Proposition 7.31 — integrating a semicontinuous function against a continuous kernel preserves semicontinuity
ProvedBertsekasShreve.SemicontSelection.prop7_31_semicontinuous_integralLet and be separable metrizable spaces, let be a continuous stochastic kernel on given (a continuous map from into the space of Borel probability measures with the weak topology), and let be Borel-measurable. Define
the integral of the extended-real-valued function in the sense with .
- If is lower semicontinuous and bounded below by a real number, then is lower semicontinuous and bounded below by a real number.
- If is upper semicontinuous and bounded above by a real number, then is upper semicontinuous and bounded above by a real number.
This is the semicontinuity of the expectation step of the dynamic programming algorithm: under a continuous transition kernel, the expected cost-to-go inherits semicontinuity from the cost-to-go.
Formalization Note The integral is the published definition DupacovaWets.Consistency.expect, which is exactly Eq. (43) of Chapter 7 with the convention (42): it equals when and (computed in EReal, hence when only is infinite) otherwise; both parts are lower Lebesgue integrals of -valued functions. Borel measurability of is with respect to the Borel σ-algebra of the product topology on . Boundedness is by a real constant, since every EReal-valued function is bounded by .
import Mathlib import Definitions.Def_DupacovaWets_Consistency_expect
namespace BertsekasShreve.SemicontSelection
open MeasureTheory TopologicalSpace
/-- Proposition 7.31 (Bertsekas & Shreve, p. 148). Let `X`, `Y` be separable metrizable spaces,
`q(dy|x)` a continuous stochastic kernel on `Y` given `X`, and `f : X × Y → R*` Borel-measurable.
Let `λ(x) = ∫ f(x, y) q(dy|x)` (the extended integral (43) of Chapter 7, `∫ f⁺ − ∫ f⁻` with `∞ − ∞ = ∞`, which is `DupacovaWets.Consistency.expect`).
(a) If `f` is lower semicontinuous and bounded below, so is `λ`.
(b) If `f` is upper semicontinuous and bounded above, so is `λ`. -/
theorem prop7_31_semicontinuous_integral {X Y : Type*}
[TopologicalSpace X] [MetrizableSpace X] [SeparableSpace X]
[TopologicalSpace Y] [MetrizableSpace Y] [SeparableSpace Y]
[MeasurableSpace Y] [BorelSpace Y]
(q : X → ProbabilityMeasure Y) (hq : Continuous q) (f : X × Y → EReal)
(hf : @Measurable (X × Y) EReal (borel (X × Y)) _ f) :
((LowerSemicontinuous f ∧ ∃ b : ℝ, ∀ z, (b : EReal) ≤ f z) →
LowerSemicontinuous (fun x => DupacovaWets.Consistency.expect (q x : Measure Y) (fun y => f (x, y))) ∧
∃ b : ℝ, ∀ x, (b : EReal) ≤ DupacovaWets.Consistency.expect (q x : Measure Y) (fun y => f (x, y))) ∧
((UpperSemicontinuous f ∧ ∃ b : ℝ, ∀ z, f z ≤ (b : EReal)) →
UpperSemicontinuous (fun x => DupacovaWets.Consistency.expect (q x : Measure Y) (fun y => f (x, y))) ∧
∃ b : ℝ, ∀ x, DupacovaWets.Consistency.expect (q x : Measure Y) (fun y => f (x, y)) ≤ (b : EReal)) := by sorry
end BertsekasShreve.SemicontSelection
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.