10
$\begingroup$

It seems to me that in a paper of Charles Fefferman (open access), it is claimed in the introduction that (3rd page of the PDF file, 'page 11', $\lambda\in(0,n)$)

$$\sup_{\xi\in\mathbb R^n}\left| \int_{2^k\le |x| <2^{k+1}} \frac{\sin |x|}{|x|^\lambda} e^{2\pi i x \cdot\xi} \, dx \right| \le 2^{(\frac{n+1}2-\lambda)k}.$$ (He claims that Plancharel's formula should be used, I expect this to mean that we should compute the $L^\infty$ norm of the symbol.)

By using asymptotics of Bessel functions, I can get the upper bound for $\xi\gg 1$, but I am worried about $\xi\approx 0$. At $\xi=0$, the Fourier transform is just $\int\frac{\sin\dots}{\dots} dx$ leading to the estimate ($C_n$ is the volume of the unit $n$-sphere) $$ \left| \int_{2^k\le |x| <2^{k+1}} \frac{\sin |x|}{|x|^\lambda} \, dx \right| \le C_n 2^{(n-\lambda)k} $$ which is bigger for $n>1$? Am I missing something simple? Calculation details can be provided on request.

$\endgroup$
2
  • $\begingroup$ In deriving your bound did you use the obvious bound on sin? $\endgroup$
    – lcv
    May 14, 2020 at 11:38
  • $\begingroup$ @lcv yes, all I did was $|\sin r|\le 1$ $\endgroup$ May 14, 2020 at 11:39

1 Answer 1

6
+200
$\begingroup$

You're ignoring the oscillations of $\sin\vert x\vert$; by the way, in the paper you can replace $[2^k,2^{k+1}]$ by a smooth version $\zeta_k(x) = \zeta(x/2^k)$, which simplify a lot the computations, so I'll do that.

By rotational symmetry we can assume that $\xi = se_n$, for $s\ge 0$. If $s\gg 1$ then it is sufficient to use Bessel functions because the oscillation of $e^{2\pi isx_n}$ are dominant. If $s\ll 1$ then the oscillation of $\sin\vert x\vert$ are dominant.

By dilation $$ \int\zeta_k\frac{\sin\vert x\vert}{\vert x\vert^\lambda}e^{2\pi isx_n}\,dx = 2^{k(n-\lambda)}\int_1^2\zeta\frac{\sin 2^k\vert x\vert}{\vert x\vert^\lambda}e^{2\pi i2^ksx_n}\,dx. $$ Since $\sin r = \frac{1}{2i}(e^{ir}-e^{-ir})$, it suffices to get the bound $$ \left\vert\int a(x)e^{2\pi i\omega(sx_n-\vert x\vert/(2\pi))}\,dx\right\vert\le C\omega^{-\frac{n-1}{2}} $$ uniformly in $s\ge 0$; here $a = \zeta\frac{1}{\vert x\vert^\lambda}$ is a smooth function supported in $[1,2]$. The term with the factor $e^{i2^k\vert x\vert}$ is bounded similarly.

After a change of variables we can write the integral also as $$ \int ae^{2\pi i\omega(sx_n-\vert x\vert/(2\pi))}\,dx = \iint_{S^{n-1}}a r^{n-1}e^{2\pi ir\omega(s\theta_n-1/(2\pi))}\,dS(\theta)dr. $$ We evaluate the integral in different regions.

$\sin\vert x\vert$ is dominant: If $s\le \frac{1}{2\pi}-c$, for some $0<c\ll 1$, then $\vert s\theta_n-\frac1{2\pi}\vert\ge c$ for every $\theta$, and then integrating by parts $N$ times we get $$ \left\vert\int\frac{ar^{n-1}}{(2\pi i \omega(s\theta_n-1/(2\pi)))^N} \partial_r^N(e^{2\pi i\omega r(s\theta_n-1/(2\pi))})\,dr\right\vert \le C_N\frac{1}{\omega^N}, $$ so let's take $N>\frac{n-1}{2}$ and integrate over the sphere.

