Recurrence Quantification Analysis: Theory, Metrics, and Workflow

Recurrence quantification analysis (RQA) is an advanced nonlinear time-series methodology that quantifies the recurrence properties of dynamical systems by extracting statistical metrics from phase space recurrence plots. Originally formulated to overcome the limitations of classical linear and spectral analysis in short, noisy, and non-stationary signals, RQA characterizes deterministic structure, laminar stability, system complexity, and phase transitions without requiring assumptions of data stationarity or infinite trajectory length.

Definition and Theoretical Scope of RQA

In dynamical systems theory, recurrence is a fundamental property of state space trajectories. First described geometrically by Henri Poincaré in 1890, recurrence states that trajectories of bounded dynamical systems eventually return arbitrarily close to previously visited states. While linear methods such as the Fourier transform and traditional autocorrelation in time series data measure periodicities and global spectral power, they implicitly assume system linearity, Gaussian noise distributions, and stationarity. When applied to nonlinear physical, biological, or physiological systems that exhibit chaotic attractors or intermittent bursts, linear tools obscure critical deterministic transitions.

To visualize high-dimensional recurrence, J.-P. Eckmann, S. O. Kamphorst, and D. Ruelle introduced the recurrence plot (RP) in 1987. A recurrence plot is a two-dimensional square matrix that records the times at which a trajectory in reconstructed phase space visits roughly the same location. Formally, given a trajectory of N state vectors x(i) for i = 1, 2, …, N, the recurrence matrix R is defined as:

R(i, j) = Θ(ε – ||x(i) – x(j)||)

where ε is a predefined spatial distance threshold (radius), ||·|| denotes a chosen distance norm (such as Euclidean, Manhattan, or Maximum norm), and Θ is the Heaviside step function, where Θ(s) = 1 if s ≥ 0, and Θ(s) = 0 if s < 0. The resulting binary matrix contains ones (recurrence points, plotted as black dots) and zeros (white spaces).

Although recurrence plots provide striking qualitative visualizations, visual inspection alone cannot quantify subtle dynamic shifts or support rigorous hypothesis testing. In 1992, Joseph P. Zbilut and Charles L. Webber Jr. developed recurrence quantification analysis (RQA) to extract objective statistical measures from the density and structural arrangements of recurrence points. Later expanded by Norbert Marwan and colleagues in 2002 and 2007 to incorporate vertical structures and cross-system dynamics, RQA has become an essential analytical standard for evaluating nonlinear complexity across laboratory experiments, physiological recordings, and physical systems.

State Space Reconstruction and Parameter Calibration

Experimental data are rarely collected as multi-variable vector trajectories; instead, researchers typically record a single scalar observable s(t). To apply RQA, the underlying multi-dimensional state space must first be reconstructed from this scalar time series using Takens’ delay embedding theorem (1981). The reconstructed state vector x(i) at time index i is given by:

x(i) = [s(i), s(i + τ), s(i + 2τ), …, s(i + (m – 1)τ)]

where m is the embedding dimension and τ is the embedding time delay (lag). Successful RQA requires rigorous, reproducible calibration of four core parameters:

• Embedding Delay (τ): Selecting τ balances coordinate independence against phase space unfolding. If τ is too small, consecutive coordinates are virtually identical (redundancy); if τ is too large, deterministic relationships are lost (irrelevance). The standard protocol determines τ using the first local minimum of the Average Mutual Information (AMI) function proposed by Fraser and Swinney (1986). While linear autocorrelation zero-crossing offers a quick estimate, AMI accounts for nonlinear dependencies that dominate complex signals. Unlike stationary spectral transforms or windowed methods like the short-time Fourier transform, delay embedding reconstructs the geometric attractor directly in phase space.
• Embedding Dimension (m): The embedding dimension must be sufficiently large to unfold the attractor without overlapping trajectories caused by projection artifacts. The canonical standard is the False Nearest Neighbors (FNN) algorithm developed by Kennel, Brown, and Abarbanel (1992). By tracking how neighbor points separate as dimension m increases to m + 1, researchers select the dimension where the percentage of false neighbors drops below 1% (or reaches a stable plateau).
• Recurrence Threshold (ε): The distance threshold defines what constitutes a “recurrence.” If ε is set too small, noise prevents points from recurring, resulting in an empty matrix; if set too large, every state becomes recurrent, saturating the plot. The accepted standard protocol sets ε to a value between 5% and 10% of the maximum phase space diameter, or between 0.1 and 0.2 times the standard deviation of the time series (σ). Alternatively, researchers calibrate ε dynamically to enforce a fixed baseline recurrence rate (such as RR = 2% to 5%) across experimental conditions.
• Theiler Window (w): In continuous or finely sampled time series, successive measurements are geometrically close simply due to temporal continuity rather than true recurrence. If uncorrected, this forms an artificial thick diagonal band along the Line of Identity (LOI where i = j). The Theiler window w establishes a temporal exclusion zone around the LOI, discarding pairs where |i – j| < w from quantification. A robust guideline sets w to the first zero-crossing of the autocorrelation function or the decorrelation time of the signal.

