Isomorphism of Free Group Factors

Isomorphism of Free Group Factors

OpenAI's claim

The manuscript claims that the factors L$F\_2$ and L$F\_3$ are isomorphic, consequently identifying all interpolated free group factors and giving each the full positive-real fundamental group. The cited manuscript argues for that result; its relation to this problem is specified below.

Mathematical statement of the claimed result

Mathematical claim 1.

Writing $\mathbb F_n$ for the free group of rank $n$ and $L(\mathbb F_n)$ for its group von Neumann factor with canonical trace, OpenAI claims \[ L(\mathbb F_2)\xrightarrow{\sim}L(\mathbb F_3) \] by a unital normal trace-preserving $*$-isomorphism. Its rank-step construction also asserts $L(\mathbb F_n)\simeq L(\mathbb F_{n+1})$ for every $n\ge3$. These are actual factor isomorphisms, not only embeddings or stable equivalences.

Outline of the argument

Manuscript 1 (family 287, 2026-09-23): Small trace-preserving automorphisms are built using finitely supported cocycles on free words. A Catalan-moment computation and truncated inverses provide coefficients that absorb an additional Haar generator with controlled error. Iterating with carefully allocated errors preserves generation as well as the free distribution; established amplification formulas transfer the consecutive-rank construction to ranks two and three.

Formal verification

For family 287, the scope page documents the following selections. The selected normal trace-preserving isomorphisms cover all interpolated free-factor parameters greater than one, including infinity. The fundamental-group consequence is separate. The documented statement covers the main claim of manuscript 1. Its selected declaration also covers the clause used for this problem. Comparator files: InterpolatedFactors.lean. Declarations: OAI.FreeGroupFactorMain.Interpolation.allInterpolatedIsomorphic.

This site has not executed the Lean files or checked their dependencies. These descriptions report the scope of the linked declarations, rather than a new verification of the proofs.

Status caveat

OpenAI attributes these results to an internal model. Its release notice cautions that results lacking full formalization can contain mistakes. This site has not independently assessed the proofs; a solved label records the scope of the reported claim.

Sources

• OpenAI, An isomorphism of the free group factors, 2026-09-23; citation key OAI:An-isomorphism-of-the-free-group-factors-September-23-2026. PDF (GitHub); PDF (direct); Manuscript page and citation; Latest version; Statement source: thm:main.
• Family 287: Lean scope.
• Comparator InterpolatedFactors.json.
• Comparator InterpolatedFactors.lean.
• Family 287: Reasoning summary.
• OpenAI catalogue entry.

Isomorphism of Free Group Factors

Definitions and mathematical statement

For a discrete group $G$, the left regular representation on $\ell^2(G)$ is $\lambda(g)\delta_h=\delta_{gh}$. Its group von Neumann algebra is $L(G)=\{\lambda(g):g\in G\}''$, the double commutant in bounded operators. Let $F_m$ be the free group on $m$ generators. Determine the isomorphism relation among $L(F_m)$ for distinct ranks in the nonabelian range, particularly \[ L(F_2)\stackrel{?}{\cong}L(F_3). \] Isomorphism here means a von Neumann algebra $*$-isomorphism, not an isomorphism between the underlying free groups. For these finite factors the normalized trace is intrinsic. The issue is whether the operator algebra remembers the number of generators.

Formulation and concept references: [S1].

Short English statement

Can free groups with different numbers of generators produce isomorphic von Neumann algebras, especially at ranks two and three?

Research attempt

Current frontier.

The random-matrix strong-convergence results discussed in [S1], Section 2.2, approximate prescribed free generators; they do not identify all possible generating families of their von Neumann algebra. Free entropy dimension distinguishes the standard free semicircular families, but the needed invariance under arbitrary von Neumann algebra generation is the obstacle to using it for the stated isomorphism problem [E2], Sections 2.4--2.5. Jung's packing formulation makes that obstacle quantitative [E1], Definition 2.1 and Corollary 2.4. The argument below investigates this route without identifying group rank with an already established factor invariant.

Attack route.

