July 15, 2010 0:13 pitfalls

International Journal of Bifurcation and Chaos

⃝c World Scientific Publishing Company

How to avoid potential pitfalls in recurrence plot based data

analysis
arXiv:1007.2215v1 [nlin.CD] 13 Jul 2010

Norbert Marwan
Potsdam Institute for Climate Impact Research, P.O. Box 601203, 14412 Potsdam, and
Interdisciplinary Centre for Dynamics of Complex Systems, University of Potsdam, 14415 Potsdam
Germany
[email protected]

Received (to be inserted by publisher)

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.

Keywords: recurrence plot, recurrence quantification analysis, time series analysis, pitfalls

1. Introduction
Since its introduction in 1987 by Eckmann et al. [1987], and the development of different quantification ap-
proaches, recurrence plots (RPs) have been widely used for the investigation of complex systems in a variety
of different disciplines, as physiology, ecology, finance or earth sciences [e.g., Marwan, 2008; Schinkel et al.,
2007; Zbilut et al., 2004; Facchini et al., 2007; Belaire-Franch, 2004; Pecar, 2003; Trauth et al., 2003;
Čermák, 2009]. RPs may attract attention because of their ability to produce beautiful or fancy pictures,
as in the case of the colourful representations of fractal sets [Mandelbrot, 1982]. The recent remarkable in-
crease of applications can be traced down in part to several free software packages available for calculating
recurrence plots and the corresponding recurrence quantification analysis (RQA). Since these methods are
also claimed to be very powerful even for short and non-stationary data, we should be careful not to con-
sider them as a kind of a magic tool, which works on all kinds of data. Owing to the fact that these methods
are indeed in some sense powerful and rather adaptable to various problems, it is really important that
the user knows how these methods work and has understood the ideas behind the RP and the measures of
complexity derived from it. Any uncritical application will lead to serious pitfalls and mis-interpretations.
As the number of applicants increases, the risk of careless application of RPs and RQA grows.
In this article we try to highlight some of the pitfalls which can occur during the application of RPs
and RQA and present future directions of research for a deep theoretical understanding of the method.

2. Recurrence plots and recurrence quantification

Although similar methods already existed before, the RP, Ri,j = Θ(ε − ||⃗xi − ⃗xj ||), for the analysis of the
dynamics of a dynamical system by using its phase space trajectory was introduced by Eckmann et al.

1
July 15, 2010 0:13 pitfalls

[1987]. This method can be used in order to visualise the recurrence of a state, i.e., all the times when this
state will recur. In the 1990’s, a heuristic approach of quantification RPs by its line structures has led to
the recurrence quantification analysis (RQA) [Webber Jr. & Zbilut, 1994; Marwan et al., 2002b]. In this
approach, the density of recurrence points as well as the histograms P (l) of the lengths l of the diagonal
and vertical lines in the RP are quantified. The density of recurrence points (recurrence rate) coincides with
the definition of the correlation dimension [Grassberger & Procaccia, 1983]. Moreover, RPs contain much
more information about the dynamics of the systems: dynamical invariants like Rényi entropy or correlation
dimensions can be derived from the structures in RPs [Faure & Korn, 1998; Thiel et al., 2004], RPs can be
used to study synchronisation [Romano et al., 2005; Senthilkumar et al., 2006] or to construct surrogate
time series [Thiel et al., 2008] and long time series from ensemble measurements [Komalapriya et al., 2010].
For a comprehensive introduction we point to [Marwan et al., 2007].

3. Pitfalls
3.1. Parameter choice for recurrence analysis
RP and RQA depend on some parameters which should be properly chosen. For the actual recurrence
analysis, a recurrence threshold is necessary. This measure is probably the most crucial one and is discussed
in the next subsection.
As already mentioned, the quantification of recurrence structures depends on lines in the RP; by
defining a minimal length of such lines, it is possible to adjust the sensitivity of line based recurrence
measures. In Subsect. 3.3 and 3.4 we will come back to this parameter.
If we start our recurrence analysis from a time series, we have first to reconstruct a phase space by using
a proper embedding, e.g., time-delay embedding [Packard et al., 1980]. This involves the proper setting of
two additional parameters: the embedding dimension m and the time-delay τ . Although the estimation of
dynamical invariants does not depend on the embedding [Thiel et al., 2004], the RQA measures depend on
the embedding. Standard approaches for finding optimal embedding parameters, like false nearest neigh-
bours for embedding dimension and auto-correlation or mutual information for time-delay, can be widely
found in the literature [e.g., Kantz & Schreiber, 1997]. However, it is recommended to visually cross-check
the embedding parameters by looking at the resulting RP. Non-optimal embedding parameters can cause
many interruptions of diagonal lines, small blocks, or even diagonal lines perpendicular to the LOI (this
corresponds to parallel trajectory segments running in opposite time direction; Fig. 1). The experience
has shown that the delay is sometimes overestimated by auto-correlation and mutual information. The
embedding dimension has also to be considered with care, as it artificially increases diagonal lines (will be
discussed in Subsect. 3.3) [Marwan et al., 2007].
In general, it is recommended to study the sensitivity (or robustness) of the results of the recurrence
analysis on the parameters (recurrence threshold, embedding parameters).
Although not really a parameter, it is worth to briefly discuss the different recurrence definitions.
The most frequently used definition is the to consider neighbours in the phase space which are smaller
than a threshold value (the recurrence threshold). Distances can be calculated using different norms, like
Maximum or Euclidean norm [Marwan et al., 2007]. Maximum norm is sometimes preferred because of its
better computational efficiency (only minor differences in the results when compared to Euclidean norm).
Another definition of recurrence considers a fixed amount nearest neighbours. This recurrence criterion
is used when the number of neighbours is important. Pitfalls related to these recurrence criteria are also
discussed in Subsect. 3.7. More interesting are combinations of the above criteria with dynamical properties
of the phase space trajectory, e.g., perpendicular RPs (Subsect. 3.3), or recurrence based on order patterns
[Groth, 2005]. Order patterns are representations of the local rank order of a given number d of values of
the time series (order pattern dimension). As the number of order patterns is equal to d!, the dimension
should not be chosen too large, because many order patterns will appear rather seldom and the RP will
be sparse. Even d = 4 is often already not appropriate, therefore, d = 3 is the best choice in most cases
(depending on the problem of interest, d = 2 may also be appropriate).
July 15, 2010 0:13 pitfalls

A B C
200 200 200

150 150 150

Time

Time

Time
100 100 100

50 50 50

0 0 0
0 50 100 150 200 0 50 100 150 200 0 50 100 150 200
Time Time Time

Fig. 1. Recurrence plots of the Rössler oscillator with parameters a = b = 0.25 and c = 40 using different embedding: (A)
m = 1, τ = 1, (B) m = 3, τ = 12, (C) m = 3, τ = 6 (adaptive recurrence threshold to ensure (A) RR = 0.1, (B, C) RR = 0.05).
Non-optimal embedding can cause line structures perpendicular to the main diagonal, wobbly or interrupted lines (A, B).

