One of the requirements for the existence of Fourier transform of $f(x)$ is that:

$\int_{-\infty}^{\infty} |f(x)| dx $ exists.

However, the table says that the Fourier transform of constant functions (\emph{i.e.}, $f(x)=1$) do exist and it is $\delta(k)$ although $\int_{-\infty}^{\infty} 1 dx = \infty$ .

Could anyone can help me to understand this? Thanks in advance.

  • $\begingroup$ It depends upon what you mean by "existence". The Dirac delta $\delta(k)$ must be defined by convention, because of course its value at $k=0$ is not finite. $\endgroup$ Mar 19 '14 at 18:26
  • $\begingroup$ Dirichlet conditions (including integrability) are sufficient, not necessary. $\endgroup$
    – user48547
    Jan 2 '15 at 10:20

The Fourier transform defined by an ordinary (Riemann or) Lebesgue integral only exists when $f \in L^1$.

It is however possible to extend the definition to tempered distributions (for example, every locally integrable function that "doesn't grow too fast" can be identified with a tempered distribution). The Fourier transform of such a thing is not in general a function though, as witnessed by your example.


I think it might be good to think the dirac as the limit of a some know function. For example when we take a zero mean Gaussian density $f(x;0, \sigma)$ with variance $\sigma^2$, we know that when $\lim_{\sigma\rightarrow 0}f(x;0, \sigma)=\delta(x)$. This is one type of the definition of the dirac delta function.

When we take the fourier transform of this function $$\int_{-\infty}^{\infty}f(x)e^{itx}\mbox{d}x=e^{i\mu t}e^{-\frac{1}{2}(\sigma t)^2}$$ and when $\sigma\rightarrow\infty$ we have $$|e^{i\mu t}e^{-\frac{1}{2}(\sigma t)^2}|\rightarrow 1$$

Another way would be to take a rectangular function on $[-t,t]$ at the frequency domain $t$ and let $t\rightarrow\infty$. At the time domain $x$ one would get a sinc function with a single main lobe. The number of zero crossings will increase, the power of side lobes will decrease and eventually go to zero, what remains will be a dirac delta function.

Mathematically, if absolute summability is not satisfied we also get something which is strangely defined; something which has an infinite power at frequency $t=0$, from Perseval's energy preservation rule.


Approach $\boldsymbol{1}$: Approximating $1$ by $e^{-\lambda x^2}$

Note that as $\lambda\to0^+$, $e^{-\lambda x^2}\to1$. Furthermore, we can explicitly compute the Fourier transform: $$ \begin{align} \int_{\mathbb{R}}e^{-\lambda x^2}e^{-2\pi ix\xi}\,\mathrm{d}x &=e^{-\pi^2\xi^2/\lambda}\int_{\mathbb{R}}e^{-\lambda(x+\pi i\xi/\lambda)^2}\,\mathrm{d}x\\ &=e^{-\pi^2\xi^2/\lambda}\int_{\mathbb{R}}e^{-\lambda x^2}\,\mathrm{d}x\\ &=\sqrt{\tfrac\pi\lambda}e^{-\pi^2\xi^2/\lambda}\tag1 \end{align} $$ For all $\lambda\gt0$, $$ \int_{\mathbb{R}}\sqrt{\tfrac\pi\lambda}e^{-\pi^2\xi^2/\lambda}\,\mathrm{d}\xi=1\tag2 $$ and as $\lambda\to0^+$, the graph of $\sqrt{\tfrac\pi\lambda}e^{-\pi^2\xi^2/\lambda}$ contracts horizontally by a factor of $\sqrt\lambda$ and expands vertically by a factor of $\frac1{\sqrt\lambda}$. That is, as $\lambda\to0^+$, $\sqrt{\tfrac\pi\lambda}e^{-\pi^2\xi^2/\lambda}$ approximates the dirac delta.

Thus, as $\lambda\to0^+$, $e^{-\lambda x^2}\to1$ and its Fourier Transform is $\sqrt{\tfrac\pi\lambda}e^{-\pi^2\xi^2/\lambda}\to\delta(\xi)$.

Approach $\boldsymbol{2}$: Compare action via the Convolution Theorem

By the Convolution Theorem $$ \widehat{f\cdot1}(\xi)=\widehat{f}\ast\widehat{1}(\xi) $$ Since multiplication by $1$ leaves $f$ alone, convolution by $\widehat{1}$ must leave $\widehat{f}$ alone. That is, we must have $\widehat{1}=\delta$.


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.