Introduction
Let { G n } n ∈ N \{G_n\}_{n\in\mathbb{N}} { G n } n ∈ N be a sequence of simple, undirected graphs with vertex set [ n ] : = { 1 , … , n } [n]:=\{1,\ldots,n\} [ n ] := { 1 , … , n } and edge set E ( G n ) E(G_n) E ( G n ) . Suppose the vertices of G n G_n G n are independently colored red with probability p ∈ ( 0 , 1 ) p\in(0,1) p ∈ ( 0 , 1 ) and blue with probability 1 − p 1-p 1 − p . Let T n T_n T n be the number of monochromatic edges in E ( G n ) E(G_n) E ( G n ) and U n U_n U n be its standardized version. Despite being elementary the statistic U n U_n U n exhibits a rich range of limiting behaviors and arises in a wide range of problems across combinatorial probability and statistical inference. In the balanced (uniform) case p = 1 2 p=\frac{1}{2} p = 2 1 , it can be expressed as a second-order Rademacher chaos, that is, a quadratic form in independent Rademacher variables (see (11) ). This is a prototypical example of a degenerate U U U -statistic of order two [17 , 30 , 35 ] , which arises in many contexts, such as the Hamiltonian of Ising models [2 , 11 ] , tests of independence in autoregressive models [3 ] , and Ramsey theory [1 , 34 ] . The unbalanced (biased) case p ≠ 1 2 p \ne\frac{1}{2} p = 2 1 also arises naturally in nonparametric two-sample testing based on geometric graphs (see § 1.3 for details) and in motif estimation problems [7 , 33 ] .
Since the statistic U n U_n U n is a sum of locally dependent indicators (see (2) ), sufficient conditions for its asymptotic normality can be obtained using standard versions of Stein’s method or other normal approximation techniques. Identifying necessary and sufficient conditions, however, is considerably more delicate. For graph sequences with ∣ E ( G n ) ∣ → ∞ |E(G_n)| \to\infty ∣ E ( G n ) ∣ → ∞ , asymptotic normality in the balanced case p = 1 2 p = \frac{1}{2} p = 2 1 is characterized by a fourth moment phenomenon [6 , 9 ] :
U n → D N ( 0 , 1 ) if and only if E [ U n 4 ] → 3. U_n \xrightarrow{D} N(0,1) \quad\text{if and only if} \quad\mathbb{E}\left[U_n^4\right] \to3. U n D N ( 0 , 1 ) if and only if E [ U n 4 ] → 3.
The unbalanced case, despite its practical relevance, has received considerably less attention.
In this paper, we develop a general framework for characterizing the asymptotic normality of U n U_n U n for every fixed p ∈ ( 0 , 1 ) p \in(0,1) p ∈ ( 0 , 1 ) and providing sufficient conditions in the regime when p = p n → 0 p = p_n \to0 p = p n → 0 . Our results reveal a new third moment phenomenon : for every fixed p ≠ 1 2 p \ne\frac{1}{2} p = 2 1 , asymptotic normality is equivalent to E [ U n 3 ] → 0 \mathbb{E}\left[U_n^3\right] \to0 E [ U n 3 ] → 0 . A particularly striking feature is the emergence of the two exceptional probabilities p = 1 ± 1 / 3 2 p = \frac{1 \pm1/\sqrt{3}}{2} p = 2 1 ± 1/ 3 . At these values, the fourth moment converges to its Gaussian value even when asymptotic normality fails (see Example 6.1), whereas the third moment continues to characterize Gaussian convergence. Outside these exceptional values, the fourth moment condition also characterizes asymptotic normality (see Theorem 1.1 ). At p = 1 2 p = \frac{1}{2} p = 2 1 , (see Example 6.2), we provide an example to show that the third moment condition is insufficient. In the sparse regime, (see Proposition 1.3, Theorem 1.2 , Example 6.3 and Example 6.4 ), where p = p n → 0 p = p_n \to0 p = p n → 0 , as n → ∞ n \to\infty n → ∞ we show that the third moment condition is sufficient for asymptotic normality of U n U_n U n . Interestingly, the third moment condition is strictly weaker than the fourth moment condition in this regime, although neither is necessary for asymptotic normality. In addition for the fixed p p p regime we provide necessary and sufficient conditions in terms of motif counts for the third and fourth moment phenomena to hold (see Theorem 1.1 ). We establish a similar result for the sparse regime, prescribing a sufficient decay rate of p n p_n p n in terms of motif counts (see Lemma 1.4 ).
Next, using the above results we derive the asymptotic null distribution of two-sample tests based on undirected K K K -nearest-neighbor (K K K -NN) graphs in the Poissonized setting. Allowing K K K to grow with the sample size has been often recommended to improve power for graph-based nonparametric tests [4 , 13 ] . Our results provide sharp general growth conditions that justify normal calibration in this regime. Specifically, we show that under the null hypothesis, the test statistic is asymptotically normal for arbitrary K K K when the limiting sample proportions are unequal, and for K = o ( n ) K = o(n) K = o ( n ) when the sample sizes are equal (see Theorem 1.6 for precise statement). The same framework also gives conditions for asymptotic normality of two-sample edge-count statistics based on general similarity graphs.
The rest of the introduction is organized as follows. In § 1.1 we define the model and statistics precisely and state our main results on the third and fourth moment phenomena for monochromatic edges. We then show asymptotic properties of graph based two-sample tests in § 1.3 . We conclude in § 1.4 by discussing the novelty and significance of our main results in relation to the existing literature and outlining the proofs of these results.
Third and Fourth Moment Phenomena for Monochromatic Edges
Let { G n } n ∈ N \{G_n\}_{n\in\mathbb{N}} { G n } n ∈ N be a sequence of simple, undirected graphs with vertex set [ n ] : = { 1 , … , n } [n] := \{1,\ldots,n\} [ n ] := { 1 , … , n } and edge set E ( G n ) E(G_n) E ( G n ) . Let d G n ( i ) d_{G_n}(i) d G n ( i ) will be the degree of the i i i -th vertex of G n G_n G n , for 1 ≤ i ≤ n 1 \le i \le n 1 ≤ i ≤ n . Suppose the vertices of G n G_n G n are independently colored red with probability p ∈ ( 0 , 1 ) p \in(0,1) p ∈ ( 0 , 1 ) and blue with probability 1 − p 1-p 1 − p . For 1 ≤ i ≤ n 1 \le i \le n 1 ≤ i ≤ n , denote by X i X_i X i the color of the vertex i i i , with X i = 1 X_i = 1 X i = 1 for red and X i = 0 X_i = 0 X i = 0 for blue. Our object of interest in this paper is the number of monochromatic edges in G n G_n G n , that is,
T n : = ∑ { i , j } ∈ E ( G n ) 1 { X i = X j } . (1) T_n := \sum_{\{i,j\}\in E(G_n)} \mathbf{1}\{X_i = X_j\}. \tag*{(1)} T n := { i , j } ∈ E ( G n ) ∑ 1 { X i = X j } . (1)
Denote the centered and rescaled version of T n T_n T n by:
U n : = T n − E [ T n ] Var [ T n ] . (2) U_n := \frac{T_n-\mathbb{E}[T_n]}{\sqrt{\operatorname{Var}[T_n]}}. \tag*{(2)} U n := Var [ T n ] T n − E [ T n ] . (2)
Throughout, for a finite graph H H H (a motif), we will denote by N ( H , G n ) N(H,G_n) N ( H , G n ) the number of copies of H H H in G n G_n G n . We will use pictograms for the motifs, so for instance N ( edge motif , G n ) N(\text{edge motif},G_n) N ( edge motif , G n ) will denote the number of edges in G n G_n G n and N ( 4-cycle motif , G n ) N(\text{4-cycle motif},G_n) N ( 4-cycle motif , G n ) will denote the number of 4-cycles on G n G_n G n . Recall that X i X_i X i is the color of the vertex i i i , for 1 ≤ i ≤ n 1 \le i \le n 1 ≤ i ≤ n . Hence,
E [ 1 { X i = X j } ] = p 2 + ( 1 − p ) 2 = 1 2 ( 1 + u ) , \mathbb{E}[\mathbf{1}\{X_i=X_j\}]=p^2+(1-p)^2=\frac{1}{2}(1+u), E [ 1 { X i = X j }] = p 2 + ( 1 − p ) 2 = 2 1 ( 1 + u ) ,
where u : = ( 2 p − 1 ) 2 ∈ [ 0 , 1 ) u:=(2p-1)^2\in[0,1) u := ( 2 p − 1 ) 2 ∈ [ 0 , 1 ) . Then, from (1) ,
E [ T n ] = 1 2 ( 1 + u ) N ( edge motif , G n ) . \mathbb{E}[T_n]=\frac{1}{2}(1+u)N(\text{edge motif},G_n). E [ T n ] = 2 1 ( 1 + u ) N ( edge motif , G n ) .
Further, since the indicators of a pair of edges being monochromatic are independent when the edges are disjoint, a direct computation gives,
Var [ T n ] = 1 − u 2 4 N ( edge motif , G n ) + u − u 2 2 N ( wedge motif , G n ) . (3) \operatorname{Var}[T_n]=\frac{1-u^2}{4}N(\text{edge motif},G_n)+\frac{u-u^2}{2}N(\text{wedge motif},G_n). \tag*{(3)} Var [ T n ] = 4 1 − u 2 N ( edge motif , G n ) + 2 u − u 2 N ( wedge motif , G n ) . (3)
Note that the second term in (3) vanishes when u = 0 u=0 u = 0 , that is, p = 1 2 p=\frac{1}{2} p = 2 1 . Consequently, the leading order of the variance changes depending on whether or not p = 1 2 p=\frac{1}{2} p = 2 1 . This dichotomy is reflected throughout our results, with the balanced case often exhibiting different asymptotic behavior.
For two positive sequences { a n } n ∈ N \{a_n\}_{n\in\mathbb{N}} { a n } n ∈ N and { b n } n ∈ N \{b_n\}_{n\in\mathbb{N}} { b n } n ∈ N we write a n = o ( b n ) a_n=o(b_n) a n = o ( b n ) if a n / b n → 0 a_n/b_n\to0 a n / b n → 0 , a n = O ( b n ) a_n=O(b_n) a n = O ( b n ) if a n / b n a_n/b_n a n / b n is bounded, and a n ≍ b n a_n\asymp b_n a n ≍ b n if both a n = O ( b n ) a_n=O(b_n) a n = O ( b n ) and b n = O ( a n ) b_n=O(a_n) b n = O ( a n ) . All asymptotics are taken as n → ∞ n\to\infty n → ∞ . We separate our results on the asymptotic normality of U n U_n U n in two regimes: the fixed p p p regime, in which p ∈ ( 0 , 1 ) p\in(0,1) p ∈ ( 0 , 1 ) is a fixed constant, and the sparse regime, in which p = p n → 0 p=p_n\to0 p = p n → 0 , as n → ∞ n\to\infty n → ∞ .
The Fixed p p p Regime: The following theorem characterizes the asymptotic normality of U n U_n U n in terms of its third and fourth moments and the growth rates of certain motif counts, when the coloring probability p ∈ ( 0 , 1 ) p\in(0,1) p ∈ ( 0 , 1 ) is fixed..
Theorem 1.1. Let { G n } n ∈ N \{G_n\}_{n\in\mathbb{N}} { G n } n ∈ N be a sequence of graphs such that N ( edge motif , G n ) → ∞ N(\text{edge motif},G_n)\to\infty N ( edge motif , G n ) → ∞ and U n U_n U n be as defined in (2) . Suppose p ∈ ( 0 , 1 ) p\in(0,1) p ∈ ( 0 , 1 ) is fixed. Then the following hold:
(i) If p ∈ ( 0 , 1 ) ∖ { 1 2 , 1 ± 1 / 3 2 } p\in(0,1)\setminus\left\{\frac{1}{2},\frac{1\pm1/\sqrt{3}}{2}\right\} p ∈ ( 0 , 1 ) ∖ { 2 1 , 2 1 ± 1/ 3 } , then the following are equivalent:
(a) U n → D N ( 0 , 1 ) U_n\xrightarrow{D}N(0,1) U n D N ( 0 , 1 ) .
(b) N ( 4-cycle motif , G n ) = o ( ( N ( edge motif , G n ) + N ( wedge motif , G n ) ) 2 ) N(\text{4-cycle motif},G_n)=o((N(\text{edge motif},G_n)+N(\text{wedge motif},G_n))^2) N ( 4-cycle motif , G n ) = o (( N ( edge motif , G n ) + N ( wedge motif , G n ) ) 2 ) .
(c) E [ U n 4 ] → 3 \mathbb{E}[U_n^4]\to3 E [ U n 4 ] → 3 .
(d) E [ U n 3 ] → 0 \mathbb{E}[U_n^3]\to0 E [ U n 3 ] → 0 .
(ii) If p = 1 ± 1 / 3 2 p=\frac{1\pm1/\sqrt{3}}{2} p = 2 1 ± 1/ 3 , then the following are equivalent:
(a) U n → D N ( 0 , 1 ) U_n\xrightarrow{D}N(0,1) U n D N ( 0 , 1 ) .
(b) N ( 4-cycle motif , G n ) = o ( ( N ( edge motif , G n ) + N ( wedge motif , G n ) ) 2 ) N(\text{4-cycle motif},G_n)=o((N(\text{edge motif},G_n)+N(\text{wedge motif},G_n))^2) N ( 4-cycle motif , G n ) = o (( N ( edge motif , G n ) + N ( wedge motif , G n ) ) 2 ) .
(c) E [ U n 3 ] → 0 \mathbb{E}[U_n^3]\to0 E [ U n 3 ] → 0 .
(iii) If p = 1 2 p=\frac{1}{2} p = 2 1 , then the following are equivalent:
(a) U n → D N ( 0 , 1 ) U_n \xrightarrow{D} N(0,1) U n D N ( 0 , 1 ) .
(b) N ( ●—● , G n ) = o ( N ( ●—● , G n ) 2 ) N(\text{●—●},G_n)=o(N(\text{●—●},G_n)^2) N ( ●—● , G n ) = o ( N ( ●—● , G n ) 2 ) .
(c) E [ U n 4 ] → 3 \mathbb{E}[U_n^4]\to3 E [ U n 4 ] → 3 .
The theorem identifies three distinct regimes for the asymptotic normality of U n U_n U n , distinguished by the exceptional probabilities p = 1 ± 1 / 3 2 p=\frac{1\pm1/\sqrt{3}}{2} p = 2 1 ± 1/ 3 and the balanced case p = 1 2 p=\frac{1}{2} p = 2 1 . Detailed remarks on the three regimes, on the novelty of the third and fourth moment phenomena and an outline of the proofs are given in § 1.4 . As remarked earlier for the two exceptional probabilities p = 1 ± 1 / 3 2 p=\frac{1\pm1/\sqrt{3}}{2} p = 2 1 ± 1/ 3 in Example 6.1, we construct a graph sequence, (n n n -star K 1 , n K_{1,n} K 1 , n ), for which the fourth moment converges to its Gaussian value even when asymptotic normality fails and for p = 1 2 p=\frac{1}{2} p = 2 1 , in Example 6.2 we consider a graph sequence, (the complete bipartite K 2 , n K_{2,n} K 2 , n ), for which the third moment condition is insufficient. The proof of Theorem 1.1 is given in Section 4.
Asymptotic Properties of Graph Based Two-Sample Tests
The Sparse p p p Regime: Next, we consider the sparse regime p n → 0 p_n\to0 p n → 0 . In this regime the third moment condition turns out to be sufficient for the asymptotic normality of U n U_n U n .
Theorem 1.2. Let { G n } n ∈ N \{G_n\}_{n\in\mathbb{N}} { G n } n ∈ N be a sequence of graphs such that N ( ●—● , G n ) → ∞ N(\text{●—●},G_n)\to\infty N ( ●—● , G n ) → ∞ and U n U_n U n be as defined in (1.2). Suppose p = p n → 0 p=p_n\to0 p = p n → 0 , as n → ∞ n\to\infty n → ∞ . Then
E [ U n 3 ] → 0 ⟹ U n → D N ( 0 , 1 ) . \mathbb{E}[U_n^3]\to0 \Longrightarrow U_n\xrightarrow{D}N(0,1). E [ U n 3 ] → 0 ⟹ U n D N ( 0 , 1 ) .
The theorem shows that in the sparse regime, the third moment condition is sufficient for asymptotic normality of U n U_n U n . Our next result states that the fourth moment condition always implies the third moment condition in this regime.
Proposition 1.3. Let { G n } n ∈ N \{G_n\}_{n\in\mathbb{N}} { G n } n ∈ N be a sequence of graphs such that N ( ●—● , G n ) → ∞ N(\text{●—●},G_n)\to\infty N ( ●—● , G n ) → ∞ , and let U n U_n U n be as defined in (1.2). Suppose p = p n → 0 p=p_n\to0 p = p n → 0 as n → ∞ n\to\infty n → ∞ . Then
E [ U n 4 ] → 3 ⟹ E [ U n 3 ] → 0. \mathbb{E}[U_n^4]\to3 \Longrightarrow\mathbb{E}[U_n^3]\to0. E [ U n 4 ] → 3 ⟹ E [ U n 3 ] → 0.
We conclude by stating a lemma that characterizes the fourth moment condition in the sparse regime in terms of motif counts.
Lemma 1.4. Let { G n } n ∈ N \{G_n\}_{n\in\mathbb{N}} { G n } n ∈ N be a sequence of graphs such that N ( ●—● , G n ) → ∞ N(\text{●—●},G_n)\to\infty N ( ●—● , G n ) → ∞ , and let U n U_n U n be as defined in (1.2). Suppose p = p n → 0 p=p_n\to0 p = p n → 0 as n → ∞ n\to\infty n → ∞ . Then the following are equivalent:
(a) E [ U n 4 ] → 3 \mathbb{E}[U_n^4]\to3 E [ U n 4 ] → 3 .
(b) p n ( N ( ●—● , G n ) + N ( ●⋎● , G n ) ) → ∞ p_n(N(\text{●—●},G_n)+N(\text{●⋎●},G_n))\to\infty p n ( N ( ●—● , G n ) + N ( ●⋎● , G n )) → ∞ and N ( ●◆● , G n ) = o ( p n ( N ( ●—● , G n ) + N ( ●⋎● , G n ) ) 2 ) N(\text{●◆●},G_n)=o(p_n(N(\text{●—●},G_n)+N(\text{●⋎●},G_n))^2) N ( ●◆● , G n ) = o ( p n ( N ( ●—● , G n ) + N ( ●⋎● , G n ) ) 2 ) .
As remarked earlier, we show in Example 6.3 that there are graph sequences, (consisting of an n 2 n^2 n 2 -star and n 5 n^5 n 5 disjoint edges, p n = 1 n 3 p_n=\frac{1}{n^3} p n = n 3 1 ), for which U n U_n U n is asymptotically normal, even though its third moment does not tend to zero. In Example 6.4 we provide a graph sequence, (an n n n -star and n 3 n^3 n 3 disjoint edges, p n = 1 n 2 p_n=\frac{1}{n^2} p n = n 2 1 ), for which the third moment condition holds, and hence U n U_n U n is asymptotically normal, while the fourth moment condition fails. Hence the implication in Proposition 1.3 is strict and that the third moment condition is strictly weaker in this regime. The proof of Theorem 1.2 is given in Section 4 while Proposition 1.3 and Lemma 1.4 are shown in § 2 .
Asymptotic Properties of Graph Based Two-Sample Tests
In this section, we will derive sharp conditions for the asymptotic normality and consistency of two-sample tests based on nearest-neighbor graphs. We begin by formally describing the two-sample problem: Given two probability measures P P P and Q Q Q in R d \mathbb{R}^{d} R d , the two-sample problem is to test the hypotheses:
H 0 : P = Q against H 1 : P ≠ Q , H_{0}: P = Q \qquad\text{against} \qquad H_{1}: P \ne Q, H 0 : P = Q against H 1 : P = Q ,
based on independent samples from the respective distributions. Throughout, we will assume P P P and Q Q Q have densities f f f and g g g with respect to the Lebesgue measure on R d \mathbb{R}^{d} R d .
Definition 1.5. Fix K ≥ 1 K \ge1 K ≥ 1 . Given a finite set S ⊂ R d S \subset\mathbb{R}^{d} S ⊂ R d with pairwise distinct distances, the undirected K K K -nearest neighbor graph (K K K -NN) is the graph with vertex set S S S and an edge { a , b } \{a,b\} { a , b } between distinct points a , b ∈ S a,b \in S a , b ∈ S if b b b is one of the K K K nearest neighbors of a a a in S ∖ { a } S\setminus\{a\} S ∖ { a } or a a a is one of the K K K nearest neighbors of b b b in S ∖ { b } S\setminus\{b\} S ∖ { b } . If ∣ S ∣ ≤ K |S| \le K ∣ S ∣ ≤ K , we adopt the convention that the graph is complete. Denote the undirected K K K -NN graph of S S S by G K ( S ) G_K(S) G K ( S ) .
Here, we will consider the two-sample problem in the Poissonized setting, where instead of taking a fixed number n 1 n_1 n 1 and n 2 n_2 n 2 samples from the two distributions, we have Poi ( n 1 ) \operatorname{Poi}(n_1) Poi ( n 1 ) from f f f and Poi ( n 2 ) \operatorname{Poi}(n_2) Poi ( n 2 ) samples from g g g . Specifically, suppose X = { X 1 , X 2 , … } \mathcal{X} = \{X_1,X_2,\ldots\} X = { X 1 , X 2 , … } and Y = { Y 1 , Y 2 , … } \mathcal{Y} = \{Y_1,Y_2,\ldots\} Y = { Y 1 , Y 2 , … } are independent collections of i.i.d. samples from f f f and g g g , respectively. Define
X n 1 = { X 1 , X 2 , … , X L n 1 } and Y n 2 = { Y 1 , Y 2 , … , Y L n 2 } , (4) \mathcal{X}_{n_1} = \{X_1,X_2,\ldots,X_{L_{n_1}}\}
\quad\text{and}\quad
\mathcal{Y}_{n_2} = \{Y_1,Y_2,\ldots,Y_{L_{n_2}}\}, \tag*{(4)} X n 1 = { X 1 , X 2 , … , X L n 1 } and Y n 2 = { Y 1 , Y 2 , … , Y L n 2 } , (4)
where L n 1 ∼ Poi ( n 1 ) L_{n_1} \sim\operatorname{Poi}(n_1) L n 1 ∼ Poi ( n 1 ) and L n 2 ∼ Poi ( n 2 ) L_{n_2} \sim\operatorname{Poi}(n_2) L n 2 ∼ Poi ( n 2 ) are independent of each other, and of X \mathcal{X} X and Y \mathcal{Y} Y . Poissonization is a common assumption in this literature [5 , 29 , 31 ] , which simplifies calculations, due to the spatial independence of the Poisson process, and yields cleaner formulas for the asymptotic variances. Then the Poissonized two-sample statistic based on the K K K -NN graph is defined as follows:1
T × ( G K ( X n 1 ∪ Y n 2 ) ) : = ∑ i = 1 L n 1 ∑ j = 1 L n 2 1 { { X i , Y j } ∈ E ( G K ( X n 1 ∪ Y n 2 ) ) } . (5) T^{\times}\left(G_K\left(\mathcal{X}_{n_1} \cup\mathcal{Y}_{n_2}\right)\right)
:= \sum_{i=1}^{L_{n_1}}\sum_{j=1}^{L_{n_2}}
\mathbf{1}\left\{\{X_i,Y_j\} \in E\left(G_K\left(\mathcal{X}_{n_1} \cup\mathcal{Y}_{n_2}\right)\right)\right\}. \tag*{(5)} T × ( G K ( X n 1 ∪ Y n 2 ) ) := i = 1 ∑ L n 1 j = 1 ∑ L n 2 1 { { X i , Y j } ∈ E ( G K ( X n 1 ∪ Y n 2 ) ) } . (5)
The distribution of this statistic in (5) can be equivalently described as follows: Let ϕ n ( x ) : = n 1 n f ( x ) + n 2 n g ( x ) \phi_n(x) := \frac{n_1}{n}f(x) + \frac{n_2}{n}g(x) ϕ n ( x ) := n n 1 f ( x ) + n n 2 g ( x ) and Z 1 , Z 2 , … Z_1,Z_2,\ldots Z 1 , Z 2 , … be independent random variables with common density ϕ n ( ⋅ ) \phi_n(\cdot) ϕ n ( ⋅ ) and L n L_n L n be an independent Poisson variable with mean n = n 1 + n 2 n = n_1+n_2 n = n 1 + n 2 . Then
Z n = { Z 1 , Z 2 , … , Z L n } , (6) \mathcal{Z}_n = \{Z_1,Z_2,\ldots,Z_{L_n}\}, \tag*{(6)} Z n = { Z 1 , Z 2 , … , Z L n } , (6)
is a nonhomogeneous Poisson process in R d \mathbb{R}^{d} R d with rate function n 1 f + n 2 g n_1f+n_2g n 1 f + n 2 g . Label each point z ∈ Z n z \in\mathcal{Z}_n z ∈ Z n independently with
c z = { 1 with probability n 1 f ( z ) n 1 f ( z ) + n 2 g ( z ) , 2 with probability n 2 g ( z ) n 1 f ( z ) + n 2 g ( z ) . c_z =
\begin{cases}
1 & \text{with probability } \dfrac{n_1f(z)}{n_1f(z)+n_2g(z)},\\
2 & \text{with probability } \dfrac{n_2g(z)}{n_1f(z)+n_2g(z)}.
\end{cases} c z = ⎩ ⎨ ⎧ 1 2 with probability n 1 f ( z ) + n 2 g ( z ) n 1 f ( z ) , with probability n 1 f ( z ) + n 2 g ( z ) n 2 g ( z ) .
Then the sets of points assigned labels 1 and 2 have the same distribution as X n 1 \mathcal{X}_{n_1} X n 1 and Y n 2 \mathcal{Y}_{n_2} Y n 2 (as in (4) ), respectively. Therefore, (5) is equal in distribution to
T × ( G K ( Z n ) ) = D ∑ { x , y } ⊆ Z n 1 { c x ≠ c y } 1 { { x , y } ∈ E ( G K ( Z n ) ) } = ∣ E ( G K ( Z n ) ) ∣ − T ∘ ( G K ( Z n ) ) , (7) \begin{aligned}
T^{\times}\left(G_K(\mathcal{Z}_n)\right)
\mathrel{\stackrel{\mathcal{D}}{=}}
\sum_{\{x,y\}\subseteq\mathcal{Z}_n}
\mathbf{1}\{c_x \ne c_y\}\mathbf{1}\left\{\{x,y\}\in E(G_K(\mathcal{Z}_n))\right\} \\
&= \left|E(G_K(\mathcal{Z}_n))\right| - T^{\circ}(G_K(\mathcal{Z}_n)), \tag*{(7)}
\end{aligned} T × ( G K ( Z n ) ) = D { x , y } ⊆ Z n ∑ 1 { c x = c y } 1 { { x , y } ∈ E ( G K ( Z n )) } = ∣ E ( G K ( Z n )) ∣ − T ∘ ( G K ( Z n )) , (7)
where the sum is over unordered pairs { x , y } \{x,y\} { x , y } , with x , y ∈ Z n x,y \in\mathcal{Z}_n x , y ∈ Z n , and ∣ E ( G K ( Z n ) ) ∣ = N ( ∙ , G K ( Z n ) ) |E(\mathcal{G}_K(\mathcal{Z}_n))| = N(\bullet,\mathcal{G}_K(\mathcal{Z}_n)) ∣ E ( G K ( Z n )) ∣ = N ( ∙ , G K ( Z n )) is the number of edges in G K ( Z n ) \mathcal{G}_K(\mathcal{Z}_n) G K ( Z n ) and
T ∘ ( G K ( Z n ) ) = ∑ { x , y } ∈ Z n 1 { c x = c y } 1 { { x , y } ∈ E ( G K ( Z n ) ) } , (8) T^\circ(\mathcal{G}_K(\mathcal{Z}_n)) = \sum_{\{x,y\}\in\mathcal{Z}_n} 1_{\{c_x=c_y\}}1_{\{\{x,y\}\in E(\mathcal{G}_K(\mathcal{Z}_n))\}}, \tag*{(8)} T ∘ ( G K ( Z n )) = { x , y } ∈ Z n ∑ 1 { c x = c y } 1 {{ x , y } ∈ E ( G K ( Z n ))} , (8)
is the number of monochromatic edges in G K ( Z n ) \mathcal{G}_K(\mathcal{Z}_n) G K ( Z n ) . Let F = σ ( Z n ) \mathcal{F}=\sigma(\mathcal{Z}_n) F = σ ( Z n ) be the sigma-algebra generated by L n L_n L n and the pooled sample in (6) . The edge count ∣ E ( G K ( Z n ) ) ∣ |E(\mathcal{G}_K(\mathcal{Z}_n))| ∣ E ( G K ( Z n )) ∣ is F \mathcal{F} F -measurable, since it is determined by the pooled sample without reference to the markings. Consequently, conditional on F \mathcal{F} F , tests based on the statistics T × ( G K ( Z n ) ) T^\times(\mathcal{G}_K(\mathcal{Z}_n)) T × ( G K ( Z n )) and T ∘ ( G K ( Z n ) ) T^\circ(\mathcal{G}_K(\mathcal{Z}_n)) T ∘ ( G K ( Z n )) are equivalent: rejecting for small values of T × ( G K ( Z n ) ) T^\times(\mathcal{G}_K(\mathcal{Z}_n)) T × ( G K ( Z n )) corresponds to rejecting for large values of T ∘ ( G K ( Z n ) ) T^\circ(\mathcal{G}_K(\mathcal{Z}_n)) T ∘ ( G K ( Z n )) .
Under the null hypothesis H 0 H_0 H 0 , let f f f denote the common density. Then Z n = { Z 1 , … , Z L n } \mathcal{Z}_n=\{Z_1,\ldots,Z_{L_n}\} Z n = { Z 1 , … , Z L n } is a Poisson process on R d \mathbb{R}^d R d with intensity function n f nf n f . Independently of the point locations, each point receives label 1 with probability λ n : = n 1 n \lambda_n:=\frac{n_1}{n} λ n := n n 1 and label 2 with probability 1 − λ n = n 2 n 1-\lambda_n=\frac{n_2}{n} 1 − λ n = n n 2 . By independent thinning, X n 1 \mathcal{X}_{n_1} X n 1 and Y n 2 \mathcal{Y}_{n_2} Y n 2 are independent Poisson processes with intensity functions n 1 f n_1f n 1 f and n 2 f n_2f n 2 f , respectively. Thus, conditional on Z n \mathcal{Z}_n Z n , G K ( Z n ) \mathcal{G}_K(\mathcal{Z}_n) G K ( Z n ) is a fixed graph with an i.i.d. Ber ( λ n ) \operatorname{Ber}(\lambda_n) Ber ( λ n ) vertex coloring, reducing the problem to the setting considered in the previous section. Specifically, conditional on F \mathcal{F} F , the expectation and variance of T × ( G K ( Z n ) ) T^\times(\mathcal{G}_K(\mathcal{Z}_n)) T × ( G K ( Z n )) under H 0 H_0 H 0 are given by:
μ F × : = E [ T × ( G K ( Z n ) ) ∣ F ] = 2 λ n ( 1 − λ n ) N ( ∙ , G K ( Z n ) ) \mu_{\mathcal{F}}^\times:= \mathbb{E}\left[T^\times(\mathcal{G}_K(\mathcal{Z}_n))\mid\mathcal{F}\right] = 2\lambda_n(1-\lambda_n)N(\bullet,\mathcal{G}_K(\mathcal{Z}_n)) μ F × := E [ T × ( G K ( Z n )) ∣ F ] = 2 λ n ( 1 − λ n ) N ( ∙ , G K ( Z n ))
and, recalling (3) ,
σ F 2 : = Var [ T × ( G K ( Z n ) ) ∣ F ] = Var [ T ∘ ( G K ( Z n ) ) ∣ F ] = 1 − ν n 2 4 N ( ∙ , G K ( Z n ) ) + ν n − ν n 2 2 N ( ∨ , G K ( Z n ) ) , (9) \sigma_{\mathcal{F}}^2 := \operatorname{Var}\left[T^\times(\mathcal{G}_K(\mathcal{Z}_n))\mid\mathcal{F}\right] = \operatorname{Var}\left[T^\circ(\mathcal{G}_K(\mathcal{Z}_n))\mid\mathcal{F}\right]
= \frac{1-\nu_n^2}{4}N(\bullet,\mathcal{G}_K(\mathcal{Z}_n))+\frac{\nu_n-\nu_n^2}{2}N(\vee,\mathcal{G}_K(\mathcal{Z}_n)), \tag*{(9)} σ F 2 := Var [ T × ( G K ( Z n )) ∣ F ] = Var [ T ∘ ( G K ( Z n )) ∣ F ] = 4 1 − ν n 2 N ( ∙ , G K ( Z n )) + 2 ν n − ν n 2 N ( ∨ , G K ( Z n )) , (9)
where ν n : = ( 2 λ n − 1 ) 2 \nu_n:=(2\lambda_n-1)^2 ν n := ( 2 λ n − 1 ) 2 . Define
U n × ( G K ( Z n ) ) : = T × ( G K ( Z n ) ) − μ F × σ F , (10) U_n^\times(\mathcal{G}_K(\mathcal{Z}_n)):=\frac{T^\times(\mathcal{G}_K(\mathcal{Z}_n))-\mu_{\mathcal{F}}^\times}{\sigma_{\mathcal{F}}}, \tag*{(10)} U n × ( G K ( Z n )) := σ F T × ( G K ( Z n )) − μ F × , (10)
where we put U n × ( G K ( Z n ) ) : = 0 U_n^\times(\mathcal{G}_K(\mathcal{Z}_n)):=0 U n × ( G K ( Z n )) := 0 on the event { σ F 2 = 0 } = { L n ≤ 1 } \{\sigma_{\mathcal{F}}^2=0\}=\{L_n\le1\} { σ F 2 = 0 } = { L n ≤ 1 } , whose probability tends to zero. The following result establishes the asymptotic normality of U n × ( G K ( Z n ) ) U_n^\times(\mathcal{G}_K(\mathcal{Z}_n)) U n × ( G K ( Z n )) under H 0 H_0 H 0 , with convergence in the Wasserstein distance.
Theorem 1.6. Let U n × ( G K ( Z n ) ) U_n^\times(\mathcal{G}_K(\mathcal{Z}_n)) U n × ( G K ( Z n )) be as defined above in (10) . Then the following hold under H 0 H_0 H 0 :
(i) Suppose λ n = n 1 n → λ ∈ ( 0 , 1 ) ∖ { 1 2 } \lambda_n=\frac{n_1}{n}\to\lambda\in(0,1)\setminus\{\frac{1}{2}\} λ n = n n 1 → λ ∈ ( 0 , 1 ) ∖ { 2 1 } . Then Wass ( U n × ( G K ( Z n ) ) , N ( 0 , 1 ) ) → 0 \operatorname{Wass}(U_n^\times(\mathcal{G}_K(\mathcal{Z}_n)),N(0,1))\to0 Wass ( U n × ( G K ( Z n )) , N ( 0 , 1 )) → 0 , for any sequence K n ≥ 1 K_n\ge1 K n ≥ 1 .
(ii) Suppose n 1 = n 2 n_1=n_2 n 1 = n 2 . Then Wass ( U n × ( G K ( Z n ) ) , N ( 0 , 1 ) ) → 0 \operatorname{Wass}(U_n^\times(\mathcal{G}_K(\mathcal{Z}_n)),N(0,1))\to0 Wass ( U n × ( G K ( Z n )) , N ( 0 , 1 )) → 0 , for any K ≥ 1 K\ge1 K ≥ 1 fixed or any sequence K n → ∞ K_n\to\infty K n → ∞ with K n = o ( n ) K_n=o(n) K n = o ( n ) .
The result distinguishes two regimes. For unbalanced sample sizes, in (i), with n 1 n → λ ∈ ( 0 , 1 ) ∖ { 1 2 } \frac{n_1}{n}\to\lambda\in(0,1)\setminus\{\frac{1}{2}\} n n 1 → λ ∈ ( 0 , 1 ) ∖ { 2 1 } , the standardized null test statistic converges to N ( 0 , 1 ) N(0,1) N ( 0 , 1 ) in Wasserstein distance for every sequence K n ≥ 1 K_n\ge1 K n ≥ 1 , including fixed K K K and sequences diverging at any rate (see (62) for an explicit convergence rate). For equal sample sizes, in (ii), the same conclusion holds for every fixed K K K and every growing sequence K n = o ( n ) K_n=o(n) K n = o ( n ) . Together, these conclusions give sharp growth conditions as general guarantees of asymptotic normality. In particular, the sublinear restriction in the balanced case cannot generally be removed. For instance, in the balanced coloring of the complete graph the centered statistic is a degenerate U U U -statistic of order two and has a non-Gaussian limit given by a scaled, centered chi-square variable (see [6 Example 7.1] ). The proof of Theorem 1.6 is given in Section 5.
