Theorem 2.7 — smallest-superharmonic-majorant criterion
DisprovedPeskirShiryaev.General.sufficient_majorantLet satisfy the strong Markov and regularity assumptions of Section 2.2 on , . Let be measurable with for all . Define over stopping times finite almost surely at the initial state. Suppose the smallest superharmonic majorant of exists, is lower semicontinuous, and is upper semicontinuous. Let
The entry time is a Markov time. If for every , then
and is optimal. If for some , there is no almost-surely finite stopping rule optimal at every initial state.
This is the sufficient majorant criterion for existence, including its nonexistence alternative.
Formalization Note This item is the infinite-horizon Theorem 2.7 itself. The candidate majorant and the expected-reward value are separate objects, and equality is a conclusion. The nonexistence clause negates a common optimal rule for all initial states, matching the global convention of (2.2.9). Empty contact sets yield infinity. The finite-horizon extension in Remark 2.8 and Corollary 2.9 is outside this mission's scope.
import Definitions.Def_PeskirShiryaev_General_StoppingProblem open MeasureTheory ProbabilityTheory Filter Set open scoped NNReal ENNReal Topology
/- Imported Lean declarations, reproduced for self-contained code review:
namespace PeskirShiryaev.General
abbrev State (d : ℕ) := Fin d → ℝ
def RightContinuousPaths {Ω A : Type*} [TopologicalSpace A]
(Y : ℝ≥0 → Ω → A) (ω : Ω) : Prop :=
∀ t, ContinuousWithinAt (fun s => Y s ω) (Ici t) t
structure MarkovModel (d : ℕ) (Ω : Type*) [mΩ : MeasurableSpace Ω] where
dimension_pos : 0 < d
P : State d → Measure Ω
probability : ∀ x, IsProbabilityMeasure (P x)
measurable_laws : ∀ A : Set Ω, MeasurableSet A → Measurable (fun x => P x A)
filtration : Filtration ℝ≥0 mΩ
filtration_rightContinuous : filtration.IsRightContinuous
X : ℝ≥0 → Ω → State d
adapted : Adapted filtration X
starts : ∀ x, ∀ᵐ ω ∂P x, X 0 ω = x
rightContinuous : ∀ ω, RightContinuousPaths X ω
strongMarkov : ∀ (x : State d) (τ : Ω → ℝ≥0∞)
(hτ : IsStoppingTime filtration τ),
(∀ᵐ ω ∂P x, τ ω < ⊤) →
∀ (s : ℝ≥0) (f : State d → ℝ), Measurable f →
Bornology.IsBounded (Set.range f) →
(P x)[fun ω => f (X ((τ ω).toNNReal + s) ω) | hτ.measurableSpace]
=ᵐ[P x] (fun ω => ∫ ω', f (X s ω') ∂P (X (τ ω).toNNReal ω))
leftContinuous_stopping : ∀ (x : State d) (τ : Ω → ℝ≥0∞)
(τn : ℕ → Ω → ℝ≥0∞),
IsStoppingTime filtration τ →
(∀ n, IsStoppingTime filtration (τn n)) →
(∀ ω, Monotone (fun n => τn n ω)) →
(∀ᵐ ω ∂P x, Tendsto (fun n => τn n ω) atTop (𝓝 (τ ω))) →
∀ᵐ ω ∂P x, τ ω < ⊤ →
Tendsto (fun n => X (τn n ω).toNNReal ω) atTop (𝓝 (X (τ ω).toNNReal ω))
end PeskirShiryaev.General
namespace PeskirShiryaev.General
variable {d : ℕ} {Ω : Type*} [MeasurableSpace Ω]
def Admissible (M : MarkovModel d Ω) (x : State d) (τ : Ω → ℝ≥0∞) : Prop :=
IsStoppingTime M.filtration τ ∧ ∀ᵐ ω ∂M.P x, τ ω < ⊤
noncomputable def gainSup (M : MarkovModel d Ω) (G : State d → ℝ) (ω : Ω) : ℝ≥0∞ :=
⨆ t : ℝ≥0, ENNReal.ofReal |G (M.X t ω)|
def GainIntegrable (M : MarkovModel d Ω) (G : State d → ℝ) : Prop :=
Measurable G ∧ ∀ x, AEMeasurable (gainSup M G) (M.P x) ∧
(∫⁻ ω, gainSup M G ω ∂M.P x) < ⊤
noncomputable def reward (M : MarkovModel d Ω) (G : State d → ℝ)
(τ : Ω → ℝ≥0∞) (ω : Ω) : ℝ :=
if τ ω = ⊤ then 0 else G (M.X (τ ω).toNNReal ω)
noncomputable def value (M : MarkovModel d Ω) (G : State d → ℝ) (x : State d) : ℝ :=
sSup {r : ℝ | ∃ τ : Ω → ℝ≥0∞, Admissible M x τ ∧ r = ∫ ω, reward M G τ ω ∂M.P x}
def Optimal (M : MarkovModel d Ω) (G : State d → ℝ) (τ : Ω → ℝ≥0∞) : Prop :=
∀ x, Admissible M x τ ∧ Integrable (reward M G τ) (M.P x) ∧
(∫ ω, reward M G τ ω ∂M.P x) = value M G x
def Superharmonic (M : MarkovModel d Ω) (F : State d → ℝ) : Prop :=
Measurable F ∧ ∀ (x : State d) (σ : Ω → ℝ≥0∞), Admissible M x σ →
Integrable (reward M F σ) (M.P x) ∧ (∫ ω, reward M F σ ω ∂M.P x) ≤ F x
def LeastMajorant (M : MarkovModel d Ω) (G F : State d → ℝ) : Prop :=
Superharmonic M F ∧ (∀ x, G x ≤ F x) ∧
∀ H : State d → ℝ, Superharmonic M H → (∀ x, G x ≤ H x) → ∀ x, F x ≤ H x
noncomputable def entryTime (M : MarkovModel d Ω) (D : Set (State d)) (ω : Ω) : ℝ≥0∞ :=
⨅ t : {t : ℝ≥0 // M.X t ω ∈ D}, (t.val : ℝ≥0∞)
noncomputable def contactTime (M : MarkovModel d Ω) (G F : State d → ℝ) : Ω → ℝ≥0∞ :=
entryTime M {x | F x = G x}
