Probabilistic analysis of recurrence plots generated by

fractional Gaussian noise

Sofiane Ramdani, Frédéric Bouchara, Annick Lesne

To cite this version:

Sofiane Ramdani, Frédéric Bouchara, Annick Lesne. Probabilistic analysis of recurrence plots gener-
ated by fractional Gaussian noise. Chaos: An Interdisciplinary Journal of Nonlinear Science, 2018, 28
(8), pp.#085721. �10.1063/1.5030522�. �lirmm-02050628�

HAL Id: lirmm-02050628

https://hal-lirmm.ccsd.cnrs.fr/lirmm-02050628
Submitted on 22 May 2019

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 CHAOS 28, 085721 (2018)

Probabilistic analysis of recurrence plots generated by fractional

Gaussian noise
Sofiane Ramdani,1,a) Frédéric Bouchara,2 and Annick Lesne3,4
1
LIRMM, University of Montpellier, CNRS, Montpellier, France
2
Université de Toulon, Aix Marseille University, CNRS, LIS, Toulon, France
3
Sorbonne Université, CNRS, Laboratoire de Physique Théorique de la Matière Condensée, LPTMC, F-75252
Paris, France
4
Institut de Génétique Moléculaire de Montpellier, University of Montpellier, CNRS, Montpellier, France
(Received 22 March 2018; accepted 10 August 2018; published online 30 August 2018)
Recurrence plots of time series generated by discrete fractional Gaussian noise (fGn) processes are
analyzed. We compute the probabilities of occurrence of consecutive recurrence points forming diag-
onals and verticals in the recurrence plot constructed without embedding. We focus on two recurrence
quantification analysis measures related to these lines, respectively, the percent determinism and the
laminarity (LAM ). The behavior of these two measures as a function of the fGn’s Hurst exponent
H is investigated. We show that the dependence of the laminarity with respect to H is monotonic
in contrast to the percent determinism. We also show that the length of the diagonal and vertical
lines involved in the computation of percent determinism and laminarity has an influence on their
dependence on H. Statistical tests performed on the LAM measure support its utility to discriminate
fGn processes with respect to their H values. These results demonstrate that recurrence plots are suit-
able for the extraction of quantitative information on the correlation structure of these widespread
stochastic processes. Published by AIP Publishing. https://doi.org/10.1063/1.5030522

Discrete fractional Gaussian noises are ubiquitous stochas-

DET is related to the predictability of the underlying process
tic processes depending on a single parameter, the Hurst
exponent H, which entirely describes their time cor- generating the analyzed data.4
relations. White noise is recovered for H= 0.5. The Some years later, Marwan et al.6 proposed new measures
variation with H of two remarkable features of their derived from the vertical structures of RPs. In particular, a
recurrence plots (the determinism and the laminarity) measure inspired by DET was defined and called laminarity
is analyzed by confronting exact probabilistic results in LAM . The recurrence points forming these vertical lines are
infinite-size and finite-size simulation. This investigation related to the concept of sojourn points introduced by Gao.7,8
provides new results about recurrence quantification anal- It has been shown that such vertical structures can be related
ysis from diagonal and vertical lines, presents a method- to intermittency phenomena and useful for the detection of
ology yielding their analytical derivation for infinite-size chaos-chaos or chaos-order transitions.4,6
recurrence plots, and overall opens a novel research direc- Although RPs and RQA were first intended to investi-
tion for the estimation of the exponent H. gate deterministic chaotic dynamics, studies have shown that
they can also be used to explore data generated by stochastic
processes. More specifically, the measures derived from the
I. INTRODUCTION diagonal structures of the RPs have been shown to be related
to the time correlations of the considered random process (see
In 1987, Eckmann, Kamphorst, and Ruelle1 introduced Refs. 4 and 9 and references therein).
the concept of recurrence plots (RPs) to analyze data produced In the present work, the statistical properties of RPs of
by nonlinear dynamical systems. Few years later, recurrence data generated by discrete fractional Gaussian noise (fGn) are
quantification analysis (RQA) was proposed by Zbilut and investigated. This stochastic process was introduced in the
Webber2,3 to extract quantitative information from recurrence seminal work of Mandelbrot and van Ness as the increment
plots (RPs). Since then, many works have shown that several process of fractional Brownian motion.10 Such processes have
measures can be derived from RPs4,5 which provide infor- been widely used to model random time series in several fields
mation on the predictability, the stationarity, the cyclicites, such as biology, physics, or finance.11 They are characterized
or the laminar nature of the underlying dynamics or process by the so-called Hurst exponent H, which is the parameter
generating the considered data. specifying the autocovariance of the process.
After the recurrence rate (REC), the measures introduced Extending a methodology introduced in a previous
next were related to diagonal structures that are parallel to paper,9 we compute the measures REC, DET, and LAM of
the main diagonal of the RP. Among these measures, the per- fGn processes RPs constructed without embedding (i.e., with
cent determinism (DET) is certainly the best acknowledged. an embedding dimension equal to 1). The effect on these
measures of varying the H exponent is explored. The results
a)
sofiane.ramdani@umontpellier.fr show that, unlike the DET measure, the laminarity LAM is

1054-1500/2018/28(8)/085721/10/$30.00 28, 085721-1 Published by AIP Publishing.