3.2. Recurrence threshold selection

The recurrence threshold ε is a crucial parameter in the RP analysis. Although several works have
contributed to this discussion [e.g., Thiel et al., 2002; Matassini et al., 2002; Marwan et al., 2007;
Schinkel et al., 2008], a general and systematic study on the recurrence threshold selection remains an
open task for future work. Nevertheless, recurrence threshold selection is a trade-off of to have a small
threshold as possible but at the same time a sufficient number of recurrences and recurrence structures.
However, the diversity of applicability of RP based methods causes a number of different criteria
for the selection of the threshold: studying dynamical properties (dynamical invariants, synchronisation)
requires a very small threshold [Marwan et al., 2007; Donner et al., 2010]; twin surrogates or trajectory
reconstruction methods may require larger thresholds [Hirata et al., 2008]; noise corrupted observation
data requires even larger thresholds [Thiel et al., 2002]; for studying dynamical transitions, the threshold
selection can be even without much importance, because the relative change of the RQA measures does
not depend too much on it in a certain range; for the detection of certain signals a specific fraction of the
phase space diameter (or standard deviation of the time series) can be required [Schinkel et al., 2008].
Several “rules of thumb” for the choice of ε have been advocated in the literature, e.g., a few per cent
of the maximum phase space diameter [Mindlin & Gilmore, 1992], a value that should not exceed 10%
of the mean or the maximum phase space ! diameter [Koebbe & Mayer-Kress, 1992; Zbilut & Webber Jr.,
1992], or that the recurrence rate RR = i,j Ri,j /N 2 is approximately 1% [Zbilut et al., 2002]. A recently
proposed criterion employing the relationship between recurrence rate and ε defines an optimal value by
using the position of the maximum of the first derivative of the recurrence rate dRR dε [Gao & Jin, 2009].
Such approach can produce ambiguous and highly unstable results, as slight variations in ε (as possible by
minor errors in finding this value or by nonstationary time series) cause high variation in the recurrence
structure. Next, the position of the maximum of dRR dε depends strongly on the chosen norm and embedding,
and may lead to an overestimation of an optimal ε. And, finally, there are systems which can have more
than one maximum [Donner et al., 2010].
Another criterion for the choice of ε takes into account that a measurement of a process is a composition
of the real signal and some observational noise with standard deviation σ [Thiel et al., 2002]. In order to
get similar results as for the noise-free situation, ε has to be chosen such that it is five times larger than
the standard deviation of the observational noise, i.e., ε > 5 σ. Although this criterion holds for a wide
class of processes, it is difficult to estimate the amount of observational noise in the signal.
For (quasi-)periodic processes, it has been suggested to use the diagonal structures within the RP in
order to find the optimal ε [Matassini et al., 2002]. In this approach, the density distribution of recurrence
points along the diagonals parallel to the LOI is investigated on dependence of ε in order to minimise the
fragmentation and thickness of the diagonal lines with respect to the threshold. However, this choice of ε
may not preserve the important distribution of the diagonal lines in the RP if observational noise is present
(the estimated threshold can be underestimated).
July 15, 2010 0:13 pitfalls

The selection of an optimal recurrence threshold ε is not straightforward and depends on the particular
problem and question.

3.3. Indicators of determinism

The length of a diagonal line in the RP corresponds to the time the system evolves very similar as during
another time, i.e., a segment of the phase space trajectory runs parallel and within an ε-tube of another
segment of the phase space trajectory. Deterministic systems are often characterised by repeated similar
state evolution (corresponding to a local predictability), yielding in a large number of diagonal lines in
the RP. In contrast, systems with independent subsequent values, like white noise, have RPs with mostly
single points. Therefore, the fraction of recurrence points forming such diagonal lines (of length l ≥ lmin )
!
l≥l lP (l)
DET = ! min (1)
i,j Ri,j
can be calculated and is, therefore, called determinism in the RQA. Somehow this measure can be inter-
preted as an indication of determinism in the data. But we should be careful in using the term determinism
in a more general or mathematical sense. In a deterministic system we can calculate the same exact state
by using given initial conditions, i.e., there is no stochastic process involved. Different methods can be used
to test for determinism in time series, e.g., a combined modelling-surrogate approach [Small & Tse, 2003]
or an analysis of the directionality of the phase space trajectory [Kaplan & Glass, 1992].
High values of DET might be an indication of determinism in the studied system, but it is just a
necessary condition, not a sufficient one. Even for non-deterministic processes we can find longer diag-
onal lines in the RP, resulting in increased DET values. For example, the following (non-deterministic)
auto-regressive process xi = 0.8xi−1 + 0.3xi−2 − 0.25xi−3 + 0.9ξ (where ξ is white Gaussian noise) has
a DET value of 0.6 (embedding dimension m = 4, delay τ = 4, and fixed recurrence rate of 0.1). As it
was shown in [Thiel et al., 2003], stochastic processes can have RPs containing longer diagonal lines just
by chance (although very rare). Moreover, due to embedding we introduce correlations in the RP and,
therefore, also uncorrelated data (e.g. from white noise process) have spurious diagonal lines [Thiel et al.,
2006; Marwan et al., 2007] (Fig. 2). Moreover, data pre-processing like low-passfiltering (smoothing) is
frequently used. Such pre-processing can also introduce spurious line structures in the RP. Therefore, from
just a high value of the RQA measure DET we have to be careful in infering that the studied system
would be deterministic. For such conclusion we need at least one further criterion included in the RP: the
directionality of the trajectory [Kaplan & Glass, 1992]. One possible solution is to use iso-directional RPs
[Horai et al., 2002] or perpendicular RPs [Choi et al., 1999]; if then the measure reaches DET ≈ 1 for a
very small recurrence density (i.e. RR < 0.05), the underlying system will be a deterministic one (like a
periodic or chaotic system).

