2

In my work, I encountered a function of the following form,

$$ f(k) = \exp\bigg[-\frac{k^2}{a^2 + k^2}\bigg] $$

In the limit $a\rightarrow 0$, the Fourier transform of this function will be $e^{-1}\delta(x)$
and in the limit $a\rightarrow\infty$ the Fourier transform of this function will be a Gaussian (by expanding the denominator binomially).

Is it possible to calculate the Fourier transform of this function for any real values of $a$?

EDIT 1:

According to Mark's answer, we can rewrite $f(k)$ as follows, $$ f(k) = e^{-1}(e^{a^2/(a^2+k^2)} - 1) + e^{-1} \\ = e^{-1}(e^{(1+(k/a)^2)^{-1}} - 1) + e^{-1} $$ Expanding the first term binomially (for $(k/a) << 1$) will give, $f(k) \approx e^{-(k^2/a^2)}$

For the limit $(k/a) << 1$, does the Fourier Transform of $f(k)$ become a Gaussian function?

user35952
  • 935

2 Answers2

3

We can write the distribution $f(k)$ as $f(k)=g(k)+e^{-1}$, where $g(k)$ is the $L^1$ function given by

$$g(k)=e^{-1}\left(e^{a^2/(a^2+k^2)}-1\right)$$

The Fourier transform of $g$ is given by

$$\begin{align} \mathscr{F}\{g\}(x)&=e^{-1}\int_{-\infty}^\infty \left(e^{a^2/(a^2+k^2)}-1\right)e^{ikx}\,dk\\\\ &=|a|e^{-1}\int_{-\infty}^\infty \left(e^{1/(1+k^2)}-1\right)e^{ik(x|a|)}\,dk \end{align}$$

According to WolframAlpha, there is no result found in terms of standard functions. You could try writing $e^{1/(1+k^2)}-1=\sum_{n=1}^\infty \frac{1}{n!(1+k^2)^n}$ and seeing if that leads to anything useful.



In this section, we find a recursive equation for the function $f_n(t)$ where

$$f_n(t)=\int_{-\infty}^\infty \frac{e^{ikt}}{(1+k^2)^n}\,dk\tag1$$

For $n=1$, it is easy to show that $f_1(t)=\pi e^{-|t|}$. Now, for $n>2$ we differentiate $(1)$ twice to find that

$$f_n''(t)-f_n(t)=-f_{n-1}(t)\tag 2$$

with $f_n'(0)=0$ and $f_n(0)=\int_{-\infty}^\infty \frac1{(1+k^2)^n}\,dk=\pi\frac{(2n-3)!!}{(2n-2)!!} $ (See THIS ANSWER).

The general solution to $(2)$ can be written in terms of the Green function (or Green's function) for the ODE $y''(x)-y(x)=\delta(x)$, along with the prescribed initial conditions. We are now equipped with a recursive algorithm to calculate $f_n(t)$ for any $n$.



EDIT:

As @Metamorphy pointed out, we can express the integral $\int_{-\infty}^\infty \frac{e^{ikx|a|}}{(1+k^2)^n}\,dk$ in terms of the second kind modified Bessel function. Specifically, we have

$$\begin{align}\int_{-\infty}^\infty \frac{e^{ikx|a|}}{(1+k^2)^n}\,dk&=2\int_0^\infty \frac{\cos(k|xa|)}{(1+k^2)^n}\,dk\\\\ &=\frac{\sqrt{\pi}\,|xa|^{n-1/2}}{2^{n-3/2}(n-1)!}K_{n-1/2}(|xa|)\tag3 \end{align}$$

Using $(3)$ we can write

$$\mathscr{F}\{f\}(x)=2\pi e^{-1}\delta(x)+|a|e^{-1}\sqrt{\pi}\sum_{n=1}^\infty \frac{K_{n-1/2}(|xa|)|xa|^{n-1/2}}{2^{n-5/2}n!(n-1)!} $$

