arXiv:0804.0747v1 [physics.

data-an] 4 Apr 2008

Fractal and Multifractal Time Series
Jan W. Kantelhardt
Institute of Physics, Martin-Luther-University
Halle-Wittenberg, 06099 Halle, Germany
April 4, 2008

Contents
1 Definition of the Subject and Its Importance 4

2 Introduction 4

3 Fractal and Multifractal Time Series 7

3.1 Fractality, Self-Affinity, and Scaling . . . . . . . . . . . . . . . 7
3.2 Persistence, Long- and Short-term Correlations . . . . . . . . 8
3.3 Crossovers and Non-stationarities in Time Series . . . . . . . . 9
3.4 Multifractal Time Series . . . . . . . . . . . . . . . . . . . . . 10

4 Methods for Stationary Fractal Time Series Analysis 11

4.1 Autocorrelation Function Analysis . . . . . . . . . . . . . . . . 11
4.2 Spectral Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 12
4.3 Hurst’s Rescaled-Range Analysis . . . . . . . . . . . . . . . . 13
4.4 Fluctuation Analysis (FA) . . . . . . . . . . . . . . . . . . . . 14

5 Methods for Non-Stationary Fractal Time-Series Analysis 15

5.1 Wavelet Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 15
5.2 Discrete Wavelet Transform (WT) Approach . . . . . . . . . . 16 5.3
Detrended Fluctuation Analysis (DFA) . . . . . . . . . . . . . 16
5.4 Detection of Trends and Crossovers with DFA . . . . . . . . . 19
5.5 Sign and Magnitude (Volatility) DFA . . . . . . . . . . . . . . 21
5.6 Further Detrending Approaches . . . . . . . . . . . . . . . . . 22
5.7 Centered Moving Average (CMA) Analysis . . . . . . . . . . . 25

6 Methods for Multifractal Time Series Analysis 26

1
6.1 The Structure Function Approach and Singularity Spectra . . 26
6.2 Wavelet Transform Modulus Maxima (WTMM) Method . . . 28
6.3 Multifractal Detrended Fluctuation Analysis (MF-DFA) . . . 29
6.4 Comparison of WTMM and MF-DFA . . . . . . . . . . . . . . 32
7 Statistics of Extreme Events in Fractal Time Series 33
7.1 Return Intervals Between Extreme Events ........... 33
7.2 Distribution of Extreme Events . . . . . . . . . . . . . . . . . 34
8 Simple Models for Fractal and Multifractal Time Series 36
8.1 Fourier Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . 36
8.2 The Schmitz-Schreiber Method ................. 37
8.3 The Extended Binomial Multifractal Model .......... 38
8.4 The Bi-fractal Model . . . . . . . . . . . . . . . . . . . . . . . 39
9 Future Directions 40
10 Bibliography 41

Glossary
• Time series: One dimensional array of numbers (xi), i = 1,...,N,
representing values of an observable x usually measured equidistant
(or nearly equidistant) in time.

• Complex system: A system consisting of many non-linearly

interacting components. It cannot be split into simpler sub-systems
without tampering the dynamical properties.

• Scaling law: A power law with a scaling exponent (e. g. α) describing

the behaviour of a quantity F (e. g., fluctuation, spectral power) as
function of a scale parameter s (e. g., time scale, frequency) at least
asymptotically: F(s) ∼ sα. The power law should be valid for a large
range of s values, e. g., at least for one order of magnitude.

• Fractal system: A system characterised by a scaling law with a

fractal, i. e., non-integer exponent. Fractal systems are characterised

2
 by self-similarity, i. e., a magnification of a small part is statstically
equivalent to the whole.

• Self-affine system: Generalization of a fractal system, where different

magnifications s and s0 = sH have to be used for different directions in
order to obtain a statistically equivalent magnification. The exponent
H is called Hurst exponent. Self-affine time series and time series
becoming self-affine upon integration are commonly denoted as
fractal using a less strict terminology.

• Multifractal system: A system characterised by scaling laws with an

infinite number of different fractal exponents. The scaling laws must
be valid for the same range of the scale parameter.

• Crossover: Change point in a scaling law, where one scaling exponent

applies for small scale parameters and another scaling exponent
applies for large scale parameters. The center of the crossover is
denoted by its characteristic scale parameter s× in this article.

• Persistence: In a persistent time series, a large value is usually (i. e.,

with high statistical preference) followed by a large value and a small
value is followed by a small value. A fractal scaling law holds at least
for a limited range of scales.

• Short-term correlations: Correlations that decay sufficiently fast

that they can be described by a characteristic correlation time scale; e.
g., exponentially decaying correlations. A crossover to uncorrelated
behaviour is observed on larger scales.

• Long-term correlations: Correlations that decay sufficiently slow

that a characteristic correlation time scale cannot be defined; e. g.,
power-law correlations with an exponent between 0 and 1. Power-
law scaling is observed on large time scales and asymptotically. The
term long-range correlations should be used if the data is not a time
series.

• Non-stationarities: If the mean or the standard deviation of the data

values change with time, the weak definition of stationarity is
violated. The strong definition of stationarity requires that all
moments remain constant, i. e., the distribution density of the values

3
 does not change with time. Non-stationarities like monotonous,
periodic, or step-like trends are often caused by external effects. In a
more general sense, changes in the dynamics of the system also
represent non-stationarities.

1 Definition of the Subject and Its Importance

Data series generated by complex systems exhibit fluctuations on a wide
range of time scales and/or broad distributions of the values. In both
equilibrium and non-equilibrium situations, the natural fluctuations are
often found to follow a scaling relation over several orders of magnitude.
Such scaling laws allow for a characterisation of the data and the generating
complex system by fractal (or multifractal) scaling exponents, which can
serve as characteristic fingerprints of the systems in comparisons with
other systems and with models. Fractal scaling behaviour has been
observed, e. g., in many data series from experimental physics, geophysics,
medicine, physiology, and even social sciences. Although the underlying
causes of the observed fractal scaling are often not known in detail, the
fractal or multifractal characterisation can be used for generating surrogate
(test) data, modelling the time series, and deriving predictions regarding
extreme events or future behaviour. The main application, however, is still
the characterisation of different states or phases of the complex system
based on the observed scaling behaviour. For example, the health status and
different physiological states of the human cardiovascular system are
represented by the fractal scaling behaviour of the time series of intervals
between successive heartbeats, and the coarsening dynamics in metal alloys
are represented by the fractal scaling of the time-dependent speckle
intensities observed in coherent X-ray spectroscopy.
In order to observe fractal and multifractal scaling behaviour in time
series, several tools have been developed. Besides older techniques
assuming stationary data, there are more recently established methods
differentiating truly fractal dynamics from fake scaling behaviour caused by
nonstationarities in the data. In addition, short-term and long-term
correlations have to be clearly distinguished to show fractal scaling
behaviour unambiguously. This article describes several methods
originating from Statistical Physics and Applied Mathematics, which have
been used for fractal and multifractal time series analysis in stationary and
non-stationary data.

4
2 Introduction
The characterisation and understanding of complex systems is a difficult
task, since they cannot be split into simpler subsystems without tampering
the dynamical properties. One approach in studying such systems is the
recording of long time series of several selected variables (observables),
which reflect the state of the system in a dimensionally reduced
representation. Some systems are characterised by periodic or nearly
periodic behaviour, which might be caused by oscillatory components or
closed-loop regulation chains. However, in truly complex systems such
periodic components are usually not limited to one or two characteristic
frequencies or frequency bands. They rather extend over a wide spectrum,
and fluctuations on many time scales as well as broad distributions of the
values are found. Often no specific lower frequency limit – or, equivalently,
upper characteristic time scale – can be observed. In these cases, the
dynamics can be characterised by scaling laws which are valid over a wide
(possibly even unlimited) range of time scales or frequencies; at least over
orders of magnitude. Such dynamics are usually denoted as fractal or
multifractal, depending on the question if they are characterised by one
scaling exponent or by a multitude of scaling exponents.
The first scientist who applied fractal analysis to natural time series is
Benoit B. Mandelbrot [1, 2, 3], who included early approaches by H.E. Hurst
regarding hydrological systems [4, 5]. For extensive introductions
describing fractal scaling in complex systems, we refer to [6, 7, 8, 9, 10, 11,
12, 13]. In the last decade, fractal and multifractal scaling behaviour has
been reported in many natural time series generated by complex systems,
including

• geophysics time series (recordings of temperature, precipitation,

water runoff, ozone levels, wind speed, seismic events, vegetational
patterns, and climate dynamics),

• medical and physiological time series (recordings of heartbeat,

respiration, blood pressure, blood flow, nerve spike intervals, human
gait, glucose levels, and gene expression data),

• DNA sequences (they are not actually time series) ,

• astrophysical time series (X-ray light sources and sunspot numbers),

5
 • technical time series (internet traffic, highway traffic, and neutronic
power from a reactor),

• social time series (finance and economy, language characteristics,

fatalities in conflicts), as well as

• physics data (also going beyond time series), e. g., surface roughness,
chaotic spectra of atoms, and photon correlation spectroscopy
recordings.

