Appendix E Eq.(E6): one-sided variance bound
ProvedQAlgorithms.Papers.Independent.Advance50.P50.misclassification_variance_boundIn the equal-prior setting of Lemma 2, let the measurable output have finite second moment under class-0 law . Suppose class-1 outputs equal , class-0 outputs lie in , and . Write and . Then
This is the one-sided variance bound with the tail direction corrected as forced by . It concerns exact outputs and includes no unspecified additive measurement-noise model.
import Definitions.Def_QAlgorithms_Papers_Independent_Advance50_P50
noncomputable section
namespace QAlgorithms.Papers.Independent.Advance50.P50
open MeasureTheory ProbabilityTheory
open scoped ComplexConjugate ENNReal
-- Exact accepted primitives; no paper-local redeclarations.
open QAlgorithms.Papers.Independent.Information100.P63 (Mat tensorPower)
open QAlgorithms.Papers.Independent.Gap48.P20 (Copies)
open QAlgorithms.Papers.Independent.Gap48.P38 (subsystemReduction)
open QAlgorithms.Papers.Independent.Information100.P32 (purity)
open QAlgorithms.Papers.Independent.Information100.P08 (IsPure)
-- Matrices carry the associative commutator bracket X*Y-Y*X.
attribute [local instance 100] LieRing.ofAssociativeRing
/-- PDF p.27, Eq.(E6): the printed upper-tail sign is corrected to the lower
tail, forced by delta=mean-c>0. The variance and posterior bounds are retained.
No unspecified finite-shot noise model is introduced.
Printed: PDF p.27 E5-E6 uses X-mean >= delta while defining delta=mean-c>0.
Changed: Use mean-X >= delta in the entire displayed probability chain. Keep delta=mean-c>0, the atom X=c, equal priors, the constant class-1 output and the same variance/(variance+delta^2) bound.
Why: At X=c, mean-X=delta, whereas X-mean=-delta. Only the lower tail contains the atom being bounded. Cantelli has the same constant for either tail, so changing this one sign is the minimal repair. -/
theorem misclassification_variance_bound {Omega : Type*} [MeasurableSpace Omega]
(mu0 mu1 : Measure Omega) [IsProbabilityMeasure mu0] [IsProbabilityMeasure mu1]
(X : Omega → ℝ) (hX : Measurable X) (hL2 : MemLp X 2 mu0)
(c b1 b2 : ℝ) (hclass1 : ∀ᵐ w ∂mu1, X w = c)
(hclass0 : ∀ᵐ w ∂mu0, X w ∈ Set.Icc b1 b2)
(hmean : c < ∫ w, X w ∂mu0) :
let delta := (∫ w, X w ∂mu0) - c
posteriorAt mu0 mu1 X c ≤ (mu0 {w | X w = c}).toReal ∧
(mu0 {w | X w = c}).toReal ≤
(mu0 {w | delta ≤ (∫ z, X z ∂mu0) - X w}).toReal ∧
(mu0 {w | delta ≤ (∫ z, X z ∂mu0) - X w}).toReal ≤
variance X mu0 / (variance X mu0 + delta ^ 2) := by sorry
end QAlgorithms.Papers.Independent.Advance50.P50
end
Read-back
What the Lean code literally says, in plain math · gpt-6.1-sol
For every measurable space , probability measures , measurable real-valued function that belongs to , and real numbers , assume that almost everywhere under , that almost everywhere under , and that . With and , the assertion is the chain
Measure values here are converted to real numbers, and the first expression is precisely the defined ratio. Zero variance and zero mass of are included; the final denominator is positive because . If , the interval hypothesis is impossible.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.