To address inequality, we must first understand its anatomy. A single summary statistic, like the Gini coefficient, can tell us that a system is unequal, but it cannot explain why. Is the disparity driven by a few wealthy individuals scattered across the population, or is it driven by structural gaps between specific regions, genders, or ethnicities?
Researchers across economics, public health, and computer science are increasingly using additive inequality decomposition to answer these questions. This mathematical framework allows researchers to "unzip" total inequality into two distinct components: the disparity occurring within specific groups and the disparity occurring between those groups.
The Power of Additive Decomposability
Most common measures of inequality lack a critical property: additive decomposability. As Ning Ma and Tsun Se Cheong point out in Evaluating Global Inequality Using Decomposition Approach, standard measures often fail to perfectly partition total disparity. Without this property, the sum of internal and external components does not equal the whole, making it impossible to precisely attribute the drivers of inequality.
To solve this, researchers turn to the Theil Index and Generalized Entropy (GE) indices. These metrics are unique because they allow for a clean mathematical split:
$$Total\ Inequality = Within\text{-}Group\ Inequality + Between\text{-}Group\ Inequality$$
This mechanism provides a "high-resolution" view of disparity. For instance, Liangliang Chen uses Theil Index decomposition in The Impact of Logistics Industry on Chinars Regional Economic Growth Based on Theil Index Analysis to show that while the logistics industry boosts the economy, the benefits are trapped within specific coastal regions, creating a massive "between-group" gap with the rest of the country.
Identifying Structural vs. Individual Drivers
The intuition behind decomposition is best seen in how it identifies "sustainability blind spots." In Measuring and Monitoring Provincial Sustainable Development Goals in China, Yongheng Yang demonstrates that even when the gap between broad regions (East vs. West) narrows, the gaps within those regions—between neighboring provinces—can actually widen.
This same logic applies to public health. In Assessing the income-related inequality in obesity among the elderly in China: A decomposition analysis, Jinpeng Xu and colleagues use decomposition to quantify exactly how much of the obesity trend is explained by income versus urban residency. By calculating the sensitivity of each factor, they can pinpoint that socioeconomic status drives nearly 70% of the observed inequality.
From Measurement to Algorithmic Fairness
Perhaps the most innovative application of this idea is in "Fair Machine Learning" (Fair-ML). Traditionally, AI researchers struggled to balance individual fairness (treating similar people similarly) with group fairness (ensuring equal outcomes for different demographics).
Krishna P. Gummadi and Hoda Heidari argue in Economic Theories of Distributive Justice for Fair Machine Learning that decomposable indices provide the mathematical bridge needed to unify these two goals. Because the Theil index can measure unfairness at both the individual and group levels simultaneously, it allows developers to see if an algorithm's bias is coming from general noise or from systemic discrimination against a protected class.
This has practical consequences for platform design. In Subgroup Fairness in Two-Sided Markets, Jakub Marecek and Robert Noel Shorten use these metrics to design market-clearing algorithms for ride-hailing services. By decomposing the inequality of driver earnings, they can independently adjust the "between-group" component to close gender pay gaps without disrupting the overall efficiency of the market.
Why It Works: The Choice Architecture
Finally, decomposition reveals how the design of digital systems—the "choice architecture"—can inadvertently create inequality. Abhijnan Chakraborty and Asia J. Biega show in On the Impact of Choice Architectures on Inequality in Online Donation Platforms that features as simple as a "location filter" can be analyzed through additive decomposition. They found that these design choices often nudge donors toward specific projects, reinforcing geographic disparities.
By using Generalized Entropy indices, they could quantify exactly how much of the funding gap was attributable to the platform's interface versus the donors' inherent preferences. This level of granularity is only possible because the math allows us to separate the "within" from the "between," turning a vague social problem into a series of addressable engineering and policy challenges.