Triangle inequality for the Mahalanobis distance
ProvedStatUnlearn.mahalanobis_trianglelinear-algebramachine-unlearningprobabilitystatistics
For a positive definite covariance and vectors of , the Mahalanobis distance
satisfies the triangle inequality
Equivalently, in the coordinates that make the quadratic form the Euclidean norm, becomes an ordinary norm and this is the triangle inequality for it. It is the single geometric input needed to show that in the feasible region of Proposition 3.1 the removal condition is automatic on the preservation ball whenever the levels satisfy .
Preamble
import Mathlib.Probability.Distributions.Gaussian.Multivariate import Mathlib.Analysis.InnerProductSpace.EuclideanDist open MeasureTheory
Formal statement
namespace StatUnlearn
theorem mahalanobis_triangle {d : ℕ} (S : Matrix (Fin d) (Fin d) ℝ) (hS : S.PosDef)
(x y : EuclideanSpace ℝ (Fin d)) :
Real.sqrt (inner ℝ (x + y) (Matrix.toEuclideanLin S⁻¹ (x + y)))
≤ Real.sqrt (inner ℝ x (Matrix.toEuclideanLin S⁻¹ x))
+ Real.sqrt (inner ℝ y (Matrix.toEuclideanLin S⁻¹ y)) := by
sorry
end StatUnlearnSource
Pandey, Kulkarni, "Statistical Unlearning of Distributions: A Hypothesis Testing Approach", arXiv:2605.16645, p. 5, proof of Proposition 3.1 (`proof:FFRG`): the Euclidean picture of the region after the change of coordinates x -> Sigma^{-1/2} x