Introduction
Let A A A be the adjacency matrix of a finite graph with eigenvalues λ 1 , … , λ N \lambda_1,\ldots,\lambda_N λ 1 , … , λ N . For a test function f f f , the linear eigenvalue statistic
Tr [ f ( A ) ] = ∑ i = 1 N f ( λ i ) \operatorname{Tr}[f(A)] = \sum_{i=1}^{N} f(\lambda_i) Tr [ f ( A )] = i = 1 ∑ N f ( λ i )
provides a basic summary of the graph spectrum. For polynomial f f f , these statistics are combinations of closed-walk counts, while for more general test functions they capture spectral information used in graph partitioning, clustering, and related learning problems; see, for instance, [29 , 24 , 30 ] . Their fluctuations are well understood for several classical random graph and random matrix models, where independence or weak dependence between matrix entries allows the use of methods from random matrix theory; see, for example, [12 , 10 , 28 , 4 ] . Much less is known for sparse spatial graphs. The shared random geometric feature introduces nontrivial dependencies between edges. These dependencies make the analysis of spectral fluctuations considerably more challenging, particularly in the sparse regime.
A fundamental example of a spatial network is the random geometric graph, in which vertices are randomly distributed and two vertices are connected whenever their distance is below a fixed threshold. We study its generalization, the random connection model , whose vertices form a Poisson point process and whose edges are sampled independently, conditional on the vertex positions, with probabilities determined by a connection profile φ ( ∥ x ∥ d ) \varphi(\lVert x\rVert^d) φ (∥ x ∥ d ) for displacement x x x . The hard-threshold profile recovers the random geometric graph, whereas profiles with unbounded support permit edges at arbitrarily large distances. We also consider a weighted extension in which every vertex carries a mark and the marks of the two endpoints modify the spatial scale of their connection. These random geometric graphs and random connection models are among the standard models of spatial networks; see [26 , 3 ] .
The soft model presents a difficulty that is absent from the hard-threshold setting. A random geometric graph has a deterministic interaction range, but a soft random connection model generally does not: a vertex can be connected to arbitrarily distant points. Moreover, inserting a deterministic vertex requires generating all of its incident edges randomly while retaining the same random edges among the original Poisson vertices. Thus, the local representations commonly used for geometric Poisson functionals do not apply directly. The question is whether the spatial structure still produces volume-order Gaussian fluctuations of the spectrum.
Our first result shows that it does. For polynomial test functions, we prove that the variance of Tr [ f ( A n ) ] \operatorname{Tr}[f(A_n)] Tr [ f ( A n )] , divided by the observation volume, has a finite limit. Under a stronger integrability assumption on the connection profile and corresponding moment assumptions on the vertex marks, we then prove a quantitative central limit theorem. When the limiting variance is positive, the standardized statistic has Wasserstein and Kolmogorov distance of order ∣ W n ∣ − 1 / 2 |W_n|^{-1/2} ∣ W n ∣ − 1/2 from a standard Gaussian. The integrability condition permits connection profiles with sufficiently fast polynomial decay.
Our second objective is to move beyond polynomials. Under an exponential moment assumption on the marks, we extend the central limit theorem to a weighted Sobolev class of general test functions. The limiting variance is obtained as the continuous extension of the variance form on polynomials. Thus, the polynomial and general results describe two complementary aspects of the same fluctuation theory: the first gives a quantitative normal approximation under comparatively weak moment assumptions, while the second gives a substantially larger class of spectral observables under stronger control of the marks.
The most closely related result is the recent central limit theorem for linear eigenvalue statistics of sparse random geometric graphs in [18 ] . The present work replaces the deterministic distance threshold by a soft connection profile and allows additional vertex marks and interaction kernels. To the best of our knowledge, this gives the first central limit theorems for linear eigenvalue statistics of sparse random connection models with genuinely soft connection profiles. Existing spectral results for spatial graphs have otherwise focused largely on limiting spectral distributions or denser regimes; see, for example, [16 , 1 ] . Spectral questions for geometric networks are also motivated by statistical problems. Higher-order spectral clustering has been studied for soft geometric block models in [2 ] . This work illustrates that spectral information in spatial networks is relevant not only to global graph structure, but also to concrete inference problems.
The proof for polynomial test functions is based on two simple structural ideas. First, we place all edge randomness on a common edge-marked probability space. Insertion differences can then be formed without resampling edges between the original Poisson vertices. Second, powers of the adjacency matrix are expanded into closed walks. Products of insertion differences therefore reduce to connected collections of finitely many walks, whose spatial integrals can be controlled by passing to spanning trees and integrating their vertices successively. This turns the dependence created by the geometry into tree integrals involving only the connection profile and moments of the marks. The resulting bounds allow us to apply the normal-approximation estimates for edge-marked Poisson processes from [21 ] ; related stabilization methods for geometric Poisson and Gibbs functionals appear in [20 , 19 ] .
The extension to general test functions uses a different mechanism. We approximate the test function by polynomials and prove a uniform variance estimate for exponentially twisted statistics. The main model-specific input is an exponential trace bound. To obtain it, we orient the edges according to the vertex marks and dominate the adjacency matrix, in quadratic-form order, by a diagonal matrix involving the resulting incoming and outgoing degrees. The exponential moment assumption on the marks then provides the required degree estimates. Together with a Fourier–Duhamel argument, this yields the variance bound needed to pass from polynomials to general test functions.
The remainder of the paper is organized as follows. § 2 introduces the random connection model, its weighted extension, and the main results. § 3 proves the variance asymptotics and gives sufficient conditions for positivity of the limiting variance. § 4 proves the quantitative central limit theorem for polynomial test functions. Finally, § 5 extends the central limit theorem to general test functions.
Model and Main Results
This section introduces the random connection model and its marked extension. It then defines the linear eigenvalue statistics and records the assumptions on the connection profile, interaction kernel, and marks. The section concludes with the variance asymptotics and central limit theorems that guide the later proofs.
The random connection model
The random connection model. The unmarked random connection model provides the basic geometric construction underlying the marked model used in the main results. Its vertices form a homogeneous Poisson point process in an expanding window, and distinct pairs are connected independently according to their distance. We first define this model and its adjacency matrix from [11 , 27 ] before introducing the vertex marks and interaction kernel.
Fix d ≥ 1 d \ge1 d ≥ 1 . For n ≥ 1 n \ge1 n ≥ 1 , let P n \mathcal{P}_n P n be a homogeneous Poisson point process with intensity 1 1 1 on W n : = [ − n 1 / d / 2 , n 1 / d / 2 ] d ⊆ R d W_n := [-n^{1/d}/2,n^{1/d}/2]^d \subseteq\mathbb{R}^d W n := [ − n 1/ d /2 , n 1/ d /2 ] d ⊆ R d . Then, ∣ W n ∣ = n |W_n| = n ∣ W n ∣ = n , and the asymptotic regime is n → ∞ n \to\infty n → ∞ with the dimension d d d fixed. We denote by N n : = ∣ P n ∣ N_n := |\mathcal{P}_n| N n := ∣ P n ∣ the random number of points in P n \mathcal{P}_n P n and choose an arbitrary enumeration P n = { X 1 , … , X N n } \mathcal{P}_n = \{X_1,\ldots,X_{N_n}\} P n = { X 1 , … , X N n } .
Let φ : [ 0 , ∞ ) → [ 0 , 1 ] \varphi: [0,\infty) \to[0,1] φ : [ 0 , ∞ ) → [ 0 , 1 ] be a non-increasing function, called the connection profile, satisfying
M φ : = ∫ R d φ ( ∥ x ∥ d ) d x < ∞ . (1) M_{\varphi} := \int_{\mathbb{R}^d} \varphi(\|x\|^d)\,\mathrm{d}x < \infty. \tag*{(1)} M φ := ∫ R d φ ( ∥ x ∥ d ) d x < ∞. (1)
More generally, for γ ∈ ( 0 , 1 ] \gamma\in(0,1] γ ∈ ( 0 , 1 ] , set
M φ , γ : = ∫ R d φ ( ∥ x ∥ d ) γ d x . (2) M_{\varphi,\gamma} := \int_{\mathbb{R}^d} \varphi(\|x\|^d)^\gamma\,\mathrm{d}x. \tag*{(2)} M φ , γ := ∫ R d φ ( ∥ x ∥ d ) γ d x . (2)
The random connection model (RCM) is the random graph with vertex set P n \mathcal{P}_n P n . Conditionally on P n \mathcal{P}_n P n , each pair of distinct points X i , X j ∈ P n X_i,X_j \in\mathcal{P}_n X i , X j ∈ P n is connected by an edge with probability φ ( ∥ X i − X j ∥ d ) \varphi(\|X_i-X_j\|^d) φ ( ∥ X i − X j ∥ d ) , independently of all other pairs. The hard threshold case φ ( r ) = 1 { r ≤ R } \varphi(r)=\mathbf{1}\{r\le R\} φ ( r ) = 1 { r ≤ R } , for fixed R > 0 R>0 R > 0 , produces the random geometric graph from [18 ] as a special case. The adjacency matrix A n ∈ R N n × N n A_n \in\mathbb{R}^{N_n\times N_n} A n ∈ R N n × N n of the RCM is defined by
A n ( i , j ) : = A n ( X i , X j ) : = 1 { { X i , X j } is an edge } , 1 ≤ i , j ≤ N n , (3) A_n(i,j):=A_n(X_i,X_j):=\mathbf{1}\{\{X_i,X_j\}\text{ is an edge}\},\qquad1\le i,j\le N_n, \tag*{(3)} A n ( i , j ) := A n ( X i , X j ) := 1 {{ X i , X j } is an edge } , 1 ≤ i , j ≤ N n , (3)
with A n ( i , i ) = 0 A_n(i,i)=0 A n ( i , i ) = 0 .
We work in an increasing-volume regime: the point intensity and the connection profile remain fixed, while the observation window expands. In particular, if a vertex is inserted at x ∈ W n x\in W_n x ∈ W n with incident edges added independently with probabilities prescribed by φ \varphi φ , then the expected degree of x x x is
∫ W n φ ( ∥ x − y ∥ d ) d y ≤ M φ , \int_{W_n}\varphi(\|x-y\|^d)\,\mathrm{d}y\le M_{\varphi}, ∫ W n φ ( ∥ x − y ∥ d ) d y ≤ M φ ,
uniformly in n n n and x x x . The expected number of edges is therefore at most M φ ∣ W n ∣ / 2 M_{\varphi}|W_n|/2 M φ ∣ W n ∣/2 , whereas the expected number of vertices is ∣ W n ∣ |W_n| ∣ W n ∣ . Thus, the RCM is sparse in the sense that the mean degree stays bounded as volume increases.
The weighted RCM
Defining the RCM as above, with the hard-threshold random geometric graph as a special case, the degree of a typical vertex follows a Poisson distribution. This is a limitation when modeling networks with degree heterogeneity, in which highly connected hubs coexist with many vertices of small degree. Power-law degree distributions are useful models for such heterogeneity [7 ] .
Vertex weights provide a way to model differences in the tendency of individual vertices to form connections. They may represent the attractiveness of a location or the connectivity capacity of a communication node. These interpretations motivate separating the spatial positions, which encode geographical constraints, from the vertex weights.
We now precisely define the weighted RCM from [13 , 14 , 15 ] . Attach to the points X 1 , … , X N n X_1,\ldots,X_{N_n} X 1 , … , X N n independent marks U 1 , … , U N n U_1,\ldots,U_{N_n} U 1 , … , U N n with common distribution μ \mu μ on [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) , independently of P n \mathcal{P}_n P n . The resulting marked Poisson process is P ~ n = { ( X i , U i ) : 1 ≤ i ≤ N n } \widetilde{\mathcal{P}}_n=\{(X_i,U_i):1\le i\le N_n\} P n = {( X i , U i ) : 1 ≤ i ≤ N n } on W ~ n : = W n × [ 1 , ∞ ) \widetilde{W}_n:=W_n\times[1,\infty) W n := W n × [ 1 , ∞ ) . We write λ ( d x , d u ) : = d x μ ( d u ) \lambda(\mathrm{d}x,\mathrm{d}u):=\mathrm{d}x\,\mu(\mathrm{d}u) λ ( d x , d u ) := d x μ ( d u ) for its intensity measure. Throughout, U U U denotes a random variable with distribution μ \mu μ .
Let φ \varphi φ be the connection profile introduced in Subsection 2.1, and let κ : [ 1 , ∞ ) 2 → [ 1 , ∞ ) \kappa:[1,\infty)^2\to[1,\infty) κ : [ 1 , ∞ ) 2 → [ 1 , ∞ ) be a symmetric measurable function, called the interaction kernel. Given the marked configuration P ~ n \widetilde{\mathcal{P}}_n P n , the weighted RCM has spatial vertex set P n \mathcal{P}_n P n . For distinct marked points ( X i , U i ) , ( X j , U j ) ∈ P ~ n (X_i,U_i),(X_j,U_j)\in\widetilde{\mathcal{P}}_n ( X i , U i ) , ( X j , U j ) ∈ P n , place an edge between X i X_i X i and X j X_j X j with probability
φ ( ∥ X i − X j ∥ d κ ( U i , U j ) ) , \varphi\left(\frac{\|X_i-X_j\|^d}{\kappa(U_i,U_j)}\right), φ ( κ ( U i , U j ) ∥ X i − X j ∥ d ) ,
independently for different unordered pairs, conditional on P ~ n \widetilde{\mathcal{P}}_n P n .
Intuitively, κ ( u , v ) \kappa(u,v) κ ( u , v ) measures the tendency for two vertices with weights u u u and v v v to connect. For interaction kernels increasing in both arguments, such as κ ( u , v ) = u v \kappa(u,v)=uv κ ( u , v ) = uv , vertices with larger weights have a greater propensity to form connections and can interact over larger spatial scales. This mechanism underlies continuum scale-free percolation [9 ] .
The unweighted RCM is recovered as a special case by taking U ≡ 1 U\equiv1 U ≡ 1 and κ ≡ 1 \kappa\equiv1 κ ≡ 1 . The effect of the weights on the degree distribution can be seen directly from the connection rule. Consider the infinite-volume model and insert a vertex at the origin with an independent mark U 0 U_0 U 0 having distribution μ \mu μ and incident edges added independently. Let D D D be its degree. For every u u u such that E [ κ ( u , U ) ] < ∞ \mathbb{E}[\kappa(u,U)]<\infty E [ κ ( u , U )] < ∞ , the neighboring vertices form an independent thinning of the marked Poisson process and hence
D ∣ { U 0 = u } ∼ Poisson ( Λ ( u ) ) , Λ ( u ) : = ∫ R d ∫ [ 1 , ∞ ) φ ( ∥ x ∥ d κ ( u , v ) ) μ ( d v ) d x = M φ E [ κ ( u , U ) ] . D \mid\{U_0=u\} \sim\operatorname{Poisson}(\Lambda(u)), \qquad\Lambda(u):=\int_{\mathbb{R}^d}\int_{[1,\infty)}\varphi\left(\frac{\lVert x\rVert^d}{\kappa(u,v)}\right)\mu(\mathrm{d}v)\,\mathrm{d}x=M_{\varphi}\mathbb{E}[\kappa(u,U)]. D ∣ { U 0 = u } ∼ Poisson ( Λ ( u )) , Λ ( u ) := ∫ R d ∫ [ 1 , ∞ ) φ ( κ ( u , v ) ∥ x ∥ d ) μ ( d v ) d x = M φ E [ κ ( u , U )] .
Consequently, whenever Λ ( U 0 ) < ∞ \Lambda(U_0)<\infty Λ ( U 0 ) < ∞ almost surely, the unconditional degree distribution is mixed Poisson. This provides a direct mechanism for introducing degree heterogeneity while retaining a homogeneous spatial distribution of vertices.
A particular example is the kernel κ ( u , v ) = u v \kappa(u,v)=uv κ ( u , v ) = uv . If 0 < M φ < ∞ 0<M_{\varphi}<\infty 0 < M φ < ∞ and E U < ∞ \mathbb{E}U<\infty E U < ∞ , then
Λ ( u ) = a u , a : = M φ E U . \Lambda(u)=au,\qquad a:=M_{\varphi}\mathbb{E}U. Λ ( u ) = a u , a := M φ E U .
Thus, the vertex weight is proportional to the conditional expected degree. For example, if the marks have a Pareto distribution, P ( U > t ) = t − β \mathbb{P}(U>t)=t^{-\beta} P ( U > t ) = t − β for t ≥ 1 t\ge1 t ≥ 1 with β > 1 \beta>1 β > 1 , then the degree distribution satisfies
P ( D ≥ k ) ∼ P ( a U ≥ k ) = a β k − β , k → ∞ . \mathbb{P}(D\ge k)\sim\mathbb{P}(aU\ge k)=a^\beta k^{-\beta},\qquad k\to\infty. P ( D ≥ k ) ∼ P ( a U ≥ k ) = a β k − β , k → ∞.
The degree therefore inherits the power-law tail of the marks. In particular, power-law degrees can arise from vertex heterogeneity even when the connection profile itself has bounded support.
Henceforth, A n A_n A n denotes the adjacency matrix of this graph. We identify a marked vertex x ˇ = ( x , u ) \check{x}=(x,u) x ˇ = ( x , u ) with its spatial coordinate when it is used as a vertex argument of a graph quantity. In particular,
A n ( x ˇ , y ˇ ) : = A n ( x , y ) . A_n(\check{x},\check{y}):=A_n(x,y). A n ( x ˇ , y ˇ ) := A n ( x , y ) .
For a graph with spatial vertex set Q Q Q and its associated vertex marks, write Q ˇ \check{Q} Q ˇ for the corresponding marked configuration.
Assumptions
The proofs require separate assumptions on the connection profile, the interaction kernel, and the mark distribution. We always assume that (1) holds for the connection profile. The quantitative CLTs further require that M φ , 1 / 4 < ∞ M_{\varphi,1/4}<\infty M φ , 1/4 < ∞ with M φ , 1 / 4 M_{\varphi,1/4} M φ , 1/4 defined in (2). This stronger condition remains compatible with long-range connections. For example, for
φ ( t ) = ( 1 + t ) − α , α > 1 , \varphi(t)=(1+t)^{-\alpha},\qquad\alpha>1, φ ( t ) = ( 1 + t ) − α , α > 1 ,
polar coordinate integration yields
M φ , γ < ∞ ⟺ α γ > 1. M_{\varphi,\gamma}<\infty\quad\Longleftrightarrow\quad\alpha\gamma>1. M φ , γ < ∞ ⟺ α γ > 1.
Thus, the condition M φ , 1 / 4 < ∞ M_{\varphi,1/4}<\infty M φ , 1/4 < ∞ permits this profile whenever α > 4 \alpha>4 α > 4 . The absence of d d d from this threshold reflects our convention of evaluating the profile at the d d d th power of the Euclidean distance.
For polynomial test functions, we assume finite moments of the vertex marks depending on the polynomial degree. For general test functions, we impose the following stronger exponential-moment condition, which is satisfied, for example, by shifted exponential distributions. We say that the mark distribution satisfies Condition (EM) if there exists η > 0 \eta>0 η > 0 such that
E [ e η U ] < ∞ . (4) \mathbb{E}[\mathrm{e}^{\eta U}]<\infty. \tag*{(4)} E [ e η U ] < ∞. (4)
The key assumption on our kernel κ \kappa κ will be that it is dominated by the product kernel . More precisely, we say that the interaction kernel satisfies Condition (PK) if there exists c κ ≥ 1 c_\kappa\ge1 c κ ≥ 1 such that
1 ≤ κ ( u , v ) ≤ c κ u v for all u , v ≥ 1. (5) 1\le\kappa(u,v)\le c_\kappa uv\qquad\text{for all }u,v\ge1. \tag*{(5)} 1 ≤ κ ( u , v ) ≤ c κ uv for all u , v ≥ 1. (5)
The assumptions separate two sources of variability. Integrability of the connection profile controls the spatial extent of interactions, whereas moments of the marks control the heterogeneity of the interaction scales. Condition (PK) links these two controls: integrating one connection probability over a spatial coordinate produces the factor M φ κ ( u , v ) M_\varphi\kappa(u,v) M φ κ ( u , v ) , which is at most c κ M φ u v c_\kappa M_\varphi uv c κ M φ uv . Repeated spatial integration therefore introduces powers of the vertex marks. This explains why the required mark moments depend on the degree of the polynomial test function (see Theorem 2 below).
These conditions are sufficient for the estimates used below; we do not claim that they are necessary for Gaussian fluctuations.
Linear eigenvalue statistics
Spectrum of the adjacency matrix A n A_n A n is the random object studied throughout the paper. This subsection introduces the associated linear eigenvalue statistics and fixes the notation used in the main results.
Let λ 1 , … , λ N n \lambda_1,\ldots,\lambda_{N_n} λ 1 , … , λ N n denote the real eigenvalues of A n A_n A n . For a test function f : R → R f:\mathbb{R}\to\mathbb{R} f : R → R , define
L n ( f ) : = ∑ i = 1 N n f ( λ i ) = Tr [ f ( A n ) ] . (6) L_n(f):=\sum_{i=1}^{N_n} f(\lambda_i)=\operatorname{Tr}[f(A_n)]. \tag*{(6)} L n ( f ) := i = 1 ∑ N n f ( λ i ) = Tr [ f ( A n )] . (6)
Our goal is to establish a finite limit for the variance of L n ( f ) L_n(f) L n ( f ) divided by ∣ W n ∣ |W_n| ∣ W n ∣ and a central limit theorem for the centered statistic normalized by ∣ W n ∣ \sqrt{|W_n|} ∣ W n ∣ . When the limiting variance is positive, this also yields a standard normal approximation under normalization by the finite-volume standard deviation.
Main theorems
This subsection states the three principal results of the paper. The first establishes variance asymptotics for polynomial test functions, the second proves a quantitative central limit theorem in that class, and the third extends qualitative convergence to a general test-function space. Their order reflects the proof structure: the variance limit enters the polynomial central limit theorem, which then provides the approximations used for general test functions.
Our first result establishes the existence of the limiting variance for polynomial test functions.
Theorem 1 (Variance asymptotics). Let f f f be a polynomial of degree m ≥ 1 m\ge1 m ≥ 1 . Assume that (PK) holds and that E [ U 3 m − 1 ] < ∞ \mathbb{E}[U^{3m-1}]<\infty E [ U 3 m − 1 ] < ∞ . Then, the following limit exists and is finite:
lim n → ∞ ∣ W n ∣ − 1 Var ( Tr [ f ( A n ) ] ) . \lim_{n\to\infty}|W_n|^{-1}\operatorname{Var}(\operatorname{Tr}[f(A_n)]). n → ∞ lim ∣ W n ∣ − 1 Var ( Tr [ f ( A n )]) .
For every polynomial satisfying the hypotheses of Theorem 1 , we denote this limit by σ f 2 \sigma_f^2 σ f 2 . Proposition 5 below provides conditions for positivity of σ f 2 \sigma_f^2 σ f 2 .
Our second result establishes a quantitative central limit theorem for polynomial test functions. For integrable real-valued random variables X X X and Y Y Y , let
d W ( X , Y ) : = sup Lip ( h ) ≤ 1 ∣ E [ h ( X ) ] − E [ h ( Y ) ] ∣ d_W(X,Y):=\sup_{\operatorname{Lip}(h)\le1}\left|\mathbb{E}[h(X)]-\mathbb{E}[h(Y)]\right| d W ( X , Y ) := Lip ( h ) ≤ 1 sup ∣ E [ h ( X )] − E [ h ( Y )] ∣
denote the Wasserstein distance between their distributions and let
d K ( X , Y ) : = sup t ∈ R ∣ P ( X ≤ t ) − P ( Y ≤ t ) ∣ d_K(X,Y):=\sup_{t\in\mathbb{R}}\left|\mathbb{P}(X\le t)-\mathbb{P}(Y\le t)\right| d K ( X , Y ) := t ∈ R sup ∣ P ( X ≤ t ) − P ( Y ≤ t ) ∣
denote their Kolmogorov distance. In addition, N ( μ , σ 2 ) \mathcal{N}(\mu,\sigma^2) N ( μ , σ 2 ) denotes the normal distribution with mean μ \mu μ and variance σ 2 \sigma^2 σ 2 .
Theorem 2 (CLT for polynomial test functions). Let f f f be a polynomial of degree m ≥ 1 m\ge1 m ≥ 1 . Assume that (PK) holds, that M φ , 1 / 4 < ∞ M_{\varphi,1/4}<\infty M φ , 1/4 < ∞ , and that E [ U 6 m ] < ∞ \mathbb{E}[U^{6m}]<\infty E [ U 6 m ] < ∞ . Then,
Tr [ f ( A n ) ] − E [ Tr [ f ( A n ) ] ] ∣ W n ∣ → d N ( 0 , σ f 2 ) . \frac{\operatorname{Tr}[f(A_n)]-\mathbb{E}[\operatorname{Tr}[f(A_n)]]}{\sqrt{|W_n|}}
\xrightarrow{d}\mathcal{N}(0,\sigma_f^2). ∣ W n ∣ Tr [ f ( A n )] − E [ Tr [ f ( A n )]] d N ( 0 , σ f 2 ) . If σ f 2 > 0 \sigma_f^2>0 σ f 2 > 0 , then there exists a constant C < ∞ C<\infty C < ∞ such that
max { d W ( Tr [ f ( A n ) ] − E [ Tr [ f ( A n ) ] ] Var ( Tr [ f ( A n ) ] ) , N ( 0 , 1 ) ) , d K ( Tr [ f ( A n ) ] − E [ Tr [ f ( A n ) ] ] Var ( Tr [ f ( A n ) ] ) , N ( 0 , 1 ) ) } ≤ C ∣ W n ∣ , \max\left\{
d_W\left(
\frac{\operatorname{Tr}[f(A_n)]-\mathbb{E}[\operatorname{Tr}[f(A_n)]]}{\sqrt{\operatorname{Var}(\operatorname{Tr}[f(A_n)])}},
\mathcal{N}(0,1)
\right),
d_K\left(
\frac{\operatorname{Tr}[f(A_n)]-\mathbb{E}[\operatorname{Tr}[f(A_n)]]}{\sqrt{\operatorname{Var}(\operatorname{Tr}[f(A_n)])}},
\mathcal{N}(0,1)
\right)
\right\}
\le\frac{C}{\sqrt{|W_n|}}, max { d W ( Var ( Tr [ f ( A n )]) Tr [ f ( A n )] − E [ Tr [ f ( A n )]] , N ( 0 , 1 ) ) , d K ( Var ( Tr [ f ( A n )]) Tr [ f ( A n )] − E [ Tr [ f ( A n )]] , N ( 0 , 1 ) ) } ≤ ∣ W n ∣ C , where C C C may depend on f f f and the fixed model parameters, but not on n n n .
Our third result extends the central limit theorem to a broader class of test functions. For c > 0 c>0 c > 0 , let H c 2 \mathcal{H}_c^2 H c 2 be the space of twice weakly differentiable functions f : R → R f:\mathbb{R}\to\mathbb{R} f : R → R such that
∥ f ∥ H c 2 2 : = ∫ R ( ∣ f ( x ) ∣ 2 + ∣ f ′ ( x ) ∣ 2 + ∣ f ′ ′ ( x ) ∣ 2 ) sech 2 ( c x ) d x < ∞ . \|f\|_{\mathcal{H}_c^2}^2:=\int_{\mathbb{R}}\left(|f(x)|^2+|f'(x)|^2+|f''(x)|^2\right)\operatorname{sech}^2(cx)\,\mathrm{d}x<\infty. ∥ f ∥ H c 2 2 := ∫ R ( ∣ f ( x ) ∣ 2 + ∣ f ′ ( x ) ∣ 2 + ∣ f ′′ ( x ) ∣ 2 ) sech 2 ( c x ) d x < ∞.
We further set H c , 0 2 : = { f ∈ H c 2 : f ( 0 ) = 0 } \mathcal{H}_{c,0}^2:=\{f\in\mathcal{H}_c^2:f(0)=0\} H c , 0 2 := { f ∈ H c 2 : f ( 0 ) = 0 } . Every element of H c 2 \mathcal{H}_c^2 H c 2 belongs to H l o c 2 ( R ) H_{\mathrm{loc}}^2(\mathbb{R}) H loc 2 ( R ) , since the weight sech 2 ( c x ) \operatorname{sech}^2(cx) sech 2 ( c x ) is bounded below on compact intervals. We always use its unique locally C 1 C^1 C 1 representative. In particular, f ( 0 ) f(0) f ( 0 ) and the spectral values f ( λ i ( A n ) ) f(\lambda_i(A_n)) f ( λ i ( A n )) refer to this representative.
The exponential weight serves two purposes. First, unlike the usual unweighted Sobolev space, H c 2 \mathcal{H}^{2}_{c} H c 2 contains every polynomial, so it provides a common space for the polynomial CLT and its extension. Second, it allows test functions that grow at infinity, while still controlling the two derivatives. The condition f ( 0 ) = 0 f(0)=0 f ( 0 ) = 0 is convenient when comparing adjacency matrices before and after a vertex insertion. Adding an isolated vertex appends a zero eigenvalue, and this condition ensures that such insertion does not change the statistic. The following theorem extends the polynomial CLT to this test-function space.
Theorem 3 (CLT for general test functions). Assume that (EM) and (PK) hold and that M φ , 1 / 4 < ∞ M_{\varphi,1/4}<\infty M φ , 1/4 < ∞ . Let c > 0 c>0 c > 0 and f ∈ H c , 0 2 f\in\mathcal{H}^{2}_{c,0} f ∈ H c , 0 2 . Then, there exists σ f 2 ≥ 0 \sigma_f^2\ge0 σ f 2 ≥ 0 such that
Tr [ f ( A n ) ] − E [ Tr [ f ( A n ) ] ] ∣ W n ∣ → d N ( 0 , σ f 2 ) as n → ∞ . \frac{\operatorname{Tr}[f(A_n)]-\mathbb{E}[\operatorname{Tr}[f(A_n)]]}{\sqrt{|W_n|}} \xrightarrow{d} \mathcal{N}(0,\sigma_f^2) \qquad\text{as } n\to\infty. ∣ W n ∣ Tr [ f ( A n )] − E [ Tr [ f ( A n )]] d N ( 0 , σ f 2 ) as n → ∞. The proof identifies σ f 2 \sigma_f^2 σ f 2 with the value at f f f of the continuous extension of the limiting variance form on polynomials.
Acknowledgments
C.H. is supported by a research grant (VIL69126) from VILLUM FONDEN. K.N. is supported by Samsung Science and Technology Foundation under Project Number SSTF-BA2202-02. M.O. is supported by the NWO Gravitation project NETWORKS under grant agreement No. 024.002.003. This article is based upon work from COST Action 24122 mSPACE, supported by COST (European Cooperation in Science and Technology).
Proof of the variance asymptotics
This section proves the existence of a limiting volume-normalized variance and records a sufficient criterion for its positivity. The proof relies on variance asymptotics results for weakly stabilizing functionals from [8 , 25 ] . Verifying the conditions of these results involves studying the difference when removing a fixed vertex from the RCM. To make this rigorous, we recall a construction from [21 ] of the RCM and its difference operators from an edge-marked Poisson process in § 3.2 . This construction will be used again for the proof of the CLT in Theorem 2 . We recall the main result from [8 , 25 ] in § 3.3 before proving the main theorem in § 3.4 .
Notation
Let G G G denote a finite graph with (deterministic) vertex set P P P and adjacency matrix A A A . We denote the entry of A A A corresponding to x 1 , x 2 ∈ P x_1,x_2\in P x 1 , x 2 ∈ P by A ( x 1 , x 2 ; G ) A(x_1,x_2;G) A ( x 1 , x 2 ; G ) . Consider the linear eigenvalue statistic L ( f ; G ) = Tr [ f ( A ) ] L(f;G)=\operatorname{Tr}[f(A)] L ( f ; G ) = Tr [ f ( A )] with a polynomial test function f f f . For integers 1 ≤ p ≤ m 1\le p\le m 1 ≤ p ≤ m , define
Σ m , p : = { π : { 1 , … , m } → { 1 , … , p } surjective } . (7) \Sigma_{m,p}:=\{\pi:\{1,\ldots,m\}\to\{1,\ldots,p\}\text{ surjective}\}. \tag*{(7)} Σ m , p := { π : { 1 , … , m } → { 1 , … , p } surjective } . (7)
This corresponds to the set of ordered partitions of { 1 , … , m } \{1,\ldots,m\} { 1 , … , m } into p p p nonempty sets. We also define
Σ m , p ∘ : = { π ∈ Σ m , p : π ( j ) ≠ π ( j + 1 ) for j = 1 , … , m } , \Sigma^\circ_{m,p}:=\{\pi\in\Sigma_{m,p}:\pi(j)\ne\pi(j+1)\text{ for }j=1,\ldots,m\}, Σ m , p ∘ := { π ∈ Σ m , p : π ( j ) = π ( j + 1 ) for j = 1 , … , m } ,
where we use the convention π ( m + 1 ) = π ( 1 ) \pi(m+1)=\pi(1) π ( m + 1 ) = π ( 1 ) . Let x 1 , … , x p x_1,\ldots,x_p x 1 , … , x p be p p p mutually distinct vertices in G G G and set
g m , p ( x 1 , … , x p ; G ) : = 1 p ! ∑ π ∈ Σ m , p ∘ ∏ j = 1 m A ( x π ( j ) , x π ( j + 1 ) ; G ) . g_{m,p}(x_1,\ldots,x_p;G):=\frac{1}{p!}\sum_{\pi\in\Sigma^\circ_{m,p}}\prod_{j=1}^{m}A(x_{\pi(j)},x_{\pi(j+1)};G). g m , p ( x 1 , … , x p ; G ) := p ! 1 π ∈ Σ m , p ∘ ∑ j = 1 ∏ m A ( x π ( j ) , x π ( j + 1 ) ; G ) .
Thus, p ! g m , p ( x 1 , … , x p ) p!g_{m,p}(x_1,\ldots,x_p) p ! g m , p ( x 1 , … , x p ) counts based closed walks of length m m m whose visited vertex set is exactly { x 1 , … , x p } \{x_1,\ldots,x_p\} { x 1 , … , x p } . For m ≥ 1 m\ge1 m ≥ 1 , one can write L ( x m ; G ) = Tr [ A m ] L(x^m;G)=\operatorname{Tr}[A^m] L ( x m ; G ) = Tr [ A m ] as a sum of m − 1 m-1 m − 1 U-statistics
L ( x m ; G ) = ∑ p = 2 m ∑ ( x 1 , … , x p ) ∈ P ( p ) 1 p ! ∑ π ∈ Σ m , p ∘ ∏ j = 1 m A ( x π ( j ) , x π ( j + 1 ) ; G ) = ∑ p = 2 m ∑ ( x 1 , … , x p ) ∈ P ( p ) g m , p ( x 1 , … , x p ; G ) , L(x^m;G)=\sum_{p=2}^{m}\ \sum_{(x_1,\ldots,x_p)\in P^{(p)}}\frac{1}{p!}\sum_{\pi\in\Sigma^\circ_{m,p}}\prod_{j=1}^{m}A(x_{\pi(j)},x_{\pi(j+1)};G)=\sum_{p=2}^{m}\ \sum_{(x_1,\ldots,x_p)\in P^{(p)}}g_{m,p}(x_1,\ldots,x_p;G), L ( x m ; G ) = p = 2 ∑ m ( x 1 , … , x p ) ∈ P ( p ) ∑ p ! 1 π ∈ Σ m , p ∘ ∑ j = 1 ∏ m A ( x π ( j ) , x π ( j + 1 ) ; G ) = p = 2 ∑ m ( x 1 , … , x p ) ∈ P ( p ) ∑ g m , p ( x 1 , … , x p ; G ) ,
where P ( p ) P^{(p)} P ( p ) denotes the set of p p p -tuples of pairwise distinct points in P P P . Thus, L ( x m ; G ) L(x^m;G) L ( x m ; G ) counts the total number of closed walks of length m m m in the graph. To see why this holds, start from the matrix-product identity
L ( x m ; G ) = Tr [ A m ] = ∑ ( v 1 , … , v m ) ∈ P m ∏ j = 1 m A ( v j , v j + 1 ; G ) , v m + 1 : = v 1 . L(x^m;G)=\operatorname{Tr}[A^m]=\sum_{(v_1,\ldots,v_m)\in P^m}\prod_{j=1}^{m}A(v_j,v_{j+1};G),\qquad v_{m+1}:=v_1. L ( x m ; G ) = Tr [ A m ] = ( v 1 , … , v m ) ∈ P m ∑ j = 1 ∏ m A ( v j , v j + 1 ; G ) , v m + 1 := v 1 .