If one finds that a complex system is characterised by fractal (or

multifractal) dynamics with particular scaling exponents, this finding will
help in obtaining predictions on the future behaviour of the system and on
its reaction to external perturbations or changes in the boundary
conditions. Phase transitions in the regulation behaviour of a complex
system are often associated with changes in their fractal dynamics, allowing
for a detection of such transitions (or the corresponding states) by fractal
analysis. One example for a successful application of this approach is the
human cardiovascular system, where the fractality of heartbeat interval
time series was shown to reflect certain cardiac impairments as well as
sleep stages [14, 15]. In addition, one can test and iteratively improve
models of the system until they reproduce the observed scaling behaviour.
One example for such an approach is climate modelling, where the models
were shown to need input from volcanos and solar radiation in order to
reproduce the long-term correlated (fractal) scaling behaviour [16]
previously found in observational temperature data [17].
Fractal (or multifractal) scaling behaviour certainly cannot be assumed a
priori, but has to be established. Hence, there is a need for refined analysis
techniques, which help to differentiate truly fractal dynamics from fake
scaling behaviour caused, e. g., by non-stationarities in the data. If
conventional statistical methods are applied for the analysis of time series
representing the dynamics of a complex system [18, 19], there are two
major problems. (i) The number of data series and their durations (lengths)
are usually very limited, making it difficult to extract significant information
on the dynamics of the system in a reliable way. (ii) If the length of the data
is extended using computer-based recording techniques or historical
(proxy) data, nonstationarities in the signals tend to be superimposed upon
the intrinsic fluctuation properties and measurement noise. Non-

6
stationarities are caused by external or internal effects that lead to either
continuous or sudden changes in the average values, standard deviations or
regulation mechanism. They are a major problem for the characterisation of
the dynamics, in particular for finding the scaling properties of given data.
Following the description of important properties of fractal and
multifractal time series and the definition of our primary quantities of
interest in Section 3, we focus on methods for the analysis of self-affine or
(mono-)fractal data in Sections 4 and 5. While Section 4 describes
traditional approaches for the analysis of stationary time series, Section 5
discusses more recently established methods applicable for non-stationary
data. Section 6 describes techniques for multifractal time series analysis.
The consequences of fractal scaling behaviour on the statistics of extreme
events and their return intervals are presented in Section 7, and a few
standard models for fractal and multifractal time series are described in
Section 8 before an outlook on future directions in Section 9.
3 Fractal and Multifractal Time Series
3.1 Fractality, Self-Affinity, and Scaling
The topic of this article is the fractality (and/or multifractality) of time
series. Since fractals and multifractals in general are discussed in many
other articles of the encyclopedia, the concept is not thoroughly explained
here. In particular, we refer to the articles ... and ... for the formalism
describing fractal and multifractal structures, respectively.
In a strict sense, most time series are one dimensional, since the values
of the considered observable are measured in homogeneous time intervals.
Hence, unless there are missing values, the fractal dimension of the support
is D(0) = 1. However, there are rare cases where most of the values of a time
series are very small or even zero, causing a dimension D(0) < 1 of the
support. In these cases, one has to be very careful in selecting appropriate
analysis techniques, since many of the methods presented in this article are
not accurate for such data; the Wavelet Transform Modulus Maxima
technique (see Section 6.2) is the most advanced applicable method.
Even if the fractal dimension of support is one, the information
dimension D(1) and the correlation dimension D(2) can be studied. As we
will see in Section 6.1, D(2) is in fact explicitly related to all exponents
studied in monofractal time series analysis. However, usually a slightly
different approach is employed based on the notion of self-affinity instead

7
of (multi-) fractality. Here, one takes into account that the time axis and the
axis of the measured values x(t) are not equivalent. Hence, a rescaling of
time t by a factor a may require rescaling of the series values x(t) by a
different factor aH in order to obtain a statistically similar (i. e., self-similar)
picture. In this case the scaling relation
x(t) → aHx(at) (1)
holds for an arbitrary factor a, describing the data as self-affine (see, e. g.,
[6]). The Hurst exponent H (after the water engineer H.E. Hurst [4])
characterises the type of self affinity. Figure 1(a) shows several examples of
self-affine time series with different H. The trace of a random walk
(Brownian motion, third line in Fig. 1(a)), for example, is characterised by H
= 0.5, implying that the position axis must be rescaled by a factor of 2 if the
time axis is rescaled by a factor of 4. Note that self-affine series are often
denoted as fractal even though they are not fractal in the strict sense. In this
article the term ”fractal” will be used in the more general sense including all
data, where a Hurst exponent H can be reasonably defined.
The scaling behaviour of self-affine data can also be characterised by
looking at their mean-square displacement. Since the mean-square
displacement of a random walker is known to increase linear in time, hx2(t)i
∼ t, deviations from this law will indicate the presence of self-affine scaling.
As we will see in Section 4.4, one can thus retrieve the Hurst (or self-
affinity) exponent H by studying the scaling behaviour of the mean-square
dispalcement, or the mean-square fluctuations hx2(t)i ∼ t2H.

3.2 Persistence, Long- and Short-term Correlations

Self-affine data are persistent in the sense that a large value is usually (i. e.,
with high statistical preference) followed by a large value and a small value
is followed by a small value. For the trace of a random walk, persistence on
all time scales is trivial, since a later position is just a former one plus some
random increment(s). The persistence holds for all time scales, where the
self-affinity relation (1) holds. However, the degree of persistence can also
vary on different time scales. Weather is a typical example: while the
weather tomorrow or in one week is probably similar to the weather today
(due to a stable general weather condition), persistence is much harder to
be seen on longer time scales.

8
 Considering the increments ∆xi = xi − xi−1 of a self-affine series, (xi), i =
1,...,N with N values measured equidistant in time, one finds that the ∆xi can
be either persistent, independent, or anti-persistent. Examples for all cases
are shown in Fig. 1(b). In our example of the random walk with H = 0.5
(third line in the figure), the increments (steps) are fully independent of
each other. Persistent and anti-persistent increments, where a positive
increment is likely to be followed by another positive or negative increment,
respectively, are also leading to persistent integrated series .
For stationary data with constant mean and standard deviation the
autocovariance function of the increments,

. (2)

can be studied to determine the degree of persistence. If C(s) is divided by

the variance h(∆xi)2i, it becomes the auto-correlation function; both are
identical if the data are normalised with unit variance. If the ∆xi are
uncorrelated (as for the random walk), C(s) is zero for s > 0. Short-range
correlations of the increments ∆xi are usually described by C(s) declining
exponentially,
C(s) ∼ exp(−s/t×) (3)
with a characteristic decay time t×. Such behaviour is typical for increments
generated by an auto-regressive (AR) process

∆xi = c∆xi−1 + εi (4)

with random uncorrelated offsets εi and c = exp(−1/t×). Figure 2(a) shows

the auto-correlation function for one configuration of an AR process with t×
= 48.
For so-called long-range correlations diverges in the limit of
infinitely long series (N → ∞). In practice, this means that t× cannot be
defined because it increases with increasing N. For example, C(s) declines as
a power-law
C(s) ∝ s−γ (5)

with an exponent 0 < γ < 1. Figure 2(b) shows C(s) for one configuration
with γ = 0.4. This type of behaviour can be modelled by the Fourier filtering
technique (see Section 8.1). Long-term correlated, i. e. persistent, behaviour

9
of the ∆xi leads to self-affine scaling behaviour of the xi, characterised by H =
1 − γ/2, as will be shown below.

3.3 Crossovers and Non-stationarities in Time Series

Short-term correlated increments ∆xi characterised by a finite characteristic
correlation decay time t× lead to a crossover in the scaling behaviour of the
integrated series , see Fig. 2(a) for an example. Since the
position of the crossover might be numerically different from t×, we denote
it by s× here. Time series with a crossover are not self-affine and there is no
unique Hurst exponent H characterising them. While H > 0.5 is observed on
small time scales (indicating correlations in the increments), the asymptotic
behaviour (for large time scales and ) is always characterised
by H = 0.5, since all correlations have decayed. Many natural recordings are
characterised by pronounced short-term correlations in addition to scaling
long-term correlations. For example, there are short-term correlations due
to particular general weather situations in temperature data and due to
respirational effects in heartbeat data. Crossovers in the scaling behaviour
of complex time series can also be caused by different regulation
mechanisms on fast and slow time scales. Fluctuations of river runoff, for
example, show different scaling behaviour on time scales below and above
approximately one year.
Non-stationarities can also cause crossovers in the scaling behaviour of
data if they are not properly taken into account. In the most strict sense,
non-stationarities are variations in the mean or the standard deviation of
the data (violating weak stationarity) or the distribution of the data values
(violating strong stationarity). Non-stationarities like monotonous, periodic
or step-like trends are often caused by external effects, e. g., by the
greenhouse warming and seasonal variations for temperature records,
different levels of activity in long-term physiological data, or unstable light
sources in photon correlation spectroscopy. Another example for non-
stationary data is a record consisting of segments with strong fluctuations
alternating with segments with weak fluctuations. Such behaviour will
cause a crossover in scaling at the time scale corresponding to the typical
duration of the homogeneous segments. Different mechanisms of regulation
during different time segments – like, e. g., different heartbeat regulation
during different sleep stages at night – can also cause crossovers; they are

10
regarded as non-stationarities here, too. Hence, if crossovers in the scaling
behaviour of data are observed, more detailed studies are needed to find
out the cause of the crossovers. One can try to obtain homogenous data by
splitting the original series and employing methods that are at least
insensitive to monotonous (polynomially shaped) trends.
To characterise a complex system based on time series, trends and
fluctuations are usually studied separately (see, e. g., [20] for a discussion).
Strong trends in data can lead to a false detection of long-range statistical
persistence if only one (non-detrending) method is used or if the results are
not carefully interpreted. Using several advanced techniques of scaling time
series analysis (as described in Chapter 5) crossovers due to trends can be
distinguished from crossovers due to different regulation mechanisms on
fast and slow time scales. The techniques can thus assists in gaining insight
into the scaling behaviour of the natural variability as well as into the kind
of trends of the considered time series.
It has to be stressed that crossovers in scaling behaviour must not be
confused with multifractality. Even though several scaling exponents are
needed, they are not applicable for the same regime (i. e., the same range of
time scales). Real multifractality, on the other hand, is characterised by
different scaling behaviour of different moments over the full range of time
scales (see next section).

