Theorem 2 — for a compact convex set of discrete distributions,
ProvedWorstCaseCVaR.Discrete.theorem_2Let be a loss, let be the scenarios of a discrete random vector, and let be a confidence level. For a probability vector put
and for a set of probability vectors let .
Theorem. Suppose that is a nonempty compact convex set of probability vectors. Then for each ,
Both extrema on the right-hand side are attained: for every the maximum over exists, and some minimizes .
The theorem turns the worst-case CVaR — a supremum over distributions of a minimum over thresholds — into a minimum over the threshold of a worst case of a function that is affine in . Minimizing over decisions therefore becomes the single minimization problem (16)–(20) in , which is a linear program under box uncertainty and a second-order cone program under ellipsoidal uncertainty.
Formalization Note The paper's hypothesis " is a compact convex set" is completed by the standing assumptions of §2.2: its elements are probability vectors (, , i.e. stdSimplex ℝ (Fin S)) and it is nonempty (a supremum over the empty set is meaningless). The confidence level satisfies . The left-hand side stays a supremum, as printed. The inner maximum is stated as IsMaxOn and its value is used as the real sSup of the image of ; the outer minimum is an explicit minimizer α₀ with IsMinOn … Set.univ.
import Mathlib import Definitions.Def_WorstCaseCVaR_Discrete_Setting
namespace WorstCaseCVaR.Discrete
/-- Theorem 2, Zhu & Fukushima (2009), p. 1159: if `𝒫_π` is a compact convex set (of probability
vectors on the scenarios `y_[1], …, y_[S]`, nonempty), then for each `x`
`WCVaR_β(x) = min_{α ∈ ℝ} max_{π ∈ 𝒫_π} G_β(x, α, π)`.
Both extrema on the right are attained: for every `α` the maximum over `𝒫_π` is attained, and
some `α₀` minimizes `α ↦ max_{π ∈ 𝒫_π} G_β(x, α, π)` over `ℝ`, with value `WCVaR_β(x)`. -/
theorem theorem_2 {n m S : ℕ} (f : (Fin n → ℝ) → (Fin m → ℝ) → ℝ)
(ys : Fin S → Fin m → ℝ) (x : Fin n → ℝ) (β : ℝ) (hβ0 : 0 < β) (hβ1 : β < 1)
(P : Set (Fin S → ℝ)) (hP : P ⊆ stdSimplex ℝ (Fin S)) (hne : P.Nonempty)
(hcpt : IsCompact P) (hcvx : Convex ℝ P) :
(∀ α : ℝ, ∃ π₀ ∈ P, IsMaxOn (fun π => G f ys β x α π) P π₀) ∧
∃ α₀ : ℝ, IsMinOn (fun α : ℝ => sSup ((fun π => G f ys β x α π) '' P)) Set.univ α₀ ∧
wcvar f ys β x P = sSup ((fun π => G f ys β x α₀ π) '' P) := by sorry
end WorstCaseCVaR.Discrete
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.