Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 6.4 — a convergence criterion for nonmonotone sequences

Proved
NumStochOpt.Nonstationary.theorem_6_4_nonmonotone_convergence_criterion

by mikedeng1 · Oct 3, 2026 · Mathlib 0df444a (Lean v4.33.1)

convergencelyapunov-functionp2o-batch-b23ap2o-gran-per-chapterp2o-plan-bookp2o-v1stochastic-optimization

Let X∗⊆RnX^* \subseteq \mathbb R^nX∗⊆Rn be closed and let (xs)s≥0(x^s)_{s \ge 0}(xs)s≥0​ be a sequence in Rn\mathbb R^nRn such that:

  1. xs∈Kx^s \in Kxs∈K for all sss, with KKK compact;
  2. for every subsequence (xsk)(x^{s_k})(xsk​) with lim⁡kxsk=x′\lim_k x^{s_k} = x'limk​xsk​=x′:
    • (a) if x′∈X∗x' \in X^*x′∈X∗, then ∥xsk+1−xsk∥→0\|x^{s_k+1} - x^{s_k}\| \to 0∥xsk​+1−xsk​∥→0 as k→∞k \to \inftyk→∞;
    • (b) if x′∉X∗x' \notin X^*x′∈/X∗, then for every sufficiently small ε>0\varepsilon > 0ε>0 and every kkk the exit time
τk=min⁡{s≥sk:∥xs−xsk∥>ε}\tau_k = \min\{s \ge s_k : \|x^s - x^{s_k}\| > \varepsilon\}τk​=min{s≥sk​:∥xs−xsk​∥>ε}

is finite; 3. there is a continuous function V:Rn→RV : \mathbb R^n \to \mathbb RV:Rn→R taking at most countably many values on X∗X^*X∗ such that, in the situation of (2)(b),

lim sup⁡k→∞V(xτk)<lim⁡k→∞V(xsk).\limsup_{k \to \infty} V(x^{\tau_k}) < \lim_{k \to \infty} V(x^{s_k}).k→∞limsup​V(xτk​)<k→∞lim​V(xsk​).

Then the sequence (V(xs))(V(x^s))(V(xs)) converges, and every accumulation point of (xs)(x^s)(xs) belongs to X∗X^*X∗.

The criterion replaces monotone decrease of a Lyapunov function by the requirement that the sequence leave any neighbourhood of a non-solution with a lower value of VVV; it applies to essentially nonmonotone procedures such as the nonstationary method (6.41), where Zangwill-type conditions are hard to verify. The book cites the proof to Ermoliev's 1976 monograph [5, p. 181].

Formalization Note The page prints (2)(b) as "τk=min⁡{s∣s≥sk,∥xsk−xs∥<ε}>∞\tau_k = \min\{s \mid s \ge s_k, \|x^{s_k} - x^s\| < \varepsilon\} > \inftyτk​=min{s∣s≥sk​,∥xsk​−xs∥<ε}>∞", which no sequence satisfies. The proof of Theorem 6.3 (pp. 155–156) shows what is intended: τk\tau_kτk​ is the first index from sks_ksk​ on with ∥xs−xsk∥>ε\|x^s - x^{s_k}\| > \varepsilon∥xs−xsk​∥>ε, and (2)(b) asserts τk<∞\tau_k < \inftyτk​<∞; that is the reading formalized. "For ε\varepsilonε sufficiently small and for any sks_ksk​" is read as: there is ε0>0\varepsilon_0 > 0ε0​>0 such that for all ε∈(0,ε0)\varepsilon \in (0, \varepsilon_0)ε∈(0,ε0​) and all kkk; condition (3) is required for the same ε\varepsilonε. The left-hand limit of (3) is read as a lim sup⁡\limsuplimsup (the proof of Theorem 6.3 prints lim⁡‾\overline{\lim}lim); the right-hand limit exists and equals V(x′)V(x')V(x′) by continuity of VVV, and is written so. The real lim sup⁡\limsuplimsup is well defined because VVV is bounded on the compact set KKK. Subsequences are strictly increasing index maps φ:N→N\varphi : \mathbb N \to \mathbb Nφ:N→N with sk=φ(k)s_k = \varphi(k)sk​=φ(k); accumulation points are cluster points of the sequence along atTop.