3.4 Multifractal Time Series

Many records do not exhibit a simple monofractal scaling behaviour, which
can be accounted for by a single scaling exponent. As discussed in the
previous section, there might exist crossover (time-) scales s× separating
regimes with different scaling exponents. In other cases, the scaling
behaviour is more complicated, and different scaling exponents are
required for different parts of the series. In even more complicated cases,
such different scaling behaviour can be observed for many interwoven
fractal subsets of the time series. In this case a multitude of scaling
exponents is required for a full description of the scaling behaviour in the
same range of time scales, and a multifractal analysis must be applied.
Two general types of multifractality in time series can be distinguished:
(i) Multifractality due to a broad probability distribution (density function)
for the values of the time series, e. g. a Levy distribution. In this case the
multifractality cannot be removed by shuffling the series. (ii) Multifractality

11
due to different long-term correlations of the small and large fluctuations. In
this case the probability density function of the values can be a regular
distribution with finite moments, e. g., a Gaussian distribution. The
corresponding shuffled series will exhibit non-multifractal scaling, since all
long-range correlations are destroyed by the shuffling procedure. Randomly
shuffling the order of the values in the time series is the easiest way of
generating surrogate data; however, there are more advanced alternatives
(see Chapter 8). If both kinds of multifractality are present, the shuffled
series will show weaker multifractality than the original series.
A multifractal analysis of time series will also reveal higher order
correlations. Multifractal scaling can be observed if, e. g., three or four-point
correlations scale differently from the standard two-point correlations
studied by classical autocorrelation analysis (Eq. (2)). In addition,
multifractal scaling is observed if the scaling behaviour of small and large
fluctuations is different. For example, extreme events might be more or less
correlated than typical events.

4 Methods for Stationary Fractal Time Series

Analysis
In this chapter we describe four traditional approaches for the fractal
analysis of stationary time series, see [21, 22, 23] for comparative studies.
The main focus is on the determination of the scaling exponents H or γ,
defined in Eqs. (1) and (5), respectively, and linked by H = 1−γ/2 in long-
term persistent data. Methods taking non-stationarities into account will be
discussed in the next chapter.

4.1 Autocorrelation Function Analysis

We consider a record (xi) of i = 1,...,N equidistant measurements. In most
applications, the index i will correspond to the time of the measurements.
We are interested in the correlation of the values xi and xi+s for different time
lags, i. e. correlations over different time scales s. In order to remove a
constant offset in the data, the mean is usually subtracted,
x˜i ≡ xi − hxi. Alternatively, the correlation properties of increments ˜xi = ∆xi =
xi − xi−1 of the original series can be studied (see also Section 3.2).
Quantitatively, correlations between ˜x-values separated by s steps are

12
defined by the (auto-) covariance function C(s) = hx˜i x˜i+si or the (auto-)
correlation function C(s)/hx˜2i i, see also Eq. (2).
As already mentioned in Section 3.2, the ˜xi are short-term correlated if
C(s) declines exponentially, C(s) ∼ exp(−s/t×), and long-term correlated if
C(s) declines as a power-law C(s) ∝ s−γ with a correlation exponent 0 < γ < 1
(see Eqs. (3) and (5), respectively). As illustrated by the two examples
shown in Fig. 2, a direct calculation of C(s) is usually not appropriate due to
noise superimposed on the data ˜xi and due to underlying non-stationarities
of unknown origin. Non-stationarities make the definition of C(s)
problematic, because the average hxi is not well-defined. Furthermore, C(s)
strongly fluctuates around zero on large scales s (see Fig. 2(b)), making it
impossible to find the correct correlation exponent γ. Thus, one has to
determine the value of γ indirectly.

4.2 Spectral Analysis

If the time series is stationary, we can apply standard spectral analysis
techniques (Fourier transform) and calculate the power spectrum S(f) of the
time series (˜xi) as a function of the frequency f to determine self-affine
scaling behaviour [24]. For long-term correlated data characterised by the
correlation exponent γ, we have

S(f) ∼ f−β with β = 1 − γ. (6)

The spectral exponent β and the correlation exponent γ can thus be

obtained by fitting a power-law to a double logarithmic plot of the power
spectrum S(f). An example is shown in Fig. 3. The relation (6) can be derived
from the Wiener-Khinchin theorem (see, e. g., [25]). If, instead of ˜xi = ∆xi the
integrated runoff time series is Fourier transformed, i. e., ˜ ,
the resulting power spectrum scales as S(f) ∼ f .−2−β

Spectral analysis, however, does not yield more reliable results than
autocorrelation analysis unless a logarithmic binning procedure is applied
to the double logarithmic plot of S(f) [21], see also Fig. 3. I. e., the average of
logS(f) is calculated in successive, logarithmically wide bands from anf0 to
an+1f0, where f0 is the minimum frequency, a > 1 is a factor (e. g., a = 1.1), and
the index n is counting the bins. Spectral analysis also requires stationarity
of the data.

13
4.3 Hurst’s Rescaled-Range Analysis
The first method for the analysis of long-term persistence in time series
based on random walk theory has been proposed by the water construction
engineer Harold Edwin Hurst (1880-1978), who developed it while working
in Egypt. His so-called rescaled range analysis (R/S analysis) [1, 2, 4, 5, 6]
begins with splitting of the time series (˜xi) into non-overlapping segments ν
of size (time scale) s (first step), yielding Ns = int(N/s) segments altogether.
In the second step, the profile (integrated data) is calculated in each
segment ν = 0,...,Ns − 1,

. (7)
By the subtraction of the local averages, piecewise constant trends in the
data are eliminated. In the third step, the differences between minimum and
maximum value (ranges) Rν(s) and the standard deviations Sν(s) in each
segment are calculated,

. (8)
Finally, the rescaled range is averaged over all segments to obtain the
fluctuation function F(s),

for , (9)
where H is the Hurst exponent already introduced in Eq. (1). One can show
[1, 24] that H is related to β and γ by 2H ≈ 1+β = 2−γ (see also Eqs. (6) and
(14)). Note that 0 < γ < 1, so that the right part of the equation does not hold
unless 0.5 < H < 1. The relationship does not hold in general for multifractal
data. Note also that H actually characterises the self-affinity of the profile
function (7), while β and γ refer to the original data.
The values of H, that can be obtained by Hurst’s rescaled range analysis,
are limited to 0 < H < 2, and significant inaccuracies are to be expected close
to the bounds. Since H can be increased or decreased by 1 if the data is
integrated (˜ ) or differentiated (˜xi → x˜i −x˜i−1), respectively, one
can always find a way to calculate H by rescaled range analysis provided the
data is stationary. While values H < 1/2 indicate long-term anti-correlated
behaviour of the data ˜xi, H > 1/2 indicates long-term positively correlated

14
behaviour. For power-law correlations decaying faster than 1/s, we have H
= 1/2 for large s values, like for uncorrelated data.
Compared with spectral analysis, Hurst’s rescaled range analysis yields
smoother curves with less effort (no binning procedure is necessary) and
works also for data with piecewise constant trends.

4.4 Fluctuation Analysis (FA)

The standard fluctuation analysis (FA) [8, 26] is also based on random walk
theory. For a time series (˜xi), i = 1,...,N, with zero mean, we consider the global
profile, i. e., the cumulative sum (cf. Eq. (7))
j

X
Y (j) = x˜i, j = 0,1,2,...,N, (10)
i=1

and study how the fluctuations of the profile, in a given time window of size
s, increase with s. The procedure is illustrated in Fig. 4 for two values of s.
We can consider the profile Y (j) as the position of a random walker on a
linear chain after j steps. The random walker starts at the origin and
performs, in the ith step, a jump of length ˜xi to the right, if ˜xi is positive, and
to the left, if ˜xi is negative.
To find how the square-fluctuations of the profile scale with s, we first
divide each record of N elements into Ns = int(N/s) non-overlapping
segments of size s starting from the beginning (see Fig. 4) and another Ns
non-overlapping segments of size s starting from the end of the considered
series. This way neither data at the end nor at the beginning of the record is
neglected. Then we determine the fluctuations in each segment ν.
In the standard FA, we obtain the fluctuations just from the values of the
profile at both endpoints of each segment ν = 0,...,Ns − 1,

FFA2 (ν,s) = [Y (νs) − Y ((ν + 1)s)]2, (11)

(see Fig. 4) and analogous for ν = Ns,...,2Ns − 1,

. (12)

Then we average ) over all subsequences to obtain the mean

fluctuation F2(s),

15
 . (13)
By definition, F2(s) can be viewed as the root-mean-square displacement of
the random walker on the chain, after s steps (the reason for the index 2 will
become clear later). For uncorrelated xi values, we obtain Fick’s diffusion
law F2(s) ∼ s1/2. For the relevant case of long-term correlations, where C(s)
follows the power-law behaviour of Eq. (5), F2(s) increases by a power law,

F2(s) ∼ sα with α ≈ H, (14)

where the fluctuation exponent α is identical with the Hurst exponent H for
mono-fractal data and related to γ and β by

2α = 1 + β = 2 − γ. (15)

The typical behaviour of F2(s) for short-term correlated and long-term

correlated data is illustrated in Fig. 2. The relation (15) can be derived
straightforwardly by inserting Eqs. (10), (2), and (5) into Eq. (11) and
separating sums over products ˜xix˜j with identical and different i and j,
respectively.
The range of the α values that can be studied by standard FA is limited to
0 < α < 1, again with significant inaccuracies close to the bounds. Regarding
integration or differentiation of the data, the same rules apply as listed for H
in the previous section. The results of FA become statistically unreliable for
scales s larger than one tenth of the length of the data, i. e. the analysis
should be limited by s < N/10.

5 Methods for Non-Stationary Fractal TimeSeries

Analysis
5.1 Wavelet Analysis
The origins of wavelet analysis come from signal theory, where frequency
decompositions of time series were studied [27, 28]. Like the Fourier
transform, the wavelet transform of a signal x(t) is a convolution integral to
be replace by a summation in case of a discrete time series (˜xi),i = 1,...,N,

16
 . (16)

Here, ψ(t) is a so-called mother wavelet, from which all daughter wavelets
ψτ,s(t) = ψ((t − τ)/s evolve by shifting and stretching of the time axis. The
wavelet coefficients Lψ(τ,s) thus depend on both, time position τ and scale s.
Hence, the local frequency decomposition of the signal is described with a
time resolution appropriate for the considered frequency f = 1/s (i. e.,
inverse time scale).

