The following might be quite straightforward, but I very rarely work in detail with unbounded operators, so I thought it would be worth seeing quickly if I have overlooked an example that is obvious from the right point of view.

Let $H=L^2(-\infty,\infty)$ and let $S:H \supset {\rm dom}(S) \to H$ be the densely-defined operator $(Sf)(x)=e^xf(x)$, with domain $$ {\rm dom}(S)=\{ f\in L^2(-\infty,\infty) \colon \int_{-\infty}^\infty e^{2x} |f(x)|^2\,dx <\infty \} $$ Then $S$ is self-adjoint.

Question. Does there exists a projection $P:H\to H$, with range $V$, satisfying the following properties?

  1. $V$ is infinite-dimensional and $V':={\rm dom}(S)\cap V$ is dense in $V$.

  2. there is a compact operator $R:V\to V$ such that $RPSP\vert_{V'}= I_{V'}$?

Some remarks:

(a) If we weakened 2. to merely require $R$ being bounded, then this should be easy by taking $V=L^2[0,\infty)$ and $R:V\to V$ to be multiplication by $e^{-x}$. But of course $R$ has continuous spectrum, so it can't be compact.

(b) If we intertwine with the Fourier transform or something similar, perhaps we can we obtain such a $V$ as the solution space to a suitable differential equation? (I have not looked at this vague idea in any detail yet.)

(c) If the answer to the question is positive, can we even get $R$ being trace-class?

  • 1
    "Projection" here means "orthogonal projection"? – Nate Eldredge Aug 17 at 3:09
  • @NateEldredge oops, yes. (Had my op alg hat on, where that is the usual shorthand) – Yemon Choi Aug 17 at 3:25
up vote 7 down vote accepted

Sure, for instance let $P$ be the orthogonal projection onto the closed span of the characteristic functions $\chi_{[n,n+1)}$ for $n \in \mathbb{N}$. You get property 1 because each of these functions is in the domain of $S$, and you get property 2 because, identifying $V$ with $l^2$ in the obvious way, the operator $PSP$ becomes multiplication by, I think, $e^{n}(e - 1)$. So this compression has compact, even trace-class inverse, namely multiplication by $e^{-n}(e-1)^{-1}$.

  • 1
    Kicking myself now for not thinking of trying to discretize the spectrum by brute force. Looks good - I'll come back tomorrow and double-check the details – Yemon Choi Aug 17 at 3:42
  • 1
    BTW, by comprssing to an increasing sequence $(n_k)$ you can make the eigenvalues of $PSP$ go to infinity as fast as you want. – Nik Weaver Aug 17 at 13:52

Your Answer


By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

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