There might be an obvious answer to the question, but it doesn't come to mind.

Suppose we have an infinite matrix $A=(a_{ij})$, which defines a bounded linear operator on $\ell^p$, i.e. for all sequences $(x_i)\in \ell^p, p>1$ $$ \sum_{i=1}^\infty\big|\sum_{j=1}^\infty a_{ij}x_j\Big|^p \leq C \Vert x \Vert_{p}^p.$$ For some positive constant $C$. Furthermore, assume that $a_{ij}=\overline{a_{ji}}$. Is it true that the spectrum of $A$ on $\ell^p $ is real ?

  • 3
    $\begingroup$ Do you know whether the point spectrum is always real? $\endgroup$ Commented Apr 1, 2020 at 22:49
  • 3
    $\begingroup$ Yes, Jochen, because the hypothesis implies that $A$ also is bounded on $\ell_q$, $1/p+1/q=1$, hence all eigenvectors are in $\ell_2$. Basically the same argument gives a positive answer to the OP's question (I think). $\endgroup$ Commented Apr 1, 2020 at 23:36
  • 3
    $\begingroup$ @BillJohnson Your argument doesn't seem to use the fact $p>1$; but there are known examples of "self-adjoint" (i.e. conjugate symmetric) convolution operators on $l^1$(free group) whose spectrum has non-empty interior $\endgroup$
    – Yemon Choi
    Commented Apr 2, 2020 at 3:38
  • 8
    $\begingroup$ There are self-adjoint operators in $L^2$, generating (analytic) semigroups in $L^p$ for all $p$, such that the spectrum is not real for $p \neq 2$. One can take $L=r^2D_{rr}+2rD_r$ in the half-line. The spectrum is a parabola which degenerates into the negative half-line when $p=2$. The resolvents and the semigroups have similar bad properties, by the spectral mapping theorem. Maybe a discrete version can be contructed using them. $\endgroup$ Commented Apr 2, 2020 at 7:57
  • 4
    $\begingroup$ I guess the only correct part of my comment above is that the point spectrum is real when $p<2$ (because the operator is bounded on $\ell_2$ and $\ell_p \subset \ell_2$). $\endgroup$ Commented Apr 2, 2020 at 13:36

1 Answer 1


It seems that I have found a counter example myself.

For the Hilbert matrix $$ H_\lambda:= \big( \frac{1}{1-\lambda+k+n} \big)_{k,n\geq 0}, \lambda < 1 $$ Rosenblum in "On the Hilbert Matrix I, Proceedings of the AMS" proves that the pointspectrum considered as an operator on $\ell^p, p>2$ contains the set $$ \{ \pi \sec(\pi u ) : | \Re ( u )| < 1/2-1/p \}. $$

If one could provide a more elementary counterexample I would be interested in looking at.

  • $\begingroup$ However, I think it is true for $1 < p \le 2$ because, for $q \ge p$, if we denote $\sigma^q(A)$ the spectrum of the unbounded operator $A$ on $\ell^q$ with domain $\{x \in \ell^q \mid Ax \in \ell^q\}$ then we can show that the spectrum is increasing with $q$ meaning that $\sigma^q(A) \subset \sigma^r(A)$ for $p \le q \le r$. If I am not mistaken, for $1 < p \le 2$, the operator $A$ induce a self-adjoint operator on $\ell^2$ so $\sigma^p(A) \subset \sigma^2(A) \subset \mathbb{R}$. $\endgroup$ Commented Apr 6, 2020 at 15:53
  • $\begingroup$ If you can provide the missing details please go ahead and post it as an answer, I will accept it. $\endgroup$ Commented Apr 6, 2020 at 16:29

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.