Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Markov property of the trajectory law in product form

Proved
TPG.integral_head_mul_tail

by Wenqian Xing · Oct 12, 2026 · Mathlib 0df444a (Lean v4.33.1)

causal-inferencemarkov-chainnonstationaritypolicy-gradienttreatment-effect

Setting. SSS is a finite state space, actions are Zt∈{0,1}Z_t\in\{0,1\}Zt​∈{0,1}, time is t=0,1,…,T−1t=0,1,\dots,T-1t=0,1,…,T−1 (the source's time t+1t+1t+1). The environment has transition matrices PztP_z^tPzt​, initial distribution ρ\rhoρ and reward laws R(⋅∣x,z)R(\cdot\mid x,z)R(⋅∣x,z) on R\mathbb RR. For a sequence q=(qt)t≥0q=(q_t)_{t\ge0}q=(qt​)t≥0​ of treatment probabilities, Pq\mathbb P_qPq​ and Eq\mathbb E_qEq​ denote the law of the trajectory (Xt,Zt,Yt)t<T(X_t,Z_t,Y_t)_{t<T}(Xt​,Zt​,Yt​)t<T​ when X0∼ρX_0\sim\rhoX0​∼ρ, the actions are independent with P(Zt=1)=qt\mathbb P(Z_t=1)=q_tP(Zt​=1)=qt​, Xt+1∼PZtt(Xt,⋅)X_{t+1}\sim P^t_{Z_t}(X_t,\cdot)Xt+1​∼PZt​t​(Xt​,⋅), and the rewards are conditionally independent with Yt∼R(⋅∣Xt,Zt)Y_t\sim R(\cdot\mid X_t,Z_t)Yt​∼R(⋅∣Xt​,Zt​) (Env.trajLaw T q); the Bernoulli policy πθ\pi_\thetaπθ​ is qt≡θq_t\equiv\thetaqt​≡θ.

Consider the horizon m+1+nm+1+nm+1+n. The head of a trajectory is its first m+1m+1m+1 time steps 0,…,m0,\dots,m0,…,m, and the tail is its last nnn steps m+1,…,m+nm+1,\dots,m+nm+1,…,m+n, re-indexed from 000. For a state aaa, let E(m+1,a)\mathbb E^{(m+1,a)}E(m+1,a) denote expectation for the process restarted at time m+1m+1m+1 from the state aaa: kernels Pzm+1+tP^{m+1+t}_zPzm+1+t​, initial state aaa, treatment probabilities qm+1+tq_{m+1+t}qm+1+t​, the same reward laws.

Lemma. Let 0≤qt≤10\le q_t\le10≤qt​≤1 for all ttt, and let FFF (a function of a trajectory of length m+1m+1m+1) and GGG (a function of a trajectory of length nnn) be bounded and measurable. Then

Eq[F(head) G(tail)]=∑a∈S Eq[F⋅PZmm(Xm,a)] E(m+1,a)[G],\mathbb E_q\bigl[F(\text{head})\,G(\text{tail})\bigr]=\sum_{a\in S}\ \mathbb E_q\bigl[F\cdot P^{m}_{Z_m}(X_m,a)\bigr]\ \mathbb E^{(m+1,a)}[G],Eq​[F(head)G(tail)]=a∈S∑​ Eq​[F⋅PZm​m​(Xm​,a)] E(m+1,a)[G],

where the first expectation on the right is taken for the horizon m+1m+1m+1.

In words: given the head, the state at time m+1m+1m+1 is drawn from the row XmX_mXm​ of PZmmP^m_{Z_m}PZm​m​, and given that state the tail is an independent copy of the restarted process. Covariance decay, strong mixing and the martingale arguments of the mission all rest on this identity.

Preamble
import Mathlib
import Definitions.Def_TPG_Model
import Definitions.Def_TPG_Markov

open MeasureTheory ProbabilityTheory
open scoped BigOperators ENNReal
Formal statement
namespace TPG
theorem integral_head_mul_tail {S : Type} [Fintype S] [DecidableEq S] [MeasurableSpace S] [MeasurableSingletonClass S]
    (E : Env S) (m n : ℕ) (q : ℕ → ℝ) (hq : ∀ t : ℕ, 0 ≤ q t ∧ q t ≤ 1)
    (F : Traj S (m + 1) → ℝ) (G : Traj S n → ℝ) (hF : Measurable F) (hG : Measurable G)
    (CF CG : ℝ) (hFb : ∀ ω, |F ω| ≤ CF) (hGb : ∀ ω, |G ω| ≤ CG) :
    ∫ ω, F (headTraj (m + 1) n ω) * G (tailTraj (m + 1) n ω) ∂(E.trajLaw (m + 1 + n) q)
      = ∑ a : S,
          (∫ ω, F ω * E.P m (Z (Fin.last m) ω) (X (Fin.last m) ω) a ∂(E.trajLaw (m + 1) q))
            * ∫ ω, G ω ∂((E.shift (m + 1) a).trajLaw n (fun t => q (m + 1 + t))) := by sorry
end TPG
Source
Johari, Peng, Xing, Estimation of Treatment Effects Under Nonstationarity via the Truncated Policy Gradient Estimator, revised manuscript of October 2026 (revision of arXiv:2506.05308v3, https://arxiv.org/abs/2506.05308); numbering follows the revised manuscript. Auxiliary statement, not a numbered result of the paper: it formalizes the Markov property of §2.2, pp. 8–9, in the form used in Appendix F.1, p. 40 (E[B_j | F_i] = (P^{(i+1)→a_j} g_j)(X_{i+1})) and in the proof of Lemma 2, p. 28.

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