🔢 mathematics

Quantitative longest-run laws for partial quotients

This paper establishes a general theorem for the quantitative longest-run statistics of fixed and arbitrary values under mixing conditions, applying it to derive explicit almost-sure logarithmic growth laws and precise double-logarithmic error bounds for continued-fraction partial quotients.

Original authors: Ying Wai Lee

Published 2026-02-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ying Wai Lee

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are watching a never-ending stream of numbers generated by a mysterious machine. This machine spits out integers one by one, like a conveyor belt of lottery balls. In the world of mathematics, this is called a continued fraction, and it's a way of representing any irrational number (like $\pi$ or $\sqrt{2}$ ) as a sequence of whole numbers.

The paper you provided is a detective story about finding patterns in this chaotic stream. Specifically, the author, Ying Wai Lee, is asking a very simple question: "How long can we expect to see the same number repeated over and over again?"

Here is the breakdown of the paper using everyday analogies:

1. The Two Types of "Streaks"

The paper looks at two different ways to count these repetitions (called "runs"):

The Specific Streak (Fixed Value): Imagine you are betting on the number 7. You watch the stream and ask, "What is the longest time I saw the number 7 appear back-to-back?"
- Example: If the stream is ... 3, 7, 7, 7, 2, 5, 7, 7, ..., your streak for the number 7 is 3.
The "Hot Hand" Streak (Maximized Value): Imagine you don't care which number repeats, you just want to know: "What is the longest streak of any number that happened?"
- Example: In the stream ... 3, 7, 7, 7, 2, 5, 5, 5, 5, 9 ..., the longest streak is 4 (because the number 5 repeated four times).

2. The Old Rules vs. The New Discovery

For a long time, mathematicians knew the rough answer. They knew that if you watch the stream for a very long time (let's say $n$ numbers), the length of the longest streak grows roughly like the logarithm of $n$ .

Think of it like this: If you watch the stream for 100 numbers, the longest streak might be 5. If you watch for 1,000 numbers, it might be 7. If you watch for 1,000,000, it might be 10. It grows, but very slowly.

The Problem: The old rules were like a weather forecast saying, "It will be sunny." They told us the general trend, but they didn't tell us how accurate that prediction was. They didn't say, "It will be sunny, give or take 2 degrees."

The New Discovery: This paper provides a quantitative forecast. It doesn't just say the streak grows; it gives a precise "error margin."

It says: "The longest streak will be exactly $X$ , plus or minus a tiny amount that grows very slowly (like the double-logarithm)."
It's like upgrading from a vague weather report to a hyper-precise GPS that tells you exactly how many seconds you will be late, down to the millisecond.

3. The "Magic Machine" (The Gauss System)

The numbers in this stream aren't random like a coin flip. They are generated by a specific mathematical rule called the Gauss Map.

The Analogy: Imagine a ball bouncing inside a curved bowl. The path it takes is deterministic (you can calculate it), but it looks chaotic and unpredictable.
The Challenge: Because the numbers aren't truly random (if you get a 7, it slightly changes the odds of getting the next number), standard probability rules don't work perfectly.
The Solution: The author proves that even though the ball is bouncing in a curved bowl, the "mixing" is so fast that after a short distance, the numbers act almost like they are independent. This allows the math to work.

4. The "Double-Logarithmic" Error

The paper's biggest achievement is calculating the error term.

Imagine you are trying to guess the height of a growing tree.
Old Math: "The tree will be about 100 feet tall."
This Paper: "The tree will be 100 feet tall, plus or minus a tiny bit that is roughly the size of a leaf."
The "double-logarithmic" part is just a fancy way of saying the error is extremely small. It grows so slowly that even if you watched the stream for the entire age of the universe, the error would still be tiny.

5. Why Does This Matter?

You might wonder, "Who cares about repeating numbers in a math sequence?"

In Nature: This helps us understand how "random" or "predictable" complex systems are.
In Computing: It helps in data compression and cryptography (hiding information).
In Math: It connects three different worlds: Probability (chance), Dynamical Systems (how things move and change), and Number Theory (properties of numbers).

Summary

Think of this paper as a precision ruler for chaos.
Previously, we knew that in a chaotic stream of numbers, long streaks of repetition happen, and we knew roughly how long they were. Ying Wai Lee has now built a ruler that measures those streaks with extreme precision, accounting for the fact that the numbers aren't perfectly random.

The result is a formula that tells you, with near-perfect certainty, exactly how long the longest streak of a specific number (or any number) will be, no matter how long you watch the stream. It turns a vague guess into a hard, mathematical fact.

1. Problem Statement

The paper investigates the statistical behavior of longest runs in the sequence of partial quotients of continued fractions. For an irrational number $x \in [0, 1)$ , let $(a_n)_{n \in \mathbb{N}}$ be its sequence of partial quotients. The author defines two primary statistics:

Fixed-symbol longest run ( $L_n(x, \lambda)$ ): The maximum length of a consecutive run of a specific integer value $\lambda$ within the first $n$ partial quotients.
Maximised longest run ( $R_n(x)$ ): The maximum length of a consecutive run of any integer value within the first $n$ partial quotients.

