2025-11-25T22:01:17.838996
Functional central limit theorem for subgraph counts in a dynamic random connection model
Hazra, Kriukov, Mandjes et al.
We prove a functional central limit theorem for subgraph counts in a dynamic version of the random connection model. To establish tightness, we develop a dynamic extension of the cumulant method.
academicFunctional Central Limit Theorem for Subgraph Counts in a Dynamic Random Connection Model
Basic Information
- Paper ID: 2511.18003
- Title: Functional central limit theorem for subgraph counts in a dynamic random connection model
- Authors: Rajat Subhra Hazra (Leiden University), Nikolai Kriukov (University of Amsterdam), Michel Mandjes (Leiden University & University of Amsterdam), Moritz Otto (Leiden University)
- Classification: math.PR (Probability Theory)
- Publication Date: November 22, 2025
- Paper Link: https://arxiv.org/abs/2511.18003
Abstract
This paper establishes a functional central limit theorem for subgraph counts in a dynamic random connection model. To establish tightness, the authors develop a dynamic extension of the cumulant method. This represents the first successful application of the cumulant method to prove functional limit theorems in dynamic random geometric graphs.
Research Background and Motivation
Core Problem
The Random Connection Model (RCM) is a fundamental random geometric model for describing spatial networks, where nodes connect with certain probabilities based on their mutual distances. The core question addressed in this paper is: What is the limiting behavior of subgraph count processes in a dynamic RCM where nodes are dynamically activated/deactivated?
Problem Significance
- Theoretical Importance: Subgraph counts (such as triangles, stars, etc.) not only capture local connectivity patterns but also play a key role in understanding higher-order structures and limiting behaviors of the model
- Practical Applications: Dynamic networks better reflect the behavior of real-world systems (communication networks, social networks, biological networks) where edges and/or vertices change randomly over time
- Methodological Contribution: Existing research primarily focuses on static networks, while the dynamic case presents greater mathematical challenges
Limitations of Existing Methods
- Static Restriction: Classical RCM studies (e.g., Penrose, Schulte & Thäle) primarily focus on asymptotic normality in static graphs
- Finite-Dimensional Convergence: The cumulant method has previously been used mainly for establishing finite-dimensional convergence, not systematically for proving tightness in functional limit theorems
- Difficulty in Dynamic Extension: Generalizing static results to dynamic settings faces technical challenges, particularly when handling temporal dependencies
Research Motivation
This paper is driven by trends in random graph literature focusing on dynamic evolving networks, aiming to:
- Generalize subgraph counting results from static RCM to dynamic settings
- Develop a dynamic extension of the cumulant method to prove functional CLT
- Provide theoretical foundations for practical applications such as clustering coefficient processes
Core Contributions
The main contributions of this paper include:
- Functional Central Limit Theorem: Establishes a functional CLT for multivariate subgraph count processes in dynamic RCM (Theorem 1), valid in both dense and sparse parameter regimes
- Dynamic Extension of Cumulant Method: For the first time, systematically applies the cumulant method to prove tightness conditions for dynamic random geometric graphs, demonstrating the broad applicability of this method
- Precise Covariance Structure: Explicitly characterizes the covariance structure of the limiting Gaussian process, distinguishing different behaviors in dense and sparse regions:
- Dense region: covariance is
- Sparse region: covariance is
