4

It is clear that for $r\in (0,1)$ fixed, the series $$f(r)=\sum_{n=0}^{\infty}2^n r^{2^n}$$ is convergent, and also it is not to hard show that $f(r)$ is not uniformly convergent in $(0,1)$. Since $\sum_{n=0}^{\infty}r^n=\frac{1}{1-r}$, I think this sum function $f(r)$ may be similar.

So can I find some upper bound such as $\frac{1}{1-r^\alpha}$ to dominate $f(r)$? Sincerely thanks!

Faye Tao
  • 1,814
  • 3
    You'll probably find everything that you need in here: https://en.m.wikipedia.org/wiki/Cauchy_condensation_test. In particular, we can bound $f(r)$ both from above and below using the geometric series of $r$ – blundered_bishop Jan 22 '22 at 08:08
  • 3
    $\sum_{n=0}^{\infty}2^n r^{2^n} < \sum_{k=0}^{\infty} k r^{k} = \frac{r}{(1-r)^2}$ .. the former is a sub-series of the later (only $k = 2^n$ terms) .. – r9m Jan 22 '22 at 08:09
  • @blundered_bishop Thx very much, that's exactly what I needed!!! – Faye Tao Jan 22 '22 at 08:22

2 Answers2

4

A crude bound is

$$f(x)<\frac{x}{(1-x)^2}\text{ for every }x\in(0,1)$$.

This can be obtained by noticing that our series is $\sum_{n=0}^\infty nx^n$ minus the terms whose index is not a power of two. For $x\in(0,1)$, the terms of $\sum_{n=0}^\infty nx^n$ are all strictly positive and the series converges, so since our series contains less terms than $\sum_{n=0}^\infty nx^n$, the following inequality holds:

$$\sum_{n=0}^\infty 2^n x^{(2^n)}<\sum_{n=0}^\infty nx^n$$

The sum of $\sum_{n=0}^\infty nx^n$ is $\frac{x}{(1-x)^2}$, a fact that can be proven by differentiating the geometric series identity $\sum_{n=0}^\infty x^n= \frac{1}{1-x}$, giving us the desired bound.

$$f(x)<\frac{x}{(1-x)^2}$$

Edit: with some thinking, we can get a better upper bound.

Notice that the only odd term of $f$'s series expansion is $x$, corresponding to the fact that $2^0=1$. Since the other powers of two $\{2^1,2^2,2^3,2^4,\dots\}$ form a subset of the even numbers, it follows that

$$f(x)-x=\sum_{n=1}^\infty 2^n x^{(2^n)}<\sum_{n=1}^\infty 2n x^{2n}\text{ for every }x\in(-1,1)\text{ with }x\neq 0$$

You may notice that the interval over which this holds was improved. This is permissible because the terms in each series consist of only even powers of $x$, so each series sums to an even function.

Finding the sum of $\sum_{n=1}^\infty 2n x^{2n}$ is also super easy, since it follows from the identity $\sum_{n=0}^\infty nx^n=\frac{x}{(1-x)^2}$:

\begin{align} \sum_{n=1}^\infty 2n x^{2n} &= \sum_{n=0}^\infty 2\cdot n(x^2)^n-2(0)(x^2)^0\\ &= \frac{2x^2}{(1-x^2)^2} \end{align}

We deduce that $f(x)<x+\frac{2x^2}{(1-x^2)^2}$ for every nonzero $x\in(-1,1)$ (they are equal at $0$).

Alann Rosas
  • 5,421
  • 1
  • 9
  • 28
1

This is a different approach with an arbitrary good bound for $x\in(0,1)$. $$f(x)=\sum_{n=0}^N 2^n x^{2^n} + 2\sum_{n=N+1}^\infty 2^{n-1} x^{2^n} = \sum_{n=0}^N 2^n x^{2^n} + 2\sum_{n=N+1}^\infty \underbrace{x^{2^n} + ... + x^{2^n}}_{2^{n-1} \text{ terms}} \\ \leq \sum_{n=0}^N 2^n x^{2^n} + 2\sum_{n=N+1}^\infty x^{2^{n-1}+1}+x^{2^{n-1}+2}+...+\underbrace{x^{2^{n-1}+2^{n-1}}}_{x^{2^n}} \\ =\sum_{n=0}^N 2^n x^{2^n} + 2\sum_{n=2^N+1}^\infty x^n = \sum_{n=0}^N 2^n x^{2^n} + \frac{2\,x^{2^N+1}}{1-x} \, .$$ For example, for $N=1$ you have $$f(x)\leq \frac{x(x+1)}{1-x} \, .$$

Aside from that, you may want to check these posts

  1. What's the limit of the series $\log_2(1-x)+x+x^2+x^4+x^8+\cdots$.
  2. Evaluate $\lim_{x\to1^-}\left(\sum_{n=0}^{\infty}\left(x^{(2^n)}\right)-\log_2\frac{1}{1-x}\right)$

which make statements about the behaviour of your function in the vicinity of $x=1$.

Since $$f(x)=\sum_{n=0}^\infty 2^n x^{2^n} = x \frac{{\rm d}}{{\rm d}x} \underbrace{\sum_{n=0}^\infty x^{2^n}}_{g(x)}$$

you may want to study $g(x)$ instead. $g(x)$ (and $f(x)$ alike) is obviously strictly increasing on $(0,1)$. Thus you may seek the behaviour as $x\rightarrow 1$. The first link shows that the function $$G(x)=g(x)+\log_2(1-x)$$ has a mean limit of $$\overline{G(1)}=\overline{\lim_{x\rightarrow 1^-} G(x)} = \frac{1}{2} - \frac{\gamma}{\log 2} \, .$$ However, $G(x)$ does not converge as $x\rightarrow 1$. If you write down the series expression for $G(1)$, you'll find it is still depending on $x$ and in the variable $t$, where $x=e^{-2^{-t}}$, it has a period of $1$. In the second link I calculate the fourier-expansion of $G$ in the variable $t$ with the result $$G(1)=\frac{1}{2}-\frac{\gamma}{\log 2} + \sum_{n=1}^\infty A_n \cos\left(2\pi nt- \varphi_n\right)$$ where \begin{align} A_n&=\frac{2\left|\Gamma\left(\frac{2\pi i \,n}{\log 2}\right)\right|}{\log 2}=\sqrt{\frac{2}{n\sinh\left(\frac{2\pi^2n}{\log 2}\right)\log 2}} \\ -\varphi_n&=\arg \left( \Gamma\left(\frac{2\pi i \,n}{\log 2}\right) \right) \, . \end{align}

Notice that the first harmonic is already a very good approximation, $A_2/A_1<10^{-6}$.

Thus in the vicinity of $x=1^-$ you can write $$g(x)=G(1)-\log_2(1-x) \\ f(x)=x \frac{{\rm d}}{{\rm d}x} g(x) = \frac{1}{\log 2} \frac{x}{1-x} - \frac{2\pi}{\log 2} \frac{-1}{\log x} \sum_{n=1}^\infty n A_n \sin\left(2\pi nt - \varphi_n \right) \\ \leq \frac{1}{\log 2} \frac{1}{1-x} \left( x + 2\pi \sum_{n=1}^\infty n A_n \right) < \frac{x+10^{-5}}{\log 2} \frac{1}{1-x}$$ with $\frac{-1}{\log x} \leq \frac{1}{1-x}$ for $0<x<1$.

Diger
  • 6,197