}_i$ and ${\sum }_i{y}_i{y}_i^{\ast }$ proves the normalization without subtracting two infinite values.

This also proves fullness of the corner. The ideal generated by $p$ contains ${\zeta }^2$, hence every integer translate of ${\zeta }^2$. Any compactly supported scalar function is a finite sum of itself times these translates, by the partition identity. Thus the ideal contains $C_0(\mathbb{R})$ and all finite Fourier sums, which are dense in the crossed product. The corner retains the whole algebraic support while fixing the trace normalization.

Turn uniqueness into a coherent sequence of actual maps

The coordinate lifts ${\phi }_n$ are only asymptotically multiplicative; an arbitrary subsequence may still oscillate in norm. The required consequence of uniqueness is stronger: for every finite set $\mathcal{E}$ and tolerance, every pair of sufficiently late coordinates is approximately unitarily conjugate on $\mathcal{E}$.

If this failed, select two increasing sequences of bad coordinates. Each subsequence still defines a model: any coordinate weights and ultrafilter on the subsequence extend to the full sequence by placing zero weights at unused indices and concentrating the ultrafilter on the chosen indices. Positive-cut ultralimits are unchanged. The uniqueness theorem therefore applies to these two models and contradicts the bad-pair choice. This quantifier argument is what permits a consistent recursive selection.

Choose successive unitaries $w_k$ with error $2^{-k}$ on increasing finite sets, and accumulate them as $W_{k+1}=W_k{w}_k$. Then ${\psi }_k=\mathrm{Ad}(W_k){\phi }_{n_k}$ has summable successive errors on a dense set. Contractivity extends convergence to every element. Its limit is multiplicative because the original lifts’ multiplicative defects vanish.

Pass the invariant through the quotient, not through unbounded evaluation

Norm convergence alone does not imply convergence of an unbounded trace. Instead, ${\psi }_k$ and its constant limit represent the same element of $Q_{\infty }$. Unitary invariance holds at every coordinate and every positive cut, so the regularized trace of this common element is exactly the prescribed value $F(\tau )(a)$. Applying this identity to positive powers gives equality of rank functions; finite matrix amplification and approximation by positive cuts extend it to the full Cuntz invariant.

Because the induced invariant map is injective, a positive element killed by the limiting homomorphism has zero Cuntz class and is itself zero. Thus the map is injective before the reverse map and approximate intertwining are constructed. These final steps depend on the earlier model-existence and uniqueness theorems; the explicit limiting argument does not replace their separate correction and absorption proofs.

🏷️ Links and reading scope

• Family 294 studies the distinction between quasitraces and traces; the present nuclear setting uses the extended trace interface in its rank invariant.
• Family 299 concerns tensor absorption through central sequences; here tensoring with $\mathcal{W}$ changes the classification invariant to the trace cone.
• The collection guide gives the operator-analysis reading sequence.

The source reading covers the full introduction, structural section, and final point-model/intertwining section, together with selected regularization, local-error, stabilizer, uniqueness, rough-cone, and corrected-cone statements and arguments. It does not verify the full 4,800-line construction or its external classification and absorption inputs.

🏷️ Manuscripts and verification scope

• The trace cone classifies Razak–Jacelon stabilizations — pinned manuscript and companion source.

No family scope page was present at lean/docs/301.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 2026. The trace cone classifies Razak–Jacelon stabilizations.