6

I would like to derive the Fourier transform of $f(x)=\ln(x^2+a^2)$, where $a\in \mathbb{R}^+$ by making use of the properties:

\begin{equation} \mathcal{F}[f'(x)]=(ik)\hat{f}(k)\\ \mathcal{F}[-ixf(x)]=\hat{f}'(k) \end{equation} For the Fourier transform I use the definition given by:

\begin{equation} \hat{f}(k)=\frac{1}{\sqrt{2\pi}}\int_{-\infty}^{\infty}f(x)e^{-ikx}dx, k \in \mathbb{R} \end{equation} Until now I found out that by taking the derivative of $f$ and finding the Fourier transform of $f'$ I can then use the relation $\mathcal{F}[f'(x)]=(ik)\hat{f}(k)$ and find $\hat{f}$. The derivative of $f$ would be: \begin{equation} f'(x)=\frac{2x}{x^2+a^2} \end{equation} and by considering $g(x)=1/(x^2+a^2)$, I then have: \begin{equation} f'(x)=2xg(x) \end{equation} Now I know that the Fourier transform of $g$ is given by:

\begin{equation} \hat{g}(k)=\frac{1}{a}\sqrt{\frac{\pi}{2}}e^{-a|k|}, a \in \mathbb{R}, k\in \mathbb{R} \end{equation} Now I must find the Fourier transform of $xg(x)$ which would be given by the derivative of $\hat{g}$ right? But how can this possible since $\hat{g}$ has no derivative?

I think I am really close now but I need that extra tip.

Thank you!

Bazinga
  • 1,463
  • 1
    $\hat{g}$ is differentiable everywhere except at $0$. It is however absolutely continuous, so the almost everywhere defined derivative is what you want. – Daniel Fischer Jun 27 '15 at 20:57
  • @DanielFischer Are you sure that I can do that? Because I think that I am not supposed to differentiate in this particular case. Is everything else correct until the point I have reached? – Bazinga Jun 27 '15 at 21:01
  • 1
    I haven't carefully checked whether you made any mistakes so far, and at the moment, I'm not in the mood to do it. Looks okay so far, however. Differentiating $g$ to get the Fourier transform of $f'$ is legitimate here (don't forget that there are some constant factors to take care of). Your problem comes when you divide by $k$ to get the Fourier transform of $f$, because what you get then is not a tempered distribution. And that indicates that something is amiss. – Daniel Fischer Jun 27 '15 at 21:09
  • @DanielFischer So, in the end I can claim that the Fourier transform of $f$ is everywhere defined except for $k=0$ and that would hold, right? – Bazinga Jun 27 '15 at 21:15
  • 1
    Ummm. $f$ is not the kind of function where you can use the integral definition of the Fourier transform. However, $f$ defines a tempered distribution, and as such it has a Fourier transform that again is a tempered distribution. Some tempered distributions are given by integration against reasonable functions, and for those you can have a notion of "almost everywhere defined". But the manipulations of pointwise (a.e.) defined functions seem to give something that is not a tempered distribution, so something is not quite right. – Daniel Fischer Jun 27 '15 at 21:22
  • 2
    @DanielFischer You are right, I do know that $f$ defines a tempered distribution (even though I cannot prove it). And I also know that the Fourier transform of an $f \in \mathcal{S}$ has a Fourier transform that is also defined in the same space. But the Fourier transform here is (finally) given by: \begin{equation} \hat{f}(k)=-\sqrt{2\pi}\frac{e^{-|ak|}}{|k|}\end{equation} which I do not know if it is a "function" of the same space. Anyway, thank you for your assistance! – Bazinga Jun 27 '15 at 21:27
  • 1
    The problem is the $\frac{1}{\lvert k\rvert}$. That's a non-integrable singularity, and since there is no sign change, we can't even go and interpret it as a principal value integral. – Daniel Fischer Jun 27 '15 at 21:35
  • 1
    Minor suggestion to the question (v2): Replace the condition $a\in\mathbb{R}$ with $a>0$ in two places. – Qmechanic Jun 28 '15 at 12:03
  • @DanielFischer Hi Daniel. I hope all is well. If we use a regularized distribution, we can find a FT for $\log(x^2+a^2)$. While the FT is not a "standard" tempered distribution, it is a distribution. I've posted a solution using this approach. As always, I'd appreciate your thoughts. – Mark Viola Jul 16 '21 at 18:10

3 Answers3

3

If $\left| x \right|^{-1}$ is the distribution defined as $$\left( \left| x \right|^{-1}, \phi \right) = \int_{\left| x \right| < 1} \frac {\phi(x) - \phi(0)} {\left| x \right|} dx + \int_{\left| x \right| > 1} \frac {\phi(x)} {\left| x \right|} dx,$$ then $${\mathcal F}\!\left[ \ln\left( x^2 +a^2 \right) \right] = \frac 1 {\sqrt{2 \pi}} \int_{-\infty}^{\infty} \ln\left( x^2 +a^2 \right) e^{-i k x} dx = -\sqrt{2 \pi} \left( \frac { e^{-a \left| k \right|}} {\left| k \right|} + 2 \gamma \delta(k) \right),$$ where $\gamma$ is Euler's constant. Probably the easiest way to prove it is to compute the inverse transform directly from the definition of $\left| x \right|^{-1}$.

