Small-world topology of functional connectivity in randomly connected dynamical systems

Characterization of real-world complex systems increasingly involves the study of their topological structure using graph theory. Among global network properties, small-world property, consisting in existence of relatively short paths together with high clustering of the network, is one of the most discussed and studied. When dealing with coupled dynamical systems, links among units of the system are commonly quantified by a measure of pairwise statistical dependence of observed time series (functional connectivity). We argue that the functional connectivity approach leads to upwardly biased estimates of small-world characteristics (with respect to commonly used random graph models) due to partial transitivity of the accepted functional connectivity measures such as the correlation coefficient. In particular, this may lead to observation of small-world characteristics in connectivity graphs estimated from generic randomly connected dynamical systems. The ubiquity and robustness of the phenomenon is documented by an extensive parameter study of its manifestation in a multivariate linear autoregressive process, with discussion of the potential relevance for nonlinear processes and measures.

💡 Research Summary

The paper investigates a subtle but pervasive source of bias in network analyses that rely on functional connectivity (FC) – the statistical dependence between observed time‑series – to infer the topology of complex dynamical systems. Small‑worldness, defined by a combination of short average path length and high clustering coefficient, is one of the most celebrated global network properties. The authors argue that FC‑based graphs systematically overestimate small‑world characteristics when compared with standard random‑graph null models, and that this overestimation can arise even in systems whose underlying structural connections are purely random.

The core of the argument rests on the partial transitivity of common FC measures such as the Pearson correlation coefficient. If node A is strongly correlated with node B, and node B with node C, then A and C will tend to exhibit a non‑trivial correlation as well, even in the absence of a direct link. This statistical “triadic closure” inflates the number of triangles in the inferred graph, boosting the clustering coefficient without substantially affecting the average shortest‑path length. Consequently, the small‑world index σ = (C/Crandom)/(L/Lrandom) becomes artificially large.

To demonstrate the phenomenon, the authors employ a multivariate linear autoregressive (AR) process as a testbed. The AR model generates N‑dimensional time series where each component evolves as a linear combination of past values of all components plus Gaussian noise. By sweeping a broad parameter space – varying the autoregressive coupling strength, noise variance, and connection density – they produce thousands of synthetic datasets. For each dataset they compute pairwise Pearson correlations, threshold them to obtain binary adjacency matrices, and then calculate the clustering coefficient C and average path length L. As a null model they use Erdős‑Rényi graphs with the same number of nodes and expected degree.

The results are strikingly consistent across the parameter space. In the majority of cases the inferred FC graphs display clustering coefficients far above the random expectation, especially when the underlying network density lies in the intermediate range (≈0.1–0.3) and the noise level is modest. The average path length remains comparable to, or slightly shorter than, that of the random graphs. Consequently, the small‑world metric σ routinely exceeds 1 by a large margin, indicating a spurious small‑world signature even though the ground‑truth structural network is completely random.

The authors extend the discussion to nonlinear dependence measures (mutual information, phase‑locking value, etc.). Although these metrics capture more complex relationships, they still inherit a form of transitivity because any two variables that share a common driver tend to exhibit mutual information or phase synchrony. Hence, the same bias is expected to persist, albeit with potentially different magnitude.

The paper concludes with methodological recommendations. Researchers should not rely solely on comparisons with simple Erdős‑Rényi null models when assessing small‑worldness of FC networks. More appropriate controls include degree‑preserving randomizations, surrogate data that preserve the temporal autocorrelation structure, or direct validation against known structural connectivity (e.g., diffusion MRI tractography in neuroscience). Moreover, the findings caution against interpreting small‑world signatures as evidence of efficient information processing without first ruling out statistical artifacts inherent to the FC estimation procedure.

Overall, the study provides a rigorous, quantitative demonstration that functional‑connectivity‑derived graphs can exhibit an illusory small‑world topology purely as a by‑product of the statistical properties of correlation‑based measures. This insight has broad implications for any field that infers network structure from time‑series data, from brain imaging to climate science and social dynamics.