Global Clustering Coefficient

• Global clustering coefficient is a metric that quantifies the ratio of closed triplets (triangles) to all triplets in a network, measuring its transitivity.
• It is defined using adjacency matrices to count closed and open triplets, and is applicable across models like random, geometric, and scale-free graphs.
• The coefficient underpins graph generation, statistical testing, and comparative network analysis by providing a scalar measure of structural cohesion.

The global clustering coefficient is a central scalar metric in network science, quantifying the prevalence of triangular motifs in a network and serving as a key indicator of structural transitivity. Mathematically, for a simple undirected graph $G = (V, E)$ with adjacency matrix $A$, the global clustering coefficient $C(G)$ is defined as the ratio of the number of closed triplets (triangles) to the total number of triplets (2-paths), formally: $$C(G) = \frac{\sum_{i \ne j \ne k} A_{ij} A_{jk} A_{ki}}{\sum_{i \ne j \ne k} A_{ij} A_{jk}}$$ Alternatively, utilizing the fact that each triangle contributes three closed triplets, this can be rewritten as $$C(G) = \frac{3 \times (\# \text{triangles})}{2 \times (\# \text{2-paths})}$$. The coefficient provides the probability that, in the network’s topology, “a friend of my friend is also my friend” (Yuan, 16 Oct 2025).

1. Formal Definition and Interpretations

The global clustering coefficient is a normalized count of transitive closed 2-paths:

• Closed triplet: three nodes, all mutually connected.
• Triplet (2-path): an ordered triple of nodes $(i, j, k)$ such that $(i,j), (j,k) \in E$.

When expressed via adjacency matrices, the numerator sums all closed triplets: $\sum_{i, j, k \in V} a_{ij} a_{jk} a_{ki}$ And the denominator sums all triplets: $\sum_{i \in V} d_i (d_i - 1)$ where $d_i$ is the degree of node $A$0 (Tuzhilin, 2024).

The global clustering coefficient, always $A$1, equals 0 if no triangles are present and 1 if every connected triple is closed.

2. Theoretical Properties and Methodological Variants

Distinct definitions of "global clustering coefficient" appear in the literature:

• Triplet-based coefficient ("transitivity"): computes the global fraction of closed triplets. This is the canonical definition in mathematical network theory and is widely adopted for analysis of motif closure and structural transitivity.
• Average local clustering coefficient: averages the fraction of closed triplets conditioned locally around each node, i.e., $A$2, where $A$3 is the local clustering at node $A$4 (Nesterov, 2024). These two are inequivalent, particularly in graphs with broad degree heterogeneity. $A$5 effectively weights local contributions by the number of 2-paths centered at each node, while the average local clustering gives each node equal weight.

Inequalities relating these definitions are established in (Tuzhilin, 2024): $A$6 provided the sequence of degree and local clustering is similarly ordered.

Extensions exist for weighted graphs, bipartite/two-mode networks, and graphs with structural constraints (Prokhorenkova, 2015, Opsahl, 2010, Touli et al., 2021). In weighted networks, definitions such as the Opsahl-Panzarasa generalization compute the weighted mean of triangle and triplet strengths, e.g., geometric or arithmetic means of edge weights (Prokhorenkova, 2015).

3. Global Clustering in Random Graphs and Model-Specific Results

Random Annulus Graphs (RAG)

In the RAG model, nodes are uniformly distributed on the unit circle, and edges are formed only if the wrap-around distance falls within an annulus $A$7. Under the sparsity regime $A$8, the limit of the global clustering coefficient is: $A$9 This interpolates between the classical random geometric graph ($C(G)$0, $C(G)$1) and low-clustering Erdős–Rényi graphs ($C(G)$2, $C(G)$3) (Yuan, 16 Oct 2025).

Random Geometric Graphs (RGG)

For uniform and non-uniform RGGs, the limiting global clustering coefficient is: $C(G)$4 with fluctuations characterized by a regime-dependent central limit theorem—Lyapunov CLT in dense, U-statistics theory in intermediate, and method of moments in sparse settings (Yuan et al., 19 Feb 2026).

Scale-Free Graphs

For graphs with infinite-variance power-law degree distributions ($C(G)$5), the global clustering coefficient vanishes asymptotically ($C(G)$6), even if high local clustering persists (Prokhorenkova et al., 2014, Oliveira et al., 2016, Prokhorenkova, 2015). Explicit upper bounds in terms of $C(G)$7 are established—e.g., $C(G)$8 for any $C(G)$9 (Prokhorenkova et al., 2014). In weighted scale-free graphs (with parallel edges/multigraphs), non-vanishing clustering can be achieved (Prokhorenkova, 2015).

Intersection Graphs

In random intersection graphs, global clustering can be tuned even in the infinite-degree-variance regime, depending on the parameters of the bipartite generation model (active, passive, inhomogeneous). In particular, for active models with $$C(G) = \frac{\sum_{i \ne j \ne k} A_{ij} A_{jk} A_{ki}}{\sum_{i \ne j \ne k} A_{ij} A_{jk}}$$0, $$C(G) = \frac{\sum_{i \ne j \ne k} A_{ij} A_{jk} A_{ki}}{\sum_{i \ne j \ne k} A_{ij} A_{jk}}$$1, but in inhomogeneous models, the limiting value can be positive and tunable (Bloznelis et al., 2016).

