2

Let $f(x,t)$, $u(x,t)$, and $S(x,t)$ be radial in $x$, i.e. $f(x, t)=f(|x|, t)$ for all $x \in \mathbf{R}^n$, and $u(x,t) = \int_{0}^{t}\int_{\mathbf{R}^n} S(x-y, t-s)f(y,s)dyds$. I want to write this expression radially. I know that $\int_{\mathbf{R}^n}f(|x|)dx = \omega_{n-1} \int_{0}^{\infty}f(r)r^{n-1}dr$ where $\omega_{n-1}$ is the surface area of the $n-1$ unit ball. I was thinking about applying this to the integrand $S(x-y, t-s)f(y,s)$, but how does this translation work given the shifted argument $x-y$?

Gary
  • 31,845
guanton
  • 325
  • You will end up with $S(|x-y|, t-s)f(|y|, s)$ in the integral. In particular, $S$ will depend on the distance between $x$ and $y$. – Giuseppe Negro May 21 '20 at 06:10
  • what would r be then? Would you first let r=|x-y| and then also define a radius variable for |y|? – guanton May 21 '20 at 06:45

1 Answers1

4

For concreteness let us consider the example $$ S(x)=e^{-\lvert x\rvert^2}, $$ since the $t$ variable plays no role in these considerations. The integral to study is $$ \int_{\mathbb R^n} e^{-\lvert x-y\rvert^2}f(y)\, dy.$$ We introduce polar coordinates $$ x=rq, \qquad y=\rho \omega, $$ where $r, \rho\ge 0$ and $q, \omega\in\mathbb S^{n-1}$. Then the integral becomes $$ \int_0^\infty f(\rho)\rho^{n-1}\underbrace{\left(\int_{\mathbb S^{n-1}} e^{-\lvert rq-\rho \omega\rvert^2} d\omega\right)}_{K(r, \rho):=}\, d\rho. $$ Note that $K$ is independent on $q$. Indeed, if you replace $q$ with $Rq$ in the integral, where $R\in SO(n)$ is a rotation matrix, the change of variable $\omega\to R\omega$ shows that you get the same result.

So this is the conclusion; our integral equals $$ \int_0^\infty K(r, \rho)\rho^{n-1} f(\rho)\, d\rho.$$ All of this actually holds for an arbitrary $S=S(|x|)$. From now on, to say something more precise, we need more information on $S$. In our case, $$\tag{1} K(r, \rho)=\int_{\mathbb S^{n-1}} e^{-(r^2+\rho^2 - 2r\rho q\cdot \omega)}\,d\omega=e^{-(r^2+\rho^2)}F(r\rho), $$ where $$\tag{*} F(t)=\lvert \mathbb S^{n-2}\rvert \int_{-1}^1 e^{2t y}(1-y^2)^\frac{n-3}{2}\, dy.$$ To compute this, I have used the following reasoning. First, as we said, the integral in (1) is actually independent of $q\in\mathbb S^{n-1}\subset \mathbb R^n$ , so we can choose $q$ however we like; for example, $q=(0, \ldots, 0,1)$, so that $\omega\cdot q=\omega_n$. Now, the surface element $d\omega^{n-1}$ on $\mathbb S^{n-1}$ satisfies the following recursive formula; $$\tag{R} d\omega^{n-1}=(1-\omega_n^2)^\frac{n-3}{2}d\omega_n d\omega^{n-2}.$$ So we conclude that $$ F(r\rho)=\int_{\mathbb S^{n-1}} e^{2r\rho q\cdot \omega}\, d\omega=\left(\int_{\mathbb S^{n-2}} d\omega^{n-2}\right) \int_{-1}^1 e^{2r\rho\omega_n}(1-\omega_n^2)^{\frac{n-3}{2}}\, d\omega_n, $$ which is exactly (*).


Some ideas on (R). It is easier to visualize that formula by letting $\omega_n=\cos \theta$, for $\theta\in [0, \pi]$. Then it reads $$ d\omega^{n-1}=(\sin \theta)^{d-2}d\theta d\omega^{n-2}.$$ And this is basically slicing $\mathbb S^{n-1}$ into a superposition of $n-2$-dimensional spheres, each one of them having radius $\sin \theta$. For more details, you can consult, for example, the introduction of the book "Analysis of spherical symmetry in Euclidean spaces" by Claus Müller.

  • 1
    Thank you! This is very helpful. The only thing that I am confused about is how you obtained the expression for $K(r, \rho)$ with the Bessel function. Could you explain that step please? – guanton May 21 '20 at 19:37
  • It's normal that you are confused, since I just dropped the result without showing any idea whatsoever on how to compute it. The truth is that I am not sure, and now that I am seeing that thing more closely, I suspect it is WRONG. Bessel functions arise from Fourier tranforms, but here there is no imaginary i, so I suspect I am off track. I'll have another look later. – Giuseppe Negro May 21 '20 at 21:30
  • I have updated the answer, removing the reference to Bessel functions. – Giuseppe Negro May 22 '20 at 10:59
  • I see, given the recursive formula, I agree with your result. However, could you give me a reference for the recurrence? I tried to look for it online but could not find it. I appreciate the help. – guanton May 22 '20 at 23:37
  • @guanton: it surely appeared on this website many times. I think I learned it in the book on spherical harmonics of Claus Muller (he has two, one old and a more recent remake). But it is also in the appendix to Classical Fourier analysis of Grafakos, I think. If needed I can compute it here – Giuseppe Negro May 22 '20 at 23:40
  • yes please, I don't quite see the intuition of the relationship – guanton May 23 '20 at 00:07
  • conceptualizing $w_n=cos(\theta)$ helps a lot by relating it back to spherical coordinates. It disambiguates the recursion significantly. If you can't find the time to make a picture, don't worry about it, I think that I can work things out – guanton May 24 '20 at 20:57