Proof of Proposition 7, p. 15 — Fréchet endpoint bounds on every attainable disparity
ProvedFairProxy.ClosedFormDD.eq_13_boundsalgorithmic-fairnessdata-combinationdemographic-disparityfrechet-hoeffdingp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1partial-identification
For valid observed marginals with both class probabilities positive, every attainable conditional class-weight function has demographic disparity between the two Fréchet endpoint expressions:
The lower inequality appears after equation (13) in the proof of Proposition 7; the paper says the upper inequality follows analogously. This gives containment of the identified set in the proposed interval.
Formalization Note The result includes both inequalities, with the group masses positive so has its intended probabilistic meaning.
Preamble
import Mathlib import Definitions.Def_FairProxy_ClosedFormDD_Model
Formal statement
namespace FairProxy.ClosedFormDD
/-- The lower bound in the proof of Proposition 7, p. 15, after equation (13),
and its stated analogous upper bound. -/
theorem eq_13_bounds {Z : Type*} [MeasurableSpace Z]
(M : Marginals Z) (hM : M.Valid)
(ha : 0 < probA M .a) (hb : 0 < probA M .b)
(w : Cls → Fin 2 → Z → ℝ) (hw : w ∈ WPD M) :
mu M .a (wL M) - mu M .b (wU M) ≤ mu M .a w - mu M .b w ∧
mu M .a w - mu M .b w ≤
mu M .a (wU M) - mu M .b (wL M) := by sorry
end FairProxy.ClosedFormDD
Source
Kallus, Mao & Zhou, Assessing Algorithmic Fairness with Unobserved Protected Class Using Data Combination, arXiv:1906.00285v2, p. 15, proof of Proposition 7, after Eq. (13)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.