- Application Examples: Applies the main results to clustering coefficient processes, proving functional CLT for subgraph ratio processes (Proposition 3)
Methodology Details
Task Definition
Input:
- Spatial domain: (with toroidal metric)
- Poisson point process: on with intensity measure
- Connection probability function: , where
- State transition rates: activation rate , deactivation rate
Output:
- Multivariate subgraph count process , where
Objective: Prove that the centered and normalized process converges to a Gaussian process
Model Architecture
Dynamic Random Connection Model
- Spatial Structure: Node locations are sampled from a homogeneous Poisson point process with intensity
- Dynamic Mechanism: Each node independently switches between active/inactive states
- Initial state: activated with probability
- Transition dynamics: inactive→active (rate ), active→inactive (rate )
- Edge Generation: Potential edges are sampled once and fixed, but appear in the graph only when both endpoints are simultaneously active
Normalization Scheme
Define normalization factors:
\varrho^{q_i}n^{(q_i-1)/2}\nu_n^{q_i-1}, & \nu_n \in \mathcal{D} \text{ (dense)} \\ \varrho^{q_i}\sqrt{nq_i\nu_n^{q_i-1}}, & \nu_n \in \mathcal{S} \text{ (sparse)} \end{cases}$$ Centered normalized process: $$\Gamma^*_{n,i}(t) = \frac{\Gamma_{n,i}(t) - \mathbb{E}[\Gamma_{n,i}(t)]}{\psi_{n,i}}$$ ### Technical Innovations #### 1. Partitions and Graph-Theoretic Structures Introduce partition sets $\Pi(q_1,\ldots,q_m)$ and their subsets: - $\tilde{\Pi}(q_1,\ldots,q_m)$: induced partitions $\sigma^*$ with only one block - $\bar{\Pi}(q_1,\ldots,q_m)$: partitions where each row has at least one element in a block of size $\geq 2$ For each partition $\sigma$, construct auxiliary graphs whose edge set $E_\sigma$ reflects the overlap structure between subgraphs. #### 2. Cumulant Graph Formula Utilize the cumulant representation for Poisson U-statistics (from Schulte & Thäle): $$\text{cum}(S_1,\ldots,S_m) = \sum_{\sigma\in\tilde{\Pi}(q_1,\ldots,q_m)} \int_{X^{|\sigma|}} (\otimes_{l=1}^m f^{(l)})_\sigma d\mu^{|\sigma|}$$ This connects cumulants to integrals of specific tensor products. #### 3. Key Estimates for Tightness Proof To prove tightness conditions (Billingsley condition): $$\mathbb{E}[\|\Gamma^*_n(r)-\Gamma^*_n(s)\|^2 \|\Gamma^*_n(s)-\Gamma^*_n(t)\|^2] \leq C(t-r)^2$$ Key steps: 1. Express fourth moments as partition sums: $$\Delta_{n,i,j}(r,s,t) = \sum_{\sigma\in\bar{\Pi}(q_i,q_i,q_j,q_j)} \int_{X^{|\sigma|}} (f^{(i)}\otimes f^{(i)}\otimes f^{(j)}\otimes f^{(j)})_\sigma d\mu_n^{|\sigma|}$$ 2. Separate spatial and temporal dependencies: - **Temporal part**: Utilize Markov jump process properties (Lemma 6) to obtain $|r-t|^2$ bounds - **Spatial part**: Apply Lemma 5 based on connectivity of auxiliary graphs 3. Connectivity analysis: - Connected graphs: $I_n(\sigma) \sim \beta_1\nu_n^{|\sigma|-1}$ - Two connected components: $I_n(\sigma) \sim \beta_2\nu_n^{|\sigma|-2}$ #### 4. Unified Treatment of Dense and Sparse Regions Through carefully designed normalization factors $\psi_{n,i}$, the proof framework is unified across both parameter regions, though the limiting covariance structures differ: - **Dense region** ($n\nu_n \to \infty$): counts of different subgraphs are completely correlated - **Sparse region** ($n\nu_n \to 0$): only isomorphic subgraphs are correlated ## Experimental Setup ### Theoretical Verification Rather Than Numerical Experiments This is a purely theoretical work without numerical experiments or simulations. Verification is accomplished through rigorous mathematical proofs. ### Parameter Configuration Theoretical results require: 1. **Basic Conditions**: $\lim_{n\to\infty}\nu_n = 0$, $\lim_{n\to\infty}n^{q_i}\nu_n^{q_i-1} = \infty$ (for all $i\in[m]$) 2. **Region Partition**: - Dense region: $n\nu_n \to \infty$ (e.g., $\nu_n = n^\gamma$, $-1<\gamma<0$) - Sparse region: $n\nu_n \to 0$ (e.g., $\nu_n = n^\gamma$, $-q_i/(q_i-1)<\gamma<-1$) ### Application