📊 statistics

A Spacing Estimator

This paper extends the known distribution of spacings between consecutive order statistics to logistic and Gumbel variates and introduces a general estimator for distributions with known inverse cumulative density functions, noting its high accuracy near the center but up to 20% degradation in the tails.

Original authors: Greg Kreider

Published 2026-01-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Greg Kreider

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 have a long, invisible ruler, and you are dropping $n$ tiny pebbles onto it at random. The spots where the pebbles land aren't evenly spaced; some clump together, and some have big gaps between them. In statistics, the distance between two neighboring pebbles is called a "spacing."

For a long time, mathematicians only had perfect, exact formulas to predict the average size of these gaps for two specific types of "rulers":

The Uniform Ruler: A ruler where every spot is equally likely to be hit (like throwing darts at a board).
The Exponential Ruler: A ruler where you are much more likely to hit the beginning, and the chances drop off quickly as you go further (like waiting for a bus that arrives randomly).

The Problem:
What if your ruler follows a different pattern? What if the pebbles are more likely to land in the middle (like a bell curve) or have a specific "heavy tail" where rare, extreme events happen? For these other shapes, calculating the exact average gap is incredibly hard. The math gets so messy that it often requires solving complex integrals that don't have simple answers.

The Solution (The "Spacing Estimator"):
Greg Kreider, the author of this paper, does two main things:

1. Cracking the Code for New Rulers

First, he successfully solved the math for two new, popular types of rulers: the Logistic and the Gumbel distributions.

The Challenge: The formulas he found are like giant, tangled knots of numbers. They involve massive factorials (multiplying huge numbers together) and long lists of terms that almost perfectly cancel each other out. To get the right answer, you need a super-precise calculator (high-precision math libraries) because if you round off even a tiny bit, the whole answer falls apart.
The Result: He provided the exact "blueprints" for the average gap size and the variation in gap size for these two specific distributions.

2. The "Shortcut" (The Quantile Estimator)

Since solving those giant knots of math is so difficult, Kreider proposes a clever shortcut for any distribution where you can reverse-engineer the probability (a distribution with an "invertible cumulative density function").

The Analogy:
Imagine you have a map of a city (the distribution). Instead of trying to calculate the exact distance between every house by measuring the winding roads (the hard math), you look at the map's grid lines.

You know that if you pick 100 random houses, they will roughly divide the city into 100 equal slices.
The shortcut simply asks: "If I take a slice of the city, how wide is it?"
It uses the inverse map (the quantile function) to estimate the gap. It's like saying, "If the map says this point is at 50% and the next is at 51%, the distance between them is just the difference in their map coordinates."

How Good is the Shortcut?
Kreider tested this shortcut against millions of computer simulations (dropping 100 million pebbles in a virtual world).

In the Middle: The shortcut is incredibly accurate. It's almost perfect for pebbles landing near the center of the distribution.
In the Tails: The shortcut gets a bit sloppy at the very edges (the tails) of the distribution. The error can grow to about 15% to 20%. It's like a GPS that is perfect for the city center but might be slightly off when you are driving out to the remote countryside.

The Verdict:

For Uniform and Exponential distributions, the shortcut is mathematically exact.
For Logistic distributions, the shortcut is also mathematically exact (and much simpler than the giant knot of formulas he derived earlier).
For Gumbel and other complex shapes, it is an approximation. It works very well for most practical purposes in the middle of the data, but you should be careful if you are looking at the extreme outliers.

In Summary:
This paper gives us the exact, complicated math for two new types of random patterns and offers a reliable, easy-to-use "GPS shortcut" for estimating gaps in almost any random pattern, warning us that the shortcut gets a little less accurate the further out you go from the center.

Technical Summary: A Spacing Estimator

Problem Statement
The distribution of spacings—the differences between consecutive order statistics—is a fundamental concept in statistical theory. While exact analytic expressions for the density, expected value, and variance of these spacings are well-established for uniform and exponential random variates, such solutions are generally unavailable for other distributions. For many distributions (e.g., normal, beta, $\chi^2$ , $t$ , Rayleigh, Weibull, Frechet), the cumulative density function (c.d.f.) lacks a closed form, or the shift $x+y$ within the spacing density integral does not separate, preventing the reduction of the problem to known integrals. Consequently, there is a need for reliable methods to estimate expected spacings and their variances for distributions where analytic solutions are intractable.

