Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 4.2 — stability of P\mathcal PP  ⟺  \iff⟺ submartingale property   ⟺  \iff⟺Ψ=Ψˉ\Psi=\bar\PsiΨ=Ψˉ  ⟺  \iff⟺ Bellman's principle

Proved
MultiperiodRisk.Bellman.stable_tfae

by mikedeng1 · Sep 26, 2026 · Mathlib 0df444a (Lean v4.33.1)

dynamic-programmingp2o-batch-p100ap2o-gran-per-chapterp2o-plan-paperp2o-v1risk-measuressnell-envelopetime-consistency

Let (Ω,F,P0)(\Omega,\mathcal F,\mathbb P_0)(Ω,F,P0​) be a probability space with a filtration (Fn)(\mathcal F_n)(Fn​) whose initial σ-algebra F0\mathcal F_0F0​ is P0\mathbb P_0P0​-trivial, fix a horizon NNN, and let P\mathcal PP be a closed convex set of test probabilities on (Ω,FN)(\Omega,\mathcal F_N)(Ω,FN​), absolutely continuous with respect to P0\mathbb P_0P0​, with Pe≠∅\mathcal P^e\ne\emptysetPe=∅. For a value process XXX and a stopping time σ≤N\sigma\le Nσ≤N let

Ψσ(X)=ess.inf⁡{EQ[Xτ∣Fσ] ∣ σ≤τ≤N stopping time, Q∈Pe},\Psi_\sigma(X)=\operatorname*{ess.inf}\bigl\{\mathbb E_{\mathbb Q}[X_\tau\mid\mathcal F_\sigma]\ \bigm|\ \sigma\le\tau\le N\text{ stopping time},\ \mathbb Q\in\mathcal P^e\bigr\},Ψσ​(X)=ess.inf{EQ​[Xτ​∣Fσ​] ​ σ≤τ≤N stopping time, Q∈Pe},

and let Ψˉ(X)\bar\Psi(X)Ψˉ(X) be the generalized Snell envelope of Theorem 4.1. The following are equivalent:

  1. P\mathcal PP is stable (Definition 3.1);
  2. for every Q∈P\mathbb Q\in\mathcal PQ∈P and every value process XXX, (Ψn(X))0≤n≤N(\Psi_n(X))_{0\le n\le N}(Ψn​(X))0≤n≤N​ is a Q\mathbb QQ-submartingale;
  3. for every value process XXX, Ψn(X)=Ψˉn(X)\Psi_n(X)=\bar\Psi_n(X)Ψn​(X)=Ψˉn​(X) a.s. for 0≤n≤N0\le n\le N0≤n≤N;
  4. (Bellman's principle) for every value process XXX and all stopping times σ≤τ≤N\sigma\le\tau\le Nσ≤τ≤N,
Ψσ(X)=Ψσ(Xτ−+Ψτ(τX)1[τ,N])a.s.\Psi_\sigma(X)=\Psi_\sigma\bigl(X^{\tau-}+\Psi_\tau({}^\tau X)\mathbf 1_{[\tau,N]}\bigr)\quad\text{a.s.}Ψσ​(X)=Ψσ​(Xτ−+Ψτ​(τX)1[τ,N]​)a.s.

The theorem identifies stability under pasting as exactly the property of the set of test probabilities that makes the risk-adjusted value time consistent: the value can be computed by backward induction, and replacing the future of a position by its risk-adjusted value does not change today's value.

Formalization Note Two conventions the page leaves implicit are made explicit. (a) F0\mathcal F_0F0​ is P0\mathbb P_0P0​-trivial: the proof evaluates Ψ0(X)\Psi_0(X)Ψ0​(X) as the number inf⁡Q∈PEQ[f]\inf_{\mathbb Q\in\mathcal P}\mathbb E_{\mathbb Q}[f]infQ∈P​EQ​[f], and without it the implications (1)–(3) ⇒\Rightarrow⇒ stability fail (for N=1N=1N=1 and a two-atom F0\mathcal F_0F0​, statements 2–3 hold for every P\mathcal PP while stability can fail). (b) Pe≠∅\mathcal P^e\neq\emptysetPe=∅, implied by the paper's convention P0∈P\mathbb P_0\in\mathcal PP0​∈P. X−1=0X_{-1}=0X−1​=0. All equalities of risk-adjusted values hold P0\mathbb P_0P0​-a.s. because essential infima are defined up to null sets.


