Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Lemma 1 — covariance inequality for bounded variables under strong mixing (Doukhan)

Proved
TPG.lemma_1_covariance_inequality

by Wenqian Xing · Oct 12, 2026 · Mathlib 0df444a (Lean v4.33.1)

causal-inferencemarkov-chainnonstationaritypolicy-gradienttreatment-effect

Let (Ω,F,P)(\Omega,\mathcal F,\mathbb P)(Ω,F,P) be a probability space and let U,V⊆F\mathcal U,\mathcal V\subseteq\mathcal FU,V⊆F be sub-σ\sigmaσ-algebras. Their strong mixing coefficient is

α(U,V)=sup⁡A∈U,  B∈V∣P(A∩B)−P(A) P(B)∣.\alpha(\mathcal U,\mathcal V)=\sup_{A\in\mathcal U,\;B\in\mathcal V}\bigl|\mathbb P(A\cap B)-\mathbb P(A)\,\mathbb P(B)\bigr| .α(U,V)=A∈U,B∈Vsup​​P(A∩B)−P(A)P(B)​.

Let XXX be a real random variable measurable with respect to U\mathcal UU and YYY one measurable with respect to V\mathcal VV, both bounded, with ∥X∥∞=inf⁡{M≥0:P(∣X∣>M)=0}\|X\|_\infty=\inf\{M\ge0:\mathbb P(|X|>M)=0\}∥X∥∞​=inf{M≥0:P(∣X∣>M)=0}.

Lemma.

∣Cov⁡(X,Y)∣≤4 α(U,V) ∥X∥∞ ∥Y∥∞.\bigl|\operatorname{Cov}(X,Y)\bigr|\le 4\,\alpha(\mathcal U,\mathcal V)\,\|X\|_\infty\,\|Y\|_\infty .​Cov(X,Y)​≤4α(U,V)∥X∥∞​∥Y∥∞​.

This inequality converts a bound on mixing coefficients into a bound on covariances of bounded functionals of separated parts of a process; the source uses it in the consistency proof of the variance estimator.

Formalization Note The formal statement takes real numbers aaa, bbb, α\alphaα with ∣X∣≤a|X|\le a∣X∣≤a and ∣Y∣≤b|Y|\le b∣Y∣≤b almost surely and ∣P(A∩B)−P(A)P(B)∣≤α|\mathbb P(A\cap B)-\mathbb P(A)\mathbb P(B)|\le\alpha∣P(A∩B)−P(A)P(B)∣≤α for all A∈UA\in\mathcal UA∈U, B∈VB\in\mathcal VB∈V, and concludes ∣Cov⁡(X,Y)∣≤4αab|\operatorname{Cov}(X,Y)|\le4\alpha ab∣Cov(X,Y)∣≤4αab. Since such a,b,αa,b,\alphaa,b,α are nonnegative and the bound is monotone in each, this is equivalent to the displayed inequality with the essential suprema and the supremum.

Preamble
import Mathlib

open MeasureTheory ProbabilityTheory
open scoped BigOperators ENNReal
Formal statement
namespace TPG
theorem lemma_1_covariance_inequality {Ω : Type} [mΩ : MeasurableSpace Ω]
    (μ : Measure Ω) [IsProbabilityMeasure μ]
    (mU mV : MeasurableSpace Ω) (hU : mU ≤ mΩ) (hV : mV ≤ mΩ)
    (X Y : Ω → ℝ) (hX : Measurable[mU] X) (hY : Measurable[mV] Y)
    (a b α : ℝ) (hXa : ∀ᵐ ω ∂μ, |X ω| ≤ a) (hYb : ∀ᵐ ω ∂μ, |Y ω| ≤ b)
    (hα : ∀ A B : Set Ω, MeasurableSet[mU] A → MeasurableSet[mV] B →
      |μ.real (A ∩ B) - μ.real A * μ.real B| ≤ α) :
    |cov[X, Y; μ]| ≤ 4 * α * a * b := by sorry
end TPG
Source
Johari, Peng, Xing, Estimation of Treatment Effects Under Nonstationarity via the Truncated Policy Gradient Estimator, revised manuscript of October 2026 (revision of arXiv:2506.05308v3, https://arxiv.org/abs/2506.05308); numbering follows the revised manuscript. Appendix A.4, p. 27, Lemma 1; cited there as Lemma 3 of P. Doukhan, Mixing: Properties and Examples, Lecture Notes in Statistics 85, Springer, 1994, https://doi.org/10.1007/978-1-4612-2642-0.
Read-back

What the Lean code literally says, in plain math · claude-opus-5-5

Objects quantified over. The statement is asserted for all of the following.

