Abstract
Recurrence networks are a powerful nonlinear tool for time series analysis of complex dynamical systems. While there are already many successful applications ranging from medicine to paleoclimatology, a solid theoretical foundation of the method has still been missing so far. Here, we interpret an ɛ-recurrence network as a discrete subnetwork of a "continuous" graph with uncountably many vertices and edges corresponding to the system's attractor. This step allows us to show that various statistical measures commonly used in complex network analysis can be seen as discrete estimators of newly defined continuous measures of certain complex geometric properties of the attractor on the scale given by ɛ. In particular, we introduce local measures such as the ɛ-clustering coefficient, mesoscopic measures such as ɛ-motif density, path-based measures such as ɛ-betweennesses, and global measures such as ɛ-efficiency. This new analytical basis for the so far heuristically motivated network measures also provides an objective criterion for the choice of ɛ via a percolation threshold, and it shows that estimation can be improved by so-called node splitting invariant versions of the measures. We finally illustrate the framework for a number of archetypical chaotic attractors such as those of the Bernoulli and logistic maps, periodic and two-dimensional quasiperiodic motions, and for hyperballs and hypercubes by deriving analytical expressions for the novel measures and comparing them with data from numerical experiments. More generally, the theoretical framework put forward in this work describes random geometric graphs and other networks with spatial constraints, which appear frequently in disciplines ranging from biology to climate science.
Introduction
Recurrence network analysis (RNA) has emerged as a significant methodology for characterizing the dynamics of complex systems. By mapping a time series into a graph—where nodes represent states in phase space and edges connect states within a distance $\varepsilon$—researchers leverage graph theory to quantify dynamical properties. Despite its success in fields like climatology and medicine, RNA has historically lacked a rigorous mathematical foundation. Metrics such as the clustering coefficient or betweenness centrality were often applied heuristically, without a clear understanding of their relationship to the underlying geometry of the system's attractor. Furthermore, the selection of the recurrence threshold $\varepsilon$ was frequently arbitrary.
The Continuous Framework
This work addresses these gaps by introducing a continuous framework for recurrence networks. Instead of relying on discrete adjacency matrices, the authors define an adjacency function $A(x, y)$ over the continuous set of points forming the system's attractor. This approach allows for the derivation of analytical definitions for continuous versions of standard network metrics, effectively treating discrete network measures as statistical estimators of these continuous integrals.
As summarized in the table above, the framework categorizes measures into neighborhood-based and path-based classes, providing heuristic interpretations that link network topology to geometric properties like local density, dimensionality, and centrality.
Geometric Interpretation and Results
The framework was validated using archetypal systems, including chaotic maps (Bernoulli, Logistic) and geometric manifolds (hyperballs, hypercubes). A key insight is that the network's topology serves as a direct proxy for the attractor's geometry. For instance, the local clustering coefficient is intimately related to local dimensionality, while path-based measures are sensitive to global shape and boundaries.
Figure 2 illustrates how these measures behave across different geometries. Notably, betweenness centrality ($b_1, b_2$) acts as a powerful tool for identifying "bottlenecks" in phase space, such as the inward corners of non-convex attractors like the L-shape shown in the bottom row. Numerical experiments confirmed that discrete estimators converge to these analytical values as the sample size $N$ increases.
Parameter Selection and Bias Reduction
The study also provides an objective criterion for selecting the threshold $\varepsilon$ using percolation theory. By identifying a critical threshold $\varepsilon_c$ where a "giant component" forms, researchers can ensure the network captures the attractor's global structure without losing fine-scale resolution. Additionally, the introduction of "node-splitting invariant" measures helps reduce statistical bias inherent in finite-sample networks, significantly improving estimation accuracy for small datasets.
Conclusion and Future Directions
While this framework provides a solid foundation for smooth manifolds, extending it to fractal geometries (strange attractors) remains a primary challenge. Future research will also need to address boundary effects in finite-length time series and the impact of non-isotropic scaling. Nevertheless, this theoretical bridge between complex network theory and nonlinear dynamics offers a principled approach to interpreting the physical meaning of recurrence network metrics.
Summary updated