Free group factors: small flows and a generating limit

Source and verification

This article explains family 287 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

The manuscript claims that all nonabelian free group factors, including the interpolated family, are isomorphic. Its constructive step absorbs one free generator into a slightly moving tuple of at least three others. A polynomial flow preserves moments, a spectral calculation supplies its small coefficients, and an adaptive limiting argument proves that the final tuple generates the entire factor. Moment preservation and generation are separate obligations.

🏷️ Tracial distributions and free generators

Let $(M,\tau )$ be a von Neumann algebra with faithful normal tracial state. We use the operator norm $\Vert x\Vert$ and the tracial $L^2$ norm $\Vert x{\Vert }_2=\tau (x^{\ast }x{)}^{1/2}$; the latter completes $M$ to a Hilbert space $L^2(M,\tau )$. Since $\tau (1)=1$, $\Vert x{\Vert }_2\le \Vert x\Vert$.

The joint tracial distribution of a tuple consists of all numbers $\tau (p(X))$, where $p$ is a noncommutative polynomial in the tuple and its adjoints. Unital subalgebras are free when every alternating product of centered elements from successively different subalgebras has trace zero. Here centered means $\tau (x)=0$.

A free Haar $n$-tuple is a tuple of unitaries $A_1,\dots ,A_n$ such that every nonidentity reduced word has trace zero. If $A_g$ denotes evaluation of $g\in {\mathbb{F}}_n$, then

$$\tau (A_h^{\ast }{A}_g)=1_{\{g=h\}}.$$

These vectors therefore form an orthonormal family. If their word algebra is dense in $L^2(M)$, the unitary ${\delta }_g\mapsto A_g$ identifies the left regular representation of ${\mathbb{F}}_n$ with left multiplication on $L^2(M)$. It yields a normal trace-preserving isomorphism $L({\mathbb{F}}_n)\cong M$.

The density requirement is essential. A free Haar tuple inside a larger factor gives the right moments for a copy of $L({\mathbb{F}}_n)$, but its generated von Neumann algebra may be proper.

Constructive rank comparison claimed by the manuscript

For every integer $n\ge 3$, $L({\mathbb{F}}_n)$ is normally and trace-preservingly isomorphic to $L({\mathbb{F}}_{n+1})$. The proof constructs a free Haar $n$-tuple generating $L({\mathbb{F}}_{n+1})$. Established amplification formulas then give $L({\mathbb{F}}_2)\cong L({\mathbb{F}}_3)$. (OpenAI, 2026)

The article explains the construction and its proof interfaces. It does not certify the complete research theorem independently.

🏷️ A bounded logarithm and support-independent norm control

Start with a freely generating Haar tuple $(A_1,\dots ,A_n,C)$ in $M=L({\mathbb{F}}_{n+1})$. Bounded Borel functional calculus defines $S=\arg (C)$, with argument in $(-\pi ,\pi ]$. Since $C$ has uniform circle distribution, $S=S^{\ast }$, $\tau (S)=0$, $\Vert S\Vert \le \pi$, and $e^{iS}=C$. Its Borel definition is legitimate in the von Neumann algebra; continuity of the argument on the circle is not asserted.

For $g\in {\mathbb{F}}_n$, put $S_g=A_{g}S{A}_g^{\ast }$. These conjugate component algebras form a free family. For example, moving adjacent group factors together in a product of centered conjugates gives alternating centered terms from $W^{\ast }(C)$ and nonidentity words in the $A_j$. Freeness makes their trace zero.

