Theorem 3, p. 12 — Z_n(x) → Z0(x) and Z̄_n(x) → Z0(x) almost surely
ProvedEmpiricalDRO.Consistency.theorem3_consistencyLet be independent and identically distributed random elements of a measurable space , with law . Fix a decision and a loss , measurable, such that
Let and let be the -quantile of the distribution. For each let
where is the empirical Burg-entropy divergence ball (19). Then, almost surely,
Theorem 3 complements the coverage result Theorem 2: the empirical DRO bounds, calibrated with the quantile, not only bracket with asymptotic probability but also shrink onto it.
Formalization Note The sample is ξ 0, …, ξ (n-1). The moment conditions are MemLp (h x ∘ ξ 0) 2 (finite mean and variance) and positive variance; the quantile is pinned by . The paper's only says where lives and is dropped. The positive variance and the value of are kept as printed ("the same conditions in Theorem 2") although the conclusion does not need them. is the published robustMean burg (q/2) (ball radius with ), and is robustLower burg (q/2).
import Mathlib import Definitions.Def_EmpiricalDRO_Consistency_Setting open MeasureTheory ProbabilityTheory Filter Topology
namespace EmpiricalDRO.Consistency
/-- Theorem 3, Lam, arXiv:1605.09349v1, p. 12: under the conditions of Theorem 2 (i.i.d. data,
`0 < Var₀(h(x;ξ)) < ∞`, `Z₀(x) = E₀[h(x;ξ)]` finite), for a fixed decision `x`, both empirical
DRO values `Z_n(x)` (min) and `Z̄_n(x)` (max) over the empirical Burg ball `U_n(χ²_{1,1−α}/(2n))`
converge almost surely to `Z₀(x)`. The sample `ξ_1, …, ξ_n` is `ξ 0, …, ξ (n-1)`, and `q` is the
`(1−α)`-quantile of `χ²₁` (the law of `Y²`, `Y ∼ N(0,1)`). -/
theorem theorem3_consistency {Ω : Type*} [MeasurableSpace Ω] {P : Measure Ω}
[IsProbabilityMeasure P] {Ξ : Type*} [MeasurableSpace Ξ]
(ξ : ℕ → Ω → Ξ) (hξ : ∀ i, Measurable (ξ i)) (hindep : iIndepFun ξ P)
(hident : ∀ i, IdentDistrib (ξ i) (ξ 0) P P)
{m : ℕ} (h : EuclideanSpace ℝ (Fin m) → Ξ → ℝ) (x : EuclideanSpace ℝ (Fin m))
(hhx : Measurable (h x))
(hL2 : MemLp (fun ω => h x (ξ 0 ω)) 2 P)
(hvar : 0 < variance (fun ω => h x (ξ 0 ω)) P)
(α : ℝ) (hα0 : 0 < α) (hα1 : α < 1) (q : ℝ)
(hq : gaussianReal 0 1 {y : ℝ | y ^ 2 ≤ q} = ENNReal.ofReal (1 - α)) :
∀ᵐ ω ∂P,
Tendsto (fun n : ℕ => EmpiricalDRO.Coverage.robustLower EmpiricalDRO.Coverage.burg (q / 2) (fun i : Fin n => h x (ξ i ω))) atTop
(𝓝 (∫ ω', h x (ξ 0 ω') ∂P)) ∧
Tendsto (fun n : ℕ =>
GenEmpLik.Expansion.robustMean EmpiricalDRO.Coverage.burg (q / 2) (fun i : Fin n => h x (ξ i ω))) atTop
(𝓝 (∫ ω', h x (ξ 0 ω') ∂P)) := by sorry
end EmpiricalDRO.Consistency
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.