085721-2 Ramdani, Bouchara, and Lesne Chaos 28, 085721 (2018)

a monotonic increasing function of H. These findings sug- A classical approach for the simulation of fGn’s sample
gest that the vertical structures of RPs of processes that can paths is based on the Cholesky decomposition of the covari-
be modeled by fGn can be used to characterize their time ance matrix , whose (i, j) entry can be written ij = γ (i − j)
correlations. due to covariance stationarity. This procedure is quite straight-
forward. Let L be the lower triangular matrix related to the
II. DEFINITIONS OF DET AND LAM MEASURES symmetric positive definite matrix  according to the equal-
ity  = LLT . Denoting η a N-dimensional column vector of
In this section, the RP construction and the definitions of white Gaussian noise samples, a time series with the desired
DET and LAM indices are briefly recalled. fGn properties is obtained by computing x = Lη.11
Let (xi ) be a time series. The first step of the RP construc- Figure 1 shows the RPs of two sample paths of fGn
tion is to perform the time delay embedding procedure12–14 to processes for two different values of the parameter H.
define the time delay vector xi of dimension d and using a
time delay τ :
IV. THEORETICAL ANALYSIS
xi = [xi , xi+τ , xi+2τ , . . . , xi+(d−1)τ ]T . (1)
Then using some norm (Euclidean or maximum norm in gen- In this section, the methodology used to compute the the-
eral), all the distances between xi and xj are compared to a oretical values of the percent determinism and laminarity of
given threshold (or radius) ε. The RP is simply the repre- time series generated by discrete fGns is described. These the-
sentation of the so-called recurrence points corresponding to oretical counterparts are denoted DETth and LAMth . We should
locations (i, j) for which the distance between xi and xj is less emphasize at this point that the term theoretical in this con-
than ε. No dot is represented otherwise. text indicates that the RQA measures correspond to theoretical
The first measure currently computed is the recurrence infinite RPs. Thus, one can refer to these theoretical quantities
rate REC, which is simply the rate of recurrence points. The as asymptotic. To achieve these computations, the derivation
second measure is the rate of recurrence points forming lines of the probabilities of the occurrence of a diagonal or a vertical
of length at least n that are parallel to the main diagonal line is necessary.
(defined by i = j). This so-called percent determinism DET The methodology is based on a previous work for which
is given by4 only diagonal lines were considered.9 It should be underlined
that this approach is not specific to fGns but is generic for the
Nv
kJk (ε) analysis of the RP measures of any discrete-time stationary
DET(ε, n) = N
k=n
, (2) Gaussian stochastic process.
v Thus, in this section, we consider that x is a real-
kJk (ε)
k=1
valued, discrete time, wide-sense stationary, centered Gaus-
sian stochastic process, with unit variance. For the sake of
where Jk (ε) is the number of diagonal lines of length exactly simplicity, the random variables and their realizations are
k and Nv is the size of the RP. denoted with the same symbols.
Similarly, one can define the rate of recurrence points For the rest of the analysis, we will consider the case of
forming vertical lines of length at least n, namely, the lami- RPs constructed without embedding. In this case, the vector
narity LAM given by4,6 xi reduces to the scalar xi [see Eq. (1)].

Nv In Secs. IV A–IV G, we will present the computations of
kVk (ε) the theoretical probabilities:
k=n
LAM (ε, n) = , (3)
Nv
• Pi,j : the probability of the occurrence of a recurrence point
kVk (ε)
k=1 at location (i, j) of the RP.
n
where Vk (ε) is the number of vertical lines of length • Pi,j : the probability of the occurrence of a diagonal of length
exactly k. n, starting from point (i, j).
• Qni,j : the probability of the occurrence of a diagonal of
III. FRACTIONAL GAUSSIAN NOISE length exactly n, starting from point (i, j).
• Ti,jn : the probability of the occurrence of a vertical of length
Fractional Gaussian noise (fGn) is defined as the incre- n, starting from point (i, j).
ment process of fractional Brownian motion (fBm) (see Refs. • Ui,jn : the probability of the occurrence of a vertical of length
10 and 11). It is a centered, stationary, and Gaussian process. exactly n, starting from point (i, j).
In the discrete-time case, the autocovariance function γ is
given by The use of an embedding dimension equal to 1 is motivated
n
by technical aspects of the computation of probabilities Pi,j
σ2   n
γ (k) = |k + 1|2H − 2|k|2H + |k − 1|2H (4) and Ti,j . As it can be seen in Secs. IV B and IV E below,
2 considering an embedding dimension d > 1 introduces an
for k ∈ Z and where σ 2 is the variance of the process (we con- overlapping of the time delay vectors involved in the compu-
sider here σ 2 = 1) and H is the Hurst exponent (0 < H < 1, tation of these probabilities, which will result in a much more
H = 0.5 corresponding to white noise). Note that complex formalization of the inequalities defining diagonal or
γ (−k) = γ (k). vertical lines. In addition, taking d = 1 automatically discard
085721-3 Ramdani, Bouchara, and Lesne Chaos 28, 085721 (2018)

