Eq. (4.6) — the residuals of Algorithm 3.1 tend to zero
ProvedAndersonAccel.Safe.residual_tendsto_zeroanderson-accelerationfixed-point-iterationp2o-batch-p200ap2o-gran-per-chapterp2o-plan-paperp2o-v1residual
Let be nonexpansive with a nonempty fixed-point set, and run Algorithm 3.1 with , and max-memory , assuming no iterate is a solution ( for all ). Then the residuals satisfy
This is Step 1 of the global convergence proof: both the averaged steps and the accepted accelerated steps drive the residual to zero.
Formalization Note The hypothesis is the paper's standing simplification at the start of §4.1.
Preamble
import Mathlib import Definitions.Def_AndersonAccel_Safe_Basic import Definitions.Def_AndersonAccel_Safe_IsAAISRun
Formal statement
namespace AndersonAccel.Safe
open Filter Topology
/-- Eq. (4.6) (p. 3181). The residuals of the iterates of Algorithm 3.1 tend to zero:
`lim_{k → ∞} ‖g_k‖ = 0`. -/
theorem residual_tendsto_zero {n : ℕ} (f : EuclideanSpace ℝ (Fin n) → EuclideanSpace ℝ (Fin n)) (hf : LipschitzWith 1 f)
(θbar τ α D ε : ℝ) (m : ℕ)
(hθ0 : 0 < θbar) (hθ1 : θbar < 1) (hτ0 : 0 < τ) (hτ1 : τ < 1) (hα0 : 0 < α) (hα1 : α < 1)
(hD : 0 < D) (hε : 0 < ε) (hm : 0 < m)
(x xt s y shat ytil : ℕ → EuclideanSpace ℝ (Fin n))
(H : ℕ → (EuclideanSpace ℝ (Fin n) →L[ℝ] EuclideanSpace ℝ (Fin n))) (mem nAA : ℕ → ℕ)
(hrun : IsAAISRun f θbar τ α D ε m x xt s y shat ytil H mem nAA)
(hX : ∃ z, f z = z)
(hnosol : ∀ k, f (x k) ≠ x k) :
Tendsto (fun k => ‖residual f (x k)‖) atTop (𝓝 0) := by sorry
end AndersonAccel.Safe
Source
Zhang, O'Donoghue, Boyd, SIAM J. Optim. 30 (2020), p. 3181, Section 4.1, Step 1, Eq. (4.6)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.