Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 1 — bias and variance bounds for the TPG estimator

Proved
TPG.theorem_1_bias_variance_bounds

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 constant γ∈(0,1)\gamma\in(0,1)γ∈(0,1): 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​]).

Let τ^k\widehat\tau_kτk​ be the truncated policy gradient (TPG) estimator with truncation size kkk, computed from one trajectory generated under the experimental policy π1/2\pi_{1/2}π1/2​:

τ^k=1T∑u=0T−1(1{Zu=1}1/2−1{Zu=0}1/2)∑t=umin⁡(u+k, T−1)Yt.\widehat\tau_k=\frac1T\sum_{u=0}^{T-1}\Bigl(\frac{\mathbf 1\{Z_u=1\}}{1/2}-\frac{\mathbf 1\{Z_u=0\}}{1/2}\Bigr)\sum_{t=u}^{\min(u+k,\,T-1)}Y_t .τk​=T1​u=0∑T−1​(1/21{Zu​=1}​−1/21{Zu​=0}​)t=u∑min(u+k,T−1)​Yt​.

Theorem. There are absolute constants C1,C2>0C_1,C_2>0C1​,C2​>0, not depending on the state space, the environment, or T,k,γ,MT,k,\gamma,MT,k,γ,M, such that under Assumption 1, for every T≥1T\ge1T≥1 and every truncation size 0≤k≤T0\le k\le T0≤k≤T,

∣E1/2[τ^k]−τ∣≤C1(1−γk(1−γ)2 δ2M+γk1−γ δM),\bigl|\mathbb E_{1/2}[\widehat\tau_k]-\tau\bigr|\le C_1\Bigl(\frac{1-\gamma^k}{(1-\gamma)^2}\,\delta^2M+\frac{\gamma^k}{1-\gamma}\,\delta M\Bigr),​E1/2​[τk​]−τ​≤C1​((1−γ)21−γk​δ2M+1−γγk​δM), Var⁡1/2(τ^k)≤C2((k+1)2M2T+γ (k+1)M2T(1−γ)).\operatorname{Var}_{1/2}(\widehat\tau_k)\le C_2\Bigl(\frac{(k+1)^2M^2}{T}+\frac{\gamma\,(k+1)M^2}{T(1-\gamma)}\Bigr).Var1/2​(τk​)≤C2​(T(k+1)2M2​+T(1−γ)γ(k+1)M2​).

The first term of the bias bound is the Taylor error of approximating a difference of policy values by a derivative, and the second is the mixing bias due to truncation. This is the main result of the source; it quantifies how the truncation size trades bias against variance.

Formalization Note The source states both bounds with O(⋅)O(\cdot)O(⋅). The formal statement renders this with two existentially quantified constants chosen before the state space, the environment and all parameters.

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 theorem_1_bias_variance_bounds :
    ∃ C₁ C₂ : ℝ, 0 < C₁ ∧ 0 < C₂ ∧
      ∀ (S : Type) [Fintype S] [MeasurableSpace S] [MeasurableSingletonClass S]
        (E : Env S) (T k : ℕ) (γ M : ℝ),
        0 < T → k ≤ T → 0 < γ → γ < 1 → E.Mixing T γ → E.RewardBound M →
          |(∫ ω, tpgEst (1 / 2) T k ω ∂(E.expLaw T (1 / 2))) - E.gate T|
              ≤ C₁ * ((1 - γ ^ k) / (1 - γ) ^ 2 * (E.delta T) ^ 2 * M
                      + γ ^ k / (1 - γ) * E.delta T * M)
            ∧ variance (tpgEst (1 / 2) T k) (E.expLaw T (1 / 2))
              ≤ C₂ * (((k : ℝ) + 1) ^ 2 * M ^ 2 / (T : ℝ)
                      + γ * ((k : ℝ) + 1) * M ^ 2 / ((T : ℝ) * (1 - 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. §3, pp. 11–12, Theorem 1, eqs. (3)–(5); proof in Appendix F.1, pp. 37–41.
Read-back

What the Lean code literally says, in plain math · claude-opus-5-5

Read-back of TPG.theorem_1_bias_variance_bounds.

Logical form. The statement asserts: there exist two real numbers C1C_1C1​ and C2C_2C2​, with C1>0C_1>0C1​>0 and C2>0C_2>0C2​>0, such that for all of the data listed below, if the six hypotheses (H1)–(H6) hold, then both inequalities (I) and (II) hold. The two constants are chosen first and everything else is universally quantified after them, so C1C_1C1​ and C2C_2C2​ are absolute: they may not depend on the set SSS, on the environment EEE, or on TTT, kkk, γ\gammaγ, MMM.

