1

Journal of Mathematical Psychology (2006), 50, 525-544

Fractal analyses for ‘short’ time series: A re-assessment of

classical methods.
Didier Delignieres1, Sofiane Ramdani1, Loïc Lemoine1, Kjerstin Torre1,
Marina Fortes2 & Grégory Ninot3

1. EA 2991 - University Montpellier I, France

2. JE 2438 – University of Nantes, France
3. JE 2416 - University Montpellier I, France

Abstract: The aim of this study was to evaluate the in a bimanual coordination task (Schmidt, Beek,
performances of some classical methods of fractal Treffner, & Turvey, 1991), or for the displacement of
analysis with short time series. We simulated exact the center-of-pressure during upright stance (Collins
fractal series to test how well methods estimate the & De Luca, 1993; Delignières, Deschamps, Legros,
Hurst exponent. We successively tested power & Caillou, 2003). Most of these variables were
spectral density analysis, detrended fluctuation previously conceived as highly stable over time, and
analysis, rescaled range analysis, dispersional fluctuations in successive measurements were
analysis, maximum likelihood estimation, and two considered as the expression of random, uncorrelated
versions of scaled windowed variance methods. All errors. For example, Epstein (1979) considered self-
methods presented different advantages and esteem as a personality trait, a highly stable reference
disadvantages, in terms of biases and variability. We value, and assigned variations in repetitive
propose in conclusion a systematic step-by-step measurements to unmeaning noise. From this point of
procedure of analysis, based on the performances of view, a sample of repeated measures is assumed to be
each method and their appropriateness regarding the normally distributed around its mean value, and noise
scientific aims that could motivate fractal analysis. can be discarded by averaging. This methodological
___________________________________________ standpoint was implicitly adopted in most classical
psychological research (Gilden, 2001; Slifkin &
A number of psychological or behavioral Newell, 1998). From this point of view, the temporal
variables were recently proven to possess fractal ordering of data points is ignored and the possible
properties, when considered from the point of view of correlation structure of fluctuations is neglected.
their evolution in time. This was the case, for In contrast, fractal analysis focuses on the time-
example, for self-esteem (Delignières, Fortes & evolutionary properties of data series and on their
Ninot, 2004), for mood (Gottschalk, Bauer, & correlation structure. Fractal processes are
Whybrow, 1995), for serial reaction time (Gilden, characterized by a complex pattern of correlations
1997; van Orden, Holden, & Turvey, 2003), for the that appears following multiple interpenetrated time
time intervals produced in finger tapping (Gilden, scales. As such, the value at a particular time is
Thornton, & Mallon, 1995; Madison, 2004), for related not just to immediately preceding values, but
stride duration during walking (Hausdorff, Peng, also to fluctuations in the remote past. Fractal series
Ladin, Wei, & Goldberger, 1995), for relative phase are also characterized by self-similarity, signifying
that the statistical properties of segments within the
____________________________ series are similar, whatever the time scale of
Address for correspondence: observation.
Pr. D. Delignières, Evidencing fractal properties in empirical time
EA 2991 Motor Efficiency and Deficiency, series has important theoretical consequences, and
Faculty of Sports and Physical Education, leads to a deep renewal of models. Fractals are
University Montpellier I, considered as the natural outcome of complex
700, avenue du Pic Saint Loup, 34090 Montpellier, dynamical systems behaving at the frontier of chaos
France.
(Bak & Chen, 1991; Marks-Tarlow, 1999).
Tel: +33 (0)4 67 41 54 57.
Fax: +33 (0)4 67 41 57 08. Psychological variables should then be conceived as
Email: [email protected] the macroscopic and dynamical products of a
complex system composed of multiple interconnected
 2
elements. Moreover, psychological and behavioral which signifies that the expected squared
time series often present fractal characteristics close displacement is a power function of the time interval
to a very special case of fractal process, called 1/f or (∆t) over which it was observed. H represents the
pink noise.‘1/f noise’ signifies that when the power typical scaling exponent of the series and can be any
spectrum of these time series is considered, each real number in the range 0 < H < 1. The aims of
frequency has power proportional to its period of fractal analysis are to check whether this scaling law
oscillation. As such, power is distributed across the holds for experimental series and to estimate the
entire spectrum and not concentrated at a certain scaling exponent. Ordinary Brownian motion
portion. Consequently, fluctuations at one time scale corresponds to the special case H = 0.5 and
are only loosely correlated with those of another time constitutes the frontier between anti-persistent (H <
scale. This relative independence of the underlying 0.5) and persistent fBms (H > 0.5).
processes acting at different time scales suggests that Fractional Gaussian noise (fGn) represents
a localized perturbation at one time scale will not another family of fractal processes, defined as the
necessarily alter the stability of the global system. In series of successive increments in an fBm. Note that
other words, 1/f noise renders the system more stable fGn and fBm are interconvertible: when an fGn is
and more adaptive to internal and external cumulatively summed, the resultant series constitutes
perturbations (West & Shlesinger, 1990). In the an fBm. Each fBm is then related to a specific fGn,
aforementioned studies, 1/f noise was evidenced in and both are characterized by the same H exponent.
most series produced by “normal” participants, These two processes possess fundamentally different
characterized as young and healthy. As such, this 1/f properties: fBm is non-stationary with time-
behavior could be considered as an indicator of the dependent variance, while fGn is a stationary process
efficiency of the system that produced the series. In with a constant expected mean value and constant
contrast, series obtained with older participants or variance over time. Examples of fBm and fGn
with patients with specific pathologies exhibited corresponding to three values of H are presented in
specific alterations in fractality (Hausdorff et al., Figure 1. The H exponent can be assessed from an
1997; Yoshinaga, Miyazima, & Mitake, 2000). fBm series as well as from the corresponding fGn,
In order to ensure better understanding of the but because of the different properties of these
following parts of this article, a deeper and more processes, the methods of estimation are necessarily
theoretical presentation of fractal processes seems different.
necessary. A good starting point for this presentation Recently a systematic evaluation of fractal
is Brownian motion, a well-known stochastic process analysis methods was undertaken by
that can be represented as the random movement of a Bassingthwaighte, Eke, and collaborators (Caccia,
single particle along a straight line. Mathematically, Percival, Cannon, Raymond, & Bassingthwaigthe,
Brownian motion is the integral of a white Gaussian 1997; Cannon, Percival, Caccia, Raymond, &
noise. As such, the most important property of Bassingthwaighte, 1997; Eke et al., 2000; Eke,
Brownian motion is that its successive increments in Hermann, Kocsis, & Kozak, 2002). This
position are uncorrelated: each displacement is methodological effort was based on the previously
independent of the former, in direction as well as in described dichotomy between fGn and fBm.
amplitude. Einstein (1905) showed that, on average, According to these authors, the first step in a fractal
this kind of motion moves a particle from its origin analysis aims at identifying the class to which the
by a distance that is proportional to the square root of analyzed series belongs, i.e. fGn or fBm. Then the
time. scaling exponent can be properly assessed, using a
Mandelbrot and van Ness (1968) defined a method relevant for the identified class. The
family of processes they called fractional Brownian evaluation proposed by these authors clearly showed
motions (fBm). The main difference from ordinary that most methods gave acceptable estimates of H
Brownian motion is that in an fBm successive when applied to a given class (fGn or fBm), but led
increments are correlated. A positive correlation to inconsistent results for the other. As claimed by
signifies that an increasing trend in the past is likely Eke et al. (2002), researchers were not aware before a
to be followed by an increasing trend in the future. recent past of the necessity of this dichotomic model.
The series is said to be persistent. Conversely, a As such, a number of former empirical analyses and
negative correlation signifies that an increasing trend theoretical interpretations remain questionable.
in the past is likely to be followed by a decreasing Psychological and behavioral research,
trend. The series is then said to be anti-persistent. nevertheless, raises a number of specific problems
Mathematically, an fBm is characterized by the that were not clearly addressed by Bassingthwaighte,
following scaling law: Eke and collaborators. The first problem is related to
the length of the series that seems required for a
<∆x²> ∝ ∆t2H (1) proper application of fractal methods. Their results
showed that the accuracy of the estimation of fractal
exponents is directly related to the length of the
 3

