14
$\begingroup$

Let $f$ be a trigonometric polynomial on the circle $\mathbb{T}$ with $\hat{f}(j) = 0$ for all $j \in \mathbb{Z}$ with $\lvert j \rvert < n$. Does there exist some $C$ independent of $n$ and $f$ such that $$ \|f''\|_p \geq Cn^2 \| f \|_p, $$ where $1 \leq p\leq \infty$?

$\endgroup$
4
  • $\begingroup$ Can C depend on p? $\endgroup$
    – Yemon Choi
    Feb 15, 2016 at 13:30
  • 2
    $\begingroup$ The more general inequality $\| f' \|_p \geq Cn \| f \|_p$ should in fact hold. This is problem 1.8 in the first volume of Classical and Multilinear Harmonic Analysis by C. Muscalu and W. Schlag. I have a thread about this on Math StackExchange which I will update accordingly. $\endgroup$
    – Eric Thoma
    Feb 22, 2016 at 6:46
  • $\begingroup$ @ChristianRemling If if helps with justification of the edit, I would be eager to see the extension of your argument to the more general inequality. I have encountered an obstacle in the extension of my argument, as the sequence $a_{n,j}$ is no longer even. $\endgroup$
    – Eric Thoma
    Feb 23, 2016 at 4:15
  • $\begingroup$ @EricThoma: Actually, I've now discovered a problem with my answer and I've deleted (wasted too much time on this already). I'm not finding an easy argument why approximations of $\sum e^{ikx}/k$ should have an $L^1$ error not worse asymptotically than approximations of characteristic functions (though "philosophically" it's clear this has to be right). $\endgroup$ Feb 23, 2016 at 4:30

2 Answers 2

3
$\begingroup$

I wish to add another proof based on the following result.

If $(a_n)_{n \in \mathbb{Z}}$ is an even sequence of nonnegative numbers with $$ a_{n+1} + a_{n-1} - 2a_n \geq 0 \quad \forall n > 0, $$ then there exists $g \in L^1(\mathbb{T})$ with $g \geq 0$ and $\hat{g}(n) = a_n$. This is lemma 1.12 in Classical and Multilinear Harmonic Analysis Vol 1 by C. Muscalu and W. Schlag. The desired function is $$ g = \sum_{n=1}^\infty n (a_{n+1} + a_{n-1} - 2a_n) K_n $$ where $K_n$ is the Fejér kernel.

Define the sequences $(a_{n,j})_{j=0}^\infty$ by $$ a_{n,j} = \begin{cases} \frac{1}{n^2} + \frac{2(n-j)}{n^3},& \text{if } j < n\\ \frac{1}{j^2}, & \text{if } j \geq n \end{cases} $$ for each $n \in \mathbb{N}$. Then (extending to $j \in \mathbb{Z}$ by $a_{n,(-j)} = a_{n,j}$) we can use the lemma to find $g_n \in L^1(\mathbb{T})$ with $g_n \geq 0$ and $\hat{g}_n(j) = a_{n,j}$.

By the monotone convergence theorem, we have $$ \|g_n \|_1 = \sum_{j=1}^\infty j(a_{n,(j+1)} + a_{n,(j-1)} - 2 a_{n,j}). $$ A computation will show that $\| g_n \|_1$ is dominated by $n^{-2}$. Furthermore, for any trigonometric polynomial $f$ with $\hat{f}(j) = 0$ for all $| j | < n$, we have $$ f = g_n \ast f'' $$ so that Young's inequality finishes the proof.

$\endgroup$
1
  • $\begingroup$ [deleted comment, I didn't see your comment to the main question] $\endgroup$
    – Yemon Choi
    Feb 22, 2016 at 18:53
3
$\begingroup$

Here is more pedestrian argument. If $T_{n}$ is a trigonometric polynomial of degree at most $n$ with total mass $\frac{1}{2\pi}\int_{-\pi}^{\pi}T_{n}=1$ then $f*T_{n}=0$, and in particular, $$ f(x) =\frac{1}{2\pi}\int_{-\pi}^{\pi}(f(x)-f(x-s))T_{n}(s)ds. $$ Therefore, by the triangle inequality $$ \|f\|_{p} \leq \frac{1}{2\pi}\int_{-\pi}^{\pi}\|f(x)-f(x-s)\|_{L^{p}(dx)}|T_{n}(t)|dt \leq \|f'\|_{p}\frac{1}{2\pi}\int_{-\pi}^{\pi}|s||T_{n}(s)|ds, $$

where the inequality $\|f(x)-f(x-s)\|_{L^{p}(dx)} \leq |s| \|f'\|_{p}$ follows, for instance from Schur test applied to the operator $(Af')(x)=\int f'(t) 1_{[x-s,x]}(t)dt$.

Now how can we make $\frac{1}{2\pi}\int_{-\pi}^{\pi}|s||T_{n}(s)|ds$ of order $\frac{1}{n}$? Let us be not too demanding and seek for $T_{n}$ among even nonnegative trigonometric polynomials to reduce the matters to $\int_{0}^{\pi}sT_{n}(s)ds$.

One immediate choice is Fejer kernel

$$ k_{n}(s) = \frac{1}{n}\left(\frac{\sin(\frac{ns}{2})}{\sin(\frac{s}{2})}\right)^{2}. $$ Now $k_{n}(s) \asymp n$ on $[0,\frac{1}{n}]$, and $k_{n}(s)<C \frac{1}{n}\frac{1}{s^{2}}$ on $[\frac{1}{n}, \pi]$, therefore $\int_{0}^{\pi}sk_{n}(s)ds \leq C' \frac{\log(n)}{n}$. Well, not too bad but not exactly what was requested.

What else can we do? Let us look at $k_{n}^{2}(s)$. It is even nonnegative trigonometric polynomial of degree $2n$. A small jump in degree is okay (we can just start from $f$ of degree $\geq 2n$). Since $k_{n}^{2}(s) \asymp n^{2}$ on $[0, \frac{1}{n}]$ its total mass is at least $\geq n$. Then

$$ \frac{k_{n}^{2}}{n} \asymp n \quad \text{on}\quad [0, \frac{1}{n}], \quad \text{and} \quad \frac{k_{n}^{2}}{n} \leq C\frac{1}{n^{3}} \frac{1}{s^{4}} \quad \text{on} \quad [\frac{1}{n}, \pi] $$ therefore $\int_{0}^{\pi} s \frac{k^{2}_{n}}{n}\leq C' \frac{1}{n}$ voila!

$\endgroup$
0

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.