For finitely supported real coefficients $k$, write $s(k)={\sum }_{g}k(g)S_g$. A central estimate is

$$\Vert s(k)\Vert \le 3\pi \Vert k{\Vert }_{{\ell }^2}.$$

The constant does not grow with the number of nonzero coefficients. To obtain it, decompose the free-product trace Hilbert space into reduced centered words. Multiplication by a centered self-adjoint component $z_j$ has three parts: prefix creation, its adjoint annihilation, and a block acting on words already starting in that component. The creation ranges are orthogonal; the remaining diagonal blocks act on orthogonal subspaces. Thus

$$\begin{array}{rl}\Vert \sum _j{z}_j\Vert & \le 2{\left(\sum _j\Vert z_j{\Vert }_2^2\right)}^{1/2}\\ & +\underset{j}{\max }\Vert z_j\Vert .\end{array}$$

Apply this to $z_g=k(g)S_g$ and use ${\max }_g|k(g)|\le \Vert k{\Vert }_{{\ell }^2}$. This estimate converts small coefficient vectors into small operator displacements later. (OpenAI, 2026)

🏷️ Polynomial flows and the trace identity

Choose finitely supported real vectors $h_1,\dots ,h_n$ on ${\mathbb{F}}_n$. Extend them to a coefficient-valued cocycle by

$$\begin{array}{rl}{D}_{x_j}& =h_j,\\ D_{gb}& =D_g+\lambda (g)D_b.\end{array}$$

Here $\lambda$ is left translation on coefficient vectors. An inverse letter contributes $D_{x_j^{-1}}=-\lambda (x_j^{-1})h_j$, so adjacent inverse letters cancel and the rule is well defined on reduced words.

The proposed flow fixes $S$ and evolves the other generators by

$$\begin{array}{rl}{\dot{A}}_j(t)& =i{s}_t(h_j)A_j(t),\\ s_t(k)& =\sum _{g}k(g)A_g(t)S{A}_g(t{)}^{\ast }.\end{array}$$

This is a finite polynomial differential equation in the real Banach space underlying $B=C^{\ast }(A_1,\dots ,A_n,S)$. Reality of the coefficients makes $s_t(h_j)$ self-adjoint. Differentiating $A_j(t{)}^{\ast }{A}_j(t)$ shows it remains the identity; faithfulness of the finite trace also gives $A_j(t)A_j(t{)}^{\ast }=1$. Bounded derivatives on this unitary trajectory allow continuation for all real times.

Global solutions alone do not show that substitution gives an automorphism. The source first proves a trace identity on the formal algebra generated by unitary symbols $a_j$ and a self-adjoint symbol $z$. Its derivation has $\delta z=0$ and $\delta a_j=i\sigma (h_j)a_j$, where $z_g=a_{g}z{a}_g^{\ast }$ and $\sigma (k)={\sum }_{g}k(g)z_g$.

Ignoring every $z$ in a monomial leaves its group label. Differentiation preserves that label. Nonidentity labels evaluate to trace zero. An identity-label monomial rewrites as a polynomial in the $z_g$, whose differentiated directions are commutators.

At the initial tuple, the direction for $S_g$ is $i[Y_g,S_g]$, where $Y_g$ belongs to the algebra generated by the other $S_b$. Conjugating $S_g$ by $e^{iu{Y}_g}$ preserves its joint law with that other algebra: inner conjugation there preserves freeness and the marginal distributions. Differentiating this scalar trace identity at zero and summing over the variables gives

$$\tau ({\mathrm{ev}}_0(\delta p))=0.$$

Applying it to ${\delta }^{r-1}p$ gives vanishing for every positive derivative order. These auxiliary one-variable conjugations only compute first variations; the argument does not assume they jointly integrate to the desired flow.

From initial derivatives to actual automorphisms

For the actual polynomial solution, repeated differentiation gives $\frac{d^r}{d{t}^r}{\mathrm{ev}}_t(p)={\mathrm{ev}}_t({\delta }^{r}p)$. Finite support bounds the growth of degrees and coefficients. For each fixed $p$, the resulting derivatives satisfy a bound of the form $\Vert {\mathrm{ev}}_t({\delta }^{r}p)\Vert /r!\le C_p{c}_p^r$, uniformly in real $t$. Taylor’s remainder then proves real analyticity of the moments.