Retired 2026-10-27: a Lean-vs-paper fidelity audit found this statement weaker than, or narrower than, the cited result. A corrected statement will be resubmitted.

Preamble
import Mathlib
import Definitions.Def_MultiperiodRisk_Bellman_StopOps
Formal statement
namespace MultiperiodRisk.Bellman

open MeasureTheory

/-- Theorem 4.2: for a closed convex set of test probabilities with `𝒫ᵉ ≠ ∅` and an initial
σ-algebra `ℱ_0` that is trivial under `P₀`, stability is equivalent to each of
(i) `Ψ(X)` is a `ℚ`-submartingale for every `ℚ ∈ 𝒫` and every value process `X`;
(ii) `Ψ = Ψ̄` on value processes;
(iii) Bellman's principle `Ψ_σ(X) = Ψ_σ(X^{τ-} + Ψ_τ(^τX) 1_{[τ,N]})` for stopping times `σ ≤ τ`. -/
theorem stable_tfae {Ω : Type*} {m : MeasurableSpace Ω} (P₀ : Measure Ω)
    [IsProbabilityMeasure P₀] (ℱ : Filtration ℕ m) (N : ℕ)
    (hℱ₀ : ∀ s, MeasurableSet[ℱ 0] s → P₀ s = 0 ∨ P₀ s = 1)
    (D : TestSet P₀ ℱ N) (hPe : (Pe D).Nonempty) :
    List.TFAE
      [ IsStable D,
        ∀ f ∈ D.set, ∀ X : ℕ → Ω → ℝ, IsValueProcess P₀ ℱ N X →
          Submartingale (fun n => PsiN D X (min n N)) ℱ (Q P₀ f),
        ∀ X : ℕ → Ω → ℝ, IsValueProcess P₀ ℱ N X →
          ∀ n ≤ N, PsiN D X n =ᵐ[P₀] PsiBar D X n,
        ∀ X : ℕ → Ω → ℝ, IsValueProcess P₀ ℱ N X →
          ∀ (σ τ : Ω → WithTop ℕ) (hσ : IsBddStoppingTime ℱ N σ)
            (hτ : IsBddStoppingTime ℱ N τ), (∀ ω, σ ω ≤ τ ω) →
            Psi D X σ hσ.1 =ᵐ[P₀] Psi D (bellmanProcess D X τ hτ.1) σ hσ.1 ] := by sorry

end MultiperiodRisk.Bellman
Source
Artzner, Delbaen, Eber, Heath, Ku, Coherent Multiperiod Risk Adjusted Values and Bellman's Principle, Ann. Oper. Res. 152 (2007); manuscript of Nov. 16, 2004, p. 11, Theorem 4.2 (proof pp. 11–12)
Read-back

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

Let (Ω,m)(\Omega, m)(Ω,m) be a measurable space, P0P_0P0​ a probability measure on it, and F=(Fn)n∈N\mathcal{F} = (\mathcal{F}_n)_{n \in \mathbb{N}}F=(Fn​)n∈N​ a filtration of sub-σ-algebras of mmm indexed by the natural numbers. Let N∈NN \in \mathbb{N}N∈N be a horizon. The statement assumes:

  • (Triviality of F0\mathcal{F}_0F0​ under P0P_0P0​.) Every F0\mathcal{F}_0F0​-measurable set sss has P0(s)=0P_0(s) = 0P0​(s)=0 or P0(s)=1P_0(s) = 1P0​(s)=1.
  • (Test set.) DDD is an object of the bundle-defined type "test set" built from P0P_0P0​, F\mathcal{F}F and NNN. Its definition is not shown here. The statement uses only its underlying collection D.setD.\mathrm{set}D.set, whose elements are written fff.
  • (Non-emptiness.) The set Pe(D)\mathcal{P}^e(D)Pe(D) is nonempty. This is a bundle-defined set attached to DDD, and its definition is not shown.