Case: Clustering Coefficient Consider $G_1$ as a triangle and $G_2$ as a wedge: - $q=3$, $a_1=6$, $a_2=2$ - Define integral constants: $$\kappa_d = \int_{\mathbb{R}^d}\phi(\|y\|_d)dy, \quad \tau_d = \int_{(\mathbb{R}^d)^2}\phi(\|y_1\|_d)\phi(\|y_2\|_d)\phi(\|y_1-y_2\|_d)d(y_1,y_2)$$ The covariance of the clustering coefficient process is: - Dense region: $\Sigma_C(s,t) = 0$ (degenerate case) - Sparse region: $\Sigma_C(s,t) = 9(Z(|t-s|))^3\left(\frac{36}{\tau_d} - \frac{90\kappa_d^2}{\tau_d^2} + \frac{54\kappa_d^4}{\tau_d^3}\right)$ ## Experimental Results ### Main Theoretical Results **Theorem 1 (Main Result)**: If $\nu_n$ satisfies condition (3), then $\Gamma^*_n(\cdot) \to \Gamma(\cdot)$ (convergence in distribution in $D([0,T],\mathbb{R}^m)$), where $\Gamma(\cdot)$ is a centered Gaussian process with covariance matrix: $$\Sigma_{i,j}(s,t) = \begin{cases} Z(|t-s|)F^+_{ij}, & \nu_n \in \mathcal{D} \\ (Z(|t-s|))^{q_i}1\{q_i=q_j\}F^-_{ij}, & \nu_n \in \mathcal{S} \end{cases}$$ where $Z(t) = 1 + (\lambda/\mu)e^{-(\lambda+\mu)t}$ characterizes the temporal correlation of the Markov jump process. **Remark 2 (Particularity of Dense Region)**: In the dense region, we can write $F^+_{i,j} = (q_iF(G_i)/a_i)(q_jF(G_j)/a_j)$, which means the components of $\Gamma(\cdot)$ are completely correlated, expressible as: $$\Gamma'_i(\cdot) = \frac{q_iF(G_i)}{a_i}\xi(\cdot)$$ where $\xi(\cdot)$ is a scalar Gaussian process. ### Application Results **Proposition 3 (Subgraph Ratio Process)**: Let $G_1, G_2$ be connected graphs satisfying $V(G_1)=V(G_2)=q$ and $G_1\subset G_2$. Define the subgraph ratio process: $$C_{n,G_1,G_2}(t) = \frac{a_1\Gamma_{n,1}(t)}{a_2\Gamma_{n,2}(t)}$$ Then the centered normalized process $C^*_{n,G_1,G_2}(\cdot)$ converges to a centered Gaussian process $C_{G_1,G_2}(\cdot)$. **In particular**, in the dense region $\Sigma_C(s,t)=0$, a degenerate phenomenon arising from the complete correlation of numerator and denominator. ### Key Technical Results 1. **Expectation Asymptotics** (Lemma 7): $$\mathbb{E}[\Gamma_{n,i}(t)] = \frac{F_n(G_i)(\varrho n)^{q_i}}{a_i}$$ 2. **Covariance Asymptotics** (Lemma 8): $$\text{Cov}(\Gamma_{n,i}(t),\Gamma_{n,j}(s)) \sim \sum_{m=1}^{q_i\wedge q_j}\sum_{H_1,H_2} \frac{n^{q_i+q_j-m}\varrho^{q_i+q_j}Z(|t-s|)^m}{m!(q_i-m)!(q_j-m)!}\nu_n^{q_i+q_j-m-1}F(H)$$ 3. **Integral Asymptotics** (Lemma 5): $$F_n(H) \sim \nu_n^{q-1}F(H)$$ where $F(H)$ is the normalized spatial integral. ### Effectiveness of Proof Strategy The proof proceeds in two steps: 1. **Finite-Dimensional Convergence** (Proposition 9): Via Cramér-Wold device and cumulant method, prove that higher-order cumulants $\text{cum}_M(S_n) \to 0$ ($M\geq 3$) 2. **Tightness** (Section 6): Verify Billingsley conditions using connectivity analysis of partitions and temporal estimates for Markov processes ## Related Work ### Static Random Connection Models 1. **Classical Literature**: - Meester & Roy (1996): Continuum percolation theory - Penrose (1991, 2003): Foundational work on random geometric graphs - Roy (2011): Percolation on random geometric graphs 2. **Subgraph Counting**: - Penrose (2003): Asymptotic normality of subgraph counts in static RCM - Schulte & Thäle (2017, 2024): Applications of cumulant method in Poisson functionals - Liu & Privault (2024): Normal approximation for subgraph counts in random connection models - Heerten et al. (2025): Moderate deviations for weighted random connection models ### Dynamic Random Graphs 1. **Temporal Random Graph Surveys**: - Holme & Saramäki (2012): Physics perspective on temporal networks 2. **Dynamic Erdős-Rényi Graphs**: - Chatterjee & Varadhan (2011): Large deviations for static ER graphs - Braunsteins et al. (2023): Sample path large deviations for dynamic ER graphs - Erdős et al. (2013): Spectral statistics of ER graphs (static) - Hazra et al. (2025a): Functional CLT for principal eigenvalues of dynamic ER graphs - Hazra et al. (2025b): Functional