Kadison's similarity conjecture

The explanation

Start with a tiny algebra. Its elements are pairs of numbers (a, b), added and multiplied slot by slot. It also has a mirror operation, called the star. The star replaces each number by its complex conjugate. For ordinary real numbers, the star changes nothing.

Now send each pair to a 2-by-2 matrix, a square table of numbers. Send (a, b) to the matrix with rows (a, a − b) and (0, b). So (2, 3) goes to rows (2, −1) and (0, 3). Products survive the trip. Multiply two pairs and convert, or convert and multiply the matrices: the answers agree. A map that keeps sums and products like this is called a homomorphism.

The mirror does not survive. The pair (1, 0) is its own mirror image. It goes to the matrix with rows (1, 1) and (0, 0). For matrices, the mirror flips the table across its diagonal. That gives rows (1, 0) and (1, 0), a different matrix. Picture the first matrix as a shadow cast on a line. It casts the shadow at a slant, not straight down. A slanted shadow can stretch arrows: here, by a factor of √2, about 1.414. A straight-down shadow never stretches anything.

A change of coordinates repairs this. Take S, the matrix with rows (1, 1) and (0, −1). It undoes itself, so S times S is the identity. Compute S, then the image of (a, b), then S again. The result has a and b on the diagonal and zeros elsewhere. Now (1, 0) casts a straight-down shadow, which equals its own mirror image. In the new coordinates the map respects the star. Two maps linked this way are called similar. Here S stretches lengths by at most about 1.618.

Kadison asked in 1955 whether such a repair always exists. The setting is a C*-algebra: a system of operators that can be added, multiplied and mirrored. An operator is a rule that moves the arrows of a Hilbert space, a space with lengths and angles. A Hilbert space may have infinitely many directions. The map must be bounded, meaning it never multiplies sizes by more than a fixed factor. The question: is every such map similar to one that respects the star? This paper proves the answer is yes, with no extra conditionsSource 1,Source 2.

Many cases were already settled. Bunce and Christensen handled nuclear algebras, a broad well-behaved class, in 1981. In 1983, Haagerup solved the cyclic case. There, one arrow, moved by the algebra, comes as close as you like to every arrow. He showed that complete boundedness is the test for similarity. Complete boundedness means the size bound still holds when the map acts on square blocks of elements, of every size. Haagerup also covered algebras without a tracial state. A tracial state is an averaging rule, like a matrix's diagonal sum divided by its size. Christensen covered type II1 factors with property Γ in 1986. A type II1 factor is an indivisible, infinite-dimensional operator algebra that has such an averaging rule. Property Γ means it contains elements, not plain numbers, that nearly commute with every element. Kirchberg proved in 1996 that the problem has an equivalent form. It asks about derivations, which measure how far an operator Y is from commuting with the algebra.

The one obstacle was a size limit that does not grow with block size. Suppose Y nearly commutes with every single element. Must it nearly commute with every block of elements, by the same margin? The paper's main estimate says yes, up to one fixed constant. That constant does not depend on the algebra, the space or the block sizeSource 3. The paper calls the hard part a constant independent of matrix size for finite factorsSource 4. A finite factor is an indivisible operator algebra with a trace, an averaging rule like a matrix's diagonal sum.

Back to the pairs. The theorem says a straightening S exists for every algebra, every space and every bounded map. It gives no limit on how much S may distort lengthsSource 5. The choice of S may depend on the whole mapSource 6. The same estimate yields one hyperreflexivity constant for all von Neumann algebras, a large class of operator algebrasSource 7. Hyperreflexive means a simple test on invariant subspaces measures distance to the algebra, up to a fixed factor. The paper does not claim its constant is the best possibleSource 8.

How sure is this? The Lean scope note says a formalization proves the similarity theorem for arbitrary Hilbert spaces. It also covers the universal hyperreflexivity constant. It links a formal statement of the uniform commutator estimate too. No specialist review of the paper has been published.

What it doesn't show. This does not bound how much the repairing change of coordinates may distort lengths.