§5.2.3, p.23 — the Snell envelope dominates and is a -supermartingale: ,
ProvedRiskStopTC.snell_dominatesdynamic-programmingoptimal-stoppingp2o-gran-per-chapterp2o-plan-paperp2o-strengthenedp2o-v1risk-measuresstochastic-programming
Throughout, is a probability space with a filtration , , , and for a fixed . Relations between random variables hold -almost surely.
Let , , be monotone, translation equivariant mappings with (5.17), let , and let be the Snell envelope (5.14). Then
- and for ;
- for , i.e. is a -supermartingale in the sense of (5.24).
These are the two properties of used in the domination step (5.18) of the proof of Theorem 5.6.
Differences from the source:
- generalization: E_t ⪰ Z_t and E_t ⪰ ϱ_t(E_{t+1}) → also concludes E_t ∈ Z_t for t ≤ T; reason: Definition 5.5 treats E as a process in Z_{0,T}; dropping that conjunct gives the page's statement
- encoding: Z_t = L^p(Ω, F_t, P) as classes of a.e.-equal functions; ⪯ and = almost surely → functions Ω → ℝ with an F_t-measurable representative in L^p (IsLp (ℱ t) P p); relations ≤ᵐ[P] and =ᵐ[P]; reason: every class has such a representative; Lean convention; same instances
- encoding: monotone translation equivariant mappings ϱ_t|F_t : Z_{t+1} → Z_t with (5.17), t = 0, …, T − 1 → one family ϱ : ℕ → (Ω → ℝ) → Ω → ℝ with IsStoppingFamily (maps-to, monotone, translation equivariant, (5.17)) on IsLp (ℱ (t+1)) arguments, t < T; reason: values off Z_{t+1} and maps with t ≥ T are never used; monotonicity makes each ϱ_t respect a.e. equality, so it acts on classes; same instances
Preamble
import Mathlib import Definitions.Def_RiskFunctionals_NestedAVaR_Compositions import Definitions.Def_MultiperiodRisk_Bellman_TestSet import Definitions.Def_RiskStopTC_IsStoppingFamily import Definitions.Def_RiskStopTC_snell open MeasureTheory RiskFunctionals.NestedAVaR
Formal statement
namespace RiskStopTC
/-- **The Snell envelope dominates the reward and is a 𝔓-supermartingale.** Pichler, Liu &
Shapiro, arXiv:1808.10807v3, §5.2.3, p.23: "It follows by the definition (5.14) that the Snell
envelope sequence {Et}0≤t≤T is 𝔓-supermartingale. It also follows from (5.14) that Et ⪰ Zt, i.e.,
Et dominates Zt." (𝔓-supermartingale is (5.24): `X_t ⪰ ϱt|Ft(X_{t+1})`, `t = 0, …, T − 1`.)
For `t ≤ T`: `E_t ∈ Z_t` and `Z_t ⪯ E_t`; for `t < T`: `ϱ_t(E_{t+1}) ⪯ E_t`.
**Formalization Note.** `E` is `snell ϱ Z T`, defined by (5.14). `E_t ∈ Z_t` is the statement that
the recursion (5.14) stays in the spaces the paper's notation presumes. -/
theorem snell_dominates {Ω : Type*} {mΩ : MeasurableSpace Ω} (ℱ : Filtration ℕ mΩ) (P : Measure[mΩ] Ω)
[IsProbabilityMeasure P] (p : ENNReal) (hp : 1 ≤ p) (T : ℕ) (hℱ : IsStandardFiltration ℱ T)
(ϱ : ℕ → (Ω → ℝ) → Ω → ℝ) (hϱ : IsStoppingFamily ℱ P p T ϱ)
(Z : ℕ → Ω → ℝ) (hZ : ∀ t ≤ T, IsLp (ℱ t) P p (Z t)) :
(∀ t ≤ T, IsLp (ℱ t) P p (snell ϱ Z T t) ∧ Z t ≤ᵐ[P] snell ϱ Z T t) ∧
∀ t < T, ϱ t (snell ϱ Z T (t + 1)) ≤ᵐ[P] snell ϱ Z T t := by sorry
end RiskStopTC
Source
Pichler, Liu & Shapiro, Risk averse stochastic programming: time consistency and optimal stopping, arXiv:1808.10807v3, §5.2.3, p.23 ("It follows by the definition (5.14) that the Snell envelope sequence {Et}0≤t≤T is 𝔓-supermartingale. It also follows from (5.14) that Et ⪰ Zt"); (5.24), p.23
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.