0

The following holds for Dirac delta function where $f(x)$ has discrete zeros at $a_i$

$$\delta\big(f(x)\big) = \sum_{i}\frac{\delta(x-a_{i})}{\left|{\frac{df}{dx}(a_{i})}\right|}$$

See: Dirac Delta Function of a Function

My question is:

But what about $f(x)$ where the zeros are not discrete, does

$\delta\big(f(x)\big) = \frac{\delta(x)}{\left|f'(x) \right|}$ , where $f'$ is not zero,

hold? If yes how is it derived?

(I would also appreciate references for the non discrete zeros case .)

--------------------edit--------------------

Re. defining $\delta(f(x))$ when the zeros are not discrete :

Here is a reference that discusses $\delta(P)$ where $P$ is a hypersurface (eg. spherical shell).

I.M. Gelfand et al "Generalized Functions, volume 1", Academic Press 1964

rough highlights:

p.209: hypersurface definition $P(x1,...xn) = 0$

p.210: require no singular points on $P=0$

p.220: definition of the form $\omega$

(?analogous to Euclidian differential combined with the Jacobian ?)

p.222: $\delta(P)$ definition:

$(\delta(P),\phi )= \int \delta(P)\phi dx)= \int_{P=0} \phi(x)\omega $

So my limited understanding of this is that if the argument of the delta is the equation

of a surface, the integral of the delta with a test function $\phi$ is the surface

integral of $\phi$ over the surface $P$. It seems to me like this could be generalized to volumes, ect..

But I do not see how any of this could lead to $\delta\big(f(x)\big) = \frac{\delta(x)}{\left|f'(x) \right|}$

user45664
  • 807
  • 1
    If $f$ has a continuum of zeros then $df/dx=0$ and the ratio is not defined. – md2perpe May 14 '19 at 19:04
  • @md2perpe I have edited the question to maybe avoid this problem. – user45664 May 15 '19 at 22:00
  • 1
    To my understanding there is no sensible way to define $\delta(f(x))$ (for $f : \mathbb{R} \to \mathbb{R}$) when the zero set of $f$ is not discrete. (The situation is sort of analogous when the domain of $f$ has higher dimension.) – Ian May 15 '19 at 22:03
  • @Ian Thanks; what is meant by (The situation is sort of analogous when the domain of f has higher dimension.)? – user45664 May 16 '19 at 01:30
  • In higher dimensions you need the zero set of $f$ to have positive codimension, so that you can define $\delta(f(x))$ by using the coarea formula. In one dimension this means the zero set needs to be zero dimensional, i.e. discrete. – Ian May 16 '19 at 01:43
  • @Ian Please see my edit re. non discrete zeros in $f(x)$ – user45664 May 16 '19 at 18:56
  • What you're talking about now is resolved by the coarea formula. The original formula at the top is only the one-dimensional case of the coarea formula. – Ian May 16 '19 at 19:10
  • Does that mean $\delta\big(f(x)\big) = \frac{\delta(x)}{\left|f'(x) \right|}$ is, or could be, correct? – user45664 May 16 '19 at 19:15
  • Not in the multidimensional case like is being discussed in your edit. – Ian May 16 '19 at 19:15

1 Answers1

1

You need to define $\delta(f(x))$. The obvious definition is for $f$ continuous and $\phi$ smooth supported on some finite interval $[a,b]$

$$\int_{-\infty}^\infty \delta(f(x))\phi(x)dx = \lim_{n \to \infty}\frac{n}{2} \int_{-\infty}^\infty 1_{|f(x)| < 1/n}\ \phi(x)dx$$ whenever the limit converges

That is $\delta(f(x)) = \lim_{n \to \infty} \frac{n}{2}1_{|f(x)| < 1/n}$ with the limit taken in the sense of distributions, assuming it does converge in the sense of distributions.

If $f$ is smooth and $Z(f) = \{c\in \Bbb{R}, f(c) = 0\}$ is discrete and for $c \in Z(f),f' \ne 0$ then the sequence does converge in the sense of distributions, to $$\delta(f(x))= \sum_{c \in Z(f)} \frac{1}{|f'(c)|} \delta(x-c)$$

Note $f$ doesn't need to be smooth as $\delta(f(x)) = \delta(|f(x)|)$.

Let $g(x) = 1 - |x| 1_{|x| < 1}$ and $$f_K(x) = \min_{k=1}^K \min_{m=1}^{2^{k-1}} g( 2^{2^k} (x-\frac{2m-1}{2^k})), \qquad f(x) = \lim_{K \to \infty} f_K(x)$$ Where $\min_{j=1}^l u_j(x) $ means the result of the sequence $v_1(x)=1, v_{j+1}(x) = \min(v_j(x),u_j(x))$.

Then $$\delta(f_K(x)) = \sum_{k=1}^k\sum_{m=1}^{2^{k-1}} \frac{1}{2^{2^k}} \delta(x-\frac{2m-1}{2^k}), \qquad \delta(f(x)) = \sum_{k=1}^\infty\sum_{m=1}^{2^{k-1}} \frac{1}{2^{2^k}} \delta(x-\frac{2m-1}{2^k})$$ which is well-defined even if $f$ has infinitely many zeros.

reuns
  • 77,999
  • 1
    Another way to define $\delta(f(x))$ is through variable change: $$ \int \delta(f(x)) , \varphi(x) , dx = { y = f(x) } = \int \delta(y) , \varphi(f^{-1}(y)) , \frac{dx}{f'(f^{-1}(y))} = \varphi(f^{-1}(0)) , \frac{dx}{f'(f^{-1}(0))} $$ This was just a sketch; a more thorough analysis needs to be done to handle functions with several zeroes. – md2perpe May 16 '19 at 20:04
  • @reuns Are the zeros then still required to be discrete? Even though the number is not limited. – user45664 May 16 '19 at 20:16
  • Being infinite and contained in $[0,1]$, $Z(f)$ is not discrete @user45664 – reuns May 16 '19 at 20:22
  • @md2perpe Could this be adapted to apply to $f$ where the zeros were not discrete? For example, the equation for a 3D sphere--that is, a single layer over a sphere $S$, $\delta(S)$ – user45664 May 16 '19 at 21:08
  • @user45664 For $\phi$ smooth compactly supported on $\Bbb{R}^2$ let $\int_{\Bbb{R}^2}\delta(x^2+y^2-1)\phi(x,y) dxdy = \frac12 \int_0^{2\pi} \phi( \cos(t),\sin(t))dt$, $ \int_{\Bbb{R}^2}\delta(x,y)\phi(x,y) dxdy = \phi(0,0)$. Do you see how to justify those definitions ? – reuns May 16 '19 at 21:28