Although we have so far focused on K K K -NN graphs for illustration, the preceding approach extends almost verbatim to two-sample edge-count tests based on general similarity graphs. Specifically, suppose
G n G_n G n is any simple graph constructed on the pooled sample Z n \mathcal{Z}_n Z n without using the sample labels. Define T × ( G n ) T^{\times}(G_n) T × ( G n ) and U n × ( G n ) U_n^{\times}(G_n) U n × ( G n ) as in (7) and (10) , respectively, with G K ( Z n ) G_K(\mathcal{Z}_n) G K ( Z n ) replaced by G n G_n G n . The formulas for the conditional mean and variance remain unchanged under this replacement. Under H 0 H_0 H 0 , conditional on F \mathcal{F} F , the vertex labels are independent Ber ( λ n ) \operatorname{Ber}(\lambda_n) Ber ( λ n ) variables. Consequently, the quantitative bounds provided by Theorem 3.3 and Proposition 3.2 in § 3 give the following general conditions for asymptotic normality for any similarity graph G n G_n G n constructed as above with N ( ⋅ , G n ) → P ∞ N(\cdot,G_n) \xrightarrow{P} \infty N ( ⋅ , G n ) P ∞ :
If λ n → λ ∈ ( 0 , 1 ) ∖ { 1 2 } \lambda_n \to\lambda\in(0,1)\setminus\left\{\frac{1}{2}\right\} λ n → λ ∈ ( 0 , 1 ) ∖ { 2 1 } , then the N ( ⋄ , G n ) ( N ( ∙ , G n ) + N ( ⋄ , G n ) ) 2 → P 0 \frac{N(\diamond,G_n)}{(N(\bullet,G_n)+N(\diamond,G_n))^2} \xrightarrow{P} 0 ( N ( ∙ , G n ) + N ( ⋄ , G n ) ) 2 N ( ⋄ , G n ) P 0 implies Wass ( U n × ( G n ) , N ( 0 , 1 ) ) → 0 \operatorname{Wass}(U_n^{\times}(G_n),N(0,1)) \to0 Wass ( U n × ( G n ) , N ( 0 , 1 )) → 0 .
If n 1 = n 2 n_1=n_2 n 1 = n 2 , the same conclusion holds when N ( □ , G n ) N ( ∙ , G n ) 2 → P 0 \frac{N(\square,G_n)}{N(\bullet,G_n)^2} \xrightarrow{P} 0 N ( ∙ , G n ) 2 N ( □ , G n ) P 0 .
Examples of commonly used similarity graphs beyond K K K -NNs, include K K K -minimum spanning trees (K K K -MSTs) [13 , 22 ] and K K K -minimum-distance matchings (K K K -matchings) [51 ] . A K K K -MST is the union of K K K successively constructed minimum spanning trees, with previously selected edges excluded at each step. Similarly, on an even number of observations, a K K K -matching is the union of K K K successively constructed minimum-weight perfect matchings, each minimizing the total dissimilarity between paired observations subject to the exclusion of previously selected edges. The case K = 1 K=1 K = 1 gives the graph underlying Rosenbaum’s cross-match test [46 ] . For the Poissonized sample, if L n L_n L n is odd, we omit one observation using a rule that does not depend on the sample labels. If K K K exceeds the number of retained observations minus one, so that K K K edge-disjoint perfect matchings do not exist, we adopt the convention that G n G_n G n is the complete graph on the retained observations, as in Definition 1.5 . Each retained vertex then has degree min { K , ∣ V ( G n ) ∣ − 1 } \min\{K,|V(G_n)|-1\} min { K , ∣ V ( G n ) ∣ − 1 } , hence, the relevant graph conditions for asymptotic normality can be easily verified. This leads to the following corollary, which follows by the same argument as in the proof of Theorem 1.6 :
Corollary 1.7. Let G n G_n G n be the K K K -matching graph described above and define U n × ( G n ) U_n^{\times}(G_n) U n × ( G n ) as in (10) . Then the following hold under H 0 H_0 H 0 :
(1) Suppose λ n = n 1 n → λ ∈ ( 0 , 1 ) ∖ { 1 2 } \lambda_n=\frac{n_1}{n}\to\lambda\in(0,1)\setminus\left\{\frac{1}{2}\right\} λ n = n n 1 → λ ∈ ( 0 , 1 ) ∖ { 2 1 } . Then Wass ( U n × ( G n ) , N ( 0 , 1 ) ) → 0 \operatorname{Wass}(U_n^{\times}(G_n),N(0,1))\to0 Wass ( U n × ( G n ) , N ( 0 , 1 )) → 0 , for any sequence K n ≥ 1 K_n\ge1 K n ≥ 1 .
(2) Suppose n 1 = n 2 n_1=n_2 n 1 = n 2 . Then Wass ( U n × ( G n ) , N ( 0 , 1 ) ) → 0 \operatorname{Wass}(U_n^{\times}(G_n),N(0,1))\to0 Wass ( U n × ( G n ) , N ( 0 , 1 )) → 0 , for any fixed K ≥ 1 K\ge1 K ≥ 1 or any sequence K n → ∞ K_n\to\infty K n → ∞ with K n = o ( n ) K_n=o(n) K n = o ( n ) .
The feature that makes two-sample tests based on geometric graphs particularly attractive is their universal consistency, that is, their power converges to one against every fixed alternative for which f f f and g g g differ on a set of positive Lebesgue measure. This is classically known for the K K K -NN-based test when K K K is fixed [49 ] (see also [28 ] ). Although the benefits of allowing K K K to grow with the sample size have often been advocated [4 , 13 ] , consistency in the diverging-K K K regime was established only recently by Kanekar [31 ] , Proposition 2.1 for the directed K K K -NN test when K = o ( n ) K=o(n) K = o ( n ) . In the appendix of this paper, we prove universal consistency for the undirected K K K -NN test under the same condition (see Proposition B.1 ). We also illustrate the finite-sample performance of the test through simulations in § 7 .
Remarks and Overview of Proofs
The fourth moment phenomenon was first discovered by Nualart and Peccati [43 ] , who showed that the convergence of the first, second, and fourth moments to 0 0 0 , 1 1 1 , and 3 3 3 , respectively, guarantees asymptotic normality for a sequence of multiple stochastic Wiener-Itô integrals of fixed order. Later, Nourdin and Peccati [39 , 41 ] provided error bounds for the fourth moment theorem. This phenomenon has since emerged as a general principle governing Gaussian limits of nonlinear functionals of random fields (see the book [40 ] for an introduction to the topic and the website https://sites.google. com/site/malliavinstein/home for a comprehensive list of the recent results). Related results for degenerate U U U -statistics of fixed order were obtained earlier by de Jong [15 , 16 ] . In this setting, the fourth moment condition generally requires an additional assumption controlling the maximum influence of the underlying independent variables (see [15 Theorem 2.1] , [16 Theorem 1] , and [19 Corollary 1.2] ). For other classical sufficient conditions for asymptotic normality and convergence rates for quadratic forms, see [12 , 23 , 24 , 27 , 48 ] and the references therein. In the context of random graph coloring, the fourth moment phenomenon was first observed for monochromatic edge counts in the balanced case by Bhattacharya et al. [6 ] , and later extended to counts of more general monochromatic subgraphs in [8 , 14 , 37 ] (see also [18 , 21 ] for related results).
An instance of the third moment phenomenon was established by Neufcourt and Viens [38 Theorem 3.2] for normalized quadratic variations of stationary Gaussian sequences under suitable covariance assumptions. For these statistics, which belong to the second Wiener chaos, both the third and fourth moment conditions characterize asymptotic normality (see also [32 ] for related results). Our results, on the other hand, establish a third moment characterization for monochromatic-edge counts on arbitrary graph sequences under biased Bernoulli coloring.
The fixed p p p regime: We shall make remarks on each of the three cases of Theorem 1.1 .
For p ∈ ( 0 , 1 ) ∖ { 1 2 , 1 ± 1 / 3 2 } p \in(0,1)\setminus\left\{\frac{1}{2},\frac{1\pm1/\sqrt{3}}{2}\right\} p ∈ ( 0 , 1 ) ∖ { 2 1 , 2 1 ± 1/ 3 } , the fourth moment characterization in Theorem 1.1 (i) is an instance of the broader phenomenon described above, but is new for biased graph colorings. The equivalent third moment characterization is a new phenomenon in its own right : convergence of an even lower-order moment suffices to determine Gaussian convergence.
At the exceptional probabilities p = 1 ± 1 / 3 2 p=\frac{1\pm1/\sqrt{3}}{2} p = 2 1 ± 1/ 3 , the coefficient of the 4-star term in the fourth cumulant E [ U n 4 ] − 3 \mathbb{E}[U_n^4]-3 E [ U n 4 ] − 3 vanishes, while all remaining terms tend to zero (see Lemma 2.3). The fourth moment therefore no longer detects the 4-star configurations responsible for non-Gaussian behavior, which is why it can converge to 3 while asymptotic normality fails, whereas the third moment, whose 4-star coefficient does not vanish, continues to characterize Gaussian convergence.
The exceptional probabilities p = 1 ± 1 / 3 2 p=\frac{1\pm1/\sqrt{3}}{2} p = 2 1 ± 1/ 3 have appeared before in [19 Corollary 1.4 and Example 1.5] in connection with fourth moment criteria for linear combinations of independent biased Rademacher variables. Their significance comes from the identity
E [ ( X − p p ( 1 − p ) ) 4 ] − 3 = 1 − 6 p ( 1 − p ) p ( 1 − p ) , \mathbb{E}\left[\left(\frac{X-p}{\sqrt{p(1-p)}}\right)^4\right]-3=\frac{1-6p(1-p)}{p(1-p)}, E ( p ( 1 − p ) X − p ) 4 − 3 = p ( 1 − p ) 1 − 6 p ( 1 − p ) ,
where X ∼ Ber ( p ) X\sim\operatorname{Ber}(p) X ∼ Ber ( p ) . Thus, precisely at these probabilities, the standardized Bernoulli variable has fourth moment 3 despite being non-Gaussian. Döbler and Krokowski [19 ] establish a fourth moment criterion for normalized linear combinations away from these values and show that the exclusion is necessary. The same cancellation explains the exceptional probabilities in Theorem 1.1 , at which the 4-star contribution to the fourth cumulant of U n U_n U n vanishes. Our result identifies the exact graph condition for asymptotic normality of monochromatic edge counts, whose expansion in centered variables contains both linear and quadratic terms under biased colorings. At the exceptional probabilities, this condition is equivalent to vanishing of the third moment, while the fourth moment condition holds automatically and therefore cannot characterize asymptotic normality. To the best of our knowledge, this provides the first example of a sequence of quadratic multilinear polynomials in independent biased Rademacher variables for which the third moment characterizes Gaussian convergence, even though the fourth moment criterion fails.
Finally, in the balanced case p = 1 2 p=\frac{1}{2} p = 2 1 , the edge indicators { 1 { X i = X j } } { i , j } ∈ E ( G n ) \left\{\mathbf{1}\{X_i=X_j\}\right\}_{\{i,j\}\in E(G_n)} { 1 { X i = X j } } { i , j } ∈ E ( G n ) are pairwise independent, and the centered statistic in (1) is a Rademacher chaos of order two, that is, a
quadratic form in i.i.d. standard Rademacher variables. Indeed, setting R i = 2 X i − 1 ∈ { ± 1 } R_i=2X_i-1\in\{\pm1\} R i = 2 X i − 1 ∈ { ± 1 } for 1 ≤ i ≤ n 1\le i\le n 1 ≤ i ≤ n , we obtain
T n − E [ T n ] = 1 4 ∑ 1 ≤ i , j ≤ n a i j R i R j , (11) T_n-\mathbb{E}[T_n]=\frac{1}{4}\sum_{1\le i,j\le n}a_{ij}R_iR_j, \tag*{(11)} T n − E [ T n ] = 4 1 1 ≤ i , j ≤ n ∑ a ij R i R j , (11)
where a i j = 1 { { i , j } ∈ E ( G n ) } a_{ij}=\mathbf{1}\{\{i,j\}\in E(G_n)\} a ij = 1 {{ i , j } ∈ E ( G n )} and the variables R 1 , … , R n R_1,\ldots,R_n R 1 , … , R n are i.i.d. standard Rademacher variables. As mentioned earlier, asymptotic normality of U n U_n U n in this setting has been extensively studied, with Theorem 1.1 (iii) established in [6 Theorem 1.4] (see also [9 , 42 ] ). Here, the fourth moment condition, which requires the number of 4-cycles to be negligible relative to its extremal order, is necessary and sufficient for asymptotic normality of U n U_n U n . As discussed earlier, the third moment condition, however, is no longer sufficient. The third moment can converge to zero, even when asymptotic normality fails (see Example 6.2).
Outline of the Paper
We now provide a brief outline of the proofs in the fixed p p p regime. In § 2 we prove equivalence of the third and fourth moment phenomena to their respective motif count conditions in Lemma 2.1 and Lemma 2.3. These lemma follow by expanding the third and fourth centered moments of T n T_n T n in terms of subgraph counts and identifying the leading contributions in each regime. In § 3 , we show convergence of U n U_n U n to Normal distribution in the Wasserstein metric under the respective motif count conditions proven in Lemma 2.1 and Lemma 2.3. For p = 1 2 p=\frac{1}{2} p = 2 1 we reprove the well known result in Proposition 3.2 and for unbalanced case in Theorem 3.3 (i). The proof of Theorem 3.3 is an application of Stein’s method and an overview is given in that section. A uniform integrability argument (see Lemma 4.1 ) ensures that asymptotic normality of U n U_n U n implies convergence of its third and fourth moments to their Gaussian values.
We also note that, since we establish fourth or third-moment phenomena, weak convergence is equivalent to convergence in the Wasserstein or Kolmogorov distance. Finding the precise rate of convergence in any metric remains an open question.
The sparse p p p regime: The proof strategy parallels the fixed p p p regime. In Lemma 2.2 we show that the third moment condition is equivalent to a motif/p n p_n p n condition which is weaker than (that is, implied by) the motif/p n p_n p n condition characterizing the fourth moment condition in Lemma 1.4 ; both lemmas are proved in § 2 , and Proposition 1.3 follows immediately from them. In Theorem 3.3 (ii) we show convergence of U n U_n U n to the normal distribution in the Wasserstein metric under the motif/p n p_n p n condition of Lemma 2.2 , which yields Theorem 1.2 . Although the overall strategy is the same as for fixed p p p , it is considerably harder to execute, because the moment bounds must now be controlled uniformly in the varying p n p_n p n . This is explained in detail in § 3 .
Graph based two-sample tests: Note that Theorem 1.6 treats two cases: unbalanced sample sizes with λ n = n 1 n → λ ∈ ( 0 , 1 ) ∖ { 1 2 } \lambda_n=\frac{n_1}{n}\to\lambda\in(0,1)\setminus\{\frac{1}{2}\} λ n = n n 1 → λ ∈ ( 0 , 1 ) ∖ { 2 1 } , and equal sample sizes with n 1 = n 2 n_1=n_2 n 1 = n 2 . It does not cover the asymptotically balanced case, where n 1 ≠ n 2 n_1\ne n_2 n 1 = n 2 , but λ n → 1 2 \lambda_n\to\frac{1}{2} λ n → 2 1 . The conclusion of Theorem 1.6 (ii) remains valid in this regime, but establishing this extension requires a more careful argument, which we omit.
Recently, Kanekar [31 ] studied a related two-sample statistic based on directed nearest-neighbor graphs in the Poissonized setting. In particular, [31 Theorem 3.1] gives a central limit theorem under H 0 H_0 H 0 for the corresponding test statistic in the diverging K K K regime, under K n = o ( n ) K_n=o(n) K n = o ( n ) . For the undirected statistic (10) considered here, Theorem 1.6 establishes asymptotic normality for arbitrary K n K_n K n in the unbalanced case λ ≠ 1 2 \lambda\ne\frac{1}{2} λ = 2 1 , and for K n = o ( n ) K_n=o(n) K n = o ( n ) in the balanced case n 1 = n 2 n_1=n_2 n 1 = n 2 . For related results on the asymptotic normality of two-sample statistics based on general similarity graphs under the fixed-sample-size permutation null, see Zhu and Chen [52 ] .
Outline of the Paper
In § 2 we express the third and fourth moments of U n U_n U n in terms of motif counts and prove the equivalences between the moment conditions and the motif conditions, in both the fixed p p p and the sparse regimes. In § 3 we prove convergence of U n U_n U n to the normal distribution in the Wasserstein metric under these motif conditions. Section 4 combines these results to prove Theorem 1.1 , Theorem 1.2 and Proposition 1.3, and Section 5 contains the proof of Theorem 1.6 . The examples showing that the moment conditions cannot be improved are collected in § 6 , and § 7 reports simulations assessing the finite-sample performance of the K-NN test. The appendix contains the motif tables used in the proofs (Appendix § 8 ) and the proof of universal consistency of the undirected K-NN test (Proposition B.1 ).
Acknowledgments: SA was supported through a Knowledge Exchange Grant, under project no. RTI4001 of the Department of Atomic Energy, Government of India. BBB was supported by NSF CAREER grant DMS 2046393, a Sloan Research Fellowship, and the National University of Singapore’s PESS and Provost’s Chair funds. The authors thank Frank den Hollander and Rajat Subhra Hazra who along with SA organized the ICTS-Networks Workshop in April 06-17th, 2026 at the ICTS-TIFR in India, where the research in this paper started. The authors also thank ICTS-TIFR for support and hospitality.
AI disclosure AI tools (Claude Fable 5.1, GPT 5.6 Sol and GPT 6 Astra) were used to compute the motif tables, verify some of the mathematical arguments, and refine grammar and typography. The authors assume responsibility for all content.
Convergence Criteria for Third and Fourth Moment of U n U_n U n
In this section we obtain a characterization of the third moment condition and fourth moment condition in all regimes in terms of motif counts. Let ( p n ) n ∈ N (p_n)_{n\in\mathbb{N}} ( p n ) n ∈ N be a sequence such that p n ∈ ( 0 , 1 ) p_n \in(0,1) p n ∈ ( 0 , 1 ) , for all n ∈ N n \in\mathbb{N} n ∈ N . In the fixed p p p regime p n = p ∈ ( 0 , 1 ) p_n=p\in(0,1) p n = p ∈ ( 0 , 1 ) , for all n ∈ N n\in\mathbb{N} n ∈ N , and in the sparse regime p n → 0 p_n\to0 p n → 0 , as n → ∞ n\to\infty n → ∞ . For each e = { i , j } ∈ E ( G n ) e=\{i,j\}\in E(G_n) e = { i , j } ∈ E ( G n ) , denote
Y e : = 1 { X i = X j } − q n , (12) Y_e := \mathbf{1}\{X_i=X_j\}-q_n, \tag*{(12)} Y e := 1 { X i = X j } − q n , (12)
where q n = E [ 1 { X i = X j } ] = p n 2 + ( 1 − p n ) 2 q_n=\mathbb{E}[\mathbf{1}\{X_i=X_j\}]=p_n^2+(1-p_n)^2 q n = E [ 1 { X i = X j }] = p n 2 + ( 1 − p n ) 2 . Recall the definition of T n T_n T n from (1) . Denote
S n : = T n − E [ T n ] = ∑ e ∈ E ( G n ) Y e . (13) S_n:=T_n-\mathbb{E}[T_n]=\sum_{e\in E(G_n)}Y_e. \tag*{(13)} S n := T n − E [ T n ] = e ∈ E ( G n ) ∑ Y e . (13)
Criteria for Convergence of Third Moment of U n U_n U n
First we compute the third moment of S n S_n S n . For this, expanding the sum in S n 3 S_n^3 S n 3 gives,
E [ S n 3 ] = E [ ∑ e 1 , e 2 , e 3 ∈ E ( G n ) Y e 1 Y e 2 Y e 3 ] . \mathbb{E}\left[S_n^3\right]=\mathbb{E}\left[\sum_{e_1,e_2,e_3\in E(G_n)}Y_{e_1}Y_{e_2}Y_{e_3}\right]. E [ S n 3 ] = E e 1 , e 2 , e 3 ∈ E ( G n ) ∑ Y e 1 Y e 2 Y e 3 .
Since E [ Y e ] = 0 \mathbb{E}[Y_e]=0 E [ Y e ] = 0 for all e ∈ E ( G n ) e\in E(G_n) e ∈ E ( G n ) , and Y e Y_e Y e is independent of Y e ′ Y_{e'} Y e ′ whenever e ∩ e ′ = ∅ e\cap e'=\varnothing e ∩ e ′ = ∅ , the only non-vanishing summands are those for which each connected component of e 1 ∪ e 2 ∪ e 3 e_1\cup e_2\cup e_3 e 1 ∪ e 2 ∪ e 3 contains at least two edges, counting multiplicities. Grouping the summands according to the isomorphism type of e 1 ∪ e 2 ∪ e 3 e_1 \cup e_2 \cup e_3 e 1 ∪ e 2 ∪ e 3 , as in the proof of Lemma 3.5, we can write
E [ S n 3 ] = ∑ H ∈ H 3 c H ( 3 ) ( u n ) N ( H , G n ) , (14) \mathbb{E}\left[S_n^3\right]=\sum_{H\in\mathcal{H}_3}c_H^{(3)}(u_n)N(H,G_n), \tag*{(14)} E [ S n 3 ] = H ∈ H 3 ∑ c H ( 3 ) ( u n ) N ( H , G n ) , (14)
where u n = ( 2 p n − 1 ) 2 u_n=(2p_n-1)^2 u n = ( 2 p n − 1 ) 2 and H 3 : = { three-edge motifs } \mathcal{H}_3:=\{\text{three-edge motifs}\} H 3 := { three-edge motifs } . Table 2 in Appendix § 8 gives the exact values of the coefficients c H ( 3 ) ( u n ) c_H^{(3)}(u_n) c H ( 3 ) ( u n ) and the asymptotic order of the subgraph count associated with each motif. Using this we can characterize the third moment condition E [ U n 3 ] → 0 \mathbb{E}[U_n^3]\to0 E [ U n 3 ] → 0 in terms of the motif counts.
Table 2: Coefficients c H ( 3 ) ( u n ) c_H^{(3)}(u_n) c H ( 3 ) ( u n ) in the expansion of E [ S n 3 ] \mathbb{E}[S_n^3] E [ S n 3 ] and the order of the motif counts.
Table 2. Coefficients c H ( 4 ) ( u n ) c_H^{(4)}(u_n) c H ( 4 ) ( u n ) in the expansion of E [ S n 4 ] \mathbb{E}\left[S_n^4\right] E [ S n 4 ] and the order of the motif counts.
The Fixed p p p Regime: We obtain equivalent condition for balanced case p = 1 2 p=\frac{1}{2} p = 2 1 involving the number of triangles and number of edges, whereas in the unbalanced case we need to control the number of 3-stars w.r.t the number of edges and 2-stars.
Lemma 2.1. Let U n U_n U n be as defined in (2) and N ( edge , G n ) → ∞ N(\text{edge},G_n)\to\infty N ( edge , G n ) → ∞ , as n → ∞ n\to\infty n → ∞ . Then the following hold.
(a) If p ∈ ( 0 , 1 ) ∖ { 1 2 } p\in(0,1)\setminus\{\frac{1}{2}\} p ∈ ( 0 , 1 ) ∖ { 2 1 } , then
E [ U n 3 ] → 0 ⟺ N ( triangle , G n ) = o ( ( N ( edge , G n ) + N ( 2-star , G n ) ) 3 2 ) . (15) \mathbb{E}\left[U_n^3\right]\to0\quad\Longleftrightarrow\quad N(\text{triangle},G_n)=o\left((N(\text{edge},G_n)+N(\text{2-star},G_n))^{\frac{3}{2}}\right). \tag*{(15)} E [ U n 3 ] → 0 ⟺ N ( triangle , G n ) = o ( ( N ( edge , G n ) + N ( 2-star , G n ) ) 2 3 ) . (15)
(b) If p = 1 2 p=\frac{1}{2} p = 2 1 , then
E [ U n 3 ] → 0 ⟺ N ( 3-star , G n ) = o ( N ( edge , G n ) 3 2 ) . (16) \mathbb{E}\left[U_n^3\right]\to0\quad\Longleftrightarrow\quad N(\text{3-star},G_n)=o\left(N(\text{edge},G_n)^{\frac{3}{2}}\right). \tag*{(16)} E [ U n 3 ] → 0 ⟺ N ( 3-star , G n ) = o ( N ( edge , G n ) 2 3 ) . (16)
Proof. First, suppose p ∈ ( 0 , 1 ) ∖ { 1 2 } p\in(0,1)\setminus\{\frac{1}{2}\} p ∈ ( 0 , 1 ) ∖ { 2 1 } . In this case, recall (31) ,
σ n 2 ≍ N ( edge , G n ) + N ( 2-star , G n ) . \sigma_n^2\asymp N(\text{edge},G_n)+N(\text{2-star},G_n). σ n 2 ≍ N ( edge , G n ) + N ( 2-star , G n ) .
Referring to the order of the motif counts in Table 2 it follows that, for all H ∈ H 3 ∖ { triangle , 3-star } H\in\mathcal{H}_3\setminus\{\text{triangle},\text{3-star}\} H ∈ H 3 ∖ { triangle , 3-star } ,
N ( H , G n ) = o ( ( N ( edge , G n ) + N ( 2-star , G n ) ) 3 2 ) . (17) N(H,G_n)=o\left((N(\text{edge},G_n)+N(\text{2-star},G_n))^{\frac{3}{2}}\right). \tag*{(17)} N ( H , G n ) = o ( ( N ( edge , G n ) + N ( 2-star , G n ) ) 2 3 ) . (17)
For H = 3-star H=\text{3-star} H = 3-star , we use the following estimate: For any t > 0 t>0 t > 0 ,
N ( H , G n ) ≤ ∑ { i , j } ∈ E ( G n ) d G n ( i ) d G n ( j ) ≤ t ∑ { i , j } ∈ E ( G n ) ( d G n ( i ) + d G n ( j ) ) + 1 2 ( ∑ i : d G n ( i ) > t d G n ( i ) ) 2 . N(H,G_n)\leq\sum_{\{i,j\}\in E(G_n)}d_{G_n}(i)d_{G_n}(j)\leq t\sum_{\{i,j\}\in E(G_n)}(d_{G_n}(i)+d_{G_n}(j))+\frac{1}{2}\left(\sum_{i:d_{G_n}(i)>t}d_{G_n}(i)\right)^2. N ( H , G n ) ≤ { i , j } ∈ E ( G n ) ∑ d G n ( i ) d G n ( j ) ≤ t { i , j } ∈ E ( G n ) ∑ ( d G n ( i ) + d G n ( j )) + 2 1 i : d G n ( i ) > t ∑ d G n ( i ) 2 .
Denote δ 2 : = ∑ i = 1 n d G n ( i ) 2 \delta_2:=\sum_{i=1}^{n}d_{G_n}(i)^2 δ 2 := ∑ i = 1 n d G n ( i ) 2 . Note that ∑ { i , j } ∈ E ( G n ) ( d G n ( i ) + d G n ( j ) ) = δ 2 \sum_{\{i,j\}\in E(G_n)}(d_{G_n}(i)+d_{G_n}(j))=\delta_2 ∑ { i , j } ∈ E ( G n ) ( d G n ( i ) + d G n ( j )) = δ 2 and
∑ i : d G n ( i ) > t d G n ( i ) ≤ 1 t ∑ i : d G n ( i ) > t d G n ( i ) 2 ≤ δ 2 t . \sum_{i:d_{G_n}(i)>t}d_{G_n}(i)\leq\frac{1}{t}\sum_{i:d_{G_n}(i)>t}d_{G_n}(i)^2\leq\frac{\delta_2}{t}. i : d G n ( i ) > t ∑ d G n ( i ) ≤ t 1 i : d G n ( i ) > t ∑ d G n ( i ) 2 ≤ t δ 2 .
Then choosing t = δ 2 1 2 t=\delta_2^{\frac{1}{2}} t = δ 2 2 1 gives,
N ( 3-star , G n ) ≤ 3 2 δ 2 3 2 = o ( δ 2 3 2 ) = o ( ( N ( edge , G n ) + N ( 2-star , G n ) ) 3 2 ) . (18) N(\text{3-star},G_n)\leq\frac{3}{2}\delta_2^{\frac{3}{2}}=o(\delta_2^{\frac{3}{2}})=o\left((N(\text{edge},G_n)+N(\text{2-star},G_n))^{\frac{3}{2}}\right). \tag*{(18)} N ( 3-star , G n ) ≤ 2 3 δ 2 2 3 = o ( δ 2 2 3 ) = o ( ( N ( edge , G n ) + N ( 2-star , G n ) ) 2 3 ) . (18)
Combining (17) and (18) with (14) and referring to Table 2 gives,
E [ U n 3 ] = − 3 u 2 ( 1 − u ) 2 N ( triangle , G n ) σ n 3 + o ( 1 ) , \mathbb{E}\left[U_n^3\right]=-\frac{3u^2(1-u)}{2}\frac{N(\text{triangle},G_n)}{\sigma_n^3}+o(1), E [ U n 3 ] = − 2 3 u 2 ( 1 − u ) σ n 3 N ( triangle , G n ) + o ( 1 ) ,
where u = ( 2 p − 1 ) 2 u=(2p-1)^2 u = ( 2 p − 1 ) 2 . This proves the result in (15) .
Criteria for Convergence of Fourth Moment of Un
Next, suppose p = 1 2 p=\frac{1}{2} p = 2 1 , that is u = 0 u=0 u = 0 . In this case, referring to Table 2 gives c H ( 3 ) ( u ) = 0 c_H^{(3)}(u)=0 c H ( 3 ) ( u ) = 0 , for all H ∈ H 3 ∖ { triangle } H\in\mathcal{H}_3\setminus\{\text{triangle}\} H ∈ H 3 ∖ { triangle } . Hence,
E [ U n 3 ] = 3 N ( 3-star , G n ) 4 σ n 3 . \mathbb{E}\left[U_n^3\right]=\frac{3N(\text{3-star},G_n)}{4\sigma_n^3}. E [ U n 3 ] = 4 σ n 3 3 N ( 3-star , G n ) .
This proves the result in (16) , since, in this case, σ n 2 ≍ N ( edge , G n ) \sigma_n^2\asymp N(\text{edge},G_n) σ n 2 ≍ N ( edge , G n ) (recall (31) ).
The Sparse p p p Regime: Here we show that the third moment condition is equivalent to a condition requiring that the rate of decay of p n p_n p n and the number of 3-stars compare suitably with the number of edges and 2-stars.
Lemma 2.2. Let { G n } n ∈ N \{G_n\}_{n\in\mathbb{N}} { G n } n ∈ N be a sequence of graphs such that N ( edge , G n ) → ∞ N(\text{edge},G_n)\to\infty N ( edge , G n ) → ∞ and U n U_n U n be as defined in (2) . Suppose p = p n → 0 p=p_n\to0 p = p n → 0 , as n → ∞ n\to\infty n → ∞ . Then the following are equivalent:
(a) E [ U n 3 ] → 0 \mathbb{E}[U_n^3]\to0 E [ U n 3 ] → 0 .
(b) p n ( N ( edge , G n ) + N ( 2-star , G n ) ) → ∞ p_n(N(\text{edge},G_n)+N(\text{2-star},G_n))\to\infty p n ( N ( edge , G n ) + N ( 2-star , G n )) → ∞ and N ( 3-star , G n ) = o ( p n 1 2 ( N ( edge , G n ) + N ( 2-star , G n ) ) 3 2 ) N(\text{3-star},G_n)=o\left(p_n^{\frac{1}{2}}(N(\text{edge},G_n)+N(\text{2-star},G_n))^{\frac{3}{2}}\right) N ( 3-star , G n ) = o ( p n 2 1 ( N ( edge , G n ) + N ( 2-star , G n ) ) 2 3 ) .
Proof. Recall that u n = ( 2 p n − 1 ) 2 u_n=(2p_n-1)^2 u n = ( 2 p n − 1 ) 2 , so that u n = 1 + O ( p n ) u_n=1+O(p_n) u n = 1 + O ( p n ) and 1 − u n = 4 p n + O ( p n 2 ) 1-u_n=4p_n+O(p_n^2) 1 − u n = 4 p n + O ( p n 2 ) . Referring to Table 2 ,
E [ S n 3 ] = − 2 p n N ( edge , G n ) − 6 p n ( N ( 2-star , G n ) + N ( 3-star , G n ) ) + Q n , \mathbb{E}[S_n^3]=-2p_nN(\text{edge},G_n)-6p_n\left(N(\text{2-star},G_n)+N(\text{3-star},G_n)\right)+Q_n, E [ S n 3 ] = − 2 p n N ( edge , G n ) − 6 p n ( N ( 2-star , G n ) + N ( 3-star , G n ) ) + Q n ,
where
∣ Q n ∣ = O ( p n 2 ( N ( edge , G n ) + N ( 2-star , G n ) + N ( 3-star , G n ) + N ( path , G n ) + N ( triangle , G n ) ) ) . |Q_n|=O\left(p_n^2\left(N(\text{edge},G_n)+N(\text{2-star},G_n)+N(\text{3-star},G_n)+N(\text{path},G_n)+N(\text{triangle},G_n)\right)\right). ∣ Q n ∣ = O ( p n 2 ( N ( edge , G n ) + N ( 2-star , G n ) + N ( 3-star , G n ) + N ( path , G n ) + N ( triangle , G n ) ) ) .
To control the error term, note that each copy of the path has a unique middle edge. Hence,
N ( path , G n ) = ∑ { i , j } ∈ E ( G n ) ( d G n ( i ) − 1 ) ( d G n ( j ) − 1 ) ≤ ∑ { i , j } ∈ E ( G n ) d G n ( i ) d G n ( j ) ≤ 1 2 ∑ { i , j } ∈ E ( G n ) ( d G n ( i ) 2 + d G n ( j ) 2 ) = 1 2 ∑ i = 1 n d G n ( i ) 3 = N ( edge , G n ) + 3 N ( 2-star , G n ) + 3 N ( 3-star , G n ) . \begin{aligned}
N(\text{path},G_n)=\sum_{\{i,j\}\in E(G_n)}(d_{G_n}(i)-1)(d_{G_n}(j)-1)\le\sum_{\{i,j\}\in E(G_n)}d_{G_n}(i)d_{G_n}(j) \\
&\le\frac{1}{2}\sum_{\{i,j\}\in E(G_n)}\left(d_{G_n}(i)^2+d_{G_n}(j)^2\right) \\
&=\frac{1}{2}\sum_{i=1}^{n}d_{G_n}(i)^3 \\
&=N(\text{edge},G_n)+3N(\text{2-star},G_n)+3N(\text{3-star},G_n).
\end{aligned} N ( path , G n ) = { i , j } ∈ E ( G n ) ∑ ( d G n ( i ) − 1 ) ( d G n ( j ) − 1 ) ≤ { i , j } ∈ E ( G n ) ∑ d G n ( i ) d G n ( j ) ≤ 2 1 { i , j } ∈ E ( G n ) ∑ ( d G n ( i ) 2 + d G n ( j ) 2 ) = 2 1 i = 1 ∑ n d G n ( i ) 3 = N ( edge , G n ) + 3 N ( 2-star , G n ) + 3 N ( 3-star , G n ) .
Moreover, 3 N ( triangle , G n ) ≤ N ( 2-star , G n ) 3N(\text{triangle},G_n)\le N(\text{2-star},G_n) 3 N ( triangle , G n ) ≤ N ( 2-star , G n ) . Hence,
∣ Q n ∣ = O ( p n 2 ( N ( edge , G n ) + 3 N ( 2-star , G n ) + 3 N ( 3-star , G n ) ) ) . |Q_n|=O\left(p_n^2\left(N(\text{edge},G_n)+3N(\text{2-star},G_n)+3N(\text{3-star},G_n)\right)\right). ∣ Q n ∣ = O ( p n 2 ( N ( edge , G n ) + 3 N ( 2-star , G n ) + 3 N ( 3-star , G n ) ) ) .
This implies,
E [ S n 3 ] = − 2 p n ( N ( edge , G n ) + 3 N ( 2-star , G n ) + 3 N ( 3-star , G n ) ) ( 1 + O ( p n ) ) . \mathbb{E}[S_n^3]=-2p_n\left(N(\text{edge},G_n)+3N(\text{2-star},G_n)+3N(\text{3-star},G_n)\right)(1+O(p_n)). E [ S n 3 ] = − 2 p n ( N ( edge , G n ) + 3 N ( 2-star , G n ) + 3 N ( 3-star , G n ) ) ( 1 + O ( p n )) .
Also, (3) gives σ n 2 = 2 p n ( N ( edge , G n ) + N ( 2-star , G n ) ) ( 1 + O ( p n ) ) \sigma_n^2=2p_n(N(\text{edge},G_n)+N(\text{2-star},G_n))(1+O(p_n)) σ n 2 = 2 p n ( N ( edge , G n ) + N ( 2-star , G n )) ( 1 + O ( p n )) . Therefore,
E [ U n 3 ] = − 1 + O ( p n ) 2 N ( edge , G n ) + 3 N ( 2-star , G n ) + 3 N ( 3-star , G n ) p n 1 2 ( N ( edge , G n ) + N ( 2-star , G n ) ) 3 2 . \mathbb{E}[U_n^3]=-\frac{1+O(p_n)}{\sqrt{2}}\frac{N(\text{edge},G_n)+3N(\text{2-star},G_n)+3N(\text{3-star},G_n)}{p_n^{\frac{1}{2}}(N(\text{edge},G_n)+N(\text{2-star},G_n))^{\frac{3}{2}}}. E [ U n 3 ] = − 2 1 + O ( p n ) p n 2 1 ( N ( edge , G n ) + N ( 2-star , G n ) ) 2 3 N ( edge , G n ) + 3 N ( 2-star , G n ) + 3 N ( 3-star , G n ) .
This proves the result in Lemma 2.2.
