Proof of Theorem 3.3, p. 590 — (16) has an optimal solution whose blocks have v_k0^i > 0 or vanish
ProvedMinimaxSLP.RhsExtremal.dual16_optimal_zero_or_positiveLet and a symmetric satisfy (Assumption 4), let , and fix the data of the objective of (16). Then (16) has an optimal solution , that is, a feasible point at which the objective
attains its maximum over the feasible set, such that for every pair either or .
The paper writes "without loss of generality, we can again assume that for all and (see Theorem 2.2)"; the blocks with are discarded from the mixture, which is the form stated here. This optimal solution is the input to the construction of the extremal distribution.
Formalization Note " for all " literally can fail at every optimum, so the zero-or-positive form is used; the zero blocks receive weight in the mixture. is assumed explicitly (with the program is infeasible); under Assumptions 2, 3 and 5 it holds automatically.
import Mathlib import Definitions.Def_MomentDRO_Conf_Setting import Definitions.Def_MinimaxSLP_RhsExtremal_Model
namespace MinimaxSLP.RhsExtremal
open MeasureTheory Matrix
/-- Proof of Theorem 3.3, p. 590: under Assumption 4 (`Q` symmetric, `Q − μμ′ ≻ 0`) and
`N ≥ 1`, the problem (16) has an optimal solution in which every block either has
`v_k0^i > 0` or is identically zero (the reading of "without loss of generality `v_k0^i > 0`
(see Theorem 2.2)"). -/
theorem dual16_optimal_zero_or_positive {r n K N : ℕ} [NeZero K]
(T : Matrix (Fin r) (Fin n) ℝ) (α β : Fin K → ℝ) (ps : Fin N → Fin r → ℝ)
(μ : Fin r → ℝ) (Q : Matrix (Fin r) (Fin r) ℝ)
(hN : 0 < N)
(hQ : Q.IsSymm) (hA4 : (Q - Matrix.vecMulVec μ μ).PosDef)
(x : Fin n → ℝ) :
∃ s ∈ feasible16 K N μ Q,
IsGreatest (obj16 T α β ps x '' feasible16 K N μ Q) (obj16 T α β ps x s) ∧
∀ k i, 0 < (s k i).2.2 ∨ s k i = 0 := by sorry
end MinimaxSLP.RhsExtremal
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.