Pure infiniteness: compatible corners and amplification removal

Source and verification

This article explains family 303 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

Proper infiniteness says that a positive operator can contain two copies of itself in Cuntz comparison. Strong pure infiniteness asks for compatible choices for two positive operators while also making a mixed entry small. Family 303 claims that individual proper infiniteness already gives this simultaneous diagonalization for every complex $C^{\ast }$-algebra. A second theorem removes a fixed amplification under exactness; separable nuclear algebras then absorb the Cuntz algebra ${\mathcal{O}}_{\infty }$. (OpenAI, 2026)

🏷️ Preliminaries: comparison, cutoffs, and ideals

A $C^{\ast }$-algebra $C$ is a norm-closed algebra of operators closed under adjoints; it may be nonunital. Its positive elements are the operators of the form $x^{\ast }x$. The compact operators $\mathcal{K}$ provide a stabilization $C\otimes \mathcal{K}$ containing all finite matrix corners.

For positive elements, write $a\precsim b$ when there are operators $v_j$ in the stabilization with $v_j^{\ast }b{v}_j\to a$ in norm. This is Cuntz subequivalence. Write $a{\sim }_{\mathrm{Cu}}b$ for comparison in both directions. Direct sums use orthogonal matrix corners.

A positive $h$ is properly infinite when

$$h\oplus h\precsim h.$$

Zero is included. This cannot happen for a nonzero finite matrix: the rank of two copies exceeds the original rank. The condition concerns approximation and support rather than an algebraic equality of operators.

The hereditary subalgebra generated by $h$ is $\mathrm{Her}(h)=\overline{hCh}$. It contains operators supported within the spectral support of $h$. The larger closed two-sided ideal generated by $h$ is denoted $\mathrm{Id}(h)$.

The positive cutoff $(h-\delta {)}_{+}$ applies $t\mapsto \max (t-\delta ,0)$ by functional calculus. Its support projection belongs to the universal bidual $C^{\ast \ast }$, a von Neumann algebra containing $C$, and need not belong to $C$ itself. Write $e_h$ for the support of $h$ there.

An irreducible representation has no proper nonzero invariant closed subspace. Its kernel is a primitive ideal $P$. The quotient norm $\Vert h+P\Vert$ detects how much of $h$ remains in that representation. Ideal membership can be tested this way: $a\in \mathrm{Id}(h)$ exactly when every irreducible representation killing $h$ also kills $a$.

🏷️ Three infiniteness conditions

The manuscript distinguishes two individual comparison hypotheses:

$$\begin{array}{rl}\text{(P)}:& h\oplus h\precsim h(h\in C_{+}),\\ {\text{(F}}_n\text{)}:& h^{\oplus n}\text{ is properly infinite}\\ & \text{for every }h\in C_{+}.\end{array}$$

In the second, one integer $n\ge 1$ is fixed for the whole algebra, before choosing $h$ or any spectral cutoff. It is the manuscript’s fixed-amplification version of weak pure infiniteness.

Strong pure infiniteness requires, for every $x,y\ge 0$ and every error tolerance, coefficients $s,t\in C$ approximately preserving $x^2,y^2$ under conjugation while making $s^{\ast }xyt$ small. There is no prescribed coefficient norm in this definition.

Individual-to-strong theorem, as stated in the manuscript

If $C$ satisfies (P), then for every $a,b\in C_{+}$, arbitrary $c\in C$, and $\epsilon >0$, there are $s,t\in C$ such that

$$\begin{array}{c}\Vert s^{\ast }as-a\Vert <\epsilon ,\\ \Vert t^{\ast }bt-b\Vert <\epsilon ,\\ \Vert s^{\ast }ct\Vert <\epsilon .\end{array}$$

No exactness, nuclearity, separability, unitality, or simplicity is assumed. Taking $(a,b,c)=(x^2,y^2,xy)$ gives strong pure infiniteness. (OpenAI, 2026)

The arbitrary mixed entry $c$ is a stronger-looking but equivalent diagonalization formulation. Finding separate coefficients for $a$ and $b$ does not control their interaction through $c$.

