Theorem 1, p. 9 — if every conditional outcome function is ϵ-smooth, the DN estimator's bias is at most d²ϵ
ProvedDiffNeighbors.Bias.theorem_1Consider units with an undirected interference graph without self-loops, maximum degree , and potential outcome functions satisfying Assumption 1 (neighborhood interference: depends only on and ). Treatments are assigned by the Bernoulli design, i.i.d. Bernoulli with . Let
and let be the Differences-in-Neighbors estimator (3). Suppose that the conditional outcome functions and are -smooth for all nodes , i.e. for all , , , . Then
The bound is second order in the strength of interference: under the paper's outcome model with interference magnitude it gives bias (Corollary 1), against first order for the difference-in-means estimator, while the DN estimator's variance stays polynomial in .
Differences from the source:
- encoding: units [N] = {1,…,N}, treatment vectors z ∈ {0,1}^N, an undirected graph without self-loops, d = max_i |N_i| → a finite type ι with N = card ι, z : ι → Bool with 1/0 read through ind, G : SimpleGraph ι, N_i = G.neighborFinset i, d = G.maxDegree; reason: a simple graph is exactly an undirected graph without self-loops; same instances
- encoding: E[·] over z_i i.i.d. Bernoulli(p) → bernExp p h = Σ_z (Π_j (p if z_j = 1 else 1 − p)) h(z), a finite sum; reason: the law of the design on the finite set {0,1}^N; same values
- encoding: a Bernoulli(p) design whose weights divide by p and 1 − p → hypotheses 0 < p and p < 1; reason: the weights 1/p and 1/(1 − p) and the design presuppose p ∈ (0,1); this only excludes undefined cases
- encoding: Assumption 1, stated "throughout this work" (p. 5) → an explicit hypothesis hf : NeighborhoodInterference G f; reason: a standing assumption of the paper made explicit; Theorem 1 is false without it
- encoding: f^a_i as a function of z_{N_i}; E[f^a(Z) | Z_j = b] and E[f^a(Z) | Z_j = b, Z_k = c] → condOutcome f i a z = f_i(z with z_i := a), a function on all of {0,1}^N; conditional expectations are bernExp of the function with the conditioned coordinates overwritten by Function.update; reason: under Assumption 1 condOutcome f i a depends only on z_{N_i}; by independence of the coordinates, conditioning on Z_j = b is the same as setting z_j := b
import Mathlib import Definitions.Def_DiffNeighbors_Bias_Model
namespace DiffNeighbors.Bias
open Finset
/-- **Theorem 1** (Peng–Ye–Zheng, arXiv:2503.02271v1, p. 9). Under neighborhood interference
(Assumption 1, standing throughout the paper) and the Bernoulli(`p`) design, if every conditional
outcome function `f^1_i`, `f^0_i` is `ε`-smooth, then the bias of the DN estimator is at most `d²ε`,
where `d = max_i |N_i|` is the maximum degree of the interference graph. -/
theorem theorem_1 {ι : Type*} [Fintype ι] [DecidableEq ι] (G : SimpleGraph ι) [DecidableRel G.Adj]
(f : ι → (ι → Bool) → ℝ) (hf : NeighborhoodInterference G f)
(p ε : ℝ) (hp0 : 0 < p) (hp1 : p < 1)
(hε : ∀ i, ∀ a : Bool, IsSmooth ε (condOutcome f i a)) :
|ate f - bernExp p (dnEstimator G p f)| ≤ (G.maxDegree : ℝ) ^ 2 * ε := by sorry
end DiffNeighbors.Bias
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.