All positive-order derivatives vanish at zero, so every moment is constant on the real line by analytic continuation. Freeness at positive times is obtained from this conclusion; it is not an input to the differential equation estimates.

Moment preservation applied to $p^{\ast }p$ shows that every relation holding at the initial tuple still holds at time $t$. Faithfulness therefore makes substitution well defined. The identity

$$\Vert b\Vert =\underset{r\to \infty }{\lim }\tau ((b^{\ast }b{)}^r{)}^{1/(2r)}$$

shows it is isometric. For surjectivity, solve the same equation backwards inside $C^{\ast }(A_1(t),\dots ,A_n(t),S)$; uniqueness recovers the original generators. The trace-Hilbert-space unitary implementing substitution then gives a normal automorphism ${\beta }_t$ of $M$. Point-norm continuity is asserted on $B$, without requiring it on every element of $M$. (OpenAI, 2026)

🏷️ Small coefficients with a large word effect

The cocycle rule differentiates a word as ${\dot{A}}_w(t)=i{s}_t(D_w)A_w(t)$. The free-sum estimate, now transported by the established automorphisms, gives

$$\begin{array}{c}\Vert {\beta }_1(A_j)-A_j\Vert \le 3\pi \Vert h_j{\Vert }_{{\ell }^2},\\ \Vert {\beta }_1(A_w)-C{A}_w\Vert \\ \le 3\pi \Vert D_w-{\delta }_e{\Vert }_{{\ell }^2}.\end{array}$$

For the second inequality, differentiate $e^{-itS}{A}_w(t)$. Its derivative has norm at most $3\pi \Vert D_w-{\delta }_e{\Vert }_{{\ell }^2}$ because $s_t({\delta }_e)=S$. The order $C{A}_w$ matters.

The coefficient problem is therefore to make every $h_j$ small while arranging $D_w$ close to the unit vector ${\delta }_e$. The manuscript uses $h_j=0$ for $j\ge 2$. With $a=x_1$, set

$$\begin{array}{rl}{b}_j& =x_2^j{x}_3{x}_2^{-j},\\ p_j& =(a{b}_1)\cdots (a{b}_j),\\ w_m& =p_{m}a,\\ T_m& =I+\sum _{j=1}^m\lambda (p_j).\end{array}$$

The $p_j$ form a free family, as follows from reduced-word cancellation and invertible changes of free basis. The selected occurrences of $a$ in $w_m$ have prefixes $e,p_1,\dots ,p_m$, while all other letters have zero coefficient. Thus $D_{w_m}=T_m{h}_1$. This step explains the restriction $n\ge 3$ in the construction.

Spectral inversion after a cutoff

Put $Q_m=T_m{T}_m^{\ast }$. Its normalized spectral measures, defined by ${\mu }_m(E)=\tau (1_E(Q_m/m))$, converge to

$$d\nu (x)=\frac{1}{2\pi }\sqrt{\frac{4-x}{x}}{1}_{(0,4)}(x)dx.$$

The proof counts moments. In the expansion of $\tau (Q_m^r)$, a word can reduce to the identity only if every nonzero index occurs at least twice. The leading terms have exactly $r$ distinct indices, paired in a noncrossing matching. Alternating adjoint signs make every such matching admissible. There are ${\mathrm{Cat}}_r$ matchings and $(m{)}_r$ distinct labelings; the other terms contribute $O_r(m^{r-1})$. Hence the normalized moments tend to the Catalan numbers.

Higher moment bounds give uniform integrability, while the first moment gives tightness. The Catalan growth bound forces any limiting measure to be supported on $[0,4]$, where polynomial approximation gives uniqueness. The displayed density realizes those moments and, crucially, has no atom at zero.

