Introduction
Social and economic networks, such as risk-sharing ties among households [26 , 25 ] and job referral networks [66 ] , shape economic outcomes [49 ] . A large literature models how they form; see Graham (2020) and de Paula et al. [20 ] for overviews. Conditionally independent dyad (CID) models, in which links form independently across pairs of agents given the agents’ observed and unobserved characteristics, are a large class of network formation models. This class is motivated by the Aldous–Hoover representation of exchangeable random graphs [1 , 45 ] and includes stochastic block models, latent space models, and models with degree heterogeneity [5 ] [42 , 32 , 54 ] . In parametric CID models, the probability that two agents link is a known function, such as a logit, of their observed characteristics and a finite-dimensional parameter; some models add agent-specific unobserved heterogeneity [32 , 51 , 23 ] (Graham 2024). These models are used to measure homophily and degree heterogeneity, to simulate counterfactual networks, and to serve as building blocks of peer-effects models with endogenous networks [52 , 47 ] . Their conclusions rest on the parametric specification, yet few tools exist to test it.
This paper develops a specification test for parametric CID models. The test uses subgraphs: the patterns of links among small groups of agents, such as triangles and stars. Under correct specification, the expected frequency of any subgraph among agents with given characteristics must equal its model-implied probability. Specifically, for a graph S S S on p p p vertices, let P ( S , x ) P(S,x) P ( S , x ) be the probability that the links among p p p agents with continuously distributed characteristics x = ( x 1 , … , x p ) x = (x_1,\ldots,x_p) x = ( x 1 , … , x p ) form S S S . Under the parametric model, links are independent given observed characteristics, and two agents with characteristics x i x_i x i and x j x_j x j link with probability e ( x i , x j ; θ ) e(x_i,x_j;\theta) e ( x i , x j ; θ ) . Hence P ( S , x ) P(S,x) P ( S , x ) is a known function of these link probabilities, which I denote by P ( S , x ; e θ ) P(S,x;e_\theta) P ( S , x ; e θ ) , where e θ : = e ( ⋅ , ⋅ ; θ ) e_\theta:= e(\cdot,\cdot;\theta) e θ := e ( ⋅ , ⋅ ; θ ) . For instance, three agents with characteristics ( x 1 , x 2 , x 3 ) (x_1,x_2,x_3) ( x 1 , x 2 , x 3 ) form a triangle with probability e ( x 1 , x 2 ; θ ) e ( x 1 , x 3 ; θ ) e ( x 2 , x 3 ; θ ) e(x_1,x_2;\theta)e(x_1,x_3;\theta)e(x_2,x_3;\theta) e ( x 1 , x 2 ; θ ) e ( x 1 , x 3 ; θ ) e ( x 2 , x 3 ; θ ) . The null hypothesis is that P ( S , x ) = P ( S , x ; e θ ) P(S,x)=P(S,x;e_\theta) P ( S , x ) = P ( S , x ; e θ ) for some θ \theta θ and almost all x x x . I estimate P ( S , x ) P(S,x) P ( S , x ) nonparametrically by the kernel-weighted frequency of S S S among tuples of agents whose characteristics are close to x x x [22 ] . I estimate θ \theta θ by the pseudo composite likelihood estimator of Graham (2024), which yields the model-implied estimate P ( S , x ; e θ ^ ) P(S,x;e_{\hat{\theta}}) P ( S , x ; e θ ^ ) . As in kernel-based specification tests (e.g., [39 ] ), the test statistic is the weighted integrated squared difference between the two estimates,
T N , S = N h N d X / 2 ∫ { P ^ N ( S , x ) − K h , N P ( S , ⋅ ; e θ ^ ) ( x ) } 2 π ( x ) d x , T_{N,S}=Nh_N^{d_X/2}\int\left\{\widehat{P}_N(S,x)-K_{h,N}P(S,\cdot;e_{\widehat{\theta}})(x)\right\}^2\pi(x)\,dx, T N , S = N h N d X /2 ∫ { P N ( S , x ) − K h , N P ( S , ⋅ ; e θ ) ( x ) } 2 π ( x ) d x ,
where N N N is the number of agents, h N h_N h N is the bandwidth, d X d_X d X is the dimension of each agent’s characteristics, P ^ N ( S , x ) \widehat{P}_N(S,x) P N ( S , x ) is the nonparametric kernel estimate, and π \pi π is a weight function. The operator K h , N K_{h,N} K h , N smooths the parametric estimate with the same kernel weights, which removes a linear term from the asymptotic expansion of the statistic. The test applies to any subgraph. Test statistics for several subgraphs can be combined by taking their maximum or a weighted average; wheels, subgraphs with one central agent and several arms, are a natural family as their moments identify stochastic block models [6 ] . The asymptotic theory in this paper is for a single subgraph.
The main technical challenge is network dependence. The observations are the subgraphs induced by p p p -tuples of agents. Two such subgraphs are independent when their tuples are disjoint but dependent when they share an agent. I first show that the parametric pseudo composite likelihood estimator is consistent and asymptotically normal at the N \sqrt{N} N rate under this dependence, provided the model is correctly specified. By the results of Ding [22 ] , the kernel estimate is uniformly consistent. I then write the test statistic as a kernel-weighted generalized U-statistic of order 2 p 2p 2 p , derive its Hoeffding decomposition, and isolate its leading term, a quadratic form. Applying the central limit theorem for generalized quadratic forms of de Jong [18 ] , I derive the asymptotic null distribution of the test statistic when its agent-level projection, the first-order Hoeffding projection onto individual agents, is nondegenerate. As with the rate of the kernel estimator in Ding [22 ] , the normalization N h N d X / 2 Nh_N^{d_X/2} N h N d X /2 depends only on the dimension d X d_X d X of a single agent’s characteristics, not on the dimension p d X p d_X p d X of all characteristics in the subgraph; this alleviates the curse of dimensionality.
Related literature. To the best of my knowledge, this paper provides the first test of parametric network formation models against nonparametric alternatives that uses subgraph probabilities conditional on the agents’ characteristics. It contributes to three literatures.
First, it extends specification testing to network data. Specification testing has a long tradition in econometrics ([65 , 40 , 72 , 62 ] ), and consistent tests of conditional moment restrictions against nonparametric alternatives originate with Bierens [8 , 9 ] . The test here builds on kernel-based comparisons of parametric and nonparametric fits ([39 , 31 , 73 , 27 , 28 ] ); related tests use series estimators and adaptive procedures ([44 , 46 , 11 ] ), integrated conditional moments and empirical processes ([71 , 2 , 68 , 15 , 24 ] ), or empirical likelihood ([70 ] ). These tests were developed for independent or weakly dependent observations ([19 ] ), and recent work extends them to spatial autoregressive models ([55 ] ) and to regressors whose distributions have singular components ([53 ] ). I extend the approach to observations that are overlapping tuples of agents in a single network.
Second, it contributes to the econometrics of network formation. Parametric CID models are studied by Graham [32 ] , Jochmans [51 ] , and Dzemski [23 ] , who allow for degree heterogeneity, and by Graham [34 ] , who studies logistic regression under sparse network asymptotics and possible misspecification. Kitamura and Laage [54 ] estimate stochastic block models with covariates, and Gao et al. [30 ] and Bonhomme and Dano [10 ] derive restrictions on model parameters that are robust to the form of unobserved heterogeneity. Models with strategic interaction in link formation, which are not CID models, are studied by Mele [59 ] , de Paula et al. [21 ] , Sheng [67 ] , and Menzel [61 ] , among others. Existing tests in this literature address specific questions: Pelican and Graham [64 ] and Graham and Pelican [37 ] test for strategic interaction, Dzemski [23 ] compares observed and predicted transitivity, and Auerbach [3 ] tests whether two networks are drawn from the same random graph model. The test proposed here instead checks the parametric specification of link probabilities against nonparametric alternatives.
Third, it builds on the statistical literature on exchangeable random graphs and graph limits ([1 , 45 , 5 , 58 ] ). Subgraph counts have long been used to describe network structure ([43 ] ), and Chandrasekhar and Jackson [13 ] model network formation directly in terms of subgraphs rather than links. Bickel et al. [6 ] use subgraph counts to estimate graphon models by the method of moments, and Bhattacharyya and Bickel [4 ] , Green and Shalizi [38 ] , and Lin et al. [57 ] develop resampling methods for them. Goodness-of-fit procedures include simulation-based diagnostics for exponential random graph models [48 ] , tests for the number of communities in stochastic block models [7 , 56 ] , and tests based on small subgraph probabilities [29 ] . These procedures use unconditional subgraph counts or target particular model classes. The test here instead conditions on the agents’ characteristics, so it allows for covariates and a parametric null. The analysis uses the Hoeffding decomposition of generalized U-statistics, whose asymptotic distributions in random graphs are studied by Janson and Nowicki [50 ] . It also relates to work on dyadic data [36 , 35 , 12 ] and on exchangeable arrays and multiway clustering [69 , 17 , 60 , 16 ] . That work mostly concerns statistics indexed by pairs of agents, whereas the statistic here involves tuples of p p p agents; for these, the analysis builds on the uniform convergence results of Ding [22 ] for the kernel estimator.
Outline. Section 2 introduces the CID model. Section 3 derives the subgraph probabilities and their conditional moment restrictions, including explicit expressions for wheels. Section 4 states the null hypothesis, defines the test statistic, establishes the properties of the parametric estimator, and states those of the nonparametric estimator. Section 5 derives the asymptotic null distribution of the test statistic. Section 6 concludes.
Model
I first introduce some graph notation commonly used in studies of networks. Let G N = ( V ( G N ) , E ( G N ) ) G_N = (\mathcal{V}(G_N),\mathcal{E}(G_N)) G N = ( V ( G N ) , E ( G N )) be an undirected network (graph) without self-loops on N N N vertices, or agents. The vertex set V ( G N ) = { 1 , … , N } \mathcal{V}(G_N) = \{1,\ldots,N\} V ( G N ) = { 1 , … , N } denotes the set of all agents in the network. The edge list E ( G N ) = { { i , j } , { k , l } , … } \mathcal{E}(G_N) = \{\{i,j\},\{k,l\},\ldots\} E ( G N ) = {{ i , j } , { k , l } , … } records the unordered pairs of agents that are linked. I represent G N G_N G N by its N × N N \times N N × N adjacency matrix D = [ D i j ] i , j ∈ V ( G N ) D = [D_{ij}]_{i,j \in\mathcal{V}(G_N)} D = [ D ij ] i , j ∈ V ( G N ) , whose ( i , j ) (i,j) ( i , j ) element is 1 if { i , j } ∈ E ( G N ) \{i,j\} \in\mathcal{E}(G_N) { i , j } ∈ E ( G N ) and 0 otherwise. See Graham (2020) for details.
Let X i ∈ R d X X_i \in\mathbb{R}^{d_X} X i ∈ R d X be the observed characteristics with support X \mathcal{X} X and U i ∈ R U_i \in\mathbb{R} U i ∈ R the unobserved characteristics of agent i i i . Assume that the pairs { ( X i , U i ) } i = 1 N \{(X_i,U_i)\}_{i=1}^{N} {( X i , U i ) } i = 1 N are independent and identically distributed across agents. Assume also that U i U_i U i is independent of X i X_i X i and has a continuous distribution with density f U f_U f U , and that X i X_i X i has a continuous distribution. Define X ( G N ) = ( X 1 , X 2 , … , X N ) ′ ∈ R N d X X(G_N) = (X_1,X_2,\ldots,X_N)^\prime\in\mathbb{R}^{Nd_X} X ( G N ) = ( X 1 , X 2 , … , X N ) ′ ∈ R N d X and U ( G N ) = ( U 1 , U 2 , … , U N ) ′ ∈ R N U(G_N) = (U_1,U_2,\ldots,U_N)^\prime\in\mathbb{R}^N U ( G N ) = ( U 1 , U 2 , … , U N ) ′ ∈ R N . I consider the conditionally independent dyad (CID) model with covariates and unobserved heterogeneity, in which
D i j ∣ X i , X j , U i , U j ∼ Bernoulli ( h ( X i , X j , U i , U j ) ) D_{ij} \mid X_i, X_j, U_i, U_j \sim\operatorname{Bernoulli}\bigl(h(X_i,X_j,U_i,U_j)\bigr) D ij ∣ X i , X j , U i , U j ∼ Bernoulli ( h ( X i , X j , U i , U j ) )
independently for every dyad { i , j } \{i,j\} { i , j } with i < j i < j i < j , where h ( X i , X j , U i , U j ) h(X_i,X_j,U_i,U_j) h ( X i , X j , U i , U j ) is the probability that $i and j j j form a link. This function h h h is called the graphon ; it is measurable, takes values in [ 0 , 1 ] [0,1] [ 0 , 1 ] , and satisfies h ( x , y , u , v ) = h ( y , x , v , u ) h(x,y,u,v)=h(y,x,v,u) h ( x , y , u , v ) = h ( y , x , v , u ) for undirected networks. CID models induce a particular dependence structure across the rows and columns of the adjacency matrix: dyads that share an agent are dependent. The model is restrictive in some respects (for example, it rules out strategic interaction among agents), but it is general enough to cover many popular network models, such as stochastic block models, latent space models, and models with degree heterogeneity.
The CID model is a nonparametric model for large networks. It is motivated by the Aldous–Hoover representation theorem for exchangeable random graphs [1 , 45 ] , under which links in the limiting network form independently conditional on agent-specific variables, here the observed characteristics X X X and the latent variables U U U . Chatterjee and Diaconis [14 ] show that dense exponential random graph models are asymptotically of this form.
Under the model, the probability of observing a network d d d conditional on X = x X=x X = x and U = u U=u U = u is
Pr ( D = d ∣ X = x , U = u ) = ∏ i < j h ( x i , x j , u i , u j ) d i j ( 1 − h ( x i , x j , u i , u j ) ) 1 − d i j . \Pr(D=d\mid X=x,U=u)=\prod_{i<j}h(x_i,x_j,u_i,u_j)^{d_{ij}}\left(1-h(x_i,x_j,u_i,u_j)\right)^{1-d_{ij}}. Pr ( D = d ∣ X = x , U = u ) = i < j ∏ h ( x i , x j , u i , u j ) d ij ( 1 − h ( x i , x j , u i , u j ) ) 1 − d ij .
Integrating out the unobserved heterogeneity, whose distribution F F F has density f U f_U f U , gives the probability of observing d d d conditional on X = x X=x X = x :
Pr ( D = d ∣ X = x ) = ∫ ⋯ ∫ ∏ i < j h ( x i , x j , u i , u j ) d i j ( 1 − h ( x i , x j , u i , u j ) ) 1 − d i j ∏ i = 1 N f U ( u i ) d u i . \Pr(D=d\mid X=x)=\int\cdots\int\prod_{i<j}h(x_i,x_j,u_i,u_j)^{d_{ij}}\left(1-h(x_i,x_j,u_i,u_j)\right)^{1-d_{ij}}\prod_{i=1}^{N}f_U(u_i)\,du_i. Pr ( D = d ∣ X = x ) = ∫ ⋯ ∫ i < j ∏ h ( x i , x j , u i , u j ) d ij ( 1 − h ( x i , x j , u i , u j ) ) 1 − d ij i = 1 ∏ N f U ( u i ) d u i .
Without loss of generality, I apply the probability integral transform to normalize F F F to the uniform distribution on [ 0 , 1 ] [0,1] [ 0 , 1 ] , so that
Pr ( D = d ∣ X = x ) = ∫ 0 1 ⋯ ∫ 0 1 ∏ i < j h ( x i , x j , u i , u j ) d i j ( 1 − h ( x i , x j , u i , u j ) ) 1 − d i j ∏ i = 1 N d u i . \Pr(D=d\mid X=x)=\int_0^1\cdots\int_0^1\prod_{i<j}h(x_i,x_j,u_i,u_j)^{d_{ij}}\left(1-h(x_i,x_j,u_i,u_j)\right)^{1-d_{ij}}\prod_{i=1}^{N}du_i. Pr ( D = d ∣ X = x ) = ∫ 0 1 ⋯ ∫ 0 1 i < j ∏ h ( x i , x j , u i , u j ) d ij ( 1 − h ( x i , x j , u i , u j ) ) 1 − d ij i = 1 ∏ N d u i .
Networks may be dense or sparse. A common way to model sparsity [6 ] is to write h N ( x i , x j , u i , u j ) = ρ N w ( x i , x j , u i , u j ) ∧ 1 h_N(x_i,x_j,u_i,u_j)=\rho_N w(x_i,x_j,u_i,u_j)\wedge1 h N ( x i , x j , u i , u j ) = ρ N w ( x i , x j , u i , u j ) ∧ 1 , where w ≥ 0 w\ge0 w ≥ 0 is a fixed function with ∫ 0 1 ∫ 0 1 w ( x i , x j , u i , u j ) d u i d u j = 1 \int_0^1\int_0^1w(x_i,x_j,u_i,u_j)\,du_i\,du_j=1 ∫ 0 1 ∫ 0 1 w ( x i , x j , u i , u j ) d u i d u j = 1 , so that, up to the truncation at one, ρ N \rho_N ρ N is the average link probability among agents with covariates ( x i , x j ) (x_i,x_j) ( x i , x j ) . In sparse networks, ρ N → 0 \rho_N\to0 ρ N → 0 , so the expected number of links grows more slowly than N 2 N^2 N 2 . This paper studies dense networks, in which h h h does not depend on N N N ; extending the test to sparse networks is left for future work.
Conditional moments using subgraph counts
This section derives conditional moment restrictions from the frequency with which a given small pattern of links, a subgraph, appears in a network. I first introduce some definitions and notation for subgraphs.
Induced subgraphs and density
Definition 3.1 (Graph Isomorphism). Graphs R R R and S S S are isomorphic, written R ≃ S R \simeq S R ≃ S , if there is a bijection φ : V ( R ) → V ( S ) \varphi:\mathcal{V}(R) \to\mathcal{V}(S) φ : V ( R ) → V ( S ) such that { a , b } ∈ E ( R ) \{a,b\} \in\mathcal{E}(R) { a , b } ∈ E ( R ) if and only if { φ ( a ) , φ ( b ) } ∈ E ( S ) \{\varphi(a),\varphi(b)\} \in\mathcal{E}(S) { φ ( a ) , φ ( b )} ∈ E ( S ) .
I say that ( S , x ) (S,x) ( S , x ) is an induced subgraph of a graph G = ( V ( G ) , E ( G ) ) G=(\mathcal{V}(G),\mathcal{E}(G)) G = ( V ( G ) , E ( G )) on p p p vertices with covariates x ∈ R p d X x \in\mathbb{R}^{pd_X} x ∈ R p d X if (i) V ( S ) ⊆ V ( G ) \mathcal{V}(S) \subseteq\mathcal{V}(G) V ( S ) ⊆ V ( G ) contains p p p vertices; (ii) E ( S ) = E ( G ) ∩ { V ( S ) × V ( S ) } \mathcal{E}(S)=\mathcal{E}(G)\cap\{\mathcal{V}(S)\times\mathcal{V}(S)\} E ( S ) = E ( G ) ∩ { V ( S ) × V ( S )} ; and (iii) the stacked covariates of the vertices in V ( S ) \mathcal{V}(S) V ( S ) equal x x x . Although X i X_i X i is continuous, it is convenient to treat it as discrete when defining the empirical subgraph density; § 4.3 handles continuous covariates. I define the induced subgraph density of S S S with covariates x ∈ R p d X x \in\mathbb{R}^{pd_X} x ∈ R p d X in G N G_N G N to be
P N ( S , x ) = ∑ i p ∈ P p , N 1 ( S ≃ G N [ i p ] , X ( G N [ i p ] ) = x ) ∑ i p ∈ P p , N 1 ( X ( G N [ i p ] ) = x ) = # of p -tuples inducing S with covariates x in G N # of p -tuples with covariates x in G N . (1) P_N(S,x)=\frac{\sum_{\mathbf{i}_p\in\mathcal{P}_{p,N}}\mathbf{1}(S\simeq G_N[\mathbf{i}_p],X(G_N[\mathbf{i}_p])=x)}{\sum_{\mathbf{i}_p\in\mathcal{P}_{p,N}}\mathbf{1}(X(G_N[\mathbf{i}_p])=x)}
=\frac{\#\text{ of }p\text{-tuples inducing }S\text{ with covariates }x\text{ in }G_N}{\#\text{ of }p\text{-tuples with covariates }x\text{ in }G_N}. \tag*{(1)} P N ( S , x ) = ∑ i p ∈ P p , N 1 ( X ( G N [ i p ]) = x ) ∑ i p ∈ P p , N 1 ( S ≃ G N [ i p ] , X ( G N [ i p ]) = x ) = # of p -tuples with covariates x in G N # of p -tuples inducing S with covariates x in G N . (1)
where P p , N : = { ( i 1 , … , i p ) : i a ∈ [ N ] , i a ≠ i b for a ≠ b } \mathcal{P}_{p,N}:=\{(i_1,\ldots,i_p):i_a\in[N],\ i_a\ne i_b\text{ for }a\ne b\} P p , N := {( i 1 , … , i p ) : i a ∈ [ N ] , i a = i b for a = b } is the set of ordered p p p -tuples i p = ( i 1 , … , i p ) \mathbf{i}_p=(i_1,\ldots,i_p) i p = ( i 1 , … , i p ) of distinct vertices in [ N ] : = { 1 , … , N } [N]:=\{1,\ldots,N\} [ N ] := { 1 , … , N } , with cardinality M N : = ∣ P p , N ∣ = N ! / ( N − p ) ! M_N:=|\mathcal{P}_{p,N}|=N!/(N-p)! M N := ∣ P p , N ∣ = N ! / ( N − p )! for N ≥ p N\ge p N ≥ p , and G N [ i p ] G_N[\mathbf{i}_p] G N [ i p ] is the subgraph of G N G_N G N induced by the vertices in i p \mathbf{i}_p i p . The covariate vector is stacked in tuple order, X i p : = ( X i 1 , … , X i p ) X_{\mathbf{i}_p}:=(X_{i_1},\ldots,X_{i_p}) X i p := ( X i 1 , … , X i p ) , and the labels 1 , … , p 1,\ldots,p 1 , … , p refer to those positions. I also define C p , N : = { q p = ( q 1 , … , q p ) : 1 ≤ q 1 < ⋯ < q p ≤ N } \mathcal{C}_{p,N}:=\{\mathbf{q}_p=(q_1,\ldots,q_p):1\le q_1<\cdots<q_p\le N\} C p , N := { q p = ( q 1 , … , q p ) : 1 ≤ q 1 < ⋯ < q p ≤ N } , the set of unordered p p p -vertex subsets, each written in increasing order. For q p ∈ C p , N \mathbf{q}_p\in\mathcal{C}_{p,N} q p ∈ C p , N , I write a permutation of q p \mathbf{q}_p q p as σ ( q p ) \sigma(\mathbf{q}_p) σ ( q p ) for σ ∈ S ( q p ) \sigma\in\mathcal{S}(\mathbf{q}_p) σ ∈ S ( q p ) , where S ( q p ) \mathcal{S}(\mathbf{q}_p) S ( q p ) denotes the set of all p ! p! p ! permutations of q p \mathbf{q}_p q p .
The population counterpart of P N ( S , x ) P_N(S,x) P N ( S , x ) is the probability that p p p agents with covariates x x x induce a copy of S S S :
P ( S , x ) : = Pr { G N [ ( 1 , … , p ) ] ≃ S ∣ X ( 1 , … , p ) = x } . P(S,x):=\Pr\{G_N[(1,\ldots,p)]\simeq S\mid X_{(1,\ldots,p)}=x\}. P ( S , x ) := Pr { G N [( 1 , … , p )] ≃ S ∣ X ( 1 , … , p ) = x } .
To compute it, label the p p p agents 1 , … , p 1,\ldots,p 1 , … , p , so that agent a a a has covariate x a x_a x a , and let E p : = { { a , b } : 1 ≤ a < b ≤ p } \mathcal{E}_p:=\{\{a,b\}:1\le a<b\le p\} E p := {{ a , b } : 1 ≤ a < b ≤ p } be the set of all pairs among them. For a labelled graph R R R on [ p ] [p] [ p ] , the probability that the links among agents 1 , … , p 1,\ldots,p 1 , … , p are exactly those of R R R (every edge of R R R present and every other pair unlinked) is
P l a b ( R , x ) : = ∫ [ 0 , 1 ] p ∏ { a , b } ∈ E ( R ) h ( x a , x b , u a , u b ) ∏ { a , b } ∈ E p ∖ E ( R ) { 1 − h ( x a , x b , u a , u b ) } d u . P_{\mathrm{lab}}(R,x):=\int_{[0,1]^p}\prod_{\{a,b\}\in\mathcal{E}(R)}h(x_a,x_b,u_a,u_b)\prod_{\{a,b\}\in\mathcal{E}_p\setminus\mathcal{E}(R)}\{1-h(x_a,x_b,u_a,u_b)\}\,\mathrm{d}u. P lab ( R , x ) := ∫ [ 0 , 1 ] p { a , b } ∈ E ( R ) ∏ h ( x a , x b , u a , u b ) { a , b } ∈ E p ∖ E ( R ) ∏ { 1 − h ( x a , x b , u a , u b )} d u .
Let O ( S ) \mathcal{O}(S) O ( S ) be the set of distinct labelled graphs on [ p ] [p] [ p ] that are isomorphic to S S S . The agents induce a copy of S S S exactly when their links form one of the graphs in O ( S ) \mathcal{O}(S) O ( S ) . These events are disjoint, so
P ( S , x ) = ∑ R ∈ O ( S ) P l a b ( R , x ) . P(S,x)=\sum_{R\in\mathcal{O}(S)}P_{\mathrm{lab}}(R,x). P ( S , x ) = R ∈ O ( S ) ∑ P lab ( R , x ) .
For example, if S S S is a path with two edges (p = 3 p=3 p = 3 ), O ( S ) \mathcal{O}(S) O ( S ) contains three graphs, one for each choice of the middle vertex.
