Proposition 3.2.1 — Tail bounds for eigenvalues
ProvedTroppMatrixConcentration.ch3_laplace_tailsFor a measurable, almost-surely Hermitian random matrix on a probability space, whenever is Bochner integrable,
Both bounds hold for every real threshold . These pointwise parameter bounds provide the source's infimum bounds over parameters with finite matrix mgf.
import Definitions.Def_TroppMatrixConcentration_probability open MeasureTheory ProbabilityTheory open scoped Matrix.Norms.L2Operator
namespace TroppMatrixConcentration
theorem ch3_laplace_tails {Ω : Type*} [MeasurableSpace Ω]
(μ : Measure Ω) [IsProbabilityMeasure μ] {d : ℕ} [NeZero d]
(Y : Ω → Matrix (Fin d) (Fin d) ℂ) (θ : ℝ)
(hMeas : Measurable Y) (hHerm : ∀ᵐ ω ∂μ, (Y ω).IsHermitian)
(hExp : Integrable (fun ω => matrixExp (θ • Y ω)) μ) :
(0 < θ → ∀ t : ℝ,
(μ {ω | t ≤ lambdaMax (Y ω)}).toReal ≤
Real.exp (-θ * t) * (∫ ω, traceExp (θ • Y ω) ∂μ)) ∧
(θ < 0 → ∀ t : ℝ,
(μ {ω | lambdaMin (Y ω) ≤ t}).toReal ≤
Real.exp (-θ * t) * (∫ ω, traceExp (θ • Y ω) ∂μ)) := by sorry
end TroppMatrixConcentration
Read-back
What the Lean code literally says, in plain math · gpt-6
For every measurable space , every probability measure on it, every nonzero natural number , every measurable map (with the Borel measurable structure on matrices), and every real number , suppose that equals its conjugate transpose for -almost every and that the matrix-valued function is Bochner integrable with respect to . Here is the normed-algebra matrix exponential, is the spectrum over the real scalars, and are the real-valued supremum and infimum used in the definitions, and . Then both of the following implications hold:
The measures in these inequalities are converted from extended nonnegative real values to real numbers; they are finite because is a probability measure. The threshold may be any real number, including zero and negative values, and both tail events include equality. Dimension zero is excluded. Hermitian symmetry is required only almost everywhere; the functions in the events are defined on all matrices by the real-spectrum supremum and infimum above, including any exceptional non-Hermitian values, with the total real supremum and infimum conventions (in particular, value zero for an empty real spectrum). When , neither antecedent holds, so the asserted conjunction imposes no tail inequality.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.