Criteria for Convergence of Fourth Moment of U n U_n U n
Next, we consider the fourth moment of S n S_n S n . Expanding as above gives,
E [ S n 4 ] = E [ ∑ e 1 , e 2 , e 3 , e 4 ∈ E ( G n ) Y e 1 Y e 2 Y e 3 Y e 4 ] . (19) \mathbb{E}\left[S_n^4\right]=\mathbb{E}\left[\sum_{e_1,e_2,e_3,e_4\in E(G_n)}Y_{e_1}Y_{e_2}Y_{e_3}Y_{e_4}\right]. \tag*{(19)} E [ S n 4 ] = E e 1 , e 2 , e 3 , e 4 ∈ E ( G n ) ∑ Y e 1 Y e 2 Y e 3 Y e 4 . (19)
As before, grouping the summands according to the isomorphism type of e 1 ∪ e 2 ∪ e 3 ∪ e 4 e_1\cup e_2\cup e_3\cup e_4 e 1 ∪ e 2 ∪ e 3 ∪ e 4 , as in the proof of Lemma 3.5, we can write
E [ S n 4 ] = ∑ H ∈ H 4 c H ( 4 ) ( u n ) N ( H , G n ) , (20) \mathbb{E}\left[S_n^4\right]=\sum_{H\in\mathcal{H}_4}c_H^{(4)}(u_n)N(H,G_n), \tag*{(20)} E [ S n 4 ] = H ∈ H 4 ∑ c H ( 4 ) ( u n ) N ( H , G n ) , (20)
where $\mathcal{H}_4:={\bullet-\bullet,\ \begin{smallmatrix}\bullet&\bullet\[-4pt]&\bullet\end{smallmatrix},\ \bullet-\bullet\sqcup\bullet-\bullet,\ \begin{smallmatrix}&\bullet&\[-4pt]\bullet-\bullet-\bullet\end{smallmatrix},\ \begin{smallmatrix}\bullet&\bullet&\bullet\[-4pt]&\bullet&\end{smallmatrix},\ \begin{smallmatrix}\bullet&\bullet\[-4pt]\bullet&\bullet\end{smallmatrix},\ \begin{smallmatrix}\bullet&\bullet&\bullet\[-4pt]\bullet&&\bullet\end{smallmatrix},\ \bullet-\bullet\sqcup\begin{smallmatrix}\bullet&\bullet\[-4pt]&\bullet\end{smallmatrix},\ \begin{smallmatrix}\bullet&\bullet&\bullet\[-4pt]&\bullet&\end{smallmatrix}\sqcup\bullet,\ \begin{smallmatrix}\bullet&\bullet\[-4pt]&\bullet\end{smallmatrix}\sqcup\begin{smallmatrix}\bullet&\bullet\[-4pt]&\bullet\end{smallmatrix},\ \begin{smallmatrix}\bullet&\bullet&\bullet\[-4pt]\bullet&&\bullet\end{smallmatrix}\sqcup\bullet,\ \begin{smallmatrix}\bullet&\bullet&\bullet\[-4pt]\bullet&&\bullet\end{smallmatrix}\sqcup\begin{smallmatrix}\bullet&\bullet\[-4pt]&\bullet\end{smallmatrix},\ \bullet-\bullet\sqcup\bullet-\bullet\sqcup\bullet-\bullet}.
{xref}`Table 3` in Appendix A gives the exact values of the coefficients $c_H^{(4)}(u_n)$ and the asymptotic order of the subgraph count associated with each motif. Using this we can characterize the fourth moment condition $\mathbb{E}[U_n^4]\to3$ in terms of the motif counts.
The Fixed p p p Regime: We obtain that as long as the edges diverge the fourth moment convergence holds for the exceptional values p = 1 ± 1 / 3 2 p=\frac{1\pm1/\sqrt{3}}{2} p = 2 1 ± 1/ 3 . We obtain equivalent conditions for balanced case p = 1 2 p=\frac{1}{2} p = 2 1 involving the number of squares and number of edges, whereas in the unbalanced case outside of the exceptional values we need to control the number of 4-stars w.r.t the number of edges and 2-stars.
Lemma 2.3. Let U n U_n U n be as defined in (1.2) and N ( ∙ − ∙ , G n ) → ∞ N(\bullet-\bullet,G_n)\to\infty N ( ∙ − ∙ , G n ) → ∞ , as n → ∞ n\to\infty n → ∞ . Then the following hold.
(a) If p ∈ ( 0 , 1 ) ∖ { 1 2 , 1 ± 1 / 3 2 } p\in(0,1)\setminus\left\{\frac{1}{2},\frac{1\pm1/\sqrt{3}}{2}\right\} p ∈ ( 0 , 1 ) ∖ { 2 1 , 2 1 ± 1/ 3 } , then
E [ U n 4 ] → 3 ⟺ N ( ∙ ∙ ∙ ∙ , G n ) = o ( ( N ( ∙ − ∙ , G n ) + N ( ∙ ∙ − ∙ − ∙ , G n ) ) 2 ) . (21) \mathbb{E}\left[U_n^4\right]\to3\quad\Longleftrightarrow\quad N\left(\begin{smallmatrix}\bullet&\bullet&\bullet\\[-4pt]&\bullet&\end{smallmatrix},G_n\right)=o\left(\left(N(\bullet-\bullet,G_n)+N\left(\begin{smallmatrix}&\bullet&\\[-4pt]\bullet-\bullet-\bullet\end{smallmatrix},G_n\right)\right)^2\right). \tag*{(21)} E [ U n 4 ] → 3 ⟺ N ( ∙ ∙ ∙ ∙ , G n ) = o ( ( N ( ∙ − ∙ , G n ) + N ( ∙−∙−∙ ∙ , G n ) ) 2 ) . (21)
(b) If p = 1 ± 1 / 3 2 p=\frac{1\pm1/\sqrt{3}}{2} p = 2 1 ± 1/ 3 , then, without any additional assumption,
E [ U n 4 ] → 3. \mathbb{E}\left[U_n^4\right]\to3. E [ U n 4 ] → 3.
(c) If p = 1 2 p=\frac{1}{2} p = 2 1 , then
E [ U n 4 ] → 3 ⟺ N ( ∙ ∙ ∙ ∙ , G n ) = o ( N ( ∙ − ∙ , G n ) 2 ) . (22) \mathbb{E}\left[U_n^4\right]\to3\quad\Longleftrightarrow\quad N\left(\begin{smallmatrix}\bullet&\bullet\\[-4pt]\bullet&\bullet\end{smallmatrix},G_n\right)=o\left(N(\bullet-\bullet,G_n)^2\right). \tag*{(22)} E [ U n 4 ] → 3 ⟺ N ( ∙ ∙ ∙ ∙ , G n ) = o ( N ( ∙ − ∙ , G n ) 2 ) . (22)
Proof. First, suppose p ∈ ( 0 , 1 ) ∖ { 1 2 } p\in(0,1)\setminus\left\{\frac{1}{2}\right\} p ∈ ( 0 , 1 ) ∖ { 2 1 } . In this case, recall (31) ,
σ n 2 ≍ N ( ∙ − ∙ , G n ) + N ( ∙ ∙ − ∙ − ∙ , G n ) . \sigma_n^2\asymp N(\bullet-\bullet,G_n)+N\left(\begin{smallmatrix}&\bullet&\\[-4pt]\bullet-\bullet-\bullet\end{smallmatrix},G_n\right). σ n 2 ≍ N ( ∙ − ∙ , G n ) + N ( ∙−∙−∙ ∙ , G n ) .
Referring to the order of the motif counts in Table 3 and by arguments as in Lemma 3.5, it follows that, for all H ∈ H 4 ∖ { ∙ − ∙ ⊔ ∙ − ∙ , ∙ − ∙ ⊔ ∙ ∙ − ∙ − ∙ , ∙ ∙ − ∙ − ∙ ⊔ ∙ ∙ − ∙ − ∙ , ∙ ∙ ∙ ∙ ∙ } H\in\mathcal{H}_4\setminus\{\bullet-\bullet\sqcup\bullet-\bullet,\ \bullet-\bullet\sqcup\begin{smallmatrix}&\bullet&\\[-4pt]\bullet-\bullet-\bullet\end{smallmatrix},\ \begin{smallmatrix}&\bullet&\\[-4pt]\bullet-\bullet-\bullet\end{smallmatrix}\sqcup\begin{smallmatrix}&\bullet&\\[-4pt]\bullet-\bullet-\bullet\end{smallmatrix},\ \begin{smallmatrix}\bullet&\bullet&\bullet\\[-4pt]\bullet&&\bullet\end{smallmatrix}\} H ∈ H 4 ∖ { ∙ − ∙ ⊔ ∙ − ∙ , ∙ − ∙ ⊔ ∙−∙−∙ ∙ , ∙−∙−∙ ∙ ⊔ ∙−∙−∙ ∙ , ∙ ∙ ∙ ∙ ∙ } ,
N ( H , G n ) = o ( ( N ( ∙ − ∙ , G n ) + N ( ∙ ∙ − ∙ − ∙ , G n ) ) 2 ) . (23) N(H,G_n)=o\left(\left(N(\bullet-\bullet,G_n)+N\left(\begin{smallmatrix}&\bullet&\\[-4pt]\bullet-\bullet-\bullet\end{smallmatrix},G_n\right)\right)^2\right). \tag*{(23)} N ( H , G n ) = o ( ( N ( ∙ − ∙ , G n ) + N ( ∙−∙−∙ ∙ , G n ) ) 2 ) . (23)
Also, observe the following identities:
( N ( ∙ − ∙ , G n ) 2 ) = N ( ∙ − ∙ ⊔ ∙ − ∙ , G n ) + N ( ∙ ∙ − ∙ − ∙ , G n ) , \binom{N(\bullet-\bullet,G_n)}{2}=N(\bullet-\bullet\sqcup\bullet-\bullet,G_n)+N\left(\begin{smallmatrix}&\bullet&\\[-4pt]\bullet-\bullet-\bullet\end{smallmatrix},G_n\right), ( 2 N ( ∙ − ∙ , G n ) ) = N ( ∙ − ∙ ⊔ ∙ − ∙ , G n ) + N ( ∙−∙−∙ ∙ , G n ) ,
where ⊔ \sqcup ⊔ denotes the disjoint union of two graphs,
N ( ∙ − ∙ , G n ) N ( ∙ ∙ − ∙ − ∙ , G n ) = N ( ∙ − ∙ ⊔ ∙ ∙ − ∙ − ∙ , G n ) + 2 N ( ∙ ∙ − ∙ − ∙ , G n ) + 3 N ( ∙ ∙ ∙ ∙ , G n ) + 2 N ( ∙ ∙ ∙ ∙ ∙ , G n ) + 3 N ( ∙ ∙ ∙ ∙ , G n ) , N(\bullet-\bullet,G_n)N\left(\begin{smallmatrix}&\bullet&\\[-4pt]\bullet-\bullet-\bullet\end{smallmatrix},G_n\right)=N\left(\bullet-\bullet\sqcup\begin{smallmatrix}&\bullet&\\[-4pt]\bullet-\bullet-\bullet\end{smallmatrix},G_n\right)+2N\left(\begin{smallmatrix}&\bullet&\\[-4pt]\bullet-\bullet-\bullet\end{smallmatrix},G_n\right)+3N\left(\begin{smallmatrix}\bullet&\bullet\\[-4pt]\bullet&\bullet\end{smallmatrix},G_n\right)+2N\left(\begin{smallmatrix}\bullet&\bullet&\bullet\\[-4pt]\bullet&&\bullet\end{smallmatrix},G_n\right)+3N\left(\begin{smallmatrix}\bullet&\bullet&\bullet\\[-4pt]&\bullet&\end{smallmatrix},G_n\right), N ( ∙ − ∙ , G n ) N ( ∙−∙−∙ ∙ , G n ) = N ( ∙ − ∙ ⊔ ∙−∙−∙ ∙ , G n ) + 2 N ( ∙−∙−∙ ∙ , G n ) + 3 N ( ∙ ∙ ∙ ∙ , G n ) + 2 N ( ∙ ∙ ∙ ∙ ∙ , G n ) + 3 N ( ∙ ∙ ∙ ∙ , G n ) ,
and
( N ( ∙ ∙ − ∙ − ∙ , G n ) 2 ) = N ( ∙ ∙ − ∙ − ∙ ⊔ ∙ ∙ − ∙ − ∙ , G n ) + 3 N ( ∙ ∙ ∙ ∙ ∙ , G n ) + o ( ( N ( ∙ − ∙ , G n ) + N ( ∙ ∙ − ∙ − ∙ , G n ) ) 2 ) , \binom{N\left(\begin{smallmatrix}&\bullet&\\[-4pt]\bullet-\bullet-\bullet\end{smallmatrix},G_n\right)}{2}=N\left(\begin{smallmatrix}&\bullet&\\[-4pt]\bullet-\bullet-\bullet\end{smallmatrix}\sqcup\begin{smallmatrix}&\bullet&\\[-4pt]\bullet-\bullet-\bullet\end{smallmatrix},G_n\right)+3N\left(\begin{smallmatrix}\bullet&\bullet&\bullet\\[-4pt]\bullet&&\bullet\end{smallmatrix},G_n\right)+o\left(\left(N(\bullet-\bullet,G_n)+N\left(\begin{smallmatrix}&\bullet&\\[-4pt]\bullet-\bullet-\bullet\end{smallmatrix},G_n\right)\right)^2\right), ( 2 N ( ∙−∙−∙ ∙ , G n ) ) = N ( ∙−∙−∙ ∙ ⊔ ∙−∙−∙ ∙ , G n ) + 3 N ( ∙ ∙ ∙ ∙ ∙ , G n ) + o ( ( N ( ∙ − ∙ , G n ) + N ( ∙−∙−∙ ∙ , G n ) ) 2 ) ,
where the o ( ⋅ ) o(\cdot) o ( ⋅ ) collects the motifs on at most four edges spanned by two overlapping wedges other than c r o s s \mathsf{cross} cross (namely p a w \mathsf{paw} paw , p a t h 3 \mathsf{path}_3 path 3 , t r i a n g l e \mathsf{triangle} triangle , s q u a r e \mathsf{square} square , t r i a n g l e - e d g e \mathsf{triangle\text{-}edge} triangle - edge , d i a m o n d \mathsf{diamond} diamond , p a t h 4 \mathsf{path}_4 path 4 ), all of which are o ( ( N ( e d g e , G n ) + N ( w e d g e , G n ) ) 2 ) o((N(\mathsf{edge},G_n)+N(\mathsf{wedge},G_n))^2) o (( N ( edge , G n ) + N ( wedge , G n ) ) 2 ) by (23) . Combining the above with (23) and by arguments as in Lemma 3.5, we obtain from (20) ,
E [ S n 4 ] = 3 ( 1 − u 2 4 N ( e d g e , G n ) + u ( 1 − u ) 2 N ( w e d g e , G n ) ) 2 + ( 3 u 2 ( 1 − u ) ( 3 u + 1 ) 2 − 3 ⋅ 3 u 2 ( 1 − u ) 2 2 ) N ( c r o s s , G n ) + o ( σ n 4 ) = 3 σ n 4 + 3 u 2 ( 1 − u ) ( 3 u − 1 ) N ( c r o s s , G n ) + o ( σ n 4 ) . (24) \begin{aligned}
\mathbb{E}\left[S_n^4\right]
&=3\left(\frac{1-u^2}{4}N(\mathsf{edge},G_n)+\frac{u(1-u)}{2}N(\mathsf{wedge},G_n)\right)^2\\
&\quad+\left(\frac{3u^2(1-u)(3u+1)}{2}-3\cdot\frac{3u^2(1-u)^2}{2}\right)N(\mathsf{cross},G_n)+o(\sigma_n^4)\\
&=3\sigma_n^4+3u^2(1-u)(3u-1)N(\mathsf{cross},G_n)+o(\sigma_n^4). \tag*{(24)}
\end{aligned} E [ S n 4 ] = 3 ( 4 1 − u 2 N ( edge , G n ) + 2 u ( 1 − u ) N ( wedge , G n ) ) 2 + ( 2 3 u 2 ( 1 − u ) ( 3 u + 1 ) − 3 ⋅ 2 3 u 2 ( 1 − u ) 2 ) N ( cross , G n ) + o ( σ n 4 ) = 3 σ n 4 + 3 u 2 ( 1 − u ) ( 3 u − 1 ) N ( cross , G n ) + o ( σ n 4 ) . (24)
Dividing by σ n 4 \sigma_n^4 σ n 4 and using σ n 4 ≍ ( N ( e d g e , G n ) + N ( w e d g e , G n ) ) 2 \sigma_n^4\asymp(N(\mathsf{edge},G_n)+N(\mathsf{wedge},G_n))^2 σ n 4 ≍ ( N ( edge , G n ) + N ( wedge , G n ) ) 2 (recall (31) ),
E [ S n 4 ] σ n 4 = 3 + 3 u 2 ( 1 − u ) ( 3 u − 1 ) N ( c r o s s , G n ) σ n 4 + o ( 1 ) . \frac{\mathbb{E}\left[S_n^4\right]}{\sigma_n^4}
=3+\frac{3u^2(1-u)(3u-1)N(\mathsf{cross},G_n)}{\sigma_n^4}+o(1). σ n 4 E [ S n 4 ] = 3 + σ n 4 3 u 2 ( 1 − u ) ( 3 u − 1 ) N ( cross , G n ) + o ( 1 ) .
Thus, provided u ≠ 1 3 u\ne\frac{1}{3} u = 3 1 , that is, p ≠ 1 ± 1 / 3 2 p\ne\frac{1\pm1/\sqrt{3}}{2} p = 2 1 ± 1/ 3 , the necessary and sufficient condition for E [ U n 4 ] → 3 \mathbb{E}\left[U_n^4\right]\to3 E [ U n 4 ] → 3 is N ( c r o s s , G n ) = o ( ( N ( e d g e , G n ) + N ( w e d g e , G n ) ) 2 ) N(\mathsf{cross},G_n)=o\left((N(\mathsf{edge},G_n)+N(\mathsf{wedge},G_n))^2\right) N ( cross , G n ) = o ( ( N ( edge , G n ) + N ( wedge , G n ) ) 2 ) . This proves (a).
Next, suppose p = 1 ± 1 / 3 2 p=\frac{1\pm1/\sqrt{3}}{2} p = 2 1 ± 1/ 3 . Then u = 1 3 u=\frac{1}{3} u = 3 1 , and the coefficient of N ( c r o s s , G n ) N(\mathsf{cross},G_n) N ( cross , G n ) in (24) vanishes and E [ U n 4 ] → 3 \mathbb{E}\left[U_n^4\right]\to3 E [ U n 4 ] → 3 with no additional condition. This proves (b).
Finally, suppose p = 1 2 p=\frac{1}{2} p = 2 1 , that is, u = 0 u=0 u = 0 . In this case, the result follows from [9 Proposition 1.11] . We will give a short proof here, for the sake of completeness. In particular, referring to Table 3 it follows that c H ( 4 ) ( u ) = 0 c_H^{(4)}(u)=0 c H ( 4 ) ( u ) = 0 , for all H ∈ H 4 ∖ { e d g e , w e d g e , e d g e ⊔ e d g e , s q u a r e } H\in\mathcal{H}_4\setminus\{\mathsf{edge},\mathsf{wedge},\mathsf{edge}\sqcup\mathsf{edge},\mathsf{square}\} H ∈ H 4 ∖ { edge , wedge , edge ⊔ edge , square } . Also, recall from (31) that σ n 4 = 1 4 N ( e d g e , G n ) \sigma_n^4=\frac{1}{4}N(\mathsf{edge},G_n) σ n 4 = 4 1 N ( edge , G n ) . Hence, by Table 3,
E [ S n 4 ] = 1 16 N ( e d g e , G n ) + 3 8 N ( w e d g e , G n ) + 3 8 ( ( N ( e d g e , G n ) 2 ) − N ( w e d g e , G n ) ) + 3 2 N ( s q u a r e , G n ) = 3 N ( e d g e , G n ) 2 − 2 N ( e d g e , G n ) 16 + 3 2 N ( s q u a r e , G n ) = 3 σ n 4 ( 1 + o ( 1 ) ) + 3 2 N ( s q u a r e , G n ) . \begin{aligned}
\mathbb{E}\left[S_n^4\right]
=\frac{1}{16}N(\mathsf{edge},G_n)+\frac{3}{8}N(\mathsf{wedge},G_n)+\frac{3}{8}\left(\binom{N(\mathsf{edge},G_n)}{2}-N(\mathsf{wedge},G_n)\right)+\frac{3}{2}N(\mathsf{square},G_n) \\
&=\frac{3N(\mathsf{edge},G_n)^2-2N(\mathsf{edge},G_n)}{16}+\frac{3}{2}N(\mathsf{square},G_n) \\
&=3\sigma_n^4(1+o(1))+\frac{3}{2}N(\mathsf{square},G_n).
\end{aligned} E [ S n 4 ] = 16 1 N ( edge , G n ) + 8 3 N ( wedge , G n ) + 8 3 ( ( 2 N ( edge , G n ) ) − N ( wedge , G n ) ) + 2 3 N ( square , G n ) = 16 3 N ( edge , G n ) 2 − 2 N ( edge , G n ) + 2 3 N ( square , G n ) = 3 σ n 4 ( 1 + o ( 1 )) + 2 3 N ( square , G n ) .
This implies that E [ U n 4 ] → 3 \mathbb{E}\left[U_n^4\right]\to3 E [ U n 4 ] → 3 if and only if N ( s q u a r e , G n ) = o ( N ( e d g e , G n ) 2 ) N(\mathsf{square},G_n)=o(N(\mathsf{edge},G_n)^2) N ( square , G n ) = o ( N ( edge , G n ) 2 ) . This proves (c). □
The Sparse p p p Regime: We have already stated the result for the sparse regime in § 1.1 . We prove Lemma 1.4.
Proof of Lemma 1.4: Recall that u n = ( 1 − 2 p n ) 2 u_n=(1-2p_n)^2 u n = ( 1 − 2 p n ) 2 and 1 − u n = 4 p n + O ( p n 2 ) 1-u_n=4p_n+O(p_n^2) 1 − u n = 4 p n + O ( p n 2 ) . We first consider the disconnected motifs in Table 3. Counting pairs of edges, pairs consisting of an edge and a wedge, and pairs of distinct wedges according to their union gives (as in the proof of Lemma 2.3),
( N ( e d g e , G n ) 2 ) = N ( e d g e ⊔ e d g e , G n ) + N ( w e d g e , G n ) , \binom{N(\mathsf{edge},G_n)}{2}=N(\mathsf{edge}\sqcup\mathsf{edge},G_n)+N(\mathsf{wedge},G_n), ( 2 N ( edge , G n ) ) = N ( edge ⊔ edge , G n ) + N ( wedge , G n ) ,
where ⊔ \sqcup ⊔ denotes the disjoint union of two graphs,
N ( e d g e , G n ) N ( w e d g e , G n ) = N ( e d g e ⊔ w e d g e , G n ) + 2 N ( w e d g e , G n ) + 3 N ( p a t h 3 , G n ) + 2 N ( w e d g e - w e d g e , G n ) + 3 N ( t r i a n g l e , G n ) , N(\mathsf{edge},G_n)N(\mathsf{wedge},G_n)=N(\mathsf{edge}\sqcup\mathsf{wedge},G_n)+2N(\mathsf{wedge},G_n)+3N(\mathsf{path}_3,G_n)+2N(\mathsf{wedge\text{-}wedge},G_n)+3N(\mathsf{triangle},G_n), N ( edge , G n ) N ( wedge , G n ) = N ( edge ⊔ wedge , G n ) + 2 N ( wedge , G n ) + 3 N ( path 3 , G n ) + 2 N ( wedge - wedge , G n ) + 3 N ( triangle , G n ) ,
and
( N ( w e d g e , G n ) 2 ) = N ( w e d g e ⊔ w e d g e , G n ) + 3 N ( p a t h 3 , G n ) + N ( w e d g e - w e d g e , G n ) + 3 N ( t r i a n g l e , G n ) \binom{N(\mathsf{wedge},G_n)}{2}
=N(\mathsf{wedge}\sqcup\mathsf{wedge},G_n)+3N(\mathsf{path}_3,G_n)+N(\mathsf{wedge\text{-}wedge},G_n)+3N(\mathsf{triangle},G_n) ( 2 N ( wedge , G n ) ) = N ( wedge ⊔ wedge , G n ) + 3 N ( path 3 , G n ) + N ( wedge - wedge , G n ) + 3 N ( triangle , G n )
+ 3 N ( c r o s s , G n ) + N ( p a t h 3 , G n ) + N ( p a t h 4 , G n ) + 2 N ( s q u a r e , G n ) + 2 N ( p a w , G n ) . \qquad+3N(\mathsf{cross},G_n)+N(\mathsf{path}_3,G_n)+N(\mathsf{path}_4,G_n)+2N(\mathsf{square},G_n)+2N(\mathsf{paw},G_n). + 3 N ( cross , G n ) + N ( path 3 , G n ) + N ( path 4 , G n ) + 2 N ( square , G n ) + 2 N ( paw , G n ) .
Substituting these identities into (20) and using (3) , the terms involving products of motif counts cancel exactly upon subtracting 3 σ n 4 3\sigma_n^4 3 σ n 4 . Expanding the remaining coefficients around u n = 1 u_n=1 u n = 1 gives
\mathbb{E}\left[S_n^4\right]-3\sigma_n^4=p_n(1+O(p_n))\left(2N(\bullet\!-\!\bullet,G_n)+14N(\mathord{\vee},G_n)+36N(\mathord{\Yup},G_n)+24N(\mathord{\times},G_n)\right)+o\left(p_n^2\left(N(\bullet\!-\!\bullet,G_n)+N(\mathord{\vee},G_n)\right)^2\right).
This follows by observing that the errors in the coefficients of the four displayed motifs are O ( p n 2 ) O(p_n^2) O ( p n 2 ) , which are absorbed into the factor 1 + O ( p n ) 1+O(p_n) 1 + O ( p n ) . Each remaining connected motif has coefficient O ( p n 2 ) O(p_n^2) O ( p n 2 ) and count o ( ( N ( ∙ − ∙ , G n ) + N ( ∨ , G n ) ) 2 ) o\left((N(\bullet\!-\!\bullet,G_n)+N(\mathord{\vee},G_n))^2\right) o ( ( N ( ∙ − ∙ , G n ) + N ( ∨ , G n ) ) 2 ) , by the combinatorial estimates used in the proof of Lemma 2.3.
Also, (3) gives,
σ n 2 = 2 p n ( N ( ∙ − ∙ , G n ) + N ( ∨ , G n ) ) ( 1 + O ( p n ) ) . \sigma_n^2=2p_n\left(N(\bullet\!-\!\bullet,G_n)+N(\mathord{\vee},G_n)\right)(1+O(p_n)). σ n 2 = 2 p n ( N ( ∙ − ∙ , G n ) + N ( ∨ , G n ) ) ( 1 + O ( p n )) .
Dividing the preceding fourth-moment expansion by σ n 4 \sigma_n^4 σ n 4 and using the non-negativity of the motif counts, we obtain
\mathbb{E}\left[S_n^4\right]\to3\quad\Longleftrightarrow\quad
\frac{N(\bullet\!-\!\bullet,G_n)+N(\mathord{\vee},G_n)+N(\mathord{\Yup},G_n)+N(\mathord{\times},G_n)}
{p_n\left(N(\bullet\!-\!\bullet,G_n)+N(\mathord{\vee},G_n)\right)^2}\to0. \tag*{(25)}
The condition on the right immediately implies (b). Conversely, separating vertices of degree 3 from those of degree at least 4 gives,
N(\mathord{\Yup},G_n)\leq N(\mathord{\vee},G_n)+4N(\mathord{\times},G_n).
Hence,
\frac{N(\bullet\!-\!\bullet,G_n)+N(\mathord{\vee},G_n)+N(\mathord{\Yup},G_n)+N(\mathord{\times},G_n)}
{p_n\left(N(\bullet\!-\!\bullet,G_n)+N(\mathord{\vee},G_n)\right)^2}
\leq
\frac{2}{p_n\left(N(\bullet\!-\!\bullet,G_n)+N(\mathord{\vee},G_n)\right)}
+
\frac{5N(\mathord{\times},G_n)}
{p_n\left(N(\bullet\!-\!\bullet,G_n)+N(\mathord{\vee},G_n)\right)^2}.
Under (b), both terms on the right tend to zero. The equivalence in (25) therefore proves the result. □
Wasserstein Convergence of U n U_n U n
In this section we establish general conditions for the asymptotic normality of U n U_n U n in Wasserstein metric. The results are key ingredients in the proofs of Theorem 1.1 and Theorem 1.2 . For this, we begin by recalling the definition of the Wasserstein distance.
Definition 3.1. Given two real-valued random variables X X X and Z Z Z with finite first moments, the Wasserstein distance between X X X and Z Z Z is defined as
Wass ( X , Z ) : = sup g ∈ L ∣ E [ g ( X ) ] − E [ g ( Z ) ] ∣ . \operatorname{Wass}(X,Z):=\sup_{g\in\mathcal{L}}\left|\mathbb{E}[g(X)]-\mathbb{E}[g(Z)]\right|. Wass ( X , Z ) := g ∈ L sup ∣ E [ g ( X )] − E [ g ( Z )] ∣ . where L \mathcal{L} L is the collection of 1 1 1 -Lipschitz functions from R → R \mathbb{R}\to\mathbb{R} R → R .
In the balanced case p = 1 2 p=\frac{1}{2} p = 2 1 , the edge indicators { 1 { X i = X j } } { i , j } ∈ E ( G n ) \{1\{X_i=X_j\}\}_{\{i,j\}\in E(G_n)} { 1 { X i = X j } } { i , j } ∈ E ( G n ) are pairwise independent, so the variance is determined solely by the number of edges (recall (3) ). In this case, using a quantitative martingale central limit theorem (CLT), Bhattacharya et al. [8 ] established asymptotic normality of U n U_n U n with convergence rates in the Kolmogorov distance. The following result, which shows convergence in the Wasserstein distance, is an easy consequence of the result. The proof also gives an explicit rate of convergence.
Proposition 3.2. Let { G n } n ∈ N \{G_n\}_{n\in\mathbb{N}} { G n } n ∈ N be a sequence of graphs with N ( ∙ − ∙ , G n ) → ∞ N(\bullet-\bullet,G_n)\to\infty N ( ∙ − ∙ , G n ) → ∞ and U n U_n U n be as defined in (2) . Suppose p = 1 2 p=\frac{1}{2} p = 2 1 and
N ( 4-star , G n ) = o ( N ( ∙ − ∙ , G n ) 2 ) . N(\text{4-star},G_n)=o\left(N(\bullet-\bullet,G_n)^2\right). N ( 4-star , G n ) = o ( N ( ∙ − ∙ , G n ) 2 ) . Then Wass ( U n , N ( 0 , 1 ) ) → 0 \operatorname{Wass}(U_n,N(0,1))\to0 Wass ( U n , N ( 0 , 1 )) → 0 .
Proof. Define
R n : = N ( ∙ − ∙ , G n ) − 1 2 + N ( 4-star , G n ) N ( ∙ − ∙ , G n ) 2 . \mathcal{R}_n:=N(\bullet-\bullet,G_n)^{-\frac{1}{2}}+\frac{N(\text{4-star},G_n)}{N(\bullet-\bullet,G_n)^2}. R n := N ( ∙ − ∙ , G n ) − 2 1 + N ( ∙ − ∙ , G n ) 2 N ( 4-star , G n ) .
By [8 Theorem 1.4] , applied with two colours,
sup x ∈ R ∣ P ( U n ≤ x ) − Φ ( x ) ∣ ≤ C R n 1 5 , (26) \sup_{x\in\mathbb{R}}\left|\mathbb{P}(U_n\le x)-\Phi(x)\right|\le C\mathcal{R}_n^{\frac{1}{5}}, \tag*{(26)} x ∈ R sup ∣ P ( U n ≤ x ) − Φ ( x ) ∣ ≤ C R n 5 1 , (26)
where C > 0 C>0 C > 0 is a constant and Φ \Phi Φ is the standard normal distribution function. We will now invoke the well known bound of the Wasserstein distance in terms of the Kolmogorov distance (See[10 Proposition A.1.2] ). That is, for two random variables X X X and Y Y Y with second moments bounded by B 2 B^2 B 2 ,
Wass ( X , Y ) ≤ 4 B sup t ∈ R ∣ P ( X ≤ t ) − P ( Y ≤ t ) ∣ . (27) \operatorname{Wass}(X,Y)\le4B\sqrt{\sup_{t\in\mathbb{R}}\left|\mathbb{P}(X\le t)-\mathbb{P}(Y\le t)\right|}. \tag*{(27)} Wass ( X , Y ) ≤ 4 B t ∈ R sup ∣ P ( X ≤ t ) − P ( Y ≤ t ) ∣ . (27)
Since E [ U n 2 ] = E [ Z 2 ] = 1 \mathbb{E}[U_n^2]=\mathbb{E}[Z^2]=1 E [ U n 2 ] = E [ Z 2 ] = 1 , where Z ∼ N ( 0 , 1 ) Z\sim N(0,1) Z ∼ N ( 0 , 1 ) , (26) and (27) gives,
Wass ( U n , Z ) ≤ 4 C R n 1 10 . (28) \operatorname{Wass}(U_n,Z)\le4\sqrt{C}\mathcal{R}_n^{\frac{1}{10}}. \tag*{(28)} Wass ( U n , Z ) ≤ 4 C R n 10 1 . (28)
Under the hypothesis, R n → 0 \mathcal{R}_n\to0 R n → 0 and the result follows.
The following result gives sufficient conditions for convergence of U n U_n U n to the standard normal distribution in the Wasserstein distance, both when p ≠ 1 2 p\ne\frac{1}{2} p = 2 1 is fixed and when p = p n → 0 p=p_n\to0 p = p n → 0 .
Theorem 3.3. Let { G n } n ∈ N \{G_n\}_{n\in\mathbb{N}} { G n } n ∈ N be a sequence of graphs with N ( ∙ − ∙ , G n ) → ∞ N(\bullet-\bullet,G_n)\to\infty N ( ∙ − ∙ , G n ) → ∞ and U n U_n U n be as defined in (2) .
Suppose p ∈ ( 0 , 1 ) ∖ { 1 2 } p\in(0,1)\setminus\left\{\frac{1}{2}\right\} p ∈ ( 0 , 1 ) ∖ { 2 1 } is fixed and
N ( 4-star , G n ) = o ( ( N ( ∙ − ∙ , G n ) + N ( 3-star , G n ) ) 2 ) . (29) N(\text{4-star},G_n)=o\left((N(\bullet-\bullet,G_n)+N(\text{3-star},G_n))^2\right). \tag*{(29)} N ( 4-star , G n ) = o ( ( N ( ∙ − ∙ , G n ) + N ( 3-star , G n ) ) 2 ) . (29) Then Wass ( U n , N ( 0 , 1 ) ) → 0 \operatorname{Wass}(U_n,N(0,1))\to0 Wass ( U n , N ( 0 , 1 )) → 0 .
Suppose p = p n → 0 p=p_n\to0 p = p n → 0 such that p n ( N ( ∙ − ∙ , G n ) + N ( 3-star , G n ) ) → ∞ p_n(N(\bullet-\bullet,G_n)+N(\text{3-star},G_n))\to\infty p n ( N ( ∙ − ∙ , G n ) + N ( 3-star , G n )) → ∞ and
N ( 3-star , G n ) = o ( p n 1 2 ( N ( ∙ − ∙ , G n ) + N ( 3-star , G n ) ) 3 2 ) . (30) N(\text{3-star},G_n)=o\left(p_n^{\frac{1}{2}}(N(\bullet-\bullet,G_n)+N(\text{3-star},G_n))^{\frac{3}{2}}\right). \tag*{(30)} N ( 3-star , G n ) = o ( p n 2 1 ( N ( ∙ − ∙ , G n ) + N ( 3-star , G n ) ) 2 3 ) . (30) Then Wass ( U n , N ( 0 , 1 ) ) → 0 \operatorname{Wass}(U_n,N(0,1))\to0 Wass ( U n , N ( 0 , 1 )) → 0 .
The proof of Theorem 3.3 uses Stein’s method with local dependence neighborhoods determined by edges sharing a vertex. The resulting Wasserstein bound consists of two terms: a cubic remainder and a fluctuation term involving local quadratic products. We present these preliminary preparation in § 3.1 . Controlling these terms requires third and fourth-order moment estimates for the summands of T n T_n T n , which translate into counts of graph motifs with up to four edges. For fixed p ≠ 1 2 p\ne\frac{1}{2} p = 2 1 , combinatorial comparisons between these counts show that the stated 4-star condition makes both error terms vanish (see § 3.2 ). The proof also provides an explicit bound for the Wasserstein distance in terms of motif counts (see Remark 3.6). The sparse regime is more delicate, as the moment estimates and variance normalization must retain their precise dependence on p n p_n p n . Here, we approximate this term by a degree-weighted sum of independent color indicators and control its mean absolute deviation by separating vertices of small and large degree. This reduces the required estimates to the stated 3-star condition, which, together with divergence of the variance, ensures that both Stein error terms vanish (see § 3.3 ).
