3

Let $\{x_k\}$ be a $\frac{1}{R}$-separated set of points on $S^{d-1}$. Then $$ \left\|\sum_ka_ke^{ix_k\cdot\xi}\right\|_{L^2\left(B\left(0,R\right)\right)}\lesssim R^{d/2} \left(\sum_k|a_k|^2\right)^{1/2}, $$ where $B\left(0,R\right)$ is the ball of radius $R$ in $\mathbb{R}^d$.

The author claims that the result follows from the almost orthogonality of $\{e^{ix_k\cdot\xi}|_{B\left(0,R\right)}\}$. I tried expanding the square, but I had troubles estimating mixed products.

I tried also using Cotlar-Stein lemma, that is, define the operators $T_k:l^{N}\to L^2\left(B\left(0,R\right)\right)$ as $T_k\left(a\right)=a_ke^{ix_k\cdot\xi}$, where $N$ is the number of points on $S^{d-1}$. Then $\|T_kT_l^*\|=c_d\delta_{kl}R^{d}$, where $c_d$ is the area of the unit ball, and $\|T_k^*T_l\|=R^d|\int_{B\left(0,1\right)}e^{iR\left(x_k-x_l\right)\cdot\xi}\,d\xi|\lesssim R^{\left(d-1\right)/2}\frac{1}{|x_k-x_l|^{\left(d+1\right)/2}}$ if $k\neq l$ and $\|T_k^*T_k\|\sim R^d$.

We have $\sum_l\|T_kT_l^*\|^{1/2}\sim R^{d/2}$, but $\sum_l\|T_k^*T_l\|^{1/2}\lesssim R^{\left(d-1\right)/4}\left(R^{\left(d+1/4\right)}+\sum_{\substack{l=1\\ l\neq k}}^N\frac{1}{|x_k-x_l|^{\left(d+1\right)/4}}\right)$ and in the worst case $N\sim R^{d-1}$, so the summation term could be rather large.

It seems that the author proves this lemma easily, but I don't know how. Thanks!

Hakim
  • 10,311
user90189
  • 1,650
  • Often, seeing a statement in wider context helps to understand what is going on. Can you give a link to the paper with this claim? –  Jul 16 '14 at 17:35
  • J. Bourgain, Besicovitch Type Maximal Operators and Applications to Four Analysis, 1(2), 1991, p. 147. Lemma 6.29. Actually, it's frustrating to read Bourgain, can you recommend a preparatory text, before reading Bourgain? – user90189 Jul 16 '14 at 22:38
  • I'd rather change the variable of integration, $\xi=R\zeta$, and let $y_k=Rx_k$, so that $y_k$ are $1$-separated and lie on sphere of radius $R$. The desired inequality simplifies to
    $$ |\sum a_ke^{i y_k \cdot \zeta}|{L^2(B(0,1))}\lesssim \left(\sum_k|a_k|^2\right)^{1/2} $$ So we have to bound the norm of the matrix with entries $c{k,l}=\int_{B(0,1)} e^{i (y_k-y_l) \cdot \zeta},d\zeta $ by a constant independent of $R$...
    –  Jul 17 '14 at 04:54
  • 1
    As you know, $|c_{k,l}|\lesssim |y_k-y_l|^{-(d+1)/2}$ when $k\ne l$. (The diagonal elements are $\lesssim 1$.) When $d=2$, Schur's test works, because (roughly) $|y_k-y_l|\ge |k-l|$, and $\sum k^{-3/2}<\infty$. But already for $d=3$ Schur's test is unsuccessful: the sum $\sum_k |y_k-y_l|^{-2}$ over a $1$-net on a sphere is $\sim \log R$. And it gets worse for $d>3$. –  Jul 17 '14 at 04:59
  • I'm guessing that the author refers to almost orthogonality of such exponentials as something "well-known in the field", which does not mean it has a straightforward proof. –  Jul 17 '14 at 06:02
  • The first I thought after reading "almost orthogonality", was "Cotlar lemma", and Schur's test doesn't work well in such a case. The problem I have with the matrix $\int_{B(0,1)}e^{i(y_k-y_l)\cdot\xi}$ is the size. Thank you for your help, I work on it (Bourgain paper). – user90189 Jul 17 '14 at 21:28
  • 1
    Fix a point $p$ on $R\cdot S^{d-1}$ that belongs to a $1$-separated set $E$ on the sphere. For every $k\lesssim \log R$ there are $\lesssim 2^{k(d-2)}$ points of $E$ at distance between $2^{k}$ and $2^{k+1}$. Each of these points contributes $2^{-k(d+1)/2}$ to the sum. So, using Schur's test we end up summing $2^{k(d-3)/2}$ over $k=0,\dots,\log R$. This works fine for $d=2$, but for $d=3$ there is logarithmic divergence, and for $d>3$ it's even worse. –  Jul 17 '14 at 21:37

