The generator problem for finite factors

The explanation

Take the 2-by-2 matrix x with rows (0, 1) and (0, 0). On its own it reaches very little. Its square is the zero matrix. So sums and products of x give only matrices of the form a·I + b·x. That is a family described by two numbers, while a 2-by-2 matrix has four entries.

Now also allow the adjoint x* (flip the matrix across its diagonal and conjugate any complex entries). Here x* has rows (0, 0) and (1, 0). The product x·x* has rows (1, 0) and (0, 0). The product x*·x has rows (0, 0) and (0, 1). With x and x*, these give the four basic building blocks. Every 2-by-2 matrix is a mix of them. The matrix with rows (3, 7) and (5, 2) is 3·xx* + 7·x + 5·x* + 2·x*x. So one matrix, with its adjoint, generates all 2-by-2 matrices.

One generator is the same as two self-adjoint ones. A self-adjoint matrix equals its own adjoint. Put a = (x + x*)/2 and b = (x − x*)/(2i). Then a has rows (0, 0.5) and (0.5, 0). And b has rows (0, −0.5i) and (0.5i, 0). Both are self-adjoint, and a + i·b gives back x.

The generator problem asks the same question in infinite dimensions. The setting is a von Neumann algebra. That is a family of operators closed under sums, products, adjoints and limits. Operators act like infinite matrices. Generated by x means: start from x and x*, take sums and products, then take limits. A separable predual is a size condition. Roughly, countably many pieces of data pin the algebra down. The question is whether every such algebra is generated by one operator.

Many cases were settled long ago. Pearcy, Suzuki and Saitô, and Wogen handled three large classes of algebras in 1962, 1963 and 1969. In 1974 Willig reduced the whole question to type II₁ factors. A factor is an algebra that cannot be split into independent pieces. A type II₁ factor is infinite-dimensional but carries a trace. The trace gives each piece a size from 0 to 1. In the 2-by-2 example, the normalized trace gives x·x* size 1/2.

Type II₁ factors then resisted a general answer. Progress used extra structure each time. Behncke handled tensor products of two such factors in 1972. A tensor product joins two factors into one, with each acting independently of the other. Popa handled factors with Cartan subalgebras in 1985. These are large commutative pieces that sit inside the factor in a special, maximal way. Ge and Popa handled factors with property Γ in 1998. Such a factor has nontrivial elements that almost commute with everything in it. Shen's generator invariant, from 2009, unified these and other classes. It is a number that measures, roughly, how many generators a factor needs. In 2012 Sherman showed that finitely many generators for every such factor would be enough. No method covered every factor.

This paper proves that every type II₁ factor with separable predual is generated by one operatorSource 1. With Willig's reduction, every von Neumann algebra with separable predual followsSource 2. The paper first proves a stronger result. Take a smaller factor P inside M, so that only plain numbers in M commute with all of PSource 3. Then almost every unitary u, together with P, generates MSource 4. A unitary is an operator that preserves lengths, like a rotation. Here almost every means a dense G_δ set: a generic set, found arbitrarily close to any unitary.

The difficulty sits in long products. The proof nudges u a little at each of its many appearances in a long word. Small nudges usually pile up as the word growsSource 5. The paper finds a bound on their total effect that does not grow with the lengthSource 6. Then it picks P generated by two self-adjoint elements, giving four for MSource 7. The 2-by-2 matrix trick from the start, with operators as entries, cuts four down to twoSource 8.

The result has limits. It concerns one factor at a time. It says nothing about generators that vary smoothly across a family of factors. Hayes built a continuous family where no finite set of continuous choices generates every memberSource 9. The paper also claims a consequence for Voiculescu's free entropy dimensionsSource 10. These numbers measure, roughly, how much room there is to approximate a list of generators by matrices. The paper says they can change when one generating list replaces another for the same algebra. How sure is this? A computer proof checker called Lean has checked formal versions of two statements. One is single generation of type II₁ factors with separable predual. The other is the dense G_δ relative generation theorem. The step to all von Neumann algebras rests on Willig's older reduction. That step and the free entropy consequence are outside the Lean scope note. No specialist review has been published yet.

What it doesn't show. This does not give generators that vary continuously across a family of factors; Hayes showed that can fail.