FIG. 1. Recurrence plots [(a),(c)] of unit-variance fGn sample paths [(b),(d)] with H = 0.20 and H = 0.80, respectively. The RPs were constructed without
embedding and using a threshold ε = 0.5.

the issue of the selection of a time delay, which is not straight- If yni,j is an n-dimensional random vector given by
forward for data generated by fGn processes. This point will
be discussed in Sec. VI. yni,j = (xi − xj , xi+1 − xj+1 , . . . , xi+n−1 − xj+n−1 )T . (6)
By definition, DETth and LAMth can be expressed as func- Thus, Pi,j n
(ε) is the probability to have yni,j ∞  ε, where
tions of Qni,j and Ui,jn . In the following, we will show that the  · ∞ is the maximum norm.
probabilities Qni,j and Ui,jn can be derived from the probabilities Note that yni,j is the difference of two Gaussian random
n
Pi,j and Ti,jn . Note also that in order to simplify the notations of vectors of dimension n, respectively, constituted of the ran-
these probabilities, their dependence on the threshold ε is not dom variables xi et xj and their (n − 1) following variables.
explicitly mentioned. Consequently, the random vector yni,j is also Gaussian as a vec-
tor composed of differences of joint normal components of the
A. Computation of the probability Pi ,j
process x.15,16
In the case of an embedding dimension d = 1, the com- Let i,j be the covariance matrix of the random vector
n
putation of the probability of the occurrence of a recurrence yi,j . The components of this matrix are defined by
point at location (i, j) of the RP was addressed in a previ- i,j
ous work.9 For a stochastic process x, with variance σ 2 and k,l = (xi+k−1 − xj+k−1 )(xi+l−1 − xj+l−1 ), (7)
autocovariance function γ , satisfying the above mentioned
for (k, l) ∈ {1, 2, . . . , n}2 , where the brackets · denote expec-
conditions, it can be written
  tation.
ε Consequently, the probability density function (pdf) asso-
Pi,j = erf  , (5)
2αi,j ciated with yni,j is given by the following multivariate Gaussian
 function:
with αi,j = 2 σ 2 − γ (i − j) for i = j. In the case i = j, we
1 1 T  i,j −1
simply get Pi,i = 1. For |i − j| large enough, the probability fi,jn (y) = exp − y  y , (8)
Pi,j is independent of (i, j) and it provides an approximation of (2π )n/2 | |1/2
i,j 2
the theoretical recurrence rate RECth .9 Thus, for |i − j| large where |i,j | is the determinant of the matrix i,j .
enough, Pi,j can be simply denoted P. The evolution of Pi,j as The probability Pi,j n
to have a diagonal of n consecutive
a function of |i − j|, shown in Sec. V, provides a quantitative recurrence points starting from a point (i, j) on the RP reads
support of this assumption.
n
Pi,j = fi,jn (y)dy, (9)
B. Computation of the probability Pin,j M(ε)

To compute Pi,j n
, a method developed in Ref. 9 for the with M(ε) = {y : y ∈ Rn , y∞  ε}.
general case of a stationary Gaussian random process x is In the generic case, this probability can not be calculated
applied. On a recurrence plot constructed without embedding, analytically, but it can be estimated numerically. To achieve
n
Pi,j is the probability of the occurrence of n joint events given this, the method proposed by Genz17 is used. A description of
by |xi − xj |  ε, |xi+1 − xj+1 |  ε,. . . ,|xi+n−1 − xj+n−1 |  ε. this method can be found in the Appendix of Ref. 9.
085721-4 Ramdani, Bouchara, and Lesne Chaos 28, 085721 (2018)

n
It can be numerically shown that Pi,j is approximately which leads to
independent of (i, j) for |i − j|  1, i.e., for points (i, j) that
kQk nPn − 2nPn+1 + nPn+2
are not close to the main diagonal of the RP. This can be justi-
k n
fied by the stationarity of the analyzed random process.9 This
will be illustrated in Sec. V and the dependence of Pi,j n
with + (n + 1)Pn+1 − (2n + 2)Pn+2 + (n + 1)Pn+3
respect to |i − j| will be shown for different values of H.
+ (n + 2)Pn+2 − (2n + 4)Pn+3 + (n + 2)Pn+4

C. Computation of the probability Qin,j + (n + 3)Pn+3 − (2n + 6)Pn+4 + (n + 3)Pn+5

The computation of Qni,j can be achieved in the general ··· .

