Experimentation.SuperPopulation.MeanCLT
This roll-up imports the construction of centered, variance-normalized network fields, the hypotheses that make those fields mean-zero and unit-variance, and the final CLT turning the abstract m-dependent network-sum the
Field 3 core · 0 supporting To make the abstract m-dependent network CLT (networkSum_clt) usable for a concrete *mean*-type estimand we must first turn raw network-dependent outcomes Y i : Ω → ℝ (with a common mean and a positive sum-variance s² = ★ centeredNormalizedField_X★ centeredNormalizedField_adj
Centered/normalized network field for a super-population mean
To make the abstract m-dependent network CLT (networkSum_clt) usable for a concrete *mean*-type
estimand we must first turn raw network-dependent outcomes Y i : Ω → ℝ (with a common mean and a
positive sum-variance s² = Var(∑ᵢ Yᵢ)) into a NetworkDependence field whose summand is the
standardized contribution Xᵢ = (Yᵢ − E[Yᵢ]) / s. This file performs that construction.
The only nontrivial field obligation is indep: the standardized tuple fun k ∈ A => Xₖ is an
affine, measurable function of the outcome tuple fun k ∈ A => Yₖ, so the outcome-level
m-dependence (non-adjacent outcome tuples independent) transfers to the standardized field by
IndepFun.comp. The remaining structure data (the reflexive/symmetric network and measurability)
is inherited verbatim from the outcome-level hypotheses.
The three field hypotheses required by networkSum_clt — mean-zero, unit total variance, and the
uniform summand bound — are derived (not re-assumed) in Hypotheses.lean; the final CLT corollary
networkMean_clt lives in MeanCLT.lean.
Given a finite population of units, a measurable sample space, a measure on that space, a real-valued outcome for every unit and sample point, a decidable relation between units, the assumption that every unit is related to itself, the assumption that this relation is symmetric, the assumption that every outcome is measurable, the assumption that outcome vectors for any two finite sets with no relation joining them are independent, and a real normalizing constant, the centered and normalized network field has, for each unit, its outcome minus its expectation under the given measure, divided by that constant; it retains the supplied relation and has independent vectors on finite sets with no relation joining them.
Definition (Lean source)
The standardized summand for unit i at sample point ω is the centered outcome Yᵢ ω − E[Yᵢ] divided by the normalizing constant s.
Formal statement
Proof (Lean source)
The interference network underlying the standardized field is exactly the supplied network adj.
Formal statement
Proof (Lean source)
Hypotheses 1 core · 2 supporting The abstract network CLT networkSum_clt consumes three facts about the summand field: the summands are mean-zero, the network sum has unit total variance, and the summands are uniformly bounded. ★ centeredNormalizedField_sq_integral
Field hypotheses for the standardized network field
The abstract network CLT networkSum_clt consumes three facts about the summand field: the
summands are mean-zero, the network sum has unit total variance, and the summands are
uniformly bounded. For the standardized field centeredNormalizedField (summand
Xᵢ = (Yᵢ − E[Yᵢ]) / s) all three are derived here from the outcome-level assumptions — they are
not re-assumed.
* centeredNormalizedField_integral_eq_zero — mean-zero: E[Xᵢ] = (E[Yᵢ] − E[Yᵢ])/s = 0.
* centeredNormalizedField_sq_integral — unit total variance: with s² = Var(∑ᵢ Yᵢ),
∫ (∑ᵢ Xᵢ)² = Var(∑ᵢ Yᵢ)/s² = 1.
* centeredNormalizedField_abs_le — uniform bound: |Xᵢ| ≤ 2c/s whenever |Yᵢ − E[Yᵢ]| ≤ c.
Unit total variance. If each outcome is square-integrable, the normalizing constant s is positive, and s² equals the variance of the network sum of outcomes, s² = Var(∑ᵢ Yᵢ), then the standardized network sum ∑ᵢ Xᵢ = (∑ᵢ Yᵢ − ∑ᵢ E[Yᵢ]) / s has unit total variance: ∫ (∑ᵢ Xᵢ)² = 1. This is the field-variance hypothesis of networkSum_clt (∫ (depSum X)² = 1), derived from the outcome sum-variance.