A $C^{\ast }$-algebra is exact when its spatial tensor product preserves short exact sequences. Nuclear algebras, whose identity maps admit completely positive finite-matrix approximations, are exact.

Amplification removal, as stated in the manuscript

An exact complex $C^{\ast }$-algebra satisfying (F${}_n$) for one fixed $n$ satisfies (P), and hence is strongly purely infinite. Separability and a unit are unnecessary. (OpenAI, 2026)

The weak-to-ordinary implication for arbitrary nonexact algebras is outside these conclusions.

🏷️ Implementing comparison inside hereditary corners

A key tool is a strengthened cutoff implementation. If $d\precsim v$ and $h=(d-\delta {)}_{+}$, the manuscript constructs a rectangular partial isometry $w$ in the bidual with

$$w^{\ast }w=e_h,w{w}^{\ast }\le e_v.$$

Its products with $\mathrm{Her}(h)$ belong to the original algebra, and conjugation transports that hereditary algebra into $\mathrm{Her}(v)$. It preserves represented norms at every matrix level. The claim is about these products, not membership of $w$ itself in $C$.

One route starts from a rectangular comparison coefficient $s$ with $\Vert d-s^{\ast }vs\Vert <\delta /2$ and forms $z=v^{1/2}sh$. The estimate

$$\frac{\delta }{2}{h}^2\le z^{\ast }z\le (\Vert d\Vert +\delta /2)h^2$$

identifies the hereditary algebra and support of $z^{\ast }z$ with those of $h$. Polar decomposition then supplies the common transport.

A second basic fact is that if $v$ is properly infinite, every positive element in its ideal is Cuntz-below $v$. Finite sums first compare a cutoff into finitely many copies of $v$; proper infiniteness absorbs those copies, and a cutoff limit completes the comparison.

🏷️ Shrinking the mixed support

For positive contractions $a,b$ and a contraction $c$, the first target is to find positive contractions $p\in \mathrm{Her}(a)$ and $q\in \mathrm{Her}(b)$ with

$$\Vert e_{p}c{e}_q\Vert <1,$$

while retaining the represented norm signals

$$\begin{array}{rl}\Vert \pi (a)\Vert >1/2& ⟹\Vert \pi (p)\Vert =1,\\ \Vert \pi (b)\Vert >1/2& ⟹\Vert \pi (q)\Vert =1.\end{array}$$

These signals prevent the construction from becoming trivial by choosing zero, and later ensure the right ideal inclusions.

The manuscript constructs a graph inside orthogonal copies of a spectral cutoff of $a$. Its finitely many free entries are chosen after certain measurements have been fixed. If the positive defect $b(1-c^{\ast }{p}^{2}c)b$ vanishes in a representation $\pi$ with $\Vert \pi (b)\Vert >1/4$, equality in contraction estimates forces every free entry to equal its fixed measurement.

Under (P), a four-entry construction excludes those equalities. Its first pair forces an isometry with nonzero defect, and its second pair contradicts the same defect. Spectral cutoffs put vectors in actual operator ranges, and directed finite tests avoid a separability hypothesis. The full four-entry exclusion is a substantive manuscript input to this exposition.

Once nonvanishing of the defect is known, ideal cutoffs convert it to a strict global support estimate. On a spectral subspace where the defect exceeds $\delta$,

$$\Vert pcb\xi {\Vert }^2\le (1-\delta )\Vert b\xi {\Vert }^2.$$

A further high cutoff of $p$ and a cutoff in the hereditary algebra of $b$ give the strict contraction while retaining both signals.

🏷️ Uniform contraction and bounded final coefficients

A merely input-dependent contraction factor is not enough for repeated use. The manuscript proves that bounded products preserve (P), with a common coefficient bound. For a positive contraction $d$ and $0<\delta <1/2$, there is a row $s$ with

$$\begin{array}{c}{s}^{\ast }ds=(d-2\delta {)}_{+}^{\oplus 2},\\ \Vert s\Vert \le {\delta }^{-1/2}.\end{array}$$

Consequently, if strict shrinking had no universal factor $\theta <1$, a product of increasingly bad examples would contradict strict shrinking in the product algebra. Coordinatewise comparison alone would not justify this argument without uniformly bounded rows.