All wavelets ψ(t) must have zero mean. They are often chosen to be
orthogonal to polynomial trends, so that the analysis method becomes
insensitive to possible trends in the data. Simple examples are derivatives of
a
Gaussian, 2), like the Mexican hat wavelet
−ψGauss(2) and the Haar wavelet,
2, and 0 otherwise. It is straightforward to construct Haar wavelet that are
orthogonal to linear, quadratic and cubic trends, e. g.,

for t ∈ [1,2), and 0 otherwise, or 2),

+3 for t ∈ [2,3), −1 for t ∈ [3,4), and 0 otherwise.

5.2 Discrete Wavelet Transform (WT) Approach

A detrending fractal analysis of time series can be easily implemented by
considering Haar wavelet coefficients of the profile Y (j), Eq. (10) [29, 17]. In
this case the convolution (16) corresponds to the addition and subtraction
¯
of mean values of Y (j) within segments of size s. Hence, defining Y ν(s) =
), the coefficients can be written as
¯ ¯
FWT0(ν,s) ≡ LψHaar(0) (νs,s) = Y ν(s) − Y ν+1(s), (17)

¯ ¯ ¯
FWT1(ν,s) ≡ LψHaar(1) (νs,s) = Y ν(s) − 2Y ν+1(s) + Y ν+2(s), and (18)

L ¯ ¯ ¯ ¯
FWT2(ν,s) ≡ ψHaar(2) (νs,s) = Y ν(s) − 3Y ν+1(s) + 3Y ν+2(s) − Y ν+3(s) (19) for
constant, linear and quadratic detrending, respectively. The generalization

17
for higher orders of detrending is obvious. The resulting mean-square
fluctuations ) are averaged over all ν to obtain the mean
fluctuation F2(s), see Eq. (13). Figure 5 shows typical results for WT analysis
of long-term correlated, short-term correlated and uncorrelated data.
Regarding trend-elimination, wavelet transform WT0 corresponds to
standard FA (see Section 4.4), and only constant trends in the profile are
eliminated. WT1 is similar to Hurst’s rescaled range analysis (see Section
4.3): linear trends in the profile and constant trends in the data are
eliminated, and the range of the fluctuation exponent α ≈ H is up to 2. In
general, WTn determines the fluctuations from the nth derivative, this way
eliminating trends described by (n−1)st-order polynomials in the data. The
results become statistically unreliable for scales s larger than one tenth of
the length of the data, just as for FA.

5.3 Detrended Fluctuation Analysis (DFA)

In the last 13 years Detrended Fluctuation Analysis (DFA), originally
introduced by Peng et al. [30], has been established as an important method
to reliably detect long-range (auto-) correlations in non-stationary time
series. The method is based on random walk theory and basically
represents a linear detrending version of FA (see Section 4.4). DFA was later
generalised for higher order detrending [15], separate analysis of sign and
magnitude series [31] (see Section 5.5), multifractal analysis [32] (see
Section 6.3), and data with more than one dimension [33]. Its features have
been studied in many articles [34, 35, 36, 37, 38, 39]. In addition, several
comparisons of DFA with other methods for stationary and non-stationary
time-series analysis have been published, see, e. g., [21, 23, 40, 41] and in
particular [22], where DFA is compared with many other established
methods for short data sets, and [42], where it is compared with recently
suggested improved methods. Altogether, there are about 450-500 papers
applying DFA (till April 2008). In most cases positive auto-correlations were
reported leaving only a few exceptions with anti-correlations, see, e. g., [43,
44, 45].
Like in the FA method, one first calculates the global profile according to
Eq. (10) and divides the profile into Ns = int(N/s) non-overlapping segments
of size s starting from the beginning and another Ns segments starting from
the end of the considered series. DFA explicitly deals with monotonous

18
trends in a detrending procedure. This is done by estimating a polynomial
trend ) within each segment ν by least-square fitting and subtracting
this trend from the original profile (‘detrending’),

. (20)
The degree of the polynomial can be varied in order to eliminate constant
(m = 0), linear (m = 1), quadratic (m = 2) or higher order trends of the
profile function [15]. Conventionally the DFA is named after the order of the
fitting polynomial (DFA0, DFA1, DFA2, ...). In DFAm, trends of order m in the
profile Y (j) and of order m − 1 in the original record ˜xi are eliminated. The
˜
variance of the detrended profile Y s(j) in each segment ν yields the mean-
square fluctuations,

. (21)

As for FA and discrete wavelet analysis, the ) are averaged over all
segments ν to obtain the mean fluctuations F2(s), see Eq. (14). Calculating
F2(s) for many s, the fluctuation scaling exponent α can be determined just
as with FA. Figure 6 shows typical results for DFA of the same long-term
correlated, short-term correlated and uncorrelated data studied already in
Fig. 5.

We note that in studies that include averaging over many records (or
one record cut into many separate pieces by the elimination of some
unreliable intermediate data points) the averaging procedure (13) must be
performed for all data. Taking the square root is always the final step after
all averaging is finished. It is not appropriate to calculate F2(s) for parts of
the data and then average the F2(s) values, since such a procedure will bias
the results towards smaller scaling exponents on large time scales.
If F2(s) increases for increasing s by F2(s) ∼ sα with 0.5 < α < 1, one finds
that the scaling exponent α ≈ H is related to the correlation exponent γ by α
= 1−γ/2 (see Eq. (15)). A value of α = 0.5 thus indicates that there are no (or
only short-range) correlations. If α > 0.5 for all scales s, the data are long-
term correlated. The higher α, the stronger the correlations in the signal are.
α > 1 indicates a non-stationary local average of the data; in this case, FA
fails and yields only α = 1. The case α < 0.5 corresponds to long-term anti-

19
correlations, meaning that large values are most likely to be followed by
small values and vice versa. α values below 0 er not possible. Since the
maximum value for α in DFAm is m + 1, higher detrending orders should be
used for very non-stationary data with large α. Like in FA and Hurst’s
analysis, α will decrease or increase by one upon additional differentiation
or integration of the data, respectively.
Small deviations from the scaling law (14), i. e. deviations from a straight
line in a double logarithmic plot, occur for small scales s, in particular for
DFAm with large detrending order m. These deviations are intrinsic to the
usual DFA method, since the scaling behaviour is only approached
asymptotically. The deviations limit the capability of DFA to determine the
correct correlation behaviour in very short records and in the regime of
small s. DFA6, e. g., is only defined for s ≥ 8, and significant deviations from
the scaling law F2(s) ∼ sα occur even up to s ≈ 30. They will lead to an
overestimation of the fluctuation exponent α, if the regime of small s is used
in a fitting procedure. An approach for correction of this systematic artefact
in DFA is described in [34].
The number of independent segments of length s is larger in DFA than in
WT, and the fluctuations in FA are larger than in DFA. Hence, the analysis
has to be based on s values lower than smax = N/4 for DFA compared with smax
= N/10 for FA and WT. The accuracy of scaling exponents α determined by
DFA was recently studied as a function of the length N of the data [42]
(fitting range s ∈ [10,N/2] was used). The results show that statistical
standard errors of α (one standard deviation) are approximately 0.1 for N =
500, 0.05 for N = 3 000, and reach 0.03 for N = 10 000. Findings of long-term
correlations with α = 0.6 in data with only 500 points are thus not
significant; and α should be at least 0.55 even for data of 10 000 points.
A generalization of DFA for two-dimensional data (or even higher
dimensions d) was recently suggested [33]. The generalization works well
when tested with synthetic surfaces including fractional Brownian surfaces
and multifractal surfaces. The two-dimensional MFDFA is also adopted to
analyse two images from nature and experiment, and nice scaling laws are
unravelled. In the 2d procedure, a double cumulative sum (profile) is
calculated by summing over both directional indices analogous with Eq.
(10), Y (k,l) =
. This surface is partitioned into squares of size s × s with in-

20
dices ν and µ, in which polynomials like
are fitted. The fluctuation function
F2(s) is again obtained by calculating the variance of the profile from the fits.

5.4 Detection of Trends and Crossovers with DFA

Frequently, the correlations of recorded data do not follow the same scaling
law for all time scales s, but one or sometimes even more crossovers
between different scaling regimes are observed (see Section 3.3). Time
series with a well-defined crossover at s× and vanishing correlations above
s× are most easily generated by Fourier filtering (see Section 8.1). The power
spectrum S(f) of an uncorrelated random series is multiplied by (f/f×)−β with
β = 2α − 1 for frequencies f > f× = 1/s× only. The series obtained by inverse
Fourier transform of this modified power spectrum exhibits power-law
correlations on time scales s < s× only, while the behaviour becomes
uncorrelated on larger time scales s > s×.
The crossover from F2(s) ∼ sα to F2(s) ∼ s1/2 is clearly visible in double
logarithmic plots of the DFA fluctuation function for such short-term
correlated data. However, it occurs at times that are different from the
original s× used for the generation of the data and that depend on the
detrending order m. This systematic deviation is most significant in the
DFAm with higher m. Extensive numerical simulations (see Fig. 3 in [34])
show that the ratios of are 1.6, 2.6, 3.6, 4.5, and 5.4 for DFA1, DFA2, ...,
DFA5, with an error bar of approximately 0.1. Note, however, that the
precise value of this ratio will depend on the method used for fitting the
crossover times
(and the method used for generating the data if generated data is
analysed). If results for different orders of DFA shall be compared, an
observed crossover can be systematically corrected dividing by the
ratio for the corresponding DFAm. If several orders of DFA are used in the
procedure, several estimates for the real s× will be obtained, which can be
checked for consistency or used for an error approximation. A real
crossover can thus be well distinguished from the effects of non-
stationarities in the data, which lead to a different dependence of an
apparent crossover on m.

