The delta function (well, delta is not really a function; it is a distribution) can be defined as a limit of (among many other approximations) the following approximates of the unity:

the heat kernel:

$$\delta(x)=\lim_{\epsilon\rightarrow 0^{+}} \frac{e^{-\frac{x^2}{2\epsilon}}}{\sqrt{2\pi\epsilon}}\qquad (1)$$

or the Poisson kernel $$\delta(x)=\lim_{\epsilon\rightarrow 0^{+}} \frac{1}{\pi} \frac{\epsilon}{\epsilon^2+x^2}\qquad (2)$$

Check out the link


Now, let $f$ be a tempered distribution. Can we interchange the limit with the integral so that

$$\int f(x)\delta(x)dx =\lim_{\epsilon\rightarrow 0^{+}}\int f(x) \frac{e^{-\frac{x^2}{2\epsilon}}}{\sqrt{2\pi\epsilon}}dx \qquad (*)$$

If $(*)$ is true, what is the justification ?

Notice that $(*)$ is true when $f$ is a smooth function with compact support, in which case both sides of $(*)$ give $f(0)$. With tempered distributions, we can not make sense of the pointwise value $f(0)$.

An easier question is when $f$ is a continuous oscillatory function, say $f(x)=e^{\dot{\imath}\alpha x^2}$, $\alpha \in \mathbb{C}$. What can we say then?


I think you're thinking about this the wrong way round. There isn't a function $\delta$ that is obtainable from kernels by pointwise convergence, and satisfies $\int_{\Bbb R}f(x)\delta (x) dx=f(0)$. Indeed, if we define $\delta$ as a pointwise limit, $\delta(x)=0$ for all $x\ne 0$ (while $\delta(0)=\infty$), so $\int_{\Bbb R}f(x)\delta (x) dx=0$. But what we can say is this: given a non-negative function $\phi(x)$ with $\phi(0)>0,\,\int_{\Bbb R}\phi(x) dx=1$, $$\lim_{\epsilon\to 0^+}\int_{\Bbb R}f(x)\tfrac{1}{\epsilon}\phi(\tfrac{x}{\epsilon})dx=\lim_{\epsilon\to 0^+}\int_{\Bbb R}f(\epsilon y)\phi(y)dy.$$If we can move the limit inside the integral, we obtain$$\int_{\Bbb R}\lim_{\epsilon\to 0^+}f(\epsilon y)\phi(y)dy=\int_{\Bbb R}f(0)\phi(y)dy=f(0).$$Note that this argument never tries to move the limit so as to write $$\int_{\Bbb R}f(x)\lim_{\epsilon\to 0^+}\tfrac{1}{\epsilon}\phi(\tfrac{x}{\epsilon})dx,$$which would be $0$ as aforesaid. The definition of the measure we call the Dirac delta is$$\int_a^b f(x)\delta(x)dx=\lim_{\epsilon\to 0^+}\int_a^bf(x)\tfrac{1}{\epsilon}\phi(\tfrac{x}{\epsilon})dx,$$not$$\delta(x)=\lim_{\epsilon\to 0^+}\tfrac{1}{\epsilon}\phi(\tfrac{x}{\epsilon}).$$The Dirac delta is not a function.

  • $\begingroup$ Also, in the pointwise sense, the limits in (1) and (2) do not exist because there is no FUNCTION to which these approximates of the identity converge to. Right ? $\endgroup$ – Medo Nov 10 '18 at 16:49
  • $\begingroup$ @Medo The functions exist, but they don't give the desired integrals; the measure we call $\delta$ is therefore not a function. In particular, no function $D(x)$ exists with the identity $\int_{\Bbb R}f(x)D(x) dx=f(0)$. $\endgroup$ – J.G. Nov 10 '18 at 16:53
  • $\begingroup$ Makes perfect sense. Thank you very much. $\endgroup$ – Medo Nov 10 '18 at 22:04

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.