§2, p.7 — the Bellman equation has the unique solution , the discounted series
ProvedMarkovEntanglementExact.bellman_unique_solutionmarkov-chainp2o-gran-per-chapterp2o-plan-paperp2o-strengthenedp2o-v1reinforcement-learningvalue-decomposition
Let be a transition matrix on a finite set, , and a reward vector. Then is invertible, and is the unique solution of the Bellman equation
Moreover
This ties the Bellman solutions used throughout the mission to the printed definition of the Q-value as an expected discounted sum, and shows that a statement quantified over solutions of the Bellman equation is about exactly one vector.
Formalization Note at an index is the expected reward at time started from that index. The series converges in the product topology on vectors.
Differences from the source:
- generalization: P^π_AB of a two-agent MDP and policy → an arbitrary transition matrix P on a product of two finite sets; reason: the argument uses only the stochasticity of P; P^π_AB is an instance by M12
- encoding: Q^π_AB as the expectation of the discounted reward → the sum of the series Σ_t γ^t P^t r; reason: P^t r at an index is the expected reward at time t from that index
Preamble
import Mathlib import Definitions.Def_markov_entanglement open scoped BigOperators open MarkovEntanglement
Formal statement
namespace MarkovEntanglementExact
/-- The Bellman equation at discount `0 ≤ γ < 1` for a transition matrix `P`: `I − γP` is
invertible, `Q = (I − γP)⁻¹ r` solves `Q = r + γPQ`, it is the only solution, and it is the sum
of the series `Σ_t γ^t P^t r`, which is the printed definition of the Q-value.
Source: Chen & Peng, Multi-agent Markov Entanglement, arXiv:2506.02385v3, §2, p.7.
**Formalization Note.** `P ^ t *ᵥ r` at an index is the expected reward at time `t` from that
index, so `Σ_t γ^t P^t r` is the expectation in the printed `Q^π_AB`. At `γ = 0` the series is
`r` alone, with `0 ^ 0 = 1`. -/
theorem bellman_unique_solution
{ι : Type*} [Fintype ι] [DecidableEq ι]
(P : Matrix ι ι ℝ) (hP : IsTransitionMatrix P)
(γ : ℝ) (hγ0 : 0 ≤ γ) (hγ1 : γ < 1) (r : ι → ℝ) :
IsUnit (1 - γ • P) ∧
IsBellmanQ P r γ ((1 - γ • P)⁻¹.mulVec r) ∧
(∀ Q : ι → ℝ, IsBellmanQ P r γ Q → Q = (1 - γ • P)⁻¹.mulVec r) ∧
HasSum (fun t : ℕ => γ ^ t • (P ^ t).mulVec r) ((1 - γ • P)⁻¹.mulVec r) := by sorry
end MarkovEntanglementExact
Source
Chen & Peng, Multi-agent Markov Entanglement, arXiv:2506.02385v3, §2, p.7
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.