CLT for simultaneous subgraph counts of dynamic ER graphs ### Unique Contributions of This Paper - **First** to extend functional CLT to dynamic random connection models (more general spatial models than ER graphs) - **First** to systematically use cumulant method for proving tightness in dynamic random geometric graphs - Provides unified theoretical framework for both dense and sparse parameter regions ## Conclusions and Discussion ### Main Conclusions 1. **Establishment of Functional CLT**: Successfully establishes a functional central limit theorem for multivariate subgraph count processes in dynamic RCM, with limiting Gaussian process whose covariance structure explicitly depends on: - Spatial structure (via $F^+_{ij}$ or $F^-_{ij}$) - Temporal correlation (via $Z(|t-s|)$) - Parameter region (dense vs. sparse) 2. **Methodological Breakthrough**: The cumulant method can be used not only for finite-dimensional convergence but also effectively handles tightness conditions in functional limit theorems, demonstrating the broad applicability of this method 3. **Practical Applications**: Functional CLT for subgraph ratio processes (such as clustering coefficients) provides theoretical foundations for analyzing statistical properties of dynamic networks ### Limitations 1. **Model Assumptions**: - Node positions are fixed; only state dynamics change (does not consider node mobility) - Independent activation/deactivation (real networks may have spatial or temporal correlations) - Toroidal metric avoids boundary effects (boundaries may be important in practical applications) 2. **Parameter Restrictions**: - Requires $\nu_n \to 0$ and $n^{q_i}\nu_n^{q_i-1} \to \infty$, excluding certain parameter ranges - Degenerate phenomena in dense region (complete correlation) limit applications 3. **Technical Limitations**: - Proof relies on connected graph assumptions - Does not address more complex graph structures (directed graphs, multigraphs) ### Future Directions While not explicitly listed in the paper, inferrable research directions include: 1. **Model Extensions**: - Consider models where node positions also change dynamically - Introduce spatially or temporally correlated activation mechanisms - Study non-Markovian state transition processes 2. **Theoretical Deepening**: - Large deviations principles (similar to Braunsteins et al.'s work) - Moderate deviations principles (similar to Heerten et al.'s work) - Refined convergence rate estimates 3. **Application Extensions**: - Other network statistics (diameter, connected component sizes) - Multilayer dynamic networks - Statistical inference on real data 4. **Computational Methods**: - Develop efficient simulation algorithms - Implementation of statistical testing methods ## In-Depth Evaluation ### Strengths 1. **Theoretical Rigor**: - Complete proofs with sufficient technical details, from cumulant graph formulas to tightness verification - Distinguishes dense and sparse regions, providing unified framework while respecting different behaviors - Lemma 5's integral asymptotics and Lemma 6's temporal estimates provide solid foundations for main results 2. **Methodological Innovation**: - **Key Innovation**: Extends cumulant method from finite-dimensional convergence to tightness proof in functional limit theorems - Connectivity analysis of partitions cleverly handles spatial dependency structures - Separation of temporal and spatial dependencies demonstrates deep technical insight 3. **Result Completeness**: - Not only proves main theorem but provides practical applications (clustering coefficient) - Explicitly gives covariance structure of limiting Gaussian process - Remark 2's observation about complete correlation in dense region is valuable 4. **Writing Clarity**: - Clear structure: motivation → model → preliminaries → expectations/covariance → finite-dimensional convergence → tightness → applications - Sufficient technical