Methodology
The paper employs two primary approaches:

Analytic Derivation: For logistic and Gumbel distributions, the author derives the exact density of the spacing using Pyke's general formulation. By utilizing known integrals involving hypergeometric functions and performing complex series expansions, closed-form expressions for the expected spacing and the second moment (required for variance) are obtained. These derivations involve intricate summations where terms nearly cancel to counteract combinatorial explosion, necessitating high-precision mathematical libraries for evaluation.
Quantile Estimator: For distributions lacking analytic spacing solutions, the paper proposes a "quantile estimator." This method leverages the inverse c.d.f. (quantile function), $F^{-1}(p)$ . Since $n$ points drawn from a distribution map to $n$ equally spaced points in the unit interval $[0,1]$ , the expected spacing is approximated by the difference in the mapped points: $\tilde{E}\{D_i\} \approx F^{-1}(p_i) - F^{-1}(p_{i-1})$ . This is refined into a finite difference approximation of the derivative of the inverse c.d.f.: $\tilde{E}\{D_i\} \approx \frac{dF^{-1}(p)}{dp} \Delta p$ .

Key Contributions and Results

New Analytic Results: The paper provides the first analytic expressions for the spacing density, expected value, and variance for logistic and Gumbel variates.
- For the logistic distribution, the expected spacing is derived as a complex series involving factorials and summations. The variance is expressed as a closed-form solution involving infinite series of logarithms and dilogarithms ( $Li_2$ ).
- For the Gumbel distribution, an expression for the expected spacing is derived involving alternating series of logarithms. The paper notes that a closed-form expression for the variance is not possible with the current derivation.
Validation of the Quantile Estimator:
- The estimator is proven to be exact for exponential and logistic distributions, matching their respective analytic series solutions.
- For other distributions with invertible c.d.f.s (Cauchy, Pareto, Rayleigh, Weibull, Frechet), the estimator serves as an approximation.
Error Analysis: Extensive numerical simulations (100 million trials) were conducted to measure the approximation error ( $|\text{mean spacing} - \text{estimate}|$ $∣ mean spacing - estimate ∣$ ) across various distributions and sample sizes ( $n=25, 75, 250$ $n = 25, 75, 250$ ).
- Accuracy: The estimator is highly accurate for points near the middle of the order statistics. Errors scale inversely with the square of the sample size ( $1/n^2$ ). For small samples ( $n \le 25$ ), errors range from $10^{-3}$ to $10^{-2}$ ; for large samples ( $n \ge 100$ ), errors improve to $10^{-5}$ to $10^{-4}$ .
- Tail Degradation: Accuracy degrades in the tails of the distributions. For asymmetric distributions, errors can reach 15–20% in the tails farthest from the peak. For example, the Pareto estimator trails the mean by up to 18% in the right tail, and the Frechet estimator shows errors of 26% at the extreme tail ( $i=n$ ).
- Specific Behaviors: The Laplace distribution shows increased error near the center ( $i=n/2$ ) due to the non-differentiability of the density at the mode. The Cauchy distribution exhibits larger errors than other distributions due to its heavy tails.

Significance
The paper claims significance primarily in expanding the known analytic results for spacing distributions to include logistic and Gumbel variates, which previously lacked such closed forms. Furthermore, it establishes the quantile estimator as a robust, general-purpose tool for distributions where analytic spacing solutions are impossible. The study quantifies the limitations of this estimator, demonstrating that while it is accurate to the limit of numerical simulations for central order statistics, it suffers from significant degradation (up to 20%) in the tails of asymmetric distributions. The work provides a practical framework for estimating spacing statistics in machine vision and other fields requiring order statistic analysis, provided the user accounts for the increased error margins in distribution tails.

1. Cracking the Code for New Rulers

2. The "Shortcut" (The Quantile Estimator)

More like this