Choose a small fixed cutoff $d>0$, then large $m$. With $E=1_{(dm,\infty )}(Q_m)$ and $R=Q_m^{-1}E$ defined by bounded spectral calculus, set $h^{(0)}=T_m^{\ast }R{\delta }_e$. The exact estimates are

$$\begin{array}{rl}{T}_m{h}^{(0)}& =E{\delta }_e,\\ \Vert T_m{h}^{(0)}-{\delta }_e{\Vert }_{{\ell }^2}^2& ={\mu }_m([0,d]),\\ \Vert h^{(0)}{\Vert }_{{\ell }^2}^2& =\tau (R)\le (dm{)}^{-1}.\end{array}$$

The first error becomes small because the limit has no atom at zero; the coefficient norm becomes small by taking $m$ large. No spectral gap or full invertibility of $Q_m$ is needed.

All shifts are real matrices in the group basis, so spectral calculus preserves the real coefficient subspace. Approximate $h^{(0)}$ by a finitely supported real vector after fixing $m$, controlling the image error by $\Vert T_m\Vert \le m+1$. For every $\eta >0$ this yields $\Vert h_1{\Vert }_{{\ell }^2}<\eta$ and $\Vert D_{w_m}-{\delta }_e{\Vert }_{{\ell }^2}<\eta$. The support and word length may grow with the requested accuracy. (OpenAI, 2026)

🏷️ Absorption of the old complementary generator

The flow makes a word close to $C{A}_w$. To make it close to $C$ itself, use the free-basis automorphism $\gamma$ fixing each $A_j$ and sending $C$ to $C{A}_w$. Then $\alpha ={\gamma }^{-1}{\beta }_1$ satisfies, for any requested tolerance $\epsilon >0$,

$$\begin{array}{c}\underset{j}{\max }\Vert \alpha (A_j)-A_j\Vert <\epsilon ,\\ \Vert \alpha (A_w)-C\Vert <\epsilon .\end{array}$$

The target is the old $C$. The new complementary coordinate is $\alpha (C)=C{A}_w^{\ast }$, since ${\beta }_1$ fixes $S$ and therefore fixes $C$. The full new tuple remains freely generating Haar because $\alpha$ is an automorphism.

Iterating with merely summable generator displacements would preserve the limiting moments, but would not by itself preserve generation. The witness words can become arbitrarily long. The source instead chooses future error budgets after each witness is known.

🏷️ Adaptive approximation and generation of the limit

Fix a countable dense family $y_1,y_2,\dots$ in $L^2(M)$ from the original rational word algebra. At stage $k$, approximate the first $k$ targets by word polynomials $F_{j,k}$ in the current completed tuple $(A^{(k-1)},C^{(k-1)})$, with error below $2^{-k}$.

For a word polynomial $F={\sum }_{\nu }{c}_{\nu }{v}_{\nu }$, telescoping products of unitaries gives the Lipschitz constant $L(F)={\sum }_{\nu }|c_{\nu }||v_{\nu }|$ for the maximum coordinate $L^2$ distance. Choose the next perturbation tolerance using the finitely many constants $L(F_{j,k})$.

The one-step construction supplies a new tuple and a word $w_k$ approximating the old completion. Substitute this word to obtain $n$-variable polynomials $G_{j,k}(X)=F_{j,k}(X,X_{w_k})$. Their current target errors are below $2\cdot 2^{-k}$. At this point the word was treated as one unitary slot of $F_{j,k}$, so the preceding tolerance did not need to know its length.

Now compute a common Lipschitz bound $L_k$ for the substituted polynomials and choose

$$0<r_k\le \min \left\{\frac{r_{k-1}}{2},\frac{2^{-k}}{1+L_k}\right\}.$$

Future displacements are chosen below half the previous budget, so the total movement after stage $k$ is at most $r_k$. The first $n$ generators converge in operator norm to a unitary tuple $A^{(\infty )}$. Every fixed word retains its Haar trace, and

$$\begin{array}{c}\Vert G_{j,k}(A^{(\infty )})-y_j{\Vert }_2\\ \le 3\cdot 2^{-k}(j\le k).\end{array}$$

