Let $V$ be an arbitrary set of infinitely many positive integers, and let: $$\varsigma_{V}\left(z\right)\overset{\textrm{def}}{=}\sum_{v\in V}z^{v}$$ Let $T_{V}$ denote the set of all $t\in\left[0,1\right)$ for which the limit:

$$c_{V}\left(t\right)\overset{\textrm{def}}{=}\lim_{x\uparrow1}\left(1-x\right)\varsigma_{V}\left(e^{2\pi it}x\right)$$ exists.

Suppose $0\in{T_{V}}$ and $c_{V}\left(0\right)\neq0$ (that is, $V$ has a positive, well-defined natural density). Are there then any $t\in\left(0,1\right)$ which must be in $T_{V}$? More generally, what, if anything, can be said about $T_{V}$ and/or $\left[0,1\right)\backslash T_{V}$? (Ideally, all of $\mathbb{Q}$ (or all but some exceptional set) would have to be a subset of $T_{V}$.) Conversely, what (if anything) does $\mathbb{Q}\subseteq T_{V}$ then force to be true about $V$? In particular, if $\varsigma_{V}\left(z\right)$ is a rational function, then $\mathbb{Q}\subseteq T_{V}$ necessarily holds. Thus, another way of phrasing the question is: does the set $\left\{ \varsigma_{V}\left(z\right):T_{V}\supseteq\mathbb{Q}\right\}$ contain functions which are not rational, and, if such functions do exist, what can be said about them and their associated $V$s?

  • 1
    $\begingroup$ I apologize for my comment is not directly related to this question, though it is related to many of the questions you have posted recently on Mathoverflow. It seems that you are interested in what happens in on the unit circle in $\Bbb C$. I am interested in the same topic too, and I appreciated some of the related questions you posed, therefore, when you are going to publish some works on this topic, I'd be glad to be informed. You can find my contact address on my MO user profile: thank you. $\endgroup$ Jun 24, 2019 at 12:06
  • 1
    $\begingroup$ No problem. I adore attention, so, the more the better. xD This is all for my doctoral dissertation. My challenge is two-fold: (1) no one at my university is doing anything close to analytic number theory; (2) I"m doing mathematics of the problem-solving kind, rather than the theory-building, so, finding the specific information applicable to my circumstances is often arduous. Asking about things online has been a life-saver. $\endgroup$
    – MCS
    Jun 24, 2019 at 18:58
  • 1
    $\begingroup$ Believe it or not, the research I'm doing isn't into these functions themselves, but rather, their application to the study of certain arithmetic dynamical systems. Of particular import is the fact that (by way of p-adic analysis, of all things!) I've found a generalization of my initial result that appears to work for all functions in the class I describe in italics at the end of my question. It's nerve-wracking, being stuck in the dark like this as to the significance my work. Hopefully, though, adding a bounty to this question will get me an answer and put me out of my misery. ;) $\endgroup$
    – MCS
    Jun 24, 2019 at 19:06
  • 1
    $\begingroup$ Before adding a bounty, let me try in a couple of days if I can find something interesting pertaining to your problem. $\endgroup$ Jun 24, 2019 at 20:23
  • 1
    $\begingroup$ Thanks, though the bounty ship has already sailed, I'm afraid. xD $\endgroup$
    – MCS
    Jun 24, 2019 at 20:27

2 Answers 2


There are lots of different questions going on here, but here are a few observations.

First of all, if you change $V$ by any set with (upper) density zero, then $c_V(t)$ doesn't change for any value of $t$. So for any $T_V$ there are uncountably many possible choices of $V$, and so almost all of them will be transcendental functions. (For an explicit example where $T_V = [0,1)$, let $V_S$ consist of the set of all integers minus any subset $S$ consisting of powers of $2$. The set of possible $S$ is uncountable.)

Second, it's also relatively easy to construct $V$ such that $T_V$ is missing infinitely many rational numbers. Let's start out with an example which is just missing $1/3$ and $2/3$. The idea is to take $V$ to consist of a string of numbers which are $1$ mod $3$, then a much longer string of numbers which is $2$ mod $3$, then a much much longer string which is $1$ mod $3$ and so on. That is,

$$\varsigma(x) = \frac{x}{1-x^3} + x^{3a_1} \frac{x^2 - x}{1-x^3} + x^{3 a_2} \frac{x - x^2}{1 - x^3} + x^{3 a_3} \frac{x^2 - x}{1-x^3} + \ldots $$

Note that

$$\varsigma(x) = \frac{x}{1-x^3} \left(1 + (x-1) x^{3a_1} - (x-1) x^{3 a_2} + (x-1) x^{3 a_3} + \ldots \right)$$

Here we imagine that the $a_n$ are growing really really fast, (say $x = n!^3$ or something). Let $t = 1/3$ and let $\omega = e^{2 \pi i t}$. Then $c_V(1/3)$ exists if and only if

$$\lim_{x \rightarrow 1^{-}} \left(1 + (\omega x-1) x^{3a_1} - (\omega x-1) x^{3 a_2} + (\omega x-1) x^{3 a_3} + \ldots \right)$$

exists. The claim is that the value of this sum will vary between $1$ and $\omega$ as $x$ approaches $1$. The reason is that when $a_n$ increases sufficiently fast, one can choose $x$ so that the first $n$ values of $x^{3 a_n}$ are very close to one and the others decrease super-exponentially. For example, choose a value of $x$ such that

$$1 - \frac{1}{n^2 a_n} < x < 1 - \frac{n}{a_{n+1}} < 1 - \frac{n^{k}}{a_{n+k}}.$$

Then for $m \le n$,

$$x^{3 a_m} \ge x^{3 a_n} \ge \left(1 - \frac{1}{n^2 a_n}\right)^{3 a_n} \sim 1 + O\left(\frac{1}{n^2}\right),$$

but for $m = n+k > n$,

$$x^{3 a_m} \le \left(1 - \frac{3 n^{k}}{a_{n+k}}\right)^{3 a_{n+k}} \sim e^{-3 n^k}$$

which are thus negligible. Thus the value of the sum is approximately

$$1 + (\omega - 1) - (\omega -1) + (\omega -1) - \ldots $$

and so will vary between $1$ and $\omega$ and not converge.