For subgraphs that are not necessarily induced, define the labelled injective homomorphism density
Q ( S , x ) : = ∫ [ 0 , 1 ] p ∏ { a , b } ∈ E ( S ) h ( x a , x b , u a , u b ) d u . Q(S,x) := \int_{[0,1]^p} \prod_{\{a,b\}\in\mathcal{E}(S)} h(x_a,x_b,u_a,u_b)\,du. Q ( S , x ) := ∫ [ 0 , 1 ] p { a , b } ∈ E ( S ) ∏ h ( x a , x b , u a , u b ) d u .
Q ( S , x ) Q(S,x) Q ( S , x ) is the probability that every edge of S S S , in its labelled position, is present, with no restriction on the other dyads. Note that P l a b ( S , x ) P_{\mathrm{lab}}(S,x) P lab ( S , x ) and Q ( S , x ) Q(S,x) Q ( S , x ) are closely related:
P l a b ( S , x ) = ∑ A ⊆ E p ∖ E ( S ) ( − 1 ) ∣ A ∣ Q ( S ∪ A , x ) , P_{\mathrm{lab}}(S,x) = \sum_{A\subseteq\mathcal{E}_p\setminus\mathcal{E}(S)} (-1)^{|A|} Q(S\cup A,x), P lab ( S , x ) = A ⊆ E p ∖ E ( S ) ∑ ( − 1 ) ∣ A ∣ Q ( S ∪ A , x ) ,
Q ( S , x ) = ∑ E ( S ) ⊆ E ( R ) ⊆ E p P l a b ( R , x ) . Q(S,x) = \sum_{\mathcal{E}(S)\subseteq\mathcal{E}(R)\subseteq\mathcal{E}_p} P_{\mathrm{lab}}(R,x). Q ( S , x ) = E ( S ) ⊆ E ( R ) ⊆ E p ∑ P lab ( R , x ) .
where S ∪ A S\cup A S ∪ A is the labelled graph with edge set E ( S ) ∪ A \mathcal{E}(S)\cup A E ( S ) ∪ A . These are the covariate-conditional versions of the identities underlying Proposition 1 in Bickel et al. [6 ] . The isomorphism probability P ( S , x ) P(S,x) P ( S , x ) then follows by summing P l a b ( R , x ) P_{\mathrm{lab}}(R,x) P lab ( R , x ) over R ∈ O ( S ) R\in O(S) R ∈ O ( S ) , as above.
Explicit formula for wheels
For some subgraphs, the homomorphism density Q ( S , x ) Q(S,x) Q ( S , x ) has an explicit form. Wheels are one example.
Definition 3.2 (Wheels). A ( k , l ) (k,l) ( k , l ) -wheel is a graph isomorphic to the graph on k l + 1 kl+1 k l + 1 vertices with the k l kl k l edges { { 1 , 2 } , … , { k , k + 1 } } ∪ { { 1 , k + 2 } , … , { 2 k , 2 k + 1 } } ∪ ⋯ ∪ { { 1 , ( l − 1 ) k + 2 } , … , { l k , l k + 1 } } \{\{1,2\},\ldots,\{k,k+1\}\}\cup\{\{1,k+2\},\ldots,\{2k,2k+1\}\}\cup\cdots\cup\{\{1,(l-1)k+2\},\ldots,\{lk,lk+1\}\} {{ 1 , 2 } , … , { k , k + 1 }} ∪ {{ 1 , k + 2 } , … , { 2 k , 2 k + 1 }} ∪ ⋯ ∪ {{ 1 , ( l − 1 ) k + 2 } , … , { l k , l k + 1 }} .
In other words, a wheel consists of a central vertex 1 and l l l arms attached to it, each a path of k k k edges. I consider only k ≥ 2 k\ge2 k ≥ 2 ; for k = 1 k=1 k = 1 , the wheel reduces to a star. For fixed covariates x , y x,y x , y , define the operator
[ T x , y f ] ( u ) : = ∫ 0 1 h ( x , y , u , v ) f ( v ) d v . [T_{x,y}f](u) := \int_0^1 h(x,y,u,v)f(v)\,dv. [ T x , y f ] ( u ) := ∫ 0 1 h ( x , y , u , v ) f ( v ) d v .
Label the noncentral vertices on arm m m m by v m , 1 , … , v m , k v_{m,1},\ldots,v_{m,k} v m , 1 , … , v m , k , and write x m , r x_{m,r} x m , r and u m , r u_{m,r} u m , r for their covariates and latent variables. Set x m , 0 = x 1 x_{m,0}=x_1 x m , 0 = x 1 and u m , 0 = u u_{m,0}=u u m , 0 = u , the shared center. Define the conditional arm moment
A m ( u ; x ) : = ∫ [ 0 , 1 ] k ∏ r = 1 k h ( x m , r − 1 , x m , r , u m , r − 1 , u m , r ) d u m , 1 ⋯ d u m , k . A_m(u;x) := \int_{[0,1]^k} \prod_{r=1}^{k} h(x_{m,r-1},x_{m,r},u_{m,r-1},u_{m,r})\,du_{m,1}\cdots du_{m,k}. A m ( u ; x ) := ∫ [ 0 , 1 ] k r = 1 ∏ k h ( x m , r − 1 , x m , r , u m , r − 1 , u m , r ) d u m , 1 ⋯ d u m , k .
Independence of the noncentral vertex variables across arms, conditional on the center, gives
Q ( S , x ) = ∫ 0 1 ∏ m = 1 l A m ( u ; x ) d u , Q(S,x) = \int_0^1 \prod_{m=1}^{l} A_m(u;x)\,du, Q ( S , x ) = ∫ 0 1 m = 1 ∏ l A m ( u ; x ) d u ,
A m ( u ; x ) = [ T x 1 , x m , 1 T x m , 1 , x m , 2 ⋯ T x m , k − 1 , x m , k 1 ] ( u ) . A_m(u;x) = [T_{x_1,x_{m,1}}T_{x_{m,1},x_{m,2}}\cdots T_{x_{m,k-1},x_{m,k}}1](u). A m ( u ; x ) = [ T x 1 , x m , 1 T x m , 1 , x m , 2 ⋯ T x m , k − 1 , x m , k 1 ] ( u ) .
The wheel moment thus reduces to a one-dimensional integral over the center of products of arm moments, each obtained by iterating the operators T x , y T_{x,y} T x , y . The induced probability P ( S , x ) P(S,x) P ( S , x ) of a wheel then follows. Wheels are also informative about the model: Theorem 1 of Bickel et al. [6 ] shows that the parameters of a stochastic block model are generically identified by finitely many wheel moments.
Network moments using subgraph counts
Write F X , N : = σ ( X 1 , … , X N ) \mathcal{F}_{X,N} := \sigma(X_1,\ldots,X_N) F X , N := σ ( X 1 , … , X N ) and, for each ordered tuple, define the induced-isomorphism and labelled homomorphism responses
Y i p , S : = 1 { G N [ i p ] ≃ S } , Z i p , S : = ∏ { a , b } ∈ E ( S ) D i a i b . Y_{i_p,S} := 1\{G_N[i_p] \simeq S\}, \qquad Z_{i_p,S} := \prod_{\{a,b\}\in E(S)} D_{i_a i_b}. Y i p , S := 1 { G N [ i p ] ≃ S } , Z i p , S := { a , b } ∈ E ( S ) ∏ D i a i b .
Under the CID model and the independence assumptions in § 2 , every ordered tuple satisfies the conditional moment restrictions
E [ Y i p , S − P ( S , X i p ) ∣ F X , N ] = 0 , \mathbb{E}[Y_{i_p,S}-P(S,X_{i_p})\mid\mathcal{F}_{X,N}]=0, E [ Y i p , S − P ( S , X i p ) ∣ F X , N ] = 0 ,
E [ Z i p , S − Q ( S , X i p ) ∣ F X , N ] = 0. \mathbb{E}[Z_{i_p,S}-Q(S,X_{i_p})\mid\mathcal{F}_{X,N}]=0. E [ Z i p , S − Q ( S , X i p ) ∣ F X , N ] = 0.
For discrete covariates, when the common denominator is positive, the conditional expectation of P N ( S , x ) P_N(S,x) P N ( S , x ) given F X , N \mathcal{F}_{X,N} F X , N is P ( S , x ) P(S,x) P ( S , x ) . With continuous covariates, exact matches occur with probability zero, and the tuple-level restrictions above motivate the kernel estimators below.
Model specification test
I define the integrated graphon function g : X × X → [ 0 , 1 ] g:\mathcal{X}\times\mathcal{X}\to[0,1] g : X × X → [ 0 , 1 ] by
g ( X i , X j ) ≡ E [ h ( X i , X j , U i , U j ) ∣ X i , X j ] g(X_i,X_j)\equiv\mathbb{E}[h(X_i,X_j,U_i,U_j)\mid X_i,X_j] g ( X i , X j ) ≡ E [ h ( X i , X j , U i , U j ) ∣ X i , X j ]
and parametrize it as g ( X i , X j ) = e ( X i , X j ; θ ) g(X_i,X_j)=e(X_i,X_j;\theta) g ( X i , X j ) = e ( X i , X j ; θ ) for a finite-dimensional θ \theta θ . A popular choice is the logit model of Graham [32 ] (2024), with θ = ( α , β ′ ) \theta=(\alpha,\beta') θ = ( α , β ′ ) :
e ( X i , X j ; θ ) = exp { α + ( X i + X j ) ′ β } 1 + exp { α + ( X i + X j ) ′ β } . e(X_i,X_j;\theta)=\frac{\exp\{\alpha+(X_i+X_j)'\beta\}}{1+\exp\{\alpha+(X_i+X_j)'\beta\}}. e ( X i , X j ; θ ) = 1 + exp { α + ( X i + X j ) ′ β } exp { α + ( X i + X j ) ′ β } .
More generally, the index can be R i j ′ θ R_{ij}'\theta R ij ′ θ for any symmetric vector of dyad regressors R i j = R j i R_{ij}=R_{ji} R ij = R j i . I maintain the hypothesis that the network follows a CID model. For a fixed subgraph S S S on p p p vertices, with X = ( X 1 , … , X p ) X=(X_1,\ldots,X_p) X = ( X 1 , … , X p ) , I test the null hypothesis
H 0 : Pr { P ( S , X ) = P ( S , X ; e θ ) } = 1 for some θ ∈ Θ , H_0:\quad\Pr\{P(S,X)=P(S,X;e_\theta)\}=1\text{ for some }\theta\in\Theta, H 0 : Pr { P ( S , X ) = P ( S , X ; e θ )} = 1 for some θ ∈ Θ ,
H 1 : Pr { P ( S , X ) = P ( S , X ; e θ ) } < 1 for all θ ∈ Θ . H_1:\quad\Pr\{P(S,X)=P(S,X;e_\theta)\}<1\text{ for all }\theta\in\Theta. H 1 : Pr { P ( S , X ) = P ( S , X ; e θ )} < 1 for all θ ∈ Θ.
Test statistics
The subgraph moments of § 3 are testable implications of the parametric model. I use them to construct a Härdle–Mammen-type test of the parametric null against a nonparametric alternative. For a fixed induced subgraph S S S , I consider the network moment P ( S , x ) P(S,x) P ( S , x ) . Under the parametric null indexed by θ \theta θ , the corresponding model-implied network moment for subgraph S S S with covariates x x x is
P ( S , x ; e θ ) = ∑ R ∈ O ( S ) ∏ { a , b } ∈ E ( R ) e ( x a , x b ; θ ) ∏ { a , b } ∈ E p ∖ E ( R ) { 1 − e ( x a , x b ; θ ) } . P(S,x;e_\theta)=\sum_{R\in\mathcal{O}(S)}\prod_{\{a,b\}\in E(R)}e(x_a,x_b;\theta)\prod_{\{a,b\}\in E_p\setminus E(R)}\{1-e(x_a,x_b;\theta)\}. P ( S , x ; e θ ) = R ∈ O ( S ) ∑ { a , b } ∈ E ( R ) ∏ e ( x a , x b ; θ ) { a , b } ∈ E p ∖ E ( R ) ∏ { 1 − e ( x a , x b ; θ )} .
Given an estimator θ ^ \hat{\theta} θ ^ , the parametric fit of P ( S , x ) P(S,x) P ( S , x ) is
P ( S , x ; e θ ^ ) = ∑ R ∈ O ( S ) ∏ { a , b } ∈ E ( R ) e ( x a , x b ; θ ^ ) ∏ { a , b } ∈ E p ∖ E ( R ) { 1 − e ( x a , x b ; θ ^ ) } . P(S,x;e_{\hat{\theta}})=\sum_{R\in\mathcal{O}(S)}\prod_{\{a,b\}\in E(R)}e(x_a,x_b;\hat{\theta})\prod_{\{a,b\}\in E_p\setminus E(R)}\{1-e(x_a,x_b;\hat{\theta})\}. P ( S , x ; e θ ^ ) = R ∈ O ( S ) ∑ { a , b } ∈ E ( R ) ∏ e ( x a , x b ; θ ^ ) { a , b } ∈ E p ∖ E ( R ) ∏ { 1 − e ( x a , x b ; θ ^ )} .
The nonparametric fit of P ( S , x ) P(S,x) P ( S , x ) is the kernel estimator P ^ N ( S , x ) \hat{P}_N(S,x) P ^ N ( S , x ) defined in § 4.3 , which also states its regularity conditions. The test statistic T N , S T_{N,S} T N , S for subgraph S S S is the weighted L 2 L^2 L 2 -distance between P ^ N ( S , x ) \hat{P}_N(S,x) P ^ N ( S , x ) and a smoothed version of P ( S , x ; e θ ^ ) P(S,x;e_{\hat{\theta}}) P ( S , x ; e θ ^ ) :
T N , S = N h N d X / 2 ∫ [ P ^ N ( S , x ) − K h , N P ( S , ⋅ ; e θ ^ ) ( x ) ] 2 π ( x ) d x , (2) T_{N,S}=Nh_N^{d_X/2}\int\left[\hat{P}_N(S,x)-\mathcal{K}_{h,N}P(S,\cdot;e_{\hat{\theta}})(x)\right]^2\pi(x)\,dx, \tag*{(2)} T N , S = N h N d X /2 ∫ [ P ^ N ( S , x ) − K h , N P ( S , ⋅ ; e θ ^ ) ( x ) ] 2 π ( x ) d x , (2)
where K h , N \mathcal{K}_{h,N} K h , N is the smoothing operator defined by
K h , N φ ( x ) = ∑ i p ∈ P p , N K h ( x − X i p ) φ ( X i p ) ∑ i p ∈ P p , N K h ( x − X i p ) . \mathcal{K}_{h,N}\varphi(x)=\frac{\sum_{i_p\in\mathcal{P}_{p,N}}K_h(x-X_{i_p})\varphi(X_{i_p})}{\sum_{i_p\in\mathcal{P}_{p,N}}K_h(x-X_{i_p})}. K h , N φ ( x ) = ∑ i p ∈ P p , N K h ( x − X i p ) ∑ i p ∈ P p , N K h ( x − X i p ) φ ( X i p ) .
with K h ( v ) : = h N − d W K ( v / h N ) K_h(v):=h_N^{-d_W}K(v/h_N) K h ( v ) := h N − d W K ( v / h N ) for a kernel K : R d W → R K:\mathbb{R}^{d_W}\to\mathbb{R} K : R d W → R on the stacked covariate space of dimension d W : = p d X d_W:=pd_X d W := p d X and a bandwidth h N → 0 h_N\to0 h N → 0 ; § 4.3 states the conditions on K K K and h N h_N h N . Note that E [ P ^ N ( S , x ) ∣ X 1 , … , X N ] = K h , N P ( S , ⋅ ) ( x ) \mathbb{E}[\hat{P}_N(S,x)\mid X_1,\ldots,X_N]=\mathcal{K}_{h,N}P(S,\cdot)(x) E [ P ^ N ( S , x ) ∣ X 1 , … , X N ] = K h , N P ( S , ⋅ ) ( x ) . Applying the same operator to P ( S , x ; e θ ^ ) P(S,x;e_{\hat{\theta}}) P ( S , x ; e θ ^ ) makes T N , S T_{N,S} T N , S asymptotically equivalent to a constant plus a quadratic form; without it, the statistic would also contain a linear term.
To use several subgraphs at once, for example all wheels in a set S \mathcal{S} S , one can take the maximum or a weighted average of the statistics over S \mathcal{S} S , with weights w ( S ) ≥ 0 w(S)\ge0 w ( S ) ≥ 0 :
T N max = N h N d X / 2 max S ∈ S ∫ ( P ^ N ( S , x ) − K h , N P ( S , ⋅ ; e θ ^ ) ( x ) ) 2 π ( x ) d x . T_N^{\max}=Nh_N^{d_X/2}\max_{S\in\mathcal{S}}\int\left(\hat{P}_N(S,x)-\mathcal{K}_{h,N}P(S,\cdot;e_{\hat{\theta}})(x)\right)^2\pi(x)\,dx. T N m a x = N h N d X /2 S ∈ S max ∫ ( P ^ N ( S , x ) − K h , N P ( S , ⋅ ; e θ ^ ) ( x ) ) 2 π ( x ) d x .
or
T N a v g = N h N d X / 2 ∑ S ∈ S w ( S ) ∫ ( P ^ N ( S , x ) − K h , N P ( S , ⋅ ; e θ ^ ) ( x ) ) 2 π ( x ) d x . T_N^{\mathrm{avg}}=Nh_N^{d_X/2}\sum_{S\in\mathcal{S}}w(S)\int\left(\hat{P}_N(S,x)-\mathcal{K}_{h,N}P(S,\cdot;e_{\hat{\theta}})(x)\right)^2\pi(x)\,dx. T N avg = N h N d X /2 S ∈ S ∑ w ( S ) ∫ ( P ^ N ( S , x ) − K h , N P ( S , ⋅ ; e θ ^ ) ( x ) ) 2 π ( x ) d x .
The asymptotic theory in § 5 is for a single subgraph S S S . The null distributions of T N max T_N^{\max} T N m a x and T ‾ N a v g \overline{T}^{\mathrm{avg}}_{N} T N avg depend on the joint limit of ( T N , S ) S ∈ S (T_{N,S})_{S\in\mathcal{S}} ( T N , S ) S ∈ S , which I do not derive here. Now I turn to estimation of the network moments.
Parametric estimation of the subgraph moments
Following Graham (2024), I estimate the parametric model by maximizing a pseudo composite log-likelihood:
θ ^ = arg max θ ∈ Θ L N ( θ ) = arg max θ ∈ Θ 1 N ( N − 1 ) ∑ i = 1 N ∑ j ≠ i ℓ i j , N ( θ ) , \hat{\theta}=\underset{\theta\in\Theta}{\arg\max}\ L_N(\theta)=\underset{\theta\in\Theta}{\arg\max}\ \frac{1}{N(N-1)}\sum_{i=1}^{N}\sum_{j\ne i}\ell_{ij,N}(\theta), θ ^ = θ ∈ Θ arg max L N ( θ ) = θ ∈ Θ arg max N ( N − 1 ) 1 i = 1 ∑ N j = i ∑ ℓ ij , N ( θ ) ,
where the logit kernel is ℓ i j , N ( θ ) ≡ ( 2 D i j − 1 ) ( R i j ′ θ ) − ln ( 1 + exp ( ( 2 D i j − 1 ) ( R i j ′ θ ) ) ) \ell_{ij,N}(\theta)\equiv(2D_{ij}-1)(R_{ij}^{\prime}\theta)-\ln\left(1+\exp\left((2D_{ij}-1)(R_{ij}^{\prime}\theta)\right)\right) ℓ ij , N ( θ ) ≡ ( 2 D ij − 1 ) ( R ij ′ θ ) − ln ( 1 + exp ( ( 2 D ij − 1 ) ( R ij ′ θ ) ) ) and R i j = ( 1 , ζ ( X i , X j ) ) ′ R_{ij}=(1,\zeta(X_i,X_j))^{\prime} R ij = ( 1 , ζ ( X i , X j ) ) ′ for a known symmetric transformation ζ \zeta ζ , that is, ζ ( x , y ) = ζ ( y , x ) \zeta(x,y)=\zeta(y,x) ζ ( x , y ) = ζ ( y , x ) . Write e i j ( θ ) : = Λ ( R i j ′ θ ) e_{ij}(\theta):=\Lambda(R_{ij}^{\prime}\theta) e ij ( θ ) := Λ ( R ij ′ θ ) with Λ ( t ) : = e t / ( 1 + e t ) \Lambda(t):=e^{t}/(1+e^{t}) Λ ( t ) := e t / ( 1 + e t ) , so that e i j ( θ ) = e ( X i , X j ; θ ) e_{ij}(\theta)=e(X_i,X_j;\theta) e ij ( θ ) = e ( X i , X j ; θ ) , and let s i j ( θ ) : = ∂ θ ℓ i j , N ( θ ) = R i j { D i j − e i j ( θ ) } s_{ij}(\theta):=\partial_{\theta}\ell_{ij,N}(\theta)=R_{ij}\{D_{ij}-e_{ij}(\theta)\} s ij ( θ ) := ∂ θ ℓ ij , N ( θ ) = R ij { D ij − e ij ( θ )} denote the dyad score. I show below that θ ^ \hat{\theta} θ ^ is N \sqrt{N} N -consistent and asymptotically normal, and I plug it into P ( S , x ; e θ ) P(S,x;e_{\theta}) P ( S , x ; e θ ) to obtain P ( S , x ; e θ ^ ) P(S,x;e_{\hat{\theta}}) P ( S , x ; e θ ^ ) . I impose the following regularity conditions.
Assumption 4.1 (Compactness and interior). The parameter space Θ \Theta Θ is compact and θ 0 ∈ int ( Θ ) \theta_0\in\operatorname{int}(\Theta) θ 0 ∈ int ( Θ ) .
Assumption 4.2 (Identification, smoothness, and nonsingularity). Define L ( θ ) ≡ E [ ℓ 12 , N ( θ ) ] \mathcal{L}(\theta)\equiv E[\ell_{12,N}(\theta)] L ( θ ) ≡ E [ ℓ 12 , N ( θ )] . The function L ( θ ) \mathcal{L}(\theta) L ( θ ) is uniquely maximized at θ 0 \theta_0 θ 0 , is twice continuously differentiable on an open neighborhood of θ 0 \theta_0 θ 0 , and H 0 ≡ ∂ 2 L ( θ 0 ) / ∂ θ ∂ θ ′ \mathcal{H}_0\equiv\partial^2\mathcal{L}(\theta_0)/\partial\theta\partial\theta^{\prime} H 0 ≡ ∂ 2 L ( θ 0 ) / ∂ θ ∂ θ ′ is nonsingular.
Assumption 4.3 (Bounded moments). Assume
E [ sup θ ∈ Θ ∥ ∂ ℓ 12 , N ( θ ) ∂ θ ∥ 2 ] < ∞ , E [ sup θ ∈ Θ ∥ ∂ 2 ℓ 12 , N ( θ ) ∂ θ ∂ θ ′ ∥ ] < ∞ . E\left[\sup_{\theta\in\Theta}\left\|\frac{\partial\ell_{12,N}(\theta)}{\partial\theta}\right\|^2\right]<\infty,\qquad E\left[\sup_{\theta\in\Theta}\left\|\frac{\partial^2\ell_{12,N}(\theta)}{\partial\theta\partial\theta^{\prime}}\right\|\right]<\infty. E [ θ ∈ Θ sup ∂ θ ∂ ℓ 12 , N ( θ ) 2 ] < ∞ , E [ θ ∈ Θ sup ∂ θ ∂ θ ′ ∂ 2 ℓ 12 , N ( θ ) ] < ∞. Moreover, for any θ , θ ˉ ∈ Θ \theta,\bar{\theta}\in\Theta θ , θ ˉ ∈ Θ , ∣ ℓ 12 , N ( θ ) − ℓ 12 , N ( θ ˉ ) ∣ ≤ L 12 ∥ θ − θ ˉ ∥ |\ell_{12,N}(\theta)-\ell_{12,N}(\bar{\theta})|\le L_{12}\|\theta-\bar{\theta}\| ∣ ℓ 12 , N ( θ ) − ℓ 12 , N ( θ ˉ ) ∣ ≤ L 12 ∥ θ − θ ˉ ∥ for some random variable L 12 L_{12} L 12 satisfying E [ L 12 ] < ∞ E[L_{12}]<\infty E [ L 12 ] < ∞ . Assume also that there exists δ > 0 \delta>0 δ > 0 such that E [ ∥ R 12 ∥ 4 + δ ] < ∞ E[\|R_{12}\|^{4+\delta}]<\infty E [ ∥ R 12 ∥ 4 + δ ] < ∞ .
Assumption 4.4 (Correct specification). Assume E [ D 12 ∣ X 1 , X 2 ] = e 12 ( θ 0 ) E[D_{12}\mid X_1,X_2]=e_{12}(\theta_0) E [ D 12 ∣ X 1 , X 2 ] = e 12 ( θ 0 ) a.s.
Proposition 4.5. Under Assumptions 4.1–4.4, θ ^ → p θ 0 \hat{\theta}\overset{p}{\to}\theta_0 θ ^ → p θ 0 .
