The Mahalanobis shell is the whole ball when the levels separate
ProvedStatUnlearn.shiftedGaussianShell_eq_ballhypothesis-testingmachine-unlearningprobabilitystatistics
For shifted Gaussians on with common positive definite covariance and levels , if
then the removal condition holds automatically on the whole preservation ball, so the shell equals that ball:
This is the "whole ball" clause of Proposition 3.1, with the ball in the Mahalanobis metric induced by ; for it is the Euclidean closed ball of the paper.
Preamble
import Mathlib.Probability.Distributions.Gaussian.Multivariate import Mathlib.Probability.Distributions.Gaussian.Real import Mathlib.Analysis.InnerProductSpace.EuclideanDist open MeasureTheory ProbabilityTheory
Formal statement
namespace StatUnlearn
theorem shiftedGaussianShell_eq_ball {d : ℕ} (S : Matrix (Fin d) (Fin d) ℝ) (hS : S.PosDef)
(μ₁ ν₁ : EuclideanSpace ℝ (Fin d)) (α ε : ℝ) (hα : 0 ≤ α) (hε : 0 ≤ ε)
(h : Real.sqrt (inner ℝ (μ₁ - ν₁) (Matrix.toEuclideanLin S⁻¹ (μ₁ - ν₁))) - ε ≥ α) :
({μ : EuclideanSpace ℝ (Fin d) |
α ≤ Real.sqrt (inner ℝ (μ₁ - μ) (Matrix.toEuclideanLin S⁻¹ (μ₁ - μ))) ∧
Real.sqrt (inner ℝ (ν₁ - μ) (Matrix.toEuclideanLin S⁻¹ (ν₁ - μ))) ≤ ε}
: Set (EuclideanSpace ℝ (Fin d)))
= ({μ : EuclideanSpace ℝ (Fin d) |
Real.sqrt (inner ℝ (ν₁ - μ) (Matrix.toEuclideanLin S⁻¹ (ν₁ - μ))) ≤ ε}
: Set (EuclideanSpace ℝ (Fin d))) := by
sorry
end StatUnlearnSource
Pandey, Kulkarni, "Statistical Unlearning of Distributions: A Hypothesis Testing Approach", arXiv:2605.16645, p. 5, Proposition 3.1 (`thm:FFR`), whole-ball case