How can I prove that the spectrum of the Laplace operator $$\Delta: H^2(\mathbb{R}^N)\subset L^2(\mathbb{R}^N)\rightarrow L^2(\mathbb{R}^N)$$ is $\sigma(\Delta)=[-\infty,0]$?


We determine the resolvent set of $(\Delta, H^2(\mathbb{R}^n))$. By definition, the complex number $\lambda$ belongs to the resolvent set if and only if the following equation (named resolvent equation) $$\tag{1}\Delta f -\lambda f = g$$ has a unique solution $f\in H^2$ for any fixed $g\in L^2$. The spectrum is the complementary of the resolvent set and its elements are called spectral values.

[EDIT: Summary. Here's two different approaches. The first uses the Fourier transform. The second computes explicitly a set of (generalized) eigenfunctions.]

Method 1. Taking the Fourier transform termwise we see that (1) is equivalent to $$(-\lvert\xi\rvert^2 - \lambda)\hat{f}=\hat{g}, $$ which has the unique formal solution $$\tag{2}\hat{f}=-\frac{\hat{g}}{\lambda+ \lvert\xi\rvert^2}.$$ Equation (2) determines a solution $f\in H^2$ of the resolvent equation (1) if and only if $f\in H^2$, that is $$\tag{3}\begin{array}{ccc} \lambda \text{ is in the resolvent set }&\iff& (1+\lvert\xi\rvert^2 )\hat{f} = -\frac{1+\lvert\xi\rvert^2}{\lambda+ \lvert\xi\rvert^2}\hat{g}\in L^2.\end{array}$$ When $\lambda\notin (-\infty, 0]$ the function $\frac{1+\lvert\xi\rvert^2}{\lambda + \lvert\xi\rvert^2}$ is bounded, so condition (3) is satisfied. Therefore $\lambda$ belongs to the resolvent set.

On the other hand, when $\lambda\in (-\infty, 0]$, the function $\frac{1+\lvert\xi\rvert^2}{\lambda + \lvert\xi\rvert^2}$ has a singularity at $\lvert \xi \rvert= -\sqrt{\lambda}$. If $\lambda \ne 0$ then we write $$ \frac{1+\lvert \xi\rvert^2}{\lambda +\lvert \xi\rvert^2}=\frac{1+\lvert \xi\rvert^2}{\sqrt{-\lambda}+\lvert \xi\rvert}\frac{1}{\sqrt{-\lambda}-\lvert \xi\rvert}.$$ We take a function $\hat{g}$ that is equal to $\lvert\sqrt{-\lambda}-\lvert\xi\rvert^2\rvert^{-\frac{N-1}{2}}$ in a neighborhood of the annulus $\lvert\xi\rvert=\sqrt{-\lambda}$ and vanishes elsewhere. This function is in $L^2$ but its product against $\frac{1+\lvert\xi\rvert^2}{\lambda +\lvert\xi\rvert^2}$ is not, because of the strong singularity around $\lvert\xi\rvert=\sqrt{-\lambda}$. Therefore condition (3) is not satisfied and $\lambda$ is not in the resolvent. The proof that $\lambda=0$ is not in the resolvent is completely analogous. (Alternatively, one can recall that the spectrum is always a closed set, so if the open half-line $(-\infty, 0)$ is contained in the spectrum, the value $\{0\}$ must be in the spectrum as well).

Method 2 (Added in response to Neal's comment below). Here's an alternative computation of the spectrum that determines a system of approximate eigenfunctions (in Weyl's sense).

Definition. Let $(A, D(A))$ be a closed operator on a Hilbert space $H$. We say that $f_k\in D(A)$ is a sequence of approximate eigenfunctions of $A$ relative to $\lambda\in \mathbb{C}$ if the following two conditions are satisfied:

  1. The sequence is normalized, meaning that $\lVert f_k\rVert =1$ for all indices $k$.
  2. It holds that: $\lVert Af_k-\lambda f_k\rVert \to 0$ as $k\to \infty$.

    If $\lambda$ admits a sequence of approximate eigenfunctions then it is a spectral value. The converse holds if $\lambda$ is on the boundary of the spectrum. (For more information see Teschl, Mathematical methods in quantum mechanics, Lemma 2.16 pag.76).