Proof. Write
L N ( θ ) = ( N 2 ) − 1 ∑ 1 ≤ i < j ≤ N ℓ ˉ i j ( θ ) L_N(\theta)=\binom{N}{2}^{-1}\sum_{1\le i<j\le N}\bar{\ell}_{ij}(\theta) L N ( θ ) = ( 2 N ) − 1 1 ≤ i < j ≤ N ∑ ℓ ˉ ij ( θ )
with ℓ ˉ i j ( θ ) : = 1 2 { ℓ i j , N ( θ ) + ℓ j i , N ( θ ) } \bar{\ell}_{ij}(\theta):=\frac{1}{2}\{\ell_{ij,N}(\theta)+\ell_{ji,N}(\theta)\} ℓ ˉ ij ( θ ) := 2 1 { ℓ ij , N ( θ ) + ℓ j i , N ( θ )} . For each fixed θ ∈ Θ \theta\in\Theta θ ∈ Θ , the dyad summands are identically distributed, and those of vertex-disjoint dyads are independent, under the CID model with i.i.d. vertex variables. They are integrable because ∣ ℓ i j , N ( θ ) ∣ ≤ log 2 + ∥ R i j ∥ ∥ θ ∥ |\ell_{ij,N}(\theta)|\le\log2+\|R_{ij}\|\|\theta\| ∣ ℓ ij , N ( θ ) ∣ ≤ log 2 + ∥ R ij ∥∥ θ ∥ and E ∥ R 12 ∥ < ∞ E\|R_{12}\|<\infty E ∥ R 12 ∥ < ∞ . The law of large numbers for U-statistics (Hoeffding, [41 ] ) therefore gives L N ( θ ) → p E [ ℓ ˉ 12 ( θ ) ] L_N(\theta)\overset{p}{\to}E[\bar{\ell}_{12}(\theta)] L N ( θ ) → p E [ ℓ ˉ 12 ( θ )] . Because ( X 1 , X 2 ) (X_1,X_2) ( X 1 , X 2 ) and ( X 2 , X 1 ) (X_2,X_1) ( X 2 , X 1 ) have the same distribution, E [ ℓ ˉ 12 ( θ ) ] = ( E [ ℓ 12 , N ( θ ) ] + E [ ℓ 21 , N ( θ ) ] ) / 2 = L ( θ ) E[\bar{\ell}_{12}(\theta)] = \left(E[\ell_{12,N}(\theta)]+E[\ell_{21,N}(\theta)]\right)/2 = \mathcal{L}(\theta) E [ ℓ ˉ 12 ( θ )] = ( E [ ℓ 12 , N ( θ )] + E [ ℓ 21 , N ( θ )] ) /2 = L ( θ ) . Next, by Assumption 4.3 , ∣ ℓ ˉ 12 ( θ ) − ℓ ˉ 12 ( θ ˉ ) ∣ ≤ L ˉ 12 ∥ θ − θ ˉ ∥ |\bar{\ell}_{12}(\theta)-\bar{\ell}_{12}(\bar{\theta})| \le\bar{L}_{12}\|\theta-\bar{\theta}\| ∣ ℓ ˉ 12 ( θ ) − ℓ ˉ 12 ( θ ˉ ) ∣ ≤ L ˉ 12 ∥ θ − θ ˉ ∥ , where L ˉ 12 : = 1 2 ( L 12 + L 21 ) \bar{L}_{12} := \frac{1}{2}(L_{12}+L_{21}) L ˉ 12 := 2 1 ( L 12 + L 21 ) satisfies E [ L ˉ 12 ] < ∞ E[\bar{L}_{12}] < \infty E [ L ˉ 12 ] < ∞ . Since Θ \Theta Θ is compact, the usual finite-covering argument yields sup θ ∈ Θ ∣ L N ( θ ) − L ( θ ) ∣ → p 0 \sup_{\theta\in\Theta}|L_N(\theta)-\mathcal{L}(\theta)| \xrightarrow{p} 0 sup θ ∈ Θ ∣ L N ( θ ) − L ( θ ) ∣ p 0 . Since L \mathcal{L} L is uniquely maximized at θ 0 \theta_0 θ 0 by Assumption 4.2 , θ ^ → p θ 0 \hat{\theta} \xrightarrow{p} \theta_0 θ ^ p θ 0 by a standard argument (Newey and McFadden, [63 ] ).
Lemma 4.6 (Uniform Hessian convergence). Under Assumptions 4.1 –4.4 , sup θ ∈ Θ ∥ H N ( θ ) − E H 12 ( θ ) ∥ → p 0 \sup_{\theta\in\Theta}\|\mathcal{H}_N(\theta)-E\mathcal{H}_{12}(\theta)\| \xrightarrow{p} 0 sup θ ∈ Θ ∥ H N ( θ ) − E H 12 ( θ ) ∥ p 0 , where
H N ( θ ) : = ∂ θ θ ′ 2 L N ( θ ) = 1 N ( N − 1 ) ∑ i ≠ j H i j ( θ ) , \mathcal{H}_N(\theta):=\partial^2_{\theta\theta'}L_N(\theta)=\frac{1}{N(N-1)}\sum_{i\ne j}\mathcal{H}_{ij}(\theta), H N ( θ ) := ∂ θ θ ′ 2 L N ( θ ) = N ( N − 1 ) 1 i = j ∑ H ij ( θ ) , H i j ( θ ) : = ∂ θ θ ′ 2 ℓ i j , N ( θ ) = − e i j ( θ ) { 1 − e i j ( θ ) } R i j R i j ′ . \mathcal{H}_{ij}(\theta):=\partial^2_{\theta\theta'}\ell_{ij,N}(\theta)=-e_{ij}(\theta)\{1-e_{ij}(\theta)\}R_{ij}R_{ij}'. H ij ( θ ) := ∂ θ θ ′ 2 ℓ ij , N ( θ ) = − e ij ( θ ) { 1 − e ij ( θ )} R ij R ij ′ .
Proof. Direct calculation gives the score, the Hessian, and the derivative of the Hessian of the logit kernel ℓ i j , N ( θ ) \ell_{ij,N}(\theta) ℓ ij , N ( θ ) :
s i j ( θ ) = ∂ θ ℓ i j , N ( θ ) = R i j { D i j − e i j ( θ ) } , s_{ij}(\theta)=\partial_\theta\ell_{ij,N}(\theta)=R_{ij}\{D_{ij}-e_{ij}(\theta)\}, s ij ( θ ) = ∂ θ ℓ ij , N ( θ ) = R ij { D ij − e ij ( θ )} ,
H i j ( θ ) = ∂ θ θ ′ 2 ℓ i j , N ( θ ) = − e i j ( θ ) { 1 − e i j ( θ ) } R i j R i j ′ , \mathcal{H}_{ij}(\theta)=\partial^2_{\theta\theta'}\ell_{ij,N}(\theta)=-e_{ij}(\theta)\{1-e_{ij}(\theta)\}R_{ij}R_{ij}', H ij ( θ ) = ∂ θ θ ′ 2 ℓ ij , N ( θ ) = − e ij ( θ ) { 1 − e ij ( θ )} R ij R ij ′ ,
and, for each coordinate $r\in{1,\ldots,\dim(\theta)},
∂ H i j ( θ ) ∂ θ r = − e i j ( θ ) { 1 − e i j ( θ ) } { 1 − 2 e i j ( θ ) } R i j , r R i j R i j ′ . \frac{\partial\mathcal{H}_{ij}(\theta)}{\partial\theta_r}=-e_{ij}(\theta)\{1-e_{ij}(\theta)\}\{1-2e_{ij}(\theta)\}R_{ij,r}R_{ij}R_{ij}'. ∂ θ r ∂ H ij ( θ ) = − e ij ( θ ) { 1 − e ij ( θ )} { 1 − 2 e ij ( θ )} R ij , r R ij R ij ′ .
Since 0 ≤ e i j ( θ ) ≤ 1 0\le e_{ij}(\theta)\le1 0 ≤ e ij ( θ ) ≤ 1 , sup θ ∈ Θ ∥ H i j ( θ ) ∥ ≤ 1 4 ∥ R i j ∥ 2 \sup_{\theta\in\Theta}\|\mathcal{H}_{ij}(\theta)\|\le\frac{1}{4}\|R_{ij}\|^2 sup θ ∈ Θ ∥ H ij ( θ ) ∥ ≤ 4 1 ∥ R ij ∥ 2 , and
sup θ ∈ Θ ∥ ∂ H i j ( θ ) ∂ θ r ∥ ≤ C ∣ R i j , r ∣ ∥ R i j ∥ 2 ≤ C ∥ R i j ∥ 3 \sup_{\theta\in\Theta}\left\|\frac{\partial\mathcal{H}_{ij}(\theta)}{\partial\theta_r}\right\|\le C|R_{ij,r}|\|R_{ij}\|^2\le C\|R_{ij}\|^3 θ ∈ Θ sup ∂ θ r ∂ H ij ( θ ) ≤ C ∣ R ij , r ∣∥ R ij ∥ 2 ≤ C ∥ R ij ∥ 3
for a finite constant C C C . Therefore, by the mean-value theorem, ∥ H i j ( θ ) − H i j ( θ ˉ ) ∥ ≤ C ∥ R i j ∥ 3 ∥ θ − θ ˉ ∥ \|\mathcal{H}_{ij}(\theta)-\mathcal{H}_{ij}(\bar{\theta})\|\le C\|R_{ij}\|^3\|\theta-\bar{\theta}\| ∥ H ij ( θ ) − H ij ( θ ˉ ) ∥ ≤ C ∥ R ij ∥ 3 ∥ θ − θ ˉ ∥ . Averaging over ordered dyads gives
∥ H N ( θ ) − H N ( θ ˉ ) ∥ ≤ ( C N ( N − 1 ) ∑ i ≠ j ∥ R i j ∥ 3 ) ∥ θ − θ ˉ ∥ . \|\mathcal{H}_N(\theta)-\mathcal{H}_N(\bar{\theta})\|\le\left(\frac{C}{N(N-1)}\sum_{i\ne j}\|R_{ij}\|^3\right)\|\theta-\bar{\theta}\|. ∥ H N ( θ ) − H N ( θ ˉ ) ∥ ≤ N ( N − 1 ) C i = j ∑ ∥ R ij ∥ 3 ∥ θ − θ ˉ ∥.
Since E ∥ R 12 ∥ 3 < ∞ E\|R_{12}\|^3<\infty E ∥ R 12 ∥ 3 < ∞ , the law of large numbers for U-statistics (Hoeffding, [41 ] ) implies
1 N ( N − 1 ) ∑ i ≠ j ∥ R i j ∥ 3 = O p ( 1 ) , \frac{1}{N(N-1)}\sum_{i\ne j}\|R_{ij}\|^3=O_p(1), N ( N − 1 ) 1 i = j ∑ ∥ R ij ∥ 3 = O p ( 1 ) ,
so { H N ( θ ) : θ ∈ Θ } \{\mathcal{H}_N(\theta):\theta\in\Theta\} { H N ( θ ) : θ ∈ Θ } is stochastically equicontinuous. For each fixed θ ∈ Θ \theta\in\Theta θ ∈ Θ ,
H N ( θ ) = ( N 2 ) − 1 ∑ 1 ≤ i < j ≤ N H ˉ i j ( θ ) , \mathcal{H}_N(\theta)=\binom{N}{2}^{-1}\sum_{1\le i<j\le N}\bar{\mathcal{H}}_{ij}(\theta), H N ( θ ) = ( 2 N ) − 1 1 ≤ i < j ≤ N ∑ H ˉ ij ( θ ) ,
with H ‾ i j ( θ ) : = 1 2 { H i j ( θ ) + H j i ( θ ) } \overline{\mathcal{H}}_{ij}(\theta) := \frac{1}{2}\{\mathcal{H}_{ij}(\theta) + \mathcal{H}_{ji}(\theta)\} H ij ( θ ) := 2 1 { H ij ( θ ) + H j i ( θ )} , hence H N ( θ ) → p E [ H ‾ 12 ( θ ) ] = { E [ H 12 ( θ ) ] + E [ H 21 ( θ ) ] } / 2 = E [ H 12 ( θ ) ] \mathcal{H}_N(\theta) \xrightarrow{p} E[\overline{\mathcal{H}}_{12}(\theta)] = \{E[\mathcal{H}_{12}(\theta)] + E[\mathcal{H}_{21}(\theta)]\}/2 = E[\mathcal{H}_{12}(\theta)] H N ( θ ) p E [ H 12 ( θ )] = { E [ H 12 ( θ )] + E [ H 21 ( θ )]} /2 = E [ H 12 ( θ )] because ( X 1 , X 2 ) (X_1, X_2) ( X 1 , X 2 ) and ( X 2 , X 1 ) (X_2, X_1) ( X 2 , X 1 ) have the same distribution. Combining pointwise convergence with stochastic equicontinuity and compactness of Θ \Theta Θ , sup θ ∈ Θ ∥ H N ( θ ) − E H 12 ( θ ) ∥ → p 0 \sup_{\theta\in\Theta}\|\mathcal{H}_N(\theta)-E\mathcal{H}_{12}(\theta)\| \xrightarrow{p} 0 sup θ ∈ Θ ∥ H N ( θ ) − E H 12 ( θ ) ∥ p 0 .
Theorem 4.7. Under Assumptions 4.1–4.4,
N ( θ ^ − θ 0 ) → d N ( 0 , H 0 − 1 Ω 0 H 0 − 1 ) , \sqrt{N}(\hat{\theta}-\theta_0) \xrightarrow{d} N(0,\mathcal{H}_0^{-1}\Omega_0\mathcal{H}_0^{-1}), N ( θ ^ − θ 0 ) d N ( 0 , H 0 − 1 Ω 0 H 0 − 1 ) , where H 0 = − E [ e 12 ( θ 0 ) { 1 − e 12 ( θ 0 ) } R 12 R 12 ′ ] \mathcal{H}_0=-E[e_{12}(\theta_0)\{1-e_{12}(\theta_0)\}R_{12}R_{12}^{\prime}] H 0 = − E [ e 12 ( θ 0 ) { 1 − e 12 ( θ 0 )} R 12 R 12 ′ ] and Ω 0 : = 4 Var [ E [ s ˉ 12 ∣ Ξ 1 ] ] \Omega_0:=4\operatorname{Var}[E[\bar{s}_{12}\mid\Xi_1]] Ω 0 := 4 Var [ E [ s ˉ 12 ∣ Ξ 1 ]] , where Ξ i : = ( X i , U i ) \Xi_i:=(X_i,U_i) Ξ i := ( X i , U i ) , and s ˉ 12 : = ( s 12 ( θ 0 ) + s 21 ( θ 0 ) ) / 2 \bar{s}_{12}:=(s_{12}(\theta_0)+s_{21}(\theta_0))/2 s ˉ 12 := ( s 12 ( θ 0 ) + s 21 ( θ 0 )) /2 .
Proof. Define the score to be s N ( θ ) : = ∂ θ L N ( θ ) = 1 N ( N − 1 ) ∑ i ≠ j s i j ( θ ) s_N(\theta):=\partial_\theta L_N(\theta)=\frac{1}{N(N-1)}\sum_{i\ne j}s_{ij}(\theta) s N ( θ ) := ∂ θ L N ( θ ) = N ( N − 1 ) 1 ∑ i = j s ij ( θ ) . By Proposition 4.5 , θ ^ → p θ 0 \hat{\theta}\xrightarrow{p}\theta_0 θ ^ p θ 0 . Choose ε > 0 \varepsilon>0 ε > 0 so that the ball B ( θ 0 , ε ) B(\theta_0,\varepsilon) B ( θ 0 , ε ) lies in int ( Θ ) \operatorname{int}(\Theta) int ( Θ ) and in the differentiability neighborhood of Assumption 4.2. On the event ∥ θ ^ − θ 0 ∥ < ε \|\hat{\theta}-\theta_0\|<\varepsilon ∥ θ ^ − θ 0 ∥ < ε , whose probability tends to one, the first-order condition s N ( θ ^ ) = 0 s_N(\hat{\theta})=0 s N ( θ ^ ) = 0 holds and the entire parameter segment lies in this ball. The fundamental theorem of calculus, applied componentwise, gives
0 = s N ( θ 0 ) + H ‾ N ( θ ^ − θ 0 ) , H ‾ N : = ∫ 0 1 H N ( θ 0 + t ( θ ^ − θ 0 ) ) d t . 0=s_N(\theta_0)+\overline{\mathcal{H}}_N(\hat{\theta}-\theta_0),\qquad\overline{\mathcal{H}}_N:=\int_0^1\mathcal{H}_N(\theta_0+t(\hat{\theta}-\theta_0))\,dt. 0 = s N ( θ 0 ) + H N ( θ ^ − θ 0 ) , H N := ∫ 0 1 H N ( θ 0 + t ( θ ^ − θ 0 )) d t .
Let H ( θ ) : = E [ H 12 ( θ ) ] \mathcal{H}(\theta):=E[\mathcal{H}_{12}(\theta)] H ( θ ) := E [ H 12 ( θ )] . The logistic bound ∥ H 12 ( θ ) ∥ ≤ ∥ R 12 ∥ 2 / 4 \|\mathcal{H}_{12}(\theta)\|\leq\|R_{12}\|^2/4 ∥ H 12 ( θ ) ∥ ≤ ∥ R 12 ∥ 2 /4 and Assumption 4.3 justify differentiation under the expectation and continuity of H \mathcal{H} H near θ 0 \theta_0 θ 0 , with H ( θ 0 ) = H 0 \mathcal{H}(\theta_0)=\mathcal{H}_0 H ( θ 0 ) = H 0 . On the event above,
∥ H ‾ N − H 0 ∥ ≤ sup θ ∈ Θ ∥ H N ( θ ) − H ( θ ) ∥ + sup ∥ θ − θ 0 ∥ ≤ ∥ θ ^ − θ 0 ∥ ∥ H ( θ ) − H ( θ 0 ) ∥ . \|\overline{\mathcal{H}}_N-\mathcal{H}_0\|\leq\sup_{\theta\in\Theta}\|\mathcal{H}_N(\theta)-\mathcal{H}(\theta)\|+\sup_{\|\theta-\theta_0\|\leq\|\hat{\theta}-\theta_0\|}\|\mathcal{H}(\theta)-\mathcal{H}(\theta_0)\|. ∥ H N − H 0 ∥ ≤ θ ∈ Θ sup ∥ H N ( θ ) − H ( θ ) ∥ + ∥ θ − θ 0 ∥ ≤ ∥ θ ^ − θ 0 ∥ sup ∥ H ( θ ) − H ( θ 0 ) ∥.
Lemma 4.6, consistency, and continuity imply H ‾ N → p H 0 \overline{\mathcal{H}}_N\xrightarrow{p}\mathcal{H}_0 H N p H 0 . Since H 0 \mathcal{H}_0 H 0 is nonsingular, H ‾ N − 1 → p H 0 − 1 \overline{\mathcal{H}}_N^{-1}\xrightarrow{p}\mathcal{H}_0^{-1} H N − 1 p H 0 − 1 , and, with probability tending to one, N ( θ ^ − θ 0 ) = − H ‾ N − 1 N s N ( θ 0 ) \sqrt{N}(\hat{\theta}-\theta_0)=-\overline{\mathcal{H}}_N^{-1}\sqrt{N}s_N(\theta_0) N ( θ ^ − θ 0 ) = − H N − 1 N s N ( θ 0 ) . It remains to show that N s N ( θ 0 ) → d N ( 0 , Ω 0 ) \sqrt{N}s_N(\theta_0)\xrightarrow{d}N(0,\Omega_0) N s N ( θ 0 ) d N ( 0 , Ω 0 ) . To this end, rewrite the score as a sum over unordered dyads:
s N ( θ 0 ) = 2 N ( N − 1 ) ∑ 1 ≤ i < j ≤ N s ˉ i j , s_N(\theta_0)=\frac{2}{N(N-1)}\sum_{1\leq i<j\leq N}\bar{s}_{ij}, s N ( θ 0 ) = N ( N − 1 ) 2 1 ≤ i < j ≤ N ∑ s ˉ ij ,
with s ˉ i j : = 1 2 { s i j ( θ 0 ) + s j i ( θ 0 ) } \bar{s}_{ij}:=\frac{1}{2}\{s_{ij}(\theta_0)+s_{ji}(\theta_0)\} s ˉ ij := 2 1 { s ij ( θ 0 ) + s j i ( θ 0 )} . By Assumption 4.4, E [ s i j ( θ 0 ) ] = E [ R i j { D i j − E [ D i j ∣ X i , X j ] } ] = 0 E[s_{ij}(\theta_0)]=E[R_{ij}\{D_{ij}-E[D_{ij}\mid X_i,X_j]\}]=0 E [ s ij ( θ 0 )] = E [ R ij { D ij − E [ D ij ∣ X i , X j ]}] = 0 , and hence E [ s ˉ 12 ] = 0 E[\bar{s}_{12}]=0 E [ s ˉ 12 ] = 0 . Define m i j : = E [ s ˉ i j ∣ Ξ i , Ξ j ] m_{ij}:=E[\bar{s}_{ij}\mid\Xi_i,\Xi_j] m ij := E [ s ˉ ij ∣ Ξ i , Ξ j ] and ψ i : = E [ m i j ∣ Ξ i ] \psi_i:=E[m_{ij}\mid\Xi_i] ψ i := E [ m ij ∣ Ξ i ] . Then the score admits the decomposition
s N ( θ 0 ) = V 1 N + V 2 N + V 3 N , s_N(\theta_0)=V_{1N}+V_{2N}+V_{3N}, s N ( θ 0 ) = V 1 N + V 2 N + V 3 N ,
where
V 1 N : = 2 N ∑ i = 1 N ψ i V_{1N}:=\frac{2}{N}\sum_{i=1}^{N}\psi_i V 1 N := N 2 i = 1 ∑ N ψ i
V 2 N : = 2 N ( N − 1 ) ∑ 1 ≤ i < j ≤ N ( m i j − ψ i − ψ j ) V_{2N}:=\frac{2}{N(N-1)}\sum_{1\leq i<j\leq N}(m_{ij}-\psi_i-\psi_j) V 2 N := N ( N − 1 ) 2 1 ≤ i < j ≤ N ∑ ( m ij − ψ i − ψ j )
and
V 3 N : = 2 N ( N − 1 ) ∑ 1 ≤ i < j ≤ N ( s ˉ i j − m i j ) . V_{3N} := \frac{2}{N(N-1)} \sum_{1 \le i<j \le N} \left(\bar{s}_{ij}-m_{ij}\right). V 3 N := N ( N − 1 ) 2 1 ≤ i < j ≤ N ∑ ( s ˉ ij − m ij ) .
I first study V 1 N V_{1N} V 1 N . Since { Ξ i } i = 1 N \{\Xi_i\}_{i=1}^{N} { Ξ i } i = 1 N are i.i.d., { ψ i } i = 1 N \{\psi_i\}_{i=1}^{N} { ψ i } i = 1 N are i.i.d. as well. Moreover, since R 12 = R 21 R_{12}=R_{21} R 12 = R 21 , s ˉ 12 \bar{s}_{12} s ˉ 12 is bounded in norm by a constant times ∥ R 12 ∥ \lVert R_{12}\rVert ∥ R 12 ∥ , so Jensen’s inequality and Assumption 4.3 imply E ∥ ψ 1 ∥ 2 + δ < ∞ E\lVert\psi_1\rVert^{2+\delta}<\infty E ∥ ψ 1 ∥ 2 + δ < ∞ . Because E [ ψ 1 ] = E [ s ˉ 12 ] = 0 E[\psi_1]=E[\bar{s}_{12}]=0 E [ ψ 1 ] = E [ s ˉ 12 ] = 0 , the multivariate i.i.d. central limit theorem gives
N V 1 N = 2 ⋅ 1 N ∑ i = 1 N ψ i → d N ( 0 , Ω 0 ) , Ω 0 = 4 Var [ ψ 1 ] . \sqrt{N}V_{1N}=2\cdot\frac{1}{\sqrt{N}}\sum_{i=1}^{N}\psi_i \xrightarrow{d} N(0,\Omega_0), \qquad\Omega_0=4\operatorname{Var}[\psi_1]. N V 1 N = 2 ⋅ N 1 i = 1 ∑ N ψ i d N ( 0 , Ω 0 ) , Ω 0 = 4 Var [ ψ 1 ] .
Next consider V 2 N V_{2N} V 2 N . Let g i j : = m i j − ψ i − ψ j g_{ij}:=m_{ij}-\psi_i-\psi_j g ij := m ij − ψ i − ψ j . Then E [ g i j ∣ Ξ i ] = 0 E[g_{ij}\mid\Xi_i]=0 E [ g ij ∣ Ξ i ] = 0 , and E [ g i j ∣ Ξ j ] = 0 E[g_{ij}\mid\Xi_j]=0 E [ g ij ∣ Ξ j ] = 0 . Fix any vector c ∈ R dim ( θ ) c\in\mathbb{R}^{\dim(\theta)} c ∈ R d i m ( θ ) . Since c ′ V 2 N c'V_{2N} c ′ V 2 N is scalar,
E [ ( c ′ V 2 N ) 2 ] = 4 N 2 ( N − 1 ) 2 ∑ 1 ≤ i < j ≤ N ∑ 1 ≤ k < ℓ ≤ N E [ ( c ′ g i j ) ( c ′ g k ℓ ) ] . E[(c'V_{2N})^2]=\frac{4}{N^2(N-1)^2}\sum_{1\le i<j\le N}\sum_{1\le k<\ell\le N}E[(c'g_{ij})(c'g_{k\ell})]. E [( c ′ V 2 N ) 2 ] = N 2 ( N − 1 ) 2 4 1 ≤ i < j ≤ N ∑ 1 ≤ k < ℓ ≤ N ∑ E [( c ′ g ij ) ( c ′ g k ℓ )] .
If { i , j } ∩ { k , ℓ } = ∅ \{i,j\}\cap\{k,\ell\}=\varnothing { i , j } ∩ { k , ℓ } = ∅ , then independence of the underlying Ξ \Xi Ξ ’s and mean zero imply E [ ( c ′ g i j ) ( c ′ g k ℓ ) ] = 0 E[(c'g_{ij})(c'g_{k\ell})]=0 E [( c ′ g ij ) ( c ′ g k ℓ )] = 0 . If the dyads share exactly one index, say i = k i=k i = k and j ≠ ℓ j\ne\ell j = ℓ , then E [ ( c ′ g i j ) ( c ′ g i ℓ ) ] = E [ E [ c ′ g i j ∣ Ξ i ] E [ c ′ g i ℓ ∣ Ξ i ] ] = 0 E[(c'g_{ij})(c'g_{i\ell})]=E[E[c'g_{ij}\mid\Xi_i]E[c'g_{i\ell}\mid\Xi_i]]=0 E [( c ′ g ij ) ( c ′ g i ℓ )] = E [ E [ c ′ g ij ∣ Ξ i ] E [ c ′ g i ℓ ∣ Ξ i ]] = 0 . Therefore only the diagonal terms remain, so
E [ ( c ′ V 2 N ) 2 ] = 4 N 2 ( N − 1 ) 2 ∑ 1 ≤ i < j ≤ N E [ ( c ′ g i j ) 2 ] = O ( N − 2 ) . E[(c'V_{2N})^2]=\frac{4}{N^2(N-1)^2}\sum_{1\le i<j\le N}E[(c'g_{ij})^2]=O(N^{-2}). E [( c ′ V 2 N ) 2 ] = N 2 ( N − 1 ) 2 4 1 ≤ i < j ≤ N ∑ E [( c ′ g ij ) 2 ] = O ( N − 2 ) .
