Theorem 7.7.1 — Intrinsic Hermitian matrix Bernstein
ProvedTroppMatrixConcentration.ch7_intrinsic_hermitian_bernsteinLet be independent measurable Hermitian complex random matrices, with finite second moments, mean zero, and almost surely, where and . Write . Let satisfy , and put . Then
The summand bound is one-sided; no lower eigenvalue or full norm bound is imposed. Finite second moments implement the source's standing integrability convention. Nonzero makes the intrinsic dimension well-defined and the displayed denominator positive.
import Definitions.Def_TroppMatrixConcentration_ch7_intrinsic open MeasureTheory ProbabilityTheory open scoped Matrix.Norms.L2Operator ComplexOrder
namespace TroppMatrixConcentration
theorem ch7_intrinsic_hermitian_bernstein {Ω : Type*} [MeasurableSpace Ω]
(μ : Measure Ω) [IsProbabilityMeasure μ] {d N : ℕ} [NeZero d]
(X : Fin N → Ω → Matrix (Fin d) (Fin d) ℂ)
(L : ℝ) (hL : 0 ≤ L) (V : Matrix (Fin d) (Fin d) ℂ) (hV : V ≠ 0)
(hMeas : ∀ k, Measurable (X k)) (hIndep : iIndepFun X μ)
(hHerm : ∀ k, ∀ᵐ ω ∂μ, (X k ω).IsHermitian)
(hL2 : ∀ k, MemLp (X k) 2 μ)
(hMean : ∀ k, (∫ ω, X k ω ∂μ) = 0)
(hBound : ∀ k, ∀ᵐ ω ∂μ, lambdaMax (X k ω) ≤ L)
(hVarianceBound : loewnerLE (∫ ω, (∑ k, X k ω) ^ 2 ∂μ) V) :
let Y := fun ω => ∑ k, X k ω
let v := spectralNorm V
(∫ ω, Y ω ^ 2 ∂μ) = ∑ k, ∫ ω, X k ω ^ 2 ∂μ ∧
∀ t : ℝ, Real.sqrt v + L / 3 ≤ t →
(μ {ω | t ≤ lambdaMax (Y ω)}).toReal ≤
4 * intrinsicDimension V * Real.exp (-(t ^ 2 / 2) / (v + L * t / 3)) := by sorry
end TroppMatrixConcentration
Read-back
What the Lean code literally says, in plain math · gpt-6
For every measurable space with a probability measure , every pair of natural numbers with , every family of functions indexed by , every real number , and every nonzero matrix , assume that each is measurable for the Borel measurable structure on matrices, the family is mutually independent under , each is Hermitian almost everywhere and belongs to for the operator norm induced by the Euclidean norm, each Bochner integral is the zero matrix, and almost everywhere for each . Here is the supremum of the real spectrum of , meaning the set of real numbers for which is not invertible; for Hermitian matrices of the positive dimension required here, it is the largest eigenvalue. On matrices with empty real spectrum the real-valued supremum convention gives , so the definition also applies to any exceptional non-Hermitian values of the random matrices. Suppose additionally that is positive semidefinite, where squares and products are matrix products. Set and . Then both and, for every real satisfying , the inequality hold. The probability on the left is expressed as a real number, which is unambiguous because is a probability measure. The matrix has no separately stated Hermitian or positivity hypothesis beyond the positive-semidefinite difference condition above. Since , one has , so the displayed denominators are nonzero for the permitted thresholds. The number of summands is allowed to be zero: then the family hypotheses are vacuous, and both sides of the integral identity are zero matrices, and the variance condition says that itself is positive semidefinite. The value and equality at the threshold are allowed.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.