---
title: Doeblin–Lenstra Law in Diophantine Approximation
url: https://www.emergentmind.com/topics/doeblin-lenstra-law
type: topic
---

# Doeblin–Lenstra Law in Diophantine Approximation

Searching arXiv for the specified papers to ground the article in current sources.
The Doeblin–Lenstra law is the limiting distribution law for the approximation coefficients of continued-fraction convergents of a typical irrational real number. In the normalization used in recent work, if $\theta\in\mathbb R\setminus\mathbb Q$ has convergents $p_k/q_k$, then the relevant statistic is
\[
q_k|\theta q_k-p_k|=q_k^2\left|\theta-\frac{p_k}{q_k}\right|,
\]
and for Lebesgue-almost every $\theta\in(0,1)$ the empirical distribution of these quantities converges to a probability measure $\nu$ on $[0,1]$. Recent work establishes effective convergence rates in this law, proves central limit theorems for the associated Diophantine statistics, and extends the framework both to higher-dimensional best approximants and to certain fractal measures, including the middle-third Cantor measure [2507.19084].

## 1. Classical statement

Let
\[
\theta = a_0 + \frac{1}{a_1+\frac{1}{a_2+\cdots}}, \qquad a_0\in\mathbb Z,\ a_j\in\mathbb N \ (j\ge 1),
\]
and let
\[
\frac{p_k}{q_k} = a_0 + \frac{1}{a_1+\frac{1}{a_2+\ddots+\frac{1}{a_k}}}
\]
be the $k$-th convergent. The approximation coefficient used in the modern formulation is
\[
q_k|\theta q_k-p_k|,
\]
which coincides with the more classical quantity
\[
\Theta_k := q_k^2\left|\theta-\frac{p_k}{q_k}\right|.
\]

Doeblin’s 1938 result is stated as follows: for Lebesgue-almost every $\theta\in(0,1)$,
\[
\lim_{N\to\infty}\frac1N\sum_{k=1}^N F\!\left(q_k|\theta q_k-p_k|\right) = \int_0^1 F(z)\,d\nu(z)
\]
for every bounded continuous $F:[0,1]\to\mathbb R$, where $\nu$ has density
\[
\frac{d\nu}{dz} = \begin{cases}
\displaystyle \frac{1}{\ln 2}, & 0\le z\le \frac12,\\[6pt]
\displaystyle \frac{1-1/z}{\ln 2}, & \frac12< z\le 1.
\end{cases}
\]
Equivalently, the empirical measures
\[
\frac1N\sum_{k=1}^N \delta_{\Theta_k}
\]
converge weak-* to $\nu$ for almost every $\theta$ [2507.19084].

The cumulative distribution function corresponding to this density is
\[
\nu([0,t])= \begin{cases}
\displaystyle \frac{t}{\ln 2}, & 0\le t\le \frac12,\\[8pt]
\displaystyle \frac{t-\ln t -1+\ln 2}{\ln 2}, & \frac12<t\le 1.
\end{cases}
\]

In this formulation, “typical” means Lebesgue-almost every, and the convergence is Cesàro convergence of observables along convergents. The law is therefore not merely a pointwise statement about a single subsequence of approximants, but an ergodic statement about the asymptotic empirical distribution of the full sequence of continued-fraction approximation coefficients.

## 2. Historical development and the conjectural effective form

The historical trajectory recorded in recent work has three stages. First, Doeblin sketched the law in 1938. Second, Lenstra independently conjectured the same phenomenon. Third, Bosma–Jager–Wiedijk proved it in 1983 via ergodic theory of the natural extension of the Gauss map [2507.19084].

What remained open was an effective version: quantitative convergence rates and fluctuation laws. In that sense, the “Doeblin–Lenstra conjecture” treated in recent work is the problem of obtaining explicit asymptotics beyond qualitative convergence. The recent paper characterizes its contribution as the first quantitative and dynamical treatment of the Doeblin–Lenstra law, with effective error terms and central limit theorems, and with extensions both to higher-dimensional best approximants and to points sampled from certain fractal measures [2507.19084].