Fractional Brownian motions

H = 0.25 H = 0.50 H = 0.75

Fractional Gaussian noises

H = 0.25 H = 0.50 H = 0.75

Figure 1: Graphical examples of fractal time series. The upper graphs represent fractional Brownian motions
(fBm) and the lower graphs, the corresponding fractional Gaussian noises (fGn), for three typical values of
the scaling exponent. The upper median graph shows an ordinary Brownian motion (H=0.5) with its
differenced series (white noise) just below. The right and left columns show respectively an anti-persistent
fBm (H=0.25) and a persistent fBm (H=0.75) and their corresponding fGns.
series. One of their main conclusions is that fractal series. A more precise and systematic evaluation of
methods cannot give reliable results with series fractal methods for short time series seems clearly
shorter than 212 data points, and in their papers, necessary, for a reliable application in psychological
especially devoted to physiological research, they and behavioral research.
focused on results obtained with very long series (217 Secondly, the evaluations performed by
data points). Such series cannot be collected in Bassingthwaighte, Eke and collaborators were based
psychological research. The application of time series on a global index, combining bias (the deviation of
analyses supposes that the system under study the mean estimated H from the true H exponent) and
remains unchanged during the whole window of standard deviation (the variability of estimations
observation, and in psychological experiments, the obtained from series of identical true H exponents).
lengthening of the task raises evident problems of According to the true aim of a specific research, both
fatigue or lack of concentration (Madison, 2001). these basic characteristics have to be clearly
Generally, the use of series of 29 or 210 data points distinguished and separately assessed. When the
was considered as an acceptable compromise problem is to analyze differences between
between the requirements of time series analyses and experimental groups, one could suppose that a small
the limitations of psychological experiments (see, for standard deviation is essential, but a (limited and
example, Chen, Ding & Kelso, 1997, 2001; systematic) bias could remain acceptable. Practically,
Delignières, Fortes et al., 2004; Gilden , 1997, 2001; one could consider that a standard deviation of 0.1
Musha, Katsurai Teramachi, 1985; M. Yamada, represents the higher acceptable limit for such inter-
1996; M. Yamada & Yonera, 2001; N. Yamada, group comparisons. In a study allowing the collection
1995). It could be useful, nevertheless, to get precise of experimental series from two groups of about 15
information about the performance of fractal methods participants, it could be possible with such standard
with shorter series (i.e. 26, 27or 28 data points), which deviation to discriminate between mean exponents
could be easier to collect in psychological or separated by about 0.08.
behavioral studies. Some earlier papers conducted On the contrary, when the goal is to obtain an
investigations about such short series (e.g. Caccia et accurate determination of the exponent that
al., 1997; Cannon et al., 1997; Pilgram & Kaplan, characterizes the system under study, bias should be
1998), but these papers were devoted to a limited set as limited as possible, but variability could be
of methods, and didn’t allow deriving a global counteracted by averaging a sample of independent
strategy for applying fractal analyses with such short assessments. Moreover, one could suppose that bias
 4
and variability, for a given method, could be different 1
according to the value of the true exponent of the Vk = S k (W2 k −1 + iW2 K ) for 1 ≤ k < M/2, (4b)
series under study. As such, prescriptions concerning 2
the relevancy of fractal methods should go beyond
the dichotomy fGn/fBm, and consider the theoretical VM / 2 = S M / 2 WM −1 ,
aims of the assessment, and the approximate location
for the empirical series in each class. Some earlier
papers provided a separate assessment of bias and Vk = VM −k * for M/2 <k ≤ M-1, (4d)
variability (e.g. Caccia et al., 1997; Cannon et al.,
1997; Pilgram & Kaplan, 1998). But as previously where * denotes that Vk and VM-k are complex
stated, these studies focused on specific methods, and conjugates.
didn’t allow supporting a global approach for fractal Finally, the first N elements of the discrete
analysis. Fourier transform of V are used to compute the
Finally, one could consider that experimental simulated series x(t):
series are systematically contaminated by random
fluctuations induced by response modes and/or by M −1
recording devices. The evaluation of methods for x(t) = 1
M
∑Ve
k =0
k
− i2πk((t −1) / M)
(5)
estimation of H rarely takes the effect of added noise
into account (see Cannon et al., 1997). This potential
effect should obviously be considered in the case of where t = 1, 2,…, N.
short series.
Power spectral density analysis (PSD)
Methods This method is widely used for assessing the
Generation of exact fractal series fractal properties of time series, and works on the
We used the algorithm proposed by Davies and basis of the periodogram obtained by the Fast Fourier
Harte (1987), for generating fGn series of length N Transform algorithm. The relation of Mandelbrot and
(N being a power of 2). The autocovariance function van Ness (1968) can be expressed as follows in the
γ(τ) of a fGn series is related to the scaling exponent frequency domain:
H according to the following equation (Mandelbrot &
van Ness, 1968): S(f) ∝ 1/f β (6)

σ² 2H 2H 2H
where f is the frequency and S(f) the
γ (τ ) = (τ + 1 − 2τ + τ −1 ), correspondent squared amplitude. β is estimated by
2 calculating the negative slope (-β) of the line relating
τ = 0,±1,±2,..., (2) log (S(f)) to log f. Obtaining a well-defined linear fit
in the log-log plot is an important indication of the
For j = 0, 1,…, N, the exact spectral power Sj presence of long-range correlation in the original
expected for this autocovariance function is series. According to Eke et al. (2000), PSD allows to
computed, from the discrete Fourier transform of the distinguish between fGn and fBm series, as fGn
following sequence of covariance values γ defined by corresponds to β exponents ranging from –1 to +1,
Eq. 2: γ0, γ1,…, γM/2-1, γM/2, γ(M/2)-1,…, γ1. and fBm to exponents from +1 to +3. β can be
converted into Ĥ according to the following
M /2 M −1 equations:
Sj = ∑ γ (τ )e
τ =0
−i 2πj (τ / M )
+ ∑ γ ( M − τ )e
τ = M / 2 +1
−i 2πj (τ / M )

Hˆ =
β +1
for fGn, (7a)
2
with i²=-1 (3)
or
β −1
It is important to check that Sj ≥ 0 for all j. Hˆ = for fBm. (7b)
Negativity would indicate that the sequence is not 2
valid. Note that in these equations and thereafter in the
Let Wk, where k is an element of {0, 1,…,M-1}, text, Ĥ represents the estimate provided by the
be a set of i.i.d. Gaussian random variables with zero analysis, and H the true exponent of the series.
mean and unit variance. The randomized spectral We also used the improved version of PSD
amplitudes, Vk, are calculated according to the proposed by Fougère (1985) and modified by Eke et
following equations: al. (2000). This method uses a combination of
preprocessing operations: First the mean of the series
is subtracted from each value, and then a parabolic
V0 = S 0 W0 , (4a) window is applied: each value in the series is
multiplied by the following function:
 5
2j Rescaled Range Analysis (R/S)
W ( j) = 1 − ( − 1)² for j = 1, 2, …, N. (8) This method was originally developed by Hurst
N +1
(1965). The x(t) series is divided into non-
overlapping intervals of length n. Within each
Thirdly a bridge detrending is performed by
interval, an integrated series X(t, n) is computed:
subtracting from the data the line connecting the first
t
and last point of the series. Finally the fitting of β
excludes the high-frequency power estimates (f > 1/8
X (t , n ) = ∑ [ x ( k ) − x ] , (13)
k =1
of maximal frequency). This method was proven by
Eke et al. (2000) to provide more reliable estimates where x is the average within each interval. In
of the spectral index β, and was designated as the classical version of R/S analysis, the range R is
low
PSDwe. computed for each interval, as the difference between
the maximum and the minimum integrated data X(t,
Detrended fluctuation analysis (DFA) n).
This method was initially proposed by Peng et al.
(1993). The x(t) series is integrated, by computing for R = max X(t, n) – min X(t, n) (14)
1≤t≤n 1≤t≤n
each t the accumulated departure from the mean of
the whole series: We used in the present paper an improved
k
version, R/S-detrended (Caccia et al., 1997), where a
X(k) = ∑ [ x(i) − x ]
i =1
(9) straight line connecting the end points of each
interval is subtracted from each point of the
cumulative sums X(t, n) before the calculation of the
This integrated series is divided into non- local range. In both methods, the range is then
overlapping intervals of length n. In each interval, a divided for normalization by the local standard
least squares line is fit to the data (representing the deviation (S) of the original series x(t). This
trend in the interval). The series X(t) is then locally computation is repeated over all possible interval
detrended by substracting the theoretical values Xn(t) lengths (in practice, the shortest length is around 10,
given by the regression. For a given interval length n, and the largest (N-1)/2, giving two adjacent
the characteristic size of fluctuation for this intervals). Finally the rescaled ranges R/S are
integrated and detrended series is calculated by: averaged for each interval length n. R/ S is related to
n by a power law:
N
1
F =
N ∑ [ X(k) −
k =1
X n(k) ]
2
(10)
R/ S ∝ nH (15)

This computation is repeated over all possible Ĥ is expressed as the slope of the double
interval lengths (in practice, the shortest length is logarithmic plot of R/ S as a function of n. R/S
around 10, and the largest N/2, giving two adjacent analysis is theoretically conceived to work on fGn
intervals). Typically, F increases with interval length signals, and should provide irrelevant results for fBm
n. A power law is expected, as signals.

F ∝ nα (11) Dispersional analysis (Disp)

This method was introduced by
α is expressed as the slope of a double Bassingthwaighte (1988). In the original algorithm,
logarithmic plot of F as a function of n. As PSD, the x(t) series is divided into non-overlapping
DFA allows to distinguish between fGn and fBm intervals of length n. The mean of each interval is
series.fGn corresponds to α exponents ranging from computed, and then the standard deviation (SD) of
0 to 1, and fBm to exponents from 1 to 2. α can be these local means, for a given length n. These
converted into Ĥ according to the following computations are repeated over all possible interval
equations: lengths. SD is related to n by a power law:

Ĥ=α for fGn, (12a) SD ∝ nH-1 (16)

