Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 5.1.1 — Matrix Chernoff, both spectral sides

Proved
TroppMatrixConcentration.ch5_matrix_chernoff

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

matrix-concentrationrandom-matrices

Let XkX_kXk​ be a finite independent family of measurable random complex d×dd\times dd×d matrices on a probability space, with d≥1d\ge1d≥1. Assume each XkX_kXk​ is Hermitian almost surely and satisfies 0≤λmin⁡(Xk)0\le\lambda_{\min}(X_k)0≤λmin​(Xk​) and λmax⁡(Xk)≤L\lambda_{\max}(X_k)\le Lλmax​(Xk​)≤L almost surely, for a common L≥0L\ge0L≥0. Define Y=∑kXkY=\sum_k X_kY=∑k​Xk​, a=λmin⁡(EY)a=\lambda_{\min}(\mathbb EY)a=λmin​(EY), and b=λmax⁡(EY)b=\lambda_{\max}(\mathbb EY)b=λmax​(EY). Then

a=λmin⁡(∑kEXk),b=λmax⁡(∑kEXk).a=\lambda_{\min}\left(\sum_k\mathbb EX_k\right),\qquad b=\lambda_{\max}\left(\sum_k\mathbb EX_k\right).a=λmin​(k∑​EXk​),b=λmax​(k∑​EXk​).

For every θ>0\theta>0θ>0, both expectation inequalities hold:

Eλmin⁡(Y)≥1−e−θθa−Llog⁡dθ,Eλmax⁡(Y)≤eθ−1θb+Llog⁡dθ.\mathbb E\lambda_{\min}(Y)\ge\frac{1-e^{-\theta}}{\theta}a-\frac{L\log d}{\theta},\qquad\mathbb E\lambda_{\max}(Y)\le\frac{e^\theta-1}{\theta}b+\frac{L\log d}{\theta}.Eλmin​(Y)≥θ1−e−θ​a−θLlogd​,Eλmax​(Y)≤θeθ−1​b+θLlogd​.

For 0≤ε<10\le\varepsilon<10≤ε<1,

P{λmin⁡(Y)≤(1−ε)a}≤d[e−ε(1−ε)1−ε]a/L.\mathbb P\{\lambda_{\min}(Y)\le(1-\varepsilon)a\}\le d\left[\frac{e^{-\varepsilon}}{(1-\varepsilon)^{1-\varepsilon}}\right]^{a/L}.P{λmin​(Y)≤(1−ε)a}≤d[(1−ε)1−εe−ε​]a/L.

For every ε≥0\varepsilon\ge0ε≥0,

P{λmax⁡(Y)≥(1+ε)b}≤d[eε(1+ε)1+ε]b/L.\mathbb P\{\lambda_{\max}(Y)\ge(1+\varepsilon)b\}\le d\left[\frac{e^\varepsilon}{(1+\varepsilon)^{1+\varepsilon}}\right]^{b/L}.P{λmax​(Y)≥(1+ε)b}≤d[(1+ε)1+εeε​]b/L.

At L=0L=0L=0 both tail expressions are explicitly ddd; all summands then vanish almost surely. The family may be empty, and either mean eigenvalue may be zero. This goal retains every conclusion of the source theorem and the distinct lower- and upper-tail deviation domains.

