Measure a model on one dataset and you ask about accuracy. Measure it separately on two groups and a different question appears: which of the equalities you would like can hold at the same time?
That turns out to be arithmetic, not policy — so this page computes instead of arguing.
See it: https://dev48.infy.uk/ml/day74-group-fairness-impossibility.html
The world is built so nothing is confounded
- The class-conditional score distribution is identical in both groups. The two ROC curves are the same curve; the AUC gap is exactly 0.
- The decision is a function of the score bin alone, so fairness through unawareness already holds for free.
- The only difference between the groups is the base rate: 30% against 8%.
At the payoff-optimal cut, both groups get true-positive rate 0.7759 and false-positive rate 0.1719 — to the last bit.
And 35.31% of group A is flagged against 22.02% of group B.
The gap is one product
selection gap = (piA - piB) * J where J = TPR - FPR (Youden's J)
= (0.30 - 0.08) * 0.6040
= 0.1329 exactly
The whole disparity is a single product of two things you already knew: the prevalence gap and how good the score is. Nothing about the model is unfair; the arithmetic simply does not permit equal selection rates and equal error rates at once when base rates differ.
The largest that gap can ever get is the total-variation distance between the two score laws — a bound, not an estimate.
The Bayes cut that does not exist
The closed-form optimal cut returns null exactly where the score is uninformative. That is not a gap in the code; it is the correct answer. There is no cut to find when the likelihood ratio is 1, and a verifier that assumes an index there crashes — which is precisely what happened when this page's verifier was written.
Verified over 120 worlds, 1,439 asserts in exact rational arithmetic, 0 failures.