1 Answers1

4

The following proof of your claim uses a somewhat different form of "orthogonality", namely of translations of a certain function on the Fourier side.

In the following, I use the version $\widehat{f}\left(\xi\right)=\int f\left(x\right)e^{-2\pi i\left\langle x,\xi\right\rangle }\, dx$ of the Fourier transform. It might be that I miss some factors of $2\pi$ below. I hope that can be excused.

Let us fix $\varphi\in\mathcal{S}(\mathbb{R}^{d})$ with the property that $\widehat{\varphi}\in C_{c}^{\infty}\left(B_{1}\left(0\right)\right)$ with $\widehat{\varphi}\geq0$ and $\varphi\left(0\right)=\int\widehat{\varphi}\, dx=2$. Here, the first equality comes from Fourier inversion. Existence of such a function is a standard result.

As $\varphi$ is continuous with $\varphi\left(0\right)=2$, there is some $r>0$ such that $\left|\varphi\left(x\right)\right|\geq1$ holds for all $\left|x\right|\leq r$. By shrinking $r$, we can assume $4\pi r\leq1$. Then, the family $\left(x_{k}\right)_{k}$ is a fortiori also $\frac{4\pi r}{R}$-separated.

Consider $$\varphi_{R}\left(x\right):=\left(\frac{r}{R}\right)^{d}\cdot\varphi\left(\frac{r}{R}x\right).$$ We then have (easy calculation) $\left\Vert \varphi_{R}\right\Vert _{2}=\left(\frac{r}{R}\right)^{d/2}\cdot\left\Vert \varphi\right\Vert _{2}$, as well as $\left|\varphi_{R}\left(x\right)\right|\geq\left(\frac{r}{R}\right)^{d}$ for $\left|x\right|\leq R$ and the Fourier transform of $\varphi_{R}$ is given by $$ \widehat{\varphi_{R}}\left(\xi\right)=\widehat{\varphi}\left(\frac{R}{r}x\right) $$ which is supported in $B_{r/R}\left(0\right)$.

For $\xi\in\mathbb{R}^{d}$ and $f:\mathbb{R}^{d}\rightarrow\mathbb{C}$, write $\left(M_{\xi}f\right)\left(y\right):=e^{2\pi i\left\langle \xi,y\right\rangle }$ and $\left(L_{\xi}f\right)\left(y\right):=f\left(y-\xi\right)$. As modulation corresponds to translation on the Fourier side (this is an easy calculation), we have, for $k\in\left\{ 1,\dots,N\right\} $: $$ \widehat{M_{\frac{x_{k}}{2\pi}}\varphi_{R}}=L_{\frac{x_{k}}{2\pi}}\widehat{\varphi_{R}}. $$ As the family $\left(x_{k}\right)_{k}$ is $\frac{4\pi r}{R}$-separated, we get \begin{eqnarray*} {\rm supp}\left(\widehat{M_{\frac{x_{k}}{2\pi}}\varphi_{R}}\right)\cap{\rm supp}\left(\widehat{M_{\frac{x_{n}}{2\pi}}\varphi_{R}}\right) & = & {\rm supp}\left(L_{\frac{x_{k}}{2\pi}}\widehat{\varphi_{R}}\right)\cap{\rm supp}\left(L_{\frac{x_{n}}{2\pi}}\widehat{\varphi_{R}}\right)\\ & \subset & \left[B_{r/R}\left(0\right)+\frac{x_{k}}{2\pi}\right]\cap\left[B_{r/R}\left(0\right)+\frac{x_{n}}{2\pi}\right]\\ & = & \frac{1}{2\pi}\cdot\left(\left[B_{2\pi r/R}\left(0\right)+x_{k}\right]\cap\left[B_{2\pi r/R}\left(0\right)+x_{n}\right]\right)=\emptyset. \end{eqnarray*} Finally, for $f_{1},\dots,f_{n}\in L^{2}(\mathbb{R}^{d})$ with disjoint supports, we have $$ \left\Vert \sum f_{j}\right\Vert _{2}^{2}=\int\left|\sum f_{j}\right|^{2}\, dx=\int\sum\left|f_{j}\right|^{2}\, dx=\sum\left\Vert f_{j}\right\Vert _{2}^{2}.\qquad\left(\dagger\right) $$

