Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 7.3.1 — Intrinsic matrix Bernstein

Proved
TroppMatrixConcentration.ch7_intrinsic_bernstein

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

matrix-concentrationrandom-matrices

Let SkS_kSk​ be a finite independent family of measurable complex m×nm\times nm×n random matrices on a probability space, m,n≥1m,n\ge1m,n≥1, with ESk=0\mathbb ES_k=0ESk​=0 and ∥Sk∥≤L\|S_k\|\le L∥Sk​∥≤L almost surely, L≥0L\ge0L≥0. Set Z=∑kSkZ=\sum_kS_kZ=∑k​Sk​. Let V1⪰EZZ∗V_1\succeq\mathbb EZZ^*V1​⪰EZZ∗ and V2⪰EZ∗ZV_2\succeq\mathbb EZ^*ZV2​⪰EZ∗Z, with at least one nonzero, and put r=r(diag⁡(V1,V2))r=r(\operatorname{diag}(V_1,V_2))r=r(diag(V1​,V2​)), v=max⁡{∥V1∥,∥V2∥}v=\max\{\|V_1\|,\|V_2\|\}v=max{∥V1​∥,∥V2​∥}. Then

EZZ∗=∑kESkSk∗,EZ∗Z=∑kESk∗Sk,\mathbb EZZ^*=\sum_k\mathbb ES_kS_k^*,\qquad\mathbb EZ^*Z=\sum_k\mathbb ES_k^*S_k,EZZ∗=k∑​ESk​Sk∗​,EZ∗Z=k∑​ESk∗​Sk​, P{∥Z∥≥t}≤4rexp⁡ ⁣(−t2/2v+Lt/3)(t≥v+L/3).\mathbb P\{\|Z\|\ge t\}\le4r\exp\!\left(-\frac{t^2/2}{v+Lt/3}\right)\qquad(t\ge\sqrt v+L/3).P{∥Z∥≥t}≤4rexp(−v+Lt/3t2/2​)(t≥v​+L/3).

The variance bounds are matrices, not merely scalar norm bounds. The nonzero block proxy keeps intrinsic dimension in the source's meaningful domain. The finite family may be empty, and deterministic zero summands remain permitted. All required moments follow from boundedness. This is the designated goal; Theorem 7.2.1 is the other main theorem in this chapter proposal.

Preamble
import Definitions.Def_TroppMatrixConcentration_ch7_intrinsic

open MeasureTheory ProbabilityTheory
open scoped Matrix.Norms.L2Operator ComplexOrder

