14
$\begingroup$

Let $a_n\in \mathbb{C}$. We can prove a mean value theorem, meaning an inequality $$\int_0^T \left|\sum_{n=1}^\infty a_n n^{-i t}\right|^2 dt \leq \sum_{n=1}^\infty (c_0 T + c_1 n + c_2) |a_n|^2.$$ Montgomery and Vaughan (1973) proved this statement with $c_0=1$, $c_1=O(1)$, $c_2=0$, and in fact one can make their work explicit and take $c_0=1$, $c_1=3\pi$, $c_2=3\pi/2$.

  1. If one uses the main inequality in Preissmann (1984), one obtains $c_0=1$, $c_1=2\pi C$, $c_2=\pi C$ with $C=\sqrt{1+\frac{2}{3} \sqrt{6/5}}=1.3154\dotsc$. Is this the record? I suppose one cannot do better than $c_0=1$, $c_1=2\pi $; is there an example showing as much?

  2. Can one prove an inequality with $c_1=2\pi$ at the cost of having $c_0$ larger than $1$?

$\endgroup$
4
  • $\begingroup$ Can you ask a precise question please? In the second part of question $1$ and in question $2$, what do you want regarding $c_2$? $\endgroup$ Commented Aug 30, 2022 at 0:54
  • $\begingroup$ The value of $c_2$ is pretty unimportant. The question is whether one can improeve on $c_0$ and $c_1$ (even at the cost at increasing $c_2$). $\endgroup$ Commented Aug 30, 2022 at 15:09
  • $\begingroup$ For the first part of question $1$, are you asking what the best $C$ is such that $(c_0,c_1,c_2) = (1,2\pi C,\pi C)$ works? $\endgroup$ Commented Aug 30, 2022 at 19:08
  • $\begingroup$ No - in question 1, I am asking whether one can prove the statement with $(1,c_1,c_2)$, $c_1<C$ and some $c_2$. I then ask whether one can find a counterexample that shows that the statement cannot hold with $c_0\leq 1$ and $c_1<2\pi$, even if one allows $c_2$ to be a large constant. $\endgroup$ Commented Aug 30, 2022 at 20:51

1 Answer 1

2
$\begingroup$

The following is an approach inspired by Selberg's treatment of the large sieve, and more directly by a sketch in Granville-Harper-Soundararajan (on the case of prime support). There remains an optimization problem to be solved at the end.


Let $\Phi:\mathbb{R}\to \mathbb{R}$ be a real-valued, symmetric, non-negative function such that $\Phi(t)\geq 1$ for $|t|\leq 1/2$, $\widehat{\Phi}(t)$ is non-increasing for $t\geq 0$, and the support of $\widehat{\Phi}(t)$ is contained in a compact interval $[-\ell,\ell]$. Let $\varphi(x) = \Phi(x/T)$; then $\widehat{\varphi}(t) = T \widehat{\Phi}(T t)$. Clearly $$\int_{-T/2}^{T/2} \left|\sum_{n=1}^\infty a_n n^{-i t}\right|^2 dt \leq \int_{-\infty}^\infty \varphi(t) \left|\sum_{n=1}^\infty a_n n^{-i t}\right|^2 dt.$$ Now we proceed in the usual way (as in, say, Iwaniec-Kowalski, proof of Thm. 9.1): expanding the square and inverting the order of summation and integration, and writing $e(t)=e^{2\pi i t}$, we see that $$\begin{aligned}\int_{-\infty}^\infty \varphi(t) \left|\sum_{n=1}^\infty a_n n^{-i t}\right|^2 dt &= \sum_{n_1=1}^\infty \sum_{n_2=1}^\infty a_{n_1} \overline{a}_{n_2} \int_{-\infty}^\infty \varphi(t) e\left(-\frac{\log(n_1/n_2)}{2\pi} t\right) dt\\ &= \sum_{n_1=1}^\infty \sum_{n_2=1}^\infty a_{n_1} \overline{a}_{n_2} \widehat{\varphi}\left(\frac{\log (n_1/n_2)}{2\pi}\right) \end{aligned}$$ Of course, $|a_{n_1} \overline{a}_{n_2}|\leq \frac{|a_{n_1}|^2 + |a_{n_2}|^2}{2}$, and $\varphi$ is real-valued and symmetric, as $\Phi$ is, as well as being non-negative. Hence $$\left|\sum_{n_1=1}^\infty \sum_{n_2=1}^\infty a_{n_1} \overline{a}_{n_2} \widehat{\varphi}\left(\frac{\log (n_1/n_2)}{2\pi}\right) \right|\leq \sum_{n_1=1}^\infty |a_{n_1}|^2 \sum_{n_2=1}^\infty \widehat{\varphi}\left(\frac{\log (n_1/n_2)}{2\pi}\right).$$