preparation (Section 3 preliminaries) - Figure 1 intuitively displays the dynamic RCM mechanism ### Weaknesses 1. **Lack of Numerical Verification**: - As a purely theoretical work, lacks numerical simulations to verify theoretical predictions - No empirical evidence of convergence rates in finite samples - Practical application cases remain at theoretical level 2. **Degeneracy in Dense Region**: - Complete correlation of different subgraph counts in dense region (Remark 2) limits result richness - Subgraph ratio process has zero covariance in dense region (Proposition 3), limiting practical significance 3. **Technical Complexity**: - Partition notation system ($\Pi, \tilde{\Pi}, \bar{\Pi}$, etc.) is quite abstract, difficult for newcomers - Dense technical details in Section 6 tightness proof reduce readability 4. **Model Realism**: - Independent activation/deactivation assumption does not hold in many real networks - Toroidal metric, while technically convenient, distances from practical applications ### Impact 1. **Contribution to Field**: - **Important Theoretical Progress**: First to extend functional CLT to dynamic random connection models - **Methodological Contribution**: Demonstrates powerful functionality of cumulant method in dynamic settings - Establishes foundations for dynamic spatial random graph theory 2. **Practical Value**: - Provides theoretical basis for statistical inference in dynamic networks - Asymptotic theory for network metrics like clustering coefficients applicable to hypothesis testing - Potential applications to wireless networks, social network analysis 3. **Reproducibility**: - Detailed theoretical proofs verifiable by specialists - Lacks code or numerical experiments; practical applications require further work - Main result conditions are explicit, facilitating citation in subsequent research ### Applicable Scenarios 1. **Theoretical Research**: - Further development of random geometric graph theory - Limit theorems for other dynamic spatial random models - Application studies of cumulant method 2. **Practical Applications**: - **Wireless Communication Networks**: Node connections depend on distance; nodes may periodically sleep - **Social Networks**: User activity changes dynamically; connection probability depends on "social distance" - **Biological Networks**: Activation/deactivation dynamics of cells or proteins 3. **Statistical Inference**: - Hypothesis testing for dynamic network data - Estimation of network parameters (activation rates, connection functions) - Change point detection in network evolution ## Key References 1. **Schulte & Thäle (2024)**: "Moderate deviations on Poisson chaos" - Core reference for cumulant method 2. **Last et al. (2014)**: "Moments and central limit theorems for some multivariate Poisson functionals" - Foundation for Poisson functional theory 3. **Penrose (2003)**: "Random Geometric Graphs" - Classical textbook on random geometric graphs 4. **Hazra et al. (2025b)**: "Functional CLT for simultaneous subgraph count of dynamic ER graphs" - Most relevant prior work 5. **Billingsley (2013)**: "Convergence of Probability Measures" - Standard reference for functional limit theorems --- ## Overall Assessment This is a **high-quality theoretical probability paper** making important contributions to dynamic random geometric graphs. Main strengths include: - First establishment of functional CLT for dynamic RCM - Innovative application of cumulant method for tightness proof - Technically rigorous and complete results Main weaknesses are lack of numerical verification and degeneracy phenomena in dense region. This work establishes solid foundations for statistical theory of dynamic spatial random networks, expected to have sustained impact on random geometric graph theory and network science. Recommended for researchers interested in limit theorems for random graphs, Poisson process theory, or dynamic network analysis.