On the other hand, it's not so hard to see that there will be no convergence issues for all other rational numbers except for $t = 1/3$ and $t = 2/3$. The point is that the $1-x$ term kills of any power series with terms that are so spread out.

But now consider:

$$\varsigma_W(x):=\varsigma_V(x) + \varsigma_V(x^3) + \varsigma_V(x^9) + \ldots $$

Since $V$ consisted only of terms $1$ and $2$ modulo $3$, there exists a corresponding set $W$. But now the analysis above shows that $T_W$ avoids the set $a/3^n$ for every $(a,3) = 1$ and $n \ge 1$. The point is that these values are screwed up exactly by $\sigma_V(x^{3^{n-1}})$. So $T_W$ is missing infinitely many rational numbers.

Some preliminary analysis suggests one can construct functions missing all rational numbers except zero but the analysis could be a bit painful. (The idea is similar to the above, but to slowly increase the modulus one is considering from $2!$ to $3!$ to $4!$ etc.)

Other observations:

Certainly a $V$ chosen randomly will, almost always (in the appropriate sense) have $S \subset T_V$ for any finite set $S$. I might even guess that for a random $V$ the set $T_V$ contains all rational numbers, did you try to prove that?

  • $\begingroup$ This is amazing. I was thinking of arranging $V$ to stay in one residue class for a long time, then another residue class for an even longer time, etc., but (ironically enough for someone who thinks of himself as an analyst) I always have trouble playing around with the growth-rate sandbox like this. As for probabilistic approaches... let's just say they're not my cup of tea. That being said... $\endgroup$
    – MCS
    Jun 26, 2019 at 2:01
  • $\begingroup$ ...does your construction (or a variant thereof) still hold if $V$ is required to contain an infinite arithmetic progression (ex: $a,a+b,a+2b,a+3b,...$)? Moreover, are there any techniques you can think of that might be helpful here, aside from the kind of brute force sequence construction that you used? $\endgroup$
    – MCS
    Jun 26, 2019 at 2:04
  • $\begingroup$ There's also one more condition that I just realized we can use: if $V$ is such that $T_{V}\cap\mathbb{Q}\neq\mathbb{Q}$, then the same needs to be true of $\mathbb{N}_{0}\backslash V$, where $\mathbb{N}_{0}$ is the set of non-negative integers, or else I win and there's nothing that any of us need to worry about. :) $\endgroup$
    – MCS
    Jun 26, 2019 at 2:23

As for probabilistic approaches... let's just say they're not my cup of tea Well, that certainly has to be fixed if one "thinks of himself as an analyst". Let me give you a quick overview of the things most relevant to your current project as a series of exercises.

1) Read https://en.wikipedia.org/wiki/Hoeffding%27s_inequality

2) Look up the proof of the Bernstein inequality $$ \|P'\|_\infty\le Cn\|P\|_\infty $$ for trigonometric polynomials $P(t)=\sum_{k=-n}^n c_ke^{ikt}$ of degree $n$. You do not need the sharp value of $C$.

3) Derive from it that $\|P\|_\infty\le C\max_{k=0}^{100n}|P(e^{(2\pi i k)/(100n)}|$

4) Using Hoeffding and 3), show that if $\xi_j$ are mean zero independent Bernoulli that are $p_j-1$ with probability $p_j$ and $p_j$ with probability $(1-p_j)$, then $$ P(\|\frac 1N\sum_{j=1}^N\xi_jz^j\|_\infty\ge N^{-0.1})\le N^{-0.1} $$ for large $N$.

5) Using Borel-Cantelli and the fact that Chesaro means of a bounded sequence can be read from any subsequence $N_k$ with $N_{k+1}/N_k\to 1$ as $k\to\infty$, show that with probability $1$, we have $$ \lim_{N\to\infty}\frac 1N \sum_{j=1}^N\xi_jz^j=0 $$ for all $z\in\mathbb T$ simultaneously.

6) Consider now the random set $V$ to which each particular integer $j$ belongs independently with probability $a_j\in[0,1]$. Show that with probability $1$, for every $t$, $$ \lim_{N\to\infty}\frac 1N\left|\sum_{v\in V, v\le N}z^v-\sum_{v=1}^N a_vz^v\right|=0 $$

Thus, it does not matter if you consider sums over sets or arbitrary sums with coefficients in $[0,1]$, but with the latter you have way more freedom in various constructions. For instance,

7) Construct a series that gives you a non-zero limit at $0$, $z$, $\bar z$ but $0$ limit everywhere else.

8) Construct a series that gives you non-zero limits at any prescribed self-conjugate set of "rational points" containing $t=0$ and zero limits everywhere else.



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.