21
 The procedure is also required if the characteristic time scale of
shortterm correlations shall be studied with DFA. If consistent (corrected)
s× values are obtained based on DFAm with different m, the existence of a
real characteristic correlation time scale is positively confirmed. Note that
lower detrending orders are advantageous in this case, since the observed
crossover time scale might become quite large and nearly reach one
forth of the total series length (N/4), where the results become statistically
inaccurate.
We would like to note that studies showing scaling long-term
correlations should not be based on DFA or variants of this method alone in
most applications. In particular, if it is not clear whether a given time series
is indeed long-term correlated or just short-term correlated with a fairly
large crossover time scale, results of DFA should be compared with other
methods. For example, one can employ wavelet methods (see, e. g., Section
5.2). Another option is to remove short-term correlations by considering
averaged series for comparison. For a time series with daily observations
and possible short-term correlations up to two years, for example, one
might consider the series of two-year averages and apply DFA together with
FA, binned power spectra analysis, and/or wavelet analysis. Only if these
methods still indicate long-term correlations, one can be sure that the data
are indeed long-term correlated.
As discussed in Section 3.3, records from real measurements are often
affected by non-stationarities, and in particular by trends. They have to be
well distinguished from the intrinsic fluctuations of the system. To
investigate the effect of trends on the DFAm fluctuation functions, one can
generate artificial series (˜xi) with smooth monotonous trends by adding
polynomials of different power p to the original record (xi),

x˜i = xi + Axp with x = i/N. (22)

For the DFAm, such trends in the data can lead to an artificial crossover in
the scaling behaviour of F2(s), i. e., the slope α is strongly increased for large
time scales s. The position of this artificial crossover depends on the
strength A and the power p of the trend. Evidently, no artificial crossover is
observed, if the detrending order m is larger than p and p is integer. The
order p of the trends in the data can be determined easily by applying the
different DFAm. If p is larger than m or p is not an integer, an artificial

22
crossover is observed, the slope αtrend in the large s regime strongly depends
on m, and the position of the artificial crossover also depends strongly on m.
The artificial crossover can thus be clearly distinguished from real
crossovers in the correlation behaviour, which result in identical slopes α
and rather similar crossover positions for all detrending orders m. For more
extensive studies of trends with non-integer powers we refer to [34, 35].
The effects of periodic trends are also studied in [34].
If the functional form of the trend in given data is not known a priori, the
fluctuation function F2(s) should be calculated for several orders m of the
fitting polynomial. If m is too low, F2(s) will show a pronounced crossover to
a regime with larger slope for large scales s [34, 35]. The maximum slope of
logF2(s) versus logs is m + 1. The crossover will move to larger scales s or
disappear when m is increased, unless it is a real crossover not due to
trends. Hence, one can find m such that detrending is sufficient. However, m
should not be larger than necessary, because shifts of the observed
crossover time scales and deviations on short scales s increase with
increasing m.

5.5 Sign and Magnitude (Volatility) DFA

To study the origin of long-term fractal correlations in a time series, the
series can be split into two parts, which are analysed separately. It is
particularly useful to split the series of increments, ∆xi = xi − xi−1, i = 1,...,N,
into a series of signs ˜xi = si = sign∆xi and a series of magnitudes ˜xi = mi = |∆xi|
[31, 46, 47]. There is an extensive interest in the magnitude time series in
economics [48, 49]. These data, usually called volatility, represents the
absolute variations in stock (or commodity) prices and are used as a
measure quantifying the risk of investments. While the actual prices are
only shortterm correlated, long-term correlations have been observed in
volatility series [48, 49].
Time series having identical distributions and long-range correlation
properties can exhibit quite different temporal organizations of the
magnitude and sign sub-series. The DFA method can be applied
independently to both of these series. Since in particular the signs are often
rather strongly anticorrelated and DFA will give incorrect results if α is too
close to zero, one often studies integrated sign and magnitude series. As
Pi
mentioned above, integration ˜xi → j=1 x˜j increases α by one.

23
 Most published results report short-term anti-correlations and no
longterm correlations in the sign series, i. e., αsign < 1/2 for the non-
integrated signs si (or αsign < 3/2 for the integrated signs) on low time scales
and αsign → 1/2 asymptotically for large s. The magnitude series, on the other
hand, are usually either uncorrelated αmagn = 1/2 (or 3/2) or positively
longterm correlated αmagn > 1/2 (or 3/2). It has been suggested that findings
of αmagn > 1/2 are related with nonlinear properties of the data and in
particular multifractality [31, 46, 47], if α < 1.5 in standard DFA. Specifically,
the results suggest that the correlation exponent of the magnitude series is
a monotonically increasing function of the multifractal spectrum (i. e., the
singularity spectrum) width of the original series (see Section 6.1). On the
other hand, the sign series mainly relates to linear properties of the original
series. At small time scales s < 16 the standard α is approximately the
average of αsign and αmagn, if integrated sign and magnitude series are
analysed. For α > 1.5 in the original series, the integrated magnitude and
sign series have approximately the same two-point scaling exponents [46].
An analytical treatment is presented in [47].

5.6 Further Detrending Approaches

A possible drawback of the DFA method is the occurrence of abrupt jumps
in the detrended profile Y˜s(j) (Eq. (20)) at the boundaries between the
segments, since the fitting polynomials in neighbouring segments are not
related. A possible way to avoid these jumps would be the calculation of
F2(s) based on polynomial fits in overlapping windows. However, this is
rather time consuming due to the polynomial fit in each segment and is
consequently not done in most applications. To overcome the problem of
jumps several modifications and extensions of the FA and DFA methods
have been suggested in the last years. These methods include

• the Detrended Moving Average technique [50, 51, 52], which we

denote by Backward Moving Average (BMA) technique (following
[53]),

• the Centred Moving Average Average (CMA) method [53], an

essentially improved version of BMA,

24
 • the Modified Detrended Fluctuation Analysis (MDFA) [54], which is
essentially a mixture of old FA and DFA,

• the continuous DFA (CDFA) technique [55, 56], which is particularly

useful for the detection of crossovers,

• the Fourier DFA [57],

• a variant of DFA based on empirical mode decomposition (EMD) [58],

• a variant of DFA based on singular value decomposition (SVD) [59,

60], and

• a variant of DFA based on high-pass filtering [61].

Detrended Moving Average techniques will be thoroughly described and

discussed in the next section. A study comparing DFA with CMA and MDFA
can be found in [42]. For studies comparing DFA and BMA, see [62, 63]; note
that [63] also discusses CMA.
The method we denote as Modified Detrended Fluctuation Analysis (MDFA)
[54], eliminates trends similar to the DFA method. A polynomial is fitted to
the profile function Y (j) in each segment ν and the deviation between the
˜
profile function and the polynomial fit is calculated, Y s(j) = Y (j) − yν,sp (j) (Eq.
(20)). To estimate correlations in the data, this method uses a derivative of
˜ ˜ ˜ ˜
Y s(j), obtained for each segment ν, by ∆Y s(j) = Y s(j+s/2)−Y s(j). Hence, the
fluctuation function (compare with Eqs. (13) and (21)) is calculated as
follows:

. (23)
As in case of DFA, MDFA can easily be generalised to remove higher order
trends in the data. Since the fitting polynomials in adjacent segments are
not related, Y˜s(j) shows abrupt jumps on their boundaries as well. This
leads to fluctuations of F2(s) for large segment sizes s and limits the
maximum usable scale to s < N/4 as for DFA. The detection of crossovers in
the data, however, is more exact with MDFA (compared with DFA), since no
correction of the estimated crossover time scales seems to be needed [42].

25
 The Fourier-detrended fluctuation analysis [57] aims to eliminate slow
oscillatory trends which are found especially in weather and climate series
due to seasonal influences. The character of these trends can be rather
periodic and regular or irregular, and their influence on the detection of
long-range correlations by means of DFA was systematically studied
previously [34]. Among other things it has been shown that slowly varying
periodic trends disturb the scaling behaviour of the results much stronger
than quickly oscillating trends and thus have to be removed prior to the
analysis. In case of periodic and regular oscillations, e. g., in temperature
fluctuations one simply removes the low frequency seasonal trend by
subtracting the daily mean temperatures from the data. Another way, which
the Fourier-detrended fluctuation analysis suggests, is to filter out the
relevant frequencies in the signals’ Fourier spectrum before applying DFA
to the filtered signal. Nevertheless, this method faces several difficulties
especially its limitation to periodic and regular trends and the need for a
priori knowledge of the interfering frequency band.
To study correlations in data with quasi-periodic or irregular oscillating
trends, empirical mode decomposition (EMD) was suggested [58]. The EMD
algorithm breaks down the signal into its intrinsic mode functions (IMFs)
which can be used to distinguish between fluctuations and background. The
background, estimated by a quasi-periodic fit containing the dominating
frequencies of a sufficiently large number of IMFs, is subtracted from the
data, yielding a slightly better scaling behaviour in the DFA curves.
However, we believe that the method might be too complicated for wide-
spread applications.
Another method which was shown to minimise the effect of periodic and
quasi-periodic trends is based on singular value decomposition (SVD) [59,
60]. In this approach, one first embeds the original signal in a matrix whose
dimension has to be much larger than the number of frequency components
of the periodic or quasi-periodic trends obtained in the power spectrum.
Applying SVD yields a diagonal matrix which can be manipulated by setting
the dominant eigen-values (associated with the trends) to zero. The filtered
matrix finally leads to the filtered data, and it has been shown that
subsequent application of DFA determines the expected scaling behaviour if
the embedding dimension is sufficiently large. None the less, the
performance of this rather complex method seems to decrease for larger