Group the vertex sequences ( v 1 , … , v m ) (v_1,\ldots,v_m) ( v 1 , … , v m ) according to the number p p p of distinct vertices they visit, where p ≥ 2 p \ge2 p ≥ 2 because diagonal entries of A A A vanish. For a fixed sequence visiting exactly p p p vertices, choose an ordered enumeration ( X 1 , … , X p ) (X_1,\ldots,X_p) ( X 1 , … , X p ) of those vertices. Then there is a unique surjection π : { 1 , … , m } → { 1 , … , p } \pi:\{1,\ldots,m\}\to\{1,\ldots,p\} π : { 1 , … , m } → { 1 , … , p } such that v j = X π ( j ) v_j=X_{\pi(j)} v j = X π ( j ) for every j j j . As diagonal entries of the adjacency matrix are zero, only terms with π ∈ Σ m , p ∘ \pi\in\Sigma_{m,p}^{\circ} π ∈ Σ m , p ∘ are non-vanishing. Conversely, the ordered tuple and the surjection recover the original vertex sequence. There are exactly p ! p! p ! possible ordered enumerations of its visited vertex set. The factor 1 / p ! 1/p! 1/ p ! removes this multiplicity.
By linearity of L ( ⋅ ; G ) L(\cdot;G) L ( ⋅ ; G ) , a polynomial f ( x ) = a 0 + ⋯ + a m x m f(x)=a_0+\cdots+a_mx^m f ( x ) = a 0 + ⋯ + a m x m satisfies
L ( f ; G ) : = Tr [ f ( A ) ] = a 0 ∣ P ∣ + ∑ q = 2 m a q ∑ p = 2 q ∑ ( X 1 , … , X p ) ∈ P ( p ) g q , p ( X 1 , … , X p ; G ) = a 0 ∣ P ∣ + ∑ p = 2 m ∑ ( X 1 , … , X p ) ∈ P ( p ) ∑ q = p m a q g q , p ( X 1 , … , X p ; G ) . (8) \begin{aligned}
L(f;G):=\operatorname{Tr}[f(A)]&=a_0|\mathcal{P}|+\sum_{q=2}^{m}a_q\sum_{p=2}^{q}\sum_{(X_1,\ldots,X_p)\in\mathcal{P}^{(p)}}g_{q,p}(X_1,\ldots,X_p;G)\\
&=a_0|\mathcal{P}|+\sum_{p=2}^{m}\sum_{(X_1,\ldots,X_p)\in\mathcal{P}^{(p)}}\sum_{q=p}^{m}a_qg_{q,p}(X_1,\ldots,X_p;G). \tag*{(8)}
\end{aligned} L ( f ; G ) := Tr [ f ( A )] = a 0 ∣ P ∣ + q = 2 ∑ m a q p = 2 ∑ q ( X 1 , … , X p ) ∈ P ( p ) ∑ g q , p ( X 1 , … , X p ; G ) = a 0 ∣ P ∣ + p = 2 ∑ m ( X 1 , … , X p ) ∈ P ( p ) ∑ q = p ∑ m a q g q , p ( X 1 , … , X p ; G ) . (8)
Construction of the RCM from edge marked sequences
We first introduce a construction of the RCM from [21 ] that allows a convenient way of adding extra fixed points to the graph. Define a measurable total ordering ⪯ \preceq ⪯ on R d \mathbb{R}^d R d as follows: order two points by their ℓ ∞ \ell_\infty ℓ ∞ -distance to the origin. Two points with the same distance to the origin are ordered lexicographically. Let P ~ \widetilde{\mathcal{P}} P be a Poisson process on R ~ d = R d × [ 1 , ∞ ) \widetilde{\mathbb{R}}^d=\mathbb{R}^d\times[1,\infty) R d = R d × [ 1 , ∞ ) with intensity measure λ \lambda λ and let P \mathcal{P} P denote the underlying Poisson process on R d \mathbb{R}^d R d . The ordering on R d \mathbb{R}^d R d almost surely allows us to label the points in P \mathcal{P} P as P = { X i , i ∈ N } \mathcal{P}=\{X_i,i\in\mathbb{N}\} P = { X i , i ∈ N } with X i ≺ X j X_i\prec X_j X i ≺ X j for i < j i<j i < j . Let [ 0 , 1 ] Z × Z [0,1]^{\mathbb{Z}\times\mathbb{Z}} [ 0 , 1 ] Z × Z denote the space of doubly indexed sequences equipped with the product of uniform measures such that a random element R = ( R i , j ) i , j ∈ Z \mathbf{R}=(R_{i,j})_{i,j\in\mathbb{Z}} R = ( R i , j ) i , j ∈ Z in [ 0 , 1 ] Z × Z [0,1]^{\mathbb{Z}\times\mathbb{Z}} [ 0 , 1 ] Z × Z corresponds to a doubly indexed sequence of i.i.d. uniform random variables R i , j R_{i,j} R i , j . Let R \mathbf{R} R be such a random element which is independent of the entire marked Poisson process P ~ \widetilde{\mathcal{P}} P . Define the edge-marked Poisson process as the pair
ξ = ( P ~ , R ) . \xi=(\widetilde{\mathcal{P}},\mathbf{R}). ξ = ( P , R ) .
We refer to R \mathbf{R} R as the edge-mark sequence; each mark R i , j R_{i,j} R i , j with i < j i<j i < j is associated with the pair of Poisson points X i X_i X i and X j X_j X j . We can build a weighted random connection model G ( ξ ) G(\xi) G ( ξ ) as the graph with vertex set P \mathcal{P} P . If ( X i , U i ) , ( X j , U j ) ∈ P ~ (X_i,U_i),(X_j,U_j)\in\widetilde{\mathcal{P}} ( X i , U i ) , ( X j , U j ) ∈ P with i < j i<j i < j , we put an edge between X i X_i X i and X j X_j X j if
R i , j ≤ φ ( ∥ X i − X j ∥ κ ( U i , U j ) ) . (9) R_{i,j}\le\varphi\left(\frac{\lVert X_i-X_j\rVert}{\kappa(U_i,U_j)}\right). \tag*{(9)} R i , j ≤ φ ( κ ( U i , U j ) ∥ X i − X j ∥ ) . (9)
Fix an ordered tuple
x ~ r : = ( x ~ 1 , … , x ~ r ) , x ~ i = ( x i , u i ) ∈ R d × [ 1 , ∞ ) , \widetilde{x}_r:=(\widetilde{x}_1,\ldots,\widetilde{x}_r),\qquad\widetilde{x}_i=(x_i,u_i)\in\mathbb{R}^d\times[1,\infty), x r := ( x 1 , … , x r ) , x i = ( x i , u i ) ∈ R d × [ 1 , ∞ ) ,
with pairwise distinct spatial coordinates. The tuple order determines negative labels − 1 , … , − r -1,\ldots,-r − 1 , … , − r ; that is, we write x ~ i = ( x i , u i ) = ( X − i , U − i ) \widetilde{x}_i=(x_i,u_i)=(X_{-i},U_{-i}) x i = ( x i , u i ) = ( X − i , U − i ) . For I ⊆ { 1 , … , r } I\subseteq\{1,\ldots,r\} I ⊆ { 1 , … , r } , we construct a random graph G I ( ξ , x ~ r ) G_I(\xi,\widetilde{x}_r) G I ( ξ , x r ) as follows: The vertex set of G I ( ξ , x ~ r ) G_I(\xi,\widetilde{x}_r) G I ( ξ , x r ) is P ∪ { X − i : i ∈ I } \mathcal{P}\cup\{X_{-i}:i\in I\} P ∪ { X − i : i ∈ I } . For i , j ∈ N ∪ { − a : a ∈ I } i,j\in\mathbb{N}\cup\{-a:a\in I\} i , j ∈ N ∪ { − a : a ∈ I } with i < j i<j i < j , we add an edge between X i X_i X i and X j X_j X j when (7) holds. This construction ensures that if J ⊆ I J\subseteq I J ⊆ I , then the graph G I ( ξ , x ~ r ) ∖ { x i : i ∈ J } G_I(\xi,\widetilde{x}_r)\setminus\{x_i:i\in J\} G I ( ξ , x r ) ∖ { x i : i ∈ J } , where we remove the deterministic vertices indexed by J J J and all incident edges from G I ( ξ , x ~ r ) G_I(\xi,\widetilde{x}_r) G I ( ξ , x r ) , equals G I ∖ J ( ξ , x ~ r ) G_{I\setminus J}(\xi,\widetilde{x}_r) G I ∖ J ( ξ , x r ) . In particular, G I ( ξ , x ~ r ) ∖ { x i : i ∈ I } = G ∅ ( ξ ) = G ( ξ ) G_I(\xi,\widetilde{x}_r)\setminus\{x_i:i\in I\}=G_\varnothing(\xi)=G(\xi) G I ( ξ , x r ) ∖ { x i : i ∈ I } = G ∅ ( ξ ) = G ( ξ ) ; that is, we recover the RCM G ( ξ ) G(\xi) G ( ξ ) if we remove all the fixed vertices in x ~ r \widetilde{x}_r x r .
This coupling of RCMs with one or more fixed points added allows a notion of difference operators adapted to the RCM setting that is useful to establish asymptotic results [21 ] . Let H H H be a real-valued measurable functional on graphs with vertices in R d \mathbb{R}^d R d , and set F : = H ( G ( ξ ) ) F:=H(G(\xi)) F := H ( G ( ξ )) . Define the difference operators
Δ x ~ 1 , … , x ~ r ; x ~ r F = ∑ I ⊆ { 1 , … , r } ( − 1 ) r − ∣ I ∣ H ( G I ( ξ , x ~ r ) ) . \Delta_{\widetilde{x}_1,\ldots,\widetilde{x}_r;\widetilde{x}_r}F=\sum_{I\subseteq\{1,\ldots,r\}}(-1)^{r-|I|}H\left(G_I(\xi,\widetilde{x}_r)\right). Δ x 1 , … , x r ; x r F = I ⊆ { 1 , … , r } ∑ ( − 1 ) r − ∣ I ∣ H ( G I ( ξ , x r ) ) .
We also consider differences with respect to a subset of the inserted vertices while retaining the full ambient tuple x ~ r \widetilde{x}_r x r to specify their reserved labels. For i , j ∈ { 1 , … , r } i,j\in\{1,\ldots,r\} i , j ∈ { 1 , … , r } with i ≠ j i\ne j i = j , we thus write
Δ x ~ i ; x ~ r F = Δ x ~ i ; x ~ r 1 F = H ( G { i } ( ξ , x ~ r ) ) − H ( G ∅ ( ξ , x ~ r ) ) , \Delta_{\widetilde{x}_i;\widetilde{x}_r}F=\Delta_{\widetilde{x}_i;\widetilde{x}_r}^{1}F=H\left(G_{\{i\}}(\xi,\widetilde{x}_r)\right)-H\left(G_{\varnothing}(\xi,\widetilde{x}_r)\right), Δ x i ; x r F = Δ x i ; x r 1 F = H ( G { i } ( ξ , x r ) ) − H ( G ∅ ( ξ , x r ) ) ,
Δ x ~ i , x ~ j ; x ~ r F = H ( G { i , j } ( ξ , x ~ r ) ) − H ( G { i } ( ξ , x ~ r ) ) − H ( G { j } ( ξ , x ~ r ) ) + H ( G ∅ ( ξ , x ~ r ) ) . \Delta_{\widetilde{x}_i,\widetilde{x}_j;\widetilde{x}_r}F=H\left(G_{\{i,j\}}(\xi,\widetilde{x}_r)\right)-H\left(G_{\{i\}}(\xi,\widetilde{x}_r)\right)-H\left(G_{\{j\}}(\xi,\widetilde{x}_r)\right)+H\left(G_{\varnothing}(\xi,\widetilde{x}_r)\right). Δ x i , x j ; x r F = H ( G { i , j } ( ξ , x r ) ) − H ( G { i } ( ξ , x r ) ) − H ( G { j } ( ξ , x r ) ) + H ( G ∅ ( ξ , x r ) ) .
For W ⊆ R d W \subseteq\mathbb{R}^{d} W ⊆ R d , we let W ~ = W × [ 1 , ∞ ) \widetilde{W}=W\times[1,\infty) W = W × [ 1 , ∞ ) and define P W = P ∩ W \mathcal{P}_{W}=\mathcal{P}\cap W P W = P ∩ W , P ~ W = P ~ ∩ W ~ \widetilde{\mathcal{P}}_{W}=\widetilde{\mathcal{P}}\cap\widetilde{W} P W = P ∩ W , and ξ W = ( P ~ W , R ) \xi_{W}=(\widetilde{\mathcal{P}}_{W},R) ξ W = ( P W , R ) . Let G W , I ( ξ , x ~ r ) G_{W,I}(\xi,\widetilde{x}_{r}) G W , I ( ξ , x r ) denote the restriction of G I ( ξ , x ~ r ) G_{I}(\xi,\widetilde{x}_{r}) G I ( ξ , x r ) to the vertex set ( P ∪ { x i : i ∈ I } ) ∩ W (\mathcal{P}\cup\{x_{i}:i\in I\})\cap W ( P ∪ { x i : i ∈ I }) ∩ W . When W = W n W=W_{n} W = W n , we simply write P n \mathcal{P}_{n} P n , ξ n \xi_{n} ξ n , G n G_{n} G n , G n , I G_{n,I} G n , I etc. Note that G n , I ( ξ , x ~ r ) = G I ( ξ n , x ~ r ) G_{n,I}(\xi,\widetilde{x}_{r})=G_{I}(\xi_{n},\widetilde{x}_{r}) G n , I ( ξ , x r ) = G I ( ξ n , x r ) when x ~ r ∈ W ~ r \widetilde{x}_{r}\in\widetilde{W}_{r} x r ∈ W r . Often, we will consider a functional F F F that depends on G ( ξ ) G(\xi) G ( ξ ) only via G W ( ξ ) G_{W}(\xi) G W ( ξ ) . In this case, we may think of the restriction as part of the functional H H H .
Multivariate Mecke formula. Fix an ordered tuple x ~ r = ( x ~ 1 , … , x ~ r ) \widetilde{x}_{r}=(\widetilde{x}_{1},\ldots,\widetilde{x}_{r}) x r = ( x 1 , … , x r ) of marked vertices in R d × [ 1 , ∞ ) \mathbb{R}^{d}\times[1,\infty) R d × [ 1 , ∞ ) with distinct spatial coordinates. Write
G n , x ~ r : = G n , { 1 , … , r } ( ξ , x ~ r ) , G n , x ~ 0 : = G ( ξ n ) . G_{n,\widetilde{x}_{r}}:=G_{n,\{1,\ldots,r\}}(\xi,\widetilde{x}_{r}),\qquad G_{n,\widetilde{x}_{0}}:=G(\xi_{n}). G n , x r := G n , { 1 , … , r } ( ξ , x r ) , G n , x 0 := G ( ξ n ) .
For any nonnegative measurable graph functional h ( y ~ k ; G ) h(\widetilde{y}_{k};G) h ( y k ; G ) , invariant under auxiliary relabeling,
E [ ∑ y ~ k ∈ P ~ n ( k ) h ( y ~ k ; G n , x ~ r ) ] = ∫ W ~ n k E [ h ( y ~ k ; G n , x ~ r , y ~ k ) ] ∏ j = 1 k λ ( d y ~ j ) . (10) \mathbb{E}\left[\sum_{\widetilde{y}_{k}\in\widetilde{\mathcal{P}}_{n}^{(k)}}h(\widetilde{y}_{k};G_{n,\widetilde{x}_{r}})\right]
=\int_{\widetilde{W}_{n}^{k}}\mathbb{E}\left[h(\widetilde{y}_{k};G_{n,\widetilde{x}_{r},\widetilde{y}_{k}})\right]\prod_{j=1}^{k}\lambda(\mathrm{d}\widetilde{y}_{j}). \tag*{(10)} E y k ∈ P n ( k ) ∑ h ( y k ; G n , x r ) = ∫ W n k E [ h ( y k ; G n , x r , y k ) ] j = 1 ∏ k λ ( d y j ) . (10)
Here G n , x ~ r , y ~ k G_{n,\widetilde{x}_{r},\widetilde{y}_{k}} G n , x r , y k denotes the graph obtained by inserting the r + k r+k r + k marked vertices in the concatenated tuple ( x ~ r , y ~ k ) (\widetilde{x}_{r},\widetilde{y}_{k}) ( x r , y k ) into G ( ξ n ) G(\xi_{n}) G ( ξ n ) (the order of the concatenated tuple specifies the labels, but does not affect the conditional distribution of the resulting graph). The sum is over ordered k k k -tuples of distinct Poisson vertices, while the inserted vertices x ~ r \widetilde{x}_{r} x r remain fixed. The case r = 0 r=0 r = 0 corresponds to the original RCM.
To justify the identity, first condition on the marked vertex configuration and average the edge randomness. The resulting function depends only on the marked vertices, so the ordinary multivariate Mecke formula [23 Theorem 4.4] applies. Conditional on the augmented marked configuration, the graph has independent edge distributions, giving the identity.
A result from the literature
Let W ⊆ R d W\subseteq\mathbb{R}^{d} W ⊆ R d be a bounded Borel set. If the origin o o o belongs to W W W and u ∈ [ 1 , ∞ ) u\in[1,\infty) u ∈ [ 1 , ∞ ) , then G W , { 1 } ( ξ , ( ( o , u ) ) ) G_{W,\{1\}}(\xi,((o,u))) G W , { 1 } ( ξ , (( o , u ))) is the weighted RCM restricted to W W W with the additional vertex ( o , u ) (o,u) ( o , u ) and its incident edges. If o ∉ W o\notin W o ∈ / W , the inserted vertex is removed by the spatial restriction, so the augmented and unaugmented restricted graphs coincide. For a polynomial f f f , define the add-one cost
Δ W , f ( u ) : = Δ ( o , u ) ; ( ( o , u ) ) 1 L ( f ; G W ( ξ ) ) = L ( f ; G W , { 1 } ( ξ , ( ( o , u ) ) ) ) − L ( f ; G W , ∅ ( ξ , ( ( o , u ) ) ) ) . (11) \Delta_{W,f}(u):=\Delta^{1}_{(o,u);((o,u))}L(f;G_{W}(\xi))=L(f;G_{W,\{1\}}(\xi,((o,u))))-L(f;G_{W,\varnothing}(\xi,((o,u)))). \tag*{(11)} Δ W , f ( u ) := Δ ( o , u ) ; (( o , u )) 1 L ( f ; G W ( ξ )) = L ( f ; G W , { 1 } ( ξ , (( o , u )))) − L ( f ; G W , ∅ ( ξ , (( o , u )))) . (11)
Let W \mathcal{W} W denote the collection of cubes of the form ∏ i = 1 d [ x i , x i + l ) \prod_{i=1}^{d}[x_{i},x_{i}+l) ∏ i = 1 d [ x i , x i + l ) , where x 1 , … , x d ∈ R x_{1},\ldots,x_{d}\in\mathbb{R} x 1 , … , x d ∈ R and l > 0 l>0 l > 0 .
The main ingredient in the proof of Theorem 1 is the following result. It is essentially the variance convergence part of the weak stabilization CLT [8 Thm. 1.1] in the unweighted setting, while the weighted case is covered by [25 Lem. 2.2] . We state it here in a version adapted to the functional L ( f ; G ) L(f;G) L ( f ; G ) .
Theorem 4 (Variance asymptotics from [8 , 25 ] ). Let f f f be a polynomial test function. Suppose that
(i) The limit Δ f ( u ) : = lim n → ∞ Δ W n , f ( u ) \Delta_{f}(u):=\lim_{n\to\infty}\Delta_{W_{n},f}(u) Δ f ( u ) := lim n → ∞ Δ W n , f ( u ) exists almost surely for any sequence of cubes W n ∈ W W^{n}\in\mathcal{W} W n ∈ W with W n ↑ R d W^{n}\uparrow\mathbb{R}^{d} W n ↑ R d and any mark u u u .
(ii) For some p > 2 p>2 p > 2 , sup W ∈ W E [ ∣ Δ W , f ( U ) ∣ p ] < ∞ \sup_{W\in\mathcal{W}}\mathbb{E}[|\Delta_{W,f}(U)|^{p}]<\infty sup W ∈ W E [ ∣ Δ W , f ( U ) ∣ p ] < ∞ , where U U U has law μ \mu μ and is independent of the original marked Poisson process and the edge randomness.
If (i) and (ii) hold, then the limit
σ f 2 : = lim n → ∞ ∣ W n ∣ − 1 Var ( L ( f ; G n ( ξ ) ) ) \sigma_{f}^{2}:=\lim_{n\to\infty}|W_{n}|^{-1}\operatorname{Var}(L(f;G_{n}(\xi))) σ f 2 := n → ∞ lim ∣ W n ∣ − 1 Var ( L ( f ; G n ( ξ ))) exists, and σ f 2 > 0 \sigma_{f}^{2}>0 σ f 2 > 0 whenever P ( Δ f ( U ) ≠ 0 ) > 0 \mathbb{P}(\Delta_{f}(U)\ne0)>0 P ( Δ f ( U ) = 0 ) > 0 .
Proof. The polynomial linear eigenvalue statistic L ( f ; G ) L(f;G) L ( f ; G ) is invariant under translations of the vertex set of G G G . Hence, the variance asymptotics part of [8 Thm. 1.1] gives the result in the unweighted case. The weighted case follows from [25 ] , where the RCM is the special case α = 1 \alpha=1 α = 1 of the random simplicial complexes introduced there. To see how the result follows, note that G { 1 } ( ξ , ( o , u ) ) G_{\{1\}}(\xi,(o,u)) G { 1 } ( ξ , ( o , u )) has the same distribution as the random graph called Δ ( o , u , 1 , V ) \Delta^{(o,u,1,V)} Δ ( o , u , 1 , V ) in [25 ] . Here V V V denotes the independent random auxiliary mark used to sample the edges. It follows that the family of random variables { Λ ( o , u , 1 , V ) L ( f , Δ W ) } W ∈ W \{\Lambda_{(o,u,1,V)}L(f,\Delta_{W})\}_{W\in\mathcal{W}} { Λ ( o , u , 1 , V ) L ( f , Δ W ) } W ∈ W in [25 ] have the same joint distribution as our { Δ W , f ( u ) } W ∈ W \{\Delta_{W,f}(u)\}_{W\in\mathcal{W}} { Δ W , f ( u ) } W ∈ W and thus Theorem 4 follows directly from [25 Lem. 2.2] .
Proof of Theorem 1
We show Theorem 1 by showing the following proposition.
Proposition 5 (Variance asymptotics and positivity). Let f ( x ) = ∑ q = 0 m a q x q f(x)=\sum_{q=0}^{m}a_qx^q f ( x ) = ∑ q = 0 m a q x q be a polynomial of degree m ≥ 1 m\ge1 m ≥ 1 . Assume that (PK) holds and that E [ U 3 m − 1 ] < ∞ \mathbb{E}[U^{3m-1}]<\infty E [ U 3 m − 1 ] < ∞ . Then, the limiting add-one cost Δ f ( u ) \Delta_f(u) Δ f ( u ) exists almost surely for every u ≥ 1 u\ge1 u ≥ 1 , and the limit lim n → ∞ ∣ W n ∣ − 1 Var ( Tr [ f ( A n ) ] ) \lim_{n\to\infty}|W_n|^{-1}\operatorname{Var}(\operatorname{Tr}[f(A_n)]) lim n → ∞ ∣ W n ∣ − 1 Var ( Tr [ f ( A n )]) exists.
If P ( Δ f ( U ) ≠ 0 ) > 0 \mathbb{P}(\Delta_f(U)\ne0)>0 P ( Δ f ( U ) = 0 ) > 0 , where U U U has law μ \mu μ , then σ f 2 > 0 \sigma_f^2>0 σ f 2 > 0 . If M φ > 0 M_\varphi>0 M φ > 0 and either a 0 ≠ 0 a_0\ne0 a 0 = 0 or all coefficients of f f f are non-negative or all are non-positive, and f f f is not of the form f ( x ) = a 1 x f(x)=a_1x f ( x ) = a 1 x , then σ f 2 > 0 \sigma_f^2>0 σ f 2 > 0 .
Proof of Theorem 1. The first conclusion of Proposition 5 gives the existence and finiteness of the limit asserted in Theorem 1 under exactly the same hypotheses. □ \square □
The proof of Proposition 5 represents polynomial spectral statistics by closed-walk counts as in § 3.1 . Their moments are controlled using the following tree estimate, which relies on an idea of [17 Theorem 2.1] .
Lemma 6 (Tree integral estimate). Let γ ∈ ( 0 , 1 ] \gamma\in(0,1] γ ∈ ( 0 , 1 ] , assume that κ \kappa κ satisfies (PK), and suppose that M φ , γ < ∞ M_{\varphi,\gamma}<\infty M φ , γ < ∞ . Let T T T be a tree with vertex set { 0 , 1 , … , k } \{0,1,\ldots,k\} { 0 , 1 , … , k } , and let d i d_i d i denote the degree of vertex i i i in T T T . If u 0 , … , u k ∈ [ 1 , ∞ ) u_0,\ldots,u_k\in[1,\infty) u 0 , … , u k ∈ [ 1 , ∞ ) and x 0 ∈ R d x_0\in\mathbb{R}^d x 0 ∈ R d , then
∫ ( R d ) k ∏ { i , j } ∈ E ( T ) φ ( ( ∥ x i − x j ∥ d κ ( u i , u j ) ) γ ) d x 1 ⋯ d x k ≤ M φ , γ k κ γ k ∏ i = 0 k u i d i . \int_{(\mathbb{R}^d)^k}\prod_{\{i,j\}\in E(T)}\varphi\left(\left(\frac{\lVert x_i-x_j\rVert^d}{\kappa(u_i,u_j)}\right)^\gamma\right)\,\mathrm{d}x_1\cdots\mathrm{d}x_k\le M_{\varphi,\gamma}^k\kappa_\gamma^k\prod_{i=0}^{k}u_i^{d_i}. ∫ ( R d ) k { i , j } ∈ E ( T ) ∏ φ ( ( κ ( u i , u j ) ∥ x i − x j ∥ d ) γ ) d x 1 ⋯ d x k ≤ M φ , γ k κ γ k i = 0 ∏ k u i d i .
Proof of Proposition 5. For the first statement, we need to show (i) and (ii) in Theorem 4 . If the origin o o o is not contained in W W W , then Δ W , f ( u ) \Delta_{W,f}(u) Δ W , f ( u ) is zero. It is therefore enough to prove the uniform moment bounds for cubes containing the origin. For such cubes, linearity gives
Δ W , f ( u ) = a 0 + a 2 Δ W , 2 ( u ) + ⋯ + a m Δ W , m ( u ) (12) \Delta_{W,f}(u)=a_0+a_2\Delta_{W,2}(u)+\cdots+a_m\Delta_{W,m}(u) \tag*{(12)} Δ W , f ( u ) = a 0 + a 2 Δ W , 2 ( u ) + ⋯ + a m Δ W , m ( u ) (12)
where Δ W , q ( u ) : = Δ W , x q ( u ) \Delta_{W,q}(u):=\Delta_{W,x^q}(u) Δ W , q ( u ) := Δ W , x q ( u ) . Here we used that Δ W , 0 ( u ) = 1 \Delta_{W,0}(u)=1 Δ W , 0 ( u ) = 1 , and Δ W , 1 ( u ) = 0 \Delta_{W,1}(u)=0 Δ W , 1 ( u ) = 0 because the graph has no loops. It follows that it is enough to show (i) and (ii) in Theorem 4 for f ( x ) = x q f(x)=x^q f ( x ) = x q where 2 ≤ q ≤ m 2\le q\le m 2 ≤ q ≤ m .
First consider (i). The loop counting interpretation of the functional L ( x q , ⋅ ) L(x^q,\cdot) L ( x q , ⋅ ) from § 3.1 implies that Δ W , q ( u ) \Delta_{W,q}(u) Δ W , q ( u ) counts the number of closed walks of length q q q in G W , { 1 } ( ξ , ( o , u ) ) G_{W,\{1\}}(\xi,(o,u)) G W , { 1 } ( ξ , ( o , u )) that pass through o o o at least once. By multivariate Mecke formula, the expected number of rooted closed walks of length q q q starting and ending at o o o is
E [ ∑ p = 2 q ∑ ( ( X 2 , U 2 ) , … , ( X p , U p ) ) ∈ ( P ∩ W ˘ ) ( p − 1 ) 1 ( p − 1 ) ! ∑ π ∈ Σ q , p π ( 1 ) = 1 ∏ j = 1 q A ( X π ( j ) , X π ( j + 1 ) ; G W , { 1 } ( ξ , ( o , u ) ) ) ] = ∑ p = 2 q 1 ( p − 1 ) ! ∑ π ∈ Σ q , p π ( 1 ) = 1 ∫ W p − 1 E [ ∏ { i , j } ∈ E ( π ) φ ( ( ∥ x i − x j ∥ d κ ( U i , U j ) ) ) ] d x 2 ⋯ d x p . (13) \begin{aligned}
\mathbb{E}\left[\sum_{p=2}^{q}\ \sum_{((X_2,U_2),\ldots,(X_p,U_p))\in(\mathcal{P}\cap\breve W)^{(p-1)}}\frac{1}{(p-1)!}\sum_{\substack{\pi\in\Sigma_{q,p}\\\pi(1)=1}}\prod_{j=1}^{q}A(X_{\pi(j)},X_{\pi(j+1)};G_{W,\{1\}}(\xi,(o,u)))\right] \\
&=\sum_{p=2}^{q}\frac{1}{(p-1)!}\sum_{\substack{\pi\in\Sigma_{q,p}\\\pi(1)=1}}\int_{W^{p-1}}\mathbb{E}\left[\prod_{\{i,j\}\in E(\pi)}\varphi\left(\left(\frac{\lVert x_i-x_j\rVert^d}{\kappa(U_i,U_j)}\right)\right)\right]\mathrm{d}x_2\cdots\mathrm{d}x_p. \tag*{(13)}
\end{aligned} E p = 2 ∑ q (( X 2 , U 2 ) , … , ( X p , U p )) ∈ ( P ∩ W ˘ ) ( p − 1 ) ∑ ( p − 1 )! 1 π ∈ Σ q , p π ( 1 ) = 1 ∑ j = 1 ∏ q A ( X π ( j ) , X π ( j + 1 ) ; G W , { 1 } ( ξ , ( o , u ))) = p = 2 ∑ q ( p − 1 )! 1 π ∈ Σ q , p π ( 1 ) = 1 ∑ ∫ W p − 1 E { i , j } ∈ E ( π ) ∏ φ ( ( κ ( U i , U j ) ∥ x i − x j ∥ d ) ) d x 2 ⋯ d x p . (13)
where X 1 = x 1 = o X_1=x_1=o X 1 = x 1 = o , U 1 = u U_1=u U 1 = u , and E ( π ) E(\pi) E ( π ) is the set of distinct undirected edges traversed by the closed walk. The graph with vertex set { 1 , … , p } \{1,\ldots,p\} { 1 , … , p } and edge set E ( π ) E(\pi) E ( π ) is connected. Choose a spanning tree T T T of this graph with edge set E ( T ) ⊆ E ( π ) E(T)\subseteq E(\pi) E ( T ) ⊆ E ( π ) . Since every connection probability lies in [ 0 , 1 ] [0,1] [ 0 , 1 ] ,
∫ W p − 1 ∏ { i , j } ∈ E ( π ) φ ( ( ∥ x i − x j ∥ d κ ( U i , U j ) ) ) d x 2 ⋯ d x p ≤ ∫ W p − 1 ∏ { i , j } ∈ E ( T ) φ ( ( ∥ x i − x j ∥ d κ ( U i , U j ) ) ) d x 2 ⋯ d x p . (14) \int_{W^{p-1}}\prod_{\{i,j\}\in E(\pi)}\varphi\left(\left(\frac{\lVert x_i-x_j\rVert^d}{\kappa(U_i,U_j)}\right)\right)\mathrm{d}x_2\cdots\mathrm{d}x_p\le\int_{W^{p-1}}\prod_{\{i,j\}\in E(T)}\varphi\left(\left(\frac{\lVert x_i-x_j\rVert^d}{\kappa(U_i,U_j)}\right)\right)\mathrm{d}x_2\cdots\mathrm{d}x_p. \tag*{(14)} ∫ W p − 1 { i , j } ∈ E ( π ) ∏ φ ( ( κ ( U i , U j ) ∥ x i − x j ∥ d ) ) d x 2 ⋯ d x p ≤ ∫ W p − 1 { i , j } ∈ E ( T ) ∏ φ ( ( κ ( U i , U j ) ∥ x i − x j ∥ d ) ) d x 2 ⋯ d x p . (14)
By Lemma 6 , the expectation of (14) is bounded by u p − 1 M φ p − 1 κ φ p − 1 ∏ i = 2 p E [ U d i ] u^{p-1}M_{\varphi}^{p-1}\kappa_{\varphi}^{p-1}\prod_{i=2}^{p}\mathbb{E}[U^{d_i}] u p − 1 M φ p − 1 κ φ p − 1 ∏ i = 2 p E [ U d i ] , where d i ≤ p − 1 d_i\le p-1 d i ≤ p − 1 . Let C q , W ( u ) C_{q,W}(u) C q , W ( u ) denote the number of length-q q q closed walks based at ( o , u ) (o,u) ( o , u ) in the common augmented graph restricted to W W W , where o ∈ W o\in W o ∈ W . Its expectation is the expression in (13) . The preceding bound is independent of W W W , and the number of terms in (13) depends only on q q q . Since all required mark moments have order at most q − 1 ≤ 3 m − 1 q-1\le3m-1 q − 1 ≤ 3 m − 1 , we have sup n E C q , W n ( u ) ≤ C q u q − 1 < ∞ \sup_n\mathbb{E}C_{q,W_n}(u)\le C_q u^{q-1}<\infty sup n E C q , W n ( u ) ≤ C q u q − 1 < ∞ for each u ≥ 1 u\ge1 u ≥ 1 .
Let C q , ∞ ( u ) C_{q,\infty}(u) C q , ∞ ( u ) count the corresponding rooted walks in the infinite augmented graph. All finite-window graphs are restrictions of this same graph, and every finite walk is contained in a sufficiently large window. Consequently, C q , W n ( u ) ↑ C q , ∞ ( u ) C_{q,W_n}(u) \uparrow C_{q,\infty}(u) C q , W n ( u ) ↑ C q , ∞ ( u ) . By monotone convergence,
E C q , ∞ ( u ) = lim n → ∞ E C q , W n ( u ) ≤ C q u q − 1 < ∞ . \mathbb{E}C_{q,\infty}(u)=\lim_{n\to\infty}\mathbb{E}C_{q,W_n}(u)\le C_q u^{q-1}<\infty. E C q , ∞ ( u ) = n → ∞ lim E C q , W n ( u ) ≤ C q u q − 1 < ∞.
Thus C q , ∞ ( u ) < ∞ C_{q,\infty}(u)<\infty C q , ∞ ( u ) < ∞ almost surely. The union of the vertices visited by these finitely many rooted walks is finite. Every based closed walk of length q q q visiting the inserted vertex is a cyclic shift of a walk based there, so it visits vertices in the same finite union. Therefore, this union contains every vertex that can contribute to the add-one cost and is contained in some random window W N W_N W N . It follows that Δ W , q ( u ) = Δ W N , q ( u ) \Delta_{W,q}(u)=\Delta_{W_N,q}(u) Δ W , q ( u ) = Δ W N , q ( u ) whenever W N ⊆ W W_N\subseteq W W N ⊆ W . Hence,
Δ q ( u ) : = lim W n ↑ R d Δ W n , q ( u ) = Δ W N , q ( u ) \Delta_q(u):=\lim_{W_n\uparrow\mathbb{R}^d}\Delta_{W_n,q}(u)=\Delta_{W_N,q}(u) Δ q ( u ) := W n ↑ R d lim Δ W n , q ( u ) = Δ W N , q ( u )
exists almost surely, which shows (i).