Hence N c ′ V 2 N → p 0 \sqrt{N}c'V_{2N}\xrightarrow{p}0 N c ′ V 2 N p 0 for every c c c , and therefore N V 2 N → p 0 \sqrt{N}V_{2N}\xrightarrow{p}0 N V 2 N p 0 .
Finally consider V 3 N V_{3N} V 3 N . Let v i j : = s ˉ i j − m i j v_{ij}:=\bar{s}_{ij}-m_{ij} v ij := s ˉ ij − m ij . Then E [ v i j ∣ Ξ i , Ξ j ] = 0 E[v_{ij}\mid\Xi_i,\Xi_j]=0 E [ v ij ∣ Ξ i , Ξ j ] = 0 . Conditional on { Ξ i } i = 1 N \{\Xi_i\}_{i=1}^{N} { Ξ i } i = 1 N , the unordered dyads { D i j : 1 ≤ i < j ≤ N } \{D_{ij}:1\le i<j\le N\} { D ij : 1 ≤ i < j ≤ N } are mutually independent under the CID model. Since v i j v_{ij} v ij depends on the sample only through ( D i j , Ξ i , Ξ j ) (D_{ij},\Xi_i,\Xi_j) ( D ij , Ξ i , Ξ j ) , it follows that conditional on { Ξ i } i = 1 N \{\Xi_i\}_{i=1}^{N} { Ξ i } i = 1 N , the collection { v i j : 1 ≤ i < j ≤ N } \{v_{ij}:1\le i<j\le N\} { v ij : 1 ≤ i < j ≤ N } is mutually independent with mean zero. Thus, for any fixed c c c ,
E [ ( c ′ V 3 N ) 2 ∣ Ξ 1 , … , Ξ N ] = 4 N 2 ( N − 1 ) 2 ∑ 1 ≤ i < j ≤ N E [ ( c ′ v i j ) 2 ∣ Ξ 1 , … , Ξ N ] . E[(c'V_{3N})^2\mid\Xi_1,\ldots,\Xi_N]=\frac{4}{N^2(N-1)^2}\sum_{1\le i<j\le N}E[(c'v_{ij})^2\mid\Xi_1,\ldots,\Xi_N]. E [( c ′ V 3 N ) 2 ∣ Ξ 1 , … , Ξ N ] = N 2 ( N − 1 ) 2 4 1 ≤ i < j ≤ N ∑ E [( c ′ v ij ) 2 ∣ Ξ 1 , … , Ξ N ] .
Taking expectations, E [ ( c ′ V 3 N ) 2 ] = O ( N − 2 ) E[(c'V_{3N})^2]=O(N^{-2}) E [( c ′ V 3 N ) 2 ] = O ( N − 2 ) . Hence N c ′ V 3 N → p 0 \sqrt{N}c'V_{3N}\xrightarrow{p}0 N c ′ V 3 N p 0 for every c c c , and therefore N V 3 N → p 0 \sqrt{N}V_{3N}\xrightarrow{p}0 N V 3 N p 0 .
Combining the three pieces, N s N ( θ 0 ) = N V 1 N + N V 2 N + N V 3 N → d N ( 0 , Ω 0 ) \sqrt{N}s_N(\theta_0)=\sqrt{N}V_{1N}+\sqrt{N}V_{2N}+\sqrt{N}V_{3N}\xrightarrow{d}N(0,\Omega_0) N s N ( θ 0 ) = N V 1 N + N V 2 N + N V 3 N d N ( 0 , Ω 0 ) . Substituting this limit into the integral-Hessian expansion and applying Slutsky’s theorem gives N ( θ ^ − θ 0 ) → d N ( 0 , H 0 − 1 Ω 0 H 0 − 1 ) \sqrt{N}(\hat{\theta}-\theta_0)\xrightarrow{d}N(0,H_0^{-1}\Omega_0H_0^{-1}) N ( θ ^ − θ 0 ) d N ( 0 , H 0 − 1 Ω 0 H 0 − 1 ) as desired.
Nonparametric estimation of the subgraph moments
With continuously distributed covariates, the empirical density P N ( S , x ) P_N(S,x) P N ( S , x ) in (1) is uninformative, because exact covariate matches occur with probability zero. I therefore smooth over nearby covariate values. Recall that d W = p d X d_W = p d_X d W = p d X is the dimension of the stacked covariate vector of a p p p -tuple, K : R d W → R K : \mathbb{R}^{d_W} \to\mathbb{R} K : R d W → R is a fixed multivariate kernel, and h N > 0 h_N > 0 h N > 0 is a bandwidth sequence with h N → 0 h_N \to0 h N → 0 . The kernel estimator is
P ^ N ( S , x ) = ∑ i p ∈ P p , N K ( x − X i p h N ) 1 { G N [ i p ] ≃ S } ∑ i p ∈ P p , N K ( x − X i p h N ) \widehat{P}_N(S,x) =
\frac{\sum_{i_p \in\mathcal{P}_{p,N}} K\left(\frac{x-X_{i_p}}{h_N}\right) 1\{G_N[i_p] \simeq S\}}
{\sum_{i_p \in\mathcal{P}_{p,N}} K\left(\frac{x-X_{i_p}}{h_N}\right)} P N ( S , x ) = ∑ i p ∈ P p , N K ( h N x − X i p ) ∑ i p ∈ P p , N K ( h N x − X i p ) 1 { G N [ i p ] ≃ S }
This estimator is the nonparametric fit of the network moment P ( S , x ) P(S,x) P ( S , x ) . It is the subgraph analogue of the Nadaraya–Watson estimator. Ding [22 ] establishes a variance bound, uniform convergence rates, and pointwise asymptotic normality for it; I use the uniform rates here.
Recall that Y i p , S = 1 { G N [ i p ] ≃ S } Y_{i_p,S} = 1\{G_N[i_p] \simeq S\} Y i p , S = 1 { G N [ i p ] ≃ S } is the response of the tuple i p i_p i p and X i p ∈ R d W X_{i_p} \in\mathbb{R}^{d_W} X i p ∈ R d W is its stacked covariate vector. The regression function of Y i p , S Y_{i_p,S} Y i p , S is m S ( w ) : = E [ Y i p , S ∣ X i p = w ] = P ( S , w ) m_S(w) := E[Y_{i_p,S} \mid X_{i_p} = w] = P(S,w) m S ( w ) := E [ Y i p , S ∣ X i p = w ] = P ( S , w ) .
Recall that P p , N \mathcal{P}_{p,N} P p , N contains the ordered tuples of distinct vertices, M N : = ∣ P p , N ∣ = N ! / ( N − p ) ! M_N := |\mathcal{P}_{p,N}| = N!/(N-p)! M N := ∣ P p , N ∣ = N ! / ( N − p )! , and C p , N \mathcal{C}_{p,N} C p , N contains their unordered vertex sets; for q p ∈ C p , N \mathbf{q}_p \in\mathcal{C}_{p,N} q p ∈ C p , N , S ( q p ) \mathcal{S}(\mathbf{q}_p) S ( q p ) denotes its p ! p! p ! permutations.
As in Theorem 4.7 , Ξ i : = ( X i , U i ) \Xi_i := (X_i,U_i) Ξ i := ( X i , U i ) denotes the i.i.d. vertex variables of the CID model; write Ξ q p : = ( Ξ q : q ∈ q p ) \Xi_{\mathbf{q}_p} := (\Xi_q : q \in\mathbf{q}_p) Ξ q p := ( Ξ q : q ∈ q p ) . Let f f f be the marginal density of X i X_i X i ; the stacked covariates then have density
f W ( w ) = ∏ a = 1 p f ( w a ) , w = ( w 1 ′ , … , w p ′ ) ′ , w a ∈ R d X . f_W(w) = \prod_{a=1}^{p} f(w_a), \qquad w = (w_1',\ldots,w_p')', \quad w_a \in\mathbb{R}^{d_X}. f W ( w ) = a = 1 ∏ p f ( w a ) , w = ( w 1 ′ , … , w p ′ ) ′ , w a ∈ R d X .
Recall that K h ( v ) = h N − d W K ( v / h N ) K_h(v) = h_N^{-d_W}K(v/h_N) K h ( v ) = h N − d W K ( v / h N ) , as defined after (2) , and set
K i p ( w ) : = K h ( w − X i p ) . K_{i_p}(w) := K_h(w-X_{i_p}). K i p ( w ) := K h ( w − X i p ) .
The normalized numerator and common density estimator are
Ψ ^ N ( w ) : = 1 M N ∑ i p ∈ P p , N K i p ( w ) Y i p , S , f ^ N ( w ) : = 1 M N ∑ i p ∈ P p , N K i p ( w ) . \widehat{\Psi}_N(w) := \frac{1}{M_N}\sum_{i_p \in\mathcal{P}_{p,N}} K_{i_p}(w)Y_{i_p,S}, \qquad\widehat{f}_N(w) := \frac{1}{M_N}\sum_{i_p \in\mathcal{P}_{p,N}} K_{i_p}(w). Ψ N ( w ) := M N 1 i p ∈ P p , N ∑ K i p ( w ) Y i p , S , f N ( w ) := M N 1 i p ∈ P p , N ∑ K i p ( w ) .
They estimate Ψ ( w ) : = P ( S , w ) f W ( w ) \Psi(w) := P(S,w)f_W(w) Ψ ( w ) := P ( S , w ) f W ( w ) and f W ( w ) f_W(w) f W ( w ) , respectively, and P ^ N ( S , w ) = Ψ ^ N ( w ) / f ^ N ( w ) \widehat{P}_N(S,w) = \widehat{\Psi}_N(w)/\widehat{f}_N(w) P N ( S , w ) = Ψ N ( w ) / f N ( w ) . In the regression notation of Ding [22 ] , this is m ^ N ( w ) \widehat{m}_N(w) m N ( w ) with response Y i p = Y i p , S Y_{i_p} = Y_{i_p,S} Y i p = Y i p , S and regression function m ( w ) = P ( S , w ) m(w) = P(S,w) m ( w ) = P ( S , w ) . That paper writes the kernel weight as h N − d W K ( ( X i p − w ) / h N ) h_N^{-d_W}K((X_{i_p}-w)/h_N) h N − d W K (( X i p − w ) / h N ) ; its results apply here with K ( − ⋅ ) K(-\cdot) K ( − ⋅ ) in place of K K K , which satisfies the same assumptions.
Assumption 4.8 (Covariate density). The covariates have a marginal density f f f satisfying sup x ∈ R d X f ( x ) ≤ B 1 < ∞ \sup_{x \in\mathbb{R}^{d_X}} f(x) \le B_1 < \infty sup x ∈ R d X f ( x ) ≤ B 1 < ∞ (without loss of generality, B 1 ≥ 1 B_1 \ge1 B 1 ≥ 1 ).
The CID model fits the framework of Ding [22 ] Section 2`): writing $D_{ij} = 1{V_{ij} \le h ( X i , X j , U i , U j ) } h(X_i, X_j, U_i, U_j)\} h ( X i , X j , U i , U j )} with i.i.d. uniform shocks V i j V_{ij} V ij independent of the vertex variables, Y i p , S Y_{i_p,S} Y i p , S is a fixed function of Ξ i p \Xi_{i_p} Ξ i p and the shocks of the dyads in i p i_p i p , as in its Example 2.1. In particular, subgraphs with disjoint vertex sets are independent. Since 0 ≤ Y i p , S ≤ 1 0 \le Y_{i_p,S} \le1 0 ≤ Y i p , S ≤ 1 , the response has bounded conditional moments of every order, so the moment conditions in its Assumptions 3.1(ii) and 3.3(i) hold automatically.2
Assumption 4.9 (Kernel). The kernel K : R d W → R K:\mathbb{R}^{d_W}\to\mathbb{R} K : R d W → R satisfies:
(i) ∫ R d W ∣ K ( u ) ∣ d u ≤ C 1 < ∞ \int_{\mathbb{R}^{d_W}} \lvert K(u)\rvert\,du\le C_1<\infty ∫ R d W ∣ K ( u )∣ d u ≤ C 1 < ∞ .
(ii) For every A ⊆ { 1 , … , p } A\subseteq\{1,\ldots,p\} A ⊆ { 1 , … , p } , sup u A ∫ ∣ K ( u 1 , … , u p ) ∣ d u A c ≤ C 2 < ∞ \sup_{u_A}\int\lvert K(u_1,\ldots,u_p)\rvert\,du_{A^c}\le C_2<\infty sup u A ∫ ∣ K ( u 1 , … , u p )∣ d u A c ≤ C 2 < ∞ , where u a ∈ R d X u_a\in\mathbb{R}^{d_X} u a ∈ R d X , u A = ( u a : a ∈ A ) u_A=(u_a:a\in A) u A = ( u a : a ∈ A ) , and d u A c du_{A^c} d u A c integrates the remaining blocks.
(iii) sup u ∈ R d W ∣ K ( u ) ∣ ≤ K max < ∞ \sup_{u\in\mathbb{R}^{d_W}}\lvert K(u)\rvert\le K_{\max}<\infty sup u ∈ R d W ∣ K ( u )∣ ≤ K m a x < ∞ .
Assumption 4.10 (Kernel regularity). For finite constants L K , Λ 1 L_K,\Lambda_1 L K , Λ 1 , the kernel satisfies K ( u ) = 0 K(u)=0 K ( u ) = 0 when ∥ u ∥ ∞ > L K \lVert u\rVert_\infty>L_K ∥ u ∥ ∞ > L K and ∣ K ( u ) − K ( v ) ∣ ≤ Λ 1 ∥ u − v ∥ ∞ \lvert K(u)-K(v)\rvert\le\Lambda_1\lVert u-v\rVert_\infty ∣ K ( u ) − K ( v )∣ ≤ Λ 1 ∥ u − v ∥ ∞ for u , v ∈ R d W u,v\in\mathbb{R}^{d_W} u , v ∈ R d W .
Definition 4.11 (Hölder smoothness). Let β , L > 0 \beta,L>0 β , L > 0 , let ⌊ β ⌋ : = max { r ∈ Z ≥ 0 : r < β } \lfloor\beta\rfloor:=\max\{r\in\mathbb{Z}_{\ge0}:r<\beta\} ⌊ β ⌋ := max { r ∈ Z ≥ 0 : r < β } be the greatest integer strictly less than β \beta β , and let α : = β − ⌊ β ⌋ ∈ ( 0 , 1 ] \alpha:=\beta-\lfloor\beta\rfloor\in(0,1] α := β − ⌊ β ⌋ ∈ ( 0 , 1 ] . For a multi-index s = ( s 1 , … , s d W ) s=(s_1,\ldots,s_{d_W}) s = ( s 1 , … , s d W ) , let ∣ s ∣ : = ∑ j s j \lvert s\rvert:=\sum_j s_j ∣ s ∣ := ∑ j s j and D s : = ∂ ∣ s ∣ / ( ∂ w 1 s 1 ⋯ ∂ w d W s d W ) D^s:=\partial^{\lvert s\rvert}/(\partial w_1^{s_1}\cdots\partial w_{d_W}^{s_{d_W}}) D s := ∂ ∣ s ∣ / ( ∂ w 1 s 1 ⋯ ∂ w d W s d W ) . A function a : R d W → R a:\mathbb{R}^{d_W}\to\mathbb{R} a : R d W → R belongs to Σ ( β , L ) \Sigma(\beta,L) Σ ( β , L ) if it is ⌊ β ⌋ \lfloor\beta\rfloor ⌊ β ⌋ times continuously differentiable and
∣ D s a ( w ) − D s a ( w ′ ) ∣ ≤ L ∥ w − w ′ ∥ ∞ α \lvert D^s a(w)-D^s a(w')\rvert\le L\lVert w-w'\rVert_\infty^\alpha ∣ D s a ( w ) − D s a ( w ′ )∣ ≤ L ∥ w − w ′ ∥ ∞ α for every ∣ s ∣ = ⌊ β ⌋ \lvert s\rvert=\lfloor\beta\rfloor ∣ s ∣ = ⌊ β ⌋ and every w , w ′ ∈ R d W w,w'\in\mathbb{R}^{d_W} w , w ′ ∈ R d W .
Assumption 4.12 (Higher-order kernel). The kernel satisfies ∫ R d W K ( u ) d u = 1 \int_{\mathbb{R}^{d_W}}K(u)\,du=1 ∫ R d W K ( u ) d u = 1 and
∫ R d W u r K ( u ) d u = 0 for every multi-index r ∈ Z + d W with 1 ≤ ∣ r ∣ < β , \int_{\mathbb{R}^{d_W}}u^rK(u)\,du=0\qquad\text{for every multi-index }r\in\mathbb{Z}_+^{d_W}\text{ with }1\le\lvert r\rvert<\beta, ∫ R d W u r K ( u ) d u = 0 for every multi-index r ∈ Z + d W with 1 ≤ ∣ r ∣ < β , where u r : = u 1 r 1 ⋯ u d W r d W u^r:=u_1^{r_1}\cdots u_{d_W}^{r_{d_W}} u r := u 1 r 1 ⋯ u d W r d W . Also, ∫ R d W ∥ u ∥ ∞ β ∣ K ( u ) ∣ d u < ∞ \int_{\mathbb{R}^{d_W}}\lVert u\rVert_\infty^\beta\lvert K(u)\rvert\,du<\infty ∫ R d W ∥ u ∥ ∞ β ∣ K ( u )∣ d u < ∞ .
Assumption 4.10 is Ding [22 ] (Assumption 3.3(ii)). Assumptions 4.8, 4.9, and 4.12 are its Assumptions 3.1(i), 3.2, and 3.4, and Definition 4.11 is its Definition 2.1.
For later use, define the stochastic and bias-adjusted rates
a N : = ( ln N N h N d X ) 1 / 2 , a N ∗ : = a N + h N β . (3) a_N:=\left(\frac{\ln N}{N h_N^{d_X}}\right)^{1/2},\qquad a_N^*:=a_N+h_N^\beta. \tag*{(3)} a N := ( N h N d X ln N ) 1/2 , a N ∗ := a N + h N β . (3)
Under the regularity and bandwidth conditions maintained in Assumption 5.1 , the uniform convergence results in Ding [22 ] give
sup w ∈ W π ∣ f ^ N ( w ) − f W ( w ) ∣ = O p ( a N ∗ ) , sup w ∈ W π ∣ P ^ N ( S , w ) − P ( S , w ) ∣ = O p ( a N ∗ ) (4) \sup_{w\in\mathcal{W}_{\pi}}\left|\widehat{f}_{N}(w)-f_{W}(w)\right|=O_{p}(a_{N}^{*}),\qquad\sup_{w\in\mathcal{W}_{\pi}}\left|\widehat{P}_{N}(S,w)-P(S,w)\right|=O_{p}(a_{N}^{*}) \tag*{(4)} w ∈ W π sup f N ( w ) − f W ( w ) = O p ( a N ∗ ) , w ∈ W π sup P N ( S , w ) − P ( S , w ) = O p ( a N ∗ ) (4)
where W π \mathcal{W}_{\pi} W π is the compact support of the test weight and f W f_{W} f W is bounded away from zero there. The density rate is Theorem 3.2 applied with Y i p ≡ 1 Y_{ip}\equiv1 Y i p ≡ 1 , combined with the smoothing-bias bound in its Lemma E.1, as in the proof of its Theorem 3.3. The rate for P ^ N \widehat{P}_{N} P N is its Theorem 3.3 with a fixed ball containing W π \mathcal{W}_{\pi} W π ; the proof of that theorem uses the lower bound on f W f_{W} f W only over the set on which the supremum is taken, so the bound on W π \mathcal{W}_{\pi} W π suffices.
Asymptotics of the test statistic
In this section I write w ∈ R d W w\in\mathbb{R}^{d_{W}} w ∈ R d W for the stacked covariate argument, written x x x in § 4 . Recall that the test statistic is
T N , S : = N h N d X / 2 ∫ { P ^ N ( S , w ) − K h , N P ( S , ⋅ ; e θ ^ ) ( w ) } 2 π ( w ) d w . T_{N,S}:=Nh_{N}^{d_{X}/2}\int\left\{\widehat{P}_{N}(S,w)-K_{h,N}P(S,\cdot;e_{\widehat{\theta}})(w)\right\}^{2}\pi(w)\,dw. T N , S := N h N d X /2 ∫ { P N ( S , w ) − K h , N P ( S , ⋅ ; e θ ) ( w ) } 2 π ( w ) d w .
Define the subgraph probability conditional on the vertex variables, which averages only over the dyad shocks V i j V_{ij} V ij :
μ S ( ξ 1 , … , ξ p ) : = E [ Y ( 1 , … , p ) , S ∣ Ξ 1 = ξ 1 , … , Ξ p = ξ p ] . \mu_{S}(\xi_{1},\ldots,\xi_{p}):=E\left[Y_{(1,\ldots,p),S}\mid\Xi_{1}=\xi_{1},\ldots,\Xi_{p}=\xi_{p}\right]. μ S ( ξ 1 , … , ξ p ) := E [ Y ( 1 , … , p ) , S ∣ Ξ 1 = ξ 1 , … , Ξ p = ξ p ] .
Thus P ( S , x 1 , … , x p ) = E [ μ S ( Ξ 1 , … , Ξ p ) ∣ X 1 = x 1 , … , X p = x p ] P(S,x_{1},\ldots,x_{p})=E\left[\mu_{S}(\Xi_{1},\ldots,\Xi_{p})\mid X_{1}=x_{1},\ldots,X_{p}=x_{p}\right] P ( S , x 1 , … , x p ) = E [ μ S ( Ξ 1 , … , Ξ p ) ∣ X 1 = x 1 , … , X p = x p ] .
To define the first-order Hoeffding projection, fix a value ξ = ( x , u ) \xi=(x,u) ξ = ( x , u ) of a vertex variable. For a = 1 , … , p a=1,\ldots,p a = 1 , … , p , let Ξ ( a ) ( ξ ) : = ( Ξ 1 , … , Ξ a − 1 , ξ , Ξ a + 1 , … , Ξ p ) \Xi^{(a)}(\xi):=(\Xi_{1},\ldots,\Xi_{a-1},\xi,\Xi_{a+1},\ldots,\Xi_{p}) Ξ ( a ) ( ξ ) := ( Ξ 1 , … , Ξ a − 1 , ξ , Ξ a + 1 , … , Ξ p ) and X ( a ) ( x ) : = ( X 1 , … , X a − 1 , x , X a + 1 , … , X p ) X^{(a)}(x):=(X_{1},\ldots,X_{a-1},x,X_{a+1},\ldots,X_{p}) X ( a ) ( x ) := ( X 1 , … , X a − 1 , x , X a + 1 , … , X p ) . Define
ℓ N , S ( ξ ; w ) : = 1 p f W ( w ) ∑ a = 1 p E [ K h { w − X ( a ) ( x ) } × { μ S ( Ξ ( a ) ( ξ ) ) − P ( S , X ( a ) ( x ) ) } ] (5) \ell_{N,S}(\xi;w):=\frac{1}{pf_{W}(w)}\sum_{a=1}^{p}E\left[K_{h}\{w-X^{(a)}(x)\}\times\left\{\mu_{S}\left(\Xi^{(a)}(\xi)\right)-P\left(S,X^{(a)}(x)\right)\right\}\right] \tag*{(5)} ℓ N , S ( ξ ; w ) := p f W ( w ) 1 a = 1 ∑ p E [ K h { w − X ( a ) ( x )} × { μ S ( Ξ ( a ) ( ξ ) ) − P ( S , X ( a ) ( x ) ) } ] (5)
where the expectation integrates over { Ξ b : b ≠ a } \{\Xi_{b}:b\ne a\} { Ξ b : b = a } . By construction, E [ ℓ N , S ( Ξ 1 ; w ) ] = 0 E[\ell_{N,S}(\Xi_{1};w)]=0 E [ ℓ N , S ( Ξ 1 ; w )] = 0 .
Define
Γ N , S ( w , z ) : = p 2 h N d X E [ ℓ N , S ( Ξ 1 ; w ) ℓ N , S ( Ξ 1 ; z ) ] , \Gamma_{N,S}(w,z):=p^{2}h_{N}^{d_{X}}E[\ell_{N,S}(\Xi_{1};w)\ell_{N,S}(\Xi_{1};z)], Γ N , S ( w , z ) := p 2 h N d X E [ ℓ N , S ( Ξ 1 ; w ) ℓ N , S ( Ξ 1 ; z )] ,
b N , S : = h N − d X / 2 ∫ Γ N , S ( w , w ) π ( w ) d w , b_{N,S}:=h_{N}^{-d_{X}/2}\int\Gamma_{N,S}(w,w)\pi(w)\,dw, b N , S := h N − d X /2 ∫ Γ N , S ( w , w ) π ( w ) d w ,
and
V N , S : = 2 h N − d X ∬ Γ N , S ( w , z ) 2 π ( w ) π ( z ) d w d z . \mathcal{V}_{N,S}:=2h_{N}^{-d_{X}}\iint\Gamma_{N,S}(w,z)^{2}\pi(w)\pi(z)\,dw\,dz. V N , S := 2 h N − d X ∬ Γ N , S ( w , z ) 2 π ( w ) π ( z ) d w d z .
Assumption 5.1 (Regularity conditions). The CID model and Assumptions 4.1 –4.4 , 4.8 , 4.9 , 4.10 , and 4.12 hold. In addition, the following conditions hold.