  • An arbitrary set Ω\OmegaΩ carrying a σ\sigmaσ-algebra F\mathcal{F}F, and a measure μ\muμ on (Ω,F)(\Omega,\mathcal{F})(Ω,F) that is a probability measure, i.e. μ(Ω)=1\mu(\Omega)=1μ(Ω)=1.
  • Two further σ\sigmaσ-algebras U\mathcal{U}U and V\mathcal{V}V on the same set Ω\OmegaΩ, each contained in F\mathcal{F}F: U⊆F\mathcal{U}\subseteq\mathcal{F}U⊆F and V⊆F\mathcal{V}\subseteq\mathcal{F}V⊆F. Nothing else is assumed about how U\mathcal{U}U and V\mathcal{V}V relate to one another (they may coincide, and no independence is assumed).
  • Two functions X,Y:Ω→RX,Y:\Omega\to\mathbb{R}X,Y:Ω→R such that XXX is measurable with respect to U\mathcal{U}U and YYY is measurable with respect to V\mathcal{V}V. In both cases the target R\mathbb{R}R carries its Borel σ\sigmaσ-algebra, and measurability is in the everywhere sense (not merely up to a null set).
  • Three real numbers aaa, bbb, α\alphaα. No sign condition on aaa, bbb or α\alphaα is written as a hypothesis.

Hypotheses.

  1. ∣X(ω)∣≤a\lvert X(\omega)\rvert\le a∣X(ω)∣≤a for μ\muμ-almost every ω∈Ω\omega\in\Omegaω∈Ω.

  2. ∣Y(ω)∣≤b\lvert Y(\omega)\rvert\le b∣Y(ω)∣≤b for μ\muμ-almost every ω∈Ω\omega\in\Omegaω∈Ω.

  3. For every pair of subsets A,B⊆ΩA,B\subseteq\OmegaA,B⊆Ω with A∈UA\in\mathcal{U}A∈U and B∈VB\in\mathcal{V}B∈V,

∣μ(A∩B)−μ(A)⋅μ(B)∣≤α.\bigl\lvert \mu(A\cap B)-\mu(A)\cdot\mu(B)\bigr\rvert\le\alpha .​μ(A∩B)−μ(A)⋅μ(B)​≤α.

Here μ(C)\mu(C)μ(C) denotes the measure of CCC read as a real number (the value in [0,∞][0,\infty][0,∞] converted to R\mathbb{R}R, with ∞\infty∞ sent to 000; since μ\muμ is a probability measure every value lies in [0,1][0,1][0,1] and this convention is never triggered).

Conclusion.

∣Cov⁡μ(X,Y)∣≤4⋅α⋅a⋅b,\bigl\lvert \operatorname{Cov}_\mu(X,Y)\bigr\rvert\le 4\cdot\alpha\cdot a\cdot b ,​Covμ​(X,Y)​≤4⋅α⋅a⋅b,

where the covariance is defined by

Cov⁡μ(X,Y)=∫Ω(X(ω)−∫ΩXdμ)⋅(Y(ω)−∫ΩYdμ)dμ(ω).\operatorname{Cov}_\mu(X,Y)=\int_\Omega\Bigl(X(\omega)-\int_\Omega X d\mu\Bigr)\cdot\Bigl(Y(\omega)-\int_\Omega Y d\mu\Bigr) d\mu(\omega).Covμ​(X,Y)=∫Ω​(X(ω)−∫Ω​Xdμ)⋅(Y(ω)−∫Ω​Ydμ)dμ(ω).

All three integrals are Lebesgue (Bochner) integrals with respect to μ\muμ on (Ω,F)(\Omega,\mathcal{F})(Ω,F), with the convention that the integral of a non-integrable function is 000. The inequality is non-strict, and the right-hand side is the product of the four real numbers 444, α\alphaα, aaa, bbb.

Edge cases and literal details.

  • The number α\alphaα is any real number that bounds the quantity in hypothesis 3 for all pairs (A,B)(A,B)(A,B); it is not required to be the supremum of those quantities.
  • Hypothesis 3 ranges over all pairs with A∈UA\in\mathcal{U}A∈U, B∈VB\in\mathcal{V}B∈V, including A=∅A=\emptysetA=∅, A=ΩA=\OmegaA=Ω, B=∅B=\emptysetB=∅, B=ΩB=\OmegaB=Ω. Taking A=B=∅A=B=\emptysetA=B=∅ shows that hypothesis 3 can only hold when α≥0\alpha\ge 0α≥0.
  • Because μ\muμ is a probability measure (so Ω\OmegaΩ is non-empty and a μ\muμ-almost-everywhere statement holds at some point), hypotheses 1 and 2 can only hold when a≥0a\ge 0a≥0 and b≥0b\ge 0b≥0. These sign facts are consequences of the hypotheses, not separate assumptions.
  • The bounds aaa and bbb are almost-everywhere bounds on ∣X∣\lvert X\rvert∣X∣ and ∣Y∣\lvert Y\rvert∣Y∣; the values a=0a=0a=0 or b=0b=0b=0 are allowed (then X=0X=0X=0 or Y=0Y=0Y=0 almost everywhere and the right-hand side is 000), and α=0\alpha=0α=0 is allowed (then the right-hand side is 000).
  • No integrability or moment hypothesis is stated separately; XXX and YYY are measurable and almost everywhere bounded on a probability space.
Human review
  • Endorsed by tianyipeng · Oct 12, 2026

  • Endorsed by Wenqian Xing · Oct 12, 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