Universally quantified data, in this order.

  1. A set SSS (a type in the lowest universe) which is finite, equipped with a σ\sigmaσ-algebra in which every singleton {x}\{x\}{x} is measurable. (As SSS is finite, this forces every subset of SSS to be measurable.) Non-emptiness of SSS is not assumed as a hypothesis.

  2. An environment EEE on SSS. This is a package consisting of:

    • for every natural number t∈Nt\in\mathbb{N}t∈N and every Boolean value z∈{1,0}z\in\{1,0\}z∈{1,0} (here and below 111 = true, 000 = false), a real matrix Pt,z=(Pt,z(x,x′))x,x′∈SP_{t,z}=(P_{t,z}(x,x'))_{x,x'\in S}Pt,z​=(Pt,z​(x,x′))x,x′∈S​ with all entries ≥0\ge 0≥0 and every row summing to 111: ∑x′∈SPt,z(x,x′)=1\sum_{x'\in S}P_{t,z}(x,x')=1∑x′∈S​Pt,z​(x,x′)=1 for all t,z,xt,z,xt,z,x;
    • a function ρ:S→R\rho:S\to\mathbb{R}ρ:S→R with ρ(x)≥0\rho(x)\ge 0ρ(x)≥0 for all xxx and ∑x∈Sρ(x)=1\sum_{x\in S}\rho(x)=1∑x∈S​ρ(x)=1;
    • for every x∈Sx\in Sx∈S and every Boolean zzz, a Borel probability measure Rx,zR_{x,z}Rx,z​ on R\mathbb{R}R.
  3. Natural numbers TTT and kkk.

  4. Real numbers γ\gammaγ and MMM.

Hypotheses.

  • (H1) 0<T0<T0<T.
  • (H2) k≤Tk\le Tk≤T.
  • (H3) 0<γ0<\gamma0<γ.
  • (H4) γ<1\gamma<1γ<1.
  • (H5) ("mixing with parameter γ\gammaγ up to TTT") For every natural number ttt with t<Tt<Tt<T, every Boolean zzz, and every pair of functions f,f′:S→Rf,f':S\to\mathbb{R}f,f′:S→R that are both probability vectors (all values ≥0\ge 0≥0 and values summing to 111),
12∑x′∈S∣∑x∈Sf′(x) Pt,z(x,x′)−∑x∈Sf(x) Pt,z(x,x′)∣  ≤  γ⋅12∑x∈S∣f′(x)−f(x)∣.\frac12\sum_{x'\in S}\Bigl|\sum_{x\in S}f'(x)\,P_{t,z}(x,x')-\sum_{x\in S}f(x)\,P_{t,z}(x,x')\Bigr|\;\le\;\gamma\cdot\frac12\sum_{x\in S}\bigl|f'(x)-f(x)\bigr| .21​x′∈S∑​​x∈S∑​f′(x)Pt,z​(x,x′)−x∈S∑​f(x)Pt,z​(x,x′)​≤γ⋅21​x∈S∑​​f′(x)−f(x)​.
  • (H6) ("reward bound MMM") For every x∈Sx\in Sx∈S and every Boolean zzz: Rx,z([−M,M])=1R_{x,z}\bigl([-M,M]\bigr)=1Rx,z​([−M,M])=1.

Objects used in the conclusion.

Trajectory space. ΩT=(S×{1,0}×R)T\Omega_T=(S\times\{1,0\}\times\mathbb{R})^{T}ΩT​=(S×{1,0}×R)T, the set of sequences ω=(ωt)t=0T−1\omega=(\omega_t)_{t=0}^{T-1}ω=(ωt​)t=0T−1​ indexed by the times t∈{0,1,…,T−1}t\in\{0,1,\dots,T-1\}t∈{0,1,…,T−1}, with the product σ\sigmaσ-algebra (the given one on SSS, the discrete one on {1,0}\{1,0\}{1,0}, the Borel one on R\mathbb{R}R). The coordinate maps are written Xt(ω)∈SX_t(\omega)\in SXt​(ω)∈S, Zt(ω)∈{1,0}Z_t(\omega)\in\{1,0\}Zt​(ω)∈{1,0}, Yt(ω)∈RY_t(\omega)\in\mathbb{R}Yt​(ω)∈R for the three components of ωt\omega_tωt​.

