Corollary 10.2.6 — three sufficient conditions for optimality of τ_{f*}
DisprovedMDPFinance.OptimalStopping.corollary_10_2_6Corollary 10.2.6 (p. 315). Suppose a stopping problem with unbounded horizon is given and is a maximizer of . Then it holds:
a) If the corresponding Markov Decision Problem has a bounding function with , then . Moreover, if for all then is an optimal stopping time. b) If , and , then for all and is an optimal stopping time. c) If is bounded from above and for all , then is an optimal stopping time and .
Theorem 10.2.3 b) needs two things — that the rule's value is -superharmonic and that the rule stops almost surely — and neither is free. This corollary is the practical entry point: three separately checkable conditions, each of which delivers one or both.
The three are genuinely different and none subsumes another. a) is a contraction argument (); b) buys almost sure finiteness from a strictly negative running reward, so waiting forever costs infinitely much, and it is the undiscounted case where a) is unavailable; c) assumes finiteness instead and buys from a one-sided bound on .
Note the direction of each implication. In a) and c) almost sure finiteness is a hypothesis; in b) it is a conclusion. Stating b)'s finiteness as a hypothesis would lose the sharpest part of the corollary.
Moderation note. The bounding function is Chapter 7's (b ≥ 0, constant c_r, Lebesgue drift bound); J − g bounded above is stated in [-∞,∞]; optimality via IsOptimal.
import Mathlib import Definitions.Def_MDPFinance_OptimalStopping_Stationary open MeasureTheory Filter Topology
namespace MDPFinance.OptimalStopping
/-- **Corollary 10.2.6** (p. 314), under (B), for a maximizer `f^* = 1_S` of `J`. a) If the
Markov Decision Model has a bounding function `b` with `βα_b < 1`, then `G_{f^*} = J`; if moreover
`ℙ_x(τ_{f^*} < ∞) = 1` for all `x`, `τ_{f^*}` is optimal. b) If `sup c < 0`, `β = 1` and
`G_{f^*} = J`, then `ℙ_x(τ_{f^*} < ∞) = 1` for all `x` and `τ_{f^*}` is optimal. c) If `J − g` is
bounded above and `ℙ_x(τ_{f^*} < ∞) = 1` for all `x`, then `τ_{f^*}` is optimal and
`G_{f^*} = J`. -/
theorem corollary_10_2_6 {E : Type*} [MeasurableSpace E] (P : StationaryProblem E)
(Pr : E → Measure (ℕ → E)) (hPr : P.IsPathLaw Pr) (hB : P.AssumptionB Pr)
(S : Set E) (hS : P.IsMaximizer S P.Jlim) :
(∀ B : P.BoundingFunction, P.beta * B.alpha < 1 →
(∀ x : E, P.Gf Pr S x = P.Jlim x) ∧
((∀ x : E, Pr x {w | P.ruleTime S w = ⊤} = 0) → P.IsOptimal Pr (P.ruleTime S))) ∧
((∃ s : ℝ, s < 0 ∧ ∀ x : E, P.c x ≤ s) → P.beta = 1 → (∀ x : E, P.Gf Pr S x = P.Jlim x) →
(∀ x : E, Pr x {w | P.ruleTime S w = ⊤} = 0) ∧ P.IsOptimal Pr (P.ruleTime S)) ∧
((∃ C : ℝ, ∀ x : E, P.Jlim x ≤ (P.g x : EReal) + (C : EReal)) →
(∀ x : E, Pr x {w | P.ruleTime S w = ⊤} = 0) →
P.IsOptimal Pr (P.ruleTime S) ∧ ∀ x : E, P.Gf Pr S x = P.Jlim x) := by sorry
end MDPFinance.OptimalStopping
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.