Proposition 4.1 — under A1 and A2 (or A2′) the interpolated process is an asymptotic pseudotrajectory of the flow of F
ProvedStochApproxDyn.Interpolation.interpAffine_isAsymptoticPseudotrajectoryLet be a continuous, globally integrable vector field, and let be the flow it induces. Let be generated by
where the step sizes satisfy , and , and are perturbations. Let be the affine interpolation of on the time scale . Assume
- A1: for all , ; and
- either A2: , or A2′: is Lipschitz and bounded on a neighbourhood of .
Then is an asymptotic pseudotrajectory of : for every ,
This is the deterministic half of the ODE method: it separates the dynamics of a stochastic approximation algorithm from its noise. Once a noise condition such as A1 has been verified (almost surely, in Propositions 4.2 and 4.4), the long-run behaviour of is governed by the flow of , and the limit set theorems for asymptotic pseudotrajectories apply.
Formalization Note The neighbourhood in A2′ is read as a uniform one: is Lipschitz and bounded on for some . Under the weaker reading "some open set containing " the conclusion can fail for unbounded iterates (the neighbourhood may thin out between consecutive iterates), so this is the reading under which the proposition holds. Time is ℝ≥0; the flow is used at times only. Sequences carry the paper's indices; and are unused.
import Mathlib import Definitions.Def_StochApproxDyn_Interpolation_AsymptoticPseudotrajectory import Definitions.Def_StochApproxDyn_Interpolation_VectorFieldFlow import Definitions.Def_StochApproxDyn_Interpolation_Scheme open scoped NNReal Topology open Filter
namespace StochApproxDyn.Interpolation
/-- Benaïm 1999, Proposition 4.1, first conclusion, pp. 12–13. Let `F : ℝ^d → ℝ^d` be continuous
and globally integrable, with flow `Φ`. Let `{x_n}` follow the recursion (7) with step sizes
satisfying the standing assumptions (`γ_n ≥ 0`, `Σ γ_n = ∞`, `γ_n → 0`). Assume A1 and either
A2 (`sup_n ‖x_n‖ < ∞`) or A2′ (`F` Lipschitz and bounded on a uniform neighbourhood of
`{x_n : n ≥ 0}`). Then the affine interpolated process `X` is an asymptotic pseudotrajectory of
`Φ`. -/
theorem interpAffine_isAsymptoticPseudotrajectory {d : ℕ}
(F : EuclideanSpace ℝ (Fin d) → EuclideanSpace ℝ (Fin d)) (hF : Continuous F)
(hint : IsGloballyIntegrable F)
(Φ : ℝ → EuclideanSpace ℝ (Fin d) → EuclideanSpace ℝ (Fin d)) (hΦ : IsFlowOf F Φ)
(γ : ℕ → ℝ) (hγ : IsStepSizeSeq γ) (x U : ℕ → EuclideanSpace ℝ (Fin d))
(hx : SatisfiesScheme F γ x U) (hA1 : SatisfiesA1 γ U)
(hA2 : SatisfiesA2 x ∨ SatisfiesA2' F x) :
IsAsymptoticPseudotrajectory Φ (fun t : ℝ≥0 => interpAffine γ x (t : ℝ)) := by sorry
end StochApproxDyn.Interpolation
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.