I've been doing some mathematical modelling in geophysics and the following integration came up. Are there any identities for the Dirac Delta that can be used to simplify this expression:

$$\iint \delta \left( t-\frac1c\left(\sqrt{(x_1-z_1)^2+z_2^2}-\sqrt{z_1^2+z_2^2}\right) \right)dz_1dz_2$$

Edit: To make matters more interesting/challenging (or worse) there are other factors in the integrend, i.e., the full expression is:

$$\iint \frac{\rho(z_1)}{\sqrt{(x_1-z_1)^2+z_2^2}\sqrt{z_1^2+z_2^2}}\delta \left( t-\frac1c\left(\sqrt{(x_1-z_1)^2+z_2^2}-\sqrt{z_1^2+z_2^2}\right) \right)dz_1dz_2$$

where $\rho$ is a some random function. So the question now boils down to: Will Fubini together with

$$\delta(f(x)) = \sum_{i}\frac{\delta(x - x_i)}{|f'(x_i)|}$$

(as pointed out in the answers below) suffice for a simplification still?

  • $\begingroup$ the range of integration for $z_1$ is $(0, \infty)$ while for $z_2$ it is the entire real line $\endgroup$
    – Iconoclast
    Sep 22, 2016 at 17:18

3 Answers 3


Judging from the setup of the problem, I'll assume $x_1, t > 0$ and only study that case.
Let $\lambda = \frac12x_1$ and $\mu = \frac12 ct$. Introduce a bunch of variables:

$$ \begin{cases} (x,y) = (z_1 - \lambda, z_2)\\ r_1 = \sqrt{(x+\lambda)^2 + y^2} = \sqrt{z_1^2+z_2^2}\\ r_2 = \sqrt{(x-\lambda)^2 + y^2} = \sqrt{(z_1-x_1)^2 + z_2^2}\\ (u,v) = \left(\frac{r_1 - r_2}{2},\frac{r_1 + r_2}{2}\right) \end{cases}$$ The actual integral you want to integral can be rewritten as

$$\mathcal{I} = \int_{[-\lambda,\infty)\times(-\infty,\infty)}\rho(x+\lambda)\delta\left(\frac2c(u+\mu)\right)\frac{dxdy}{r_1r_2}$$

Since the integrand is symmetric with respect to the $y$-axis, we can replace the integral by one over the upper $(x,y)$ plane: $$\int_{[-\lambda,\infty)\times(-\infty,\infty)}\cdots\quad\leftrightarrow\quad 2\int_{[-\lambda,\infty)\times [0,\infty)}\cdots$$ Change variables to $(u,v)$, the upper $(x,y)$ plane corresponds to the strip $[-\lambda,\lambda] \times [\lambda,\infty)$ in the $(u,v)$ plane. Notice $$\begin{align} 4uv &= (r_1-r_2)(r_1+r_2) = r_1^2 - r_2^2 = 4\lambda x\\ u^2 + v^2 &= \frac14\left((r_1-r_2)^2+(r_1+r_2)^2\right) = \frac12\left(r_1^2+r_2^2\right) = x^2+y^2+\lambda^2 \end{align} $$ A point $(x,y) \in [-\lambda,\infty) \times [0,\infty)$ is equivalent to corresponding $(u,v)$ belongs to following region: $$\Omega \stackrel{def}{=} \{ (u,v) \in [ -\lambda,\lambda ] \times [ \lambda, \infty ) : uv \ge -\lambda^2 \}$$

In terms of $(u,v)$, the area element has the form $$dx \wedge dy = dx \wedge \frac{dy^2}{2y} = \frac{1}{2\lambda y} d(\lambda x)\wedge d(x^2+y^2+\lambda^2) = \frac{1}{2\lambda y} d(uv)\wedge d(u^2+v^2)\\ = \frac{1}{\lambda y} (udv + vdu)\wedge(udu + vdv) = \frac{v^2-u^2}{\lambda y} du\wedge dv = \frac{r_1r_2}{\lambda y} du\wedge dv $$ Since $$y^2 = u^2 + v^2 - x^2 - \lambda^2 = u^2 + v^2 - \frac{u^2v^2}{\lambda^2} - \lambda^2 = \frac{1}{\lambda^2}(v^2 - \lambda^2)(\lambda^2-u^2)$$ We find $$\frac{dxdy}{r_1r_2} = \frac{dudv}{\lambda y} = \frac{dudv}{\sqrt{(v^2-\lambda^2)(\lambda^2-u^2)}} $$ and hence $$\mathcal{I} = 2\int_\Omega \rho(x+\lambda)\frac{\delta(\frac{2}{c}(u+\mu))dudv}{\sqrt{(v^2-\lambda)^2(\lambda^2-u^2)}}$$

Apply the formula mentioned in other answers, we have $$\delta\left(\frac2c(u+\mu)\right) = \frac{c}{2}\delta(u + \mu)$$

Integrate the delta function over $u$, the $u$ get fixed to $-\mu$.

When $\mu > \lambda$, the line $u = -\mu$ lies outside the strip $[-\lambda,\lambda] \times [\lambda,\infty)$ and $\mathcal{I}$ vanishes.

For $0 < \mu < \lambda$, the integral reduces to

$$\mathcal{I} = \frac{c}{\sqrt{\lambda^2-\mu^2}} \int_\lambda^{\lambda^2/\mu}\frac{\rho\left(\lambda - \frac{\mu v}{\lambda}\right) d v}{\sqrt{v^2-\lambda^2}} = \frac{c}{\sqrt{\lambda^2-\mu^2}}\int_0^{\lambda-\mu} \frac{\rho(z_1)dz_1}{\sqrt{(\lambda-z_1)^2-\mu^2}} $$ The rest of the integral depends on the actual form of $\rho(\cdot)$ and I'll leave that for you.

  • $\begingroup$ Thank you. I'd like to ask: how can I modify the answer if the original integral looks like this $$\iiint \frac{\rho(z_1)}{\sqrt{(x_1-z_1)^2+z_2^2 +z_3^2}\sqrt{z_1^2+z_2^2+z_3^2}}\delta \left( t-\frac1c\left(\sqrt{(x_1-z_1)^2+z_2^2+z_3^2}-\sqrt{z_1^2+z_2^2+z_3^2}\right) \right)dz_1dz_2dz_3$$ $\endgroup$
    – Iconoclast
    Oct 7, 2016 at 12:54
  • $\begingroup$ @Iconoclast That case will be even simpler. If one parametrize $(z_1,z_2,z_3)$ as $(x+\lambda,y\cos\theta,y\sin\theta)$, then $$\frac{dz_1 dz_2 dz_3}{r_1r_2} = \frac{dx dy}{r_1 r_2}yd\theta = \frac{du dv}{\lambda y}yd\theta = \frac{1}{\lambda}du dv d\theta$$ Integrate over $\theta$ and then over $u$, you will get back essentially the same integral with the ugly square root at the denominator replaced by some constant. $\endgroup$ Oct 7, 2016 at 14:04
  • $\begingroup$ One last question, you are assuming that $z_1>0$ right? What happens if that assumption is invalid, i.e., $z_1 \in (-\infty, \infty)$? $\endgroup$
    – Iconoclast
    Oct 7, 2016 at 14:16
  • $\begingroup$ @Iconoclast If $z_1$ can be any value, the integral of delta function will be over the whole strip $[-\lambda,\lambda]\times[\lambda,\infty)$ in $uv$-plane and the final integral over $z_1$ will be over $(-\infty,\lambda - \mu)$ when $\mu < \lambda$ and identically zero when $\mu > \lambda$. $\endgroup$ Oct 7, 2016 at 14:24
  • $\begingroup$ how do you determine that the region $[-\lambda, \infty) \times (-\infty, \infty)$ in the $xy$-plane corresponds to the region $[-\lambda, \lambda] \times [\lambda, \infty)$ in the $uv$-plane? By plotting? $\endgroup$
    – Iconoclast
    Oct 13, 2016 at 16:19


$$ \delta \left [g(x) \right ] = \sum_{k=1}^N \frac{\delta(x-x_k)}{|g'(x_k)|}$$

where $g(x_k)=0$ for $k=1,2,\ldots,N$.

  • $\begingroup$ I guess what i have is $\delta[g(x_1, x_2)]$. How would the above expression change in case of two variables. $\endgroup$
    – Iconoclast
    Sep 22, 2016 at 16:59
  • $\begingroup$ Well, you are still integrating over one of the variables first. Let that variable be $x$ in my notation. $\endgroup$
    – Ron Gordon
    Sep 22, 2016 at 17:01
  • $\begingroup$ Just a question: Shouldn't it be $|g'(x_k)|$? $\endgroup$
    – Laplacian
    Sep 22, 2016 at 17:03
  • 1
    $\begingroup$ @FourierTransform: yes, thanks. $\endgroup$
    – Ron Gordon
    Sep 22, 2016 at 17:03

First of all use Fubinit's theorem to "change" the integration:

$$\int \text{d}z_1 \int \delta \left( t-\frac{\sqrt{(x_1-z_1)^2+z_2^2}-\sqrt{(z_1^2+z_2^2)}}{c} \right)\text{d}z_2$$

Now the integration of a Delta "function" of the type


can be done via the definition of the delta "function" of a function:

$$\delta(f(x)) = \sum_{i}\frac{\delta(x - x_i)}{|f'(x_i)|}$$

Where $i$ runs over all the possible number of roots of the equation

$$f(x_i) = 0$$

Can you proceed?


Despite you have two variables of integration, thanks to Fubini you can first integrate in $z_2$ (or $z_1$) and then in the other one, so first of all, treat one of the two as a constant and find the roots in terms of $z_2$.


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.