We now make a remark on the equivalence of (29) and (30) when p ≠ 1 2 p \ne\frac{1}{2} p = 2 1 is fixed. Note that when p n = p ∈ ( 0 , 1 ) p_n=p\in(0,1) p n = p ∈ ( 0 , 1 ) is fixed, the condition p n ( N ( ∙ − ∙ , G n ) + N ( ∙ ∨ ∙ , G n ) ) → ∞ p_n\bigl(N(\bullet-\bullet,G_n)+N(\bullet\vee\bullet,G_n)\bigr)\to\infty p n ( N ( ∙ − ∙ , G n ) + N ( ∙ ∨ ∙ , G n ) ) → ∞ holds automatically from our standing assumption N ( ∙ − ∙ , G n ) → ∞ N(\bullet-\bullet,G_n)\to\infty N ( ∙ − ∙ , G n ) → ∞ . The remaining condition in (30) then reduces to
N ( ∙ ∙ , G n ) = o ( ( N ( ∙ − ∙ , G n ) + N ( ∙ ∨ ∙ , G n ) ) 3 2 ) . N(\bullet\!\bullet,G_n)=o\left(\left(N(\bullet-\bullet,G_n)+N(\bullet\vee\bullet,G_n)\right)^{\frac{3}{2}}\right). N ( ∙ ∙ , G n ) = o ( ( N ( ∙ − ∙ , G n ) + N ( ∙ ∨ ∙ , G n ) ) 2 3 ) .
For fixed p ≠ 1 2 p\ne\frac{1}{2} p = 2 1 , this is equivalent to the condition in (29) , by Lemma 3.4 and (56) . Hence, the 3-star and 4-star criteria when p ≠ 1 2 p\ne\frac{1}{2} p = 2 1 is fixed, are equivalent, showing that the two cases of Theorem 3.3 are consistent.
Preliminaries for Proof of Theorem 3.3
Let ( p n ) n ∈ N (p_n)_{n\in\mathbb{N}} ( p n ) n ∈ N be a sequence such that p n ∈ ( 0 , 1 ) p_n\in(0,1) p n ∈ ( 0 , 1 ) , for all n ∈ N n\in\mathbb{N} n ∈ N . (In the fixed p p p regime p n = p ∈ ( 0 , 1 ) ∖ { 1 2 } p_n=p\in(0,1)\setminus\{\frac{1}{2}\} p n = p ∈ ( 0 , 1 ) ∖ { 2 1 } , for all n ∈ N n\in\mathbb{N} n ∈ N , and in the sparse regime p n → 0 p_n\to0 p n → 0 , as n → ∞ n\to\infty n → ∞ .) For each e = { i , j } ∈ E ( G n ) e=\{i,j\}\in E(G_n) e = { i , j } ∈ E ( G n ) , recall Y e Y_e Y e from (12) . Then the normalised statistic U n U_n U n (recall (2) ) can be rewritten as
U n = 1 σ n ∑ e ∈ E ( G n ) Y e . U_n=\frac{1}{\sigma_n}\sum_{e\in E(G_n)}Y_e. U n = σ n 1 e ∈ E ( G n ) ∑ Y e .
where σ n 2 : = Var [ T n ] \sigma_n^2:=\operatorname{Var}[T_n] σ n 2 := Var [ T n ] . Note that E [ Y e ] = 0 \mathbb{E}[Y_e]=0 E [ Y e ] = 0 and Var [ Y e ] = q n ( 1 − q n ) = 1 4 ( 1 − u n 2 ) \operatorname{Var}[Y_e]=q_n(1-q_n)=\frac{1}{4}(1-u_n^2) Var [ Y e ] = q n ( 1 − q n ) = 4 1 ( 1 − u n 2 ) . Further, Y e Y_e Y e and Y e ′ Y_{e'} Y e ′ are independent whenever e ∩ e ′ = ∅ e\cap e'=\varnothing e ∩ e ′ = ∅ , and for edges e , e ′ e,e' e , e ′ sharing exactly one vertex
Cov [ Y e , Y e ′ ] = p n 3 + ( 1 − p n ) 3 − q n 2 = 1 4 ( u n − u n 2 ) . \operatorname{Cov}[Y_e,Y_{e'}]=p_n^3+(1-p_n)^3-q_n^2=\frac{1}{4}(u_n-u_n^2). Cov [ Y e , Y e ′ ] = p n 3 + ( 1 − p n ) 3 − q n 2 = 4 1 ( u n − u n 2 ) .
Hence,
σ n 2 = ∑ e 1 , e 2 ∈ E ( G n ) Cov [ Y e 1 , Y e 2 ] = 1 − u n 2 4 N ( ∙ − ∙ , G n ) + u n − u n 2 2 N ( ∙ ∨ ∙ , G n ) , (31) \sigma_n^2=\sum_{e_1,e_2\in E(G_n)}\operatorname{Cov}[Y_{e_1},Y_{e_2}]=\frac{1-u_n^2}{4}N(\bullet-\bullet,G_n)+\frac{u_n-u_n^2}{2}N(\bullet\vee\bullet,G_n), \tag*{(31)} σ n 2 = e 1 , e 2 ∈ E ( G n ) ∑ Cov [ Y e 1 , Y e 2 ] = 4 1 − u n 2 N ( ∙ − ∙ , G n ) + 2 u n − u n 2 N ( ∙ ∨ ∙ , G n ) , (31)
where u n = ( 2 p n − 1 ) 2 u_n=(2p_n-1)^2 u n = ( 2 p n − 1 ) 2 . (This is the same as the expression in (3) with u u u replaced by u n u_n u n , to highlight the dependence on n n n in the sparse regime). We now use the Wasserstein distance bound obtained from Stein’s operator (see, for example, Ross [47, Theorem 3.1]):
d W a s s ( U n , N ( 0 , 1 ) ) ≤ sup f ∈ F ∣ E [ U n f ( U n ) ] − E [ f ′ ( U n ) ] ∣ , (32) d_{\mathrm{Wass}}(U_n,N(0,1))\leq\sup_{f\in\mathcal{F}}\left|\mathbb{E}[U_nf(U_n)]-\mathbb{E}[f'(U_n)]\right|, \tag*{(32)} d Wass ( U n , N ( 0 , 1 )) ≤ f ∈ F sup ∣ E [ U n f ( U n )] − E [ f ′ ( U n )] ∣ , (32)
where F = { f : R → R such that ∥ f ∥ ∞ ≤ 1 , ∥ f ′ ∥ ∞ ≤ 2 π , ∥ f ′ ′ ∥ ∞ ≤ 2 } \mathcal{F}=\{f:\mathbb{R}\to\mathbb{R}\text{ such that }\|f\|_\infty\leq1,\ \|f'\|_\infty\leq\sqrt{\frac{2}{\pi}},\ \|f''\|_\infty\leq2\} F = { f : R → R such that ∥ f ∥ ∞ ≤ 1 , ∥ f ′ ∥ ∞ ≤ π 2 , ∥ f ′′ ∥ ∞ ≤ 2 } . Fix f ∈ F f\in\mathcal{F} f ∈ F and for each edge e = { i , j } ∈ E ( G n ) e=\{i,j\}\in E(G_n) e = { i , j } ∈ E ( G n ) , set
U e = 1 σ n ∑ e ′ ∈ E ( G n ) e ′ ∩ e = ∅ Y e ′ . U_e=\frac{1}{\sigma_n}\sum_{\substack{e'\in E(G_n)\\e'\cap e=\varnothing}}Y_{e'}. U e = σ n 1 e ′ ∈ E ( G n ) e ′ ∩ e = ∅ ∑ Y e ′ .
Then
E [ U n f ( U n ) ] − E [ f ′ ( U n ) ] = 1 σ n ∑ e ∈ E ( G n ) E [ Y e f ( U n ) ] − E [ f ′ ( U n ) ] = 1 σ n ∑ e ∈ E ( G n ) E [ Y e ( f ( U n ) − f ( U e ) ) ] − E [ f ′ ( U n ) ] , (33) \begin{aligned}
\mathbb{E}[U_nf(U_n)]-\mathbb{E}[f'(U_n)]
=\frac{1}{\sigma_n}\sum_{e\in E(G_n)}\mathbb{E}[Y_ef(U_n)]-\mathbb{E}[f'(U_n)] \\
&=\frac{1}{\sigma_n}\sum_{e\in E(G_n)}\mathbb{E}\left[Y_e\bigl(f(U_n)-f(U_e)\bigr)\right]-\mathbb{E}[f'(U_n)], \tag*{(33)}
\end{aligned} E [ U n f ( U n )] − E [ f ′ ( U n )] = σ n 1 e ∈ E ( G n ) ∑ E [ Y e f ( U n )] − E [ f ′ ( U n )] = σ n 1 e ∈ E ( G n ) ∑ E [ Y e ( f ( U n ) − f ( U e ) ) ] − E [ f ′ ( U n )] , (33)
where (33) follows from the fact
E [ Y e f ( U e ) ] = E [ Y e ] E [ f ( U e ) ] = 0 , \mathbb{E}[Y_ef(U_e)]=\mathbb{E}[Y_e]\mathbb{E}[f(U_e)]=0, E [ Y e f ( U e )] = E [ Y e ] E [ f ( U e )] = 0 ,
as Y e Y_e Y e is independent of U e U_e U e . Adding and subtracting 1 σ n ∑ e ∈ E ( G n ) E [ Y e ( U n − U e ) f ′ ( U n ) ] \frac{1}{\sigma_n}\sum_{e\in E(G_n)}\mathbb{E}[Y_e(U_n-U_e)f'(U_n)] σ n 1 ∑ e ∈ E ( G n ) E [ Y e ( U n − U e ) f ′ ( U n )] on the right-hand side of (33) , we have
E [ U n f ( U n ) ] − E [ f ′ ( U n ) ] = 1 σ n ∑ e ∈ E ( G n ) E [ Y e ( f ( U n ) − f ( U e ) − ( U n − U e ) f ′ ( U n ) ) ] + 1 σ n ∑ e ∈ E ( G n ) E [ Y e ( U n − U e ) f ′ ( U n ) ] − E [ f ′ ( U n ) ] . (34) \mathbb{E}[U_nf(U_n)]-\mathbb{E}[f'(U_n)]
=\frac{1}{\sigma_n}\sum_{e\in E(G_n)}\mathbb{E}\left[Y_e\left(f(U_n)-f(U_e)-(U_n-U_e)f'(U_n)\right)\right]
+\frac{1}{\sigma_n}\sum_{e\in E(G_n)}\mathbb{E}\left[Y_e(U_n-U_e)f'(U_n)\right]-\mathbb{E}[f'(U_n)]. \tag*{(34)} E [ U n f ( U n )] − E [ f ′ ( U n )] = σ n 1 e ∈ E ( G n ) ∑ E [ Y e ( f ( U n ) − f ( U e ) − ( U n − U e ) f ′ ( U n ) ) ] + σ n 1 e ∈ E ( G n ) ∑ E [ Y e ( U n − U e ) f ′ ( U n ) ] − E [ f ′ ( U n )] . (34)
Define
V n : = ∑ e 1 , e 2 ∈ E ( G n ) e 1 ∩ e 2 ≠ ∅ Y e 1 Y e 2 = σ n ∑ e ∈ E ( G n ) Y e ( U n − U e ) . (35) V_n:=\sum_{\substack{e_1,e_2\in E(G_n)\\e_1\cap e_2\ne\varnothing}}Y_{e_1}Y_{e_2}=\sigma_n\sum_{e\in E(G_n)}Y_e(U_n-U_e). \tag*{(35)} V n := e 1 , e 2 ∈ E ( G n ) e 1 ∩ e 2 = ∅ ∑ Y e 1 Y e 2 = σ n e ∈ E ( G n ) ∑ Y e ( U n − U e ) . (35)
Using the fact that E [ V n ] = σ n 2 \mathbb{E}[V_n]=\sigma_n^2 E [ V n ] = σ n 2 , we may rewrite (34) as
E [ U n f ( U n ) ] − E [ f ′ ( U n ) ] = 1 σ n ∑ e ∈ E ( G n ) E [ Y e ( f ( U n ) − f ( U e ) − ( U n − U e ) f ′ ( U n ) ) ] + E [ f ′ ( U n ) ( V n − E [ V n ] σ n 2 ) ] = : A 1 + A 2 . (36) \begin{aligned}
\mathbb{E}[U_nf(U_n)]-\mathbb{E}[f'(U_n)]
&=\frac{1}{\sigma_n}\sum_{e\in E(G_n)}\mathbb{E}\left[Y_e\left(f(U_n)-f(U_e)-(U_n-U_e)f'(U_n)\right)\right]\\
&\quad+\mathbb{E}\left[f'(U_n)\left(\frac{V_n-\mathbb{E}[V_n]}{\sigma_n^2}\right)\right]\\
&=:A_1+A_2. \tag*{(36)}
\end{aligned} E [ U n f ( U n )] − E [ f ′ ( U n )] = σ n 1 e ∈ E ( G n ) ∑ E [ Y e ( f ( U n ) − f ( U e ) − ( U n − U e ) f ′ ( U n ) ) ] + E [ f ′ ( U n ) ( σ n 2 V n − E [ V n ] ) ] =: A 1 + A 2 . (36)
where
A 1 : = 1 σ n ∑ e ∈ E ( G n ) E [ Y e ( f ( U n ) − f ( U e ) − ( U n − U e ) f ′ ( U n ) ) ] A_1:=\frac{1}{\sigma_n}\sum_{e\in E(G_n)}\mathbb{E}\left[Y_e\left(f(U_n)-f(U_e)-(U_n-U_e)f'(U_n)\right)\right] A 1 := σ n 1 e ∈ E ( G n ) ∑ E [ Y e ( f ( U n ) − f ( U e ) − ( U n − U e ) f ′ ( U n ) ) ]
and
A 2 : = E [ f ′ ( U n ) ( V n − E [ V n ] σ n 2 ) ] . A_2:=\mathbb{E}\left[f'(U_n)\left(\frac{V_n-\mathbb{E}[V_n]}{\sigma_n^2}\right)\right]. A 2 := E [ f ′ ( U n ) ( σ n 2 V n − E [ V n ] ) ] .
Hence, by (32) , to prove Theorem 3.3 it suffices to show that both A 1 A_1 A 1 and A 2 A_2 A 2 are o ( 1 ) o(1) o ( 1 ) , uniformly in f ∈ F f\in\mathcal{F} f ∈ F , in the respective settings.
Proof of Theorem 3.3 (i)
We will need two lemmas. One which bounds the number of 3-stars in a graph in terms of the numbers of 2-stars and 4-stars, and a second which computes the variance of V n V_n V n as defined in (35) .
Lemma 3.4. For every finite simple graph G n G_n G n ,
N ( K 1 , 3 , G n ) ≤ N ( K 1 , 2 , G n ) + 8 3 N ( K 1 , 2 , G n ) N ( K 1 , 4 , G n ) . (37) N(K_{1,3},G_n)\le N(K_{1,2},G_n)+\sqrt{\frac{8}{3}N(K_{1,2},G_n)N(K_{1,4},G_n)}. \tag*{(37)} N ( K 1 , 3 , G n ) ≤ N ( K 1 , 2 , G n ) + 3 8 N ( K 1 , 2 , G n ) N ( K 1 , 4 , G n ) . (37)
Lemma 3.5. Let { G n } n ∈ N \{G_n\}_{n\in\mathbb{N}} { G n } n ∈ N be a sequence of graphs with N ( K 1 , 2 , G n ) → ∞ N(K_{1,2},G_n)\to\infty N ( K 1 , 2 , G n ) → ∞ . Let V n V_n V n be as defined in (35) and p ∈ ( 0 , 1 ) ∖ { 1 2 } p\in(0,1)\setminus\left\{\frac{1}{2}\right\} p ∈ ( 0 , 1 ) ∖ { 2 1 } .
Var [ V n ] = c K 1 , 4 N ( K 1 , 4 , G n ) + o ( ( N ( K 1 , 2 , G n ) + N ( K 1 , 3 , G n ) ) 2 ) , (38) \operatorname{Var}[V_n]=c_{K_{1,4}}N(K_{1,4},G_n)+o\left((N(K_{1,2},G_n)+N(K_{1,3},G_n))^2\right), \tag*{(38)} Var [ V n ] = c K 1 , 4 N ( K 1 , 4 , G n ) + o ( ( N ( K 1 , 2 , G n ) + N ( K 1 , 3 , G n ) ) 2 ) , (38) where c K 1 , 4 = 6 u 3 ( 1 − u ) c_{K_{1,4}}=6u^3(1-u) c K 1 , 4 = 6 u 3 ( 1 − u ) and u = ( 2 p − 1 ) 2 u=(2p-1)^2 u = ( 2 p − 1 ) 2 .
We shall defer the proof of this lemma to the end of this section and prove Theorem 3.3(i) assuming the above result.
Proof of Theorem 3.3(i): We now proceed with the proof of Theorem 3.3 (i). We first estimate A 2 A_2 A 2 . By (36) , the Cauchy–Schwarz inequality, the fact that f ∈ F f \in\mathcal{F} f ∈ F , and Lemma 3.5,
∣ A 2 ∣ ≤ ∥ f ′ ′ ∥ ∞ σ n 2 Var [ V n ] ≤ 1 σ n 2 O ( N ( ⋆ 3 , G n ) ) + o ( ( N ( ∙ − ∙ , G n ) + N ( ∨ , G n ) ) 2 ) . (39) \lvert A_2\rvert\le\frac{\lVert f''\rVert_\infty}{\sigma_n^2}\sqrt{\operatorname{Var}[V_n]} \le\frac{1}{\sigma_n^2}\sqrt{O(N(\star_3,G_n))+o\left((N(\bullet\!-\!\bullet,G_n)+N(\vee,G_n))^2\right)}. \tag*{(39)} ∣ A 2 ∣ ≤ σ n 2 ∥ f ′′ ∥ ∞ Var [ V n ] ≤ σ n 2 1 O ( N ( ⋆ 3 , G n )) + o ( ( N ( ∙ − ∙ , G n ) + N ( ∨ , G n ) ) 2 ) . (39)
In this case, recalling (31) ,
σ n 2 ≍ N ( ∙ − ∙ , G n ) + N ( ∨ , G n ) . (40) \sigma_n^2 \asymp N(\bullet\!-\!\bullet,G_n)+N(\vee,G_n). \tag*{(40)} σ n 2 ≍ N ( ∙ − ∙ , G n ) + N ( ∨ , G n ) . (40)
Using (40) and the assumption N ( ⋆ 3 , G n ) = o ( ( N ( ∙ − ∙ , G n ) + N ( ∨ , G n ) ) 2 ) N(\star_3,G_n)=o\left((N(\bullet\!-\!\bullet,G_n)+N(\vee,G_n))^2\right) N ( ⋆ 3 , G n ) = o ( ( N ( ∙ − ∙ , G n ) + N ( ∨ , G n ) ) 2 ) in (39) we have ∣ A 2 ∣ = o ( 1 ) \lvert A_2\rvert=o(1) ∣ A 2 ∣ = o ( 1 ) .
Next, we consider A 1 A_1 A 1 . By a Taylor expansion,
∣ A 1 ∣ ≤ 1 σ n ⋅ ∥ f ′ ′ ∥ ∞ 2 ∑ e ∈ E ( G n ) E [ ∣ Y e ∣ ⋅ ( U n − U e ) 2 ] ≤ 1 σ n ∑ e ∈ E ( G n ) E [ ∣ Y e ∣ ⋅ ( U n − U e ) 2 ] . (41) \lvert A_1\rvert\le\frac{1}{\sigma_n}\cdot\frac{\lVert f''\rVert_\infty}{2}\sum_{e\in E(G_n)}\mathbb{E}\left[\lvert Y_e\rvert\cdot(U_n-U_e)^2\right]\le\frac{1}{\sigma_n}\sum_{e\in E(G_n)}\mathbb{E}\left[\lvert Y_e\rvert\cdot(U_n-U_e)^2\right]. \tag*{(41)} ∣ A 1 ∣ ≤ σ n 1 ⋅ 2 ∥ f ′′ ∥ ∞ e ∈ E ( G n ) ∑ E [ ∣ Y e ∣ ⋅ ( U n − U e ) 2 ] ≤ σ n 1 e ∈ E ( G n ) ∑ E [ ∣ Y e ∣ ⋅ ( U n − U e ) 2 ] . (41)
Recall that d G n ( i ) d_{G_n}(i) d G n ( i ) is the degree of the i i i -th vertex of G n G_n G n , for 1 ≤ i ≤ n 1\le i\le n 1 ≤ i ≤ n . Then the sum in the right-hand side of (41) can be expressed as:
∑ e ∈ E ( G n ) E [ ∣ Y e ∣ ⋅ ( U n − U e ) 2 ] = 1 σ n 2 ∑ e ∈ E ( G n ) E [ ∣ Y e ∣ ⋅ ( ∑ e ′ ∈ E ( G n ) e ′ ∩ e ≠ ∅ Y e ′ ) 2 ] ≤ 1 σ n 2 ∑ e ∈ E ( G n ) ∣ { e ′ ∈ E ( G n ) : e ′ ∩ e ≠ ∅ } ∣ 2 ≤ 1 σ n 2 ∑ { i , j } ∈ E ( G n ) ( d G n ( i ) + d G n ( j ) ) 2 ≤ 2 σ n 2 ∑ { i , j } ∈ E ( G n ) ( d G n ( i ) 2 + d G n ( j ) 2 ) = 2 σ n 2 ∑ i = 1 n d G n ( i ) 3 . \begin{aligned}
\sum_{e\in E(G_n)}\mathbb{E}\left[\lvert Y_e\rvert\cdot(U_n-U_e)^2\right]
&=\frac{1}{\sigma_n^2}\sum_{e\in E(G_n)}\mathbb{E}\left[\lvert Y_e\rvert\cdot\left(\sum_{\substack{e'\in E(G_n)\\e'\cap e\ne\varnothing}}Y_{e'}\right)^2\right]\\
&\le\frac{1}{\sigma_n^2}\sum_{e\in E(G_n)}\left\lvert\{e'\in E(G_n):e'\cap e\ne\varnothing\}\right\rvert^2\\
&\le\frac{1}{\sigma_n^2}\sum_{\{i,j\}\in E(G_n)}\left(d_{G_n}(i)+d_{G_n}(j)\right)^2\\
&\le\frac{2}{\sigma_n^2}\sum_{\{i,j\}\in E(G_n)}\left(d_{G_n}(i)^2+d_{G_n}(j)^2\right)=\frac{2}{\sigma_n^2}\sum_{i=1}^{n}d_{G_n}(i)^3.
\end{aligned} e ∈ E ( G n ) ∑ E [ ∣ Y e ∣ ⋅ ( U n − U e ) 2 ] = σ n 2 1 e ∈ E ( G n ) ∑ E ∣ Y e ∣ ⋅ e ′ ∈ E ( G n ) e ′ ∩ e = ∅ ∑ Y e ′ 2 ≤ σ n 2 1 e ∈ E ( G n ) ∑ ∣ { e ′ ∈ E ( G n ) : e ′ ∩ e = ∅ } ∣ 2 ≤ σ n 2 1 { i , j } ∈ E ( G n ) ∑ ( d G n ( i ) + d G n ( j ) ) 2 ≤ σ n 2 2 { i , j } ∈ E ( G n ) ∑ ( d G n ( i ) 2 + d G n ( j ) 2 ) = σ n 2 2 i = 1 ∑ n d G n ( i ) 3 .
where the first inequality uses ∣ Y e ∣ ≤ 1 \lvert Y_e\rvert\le1 ∣ Y e ∣ ≤ 1 , for all e e e . Substituting this back into (41) gives,
∣ A 1 ∣ ≤ 2 σ n 3 ∑ i = 1 n d G n ( i ) 3 = 2 σ n 3 ∑ i = 1 n ( 6 ( d G n ( i ) 3 ) + 6 ( d G n ( i ) 2 ) + d G n ( i ) ) = 1 σ n 3 ( 12 N ( ⋆ 3 , G n ) + 12 N ( ∨ , G n ) + 4 N ( ∙ − ∙ , G n ) ) ≤ 1 σ n 3 ( 24 N ( ∨ , G n ) + 12 8 3 N ( ∨ , G n ) N ( ⋆ 3 , G n ) + 4 N ( ∙ − ∙ , G n ) ) , \begin{aligned}
\lvert A_1\rvert&\le\frac{2}{\sigma_n^3}\sum_{i=1}^{n}d_{G_n}(i)^3\\
&=\frac{2}{\sigma_n^3}\sum_{i=1}^{n}\left(6\binom{d_{G_n}(i)}{3}+6\binom{d_{G_n}(i)}{2}+d_{G_n}(i)\right)\\
&=\frac{1}{\sigma_n^3}\left(12N(\star_3,G_n)+12N(\vee,G_n)+4N(\bullet\!-\!\bullet,G_n)\right)\\
&\le\frac{1}{\sigma_n^3}\left(24N(\vee,G_n)+12\sqrt{\frac{8}{3}N(\vee,G_n)N(\star_3,G_n)}+4N(\bullet\!-\!\bullet,G_n)\right),
\end{aligned} ∣ A 1 ∣ ≤ σ n 3 2 i = 1 ∑ n d G n ( i ) 3 = σ n 3 2 i = 1 ∑ n ( 6 ( 3 d G n ( i ) ) + 6 ( 2 d G n ( i ) ) + d G n ( i ) ) = σ n 3 1 ( 12 N ( ⋆ 3 , G n ) + 12 N ( ∨ , G n ) + 4 N ( ∙ − ∙ , G n ) ) ≤ σ n 3 1 ( 24 N ( ∨ , G n ) + 12 3 8 N ( ∨ , G n ) N ( ⋆ 3 , G n ) + 4 N ( ∙ − ∙ , G n ) ) ,
where the last inequality follows from Lemma 3.4. Using (40) and the assumption N ( ⋆ 3 , G n ) = o ( ( N ( ∙ − ∙ , G n ) + N ( ∨ , G n ) ) 2 ) N(\star_3,G_n)=o\left((N(\bullet\!-\!\bullet,G_n)+N(\vee,G_n))^2\right) N ( ⋆ 3 , G n ) = o ( ( N ( ∙ − ∙ , G n ) + N ( ∨ , G n ) ) 2 ) gives ∣ A 1 ∣ = o ( 1 ) \lvert A_1\rvert=o(1) ∣ A 1 ∣ = o ( 1 ) . This completes the proof of Theorem 3.3 (i). □
Remark 3.6. By tracking the constants in the proof of Theorem 3.3 (i) and replacing the asymptotic estimate in Lemma 3.5 with a quantitative bound on Var [ V n ] \operatorname{Var}[V_n] Var [ V n ] , one obtains the following rate: For every fixed p ∈ ( 0 , 1 ) ∖ { 1 2 } p\in(0,1)\setminus\left\{\frac{1}{2}\right\} p ∈ ( 0 , 1 ) ∖ { 2 1 } ,
Wass ( U n , N ( 0 , 1 ) ) ≤ C p ( 1 N ( ∙ − ∙ , G n ) + N ( ∨ , G n ) + N ( ⋆ 3 , G n ) N ( ∙ − ∙ , G n ) + N ( ∨ , G n ) ) . (42) \operatorname{Wass}(U_n,N(0,1))\le C_p\left(\frac{1}{\sqrt{N(\bullet\!-\!\bullet,G_n)+N(\vee,G_n)}}+\frac{\sqrt{N(\star_3,G_n)}}{N(\bullet\!-\!\bullet,G_n)+N(\vee,G_n)}\right). \tag*{(42)} Wass ( U n , N ( 0 , 1 )) ≤ C p ( N ( ∙ − ∙ , G n ) + N ( ∨ , G n ) 1 + N ( ∙ − ∙ , G n ) + N ( ∨ , G n ) N ( ⋆ 3 , G n ) ) . (42)
for some constant C p > 0 C_p > 0 C p > 0 depending only on p p p . This bound holds for every n n n such that G n G_n G n has at least one edge.
To complete the proof of Theorem 3.3 (i) we prove Lemma 3.5 and Lemma 3.4.
Proof of Lemma 3.5: Recall the definition of V n V_n V n from (35) . Expanding the variance of V n V_n V n gives
Var [ V n ] = ∑ e 1 , e 2 ∈ E ( G n ) e 1 ∩ e 2 ≠ ∅ ∑ e 3 , e 4 ∈ E ( G n ) e 3 ∩ e 4 ≠ ∅ Cov [ Y e 1 Y e 2 , Y e 3 Y e 4 ] . (43) \operatorname{Var}[V_n] =
\sum_{\substack{e_1,e_2\in E(G_n)\\ e_1\cap e_2\ne\varnothing}}
\sum_{\substack{e_3,e_4\in E(G_n)\\ e_3\cap e_4\ne\varnothing}}
\operatorname{Cov}[Y_{e_1}Y_{e_2},Y_{e_3}Y_{e_4}]. \tag*{(43)} Var [ V n ] = e 1 , e 2 ∈ E ( G n ) e 1 ∩ e 2 = ∅ ∑ e 3 , e 4 ∈ E ( G n ) e 3 ∩ e 4 = ∅ ∑ Cov [ Y e 1 Y e 2 , Y e 3 Y e 4 ] . (43)
Note that the covariance vanishes unless ( e 1 ∪ e 2 ) ∩ ( e 3 ∪ e 4 ) ≠ ∅ (e_1\cup e_2)\cap(e_3\cup e_4)\ne\varnothing ( e 1 ∪ e 2 ) ∩ ( e 3 ∪ e 4 ) = ∅ , hence the sum in (43) runs over ordered 4-tuples of edges ( e 1 , e 2 , e 3 , e 4 ) (e_1,e_2,e_3,e_4) ( e 1 , e 2 , e 3 , e 4 ) such that e 1 ∩ e 2 ≠ ∅ e_1\cap e_2\ne\varnothing e 1 ∩ e 2 = ∅ , e 3 ∩ e 4 ≠ ∅ e_3\cap e_4\ne\varnothing e 3 ∩ e 4 = ∅ and ( e 1 ∪ e 2 ) ∩ ( e 3 ∪ e 4 ) ≠ ∅ (e_1\cup e_2)\cap(e_3\cup e_4)\ne\varnothing ( e 1 ∪ e 2 ) ∩ ( e 3 ∪ e 4 ) = ∅ . We split the summation in (43) according to the isomorphism type of the subgraph e 1 ∪ e 2 ∪ e 3 ∪ e 4 e_1\cup e_2\cup e_3\cup e_4 e 1 ∪ e 2 ∪ e 3 ∪ e 4 . The possible subgraphs form the following collection:
H = { K 2 , P 3 , K 3 , P 4 , K 1 , 3 , P 5 , K 1 , 4 , T 3 , 2 , C 4 , paw } . \mathcal{H}=\{K_2,P_3,K_3,P_4,K_{1,3},P_5,K_{1,4},T_{3,2},C_4,\operatorname{paw}\}. H = { K 2 , P 3 , K 3 , P 4 , K 1 , 3 , P 5 , K 1 , 4 , T 3 , 2 , C 4 , paw } .
For any H ∈ H H\in\mathcal{H} H ∈ H , the total contribution of all tuples spanning a given copy of H H H depends only on H H H and u n u_n u n . We denote it by c H = c H ( u n ) c_H=c_H(u_n) c H = c H ( u n ) . Thus, we can write
Var [ V n ] = ∑ H ∈ H c H N ( H , G n ) . (44) \operatorname{Var}[V_n]=\sum_{H\in\mathcal{H}}c_HN(H,G_n). \tag*{(44)} Var [ V n ] = H ∈ H ∑ c H N ( H , G n ) . (44)
To evaluate the coefficients c H c_H c H , one computes Cov [ Y e 1 Y e 2 , Y e 3 Y e 4 ] \operatorname{Cov}[Y_{e_1}Y_{e_2},Y_{e_3}Y_{e_4}] Cov [ Y e 1 Y e 2 , Y e 3 Y e 4 ] for all tuples of edges ( e 1 , e 2 , e 3 , e 4 ) (e_1,e_2,e_3,e_4) ( e 1 , e 2 , e 3 , e 4 ) whose union is a fixed copy of H H H . For instance, when H = K 2 H=K_2 H = K 2 , all four edges coincide, and c K 2 = Var Y e 2 c_{K_2}=\operatorname{Var}Y_e^2 c K 2 = Var Y e 2 . The remaining coefficients are computed similarly; the results, together with the asymptotic order of the corresponding subgraph counts (obtained from ∑ i = 1 n d G n ( i ) = 2 N ( K 2 , G n ) \sum_{i=1}^n d_{G_n}(i)=2N(K_2,G_n) ∑ i = 1 n d G n ( i ) = 2 N ( K 2 , G n ) , ∑ i = 1 n ( d G n ( i ) 2 ) = N ( P 3 , G n ) \sum_{i=1}^n\binom{d_{G_n}(i)}{2}=N(P_3,G_n) ∑ i = 1 n ( 2 d G n ( i ) ) = N ( P 3 , G n ) and elementary bounds), are displayed in Table 1 in Appendix § 8 . The following stronger bound will be required in the proof:
Table 1: Coefficients c H ( u n ) c_H(u_n) c H ( u n ) in (44) and the order of the motif counts.
Table 1.
N ( P 5 , G n ) + N ( K 1 , 4 , G n ) = O ( N ( P 3 , G n ) 7 4 ) . (45) N(P_5,G_n)+N(K_{1,4},G_n)=O\left(N(P_3,G_n)^{\frac{7}{4}}\right). \tag*{(45)} N ( P 5 , G n ) + N ( K 1 , 4 , G n ) = O ( N ( P 3 , G n ) 4 7 ) . (45)
Note that if N ( P 3 , G n ) = 0 N(P_3,G_n)=0 N ( P 3 , G n ) = 0 , then both terms on the left of (45) vanish. We therefore assume N ( P 3 , G n ) ≥ 1 N(P_3,G_n)\ge1 N ( P 3 , G n ) ≥ 1 . Using ( d G n ( i ) 2 ) ≤ N ( P 3 , G n ) \binom{d_{G_n}(i)}{2}\le N(P_3,G_n) ( 2 d G n ( i ) ) ≤ N ( P 3 , G n ) , it follows that d G n ( i ) = O ( N ( P 3 , G n ) 1 2 ) d_{G_n}(i)=O\left(N(P_3,G_n)^{\frac{1}{2}}\right) d G n ( i ) = O ( N ( P 3 , G n ) 2 1 ) , uniformly in 1 ≤ i ≤ n 1\le i\le n 1 ≤ i ≤ n . Hence, by the Cauchy–Schwarz inequality,
∑ j : { i , j } ∈ E ( G n ) ( d G n ( j ) − 1 ) ≤ ( 2 d G n ( i ) N ( P 3 , G n ) ) 1 2 = O ( N ( P 3 , G n ) 3 4 ) , \sum_{j:\{i,j\}\in E(G_n)}(d_{G_n}(j)-1)
\le\left(2d_{G_n}(i)N(P_3,G_n)\right)^{\frac{1}{2}}
=O\left(N(P_3,G_n)^{\frac{3}{4}}\right), j : { i , j } ∈ E ( G n ) ∑ ( d G n ( j ) − 1 ) ≤ ( 2 d G n ( i ) N ( P 3 , G n ) ) 2 1 = O ( N ( P 3 , G n ) 4 3 ) ,
where we used ∑ j = 1 n d G n ( j ) ( d G n ( j ) − 1 ) = 2 N ( P 3 , G n ) \sum_{j=1}^n d_{G_n}(j)(d_{G_n}(j)-1)=2N(P_3,G_n) ∑ j = 1 n d G n ( j ) ( d G n ( j ) − 1 ) = 2 N ( P 3 , G n ) . Counting copies of P 4 P_4 P 4 according to their middle vertex gives,
N ( P 4 , G n ) ≤ 1 2 ∑ i = 1 n ( ∑ j : { i , j } ∈ E ( G n ) ( d G n ( j ) − 1 ) ) 2 = O ( N ( P 3 , G n ) 7 4 ) , N(P_4,G_n)\le\frac{1}{2}\sum_{i=1}^n\left(\sum_{j:\{i,j\}\in E(G_n)}(d_{G_n}(j)-1)\right)^2
=O\left(N(P_3,G_n)^{\frac{7}{4}}\right), N ( P 4 , G n ) ≤ 2 1 i = 1 ∑ n j : { i , j } ∈ E ( G n ) ∑ ( d G n ( j ) − 1 ) 2 = O ( N ( P 3 , G n ) 4 7 ) ,
since ∑ i = 1 n ∑ j : { i , j } ∈ E ( G n ) ( d G n ( j ) − 1 ) = 2 N ( P 3 , G n ) \sum_{i=1}^n\sum_{j:\{i,j\}\in E(G_n)}(d_{G_n}(j)-1)=2N(P_3,G_n) ∑ i = 1 n ∑ j : { i , j } ∈ E ( G n ) ( d G n ( j ) − 1 ) = 2 N ( P 3 , G n ) . Similarly, counting copies of K 1 , 4 K_{1,4} K 1 , 4 according to their unique vertex of degree three gives,
N ( K 1 , 4 , G n ) ≤ ∑ i : d G n ( i ) ≥ 3 ( d G n ( i ) − 1 2 ) ∑ j : { i , j } ∈ E ( G n ) ( d G n ( j ) − 1 ) = O ( N ( P 3 , G n ) 7 4 ) , N(K_{1,4},G_n)\le
\sum_{i:d_{G_n}(i)\ge3}\binom{d_{G_n}(i)-1}{2}
\sum_{j:\{i,j\}\in E(G_n)}(d_{G_n}(j)-1)
=O\left(N(P_3,G_n)^{\frac{7}{4}}\right), N ( K 1 , 4 , G n ) ≤ i : d G n ( i ) ≥ 3 ∑ ( 2 d G n ( i ) − 1 ) j : { i , j } ∈ E ( G n ) ∑ ( d G n ( j ) − 1 ) = O ( N ( P 3 , G n ) 4 7 ) ,
because ∑ i : d G n ( i ) ≥ 3 ( d G n ( i ) − 1 2 ) ≤ N ( P 3 , G n ) \sum_{i:d_{G_n}(i)\ge3}\binom{d_{G_n}(i)-1}{2}\le N(P_3,G_n) ∑ i : d G n ( i ) ≥ 3 ( 2 d G n ( i ) − 1 ) ≤ N ( P 3 , G n ) . This proves (45) .
