Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Finite-sum form of the bias bound for the reward-corrected estimator, valid for every γ≥0\gamma\ge0γ≥0

Proved
TPG.rc_bias_finite_sum

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

causal-inferencemarkov-chainnonstationaritypolicy-gradienttreatment-effect

Setting. SSS is a finite state space and the action Zt∈{0,1}Z_t\in\{0,1\}Zt​∈{0,1} is control or treatment. Time is t=0,1,…,T−1t=0,1,\dots,T-1t=0,1,…,T−1, which is the source's time t+1t+1t+1. Under action zzz at time ttt the state moves according to a transition matrix PztP_z^tPzt​, i.e. P(Xt+1=x′∣Xt=x,Zt=z)=Pzt(x,x′)\mathbb P(X_{t+1}=x'\mid X_t=x,Z_t=z)=P_z^t(x,x')P(Xt+1​=x′∣Xt​=x,Zt​=z)=Pzt​(x,x′); the matrices may depend on ttt arbitrarily. The initial state X0X_0X0​ has distribution ρ\rhoρ, and the reward YtY_tYt​ has distribution R(⋅∣Xt,Zt)R(\cdot\mid X_t,Z_t)R(⋅∣Xt​,Zt​) on R\mathbb RR, with R([−M,M]∣x,z)=1R([-M,M]\mid x,z)=1R([−M,M]∣x,z)=1 for all x,zx,zx,z and conditional independence from everything else given (Xt,Zt)(X_t,Z_t)(Xt​,Zt​). Under the policy πθ\pi_\thetaπθ​ every action is treatment with probability θ\thetaθ, independently of everything else; Eθ\mathbb E_\thetaEθ​, Pθ\mathbb P_\thetaPθ​ and Var⁡θ\operatorname{Var}_\thetaVarθ​ refer to the law of the trajectory (Xt,Zt,Yt)t<T(X_t,Z_t,Y_t)_{t<T}(Xt​,Zt​,Yt​)t<T​ under πθ\pi_\thetaπθ​.

Assumption 1 (mixing time) with a constant γ≥0\gamma\ge0γ≥0 (not assumed smaller than one): for every time t<Tt<Tt<T, every action zzz and every pair of probability vectors f,f′f,f'f,f′ on SSS, ∥f′Pzt−fPzt∥TV≤γ ∥f′−f∥TV\|f'P_z^t-fP_z^t\|_{\mathrm{TV}}\le\gamma\,\|f'-f\|_{\mathrm{TV}}∥f′Pzt​−fPzt​∥TV​≤γ∥f′−f∥TV​, where ∥f−f′∥TV=12∥f−f′∥1\|f-f'\|_{\mathrm{TV}}=\tfrac12\|f-f'\|_1∥f−f′∥TV​=21​∥f−f′∥1​. The kernel deviation is δ=sup⁡t<Tsup⁡x∈S∥P1t(x,⋅)−P0t(x,⋅)∥TV\delta=\sup_{t<T}\sup_{x\in S}\|P_1^t(x,\cdot)-P_0^t(x,\cdot)\|_{\mathrm{TV}}δ=supt<T​supx∈S​∥P1t​(x,⋅)−P0t​(x,⋅)∥TV​.

The global average treatment effect is τ=1T∑t<T(E1[Yt]−E0[Yt])\tau=\frac1T\sum_{t<T}\bigl(\mathbb E_1[Y_t]-\mathbb E_0[Y_t]\bigr)τ=T1​∑t<T​(E1​[Yt​]−E0​[Yt​]).