26
values of the scaling exponent. Furthermore SVD-DFA assumes that trends
are deterministic and narrow banded.
The detrending procedure in DFA (Eq. (20)) can be regarded as a
scaledependent high-pass filter since (low-frequency) fluctuations
exceeding a specific scale s are eliminated. Therefore, it has been suggested
˜
to obtain the detrended profile Y s(j) for each scale s directly by applying
digital highpass filters [61]. In particular, Butterworth, Chebyshev-I,
Chebyshev-II, and an elliptical filter were suggested. While the elliptical
filter showed the best performance in detecting long-range correlations in
artificial data, the Chebyshev-II filter was found to be problematic.
Additionally, in order to avoid a time shift between filtered and original
profile, the average of the directly filtered signal and the time reversed
filtered signal is considered. The effects of these complicated filters on the
scaling behaviour are, however, not fully understood.
Finally, a continuous DFA method has been suggested in the context of
studying heartbeat data during sleep [55, 56]. The method compares
unnormalised fluctuation functions F2(s) for increasing length of the data. I.
e., one starts with a very short recording and subsequently adds more
points of data. The method is particularly suitable for the detection of
change points in the data, e. g., physiological transitions between different
activity or sleep stages. Since the main objective of the method is not the
study of scaling behaviour, we do not discuss it in detail here.
5.7 Centered Moving Average (CMA) Analysis
Particular attractive modifications of DFA are the Detrended Moving
Average (DMA) methods, where running averages replace the polynomial
fits. The first suggested version, the Backward Moving Average (BMA)
method [50, 51, 52], however, suffers from severe problems, because an
artificial time shift of s between the original signal and the moving average
is introduced. This time shift leads to an additional contribution to the
detrended profile Y˜s(j), which causes a larger fluctuation function F2(s) in
particular for small scales in the case of long-term correlated data. Hence,
the scaling exponent α is systematically underestimated [62]. In addition,
the BMA method preforms even worse for data with trends [63], and its
slope is limited by α < 1 just as for the non-detrending method FA.

27
 It was soon recognised that the intrinsic error of BMA can be overcome
by eliminating the artificial time shift. This leads to the Centred Moving
˜
Average (CMA) method [53], where Y s(j) is calculated as

, (24)

replacing Eq. (20) while Eq. (21) and the rest of the DFA procedure
described in Section 5.3 stay the same. Unlike DFA, the CMA method cannot
easily be generalised to remove linear and higher order trends in the data.
It was recently proposed [42] that the scaling behaviour of the CMA
method is more stable than for DFA1 and MDFA1, suggesting that CMA
could be used for reliable computation of α even for scales s < 10 (without
correction of any systematic deviations needed in DFA for this regime) and
up to smax = N/2. The standard errors in determining the scaling exponent α
by fitting straight lines to the double logarithmic plots of F2(s) have been
studied in [42]; they are comparable with DFA1 (see end of Section 5.3).
Regarding the determination of crossovers, CMA is comparable to DFA1.
Ultimately, the CMA seems to be a good alternative to DFA1 when analysing
the scaling properties in short data sets without trends. Nevertheless for
data with possible unknown trends we recommend the application of
standard DFA with several different detrending polynomial orders in order
to distinguish real crossovers from artificial crossovers due to trends. In
addition, an independent approach (e. g., wavelet analysis) should be used
to confirm findings of long-term correlations (see also Section 5.4).
6 Methods for Multifractal Time Series Analysis
This chapter describes the multifractal characterisation of time series, for
an introduction, see Section 3.4. The simplest type of multifractal analysis is
based upon the standard partition function multifractal formalism, which
has been developed for the multifractal characterisation of normalised,
stationary measures [6, 12, 64, 65]. Unfortunately, this standard formalism
does not give correct results for non-stationary time series that are affected
by trends or that cannot be normalised. Thus, in the early 1990s an
improved multifractal formalism has been developed, the wavelet
transform modulus maxima (WTMM) method [66, 67, 68, 69, 70], which is
based on wavelet analysis and involves tracing the maxima lines in the
continuous wavelet transform over all scales. An important alternative is

28
the multifractal DFA (MF-DFA) algorithm [32], which does not require the
modulus maxima procedure, and hence involves little more effort in
programming than the conventional DFA. For studies comparing methods
for detrending multifractal analysis (multifractal DFA (MF-DFA) and
wavelet transform modulus maxima (WTMM) method), see [32, 71, 72].

6.1 The Structure Function Approach and Singularity

Spectra
In the general multifractal formalism, one considers a normalised measure
µ(t), t ∈ [0,1], and defines the box probabilities ˜ in
neighbourhoods of (scale) length 1 around t. The multifractal approach
is then introduced by the partition function

for , (25)
where τ(q) is the Renyi scaling exponent and q is a real parameter that can
take positive as well as negative moments. Note that τ(q) is sometimes
defined with opposite sign (see, e. g., [6]). A record is called monofractal (or
self-affine), when the Renyi scaling exponent τ(q) depends linearly on q;
otherwise it is called multifractal. The generalised multifractal dimensions
D(q) (see also Section 3.4) are related to τ(q) by D(q) = τ(q)/q − 1, such that
the fractal dimension of the support is D(0) = −τ(0) and the correlation
dimension is D(2) = τ(2).
In time series, a discrete version has to be used, and the considered data
(xi), i = 1,...,N may usually include negative values. Hence, setting Ns =
int(N/s) and 1 we can define [6,
12],
Ns−1

X
Zq(s) = |X(ν,s)|q ∼ sτ(q) for s > 1. (26)
ν=0

Inserting the profile Y (j) and FFA(ν,s) from Eqs. (10) and (11), respectively,
we obtain
Ns−1 Ns−1
X o
n 2 q/2 X q/2

Zq(s) = [Y ((ν + 1)s) − Y (νs)] = FFA (ν,s). (27)

29
 ν=0 ν=0

Comparing Eqs. (27) with (13), we see that this multifractal approach can
be considered as a generalised version of the Fluctuation Analysis (FA)
method, where the exponent 2 is replaced by q. In particular we find
(disregarding the summation over the second partition of the time series)

We thus see that all methods for (mono-)fractal time analysis (discussed in
Chapters 4 and 5) in fact study the correlation dimension D(2) = 2α − 1 = β =
1 − γ (see Eq. (15)).
It is straightforward to define a generalised (multifractal) Hurst
exponent h(q) for the scaling behaviour of the qth moments of the
fluctuations [64, 65],

with h(2) = α ≈ H. In the following, we will use only h(2) for the standard
fluctuation exponent (denoted by α in the previous chapters), and reserve
the letter α for the Ho¨lder exponent.
Another way to characterise a multifractal series is the singularity
spectrum f(α), that is related to τ(q) via a Legendre transform [6, 12],

) and f(α) = qα − τ(q). (30)

Here, α is the singularity strength or H¨older exponent (see also articles ... in
the encyclopedia), while f(α) denotes the dimension of the subset of the
series that is characterised by α. Note that α is not the fluctuation scaling
exponent in this section, although the same letter is traditionally used for
both. Using Eq. (29), we can directly relate α and f(α) to h(q),

α = h(q) + qh0(q) and f(α) = q[α − h(q)] + 1. (31)

6.2 Wavelet Transform Modulus Maxima (WTMM) Method
The wavelet transform modulus maxima (WTMM) method [66, 67, 68, 69,
70] is a well-known method to investigate the multifractal scaling
properties of fractal and self-affine objects in the presence of non-
stationarities. For applications, see e. g. [73, 74]. It is based upon the

30
wavelet transform with continuous basis functions as defined in Section 5.1,
Eq. (16). Note that in this case the series ˜xi are analysed directly instead of
the profile Y (j) defined in Eq. (10). Using wavelets orthogonal to mth order
polynomials, the corresponding trends are elliminated.
Instead of averaging over all wavelet coefficients Lψ(τ,s), one averages,
within the modulo-maxima method, only the local maxima of |Lψ(τ,s)|. First,
one determines for a given scale s, the positions τj of the local maxima of |
W(τ,s)| as function of τ, so that |Lψ(τj − 1,s)| < |Lψ(τj,s)| ≥ |Lψ(τj + 1,s)| for j =
1,...,jmax. This maxima procedure is demonstrated in Fig. 7. Then one sums up
the qth power of the maxima,
jmax

X
Z(q,s) = |Lψ(τj,s)|q. (32)
j=1

The reason for the maxima procedure is that the absolute wavelet
coefficients |Lψ(τ,s)| can become arbitrarily small. The analysing wavelet
ψ(x) must always have positive values for some x and negative values for
other x, since it has to be orthogonal to possible constant trends. Hence
there are always positive and negative terms in the sum (16), and these
terms might cancel. If that happens, |Lψ(τ,s)| can become close to zero. Since
such small terms would spoil the calculation of negative moments in Eq.
(32), they have to be eliminated by the maxima procedure.
In fluctuation analysis, on the other hand, the calculation of the
variances F2(ν,s), e. g. in Eq. (11), involves only positive terms under the
summation. The variances cannot become arbitrarily small, and hence no
maximum procedure is required for series with compact support. In
addition, the variances will always increase if the segment length s is
increased, because the fit will always be worse for a longer segment. In the
WTMM method, in contrast, the absolute wavelet coefficients |Lψ(τ,s)| need
not increase with increasing scale s, even if only the local maxima are
considered. The values |Lψ(τ,s)| might become smaller for increasing s since
just more (positive and negative) terms are included in the summation (16),
and these might cancel even better. Thus, an additional supremum
procedure has been introduced in the WTMM method in order to keep the
dependence of Z(q,s) on s monotonous. If, for a given scale s, a maximum at a
certain position τj happens to be smaller than a maximum at τj0 ≈ τj for a
lower scale s0 < s, then Lψ(τj,s) is replaced by Lψ(τj0,s0) in Eq. (32).

31
 Often, scaling behaviour is observed for Z(q,s), and scaling exponents
τˆ(q) can be defined that describe how Z(q,s) scales with s,

Z(q,s) ∼ sτˆ(q). (33)

The exponents ˆτ(q) characterise the multifractal properties of the series

under investigation, and theoretically they are identical with the τ(q)
defined in Eq. (26) [66, 67, 68, 70] and related to h(q) by Eq. (29).