A compression argument would require an isomorphism invariant with a known nontrivial scaling law. A generator-count argument must be invariant under changing to an arbitrary generating tuple, not just a polynomial change of coordinates. We pursue the metric-entropy version of the second route: quantify how accurately and how gently one must reconstruct one microstate tuple from another. The missing lemma will be a uniform, subpower-distortion transport statement. A scalar polynomial example tests the proposed shortcut from approximation alone to that statement.

Attempt: a finite-matrix packing lemma.

For tuples of self-adjoint $k\times k$ matrices use \[ d_k(A,B)^2=\sum_i k^{-1}\operatorname{Tr}((A_i-B_i)^2). \] Let $\mathcal X_k,\mathcal Y_k$ be bounded sets of such tuples, possibly of different lengths, and let $\Phi_k:\mathcal X_k\to\mathcal Y_k$. Suppose there is a map $\Psi_k$ defined on $\Phi_k(\mathcal X_k)$ with \[ d_k(\Psi_k\Phi_k(A),A)\le\eta, \qquad d_k(\Psi_k(B),\Psi_k(B'))\le Ld_k(B,B'). \] Let $\operatorname{sep}_\varepsilon$ denote the largest cardinality of a set whose pairwise distances exceed $\varepsilon$. If $2\eta<\varepsilon$, then \begin{equation}\label{wsref-0-65-71-31-37-34-2d-70-61-63-6b-69-6e-67} \operatorname{sep}_\varepsilon(\mathcal X_k) \le\operatorname{sep}_{(\varepsilon-2\eta)/L}(\mathcal Y_k). \end{equation} Indeed, the triangle inequality bounds $d_k(A,A')$ by $2\eta+L d_k(\Phi_k(A),\Phi_k(A'))$. Thus images of an $\varepsilon$-separated set remain separated at the claimed scale. Injectivity on that set follows at once. This proof has no dependence on $k$ and requires no differentiability or volume-preservation assumption.

On a fixed operator-norm ball of radius $R\ge1$, a noncommutative polynomial $P=\sum_w c_w w$ is Lipschitz in this metric, with a bound obtained by summing $|c_w||w|R^{|w|-1}$, up to the fixed tuple-norm constant. To check this, telescope the difference of each monomial, one factor at a time, and use $\|ABC\|_2\le\|A\|\|B\|_2\|C\|$. Thus polynomial reconstruction is suitable for (eq174-packing), but its coefficient sizes and degrees matter.

Attempt: the precise transport hypothesis.

For a self-adjoint tuple $X$, let $\Gamma_R(X;m,k,\gamma)$ consist of operator-norm-$R$ matrix tuples whose normalized traces of words of length at most $m$ differ from those of $X$ by less than $\gamma$. Define the normalized packing quantity by taking $\limsup_{k\to\infty}$ of $k^{-2}\log\operatorname{sep}_\varepsilon$, then the infimum over $(m,\gamma)$. Denote it by $\mathcal P_{\varepsilon,R}(X)$. The packing description of $\delta_0$ takes a further small-scale limit; fixed changes in the separation convention do not change the dimension [E1], Corollary 2.4.

Here is a sufficient input for transporting that dimension from $X$ to $Y$. Fix cutoffs $R_X,R_Y$ larger than the operator norms of the tuples. For every small $\varepsilon$ suppose that, for every target moment condition $(m_Y,\gamma_Y)$, sufficiently stringent source moment conditions $(m_X,\gamma_X)$ admit maps as in (eq174-packing) between the corresponding microstate sets, for all sufficiently large $k$. Require \begin{equation}\label{wsref-0-65-71-31-37-34-2d-64-69-73-74-6f-72-74-69-6f-6e} \eta\le\varepsilon/4,\qquad L=L(\varepsilon)\ge1, \qquad \log L(\varepsilon)=o(\log(1/\varepsilon)). \end{equation} The constants and cutoffs are independent of the matrix dimension and the target moment accuracy. The maps themselves may depend on those moment conditions. Then (eq174-packing), followed by the indicated limits and infima, gives \begin{equation}\label{wsref-0-65-71-31-37-34-2d-65-6e-74-72-6f-70-79} \mathcal P_{\varepsilon,R_X}(X) \le \mathcal P_{\varepsilon/(2L(\varepsilon)),R_Y}(Y). \end{equation} The order of quantifiers is essential: a map into one fixed coarse target microstate set would not justify taking the target infimum.

