This is my question:

Is $\mathcal{G}(\ell_2)$ is dense in $\mathcal{B}(\ell_2)$?

I am attempting to show that it is not by showing that the right-shift - call it $T:\ell_2 \rightarrow \ell_2$ - cannot be made invertible via a small perturbation. We know that the spectrum of $T$ is $$\sigma(T) = \overline {\mathcal D} = \{z \in \Bbb C \ | \ |z| \le 1 \}.$$ This includes $0$, showing that $T$ itself is not invertible (pretty clear as not surjective, although is injective). Also, it shows us that if $T$ is perturbed by a small multiple of the identity (in fact, any multiple less than or equal to $1$), then it is still non-invertible.

The issue that I'm having is showing that for any $S \in \mathcal{B}(\ell_2)$ with $\|S\|=1$, given $\epsilon > 0$ small enough, $T + \epsilon S$ is not invertible. If we choose $S$ injective, then $T + \epsilon S$ is injective, so we need $S$ not surjective. We know by the open mapping theorem that any surjective (linear and continuous) operator is an open map - can I use this to show obtain a contradiction if $T + \epsilon S$ is invertible?

If I could be given a hint, then I'd be most appreciative. I would, however, only like a small hint, as I would like to learn from this question, so please don't just blast out the answer!
As a note, the first part of the question asked to show that $\mathcal{F}(\ell_2)$ (the set of finite rank operators on $\ell_2$) is dense in $\mathcal{K}(\ell_2)$ (the set of compact operators on $\ell_2$). I can't see how this could be used, but maybe...?
Thank you! :)

As a final note, hopefully this isn't a duplicate question: it seems a bit odd to me that no-one has asked it before, but I couldn't find it. Apologies if it is.

  • $\begingroup$ The set of all invertible operators is open. You can prove this by showing that if $\|T\|<1$, then $I-T$ is invertible. $\endgroup$ – Vincent Boelens Dec 1 '14 at 20:23
  • $\begingroup$ Yes, that is one of the things that I've covered. Just wasn't sure how to use this! Need to find a ball (open or closed) about the right shift for which no element in it is invertible... $\endgroup$ – Sam T Dec 1 '14 at 21:12
  • $\begingroup$ Sorry, I was confused before, openness of the invertible operators doesn't help of course. $\endgroup$ – Vincent Boelens Dec 1 '14 at 21:23
  • $\begingroup$ It was my first port of call for thought, but didn't help. No worries - think I've got it now though. :) $\endgroup$ – Sam T Dec 1 '14 at 21:34

I suppose that $0\in\mathbb{N}$. If you follow the other convention, replace $e_0$ with $e_1$ in the following.

If $T+ \epsilon S$ is surjective, then there is an $x\in \ell_2$ with $\lVert x\rVert = 1$ and

$$(T+\epsilon S)(x) = c\cdot e_0$$

for some $c\neq 0$.

Then $\lVert\epsilon Sx\rVert^2 = \lVert c\cdot e_0 - Tx\rVert^2 = \lvert c\rvert^2 + 1$ and hence $\lVert \epsilon S\rVert \geqslant \sqrt{1+\lvert c\rvert^2} > 1$, therefore $\epsilon > 1$.

| cite | improve this answer | |
  • $\begingroup$ This was the sort of thing that I was thinking. I then know that $(\epsilon S x)_1 = c$ (meaning first component) because $(Tx)_1 = 0$. Surely this is ok? Or does it mean that if it is the case, then $T + \epsilon S$ is not injective, ie non-trivial kernet? $\endgroup$ – Sam T Dec 1 '14 at 21:14
  • $\begingroup$ (Just typing a comment now explaining that I think I've got it, so if you see this before it's arrived, hold fire :)!) $\endgroup$ – Sam T Dec 1 '14 at 21:23
  • $\begingroup$ So that I can fit it in in one go, I've made it an answer (as hopefully it's also correct!). If you could check it, then I'd be most appreciative! :) $\endgroup$ – Sam T Dec 1 '14 at 21:34

Ok, so I think I've got it (I use $\Bbb N = \{1,2,...\}$): Let $S' \in \mathcal{B}$ be given by $S'y = Sy - \langle e_1,Sy \rangle e_1$. $T + \epsilon S'$ is an isomorphism for small enough $\epsilon \ (>0)$ from $H_0 = \ell_2$ to $H_1 = \{ x \in \ell_2 | \langle e_1,x \rangle = 0 \}$. Now suppose that $e_1 \in \text{range}(T + \epsilon S)$: there exists $x \in B_H$ (ie $x \in \ell_2$ with $\|x\| = 1$) such that $(T + \epsilon S)x = c e_1$, for $c \neq 0$; in fact, $c = \langle e_1, Sx \rangle$. Then $$(T + \epsilon S)x = Tx + \epsilon S' + \langle e_1, Sx \rangle e_1$$ $$\Rightarrow (T + \epsilon S')x = 0$$ by definition of $c$. This is a contradiction since $x \neq 0$ and $T + \epsilon S$ is an isomorphism, so in particular is injective. Hence $e_1 \not\in \text{range}(T + \epsilon S)$, and so $T + \epsilon S$ is not surjective, and hence not invertible.

| cite | improve this answer | |
  • $\begingroup$ Correct. I've added the argument I had in mind to my answer, it amounts to the same, but is shorter. $\endgroup$ – Daniel Fischer Dec 2 '14 at 10:15
  • $\begingroup$ Ah yes, I see, because it's orthogonal so bring out Pythagoras. Interesting that your shows that we require $\epsilon > 1$, whereas I just showed that it didn't work for $\epsilon$ "small enough". I guess that $1$ is an intuitive bound as the set of eigenvalues is the unit disc. Thank you. :) $\endgroup$ – Sam T Dec 2 '14 at 11:29

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.