Preamble
import Mathlib
import Definitions.Def_NumStochOpt_Nonstationary_ExitTime

open Filter Topology
Formal statement
namespace NumStochOpt.Nonstationary

/-- Theorem 6.4 (Ermoliev, Ch. 6 of Ermoliev & Wets (1988), p. 155, citing [5, p. 181]): a
convergence criterion for nonmonotone sequences. Let `X* ⊆ ℝⁿ` be closed and `(x^s)` a sequence
in `ℝⁿ` such that
(1) all `x^s` lie in a compact set `K`;
(2) for every subsequence `x^{s_k} → x'`:
  (a) if `x' ∈ X*` then `‖x^{s_k+1} - x^{s_k}‖ → 0`;
  (b) if `x' ∉ X*` then for all sufficiently small `ε > 0` and every `k` the exit time
      `τ_k = min {s ≥ s_k | ‖x^s - x^{s_k}‖ > ε}` is finite, and
(3) for a continuous `V` taking at most countably many values on `X*`,
    `limsup_k V(x^{τ_k}) < lim_k V(x^{s_k}) = V(x')`.
Then `V(x^s)` converges and every accumulation point of `(x^s)` lies in `X*`.
The page prints `τ_k = min{s | s ≥ s_k, ‖x^{s_k} - x^s‖ < ε} > ∞`; the reading used is the
one the proof of Theorem 6.3 (pp. 155–156) verifies. -/
theorem theorem_6_4_nonmonotone_convergence_criterion {n : ℕ}
    (Xstar K : Set (EuclideanSpace ℝ (Fin n))) (x : ℕ → EuclideanSpace ℝ (Fin n))
    (V : EuclideanSpace ℝ (Fin n) → ℝ)
    (hXstar : IsClosed Xstar)
    (hK : IsCompact K) (hxK : ∀ s, x s ∈ K)
    (h2a : ∀ (φ : ℕ → ℕ) (x' : EuclideanSpace ℝ (Fin n)), StrictMono φ →
      Tendsto (x ∘ φ) atTop (𝓝 x') → x' ∈ Xstar →
        Tendsto (fun k => ‖x (φ k + 1) - x (φ k)‖) atTop (𝓝 0))
    (hV : Continuous V) (hVcount : (V '' Xstar).Countable)
    (h2b3 : ∀ (φ : ℕ → ℕ) (x' : EuclideanSpace ℝ (Fin n)), StrictMono φ →
      Tendsto (x ∘ φ) atTop (𝓝 x') → x' ∉ Xstar →
        ∃ ε₀ : ℝ, 0 < ε₀ ∧ ∀ ε : ℝ, 0 < ε → ε < ε₀ →
          (∀ k, ∃ s, φ k ≤ s ∧ ε < ‖x s - x (φ k)‖) ∧
          limsup (fun k => V (x (exitTime x (φ k) ε))) atTop < V x') :
    (∃ L : ℝ, Tendsto (fun s => V (x s)) atTop (𝓝 L)) ∧
      ∀ x' : EuclideanSpace ℝ (Fin n), MapClusterPt x' atTop x → x' ∈ Xstar := by sorry

end NumStochOpt.Nonstationary
Source
Yu. Ermoliev, "Stochastic Quasigradient Methods", in Ermoliev & Wets (eds.), Numerical Techniques for Stochastic Optimization, Springer 1988, Ch. 6, p. 155, Theorem 6.4 (citing Yu. Ermoliev, Methods of Stochastic Programming, Nauka 1976, p. 181)
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