📊 statistics

On classes of distributions on the unit interval: structural properties and application to inequality data

This paper introduces two novel families of unit-interval distributions derived from gamma ratio transformations that not only provide closed-form statistical properties and maximum likelihood estimation methods but also establish a direct link to sample-based estimators of the Gini and Atkinson inequality indices, demonstrating their practical utility through simulation and real-world cross-country data analysis.

Original authors: Roberto Vila, Helton Saulo, Poliana Matos, Subhankar Dutta

Published 2026-03-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Roberto Vila, Helton Saulo, Poliana Matos, Subhankar Dutta

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 trying to describe the shape of a crowd. Are they all standing in a neat line? Are they clustered in one corner? Or are they spread out wildly? In the world of statistics, we use special mathematical "shapes" (called distributions) to describe how data behaves.

Most of the time, when we look at things that are measured as a percentage or a ratio—like the Gini Index (which measures how unequal a country's income is, ranging from 0% to 100%)—we use standard shapes like the Beta or Kumaraswamy distributions. Think of these as the "classic t-shirts" of statistics: they fit well in many situations, but sometimes they don't quite capture the unique quirks of the data.

This paper introduces two new, custom-tailored suits for this specific job. Here is the simple breakdown of what the authors did:

1. The Ingredients: Mixing Two "Gamma" Cakes

Imagine you have two bakers, X and Y. They both bake cakes, but the size of their cakes is random and follows a specific pattern called a "Gamma distribution."

Usually, statisticians look at the ratio of one cake to the total (X divided by X+Y). This gives them a standard Beta distribution.
The authors of this paper said, "What if we take that ratio and twist it?"

2. The Twist: The "Magic Mirror"

The authors invented two special "magic mirrors" (mathematical transformations) named $S_r$ and $T_r$ .

When you look at the ratio of the cakes through these mirrors, the image gets distorted in a very specific, non-linear way.
The Analogy: Imagine taking a straight line of people and folding it in half, then squishing the ends together. The people in the middle get pushed apart, and the people at the edges get pulled in. This creates a new, unique shape that the old "Beta" t-shirt couldn't make.

3. The "Aha!" Moment: Connecting to Inequality

Here is the coolest part of the paper. The authors found that if you set the "twist" parameter ( $r$ ) to a specific number (1/2), these new shapes literally become the formulas for famous inequality measures:

One shape becomes the Gini Coefficient (the most famous measure of wealth inequality).
The other becomes the Atkinson Index (another measure of inequality).

Why does this matter?
Before this, if you wanted to model inequality data, you had to use a generic shape and hope it fit. Now, the authors have built a statistical model that is born from the math of inequality itself. It's like building a house using the exact blueprints of the terrain it sits on, rather than trying to force a square peg into a round hole.

4. Testing the New Suits (Simulation)

The authors put these new models through a rigorous "stress test" using computer simulations (like a wind tunnel for statistics).

Model W (The $S_r$ mirror): This one was very stable. It handled different types of data well, even when the data was messy or skewed. It was like a reliable, sturdy pair of boots.
Model Z (The $T_r$ mirror): This one was a bit more sensitive and tricky to tune, especially when the data was very lopsided. It was like a high-performance sports car: fast and flexible, but harder to drive if you aren't an expert.

5. The Real-World Test: 2021 Global Income Data

Finally, they took these models out for a drive with real data: the Gini Index for 78 countries in 2021.

They compared their new models against the old "classic t-shirts" (Beta and Kumaraswamy).
The Result: The new Model Z was the clear winner. It fit the real-world data better than the others, capturing the specific "hump" and "tail" of how wealth is distributed across the globe.
The Catch: The new models have more "knobs" to turn (more parameters), which makes them harder to calculate, but they fit the data so much better that the extra effort was worth it.

The Bottom Line

This paper is about customizing the tools of the trade.
Instead of using a generic ruler to measure the complex, jagged landscape of global wealth inequality, the authors built a flexible, shape-shifting ruler that understands the very nature of inequality. They proved that by twisting the math of simple random numbers (Gamma distributions), you can create powerful new tools that describe the real world more accurately than ever before.

In short: They took a standard mathematical recipe, added a secret ingredient (a non-linear twist), and discovered that the resulting dish tastes exactly like the real-world problem of income inequality.

1. Problem Statement

Probability distributions supported on the unit interval $(0, 1)$ are essential in fields such as econometrics, reliability, and natural sciences. Traditional models like the Beta and Kumaraswamy distributions are widely used due to their flexibility. However, the authors argue that most existing models are phenomenological, meaning they are chosen for their shape-fitting capabilities but lack a deep theoretical connection to the stochastic mechanisms underlying classical economic inequality measures (e.g., the Gini and Atkinson indices).

The paper addresses the need for unit-interval distributions that:

Are derived from fundamental stochastic processes (specifically ratios of Gamma variables).
Provide a direct theoretical link to inequality indices.
Offer flexible shapes (including bimodality and skewness) through non-injective transformations.

2. Methodology

2.1 Model Construction

The authors define two new families of random variables, $W$ and $Z$ , by applying non-injective transformations to the ratio of two independent Gamma-distributed random variables, $X \sim \text{Gamma}(\alpha_1, \lambda_1)$ and $Y \sim \text{Gamma}(\alpha_2, \lambda_2)$ .