case and without any assumption on the nature of the pdf of (15)
the stochastic process. It can be expressed as follows: Through this last expression, one can observe that the inter-
mediary terms Pn+2 , Pn+3 , Pn+4 , . . . cancel out. Thus, an
Qni,j = (Pi,j
n
− Pi,j
n+1
) − (Pi−1,j−1
n+1
− Pi−1,j−1
n+2
), (10) approximation of the numerator of DETth can be written

where Pi,jn
is given by Eq. (9). kQk nPn − (n − 1)Pn+1 . (16)
k n
In order to compute the theoretical value DETth , we
n
also use the fact that Pi,j is approximately independent of Consequently, for the denominator of DETth [see Eq. (12)],
(i, j) for |i − j|  1. Thus, the notations Pi,j
n
and Qni,j can be, we get
n n
respectively, replaced with P and Q . Equation (10) then
leads to kQk P1 . (17)
k 1
Qn Pn − 2Pn+1 + Pn+2 , (11)
Note that P1 is, in fact, equal to the probability P, that
is, the theoretical recurrence rate [see Eq. (5)]. Finally, an
where Pn , Pn+1 , and Pn+2 are estimated using Eq. (9) for
approximation of DETth is given by
|i − j|  1.
nPn − (n − 1)Pn+1
DETth (ε, n) . (18)
D. Computation of the theoretical percent determinism P1
DETth
E. Computation of the probability Tin,j
After a normalization of the numerator and denominator
in the definition, Eq. (2), of the empirical DET, the following The probability of the occurrence of a vertical of length n,
expression for its theoretical (i.e., asymptotic) counterpart is starting from point (i, j) corresponds to the joint events
obtained: defined by |xi − xj |  ε, |xi − xj+1 |  ε,. . . ,|xi − xj+n−1 |  ε.
 k Similar to the case of diagonal lines, we introduce a new
kQ n-dimensional random vector zni,j defined by
k n
DETth (ε, n) =  . (12)
kQk zni,j = (xi − xj , xi − xj+1 , . . . , xi − xj+n−1 )T . (19)
k 1
Consequently, Ti,jn (ε) is the probability to have zni,j ∞  ε.
For the numerical computation of DETth , a reasonable approx- Denoting i,j the covariance matrix of the Gaussian
imation can be obtained by considering only the first terms of random vector zni,j , we have
the involved sums. This is a consequence of the fast vanishing
i,j
of probability Qk with k, especially for small ε values. A dif- k,l = (xi − xj+k−1 )(xi − xj+l−1 ), (20)
ferent and refined approximation can be derived by replacing
Qk with the expression given by Eq. (11). According to this for (k, l) ∈ {1, 2, . . . , n}2 . The pdf associated with zni,j is also
result, the numerator of DETth can be written given by a multivariate Gaussian function
1 1  −1
kQk k(Pk − 2Pk+1 + Pk+2 ). (13)
n
gi,j (z) = exp − zT i,j z . (21)
(2π )n/2 |i,j |1/2 2
k n k n
Finally, the probability Ti,jn to observe a vertical line of n recur-
If we detail the first terms of the involved sum, we get rence points starting from a point (i, j) on the RP can be
written
kQk n(Pn − 2Pn+1 + Pn+2 )
k n Ti,jn = n
gi,j (z)dz, (22)
D(ε)
+ (n + 1)(Pn+1 − 2Pn+2 + Pn+3 )
with D(ε) = {z : z ∈ Rn , z∞  ε}.
(14)
+ (n + 2)(P n+2
− 2P n+3
+P n+4
) As done for Pi,jn
, we exploit the fact that Ti,jn is approx-
imately independent of (i, j) for |i − j|  1. This will be
+ (n + 3)(Pn+3 − 2Pn+4 + Pn+5 )
numerically validated in Sec. V where the dependence of Ti,jn
··· , as function of |i − j| is depicted for different values of H.
085721-5 Ramdani, Bouchara, and Lesne Chaos 28, 085721 (2018)

F. Computation of the probability Uin,j matrix of the process. Alternatives are based on the so-
called Lowen’s circulant method,19 truncated symmetric mov-
The computation of the probability Ui,jn of the occurrence
ing average filters,20 or wavelet-based synthesis.21 For our
of a vertical of length exactly n, starting from point (i, j), can
study, we performed simulations using these three approaches
be performed thanks to the reasoning used for the case of
and obtained very similar qualitative results. Thus, we only
diagonals. Thus, we have
present here the outcomes of the simulations based on the
Ui,jn = (Ti,jn − Ti,jn+1 ) − (Ti−1,j−1
n+1
− Ti−1,j−1
n+2
). (23) Cholesky decomposition of the covariance matrix.
According to Eqs. (18) and (29), the computation of
Once again, the fact that Ti,jn is approximately independent of
DETth and LAMth relies, respectively, on the computation of
(i, j), for |i − j|  1, is used. Replacing the notation Ti,jn with
the probabilities Pn and T n . These probabilities are estimated
T n and Ui,jn with U n , we then get from Eq. (23)
thanks to Eqs. (9) and (22) for |i − j|  1 and are obtained
Un T n − 2T n+1 + T n+2 , (24) through the integration of the multivariate Gaussian functions
defined by Eqs. (11) and (29). This can be achieved using the
n n+1 n+2
where T , T , and T are computed using Eq. (22) for approach proposed by Genz17 and detailed in the Appendix of
|i − j|  1. Ref. 9.
The covariance matrices involved in the functions defined
G. Computation of the theoretical laminarity LAMth by Eqs. (8) and (21) were, respectively, denoted by i,j
From Eq. (3) defining the empirical LAM , its theoretical and i,j . According to Eqs. (7) and (20), these matrices
(i.e., asymptotic) counterpart can be defined by are composed of terms that can be expressed through the
 autocovariance function γ (defined in Sec. III) as follows:
kU k
k n i,j
LAMth (ε, n) =  . (25) k,l = 2γ (k − l) − γ (i − j + k − l) − γ (j − i + k − l),
kU k (30)
k 1