It is easy to see that the sum over the integers of a function $F$ of compact support with its only maximum at some $t=t_1$ such that $F(t)$ non-increasing for $t\geq t_1$ and non-decreasing for $t\leq t_1$ is at most $$F(0) + \int_0^\infty F(x) dx + \int_{-\infty}^0 F(x) dx =F(0) + \int_{-\infty}^\infty F(x) dx.$$ We apply this simple statement to the function $f(t) = \widehat{\varphi}\left(\frac{\log(n_1/t)}{n_2}\right)$. We obtain $$\sum_{n_2=1}^\infty \widehat{\varphi}\left(\frac{\log (n_1/n_2)}{2\pi}\right) = \widehat{\varphi}(0) + \int_{-\infty}^\infty \widehat{\varphi}\left(\frac{\log(n_1/t)}{2\pi}\right) dt.$$ It is clear that $\widehat{\varphi}(0) = T \widehat{\Phi}(0)$.

Now, in general, for $f:\mathbb{R}\to [0,\infty)$ of compact support and $g:\mathbb{R}\to \mathbb{R}$ differentiable and monotonic, $$|f\circ g|_1 \leq \frac{|f|_1}{\inf_{t: f(g(t))\ne 0} |g'(t)|}.$$ Let $f = \widehat{\varphi}$ and $g(t) = \log(n_1/t)/2\pi$. Then $f(g(t))\ne 0$ implies $|g(t)|\leq \ell/T$, and hence $$ e^{-\frac{2\pi \ell}{T}} n_1 \leq t\leq e^{\frac{2\pi \ell}{T}} n_1,$$ and so $$|g'(t)| = \frac{1}{2\pi |t|} \geq \frac{e^{-\frac{2\pi \ell}{T}}}{2\pi n_1}.$$ Hence, $|f\circ g|_1 \leq e^{\frac{2\pi \ell}{T}} 2\pi n_1 |f|_1$, i.e., $$\int_{-\infty}^\infty \widehat{\varphi}\left(\frac{\log(n_1/t)}{2\pi}\right) dt \leq e^{\frac{2\pi \ell}{T}} 2\pi n_1 \int_{-\infty}^\infty \widehat{\varphi}(t) dt.$$ Finally, $$\int_{-\infty}^\infty \widehat{\varphi}(t) dt = \varphi(0) = \Phi(0).$$

We conclude that $$\int_{-T/2}^{T/2} \left|\sum_{n=1}^\infty a_n n^{-i t}\right|^2 dt \leq \widehat{\Phi}(0) T \sum_n \left|a_n\right|^2 + e^{\frac{2\pi \ell}{T}} 2\pi \Phi(0) \sum_n \left|a_n\right|^2 n.$$

(We can replace the integral from $-T/2$ to $T/2$ by an integral from $0$ to $T$ if we prefer, simply by applying the above with $a_n n^{-i T/2}$ instead of $a_n$.)

It is clear, given our assumptions, that $\Phi(0)\geq 1$ and $\widehat{\Phi}(0) = |\Phi|_1 \geq 1$.


It remains to choose $\Phi$. One option is to let $\Phi = \widehat{F\ast F}$, where $F$ is a non-negative symmetric function of compact support with $F(x)$ non-increasing for $x\geq 0$. Then $\Phi$ equals $\widehat{F}^2 = |\widehat{F}|^2$, and so is non-negative and real-valued, besides being symmetric; since $\widehat{\Phi} = F\ast F$, we see easily that $\widehat{\Phi}$ is of compact support and $\widehat{\Phi}(t)$ is non-increasing for $t\geq 0$. Clearly $\Phi(0) = |F|_1^2$ and $\widehat{\Phi}(0) = |F|_2^2$. It would remain to verify the condition $\Phi(t) \geq 1$ for $|t|\leq 1/2$, i.e., $|\widehat{F}(t)|\geq 1$ for $|t|\leq 1/2$.

$\endgroup$
0

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .