Given an integral operator $K$ on $L^2(\mathbb R)$ with kernel $k(x, y)$, consider the integral operator $L$ on $L^2(\mathbb R)$, whose kernel has the form $k(\alpha x+\beta y, \gamma x+\delta y)$, where $\alpha\delta-\beta\gamma\ne 0$. It is well known that an integral operator belongs to the Hilbert-Schmidt class $\mathfrak S_2$ if and only if its kernel is square-summable; therefore, $K\in\mathfrak S_2$ if and only if $L\in\mathfrak S_2$. Is the same true if we replace the class $\mathfrak S_2$ by $\mathfrak S_1$, or by $\mathfrak S_p$ with $p\ge 1$? Can such an equivalence be proved for other properties? (References to known results would be appreciated.)

This question was not answered at SE. I was not able to use hints suggested in the discussion there.

In particular, this contains another question as a special case: does the integral operator on $L^2(\mathbb R)$, whose kernel is the indicator of the rhombus $\{|x|+|y|<1\}$, belong to the trace class?

  • 1
    $\begingroup$ My impression is that rotations are not relevant to the theory. Most of the interesting kernels are singular on the diagonal $x=y$, but this locus is not invariant under rotations. $\endgroup$ Oct 9, 2014 at 11:37
  • 2
    $\begingroup$ Integral operators discussed here belong to the Hilbert-schmidt class and thus their kernels are regular. Moreover, there are various serious questons about operators between two different $L^2$-spaces, and then there is no diagonal at all. My question stems from a concrete setting, and I do not pretend that it cover all interesting kernels. $\endgroup$
    – limanac
    Oct 9, 2014 at 13:00
  • $\begingroup$ I voted to close but this was based on an error in my thinking and on 2nd thought I don't see how the comments at MSE can be used to answer the question $\endgroup$
    – Yemon Choi
    Oct 9, 2014 at 17:45

1 Answer 1


Consider the characteristic function of the rhombus $\{|x|+|y| <1\}$ as a kernel on $L^2(I)$, with $I:=[-1,1]$. Since this kernel is symmetric, the trace norm of the corresponding compact operator $T$ is just the $\ell^1$ norm of the sequence of its eigenvalues. Let's find them.

We have for any $x\in I$ $$Tu(x)=\int^{1-|x|}_{-1+|x|}u(y)dy.$$

Let $v(x):=\int_0^xu(y)dy$ and $\lambda\neq0$. The eigenvalue equation $Tu=\lambda u$ then writes $$v(1-x)-v(-1+x)=\lambda v'(x)\quad \mathrm{for\quad } x >0\; , $$ $$v(1+x)-v(-1-x)=\lambda v'(x)\quad \mathrm{for\quad } x <0\; , $$ with $v(0)=0$. Equivalently, $v(x)$ is an odd function satisfying $$2v(1-x)=\lambda v'(x).$$ If we derive both sides and use the equation again we find $$v''(x)=-(2/\lambda)^2v(x),$$ with boundary conditions $v(0)=0=v'(1).$ Hence the eigenfunctions are cosine functions, with eigenvalues $\lambda_n\sim C/n$, and $T$ is not in the trace class.

  • $\begingroup$ Double check: from $\lambda_n=4/(2n+1)\pi$ one finds for the HS norm of $T$ $\sum_{n\ge1}\lambda_n^2=2$, as it has to be. $\endgroup$ Oct 9, 2014 at 22:32
  • $\begingroup$ Does it help to point out (the obvious fact) that $Tu=0$ if $u$ is odd; this confused me for a moment. $\endgroup$ Oct 9, 2014 at 23:27
  • $\begingroup$ Also note that $T$ is in $\mathfrak{S}_p$ for all $p>1$, so the question remains open for $2\neq p>1$. $\endgroup$ Oct 10, 2014 at 10:41
  • $\begingroup$ Yes, thanks. The case $p=1$ was the most important one for me. $\endgroup$
    – limanac
    Oct 10, 2014 at 10:49

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.