Theorem 1 — the Bregman proximal point algorithm converges to a zero of a maximal monotone T
ProvedBregmanPPA.ProxMult.theorem1Let be a finite-dimensional real inner product space, a maximal monotone operator, and a Bregman function with zone , . Let be positive scalars with , and let be an infinite sequence conforming to
Suppose that (C1) , or (C2) for a proper lower semicontinuous convex . Then if has a zero, converges to one of them. If has no zero and (C1) holds, is unbounded.
This is the paper's central theorem, restated in this mission so that it can be applied to the saddle operator on .
Formalization Note The paper states the theorem on (and prints "maximal monontone operator of "); the restatement is over any finite-dimensional real inner product space, which is up to isometry and includes as WithLp 2 (E n × E m). (3) is encoded by (4) with . "Converges to one of them" is ; "unbounded" is "the range of is not bounded".
import Mathlib import Definitions.Def_BregmanPPA_Convergence_Model open Filter Topology ThreeOpSplitting.Convergence InertialFB.IFB
namespace BregmanPPA.ProxMult
variable {H : Type*} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [FiniteDimensional ℝ H]
/-- Theorem 1, pp. 206–207, restated on a finite-dimensional inner product space `H` (it is
applied on `ℝⁿ⁺ᵐ`): convergence to a zero, or unboundedness when there is none under (C1). -/
theorem theorem1 (T : H → Set H) (S : Set H) (h : H → ℝ) (c : ℕ → ℝ) (x : ℕ → H)
(hT : IsMaximalMonotone T) (hh : BregmanPPA.Convergence.IsBregmanFunction S h)
(hdom : dom T ⊆ closure S) (hc : ∀ k, 0 < c k)
(hcinf : ∃ ε : ℝ, 0 < ε ∧ ∀ k, ε ≤ c k)
(hx : BregmanPPA.Convergence.IsBregmanPPARun S h T c x)
(hC : closure (dom T) ⊆ S ∨ ∃ f : H → EReal,
IsProperFn f ∧ IsConvexFn f ∧ LowerSemicontinuous f ∧ T = BregmanPPA.Convergence.subdiffOp f) :
((zer T).Nonempty → ∃ z : H, z ∈ zer T ∧ Tendsto x atTop (𝓝 z)) ∧
(zer T = ∅ → closure (dom T) ⊆ S → ¬ Bornology.IsBounded (Set.range x)) := by sorry
end BregmanPPA.ProxMult
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.