noncomputable def stoppedValue (M : MarkovModel d Ω) (F : State d → ℝ)
(τ : Ω → ℝ≥0∞) (t : ℝ≥0) (ω : Ω) : ℝ :=
F (M.X (min (t : ℝ≥0∞) (τ ω)).toNNReal ω)
end PeskirShiryaev.General
-/
namespace PeskirShiryaev.General
theorem sufficient_majorant {d : ℕ} {Ω : Type*} [MeasurableSpace Ω]
(M : MarkovModel d Ω) (G Vhat : State d → ℝ) (hG : GainIntegrable M G)
(hmajorant : LeastMajorant M G Vhat)
(hlsc : LowerSemicontinuous Vhat) (husc : UpperSemicontinuous G) :
IsStoppingTime M.filtration (contactTime M G Vhat) ∧
(((∀ x, ∀ᵐ ω ∂M.P x, contactTime M G Vhat ω < ⊤) →
Vhat = value M G ∧ Optimal M G (contactTime M G Vhat)) ∧
((∃ x, M.P x {ω | contactTime M G Vhat ω < ⊤} < 1) →
¬ ∃ τ : Ω → ℝ≥0∞, Optimal M G τ)) := by sorry
end PeskirShiryaev.General
Read-back
What the Lean code literally says, in plain math · claude-opus-5-5
Setting. Fix a natural number , a measurable space , and a "Markov model" on with state space (functions , with the product topology). The model consists of the following data and requirements:
- .
- A family of probability measures on , such that is measurable for every measurable .
- A filtration indexed by , contained in and right-continuous.
- A process for , adapted to .
- holds -almost surely for every .
- For every , the path is right-continuous at every .
- Strong Markov property. Take any , any -valued -stopping time with -a.s., any , and any measurable bounded . Then -a.s.
Here "" means evaluated at converted to a finite time, with the convention that is converted to time .
- Left-continuity along stopping times. Take any , any stopping time , and any stopping times such that is nondecreasing for every and -a.s. Then -a.s. on we have , with the same convention that an infinite time is read as .
Auxiliary notions.
- A map is admissible at if it is an -stopping time and -a.s.
- For , the reward of is if , and if .
- The value is
This is the supremum in , which by convention equals if the set is unbounded above or empty. Each integral is the Lebesgue (Bochner) integral, which by convention equals when is not -integrable.
- is superharmonic if is measurable and, for every and every admissible at , is -integrable with .
- A stopping time is optimal if, for every , it is admissible at , is -integrable, and .
- For , the contact time is
The infimum is taken in , so it equals when the path never enters , and it need not be attained.
Hypotheses. Let , and assume:
- is measurable, and for every the function is -a.e. measurable with finite -expectation.
- is superharmonic and pointwise.
- Every superharmonic with pointwise satisfies pointwise.
- is lower semicontinuous and is upper semicontinuous.
Conclusion. Three statements hold together:
- (a) is an -stopping time, meaning for all .
- (b) If -a.s. for every , then for all , and is optimal in the sense above.
- (c) If there is some with , then no optimal stopping time exists. That is, there is no that is admissible at every , has -integrable reward for every , and attains for every .
Part (b) is an implication, not an equivalence, and part (c) asserts only the non-existence of a stopping time that is optimal at all starting points simultaneously.
Degenerate cases.
- is excluded by the model, so the state space is with .
- No model exists when is empty, since probability measures are required. No model exists when is a single point either: -a.s. for every would force one point to equal every . The model's hypotheses are therefore satisfiable only on richer .
- If is empty, or a path never enters , then there. If is empty, every , so the hypothesis of (b) fails and (c) asserts that no optimal stopping time exists.
- On :
- the reward is defined to be , not a limit of ;
- in the strong Markov and left-continuity axioms, silently means .
- wherever .
- The value uses the real-supremum convention: it would be if the set of expected rewards were unbounded above. The set is nonempty, because the constant time is admissible.
- An expected reward is taken as whenever the reward is not integrable or not measurable.
- is the measure (outer measure, if non-measurable) of that set.
- The "pointwise" inequalities and the semicontinuity requirements are imposed at every point of , including points the process may never visit.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.