Suppose $f:\mathbb{R}^3\to\mathbb{R}$ is an integrable function (I'm open to further restricting $f$, e.g. $f$ can be assumed to be smooth).

For some constants $0<r\leq R$, is there a way to simplify the expression $$ \int_{\lvert y \rvert = R}\int_{\lvert x\rvert =r} f(x+y) \,\mathrm{d}S(x)\mathrm{d}S(y) $$ to an expression not involving surface integrals?

  • $\begingroup$ This is a little too general for me... Maybe you could give some examples? $\endgroup$
    – Yuriy S
    Mar 21, 2019 at 0:54
  • 1
    $\begingroup$ @YuriyS I was thinking the double integral might be equal to something of the form $$C \int_{S} f(x) dx$$ where $S\subseteq \mathbb{R}^3$ and $C>0$ is a constant. I'm not entirely sure if this is possible, but my intuition tells me it might work $\endgroup$
    – Quoka
    Mar 21, 2019 at 1:19
  • $\begingroup$ @Quoka But that would be a surface integral, right? $\endgroup$
    – zhw.
    Mar 28, 2019 at 1:49
  • $\begingroup$ @zhw. Which integral are you referring to? $\endgroup$
    – Quoka
    Mar 28, 2019 at 2:25

3 Answers 3


Let's start by writing the integral as $$ I = \int_{\mathbb{R}^3} \int_{\mathbb{R}^3} f(x+y) \delta(|x|-r) \delta(|y|-R)\, {\rm d}^3x {\rm d}^3y$$ where $\delta$ is Dirac's delta function. Changing the variables to $(z,y) = (x+y,y)$ we get \begin{align} I &= \int_{\mathbb{R}^3} \int_{\mathbb{R}^3} f(z) \delta(|z-y|-r) \delta(|y|-R)\, {\rm d}^3y {\rm d}^3z \\ &= \int_{\mathbb{R}^3} f(z) \Big(\int_{\mathbb{R}^3}\delta(|z-y|-r) \delta(|y|-R)\, {\rm d}^3y \Big) {\rm d}^3z\end{align} For $0\le r \le R$, it can be calculated that To calculate $ \int_{\mathbb{R}^3}\delta(|z-y|-r) \delta(|y|-R)\, {\rm d}^3y $ that let us use spherical coordinates for $y$ where angle $\theta$ is measured from the direction of $z$, so that $$|y| = \rho $$ $$|z-y| = \sqrt{(|z|-\rho\cos\theta)^2 + (\rho\sin\theta)^2} = \sqrt{\rho^2 - 2\rho|z|\cos\theta + |z|^2}$$ We have then \begin{align} & \int_{\mathbb{R}^3}\delta(|z-y|-r) \delta(|y|-R)\, {\rm d}^3y = \\ &\quad = \int_0^\infty {\rm d}\rho\int_0^{\pi}{\rm d}\theta\int_0^{2\pi}{\rm d}\phi\,\rho^2\sin\theta \delta(\sqrt{\rho^2 - 2\rho|z|\cos\theta + |z|^2}-r) \delta(\rho-R) = \\ &\quad = 2\pi R^2 \int_0^{\pi}{\rm d}\theta \sin\theta \;\delta(\sqrt{R^2 - 2R|z|\cos\theta + |z|^2}-r)=^{u:=\sqrt{R^2 - 2R|z|\cos\theta + |z|^2}}\\ &\quad = 2\pi R^2 \int_{|R-|z||}^{R+|z|} {\rm d}u \frac{u}{R|z|} \delta(u-r)= \\ &\quad = \frac{2\pi R}{|z|} \int_{-\infty}^\infty {\rm d}u \,u \,\theta((R+|z|)-u)\theta(u - \big|R-|z|\big|)\delta(u-r)=\\ &\quad = \frac{2\pi R r}{|z|} \,\theta((R+|z|)-r)\theta(r - \big|R-|z|\big|)=\\ &\quad = \frac{2\pi R r}{|z|} \,\theta((R+|z|)-r)\theta(r - (R-|z|))\theta(r - (|z|-R))=\\ &\quad = \frac{2\pi R r}{|z|} \,\theta(|z|+(R-r))\theta(|z| - (R-r))\theta(R+r - |z|) = ^{\text{since }|z|>0, R-r>0}\\ &\quad = \frac{2\pi R r}{| z|} \theta\big(|z|-(R-r)\big)\theta\big(R+r-|z|\big) \end{align} where $\theta$ is Heaviside's theta function.

