I'm working through some notes on showing the essential self-adjointness of the Laplace operator on $\Bbb R$ via the Fourier transform (see here) but there seems to be a little bit of liberty taken at one point which I can't quite justify. It is shown that $-\Delta$ is symmetric on $C_c^{\infty}(\Bbb R)$ and is also positive. These are pretty straightforward as is often the case with differential operators. The part which bothers me is showing that it is in fact essentially self-adjoint.

The easiest way (I think) to do it is via the Fourier transform since the Fourier transform turns the Laplace operator into a multiplication operator. The author wants to show that $\operatorname{Im}(-\Delta\pm i)$ is dense in $L^2(\Bbb R)$, or equivalently $\operatorname{Im}(-\Delta \pm i)^{\perp} = \{0\}$, which is of course one of the equivalent conditions for a symmetric operator to be essentially self-adjoint. I'll replicate the argument here and explain where I think some rigor is missing.

Suppose $g\in L^2(\Bbb R)$ is such that

$$0 = \langle (-\Delta \pm i)f,g\rangle \tag{1}$$

for all $f\in C_c^{\infty}(\Bbb R)$. Since the Fourier transform is unitary, we have

$$0 = \langle \mathcal{F}(-\Delta\pm i)f,\mathcal{F}g\rangle.$$

For $f\in C_c^{\infty}(\Bbb R)$, $-\mathcal{F}\Delta f(x) = 4\pi^2|x|^2\mathcal{F}f(x)$ so this becomes

$$ 0 = \langle (4\pi^2 |x|^2\pm i)\mathcal{F}f,\mathcal{F}g\rangle.$$

We want to show that $g$ is zero, so it stands to reason that we would push $4\pi^2|x|^2\pm i$ onto $\mathcal{F}g$ to then get that

$$ 0 = \langle \mathcal{F}f, (4\pi^2|x|^2\mp i)\mathcal{F}g\rangle.$$

Since the Fourier transform is unitary, $\mathcal{F}(C_c^{\infty}(\Bbb R))$ is dense in $L^2(\Bbb R)$ so this would say that $(4\pi^2|x|^2\mp i)\mathcal{F}g = 0$, giving that $\mathcal{F}g=0$ and so $g=0$ since the Fourier transform is unitary.

Herein lies my problem. There is no guarantee a priori that $(4\pi^2|x|^2\mp i)\mathcal{F}g$ is in $L^2(\Bbb R)$ so we cannot apply this reasoning. The obvious way around this is to require that $g$ be such that $g\in L^2(\Bbb R)$ and $(4\pi^2|x|^2\mp i)\mathcal{F}g\in L^2(\Bbb R)$. The set of $g$ satisfying this is dense since it contains $C_c^{\infty}(\Bbb R)$ by Paley-Wiener (or, better yet, contains the Schwartz space). For such $g$, the above argument would show that $g$ must indeed be the zero function. However it isn't clear to me that this is sufficient to say that any such $g$ satisfying $(1)$ must be zero. Is there something I am missing? Can density arguments be used somehow or is this a fundamentally flawed argument?

  • $\begingroup$ Related $\endgroup$ – Giuseppe Negro Feb 4 '15 at 20:45
  • 1
    $\begingroup$ I agree with your doubts. Moreover this proof does not use in any way the symbol of $-\Delta$, it could in principle work with any differential operator. Something is wrong $\endgroup$ – Giuseppe Negro Feb 4 '15 at 20:49
  • $\begingroup$ @GiuseppeNegro I think the only way the symbol really comes about is insofar that $|x|^2$ doesn't have any eigenfunctions in $L^2(\Bbb R)$ (giving that $\mathcal{F}g$ must be zero), which is quite elegant but there is definitely something missing. This domain mismatch issue is of course closely related to the domain of $(-\Delta)^*$ but I was hoping that somehow it could be skirted via the Fourier transform. I've seen other texts that go through the details surrounding $(-\Delta)^*$ and do it without the Fourier transform but it's not very clean by any means. $\endgroup$ – Cameron Williams Feb 4 '15 at 20:54

It is actually not too difficult to repair the argument: You know $0=\left\langle (4\pi^2 |x|^2\pm i)\hat{f},\hat{g}\right\rangle$; now the set $\{ (4\pi^2 |x|^2\pm i)\hat{f} \ | \ f\in C^\infty_0\}$ is dense in $L^2$, whence $\hat{g}=0$ and $g=0$. Why is this set dense? Adapt the proof (last paragraph) of Proposition 1.1.15 (page 5) here.

[It is not too difficult to find other references which carry out a correct proof using the Fourier transform, for instance http://people.math.gatech.edu/~loss/14SPRINGTEA/laplacian.pdf or Davies' "Spectral theory and differential operators" (Theorem 3.5.3). In Davies' book the result is generalized to a certain class of constant coefficient differential operators (depending on the symbol, as it has been mentioned in the comments by G.N.).]

  • 2
    $\begingroup$ I saw the notes by Loss but it dealt a little too heavily with some of the technical details so I was hoping there was a way to repair this argument. The proposition you pointed out is exactly what I was going for. I had a sneaking suspicion that that set was dense but I assumed it was a much lengthier proof. Thanks a bunch! $\endgroup$ – Cameron Williams Feb 7 '15 at 14:39

This started as a comment but soon became too big, so I am making it an answer.

The Fourier transform makes it clear that $-\Delta$ is self-adjoint on its natural domain in free space $H^2(\mathbb{R}^n)$ because it unitarily diagonalizes it. That is, the Fourier transform is unitary and it intertwines $\mathcal{F}\circ(-\Delta) = a(\xi)\circ\mathcal{F}$, where $a$ is a real valued function acting as a multiplication. (Precisely $a(\xi)=\lvert \xi\rvert^2$, but only the fact that it is real valued matters here).

Therefore, $-\Delta$ is self-adjoint on its natural domain iff $a$ is. The natural domain of this last operator is

$$ \text{Dom}(a)=\left\{ \phi\in L^2(\mathbb{R}^n)\ :\ a(\xi)\phi(\xi)\in L^2(\mathbb{R}^n)\right\}. $$ It is clear from realness of $a(\xi)$ that $\text{Dom}(a)\subset \text{Dom}(a^\star)$, that is, that $a$ is symmetric. The other inclusion is Riesz's representation theorem: if $\psi \in \text{Dom}(a^\star)$, then by definition $$ \frac{\left\lvert \int \phi(\xi)\overline{a(\xi)\psi(\xi)}\, d\xi\right\rvert}{\lVert \phi\rVert_2}\le C $$ for all $\phi\in \text{Dom}(a)$, which surely is a dense subset of $L^2$ space. By Riesz's representation theorem the supremum of this ratio is exactly $\lVert a\psi\rVert_2$, so this norm is finite and $\psi\in \text{Dom}(a)$.

You probably already knew this stuff, but I am posting here anyway hoping that it might help.


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.