Definition 2 — distributional unlearning
Definitionsu_unlearningmeasure-theoryprobabilitystatistics
For baseline trade-off functions (removal) and (preservation), a distribution satisfies -distributional unlearning with respect to an unwanted distribution and a retained distribution when
both required to hold at every type-I level . The removal condition says the edited data can be told apart from the unwanted domain at least as easily as the baseline pair defining ; the preservation condition says it cannot be told apart from the retained domain more easily than the baseline pair defining . In Lean this is StatUnlearn.IsUnlearning p p₁ q₁ f_d f_c.
Definition code
/-
Mission: Statistical Unlearning of Distributions
Source: A. Pandey, S. Kulkarni, arXiv:2605.16645 — Definition 2 (eq. (2.2)).
-/
import Mathlib.MeasureTheory.Measure.MeasureSpaceDef
import Mathlib.Data.Real.Basic
import Definitions.Def_su_tof
open MeasureTheory
namespace StatUnlearn
/-- **`(f_d, f_c)` distributional unlearning** (Definition 2 of the paper, eq. (2.2)).
A distribution `p` satisfies `(f_d, f_c)`-unlearning with respect to the unwanted
distribution `p₁` and the retained distribution `q₁` when the removal condition
`T(p, p₁) ≤ f_d` and the preservation condition `T(p, q₁) ≥ f_c` hold for every
type-I error level `t`. -/
def IsUnlearning {W : Type*} [MeasurableSpace W] (p p₁ q₁ : Measure W)
(f_d f_c : ℝ → ℝ) : Prop :=
(∀ t, tof p p₁ t ≤ f_d t) ∧ ∀ t, f_c t ≤ tof p q₁ t
end StatUnlearn
Source
Pandey, Kulkarni, "Statistical Unlearning of Distributions: A Hypothesis Testing Approach", arXiv:2605.16645, p. 3, Definition 2, eq. (2.2)