Lower one-sided coverage limit for the optimal value
ProvedGenEmpLik.Coverage.one_sided_limit_lowerconfidence-regionempirical-likelihoodstochastic-optimization
Let be nonempty and compact, let the losses be measurable and satisfy Assumption B with , and suppose for some . Let satisfy Assumption A, let be i.i.d. with law , and suppose the population optimizer of is unique with . For , the lower endpoint of the empirical likelihood interval satisfies
Here the infimum runs over probability weight vectors on the sample points with , and . This is the lower counterpart of the one-sided limit ; together the two give Theorem 3.
Preamble
import Mathlib import Definitions.Def_GenEmpLik_Coverage_AssumptionA import Definitions.Def_GenEmpLik_Coverage_AssumptionB import Definitions.Def_GenEmpLik_Coverage_optimalValue import Definitions.Def_GenEmpLik_Coverage_empiricalOptimalValue import Definitions.Def_GenEmpLik_Coverage_divergenceBall import Definitions.Def_GenEmpLik_Coverage_confidenceSet import Definitions.Def_GenEmpLik_Coverage_influence import Definitions.Def_GenEmpLik_Coverage_remainder import Definitions.Def_GenEmpLik_Coverage_robustUpper import Definitions.Def_GenEmpLik_Coverage_uniqueOptimizer import Definitions.Def_VarianceRegularization_Expansion_empMean import Definitions.Def_VarianceRegularization_Expansion_empVar open MeasureTheory ProbabilityTheory Filter Topology
Formal statement
namespace GenEmpLik.Coverage
/-- The lower one-sided coverage limit in Appendix B.4, p. 37: by the argument symmetric to the
upper one-sided limit, the probability that the lower endpoint inf {T_opt(P) : D_f(P||P̂_n) ≤ ρ/n}
lies below T_opt(P₀) converges to P(N(0,1) ≤ √ρ). -/
theorem one_sided_limit_lower
{Ω Ξ : Type*} [MeasurableSpace Ω] [MeasurableSpace Ξ]
(μ : Measure Ω) [IsProbabilityMeasure μ]
(ξ : ℕ → Ω → Ξ) (d : ℕ)
(X : Set (EuclideanSpace ℝ (Fin d)))
(ℓ : EuclideanSpace ℝ (Fin d) → Ξ → ℝ)
(M : Ξ → ℝ) (f : ℝ → EReal) (ρ : ℝ)
(hρ : 0 ≤ ρ) (hA : AssumptionA f)
(hX : X.Nonempty) (hB : AssumptionB X ℓ M)
(hℓmeas : ∀ x, Measurable (ℓ x))
(hM2 : Integrable (fun z => M z ^ 2) (μ.map (ξ 0)))
(hℓ2 : ∃ x0 ∈ X, Integrable (fun z => ℓ x0 z ^ 2) (μ.map (ξ 0)))
(hξmeas : ∀ i, Measurable (ξ i))
(hξindep : iIndepFun ξ μ)
(hξident : ∀ i, IdentDistrib (ξ i) (ξ 0) μ μ)
(hunique : HasUniqueOptimizer X ℓ (μ.map (ξ 0)))
(hvar : 0 < variance (ℓ (uniqueOptimizer X ℓ (μ.map (ξ 0)) hunique))
(μ.map (ξ 0))) :
Tendsto (fun n : ℕ => μ {ω |
sInf (empiricalOptimalValue X ℓ (fun i : Fin n => ξ i ω) '' divergenceBall f ρ n) ≤
optimalValue X ℓ (μ.map (ξ 0))})
atTop (𝓝 (gaussianReal 0 1 (Set.Iic (Real.sqrt ρ)))) := by sorry
end GenEmpLik.CoverageSource
Duchi, Glynn & Namkoong, Statistics of Robust Optimization: A Generalized Empirical Likelihood Approach, arXiv:1610.03425v3, Appendix B.4 (proof of Theorem 10), p. 37: "By an exactly symmetric argument on -T(P0), we similarly have P(T(P0) >= inf_{P: D_f(P||P_n) <= rho/n} T(P)) -> P(N(0,1) <= sqrt rho)"; specialized to T_opt via Appendix C.1 (proof of Theorem 3).