Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Eq. (10.42) — under (F^+)(\hat F^+)(F^+) the (PSI) optimal cost is >−∞>-\infty>−∞

Open
BertsekasShreve.ImperfectInfo.optCostHat_gt_bot

by tianyipeng · Nov 7, 2026 · Mathlib 0df444a (Lean v4.33.1)

dynamic-programmingimperfect-informationoptimal-controlsufficient-statistic

Let an (ISI) model with finite horizon NNN be given, together with a statistic (η0,…,ηN−1)(\eta_0,\dots,\eta_{N-1})(η0​,…,ηN−1​) sufficient for control (Definition 10.6) and the associated perfect state information model (PSI) of Definition 10.7: states YkY_kYk​, control space CCC, constraint sets U^k(yk)=(Γ^k)yk\hat U_k(y_k)=(\hat\Gamma_k)_{y_k}U^k​(yk​)=(Γ^k​)yk​​, one-stage costs g^k:Γ^k→R∗\hat g_k:\hat\Gamma_k\to\mathbb R^*g^​k​:Γ^k​→R∗ and transition kernels t^k(dyk+1∣yk,uk)\hat t_k(dy_{k+1}\mid y_k,u_k)t^k​(dyk+1​∣yk​,uk​). For a (PSI) policy π^∈Π^′\hat\pi\in\hat\Pi'π^∈Π^′ (universally measurable kernels μ^k(duk∣y0,u0,…,yk)\hat\mu_k(du_k\mid y_0,u_0,\dots,y_k)μ^​k​(duk​∣y0​,u0​,…,yk​) with μ^k(U^k(yk)∣⋅)=1\hat\mu_k(\hat U_k(y_k)\mid\cdot)=1μ^​k​(U^k​(yk​)∣⋅)=1), J^N,π^(y)\hat J_{N,\hat\pi}(y)J^N,π^​(y) denotes its (0-originating) cost at y∈Y0y\in Y_0y∈Y0​ (eq. (31)), and J^N∗(y)=inf⁡π^∈Π^′J^N,π^(y)\hat J^*_N(y)=\inf_{\hat\pi\in\hat\Pi'}\hat J_{N,\hat\pi}(y)J^N∗​(y)=infπ^∈Π^′​J^N,π^​(y) is the optimal cost (eq. (32)).

If assumption (F^+)(\hat F^+)(F^+) holds for (PSI), i.e. ∫g^j− dq^j(π^k,pyk)<∞\int \hat g_j^-\,d\hat q_j(\hat\pi^k,p_{y_k})<\infty∫g^​j−​dq^​j​(π^k,pyk​​)<∞ for all kkk-originating policies, all yk∈Yky_k\in Y_kyk​∈Yk​ and k≤j≤N−1k\le j\le N-1k≤j≤N−1, then

J^N∗(y0)>−∞∀ y0∈Y0.\hat J^*_N(y_0)>-\infty\qquad\forall\,y_0\in Y_0.J^N∗​(y0​)>−∞∀y0​∈Y0​.

This is equation (42) in the proof of Proposition 10.3; it is Lemma 8.3 combined with Proposition 8.2 (JN∗=TN(J0)J^*_N=T^N(J_0)JN∗​=TN(J0​)) applied to the model (PSI). It is what makes an ε\varepsilonε-optimal (PSI) policy satisfy J^N,π^≤J^N∗+ε\hat J_{N,\hat\pi}\le\hat J^*_N+\varepsilonJ^N,π^​≤J^N∗​+ε everywhere.

Formalization Note The model (PSI) is the one bundled in SuffStat (Definitions 10.6–10.8); J^N∗\hat J^*_NJ^N∗​ is σ.optCostHat.

Preamble
import Mathlib
import Definitions.Def_BertsekasShreve_ImperfectInfo_SufficientStatistic

open MeasureTheory ProbabilityTheory
Formal statement
namespace BertsekasShreve.ImperfectInfo

theorem optCostHat_gt_bot {S C Z : Type}
    [TopologicalSpace S] [MeasurableSpace S] [BorelSpace S]
    [TopologicalSpace C] [MeasurableSpace C] [BorelSpace C]
    [TopologicalSpace Z] [MeasurableSpace Z] [BorelSpace Z]
    (M : ISIModel S C Z) {Y : ℕ → Type} [∀ k, TopologicalSpace (Y k)] [∀ k, MeasurableSpace (Y k)]
    (σ : SuffStat M Y)
    (hF : σ.FhatPlus) (y : Y 0) :
    ⊥ < σ.optCostHat y := by sorry

end BertsekasShreve.ImperfectInfo
Source
Bertsekas & Shreve, Stochastic Optimal Control: The Discrete-Time Case, Athena Scientific 1996, p. 257, eq. (42) in the proof of Proposition 10.3 (citing Lemma 8.3, p. 196, and Proposition 8.2, p. 198)

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