Theorem 3.4, p. 10 — with linesearch and locally Lipschitz B, the method converges to a zero of A + B in finite dimensions
ProvedFRBSplitting.Linesearch.theorem_3_4Let be a finite-dimensional real inner product space, maximally monotone with resolvents , and monotone and locally Lipschitz continuous, and suppose . Let , , and , and let be generated by Algorithm 1: in every iteration some is chosen, with the smallest nonnegative integer such that
Then there is with
The linesearch removes the need to know a Lipschitz constant of , and the result covers operators that are only locally Lipschitz.
Formalization Note. The resolvent is a hypothesis: J : ℝ → H → H with IsResolventFamily A J (for maximally monotone it exists and is unique on ). The run predicate requires, at every iteration, the chosen and the least accepted index ; may vary between iterations. Indices are shifted: x j , lam j ; Algorithm 1's is never used by the iteration and is not a parameter. Convergence is in norm (equivalently in any norm, being finite dimensional).
import Mathlib import Definitions.Def_ThreeOpSplitting_Accel_MonotoneOperators import Definitions.Def_FRBSplitting_Linesearch_Setting open InnerProductSpace ThreeOpSplitting.Accel
namespace FRBSplitting.Linesearch
theorem theorem_3_4 {H : Type*} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [FiniteDimensional ℝ H]
(A : H → Set H) (B : H → H) (J : ℝ → H → H) (δ σ : ℝ)
(hA : IsMaximalMonotone A)
(hJ : IsResolventFamily A J)
(hBm : IsMonotoneFun B) (hBloc : LocallyLipschitz B)
(hzer : (FRBSplitting.Weak.zeroSet A B).Nonempty)
(hδ0 : 0 < δ) (hδ1 : δ < 1) (hσ0 : 0 < σ) (hσ1 : σ < 1)
(lam : ℕ → ℝ) (x : ℕ → H) (hlam0 : 0 < lam 0) (hrun : IsLinesearchRun J B δ σ lam x) :
∃ xs ∈ FRBSplitting.Weak.zeroSet A B, Filter.Tendsto x Filter.atTop (nhds xs) := by sorry
end FRBSplitting.Linesearch
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.