Equation (8.5) — the tangent plane at any feasible point lies above the optimum
ProvedKellyStochasticNetworks.alpha_fair_tangentLet maximize the -fair objective over the feasible set, and let be any other feasible point with positive components. Then
The left-hand side is , since . The inequality is concavity read backwards: the tangent plane to at lies above everywhere, so , and since is the maximum, ; chaining the two gives .
This is equation (8.5), and it is the entire mechanism of the stability proof. The drift of the Lyapunov function has exactly the shape with the load vector ; the inequality applies as soon as is feasible, which is what the stability condition (8.2) asserts. The choice of is the whole idea — no term-by-term estimate would work, because some routes are over-served and some under-served.
Formalization Note The inequality is stated in the form it is used, as a bound on a sum rather than as an inner product, and for a fixed feasible rather than universally. Positivity of is needed for the real power to be the intended quantity.
import Mathlib import Definitions.Def_KellyStochasticNetworks_Wardrop import Definitions.Def_KellyStochasticNetworks_Congestion import Definitions.Def_KellyStochasticNetworks_FlowLevel
namespace KellyStochasticNetworks
theorem alpha_fair_tangent {J R : ℕ} (A : Fin J → Fin R → ℝ) (C : Fin J → ℝ)
(w n : Fin R → ℝ) (α : ℝ) (hα : 0 < α) (hw : ∀ r, 0 < w r) (hn : ∀ r, 0 < n r)
(X U : Fin R → ℝ) (hXpos : ∀ r, 0 < X r) (hUpos : ∀ r, 0 < U r)
(hXfeas : X ∈ networkFeasible A C) (hUfeas : U ∈ networkFeasible A C)
(hmax : IsMaxOn (alphaFairObjective w n α)
(networkFeasible A C ∩ {Y | ∀ r, 0 < Y r}) X) :
(∑ r, w r * n r ^ α * U r ^ (-α) * (U r - X r)) ≤ 0 := by sorry
end KellyStochasticNetworksConfirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.