betaFeedback

WorksMathematics

Central limit theorem for linear eigenvalue statistics of random spatial networks

Abstract

The random connection model provides a flexible framework for spatial networks in which connectivity is determined by both geometry and randomness. Its vertices form a Poisson point process, and pairs are connected independently, conditional on their positions, with distance-dependent probabilities. This generalizes the hard distance threshold of random geometric graphs and accommodates long-range connections. We also consider a weighted extension in which each vertex carries a positive random weight that influences its probability of connecting to other vertices. This combination of spatial structure and heterogeneous connectivity raises the question of whether spectral fluctuations exhibit Gaussian behavior on large scales. We address this question by studying linear eigenvalue statistics Tr⁡[f(An)]\operatorname{Tr}[f(A_n)] of the adjacency matrix AnA_n of the random connection model in expanding windows WnW_n. For polynomial test functions, we establish a finite limit for the normalized variance and prove a central limit theorem under suitable connection-profile integrability and vertex mark moment assumptions. When the limiting variance is positive, the normalized statistics admit an O(∣Wn∣−1/2)O(|W_n|^{-1/2}) normal-approximation bound in both Wasserstein and Kolmogorov distance. Furthermore, under an exponential moment assumption on the vertex marks, we extend the central limit theorem to a general class of test functions ff.

Introduction

Let AA be the adjacency matrix of a finite graph with eigenvalues λ1,…,λN\lambda_1,\ldots,\lambda_N. For a test function ff, the linear eigenvalue statistic

Tr⁡[f(A)]=∑i=1Nf(λi)\operatorname{Tr}[f(A)] = \sum_{i=1}^{N} f(\lambda_i)

provides a basic summary of the graph spectrum. For polynomial ff, 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) for displacement xx. 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(An)]\operatorname{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 ∣Wn∣−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≥1d \ge1. For n≥1n \ge1, let Pn\mathcal{P}_n be a homogeneous Poisson point process with intensity 11 on Wn:=[−n1/d/2,n1/d/2]d⊆RdW_n := [-n^{1/d}/2,n^{1/d}/2]^d \subseteq\mathbb{R}^d. Then, ∣Wn∣=n|W_n| = n, and the asymptotic regime is n→∞n \to\infty with the dimension dd fixed. We denote by Nn:=∣Pn∣N_n := |\mathcal{P}_n| the random number of points in Pn\mathcal{P}_n and choose an arbitrary enumeration Pn={X1,…,XNn}\mathcal{P}_n = \{X_1,\ldots,X_{N_n}\}.

Let φ:[0,∞)→[0,1]\varphi: [0,\infty) \to[0,1] be a non-increasing function, called the connection profile, satisfying

Mφ:=∫Rdφ(∥x∥d) dx<∞.(1)M_{\varphi} := \int_{\mathbb{R}^d} \varphi(\|x\|^d)\,\mathrm{d}x < \infty. \tag*{(1)}

More generally, for γ∈(0,1]\gamma\in(0,1], set

Mφ,γ:=∫Rdφ(∥x∥d)γ dx.(2)M_{\varphi,\gamma} := \int_{\mathbb{R}^d} \varphi(\|x\|^d)^\gamma\,\mathrm{d}x. \tag*{(2)}

The random connection model (RCM) is the random graph with vertex set Pn\mathcal{P}_n. Conditionally on Pn\mathcal{P}_n, each pair of distinct points Xi,Xj∈PnX_i,X_j \in\mathcal{P}_n is connected by an edge with probability φ(∥Xi−Xj∥d)\varphi(\|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\}, for fixed R>0R>0, produces the random geometric graph from [18] as a special case. The adjacency matrix An∈RNn×NnA_n \in\mathbb{R}^{N_n\times N_n} of the RCM is defined by

An(i,j):=An(Xi,Xj):=1{{Xi,Xj} is an edge},1≤i,j≤Nn,(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)}

with An(i,i)=0A_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∈Wnx\in W_n with incident edges added independently with probabilities prescribed by φ\varphi, then the expected degree of xx is

∫Wnφ(∥x−y∥d) dy≤Mφ,\int_{W_n}\varphi(\|x-y\|^d)\,\mathrm{d}y\le M_{\varphi},

uniformly in nn and xx. The expected number of edges is therefore at most Mφ∣Wn∣/2M_{\varphi}|W_n|/2, whereas the expected number of vertices is ∣Wn∣|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 X1,…,XNnX_1,\ldots,X_{N_n} independent marks U1,…,UNnU_1,\ldots,U_{N_n} with common distribution μ\mu on [1,∞)[1,\infty), independently of Pn\mathcal{P}_n. The resulting marked Poisson process is P~n={(Xi,Ui):1≤i≤Nn}\widetilde{\mathcal{P}}_n=\{(X_i,U_i):1\le i\le N_n\} on W~n:=Wn×[1,∞)\widetilde{W}_n:=W_n\times[1,\infty). We write λ(dx,du):=dx μ(du)\lambda(\mathrm{d}x,\mathrm{d}u):=\mathrm{d}x\,\mu(\mathrm{d}u) for its intensity measure. Throughout, UU denotes a random variable with distribution μ\mu.