We consider the following cases depending on the order of N ( P 3 , G n ) N(P_3,G_n) N ( P 3 , G n ) .
N ( ∙ − ∙ , G n ) = o ( N ( ∨ , G n ) ) N(\bullet\!\!-\!\!\bullet,G_n)=o(N(\mathord{\vee},G_n)) N ( ∙ − ∙ , G n ) = o ( N ( ∨ , G n )) : Then (38) follows by plugging the values for c H c_H c H and the orders of N ( H , G n ) N(H,G_n) N ( H , G n ) from Table 1 into (44) , since every count other than N ( × , G n ) N(\mathord{\times},G_n) N ( × , G n ) is
O ( N ( ∙ − ∙ , G n ) N ( ∨ , G n ) ) + o ( N ( ∨ , G n ) 2 ) . O(N(\bullet\!\!-\!\!\bullet,G_n)N(\mathord{\vee},G_n))+o(N(\mathord{\vee},G_n)^2). O ( N ( ∙ − ∙ , G n ) N ( ∨ , G n )) + o ( N ( ∨ , G n ) 2 ) .
N ( ∨ , G n ) ≤ C N ( ∙ − ∙ , G n ) N(\mathord{\vee},G_n)\leq CN(\bullet\!\!-\!\!\bullet,G_n) N ( ∨ , G n ) ≤ C N ( ∙ − ∙ , G n ) , for some C > 0 C>0 C > 0 : Note that
Δ ( G n ) : = max 1 ≤ i ≤ n d G n ( i ) = O ( N ( ∙ − ∙ , G n ) ) , \Delta(G_n):=\max_{1\leq i\leq n}d_{G_n}(i)=O(\sqrt{N(\bullet\!\!-\!\!\bullet,G_n)}), Δ ( G n ) := 1 ≤ i ≤ n max d G n ( i ) = O ( N ( ∙ − ∙ , G n ) ) ,
since ( ( Δ ( G n ) 2 ) ≤ N ( ∨ , G n ) ≤ C N ( ∙ − ∙ , G n ) ) (\Delta(G_n)\choose2)\leq N(\mathord{\vee},G_n)\leq CN(\bullet\!\!-\!\!\bullet,G_n) ( 2 ) ≤ N ( ∨ , G n ) ≤ C N ( ∙ − ∙ , G n ) ( Δ ( G n ) ) . Using this bound and (45) , each motif count in Table 1 other than N ( × , G n ) N(\mathord{\times},G_n) N ( × , G n ) is seen to be o ( N ( ∙ − ∙ , G n ) 2 ) o(N(\bullet\!\!-\!\!\bullet,G_n)^2) o ( N ( ∙ − ∙ , G n ) 2 ) : for instance
N ( □ , G n ) ≤ N ( ∨ , G n ) Δ ( G n ) = O ( N ( ∙ − ∙ , G n ) 3 / 2 ) N(\mathord{\square},G_n)\leq N(\mathord{\vee},G_n)\Delta(G_n)=O(N(\bullet\!\!-\!\!\bullet,G_n)^{3/2}) N ( □ , G n ) ≤ N ( ∨ , G n ) Δ ( G n ) = O ( N ( ∙ − ∙ , G n ) 3/2 )
and
N ( △ , G n ) ≤ N ( ∨ , G n ) Δ ( G n ) = O ( N ( ∙ − ∙ , G n ) 3 / 2 ) . N(\mathord{\triangle},G_n)\leq N(\mathord{\vee},G_n)\Delta(G_n)=O(N(\bullet\!\!-\!\!\bullet,G_n)^{3/2}). N ( △ , G n ) ≤ N ( ∨ , G n ) Δ ( G n ) = O ( N ( ∙ − ∙ , G n ) 3/2 ) .
Hence, (38) also holds in this case.
Proof of Lemma 3.4. Since N ( △ , G n ) = ∑ i = 1 n ( d G n ( i ) 3 ) N(\mathord{\triangle},G_n)=\sum_{i=1}^{n}{d_{G_n}(i)\choose3} N ( △ , G n ) = ∑ i = 1 n ( 3 d G n ( i ) ) , we split this sum according to whether d G n ( i ) = 3 d_{G_n}(i)=3 d G n ( i ) = 3 or d G n ( i ) ≥ 4 d_{G_n}(i)\geq4 d G n ( i ) ≥ 4 . For d G n ( i ) = 3 d_{G_n}(i)=3 d G n ( i ) = 3 , use ( d G n ( i ) 3 ) = 1 ≤ ( d G n ( i ) 2 ) {d_{G_n}(i)\choose3}=1\leq{d_{G_n}(i)\choose2} ( 3 d G n ( i ) ) = 1 ≤ ( 2 d G n ( i ) ) , so the sum over all vertices of degree 3 is at most N ( ∨ , G n ) N(\mathord{\vee},G_n) N ( ∨ , G n ) . For d G n ( i ) ≥ 4 d_{G_n}(i)\geq4 d G n ( i ) ≥ 4 , we can write
( d G n ( i ) 3 ) = d G n ( i ) − 2 3 ( d G n ( i ) 2 ) . {d_{G_n}(i)\choose3}=\frac{d_{G_n}(i)-2}{3}{d_{G_n}(i)\choose2}. ( 3 d G n ( i ) ) = 3 d G n ( i ) − 2 ( 2 d G n ( i ) ) .
Summing over all vertices i i i with d G n ( i ) ≥ 4 d_{G_n}(i)\geq4 d G n ( i ) ≥ 4 and applying the Cauchy–Schwarz inequality gives
∑ i : d G n ( i ) ≥ 4 ( d G n ( i ) 3 ) ≤ ∑ i : d G n ( i ) ≥ 4 ( d G n ( i ) 2 ) ∑ i : d G n ( i ) ≥ 4 ( d G n ( i ) 2 ) ( d G n ( i ) − 2 3 ) 2 ≤ N ( ∨ , G n ) ∑ i : d G n ( i ) ≥ 4 4 3 ( d G n ( i ) 4 ) d G n ( i ) − 2 d G n ( i ) − 3 ≤ 8 3 N ( ∨ , G n ) ∑ i = 1 n ( d G n ( i ) 4 ) , \begin{aligned}
\sum_{i:d_{G_n}(i)\geq4}{d_{G_n}(i)\choose3}
\leq
\sqrt{\sum_{i:d_{G_n}(i)\geq4}{d_{G_n}(i)\choose2}}
\sqrt{\sum_{i:d_{G_n}(i)\geq4}{d_{G_n}(i)\choose2}\left(\frac{d_{G_n}(i)-2}{3}\right)^2} \\
&\leq
\sqrt{N(\mathord{\vee},G_n)}
\sqrt{\sum_{i:d_{G_n}(i)\geq4}\frac{4}{3}{d_{G_n}(i)\choose4}\frac{d_{G_n}(i)-2}{d_{G_n}(i)-3}} \\
&\leq
\sqrt{\frac{8}{3}N(\mathord{\vee},G_n)\sum_{i=1}^{n}{d_{G_n}(i)\choose4}},
\end{aligned} i : d G n ( i ) ≥ 4 ∑ ( 3 d G n ( i ) ) ≤ i : d G n ( i ) ≥ 4 ∑ ( 2 d G n ( i ) ) i : d G n ( i ) ≥ 4 ∑ ( 2 d G n ( i ) ) ( 3 d G n ( i ) − 2 ) 2 ≤ N ( ∨ , G n ) i : d G n ( i ) ≥ 4 ∑ 3 4 ( 4 d G n ( i ) ) d G n ( i ) − 3 d G n ( i ) − 2 ≤ 3 8 N ( ∨ , G n ) i = 1 ∑ n ( 4 d G n ( i ) ) ,
where we used ( d G n ( i ) 2 ) ( d G n ( i ) − 2 ) 2 9 = 4 3 ( d G n ( i ) 4 ) d G n ( i ) − 2 d G n ( i ) − 3 {d_{G_n}(i)\choose2}\frac{(d_{G_n}(i)-2)^2}{9}=\frac{4}{3}{d_{G_n}(i)\choose4}\frac{d_{G_n}(i)-2}{d_{G_n}(i)-3} ( 2 d G n ( i ) ) 9 ( d G n ( i ) − 2 ) 2 = 3 4 ( 4 d G n ( i ) ) d G n ( i ) − 3 d G n ( i ) − 2 and the fact that d G n ( i ) − 2 d G n ( i ) − 3 ≤ 2 \frac{d_{G_n}(i)-2}{d_{G_n}(i)-3}\leq2 d G n ( i ) − 3 d G n ( i ) − 2 ≤ 2 , for d G n ( i ) ≥ 4 d_{G_n}(i)\geq4 d G n ( i ) ≥ 4 . Combining the above bounds the result follows.
Proof of Theorem 3.3 (ii)
Lemma 3.7. Under the hypotheses of Theorem 3.3 (ii),
∑ i = 1 n d G n ( i ) 3 = o ( n 1 n 2 ( N ( ∙ − ∙ , G n ) + N ( ∨ , G n ) ) 3 2 ) . \sum_{i=1}^{n}d_{G_n}(i)^3=o\left(n^{\frac{1}{n^2}}\left(N(\bullet\!\!-\!\!\bullet,G_n)+N(\mathord{\vee},G_n)\right)^{\frac{3}{2}}\right). i = 1 ∑ n d G n ( i ) 3 = o ( n n 2 1 ( N ( ∙ − ∙ , G n ) + N ( ∨ , G n ) ) 2 3 ) .
Proof. Using the identity d G n ( i ) 3 = 6 ( d G n ( i ) 3 ) + 6 ( d G n ( i ) 2 ) + d G n ( i ) d_{G_n}(i)^3=6\binom{d_{G_n}(i)}{3}+6\binom{d_{G_n}(i)}{2}+d_{G_n}(i) d G n ( i ) 3 = 6 ( 3 d G n ( i ) ) + 6 ( 2 d G n ( i ) ) + d G n ( i ) , we obtain
∑ i = 1 n d G n ( i ) 3 = 6 N ( ∙ − ∙ , G n ) + 6 N ( ∨ , G n ) + 2 N ( − ∙ , G n ) \sum_{i=1}^{n}d_{G_n}(i)^3=6N(\bullet\!-\!\bullet,G_n)+6N(\mathord{\vee},G_n)+2N(\mathord{-\!\bullet},G_n) i = 1 ∑ n d G n ( i ) 3 = 6 N ( ∙ − ∙ , G n ) + 6 N ( ∨ , G n ) + 2 N ( − ∙ , G n )
≤ 6 ( N ( ∙ − ∙ , G n ) + N ( ∨ , G n ) + N ( − ∙ , G n ) ) . \leq6\left(N(\bullet\!-\!\bullet,G_n)+N(\mathord{\vee},G_n)+N(\mathord{-\!\bullet},G_n)\right). ≤ 6 ( N ( ∙ − ∙ , G n ) + N ( ∨ , G n ) + N ( − ∙ , G n ) ) .
By the second condition in Theorem 3.3 (ii),
N ( ∙ − ∙ , G n ) = o ( p n 1 2 ( N ( − ∙ , G n ) + N ( ∨ , G n ) ) 3 2 ) . N(\bullet\!-\!\bullet,G_n)=o\left(p_n^{\frac{1}{2}}\left(N(\mathord{-\!\bullet},G_n)+N(\mathord{\vee},G_n)\right)^{\frac{3}{2}}\right). N ( ∙ − ∙ , G n ) = o ( p n 2 1 ( N ( − ∙ , G n ) + N ( ∨ , G n ) ) 2 3 ) .
Combining the above with the first condition in Theorem 3.3 (ii) gives the required result. □ \square □
Proof of Theorem 3.3 (ii): We first bound A 1 A_1 A 1 . Recall from (3) that
σ n 2 = ( 1 − u n ) ( 1 + u n ) 4 N ( − ∙ , G n ) + u n ( 1 − u n ) 2 N ( ∨ , G n ) = 2 p n ( N ( − ∙ , G n ) + N ( ∨ , G n ) ) ( 1 + o ( 1 ) ) , (46) \begin{aligned}
\sigma_n^2=\frac{(1-u_n)(1+u_n)}{4}N(\mathord{-\!\bullet},G_n)+\frac{u_n(1-u_n)}{2}N(\mathord{\vee},G_n) \\
&=2p_n\left(N(\mathord{-\!\bullet},G_n)+N(\mathord{\vee},G_n)\right)(1+o(1)), \tag*{(46)}
\end{aligned} σ n 2 = 4 ( 1 − u n ) ( 1 + u n ) N ( − ∙ , G n ) + 2 u n ( 1 − u n ) N ( ∨ , G n ) = 2 p n ( N ( − ∙ , G n ) + N ( ∨ , G n ) ) ( 1 + o ( 1 )) , (46)
since 1 − u n = 4 p n ( 1 − p n ) = 4 p n ( 1 + o ( 1 ) ) 1-u_n=4p_n(1-p_n)=4p_n(1+o(1)) 1 − u n = 4 p n ( 1 − p n ) = 4 p n ( 1 + o ( 1 )) and u n → 1 u_n\to1 u n → 1 , as p n → 0 p_n\to0 p n → 0 . Hence, under the first condition in Theorem 3.3 (ii), σ n 2 \sigma_n^2 σ n 2 diverges. Now, from (36) and (41) , we have
∣ A 1 ∣ ≤ 1 σ n ∑ e ∈ E ( G n ) E [ ∣ Y e ∣ ⋅ ( U n − U e ) 2 ] ≤ 1 σ n 3 ∑ e ∈ E ( G n ) ∑ e 1 ∈ E ( G n ) e 1 ∩ e ≠ ∅ ∑ e 2 ∈ E ( G n ) e 2 ∩ e ≠ ∅ E [ ∣ Y e Y e 1 Y e 2 ∣ ] . (47) \begin{aligned}
|A_1|\leq\frac{1}{\sigma_n}\sum_{e\in E(G_n)}\mathbb{E}\left[|Y_e|\cdot(U_n-U_e)^2\right] \\
&\leq\frac{1}{\sigma_n^3}\sum_{e\in E(G_n)}\sum_{\substack{e_1\in E(G_n)\\e_1\cap e\neq\varnothing}}\sum_{\substack{e_2\in E(G_n)\\e_2\cap e\neq\varnothing}}\mathbb{E}\left[|Y_eY_{e_1}Y_{e_2}|\right]. \tag*{(47)}
\end{aligned} ∣ A 1 ∣ ≤ σ n 1 e ∈ E ( G n ) ∑ E [ ∣ Y e ∣ ⋅ ( U n − U e ) 2 ] ≤ σ n 3 1 e ∈ E ( G n ) ∑ e 1 ∈ E ( G n ) e 1 ∩ e = ∅ ∑ e 2 ∈ E ( G n ) e 2 ∩ e = ∅ ∑ E [ ∣ Y e Y e 1 Y e 2 ∣ ] . (47)
To estimate the expectation of this triple product, for each e = { i , j } ∈ E ( G n ) e=\{i,j\}\in E(G_n) e = { i , j } ∈ E ( G n ) define Z e : = 1 { X i ≠ X j } Z_e:=1\{X_i\neq X_j\} Z e := 1 { X i = X j } , so that
E [ Z e ] = 2 p n ( 1 − p n ) ≤ 2 p n . (48) \mathbb{E}[Z_e]=2p_n(1-p_n)\leq2p_n. \tag*{(48)} E [ Z e ] = 2 p n ( 1 − p n ) ≤ 2 p n . (48)
Recall that q n = E [ 1 { X i = X j } ] = p n 2 + ( 1 − p n ) 2 q_n=\mathbb{E}[1\{X_i=X_j\}]=p_n^2+(1-p_n)^2 q n = E [ 1 { X i = X j }] = p n 2 + ( 1 − p n ) 2 and, hence, Y e = 1 { X i = X j } − q n = ( 1 − q n ) − Z e Y_e=1\{X_i=X_j\}-q_n=(1-q_n)-Z_e Y e = 1 { X i = X j } − q n = ( 1 − q n ) − Z e and 1 − q n = 2 p n ( 1 − p n ) ≤ 2 p n 1-q_n=2p_n(1-p_n)\leq2p_n 1 − q n = 2 p n ( 1 − p n ) ≤ 2 p n . Further, ∣ Y e ∣ = 1 − q n |Y_e|=1-q_n ∣ Y e ∣ = 1 − q n on { Z e = 0 } \{Z_e=0\} { Z e = 0 } and ∣ Y e ∣ = q n ≤ 1 |Y_e|=q_n\leq1 ∣ Y e ∣ = q n ≤ 1 on { Z e = 1 } \{Z_e=1\} { Z e = 1 } . This gives the bound,
∣ Y e ∣ ≤ 2 p n + Z e . |Y_e|\leq2p_n+Z_e. ∣ Y e ∣ ≤ 2 p n + Z e .
Using this bound on ∣ Y e ∣ |Y_e| ∣ Y e ∣ , the triple product in (47) satisfies
∣ Y e Y e 1 Y e 2 ∣ ≤ ( 2 p n + Z e ) ( 2 p n + Z e 1 ) ( 2 p n + Z e 2 ) |Y_eY_{e_1}Y_{e_2}|\leq(2p_n+Z_e)(2p_n+Z_{e_1})(2p_n+Z_{e_2}) ∣ Y e Y e 1 Y e 2 ∣ ≤ ( 2 p n + Z e ) ( 2 p n + Z e 1 ) ( 2 p n + Z e 2 )
≤ 8 p n 3 + 4 p n 2 ( Z e + Z e 1 + Z e 2 ) + 2 p n ( Z e Z e 1 + Z e Z e 2 + Z e 1 Z e 2 ) + Z e Z e 1 Z e 2 . \leq8p_n^3+4p_n^2(Z_e+Z_{e_1}+Z_{e_2})+2p_n(Z_eZ_{e_1}+Z_eZ_{e_2}+Z_{e_1}Z_{e_2})+Z_eZ_{e_1}Z_{e_2}. ≤ 8 p n 3 + 4 p n 2 ( Z e + Z e 1 + Z e 2 ) + 2 p n ( Z e Z e 1 + Z e Z e 2 + Z e 1 Z e 2 ) + Z e Z e 1 Z e 2 .
Taking expectations and using (48) together with E [ Z a Z b ] ≤ E [ Z a ] ≤ 2 p n \mathbb{E}[Z_aZ_b]\leq\mathbb{E}[Z_a]\leq2p_n E [ Z a Z b ] ≤ E [ Z a ] ≤ 2 p n and E [ Z a Z b Z c ] ≤ E [ Z a ] ≤ 2 p n \mathbb{E}[Z_aZ_bZ_c]\leq\mathbb{E}[Z_a]\leq2p_n E [ Z a Z b Z c ] ≤ E [ Z a ] ≤ 2 p n , for all a , b , c ∈ E ( G n ) a,b,c\in E(G_n) a , b , c ∈ E ( G n ) ,
E [ ∣ Y e Y e 1 Y e 2 ∣ ] ≤ 8 p n 3 + 24 p n 3 + 12 p n 2 + 2 p n ≤ C p n , (49) \mathbb{E}[|Y_eY_{e_1}Y_{e_2}|]\leq8p_n^3+24p_n^3+12p_n^2+2p_n\leq Cp_n, \tag*{(49)} E [ ∣ Y e Y e 1 Y e 2 ∣ ] ≤ 8 p n 3 + 24 p n 3 + 12 p n 2 + 2 p n ≤ C p n , (49)
for an absolute constant C C C and all n n n large enough. Combining (46) and (49) , together with (47) , gives,
∣ A 1 ∣ ≤ C p n σ n 3 ∑ e ∈ E ( G n ) ∣ { e ′ ∈ E ( G n ) : e ′ ∩ e ≠ ∅ } ∣ 2 ≤ C ′ ∑ i = 1 n d G n ( i ) 3 p n 1 / 2 ( N ( ∨ , G n ) + N ( − ∙ , G n ) ) 3 / 2 , (50) |A_1|\leq\frac{Cp_n}{\sigma_n^3}\sum_{e\in E(G_n)}\left|\{e'\in E(G_n):e'\cap e\neq\varnothing\}\right|^2
\leq\frac{C'\sum_{i=1}^{n}d_{G_n}(i)^3}{p_n^{1/2}\left(N(\mathord{\vee},G_n)+N(\mathord{-\!\bullet},G_n)\right)^{3/2}}, \tag*{(50)} ∣ A 1 ∣ ≤ σ n 3 C p n e ∈ E ( G n ) ∑ ∣ { e ′ ∈ E ( G n ) : e ′ ∩ e = ∅ } ∣ 2 ≤ p n 1/2 ( N ( ∨ , G n ) + N ( − ∙ , G n ) ) 3/2 C ′ ∑ i = 1 n d G n ( i ) 3 , (50)
for some constant C ′ > 0 C'>0 C ′ > 0 . Using Lemma 3.7 in (50) , we get ∣ A 1 ∣ = o ( 1 ) |A_1|=o(1) ∣ A 1 ∣ = o ( 1 ) .
We now bound ∣ A 2 ∣ |A_2| ∣ A 2 ∣ . By (36) and the definition of the function class F \mathcal{F} F ,
∣ A 2 ∣ ≤ 2 π 1 σ n 2 E [ ∣ V n − E [ V n ] ∣ ] . (51) |A_2| \leq\sqrt{\frac{2}{\pi}\frac{1}{\sigma_n^2}\mathbb{E}\left[|V_n-\mathbb{E}[V_n]|\right]}. \tag*{(51)} ∣ A 2 ∣ ≤ π 2 σ n 2 1 E [ ∣ V n − E [ V n ] ∣ ] . (51)
Define V ~ n : = ∑ i = 1 n d G n ( i ) 2 X i \widetilde{V}_n := \sum_{i=1}^n d_{G_n}(i)^2X_i V n := ∑ i = 1 n d G n ( i ) 2 X i , where d G n ( i ) d_{G_n}(i) d G n ( i ) is the degree of the i i i -th vertex in G n G_n G n . We first show that
E [ ∣ V n − V ~ n ∣ ] ≤ C p n 2 ∑ i = 1 n d G n ( i ) 2 , (52) \mathbb{E}\left[|V_n-\widetilde{V}_n|\right] \leq Cp_n^2\sum_{i=1}^n d_{G_n}(i)^2, \tag*{(52)} E [ ∣ V n − V n ∣ ] ≤ C p n 2 i = 1 ∑ n d G n ( i ) 2 , (52)
for some constant C > 0 C>0 C > 0 . Recall that Y e = ( 1 − q n ) − Z e Y_e=(1-q_n)-Z_e Y e = ( 1 − q n ) − Z e and E [ Z e ] = 1 − q n = 2 p n ( 1 − p n ) \mathbb{E}[Z_e]=1-q_n=2p_n(1-p_n) E [ Z e ] = 1 − q n = 2 p n ( 1 − p n ) . Hence, for any two edges e 1 , e 2 ∈ E ( G n ) e_1,e_2\in E(G_n) e 1 , e 2 ∈ E ( G n ) ,
E [ ∣ Y e 1 Y e 2 − Z e 1 Z e 2 ∣ ] ≤ ( 1 − q n ) E [ Z e 1 + Z e 2 ] + ( 1 − q n ) 2 = 3 ( 1 − q n ) 2 ≤ 12 p n 2 . \mathbb{E}\left[|Y_{e_1}Y_{e_2}-Z_{e_1}Z_{e_2}|\right]\leq(1-q_n)\mathbb{E}\left[Z_{e_1}+Z_{e_2}\right]+(1-q_n)^2=3(1-q_n)^2\leq12p_n^2. E [ ∣ Y e 1 Y e 2 − Z e 1 Z e 2 ∣ ] ≤ ( 1 − q n ) E [ Z e 1 + Z e 2 ] + ( 1 − q n ) 2 = 3 ( 1 − q n ) 2 ≤ 12 p n 2 .
Further, for e 1 = { i , j } e_1=\{i,j\} e 1 = { i , j } and e 2 = { j , k } e_2=\{j,k\} e 2 = { j , k } such that the indices i , j , k i,j,k i , j , k are distinct,
Z e 1 2 = X i + X j − 2 X i X j and Z e 1 Z e 2 = X j − X i X j − X j X k + X i X k . Z_{e_1}^2=X_i+X_j-2X_iX_j\quad\text{and}\quad Z_{e_1}Z_{e_2}=X_j-X_iX_j-X_jX_k+X_iX_k. Z e 1 2 = X i + X j − 2 X i X j and Z e 1 Z e 2 = X j − X i X j − X j X k + X i X k .
Hence, in the sum
∑ e 1 , e 2 ∈ E ( G n ) e 1 ∩ e 2 ≠ ∅ Z e 1 Z e 2 , \sum_{\substack{e_1,e_2\in E(G_n)\\ e_1\cap e_2\neq\varnothing}} Z_{e_1}Z_{e_2}, e 1 , e 2 ∈ E ( G n ) e 1 ∩ e 2 = ∅ ∑ Z e 1 Z e 2 ,
the linear terms add to ∑ i = 1 n ( d G n ( i ) + d G n ( i ) ( d G n ( i ) − 1 ) ) X i = V ~ n \sum_{i=1}^n(d_{G_n}(i)+d_{G_n}(i)(d_{G_n}(i)-1))X_i=\widetilde{V}_n ∑ i = 1 n ( d G n ( i ) + d G n ( i ) ( d G n ( i ) − 1 )) X i = V n . By the triangle inequality,
E [ ∣ V n − V ~ n ∣ ] ≤ E [ ∣ ∑ e 1 , e 2 ∈ E ( G n ) e 1 ∩ e 2 ≠ ∅ ( Y e 1 Y e 2 − Z e 1 Z e 2 ) ∣ ] + E [ ∣ ∑ e 1 , e 2 ∈ E ( G n ) e 1 ∩ e 2 ≠ ∅ Z e 1 Z e 2 − V ~ n ∣ ] ≤ ∑ e 1 , e 2 ∈ E ( G n ) e 1 ∩ e 2 ≠ ∅ E [ ∣ Y e 1 Y e 2 − Z e 1 Z e 2 ∣ ] + 3 p n 2 ( N ( ∙ − ∙ , G n ) + 2 N ( ∨ , G n ) ) ≤ 15 p n 2 ( N ( ∙ − ∙ , G n ) + 2 N ( ∨ , G n ) ) ≤ 15 p n 2 ∑ i = 1 n d G n ( i ) 2 . \begin{aligned}
\mathbb{E}\left[|V_n-\widetilde{V}_n|\right]
&\leq\mathbb{E}\left[\left|\sum_{\substack{e_1,e_2\in E(G_n)\\e_1\cap e_2\neq\varnothing}}(Y_{e_1}Y_{e_2}-Z_{e_1}Z_{e_2})\right|\right]
+\mathbb{E}\left[\left|\sum_{\substack{e_1,e_2\in E(G_n)\\e_1\cap e_2\neq\varnothing}}Z_{e_1}Z_{e_2}-\widetilde{V}_n\right|\right]\\
&\leq\sum_{\substack{e_1,e_2\in E(G_n)\\e_1\cap e_2\neq\varnothing}}\mathbb{E}\left[|Y_{e_1}Y_{e_2}-Z_{e_1}Z_{e_2}|\right]
+3p_n^2\left(N(\bullet\!\!-\!\!\bullet,G_n)+2N(\mathord{\vee},G_n)\right)\\
&\leq15p_n^2\left(N(\bullet\!\!-\!\!\bullet,G_n)+2N(\mathord{\vee},G_n)\right)\\
&\leq15p_n^2\sum_{i=1}^n d_{G_n}(i)^2.
\end{aligned} E [ ∣ V n − V n ∣ ] ≤ E e 1 , e 2 ∈ E ( G n ) e 1 ∩ e 2 = ∅ ∑ ( Y e 1 Y e 2 − Z e 1 Z e 2 ) + E e 1 , e 2 ∈ E ( G n ) e 1 ∩ e 2 = ∅ ∑ Z e 1 Z e 2 − V n ≤ e 1 , e 2 ∈ E ( G n ) e 1 ∩ e 2 = ∅ ∑ E [ ∣ Y e 1 Y e 2 − Z e 1 Z e 2 ∣ ] + 3 p n 2 ( N ( ∙ − ∙ , G n ) + 2 N ( ∨ , G n ) ) ≤ 15 p n 2 ( N ( ∙ − ∙ , G n ) + 2 N ( ∨ , G n ) ) ≤ 15 p n 2 i = 1 ∑ n d G n ( i ) 2 .
This proves (52) . Then by triangle inequality,
E [ ∣ V n − E [ V n ] ∣ ] ≤ E [ ∣ V ~ n − E [ V ~ n ] ∣ ] + 2 E [ ∣ V n − V ~ n ∣ ] ≤ E [ ∣ ∑ i = 1 n d G n ( i ) 2 ( X i − p n ) ∣ ] + 30 p n 2 ∑ i = 1 n d G n ( i ) 2 . (53) \begin{aligned}
\mathbb{E}\left[|V_n-\mathbb{E}[V_n]|\right]
&\leq\mathbb{E}\left[|\widetilde{V}_n-\mathbb{E}[\widetilde{V}_n]|\right]+2\mathbb{E}\left[|V_n-\widetilde{V}_n|\right]\\
&\leq\mathbb{E}\left[\left|\sum_{i=1}^n d_{G_n}(i)^2(X_i-p_n)\right|\right]+30p_n^2\sum_{i=1}^n d_{G_n}(i)^2. \tag*{(53)}
\end{aligned} E [ ∣ V n − E [ V n ] ∣ ] ≤ E [ ∣ V n − E [ V n ] ∣ ] + 2 E [ ∣ V n − V n ∣ ] ≤ E [ i = 1 ∑ n d G n ( i ) 2 ( X i − p n ) ] + 30 p n 2 i = 1 ∑ n d G n ( i ) 2 . (53)
To bound the first term, set t n : = ( p n ∑ i = 1 n d G n ( i ) 3 ) 1 3 t_n:=(p_n\sum_{i=1}^n d_{G_n}(i)^3)^{\frac{1}{3}} t n := ( p n ∑ i = 1 n d G n ( i ) 3 ) 3 1 and first consider the indices where d G n ( i ) ≤ t n d_{G_n}(i)\leq t_n d G n ( i ) ≤ t n . By the Cauchy–Schwarz inequality and independence,
( E [ ∣ ∑ i : d G n ( i ) ≤ t n d G n ( i ) 2 ( X i − p n ) ∣ ] ) 2 ≤ E [ ( ∑ i : d G n ( i ) ≤ t n d G n ( i ) 2 ( X i − p n ) ) 2 ] ≤ p n ∑ i : d G n ( i ) ≤ t n d G n ( i ) 4 . \left(\mathbb{E}\left[\left|\sum_{\substack{i:d_{G_n}(i)\leq t_n}}d_{G_n}(i)^2(X_i-p_n)\right|\right]\right)^2
\leq\mathbb{E}\left[\left(\sum_{\substack{i:d_{G_n}(i)\leq t_n}}d_{G_n}(i)^2(X_i-p_n)\right)^2\right]
\leq p_n\sum_{\substack{i:d_{G_n}(i)\leq t_n}}d_{G_n}(i)^4. E i : d G n ( i ) ≤ t n ∑ d G n ( i ) 2 ( X i − p n ) 2 ≤ E i : d G n ( i ) ≤ t n ∑ d G n ( i ) 2 ( X i − p n ) 2 ≤ p n i : d G n ( i ) ≤ t n ∑ d G n ( i ) 4 .
For the indices where d G n ( i ) > t n d_{G_n}(i)>t_n d G n ( i ) > t n we use the bound
E [ ∣ ∑ i : d G n ( i ) > t n d G n ( i ) 2 ( X i − p n ) ∣ ] ≤ ∑ i : d G n ( i ) > t n d G n ( i ) 2 E [ ∣ X i − p n ∣ ] ≤ 2 p n ∑ i : d G n ( i ) > t n d G n ( i ) 2 . \mathbb{E}\left[\left|\sum_{i:d_{G_n}(i)>t_n}d_{G_n}(i)^2(X_i-p_n)\right|\right]\leq\sum_{i:d_{G_n}(i)>t_n}d_{G_n}(i)^2\mathbb{E}[|X_i-p_n|]\leq2p_n\sum_{i:d_{G_n}(i)>t_n}d_{G_n}(i)^2. E i : d G n ( i ) > t n ∑ d G n ( i ) 2 ( X i − p n ) ≤ i : d G n ( i ) > t n ∑ d G n ( i ) 2 E [ ∣ X i − p n ∣ ] ≤ 2 p n i : d G n ( i ) > t n ∑ d G n ( i ) 2 .
Hence,
E [ ∣ ∑ i = 1 n d G n ( i ) 2 ( X i − p n ) ∣ ] ≤ ( p n ∑ i : d G n ( i ) ≤ t n d G n ( i ) 4 ) 1 2 + 2 p n ∑ i : d G n ( i ) > t n d G n ( i ) 2 ≤ ( p n t n ∑ i = 1 n d G n ( i ) 3 ) 1 2 + 2 p n t n ∑ i = 1 n d G n ( i ) 3 = 3 ( p n ∑ i = 1 n d G n ( i ) 3 ) 2 3 . (54) \begin{aligned}
\mathbb{E}\left[\left|\sum_{i=1}^n d_{G_n}(i)^2(X_i-p_n)\right|\right]
&\leq\left(p_n\sum_{i:d_{G_n}(i)\leq t_n}d_{G_n}(i)^4\right)^{\frac{1}{2}}+2p_n\sum_{i:d_{G_n}(i)>t_n}d_{G_n}(i)^2\\
&\leq\left(p_nt_n\sum_{i=1}^n d_{G_n}(i)^3\right)^{\frac{1}{2}}+\frac{2p_n}{t_n}\sum_{i=1}^n d_{G_n}(i)^3\\
&=3\left(p_n\sum_{i=1}^n d_{G_n}(i)^3\right)^{\frac{2}{3}}. \tag*{(54)}
\end{aligned} E [ i = 1 ∑ n d G n ( i ) 2 ( X i − p n ) ] ≤ p n i : d G n ( i ) ≤ t n ∑ d G n ( i ) 4 2 1 + 2 p n i : d G n ( i ) > t n ∑ d G n ( i ) 2 ≤ ( p n t n i = 1 ∑ n d G n ( i ) 3 ) 2 1 + t n 2 p n i = 1 ∑ n d G n ( i ) 3 = 3 ( p n i = 1 ∑ n d G n ( i ) 3 ) 3 2 . (54)
Combining (51) , (53) , and (54) , with (46) and ∑ i = 1 n d G n ( i ) 2 = 2 ( N ( ∙ − ∙ , G n ) + N ( ∙ ∨ ∙ , G n ) ) \sum_{i=1}^n d_{G_n}(i)^2=2(N(\bullet\!-\!\bullet,G_n)+N(\bullet\!\vee\!\bullet,G_n)) ∑ i = 1 n d G n ( i ) 2 = 2 ( N ( ∙ − ∙ , G n ) + N ( ∙ ∨ ∙ , G n )) , for some constant C > 0 C>0 C > 0 ,
∣ A 2 ∣ ≤ C σ n 2 [ ( p n ∑ i = 1 n d G n ( i ) 3 ) 2 3 + p n 2 ∑ i = 1 n d G n ( i ) 2 ] ≤ C [ ( ∑ i = 1 n d G n ( i ) 3 p n 1 2 ( N ( ∙ − ∙ , G n ) + N ( ∙ ∨ ∙ , G n ) ) 3 2 ) 2 3 + p n ] = o ( 1 ) , \begin{aligned}
|A_2|\leq\frac{C}{\sigma_n^2}\left[\left(p_n\sum_{i=1}^n d_{G_n}(i)^3\right)^{\frac{2}{3}}+p_n^2\sum_{i=1}^n d_{G_n}(i)^2\right] \\
&\leq C\left[\left(\frac{\sum_{i=1}^n d_{G_n}(i)^3}{p_n^{\frac{1}{2}}(N(\bullet\!-\!\bullet,G_n)+N(\bullet\!\vee\!\bullet,G_n))^{\frac{3}{2}}}\right)^{\frac{2}{3}}+p_n\right]=o(1),
\end{aligned} ∣ A 2 ∣ ≤ σ n 2 C ( p n i = 1 ∑ n d G n ( i ) 3 ) 3 2 + p n 2 i = 1 ∑ n d G n ( i ) 2 ≤ C ( p n 2 1 ( N ( ∙ − ∙ , G n ) + N ( ∙ ∨ ∙ , G n ) ) 2 3 ∑ i = 1 n d G n ( i ) 3 ) 3 2 + p n = o ( 1 ) ,
where the last step follows from Lemma 3.7 and p n → 0 p_n\to0 p n → 0 . This completes the proof of Theorem 3.3 (ii).
Proof of thm:Un, ppn:clthalfsparse and ppn:thirdfourthsparse
We will prove the results in reverse order. The Proof of Proposition 1.3 follows readily from Lemma 1.4 and Lemma 2.2. We provide the details below.