(i) The weight π \pi π is nonnegative and bounded with compact support W π \mathcal{W}_{\pi} W π , and there is a constant f ‾ > 0 \underline{f} > 0 f > 0 such that inf w ∈ W π f W ( w ) ≥ f ‾ \inf_{w \in\mathcal{W}_{\pi}} f_W(w) \ge\underline{f} inf w ∈ W π f W ( w ) ≥ f .
(ii) The functions f W f_W f W , P ( S , ⋅ ) P(S,\cdot) P ( S , ⋅ ) , and the first two derivatives of P ( S , ⋅ ; e θ ) P(S,\cdot;\boldsymbol{e}_{\theta}) P ( S , ⋅ ; e θ ) with respect to θ \theta θ belong to Σ ( β , L ) \Sigma(\beta,L) Σ ( β , L ) , uniformly for θ \theta θ in a neighborhood of θ 0 \theta_0 θ 0 .
(iii) The bandwidth satisfies h N → 0 h_N \to0 h N → 0 and a N ∗ h N − d X / 2 → 0 a_N^{*}h_N^{-d_X/2} \to0 a N ∗ h N − d X /2 → 0 . Equivalently, N h N 2 d X / ln N → ∞ Nh_N^{2d_X}/\ln N \to\infty N h N 2 d X / ln N → ∞ and β > d X / 2 \beta> d_X/2 β > d X /2 .
(iv) The variance constant V N , S \mathcal{V}_{N,S} V N , S converges to a limit V S ∈ ( 0 , ∞ ) \mathcal{V}_S \in(0,\infty) V S ∈ ( 0 , ∞ ) .
Theorem 5.2 (Asymptotic null distribution). Suppose Assumption 5.1 holds. Then, under H 0 H_0 H 0 ,
T N , S − b N , S → d N ( 0 , V S ) . T_{N,S}-b_{N,S}\to_d N(0,\mathcal{V}_S). T N , S − b N , S → d N ( 0 , V S ) . Proof. Throughout the proof, C C C denotes a finite constant that does not depend on N N N and may change from line to line. Under H 0 H_0 H 0 , set η i p : = Y i p , S − P ( S , X i p ) \eta_{ip}:=Y_{ip,S}-P(S,X_{ip}) η i p := Y i p , S − P ( S , X i p ) , so that
Y i p , S − P ( S , X i p ; e ^ θ ^ ) = η i p − { P ( S , X i p ; e ^ θ ^ ) − P ( S , X i p ; e θ 0 ) } . Y_{ip,S}-P(S,X_{ip};\widehat{\boldsymbol{e}}_{\widehat{\theta}})=\eta_{ip}-\left\{P(S,X_{ip};\widehat{\boldsymbol{e}}_{\widehat{\theta}})-P(S,X_{ip};\boldsymbol{e}_{\theta_0})\right\}. Y i p , S − P ( S , X i p ; e θ ) = η i p − { P ( S , X i p ; e θ ) − P ( S , X i p ; e θ 0 ) } . Define
U N 1 ( w ) : = 1 M N f W ( w ) ∑ i p ∈ P p , N K h ( w − X i p ) η i p , U_{N1}(w):=\frac{1}{M_Nf_W(w)}\sum_{ip\in\mathcal{P}_{p,N}}K_h(w-X_{ip})\eta_{ip}, U N 1 ( w ) := M N f W ( w ) 1 i p ∈ P p , N ∑ K h ( w − X i p ) η i p , U N 2 ( w ) : = − 1 M N f W ( w ) ∑ i p ∈ P p , N K h ( w − X i p ) { P ( S , X i p ; e ^ θ ^ ) − P ( S , X i p ; e θ 0 ) } . U_{N2}(w):=-\frac{1}{M_Nf_W(w)}\sum_{ip\in\mathcal{P}_{p,N}}K_h(w-X_{ip})\left\{P(S,X_{ip};\widehat{\boldsymbol{e}}_{\widehat{\theta}})-P(S,X_{ip};\boldsymbol{e}_{\theta_0})\right\}. U N 2 ( w ) := − M N f W ( w ) 1 i p ∈ P p , N ∑ K h ( w − X i p ) { P ( S , X i p ; e θ ) − P ( S , X i p ; e θ 0 ) } . Since the nonparametric estimator and the smoothed parametric estimator have the same denominator,
P ^ N ( S , w ) − K h , N P ( S , ⋅ ; e ^ θ ^ ) ( w ) = f W ( w ) f ^ N ( w ) { U N 1 ( w ) + U N 2 ( w ) } . (6) \widehat{P}_N(S,w)-K_{h,N}P(S,\cdot;\widehat{\boldsymbol{e}}_{\widehat{\theta}})(w)=\frac{f_W(w)}{\widehat{f}_N(w)}\{U_{N1}(w)+U_{N2}(w)\}. \tag*{(6)} P N ( S , w ) − K h , N P ( S , ⋅ ; e θ ) ( w ) = f N ( w ) f W ( w ) { U N 1 ( w ) + U N 2 ( w )} . (6) Step 1: Replacement of the random denominator.
Let δ N : = ∥ f ^ N − f W ∥ ∞ , W π \delta_N:=\|\widehat{f}_N-f_W\|_{\infty,\mathcal{W}_{\pi}} δ N := ∥ f N − f W ∥ ∞ , W π . By (4) , δ N = O p ( a N ∗ ) \delta_N=O_p(a_N^{*}) δ N = O p ( a N ∗ ) and δ N h N − d X / 2 = o p ( 1 ) \delta_Nh_N^{-d_X/2}=o_p(1) δ N h N − d X /2 = o p ( 1 ) . Since f W ≥ f ‾ f_W\ge\underline{f} f W ≥ f on W π \mathcal{W}_{\pi} W π ,
Pr ( inf w ∈ W π f ^ N ( w ) ≥ f ‾ 2 ) → 1. \Pr\left(\inf_{w\in\mathcal{W}_{\pi}}\widehat{f}_N(w)\ge\frac{\underline{f}}{2}\right)\to1. Pr ( w ∈ W π inf f N ( w ) ≥ 2 f ) → 1. On this event, equation (6) becomes
P ^ N ( S , w ) − K h , N P ( S , ⋅ ; e ^ θ ^ ) ( w ) = U N 1 ( w ) + U N 2 ( w ) + R f , N ( w ) , \widehat{P}_N(S,w)-K_{h,N}P(S,\cdot;\widehat{\boldsymbol{e}}_{\widehat{\theta}})(w)=U_{N1}(w)+U_{N2}(w)+R_{f,N}(w), P N ( S , w ) − K h , N P ( S , ⋅ ; e θ ) ( w ) = U N 1 ( w ) + U N 2 ( w ) + R f , N ( w ) ,
where
R f , N ( w ) = { U N 1 ( w ) + U N 2 ( w ) } { f W ( w ) f ^ N ( w ) − 1 } . R_{f,N}(w)=\{U_{N1}(w)+U_{N2}(w)\}\left\{\frac{f_W(w)}{\widehat{f}_N(w)}-1\right\}. R f , N ( w ) = { U N 1 ( w ) + U N 2 ( w )} { f N ( w ) f W ( w ) − 1 } .
Moreover,
∣ f W ( w ) f ^ N ( w ) − 1 ∣ = ∣ f W ( w ) − f ^ N ( w ) ∣ f ^ N ( w ) ≤ 2 f δ N . \left|\frac{f_W(w)}{\widehat{f}_N(w)}-1\right|=\frac{|f_W(w)-\widehat{f}_N(w)|}{\widehat{f}_N(w)}\leq\frac{2}{f}\delta_N. f N ( w ) f W ( w ) − 1 = f N ( w ) ∣ f W ( w ) − f N ( w ) ∣ ≤ f 2 δ N .
Thus,
∣ R f , N ( w ) ∣ ≤ 2 f δ N { ∣ U N 1 ( w ) ∣ + ∣ U N 2 ( w ) ∣ } . (7) |R_{f,N}(w)|\leq\frac{2}{f}\delta_N\{|U_{N1}(w)|+|U_{N2}(w)|\}. \tag*{(7)} ∣ R f , N ( w ) ∣ ≤ f 2 δ N { ∣ U N 1 ( w ) ∣ + ∣ U N 2 ( w ) ∣ } . (7)
I first bound the integrated square of U N 1 U_{N1} U N 1 . Because E [ η i p ∣ X i p ] = 0 E[\eta_{ip}\mid X_{ip}]=0 E [ η i p ∣ X i p ] = 0 , disjoint tuples have zero covariance. There are O ( N 2 p − t ) O(N^{2p-t}) O ( N 2 p − t ) ordered pairs of tuples sharing exactly t t t vertices. For such a pair, integration over the 2 p − t 2p-t 2 p − t distinct vertex covariates, as in the proof of Ding [22 ] , Theorem 3.1, yields
E ∫ ∣ K h ( w − X i p ) K h ( w − X k p ) η i p η k p ∣ π ( w ) d w ≤ C h N − t d X . E\int|K_h(w-X_{ip})K_h(w-X_{kp})\eta_{ip}\eta_{kp}|\pi(w)\,dw\leq Ch_N^{-td_X}. E ∫ ∣ K h ( w − X i p ) K h ( w − X k p ) η i p η k p ∣ π ( w ) d w ≤ C h N − t d X .
Since M N 2 ≍ N 2 p M_N^2\asymp N^{2p} M N 2 ≍ N 2 p ,
E ∫ U N 1 ( w ) 2 π ( w ) d w ≤ C ∑ t = 1 p N − t h N − t d X = O ( 1 N h N d X ) , E\int U_{N1}(w)^2\pi(w)\,dw\leq C\sum_{t=1}^{p}N^{-t}h_N^{-td_X}=O\left(\frac{1}{Nh_N^{d_X}}\right), E ∫ U N 1 ( w ) 2 π ( w ) d w ≤ C t = 1 ∑ p N − t h N − t d X = O ( N h N d X 1 ) ,
where the t = 1 t=1 t = 1 term dominates because N h N d X → ∞ Nh_N^{d_X}\to\infty N h N d X → ∞ . Consequently,
N h N d X / 2 ∫ U N 1 ( w ) 2 π ( w ) d w = O p ( h N − d X / 2 ) . (8) Nh_N^{d_X/2}\int U_{N1}(w)^2\pi(w)\,dw=O_p(h_N^{-d_X/2}). \tag*{(8)} N h N d X /2 ∫ U N 1 ( w ) 2 π ( w ) d w = O p ( h N − d X /2 ) . (8)
Step 2 below shows that
N h N d X / 2 ∫ U N 2 ( w ) 2 π ( w ) d w = o p ( 1 ) . (9) Nh_N^{d_X/2}\int U_{N2}(w)^2\pi(w)\,dw=o_p(1). \tag*{(9)} N h N d X /2 ∫ U N 2 ( w ) 2 π ( w ) d w = o p ( 1 ) . (9)
Using (7) and ( a + b ) 2 ≤ 2 a 2 + 2 b 2 (a+b)^2\leq2a^2+2b^2 ( a + b ) 2 ≤ 2 a 2 + 2 b 2 ,
N h N d X / 2 ∫ R f , N ( w ) 2 π ( w ) d w ≤ C δ N 2 N h N d X / 2 ∫ { U N 1 ( w ) 2 + U N 2 ( w ) 2 } π ( w ) d w = O p ( δ N 2 h N − d X / 2 ) = o p ( 1 ) . \begin{aligned}
Nh_N^{d_X/2}\int R_{f,N}(w)^2\pi(w)\,dw\leq C\delta_N^2Nh_N^{d_X/2}\int\{U_{N1}(w)^2+U_{N2}(w)^2\}\pi(w)\,dw \\
&=O_p(\delta_N^2h_N^{-d_X/2})=o_p(1).
\end{aligned} N h N d X /2 ∫ R f , N ( w ) 2 π ( w ) d w ≤ C δ N 2 N h N d X /2 ∫ { U N 1 ( w ) 2 + U N 2 ( w ) 2 } π ( w ) d w = O p ( δ N 2 h N − d X /2 ) = o p ( 1 ) .
The last conclusion follows because δ N h N − d X / 2 = o p ( 1 ) \delta_Nh_N^{-d_X/2}=o_p(1) δ N h N − d X /2 = o p ( 1 ) and δ N = o p ( 1 ) \delta_N=o_p(1) δ N = o p ( 1 ) .
Similarly, (7) , (8) , and (9) imply
N h N d X / 2 ∫ ∣ R f , N ( w ) ∣ { ∣ U N 1 ( w ) ∣ + ∣ U N 2 ( w ) ∣ } π ( w ) d w ≤ C δ N N h N d X / 2 ∫ { ∣ U N 1 ( w ) ∣ + ∣ U N 2 ( w ) ∣ } 2 π ( w ) d w ≤ 2 C δ N N h N d X / 2 ∫ { U N 1 ( w ) 2 + U N 2 ( w ) 2 } π ( w ) d w = O p ( δ N h N − d X / 2 ) = o p ( 1 ) . N h_N^{d_X/2}\int\left|R_{f,N}(w)\right|\left\{\left|U_{N1}(w)\right|+\left|U_{N2}(w)\right|\right\}\pi(w)\,dw
\le C\delta_N N h_N^{d_X/2}\int\left\{\left|U_{N1}(w)\right|+\left|U_{N2}(w)\right|\right\}^{2}\pi(w)\,dw
\le2C\delta_N N h_N^{d_X/2}\int\left\{U_{N1}(w)^2+U_{N2}(w)^2\right\}\pi(w)\,dw
=O_p\left(\delta_N h_N^{-d_X/2}\right)=o_p(1). N h N d X /2 ∫ ∣ R f , N ( w ) ∣ { ∣ U N 1 ( w ) ∣ + ∣ U N 2 ( w ) ∣ } π ( w ) d w ≤ C δ N N h N d X /2 ∫ { ∣ U N 1 ( w ) ∣ + ∣ U N 2 ( w ) ∣ } 2 π ( w ) d w ≤ 2 C δ N N h N d X /2 ∫ { U N 1 ( w ) 2 + U N 2 ( w ) 2 } π ( w ) d w = O p ( δ N h N − d X /2 ) = o p ( 1 ) .
Therefore,
T N , S = N h N d X / 2 ∫ { U N 1 ( w ) + U N 2 ( w ) } 2 π ( w ) d w + o p ( 1 ) . (10) T_{N,S}=N h_N^{d_X/2}\int\left\{U_{N1}(w)+U_{N2}(w)\right\}^{2}\pi(w)\,dw+o_p(1). \tag*{(10)} T N , S = N h N d X /2 ∫ { U N 1 ( w ) + U N 2 ( w ) } 2 π ( w ) d w + o p ( 1 ) . (10)
Step 2: The squared parametric estimation deviation.
Only tuples with X i p \mathbf{X}_{i_p} X i p within L K h N L_Kh_N L K h N of W π \mathcal{W}_\pi W π contribute to U N 2 ( w ) U_{N2}(w) U N 2 ( w ) for w ∈ W π w\in\mathcal{W}_\pi w ∈ W π , because K K K vanishes outside { ∥ u ∥ ∞ ≤ L K } \{\lVert u\rVert_\infty\le L_K\} {∥ u ∥ ∞ ≤ L K } by Assumption 4.10 and π \pi π is supported on the compact set W π \mathcal{W}_\pi W π . On this compact neighborhood of W π \mathcal{W}_\pi W π , the first two θ \theta θ -derivatives of P ( S , ⋅ ; e θ ) P(S,\cdot;e_\theta) P ( S , ⋅ ; e θ ) , which are continuous by Assumption 5.1 (ii), are bounded uniformly for θ \theta θ in a neighborhood of θ 0 \theta_0 θ 0 . A second-order Taylor expansion in θ \theta θ gives
P ( S , X i p ; e θ ^ ) − P ( S , X i p ; e θ 0 ) = ∂ θ P ( S , X i p ; e θ 0 ) ′ ( θ ^ − θ 0 ) + R i p θ , P(S,X_{i_p};e_{\widehat{\theta}})-P(S,X_{i_p};e_{\theta_0})
=\partial_\theta P(S,X_{i_p};e_{\theta_0})'(\widehat{\theta}-\theta_0)+R_{i_p}^{\theta}, P ( S , X i p ; e θ ) − P ( S , X i p ; e θ 0 ) = ∂ θ P ( S , X i p ; e θ 0 ) ′ ( θ − θ 0 ) + R i p θ ,
where, uniformly over i p i_p i p ,
∣ R i p θ ∣ ≤ C ∥ θ ^ − θ 0 ∥ 2 = O p ( N − 1 ) . \left|R_{i_p}^{\theta}\right|\le C\lVert\widehat{\theta}-\theta_0\rVert^2=O_p(N^{-1}). R i p θ ≤ C ∥ θ − θ 0 ∥ 2 = O p ( N − 1 ) .
Let
A N ( w ) : = 1 M N ∑ i p ∈ P p , N K h N ( w − X i p ) ∂ θ P ( S , X i p ; e θ 0 ) f W ( w ) A_N(w):=\frac{1}{M_N}\sum_{i_p\in\mathcal{P}_{p,N}}\frac{K_{h_N}(w-X_{i_p})\partial_\theta P(S,X_{i_p};e_{\theta_0})}{f_W(w)} A N ( w ) := M N 1 i p ∈ P p , N ∑ f W ( w ) K h N ( w − X i p ) ∂ θ P ( S , X i p ; e θ 0 )
and define R N θ ( w ) R_N^\theta(w) R N θ ( w ) analogously with R i p θ R_{i_p}^\theta R i p θ in place of ∂ θ P ( S , X i p ; e θ 0 ) \partial_\theta P(S,X_{i_p};e_{\theta_0}) ∂ θ P ( S , X i p ; e θ 0 ) . Then
U N 2 ( w ) = − A N ( w ) ′ ( θ ^ − θ 0 ) − R N θ ( w ) . (11) U_{N2}(w)=-A_N(w)'(\widehat{\theta}-\theta_0)-R_N^\theta(w). \tag*{(11)} U N 2 ( w ) = − A N ( w ) ′ ( θ − θ 0 ) − R N θ ( w ) . (11)
Write A N ( w ) = A ‾ N ( w ) + A ~ N ( w ) A_N(w)=\overline{A}_N(w)+\widetilde{A}_N(w) A N ( w ) = A N ( w ) + A N ( w ) with A ‾ N ( w ) : = E [ A N ( w ) ] \overline{A}_N(w):=E[A_N(w)] A N ( w ) := E [ A N ( w )] . Boundedness of ∂ θ P ( S , ⋅ ; e θ 0 ) \partial_\theta P(S,\cdot;e_{\theta_0}) ∂ θ P ( S , ⋅ ; e θ 0 ) , f W − 1 f_W^{-1} f W − 1 , and ∫ ∣ K h N ( w − z ) ∣ d z \int\left|K_{h_N}(w-z)\right|\,dz ∫ ∣ K h N ( w − z ) ∣ d z implies sup w ∈ W π ∣ A ‾ N ( w ) ∣ ≤ C \sup_{w\in\mathcal{W}_\pi}\left|\overline{A}_N(w)\right|\le C sup w ∈ W π A N ( w ) ≤ C . I next calculate the integrated variance of A ~ N \widetilde{A}_N A N . Expanding the square gives a double sum over ordered tuples. Terms indexed by disjoint tuples have zero covariance because they depend on disjoint sets of i.i.d. vertex covariates. If two tuples share t ≥ 1 t\ge1 t ≥ 1 vertices, partial kernel integration gives an integrated covariance of order at most h N − t d X h_N^{-td_X} h N − t d X . Hence
E ∥ A ~ N ∥ L 2 ( π ) 2 ≤ C M N 2 ∑ t = 1 p N 2 p − t h N − t d X ≤ C ∑ t = 1 p N − t h N − t d X = O ( 1 N h N d X ) . E\left\|\widetilde{A}_N\right\|_{L^2(\pi)}^2
\le\frac{C}{M_N^2}\sum_{t=1}^{p}N^{2p-t}h_N^{-td_X}
\le C\sum_{t=1}^{p}N^{-t}h_N^{-td_X}
=O\left(\frac{1}{N h_N^{d_X}}\right). E A N L 2 ( π ) 2 ≤ M N 2 C t = 1 ∑ p N 2 p − t h N − t d X ≤ C t = 1 ∑ p N − t h N − t d X = O ( N h N d X 1 ) .
The final equality follows because the t = 1 t=1 t = 1 term dominates. Thus ∥ A ~ N ∥ L 2 ( π ) = O p ( ( N h N d X ) − 1 / 2 ) = o p ( 1 ) \|\widetilde{A}_N\|_{L^2(\pi)}=O_p\left((Nh_N^{d_X})^{-1/2}\right)=o_p(1) ∥ A N ∥ L 2 ( π ) = O p ( ( N h N d X ) − 1/2 ) = o p ( 1 ) . Since π \pi π is bounded with compact support, ∥ A ‾ N ∥ L 2 ( π ) ≤ C \|\overline{A}_N\|_{L^2(\pi)}\le C ∥ A N ∥ L 2 ( π ) ≤ C , and hence ∥ A N ∥ L 2 ( π ) = O p ( 1 ) \|A_N\|_{L^2(\pi)}=O_p(1) ∥ A N ∥ L 2 ( π ) = O p ( 1 ) .
Also, the average absolute kernel weight M N − 1 ∑ i p ∣ K h ( w − X i p ) ∣ M_N^{-1}\sum_{i_p}|K_h(w-X_{i_p})| M N − 1 ∑ i p ∣ K h ( w − X i p ) ∣ is O p ( 1 ) O_p(1) O p ( 1 ) uniformly over W π \mathcal{W}_\pi W π by (4) applied with ∣ K ∣ |K| ∣ K ∣ in place of K K K , which together with sup i p ∣ R N θ ∣ = O p ( N − 1 ) \sup_{i_p}|R_{N}^{\theta}|=O_p(N^{-1}) sup i p ∣ R N θ ∣ = O p ( N − 1 ) implies sup w ∈ W π ∣ R N θ ( w ) ∣ = O p ( N − 1 ) \sup_{w\in\mathcal{W}_\pi}|R_N^\theta(w)|=O_p(N^{-1}) sup w ∈ W π ∣ R N θ ( w ) ∣ = O p ( N − 1 ) . Using (11) , ( a + b ) 2 ≤ 2 a 2 + 2 b 2 (a+b)^2\le2a^2+2b^2 ( a + b ) 2 ≤ 2 a 2 + 2 b 2 , and ∥ θ ^ − θ 0 ∥ = O p ( N − 1 / 2 ) \|\widehat{\theta}-\theta_0\|=O_p(N^{-1/2}) ∥ θ − θ 0 ∥ = O p ( N − 1/2 ) from Theorem 4.7 ,
∫ U N 2 ( w ) 2 π ( w ) d w ≤ 2 ∥ θ ^ − θ 0 ∥ 2 ∥ A N ∥ L 2 ( π ) 2 + 2 ∥ R N θ ∥ L 2 ( π ) 2 = O p ( N − 1 ) , \int U_{N2}(w)^2\pi(w)\,dw\le2\|\widehat{\theta}-\theta_0\|^2\|A_N\|_{L^2(\pi)}^2+2\|R_N^\theta\|_{L^2(\pi)}^2=O_p(N^{-1}), ∫ U N 2 ( w ) 2 π ( w ) d w ≤ 2∥ θ − θ 0 ∥ 2 ∥ A N ∥ L 2 ( π ) 2 + 2∥ R N θ ∥ L 2 ( π ) 2 = O p ( N − 1 ) ,
and therefore, establishing (9) , N h N d X / 2 ∫ U N 2 ( w ) 2 π ( w ) d w = O p ( h N d X / 2 ) = o p ( 1 ) N h_N^{d_X/2}\int U_{N2}(w)^2\pi(w)\,dw=O_p(h_N^{d_X/2})=o_p(1) N h N d X /2 ∫ U N 2 ( w ) 2 π ( w ) d w = O p ( h N d X /2 ) = o p ( 1 ) .
Step 3: The nonparametric–parametric cross term.
By (11) , the nonparametric–parametric cross term is
∫ U N 1 ( w ) U N 2 ( w ) π ( w ) d w = − ( θ ^ − θ 0 ) ′ ∫ U N 1 ( w ) A N ( w ) π ( w ) d w − ∫ U N 1 ( w ) R N θ ( w ) π ( w ) d w . \int U_{N1}(w)U_{N2}(w)\pi(w)\,dw=-(\widehat{\theta}-\theta_0)^\prime\int U_{N1}(w)A_N(w)\pi(w)\,dw-\int U_{N1}(w)R_N^\theta(w)\pi(w)\,dw. ∫ U N 1 ( w ) U N 2 ( w ) π ( w ) d w = − ( θ − θ 0 ) ′ ∫ U N 1 ( w ) A N ( w ) π ( w ) d w − ∫ U N 1 ( w ) R N θ ( w ) π ( w ) d w .
I bound the first term again using the decomposition A N = A ‾ N + A ~ N A_N=\overline{A}_N+\widetilde{A}_N A N = A N + A N . Fix a coordinate m ∈ { 1 , … , dim ( θ ) } m\in\{1,\ldots,\dim(\theta)\} m ∈ { 1 , … , dim ( θ )} . Substituting the definition of U N 1 U_{N1} U N 1 ,
∫ U N 1 ( w ) A ‾ N , m ( w ) π ( w ) d w = 1 M N ∑ i p ∈ P p , N η i p ∫ K h N ( w − X i p ) A ‾ N , m ( w ) π ( w ) f W ( w ) d w . \int U_{N1}(w)\overline{A}_{N,m}(w)\pi(w)\,dw=\frac{1}{M_N}\sum_{i_p\in\mathcal{P}_{p,N}}\eta_{i_p}\int\frac{K_{h_N}(w-X_{i_p})\overline{A}_{N,m}(w)\pi(w)}{f_W(w)}\,dw. ∫ U N 1 ( w ) A N , m ( w ) π ( w ) d w = M N 1 i p ∈ P p , N ∑ η i p ∫ f W ( w ) K h N ( w − X i p ) A N , m ( w ) π ( w ) d w .
