Theorem 6, p. 24 — for every τ > 0, E[Δ̂^τ_SP] = Ψ′_τ(ρ)
ProvedMarketInterference.FiniteSample.theorem_6Fix a matching network (any finite set of intermediate nodes, any arc values, any capacities ), a control demand rate , a treatment effect with , a supply rate , a treatment fraction and a density . Let be the finite-sample partial-treatment value function, and let be the shadow-price estimator computed from the experiment's independent control counts , treatment counts and supply counts , with the negative-demand subgradient of the experiment LP. Then is differentiable at and
In expectation, the shadow-price estimator measures the slope of the partial-treatment value function at the experiment state, at every finite market size and with no uniqueness assumption on the optimal duals. Combined with a bound on it bounds the estimator's expected bias (Corollary 2).
Formalization Note The expectation of is taken under the product law of the control, treatment and supply counts (not under the law of the total demand). The derivative is asserted with HasDerivAt, so differentiability is part of the conclusion. Assumptions 1–3 of the paper (random tie-breaking, unique optima of the fluid LP, sign-consistent ) are not hypotheses: the proof on pp. 53–55 uses none of them, and the selection is what makes the identity hold without uniqueness (p. 25). No integrality of capacities is assumed.
import Mathlib import Definitions.Def_MarketInterference_FiniteSample_Experiment
namespace MarketInterference.FiniteSample
open MeasureTheory ProbabilityTheory
theorem theorem_6 {nd ns : ℕ} {I E : Type} [Fintype E] [DecidableEq I]
(N : MarketInterference.FluidBias.Network nd ns I E) (lam beta : Fin nd → ℝ) (pi : Fin ns → ℝ)
(hlam : 0 ≤ lam) (hlamb : 0 ≤ lam + beta) (hpi : 0 ≤ pi)
(ρ : ℝ) (hρ0 : 0 < ρ) (hρ1 : ρ < 1) (τ : ℝ) (hτ : 0 < τ) :
HasDerivAt (PsiTau N lam beta pi τ)
(∫ z, SPest N ρ τ z.1.1 z.1.2 z.2 ∂(experimentLaw lam beta pi ρ τ)) ρ := by sorry
end MarketInterference.FiniteSample
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.