Trajectory law with constant parameter θ∈R\theta\in\mathbb{R}θ∈R. For a state sequence x=(x0,…,xT−1)∈STx=(x_0,\dots,x_{T-1})\in S^Tx=(x0​,…,xT−1​)∈ST and a Boolean sequence z=(z0,…,zT−1)∈{1,0}Tz=(z_0,\dots,z_{T-1})\in\{1,0\}^Tz=(z0​,…,zT−1​)∈{1,0}T define the real weight

wθ(x,z)  =  ρ(x0)  ⋅  ∏s=0T−2Ps,zs(xs,xs+1)  ⋅  ∏t=0T−1{θ,zt=1,1−θ,zt=0.w_\theta(x,z)\;=\;\rho(x_0)\;\cdot\;\prod_{s=0}^{T-2}P_{s,z_s}(x_s,x_{s+1})\;\cdot\;\prod_{t=0}^{T-1}\begin{cases}\theta,& z_t=1,\\ 1-\theta,& z_t=0.\end{cases}wθ​(x,z)=ρ(x0​)⋅s=0∏T−2​Ps,zs​​(xs​,xs+1​)⋅t=0∏T−1​{θ,1−θ,​zt​=1,zt​=0.​

(The code writes this as three products over all time indices with indicator conditions "t=0t=0t=0" and "s+1=ts+1=ts+1=t"; for T≥1T\ge 1T≥1 it equals the displayed expression. The transition from time sss to time s+1s+1s+1 uses the matrix with time index sss and the Boolean zsz_szs​.) The measure on ΩT\Omega_TΩT​ is the finite mixture

PT,θ  =  ∑x∈ST ∑z∈{1,0}T max⁡{wθ(x,z), 0}  ⋅  ⨂t=0T−1(Dirac(xt)⊗Dirac(zt)⊗Rxt,zt),\mathbb{P}_{T,\theta}\;=\;\sum_{x\in S^T}\ \sum_{z\in\{1,0\}^T}\ \max\{w_\theta(x,z),\,0\}\;\cdot\;\bigotimes_{t=0}^{T-1}\Bigl(\mathrm{Dirac}(x_t)\otimes\mathrm{Dirac}(z_t)\otimes R_{x_t,z_t}\Bigr),PT,θ​=x∈ST∑​ z∈{1,0}T∑​ max{wθ​(x,z),0}⋅t=0⨂T−1​(Dirac(xt​)⊗Dirac(zt​)⊗Rxt​,zt​​),

where Dirac(a)\mathrm{Dirac}(a)Dirac(a) is the unit point mass at aaa and ⨂\bigotimes⨂ is the product measure over the TTT time coordinates. A negative weight is replaced by 000. For θ∈[0,1]\theta\in[0,1]θ∈[0,1] (the only values used in this theorem are θ=12,1,0\theta=\tfrac12,1,0θ=21​,1,0) all weights are non-negative and sum to 111, and PT,θ\mathbb{P}_{T,\theta}PT,θ​ is the probability law under which: X0X_0X0​ has distribution ρ\rhoρ; Z0,…,ZT−1Z_0,\dots,Z_{T-1}Z0​,…,ZT−1​ are independent of X0X_0X0​ and i.i.d. with Pr⁡(Zt=1)=θ\Pr(Z_t=1)=\thetaPr(Zt​=1)=θ; given the past, Xs+1X_{s+1}Xs+1​ is drawn from row XsX_sXs​ of Ps,ZsP_{s,Z_s}Ps,Zs​​; and, given the whole sequences (Xt)(X_t)(Xt​) and (Zt)(Z_t)(Zt​), the YtY_tYt​ are independent with Yt∼RXt,ZtY_t\sim R_{X_t,Z_t}Yt​∼RXt​,Zt​​. For θ=12\theta=\tfrac12θ=21​ the weight is 2−Tρ(x0)∏s=0T−2Ps,zs(xs,xs+1)2^{-T}\rho(x_0)\prod_{s=0}^{T-2}P_{s,z_s}(x_s,x_{s+1})2−Tρ(x0​)∏s=0T−2​Ps,zs​​(xs​,xs+1​). For θ=1\theta=1θ=1 the weight is 000 unless every zt=1z_t=1zt​=1; for θ=0\theta=0θ=0 it is 000 unless every zt=0z_t=0zt​=0.

