The Impossibility of Simultaneous Fairness
What you'll learn: Why you mathematically cannot satisfy multiple fairness definitions at the same time—and what that means for real-world ML systems.
The Core Problem
You've learned several fairness metrics: demographic parity, equalized odds, predictive parity, and calibration. Here's the uncomfortable truth: except in trivial cases, you cannot satisfy all of them simultaneously. This isn't a matter of better algorithms or more data—it's a mathematical impossibility proven by formal theorems.
Why It's Impossible
The constraints imposed by different fairness definitions conflict at a fundamental level. For example:
- Demographic parity requires equal positive prediction rates across groups
- Equalized odds requires equal true positive and false positive rates across groups
- Predictive parity requires equal precision (positive predictive value) across groups
If the base rates (actual prevalence of the positive class) differ between groups—which they almost always do in real data—satisfying one metric forces you to violate another. It's like trying to make a triangle where all three angles are 90 degrees: geometrically impossible.
What This Means in Practice
You must make choices. There is no "perfectly fair" model that satisfies everyone's definition of fairness. Different stakeholders care about different metrics, and your choice reflects a value judgment about which harms to prioritize minimizing. A loan approval system optimized for demographic parity makes different tradeoffs than one optimized for equalized odds.
Key Takeaway: Multiple fairness criteria cannot be satisfied simultaneously due to mathematical constraints—fairness in ML always involves explicit tradeoffs and value choices, not purely technical solutions.