As for the theoretical percent determinism, this expression can i,j

be simplified. Indeed, by replacing U n with the expression k,l = 1 − γ [i − (j + l − 1)] − γ [(j + k − 1) − i] + γ (k − l),
given by Eq. (24) involving T n , the following approximation (31)
for the denominator of LAMth is obtained:
for (k, l) ∈ {1, 2, . . . , n}2 .
kU k k(T k − 2T k+1 + T k+2 ). (26) For the computations of RECth , DETth , and LAMth , we set
k n k n |i − j| = 500 and ε = 0.5. This value of the threshold radius
After some algebra detailed in Eqs. (14) and (15) for the case ε ensured large enough probabilities of occurrence of diago-
of diagonal lines and for the computation of DETth , we get nal and vertical lines in almost all numerically explored cases
(values of H and n). Note that we consider zero-mean and
kU k nT n − (n − 1)T n+1 . (27) unit variance processes. This implies that the selected relative
k n radius corresponds to 50% of the standard deviation of the
Similarly, the denominator of LAMth is given by theoretical process. This value was also selected according to
our previously published results regarding a sensitivity analy-
kU k T 1. (28) sis with respect to the threshold value.9 Indeed, for 1000-point
k 1 sample paths, the findings support that such value ensures reli-
The approximation of LAMth then can be written able estimations of RQA measures, i.e., consistent with their
theoretical values and with low variability.
nT n − (n − 1)T n+1
LAMth (ε, n) . (29)
T1

V. NUMERICAL RESULTS
In this section, we present the results of the computation
n
of the theoretical probabilities Pi,j , Pi,j , and Ti,jn as a function
of |i − j| for different values of the parameter H and for n = 2
and n = 4. Then, we show the results of the numerical compu-
tations of the theoretical values RECth , DETth , and LAMth for
a range of H values (from 0.05 to 0.95) and compare them to
the empirical REC, DET, and LAM values obtained for simu-
lated paths of fGn processes. In the case of DETth and LAMth ,
we also explore the effect of the diagonal and vertical minimal
length n on the obtained values.
The simulation of discrete fGn time series can be per-
FIG. 2. Probability Pi,j as a function of |i − j| obtained for H values ranging
formed through different methods.11,18 A classical approach from 0.1 to 0.9, with a step of 0.1. The curves for H  0.5 are almost identical
relies on the Cholesky decomposition of the covariance for largest values of |i − j|.
085721-6 Ramdani, Bouchara, and Lesne Chaos 28, 085721 (2018)

Figure 2 shows the values of the probability Pi,j (which value of each measure (over thirty simulated sample paths)
provides an estimation of RECth ) for |i − j| ranging from 0 to is shown and the error bars correspond to the standard devia-
500 for different values of the parameter H. Figures 3 and 4 tions. This figure also shows these values for different values
n
show similar results for the probabilities Pi,j and Ti,jn , respec- of n (2, 3 and 4), namely, the minimal length of the diagonal
tively, involved in the computation of DETth and LAMth , for and vertical lines, respectively, considered for the computa-
n = 2 and n = 4. The results depicted in these figures con- tion of DET and LAM measures. As expected, the DET and
firm the numerically observed asymptotic independence of the LAM values decrease with n. Note that, in Fig. 4, the scale for
theoretical probabilities of (i, j) for |i − j|  500. Note that as the REC values is reduced in comparison to the scales used to
expected and unlike Ti,jn , the probability Pi,jn
is equal to 1 for depict DET and LAM values.
|i − j| = 0, which corresponds to the main diagonal of the RP. In order to get a deeper insight into the potential discrimi-
For the estimation of the empirical values of REC, DET, native power of empirical LAM measure, which is monotonic
and LAM , we set the embedding dimension d = 1 and the with respect to H, we also performed statistical tests. We used
same threshold ε = 0.5. We excluded the main diagonal of a pairwise two-tailed Student’s test for independent samples
the RPs for these estimations (for Figs. 1 and 6, displaying to compare, for all possible pairs of H values (19 values rang-
RPs, the main diagonal was not discarded). For each value ing from 0.05 to 0.95), the sample means of LAM values
of H, these computations were performed for simulated time obtained from 30 simulated time series. The results for three
series of 1000-point length. Thirty realizations of each process different values of n, namely, 2, 3, and 4, are presented in
were simulated in order to get statistics (mean and standard panels (a), (b), and (c) of Fig. 5, in a graphical matrix form
deviation) of the REC, DET, and LAM values. We performed using a 2-D binary plot where the white squares represent the
the comparison between the theoretical and empirical RQA cases for which the null hypothesis of equality of means was
(DET and LAM ) measures for three values of the diagonal rejected. A black square is plotted when the null hypothesis
and vertical lines length, namely, n = 2, 3, 4. was not rejected. The significance level α value was set to 0.05
The empirical RQA measures were obtained by means of for these tests. The use of the Student test was supported by
the Cross Recurrence Plot Toolbox developed by N. Marwan6 checking the normality of each sample of 30 empirical LAM
(available at http://www.agnld.uni-potsdam.de/∼marwan/ values using the Kolmogorov-Smirnov test: for all H and
toolbox.php). n values, we found that the null hypothesis of a normal under-
In Fig. 4, the empirical and theoretical values of REC, lying distribution was not rejected. In panels (d), (e), and (f)
DET, and LAM are depicted for H values ranging from 0.05 of Fig. 5, we present the power of these statistical tests using
to 0.95 with a step of 0.05. For the empirical values, the mean heatmap plots, for each value of n.