For the moment condition (ii), we show that sup W ∈ W E [ ∣ Δ W , q ( U ) ∣ 3 ] < ∞ \sup_{W\in\mathcal{W}}\mathbb{E}[|\Delta_{W,q}(U)|^3]<\infty sup W ∈ W E [ ∣ Δ W , q ( U ) ∣ 3 ] < ∞ . We have
E [ ∣ Δ W , q ( U ) ∣ 3 ] ≤ E [ q 3 ( ∑ p = 2 q ∑ ( ( X 2 , U 2 ) , … , ( X p , U p ) ) ∈ ( P ∩ W ˘ ) ( p − 1 ) 1 ( p − 1 ) ! ∑ π ∈ S q , p ∘ π ( 1 ) = 1 ∏ j = 1 q A ( X π ( j ) , X π ( j + 1 ) ; G W , { 1 } ( ξ , ( o , U ) ) ) ) 3 ] , (15) \mathbb{E}[|\Delta_{W,q}(U)|^3]
\le\mathbb{E}\left[q^3\left(\sum_{p=2}^{q}\sum_{((X_2,U_2),\ldots,(X_p,U_p))\in(\mathcal{P}\cap\breve{W})^{(p-1)}}\frac{1}{(p-1)!}\sum_{\pi\in\mathfrak{S}_{q,p}^{\circ}\atop\pi(1)=1}\prod_{j=1}^{q}A(X_{\pi(j)},X_{\pi(j+1)};G_{W,\{1\}}(\xi,(o,U)))\right)^3\right], \tag*{(15)} E [ ∣ Δ W , q ( U ) ∣ 3 ] ≤ E q 3 p = 2 ∑ q (( X 2 , U 2 ) , … , ( X p , U p )) ∈ ( P ∩ W ˘ ) ( p − 1 ) ∑ ( p − 1 )! 1 π ( 1 ) = 1 π ∈ S q , p ∘ ∑ j = 1 ∏ q A ( X π ( j ) , X π ( j + 1 ) ; G W , { 1 } ( ξ , ( o , U ))) 3 , (15)
where the factor q 3 q^3 q 3 accounts for the at most q q q possible starting points of the same closed walk. Expanding (15) , we obtain a sum in which each term corresponds to three rooted closed walks of length q q q , each starting and ending at o o o . Concatenating them at o o o produces a closed walk of length 3 q 3q 3 q that starts and ends at o o o . Thus, (15) is bounded by q 3 q^3 q 3 times the expected total number of rooted closed walks of length 3 q 3q 3 q starting and ending at o o o . For fixed u u u , this was bounded by (13) with q q q replaced by 3 q 3q 3 q , and each integral was again bounded. After concatenation, the resulting walk has length 3 q 3q 3 q , so the weighted tree estimate involves only moments of order at most 3 q − 1 ≤ 3 m − 1 3q-1\le3m-1 3 q − 1 ≤ 3 m − 1 . Recalling the condition E [ U 3 m − 1 ] < ∞ \mathbb{E}[U^{3m-1}]<\infty E [ U 3 m − 1 ] < ∞ , taking expectation with respect to U U U in (13) as above therefore yields a finite bound independent of W W W . This shows (ii).
It remains to verify the conditions for variance positivity. Clearly, Δ 0 = 1 \Delta_0=1 Δ 0 = 1 and Δ 1 = 0 \Delta_1=0 Δ 1 = 0 . The probability that there are no edges incident to ( o , u ) (o,u) ( o , u ) , for fixed u u u , is
E [ ∏ ( X i , U i ) ∈ P ˘ ( 1 − A ( o , X i ) ) ] = E [ ∏ ( X i , U i ) ∈ P ˘ ( 1 − φ ( ∥ X i ∥ d κ ( u , U i ) ) ) ] = exp ( − ∫ R d E [ φ ( ∥ x ∥ d κ ( u , U ) ) ] d x ) ≥ exp ( − c κ M φ u E [ U ] ) \begin{aligned}
\mathbb{E}\left[\prod_{(X_i,U_i)\in\breve{\mathcal{P}}}(1-A(o,X_i))\right]
=\mathbb{E}\left[\prod_{(X_i,U_i)\in\breve{\mathcal{P}}}\left(1-\varphi\left(\frac{\lVert X_i\rVert^d}{\kappa(u,U_i)}\right)\right)\right] \\
&=\exp\left(-\int_{\mathbb{R}^d}\mathbb{E}\left[\varphi\left(\frac{\lVert x\rVert^d}{\kappa(u,U)}\right)\right]\mathrm{d}x\right)
\ge\exp(-c_\kappa M_\varphi u\mathbb{E}[U])
\end{aligned} E ( X i , U i ) ∈ P ˘ ∏ ( 1 − A ( o , X i )) = E ( X i , U i ) ∈ P ˘ ∏ ( 1 − φ ( κ ( u , U i ) ∥ X i ∥ d ) ) = exp ( − ∫ R d E [ φ ( κ ( u , U ) ∥ x ∥ d ) ] d x ) ≥ exp ( − c κ M φ u E [ U ])
by Lemma 6 . Since M φ < ∞ M_\varphi<\infty M φ < ∞ by the assumption (1) , there is a positive probability that no edges are incident to ( o , u ) (o,u) ( o , u ) , in which case Δ q ( u ) = 0 \Delta_q(u)=0 Δ q ( u ) = 0 for all q > 0 q>0 q > 0 , and hence P ( Δ f ( U ) = a 0 ) > 0 \mathbb{P}(\Delta_f(U)=a_0)>0 P ( Δ f ( U ) = a 0 ) > 0 . If a 0 ≠ 0 a_0\ne0 a 0 = 0 , this implies σ f 2 > 0 \sigma_f^2>0 σ f 2 > 0 .
For the second case, note that Δ q ( u ) ≥ 0 \Delta_q(u)\ge0 Δ q ( u ) ≥ 0 for every q ≥ 2 q\ge2 q ≥ 2 , because adding a vertex at ( o , u ) (o,u) ( o , u ) together with its incident edges does not destroy existing closed walks but may add extra closed walks that pass through o o o . A walk of length 3 starting and ending at o o o must pass through exactly two other points; hence the expected number of such paths is
E [ ∑ ( ( X 1 , U 1 ) , ( X 2 , U 2 ) ) ∈ P ˘ ( 2 ) A ( o , X 1 ) A ( X 1 , X 2 ) A ( X 2 , o ) ] ≥ ∫ R d ∫ R d φ ( ∥ x 1 ∥ d ) φ ( ∥ x 1 − x 2 ∥ d ) φ ( ∥ x 2 ∥ d ) d x 1 d x 2 , \mathbb{E}\left[\sum_{((X_1,U_1),(X_2,U_2))\in\breve{\mathcal{P}}^{(2)}}A(o,X_1)A(X_1,X_2)A(X_2,o)\right]
\ge\int_{\mathbb{R}^d}\int_{\mathbb{R}^d}\varphi(\lVert x_1\rVert^d)\varphi(\lVert x_1-x_2\rVert^d)\varphi(\lVert x_2\rVert^d)\,\mathrm{d}x_1\mathrm{d}x_2, E (( X 1 , U 1 ) , ( X 2 , U 2 )) ∈ P ˘ ( 2 ) ∑ A ( o , X 1 ) A ( X 1 , X 2 ) A ( X 2 , o ) ≥ ∫ R d ∫ R d φ (∥ x 1 ∥ d ) φ (∥ x 1 − x 2 ∥ d ) φ (∥ x 2 ∥ d ) d x 1 d x 2 ,
where we used that φ \varphi φ is nonincreasing and κ ( u , v ) ≥ 1 \kappa(u,v)\ge1 κ ( u , v ) ≥ 1 . Since φ \varphi φ is nonincreasing and M φ > 0 M_\varphi>0 M φ > 0 , φ \varphi φ must be positive on a ball of (possibly infinite) radius ρ > 0 \rho>0 ρ > 0 . It follows that the integrand above must be positive on B ρ / 2 ( o ) × B ρ / 2 ( o ) B_{\rho/2}(o)\times B_{\rho/2}(o) B ρ /2 ( o ) × B ρ /2 ( o ) and hence the integral is positive. Thus, P ( Δ 3 ( U ) > 0 ) > 0 \mathbb{P}(\Delta_3(U)>0)>0 P ( Δ 3 ( U ) > 0 ) > 0 . Going back and forth along an edge incident to o o o creates a closed walk of length 2. For any q > 3 q>3 q > 3 , a closed walk of length q q q can be constructed by concatenating closed walks of lengths 2 and 3. Therefore, Δ 3 ( u ) > 0 \Delta_3(u)>0 Δ 3 ( u ) > 0 implies Δ q ( u ) > 0 \Delta_q(u)>0 Δ q ( u ) > 0 for all q > 1 q>1 q > 1 . It follows from (12) that if all coefficients are non-negative (resp. non-positive) and f f f is not of the form a 1 x a_1x a 1 x , then Δ f ( u ) \Delta_f(u) Δ f ( u ) is also positive (resp. negative) on the event Δ 3 ( u ) > 0 \Delta_3(u)>0 Δ 3 ( u ) > 0 .
Proof of Lemma 6. We integrate the spatial variables successively, beginning with a leaf of T \mathcal{T} T other than the fixed vertex 0. If j j j is a leaf with parent i i i , then the change of variables z = κ ( u i , u j ) − 1 / d ( x j − x i ) z=\kappa(u_i,u_j)^{-1/d}(x_j-x_i) z = κ ( u i , u j ) − 1/ d ( x j − x i ) gives
∫ R d φ ( ( ∥ x j − x i ∥ κ ( u i , u j ) ) d ) d x j = κ ( u i , u j ) M φ , γ . \int_{\mathbb{R}^d}\varphi\left(\left(\frac{\lVert x_j-x_i\rVert}{\kappa(u_i,u_j)}\right)^d\right)\,dx_j=\kappa(u_i,u_j)M_{\varphi,\gamma}. ∫ R d φ ( ( κ ( u i , u j ) ∥ x j − x i ∥ ) d ) d x j = κ ( u i , u j ) M φ , γ .
Iterating over the tree yields M φ , γ k ∏ { i , j } ∈ E ( T ) κ ( u i , u j ) M_{\varphi,\gamma}^k\prod_{\{i,j\}\in E(\mathcal{T})}\kappa(u_i,u_j) M φ , γ k ∏ { i , j } ∈ E ( T ) κ ( u i , u j ) . Since d i d_i d i is the degree of vertex i i i , Condition (PK) gives ∏ { i , j } ∈ E ( T ) κ ( u i , u j ) ≤ c k ∏ i = 0 k u i d i \prod_{\{i,j\}\in E(\mathcal{T})}\kappa(u_i,u_j)\le c^k\prod_{i=0}^k u_i^{d_i} ∏ { i , j } ∈ E ( T ) κ ( u i , u j ) ≤ c k ∏ i = 0 k u i d i , which proves the claim. □
Proof for polynomial test functions
In this section, we prove the CLT for polynomial test functions stated in Theorem 2 . For functionals of the Poisson process alone, optimal convergence rates are typically derived from the general results of [22 ] stated in terms of finite difference operators when inserting one or more fixed points. However, when working with functionals of an RCM, one needs to keep track of how edges are added to the extra points. We therefore use normal approximation bounds for edge-marked Poisson processes developed in [21 ] which are analogues of [22 ] but with the point-wise difference operators replaced by difference operators for edge marked Poisson processes introduced in § 3.2 .
Section 4.1 records the Poincaré inequality, Wasserstein, and Kolmogorov bounds for normal approximations [21 ] . Section 4.2 then proves the required moment and spatial integrability estimates for the first- and second-order insertion differences and combines them with the variance asymptotics from § 3 to obtain both the quantitative normal approximation and the asserted central limit theorem.
The CLT from nestmann
Recall the construction of the RCM and the difference operators from § 3.2 . For a functional F n = H ( G ( ξ n ) ) F_n=H(G(\xi_n)) F n = H ( G ( ξ n )) , we have the following Poincaré inequality from [21 ] .
Theorem 7 (Thm. 5.2 in [21 ] ). Suppose E F n 2 < ∞ \mathbb{E}F_n^2<\infty E F n 2 < ∞ . Then , Var F n ≤ ∫ W ~ n E [ ( Δ x ~ 1 ; x ~ 1 F n ) 2 ] λ ( d x ~ 1 ) \operatorname{Var}F_n\le\int_{\widetilde{W}_n}\mathbb{E}[(\Delta_{\widetilde{x}_1;\widetilde{x}_1}F_n)^2]\lambda(d\widetilde{x}_1) Var F n ≤ ∫ W n E [( Δ x 1 ; x 1 F n ) 2 ] λ ( d x 1 ) .
Moreover, we have the following CLT:
Theorem 8 (Thm. 6.1 in [21 ] ). Suppose E F n = 0 \mathbb{E}F_n=0 E F n = 0 , Var F n = 1 \operatorname{Var}F_n=1 Var F n = 1 and E F n 4 < ∞ \mathbb{E}F_n^4<\infty E F n 4 < ∞ . Let γ 1 , … , γ 6 \gamma_1,\ldots,\gamma_6 γ 1 , … , γ 6 be as in (16) below. If γ 5 , γ 6 < ∞ \gamma_5,\gamma_6<\infty γ 5 , γ 6 < ∞ and the condition (17) below holds, then
d W ( F n , N ( 0 , 1 ) ) ≤ γ 1 + γ 2 + γ 3 d_W(F_n,\mathcal{N}(0,1))\le\gamma_1+\gamma_2+\gamma_3 d W ( F n , N ( 0 , 1 )) ≤ γ 1 + γ 2 + γ 3 and
d K ( F n , N ( 0 , 1 ) ) ≤ γ 1 + γ 2 + γ 3 + γ 4 + γ 5 + γ 6 . d_K(F_n,\mathcal{N}(0,1))\le\gamma_1+\gamma_2+\gamma_3+\gamma_4+\gamma_5+\gamma_6. d K ( F n , N ( 0 , 1 )) ≤ γ 1 + γ 2 + γ 3 + γ 4 + γ 5 + γ 6 .
The quantities γ i \gamma_i γ i have the usual form from [22 ] , but they must be interpreted through the difference operators for edge-marked Poisson processes. Precisely,
γ 1 = 2 ( ∫ W ~ n 3 E [ ( Δ x ~ 1 , x ~ 3 ; x ~ 3 F n ) 2 ( Δ x ~ 2 , x ~ 3 ; x ~ 3 F n ) 2 ] 1 / 2 E [ ( Δ x ~ 1 , x ~ 3 ; x ~ 3 2 F n ) 2 ( Δ x ~ 2 , x ~ 3 ; x ~ 3 2 F n ) 2 ] 1 / 2 λ 3 ( d ( x ~ 1 , x ~ 2 , x ~ 3 ) ) ) 1 / 2 \gamma_1=2\left(\int_{\widetilde{W}_n^3}\mathbb{E}[(\Delta_{\widetilde{x}_1,\widetilde{x}_3;\widetilde{x}_3}F_n)^2(\Delta_{\widetilde{x}_2,\widetilde{x}_3;\widetilde{x}_3}F_n)^2]^{1/2}\mathbb{E}[(\Delta^2_{\widetilde{x}_1,\widetilde{x}_3;\widetilde{x}_3}F_n)^2(\Delta^2_{\widetilde{x}_2,\widetilde{x}_3;\widetilde{x}_3}F_n)^2]^{1/2}\lambda^3(d(\widetilde{x}_1,\widetilde{x}_2,\widetilde{x}_3))\right)^{1/2} γ 1 = 2 ( ∫ W n 3 E [( Δ x 1 , x 3 ; x 3 F n ) 2 ( Δ x 2 , x 3 ; x 3 F n ) 2 ] 1/2 E [( Δ x 1 , x 3 ; x 3 2 F n ) 2 ( Δ x 2 , x 3 ; x 3 2 F n ) 2 ] 1/2 λ 3 ( d ( x 1 , x 2 , x 3 )) ) 1/2
γ 2 = ( ∫ W ~ n 3 E [ ( Δ x ~ 1 , x ~ 3 ; x ~ 3 2 F n ) 2 ( Δ x ~ 2 , x ~ 3 ; x ~ 3 2 F n ) 2 ] λ 3 ( d ( x ~ 1 , x ~ 2 , x ~ 3 ) ) ) 1 / 2 \gamma_2=\left(\int_{\widetilde{W}_n^3}\mathbb{E}[(\Delta^2_{\widetilde{x}_1,\widetilde{x}_3;\widetilde{x}_3}F_n)^2(\Delta^2_{\widetilde{x}_2,\widetilde{x}_3;\widetilde{x}_3}F_n)^2]\lambda^3(d(\widetilde{x}_1,\widetilde{x}_2,\widetilde{x}_3))\right)^{1/2} γ 2 = ( ∫ W n 3 E [( Δ x 1 , x 3 ; x 3 2 F n ) 2 ( Δ x 2 , x 3 ; x 3 2 F n ) 2 ] λ 3 ( d ( x 1 , x 2 , x 3 )) ) 1/2
γ 3 = ∫ W ~ n E [ ∣ Δ x ~ 1 ; x ~ 1 F n ∣ 3 ] λ ( d ( x ~ 1 ) ) \gamma_3=\int_{\widetilde{W}_n}\mathbb{E}[|\Delta_{\widetilde{x}_1;\widetilde{x}_1}F_n|^3]\lambda(d(\widetilde{x}_1)) γ 3 = ∫ W n E [ ∣ Δ x 1 ; x 1 F n ∣ 3 ] λ ( d ( x 1 ))
and
γ 4 = 1 2 E [ F n 4 ] 1 / 4 ( ∫ W ~ n E [ ( Δ x ~ 1 ; x ~ 1 F n ) 4 ] 3 / 4 λ ( d x ~ 1 ) ) \gamma_4=\frac{1}{2}\mathbb{E}[F_n^4]^{1/4}\left(\int_{\widetilde{W}_n}\mathbb{E}[(\Delta_{\widetilde{x}_1;\widetilde{x}_1}F_n)^4]^{3/4}\lambda(d\widetilde{x}_1)\right) γ 4 = 2 1 E [ F n 4 ] 1/4 ( ∫ W n E [( Δ x 1 ; x 1 F n ) 4 ] 3/4 λ ( d x 1 ) )
γ 5 = ( ∫ W ~ n E [ ( Δ x ~ 1 ; x ~ 1 F n ) 4 ] λ ( d x ~ 1 ) ) 1 / 2 \gamma_5=\left(\int_{\widetilde{W}_n}\mathbb{E}[(\Delta_{\widetilde{x}_1;\widetilde{x}_1}F_n)^4]\lambda(d\widetilde{x}_1)\right)^{1/2} γ 5 = ( ∫ W n E [( Δ x 1 ; x 1 F n ) 4 ] λ ( d x 1 ) ) 1/2
γ 6 = ( ∫ W ~ n 2 ( 6 E [ ( Δ x ~ 1 ; x ~ 2 F n ) 4 ] 1 / 2 E [ ( Δ x ~ 1 , x ~ 2 ; x ~ 2 2 F n ) 4 ] 1 / 2 + 3 E [ ( Δ x ~ 1 , x ~ 2 ; x ~ 2 2 F n ) 4 ] ) λ 2 ( d ( x ~ 1 , x ~ 2 ) ) ) 1 / 2 . (16) \gamma_6=\left(\int_{\widetilde{W}_n^2}\left(6\mathbb{E}[(\Delta_{\widetilde{x}_1;\widetilde{x}_2}F_n)^4]^{1/2}\mathbb{E}[(\Delta^2_{\widetilde{x}_1,\widetilde{x}_2;\widetilde{x}_2}F_n)^4]^{1/2}+3\mathbb{E}[(\Delta^2_{\widetilde{x}_1,\widetilde{x}_2;\widetilde{x}_2}F_n)^4]\right)\lambda^2(d(\widetilde{x}_1,\widetilde{x}_2))\right)^{1/2}. \tag*{(16)} γ 6 = ( ∫ W n 2 ( 6 E [( Δ x 1 ; x 2 F n ) 4 ] 1/2 E [( Δ x 1 , x 2 ; x 2 2 F n ) 4 ] 1/2 + 3 E [( Δ x 1 , x 2 ; x 2 2 F n ) 4 ] ) λ 2 ( d ( x 1 , x 2 )) ) 1/2 . (16)
Moreover, the following two finiteness conditions are required:
∫ W ~ n 3 E [ ( Δ x ~ 1 ; x ~ 3 F n ) 4 ] 1 / 4 E [ ( Δ x ~ 2 ; x ~ 3 F n ) 4 ] 1 / 4 E [ ( Δ x ~ 1 , x ~ 3 ; x ~ 3 2 F n ) 4 ] 1 / 4 E [ ( Δ x ~ 2 , x ~ 3 ; x ~ 3 2 F n ) 4 ] 1 / 4 λ 3 ( d ( x ~ 1 , x ~ 2 , x ~ 3 ) ) < ∞ , \int_{\widetilde{W}_n^3} \mathbb{E}\left[(\Delta_{\widetilde{x}_1;\widetilde{x}_3}F_n)^4\right]^{1/4}\mathbb{E}\left[(\Delta_{\widetilde{x}_2;\widetilde{x}_3}F_n)^4\right]^{1/4}\mathbb{E}\left[(\Delta^2_{\widetilde{x}_1,\widetilde{x}_3;\widetilde{x}_3}F_n)^4\right]^{1/4}\mathbb{E}\left[(\Delta^2_{\widetilde{x}_2,\widetilde{x}_3;\widetilde{x}_3}F_n)^4\right]^{1/4}\lambda^3(d(\widetilde{x}_1,\widetilde{x}_2,\widetilde{x}_3)) < \infty, ∫ W n 3 E [ ( Δ x 1 ; x 3 F n ) 4 ] 1/4 E [ ( Δ x 2 ; x 3 F n ) 4 ] 1/4 E [ ( Δ x 1 , x 3 ; x 3 2 F n ) 4 ] 1/4 E [ ( Δ x 2 , x 3 ; x 3 2 F n ) 4 ] 1/4 λ 3 ( d ( x 1 , x 2 , x 3 )) < ∞ ,
∫ W ~ n 3 E [ ( Δ x ~ 1 , x ~ 3 ; x ~ 3 2 F n ) 4 ] 1 / 2 E [ ( Δ x ~ 2 , x ~ 3 ; x ~ 3 2 F n ) 4 ] 1 / 2 λ 3 ( d ( x ~ 1 , x ~ 2 , x ~ 3 ) ) < ∞ . (17) \int_{\widetilde{W}_n^3} \mathbb{E}\left[(\Delta^2_{\widetilde{x}_1,\widetilde{x}_3;\widetilde{x}_3}F_n)^4\right]^{1/2}\mathbb{E}\left[(\Delta^2_{\widetilde{x}_2,\widetilde{x}_3;\widetilde{x}_3}F_n)^4\right]^{1/2}\lambda^3(d(\widetilde{x}_1,\widetilde{x}_2,\widetilde{x}_3)) < \infty. \tag*{(17)} ∫ W n 3 E [ ( Δ x 1 , x 3 ; x 3 2 F n ) 4 ] 1/2 E [ ( Δ x 2 , x 3 ; x 3 2 F n ) 4 ] 1/2 λ 3 ( d ( x 1 , x 2 , x 3 )) < ∞. (17)
To bound E [ F n 4 ] \mathbb{E}[F_n^4] E [ F n 4 ] in γ 4 \gamma_4 γ 4 we may use:
Lemma 9 (Lemma 6.3 in [21 ] ). If E [ F n 4 ] < ∞ \mathbb{E}[F_n^4] < \infty E [ F n 4 ] < ∞ , E F n = 0 \mathbb{E}F_n = 0 E F n = 0 and Var ( F n ) = 1 \operatorname{Var}(F_n) = 1 Var ( F n ) = 1 , then
E [ F n 4 ] ≤ max { 256 ( ∫ W ~ n E [ ( Δ x ~ 1 ; x ~ 1 F n ) 4 ] 1 / 2 λ ( d x ~ 1 ) ) 2 , 4 ∫ W ~ n E [ ( Δ x ~ 1 ; x ~ 1 F n ) 4 ] λ ( d x ~ 1 ) + 2 } . \mathbb{E}[F_n^4] \le\max\left\{256\left(\int_{\widetilde{W}_n}\mathbb{E}\left[(\Delta_{\widetilde{x}_1;\widetilde{x}_1}F_n)^4\right]^{1/2}\lambda(d\widetilde{x}_1)\right)^2,4\int_{\widetilde{W}_n}\mathbb{E}\left[(\Delta_{\widetilde{x}_1;\widetilde{x}_1}F_n)^4\right]\lambda(d\widetilde{x}_1)+2\right\}. E [ F n 4 ] ≤ max { 256 ( ∫ W n E [ ( Δ x 1 ; x 1 F n ) 4 ] 1/2 λ ( d x 1 ) ) 2 , 4 ∫ W n E [ ( Δ x 1 ; x 1 F n ) 4 ] λ ( d x 1 ) + 2 } .
In the following, set
F n : = F ( ξ n ) : = L ( f ; G ( ξ n ) ) , (18) F_n := F(\xi_n) := L(f;G(\xi_n)), \tag*{(18)} F n := F ( ξ n ) := L ( f ; G ( ξ n )) , (18)
and whenever Var ( F n ) > 0 \operatorname{Var}(F_n)>0 Var ( F n ) > 0 , define
F ^ n : = F ( ξ n ) − E [ F ( ξ n ) ] Var ( F ( ξ n ) ) . \widehat{F}_n := \frac{F(\xi_n)-\mathbb{E}[F(\xi_n)]}{\sqrt{\operatorname{Var}(F(\xi_n))}}. F n := Var ( F ( ξ n )) F ( ξ n ) − E [ F ( ξ n )] .
Then we have F n ∈ L 4 F_n \in L^4 F n ∈ L 4 and hence F ^ n ∈ L 4 \widehat{F}_n \in L^4 F n ∈ L 4 for every fixed n n n . Indeed, letting A n A_n A n denote the adjacency matrix of G ( ξ n ) G(\xi_n) G ( ξ n ) , the bounds ∥ A n ∥ o p ≤ N n \lVert A_n\rVert_{\mathrm{op}}\leq N_n ∥ A n ∥ op ≤ N n and ∣ f ( t ) ∣ ≤ C ( 1 + ∣ t ∣ m ) |f(t)|\leq C(1+|t|^m) ∣ f ( t ) ∣ ≤ C ( 1 + ∣ t ∣ m ) give ∣ F n ∣ ≤ C N n ( 1 + N n m ) |F_n|\leq CN_n(1+N_n^m) ∣ F n ∣ ≤ C N n ( 1 + N n m ) , and all moments of the Poisson variable N n N_n N n are finite. We may thus apply Theorem 8 to F ^ n \widehat{F}_n F n and get
d W ( F ^ n , N ( 0 , 1 ) ) ≤ Var ( F n ) − 1 γ 1 + Var ( F n ) − 1 γ 2 + Var ( F n ) − 3 / 2 γ 3 . (19) d_W(\widehat{F}_n,\mathcal{N}(0,1)) \leq\operatorname{Var}(F_n)^{-1}\gamma_1+\operatorname{Var}(F_n)^{-1}\gamma_2+\operatorname{Var}(F_n)^{-3/2}\gamma_3. \tag*{(19)} d W ( F n , N ( 0 , 1 )) ≤ Var ( F n ) − 1 γ 1 + Var ( F n ) − 1 γ 2 + Var ( F n ) − 3/2 γ 3 . (19)
and
d K ( F ^ n , N ( 0 , 1 ) ) ≤ Var ( F n ) − 1 γ 1 + Var ( F n ) − 1 γ 2 + Var ( F n ) − 3 / 2 γ 3 + γ ^ 4 + Var ( F n ) − 1 γ 5 + Var ( F n ) − 1 γ 6 (20) d_K(\widehat{F}_n,\mathcal{N}(0,1)) \leq\operatorname{Var}(F_n)^{-1}\gamma_1+\operatorname{Var}(F_n)^{-1}\gamma_2+\operatorname{Var}(F_n)^{-3/2}\gamma_3+\widehat{\gamma}_4+\operatorname{Var}(F_n)^{-1}\gamma_5+\operatorname{Var}(F_n)^{-1}\gamma_6 \tag*{(20)} d K ( F n , N ( 0 , 1 )) ≤ Var ( F n ) − 1 γ 1 + Var ( F n ) − 1 γ 2 + Var ( F n ) − 3/2 γ 3 + γ 4 + Var ( F n ) − 1 γ 5 + Var ( F n ) − 1 γ 6 (20)
where the γ i \gamma_i γ i are computed for F n F_n F n and γ ^ 4 \widehat{\gamma}_4 γ 4 is computed for F ^ n \widehat{F}_n F n . The powers of the variance in these bounds follow directly from the homogeneity of the difference operators. Indeed, writing s n : = Var ( F n ) > 0 s_n:=\sqrt{\operatorname{Var}(F_n)}>0 s n := Var ( F n ) > 0 ,
Δ x ~ 1 ; x ~ 1 F ^ n = s n − 1 Δ x ~ 1 ; x ~ 1 F n , Δ x ~ 1 , x ~ 2 ; x ~ 2 2 F ^ n = s n − 1 Δ x ~ 1 , x ~ 2 ; x ~ 2 2 F n . \Delta_{\widetilde{x}_1;\widetilde{x}_1}\widehat{F}_n=s_n^{-1}\Delta_{\widetilde{x}_1;\widetilde{x}_1}F_n,\qquad\Delta^2_{\widetilde{x}_1,\widetilde{x}_2;\widetilde{x}_2}\widehat{F}_n=s_n^{-1}\Delta^2_{\widetilde{x}_1,\widetilde{x}_2;\widetilde{x}_2}F_n. Δ x 1 ; x 1 F n = s n − 1 Δ x 1 ; x 1 F n , Δ x 1 , x 2 ; x 2 2 F n = s n − 1 Δ x 1 , x 2 ; x 2 2 F n .
Inspection of (16) therefore gives γ i ( F ^ n ) = s n − 2 γ i ( F n ) \gamma_i(\widehat{F}_n)=s_n^{-2}\gamma_i(F_n) γ i ( F n ) = s n − 2 γ i ( F n ) for i ∈ { 1 , 2 , 5 , 6 } i\in\{1,2,5,6\} i ∈ { 1 , 2 , 5 , 6 } , and γ 3 ( F ^ n ) = s n − 3 γ 3 ( F n ) \gamma_3(\widehat{F}_n)=s_n^{-3}\gamma_3(F_n) γ 3 ( F n ) = s n − 3 γ 3 ( F n ) . For γ 4 \gamma_4 γ 4 , the factor E [ F ^ n 4 ] 1 / 4 \mathbb{E}[\widehat{F}_n^4]^{1/4} E [ F n 4 ] 1/4 must also be controlled, so we keep the notation γ ^ 4 \widehat{\gamma}_4 γ 4 and estimate it separately below. It remains to control the moments of the first- and second-order difference operators of F n F_n F n .
Quantitative CLT
To prove Theorem 2 , we first establish moment bounds for the first- and second-order add-one cost operators Δ x ~ 1 ; x ~ 1 F n \Delta_{\widetilde{x}_1;\widetilde{x}_1}F_n Δ x 1 ; x 1 F n and Δ x ~ 1 , x ~ 2 ; x ~ 2 2 F n \Delta^2_{\widetilde{x}_1,\widetilde{x}_2;\widetilde{x}_2}F_n Δ x 1 , x 2 ; x 2 2 F n . Set
M ‾ φ : = 1 ∨ M φ , 1 , Ψ ℓ , m ( x , y ) : = E [ φ ( ∥ x − y ∥ d ( ℓ m ) κ ( U , V ) ) U ℓ m − 1 V ℓ m − 1 ] , \overline{M}_{\varphi}:=1\vee M_{\varphi,1},\qquad\Psi_{\ell,m}(x,y):=\mathbb{E}\left[\varphi\left(\frac{\lVert x-y\rVert^d}{(\ell m)\kappa(U,V)}\right)U^{\ell m-1}V^{\ell m-1}\right], M φ := 1 ∨ M φ , 1 , Ψ ℓ , m ( x , y ) := E [ φ ( ( ℓ m ) κ ( U , V ) ∥ x − y ∥ d ) U ℓ m − 1 V ℓ m − 1 ] ,
where U , V U,V U , V are independent random variables with distribution μ \mu μ .
Lemma 10 (Uniform fourth-moment bounds for difference operators). Let f f f be a polynomial of degree m ≥ 1 m\geq1 m ≥ 1 . Assume that (PK) holds, that M φ , 1 / 4 < ∞ M_{\varphi,1/4}<\infty M φ , 1/4 < ∞ , and that E [ U 4 m − 1 ] < ∞ \mathbb{E}[U^{4m-1}]<\infty E [ U 4 m − 1 ] < ∞ . There exists a constant C < ∞ C<\infty C < ∞ , independent of n n n , such that the following bounds hold for every n ≥ 1 n\geq1 n ≥ 1 .
(i) First Difference Operator: For every x ~ 1 : = x ~ = ( x , u ) ∈ W ~ n \widetilde{x}_1:=\widetilde{x}=(x,u)\in\widetilde{W}_n x 1 := x = ( x , u ) ∈ W n ,
E [ ∣ Δ ( x , u ) ; x ~ 1 F n ∣ 4 ] ≤ C u 4 m − 1 M ‾ φ 4 m − 1 . (21) \mathbb{E}\left[|\Delta_{(x,u);\widetilde{x}_1}F_n|^4\right]\leq Cu^{4m-1}\overline{M}_{\varphi}^{4m-1}. \tag*{(21)} E [ ∣ Δ ( x , u ) ; x 1 F n ∣ 4 ] ≤ C u 4 m − 1 M φ 4 m − 1 . (21) (ii) Second Difference Operator: For every pair of distinct spatial points x 1 , x 2 ∈ W n x_1,x_2\in W_n x 1 , x 2 ∈ W n ,
∫ ∫ E [ ∣ Δ ( x 1 , u 1 ) , ( x 2 , u 2 ) ; x ~ 2 2 F n ∣ 4 ] μ ( d u 1 ) μ ( d u 2 ) ≤ C Ψ 4 , m ( x 1 , x 2 ) . (22) \int\!\!\int\mathbb{E}\left[|\Delta^2_{(x_1,u_1),(x_2,u_2);\widetilde{x}_2}F_n|^4\right]\mu(du_1)\mu(du_2)\leq C\Psi_{4,m}(x_1,x_2). \tag*{(22)} ∫ ∫ E [ ∣ Δ ( x 1 , u 1 ) , ( x 2 , u 2 ) ; x 2 2 F n ∣ 4 ] μ ( d u 1 ) μ ( d u 2 ) ≤ C Ψ 4 , m ( x 1 , x 2 ) . (22)
The spatial integrations in the normal-approximation bounds use the following scale-mixture estimate.
Lemma 11 (Scale-mixture integrability). Let γ ∈ ( 0 , 1 ] \gamma\in(0,1] γ ∈ ( 0 , 1 ] , ℓ , m ≥ 1 \ell,m \ge1 ℓ , m ≥ 1 , and set a = ℓ m − 1 a=\ell m-1 a = ℓ m − 1 . Assume that (PK) holds, that M φ , γ < ∞ M_{\varphi,\gamma}<\infty M φ , γ < ∞ , and that E [ U b ] < ∞ \mathbb{E}[U^b]<\infty E [ U b ] < ∞ for some b > a b>a b > a satisfying γ ( b − a ) > 1 \gamma(b-a)>1 γ ( b − a ) > 1 . Then,
∫ R d Ψ ℓ , m ( 0 , z ) γ d z < ∞ . \int_{\mathbb{R}^d}\Psi_{\ell,m}(0,z)^\gamma\,\mathrm{d}z<\infty. ∫ R d Ψ ℓ , m ( 0 , z ) γ d z < ∞.
Proof. Since κ ( U , V ) ≤ c κ U V \kappa(U,V)\le c_\kappa UV κ ( U , V ) ≤ c κ U V , decompose according to the events 2 k ≤ U V < 2 k + 1 2^k\le UV<2^{k+1} 2 k ≤ U V < 2 k + 1 , k ≥ 0 k\ge0 k ≥ 0 . Monotonicity of φ \varphi φ gives
Ψ ℓ , m ( 0 , z ) ≤ E [ ( U V ) b ] ∑ k = 0 ∞ 2 − k ( b − a ) φ ( ∥ z ∥ d 2 c κ ( ℓ m ) d 2 k ) . \Psi_{\ell,m}(0,z)\le\mathbb{E}[(UV)^b]\sum_{k=0}^{\infty}2^{-k(b-a)}\varphi\left(\frac{\lVert z\rVert^d}{2c_\kappa(\ell m)^d2^k}\right). Ψ ℓ , m ( 0 , z ) ≤ E [( U V ) b ] k = 0 ∑ ∞ 2 − k ( b − a ) φ ( 2 c κ ( ℓ m ) d 2 k ∥ z ∥ d ) .
For 0 < γ ≤ 1 0<\gamma\le1 0 < γ ≤ 1 and nonnegative numbers a k a_k a k , we have ( ∑ k a k ) γ ≤ ∑ k a k γ (\sum_k a_k)^\gamma\le\sum_k a_k^\gamma ( ∑ k a k ) γ ≤ ∑ k a k γ . Applying this inequality to the preceding bound for Ψ ℓ , m ( 0 , z ) \Psi_{\ell,m}(0,z) Ψ ℓ , m ( 0 , z ) , we have
∫ R d Ψ ℓ , m ( 0 , z ) γ d z ≤ E [ ( U V ) b ] γ ∑ k = 0 ∞ 2 − k γ ( b − a ) ∫ R d φ ( ∥ z ∥ d 2 c κ ( ℓ m ) d 2 k ) γ d z . \int_{\mathbb{R}^d}\Psi_{\ell,m}(0,z)^\gamma\,\mathrm{d}z\le\mathbb{E}[(UV)^b]^\gamma\sum_{k=0}^{\infty}2^{-k\gamma(b-a)}\int_{\mathbb{R}^d}\varphi\left(\frac{\lVert z\rVert^d}{2c_\kappa(\ell m)^d2^k}\right)^\gamma\,\mathrm{d}z. ∫ R d Ψ ℓ , m ( 0 , z ) γ d z ≤ E [( U V ) b ] γ k = 0 ∑ ∞ 2 − k γ ( b − a ) ∫ R d φ ( 2 c κ ( ℓ m ) d 2 k ∥ z ∥ d ) γ d z .