We now employ $\left(\dagger\right)$, the Plancherel theorem, as well as $\left|\varphi_{R}\left(x\right)\right|\geq\left(\frac{r}{R}\right)^{d}$ on $B_{R}\left(0\right)$, to derive \begin{eqnarray*} \left\Vert \sum_{k}a_{k}e^{i\left\langle x_{k},\xi\right\rangle }\right\Vert _{L^{2}\left(B_{R}\left(0\right)\right)} & \leq & \left(\frac{R}{r}\right)^{d}\cdot\left\Vert \sum_{k}a_{k}e^{i\left\langle x_{k},\xi\right\rangle }\varphi_{R}\left(\xi\right)\right\Vert _{L^{2}}\\ & = & \left(\frac{R}{r}\right)^{d}\cdot\left\Vert \sum_{k}a_{k}\cdot M_{\frac{x_{k}}{2\pi}}\varphi_{R}\right\Vert _{L^{2}}\\ & \overset{\text{Plancherel}}{=} & \left(\frac{R}{r}\right)^{d}\cdot\left\Vert \sum_{k}a_{k}\cdot\widehat{M_{\frac{x_{k}}{2\pi}}\varphi_{R}}\right\Vert _{L^{2}}\\ & \overset{\left(\dagger\right)}{=} & \left(\frac{R}{r}\right)^{d}\cdot\sqrt{\sum_{k}\left\Vert a_{k}\cdot\widehat{M_{\frac{x_{k}}{2\pi}}\varphi_{R}}\right\Vert _{L^{2}}^{2}}\\ & \overset{\text{Plancherel}}{=} & \left(\frac{R}{r}\right)^{d}\cdot\sqrt{\sum_{k}\left|a_{k}\right|^{2}\cdot\left\Vert \varphi_{R}\right\Vert _{2}^{2}}\\ & = & \left(\frac{R}{r}\right)^{d}\cdot\left(\frac{r}{R}\right)^{d/2}\cdot\left\Vert \varphi\right\Vert _{2}\cdot\sqrt{\sum_{k}\left|a_{k}\right|^{2}}\\ & = & \frac{\left\Vert \varphi\right\Vert _{2}}{r^{d/2}}\cdot R^{d/2}\cdot\sqrt{\sum_{k}\left|a_{k}\right|^{2}}. \end{eqnarray*} This proves the claim, as long as we allow the implied constant to depend on the dimension $d$, as $\varphi,r$ can be chosen once and for all for each dimension $d\in\mathbb{N}$. Also note that we did not make use of the assumption $\left|x_{k}\right|=1$ for all $k$.

PhoemueX
  • 36,211
  • Such a lovely trick, you just replaced $B(0,R)$ by a cut-off function. Thank you! – user90189 Jul 27 '14 at 19:50
  • More or less yes, but you cannot "just" use any cutoff-function, the support on the Fourier-side has to be chosen correctly. :) – PhoemueX Jul 29 '14 at 17:16