The docstring describes the test set as "closed convex". No closedness or convexity hypothesis appears in the statement itself. Whether these properties hold depends entirely on the unseen definition of a test set.

The conclusion is that the following four propositions are pairwise logically equivalent. Each one implies every other one.

  1. Stability. DDD satisfies the bundle-defined predicate "stable" (definition not shown).

  2. Submartingale property. Take any f∈D.setf \in D.\mathrm{set}f∈D.set and any X:N×Ω→RX : \mathbb{N} \times \Omega \to \mathbb{R}X:N×Ω→R, (n,ω)↦Xn(ω)(n,\omega) \mapsto X_n(\omega)(n,ω)↦Xn​(ω), that satisfies the bundle-defined predicate "value process" relative to P0P_0P0​, F\mathcal{F}F, NNN. For every such pair, the real-valued process

n  ⟼  Ψmin⁡(n,N)N(X)n \;\longmapsto\; \Psi^N_{\min(n,N)}(X)n⟼Ψmin(n,N)N​(X)

is a submartingale with respect to F\mathcal{F}F under the measure Qf:=Q(P0,f)Q_f := Q(P_0, f)Qf​:=Q(P0​,f). Here:

  • ΨN\Psi^NΨN is the bundle-defined map PsiN(D,X,⋅)\mathrm{PsiN}(D, X, \cdot)PsiN(D,X,⋅), which gives a random variable for each time index.
  • Q(P0,f)Q(P_0, f)Q(P0​,f) is a bundle-defined measure on (Ω,m)(\Omega, m)(Ω,m).
  • The process is frozen at its time-NNN value for all n≥Nn \ge Nn≥N.
  • "Submartingale" means three things. (a) For each nnn, the time-nnn value is strongly Fn\mathcal{F}_nFn​-measurable. (b) For each nnn, it is QfQ_fQf​-integrable. (c) For all i≤ji \le ji≤j, the time-iii value is at most the QfQ_fQf​-conditional expectation of the time-jjj value given Fi\mathcal{F}_iFi​, QfQ_fQf​-almost everywhere.
  1. Agreement of the two operators. For every value process XXX and every n≤Nn \le Nn≤N,
ΨnN(X)=Ψˉn(X)P0-almost everywhere,\Psi^N_n(X) = \bar\Psi_n(X) \quad P_0\text{-almost everywhere},ΨnN​(X)=Ψˉn​(X)P0​-almost everywhere,

where Ψˉ\bar\PsiΨˉ is the bundle-defined map PsiBar(D,X,⋅)\mathrm{PsiBar}(D, X, \cdot)PsiBar(D,X,⋅).

  1. Bellman principle. Take any value process XXX and any two random times σ,τ:Ω→N∪{∞}\sigma, \tau : \Omega \to \mathbb{N} \cup \{\infty\}σ,τ:Ω→N∪{∞} that satisfy:

    • both satisfy the bundle-defined predicate "bounded stopping time" relative to F\mathcal{F}F and NNN;
    • σ(ω)≤τ(ω)\sigma(\omega) \le \tau(\omega)σ(ω)≤τ(ω) for every ω\omegaω, not merely almost every ω\omegaω.

    For all such XXX, σ\sigmaσ, τ\tauτ,

Ψσ(X)  =  Ψσ(Bτ(X))P0-almost everywhere.\Psi_\sigma(X) \;=\; \Psi_\sigma\big(B_\tau(X)\big) \quad P_0\text{-almost everywhere}.Ψσ​(X)=Ψσ​(Bτ​(X))P0​-almost everywhere.