The estimator. For ω∈ΩT\omega\in\Omega_Tω∈ΩT​,

G^T,k(ω)  =  1T∑u=0T−1σ(Zu(ω))∑t∈{0,…,T−1}u≤t≤u+kYt(ω),σ(1)=11/2=2,σ(0)=−11−1/2=−2.\widehat G_{T,k}(\omega)\;=\;\frac1T\sum_{u=0}^{T-1}\sigma\bigl(Z_u(\omega)\bigr)\sum_{\substack{t\in\{0,\dots,T-1\}\\ u\le t\le u+k}}Y_t(\omega),\qquad \sigma(1)=\frac{1}{1/2}=2,\quad \sigma(0)=-\frac{1}{1-1/2}=-2 .GT,k​(ω)=T1​u=0∑T−1​σ(Zu​(ω))t∈{0,…,T−1}u≤t≤u+k​∑​Yt​(ω),σ(1)=1/21​=2,σ(0)=−1−1/21​=−2.

The inner sum runs over the times ttt from uuu to min⁡(u+k, T−1)\min(u+k,\,T-1)min(u+k,T−1).

The comparison quantity (called gate in the code).

GT(E)  =  1T∑t=0T−1(∫ΩTYt dPT,1  −  ∫ΩTYt dPT,0).G_T(E)\;=\;\frac1T\sum_{t=0}^{T-1}\Bigl(\int_{\Omega_T}Y_t\,d\mathbb{P}_{T,1}\;-\;\int_{\Omega_T}Y_t\,d\mathbb{P}_{T,0}\Bigr).GT​(E)=T1​t=0∑T−1​(∫ΩT​​Yt​dPT,1​−∫ΩT​​Yt​dPT,0​).

The quantity δT(E)\delta_T(E)δT​(E) (called delta in the code).

δT(E)  =  sup⁡t∈{0,…,T−1} sup⁡x∈S 12∑x′∈S∣Pt,1(x,x′)−Pt,0(x,x′)∣,\delta_T(E)\;=\;\sup_{t\in\{0,\dots,T-1\}}\ \sup_{x\in S}\ \frac12\sum_{x'\in S}\bigl|P_{t,1}(x,x')-P_{t,0}(x,x')\bigr| ,δT​(E)=t∈{0,…,T−1}sup​ x∈Ssup​ 21​x′∈S∑​​Pt,1​(x,x′)−Pt,0​(x,x′)​,

i.e. the largest, over times t<Tt<Tt<T and states xxx, of half the ℓ1\ell^1ℓ1-distance between row xxx of Pt,1P_{t,1}Pt,1​ and row xxx of Pt,0P_{t,0}Pt,0​. (The supremum is the real-number supremum with the convention that an empty supremum is 000; under (H1), and because an environment can only exist on a non-empty SSS, both index sets are finite and non-empty, so this is a maximum, and it lies in [0,1][0,1][0,1].)

Integral and variance. ∫\int∫ denotes the Lebesgue (Bochner) integral, which by convention is 000 for a non-integrable integrand. The variance of a function FFF under a measure μ\muμ is

Var⁡μ(F)  =  ∫ΩT∣F−∫F dμ∣2dμ,\operatorname{Var}_\mu(F)\;=\;\int_{\Omega_T}\Bigl|F-\int F\,d\mu\Bigr|^2 d\mu ,Varμ​(F)=∫ΩT​​​F−∫Fdμ​2dμ,

computed as an integral with values in [0,∞][0,\infty][0,∞] and then converted to a real number, with the value ∞\infty∞ converted to 000. (Under (H6) every YtY_tYt​ lies in [−M,M][-M,M][−M,M] almost surely for each of the measures above, so all integrands here are almost surely bounded.)

Conclusion. Under (H1)–(H6), both of the following hold.

(I)

∣∫ΩTG^T,k dPT,1/2  −  GT(E)∣  ≤  C1(1−γk(1−γ)2 δT(E)2 M  +  γk1−γ δT(E) M).\Bigl|\int_{\Omega_T}\widehat G_{T,k}\,d\mathbb{P}_{T,1/2}\;-\;G_T(E)\Bigr|\;\le\;C_1\left(\frac{1-\gamma^{k}}{(1-\gamma)^{2}}\,\delta_T(E)^{2}\,M\;+\;\frac{\gamma^{k}}{1-\gamma}\,\delta_T(E)\,M\right).​∫ΩT​​GT,k​dPT,1/2​−GT​(E)​≤C1​((1−γ)21−γk​δT​(E)2M+1−γγk​δT​(E)M).

