Theorem 3.3 — under the Strong Hypotheses
ProvedLewisTorczon.BoundPS.theorem_3_3Consider the bound constrained problem on , with for every (infinite bounds allowed), and let , where is the projection onto . Let be a run of the generalized pattern search method for bound constrained minimization (Algorithm 1 with Algorithm 2). Suppose that
- is compact;
- is continuously differentiable on an open set containing ;
- the columns of the generating matrices are uniformly bounded in norm;
- ;
- the run enforces the Strong Hypotheses on Bound Constrained Exploratory Moves.
Then
Since exactly at stationary points of the bound constrained problem, every limit point of the iterates is then a stationary point: the whole sequence, not just a subsequence, approaches first-order stationarity.
Formalization Note Two departures from the page, both disclosed in the mission description. (i) The page assumes continuously differentiable on ; here it is assumed on an open set , which the proof needs (it uses at trial points in outside , and may have empty interior). (ii) The third Strong Hypothesis is encoded with "" instead of the printed "" (the printed strict form excludes the paper's own examples; "" is the weaker hypothesis). Norms are Euclidean.
import Mathlib import Definitions.Def_LewisTorczon_BoundPS_Box import Definitions.Def_LewisTorczon_BoundPS_ProjQ import Definitions.Def_LewisTorczon_BoundPS_GPS
namespace LewisTorczon.BoundPS
/-- **Theorem 3.3**, p. 7: if `L_Ω(x_0)` is compact, `f` is `C¹`, the columns of the generating
matrices are uniformly bounded in norm, `Δ_k → 0`, and the method enforces the Strong Hypotheses
on Bound Constrained Exploratory Moves, then `‖q(x_k)‖ → 0`. -/
theorem theorem_3_3 {n m : ℕ} (P : GPSParams n) (lo hi : Fin n → EReal)
(hlohi : ∀ j, lo j < hi j) (f : EuclideanSpace ℝ (Fin n) → ℝ)
(U : Set (EuclideanSpace ℝ (Fin n))) (hU : IsOpen U) (hΩU : box lo hi ⊆ U)
(hf : ContDiffOn ℝ 1 f U) (R : GPSRun n m) (hR : IsGPSRun P lo hi f R)
(hcpt : IsCompact (levelSet lo hi f (R.x 0))) (hbdd : BoundedCols R)
(hΔ : Filter.Tendsto R.Δ Filter.atTop (nhds 0)) (hstrong : StrongHyp P lo hi f R) :
Filter.Tendsto (fun k => ‖projQ lo hi f (R.x k)‖) Filter.atTop (nhds 0) := by sorry
end LewisTorczon.BoundPS
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.