I consider on $M_n(\mathbb C)$ the normalized $2$-norm, i.e. the norm given by $\|A\|_2 = \sqrt{\mathrm{Tr}(A^* A)/n}$.

My question is whether a $k$-uple of hermitian matrices that are almost commuting (with respect to the $2$-norm) is close to a $k$-uple of commuting matrices (again with respect to the $2$-norm). More precisely, for an integer $k$, is the following statement true?

For every $\varepsilon>0$, there exists $\delta>0$ such that for any $n$ and any matrices $A_1,\dots, A_k\in M_n(\mathbb C)$ satisfying $0\leq A_i\leq 1$ and $\|A_iA_j - A_j A_i\|_2 \leq \delta$, there are commuting matrices $\tilde A_1,\dots,\tilde A_k$ satisfying $0\leq \tilde A_i\leq 1$ and such that $\|A_i - \tilde A_i\|_2 \leq \varepsilon$.

The important point is that $\delta$ does not depend on $n$.

I could not find a reference to this problem in the litterature. However, this question with the $2$-norm replaced by the operator norm is well-studied. And the answer is known to be true if $k=2$ (a result due to Lin) and false for $k=3$, and hence $k\geq 3$ (a result of Voiculescu).


4 Answers 4


There is a recent paper by Glebsky titled "Almost commuting matrices with respect to normalized Hilbert-Schmidt norm" which shows that this is indeed true for any $k$ for Hermitian matrices (and in fact also unitary and normal matrices).

  • 2
    $\begingroup$ Thanks a lot for your answer. I did not yet have time to read the whole proof, but this seems to completely answer my question. I just wonder why this paper was posted in the algebraic geometry section of arXiv. $\endgroup$ Jun 21, 2011 at 14:31

The answer is yes, and much more is true. Any hyperfinite von Neumann algebra (with separable predual) has a unique embedding (up to conjugacy) into the ultra-product of the hyperfinite $II_1$-factor.

This implies in particular, that almost commuting matrices in Hilbert-Schmidt are close to commuting matrices. The proofs goes by contradiction; assume that there is a sequence of counterexamples and construct non-conjugate embeddings. Since any abelian von Neumann algebra is hyperfinite, this yields a contradiction.

Kenley Jung showed that uniqueness of the embedding also implies that the algebra is hyperfinite.

  • $\begingroup$ Andreas, is your argument presupposing that the matrices almost commute in HS-norm and are uniformly bounded in operator norm? $\endgroup$
    – Yemon Choi
    Jun 21, 2011 at 21:16
  • $\begingroup$ Thanks Andreas. I already accepted Ashley's answer, but your answer is more what I was hoping for (Glebsky's proof has however the advantage of giving an explicit bound on $\delta$ depending on $k$ and $\varepsilon$). $\endgroup$ Jun 22, 2011 at 6:42
  • $\begingroup$ I guess that the precise statement is :"Given a non principal ultrafilter $\mathcal U$, any two embeddiing of a hyperfinite von Neumann algebra into $\prod_{\mathcal U} M_n$ (or $R^{\mathcal U}$) are conjugate". (unless I missed something, the ultrapower might depend on $\mathcal U$). By the way, do you have a reference for this statement? $\endgroup$ Jun 22, 2011 at 6:57
  • $\begingroup$ Yemon: yes, just as in the question. $\endgroup$ Jun 22, 2011 at 7:32
  • 2
    $\begingroup$ Mikael: you are right, this might depend on the ultrafilter. I would try looking at the paper by Kenley Jung (Math. Ann. 2007 vol. 338 (1) pp. 241-248) and see whether he gives a reference. This is the easy part anyway, it is based on the fact that a full matrix-algebras embeds (unitally) in a unique way (up to conjugacy) into each other. The same holds if you take block-sums of matrix-algebras and remember their relative multiplicities. $\endgroup$ Jun 22, 2011 at 7:36

I just found the discussion. In this paper by Filonov and Kachkovskiy there are better estimates than mines and it contains citations of proofs using von Neumann algebras.

(It was a surprise for me too why my paper is in Algebraic Geometry. Probably it is my error. I have not found an easy way to fix it.)


As part of my dissertation, "Almost Commuting Operators on von Neumann Algebras," I have extended Glebsky's result to the normalized Schatten class for $1 \leq p < \infty$. Moreover, for $p=2$ we recover Filonov and Kachkovskiy's theorem with the same estimate. In our work, however, we use different techniques as the normalized Schatten p-norm does not arise from an inner product for $p \neq 2$.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.