n
FIG. 3. Probabilities Pi,j [panels (a) and (c)] and Ti,jn [panels (b) and (d)] as functions of |i − j| for H values ranging from 0.1 to 0.9, with n = 2 [(a) and (b)] and
n = 4 [(c) and (d)]. The value of probability Ti,jn shown in panels (b) and (d) increases with H.
085721-7 Ramdani, Bouchara, and Lesne Chaos 28, 085721 (2018)

0.4

FIG. 4. Empirical and theoretical REC (a), DET (b), and LAM (c) measures obtained for H values ranging from 0.05 to 0.95. In the case of DET and LAM , the
values obtained for n = 2, n = 3, and n = 4 are reported. The top-most curve corresponds to n = 2 and the bottom-most to n = 4. Note that the scale used to
depict REC values is reduced in comparison to the scales of DET and LAM .

VI. DISCUSSION The results shown in Figs. 2 and 3 indicate that the setting of
|i − j| = 500 for the computation of RECth , DETth , and LAMth
The computation of the theoretical RQA measures RECth , ensures a very good approximation in almost all cases.
DETth , and LAMth is directly related to the computation of the We also observe from Fig. 3 that the probability related
n
probabilities Pi,j , Pi,j , and Ti,jn [see Eqs. (5), (18), and (29)]. to the vertical lines Ti,jn [shown in panels (b) and (d)] is more

FIG. 5. 2-D binary plots [(a)–(c)] and heatmap plots [(d)–(f)], respectively, displaying t-test results and their corresponding power for n = 2, 3, and 4 (from left
to right). These plots were obtained by comparing empirical LAM means for all possible pairs of H values ranging from 0.05 to 0.95. Note that the diagonal
black squares result from the comparison of two identical samples of LAM values and thus they do not convey a statistical meaning.
085721-8 Ramdani, Bouchara, and Lesne Chaos 28, 085721 (2018)

n
relevant than Pi,j [depicted in panels (a) and (c), in which a be used to distinguish different values of H (with a 0.05 step
zoom is added] to distinguish H values lower than 0.5 to those resolution) when H  0.5 [see Figs. 5(a) and 5(d)]. For n = 3
larger than 0.5 (H = 0.5 corresponds to the case of white and n = 4, this would require higher values of H, namely,
noise). This can be confirmed after an adjustment of the scale H > 0.6 [see Figs. 5(b), 5(c), 5(e), and 5(f)]. As for a pos-
n
of Pi,j to match the Ti,jn scale. sible limitation coming from data length, and specifically for
The results displayed in Fig. 4 show that the theoretical the results of Fig. 4, simulations performed with 200-point
and empirical values of the three considered RQA measures fGn sample paths provided consistent findings (not shown)
are very coherent. For the REC measure [Fig. 4, panel (a)], with those presented here using 1000-point length time series.
we observe a moderate overestimation for H  0.85. We also The main differences were a moderately larger overestima-
observe a very stable value of the theoretical value RECth for tion of the REC measure for values of H larger than 0.80
H  0.80. For larger values of H, low-frequency components and an expected increase of the variability of the empirical
(local trends) appear in the time series, inducing modifications estimations. Concerning the results of the statistical tests, as
on the expected statistical properties, e.g., mean and standard expected, the discriminative power of the LAM measure was
deviation, of the simulated sample paths. This could explain weakened so that the significant statistical differences were
the overestimation of the REC values of the simulated signals only found for H values larger than 0.75.
with respect to their theoretical counterparts. This probable In order to confirm our results regarding the potential
explanation should be confirmed in the future by a dedicated effect of the quality of the simulated fGn paths, we also inves-
study. Note that the local trends induce white bands and less tigated the RQA measures as functions of the a posteriori
homogeneous RPs, as apparent in the two RPs displayed in estimated H values instead of using the preset theoretical H
Fig. 6, generated by two fGn sample paths, with H = 0.7 and values. To estimate these values of H, we used two classi-
H = 0.9 (see also Ref. 4 and references therein regarding non- cal methods, namely, detrended fluctuation analysis (DFA)
stationarity issues). and an approach using a wavelet-based discrete second-order
We observe that the variability of the empirical DET derivative estimator (see, for instance, Refs. 22–26). We thus
(quantified by the standard deviation) is, in general, lower than obtained H for each 1000-point simulated path and then aver-
the variability obtained for the LAM measure [see Fig. 4, pan- aged over the 30 realizations to get a unique statistic. The
els (b) and (c)]. We also note that the variability of DET is dependence of the three considered RQA measures on these
decreased when the number of the minimal length n of the estimated H values was similar and consistent with the one
considered diagonals is increased. based on the predefined H values. Actually, the estimated H
Figure 4 shows that unlike DETth , LAMth is monotonic. values are only slightly shifted with respect to the predefined
The DET measure reaches a minimal value for H = 0.5, H values. The results of the second estimation method are
which corresponds to white Gaussian noise. This is expected displayed in Fig. 7.
as DET is positively correlated to the predictability of the Few studies have addressed the theoretical derivation
process. The statistical results displayed in Fig. 5 suggest of RQA measures for large classes of stochastic processes.
that the LAM measure computed for n = 2 can potentially These studies had different objectives and were based on

