Proof of Theorem 2.2, pp. 586–587 — distributions in approach the value of any feasible point of (9)
ProvedMinimaxSLP.ObjExtremal.feasible9_approxConsider the two-stage minimax stochastic linear program with random objective of §2: recourse set , second-stage value , piecewise-linear disutility with , and moment class of probability measures on (with finite second moments) having mean and second-moment matrix . Assume and Assumption 3: for every .
Let be any feasible solution of the semidefinite program (9), i.e.
Then for every there is a distribution such that is -integrable and
This is the construction step of the proof of Theorem 2.2: the paper builds, from an optimal solution of (9), distributions that put mass near the conditional mean with a rare large perturbation carrying the remaining second moment, and uses homogeneity and superadditivity of together with to bound the expected disutility from below by the objective of (9) up to an error that vanishes as . It gives . The statement here is for an arbitrary feasible point of (9) (no optimality or attainment is assumed).
import Mathlib import Definitions.Def_MomentDRO_Conf_Setting import Definitions.Def_MinimaxSLP_ObjExtremal_Model open MeasureTheory Matrix Filter Topology
namespace MinimaxSLP.ObjExtremal
/-- Proof of Theorem 2.2 (pp. 586–587), the construction step: every feasible solution
`(V_k, v_k, v_k0, p_k)_k` of (9) is approached from below by distributions of the moment class
`𝒫`: for every `δ > 0` there is `P ∈ 𝒫` with `𝕌(𝒬(·, x))` integrable and
`E_P[𝕌(𝒬(q̃, x))] ≥ Σ_k ((h − T x)′p_k + β_k v_k0) − δ`. (The paper applies it to an optimal
solution of (9); it yields `Z(x) ≥ Z_DD(x)`.) -/
theorem feasible9_approx {n r d K : ℕ} [NeZero K]
(W : Matrix (Fin r) (Fin d) ℝ) (T : Matrix (Fin r) (Fin n) ℝ) (h : Fin r → ℝ)
(α β : Fin K → ℝ) (hα : ∀ k, 0 ≤ α k)
(μ : Fin d → ℝ) (Q : Matrix (Fin d) (Fin d) ℝ)
(hA3 : ∀ q : Fin d → ℝ, ∃ π : Fin r → ℝ, Wᵀ *ᵥ π ≤ q)
(x : Fin n → ℝ) (hrec : (MinimaxSLP.ObjSDP.recourseSet W T h x).Nonempty)
(s : (Fin K → Matrix (Fin d) (Fin d) ℝ) × (Fin K → Fin d → ℝ) × (Fin K → ℝ) ×
(Fin K → Fin r → ℝ))
(hs : s ∈ feasible9 W α μ Q) (δ : ℝ) (hδ : 0 < δ) :
∃ P ∈ MinimaxSLP.ObjSDP.momentClass μ Q,
Integrable (fun q => MinimaxSLP.ObjSDP.disutility α β (MinimaxSLP.ObjSDP.Qval W T h q x)) P ∧
obj9 T h β x s - δ ≤
∫ q, MinimaxSLP.ObjSDP.disutility α β (MinimaxSLP.ObjSDP.Qval W T h q x) ∂P := by sorry
end MinimaxSLP.ObjExtremal