For the integral, setting s k : = 2 c κ ( ℓ m ) d 2 k s_k:=2c_\kappa(\ell m)^d2^k s k := 2 c κ ( ℓ m ) d 2 k ,
∫ R d φ ( ∥ z ∥ d s k ) γ d z = s k ∫ R d φ ( ∥ y ∥ d ) γ d y = 2 c κ ( ℓ m ) d 2 k M φ , γ . \int_{\mathbb{R}^d}\varphi\left(\frac{\lVert z\rVert^d}{s_k}\right)^\gamma\,\mathrm{d}z=s_k\int_{\mathbb{R}^d}\varphi(\lVert y\rVert^d)^\gamma\,\mathrm{d}y=2c_\kappa(\ell m)^d2^kM_{\varphi,\gamma}. ∫ R d φ ( s k ∥ z ∥ d ) γ d z = s k ∫ R d φ (∥ y ∥ d ) γ d y = 2 c κ ( ℓ m ) d 2 k M φ , γ .
It follows that
∫ R d Ψ ℓ , m ( 0 , z ) γ d z ≤ 2 c κ ( ℓ m ) d E [ U b ] 2 γ M φ , γ ∑ k = 0 ∞ 2 − k ( γ ( b − a ) − 1 ) . \int_{\mathbb{R}^d}\Psi_{\ell,m}(0,z)^\gamma\,\mathrm{d}z\le2c_\kappa(\ell m)^d\mathbb{E}[U^b]^{2\gamma}M_{\varphi,\gamma}\sum_{k=0}^{\infty}2^{-k(\gamma(b-a)-1)}. ∫ R d Ψ ℓ , m ( 0 , z ) γ d z ≤ 2 c κ ( ℓ m ) d E [ U b ] 2 γ M φ , γ k = 0 ∑ ∞ 2 − k ( γ ( b − a ) − 1 ) .
The geometric series converges because γ ( b − a ) > 1 \gamma(b-a)>1 γ ( b − a ) > 1 . This proves the lemma. □ \square □
Given Lemma 10 , we conclude the proof of Theorem 2 .
Proof of Theorem 2 . Let f f f be a polynomial of degree m m m , and let F n F_n F n be defined in (18) . By Lemma 10 ,
A x ( u ) : = E [ ∣ Δ ( x , u ) ; x ˘ 1 F n ∣ 4 ] ≤ C u 4 m − 1 , (23) A_x(u):=\mathbb{E}\left[|\Delta_{(x,u);\breve{x}_1}F_n|^4\right]\le Cu^{4m-1}, \tag*{(23)} A x ( u ) := E [ ∣ Δ ( x , u ) ; x ˘ 1 F n ∣ 4 ] ≤ C u 4 m − 1 , (23)
and for B x 1 , x 2 ( u 1 , u 2 ) : = E [ ∣ Δ ( x 1 , u 1 ) , ( x 2 , u 2 ) ; x ˘ 2 2 F n ∣ 4 ] B_{x_1,x_2}(u_1,u_2):=\mathbb{E}\left[|\Delta^2_{(x_1,u_1),(x_2,u_2);\breve{x}_2}F_n|^4\right] B x 1 , x 2 ( u 1 , u 2 ) := E [ ∣ Δ ( x 1 , u 1 ) , ( x 2 , u 2 ) ; x ˘ 2 2 F n ∣ 4 ] , we have
∬ B x 1 , x 2 ( u 1 , u 2 ) μ ( d u 1 ) μ ( d u 2 ) ≤ C Ψ 4 , m ( x 1 , x 2 ) . (24) \iint B_{x_1,x_2}(u_1,u_2)\,\mu(\mathrm{d}u_1)\mu(\mathrm{d}u_2)\le C\Psi_{4,m}(x_1,x_2). \tag*{(24)} ∬ B x 1 , x 2 ( u 1 , u 2 ) μ ( d u 1 ) μ ( d u 2 ) ≤ C Ψ 4 , m ( x 1 , x 2 ) . (24)
Assume m ≥ 2 m\ge2 m ≥ 2 . We first bound γ 1 \gamma_1 γ 1 . For fixed inserted positions and marks, Cauchy–Schwarz gives
E [ ( Δ x ˘ 1 ; x ˘ 3 F n ) 2 ( Δ x ˘ 2 ; x ˘ 3 F n ) 2 ] 1 / 2 ≤ A x 1 ( u 1 ) 1 / 4 A x 2 ( u 2 ) 1 / 4 , \mathbb{E}\left[(\Delta_{\breve{x}_1;\breve{x}_3}F_n)^2(\Delta_{\breve{x}_2;\breve{x}_3}F_n)^2\right]^{1/2}\le A_{x_1}(u_1)^{1/4}A_{x_2}(u_2)^{1/4}, E [ ( Δ x ˘ 1 ; x ˘ 3 F n ) 2 ( Δ x ˘ 2 ; x ˘ 3 F n ) 2 ] 1/2 ≤ A x 1 ( u 1 ) 1/4 A x 2 ( u 2 ) 1/4 ,
and similarly
E [ ( Δ x ˘ 1 , x ˘ 3 ; x ˘ 3 2 F n ) 2 ( Δ x ˘ 2 , x ˘ 3 ; x ˘ 3 2 F n ) 2 ] 1 / 2 ≤ B x 1 , x 3 ( u 1 , u 3 ) 1 / 4 B x 2 , x 3 ( u 2 , u 3 ) 1 / 4 . \mathbb{E}\left[(\Delta^2_{\breve{x}_1,\breve{x}_3;\breve{x}_3}F_n)^2(\Delta^2_{\breve{x}_2,\breve{x}_3;\breve{x}_3}F_n)^2\right]^{1/2}\le B_{x_1,x_3}(u_1,u_3)^{1/4}B_{x_2,x_3}(u_2,u_3)^{1/4}. E [ ( Δ x ˘ 1 , x ˘ 3 ; x ˘ 3 2 F n ) 2 ( Δ x ˘ 2 , x ˘ 3 ; x ˘ 3 2 F n ) 2 ] 1/2 ≤ B x 1 , x 3 ( u 1 , u 3 ) 1/4 B x 2 , x 3 ( u 2 , u 3 ) 1/4 .
Substituting these inequalities into the definition of γ 1 \gamma_1 γ 1 gives
γ 1 2 4 ≤ ∫ W n 3 ∫ A x 1 ( u 1 ) 1 / 4 A x 2 ( u 2 ) 1 / 4 B x 1 , x 3 ( u 1 , u 3 ) 1 / 4 B x 2 , x 3 ( u 2 , u 3 ) 1 / 4 μ 3 ( d ( u 1 , u 2 , u 3 ) ) d x 1 d x 2 d x 3 . \frac{\gamma_1^2}{4}\le\int_{W_{n^3}}\int A_{x_1}(u_1)^{1/4}A_{x_2}(u_2)^{1/4}B_{x_1,x_3}(u_1,u_3)^{1/4}B_{x_2,x_3}(u_2,u_3)^{1/4}\,\mu^3(\mathrm{d}(u_1,u_2,u_3))\,\mathrm{d}x_1\mathrm{d}x_2\mathrm{d}x_3. 4 γ 1 2 ≤ ∫ W n 3 ∫ A x 1 ( u 1 ) 1/4 A x 2 ( u 2 ) 1/4 B x 1 , x 3 ( u 1 , u 3 ) 1/4 B x 2 , x 3 ( u 2 , u 3 ) 1/4 μ 3 ( d ( u 1 , u 2 , u 3 )) d x 1 d x 2 d x 3 .
For fixed x 1 , x 2 , x 3 x_1,x_2,x_3 x 1 , x 2 , x 3 , apply Hölder’s inequality with four exponents equal to 4 on the common mark space:
∫ A x 1 ( u 1 ) 1 / 4 A x 2 ( u 2 ) 1 / 4 B x 1 , x 3 ( u 1 , u 3 ) 1 / 4 B x 2 , x 3 ( u 2 , u 3 ) 1 / 4 d μ 3 ≤ ( ∫ A x 1 ( u 1 ) d μ 3 ) 1 / 4 ( ∫ A x 2 ( u 2 ) d μ 3 ) 1 / 4 ( ∫ B x 1 , x 3 ( u 1 , u 3 ) d μ 3 ) 1 / 4 ( ∫ B x 2 , x 3 ( u 2 , u 3 ) d μ 3 ) 1 / 4 . \begin{aligned}
&\int A_{x_1}(u_1)^{1/4}A_{x_2}(u_2)^{1/4}B_{x_1,x_3}(u_1,u_3)^{1/4}B_{x_2,x_3}(u_2,u_3)^{1/4}\,\mathrm{d}\mu^3 \\
&\le\left(\int A_{x_1}(u_1)\,\mathrm{d}\mu^3\right)^{1/4}\left(\int A_{x_2}(u_2)\,\mathrm{d}\mu^3\right)^{1/4}\left(\int B_{x_1,x_3}(u_1,u_3)\,\mathrm{d}\mu^3\right)^{1/4}\left(\int B_{x_2,x_3}(u_2,u_3)\,\mathrm{d}\mu^3\right)^{1/4}.
\end{aligned} ∫ A x 1 ( u 1 ) 1/4 A x 2 ( u 2 ) 1/4 B x 1 , x 3 ( u 1 , u 3 ) 1/4 B x 2 , x 3 ( u 2 , u 3 ) 1/4 d μ 3 ≤ ( ∫ A x 1 ( u 1 ) d μ 3 ) 1/4 ( ∫ A x 2 ( u 2 ) d μ 3 ) 1/4 ( ∫ B x 1 , x 3 ( u 1 , u 3 ) d μ 3 ) 1/4 ( ∫ B x 2 , x 3 ( u 2 , u 3 ) d μ 3 ) 1/4 .
Integration over a mark absent from an integrand contributes a factor 1. Thus, by (23)
∫ A x i ( u i ) d μ 3 = ∫ A x i ( u i ) μ ( d u i ) ≤ C E [ U 4 m − 1 ] < ∞ , (25) \int A_{x_i}(u_i)\,\mathrm{d}\mu^3=\int A_{x_i}(u_i)\,\mu(\mathrm{d}u_i)\le C\mathbb{E}[U^{4m-1}]<\infty, \tag*{(25)} ∫ A x i ( u i ) d μ 3 = ∫ A x i ( u i ) μ ( d u i ) ≤ C E [ U 4 m − 1 ] < ∞ , (25)
whereas (24) gives
∫ B x i , x 3 ( u i , u 3 ) d μ 3 ≤ C Ψ 4 , m ( x i , x 3 ) , i = 1 , 2. \int B_{x_i,x_3}(u_i,u_3)\,d\mu^3 \leq C\Psi_{4,m}(x_i,x_3), \qquad i=1,2. ∫ B x i , x 3 ( u i , u 3 ) d μ 3 ≤ C Ψ 4 , m ( x i , x 3 ) , i = 1 , 2.
Therefore,
γ 1 2 ≤ C ∫ W n 3 Ψ 4 , m ( x 1 , x 3 ) 1 / 4 Ψ 4 , m ( x 2 , x 3 ) 1 / 4 d x 1 d x 2 d x 3 . \gamma_1^2 \leq C\int_{W_n^3}\Psi_{4,m}(x_1,x_3)^{1/4}\Psi_{4,m}(x_2,x_3)^{1/4}\,dx_1dx_2dx_3. γ 1 2 ≤ C ∫ W n 3 Ψ 4 , m ( x 1 , x 3 ) 1/4 Ψ 4 , m ( x 2 , x 3 ) 1/4 d x 1 d x 2 d x 3 .
For m ≥ 2 m\geq2 m ≥ 2 , recalling the condition E [ U 6 m ] < ∞ \mathbb{E}[U^{6m}]<\infty E [ U 6 m ] < ∞ , Lemma 11 applies with ℓ = 4 \ell=4 ℓ = 4 , γ = 1 / 4 \gamma=1/4 γ = 1/4 , and b = 6 m b=6m b = 6 m , since ( 6 m − ( 4 m − 1 ) ) = 2 m + 1 > 4 (6m-(4m-1))=2m+1>4 ( 6 m − ( 4 m − 1 )) = 2 m + 1 > 4 . Hence by translation invariance of Ψ 4 , m \Psi_{4,m} Ψ 4 , m ,
γ 1 2 ≤ C ∫ W n ( ∫ W n Ψ 4 , m ( x 1 , x 3 ) 1 / 4 d x 1 ) ( ∫ W n Ψ 4 , m ( x 2 , x 3 ) 1 / 4 d x 2 ) d x 3 ≤ C ∣ W n ∣ . \gamma_1^2 \leq C\int_{W_n}\left(\int_{W_n}\Psi_{4,m}(x_1,x_3)^{1/4}\,dx_1\right)\left(\int_{W_n}\Psi_{4,m}(x_2,x_3)^{1/4}\,dx_2\right)\,dx_3 \leq C|W_n|. γ 1 2 ≤ C ∫ W n ( ∫ W n Ψ 4 , m ( x 1 , x 3 ) 1/4 d x 1 ) ( ∫ W n Ψ 4 , m ( x 2 , x 3 ) 1/4 d x 2 ) d x 3 ≤ C ∣ W n ∣.
For γ 2 \gamma_2 γ 2 , we similarly obtain
γ 2 2 ≤ ∫ W n 3 ∫ B x 1 , x 3 ( u 1 , u 3 ) 1 / 2 B x 2 , x 3 ( u 2 , u 3 ) 1 / 2 μ 3 ( d ( u 1 , u 2 , u 3 ) ) d ( x 1 , x 2 , x 3 ) ≤ C ∫ W n 3 Ψ 4 , m ( x 1 , x 3 ) 1 / 2 Ψ 4 , m ( x 2 , x 3 ) 1 / 2 d ( x 1 , x 2 , x 3 ) , \begin{aligned}
\gamma_2^2 \leq\int_{W_n^3}\int B_{x_1,x_3}(u_1,u_3)^{1/2}B_{x_2,x_3}(u_2,u_3)^{1/2}\,\mu^3(d(u_1,u_2,u_3))\,d(x_1,x_2,x_3) \\
&\leq C\int_{W_n^3}\Psi_{4,m}(x_1,x_3)^{1/2}\Psi_{4,m}(x_2,x_3)^{1/2}\,d(x_1,x_2,x_3),
\end{aligned} γ 2 2 ≤ ∫ W n 3 ∫ B x 1 , x 3 ( u 1 , u 3 ) 1/2 B x 2 , x 3 ( u 2 , u 3 ) 1/2 μ 3 ( d ( u 1 , u 2 , u 3 )) d ( x 1 , x 2 , x 3 ) ≤ C ∫ W n 3 Ψ 4 , m ( x 1 , x 3 ) 1/2 Ψ 4 , m ( x 2 , x 3 ) 1/2 d ( x 1 , x 2 , x 3 ) ,
where the second inequality follows from Cauchy–Schwarz in the three mark variables. Since M φ , 1 / 2 ≤ M φ , 1 / 4 M_{\varphi,1/2}\leq M_{\varphi,1/4} M φ , 1/2 ≤ M φ , 1/4 , Lemma 11 with γ = 1 / 2 \gamma=1/2 γ = 1/2 and b = 6 m b=6m b = 6 m yields γ 2 2 ∈ O ( ∣ W n ∣ ) \gamma_2^2\in O(|W_n|) γ 2 2 ∈ O ( ∣ W n ∣ ) . The same estimates also verify the two finiteness conditions in (17) . Next, by (23) , we have γ 3 ≤ ∫ W ~ n A x ( u ) 3 / 4 λ ( d ( x , u ) ) ∈ O ( ∣ W n ∣ ) \gamma_3\leq\int_{\widetilde W_n}A_x(u)^{3/4}\lambda(d(x,u))\in O(|W_n|) γ 3 ≤ ∫ W n A x ( u ) 3/4 λ ( d ( x , u )) ∈ O ( ∣ W n ∣ ) , since E [ U 3 ( 4 m − 1 ) / 4 ] < ∞ \mathbb{E}[U^{3(4m-1)/4}]<\infty E [ U 3 ( 4 m − 1 ) /4 ] < ∞ .
We now consider the further terms needed for the Kolmogorov distance. Here, ∫ W ~ n A x ( u ) λ ( d ( x , u ) ) ∈ O ( ∣ W n ∣ ) \int_{\widetilde W_n}A_x(u)\lambda(d(x,u))\in O(|W_n|) ∫ W n A x ( u ) λ ( d ( x , u )) ∈ O ( ∣ W n ∣ ) and ∫ W ~ n A x ( u ) 1 / 2 λ ( d ( x , u ) ) ∈ O ( ∣ W n ∣ ) \int_{\widetilde W_n}A_x(u)^{1/2}\lambda(d(x,u))\in O(|W_n|) ∫ W n A x ( u ) 1/2 λ ( d ( x , u )) ∈ O ( ∣ W n ∣ ) . If σ f 2 > 0 \sigma_f^2>0 σ f 2 > 0 , then by Theorem 1 , Var ( F n ) ≍ ∣ W n ∣ \operatorname{Var}(F_n)\asymp|W_n| Var ( F n ) ≍ ∣ W n ∣ . The two integrals in Lemma 9 satisfy
∫ W ~ n E [ ( Δ x ~ ; ( x ~ ) F ^ n ) 4 ] 1 / 2 λ ( d x ~ ) ≤ C ∣ W n ∣ Var ( F n ) , ∫ W ~ n E [ ( Δ x ~ ; ( x ~ ) F ^ n ) 4 ] λ ( d x ~ ) ≤ C ∣ W n ∣ Var ( F n ) 2 . \int_{\widetilde W_n}\mathbb{E}\left[(\Delta_{\widetilde x;(\widetilde x)}\widehat F_n)^4\right]^{1/2}\lambda(d\widetilde x)\leq C\frac{|W_n|}{\operatorname{Var}(F_n)},
\qquad
\int_{\widetilde W_n}\mathbb{E}\left[(\Delta_{\widetilde x;(\widetilde x)}\widehat F_n)^4\right]\lambda(d\widetilde x)\leq C\frac{|W_n|}{\operatorname{Var}(F_n)^2}. ∫ W n E [ ( Δ x ; ( x ) F n ) 4 ] 1/2 λ ( d x ) ≤ C Var ( F n ) ∣ W n ∣ , ∫ W n E [ ( Δ x ; ( x ) F n ) 4 ] λ ( d x ) ≤ C Var ( F n ) 2 ∣ W n ∣ .
Lemma 9 therefore gives E [ F ^ n 4 ] ≤ C \mathbb{E}[\widehat F_n^4]\leq C E [ F n 4 ] ≤ C . It follows that
γ ^ 4 = 1 2 E [ F ^ n 4 ] 1 / 4 Var ( F n ) − 3 / 2 ∫ W ~ n A x ( u ) 3 / 4 λ ( d x ~ ) ≤ C Var ( F n ) − 3 / 2 ∣ W n ∣ ≤ C ∣ W n ∣ − 1 / 2 . \widehat\gamma_4=\frac{1}{2}\mathbb{E}[\widehat F_n^4]^{1/4}\operatorname{Var}(F_n)^{-3/2}\int_{\widetilde W_n}A_x(u)^{3/4}\lambda(d\widetilde x)\leq C\operatorname{Var}(F_n)^{-3/2}|W_n|\leq C|W_n|^{-1/2}. γ 4 = 2 1 E [ F n 4 ] 1/4 Var ( F n ) − 3/2 ∫ W n A x ( u ) 3/4 λ ( d x ) ≤ C Var ( F n ) − 3/2 ∣ W n ∣ ≤ C ∣ W n ∣ − 1/2 .
Moreover, γ 5 2 = ∫ W ~ n A x ( u ) λ ( d ( x , u ) ) ∈ O ( ∣ W n ∣ ) \gamma_5^2=\int_{\widetilde W_n}A_x(u)\lambda(d(x,u))\in O(|W_n|) γ 5 2 = ∫ W n A x ( u ) λ ( d ( x , u )) ∈ O ( ∣ W n ∣ ) .
Finally, consider γ 6 \gamma_6 γ 6 . By Cauchy–Schwarz in the two mark variables,
∬ A x 1 ( u 1 ) 1 / 2 B x 1 , x 2 ( u 1 , u 2 ) 1 / 2 μ ( d u 1 ) μ ( d u 2 ) ≤ ( ∫ A x 1 ( u 1 ) μ ( d u 1 ) ) 1 / 2 ( ∬ B x 1 , x 2 ( u 1 , u 2 ) μ ( d u 1 ) μ ( d u 2 ) ) 1 / 2 ≤ C Ψ 4 , m ( x 1 , x 2 ) 1 / 2 , \begin{aligned}
\iint A_{x_1}(u_1)^{1/2}B_{x_1,x_2}(u_1,u_2)^{1/2}\,\mu(du_1)\mu(du_2) \\
&\leq\left(\int A_{x_1}(u_1)\,\mu(du_1)\right)^{1/2}\left(\iint B_{x_1,x_2}(u_1,u_2)\,\mu(du_1)\mu(du_2)\right)^{1/2}\leq C\Psi_{4,m}(x_1,x_2)^{1/2},
\end{aligned} ∬ A x 1 ( u 1 ) 1/2 B x 1 , x 2 ( u 1 , u 2 ) 1/2 μ ( d u 1 ) μ ( d u 2 ) ≤ ( ∫ A x 1 ( u 1 ) μ ( d u 1 ) ) 1/2 ( ∬ B x 1 , x 2 ( u 1 , u 2 ) μ ( d u 1 ) μ ( d u 2 ) ) 1/2 ≤ C Ψ 4 , m ( x 1 , x 2 ) 1/2 ,
whereas directly from (24) , we get that ∬ B x 1 , x 2 ( u 1 , u 2 ) μ ( d u 1 ) μ ( d u 2 ) ≤ C Ψ 4 , m ( x 1 , x 2 ) \iint B_{x_1,x_2}(u_1,u_2)\,\mu(du_1)\mu(du_2)\leq C\Psi_{4,m}(x_1,x_2) ∬ B x 1 , x 2 ( u 1 , u 2 ) μ ( d u 1 ) μ ( d u 2 ) ≤ C Ψ 4 , m ( x 1 , x 2 ) . It follows that
γ 6 2 ≤ C ∫ W n 2 ( Ψ 4 , m ( x 1 , x 2 ) 1 / 2 + Ψ 4 , m ( x 1 , x 2 ) ) d ( x 1 , x 2 ) ∈ O ( ∣ W n ∣ ) , \gamma_6^2\leq C\int_{W_n^2}\left(\Psi_{4,m}(x_1,x_2)^{1/2}+\Psi_{4,m}(x_1,x_2)\right)\,d(x_1,x_2)\in O(|W_n|), γ 6 2 ≤ C ∫ W n 2 ( Ψ 4 , m ( x 1 , x 2 ) 1/2 + Ψ 4 , m ( x 1 , x 2 ) ) d ( x 1 , x 2 ) ∈ O ( ∣ W n ∣ ) ,
by Lemma 11 with γ = 1 / 2 \gamma=1/2 γ = 1/2 and γ = 1 \gamma=1 γ = 1 , respectively.
If m ∈ { 0 , 1 } m\in\{0,1\} m ∈ { 0 , 1 } , then F n = a 0 N n F_n=a_0N_n F n = a 0 N n , because Tr ( A n ) = 0 \operatorname{Tr}(A_n)=0 Tr ( A n ) = 0 . It follows that Δ x ~ ; ( x ~ ) F n = a 0 \Delta_{\widetilde x;(\widetilde x)}F_n=a_0 Δ x ; ( x ) F n = a 0 and Δ x ~ 1 , x ~ 2 ; x ~ 2 2 F n = 0 \Delta^2_{\widetilde x_1,\widetilde x_2;\widetilde x_2}F_n=0 Δ x 1 , x 2 ; x 2 2 F n = 0 , from which we obtain A x ( u ) ≤ a 0 4 A_x(u)\leq a_0^4 A x ( u ) ≤ a 0 4 and B x 1 , x 2 ( u 1 , u 2 ) = 0 B_{x_1,x_2}(u_1,u_2)=0 B x 1 , x 2 ( u 1 , u 2 ) = 0 . This can be inserted in the estimates above to obtain (at most) the same order of the γ i \gamma_i γ i .
Assume first that σ f 2 > 0 \sigma_f^2>0 σ f 2 > 0 . Theorem 1 gives Var ( F n ) ≥ σ f 2 ∣ W n ∣ / 2 \operatorname{Var}(F_n)\geq\sigma_f^2|W_n|/2 Var ( F n ) ≥ σ f 2 ∣ W n ∣/2 for all sufficiently large n n n . Inserting the preceding estimates into (19) yields d W ( F ^ n , N ( 0 , 1 ) ) ≤ C ∣ W n ∣ − 1 / 2 d_W(\widehat F_n,\mathcal{N}(0,1))\leq C|W_n|^{-1/2} d W ( F n , N ( 0 , 1 )) ≤ C ∣ W n ∣ − 1/2 . For the Kolmogorov distance, the preceding estimates give γ ^ 4 ≤ C ∣ W n ∣ − 1 / 2 \widehat\gamma_4\leq C|W_n|^{-1/2} γ 4 ≤ C ∣ W n ∣ − 1/2 and γ 5 + γ 6 ≤ C ∣ W n ∣ 1 / 2 \gamma_5+\gamma_6\leq C|W_n|^{1/2} γ 5 + γ 6 ≤ C ∣ W n ∣ 1/2 , so their normalized contributions in (20) are also at most C ∣ W n ∣ − 1 / 2 C|W_n|^{-1/2} C ∣ W n ∣ − 1/2 , yielding d K ( F ^ n , N ( 0 , 1 ) ) ≤ C ∣ W n ∣ − 1 / 2 d_K(\widehat F_n,\mathcal{N}(0,1))\leq C|W_n|^{-1/2} d K ( F n , N ( 0 , 1 )) ≤ C ∣ W n ∣ − 1/2 . Moreover, ( F n − E [ F n ] ) / ∣ W n ∣ = F ^ n Var ( F n ) / ∣ W n ∣ (F_n-\mathbb{E}[F_n])/\sqrt{|W_n|}=\widehat F_n\sqrt{\operatorname{Var}(F_n)/|W_n|} ( F n − E [ F n ]) / ∣ W n ∣ = F n Var ( F n ) /∣ W n ∣ , and Slutsky’s theorem together with the variance limit proves convergence to N ( 0 , σ f 2 ) \mathcal{N}(0,\sigma_f^2) N ( 0 , σ f 2 ) .
If σ f 2 = 0 \sigma_f^2=0 σ f 2 = 0 , then we have E [ ( ( F n − E [ F n ] ) / ∣ W n ∣ ) 2 ] = Var ( F n ) / ∣ W n ∣ ⟶ 0 \mathbb{E}\left[\left((F_n-\mathbb{E}[F_n])/\sqrt{|W_n|}\right)^2\right]=\operatorname{Var}(F_n)/|W_n|\longrightarrow0 E [ ( ( F n − E [ F n ]) / ∣ W n ∣ ) 2 ] = Var ( F n ) /∣ W n ∣ ⟶ 0 . Thus, the volume-normalized statistic converges in distribution to N ( 0 , 0 ) \mathcal{N}(0,0) N ( 0 , 0 ) , which completes the proof. □
It remains to show Lemma 10. Observe that, by (8) , F n F_n F n can be expressed as a sum of scores
L ( f ; G ( ξ n ) ) = ∑ X ˇ i ∈ P ˇ n g f ( X ˇ i , G ( ξ n ) ) , (26) L(f;G(\xi_n))=\sum_{\check{X}_i\in\check{P}_n}g_f(\check{X}_i,G(\xi_n)), \tag*{(26)} L ( f ; G ( ξ n )) = X ˇ i ∈ P ˇ n ∑ g f ( X ˇ i , G ( ξ n )) , (26)
where the score function g f g_f g f is given for a finite graph G G G with marked vertex set Q ˇ ⊆ R d × [ 1 , ∞ ) \check{Q}\subseteq\mathbb{R}^d\times[1,\infty) Q ˇ ⊆ R d × [ 1 , ∞ ) and a marked vertex x ˇ 1 = ( x 1 , u 1 ) ∈ Q ˇ \check{x}_1=(x_1,u_1)\in\check{Q} x ˇ 1 = ( x 1 , u 1 ) ∈ Q ˇ by
g f ( x ˇ 1 , G ) : = a 0 + ∑ p = 2 m ∑ ( x ˇ 2 , … , x ˇ p ) ∈ ( Q ˇ ∖ { x ˇ 1 } ) ( p − 1 ) ∑ q = p m a q 1 ( p − 1 ) ! ∑ π ∈ S q , p ∏ j = 1 q A ( x ˇ π ( j ) , x ˇ π ( j + 1 ) ; G ) . (27) g_f(\check{x}_1,G):=a_0+\sum_{p=2}^{m}\sum_{(\check{x}_2,\ldots,\check{x}_p)\in(\check{Q}\setminus\{\check{x}_1\})^{(p-1)}}\sum_{q=p}^{m}a_q\frac{1}{(p-1)!}\sum_{\pi\in\mathfrak{S}_{q,p}}\prod_{j=1}^{q}A(\check{x}_{\pi(j)},\check{x}_{\pi(j+1)};G). \tag*{(27)} g f ( x ˇ 1 , G ) := a 0 + p = 2 ∑ m ( x ˇ 2 , … , x ˇ p ) ∈ ( Q ˇ ∖ { x ˇ 1 } ) ( p − 1 ) ∑ q = p ∑ m a q ( p − 1 )! 1 π ∈ S q , p ∑ j = 1 ∏ q A ( x ˇ π ( j ) , x ˇ π ( j + 1 ) ; G ) . (27)
A non-zero term can be interpreted as a closed walk of length q q q starting and ending at x ˇ 1 \check{x}_1 x ˇ 1 visiting exactly p p p distinct points. Therefore, there are only ( p − 1 ) ! (p-1)! ( p − 1 )! enumerations of the remaining vertices. This explains the coefficient 1 / ( p − 1 ) ! 1/(p-1)! 1/ ( p − 1 )! in the above expression. Summing over all possible starting points gives that (23) is equivalent to (8) .
To derive Lemma 10, we first establish corresponding moment bounds for the scores g f g_f g f in Lemma 12. We begin by introducing notations. Fix an ordered tuple
x ˇ r = ( x ˇ 1 , … , x ˇ r ) , x ˇ i = ( x i , u i ) ∈ R d × [ 1 , ∞ ) , \check{x}_r=(\check{x}_1,\ldots,\check{x}_r),\qquad\check{x}_i=(x_i,u_i)\in\mathbb{R}^d\times[1,\infty), x ˇ r = ( x ˇ 1 , … , x ˇ r ) , x ˇ i = ( x i , u i ) ∈ R d × [ 1 , ∞ ) ,
with distinct spatial coordinates. For I ⊆ { 1 , … , r } I\subseteq\{1,\ldots,r\} I ⊆ { 1 , … , r } , write
G n , I , x ˇ r : = G n , I ( ξ , x ˇ r ) , G n , x ˇ r : = G n , { 1 , … , r } , x ˇ r G_{n,I,\check{x}_r}:=G_{n,I}(\xi,\check{x}_r),\qquad G_{n,\check{x}_r}:=G_{n,\{1,\ldots,r\},\check{x}_r} G n , I , x ˇ r := G n , I ( ξ , x ˇ r ) , G n , x ˇ r := G n , { 1 , … , r } , x ˇ r
(see § 3.2 for the definition of G n , I G_{n,I} G n , I ).
We now define deletion differences for score functions. For a finite graph G G G containing distinct marked vertices z ˇ \check{z} z ˇ and x ˇ \check{x} x ˇ , define the first-order difference
∂ x ˇ g f ( z ˇ , G ) : = g f ( z ˇ , G ) − g f ( z ˇ , G ∖ { x ˇ } ) . \partial_{\check{x}}g_f(\check{z},G):=g_f(\check{z},G)-g_f(\check{z},G\setminus\{\check{x}\}). ∂ x ˇ g f ( z ˇ , G ) := g f ( z ˇ , G ) − g f ( z ˇ , G ∖ { x ˇ }) .
Deleting a marked vertex means deleting its spatial vertex and all incident edges, while retaining every other vertex, mark, and edge. Similarly for distinct marked vertices z ˇ , x ˇ , y ˇ \check{z},\check{x},\check{y} z ˇ , x ˇ , y ˇ of G G G , define the second-order difference
∂ x ˇ , y ˇ 2 g f ( z ˇ , G ) : = g f ( z ˇ , G ) − g f ( z ˇ , G ∖ { x ˇ } ) − g f ( z ˇ , G ∖ { y ˇ } ) + g f ( z ˇ , G ∖ { x ˇ , y ˇ } ) . \partial^2_{\check{x},\check{y}}g_f(\check{z},G):=g_f(\check{z},G)-g_f(\check{z},G\setminus\{\check{x}\})-g_f(\check{z},G\setminus\{\check{y}\})+g_f(\check{z},G\setminus\{\check{x},\check{y}\}). ∂ x ˇ , y ˇ 2 g f ( z ˇ , G ) := g f ( z ˇ , G ) − g f ( z ˇ , G ∖ { x ˇ }) − g f ( z ˇ , G ∖ { y ˇ }) + g f ( z ˇ , G ∖ { x ˇ , y ˇ }) .
For example,
∂ x ˇ i g f ( z ˇ , G n , x ˇ r ) = g f ( z ˇ , G n , { 1 , … , r } , x ˇ r ) − g f ( z ˇ , G n , { 1 , … , r } ∖ { i } , x ˇ r ) . \partial_{\check{x}_i}g_f(\check{z},G_{n,\check{x}_r})=g_f(\check{z},G_{n,\{1,\ldots,r\},\check{x}_r})-g_f(\check{z},G_{n,\{1,\ldots,r\}\setminus\{i\},\check{x}_r}). ∂ x ˇ i g f ( z ˇ , G n , x ˇ r ) = g f ( z ˇ , G n , { 1 , … , r } , x ˇ r ) − g f ( z ˇ , G n , { 1 , … , r } ∖ { i } , x ˇ r ) .
Lemma 12 (Moment bounds). Let ℓ , m ∈ N \ell,m\in\mathbb{N} ℓ , m ∈ N and let f f f be a polynomial of degree m m m . Assume that (PK) holds, that M φ , 1 < ∞ M_{\varphi,1}<\infty M φ , 1 < ∞ , and that E [ U ℓ m − 1 ] < ∞ \mathbb{E}[U^{\ell m-1}]<\infty E [ U ℓ m − 1 ] < ∞ . Let x ˇ 1 = ( x 1 , u 1 ) , x ˇ 2 = ( x 2 , u 2 ) ∈ W ˇ n \check{x}_1=(x_1,u_1),\check{x}_2=(x_2,u_2)\in\check{W}_n x ˇ 1 = ( x 1 , u 1 ) , x ˇ 2 = ( x 2 , u 2 ) ∈ W ˇ n with x 1 ≠ x 2 x_1\ne x_2 x 1 = x 2 . Then, there exists a constant C > 0 C>0 C > 0 (depending on ℓ \ell ℓ and f f f , but not on n , x ˇ 1 , x ˇ 2 n,\check{x}_1,\check{x}_2 n , x ˇ 1 , x ˇ 2 ) such that, for every n ∈ N n\in\mathbb{N} n ∈ N , the following holds:
(i) E [ ∣ g f ( x ˇ 1 , G n , x ˇ 1 ) ∣ ℓ ] ≤ C u 1 ℓ m − 1 M φ ℓ m − 1 \mathbb{E}[|g_f(\check{x}_1,G_{n,\check{x}_1})|^\ell]\le Cu_1^{\ell m-1}M_\varphi^{\ell m-1} E [ ∣ g f ( x ˇ 1 , G n , x ˇ 1 ) ∣ ℓ ] ≤ C u 1 ℓ m − 1 M φ ℓ m − 1 .
(ii) E [ ∣ ∑ X ˇ i ∈ P ˇ n ∂ x ˇ 1 g f ( X ˇ i , G n , x ˇ 1 ) ∣ ℓ ] ≤ C u 1 ℓ m − 1 M φ ℓ m − 1 \mathbb{E}\left[\left|\sum_{\check{X}_i\in\check{P}_n}\partial_{\check{x}_1}g_f(\check{X}_i,G_{n,\check{x}_1})\right|^\ell\right]\le Cu_1^{\ell m-1}M_\varphi^{\ell m-1} E [ ∑ X ˇ i ∈ P ˇ n ∂ x ˇ 1 g f ( X ˇ i , G n , x ˇ 1 ) ℓ ] ≤ C u 1 ℓ m − 1 M φ ℓ m − 1 .
(iii) ∬ E [ ∣ ∂ x ˇ 1 g f ( x ˇ 2 , G n , x ˇ 2 ) ∣ ℓ ] μ ( d u 1 ) μ ( d u 2 ) ≤ C M φ ℓ m − 2 Ψ ℓ , m ( x 1 , x 2 ) \iint\mathbb{E}[|\partial_{\check{x}_1}g_f(\check{x}_2,G_{n,\check{x}_2})|^\ell]\,\mu(du_1)\mu(du_2)\le CM_\varphi^{\ell m-2}\Psi_{\ell,m}(x_1,x_2) ∬ E [ ∣ ∂ x ˇ 1 g f ( x ˇ 2 , G n , x ˇ 2 ) ∣ ℓ ] μ ( d u 1 ) μ ( d u 2 ) ≤ C M φ ℓ m − 2 Ψ ℓ , m ( x 1 , x 2 ) .
