Eq. (13): Burkholder's inequality for the noise sums of a Robbins–Monro algorithm
ProvedStochApproxDyn.MartingaleNoise.burkholder_eq13Fix a dimension and an exponent . There is a constant , depending only on and , with the following property. Let , given by (7), be a Robbins–Monro algorithm on a probability space with filtration , step sequence and noise in , with times and inverse . Then for every and every
This is the instance of Burkholder's martingale inequality used in the proof of Proposition 4.2: the partial sums form a martingale in , and the right-hand side is the -th moment of its square function.
Formalization Note Both expectations are integrals in (either side may be ). The supremum over an empty range () is . The constant is chosen before the probability space, the algorithm, and , so it is universal. The paper writes "for any " before (13); the statement is about the index and the window length .
import Mathlib import Definitions.Def_StochApproxDyn_MartingaleNoise_Interpolation import Definitions.Def_StochApproxDyn_MartingaleNoise_AssumptionA1 import Definitions.Def_StochApproxDyn_MartingaleNoise_RobbinsMonro open MeasureTheory ProbabilityTheory Filter Topology open scoped ENNReal
namespace StochApproxDyn.MartingaleNoise
universe u
/-- Benaïm 1999, §4.2, proof of Proposition 4.2, Eq. (13), p. 15: the instance of Burkholder's
inequality used in the proof. For `q ≥ 2` there is a constant `C_q > 0` (depending only on `q`
and the dimension `d`) such that for every Robbins–Monro algorithm, every `n` and every `T > 0`,
`E sup_{n<k≤m(τ_n+T)} ‖∑_{i=n}^{k-1} γ_{i+1}U_{i+1}‖^q ≤ C_q E[(∑_{i=n}^{m(τ_n+T)-1} γ_{i+1}² ‖U_{i+1}‖²)^{q/2}]`. -/
theorem burkholder_eq13 (d : ℕ) (q : ℝ) (hq : 2 ≤ q) :
∃ C : ℝ, 0 < C ∧
∀ {Ω : Type u} {m0 : MeasurableSpace Ω} (P : Measure Ω) [IsProbabilityMeasure P]
(ℱ : Filtration ℕ m0) (F : EuclideanSpace ℝ (Fin d) → EuclideanSpace ℝ (Fin d))
(γ : ℕ → ℝ) (x U : ℕ → Ω → EuclideanSpace ℝ (Fin d)),
IsRobbinsMonro P ℱ F γ x U →
∀ (n : ℕ) (T : ℝ), 0 < T →
∫⁻ ω, (Finset.Ioc n (stepIndex γ (StochApproxDyn.Interpolation.tau γ n + T))).sup
(fun k => ‖∑ i ∈ Finset.Ico n k, γ (i + 1) • U (i + 1) ω‖ₑ ^ q) ∂P
≤ ENNReal.ofReal C *
∫⁻ ω, (∑ i ∈ Finset.Ico n (stepIndex γ (StochApproxDyn.Interpolation.tau γ n + T)),
ENNReal.ofReal (γ (i + 1) ^ 2) * ‖U (i + 1) ω‖ₑ ^ (2 : ℝ)) ^ (q / 2) ∂P := by sorry
end StochApproxDyn.MartingaleNoise
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.