Definition 1 — the trade-off function
Definitionsu_tofmeasure-theoryprobabilitystatistics
Given two probability distributions and on a measurable space , the trade-off function is the map
where the infimum runs over all measurable functions , with the type-I error and the type-II error. Intuitively, is the smallest type-II error an adversary can achieve at type-I level , so a larger trade-off function means the two distributions are harder to distinguish; is the largest possible value. In Lean this is StatUnlearn.tof P Q t, an sInf over measurable [0,1]-valued tests against P and Q.
Definition code
/-
Mission: Statistical Unlearning of Distributions
Source: A. Pandey, S. Kulkarni, "Statistical Unlearning of Distributions: A Hypothesis
Testing Approach", arXiv:2605.16645 — Definition 1 (eq. (2.1)).
-/
import Mathlib.MeasureTheory.Integral.Bochner.Basic
import Mathlib.MeasureTheory.Measure.MeasureSpaceDef
import Mathlib.Data.Real.Basic
import Mathlib.Order.ConditionallyCompleteLattice.Basic
open MeasureTheory
namespace StatUnlearn
/-- **Trade-off function** (Definition 1 of the paper).
Given two probability distributions `P`, `Q` on a measurable space `W`, and a level
`t` of type-I error, `tof P Q t` is the infimum of the achievable type-II errors
`β_φ = ∫ (1 - φ) dQ` over all randomized tests `φ : W → [0,1]` that are measurable and
satisfy `α_φ = ∫ φ dP ≤ t`.
This is the paper's
`T(P,Q)(α) := inf { β_φ : α_φ ≤ α }` of equation (2.1). -/
noncomputable def tof {W : Type*} [MeasurableSpace W] (P Q : Measure W) (t : ℝ) : ℝ :=
sInf {β : ℝ | ∃ φ : W → ℝ, Measurable φ ∧ (∀ w, φ w ∈ Set.Icc (0 : ℝ) 1) ∧
(∫ w, φ w ∂P) ≤ t ∧ β = ∫ w, (1 - φ w) ∂Q}
end StatUnlearn
Source
Pandey, Kulkarni, "Statistical Unlearning of Distributions: A Hypothesis Testing Approach", arXiv:2605.16645, p. 3, Definition 1, eq. (2.1)