Alternate definitions of Gini, Hoover and Lorenz measures of inequalities and convergence with respect to the Wasserstein W1 metric

TL;DR

This paper explores alternative definitions of Gini, Hoover, and Lorenz measures of inequality, and establishes their convergence properties under the Wasserstein W1 metric, linking distribution approximation to inequality measure stability.

Contribution

It introduces generalized definitions of inequality measures and proves their convergence under the Wasserstein W1 metric, connecting distribution approximation with the stability of these measures.

Findings

Wasserstein W1 convergence implies uniform convergence of Lorenz curves.

Empirical and partial information-based measures converge to true inequality measures.

The paper discusses conditions where W1 convergence may not hold but weaker assumptions suffice.

Abstract

This article focuses on some properties of three tools used to measure economic inequalities with respect to a distribution of wealth $\mu$: Gini coefficient $G$, Hoover coefficient or Robin Hood coefficient $H$, and the Lorenz concentration curve $L$. To express the distributions of resources, we use the framework of random variables and abstract Borel measures. In the first part (sections 1-4), we discuss alternate definitions of $G$, $H$ and $L$ that can be found in economics literature. Gini and Hoover coefficients are defined as mean deviation and mean absolute differences, and interpreted as geometrical properties of the Lorenz curve. In particular, we give a more general and straightforward proof of the main result of [Dorfman, 1979]. The second part of the article (section 5-7) focuses on the consistency of $G(\mu)$, $H(\mu)$ and $L_\mu$ as $\mu$ is approximated or perturbated. The relevant tool to use is the Wasserstein metric $\mathrm{W}_1$, i.e. the $\mathrm{L}^1$ metric between quantile functions. Our main theorem shows that if $\mathrm{W}_1(\mu_n, \mu_\infty) \to 0$ if and only if $L_{\mu_n} \to L_{\mu_\infty}$ uniformly. We discuss the topological implications of this fact. Thus, we show that the empirical Gini, Hoover indexes and Lorenz curves computed on a sample or rebuilt with partial information converge to the real Gini, Hoover indexes and Lorenz curve as information increases in several cases. Eventually, we discuss the situations where the $\mathrm{W}_1$ convergence is not matched but weaker asumptions can be made.