Let $1\le p<q\le\infty$, $L:\ell^p\to\ell^p,x\mapsto(x_2,x_3,\ldots)$ and $R:\ell^q\to\ell^q,y\mapsto(0,x_1,x_2,\ldots)$.

It's easy to show $L'=R$, $\sigma(L)=\sigma(R)=\overline B_1(0)$, $\sigma_p(L)=B_1(0)$, $\sigma_r(L)=\emptyset$, $\sigma_c(L)=\partial B_1(0)$, $\sigma_p(R)=\emptyset$ and $B_1(0)\subseteq\sigma_r(R)$.

But I'm failing to show that $\sigma_c(R)=\emptyset$. How can we do that?

Clearly, if $\lambda\in\sigma_c(R)$, then $\mathcal R(\lambda-R)$ is dense and hence $\lambda\not\in\sigma_p(L)$, i.e. $\lambda\in\partial B_1(0)$ ... But I'm not able to derive a contradiction from that.


Since $q>1$, you have $R=L'$, considering $L$ acting on $L^p$ for $p$ conjugate to $q$. From $$ \|(R-\lambda I)x\|_p=\|Rx-\lambda x\|_p\geq\|Rx\|_p-|\lambda|\,\|x\|_p=(1-|\lambda|)\,\|x\|_p, $$ we get that $R-\lambda I$ is bounded below for all $\lambda\in B_1(0)$. This precludes $\lambda$ from being an approximate eigenvalue. Combined with the fact that the boundary of the spectrum consists of approximate eigenvalues shows that $\sigma_{\rm ap}(R)=\partial B_1(0)$.

From $\sigma(R)=\sigma_r(R)\cup\sigma_{\rm ap}(R)$ we conclude that $B_1(0)\subset\sigma_r(R)$.

If $\lambda\in\partial B_1(0)$, then $ \lambda\not\in\sigma_p(L)$ (here we use $p<\infty$). This means that $\ker(L-\lambda I)=\{0\}$. It is well-known (proof at the end) that if an operator is injective, then its adjoint has dense range. So $\overline{\operatorname{ran}(R-\lambda I)}=\ell^q$, showing that $\lambda\in\sigma_c(R)$. That is, $\partial B_1(0)\subset \sigma_c(R)$. We have $$ \overline{B_1(0)}=\sigma(R)=\sigma_r(R)\cup\sigma_c(R) $$ (since $\sigma_p(R)=\emptyset$). As $\sigma_r(R)$ and $\sigma_c(R)$ are disjoint, with $B_1(0)\subset\sigma_r(R)$ and $\partial B_1(0)\subset\sigma_c(R)$, it follows that $$ \sigma_r(R)=B_1(0),\qquad \sigma_c(R)=\partial B_1(0). $$

Proof that $T:X\to Y$ is a bounded operator and $T'$ is injective, then $T$ has dense range.

If $\operatorname{ran}T$ were not dense, there exists $y\in Y\setminus\overline{\operatorname{ran}T}$. By Hahn-Banach there exists $g\in Y'$ with $g(y)=1$ and $g(Tx)=0$ for all $x\in X$. This means that $$ (T'g)(x)=g(Tx)=0 $$ for all $x$, so $T'g=0$. As $T'$ is injective, we get that $g=0$ a contradiction.

  • $\begingroup$ Thank you for your answer. It's clear to me that $B_1(0)\subseteq\sigma_r(R)$ (missed that in the question). But how do we prove $\overline{\operatorname{ran}(R-\lambda I)}=\ker(T-\overline\lambda I)^\perp$ (and what is $T$? $T=L$ I guess)? (I'm curious, by the way, why you're considering $\overline\lambda$. I mean, we know that if $\lambda\in\partial B_1(0)$, then $\lambda\not\in B_1(0)=\sigma_p(L)$ and clearly $\overline\lambda\not\in\sigma_p(L)$ as well. So, I guess the point is the former identity ...) $\endgroup$ – 0xbadf00d Feb 2 at 18:21
  • $\begingroup$ The formula that I know, for a general bounded linear operator $T$ on a complex Banach space $X$, is $\overline{\mathcal R(T)}=(\mathcal N(T'))_\perp$. Now, if $\lambda\in\mathbb C$, then $(\lambda-T)'=\lambda-T'$ (not $\overline\lambda-T'$). Did you confuse that with the Hilbert space adjoint? My expectation would be $\sigma_c(R)=\emptyset$. $\endgroup$ – 0xbadf00d Feb 2 at 18:24
  • $\begingroup$ Yes, you are right. I tend to think about Hilbert spaces. $\endgroup$ – Martin Argerami Feb 2 at 18:27
  • $\begingroup$ I'm on my phone right now, so I'll get back to this a bit later. $\endgroup$ – Martin Argerami Feb 2 at 18:31
  • $\begingroup$ Thank you in advance for your effort. $\endgroup$ – 0xbadf00d Feb 2 at 18:32

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.