Formal statement
namespace TroppMatrixConcentration

theorem ch7_intrinsic_bernstein {Ω : Type*} [MeasurableSpace Ω]
    (μ : Measure Ω) [IsProbabilityMeasure μ] {m n N : ℕ} [NeZero m] [NeZero n]
    (S : Fin N → Ω → Matrix (Fin m) (Fin n) ℂ) (L : ℝ) (hL : 0 ≤ L)
    (V₁ : Matrix (Fin m) (Fin m) ℂ) (V₂ : Matrix (Fin n) (Fin n) ℂ)
    (hV : V₁ ≠ 0 ∨ V₂ ≠ 0)
    (hMeas : ∀ k, Measurable (S k)) (hIndep : iIndepFun S μ)
    (hMean : ∀ k, (∫ ω, S k ω ∂μ) = 0)
    (hBound : ∀ k, ∀ᵐ ω ∂μ, spectralNorm (S k ω) ≤ L)
    (hVariance₁ : loewnerLE
      (∫ ω, (∑ k, S k ω) * (∑ k, S k ω).conjTranspose ∂μ) V₁)
    (hVariance₂ : loewnerLE
      (∫ ω, (∑ k, S k ω).conjTranspose * (∑ k, S k ω) ∂μ) V₂) :
    let Z := fun ω => ∑ k, S k ω
    let r := intrinsicDimension (Matrix.fromBlocks V₁ 0 0 V₂)
    let v := max (spectralNorm V₁) (spectralNorm V₂)
    (∫ ω, Z ω * (Z ω).conjTranspose ∂μ) =
      ∑ k, ∫ ω, S k ω * (S k ω).conjTranspose ∂μ ∧
    (∫ ω, (Z ω).conjTranspose * Z ω ∂μ) =
      ∑ k, ∫ ω, (S k ω).conjTranspose * S k ω ∂μ ∧
    ∀ t : ℝ, Real.sqrt v + L / 3 ≤ t →
      (μ {ω | t ≤ spectralNorm (Z ω)}).toReal ≤
        4 * r * Real.exp (-(t ^ 2 / 2) / (v + L * t / 3)) := 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 7.3.1, equations (7.3.1–2), printed p. 108; independence confirmed in Section 7.7.3, printed p. 117.
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 natural numbers m,nm,nm,n, every natural number NNN (including N=0N=0N=0), every family of measurable, mutually independent random matrices Sk:Ω→Cm×nS_k:\Omega\to\mathbb C^{m\times n}Sk​:Ω→Cm×n indexed by k∈{0,…,N−1}k\in\{0,\ldots,N-1\}k∈{0,…,N−1} (with the Borel measurable structure on matrices and independence taken with respect to μ\muμ), every real number L≥0L\ge0L≥0, and every pair of matrices V1∈Cm×mV_1\in\mathbb C^{m\times m}V1​∈Cm×m and V2∈Cn×nV_2\in\mathbb C^{n\times n}V2​∈Cn×n with V1≠0V_1\ne0V1​=0 or V2≠0V_2\ne0V2​=0, assume that each matrix-valued Bochner integral ∫Sk dμ\int S_k\,d\mu∫Sk​dμ equals zero and that ∥Sk(ω)∥2→2≤L\|S_k(\omega)\|_{2\to2}\le L∥Sk​(ω)∥2→2​≤L for μ\muμ-almost every ω\omegaω, separately for every kkk. Write Z(ω)=∑k=0N−1Sk(ω)Z(\omega)=\sum_{k=0}^{N-1}S_k(\omega)Z(ω)=∑k=0N−1​Sk​(ω), and further assume that both V1−∫ZZ∗ dμV_1-\int ZZ^*\,d\muV1​−∫ZZ∗dμ and V2−∫Z∗Z dμV_2-\int Z^*Z\,d\muV2​−∫Z∗Zdμ are positive semidefinite, meaning Hermitian with nonnegative quadratic forms; here A∗A^*A∗ is the conjugate transpose and ∥A∥2→2\|A\|_{2\to2}∥A∥2→2​ is the operator norm for the Euclidean norms. Set D=(V100V2)D=\begin{pmatrix}V_1&0\\0&V_2\end{pmatrix}D=(V1​0​0V2​​), r=Re⁡(tr⁡D)/∥D∥2→2r=\operatorname{Re}(\operatorname{tr}D)/\|D\|_{2\to2}r=Re(trD)/∥D∥2→2​, and v=max⁡{∥V1∥2→2,∥V2∥2→2}v=\max\{\|V_1\|_{2\to2},\|V_2\|_{2\to2}\}v=max{∥V1​∥2→2​,∥V2​∥2→2​}. Then all three of the following hold: ∫ZZ∗ dμ=∑k=0N−1∫SkSk∗ dμ\int ZZ^*\,d\mu=\sum_{k=0}^{N-1}\int S_kS_k^*\,d\mu∫ZZ∗dμ=∑k=0N−1​∫Sk​Sk∗​dμ; ∫Z∗Z dμ=∑k=0N−1∫Sk∗Sk dμ\int Z^*Z\,d\mu=\sum_{k=0}^{N-1}\int S_k^*S_k\,d\mu∫Z∗Zdμ=∑k=0N−1​∫Sk∗​Sk​dμ; and, for every real ttt satisfying v+L/3≤t\sqrt v+L/3\le tv​+L/3≤t, the probability, converted from an extended nonnegative real to a real number, satisfies μ{ω:t≤∥Z(ω)∥2→2}≤4rexp⁡ ⁣(−t2/2v+Lt/3)\mu\{\omega:t\le\|Z(\omega)\|_{2\to2}\}\le4r\exp\!\left(-\dfrac{t^2/2}{v+Lt/3}\right)μ{ω:t≤∥Z(ω)∥2→2​}≤4rexp(−v+Lt/3t2/2​). When N=0N=0N=0, all indexed hypotheses are vacuous and all sums are zero; the same nonzero requirement on the pair (V1,V2)(V_1,V_2)(V1​,V2​) and the two positive semidefinite difference hypotheses still apply. The case L=0L=0L=0 is included. The nonzero requirement ensures ∥D∥2→2>0\|D\|_{2\to2}>0∥D∥2→2​>0 and v>0v>0v>0, so the displayed divisions have nonzero denominators for every permitted ttt.

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