Or The quantity (H-1) is expressed as the slope of

the double logarithmic plot of SD as a function of n.
Ĥ=α-1 for fBm. (12b) Obviously, as the number of means involved in the
calculation depends on the number of available
intervals, the SD’s calculated from the highest values
of n tend to fall below the regression line and bias the
 6
estimate. Caccia et al. (1997) suggested to ignore arising from a Gaussian underlying process with the
measures obtained from the longest intervals. In the autocorrelation function rk ( β ) .
present paper, we considered only the standard If the time series has a variance v, this probability
deviations obtained on the means of at least 6 non- is:
overlapping intervals. As R/S analysis, Disp is
theoretically conceived to work on fGn signals, and
1  1 
p ( x, β ) = n/2 1/ 2
exp − x t [ Rv ( β )]−1 x 
should provide irrelevant results for fBm signals. (2π ) det( Rv ( β ))  2 
Caccia et al. (1997) proposed two techniques for (19)
improving this original algorithm. Disps allows where x = ( x1 , x 2 ,...., x n ) t is the vector of the
obtaining multiple estimates of SD for a given
interval length. For each interval length, several analyzed time series and Rv ( β ) is a covariance
partitions of non-overlapping intervals are obtained matrix of size n × n given by:
by shifting the starting position by one point.
Theoretically, n-1 partitions can be obtained for an Rv ( β ) = v.R ( β ) (20)
interval length n, but practically the number of
partitions is limited to 16. The multiple estimates of
where R ( β ) is defined by the symmetrical matrix:
SD are then averaged for a given interval length.
Caccia et al. (1997) proposed another
modification, Dispr, where SD is estimated  r0 ( β ) r1 ( β ) . . . rn −1 ( β ) 
 r (β ) . . . rn − 2 ( β )
iteratively. They showed (see Caccia et al., 1997, p.  1 r0 ( β )
615-616 for details) that the following expression:  . . . . . . 
R( β ) = [ri − j ( β )]1≤ i , j ≤ n = 
 . . . . . . 
1 k 2   . . 
2 H +1 ∑
var(n) = X i + kX  (17) . . . .
 
k+k  i =1  rn −1 ( β ) rn − 2 ( β ) . . . r0 ( β ) 
(21)
where k represents the number of non-
overlapping intervals of length n, provided an The MLE principle states that the optimal β is
unbiased estimate of the means’ variance of intervals the one maximizing the value of the probability
of length n, with a known H exponent. In a first step, p ( x, β ) . After substituting the variance v, the
H is arbitrarily set at H = 0.99. SD is then estimated
for each interval length n, according to Eq. 17, and an function L( x, β ) to be maximized is generally
estimate of H is obtained from Eq. 16. This process is written using a logarithm (Deriche & Tewfik, 1993).
repeated six times, using the obtained estimate of H
in Eq. 17. In the present study we used n 1
simultaneously both modifications. The resultant 2
( 2
)
L( x, β ) = − log x t [ R ( β )]−1 x − log(det( R ( β )) )
Dispsr method was showed to reduce bias and (22)
variance in the estimation of H (Caccia et al., 1997).
Maximum likelihood estimation (MLE) Hence, for a given time series defined by a vector
We used the maximum likelihood estimator x of length n, one can compute the optimal β
proposed by Deriche and Tewfik (1993). This
method works only on fGn series and is known to parameter maximizing L( x, β ) .
provide a low variability in H estimates (Pilgram & Technically, the estimation of the Hurst exponent
Kaplan, 1998). using the MLE method is time consuming essentially
The autocorrelation function associated to fGn is because of the size of the covariance matrix R ( β ) .
given by: As an example, for a time series of length 512
(matrix R ( β ) of size 512 × 512), it will take
(−1) k Γ(1 − β ) (18) approximately 10 minutes to get the result when the
rk ( β ) =
 β  β MLE is processed on a 3.06 GHz Pentium 4 CPU
Γ1 + k − .Γ1 − k −  computer (with a RAM of 1024 Mo). Considering
 2  2
these limitations, we didn’t apply this method to
series longer than 512 points.
This expression naturally leads to the definition
of a maximum likelihood criteria for the estimation
Scaled windowed variance methods (SWV)
of the parameter β (Pilgram & Kaplan, 1998). This These methods were developed by Cannon et al.
technique is a classical in signal processing and is (1997). The x(t) series is divided into non-
generally efficient. overlapping intervals of length n. Then the standard
The first step of this approach is to express the
deviation is calculated within each interval using the
probability of observing a time series of length n formula:
 7
n The spectral index β provided by PSD, lowPSDwe
∑ [ x(t ) − x ]² and MLE, and the α exponent of DFA were
SD = t =1
(23) converted into Ĥ using the previously described
n −1 equations (Eq. 7 and 12). We then obtained for each
method one sample of 40 estimates (Ĥ), for each true
where x is the average within each interval. H value, each class of signal, and each series length.
The means and standard deviations of these samples
Finally the average standard deviation ( SD ) of all were computed, in order to assess, respectively, bias
intervals of length n is computed. This computation is and variability. These two indicators were considered
repeated over all possible interval lengths. For a separately. Note that our goal was to roughly
fractal series SD is related to n by a power law: characterize and localize biases and variability, and
not to accurately describe the mathematical
relationships between true H and Ĥ, or between true
SD ∝ nH (24)
H and standard deviation. As such, the use of 40
simulated series per condition was considered as
Ĥ is expressed as the slope of the log-log plot of sufficient for contrasting means and standard
SD as a function of n. Cannon et al. (1997) showed deviations.
that a detrending of the series within each interval In a second step, we added to each original series
before the calculation of the standard deviation (fGn and fBm) a white noise series (fGn with H =
provided better estimates of H, especially with short 0.5). The added white noise series were different for
series. In this paper we tested the two detrending each fGn or fBm series. We tested four noise/signal
techniques proposed by the authors: the linear SD ratios: 0.00 (no added white noise), 0.33, 0.66,
detrending (ldSWV) is performed by removing the and 1.00 (equal variance for white noise and signal).
regression line within each considered interval, and Note that in the case of fBm series, this ratio does not
bridge detrending (bdSWV) by removing the line express the ratio between the SD of white noise and
connecting the first and last points of the interval. the SD of the fBm series, but the ratio between the
Exploiting the diffusion properties of signals (i.e. the SD of white noise and the SD of the fGn that was
expected increase of variance over time, expressed by summed to obtain the fBm. All methods were then
Eq. 1), SWV methods are conceived to work properly applied to these contaminated signals. Nevertheless,
on fBm, but should provide irrelevant results on fGn. we restricted in this second step the application of the
SWV methods can also be used to distinguish methods to the cases where they were previously
between fGn and fBm near the 1/f boundary. Eke et proven to be relevant. These tests were performed for
al. (2000) proposed a method called Signal a single series length (512 points).
Summation Conversion method (SSC), based on the Finally, we assessed the capacity of four methods
application of SWV on the cumulative sum of the (PSD, lowPSDwe, DFA, and SSC) to distinguish
original signal. If the obtained Ĥ is lower than 1.0 the between fGn and fBm near the 1/f boundary, by the
original series is an fGn (in this case the cumulant analysis of misclassification rates for fGn series with
series is the corresponding fBm). If Ĥ is higher than H = 0.8 and H = 0.9 and fBm series with H = 0.1 and
1.0 the original series is a fBm. H = 0.2.

Procedure Results
We generated 40 fGn series of 2048 data points Power spectral density analyses
for each of 9 values of H ranging from 0.1 to 0.9 by Results concerning the effect of series length on
steps of 0.1. These series were then cumulatively bias and variability in H estimation with PSD are
summed to obtain the corresponding fBm series. We displayed in figure 2. As can be seen, PSD works
then applied on all series (fGn and fBm) the quite well for fGn series, despite an underestimation
previously described methods: PSD, lowPSDwe, DFA, of H for anti-persistent noises (H < 0.4), and a slight
R/S, Dispsr, MLE, ldSWV, and bdSWV. In order to overestimation for H>0.7. The length of the series
test the effect of series length on H estimation, each didn’t seem to have a great influence on the
method was applied on the entire series (2048 magnitude of these biases, except for the shortest one
points), and then on the first 1024, 512, 256, 128 and (26 points), for the highest and lowest H values. The
64 points (i.e. series of 211, 210, 29, 28, 27 and 26 variability of estimation was quite low (around
points). The choice of series lengths that are powers 0.025) for the longest series, but increased as series
of 2 was motivated by the requirements of spectral length decreased. Variability was particularly high
methods. In order to facilitate comparisons, we for the shortest series (26 and 27 points). The results
adopted the same series lengths for all methods. The were less convincing for fBm series. PSD presented
only exception to these general rules was MLE, for fBm series a global bias of underestimation,
which was exclusively applied to fGn series, and for which was strangely reduced as series length
lengths ranging from 64 to 512 data points. decreased. This bias was particularly important for
 8
fGn
1.2 0.3
2048
1 2048 1024
0.25
mean estimated Hurst exponent

1024 512
512 256
0.8 128
256

standard deviation
0.2 64
128
0.6 64
0.15
0.4
0.1
0.2

0.05
0

-0.2 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

true Hurst exponent true Hurst exponent

fBm

1 0.3
2048
2048 1024
0.8 0.25
mean estimated Hurst exponent

1024 512
512 256
128
standard deviation

0.6 256 0.2

128 64
64
0.4 0.15

0.2 0.1

0 0.05

-0.2 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
true Hurst exponent true Hurst exponent

Figure 2: PSD analysis. Plots of mean Ĥ versus H (upper panels), and Ĥ standard deviation versus
H (lower panels). Results for fGn series are displayed in the left column, and for fBm series in the
right column.

