Network Scientists Show the Standard Measure of Social Connectivity Fails in Heavy-Tailed Networks

Networks are everywhere in modern science, from the wiring of the human brain to the architecture of the internet, from food webs to friendship graphs on social media. One of the most studied properties of any network is assortativity, the tendency of similar nodes to connect to one another. In social networks, assortativity usually means that well-connected people tend to befriend other well-connected people, while peripheral individuals cluster together at the margins. In technological and biological networks, the opposite pattern, called disassortativity, is often observed, with hubs linking preferentially to low-degree nodes. For decades, researchers have summarized this behavior with a single number, the Pearson assortativity coefficient, and used it to compare networks, to validate models, and to reason about robustness and epidemic spreading. A new study published in PLOS Complex Systems by Marc Kaufmann, Ulysse Schaller, Thomas Bläsius and Johannes Lengler now delivers a rigorous mathematical warning: that single number, in networks with heavy-tailed degree distributions, does not measure assortativity at all.

The degree of a node is simply the number of connections it has. Real-world networks almost universally exhibit heavy-tailed degree distributions, meaning that most nodes have only a handful of links while a small minority of hubs accumulate enormous numbers of them. This skewness is not a curiosity; it is the defining structural signature of scale-free networks, and it underlies phenomena as diverse as the resilience of the internet to random failures and the explosive speed at which ideas or pathogens can propagate through populations. Assortativity interacts directly with these phenomena. A network in which hubs connect to hubs can fragment in characteristic ways under targeted attack, while a network in which hubs shield the periphery can either suppress or accelerate spreading processes depending on the details of the wiring. Getting assortativity right, therefore, is not an academic exercise but a prerequisite for credible modeling.

The standard tool for quantifying assortativity is the Pearson correlation coefficient of the degrees at the two ends of each edge. It is easy to compute, bounded between minus one and one, and has accumulated an intuitive interpretation over decades of use. Yet practitioners have long harbored reservations. The coefficient is known to be sensitive to the extreme values that heavy tails produce, and its value depends on aspects of the degree distribution that have nothing to do with wiring preferences. What distinguishes the new work is that these reservations are no longer informal caveats. Kaufmann and colleagues prove mathematically that in any network whose degree distribution is sufficiently heavy-tailed, a condition typical of real-world systems, the Pearson assortativity coefficient fails to measure assortativity in a meaningful way. The number that has been reported in thousands of papers is, in a precise sense, not measuring what researchers thought it was measuring.

The problem is not confined to Pearson’s coefficient. The authors examine other single-valued assortativity coefficients and show that they, too, are inadequate for heavy-tailed networks. The fundamental difficulty is that wiring preferences in real networks are not captured by one scalar. Nodes of different degrees can behave in radically different ways: low-degree nodes may preferentially attach to hubs while mid-degree nodes prefer their own kind, or vice versa. A single average over all edges collapses this rich, degree-dependent structure into one number, and in the presence of heavy tails the averaging process is dominated by precisely the nodes whose behavior is least representative. The consequence is that two networks with identical coefficients can have profoundly different local organization, and two networks with similar organization can report wildly different coefficients.

To escape this trap, the team adopted a more fine-grained approach. Rather than compressing assortativity into one value, they analyzed conditional and joint distributions of the degrees and weights of connected node pairs. In practical terms, they asked questions such as: given a node of degree k, what is the distribution of degrees among its neighbors? How does the probability of an edge between two nodes depend on both of their degrees simultaneously? How do edge weights, when available, correlate with the degrees of the endpoints? They pursued this program both numerically, by measuring these distributions in real-world networks, and mathematically, by deriving them in generative graph models. They also developed visualization methods that make the resulting high-dimensional structure legible, allowing researchers to see at a glance how connection preferences vary across the degree spectrum.