Maxim
  • 10,764
  • Hi "Maxim." Under the $\lambda_1$ distribution, the scale factor $-\sqrt{2\pi}2\gamma$ on $\delta(k)$ is unfortunately incorrect. The scale factor should be instead $2\sqrt{2\pi}(\log(a)+\text{Ei}(-a))$. As $a\to 0$, this approaches $-2\sqrt{2\pi}\gamma$. I've posted a solution herein and provided a section to show that the Fourier Transform has as its inverse $\log(x^2+a^2)$. – Mark Viola Jul 16 '21 at 18:08
  • There is an issue with this answer, but the issue is completely different. Namely, $e^{-a |k|}/|k|$ is a sloppy notation because I've only defined $|k|^{-1}$ and, in general, a distribution cannot be multiplied by $e^{-a |k|}$. A cleaner way to write the result is $$\mathcal F[\ln(x^2 + a^2)] = -\sqrt {2 \pi} \left( |k|^{-1} + 2 \gamma \delta(k) - \frac {1 - e^{-a |k|}} {|k|} \right),$$ where the last term is an ordinary function (or a regular distribution). This is unambiguous. The distribution in your answer is identical to this one. – Maxim Jul 17 '21 at 19:32
  • Yes, the distribution in your comment and my answer(for $\nu =1$) are identical. And it's crystal clear. Again, much appreciate your time. – Mark Viola Jul 18 '21 at 00:02
  • Any $\nu > 0$ gives the same distribution $\mathcal F[\ln(x^2 + a^2)]$. – Maxim Jul 18 '21 at 00:12
  • Yes that is true. $C\nu$ and $\left(\cdots\right)\nu$ depend on $\nu$ such that the sum of $C\nu(a) \delta(k)-\sqrt{2\pi}\left(\frac{e^{-|la|}}{|k|}\right)\nu$ is independent of $\nu$. – Mark Viola Jul 18 '21 at 00:38
  • What is the defn of the distn $1/|x|$? – paul garrett Jul 31 '21 at 04:37
  • @paulgarrett If you mean the distribution $|x|^{-1}$ that I'm using, it's defined by the first formula in the answer (a tempered distribution is defined by its action on an arbitrary Schwartz function; it can be verified that the formula gives a linear continuous functional, so the definition is consistent). We have to specify how we define $|x|^{-1}$ because the ordinary function $1/|x|$ has many different regularizations (the FT of a distribution is, of course, unique). – Maxim Jul 31 '21 at 06:16
  • @paulgarrett The difference between Max's solution and mine posted herein (and one posted HERE) is that I am using the distribution $\left(\frac{e^{-|ak|}}{|k|}\right)\nu$ (as defined in my post) while Max is using $\left(\frac1{|k|}\right)\nu+\frac{e^{-|ka|}-1}{|k|}$ with $\nu=1$.Note that the differences between our solutions result in different coefficients on the Dirac Delta contributions – Mark Viola Jul 31 '21 at 20:23
1

If $1/|x|$ means even and homog degree $-1$, the only such distn is $\delta$. Attempting to alter this fact by arbitrary truncations can indeed produce several different seeming answers. That’s not good.

I’d prefer to continue the approach in the question: up to the constant there, the desired Fourier transform is the derivative of $e^{-|ak|}$ in $k$. Without loss of generality, $a>0$. Then the deriv is $-sgn(x)ae^{-|ak|}-2a\delta$. Then mult by the ${1\over a}\sqrt{\pi/2}$.

paul garrett
  • 52,465
0

I thought it might be instructive to present an approach that develops the Fourier Transform directly and is distinct and complements the approach I adopted in THIS ANSWER. To that end, we now proceed.


Let $f(x)=\log(x^2+a^2)$ and let $F(k)$ denote its Fourier Transform. Then, we write

$$F(k)=\mathscr{F}\{f\}(k)\tag 1$$

where $(1)$ is interpreted as a Tempered Distribution. That is, for any $\phi \in \mathbb{S}$, we can write for any $\nu>0$

