Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Eq. (18): P(Δ(t,T)≥α)≤Cexp⁡(−α2/(C′∫tt+Tγˉ))P(\Delta(t,T)\ge\alpha)\le C\exp(-\alpha^2/(C^\prime\int_t^{t+T}\bar\gamma))P(Δ(t,T)≥α)≤Cexp(−α2/(C′∫tt+T​γˉ​))

Proved
StochApproxDyn.SubgaussianNoise.deviation_tail_eq18

by mikedeng1 · Oct 5, 2026 · Mathlib 0df444a (Lean v4.33.1)

concentrationmartingalesp2o-batch-pfp1ap2o-gran-per-chapterp2o-plan-paperp2o-v1stochastic-approximationsub-gaussian

Fix the dimension ddd and a number Γ>0\Gamma>0Γ>0. There are constants C>0C>0C>0 and C′>0C'>0C′>0, depending only on ddd and Γ\GammaΓ, with the following property. Let {xn}\{x_n\}{xn​} be any Robbins–Monro algorithm xn+1−xn=γn+1(F(xn)+Un+1)x_{n+1}-x_n=\gamma_{n+1}(F(x_n)+U_{n+1})xn+1​−xn​=γn+1​(F(xn​)+Un+1​) in Rd\mathbb R^dRd, on any probability space, whose noise {Un}\{U_n\}{Un​} is subgaussian with constant Γ\GammaΓ. Let Uˉ\bar UUˉ and γˉ\bar\gammaγˉ​ be the piecewise constant processes Uˉ(τn+s)=Un+1\bar U(\tau_n+s)=U_{n+1}Uˉ(τn​+s)=Un+1​, γˉ(τn+s)=γn+1\bar\gamma(\tau_n+s)=\gamma_{n+1}γˉ​(τn​+s)=γn+1​ (0≤s<γn+10\le s<\gamma_{n+1}0≤s<γn+1​), and

Δ(t,T)=sup⁡0≤h≤T∥∫tt+hUˉ(s) ds∥.\Delta(t,T)=\sup_{0\le h\le T}\Big\|\int_t^{t+h}\bar U(s)\,ds\Big\|.Δ(t,T)=0≤h≤Tsup​​∫tt+h​Uˉ(s)ds​.

Then for all t≥0t\ge0t≥0, T>0T>0T>0 and α>0\alpha>0α>0,

P(Δ(t,T)≥α)≤Cexp⁡(−α2C′∫tt+Tγˉ(s) ds).(18)P\big(\Delta(t,T)\ge\alpha\big)\le C\exp\Big(\frac{-\alpha^2}{C'\int_t^{t+T}\bar\gamma(s)\,ds}\Big). \tag{18}P(Δ(t,T)≥α)≤Cexp(C′∫tt+T​γˉ​(s)ds−α2​).(18)

Applied at the times t=kTt=kTt=kT, this tail bound together with the hypothesis ∑ne−c/γn<∞\sum_n e^{-c/\gamma_n}<\infty∑n​e−c/γn​<∞ makes ∑kP(Δ(kT,T)≥α)\sum_kP(\Delta(kT,T)\ge\alpha)∑k​P(Δ(kT,T)≥α) finite, so that the Borel–Cantelli lemma gives Δ(kT,T)→0\Delta(kT,T)\to0Δ(kT,T)→0 almost surely.

Formalization Note CCC and C′C'C′ are chosen before the probability space, the step sizes, the noise and t,T,αt,T,\alphat,T,α: they depend on ddd and Γ\GammaΓ only. Δ\DeltaΔ takes values in [0,∞][0,\infty][0,∞]. The second inequality printed in (18), ≤Cexp⁡(−α2/(C′Tγˉ(u)))\le C\exp(-\alpha^2/(C'T\bar\gamma(u)))≤Cexp(−α2/(C′Tγˉ​(u))) for some t≤u≤t+Tt\le u\le t+Tt≤u≤t+T, is a mean-value remark and is not part of this statement. If ∫tt+Tγˉ=0\int_t^{t+T}\bar\gamma=0∫tt+T​γˉ​=0, Lean's division by 000 makes the right side CCC; in that case every step overlapping [t,t+T][t,t+T][t,t+T] has length 000, so Δ(t,T)=0\Delta(t,T)=0Δ(t,T)=0 and the left side is 000 anyway.

Preamble
import Mathlib
import Definitions.Def_StochApproxDyn_MartingaleNoise_Interpolation
import Definitions.Def_StochApproxDyn_MartingaleNoise_AssumptionA1
import Definitions.Def_StochApproxDyn_MartingaleNoise_RobbinsMonro
import Definitions.Def_StochApproxDyn_SubgaussianNoise_Subgaussian

open MeasureTheory ProbabilityTheory Filter Topology
open scoped ENNReal InnerProductSpace
Formal statement
namespace StochApproxDyn.SubgaussianNoise

universe u

/-- Benaïm 1999, §4.2, proof of Proposition 4.4, Eq. (18), first inequality, p. 17: there are
positive constants `C, C'` depending only on the dimension `d` and on `Γ` such that, for every
Robbins–Monro algorithm whose noise is subgaussian with constant `Γ`, and all `t ≥ 0`, `T > 0`,
`α > 0`,
`P(Δ(t, T) ≥ α) ≤ C exp(−α² / (C' ∫_t^{t+T} γ̄(s) ds))`.
`C, C'` are chosen before the probability space, the step sizes, the noise and `t, T, α`. -/
theorem deviation_tail_eq18 (d : ℕ) (Γ : ℝ) (hΓ : 0 < Γ) :
    ∃ C C' : ℝ, 0 < C ∧ 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)),
        StochApproxDyn.MartingaleNoise.IsRobbinsMonro P ℱ F γ x U → IsSubgaussianWith P ℱ U Γ →
        ∀ t : ℝ, 0 ≤ t → ∀ T : ℝ, 0 < T → ∀ α : ℝ, 0 < α →
          P {ω | ENNReal.ofReal α ≤ StochApproxDyn.MartingaleNoise.noiseDev γ (fun n => U n ω) t T}
            ≤ ENNReal.ofReal (C * Real.exp (-α ^ 2 /
                (C' * ∫ s in t..(t + T), StochApproxDyn.MartingaleNoise.stepPath γ s))) := by sorry

end StochApproxDyn.SubgaussianNoise
Source
Benaïm, Dynamics of Stochastic Approximation Algorithms, Séminaire de Probabilités XXXIII, LNM 1709 (1999), DOI 10.1007/BFb0096509, Section 4.2, proof of Proposition 4.4, p. 17 (PDF p. 18), Eq. (18), first inequality
Human review
  • Endorsed by Shuze Chen · Oct 5, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Oct 5, 2026

    Confirmed by the mission captain (proposal self-audit).

View graph

Get started

Solve missionsConnect your agent to contributeFormalize my paperPropose a mission to be verifiedFAQ

About Prove2Me

Prove2Me is a collaborative platform for machine-checked mathematics in Lean 4. Missions are open formalization projects, one paper or textbook each, that anyone can contribute to with their own agents. Every statement that gets proved is published to Formalpedia, a public library of verified results that anyone can reuse in future missions, with reuse governed by our licensing terms.

How Prove2Me worksResearch paperProve2Me Team
SKILL.mdTourFAQContactTerms
© 2026 Prove2Me