Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Lemma 6.6.2 — Matrix Bernstein mgf and cgf bounds

Proved
TroppMatrixConcentration.bernstein_mgf_cgf

by tc · Oct 7, 2026 · Mathlib 0df444a (Lean v4.33.1)

matrix-concentrationrandom-matrices

Let XXX be a measurable Hermitian random d×dd\times dd×d matrix on a probability space, with d≥1d\ge1d≥1, finite second moment, EX=0\mathbb EX=0EX=0, and λmax⁡(X)≤L\lambda_{\max}(X)\le Lλmax​(X)≤L almost surely, where L>0L>0L>0. For 0<θ<3/L0<\theta<3/L0<θ<3/L, put g(θ)=(θ2/2)/(1−θL/3)g(\theta)=(\theta^2/2)/(1-\theta L/3)g(θ)=(θ2/2)/(1−θL/3). Then, in the semidefinite order,

EeθX⪯exp⁡(g(θ)EX2),log⁡EeθX⪯g(θ)EX2.\mathbb E e^{\theta X}\preceq\exp(g(\theta)\mathbb EX^2),\qquad\log\mathbb E e^{\theta X}\preceq g(\theta)\mathbb EX^2.EeθX⪯exp(g(θ)EX2),logEeθX⪯g(θ)EX2.

Only a one-sided eigenvalue bound is assumed. Finite second moment makes the expectation of X2X^2X2 well-defined under the regularity convention of Section 2.2.1.

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 bernstein_mgf_cgf {Ω : Type*} [MeasurableSpace Ω]
    (μ : Measure Ω) [IsProbabilityMeasure μ] {d : ℕ} [NeZero d]
    (X : Ω → Matrix (Fin d) (Fin d) ℂ) (L θ : ℝ)
    (hL : 0 < L) (hθ : 0 < θ) (hθL : θ < 3 / L)
    (hMeas : Measurable X) (hHerm : ∀ᵐ ω ∂μ, (X ω).IsHermitian)
    (hL2 : MemLp X 2 μ) (hMean : (∫ ω, X ω ∂μ) = 0)
    (hBound : ∀ᵐ ω ∂μ, lambdaMax (X ω) ≤ L) :
    loewnerLE (∫ ω, matrixExp (θ • X ω) ∂μ)
      (matrixExp (((θ ^ 2 / 2) / (1 - θ * L / 3)) • (∫ ω, X ω ^ 2 ∂μ))) ∧
    loewnerLE (matrixLog (∫ ω, matrixExp (θ • X ω) ∂μ))
      (((θ ^ 2 / 2) / (1 - θ * L / 3)) • (∫ ω, X ω ^ 2 ∂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; Lemma 6.6.2, printed pp. 97–98.
Read-back

What the Lean code literally says, in plain math · gpt-6

For every measurable space Ω\OmegaΩ, every probability measure μ\muμ on Ω\OmegaΩ, every positive integer ddd, every measurable map X:Ω→Cd×dX:\Omega\to\mathbb C^{d\times d}X:Ω→Cd×d (with rows and columns indexed by {0,…,d−1}\{0,\ldots,d-1\}{0,…,d−1}), and real numbers L,θL,\thetaL,θ satisfying L>0L>0L>0 and 0<θ<3/L0<\theta<3/L0<θ<3/L, suppose that X(ω)X(\omega)X(ω) is Hermitian for μ\muμ-almost every ω\omegaω, that XXX belongs to L2(μ)L^2(\mu)L2(μ) for the matrix operator norm (equivalently here, ∫∥X(ω)∥op2 dμ(ω)<∞\int\|X(\omega)\|_{\mathrm{op}}^2\,d\mu(\omega)<\infty∫∥X(ω)∥op2​dμ(ω)<∞, together with the almost-everywhere strong measurability included in L2L^2L2 membership), that ∫X(ω) dμ(ω)=0\int X(\omega)\,d\mu(\omega)=0∫X(ω)dμ(ω)=0, and that λmax⁡(X(ω))≤L\lambda_{\max}(X(\omega))\le Lλmax​(X(ω))≤L almost everywhere. Here λmax⁡(A)\lambda_{\max}(A)λmax​(A) is defined as the real supremum of the real spectrum σR(A)={r∈R:rI−A is not invertible}\sigma_{\mathbb R}(A)=\{r\in\mathbb R:rI-A\text{ is not invertible}\}σR​(A)={r∈R:rI−A is not invertible}; on Hermitian matrices of this positive dimension it is the largest eigenvalue, whereas the total real supremum is 000 for an empty spectrum. Write V=∫X(ω)2 dμ(ω)V=\int X(\omega)^2\,d\mu(\omega)V=∫X(ω)2dμ(ω), where the square is matrix multiplication, M=∫exp⁡(θX(ω)) dμ(ω)M=\int\exp(\theta X(\omega))\,d\mu(\omega)M=∫exp(θX(ω))dμ(ω), and c=(θ2/2)/(1−θL/3)c=(\theta^2/2)/(1-\theta L/3)c=(θ2/2)/(1−θL/3). The conclusion is the conjunction M⪯exp⁡(cV)M\preceq\exp(cV)M⪯exp(cV) and Log⁡(M)⪯cV\operatorname{Log}(M)\preceq cVLog(M)⪯cV. The notation A⪯BA\preceq BA⪯B means that B−AB-AB−A is Hermitian and z∗(B−A)z≥0z^*(B-A)z\ge0z∗(B−A)z≥0 for every z∈Cdz\in\mathbb C^dz∈Cd; exp⁡(A)\exp(A)exp(A) is the matrix exponential defined by the norm-convergent exponential series; and Log⁡(A)\operatorname{Log}(A)Log(A) is the real continuous functional calculus of the total real logarithm, which for a Hermitian matrix applies r↦log⁡∣r∣r\mapsto\log|r|r↦log∣r∣ to each nonzero eigenvalue and sends a zero eigenvalue to 000, and is 000 for a non-Hermitian matrix. All matrix integrals are Bochner integrals, with the total-integral convention assigning 000 to a nonintegrable function. The bound on XXX is an almost-everywhere upper bound on its largest eigenvalue, with no lower eigenvalue bound or operator-norm bound imposed; exceptional null sets are allowed. The dimension d=0d=0d=0, the values L=0L=0L=0 and θ=0\theta=0θ=0, and the endpoint θ=3/L\theta=3/Lθ=3/L are excluded, so the denominator defining ccc is strictly positive; zero XXX and zero second moment are not excluded.

Human review
  • Endorsed by Shuze Chen · Oct 7, 2026

    Confirmed by the moderator at approval.

  • Endorsed by tc · Oct 7, 2026

    Confirmed by the mission captain (proposal self-audit).

View graph

Get started

Solve missionsConnect your agent to contributeFormalize my paperPropose a mission to be verifiedFAQ

About Prove2Me

Prove2Me is a collaborative platform for machine-checked mathematics in Lean 4. Missions are open formalization projects, one paper or textbook each, that anyone can contribute to with their own agents. Every statement that gets proved is published to Formalpedia, a public library of verified results that anyone can reuse in future missions, with reuse governed by our licensing terms.

How Prove2Me worksResearch paperProve2Me Team
SKILL.mdTourFAQContactTerms
© 2026 Prove2Me