A central distinction in the modern literature is that the Doeblin–Lenstra law belongs to metrical number theory, continued fractions, and Diophantine approximation, whereas other uses of the name “Doeblin” arise from Markov kernels, contraction coefficients, and strong data-processing inequalities. A recent quantum-information paper explicitly notes that its “Doeblin coefficient” is not the Doeblin–Lenstra law and is instead rooted in Doeblin’s work on Markov chains and stochastic kernels [2503.22823]. This distinction is bibliographically substantive: the Doeblin–Lenstra law is associated with continued fractions and Euclidean-algorithm statistics, not with channel ergodicity coefficients.

## 3. Effective convergence rates and central limit theorems

The principal effective one-dimensional result states that if $F:\mathbb R\to\mathbb R$ is differentiable with bounded derivative, then for any $\varepsilon>0$, for Lebesgue-almost every $\theta$,
\[
\frac{1}{N}\sum_{k=1}^N F\bigl(q_k|\theta q_k-p_k|\bigr) = \int_0^1 F(z)\,d\nu(z) + O\!\left(N^{-1/2}\log^{3/2+\varepsilon}N\right).
\]
Thus the classical limiting law is strengthened to an almost-sure effective asymptotic with rate
\[
O\!\left(N^{-1/2}\log^{3/2+\varepsilon}N\right)
\]
[2507.19084].

The paper proves a more general counting theorem in logarithmic “flow time” $T$. If
\[
F_F(\theta,T) := \sum_{\substack{(p,q)\text{ best approximate}\\ \|q\|<e^T}}
F\!\left( \|p+\theta q\|^m\|q\|^n,\,
\frac{p_k+\theta q_k}{\|p_k+\theta q_k\|},\,
\frac{q_k}{\|q_k\|} \right),
\]
then there exist $\gamma,\sigma>0$ such that for any $\varepsilon>0$, for $\mu$-almost every $\theta$,
\[
F_F(\theta,T)=\gamma T + O\!\left(T^{1/2}\log^{3/2+\varepsilon}T\right),
\]
and a central limit theorem holds:
\[
\mu\!\left(
\left\{\theta:
\frac{F_F(\theta,T)-\gamma T}{T^{1/2}}<\xi
\right\}
\right)\to \mathrm{Norm}_\sigma(\xi)
\quad (T\to\infty).
\]

The central limit theorem applies to weighted counts of best approximants. Choosing $F\equiv 1$ yields a CLT for the number of best approximants up to logarithmic scale $T$, while general $F$ gives a CLT for weighted counts according to approximation quality and direction. The variance is given by the Green–Kubo-type sum
\[
\sigma^2 = \sum_{s\in\mathbb Z} \left( \int_X f(a_s\Lambda)f(\Lambda)\,d\mu_X(\Lambda)-\mu_X(f)^2 \right).
\]

A plausible implication is that the effective theory is structurally stronger than the classical law in two distinct senses: it quantifies the rate of convergence of empirical distributions, and it identifies Gaussian fluctuation behavior for naturally associated counting statistics.

## 4. Dynamical reformulation on spaces of lattices

The modern proof strategy reformulates the problem in homogeneous dynamics. The ambient space is
\[
G=\mathrm{SL}_{m+n}(\mathbb R),\qquad \Gamma=\mathrm{SL}_{m+n}(\mathbb Z),\qquad X=G/\Gamma,
\]
identified with the space of unimodular lattices in $\mathbb R^{m+n}$ [2507.19084].

For $\theta\in M_{m\times n}(\mathbb R)$ and $t\in\mathbb R$, the relevant matrices are
\[
u(\theta)= \begin{pmatrix} I_m & \theta\\ 0 & I_n \end{pmatrix}, \qquad
a_t= \begin{pmatrix} e^{\frac{n}{m}t}I_m & 0\\ 0 & e^{-t}I_n \end{pmatrix}.
\]
The lattice $u(\theta)\Gamma$ encodes the Diophantine data of $\theta$, and the diagonal flow $a_t$ rescales error and denominator coordinates so that best approximants appear as primitive lattice vectors in a fixed window.

