The arithmetic-harmonic inequality index: Theory, inference, and finite-sample analysis

1. Problem Statement and Motivation

The paper addresses the need for a bounded, scale-invariant measure of dispersion for positive random variables. While classical measures like the Coefficient of Variation (CV) are widely used, they are unbounded and can be difficult to interpret in extreme cases of variability.

The authors focus on the Arithmetic-Harmonic Inequality (AHI) index, defined for a positive random variable $X$ with finite mean $E[X]$ and finite reciprocal mean $E[1/X]$ as:
$$J = 1 - \frac{1}{E[X]E[1/X]}$$
This index quantifies the deviation of the average ratio of independent copies of $X$ from unity. It is bounded within $[0, 1)$, where $J=0$ implies a degenerate distribution (no dispersion), and $J \to 1$ as dispersion increases. The paper aims to:

1. Derive analytical expressions for $J$ under specific distribution families.
2. Develop a rigorous statistical inference framework for the plug-in estimator $\hat{J}$.
3. Analyze the finite-sample properties (bias and variance) of the estimator.

2. Methodology

Theoretical Derivation

The authors investigate the behavior of $J$ under the Generalized Inverse Gaussian (GIG) family, which serves as a unifying framework. They derive closed-form expressions for $J$ using properties of modified Bessel functions of the second kind ($K_p$).

• Special Cases: The analysis specifically covers the Inverse Gaussian (IG) and Gamma distributions as limiting or special cases of the GIG family.
• Approximation: A second-order Taylor expansion is used to show that for small variability, $J \approx \frac{CV^2}{1+CV^2}$, establishing a direct link to the squared coefficient of variation.

Estimator Construction

Given an i.i.d. sample $X_1, \dots, X_n$, the plug-in estimator is defined as:
$$\hat{J} = 1 - \frac{1}{\bar{X} \cdot \overline{(1/X)}}$$
where $\bar{X}$ is the sample arithmetic mean and $\overline{(1/X)}$ is the sample harmonic mean.

Expectation and Bias Analysis

To evaluate the finite-sample performance, the authors employ Laplace transform techniques and Tonelli's theorem.

• They derive an exact integral representation for the expectation $E[\hat{J}]$ involving the joint Laplace transform of $(X, 1/X)$.
• For the GIG family, this results in a complex integral involving Bessel functions.
• Specific closed-form or semi-closed-form expressions for the bias are derived for the Inverse Gaussian and Gamma distributions using properties of incomplete gamma functions and asymptotic limits.

Asymptotic Theory

The paper establishes large-sample properties using the Multivariate Central Limit Theorem and the Delta Method:

• Consistency: $\hat{J}$ is shown to be strongly consistent ($\hat{J} \xrightarrow{a.s.} J$).
• Asymptotic Normality: $\sqrt{n}(\hat{J} - J)$ converges in distribution to a normal distribution $N(0, \sigma_J^2)$. The asymptotic variance $\sigma_J^2$ depends on the variances of $X$ and $1/X$, as well as their covariance.
• Bias Expansion: A first-order bias expansion of order $O(1/n)$ is derived, providing a correction term for the estimator in finite samples.

Simulation Study

A Monte Carlo simulation study was conducted to validate the theoretical findings.

• Design: 4,000 replications ($R=4000$) for sample sizes $n \in \{20, 50, 75, 100, 200\}$.
• Models: Inverse Gaussian ($IG(\mu, \lambda)$) and Gamma ($Gamma(\alpha, \beta)$) distributions with varying parameters to test different levels of dispersion.
• Metrics: Empirical bias and Mean Squared Error (MSE) were computed.

3. Key Contributions

1. Analytical Expressions for GIG Family: The paper provides the first explicit analytical formulas for the AHI index within the Generalized Inverse Gaussian family, including the Inverse Gaussian and Gamma distributions.
2. Exact Integral Representation for Expectation: A novel derivation of the exact expectation of the estimator $\hat{J}$ using joint Laplace transforms, allowing for precise bias calculation without relying solely on asymptotic approximations.
3. Bias Approximations: Explicit first-order bias expressions are derived for the Inverse Gaussian and Gamma cases. Notably, for the Gamma distribution with shape parameter $\alpha$, the bias is shown to be approximately $-\frac{\alpha-1}{n\alpha(\alpha-2)}$.
4. Asymptotic Variance Structure: The derivation of the asymptotic variance highlights the critical role of the covariance between $X$ and $1/X$ in the dispersion measure.
5. Finite-Sample Validation: The simulation study confirms that the estimator is negatively biased (underestimates the true index) in small samples, but both bias and MSE decrease rapidly as sample size increases, validating the theoretical asymptotic results.

4. Key Results

• Inverse Gaussian Distribution: For $X \sim IG(\mu, \lambda)$, the population index is $J = \frac{\mu}{\mu + \lambda}$. The bias of the estimator depends on the ratio $\lambda/\mu$.
• Gamma Distribution: For $X \sim Gamma(\alpha, \beta)$, the population index is $J = 1/\alpha$ (for $\alpha > 1$). The index is independent of the rate parameter $\beta$. The bias is significant for small $\alpha$ (high dispersion) and small $n$.
• Levy Distribution: As a limiting case of the Inverse Gaussian ($\mu \to \infty$), the estimator $\hat{J}$ is found to be unbiased.
• Simulation Findings:
  • The estimator exhibits negative bias in all tested scenarios.
  • The magnitude of the bias decreases as $n$ increases.
  • The Mean Squared Error (MSE) follows the expected $O(1/n)$ decay pattern.
  • The theoretical bias approximations derived in Section 4 align closely with the empirical bias observed in the simulations.

5. Significance and Conclusion

This paper provides a comprehensive statistical foundation for the Arithmetic-Harmonic Inequality index, transforming it from a theoretical concept into a practical tool for statistical inference.

• Interpretability: By linking $J$ to the Atkinson index (with $\epsilon=2$) and the Coefficient of Variation, the authors provide a clear economic and statistical interpretation of the index as a bounded measure of inequality.
• Practical Utility: The derivation of bias corrections allows practitioners to adjust estimates in small-sample scenarios, which is crucial in fields like economics, finance, and reliability engineering where data may be limited.
• Theoretical Rigor: The use of Laplace transforms to handle the expectation of the ratio of means offers a robust methodological approach that can potentially be extended to other bounded dispersion indices.

In conclusion, Vila and Saulo successfully establish the AHI index as a robust, scale-invariant measure of dispersion with well-understood asymptotic and finite-sample properties, making it a viable alternative to traditional unbounded dispersion metrics.