Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Equation (5.2): P_χ = p₁ (Pr_{EX_CN(c,D₁)}[χ = 1] − η)/(1 − 2η) + Pr_{EX_CN(c,D)}[χ = 1 ∧ x ∈ X₂], simulating a statistical query from noisy examples

Proved
ComputationalLearning.noisy_query_decomposition

by naimengye · Sep 24, 2026 · Mathlib 0df444a (Lean v4.33.1)

classification-noisepac-learningstatistical-queries

Equation (5.2) (p. 113). Let X1X_1X1​ be the set of inputs xxx with χ(x,0)≠χ(x,1)\chi(x, 0) \ne \chi(x, 1)χ(x,0)=χ(x,1) and X2X_2X2​ its complement; let p1=Pr⁡x∼D[x∈X1]p_1 = \Pr_{x \sim D}[x \in X_1]p1​=Prx∼D​[x∈X1​] and D1D_1D1​ be DDD restricted to X1X_1X1​. Then

Pχ=p1 Pr⁡EXCNη(c,D1)[χ=1]−η1−2η+Pr⁡EXCNη(c,D)[(χ=1)∧(x∈X2)],P_\chi = p_1\,\frac{\Pr_{EX^\eta_{CN}(c, D_1)}[\chi = 1] - \eta}{1 - 2\eta} + \Pr_{EX^\eta_{CN}(c, D)}[(\chi = 1) \wedge (x \in X_2)],Pχ​=p1​1−2ηPrEXCNη​(c,D1​)​[χ=1]−η​+EXCNη​(c,D)Pr​[(χ=1)∧(x∈X2​)],

obtained from Pχ=p1Pr⁡EX(c,D1)[χ=1]+Pr⁡EX(c,D)[(χ=1)∧(x∈X2)]P_\chi = p_1 \Pr_{EX(c, D_1)}[\chi = 1] + \Pr_{EX(c,D)}[(\chi = 1) \wedge (x \in X_2)]Pχ​=p1​PrEX(c,D1​)​[χ=1]+PrEX(c,D)​[(χ=1)∧(x∈X2​)] (5.1), the replacement of the correct label by a noisy one on X2X_2X2​, and Pr⁡EXCN(c,D1)[χ=1]=η+(1−2η)Pr⁡EX(c,D1)[χ=1]\Pr_{EX_{CN}(c, D_1)}[\chi = 1] = \eta + (1 - 2\eta)\Pr_{EX(c, D_1)}[\chi = 1]PrEXCN​(c,D1​)​[χ=1]=η+(1−2η)PrEX(c,D1​)​[χ=1] on X1X_1X1​. Every quantity on the right can be estimated from the noisy oracle, which is the key to simulating statistical query algorithms in the presence of classification noise (Theorem 5.3).

Formally: for a distribution DDD, measurable ccc and χ\chiχ, and 0≤η<1/20 \le \eta < 1/20≤η<1/2, the identity above with D1D_1D1​ the conditional measure D[⋅∣X1]D[\cdot \mid X_1]D[⋅∣X1​] (the zero measure if D(X1)=0D(X_1) = 0D(X1​)=0, in which case the first term vanishes) and the probabilities taken under the noisy example laws.

Preamble
import Definitions.Def_ComputationalLearning_Noise

open MeasureTheory ProbabilityTheory
Formal statement
namespace ComputationalLearning

/-- **Equation (5.2)** (§5.4.1, p. 113), the decomposition of `P_χ` in terms of quantities
estimable from the noisy oracle. For a statistical query `χ`, let `X₁` be the inputs on which
the label matters, `p₁ = Pr_{x ~ D}[x ∈ X₁]`, and `D₁` the distribution `D` restricted to `X₁`.
Then
`P_χ = p₁ · (Pr_{EX_CN(c, D₁)}[χ = 1] − η)/(1 − 2η) + Pr_{EX_CN(c, D)}[(χ = 1) ∧ (x ∈ X₂)]`,
where the first probability is under the noisy oracle on `D₁` (Equation (5.1) and the identity
`Pr_{EX_CN(c,D₁)}[χ = 1] = η + (1 − 2η) Pr_{EX(c,D₁)}[χ = 1]`) and the second under the noisy
oracle on `D` (on `X₂` a noisy label can replace the correct one). Stated for `0 ≤ η < 1/2`,
measurable `c` and `χ`; if `p₁ = 0` the conditional is the zero measure and the first term
vanishes. -/
theorem noisy_query_decomposition {X : Type*} [MeasurableSpace X] (D : Measure X)
    [IsProbabilityMeasure D] (c : X → Bool) (hc : Measurable c) (χ : X × Bool → Bool)
    (hχ : Measurable χ) {η : ℝ} (hη0 : 0 ≤ η) (hη : η < 1 / 2) :
    queryProb D c χ =
      (D (labelSensitive χ)).toReal *
        ((noisyExampleLaw (D[|labelSensitive χ]) c η {p | χ p = true}).toReal - η) / (1 - 2 * η) +
      (noisyExampleLaw D c η {p | χ p = true ∧ p.1 ∉ labelSensitive χ}).toReal := by sorry

end ComputationalLearning
Source
Kearns and Vazirani, An Introduction to Computational Learning Theory, MIT Press 1994, doi:10.7551/mitpress/3897.001.0001, §5.4.1 pp. 112-113, the decomposition of P_χ, Equations (5.1)-(5.2), the key idea of Theorem 5.3
Human review
  • Endorsed by Shuze Chen · Sep 25, 2026

    Confirmed by the moderator at approval.

  • Endorsed by naimengye · Sep 25, 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