The lattice observable is defined by
\[
f(\Lambda)=\sum_{v\in S_\Lambda}(F\circ\phi)(v),
\qquad
\phi(x,y)= \left( \|x\|^m\|y\|^n,\ \frac{x}{\|x\|},\ \frac{y}{\|y\|} \right),
\]
where $S_\Lambda$ consists of primitive vectors satisfying a shell condition and a box-minimality condition. The exact one-shell correspondence is
\[
F_F(\theta,l,l+1)=f(a_lu(\theta)\Gamma),
\]
and more generally
\[
\sum_{i=1}^{\lfloor T\rfloor} f(a_i u(\theta)\Gamma) \le F_F(\theta,T)\le \sum_{i=1}^{\lfloor T\rfloor+1} f(a_i u(\theta)\Gamma).
\]

This correspondence turns the Doeblin–Lenstra law into an ergodic theorem for Birkhoff sums of $f$ along the diagonal orbit $a_tu(\theta)\Gamma$. In dimension one, taking $F(z,\omega_1,\omega_2)=\widetilde F(z)$ recovers the scalar approximation-coefficient law. The significance of this reformulation is methodological: it replaces cross-section and natural-extension methods by a diagonal-flow framework that is adapted to quantitative estimates.

## 5. Discontinuous observables and effective ergodic inputs

A central technical issue is that the observable
\[
f(\Lambda)=\sum_{v\in S_\Lambda}(F\circ\phi)(v)
\]
is bounded but not continuous. The discontinuity arises because vectors may enter or leave the shell
\[
1\le \|\pi_2(v)\|<e,\qquad \|\pi_1(v)\|\le 1,
\]
and because small perturbations may alter minimality relations in the box order [2507.19084].

The abstract criterion used in the proof requires “average regularity under perturbations.” One must construct nonnegative measurable $\tau_\varepsilon$ such that
\[
\int_X \tau_\varepsilon\,d\mu_X \le C\varepsilon
\]
and such that for $g$ in an $\varepsilon$-ball around the identity,
\[
|f(g\Lambda)-f(\Lambda)|\le \tau_\varepsilon(\Lambda).
\]
The perturbation bound is reduced to auxiliary indicator functions $\varphi_\delta$ and $\Phi_\delta$, where $\varphi_\delta$ detects vectors near the boundary of the shell/window and $\Phi_\delta$ detects pairs of primitive vectors near tie situations.

The necessary average bounds are then verified by geometry of numbers. Specifically, Siegel’s mean value theorem controls
\[
\int_X \sum_{v\in\Lambda_{\mathrm{prim}}}\varphi_\varepsilon(v)\,d\mu_X,
\]
and Rogers’ second moment formula controls
\[
\int_X \sum_{\substack{v,w\in\Lambda_{\mathrm{prim}}\\ w\neq\pm v}}\Phi_\varepsilon(v,w)\,d\mu_X.
\]
Both contributions are shown to be $O(\varepsilon)$.

This establishes that the discontinuous observable is sufficiently regular on average for effective Birkhoff theorems and central limit theorems to apply. The paper emphasizes that this average-regularity mechanism is one of the technically central ideas of the argument.

## 6. Higher-dimensional and fractal extensions

The higher-dimensional theory replaces continued-fraction convergents by best approximants for $\theta\in M_{m\times n}(\mathbb R)$. A pair
\[
(p,q)\in \mathbb Z^m\times (\mathbb Z^n\setminus\{0\})
\]
is a best approximation if there is no other $(p',q')\neq (\pm p,\pm q)$ such that
\[
\|p'+\theta q'\|_{\mathbb R^m}\le \|p+\theta q\|_{\mathbb R^m}
\quad\text{and}\quad
\|q'\|_{\mathbb R^n}\le \|q\|_{\mathbb R^n}.
\]
The scalar analogue of the classical approximation coefficient is
\[
\|q\|_{\mathbb R^n}^n\,\|p+\theta q\|_{\mathbb R^m}^m.
\]

