Suppose that $T \in TC(l^2( \mathbb{Z}))$ is trace class. Consider its kernel $ T(i,j) = \langle e_i, T e_j \rangle $ where $ \{e_i\}_{i \in \mathbb{Z}}$ is an ONB for $l^2( \mathbb{Z})$. Now, consider the operator given by the kernel $T(i,j) K(i,j) $ for some numbers $K(i,j)$ such that $\sup_{i,j} \vert K(i,j)\vert < \infty$.

Is this operator still trace class?

My thoughts: The new operator is the Hadamard/Schur product of $K \circ T$ for some operator $K$ which we do not know is bounded. If $K$ is bounded in operator norm then $\vert \vert K \circ T \vert \vert_1 \leq \vert \vert K \vert \vert_\infty \vert \vert T \vert \vert_1 $. But our condition on $K$ is not enough to ensure that.

We can split $K$ and $T$ up into real and imaginary parts and then their real and imaginary parts into positive and negative parts. By a triangle inequality we can estimate each of the 16 terms $K_i \circ T_j$ where $K_i$ and $T_j$ are positive. By the Schur product theorem then $K_i \circ T_j$ is also positive. Hence we can compute its trace norm by \begin{align*} \vert \vert K_i \circ T_j \vert \vert_1 &= Tr( K_i \circ T_j ) = \sum_n K_i(n,n) T_j(n,n)\\ &\leq \sup_{n} K_i(n,n) \sum_n T_j(n,n) \leq \sup_{n} K_i(n,n) \vert \vert T_j \vert \vert_1 \leq \sup_{n} K_i(n,n) \vert \vert T \vert \vert_1 \end{align*}  where $K_i(n,n), T_j(n,n) \geq 0$ since $K_i$ and $T_j$ are positive operators. We can bound the matrix elements of the real and imaginary parts of $K$, but unfortunately we can't bound the the matrix elements of the positive and negative parts. I suspect one can use this insight to construct a counterexample.


2 Answers 2


It's a little more complicated than I thought! Frederik Ravn Klausen pointed out an error. Still, I maintain that the product needn't even be bounded.

As the answer to this question shows, in $M_n$ you can find a unitary $U$ and an matrix $A$ whose entries all have modulus 1, whose Hadamard product $A\bullet U$ has operator norm $\sqrt{n}$. Using the duality between operator norm and trace class norm, find a trace class matrix $B$ with trace norm 1 such that ${\rm tr}(B^*(A\bullet U)) = \sqrt{n}$.

We have an identity ${\rm tr}(B^*(A\bullet U)) = {\rm tr}((A\bullet B)^*U)$. Therefore the trace norm of $A\bullet B$ is at least (in fact, exactly) $\sqrt{n}$. Now $\bigoplus 2^{-n} B_{5^n}$ is trace class, and its Hadamard product with $\bigoplus A_{5^n}$ is unbounded.

  • $\begingroup$ Great example! Do you need $8^n$ instead of $4^n$ to make sure that the operator norm is infinite? $\endgroup$ Oct 21, 2021 at 12:18
  • $\begingroup$ Oh right, I'll correct it. $\endgroup$
    – Nik Weaver
    Oct 21, 2021 at 14:20
  • $\begingroup$ Wait, the trace norm of a unitary $U$ is $ Tr( \sqrt{ U^* U}) = Tr( 1) = n$. So the operator $\oplus 2^{-n} U_{8^n} $ is not trace class? $\endgroup$ Oct 22, 2021 at 10:21
  • $\begingroup$ Right again! But still fixable ... $\endgroup$
    – Nik Weaver
    Oct 22, 2021 at 11:43
  • 1
    $\begingroup$ Thanks, if you just want to get that the trace norm is not bounded one can look at $A_n = \sqrt{n} U_n$ where $U_n$ is the unitary and then let $B_n$ be the matrix of all $\frac{1}{n}$. Then $\oplus 2^{-n} B_n$ is trace class with trace norm 1 and $\oplus \sqrt{n} U_n \circ B_n = \oplus \frac{1}{\sqrt{n}} U_n $ which has trace norm $\sum_n \sqrt{n} = \infty$. $\endgroup$ Oct 22, 2021 at 12:11

Nik Weaver's answer gives a nice counter-example. Let me just say a few words of context. Kernels $K$ for which $KT$ is trace-class for all trace-class $T$ are called Schur multipliers. (Not to be confused with an unrelated, and more common, term from group theory). I believe this is because of Schur's 1911 paper (Crelle's journl, in German).

A more modern approach recognises links with completely bounded maps, the Haagerup tensor product, and so forth. One comment is that as the bounded linear operators and the trace-class operators are in duality, we can instead consider kernels which multiply bounded operators to bounded operators, and this is the more common framing of the question. References I know are Pisier's work (JSTOR) and Spronk's paper (PLMS). In particular, we have the following characterisation of such kernels $K$: there must be some Hilbert space $H$ and vectors $\xi_i, \eta_j$ in $H$ with $\sup_i\|\xi_i\|<\infty, \sup_j\|\eta_j\|<\infty$ and $K(i,j) = (\xi_i|\eta_j)$ for all $i,j$. This open up the idea of using the theory of tensor norms on Banach spaces to study such maps, which is (roughly speaking) what Pisier does in the cited paper.

I am not aware of a canonical reference; perhaps others are?

  • 1
    $\begingroup$ The reference I like to give is Pisier's book Similarity problems and completely bounded maps. $\endgroup$ Oct 21, 2021 at 12:04
  • $\begingroup$ Thanks for providing this context, so if I have a concrete kernel and I want to check whether it is a Schur multiplier can the theory help me? $\endgroup$ Oct 23, 2021 at 18:03
  • $\begingroup$ Well, you need to prove if the kernel can be written in the way stated. I guess I would have my doubts about whether this is always a practical test... $\endgroup$ Oct 23, 2021 at 19:49

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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