Theorem 4.2.2 — Strassen's ≤_cx theorem
ProvedPalmQueueing.Ordering.strassen_cxTheorem 4.2.2 (Strassen's theorem). Let and be two integrable distributions in ; (resp. ) if and only if there exist two -valued random variables and defined on the same probability space with distributions and respectively, and such that (resp. ) a.s.
The convex order holds exactly when the larger distribution is a martingale dilation of the smaller: is obtained from by spreading each point out without moving its conditional mean. This is the classical Strassen representation, and it is what makes the convex order tractable — comparison results for queues become induction arguments on a coupling rather than analytic manipulations of convolutions of c.d.f.'s.
Both clauses are stated, as the book states them together: with and with .
Both directions are asserted. The hard one is the construction of the coupling; a one-directional statement would be the easy half, since Jensen's inequality gives from any such pair immediately.
is a conditional expectation, not an equality of expectations: the content is that the identity holds a.s. given .
Integrability is a hypothesis here and not in Theorem 4.2.1. The convex order needs first moments; the stochastic order does not. The two are not harmonised.
The book attributes both theorems to Strassen (1965) and proves neither. Example 4.2.3 on the same page shows the construction at work: for the comparison of exponential and Erlang distributions, symmetry makes a function independent of , and a.s.
import Mathlib import Definitions.Def_PalmQueueing_Ordering_PartialOrders import Definitions.Def_PalmQueueing_Ordering_IntegralOrders /-! # Theorem 4.2.2: Strassen's `≤_cx` theorem (§4.2.3, p.278) -/
namespace PalmQueueing.Ordering
open MeasureTheory
/-- **Theorem 4.2.2 (Strassen's `≤_cx` theorem)** (p.278). Let `F` and `G` be two integrable
distributions in `𝒟(ℝⁿ)`; `F ≤_cx G` (resp. `F ≤_icx G`) **if and only if** there exist two
`ℝⁿ`-valued random variables `X` and `Y` defined on the same probability space with distributions
`F` and `G` respectively, and such that `E[Y | X] = X` (resp. `E[Y | X] ≥ X`) a.s.
The convex order holds exactly when the larger distribution is a **martingale dilation** of the
smaller: `G` is obtained from `F` by spreading each point out without moving its conditional mean.
That is the classical Strassen representation, and it is what makes the convex order tractable —
comparison results for queues become induction arguments on a coupling, rather than analytic
manipulations of convolutions of c.d.f.'s.
Both clauses are in the statement, as the book states them together: the `≤_cx` case with
`E[Y | X] = X` and the `≤_icx` case with `E[Y | X] ≥ X`.
**Both directions are asserted.** The hard one is the construction of the coupling; a
one-directional statement would be the easy half, since Jensen's inequality gives `F ≤_cx G` from
any such pair immediately.
`E[Y | X]` is a **conditional expectation**, not an equality of expectations: the whole content is
that the identity holds pointwise a.s. given `X`.
**Integrability is a hypothesis here and not in Theorem 4.2.1.** The convex order needs first
moments; the stochastic order does not. The two theorems are not harmonised.
The book attributes both to Strassen and does not prove them: "The next two theorems on the
pathwise representation of stochastic orders are proved in Strassen (1965)." -/
theorem strassen_cx {n : ℕ} (F G : Measure (Fin n → ℝ))
(hF : IsDistribution F) (hG : IsDistribution G)
(hFi : IsIntegrableDist F) (hGi : IsIntegrableDist G) :
(CxLe F G ↔
∃ (Ω : Type) (_ : MeasurableSpace Ω) (P : Measure Ω) (_ : IsProbabilityMeasure P)
(X Y : Ω → Fin n → ℝ),
Measurable X ∧ Measurable Y ∧
Measure.map X P = F ∧ Measure.map Y P = G ∧
condExp (MeasurableSpace.comap X inferInstance) P Y =ᵐ[P] X) ∧
(IcxLe F G ↔
∃ (Ω : Type) (_ : MeasurableSpace Ω) (P : Measure Ω) (_ : IsProbabilityMeasure P)
(X Y : Ω → Fin n → ℝ),
Measurable X ∧ Measurable Y ∧
Measure.map X P = F ∧ Measure.map Y P = G ∧
∀ᵐ ω ∂P, CoordLe (X ω)
(condExp (MeasurableSpace.comap X inferInstance) P Y ω)) := by sorry
end PalmQueueing.Ordering
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.