3.4. Indicators of periodic systems

As explained in the previous section, deterministic systems cause a high value in the RQA measure
DET . This measure has been successfully used to detect transitions in the dynamics of complex sys-
tems [Trulla et al., 1996]. A frequently used example in order to present this ability is the study of the
different dynamical regimes of the logistic map, where DET is able to detect the periodic windows (by
values DET = 1). Therefore, it is often claimed that this measure is able to detect chaos-period transitions.
However, we can also find such high DET values for non-periodic, but chaotic systems. For example,
the Rössler system [Rössler, 1976],
" #
dx dy dz $ %
, , = −y − z, x + 0.25y, 0.25 + z(x − c) , (2)
dt dt dt
exhibits in the parameter interval c ∈ [35, 45] a transition from periodic to chaotic states (Fig. 3A). But
due to the smooth phase space trajectory and high sampling frequency (sampling time ∆t = 0.1), the RP
for the chaotic trajectory consists almost exclusively on diagonal line structures (Fig. 4), resulting in a
high value of DET , i.e., DET ≈ 1 (Fig. 3B).
July 15, 2010 0:13 pitfalls

A 1
70
B C 60
70
60 60 0.8 50

50 50 40

Correlation

Frequency
0.6

Time
Time

40 40 30
30 30 0.4
20
20 20
0.2
10
10 10
0 0
20 40 60 20 40 60 0 10 20 30
Time Time Diagonal line length

Fig. 2. (A) Recurrence plot of one realisation of Gaussian white noise, calculated using embeddeding dimension m = 6, delay
τ = 1, and a recurrence threshold of ε = 0.2. The embedding causes a number of long lines. (B) Correlation between a single
recurrence point at (15, 30) and other recurrence points in the RP of white noise demonstrating the effect of embedding for a
bogusly creation of long diagonal lines (estimated from 1,000 realisations). (C) The histogram of line lengths found in the RP
shown in (A). The maximum length is Lmax = 17, a value, which would not be uncommon for a deterministic process.

A 0.4

0.2
λ1,2

−0.2

−0.4
35 36 37 38 39 40 41 42 43 44 45

B 1

0.95

0.9
DET

0.85

0.8

0.75
35 36 37 38 39 40 41 42 43 44 45
Control parameter c

Fig. 3. (A) 1st and 2nd positive Lyapunov exponents of the Rössler oscillator with parameters a = b = 0.25 and c ∈ [35, 45].
A periodic window occurs between c = 36.56 and c = 37.25. However, the DET measures reveals an almost constant very
high value of approximately DET = 0.94. Used RP parameters: dimension m = 3, delay τ = 6, adaptive recurrence threshold
to ensure a RR = 0.05.

A very high value of DET is not a clear or even sufficient indication of a periodic system. High
values can be caused by very smooth phase space trajectories. This should also be considered when looking
for indications of unstable periodic orbits (UPOs), where DET or mean and maximal line lengths L
and Lmax may not be sufficient. A solution could be to increase the minimal length lmin of a diagonal
recurrence structure which is considered to be a line. However, a better solution is to look at the cumulative
distribution of the diagonal line lengths and estimate the K2 entropy (but this requires much longer time
series, cf. Subsect. 3.9). Recent work has shown that measures coming from complex network theory, like
July 15, 2010 0:13 pitfalls

450

400

350

Time 300

250

200

150

100

50

50 100 150 200 250 300 350 400 450

Time

Fig. 4. Recurrence plot of the Rössler oscillator with parameters a = b = 0.25 and c = 40. For this parameters, the Rössler
system is in a chaotic regime (λ1 = 0.14), but the RP consists almost only on diagonal lines. Used RP parameters: dimension
m = 3, delay τ = 6, adaptive recurrence threshold to ensure a RR = 0.05.

clustering coefficient, applied to recurrence matrices are more powerful and reliable for the detection of
periodic dynamics [Marwan et al., 2009; Zou et al., subm.; Donner et al., XXXX].

3.5. Indicators of chaos

The RP visualises the recurrence structure of the considered system (based on the phase space trajectory).
The basic idea behind RPs comes, in general, from the study of chaos. Therefore it can be considered as
a nonlinear tool for data analysis. But this cannot be a criterion to understand complex structures in the
RP or high values of RQA measures as indicators of chaos or nonlinearity in the dynamical system.
As mentioned above, uncorrelated stochastic systems have mostly short or almost no diagonal line
structures in their RPs, whereas deterministic and regular systems, like periodic processes, have mostly
long and continuous diagonal line structures. Chaotic processes have also diagonal, but shorter lines, and
can have single recurrence points. Nevertheless, only by looking at the appearance of an RP it is difficult
(almost impossible) to infer about the type of dynamics; only periodic and white noise processes can be
identified with some certainty.
The alternative is to look at the RQA measures quantifying the structures in an RP which are related
to some dynamical characteristics of the system. As diagonal lines in the RP correspond to parallel running
trajectory segments, it is clear that the length of these lines is somehow related to the divergence behaviour
of the dynamical system. Divergence rate of phase space trajectories is measured by the Lyapunov exponent.
In fact, the lengths of the diagonal lines are directly related to dynamical invariants as K2 entropy or D2
correlation dimension [Faure & Korn, 1998; Thiel et al., 2004]. The K2 entropy is the lower limit of the
sum of the positive Lyapunov exponents.
For example, RQA measures based on the length of the diagonal lines, like determinism DET and mean
line length L, also depend on the type of the dynamics of the systems (rather low values for uncorrelated
stochastic (white noise) systems, higher values for more regular, correlated and also chaotic systems).
It has been suggested to measure the length of the longest diagonal line Lmax and interpret its inverse
DIV = 1/Lmax as an estimator of the maximal Lyapunov exponent [Trulla et al., 1996]. However, this
interpretation incorporates high potential of erroneous conclusions derived from RQA.
First, the main diagonal in the RP (i.e., the line of identity, LOI) is naturally the longest diagonal
July 15, 2010 0:13 pitfalls

line, wherefore it is usually excluded from the analysis. However, due to the tangential motion of the phase
space trajectory1 , subsequent phase space vectors are often also considered as recurrence points (known
as sojourn points) [Marwan et al., 2007]. These recurrence points lead to further continuous diagonal lines
directly close to the LOI. Without excluding an appropriate corridor along the LOI (the Theiler window),
Lmax will be artificially large (≈ N ) and DIV too small.
Second, as explained above, even white noise can have long diagonal lines [Thiel et al., 2003], leading
to a small DIV value just by chance (Fig. 2). Although the probability for the occurrence of such long
lines is rather small, the probability that lines of length two occur in RPs of stochastic processes is, on the
contrary, rather high. Only one line of length two is enough to get a finite value of DIV which might be
mis-interpreted as a finite Lyapunov exponent and that the system would be chaotic instead stochastic.
Therefore, we have to be careful in interpreting the RQA measures themselves as indicators of chaos.
Moreover, such conclusion cannot be drawn by applying a simple surrogate test where the data points are
simply shuffled (such a test would only destroy the correlation structure within the data, and, thus, the
frequency information).
RP or RQA alone cannot be used to infer nonlinearity from a time series. For this purpose, advanced
surrogate techniques are more appropriate [Schreiber & Schmitz, 2000; Rapp et al., 2001].