Core Recurrence Quantification Metrics

RQA extracts quantitative indices based on the density of recurrence points, the distribution of diagonal lines, and the distribution of vertical/horizontal lines in the recurrence plot. Each geometric feature maps directly to a specific physical or dynamical property of the underlying system.

Point Density and Diagonal Line Metrics

Diagonal lines in a recurrence plot run parallel to the Line of Identity (LOI). A diagonal line of length l indicates that two segments of the trajectory remain close to each other for l consecutive time steps, representing deterministic and predictable evolution.

• Recurrence Rate (RR): The ratio of all recurrence points to the total number of possible points in the recurrence matrix. Excluding the LOI, RR = [1 / (N(N – 1))] Σ_i≠j R(i, j). It quantifies the overall probability that the system revisits a previous state.
• Determinism (DET): The percentage of recurrence points that form diagonal lines of length at least l_min (typically l_min = 2). Formally, DET = [Σ_l=lmin^N l P(l)] / [Σ_i,j=1^N R(i, j)], where P(l) is the frequency distribution of diagonal line lengths. Pure stochastic noise produces isolated points and very low DET (near 0%), whereas periodic or deterministic chaotic systems yield high DET (often exceeding 80% to 90%).
• Average Diagonal Line Length (L): The mean length of deterministic diagonal structures, calculated as L = [Σ_l=lmin^N l P(l)] / [Σ_l=lmin^N P(l)]. L reflects the average duration over which two nearby trajectories parallel each other, serving as a measure of mean predictability time.
• Longest Diagonal Line (L_max) and Divergence (DIV): L_max is the length of the longest diagonal line in the plot (excluding the LOI). Its mathematical inverse, Divergence (DIV = 1 / L_max), is proportional to the largest positive Lyapunov exponent (λ₁) of the system. Faster trajectory divergence indicates stronger chaotic instability and shorter L_max.
• Shannon Entropy of Line Lengths (ENTR): The Shannon entropy of the probability distribution of diagonal line lengths p(l) = P(l) / Σ_l=lmin^N P(l), defined as ENTR = -Σ_l=lmin^N p(l) ln p(l). ENTR measures the structural complexity of deterministic behaviors; a single periodic orbit has low entropy, while complex chaotic regimes exhibit diverse line length distributions and high entropy.
• Trend (TREND): A linear regression slope comparing recurrence point density in diagonal bands as distance from the LOI increases. TREND detects non-stationarity, drifts, or non-constant baseline shifts in the experimental protocol.

Vertical and Horizontal Line Metrics (Laminar Dynamics)

Introduced by Marwan et al. (2002), vertical and horizontal lines in a recurrence plot correspond to time intervals during which the system state changes very slowly or remains trapped in a localized phase space region. These features characterize laminar states and intermittency:

• Laminarity (LAM): The fraction of recurrence points that form vertical line structures of length at least v_min (typically v_min = 2). Formally, LAM = [Σ_v=vmin^N v P(v)] / [Σ_v=1^N v P(v)], where P(v) is the vertical line length distribution. High LAM indicates that the system frequently enters steady, stagnant, or laminar regimes.
• Trapping Time (TT): The average length of vertical line segments, calculated as TT = [Σ_v=vmin^N v P(v)] / [Σ_v=vmin^N P(v)]. TT estimates the mean duration that the physical or biological system remains trapped within a single state space neighborhood.
• Maximum Vertical Line Length (V_max): The length of the longest vertical structure, indicating the maximal duration of a laminar state.