For a function
\[
F:\mathbb R_{\ge 0}\times S^m\times S^n\to\mathbb R
\]
with bounded first derivative, there exists $\beta>0$ such that for any $\varepsilon>0$ and for $\mu$-almost every $\theta$,
\[
\frac1N\sum_{k=1}^N F\!\left( \|p_k+\theta q_k\|^m\|q_k\|^n,\,
\frac{p_k+\theta q_k}{\|p_k+\theta q_k\|},\,
\frac{q_k}{\|q_k\|} \right)
= \beta + O\!\left(N^{-1/2}\log^{3/2+\varepsilon}N\right),
\]
provided $\mu$ satisfies Condition (EMEI). In the scalar higher-dimensional specialization,
\[
\frac1N\sum_{k=1}^N F\bigl(\|q_k\|_{R^n}^n\,\|p_k+\theta q_k\|_{R^m}^m\bigr)
=
\int_{\mathbb R}F(z)\,d\nu_{m,n}(z) + O\!\left(N^{-1/2}\log^{3/2+\varepsilon}N\right)
\]
for Lebesgue-almost every $\theta$ [2507.19084].

The same paper extends the law to self-similar measures on $\mathbb R$ generated by similarities with a common contraction ratio. In particular, if $\mu$ is a non-atomic self-similar measure on $\mathbb R$, then for differentiable $F$ with bounded derivative and any $\varepsilon>0$, for $\mu$-almost every $\theta$,
\[
\frac1N\sum_{k=1}^N F\bigl(q_k|\theta q_k-p_k|\bigr) = \int_0^1F(z)\,d\nu(z) + O\!\left(N^{-1/2}\log^{3/2+\varepsilon}N\right),
\]
and the same CLT mechanism applies. The middle-third Cantor measure is explicitly included in this class.

The dynamical input is Condition (EMEI), an effective multi-correlation estimate of the form
\[
\int F_0(\theta)\prod_{i=1}^r F_i(a_{t_i}u(\theta)\Gamma)\,d\mu(\theta)
=
\mu(F_0)\prod_{i=1}^r \mu_X(F_i)
+
O_r\!\left(
e^{-\delta D(t_1,\dots,t_r)}
\|F_0\|_{C^k}\prod_{i=1}^r \|F_i\|_{C^k}
\right),
\]
where
\[
D(t_1,\dots,t_r)=\min\{t_i,\ |t_i-t_j|: i\neq j\}.
\]
Lebesgue measure on a compact box satisfies this condition by earlier homogeneous-dynamics results, and the recent paper proves it for the relevant class of self-similar measures by upgrading effective single equidistribution to effective multi-equidistribution [2507.19084].

## 7. Scope, significance, and disambiguation

The present state of the subject may be summarized in three layers. At the classical level, the Doeblin–Lenstra law identifies the limiting distribution of
\[
\Theta_k=q_k^2\left|\theta-\frac{p_k}{q_k}\right|
\]
for convergents of a typical irrational. At the quantitative level, recent work proves the first effective convergence rate
\[
O\!\left(N^{-1/2}\log^{3/2+\varepsilon}N\right)
\]
and establishes central limit theorems for weighted counts of best approximants. At the structural level, the law is now embedded in a homogeneous-dynamics framework that also covers higher-dimensional best approximation and self-similar fractal measures [2507.19084].

The significance of these developments lies in the unification of several previously separate themes: one-dimensional continued fractions, higher-dimensional Diophantine approximation, and equidistribution of fractal measures under diagonal flows. A plausible implication is that the Doeblin–Lenstra law is best viewed not only as a result about continued fractions, but as a manifestation of a broader correspondence between Diophantine approximation statistics and ergodic averages on spaces of lattices.

It is also important to distinguish the Doeblin–Lenstra law from other mathematical objects carrying Doeblin’s name. A recent paper on quantum channels makes this explicit: its “Doeblin coefficients” concern channel contraction, minorization, and ergodicity coefficients, and it states that this is not the Doeblin–Lenstra law from continued fractions, Euclidean algorithms, or number theory [2503.22823]. The two topics therefore belong to different branches of Doeblin’s legacy: one in metrical number theory and homogeneous dynamics, the other in Markov kernels, information theory, and quantum channels.

Source: https://www.emergentmind.com/topics/doeblin-lenstra-law