Proof of Proposition 1.3. Suppose E [ U n 4 ] → 3 \mathbb{E}[U_n^4]\to3 E [ U n 4 ] → 3 . By Lemma 1.4,
p n ( N ( ∙ − ∙ , G n ) + N ( ∙ ∨ ∙ , G n ) ) → ∞ and N ( ∙ ∨ ∙ , G n ) = o ( p n ( N ( ∙ − ∙ , G n ) + N ( ∙ ∨ ∙ , G n ) ) 2 ) . p_n(N(\bullet\!-\!\bullet,G_n)+N(\bullet\!\vee\!\bullet,G_n))\to\infty
\quad\text{and}\quad
N(\bullet\!\vee\!\bullet,G_n)=o\left(p_n(N(\bullet\!-\!\bullet,G_n)+N(\bullet\!\vee\!\bullet,G_n))^2\right). p n ( N ( ∙ − ∙ , G n ) + N ( ∙ ∨ ∙ , G n )) → ∞ and N ( ∙ ∨ ∙ , G n ) = o ( p n ( N ( ∙ − ∙ , G n ) + N ( ∙ ∨ ∙ , G n ) ) 2 ) .
Hence, by (37) ,
N ( ∙ ∙ ∙ , G n ) p n 1 2 ( N ( ∙ − ∙ , G n ) + N ( ∙ ∨ ∙ , G n ) ) 3 2 → 0. \frac{N(\bullet\!\bullet\!\bullet,G_n)}
{p_n^{\frac{1}{2}}(N(\bullet\!-\!\bullet,G_n)+N(\bullet\!\vee\!\bullet,G_n))^{\frac{3}{2}}}\to0. p n 2 1 ( N ( ∙ − ∙ , G n ) + N ( ∙ ∨ ∙ , G n ) ) 2 3 N ( ∙ ∙ ∙ , G n ) → 0.
The conclusion now follows from Lemma 2.2.
The proof of Theorem 1.2 also follows immediately.
Proof of Theorem 1.2. The result in Theorem 1.2 is an immediate consequence of Theorem 3.3(ii) and Lemma 2.2.
The proof of Theorem 1.1 requires one more lemma which ensures the uniform integrability of U n 4 U_n^4 U n 4 . We state the lemma next.
Lemma 4.1. Fix p ∈ ( 0 , 1 ) p \in(0,1) p ∈ ( 0 , 1 ) and let U n U_n U n be as defined in (2) . Then there exists a constant C p < ∞ C_p < \infty C p < ∞ , depending only on p p p , such that
sup n ≥ 1 E [ ∣ U n ∣ 6 ] < C p . (55) \sup_{n \ge1} \mathbb{E}\left[|U_n|^6\right] < C_p. \tag*{(55)} n ≥ 1 sup E [ ∣ U n ∣ 6 ] < C p . (55)
Note that the bound is for fixed p p p . Uniform integrability of U n 4 U_n^4 U n 4 can fail when p n → 0 p_n \to0 p n → 0 , which is why asymptotic normality does not force the moment conditions in the sparse regime (see Example 6.3 ). We shall defer the proof of this lemma to the end of this section and prove Theorem 1.1 assuming the above result.
Proof of Theorem 1.1 : (i) Fix p ∈ ( 0 , 1 ) ∖ { 1 2 , 1 ± 1 / 3 2 } p \in(0,1) \setminus\left\{\frac{1}{2},\frac{1 \pm1/\sqrt{3}}{2}\right\} p ∈ ( 0 , 1 ) ∖ { 2 1 , 2 1 ± 1/ 3 } . Then (b) implies (a) follows from Theorem 3.3(i). The equivalence of (b) and (c) follows from (21) in Lemma 2.3(a). Further, using the bound (37) from Lemma 3.4 and (15) from Lemma 2.1 it follows that (b) implies (d). To prove (d) implies (b), we use the bound
( d G n ( i ) − 3 ) 3 ≤ d G n ( i ) ( d G n ( i ) − 1 ) ( d G n ( i ) − 2 ) = 6 ( d G n ( i ) 3 ) , \left(d_{G_n}(i)-3\right)^3 \le d_{G_n}(i)\left(d_{G_n}(i)-1\right)\left(d_{G_n}(i)-2\right)=6\binom{d_{G_n}(i)}{3}, ( d G n ( i ) − 3 ) 3 ≤ d G n ( i ) ( d G n ( i ) − 1 ) ( d G n ( i ) − 2 ) = 6 ( 3 d G n ( i ) ) ,
for any d G n ( i ) ≥ 1 d_{G_n}(i) \ge1 d G n ( i ) ≥ 1 . This yields,
N ( K 4 , G n ) = ∑ i = 1 n ( d G n ( i ) 4 ) ≤ 6 1 / 3 4 ∑ i = 1 n ( d G n ( i ) 3 ) 4 / 3 ≤ 6 1 / 3 4 ( ∑ i = 1 n ( d G n ( i ) 3 ) ) 4 / 3 = 6 1 / 3 4 ( N ( K 1 , 2 , G n ) ) 4 / 3 . N(K_4,G_n)=\sum_{i=1}^{n}\binom{d_{G_n}(i)}{4}\le\frac{6^{1/3}}{4}\sum_{i=1}^{n}\binom{d_{G_n}(i)}{3}^{4/3}\le\frac{6^{1/3}}{4}\left(\sum_{i=1}^{n}\binom{d_{G_n}(i)}{3}\right)^{4/3}=\frac{6^{1/3}}{4}\left(N(K_{1,2},G_n)\right)^{4/3}. N ( K 4 , G n ) = i = 1 ∑ n ( 4 d G n ( i ) ) ≤ 4 6 1/3 i = 1 ∑ n ( 3 d G n ( i ) ) 4/3 ≤ 4 6 1/3 ( i = 1 ∑ n ( 3 d G n ( i ) ) ) 4/3 = 4 6 1/3 ( N ( K 1 , 2 , G n ) ) 4/3 .
Hence,
N ( K 4 , G n ) ( N ( K 2 , G n ) + N ( K 1 , 2 , G n ) ) 2 ≤ 6 1 / 3 4 ( N ( K 1 , 2 , G n ) ( N ( K 2 , G n ) + N ( K 1 , 2 , G n ) ) 3 / 2 ) 4 / 3 = o ( 1 ) . (56) \frac{N(K_4,G_n)}{\left(N(K_2,G_n)+N(K_{1,2},G_n)\right)^2}\le\frac{6^{1/3}}{4}\left(\frac{N(K_{1,2},G_n)}{\left(N(K_2,G_n)+N(K_{1,2},G_n)\right)^{3/2}}\right)^{4/3}=o(1). \tag*{(56)} ( N ( K 2 , G n ) + N ( K 1 , 2 , G n ) ) 2 N ( K 4 , G n ) ≤ 4 6 1/3 ( ( N ( K 2 , G n ) + N ( K 1 , 2 , G n ) ) 3/2 N ( K 1 , 2 , G n ) ) 4/3 = o ( 1 ) . (56)
Again employing (15) yields the implication. Finally, since { U n 4 } n ≥ 1 \{U_n^4\}_{n\ge1} { U n 4 } n ≥ 1 is uniformly integrable by Lemma 4.1, we have (a) implies both (c) and (d). This proves all the implications in Theorem 1.1 (i).
(ii) Fix p = 1 ± 1 / 3 2 p=\frac{1\pm1/\sqrt{3}}{2} p = 2 1 ± 1/ 3 . Then (b) implies (a) follows from Theorem 3.3(i). The equivalence of (b) and (c) follows from (15) , Lemma 3.4, and (56) . Finally, since { U n 4 } n ≥ 1 \{U_n^4\}_{n\ge1} { U n 4 } n ≥ 1 is uniformly integrable by Lemma 4.1, we have (a) implies (c). This proves all the implications in Theorem 1.1 (ii).
(iii) Fix p = 1 2 p=\frac{1}{2} p = 2 1 . Then (b) implies (a) follows from Proposition 3.2. The equivalence of (b) and (c) follows from (22) in Lemma 2.3 (c). Finally, since { U n 4 } n ≥ 1 \{U_n^4\}_{n\ge1} { U n 4 } n ≥ 1 is uniformly integrable by Lemma 4.1, we have (a) implies (c). This proves all the implications in Theorem 1.1 (iii). □
To complete the proof of Theorem 1.1 we now prove Lemma 4.1.
Proof of Lemma 4.1: Let S n S_n S n be as in (13) . For 1 ≤ i ≤ n 1 \le i \le n 1 ≤ i ≤ n , define
Z i : = X i − p p ( 1 − p ) . Z_i:=\frac{X_i-p}{\sqrt{p(1-p)}}. Z i := p ( 1 − p ) X i − p .
Then S n S_n S n can be written as
S n = ( 2 p − 1 ) p ( 1 − p ) ∑ i = 1 n d G n ( i ) Z i + 2 p ( 1 − p ) ∑ { i , j } ∈ E ( G n ) Z i Z j = ∑ i = 1 n c i Z i + ∑ 1 ≤ i < j ≤ n b i j Z i Z j \begin{aligned}
S_n=(2p-1)\sqrt{p(1-p)}\sum_{i=1}^{n}d_{G_n}(i)Z_i+2p(1-p)\sum_{\{i,j\}\in E(G_n)}Z_iZ_j \\
&=\sum_{i=1}^{n}c_iZ_i+\sum_{1\leq i<j\leq n}b_{ij}Z_iZ_j
\end{aligned} S n = ( 2 p − 1 ) p ( 1 − p ) i = 1 ∑ n d G n ( i ) Z i + 2 p ( 1 − p ) { i , j } ∈ E ( G n ) ∑ Z i Z j = i = 1 ∑ n c i Z i + 1 ≤ i < j ≤ n ∑ b ij Z i Z j
where b i j = 2 p ( 1 − p ) 1 { { i , j } ∈ E ( G n ) } b_{ij}=2p(1-p)1\{\{i,j\}\in E(G_n)\} b ij = 2 p ( 1 − p ) 1 {{ i , j } ∈ E ( G n )} and c i = ( 2 p − 1 ) p ( 1 − p ) d G n ( i ) c_i=(2p-1)\sqrt{p(1-p)}d_{G_n}(i) c i = ( 2 p − 1 ) p ( 1 − p ) d G n ( i ) . Note that the { Z i } 1 ≤ i ≤ n \{Z_i\}_{1\leq i\leq n} { Z i } 1 ≤ i ≤ n and { Z i Z j } 1 ≤ i < j ≤ n \{Z_iZ_j\}_{1\leq i<j\leq n} { Z i Z j } 1 ≤ i < j ≤ n are mutually uncorrelated, and each has second moment one. Hence,
σ n 2 = Var [ S n ] = ∑ i = 1 n c i 2 + ∑ 1 ≤ i < j ≤ n b i j 2 . \sigma_n^2=\operatorname{Var}[S_n]=\sum_{i=1}^{n}c_i^2+\sum_{1\leq i<j\leq n}b_{ij}^2. σ n 2 = Var [ S n ] = i = 1 ∑ n c i 2 + 1 ≤ i < j ≤ n ∑ b ij 2 .
Since U n = S n σ n U_n=\frac{S_n}{\sigma_n} U n = σ n S n , the result in (55) then follows if we show the following:
E [ ∣ ∑ i = 1 n c i Z i ∣ 6 ] ≤ C p ′ ( ∑ i = 1 n c i 2 ) 3 and E [ ∣ ∑ 1 ≤ i < j ≤ n b i j Z i Z j ∣ 6 ] ≤ C p ′ ′ ( ∑ 1 ≤ i < j ≤ n b i j 2 ) 3 , (57) \mathbb{E}\left[\left|\sum_{i=1}^{n}c_iZ_i\right|^6\right]\leq C'_p\left(\sum_{i=1}^{n}c_i^2\right)^3\quad\text{and}\quad\mathbb{E}\left[\left|\sum_{1\leq i<j\leq n}b_{ij}Z_iZ_j\right|^6\right]\leq C''_p\left(\sum_{1\leq i<j\leq n}b_{ij}^2\right)^3, \tag*{(57)} E i = 1 ∑ n c i Z i 6 ≤ C p ′ ( i = 1 ∑ n c i 2 ) 3 and E 1 ≤ i < j ≤ n ∑ b ij Z i Z j 6 ≤ C p ′′ ( 1 ≤ i < j ≤ n ∑ b ij 2 ) 3 , (57)
for constants C p ′ , C p ′ ′ > 0 C'_p,C''_p>0 C p ′ , C p ′′ > 0 , depending only on p p p .
To establish the first inequality in (57) , note that every term in the sixth moment expansion in which an index appears exactly once has expectation zero. Hence,
E ∣ ∑ i = 1 n c i Z i ∣ 6 ≤ 90 M p 6 ( ∑ i = 1 n ∣ c i ∣ 6 + ∑ 1 ≤ i ≠ j ≤ n ∣ c i ∣ 4 ∣ c j ∣ 2 + ∑ 1 ≤ i < j ≤ n ∣ c i ∣ 3 ∣ c j ∣ 3 + ∑ 1 ≤ i < j < k ≤ n ∣ c i ∣ 2 ∣ c j ∣ 2 ∣ c k ∣ 2 ) , \mathbb{E}\left|\sum_{i=1}^{n}c_iZ_i\right|^6
\leq90M_p^6\left(\sum_{i=1}^{n}|c_i|^6+\sum_{1\leq i\ne j\leq n}|c_i|^4|c_j|^2+\sum_{1\leq i<j\leq n}|c_i|^3|c_j|^3+\sum_{1\leq i<j<k\leq n}|c_i|^2|c_j|^2|c_k|^2\right), E i = 1 ∑ n c i Z i 6 ≤ 90 M p 6 i = 1 ∑ n ∣ c i ∣ 6 + 1 ≤ i = j ≤ n ∑ ∣ c i ∣ 4 ∣ c j ∣ 2 + 1 ≤ i < j ≤ n ∑ ∣ c i ∣ 3 ∣ c j ∣ 3 + 1 ≤ i < j < k ≤ n ∑ ∣ c i ∣ 2 ∣ c j ∣ 2 ∣ c k ∣ 2 ,
where we use the bound ∣ Z i ∣ ≤ M p : = max { p , 1 − p } p ( 1 − p ) |Z_i|\leq M_p:=\frac{\max\{p,1-p\}}{\sqrt{p(1-p)}} ∣ Z i ∣ ≤ M p := p ( 1 − p ) m a x { p , 1 − p } . Using the bound ∑ i = 1 n ∣ c i ∣ r ≤ ( ∑ i = 1 n c i 2 ) r 2 \sum_{i=1}^{n}|c_i|^r\leq\left(\sum_{i=1}^{n}c_i^2\right)^{\frac{r}{2}} ∑ i = 1 n ∣ c i ∣ r ≤ ( ∑ i = 1 n c i 2 ) 2 r , for r ≥ 2 r\geq2 r ≥ 2 , the first inequality in (57) follows.
For the second inequality, define β : = 2 p ( 1 − p ) \beta:=2p(1-p) β := 2 p ( 1 − p ) and
Q n : = ∑ 1 ≤ i < j ≤ n b i j Z i Z j = β ∑ { i , j } ∈ E ( G n ) Z i Z j . Q_n:=\sum_{1\leq i<j\leq n}b_{ij}Z_iZ_j=\beta\sum_{\{i,j\}\in E(G_n)}Z_iZ_j. Q n := 1 ≤ i < j ≤ n ∑ b ij Z i Z j = β { i , j } ∈ E ( G n ) ∑ Z i Z j .
Expanding the sixth power gives
E ∣ Q n ∣ 6 = β 6 ∑ { i 1 , j 1 } , { i 2 , j 2 } , … , { i 6 , j 6 } ∈ E ( G n ) E [ ∏ r = 1 6 Z i r Z j r ] . \mathbb{E}|Q_n|^6=\beta^6\sum_{\{i_1,j_1\},\{i_2,j_2\},\ldots,\{i_6,j_6\}\in E(G_n)}\mathbb{E}\left[\prod_{r=1}^{6}Z_{i_r}Z_{j_r}\right]. E ∣ Q n ∣ 6 = β 6 { i 1 , j 1 } , { i 2 , j 2 } , … , { i 6 , j 6 } ∈ E ( G n ) ∑ E [ r = 1 ∏ 6 Z i r Z j r ] .
Associate with each ordered tuple of edges { i 1 , j 1 } , { i 2 , j 2 } , … , { i 6 , j 6 } \{i_1,j_1\},\{i_2,j_2\},\ldots,\{i_6,j_6\} { i 1 , j 1 } , { i 2 , j 2 } , … , { i 6 , j 6 } the loop-less multigraph H H H formed by these six edges, retaining their multiplicities. Then
E H : = E [ ∏ r = 1 6 Z i r Z j r ] = ∏ v ∈ V ( H ) E [ Z d H ( v ) ] , E_H:=\mathbb{E}\left[\prod_{r=1}^{6}Z_{i_r}Z_{j_r}\right]=\prod_{v\in V(H)}\mathbb{E}[Z^{d_H(v)}], E H := E [ r = 1 ∏ 6 Z i r Z j r ] = v ∈ V ( H ) ∏ E [ Z d H ( v ) ] ,
where d H ( v ) d_H(v) d H ( v ) is the degree of the vertex v v v in H H H . Note that E H = 0 E_H=0 E H = 0 whenever H H H has a vertex of degree one and ∣ E H ∣ ≤ M p 12 |E_H|\leq M_p^{12} ∣ E H ∣ ≤ M p 12 . Denote by H 6 \mathcal{H}_6 H 6 the collection of all multigraphs with six edges up to isomorphism such that min v ∈ V ( H ) d H ( v ) ≥ 2 \min_{v\in V(H)}d_H(v)\geq2 min v ∈ V ( H ) d H ( v ) ≥ 2 . Then, by the above discussion,
E ∣ Q n ∣ 6 ≤ A p ∑ H ∈ H 6 N ( H , G n ) , (58) \mathbb{E}|Q_n|^6\leq A_p\sum_{H\in\mathcal{H}_6}N(H,G_n), \tag*{(58)} E ∣ Q n ∣ 6 ≤ A p H ∈ H 6 ∑ N ( H , G n ) , (58)
for some constant A p > 0 A_p>0 A p > 0 . We claim that there exists a constant C > 0 C>0 C > 0 such that
N ( H , G n ) ≤ C N ( ∙ , G n ) 3 , for every H ∈ H 6 . (59) N(H,G_n)\leq C N(\bullet,G_n)^3,\qquad\text{for every }H\in\mathcal{H}_6. \tag*{(59)} N ( H , G n ) ≤ C N ( ∙ , G n ) 3 , for every H ∈ H 6 . (59)
Note that the claim follows immediately if every H ∈ H 6 H\in\mathcal{H}_6 H ∈ H 6 has an edge-cover consisting of at most three edges. (Recall that an edge cover of H H H is a collection of edges of H H H incident to every vertex.) Through the following cases we show that such an edge cover exists, with one exception.
∣ V ( H ) ∣ ≤ 4 |V(H)|\leq4 ∣ V ( H ) ∣ ≤ 4 : In this case a spanning forest of its underlying simple graph is an edge cover with at most three edges.
∣ V ( H ) ∣ = 5 |V(H)|=5 ∣ V ( H ) ∣ = 5 . In this case, the underlying simple graph contains a matching of size two. If not, the edges of the underlying simple graph would be pairwise intersect. A family of pairwise intersecting edges either has a common endpoint or is contained in a triangle. Since all five vertices are incident to an edge, the underlying graph would therefore be the 4 4 4 -star K 1 , 4 K_{1,4} K 1 , 4 . Each of its four leaves would require an incident edge of multiplicity at least two, contradicting the fact that H H H has only six edges. Hence, the two edges forming the matching together with any edge incident to the remaining vertex, give the required three-edge cover.
If ∣ V ( H ) ∣ = 6 |V(H)|=6 ∣ V ( H ) ∣ = 6 , then every vertex of H H H has degree exactly two, because the sum of the degrees is twelve. Hence, H H H is a disjoint union of cycles, with a doubled edge regarded as a cycle of length two. Therefore, the possibilities are a 6 6 6 -cycle, a disjoint union of a 4 4 4 -cycle and a 2 2 2 -cycle, a disjoint union of two 3 3 3 -cycles ( △ ⊔ △ ) (\triangle\sqcup\triangle) ( △ ⊔ △ ) , and a disjoint union of three 2 2 2 -cycles. All of these except △ ⊔ △ \triangle\sqcup\triangle △ ⊔ △ have an edge cover consisting of three edges.
To deal with the case H = △ ⊔ △ H=\triangle\sqcup\triangle H = △ ⊔ △ , first recall that by the Kruskal-Katona theorem (see, for example, [36 Equation (2.3)] ),
N ( △ , G n ) ≤ C ′ N ( ∙ , G n ) 3 2 , N(\triangle,G_n)\leq C'N(\bullet,G_n)^{\frac{3}{2}}, N ( △ , G n ) ≤ C ′ N ( ∙ , G n ) 2 3 ,
for some constant C ′ > 0 C'>0 C ′ > 0 . Hence,
N ( △ ⊔ △ , G n ) ≤ 2 N ( △ , G n ) 2 ≤ C N ( ∙ , G n ) 3 . N(\triangle\sqcup\triangle,G_n)\leq2N(\triangle,G_n)^2\leq C N(\bullet,G_n)^3. N ( △ ⊔ △ , G n ) ≤ 2 N ( △ , G n ) 2 ≤ C N ( ∙ , G n ) 3 .
The above discussion shows that (59) holds for H ∈ H 6 H\in\mathcal{H}_6 H ∈ H 6 . Therefore, using (58) ,
E ∣ Q n ∣ 6 ≤ C A p N ( ∙ , G n ) 3 = C p ′ ′ ( ∑ 1 ≤ i < j ≤ n b i j 2 ) 3 , \mathbb{E}|Q_n|^6\leq C A_p N(\bullet,G_n)^3=C_p''\left(\sum_{1\leq i<j\leq n}b_{ij}^2\right)^3, E ∣ Q n ∣ 6 ≤ C A p N ( ∙ , G n ) 3 = C p ′′ ( 1 ≤ i < j ≤ n ∑ b ij 2 ) 3 ,
which proves the second inequality in (57) .
Proof of thm:cltH0
Recall that given L n ∼ Poi ( n 1 + n 2 ) L_n\sim\operatorname{Poi}(n_1+n_2) L n ∼ Poi ( n 1 + n 2 ) , Z n = { Z 1 , Z 2 , … , Z L n } \mathcal{Z}_n=\{Z_1,Z_2,\ldots,Z_{L_n}\} Z n = { Z 1 , Z 2 , … , Z L n } is an i.i.d. collection from the distribution with density n 1 n f + n 2 n g \frac{n_1}{n}f+\frac{n_2}{n}g n n 1 f + n n 2 g . Then Z n \mathcal{Z}_n Z n is a nonhomogeneous Poisson process in R d \mathbb{R}^d R d with rate function n ϕ n = n 1 f + n 2 g n\phi_n=n_1f+n_2g n ϕ n = n 1 f + n 2 g . Denote by G K ( Z n ) G_K(\mathcal{Z}_n) G K ( Z n ) the K K K -NN graph on the pooled sample Z n \mathcal{Z}_n Z n . The following establishes bounds on the relevant motif counts in G K ( Z n ) G_K(\mathcal{Z}_n) G K ( Z n ) .
Lemma 5.1. For an integer K ≥ 1 K\geq1 K ≥ 1 , let K ~ : = min { K , max { L n − 1 , 0 } } \widetilde K:=\min\{K,\max\{L_n-1,0\}\} K := min { K , max { L n − 1 , 0 }} . There exists a constant B d > 0 B_d>0 B d > 0 , depending only on d d d , such that the following hold almost surely,
1 2 L n K ~ ≤ N ( ∙ , G K ( Z n ) ) ≤ L n K ~ \frac{1}{2}L_n\widetilde K\leq N(\bullet,G_K(\mathcal{Z}_n))\leq L_n\widetilde K 2 1 L n K ≤ N ( ∙ , G K ( Z n )) ≤ L n K ,
L n ( K ~ 2 ) ≤ N ( ∨ , G K ( Z n ) ) ≤ 1 2 B d 2 L n K ~ 2 L_n\binom{\widetilde K}{2}\leq N(\mathord{\vee},G_K(\mathcal{Z}_n))\leq\frac{1}{2}B_d^2L_n\widetilde K^2 L n ( 2 K ) ≤ N ( ∨ , G K ( Z n )) ≤ 2 1 B d 2 L n K 2 ,
(iii) N ( ∙ − ∙ , G K ( Z n ) ) ≤ 1 24 B d 4 L n K ~ 4 N(\bullet\!-\!\bullet,\mathcal{G}_K(\mathbb{Z}_n)) \le\frac{1}{24}B_d^4L_n\widetilde{K}^4 N ( ∙ − ∙ , G K ( Z n )) ≤ 24 1 B d 4 L n K 4 .
Further, for every K ≥ 1 K \ge1 K ≥ 1 ,
1 2 L n K ~ 2 ≤ N ( ∙ − ∙ , G K ( Z n ) ) + N ( ∨ , G K ( Z n ) ) ≤ 1 2 B d 2 L n K ~ 2 . (60) \frac{1}{2}L_n\widetilde{K}^2 \le N(\bullet\!-\!\bullet,\mathcal{G}_K(\mathbb{Z}_n)) + N(\vee,\mathcal{G}_K(\mathbb{Z}_n)) \le\frac{1}{2}B_d^2L_n\widetilde{K}^2. \tag*{(60)} 2 1 L n K 2 ≤ N ( ∙ − ∙ , G K ( Z n )) + N ( ∨ , G K ( Z n )) ≤ 2 1 B d 2 L n K 2 . (60)
Proof. Note that the bounds are immediate if L n ≤ 1 L_n \le1 L n ≤ 1 . Therefore, we assume L n ≥ 2 L_n \ge2 L n ≥ 2 . We will use the following standard degree-bound: For x ∈ Z n x \in\mathbb{Z}_n x ∈ Z n , the degree d x d_x d x of the vertex x x x in the graph G K ( Z n ) \mathcal{G}_K(\mathbb{Z}_n) G K ( Z n ) satisfies:
K ~ ≤ d x ≤ B d K ~ , (61) \widetilde{K} \le d_x \le B_d\widetilde{K}, \tag*{(61)} K ≤ d x ≤ B d K , (61)
for some constant B d > 0 B_d > 0 B d > 0 , depending only on d d d . (Note that the lower bound is immediate from the definition and for the upper bound see [50 Proposition 12, Section 11] and also Györfi et al. [26 Chapter 6] ).
Using the degree bound the proof can be completed as follows:
There are exactly L n K ~ L_n\widetilde{K} L n K directed nearest-neighbor relations. Each undirected edge corresponds to either one or two such relations, giving 1 2 L n K ~ ≤ N ( ∙ − ∙ , G K ( Z n ) ) ≤ L n K ~ \frac{1}{2}L_n\widetilde{K} \le N(\bullet\!-\!\bullet,\mathcal{G}_K(\mathbb{Z}_n)) \le L_n\widetilde{K} 2 1 L n K ≤ N ( ∙ − ∙ , G K ( Z n )) ≤ L n K .
L n ( K ~ 2 ) ≤ N ( ∨ , G K ( Z n ) ) = ∑ x ∈ Z n ( d x 2 ) ≤ B d 2 2 L n K ~ 2 . L_n\binom{\widetilde{K}}{2} \le N(\vee,\mathcal{G}_K(\mathbb{Z}_n)) = \sum_{x\in\mathbb{Z}_n}\binom{d_x}{2} \le\frac{B_d^2}{2}L_n\widetilde{K}^2. L n ( 2 K ) ≤ N ( ∨ , G K ( Z n )) = x ∈ Z n ∑ ( 2 d x ) ≤ 2 B d 2 L n K 2 .
N ( K 4 , G K ( Z n ) ) = ∑ x ∈ Z n ( d x 4 ) ≤ B d 4 24 L n K ~ 4 . N(K_4,\mathcal{G}_K(\mathbb{Z}_n)) = \sum_{x\in\mathbb{Z}_n}\binom{d_x}{4} \le\frac{B_d^4}{24}L_n\widetilde{K}^4. N ( K 4 , G K ( Z n )) = x ∈ Z n ∑ ( 4 d x ) ≤ 24 B d 4 L n K 4 .
Also, (60) follows from the degree bound and the observation:
N ( ∙ − ∙ , G K ( Z n ) ) + N ( ∨ , G K ( Z n ) ) = 1 2 ∑ x ∈ Z n d x 2 . N(\bullet\!-\!\bullet,\mathcal{G}_K(\mathbb{Z}_n)) + N(\vee,\mathcal{G}_K(\mathbb{Z}_n)) = \frac{1}{2}\sum_{x\in\mathbb{Z}_n}d_x^2. N ( ∙ − ∙ , G K ( Z n )) + N ( ∨ , G K ( Z n )) = 2 1 x ∈ Z n ∑ d x 2 .
This completes the proof.
To prove Theorem 1.6 it is more convenient to work with T ∘ ( G K ( Z n ) ) T^\circ(\mathcal{G}_K(\mathbb{Z}_n)) T ∘ ( G K ( Z n )) , the number of monochromatic edges (recall (8) ). Note that
U n × ( G K ( Z n ) ) = T × ( G K ( Z n ) ) − μ F × σ F = T ∘ ( G K ( Z n ) ) − μ F ∘ σ F , U_n^\times(\mathcal{G}_K(\mathbb{Z}_n)) = \frac{T^\times(\mathcal{G}_K(\mathbb{Z}_n))-\mu^\times_{\mathcal{F}}}{\sigma_{\mathcal{F}}} = \frac{T^\circ(\mathcal{G}_K(\mathbb{Z}_n))-\mu^\circ_{\mathcal{F}}}{\sigma_{\mathcal{F}}}, U n × ( G K ( Z n )) = σ F T × ( G K ( Z n )) − μ F × = σ F T ∘ ( G K ( Z n )) − μ F ∘ ,
where
μ F ∘ : = E [ T ∘ ( G K ( Z n ) ) ∣ F ] = ( λ n 2 + ( 1 − λ n ) 2 ) N ( ∙ − ∙ , G K ( Z n ) ) = N ( ∙ − ∙ , G K ( Z n ) ) − μ F × . \mu^\circ_{\mathcal{F}} := \mathbb{E}[T^\circ(\mathcal{G}_K(\mathbb{Z}_n))\mid\mathcal{F}] = (\lambda_n^2+(1-\lambda_n)^2)N(\bullet\!-\!\bullet,\mathcal{G}_K(\mathbb{Z}_n)) = N(\bullet\!-\!\bullet,\mathcal{G}_K(\mathbb{Z}_n))-\mu^\times_{\mathcal{F}}. μ F ∘ := E [ T ∘ ( G K ( Z n )) ∣ F ] = ( λ n 2 + ( 1 − λ n ) 2 ) N ( ∙ − ∙ , G K ( Z n )) = N ( ∙ − ∙ , G K ( Z n )) − μ F × .
We will now apply the results in § 3 by conditioning on sigma-algebra F = σ ( L n , Z n ) \mathcal{F}=\sigma(L_n,\mathbb{Z}_n) F = σ ( L n , Z n ) . For this, we will require the following notion of the conditional Wasserstein distance.
Definition 5.2. Given two real-valued random variables X X X and Z Z Z with finite first moments, and a σ \sigma σ -algebra F \mathcal{F} F , the Wasserstein distance between X ∣ F X \mid\mathcal{F} X ∣ F (the conditional law of X X X given the σ \sigma σ -algebra F \mathcal{F} F ) and Z Z Z is defined as
Wass ( X ∣ F , Z ) : = sup g ∈ L ∣ E [ g ( X ) ∣ F ] − E [ g ( Z ) ] ∣ , \operatorname{Wass}(X \mid\mathcal{F},Z):=\sup_{g\in\mathcal{L}}\left|\mathbb{E}[g(X)\mid\mathcal{F}]-\mathbb{E}[g(Z)]\right|, Wass ( X ∣ F , Z ) := g ∈ L sup ∣ E [ g ( X ) ∣ F ] − E [ g ( Z )] ∣ , where L \mathcal{L} L is the collection of 1-Lipschitz functions from R → R \mathbb{R}\to\mathbb{R} R → R .
Recall that λ n = n 1 n \lambda_n=\frac{n_1}{n} λ n = n n 1 . First we consider the case λ n → λ ∈ ( 0 , 1 ) ∖ { 1 2 } \lambda_n\to\lambda\in(0,1)\setminus\left\{\frac{1}{2}\right\} λ n → λ ∈ ( 0 , 1 ) ∖ { 2 1 } . Note that conditional on Z n Z_n Z n , G K ( Z n ) \mathcal{G}_K(Z_n) G K ( Z n ) is a fixed graph with an i.i.d. Ber ( λ n ) \operatorname{Ber}(\lambda_n) Ber ( λ n ) vertex coloring. Hence, from (42) , whenever L n ≥ 2 L_n\ge2 L n ≥ 2 ,
Wass ( U n × ( G K ( Z n ) ) ∣ F , N ( 0 , 1 ) ) ≤ C λ ( 1 N ( ∙ − ∙ , G K ( Z n ) ) + N ( ∙ − ∙ , G K ( Z n ) ) + N ( ∙ − ∙ , G K ( Z n ) ) N ( ∙ − ∙ , G K ( Z n ) ) + N ( ∙ − ∙ , G K ( Z n ) ) ) . \operatorname{Wass}\left(U_n^\times(\mathcal{G}_K(Z_n))\mid\mathcal{F},N(0,1)\right)\le C_\lambda\left(\frac{1}{\sqrt{N(\bullet\!\!-\!\!\bullet,\mathcal{G}_K(Z_n))+N(\bullet\!\!-\!\!\bullet,\mathcal{G}_K(Z_n))}}+\frac{\sqrt{N(\bullet\!\!-\!\!\bullet,\mathcal{G}_K(Z_n))}}{N(\bullet\!\!-\!\!\bullet,\mathcal{G}_K(Z_n))+N(\bullet\!\!-\!\!\bullet,\mathcal{G}_K(Z_n))}\right). Wass ( U n × ( G K ( Z n )) ∣ F , N ( 0 , 1 ) ) ≤ C λ ( N ( ∙ − ∙ , G K ( Z n )) + N ( ∙ − ∙ , G K ( Z n )) 1 + N ( ∙ − ∙ , G K ( Z n )) + N ( ∙ − ∙ , G K ( Z n )) N ( ∙ − ∙ , G K ( Z n )) ) .
where C λ > 0 C_\lambda>0 C λ > 0 is a constant depending only on λ \lambda λ . Applying the bounds in Lemma 5.1 then gives,
Wass ( U n × ( G K ( Z n ) ) ∣ F , N ( 0 , 1 ) ) ≤ C λ , d L n , (62) \operatorname{Wass}\left(U_n^\times(\mathcal{G}_K(Z_n))\mid\mathcal{F},N(0,1)\right)\le\frac{C_{\lambda,d}}{\sqrt{L_n}}, \tag*{(62)} Wass ( U n × ( G K ( Z n )) ∣ F , N ( 0 , 1 ) ) ≤ L n C λ , d , (62)
almost surely, for some constant C λ , d > 0 C_{\lambda,d}>0 C λ , d > 0 , that depends only on d d d and λ \lambda λ .
Next, we consider the case where n 1 = n 2 n_1=n_2 n 1 = n 2 . This means, λ n = 1 2 \lambda_n=\frac{1}{2} λ n = 2 1 . In this case, by (28) , whenever L n ≥ 2 L_n\ge2 L n ≥ 2 ,
Wass ( U n × ( G K ( Z n ) ) ∣ F , N ( 0 , 1 ) ) ≤ 4 C ( N ( ∙ − ∙ , G K ( Z n ) ) − 1 / 2 + N ( ■ , G K ( Z n ) ) N ( ∙ − ∙ , G K ( Z n ) ) 2 ) 1 / 10 . \operatorname{Wass}\left(U_n^\times(\mathcal{G}_K(Z_n))\mid\mathcal{F},N(0,1)\right)\le4\sqrt{C}\left(N(\bullet\!\!-\!\!\bullet,\mathcal{G}_K(Z_n))^{-1/2}+\frac{N(\blacksquare,\mathcal{G}_K(Z_n))}{N(\bullet\!\!-\!\!\bullet,\mathcal{G}_K(Z_n))^2}\right)^{1/10}. Wass ( U n × ( G K ( Z n )) ∣ F , N ( 0 , 1 ) ) ≤ 4 C ( N ( ∙ − ∙ , G K ( Z n ) ) − 1/2 + N ( ∙ − ∙ , G K ( Z n ) ) 2 N ( ■ , G K ( Z n )) ) 1/10 .
Applying the bounds in Lemma 5.1 and N ( ■ , G K ( Z n ) ) = O ( N ( ∙ − ∙ , G K ( Z n ) ) K ~ 2 ) = O ( L n K ~ 3 ) N(\blacksquare,\mathcal{G}_K(Z_n))=O(N(\bullet\!\!-\!\!\bullet,\mathcal{G}_K(Z_n))\widetilde{K}^2)=O(L_n\widetilde{K}^3) N ( ■ , G K ( Z n )) = O ( N ( ∙ − ∙ , G K ( Z n )) K 2 ) = O ( L n K 3 ) gives,
Wass ( U n × ( G K ( Z n ) ) ∣ F , N ( 0 , 1 ) ) ≤ C ‾ d ( 1 L n K ~ + K ~ L n ) 1 / 10 , (63) \operatorname{Wass}\left(U_n^\times(\mathcal{G}_K(Z_n))\mid\mathcal{F},N(0,1)\right)\le\overline{C}_d\left(\frac{1}{\sqrt{L_n\widetilde{K}}}+\frac{\widetilde{K}}{L_n}\right)^{1/10}, \tag*{(63)} Wass ( U n × ( G K ( Z n )) ∣ F , N ( 0 , 1 ) ) ≤ C d ( L n K 1 + L n K ) 1/10 , (63)
where C ‾ d > 0 \overline{C}_d>0 C d > 0 is a constant depending on d d d .
