Abstract
The appropriate selection of recurrence thresholds is a key problem in applications of recurrence quantification analysis and related methods across disciplines. Here, we discuss the distribution of pairwise distances between state vectors in the studied system's state space reconstructed by means of time-delay embedding as the key characteristic that should guide the corresponding choice for obtaining an adequate resolution of a recurrence plot. Specifically, we present an empirical description of the distance distribution, focusing on characteristic changes of its shape with increasing embedding dimension. Our results suggest that selecting the recurrence threshold according to a fixed percentile of this distribution reduces the dependence of recurrence characteristics on the embedding dimension in comparison with other commonly used threshold selection methods. Numerical investigations on some paradigmatic model systems with time-dependent parameters support these empirical findings.
Introduction
When analyzing complex systems using Recurrence Quantification Analysis (RQA), choosing the right distance threshold ($\epsilon$) is notoriously difficult. To analyze a dynamical system from a single observed time series, researchers use "time-delay embedding" to reconstruct the system's multi-dimensional state space. A key tool in this field is the Recurrence Plot (RP), which marks pairs of points in that space that are "close" to each other—closer than some threshold $\epsilon$.
The problem is that the choice of $\epsilon$ is often arbitrary. Common rules of thumb—like picking 10% of the maximum distance or a multiple of the standard deviation—fail when the embedding dimension ($m$) changes. As dimensions are added, the "curse of dimensionality" causes pairwise distances to increase and the shape of their distribution to shift. If the threshold selection doesn't account for this, the recurrence plot may lose its informative value, making it impossible to compare results across different embedding choices.
Methods and Distance Distributions
The study investigates how the distribution of pairwise distances changes as the embedding dimension $m$ increases, focusing on two common metrics: the Euclidean ($L_2$) norm and the Maximum ($L_\infty$) norm.
Mathematical analysis shows that in $L_\infty$, while the maximum distance stays constant as $m$ grows, the mean distance increases and the variance decreases. In $L_2$, both the mean and maximum distances grow with $m$. To validate a robust solution, the authors propose fixing the recurrence rate (RR) by choosing a specific percentile of the distance distribution. This was tested using a sliding-window analysis to calculate Recurrence Time Entropy (RTE), a measure sensitive to transitions between chaotic and periodic behavior.
Key Insights: The Concentration of Distribution
As the embedding dimension increases, the distribution of distances doesn't just shift; it "concentrates" or narrows. Points that were "close" in a lower dimension might be "far" in a higher one, but the relative variability of these distances shrinks.
By choosing a threshold that always captures a fixed percentile (e.g., the bottom 5% of distances), researchers can effectively "normalize" for the change in dimensionality. This preserves the density of the recurrence plot and ensures robustness even when the system is over-embedded (i.e., when $m$ is chosen higher than the system's true dimension).
Results from the Lorenz System
The most striking evidence for the fixed-percentile method comes from a non-stationary Lorenz-63 system. In this model, a control parameter drifts over time, causing transitions between chaos and limit cycles.
When using a threshold based on a fixed percentage of the maximum or mean distance, the RTE signal broke down as $m$ increased from 3 to 10, as recurrence plots became nearly empty. Conversely, when the threshold was set to a fixed percentile (e.g., the 4th percentile), the RTE estimates remained remarkably consistent across all embedding dimensions, accurately capturing the transitions regardless of the choice of $m$.
Conclusion and Future Work
Selecting $\epsilon$ based on a fixed percentile of the distance distribution is a robust way to ensure RQA results remain consistent across varying embedding dimensions. While this study focused on time-delay embedding and RTE, future work could explore other embedding methods (like derivative embedding) and other RQA measures (like determinism or laminarity). Additionally, theoretical work is needed to confirm if distance distributions for correlated time series converge to specific forms, such as the reversed Weibull distribution suggested for the $L_\infty$ norm at high dimensions.
Summary updated