v21 Lemma 7: general secant degeneration
Lean contributors (project-level attribution): Zichen Wang.
public sectionopen Filteropen DFP.PlanarConvergence.SecantDegenerationopen scoped Topology ContDiffnamespace DFPWolfe.V21Helper for lem:planar-degeneration and eq:planar-scalar-definition: the source step and secant are the actual displacement and gradient difference.
theorem planarStepSecantIdentities (iteration : DFP.SecantIteration 2) (k : ℕ) :
step iteration k = iteration.point (k + 1) - iteration.point k ∧
secant iteration k = gradient iteration.objective (iteration.point (k + 1)) -
gradient iteration.objective (iteration.point k) := by
exact ⟨step_def iteration k, secant_def iteration k⟩
Helper for lem:planar-degeneration and eq:planar-scalar-definition: the
curvature is , and the original coefficient for k≥1 is
. Natural subtraction merely totalizes the latter at zero.
theorem planarScalarDefinitions (iteration : DFP.SecantIteration 2) (k : ℕ) :
curvature iteration k = inner ℝ (step iteration k) (secant iteration k) /
‖step iteration k‖ ^ 2 ∧
freeCoefficient iteration k =
curvature iteration (k - 1) * (iteration.inverseHessian k).det := by
exact ⟨curvature_def iteration k, freeCoefficient_def iteration k⟩
lem:planar-degeneration (v21 Lemma 7). A C² globally μ-strongly convex planar
objective has a unique global minimizer. For nonstationary iterates generated by
symmetric positive definite search matrices satisfying the previous secant equation,
with positive step lengths and fixed weak Wolfe conditions, failure of convergence to
this minimizer forces degeneration of the source free coefficient and smallest eigenvalue.
The predicate LineSearch.IsWeakWolfe includes and both source inequalities.
With the exact scalars in planarScalarDefinitions, eq:planar-degeneration is expressed
by the projections of DegenerationEstimates. positive gives for
every k≥1; product_bound gives eventually, with one C>0;
coefficient_limit gives . operator_bound bounds the induced Euclidean norm
‖Matrix.toEuclideanCLM H_k‖ for every k, including H₀. determinant_limit gives
det H_k→0; smallest_eigenvalue_limit gives convergence to zero of the canonical
decreasingly ordered Hermitian eigenvalue at index 1, the planar minimum.
The accepted proof follows eq:planar-secant-representation and its energy absorption
argument. The general DFP.SecantIteration interface requires neither a DFP update
nor Hessian Lipschitz continuity.
theorem planarDegeneration (iteration : DFP.SecantIteration 2) {μ c₁ c₂ : ℝ}
(hf : ContDiff ℝ 2 iteration.objective)
(hstrong : StrongConvexOn Set.univ μ iteration.objective)
(hμ : 0 < μ) (hStep : ∀ k, 0 < iteration.stepLength k)
(hw : ∀ k, LineSearch.IsWeakWolfe c₁ c₂ iteration.objective
(iteration.point k) (iteration.point (k + 1) - iteration.point k)) :
∃! xstar : EuclideanSpace ℝ (Fin 2),
(∀ z, iteration.objective xstar ≤ iteration.objective z) ∧
(¬ Tendsto iteration.point atTop (𝓝 xstar) → DegenerationEstimates iteration) := by
exact DFP.PlanarConvergence.SecantDegeneration.planarDegeneration
iteration hf hstrong hμ hStep hwend DFPWolfe.V21