4. Asymptotics, Statistical Fluctuations, and Central Limit Theorems

For spatial and geometric random graphs, central limit theorems (CLTs) characterize the fluctuations of the global clustering coefficient about its deterministic limit:

• In the RAG, with proper centering and scaling, the standardized statistic $$C(G) = \frac{\sum_{i \ne j \ne k} A_{ij} A_{jk} A_{ki}}{\sum_{i \ne j \ne k} A_{ij} A_{jk}}$$2 converges in distribution to standard normal, where $$C(G) = \frac{\sum_{i \ne j \ne k} A_{ij} A_{jk} A_{ki}}{\sum_{i \ne j \ne k} A_{ij} A_{jk}}$$3 is the limiting mean and $$C(G) = \frac{\sum_{i \ne j \ne k} A_{ij} A_{jk} A_{ki}}{\sum_{i \ne j \ne k} A_{ij} A_{jk}}$$4 is the variance obtained via degenerate U-statistics with sample-size-dependent kernels (Yuan, 16 Oct 2025).
• Analogous CLTs hold in random geometric graphs, with explicit expressions for centering, scaling, and details on variance asymptotics (Yuan et al., 19 Feb 2026).

These results enable the construction of confidence intervals and hypothesis tests for clustering levels, and provide analytic understanding of motif-based statistics under random graph models.

5. Extensions and Variants: Two-Mode, Relative, and Model-Aware Coefficients

Two-Mode and Bipartite Networks

Standard global clustering coefficients suffer from substantial bias when applied naïvely to projected one-mode representations of bipartite/two-mode networks (Opsahl, 2010). To avoid inflation by “automatic” cliques, a redefinition based on 4-paths is introduced: $$C(G) = \frac{\sum_{i \ne j \ne k} A_{ij} A_{jk} A_{ki}}{\sum_{i \ne j \ne k} A_{ij} A_{jk}}$$5 where a 4-path is closed if its endpoints share an additional common neighbor other than those already traversed. Weighted extensions preserve interpretability and robustness.

Relative Clustering Coefficient

The relative clustering coefficient (RCC) is designed for situations where certain edges are structurally forbidden (zero probability of existing). RCC restricts the denominator in the clustering calculation to only those triplets where all three edges are feasible, yielding: $$C(G) = \frac{\sum_{i \ne j \ne k} A_{ij} A_{jk} A_{ki}}{\sum_{i \ne j \ne k} A_{ij} A_{jk}}$$6 where $$C(G) = \frac{\sum_{i \ne j \ne k} A_{ij} A_{jk} A_{ki}}{\sum_{i \ne j \ne k} A_{ij} A_{jk}}$$7 and $$C(G) = \frac{\sum_{i \ne j \ne k} A_{ij} A_{jk} A_{ki}}{\sum_{i \ne j \ne k} A_{ij} A_{jk}}$$8 count closed and open triplets with all edges feasible (Touli et al., 2021).

6. Applications and Enforcement in Graph Generation

In constrained random graph generation, the global clustering coefficient serves as a strict structural constraint. Hybrid frameworks such as Ant Colony Optimization (ACO) combined with MCMC samplers can efficiently find graphs that meet prescribed global clustering and other structural targets. Local and incremental update rules allow for efficient enforcement of the constraint in graph rewriting procedures. Experimental results show that by strictly controlling C(G), structurally diverse ensembles of graphs can be generated under fixed global transitivity (Ferenczi et al., 23 Feb 2026).

7. Network Structure, Centralities, and Comparative Analysis

The global clustering coefficient structurally relates to other network properties such as local efficiency, various centralities (betweenness, stress), closeness, and radiality. Inequalities bound C(G) via these centralities under monotonicity or ordering conditions on the degree and clustering sequences (Tuzhilin, 2024). For example, Theorem 6 in (Tuzhilin, 2024) establishes that in degree—clustering-comonotonic graphs, the average local clustering is bounded above by the global clustering coefficient.

A plausible implication is that in networks where local path-based centralities are high, a substantial global clustering coefficient must exist, enforcing a minimum level of transitive closure at the network scale.

---

References:

• (Yuan, 16 Oct 2025): Asymptotic distribution of the global clustering coefficient in a random annulus graph
• (Yuan et al., 19 Feb 2026): Central limit theorem for the global clustering coefficient of random geometric graphs
• (Prokhorenkova et al., 2014): Global clustering coefficient in scale-free networks
• (Prokhorenkova, 2015): Global clustering coefficient in scale-free weighted and unweighted networks
• (Bloznelis et al., 2016): Clustering coefficient of random intersection graphs with infinite degree variance
• (Opsahl, 2010): Triadic closure in two-mode networks: Redefining the global and local clustering coefficients
• (Touli et al., 2021): Relative Clustering Coefficient
• (Ferenczi et al., 23 Feb 2026): Constrained graph generation: Preserving diameter and clustering coefficient simultaneously
• (Nesterov, 2024, Tuzhilin, 2024, Oliveira et al., 2016)