Let φ\varphi be the connection profile introduced in Subsection 2.1, and let κ:[1,∞)2→[1,∞)\kappa:[1,\infty)^2\to[1,\infty) be a symmetric measurable function, called the interaction kernel. Given the marked configuration P~n\widetilde{\mathcal{P}}_n, the weighted RCM has spatial vertex set Pn\mathcal{P}_n. For distinct marked points (Xi,Ui),(Xj,Uj)∈P~n(X_i,U_i),(X_j,U_j)\in\widetilde{\mathcal{P}}_n, place an edge between XiX_i and XjX_j with probability

φ(∥Xi−Xj∥dκ(Ui,Uj)),\varphi\left(\frac{\|X_i-X_j\|^d}{\kappa(U_i,U_j)}\right),

independently for different unordered pairs, conditional on P~n\widetilde{\mathcal{P}}_n.

Intuitively, κ(u,v)\kappa(u,v) measures the tendency for two vertices with weights uu and vv to connect. For interaction kernels increasing in both arguments, such as κ(u,v)=uv\kappa(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≡1U\equiv1 and κ≡1\kappa\equiv1. 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 U0U_0 having distribution μ\mu and incident edges added independently. Let DD be its degree. For every uu such that E[κ(u,U)]<∞\mathbb{E}[\kappa(u,U)]<\infty, the neighboring vertices form an independent thinning of the marked Poisson process and hence

D∣{U0=u}∼Poisson⁡(Λ(u)),Λ(u):=∫Rd∫[1,∞)φ(∥x∥dκ(u,v))μ(dv) dx=MφE[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)].

Consequently, whenever Λ(U0)<∞\Lambda(U_0)<\infty 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)=uv\kappa(u,v)=uv. If 0<Mφ<∞0<M_{\varphi}<\infty and EU<∞\mathbb{E}U<\infty, then

Λ(u)=au,a:=MφEU.\Lambda(u)=au,\qquad a:=M_{\varphi}\mathbb{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} for t≥1t\ge1 with β>1\beta>1, then the degree distribution satisfies

P(D≥k)∼P(aU≥k)=aβk−β,k→∞.\mathbb{P}(D\ge k)\sim\mathbb{P}(aU\ge k)=a^\beta k^{-\beta},\qquad k\to\infty.

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, AnA_n denotes the adjacency matrix of this graph. We identify a marked vertex xˇ=(x,u)\check{x}=(x,u) with its spatial coordinate when it is used as a vertex argument of a graph quantity. In particular,

An(xˇ,yˇ):=An(x,y).A_n(\check{x},\check{y}):=A_n(x,y).

For a graph with spatial vertex set QQ and its associated vertex marks, write Qˇ\check{Q} 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 Mφ,1/4<∞M_{\varphi,1/4}<\infty with Mφ,1/4M_{\varphi,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,

polar coordinate integration yields

Mφ,γ<∞⟺αγ>1.M_{\varphi,\gamma}<\infty\quad\Longleftrightarrow\quad\alpha\gamma>1.

Thus, the condition Mφ,1/4<∞M_{\varphi,1/4}<\infty permits this profile whenever α>4\alpha>4. The absence of dd from this threshold reflects our convention of evaluating the profile at the ddth 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 such that

E[eηU]<∞.(4)\mathbb{E}[\mathrm{e}^{\eta U}]<\infty. \tag*{(4)}

The key assumption on our kernel κ\kappa will be that it is dominated by the product kernel. More precisely, we say that the interaction kernel satisfies Condition (PK) if there exists cκ≥1c_\kappa\ge1 such that

1≤κ(u,v)≤cκuvfor all u,v≥1.(5)1\le\kappa(u,v)\le c_\kappa uv\qquad\text{for all }u,v\ge1. \tag*{(5)}

