Suppose I have an operator on $L^2(\mathbb{R}^3)$ of the form $A=I-B$, where:

1) $B$ is bounded on $L^2(\mathbb{R}^3)$, moreover $\Vert B\Vert_{L^2\rightarrow L^2}=1$,

2) $B$ has a positive integral kernel,

3) $A$ is injective.

Can i conclude that $A$ is invertible on $L^2(\mathbb{R}^3)$?

  • $\begingroup$ @Anton: I'm confused, if the integral kernel of $B$ is in $L^2$ then $B$ is a compact operator, and since $I-B$ is injective by Freedholm alternative we get that it's invertible. I'm missing something? Neverthless in my specific example the integral kernel of $B$ only belongs to $L^{2,\infty}$ $\endgroup$ – Capublanca Feb 10 '17 at 9:10
  • $\begingroup$ Anton, if the kernel is $L^2$, $B$ is Hilbert-Schmidt hence compact. Then $A$ is Fredholm of index zero. Since it is injective, it is invertible. $\endgroup$ – Denis Serre Feb 10 '17 at 9:10
  • $\begingroup$ Capublanca, you beat me for 20 seconds. $\endgroup$ – Denis Serre Feb 10 '17 at 9:11
  • $\begingroup$ @DenisSerre: Yeah, by photofinish :D $\endgroup$ – Capublanca Feb 10 '17 at 9:14

This is never true if the kernel of $B$ is symmetric and (say) positive-definite, as follows from the spectral theorem. In this case, $A$ has dense range for $\overline{\operatorname{ran}A} = (\operatorname{ker}A)^\perp=L^2(\mathbb R^3)$, but it is not surjective and $A^{-1}$ is unbounded. That $A$ is injective excludes $0$ from being an eigenvalue of $A$, but $0$ is still in the spectrum of $A$.

Indeed, $\sigma(B)\subseteq [0,1]$ with $1$ belonging to the spectrum of $B$ (since $\|B\|_{L^2\to L^2} =1$) implies that $\sigma(A)\subseteq [0,1]$ with $0$ belonging to the spectrum of $A=I-B$.

Example. Take the Gauss-Weierstrass kernel $(4\pi)^{-3/2}e^{-|x-y|^2/4}$ as kernel of $B$. There are two possibilities to see that $1$ is not an eigenvalue:

1) It holds $Bf= k\ast f$ with $k(x) = (4\pi)^{-3/2}e^{-|x|^2/4}$. If $f$ was a (normalized) eigenfunction, then $\bigl(\hat k(\xi)-1\bigr)\,\hat{f}(\xi)=0$ a.e., with $\hat{f}$ being the Fourier transform of $f$. But $\hat k(\xi) = e^{-|\xi|^2}$, so $\hat f(\xi) = 0$ a.e., a contradiction.

2) It holds $B= e^\Delta$, i.e., $B$ is the solution operator of the initial-value problem for the heat equation at time $1$. If $f$ was an eigenfunction of $B$, then $f$ would be an eigenfunction of the generator $\Delta$ (with eigenvalue $0$ as $e^0=1$), which again is a contradiction.

  • $\begingroup$ i think there is some problem with your argument. Take $B$ a compact operator with norm 1. In that case $A$ is invertible by Freedholm theory. $\endgroup$ – Capublanca Feb 10 '17 at 9:30
  • $\begingroup$ If $B$ is compact, then $1$ is an eigenvalue of $B$ and $A$ is not injective. $\endgroup$ – ifw Feb 10 '17 at 9:34
  • $\begingroup$ You still need to give an example where 1 is not an eigenvalue of B. $\endgroup$ – user1688 Feb 10 '17 at 11:07
  • $\begingroup$ @Anton $1$ cannot be an eigenvalue of $B$, as $A$ is supposed to be injective. Do you mean an operator $B$ meeting all the assumptions? $\endgroup$ – ifw Feb 10 '17 at 11:53
  • 1
    $\begingroup$ Yes. That's why I said "example". $\endgroup$ – user1688 Feb 10 '17 at 12:59

The answer is no and an example follows by amenability of $\mathbb{R}^n$ as a locally compact group.

Let $\mu$ be a compactly supported probability measure defined by a density function $f$ with respect to the Lebesgue measure and let $T$ be the operator given by the kernel $k(x,y)=f(y-x)$ (i.e. $T$ acts by convolution with $f$). $T$ has norm 1.

For $B_n=$ the ball of radius $n$ in $\mathbb{R}^3$ the operator satisfies $T\chi_{B_n}-\chi_{B_n}\to 0$ in norm. Thus the range of $I-T$ is not closed.

If the support of $f$ contains a neighborhood of the the origin then $I-T$ is injective since constant functions are not in $L^2$. However $T$ is not invertible.


I'm not sure if my unswer would be useful, but I have the following thing in mind.

  1. If B - is self-adjoint (symmetric) then it has a countable number of eigenvectors $e_j$ with eigenvalues which form a basis in $L^2(\mathbb{R}^3)$.
  2. Since $A$ is claimed to be injective then $B$ cannot have $\lambda=1$ as an eigenvalue.
  3. Then we can rewrite equation $(I-B)x=y$ in the following way: \begin{equation} \sum\limits_j (I-B)e_jx_j = \sum\limits_{j}e_jy_j, \end{equation} and consequently define the quasi-inverse: \begin{equation} x_j = \dfrac{y_j}{1-\lambda_j}, \, j = \overline{1,\infty}. \end{equation} Now the question is if $x = (x_j)\in L^2(\mathbb{R}^3)$? The answer is not necessarily if $\lambda_j$ have a convergent subsequence $\lambda_{j(k)}$: \begin{equation} \lambda_{j(k)}\rightarrow 1, \, k\rightarrow \infty. \end{equation} To prove this I will do the following: since $B$ has an eigenvalue basis, I will define the following symmetric operator $Q:L^2(\mathbb{R}^3)\rightarrow L^2(\mathbb{R}^3)$: \begin{equation} Qe_j = \left(1-\dfrac{1}{j+1}\right)e_j, \end{equation} where $\{e_j\}$ is the spectral basis for $B$. It is easy to see that $Q$ is symmetric, self-adjoint, positive and \begin{equation} \|Q\| = 1. \end{equation} Also $I-Q$ is injective. Indeed, if $I-Q$ is not injective then for some $x\neq 0$ it holds: $$ Qx = x \rightarrow \|x-Qx\|^2_{L^2} = \sum\limits_{j=1}^{\infty} \dfrac{x^2_j}{(j+1)^2} > 0, $$ which makes a contradiction.

Applying the same process for $I-Q$ we obtain that series $\{x_j(y_j)\}_j$ will explode in infinity for some chosen $y$. Hence we have an example when operator I-Q is not invertible, but all conditions are satisfied.


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.