$$\begin{align} \langle \mathscr{F}\{f\}, \phi\rangle &=\langle f, \mathscr{F}\{\phi\}\rangle\\\\ &=\frac1{\sqrt{2\pi}}\int_{-\infty}^\infty \log(x^2+a^2)\int_{-\infty}^\infty \phi(k)e^{-ikx}\,dk\,dx\\\\ &=\frac2{\sqrt{2\pi}}\int_0^\infty \log(x^2+a^2)\int_{-\infty}^\infty \phi(k)\cos(kx)\,dk\,dx\\\\ &=\frac4{\sqrt{2\pi}}\phi(0)\int_0^\infty \frac{\sin(\nu x)}{x}\,\log(x^2+a^2)\,dx\\\\ &+\frac2{\sqrt{2\pi}}\int_0^\infty \log(x^2+a^2) \left(\int_{-\infty}^\infty (\phi(k)-\phi(0)\xi_{[-\nu,\nu]}(k))\cos(kx)\,dk\right)\,dx\\\\ &=C_\nu(a) \phi(0)\tag2\\\\ &+\frac2{\sqrt{2\pi}}\lim_{L\to \infty } \int_{-\infty}^\infty \frac{\phi(k)-\phi(0)\xi_{[-\nu,\nu]}(k)}k\left(\log(L^2+a^2)\sin(kL)\\-\int_0^L \frac{2x\sin(kx)}{x^2+a^2}\,dx\right)\,dk\\\\ &=C_\nu(a) \phi(0)-\sqrt{2\pi}\int_{-\infty}^\infty \left(\phi(k)-\phi(0)\xi_{[-\nu,\nu]}(k)\right)\frac{e^{-|ka|}}{|k|}\,dk\tag3 \end{align}$$

where the constant $C_\nu(a)$ is given by

$$\begin{align} C_\nu(a)&=-2\sqrt{2\pi} \log(\nu)+\frac4{\sqrt{2\pi}}\int_0^\infty \frac{\sin(x)}{x}\,\log(x^2+\nu^2a^2)\,dx\\\\ &=2\sqrt{2\pi} (\log(a)-\text{Ei}(-\nu a)) \end{align}$$



NOTES:

In arriving at $(2)$, used Fubini's theorem to justify interchanging integral $\int_0^L \,dx$ with the integral $\int_{-\infty}^\infty \,dk$.

In arriving at $(3)$, we used integration by parts to show that the limit of the term involving $\log(L^2+a^2)$ is $0$ and we appealed the the Dominated Convergence Theorem to justify interchanging the limit with the integration over $k$ for the second term on the right-hand side of $(3)$.

For $\nu=1$ and $a\to 0$, $C_1(a\to 0)=-2\sqrt{2\pi}\gamma$, which agrees with the results in the previously referenced answer (modulo the scale factor $1/\sqrt{2\pi}$ here from the definition of the Fourier Transform).



From $(3)$, we deduce the Fourier Transform of $\psi$ in distribution

$$\bbox[5px,border:2px solid #C0A000] {\mathscr{F}\{f\}(k)=C_\nu(a) \delta(k)-\sqrt{2\pi}\left(\frac{e^{-|ka|}}{|k|}\right)_\nu}\tag4$$

where the distribution $\left(\frac{e^{-|ka|}}{|k|}\right)_\nu$ is defined by its action on any $\phi\in \mathbb{S}$

$$\int_{-\infty}^\infty \left(\frac{e^{-|ka|}}{|k|}\right)_\nu\phi(k)\,dk=\int_{-\infty}^\infty \left(\phi(k)-\phi(0)\xi_{[-\nu,\nu]}(k)\right)\frac{e^{-|ka|}}{|k|}\,dk$$



SOLUTION VERIFICATION:

In this section, we take the inverse Fourier Transform of the distribution in $(6)$ and show that it is indeed equal to $\log(x^2+a^2)$. To begin, we write for any $\phi\in \mathbb{S}$

$$\begin{align} \langle \mathscr{F}^{-1}\{F\},\phi\rangle&=\langle F,\mathscr{F}^{-1}\{\phi\}\rangle \\\\ &=C_\nu(a)\mathscr{F}^{-1}\{\phi\}(0)-\int_{-\infty}^\infty \frac{e^{-|ka|}}{|k|}\int_{-\infty}^\infty \phi(x)\left(e^{ikx}-\xi_{[-\nu,\nu]}(k)\right)\,dx\,dk\\\\ &=C_\nu(a)\mathscr{F}^{-1}\{\phi\}(0)\\\\ &-2\int_{-\infty}^\infty \phi(x)\int_{0}^\infty \frac{e^{-ka}}{k}\left(\cos(kx)-\xi_{[-\nu,\nu]}(k)\right)\,dk\,dx\tag5 \end{align}$$

Using "Feynman's Trick," we find that the inner integral on the right-hand side of $(5)$ is given by

$$\begin{align} \int_{0}^\infty \frac{e^{-ka}}{k}\left(\cos(kx)-\xi_{[-\nu,\nu]}(k)\right)\,dk&=\int_\infty^a \left(\frac{1-e^{-\nu a'}}{a'}-\frac{a'}{x^2+a'^2}\right)\,da'\\\\ &=\log(a)+\int_{\nu a}^\infty \frac{e^{-a' }}{a'}\,da'-\frac12\log(x^2+a^2))\\\\ &=-\frac12\log(x^2+a^2)+\log(a)+\text{Ei}(-\nu a)\tag6 \end{align}$$

Finally, substituting $(6)$ in $(5)$ yields the anticipated result that

$$\langle \mathscr{F}^{-1}\{F\},\phi\rangle=\int_{-\infty}^\infty \phi(x)\log(x^2+a^2)\,dx$$

from which we conclude that

$$f(x)=\mathscr{F}^{-1}\{F\}(x)=\log(x^2+a^2)$$

Mark Viola
  • 179,405