References

  1. [1]K. Adhikari, R. J. Adler, O. Bobrowski, and R. Rosenthal. On the spectrum of dense random geometric graphs. Ann. Appl. Probab., 32(3):1734–1773, 2022.arxiv.org/abs/2004.04967
  2. [2]K. Avramchuk, A. Bobu, and M. Dreveton. Higher-order spectral clustering for geometric graphs. J. Fourier Anal. Appl., 27:22, 2021.
  3. [3]M. Barthélemy. Spatial networks. Phys. Rep., 499(1–3):1–101, 2011.
  4. [4]F. Benaych-Georges, A. Guionnet, and C. Male. Central limit theorems for linear statistics of heavy tailed random matrices. Comm. Math. Phys., 329(2):641–686, 2014.
  5. [5]R. Bhatia. Matrix Analysis, volume 169 of Graduate Texts in Mathematics. Springer, New York, 1997.
  6. [6]K. Bogdan, J. Rosiński, G. Serafin, and L. Wojciechowski. Lévy systems and moment formulas for mixed Poisson integrals. In Stochastic Analysis and Related Topics: A Festschrift in Honor of Rodrigo Bańuelos, pages 139–164. Springer, 2017.
  7. [7]A. D. Broido and A. Clauset. Scale-free networks are rare. Nat. Commun., 10:1017, 2019.
  8. [8]V. H. Can and K. D. Trinh. Random connection models in the thermodynamic regime: central limit theorems for add-one cost stabilizing functionals. Electron. J. Probab., 27:Paper No. 36, 40, 2022.arxiv.org/abs/2004.06313
  9. [9]P. Deprez and M. V. Wüthrich. Scale-free percolation in continuum space. Commun. Math. Stat., 7(3):269–308, 2019.arxiv.org/abs/1312.1948
  10. [10]L. Erdős, A. Knowles, H.-T. Yau, and J. Yin. Spectral statistics of Erdős–Rényi graphs I: Local semicircle law. Ann. Probab., 41(3B):2279–2375, 2013.
  11. [11]M. Franceschetti and R. Meester. Random Networks for Communication. Cambridge University Press, Cambridge, 2007.
  12. [12]Z. Füredi and J. Komlós. The eigenvalues of random symmetric matrices. Combinatorica, 1(3):233–241, 1981.DOI
  13. [13]P. Gracar, A. Grauer, L. Lüchtrath, and P. Mörters. The age-dependent random connection model. Queueing Syst., 93(3–4):309–331, 2019.DOI
  14. [14]P. Gracar, M. Heydenreich, C. Mönch, and P. Mörters. Recurrence versus transience for weight-dependent random connection models. Electron. J. Probab., 27:Paper No. 60, 31, 2022.DOI
  15. [15]P. Gracar, L. Lüchtrath, and C. Mönch. Finiteness of the percolation threshold for inhomogeneous long-range models in one dimension. Electron. J. Probab., 30:Paper No. 134, 29, 2025.DOI
  16. [16]M. Hamidouche, L. Cottatellucci, and K. Avrachenkov. Spectral analysis of the adjacency matrix of random geometric graphs. In 2019 57th Annual Allerton Conference on Communication, Control, and Computing, pages 208–214, 2019.arxiv.org/abs/1910.08871
  17. [17]N. Heerten, C. Hirsch, and M. Otto. Moderate deviations for weight-dependent random connection models. J. Appl. Probab., 62(4):1406–1428, 2025.DOI
  18. [18]C. Hirsch, K. Nam, and M. Otto. Central limit theorem for linear eigenvalue statistics of random geometric graphs. arXiv preprint arXiv:2602.17006, 2026.arxiv.org/abs/2602.17006
  19. [19]C. Hirsch, M. Otto, and A. M. Svane. Normal approximation for Gibbs processes via disagreement couplings. Electron. J. Probab., 30, 2025.arxiv.org/abs/2309.00394
  20. [20]R. Lachièze-Rey, M. Schulte, and J. E. Yukich. Normal approximation for stabilizing functionals. Ann. Appl. Probab., 29(2):931–993, 2019.
  21. [21]G. Last, F. Nestmann, and M. Schulte. The random connection model and functions of edge-marked Poisson processes: second order properties and normal approximation. Ann. Appl. Probab., 31(1):128–168, 2021.arxiv.org/abs/1808.01203
  22. [22]G. Last, G. Peccati, and M. Schulte. Normal approximation on Poisson spaces: Mehler’s formula, second order Poincaré inequalities and stabilization. Probab. Theory Related Fields, 165(3-4):667–723, 2016.
  23. [23]G. Last and M. Penrose. Lectures on the Poisson process, volume 7. Cambridge University Press, 2018.DOI
  24. [24]A. Y. Ng, M. I. Jordan, and Y. Weiss. On spectral clustering: Analysis and an algorithm. In Advances in Neural Information Processing Systems 14, pages 849–856, 2002.
  25. [25]D. Pabst. Betti numbers in the random connection model for higher-dimensional simplicial complexes and the Boolean model. Stochastic Process. Appl., 201:105071, 2026.DOI
  26. [26]M. Penrose. Random Geometric Graphs, volume 5 of Oxford Studies in Probability. Oxford University Press, 2003.DOI
  27. [27]M. D. Penrose. On a continuum percolation model. Adv. Appl. Probab., 23(3):536–556, 1991.DOI
  28. [28]M. Shcherbina and B. Tirozzi. Central limit theorem for fluctuations of linear eigenvalue statistics of large random graphs. J. Math. Phys., 51(2):023523, 20, 2010.
  29. [29]J. Shi and J. Malik. Normalized cuts and image segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 22(8):888–905, 2000.
  30. [30]U. von Luxburg. A tutorial on spectral clustering. Stat. Comput., 17(4):395–416, 2007.arxiv.org/abs/0711.0189

Paper details

vlnc:W_p46p9ug4
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents