Dirac delta function have this property:
\begin{equation} \delta(f(x))=\textstyle \sum_i\frac{\delta(x-a_i)}{\lvert f^\prime(a_i)\rvert}. \end{equation} And its derivation is: \begin{eqnarray} \int_{-\infty}^{\infty}g(x)\delta(f(x))&=&\sum_i\int_{a_i-\epsilon}^{a_i+\epsilon}g(x)\delta(f(x)),\qquad f(a_i)=0\cr &=&\sum_i\int_{f(a_i-\epsilon)}^{f(a_i+\epsilon)}g(f^{-1}(y))\delta(y)\frac{1}{\lvert f^\prime(f^{-1}(y))\rvert}dy,\qquad x=f^{-1}(y)\cr &=&\sum_i\frac{a_i}{\lvert f^\prime(a_i)\rvert}.\cr &&\therefore\ \delta(f(x))=\sum_i\frac{\delta(x-a_i)}{\lvert f^\prime(a_i)\rvert}. \end{eqnarray} Here are questions:

  1. How about if $f^\prime(a_i)$ is 0?
  2. I think $\int_{-\infty}^{\infty}g(x)\delta(f(x))=\sum_i g(a_i)$ because Dirac delta function $\delta(f(x))$ in LHS is zero at $a_i$, and so $\int\delta(x)f(x)\ dx$ is just $f(x)$. What's wrong with this?
  1. Firstly, OP is essentially asking the following question.

    What is $$I~=~\int_{\mathbb{R}}\! dx~g(x)~\delta(f(x)) \tag{A}$$ if the function $f$ only has isolated zeros $a_i$ with multiplicity $n_i$, where the index $i$ labels the zeros?

    That is a good question. Below we will take a heuristic approach, and leave it to the reader to insert adequate assumptions to justify the various steps in the calculation. The answer is then

    $$I~=~\sum_i \frac{(n_i-1)!}{|f^{(n_i)}(a_i)|} \lim_{y\to 0^+}\frac{g(a_i+\sqrt[n_i]{y})+g(a_i-\sqrt[n_i]{y})}{2y^{1-\frac{1}{n_i}}}. \tag{B}$$

    Obviously, the limits in eq. (B) might not exist.

    Sketched proof of formula (B). Since each zero is isolated, it is enough to just consider one zero $a$. By translational symmetry of the Lebesgue measure, we can assume that the zero is at the origin $a=0$. It follows that it is enough to consider $$f(x)~=~Ax^n, \tag{C}$$ where $$A~=~\frac{f^{(n)}(0)}{n!}~\neq~ 0.\tag{D}$$ We can throw the constant $A$ outside of the Dirac delta distribution as $\frac{1}{|A|}$. Then $$I~=~\frac{1}{|A|}\int_{\mathbb{R}}\! dx~g(x)\delta(x^n) ~=~\frac{1}{|A|}\int_{\mathbb{R}_+}\! dx~[g(x)+g(-x)]\delta(x^n)$$ $$~\stackrel{y=x^n}{=}~\frac{1}{|A|}\int_{\mathbb{R}_+}\! \frac{dy}{ny^{1-\frac{1}{n}}}~[g(\sqrt[n]{y})+g(-\sqrt[n]{y})]\delta(y) ~=~\lim_{y\to 0^+}\frac{g(\sqrt[n]{y})+g(-\sqrt[n]{y})}{2n|A|y^{1-\frac{1}{n}}}, \tag{E}$$ which reproduces the familiar result $$I~=~\frac{g(0)}{|f^{\prime}(0)|}\quad\text{for}\quad n~=~1.\tag{F}$$ $\Box$

  2. Secondly, OP is asking what is wrong with his second calculation? For starters, it seems that OP is forgetting the Jacobian factor under integration by substitution.


If $f'(a_i) = 0$ then the substitution $x = f^{-1}(y)$ is invalid because $f^{-1}$ is not differentiable in $a_i$.

  • $\begingroup$ So in that case, above property is not available? How about this case that some $f^\prime(a_i)=0$ and the others is not? $\endgroup$ – Orient Oct 26 '17 at 9:13
  • $\begingroup$ @bowstring If at any point both $f$ and $f'$ are $0$ then you run into this problem. Of course you don't care if $f'(x)=0$ but $f(x)\neq 0$, since the Dirac Delta ignores those points. $\endgroup$ – Hyperplane Oct 26 '17 at 9:36

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.