The Mathematical Incompatibility of Algorithmic Fairness Metrics

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2358 papers568 conceptsfrozen 2026-10-04

In the effort to build ethical AI, researchers have proposed dozens of mathematical definitions for what it means for an algorithm to be "fair." However, a central finding across modern machine learning research is that these definitions are often mutually exclusive. We cannot simply "fix" bias by checking every box; instead, we are forced to navigate a landscape of unavoidable mathematical trade-offs.

The Intuition of Fairness Friction

The core problem arises because different stakeholders prioritize different statistical properties. A bank might want Predictive Parity (ensuring a 700 credit score means the same thing regardless of race), while a regulator might demand Demographic Parity (ensuring loan approval rates are equal across groups).

As Solon Barocas, Elizabeth Bradley, Vasant G Honavar, and Foster J. Provost argue in Big Data, Data Science, and Civil Rights, it is mathematically impossible to satisfy these common definitions simultaneously when the underlying populations have different "base rates" for the target behavior. If one group historically has a higher rate of default due to systemic factors, a model cannot be both perfectly calibrated and perfectly equal in its outcomes.

Why Fairness Definitions Conflict

To understand why this happens, consider the relationship between precision, recall, and prevalence. If you force an algorithm to have the same false positive rate and false negative rate across two groups (Equalized Odds), but those groups have different underlying frequencies of the outcome, the model’s predictive value will necessarily differ between them.

This "no free lunch" theorem is a recurring theme. Yijin Ni and Xiaoming Huo demonstrate in Kernel-based Equalized Odds that even in advanced representation learning—where we try to "scrub" sensitive attributes from data—satisfying one criterion like demographic parity often forces the violation of others.

The conflict isn't just a quirk of simple classifiers; it scales with complexity. In The Impossibility of Fair LLMs, Jacy Reese Anthis, Kristian Lum, and colleagues explain that for general-purpose models, the combinatorial explosion of possible contexts makes universal fairness guarantees "intractable." A model optimized for fairness in one cultural context will almost certainly introduce distortions in another.

Mapping the Trade-off Space

Rather than searching for a single "fair" model, recent work focuses on quantifying the cost of these trade-offs. Researchers are moving from binary "fair or unfair" labels to a more nuanced "feasibility region."

From Mathematical Limits to Policy Realities

The realization that fairness is a zero-sum game in many statistical settings has profound implications for law and regulation. As noted by Jakob Mökander and David S. Watson in their analysis of the US Algorithmic Accountability Act vs. the EU AI Act, legislative mandates often fail to account for these mathematical contradictions. Laws that demand both "equal treatment" (individual fairness) and "equal outcomes" (group fairness) may be asking for something that is technically impossible to deliver.

Ultimately, the research suggests that "fairness" cannot be solved by an algorithm alone. Because we cannot satisfy every definition, the choice of which metric to prioritize is a subjective, ethical judgment. As Rik Adriaensen and Jessa Bekker suggest with their neurosymbolic approach in ProbLog4Fairness, the path forward involves explicitly modeling the process of how bias enters data, allowing humans to make principled decisions about which trade-offs are acceptable in a given social context.

Papers in this version

Showing 60 of 358 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 521 papers
  2. Version 1 · 2026-10-04 · 358 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_fairness_metric_incompatibility_dfaccaa8,
  title = {fairness metric incompatibility},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/fairness-metric-incompatibility/rdr_dfaccaa86e5a49259035a81948b071fc},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_the_mathematical_incompatibility_of_algo_dfaccaa8_v1,
  title = {The Mathematical Incompatibility of Algorithmic Fairness Metrics},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/the-mathematical-incompatibility-of-algorithmic-fairness-metrics/rdr_dfaccaa86e5a49259035a81948b071fc/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}