Iteration gives $\Vert e_{p}c{e}_q\Vert <\zeta$ for any prescribed $\zeta >0$, retaining the original norm signals. Nested cutoff functions keep the normalized mixed entry in $C$ and preserve norm one in all relevant representations.

The final quantitative statement for positive contractions $a,b$ is

$$\begin{array}{c}\Vert s^{\ast }as-a\Vert ,\Vert t^{\ast }bt-b\Vert <2\delta +\rho ,\\ \Vert s^{\ast }ct\Vert <\zeta ,\Vert s\Vert ,\Vert t\Vert \le {\delta }^{-1/2},\end{array}$$

where $0<\delta <1/2$ and $0<\rho \le 2\delta$. The coefficient bound is independent of the requested mixed-entry accuracy.

To see the order of choices, first shrink the support compression below $\delta \zeta$. The norm signals place $(a-2\delta {)}_{+}$ in $\mathrm{Id}(p)$. Since $p$ is supported where $a>\delta$, one has $pap\ge \delta p^2$, and its ideal equals that of $p$. Compare the diagonal cutoff into $pap$, writing $s=pw$. Then

$$\begin{array}{c}\delta s^{\ast }s\le s^{\ast }as,\\ \Vert s{\Vert }^2\le \frac{1-2\delta +\rho }{\delta }\le {\delta }^{-1}.\end{array}$$

The same holds for $t$, so

$$\Vert s^{\ast }ct\Vert \le {\delta }^{-1}\Vert e_{p}c{e}_q\Vert <\zeta .$$

This bound ensures the later implementation cannot undo the earlier separation. Scaling and choosing $2\delta +\rho$ sufficiently small prove the theorem.

🏷️ Exactness and a uniform obstruction

The exact route to removing $n$ begins with the universal unital algebra generated by one contraction. The manuscript uses its nonexactness to obtain a universal $\gamma >0$: no contraction $v$ in a unital exact algebra can approximately realize every finite matrix contraction $D$ through an isometry $J$ with both

$$\begin{array}{c}\Vert \pi (v)J-JD\Vert <\gamma ,\\ \Vert \pi (v{)}^{\ast }J-J{D}^{\ast }\Vert <\gamma .\end{array}$$

The representation and $J$ may depend on $D$. Both inequalities must hold for the same copy; they control words involving the generator and its adjoint.

Combined with (F${}_n$), this obstruction gives perturbations that remain nonzero simultaneously on a prescribed primitive norm slice. The comparison data are fixed before each finite matrix test, so a hypothetical failure would realize all forbidden matrices using one element of the same exact target.

The slice

$$\begin{array}{c}K(a,r)=\{P\in \mathrm{Prim}(C):\\ \Vert a+P\Vert \ge r\}\end{array}$$

is quasi-compact for $r>0$, even when the primitive ideal space is not Hausdorff. A finite-cover proof uses norm attainment in a quotient and approximation by a finite sum of ideals. It does not require this slice to be closed.

A finite coupling argument preserves previously obtained nonzero corners. If the $i$th perturbation has size ${\eta }_i={\theta }^{i-1}$ with $\theta =\sigma /4$, later changes have total norm below $\sigma {\eta }_i/2$, while the $i$th corner was established with norm at least $\sigma {\eta }_i$. The signal survives.

The manuscript uses three such coupling stages to obtain a square-zero element $x$ in a prescribed hereditary subalgebra with $x+P\ne 0$ throughout the slice. The complete perturbation and localization proofs are further inputs; the final amplification removal is transparent once this square-zero construction is available.

