Proposition 2 (a) and λ-part of (b), (12c)–(12d), p. 16 — one-sided λ-derivatives of ℓ_rob are min/max of −√δ(γ²βᵀA(x)⁻¹β − 1) over Γ*
ProvedOTDRO.WorstCase.ellRob_lambda_derivativesAssume Assumption 1, , convex, and let satisfy Assumption 2 and be of the form for a positive integer and continuously differentiable . Let and, for ,
Then for -almost every , for every :
- (a) ;
- is finite near and absolutely continuous on every interval with endpoints in ;
- (12c), (12d): its left and right derivatives are
both minimum and maximum being attained; 4. is differentiable at if and only if is a singleton.
This is the envelope-theorem description of the robust loss in the multiplier ; integrated over it yields the one-sided derivatives of the dual objective used to locate the dual optimizer.
Formalization Note Only part (a) and the -part of part (b) are formalized; the -derivatives (12a), (12b), the absolute continuity in , the -differentiability criterion and (13) are omitted. Because is affine in , both one-sided partial derivatives equal , so the page's set is the image written here. One-sided derivatives are HasDerivWithinAt on and of the real-valued map , which is finite near (stated as a conjunct). "Absolutely continuous for " is read as absolute continuity on every compact interval inside . The "-almost every " quantifier is placed before , as on the page: one null set serves all . is written as: is the greatest element of .
import Mathlib import Definitions.Def_OTDRO_Dual_Setting open MeasureTheory Filter Topology
namespace OTDRO.WorstCase
/-- **Proposition 2**, part (a) and the `λ`-part of (b) (arXiv:1810.02403v3, p. 16; proof
pp. 48–49): for `ℓ = max_{i ≤ K} ℓᵢ` with `ℓᵢ` continuously differentiable and `ℓ` satisfying
Assumption 2, for `P₀`-almost every `x` and every `(β, λ) ∈ 𝕌₁`: `Γ*(β, λ; x) ≠ ∅`;
`λ ↦ ℓ_rob(β, λ; x)` is finite near `λ` and absolutely continuous on every interval inside
`(λ_thr(β), ∞)`; its left and right derivatives are the minimum (12c) and the maximum (12d) of
`−√δ(γ² βᵀA(x)⁻¹β − 1)` over `γ ∈ Γ*(β, λ; x)`; and it is differentiable at `λ` if and only if
the set `{∂±F/∂λ(γ, β, λ; x) : γ ∈ Γ*(β, λ; x)} = {−√δ(γ²βᵀA(x)⁻¹β − 1) : γ ∈ Γ*}` is a
singleton. -/
theorem ellRob_lambda_derivatives {d : ℕ} (P0 : Measure (EuclideanSpace ℝ (Fin d)))
[IsProbabilityMeasure P0] (A : EuclideanSpace ℝ (Fin d) → Matrix (Fin d) (Fin d) ℝ)
(ρmin ρmax : ℝ) (hA : OTDRO.Dual.Assumption1 P0 A ρmin ρmax) (ℓ : ℝ → ℝ) (h2 : OTDRO.Dual.Assumption2 P0 ℓ)
(δ : ℝ) (hδ : 0 < δ) (B : Set (EuclideanSpace ℝ (Fin d))) (hB : Convex ℝ B)
(K : ℕ) (hK : 0 < K) (ℓi : Fin K → ℝ → ℝ) (hℓi : ∀ i, ContDiff ℝ 1 (ℓi i))
(hmax : ∀ u, IsGreatest (Set.range fun i => ℓi i u) (ℓ u)) :
∀ᵐ x ∂P0, ∀ β ∈ B, ∀ lam : ℝ, OTDRO.Dual.lamThr P0 ℓ A δ β < lam →
(OTDRO.Dual.maximizers ℓ A δ β lam x).Nonempty ∧
(∀ᶠ l in 𝓝 lam, OTDRO.Dual.ellRob ℓ A δ β l x ≠ ⊤) ∧
(∀ a b : ℝ, OTDRO.Dual.lamThr P0 ℓ A δ β < a → OTDRO.Dual.lamThr P0 ℓ A δ β < b →
AbsolutelyContinuousOnInterval (fun l => (OTDRO.Dual.ellRob ℓ A δ β l x).toReal) a b) ∧
(∃ m : ℝ, IsLeast ((fun γ => -Real.sqrt δ * (γ ^ 2 * OTDRO.Dual.quadInv A β x - 1)) ''
OTDRO.Dual.maximizers ℓ A δ β lam x) m ∧
HasDerivWithinAt (fun l => (OTDRO.Dual.ellRob ℓ A δ β l x).toReal) m (Set.Iic lam) lam) ∧
(∃ M : ℝ, IsGreatest ((fun γ => -Real.sqrt δ * (γ ^ 2 * OTDRO.Dual.quadInv A β x - 1)) ''
OTDRO.Dual.maximizers ℓ A δ β lam x) M ∧
HasDerivWithinAt (fun l => (OTDRO.Dual.ellRob ℓ A δ β l x).toReal) M (Set.Ici lam) lam) ∧
(DifferentiableAt ℝ (fun l => (OTDRO.Dual.ellRob ℓ A δ β l x).toReal) lam ↔
∃ s : ℝ, (fun γ => -Real.sqrt δ * (γ ^ 2 * OTDRO.Dual.quadInv A β x - 1)) ''
OTDRO.Dual.maximizers ℓ A δ β lam x = {s}) := by sorry
end OTDRO.WorstCase
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.