Strange property of irrational numbers, linked to continued fractions

113 Views Asked by At

Convergents of continued fractions provide in some sense the best rational approximations to an irrational number. In my research trying to prove that the binary digits of $\sqrt{2}$ are 50% one, 50% zero, I came up with something interesting.

I analyzed the successive records (minima) of $w_n = |\{n\alpha\} - \beta|$ where $\{n\alpha\}$ is $n\alpha$ modulo $1$ (the brackets represent the fractional part function,) $n =1, 2$ and so on, $\beta = 1/2$ and $\alpha = \log_2 3$. I found how the arrival times $t_n$ of these records are related to the denominators of the convergents of $\alpha$. More specifically, the $\Delta t_n$ are linked to these convergents, and also relevant to musical theory. My findings are summarized here.

I did some more reseach and found that something rather similar happens with $\beta = e^{-1}$, and probably with almost all $\beta \in [0, 1]$. I posted it as an exercise:

Exercise:

Let $\alpha = \log_2 3$ and $w_n = | \{ n\alpha \} - 1/2 |$. Find the arrival times of the first records for $w_n$, focusing on minima rather than maxima. Let $t_n$ denotes the arrival time of the $n$-th record. The sequence $\Delta t_n = t_n - t_{n-1}$ is a subsequence of the one listed here.

Also,

  • $\Delta t_6 = \Delta t_7 = \cdots = \Delta t_{17} = 665$, and
  • $\Delta t_{20} = \Delta t_{21} = \cdots = \Delta t_{47} = 190537$.

Now do the same with $w_n = | \{n \alpha\} - 1/e |$. The same $\Delta t_n$ show up, in particular

  • $\Delta t_9 = \Delta t_{10} = \cdots = \Delta t_{17} = 665$, and
  • $\Delta t_{19} = \Delta t_{20} = \cdots = \Delta t_{38} = 190537$.

However $\Delta t_8 = 4296$ is unusually large. Is there a pattern here? Does it generalize to any $\beta \in [0, 1]$? Is there an explanation for this?

Note

This problem can be re-phrased as follows: we are trying to find the best approximation of the form $p\alpha - q$ for the number $\beta$, with $p, q$ being strictly positive integers. The $n$-th best approximant for $\beta$ is $\{t_n \alpha\}$, and of course $\{t_n\alpha\} \rightarrow \beta$.