Cross-Recurrence and Joint Recurrence Analysis

While standard RQA evaluates the self-recurrence of a single time series, its bivariate extensions analyze coupling between two distinct systems:

• Cross-Recurrence Quantification Analysis (CRQA): Evaluates recurrence between two different time series x and y embedded in the same phase space: CR(i, j) = Θ(ε – ||x(i) – y(j)||). Unlike the symmetric auto-recurrence matrix, CR is generally asymmetric. Diagonal lines displaced above or below the main diagonal reveal time lags, leader-follower dynamics, and phase synchronization between interacting agents or organs.
• Joint Recurrence Quantification Analysis (JRQA): Examines two systems that may have different physical dimensions or units by computing the Hadamard (element-wise) product of their individual recurrence matrices: JR(i, j) = R_x(i, j) × R_y(i, j). JRQA detects generalized synchronization when both systems simultaneously revisit their respective state spaces.

Worked Numerical Example: Calculating RQA Metrics

To understand the mechanical calculation of RQA metrics from a raw recurrence matrix, consider a compact illustrative trajectory of 5 time steps (N = 5). Suppose the reconstructed state vectors yield the following symmetric 5 × 5 binary recurrence matrix R:

Time Step (i / j)	j = 1	j = 2	j = 3	j = 4	j = 5
i = 1	1	1	0	0	0
i = 2	1	1	1	0	1
i = 3	0	1	1	1	0
i = 4	0	0	1	1	1
i = 5	0	1	0	1	1

In this matrix, the total number of elements is N² = 25. The Line of Identity (LOI) corresponds to the main diagonal entries (1,1), (2,2), (3,3), (4,4), (5,5), each equal to 1.

Step 1: Count Recurrence Points
The total number of recurrence points (entries with value 1) in the matrix is 13. When excluding the LOI (5 points), there are 8 off-diagonal recurrence points. Under the standard definition including the LOI, the recurrence rate is RR = 13 / 25 = 0.52 (52.0%). Under the convention excluding the LOI, RR = 8 / (5 × 4) = 8 / 20 = 0.40 (40.0%).

Step 2: Extract Diagonal Lines (Parallel to LOI)
Let us evaluate diagonal lines parallel to the main diagonal, using standard threshold l_min = 2. Looking at the upper triangle off-diagonals:

• First sub-diagonal (j – i = 1): entries are (1,2)=1, (2,3)=1, (3,4)=1, (4,5)=1. All 4 points are connected, forming a single diagonal line of length l = 4.
• Second sub-diagonal (j – i = 2): entries are (1,3)=0, (2,4)=0, (3,5)=0. No points.
• Third sub-diagonal (j – i = 3): entries are (1,4)=0, (2,5)=1. A single isolated point of length l = 1.
• Fourth sub-diagonal (j – i = 4): entry is (1,5)=0. No points.

Due to matrix symmetry, the lower triangle exhibits identical structures: one diagonal line of length l = 4 and one isolated point of length l = 1. Excluding the LOI, the diagonal line length distribution across both triangles is P(1) = 2, P(2) = 0, P(3) = 0, P(4) = 2. The points belonging to diagonal structures of length l ≥ 2 total (2 × 4) = 8 points.

Step 3: Calculate Determinism (DET)
Points belonging to diagonal lines of length l ≥ 2 comprise the two lines of length 4 (one above and one below the LOI), totaling 8 points. The total off-diagonal recurrence points equal 8. Therefore, DET = 8 / 8 = 1.0 (100.0%). If isolated points were distributed across off-diagonals, DET would measure the exact fraction of points trapped within contiguous trajectories.

Step 4: Extract Vertical Lines and Compute Laminarity (LAM)
Examining vertical columns with threshold v_min = 2:

• Column 1: points at i = 1 and i = 2 form a vertical line of length v = 2.
• Column 2: points at i = 1, 2, 3 form a vertical line of length v = 3; point at i = 5 is isolated (v = 1).
• Column 3: points at i = 2, 3, 4 form a vertical line of length v = 3.
• Column 4: points at i = 3, 4, 5 form a vertical line of length v = 3.
• Column 5: points at i = 4, 5 form a vertical line of length v = 2; point at i = 2 is isolated (v = 1).

