5

From this question I know that for every $a\in\mathbb{R}$ there exists a unique radial positive homogeneous tempered distribution of degree $a$, up to a multiplicative constant.

Also, it is easily proven that if $g$ is a radial positive homogeneous tempered distribution of degree $a$, then its Fourier transform $\mathcal{F}(g)$ is a radial positive homogeneous tempered distribution of degree $-1-a$.

Now, it is clear that if $n\in\{1,2,...\}$ and $a\in\left(n,n+1\right)$, the following: $$f\mapsto\int_\mathbb{R} \frac{f(t)-\sum_{k=0}^{n-1}f^{(k)}(0)t^k}{|t|^{a}}\operatorname{d}t$$ is a non-null radial positive homogeneous tempered distribution of degree $-a$, while: $$x\mapsto|x|^{-1+a}$$ is locally integrable and of moderate growth, so it represents a tempered distribution that is also non-null, radial and positive homogeneous of degree $-1+a$ and then its Fourier transform is a radial positive homogeneous distribution of degree $-a$. So, if $\mathcal{S}$ is the space of Schwartz test functions we have that:

$$\forall n\in\{1,2,...\}, \forall a\in(n,n+1),\exists c_a\in\mathbb{C},\forall f\in\mathcal{S},\\ \left(f\mapsto\int_\mathbb{R} \frac{f(t)-\sum_{k=0}^{n-1}f^{(k)}(0)t^k}{|t|^{a}}\operatorname{d}t\right)=c_a\mathcal{F}_x(|x|^{-1+a}),$$ How can we explicitly calculate $c_a$?

Bob
  • 5,995

1 Answers1

4

The following is from (informal) lecture notes of a course I taught on distributions. The calculation corresponds to the easy case $0<a<1$ which was left out from the OP's question. I would try to adapt the proof below to see what happens for $1<a<2$.

The notation below is as follows: For a Schwartz function $f$ a multiindex $\beta$ and an integer $k\ge 0$, we use the seminorm $||f||_{\beta,k}=\sup_{x\in\mathbb{R}^d} \langle x\rangle^k |\partial^{\beta} f(x)|$. Here $\langle x\rangle=\sqrt{1+\sum_{1\le i\le d} x_i^2}$ is the so-called "Japanese bracket". The fourier transform is $\widehat{f}(\xi)=\mathcal{F}[f](\xi)=\int_{\mathbb{R}^d} e^{-i\xi x}f(x)\ d^dx$.


Example: Let $0<\alpha <d$. Then

$``\displaystyle \int _{\mathbb{R}^d}e^{-i\xi x}\frac{1}{|x|^\alpha}\ d^d x=\frac{\Gamma(\frac{d-\alpha}{2})}{\Gamma(\frac{d}{2})}2^{d-\alpha} \pi^{\frac{d}{2}}\frac{1}{|\xi|^{d-\alpha}}" $

Proof:

Note that L.H.S. does not make sense as Lebesgue integral as $$ \int_{\mathbb{R}^d}|x|^{-\alpha} d^dx=\infty\ . $$ But we will work in the language of distributions. Define

$\displaystyle \phi(x)= \begin{cases} \frac{1}{|x|^\alpha} & x\neq 0,\\ 0 & x=0 \end{cases}$

Define the distribution associated to $\phi$, say $T$, i.e. given $f\in S'$, we have $\displaystyle \langle T,f \rangle =\int _{\mathbb{R}^d}\phi(x) \, f(x) \, d^dx$.

$$ \int |\phi f|=\int \frac{1}{|x|^\alpha}\frac{\langle x \rangle^{d+1}}{\langle x \rangle^{d+1}}|f(x)|d^dx\leq ||f||_{0,d+1}\int_{\mathbb{R}^d}\frac{d^dx}{|x|^\alpha \langle x \rangle^{d+1}}<\infty $$

Thus $T$ is well-defined and continuous. So, $T\in S'$.

$\displaystyle \forall f\in S'$,

\begin{align*} \langle \widehat{T},f \rangle := \langle T,\widehat{f} \rangle &=\int_{\mathbb{R}^d\setminus\{0\}}d^dx\frac{1}{|x|^\alpha}\widehat{f}(x)\\ &=\int_{x\neq0}d^dx\Big(\frac{1}{\Gamma(\frac{\alpha}{2})}\int_0^\infty\frac{dt}{t}t^{\frac{\alpha}{2}}e^{-t|x|^2}\Big)\widehat{f}(x)\\ & \overset{\text{Fubini}}{=}\frac{1}{\Gamma(\frac{\alpha}{2})}\int_0^\infty \frac{dt}{t}t^{\frac{\alpha}{2}}\int_{\mathbb{R}^d\setminus\{0\}}d^dx\ \overline{e^{-t|x|^2}}\widehat{f}(x)\quad (*) \end{align*}

