Second moment of an independent centered matrix sum: cross terms vanish
ProvedTroppMatrixConcentration.ch6_independent_sum_second_momentmatrix-concentrationprobabilityrandom-matrices
Let be independent measurable random complex matrices on a probability space, and let and be measurable matrix-valued maps (into and matrices) such that each and each is square integrable for the spectral norm and for every . Then
The cross terms with vanish because and are independent and has mean zero. Taking gives the additivity of the matrix variance of an independent centered Hermitian sum (Tropp, eq. (2.2.5)); taking , or , gives the two rectangular variance identities used in the matrix Bernstein inequalities (Tropp, eqs. (6.1.2) and (7.3.2)). Expectations are Bochner integrals with respect to the spectral norm.
Preamble
import Definitions.Def_TroppMatrixConcentration_probability import Definitions.Def_TroppMatrixConcentration_dilation import Mathlib.Analysis.Convex.Function open MeasureTheory ProbabilityTheory open scoped Matrix.Norms.L2Operator
Formal statement
namespace TroppMatrixConcentration
theorem ch6_independent_sum_second_moment {Ω : Type*} [MeasurableSpace Ω]
(μ : Measure Ω) [IsProbabilityMeasure μ] {m n a b c N : ℕ}
(X : Fin N → Ω → Matrix (Fin m) (Fin n) ℂ)
(f : Matrix (Fin m) (Fin n) ℂ → Matrix (Fin a) (Fin b) ℂ)
(g : Matrix (Fin m) (Fin n) ℂ → Matrix (Fin b) (Fin c) ℂ)
(hf : Measurable f) (hg : Measurable g)
(hMeas : ∀ k, Measurable (X k)) (hIndep : iIndepFun X μ)
(hfL2 : ∀ k, MemLp (fun ω => f (X k ω)) 2 μ)
(hgL2 : ∀ k, MemLp (fun ω => g (X k ω)) 2 μ)
(hfMean : ∀ k, (∫ ω, f (X k ω) ∂μ) = 0) :
(∫ ω, (∑ k, f (X k ω)) * (∑ k, g (X k ω)) ∂μ) =
∑ k, ∫ ω, f (X k ω) * g (X k ω) ∂μ := by sorry
end TroppMatrixConcentrationSource
Joel A. Tropp, An Introduction to Matrix Concentration Inequalities, arXiv:1501.01571v1 (7 January 2015); https://arxiv.org/abs/1501.01571v1; Section 2.2.3, equations (2.2.5)-(2.2.6), printed pp. 27-28 (additivity of the variance of an independent sum; the cross terms vanish by independence and zero mean); used in Theorem 6.1.1 eq. (6.1.2), Theorem 6.6.1, and Theorem 7.3.1 eq. (7.3.2).