Vertical lines with v ≥ 2 include: two lines of length 2 and three lines of length 3, contributing (2 × 2) + (3 × 3) = 4 + 9 = 13 points. Total points across all vertical segments (including isolated points) = 13 + 2 = 15. Thus, Laminarity LAM = 13 / 15 ≈ 0.867 (86.7%), and Trapping Time TT = 13 / 5 = 2.60 time steps.

Applications Across Laboratory Protocols and Experimental Research

Because RQA operates effectively on short, non-stationary, and noisy measurements, it is widely utilized across laboratory science, clinical research, and engineering protocols:

• Cardiovascular and Physiological Monitoring: In electrocardiography (ECG) and heart rate variability (HRV) analysis, autonomic nervous control induces non-linear fluctuations. Marwan et al. (2002) demonstrated that drops in determinism and surges in laminarity precede ventricular tachyarrhythmias, providing predictive indicators where classical spectral measures fail.
• Neuromuscular and Postural Control: Postural sway data recorded via force plates exhibit subtle intermittency. Trapping time and determinism quantify fatigue, neurological impairment (e.g., Parkinson’s disease), and motor control adjustments without requiring artificial stationarity filtering.
• Chemical Oscillations and Fluid Dynamics: Chemical reactors displaying Belousov-Zhabotinsky reactions and microfluidic mixing channels utilize RQA to detect bifurcations, laminar-to-turbulent transitions, and chaotic mixing regimes.
• Linguistics and Behavioral Coordination: Using Cross-RQA (CRQA) implemented in packages like R’s crqa, cognitive scientists quantify interpersonal coordination, conversational turn-taking, and gaze-tracking alignment between experimental subjects.

Implementing RQA on high-frequency signals requires addressing algorithmic scalability. Because constructing the distance matrix requires pairwise calculations, naive algorithms scale with quadratic computational complexity (O(N²)). When analyzing datasets where N exceeds 100,000 points, researchers utilize optimized libraries such as PyRQA (Rawald et al., 2017), which utilizes OpenCL parallel processing across GPUs and multi-core clusters in a standardized computational environment.

Comparative Analysis Across Dynamical Regimes

The diagnostic power of RQA lies in its ability to cleanly discriminate between distinct dynamical regimes. The data table below compares the characteristic quantitative signatures of canonical mathematical systems under standardized RQA parameters.

As evident from mathematical modeling, uncorrelated Gaussian white noise exhibits low recurrence density, negligible determinism, and minimal line lengths. Periodic signals generate near-total determinism (DET approaching 100%) with zero laminarity, as harmonic trajectories circulate without pausing in localized phase space regions. In contrast, chaotic systems like the Lorenz and Hénon attractors produce intermediate determinism alongside high Shannon entropy, reflecting diverse deterministic line distributions. Finally, intermittent regimes exhibit pronounced laminarity and elevated trapping times, directly capturing the alternation between chaotic bursts and laminar pauses.

Common Methodological Mistakes and Diagnostic Checks

To ensure rigorous, reproducible findings that comply with institutional methods reproducibility standards, researchers must avoid several common pitfalls in RQA execution:

• Omitting the Theiler Window: Failing to apply a Theiler window in finely sampled continuous signals causes spurious autocorrelations to inflate DET, L, and ENTR. Always check that the diagonal exclusion band matches or exceeds the decorrelation time.
• Inappropriate Threshold Selection: Setting an arbitrary threshold without calibration leads to unstable metrics. Researchers must perform threshold sensitivity sweeps or adopt fixed recurrence rates (e.g., RR = 5%) to ensure valid cross-condition comparisons.
• Under-Embedding the Phase Space: Setting an embedding dimension m lower than indicated by False Nearest Neighbors creates false trajectory crossings, corrupting diagonal line statistics. Always verify the FNN plateau before computing RQA.
• Conflating Non-Stationary Drift with Determinism: Slow linear baseline drifts can artificially generate long diagonal structures. Researchers should inspect the TREND metric; if TREND deviates substantially from zero, trend removal or windowed (epoch-based) RQA must be employed.
• Disregarding Minimum Line Length Conventions: Altering l_min or v_min between comparative experimental groups invalidates relative statistical comparisons. Adhere strictly to the standard literature conventions (l_min = 2, v_min = 2) unless explicitly justified.