Thus the limiting word algebra is dense in all of $L^2(M)$. The trace-preserving conditional expectation onto $W^{\ast }(A^{(\infty )})$ is the orthogonal $L^2$ projection onto that subalgebra. It fixes a dense set and therefore is the identity. This proves generation and completes the claimed rank step. The complementary coordinates need not converge, and the approximating polynomials need not have uniformly bounded operator norms. (OpenAI, 2026)

🏷️ Rank two, interpolation, and entropy consequences

Write $N_r=L({\mathbb{F}}_r)$ for the interpolated free group factors. The imported amplification formula is $N_r^t\cong N_{1+(r-1)t^{-2}}$, where the corner trace is normalized. The constructive rank steps give $N_3\cong N_4\cong N_5$. Amplifying their composite by $\sqrt{2}$ gives

$$N_2\cong N_3^{\sqrt{2}}\cong N_5^{\sqrt{2}}\cong N_3.$$

The manuscript then invokes the classical dichotomy, including the infinite parameter: either all these factors are isomorphic or all distinct parameters give nonisomorphic factors. Its rank-two comparison selects the first alternative. The amplification formula then gives fundamental group ${\mathbb{R}}_{>0}$ for the common factor. These classical inputs are distinct from the constructive argument.

A further consequence concerns free entropy dimensions, numerical quantities attached to finite self-adjoint tuples. The microstates versions $\delta ,{\delta }_0$ measure asymptotic matrix-approximation size under small noise; the nonmicrostates versions ${\delta }^{\ast },{\delta }^{\star }$ use free entropy and free Fisher information. Each depends on joint tracial distribution.

The manuscript transports standard generating tuples from $L({\mathbb{F}}_n)$ to the single factor $L({\mathbb{F}}_2)$. Semicircular $n$-tuples give $\delta ={\delta }_0=n$. For the nonmicrostates calculation it uses the $2n$ self-adjoint real and imaginary parts of the canonical group unitaries, giving ${\delta }^{\ast }={\delta }^{\star }=n$. These two tuple families need not coincide. The conclusion concerns changing finite von Neumann generating tuples; their algebraic $\ast$-algebras are not held fixed. The external entropy calculations are used as inputs, rather than proved here. (OpenAI, 2026)

🏷️ Review scope

The README, main file, all six principal sections, and family scope page were read. The exposition follows the free-sum estimate, polynomial trace argument, Catalan spectral limit, cutoff inversion, old-completion absorption, and adaptive density proof. This source audit does not independently certify every analytic step or the cited amplification, dichotomy, and entropy theorems.

🏷️ Key steps and proof outlines

The one-step approximation targets the old completion

Let $(A_1,\dots ,A_n,C)$ be the current freely generating Haar tuple. The small-cocycle and flow inputs produce an automorphism $\beta$ with

$$\Vert \beta (A_j)-A_j\Vert <\epsilon ,\Vert \beta (A_w)-C{A}_w\Vert <\epsilon ,\beta (C)=C.$$

The multiplication order $C{A}_w$ matters. Let $\gamma$ fix the $A_j$ and send $C$ to $C{A}_w$, and put $\alpha ={\gamma }^{-1}\beta$. Then isometry of ${\gamma }^{-1}$ gives

$$\Vert \alpha (A_w)-C\Vert <\epsilon ,\alpha (C)=C{A}_w^{\ast }.$$

Thus a word in the new first $n$ generators approximates the old completion, while the new completion remains available to generate the ambient algebra at the next stage. Merely approximating the moving new completion would not preserve approximations already expressed in the old tuple. (OpenAI, 2026)

