Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 4.2.4 — an extreme optimal solution of the multichain LP yields a pure stationary average optimal policy

Proved
KallenbergLP.AverageLP.theorem_4_2_4

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

average-rewarddynamic-programminglinear-programmingmarkov-decision-processp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-bookp2o-v1

Consider a finite Markov decision model with state set EEE, finite action sets A(i)A(i)A(i), rewards riar_{ia}ria​ and transition probabilities with ∑jpiaj=1\sum_jp_{iaj}=1∑j​piaj​=1, with no assumption on the chain structure. Fix βj>0\beta_j>0βj​>0 with ∑jβj=1\sum_j\beta_j=1∑j​βj​=1 and consider the linear program (4.2.11)

max⁡{∑i∑ariaxia ∣ ∑i∑a(δij−piaj)xia=0,  ∑axja+∑i∑a(δij−piaj)yia=βj (j∈E); x,y≥0}.\max\Big\{\sum_i\sum_ar_{ia}x_{ia}\ \Big|\ \sum_i\sum_a(\delta_{ij}-p_{iaj})x_{ia}=0,\ \ \sum_ax_{ja}+\sum_i\sum_a(\delta_{ij}-p_{iaj})y_{ia}=\beta_j\ (j\in E);\ x,y\ge0\Big\}.max{i∑​a∑​ria​xia​ ​ i∑​a∑​(δij​−piaj​)xia​=0,  a∑​xja​+i∑​a∑​(δij​−piaj​)yia​=βj​ (j∈E); x,y≥0}.

Let (x∗,y∗)(x^*,y^*)(x∗,y∗) be an optimal solution that is an extreme point of the set of feasible solutions, and let Ex∗={i∣∑axia∗>0}E_{x^*}=\{i\mid\sum_ax^*_{ia}>0\}Ex∗​={i∣∑a​xia∗​>0}. Then a decision rule f∗f_*f∗​ with

xif∗(i)∗>0  (i∈Ex∗),yif∗(i)∗>0  (i∈E∖Ex∗)x^*_{if_*(i)}>0\ \ (i\in E_{x^*}),\qquad y^*_{if_*(i)}>0\ \ (i\in E\setminus E_{x^*})xif∗​(i)∗​>0  (i∈Ex∗​),yif∗​(i)∗​>0  (i∈E∖Ex∗​)

exists, and for every such f∗f_*f∗​ the pure stationary policy f∗∞f_*^\inftyf∗∞​ is average optimal: ϕi(f∗∞)=sup⁡Rϕi(R)\phi_i(f_*^\infty)=\sup_R\phi_i(R)ϕi​(f∗∞​)=supR​ϕi​(R) for every i∈Ei\in Ei∈E, the supremum running over all policies.

Since the simplex method returns an extreme optimal solution, this turns one linear program into an average optimal policy for a general multichain model.

Formalization Note Every f∗f_*f∗​ obeying the rule is covered (Remark 4.2.4: an arbitrary positive variable may be chosen). The average reward is the lim inf criterion.

Preamble
import Mathlib
import Definitions.Def_MarkovDecisionProcesses_AverageReward
import Definitions.Def_BlackwellDiscreteDP_NearOne_LimitMatrix
import Definitions.Def_KallenbergLP_AverageLP_Model
open MarkovDecisionProcesses BlackwellDiscreteDP.NearOne Matrix Filter
Formal statement
namespace KallenbergLP.AverageLP

variable {S A : Type*} [Fintype S] [Fintype A] [DecidableEq S] [DecidableEq A]

/-- **Theorem 4.2.4.** If `(x*, y*)` is an optimal solution of the linear program (4.2.11) such that
`(x*, y*)` is an extreme point of the set of feasible solutions, then the policy `f_*^∞`, where
`f_*(i) := a_i` such that `x*_{i a_i} > 0` for `i ∈ E_{x*}` and `y*_{i a_i} > 0` for
`i ∈ E ∖ E_{x*}`, is an average optimal policy.

Kallenberg, Linear Programming and Finite Markovian Control Problems, Mathematical Centre Tracts 148, Mathematisch Centrum, Amsterdam 1983, p. 103, Theorem 4.2.4; `β_j > 0`, `Σ_j β_j = 1` from (4.2.10), p. 102.

**Formalization Note.** The conclusion has two parts: such an `f_*` exists (the book's "hence
the policy `f_*^∞` is well-defined", p. 103), and *every* `f_*` obeying the rule gives an average
optimal policy (Remark 4.2.4). Average optimality is `φ(f_*^∞) = φ` against all
history-dependent randomized policies. No unichain or ergodicity assumption is made. -/
theorem theorem_4_2_4 (M : StationaryMDP S A)
    (β : S → ℝ) (hβpos : ∀ j, 0 < β j) (hβsum : ∑ j, β j = 1)
    (z : (Pair M → ℝ) × (Pair M → ℝ)) (hext : z ∈ Set.extremePoints ℝ (dualFeasible M β))
    (hopt : IsDualOptimal M β z) :
    (∃ (f : S → A) (hf : ∀ i, f i ∈ M.admissible i), IsSelection M z.1 z.2 f hf) ∧
    ∀ (f : S → A) (hf : ∀ i, f i ∈ M.admissible i), IsSelection M z.1 z.2 f hf →
      IsAvgOptimal M (stationaryPolicy M f hf) := by sorry

end KallenbergLP.AverageLP
Source
Kallenberg, Linear Programming and Finite Markovian Control Problems, Mathematical Centre Tracts 148, Mathematisch Centrum, Amsterdam 1983, p. 103, Theorem 4.2.4; (4.2.10)–(4.2.11), p. 102
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