Can an operator algebra forget how many free generators it had?

← All digests

Family 287 · Operator algebras
Importance 5/5 · Provisional editorial rating
Sources consulted 2026-10-25

1. The question

The free group on two generators and the free group on three generators are different groups. Their elements are words, with cancellation of a letter against its inverse as the only rule. The number of independent generators is an essential algebraic distinction.

Now replace each group element by its translation operator on square-summable functions, and take the weak operator closure. Does this much larger algebra still remember whether the original group had two or three generators?

Family 287 claims that it does not:

$$L(\mathbb F_2)\cong L(\mathbb F_3).$$

This would resolve the free group factor isomorphism problem, a foundational question in operator algebras. Through earlier classification results, the manuscript derives that all interpolated free group factors, including the countably infinite-rank one, are isomorphic. The proposed mechanism gradually absorbs an extra free generator while preserving the statistical behavior and eventual generating power of the remaining generators.

2. Background and essential definitions

For a discrete group $G$, left translation acts on $\ell^2(G)$ by

$$(\lambda(g)\xi)(h)=\xi(g^{-1}h).$$

The group von Neumann algebra $L(G)$ is the weak operator closure of finite linear combinations of these translations. Weak closure permits limits tested through inner products of vectors. The canonical trace is $\tau(x)=\langle x\delta_e,\delta_e\rangle$, with $\tau(1)=1$.

For the nonabelian free groups $\mathbb F_n$, $n\geq2$, this algebra is a type $\mathrm{II}_1$ factor: its center is just the scalars, it has a finite normalized trace, and it has no minimal nonzero projections. The trace gives an additional norm

$$\|x\|_2=\tau(x^*x)^{1/2},$$

which measures average quadratic size. It differs from the operator norm, which measures the largest effect on a unit vector.

A unitary $U$ is Haar when $\tau(U^k)=0$ for every nonzero integer $k$, analogous to a uniformly distributed point on the unit circle. A tuple $U_1,\ldots,U_n$ is a free Haar tuple when every nontrivial reduced word in the unitaries and their adjoints has trace zero. It is freely generating if its weakly closed algebra is the whole factor.

That final requirement matters. A free Haar tuple can sit inside a much larger algebra without generating it. Matching all its moments identifies the algebra it generates, not automatically the surrounding algebra.

An amplification $M^t$, $t>0$, is a normalized corner of a matrix algebra over $M$, cut out by a projection with unnormalized trace $t$. The fundamental group $\mathcal F(M)$ consists of scales $t$ for which $M^t\cong M$. Here “fundamental group” names an invariant of a factor, not the loop group of a topological space.

3. Development history and previous partial results

Free group factors appeared in Murray and von Neumann’s foundational work. Their nonhyperfiniteness separated them from the hyperfinite factor, which is built by approximating with finite-dimensional algebras. But this distinction did not distinguish one free rank from another. Kadison is credited in the historical literature with raising the rank-isomorphism question.

Voiculescu’s free probability supplied models in terms of freely independent noncommuting random variables. Semicircular variables, the free-probability counterparts of Gaussian variables, generate free group factors. Compression calculations related the number of such variables to the size of a corner.

Dykema and Rădulescu independently introduced interpolated factors $L(\mathbb F_r)$ for real $r>1$. The noninteger symbol is a parameter for a factor; it does not mean an ordinary free group with a fractional number of generators. Their amplification formula is

$$L(\mathbb F_r)^t\cong L(\mathbb F_{1+(r-1)/t^2}).$$

The classical alternative says that these factors are either all isomorphic or pairwise nonisomorphic. Rădulescu’s version includes the infinite parameter. Thus a single isomorphism between distinct finite ranks would settle the alternative.

Free entropy dimension offered a possible route toward distinguishing ranks. A semicircular $n$-tuple has dimension $n$. If that dimension depended only on the generated factor, different ranks would be different. However, this required a generator-invariance statement, not merely computing the dimension for familiar tuples.

Other methods constructed isomorphisms in perturbative settings. Guionnet–Shlyakhtenko’s free monotone transport changes noncommuting variables while retaining the generated algebra under appropriate analytic assumptions. M1 also discusses a recent Shlyakhtenko construction involving Fuchsian group factors, but states that it uses no result from that preprint. The present claim specifically requires changing the number of free generators.

4. What the new manuscripts claim

M1’s direct construction gives, for every integer $n\geq3$, a unital normal trace-preserving *-isomorphism

$$L(\mathbb F_n)\cong L(\mathbb F_{n+1}).$$

A *-isomorphism preserves multiplication, complex linear combinations, adjoints and the identity. Normality gives the appropriate continuity for von Neumann algebras. The claim is an actual isomorphism, not only agreement of selected invariants or an approximation by maps.