To remove the conditioning, define E n : = { L n ≥ n / 2 } \mathcal{E}_n:=\{L_n\ge n/2\} E n := { L n ≥ n /2 } . Then,
Wass ( U n × ( G K ( Z n ) ) , N ( 0 , 1 ) ) ≤ E [ Wass ( U n × ( G K ( Z n ) ) ∣ F , N ( 0 , 1 ) ) 1 { E n } ] + 2 P ( E n c ) , \operatorname{Wass}\left(U_n^\times(\mathcal{G}_K(Z_n)),N(0,1)\right)\le\mathbb{E}\left[\operatorname{Wass}\left(U_n^\times(\mathcal{G}_K(Z_n))\mid\mathcal{F},N(0,1)\right)\mathbf{1}\{\mathcal{E}_n\}\right]+2\mathbb{P}(\mathcal{E}_n^c), Wass ( U n × ( G K ( Z n )) , N ( 0 , 1 ) ) ≤ E [ Wass ( U n × ( G K ( Z n )) ∣ F , N ( 0 , 1 ) ) 1 { E n } ] + 2 P ( E n c ) ,
where for the second term we use the bound Wass ( U n × ( G K ( Z n ) ) ∣ F , N ( 0 , 1 ) ) ≤ 2 \operatorname{Wass}(U_n^\times(\mathcal{G}_K(Z_n))\mid\mathcal{F},N(0,1))\le2 Wass ( U n × ( G K ( Z n )) ∣ F , N ( 0 , 1 )) ≤ 2 . Since E [ L n ] = Var [ L n ] = n \mathbb{E}[L_n]=\operatorname{Var}[L_n]=n E [ L n ] = Var [ L n ] = n , Chebyshev’s inequality gives P ( E n c ) ≤ P ( ∣ L n − n ∣ ≥ n / 2 ) ≤ 4 n \mathbb{P}(\mathcal{E}_n^c)\le\mathbb{P}(|L_n-n|\ge n/2)\le\frac{4}{n} P ( E n c ) ≤ P ( ∣ L n − n ∣ ≥ n /2 ) ≤ n 4 . Combining the above with (62) shows that Wass ( U n × ( G K ( Z n ) ) , N ( 0 , 1 ) ) → 0 \operatorname{Wass}(U_n^\times(\mathcal{G}_K(Z_n)),N(0,1))\to0 Wass ( U n × ( G K ( Z n )) , N ( 0 , 1 )) → 0 , for λ ≠ 1 2 \lambda\ne\frac{1}{2} λ = 2 1 and irrespective of the choice of K K K . For n 1 = n 2 n_1=n_2 n 1 = n 2 , the same argument with (63) shows that Wass ( U n × ( G K ( Z n ) ) , N ( 0 , 1 ) ) → 0 \operatorname{Wass}(U_n^\times(\mathcal{G}_K(Z_n)),N(0,1))\to0 Wass ( U n × ( G K ( Z n )) , N ( 0 , 1 )) → 0 , whenever K = o ( n ) K=o(n) K = o ( n ) . □
Examples
In this section, we present various examples illustrating the necessity of the conditions in the preceding results. We begin with an example showing that the fourth moment condition is not enough for the asymptotic normality of U n U_n U n when p = 1 ± 1 / 3 2 p=\frac{1\pm1/\sqrt{3}}{2} p = 2 1 ± 1/ 3 . Recall from Theorem 1.1 (ii), that for these two values of p p p , the third moment condition E [ U n 3 ] → 0 \mathbb{E}[U_n^3]\to0 E [ U n 3 ] → 0 is necessary and sufficient for the asymptotic normality of U n U_n U n . Departing slightly from the notation in the preceding sections, we use n n n here to index the graph sequence rather than to denote the number of vertices.
Example 6.1. (The fourth moment condition does not suffice for p = 1 ± 1 / 3 2 p=\frac{1\pm1/\sqrt{3}}{2} p = 2 1 ± 1/ 3 .) For n ≥ 1 n\ge1 n ≥ 1 , let G n = K 1 , n G_n=K_{1,n} G n = K 1 , n on the vertex set { 0 , 1 , … , n } \{0,1,\ldots,n\} { 0 , 1 , … , n } , with the central vertex labeled 0. Assume p = 1 + 1 / 3 2 p=\frac{1+1/\sqrt{3}}{2} p = 2 1 + 1/ 3 . The case p = 1 − 1 / 3 2 p=\frac{1-1/\sqrt{3}}{2} p = 2 1 − 1/ 3 is similar. Then E [ U n 4 ] → 3 \mathbb{E}\left[U_n^4\right]\to3 E [ U n 4 ] → 3 , from Lemma 2.3 (b). However, in this case U n U_n U n does not have a limiting normal distribution. In particular, we claim that, as n → ∞ n\to\infty n → ∞ ,
U n → D p δ a + ( 1 − p ) δ b , where a = 3 − 1 2 and b = − 3 + 1 2 . (64) U_n\xrightarrow{\mathcal{D}}p\delta_a+(1-p)\delta_b,\qquad\text{where }a=\frac{\sqrt{3}-1}{\sqrt{2}}\text{ and }b=-\frac{\sqrt{3}+1}{\sqrt{2}}. \tag*{(64)} U n D p δ a + ( 1 − p ) δ b , where a = 2 3 − 1 and b = − 2 3 + 1 . (64) To see this, define R i = 2 X i − 1 R_i=2X_i-1 R i = 2 X i − 1 , for 0 ≤ i ≤ n 0\le i\le n 0 ≤ i ≤ n . Note that { R i } 0 ≤ i ≤ n \{R_i\}_{0\le i\le n} { R i } 0 ≤ i ≤ n are i.i.d. ± 1 \pm1 ± 1 -valued with mean 1 / 3 1/\sqrt{3} 1/ 3 and variance 2 / 3 2/3 2/3 . Since 1 { X 0 = X i } = ( 1 + R 0 R i ) / 2 \mathbb{1}\{X_0=X_i\}=(1+R_0R_i)/2 1 { X 0 = X i } = ( 1 + R 0 R i ) /2 and E [ R 0 R i ] = 1 / 3 \mathbb{E}[R_0R_i]=1/3 E [ R 0 R i ] = 1/3 ,
U n = 1 2 σ n ∑ i = 1 n ( R 0 R i − 1 3 ) = R 0 2 σ n ( ∑ i = 1 n R i − n 3 ) + ( n R 0 / 3 − n / 3 ) 2 σ n . U_n=\frac{1}{2\sigma_n}\sum_{i=1}^{n}\left(R_0R_i-\frac{1}{3}\right)=\frac{R_0}{2\sigma_n}\left(\sum_{i=1}^{n}R_i-\frac{n}{\sqrt{3}}\right)+\frac{(nR_0/\sqrt{3}-n/3)}{2\sigma_n}. U n = 2 σ n 1 i = 1 ∑ n ( R 0 R i − 3 1 ) = 2 σ n R 0 ( i = 1 ∑ n R i − 3 n ) + 2 σ n ( n R 0 / 3 − n /3 ) . Note that
( n R 0 / 3 − n / 3 ) 2 σ n → D p δ a + ( 1 − p ) δ b , \frac{(nR_0/\sqrt{3}-n/3)}{2\sigma_n}\xrightarrow{\mathcal{D}}p\delta_a+(1-p)\delta_b, 2 σ n ( n R 0 / 3 − n /3 ) D p δ a + ( 1 − p ) δ b , for a a a and b b b as defined in (64) . Further, by the classical CLT,
∑ i = 1 n R i − n / 3 2 n / 3 → D N ( 0 , 1 ) . \frac{\sum_{i=1}^{n}R_i-n/\sqrt{3}}{\sqrt{2n/3}}\xrightarrow{\mathcal{D}}N(0,1). 2 n /3 ∑ i = 1 n R i − n / 3 D N ( 0 , 1 ) . Since, by (3) , σ n = 1 3 2 n + o ( n ) \sigma_n=\frac{1}{3\sqrt{2}}n+o(n) σ n = 3 2 1 n + o ( n ) , this implies,
R 0 2 σ n ( ∑ i = 1 n R i − n 3 ) → P 0. \frac{R_0}{2\sigma_n}\left(\sum_{i=1}^{n}R_i-\frac{n}{\sqrt{3}}\right)\xrightarrow{\mathbb{P}}0. 2 σ n R 0 ( i = 1 ∑ n R i − 3 n ) P 0. Combining the above proves the result in (64) .
Next, we show that the third moment condition is not enough for the asymptotic normality of U n U_n U n when p = 1 2 p=\frac{1}{2} p = 2 1 . Recall from Theorem 1.1 (iii) that in this case, the fourth moment condition E [ U n 4 ] → 3 \mathbb{E}\left[U_n^4\right]\to3 E [ U n 4 ] → 3 is necessary and sufficient for the asymptotic normality of U n U_n U n .
Example 6.2. (The third moment condition does not suffice for p = 1 2 p=\frac{1}{2} p = 2 1 .) For n ≥ 1 n\ge1 n ≥ 1 , let G n = K 2 , n G_n=K_{2,n} G n = K 2 , n , the complete bipartite graph with two vertices on one side and n n n vertices on the other side. Label the vertices in the two sides by { 1 ′ , 2 ′ } \{1',2'\} { 1 ′ , 2 ′ } and { 1 , … , n } \{1,\ldots,n\} { 1 , … , n } , respectively. Assume p = 1 2 p=\frac{1}{2} p = 2 1 . Then, since K 2 , n K_{2,n} K 2 , n is triangle-free, by Lemma 2.1 (b), E [ U n 4 ] → 0 \mathbb{E}\left[U_n^4\right]\to0 E [ U n 4 ] → 0 . However, in this case U n U_n U n does not have a limiting normal distribution. To see this, note that there are two possibilities:
If the vertices { 1 ′ , 2 ′ } \{1',2'\} { 1 ′ , 2 ′ } receive different colors (which happens with probability 1 2 \frac{1}{2} 2 1 ), then the number of monochromatic edges T n = n T_n=n T n = n .
If the vertices { 1 ′ , 2 ′ } \{1',2'\} { 1 ′ , 2 ′ } receive the same color (which also happens with probability 1 2 \frac{1}{2} 2 1 ), then the number of monochromatic edges T n = D 2 M n T_n\overset{\mathcal{D}}{=}2M_n T n = D 2 M n , where M n ∼ Bin ( n , 1 2 ) M_n\sim\operatorname{Bin}(n,\frac{1}{2}) M n ∼ Bin ( n , 2 1 ) .
Hence, E [ T n ] = n \mathbb{E}[T_n]=n E [ T n ] = n , σ n 2 = Var [ T n ] = n 2 \sigma_n^2=\operatorname{Var}[T_n]=\frac{n}{2} σ n 2 = Var [ T n ] = 2 n (recall (3) ), and
U n = T n − E [ T n ] σ n → D 1 2 N ( 0 , 2 ) + 1 2 δ 0 , U_n=\frac{T_n-\mathbb{E}[T_n]}{\sigma_n}\xrightarrow{\mathcal{D}}\frac{1}{2}N(0,2)+\frac{1}{2}\delta_0, U n = σ n T n − E [ T n ] D 2 1 N ( 0 , 2 ) + 2 1 δ 0 , which is a mixture of a normal distribution and a point mass at 0.
The next example shows that in the sparse regime p n → 0 p_n\to0 p n → 0 , as n → ∞ n\to\infty n → ∞ , the third moment condition, although sufficient, is not necessary for the asymptotic normality of U n U_n U n .
Example 6.3 (Third moment condition not necessary in the sparse regime). Suppose G n G_n G n consists of an n 2 n^2 n 2 -star K 1 , n 2 K_{1,n^2} K 1 , n 2 and n 5 n^5 n 5 disjoint edges. Choose p n = 1 n 3 p_n = \frac{1}{n^3} p n = n 3 1 . In this case, it is easy to check that condition (b) of Lemma 2.2 does not hold. Hence, E [ U n 3 ] ↛ 0 \mathbb{E}[U_n^3] \nrightarrow0 E [ U n 3 ] ↛ 0 . In particular, by direct calculation using Table 2 it follows that E [ U n 3 ] → − 1 2 2 \mathbb{E}[U_n^3] \to-\frac{1}{2\sqrt{2}} E [ U n 3 ] → − 2 2 1 . However, U n U_n U n is asymptotically normal. To see this, recall the definition of T n T_n T n from (1) and denote
T n = A n + B n , T_n = A_n + B_n, T n = A n + B n , where A n A_n A n and B n B_n B n count the number of monochromatic edges in the star and the disjoint edges, respectively. Let u n = ( 2 p n − 1 ) 2 u_n=(2p_n-1)^2 u n = ( 2 p n − 1 ) 2 . Note that B n ∼ Bin ( n 5 , p n 2 + ( 1 − p n ) 2 ) B_n\sim\operatorname{Bin}(n^5,p_n^2+(1-p_n)^2) B n ∼ Bin ( n 5 , p n 2 + ( 1 − p n ) 2 ) and Var [ B n ] ∼ 2 n 2 \operatorname{Var}[B_n]\sim2n^2 Var [ B n ] ∼ 2 n 2 . Also, from (31) we have,
Var [ A n ] = p n ( 1 − p n ) ( n 2 + u n n 4 ) ∼ n . \operatorname{Var}[A_n]=p_n(1-p_n)(n^2+u_n n^4)\sim n. Var [ A n ] = p n ( 1 − p n ) ( n 2 + u n n 4 ) ∼ n . Since A n A_n A n and B n B_n B n are independent, σ n 2 = Var [ T n ] ∼ 2 n 2 \sigma_n^2=\operatorname{Var}[T_n]\sim2n^2 σ n 2 = Var [ T n ] ∼ 2 n 2 . Hence,
E [ ( A n − E [ A n ] σ n ) 2 ] = Var [ A n ] σ n 2 → 0. \mathbb{E}\left[\left(\frac{A_n-\mathbb{E}[A_n]}{\sigma_n}\right)^2\right]=\frac{\operatorname{Var}[A_n]}{\sigma_n^2}\to0. E [ ( σ n A n − E [ A n ] ) 2 ] = σ n 2 Var [ A n ] → 0. Therefore, by the CLT and Slutsky’s theorem,
U n = A n − E [ A n ] σ n + B n − E [ B n ] σ n → D N ( 0 , 1 ) . U_n=\frac{A_n-\mathbb{E}[A_n]}{\sigma_n}+\frac{B_n-\mathbb{E}[B_n]}{\sigma_n}\xrightarrow{D}N(0,1). U n = σ n A n − E [ A n ] + σ n B n − E [ B n ] D N ( 0 , 1 ) . Hence, U n → D N ( 0 , 1 ) U_n\xrightarrow{D}N(0,1) U n D N ( 0 , 1 ) , even though E [ U n 3 ] ↛ 0 \mathbb{E}[U_n^3]\nrightarrow0 E [ U n 3 ] ↛ 0 .
Finally, we give an example in the sparse regime where the third moment condition holds, hence, U n U_n U n is asymptotically normal, but the fourth moment condition does not. We show in Proposition 1.3 that the fourth moment condition always implies the third moment condition in the sparse case. This together with the following example implies that the third moment condition is strictly weaker than the fourth moment condition in the sparse regime.
Example 6.4 (Third moment condition does not imply the fourth moment condition). Suppose G n G_n G n consists of an n n n -star K 1 , n K_{1,n} K 1 , n and n 3 n^3 n 3 disjoint edges. Choose p n = 1 n 2 p_n=\frac{1}{n^2} p n = n 2 1 . In this case, it is easy to check that condition (b) of Lemma 2.2 holds, which implies
E [ U n 3 ] → 0. \mathbb{E}[U_n^3]\to0. E [ U n 3 ] → 0. Hence, by Theorem 3.3 U n → D N ( 0 , 1 ) U_n\xrightarrow{D}N(0,1) U n D N ( 0 , 1 ) . However, condition (b) of Lemma 1.4 does not hold. Hence, E [ U n 4 ] ↛ 3 \mathbb{E}[U_n^4]\nrightarrow3 E [ U n 4 ] ↛ 3 . In particular, by direct calculation using Table 3 it follows that E [ U n 4 ] → 13 4 \mathbb{E}[U_n^4]\to\frac{13}{4} E [ U n 4 ] → 4 13 . This shows that the third moment condition does not imply the fourth moment condition in the sparse regime.
Simulations
In this section, we investigate the effect of K K K , the number of nearest neighbors, on the empirical performance of the K K K -NN-based two-sample test discussed in § 1.3 . We consider the following two examples. In both cases, the sample sizes are chosen to be independent Poisson random variables, each with mean 100, the nominal level α = 0.05 \alpha=0.05 α = 0.05 and K ∈ { 5 , 10 , 15 } K\in\{5,10,15\} K ∈ { 5 , 10 , 15 } . We construct a two-sample test with precise control on the asymptotic level as follows: Fix α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) and consider the test function that rejects the null hypothesis H 0 H_0 H 0 as follows:
ψ n : = 1 { T × ( G K ( Z n ) ) < μ F × − z α σ F } , (65) \psi_n:=\mathbf{1}\{T^{\times}(G_K(Z_n))<\mu_F^{\times}-z_\alpha\sigma_F\}, \tag*{(65)} ψ n := 1 { T × ( G K ( Z n )) < μ F × − z α σ F } , (65)
where z α z_\alpha z α is the ( 1 − α ) (1-\alpha) ( 1 − α ) -th quantile of the standard normal distribution.
We estimate the empirical power of the test ψ n \psi_n ψ n using 2000 repetitions for each configuration.
(E1) Location shifts: Here,
P = N ( 0 , I 10 ) and Q = N ( μ u , I 10 ) , P = \mathcal{N}(0, I_{10}) \quad\text{and} \quad Q = \mathcal{N}(\mu\mathbf{u}, I_{10}), P = N ( 0 , I 10 ) and Q = N ( μ u , I 10 ) ,
where u = 1 10 ( 1 , … , 1 ) T \mathbf{u} = \frac{1}{\sqrt{10}}(1,\ldots,1)^{\mathsf{T}} u = 10 1 ( 1 , … , 1 ) T . This normalization ensures that u \mathbf{u} u is a unit vector, so that μ ≥ 0 \mu\ge0 μ ≥ 0 is the Euclidean distance between the two means. Figure 1 (a) shows the empirical power for a grid of values of μ \mu μ between 0 and 1.5, for each choice of K K K .
Figure 1. Empirical power of the test ψ n \psi_n ψ n for Example (E1) (left panel) and Example (E2) (right panel). The dashed lines show the nominal level α = 0.05 \alpha= 0.05 α = 0.05 .
(E2) Heavy-tailed mixture alternatives: Here,
P = N ( 0 , I 2 ) and Q = ( 1 − ε ) N ( 0 , I 2 ) + ε R , P = \mathcal{N}(0, I_{2}) \quad\text{and} \quad Q = (1-\varepsilon)\mathcal{N}(0, I_{2}) + \varepsilon R, P = N ( 0 , I 2 ) and Q = ( 1 − ε ) N ( 0 , I 2 ) + εR ,
where R R R is the distribution of 1 3 ( T 1 , T 2 ) T \frac{1}{\sqrt{3}}(T_{1},T_{2})^{\mathsf{T}} 3 1 ( T 1 , T 2 ) T , and T 1 T_{1} T 1 and T 2 T_{2} T 2 are independent t t t -distributed random variables with 3 degrees of freedom. Since each T i T_i T i has variance 3, scaling by 1 3 \frac{1}{\sqrt{3}} 3 1 gives each coordinate unit variance. Thus, P P P and Q Q Q both have mean 0 and covariance matrix I 2 I_{2} I 2 . Figure 1 (b) shows the empirical power for a grid of values of ε \varepsilon ε between 0 and 1, for each choice of K K K .
In both examples, the empirical power increases as the alternative moves farther from the null, as expected. It also generally increases with K K K , illustrating the benefit of using more neighbors in these settings. More generally, the dependence of power on K K K is subtle and can vary with the underlying distributions, leading to different detection rates. This dependence was investigated by Bhattacharya [5 ] for fixed K K K , and recently extended by Kanekar [31 ] to the regime K = o ( n 1 / 4 ) K = o(n^{1/4}) K = o ( n 1/4 ) . A more complete understanding of how K K K affects power and detection rates for specific alternatives remains an interesting direction for future research.
Motif Tables
In this section, we collect the coefficients appearing in the motif expansions used in the proofs, together with elementary bounds on the corresponding motif counts.
(i) Table 1 lists the coefficients c H ( u n ) c_H(u_n) c H ( u n ) in the motif expansion (44) of Var [ V n ] \operatorname{Var}[V_n] Var [ V n ] , used in the proof of Lemma 3.5.
(ii) Table 2 lists the coefficients c H ( 3 ) ( u n ) c_H^{(3)}(u_n) c H ( 3 ) ( u n ) in the motif expansion (14) of E [ S n 3 ] \mathbb{E}[S_n^3] E [ S n 3 ] .
(iii) Table 3 lists the coefficients c H ( 4 ) ( u n ) c_H^{(4)}(u_n) c H ( 4 ) ( u n ) in the motif expansion (20) of E [ S n 4 ] \mathbb{E}[S_n^4] E [ S n 4 ] .
In each table, the last column gives an asymptotic bound on the corresponding motif count, obtained using the identities ∑ i = 1 n d G n ( i ) = 2 N ( ∙ − ∙ , G n ) \sum_{i=1}^{n} d_{G_n}(i)=2N(\bullet\!-\!\bullet,G_n) ∑ i = 1 n d G n ( i ) = 2 N ( ∙ − ∙ , G n ) and ∑ i = 1 n ( d G n ( i ) 2 ) = N ( ∙ ╱ ╲ ∙ , G n ) \sum_{i=1}^{n}\binom{d_{G_n}(i)}{2}=N(\begin{smallmatrix}\bullet\\[-2pt]\diagup\!\!\diagdown\\[-2pt]\bullet\end{smallmatrix},G_n) ∑ i = 1 n ( 2 d G n ( i ) ) = N ( ∙ ╱ ╲ ∙ , G n ) and elementary counting arguments.
Universal Consistency
In this section we establish universal consistency of the undirected K K K -NN test.
Proposition B.1. Let T × ( G K ( Z n ) ) T^\times(\mathcal{G}_K(\mathcal{Z}_n)) T × ( G K ( Z n )) be as defined in (7) . Suppose λ n : = n 1 n → λ ∈ ( 0 , 1 ) \lambda_n := \frac{n_1}{n} \to\lambda\in(0,1) λ n := n n 1 → λ ∈ ( 0 , 1 ) and either K ≥ 1 K \ge1 K ≥ 1 fixed or K n → ∞ K_n \to\infty K n → ∞ with K n = o ( n ) K_n=o(n) K n = o ( n ) . Then, as n → ∞ n \to\infty n → ∞ ,
T × ( G K ( Z n ) ) N ( ∙ — ∙ , G K ( Z n ) ) → P δ λ ( f , g ) : = ∫ 2 λ ( 1 − λ ) f ( x ) g ( x ) λ f ( x ) + ( 1 − λ ) g ( x ) d x . (66) \frac{T^\times(\mathcal{G}_K(\mathcal{Z}_n))}{N(\bullet\text{—}\bullet,\mathcal{G}_K(\mathcal{Z}_n))} \xrightarrow{P} \delta_\lambda(f,g) := \int\frac{2\lambda(1-\lambda)f(x)g(x)}{\lambda f(x)+(1-\lambda)g(x)}\,\mathrm{d}x. \tag*{(66)} N ( ∙ — ∙ , G K ( Z n )) T × ( G K ( Z n )) P δ λ ( f , g ) := ∫ λ f ( x ) + ( 1 − λ ) g ( x ) 2 λ ( 1 − λ ) f ( x ) g ( x ) d x . (66) Further, δ λ ( f , g ) ≤ δ λ ( f , f ) = 2 λ ( 1 − λ ) \delta_\lambda(f,g) \le\delta_\lambda(f,f)=2\lambda(1-\lambda) δ λ ( f , g ) ≤ δ λ ( f , f ) = 2 λ ( 1 − λ ) and equality holds if and only if f = g f=g f = g almost everywhere.
The proof, given in Section B.1, works directly with Lebesgue points of f f f and g g g and, unlike in [31 ] , requires no additional continuity, smoothness, or support assumptions on the densities.
Recall that we can construct a universally consistent two-sample test with precise control on the asymptotic level as follows: Fix α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) and consider the test function that rejects the null hypothesis H 0 H_0 H 0 as follows:
ψ n : = 1 { T × ( G K ( Z n ) ) < μ F × − z α σ F } , (67) \psi_n := \mathbf{1}\{T^{\times}(\mathcal{G}_K(\mathcal{Z}_n)) < \mu^{\times}_{\mathcal{F}} - z_\alpha\sigma_{\mathcal{F}}\}, \tag*{(67)} ψ n := 1 { T × ( G K ( Z n )) < μ F × − z α σ F } , (67)
where z α z_\alpha z α is the ( 1 − α ) (1-\alpha) ( 1 − α ) -th quantile of the standard normal distribution. The following result is a direct corollary of the above results.
Corollary B.2. Fix α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) and let ψ n \psi_n ψ n be as defined above. Suppose either K ≥ 1 K \ge1 K ≥ 1 is fixed or K n → ∞ K_n \to\infty K n → ∞ with K n = o ( n ) K_n = o(n) K n = o ( n ) . Then the following hold, whenever λ n = n 1 n → λ ∈ ( 0 , 1 ) ∖ { 1 2 } \lambda_n = \frac{n_1}{n} \to\lambda\in(0,1)\setminus\left\{\frac{1}{2}\right\} λ n = n n 1 → λ ∈ ( 0 , 1 ) ∖ { 2 1 } or n 1 = n 2 n_1=n_2 n 1 = n 2 :
(i) (Asymptotic Type I error control) lim n → ∞ P H 0 ( ψ n = 1 ) = α \lim_{n\to\infty} \mathbb{P}_{H_0}(\psi_n=1)=\alpha lim n → ∞ P H 0 ( ψ n = 1 ) = α
(ii) (Universal consistency) For any two distributions f ≠ g f \ne g f = g in H 1 H_1 H 1 , lim n → ∞ P H 1 ( ψ n = 1 ) = 1 \lim_{n\to\infty}\mathbb{P}_{H_1}(\psi_n=1)=1 lim n → ∞ P H 1 ( ψ n = 1 ) = 1 .
The proof of Corollary B.2 is given in Section B.3.
Proof of ppn:KNNconsistency
Let P 1 \mathcal{P}_1 P 1 denote a homogeneous Poisson process of unit intensity in R d \mathbb{R}^d R d , and let P n ϕ n \mathcal{P}_{n\phi_n} P n ϕ n denote an inhomogeneous Poisson process in R d \mathbb{R}^d R d with intensity function n ϕ n n\phi_n n ϕ n , where ϕ n ( z ) = λ n f ( z ) + ( 1 − λ n ) g ( z ) \phi_n(z)=\lambda_n f(z)+(1-\lambda_n)g(z) ϕ n ( z ) = λ n f ( z ) + ( 1 − λ n ) g ( z ) . For a finite set S ⊂ R d S\subset\mathbb{R}^d S ⊂ R d , we denote by P 1 S \mathcal{P}_1^S P 1 S and P n ϕ n S \mathcal{P}_{n\phi_n}^S P n ϕ n S the respective point processes obtained by adding to them the points of S S S . The proof of Proposition B.1 will require the following result about the limiting value of the expected degree of a vertex in a K K K -NN graph, in the diverging K K K regime.
Lemma B.3. Suppose K = K n → ∞ K=K_n\to\infty K = K n → ∞ such that K = o ( n ) K=o(n) K = o ( n ) . Then, for almost every z ∈ R d z\in\mathbb{R}^d z ∈ R d with ϕ ( z ) : = λ f ( z ) + ( 1 − λ ) g ( z ) > 0 \phi(z):=\lambda f(z)+(1-\lambda)g(z)>0 ϕ ( z ) := λ f ( z ) + ( 1 − λ ) g ( z ) > 0 ,
1 K E [ d G K ( P n ϕ n z ) ( z ) ] → 1. (68) \frac{1}{K}\mathbb{E}\left[d_{\mathcal{G}_K(\mathcal{P}_{n\phi_n}^z)}(z)\right]\to1. \tag*{(68)} K 1 E [ d G K ( P n ϕ n z ) ( z ) ] → 1. (68)
The proof of Lemma B.3 is given later in Section B.2. We first proceed with the proof of Proposition B.1 , assuming this result. We begin by computing the expected value of T × ( G K ( Z n ) ) T^{\times}(\mathcal{G}_K(\mathcal{Z}_n)) T × ( G K ( Z n )) . To this end, define,
h n ( x , y ) = n 1 n 2 f ( x ) g ( y ) ( n 1 f ( x ) + n 2 g ( x ) ) ( n 1 f ( y ) + n 2 g ( y ) ) h_n(x,y)=\frac{n_1n_2f(x)g(y)}{(n_1f(x)+n_2g(x))(n_1f(y)+n_2g(y))} h n ( x , y ) = ( n 1 f ( x ) + n 2 g ( x )) ( n 1 f ( y ) + n 2 g ( y )) n 1 n 2 f ( x ) g ( y )
and
h ( x , y ) = λ ( 1 − λ ) f ( x ) g ( y ) ( λ f ( x ) + ( 1 − λ ) g ( x ) ) ( λ f ( y ) + ( 1 − λ ) g ( y ) ) . (69) h(x,y)=\frac{\lambda(1-\lambda)f(x)g(y)}{(\lambda f(x)+(1-\lambda)g(x))(\lambda f(y)+(1-\lambda)g(y))}. \tag*{(69)} h ( x , y ) = ( λ f ( x ) + ( 1 − λ ) g ( x )) ( λ f ( y ) + ( 1 − λ ) g ( y )) λ ( 1 − λ ) f ( x ) g ( y ) . (69)
Note that ∥ h n − h ∥ ∞ → 0 \|h_n-h\|_\infty\to0 ∥ h n − h ∥ ∞ → 0 . Conditionally on F \mathcal{F} F , the labels are independent, and h n ( Z i , Z j ) h_n(Z_i,Z_j) h n ( Z i , Z j ) is the probability that Z i Z_i Z i has label 1 and Z j Z_j Z j has label 2. Therefore,
E [ T × ( G K ( Z n ) ) ∣ F ] = ∑ 1 ≤ i ≠ j ≤ L n h n ( Z i , Z j ) 1 { { Z i , Z j } ∈ E ( G K ( Z n ) ) } . (70) \mathbb{E}\left[T^{\times}(\mathcal{G}_K(\mathcal{Z}_n))\mid\mathcal{F}\right]
=
\sum_{1\le i\ne j\le L_n}h_n(Z_i,Z_j)\mathbf{1}\{\{Z_i,Z_j\}\in E(\mathcal{G}_K(\mathcal{Z}_n))\}. \tag*{(70)} E [ T × ( G K ( Z n )) ∣ F ] = 1 ≤ i = j ≤ L n ∑ h n ( Z i , Z j ) 1 {{ Z i , Z j } ∈ E ( G K ( Z n ))} . (70)
Note that
∣ 1 n K ∑ 1 ≤ i < j ≤ L n { h n ( Z i , Z j ) − h ( Z i , Z j ) } 1 { { Z i , Z j } ∈ E ( G K ( Z n ) ) } ∣ ≤ 2 ∥ h n − h ∥ ∞ ∣ E ( G K ( Z n ) ) ∣ n K ≤ 2 ∥ h n − h ∥ ∞ L n n = o L 1 ( 1 ) , (71) \left|\frac{1}{nK}\sum_{1\le i<j\le L_n}\{h_n(Z_i,Z_j)-h(Z_i,Z_j)\}1\{\{Z_i,Z_j\}\in E(G_K(\mathcal{Z}_n))\}\right|
\le2\lVert h_n-h\rVert_\infty\frac{|E(G_K(\mathcal{Z}_n))|}{nK}
\le2\lVert h_n-h\rVert_\infty\frac{L_n}{n}=o_{L^1}(1), \tag*{(71)} n K 1 1 ≤ i < j ≤ L n ∑ { h n ( Z i , Z j ) − h ( Z i , Z j )} 1 {{ Z i , Z j } ∈ E ( G K ( Z n ))} ≤ 2 ∥ h n − h ∥ ∞ n K ∣ E ( G K ( Z n )) ∣ ≤ 2 ∥ h n − h ∥ ∞ n L n = o L 1 ( 1 ) , (71)
where we used ∣ E ( G K ( Z n ) ) ∣ ≤ K L n |E(G_K(\mathcal{Z}_n))|\le KL_n ∣ E ( G K ( Z n )) ∣ ≤ K L n and E L n = n \mathbb{E}L_n=n E L n = n . Hence, by (B.5),
1 n K E [ T × ( G K ( Z n ) ) ∣ F ] = 1 n K ∑ 1 ≤ i < j ≤ L n h ( Z i , Z j ) 1 { { Z i , Z j } ∈ E ( G K ( Z n ) ) } + o L 1 ( 1 ) . \frac{1}{nK}\mathbb{E}\left[T^{\times}(G_K(\mathcal{Z}_n))\mid\mathcal{F}\right]
=\frac{1}{nK}\sum_{1\le i<j\le L_n}h(Z_i,Z_j)1\{\{Z_i,Z_j\}\in E(G_K(\mathcal{Z}_n))\}+o_{L^1}(1). n K 1 E [ T × ( G K ( Z n )) ∣ F ] = n K 1 1 ≤ i < j ≤ L n ∑ h ( Z i , Z j ) 1 {{ Z i , Z j } ∈ E ( G K ( Z n ))} + o L 1 ( 1 ) .
The following lemma computes the limit of the expected value of the sum above.
Lemma B.4. Suppose either K K K is fixed or K = K n → ∞ K=K_n\to\infty K = K n → ∞ with K n = o ( n ) K_n=o(n) K n = o ( n ) . Let H : R d × R d → R H:\mathbb{R}^d\times\mathbb{R}^d\to\mathbb{R} H : R d × R d → R be a jointly measurable and bounded function, such that for almost every z ∈ R d z\in\mathbb{R}^d z ∈ R d , H ( z , ⋅ ) H(z,\cdot) H ( z , ⋅ ) is measurable and z z z is a Lebesgue point of the function ϕ ( ⋅ ) H ( z , ⋅ ) \phi(\cdot)H(z,\cdot) ϕ ( ⋅ ) H ( z , ⋅ ) . Then
1 n K E [ ∑ 1 ≤ i < j ≤ L n H ( Z i , Z j ) 1 { { Z i , Z j } ∈ E ( G K ( Z n ) ) } ] → γ ∫ R d H ( z , z ) ϕ ( z ) d z . \frac{1}{nK}\mathbb{E}\left[\sum_{1\le i<j\le L_n}H(Z_i,Z_j)1\{\{Z_i,Z_j\}\in E(G_K(\mathcal{Z}_n))\}\right]\to\gamma\int_{\mathbb{R}^d}H(z,z)\phi(z)\,dz. n K 1 E 1 ≤ i < j ≤ L n ∑ H ( Z i , Z j ) 1 {{ Z i , Z j } ∈ E ( G K ( Z n ))} → γ ∫ R d H ( z , z ) ϕ ( z ) d z . γ = { 1 K E [ d G K ( P 1 0 ) ( 0 ) ] , if K is fixed , 1 , if K = K n → ∞ . \gamma=
\begin{cases}
\frac{1}{K}\mathbb{E}\left[d_{G_K(\mathcal{P}_1^0)}(0)\right], & \text{if }K\text{ is fixed},\\
1, & \text{if }K=K_n\to\infty.
\end{cases} γ = { K 1 E [ d G K ( P 1 0 ) ( 0 ) ] , 1 , if K is fixed , if K = K n → ∞.
Proof. By the Palm theory for Poisson processes [44 Theorem 1.6] ,
E [ ∑ 1 ≤ i ≠ j ≤ L n H ( Z i , Z j ) 1 { { Z i , Z j } ∈ E ( G K ( Z n ) ) } ] = n ∫ R d w n ( z ) ϕ n ( z ) d z , \mathbb{E}\left[\sum_{1\le i\ne j\le L_n}H(Z_i,Z_j)1\{\{Z_i,Z_j\}\in E(G_K(\mathcal{Z}_n))\}\right]
=n\int_{\mathbb{R}^d}w_n(z)\phi_n(z)\,dz, E 1 ≤ i = j ≤ L n ∑ H ( Z i , Z j ) 1 {{ Z i , Z j } ∈ E ( G K ( Z n ))} = n ∫ R d w n ( z ) ϕ n ( z ) d z ,
where, for z ∈ R d z\in\mathbb{R}^d z ∈ R d ,
w n ( z ) : = E [ ∑ z ′ ∈ P ϕ n H ( z , z ′ ) 1 { { z , z ′ } ∈ E ( G K ( P ϕ n z ) ) } ] . w_n(z):=\mathbb{E}\left[\sum_{z'\in\mathcal{P}_{\phi_n}}H(z,z')1\{\{z,z'\}\in E(G_K(\mathcal{P}_{\phi_n}^{z}))\}\right]. w n ( z ) := E z ′ ∈ P ϕ n ∑ H ( z , z ′ ) 1 {{ z , z ′ } ∈ E ( G K ( P ϕ n z ))} .
