resample_measure_preserving
Provedconcentration-inequalitiesefron-steinprobabilityproduct-measureresamplingvariance
The single-coordinate resample map is measure-preserving on a product cube. Let be a product of probability measures on (finite index set), and fix a coordinate . The map , which replaces the -th coordinate of by (leaving the other coordinates of unchanged), pushes the product measure forward to . Equivalently: resampling one coordinate of a draw from a product measure, using an independent copy, leaves the distribution unchanged. This is the measure-theoretic backbone of the resampling form of the Efron–Stein inequality: it makes and identically distributed, so the resampled difference is well-behaved.
Preamble
import Mathlib.Probability.CondVar import Mathlib.Probability.Moments.Variance import Mathlib.Probability.Independence.Basic import Mathlib.MeasureTheory.Constructions.Pi import Mathlib.Probability.Independence.Integration import Mathlib.Probability.IdentDistrib open MeasureTheory ProbabilityTheory Filter Set Function open scoped ENNReal NNReal BigOperators
Formal statement
theorem resample_measure_preserving
{ι : Type*} [Fintype ι] [DecidableEq ι]
{α : ι → Type*} [∀ i, MeasurableSpace (α i)]
(μ : ∀ i, Measure (α i)) [∀ i, IsProbabilityMeasure (μ i)] (i : ι) :
MeasurePreserving (fun p : (∀ j, α j) × (∀ j, α j) => Function.update p.1 i (p.2 i))
((Measure.pi μ).prod (Measure.pi μ)) (Measure.pi μ) := by sorrySource
R. van Handel, Probability in High Dimension (APC 550 lecture notes, Princeton), §2.1 Tensorization and bounded differences (Efron-Stein via the resampling / symmetrization identity Var(W)=½E[(W-W')^2] for an independent copy); Boucheron-Lugosi-Massart, Concentration Inequalities (OUP 2013), Ch. 3, Theorem 3.1 (the resampling form of the Efron-Stein inequality on a product space).