3.6. Discrimination analysis and detection of deterministic signals

RQA is also a powerful tool in order to distinguish between different types of signals, different groups of
dynamical regimes etc. [e.g., Zbilut et al., 1998; Marwan et al., 2002b; Facchini et al., 2007; Litak et al.,
2010]. However, the selection of applicable RQA measures is a crucial task. Not all measures will be useful
for all questions. Their application needs justification in terms of the purpose of the intended analysis. For
example, for processes which does not contain laminar regimes, or if we are not interested in the detection
of such laminar regimes, it would not make sense to use RQA measures basing on vertical recurrence
structures (like laminarity or trapping time) [Marwan & Kurths, 2009].

3.7. Indicators of nonstationarity and transition analysis

RQA is powerful for the analysis of slight changes and transitions in the dynamics of a complex system.
For this purpose we need a time-dependend RQA (a RQA series) what can be realised in two ways (Fig. 5):
(1) The RP is covered with small overlapping windows of size w spreading along the LOI and in which
the RQA will be calculated, Ri,j |k+w−1
i,j=k .
(2) The time series (or phase space trajectory) is divided into overlapping segments xi |k+w−1 i=k from
which RPs and subsequent RQA will be calculated separately.
Such time dependent approach can also be used to analyse the stationarity of the dynamical system.
Here we should note the following important points. The time scale of the RQA values depends on the
choice which point in the window should be considered as the corresponding time point. Selecting the first
point k of the window as the time point of the RQA measures allows to directly transfer the time scale of
the time series to the RQA series. However, the window reaches into the future of the current time point
and, thus, the RQA measures represent a state which lies in the future. Variations in the RQA measures
can be misinterpreted as early signs of later state transitions (like a prediction). A better choice is therefore
to select the centre of the window as the current time point of the RQA. Then the RQA considers states in
the past and in the future. If strict causality is required (crucial when attempting to detect subtle changes
in the dynamics just prior the onset of dramatic state changes), it might be even useful to select the end
point of the window as the current time point of the RQA (using embedding we have to add (m − 1)τ − 1).
For most applications the centre point should be appropriate.
Another important issue can rise from the different windowing methods (1) or (2), which are only
equivalent when we do not normalise the time series (or its pieces) from which the RP is calculated and
when we chose a fixed threshold recurrence criterion. If we normalise the time series just before the RP

1
Tangential motion becomes even more crucial and influential for highly sampled or smooth systems.
July 15, 2010 0:13 pitfalls

Fig. 5. Two possibilities of windowed RQA: (A) Windowing of time series and (B) windowing of RP. The example is an
auto-regressive process: xi = 0.95xi−1 + 0.05xi−2 + 0.9ξ (where ξ is white Gaussian noise), the RP is calculated using a
constant number of neighbours (10% of all points) and without embedding. The sub-RPs at the bottom clearly demonstrate
the differences between the two approaches.

Table 1. Selected RQA measures derived from windowing of time series (left) and windowing of RP (right) of
an auto-regressive process and windowing as shown in Fig. 5.
Window 1–250 251–500 501–750 751–1000 Window 1–250 251–500 501–750 751–1000
RR 0.10 0.10 0.10 0.10 RR 0.18 0.12 0.20 0.19
DET 0.62 0.74 0.48 0.79 DET 0.81 0.81 0.69 0.95
L 3.13 3.69 2.75 3.75 L 3.78 4.27 2.90 9.50

calculation, we get differently normalised segments resulting in different sub-RPs (and thus different RQA
results) than such derived directly by moving windows from the RP of the entire time series (Fig. 5 and
Tab. 1). A similar problem arises when we use a fixed number of nearest neighbours for the definition of
recurrence, because it is a big difference considering the entire time series in order to find the k nearest
neighbours or just a small piece of it. Nevertheless, both approaches (1) and (2) can be useful and depend
on the given question. If we know that the time series shows some nonstationarities or trends which are
not of interest, then approach (2) can help to find transitions neglecting these nonstationarities. But, if we
are interested in the detection of the overall changes (e.g., to test for nonstationarity), we should keep the
numerical conditions for the entire available time constant and chose approach (1). Anyway, for each RQA
we should explicitly state how the windowing procedure has been performed.
The choice of the window size itself needs the same attention. Because the RQA measures are statistical
measures derived from histograms, the window should be large enough to cover a sufficient number of
recurrence lines or orbits. A too small window can pretend strong fluctuations in the RQA measures just
by weak statistical significance (the RQA measure T REN D is very sensitive to the window size and can
reveal even contrary results, cp. Fig. 7B). Therefore, conclusions about nonstationarity of the system should
be drawn with much care. Moreover, statements on stationarity of the system itself are questionable at all
(if not enough knowledge about the system is available), because detected nonstationarity in an observed
finite time series does not mean automatically nonstationarity in the underlying system. For example, an
auto-regressive process is stationary by definition, but its RP and RQA can reveal a nonstationary signal
(Figs. 6 and 7).
July 15, 2010 0:13 pitfalls

1000

900

800

700

600
Time (a.u.)
500

400

300

200

100

200 400 600 800 1000

Time (a.u.)

Fig. 6. RP of the same auto-regressive process as presented in Fig. 5, which is by definition stationary. The RP is calculated
using maximum norm, ε = 2 and without embedding.

A 0.15 B 0
75
150
−0.5 250
0.1
TREND
RR

−1

0.05
−1.5

0 −2
0 200 400 600 800 1000 0 200 400 600 800 1000
Time (a.u.) Time (a.u.)

Fig. 7. Two exemplary RQA measures, (A) recurrence rate RR and (B) paling trend (T REN D), of the auto-regressive
process as presented in Fig. 5 for three different window sizes w (w = 75, 150, 250). (A) The strong variation in RR pretends
a nonstationarity in the signal. (B) T REN D depends rather strongly on w, resulting in contrary outcomes, e.g., revealing
high values for w = 250, but small values for w = 75 at the same time period t = 700 . . . 800. The RQA is calculated using
maximum norm, ε = 0.3 and without embedding (the windows are moved by w/2, i.e., 50% overlap; the RQA time point is
set to the centre of the RQA window).

