Lemma 2.6: and the same bound under
ProvedPrivBayesTest.lemma_2_6differential-privacyerror-exponentshypothesis-testinginformation-theoryp2o-gran-per-chapterp2o-plan-paperp2o-strengthenedp2o-v1
Let be probability measures on , , , and let be bounded and measurable with . Let and with independent of . For every and ,
These moment bounds control both error probabilities of the Laplace test and its Chernoff information (Corollary 2.7, Lemma 3.3).
Differences from the source:
- generalization: ℓ̄ the clipped log-likelihood ratio of the dual optimizer at α → any bounded measurable ℓ̄ : X → ℝ with ∫ e^{(α−1)ℓ̄} dP = ∫ e^{αℓ̄} dQ; reason: the proof uses only the balancing condition and the Laplace scale; the paper's ℓ̄ is bounded (clipped) and balanced by Proposition 2.5 item 4
- encoding: E_Q[e^{γαM_Lap(X)}] → lower integral of e^{γαy} against the law lapLaw ℓ̄ (1/(1 − α̃)) n Q of M_Lap(X); reason: same expectation of a nonnegative variable
Preamble
import Mathlib import Definitions.Def_PrivBayesTest_klTvValue import Definitions.Def_PrivBayesTest_lapLaw open MeasureTheory
Formal statement
namespace PrivBayesTest
/-- **Lemma 2.6** of Jacobsen, *Optimal Error Exponents for Private Bayesian Hypothesis Testing*, STOC 2027 submission (manuscript sha256 abff2fdd…124ef), p.12. Let `V = −log E_Q[e^{αℓ̄}] = −log E_P[e^{(α−1)ℓ̄}]`. For every `n ≥ 1` and
every `γ ∈ (0,1)`,
`E_Q[e^{γα M_Lap(X)}] ≤ e^{−γnV}/(1 − γ²)` and `E_P[e^{−γ(1−α) M_Lap(X)}] ≤ e^{−γnV}/(1 − γ²)`,
where `M_Lap(X) = Σᵢ ℓ̄(Xᵢ) + Z`, `Z ∼ Lap(1/(1 − α̃))`.
**Formalization Note** Stated for every bounded measurable `ℓ̄ : X → ℝ` satisfying the balancing
condition `∫ e^{(α−1)ℓ̄} dP = ∫ e^{αℓ̄} dQ` (this contains the paper's clipped log-likelihood ratio,
Proposition 2.5). The integrands of the balancing condition and of `V` are bounded and measurable, so
the Bochner integrals are the true expectations. The law of `M_Lap(X)` under `X ∼ Qⁿ` is
`lapLaw ℓ̄ (1/(1 − α̃)) n Q`; the left-hand expectations are lower integrals in `ENNReal`. -/
theorem lemma_2_6 {X : Type*} [MeasurableSpace X] (P Q : Measure X) [IsProbabilityMeasure P]
[IsProbabilityMeasure Q] (α : ℝ) (hα : α ∈ Set.Ioo (0 : ℝ) 1) (ℓbar : X → ℝ)
(hℓm : Measurable ℓbar) (hℓb : ∃ C : ℝ, ∀ x, |ℓbar x| ≤ C)
(hbal : ∫ x, Real.exp ((α - 1) * ℓbar x) ∂P = ∫ x, Real.exp (α * ℓbar x) ∂Q)
(n : ℕ) (hn : 1 ≤ n) (γ : ℝ) (hγ : γ ∈ Set.Ioo (0 : ℝ) 1) :
∫⁻ y, ENNReal.ofReal (Real.exp (γ * α * y)) ∂(lapLaw ℓbar (1 / (1 - alphaTilde α)) n Q) ≤
ENNReal.ofReal (Real.exp (-γ * n * -Real.log (∫ x, Real.exp (α * ℓbar x) ∂Q)) / (1 - γ ^ 2)) ∧
∫⁻ y, ENNReal.ofReal (Real.exp (-γ * (1 - α) * y)) ∂(lapLaw ℓbar (1 / (1 - alphaTilde α)) n P) ≤
ENNReal.ofReal (Real.exp (-γ * n * -Real.log (∫ x, Real.exp (α * ℓbar x) ∂Q)) / (1 - γ ^ 2)) := by sorry
end PrivBayesTest
Source
Jacobsen, Optimal Error Exponents for Private Bayesian Hypothesis Testing, STOC 2027 submission (manuscript sha256 abff2fdd66827f43…124ef), Lemma 2.6, p.12
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.