Note that the geometric degree bound gives ∣ w n ( z ) ∣ ≤ C d K ∥ H ∥ ∞ |w_n(z)|\le C_dK\lVert H\rVert_\infty ∣ w n ( z ) ∣ ≤ C d K ∥ H ∥ ∞ . Also, for all sufficiently large n n n , ϕ n ≤ C λ ϕ \phi_n\le C_\lambda\phi ϕ n ≤ C λ ϕ . Therefore, by dominated convergence, it suffices to prove
w n ( z ) K → γ H ( z , z ) , (72) \frac{w_n(z)}{K}\to\gamma H(z,z), \tag*{(72)} K w n ( z ) → γ H ( z , z ) , (72)
for almost every z ∈ R d z\in\mathbb{R}^d z ∈ R d with ϕ ( z ) > 0 \phi(z)>0 ϕ ( z ) > 0 . Fix a common Lebesgue point z z z of f f f and g g g such that ϕ ( z ) > 0 \phi(z)>0 ϕ ( z ) > 0 . Fix a sufficiently large constant A A A , and choose r n r_n r n so that
n ϕ n ( z ) ∣ B ( z , r n ) ∣ = A K , n\phi_n(z)|B(z,r_n)|=AK, n ϕ n ( z ) ∣ B ( z , r n ) ∣ = A K ,
where B ( z , r n ) B(z,r_n) B ( z , r n ) is the d d d -dimensional Euclidean ball with center at z z z and radius r n r_n r n . Since K = o ( n ) K=o(n) K = o ( n ) , we have r n → 0 r_n\to0 r n → 0 . Cover the directions from z z z by a finite collection of cones C \mathcal{C} C each of angular diameter strictly less than π 3 \frac{\pi}{3} 3 π . Fix one cone C j ∈ C C_j\in\mathcal{C} C j ∈ C . By invariance under dilation,
∣ C j ∩ B ( z , r ) ∣ = c j ∣ B ( z , r ) ∣ , |C_j\cap B(z,r)|=c_j|B(z,r)|, ∣ C j ∩ B ( z , r ) ∣ = c j ∣ B ( z , r ) ∣ ,
where c j : = ∣ C j ∩ B ( z , 1 ) ∣ ∣ B ( z , 1 ) ∣ > 0 c_j:=\frac{|C_j\cap B(z,1)|}{|B(z,1)|}>0 c j := ∣ B ( z , 1 ) ∣ ∣ C j ∩ B ( z , 1 ) ∣ > 0 . Define
N j : = ∣ P n ϕ n ∩ C j ∩ B ( z , r n ) ∣ ∼ Poi ( n ∫ C j ∩ B ( z , r n ) ϕ n ( y ) d y ) . N_j:=|\mathcal{P}_{n\phi_n}\cap C_j\cap B(z,r_n)|\sim\operatorname{Poi}\left(n\int_{C_j\cap B(z,r_n)}\phi_n(y)\,dy\right). N j := ∣ P n ϕ n ∩ C j ∩ B ( z , r n ) ∣ ∼ Poi ( n ∫ C j ∩ B ( z , r n ) ϕ n ( y ) d y ) .
Hence, by the Lebesgue differentiation theorem,
∣ ∫ C j ∩ B ( z , r n ) ( ϕ n ( y ) − ϕ n ( z ) ) d y ∣ ≤ ∫ B ( z , r n ) ∣ ϕ n ( y ) − ϕ n ( z ) ∣ d y = o ( r n d ) . \left|\int_{C_j\cap B(z,r_n)}(\phi_n(y)-\phi_n(z))\,dy\right|\leq\int_{B(z,r_n)}|\phi_n(y)-\phi_n(z)|\,dy=o(r_n^d). ∫ C j ∩ B ( z , r n ) ( ϕ n ( y ) − ϕ n ( z )) d y ≤ ∫ B ( z , r n ) ∣ ϕ n ( y ) − ϕ n ( z ) ∣ d y = o ( r n d ) .
Therefore, using the choice n ϕ n ( z ) ∣ B ( z , r n ) ∣ = A K n\phi_n(z)|B(z,r_n)|=AK n ϕ n ( z ) ∣ B ( z , r n ) ∣ = A K , we obtain
E [ N j ] = n ∫ C j ∩ B ( z , r n ) ϕ n ( y ) d y = n ϕ n ( z ) ∣ C j ∩ B ( z , r n ) ∣ + n o ( r n d ) = c j n ϕ n ( z ) ∣ B ( z , r n ) ∣ + n o ( r n d ) = ( c j A + o ( 1 ) ) K , \begin{aligned}
\mathbb{E}[N_j]&=n\int_{C_j\cap B(z,r_n)}\phi_n(y)\,dy=n\phi_n(z)|C_j\cap B(z,r_n)|+no(r_n^d)\\
&=c_jn\phi_n(z)|B(z,r_n)|+no(r_n^d)\\
&=(c_jA+o(1))K,
\end{aligned} E [ N j ] = n ∫ C j ∩ B ( z , r n ) ϕ n ( y ) d y = n ϕ n ( z ) ∣ C j ∩ B ( z , r n ) ∣ + n o ( r n d ) = c j n ϕ n ( z ) ∣ B ( z , r n ) ∣ + n o ( r n d ) = ( c j A + o ( 1 )) K ,
using ϕ n ( z ) → ϕ ( z ) > 0 \phi_n(z)\to\phi(z)>0 ϕ n ( z ) → ϕ ( z ) > 0 and n r n d = O ( K ) nr_n^d=O(K) n r n d = O ( K ) , which justifies n o ( r n d ) = o ( K ) no(r_n^d)=o(K) n o ( r n d ) = o ( K ) . Note that the constants c j c_j c j may differ between cones. However, there are only finitely many cones, so the minimum of these constants is positive. Hence, by choosing A A A sufficiently large, we can make E [ N j ] ≥ 1 2 c j A K ≥ 2 K \mathbb{E}[N_j]\geq\frac{1}{2}c_jAK\geq2K E [ N j ] ≥ 2 1 c j A K ≥ 2 K , for all C j ∈ C C_j\in\mathcal{C} C j ∈ C , for sufficiently large n n n . Hence, by Chebyshev’s inequality and a union bound, for sufficiently large n n n ,
P ( ∃ C j ∈ C : N j < K ) ≤ ∑ C j ∈ C P ( N j < K ) ≤ ∑ C j ∈ C E [ N j ] ( E [ N j ] − K ) 2 ≤ ∑ C j ∈ C 4 E [ N j ] ≤ C d A K , (73) \mathbb{P}(\exists\,C_j\in\mathcal{C}:N_j<K)\leq\sum_{C_j\in\mathcal{C}}\mathbb{P}(N_j<K)\leq\sum_{C_j\in\mathcal{C}}\frac{\mathbb{E}[N_j]}{(\mathbb{E}[N_j]-K)^2}\leq\sum_{C_j\in\mathcal{C}}\frac{4}{\mathbb{E}[N_j]}
\leq\frac{C_d}{AK}, \tag*{(73)} P ( ∃ C j ∈ C : N j < K ) ≤ C j ∈ C ∑ P ( N j < K ) ≤ C j ∈ C ∑ ( E [ N j ] − K ) 2 E [ N j ] ≤ C j ∈ C ∑ E [ N j ] 4 ≤ A K C d , (73)
with C d = 8 ∑ C j ∈ C 1 c j C_d=8\sum_{C_j\in\mathcal{C}}\frac{1}{c_j} C d = 8 ∑ C j ∈ C c j 1 . Let E n = { N j ≥ K for all C j ∈ C } E_n=\{N_j\geq K\text{ for all }C_j\in\mathcal{C}\} E n = { N j ≥ K for all C j ∈ C } be the event that every cone contains at least K K K points in B ( z , r n ) B(z,r_n) B ( z , r n ) . Note that
∣ w n ( z ) − H ( z , z ) E [ d G K ( P n ϕ n z ) ( z ) ] ∣ ≤ E [ ∑ z ′ ∈ P n ϕ n ∣ H ( z , z ′ ) − H ( z , z ) ∣ 1 { { z , z ′ } ∈ E ( G K ( P n ϕ n z ) ) } ] . \left|w_n(z)-H(z,z)\mathbb{E}\left[d_{\mathcal{G}_K(\mathcal{P}_{n\phi_n}^z)}(z)\right]\right|
\leq\mathbb{E}\left[\sum_{z'\in\mathcal{P}_{n\phi_n}}|H(z,z')-H(z,z)|1\{\{z,z'\}\in E(\mathcal{G}_K(\mathcal{P}_{n\phi_n}^z))\}\right]. w n ( z ) − H ( z , z ) E [ d G K ( P n ϕ n z ) ( z ) ] ≤ E z ′ ∈ P n ϕ n ∑ ∣ H ( z , z ′ ) − H ( z , z ) ∣1 {{ z , z ′ } ∈ E ( G K ( P n ϕ n z ))} .
Note that on E n E_n E n , every undirected neighbor of z z z lies in B ( z , r n ) B(z,r_n) B ( z , r n ) . Also, on E n c E_n^c E n c , each summand above is at most 2 ∥ H ∥ ∞ 2\|H\|_\infty 2∥ H ∥ ∞ and the degree bound gives at most B d K B_dK B d K terms. Hence,
∣ w n ( z ) − H ( z , z ) E [ d G K ( P n ϕ n z ) ( z ) ] ∣ ≤ E [ ∑ z ′ ∈ P n ϕ n ∩ B ( z , r n ) ∣ H ( z , z ′ ) − H ( z , z ) ∣ ] + 2 B d K ∥ H ∥ ∞ P ( E n c ) . \left|w_n(z)-H(z,z)\mathbb{E}\left[d_{\mathcal{G}_K(\mathcal{P}_{n\phi_n}^z)}(z)\right]\right|
\leq\mathbb{E}\left[\sum_{z'\in\mathcal{P}_{n\phi_n}\cap B(z,r_n)}|H(z,z')-H(z,z)|\right]+2B_dK\|H\|_\infty\mathbb{P}(E_n^c). w n ( z ) − H ( z , z ) E [ d G K ( P n ϕ n z ) ( z ) ] ≤ E z ′ ∈ P n ϕ n ∩ B ( z , r n ) ∑ ∣ H ( z , z ′ ) − H ( z , z ) ∣ + 2 B d K ∥ H ∥ ∞ P ( E n c ) .
Dividing by K K K , we conclude that
1 K ∣ w n ( z ) − H ( z , z ) E [ d G K ( P n ϕ n z ) ( z ) ] ∣ ≤ n K ∫ B ( z , r n ) ∣ H ( z , y ) − H ( z , z ) ∣ ϕ n ( y ) d y + 2 C d ∥ H ∥ ∞ P ( E n c ) . (74) \frac{1}{K}\left|w_n(z)-H(z,z)\mathbb{E}\left[d_{\mathcal{G}_K(\mathcal{P}_{n\phi_n}^{z})}(z)\right]\right|
\leq\frac{n}{K}\int_{B(z,r_n)}\left|H(z,y)-H(z,z)\right|\phi_n(y)\,dy+2C_d\lVert H\rVert_\infty\mathbb{P}(\mathcal{E}_n^c). \tag*{(74)} K 1 w n ( z ) − H ( z , z ) E [ d G K ( P n ϕ n z ) ( z ) ] ≤ K n ∫ B ( z , r n ) ∣ H ( z , y ) − H ( z , z ) ∣ ϕ n ( y ) d y + 2 C d ∥ H ∥ ∞ P ( E n c ) . (74)
Using ϕ n ≤ C λ ϕ \phi_n\leq C_\lambda\phi ϕ n ≤ C λ ϕ and the Lebesgue differentiation theorem
∫ B ( z , r n ) ∣ H ( z , y ) − H ( z , z ) ∣ ϕ n ( y ) d y = o ( r n d ) . \int_{B(z,r_n)}\left|H(z,y)-H(z,z)\right|\phi_n(y)\,dy=o(r_n^d). ∫ B ( z , r n ) ∣ H ( z , y ) − H ( z , z ) ∣ ϕ n ( y ) d y = o ( r n d ) .
Hence, for fixed A A A , the first term in (B.9) goes to zero as n → ∞ n\to\infty n → ∞ , since n o ( r n d ) = o ( K ) no(r_n^d)=o(K) n o ( r n d ) = o ( K ) . The second term in (B.9 also goes to zero as n → ∞ n\to\infty n → ∞ followed by A → ∞ A\to\infty A → ∞ , by using (B.8). Hence,
1 K w n ( z ) − H ( z , z ) 1 K E [ d G K ( P n ϕ n z ) ( z ) ] → 0. \frac{1}{K}w_n(z)-H(z,z)\frac{1}{K}\mathbb{E}\left[d_{\mathcal{G}_K(\mathcal{P}_{n\phi_n}^{z})}(z)\right]\to0. K 1 w n ( z ) − H ( z , z ) K 1 E [ d G K ( P n ϕ n z ) ( z ) ] → 0.
Now, for K K K fixed the result in (B.7) follows by invoking standard results of stabilization of the K K K -NN graph (see, for example, [45 Lemma 3.2] ). For K → ∞ K\to\infty K → ∞ , the result in (B.7) follows by invoking Lemma B.3. This completes the proof of Lemma B.4.
By (B.6) and Lemma B.4 with H = h H=h H = h as in (B.4),
1 n K E [ T × ( G K ( Z n ) ) ] → γ ∫ λ ( 1 − λ ) f ( x ) g ( x ) λ f ( x ) + ( 1 − λ ) g ( x ) d x . (75) \frac{1}{nK}\mathbb{E}\left[T^\times(\mathcal{G}_K(\mathcal{Z}_n))\right]\to\gamma\int\frac{\lambda(1-\lambda)f(x)g(x)}{\lambda f(x)+(1-\lambda)g(x)}\,dx. \tag*{(75)} n K 1 E [ T × ( G K ( Z n )) ] → γ ∫ λ f ( x ) + ( 1 − λ ) g ( x ) λ ( 1 − λ ) f ( x ) g ( x ) d x . (75)
The variance bound in the next lemma shows that this convergence holds in L 2 L^2 L 2 .
Lemma B.5. Suppose either K K K is fixed or K = K n → ∞ K=K_n\to\infty K = K n → ∞ . Then
Var [ T × ( G K ( Z n ) ) ] ≤ C d n K 2 , \operatorname{Var}\left[T^\times(\mathcal{G}_K(\mathcal{Z}_n))\right]\leq C_d nK^2, Var [ T × ( G K ( Z n )) ] ≤ C d n K 2 , for some constant C d > 0 C_d>0 C d > 0 , depending on d d d .
Proof. Let B d K B_dK B d K be the geometric upper bound on every vertex degree as in (61) . Note that adding one marked point introduces at most B d K B_dK B d K edges incident to the new point. Further, an existing vertex changes its nearest-neighbor list only if it selects the new point, and each such vertex loses at most one previous neighbor. There are at most B d K B_dK B d K such vertices, so at most B d K B_dK B d K old edges disappear. Hence, adding or deleting one marked point changes the cross-edge count by at most 2 B d K 2B_dK 2 B d K . Consequently, replacing one marked observation, including its location and label, changes the cross-edge count by at most 4 B d K 4B_dK 4 B d K .
Condition on L n = ℓ L_n=\ell L n = ℓ , denote T = T × ( G K ( Z n ) ) T=T^\times(\mathcal{G}_K(\mathcal{Z}_n)) T = T × ( G K ( Z n )) and T ( i ) = T × ( G K ( Z n ( i ) ) ) T^{(i)}=T^\times(\mathcal{G}_K(\mathcal{Z}_n^{(i)})) T ( i ) = T × ( G K ( Z n ( i ) )) , where Z n ( i ) \mathcal{Z}_n^{(i)} Z n ( i ) is obtained from Z n \mathcal{Z}_n Z n by replacing the i i i -th marked observation with an independent copy and rebuilding the graph. Then by the Efron–Stein inequality [20 ] ,
Var [ T ∣ L n = ℓ ] ≤ 1 2 ∑ i = 1 ℓ E [ ( T − T ( i ) ) 2 ∣ L n = ℓ ] . (76) \operatorname{Var}[T\mid L_n=\ell]\leq\frac{1}{2}\sum_{i=1}^{\ell}\mathbb{E}\left[(T-T^{(i)})^2\mid L_n=\ell\right]. \tag*{(76)} Var [ T ∣ L n = ℓ ] ≤ 2 1 i = 1 ∑ ℓ E [ ( T − T ( i ) ) 2 ∣ L n = ℓ ] . (76)
Replacing one marked observation can be performed by first deleting the original point and then inserting its replacement. Each operation changes the cross-edge count by at most 2 B d K 2B_dK 2 B d K , as discussed above. Hence, by triangle inequality, ∣ T − T ( i ) ∣ ≤ 4 B d K \lvert T-T^{(i)}\rvert\le4B_dK ∣ T − T ( i ) ∣ ≤ 4 B d K . Therefore,
E [ ( T − T ( i ) ) 2 ∣ L n = ℓ ] ≤ ( 4 B d K ) 2 . \mathbb{E}\left[(T-T^{(i)})^2\mid L_n=\ell\right]\le(4B_dK)^2. E [ ( T − T ( i ) ) 2 ∣ L n = ℓ ] ≤ ( 4 B d K ) 2 .
Applying this bound in (B.11) gives,
Var [ T × ( G K ( Z n ) ) ∣ L n = ℓ ] ≤ 1 2 ∑ i = 1 ℓ ( 4 B d K ) 2 = 8 B d 2 ℓ K 2 . \operatorname{Var}\left[T^{\times}(\mathcal{G}_K(\mathcal{Z}_n))\mid L_n=\ell\right]\le\frac{1}{2}\sum_{i=1}^{\ell}(4B_dK)^2=8B_d^2\ell K^2. Var [ T × ( G K ( Z n )) ∣ L n = ℓ ] ≤ 2 1 i = 1 ∑ ℓ ( 4 B d K ) 2 = 8 B d 2 ℓ K 2 .
Since E L n = n \mathbb{E}L_n=n E L n = n , it follows that
E [ Var [ T × ( G K ( Z n ) ) ∣ L n ] ] ≤ 8 B d 2 n K 2 . (77) \mathbb{E}\left[\operatorname{Var}\left[T^{\times}(\mathcal{G}_K(\mathcal{Z}_n))\mid L_n\right]\right]\le8B_d^2nK^2. \tag*{(77)} E [ Var [ T × ( G K ( Z n )) ∣ L n ] ] ≤ 8 B d 2 n K 2 . (77)
To control the variance of the conditional mean, define
a ℓ : = E [ T × ( G K ( Z n ) ) ∣ L n = ℓ ] , a_\ell:=\mathbb{E}\left[T^{\times}(\mathcal{G}_K(\mathcal{Z}_n))\mid L_n=\ell\right], a ℓ := E [ T × ( G K ( Z n )) ∣ L n = ℓ ] ,
for ℓ ≥ 0 \ell\ge0 ℓ ≥ 0 . By the same argument as above, ∣ a ℓ + 1 − a ℓ ∣ ≤ 2 B d K \lvert a_{\ell+1}-a_\ell\rvert\le2B_dK ∣ a ℓ + 1 − a ℓ ∣ ≤ 2 B d K . Therefore, ∣ a ℓ − a ℓ ′ ∣ ≤ 2 B d K ∣ ℓ − ℓ ′ ∣ \lvert a_\ell-a_{\ell'}\rvert\le2B_dK\lvert\ell-\ell'\rvert ∣ a ℓ − a ℓ ′ ∣ ≤ 2 B d K ∣ ℓ − ℓ ′ ∣ , for ℓ , ℓ ′ ≥ 0 \ell,\ell'\ge0 ℓ , ℓ ′ ≥ 0 . Let L n ′ L'_n L n ′ be an independent copy of L n L_n L n . Since L n ∼ Poi ( n ) L_n\sim\operatorname{Poi}(n) L n ∼ Poi ( n ) ,
Var [ a L n ] = 1 2 E [ ( a L n − a L n ′ ) 2 ] ≤ ( 2 B d K ) 2 2 E [ ( L n − L n ′ ) 2 ] = 4 B d 2 K 2 Var [ L n ] = 4 B d 2 n K 2 . (78) \operatorname{Var}[a_{L_n}]=\frac{1}{2}\mathbb{E}\left[(a_{L_n}-a_{L'_n})^2\right]\le\frac{(2B_dK)^2}{2}\mathbb{E}\left[(L_n-L'_n)^2\right]=4B_d^2K^2\operatorname{Var}[L_n]=4B_d^2nK^2. \tag*{(78)} Var [ a L n ] = 2 1 E [ ( a L n − a L n ′ ) 2 ] ≤ 2 ( 2 B d K ) 2 E [ ( L n − L n ′ ) 2 ] = 4 B d 2 K 2 Var [ L n ] = 4 B d 2 n K 2 . (78)
Combining (B.12) and (B.13, the result in Lemma B.5 follows. □
Combining (B.10) with Lemma B.5 gives,
1 n K T × ( G K ( Z n ) ) → P γ ∫ λ ( 1 − λ ) f ( x ) g ( x ) λ f ( x ) + ( 1 − λ ) g ( x ) d x . \frac{1}{nK}T^{\times}(\mathcal{G}_K(\mathcal{Z}_n))\xrightarrow{P}\gamma\int\frac{\lambda(1-\lambda)f(x)g(x)}{\lambda f(x)+(1-\lambda)g(x)}\,dx. n K 1 T × ( G K ( Z n )) P γ ∫ λ f ( x ) + ( 1 − λ ) g ( x ) λ ( 1 − λ ) f ( x ) g ( x ) d x .
Also, applying Lemma B.4 with H = 1 H=1 H = 1 and the same variance argument as in Lemma B.5 gives,
N ( ∙ , G K ( Z n ) ) n K → P γ 2 . \frac{N(\bullet,\mathcal{G}_K(\mathcal{Z}_n))}{nK}\xrightarrow{P}\frac{\gamma}{2}. n K N ( ∙ , G K ( Z n )) P 2 γ .
Combining the above proves the result in (B.1).
Finally, the fact that δ λ ( f , g ) ≤ δ λ ( f , f ) = 2 λ ( 1 − λ ) \delta_\lambda(f,g)\le\delta_\lambda(f,f)=2\lambda(1-\lambda) δ λ ( f , g ) ≤ δ λ ( f , f ) = 2 λ ( 1 − λ ) and equality holds if and only if f = g f=g f = g , follows from [25 Theorem 1 and Corollary 1] . This completes the proof of Proposition B.1 . □
Proof of lem:degree
The lower bound in (B.3) is immediate: Since the degree of any vertex is at least the outgoing degree,
E [ d G K ( P n ϕ n ( z ) ) ] ≥ E [ min { K , ∣ P n ϕ n ∣ } ] = K − o ( K ) , \mathbb{E}\left[d_{\mathcal{G}_K}(\mathcal{P}_{n\phi_n}(z))\right]\ge\mathbb{E}\left[\min\{K,\lvert\mathcal{P}_{n\phi_n}\rvert\}\right]=K-o(K), E [ d G K ( P n ϕ n ( z )) ] ≥ E [ min { K , ∣ P n ϕ n ∣} ] = K − o ( K ) ,
Proof of cor:levelconsistency
since ∣ P n ϕ n ∣ ∼ Pois ( n ) |\mathcal{P}_{n\phi_n}| \sim\operatorname{Pois}(n) ∣ P n ϕ n ∣ ∼ Pois ( n ) and K = o ( n ) K = o(n) K = o ( n ) .
For the upper bound, fix ε > 0 \varepsilon> 0 ε > 0 , and now choose r n r_n r n such that
n ϕ n ( z ) ∣ B ( z , r n ) ∣ = ( 1 + ε ) K . (79) n\phi_n(z)|B(z,r_n)| = (1+\varepsilon)K. \tag*{(79)} n ϕ n ( z ) ∣ B ( z , r n ) ∣ = ( 1 + ε ) K . (79)
Choose δ > 0 \delta> 0 δ > 0 sufficiently small that ( 1 + ε ) ( 1 − 2 δ ) d > 1 + ε 2 (1+\varepsilon)(1-2\delta)^d > 1+\frac{\varepsilon}{2} ( 1 + ε ) ( 1 − 2 δ ) d > 1 + 2 ε , and choose a finite δ \delta δ -net { u s } s ∈ S \{u_s\}_{s\in S} { u s } s ∈ S of the unit sphere. Fix a common Lebesgue point z z z of f f f and g g g (so that ∫ B ( z , r ) ∣ ϕ n ( y ) − ϕ n ( z ) ∣ d y = o ( r d ) \int_{B(z,r)}|\phi_n(y)-\phi_n(z)|\,dy=o(r^d) ∫ B ( z , r ) ∣ ϕ n ( y ) − ϕ n ( z ) ∣ d y = o ( r d ) uniformly in n n n ) with ϕ ( z ) > 0 \phi(z)>0 ϕ ( z ) > 0 and define
N s : = ∣ B ( z + r n u s , ( 1 − 2 δ ) r n ) ∩ P n ϕ n z ∣ . N_s := \left|B(z+r_nu_s,(1-2\delta)r_n)\cap\mathcal{P}^{z}_{n\phi_n}\right|. N s := B ( z + r n u s , ( 1 − 2 δ ) r n ) ∩ P n ϕ n z .
the number of points of the Poisson process P n ϕ n z \mathcal{P}^{z}_{n\phi_n} P n ϕ n z in the ball B ( z + r n u s , ( 1 − 2 δ ) r n ) B(z+r_nu_s,(1-2\delta)r_n) B ( z + r n u s , ( 1 − 2 δ ) r n ) . (Note that the inserted point z z z lies outside this ball B ( z + r n u s , ( 1 − 2 δ ) r n ) B(z+r_nu_s,(1-2\delta)r_n) B ( z + r n u s , ( 1 − 2 δ ) r n ) .) By the Lebesgue differentiation theorem and (B.14),
E [ N s ] = n ∫ B ( z + r n u s , ( 1 − 2 δ ) r n ) ϕ n ( y ) d y = ( ( 1 + ε ) ( 1 − 2 δ ) d + o ( 1 ) ) K ≥ ( 1 + ε 2 ) K , \mathbb{E}[N_s] = n\int_{B(z+r_nu_s,(1-2\delta)r_n)}\phi_n(y)\,dy = \left((1+\varepsilon)(1-2\delta)^d+o(1)\right)K \ge\left(1+\frac{\varepsilon}{2}\right)K, E [ N s ] = n ∫ B ( z + r n u s , ( 1 − 2 δ ) r n ) ϕ n ( y ) d y = ( ( 1 + ε ) ( 1 − 2 δ ) d + o ( 1 ) ) K ≥ ( 1 + 2 ε ) K ,
for sufficiently large n n n and all s ∈ S s\in S s ∈ S . Then Chebyshev’s inequality gives
P ( N s < K + 1 ) = P ( N s ≤ K ) ≤ P ( ∣ N s − E [ N s ] ∣ ≥ E [ N s ] − K ) ≤ E [ N s ] ( E [ N s ] − K ) 2 ≤ 4 E [ N s ] ε 2 K 2 = O ( 1 K ) . \begin{aligned}
\mathbb{P}(N_s<K+1)=\mathbb{P}(N_s\le K)\le\mathbb{P}\left(|N_s-\mathbb{E}[N_s]|\ge\mathbb{E}[N_s]-K\right)\le\frac{\mathbb{E}[N_s]}{(\mathbb{E}[N_s]-K)^2} \\
&\le\frac{4\mathbb{E}[N_s]}{\varepsilon^2K^2}=O\left(\frac{1}{K}\right).
\end{aligned} P ( N s < K + 1 ) = P ( N s ≤ K ) ≤ P ( ∣ N s − E [ N s ] ∣ ≥ E [ N s ] − K ) ≤ ( E [ N s ] − K ) 2 E [ N s ] ≤ ε 2 K 2 4 E [ N s ] = O ( K 1 ) .
Also, define N 0 : = ∣ B ( z , r n ) ∩ P n ϕ n ∣ N_0:=|B(z,r_n)\cap\mathcal{P}_{n\phi_n}| N 0 := ∣ B ( z , r n ) ∩ P n ϕ n ∣ . Note that the inserted point z z z is excluded from this count. Applying Lebesgue differentiation theorem and (B.14) again gives
E N 0 = n ∫ B ( z , r n ) ϕ n ( y ) d y = ( 1 + ε + o ( 1 ) ) K . (80) \mathbb{E}N_0=n\int_{B(z,r_n)}\phi_n(y)\,dy=(1+\varepsilon+o(1))K. \tag*{(80)} E N 0 = n ∫ B ( z , r n ) ϕ n ( y ) d y = ( 1 + ε + o ( 1 )) K . (80)
Again, by Chebyshev’s inequality, P ( N 0 < K ) ≤ E N 0 ( E N 0 − K ) 2 = O ( 1 K ) \mathbb{P}(N_0<K)\le\frac{\mathbb{E}N_0}{(\mathbb{E}N_0-K)^2}=O\left(\frac{1}{K}\right) P ( N 0 < K ) ≤ ( E N 0 − K ) 2 E N 0 = O ( K 1 ) . Define
E n : = { N 0 ≥ K } ∩ ( ⋂ s ∈ S { N s ≥ K + 1 } ) . E_n:=\{N_0\ge K\}\cap\left(\bigcap_{s\in S}\{N_s\ge K+1\}\right). E n := { N 0 ≥ K } ∩ ( s ∈ S ⋂ { N s ≥ K + 1 } ) .
Since S S S is finite and K = K n → ∞ K=K_n\to\infty K = K n → ∞ , a union bound gives,
P ( E n c ) ≤ P ( N 0 < K ) + ∑ s ∈ S P ( N s < K + 1 ) = O ( 1 K ) → 0. \mathbb{P}(E_n^c)\le\mathbb{P}(N_0<K)+\sum_{s\in S}\mathbb{P}(N_s<K+1)=O\left(\frac{1}{K}\right)\to0. P ( E n c ) ≤ P ( N 0 < K ) + s ∈ S ∑ P ( N s < K + 1 ) = O ( K 1 ) → 0.
Note that on the event E n E_n E n every undirected neighbor of z z z lies in B ( z , r n ) B(z,r_n) B ( z , r n ) , hence the degree in this case is at most N 0 N_0 N 0 . Then using the degree bound (61) on the event E n c E_n^c E n c gives,
E [ d G K ( P n ϕ n ) ( z ) ] ≤ E [ N 0 ] + B d K P ( E n c ) ≤ n ∫ B ( z , r n ) ϕ n ( y ) d y + O ( 1 ) = ( 1 + ε + o ( 1 ) ) K , \mathbb{E}\left[d_{G_K(\mathcal{P}_{n\phi_n})}(z)\right]\le\mathbb{E}[N_0]+B_dK\mathbb{P}(E_n^c)\le n\int_{B(z,r_n)}\phi_n(y)\,dy+O(1)=(1+\varepsilon+o(1))K, E [ d G K ( P n ϕ n ) ( z ) ] ≤ E [ N 0 ] + B d K P ( E n c ) ≤ n ∫ B ( z , r n ) ϕ n ( y ) d y + O ( 1 ) = ( 1 + ε + o ( 1 )) K ,
where the last step uses (B.15). Then dividing both sides by K K K and taking limits, the result follows. □
Proof of Corollary B.2
The result in (1) is immediate from Theorem 1.6 .
For the consistency result, we first show that
σ G N ( ∙ , G K ( Z n ) ) = O P ( 1 n ) . (81) \frac{\sigma_{\mathcal{G}}}{N(\bullet,\mathcal{G}_K(\mathcal{Z}_n))}=O_P\left(\frac{1}{\sqrt{n}}\right). \tag*{(81)} N ( ∙ , G K ( Z n )) σ G = O P ( n 1 ) . (81)
For this, recalling (9) and denoting d z = d G K ( Z n ) ( z ) d_z=d_{\mathcal{G}_K(\mathcal{Z}_n)}(z) d z = d G K ( Z n ) ( z ) gives,
σ G 2 ≤ 1 4 ( N ( ∙ , G K ( Z n ) ) + ∑ z ∈ Z n ( d z 2 ) ) = 1 8 ∑ z ∈ Z n d z 2 . \sigma_{\mathcal{G}}^2\leq\frac{1}{4}\left(N(\bullet,\mathcal{G}_K(\mathcal{Z}_n))+\sum_{z\in\mathcal{Z}_n}\binom{d_z}{2}\right)=\frac{1}{8}\sum_{z\in\mathcal{Z}_n}d_z^2. σ G 2 ≤ 4 1 ( N ( ∙ , G K ( Z n )) + z ∈ Z n ∑ ( 2 d z ) ) = 8 1 z ∈ Z n ∑ d z 2 .
Let K ~ = min { K , max { L n − 1 , 0 } } \widetilde{K}=\min\{K,\max\{L_n-1,0\}\} K = min { K , max { L n − 1 , 0 }} . Then by Lemma 5.1,
σ G 2 N ( ∙ , G K ( Z n ) ) 2 ≤ B d 2 2 L n = O P ( 1 n ) , \frac{\sigma_{\mathcal{G}}^2}{N(\bullet,\mathcal{G}_K(\mathcal{Z}_n))^2}\leq\frac{B_d^2}{2L_n}=O_P\left(\frac{1}{n}\right), N ( ∙ , G K ( Z n ) ) 2 σ G 2 ≤ 2 L n B d 2 = O P ( n 1 ) ,
since L n n → 1 \frac{L_n}{n}\to1 n L n → 1 in probability. This proves (B.16). Now fix an alternative with f f f and g g g differing on a set of positive Lebesgue measure. By Proposition B.1,
T × ( G K ( Z n ) ) N ( ∙ , G K ( Z n ) ) → P δ λ ( f , g ) < 2 λ ( 1 − λ ) . \frac{T^\times(\mathcal{G}_K(\mathcal{Z}_n))}{N(\bullet,\mathcal{G}_K(\mathcal{Z}_n))}\xrightarrow{P}\delta_\lambda(f,g)<2\lambda(1-\lambda). N ( ∙ , G K ( Z n )) T × ( G K ( Z n )) P δ λ ( f , g ) < 2 λ ( 1 − λ ) .
Also, μ G × N ( ∙ , G K ( Z n ) ) → 2 λ ( 1 − λ ) \frac{\mu^\times_{\mathcal{G}}}{N(\bullet,\mathcal{G}_K(\mathcal{Z}_n))}\to2\lambda(1-\lambda) N ( ∙ , G K ( Z n )) μ G × → 2 λ ( 1 − λ ) . Combining the above with (B.16) gives,
μ G × − z α σ G − T × ( G K ( Z n ) ) N ( ∙ , G K ( Z n ) ) → P 2 λ ( 1 − λ ) − δ λ ( f , g ) > 0. \frac{\mu^\times_{\mathcal{G}}-z_\alpha\sigma_{\mathcal{G}}-T^\times(\mathcal{G}_K(\mathcal{Z}_n))}{N(\bullet,\mathcal{G}_K(\mathcal{Z}_n))}\xrightarrow{P}2\lambda(1-\lambda)-\delta_\lambda(f,g)>0. N ( ∙ , G K ( Z n )) μ G × − z α σ G − T × ( G K ( Z n )) P 2 λ ( 1 − λ ) − δ λ ( f , g ) > 0.
Therefore,
P H 1 ( ψ n = 1 ) = P H 1 ( T × ( G K ( Z n ) ) < μ G × − z α σ G ) → 1. \mathbb{P}_{H_1}(\psi_n=1)=\mathbb{P}_{H_1}\left(T^\times(\mathcal{G}_K(\mathcal{Z}_n))<\mu^\times_{\mathcal{G}}-z_\alpha\sigma_{\mathcal{G}}\right)\to1. P H 1 ( ψ n = 1 ) = P H 1 ( T × ( G K ( Z n )) < μ G × − z α σ G ) → 1.
This proves the result in (2). □
Siva Athreya, International Centre for Theoretical Sciences, Tata Institute of Fundamental Research, Survey No. 151, Shivakote, Hesaraghatta Hobli, Bengaluru–560089, India. athreya@icts.res.in
Bhaswar B. Bhattacharya, Department of Statistics and Data Science, The Wharton School, University of Pennsylvania, Academic Research Building, 265 South 37th Street, 3rd and 4th Floors, Philadelphia, PA 19104–1686, USA, and Department of Statistics and Data Science, Faculty of Science, National University of Singapore, Block S16, Level 7, 6 Science Drive 2, Singapore 117546. bhaswar.bhattacharya@gmail.com
Yogesh Dahiya, Department of Mathematical Sciences, Indian Institute of Science Education and Research Mohali, Knowledge City, Sector 81, SAS Nagar, Manauli PO, Punjab 140306, India.
ph20002@iisermohali.ac.in
Karambir Das, Department of Mathematics, Indian Institute of Science, Bengaluru 560012, India.
karambirdas@iisc.ac.in
Spandan Ghoshal, Indian Statistical Institute, 203 B. T. Road, Kolkata 700108, India.
spandan.ghoshal_r@isical.ac.in
Sourish Maniyar, Department of Mathematics, Indian Institute of Science, Bengaluru 560012, India.
sourishparag@iisc.ac.in
Elizabeth Sara Roy, International Centre for Theoretical Sciences, Tata Institute of Fundamental Research, Survey No. 151, Shivakote, Hesaraghatta Hobli, Bengaluru–560089, India.
elizabeth.sararoy@icts.res.in