Choose tolerances before the witness word, then budget its later sensitivity

Fix an $L^2$-dense countable family $(y_j)$ in the ambient factor. At stage $k$, use the current completed tuple to choose word polynomials $F_{j,k}$, for $j\le k$, approximating those targets to within $2^{-k}$. For a word polynomial $F={\sum }_{\nu }{c}_{\nu }{v}_{\nu }$, define

$$L(F)=\sum _{\nu }|c_{\nu }||v_{\nu }|.$$

Telescoping products of unitaries gives $\Vert F(U)-F(V){\Vert }_2\le L(F){\max }_i\Vert U_i-V_i{\Vert }_2$. Let $H_k$ bound the constants of the finitely many $F_{j,k}$. Choose

$${\epsilon }_k<\min \left\{r_{k-1}/2,\frac{2^{-k}}{1+H_k}\right\},$$

then apply the one-step result and obtain its word $w_k$. Substitute it into the completion slot:

$$G_{j,k}(X)=F_{j,k}(X,X_{w_k}).$$

When comparing with the old tuple, $X_{w_k}$ is one unitary slot, so the error is bounded by $H_k{\epsilon }_k$ independently of the length of the as-yet-unknown word. Together with the old approximation this gives $\Vert G_{j,k}(A^{(k)})-y_j{\Vert }_2<2\cdot 2^{-k}$.

Only now compute a common Lipschitz bound $L_k$ for the substituted polynomials and choose

$$r_k\le \min \left\{r_{k-1}/2,\frac{2^{-k}}{1+L_k}\right\}.$$

A long witness word can force a tiny future budget, but cannot invalidate the tolerance already used to obtain it.

flowchart TB
    F["Choose target polynomials F"] --> E["Choose current tolerance"]
    E --> W["Obtain witness word w"]
    W --> G["Substitute w to form G"]
    G --> R["Choose future budget using lengths in G"]
    R --> N["All later perturbations preserve these targets"]

One summable tail preserves every finite target

The geometric budgets imply ${\sum }_{\ell >k}{\epsilon }_{\ell }\le r_k$. Therefore the first $n$ generators converge in operator norm to unitaries $A^{(\infty )}$, and their tail displacement after stage $k$ is at most $r_k$. Every fixed nonidentity reduced word has trace zero at every stage. Norm continuity of that fixed word and of the trace passes this identity to the limit, preserving the free Haar distribution.

For every $j\le k$, the same tail estimate gives

$$\Vert G_{j,k}(A^{(\infty )})-y_j{\Vert }_2\le L_k{r}_k+2\cdot 2^{-k}\le 3\cdot 2^{-k}.$$

Thus the limiting word algebra is $L^2$-dense. This is a separate property from preservation of joint distribution: a free tuple could otherwise generate only a proper subfactor.

Conditional expectation converts density into generation

Put $N=W^{\ast }(A^{(\infty )})$. The trace-preserving conditional expectation $E_N$ is the orthogonal projection onto $L^2(N)$. It fixes each approximating polynomial and is $L^2$-contractive, so

$$\Vert E_N(y_j)-y_j{\Vert }_2\le 2\Vert G_{j,k}(A^{(\infty )})-y_j{\Vert }_2\to 0.$$

It therefore fixes a dense set and is the identity. Faithfulness of the trace yields $N=M$. No uniform operator-norm bound on the approximating polynomials is required, and the sequence of completion generators need not converge. These are the precise reasons the generation section can combine a norm-convergent tuple with only $L^2$ target approximations. The flow and small-cocycle results remain the substantive inputs supplying each perturbation step.

See Also

• Lattice factors and stable recovery explains amplification and module dimensions in a rigidity setting; its manuscript class has trivial fundamental group, while the present family asserts the opposite amplification behavior for free group factors.
• Projections and traces distinguishes continuous from Borel spectral calculus. The cutoff here is used in a von Neumann algebra and does not require a reduced-algebra spectral gap.

🏷️ Manuscripts and verification scope

• An isomorphism of the free group factors — pinned manuscript and companion source.

The family scope page describes selected formalization coverage. The documented selection is normal trace-preserving isomorphism of all interpolated free group factors, including the infinite parameter. The fundamental-group conclusion is described as a further consequence rather than a separate selected statement; the entropy-dimension consequences are not listed in 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. An isomorphism of the free group factors.