$e^{2\pi isx_n}$ is dominant: If $s\ge \frac1{2\pi}+c$, then $\vert s-\frac{x_n}{2\pi\vert x\vert}\vert \ge c$ and then again using integration by parts we get $$ \left\vert\int a e^{2\pi i\omega(sx_n-\vert x\vert/(2\pi))}\,dx\right\vert\le \frac{c}{\omega^N}; $$ here, we used $\partial_ne^{2\pi i\omega(sx_n-\vert x\vert/(2\pi))} = 2\pi i\omega(s-x_n/(2\pi\vert x\vert))e^{2\pi i\omega(sx_n-\vert x\vert/(2\pi))}$.

The hard part: The case $\vert s-\frac1{2\pi}\vert \le c\ll 1$ is hard and here you should expect the upper bound $C/\omega^\frac{n-1}{2}$. The phase of your oscillatory integral is $f(x) = sx_n-\vert x\vert/(2\pi)$, and Stein's book Harmonic Analysis: real-variable methods, orthogonality, and oscillatory integrals, Ch. VIII-IX is the standard reference.

Since $\nabla f = (-\frac{x_1}{2\pi\vert x\vert},\cdots,-\frac{x_{n-1}}{2\pi\vert x\vert}, s-\frac{x_n}{2\pi\vert x\vert})$, then the principle of non-stationary phase suggests that we should divide $\mathbb{R}^d$ into a tube $T = \{(x',x_n)\mid \vert x'\vert\le c, x_n>0\}$ and its complement. Let $\psi$ be a cut-off of $T$, so by non-stationary phase, i.e. repeated integration by parts as above, we get $$ \left\vert\int a(1-\psi)e^{2\pi i \omega f}\,dx\right\vert \le C_N\frac{1}{\omega^N}. $$ To bound the last contribution from the region $T$, fix $x_n>0$ and write $$ \int_{\mathbb{R}^{n-1}} a\psi e^{2\pi i \omega f}\,dx' = c_ne^{2\pi i\omega (s-1/(2\pi))x_n} \int_0^\infty a\psi r^{n-2} e^{-i\omega (\sqrt{r^2+x_n^2}-x_n)}\,dr. $$ The coefficient $e^{2\pi i\omega (s-1/(2\pi))x_n}$ has absolute value 1, so we can ignore it. Now make the change of variables $t = \omega(\sqrt{r^2+x_n^2}-x_n)$ to see that $$ \int_0^\infty a\psi r^{n-2} e^{-i\omega (\sqrt{r^2+x_n^2}-x_n)}\,dr = \frac{1}{\omega^\frac{n-1}{2}}\int_0^\infty \tilde{a}(t/\omega)t^\frac{n-3}{2}e^{-it}\,dt; $$ here $\tilde{a}$ is a smooth function supported in $\vert x'\vert\le c$. Hence, it suffices to get the upper bound $$ \left\vert \int_0^\infty \tilde{a}(t/\omega)t^\frac{n-3}{2}e^{-it}\,dt\right\vert\le C, $$ uniformly in $\omega$, which after integration in $x_n$, suffices to get the desired result. Using repeated integrations by parts we get $$ \vert \int_0^\infty \tilde{a}(t/\omega)t^\frac{n-3}{2}e^{-it}\,dt\vert = \vert\int_0^\infty \partial^N_t(\tilde{a}(t/\omega)t^\frac{n-3}{2})e^{-it}\,dt\vert, $$ whenever $N<\frac{n-3}{2}$. When you reach $N = \lfloor\frac{n-3}{2}\rfloor$, you have to be careful at the next integration by parts because you get a boundary term at 0. The typical term in the derivative after $N_1+N_2 = N$ integration by parts is $\omega^{-N_1}\partial^{N_1}\tilde{a}(t/\omega)\cdot t^{\frac{n-3}{2}-N_2}$, whose absolute value is less than $\omega^{\frac{n-3}{2}-N}$ in the support of $\tilde{a}(t/\omega)$. Hence, $$ \left\vert \int_0^\infty \tilde{a}(t/\omega)t^\frac{n-3}{2}e^{-it}\,dt\right\vert \le C, $$ with $C$ uniform in $\omega$. Taking $N\gg 1$ we conclude the estimations.

$\endgroup$
2
  • $\begingroup$ Thank you very much, especially for finding the time to finish the proof :) I'm still reading and will take some time to process $\endgroup$ May 15, 2020 at 1:03
  • 1
    $\begingroup$ Please, take your time, there are many computations. $\endgroup$
    – user90189
    May 15, 2020 at 9:32

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.