# When does the Fourier transform "looks like" the original function?

I have just learned about the Fourier transform, and a I have thought of a nice question i cannot answer. The Fourier transform of a Gaussian curve, $$e^{-{\alpha}x^2}$$ is $$\frac{1}{\sqrt{2\alpha}}e^{-\frac{\omega^2}{4\alpha}}$$ which is essentially the same "shape". Another example which we saw is a chain of delta functions, $$\sum_{n=-\infty}^{\infty}{\delta(x-nA)}$$ (for some $$A\in\mathbb{R}$$), which its Fourier transform is $$\frac{\sqrt{2\pi}}{A}\sum_{n=-\infty}^{\infty}{\delta(\omega-\frac{2\pi}{A}n)}$$. Again, the same "shape". What's I'm asking is are there any other known functions which theirs Fourier transform has the same shape?

By "shape" I mean that if $$f(x)=af(bx+\phi)$$ then the Fourier transform of $$f$$ will be $$\hat{a}f(\hat{b}\omega+\hat{\phi})$$ for some $$\hat{a}, \hat{b}, \hat{\phi}$$.

• You'll need to be more precise for "shape". If you mean that they are same up to multiplicative factor, then you're looking for eigenvectors of the Fourier transform which are Hermite polynomials multiplied by a gaussian. If you also grant the liberty of a scaling in the argument, then this changes the result.
– LPZ
Commented Jan 30 at 12:06
• @DominikS Yeah, the Fourier transform of a single delta function isn't a delta function, but I am talking about a chain of them. Commented Jan 30 at 12:15
• @ShaharSella Indeed, you are right! Commented Jan 30 at 12:17
• The Fourier transform of an even function is even, and the Fourier transform of an odd function is odd, but I'm not sure how you define shape. Commented Jan 30 at 21:07
• Another example is $1/\sqrt{|x|}$ whose Fourier transform is proportional to itself ... see more generally mathoverflow.net/questions/12045/…, or math.stackexchange.com/questions/118078/… Commented Feb 4 at 11:12

TL DR:

I will show that the only integrable functions that satisfy the relation are the eigenvectors of the Fourier transform (either the continuos or the discrete one) that are characterized here.

As the Fourier transform is linear we can assume $$a =1$$. Let $$b\ne 0$$ and assume $$\mathcal{F}[f(bx+\phi)]=a'f(b'\xi+\phi')$$ By the involutive property of the Fourier transform $$\mathcal{F}^2f(bx+\phi)=f(-bx-\phi)=\mathcal{F} \left[a'f(b'\xi+\phi')\right]=\mathcal{F}\left[a'f\left(b\left(\frac{b'\xi+\phi'-\phi}{b}\right)+\phi\right)\right]$$ so $$f(-bx-\phi)=a'^2f\left(\frac{b'^2}{b} x +\left(\frac{b'}{b}+1\right) \phi' -\frac{b'}{b} \phi\right)$$ or $$f(x)=a'^2f\left(-\left(\frac{b'}{b}\right)^2 x -\frac{1}{b}\left(\frac{b'}{b}+1\right) \phi' +\frac{1}{b}\left(\frac{b'}{b}-1\right) \phi\right)$$

If $$\frac{b'}{b} \ne \pm 1$$ it is easy to see that the function is continuos iif it is constant. As a $$L^1(\mathbb{R}^n)$$ function has a uniformly continuous Fourier transform, this implies that there are no $$L^1$$ functions that satisfy the above inequalities.

Suppose $$\frac{b'}{b}=1$$. the equation becames: $$f(x)=a'^2 f\left(x+\frac{2}{b} \phi'\right)$$

If $$\phi' \ne 0$$ then, suppose $$a'^2 \ne \pm 1$$ and consider the sequence $$x_0=x$$, $$x_n= x +n\frac{2}{b} \phi'$$. We have $$\lim_{n \to \infty} x_n = \pm \infty$$ (depending from the sign of $$\frac{\phi'}{b}$$ and $$f(x_n)= a'^{2n} f(x_0)$$ By the continuity of $$f$$ it follows that it growth exponentially at infinity i.e. it isn't a tempered distribution, but this is impossible as $$f \in L^1$$. So $$(a')^2= \pm 1$$ that implies that $$f$$ is periodic. In particular, it is an eigenvector of the discrete Fourier transform (Indeed $$a'= \pm 1$$ or $$\pm i$$).

If $$\phi'=0$$. As above we have $$(\mathcal{F}^{-1})^2 f(b' \xi) = f (-b' \xi) = \cdots =a'^2 f\left(bx+ 2 \phi \right)$$ From this you obtain that if $$\phi \ne 0$$ you have $$a'= \pm 1$$ or $$\pm i$$ and $$f$$ is an eigenvector of the discrete Fourier transform.

If $$\phi=\phi'=0$$ you obtain $$\mathcal{F}f(bx)=a'f(bx)$$ So $$f$$ is an eigenvector of the continuos Fourier transform.

If $$\frac{b}{b'}=-1$$, by interchanging the roles of $$\phi$$ and $$\phi'$$ in the previous proof we see that if at least one of the two phases is not zero, than $$f$$ is an eigenvector of the discrete Fourier transform.

If $$\phi=\phi'=0$$ then

$$\mathcal{F} f(bx)= a'f(-bx)$$ By the involutive property of the Fourier transform it follow that $$f$$ is even and so it is an eigenvector of the continuous Fourier transform