Divide by $\log(1/\varepsilon)$. By (eq174-distortion), \[ \frac{\log(2L(\varepsilon)/\varepsilon)}{\log(1/\varepsilon)}\longrightarrow1. \] The usual fixed-cutoff packing formula therefore yields $\delta_0(X)\le\delta_0(Y)$. A reverse transport with the same hypotheses yields equality. This proves the conditional dimension-transfer statement; there is no exchange of the infimum over moment accuracy with a matrix limit that was not part of the definition.

In particular, applying this transport to every pair of generating tuples of an isomorphic pair of free group factors would make $\delta_0$ a factor invariant in the needed cases. Its value on a free semicircular $r$-tuple is $r$, so this would distinguish ranks two and three, and more generally distinct finite nonabelian ranks; see [E2] for the free-entropy calculation and its intended application. This is a sufficient deep input, not a property established merely from an isomorphism of factors.

Attempt: breaking the approximation-to-transport shortcut.

Approximation of actual operators does not automatically control its Lipschitz cost. Even on $[-1,1]$, put \[ P_n(x)=x+n^{-1}T_{n^2}(x), \] where $T_j(\cos\theta)=\cos(j\theta)$. Then $\|P_n-\mathrm{id}\|_\infty\le1/n$, while \[ P_n'(1)=1+n^{-1}(n^2)^2=1+n^3. \] The derivative identity follows by taking the limit of $j\sin(j\theta)/\sin\theta$ as $\theta\to0$. For an approximation scale $\varepsilon=1/n$, this Lipschitz cost is a positive power of $1/\varepsilon$, not the subpower cost in (eq174-distortion). If $L\asymp\varepsilon^{-c}$ in (eq174-entropy), the elementary dimension comparison can lose the factor $1+c$ instead of proving equality.

This example does not say that every approximation to the identity must be bad: the identity itself is an excellent choice. It disproves the general inference that an accurate sequence of polynomial approximants comes with adequate quantitative control. Von Neumann generation supplies strong or $L^2$ approximation, not the fixed-cutoff, uniformly reversible, subpower-distortion microstate transport asserted above. Controlling only one actual operator tuple also does not control every matrix tuple with sufficiently many matching moments unless the relevant estimates are transferred with their tolerances.

Stress test.

The normalized Hilbert--Schmidt metric, rather than unnormalized Euclidean matrix distance, is used throughout. A dimension-dependent Lipschitz constant could survive division by $k^2$ and invalidate the argument. Operator-norm cutoffs cannot be allowed to diverge arbitrarily with the small scale without checking the entropy definition again.

The entropy route studies microstates with moment tolerances. Strong convergence of a selected random-matrix sequence [S1] is valuable but is not a uniformly invertible parametrization of those entire spaces. Nor would a group-theoretic invariant of $F_r$ automatically be an invariant of $L(F_r)$: isomorphisms need not carry canonical group unitaries to group unitaries. The argument makes no assertion about countably infinite rank from the finite-tuple calculation alone.

Outcome.

Outcome: Conditional reduction.

The attempt proves a dimension-free packing inequality and a quantified sufficient condition for invariance of free entropy dimension under change of generators. A concrete polynomial test shows why approximation alone cannot supply the condition. The unproved input is the uniform transport statement for arbitrary generating tuples; obtaining it in the free-factor cases would resolve the finite-rank distinction. No factor isomorphism or nonisomorphism is established here, and no novelty is claimed for the standard packing interpretation itself.

Sources

• S1 Ramon van Handel, Strong Convergence: A Short Survey $2025$, section 2.2. https://arxiv.org/html/2510.12520v1. Catalogue formulation source.
• E1 Kenley Jung, A free entropy dimension lemma, Pacific Journal of Mathematics 211$2$ $2003$, 265--271; arXiv:math/0207149v2, Definition 2.1, Remark 2.3, and Corollary 2.4. https://arxiv.org/html/math/0207149. Consulted 27 September 2026; normalized microstate packing and the order of its limits.
• E2 Dan Voiculescu, Free Entropy, arXiv:math/0103168, 2001, Sections 2.4--2.5 and the generator-invariance problem. https://arxiv.org/html/math/0103168. Consulted 27 September 2026; free-family dimension and the distinction between a tuple invariant and a von Neumann algebra invariant.