(P_k), §3.3, p. 487 — ∂g*(y) is the set of minimizers of g − ⟨·, y⟩
ProvedTaoAnDCA.TRS.subdiff_conj_eq_argminconjugateconvex-analysisdcap2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1subdifferential
Let and . Then the subdifferential of the conjugate at is the set of minimizers of tilted by :
where only points with count as minimizers.
In the DCA, applied with , this is the identification (P_k): the step amounts to solving the convex program obtained from (P) by replacing with its affine minorization .
Formalization Note. The paper writes the objective as ; it differs from by the constant , which does not change the set of minimizers, so the statement is given in this general form. The minimizers are required to lie in , as the subdifferential is empty off and a minimizer of a proper function has a finite value.
Preamble
import Mathlib import Definitions.Def_TaoAnDCA_TRS_Setting open Filter Topology
Formal statement
namespace TaoAnDCA.TRS
theorem subdiff_conj_eq_argmin {n : ℕ} (g : EuclideanSpace ℝ (Fin n) → EReal)
(hg : ThreeOpSplitting.ConvexRates.IsProperClosedConvex g) (y : EuclideanSpace ℝ (Fin n)) :
TaoAnDCA.GlobalOpt.subdiff (CondatPD.FinDim.conj g) y =
{x | g x ≠ ⊤ ∧ ∀ z, g x - ((inner ℝ x y : ℝ) : EReal) ≤ g z - ((inner ℝ z y : ℝ) : EReal)} := by sorry
end TaoAnDCA.TRS
Source
Pham Dinh & Le Thi, A d.c. optimization algorithm for solving the trust-region subproblem, SIAM J. Optim. 8 (1998), p. 487, §3.3, display (P_k)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.