(25), p. 32 — sup_Δ {ℓ(βᵀ(x+Δ)) − λ'(ΔᵀA(x)Δ − δ)} = ℓ_rob(β, λ; x) with λ' = λ/√δ
ProvedOTDRO.Cvx.sup_transport_eq_ellRobLet be such that is positive definite, let be any function, , and . Then
as extended real numbers.
This is display (25): it rewrites the inner supremum of the transport dual over the perturbation as a one-dimensional supremum over . The proof of Lemma 2 (convexity of ) starts from the left-hand side, which is visibly a supremum of functions convex in when is convex.
Formalization Note On the page the multiplier on the left of (25) is the dual variable before the change of variables "from to "; here it is written explicitly as , so that one symbol is not used for two different multipliers. The identity holds for every , including , where both sides equal . The same identity is also drafted in mission 1 of this series (OTDRO.Dual.sup_transport_eq_ellRob), because draft items of parallel missions cannot import each other.
import Mathlib import Definitions.Def_OTDRO_Dual_Setting
namespace OTDRO.Cvx
open Matrix
/-- Display (25), p. 32: for positive definite `A(x)`, `δ > 0` and `λ ≥ 0`,
`sup_Δ {ℓ(βᵀ(x + Δ)) − λ'(ΔᵀA(x)Δ − δ)} = sup_γ F(γ, β, λ; x) = ℓ_rob(β, λ; x)`, where the
transport multiplier on the left is the *old* dual variable `λ' = λ / √δ` (the page's change of
variables "from `λ√δ` to `λ`"). Valid for every `β`, including `β = 0`. -/
theorem sup_transport_eq_ellRob {d : ℕ}
(A : EuclideanSpace ℝ (Fin d) → Matrix (Fin d) (Fin d) ℝ)
(x : EuclideanSpace ℝ (Fin d)) (hQ : (A x).PosDef)
(ℓ : ℝ → ℝ) (δ : ℝ) (hδ : 0 < δ)
(β : EuclideanSpace ℝ (Fin d)) (lam : ℝ) (hlam : 0 ≤ lam) :
(⨆ Δ : EuclideanSpace ℝ (Fin d),
((ℓ (inner ℝ β (x + Δ)) - (lam / Real.sqrt δ) *
(dotProduct Δ.ofLp (A x *ᵥ Δ.ofLp) - δ) : ℝ) : EReal)) =
OTDRO.Dual.ellRob ℓ A δ β lam x := by sorry
end OTDRO.Cvx
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.