6.3 Multifractal Detrended Fluctuation Analysis (MFDFA)

The multifractal DFA (MF-DFA) procedure consists of five steps [32]. The
first three steps are essentially identical to the conventional DFA procedure
(see Section 5.3 and Fig. 4). Let us assume that (˜xi) is a series of length N,
and that this series is of compact support. The support can be defined as the
set of the indices j with nonzero values ˜xj, and it is compact if ˜xj = 0 for an
insignificant fraction of the series only. The value of ˜xj = 0 is interpreted as
having no value at this j. Note that we are not discussing the fractal or
multifractal features of the plot of the time series in a two-dimensional
graph (see also the discussion in Section 3.1), but analysing time series as
one-dimensional structures with values assigned to each point. Since real
time series always have finite length N, we explicitly want to determine the
multifractality of finite series, and we are not discussing the limit for N → ∞
here (see also Section 6.1).

• Step 1: Calculate the profile Y (j, Eq. (10), by integrating the time
series.

• Step 2: Divide the profile Y (j) into Ns = int(N/s) non-overlapping

segments of equal length s. Since the length N of the series is often not
a multiple of the considered time scale s, the same procedure can be
repeated starting from the opposite end. Thereby, 2Ns segments are
obtained altogether.

• Step 3: Calculate the local trend for each of the 2Ns segments by a
least-square fit of the profile. Then determine the variance by Eqs.
(20) and (21) for each segment ν = 0,...,2Ns − 1 Again, linear, quadratic,

32
 cubic, or higher order polynomials can be used in the fitting
procedure, and the corresponding methods are thus called MF-DFA1,
MF-DFA2, MF-DFA3, ...) [32]. In (MF-)DFAm [mth order (MF-)DFA]
trends of order mp in the profile (or, equivalently, of order m − 1 in
the original series) are eliminated. Thus a comparison of the results
for different orders of DFA allows one to estimate the type of the
polynomial trend in the time series [34, 35].

• Step 4: Average over all segments to obtain the qth order fluctuation
function

, (34)
This is the generalization of Eq. (13) suggested by the relations
derived in Section 6.1. For q = 2, the standard DFA procedure is
retrieved. One is interested in how the generalised q dependent
fluctuation functions Fq(s) depend on the time scale s for different
values of q. Hence, we must repeat steps 2 to 4 for several time scales
s. It is apparent that Fq(s) will increase with increasing s. Of course,
Fq(s) depends on the order m. By construction, Fq(s) is only defined for
s ≥ m + 2.

• Step 5: Determine the scaling behaviour of the fluctuation functions by

analysing log-log plots Fq(s) versus s for each value of q. If the series x˜i
are long-range power-law correlated, Fq(s) increases, for large values
of s, as a power-law,

Fq(s) ∼ sh(q) with . (35)

For very large scales, s > N/4, Fq(s) becomes statistically unreliable
because the number of segments Ns for the averaging procedure in step 4
becomes very small. Thus, scales s > N/4 should be excluded from the fitting
procedure determining h(q). Besides that, systematic deviations from the
scaling behaviour in Eq. (35), which can be corrected, occur for small scales
s ≈ 10.

33
 The value of h(0), which corresponds to the limit h(q) for q → 0, cannot
be determined directly using the averaging procedure in Eq. (34) because of
the diverging exponent. Instead, a logarithmic averaging procedure has to
be employed,

. (36)
Note that h(0) cannot be defined for time series with fractal support, where
h(q) diverges for q → 0.
For monofractal time series with compact support, h(q) is independent
of q, since the scaling behaviour of the variances ) is identical for
all segments ν, and the averaging procedure in Eq. (34) will give just this
identical scaling behaviour for all values of q. Only if small and large
fluctuations scale differently, there will be a significant dependence of h(q)
on q: If we consider positive values of q, the segments ν with large variance
F2(ν,s) (i. e. large deviations from the corresponding fit) will dominate the
average Fq(s). Thus, for positive values of q, h(q) describes the scaling
behaviour of the segments with large fluctuations. On the contrary, for
negative values of q, the segments ν with small variance ) will
dominate the average Fq(s). Hence, for negative values of q, h(q) describes
the scaling behaviour of the segments with small fluctuations. Figure 8
shows typical results obtained for Fq(s) in the MF-DFA procedure.
Usually the large fluctuations are characterised by a smaller scaling
exponent h(q) for multifractal series than the small fluctuations. This can be
understood from the following arguments: For the maximum scale s = N the
fluctuation function Fq(s) is independent of q, since the sum in Eq. (34) runs
over only two identical segments. For smaller scales the averaging
procedure runs over several segments, and the average value Fq(s) will be
dominated by the F2(ν,s) from the segments with small (large) fluctuations if
q < 0 (q > 0). Thus, for ) with q < 0 will be smaller than
Fq(s) with q > 0, while both become equal for s = N. Hence, if we assume an
homogeneous scaling behaviour of Fq(s) following Eq. (35), the slope h(q) in
a log-log plot of Fq(s) with q < 0 versus s must be larger than the
corresponding slope for Fq(s) with q > 0. Thus, h(q) for q < 0 will usually be
larger than h(q) for q > 0.
However, the MF-DFA method can only determine positive generalised
Hurst exponents h(q), and it already becomes inaccurate for strongly

34
anticorrelated signals when h(q) is close to zero. In such cases, a modified
(MF)DFA technique has to be used. The most simple way to analyse such
data is to integrate the time series before the MF-DFA procedure. Following
the MF-DFA procedure as described above, we obtain a generalised
˜
fluctuation functions described by a scaling law with h (q) = h(q) + 1. The
scaling behaviour can thus be accurately determined even for h(q) which
are smaller than zero for some values of q.
The accuracy of h(q) determined by MF-DFA certainly depends on the
length N of the data. For q = ±10 and data with N = 10 000 and 100 000,
systematic and statistical error bars (standard deviations) up to ∆h(q) ≈
±0.1 and ≈ ±0.05 should be expected, respectively [32]. A difference of
h(−10) − h(+10) = 0.2, corresponding to an even larger width ∆α of the
singularity spectrum f(α) defined in Eq. (30) is thus not significant unless
the data was longer than N = 10 000 points. Hence, one has to be very
careful when concluding multifractal properties from differences in h(q).
As already mentioned in the introduction, two types of multifractality in
time series can be distinguished. Both of them require a multitude of scaling
exponents for small and large fluctuations: (i) Multifractality of a time series
can be due to a broad probability density function for the values of the time
series, and (ii) multifractality can also be due to different longrange
correlations for small and large fluctuations. The most easy way to
distinguish between these two types is by analysing also the corresponding
randomly shuffled series [32]. In the shuffling procedure the values are put
into random order, and thus all correlations are destroyed. Hence the
shuffled series from multifractals of type (ii) will exhibit simple random
behaviour, hshuf(q) = 0.5, i. e. non-multifractal scaling. For multifractals of
type (i), on the contrary, the original h(q) dependence is not changed, h(q) =
hshuf(q), since the multifractality is due to the probability density, which is
not affected by the shuffling procedure. If both kinds of multifractality are
present in a given series, the shuffled series will show weaker
multifractality than the original one.

6.4 Comparison of WTMM and MF-DFA

The MF-DFA results turn out to be slightly more reliable than the WTMM
results [32, 71, 72]. In particular, the MF-DFA has slight advantages for
negative q values and short series. In the other cases the results of the two

35
methods are rather equivalent. Besides that, the main advantage of the
MFDFA method compared with the WTMM method lies in the simplicity of
the MF-DFA method. However, contrary to WTMM, MF-DFA is restricted to
studies of data with full one-dimensional support, while WTMM is not. Both,
WTMM and MF-DFA have been generalised for higher dimensional data, see
[33] for higher dimensional MF-DFA and, e. g., [70] for higher dimensional
WTMM. Studies of other generalisations of detrending methods like the
discrete WT approach (see Section 5.2) and the CMA method (see Section
5.7) are currently under investigation [75].

7 Statistics of Extreme Events in Fractal Time

Series
The statistics of return intervals between well defined extremal events is a
powerful tool to characterise the temporal scaling properties of observed
time series and to derive quantities for the estimation of the risk for
hazardous events like floods, very high temperatures, or earthquakes. It was
shown recently that long-term correlations represent a natural mechanism
for the clustering of the hazardous events [76]. In this chapter we will
discuss the most important consequences of long-term correlations and
fractal scaling of time series upon the statistics of extreme events [76, 77,
78, 79, 80]. Corresponding work regarding multifractal data [81] is not
discussed here.

7.1 Return Intervals Between Extreme Events

To study the statistics of return intervals we consider again a time series
(xi), i = 1,...,N with fractal scaling behaviour, sampled homogeneously and
normalised to zero mean and unit variance. For describing the reoccurrence
of rare events exceeding a certain threshold q, we investigate the return
intervals rq between these events, see Fig. 9. The average return interval Rq =
hrqi increases as a function of the threshold q (see, e. g. [82]). It is known
that for uncorrelated records (’white noise’), the return intervals are also
uncorrelated and distributed according to the Poisson distribution,
). For fractal (long-term correlated) data with auto-
correlations following Eq. (5), we obtain a stretched exponential [76, 77, 78,
79, 83],

36
 , (37)
This behaviour is shown in Fig. 10. The exponent γ is the correlation
exponent from C(s), and the parameters aγ and bγ are independent of q. They
R
can be determined from the normalization conditions for Pq(r), i. e., Pq(r)dr
R
= 1 and rPq(r)dr = Rq. The form of the distribution (37) indicates that
return intervals both well below and well above their average value Rq
(which is independent of γ) are considerably more frequent for long-term
correlated than for uncorrelated data. It has to be noted that there are
deviations from the stretched exponential law (37) for very small r
(discretization effects and an additional power-law regime) and for very
large r (finite size effects), see Fig. 10. The extent of the deviations from Eq.
(37) depends on the distribution of the values xi of the time series. For a
discussion of these effects, see [79].