Let $U = \frac{X}{X+Y}$ . The authors introduce two transformation functions parameterized by $r > 0$ :

Transformation $S_r$ : $S_r(u) = [1 - 4u(1-u)]^r$
Transformation $T_r$ : $T_r(u) = 1 - [4u(1-u)]^r$

The new random variables are defined as:

$W = S_r\left(\frac{X}{X+Y}\right) = \left(\frac{Y-X}{X+Y}\right)^{2r}$
$Z = T_r\left(\frac{X}{X+Y}\right) = 1 - \frac{(XY)^r}{[\frac{1}{2}(X+Y)]^{2r}}$

Special Case ( $r=1/2$ ):
When $r=1/2$ , these variables correspond exactly to sample-based estimators of classical inequality indices for a sample of size $n=2$ :

$W = \frac{|X-Y|}{X+Y}$ corresponds to the Gini coefficient estimator.
$Z = 1 - \frac{\sqrt{XY}}{\frac{1}{2}(X+Y)}$ corresponds to the Atkinson index estimator.

2.2 Theoretical Derivations

The paper derives the complete probabilistic structure of $W$ and $Z$ :

Probability Density Functions (PDF) and Cumulative Distribution Functions (CDF): Closed-form expressions are derived using the density of the ratio $X/(X+Y)$ (which follows a generalized Beta distribution). The PDFs involve the function $G(x)$ , which combines terms from the underlying Gamma densities.
Quantile Functions: Implicit formulas are provided, requiring numerical inversion.
Raw Moments: General integral expressions for the $n$ $n$ -th raw moments are given. The authors derive specific closed-form analytical expressions for:
- The first moment of $W$ when $r=1/2$ and specific conditions on $\lambda_1, \lambda_2$ hold.
- The first moment of $Z$ when $r=1/2$ .
- The $n$ -th moment for both models when $r \in \mathbb{N}$ and $\lambda_1 = \lambda_2$ .
Maximum Likelihood Estimation (MLE): The log-likelihood functions for both models are formulated. The partial derivatives with respect to parameters $(\alpha_1, \alpha_2, \lambda_1, \lambda_2, r)$ are derived, leading to a system of non-linear equations solved numerically (using the L-BFGS-B algorithm).

3. Key Contributions

Novel Distribution Families: Introduction of two 5-parameter families ( $W$ and $Z$ ) on the unit interval derived from Gamma ratios via non-injective transformations.
Theoretical Link to Inequality: Establishing a rigorous mathematical connection where specific parameter settings ( $r=1/2$ ) recover the Gini and Atkinson indices, moving beyond purely phenomenological modeling.
Analytical Tractability: Providing closed-form expressions for PDFs, CDFs, and specific moments, which is rare for complex transformations of Gamma ratios.
Comprehensive Estimation Framework: Deriving the likelihood equations and discussing numerical strategies for MLE.

4. Results

4.1 Simulation Study

A Monte Carlo simulation study (500 replications) was conducted to assess the finite-sample performance of the MLEs for various sample sizes ( $n=50, 100, 250, 500$ ) and parameter configurations.

Model $W$ : Exhibited globally stable performance. Shape and scale estimators showed progressive convergence with increasing $n$ . Bias and RMSE decreased systematically. The transformation parameter $r$ was recovered with good accuracy.
Model $Z$ : Proved more challenging to estimate. The transformation $T_r$ reduces information concentration, leading to higher sensitivity, particularly in asymmetric scenarios or when shape parameters are low ( $<1$ ). While scale parameters were estimated well, shape parameters ( $\alpha_1, \alpha_2$ ) and the transformation parameter $r$ exhibited higher bias and RMSE, even at $n=500$ .
Comparison: Model $W$ is numerically more robust and suitable for a broader range of configurations, whereas Model $Z$ requires larger samples and more regular (symmetric) data structures for precise parameter identification.

4.2 Real Data Application

The models were applied to 2021 Gini index data from 78 countries.

Competitors: The new models were compared against the standard Beta and Kumaraswamy distributions.
Performance Metrics:
- Model $Z$ achieved the highest log-likelihood ( $\approx 108.96$ ) and the lowest AIC ($-207.93$), indicating the best fit to the data's skewness and tail behavior.
- Beta Distribution ranked second in AIC but was favored by BIC ($-201.09$) due to its parsimony (fewer parameters).
- Model $W$ showed intermediate performance.
- Kumaraswamy performed the worst, with numerical instability in parameter estimation.
Diagnostic Plots: Quantile-Quantile (Q-Q) plots of randomized and Cox-Snell residuals confirmed that Models $Z$ and Beta provided the best adherence to the empirical distribution, particularly in the tails.

5. Significance and Conclusion

The paper makes a significant contribution to the literature on unit-interval distributions by bridging the gap between stochastic theory (Gamma ratios) and applied economics (inequality measurement).

Practical Relevance: The models offer a flexible alternative to Beta and Kumaraswamy distributions, particularly for data exhibiting complex shapes or specific connections to inequality metrics.
Theoretical Insight: The derivation of the Gini and Atkinson indices as specific cases of these distributions validates the use of these models for analyzing inequality data, providing a mechanistic justification rather than just a curve-fitting exercise.
Future Directions: The authors suggest extending these models to correlated Gamma components, multivariate settings, and Bayesian estimation frameworks.

In summary, the proposed distributions, particularly Model $Z$ , demonstrate superior flexibility and fit for real-world inequality data, while the theoretical framework provides a robust foundation for understanding the stochastic nature of inequality indices.