Let us fix the following convention for the Fourier transform in $L^1(\mathbb{R})$ space: $$\hat{f}(\xi)=\int_{-\infty}^\infty f(x)\, e^{-2\pi i x\xi}\, dx.$$ We then have the following properties: \begin{align}\tag{1} \displaystyle \left[\frac{df}{dx}\right]^\hat{}(\xi)=2\pi i\xi\, \hat{f}(\xi); \\ \tag{2} \displaystyle \left[ -2\pi i x\, f\right]^\hat{}(\xi)=\frac{d\hat{f}}{d\xi}(\xi);\\ \tag{3} \displaystyle \hat{f}(0)=\int_{-\infty}^\infty f(x)\, dx. \end{align}

Question Let $K(x, \xi)$ be a bounded function. Suppose that the integral transform $$Tf(\xi)=\int_{-\infty}^{\infty}f(x)K(x, \xi)\, dx,\quad f \in L^1(\mathbb{R})$$ satisfies properties (1), (2) and (3). Is it true that $K(x, \xi)=\exp(-2\pi i x\xi)$?

Motivation for this question comes from the fact that one can evaluate $$\hat{G}(\xi)=\left[ \exp(-\pi x^2)\right]^\hat{}$$ by using only the properties (1), (2) and (3). This is done by Fourier transforming both sides of the differential identity $$\frac{d}{dx}e^{-\pi x^2}=-2\pi x\, e^{-\pi x^2},$$ obtaining the Cauchy problem $$ \begin{cases} -2\pi \xi\, \hat{G}(\xi)=\frac{d}{d\xi}\hat{G}(\xi) \\ \hat{G}(0)=1 \end{cases} $$ whose unique solution is $$\hat{G}(\xi)=\exp(-\pi\,\xi^2).$$

  • 1
    $\begingroup$ Very nice question. I'd guess the answer is negative, since I'd expect to read such a theorem on books in harmonic analysis. But I have no counter-example... $\endgroup$ – Siminore Oct 24 '12 at 14:12

Apart from some quibbling about making sure $K$ is sufficiently integrable or something... this is true. E.g., for precision, take $K$ to be a tempered distribution in two variables. Using the hypothesis that $2\pi ix\cdot T=T\cdot {d\over dx}$ as operators on Schwartz functions, (thinking of "$x$" as multiplication-by-$x$), integration by parts in the integral for $T$ gives ${\partial\over \partial y}K(x,y)=2\pi i x\cdot K(x,y)$ as tempered distribution in two variables. This has obvious classical solutions $C\cdot e^{2\pi ixy}$, as expected. To show that there are no others, among tempered distributions, one way is to divide $K(x,y)$ by $e^{2\pi ixy}$, so the equation becomes ${\partial \over \partial y}K(x,y)=0$. By symmetry, ${\partial\over \partial x}K(x,y)=0$. Integrating, $K(x,y)$ is a translation-invariant tempered distribution in two variables. It is a separate exercise to see that all such are constants.

Edit: in response to @GiuseppeNegro's comment (apart from correcting the sign), the secondary exercise of proving that vanishing first partials implies that a tempered distribution is (integrate-against-) a constant has different solutions depending on one's context, I think. Even in just a single variable, while we can instantly invoke the mean value theorem to prove that a function with pointwise values is constant when it is differentiable and has vanishing derivative, that literal argument does not immediately apply to distributions. In a single variable, integration by parts proves that $u'=0$ for distribution $u$ implies that $u(f')=0\,$ for all test functions $f$, and we can characterize such $f$, namely, that their integrals over the whole line are $0$, from which a small further argument proves that $u$ is a constant. This sort of argument seems to become a little uglier in more than one variable... and, in any case, I tend to favor a slightly different argument that is a special case of proving uniqueness of various group-invariant functionals. E.g., on a real Lie group $G$, there is a unique right $G$-invariant distribution (=functional on test functions), and it is integration-against right Haar measure. The argument is essentially just interchange of the functional and integration against an approximate identity, justified in the context of Gelfand-Pettis (weak) integrals. Probably there are more elementary arguments, but this sort of approach seems clearer and more persuasive in the long run.

Edit-Edit: on a Lie group $G$, to prove that all distributions annihilated by the left $G$-invariant differential operators attached to the Lie algebra $\mathfrak g$ (acting on the right) are (integrate-against-) constants: Let $f_n$ be a Dirac sequence of test functions. A test function $f$ acts on distributions by $f\cdot u=\int_G f(g)\,R_gu\;dg$, where $R$ is right translation, and the integral is distribution-valued (e.g., Gelfand-Pettis). A basic property of vector-valued integrals is that $f_n\cdot u\to u$ in the topology on distributions. At the same time, the distribution $f_n\cdot u$ is (integration-against) a smooth function. It is annihilated by all invariant first-order operators, so by the Mean Value Theorem it is (integration against) a constant. The distributional limit of constants is a constant.

  • $\begingroup$ Very nice! The fact that all translation-invariant tempered distributions are constants must surely be a consequence of the fact that their derivative is zero, right? (PS: There is a small typo: it is $2\pi i xT=T\frac{d}{dy}$, not $-2\pi i x T=T\frac{d}{dx}$). Thank you very much! $\endgroup$ – Giuseppe Negro Oct 24 '12 at 18:07
  • $\begingroup$ I know this is an old post, but I have a question: you said that "on a real Lie group $G$, there is a unique right $G$-invariant distribution, which is integration-against right Haar measure". How do you deduce from that fact that the only distributions on $\mathbb{R}^n$ whose all partial (distributional) derivatives are zero are the constants? (I don't see the relation between the first sentence and the corollary; Are you saying that every distribution with zero derivative must be translation-invariant? Why is that? Thanks). $\endgroup$ – Asaf Shachar Nov 5 '18 at 9:01

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.