(iv) ∬ E [ ∣ ∑ X ˇ i ∈ P ˇ n ∂ x ˇ 1 , x ˇ 2 2 g f ( X ˇ i , G n , x ˇ 2 ) ∣ ℓ ] μ ( d u 1 ) μ ( d u 2 ) ≤ C M φ ℓ m − 2 Ψ ℓ , m ( x 1 , x 2 ) \iint\mathbb{E}\left[\left|\sum_{\check{X}_i\in\check{P}_n}\partial^2_{\check{x}_1,\check{x}_2}g_f(\check{X}_i,G_{n,\check{x}_2})\right|^\ell\right]\,\mu(du_1)\mu(du_2)\le CM_\varphi^{\ell m-2}\Psi_{\ell,m}(x_1,x_2) ∬ E [ ∑ X ˇ i ∈ P ˇ n ∂ x ˇ 1 , x ˇ 2 2 g f ( X ˇ i , G n , x ˇ 2 ) ℓ ] μ ( d u 1 ) μ ( d u 2 ) ≤ C M φ ℓ m − 2 Ψ ℓ , m ( x 1 , x 2 ) .
Proof. Write f ( x ) = ∑ q = 0 m a q x q f(x)=\sum_{q=0}^{m}a_qx^q f ( x ) = ∑ q = 0 m a q x q and let g q : = g x q g_q:=g_{x^q} g q := g x q denote the score associated with x q x^q x q . The score and its difference operators are linear in f f f , and for every finite family ( Y q ) q = 0 m (Y_q)_{q=0}^{m} ( Y q ) q = 0 m , we have ∣ ∑ q = 0 m a q Y q ∣ ℓ ≤ C f , ℓ ∑ q = 0 m ∣ Y q ∣ ℓ \left|\sum_{q=0}^{m}a_qY_q\right|^\ell\le C_{f,\ell}\sum_{q=0}^{m}|Y_q|^\ell ∑ q = 0 m a q Y q ℓ ≤ C f , ℓ ∑ q = 0 m ∣ Y q ∣ ℓ . It is therefore enough to establish the estimates for a monomial. The constant term has zero score differences, and the linear term contributes nothing because the adjacency matrices have zero diagonal. We henceforth take f ( x ) = x m f(x)=x^m f ( x ) = x m with m ≥ 2 m\ge2 m ≥ 2 , for which the score in (27) is given by
g f ( x ~ 1 , G ) : = ∑ p = 2 m ∑ ( x ~ 2 , … , x ~ p ) ∈ ( Q ∖ { x ~ 1 } ) ( p − 1 ) 1 ( p − 1 ) ! ∑ π ∈ Σ m , p ∘ π ( 1 ) = 1 ∏ j = 1 m A ( x ~ π ( j ) , x ~ π ( j + 1 ) ; G ) . (28) g_f(\tilde{x}_1,G):=\sum_{p=2}^{m}\ \sum_{(\tilde{x}_2,\ldots,\tilde{x}_p)\in(\mathcal{Q}\setminus\{\tilde{x}_1\})^{(p-1)}}\frac{1}{(p-1)!}\sum_{\substack{\pi\in\Sigma^\circ_{m,p}\\\pi(1)=1}}\prod_{j=1}^{m}A(\tilde{x}_{\pi(j)},\tilde{x}_{\pi(j+1)};G). \tag*{(28)} g f ( x ~ 1 , G ) := p = 2 ∑ m ( x ~ 2 , … , x ~ p ) ∈ ( Q ∖ { x ~ 1 } ) ( p − 1 ) ∑ ( p − 1 )! 1 π ∈ Σ m , p ∘ π ( 1 ) = 1 ∑ j = 1 ∏ m A ( x ~ π ( j ) , x ~ π ( j + 1 ) ; G ) . (28)
We prove the four parts of the lemma separately.
Part (i). Since we have reduced to f ( x ) = x m f(x)=x^m f ( x ) = x m , the score g f ( x ~ 1 , G ) g_f(\tilde{x}_1,G) g f ( x ~ 1 , G ) counts rooted closed walks of length m m m starting and ending at x ~ 1 \tilde{x}_1 x ~ 1 . Consequently, its ℓ \ell ℓ -th power counts ordered ℓ \ell ℓ -tuples of such walks. Concatenating the ℓ \ell ℓ walks produces a closed walk of length ℓ m \ell m ℓ m starting and ending at x ~ 1 \tilde{x}_1 x ~ 1 . This map is injective, since the original walks are recovered by cutting the concatenated walk after every m m m steps. Hence, writing X ~ − 1 : = x ~ 1 \widetilde{X}_{-1}:=\tilde{x}_1 X − 1 := x ~ 1 , i 1 : = − 1 i_1:=-1 i 1 := − 1 and π ( ℓ m + 1 ) : = π ( 1 ) \pi(\ell m+1):=\pi(1) π ( ℓ m + 1 ) := π ( 1 ) ,
∣ g f ( x ~ 1 , G n , x ~ 1 ) ∣ ℓ ≤ ∑ p = 2 ℓ m ∑ ( X ~ i 2 , … , X ~ i p ) ∈ P ~ n ( p − 1 ) 1 ( p − 1 ) ! ∑ π ∈ Σ ℓ m , p ∘ π ( 1 ) = 1 ∏ j = 1 ℓ m A ( X ~ i π ( j ) , X ~ i π ( j + 1 ) ; G n , x ~ 1 ) . (29) |g_f(\tilde{x}_1,G_{n,\tilde{x}_1})|^\ell\le\sum_{p=2}^{\ell m}\ \sum_{(\widetilde{X}_{i_2},\ldots,\widetilde{X}_{i_p})\in\widetilde{\mathcal{P}}_n^{(p-1)}}\frac{1}{(p-1)!}\sum_{\substack{\pi\in\Sigma^\circ_{\ell m,p}\\\pi(1)=1}}\prod_{j=1}^{\ell m}A(\widetilde{X}_{i_{\pi(j)}},\widetilde{X}_{i_{\pi(j+1)}};G_{n,\tilde{x}_1}). \tag*{(29)} ∣ g f ( x ~ 1 , G n , x ~ 1 ) ∣ ℓ ≤ p = 2 ∑ ℓ m ( X i 2 , … , X i p ) ∈ P n ( p − 1 ) ∑ ( p − 1 )! 1 π ∈ Σ ℓ m , p ∘ π ( 1 ) = 1 ∑ j = 1 ∏ ℓ m A ( X i π ( j ) , X i π ( j + 1 ) ; G n , x ~ 1 ) . (29)
For π ∈ Σ ℓ m , p ∘ \pi\in\Sigma^\circ_{\ell m,p} π ∈ Σ ℓ m , p ∘ , consider the connected simple undirected graph formed by the distinct edges traversed by the corresponding closed walk. For each π \pi π , choose its spanning tree T π T_\pi T π . Since every entry of the adjacency matrix is either 0 or 1,
∏ j = 1 ℓ m A ( X ~ i π ( j ) , X ~ i π ( j + 1 ) ; G n , x ~ 1 ) ≤ ∏ { j , k } ∈ E ( T π ) A ( X ~ i j , X ~ i k ; G n , x ~ 1 ) , (30) \prod_{j=1}^{\ell m}A(\widetilde{X}_{i_{\pi(j)}},\widetilde{X}_{i_{\pi(j+1)}};G_{n,\tilde{x}_1})\le\prod_{\{j,k\}\in E(T_\pi)}A(\widetilde{X}_{i_j},\widetilde{X}_{i_k};G_{n,\tilde{x}_1}), \tag*{(30)} j = 1 ∏ ℓ m A ( X i π ( j ) , X i π ( j + 1 ) ; G n , x ~ 1 ) ≤ { j , k } ∈ E ( T π ) ∏ A ( X i j , X i k ; G n , x ~ 1 ) , (30)
where E ( T π ) E(T_\pi) E ( T π ) denotes the collection of (undirected) edges in T π T_\pi T π and { j , k } \{j,k\} { j , k } represents the (undirected) edge between vertices j j j and k k k . Thus, using (29) and (30) ,
∣ g f ( x ~ 1 , G n , x ~ 1 ) ∣ ℓ ≤ C ℓ , m ∑ p = 2 ℓ m ∑ ( X ~ i 2 , … , X ~ i p ) ∈ P ~ n ( p − 1 ) 1 ( p − 1 ) ! ∑ π ∈ Σ ℓ m , p ∘ π ( 1 ) = 1 ∏ { j , k } ∈ E ( T π ) A ( X ~ i j , X ~ i k ; G n , x ~ 1 ) . (31) |g_f(\tilde{x}_1,G_{n,\tilde{x}_1})|^\ell\le C_{\ell,m}\sum_{p=2}^{\ell m}\ \sum_{(\widetilde{X}_{i_2},\ldots,\widetilde{X}_{i_p})\in\widetilde{\mathcal{P}}_n^{(p-1)}}\frac{1}{(p-1)!}\sum_{\substack{\pi\in\Sigma^\circ_{\ell m,p}\\\pi(1)=1}}\prod_{\{j,k\}\in E(T_\pi)}A(\widetilde{X}_{i_j},\widetilde{X}_{i_k};G_{n,\tilde{x}_1}). \tag*{(31)} ∣ g f ( x ~ 1 , G n , x ~ 1 ) ∣ ℓ ≤ C ℓ , m p = 2 ∑ ℓ m ( X i 2 , … , X i p ) ∈ P n ( p − 1 ) ∑ ( p − 1 )! 1 π ∈ Σ ℓ m , p ∘ π ( 1 ) = 1 ∑ { j , k } ∈ E ( T π ) ∏ A ( X i j , X i k ; G n , x ~ 1 ) . (31)
For a fixed π ∈ Σ ℓ m , p ∘ \pi\in\Sigma^\circ_{\ell m,p} π ∈ Σ ℓ m , p ∘ , the multivariate Mecke formula gives, writing x ~ k = ( x k , u k ) \tilde{x}_k=(x_k,u_k) x ~ k = ( x k , u k ) for k ∈ { 2 , … , p } k\in\{2,\ldots,p\} k ∈ { 2 , … , p } ,
E [ ∑ ( X ~ i 2 , … , X ~ i p ) ∈ P ~ n ( p − 1 ) ∏ { j , k } ∈ E ( T π ) A ( X ~ i j , X ~ i k ; G n , x ~ 1 ) ] = ∫ W n p − 1 ∫ [ 1 , ∞ ) p − 1 E [ ∏ { j , k } ∈ E ( T π ) A ( x ~ j , x ~ k ; G n , x ~ p ) ] ∏ k = 2 p d x k ∏ k = 2 p μ ( d u k ) . (32) \mathbb{E}\left[\sum_{(\widetilde{X}_{i_2},\ldots,\widetilde{X}_{i_p})\in\widetilde{\mathcal{P}}_n^{(p-1)}}\prod_{\{j,k\}\in E(T_\pi)}A(\widetilde{X}_{i_j},\widetilde{X}_{i_k};G_{n,\tilde{x}_1})\right]
=\int_{W_n^{p-1}}\int_{[1,\infty)^{p-1}}\mathbb{E}\left[\prod_{\{j,k\}\in E(T_\pi)}A(\tilde{x}_j,\tilde{x}_k;G_{n,\tilde{x}_p})\right]\prod_{k=2}^{p}\mathrm{d}x_k\prod_{k=2}^{p}\mu(\mathrm{d}u_k). \tag*{(32)} E ( X i 2 , … , X i p ) ∈ P n ( p − 1 ) ∑ { j , k } ∈ E ( T π ) ∏ A ( X i j , X i k ; G n , x ~ 1 ) = ∫ W n p − 1 ∫ [ 1 , ∞ ) p − 1 E { j , k } ∈ E ( T π ) ∏ A ( x ~ j , x ~ k ; G n , x ~ p ) k = 2 ∏ p d x k k = 2 ∏ p μ ( d u k ) . (32)
Recalling the connection rule, this is equal to
∫ W n p − 1 ∫ [ 1 , ∞ ) p − 1 ∏ { j , k } ∈ E ( T π ) k < j φ ( ∥ x j − x k ∥ d κ ( u j , u k ) ) ∏ k = 2 p d x k ∏ k = 2 p μ ( d u k ) . (33) \int_{W_n^{p-1}}\int_{[1,\infty)^{p-1}}\prod_{\substack{\{j,k\}\in E(T_\pi)\\k<j}}\varphi\left(\frac{\lVert x_j-x_k\rVert^d}{\kappa(u_j,u_k)}\right)\prod_{k=2}^{p}\mathrm{d}x_k\prod_{k=2}^{p}\mu(\mathrm{d}u_k). \tag*{(33)} ∫ W n p − 1 ∫ [ 1 , ∞ ) p − 1 { j , k } ∈ E ( T π ) k < j ∏ φ ( κ ( u j , u k ) ∥ x j − x k ∥ d ) k = 2 ∏ p d x k k = 2 ∏ p μ ( d u k ) . (33)
The argument now proceeds by integrating over leaf vertices of the tree T π T_\pi T π . Lemma 6 , applied conditionally on the marks, bounds the spatial integral by M φ p − 1 c φ p − 1 u 1 d 1 ∏ i = 2 p u i d i M_\varphi^{p-1}c_\varphi^{p-1}u_1^{d_1}\prod_{i=2}^{p}u_i^{d_i} M φ p − 1 c φ p − 1 u 1 d 1 ∏ i = 2 p u i d i , where d i d_i d i is the degree of vertex i i i in T π T_\pi T π . Since d i ≤ p − 1 ≤ ℓ m − 1 d_i\le p-1\le\ell m-1 d i ≤ p − 1 ≤ ℓ m − 1 , direct integration of the mark powers gives
E [ ∑ ( X ~ i 2 , … , X ~ i p ) ∈ P ~ n ( p − 1 ) ∏ { j , k } ∈ E ( T π ) A ( X ~ i j , X ~ i k ; G n , x ~ 1 ) ] ≤ C M φ p − 1 u 1 ℓ m − 1 , \mathbb{E}\left[\sum_{(\widetilde{X}_{i_2},\ldots,\widetilde{X}_{i_p})\in\widetilde{\mathcal{P}}_n^{(p-1)}}\prod_{\{j,k\}\in E(T_\pi)}A(\widetilde{X}_{i_j},\widetilde{X}_{i_k};G_{n,\tilde{x}_1})\right]\le C M_\varphi^{p-1}u_1^{\ell m-1}, E ( X i 2 , … , X i p ) ∈ P n ( p − 1 ) ∑ { j , k } ∈ E ( T π ) ∏ A ( X i j , X i k ; G n , x ~ 1 ) ≤ C M φ p − 1 u 1 ℓ m − 1 ,
provided that E [ U ℓ m − 1 ] < ∞ \mathbb{E}[U^{\ell m-1}]<\infty E [ U ℓ m − 1 ] < ∞ . Returning to (31) , take expectations and sum this bound over the finitely many values of p p p and π \pi π :
E ∣ g f ( x ~ 1 , G n , x ~ 1 ) ∣ ℓ ≤ C u 1 ℓ m − 1 ∑ p = 2 ℓ m ∣ Σ ℓ m , p ∘ ∣ ( p − 1 ) ! M φ p − 1 ≤ C ′ u 1 ℓ m − 1 M φ ℓ m − 1 . \mathbb{E}|g_f(\tilde{x}_1,G_{n,\tilde{x}_1})|^\ell\le C u_1^{\ell m-1}\sum_{p=2}^{\ell m}\frac{|\Sigma^\circ_{\ell m,p}|}{(p-1)!}M_\varphi^{p-1}\le C' u_1^{\ell m-1}M_\varphi^{\ell m-1}. E ∣ g f ( x ~ 1 , G n , x ~ 1 ) ∣ ℓ ≤ C u 1 ℓ m − 1 p = 2 ∑ ℓ m ( p − 1 )! ∣ Σ ℓ m , p ∘ ∣ M φ p − 1 ≤ C ′ u 1 ℓ m − 1 M φ ℓ m − 1 .
This proves Part (i).
Part (ii). Fix X ~ i ∈ P ~ n \widetilde{X}_i \in\widetilde{P}_n X i ∈ P n and set i 1 : = i i_1 := i i 1 := i . Recall that X ~ − 1 : = x ~ 1 \widetilde{X}_{-1} := \widetilde{x}_1 X − 1 := x 1 is the inserted vertex. Write G n : = G ( ξ n ) G_n := G(\xi_n) G n := G ( ξ n ) and P ~ n , x ~ 1 : = P ~ n ∪ { x ~ 1 } \widetilde{P}_{n,\widetilde{x}_1} := \widetilde{P}_n \cup\{\widetilde{x}_1\} P n , x 1 := P n ∪ { x 1 } . By the definition of the score deletion difference,
∂ x ~ 1 g f ( X ~ i , G n , x ~ 1 ) = g f ( X ~ i , G n , x ~ 1 ) − g f ( X ~ i , G n ) . \partial_{\widetilde{x}_1} g_f(\widetilde{X}_i,G_{n,\widetilde{x}_1}) = g_f(\widetilde{X}_i,G_{n,\widetilde{x}_1}) - g_f(\widetilde{X}_i,G_n). ∂ x 1 g f ( X i , G n , x 1 ) = g f ( X i , G n , x 1 ) − g f ( X i , G n ) .
Using (28) with π ( m + 1 ) = π ( 1 ) \pi(m+1)=\pi(1) π ( m + 1 ) = π ( 1 ) , we have
g f ( X ~ i , G n ) = ∑ p = 2 m ∑ ( X ~ i 2 , … , X ~ i p ) ∈ ( P ~ n ∖ { X ~ i } ) ( p − 1 ) 1 ( p − 1 ) ! ∑ π ∈ Σ m , p π ( 1 ) = 1 ∏ j = 1 m A ( X ~ i π ( j ) , X ~ i π ( j + 1 ) ; G n ) . g_f(\widetilde{X}_i,G_n)=\sum_{p=2}^{m}\ \sum_{(\widetilde{X}_{i_2},\ldots,\widetilde{X}_{i_p})\in(\widetilde{P}_n\setminus\{\widetilde{X}_i\})^{(p-1)}}\frac{1}{(p-1)!}\sum_{\substack{\pi\in\Sigma_{m,p}\\\pi(1)=1}}\prod_{j=1}^{m}A(\widetilde{X}_{i_{\pi(j)}},\widetilde{X}_{i_{\pi(j+1)}};G_n). g f ( X i , G n ) = p = 2 ∑ m ( X i 2 , … , X i p ) ∈ ( P n ∖ { X i } ) ( p − 1 ) ∑ ( p − 1 )! 1 π ∈ Σ m , p π ( 1 ) = 1 ∑ j = 1 ∏ m A ( X i π ( j ) , X i π ( j + 1 ) ; G n ) .
and
g f ( X ~ i , G n , x ~ 1 ) = ∑ p = 2 m ∑ ( X ~ i 2 , … , X ~ i p ) ∈ ( P ~ n , x ~ 1 ∖ { X ~ i } ) ( p − 1 ) 1 ( p − 1 ) ! ∑ π ∈ Σ m , p π ( 1 ) = 1 ∏ j = 1 m A ( X ~ i π ( j ) , X ~ i π ( j + 1 ) ; G n , x ~ 1 ) . g_f(\widetilde{X}_i,G_{n,\widetilde{x}_1})=\sum_{p=2}^{m}\ \sum_{(\widetilde{X}_{i_2},\ldots,\widetilde{X}_{i_p})\in(\widetilde{P}_{n,\widetilde{x}_1}\setminus\{\widetilde{X}_i\})^{(p-1)}}\frac{1}{(p-1)!}\sum_{\substack{\pi\in\Sigma_{m,p}\\\pi(1)=1}}\prod_{j=1}^{m}A(\widetilde{X}_{i_{\pi(j)}},\widetilde{X}_{i_{\pi(j+1)}};G_{n,\widetilde{x}_1}). g f ( X i , G n , x 1 ) = p = 2 ∑ m ( X i 2 , … , X i p ) ∈ ( P n , x 1 ∖ { X i } ) ( p − 1 ) ∑ ( p − 1 )! 1 π ∈ Σ m , p π ( 1 ) = 1 ∑ j = 1 ∏ m A ( X i π ( j ) , X i π ( j + 1 ) ; G n , x 1 ) .
The induced subgraph of G n , x ~ 1 G_{n,\widetilde{x}_1} G n , x 1 on the old vertices is exactly G n G_n G n , with the same edge indicators. Consequently, every term whose auxiliary tuple ( X ~ i 2 , … , X ~ i p ) (\widetilde{X}_{i_2},\ldots,\widetilde{X}_{i_p}) ( X i 2 , … , X i p ) avoids x ~ 1 \widetilde{x}_1 x 1 cancels between these two expansions. The remaining terms are precisely those whose auxiliary tuple contains x ~ 1 \widetilde{x}_1 x 1 . They are nonnegative and count closed walks based at X ~ i \widetilde{X}_i X i that visit the inserted vertex. Therefore,
∂ x ~ 1 g f ( X ~ i , G n , x ~ 1 ) = ∑ p = 2 m ∑ ( X ~ i 2 , … , X ~ i p ) ∈ ( P ~ n , x ~ 1 ∖ { X ~ i } ) ( p − 1 ) x ~ 1 ∈ { X ~ i 2 , … , X ~ i p } 1 ( p − 1 ) ! ∑ π ∈ Σ m , p π ( 1 ) = 1 ∏ j = 1 m A ( X ~ i π ( j ) , X ~ i π ( j + 1 ) ; G n , x ~ 1 ) . \partial_{\widetilde{x}_1}g_f(\widetilde{X}_i,G_{n,\widetilde{x}_1})=\sum_{p=2}^{m}\ \sum_{\substack{(\widetilde{X}_{i_2},\ldots,\widetilde{X}_{i_p})\in(\widetilde{P}_{n,\widetilde{x}_1}\setminus\{\widetilde{X}_i\})^{(p-1)}\\\widetilde{x}_1\in\{\widetilde{X}_{i_2},\ldots,\widetilde{X}_{i_p}\}}}\frac{1}{(p-1)!}\sum_{\substack{\pi\in\Sigma_{m,p}\\\pi(1)=1}}\prod_{j=1}^{m}A(\widetilde{X}_{i_{\pi(j)}},\widetilde{X}_{i_{\pi(j+1)}};G_{n,\widetilde{x}_1}). ∂ x 1 g f ( X i , G n , x 1 ) = p = 2 ∑ m ( X i 2 , … , X i p ) ∈ ( P n , x 1 ∖ { X i } ) ( p − 1 ) x 1 ∈ { X i 2 , … , X i p } ∑ ( p − 1 )! 1 π ∈ Σ m , p π ( 1 ) = 1 ∑ j = 1 ∏ m A ( X i π ( j ) , X i π ( j + 1 ) ; G n , x 1 ) .
Thus,
( ∑ X ~ i ∈ P ~ n ∂ x ~ 1 g f ( X ~ i , G n , x ~ 1 ) ) ℓ = ∏ h = 1 ℓ ( ∑ p h = 2 m ∑ ( X ~ i h , 1 , … , X ~ i h , p h ) ∈ P ~ n , x ~ 1 ( p h ) 1 ( p h − 1 ) ! ∑ π h ∈ Σ m , p h π h ( 1 ) = 1 ∏ j = 1 m A ( X ~ i h , π h ( j ) , X ~ i h , π h ( j + 1 ) ; G n , x ~ 1 ) ) . \left(\sum_{\widetilde{X}_i\in\widetilde{P}_n}\partial_{\widetilde{x}_1}g_f(\widetilde{X}_i,G_{n,\widetilde{x}_1})\right)^\ell
=
\prod_{h=1}^{\ell}\left(\sum_{p_h=2}^{m}\ \sum_{(\widetilde{X}_{i_{h,1}},\ldots,\widetilde{X}_{i_{h,p_h}})\in\widetilde{P}_{n,\widetilde{x}_1}^{(p_h)}}\frac{1}{(p_h-1)!}\sum_{\substack{\pi_h\in\Sigma_{m,p_h}\\\pi_h(1)=1}}\prod_{j=1}^{m}A(\widetilde{X}_{i_{h,\pi_h(j)}},\widetilde{X}_{i_{h,\pi_h(j+1)}};G_{n,\widetilde{x}_1})\right). X i ∈ P n ∑ ∂ x 1 g f ( X i , G n , x 1 ) ℓ = h = 1 ∏ ℓ p h = 2 ∑ m ( X i h , 1 , … , X i h , p h ) ∈ P n , x 1 ( p h ) ∑ ( p h − 1 )! 1 π h ∈ Σ m , p h π h ( 1 ) = 1 ∑ j = 1 ∏ m A ( X i h , π h ( j ) , X i h , π h ( j + 1 ) ; G n , x 1 ) .
Indeed, summing over X ~ i ∈ P ~ n \widetilde{X}_i\in\widetilde{P}_n X i ∈ P n amounts to including the starting vertex as the first coordinate X ~ i h , 1 \widetilde{X}_{i_{h,1}} X i h , 1 of each ordered p h p_h p h -tuple. Thus, the summation over X ~ i \widetilde{X}_i X i is incorporated into the summation over P ~ n , x ~ 1 ( p h ) \widetilde{P}_{n,\widetilde{x}_1}^{(p_h)} P n , x 1 ( p h ) . For each h h h , there is a unique position s h ∈ { 2 , … , p h } s_h\in\{2,\ldots,p_h\} s h ∈ { 2 , … , p h } occupied by x ~ 1 \widetilde{x}_1 x 1 . Separating the sum according to this position rewrites the preceding right-hand side as
∏ h = 1 ℓ [ ∑ p h = 2 m ∑ s h = 2 p h 1 ( p h − 1 ) ! ∑ π h ∈ Σ m , p h π h ( 1 ) = 1 ∑ ( X ~ i h , k ) k ∈ [ p h ] ∖ { s h } ∈ P ~ n ( p h − 1 ) ∏ j = 1 m A ( X ^ h , π h ( j ) , X ^ h , π h ( j + 1 ) ; G n , x ~ 1 ) ] , (34) \prod_{h=1}^{\ell}\left[\sum_{p_h=2}^{m}\sum_{s_h=2}^{p_h}\frac{1}{(p_h-1)!}\sum_{\substack{\pi_h\in\Sigma_{m,p_h}\\\pi_h(1)=1}}\sum_{(\widetilde{X}_{i_{h,k}})_{k\in[p_h]\setminus\{s_h\}}\in\widetilde{P}_n^{(p_h-1)}}\prod_{j=1}^{m}A(\widehat{X}_{h,\pi_h(j)},\widehat{X}_{h,\pi_h(j+1)};G_{n,\widetilde{x}_1})\right], \tag*{(34)} h = 1 ∏ ℓ p h = 2 ∑ m s h = 2 ∑ p h ( p h − 1 )! 1 π h ∈ Σ m , p h π h ( 1 ) = 1 ∑ ( X i h , k ) k ∈ [ p h ] ∖ { s h } ∈ P n ( p h − 1 ) ∑ j = 1 ∏ m A ( X h , π h ( j ) , X h , π h ( j + 1 ) ; G n , x 1 ) , (34)
where [ p h ] : = { 1 , … , p h } [p_h]:=\{1,\ldots,p_h\} [ p h ] := { 1 , … , p h } , π h ( m + 1 ) : = π h ( 1 ) \pi_h(m+1):=\pi_h(1) π h ( m + 1 ) := π h ( 1 ) , and
X ^ h , j : = { X ~ i h , j , j ≠ s h , X ~ − 1 = x ~ 1 , j = s h . \widehat{X}_{h,j}:=
\begin{cases}
\widetilde{X}_{i_{h,j}}, & j\ne s_h,\\
\widetilde{X}_{-1}=\widetilde{x}_1, & j=s_h.
\end{cases} X h , j := { X i h , j , X − 1 = x 1 , j = s h , j = s h .
The tuple ( X ~ i h , k ) k ∈ [ p h ] ∖ { s h } (\widetilde{X}_{i_{h,k}})_{k\in[p_h]\setminus\{s_h\}} ( X i h , k ) k ∈ [ p h ] ∖ { s h } is ordered by increasing k ∈ [ p h ] ∖ { s h } k\in[p_h]\setminus\{s_h\} k ∈ [ p h ] ∖ { s h } . For fixed p h , s h , π h p_h,s_h,\pi_h p h , s h , π h , write
g h ( ( z ~ h , k ) k ∈ [ p h ] ∖ { s h } ; G ) : = ∏ j = 1 m A ( z ^ h , π h ( j ) , z ^ h , π h ( j + 1 ) ; G ) , g_h\left((\widetilde{z}_{h,k})_{k\in[p_h]\setminus\{s_h\}};G\right):=\prod_{j=1}^{m}A(\widehat{z}_{h,\pi_h(j)},\widehat{z}_{h,\pi_h(j+1)};G), g h ( ( z h , k ) k ∈ [ p h ] ∖ { s h } ; G ) := j = 1 ∏ m A ( z h , π h ( j ) , z h , π h ( j + 1 ) ; G ) ,
where z ^ h , j : = z ~ h , j \widehat{z}_{h,j}:=\widetilde{z}_{h,j} z h , j := z h , j for j ≠ s h j\ne s_h j = s h and z ^ h , j : = x ~ 1 \widehat{z}_{h,j}:=\widetilde{x}_1 z h , j := x 1 for j = s h j=s_h j = s h . Consider ℓ \ell ℓ rows, where row h h h corresponds to the h h h th walk and contains p h − 1 p_h-1 p h − 1 Poisson vertices, indexed by ( h , j ) (h,j) ( h , j ) with j ∈ [ p h ] ∖ { s h } j \in[p_h] \setminus\{s_h\} j ∈ [ p h ] ∖ { s h } . When expanding the product in (34) , the Poisson vertices within each row are distinct, but those from different rows may coincide. We encode these coincidences by partitions of
J : = { ( h , j ) : 1 ≤ h ≤ ℓ , j ∈ [ p h ] ∖ { s h } } . \mathcal{J}:=\{(h,j):1\le h\le\ell,\ j\in[p_h]\setminus\{s_h\}\}. J := {( h , j ) : 1 ≤ h ≤ ℓ , j ∈ [ p h ] ∖ { s h }} .
Let Π ( p 1 − 1 , … , p ℓ − 1 ) \Pi(p_1-1,\ldots,p_\ell-1) Π ( p 1 − 1 , … , p ℓ − 1 ) denote the set of partitions σ \sigma σ of J \mathcal{J} J such that each block contains at most one element from each row. Each block represents one distinct Poisson vertex. Write ∣ σ ∣ |\sigma| ∣ σ ∣ for the number of blocks, order them by their lexicographically smallest elements, and let b σ ( h , j ) ∈ { 1 , … , ∣ σ ∣ } b_\sigma(h,j)\in\{1,\ldots,|\sigma|\} b σ ( h , j ) ∈ { 1 , … , ∣ σ ∣ } be the index of the block containing ( h , j ) (h,j) ( h , j ) . For y ˇ ∣ σ ∣ = ( y ˇ 1 , … , y ˇ ∣ σ ∣ ) \check{\mathbf{y}}_{|\sigma|}=(\check{y}_1,\ldots,\check{y}_{|\sigma|}) y ˇ ∣ σ ∣ = ( y ˇ 1 , … , y ˇ ∣ σ ∣ ) , set
( ⨂ h = 1 ℓ g h ) σ ( y ˇ ∣ σ ∣ ; G ) : = ∏ h = 1 ℓ g h ( ( y ˇ b σ ( h , j ) ) j ∈ [ p h ] ∖ { s h } ; G ) . (35) \left(\bigotimes_{h=1}^{\ell}g_h\right)_{\sigma}(\check{\mathbf{y}}_{|\sigma|};G):=\prod_{h=1}^{\ell}g_h\left((\check{y}_{b_\sigma(h,j)})_{j\in[p_h]\setminus\{s_h\}};G\right). \tag*{(35)} ( h = 1 ⨂ ℓ g h ) σ ( y ˇ ∣ σ ∣ ; G ) := h = 1 ∏ ℓ g h ( ( y ˇ b σ ( h , j ) ) j ∈ [ p h ] ∖ { s h } ; G ) . (35)
Also define G σ , n : = G { 1 , … , ∣ σ ∣ + 1 } ( ξ ˇ n , ( x ˇ 1 , y ˇ 1 , … , y ˇ ∣ σ ∣ ) ) G_{\sigma,n}:=G_{\{1,\ldots,|\sigma|+1\}}(\check{\xi}_n,(\check{x}_1,\check{y}_1,\ldots,\check{y}_{|\sigma|})) G σ , n := G { 1 , … , ∣ σ ∣ + 1 } ( ξ ˇ n , ( x ˇ 1 , y ˇ 1 , … , y ˇ ∣ σ ∣ )) . By multivariate Mecke formula [6 Theorem 3.1] (y ˇ 1 , … , y ˇ ∣ σ ∣ \check{y}_1,\ldots,\check{y}_{|\sigma|} y ˇ 1 , … , y ˇ ∣ σ ∣ are integration variables), the expectation of (34) is given by
∑ ( p h , s h , π h ) h = 1 ℓ : 2 ≤ p h ≤ m , 2 ≤ s h ≤ p h , π h ∈ Σ m , p h , π h ( 1 ) = 1 for all h ( ∏ h = 1 ℓ 1 ( p h − 1 ) ! ) ∑ σ ∈ Π ( p 1 − 1 , … , p ℓ − 1 ) ∫ W ∣ σ ∣ ∣ σ ∣ E [ ( ⨂ h = 1 ℓ g h ) σ ( y ˇ ∣ σ ∣ ; G σ , n ) ] ∏ a = 1 ∣ σ ∣ λ ( d y ˇ a ) . \sum_{\substack{(p_h,s_h,\pi_h)_{h=1}^{\ell}:\\2\le p_h\le m,\ 2\le s_h\le p_h,\\\pi_h\in\Sigma_{m,p_h},\ \pi_h(1)=1\ \text{for all }h}}
\left(\prod_{h=1}^{\ell}\frac{1}{(p_h-1)!}\right)
\sum_{\sigma\in\Pi(p_1-1,\ldots,p_\ell-1)}
\int_{W_{|\sigma|}^{|\sigma|}}
\mathbb{E}\left[\left(\bigotimes_{h=1}^{\ell}g_h\right)_{\sigma}(\check{\mathbf{y}}_{|\sigma|};G_{\sigma,n})\right]\prod_{a=1}^{|\sigma|}\lambda(\mathrm{d}\check{y}_a). ( p h , s h , π h ) h = 1 ℓ : 2 ≤ p h ≤ m , 2 ≤ s h ≤ p h , π h ∈ Σ m , p h , π h ( 1 ) = 1 for all h ∑ ( h = 1 ∏ ℓ ( p h − 1 )! 1 ) σ ∈ Π ( p 1 − 1 , … , p ℓ − 1 ) ∑ ∫ W ∣ σ ∣ ∣ σ ∣ E [ ( h = 1 ⨂ ℓ g h ) σ ( y ˇ ∣ σ ∣ ; G σ , n ) ] a = 1 ∏ ∣ σ ∣ λ ( d y ˇ a ) .
(recall λ ( d y ˇ ) = d y μ ( d v ) \lambda(\mathrm{d}\check{y})=\mathrm{d}y\,\mu(\mathrm{d}v) λ ( d y ˇ ) = d y μ ( d v ) ). Using the block ordering defined above, write w ˇ 1 : = x ˇ 1 \check{w}_1:=\check{x}_1 w ˇ 1 := x ˇ 1 and w ˇ a + 1 : = y ˇ a \check{w}_{a+1}:=\check{y}_a w ˇ a + 1 := y ˇ a (1 ≤ a ≤ ∣ σ ∣ 1\le a\le|\sigma| 1 ≤ a ≤ ∣ σ ∣ ). Let H σ H_\sigma H σ be the (undirected) graph on { 1 , … , ∣ σ ∣ + 1 } \{1,\ldots,|\sigma|+1\} { 1 , … , ∣ σ ∣ + 1 } whose edges are the distinct edges traversed by the rows after the identifications prescribed by σ \sigma σ . Choose its spanning tree T σ T_\sigma T σ . Then,
( ⨂ h = 1 ℓ g h ) σ ( y ˇ ∣ σ ∣ ; G σ , n ) ≤ ∏ { j , k } ∈ E ( T σ ) A ( w ˇ j , w ˇ k ; G σ , n ) . \left(\bigotimes_{h=1}^{\ell}g_h\right)_{\sigma}(\check{\mathbf{y}}_{|\sigma|};G_{\sigma,n})
\le\prod_{\{j,k\}\in E(T_\sigma)}A(\check{w}_j,\check{w}_k;G_{\sigma,n}). ( h = 1 ⨂ ℓ g h ) σ ( y ˇ ∣ σ ∣ ; G σ , n ) ≤ { j , k } ∈ E ( T σ ) ∏ A ( w ˇ j , w ˇ k ; G σ , n ) .
