How to avoid potential pitfalls in recurrence plot based data analysis

Recurrence plots and recurrence quantification analysis have become popular in the last two decades. Recurrence based methods have on the one hand a deep foundation in the theory of dynamical systems and are on the other hand powerful tools for the investigation of a variety of problems. The increasing interest encompasses the growing risk of misuse and uncritical application of these methods. Therefore, we point out potential problems and pitfalls related to different aspects of the application of recurrence plots and recurrence quantification analysis.

💡 Research Summary

The paper provides a comprehensive warning and guidance on the growing misuse of Recurrence Plots (RPs) and Recurrence Quantification Analysis (RQA) across many scientific fields. It begins by outlining the theoretical foundation of RPs: the construction of a binary matrix that records when state vectors in a reconstructed phase space are within a chosen distance ε of each other. The authors emphasize that three parameters—embedding dimension (m), time delay (τ), and distance threshold (ε)—govern the quality of the plot and the derived RQA measures.

The first major pitfall discussed is inadequate data preprocessing. Real‑world time series often contain trends, seasonal components, non‑stationarities, and measurement noise. If these features are not removed or appropriately modeled, the resulting RP can display spurious structures that do not reflect the underlying dynamics. The authors recommend detrending, seasonal adjustment, and, when necessary, non‑linear filtering, but caution that over‑filtering can erase essential dynamical information.

Next, the paper scrutinizes the embedding step. Selecting m and τ using standard algorithms such as False Nearest Neighbors (FNN) and average mutual information is common, yet both methods are highly sensitive to data length and noise level. The authors propose a multi‑criterion approach: run several candidate values, visualize the resulting attractors, and assess the stability of RQA metrics across these choices. Failure to embed correctly leads to “topological folding” where distinct regions of the true attractor are projected onto each other, producing misleading recurrence patterns.

The choice of ε is identified as the most delicate decision. A fixed ε may yield either an overly sparse RP (few recurrence points) or an overly dense one (loss of structural detail). The authors advocate fixing the global recurrence rate (RR) to a target value (e.g., 2–5 %) and adjusting ε accordingly, but they also note that systems with heterogeneous scaling may require local or adaptive thresholds. Multi‑scale RPs, where several ε values are examined simultaneously, can reveal scale‑dependent dynamics that a single ε would miss.

Statistical validation of RQA measures forms the fourth pillar of the analysis. Determinism (DET), laminarity (LAM), entropy (ENTR), and other indices are not independent; they are influenced by sample size, autocorrelation, and the choice of embedding parameters. The authors recommend surrogate data testing (phase‑randomized or amplitude‑adjusted surrogates) to establish a null distribution, and bootstrap resampling to compute confidence intervals. When multiple RQA metrics are examined, corrections for multiple comparisons (e.g., Bonferroni or false discovery rate) should be applied to avoid inflated Type I error rates.

Interpretation pitfalls are addressed in the fifth section. While RP visualizations can highlight periodicities, drift, or regime shifts, they do not prove causality. The authors stress that RP‑based findings should be cross‑validated with complementary nonlinear tools such as Lyapunov exponents, correlation dimension, or entropy measures. Moreover, the presence of non‑stationarity can masquerade as deterministic structure; therefore, stationarity tests and, if needed, windowed RP analyses are essential.

The final practical concern concerns software implementation. Different packages (MATLAB’s CRP toolbox, Python’s pyunicorn, R’s ‘crqa’) have distinct default settings for distance metrics, normalization, and plotting. These defaults can subtly alter the recurrence matrix and consequently the RQA outcomes. The authors call for full transparency: researchers must publish the exact code, parameter values, and preprocessing steps used, enabling reproducibility and peer verification.

In summary, the paper proposes a ten‑point checklist for reliable RP/RQA practice: (1) rigorous preprocessing and stationarity assessment; (2) careful selection and validation of embedding parameters; (3) adaptive or scale‑aware ε determination; (4) maintenance of a consistent recurrence rate; (5) statistical testing with surrogates and bootstraps; (6) correction for multiple hypothesis testing; (7) cross‑validation with other nonlinear diagnostics; (8) use of windowed or multi‑scale RPs for non‑stationary data; (9) explicit reporting of software, code, and parameter choices; and (10) cautious interpretation that avoids over‑stating causal claims. By adhering to these guidelines, researchers can mitigate the most common sources of error and harness the full analytical power of recurrence‑based methods.