anti-persistent motions (H<0.4): note that the mean Ĥ For fBm series, the bias remained qualitatively the
for fBm series with H = 0.1 was lower than 0. This same, with a global underestimation of H, but
bias was dramatic for persistent motions (H>0.5), increased in magnitude as the ratio increased.
with a global underestimation toward Ĥ = 0.5. Results concerning the effect of series length on
Despite these important biases, estimation variability bias and variability in H estimation with lowPSDwe are
remained moderate, for the longest series (29 to 211 displayed in figure 4. As suggested by Eke et al.
points), but increased as series length decreased, (2000), the combination of preprocessing operations
especially for anti-persistent motions (H<0.4). and the exclusion of the high-frequency power
Results concerning the influence of an added estimates in the fitting procedure led to a quite good
white noise on bias and variability in Ĥ with PSD are correction of biases. The estimation was quite
displayed in figure 3. For fGn series, noise seemed to accurate for series of 2048 and 1024 points, despite a
induce biases opposite to the intrinsic biases of PSD, slight underestimation when H < 0.3, for fGn as well
with overestimations for H < 0.4 and as for fBm. With shorter series, an underestimation
underestimations for H > 0.6. As a consequence, PSD appeared for fGn, especially for H > 0.3. This bias
gave precise assessments for a ratio of 0.33 between became dramatic for the shortest series (26 and 27
the SD of white noise and the SD of fGn. For higher points). The effect of series length on H estimation
ratios, a global bias toward Ĥ = 0.5 was observed. for fBm series was less evident, except for the
 9
fGn

1.2 0.3
mean estimated Hurst exponent

SD whit e noise / SD fGn SD white noise / SD f Gn

1 0.25
0.00 0.33 0.00 0.33

standard deviation
0.8 0.66 1.00 0.66 1.00
0.2
0.6
0.15
0.4
0.1
0.2

0 0.05

-0.2 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

true Hurst exponent true Hurst exponent

fBm

1 0.3
mean estimated Hurst exponent

SD white noise / SD fGn

0.8 0.25
SD white noise / SD f Gn
0.00 0.33
0.66 1.00 0.00 0.33
standard deviation

0.6
0.2 0.66 1.00
0.4
0.15
0.2
0.1
0

-0.2 0.05

-0.4 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
true Hurst exponent true Hurst exponent

Figure 3: PSD method. Influence of noise/signal ratio on bias and variability in H estimation.
Results for fGn series are displayed in the left column, and for fBm series in the right column.

shortest lengths (26 and 27 points) that led to a global increase as H increased, and especially for short
underestimation of H. But the most important result series of persistent noise (H > 0.5). The results were
was the dramatic increase of variability, which less convincing for fBm series, with a global
reached unacceptable levels for series lengths lower underestimation of H, which seemed to affect the
than 1024 points. whole range of fBm (except Brownian motion, for H
Results concerning the influence of an added = 0.5). This underestimation bias was particularly
white noise on bias and variability in Ĥ with lowPSDwe important for the shortest series (64 points).
are displayed in figure 5. The effects were quite Moreover, the variability of estimates reached
similar to those observed with PSD, with a reversal unacceptable levels, whatever the location of the
of biases for fGn series, and the appearance of an series in the fBm continuum, and whatever series
underestimation bias for fBm series, as the ratio length.
increased. Finally, the addition of white noise seemed Results concerning the influence of an added
to have no effect on estimation variability, for fGn as white noise on bias and variability in H estimation
well as for fBm series. with DFA are displayed in figure 7. Noise induced a
global bias toward Ĥ = 0.5 for anti-persistent noises,
Detrended fluctuation analysis with a concomitant increase of variability. One could
Results concerning the effect of series length on note also a slight underestimation bias for persistent
bias and variability in H estimation with DFA are noises. Finally, the addition of noise had no effect for
displayed in figure 6. DFA worked particularly well fBm series, neither for bias nor for variability.
with fGn series, with no apparent bias and no effect
of series length, whatever the true value of H. The
variability of estimates tended nevertheless to
 10
fGn
1 0.8
2048 2048 1024
512 256
1024 0.7 128 64
0.8
mean estimated Hurst exponent

512
256 0.6
128

standard deviation
0.6
64 0.5

0.4 0.4

0.3
0.2
0.2
0
0.1

-0.2 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
true Hurst exponent true Hurst exponent

fBm

1.2 0.8
2048 2048 1024
1 0.7 512 256
1024 128 64
mean estimated Hurst exponent

512
0.8 0.6
256
128
standard deviation

0.6
64 0.5
0.4
0.4
0.2
0.3
0
0.2
-0.2

-0.4 0.1

-0.6 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
true Hurst exponent true Hurst exponent

Figure 4: lowPSDwe method. Plots of mean Ĥ versus H (upper panels), and Ĥ standard deviation
versus H (lower panels). Results for fGn series are displayed in the left column, and for fBm series
in the right column.

R/S detrended analysis Results concerning the influence of an added

Results concerning the effect of series length on white noise on bias and variability in H estimation
bias and variability in H estimation with R/S are with R/S are displayed in figure 9. These tests were
displayed in figure 8. A systematic overestimation only conducted with fGn series. As can be seen, the
appeared for H < 0.7, and especially for H < 0.4, and addition of white noise increased the overestimation
tended to slightly increase as series length decreased. bias for H < 0.5, leading to a global bias toward Ĥ =
The variability of estimates was quite low for H <0.5, 0.5. For persistent noises (H > 0.5), the addition of
tended to increase for H >0.5, and was moderately white noise induced a slight underestimation. In
affected by series length. As hypothesized, R/S general, this addition had no marked effect on
analysis gave irrelevant results for fBm series, with a estimation variability.
global bias toward Ĥ = 1.
 11
fGn

1 0.3
mean estimated Hurst exponent

SD whit e noise / SD f Gn SD whit e noise / SD f Gn

0.8 0.00 0.33 0.25 0.00 0.33

standard deviation
0.66 1.00 0.66 1.00
0.6 0.2

0.4 0.15

0.2 0.1

0 0.05

-0.2 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
true Hurst exponent true Hurst exponent

fBm

1.2 0.3
mean estimated Hurst exponent

SD whit e noise / SD f Gn SD whit e noise / SD f Gn

1 0.00 0.33 0.25 0.00 0.33
0.66 1.00
standard deviation

0.8 0.66 1.00

0.2
0.6
0.15
0.4
0.1
0.2

0 0.05

-0.2 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
true Hurst exponent true Hurst exponent

Figure 5: lowPSDwe method. Influence of noise/signal ratio on bias and variability in H estimation.
Results for fGn series are displayed in the left column, and for fBm series in the right column.

Dispersional analysis Results concerning the influence of an added

Results concerning the effect of series length on white noise on bias and variability in H estimation
bias and variability in H estimation with Dispsr are with Dispsr are displayed in figure 11. These tests
displayed in figure 10. As can be seen, this method were only conducted with fGn series. The addition of
seemed characterized by a global underestimation noise induced a slight increase of the underestimation
bias. We also applied to our series the original bias for the higher values of H, but introduced a
version of Disp analysis: this method produced a global bias toward Ĥ = 0.5 for series with H < 0.5.
more pronounced underestimation bias for persistent
noises (H > 0.5). This default was partly corrected Maximum likelihood estimation
with the improved version presently used. The Results concerning the effect of series length on
decrease of series length had no effect on the bias and variability in H estimation with MLE are
magnitude of this bias, except for the shortest series displayed in figure 12. As previously explained, these
(64 points). Estimation variability was in general analyses were restricted to the shortest time series, up
higher than that observed for R/S analysis, and was to 512 points. This method was characterized by an
particularly important for the shortest series (128 and underestimation bias for anti-persistent noises,
64 points). Finally, as observed for R/S analysis, especially for the lowest H values. Conversely, this
Dispsr gave irrelevant results for fBm series, with a method appeared accurate for persistent fGn, despite
global bias toward Ĥ = 1. a slight positive bias. Series length didn’t seem to
have any effect on estimation accuracy.
 12
fGn

1 0.4

2048 0.35 2048

mean estimated Hurst exponent

0.8 1024 1024

512 0.3 512

standard deviation
256 256
0.25
0.6 128 128
64 64
0.2
0.4 0.15

0.1
0.2
0.05

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

true Hurst exponent true Hurst exponent

fBm

1 0.4

2048 0.35
mean estimated Hurst exponent

0.8 1024
512 0.3
standard deviation

0.6 256
128 0.25
64
0.4 0.2

0.15
0.2
0.1
0 2048 1024 512
0.05
256 128 64
-0.2 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

true Hurst exponent true Hurst exponent

Figure 6: DFA. Plots of mean Ĥ versus true H (upper panels), and Ĥ standard deviation versus H
(lower panels). Results for fGn series are displayed in
R/S analysis
MLE
Moreover, variability remained limited, even which seems more convincing, from a strictly
with the shortest series. mathematical point of view, for controlling local
Results concerning the influence of an added trends in the series (figure 14). As hypothesized,
white noise on bias and variability in H estimation SWV analyses gave irrelevant results for fGn series,
with MLE are displayed in figure 13. Noise with a global bias toward Ĥ = 0. For fBm series, no
introduced a global bias toward Ĥ = 0.5, inducing apparent bias was noticeable, whatever series length.
overestimation for anti-persistent series, and These methods seemed to provide very accurate
underestimation of persistent series. Surprisingly, mean estimates of H, even with very short series. The
noise tended to reverse the intrinsic biases of MLE, variability of estimates tended nevertheless to
leading to a quite perfect H estimation for a moderate increase as H increased, especially for short series of
percentage of noise contamination (33%). persistent motion (H > 0.5).
Results concerning the influence of an added
Scaled windowed variance analyses white noise on bias and variability in H estimation
The two tested methods, ldSWV and bdSWV, with ldSWV are displayed in figure 15. These
gave essentially similar results. We present here only analyses were only conducted for fBm series. As can
the results concerning the linear detrended method, be seen, the addition of noise induced a slight bias of
 13
fGn