The tree T σ T_\sigma T σ has ∣ σ ∣ |\sigma| ∣ σ ∣ edges. Let d j d_j d j be the degree of vertex j j j in T σ T_\sigma T σ . By conditionally averaging the distinct edge indicators and applying Lemma 6 , we obtain
∫ W ∣ σ ∣ ∣ σ ∣ E [ ( ⨂ h = 1 ℓ g h ) σ ( y ˇ ∣ σ ∣ ; G σ , n ) ] ∏ a = 1 ∣ σ ∣ λ ( d y ˇ a ) ≤ c κ ∣ σ ∣ M φ ∣ σ ∣ u 1 d 1 ∏ j = 2 ∣ σ ∣ + 1 E [ U d j ] . \int_{W_{|\sigma|}^{|\sigma|}}\mathbb{E}\left[\left(\bigotimes_{h=1}^{\ell}g_h\right)_{\sigma}(\check{\mathbf{y}}_{|\sigma|};G_{\sigma,n})\right]\prod_{a=1}^{|\sigma|}\lambda(\mathrm{d}\check{y}_a)
\le c_\kappa^{|\sigma|}M_\varphi^{|\sigma|}u_1^{d_1}\prod_{j=2}^{|\sigma|+1}\mathbb{E}[U^{d_j}]. ∫ W ∣ σ ∣ ∣ σ ∣ E [ ( h = 1 ⨂ ℓ g h ) σ ( y ˇ ∣ σ ∣ ; G σ , n ) ] a = 1 ∏ ∣ σ ∣ λ ( d y ˇ a ) ≤ c κ ∣ σ ∣ M φ ∣ σ ∣ u 1 d 1 j = 2 ∏ ∣ σ ∣ + 1 E [ U d j ] .
The root position x 1 x_1 x 1 and root mark u 1 u_1 u 1 remain fixed. Since ∣ σ ∣ ≤ ∑ h = 1 ℓ ( p h − 1 ) ≤ ℓ ( m − 1 ) ≤ ℓ m − 1 |\sigma|\le\sum_{h=1}^{\ell}(p_h-1)\le\ell(m-1)\le\ell m-1 ∣ σ ∣ ≤ ∑ h = 1 ℓ ( p h − 1 ) ≤ ℓ ( m − 1 ) ≤ ℓ m − 1 and d j ≤ ∣ σ ∣ d_j\le|\sigma| d j ≤ ∣ σ ∣ , the moment assumption implies the bound above is at most C ℓ , m u 1 ℓ m − 1 M φ ℓ m − 1 C_{\ell,m}u_1^{\ell m-1}M_\varphi^{\ell m-1} C ℓ , m u 1 ℓ m − 1 M φ ℓ m − 1 . Summing these bounds in the preceding diagram formula proves Part (ii).
Part (iii). Set L : = ℓ m L:=\ell m L := ℓ m . For f ( x ) = x m f(x)=x^m f ( x ) = x m , the quantity ∂ x ˇ 1 g f ( x ˇ 2 , G n , x ˇ 2 ) \partial_{\check{x}_1}g_f(\check{x}_2,G_{n,\check{x}_2}) ∂ x ˇ 1 g f ( x ˇ 2 , G n , x ˇ 2 ) counts length-m m m closed walks based at x 2 x_2 x 2 that visit x 1 x_1 x 1 . It is nonnegative, and its ℓ \ell ℓ th power counts ordered collections of ℓ \ell ℓ such walks. Concatenating these walks at x 2 x_2 x 2 produces a closed walk of length L L L visiting both x 1 x_1 x 1 and x 2 x_2 x 2 . Let p p p denote the number of distinct vertices visited by this concatenated walk. Then 2 ≤ p ≤ L 2\le p\le L 2 ≤ p ≤ L . Label these vertices by { 1 , … , p } \{1,\ldots,p\} { 1 , … , p } , assigning labels 1 1 1 and 2 2 2 to x 1 x_1 x 1 and x 2 x_2 x 2 , respectively. Let H H H be the simple (undirected) graph on { 1 , … , p } \{1,\ldots,p\} { 1 , … , p } whose edges are traversed by the walk.
Apply the multivariate Mecke formula to the p − 2 p-2 p − 2 vertices labeled 3 , … , p 3,\ldots,p 3 , … , p , which belong to the original Poisson process, while keeping the two distinguished vertices x ˇ 1 \check{x}_1 x ˇ 1 and x ˇ 2 \check{x}_2 x ˇ 2 fixed. Then average the independent edges. After also integrating the two inserted marks, the contribution of this pattern is
∫ [ 1 , ∞ ) p ∫ W p − 2 ∏ { i , j } ∈ E ( H ) φ ( ∥ z i − z j ∥ d κ ( u i , u j ) ) d z 3 ⋯ d z p ∏ v = 1 p μ ( d u v ) , \int_{[1,\infty)^p}\int_{W^{p-2}}
\prod_{\{i,j\}\in E(H)}
\varphi\left(\frac{\lVert z_i-z_j\rVert^d}{\kappa(u_i,u_j)}\right)
\,\mathrm{d}z_3\cdots\mathrm{d}z_p
\prod_{v=1}^{p}\mu(\mathrm{d}u_v), ∫ [ 1 , ∞ ) p ∫ W p − 2 { i , j } ∈ E ( H ) ∏ φ ( κ ( u i , u j ) ∥ z i − z j ∥ d ) d z 3 ⋯ d z p v = 1 ∏ p μ ( d u v ) ,
where z 1 = x 1 z_1=x_1 z 1 = x 1 and z 2 = x 2 z_2=x_2 z 2 = x 2 are fixed. There are finitely many patterns, depending only on ℓ , m \ell,m ℓ , m , so it suffices to bound each such integral. Fix a spanning tree T T T of the graph H H H . Write
p { i , j } : = φ ( ∥ z i − z j ∥ d κ ( u i , u j ) ) . p_{\{i,j\}}:=\varphi\left(\frac{\lVert z_i-z_j\rVert^d}{\kappa(u_i,u_j)}\right). p { i , j } := φ ( κ ( u i , u j ) ∥ z i − z j ∥ d ) .
@արաբաղ Since these factors lie in [ 0 , 1 ] [0,1] [ 0 , 1 ] , the product over E ( H ) E(H) E ( H ) is bounded by the product over E ( T ) E(T) E ( T ) .
Let P T ( x 1 , x 2 ) P_T(x_1,x_2) P T ( x 1 , x 2 ) be the unique path joining x 1 x_1 x 1 and x 2 x_2 x 2 . For every spatial configuration, the triangle inequality therefore guarantees at least one path edge of length at least ∥ x 1 − x 2 ∥ / L \lVert x_1-x_2\rVert/L ∥ x 1 − x 2 ∥ / L . Thus
∏ h ∈ E ( T ) p h ≤ ∑ e = { a , b } ∈ P T ( x 1 , x 2 ) 1 { ∥ z a − z b ∥ ≥ ∥ x 1 − x 2 ∥ / L } ∏ h ∈ E ( T ) p h . \prod_{h\in E(T)}p_h \le\sum_{e=\{a,b\}\in P_T(x_1,x_2)}1\{\lVert z_a-z_b\rVert\ge\lVert x_1-x_2\rVert/L\}\prod_{h\in E(T)}p_h. h ∈ E ( T ) ∏ p h ≤ e = { a , b } ∈ P T ( x 1 , x 2 ) ∑ 1 {∥ z a − z b ∥ ≥ ∥ x 1 − x 2 ∥ / L } h ∈ E ( T ) ∏ p h .
For each fixed summand, monotonicity of φ \varphi φ gives 1 { ∥ z a − z b ∥ ≥ ∥ x 1 − x 2 ∥ / L } p e ≤ φ ( ∥ x 1 − x 2 ∥ d L d κ ( u a , u b ) ) 1\{\lVert z_a-z_b\rVert\ge\lVert x_1-x_2\rVert/L\}p_e\le\varphi\left(\frac{\lVert x_1-x_2\rVert^d}{L^d\kappa(u_a,u_b)}\right) 1 {∥ z a − z b ∥ ≥ ∥ x 1 − x 2 ∥ / L } p e ≤ φ ( L d κ ( u a , u b ) ∥ x 1 − x 2 ∥ d ) . Consequently,
∏ h ∈ E ( T ) p h ≤ ∑ e = { a , b } ∈ P T ( x 1 , x 2 ) φ ( ∥ x 1 − x 2 ∥ d L d κ ( u a , u b ) ) ∏ h ∈ E ( T ) ∖ { e } p h . \prod_{h\in E(T)}p_h\le\sum_{e=\{a,b\}\in P_T(x_1,x_2)}\varphi\left(\frac{\lVert x_1-x_2\rVert^d}{L^d\kappa(u_a,u_b)}\right)\prod_{h\in E(T)\setminus\{e\}}p_h. h ∈ E ( T ) ∏ p h ≤ e = { a , b } ∈ P T ( x 1 , x 2 ) ∑ φ ( L d κ ( u a , u b ) ∥ x 1 − x 2 ∥ d ) h ∈ E ( T ) ∖ { e } ∏ p h .
Fix an edge e = { a , b } e=\{a,b\} e = { a , b } in this sum. Its deletion leaves two trees, each containing exactly one fixed root. After enlarging the spatial domain to ( R d ) p − 2 (\mathbb{R}^d)^{p-2} ( R d ) p − 2 , integrate all nonroot vertices successively from the leaves of these two trees. Lemma 6 gives the bound
κ p − 2 M φ p − 2 φ ( ∥ x 1 − x 2 ∥ d L d κ ( u a , u b ) ) ∏ v = 1 p d v ′ , \kappa^{p-2}M_\varphi^{p-2}\varphi\left(\frac{\lVert x_1-x_2\rVert^d}{L^d\kappa(u_a,u_b)}\right)\prod_{v=1}^{p}d'_v, κ p − 2 M φ p − 2 φ ( L d κ ( u a , u b ) ∥ x 1 − x 2 ∥ d ) v = 1 ∏ p d v ′ ,
where d v ′ d'_v d v ′ is the degree in the forest T ∖ { e } T\setminus\{e\} T ∖ { e } . The endpoint labels a , b a,b a , b and the forest are now fixed within this summand. Independence of marks gives
∫ [ 1 , ∞ ) p φ ( ∥ x 1 − x 2 ∥ d L d κ ( u a , u b ) ) ∏ v = 1 p u v d v ′ ∏ v = 1 p μ ( d u v ) = ( ∏ v ∉ { a , b } E [ U d v ′ ] ) E [ φ ( ∥ x 1 − x 2 ∥ d L d κ ( U , V ) ) U d a ′ V d b ′ ] , \int_{[1,\infty)^p}\varphi\left(\frac{\lVert x_1-x_2\rVert^d}{L^d\kappa(u_a,u_b)}\right)\prod_{v=1}^{p}u_v^{d'_v}\prod_{v=1}^{p}\mu(du_v)
=
\left(\prod_{v\notin\{a,b\}}\mathbb{E}[U^{d'_v}]\right)\mathbb{E}\left[\varphi\left(\frac{\lVert x_1-x_2\rVert^d}{L^d\kappa(U,V)}\right)U^{d'_a}V^{d'_b}\right], ∫ [ 1 , ∞ ) p φ ( L d κ ( u a , u b ) ∥ x 1 − x 2 ∥ d ) v = 1 ∏ p u v d v ′ v = 1 ∏ p μ ( d u v ) = v ∈ / { a , b } ∏ E [ U d v ′ ] E [ φ ( L d κ ( U , V ) ∥ x 1 − x 2 ∥ d ) U d a ′ V d b ′ ] ,
where U , V U,V U , V are independent random variables with distribution μ \mu μ . As d v ′ ≤ p − 1 ≤ L − 1 d'_v\le p-1\le L-1 d v ′ ≤ p − 1 ≤ L − 1 , the first factor is bounded by a finite constant. For the second factor, as U , V ≥ 1 U,V\ge1 U , V ≥ 1 ,
E [ φ ( ∥ x 1 − x 2 ∥ d L d κ ( U , V ) ) U d a ′ V d b ′ ] ≤ E [ φ ( ∥ x 1 − x 2 ∥ d L d κ ( U , V ) ) U L − 1 V L − 1 ] = Ψ ℓ , m ( x 1 , x 2 ) . \mathbb{E}\left[\varphi\left(\frac{\lVert x_1-x_2\rVert^d}{L^d\kappa(U,V)}\right)U^{d'_a}V^{d'_b}\right]
\le\mathbb{E}\left[\varphi\left(\frac{\lVert x_1-x_2\rVert^d}{L^d\kappa(U,V)}\right)U^{L-1}V^{L-1}\right]=\Psi_{\ell,m}(x_1,x_2). E [ φ ( L d κ ( U , V ) ∥ x 1 − x 2 ∥ d ) U d a ′ V d b ′ ] ≤ E [ φ ( L d κ ( U , V ) ∥ x 1 − x 2 ∥ d ) U L − 1 V L − 1 ] = Ψ ℓ , m ( x 1 , x 2 ) .
Finally, M φ p − 2 ≤ M ‾ φ L − 2 M_\varphi^{p-2}\le\overline{M}_\varphi^{L-2} M φ p − 2 ≤ M φ L − 2 . Summing over the finitely many path edges and patterns concludes the proof.
Part (iv). Continue with f ( x ) = x m f(x)=x^m f ( x ) = x m , m ≥ 2 m\ge2 m ≥ 2 , and set L : = ℓ m L:=\ell m L := ℓ m and G : = G n , x ~ 2 G:=G_{n,\tilde x_2} G := G n , x ~ 2 . Write
S m : = ∑ z ~ ∈ P ~ n ∂ x ~ 1 , x ~ 2 2 g f ( z ~ , G ) . S_m:=\sum_{\tilde z\in\widetilde{\mathcal P}_n}\partial^2_{\tilde x_1,\tilde x_2}g_f(\tilde z,G). S m := z ~ ∈ P n ∑ ∂ x ~ 1 , x ~ 2 2 g f ( z ~ , G ) .
For a vertex z ~ \tilde z z ~ , the second score difference counts length-m m m closed walks based at z z z that visit both inserted vertices. Indeed, every walk missing at least one inserted vertex cancels in the alternating deletion difference. Thus S m ≥ 0 S_m\ge0 S m ≥ 0 , and S m ℓ S_m^\ell S m ℓ counts ordered ℓ \ell ℓ -tuples of such walks. Unlike in Part (iii), their starting vertices are old Poisson vertices and need not coincide.
Group these trajectories according to the equality patterns of their ℓ m \ell m ℓ m indexed walk positions. Each pattern distinguishes the two blocks occupied by x 1 , x 2 x_1,x_2 x 1 , x 2 , both of which meet every walk; each starting position belongs to a nonroot block. Following the notation in Part (iii), let p p p be the number of distinct visited vertices, label x 1 , x 2 x_1,x_2 x 1 , x 2 by 1 , 2 1,2 1 , 2 , and label the remaining vertices in order of first appearance across the ordered walks. Let H H H be the simple undirected graph formed by the distinct traversed edges. Then p ≤ L = ℓ m p\le L=\ell m p ≤ L = ℓ m , and H H H is connected.
For each fixed pattern, average the edge randomness and apply the multivariate Mecke formula as in Part (iii). After also integrating the two inserted marks, its contribution is the same two-root graph integral considered there, with z 1 = x 1 z_1=x_1 z 1 = x 1 , z 2 = x 2 z_2=x_2 z 2 = x 2 , and p − 2 p-2 p − 2 Poisson vertices. Applying the aforementioned argument to bound each pattern’s contribution by
C m M ‾ φ ℓ m − 2 Ψ ℓ , m ( x 1 , x 2 ) . C_m\overline{M}_\varphi^{\ell m-2}\Psi_{\ell,m}(x_1,x_2). C m M φ ℓ m − 2 Ψ ℓ , m ( x 1 , x 2 ) .
The required moment E [ U ℓ m − 1 ] < ∞ \mathbb{E}[U^{\ell m-1}]<\infty E [ U ℓ m − 1 ] < ∞ is included in the assumptions. Since the number of equality patterns depends only on ℓ , m \ell,m ℓ , m , summing their contributions concludes the proof.
Finally, we proceed with the proof of Lemma 10, inspired by the argument of [20 Lemma 5.5] .
Proof of Lemma 10. We prove the two parts separately.
Part (i). Recalling the definition of the difference operator, we get as in [20 Lemma 5.2] that
Δ x ˇ 1 ; x ˇ 1 F n = g f ( x ˇ 1 , G n , x ˇ 1 ) + ∑ X ˇ i ∈ P ˇ n ∂ x ˇ 1 g f ( X ˇ i , G n , x ˇ 1 ) . (36) \Delta_{\check{x}_1;\check{x}_1}F_n=g_f(\check{x}_1,G_{n,\check{x}_1})+\sum_{\check{X}_i\in\check{\mathcal{P}}_n}\partial_{\check{x}_1}g_f(\check{X}_i,G_{n,\check{x}_1}). \tag*{(36)} Δ x ˇ 1 ; x ˇ 1 F n = g f ( x ˇ 1 , G n , x ˇ 1 ) + X ˇ i ∈ P ˇ n ∑ ∂ x ˇ 1 g f ( X ˇ i , G n , x ˇ 1 ) . (36)
Thus, E [ ∣ Δ x ˇ 1 ; x ˇ 1 F n ∣ 4 ] ≤ 16 E [ g f ( x ˇ 1 , G n , x ˇ 1 ) 4 ] + 16 E [ ( ∑ X ˇ i ∈ P ˇ n ∂ x ˇ 1 g f ( X ˇ i , G n , x ˇ 1 ) ) 4 ] \mathbb{E}[|\Delta_{\check{x}_1;\check{x}_1}F_n|^4]\leq16\mathbb{E}[g_f(\check{x}_1,G_{n,\check{x}_1})^4]+16\mathbb{E}\left[\left(\sum_{\check{X}_i\in\check{\mathcal{P}}_n}\partial_{\check{x}_1}g_f(\check{X}_i,G_{n,\check{x}_1})\right)^4\right] E [ ∣ Δ x ˇ 1 ; x ˇ 1 F n ∣ 4 ] ≤ 16 E [ g f ( x ˇ 1 , G n , x ˇ 1 ) 4 ] + 16 E [ ( ∑ X ˇ i ∈ P ˇ n ∂ x ˇ 1 g f ( X ˇ i , G n , x ˇ 1 ) ) 4 ] . By Lemma 12(i), the first term is controlled as E [ g f ( x ˇ 1 , G n , x ˇ 1 ) 4 ] ≤ C u 1 4 m − 1 M ‾ φ 4 m − 1 \mathbb{E}[g_f(\check{x}_1,G_{n,\check{x}_1})^4]\leq C u_1^{4m-1}\overline{M}_{\varphi}^{4m-1} E [ g f ( x ˇ 1 , G n , x ˇ 1 ) 4 ] ≤ C u 1 4 m − 1 M φ 4 m − 1 . By Lemma 12(ii), the second term is bounded by
E ∣ ∑ X ˇ i ∈ P ˇ n ∂ x ˇ 1 g f ( X ˇ i , G n , x ˇ 1 ) ∣ 4 ≤ C u 1 4 m − 1 M ‾ φ 4 m − 1 . \mathbb{E}\left|\sum_{\check{X}_i\in\check{\mathcal{P}}_n}\partial_{\check{x}_1}g_f(\check{X}_i,G_{n,\check{x}_1})\right|^4\leq C u_1^{4m-1}\overline{M}_{\varphi}^{4m-1}. E X ˇ i ∈ P ˇ n ∑ ∂ x ˇ 1 g f ( X ˇ i , G n , x ˇ 1 ) 4 ≤ C u 1 4 m − 1 M φ 4 m − 1 .
This proves Part (i).
Part (ii). Recalling the definition of second-order difference operator, we have
Δ x ˇ 1 , x ˇ 2 ; x ˇ 2 2 F n = ∂ x ˇ 1 g f ( x ˇ 2 , G n , x ˇ 2 ) + ∂ x ˇ 2 g f ( x ˇ 1 , G n , x ˇ 2 ) + ∑ X ˇ i ∈ P ˇ n ∂ x ˇ 1 , x ˇ 2 2 g f ( X ˇ i , G n , x ˇ 2 ) \Delta^2_{\check{x}_1,\check{x}_2;\check{x}_2}F_n=\partial_{\check{x}_1}g_f(\check{x}_2,G_{n,\check{x}_2})+\partial_{\check{x}_2}g_f(\check{x}_1,G_{n,\check{x}_2})+\sum_{\check{X}_i\in\check{\mathcal{P}}_n}\partial^2_{\check{x}_1,\check{x}_2}g_f(\check{X}_i,G_{n,\check{x}_2}) Δ x ˇ 1 , x ˇ 2 ; x ˇ 2 2 F n = ∂ x ˇ 1 g f ( x ˇ 2 , G n , x ˇ 2 ) + ∂ x ˇ 2 g f ( x ˇ 1 , G n , x ˇ 2 ) + X ˇ i ∈ P ˇ n ∑ ∂ x ˇ 1 , x ˇ 2 2 g f ( X ˇ i , G n , x ˇ 2 )
(see [20 Lemma 5.2] for details), implying that
∫ [ 1 , ∞ ) 2 E ∣ Δ ( x 1 , u 1 ) , ( x 2 , u 2 ) ; x ˇ 2 2 F n ∣ 4 μ ( d u 1 ) μ ( d u 2 ) ≤ C ∫ [ 1 , ∞ ) 2 ( E ∣ ∂ x ˇ 1 g f ( x ˇ 2 , G n , x ˇ 2 ) ∣ 4 + E ∣ ∂ x ˇ 2 g f ( x ˇ 1 , G n , x ˇ 2 ) ∣ 4 + E ∣ ∑ X ˇ i ∈ P ˇ n ∂ x ˇ 1 , x ˇ 2 2 g f ( X ˇ i , G n , x ˇ 2 ) ∣ 4 ) μ ( d u 1 ) μ ( d u 2 ) . \int_{[1,\infty)^2}\mathbb{E}\left|\Delta^2_{(x_1,u_1),(x_2,u_2);\check{x}_2}F_n\right|^4\mu(\mathrm{d}u_1)\mu(\mathrm{d}u_2)
\leq C\int_{[1,\infty)^2}\left(\mathbb{E}|\partial_{\check{x}_1}g_f(\check{x}_2,G_{n,\check{x}_2})|^4+\mathbb{E}|\partial_{\check{x}_2}g_f(\check{x}_1,G_{n,\check{x}_2})|^4+\mathbb{E}\left|\sum_{\check{X}_i\in\check{\mathcal{P}}_n}\partial^2_{\check{x}_1,\check{x}_2}g_f(\check{X}_i,G_{n,\check{x}_2})\right|^4\right)\mu(\mathrm{d}u_1)\mu(\mathrm{d}u_2). ∫ [ 1 , ∞ ) 2 E Δ ( x 1 , u 1 ) , ( x 2 , u 2 ) ; x ˇ 2 2 F n 4 μ ( d u 1 ) μ ( d u 2 ) ≤ C ∫ [ 1 , ∞ ) 2 E ∣ ∂ x ˇ 1 g f ( x ˇ 2 , G n , x ˇ 2 ) ∣ 4 + E ∣ ∂ x ˇ 2 g f ( x ˇ 1 , G n , x ˇ 2 ) ∣ 4 + E X ˇ i ∈ P ˇ n ∑ ∂ x ˇ 1 , x ˇ 2 2 g f ( X ˇ i , G n , x ˇ 2 ) 4 μ ( d u 1 ) μ ( d u 2 ) .
By Lemma 12(iii), the first and second terms are both bounded by C M ‾ φ 4 m − 2 Ψ 4 , m ( x 1 , x 2 ) C\overline{M}_{\varphi}^{4m-2}\Psi_{4,m}(x_1,x_2) C M φ 4 m − 2 Ψ 4 , m ( x 1 , x 2 ) , where C > 0 C>0 C > 0 is a constant. For the last term, we obtain by Lemma 12(iv),
∫ [ 1 , ∞ ) 2 E [ ( ∑ X ˇ i ∈ P ˇ n ∂ x ˇ 1 , x ˇ 2 2 g f ( X ˇ i , G n , x ˇ 2 ) ) 4 ] μ ( d u 1 ) μ ( d u 2 ) ≤ C M ‾ φ 4 m − 2 Ψ 4 , m ( x 1 , x 2 ) . \int_{[1,\infty)^2}\mathbb{E}\left[\left(\sum_{\check{X}_i\in\check{\mathcal{P}}_n}\partial^2_{\check{x}_1,\check{x}_2}g_f(\check{X}_i,G_{n,\check{x}_2})\right)^4\right]\mu(\mathrm{d}u_1)\mu(\mathrm{d}u_2)\leq C\overline{M}_{\varphi}^{4m-2}\Psi_{4,m}(x_1,x_2). ∫ [ 1 , ∞ ) 2 E X ˇ i ∈ P ˇ n ∑ ∂ x ˇ 1 , x ˇ 2 2 g f ( X ˇ i , G n , x ˇ 2 ) 4 μ ( d u 1 ) μ ( d u 2 ) ≤ C M φ 4 m − 2 Ψ 4 , m ( x 1 , x 2 ) .
This gives the assertion.
Proof for general test functions
We now extend the CLT for polynomial test functions in Theorem 2 to the broader class of test functions H c 0 2 \mathcal{H}_{c_0}^2 H c 0 2 . This relies on an abstract approximation principle stated in Proposition 13 below that allows us to pass from polynomials to general test functions. The argument is essentially the same as in [18 Section 5.1] and is not specific to the present model.
In Section 5.1 we state the abstract approximation principle and show how Theorem 3 follows through two key intermediate results. These key results are proved in Section 5.2 with the main difference for the weighted RCM being that the variance estimate (Proposition 15) is derived from the RCM version of the Poincaré inequality stated in Theorem 7 involving the RCM difference operator Δ x ˇ \Delta_{\check{x}} Δ x ˇ rather than the point-wise difference operator for Poisson input. Otherwise, the proof follows the Fourier-Duhamel argument of [18 Thm. 5.2] . The only model-specific input is an exponential trace estimate stated separately (Lemma 16) obtained by orienting edges according to the vertex marks.
Key ingredients and proof of Theorem 3
We first state the following abstract approximation principle used to pass from polynomial test functions to general test functions. All constants in the estimates below are independent of n n n , unless explicitly stated otherwise.
Proposition 13 (Abstract approximation principle). Let L \mathcal{L} L be a real vector space of functions equipped with a norm ∥ ⋅ ∥ L \|\cdot\|_{\mathcal{L}} ∥ ⋅ ∥ L , and let Z n : L → L 2 ( Ω ) Z_n:\mathcal{L}\to L^2(\Omega) Z n : L → L 2 ( Ω ) be linear maps. Assume that there is a dense subspace L ~ ⊆ L \widetilde{\mathcal{L}}\subseteq\mathcal{L} L ⊆ L and a quadratic form V : L ~ → R ≥ 0 V:\widetilde{\mathcal{L}}\to\mathbb{R}_{\geq0} V : L → R ≥ 0 such that Z n ( φ ) → d N ( 0 , V ( φ ) ) Z_n(\varphi)\xrightarrow{d}\mathcal{N}(0,V(\varphi)) Z n ( φ ) d N ( 0 , V ( φ )) for every φ ∈ L ~ \varphi\in\widetilde{\mathcal{L}} φ ∈ L . Assume also that, for some C > 0 C>0 C > 0 ,
sup n ≥ 1 E [ Z n ( φ ) 2 ] ≤ C ∥ φ ∥ L 2 for all φ ∈ L . (37) \sup_{n\geq1}\mathbb{E}[Z_n(\varphi)^2]\leq C\|\varphi\|_{\mathcal{L}}^2 \qquad\text{for all }\varphi\in\mathcal{L}. \tag*{(37)} n ≥ 1 sup E [ Z n ( φ ) 2 ] ≤ C ∥ φ ∥ L 2 for all φ ∈ L . (37) Then, V V V extends uniquely to a continuous quadratic form on L \mathcal{L} L , and Z n ( φ ) → d N ( 0 , V ( φ ) ) Z_n(\varphi)\xrightarrow{d}\mathcal{N}(0,V(\varphi)) Z n ( φ ) d N ( 0 , V ( φ )) for every φ ∈ L \varphi\in\mathcal{L} φ ∈ L .
This is similar to [18 Prop. 5.1] adapted from [4 ] . We give a proof for the convenience of the reader.
Proof. For φ , ψ ∈ L ~ \varphi,\psi\in\widetilde{\mathcal{L}} φ , ψ ∈ L , the second moment bound and the assumed convergence on L ~ \widetilde{\mathcal{L}} L give
V ( φ − ψ ) ≤ lim inf n → ∞ E [ Z n ( φ − ψ ) 2 ] ≤ C ∥ φ − ψ ∥ L 2 . V(\varphi-\psi)\leq\liminf_{n\to\infty}\mathbb{E}[Z_n(\varphi-\psi)^2]\leq C\|\varphi-\psi\|_{\mathcal{L}}^2. V ( φ − ψ ) ≤ n → ∞ lim inf E [ Z n ( φ − ψ ) 2 ] ≤ C ∥ φ − ψ ∥ L 2 .
Hence, V V V is continuous on L ~ \widetilde{\mathcal{L}} L , and it extends uniquely by continuity to L \mathcal{L} L .
Fix φ ∈ L \varphi\in\mathcal{L} φ ∈ L , and choose φ k ∈ L ~ \varphi_k\in\widetilde{\mathcal{L}} φ k ∈ L with ∥ φ k − φ ∥ L → 0 \|\varphi_k-\varphi\|_{\mathcal{L}}\to0 ∥ φ k − φ ∥ L → 0 . The same variance bound gives
sup n ≥ 1 E [ ∣ Z n ( φ ) − Z n ( φ k ) ∣ 2 ] ≤ C ∥ φ − φ k ∥ L 2 . \sup_{n\geq1}\mathbb{E}[|Z_n(\varphi)-Z_n(\varphi_k)|^2]\leq C\|\varphi-\varphi_k\|_{\mathcal{L}}^2. n ≥ 1 sup E [ ∣ Z n ( φ ) − Z n ( φ k ) ∣ 2 ] ≤ C ∥ φ − φ k ∥ L 2 .
For each fixed k k k , Z n ( φ k ) → d N ( 0 , V ( φ k ) ) Z_n(\varphi_k)\xrightarrow{d}\mathcal{N}(0,V(\varphi_k)) Z n ( φ k ) d N ( 0 , V ( φ k )) , and V ( φ k ) → V ( φ ) V(\varphi_k)\to V(\varphi) V ( φ k ) → V ( φ ) . Let Y k ∼ N ( 0 , V ( φ k ) ) Y_k\sim\mathcal{N}(0,V(\varphi_k)) Y k ∼ N ( 0 , V ( φ k )) and Y ∼ N ( 0 , V ( φ ) ) Y\sim\mathcal{N}(0,V(\varphi)) Y ∼ N ( 0 , V ( φ )) . If h h h is bounded by 1 1 1 and 1 1 1 -Lipschitz, then
∣ E h ( Z n ( φ ) ) − E h ( Y ) ∣ ≤ E [ ∣ Z n ( φ ) − Z n ( φ k ) ∣ ] + ∣ E h ( Z n ( φ k ) ) − E h ( Y k ) ∣ + ∣ E h ( Y k ) − E h ( Y ) ∣ . |\mathbb{E}h(Z_n(\varphi))-\mathbb{E}h(Y)|\leq\mathbb{E}[|Z_n(\varphi)-Z_n(\varphi_k)|]+|\mathbb{E}h(Z_n(\varphi_k))-\mathbb{E}h(Y_k)|+|\mathbb{E}h(Y_k)-\mathbb{E}h(Y)|. ∣ E h ( Z n ( φ )) − E h ( Y ) ∣ ≤ E [ ∣ Z n ( φ ) − Z n ( φ k ) ∣ ] + ∣ E h ( Z n ( φ k )) − E h ( Y k ) ∣ + ∣ E h ( Y k ) − E h ( Y ) ∣.
Taking lim sup n → ∞ \limsup_{n\to\infty} lim sup n → ∞ and then letting k → ∞ k\to\infty k → ∞ shows weak convergence. □
To use the abstract approximation to show Theorem 3 , we need two input results. The first input is a lemma that allows us to use Proposition 13 to pass from polynomials to H c , 0 2 \mathcal{H}_{c,0}^2 H c , 0 2 .
Lemma 14 (Polynomial density in H c , 0 2 \mathcal{H}_{c,0}^2 H c , 0 2 ). Let c > 0 c>0 c > 0 . Then, polynomials p p p satisfying p ( 0 ) = 0 p(0)=0 p ( 0 ) = 0 are dense in H c , 0 2 \mathcal{H}_{c,0}^2 H c , 0 2 , that is, for every f ∈ H c , 0 2 f\in\mathcal{H}_{c,0}^2 f ∈ H c , 0 2 , there are polynomials p k p_k p k with p k ( 0 ) = 0 p_k(0)=0 p k ( 0 ) = 0 such that
∥ f − p k ∥ H c 2 → 0. \|f-p_k\|_{\mathcal{H}_{c}^2}\to0. ∥ f − p k ∥ H c 2 → 0.
The second input controls the variance of exponentially twisted statistics and allows us to verify (37) .
Proposition 15 (Variance estimate for exponentially twisted statistics). Assume that (EM) and (PK) hold and that M φ , 1 < ∞ M_{\varphi,1}<\infty M φ , 1 < ∞ . For every c ≠ 0 c\ne0 c = 0 , there exists C > 0 C>0 C > 0 , depending only on c c c and the model parameters, such that for every f ∈ H 2 ( R ) f\in H^2(\mathbb{R}) f ∈ H 2 ( R ) satisfying f ( 0 ) = 0 f(0)=0 f ( 0 ) = 0 ,
Var ( Tr [ f ( A n ) e c A n ] ) ≤ C ∣ W n ∣ ( ∥ f ∥ 2 2 + ∥ f ′ ′ ∥ 2 2 ) . \operatorname{Var}(\operatorname{Tr}[f(A_n)e^{cA_n}])\leq C|W_n|(\|f\|_2^2+\|f''\|_2^2). Var ( Tr [ f ( A n ) e c A n ]) ≤ C ∣ W n ∣ ( ∥ f ∥ 2 2 + ∥ f ′′ ∥ 2 2 ) .
The two preceding inputs now give the CLT for the general test-functions.
Proof of Theorem 3 . We apply Proposition 13 with L = H c , 0 2 \mathcal{L}= \mathcal{H}_{c,0}^2 L = H c , 0 2 . By Lemma 14 , we may take L ~ \widetilde{\mathcal{L}} L to be the space of polynomials p p p satisfying p ( 0 ) = 0 p(0)=0 p ( 0 ) = 0 . For f ∈ H c , 0 2 f\in\mathcal{H}_{c,0}^2 f ∈ H c , 0 2 , define
Z n ( f ) : = ∣ W n ∣ − 1 / 2 ( Tr [ f ( A n ) ] − E Tr [ f ( A n ) ] ) . Z_n(f):=|W_n|^{-1/2}\left(\operatorname{Tr}[f(A_n)]-\mathbb{E}\operatorname{Tr}[f(A_n)]\right). Z n ( f ) := ∣ W n ∣ − 1/2 ( Tr [ f ( A n )] − E Tr [ f ( A n )] ) .
By Theorem 2 , Z n ( p ) Z_n(p) Z n ( p ) converges to a centered Gaussian random variable for every such polynomial p p p . The limiting variance σ p 2 \sigma_p^2 σ p 2 is a quadratic form in p p p since it is the limit of quadratic forms.
It remains to verify (37) . Let ψ ( x ) : = f ( x ) sech ( c x ) \psi(x):=f(x)\operatorname{sech}(cx) ψ ( x ) := f ( x ) sech ( c x ) . Then, ψ ( 0 ) = 0 \psi(0)=0 ψ ( 0 ) = 0 and f ( x ) = ψ ( x ) cosh ( c x ) f(x)=\psi(x)\cosh(cx) f ( x ) = ψ ( x ) cosh ( c x ) , and hence
Tr [ f ( A n ) ] = ( Tr [ ψ ( A n ) e c A n ] + Tr [ ψ ( A n ) e − c A n ] ) / 2. \operatorname{Tr}[f(A_n)]=\left(\operatorname{Tr}[\psi(A_n)e^{cA_n}]+\operatorname{Tr}[\psi(A_n)e^{-cA_n}]\right)/2. Tr [ f ( A n )] = ( Tr [ ψ ( A n ) e c A n ] + Tr [ ψ ( A n ) e − c A n ] ) /2.