In the end you get $$ I = 2\pi R r \int_{|z|\in[R-r, R+r]} \frac{f(z)}{|z|} {\rm d}^3z$$

  • 1
    $\begingroup$ I'm having a hard time justifying your use of the Dirac-delta function since it is technically a distribution. Do you have any references explaining why I can use it this way? $\endgroup$
    – Quoka
    Mar 26, 2019 at 0:10
  • $\begingroup$ It is a distribution, but as long as it's integrated with a continuous function, I don't believe there would be any problems with it. I think you can avoid using distributions and get the same result, although in a more complicated way. $\endgroup$ Mar 26, 2019 at 0:30
  • 1
    $\begingroup$ I'm not sure how integration against a distribution is generally defined, could you elaborate on that a bit? Also, how did you obtain the equality $$\int_{\mathbb{R}^3}\delta(|z-y|-r) \delta(|y|-R)\, {\rm d}^3y = \frac{2\pi R r}{| z|} \theta\big(|z|-(R-r)\big)\theta\big(R+r-|z|\big)?$$ $\endgroup$
    – Quoka
    Mar 26, 2019 at 1:27
  • $\begingroup$ Distributions are formally defined as linear functionals on some space of functions. Dirac's delta is defined on the space of continuous functions, in a following way: $$ \langle \delta_a, f\rangle = f(a) $$ Denoting $$ \langle \delta_a, f\rangle = \int_{-\infty}^\infty {\rm d}x\, f(x)\delta(x-a) $$ is just a useful notation, but it has all necessary properties of an integral. $\endgroup$ Mar 26, 2019 at 10:49
  • $\begingroup$ As for how I obtained the equality, I've edited the answer, as there's little space in the comment. $\endgroup$ Mar 26, 2019 at 10:57

Here's a re-derivation of Adam Latosiński's formula without using distributions. Ultimately what needs to be done here is just a change of variables, but I'm going to use probabilistic language because I think it makes the whole thing easier to understand.

Observe that $$ E[f] = \frac{1}{4\pi R^2} \int_{|y|=R} \frac{1}{4\pi r^2}\int_{|x|=r} f(x+y)\,dS(x)\,dS(y) $$ is the expected value of $f(x+y)$ if we randomly choose vectors $x$ and $y$ for which $|x|=r$ and $|y|=R$. We want to change variables to $$ \rho=|x+y|\qquad\text{and}\qquad u = \frac{x+y}{\rho}. $$ Note that $u$ is a unit vector which can be chosen uniformly at random from the unit sphere in $\mathbb{R}^3$. Clearly $\rho \in [R-r,R+r]$, but the probability distribution of $\rho$ is not uniform on this interval, so we have to do some work to figure out how to sample $\rho$.

For this purpose, we may assume by symmetry that $y=(R,0,0)$ and $x$ is uniformly chosen on the sphere $|x|=r$. Now, it is a famous observation of Archimedes that each coordinate of the unit sphere in $\mathbb{R}^3$ is uniformly distributed on the interval $[-1,1]$. It follows that the first coordinate $t$ of $x$ is uniformly distributed on the interval $[-r,r]$. But $$ \rho = \sqrt{(R+t)^2 + (r^2 -t^2)} = \sqrt{R^2+r^2 + 2Rt}\tag*{(*)} $$ so we can sample $\rho$ by choosing $t$ uniformly at random from $[-r,r]$ and then using this formula.

We conclude that $$ E[f] = \frac{1}{2r}\int_{-r}^r \frac{1}{4\pi} \int_{|u|=1} f(\rho u)\,dS(u)\,dt $$ where $\rho$ is the formula (*) above. Note that the $t$ here actually represents the component of $x$ in the direction of $y$, i.e. $t = (x\cdot y)/R$. Combining this with our original formula for $E[f]$, we deduce that $$ \int_{|y|=R} \int_{|x|=r} f(x+y)\,dS(x)\,dS(y) = 2\pi rR^2 \int_{-r}^r \int_{|u|=1} f(\rho u)\,dS(u)\,dt $$ where $\rho$ is given by the formula above. Now observe that $$ d\rho = \frac{d}{dt}\Bigl[\sqrt{R^2+r^2 + 2Rt} \Bigr]\,dt = \frac{R}{\sqrt{R^2+r^2+2Rt}}\,dt = \frac{R}{\rho}\,dt. $$ Substituting $dt = (\rho/R)\,d\rho$, the integral becomes $$ 2\pi rR \int_{R-r}^{R+r} \int_{|u|=1} \rho\,f(\rho u)\,dS(u)\,d\rho $$ If we like, we can change this to a single volume integral over $z=\rho u$. Since $dV(z) = \rho^2\,dS(u)\,d\rho$ and $\rho=|z|$, the integral is $$ 2\pi rR \int_{|z|\in[R-r,R+r]} \frac{f(z)}{|z|}\,dV(z) $$ where $z$ represents $x+y$.


If you work in Fourier space you can write $f(x+y)=\int e^{ik(x+y)}\hat{f}_k dk$ and you can separate the integral to $\int dk \hat{f}_k \int e^{ikx} dS(x) \int e^{iky}dS(y) $.


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.