Fix θ∈(0,1)\theta\in(0,1)θ∈(0,1) and let τ^θ,kRC=1T∑u<Tsθ(Zu)[Yu+∑t=u+1min⁡(u+k,T−1)wθ(Zt)Yt]\widehat\tau^{\mathrm{RC}}_{\theta,k}=\frac1T\sum_{u<T}s_\theta(Z_u)\bigl[Y_u+\sum_{t=u+1}^{\min(u+k,T-1)}w_\theta(Z_t)Y_t\bigr]τθ,kRC​=T1​∑u<T​sθ​(Zu​)[Yu​+∑t=u+1min(u+k,T−1)​wθ​(Zt​)Yt​] be the reward-corrected estimator, with sθ(z)=zθ−1−z1−θs_\theta(z)=\frac z\theta-\frac{1-z}{1-\theta}sθ​(z)=θz​−1−θ1−z​ and wθ(z)=(1−θ)zθ+θ(1−z)1−θw_\theta(z)=\frac{(1-\theta)z}{\theta}+\frac{\theta(1-z)}{1-\theta}wθ​(z)=θ(1−θ)z​+1−θθ(1−z)​, computed under πθ\pi_\thetaπθ​. Write Gn(γ)=∑j<nγjG_n(\gamma)=\sum_{j<n}\gamma^jGn​(γ)=∑j<n​γj and Ψn(γ)=∑j=0n−2(j+1)γj\Psi_n(\gamma)=\sum_{j=0}^{n-2}(j+1)\gamma^jΨn​(γ)=∑j=0n−2​(j+1)γj (so Ψ0=Ψ1=0\Psi_0=\Psi_1=0Ψ0​=Ψ1​=0), and for a reward time ttt let st=max⁡(0,t−k)s_t=\max(0,t-k)st​=max(0,t−k) be the start of its truncation window and ℓt=min⁡(t,k)=t−st\ell_t=\min(t,k)=t-s_tℓt​=min(t,k)=t−st​ its length.

Lemma. For every T≥1T\ge1T≥1 and 0≤k≤T0\le k\le T0≤k≤T,

∣Eθ[τ^θ,kRC]−τ∣≤1T∑t<T(2δM γℓt Gst(γ)+4[θ2+(1−θ)2]δ2M Ψℓt(γ)).\Bigl|\mathbb E_\theta\bigl[\widehat\tau^{\mathrm{RC}}_{\theta,k}\bigr]-\tau\Bigr|\le\frac1T\sum_{t<T}\Bigl(2\delta M\,\gamma^{\ell_t}\,G_{s_t}(\gamma)+4\bigl[\theta^2+(1-\theta)^2\bigr]\delta^2M\,\Psi_{\ell_t}(\gamma)\Bigr).​Eθ​[τθ,kRC​]−τ​≤T1​t<T∑​(2δMγℓt​Gst​​(γ)+4[θ2+(1−θ)2]δ2MΨℓt​​(γ)).

Compared with the TPG estimator there is no first-order term in δ\deltaδ. For γ<1\gamma<1γ<1 the bound implies the bias part of Proposition 6.

Formalization Note t−kt-kt−k is a truncated subtraction of natural numbers.

Preamble
import Mathlib
import Definitions.Def_TPG_Model
import Definitions.Def_TPG_Estimators

open MeasureTheory ProbabilityTheory
open scoped BigOperators ENNReal
Formal statement
namespace TPG
theorem rc_bias_finite_sum {S : Type} [Fintype S] [MeasurableSpace S] [MeasurableSingletonClass S]
    (E : Env S) (T k : ℕ) (θ γ M : ℝ)
    (hT : 0 < T) (hk : k ≤ T) (hθ0 : 0 < θ) (hθ1 : θ < 1) (hγ0 : 0 ≤ γ)
    (hmix : E.Mixing T γ) (hM : E.RewardBound M) :
    |(∫ ω, rcEst θ T k ω ∂(E.expLaw T θ)) - E.gate T|
      ≤ (1 / (T : ℝ)) * ∑ t : Fin T,
          (2 * E.delta T * M * γ ^ (min (t : ℕ) k) * (∑ j ∈ Finset.range ((t : ℕ) - k), γ ^ j)
            + 4 * (θ ^ 2 + (1 - θ) ^ 2) * (E.delta T) ^ 2 * M * Psi (min (t : ℕ) k) γ) := 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 is the estimate obtained in Appendix G.6, pp. 48–51 (proof of Proposition 6) before the geometric sums are bounded by their limits.

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