Let $f:\mathbb R\to\mathbb C$ be a sufficiently smooth and $1$-periodic function of average zero (i.e., $\int_0^{1}f(x)\,dx=0$), and let $\alpha\in(0,1)\smallsetminus\mathbb Q$. We know that

$$ \lim_{N\to\infty}\frac{1}{N} \sum_{k=1}^N f(k\alpha)=0. $$

Let $s_n=\sum_{k=1}^n f(k\alpha)$. Is it true that the sequence $\{s_n\}_{n\in\mathbb N}$ is bounded?


I believe the answer depends on the irrationality measure of $\alpha$.

Per Fourier, Write $f= \sum_{m=-\infty}^\infty a_m e^{ 2 \pi i m x}$, then

$$s_n = \sum_{m=-\infty}^\infty a_m \frac{ e^{2\pi i m (n+1) \alpha} - 1}{e^{2 \pi i m \alpha} - 1}$$

so we have the bound:

$$ |s_n| < 2\sum_{m=-\infty}^\infty \frac{ |a_m|} { | e^{ 2 \pi i m \alpha} - 1 |} $$

If for some number $\mu$, the equation:

$$ \left| \alpha- \frac{p}{q} \right| < \frac{1}{q^\mu} $$

has finitely many solutions, then we have the bound $1/ | e^{ 2 \pi i m \alpha} - 1 | = O( m^{\mu-1} ) $, by setting $q=m$ and $p=$ the closest integer approximation to $m\alpha$. If $f$ is $\lfloor\mu\rfloor+1$ times continuously differentiable, we have the bound $|a_m|=o(m^{\lfloor\mu\rfloor+1})$. Multiplying these, we see that the terms of our sum are $o(1/m^{2+\lfloor \mu\rfloor - \mu})= o(1/m^{1+\epsilon})$, so the sum converges. Working more carefully, one could probably improve this to $\lfloor \mu \rfloor$ times differentiable using the fact that solutions $p,q$ are very rare.

However, for Liouville numbers, this bound does not hold for any $\mu$. Specifically, if $\alpha$ is the Liouville constant then for $m=10^{a!}$, $1/|e^{2\pi i m \alpha} - 1|$ is about $10^{(a+1)!}=m^{a+1}$, so even for smooth $f$ this bound is useless. In fact, for $\alpha$ the Liouville constant, $s_n$ cannot be bounded at all:

Take $f$ to be the smooth function

$$f(x) = \sum_{t=1}^\infty \frac{e^{ 2 \pi i 10^{t!}x} }{10^{ (t+1) !}}$$

Then by choosing $n = \sum_{b=1}^c \frac{ 10^{(b+1)! - b!}}{2} $, the first $c$ turns will all be around $(1 - e^{ 2\pi i/2})/i=-2i$ and the remaining terms will be vanishingly small, so the sum will go to $\infty$ as $c$ goes to $\infty$.

Suppose that for every $f$ that has $r$ derivatives with the last derivative $L^1$, $s_n(f)$ is bounded independent of $n$. Then the irrationality measure of $\alpha$ is at most $r+1$.

Let $X$ be the Banach space of functions that have $r$ derivatives with the last derivative $L^1$. Consider the sequence of linear functions from $X$ to $\mathbb R$ whose $n$th element sends $f \to s_n(f)$. If this sequence is bounded when applied to any element of $X$, then by the uniform boundedness principle it is uniformly bounded, hence there is an absolute constant $C$ such that $s_n( e^{2\pi i m x} ) < C m^r$ for all $n,m$. By ordinary equidstribution, the supremum over all $n$ of $s_n(e^{ 2 \pi i q x})$ is $2/(1 - e^{ 2 pi i q\alpha} = 1/ O (|q\alpha-p|)$ for intgers $p$, which gives us the inequality

$$|q\alpha-p| > C' q^{-r}$$

$$|\alpha-\frac{p}{q} | > C' q^{-r-1}$$

So the irrationality measure of $\alpha$ is at most $r+1$.

  • $\begingroup$ Very good counterexample. I am wondering whether the following is true: Is it possible for every smooth 1-periodic and 0-average function (which is not a trigonometric polynomial) to find an irrational, such that the sequence in my question is not bounded? $\endgroup$
    – smyrlis
    Dec 5 '13 at 12:14
  • $\begingroup$ I believe so, by restricting to successive intervals. Suppose we have already chosen $\alpha$ to lie in some very small interval such that for some number of $m$, $e^{2 \pi i m \alpha} - 1$ is quite small, and for some $n$, $s_n$ is very large. Let $w$ be the width of the interval. Then for all $m> 1/w $, we can choose an appropriate $\alpha$ such that $e^{ 2\pi m \alpha}$ is any point on the unit circle we want. So fix some $m>1/w$ such that $a_m\neq 0$, and choose an $\alpha$ such that $e^{2 \pi m \alpha}$ is incredibly close to $1$, making some appropriate $s_n$ arbitrarily large. $\endgroup$
    – Will Sawin
    Dec 5 '13 at 15:31
  • $\begingroup$ Actually choose an interval of appropriate $\alpha$s. Then we can repeat the process arbitrarily often, making a bigger $s_n$ each time. The only difficulty is making sure that later $s_n$s don't mess up earlier $s_n$s. But we control the contribution from $m$ that come just after our $m$, because $e^{2 \pi i m\alpha]-1$ will be large for there, and we can control the contribution from far-away terms because the trigonometric sum will be at most $n$ and $a_m$ will be very small. So it should be possible to do this, but the calculation will probably be messy. $\endgroup$
    – Will Sawin
    Dec 5 '13 at 15:59

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.