3.8. Significance of RQA measures

Related to the preceding issue on windowed RQA is the question on the significance of the RQA variation.
A sub-optimal scaling of the variation of the RQA measures can mislead to conclusions that the studied
system has changed its regime or that it would be nonstationary (Fig. 8A, B). Therefore, it is strongly
recommended to cross-check the scaling of the presentation and to present confidence intervals (Fig. 8C,
D). Confidence intervals can be calculated in various ways, but we should avoid to derive them by simply
shuffling the original data. One approach could be a bootstrap resampling of the line structures in the RP
[Marwan et al., 2008; Schinkel et al., 2009]. Another approach fits the probability of serial dependences
(diagonal lines) to a binomial distribution [Hirata & Aihara, XXXX]. Whatever approach we chose, the
estimation of the confidence intervals is not a trivial task, but in the future the standard software for RQA
should include such tests.
July 15, 2010 0:13 pitfalls

10

A 0.4 B 0.5

0.38

0.36 0.45

LAM
DET

0.34

0.32 0.4

0.3
0 200 400 600 800 1000 0 200 400 600 800 1000
Time (a.u.) Time (a.u.)

C 1 D 1

0.8 0.8

0.6 0.6

LAM
DET

0.4 0.4

0.2 0.2

0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Time (a.u.) Time (a.u.)

Fig. 8. Two exemplary RQA measures, (A, C) determinism DET and (B, D) laminarity (LAM ), of the auto-regressive
process as presented in Fig. 5. (A, B) The scaling of the y-axis is affecting a strong variation in the RQA measures – a
potential of wrong conclusions. (C, D) Considering a 5% confidence interval of the RQA measures (details can be found in
[Marwan et al., 2008]) and a better value range for the y-axis, we cannot infer that the values of the RQA measures as shown
in (A) and (B) significantly vary. The RQA is calculated using a window size of w = 250 and a window step of ws = 20, using
maximum norm, ε = 0.3 and without embedding (the RQA time point is set to the centre of the RQA window). LAM is the
fraction of recurrence points forming vertical lines in an RP (analogously as DET for the diagonal lines).

A common statement on recurrence analysis is that it is useful to analyse short data series. But we
have to ask, how short is short? The required length for the estimation of dynamical invariants will be
discussed in the following Subsect. Applying RQA analysis we should be aware that the RQA measures
are statistical measures (like an average) and need some minimal length that a variation can be considered
to be significant.

3.9. Dynamical invariants from short time series

An RP analysis is appropriate for analysing short and nonstationary time series, as it is often stated in
many reports [Fabretti & Ausloos, 2005; Schinkel et al., 2007; Zbilut et al., 1998]. However, this statement
holds actually only for the heuristic measures of complexity as introduced for the RQA or for the detection
of differences or transitions in data series. If we are interested in the dynamical invariants derived from
RPs, the length N of the time series becomes a more crucial part like it is for the standard methods of
nonlinear data analysis.
The derivations of dimensions (D1 , D2 ) and dynamical invariants (like K2 ) from the RPs hold only
in the limit N → ∞ and small ε (ε → 0). Nevertheless, an estimation of dynamical invariants from
shorter time series can be feasible. We have to regard the following factors if discussing the time series
length: the number of orbits representing stretching, the number of recurrences filling out a sufficient part
of the attractor, and the number of data points necessary for an acceptable phase space reconstruction
[Wolf et al., 1985]. Since these factors may require different minimal lengths, the largest of these lengths
should be considered.
For example, numerical considerations for the estimation of the attractor (correlation) dimension
D2 using the Grassberger-Procaccia algorithm [Grassberger & Procaccia, 1983] lead to the requirement
July 15, 2010 0:13 pitfalls

11

log N > D22 log( ϱ1 ) (where ϱ = Sε is the fraction the recurrence neighbourhood of size ε covers on the entire
phase space of diameter S) [Eckmann & Ruelle, 1998]. Considering a ϱ = 0.1 and a decimal logarithm, for
finding a D2 = 10 we need at least N = 100, 000 data points. Furthermore, a ϱ = 0.1 is actually too large
and we need much smaller ε, which consequently provokes that again a larger N is required.
For Lyapunov exponents (and analogously for K2 ), a rough estimate based on the mentioned require-
ments suggests minimal time series lengths of 10D2 to 30D2 (with attractor dimension D2 ) [Wolf et al.,
1985]. Accordingly, a system with D2 = 3 requires 1000–30,000 data points (a more strict consideration
even requires log N > D2 log( ϱ1 ) [Eckmann & Ruelle, 1998]).
Therefore, too guarantee useful results we need long time series. If we calculate dimensions or K2 from
short time series the results are probably worthless.

3.10. Synchronisation and line of synchronisation

Cross recurrence plots (CRPs) can be used for the investigation of the simultaneous evolution of two
different phase space trajectories [Marwan et al., 2002a; Marwan & Kurths, 2002; Zolotova et al., 2009;
Ihrke et al., 2009]. The line of identity (LOI) in the RP becomes a line of synchronisation (LOS) in the
CRP. Two more-or-less identical systems but with differences on the time-scale will reveal a bowed LOS
[Marwan et al., 2002a; Marwan & Kurths, 2005]. An off-set of the LOS away from the main diagonal is an
indication of a phase shift or a delay between the two considered systems.
However, because this method tests if the two trajectories visit the same region in the phase space, it can
be used only to study complete synchronisation (CS) or a kind of a generalised correlation (although with
possible delays), or to get the relation between the transformations between their time-scales. Moreover,
the data under consideration should be from the same (or a very comparable) process and, actually, should
represent the same observable. Therefore, the reconstructed phase space should be the same.
For the study of the LOS the distance matrix may be more appropriate because it contains more infor-
mation, especially if the data series show nonstationarities. Then, the LOS can be found by using efficient
algorithms like dynamic time warping [Sakoe & Chiba, 1978]. Nevertheless, it is always very important to
check if the found LOS makes sense; for instance, it is possible to find several LOS (cp. Application in
magneto-stratigraphy in [Marwan et al., 2007]).

3.11. Macrostructures and sampling

For the visual interpretation of an RP and also for a reliable RQA we should remember that our data
are discretised time or data series. The sampling of the signal has an importance which should not be
underestimated. If the sampling frequency is just one magnitude higher than the system’s main frequencies,
and their ratio is not a multiple of an integer (i.e., we have an intrinsic phase error), an interference triggered
by the sampling of the continuous signal can produce large empty regions in the recurrence matrix, although
they should be there [Facchini et al., 2005; Facchini & Kantz, 2007]. Nonstationarities or modulations in
frequency or phase cause non-trivial gaps or macrostructures in the recurrence matrix (Fig. 9). We should
be aware that such gaps can occur in particular when we use a low sampling frequency. The recurrence
structure of interest can appear rather different; diagonal lines can vanish or can be reduced to just single
points yielding in biased RQA measures (Fig. 10).
Nevertheless, tiny modulations in frequency or phase in oscillating signals can be detected by RPs,
which are non-detectable by standard methods (spectral or wavelet analysis). This turns RPs to a powerful
tool for the analysis of slight modulations in oscillatory signals like audio signals.
Please note that macrostructures are also an apparent problem when displaying large RPs on a com-
puter screen (and up to a certain amount on print outs). The resolution of modern computer screens is
around 72 ppi (points per inch, 72 ppi corresponds to around 28 points per centimetre). The presenta-
tion of RPs in a window of, e.g., 6 inch allows only the display of around 430 points. Larger RPs will be
rendered using downsampling or interpolation, resulting in similar interference effects and artificial sec-
ondary macrostructures as described above; such macrostructures will even change for different window
sizes (Fig. 11). Therefore, we should take care in visual interpretation of patterns found in large RPs which
are represented on computer screens.
July 15, 2010 0:13 pitfalls

12

A 0.05 B 0.05

0.04 0.04

0.03 0.03
Time

Time
0.02 0.02

0.01 0.01

0 0
0 0.01 0.02 0.03 0.04 0.05 0 0.01 0.02 0.03 0.04 0.05
Time Time

Fig. 9. RP of a modulated harmonic oscillation sin(2π1000(π + t) + 2π sin(2π44t)t). (A) Non-trivial macrostructures (gaps)
in the RP due to the interference of the sampling frequency of 1 kHz and the frequency of the modulated harmonic signal. (B)
Corresponding RP as shown in (A), but for a higher sampling frequency of 10 kHz. As expected, the entire RP now consists
of the periodic line structures due to the oscillation. Used RP parameters: dimension m = 3, delay τ = 1, recurrence threshold
ε = 0.05σ, L∞ -norm.

A B
140 140

120 120

100 100
Time (s)

Time (s)

80 80

60 60

40 40

20 20

0 0
0 50 100 0 50 100
Time (s) Time (s)

Fig. 10. RP of the x-component of the Rössler oscillator, Eq. (2), with parameters a = b = 0.2, c = 5.7. The sampling time
is (A) ∆t = 0.05 and (B) ∆t = 1. The embedding was chosen in both settings to be equivalent: dimension in (A) and (B) is
m = 3, the delay in (A) τ = 20, but in (B) τ = 1; recurrence threshold ε = 1.5 (maximum norm). Due to the low sampling in
(B), many diagonal lines vanish.

4. Conclusions
We have illustrated several problems regarding the application of recurrence plots (RPs) and recurrence
quantification analysis (RQA) which need our attention in order to avoid wrong results. The uncritical
application of these methods can yield to serious pitfalls. Therefore, it is important to understand the basic
principles and ideas behind the measures of complexity forming the RQA and the different techniques to
study the numerous phenomena of complex systems, like transitions, synchronisation, etc. Nevertheless,
the recurrence plot based techniques are still a rather young field in nonlinear time series analysis, and
many open questions remain. For example, systematic research is necessary to define reliable criteria for
July 15, 2010 0:13 pitfalls

REFERENCES 13

Fig. 11. Screenshot of the RP as shown in Fig. 9B for two different window sizes of display on a computer screen (using
Matlab⃝ R
). Although the RP consists only on continuous diagonal lines as represented in Fig. 9B, its size (N = 5511) exceeds
the screen resolution and requires downsampling, leading to artifical macrostructures.

the selection of the recurrence threshold, and the estimation of the confidence of the RQA measures will
be a hot topic in the near future.

Acknowledgments
The work has been supported by the Potsdam Research Cluster for Georisk Analysis, Environmental
Change and Sustainability (PROGRESS).

References
Belaire-Franch, J. [2004] “Testing for non-linearity in an artificial financial market: a recurrence quantifica-
tion approach,” Journal of Economic Behavior & Organization 54, 483–494, doi:10.1016/j.jebo.2003.
05.001.
Čermák, V. [2009] “Recurrence Quantification Analysis of Borehole Temperatures: Evidence of
Fluid Convection,” International Journal of Bifurcation and Chaos 19, 889–902, doi:10.1142/
S0218127409023366.
Choi, J. M., Bae, B. H. & Kim, S. Y. [1999] “Divergence in perpendicular recurrence plot; quantification of
dynamical divergence from short chaotic time series,” Physics Letters A 263, 299–306, doi:10.1016/
S0375-9601(99)00751-3.
Donner, R. V., Small, M., Donges, J. F., Marwan, N., Zou, Y. & Kurths, J. [XXXX] “Recurrence-based
time series analysis by means of complex network methods,” International Journal of Bifurcation and
Chaos (this special issue).
Donner, R. V., Zou, Y., Donges, J. F., Marwan, N. & Kurths, J. [2010] “Ambiguities in recurrence-based
complex network representations of time series,” Physical Review E 81, 015101(R), doi:10.1103/
PhysRevE.81.015101.
Eckmann, J.-P., Oliffson Kamphorst, S. & Ruelle, D. [1987] “Recurrence Plots of Dynamical Systems,”
Europhysics Letters 5, 973–977.
Eckmann, J.-P. & Ruelle, D. [1998] “Fundamental limitations for estimating dimensions and Lyapunov
exponents in dynamical systems,” Physica D 56, 185–187, doi:10.1016/0167-2789(92)90023-G.
Fabretti, A. & Ausloos, M. [2005] “Recurrence plot and recurrence quantification analysis techniques for
July 15, 2010 0:13 pitfalls

14 REFERENCES

detecting a critical regime. Examples from financial market indices,” International Journal of Modern
Physics C 16, 671–706, doi:10.1142/S0129183105007492.
Facchini, A. & Kantz, H. [2007] “Curved structures in recurrence plots: The role of the sampling time,”
Physical Review E 75, 036215, doi:10.1103/PhysRevE.75.036215.
Facchini, A., Kantz, H. & Tiezzi, E. B. P. [2005] “Recurrence plot analysis of nonstationary data: The
understanding of curved patterns,” Physical Review E 72, 021915, doi:10.1103/PhysRevE.72.021915.
Facchini, A., Mocenni, C., Marwan, N., Vicino, A. & Tiezzi, E. B. P. [2007] “Nonlinear time series analysis
of dissolved oxygen in the Orbetello Lagoon (Italy),” Ecological Modelling 203, 339–348, doi:10.1016/
j.ecolmodel.2006.12.001.
Faure, P. & Korn, H. [1998] “A new method to estimate the Kolmogorov entropy from recurrence plots:
its application to neuronal signals,” Physica D 122, 265–279, doi:10.1016/S0167-2789(98)00177-8.
Gao, Z. & Jin, N. [2009] “Flow-pattern identification and nonlinear dynamics of gas-liquid two-phase flow
in complex networks,” Physical Review E 79, 066303, doi:10.1103/PhysRevE.79.066303.
Grassberger, P. & Procaccia, I. [1983] “Characterization of strange attractors,” Physical Review Letters
50, 346–349, doi:10.1103/PhysRevLett.50.346.
Groth, A. [2005] “Visualization of coupling in time series by order recurrence plots,” Physical Review E
72, 046220, doi:10.1103/PhysRevE.72.046220.
Hirata, Y. & Aihara, K. [XXXX] “Statistical tests for serial dependence and laminarity on recurrence
plots,” International Journal of Bifurcation and Chaos (this special issue).
Hirata, Y., Horai, S. & Aihara, K. [2008] “Reproduction of distance matrices from recurrence plots
and its applications,” European Physical Journal – Special Topics 164, 13–22, doi:10.1140/epjst/
e2008-00830-8.
Horai, S., Yamada, T. & Aihara, K. [2002] “Determinism Analysis with Iso-Directional Recurrence Plots,”
IEEE Transactions - Institute of Electrical Engineers of Japan C 122, 141–147.
Ihrke, M., Schrobsdorff, H. & Herrmann, J. M. [2009] “Recurrence-Based Synchronization of Single Tri-
als for EEG-Data Analysis,” Lecture Notes in Computer Science: Intelligent Data Engineering and
Automated Learning – IDEAL 2009 5788, 118–125, doi:10.1007/978-3-642-04394-9.
Kantz, H. & Schreiber, T. [1997] Nonlinear Time Series Analysis (University Press, Cambridge).
Kaplan, D. T. & Glass, L. [1992] “Direct test for determinism in a time series,” Physical Review Letters
68, 427–430.
Koebbe, M. & Mayer-Kress, G. [1992] “Use of Recurrence Plots in the Analysis of Time-Series Data,”
Proceedings of SFI Studies in the Science of Complexity, eds. Casdagli, M. & Eubank, S. (Addison-
Wesley, Redwood City), pp. 361–378.
Komalapriya, C., Romano, M. C., Thiel, M., Marwan, N., Kurths, J., Kiss, I. Z. & Hudson, J. L. [2010]
“An automated algorithm for the generation of dynamically reconstructed trajectories,” Chaos 20,
013107, doi:10.1063/1.3279680.
Litak, G., Wiercigroch, M., Horton, B. W. & Xu, X. [2010] “Transient chaotic behaviour versus periodic
motion of a parametric pendulum by recurrence plots,” ZAMM – Journal of Applied Mathematics
and Mechanics/ Zeitschrift fr Angewandte Mathematik und Mechanik 90, 33–41, doi:10.1002/zamm.
200900290.
Mandelbrot, B. B. [1982] The fractal geometry of nature (Freeman, San Francisco).
Marwan, N. [2008] “A Historical Review of Recurrence Plots,” European Physical Journal – Special Topics
164, 3–12, doi:10.1140/epjst/e2008-00829-1.
Marwan, N., Donges, J. F., Zou, Y., Donner, R. V. & Kurths, J. [2009] “Complex network approach for
recurrence analysis of time series,” Physics Letters A 373, 4246–4254, doi:10.1016/j.physleta.2009.09.
042.
Marwan, N. & Kurths, J. [2002] “Nonlinear analysis of bivariate data with cross recurrence plots,” Physics
Letters A 302, 299–307, doi:10.1016/S0375-9601(02)01170-2.
Marwan, N. & Kurths, J. [2005] “Line structures in recurrence plots,” Physics Letters A 336, 349–357,
doi:10.1016/j.physleta.2004.12.056.
Marwan, N. & Kurths, J. [2009] “Comment on “Stochastic analysis of recurrence plots with applications
to the detection of deterministic signals” by Rohde et al. [Physica D 237 (2008) 619–629],” Physica
July 15, 2010 0:13 pitfalls

REFERENCES 15

D 238, 1711–1715, doi:10.1016/j.physd.2009.04.018.