Formal statement
Proof (Lean source)
2 supporting declarations (lemmas, instances)
-
centeredNormalizedField_integral_eq_zerotheorem — Mean-zero summands. Each standardized summand has integral zero: E[Xᵢ] = (E[Yᵢ] − E[Yᵢ]) / s = 0. (Uses integrability of Yᵢ, from MemLp Yᵢ 2.)hypothesesV :sharedType u_1Y :sharedV → Ω → ℝadj :sharedV → V → Prophrefl :shared∀ (i : V), adj i ihsymm :shared∀ (i j : V)ifadj i jthenadj j ihindepY :shared∀ (A B : Finset V)if(∀ a ∈ A, ∀ b ∈ B, ¬adj a b)thenIndepFun (fun ω k => Y (↑k) ω) (fun ω k => Y (↑k) ω) μhL2 :∀ i, MemLp (Y i) 2 μs :ℝi :Vconclusion∫ ω, (centeredNormalizedField Y adj hrefl hsymm hmeasY hindepY s).X i ω ∂μ = 0Proof (Lean source)
theorem centeredNormalizedField_integral_eq_zero (hL2 : ∀ i, MemLp (Y i) 2 μ) (s : ℝ) (i : V) : ∫ ω, (centeredNormalizedField Y adj hrefl hsymm hmeasY hindepY s).X i ω ∂μ = 0 := by simp only [centeredNormalizedField_X] rw [integral_div] rw [integral_sub ((hL2 i).integrable (by norm_num)) (integrable_const _)] simp [integral_const, probReal_univ] -
centeredNormalizedField_abs_letheorem — Uniform summand bound. If the centered outcomes are bounded, |Yᵢ − E[Yᵢ]| ≤ c, then each standardized summand satisfies |Xᵢ| ≤ 2c/s (with s > 0). The tight bound is c/s, weakened to 2c/s to match the engine's card·Bₙ³ → 0 smallness with Bₙ = 2c/s.hypothesesV :sharedType u_1Y :sharedV → Ω → ℝadj :sharedV → V → Prophrefl :shared∀ (i : V), adj i ihsymm :shared∀ (i j : V)ifadj i jthenadj j ihindepY :shared∀ (A B : Finset V)if(∀ a ∈ A, ∀ b ∈ B, ¬adj a b)thenIndepFun (fun ω k => Y (↑k) ω) (fun ω k => Y (↑k) ω) μs :ℝhs_pos :0 < sc :ℝhc :∀ i ω, |Y i ω - ∫ x, Y i x ∂μ| ≤ ci :Vω :Ωconclusion|(centeredNormalizedField Y adj hrefl hsymm hmeasY hindepY s).X i ω| ≤ 2 * c / sProof (Lean source)
theorem centeredNormalizedField_abs_le (s : ℝ) (hs_pos : 0 < s) (c : ℝ) (hc : ∀ i ω, |Y i ω - ∫ x, Y i x ∂μ| ≤ c) (i : V) (ω : Ω) : |(centeredNormalizedField Y adj hrefl hsymm hmeasY hindepY s).X i ω| ≤ 2 * c / s := by simp only [centeredNormalizedField_X] rw [abs_div, abs_of_pos hs_pos] have hc_nonneg : 0 ≤ c := le_trans (abs_nonneg _) (hc i ω) calc |Y i ω - ∫ x, Y i x ∂μ| / s ≤ c / s := by exact div_le_div_of_nonneg_right (hc i ω) hs_pos.le _ ≤ 2 * c / s := by exact div_le_div_of_nonneg_right (by nlinarith) hs_pos.le
MeanCLT 1 core · 0 supporting This is the consumer-facing bridge that makes the abstract m-dependent network CLT (Causalean.Experimentation.SuperPopulation.networkSum_clt) usable for raw network-dependent outcomes with positive sum-variance. ★ networkMean_clt
Asymptotic normality of a centered super-population network sum
This is the consumer-facing bridge that makes the abstract m-dependent network CLT
(Causalean.Experimentation.SuperPopulation.networkSum_clt) usable for raw network-dependent
outcomes with positive sum-variance. It proves standardized asymptotic normality for the centered
network sum, subtracting the sum of the individual outcome means; a sample-mean formulation requires
an additional external common-mean or normalization rewrite.
networkMean_clt assembles the pieces: it builds the standardized field
(centeredNormalizedField), discharges the engine's three hypotheses (mean-zero, unit total
variance, uniform bound — proved in Hypotheses.lean), derives the summand-size negligibility
B n → 0 from the population-level smallness card(Vₙ)·(cₙ/sₙ)³ → 0 (using that a positive
sum-variance forces a nonempty population), feeds networkSum_clt, and rewrites the engine's
pushforward CDF into the studentized probability set
{ω | (∑ᵢ Yₙ ᵢ − ∑ᵢ E[Yₙ ᵢ]) / sₙ ≤ t}.
The reduction follows Causalean.Experimentation.DesignBased.prodDesign_clt.
Asymptotic normality of a centered super-population network sum. Fix a family of outcomes Y n over probability spaces with measures μ n and a reflexive, symmetric adjacency relation adj n recording which units interfere, with every outcome measurable, outcome tuples on non-adjacent unit sets independent — i.e. m-dependence — and adjacency degree bounded by m. Assume every outcome is square-integrable, a positive normalizing constant s n with (s n)² = Var(∑ᵢ Yₙᵢ), outcomes uniformly bounded around their means by a sequence c n, and the negligibility rate card(Vₙ)·(cₙ/sₙ)³ → 0. Then the network sum, centered by subtracting the sum of the individual outcome means and divided by s n, converges in distribution to the standard normal.