Since ψ ∈ H 2 ( R ) \psi\in H^2(\mathbb{R}) ψ ∈ H 2 ( R ) , we may apply Proposition 15 to ψ \psi ψ with constants c c c and − c -c − c . Together with the inequality Var ( ( X + Y ) / 2 ) ≤ ( Var X + Var Y ) / 2 \operatorname{Var}((X+Y)/2)\leq(\operatorname{Var}X+\operatorname{Var}Y)/2 Var (( X + Y ) /2 ) ≤ ( Var X + Var Y ) /2 , this gives
∣ W n ∣ − 1 Var ( Tr [ f ( A n ) ] ) = ∣ W n ∣ − 1 Var ( ( Tr [ ψ ( A n ) e c A n ] + Tr [ ψ ( A n ) e − c A n ] ) / 2 ) ≤ C ( ∥ ψ ∥ 2 2 + ∥ ψ ′ ′ ∥ 2 2 ) . |W_n|^{-1}\operatorname{Var}(\operatorname{Tr}[f(A_n)])=|W_n|^{-1}\operatorname{Var}\left((\operatorname{Tr}[\psi(A_n)e^{cA_n}]+\operatorname{Tr}[\psi(A_n)e^{-cA_n}])/2\right)\leq C(\|\psi\|_2^2+\|\psi''\|_2^2). ∣ W n ∣ − 1 Var ( Tr [ f ( A n )]) = ∣ W n ∣ − 1 Var ( ( Tr [ ψ ( A n ) e c A n ] + Tr [ ψ ( A n ) e − c A n ]) /2 ) ≤ C ( ∥ ψ ∥ 2 2 + ∥ ψ ′′ ∥ 2 2 ) .
If h ( x ) : = sech ( c x ) h(x):=\operatorname{sech}(cx) h ( x ) := sech ( c x ) , then ∣ h ′ ∣ ≤ ∣ c ∣ h |h'|\leq|c|h ∣ h ′ ∣ ≤ ∣ c ∣ h and ∣ h ′ ′ ∣ ≤ c 2 h |h''|\leq c^2h ∣ h ′′ ∣ ≤ c 2 h . Since ψ = f h \psi=fh ψ = f h ,
∥ ψ ∥ 2 2 + ∥ ψ ′ ′ ∥ 2 2 ≤ C c ( ∥ f h ∥ 2 2 + ∥ f ′ h ′ ∥ 2 2 + ∥ f ′ ′ h ∥ 2 2 ) ≤ C c ∥ f ∥ H c 2 2 . \|\psi\|_2^2+\|\psi''\|_2^2\leq C_c(\|fh\|_2^2+\|f'h'\|_2^2+\|f''h\|_2^2)\leq C_c\|f\|_{\mathcal{H}_{c}^2}^2. ∥ ψ ∥ 2 2 + ∥ ψ ′′ ∥ 2 2 ≤ C c ( ∥ f h ∥ 2 2 + ∥ f ′ h ′ ∥ 2 2 + ∥ f ′′ h ∥ 2 2 ) ≤ C c ∥ f ∥ H c 2 2 .
Hence, sup n ≥ 1 E [ Z n ( f ) 2 ] ≤ C ∥ f ∥ H c 2 2 \sup_{n\geq1}\mathbb{E}[Z_n(f)^2]\leq C\|f\|_{\mathcal{H}_{c}^2}^2 sup n ≥ 1 E [ Z n ( f ) 2 ] ≤ C ∥ f ∥ H c 2 2 for all f ∈ H c , 0 2 f\in\mathcal{H}_{c,0}^2 f ∈ H c , 0 2 . Proposition 13 therefore extends the polynomial CLT to every f ∈ H c , 0 2 f\in\mathcal{H}_{c,0}^2 f ∈ H c , 0 2 and Theorem 3 follows. □
Proof of key ingredients
We now prove Lemma 14 .
Proof of Lemma 14 . Fix c > 0 c>0 c > 0 and set w ( x ) : = sech 2 ( c x ) w(x):=\operatorname{sech}^{2}(cx) w ( x ) := sech 2 ( c x ) . Then e − 2 c ∣ x ∣ ≤ w ( x ) ≤ 4 e − 2 c ∣ x ∣ e^{-2c|x|}\le w(x)\le4e^{-2c|x|} e − 2 c ∣ x ∣ ≤ w ( x ) ≤ 4 e − 2 c ∣ x ∣ .
We first prove that the set of polynomials is dense in L 2 ( w d x ) L^{2}(w\,dx) L 2 ( w d x ) . Suppose that g ∈ L 2 ( w d x ) g\in L^{2}(w\,dx) g ∈ L 2 ( w d x ) is orthogonal to every polynomial. For ∣ Re z ∣ < c |\operatorname{Re}z|<c ∣ Re z ∣ < c , define
F ( z ) : = ∫ R g ( x ) w ( x ) e z x d x . F(z):=\int_{\mathbb{R}}g(x)w(x)e^{zx}\,dx. F ( z ) := ∫ R g ( x ) w ( x ) e z x d x .
Then F F F is holomorphic on { z : ∣ Re z ∣ < c } \{z:|\operatorname{Re}z|<c\} { z : ∣ Re z ∣ < c } , and orthogonality gives F ( k ) ( 0 ) = ∫ R x k g ( x ) w ( x ) d x = 0 F^{(k)}(0)=\int_{\mathbb{R}}x^{k}g(x)w(x)\,dx=0 F ( k ) ( 0 ) = ∫ R x k g ( x ) w ( x ) d x = 0 for all k ≥ 0 k\ge0 k ≥ 0 . Hence F ≡ 0 F\equiv0 F ≡ 0 . In particular, the Fourier transform of g w gw g w vanishes, implying g w = 0 gw=0 g w = 0 a.e. As w > 0 w>0 w > 0 , we have g = 0 g=0 g = 0 a.e. The orthogonal complement of the polynomials is trivial, proving density in L 2 ( w d x ) L^{2}(w\,dx) L 2 ( w d x ) .
Now, define ( J g ) ( x ) : = ∫ 0 x g ( t ) d t (Jg)(x):=\int_{0}^{x}g(t)\,dt ( J g ) ( x ) := ∫ 0 x g ( t ) d t . Then e − c x ( J g ) ( x ) = ∫ 0 x e − c ( x − t ) e − c t g ( t ) d t e^{-cx}(Jg)(x)=\int_{0}^{x}e^{-c(x-t)}e^{-ct}g(t)\,dt e − c x ( J g ) ( x ) = ∫ 0 x e − c ( x − t ) e − c t g ( t ) d t . The convolution kernel e − c s 1 { s ≥ 0 } e^{-cs}\mathbf{1}_{\{s\ge0\}} e − cs 1 { s ≥ 0 } has L 1 L^{1} L 1 norm 1 / c 1/c 1/ c . By Young’s convolution inequality,
∫ 0 ∞ e − 2 c x ∣ J g ( x ) ∣ 2 d x ≤ c − 2 ∫ 0 ∞ e − 2 c x ∣ g ( x ) ∣ 2 d x . \int_{0}^{\infty}e^{-2cx}|Jg(x)|^{2}\,dx\le c^{-2}\int_{0}^{\infty}e^{-2cx}|g(x)|^{2}\,dx. ∫ 0 ∞ e − 2 c x ∣ J g ( x ) ∣ 2 d x ≤ c − 2 ∫ 0 ∞ e − 2 c x ∣ g ( x ) ∣ 2 d x .
By symmetry we have the corresponding inequality on the negative domain. Hence,
∥ J g ∥ L 2 ( w ) ≤ 2 c ∥ g ∥ L 2 ( w ) . (38) \lVert Jg\rVert_{L^{2}(w)}\le\frac{2}{c}\lVert g\rVert_{L^{2}(w)}. \tag*{(38)} ∥ J g ∥ L 2 ( w ) ≤ c 2 ∥ g ∥ L 2 ( w ) . (38)
Let f ∈ H c 2 f\in H_{c}^{2} f ∈ H c 2 . Choose polynomials q k q_{k} q k such that q k → f ′ ′ q_{k}\to f^{\prime\prime} q k → f ′′ in L 2 ( w d x ) L^{2}(w\,dx) L 2 ( w d x ) , and set
p k ( x ) : = f ( 0 ) + f ′ ( 0 ) x + ∫ 0 x ( x − t ) q k ( t ) d t . p_{k}(x):=f(0)+f^{\prime}(0)x+\int_{0}^{x}(x-t)q_{k}(t)\,dt. p k ( x ) := f ( 0 ) + f ′ ( 0 ) x + ∫ 0 x ( x − t ) q k ( t ) d t .
Each p k p_{k} p k is a polynomial, with p k ( 0 ) = f ( 0 ) p_{k}(0)=f(0) p k ( 0 ) = f ( 0 ) , p k ′ ( 0 ) = f ′ ( 0 ) p_{k}^{\prime}(0)=f^{\prime}(0) p k ′ ( 0 ) = f ′ ( 0 ) , and p k ′ ′ = q k p_{k}^{\prime\prime}=q_{k} p k ′′ = q k . Thus
( f − p k ) ′ ′ = f ′ ′ − q k , ( f − p k ) ′ = J ( f ′ ′ − q k ) , f − p k = J 2 ( f ′ ′ − q k ) . (f-p_{k})^{\prime\prime}=f^{\prime\prime}-q_{k},\qquad(f-p_{k})^{\prime}=J(f^{\prime\prime}-q_{k}),\qquad f-p_{k}=J^{2}(f^{\prime\prime}-q_{k}). ( f − p k ) ′′ = f ′′ − q k , ( f − p k ) ′ = J ( f ′′ − q k ) , f − p k = J 2 ( f ′′ − q k ) .
Hence using (38) ,
∥ f − p k ∥ H c 2 2 ≤ ( 1 + 4 c 2 + 16 c 4 ) ∥ f ′ ′ − q k ∥ L 2 ( w ) 2 ⟶ 0. \lVert f-p_{k}\rVert_{H_{c}^{2}}^{2}\le\left(1+\frac{4}{c^{2}}+\frac{16}{c^{4}}\right)\lVert f^{\prime\prime}-q_{k}\rVert_{L^{2}(w)}^{2}\longrightarrow0. ∥ f − p k ∥ H c 2 2 ≤ ( 1 + c 2 4 + c 4 16 ) ∥ f ′′ − q k ∥ L 2 ( w ) 2 ⟶ 0.
If f ( 0 ) = 0 f(0)=0 f ( 0 ) = 0 , then p k ( 0 ) = 0 p_{k}(0)=0 p k ( 0 ) = 0 for every k k k . This concludes the proof.
We next prove exponential trace bounds that are needed for the Fourier–Duhamel argument in the proof of Proposition 15 . Recall that A n A_{n} A n denotes the adjacency matrix of the weighted RCM G ( ξ n ) G(\xi_{n}) G ( ξ n ) . Let D n D_{n} D n denote the diagonal matrix containing the degrees of G ( ξ n ) G(\xi_{n}) G ( ξ n ) . For x ˇ ∈ W ^ n \check{x}\in\widehat{W}_{n} x ˇ ∈ W n , let A ^ n x ˇ \widehat{A}_{n}^{\check{x}} A n x ˇ be the adjacency matrix of G n , x ˇ : = G { 1 } ( ξ n , ( x ˇ ) ) G_{n,\check{x}}:=G_{\{1\}}(\xi_{n},(\check{x})) G n , x ˇ := G { 1 } ( ξ n , ( x ˇ )) . We write e x ˇ e_{\check{x}} e x ˇ for the standard basis vector corresponding to the x ˇ \check{x} x ˇ -coordinate and set
a x ˇ : = A ^ n x ˇ e x ˇ . a_{\check{x}}:=\widehat{A}_{n}^{\check{x}}e_{\check{x}}. a x ˇ := A n x ˇ e x ˇ .
Lemma 16 (Trace bounds). Assume that (EM) and (PK) hold and that M φ , 1 < ∞ M_{\varphi,1}<\infty M φ , 1 < ∞ . For every R > 0 R>0 R > 0 ,
sup n ≥ 1 sup 0 ≤ r ≤ R ∣ W n ∣ − 1 E [ Tr ( e r A n ) + Tr ( e − r A n ) ] < ∞ , (39) \sup_{n\ge1}\sup_{0\le r\le R}|W_{n}|^{-1}\mathbb{E}\left[\operatorname{Tr}(e^{rA_{n}})+\operatorname{Tr}(e^{-rA_{n}})\right]<\infty, \tag*{(39)} n ≥ 1 sup 0 ≤ r ≤ R sup ∣ W n ∣ − 1 E [ Tr ( e r A n ) + Tr ( e − r A n ) ] < ∞ , (39) sup n ≥ 1 sup 0 ≤ r ≤ R ∣ W n ∣ − 1 E [ Tr ( D n A n 2 e r A n ) ] < ∞ . (40) \sup_{n\ge1}\sup_{0\le r\le R}|W_{n}|^{-1}\mathbb{E}\left[\operatorname{Tr}(D_{n}A_{n}^{2}e^{rA_{n}})\right]<\infty. \tag*{(40)} n ≥ 1 sup 0 ≤ r ≤ R sup ∣ W n ∣ − 1 E [ Tr ( D n A n 2 e r A n ) ] < ∞. (40) Moreover,
sup n ≥ 1 sup 0 ≤ r ≤ R ∣ W n ∣ − 1 ∫ W ^ n E [ ⟨ e r A ^ n x ˇ e x ˇ , e x ˇ ⟩ ] λ ( d x ˇ ) < ∞ , (41) \sup_{n\ge1}\sup_{0\le r\le R}|W_{n}|^{-1}\int_{\widehat{W}_{n}}\mathbb{E}\left[\left\langle e^{r\widehat{A}_{n}^{\check{x}}}e_{\check{x}},e_{\check{x}}\right\rangle\right]\lambda(d\check{x})<\infty, \tag*{(41)} n ≥ 1 sup 0 ≤ r ≤ R sup ∣ W n ∣ − 1 ∫ W n E [ ⟨ e r A n x ˇ e x ˇ , e x ˇ ⟩ ] λ ( d x ˇ ) < ∞ , (41) sup n ≥ 1 sup 0 ≤ r ≤ R ∣ W n ∣ − 1 ∫ W ^ n E [ ⟨ e r A ^ n x ˇ a x ˇ , a x ˇ ⟩ ⟨ a x ˇ , a x ˇ ⟩ ] λ ( d x ˇ ) < ∞ . (42) \sup_{n\ge1}\sup_{0\le r\le R}|W_{n}|^{-1}\int_{\widehat{W}_{n}}\mathbb{E}\left[\left\langle e^{r\widehat{A}_{n}^{\check{x}}}a_{\check{x}},a_{\check{x}}\right\rangle\langle a_{\check{x}},a_{\check{x}}\rangle\right]\lambda(d\check{x})<\infty. \tag*{(42)} n ≥ 1 sup 0 ≤ r ≤ R sup ∣ W n ∣ − 1 ∫ W n E [ ⟨ e r A n x ˇ a x ˇ , a x ˇ ⟩ ⟨ a x ˇ , a x ˇ ⟩ ] λ ( d x ˇ ) < ∞. (42) Proof. We first derive the estimates (41) and (42) , given (39) and (40) . By the Mecke formula (10) ,
E Tr ( e r A n ) = E ∑ x ∈ P n ⟨ e r A n e x , e x ⟩ = ∫ W ^ n E [ ⟨ e r A ^ n x ˇ e x ˇ , e x ˇ ⟩ ] λ ( d x ˇ ) . \mathbb{E}\operatorname{Tr}(e^{rA_{n}})=\mathbb{E}\sum_{x\in P_{n}}\langle e^{rA_{n}}e_{x},e_{x}\rangle=\int_{\widehat{W}_{n}}\mathbb{E}\left[\left\langle e^{r\widehat{A}_{n}^{\check{x}}}e_{\check{x}},e_{\check{x}}\right\rangle\right]\lambda(d\check{x}). E Tr ( e r A n ) = E x ∈ P n ∑ ⟨ e r A n e x , e x ⟩ = ∫ W n E [ ⟨ e r A n x ˇ e x ˇ , e x ˇ ⟩ ] λ ( d x ˇ ) .
Thus, (38) follows from (36). To prove (39), observe that ⟨ a x ~ , a x ~ ⟩ = d x ~ \langle a_{\tilde{x}},a_{\tilde{x}}\rangle=d_{\tilde{x}} ⟨ a x ~ , a x ~ ⟩ = d x ~ and, since a x ~ = A n x ~ e x ~ a_{\tilde{x}}=A_n^{\tilde{x}}e_{\tilde{x}} a x ~ = A n x ~ e x ~ , we have
⟨ e r A n x ~ a x ~ , a x ~ ⟩ = ( ( A n x ~ ) 2 e r A n x ~ ) x ~ x ~ . \left\langle e^{rA_n^{\tilde{x}}}a_{\tilde{x}},a_{\tilde{x}}\right\rangle=\left((A_n^{\tilde{x}})^2e^{rA_n^{\tilde{x}}}\right)_{\tilde{x}\tilde{x}}. ⟨ e r A n x ~ a x ~ , a x ~ ⟩ = ( ( A n x ~ ) 2 e r A n x ~ ) x ~ x ~ .
Mecke formula therefore gives
∫ W n E [ ⟨ e r A n x ~ a x ~ , a x ~ ⟩ ∣ ⟨ a x ~ , a x ~ ⟩ ] λ ( d x ~ ) = E ∑ x ∈ P n d x ( A n 2 e r A n ) x x = E [ Tr ( D n A n 2 e r A n ) ] . \int_{W_n}\mathbb{E}\left[\left\langle e^{rA_n^{\tilde{x}}}a_{\tilde{x}},a_{\tilde{x}}\right\rangle\mid\langle a_{\tilde{x}},a_{\tilde{x}}\rangle\right]\lambda(\mathrm{d}\tilde{x})=\mathbb{E}\sum_{x\in P_n}d_x\left(A_n^2e^{rA_n}\right)_{xx}=\mathbb{E}\left[\operatorname{Tr}\left(D_nA_n^2e^{rA_n}\right)\right]. ∫ W n E [ ⟨ e r A n x ~ a x ~ , a x ~ ⟩ ∣ ⟨ a x ~ , a x ~ ⟩ ] λ ( d x ~ ) = E x ∈ P n ∑ d x ( A n 2 e r A n ) xx = E [ Tr ( D n A n 2 e r A n ) ] .
The bound (39) now follows from (37).
We now verify (36) and (37). A direct way to bound the adjacency matrix is through total degrees: ∣ z ⊤ A n z ∣ ≤ ∑ x ∈ P n d x z x 2 = z ⊤ D n z |z^\top A_nz|\le\sum_{x\in P_n}d_xz_x^2=z^\top D_nz ∣ z ⊤ A n z ∣ ≤ ∑ x ∈ P n d x z x 2 = z ⊤ D n z . The resulting exponential trace estimate would require control of exponential moments of the total degree. The degree of a vertex with mark u u u is Poisson with mean at most C 0 u C_0u C 0 u , where C 0 : = c κ M φ , 1 E U C_0:=c_\kappa M_{\varphi,1}\mathbb{E}U C 0 := c κ M φ , 1 E U with c κ c_\kappa c κ defined in (PK). Consequently, for fixed R > 0 R>0 R > 0 ,
E [ e R d x ∣ U x = u ] ≤ exp ( C 0 ( e R − 1 ) u ) . \mathbb{E}[e^{Rd_x}\mid U_x=u]\le\exp(C_0(e^R-1)u). E [ e R d x ∣ U x = u ] ≤ exp ( C 0 ( e R − 1 ) u ) .
Averaging this bound would require E [ e C 0 ( e R − 1 ) U ] < ∞ \mathbb{E}[e^{C_0(e^R-1)U}]<\infty E [ e C 0 ( e R − 1 ) U ] < ∞ . Condition (EM) guarantees a finite exponential moment only at some η \eta η , so this approach does not provide the desired estimate for arbitrary R > 0 R>0 R > 0 .
We instead distinguish neighbors with smaller marks from those with larger marks. For a neighbor mark v ≥ u v\ge u v ≥ u , the factor u v uv uv , by the condition (PK), is bounded by v 2 v^2 v 2 . Thus, a finite second mark moment already gives a uniform bound for the conditional mean of neighbors with no smaller mark. The conditional mean of neighbors with smaller marks may grow linearly in u u u . Orienting edges toward larger marks separates these two contributions. The finite exponential moment assumption (EM) will be used for the incoming contribution after its coefficient has been made sufficiently small.
To specify the orientation also when marks are equal, attach to each Poisson point an independent auxiliary variable T ∼ Unif [ 0 , 1 ] T\sim\operatorname{Unif}[0,1] T ∼ Unif [ 0 , 1 ] , independently of everything else. Write ( U x , T x ) (U_x,T_x) ( U x , T x ) for the enlarged mark of vertex x x x , and orient each existing edge from the smaller pair to the larger pair in lexicographic order. The variable T T T is used only to break ties in U U U , whose distribution may have atoms, including the case U ≡ 1 U\equiv1 U ≡ 1 . Thus, the adjacency matrix is not changed.
For a vertex x x x , let d x + d_x^+ d x + and d x − d_x^- d x − denote its outgoing and incoming degrees, respectively, and let D + D^+ D + and D − D^- D − be the corresponding diagonal matrices. Then D n = D + + D − D_n=D^++D^- D n = D + + D − . For τ > 0 \tau>0 τ > 0 , set
Q τ : = τ D + + τ − 1 D − . Q_\tau:=\tau D^++\tau^{-1}D^-. Q τ := τ D + + τ − 1 D − .
The asymmetric coefficients are chosen to match the different moment bounds for the two degrees. We will take τ \tau τ large below.
Now, for every oriented edge x → y x\to y x → y , we have 2 ∣ z x z y ∣ ≤ τ z x 2 + τ − 1 z y 2 2|z_xz_y|\le\tau z_x^2+\tau^{-1}z_y^2 2∣ z x z y ∣ ≤ τ z x 2 + τ − 1 z y 2 . Summing this inequality over the oriented edges yields
∣ z ⊤ A n z ∣ ≤ ∑ x → y 2 ∣ z x z y ∣ ≤ ∑ x → y ( τ z x 2 + τ − 1 z y 2 ) = z ⊤ Q τ z . |z^\top A_nz|\le\sum_{x\to y}2|z_xz_y|\le\sum_{x\to y}\left(\tau z_x^2+\tau^{-1}z_y^2\right)=z^\top Q_\tau z. ∣ z ⊤ A n z ∣ ≤ x → y ∑ 2∣ z x z y ∣ ≤ x → y ∑ ( τ z x 2 + τ − 1 z y 2 ) = z ⊤ Q τ z .
Thus, z ⊤ ( ± A n ) z ≤ z ⊤ Q τ z z^\top(\pm A_n)z\le z^\top Q_\tau z z ⊤ ( ± A n ) z ≤ z ⊤ Q τ z for every z z z . Recall that if B , C B,C B , C are symmetric matrices of the same size and z ⊤ B z ≤ z ⊤ C z z^\top Bz\le z^\top Cz z ⊤ B z ≤ z ⊤ C z for every z z z , then λ j ( B ) ≤ λ j ( C ) \lambda_j(B)\le\lambda_j(C) λ j ( B ) ≤ λ j ( C ) for every j j j ; see [5 Corollary III.2.3] . Applying this result with B = ± A n B=\pm A_n B = ± A n and C = Q τ C=Q_\tau C = Q τ , we obtain Tr ( e r A n ) + Tr ( e − r A n ) ≤ 2 Tr ( e r Q τ ) = 2 ∑ x ∈ P n e r ( τ d x + + τ − 1 d x − ) \operatorname{Tr}(e^{rA_n})+\operatorname{Tr}(e^{-rA_n})\le2\operatorname{Tr}(e^{rQ_\tau})=2\sum_{x\in P_n}e^{r(\tau d_x^++\tau^{-1}d_x^-)} Tr ( e r A n ) + Tr ( e − r A n ) ≤ 2 Tr ( e r Q τ ) = 2 ∑ x ∈ P n e r ( τ d x + + τ − 1 d x − ) . Consequently,
∣ W n ∣ − 1 E [ Tr ( e r A n ) + Tr ( e − r A n ) ] ≤ 2 ∣ W n ∣ − 1 E ∑ x ∈ P n e r ( τ d x + + τ − 1 d x − ) . (43) |W_n|^{-1}\mathbb{E}\left[\operatorname{Tr}(e^{rA_n})+\operatorname{Tr}(e^{-rA_n})\right]\le2|W_n|^{-1}\mathbb{E}\sum_{x\in P_n}e^{r(\tau d_x^++\tau^{-1}d_x^-)}. \tag*{(43)} ∣ W n ∣ − 1 E [ Tr ( e r A n ) + Tr ( e − r A n ) ] ≤ 2∣ W n ∣ − 1 E x ∈ P n ∑ e r ( τ d x + + τ − 1 d x − ) . (43)
It therefore remains to bound the averaged exponential moment of the oriented degrees.
For x ∈ W n x\in W_n x ∈ W n , u ≥ 1 u\ge1 u ≥ 1 , and t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] , let E x , u , t \mathbb{E}^{x,u,t} E x , u , t denote expectation after inserting a vertex at x x x with enlarged mark ( u , t ) (u,t) ( u , t ) into an independent marked Poisson configuration in W n W_n W n and sampling its incident edges according to the connection rule. Independent Poisson thinning implies that d x + d_x^+ d x + and d x − d_x^- d x − are independent Poisson variables under E x , u , t \mathbb{E}^{x,u,t} E x , u , t . Write their means as m n , + ( x , u , t ) m_{n,+}(x,u,t) m n , + ( x , u , t ) and m n , − ( x , u , t ) m_{n,-}(x,u,t) m n , − ( x , u , t ) . We derive the bounds on these means explicitly. Fix the inserted root ( x , u , t ) (x,u,t) ( x , u , t ) . The remaining, enlarged marked vertices form a Poisson process on W n × [ 1 , ∞ ) × [ 0 , 1 ] W_n\times[1,\infty)\times[0,1] W n × [ 1 , ∞ ) × [ 0 , 1 ] with intensity d y μ ( d v ) d s \mathrm{d}y\,\mu(\mathrm{d}v)\,\mathrm{d}s d y μ ( d v ) d s . For a potential neighbor ( y , v , s ) (y,v,s) ( y , v , s ) , write
p x , u ( y , v ) : = φ ( ∥ y − x ∥ d κ ( u , v ) ) . p_{x,u}(y,v):=\varphi\left(\frac{\|y-x\|^d}{\kappa(u,v)}\right). p x , u ( y , v ) := φ ( κ ( u , v ) ∥ y − x ∥ d ) .
Conditional on the remaining marked configuration, the expected contribution of ( y , v , s ) (y,v,s) ( y , v , s ) to the incoming degree is p x , u ( y , v ) 1 { ( v , s ) ≺ l e x ( u , t ) } p_{x,u}(y,v)\mathbf{1}_{\{(v,s)\prec_{\mathrm{lex}}(u,t)\}} p x , u ( y , v ) 1 {( v , s ) ≺ lex ( u , t )} , and its expected contribution to the outgoing degree is p x , u ( y , v ) 1 { ( u , t ) ≺ l e x ( v , s ) } p_{x,u}(y,v)\mathbf{1}_{\{(u,t)\prec_{\mathrm{lex}}(v,s)\}} p x , u ( y , v ) 1 {( u , t ) ≺ lex ( v , s )} . Averaging these sums over the Poisson configuration gives
m n , − ( x , u , t ) = ∫ W n ∫ [ 1 , ∞ ) ∫ 0 1 p x , u ( y , v ) 1 { ( v , s ) ≺ l e x ( u , t ) } d s μ ( d v ) d y , m_{n,-}(x,u,t)=\int_{W_n}\int_{[1,\infty)}\int_0^1 p_{x,u}(y,v)\mathbf{1}_{\{(v,s)\prec_{\mathrm{lex}}(u,t)\}}\,ds\,\mu(dv)\,dy, m n , − ( x , u , t ) = ∫ W n ∫ [ 1 , ∞ ) ∫ 0 1 p x , u ( y , v ) 1 {( v , s ) ≺ lex ( u , t )} d s μ ( d v ) d y ,
m n , + ( x , u , t ) = ∫ W n ∫ [ 1 , ∞ ) ∫ 0 1 p x , u ( y , v ) 1 { ( u , t ) ≺ l e x ( v , s ) } d s μ ( d v ) d y . m_{n,+}(x,u,t)=\int_{W_n}\int_{[1,\infty)}\int_0^1 p_{x,u}(y,v)\mathbf{1}_{\{(u,t)\prec_{\mathrm{lex}}(v,s)\}}\,ds\,\mu(dv)\,dy. m n , + ( x , u , t ) = ∫ W n ∫ [ 1 , ∞ ) ∫ 0 1 p x , u ( y , v ) 1 {( u , t ) ≺ lex ( v , s )} d s μ ( d v ) d y .
The lexicographic order gives
∫ 0 1 1 { ( v , s ) ≺ l e x ( u , t ) } d s ≤ 1 { v ≤ u } , ∫ 0 1 1 { ( u , t ) ≺ l e x ( v , s ) } d s ≤ 1 { v ≥ u } . \int_0^1\mathbf{1}_{\{(v,s)\prec_{\mathrm{lex}}(u,t)\}}\,ds\leq\mathbf{1}_{\{v\leq u\}},\qquad
\int_0^1\mathbf{1}_{\{(u,t)\prec_{\mathrm{lex}}(v,s)\}}\,ds\leq\mathbf{1}_{\{v\geq u\}}. ∫ 0 1 1 {( v , s ) ≺ lex ( u , t )} d s ≤ 1 { v ≤ u } , ∫ 0 1 1 {( u , t ) ≺ lex ( v , s )} d s ≤ 1 { v ≥ u } .
For fixed u , v ≥ 1 u,v\geq1 u , v ≥ 1 ,
∫ W n p x , u ( y , v ) d y ≤ ∫ R d φ ( ∥ y − x ∥ d κ ( u , v ) ) d y = κ ( u , v ) ∫ R d φ ( ∥ z ∥ d ) d z = κ ( u , v ) M φ , 1 . \int_{W_n}p_{x,u}(y,v)\,dy\leq\int_{\mathbb{R}^d}\varphi\left(\frac{\lVert y-x\rVert^d}{\kappa(u,v)}\right)\,dy=\kappa(u,v)\int_{\mathbb{R}^d}\varphi(\lVert z\rVert^d)\,dz=\kappa(u,v)M_{\varphi,1}. ∫ W n p x , u ( y , v ) d y ≤ ∫ R d φ ( κ ( u , v ) ∥ y − x ∥ d ) d y = κ ( u , v ) ∫ R d φ (∥ z ∥ d ) d z = κ ( u , v ) M φ , 1 .
Define
Λ − : = c κ M φ , 1 E [ U ] , Λ + : = c κ M φ , 1 E [ U 2 ] . \Lambda_-:=c_\kappa M_{\varphi,1}\mathbb{E}[U],\qquad\Lambda_+:=c_\kappa M_{\varphi,1}\mathbb{E}[U^2]. Λ − := c κ M φ , 1 E [ U ] , Λ + := c κ M φ , 1 E [ U 2 ] .
These constants are finite under Condition (EM). Using Condition (PK), we obtain
m n , − ( x , u , t ) ≤ ∫ [ 1 , u ] ( ∫ W n p x , u ( y , v ) d y ) μ ( d v ) ≤ M φ , 1 ∫ [ 1 , u ] κ ( u , v ) μ ( d v ) ≤ c κ M φ , 1 u ∫ [ 1 , u ] v μ ( d v ) = c κ M φ , 1 u E [ U 1 { U ≤ u } ] ≤ Λ − u . \begin{aligned}
m_{n,-}(x,u,t)\leq\int_{[1,u]}\left(\int_{W_n}p_{x,u}(y,v)\,dy\right)\mu(dv)\leq M_{\varphi,1}\int_{[1,u]}\kappa(u,v)\mu(dv) \\
&\leq c_\kappa M_{\varphi,1}u\int_{[1,u]}v\,\mu(dv)=c_\kappa M_{\varphi,1}u\mathbb{E}[U\mathbf{1}_{\{U\leq u\}}]\leq\Lambda_-u.
\end{aligned} m n , − ( x , u , t ) ≤ ∫ [ 1 , u ] ( ∫ W n p x , u ( y , v ) d y ) μ ( d v ) ≤ M φ , 1 ∫ [ 1 , u ] κ ( u , v ) μ ( d v ) ≤ c κ M φ , 1 u ∫ [ 1 , u ] v μ ( d v ) = c κ M φ , 1 u E [ U 1 { U ≤ u } ] ≤ Λ − u .
Similarly, for the outgoing degree,
m n , + ( x , u , t ) ≤ ∫ [ u , ∞ ) ( ∫ W n p x , u ( y , v ) d y ) μ ( d v ) ≤ M φ , 1 ∫ [ u , ∞ ) κ ( u , v ) μ ( d v ) ≤ c κ M φ , 1 u E [ U 1 { U ≥ u } ] ≤ c κ M φ , 1 E [ U 2 1 { U ≥ u } ] ≤ Λ + . \begin{aligned}
m_{n,+}(x,u,t)\leq\int_{[u,\infty)}\left(\int_{W_n}p_{x,u}(y,v)\,dy\right)\mu(dv)\leq M_{\varphi,1}\int_{[u,\infty)}\kappa(u,v)\mu(dv) \\
&\leq c_\kappa M_{\varphi,1}u\mathbb{E}[U\mathbf{1}_{\{U\geq u\}}]\leq c_\kappa M_{\varphi,1}\mathbb{E}[U^2\mathbf{1}_{\{U\geq u\}}]\leq\Lambda_+.
\end{aligned} m n , + ( x , u , t ) ≤ ∫ [ u , ∞ ) ( ∫ W n p x , u ( y , v ) d y ) μ ( d v ) ≤ M φ , 1 ∫ [ u , ∞ ) κ ( u , v ) μ ( d v ) ≤ c κ M φ , 1 u E [ U 1 { U ≥ u } ] ≤ c κ M φ , 1 E [ U 2 1 { U ≥ u } ] ≤ Λ + .
The Poisson exponential-moment formula now gives, for α , β ≥ 0 \alpha,\beta\geq0 α , β ≥ 0 ,
E x , u , t [ e α d x + + β d x − ] = exp { m n , + ( x , u , t ) ( e α − 1 ) + m n , − ( x , u , t ) ( e β − 1 ) } ≤ exp { Λ + ( e α − 1 ) + Λ − u ( e β − 1 ) } . \mathbb{E}^{x,u,t}\left[e^{\alpha d_x^++\beta d_x^-}\right]=\exp\left\{m_{n,+}(x,u,t)(e^\alpha-1)+m_{n,-}(x,u,t)(e^\beta-1)\right\}\leq\exp\left\{\Lambda_+(e^\alpha-1)+\Lambda_-u(e^\beta-1)\right\}. E x , u , t [ e α d x + + β d x − ] = exp { m n , + ( x , u , t ) ( e α − 1 ) + m n , − ( x , u , t ) ( e β − 1 ) } ≤ exp { Λ + ( e α − 1 ) + Λ − u ( e β − 1 ) } .
{ ( X i , U i , T i ) : 1 ≤ i ≤ N n } \{(X_i,U_i,T_i):1\leq i\leq N_n\} {( X i , U i , T i ) : 1 ≤ i ≤ N n } is Poisson with intensity d x μ ( d u ) d t dx\,\mu(du)\,dt d x μ ( d u ) d t . By Mecke formula, using the above bound,
∣ W n ∣ − 1 E ∑ x ∈ P n e α d x + + β d x − = ∣ W n ∣ − 1 ∫ W n ∫ [ 1 , ∞ ) ∫ 0 1 E x , u , t [ e α d x + + β d x − ] d t μ ( d u ) d x ≤ ∣ W n ∣ − 1 ∫ W n ∫ [ 1 , ∞ ) ∫ 0 1 e Λ + ( e α − 1 ) + Λ − u ( e β − 1 ) d t μ ( d u ) d x = e Λ + ( e α − 1 ) E [ e Λ − U ( e β − 1 ) ] . (44) \begin{aligned}
|W_n|^{-1}\mathbb{E}\sum_{x\in\mathcal{P}_n}e^{\alpha d_x^++\beta d_x^-}
=|W_n|^{-1}\int_{W_n}\int_{[1,\infty)}\int_0^1\mathbb{E}^{x,u,t}\left[e^{\alpha d_x^++\beta d_x^-}\right]\,dt\,\mu(du)\,dx \\
&\leq|W_n|^{-1}\int_{W_n}\int_{[1,\infty)}\int_0^1e^{\Lambda_+(e^\alpha-1)+\Lambda_-u(e^\beta-1)}\,dt\,\mu(du)\,dx \\
&=e^{\Lambda_+(e^\alpha-1)}\mathbb{E}\left[e^{\Lambda_-U(e^\beta-1)}\right]. \tag*{(44)}
\end{aligned} ∣ W n ∣ − 1 E x ∈ P n ∑ e α d x + + β d x − = ∣ W n ∣ − 1 ∫ W n ∫ [ 1 , ∞ ) ∫ 0 1 E x , u , t [ e α d x + + β d x − ] d t μ ( d u ) d x ≤ ∣ W n ∣ − 1 ∫ W n ∫ [ 1 , ∞ ) ∫ 0 1 e Λ + ( e α − 1 ) + Λ − u ( e β − 1 ) d t μ ( d u ) d x = e Λ + ( e α − 1 ) E [ e Λ − U ( e β − 1 ) ] . (44)
Choose τ R ≥ 1 \tau_R\geq1 τ R ≥ 1 sufficiently large that Λ − ( e R / τ R − 1 ) ≤ η / 2 \Lambda_-(e^{R/\tau_R}-1)\leq\eta/2 Λ − ( e R / τ R − 1 ) ≤ η /2 . For 0 ≤ r ≤ R 0\leq r\leq R 0 ≤ r ≤ R , set α = r τ R \alpha=r\tau_R α = r τ R and β = r / τ R \beta=r/\tau_R β = r / τ R . The right-hand side is then at most exp { Λ + ( e r τ R − 1 ) } E [ e η U / 2 ] < ∞ \exp\{\Lambda_+(e^{r\tau_R}-1)\}\mathbb{E}[e^{\eta U/2}]<\infty exp { Λ + ( e r τ R − 1 )} E [ e η U /2 ] < ∞ . Inserting this into (40) proves (36).
