Proof of Corollary 6.2.1 — estimator variance bound
ProvedIntroMatrixConc.Sampling.variance_boundmatrix-concentrationp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-bookp2o-v1samplingvariance
Let be independent copies of a measurable complex random matrix , with and . Suppose and almost surely. Write and . Then
where is the rectangular second-moment statistic of the centered estimator error. This bounds the variance parameter in the matrix Bernstein estimate by one sample's second moment.
Formalization Note The almost-sure spectral norm bound ensures that all matrix-valued second moments are integrable.
Preamble
import Definitions.Def_IntroMatrixConc_Sampling_Defs import Mathlib.Probability.IdentDistrib open MeasureTheory ProbabilityTheory TroppMatrixConcentration open scoped Matrix.Norms.L2Operator noncomputable section
Formal statement
namespace IntroMatrixConc.Sampling
/-- Tropp, proof of Corollary 6.2.1, p. 82 (PDF p. 88): `v(Z) ≤ m₂(R)/n`. -/
theorem variance_bound {Ω : Type*} [MeasurableSpace Ω]
(μ : Measure Ω) [IsProbabilityMeasure μ]
{d₁ d₂ n : ℕ} [NeZero d₁] [NeZero d₂]
(R₀ : Ω → Matrix (Fin d₁) (Fin d₂) ℂ)
(R : Fin n → Ω → Matrix (Fin d₁) (Fin d₂) ℂ)
(B : Matrix (Fin d₁) (Fin d₂) ℂ) (L : ℝ)
(hn : 0 < n)
(hMeas₀ : Measurable R₀) (hMeas : ∀ k, Measurable (R k))
(hIndep : iIndepFun R μ)
(hCopies : ∀ k, IdentDistrib (R k) R₀ μ μ)
(hMean : (∫ ω, R₀ ω ∂μ) = B)
(hBound : ∀ᵐ ω ∂μ, spectralNorm (R₀ ω) ≤ L) :
rectSecondMoment μ (fun ω => samplingEstimator R ω - B) ≤
rectSecondMoment μ R₀ / n := by sorry
end IntroMatrixConc.Sampling
Source
Tropp, arXiv:1501.01571v1, §6.2.2, proof of Corollary 6.2.1, p. 82 (PDF p. 88), 'In summary' display
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.