Lemma B.1: for Z ∈ [0,1] with E[Z] = μ and a ∈ (0,1), P[Z > 1 − a] ≥ (μ − (1 − a))/a and P[Z > a] ≥ (μ − a)/(1 − a) ≥ μ − a
ProvedUnderstandingML.reverse_markovappendix-bmarkov-inequalityprobability
Lemma B.1. Let be a random variable that takes values in . Assume that . Then, for any ,
This also implies that for every ,
Formally: for a probability measure on , a measurable with -almost surely, and , the three inequalities hold with .
Preamble
import Definitions.Def_UnderstandingML_Framework import Mathlib.SetTheory.Cardinal.Finite open MeasureTheory
Formal statement
namespace UnderstandingML
/-- **Lemma B.1** (p. 422). Let `Z` be a random variable that takes values in `[0, 1]` and assume
that `E[Z] = μ`. Then for any `a ∈ (0, 1)`, `P[Z > 1 − a] ≥ (μ − (1 − a))/a`. This also implies
that for every `a ∈ (0, 1)`, `P[Z > a] ≥ (μ − a)/(1 − a) ≥ μ − a`. Stated for a measurable
`θ : Ω → ℝ` with values in `[0, 1]` almost surely under a probability measure `D`, with
`μ = ∫ θ dD`. -/
theorem reverse_markov {Ω : Type*} [MeasurableSpace Ω] (D : Measure Ω) [IsProbabilityMeasure D]
(θ : Ω → ℝ) (hθ : Measurable θ) (hrange : ∀ᵐ ω ∂D, θ ω ∈ Set.Icc (0 : ℝ) 1) {a : ℝ}
(ha0 : 0 < a) (ha1 : a < 1) :
((∫ ω, θ ω ∂D) - (1 - a)) / a ≤ (D {ω | 1 - a < θ ω}).toReal ∧
((∫ ω, θ ω ∂D) - a) / (1 - a) ≤ (D {ω | a < θ ω}).toReal ∧
(∫ ω, θ ω ∂D) - a ≤ (D {ω | a < θ ω}).toReal := by sorry
end UnderstandingML
Source
Shalev-Shwartz and Ben-David, Understanding Machine Learning: From Theory to Algorithms, Cambridge University Press 2014, doi:10.1017/CBO9781107298019, Appendix B.1 p. 422, Lemma B.1 with its proof (from Markov's inequality (B.3))
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.