Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Equation (12) — finite true value parameter and consistency

Proved
LeastSquaresTD.Absorbing.value_consistency

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

markov-chainp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1reinforcement-learningtemporal-difference-learning

For an absorbing finite Markov chain with rewards zero between absorbing states, take 0≤γ≤10\le\gamma\le10≤γ≤1. Suppose the non-absorbing feature vectors are linearly independent, absorbing feature vectors vanish, and the feature dimension is the number of non-absorbing states. Then the expected-return series converges at every state, and there is a parameter θ∗\theta^*θ∗ representing its value function such that

V(x)=ϕx⊤θ∗(x∈X),rˉ=(I−γP)Φθ∗.V(x)=\phi_x^\top\theta^*\quad(x\in X),\qquad \bar r=(I-\gamma P)\Phi\theta^*.V(x)=ϕx⊤​θ∗(x∈X),rˉ=(I−γP)Φθ∗.

This supplies the finite true parameter and the consistency relation used to identify the LS TD limit with it.

Formalization Note At γ=1\gamma=1γ=1, summability follows from absorption and the zero-reward condition; it is stated explicitly because a divergent real tsum would otherwise have a default value. The parameter is existential, not defined by the limiting LS TD matrix.

Preamble
import Definitions.Def_LeastSquaresTD_Absorbing_Estimator
Formal statement
namespace LeastSquaresTD.Absorbing

variable {m : ℕ}

/-- Proof of Theorem 1, p. 44, Eq. (12), together with the paper's
claim that `θ*` is finite. The value function is the return series, so
summability at `γ = 1` remains part of this conclusion. -/
theorem value_consistency
    {X : Type*} [Fintype X] [DecidableEq X] [Nonempty X]
    (C : Chain X) (R : X → X → ℝ) (φ : X → Fin m → ℝ) (γ : ℝ)
    (habs : C.IsAbsorbing) (hRabs : ∀ x y, C.P x x = 1 → C.P y y = 1 → R x y = 0)
    (hm : m = Fintype.card C.Nonabsorbing)
    (hli : LinearIndependent ℝ (fun x : C.Nonabsorbing => φ x.val))
    (hφabs : ∀ x, C.P x x = 1 → φ x = 0)
    (hγ0 : 0 ≤ γ) (hγ1 : γ ≤ 1) :
    ∃ θstar : Fin m → ℝ,
      (∀ x, Summable (fun k : ℕ => γ ^ k *
        (Matrix.mulVec (C.P ^ k) (C.rbar R)) x)) ∧
      (∀ x, C.value R γ x = ∑ i, φ x i * θstar i) ∧
      C.rbar R = Matrix.mulVec (1 - γ • C.P)
        (Matrix.mulVec (Matrix.of φ) θstar) := by sorry

end LeastSquaresTD.Absorbing
Source
Bradtke and Barto, Linear Least-Squares Algorithms for Temporal Difference Learning, Machine Learning 22 (1996), https://doi.org/10.1023/A:1018056104778, p. 44, Proof of Theorem 1, Eq. (12) and following sentence; p. 41, Eq. (9)
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