1 0.3
mean estimated Hurst exponent

SD white noise / SD f Gn
SD whit e noise / SD f Gn
0.00 0.33 0.25 0.00 0.33
0.8

standard deviation
0.66 1.00 0.2 0.66 1.00
0.6
0.15
0.4
0.1
0.2 0.05

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
true Hurst exponent true Hurst exponent

fBm

1 0.3
mean estimated Hurst exponent

SD white noise / SD f Gn SD white noise / SD f Gn

0.8 0.00 0.33 0.25 0.00 0.33
standard deviation

0.66 1.00 0.66 1.00

0.6 0.2

0.4 0.15

0.2 0.1

0 0.05

-0.2 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

true Hurst exponent true Hurst exponent

Figure 7: DFA. Influence of noise/signal ratio on bias and variability in H estimation. Results for
fGn series are displayed in the left column, and for fBm series in the right column.

underestimation, but had no effect on estimation classification on lowPSDwe or SSC, and then to apply
variability. Disp on fGn series and SWV on fBm series. These
The performances of the four methods (PSD, two methods were presented as the best tools for
low
PSDwe, DFA, and SSC) able to classify series in providing the most reliable estimates of H. The
fGn or fBm near the 1/f boundary can be compared in present work, focusing on short time series, led us to
Table 1. PSD completely failed to recognized as fBm more complex conclusions. As suggested in the
original fBm series with H = 0.1. On the other hand, introduction, each method seemed to present specific
this method worked quite well with fGn with H = 0.9, advantages and drawbacks, in terms of bias or
at least with series of 2048 or 1024 data points. variability. Then the decision to apply a given
low
PSDwe gave acceptable results for fGn, except for method should consider the precise aim of the
the shortest series, but the percentage of research (e.g.. precise estimation of exponents, or
misclassifications for fBm was clearly unacceptable. means comparison). Moreover, biases or variability
Finally DFA and SSC gave similar results, with levels are not identical over the entire range of H, and
moderate percentages of misclassifications for fGn then the relevancy of each method could be defined
when series were sufficiently long, but unacceptable not only in terms of class (fGn or fBm), but more
percentages for fBm, whatever series length. precisely in terms of H intervals within a given class.
Methods applicability
Discussion
Eke et al. (2000) highlighted the necessity to
classify signals as fGn or fBm before the application
of fractal analyses. They proposed to base this
 14

1 0.4

2048 0.35 2048

mean estimated Hurst exponent

0.8 1024 1024

512 0.3 512

standard deviation
256 256
0.25
0.6 128 128
64 64
0.2
0.4 0.15

0.1
0.2
0.05

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

true Hurst exponent true Hurst exponent

Figure 8: R/S detrended analysis. Plots of mean Ĥ versus H (left), and Ĥ standard deviation versus H
(right). Results are given for fGn series.

1 0.3
mean estimated Hurst exponent

SD white noise / SD f Gn SD whit e noise / SD f Gn

0.00 0.33 0.25 0.00 0.33

0.8
0.66 1.00
standard deviation

0.66 1.00
0.2
0.6
0.15
0.4
0.1

0.2
0.05

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
true Hurst exponent true Hurst exponent

Figure 9: R/S detrended analysis. Influence of noise/signal ratio on bias and variability in H
estimation. Results are given for fGn series.

1 0.4
2048
2048 0.35 1024
mean estimated Hurst exponent

0.8 1024 512

512 0.3 256
standard deviation

256 128
0.25 64
0.6 128
64 0.2

0.4 0.15

0.1
0.2
0.05

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

true Hurst exponent true Hurst exponent

Figure 10: Dispersional analysis. Plots of mean Ĥ versus H (left), and Ĥ standard deviation versus H
(right). Results are given for fGn series.
 15

1 0.3
mean estimated Hurst exponent SD whit e noise / SD fGn
SD white noise / SD fGn
0.00 0.33 0.25
0.8 0.00 0.33

standard deviation
0.66 1.00 0.66 1.00
0.2
0.6
0.15
0.4
0.1
0.2
0.05

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
true Hurst exponent true Hurst exponent

Figure 11: Dispersional analysis. Influence of noise/signal ratio on bias and variability in H
estimation. Results are given for fGn series.

1.2 0.3
512
1 512
0.25 256
mean estimated Hurst exponent

256 128
0.8
standard deviation

128 0.2 64
0.6 64
0.15
0.4
0.1
0.2

0.05
0

-0.2 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
true Hurst exponent true Hurst exponent

Figure 12: Maximum likelihood estimation. Plots of mean Ĥ versus H (left), and Ĥ standard
deviation versus H (right). Results are given for fGn series.

1.2 0.3
mean estimated Hurst exponent

SD whit e noise / SD f Gn SD whit e noise / SD f Gn

1 0.00 0.33 0.25 0.00 0.33
standard deviation

0.66 1.00 0.66 1.00

0.8 0.2

0.6 0.15

0.4 0.1

0.2 0.05

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
true Hurst exponent true Hurst exponent

Figure 13: Maximum likelihood estimation. Influence of noise/signal ratio on bias and variability in
H estimation. Results are given for fGn series.
 16
1 0.4
2048
2048 0.35
mean estimated Hurst exponent
1024
0.8 1024
0.3 512
512

standard deviation
256 256
0.25 128
0.6 128
64 64
0.2

0.4 0.15

0.1
0.2
0.05

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

true Hurst exponent true Hurst exponent

Figure 14: ldSWV analysis. Plots of mean Ĥ versus H (left), and Ĥ standard deviation versus H
(right). Results are given for fBm series.

1 0.3
mean estimated Hurst exponent

SD whit e noise / SD fGn SD white noise / SD f Gn

0.00 0.33 0.25 0.00 0.33

0.8
standard deviation

0.66 1.00 0.66 1.00

0.2
0.6
0.15
0.4
0.1

0.2
0.05

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
true Hurst exponent true Hurst exponent

Figure 15: ldSWV analysis. Influence of noise/signal ratio on bias and variability in H estimation.
Results are given for fBm series.

Some methods appeared inapplicable for a given sums within each interval: the method is applied on
class of signals, as for example R/S and Disp for fBm fGn, but actually works on the corresponding fBm. A
series, and SWV for fGn series. The underlying similar integration procedure is performed at the first
algorithms can easily explain these incompatibilities. step of the DFA algorithm. DFA can thus be applied
Note that we applied in this paper the classical on fGn but works actually on the corresponding fBm,
algorithms, as commonly reported and used in the as does R/S. SWV methods do not perform this
literature. We thought important to test methods just integration procedure, and work directly on raw data.
as they were used in previous papers, in order to That is why they gave reliable results when directly
allow a posteriori reexaminations. We are convinced, applied on fBm.
nevertheless, that these methods remain improvable: The Mandelbrot and van Ness (1968)’s
the causes of their specific biases have to be clearly scaling law (Eq. 1) holds also for cumulant fBm, but
identified, and their algorithms modified in in this case the exponent is comprised between 1.0
consequence. and 2.0, 1.5 corresponding to the cumulative sum of a
R/S, SWV and DFA exploit the diffusion Brownian motion. This property is exploited by DFA
property of fBm series, according to which variance when applied to fBm series, and also by the SSC
is a power function of the length of the interval of method. One could then wonder why R/S analysis,
observation (Eq. 1). R/S analysis expresses local when applied on fBm series, didn’t give exponents
variance through the rescaled range, and SWV and comprised between 1.0 and 2.0, but appeared
DFA through standard deviation. fGn series, as bounded to 1.0? This bounding effect is due to the
stationary processes, do not possess this diffusion normalization procedure that occurs at the end of the
property, and cannot be directly assessed by this algorithm, when the range is divided by the local
mean. That is why R/S analysis computes cumulative standard deviation of the original series. When this
 17
normalization procedure is omitted, the bounding where n is the length of the subsets, and σ² the series’
effect disappears and Hurst method could then variance (Caccia et al., 1997). This property doesn’t
eventually be applied on fBm, giving in this case hold for fBm, but one could propose to first
exponents comprised between 1.0 and 2.0. differentiate an fBm series, and then to apply Disp on
Disp is only applicable on fGn, but exploits a the corresponding fGn. As well, SWV methods could
different property: the variance of the mean of a become applicable to fGn by introducing an
subset of an fGn series is integration procedure as the first step of the
algorithm. Such procedures could allow to extend the
1 n applicability of all methods to both fGn and fBm
Var[ ∑ x(i)] = σ ²n 2 H −2
n i =1
(25) series (Delignières, Fortes et al., 2004).

Table 1: Percentages of misclassifications for fGn series with H = 0.8 and H = 0.9 (misclassified as fBm), and fBm
series with H = 0.1 and H = 0.2 (misclassified as fGn), observed with PSD, lowPSDwe, DFA, and SSC, and six series
lengths.

