§3.2, p.10 — the marginal identity; is a transition matrix
ProvedMarkovEntanglementExact.marginal_identitymarkov-chainp2o-gran-per-chapterp2o-plan-paperp2o-strengthenedp2o-v1reinforcement-learningvalue-decomposition
Let be a transition matrix on with , as in Definition 1, where are nonempty. Then for all ,
so the left side does not depend on . The symmetric identity holds for agent . In particular, and are transition matrices.
The second conclusion is not automatic, since the may be negative: the nonnegativity of the sum comes from that of . The identity is also a test for entanglement, since a row marginal of that depends on rules out separability.
Formalization Note Nonemptiness of the two sets excludes only the case of an empty factor, where is the empty matrix and gives no information on the signed sum of the other factor.
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
- generalization: the identity appears as the middle of a computation of P^π_A, with μ summed out → the row-marginal identity at every b, plus the conclusion that both signed sums are transition matrices; reason: it implies the printed computation, and the transition-matrix conclusion follows from P ≥ 0
- encoding: state-action spaces of an MDP, implicitly nonempty → [Nonempty SA] [Nonempty SB]; reason: this only excludes an empty factor, for which the signed sum of the other factor can have negative entries
Preamble
import Mathlib import Definitions.Def_markov_entanglement import Definitions.Def_MarkovEntanglementExact_IsSeparableWitness open scoped BigOperators open MarkovEntanglement
Formal statement
namespace MarkovEntanglementExact
/-- The marginal identity: if a transition matrix `P` decomposes as
`Σ_j x_j P^(j)_A ⊗ P^(j)_B`, then at every `(a, b)` the row of `P` summed over `b'` is
`Σ_j x_j P^(j)_A(a, ·)`, not depending on `b`, and symmetrically for agent `B`. Hence
`Σ_j x_j P^(j)_A` and `Σ_j x_j P^(j)_B` are transition matrices, even though the `x_j` are signed.
Source: Chen & Peng, Multi-agent Markov Entanglement, arXiv:2506.02385v3, §3.2, p.10.
**Formalization Note.** The state-action spaces are nonempty. This excludes only the degenerate
case of an empty factor, where `P` is the empty matrix and says nothing about the other factor's
signed sum. -/
theorem marginal_identity
{SA SB : Type*} [Fintype SA] [Fintype SB] [DecidableEq SA] [DecidableEq SB]
[Nonempty SA] [Nonempty SB]
(P : Matrix (SA × SB) (SA × SB) ℝ) (hP : IsTransitionMatrix P)
{K : ℕ} {x : Fin K → ℝ} {PAj : Fin K → Matrix SA SA ℝ} {PBj : Fin K → Matrix SB SB ℝ}
(hw : IsSeparableWitness P K x PAj PBj) :
(∀ a a' b, ∑ b', P (a, b) (a', b') = (∑ k, x k • PAj k) a a') ∧
(∀ a b b', ∑ a', P (a, b) (a', b') = (∑ k, x k • PBj k) b b') ∧
IsTransitionMatrix (∑ k, x k • PAj k) ∧
IsTransitionMatrix (∑ k, x k • PBj k) := by sorry
end MarkovEntanglementExact
Source
Chen & Peng, Multi-agent Markov Entanglement, arXiv:2506.02385v3, §3.2, p.10
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.