Mark Viola
  • 179,405
  • Thanks for this input. Upto the first-order it looks like a symmetric exponential decay function in $x$, am I right? Any ideas on how to integrate higher ($n>1$) orders of the form $(1+k)^{-n}$? – user35952 Nov 22 '22 at 02:59
  • I just obtained some result from Mathematica on that, will post shortly if find a complete solution from the series. – user35952 Nov 22 '22 at 03:01
  • @user35952 You're welcome. My pleasure. I've added a section that develops a recursive algorithm to find the Fourier transform of $\frac1{(1+k^2)^n}$ for any $n$. Let me know if this helps. – Mark Viola Nov 22 '22 at 04:36
  • 1
    Your $f_n$ is (a particular case of) Basset's integral. – metamorphy Nov 22 '22 at 05:49
  • @metamorphy Yes, I had intended to pursue this, but ran out of time. I've edited to include this. Much appreciated. – Mark Viola Nov 22 '22 at 15:20
  • @metamorphy Thanks for integral relation.
    Mark: This looks like a complete answer, I just have one question, can you comment on what happens in the case of $a >> k$? I am inclined to believe that $f(k)$ can be approximated to Gaussian in that case, am I right?
    – user35952 Nov 23 '22 at 03:02
  • @user35952 I don't believe that for "large" $a$, the Fourier transform of the regular part of $f$ becomes asymptotically Gaussian. – Mark Viola Nov 23 '22 at 03:26
  • @MarkViola I have added a small edit in my question, can you kindly have a look? – user35952 Nov 23 '22 at 05:35
  • @user35952 I have had a look. The one thing to understand is that $a$ is fixed while $k$ spans the real numbers. So the expansion you used is only valid for those values of $k$ much less than $a$. It is not valid for $k\ge a$. – Mark Viola Nov 23 '22 at 15:03
  • @MarkViola Now I get it, thanks! It was my bad that I didn't state the conditions clearly. There's one tiny thing that bothers me though, as a physicist I find it inappropriate to have terms like $e^{i k x/|a|}$ because the factors in the exponential wouldn't be dimensionless ($k$ and $a$ have the same dimension, while $x$ has inverse dimension of $k$), How would the last expression really change if I chose not use a rescaled version of Fourier transform as mentioned just before eq. (1)? – user35952 Nov 24 '22 at 05:51
  • @MarkViola Is the integral before eq. (1) supposed to be $ |a|e^{-1}\int_{-\infty}^\infty \left(e^{1/(1+k^2)}-1\right)e^{ik(x|a|)} dk $ rather than $ |a|e^{-1}\int_{-\infty}^\infty \left(e^{1/(1+k^2)}-1\right)e^{ik(x/|a|)} dk $ – user35952 Nov 24 '22 at 08:10
  • @user35952 Yes, you are correct. I've edited accordingly. – Mark Viola Nov 24 '22 at 14:23
2

Since there are several choices of parameters for the Fourier transform, the answer will depend on the version of the Fourier transform adopted. Now, a residue calculation shows that

\begin{align*} \hat{f}(\xi)&=\int_{-\infty}^{\infty} \exp\left(-\frac{k^2}{k^2+a^2}\right)e^{-ik\xi} \, \mathrm{d}k \\ &= 2\pi e^{-1} \Biggl[ \delta(\xi) + e^{-|a\xi|} \sum_{n=0}^{\infty} \sum_{j=0}^{\infty} \frac{(n+2j)!}{(n+j+1)!(n+j)!j!n!} \frac{|a|^{n+1} |\xi|^n}{2^{n+2j+1}} \Biggr] \tag{1} \end{align*}

in distribution sense. I utilized contour integral to compute the integral, but this is essentially what we would obtain from @Mark Viola's approach as well. As a sanity check, we find that $\text{(1)}$ correctly produces the result

$$ \hat{f}(\xi) = \int_{-\infty}^{\infty} \exp\left(-\frac{k^2}{k^2+a^2}\right)e^{-ik\xi} \, \mathrm{d}k \to 2\pi e^{-1} \delta(\xi) $$

as $a \to 0$. To verify $\text{(1)}$, note that the integral is invariant under the substitution $a \mapsto |a|$ and $\xi \mapsto |\xi|$ by the symmetry. So, we may assume that $a > 0$ and $\zeta \geq 0$. Then we decompose the left-hand side as

\begin{align*} \hat{f}(\xi) &= e^{-1} \int_{-\infty}^{\infty} \exp\left(\frac{a^2}{k^2+a^2}\right)e^{-ik\xi} \, \mathrm{d}k \\ &= e^{-1} \Biggl[ 2\pi \delta(\xi) + \int_{-\infty}^{\infty} \biggl( \exp\left(\frac{a^2}{k^2+a^2}\right) - 1 \biggr) e^{-ik\xi} \, \mathrm{d}k \Biggr] \\ &= e^{-1} \Biggl[ 2\pi \delta(\xi) + \sum_{m=1}^{\infty} \frac{1}{m!} \int_{-\infty}^{\infty} \left( \frac{a^2}{k^2 + a^2} \right)^m e^{-ik\xi} \, \mathrm{d}k \Biggr] \\ &= e^{-1} \Biggl[ 2\pi \delta(\xi) + \sum_{m=1}^{\infty} \frac{1}{m!} (-2\pi i) \underset{k=-ia}{\mathrm{Res}} \left( \frac{a^2}{k^2 + a^2} \right)^m e^{-ik\xi} \Biggr]. \end{align*}

Here, the last line follows by employing the lower semicircular contour. Computing and simplifying the residue then leads to $\text{(1)}$.

Sangchul Lee
  • 167,468
  • Can you please elaborate on how to use the residue calculation here? Also, can you comment on the other limit, in which the series should converge to Gaussian? – user35952 Nov 22 '22 at 05:50
  • @user35952, For the asymptotic behavior as $a \to \infty$, I am skeptical about the idea that the transform is asymptotically gaussian. Where did you hear that? – Sangchul Lee Nov 22 '22 at 08:04
  • What about the case when $a >> k$? In that scenario, can't the argument in the exponential can be approximated to Gaussian $e^{(-k^2/a^2)}$ – user35952 Nov 23 '22 at 02:56