The inner integral is uniformly bounded in X i p X_{i_p} X i p , because A ‾ N \overline{A}_N A N , π \pi π , and f W − 1 f_W^{-1} f W − 1 are bounded and the L 1 L^1 L 1 -norm of K h N K_{h_N} K h N is constant in h N h_N h N . Each summand is centered since E [ η i p ∣ X i p ] = 0 E[\eta_{i_p}\mid X_{i_p}]=0 E [ η i p ∣ X i p ] = 0 . Pairs of disjoint tuples have zero covariance. For each t ≥ 1 t\ge1 t ≥ 1 , there are O ( N 2 p − t ) O(N^{2p-t}) O ( N 2 p − t ) ordered pairs sharing exactly t t t vertices, and the covariance of the corresponding bounded summands is bounded by a constant. Therefore, using M N 2 ≍ N 2 p M_N^2\asymp N^{2p} M N 2 ≍ N 2 p ,
Var [ ∫ U N 1 ( w ) A ‾ N , m ( w ) π ( w ) d w ] ≤ C M N 2 ∑ t = 1 p N 2 p − t ≤ C N − 2 p ∑ t = 1 p N 2 p − t = C ∑ t = 1 p N − t = O ( N − 1 ) . \operatorname{Var}\left[\int U_{N1}(w)\overline{A}_{N,m}(w)\pi(w)\,dw\right]\le\frac{C}{M_N^2}\sum_{t=1}^{p}N^{2p-t}\le CN^{-2p}\sum_{t=1}^{p}N^{2p-t}=C\sum_{t=1}^{p}N^{-t}=O(N^{-1}). Var [ ∫ U N 1 ( w ) A N , m ( w ) π ( w ) d w ] ≤ M N 2 C t = 1 ∑ p N 2 p − t ≤ C N − 2 p t = 1 ∑ p N 2 p − t = C t = 1 ∑ p N − t = O ( N − 1 ) .
Since this holds for every coordinate m m m ,
∫ U N 1 ( w ) A ‾ N ( w ) π ( w ) d w = O p ( N − 1 / 2 ) . \int U_{N1}(w)\overline{A}_N(w)\pi(w)\,dw=O_p(N^{-1/2}). ∫ U N 1 ( w ) A N ( w ) π ( w ) d w = O p ( N − 1/2 ) .
For the random part, Cauchy–Schwarz gives
∣ ∫ U N 1 ( w ) A ~ N ( w ) π ( w ) d w ∣ ≤ ∥ U N 1 ∥ L 2 ( π ) ∥ A ~ N ∥ L 2 ( π ) = O p ( ( N h N d X ) − 1 / 2 ) O p ( ( N h N d X ) − 1 / 2 ) = O p ( 1 N h N d X ) . \left|\int U_{N1}(w)\widetilde{A}_{N}(w)\pi(w)\,dw\right| \le\|U_{N1}\|_{L^2(\pi)}\|\widetilde{A}_{N}\|_{L^2(\pi)}
= O_p\left((Nh_N^{d_X})^{-1/2}\right)O_p\left((Nh_N^{d_X})^{-1/2}\right)=O_p\left(\frac{1}{Nh_N^{d_X}}\right). ∫ U N 1 ( w ) A N ( w ) π ( w ) d w ≤ ∥ U N 1 ∥ L 2 ( π ) ∥ A N ∥ L 2 ( π ) = O p ( ( N h N d X ) − 1/2 ) O p ( ( N h N d X ) − 1/2 ) = O p ( N h N d X 1 ) .
The two norm orders follow respectively from the integrated second-moment bound for U N 1 U_{N1} U N 1 in Step 1 and the integrated variance calculation for A ^ N \widehat{A}_N A N . Hence
∫ U N 1 ( w ) A N ( w ) π ( w ) d w = O p ( N − 1 / 2 ) + O p ( 1 N h N d X ) . (12) \int U_{N1}(w)A_N(w)\pi(w)\,dw=O_p(N^{-1/2})+O_p\left(\frac{1}{Nh_N^{d_X}}\right). \tag*{(12)} ∫ U N 1 ( w ) A N ( w ) π ( w ) d w = O p ( N − 1/2 ) + O p ( N h N d X 1 ) . (12)
Combining (11) and (12) , and using ∥ θ ^ − θ 0 ∥ = O p ( N − 1 / 2 ) \|\widehat{\theta}-\theta_0\|=O_p(N^{-1/2}) ∥ θ − θ 0 ∥ = O p ( N − 1/2 ) ,
N h N d X / 2 ∣ ( θ ^ − θ 0 ) ′ ∫ U N 1 ( w ) A N ( w ) π ( w ) d w ∣ ≤ N h N d X / 2 O p ( N − 1 / 2 ) { O p ( N − 1 / 2 ) + O p ( 1 N h N d X ) } = O p ( h N d X / 2 ) + O p ( 1 N h N d X ) = o p ( 1 ) . \begin{aligned}
Nh_N^{d_X/2}\left|(\widehat{\theta}-\theta_0)^\prime\int U_{N1}(w)A_N(w)\pi(w)\,dw\right| \\
&\le Nh_N^{d_X/2}O_p(N^{-1/2})\left\{O_p(N^{-1/2})+O_p\left(\frac{1}{Nh_N^{d_X}}\right)\right\} \\
&=O_p(h_N^{d_X/2})+O_p\left(\frac{1}{\sqrt{Nh_N^{d_X}}}\right)=o_p(1).
\end{aligned} N h N d X /2 ( θ − θ 0 ) ′ ∫ U N 1 ( w ) A N ( w ) π ( w ) d w ≤ N h N d X /2 O p ( N − 1/2 ) { O p ( N − 1/2 ) + O p ( N h N d X 1 ) } = O p ( h N d X /2 ) + O p N h N d X 1 = o p ( 1 ) .
For the Taylor remainder,
N h N d X / 2 ∣ ∫ U N 1 ( w ) R N θ ( w ) π ( w ) d w ∣ ≤ N h N d X / 2 ∥ U N 1 ∥ L 2 ( π ) ∥ R N θ ∥ L 2 ( π ) = N h N d X / 2 O p ( ( N h N d X ) − 1 / 2 ) O p ( N − 1 ) = O p ( N − 1 / 2 ) = o p ( 1 ) . \begin{aligned}
Nh_N^{d_X/2}\left|\int U_{N1}(w)R_N^\theta(w)\pi(w)\,dw\right|\le Nh_N^{d_X/2}\|U_{N1}\|_{L^2(\pi)}\|R_N^\theta\|_{L^2(\pi)} \\
&=Nh_N^{d_X/2}O_p\left((Nh_N^{d_X})^{-1/2}\right)O_p(N^{-1}) \\
&=O_p(N^{-1/2})=o_p(1).
\end{aligned} N h N d X /2 ∫ U N 1 ( w ) R N θ ( w ) π ( w ) d w ≤ N h N d X /2 ∥ U N 1 ∥ L 2 ( π ) ∥ R N θ ∥ L 2 ( π ) = N h N d X /2 O p ( ( N h N d X ) − 1/2 ) O p ( N − 1 ) = O p ( N − 1/2 ) = o p ( 1 ) .
Consequently, N h N d X / 2 ∫ U N 1 ( w ) U N 2 ( w ) π ( w ) d w = o p ( 1 ) Nh_N^{d_X/2}\int U_{N1}(w)U_{N2}(w)\pi(w)\,dw=o_p(1) N h N d X /2 ∫ U N 1 ( w ) U N 2 ( w ) π ( w ) d w = o p ( 1 ) . Together with (10) and (9) , this gives
T N , S = N h N d X / 2 ∫ U N 1 ( w ) 2 π ( w ) d w + o p ( 1 ) . (13) T_{N,S}=Nh_N^{d_X/2}\int U_{N1}(w)^2\pi(w)\,dw+o_p(1). \tag*{(13)} T N , S = N h N d X /2 ∫ U N 1 ( w ) 2 π ( w ) d w + o p ( 1 ) . (13)
Step 4: Reduction of U N 1 U_{N1} U N 1 to its Hájek projection.
For q p ∈ C p , N \mathbf{q}_p\in C_{p,N} q p ∈ C p , N , define the symmetrized residualized kernel
γ N , q p 0 ( w ) : = 1 p ! ∑ σ ∈ S ( q p ) K h ( w − X σ ( q p ) ) f W ( w ) × { Y σ ( q p ) , S − P ( S , X σ ( q p ) ) } . \gamma^0_{N,\mathbf{q}_p}(w):=\frac{1}{p!}\sum_{\sigma\in\mathfrak{S}_{(\mathbf{q}_p)}}\frac{K_h(w-X_{\sigma(\mathbf{q}_p)})}{f_W(w)}\times\left\{Y_{\sigma(\mathbf{q}_p),S}-P(S,X_{\sigma(\mathbf{q}_p)})\right\}. γ N , q p 0 ( w ) := p ! 1 σ ∈ S ( q p ) ∑ f W ( w ) K h ( w − X σ ( q p ) ) × { Y σ ( q p ) , S − P ( S , X σ ( q p ) ) } .
Then U N 1 ( w ) = ( N p ) − 1 ∑ q p ∈ C p , N γ N , q p 0 ( w ) U_{N1}(w)=\binom{N}{p}^{-1}\sum_{\mathbf{q}_p\in C_{p,N}}\gamma^0_{N,\mathbf{q}_p}(w) U N 1 ( w ) = ( p N ) − 1 ∑ q p ∈ C p , N γ N , q p 0 ( w ) . To average out the dyad shocks, set Γ ‾ N , q p 0 ( w ) : = \overline{\Gamma}^0_{N,\mathbf{q}_p}(w):= Γ N , q p 0 ( w ) := E [ Υ ‾ N , q p 0 ( w ) ∣ Ξ q p ] E[\overline{\Upsilon}^{0}_{N,q_p}(w)\mid\Xi_{q_p}] E [ Υ N , q p 0 ( w ) ∣ Ξ q p ] . By the definition of μ S \mu_S μ S ,
Υ ‾ N , q p 0 ( w ) = 1 p ! ∑ σ ∈ S ( q p ) K h ( w − X σ ( q p ) ) f W ( w ) × { μ S ( Ξ ‾ σ ( q p ) ) − P ( S , X σ ( q p ) ) } . \overline{\Upsilon}^{0}_{N,q_p}(w)=\frac{1}{p!}\sum_{\sigma\in\mathfrak{S}(q_p)}\frac{K_h(w-X_{\sigma(q_p)})}{f_W(w)}\times\left\{\mu_S(\overline{\Xi}_{\sigma(q_p)})-P(S,X_{\sigma(q_p)})\right\}. Υ N , q p 0 ( w ) = p ! 1 σ ∈ S ( q p ) ∑ f W ( w ) K h ( w − X σ ( q p ) ) × { μ S ( Ξ σ ( q p ) ) − P ( S , X σ ( q p ) ) } .
Moreover, E [ Υ ‾ N , q p 0 ( w ) ] = 0 E[\overline{\Upsilon}^{0}_{N,q_p}(w)]=0 E [ Υ N , q p 0 ( w )] = 0 . Write
U N 1 ( w ) = ( N p ) − 1 ∑ q p ∈ C p , N Υ ‾ N , q p 0 ( w ) + V N , S e ( w ) , U_{N1}(w)=\binom{N}{p}^{-1}\sum_{q_p\in\mathcal{C}_{p,N}}\overline{\Upsilon}^{0}_{N,q_p}(w)+V^{e}_{N,S}(w), U N 1 ( w ) = ( p N ) − 1 q p ∈ C p , N ∑ Υ N , q p 0 ( w ) + V N , S e ( w ) ,
where V N , S e V^{e}_{N,S} V N , S e , the edge-shock remainder, is the average of Υ N , q p 0 − Υ ‾ N , q p 0 \Upsilon^{0}_{N,q_p}-\overline{\Upsilon}^{0}_{N,q_p} Υ N , q p 0 − Υ N , q p 0 , so that E [ V N , S e ( w ) ∣ Ξ 1 , … , Ξ N ] = 0 E[V^{e}_{N,S}(w)\mid\Xi_1,\ldots,\Xi_N]=0 E [ V N , S e ( w ) ∣ Ξ 1 , … , Ξ N ] = 0 . The first term is an order-p p p U-statistic in the i.i.d. vertex variables, and the Hoeffding decomposition of this U-statistic is ∑ k = 1 p U N , S , k ( w ) \sum_{k=1}^{p}U_{N,S,k}(w) ∑ k = 1 p U N , S , k ( w ) with
U N , S , k ( w ) : = ( p k ) ( N k ) ∑ q k ∈ C k , N ϕ N , S , k 0 ( Ξ q k ; w ) . U_{N,S,k}(w):=\frac{\binom{p}{k}}{\binom{N}{k}}\sum_{q_k\in\mathcal{C}_{k,N}}\phi^{0}_{N,S,k}(\Xi_{q_k};w). U N , S , k ( w ) := ( k N ) ( k p ) q k ∈ C k , N ∑ ϕ N , S , k 0 ( Ξ q k ; w ) .
Here ϕ N , S , k 0 \phi^{0}_{N,S,k} ϕ N , S , k 0 is the canonical kernel of order k k k , which I now define. For A ⊆ { 1 , … , p } A\subseteq\{1,\ldots,p\} A ⊆ { 1 , … , p } and ξ A : = ( ξ a ) a ∈ A \xi_A:=(\xi_a)_{a\in A} ξ A := ( ξ a ) a ∈ A , let
ψ N , S , A ( ξ A ; w ) : = E [ Υ ‾ N , { 1 , … , p } 0 ( w ) ∣ Ξ a = ξ a , a ∈ A ] \psi_{N,S,A}(\xi_A;w):=E\left[\overline{\Upsilon}^{0}_{N,\{1,\ldots,p\}}(w)\mid\Xi_a=\xi_a,\ a\in A\right] ψ N , S , A ( ξ A ; w ) := E [ Υ N , { 1 , … , p } 0 ( w ) ∣ Ξ a = ξ a , a ∈ A ]
be the projection of the kernel onto the vertices in A A A . Since Υ ‾ N , { 1 , … , p } 0 ( w ) \overline{\Upsilon}^{0}_{N,\{1,\ldots,p\}}(w) Υ N , { 1 , … , p } 0 ( w ) is a symmetric function of Ξ 1 , … , Ξ p \Xi_1,\ldots,\Xi_p Ξ 1 , … , Ξ p , ψ N , S , A \psi_{N,S,A} ψ N , S , A depends on A A A only through ∣ A ∣ |A| ∣ A ∣ , and ψ N , S , ∅ ( w ) = E [ Υ ‾ N , { 1 , … , p } 0 ( w ) ] = 0 \psi_{N,S,\varnothing}(w)=E[\overline{\Upsilon}^{0}_{N,\{1,\ldots,p\}}(w)]=0 ψ N , S , ∅ ( w ) = E [ Υ N , { 1 , … , p } 0 ( w )] = 0 . The canonical kernels are obtained by Möbius inversion of the ψ N , S , A \psi_{N,S,A} ψ N , S , A over subsets,
ϕ N , S , k 0 ( ξ 1 , … , ξ k ; w ) : = ∑ A ⊆ { 1 , … , k } ( − 1 ) k − ∣ A ∣ ψ N , S , A ( ξ A ; w ) , k = 1 , … , p . (14) \phi^{0}_{N,S,k}(\xi_1,\ldots,\xi_k;w):=\sum_{A\subseteq\{1,\ldots,k\}}(-1)^{k-|A|}\psi_{N,S,A}(\xi_A;w),\qquad k=1,\ldots,p. \tag*{(14)} ϕ N , S , k 0 ( ξ 1 , … , ξ k ; w ) := A ⊆ { 1 , … , k } ∑ ( − 1 ) k − ∣ A ∣ ψ N , S , A ( ξ A ; w ) , k = 1 , … , p . (14)
They are canonical in the sense that ϕ N , S , k 0 \phi^{0}_{N,S,k} ϕ N , S , k 0 has conditional mean zero given any k − 1 k-1 k − 1 of its arguments, as I verify below for k ≥ 2 k\ge2 k ≥ 2 .
For k = 1 k=1 k = 1 , (14) gives ϕ N , S , 1 0 ( ξ ; w ) = ψ N , S , { 1 } ( ξ ; w ) = E [ Υ ‾ N , { 1 , … , p } 0 ( w ) ∣ Ξ 1 = ξ ] \phi^{0}_{N,S,1}(\xi;w)=\psi_{N,S,\{1\}}(\xi;w)=E[\overline{\Upsilon}^{0}_{N,\{1,\ldots,p\}}(w)\mid\Xi_1=\xi] ϕ N , S , 1 0 ( ξ ; w ) = ψ N , S , { 1 } ( ξ ; w ) = E [ Υ N , { 1 , … , p } 0 ( w ) ∣ Ξ 1 = ξ ] . In the symmetrization over σ ∈ S ( { 1 , … , p } ) \sigma\in\mathfrak{S}(\{1,\ldots,p\}) σ ∈ S ({ 1 , … , p }) , vertex 1 occupies coordinate block a a a of the kernel argument in exactly ( p − 1 ) ! (p-1)! ( p − 1 )! of the p ! p! p ! permutations, for each a = 1 , … , p a=1,\ldots,p a = 1 , … , p . Conditioning on Ξ 1 = ξ = ( x , u ) \Xi_1=\xi=(x,u) Ξ 1 = ξ = ( x , u ) and integrating out the other p − 1 p-1 p − 1 vertices therefore gives
ϕ N , S , 1 0 ( ξ ; w ) = 1 p f W ( w ) ∑ a = 1 p E [ K h { w − X ( a ) ( x ) } { μ S ( Ξ ( a ) ( ξ ) ) − P ( S , X ( a ) ( x ) ) } ] = ℓ N , S ( ξ ; w ) , \phi^{0}_{N,S,1}(\xi;w)=\frac{1}{pf_W(w)}\sum_{a=1}^{p}E\left[K_h\{w-X^{(a)}(x)\}\left\{\mu_S(\Xi^{(a)}(\xi))-P(S,X^{(a)}(x))\right\}\right]=\ell_{N,S}(\xi;w), ϕ N , S , 1 0 ( ξ ; w ) = p f W ( w ) 1 a = 1 ∑ p E [ K h { w − X ( a ) ( x )} { μ S ( Ξ ( a ) ( ξ )) − P ( S , X ( a ) ( x )) } ] = ℓ N , S ( ξ ; w ) ,
the first-order projection defined in (5) . Hence the Hájek projection of the U-statistic is
U N , S , 1 ( w ) = p N ∑ i = 1 N ℓ N , S ( Ξ i ; w ) . (15) U_{N,S,1}(w)=\frac{p}{N}\sum_{i=1}^{N}\ell_{N,S}(\Xi_i;w). \tag*{(15)} U N , S , 1 ( w ) = N p i = 1 ∑ N ℓ N , S ( Ξ i ; w ) . (15)
For k ≥ 2 k \ge2 k ≥ 2 , the Möbius representation makes every ϕ N , S , k 0 \phi^{0}_{N,S,k} ϕ N , S , k 0 degenerate. To see this, condition on Ξ 1 , … , Ξ k − 1 \Xi_1,\ldots,\Xi_{k-1} Ξ 1 , … , Ξ k − 1 and pair every A ⊆ { 1 , … , k − 1 } A \subseteq\{1,\ldots,k-1\} A ⊆ { 1 , … , k − 1 } with A ∪ { k } A \cup\{k\} A ∪ { k } . Iterated expectations give
E [ ψ N , S , A ∪ { k } ( Ξ A ∪ { k } ; w ) ∣ Ξ 1 , … , Ξ k − 1 ] = ψ N , S , A ( Ξ A ; w ) , E\left[\psi_{N,S,A\cup\{k\}}(\Xi_{A\cup\{k\}};w)\mid\Xi_1,\ldots,\Xi_{k-1}\right]=\psi_{N,S,A}(\Xi_A;w), E [ ψ N , S , A ∪ { k } ( Ξ A ∪ { k } ; w ) ∣ Ξ 1 , … , Ξ k − 1 ] = ψ N , S , A ( Ξ A ; w ) ,
while the two Möbius coefficients have opposite signs. Therefore,
E [ ϕ N , S , k 0 ( Ξ 1 , … , Ξ k ; w ) ∣ Ξ 1 , … , Ξ k − 1 ] = 0. E\left[\phi^{0}_{N,S,k}(\Xi_1,\ldots,\Xi_k;w)\mid\Xi_1,\ldots,\Xi_{k-1}\right]=0. E [ ϕ N , S , k 0 ( Ξ 1 , … , Ξ k ; w ) ∣ Ξ 1 , … , Ξ k − 1 ] = 0.
