Finite factor generators: uniform perturbations and matrix encoding

Source and verification

This article explains family 296 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: operator generation and finite factors

For operators $S$ inside a von Neumann algebra $M$, write $W^{\ast }(S)$ for the smallest unital algebra containing $S$ and closed under adjoints and weak operator limits. A single generator is a bounded operator $x$ with $W^{\ast }(x)=M$.

This is equivalent to generation by two selfadjoint operators. Indeed,

$$\begin{array}{c}a=\frac{x+x^{\ast }}{2},b=\frac{x-x^{\ast }}{2i},\\ x=a+ib.\end{array}$$

give $W^{\ast }(x)=W^{\ast }(a,b)$. A single selfadjoint operator, or even a single normal operator, generates a commutative algebra. A generator of a noncommutative factor must therefore be nonnormal.

A factor is a von Neumann algebra with center $\mathbb{C}1$. A type ${\mathrm{II}}_1$ factor is an infinite-dimensional finite factor with a normalized faithful normal trace $\tau$. Here $\tau (1)=1$, $\tau (xy)=\tau (yx)$, and $\tau (x^{\ast }x)>0$ for $x\ne 0$. Its trace Hilbert space is the completion of $M$ in the norm

$$\Vert x{\Vert }_2=\tau (x^{\ast }x{)}^{1/2}.$$

Normality of the trace means that it preserves increasing bounded suprema of positive operators. The trace supplies a notion of small perturbation weaker than the operator norm.

The predual $M_{\ast }$ consists of the normal linear functionals on $M$. Separable predual is the countability assumption used here; it does not mean that $M$ is separable in operator norm, and it does not impose separability on a chosen representing Hilbert space.

For a unital inclusion of factors $P\subset M$, the relative commutant $P^{\prime }\cap M$ consists of the elements of $M$ commuting with every element of $P$. The inclusion is irreducible if this commutant equals $\mathbb{C}1$. The relative generator question asks whether one extra unitary $u$ suffices to give $W^{\ast }(P,u)=M$.

🏷️ The two generation theorems

Relative generation

The manuscript states that for an irreducible inclusion $P\subset M$ of type ${\mathrm{II}}_1$ factors, with $M$ of separable predual, the set

$$\{u\in \mathcal{U}(M):W^{\ast }(P,u)=M\}$$

is a dense $G_{\delta }$ subset of $\mathcal{U}(M)$ in the normalized trace $2$-norm topology. Here $G_{\delta }$ means a countable intersection of open sets. (OpenAI, 2026)

Thus every unitary can be approximated in trace $2$-norm by a relative generator, and relative generation holds on a Baire-generic set. This is a topological conclusion, not a probability-one statement for a specified distribution of unitaries.

Single generation

Every type ${\mathrm{II}}_1$ factor with separable predual has two selfadjoint generators, equivalently one bounded operator generator. (OpenAI, 2026)

The paper obtains this second assertion by applying relative generation over an irreducible hyperfinite subfactor, followed by an explicit matrix encoding. It also invokes a classical direct-integral reduction to deduce single generation for all von Neumann algebras with separable predual. The finite-factor proof and the imported general reduction should be distinguished.

🏷️ Projection norms and Baire continuity

For a unitary $u$, put $N_u=W^{\ast }(P,u)$ and let $e_u$ be the orthogonal projection from $L^2(M,\tau )$ onto $L^2(N_u,\tau )$. For a fixed vector $\xi$, consider

$$f_{\xi }(u)=\Vert e_u\xi {\Vert }_2.$$

This measures how much of $\xi$ can be seen inside the algebra generated by $P$ and $u$. The distance to that subspace satisfies

$$\begin{array}{c}{d}_{\xi }(u)=\underset{b}{\inf }\Vert \xi -b(u){\Vert }_2,\\ f_{\xi }(u{)}^2=\Vert \xi {\Vert }_2^2-d_{\xi }(u{)}^2,\end{array}$$

where $b$ ranges over finite noncommutative polynomials in $u,u^{\ast }$ with coefficients in $P$. Such polynomials are dense in $L^2(N_u,\tau )$. For each fixed polynomial, telescoping products shows that $u\mapsto b(u)$ is $2$-norm continuous. The infimum is therefore upper semicontinuous, so $f_{\xi }$ is lower semicontinuous.