Marwan, N., Romano, M. C., Thiel, M. & Kurths, J. [2007] “Recurrence Plots for the Analysis of Complex
Systems,” Physics Reports 438, 237–329, doi:10.1016/j.physrep.2006.11.001.
Marwan, N., Schinkel, S. & Kurths, J. [2008] “Significance for a recurrence based transition analysis,”
Proceedings of the International Symposium on Nonlinear Theory and its Applications (NOLTA2008),
Budapest (Budapest, Hungary), pp. 412–415.
Marwan, N., Thiel, M. & Nowaczyk, N. R. [2002a] “Cross Recurrence Plot Based Synchronization of Time
Series,” Nonlinear Processes in Geophysics 9, 325–331.
Marwan, N., Wessel, N., Meyerfeldt, U., Schirdewan, A. & Kurths, J. [2002b] “Recurrence Plot Based
Measures of Complexity and its Application to Heart Rate Variability Data,” Physical Review E 66,
026702, doi:10.1103/PhysRevE.66.026702.
Matassini, L., Kantz, H., Ho&lyst, J. A. & Hegger, R. [2002] “Optimizing of recurrence plots for noise
reduction,” Physical Review E 65, 021102, doi:10.1103/PhysRevE.65.021102.
Mindlin, G. M. & Gilmore, R. [1992] “Topological analysis and synthesis of chaotic time series,” Physica
D 58, 229–242, doi:10.1016/0167-2789(92)90111-Y.
Packard, N. H., Crutchfield, J. P., Farmer, J. D. & Shaw, R. S. [1980] “Geometry from a Time Series,”
Physical Review Letters 45, 712–716, doi:10.1103/PhysRevLett.45.712.
Pecar, B. [2003] “The use of Visual Recurrence Analysis and Hurst exponents as qualitative tools for
analysing financial time series,” .
Rapp, P. E., Cellucci, C. J., Korslund, K. E., Watanabe, T. A. & Jimenez-Montano, M. A. [2001] “Effective
normalization of complexity measurements for epoch length and sampling frequency,” Physical Review
E 64, 16209, doi:10.1103/PhysRevE.64.016209.
Romano, M. C., Thiel, M., Kurths, J., Kiss, I. Z. & Hudson, J. [2005] “Detection of synchronization
for non-phase-coherent and non-stationary data,” Europhysics Letters 71, 466–472, doi:10.1209/epl/
i2005-10095-1.
Rössler, O. E. [1976] “An equation for continuous chaos,” Physics Letters A 57, 397–398, doi:10.1016/
0375-9601(76)90101-8.
Sakoe, H. & Chiba, S. [1978] “Dynamic Programming Algorithm Optimization for Spoken Word Recogni-
tion,” IEEE Transactions on Acoustics, Speech and Signal Processing ASSP-26, 43–49.
Schinkel, S., Dimigen, O. & Marwan, N. [2008] “Selection of recurrence threshold for signal detection,”
European Physical Journal – Special Topics 164, 45–53, doi:10.1140/epjst/e2008-00833-5.
Schinkel, S., Marwan, N., Dimigen, O. & Kurths, J. [2009] “Confidence bounds of recurrence-based com-
plexity measures,” Physics Letters A 373, 2245–2250, doi:10.1016/j.physleta.2009.04.045.
Schinkel, S., Marwan, N. & Kurths, J. [2007] “Order patterns recurrence plots in the anaylsis of ERP
data,” Cognitive Neurodynamics 1, 317–325, doi:10.1007/s11571-007-9023-z.
Schreiber, T. & Schmitz, A. [2000] “Surrogate time series,” Physica D 142, 346–382, doi:10.1016/
S0167-2789(00)00043-9.
Senthilkumar, D. V., Lakshmanan, M. & Kurths, J. [2006] “Phase synchronization in time-delay systems,”
Physical Review E 74, 035205, doi:10.1103/PhysRevE.74.035205.
Small, M. & Tse, C. K. [2003] “Detecting Determinism in Time Series: The Method of Surrogate Data,”
IEEE Transactions on Circuits and Systems: Fundamental Theory and Applications 50, 663–672.
Thiel, M., Romano, M. C. & Kurths, J. [2003] “Analytical Description of Recurrence Plots of white noise
and chaotic processes,” Izvestija vyssich ucebnych zavedenij/ Prikladnaja nelinejnaja dinamika – Ap-
plied Nonlinear Dynamics 11, 20–30.
Thiel, M., Romano, M. C. & Kurths, J. [2006] “Spurious Structures in Recurrence Plots Induced by
Embedding,” Nonlinear Dynamics 44, 299–305, doi:10.1007/s11071-006-2010-9.
Thiel, M., Romano, M. C., Kurths, J., Meucci, R., Allaria, E. & Arecchi, F. T. [2002] “Influence of
observational noise on the recurrence quantification analysis,” Physica D 171, 138–152, doi:10.1016/
S0167-2789(02)00586-9.
Thiel, M., Romano, M. C., Kurths, J., Rolfs, M. & Kliegl, R. [2008] “Generating surrogates from recur-
rences,” Philosophical Transactions of the Royal Society A 366, 545–557, doi:10.1098/rsta.2007.2109.
Thiel, M., Romano, M. C., Read, P. L. & Kurths, J. [2004] “Estimation of dynamical invariants without
July 15, 2010 0:13 pitfalls

16 REFERENCES

embedding by recurrence plots,” Chaos 14, 234–243, doi:10.1063/1.1667633.

Trauth, M. H., Bookhagen, B., Marwan, N. & Strecker, M. R. [2003] “Multiple landslide clusters record
Quaternary climate changes in the northwestern Argentine Andes,” Palaeogeography Palaeoclimatology
Palaeoecology 194, 109–121, doi:10.1016/S0031-0182(03)00273-6.
Trulla, L. L., Giuliani, A., Zbilut, J. P. & Webber Jr., C. L. [1996] “Recurrence quantification analysis of
the logistic equation with transients,” Physics Letters A 223, 255–260, doi:10.1016/S0375-9601(96)
00741-4.
Webber Jr., C. L. & Zbilut, J. P. [1994] “Dynamical assessment of physiological systems and states using
recurrence plot strategies,” Journal of Applied Physiology 76, 965–973.
Wolf, A., Swift, J. B., Swinney, H. L. & Vastano, J. A. [1985] “Determining Lyapunov Exponents from a
Time Series,” Physica D 16, 285–317, doi:10.1016/0167-2789(85)90011-9.
Zbilut, J. P., Giuliani, A., Colosimo, A., Mitchell, J. C., Colafranceschi, M., Marwan, N., Uversky, V. N.
& Webber Jr., C. L. [2004] “Charge and Hydrophobicity Patterning along the Sequence Predicts the
Folding Mechanism and Aggregation of Proteins: A Computational Approach,” Journal of Proteome
Research 3, 1243–1253, doi:10.1021/pr049883+.
Zbilut, J. P., Giuliani, A. & Webber Jr., C. L. [1998] “Detecting deterministic signals in exceptionally
noisy environments using cross-recurrence quantification,” Physics Letters A 246, 122–128, doi:10.
1016/S0375-9601(98)00457-5.
Zbilut, J. P. & Webber Jr., C. L. [1992] “Embeddings and delays as derived from quantification of recurrence
plots,” Physics Letters A 171, 199–203, doi:10.1016/0375-9601(92)90426-M.
Zbilut, J. P., Zaldı́var-Comenges, J.-M. & Strozzi, F. [2002] “Recurrence quantification based Liapunov
exponents for monitoring divergence in experimental data,” Physics Letters A 297, 173–181, doi:
10.1016/S0375-9601(02)00436-X.
Zolotova, N. V., Ponyavin, D. I., Marwan, N. & Kurths, J. [2009] “Long-term asymmetry in the wings of
the butterfly diagram,” Astronomy & Astrophysics 505, 197–201, doi:10.1051/0004-6361/200811430.
Zou, Y., Donner, R. V., Donges, J. F., Marwan, N. & Kurths, J. [subm.] “Identifying shrimps in continuous
dynamical systems using recurrence-based methods,” Physical Review E .