§3.3, p. 11 — ∂²F/∂w_j∂w_k = E[Δ_jk h(Z)] for j ≠ k, and ∂²F/∂w_j² = 0
ProvedDiffNeighbors.Bias.second_partialdifferences-in-neighborsmultilinear-extensionp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-strengthenedp2o-v1second-order-difference
Let with multilinear extension , and write
for the first partial derivative , where is the expectation under independent , extended multilinearly to all . Then for every :
- for ,
- .
So the Hessian of has the expected second-order finite differences off the diagonal and zeros on it; the proof of Theorem 1 bounds the Taylor remainder through these entries.
Formalization Note The second derivative is the derivative of in the coordinate , stated as HasDerivAt of at , with written out as the difference of the two multilinear extensions.
Differences from the source:
- generalization: F^a of node i, for j, k ∈ N_i → the multilinear extension of any h : {0,1}^ι → ℝ, at any w ∈ ℝ^ι, for all coordinates j, k; reason: specializes with h = condOutcome f i a and j, k ∈ N_i
- encoding: ∂²F/∂w_j∂w_k and E[Δ_jk f^a(Z)] → HasDerivAt in the k-th coordinate of the first partial D_j of first_partial, with value mlExt (secondDiff h j k) w; reason: the mixed partial derivative; E_w[Δ_jk h(Z)] equals the four conditional expectations by linearity and independence
- 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
Preamble
import Mathlib import Definitions.Def_DiffNeighbors_Bias_Model
Formal statement
namespace DiffNeighbors.Bias
open Finset
/-- **Second partial derivatives of the multilinear extension**, §3.3, p. 11. Write
`D_j(w) = E_w[h(Z) | Z_j = 1] − E_w[h(Z) | Z_j = 0]` for the first partial derivative `∂F/∂w_j`
(p. 10). Then (i) for `j ≠ k`, `∂²F/∂w_j∂w_k (w) = E_w[Δ_jk h(Z)]`, and (ii) `∂²F/∂w_j² = 0`. -/
theorem second_partial {ι : Type*} [Fintype ι] [DecidableEq ι] (h : (ι → Bool) → ℝ) (w : ι → ℝ)
(j : ι) :
(∀ k, j ≠ k →
HasDerivAt (fun t => mlExt (fun z => h (Function.update z j true)) (Function.update w k t)
- mlExt (fun z => h (Function.update z j false)) (Function.update w k t))
(mlExt (fun z => secondDiff h j k z) w) (w k)) ∧
HasDerivAt (fun t => mlExt (fun z => h (Function.update z j true)) (Function.update w j t)
- mlExt (fun z => h (Function.update z j false)) (Function.update w j t))
0 (w j) := by sorry
end DiffNeighbors.Bias
Source
Peng, Ye, Zheng, Differences-in-Neighbors for Network Interference in Experiments, arXiv:2503.02271v1, p. 11, §3.3 (proof of Theorem 1), display '∂²F^a/∂w_j∂w_k = … = E[Δ_jk f^a(Z)] for j, k ∈ N_i and j ≠ k, and ∂²F/∂w_j² = 0 for j = k'
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.