The construction produces a freely generating Haar $n$-tuple inside $L(\mathbb F_{n+1})$. To reach ranks two and three, M1 first combines the direct comparisons at ranks three and four, obtaining $L(\mathbb F_3)\cong L(\mathbb F_5)$. Amplifying both sides by $\sqrt2$ and applying the classical formula yields ranks two and three.

The earlier all-or-nothing alternative then implies

$$L(\mathbb F_r)\cong L(\mathbb F_s),\qquad 1<r,s\leq\infty,$$

and

$$\mathcal F(L(\mathbb F_r))=\mathbb R_{>0}.$$

M1 also states that four free entropy dimensions, $\delta,\delta_0,\delta^*,\delta^\star$, fail to depend only on the generated von Neumann algebra. Each takes every integer value $n\geq2$ on some finite self-adjoint generating tuple of $L(\mathbb F_2)$. The tuples used for the microstates and nonmicrostates versions need not be the same. This does not claim that all four dimensions disagree on every tuple.

5. Why the advance matters

Importance: 5/5. If correct, this settles one of the central classification problems for finite factors and determines the isomorphism type of an entire interpolated family. It would establish a striking loss of information: the completed operator algebra forgets a rank distinction that the original free group retains.

It would also change how free entropy dimensions can be used. A quantity may remain highly informative about a chosen family of noncommuting variables while failing to be an invariant of the algebra they generate. The claimed result separates those two roles decisively.

The conclusion does not make all $\mathrm{II}_1$ factors isomorphic. Free group factors would still differ from the hyperfinite factor and from many other factors. What collapses is this particular rank parameter, not the entire classification theory.

6. The main idea

Move generators without changing their law

Begin with a free Haar tuple $(A_1,\ldots,A_n,C)$ generating $L(\mathbb F_{n+1})$. The goal is to hide the information carried by $C$ inside small changes to the first $n$ variables.

The manuscript takes a bounded logarithm $S=\arg(C)$, so $e^{iS}=C$, and uses its conjugates by words in the $A_j$. These feed a polynomial differential equation. M1’s moment identity and analyticity argument show that the resulting flow preserves traces of words and extends to automorphisms of the factor.

This is more specific than a formal change of symbols. The maps must respect the actual operator relations and norms before they can serve as automorphisms.

Make a long word carry the missing variable

The flow is controlled by finitely supported coefficient vectors on a free group. Their values on generators are small, while a suitable long word accumulates a coefficient close to a unit point mass. A cocycle identity organizes that accumulation, much as repeatedly applying a product rule organizes a long algebraic expression.

M1 obtains the coefficients using spectral estimates for sums of free translations and a truncated inverse. Small generator coefficients keep each $A_j$ almost fixed in operator norm, yet the selected word can move enough to encode the extra variable. After a change of free basis, a word in the moved first $n$ coordinates approximates the old $C$.

Take a limit and preserve generation

The absorption step is repeated with decreasing errors. The first $n$ coordinates converge in norm; their joint free Haar moments therefore persist in the limit. The complementary coordinate is allowed to change at each stage.

Generation requires a separate argument. Once an approximating word has been chosen, its length controls how sensitive it is to future changes in the generators. M1 chooses later error budgets only after knowing that word. This keeps earlier approximations useful throughout the iteration.

The resulting word algebra is claimed to be dense in the ambient tracial $L^2$ space. A trace-preserving conditional expectation, which acts as orthogonal projection onto a von Neumann subalgebra’s $L^2$ space, then identifies the limiting generated algebra with the whole factor. The construction thus needs both preservation of distribution and recovery of all ambient operators.

7. Scope, limitations, and open questions

The direct absorption theorem begins at $n=3$. The ranks-two-and-three statement uses the established amplification calculus afterward. The extension to every real parameter and infinite rank uses the classical dichotomy, rather than a separate direct construction at each parameter.

The theorem concerns tracial von Neumann algebras. It does not identify the original free groups, their algebraic group rings, or their reduced group C*-algebras. Taking weak closure is essential to the setting.

The argument is constructive in its mathematical presentation but does not provide a practical finite formula evaluating the isomorphism on arbitrary operators. It relies on limiting procedures and generation in a topology weaker than norm closure of the word algebra.

Finally, failure of generator invariance for free entropy dimensions leaves many questions about their behavior on particular tuples. Their definitions and local analytic uses remain meaningful. M1’s conclusion is that the generated factor alone cannot determine these four numerical values.

8. Sources and further reading

• Dykema, Interpolated free group factors, for interpolation and amplification theory.
• Guionnet and Shlyakhtenko, Free Monotone Transport, for an earlier analytic method of constructing isomorphisms through changes of variables.
• M1’s Sections 6–7 specify the classical dichotomy used, its inclusion of infinite rank, and the separate consequences for the four entropy dimensions.

Source manuscripts

• M1. OpenAI, An isomorphism of the free group factors (September 23, 2026).

Sources consulted on 25 October 2026 at repository revision 1581d4c. Download the manuscript citations.