Lemma 2, p. 16 — (β, λ) ↦ ℓ_rob(β, λ; x) is convex on B × ℝ₊ for every x
ProvedOTDRO.Cvx.ellRob_convexconvex-analysisdistributionally-robust-optimizationoptimal-transportp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Let take positive definite values with lower semicontinuous cost (Assumption 1a), let be convex, , and let be convex. Then for every the map is convex on : for and ,
Pointwise convexity of the robust loss is the step from which convexity of the dual objective follows by integration.
Formalization Note takes values in , so convexity is written as the displayed inequality in EReal rather than with ConvexOn; with Mathlib's convention the inequality reduces to an equality at . The paper assumes "Assumptions 1a and 2"; only the convexity of from Assumption 2 is kept, and Assumption 1(b) is not assumed, which makes the statement stronger. No probability measure appears.
Preamble
import Mathlib import Definitions.Def_OTDRO_Dual_Setting
Formal statement
namespace OTDRO.Cvx
open Matrix
/-- **Lemma 2**, p. 16: under Assumption 1a (`A(x)` positive definite for every `x`, the cost
lower semicontinuous) and the convexity of `ℓ` from Assumption 2, the map
`(β, λ) ↦ ℓ_rob(β, λ; x)` is convex on `B × ℝ₊` for every `x ∈ ℝ^d`. Convexity of the
`EReal`-valued map is written out: for `θᵢ = (βᵢ, λᵢ) ∈ B × ℝ₊` and `α ∈ [0, 1]`,
`ℓ_rob(αθ₁ + (1 − α)θ₂; x) ≤ α ℓ_rob(θ₁; x) + (1 − α) ℓ_rob(θ₂; x)`. -/
theorem ellRob_convex {d : ℕ}
(A : EuclideanSpace ℝ (Fin d) → Matrix (Fin d) (Fin d) ℝ)
(hPD : ∀ x, (A x).PosDef)
(hlsc : LowerSemicontinuous (fun p : EuclideanSpace ℝ (Fin d) × EuclideanSpace ℝ (Fin d) =>
OTDRO.Dual.mahalCost A p.1 p.2))
(ℓ : ℝ → ℝ) (hℓ : ConvexOn ℝ Set.univ ℓ) (δ : ℝ) (hδ : 0 < δ)
(B : Set (EuclideanSpace ℝ (Fin d))) (hB : Convex ℝ B) :
∀ x : EuclideanSpace ℝ (Fin d),
∀ β₁ ∈ B, ∀ β₂ ∈ B, ∀ lam₁ lam₂ : ℝ, 0 ≤ lam₁ → 0 ≤ lam₂ →
∀ α : ℝ, 0 ≤ α → α ≤ 1 →
OTDRO.Dual.ellRob ℓ A δ (α • β₁ + (1 - α) • β₂) (α * lam₁ + (1 - α) * lam₂) x ≤
(α : EReal) * OTDRO.Dual.ellRob ℓ A δ β₁ lam₁ x +
((1 - α : ℝ) : EReal) * OTDRO.Dual.ellRob ℓ A δ β₂ lam₂ x := by sorry
end OTDRO.Cvx
Source
arXiv:1810.02403v3, Lemma 2, p. 16 (proof in §5.1, p. 32)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.