10
$\begingroup$

Consider for $X = 1,2, \ldots$ the set $\mathcal{S}_X$ of trigonometric polynomials $f(t) := \sum_{|k| \leq X} c_k e^{2\pi i kt}$ on the circle $\mathbb{T} := \mathbb{R}/\mathbb{Z}$ of degree $\leq X$ (Fourier transform supported on $\{-X,\ldots, X\}$), such that $f(0) = 1$ and $c_0 = 0$ (the last assumption is probably inessential). Let $$ M_X(f) := \sup_{\mathbb{T} \setminus [-\frac{1}{X},\frac{1}{X}] }|f|. $$ and $$ B_X := \inf_{f \in \mathcal{S}_X} M_X(f). $$

Is the $X \to \infty$ limit of $B_X$ strictly positive, or zero?

The motivation is the same as in this question that I have asked previously. Observe that, for all $n \leq X$, the average of $f(q)$ over $\mu_{n} \setminus \{1\}$, $q := e^{2\pi i t}$, does not exceed $M_X(f)$ in magnitude. As a result, Chebyshev's bound, the orthogonality relation in $\mathbb{Z} / n$ and the Dirichlet convolution identity (definition of the Mangoldt function) $\log = 1 * \Lambda$ yield for any $f \in \mathcal{S}_X$ the estimate $$ \sum_{n \leq X} \frac{\Lambda(n)}{n} = \sum_{k \neq 0} c_k \log{|k|} + O(M_X(f)), $$ where the implied coefficient is absolute and explicit. This would be an asymptotic formula if the $O(\cdot)$ term could be made to approach zero as $X \to \infty$, suggesting that $M_X(f)$ should perhaps be bounded away from zero. If not, then of course the next question would be to ask for the asymptotic computation of the extremal sequence $f_X$ and its decay rate $M_X(f_X) = B_X$.

It occurred to me that the linked question may have possibly been about functions on the circle $\mathbb{T} \leftrightarrow \mathbb{Z}$ rather than on the real line $\mathbb{R} \leftrightarrow \mathbb{R}$. For (I could be wrong about this) it seems to be a rather special situation to have $c_k \sim \frac{1}{X}\varphi(k/X)$ with $\varphi \in \mathcal{S}(\mathbb{R})$ a fixed Schwartz function supported on $[-1,1]$ and with $\varphi(0) = 0$ and $\widehat{\varphi}(0) = 1$. It does follow from the same argument as in Terry Tao's solution of the linked problem that there is an absolute $\epsilon_0 > 0$ such that $\lim_{X \to \infty} M_X(\sum_{k} \frac{1}{X} \varphi(k/X) e^{2\pi i kt}) \geq \epsilon_0$ for all such $\varphi$; for this unpacks to stating that $\sup_{\mathbb{R} \setminus [-1,1]} |\widehat{\varphi}| \geq \epsilon_0 > 0$ whenever $\mathbb{supp}(\varphi) \subset [-1,1]$ and $\varphi(0) = 0, \widehat{\varphi}(0) = 1$: a version of the uncertainty principle on the real line. But it isn't clear to me whether the sequence of solutions to our extremal problem should have such a limiting distribution $\varphi$. Also it would be nice to know of an argument that is directly about trigonometric polynomials.

Is there a version of the uncertainty principle on the circle that would yield the $M_X(f) \geq \epsilon_0 > 0$ answer in the present question too?

[Note: The conditions $c_0 = 0$ and $\varphi(0) = 0$ are probably irrelevant to the discussion; but they are convenient, so let me impose them for concreteness' sake. Other natural choices would be to take $c_0 = 1/X$ (corresponding to $\varphi(0) = 1$), or to drop them altogether. ]

$\endgroup$
2
  • $\begingroup$ @reuns: It doesn't exist, according to the answer. All this amounts is to observe and prove that a degree $X$ trigonometric polynomial having $f(0) = 1$ is not too small (i.e., is bounded below by a positive absolute constant) in the complement of the $1/X$-neighborhood of $0$. $\endgroup$ Dec 8, 2017 at 20:06
  • $\begingroup$ So there is no concrete number-theoretic $f$ you really want to study ? If I ask it is because that's the kind of questions I'm interested in about $\sum_{n \le N}\frac{ \mu(n)}{n} \{ nx\}, \sum_n e^{2i \pi n x} 1_{\gcd(n,N!)=1}$, $\min_a |1-\zeta(s)\sum_{n=1}^N a_n(N) n^{-s}|^2$ $\endgroup$
    – reuns
    Dec 8, 2017 at 20:10

1 Answer 1

11
$\begingroup$

A compactness argument shows that for sufficiently large $X$ one has the bound $$ \sup_{x \in {\mathbb T} \backslash [-1/X,1/X]} |f(x)| \gg \sup_{x \in [-1/X,1/X]} |f(x)|$$ whenever $f$ is a trigonometric polynomial of degree at most $X$; this would imply that $B_X \gg X$. (Perhaps there is a normalising factor of $1/X$ missing in your question?)

Proof: Suppose the claim failed, then one could (after normalising) find a sequence $X_n \to \infty$ and a sequence $f_n$ of trigonometric polynomials of degree at most $X_n$ such that $$ \sup_{x \in [-1/X_n,1/X_n]} |f_n(x)| = 1$$ and $$ \sup_{x \in {\mathbb T} \backslash [-1/X_n,1/X_n]} |f_n(x)| = o(1).$$ We can find $x_n \in [-1/X_n,1/X_n]$ such that $|f_n(x_n)|=1$. Writing $F_n: {\mathbf R} \to {\mathbf C}$ for the $X_n$-periodic function $$ F_n(x) := f_n( \frac{x}{X_n} - x_n \hbox{ mod } 1)$$ we see that $|F_n(0)|=\|F_n\|_{L^\infty}=1$, that $\sup_{2 \leq |x| \leq X_n/2} |F_n(x)| = o(1)$, and that $F_n$ is band-limited to $[-1,1]$ (i.e., its Fourier transform is supported in $[-1,1]$). In particular, if $\varphi$ is a Schwartz function whose Fourier transform equals $1$ on $[-1,1]$, then $|\langle F_n, \varphi \rangle| = |F_n(0)| = 1$.

By passing to a subsequence one can assume that $F_n$ converges weakly to another function $F$ in the unit ball of $L^\infty$, which is then non-zero by testing against $\varphi$. On the other hand, $F_n$ vanishes outside of $[-2,2]$ and its distributional Fourier transform is supported on $[-1,1]$, which is a contradiction as the Fourier transform is also analytic.

The compactness argument does not give an effective bound for the implied constant, but presumably one can do so if one uses a suitable effective version of the uncertainty principle, such as the one in this blog post of mine.

$\endgroup$
1
  • 1
    $\begingroup$ So it is the same as in $\mathbb{R}$: extending functions by periodicity, and applying the same kind of compactness argument. Hence I was wrong to think it could be a different problem. Yes, I had neglected the $1/X$ factor: it should be simply about the bound $M_X(f) \gg 1$ among the degree-$X$ trigonometric polynomials with $f(0) = 1$. I edited my post to reflect that. Thanks for this very quick answer! $\endgroup$ Dec 8, 2017 at 6:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.