The unitary group is complete in its $2$-norm metric. Baire’s theorem gives a dense $G_{\delta }$ set on which any bounded lower semicontinuous real function is continuous. Apply this to a countable dense family of vectors in $L^2(M,\tau )$. The estimate

$$|f_{\xi }(u)-f_{\eta }(u)|\le \Vert \xi -\eta {\Vert }_2$$

extends simultaneous continuity to every vector $\xi$ at each point of the resulting dense set $\mathcal{C}$.

The proof will show that each $u\in \mathcal{C}$ is a relative generator. If $N_u$ is proper, choose a unitary $z\notin N_u$ and let

$$\xi =z-E_{N_u}(z),s=\Vert \xi {\Vert }_2>0,$$

where $E_{N_u}$ is the trace-preserving conditional expectation. On bounded elements it is the orthogonal projection $e_u$, and it satisfies $E_{N_u}(axb)=a{E}_{N_u}(x)b$ for $a,b\in N_u$. Thus

$$e_u\xi =0,\tau ({\xi }^{\ast }z)=s^2.$$

The remaining task is to perturb $u$ slightly so that a word in the new generated algebra has a fixed positive correlation with $\xi$. That would contradict continuity of $f_{\xi }$ at $u$.

🏷️ A perturbation detected by long words

The manuscript proves the following estimate. Given unitaries $u,z$, a bounded $\xi \in M$, and $0<\alpha <1/2$, for every integer $\ell \ge 1$ there are a unitary $\tilde{u}$ and unitaries ${\tilde{t}}_0,\dots ,{\tilde{t}}_{\ell }\in P$ such that

$$\begin{array}{c}\tilde{W}={\tilde{t}}_0\prod _{k=1}^{\ell }(\tilde{u}{\tilde{t}}_k),\\ \Vert \tilde{u}-u{\Vert }_2\le \frac{\alpha }{\sqrt{2\ell }}+\frac{1}{\ell },\end{array}$$

and

$$\begin{array}{c}\left|\tau ({\xi }^{\ast }\tilde{W})-\frac{\alpha }{2}\tau ({\xi }^{\ast }z)\right|\\ \le \Vert \xi {\Vert }_2\frac{4{\alpha }^2}{1-2\alpha }+\frac{1}{\ell }.\end{array}$$

Products are ordered from left to right in increasing $k$. The perturbation tends to zero as $\ell$ grows, but the desired linear correlation remains of order $\alpha$. The nonlinear error is of order ${\alpha }^2$, uniformly in the word length. Establishing that uniformity is the central estimate.

The construction first takes place in the tracial ultrapower $M^{\omega }$: bounded sequences are identified when their $2$-norm difference tends to zero along a free ultrafilter. This permits exact independence relations before a finite list of estimates is transferred back to $M$.

An imported independence theorem of Popa supplies successive unitaries $t_j\in P^{\omega }$ that are Haar and free from the prescribed data and earlier choices. A Haar unitary has $\tau (t^n)=0$ for every nonzero integer $n$. Freeness means that alternating products of centered elements from different participating algebras have trace zero.

Here irreducibility must hold for the full inclusion $P^{\omega }\subset M^{\omega }$. The paper proves this using the uniform conjugation displacement

$${\delta }_n=\underset{w\in \mathcal{U}(P)}{\sup }\Vert w{b}_n{w}^{\ast }-b_n{\Vert }_2.$$

An element commuting with every sequence of unitaries has ${\delta }_n{\to }_{\omega }0$. The minimum-norm point in the convex hull of the conjugates of $b_n$ is scalar by irreducibility of $P\subset M$. Since it equals $\tau (b_n)1$, the distance of $b_n$ from the scalars tends to zero. This concerns all of $P^{\omega }$, not merely the constant copy of $P$.

🏷️ Coordinating the repeated perturbations

Work in the ultrapower for a fixed $\ell$. Choose the free unitaries described above, starting with the data algebra $C^{\ast }(1,u,z,\xi )$, and set

$$W=t_0\prod _{k=1}^{\ell }(u{t}_k).$$

Let $g_k$ be the prefix immediately before the $k$th occurrence of $u$. Thus $g_1=t_0$ and $g_{k+1}=g_{k}u{t}_k$. The last fresh Haar unitary ensures that $W$ is Haar and free from the algebra containing the data and the prefixes. So is $y=z{W}^{\ast }$.

Define a selfadjoint perturbation by

$$\begin{array}{c}h=\frac{\alpha }{\ell }\sum _{j=1}^{\ell }\mathrm{Im}(g_j^{-1}y{g}_j),\\ u^{\prime }=e^{ih}u,W^{\prime }=t_0\prod _{k=1}^{\ell }(u^{\prime }{t}_k).\end{array}$$

The conjugates appearing in this sum form a free Haar family. Their imaginary parts are orthogonal in $L^2$, each with squared norm $1/2$. Hence

$$\Vert h\Vert \le \alpha ,\Vert h{\Vert }_2=\frac{\alpha }{\sqrt{2\ell }}.$$

Moving the perturbation at each occurrence to the front of the word gives the exact ordered identity

$$\begin{array}{c}{W}^{\prime }=\left(\prod _{k=1}^{\ell }{e}^{i{H}_k}\right)W,\\ H_k=g_{k}h{g}_k^{-1}.\end{array}$$

Every diagonal choice $j=k$ contributes $y$ before multiplication by $W$, hence contributes $z$. Its coefficient is $\alpha /(2\ell )$. The $\ell$ such contributions add coherently, while freeness makes all other linear terms vanish against $\xi$. With $C_1={\sum }_{k}i{H}_k$,

$$\tau ({\xi }^{\ast }(1+C_1)W)=\frac{\alpha }{2}\tau ({\xi }^{\ast }z).$$

The same perturbation is used at every occurrence of $u$. Independent small perturbations at the separate positions would not establish this identity for a word in one new unitary.

🏷️ A remainder bound independent of word length

To control the exponential remainder, write $p_{kj}=g_k{g}_j^{-1}$. The free prefix letters show that all off-diagonal $p_{kj}$ are distinct nonidentity words. The identity occurs exactly $\ell$ times. If $n_p$ is the number of occurrences of $p$, then

$$\sum _p{n}_p^2=2{\ell }^2-\ell .$$

The conjugates $Y_p=py{p}^{-1}$ form a free Haar family. Introduce a formal free generator $T_p$ for each label $p$ and put $c=\alpha /(2\ell )$. Even the label $p=e$ corresponds to a generator $T_e$, not the identity of this new free group.

Replace every coefficient in the linear exponentials by its absolute value and collect the resulting positive majorant

$$A=c\sum _p{n}_p(T_p+T_p^{-1}).$$

The length-one free convolution inequality gives

$$\begin{array}{rl}\Vert \lambda (A)\Vert & \le 2{\left(2c^2\sum _p{n}_p^2\right)}^{1/2}\\ & \le 2\alpha <1.\end{array}$$

Here $\lambda (A)$ acts by left convolution on square-summable functions on the formal free group. The basic inequality follows by splitting each letter’s action according to whether it cancels the first letter of the input word: orthogonal ranges control the noncancellation part, and Cauchy–Schwarz controls the cancellation part.

Expand the ordered product of exponentials by total degree before free reduction, and denote its degree-$d$ contribution by $C_d$. Taking coefficientwise absolute values bounds it by $A^d$. This comparison includes contributions that reduce to shorter words, including the identity. It uses neither commutation of the factors nor orthogonality between different degrees.

Freeness identifies the $2$-norm of a finite sum of words in the $Y_p$ with the Euclidean norm of its reduced-word coefficients. Consequently

$$\Vert C_d{\Vert }_2\le \Vert \lambda (A){\Vert }^d\le (2\alpha {)}^d.$$

Summing the geometric tail yields

$$\begin{array}{c}{\Vert \prod _{k=1}^{\ell }{e}^{i{H}_k}-1-C_1\Vert }_2\\ \le \sum _{d\ge 2}(2\alpha {)}^d=\frac{4{\alpha }^2}{1-2\alpha }.\end{array}$$

For each fixed $\ell$, the exponential expansion converges in operator norm, so its regrouping and passage to the $2$-norm tail are justified. The displayed estimate is the separate bound that remains uniform as $\ell$ varies.

Cauchy–Schwarz now gives the desired correlation error in the ultrapower. Choose unitary representatives in $P$ and one coordinate satisfying the finitely many required estimates within $1/\ell$. The scalar inequality $|e^{it}-1|\le |t|$ controls $\Vert u^{\prime }-u{\Vert }_2$. A coordinate is chosen separately for each $\ell$; no uniform coordinate or exact free family inside $M$ is claimed.

🏷️ From the estimate to relative generation

Return to the nonzero vector $\xi =z-E_{N_u}(z)$ at a common continuity point. Choose $\alpha$ once, with

$$0<\alpha <\min \{1/4,s/32\}.$$

Then the nonlinear error is less than $\alpha s^2/4$. For all sufficiently large $\ell$, the perturbation estimate supplies ${\tilde{u}}_{\ell }\to u$ and a unitary word ${\tilde{W}}_{\ell }\in N_{{\tilde{u}}_{\ell }}$ with

$$\begin{array}{c}|\tau ({\xi }^{\ast }{\tilde{W}}_{\ell })|\ge \frac{\alpha s^2}{8},\\ f_{\xi }({\tilde{u}}_{\ell })\ge |\tau ({\xi }^{\ast }{\tilde{W}}_{\ell })|.\end{array}$$

The second inequality uses $\Vert {\tilde{W}}_{\ell }{\Vert }_2=1$ and orthogonal projection. It contradicts continuity at $u$, where $f_{\xi }(u)=0$. Thus every point of $\mathcal{C}$ generates relatively.

To identify the entire relative-generator set as $G_{\delta }$, take a dense sequence ${\eta }_j$ in $L^2(M,\tau )$ and write it as

$$\bigcap _{j,n\ge 1}\{v\in \mathcal{U}(M):d_{{\eta }_j}(v)<1/n\}.$$

Each set is open by upper semicontinuity. Vanishing of all these distances means $L^2(N_v)=L^2(M)$, and the conditional expectation then gives $N_v=M$.

🏷️ Encoding four generators into two

An imported theorem of Popa places an irreducible hyperfinite subfactor $R$ inside every separable-predual type ${\mathrm{II}}_1$ factor. “Hyperfinite” means that finite-dimensional subalgebras approximate it in the von Neumann topology. Its standard tensor model is the infinite tensor product of $M_2(\mathbb{C})$ with normalized traces.

At tensor site $m$, let $d_m$ be the projection onto the first coordinate and $q_m$ the projection onto $(1,1)$. These generate the site’s matrix algebra. Each family separately consists of commuting projections. Encode a family $(e_m)$ by the norm-convergent series

$$b=\sum _{m\ge 1}23^{-m}{e}_m.$$

After earlier projections have been recovered, the residual spectrum lies in $[0,3^{-m}]$ on $1-e_m$ and in $[23^{-m},33^{-m}]$ on $e_m$. A spectral cut between these intervals recovers $e_m$. Applying this separately to $(d_m)$ and $(q_m)$ gives two selfadjoint generators of $R$.

Relative generation adds one unitary. Its real and imaginary parts therefore give a uniform bound of four selfadjoint generators for every factor in the class.

Apply that bound to the half-trace corner $N=pMp$, where $\tau (p)=1/2$ and the normalized corner trace is ${\tau }_N=2\tau {|}_N$. Equivalent projections $p$ and $1-p$ give $M\cong M_2(N)$. Write $N=W^{\ast }(a_1,a_2,a_3,a_4)$ with selfadjoint $a_j$, and choose $C>\Vert a_1\Vert +\Vert a_2\Vert$. Set

$$\begin{array}{c}A=\left(\begin{array}{cc}{a}_1& 0\\ 0& a_2+C1\end{array}\right),\\ B=\left(\begin{array}{cc}{a}_3& 1\\ 1& a_4\end{array}\right).\end{array}$$

The two diagonal spectral intervals of $A$ are disjoint, so spectral calculus recovers the diagonal matrix projections $E_{11},E_{22}$. Cutting $B$ gives $E_{12}=E_{11}B{E}_{22}$ and its adjoint. These matrix units transfer all four $a_j$ into the first corner; for example,

$$\begin{array}{rl}{a}_2{E}_{11}& =E_{12}A{E}_{21}-C{E}_{11},\\ a_4{E}_{11}& =E_{12}B{E}_{21}.\end{array}$$

Thus $W^{\ast }(A,B)=M_2(N)$. The operator $A+iB$, transported through the corner isomorphism, is the asserted single generator. This is an explicit recovery argument once the four corner generators are given.

🏷️ Key steps and proof outlines

Verify irreducibility for the whole ultrapower inclusion

Suppose $b=[(b_n)]$ commutes with every element of $P^{\omega }$, and put

$${\delta }_n=\underset{w\in \mathcal{U}(P)}{\sup }\Vert w{b}_n{w}^{\ast }-b_n{\Vert }_2.$$

If ${\delta }_n$ failed to tend to zero along the ultrafilter, choose at each bad coordinate a witnessing unitary $w_n$. Their sequence would be a unitary in $P^{\omega }$ not commuting with $b$. This coordinatewise choice is legitimate because the commutation assumption concerns every sequence, not only constant unitaries. (OpenAI, 2026)

For each $n$, take the minimum-$2$-norm point of the closed convex hull of the conjugates of $b_n$. Uniqueness makes it invariant under conjugation by $P$. The hull is operator-norm bounded, so its $2$-norm limit remains in $M$. Irreducibility then makes that point scalar, and trace preservation identifies it as $\tau (b_n)1$. Every point of the hull is within ${\delta }_n$ of $b_n$, proving

$$\Vert b_n-\tau (b_n)1{\Vert }_2\le {\delta }_n{\to }_{\omega }0.$$

This verifies the precise relative-commutant hypothesis needed before importing free Haar unitaries into the construction.

Coherent linear terms survive while repeated labels control the error

The perturbation uses one selfadjoint $h$ repeatedly in a long word. After moving each occurrence to the front, the conjugation labels are $p_{kj}=g_k{g}_j^{-1}$. All off-diagonal labels are distinct; the identity occurs $\ell$ times. If $n_p$ is a label’s multiplicity, then

$$\sum _p{n}_p^2=({\ell }^2-\ell )+{\ell }^2=2{\ell }^2-\ell .$$

That square sum is why the nonlinear estimate stays bounded as word length grows. With coefficient $c=\alpha /(2\ell )$, the free convolution bound sees the square root of the multiplicity sum, cancelling the denominator $\ell$. It gives a positive coefficient majorant of norm at most $2\alpha$.

