10
$\begingroup$

Let $X$ be a Banach space and $K:X\to X$ be a compact operator. If $I+K$ is injective then it is onto and hence the inverse $(I+K)^{-1}$ is bounded. What kind of qualitative or quantitative assumptions on $K$ are necessary in order to find an explicit bound on the operator norm of $(I+K)^{-1}$?

In such generality, there is probably no answer (or there are too many), so let me be more specific and explain the situation I've encountered this. The Banach space is either $L^\infty$ or a weighted $L^2$ with weight $(1+|x|)^s$ on $\mathbb{R}^n$, these cases are all of interest. Call $H_0$ the operator $-\Delta$, $V\ $ the multiplication by a function $V(x)$, and $H=H_0+V$. This should all be properly defined but I leave the definitions aside so to accommodate as many additional assumptions as necessary. A standard procedure in spectral theory is to express the resolvent operator $R(z)=(H-z)^{-1}$ as a perturbation of $R_0(z)=(H_0-z)^{-1}$ through the formula $$ R_0(z)=R(z)(I+VR_0(z)). $$ Now the main step: assume we know that $VR_0$ is a compact operator and $I+VR_0$ is injective (this typically amounts to requiring that $H$ has no eigenvalues, or resonances). Then we can invert $I+VR_0$ and write $$ R(z)=R_0(z)(I+VR_0(z))^{-1} $$ which is a very powerful formula, but, alas!, not quantitative. Is there a way to bound the norm of the inverse based on any reasonable assumption on $H_0$ and $V$? Notice that the operator $K=VR_0$ is very explicit, one can write an explicit integral kernel and extract any kind of detailed information on it. Additionally, one can make any sort of assumption on the perturbed operator $H=H_0+V$ (e.g., about the distance of the eigenvalues from the continuous spectrum etc.)

EDIT: I might add that in the example the main point is to obtain a uniform estimate for the norm of $R(z)$ as $z$ approaches the real axis, which contains the spectrum of $H$. Otherwise it might seem odd to look for an estimate on $R(z)$, which is already a well defined and bounded operator on $L^2$, with known bounds. But this procedure gives a reasonable definition of $R(z)$ also for real $z$, certainly not as a bounded operator on $L^2$ but on different spaces.

$\endgroup$
2
  • $\begingroup$ Such a qualitative theory exists, in the following sense: You are describing a natural phenomenon that many people have to study. In order to do so, they develop methods. If you want to call them a theory, is a matter of taste. I wouldn't because these methods aren't simple and one needs to fine tune them on a case by case basis. One thing on relies on is what Florian mentions that one needs to control the distance of the spectrum of K to -1. The other thing, I can think of is that there are often better bounds available on P (1 + K)^{-1} Q then on (1 + K)^{-1}, where P,Q are projections. $\endgroup$
    – Helge
    Sep 30, 2010 at 16:06
  • 2
    $\begingroup$ I have some old notes on this problem at math.ucla.edu/~tao/preprints/Expository/compact_operator.dvi . If you assume some upper bound on the metric entropy of the image of $K$, and also bound the operator norm of $K$, one can control $(1+K)^{-1}$ quite well after projecting away a bounded dimensional bad subspace as Helge indicates, and this implies some (worse) bound on $(1+K)^{-1}$ itself involving the distance to the spectrum. $\endgroup$
    – Terry Tao
    Oct 27, 2022 at 20:00

4 Answers 4

2
$\begingroup$

Let $S$ denote the spectrum of $K$ (which consists of eigenvalues and 0). Then $\|(I+K)^{-1}\| = d(S,-1)^{-1}$. The statement $-1\notin S$ is equivalent to requiring that $K$ is injective, so your request for a "quantitative version" of the inversion formula is equivalent to a quantitative version of the injectivity of $I+K$, namely a lower bound on $d(S,-1)$.

A simple criterion would be: if $\|K\|\le C < 1$ then $d(S,-1)\ge 1-C$.

I think this is the best that can be said in the general case without knowing the operators.

Edit: As Jeff points out, this is valid only for normal $K$ acting on a Hilbert space. For general operators however it's still true that $\|K\|\le C < 1$ implies that $(I+K)$ is invertible and $\|(I+K)^{-1}\|\le (1-C)^{-1}$.

$\endgroup$
1
  • 4
    $\begingroup$ The norm identity is $\| (I+K)^{-1} \| = d(S,-1)$ is true for normal operators, but is not true in general! A simple example is already given by $K$ a Jordan block of size $n$ and eigenvalue $\lambda$. The norm in that case is of order $1/|\lambda +1|^n$ as $\lambda \rightarrow -1$. $\endgroup$ Sep 30, 2010 at 15:24
2
$\begingroup$

A bit similar but different to the answer of Florian. This is not a solution to the problem, but it is a start.

Suppose $\|K\|\le 1$ and let $r(K)$ denote the spectral radius of $K$. If $r(K)<1$ then certainly $H=(1-K)^{-1}$ exists. The most natural way to find (norm-)bounds on $H$ would be to bound $K^n$ and then use block summation of the geometric series. To bound $K^n$ we would need something more than the spectral radius formula, we would need some uniform spectral radius formula.

For further development I would start here (where the problem is considered in a general commutative Banach algebra):

Jan-Erik Björk: On the spectral radius formula in Banach algebras. Pacific J. Math. Volume 40, Number 2 (1972), 279-284.

$\endgroup$
1
  • $\begingroup$ Yes it really is! Also, compactness is a good to have in these problems. $\endgroup$ Oct 7, 2010 at 13:09
0
$\begingroup$

Suppose $N=\|(I+K)^{-1}\|$. In order to calculate $N$ one can do the following. Write $$N^2= \|(I+K)^{-1}(I+K^*)^{-1}\| = \|(1+K+K^*+K^* K)^{-1}\| = 1/(1+\lambda),$$ where $\lambda$ is the smallest eigenvalue of $K+K^*+K K^*$. Try to calculate it using your explicit formulas.

$\endgroup$
1
  • 1
    $\begingroup$ It would be easier to compute the eigenvalues of K and check which one is closer to −1, but is is exactly like computing the norm of $(I+K)^{−1}$ directly. $\endgroup$ Sep 30, 2010 at 13:35
-1
$\begingroup$

The second resolvent equation you are studying is an important tool in quantum scattering theory.

See M. Reed and B. Simon, "Methods of Modern Mathematical Physics", Vol IV, Academic Press

$\endgroup$

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.