🏷️ Repeated halving removes the fixed amplification

For a square-zero $x$, the elements $x^{\ast }x$ and $x{x}^{\ast }$ are orthogonal and Cuntz equivalent. Starting in $\mathrm{Her}(a)$ and repeating on the same slice $K(a,r)$ produces positives $z_j$ nonzero throughout that slice with

$$z_{j+1}^{\oplus 2}\precsim z_j,z_j^{\oplus 2^j}\precsim a.$$

Choose $j$ with $2^j\ge n$. Then $z_j^{\oplus n}\precsim a$, and (F${}_n$) makes this amplification properly infinite.

Every primitive representation killing $z_j$ has $\Vert a+P\Vert <r$, so it kills $(a-r{)}_{+}$. Thus $(a-r{)}_{+}\in \mathrm{Id}(z_j)$, and proper-target ideal comparison gives

$$((a-r{)}_{+}{)}^{\oplus 2}\precsim z_j^{\oplus n}\precsim a.$$

Let $r\downarrow 0$ to obtain $a\oplus a\precsim a$. Only scalar positives $z_j\in C$ were tested against the fixed-amplification hypothesis. This avoids assuming the desired conclusion for arbitrary matrix positives.

The manuscript also gives a separate nuclear proof using a finite tuple of matrix norms, representation-dependent hereditary witnesses, and a merger preserving their norm signals. It is an alternative route, not an additional hypothesis for the exact theorem.

🏷️ Tensor absorption and scope

For separable nuclear $A$ satisfying $(F_n)$ for some fixed $n$ (or already strongly purely infinite), the manuscript combines strong pure infiniteness with an imported central-sequence absorption theorem to conclude

$$A\cong A{\otimes }_{\min }{\mathcal{O}}_{\infty }.$$

The Cuntz algebra ${\mathcal{O}}_{\infty }$ is generated by countably many isometries with orthogonal ranges. In the nonunital case, the relevant central-sequence algebra is the relative commutant in the norm ultrapower modulo the annihilator of $A$. That quotient supplies the unit needed for the absorption interface.

These separability and nuclearity hypotheses belong to the absorption conclusion, not to the individual-to-strong theorem. The zero algebra is included.

As a concrete nonunital example, $C_0((0,1])\otimes {\mathcal{O}}_{\infty }$ absorbs ${\mathcal{O}}_{\infty }$. It has no nonzero projections: the norm of a continuous projection-valued function is either zero or one, is constant on the connected interval, and must tend to zero at the missing endpoint.

🏷️ Key steps and proof outlines

Uniformly bounded witnesses make the product contradiction legitimate

The source’s “Uniform separation” section first proves permanence of individual proper infiniteness under arbitrary bounded products. For a positive contraction $d_i$, put $h_i=(d_i-\delta {)}_{+}$ and

$$k_i=(d_i-2\delta {)}_{+}\oplus (d_i-2\delta {)}_{+}.$$

Proper infiniteness and hereditary transport give a rectangular $r_i$ in the original algebra with $r_i^{\ast }{r}_i=k_i$, $e_{h_i}{r}_i=r_i$, and $\Vert r_i\Vert \le 1$. Its construction may use a bidual partial isometry, but multiplying by $k_i^{1/2}$ ensures the resulting $r_i$ belongs to the algebra. Define the continuous function

$${\varphi }_{\delta }(t)=\{\begin{array}{ll}t/{\delta }^{3/2},& 0\le t\le \delta ,\\ t^{-1/2},& t\ge \delta .\end{array}$$

It vanishes at zero and has norm at most ${\delta }^{-1/2}$. Since $r_i$ is supported where $d_i>\delta$, the row $s_i={\varphi }_{\delta }(d_i)r_i$ satisfies $d_i^{1/2}{s}_i=r_i$. Hence $s_i^{\ast }{d}_i{s}_i=k_i$ with a bound independent of $i$. This places the whole row sequence in the bounded product. Letting $\delta \downarrow 0$ gives proper infiniteness there. (OpenAI, 2026)

Now suppose no universal contraction factor existed. Choose the $m$th example so that every admissible pair of hereditary supports has mixed norm greater than $1-1/m$. Strict shrinking in their product produces one pair with mixed norm $\kappa <1$. The normal extensions of coordinate maps carry support projections to their coordinate supports, so every coordinate has mixed norm at most $\kappa$. Composing a coordinate representation with the product map preserves all the required norm signals. For large $m$, $1-1/m>\kappa$, a contradiction. This uses the coordinate maps; it does not identify the product’s bidual with a product of biduals.