We next prove (37). For 0 ≤ r ≤ R 0\leq r\leq R 0 ≤ r ≤ R , set B r : = A n 2 e r A n B_r:=A_n^2e^{rA_n} B r := A n 2 e r A n . The matrix B r B_r B r is positive semi-definite because its eigenvalues are λ 2 e r λ \lambda^2e^{r\lambda} λ 2 e r λ where λ \lambda λ are the eigenvalues of A n A_n A n . The Cauchy–Schwarz inequality (in Hilbert–Schmidt norm) gives Tr ( D n B r ) ≤ Tr ( D n 2 ) 1 / 2 Tr ( B r 2 ) 1 / 2 \operatorname{Tr}(D_nB_r)\leq\operatorname{Tr}(D_n^2)^{1/2}\operatorname{Tr}(B_r^2)^{1/2} Tr ( D n B r ) ≤ Tr ( D n 2 ) 1/2 Tr ( B r 2 ) 1/2 . Taking expectations and applying Cauchy–Schwarz,
E [ Tr ( D n A n 2 e r A n ) ] ≤ E [ Tr ( D n 2 ) ] 1 / 2 E [ Tr ( B r 2 ) ] 1 / 2 . (45) \mathbb{E}\left[\operatorname{Tr}(D_nA_n^2e^{rA_n})\right]\leq\mathbb{E}\left[\operatorname{Tr}(D_n^2)\right]^{1/2}\mathbb{E}\left[\operatorname{Tr}(B_r^2)\right]^{1/2}. \tag*{(45)} E [ Tr ( D n A n 2 e r A n ) ] ≤ E [ Tr ( D n 2 ) ] 1/2 E [ Tr ( B r 2 ) ] 1/2 . (45)
We bound each factor respectively. For a root inserted at x x x with mark u u u , its total degree has Poisson mean
m n ( x , u ) = ∫ W n ∫ [ 1 , ∞ ) φ ( ∥ x − y ∥ d κ ( u , v ) ) μ ( d v ) d y ≤ c κ M φ , 1 u E U = Λ − u . m_n(x,u)=\int_{W_n}\int_{[1,\infty)}\varphi\left(\frac{\lVert x-y\rVert^d}{\kappa(u,v)}\right)\mu(dv)\,dy\leq c_\kappa M_{\varphi,1}u\mathbb{E}U=\Lambda_-u. m n ( x , u ) = ∫ W n ∫ [ 1 , ∞ ) φ ( κ ( u , v ) ∥ x − y ∥ d ) μ ( d v ) d y ≤ c κ M φ , 1 u E U = Λ − u .
This is a bound for the total mean itself. Thus, E x , u , t [ d x 2 ] ≤ Λ − u + Λ 2 u 2 \mathbb{E}^{x,u,t}[d_x^2] \le\Lambda^{-u}+\Lambda^2u^2 E x , u , t [ d x 2 ] ≤ Λ − u + Λ 2 u 2 , uniformly also in the auxiliary root mark t t t . Applying Mecke formula,
E Tr ( D n 2 ) = ∫ W n ∫ [ 1 , ∞ ) ∫ 0 1 E x , u , t [ d x 2 ] d t μ ( d u ) d x ≤ ∣ W n ∣ ( Λ − E U + Λ 2 E [ U 2 ] ) ≤ C ∣ W n ∣ . \mathbb{E}\operatorname{Tr}(D_n^2)=\int_{W_n}\int_{[1,\infty)}\int_0^1\mathbb{E}^{x,u,t}[d_x^2]\,\mathrm{d}t\,\mu(\mathrm{d}u)\,\mathrm{d}x\le|W_n|\left(\Lambda_{-}\mathbb{E}U+\Lambda^2\mathbb{E}[U^2]\right)\le C|W_n|. E Tr ( D n 2 ) = ∫ W n ∫ [ 1 , ∞ ) ∫ 0 1 E x , u , t [ d x 2 ] d t μ ( d u ) d x ≤ ∣ W n ∣ ( Λ − E U + Λ 2 E [ U 2 ] ) ≤ C ∣ W n ∣.
To control the second factor in (45) , choose S > 2 R S>2R S > 2 R . There is a constant C R , S < ∞ C_{R,S}<\infty C R , S < ∞ such that
sup 0 ≤ r ≤ R λ 4 e 2 r λ ≤ C R , S ( e S λ + e − S λ ) (46) \sup_{0\le r\le R}\lambda^{4}e^{2r\lambda}\le C_{R,S}\left(e^{S\lambda}+e^{-S\lambda}\right) \tag*{(46)} 0 ≤ r ≤ R sup λ 4 e 2 r λ ≤ C R , S ( e S λ + e − S λ ) (46)
for every λ ∈ R \lambda\in\mathbb{R} λ ∈ R . Indeed, for λ ≥ 0 \lambda\ge0 λ ≥ 0 , the exponential e S λ e^{S\lambda} e S λ absorbs λ 4 e 2 R λ \lambda^4e^{2R\lambda} λ 4 e 2 R λ because S > 2 R S>2R S > 2 R . For λ < 0 \lambda<0 λ < 0 , we have e 2 r λ ≤ 1 e^{2r\lambda}\le1 e 2 r λ ≤ 1 , and e − S λ e^{-S\lambda} e − S λ absorbs the polynomial factor. Since A n A_n A n commutes with e r A n e^{rA_n} e r A n , we have B r 2 = A n 4 e 2 r A n B_r^2=A_n^4e^{2rA_n} B r 2 = A n 4 e 2 r A n . Applying the inequality (46) to the eigenvalues of A n A_n A n gives
sup 0 ≤ r ≤ R Tr ( B r 2 ) ≤ C R , S ( Tr ( e S A n ) + Tr ( e − S A n ) ) . \sup_{0\le r\le R}\operatorname{Tr}(B_r^2)\le C_{R,S}\left(\operatorname{Tr}(e^{SA_n})+\operatorname{Tr}(e^{-SA_n})\right). 0 ≤ r ≤ R sup Tr ( B r 2 ) ≤ C R , S ( Tr ( e S A n ) + Tr ( e − S A n ) ) .
Therefore, (39) , applied with S S S in place of R R R , implies sup n ≥ 1 sup 0 ≤ r ≤ R ∣ W n ∣ − 1 E [ Tr ( B r 2 ) ] < ∞ \sup_{n\ge1}\sup_{0\le r\le R}|W_n|^{-1}\mathbb{E}[\operatorname{Tr}(B_r^2)]<\infty sup n ≥ 1 sup 0 ≤ r ≤ R ∣ W n ∣ − 1 E [ Tr ( B r 2 )] < ∞ . Inserting this and sup n ≥ 1 ∣ W n ∣ − 1 E [ Tr ( D n 2 ) ] < ∞ \sup_{n\ge1}|W_n|^{-1}\mathbb{E}[\operatorname{Tr}(D_n^2)]<\infty sup n ≥ 1 ∣ W n ∣ − 1 E [ Tr ( D n 2 )] < ∞ into (45) yields (40) .
We now prove the variance estimate for exponentially twisted linear statistics in Proposition 15 . The proof follows the reduction of [18 ] ; the all-walk series estimate [18, Lem. 5.3] is replaced by the trace bounds of Lemma 16. We sketch the proof and refer the reader to the proof of [18 Thm. 5.2] for details.
Proof of Proposition 15 . Set H n ( f ) : = Tr [ f ( A n ) e c A n ] H_n(f):=\operatorname{Tr}[f(A_n)e^{cA_n}] H n ( f ) := Tr [ f ( A n ) e c A n ] . For x ˇ ∈ W ˇ n \check{x}\in\check{W}_n x ˇ ∈ W ˇ n , we identify A n A_n A n with the adjacency matrix of G ∅ ( ξ n , ( x ˇ ) ) G_{\varnothing}(\xi_n,(\check{x})) G ∅ ( ξ n , ( x ˇ )) , and we write A ^ n : = A n ⊕ 0 \widehat{A}_n:=A_n\oplus0 A n := A n ⊕ 0 for the matrix corresponding to adding one isolated vertex at x ˇ \check{x} x ˇ to G ( ξ n ) G(\xi_n) G ( ξ n ) . Use the same ordering of the old vertices in A n x ˇ A_n^{\check{x}} A n x ˇ and A ^ n \widehat{A}_n A n , with the inserted vertex placed in the final matrix coordinate. This matrix-coordinate convention is independent of its reserved label in the edge-mark array. In these coordinates, the upper-left block of A n x ˇ A_n^{\check{x}} A n x ˇ is exactly A n A_n A n .
Since f ( 0 ) = 0 f(0)=0 f ( 0 ) = 0 , the isolated vertex contributes f ( 0 ) e 0 = 0 f(0)e^0=0 f ( 0 ) e 0 = 0 to the trace, and therefore, Tr [ f ( A ^ n ) e c A ^ n ] = Tr [ f ( A n ) e c A n ] \operatorname{Tr}[f(\widehat{A}_n)e^{c\widehat{A}_n}]=\operatorname{Tr}[f(A_n)e^{cA_n}] Tr [ f ( A n ) e c A n ] = Tr [ f ( A n ) e c A n ] . Thus, Δ x ˇ ; ( x ˇ ) H n ( f ) = Tr [ f ( A n x ˇ ) e c A n x ˇ ] − Tr [ f ( A ^ n ) e c A ^ n ] \Delta_{\check{x};(\check{x})}H_n(f)=\operatorname{Tr}[f(A_n^{\check{x}})e^{cA_n^{\check{x}}}]-\operatorname{Tr}[f(\widehat{A}_n)e^{c\widehat{A}_n}] Δ x ˇ ; ( x ˇ ) H n ( f ) = Tr [ f ( A n x ˇ ) e c A n x ˇ ] − Tr [ f ( A n ) e c A n ] . Let B x ˇ : = A n x ˇ − A ^ n B_{\check{x}}:=A_n^{\check{x}}-\widehat{A}_n B x ˇ := A n x ˇ − A n . By the construction of G { 1 } ( ξ n , x ˇ ) G_{\{1\}}(\xi_n,\check{x}) G { 1 } ( ξ n , x ˇ ) , the induced graph on P n \mathcal{P}_n P n is exactly G ∅ ( ξ n , ( x ˇ ) ) G_{\varnothing}(\xi_n,(\check{x})) G ∅ ( ξ n , ( x ˇ )) . Hence, B x ˇ B_{\check{x}} B x ˇ is supported only on the row and column corresponding to the inserted vertex x ˇ \check{x} x ˇ . Let e x ˇ e_{\check{x}} e x ˇ be the coordinate vector of the inserted vertex. Since A ^ n e x ˇ = 0 \widehat{A}_ne_{\check{x}}=0 A n e x ˇ = 0 , we have B x ˇ e x ˇ = A n x ˇ e x ˇ = a x ˇ B_{\check{x}}e_{\check{x}}=A_n^{\check{x}}e_{\check{x}}=a_{\check{x}} B x ˇ e x ˇ = A n x ˇ e x ˇ = a x ˇ , where this vector is the same as the above defined a x ˇ a_{\check{x}} a x ˇ . It is supported on the neighbors of x ˇ \check{x} x ˇ in G { 1 } ( ξ n , x ˇ ) G_{\{1\}}(\xi_n,\check{x}) G { 1 } ( ξ n , x ˇ ) , and B x ˇ = e x ˇ a x ˇ T + a x ˇ e x ˇ T B_{\check{x}}=e_{\check{x}}a_{\check{x}}^{\mathsf{T}}+a_{\check{x}}e_{\check{x}}^{\mathsf{T}} B x ˇ = e x ˇ a x ˇ T + a x ˇ e x ˇ T . For fixed n n n , we have H n ( f ) ∈ L 2 H_n(f)\in L^2 H n ( f ) ∈ L 2 . Indeed, since H 2 ( R ) ⊆ C b ( R ) H^2(\mathbb{R})\subseteq C_b(\mathbb{R}) H 2 ( R ) ⊆ C b ( R ) and ∥ A n ∥ o p ≤ N n \|A_n\|_{\mathrm{op}}\le N_n ∥ A n ∥ op ≤ N n , we have ∣ H n ( f ) ∣ ≤ ∥ f ∥ ∞ N n e ∣ c ∣ N n |H_n(f)|\le\|f\|_\infty N_ne^{|c|N_n} ∣ H n ( f ) ∣ ≤ ∥ f ∥ ∞ N n e ∣ c ∣ N n , whose square has finite expectation because N n N_n N n is Poisson. Hence, the Poincaré inequality from Theorem 7 gives Var ( H n ( f ) ) ≤ ∫ W ^ n E [ ( Δ x ˇ ; ( x ˇ ) H n ( f ) ) 2 ] λ ( d x ˇ ) \operatorname{Var}(H_n(f))\le\int_{\widehat{W}_n}\mathbb{E}[(\Delta_{\check{x};(\check{x})}H_n(f))^2]\,\lambda(\mathrm{d}\check{x}) Var ( H n ( f )) ≤ ∫ W n E [( Δ x ˇ ; ( x ˇ ) H n ( f ) ) 2 ] λ ( d x ˇ ) .
For f ∈ L 1 ( R ) f\in L^1(\mathbb{R}) f ∈ L 1 ( R ) , we use the Fourier transform f ^ ( ξ ) : = 1 2 π ∫ R f ( x ) e − i ξ x d x \widehat{f}(\xi):=\frac{1}{2\pi}\int_{\mathbb{R}}f(x)e^{-i\xi x}\,\mathrm{d}x f ( ξ ) := 2 π 1 ∫ R f ( x ) e − i ξ x d x , so that f ( x ) = ∫ R f ^ ( ξ ) e i ξ x d ξ f(x)=\int_{\mathbb{R}}\widehat{f}(\xi)e^{i\xi x}\,\mathrm{d}\xi f ( x ) = ∫ R f ( ξ ) e i ξ x d ξ whenever Fourier inversion applies. To ensure this, assume first that f ∈ L 1 ( R ) ∩ H 2 ( R ) f\in L^1(\mathbb{R})\cap H^2(\mathbb{R}) f ∈ L 1 ( R ) ∩ H 2 ( R ) , and put
I c ( f ) : = ∫ R ∣ f ^ ( ξ ) ∣ ( ∣ ξ ∣ + ∣ c ∣ ) d ξ . I_c(f):=\int_{\mathbb{R}}|\widehat{f}(\xi)|(|\xi|+|c|)\,\mathrm{d}\xi. I c ( f ) := ∫ R ∣ f ( ξ ) ∣ ( ∣ ξ ∣ + ∣ c ∣ ) d ξ .
For s = c + i ξ s=c+i\xi s = c + i ξ , Fourier inversion gives Tr [ f ( A n x ˇ ) e c A n x ˇ ] − Tr [ f ( A ^ n ) e c A ^ n ] = ∫ R f ^ ( ξ ) Tr [ e s A n x ˇ − e s A ^ n ] d ξ \operatorname{Tr}[f(A_n^{\check{x}})e^{cA_n^{\check{x}}}]-\operatorname{Tr}[f(\widehat{A}_n)e^{c\widehat{A}_n}]=\int_{\mathbb{R}}\widehat{f}(\xi)\operatorname{Tr}[e^{sA_n^{\check{x}}}-e^{s\widehat{A}_n}]\,\mathrm{d}\xi Tr [ f ( A n x ˇ ) e c A n x ˇ ] − Tr [ f ( A n ) e c A n ] = ∫ R f ( ξ ) Tr [ e s A n x ˇ − e s A n ] d ξ . As in [18 Eq. (50)] , Duhamel’s formula yields that Tr [ e s A n x ˇ − e s A ^ n ] = s ∫ 0 1 Tr [ e t s A n x ˇ B x ˇ e ( 1 − t ) s A ^ n ] d t \operatorname{Tr}[e^{sA_n^{\check{x}}}-e^{s\widehat{A}_n}]=s\int_0^1\operatorname{Tr}[e^{tsA_n^{\check{x}}}B_{\check{x}}e^{(1-t)s\widehat{A}_n}]\,\mathrm{d}t Tr [ e s A n x ˇ − e s A n ] = s ∫ 0 1 Tr [ e t s A n x ˇ B x ˇ e ( 1 − t ) s A n ] d t . Using the representation of B x ˇ B_{\check{x}} B x ˇ and the rank-one trace identity Tr ( M u v T N ) = v T N M u \operatorname{Tr}(Muv^{\mathsf{T}}N)=v^{\mathsf{T}}NM u Tr ( M u v T N ) = v T N M u , we obtain Tr [ e t s A n x ˇ B x ˇ e ( 1 − t ) s A ^ n ] = a x ˇ T e ( 1 − t ) s A ^ n e t s A n x ˇ e x ˇ + e x ˇ T e ( 1 − t ) s A ^ n e t s A n x ˇ a x ˇ \operatorname{Tr}[e^{tsA_n^{\check{x}}}B_{\check{x}}e^{(1-t)s\widehat{A}_n}]=a_{\check{x}}^{\mathsf{T}}e^{(1-t)s\widehat{A}_n}e^{tsA_n^{\check{x}}}e_{\check{x}}+e_{\check{x}}^{\mathsf{T}}e^{(1-t)s\widehat{A}_n}e^{tsA_n^{\check{x}}}a_{\check{x}} Tr [ e t s A n x ˇ B x ˇ e ( 1 − t ) s A n ] = a x ˇ T e ( 1 − t ) s A n e t s A n x ˇ e x ˇ + e x ˇ T e ( 1 − t ) s A n e t s A n x ˇ a x ˇ . Cauchy’s inequality therefore gives
∣ Tr [ e t s A n x ˇ B x ˇ e ( 1 − t ) s A ^ n ] ∣ ≤ ∥ e t s A n x ˇ e x ˇ ∥ 2 ∥ a x ˇ T e ( 1 − t ) s A ^ n ∥ 2 + ∥ e t s A n x ˇ a x ˇ ∥ 2 ∥ e ( 1 − t ) s A ^ n e x ˇ ∥ 2 . \left|\operatorname{Tr}[e^{tsA_n^{\check{x}}}B_{\check{x}}e^{(1-t)s\widehat{A}_n}]\right|\le\|e^{tsA_n^{\check{x}}}e_{\check{x}}\|_2\|a_{\check{x}}^{\mathsf{T}}e^{(1-t)s\widehat{A}_n}\|_2+\|e^{tsA_n^{\check{x}}}a_{\check{x}}\|_2\|e^{(1-t)s\widehat{A}_ne_{\check{x}}}\|_2. Tr [ e t s A n x ˇ B x ˇ e ( 1 − t ) s A n ] ≤ ∥ e t s A n x ˇ e x ˇ ∥ 2 ∥ a x ˇ T e ( 1 − t ) s A n ∥ 2 + ∥ e t s A n x ˇ a x ˇ ∥ 2 ∥ e ( 1 − t ) s A n e x ˇ ∥ 2 .
Since A n x ˇ A_n^{\check{x}} A n x ˇ and A ^ n \widehat{A}_n A n are real symmetric, the unitary parts e i t ξ A n x ˇ e^{it\xi A_n^{\check{x}}} e i t ξ A n x ˇ and e i ( 1 − t ) ξ A ^ n e^{i(1-t)\xi\widehat{A}_n} e i ( 1 − t ) ξ A n do not change Euclidean norms. The vectors e x ˇ e_{\check{x}} e x ˇ and a x ˇ a_{\check{x}} a x ˇ have non-negative entries, and both adjacency matrices have non-negative entries. Hence, for either of these vectors v v v , the exponential series gives the entrywise bound ∣ ( e t c A v ) i ∣ ≤ ( e t ∣ c ∣ A v ) i |(e^{tcA}v)_i|\le(e^{t|c|A}v)_i ∣ ( e t c A v ) i ∣ ≤ ( e t ∣ c ∣ A v ) i . Moreover, A ^ n e x ˇ = 0 \widehat{A}_ne_{\check{x}}=0 A n e x ˇ = 0 . Thus, the preceding display is bounded by X t ( x ˇ ) 1 / 2 Y ‾ 1 − t ( x ˇ ) 1 / 2 + Y t ( x ˇ ) 1 / 2 X_t(\check{x})^{1/2}\overline{Y}_{1-t}(\check{x})^{1/2}+Y_t(\check{x})^{1/2} X t ( x ˇ ) 1/2 Y 1 − t ( x ˇ ) 1/2 + Y t ( x ˇ ) 1/2 , where
X t ( x ˇ ) : = ⟨ e 2 t ∣ c ∣ A n x ˇ e x ˇ , e x ˇ ⟩ , Y ‾ t ( x ˇ ) : = ⟨ e 2 t ∣ c ∣ A ^ n a x ˇ , a x ˇ ⟩ , Y t ( x ˇ ) : = ⟨ e 2 t ∣ c ∣ A n x ˇ a x ˇ , a x ˇ ⟩ . X_t(\check{x}):=\langle e^{2t|c|A_n^{\check{x}}}e_{\check{x}},e_{\check{x}}\rangle,\qquad\overline{Y}_t(\check{x}):=\langle e^{2t|c|\widehat{A}_n}a_{\check{x}},a_{\check{x}}\rangle,\qquad Y_t(\check{x}):=\langle e^{2t|c|A_n^{\check{x}}}a_{\check{x}},a_{\check{x}}\rangle. X t ( x ˇ ) := ⟨ e 2 t ∣ c ∣ A n x ˇ e x ˇ , e x ˇ ⟩ , Y t ( x ˇ ) := ⟨ e 2 t ∣ c ∣ A n a x ˇ , a x ˇ ⟩ , Y t ( x ˇ ) := ⟨ e 2 t ∣ c ∣ A n x ˇ a x ˇ , a x ˇ ⟩ .
Consequently, using ∣ s ∣ ≤ ∣ ξ ∣ + ∣ c ∣ |s| \le|\xi| + |c| ∣ s ∣ ≤ ∣ ξ ∣ + ∣ c ∣ together with Jensen’s inequality in the t t t -variable, we obtain the analogue of [18 Proof of Theorem 5.2, Step 1] :
E [ ( Δ x ~ ; ( x ~ ) H n ( f ) ) 2 ] ≤ C I c ( f ) 2 ∫ 0 1 E [ X t ( x ~ ) Y ‾ 1 − t ( x ~ ) + Y t ( x ~ ) ] d t . \mathbb{E}\left[(\Delta_{\tilde{x};(\tilde{x})}H_n(f))^2\right] \le C I_c(f)^2 \int_0^1 \mathbb{E}\left[X_t(\tilde{x})\overline{Y}_{1-t}(\tilde{x})+Y_t(\tilde{x})\right]\,dt. E [ ( Δ x ~ ; ( x ~ ) H n ( f ) ) 2 ] ≤ C I c ( f ) 2 ∫ 0 1 E [ X t ( x ~ ) Y 1 − t ( x ~ ) + Y t ( x ~ ) ] d t .
Since A ~ n x ~ \widetilde{A}_n^{\tilde{x}} A n x ~ and A ^ n \widehat{A}_n A n have non-negative entries and A ^ n ≤ A ~ n x ~ \widehat{A}_n \le\widetilde{A}_n^{\tilde{x}} A n ≤ A n x ~ entry-wise, the exponential series gives e α A ^ n ≤ e α A ~ n x ~ e^{\alpha\widehat{A}_n} \le e^{\alpha\widetilde{A}_n^{\tilde{x}}} e α A n ≤ e α A n x ~ entry-wise for every α ≥ 0 \alpha\ge0 α ≥ 0 . As a x ~ a_{\tilde{x}} a x ~ has non-negative entries, it follows that Y ‾ t ( x ~ ) ≤ Y t ( x ~ ) \overline{Y}_t(\tilde{x}) \le Y_t(\tilde{x}) Y t ( x ~ ) ≤ Y t ( x ~ ) . We next bound the two matrix elements directly. Cauchy’s inequality gives X t ( x ~ ) 2 ≤ ⟨ e 4 t ∣ c ∣ ∣ A ^ n ∣ e ~ x ~ , e ~ x ~ ⟩ X_t(\tilde{x})^2 \le\langle e^{4t|c|\left|\widehat{A}_n\right|}\tilde{e}_{\tilde{x}},\tilde{e}_{\tilde{x}}\rangle X t ( x ~ ) 2 ≤ ⟨ e 4 t ∣ c ∣ ∣ A n ∣ e ~ x ~ , e ~ x ~ ⟩ . Moreover, with a x ~ = A ^ n x ~ e ~ x ~ a_{\tilde{x}}=\widehat{A}_n^{\tilde{x}}\tilde{e}_{\tilde{x}} a x ~ = A n x ~ e ~ x ~ , we have Y t ( x ~ ) 2 ≤ ⟨ e 4 t ∣ c ∣ A ^ n x ~ a x ~ , a x ~ ⟩ / ⟨ a x ~ , a x ~ ⟩ Y_t(\tilde{x})^2 \le\langle e^{4t|c|\widehat{A}_n^{\tilde{x}}}a_{\tilde{x}},a_{\tilde{x}}\rangle/\langle a_{\tilde{x}},a_{\tilde{x}}\rangle Y t ( x ~ ) 2 ≤ ⟨ e 4 t ∣ c ∣ A n x ~ a x ~ , a x ~ ⟩ / ⟨ a x ~ , a x ~ ⟩ . Since 4 t ∣ c ∣ ≤ 4 ∣ c ∣ 4t|c| \le4|c| 4 t ∣ c ∣ ≤ 4∣ c ∣ , Lemma 16, with R = 4 ∣ c ∣ R=4|c| R = 4∣ c ∣ , yields
sup 0 ≤ t ≤ 1 ∫ W ~ n E [ X t ( x ~ ) 2 ] λ ( d x ~ ) + sup 0 ≤ t ≤ 1 ∫ W ~ n E [ Y t ( x ~ ) 2 ] λ ( d x ~ ) ∈ O ( ∣ W n ∣ ) . \sup_{0\le t\le1}\int_{\widetilde{W}_n}\mathbb{E}\left[X_t(\tilde{x})^2\right]\lambda(d\tilde{x})+\sup_{0\le t\le1}\int_{\widetilde{W}_n}\mathbb{E}\left[Y_t(\tilde{x})^2\right]\lambda(d\tilde{x})\in O(|W_n|). 0 ≤ t ≤ 1 sup ∫ W n E [ X t ( x ~ ) 2 ] λ ( d x ~ ) + 0 ≤ t ≤ 1 sup ∫ W n E [ Y t ( x ~ ) 2 ] λ ( d x ~ ) ∈ O ( ∣ W n ∣ ) .
Combining these two estimates and using Cauchy’s inequality gives
∫ W ~ n E [ X t ( x ~ ) Y ‾ 1 − t ( x ~ ) ] λ ( d x ~ ) ≤ ∫ W ~ n E [ X t ( x ~ ) Y 1 − t ( x ~ ) ] λ ( d x ~ ) ≤ ∫ W ~ n E [ X t ( x ~ ) 2 ] λ ( d x ~ ) ∫ W ~ n E [ Y 1 − t ( x ~ ) 2 ] λ ( d x ~ ) \int_{\widetilde{W}_n}\mathbb{E}\left[X_t(\tilde{x})\overline{Y}_{1-t}(\tilde{x})\right]\lambda(d\tilde{x})\le\int_{\widetilde{W}_n}\mathbb{E}\left[X_t(\tilde{x})Y_{1-t}(\tilde{x})\right]\lambda(d\tilde{x})\le\sqrt{\int_{\widetilde{W}_n}\mathbb{E}\left[X_t(\tilde{x})^2\right]\lambda(d\tilde{x})}\sqrt{\int_{\widetilde{W}_n}\mathbb{E}\left[Y_{1-t}(\tilde{x})^2\right]\lambda(d\tilde{x})} ∫ W n E [ X t ( x ~ ) Y 1 − t ( x ~ ) ] λ ( d x ~ ) ≤ ∫ W n E [ X t ( x ~ ) Y 1 − t ( x ~ ) ] λ ( d x ~ ) ≤ ∫ W n E [ X t ( x ~ ) 2 ] λ ( d x ~ ) ∫ W n E [ Y 1 − t ( x ~ ) 2 ] λ ( d x ~ )
which is O ( ∣ W n ∣ ) O(|W_n|) O ( ∣ W n ∣ ) . Moreover, since λ = d x ⊗ μ \lambda=dx\otimes\mu λ = d x ⊗ μ on W ~ n \widetilde{W}_n W n and μ \mu μ is a probability measure, λ ( W ~ n ) = ∣ W n ∣ μ ( [ 1 , ∞ ) ) = ∣ W n ∣ \lambda(\widetilde{W}_n)=|W_n|\mu([1,\infty))=|W_n| λ ( W n ) = ∣ W n ∣ μ ([ 1 , ∞ )) = ∣ W n ∣ . Hence, Cauchy’s inequality and the second-moment estimate for Y t Y_t Y t give
∫ W ~ n E [ Y t ( x ~ ) ] λ ( d x ~ ) ≤ λ ( W ~ n ) ∫ W ~ n E [ Y t ( x ~ ) 2 ] λ ( d x ~ ) ∈ O ( ∣ W n ∣ ) . \int_{\widetilde{W}_n}\mathbb{E}\left[Y_t(\tilde{x})\right]\lambda(d\tilde{x})\le\sqrt{\lambda(\widetilde{W}_n)}\sqrt{\int_{\widetilde{W}_n}\mathbb{E}\left[Y_t(\tilde{x})^2\right]\lambda(d\tilde{x})}\in O(|W_n|). ∫ W n E [ Y t ( x ~ ) ] λ ( d x ~ ) ≤ λ ( W n ) ∫ W n E [ Y t ( x ~ ) 2 ] λ ( d x ~ ) ∈ O ( ∣ W n ∣ ) .
Therefore, ∫ W ~ n E [ ( Δ x ~ ; ( x ~ ) H n ( f ) ) 2 ] λ ( d x ~ ) ≤ C ∣ W n ∣ I c ( f ) 2 \int_{\widetilde{W}_n}\mathbb{E}\left[(\Delta_{\tilde{x};(\tilde{x})}H_n(f))^2\right]\lambda(d\tilde{x})\le C|W_n|I_c(f)^2 ∫ W n E [ ( Δ x ~ ; ( x ~ ) H n ( f ) ) 2 ] λ ( d x ~ ) ≤ C ∣ W n ∣ I c ( f ) 2 . By Cauchy–Schwarz and Plancherel’s identity,
I c ( f ) 2 = ( ∫ R ∣ f ^ ( ξ ) ∣ ( ∣ ξ ∣ + ∣ c ∣ ) d ξ ) 2 ≤ ( ∫ R ( 1 + ξ 4 ) ∣ f ^ ( ξ ) ∣ 2 d ξ ) ( ∫ R ( ∣ ξ ∣ + ∣ c ∣ ) 2 1 + ξ 4 d ξ ) ≤ C ( ∥ f ∥ 2 2 + ∥ f ′ ′ ∥ 2 2 ) . I_c(f)^2=\left(\int_{\mathbb{R}}|\widehat{f}(\xi)|(|\xi|+|c|)\,d\xi\right)^2\le\left(\int_{\mathbb{R}}(1+\xi^4)|\widehat{f}(\xi)|^2\,d\xi\right)\left(\int_{\mathbb{R}}\frac{(|\xi|+|c|)^2}{1+\xi^4}\,d\xi\right)\le C\left(\|f\|_2^2+\|f''\|_2^2\right). I c ( f ) 2 = ( ∫ R ∣ f ( ξ ) ∣ ( ∣ ξ ∣ + ∣ c ∣ ) d ξ ) 2 ≤ ( ∫ R ( 1 + ξ 4 ) ∣ f ( ξ ) ∣ 2 d ξ ) ( ∫ R 1 + ξ 4 ( ∣ ξ ∣ + ∣ c ∣ ) 2 d ξ ) ≤ C ( ∥ f ∥ 2 2 + ∥ f ′′ ∥ 2 2 ) .
Combining this estimate with the Poincaré inequality proves the assertion for f ∈ L 1 ( R ) ∩ H 2 ( R ) f\in L^1(\mathbb{R})\cap H^2(\mathbb{R}) f ∈ L 1 ( R ) ∩ H 2 ( R ) with f ( 0 ) = 0 f(0)=0 f ( 0 ) = 0 .
Now let f ∈ H 2 ( R ) f\in H^2(\mathbb{R}) f ∈ H 2 ( R ) satisfy f ( 0 ) = 0 f(0)=0 f ( 0 ) = 0 . Choose χ ∈ C ∞ ( R ) \chi\in C^\infty(\mathbb{R}) χ ∈ C ∞ ( R ) with 0 ≤ χ ≤ 1 0\le\chi\le1 0 ≤ χ ≤ 1 and χ = 1 \chi=1 χ = 1 on [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] , and set
f R ( x ) : = χ ( x / R ) f ( x ) , R ≥ 1. f_R(x):=\chi(x/R)f(x),\qquad R\ge1. f R ( x ) := χ ( x / R ) f ( x ) , R ≥ 1.
Then f R ∈ L 1 ( R ) ∩ H 2 ( R ) f_R\in L^1(\mathbb{R})\cap H^2(\mathbb{R}) f R ∈ L 1 ( R ) ∩ H 2 ( R ) and f R ( 0 ) = 0 f_R(0)=0 f R ( 0 ) = 0 . Dominated convergence shows that ∥ f R − f ∥ H 2 ( R ) → 0 \|f_R-f\|_{H^2(\mathbb{R})}\to0 ∥ f R − f ∥ H 2 ( R ) → 0 . Sobolev embedding also gives ∥ f R − f ∥ ∞ → 0 \|f_R-f\|_\infty\to0 ∥ f R − f ∥ ∞ → 0 . For each fixed n n n , use ∥ A n ∥ o p ≤ N n \|A_n\|_{\mathrm{op}}\le N_n ∥ A n ∥ op ≤ N n to obtain
E [ ∣ Tr [ f R ( A n ) e c A n ] − Tr [ f ( A n ) e c A n ] ∣ 2 ] ≤ ∥ f R − f ∥ ∞ 2 E [ N n 2 e 2 ∣ c ∣ N n ] ⟶ 0. \mathbb{E}\left[\left|\operatorname{Tr}[f_R(A_n)e^{cA_n}]-\operatorname{Tr}[f(A_n)e^{cA_n}]\right|^2\right]\le\|f_R-f\|_\infty^2\mathbb{E}\left[N_n^2e^{2|c|N_n}\right]\longrightarrow0. E [ Tr [ f R ( A n ) e c A n ] − Tr [ f ( A n ) e c A n ] 2 ] ≤ ∥ f R − f ∥ ∞ 2 E [ N n 2 e 2∣ c ∣ N n ] ⟶ 0.
The expectation on the right is finite as N n N_n N n is Poisson. Thus the trace statistics converge in L 2 L^2 L 2 for each fixed n n n , and their variances converge. Passing to R → ∞ R\to\infty R → ∞ in the previous bound for f R f_R f R yields Var [ Tr [ f ( A n ) e c A n ] ] ≤ C ∣ W n ∣ ( ∥ f ∥ 2 2 + ∥ f ′ ′ ∥ 2 2 ) \operatorname{Var}\left[\operatorname{Tr}[f(A_n)e^{cA_n}]\right]\le C|W_n|(\|f\|_2^2+\|f''\|_2^2) Var [ Tr [ f ( A n ) e c A n ] ] ≤ C ∣ W n ∣ ( ∥ f ∥ 2 2 + ∥ f ′′ ∥ 2 2 ) . The constant is the same as for f R f_R f R , and thus independent of n n n . □
Statement on AI use
Statement on AI use. The original research ideas and the first drafts of all sections were our own. During subsequent revisions, we used AI, more precisely the latest versions of ChatGPT, to identify possible mathematical errors and suggest corrections. Moreover, our original draft needed an exponential moment assumption that was stronger than the currently used (EM) assumption. Here, AI suggested a preliminary version of Lemma 16 that allows us to prove the results in the stated generality. After that, we carefully went through this auxiliary result and revised its formulation so as to make it accessible. All mathematical statements, proofs, computations, references, and conclusions were independently checked and verified by the authors, who take full responsibility for the content of the manuscript.
References