For $a>0$,

\begin{align*} \mathcal{F}[\xi \to e^{-a\xi^2}](x) &=\int_{\mathbb{R}^d}e^{-ix\xi}e^{-a\xi^2}d^d\xi\\ &=(2a)^{-\frac{d}{2}}\int e^{-\frac{\eta^2}{2}-i\frac{x}{\sqrt{2a}\eta}}d^d\eta , ~~\eta=\sqrt{2a}\xi\\ &=(\frac{\pi}{a})^{\frac{d}{2}}e^{-\frac{x^2}{4a}} \end{align*}

Take $\frac{1}{4a}=t$; $\displaystyle e^{-t|x|^2}=\mathcal{F}[\xi\to (4\pi t)^{-\frac{d}{2}}e^{-\frac{\xi^2}{4t}}](x)$

Substitute into (*) and use Plancherel,

\begin{align*} \langle T,\hat{f} \rangle & =\frac{1}{\Gamma(\frac{d}{2})}\int_0^\infty \frac{dt}{t} t^{\frac{\alpha}{2}}(2\pi)^d\int_{\mathbb{R}^d\setminus\{0\}}d^d\xi\ (4\pi)^{-\frac{d}{2}}e^{-\frac{\xi^2}{4t}}f(\xi) \\ & \overset{Fubini}{=}\frac{1}{\Gamma(\frac{\alpha}{2})}\int_{\xi \neq 0}d^d \xi \ f(\xi) \pi^{\frac{d}{2}}\int_0^\infty \frac{dt}{t}t^{\frac{\alpha-d}{2}}e^{-\frac{\xi^2}{4t}} \\ & =\frac{1}{\Gamma(\frac{\alpha}{2})}\int_{\xi \neq 0}d^d \xi \ f(\xi) \pi^{\frac{d}{2}}(\frac{\xi^2}{4})^{\frac{\alpha-d}{2}}\int_0^\infty \frac{ds}{s}s^{\frac{d-\alpha}{2}}e^{-s},\quad s=\frac{\xi^2}{4t} \end{align*} Thus, $$ \langle \widehat{T},f \rangle =\frac{\Gamma(\frac{d-\alpha}{2})}{\Gamma(\frac{\alpha}{2})} 2^{d-\alpha}\pi^{\frac{d}{2}} \int_{\xi \neq 0}d^d\xi\ \frac{1}{|\xi|^{d-\alpha}}f(\xi) $$

The last part does define an element in $S'$ as $d-\alpha<d\iff \alpha>0$.

  • 1
    What do you mean with $$? – md2perpe Aug 20 '19 at 19:23
  • 1
    Sorry, that's the standard notation for $\sqrt{1+x^2}$ where $x^2=\sum_{i=1}^{d} x_i^2$. I have seen people call this the "Japanese bracket" but I don't know why. – Abdelmalek Abdesselam Aug 21 '19 at 15:03
  • Thanks. I've never seen it before. Googling on "Japanese bracket math" showed that I'm not the only one: https://math.stackexchange.com/a/1718676/168433 – md2perpe Aug 21 '19 at 22:57
  • Hi! Sorry for asking about this after such a long time, but here the Fourier transform you calculated for $|x|^{-\alpha}$ is not the real $|x|^{-\alpha}$, it is $\phi$ the modification of $|x|^{-\alpha}$ that you are transforming, right? In this case no matter what value $\alpha$ takes, the Fourier transform of $\phi$ is always well-defined? – Chang Apr 16 '24 at 15:54
  • @Chang: I don't understand what you said. What is for you "the real" $|x|^{-\alpha}$? and in what detailed sense is it different from my $\phi$? – Abdelmalek Abdesselam Apr 19 '24 at 14:49
  • Thanks for your reply! In your definition $\phi$ takes value $0$ at the origin, however, the true $|x|^{-\alpha}$ is undefined for $x=0$. So I want to ask to calculate the Fourier of $|x|^{-\alpha}$, do we have to insert the value of $x=0$ in such a way, so that a tempered distribution could be defined properly? – Chang Apr 20 '24 at 18:16
  • 1
    As far as what distribution is associated to $|x|^{-\alpha}$, it makes no difference if one uses what you call the true $|x|^{-\alpha}$ or $\phi(x)$. I introduced $\phi$ just as a matter of convenience so as to use $\mathbb{R}$ instead or $\mathbb{R}\backslash{0}$ as a domain of integration. – Abdelmalek Abdesselam Apr 22 '24 at 22:37