Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Lemma 5.6 — A_g approximates E[g] within ±τ with probability ≥ 1 − β from n ≥ c·log(1/β)b²/(ε²τ²) samples

Proved
PrivLearn.LocalSim.lemma_5_6

by mikedeng1 · Oct 7, 2026 · Mathlib 0df444a (Lean v4.33.1)

concentrationlocal-privacyp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1statistical-queries

There is an absolute constant c>0c>0c>0 with the following property. Let PPP be a probability distribution on a domain DDD, let g:D→[−b,b]g:D\to[-b,b]g:D→[−b,b] be measurable, and let b,ε,τ>0b,\varepsilon,\tau>0b,ε,τ>0, 0<β<10<\beta<10<β<1 with ε≤1\varepsilon\le1ε≤1 and ετ≤4b\varepsilon\tau\le4bετ≤4b. If the database z=(z1,…,zn)z=(z_1,\dots,z_n)z=(z1​,…,zn​) has

n ≥ c⋅ln⁡(1/β) b2ε2τ2n\ \ge\ c\cdot\frac{\ln(1/\beta)\,b^2}{\varepsilon^2\tau^2}n ≥ c⋅ε2τ2ln(1/β)b2​

entries drawn i.i.d. from PPP, then the local algorithm Ag(n,ε,LRz)\mathcal A_g(n,\varepsilon,LR_z)Ag​(n,ε,LRz​) approximates Eu∼P[g(u)]\mathbb E_{u\sim P}[g(u)]Eu∼P​[g(u)] within additive error ±τ\pm\tau±τ with probability at least 1−β1-\beta1−β:

Pr⁡z∼Pn, η∼Lap(2b/ε)n[∣1n∑i=1n(g(zi)+ηi)−Eu∼P[g(u)]∣≤τ] ≥ 1−β.\Pr_{z\sim P^n,\ \eta\sim\mathrm{Lap}(2b/\varepsilon)^n}\Bigl[\Bigl|\frac1n\sum_{i=1}^n\bigl(g(z_i)+\eta_i\bigr)-\mathbb E_{u\sim P}[g(u)]\Bigr|\le\tau\Bigr]\ \ge\ 1-\beta.z∼Pn, η∼Lap(2b/ε)nPr​[​n1​i=1∑n​(g(zi​)+ηi​)−Eu∼P​[g(u)]​≤τ] ≥ 1−β.

One run of Ag\mathcal A_gAg​ thus answers one statistical query; Theorem 5.7 applies it once per query on fresh data.

Formalization Note. "For sufficiently large constant ccc" is stated as ∃c>0\exists c>0∃c>0 before every other quantifier. The paper's log⁡\loglog is base 2; we use the natural logarithm, which only changes ccc. The probability is over the database and the Laplace noises jointly (product measure). Two hypotheses are added because the printed lemma is false without them. (i) ετ≤4b\varepsilon\tau\le4bετ≤4b: with g≡bg\equiv bg≡b and K=ετ/bK=\varepsilon\tau/bK=ετ/b large, n=1n=1n=1 is allowed for β=e−K2/c\beta=e^{-K^2/c}β=e−K2/c and the error probability e−K/2e^{-K/2}e−K/2 exceeds β\betaβ. (ii) ε≤1\varepsilon\le1ε≤1: for large ε\varepsilonε the threshold cln⁡(1/β)b2/(ε2τ2)c\ln(1/\beta)b^2/(\varepsilon^2\tau^2)cln(1/β)b2/(ε2τ2) falls below the Θ(b2ln⁡(1/β)/τ2)\Theta(b^2\ln(1/\beta)/\tau^2)Θ(b2ln(1/β)/τ2) samples needed to control the sampling error of ggg (e.g. g=±bg=\pm bg=±b fair coin, τ≪b\tau\ll bτ≪b, ε=4b/τ\varepsilon=4b/\tauε=4b/τ); the paper's remark that the sampling bound "is smaller than the lower bound on nnn in the lemma by a factor of O(ε−2)O(\varepsilon^{-2})O(ε−2)" presumes ε=O(1)\varepsilon=O(1)ε=O(1). Every use in the paper satisfies both.

Preamble
import Mathlib
import Definitions.Def_PrivLearn_LocalSim_Privacy
import Definitions.Def_PrivLearn_LocalSim_Simulation
Formal statement
namespace PrivLearn.LocalSim

open MeasureTheory

/-- Lemma 5.6 (p. 20): there is an absolute constant `c > 0` such that if the database
`z = (z_1, …, z_n)` has `n ≥ c · ln(1/β) b² / (ε² τ²)` entries drawn i.i.d. from `P`, then for a
measurable query `g : Dom → [−b, b]` the output of `A_g(n, ε, LR_z)` is within `τ` of
`E_{u∼P}[g(u)]` with probability at least `1 − β` (over the database and the Laplace noises).
Stated for `ε ≤ 1` and in the regime `ε τ ≤ 4 b`, without which it is false. -/
theorem lemma_5_6 :
    ∃ c : ℝ, 0 < c ∧
      ∀ (Dom : Type) [MeasurableSpace Dom] (P : Measure Dom) [IsProbabilityMeasure P]
        (g : Dom → ℝ) (b ε τ β : ℝ) (n : ℕ),
        Measurable g → (∀ u, |g u| ≤ b) →
        0 < b → 0 < ε → ε ≤ 1 → 0 < τ → ε * τ ≤ 4 * b → 0 < β → β < 1 →
        c * Real.log (1 / β) * b ^ 2 / (ε ^ 2 * τ ^ 2) ≤ n →
        ENNReal.ofReal (1 - β) ≤
          ((Measure.pi fun _ : Fin n => P).prod
              (Measure.pi fun _ : Fin n => PrivLearn.Generic.laplace (2 * b / ε)))
            {p | |avgResp g p.1 p.2 - ∫ x, g x ∂P| ≤ τ} := by sorry

end PrivLearn.LocalSim
Source
Kasiviswanathan, Lee, Nissim, Raskhodnikova and Smith, What Can We Learn Privately?, arXiv:0803.0924v3, p. 20, Lemma 5.6
Human review
  • Endorsed by Shuze Chen · Oct 7, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · 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