Proof of Theorem 2(i), pp. 39–40 — the P-SCP satisfies condition (49)
OpenManyServerQED.Scheduling.pscp_condition_49Consider the sequence of -class, -server queues with abandonment under Assumptions 1(i), 2 (cost , Hölder exponent , growth exponent ) and 3 (moment exponent ), with discount rate . Let be a classical solution of the HJB equation (41), and let be measurable with
Let with , and for each let solve the system equations (7) from under the proposed P-SCP , where when (31) holds and is given by the static priority rule otherwise (pp. 18–19). Fix and let be given by (47):
With
where , the P-SCP satisfies condition (49):
Since the limit is the zero path, convergence in distribution is stated equivalently as uniform convergence on compact time intervals in probability: for every and , .
This is the key step of the proof of Theorem 2(i) (pp. 39–40): it verifies that the P-SCP meets the hypothesis (49) of Theorem 4(ii), which together with Theorem 4(i) yields the asymptotic optimality of the P-SCP.
Formalization notes. The hypotheses are exactly those of the goal theorem pscp_asymptotically_optimal (Theorem 2(i)), plus the arbitrary of (47). The solution is hypothesized to solve (7) under the P-SCP feedback; Proposition 1 makes it unique and admissible.
import Mathlib import Definitions.Def_BellWilliams2001_ThresholdPolicy_Paths import Definitions.Def_ManyServerQED_Scheduling_Diffusion import Definitions.Def_ManyServerQED_Scheduling_Model import Definitions.Def_ManyServerQED_Scheduling_Policy open MeasureTheory ProbabilityTheory Filter Topology open scoped ENNReal NNReal
namespace ManyServerQED.Scheduling
open BellWilliams2001.ThresholdPolicy
/-- Proof of Theorem 2(i) (pp. 39–40), "(i) The P-SCP": under Assumptions 1(i), 2 and 3, with `f`
a `C²_pol` solution of the HJB equation (41) and `h : ℝ^k → 𝕊^k` a measurable selection of its
minimizers, the processes `X^{n,∗}` controlled by the proposed P-SCP
`Ψ^{n,∗} = X^{n,∗} − Φ̃ₙ(X^{n,∗})` (pp. 18–19), started from `X^{0,n}` with `X̂^{0,n} → x`, satisfy
condition (49): `∫₀^· e^{−γs} Kⁿ_s ds ⇒ 0`, where
`Kⁿ_s = b(X̂ⁿ_s, uⁿ_s) · Df(X̂ⁿ_s) + L(X̂ⁿ_s, uⁿ_s) − H(X̂ⁿ_s, Df(X̂ⁿ_s))` (48) and `uⁿ` is given by
(47) with a fixed `u₀ ∈ 𝕊^k`. Convergence to the zero path is stated as u.o.c. convergence in
probability. -/
theorem pscp_condition_49 {Ω : Type*} [MeasurableSpace Ω] {k : ℕ} [NeZero k]
(M : SystemSequence Ω k)
(Lt : (Fin k → ℝ) → (Fin k → ℝ) → ℝ) (ϱ mL mU γ : ℝ)
(hA2 : Assumption2 Lt ϱ mL) (hA3 : M.Assumption3 mL mU) (hγ : 0 < γ)
(f : (Fin k → ℝ) → ℝ) (hf : IsC2Pol f)
(hHJB : IsHJBSolution M.diffData (costOfTilde Lt) γ f)
(h : (Fin k → ℝ) → (Fin k → ℝ)) (hh_meas : Measurable h)
(hh_mem : ∀ y, h y ∈ stdSimplex ℝ (Fin k))
(hh_min : ∀ y, ∑ i, drift M.diffData y (h y) i * grad f y i + costOfTilde Lt y (h y) =
hamiltonian M.diffData (costOfTilde Lt) y (grad f y))
(u0 : Fin k → ℝ) (hu0 : u0 ∈ stdSimplex ℝ (Fin k))
(X0 : ℕ → Fin k → ℕ) (x : Fin k → ℝ)
(hX0 : Tendsto (fun n => M.Xhat0 n (X0 n)) atTop (𝓝 x))
(Xs : ℕ → Ω → ℝ → Fin k → ℕ)
(hXs : ∀ n, 1 ≤ n → M.IsControlled n (X0 n) (fun ω t i => (Xs n ω t i : ℝ))
(fun ω t i => (Xs n ω t i : ℝ) - M.pscpPhi h n (Xs n ω t) i)) :
UocInProb M.P
(fun n ω t => ![∫ s in (0 : ℝ)..t, Real.exp (-γ * s) *
(∑ i, drift M.diffData (M.Xhat n (fun ω t i => (Xs n ω t i : ℝ)) ω s)
(uProc n u0 (fun ω t i => (Xs n ω t i : ℝ))
(fun ω t i => (Xs n ω t i : ℝ) - M.pscpPhi h n (Xs n ω t) i) ω s) i *
grad f (M.Xhat n (fun ω t i => (Xs n ω t i : ℝ)) ω s) i +
costOfTilde Lt (M.Xhat n (fun ω t i => (Xs n ω t i : ℝ)) ω s)
(uProc n u0 (fun ω t i => (Xs n ω t i : ℝ))
(fun ω t i => (Xs n ω t i : ℝ) - M.pscpPhi h n (Xs n ω t) i) ω s) -
hamiltonian M.diffData (costOfTilde Lt) (M.Xhat n (fun ω t i => (Xs n ω t i : ℝ)) ω s)
(grad f (M.Xhat n (fun ω t i => (Xs n ω t i : ℝ)) ω s)))])
(fun _ => 0) := by sorry
end ManyServerQED.Scheduling