At first order, precisely the repeated identity label gives the desired $\ell$ coherent copies of $z$, each of coefficient $\alpha /(2\ell )$. All other linear terms vanish against the chosen test by freeness. At higher orders, take absolute coefficients before reducing words. Reduction can merge many terms and can produce the identity, so orthogonality by total degree cannot be assumed. The coefficientwise majorant instead bounds the degree-$d$ term in $2$-norm by $(2\alpha {)}^d$. Summing $d\ge 2$ gives $4{\alpha }^2/(1-2\alpha )$ uniformly in $\ell$.

Fix the signal-to-error ratio before shrinking the perturbation

At a Baire continuity point that does not generate, take $\xi =z-E_{N_u}(z)$ and write $s=\Vert \xi {\Vert }_2>0$. Then $\tau ({\xi }^{\ast }z)=s^2$, and $e_u\xi =0$. Choose $0<\alpha <\min (1/4,s/32)$ once. Since $1-2\alpha >1/2$,

$$s\frac{4{\alpha }^2}{1-2\alpha }<8s{\alpha }^2<\frac{\alpha s^2}{4}.$$

The linear correlation is $\alpha s^2/2$, leaving a positive margin before choosing $\ell$. Now let $\ell \to \infty$: the perturbation size $\alpha /\sqrt{2\ell }+1/\ell$ tends to zero, while the correlation remains at least $\alpha s^2/8$ eventually. Because the witness word is a unit vector in the new generated algebra, its correlation bounds the projection norm $f_{\xi }$ from below. This contradicts continuity where $f_{\xi }(u)=0$.

Choosing $\alpha$ to shrink with $\ell$ would risk making that lower bound vanish and would not give the same continuity contradiction. The two parameters have distinct roles: $\alpha$ separates signal from nonlinear error, while $\ell$ makes the perturbation arbitrarily small.

The final encoding is a recovery procedure

After relative generation supplies four selfadjoint generators for a half-trace corner $N$, the two matrices $A,B$ above encode them in $M_2(N)$. The scalar shift in the second diagonal block of $A$ separates its spectrum from the first, so functional calculus recovers the two diagonal matrix units. The off-diagonal identity in $B$ then recovers the remaining matrix units.

Cutting and transporting diagonal blocks recovers all four original corner generators. Thus the encoding proves equality of generated algebras by explicitly decoding every ingredient. Taking $A+iB$ finishes single generation. This elementary final step depends on the earlier uniform four-generator bound applying again to the half-trace corner; it does not itself prove relative generation.

🏷️ Consequences and connections

The manuscript’s general von Neumann algebra conclusion uses the established direct-integral reduction in addition to the finite-factor theorem. It does not assert that generators can be chosen continuously across a family of factors.

It also records a consequence for microstates free entropy dimensions, which measure asymptotic dimensions of spaces of matrix approximations to a selfadjoint tuple. A free semicircular $n$-tuple has both dimensions $\delta ={\delta }_0=n$ and generates the free group factor $L({\mathbb{F}}_n)$. A generating selfadjoint pair supplied here has both dimensions at most two. For $n\ge 3$, these give different dimension values on tuples generating the same tracial factor. The argument does not determine the pair’s exact dimensions. (OpenAI, 2026)

Bounded cohomology and harmonic rigidity also uses exact free unitaries in a tracial ultrapower and free-word estimates. Here they control a nonlinear perturbation uniformly in word length, enabling the Baire argument. The final matrix encoding is separate from that analytic construction.

🏷️ Manuscripts and verification scope

• Relative generation and the generator problem for finite factors — pinned manuscript and companion source.

The family scope page describes selected formalization coverage. The documented selections cover single generation of type II1 factors with separable predual, and dense G-delta relative unitary generators for irreducible inclusions with separable predual for the larger factor, in normalized trace two-norm. They impose no operator-norm separability or separability of a representing Hilbert space. The general von Neumann algebra consequence additionally uses an imported direct-integral reduction. The entropy-dimension comparison is a further manuscript consequence, rather than a selected statement on that scope page. 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. Relative generation and the generator problem for finite factors.