Frequently asked questions

What is the difference between a recurrence plot and recurrence quantification analysis?

A recurrence plot is a two-dimensional graphical representation that visualizes when a trajectory in phase space revisits previous states, producing qualitative patterns such as dots, diagonals, and vertical lines. Recurrence quantification analysis (RQA) is the objective mathematical framework that extracts statistical metrics (such as recurrence rate, determinism, laminarity, and trapping time) from those graphical structures to characterize system dynamics quantitatively.

How do you select the threshold radius epsilon for RQA?

The recurrence threshold radius ε is typically selected using one of two established protocols: choosing a fixed percentage (commonly 5% to 10%) of the maximum phase space diameter or 0.1 to 0.2 times the standard deviation of the signal, or calibrating ε to yield a constant baseline recurrence rate (such as 2% to 5%) across all experimental conditions to permit unbiased comparative analysis.

Why is the Theiler window necessary in recurrence quantification?

The Theiler window is an exclusion zone around the main diagonal of a recurrence plot that removes points occurring close in time due solely to high sampling rates or linear autocorrelation rather than true dynamic recurrence. Omitting the Theiler window artificially inflates determinism, average line length, and entropy metrics.

Can RQA be used on short or non-stationary time series?

Yes. One of the primary advantages of RQA over classical Fourier and spectral methods is that it does not assume mathematical stationarity or require long data records. By utilizing windowed (sliding-epoch) RQA, researchers can track non-stationary transitions and transient states in short, noisy physiological and physical recordings.

What does determinism (DET) measure in RQA compared to random noise?

Determinism (DET) measures the proportion of recurrence points that form contiguous diagonal lines parallel to the line of identity. In purely stochastic random noise, recurrence points are isolated dots, yielding determinism values near zero. In deterministic periodic or chaotic systems, trajectories follow structured paths, resulting in high determinism values typically exceeding 80% to 90%.

What is the difference between auto-RQA and cross-RQA?

Auto-RQA analyzes the recurrence of a single system trajectory with itself to evaluate internal dynamics and stability. Cross-RQA (CRQA) compares the trajectories of two distinct time series embedded in a shared phase space, quantifying synchronization, coupling strength, and temporal lead-lag relationships between two interacting processes.

Sources

• Eckmann, J.-P., Kamphorst, S. O., and Ruelle, D. (1987). Recurrence Plots of Dynamical Systems. Europhysics Letters, 4(9), 973–977. https://doi.org/10.1209/0295-5075/4/9/004 (Accessed 2026-11-14)
• Zbilut, J. P., and Webber, C. L., Jr. (1992). Embeddings and delays in recurrence plots: Analyzing non-stationary physiological data. Physics Letters A, 171(3–4), 199–203. https://doi.org/10.1016/0375-9601(92)90426-M (Accessed 2026-11-14)
• Webber, C. L., Jr., and Zbilut, J. P. (1994). Dynamical assessment of physiological systems and states using recurrence plot strategies. Journal of Applied Physiology, 76(2), 965–973. https://doi.org/10.1152/jappl.1994.76.2.965 (Accessed 2026-11-14)
• Marwan, N., Wessel, N., Meyerfeldt, U., Schirdewan, A., and Kurths, J. (2002). Recurrence-plot-based measures of complexity and their application to heart-rate-variability data. Physical Review E, 66(2), 026702. https://doi.org/10.1103/PhysRevE.66.026702 (Accessed 2026-11-14)
• Marwan, N., Romano, M. C., Thiel, M., and Kurths, J. (2007). Recurrence plots for the analysis of complex systems. Physics Reports, 438(5–6), 237–329. https://doi.org/10.1016/j.physrep.2006.11.001 (Accessed 2026-11-14)
• Rawald, T., Sips, M., and Marwan, N. (2017). PyRQA – Conducting recurrence quantification analysis on very long time series efficiently. Computers and Geosciences, 104, 101–108. https://doi.org/10.1016/j.cageo.2016.11.016 (Accessed 2026-11-14)
• Coco, M. I., and Dale, R. (2014). Cross-recurrence quantification analysis of categorical and continuous time series: an R package. Frontiers in Psychology, 5, 510. https://doi.org/10.3389/fpsyg.2014.00510 (Accessed 2026-11-14)