Lemma 3.1, p. 345 — three-point property: φ(x) + D_h(x,z) ≥ φ(z⁺) + D_h(z⁺,z) + D_h(x,z⁺)
ProvedRelSmoothFOM.PrimalGrad.lemma_3_1This is the three-point property of Tseng.
Let be a convex subset of a finite-dimensional real inner-product space , let be differentiable at every point of and convex on , and let be its Bregman distance. Let be convex on , let , and let be a minimiser over of . Then
The lemma quantifies how much the proximal objective grows when one moves from the minimiser to any other feasible point, in terms of the Bregman distance from . It is the step that turns one iteration of a Bregman proximal method into a telescoping inequality.
Formalization Note The minimiser is a hypothesis on a given point (it minimises over ), not a value of an argmin operator, so no existence or uniqueness of minimisers is assumed or claimed. The function need not be differentiable. is assumed because the paper's Bregman distance (7) is defined on .
import Mathlib import Definitions.Def_RelSmoothFOM_PrimalGrad_Setting
namespace RelSmoothFOM.PrimalGrad
theorem lemma_3_1 {E : Type*} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
[FiniteDimensional ℝ E]
(Q : Set E) (hQ : Convex ℝ Q) (φ : E → ℝ) (hφ : ConvexOn ℝ Q φ) (h : E → ℝ)
(hhd : ∀ x ∈ Q, DifferentiableAt ℝ h x) (hhc : ConvexOn ℝ Q h)
(z zp : E) (hz : z ∈ Q) (hzp : zp ∈ Q)
(hmin : ∀ u ∈ Q, φ zp + bregman h zp z ≤ φ u + bregman h u z) :
∀ u ∈ Q, φ u + bregman h u z ≥ φ zp + bregman h zp z + bregman h u zp := by sorry
end RelSmoothFOM.PrimalGrad
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.