Preamble
import Definitions.Def_TroppMatrixConcentration_ch5_chernoff_functions

open MeasureTheory ProbabilityTheory
open scoped Matrix.Norms.L2Operator

Formal statement
namespace TroppMatrixConcentration

theorem ch5_matrix_chernoff {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)
    (hMeas : ∀ k, Measurable (X k)) (hIndep : iIndepFun X μ)
    (hHerm : ∀ k, ∀ᵐ ω ∂μ, (X k ω).IsHermitian)
    (hBound : ∀ k, ∀ᵐ ω ∂μ, 0 ≤ lambdaMin (X k ω) ∧ lambdaMax (X k ω) ≤ L) :
    let Y := fun ω => ∑ k, X k ω
    let a := lambdaMin (∫ ω, Y ω ∂μ)
    let b := lambdaMax (∫ ω, Y ω ∂μ)
    a = lambdaMin (∑ k, ∫ ω, X k ω ∂μ) ∧
    b = lambdaMax (∑ k, ∫ ω, X k ω ∂μ) ∧
    (∀ θ : ℝ, 0 < θ →
      (1 - Real.exp (-θ)) / θ * a - L * Real.log d / θ ≤
        (∫ ω, lambdaMin (Y ω) ∂μ) ∧
      (∫ ω, lambdaMax (Y ω) ∂μ) ≤
        (Real.exp θ - 1) / θ * b + L * Real.log d / θ) ∧
    (∀ ε : ℝ, 0 ≤ ε → ε < 1 →
      (μ {ω | lambdaMin (Y ω) ≤ (1 - ε) * a}).toReal ≤
        chernoffLowerTail d a L ε) ∧
    (∀ ε : ℝ, 0 ≤ ε →
      (μ {ω | (1 + ε) * b ≤ lambdaMax (Y ω)}).toReal ≤
        chernoffUpperTail d b L ε) := 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; Theorem 5.1.1, equations (5.1.1–6), printed p. 60.
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 it, natural numbers d≥1d\geq 1d≥1 and N≥0N\geq 0N≥0, a family Xk:Ω→Cd×dX_k:\Omega\to\mathbb C^{d\times d}Xk​:Ω→Cd×d indexed by k∈{0,…,N−1}k\in\{0,\ldots,N-1\}k∈{0,…,N−1}, and a real number L≥0L\geq0L≥0, assume that each XkX_kXk​ is measurable for the Borel measurable structure on complex matrices, that the whole family is jointly independent under μ\muμ, and that for each kkk, almost surely Xk=Xk∗X_k=X_k^*Xk​=Xk∗​, and almost surely both 0≤λ−(Xk)0\leq\lambda_-(X_k)0≤λ−​(Xk​) and λ+(Xk)≤L\lambda_+(X_k)\leq Lλ+​(Xk​)≤L. Here, for any complex square matrix AAA, λ−(A)=inf⁡σR(A)\lambda_-(A)=\inf\sigma_{\mathbb R}(A)λ−​(A)=infσR​(A) and λ+(A)=sup⁡σR(A)\lambda_+(A)=\sup\sigma_{\mathbb R}(A)λ+​(A)=supσR​(A), where 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}; these are the least and greatest eigenvalues for Hermitian AAA, and the infimum and supremum of the empty real spectrum are both assigned the value zero. Set Y(ω)=∑k=0N−1Xk(ω)Y(\omega)=\sum_{k=0}^{N-1}X_k(\omega)Y(ω)=∑k=0N−1​Xk​(ω), a=λ−(∫Y dμ)a=\lambda_-(\int Y\,d\mu)a=λ−​(∫Ydμ), and b=λ+(∫Y dμ)b=\lambda_+(\int Y\,d\mu)b=λ+​(∫Ydμ), using Bochner integrals for matrices and real-valued functions (with the integral assigned zero when the function is not integrable). Then a=λ−(∑k=0N−1∫Xk dμ)a=\lambda_-(\sum_{k=0}^{N-1}\int X_k\,d\mu)a=λ−​(∑k=0N−1​∫Xk​dμ) and b=λ+(∑k=0N−1∫Xk dμ)b=\lambda_+(\sum_{k=0}^{N-1}\int X_k\,d\mu)b=λ+​(∑k=0N−1​∫Xk​dμ); for every real θ>0\theta>0θ>0, both 1−e−θθa−Llog⁡dθ≤∫λ−(Y) dμ\frac{1-e^{-\theta}}{\theta}a-\frac{L\log d}{\theta}\leq\int\lambda_-(Y)\,d\muθ1−e−θ​a−θLlogd​≤∫λ−​(Y)dμ and ∫λ+(Y) dμ≤eθ−1θb+Llog⁡dθ\int\lambda_+(Y)\,d\mu\leq\frac{e^{\theta}-1}{\theta}b+\frac{L\log d}{\theta}∫λ+​(Y)dμ≤θeθ−1​b+θLlogd​ hold; for every real 0≤ε<10\leq\varepsilon<10≤ε<1, μ{ω:λ−(Y(ω))≤(1−ε)a}≤d\mu\{\omega:\lambda_-(Y(\omega))\leq(1-\varepsilon)a\}\leq dμ{ω:λ−​(Y(ω))≤(1−ε)a}≤d if L=0L=0L=0, and otherwise this probability is at most d(e−ε/(1−ε)1−ε)a/Ld\bigl(e^{-\varepsilon}/(1-\varepsilon)^{1-\varepsilon}\bigr)^{a/L}d(e−ε/(1−ε)1−ε)a/L; and for every real ε≥0\varepsilon\geq0ε≥0, μ{ω:(1+ε)b≤λ+(Y(ω))}≤d\mu\{\omega:(1+\varepsilon)b\leq\lambda_+(Y(\omega))\}\leq dμ{ω:(1+ε)b≤λ+​(Y(ω))}≤d if L=0L=0L=0, and otherwise this probability is at most d(eε/(1+ε)1+ε)b/Ld\bigl(e^{\varepsilon}/(1+\varepsilon)^{1+\varepsilon}\bigr)^{b/L}d(eε/(1+ε)1+ε)b/L. Probabilities in these inequalities are the real values of their finite measure values, all displayed powers are real powers, and log⁡\loglog is the natural logarithm. The hypotheses require no positive lower bound on LLL, aaa, or bbb; the L=0L=0L=0 branches use exactly ddd without division by LLL. The allowed value N=0N=0N=0 makes all family hypotheses vacuous, gives Y=0Y=0Y=0 and a=b=0a=b=0a=b=0, and makes both tail events the whole space with probability one and both tail bounds equal to ddd; d=0d=0d=0 is excluded, d=1d=1d=1 is allowed and gives log⁡d=0\log d=0logd=0, θ=0\theta=0θ=0 is excluded, ε=0\varepsilon=0ε=0 is included in both tail assertions and gives bound ddd, and ε=1\varepsilon=1ε=1 is excluded only from the lower-tail assertion. Both tail events include equality, and every asserted inequality is non-strict.

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).

  • Endorsed by iris · Nov 6, 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