Diophantine approximation by negative continued fraction (Q6171970)

Given the classical continued fraction \(\alpha = [0;a_1,a_2,\ldots]\in [0,1)\) with convergents \(\frac{p_n}{q_n}\), the convergence of various entities that can be expressed by partial quotients or convergents are known, both in almost sure convergence (with respect to Lebesgue measure) and in measure (denoted by \(\xrightarrow{\lambda}\)). In this article, the author (re-)proves several analogous convergence results for for the negative continued fraction expansion \(\alpha = [1;a_1',a_2',\ldots]_{-1} := 1 - \frac{1}{a_1' - \frac{1}{a_2' - \frac{1}{a_3'\ldots}}} \in [0,1), (a_i' \geq 2)\) with convergents \(\frac{P_n}{Q_n}\). For the arithmetic mean, the author provides alternative proofs (Theorem 2.1) to show that \[ \frac{a_1' + a_2' +\ldots + a_n'}{n} \xrightarrow{\lambda} 3 \] (first proven by J. Aaronson [Ann. Probab. 14, 1037--1057 (1986; Zbl 0658.60050)]), but almost surely, \[ \liminf_{n \to\infty} \frac{a_1' + a_2' +\ldots + a_n'}{n} = 2, \quad \limsup_{n \to\infty} \frac{a_1' + a_2' +\ldots + a_n'}{n} \to \infty \] (first proven by \textit{J. Aaronson} and \textit{H. Nakada} [Stochastic Processes Appl. 104, No. 2, 173--192 (2003; Zbl 1075.60511)]). In particular, this shows that the arithmetic mean converges in measure, but not almost surely, which is in accordance with the behaviour for the classical continued fractions. The main results in this article consider the entities \(\frac{\log n}{n} \log \left\lvert x - \frac{P_n}{Q_n}\right\rvert\) and \(\frac{\log n}{n}\log Q_n\). The author proves that (Theorem 1.2) \[ \begin{split} \frac{\log n}{n}\log Q_n \xrightarrow{\lambda} \frac{\pi^2}{6}, \qquad \frac{\log n}{n} \log \left\lvert x - \frac{P_n}{Q_n}\right\rvert &\xrightarrow{\lambda} \frac{\pi^2}{6}, \end{split} \] but almost surely (Theorem 3.4 and Theorem 3.5), \begin{alignat*}{2} \\ &\liminf_{n \to\infty} \frac{\log n}{n}\log Q_n = 0, &&\limsup_{n \to\infty} \frac{\log n}{n}\log Q_n = \frac{\pi^2}{6}, \\ &\liminf_{n \to \infty} \frac{\log n}{n} \log \left\lvert x - \frac{P_n}{Q_n}\right\rvert = -\frac{\pi^2}{3}, \qquad &&\limsup_{n \to \infty} \frac{\log n}{n} \log \left\lvert x - \frac{P_n}{Q_n}\right\rvert = 0. \end{alignat*} So again, there is convergence in measure, but no almost sure convergence. This is in contrast to the classical continued fractions where we have an almost sure convergence in the form of \[ \begin{split} &\lim_{n \to \infty} \frac{\log q_n}{n} = \frac{\pi^2}{12 \log 2},\\ &\lim_{n \to \infty} \frac{1}{n} \log \left\lvert x - \frac{p_n}{q_n}\right\rvert = -\frac{\pi^2}{6\log 2}. \end{split} \] Note that in these limits, the scaling factor is \(\frac{1}{n}\) in comparison to \(\frac{\log n}{n}\) in the negative continued fraction case. This shows that the quality of approximation by negative continued fractions is usually much worse than when approximating with classical continued fractions. As a final result, a Gauss-Kuzmin-type analogue for negative continued fraction is examined. The author proves (Proposition 3.8) that almost surely, \[ \begin{split} \lim_{n \to \infty}\frac{1}{n}\#\{1 \leq i \leq n: a_i' = m\} = \begin{cases} 1 \text{ if } m = 2,\\ 0 \text{ otherwise}. \end{cases} \end{split} \] For \(m \geq 3\), \[ \begin{split} \frac{\log}{n}\#\{1 \leq i \leq n: a_i' = m\} &\xrightarrow{\lambda} \log\left(1 + \frac{1}{(m-2)m}\right). \end{split} \] This again is different to the classical case where for all \(m \in \mathbb{N}\) almost surely, \[ \begin{split} \lim_{n \to \infty}\frac{1}{n}\#\{1 \leq i \leq n: a_i' = m\} = \frac{1}{\log 2}\log\left(1 + \frac{1}{m(m+2)}\right). \end{split} \] Note that the ergodic measure \(\mu_{-1}\) for the operator that generates the negative continued fractions is not finite on \([0,1)\), hence a standard method via Birkhoff's ergodic theorem (that works for many results about classical continued fractions) is not directly applicable for negative continued fractions. Among other arguments, the proofs make use of a recoding formula that allows to translate a negative continued fraction in a classical one. Using several theorems from the classical world (such as Borel-Bernstein or Diamond-Vaaler), results about the negative continued fraction are deduced.