In the case at hand the eigenfunction equation $$ \Delta f-\lambda f=0 $$

is formally solved by the plane wave profiles $e^{i\sqrt{-\lambda}\xi_0\cdot x}$, where $\xi_0$ is a unit vector. However, those profiles have not a finite $L^2$ norm. To circumvent this difficulty we fix a smooth bump function $\Phi$, normalized so that $\lVert \Phi\rVert_2=1$, and we define \begin{align*} f^0_k(x)&=\frac{1}{k^{N/2}}\Phi\left(\frac{x}{k}\right); \\ f^{\lambda}_k(x)&=e^{i\sqrt{-\lambda}\xi_0\cdot x} f^0_k(x). \end{align*} Here $\xi_0\in \mathbb{S}^{N-1}$ is arbitrarily chosen.

Claim. For any fixed $\lambda\in (-\infty, 0]$ the sequence $f^\lambda_k$ consists of approximate eigenfunctions of $(\Delta, H^2)$ relative to the spectral value $\lambda$.

Proof. We consider the case $\lambda=0$ first. It is obvious that the sequence $f^0_k$ belongs to $H^2$ and that it is normalized. We have \begin{equation*} \lVert \Delta f^0_k\rVert_2=\frac{1}{k^2}\left\lVert \frac{1}{k^{N/2}} \Delta\Phi\left(\frac{x}{k}\right) \right\rVert_2 =\frac{C}{k^2}\to 0. \end{equation*} So $f^0_k$ is a sequence of approximate eigenfunctions. We note also that \begin{equation*} \left\lVert \nabla f^0_k\right\rVert_2=\frac{1}{k}\left\lVert \frac{1}{k^{N/2}} \nabla\Phi\left(\frac{x}{k}\right) \right\rVert_2\to 0; \end{equation*} this will be useful later. Now we consider the sequence $f^\lambda_k$ for a $\lambda\in (-\infty, 0)$. It is again obvious that $f^\lambda_k$ is normalized and that it belongs to $H^2$. Using the Leibniz formula for the Laplacian (that is, $\Delta(fg)=\Delta f g + 2 \nabla f\cdot\nabla g + f \Delta g$), we compute \begin{equation*} \begin{split} \lVert \Delta f^\lambda_k -\lambda f^\lambda_k \rVert_2 &=\left\lVert 2\nabla \left( e^{i\sqrt{-\lambda} \xi_0\cdot x}\right)\cdot \nabla f^0_k + e^{i\sqrt{-\lambda}\xi_0\cdot x} \Delta f^0_k \right\rVert_2 \\ &\le2\sqrt{-\lambda}\lVert \xi_0\cdot \nabla f^0_k\rVert_2 + \lVert \Delta f^0_k\rVert_2 \\ &\le 2\sqrt{-\lambda} \lVert \nabla f^0_k\rVert_2 +\lVert \Delta f^0_k\rVert_2 \to 0. \end{split} \end{equation*} This proves our claim.

  • $\begingroup$ Do you know of any constructive proofs, e.g., constructing a sequence of $\psi_k\in H^2(\mathbb{R}^n)$ satisfying Weyl's criterion? $\endgroup$ – Neal May 11 '14 at 17:32
  • 3
    $\begingroup$ Actually, the "eigenfunctions" of $\Delta$ are the plane wave profiles $f_\lambda(x)=e^{i\sqrt{-\lambda}\xi_0\cdot x}$, where $\xi_0\in\mathbb{R}^n$ is a unit vector. At Fourier side those correspond roughly to Dirac deltas concentrated at the point $\sqrt{-\lambda} \xi_0$. So a sequence $f_k\in H^2$ such that $\hat{f}_k$ is a bump function that concentrates more and more at $\sqrt{-\lambda} \xi_0$ should do the trick here. I have not checked the details though $\endgroup$ – Giuseppe Negro May 11 '14 at 17:58
  • $\begingroup$ Thank you! That makes perfect sense. $\endgroup$ – Neal May 11 '14 at 20:39
  • $\begingroup$ I am adding a commentary to this answer to account for your observation. $\endgroup$ – Giuseppe Negro May 11 '14 at 20:45
  • 1
    $\begingroup$ Quite clear! Thank you for the alternative approach. It looks like the Fourier transform of your argument should be what you sketched in your comment. $\endgroup$ – Neal May 11 '14 at 23:52

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.