Iterate inside the algebra while retaining every norm signal

A support projection may lie only in the bidual, so the next normalized mixed input cannot simply be an arbitrary support compression. Use instead

$$f(t)=(2t-1{)}_{+},g(t)=\min \{1,(4t-1{)}_{+}\}.$$

Both functions vanish near zero, $f(1)=1$, and $g=1$ on $[1/2,1]$. Suppose the current supports satisfy $\Vert e_{p_k}c{e}_{q_k}\Vert \le {\theta }^k$. Then

$$c^{(k)}={\theta }^{-k}g(p_k)cg(q_k)$$

belongs to the original algebra and is a contraction. Apply the uniform shrinking lemma to $f(p_k),f(q_k),c^{(k)}$. The new supports lie where $g(p_k)$ and $g(q_k)$ act as the identity, so

$$e_{p_{k+1}}c{e}_{q_{k+1}}={\theta }^k{e}_{p_{k+1}}{c}^{(k)}{e}_{q_{k+1}}.$$

Their mixed norm is at most ${\theta }^{k+1}$. Hereditary membership is nested. If a representation gave $p_k$ norm one, spectral mapping gives $f(p_k)$ norm one; the shrinking lemma therefore gives $p_{k+1}$ norm one in that same representation. Thus the original signals persist through all iterations, not merely in a representation selected afterward.

Reserve enough contraction for the final comparison coefficients

The final implementation multiplies the small support compression by coefficients whose norms are bounded by ${\delta }^{-1/2}$. Accordingly, shrink first below $\delta \zeta$, where $\zeta$ is the desired final mixed error. The retained signals ensure that the positive diagonal cuts belong to the ideals generated by the new supports. Proper-target comparison then implements them, while the lower spectral bound $a>\delta$ on the chosen support gives $\delta s^{\ast }s\le s^{\ast }as$ and the coefficient bound.

Both sides together can enlarge the mixed norm by at most ${\delta }^{-1}$, leaving it below $\zeta$. The diagonal error is separately at most $2\delta +\rho$. This ordering permits arbitrarily accurate mixed separation without sending the coefficient bounds to infinity for that reason alone. The distinct exactness argument is needed only to obtain individual proper infiniteness from a fixed amplification; the product and iteration mechanism above starts with individual proper infiniteness already available.

🏷️ Links and reading scope

• Family 299 also passes from comparison constructions to absorption through norm central sequences; the absorbing algebra and hypotheses differ.
• Family 301 uses extended traces to retain ideal information; here represented norm signals retain the ideals needed for comparison.
• The collection guide gives the surrounding operator-analysis sequence.

The selected-source audit covers the introduction, the separation and bounded-coefficient argument, selected comparison/transport and graph passages, the exact obstruction and perturbation statements, primitive slices and the square-zero interface, the final halving proof, and the absorption interface. The complete four-entry, exact perturbation, localization, independent nuclear, and alternate quantitative proofs were not independently verified.

🏷️ Manuscripts and verification scope

• Weak pure infiniteness and O-infinity absorption — pinned manuscript and companion source.

The family scope page describes selected formalization coverage. The documented selections are individual proper infiniteness implying full positive-element diagonalization with no exactness or nuclearity assumption, and fixed-amplification removal under exactness. The tensor-absorption corollary additionally requires separability and nuclearity and uses an external central-sequence absorption interface. The manuscript does not settle weak-to-ordinary pure infiniteness for arbitrary nonexact algebras. 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. Weak pure infiniteness and O∞ absorption.