Context & Gap:

Previous work by Song and Zhou (2018) and Wang and Wu established first-order logarithmic laws (limit theorems) for these statistics. Specifically, they showed that $L_n(x, \lambda) \sim \frac{\log n}{2 \log \tau(\lambda)}$ and $R_n(x) \sim \frac{\log n}{2 \log \varphi}$ almost everywhere, where $\tau(\lambda)$ and $\varphi$ are specific constants.
The Gap: These previous results only described the asymptotic ratio (growth rate). They did not provide quantitative error bounds or explicit convergence rates. The paper aims to refine these laws by establishing eventual two-sided additive bounds with explicit double-logarithmic error terms.

2. Methodology

The author employs a probabilistic approach within the framework of ergodic theory and dynamical systems, specifically utilizing the Gauss map and Gauss measure.

A. General Theoretical Framework

The paper first establishes a general theorem (Theorem 1) for shift-invariant symbolic processes $(\Omega, \mathcal{F}, \mu, T)$ with a countable alphabet. The proof relies on three key assumptions:

Quantitative Mixing (Assumption 1): The system must satisfy a strong mixing condition with an exponential decay rate ( $\theta^g$ ) for events separated by a gap $g$ . This allows treating separated blocks as nearly independent.
Exponential Cylinder Estimates (Assumption 2): For a fixed symbol $m^*$ , the measure of the cylinder set where the symbol repeats $k$ times ( $\Delta_k(m^*)$ ) must decay exponentially as $c \rho^{-2k}$ .
Uniform Summed Bound (Assumption 3): The sum of measures of constant-word cylinders over all symbols must also decay exponentially.

B. Proof Techniques

Upper Bounds: Established using the Borel-Cantelli Lemma. By estimating the probability that a run of length $k$ occurs and summing over time, the author shows that runs exceeding a certain threshold occur only finitely often.
Lower Bounds: Established using separated trials and the mixing assumption. The author constructs independent (or nearly independent) trials separated by gaps large enough to satisfy the mixing condition, ensuring that runs of a certain minimum length occur infinitely often.
Scaling: The proof utilizes a dyadic subsequence ( $n_j = 2^j$ ) to interpolate between discrete steps and derive bounds for all $n$ .

C. Application to Continued Fractions

The general theorem is applied to the Gauss system (the dynamical system generating continued fractions):

Mixing: The Gauss map is known to be exponentially $\psi$ -mixing (Lemma 1).
Cylinder Estimates: The author derives explicit bounds for the measure of cylinders where a specific digit $\lambda$ $λ$ repeats $k$ $k$ times. This involves analyzing the denominators of convergents and the geometry of continued fraction intervals (Lemmas 2 and 3).
- For a fixed $\lambda$ , the decay rate is governed by $\tau(\lambda) = \frac{\lambda + \sqrt{\lambda^2+4}}{2}$ .
- For the maximised run, the decay rate is governed by the golden ratio $\varphi = \frac{1+\sqrt{5}}{2}$ , corresponding to the most probable digit ( $\lambda=1$ ).

3. Key Contributions

Quantitative Refinement of Limit Laws: The paper moves beyond asymptotic ratios to provide explicit two-sided additive bounds. It proves that for almost every $x$ , the longest run length is concentrated around the logarithmic term with an error of order $O(\log \log n)$ .
Explicit Constants: The paper derives explicit centering constants and error coefficients involving $\log \log n$ , sharpening the results of Song-Zhou and Wang-Wu.
General Theorem for Countable Alphabets: Theorem 1 provides a robust template for analyzing longest runs in systems with countable alphabets under mixing conditions, which is more complex than the finite alphabet case due to the need for uniform summation bounds (Assumption 3).
Verification for Gauss Measure: The paper rigorously verifies that the Gauss system satisfies the necessary mixing and exponential decay conditions, bridging abstract probability theory with metric number theory.

4. Main Results

The paper proves the following quantitative laws (Theorem 2):

For Fixed Symbol $\lambda$ :
For any $c > 1/2$ and almost every $x$ , there exists $n_{c,x}$ such that for all $n \geq n_{c,x}$ :
$\left| L_n(x, \lambda) - \frac{\log n}{2 \log \tau(\lambda)} \right| \leq c \frac{\log \log n}{\log \tau(\lambda)}$

For Maximised Run:
For any $c > 1/2$ and almost every $x$ , there exists $n_{c,x}$ such that for all $n \geq n_{c,x}$ :
$\left| R_n(x) - \frac{\log n}{2 \log \varphi} \right| \leq c \frac{\log \log n}{\log \varphi}$

Where:

$\tau(\lambda) = \frac{\lambda + \sqrt{\lambda^2+4}}{2}$
$\varphi = \frac{1+\sqrt{5}}{2}$ (the golden ratio)
The error term is double-logarithmic ( $\log \log n$ ), which is significantly sharper than previous non-quantitative results.

5. Significance

Metric Number Theory: The results provide a deeper understanding of the "persistence" of digits in continued fraction expansions. While magnitude-based statistics (like the growth rate of partial quotients) are well-studied, this work quantifies the consecutive repetition of digits.
Dynamical Systems: It demonstrates how quantitative mixing properties translate into precise statistical laws for non-additive observables (longest runs are non-additive).
Methodological Advance: The derivation of double-logarithmic error bounds for systems with countable alphabets is a technical achievement, as the infinite nature of the alphabet requires careful handling of the "maximised" statistic to ensure the sum of probabilities converges appropriately.
Universality: The framework suggests that similar quantitative laws can be derived for other dynamical systems satisfying the mixing and cylinder estimate conditions, extending the applicability of Erdős–Rényi type laws to dependent, infinite-alphabet processes.