______________________________ ______________________________
fGn series (H=0.8) fGn series (H=0.9)
______________________________ ______________________________
N PSD lowPSDwe DFA CSS PSD lowPSDwe DFA CSS
__________________________________________ ______________________________
2048 0.0 0.0 0.0 0.0 0.0 12.5 5.0 10.0
1024 0.0 0.0 0.0 2.5 2.5 0.0 10.0 12.5
512 0.0 0.0 0.0 0.0 30.0 2.5 22.5 22.5
256 0.0 2.5 2.5 5.0 47.5 15.0 22.5 30.0
128 5.0 10.0 7.5 7.5 62.5 10.0 25.0 30.0
64 25.0 12.5 27.5 22.5 72.5 15.0 25.0 22.5
__________________________________________ ______________________________
fBm series (H=0.1) fBm series (H=0.2)
______________________________ ______________________________
N PSD lowPSDwe DFA CSS PSD lowPSDwe DFA CSS
__________________________________________ ______________________________
2048 100.0 35.0 22.5 22.5 2.5 0.0 12.5 7.5
1024 100.0 47.5 45.0 47.5 2.5 2.5 10.0 10.0
512 100.0 60.0 50.0 37.5 12.5 15.0 15.0 15.0
256 100.0 67.5 37.5 37.5 17.5 32.5 32.5 27.5
128 62.5 67.5 75.0 70.0 15.0 45.0 27.5 27.5
64 55.0 82.5 72.5 75.0 17.5 52.5 45.0 47.5
__________________________________________ ______________________________

Series classification bias that characterizes all methods for fBm series
The preliminary classification of series as fGn or with low H exponents. This negative bias is
fBm is a crucial step in fractal analysis. This particularly salient for PSD: all fBm series with H =
procedure requires methods that can be applied to 0.1 were classified as fGn using this method. PSD
both classes of signals. We tested in this study four of worked better for H = 0.2, despite the negative bias,
these methods: PSD, lowPSDwe, DFA, and SSC. because of a low variability in H estimation. Note
All these methods seemed able to distinguish that the addition of noise dramatically increases the
between fGn and fBm, at least when true H negative bias (figure 3). lowPSDwe also presents a
exponents were sufficiently far from the 1/f negative bias for fBm series with H = 0.1 or H = 0.2
boundary. Nevertheless, a zone of uncertainty (figure 4). This bias is lesser than for PSD, but
remains, as can be seen from the results reported in increases when series length decreases (figure 4) and
table 1, for fGn with H = 0.9, and for fBm with H = when noise is added (figure 5). Note also that Ĥ
0.1 and H = 0.2. These results should be understood variability was very high with the shortest
as follows: a number of series classified as fGn with series(figure 4). This increase in variability is related
exponents close to 1 are in fact fBm processes. The to the few number of points that are involved in the
opposite can also be observed, but to a lesser extend. fitting for Ĥ with this method. DFA gave quite
This asymmetry results from the important negative similar results in terms of misclassification
 18
percentages, because of a global negative bias for R/S analysis presents a positive bias for series
fBm series and a rather high variability in H with H < 0.4, and this bias tends to increase toward Ĥ
estimation, whatever series length (figure 6). Finally = 0.5 in the presence of added noise. This bias for
SSC appeared unable to provide a better signal fGn is known and was already described by Caccia et
classification in this uncertainty range, but as al. (1997). This phenomenon seems related to the use
previously indicated, this method exploits similar of range, instead of standard deviation, for estimating
mathematics as DFA. the diffusion property of the integrated version of the
This difficulty to distinguish between fGn and signal. On the other hand, this method presents quite
fBm around the 1/f boundary is problematic, as a limited biases for H ≥ 0.5, and, moreover, a low
number of empirical series produced by variability within this range (figures 8 and 9). Disp
psychological or behavioral systems falls into this could also be proposed for the analysis of fGn series
particular range (e.g. Delignières, Fortes et al., 2004; with H ≤ 0.5. Nevertheless the level of variability
Gilden et al., 1995; Gilden, 2001; Haussdorff et al., seems higher for Disp than for DFA or SWV within
1997). Finally, the best solution when series fall into this H range and Disp is severely biased toward Ĥ =
this uncertainty range could be to restrain analyses to 0.5 for series with H ≤ 0.5 in the presence of white
methods insensitive to the fGn/fBm dichotomy, such noise (figure 10 and 11). Our results concerning Disp
as lowPSDwe or DFA. In other terms, the solution are clearly disappointing, as this method was selected
could be to work directly on β or α exponents, by Eke et al. (2000) as the most relevant for the
without trying to convert them into H metrics. This analysis of fGn series. Finally MLE could give a
could be necessary, for example, when the goal is to possible alternative for persistent noise, despite a
determine the mean fractal exponent of a sample of slight positive bias.
series, and when some series are classified as fGn, In conclusion, the accurate estimation of H
and the others as fBm (see, for example, Delignières, (specifically affected by bias) should follow different
Fortes et al., 2004). The mean exponent can be ways according to the nature of the series. For
computed in this case on the basis of the samples of β antipersistent noises (H < 0.5), the best strategy is to
or α obtained by lowPSDwe or DFA, and then calculate the cumulative sum of the series, and then
eventually converted into H. DFA seems preferable to apply ldSWV or bdSWV. For persistent noise, R/S
in this case, as this method presents lower biases than analysis provides the best results.
spectral analyses. The high variability of DFA should
be compensated, nevertheless, by a sufficient number Means comparisons for fGn series
of series in the sample. The main requirement for means comparison is
Another critical case is when one has to compare to obtain a low variability in H estimation. Limited
two or more mean exponents, and when one of the biases can be accepted, if they do not interfere with
sample meansfalls into the uncertainty range. This the capability of the method to distinguish between
was the case, for example, in studies by Hausdorff et exponents. With this regard, MLE seems the best
al. (1997), Gottschalk et al. (1995), or Peng, Havlin, candidate for fGn series. Despite a negative bias for
Stanley and Goldberger (1995). For optimizing these low values of H, and a slight positive bias for high
means comparisons, H estimation should present a values, the variability in H estimation remains very
low variability and as such lowPSDwe should be low, even for short series (see figure 12). This
preferred to DFA. Remember, nevertheless, that method, nevertheless, is severely time-consuming,
variability reached high levels with lowPSDwe for the and is difficult to use with series longer than 512
shortest series (below 512 data points, see figure 4). points (Pilgram & Kaplan, 1998). PSD can offer an
Means comparisons in this uncertainty range require alternative, despite the presence of similar biases: the
longer series, and one could consider 2048 points as variability in H estimation remains low with PSD,
the shortest acceptable series length. even for short series (see figure 2). Nevertheless,
PSD seemed highly affected by the addition of noise,
Estimating H for fGn series and this method cannot be used when series are
When a series is clearly classified as fGn, a suspected to be contaminated by such random
number of methods are available for a more accurate fluctuations. White noise induces a flattening of the
estimation of its fractal exponent. Clearly the least log-log power spectrum, especially in the high
biased method for fGn series is DFA. Alternatively, frequencies, leading to a typical bias toward Ĥ = 0.5.
one could use SWV methods on the cumulative sum When noise is present, SWV methods, applied on the
of the original series. For these two methods, the bias cumulative sums of the original series, could
remains limited over the whole range of H, and constitute a valuable alternative when H < 0.5. For
variability seems acceptable, at least for H ≤ 0.5. persistent fGn (H > 0.5), R/S analysis seems to be the
Note, nevertheless, that DFA is severely biased best choice.
toward Ĥ = 0.5 for series with H ≤ 0.5 when white
noise is added. This effect did not appear with SWV
methods (see figure 15).
 19
Estimating H for fBm series As can be seen, we propose a procedure quite
Clearly the best methods for fBm series are SWV different than that of Eke et al. (2000). We selected
methods: biases are limited over the whole range of methods on the basis of multiple criteria, according to
H values, and variability remains low, especially for the specific experimental goals that could motivate
H < 0.5. These methods are affected little by series fractal analyses. We showed the necessity to use
length, and by the addition of noise. In contrast, DFA different methods for obtaining an accurate estimate
presents a systematic negative bias and a high level of fractal exponents, or to compare the mean
of variability. For understanding these bad results of exponents obtained in different experimental groups.
DFA with fBm, as compared with SWV, it is We also highlighted the necessity to use different
important to keep in mind that DFA actually works methods, within each class of signal, according to the
on integrated series, and in this case on integrated rough value of H, below or above 0.5. It is important
fBm. This family of over-diffusive processes is not to remember that these methods were selected on the
well known, and the diffusion property exploited by basis of their performances with quite short series.
DFA seems moderately appropriate with such series. Evaluations performed with longer series could
low
PSDwe could represent an interesting alternative, obviously lead to different conclusions.
but is characterized by higher levels of variability
than SWV methods, and some systematic biases for An empirical example
very low and very high H values. As an example, we present in figure 17 (panel a)
a series of 1458 points. This series represents the
Means comparisons for fBm series evolution of self-esteem in an adult participant (age
For sub-diffusive fBm series (H < 0.5), SWV 43) over 729 consecutive days. These data were
methods present the best results: variability remains collected through the bi-daily completion (morning
limited (below 0.1) and biases are absent. Moreover, and evening) of the Physical-Self Inventory (PSI-6,
these methods are not affected by the addition of Ninot, Fortes & Delignières, 2001), a six-item
noise. The choice is more difficult concerning over- questionnaire especially devoted to repeated
diffusive fBm series (H > 0.5), because all methods measurements. Each item measures a specific
present high levels of variability within this range. dimension of the physical self: global self-esteem,
The best choice seems to be lowPSDwe, but the use of physical self-worth, physical condition, sport
time series longer than 1024 point is highly competence, attractive body and physical strength.
recommended. Each item is a simple declarative statement, to which
participants respond using an analog visual scale. The
Noise detection series presented in figure 17 corresponds to the
We frequently evoked in this paper the possible responses to the global self-esteem item.
contamination of empirical series by noise. We As can be seen, the series appears rather
argued that according to the level of contamination, stationary in the long term ( M = 6.80, SD = 0.68),
different methods could be preferred for fractal but presents important fluctuations, in the form of
analyses. The detection of the presence of such multiple interpenetrated ‘waves’. The application of
random fluctuations thus constitutes an important low
PSDwe method to this series suggested the presence
step for such analyses. This can be performed by the of long-term correlation, with a typical linear trend in
inspection of the double-logarithmic plot of the the double-logarithmic plot of the power spectrum
power spectrum provided by PSD. White noise is (Figure 17, panel b). The slope of this spectrum, in
revealed by a flattening of the slope in the high the low-frequency region, was about -1.30,
frequency region. On some occasions, for example in suggesting that the series could be modeled as fBm.
tapping experiments, the log-log plot of power DFA confirmed this diagnostic, with the obtaining of
spectrum presents a positive slope in the high an α exponent of 1.22 (figure 17, panel c). These
frequency region: this suggests the presence of a methods could provide first estimates of H (according
differenced white noise added to the fractal signal to Eq. 7 and 12, Ĥ = 0.15 for lowPSDwe, and Ĥ =0.22
(Gilden et al., 1995; Gilden, 2001). These spectra can for DFA). Note that these two estimates could be
be characterized on the basis of the slopes observed suspected of negative biases. The final estimation of
in the high-frequency and the low-frequency regions, H was then performed with ldSWV, which gave a
and the critical frequency corresponding to the point value of about 0.21 (figure 17, panel d). These
of inflexion between the two portions of the analyses suggested the presence of long-term
spectrum. This information could allow the correlation in this self-esteem series, which could be
assessment of the approximate ratio between the SD considered as an anti-persistent fractional Brownian
of the fractal part of the signal and the SD of noise motion. Further theoretical considerations about
(Delignières, Lemoine & Torre, 2004), thus allowing similar results can be found in Delignières, Fortes et
an appropriate choice of methods of analysis. al. (2004).
The main conclusions of this study are presented
in the flowchart of figure 16.
 20
Detection of noise in the series
PSD

