2

We are asked to solve the following PDE:

$$ \Delta u(x) = \phi(x) \hspace{.3 cm} x\in \mathbb{R}^3 $$ $$ \phi(x) = \begin{cases} \exp\left(-\dfrac{1}{1−x^2}\right) & |x|<1 \\ 0 & |x| \geq 1 \end{cases} $$

I know in order to get the final solution, I know I need to do the convolution of the fundamental solution with the inhomogeneous part but I genuinely do not know if the convolution it self is integrable at this point. Any help is extremely appreciated.

BrazyOski
  • 121
  • As mentioned in the comments to the Answer of K, your $\phi$ is the standard example of a bump function, i.e $\phi \in C^\infty_c$ is a smooth function with compact support. The proof that it is smooth, despite it being piecewise defined and compactly supported can be found in wikipedia, in the wiki page "non-analytic smooth functions" (can't get link now sorry). As for the convolution making sense, it is proven in Evans' PDEs book, chapter 2 (in the generality of $C^2_c$ functions) – Calvin Khor Jul 12 '20 at 13:23

1 Answers1

2

I'm going to use slightly different notation, because I am an applied, not pure, mathematician. We want solutions of $$\nabla^2u(\mathbf{r})=f(\mathbf{r})$$ Here $\mathbf{r}=(x,y,z)$, the position vector. We know $$ f(\mathbf{r}) = \begin{cases} \exp\left(-\dfrac{1}{1−\Vert\mathbf{r}\Vert^2}\right) & \Vert\mathbf{r}\Vert<1 \\ 0 & \Vert\mathbf{r}\Vert \geq 1 \end{cases} $$ Let's do the $\Vert\mathbf{r}\Vert\geq1$ case first. Call this solution $u_2$. Since $f$ is spherically symmetric, it makes sense to express our problem in spherical polar coordinates. We want solutions to $$\frac{1}{r^2}\frac{\partial}{\partial r}\left(r^2\frac{\partial u_2}{\partial r}\right)=0$$ It's fairly elementary to see that our solution is $$u_2(r)=a_1+\frac{a_2}{r}$$

Now onto the $\Vert\mathbf{r}\Vert<1$ case. Call this solution $u_1$. Now we want solutions to $$\frac{1}{r^2}\frac{\partial}{\partial r}\left(r^2\frac{\partial u_1}{\partial r}\right)=\exp\left(\frac{-1}{1-r^2}\right)$$ Since we've already found the solution of the homogeneous case, all we need is a particular solution of the above and we can combine it with the homogeneous solution for a full answer. Let's expand the L.H.S, $$\frac{1}{r^2}\frac{\partial}{\partial r}\left(r^2\frac{\partial u_1}{\partial r}\right)=\frac{1}{r^2}\left(2r\frac{\partial u_1}{\partial r}+r^2\frac{\partial^2u_1}{\partial r^2}\right)=\frac{\partial^2u_1}{\partial r^2}+\frac{2}{r}\frac{\partial u_1}{\partial r}$$ So we need to study the ODE $${u_1}''(r)+\frac{2}{r}{u_1}'(r)=\exp\left(\frac{-1}{1-r^2}\right)$$

Let $v(r)=u_1'(r).$ We now have the first order ODE $$v'(r)+\frac{2}{r}v(r)=\exp\left(\frac{-1}{1-r^2}\right)$$ This is a first order linear ODE, so $$v(r)=\frac{1}{\mu(r)}\int \mu(r)\exp\left(\frac{-1}{1-r^2}\right)\mathrm{d}r$$ Where $\mu(r)=\exp\left(\int\frac{2}{r}\mathrm{d}r\right)=\exp\left(2\ln(r)\right)=r^2$. So we need to take a look at the integral $$\int r^2\exp\left(\frac{-1}{1-r^2}\right)\mathrm{d}r$$ Unfortunately, this integral can't be expressed in terms of elementary functions, so we'll use a series expansion instead. A Maclaurin series shows us that $$\exp\left(\frac{-1}{1-x^2}\right)\approx \frac{1}{e}\left(1-x^2-\frac{x^4}{2}\right)$$

Therefore $$v(r)\approx \frac{1}{r^2}\frac{1}{e}\int r^2\left(1-r^2-\frac{r^4}{2}\right)\mathrm{d}r$$ The algebra is rather tedious, but at the end we find $$v(r)\approx \frac{1}{e\cdot r^2}\left(\frac{r^3}{3}-\frac{r^5}{5}+\frac{r^7}{14}+b_1\right)=\frac{b_1}{r^2}+\frac{r}{e}-\frac{r^3}{5e}+\frac{r^5}{14e}$$ Since $v(r)={u_1}'(r)$, $$u_1(r)\approx\int \left(\frac{b_1}{r^2}+\frac{r}{e}-\frac{r^3}{5e}+\frac{r^5}{14e}\right)\mathrm{d}r=-\frac{b_1}{r}+\frac{r^2}{2e}+\frac{e}{420}r^4(5r^2-21)+b_2$$

Thus, finally, our general answer is $$ u(r) \approx \begin{cases} u_1(r)=-\frac{b_1}{r}+\frac{r^2}{2e}+\frac{e}{420}r^4(5r^2-21)+b_2 & r<1 \\ u_2(r)=a_1+\frac{a_2}{r} & r \geq 1 \end{cases} $$ We want $u$ to be smooth at the boundary, that is $u_1(1)=u_2(1)~;~{u_1}'(1)={u_2}'(1)$ We can compute $${u_1}'(r)=\frac{b_1}{r^2}+\frac{e}{70}r^3(5r^2-14)+\frac{x}{e} ~;~ {u_2}'(r)=\frac{-a_2}{r^2}$$ Therefore $${u_1}'(1)=b_1-\frac{9e}{70}+\frac{1}{e}=-a_2={u_2}'(1)$$ And $$a_1+a_2=-b_1+\frac{1}{2e}-\frac{4e}{105}+b_2$$ Since we have two equations in four variables, we don't have sufficient information to work out $a_1,a_2,b_1,b_2$. Other IC's and BC's are needed to compute these.

Hope this helped!

EDIT: Linked below is an interactive plot of the solution curves. $b_1$ and $b_2$ are free parameters and $a_1,a_2$ will vary with these parameters to keep the curve smooth.

DESMOS

K.defaoite
  • 12,536
  • 1
    The theoretical answer is fully determined so anything left undetermined is either work not yet done or an artifact of your approximation. Note that although the perturbation is piecewise defined it’s actually $C^\infty$ smooth, in fact I would say it is the standard example of a bump function – Calvin Khor Jul 12 '20 at 02:58
  • I would guess a more suitable expansion, or more than one expansion should be used to match the boundary since the bump function loses analyticity at $|x|<1$ – Calvin Khor Jul 12 '20 at 03:05
  • @Calvin Khor I actually wasn't aware of the bump function; it seems rather well studied. Is there perhaps a more suitable expansion one could use? Maybe the Fourier series? – K.defaoite Jul 12 '20 at 03:12
  • @K.defaoite This is fantastic! Very much appreciated. :) – BrazyOski Jul 12 '20 at 03:22
  • @BrazyOski Oh goodness! I've found a mistake. I forgot to add the homogeneous solution to the particular solution of $u_1$. I'll correct it. – K.defaoite Jul 12 '20 at 03:24
  • @K.defaoite Nice catching. No worries, I was barely in the first few steps trying to recreate what you had written. – BrazyOski Jul 12 '20 at 03:25
  • 1
    @BrazyOski Actually nevermind, the $u_1$ function already has a $\frac{\text{constant}}{r}$ and $\text{constant}$ term, so its fully general. – K.defaoite Jul 12 '20 at 03:33
  • 1
    @BrazyOski I added an interactive plot on desmos as well. – K.defaoite Jul 12 '20 at 03:55
  • 1
    @K.defaoite Wow! I'm at a loss for words, thank you so much kind stranger. This has helped a lot. – BrazyOski Jul 12 '20 at 04:17
  • 1
    @BrazyOski No problem! Unfortunately your other question regarding the convolution of the fundamental solution of the heat equation is much more difficult, and I don't think I'll be of much help on that one. – K.defaoite Jul 12 '20 at 04:19
  • @K.defaoite Oh, that one is definitely extremely challenging, but I am somewhat sure that I will have to use Fourier transform at some point instead of outright solving it. But I'll leave it open for now to see if anyone does end up writing a full proof solution. – BrazyOski Jul 12 '20 at 04:23
  • I don’t know what other expansion to use. I have some ideas for numerics but at that point the convolution integral is much cleaner. The method of matched expansions or boundary layers might work but I don’t know this theory well enough to tell you how. At the very least, you could determine your constants by minimising in some norm – Calvin Khor Jul 12 '20 at 05:18
  • Also you'd want to have $b_1=0$ (else the PDE solved is really $\Delta u = \phi + \tilde b_1 \delta_0$), and there seems to be a choice of $b_2$ that makes the solution integrable. If you don't want to enforce integrability or something else, your general solution is missing half of the freedom to add a harmonic polynomial $\mathbf b_3\cdot \mathbf r+b_2$, which you can't see if you restrict to radial solutions – Calvin Khor Jul 12 '20 at 11:32