Symmetry gives the same result after conditioning on any k − 1 k-1 k − 1 arguments. I now derive the orders of the higher-order projections to show that
E ∫ ϕ N , S , k 0 ( Ξ 1 , … , Ξ k ; w ) 2 π ( w ) d w ≤ C h N − k d X , k = 1 , … , p . (16) E\int\phi^{0}_{N,S,k}(\Xi_1,\ldots,\Xi_k;w)^2\pi(w)\,dw\le Ch_N^{-kd_X},\qquad k=1,\ldots,p. \tag*{(16)} E ∫ ϕ N , S , k 0 ( Ξ 1 , … , Ξ k ; w ) 2 π ( w ) d w ≤ C h N − k d X , k = 1 , … , p . (16)
Since ϕ N , S , k 0 \phi^{0}_{N,S,k} ϕ N , S , k 0 is a signed sum of the ψ N , S , A \psi_{N,S,A} ψ N , S , A with A ⊆ { 1 , … , k } A\subseteq\{1,\ldots,k\} A ⊆ { 1 , … , k } by (14) , and h N − r d X ≤ h N − k d X h_N^{-rd_X}\le h_N^{-kd_X} h N − r d X ≤ h N − k d X for r ≤ k r\le k r ≤ k once h N ≤ 1 h_N\le1 h N ≤ 1 , it suffices to show that
E ∫ ψ N , S , A ( Ξ A ; w ) 2 π ( w ) d w ≤ C h N − ∣ A ∣ d X for every A ⊆ { 1 , … , p } . (17) E\int\psi_{N,S,A}(\Xi_A;w)^2\pi(w)\,dw\le Ch_N^{-|A|d_X}\qquad\text{for every }A\subseteq\{1,\ldots,p\}. \tag*{(17)} E ∫ ψ N , S , A ( Ξ A ; w ) 2 π ( w ) d w ≤ C h N − ∣ A ∣ d X for every A ⊆ { 1 , … , p } . (17)
For B ⊆ { 1 , … , p } B\subseteq\{1,\ldots,p\} B ⊆ { 1 , … , p } , write z = ( z 1 , … , z p ) ∈ R d W z=(z_1,\ldots,z_p)\in\mathbb{R}^{d_W} z = ( z 1 , … , z p ) ∈ R d W with blocks z a ∈ R d X z_a\in\mathbb{R}^{d_X} z a ∈ R d X , let z B : = ( z a ) a ∈ B z_B:=(z_a)_{a\in B} z B := ( z a ) a ∈ B , and define the partially integrated kernel envelope
K ‾ B ( z B ) : = ∫ ∣ K ( z ) ∣ d z B c , \overline{K}_B(z_B):=\int|K(z)|\,dz_{B^c}, K B ( z B ) := ∫ ∣ K ( z ) ∣ d z B c ,
the integral of ∣ K ∣ |K| ∣ K ∣ over the blocks not in B B B . By Assumption 4.9, sup z B K ‾ B ( z B ) ≤ C 2 \sup_{z_B}\overline{K}_B(z_B)\le C_2 sup z B K B ( z B ) ≤ C 2 and ∫ K ‾ B ( z B ) d z B = ∫ ∣ K ( z ) ∣ d z ≤ C 1 \int\overline{K}_B(z_B)\,dz_B=\int|K(z)|\,dz\le C_1 ∫ K B ( z B ) d z B = ∫ ∣ K ( z ) ∣ d z ≤ C 1 ; hence ∥ K ‾ B ∥ L 2 2 ≤ C 1 C 2 \|\overline{K}_B\|_{L^2}^2\le C_1C_2 ∥ K B ∥ L 2 2 ≤ C 1 C 2 . Fix A ⊆ { 1 , … , p } A\subseteq\{1,\ldots,p\} A ⊆ { 1 , … , p } with ∣ A ∣ = r ≥ 1 |A|=r\ge1 ∣ A ∣ = r ≥ 1 , and write its conditioned arguments as ξ 1 , … , ξ r \xi_1,\ldots,\xi_r ξ 1 , … , ξ r with ξ j = ( x j , u j ) \xi_j=(x_j,u_j) ξ j = ( x j , u j ) . In the definition of ψ N , S , A \psi_{N,S,A} ψ N , S , A , each permutation in the symmetrization of Υ ‾ N , { 1 , … , p } 0 \overline{\Upsilon}^{0}_{N,\{1,\ldots,p\}} Υ N , { 1 , … , p } 0 places the r r r conditioned vertices into r r r of the p p p coordinate blocks of w = ( w 1 , … , w p ) w=(w_1,\ldots,w_p) w = ( w 1 , … , w p ) , and the conditional expectation integrates the remaining p − r p-r p − r vertices against the kernel. Permutations that place the conditioned vertices in the same blocks in the same order give the same conditional expectation, so ψ N , S , A ( ξ A ; w ) \psi_{N,S,A}(\xi_A;w) ψ N , S , A ( ξ A ; w ) is an average of at most p ! / ( p − r ) ! p!/(p-r)! p ! / ( p − r )! distinct terms, one for each placement τ \tau τ of the conditioned vertices into distinct blocks; write B τ B_\tau B τ for the set of blocks used, w B τ w_{B_\tau} w B τ for those blocks of w w w , and x B τ x_{B_\tau} x B τ for the conditioned covariates placed in them. For w ∈ W π w\in\mathcal{W}_\pi w ∈ W π , since ∣ μ S − P ∣ ≤ 1 |\mu_S-P|\le1 ∣ μ S − P ∣ ≤ 1 and f W ( w ) − 1 ≤ f − 1 f_W(w)^{-1}\le f^{-1} f W ( w ) − 1 ≤ f − 1 by Assumption 5.1(i), I bound the densities of the p − r p-r p − r free vertices by B 1 B_1 B 1 by Assumption 4.8, and substitute z B τ = ( w B τ − x B τ ) / h N z_{B_\tau}=(w_{B_\tau}-x_{B_\tau})/h_N z B τ = ( w B τ − x B τ ) / h N for change of variables, which gives
∣ ψ N , S , A ( ξ A ; w ) ∣ ≤ C h N − d W h N ( p − r ) d X ∑ τ K ‾ B τ ( w B τ − x B τ h N ) = C h N − r d X ∑ τ K ‾ B τ ( w B τ − x B τ h N ) , \left|\psi_{N,S,A}(\xi_A;w)\right|\le C h_N^{-d_W}h_N^{(p-r)d_X}\sum_\tau\overline{K}_{B_\tau}\left(\frac{w_{B_\tau}-x_{B_\tau}}{h_N}\right)=C h_N^{-rd_X}\sum_\tau\overline{K}_{B_\tau}\left(\frac{w_{B_\tau}-x_{B_\tau}}{h_N}\right), ∣ ψ N , S , A ( ξ A ; w ) ∣ ≤ C h N − d W h N ( p − r ) d X τ ∑ K B τ ( h N w B τ − x B τ ) = C h N − r d X τ ∑ K B τ ( h N w B τ − x B τ ) ,
where h N − d W h_N^{-d_W} h N − d W is the factor in K h K_h K h and h N ( p − r ) d X h_N^{(p-r)d_X} h N ( p − r ) d X the Jacobian of the substitution. Squaring, using ( ∑ τ a τ ) 2 ≤ M ∑ τ a τ 2 \left(\sum_\tau a_\tau\right)^2\le M\sum_\tau a_\tau^2 ( ∑ τ a τ ) 2 ≤ M ∑ τ a τ 2 with M ≤ p ! / ( p − r ) ! M\le p!/(p-r)! M ≤ p ! / ( p − r )! , and integrating against π \pi π gives, for every ξ A \xi_A ξ A ,
∫ ψ N , S , A ( ξ A ; w ) 2 π ( w ) d w ≤ C h N − 2 r d X ∑ τ ∫ K ‾ B τ ( w B τ − x B τ h N ) 2 π ( w ) d w ≤ C h N − 2 r d X ∑ τ ∫ K ‾ B τ ( w B τ − x B τ h N ) 2 d w B τ = C h N − 2 r d X ⋅ h N r d X ∑ τ ∥ K ‾ B τ ∥ L 2 2 ≤ C h N − r d X , \begin{aligned}
\int\psi_{N,S,A}(\xi_A;w)^2\pi(w)\,dw\le C h_N^{-2rd_X}\sum_\tau\int\overline{K}_{B_\tau}\left(\frac{w_{B_\tau}-x_{B_\tau}}{h_N}\right)^2\pi(w)\,dw \\
&\le C h_N^{-2rd_X}\sum_\tau\int\overline{K}_{B_\tau}\left(\frac{w_{B_\tau}-x_{B_\tau}}{h_N}\right)^2dw_{B_\tau} \\
&=C h_N^{-2rd_X}\cdot h_N^{rd_X}\sum_\tau\left\|\overline{K}_{B_\tau}\right\|_{L^2}^2\le C h_N^{-rd_X},
\end{aligned} ∫ ψ N , S , A ( ξ A ; w ) 2 π ( w ) d w ≤ C h N − 2 r d X τ ∑ ∫ K B τ ( h N w B τ − x B τ ) 2 π ( w ) d w ≤ C h N − 2 r d X τ ∑ ∫ K B τ ( h N w B τ − x B τ ) 2 d w B τ = C h N − 2 r d X ⋅ h N r d X τ ∑ K B τ L 2 2 ≤ C h N − r d X ,
where the second inequality integrates the p − r p-r p − r blocks of w w w outside B τ B_\tau B τ against π \pi π , which is bounded with compact support, and the equality substitutes u = ( w B τ − x B τ ) / h N u=(w_{B_\tau}-x_{B_\tau})/h_N u = ( w B τ − x B τ ) / h N in the r r r blocks of B τ B_\tau B τ . The bound does not depend on ξ A \xi_A ξ A , so
E ∫ ψ N , S , A ( Ξ A ; w ) 2 π ( w ) d w ≤ C h N − r d X , E\int\psi_{N,S,A}(\Xi_A;w)^2\pi(w)\,dw\le C h_N^{-rd_X}, E ∫ ψ N , S , A ( Ξ A ; w ) 2 π ( w ) d w ≤ C h N − r d X ,
which is (17) , and hence (16) . Next, terms indexed by distinct q k ≠ r k q_k\ne r_k q k = r k in C k , N C_{k,N} C k , N are uncorrelated. Pick i ∈ q k ∖ r k i\in q_k\setminus r_k i ∈ q k ∖ r k and condition on { Ξ j : j ≠ i } \{\Xi_j:j\ne i\} { Ξ j : j = i } : the term indexed by r k r_k r k is a function of the conditioning variables, while the term indexed by q k q_k q k has conditional mean zero, because ϕ N , S , k 0 \phi^0_{N,S,k} ϕ N , S , k 0 is canonical and Ξ i \Xi_i Ξ i is independent of the other vertex variables. Consequently,
E ∫ U N , S , k ( w ) 2 π ( w ) d w = ( p k ) 2 ( N k ) E ∫ ϕ N , S , k 0 ( Ξ 1 , … , Ξ k ; w ) 2 π ( w ) d w ≤ C N − k h N − k d X . E\int U_{N,S,k}(w)^2\pi(w)\,dw=\frac{\binom{p}{k}^2}{\binom{N}{k}}E\int\phi^0_{N,S,k}(\Xi_1,\ldots,\Xi_k;w)^2\pi(w)\,dw\le CN^{-k}h_N^{-kd_X}. E ∫ U N , S , k ( w ) 2 π ( w ) d w = ( k N ) ( k p ) 2 E ∫ ϕ N , S , k 0 ( Ξ 1 , … , Ξ k ; w ) 2 π ( w ) d w ≤ C N − k h N − k d X .
In particular, for k = 1 k=1 k = 1 , N h N d X / 2 ∫ U N , S , 1 ( w ) 2 π ( w ) d w = O p ( h N − d X / 2 ) Nh_N^{d_X/2}\int U_{N,S,1}(w)^2\pi(w)\,dw=O_p(h_N^{-d_X/2}) N h N d X /2 ∫ U N , S , 1 ( w ) 2 π ( w ) d w = O p ( h N − d X /2 ) . For k ≥ 2 k\ge2 k ≥ 2 , orthogonality across Hoeffding orders, which follows by the same argument, implies
E ∫ { ∑ k = 2 p U N , S , k ( w ) } 2 π ( w ) d w ≤ C ∑ k = 2 p N − k h N − k d X . E\int\left\{\sum_{k=2}^p U_{N,S,k}(w)\right\}^2\pi(w)\,dw\le C\sum_{k=2}^pN^{-k}h_N^{-kd_X}. E ∫ { k = 2 ∑ p U N , S , k ( w ) } 2 π ( w ) d w ≤ C k = 2 ∑ p N − k h N − k d X .
For the edge-shock remainder, conditional independence implies that its conditional covariance is zero for two p p p -subsets sharing at most one vertex, because their induced subgraphs contain no common dyadic shock. For a pair sharing t ≥ 2 t\ge2 t ≥ 2 vertices, partial kernel integration gives a bound C h N − t d X Ch_N^{-td_X} C h N − t d X , and there are O ( N 2 p − t ) O(N^{2p-t}) O ( N 2 p − t ) such pairs. Hence
E ∫ [ V N , S e ( w ) ] 2 π ( w ) d w ≤ C ∑ t = 2 p N 1 − t h N − t d X . E\int\left[V^e_{N,S}(w)\right]^2\pi(w)\,dw \le C\sum_{t=2}^{p}N^{1-t}h_N^{-td_X}. E ∫ [ V N , S e ( w ) ] 2 π ( w ) d w ≤ C t = 2 ∑ p N 1 − t h N − t d X .
Set R N , S H ( w ) : = ∑ k = 2 p U N , S , k ( w ) + V N , S e ( w ) R^H_{N,S}(w):=\sum_{k=2}^{p}U_{N,S,k}(w)+V^e_{N,S}(w) R N , S H ( w ) := ∑ k = 2 p U N , S , k ( w ) + V N , S e ( w ) , so that U N 1 = U N , S , 1 + R N , S H U_{N1}=U_{N,S,1}+R^H_{N,S} U N 1 = U N , S , 1 + R N , S H . Since ∑ k = 2 p U N , S , k \sum_{k=2}^{p}U_{N,S,k} ∑ k = 2 p U N , S , k is a function of Ξ 1 , … , Ξ N \Xi_1,\ldots,\Xi_N Ξ 1 , … , Ξ N alone whereas E [ V N , S e ∣ Ξ 1 , … , Ξ N ] = 0 E[V^e_{N,S}\mid\Xi_1,\ldots,\Xi_N]=0 E [ V N , S e ∣ Ξ 1 , … , Ξ N ] = 0 , the two parts of R N , S H R^H_{N,S} R N , S H are orthogonal. Therefore,
E [ N h N d X / 2 ∫ [ R N , S H ( w ) ] 2 π ( w ) d w ] ≤ C ∑ k = 2 p N 1 − k h N − ( k − 1 / 2 ) d X = O ( 1 N h N 3 d X / 2 ) . E\left[Nh_N^{d_X/2}\int\left[R^H_{N,S}(w)\right]^2\pi(w)\,dw\right]\le C\sum_{k=2}^{p}N^{1-k}h_N^{-(k-1/2)d_X}=O\left(\frac{1}{Nh_N^{3d_X/2}}\right). E [ N h N d X /2 ∫ [ R N , S H ( w ) ] 2 π ( w ) d w ] ≤ C k = 2 ∑ p N 1 − k h N − ( k − 1/2 ) d X = O ( N h N 3 d X /2 1 ) .
Thus
N h N d X / 2 ∫ [ R N , S H ( w ) ] 2 π ( w ) d w = O p ( 1 N h N 3 d X / 2 ) = o p ( 1 ) . (18) Nh_N^{d_X/2}\int\left[R^H_{N,S}(w)\right]^2\pi(w)\,dw=O_p\left(\frac{1}{Nh_N^{3d_X/2}}\right)=o_p(1). \tag*{(18)} N h N d X /2 ∫ [ R N , S H ( w ) ] 2 π ( w ) d w = O p ( N h N 3 d X /2 1 ) = o p ( 1 ) . (18)
By the Cauchy–Schwarz inequality and the bounds on U N , S , 1 U_{N,S,1} U N , S , 1 and R N , S H R^H_{N,S} R N , S H above,
∣ N h N d X / 2 ∫ U N , S , 1 ( w ) R N , S H ( w ) π ( w ) d w ∣ ≤ [ N h N d X / 2 ∫ U N , S , 1 ( w ) 2 π ( w ) d w ] 1 / 2 × [ N h N d X / 2 ∫ R N , S H ( w ) 2 π ( w ) d w ] 1 / 2 = O p ( h N − d X / 4 ) O p ( N − 1 / 2 h N − 3 d X / 4 ) = O p ( N − 1 / 2 h N − d X ) = O p ( 1 N 2 h N 2 d X ) = o p ( 1 ) . \begin{aligned}
\left|Nh_N^{d_X/2}\int U_{N,S,1}(w)R^H_{N,S}(w)\pi(w)\,dw\right|
&\le\left[Nh_N^{d_X/2}\int U_{N,S,1}(w)^2\pi(w)\,dw\right]^{1/2}\\
&\quad{}\times\left[Nh_N^{d_X/2}\int R^H_{N,S}(w)^2\pi(w)\,dw\right]^{1/2}\\
&=O_p(h_N^{-d_X/4})O_p\left(N^{-1/2}h_N^{-3d_X/4}\right)\\
&=O_p\left(N^{-1/2}h_N^{-d_X}\right)=O_p\left(\frac{1}{\sqrt{N^2h_N^{2d_X}}}\right)=o_p(1).
\end{aligned} N h N d X /2 ∫ U N , S , 1 ( w ) R N , S H ( w ) π ( w ) d w ≤ [ N h N d X /2 ∫ U N , S , 1 ( w ) 2 π ( w ) d w ] 1/2 × [ N h N d X /2 ∫ R N , S H ( w ) 2 π ( w ) d w ] 1/2 = O p ( h N − d X /4 ) O p ( N − 1/2 h N − 3 d X /4 ) = O p ( N − 1/2 h N − d X ) = O p N 2 h N 2 d X 1 = o p ( 1 ) .
Thus,
N h N d X / 2 ∫ U N 1 ( w ) 2 π ( w ) d w = N h N d X / 2 ∫ U N , S , 1 ( w ) 2 π ( w ) d w + o p ( 1 ) . (19) Nh_N^{d_X/2}\int U_{N1}(w)^2\pi(w)\,dw=Nh_N^{d_X/2}\int U_{N,S,1}(w)^2\pi(w)\,dw+o_p(1). \tag*{(19)} N h N d X /2 ∫ U N 1 ( w ) 2 π ( w ) d w = N h N d X /2 ∫ U N , S , 1 ( w ) 2 π ( w ) d w + o p ( 1 ) . (19)
Step 5: Diagonal and off-diagonal decomposition.
Substituting (15) into (19) and separating the diagonal from the off-diagonal terms gives
N h N d X / 2 ∫ U N , S , 1 ( w ) 2 π ( w ) d w = B N , S + W N , S , Nh_N^{d_X/2}\int U_{N,S,1}(w)^2\pi(w)\,dw=B_{N,S}+W_{N,S}, N h N d X /2 ∫ U N , S , 1 ( w ) 2 π ( w ) d w = B N , S + W N , S ,
where, with c N : = 2 p 2 h N d X / 2 / N c_N:=2p^2h_N^{d_X/2}/N c N := 2 p 2 h N d X /2 / N ,
B N , S : = c N 2 ∑ i = 1 N ∫ ℓ N , S ( Ξ i ; w ) 2 π ( w ) d w , B_{N,S}:=\frac{c_N}{2}\sum_{i=1}^{N}\int\ell_{N,S}(\Xi_i;w)^2\pi(w)\,dw, B N , S := 2 c N i = 1 ∑ N ∫ ℓ N , S ( Ξ i ; w ) 2 π ( w ) d w ,
and
W N , S : = c N ∑ 1 ≤ i < j ≤ N ∫ ℓ N , S ( Ξ i ; w ) ℓ N , S ( Ξ j ; w ) π ( w ) d w . W_{N,S} := c_N \sum_{1\le i<j\le N} \int\ell_{N,S}(\Xi_i;w)\ell_{N,S}(\Xi_j;w)\pi(w)\,dw. W N , S := c N 1 ≤ i < j ≤ N ∑ ∫ ℓ N , S ( Ξ i ; w ) ℓ N , S ( Ξ j ; w ) π ( w ) d w .
By the definition of Γ N , S \Gamma_{N,S} Γ N , S ,
E [ B N , S ] = p 2 h N d X / 2 ∫ E [ ℓ N , S ( Ξ 1 ; w ) 2 ] π ( w ) d w = h N − d X / 2 ∫ Γ N , S ( w , w ) π ( w ) d w = b N , S . E[B_{N,S}] = p^2 h_N^{d_X/2} \int E[\ell_{N,S}(\Xi_1;w)^2]\pi(w)\,dw = h_N^{-d_X/2}\int\Gamma_{N,S}(w,w)\pi(w)\,dw = b_{N,S}. E [ B N , S ] = p 2 h N d X /2 ∫ E [ ℓ N , S ( Ξ 1 ; w ) 2 ] π ( w ) d w = h N − d X /2 ∫ Γ N , S ( w , w ) π ( w ) d w = b N , S .
To bound Var [ B N , S ] \operatorname{Var}[B_{N,S}] Var [ B N , S ] , I first bound ℓ N , S \ell_{N,S} ℓ N , S pointwise. For w ∈ W π w \in\mathcal{W}_{\pi} w ∈ W π , since ∣ μ S − P ∣ ≤ 1 |\mu_S-P|\le1 ∣ μ S − P ∣ ≤ 1 and f W ( w ) ≥ f ‾ f_W(w)\ge\underline{f} f W ( w ) ≥ f by Assumption 5.1(i), the definition (5) gives
∣ ℓ N , S ( Ξ i ; w ) ∣ ≤ 1 p f ‾ ∑ a = 1 p E ∣ K { a } { w − X ( a ) ( Ξ i ) } ∣ ≤ B 1 p − 1 p f ‾ h N − d W ∑ a = 1 p ∫ ∣ K ( w − x ( a ) h N ) ∣ d x { a } c = B 1 p − 1 p f ‾ h N − d W h N ( p − 1 ) d X ∑ a = 1 p K ‾ { a } ( w a − X i h N ) ≤ C h N − d X ∑ a = 1 p K ‾ { a } ( w a − X i h N ) , \begin{aligned}
\left|\ell_{N,S}(\Xi_i;w)\right| \le\frac{1}{p\underline{f}}\sum_{a=1}^{p} E\left|K_{\{a\}}\{w-X^{(a)}(\Xi_i)\}\right| \\
&\le\frac{B_1^{p-1}}{p\underline{f}}h_N^{-d_W}\sum_{a=1}^{p}\int\left|K\left(\frac{w-x^{(a)}}{h_N}\right)\right|dx_{\{a\}^{c}} \\
&= \frac{B_1^{p-1}}{p\underline{f}}h_N^{-d_W}h_N^{(p-1)d_X}\sum_{a=1}^{p}\overline{K}_{\{a\}}\left(\frac{w_a-X_i}{h_N}\right)\le Ch_N^{-d_X}\sum_{a=1}^{p}\overline{K}_{\{a\}}\left(\frac{w_a-X_i}{h_N}\right),
\end{aligned} ∣ ℓ N , S ( Ξ i ; w ) ∣ ≤ p f 1 a = 1 ∑ p E K { a } { w − X ( a ) ( Ξ i )} ≤ p f B 1 p − 1 h N − d W a = 1 ∑ p ∫ K ( h N w − x ( a ) ) d x { a } c = p f B 1 p − 1 h N − d W h N ( p − 1 ) d X a = 1 ∑ p K { a } ( h N w a − X i ) ≤ C h N − d X a = 1 ∑ p K { a } ( h N w a − X i ) ,
where x ( a ) x^{(a)} x ( a ) has X i X_i X i in block a a a and the integration variables x b x_b x b , b ≠ a b\ne a b = a , in the other blocks; the second inequality bounds the densities of the p − 1 p-1 p − 1 free vertices by B 1 B_1 B 1 (Assumption 4.8), and the equality substitutes u b = ( w b − x b ) / h N u_b=(w_b-x_b)/h_N u b = ( w b − x b ) / h N for b ≠ a b\ne a b = a , with Jacobian h N ( p − 1 ) d X h_N^{(p-1)d_X} h N ( p − 1 ) d X . Then,
∫ ℓ N , S ( Ξ i ; w ) 2 π ( w ) d w ≤ C p h N − 2 d X ∑ a = 1 p ∫ K ‾ { a } ( w a − X i h N ) 2 π ( w ) d w ≤ C h N − 2 d X ∑ a = 1 p ∫ K ‾ { a } ( w a − X i h N ) 2 d w a = C h N − 2 d X ⋅ h N d X ∑ a = 1 p ∥ K ‾ { a } ∥ L 2 2 ≤ C h N − d X , \begin{aligned}
\int\ell_{N,S}(\Xi_i;w)^2\pi(w)\,dw\le Cp h_N^{-2d_X}\sum_{a=1}^{p}\int\overline{K}_{\{a\}}\left(\frac{w_a-X_i}{h_N}\right)^2\pi(w)\,dw \\
&\le Ch_N^{-2d_X}\sum_{a=1}^{p}\int\overline{K}_{\{a\}}\left(\frac{w_a-X_i}{h_N}\right)^2dw_a \\
&= Ch_N^{-2d_X}\cdot h_N^{d_X}\sum_{a=1}^{p}\left\|\overline{K}_{\{a\}}\right\|_{L^2}^2\le Ch_N^{-d_X},
\end{aligned} ∫ ℓ N , S ( Ξ i ; w ) 2 π ( w ) d w ≤ C p h N − 2 d X a = 1 ∑ p ∫ K { a } ( h N w a − X i ) 2 π ( w ) d w ≤ C h N − 2 d X a = 1 ∑ p ∫ K { a } ( h N w a − X i ) 2 d w a = C h N − 2 d X ⋅ h N d X a = 1 ∑ p K { a } L 2 2 ≤ C h N − d X ,
where the second inequality integrates the p − 1 p-1 p − 1 blocks of w w w other than w a w_a w a against π \pi π , which is bounded with compact support; the equality substitutes u = ( w a − X i ) / h N u=(w_a-X_i)/h_N u = ( w a − X i ) / h N in block $a; and the last inequality uses ∥ K ‾ { a } ∥ L 2 2 ≤ C 1 C 2 \left\|\overline{K}_{\{a\}}\right\|_{L^2}^2\le C_1C_2 K { a } L 2 2 ≤ C 1 C 2 from Step 4. Since this bound holds for every realization of Ξ i \Xi_i Ξ i , squaring it and taking expectations gives
E [ { ∫ ℓ N , S ( Ξ i ; w ) 2 π ( w ) d w } 2 ] ≤ C h N − 2 d X . E\left[\left\{\int\ell_{N,S}(\Xi_i;w)^2\pi(w)\,dw\right\}^2\right]\le Ch_N^{-2d_X}. E [ { ∫ ℓ N , S ( Ξ i ; w ) 2 π ( w ) d w } 2 ] ≤ C h N − 2 d X .
Since the summands in B N , S B_{N,S} B N , S are i.i.d.,
Var [ B N , S ] ≤ p 4 h N d X N C h N − 2 d X = O ( 1 N h N d X ) = o ( 1 ) . \operatorname{Var}[B_{N,S}]\le\frac{p^4h_N^{d_X}}{N}Ch_N^{-2d_X}=O\left(\frac{1}{Nh_N^{d_X}}\right)=o(1). Var [ B N , S ] ≤ N p 4 h N d X C h N − 2 d X = O ( N h N d X 1 ) = o ( 1 ) .
Therefore, B N , S = b N , S + o p ( 1 ) B_{N,S}=b_{N,S}+o_p(1) B N , S = b N , S + o p ( 1 ) .
Step 6: de Jong's CLT for the off-diagonal quadratic term.
For i < j i<j i < j , define
W N , i j , S : = c N ∫ ℓ N , S ( Ξ i ; w ) ℓ N , S ( Ξ j ; w ) π ( w ) d w . W_{N,ij,S}:=c_N\int\ell_{N,S}(\Xi_i;w)\ell_{N,S}(\Xi_j;w)\pi(w)\,dw. W N , ij , S := c N ∫ ℓ N , S ( Ξ i ; w ) ℓ N , S ( Ξ j ; w ) π ( w ) d w .
Then W N , S = ∑ i < j W N , i j , S W_{N,S}=\sum_{i<j}W_{N,ij,S} W N , S = ∑ i < j W N , ij , S . I apply the central limit theorem for clean quadratic forms of de Jong [18 ] by verifying its three conditions.
First, the array is clean, that is, E [ W N , i j , S ∣ Ξ i ] = E [ W N , i j , S ∣ Ξ j ] = 0 E[W_{N,ij,S}\mid\Xi_i]=E[W_{N,ij,S}\mid\Xi_j]=0 E [ W N , ij , S ∣ Ξ i ] = E [ W N , ij , S ∣ Ξ j ] = 0 for i ≠ j i\ne j i = j . Indeed, since Ξ j \Xi_j Ξ j is independent of Ξ i \Xi_i Ξ i and E [ ℓ N , S ( Ξ j ; w ) ] = 0 E[\ell_{N,S}(\Xi_j;w)]=0 E [ ℓ N , S ( Ξ j ; w )] = 0 for every w w w by (5) ,
E [ W N , i j , S ∣ Ξ i ] = c N ∫ ℓ N , S ( Ξ i ; w ) E [ ℓ N , S ( Ξ j ; w ) ] π ( w ) d w = 0 , E[W_{N,ij,S}\mid\Xi_i]=c_N\int\ell_{N,S}(\Xi_i;w)E[\ell_{N,S}(\Xi_j;w)]\pi(w)\,dw=0, E [ W N , ij , S ∣ Ξ i ] = c N ∫ ℓ N , S ( Ξ i ; w ) E [ ℓ N , S ( Ξ j ; w )] π ( w ) d w = 0 ,
and the same argument gives E [ W N , i j , S ∣ Ξ j ] = 0 E[W_{N,ij,S}\mid\Xi_j]=0 E [ W N , ij , S ∣ Ξ j ] = 0 .
Second, let s N , S 2 : = Var [ W N , S ] s_{N,S}^2:=\operatorname{Var}[W_{N,S}] s N , S 2 := Var [ W N , S ] . If two unordered index pairs are disjoint, the corresponding W W W 's are independent; if they share one index, cleanness gives zero covariance by conditioning on that index. Hence
s N , S 2 = ( N 2 ) E [ W N , 12 , S 2 ] . s_{N,S}^2=\binom{N}{2}E[W_{N,12,S}^2]. s N , S 2 = ( 2 N ) E [ W N , 12 , S 2 ] .
Independence of Ξ 1 \Xi_1 Ξ 1 and Ξ 2 \Xi_2 Ξ 2 gives
E [ W N , 12 , S 2 ] = c N 2 ∬ E [ ℓ N , S ( Ξ 1 ; w ) ℓ N , S ( Ξ 1 ; z ) ] 2 π ( w ) π ( z ) d w d z = 4 h N 2 d X N 2 ∬ Γ N , S ( w , z ) 2 π ( w ) π ( z ) d w d z . \begin{aligned}
E[W_{N,12,S}^2]
&=c_N^2\iint E[\ell_{N,S}(\Xi_1;w)\ell_{N,S}(\Xi_1;z)]^2\pi(w)\pi(z)\,dw\,dz\\
&=\frac{4h_N^{2d_X}}{N^2}\iint\Gamma_{N,S}(w,z)^2\pi(w)\pi(z)\,dw\,dz.
