Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Lemma 5.2, p. 23 — NPG improvement: V⁽ᵗ⁺¹⁾(µ) − V⁽ᵗ⁾(µ) ≥ ((1−γ)/η) E_{s∼µ} log Z_t(s) ≥ 0

Proved
PolicyGradTheory.NPG.npg_improvement_lower_bound

by mikedeng1 · Oct 6, 2026 · Mathlib 0df444a (Lean v4.33.1)

natural-policy-gradientp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1policy-improvementsoftmax-policy

This is Lemma 5.2 (improvement lower bound for NPG).

Let (P,r,γ)(P,r,\gamma)(P,r,γ) be a finite MDP with a nonempty action set, ρ∈Δ(S)\rho\in\Delta(\mathcal S)ρ∈Δ(S) a start distribution from which every state is reachable, and η>0\eta>0η>0. Let (θ(t))t≥0(\theta^{(t)})_{t\ge0}(θ(t))t≥0​ be any run of the natural policy gradient updates (15) with softmax policies, write V(t)V^{(t)}V(t) for the value of π(t)=πθ(t)\pi^{(t)}=\pi_{\theta^{(t)}}π(t)=πθ(t)​ and Zt(s)=∑aπ(t)(a∣s)exp⁡(ηA(t)(s,a)/(1−γ))Z_t(s)=\sum_a\pi^{(t)}(a\mid s)\exp(\eta A^{(t)}(s,a)/(1-\gamma))Zt​(s)=∑a​π(t)(a∣s)exp(ηA(t)(s,a)/(1−γ)). Then for every starting state distribution μ∈Δ(S)\mu\in\Delta(\mathcal S)μ∈Δ(S) and every ttt,

V(t+1)(μ)−V(t)(μ)  ≥  1−γη Es∼μlog⁡Zt(s)  ≥  0.V^{(t+1)}(\mu)-V^{(t)}(\mu)\;\ge\;\frac{1-\gamma}{\eta}\,\mathbb E_{s\sim\mu}\log Z_t(s)\;\ge\;0.V(t+1)(μ)−V(t)(μ)≥η1−γ​Es∼μ​logZt​(s)≥0.

The lemma shows that NPG improves the value at every start distribution, and quantifies the improvement by the log-normalizers; in the proof of Theorem 5.3 it bounds Es∼d⋆log⁡Zt(s)\mathbb E_{s\sim d^\star}\log Z_t(s)Es∼d⋆​logZt​(s) by a telescoping value difference.

Formalization Note Both inequalities of the chain are stated. Reachability of every state from ρ\rhoρ is the standing restriction of §5.3 (footnote 6); without it the bound can fail for μ\muμ charging states that ρ\rhoρ never reaches. The run may start from any θ(0)\theta^{(0)}θ(0).

Preamble
import Mathlib
import Definitions.Def_PolicyGradTheory_ProjGA_MDP
import Definitions.Def_PolicyGradTheory_NPG_Algorithm
import Definitions.Def_FoundationsML_ReinforcementLearning_IsOptimalPolicy

open FoundationsML.ReinforcementLearning
Formal statement
namespace PolicyGradTheory.NPG

/-- Lemma 5.2 (improvement lower bound for NPG; arXiv:1908.00261v5, p. 23): for the NPG
iterates and every starting state distribution `µ`,
`V^{(t+1)}(µ) − V^{(t)}(µ) ≥ ((1−γ)/η) E_{s∼µ} log Z_t(s) ≥ 0`. -/
theorem npg_improvement_lower_bound {S A : Type*} [Fintype S] [DecidableEq S] [Fintype A] [DecidableEq A]
    (P : S → A → S → ℝ) (r : S → A → ℝ) (γ : ℝ) (hM : PolicyGradTheory.ProjGA.IsFiniteMDP P r γ)
    [Nonempty A] (ρ : S → ℝ) (hρ : PolicyGradTheory.ProjGA.IsDist ρ) (hreach : AllReachable P γ ρ)
    (η : ℝ) (hη : 0 < η) (θ : ℕ → EuclideanSpace ℝ (S × A)) (hrun : IsNPGRun P r γ ρ η θ)
    (μ : S → ℝ) (hμ : PolicyGradTheory.ProjGA.IsDist μ) :
    ∀ t : ℕ,
      (1 - γ) / η * ∑ s, μ s * Real.log (npgNormalizer P r γ η (θ t) s) ≤
          PolicyGradTheory.ProjGA.valueAt (softmaxPolicy (θ (t + 1))) P r γ μ - PolicyGradTheory.ProjGA.valueAt (softmaxPolicy (θ t)) P r γ μ ∧
        0 ≤ (1 - γ) / η * ∑ s, μ s * Real.log (npgNormalizer P r γ η (θ t) s) := by sorry

end PolicyGradTheory.NPG
Source
arXiv:1908.00261v5, Lemma 5.2, p. 23
Human review
  • Endorsed by Shuze Chen · Oct 6, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Oct 6, 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