If I have a normal bounded operator $T$ on a (say separable) Hilbert space $H$ then the following resolvent growth condition holds:


where $R(z,T)=(T-zI)^{-1}$ is the resolvent and $\sigma(T)$ denotes the spectrum of $T$ and $r$ the spectral radius. In the non-normal case things are not so pleasant and we may have


(note that we always have $\geq$ in the above). My question is whether there are known cases (with $H$ infinite dimensional and $T$ not compact) where growth of $R(z,T)$ is known in terms of $dist(z,\sigma(T))$ which isn't just the linear case above. For instance, replacing the linear growth by quadratic growth?

  • $\begingroup$ Why does the equality hold for normal operators? I can see it for self adjoint, but not sure why it works for a general normal operator $\endgroup$
    – Niebla
    Dec 18 '17 at 7:29

Start with a matrix on $\mathbb{C}^2$ (you can extend to a non-compact operator in $\ell^2$ later) defined as $$ A = \left[\begin{array}{ccc}0 & 1 & 0 \\ 0 & 0 & 1 \\0 & 0 & 0\end{array}\right]. $$ Then $A^3=0$ and $A^2 \ne 0$, as $A$ maps the third basis element to the second, the second to the first, and the first to the zero vector. You can directly verify that the resolvent of $A$ is defined for all $\lambda\ne 0$ by $$ (A-\lambda I)^{-1}=-\frac{1}{\lambda^3}(A^2+\lambda A + \lambda^2 I) $$ Near $\lambda=0$, the resolvent has a $1/\lambda^3$ behavior. For example, if you apply the above to the third basis vector, you $-1/\lambda^3$ times the first basis vector plus $-1/\lambda^2$ times the second basis vector and $-1/\lambda$ times the third basis vector. You can certainly operator norm bounded $(A-\lambda I)^{-1}$ by a constant times $1/|\lambda|^3$, and that order of bound is achieved.

Finally, you can put this into $\ell^2$ in such a way that the result is not compact. To do this, define $A$ on the first three basis vectors, and let $A$ be the identity on the subspace spanned by the remaining basis vectors. Then $A$ is not compact. The resolvent of the extended operator is defined everywhere except for $\{0,1\}$. If you use a higher-order nilpotent matrix on a higher-dimensional subspace, you end up with a higher-order pole of the resolvent. That way you can get any reciprocal polynomial behavior you want for the resolvent.

Why this is general: This is basically the general case for an isolated point of the spectrum. For example, if $\mu$ is an isolated point of the spectrum, then you can choose any positively oriented circular contour in the resolvent set that encloses only $\mu$ from the spectrum in its interior, and then show that the following is a projection $$ P_{\mu} = \frac{1}{2\pi i}\oint_{C} \frac{1}{\lambda I-A}d\lambda. $$ Then, if you can choose $n$ large enough that $(\lambda-\mu)^{n}(\lambda I-A)^{-1}$ is non-singular, you obtain a nilpotent condition: $$ (A-\mu I)^n P_{\mu} = \frac{1}{2\pi i}\oint_{C}\frac{(\lambda-\mu)^{n}}{\lambda I-A}d\lambda = 0 $$ In other words, $A-\lambda I$ is nilpotent of order $n$ on $P_{\mu}X$ iff $(A-\lambda I)^{-1}$ has an isolated singularity of order $n$ at $\mu$. I gave you an example of this, but it's basically the general case for an isolated point of the spectrum. (You can use this in finite-dimensional spaces to derive the Jordan blocks associated with $\mu$ because the singularities of the resolvent are all isolated singularities of finite order in finite-dimensional cases.)

It should be noted that in infinite-dimensional spaces, the resolvent can have an isolated essential singularity (i.e., of infinite order.)

  • $\begingroup$ Nice! But this seems like cheating a bit since we can split the operator in a trivial way. $\endgroup$
    – Mathmo
    Sep 7 '16 at 18:54
  • $\begingroup$ @Mathmo : I added more so that you can see why this is more general than you might think. $\endgroup$ Sep 7 '16 at 19:32
  • $\begingroup$ Ah, I see yes. Yes it's making more sense. Are there conditions that ensure we don't have an isolated essential singularity? For instance, compact perturbations of normal operators? $\endgroup$
    – Mathmo
    Sep 9 '16 at 7:10
  • $\begingroup$ @Mathmo : If a normal operator has a resolvent with an isolated singularity at $\lambda$, then the order of the pole is 1 because $(\lambda I-A)^{-2}P_{\mu}=0$ implies $(\lambda I- A)^{-1}P_{\mu}=0$. I'll let you consider the Volterra operator $Vf = \int_{0}^{x}f(t)dt$ on $C[0,1]$ or on $L^2[0,1]$. $V$ is compact, which makes it a compact perturbation of $0$. Try it, and see what you get for the resolvent. $\endgroup$ Sep 9 '16 at 12:31

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.