Signal classification
low
PSDwe, DFA, or SSC

fGn uncertainty range fBm

0 < H < 0.9 HfGn > 0.9 or HfBm <0.2 0.2 <H <1.0

Estimating H

H < 0.5 H > 0.5 DFA or lowPSDwe H < 0.5 H > 0.5
DFA R/S SWV SWV
SWV on analysis
cumulated MLE
series

Comparing mean Hs

low
PSD, or PSDwe H < 0.5 H < 0.5
low
SWV if added noise (long series required) SWV PSDwe
MLE for short series (long series
required)

Figure 16: Flowchart for fractal analysis. See text for details.
These observations are very important, because
Series length of the difficulty to obtain long time series in
A last important, and quite unexpected result was psychological and behavioral experiments. As stated
the good performance of most methods in H in the introduction, the validity of H estimation
estimation with very short series. We expected, in supposes that the system remains invariant during the
fact, to find a dramatic increase of biases and whole window of observation. This condition seems
variability with series shorter than 1024 data points. difficult to assure, because of potential problems of
These results were generally present, but with rather fatigue, lack of concentration, etc. Our results
moderate amplitudes. Only lowPSDwe appeared suggests that a better estimate of H could be
severely affected by the shortening of series. The obtained, with a similar time on the experimental
other methods gave acceptable results, at least for task, from the average of four exponents derived
series lengths superior or equal to 256 points. For the from distinct 256 data points series (with an
shortest series (especially 64 points) variability appropriate period of rest between two successive
generally reached unacceptable levels. One can note sessions), than from a single session providing 1024
the exception of MLE, which seemed slightly data points. This conclusion could open new
affected by series length (see figure 12), and should perspectives of research in areas that was until now
be recommended for the analysis of very short series, reticent for using this kind of analyses.
especially for persistent noises.
 21
0
β= 1.30
10 -1

9
-2
8
-3
Global self-esteem

log power
6 -4

5 -5
4
-6
3
-7
2

1 -8

0 -9
0 200 400 600 800 1000 1200 1400 -3.5 -3 -2.5 -2 -1.5 -1 -0.5 0
Observations log frequency

a. b.

2 -0.2

-0.25
1.5
α = 1.22 -0.3 H = 0.21

1 -0.35
log F(n)

log SD

-0.4
0.5
-0.45

0 -0.5

-0.55
-0.5
-0.6

-1 -0.65
1 1.5 2 2.5 3 1 1.5 2 2.5 3

log n log n
c. d.

Figure 17: An example of fractal analysis, performed on a series of self-esteem self-assessments

(1458 data points). Panel a: the raw time series. Panel b, lowPSDwe method: double logarithmic plot
of power against frequency. Panel c, DFA: double logarithmic plot of F(n) against interval length.
Panel c, ldSWV method: double logarithmic plot of the averaged standard deviation against interval
length. See text for the detail of the methods.