FIG. 6. Recurrence plots [(a),(c)] of unit variance fGn sample paths [(b),(d)] with H = 0.70 and H = 0.90, respectively. The RPs were constructed without
embedding and using a threshold ε = 0.5.
085721-9 Ramdani, Bouchara, and Lesne Chaos 28, 085721 (2018)

FIG. 7. Empirical and theoretical REC (a), DET (b), and LAM (c) measures obtained for H values ranging from 0.05 to 0.95. Values obtained for n = 2, 3, and
4 are reported for DET and LAM measures. The top-most curve corresponds to n = 2 and the bottom-most to n = 4. For the empirical estimations of the three
measures, the corresponding H values shown on x-axis were computed a posteriori using a wavelet-based discrete second-order derivative method, explaining
the slight horizontal shift between the points of the blue and red curves.

completely different methodological approaches than ours. ACKNOWLEDGMENTS

They were not specific to fGn processes. In Ref. 27, a
mathematical and statistical analysis of some specific RQA S.R. would like to thank V. Kleptsyn for useful
measures (the k-recurrence rate, the percent determinism, and discussions concerning probabilistic computations. The
the average length of diagonal lines) was performed. The authors would like to thank two anonymous referees for their
authors exploited correlation sums to analytically express the constructive criticisms and insightful comments.
asymptotic values of these measures and applied them to 1
J. Eckmann, S. Kamphorst, and D. Ruelle, “Recurrence plots of dynamical
i.i.d. processes, Markov chains, and autoregressive processes. systems,” Europhys. Lett. 4, 973–977 (1987).
In another work, Schultz et al.28 derived approximations 2
J. Zbilut and C. Webber, “Embeddings and delays as derived from quan-
of diagonal line based RQA-measures including the percent tification of recurrence plots,” Phys. Lett. A 171, 199–203 (1992).
3
C. Webber and J. Zbilut, “Dynamical assessment of physiological systems
determinism. The results were expressed in terms of pair-
and states using recurrence plot strategies,” J. Appl. Physiol. 76, 965–973
wise proximity measures, which basically count the number of (1994).
pairs of neighboring states in the reconstructed phase space. In 4
N. Marwan, M. Romano, M. Thiel, and J. Kurths, “Recurrence plots for
the same line, Spiegel et al.29 extended this result to approxi- 5
the analysis of complex systems,” Phys. Rep. 438, 237–329 (2007).
C. Webber and N. Marwan, Recurrence Quantification Analysis (Springer
mate the laminarity using a measure of stationary states of the International Publishing, 2015).
embedded trajectory in a phase space with distances defined 6
N. Marwan, N. Wessel, U. Meyerfeldt, A. Schirdewan, and J. Kurths,
by maximum norm. In Ref. 30, fGn processes were investi- “Recurrence plot based measures of complexity and its application to heart
gated through recurrence networks. The authors empirically rate variability data,” Phys. Rev. E 66, 026702 (2002).
7
J. B. Gao, “Recurrence time statistics for chaotic systems and their
showed that these networks can be useful to analyze such applications,” Phys. Rev. Lett. 83, 3178–3181 (1999).
processes provided that the embedding dimension and delay 8
J. Gao and H. Cai, “On the structures and quantification of recurrence
are properly selected. They also emphasized that selecting plots,” Phys. Lett. A 270, 75–87 (2000).
9
the embedding dimension is not straightforward for fGn pro- S. Ramdani, F. Bouchara, J. Lagarde, and A. Lesne, “Recurrence plots
of discrete-time gaussian stochastic processes,” Physica D 330, 17–31
cesses. One advantage of our approach is that the embedding (2016).
dimension is set to 1 and thus no time delay is required to 10
B. Mandelbrot and J. van Ness, “Fractional brownian motions fractional
construct the RPs. noises and applications,” SIAM Rev. 10, 422–437 (1968).
11
Overall, our results establish an analytical relationship J. Beran, Statistics for Long-Memory Processes (Chapman & Hall, 1994).
12
N. Packard, J. Crutchfield, J. Farmer, and R. Shaw, “Geometry from a time
between the main RQA measures and the covariance struc- series,” Phys. Rev. Lett. 45, 712–716 (1980).
ture of a wide class of stationary Gaussian processes. These 13
F. Takens, “Detecting strange attractors in turbulence,” in Dynamical
findings demonstrate that measures extracted from RPs can Systems and Turbulence, Warwick 1980 (Springer, 1981), pp. 712–716.
14
H. Kantz and T. Schreiber, Nonlinear Time Series Analysis, 2nd ed.
be exploited to identify and distinguish fGn processes. The
(Cambridge University Press, 2004).
theoretical values of specific RQA measures can be used to 15
A. Papoulis and S. Pillai, Probability, Random Variables and Stochastic
control the quality of the estimation of the H exponent. Processes, 4th ed. (McGraw-Hill Higher Education, 2002).
085721-10 Ramdani, Bouchara, and Lesne Chaos 28, 085721 (2018)

