Lemma 3.7 — a drift condition sufficient for positive recurrence
DisprovedProcessingNetworks.Stability.drift_condition_implies_positive_recurrentFor a state of the ambient Markov chain of a Markov representation (Assumption 3.1), write where (Eq. 3.4), and let denote expectation under .
Lemma 3.7. If there is a such that
then the ambient Markov chain is positive recurrent.
This is the book's bridge from a fluid-scale drift condition to the probabilistic notion of stability, foreshadowing the whole fluid-model machinery of Chapter 6: the hypothesis says that, started from a state of total size , the expected buffer content at the fluid-scaled time is — vanishing relative to the initial size.
Formalization note. is M.size x (Eq. 3.4); the limit
Tendsto ... (Filter.comap M.size Filter.atTop) ... formalizes "as ", taken
componentwise in ; and is the expectation under the conditional
measure ℙ[|{ω | M.X 0 ω = x}], following the book's own convention
(Definition D.2), with the hypothesis hsupp that every state carries positive initial mass so
that each is defined. The expectation is taken as a Lebesgue integral in
, so an infinite violates (3.5) as it should, rather
than being read as . The conclusion is positive recurrence of the continuous-time chain
(Definition D.15, PositiveRecurrent M.jump M.rate).
import Mathlib import Definitions.Def_ProcessingNetworks_Stability_MarkovRepresentation import Definitions.Def_ProcessingNetworks_Stability_StabilityConditions
namespace ProcessingNetworks.Stability
open MeasureTheory ProbabilityTheory Filter
open scoped ENNReal
/-- Lemma 3.7 (a drift-type sufficient condition for positive recurrence), Dai & Harrison, p. 48:
if there is a `δ > 0` such that, as `|x| → ∞` over states `x` of the ambient chain,
`(1/|x|) E_x[Z(|x|δ)] → 0` componentwise, then the ambient chain is positive recurrent. Here
`M.size x` is the book's `|x| = ∑ᵢ zᵢ` for `f x = (n, z)` (Eq. 3.4), `E_x[·]` is the expectation
under the chain started at `x`, i.e. under `ℙ[|{X(0) = x}]` (every state is given positive
initial mass by `hsupp`, so that these conditional laws are all defined), and the expectation is
the Lebesgue integral in `ℝ≥0∞`, so an infinite expectation counts as such rather than as `0`. -/
theorem drift_condition_implies_positive_recurrent
{Xstate : Type*} [Countable Xstate] {Ω : Type*} [MeasureSpace Ω] {I J : ℕ}
{N : ℝ → Ω → Fin J → ℕ} {Z : ℝ → Ω → Fin I → ℕ}
(M : MarkovRepresentation Xstate I J N Z)
(hsupp : ∀ x : Xstate, ℙ {ω | M.X 0 ω = x} ≠ 0)
(δ : ℝ) (hδ : 0 < δ)
(hdrift :
Tendsto (fun x : Xstate => fun i : Fin I =>
(∫⁻ ω, ((Z (M.size x * δ) ω i : ℕ) : ℝ≥0∞) ∂(ℙ[|{ω | M.X 0 ω = x}])) /
ENNReal.ofReal (M.size x))
(Filter.comap M.size Filter.atTop) (nhds (0 : Fin I → ℝ≥0∞))) :
PositiveRecurrent M.jump M.rate := by sorry
end ProcessingNetworks.Stability
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.