Here:

  • Ψσ\Psi_\sigmaΨσ​ is the bundle-defined map Psi(D,⋅,σ,⋅)\mathrm{Psi}(D, \cdot, \sigma, \cdot)Psi(D,⋅,σ,⋅). It also takes as input the first component of the proof that σ\sigmaσ is a bounded stopping time.
  • Bτ(X)B_\tau(X)Bτ​(X) is the bundle-defined process bellmanProcess(D,X,τ,⋅)\mathrm{bellmanProcess}(D, X, \tau, \cdot)bellmanProcess(D,X,τ,⋅). It likewise takes the first component of the bounded-stopping-time proof for τ\tauτ.
  • The docstring describes Bτ(X)B_\tau(X)Bτ​(X) as Xτ−+Ψτ(τX) 1[τ,N]X^{\tau-} + \Psi_\tau({}^\tau X)\,\mathbf{1}_{[\tau,N]}Xτ−+Ψτ​(τX)1[τ,N]​. That description is not visible in the statement.

Measures used in each condition.

  • The almost-everywhere equalities in (3) and (4) are with respect to P0P_0P0​.
  • The submartingale inequalities in (2) are with respect to each QfQ_fQf​.
  • The triviality hypothesis on F0\mathcal{F}_0F0​ is imposed only under P0P_0P0​. Nothing is assumed about how QfQ_fQf​ behaves on F0\mathcal{F}_0F0​.

Degenerate cases.

  • Ω\OmegaΩ cannot be empty, because P0P_0P0​ is a probability measure.
  • If Ω\OmegaΩ has one element, or F0\mathcal{F}_0F0​ is the trivial σ-algebra, the triviality hypothesis holds automatically.
  • No value processes. If no XXX satisfies the value-process predicate, then (2), (3) and (4) are all vacuously true. The theorem then asserts outright that DDD is stable.
  • Empty D.setD.\mathrm{set}D.set. If D.setD.\mathrm{set}D.set is empty, (2) is vacuously true, and the theorem asserts that (1), (3) and (4) all hold. Whether this case is compatible with Pe(D)≠∅\mathcal{P}^e(D) \neq \emptysetPe(D)=∅ depends on the unseen definitions.
  • Empty Pe(D)\mathcal{P}^e(D)Pe(D). If Pe(D)\mathcal{P}^e(D)Pe(D) can never be nonempty for any test set, the theorem is vacuous.
  • Horizon N=0N = 0N=0.
    • In (2), the process is the constant-in-time process Ψ0N(X)\Psi^N_0(X)Ψ0N​(X). The condition reduces to three requirements on Ψ0N(X)\Psi^N_0(X)Ψ0N​(X): it is F0\mathcal{F}_0F0​-measurable, it is QfQ_fQf​-integrable, and QfQ_fQf​-a.e. it is at most its own QfQ_fQf​-conditional expectation given each Fi\mathcal{F}_iFi​.
    • In (3), only n=0n = 0n=0 is tested.
    • In (4), only stopping times allowed by the bounded-stopping-time predicate with bound 000 are tested.
  • Times after NNN. The process XXX is defined for all times n∈Nn \in \mathbb{N}n∈N. Whether values after NNN matter depends on the unseen definitions.
  • Infinite random times. The random times take values in N∪{∞}\mathbb{N} \cup \{\infty\}N∪{∞}. Whether the value ∞\infty∞ is excluded depends on the unseen "bounded stopping time" predicate.
  • Default conditional expectation. The QfQ_fQf​-conditional expectation given Fi\mathcal{F}_iFi​ is the formal default 000 whenever QfQ_fQf​ restricted to Fi\mathcal{F}_iFi​ is not σ-finite. In that case, condition (c) of (2) requires the time-iii value to be ≤0\le 0≤0 QfQ_fQf​-a.e.
  • Nature of QfQ_fQf​. No hypothesis states that QfQ_fQf​ is a probability measure, or even finite. That depends on the unseen definitions of QQQ and of the test set.
Human review
  • Endorsed by Shuze Chen · Sep 27, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Sep 27, 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