(6.2) — the proximal term does not change the generalized gradient at X*(s): ∂f̃_s(X*(s)) = ∂f_s(X*(s))
ProvedProgHedging.Nonconvex.eq_6_2generalized-gradientnonsmooth-analysisp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Let , let be a policy and , and let . Writing for Clarke's generalized gradient,
- for every ,
- consequently
Because of (6.2), the optimality conditions of Theorem 4.1 for the modified problem and for the original problem (P) coincide at ; this is how the paper turns the multiplier statement for into stationarity for (P).
Formalization Note The first identity is the sum rule for the generalized gradient of a locally Lipschitz function plus a smooth function, which the paper cites from Clarke's book (Corollary 2.4.2 of [7] as printed). It is stated for every real ; the paper's is positive. The paper prints on the left of (6.2); it means .
Preamble
import Mathlib import Definitions.Def_ProgHedging_Convex_Problem import Definitions.Def_ProgHedging_Nonconvex_Algorithm import Definitions.Def_ClarkeGradients_Shared_generalizedGradient import Definitions.Def_ClarkeGradients_FlowInvariance_normalCone open scoped Pointwise
Formal statement
namespace ProgHedging.Nonconvex
/-- Proof of Theorem 6.1, pp. 30–31, and (6.2): `∂f̃_s(x) = ∂f_s(x) + r(x − X*(s))` for every `x`,
and consequently `∂f̃_s(X*(s)) = ∂f_s(X*(s))` (Clarke generalized gradients). -/
theorem eq_6_2 {S : Type*} [Fintype S] {n T : ℕ} (pr : Problem S n T) (r : ℝ)
(Xs : ProgHedging.Convex.Policy S n) (s : S) :
(∀ x : EuclideanSpace ℝ (Fin n),
ClarkeGradients.Shared.generalizedGradient (pr.fTilde r Xs s) x =
ClarkeGradients.Shared.generalizedGradient (pr.f s) x + {r • (x - Xs s)}) ∧
ClarkeGradients.Shared.generalizedGradient (pr.fTilde r Xs s) (Xs s) =
ClarkeGradients.Shared.generalizedGradient (pr.f s) (Xs s) := by sorry
end ProgHedging.Nonconvex
Source
Rockafellar and Wets, Scenarios and policy aggregation in optimization under uncertainty, IIASA Working Paper WP-87-119 (1987), pp. 30–31, proof of Theorem 6.1, display before (6.2) and (6.2)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.