I was looking at a problem which involves the Green's function of a fractional Poisson equation. To fix notation, let $D\subset \mathbb{R}^n$ very nice, i.e. a hypercube, and \begin{equation} \begin{aligned} (-\Delta)^{\alpha/2} u &= f \quad in \; D\\ \end{aligned} \end{equation} with periodic boundary conditions.

Define $(-\Delta)^{\alpha/2}$ in the spectral way, i.e. given $\alpha>0$, \begin{equation} (-\Delta)^{\alpha/2} v = \sum_{k=1}^{\infty} \lambda^{\alpha/2}_k \langle v, e_k\rangle e_k \\ \end{equation}

My questions are very simple and for sure well-known in the expert community, so I hope I can get some good reference by you:

1) is it true that the green's function looks like \begin{equation} G(x,y) = \sum_{k=1}^{\infty} \lambda_k^{-\alpha/2} e_k(x) e_k(y) ? \end{equation}

2) Are there any results about the order of the singularity of a Green's function to the fractional Laplacian? For the usual Poisson equation, we know that it has a logarithmic singularity, so I guess it should be sith like G(x,y) = |x-y|^? + g(x,y) and the little g has something to do with the boundary?

I'd be grateful for any literature advice you may have. I read https://arxiv.org/abs/1507.07356 , but the topic was not adressed, and the articles by Bogdan, Chen, Song or others are already very specialized for domains with $C^{1,1}$ or Lipschitz boundary, and I wasn't able to strip them down to the easy question I have.


Oh, well, I'll try my best to help.

(0) Hypercube is not really super-nice, it is merely Lipschitz! This is of course a joke, because periodic boundary conditions indeed make it perfectly regular. But then...

(1) Even for the usual Laplace operator the Green function (in the strict sense: as the kernel of the inverse operator $(-\Delta_{\mathbb{T}})^{-1}$) is not defined, because $\Delta_{\mathbb{T}}$ has eigenvalue $0$ (corresponding to a constant eigenfunction $e_1$; I use ${\mathbb{T}}$ for the hypercube here and below). The same is clearly true for the fractional power $(-\Delta_{\mathbb{T}})^{\alpha/2}$, if it is defined via spectral theory (in this case: Fourier series), as in the second display in the statement of the question.

(2) Noteworthy, it does not matter whether we take the fractional power of $-\Delta$ first and then periodize (is this the right word?), or vice versa. The easiest way to see this is to note that Bochner's subordination formula for the heat kernel of a fractional power commutes with periodization, so to say.

(3) The inverse operator $G^{(\alpha)}_{\mathbb{T}} = (-\Delta_{\mathbb{T}})^{-\alpha/2}$ is well-defined on the orthogonal complement of constants. Of course it can be expressed in terms of spectral theory: $$G^{(\alpha)}_{\mathbb{T}} v = \sum_{k = 2}^\infty \lambda_k^{-\alpha/2} \langle v, e_k \rangle e_k ,$$ but there is a more useful expression: the kernel of $G$ is a "compensated" integral of the heat kernel: $$G^{(\alpha)}_{\mathbb{T}}(x, y) = \int_0^\infty (p^{(\alpha)}_{\mathbb{T}}(t, x, y) - 1) dt .$$ Here $p^{(\alpha)}_{\mathbb{T}}(t, x, y)$ is the heat kernel of $(-\Delta_{\mathbb{T}})^{\alpha/2}$. Indeed, in order to prove the above formula, one simply expands the heat kernel in terms of the eigenfunctions and uses Fubini.

(4) Since we have already observed that $p^{(\alpha)}_{\mathbb{T}}(t, x, y)$ is the periodization of the heat kernel $p^{(\alpha)}(t, x, y)$ in full space, it is not hard to see that the order of the singularity of $G^{(\alpha)}(x, y)$ at the diagonal is preserved. Namely, $$G^{(\alpha)}_{\mathbb{T}}(x, y) - G^{(\alpha)}(x - y) = \int_0^\infty \left(\sum_{k \in \mathbb{Z}^d \setminus \{0\}} p^{(\alpha)}(t, x, y + k) - 1\right) dt$$ is uniformly bounded due to well-known bounds on the heat kernel.

(5) In fact the order of singularity of the Green function of $(-\Delta)^{\alpha/2}$ is the same in essentially every reasonable space. One only needs Gaussian (or even sub-Gaussian!) bounds for the heat kernel of $(-\Delta)^{\alpha/2}$ for small times, and the subordination formula. I suppose this can be found in Andrzej Stós's article Symmetric $\alpha$-stable processes on $d$-sets, which unfortunately does not seem to be easily available online.

Finally, let me stress that I do not really know the literature on fractional Laplace operators in spaces other than $\mathbb{R}^n$.

  • $\begingroup$ Thanks a lot, Mateusz Kwasnicki, that were helpful hints to me (though indeed Stos' article I couldn't find online). May I ask one mor thing, regarding my second question: You wrote in this post mathoverflow.net/questions/203578/… that it is known for the Green's function that $G(x)≈min(|x|−n−α,|x|−n+α)$ (where n is the dimension) - that is indeed what a first rule of thumb, starting from the Green's function of the usual Laplacian, would lead to. You also said that "there is no closed-form expression" - do you have any (elementary) reference? $\endgroup$
    – Kira G.
    Jul 6 '18 at 15:41
  • $\begingroup$ @KiraG.: No, I do not. In fact I do not even think someone tried proving that a closed-form expression does not exist. As I understand my previous comment now, what I meant is: if you open a standard table of integrals (like Gradshteyn–Ryzhik) or ask Mathematica, you will not find nor get a closed-form answer. $\endgroup$ Jul 6 '18 at 17:58

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.