Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 6.1.1 — Matrix Bernstein for rectangular matrices

Proved
TroppMatrixConcentration.matrix_bernstein

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

matrix-concentrationrandom-matrices

Let S1,…,SNS_1,\ldots,S_NS1​,…,SN​ be independent measurable complex m×nm\times nm×n random matrices on a probability space, with m,n≥1m,n\ge1m,n≥1. Assume ESk=0\mathbb ES_k=0ESk​=0 and ∥Sk∥≤L\|S_k\|\le L∥Sk​∥≤L almost surely, where L≥0L\ge0L≥0 and the norm is the spectral norm. Set Z=∑kSkZ=\sum_kS_kZ=∑k​Sk​. Then

v=max⁡{∥EZZ∗∥,∥EZ∗Z∥}=max⁡{∥∑kESkSk∗∥,∥∑kESk∗Sk∥},v=\max\{\|\mathbb EZZ^*\|,\|\mathbb EZ^*Z\|\}=\max\left\{\left\|\sum_k\mathbb ES_kS_k^*\right\|,\left\|\sum_k\mathbb ES_k^*S_k\right\|\right\},v=max{∥EZZ∗∥,∥EZ∗Z∥}=max{​k∑​ESk​Sk∗​​,​k∑​ESk∗​Sk​​}, E∥Z∥≤2vlog⁡(m+n)+L3log⁡(m+n),\mathbb E\|Z\|\le\sqrt{2v\log(m+n)}+\frac L3\log(m+n),E∥Z∥≤2vlog(m+n)​+3L​log(m+n), P{∥Z∥≥t}≤(m+n)exp⁡(−t2/2v+Lt/3)(t≥0).\mathbb P\{\|Z\|\ge t\}\le(m+n)\exp\left(-\frac{t^2/2}{v+Lt/3}\right)\quad(t\ge0).P{∥Z∥≥t}≤(m+n)exp(−v+Lt/3t2/2​)(t≥0).

At a zero denominator the tail right-hand side is m+nm+nm+n for t=0t=0t=0 and zero for t>0t>0t>0. The finite family may be empty. All moments needed in this statement exist by boundedness. This is the mission's goal, retaining both the expectation and tail conclusions.

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 matrix_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)
    (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) :
    let Z := fun ω => ∑ k, S k ω
    let v := rectSecondMoment μ Z
    v = max (spectralNorm (∑ k, ∫ ω, S k ω * (S k ω).conjTranspose ∂μ))
      (spectralNorm (∑ k, ∫ ω, (S k ω).conjTranspose * S k ω ∂μ)) ∧
    (∫ ω, spectralNorm (Z ω) ∂μ) ≤
      Real.sqrt (2 * v * Real.log (m + n)) + L * Real.log (m + n) / 3 ∧
    ∀ t : ℝ, 0 ≤ t → (μ {ω | t ≤ spectralNorm (Z ω)}).toReal ≤
      bernsteinTail (m + n) v L t := 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 6.1.1, equations (6.1.1–4), printed pp. 76.
Read-back

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

Let Ω\OmegaΩ be any measurable space, let μ\muμ be a probability measure on it, let m,nm,nm,n be positive integers and NNN any nonnegative integer, and let 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}, be measurable and jointly independent with respect to μ\muμ. Write A∗A^*A∗ for conjugate transpose and ∥A∥\|A\|∥A∥ for the operator norm induced by the Euclidean norms on the complex coordinate spaces, and interpret all matrix-valued integrals as Bochner integrals. Suppose L∈RL\in\mathbb RL∈R satisfies L≥0L\ge0L≥0, each ∫Sk dμ=0\int S_k\,d\mu=0∫Sk​dμ=0, and, for every kkk, ∥Sk(ω)∥≤L\|S_k(\omega)\|\le L∥Sk​(ω)∥≤L for μ\muμ-almost every ω\omegaω. Define Z(ω)=∑k=0N−1Sk(ω)Z(\omega)=\sum_{k=0}^{N-1}S_k(\omega)Z(ω)=∑k=0N−1​Sk​(ω) and v=max⁡{∥∫ZZ∗ dμ∥,∥∫Z∗Z dμ∥}v=\max\{\|\int ZZ^*\,d\mu\|,\|\int Z^*Z\,d\mu\|\}v=max{∥∫ZZ∗dμ∥,∥∫Z∗Zdμ∥}. Then all three of the following hold: v=max⁡{∥∑k=0N−1∫SkSk∗ dμ∥,∥∑k=0N−1∫Sk∗Sk dμ∥}v=\max\{\|\sum_{k=0}^{N-1}\int S_kS_k^*\,d\mu\|,\|\sum_{k=0}^{N-1}\int S_k^*S_k\,d\mu\|\}v=max{∥∑k=0N−1​∫Sk​Sk∗​dμ∥,∥∑k=0N−1​∫Sk∗​Sk​dμ∥}; ∫∥Z(ω)∥ dμ(ω)≤2vlog⁡(m+n)+Llog⁡(m+n)/3\int\|Z(\omega)\|\,d\mu(\omega)\le\sqrt{2v\log(m+n)}+L\log(m+n)/3∫∥Z(ω)∥dμ(ω)≤2vlog(m+n)​+Llog(m+n)/3, where log⁡\loglog is the natural logarithm; and for every real t≥0t\ge0t≥0, μ({ω:t≤∥Z(ω)∥})≤B(t)\mu(\{\omega:t\le\|Z(\omega)\|\})\le B(t)μ({ω:t≤∥Z(ω)∥})≤B(t), with the probability viewed as a real number and with B(t)B(t)B(t) defined to be m+nm+nm+n if v+Lt/3=0v+Lt/3=0v+Lt/3=0 and t=0t=0t=0, to be 000 if v+Lt/3=0v+Lt/3=0v+Lt/3=0 and t≠0t\ne0t=0, and otherwise to be (m+n)exp⁡(−(t2/2)/(v+Lt/3))(m+n)\exp(-(t^2/2)/(v+Lt/3))(m+n)exp(−(t2/2)/(v+Lt/3)). The quantifiers include L=0L=0L=0, v=0v=0v=0, and N=0N=0N=0; when N=0N=0N=0 the sums are empty, Z=0Z=0Z=0 and v=0v=0v=0, and the assumptions indexed by kkk are vacuous. In particular t=0t=0t=0 is included, its event is all of Ω\OmegaΩ, and its asserted upper bound is m+nm+nm+n in every case; zero values of v+Lt/3v+Lt/3v+Lt/3 are governed by the stated branches rather than division by zero.

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