Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Lemma 3.4 — population grid rounding rule (for the greedy partition of its proof)

Proved
CalibratedMA.Randomized.lemma_3_4

by mikedeng1 · Nov 10, 2026 · Mathlib 0df444a (Lean v4.33.1)

calibrationp2o-batch-stoc27ap2o-gran-per-chapterp2o-plan-paperp2o-strengthenedp2o-v1randomized-rounding

Let ν\nuν be any distribution on [0,1][0,1][0,1], let m∈Rm\in\mathbb Rm∈R with m≥2m\ge2m≥2, and let ρ>0\rho>0ρ>0 with 1/ρ1/\rho1/ρ an integer and ρ≤m−2\rho\le m^{-2}ρ≤m−2. Let Γ\GammaΓ be the greedy partition of ν\nuν with threshold τ=1/(64m2)\tau=1/(64m^2)τ=1/(64m2) on the grid Λρ\Lambda_\rhoΛρ​, the construction in the proof of Lemma 3.4 that Algorithm 1 uses in step 4. Then the partition points of Γ\GammaΓ belong to Λρ={0,ρ,2ρ,…,1}\Lambda_\rho=\{0,\rho,2\rho,\dots,1\}Λρ​={0,ρ,2ρ,…,1}, Γ\GammaΓ has at most 4m+24m+24m+2 intervals, and, for a universal constant CCC,

RoundErr(ν,Γ)≤Cm2.\mathrm{RoundErr}(\nu,\Gamma)\le\frac{C}{m^2}.RoundErr(ν,Γ)≤m2C​.

The point of the m−2m^{-2}m−2 rate is that rounding to about mmm values costs only O(m−2)O(m^{-2})O(m−2) in calibration, rather than the O(m−1)O(m^{-1})O(m−1) obtained by summing interval contributions separately. Theorem 4.1 applies this lemma to the empirical score distribution ν^\hat\nuν^.

Formalization Note. ρ=1/K\rho=1/Kρ=1/K with K≥1K\ge1K≥1 a natural number, so ρ≤m−2\rho\le m^{-2}ρ≤m−2 reads 1/K≤(m2)−11/K\le(m^2)^{-1}1/K≤(m2)−1. The conclusion includes that the greedy output is a grid partition (IsGridPartition). CCC is quantified before ν\nuν, mmm and KKK.

Differences from the source:

  • generalization: "There is a partition Γ\GammaΓ" → the conclusion holds for the explicit greedy partition of the proof; reason: this is the partition Algorithm 1 uses, and the existence statement follows by taking it.
  • encoding: ρ>0\rho>0ρ>0 with 1/ρ1/\rho1/ρ an integer → K∈NK\in\mathbb NK∈N, K≥1K\ge1K≥1, ρ=1/K\rho=1/Kρ=1/K; reason: same instances.
  • encoding: a distribution on [0,1] → a probability measure on ℝ assigning zero mass to the complement of [0,1]; reason: extension by zero and restriction identify the same distributions.
Preamble
import Mathlib
import Definitions.Def_CalibratedMA_Randomized_Rounding
import Definitions.Def_CalibratedMA_Randomized_Algorithm1

open MeasureTheory
Formal statement
namespace CalibratedMA.Randomized

/-- Lemma 3.4 (p. 16), population grid rounding rule, for the greedy partition constructed in
its proof: for every distribution `ν` on `[0,1]`, every `m ≥ 2` and every grid `Λ_ρ` with
`1/ρ = K` an integer and `ρ ≤ m^{-2}`, the greedy partition is a grid partition of `Λ_ρ` with at
most `4m + 2` intervals and `RoundErr(ν, Γ) ≤ C/m²`, for a universal constant `C`. -/
theorem lemma_3_4 :
    ∃ C : ℝ, 0 < C ∧
      ∀ (ν : Measure ℝ), IsProbabilityMeasure ν → ν (Set.Icc (0 : ℝ) 1)ᶜ = 0 →
      ∀ (m : ℝ) (K : ℕ), 2 ≤ m → 1 ≤ K → (1 : ℝ) / K ≤ (m ^ 2)⁻¹ →
        IsGridPartition K (greedyPartition ν m K).1 (greedyPartition ν m K).2 ∧
        ((greedyPartition ν m K).1 : ℝ) ≤ 4 * m + 2 ∧
        RoundErr ν (greedyPartition ν m K).1 (greedyPartition ν m K).2 ≤ C / m ^ 2 := by sorry

end CalibratedMA.Randomized
Source
Jiuyao Lu, Georgy Noarov, Aaron Roth, Calibrated Multiaccuracy: Deterministic and Randomized Sample Complexity, STOC 2027 submission #900 (October 2026), p. 16, Lemma 3.4 (greedy construction in its proof, pp. 16–17)
Human review
  • Endorsed by Shuze Chen · Nov 10, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Nov 10, 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