§2.2, pp. 584–585 — conic duality: under Assumption 4
ProvedMinimaxSLP.ObjExtremal.dual6_le_dual9_upperConsider the two-stage minimax stochastic linear program with random objective of §2: recourse set , second-stage value , disutility with , mean and second-moment matrix (symmetric). Assume , Assumption 3 ( for all ) and Assumption 4 ().
The dual (6) of the moment problem is
and the semidefinite program (9) is
Statement. If is an upper bound for the objective values of (9), then for every there is a feasible point of (6), with symmetric, such that
In particular .
This is the conic-duality half of the identity in Theorem 2.2. Program (9) is the semidefinite dual of (8), whose value is , and Assumption 4 makes positive definite. That is a Slater point for (9), so there is no duality gap. Together with weak duality and the extremal-distribution construction it gives .
import Mathlib import Definitions.Def_MomentDRO_Conf_Setting import Definitions.Def_MinimaxSLP_ObjExtremal_Model open MeasureTheory Matrix Filter Topology
namespace MinimaxSLP.ObjExtremal
/-- §2.2 (pp. 583–585): conic duality between the dual (6) of the moment problem and the
semidefinite program (9). Under Assumption 4 (`Q − μμ′ ≻ 0`, the Slater condition for (9)), every
upper bound `u` of the objective values of (9) is, up to any `ε > 0`, the objective value of a
feasible point of (6); in particular `Z_D(x) ≤ Z_DD(x)`. -/
theorem dual6_le_dual9_upper {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)
(hQ : Q.IsSymm) (hA4 : (Q - Matrix.vecMulVec μ μ).PosDef)
(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)
(u : ℝ) (hu : u ∈ upperBounds (dualValues9 W T h α β μ Q x)) (ε : ℝ) (hε : 0 < ε) :
∃ s ∈ MinimaxSLP.ObjSDP.dualFeasible6 W T h α β x,
MinimaxSLP.ObjSDP.frob Q s.1 + μ ⬝ᵥ s.2.1 + s.2.2 ≤ u + ε := by sorry
end MinimaxSLP.ObjExtremal