(2.3), p. 4 — for a convex problem, a sharp unique minimum at x̄ iff g′(x̄, d) > 0 on T_Θ(x̄) \ {0}
ProvedSAARate.Sharp.sharp_iff_dirDeriv_posLet be convex, let be convex and let . Write for the directional derivative of at in the direction , and for the tangent cone to at . Then the following are equivalent:
- is the unique minimizer of over , and there is with for all (a sharp minimum);
- the directional derivative is positive on all nonzero tangent directions:
For this is the paper's statement that Assumption (A) holds iff (2.3). It is used twice in the proof of Theorem 2.1: at the true problem, to turn (A) into a condition on , and at the approximating problem, to turn positivity of into a sharp, hence unique, minimum at .
Formalization Note. The statement is made for an arbitrary finite convex function , not only for the expected value function ; this is stronger than the page and covers both uses. The tangent cone is Mathlib's posTangentConeAt Θ x̄, the set of limits of with , and ; for convex this is the closure of , the tangent cone of convex analysis that the paper uses. (Mathlib's tangentConeAt ℝ, which allows negative , is a different, symmetric cone and is not used.) Closedness of is not needed and not assumed.
import Mathlib import Definitions.Def_SAARate_Sharp_Setting
namespace SAARate.Sharp
open MeasureTheory ProbabilityTheory Filter Topology
theorem sharp_iff_dirDeriv_pos {m : ℕ} (g : E m → ℝ) (hg : ConvexOn ℝ Set.univ g)
(Θ : Set (E m)) (hΘ : Convex ℝ Θ) (xbar : E m) (hx : xbar ∈ Θ) :
(argminOn g Θ = {xbar} ∧ ∃ c > 0, ∀ x ∈ Θ, g x ≥ g xbar + c * ‖x - xbar‖) ↔
∀ d ∈ posTangentConeAt Θ xbar, d ≠ 0 → 0 < dirDeriv g xbar d := by sorry
end SAARate.Sharp
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.