(II)

Var⁡PT,1/2(G^T,k)  ≤  C2((k+1)2M2T  +  γ (k+1) M2T (1−γ)).\operatorname{Var}_{\mathbb{P}_{T,1/2}}\bigl(\widehat G_{T,k}\bigr)\;\le\;C_2\left(\frac{(k+1)^{2}M^{2}}{T}\;+\;\frac{\gamma\,(k+1)\,M^{2}}{T\,(1-\gamma)}\right).VarPT,1/2​​(GT,k​)≤C2​(T(k+1)2M2​+T(1−γ)γ(k+1)M2​).

Both inequalities are non-strict. The estimator's parameter and the sampling law's parameter are both the fixed number 12\tfrac1221​; no other value of the parameter is covered.

Boundary and degenerate cases as the statement literally has them.

  • Empty SSS. If SSS is empty there is no environment on SSS (the condition ∑xρ(x)=1\sum_x\rho(x)=1∑x​ρ(x)=1 cannot hold), so nothing is asserted for empty SSS.
  • Sign of MMM. MMM ranges over all reals. For M<0M<0M<0 the interval [−M,M][-M,M][−M,M] is empty, (H6) cannot hold, and nothing is asserted. For M=0M=0M=0, (H6) says every Rx,zR_{x,z}Rx,z​ is the point mass at 000, and both right-hand sides are 000.
  • k=0k=0k=0 is allowed. Then the inner sum of the estimator is the single term YuY_uYu​, γ0=1\gamma^{0}=1γ0=1, and the right-hand side of (I) is C1 δT(E) M/(1−γ)C_1\,\delta_T(E)\,M/(1-\gamma)C1​δT​(E)M/(1−γ).
  • Large kkk. For k≥T−1k\ge T-1k≥T−1 the window condition u≤t≤u+ku\le t\le u+ku≤t≤u+k is just t≥ut\ge ut≥u, so k=T−1k=T-1k=T−1 and k=Tk=Tk=T give the same estimator; the right-hand sides still depend on kkk. Values k>Tk>Tk>T are excluded by (H2).
  • Range of γ\gammaγ. γ\gammaγ is strictly between 000 and 111; γ=0\gamma=0γ=0 and γ=1\gamma=1γ=1 are excluded, so no denominator 1−γ1-\gamma1−γ vanishes.
  • Which matrices are constrained and which are used. The law PT,θ\mathbb{P}_{T,\theta}PT,θ​ involves only Ps,zP_{s,z}Ps,z​ for 0≤s≤T−20\le s\le T-20≤s≤T−2. Hypothesis (H5) and the quantity δT(E)\delta_T(E)δT​(E) involve Pt,zP_{t,z}Pt,z​ for 0≤t≤T−10\le t\le T-10≤t≤T−1, i.e. they also involve PT−1,zP_{T-1,z}PT−1,z​, which does not enter the law. The matrices Pt,zP_{t,z}Pt,z​ with t≥Tt\ge Tt≥T appear nowhere. In particular for T=1T=1T=1 no transition matrix enters the law, while (H5) and δ1(E)\delta_1(E)δ1​(E) concern P0,zP_{0,z}P0,z​.
  • What δT(E)\delta_T(E)δT​(E) measures. Only the transition matrices enter δT(E)\delta_T(E)δT​(E); the measures Rx,1R_{x,1}Rx,1​ and Rx,0R_{x,0}Rx,0​ do not. When δT(E)=0\delta_T(E)=0δT​(E)=0 the right-hand side of (I) is 000, so (I) then asserts equality of the two quantities.
  • The mixing hypothesis is a one-step condition imposed separately for each t<Tt<Tt<T and each Boolean zzz, over pairs of probability vectors only. If SSS has exactly one element it holds for every γ\gammaγ (both sides are 000).
  • The last Boolean coordinate. zT−1z_{T-1}zT−1​ enters no transition factor of the weight; it affects only the law RxT−1,zT−1R_{x_{T-1},z_{T-1}}RxT−1​,zT−1​​ of YT−1Y_{T-1}YT−1​.
Human review
  • Endorsed by tianyipeng · Oct 12, 2026

  • Endorsed by Wenqian Xing · Oct 12, 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