16 24
C. Rasmussen and C. Williams, Gaussian Processes for Machine Learning D. Delignieres, S. Ramdani, L. Lemoine, K. Torre, M. Fortes, and G.
(MIT Press, 2006). Ninot, “Fractal analyses for ‘short’ time series: A re-assessment of classical
17
A. Genz, “Numerical computation of multivariate normal probabilities,” methods,” J. Math. Psychol. 50, 525–544 (2006).
25
J. Comput. Graph. Stat. 1, 141–149 (1992). P. Abry, P. Flandrin, M. Taqqu, and D. Veitch, “t,” in Theory and Applica-
18
J.-M. Bardet, G. Lang, G. Oppenheim, A. Philippe, S. Stoev, and M. Taqqu, tions of Long-Range Dependence (Birkhäuser, 2003), pp. 527–556.
26
“Generators of long-range dependence processes: A survey,” in Theory and J.-M. Bardet, G. Lang, G. Oppenheim, A. Philippe, S. Stoev, and
Applications of Long-Range Dependence (Birkhäuser, 2003), pp. 579–623. M. Taqqu, “Semi-parametric estimation of the long-range dependence
19
S. Lowen, “Efficient generation of fractional brownian motion for simu- parameter: A survey,” in Theory and Applications of Long-Range Depen-
lation of infrared focal-plane array calibration drift,” Methodol. Comput. dence (Birkhäuser, 2003), pp. 557–577.
27
Appl. 1, 445–456 (1999). M. Grendár, J. Majerová, and V. Špitalský, “Strong laws for recurrence
20
S. Stoev and M. Taqqu, “Simulation methods for linear fractional sta- quantification analysis,” Int. J. Bifurcat. Chaos 23, 1350147 (2013).
28
ble motion and FARIMA using the fast Fourier transform,” Fractals 12, D. Schultz, S. Spiegel, N. Marwan, and S. Albayrak, “Approximation of
95–121 (2003). diagonal line based measures in recurrence quantification analysis,” Phys.
21
P. Abry and F. Sellan, “The wavelet-based synthesis for the fractional Lett. A 379, 997–1011 (2015).
29
brownian motion proposed by F. Sellan and Y. Meyer: Remarks and fast S. Spiegel, D. Schultz, and N. Marwan, “Approximate recurrence quan-
implementation,” Appl. Comput. Harmon. Anal. 3, 377–383 (1996). tification analysis (arqa) in code of best practice,” in Recurrence Plots
22
C. Peng, J. Mietus, J. Hausdorff, S. Havlin, H. Stanley, and A. Gold- and Their Quantifications: Expanding Horizons, Springer Proceedings in
berger, “Long-range anti-correlations and non-Gaussian behavior of the Physics Vol. 180, edited by C. Webber, C. Ioana, and N. Marwan (Springer,
heartbeat,” Phys. Rev. Lett. 70, 1343–1346 (1993). Cham, 2016), pp. 113–136.
23 30
J. Moreira, J. Kamphorst, L. da Silva, and S. Kamphorst, “On the Y. Zou, R. V. Donner, and J. Kurths, “Analyzing long-term correlated
fractal dimension of self-affine profiles,” J. Phys. A 27, 8079–8089 stochastic processes by means of recurrence networks: Potentials and
(1994). pitfalls,” Phys. Rev. E 91, 022926 (2015).