Lemma 1 — the Bregman resolvent is single-valued and decreases distance to a zero
ProvedBregmanPPA.Convergence.lemma1bregman-functionmonotone-operatorp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1
Let be a monotone operator on a finite-dimensional real inner-product space, and let be a Bregman function with open zone . For , any point satisfying is unique. If is a zero of and such a exists, then
This is the resolvent inequality used to control every step of the proximal-point run.
Formalization Note The resolvent is expressed as a relation on , which preserves its possible nonexistence. The gradient is evaluated only in ; the zero may lie on the boundary .
Preamble
import Mathlib import Definitions.Def_BregmanPPA_Convergence_Model open ThreeOpSplitting.Convergence Filter Topology
Formal statement
namespace BregmanPPA.Convergence
variable {H : Type*} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [FiniteDimensional ℝ H]
/-- Lemma 1: uniqueness of the Bregman resolvent and its three-point inequality. -/
theorem lemma1 (T : H → Set H) (S : Set H) (h : H → ℝ)
(hT : IsMonotoneOp T) (hh : IsBregmanFunction S h) :
(∀ (y p q : H), y ∈ S → p ∈ S → q ∈ S →
gradient h y - gradient h p ∈ T p →
gradient h y - gradient h q ∈ T q → p = q) ∧
(∀ (z y p : H), z ∈ zer T → z ∈ closure S → y ∈ S → p ∈ S →
gradient h y - gradient h p ∈ T p →
bregmanD h z p ≤ bregmanD h z y - bregmanD h p y) := by sorry
end BregmanPPA.Convergence
Source
Eckstein, Nonlinear proximal point algorithms using Bregman functions, with applications to convex programming, Math. Oper. Res. 18(1) (1993), p. 206, Lemma 1, https://doi.org/10.1287/moor.18.1.202
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.