Testing for the presence of fractal process application of a unique method in fractal analysis.
All these methods were classically considered as They developed a series of examples showing how
sufficient for evidencing long-range dependences in spectral or time-related methods, applied in isolation,
the analyzed series, through the visual inspection of could lead to false identification of long-range
power spectrum in the frequency domain, or of the dependence. They showed, for example, that a series
diffusion plot in the time domain. Often researchers composed by the superposition of an exponential
applied a unique method (in the frequency domain or trend over a white noise gives a perfect linear fit in
in the time domain), and based their conclusions on the diffusion plot obtained through R/S analysis. The
this visual, and qualitative, observation of a linear spectral method, conversely, provided a flat spectrum
regression in double-logarithmic plots. This apparent revealing the absence of serial correlation in the
simplicity is highly questionable and raises a number series. A first-order auto-regressive process could be
of methodological and theoretical problems. A identified as a fractal series on the basis of R/S
simulated or experimental time series, while having analysis: the diffusion plot presents in this case also a
no long memory property, can mimic the expected perfect linear fit. The absence of long-range
linear fit in log-log plots, and lead to false claims correlation is nevertheless attested by the power
about the presence of underlying fractal processes spectrum, with a typical flattening at low frequencies.
(Thornton & Gilden, 2004; Wagenmakers, Farrell & Rangarajan and Ding (2000) concluded with the
Ratcliff, 2004). necessity of an integrated approach, based on the
Rangarajan and Ding (2000) highlighted the consistent use of several tools, in the frequency as
possible misinterpretations that could arise from the well as in the time domain. The identification of
 22
long-range correlation requires the obtaining of the Lemoine (in press) showed that this method allowed
typical graphical signature with several methods, and an efficient detection of long-range dependence in
also the consistency of the obtained slopes (this simulated series, and could also provide relevant
consistency is assessable through Eqs. 3 and 8). estimates of H. ARFIMA modeling could thus
This integrated approach, nevertheless, remains complete the battery of methods proposed in the
limited to the qualitative analysis of spectral and present study.
diffusion plots, and doesn’t include any test aiming at Thornton and Gilden (2004) proposed to contrast,
statistically evidencing the presence of long-range more directly, on the basis of the obtained power
correlation. Some authors have proposed the spectra, short-range (ARMA) processes and long-
application of surrogate tests, in order to differentiate range fractal processes. They constructed an optimal
between long-range scaling and a random process Bayesian classifier that discriminates between the
with no long-range correlation (see, for example, two families of processes, and showed that this
Haussdorf, Peng, Ladin, Wei, and Goldberger, 1995). classifier had sufficient sensitivity to avoid false
Surrogate data sets are obtained by randomly identifications.
shuffling the original time series. Each surrogate data As can be seen, the identification of true long-
set has the same mean and variance as the range correlation in a series is not so straightforward,
corresponding original series, and differs only in the and remains a current theoretical and methodological
sequential ordering. The scaling exponents of the debate. The simple presence of a linear trend in the
surrogate data sets are then statistically compared to log-log power spectrum, or in a part of this spectrum,
those of the original series. Nevertheless, the interest cannot be per se considered as a definitive proof of
of these tests remains limited, because considering underlying long-range dependence. As such, the
their null hypothesis, they allow attesting for the present results have to be considered as an
presence of correlations in the series, but they are assessment of the ability of classical methods for
unable to certify their long-range nature. estimating the fractal exponent of series whose fractal
This problem was addressed by several recent nature has been already established. An accurate
papers (Thornton & Gilden, 2004; Farrell, estimation of these exponents remains essential, in
Wagenmakers & Ratcliff, 2004; Torre, Delignières & order to assess the effects of experimental conditions
Lemoine, in press; Wagenmakers, Farrell & Ratcliff, and/or participants’ characteristics on fractal
2004). According to these authors, the main question behavior.
is to statistically distinguish between short-term and
long-term dependence in the series. Short-term Conclusion
dependence signifies that the current value in the The present study was based on the dichotomous
series is only determined by a few numbers of model (fGn/fBm) emphasized by Eke et al. (2000).
preceding values. These short-term dependences are We tested similar methods as did these authors, but a
generally modeled by the ARMA models developed separate analysis of bias and variability in H
by Box and Jenkins (1976), which are composed by a estimation, and a focus on short time series led us to
combination of auto-regressive and moving-average slightly different practical considerations. The
terms. A quite simple solution could be to compare procedure we proposed in conclusion is based on the
the shape of the auto-correlation function, which is results obtained in the present study, with the original
supposed to be exponential in the case of a short-term and commonly used algorithms of each method. This
memory process, and to decay according to a power evaluation of the classical versions of fractal analysis
law in the case of long-term dependence. This tools was necessary for a possible reassessing of
comparison, nevertheless, remains qualitative, and previous studies. Nevertheless, we think that all
auto-correlation functions do not present sufficient methods could be improved, for correcting specific
information to give support to unequivocal statistical shortcomings in terms of local biases or variability in
tests. H estimation. Preprocessing operations and
Wagenmakers et al. (2004) based their approach refinements of algorithms could allow coping with
on the so-called ARFIMA models, which are specific problems (trends, noise, etc.) that frequently
frequently used in the domain of econometry for occur in experimental series (Hu, Ivanov, Chen,
modeling long-range dependence (see, for example Carpena, & Stanley, 2001; Chen, Ivanov, Hu, &
Diebolt and Guiraud, 2005). ARFIMA is the acronym Stanley, 2002). A methodological effort is clearly
of autoregressive fractionally integrated moving required for improving the reliability of these
average, and these models differ from the traditional methods for short time series, for a possible
ARMA models by the inclusion of an additional development of fractal approaches in psychological
term, d, corresponding to a fractional integration and behavioral studies.
process. Wagenmakers et al. (2004) proposed to test
the null hypothesis d = 0, in order to determine Authors’notes
whether the analyzed series belongs to the ARFIMA We thank Stefano Lazzari for his helpful contribution
or to the ARMA families. Torre, Delignières and in the implementation of MLE.
 23
References Eke, A., Herman, P., Bassingthwaighte, J.B.,
Bak, P., & Chen, K. (1991). Self-organized Raymond, G.M., Percival, D.B., Cannon, M.,
criticality. Scientific American, 264, 46-53. Balla, I., & Ikrényi, C. (2000). Physiological time
Bassingthwaighte, J.B. (1988). Physiological series: distinguishing fractal noises from motions.
heterogeneity: fractals link determinism and Pflügers Archives, 439, 403-415.
randomness in structure and function. News in Eke, A., Hermann, P., Kocsis, L. & Kozak, L.R.
Physiological Sciences, 3, 5-10. (2002). Fractal characterization of complexity in
Caccia, D.C., Percival, D., Cannon, M.J., Raymond, temporal physiological signals. Physiological
G., & Bassingthwaigthe, J.B. (1997). Analyzing Measurement, 23, R1-R38.
exact fractal time series: evaluating dispersional Epstein, S. (1979). The stability of behaviour: on
analysis and rescaled range methods. Physica A, predicting most of the people much of the time.
246, 609-632. Journal of Personality and Social Psychology, 37,
Cannon, M.J., Percival, D.B., Caccia, D.C., 1097-1126.
Raymond, G.M., & Bassingthwaighte, J.B. (1997). Farrell, S., Wagenmakers, E.-J., & Ratcliff, R.
Evaluating scaled windowed variance methods for (2004). ARFIMA time series modeling of serial
estimating the Hurst coefficient of time series. correlations in human performance. Manuscript
Physica A, 241, 606-626. submitted for publication.
Chen, Y., Ding, M. & Kelso, J.A.S. (1997). Long Fougère, P.F. (1985). On the accuracy of spectrum
memory processes (1/fα type) in human analysis of red noise processes using maximum
coordination. Physical Review Letters, 79, 4501- entropy and periodogram methods : simulation
4504. studies and application to geographical data.
Chen, Z., Ivanov, P.C., Hu, K., & Stanley, H.E. Journal of Geographical Research, 90 (A5), 4355-
(2002). Effects of nonstationarities on detrended 4366.
fluctuation analysis. Physical Review E, 65, Gilden, D.L. (1997). Fluctuations in the time required
041107. for elementary decisions. Psychological Science, 8,
Collins, J.J., & De Luca, C.D. (1993). Open-loop and 296-301.
closed-loop control of posture: A random-walk Gilden, D.L. (2001). Cognitive emissions of 1/f
analysis of center-of-pressure trajectories. noise. Psychological Review, 108, 33-56.
Experimental Brain Research, 95, 308-318. Gilden, D.L., Thornton, T. & Mallon, M.W. (1995).
Davies, R.B., & Harte, D.S. (1987). Tests for Hurst 1/f noise in human cognition. Science, 267, 1837-
effect. Biometrika, 74, 95-101 1839.
Delignières, D., Deschamps, T., Legros, A., & Gottschalk, A., Bauer, M.S., & Whybrow, P.C.
Caillou, N. (2003). A methodological note on (1995). Evidence of chaotic mood variation in
nonlinear time series analysis: Is Collins and De bipolar disorder. Archives of General Psychiatry,
Luca (1993)'s open- and closed-loop model a 52, 947-959.
statistical artifact? Journal of Motor Behavior, 35, Hausdorff, J.M., Mitchell, S.L., Firtion, R., Peng,
86-96. C.K., Cudkowicz, M.E., Wei, J.Y., & Goldberger,
Delignières, D., Fortes, M., & Ninot, G. (2004). The A.L. (1997). Altered fractal dynamics of gait:
fractal dynamics of self-esteem and physical self. reduced stride-interval correlations with aging and
Nonlinear Dynamics in Psychology and Life Huntington's disease. Journal of Applied
Sciences, 8, 479-510. Physiology, 82, 262-269.
Delignières, D., Lemoine, L., & Torre, K. (2004). Hausdorff, J.M., Peng, C.K., Ladin, Z., Wei, J.Y., &
Time intervals production in tapping and Goldberger, A.R. (1995). Is walking a random
oscillatory motion. Human Movement Science, 23, walk? Evidence for long-range correlations in
87-103. stride interval of human gait. Journal of Applied
Deriche, M. & Tewfik, A.H. (1993). Maximum Physiology, 78, 349-358.
likelihood estimation of the parameters of discrete Hu, K., Ivanov, P.C., Chen, Z., Carpena, P. &
fractionally differenced Gaussian noise process. Stanley, H.E. (2001). Effect of trends on detrended
IEEE Transactions on Signal Processing, 41, 2977- fluctuation analysis. Physical Review E, 64,
2989. 011114.
Diebolt, C., & Guiraud, V (2005). A note on long Hurst, H.E. (1965). Long-term storage: An
memory time series. Quality & Quantity, 39, 827- experimental study. London: Constable.
836. Madison, G. (2001). Variability in isochronous
Einstein, A. (1905). Über die von der tapping: Higher order dependencies as a function of
molekularkinetischen Theorie der Wärme intertap interval. Journal of Experimental
geforderte Bewegung von in ruhenden Psychology: Human Perception and performance,
Flüssigkeiten suspendieren Teilchen. Annalen der 27, 411-422.
Physik, 322, 549-560.
 24
Madison, G. (2004). Fractal modeling of human Experimenal Psychology: Human Perception and
isochronous serial interval production. Biological Performance, 17, 635-651.
Cybernetics, 90, 105, 112. Thornton, T., & Gilden, D.L. (2004). Provenance of
Mandelbrot, B.B., & van Ness, J.W. (1968). correlations in psychophysical data. Manuscript
Fractional Brownian motions, fractional noises and submitted for publication.
applications. SIAM Review, 10, 422-437. Torre, K, Delignières, D., & Lemoine, L. (in press).
Marks-Tarlow, T. (1999). The self as a dynamical Detection of long-range dependence and estimation
system. Nonlinear Dynamics, Psychology, and Life of fractal exponents through ARFIMA modelling.
Sciences, 3, 311-345. British Journal of Mathematical and Statistical
Musha, T., Katsurai, K. & Teramachi, Y. (1985). Psychology.
Fluctuations of human tapping intervals. IEEE Van Orden, G.C., Holden, J.C., & Turvey, M.T.
Transactions on Biomedical Engineering, BME-32, (2003). Self-organization of cognitive performance.
578-582. Journal of Experimental Psychology: General, 132,
Peng, C.K., Havlin, S., Stanley, H.E., & Goldberger, 331-350.
A.L. (1995). Quantification of scaling exponents Wagenmakers, E.J., Farrell, S., & Ratcliff, R. (2004).
and crossover phenomena in non stationary Estimation and interpretation of 1/fα noise in
heartbeat time series. Chaos, 5, 82-87. human cognition. Psychonomic Bulletin & Review,
Peng, C.K., Mietus, J., Hausdorff, J.M., Havlin, S., 11, 579-615.
Stanley, H.E., & Goldberger, AL. (1993). Long- West, B.J., & Shlesinger, M.F. (1989). On the
range anti-correlations and non-Gaussian behavior ubiquity of 1/f noise. International Journal of
of the heartbeat. Physical Review Letter, 70, 1343- Modern Physics B, 3, 795-819.
1346. Yamada, M. (1996). Temporal control mechanism in
Pilgram, B. & Kaplan, D.T. (1998). A comparison of equalled interval tapping. Applied Human Science,
estimators for 1/f noise. Physica D, 114, 108-122. 15, 105-110.
Rangarajan, G., & Ding, M. (2000). Integrated Yamada, M., & Yonera, S. (2001). Temporal control
approach to the assessment of long range mechanism of repetitive tapping with simple
correlation in time series data. Physical Review E, rhythmic patterns. Acoustical Science and
61, 4991-5001. Technology, 22, 245-252.
Slifkin, A.B,. & Newell, K.M. (1998). Is variability Yamada, N. (1995). Nature of variability in
in human performance a reflection of system noise? rhythmical movement. Human Movement Science,
Current Directions in Psychological Science, 7, 14, 371-384.
170-177. Yoshinaga, H., Miyazima, S., & Mitake, S. (2000).
Schmidt, R.C., Beek, P.J., Treffner, P.J. & Turvey, Fluctuation of biological rhythm in finger tapping.
M.T. (1991). Dynamical substructure of Physica A, 280, 582-586.
coordinated rhythmic movements. Journal of