Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Lemma 2 — unchanged sub-Gaussian parameter under bin mixing

Disproved
ChenGallegoPricing.ABE.lemma_2

by Shuze Chen · Oct 29, 2026 · Mathlib 0df444a (Lean v4.33.1)

bandit-algorithmsdynamic-pricingoperations-researchp2o-gran-per-chapterp2o-plan-paperp2o-strengthenedp2o-v1

Let μ\muμ be a probability measure on a measurable space XXX, let BBB be measurable with μ(B)>0\mu(B)>0μ(B)>0, and let κx\kappa_xκx​ be a measurable family of real probability laws. Fix σ≥0\sigma\ge0σ≥0. Suppose each component with x∈Bx\in Bx∈B has a finite first moment m(x)m(x)m(x) and satisfies

Eκxet(Z−m(x))≤eσt2(t∈R),\mathbb E_{\kappa_x}e^{t(Z-m(x))}\le e^{\sigma t^2}\qquad(t\in\mathbb R),Eκx​​et(Z−m(x))≤eσt2(t∈R),

with finite exponential moments. Let η=∫κx μ(dx∣B)\eta=\int\kappa_x\,\mu(dx\mid B)η=∫κx​μ(dx∣B) and assume its first moment is finite. The printed assertion is that

Eηet(Z−EηZ)≤eσt2(t∈R),\mathbb E_\eta e^{t(Z-\mathbb E_\eta Z)}\le e^{\sigma t^2}\qquad(t\in\mathbb R),Eη​et(Z−Eη​Z)≤eσt2(t∈R),

including existence of the finite exponential moments. No equal-component-means hypothesis is imposed. The price is fixed throughout and is suppressed from the notation. This is the source's assertion; insight I1 records its bounded counterexample.

Differences from the source:

  • encoding: the displayed means and conditional distribution → finite first moments and conditioning on a positive-probability measurable event; a Markov kernel represents the family at the fixed price; reason: these are the domains where the source’s displayed expectations and conditioning are defined
  • encoding: the bound exp(σt²) → Mathlib HasSubgaussianMGF for the centered variable with variance proxy 2σ, including integrability of every exponential; reason: Mathlib uses exp(c t²/2); this preserves σ exactly and avoids default-zero nonintegrable expectations
  • generalization: a componentwise bound for given covariates in the model → arbitrary measurable mixing space and component hypotheses required only on B; reason: values outside the conditioning event have no effect; the original covariate-space claim is a specialization
Preamble
import Mathlib

open MeasureTheory ProbabilityTheory
open scoped NNReal
Formal statement
namespace ChenGallegoPricing.ABE

theorem lemma_2 {X : Type*} [MeasurableSpace X]
    (μ : Measure X) [IsProbabilityMeasure μ] (B : Set X)
    (hB : MeasurableSet B) (hpos : μ B ≠ 0)
    (κ : Kernel X ℝ) [IsMarkovKernel κ] (σ : ℝ≥0)
    (hcomponents : ∀ x ∈ B, Integrable id (κ x) ∧
      HasSubgaussianMGF (fun z => z - ∫ y, y ∂κ x) (2 * σ) (κ x))
    (hmean : Integrable id (κ ∘ₘ ProbabilityTheory.cond μ B)) :
    HasSubgaussianMGF
      (fun z => z - ∫ y, y ∂(κ ∘ₘ ProbabilityTheory.cond μ B))
      (2 * σ) (κ ∘ₘ ProbabilityTheory.cond μ B) := by sorry

end ChenGallegoPricing.ABE
Source
Chen and Gallego, Nonparametric Pricing Analytics with Customer Covariates, arXiv:1805.01136v3 (16 February 2020), https://arxiv.org/abs/1805.01136v3; Lemma 2, printed/PDF p. 27
Read-back

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

Setting.

  • XXX is any measurable space.
  • μ\muμ is a probability measure on XXX.
  • B⊆XB\subseteq XB⊆X is a measurable set with μ(B)≠0\mu(B)\ne0μ(B)=0.
  • μ(⋅∣B)=μ(B)−1μ∣B\mu(\cdot\mid B)=\mu(B)^{-1}\mu|_Bμ(⋅∣B)=μ(B)−1μ∣B​ is the conditional probability given BBB.
  • κ\kappaκ is a Markov kernel from XXX to R\mathbb{R}R: each κx\kappa_xκx​ is a probability measure on R\mathbb{R}R, measurable in xxx.
  • σ≥0\sigma\ge0σ≥0 is a real number.
  • κˉ=∫κx μ(dx∣B)\bar\kappa=\int \kappa_x\,\mu(dx\mid B)κˉ=∫κx​μ(dx∣B) is the mixture probability measure on R\mathbb{R}R.

Hypotheses.

  1. For every x∈Bx\in Bx∈B:
    • κx\kappa_xκx​ has a finite first moment, with mean mx=∫y κx(dy)m_x=\int y\,\kappa_x(dy)mx​=∫yκx​(dy);
    • the centred variable z↦z−mxz\mapsto z-m_xz↦z−mx​ under κx\kappa_xκx​ is sub-Gaussian with variance proxy 2σ2\sigma2σ. That is, for every t∈Rt\in\mathbb{R}t∈R, z↦et(z−mx)z\mapsto e^{t(z-m_x)}z↦et(z−mx​) is κx\kappa_xκx​-integrable and
∫et(z−mx) κx(dz)≤exp⁡ ⁣(2σ t22)=eσt2.\int e^{t(z-m_x)}\,\kappa_x(dz)\le \exp\!\Big(\frac{2\sigma\,t^2}{2}\Big)=e^{\sigma t^2}.∫et(z−mx​)κx​(dz)≤exp(22σt2​)=eσt2.
  1. κˉ\bar\kappaκˉ has a finite first moment.

Conclusion. Let mˉ=∫y κˉ(dy)\bar m=\int y\,\bar\kappa(dy)mˉ=∫yκˉ(dy). For every t∈Rt\in\mathbb{R}t∈R, z↦et(z−mˉ)z\mapsto e^{t(z-\bar m)}z↦et(z−mˉ) is κˉ\bar\kappaκˉ-integrable and

∫et(z−mˉ) κˉ(dz)≤eσt2.\int e^{t(z-\bar m)}\,\bar\kappa(dz)\le e^{\sigma t^2}.∫et(z−mˉ)κˉ(dz)≤eσt2.

In other words, the mixture, centred at its own mean, is sub-Gaussian with variance proxy 2σ2\sigma2σ, the same constant as the components. No hypothesis constrains the means mxm_xmx​ to agree or to be close.

Degenerate cases.

  • Mass zero. μ(B)=0\mu(B)=0μ(B)=0 is excluded, so BBB is nonempty and the conditioning involves no division by zero.
  • σ=0\sigma=0σ=0. The hypothesis bounds each component's centred moment generating function by 111 for all ttt. The conclusion asserts the same bound 111 for the mixture.
  • B=XB=XB=X. The mixture is taken against μ\muμ itself.
  • One-point XXX. The mixture is the single component κx\kappa_xκx​, and the statement reduces to its own hypothesis.
  • Points outside BBB. Nothing is assumed about κx\kappa_xκx​ for x∉Bx\notin Bx∈/B. Such points carry zero conditional mass.
Human review
  • Endorsed by Community (Bot) · Oct 29, 2026

    Confirmed by the moderator at approval.

  • Endorsed by Shuze Chen · Oct 29, 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