§3.1, pp. 436–437 — at X̄ ∈ int(dom ρ) the directional derivative ρ′(X̄,·) is finite, positively homogeneous, convex, and ρ(X) ≥ ρ(X̄) + ρ′(X̄, X − X̄)
ProvedConvexRiskFn.Cont.directional_derivative_propsLet be a real normed space and a proper function satisfying (A1) (convexity). Let be an interior point of . Then for every the one-sided directional derivative
exists as a real number, and the function satisfies:
- is positively homogeneous: for all ;
- is convex on ;
- for every .
This is the construction from which the paper obtains an algebraic subgradient at interior points of the domain.
Formalization Note The statement asserts the existence of a real-valued to which the difference quotients converge as , so is finite valued and is the directional derivative. The quotient is written with EReal.toReal; since is interior to the domain and is proper, is finite for all small , so this agrees with the paper's quotient along the limit. The paper also states convexity of and the inequality for an arbitrary (where may take infinite values); only the case , the one the paper's argument uses, is formalized.
import Mathlib import Definitions.Def_ConvexRiskFn_Cont_Setting open Filter Topology
namespace ConvexRiskFn.Cont
theorem directional_derivative_props {E : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E]
(ρ : E → EReal) (hρ : IsProper ρ) (h1 : ConvexRiskFn.Dual.A1 ρ) (Xbar : E)
(hXbar : Xbar ∈ interior (ConvexRiskFn.Dual.dom ρ)) :
∃ δ : E → ℝ,
(∀ X : E, Tendsto (fun t : ℝ => ((ρ (Xbar + t • X)).toReal - (ρ Xbar).toReal) / t)
(𝓝[>] 0) (𝓝 (δ X))) ∧
(∀ X : E, ∀ t : ℝ, 0 < t → δ (t • X) = t * δ X) ∧
ConvexOn ℝ Set.univ δ ∧
(∀ X : E, ρ Xbar + ((δ (X - Xbar) : ℝ) : EReal) ≤ ρ X) := by sorry
end ConvexRiskFn.Cont
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.