Abstract
The appropriate selection of recurrence thresholds is a key problem in applications of recurrence quantification analysis (RQA) 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. Based on our results, we recommend selecting the recurrence threshold adaptively according to a fixed quantile of this distribution. We highlight the advantages of this strategy over other previously suggested approaches by discussing the performance of selected RQA measures in detecting chaos--chaos transitions in some prototypical model system.
Introduction
Recurrence Plots (RPs) are a powerful tool for visualizing the dynamics of a system by marking when a trajectory returns to a previously visited state. However, most real-world data is univariate, requiring "time-delay embedding" to reconstruct a multi-dimensional state space. A major bottleneck in this process is selecting the threshold distance $\epsilon$, which defines how "close" two points must be to count as a recurrence.
Traditional rules of thumb—like picking $\epsilon$ as 5-10% of the maximum distance in the phase space—break down as the embedding dimension ($m$) increases. Due to the "curse of dimensionality," distances between points generally grow and concentrate differently depending on the chosen norm. This often results in "white" (empty) recurrence plots at higher dimensions, rendering RQA measures meaningless.
Methodology
The authors investigate how the distribution of pairwise distances changes as the embedding dimension $m$ increases, comparing two common metrics: the Maximum Norm ($L_\infty$) and the Euclidean Norm ($L_2$).
As shown in Figure 2, in the $L_2$ norm, both the maximum and mean distances grow with $\sqrt{m}$. In contrast, for the $L_\infty$ norm, the maximum possible distance stays constant while the mean distance increases. To address this, the paper proposes setting $\epsilon$ to a specific lower quantile (e.g., the 4th percentile) of the distance distribution. This ensures a constant Recurrence Rate (RR) regardless of the dimension $m$ or the norm used.
The study evaluates this method using random noise (uniform and Gaussian), harmonic oscillations, and the chaotic Rössler system. For the Rössler system, a non-stationary version is used where a control parameter $a$ increases over time, triggering a transition from "spiral chaos" to "screw-type chaos."
Key Insights and Results
As $m$ increases, the distance distribution shifts and changes shape. For $L_2$, it tends toward a Normal distribution (consistent with the Central Limit Theorem), while for $L_\infty$, it tends toward a reversed Weibull distribution (consistent with Extreme Value Theory).
Figure 6 illustrates how the distribution for the Rössler system shifts toward higher distances as $m$ increases. By fixing the quantile, the "density" of the recurrence plot remains invariant. This makes RQA measures like Laminarity and Transitivity comparable across different embedding choices, effectively decoupling the threshold selection from the "curse of dimensionality."
In comparative tests on the Rössler system:
- At low dimension ($m=3$): Traditional methods (percentage of max/mean) and the quantile method performed similarly, successfully detecting the chaos-chaos transition.
- At high dimension ($m=10$): Traditional methods failed. A threshold based on 11% of the max distance saw the Recurrence Rate drop from 4% to 0.5%, obscuring the dynamical transition.
- Quantile Success: The 4th-percentile threshold maintained a steady 4% Recurrence Rate at $m=10$, allowing RQA measures to clearly identify the transition just as well as they did at lower dimensions.
Conclusion
The quantile-based approach provides a dimension-scalable solution for RQA. While it solves the scaling problem, future work is needed to automate the initial selection of the specific quantile and to investigate the impact of serial correlations in real-world data on the convergence to limiting distributions.
Summary updated