Lieb replacement for an independent random matrix offset
ProvedTroppMatrixConcentration.ch3_cgf_independent_replacementmatrix-concentrationrandom-matrices
Let and be independent measurable, almost surely Hermitian random complex matrices on a probability space, with . Assume is Bochner integrable, and both trace-exponential expressions below are integrable. Put . Then
Traces mean their real parts. The offset is random; independence permits probabilistic Lieb to be applied at each value of the offset. This is the one-summand replacement step used to derive cgf subadditivity.
Preamble
import Definitions.Def_TroppMatrixConcentration_probability open MeasureTheory ProbabilityTheory open scoped Matrix.Norms.L2Operator set_option autoImplicit false
Formal statement
namespace TroppMatrixConcentration
theorem ch3_cgf_independent_replacement {Ω : Type*} [MeasurableSpace Ω]
(μ : Measure Ω) [IsProbabilityMeasure μ] {d : ℕ} [NeZero d]
(H X : Ω → Matrix (Fin d) (Fin d) ℂ)
(hMeasH : Measurable H) (hMeasX : Measurable X)
(hHermH : ∀ᵐ ω ∂μ, (H ω).IsHermitian)
(hHermX : ∀ᵐ ω ∂μ, (X ω).IsHermitian)
(hIndep : IndepFun H X μ)
(hExpX : Integrable (fun ω => matrixExp (X ω)) μ)
(hIntTotal : Integrable (fun ω => traceExp (H ω + X ω)) μ)
(hIntReplaced : Integrable (fun ω =>
traceExp (H ω + matrixLog (∫ u, matrixExp (X u) ∂μ))) μ) :
(∫ ω, traceExp (H ω + X ω) ∂μ) ≤
∫ ω, traceExp (H ω + matrixLog (∫ u, matrixExp (X u) ∂μ)) ∂μ := by sorry
end TroppMatrixConcentration
Source
Joel A. Tropp, An Introduction to Matrix Concentration Inequalities, arXiv:1501.01571v1 (7 January 2015), https://arxiv.org/abs/1501.01571v1; Proof of Lemma 3.5.1, printed pp. 35–36, applying Corollary 3.4.2.