Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Lemma 4.1, p. 14 — gradient domination: V⋆(ρ) − V^π(ρ) ≤ ‖d^{π⋆}_ρ/d^π_μ‖_∞ max_π̄ (π̄−π)ᵀ∇V ≤ (1/(1−γ))‖d^{π⋆}_ρ/μ‖_∞ max_π̄ (π̄−π)ᵀ∇V

Proved
PolicyGradTheory.ProjGA.gradient_domination

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

distribution-mismatchgradient-dominationp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1policy-gradient

Let (S,A,P,r,γ)(\mathcal S,\mathcal A,P,r,\gamma)(S,A,P,r,γ) be a finite discounted MDP with rewards in [0,1][0,1][0,1] and γ∈[0,1)\gamma\in[0,1)γ∈[0,1), let μ,ρ∈Δ(S)\mu,\rho\in\Delta(\mathcal S)μ,ρ∈Δ(S) be state distributions, and let π⋆\pi^\starπ⋆ be an optimal policy, with V⋆=Vπ⋆V^\star=V^{\pi^\star}V⋆=Vπ⋆. Use the direct parameterization, and write ∇πVπ(μ)\nabla_\pi V^\pi(\mu)∇π​Vπ(μ) for the gradient of π↦Vπ(μ)\pi\mapsto V^\pi(\mu)π↦Vπ(μ).

Lemma (gradient domination). For every policy π∈Δ(A)∣S∣\pi\in\Delta(\mathcal A)^{|\mathcal S|}π∈Δ(A)∣S∣,

V⋆(ρ)−Vπ(ρ)  ≤  ∥dρπ⋆dμπ∥∞max⁡πˉ (πˉ−π)⊤∇πVπ(μ)  ≤  11−γ∥dρπ⋆μ∥∞max⁡πˉ (πˉ−π)⊤∇πVπ(μ),V^\star(\rho)-V^\pi(\rho)\;\le\;\Big\|\frac{d^{\pi^\star}_\rho}{d^\pi_\mu}\Big\|_\infty\max_{\bar\pi}\,(\bar\pi-\pi)^\top\nabla_\pi V^\pi(\mu)\;\le\;\frac1{1-\gamma}\Big\|\frac{d^{\pi^\star}_\rho}{\mu}\Big\|_\infty\max_{\bar\pi}\,(\bar\pi-\pi)^\top\nabla_\pi V^\pi(\mu),V⋆(ρ)−Vπ(ρ)≤​dμπ​dρπ⋆​​​∞​πˉmax​(πˉ−π)⊤∇π​Vπ(μ)≤1−γ1​​μdρπ⋆​​​∞​πˉmax​(πˉ−π)⊤∇π​Vπ(μ),

where the maximum is over all policies πˉ∈Δ(A)∣S∣\bar\pi\in\Delta(\mathcal A)^{|\mathcal S|}πˉ∈Δ(A)∣S∣ and the ratios are componentwise.

The lemma says that a policy that is nearly first-order stationary for the objective Vπ(μ)V^\pi(\mu)Vπ(μ) is nearly optimal for every start distribution ρ\rhoρ, provided μ\muμ covers the states visited by an optimal policy. It is the reason projected gradient ascent converges globally despite the non-concavity of VπV^\piVπ.

Formalization Note The maximum is represented by an arbitrary upper bound G≥(πˉ−π)⊤∇πVπ(μ)G\ge(\bar\pi-\pi)^\top\nabla_\pi V^\pi(\mu)G≥(πˉ−π)⊤∇π​Vπ(μ) for all policies πˉ\bar\piπˉ (the maximum is attained on the compact simplex, so this is equivalent). The two mismatch coefficients are represented by arbitrary constants D1D_1D1​, DDD with dρπ⋆(s)≤D1 dμπ(s)d^{\pi^\star}_\rho(s)\le D_1\,d^\pi_\mu(s)dρπ⋆​(s)≤D1​dμπ​(s) and dρπ⋆(s)≤D μ(s)d^{\pi^\star}_\rho(s)\le D\,\mu(s)dρπ⋆​(s)≤Dμ(s) for all sss, which avoids Lean's convention x/0=0x/0=0x/0=0. The conclusion is the two inequalities V⋆(ρ)−Vπ(ρ)≤D1GV^\star(\rho)-V^\pi(\rho)\le D_1GV⋆(ρ)−Vπ(ρ)≤D1​G and V⋆(ρ)−Vπ(ρ)≤11−γDGV^\star(\rho)-V^\pi(\rho)\le\frac{1}{1-\gamma}DGV⋆(ρ)−Vπ(ρ)≤1−γ1​DG. Optimality of π⋆\pi^\starπ⋆ means Vπ′(s)≤Vπ⋆(s)V^{\pi'}(s)\le V^{\pi^\star}(s)Vπ′(s)≤Vπ⋆(s) for every policy π′\pi'π′ and state sss.

Preamble
import Mathlib
import Definitions.Def_FoundationsML_ReinforcementLearning_IsPolicy
import Definitions.Def_FoundationsML_ReinforcementLearning_IsOptimalPolicy
import Definitions.Def_PolicyGradTheory_ProjGA_MDP
import Definitions.Def_PolicyGradTheory_ProjGA_Algorithm

open FoundationsML.ReinforcementLearning
Formal statement
namespace PolicyGradTheory.ProjGA

/-- Lemma 4.1 (gradient domination), arXiv:1908.00261v5, p. 14: for the direct parameterization,
all state distributions `μ, ρ` and every policy `π`,
`V⋆(ρ) − V^π(ρ) ≤ ‖d^{π⋆}_ρ/d^π_μ‖_∞ max_{π̄} (π̄ − π)ᵀ∇_π V^π(μ)
  ≤ (1/(1−γ)) ‖d^{π⋆}_ρ/μ‖_∞ max_{π̄} (π̄ − π)ᵀ∇_π V^π(μ)`, the max over all policies `π̄`.
The max is passed as any upper bound `G`, the two mismatch coefficients as any `D₁`, `D` with
`d^{π⋆}_ρ ≤ D₁ d^π_μ` and `d^{π⋆}_ρ ≤ D μ` componentwise. -/
theorem gradient_domination {S A : Type*} [Fintype S] [DecidableEq S] [Fintype A]
    [DecidableEq A] (P : S → A → S → ℝ) (r : S → A → ℝ) (γ : ℝ) (hM : IsFiniteMDP P r γ)
    (μ : S → ℝ) (hμ : IsDist μ) (ρ : S → ℝ) (hρ : IsDist ρ)
    (πstar : S → A → ℝ) (hπstar : IsPolicy πstar) (hopt : IsOptimalPolicy πstar P r γ)
    (π : EuclideanSpace ℝ (S × A)) (hπ : π ∈ simplexSet S A)
    (G : ℝ) (hG : ∀ πbar ∈ simplexSet S A,
      inner ℝ (πbar - π) (gradient (directValue P r γ μ) π) ≤ G)
    (D₁ : ℝ) (hD₁ : ∀ s, visitation πstar P γ ρ s ≤ D₁ * visitation (asPolicy π) P γ μ s)
    (D : ℝ) (hD : ∀ s, visitation πstar P γ ρ s ≤ D * μ s) :
    valueAt πstar P r γ ρ - valueAt (asPolicy π) P r γ ρ ≤ D₁ * G ∧
      valueAt πstar P r γ ρ - valueAt (asPolicy π) P r γ ρ ≤ 1 / (1 - γ) * D * G := by sorry

end PolicyGradTheory.ProjGA
Source
arXiv:1908.00261v5, Lemma 4.1, p. 14 (proof pp. 14–15)
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