When evaluating the fairness of an AI model or the economic development of a region, we often turn to summary statistics like the Gini coefficient or variance. These metrics collapse an entire population's distribution into a single number. While helpful, a solitary index treats the population as a monolithic pool, hiding critical structural details.
If inequality or unfairness increases, is it because the gap between different demographic groups is widening, or because the disparity within those groups is exploding? Standard metrics cannot answer this question because they lack a vital mathematical property: additive decomposability. Without this property, it is impossible to perfectly partition total disparity into distinct, non-overlapping structural components without introducing messy residual terms.
Additive inequality decomposition solves this by leveraging specific mathematical frameworks—primarily rooted in information theory—to break down a single global metric into mutually exclusive, exhaustive sub-components. This approach allows researchers to look inside the black box of a summary statistic and discover exactly where and why inequality occurs.
The Intuition of Additive Decomposability
An inequality index is additively decomposable if the total measured inequality ($I_{\text{total}}$) of a population can be expressed exactly as the sum of inequality between defined subgroups ($I_{\text{between}}$) and a weighted average of the inequality within those subgroups ($I_{\text{within}}$):
$$I_{\text{total}} = I_{\text{between}} + I_{\text{within}}$$
The most prominent examples of such metrics are entropy-based indices, such as the Theil Index and the broader family of Generalized Entropy (GE) indices. Because these indices are derived from information theory, they measure inequality as the "informational distance" or "surprise" of a distribution relative to a state of perfect equality.
To understand why this works so cleanly, imagine a hierarchical tree. If you group individuals into categories (such as geographic regions or demographic classes), the total variation in the system is equal to the variation between the group averages plus the remaining variation inside each individual group. As noted by Ma and Cheong (2022), standard aggregate measures often lack this property, which prevents researchers from cleanly separating internal group dynamics from structural gaps between different groups. Entropy-based indices inherently respect this hierarchy, ensuring that the internal and external components perfectly sum to the total.
Dissecting Spatial and Structural Disparities
This mathematical property makes additive decomposition an invaluable diagnostic tool for economists and policy analysts tracking regional development. A policy might successfully stimulate a lagging region, but a flat summary statistic would fail to show if that growth was evenly distributed or captured by a select few.
For example, when evaluating large-scale economic interventions like China's Western Development Strategy, Lu (2013) used entropy-based indices to show that while the macro-level gap between the coastal East and the inland West narrowed, internal inequalities within sub-regions actually intensified. Similarly, Yang (2024) leveraged multi-level Theil index decomposition to reveal "sustainability blind spots," demonstrating that while broad macro-regional economic gaps were shrinking, provincial-level disparities in public services and innovation were widening. By partitioning the data, the author could pinpoint whether development gaps were driven by macro-regional shifts or localized, neighbor-to-neighbor imbalances.
This technique has been widely adopted across diverse domains to isolate the primary drivers of disparity:
- Resource Allocation: Du et al. (2024) applied Theil decomposition to nursing human resources to distinguish whether healthcare access disparities primarily arose from differences within subgroups or between distinct groups.
- Industrial Economics: Chen (2018) isolated whether logistics industry disparities were concentrated within specific sectors or distributed across broader geographic regions.
- Alternative Disparity Metrics: While the standard Gini coefficient is notoriously difficult to decompose additively due to overlapping distributions, the Dagum Gini coefficient decomposition solves this by introducing a third component—transvariation (or hyperdensity)—which accounts for the overlap between groups. This approach has been used to study spatial differences in urban ecological performance (Zhang, 2022), digital innovation ecosystems (Lin et al., 2025), green competitiveness (Xu and Ci, 2023), and agricultural productivity (Sheng and Ma, 2026).
Unifying Individual and Group Algorithmic Fairness
In recent years, computer scientists have recognized that the mathematical machinery used to dissect regional economic inequality can solve a fundamental crisis in algorithmic fairness.
Historically, machine learning researchers treated individual fairness (treating similar individuals similarly) and group fairness (ensuring equal outcomes across demographic groups like race or gender) as separate, often conflicting objectives. As Gummadi and Heidari (2019) demonstrate, axiomatic inequality indices from economics provide the exact moral and mathematical framework needed to unify these concepts. By viewing algorithmic unfairness through the lens of additively decomposable indices, a profound realization emerges:
- Between-group inequality corresponds precisely to group unfairness.
- Within-group inequality corresponds precisely to individual unfairness.
Using this insight, Speicher et al. (2018) developed a unified framework that calculates the total unfairness of an algorithmic decision system and decomposes it. This reveals the exact trade-offs at play. For example, a system can be modified to achieve perfect group fairness (reducing between-group inequality to zero), but this might inadvertently increase within-group inequality, treating individuals within those groups highly unequally.
Without additive decomposition, this internal variance remains invisible. As Rampisela (2025) demonstrates in the context of Large Language Model (LLM) recommenders, a system can appear perfectly fair at a group level while masking severe disparities among individual users within those groups. Similarly, Chakraborty et al. (2019) used Generalized Entropy indices to show how subtle changes in the "choice architecture" of online donation platforms can quietly shift the balance between geographic group equality and individual project funding.
Whether applied to satellite data of nighttime lights (Zhou et al., 2015), public health trends (Xu et al., 2022; Xu et al., 2023), or machine learning models, additive inequality decomposition transforms a flat, one-dimensional metric into a multi-dimensional diagnostic tool. It works because it respects the mathematical structure of subgroups, allowing us to see not just that a system is unequal, but exactly how that inequality is constructed.
Key Foundational Work
"Regional Policy And Regional Development: A Case Study Of China‘s Western Development Strategy " (2013) — contributed 1 concept elements to this direction.
Nighttime Light Derived Assessment of Regional Inequality of Socioeconomic Development in China (2015) — contributed 1 concept elements to this direction.