\end{aligned} E [ W N , 12 , S 2 ] = c N 2 ∬ E [ ℓ N , S ( Ξ 1 ; w ) ℓ N , S ( Ξ 1 ; z ) ] 2 π ( w ) π ( z ) d w d z = N 2 4 h N 2 d X ∬ Γ N , S ( w , z ) 2 π ( w ) π ( z ) d w d z .
Thus
s N , S 2 = N − 1 N ν N , S → ν S . s_{N,S}^2=\frac{N-1}{N}\nu_{N,S}\to\nu_S. s N , S 2 = N N − 1 ν N , S → ν S .
For every i i i ,
∑ j ≠ i E [ W N , i j , S 2 ] = ( N − 1 ) E [ W N , 12 , S 2 ] = 2 N s N , S 2 . \sum_{j\ne i}E[W_{N,ij,S}^2]=(N-1)E[W_{N,12,S}^2]=\frac{2}{N}s_{N,S}^2. j = i ∑ E [ W N , ij , S 2 ] = ( N − 1 ) E [ W N , 12 , S 2 ] = N 2 s N , S 2 .
Therefore de Jong's row-negligibility condition holds:
1 s N , S 2 max i ∑ j ≠ i E [ W N , i j , S 2 ] = 2 N → 0. \frac{1}{s_{N,S}^2}\max_i\sum_{j\ne i}E[W_{N,ij,S}^2]=\frac{2}{N}\to0. s N , S 2 1 i max j = i ∑ E [ W N , ij , S 2 ] = N 2 → 0.
Finally, I verify the fourth-moment condition. Substituting the pointwise bound on ℓ N , S \ell_{N,S} ℓ N , S from Step 5 gives
∣ ∫ ℓ N , S ( Ξ i ; w ) ℓ N , S ( Ξ j ; w ) π ( w ) d w ∣ ≤ C h N − 2 d X ∑ a , b = 1 p ∫ K ‾ { a } ( w a − X i h N ) K ‾ { b } ( w b − X j h N ) π ( w ) d w . \left|\int\ell_{N,S}(\Xi_i;w)\ell_{N,S}(\Xi_j;w)\pi(w)\,dw\right|
\le Ch_N^{-2d_X}\sum_{a,b=1}^{p}\int\overline{K}_{\{a\}}\left(\frac{w_a-X_i}{h_N}\right)\overline{K}_{\{b\}}\left(\frac{w_b-X_j}{h_N}\right)\pi(w)\,dw. ∫ ℓ N , S ( Ξ i ; w ) ℓ N , S ( Ξ j ; w ) π ( w ) d w ≤ C h N − 2 d X a , b = 1 ∑ p ∫ K { a } ( h N w a − X i ) K { b } ( h N w b − X j ) π ( w ) d w .
For a ≠ b a \ne b a = b , the integrand depends on w w w through two different blocks; integrating the other blocks against π \pi π and substituting u a = ( w a − X i ) / h N u_a=(w_a-X_i)/h_N u a = ( w a − X i ) / h N and u b = ( w b − X j ) / h N u_b=(w_b-X_j)/h_N u b = ( w b − X j ) / h N gives a Jacobian h N 2 d X h_N^{2d_X} h N 2 d X and leaves ∥ K ‾ { a } ∥ L 1 ∥ K ‾ { b } ∥ L 1 ≤ C 1 2 \lVert\overline{K}_{\{a\}}\rVert_{L^1}\lVert\overline{K}_{\{b\}}\rVert_{L^1}\le C_1^2 ∥ K { a } ∥ L 1 ∥ K { b } ∥ L 1 ≤ C 1 2 , so these terms are bounded by a constant. For a = b a=b a = b , the same substitution in block a a a alone gives a Jacobian h N d X h_N^{d_X} h N d X and leaves κ a ( ( X i − X j ) / h N ) \kappa_a((X_i-X_j)/h_N) κ a (( X i − X j ) / h N ) , where
κ a ( v ) : = ∫ K ‾ { a } ( u ) K ‾ { a } ( u + v ) d u \kappa_a(v):=\int\overline{K}_{\{a\}}(u)\overline{K}_{\{a\}}(u+v)\,du κ a ( v ) := ∫ K { a } ( u ) K { a } ( u + v ) d u
is bounded by C 1 C 2 C_1C_2 C 1 C 2 and integrable with ∫ κ a = ∥ K ‾ { a } ∥ L 1 2 ≤ C 1 2 \int\kappa_a=\lVert\overline{K}_{\{a\}}\rVert_{L^1}^2\le C_1^2 ∫ κ a = ∥ K { a } ∥ L 1 2 ≤ C 1 2 . With κ : = ∑ a = 1 p κ a \kappa:=\sum_{a=1}^p\kappa_a κ := ∑ a = 1 p κ a , therefore,
∣ ∫ ℓ N , S ( Ξ i ; w ) ℓ N , S ( Ξ j ; w ) π ( w ) d w ∣ ≤ C [ 1 + h N − d X κ ( X i − X j h N ) ] . (20) \left|\int\ell_{N,S}(\Xi_i;w)\ell_{N,S}(\Xi_j;w)\pi(w)\,dw\right|\le C\left[1+h_N^{-d_X}\kappa\left(\frac{X_i-X_j}{h_N}\right)\right]. \tag*{(20)} ∫ ℓ N , S ( Ξ i ; w ) ℓ N , S ( Ξ j ; w ) π ( w ) d w ≤ C [ 1 + h N − d X κ ( h N X i − X j ) ] . (20)
For r ≥ 1 r\ge1 r ≥ 1 and j ≠ 1 j\ne1 j = 1 , the substitution X j = X 1 − h N u X_j=X_1-h_Nu X j = X 1 − h N u gives, uniformly in X 1 X_1 X 1 ,
E [ κ ( X 1 − X j h N ) r | X 1 ] = h N d X ∫ κ ( u ) r f ( X 1 − h N u ) d u ≤ B 1 ∥ κ ∥ ∞ r − 1 ∥ κ ∥ L 1 h N d X , E\left[\kappa\left(\frac{X_1-X_j}{h_N}\right)^r\middle|X_1\right]=h_N^{d_X}\int\kappa(u)^rf(X_1-h_Nu)\,du\le B_1\lVert\kappa\rVert_\infty^{r-1}\lVert\kappa\rVert_{L^1}h_N^{d_X}, E [ κ ( h N X 1 − X j ) r X 1 ] = h N d X ∫ κ ( u ) r f ( X 1 − h N u ) d u ≤ B 1 ∥ κ ∥ ∞ r − 1 ∥ κ ∥ L 1 h N d X ,
so that, by ( 1 + x ) r ≤ 2 r − 1 ( 1 + x r ) (1+x)^r\le2^{r-1}(1+x^r) ( 1 + x ) r ≤ 2 r − 1 ( 1 + x r ) , E [ { 1 + h N − d X κ ( X 1 − X j h N ) } r | X 1 ] = O ( h N − ( r − 1 ) d X ) E\left[\left\{1+h_N^{-d_X}\kappa\left(\frac{X_1-X_j}{h_N}\right)\right\}^r\middle|X_1\right]=O(h_N^{-(r-1)d_X}) E [ { 1 + h N − d X κ ( h N X 1 − X j ) } r X 1 ] = O ( h N − ( r − 1 ) d X ) uniformly in X 1 X_1 X 1 . Combined with (20) , this gives
E [ { ∫ ℓ N , S ( Ξ 1 ; w ) ℓ N , S ( Ξ 2 ; w ) π ( w ) d w } r ] = O ( h N − ( r − 1 ) d X ) , (21) E\left[\left\{\int\ell_{N,S}(\Xi_1;w)\ell_{N,S}(\Xi_2;w)\pi(w)\,dw\right\}^r\right]=O(h_N^{-(r-1)d_X}), \tag*{(21)} E [ { ∫ ℓ N , S ( Ξ 1 ; w ) ℓ N , S ( Ξ 2 ; w ) π ( w ) d w } r ] = O ( h N − ( r − 1 ) d X ) , (21)
in particular O ( h N − d X ) O(h_N^{-d_X}) O ( h N − d X ) for r = 2 r=2 r = 2 and O ( h N − 3 d X ) O(h_N^{-3d_X}) O ( h N − 3 d X ) for r = 4 r=4 r = 4 .
Conditional on X 1 X_1 X 1 , the factors involving X 2 X_2 X 2 and X 3 X_3 X 3 are independent, so (20) and the conditional bound with r = 2 r=2 r = 2 , applied to each factor, give
E [ { ∫ ℓ N , S ( Ξ 1 ; w ) ℓ N , S ( Ξ 2 ; w ) π ( w ) d w } 2 { ∫ ℓ N , S ( Ξ 1 ; w ) ℓ N , S ( Ξ 3 ; w ) π ( w ) d w } 2 ] = O ( h N − d X ) ⋅ O ( h N − d X ) = O ( h N − 2 d X ) . (22) \begin{aligned}
E\left[\left\{\int\ell_{N,S}(\Xi_1;w)\ell_{N,S}(\Xi_2;w)\pi(w)\,dw\right\}^2\left\{\int\ell_{N,S}(\Xi_1;w)\ell_{N,S}(\Xi_3;w)\pi(w)\,dw\right\}^2\right] \\
&=O(h_N^{-d_X})\cdot O(h_N^{-d_X})=O(h_N^{-2d_X}). \tag*{(22)}
\end{aligned} E [ { ∫ ℓ N , S ( Ξ 1 ; w ) ℓ N , S ( Ξ 2 ; w ) π ( w ) d w } 2 { ∫ ℓ N , S ( Ξ 1 ; w ) ℓ N , S ( Ξ 3 ; w ) π ( w ) d w } 2 ] = O ( h N − d X ) ⋅ O ( h N − d X ) = O ( h N − 2 d X ) . (22)
For the four-cycle, (20) bounds the product of the four inner products by
C ∏ ( i , j ) ∈ { ( 1 , 2 ) , ( 2 , 3 ) , ( 3 , 4 ) , ( 4 , 1 ) } { 1 + h N − d X κ ( X i − X j h N ) } . C\prod_{(i,j)\in\{(1,2),(2,3),(3,4),(4,1)\}}\left\{1+h_N^{-d_X}\kappa\left(\frac{X_i-X_j}{h_N}\right)\right\}. C ( i , j ) ∈ {( 1 , 2 ) , ( 2 , 3 ) , ( 3 , 4 ) , ( 4 , 1 )} ∏ { 1 + h N − d X κ ( h N X i − X j ) } .
Expanding the product, the term with all four localized factors is, after the substitution X 1 = v X_1=v X 1 = v and X i = v + h N u i X_i=v+h_Nu_i X i = v + h N u i for i = 2 , 3 , 4 i=2,3,4 i = 2 , 3 , 4 with Jacobian h N 3 d X h_N^{3d_X} h N 3 d X ,
h N − 4 d X E [ κ ( X 1 − X 2 h N ) κ ( X 2 − X 3 h N ) κ ( X 3 − X 4 h N ) κ ( X 4 − X 1 h N ) ] = h N − d X ∫ κ ( − u 2 ) κ ( u 2 − u 3 ) κ ( u 3 − u 4 ) κ ( u 4 ) f ( v ) ∏ i = 2 4 f ( v + h N u i ) d u 2 d u 3 d u 4 d v = O ( h N − d X ) , h_N^{-4d_X}E\left[\kappa\left(\frac{X_1-X_2}{h_N}\right)\kappa\left(\frac{X_2-X_3}{h_N}\right)\kappa\left(\frac{X_3-X_4}{h_N}\right)\kappa\left(\frac{X_4-X_1}{h_N}\right)\right]
= h_N^{-d_X}\int\kappa(-u_2)\kappa(u_2-u_3)\kappa(u_3-u_4)\kappa(u_4)f(v)\prod_{i=2}^{4}f(v+h_Nu_i)\,du_2\,du_3\,du_4\,dv=O(h_N^{-d_X}), h N − 4 d X E [ κ ( h N X 1 − X 2 ) κ ( h N X 2 − X 3 ) κ ( h N X 3 − X 4 ) κ ( h N X 4 − X 1 ) ] = h N − d X ∫ κ ( − u 2 ) κ ( u 2 − u 3 ) κ ( u 3 − u 4 ) κ ( u 4 ) f ( v ) i = 2 ∏ 4 f ( v + h N u i ) d u 2 d u 3 d u 4 d v = O ( h N − d X ) ,
since f ≤ B 1 f\le B_1 f ≤ B 1 and κ \kappa κ is bounded and integrable. Terms with fewer localized factors are O ( 1 ) O(1) O ( 1 ) : their edges contain no cycle, so each localized factor can be substituted separately with Jacobian h N d X h_N^{d_X} h N d X . Hence
∣ E [ ∏ ( i , j ) ∈ { ( 1 , 2 ) , ( 2 , 3 ) , ( 3 , 4 ) , ( 4 , 1 ) } ∫ ℓ N , S ( Ξ i ; w ) ℓ N , S ( Ξ j ; w ) π ( w ) d w ] ∣ = O ( h N − d X ) . (23) \left|E\left[\prod_{(i,j)\in\{(1,2),(2,3),(3,4),(4,1)\}}\int\ell_{N,S}(\Xi_i;w)\ell_{N,S}(\Xi_j;w)\pi(w)\,dw\right]\right|=O(h_N^{-d_X}). \tag*{(23)} E ( i , j ) ∈ {( 1 , 2 ) , ( 2 , 3 ) , ( 3 , 4 ) , ( 4 , 1 )} ∏ ∫ ℓ N , S ( Ξ i ; w ) ℓ N , S ( Ξ j ; w ) π ( w ) d w = O ( h N − d X ) . (23)
I now expand the fourth moment
E [ W N , S 4 ] = ∑ e 1 , e 2 , e 3 , e 4 E [ W e 1 W e 2 W e 3 W e 4 ] , E[W_{N,S}^{4}]=\sum_{e_1,e_2,e_3,e_4}E[W_{e_1}W_{e_2}W_{e_3}W_{e_4}], E [ W N , S 4 ] = e 1 , e 2 , e 3 , e 4 ∑ E [ W e 1 W e 2 W e 3 W e 4 ] ,
where each e r e_r e r ranges over the unordered vertex pairs and W e : = W N , i j , S W_e:=W_{N,ij,S} W e := W N , ij , S for e = { i , j } e=\{i,j\} e = { i , j } with i < j i<j i < j . View ( e 1 , … , e 4 ) (e_1,\ldots,e_4) ( e 1 , … , e 4 ) as a multigraph on the vertices it covers. If some vertex lies in exactly one of the four edges, say e r = { i , j } e_r=\{i,j\} e r = { i , j } with j j j in no other edge, then, conditioning on { Ξ m : m ≠ j } \{\Xi_m:m\ne j\} { Ξ m : m = j } , the other three factors are functions of the conditioning variables while E [ W N , i j , S ∣ { Ξ m : m ≠ j } ] = E [ W N , i j , S ∣ Ξ i ] = 0 E[W_{N,ij,S}\mid\{\Xi_m:m\ne j\}]=E[W_{N,ij,S}\mid\Xi_i]=0 E [ W N , ij , S ∣ { Ξ m : m = j }] = E [ W N , ij , S ∣ Ξ i ] = 0 by cleanness, so the expectation vanishes. Hence only multigraphs in which every vertex has degree at least two contribute.
Up to relabeling, these are of five types: (i) all four edges equal, with O ( N 2 ) O(N^2) O ( N 2 ) configurations; (ii) two distinct doubled edges sharing one vertex, with O ( N 3 ) O(N^3) O ( N 3 ) ; (iii) a triangle with one doubled edge, with O ( N 3 ) O(N^3) O ( N 3 ) ; (iv) a four-cycle, with O ( N 4 ) O(N^4) O ( N 4 ) ; and (v) two distinct disjoint doubled edges, the Gaussian pairings. Since W N , i j , S = c N ∫ ℓ N , S ( Ξ i ; w ) ℓ N , S ( Ξ j ; w ) π ( w ) d w W_{N,ij,S}=c_N\int\ell_{N,S}(\Xi_i;w)\ell_{N,S}(\Xi_j;w)\pi(w)\,dw W N , ij , S = c N ∫ ℓ N , S ( Ξ i ; w ) ℓ N , S ( Ξ j ; w ) π ( w ) d w with c N 4 = 16 p 8 h N 2 d X / N 4 c_N^4=16p^8h_N^{2d_X}/N^4 c N 4 = 16 p 8 h N 2 d X / N 4 , the moment bounds (21) with r = 4 r=4 r = 4 , (22), and (23) give
G I = O ( N 2 ) E [ W N , 12 , S 4 ] = O ( N 2 ) c N 4 O ( h N − 3 d X ) = O ( N − 2 h N − d X ) , G I I = O ( N 3 ) E [ W N , 12 , S 2 W N , 13 , S 2 ] = O ( N 3 ) c N 4 O ( h N − 2 d X ) = O ( N − 1 ) , G I I I = O ( N 3 ) ∣ E [ W N , 12 , S 2 W N , 13 , S W N , 23 , S ] ∣ ≤ O ( N 3 ) ( E [ W N , 12 , S 2 W N , 13 , S 2 ] ) 1 / 2 ( E [ W N , 12 , S 2 W N , 23 , S 2 ] ) 1 / 2 ≤ O ( N 3 ) c N 4 O ( h N − 2 d X ) = O ( N − 1 ) , G I V = O ( N 4 ) ∣ E [ W N , 12 , S W N , 23 , S W N , 34 , S W N , 41 , S ] ∣ = O ( N 4 ) c N 4 O ( h N − d X ) = O ( h N d X ) . \begin{aligned}
G_I&=O(N^2)E[W_{N,12,S}^4]=O(N^2)c_N^4O(h_N^{-3d_X})=O(N^{-2}h_N^{-d_X}),\\
G_{II}&=O(N^3)E[W_{N,12,S}^2W_{N,13,S}^2]=O(N^3)c_N^4O(h_N^{-2d_X})=O(N^{-1}),\\
G_{III}&=O(N^3)\left|E[W_{N,12,S}^2W_{N,13,S}W_{N,23,S}]\right|\\
&\le O(N^3)\left(E[W_{N,12,S}^2W_{N,13,S}^2]\right)^{1/2}\left(E[W_{N,12,S}^2W_{N,23,S}^2]\right)^{1/2}\\
&\le O(N^3)c_N^4O(h_N^{-2d_X})=O(N^{-1}),\\
G_{IV}&=O(N^4)\left|E[W_{N,12,S}W_{N,23,S}W_{N,34,S}W_{N,41,S}]\right|=O(N^4)c_N^4O(h_N^{-d_X})=O(h_N^{d_X}).
\end{aligned} G I G I I G I I I G I V = O ( N 2 ) E [ W N , 12 , S 4 ] = O ( N 2 ) c N 4 O ( h N − 3 d X ) = O ( N − 2 h N − d X ) , = O ( N 3 ) E [ W N , 12 , S 2 W N , 13 , S 2 ] = O ( N 3 ) c N 4 O ( h N − 2 d X ) = O ( N − 1 ) , = O ( N 3 ) E [ W N , 12 , S 2 W N , 13 , S W N , 23 , S ] ≤ O ( N 3 ) ( E [ W N , 12 , S 2 W N , 13 , S 2 ] ) 1/2 ( E [ W N , 12 , S 2 W N , 23 , S 2 ] ) 1/2 ≤ O ( N 3 ) c N 4 O ( h N − 2 d X ) = O ( N − 1 ) , = O ( N 4 ) ∣ E [ W N , 12 , S W N , 23 , S W N , 34 , S W N , 41 , S ] ∣ = O ( N 4 ) c N 4 O ( h N − d X ) = O ( h N d X ) .
All four are o ( 1 ) o(1) o ( 1 ) , since h N → 0 h_N\to0 h N → 0 and N h N d X → ∞ Nh_N^{d_X}\to\infty N h N d X → ∞ . For type (v), fix two distinct disjoint pairs { i , j } \{i,j\} { i , j } and { k , l } \{k,l\} { k , l } . There are ( 4 2 ) = 6 \binom{4}{2}=6 ( 2 4 ) = 6 ways to assign them to the positions r = 1 , … , 4 r=1,\ldots,4 r = 1 , … , 4 , and E [ W N , i j , S 2 W N , k l , S 2 ] = E [ W N , i j , S 2 ] E [ W N , k l , S 2 ] E[W_{N,ij,S}^2W_{N,kl,S}^2]=E[W_{N,ij,S}^2]E[W_{N,kl,S}^2] E [ W N , ij , S 2 W N , k l , S 2 ] = E [ W N , ij , S 2 ] E [ W N , k l , S 2 ] because disjoint pairs involve independent vertex variables. Splitting ( ∑ i < j E [ W N , i j , S 2 ] ) 2 \left(\sum_{i<j} E[W^{2}_{N,ij,S}]\right)^{2} ( ∑ i < j E [ W N , ij , S 2 ] ) 2 into its diagonal terms, its disjoint pairs, and its pairs sharing one vertex, the total contribution of type (v) is
6 ∑ { i , j } ∩ { k , l } = ∅ E [ W N , i j , S 2 ] E [ W N , k l , S 2 ] = 3 ( ∑ i < j E [ W N , i j , S 2 ] ) 2 − 3 ∑ i < j ( E [ W N , i j , S 2 ] ) 2 − 6 ∑ ∣ { i , j } ∩ { k , l } ∣ = 1 E [ W N , i j , S 2 ] E [ W N , k l , S 2 ] = 3 s N , S 4 + o ( 1 ) , 6\sum_{\{i,j\}\cap\{k,l\}=\varnothing} E[W^{2}_{N,ij,S}]E[W^{2}_{N,kl,S}]
=3\left(\sum_{i<j}E[W^{2}_{N,ij,S}]\right)^{2}-3\sum_{i<j}\left(E[W^{2}_{N,ij,S}]\right)^{2}-6\sum_{|\{i,j\}\cap\{k,l\}|=1}E[W^{2}_{N,ij,S}]E[W^{2}_{N,kl,S}]
=3s^{4}_{N,S}+o(1), 6 { i , j } ∩ { k , l } = ∅ ∑ E [ W N , ij , S 2 ] E [ W N , k l , S 2 ] = 3 ( i < j ∑ E [ W N , ij , S 2 ] ) 2 − 3 i < j ∑ ( E [ W N , ij , S 2 ] ) 2 − 6 ∣ { i , j } ∩ { k , l } ∣ = 1 ∑ E [ W N , ij , S 2 ] E [ W N , k l , S 2 ] = 3 s N , S 4 + o ( 1 ) ,
where the last equality uses ∑ i < j E [ W N , i j , S 2 ] = s N , S 2 \sum_{i<j}E[W^{2}_{N,ij,S}]=s^{2}_{N,S} ∑ i < j E [ W N , ij , S 2 ] = s N , S 2 and E [ W N , i j , S 2 ] = E [ W N , 12 , S 2 ] = 2 V N , S / N 2 = O ( N − 2 ) E[W^{2}_{N,ij,S}]=E[W^{2}_{N,12,S}]=2\mathcal{V}_{N,S}/N^{2}=O(N^{-2}) E [ W N , ij , S 2 ] = E [ W N , 12 , S 2 ] = 2 V N , S / N 2 = O ( N − 2 ) . The diagonal sum is O ( N 2 ⋅ N − 4 ) = O ( N − 2 ) O(N^{2}\cdot N^{-4})=O(N^{-2}) O ( N 2 ⋅ N − 4 ) = O ( N − 2 ) , and the sum over the O ( N 3 ) O(N^{3}) O ( N 3 ) pairs sharing one vertex is O ( N 3 ⋅ N − 4 ) = O ( N − 1 ) O(N^{3}\cdot N^{-4})=O(N^{-1}) O ( N 3 ⋅ N − 4 ) = O ( N − 1 ) .
Consequently, since s N , S 2 → V S > 0 s^{2}_{N,S}\to\mathcal{V}_{S}>0 s N , S 2 → V S > 0 , E [ W N , S 4 ] / s N , S 4 → 3 E[W^{4}_{N,S}]/s^{4}_{N,S}\to3 E [ W N , S 4 ] / s N , S 4 → 3 . De Jong’s central limit theorem for clean generalized quadratic forms now gives
W N , S → d N ( 0 , V S ) . (24) W_{N,S}\to_{d}N(0,\mathcal{V}_{S}). \tag*{(24)} W N , S → d N ( 0 , V S ) . (24)
Step 7: Conclusion.
By (13) , (19) , B N , S = b N , S + o p ( 1 ) B_{N,S}=b_{N,S}+o_{p}(1) B N , S = b N , S + o p ( 1 ) , and (24) , T N , S = B N , S + W N , S + o p ( 1 ) = b N , S + W N , S + o p ( 1 ) T_{N,S}=B_{N,S}+W_{N,S}+o_{p}(1)=b_{N,S}+W_{N,S}+o_{p}(1) T N , S = B N , S + W N , S + o p ( 1 ) = b N , S + W N , S + o p ( 1 ) ; hence T N , S − b N , S → d N ( 0 , V S ) T_{N,S}-b_{N,S}\to_{d}N(0,\mathcal{V}_{S}) T N , S − b N , S → d N ( 0 , V S ) .
Conclusion
This paper develops a specification test for parametric conditionally independent dyad models of network formation. The test is based on the restrictions that a parametric model places on subgraphs. Under the null hypothesis and a nondegeneracy condition, the centered test statistic is asymptotically normal, with an explicit centering constant and variance. The asymptotic theory accounts for the dependence among overlapping tuples of agents in a single network, and the normalization of the statistic depends only on the dimension of a single agent’s characteristics, which alleviates the curse of dimensionality.