On the modeling side, the study concentrated on two prominent families of random graphs built on latent spaces: Chung-Lu graphs and Geometric Inhomogeneous Random Graphs, abbreviated GIRGs. Both models generate networks with heavy-tailed degree distributions by assigning each node a weight that determines its expected degree, and both embed nodes in an underlying space, a latent similarity space in the case of GIRGs, where nearby nodes are more likely to connect. GIRGs in particular have attracted attention because they reproduce a long list of desirable properties of real networks simultaneously: heavy tails, clustering, small distances, and a geometric structure that supports realistic navigation and spreading processes. They are, in many respects, the current gold standard among parsimonious network models.

The analysis of these models yielded a striking result: Chung-Lu graphs and GIRGs are assortativity-neutral. Their wiring preferences, examined through the fine-grained conditional distributions, show no systematic tendency of similar nodes to connect beyond what the degree distribution alone dictates. This neutrality is a theorem, not an empirical observation, and it has immediate consequences. Many real-world networks, by contrast, are demonstrably not assortativity-neutral when measured with the fine-grained tools. This mismatch means that when researchers have used Chung-Lu graphs or GIRGs as null models and compared their assortativity coefficients against real data, the comparison has been distorted by the failure of the coefficient itself, and the genuine structural differences between model and reality have been obscured rather than illuminated.

Recognizing this gap, the authors proposed an extension of the GIRG model that retains everything that makes the original model attractive, the heavy-tailed degree distribution, the latent geometric space, and the cascade of desirable properties that follow from them, while adding a new capability: tunable assortativity. The extended model introduces a control that shifts the wiring preferences of nodes across the degree spectrum, allowing the generation of networks that range from assortative to disassortative in a controlled and mathematically transparent way. The team analyzed the resulting model rigorously and provided a fine-grained quantification of its assortativity, characterizing exactly how the conditional degree distributions of connected pairs respond as the new parameter is varied. This gives experimentalists and theorists alike a principled instrument for generating networks with prescribed connection preferences, something single-number coefficients could never reliably specify.

The broader implications reach into every field that leans on network models. Epidemiologists who simulate disease spread on contact networks, physicists who study percolation and robustness, and social scientists who model information cascades all routinely calibrate their models against assortativity measurements. If those measurements are unreliable in heavy-tailed networks, and the new proof shows they are, then a substantial body of comparative results needs reexamination. The fine-grained distributions proposed in this study offer a replacement: instead of asking whether a network is assortative on average, researchers can now ask precisely which classes of nodes prefer which classes of partners, and how those preferences shape dynamics. The visualization methods accompanying the analysis make this practical, turning what could be an abstract statistical exercise into a diagnostic that can be applied to empirical data.

The study also carries a methodological lesson about the relationship between models and reality. Generative models are valuable not because they replicate every feature of real networks but because they provide controlled settings in which individual mechanisms can be isolated. The demonstration that GIRGs are assortativity-neutral sharpens their role: any assortative or disassortative structure observed in real data is now known to be an addition that the base model does not supply, and the extended GIRG provides a minimal way to supply it. In this sense the work does not discard the existing modeling toolkit but recalibrates it, replacing a single, provably misleading number with a distributional picture that is mathematically sound, empirically measurable, and tunable at will. For a field that has summarized half a century of network comparison in one coefficient, that is a consequential shift, and one that is likely to reshape how assortativity is reported, modeled, and interpreted across the network sciences.

Subject of Research: Degree assortativity and its measurement in scale-free and geometric random networks

Article Title: Assortativity in geometric and scale-free networks

Article References: Kaufmann, M., Schaller, U., Bläsius, T., & Lengler, J. (2026). Assortativity in geometric and scale-free networks. PLOS Complex Systems, 3(4), e0000097. https://doi.org/10.1371/journal.pcsy.0000097

Image Credits: AI Generated

DOI: 10.1371/journal.pcsy.0000097

Keywords: assortativity, networks, scale-free networks, heavy-tailed degree distribution, Pearson coefficient, geometric inhomogeneous random graphs, Chung-Lu graphs, latent space models, network robustness, degree distribution, random graph models, complex systems