Equation (37) does not quantify, however, if the return intervals

themselves are arranged in a correlated or in an uncorrelated fashion, and if
clustering of rare events may be induced by long-term correlations. To
study this question, one has to evaluate the auto-correlation function Cr(s) =
of the return intervals. The results for model data
suggests that also the return intervals are long-term power-law correlated,
with the same exponent γ as the original record. Accordingly, large and
small return intervals are not arranged in a random fashion but are
expected to form clusters. As a consequence, the probability of finding a
certain return interval r depends on the value of the preceding interval r0,
and this effect has to be taken into account in predictions and risk
estimations [76, 79].
The conditional distribution function Pq(r|r0) is a basic quantity, from
which the relevant quantities in risk estimations can be derived [82]. For
example, the first moment of Pq(r|r0) is the average value Rq(r0) of those
return intervals that directly follow r0. By definition, Rq(r0) is the expected
waiting time to the next event, when the two events before were separated
by r0. The more general quantity is the expected waiting time τq(x|r0) to the
next event, when the time x has been elapsed. For x = 0, τq(0|r0) is identical
to Rq(r0). In general, τq(x|r0) is related to Pq(r|r0) by

(38)

37
 For uncorrelated records, τq(x|r0)/Rq = 1 (except for discreteness effects
that lead to τq(x|r0)/Rq > 1 for x > 0, see [84]). Due to the scaling of Pq(r|r0),
also τq(x|r0)/Rq scales with r0/Rq and x/Rq. Small and large return intervals
are more likely to be followed by small and large ones, respectively, and
hence τq(0|r0)/Rq = Rq(r0)/Rq is well below (above) one for r0/Rq well below
(above) one. With increasing x, the expected residual time to the next event
increases. Note that only for an infinite long-term correlated record, the
value of τq(x|r0) will increase indefinitely with x and r0. For real (finite)
records, there exists a maximum return interval which limits the values of x,
r0 and τq(x|r0).

7.2 Distribution of Extreme Events

In this section we describe how the presence of fractal long-term
correlations affects the statistics of the extreme events, i. e., maxima within
time segments of fixed duration R, see Fig. 11 for illustration. By definition,
extreme events are rare occurrences of extraordinary nature, such as, e. g.
floods, very high temperatures, or earthquakes. In hydrological engineering
such conventional extreme value statistics are commonly applied to decide
what building projects are required to protect riverside areas against
typical floods that occur for example once in 100 years. Most of these
results are based on statistically independent values xi and hold only in the
limit R → ∞. However, both of these assumptions are not strictly fulfilled for
correlated fractal scaling data.
In classical extreme value statistics one assumes that records (xi) consist
of i.i.d. data, described by density distributions P(x), which can be, e. g., a
Gaussian or an exponential distribution. One is interested in the distribution
density function PR(m) of the maxima (mj) determined in segments of length
R in the original series (xi), see Fig. 11. Note that all maxima are also
elements of the original data. The corresponding integrated maxima
distribution GR(m) is defined as
Zm
GR(m) = 1 − ER(m) = PR(m0) dm0. (39)
−∞

Since GR(m) is the probability of finding a maximum smaller than m, ER(m)

denotes the probability of finding a maximum that exceeds m. One of the
main results of traditional extreme value statistics states that for

38
independently and identically distributed (i.i.d.) data (xi) with Gaussian or
exponential distribution density function P(x) the integrated distribution
GR(m) converges to a double exponential (Fisher-Tippet-Gumbel)
distribution (often labelled as Type I) [85, 86, 87, 88, 89], i. e.,

(40)
for R → ∞, where α is the scale parameter and u the location parameter. By
the method of moments those parameters are given by and u =
mR − neα with the Euler constant ne = 0.577216 [88, 90, 91, 92]. Here mR and
σR denote the (R-dependent) mean maximum and the standard deviation,
respectively. Note that different asymptotics will be reached for broader
distributions of data (xi) that belong to other domains of attraction [88]. For
example, for data following a power-law distribution (or Pareto
distribution), P(x) = (x/x0)−k, GR(m) converges to a Fr´echet distribution,
often labelled as Type II. For data following a distribution with finite upper
endpoint, for example the uniform distribution P(x) = 1 for 0 ≤ x ≤ 1 , GR(m)
converges to a Weibull distribution, often labelled as Type III. We do not
consider the latter two types of asymptotics here.
Numerical studies of fractal model data have recently shown that the
distribution P(x) of the original data has a much stronger effect upon the
convergence towards the Gumbel distribution than the long-term
correlations in the data. Long-term correlations just slightly delay the
convergence of GR(m) towards the Gumbel distribution (40). This can be
observed very clearly in a plot of the integrated and scaled distribution
GR(m) on logarithmic scale [80].
Furthermore, it was found numerically that (i) the maxima series (mj)
exhibit long-term correlations similar to those of the original data (xi), and
most notably (ii) the maxima distribution as well as the mean maxima
significantly depend on the history, in particular on the previous maximum
[80]. The last item implies that conditional mean maxima and conditional
maxima distributions should be considered for improved extreme event
predictions.

39
8 Simple Models for Fractal and Multifractal Time
Series
8.1 Fourier Filtering
Fractal scaling with long-term correlations can be introduced most easily
into time series by the Fourier-filtering technique, see, e. g., [93, 94, 95]. The
Fourier filtering technique is not limited to the generation of long-term
correlated data characterised by a power-law auto-correlation function C(s)
∼ x−γ with 0 < γ < 1. All values of the scaling exponents α = h(2) ≈ H or β = 2α
− 1 can be obtained, even those that cannot be found directly by the fractal
analysis techniques described in Chapters 4 and 5 (e. g. α < 1). Note,
however, that Fourier filtering will always yield Gaussian distributed data
values and that no nonlinear or multifractal properties can be achieved (see
also Sections 3.4, 5.5, and Chapter 6). In Section 5.4, we have briefly
described a modification of Fourier filtering for obtaining reliable short-
term correlated data.
For the generation of data characterised by fractal scaling with β = 2α−1
[93, 94] we start with uncorrelated Gaussian distributed random numbers
xi from an i.i.d. generator. Transforming a series of such numbers into
frequency space with discrete Fourier transform or FFT (for suitable series
lengths N) yields a flat power spectrum, since random numbers correspond
to white noise. Multiplying the (complex) Fourier coefficients by f−β/2, where
f ∝ 1/s is the frequency, will rescale the power spectrum S(f) to follow Eq.
(6), as expected for time series with fractal scaling. After transforming back
to the time domain (using inverse Fourier transform or inverse FFT) we will
thus obtain the desired long-term correlated data ˜xi. The final step is the
normalization of this data.
The Fourier filtering method can be improved using modified Bessel
functions instead of the simple factors f−β/2 in modifying the Fourier
coefficients [95]. This way problems with the divergence of the
autocorrelation function C(s) at s = 0 can be avoided.
An alternative method to the Fourier filtering technique, the random
midpoint displacement method, is based on the construction of self-affine
surfaces by an iterative procedure, see, e. g. [6]. Starting with one interval
with constant values, the intervals are iterative split in the middle and the
midpoint is displaced by a random offset. The amplitude of this offset is

40
scaled according to the length of the interval. Since the method generates a
self-affine surface xi characterised by a Hurst exponent H, the differentiated
series ∆xi can be used as long-term correlated or anti-correlated random
numbers. Note, however, that the correlations do not persist for the whole
length of the data generated this way. Another option is the use of wavelet
synthesis, the reverse of wavelet analysis described in Section 5.1. In that
method, the scaling law is introduced by setting the magnitudes of the
wavelet coefficients according to the corresponding time scale s.

8.2 The Schmitz-Schreiber Method

When long-term correlations in random numbers are introduced by the
Fourierfiltering technique (see previous section), the original distribution
P(x) of the time series values xi is always modified such that it becomes closer
to a Gaussian. Hence, no series (xi) with broad distributions of the values and
fractal scaling can be generated. In these cases an iterative algorithm
introduced by Schreiber and Schmitz [97, 98] must be applied.
The algorithm consists of the following steps: First one creates a
Gaussian distributed long-term correlated data set with the desired
correlation exponent γ by standard Fourier-filtering [95]. The power
spectrum SG(f) =
) of this data set is considered as reference spectrum (where f
denotes the frequency in Fourier space and the FG(f) are the complex
Fourier coefficients). Next one creates an uncorrelated sequence of random
numbers (xrefi ), following a desired distribution P(x). The (complex) Fourier
transform F(f) of the (xrefi ) is now divided by its absolute value and
multiplied by the square root of the reference spectrum,

. (41)

After the Fourier back-transformation of Fnew(f), the new sequence (xnewi )

has the desired correlations (i. e. the desired γ), but the shape of the
distribution has changed towards a (more or less) Gaussian distribution. In
order to enforce the desired distribution, we exchange the (xnewi ) by the (xrefi
), such that the largest value of the new set is replaced by the largest value
of the reference set, the second largest of the new set by the second largest
of the reference set and so on. After this the new sequence has the desired

41
distribution and is clearly correlated. However, due to the exchange
algorithm the perfect long-term correlations of the new data sequence were
slightly altered again. So the procedure is repeated: the new sequence is
Fourier transformed followed by spectrum adjustment, and the exchange
algorithm is applied to the Fourier back-transformed data set. These steps
are repeated several times, until the desired quality (or the best possible
quality) of the spectrum of the new data series is achieved.

8.3 The Extended Binomial Multifractal Model

The multifractal cascade model [6, 64, 32] is a standard model for
multifractal data, which is often applied, e. g., in hydrology [96]. In the
model, a record xi of length N = 2nmax is constructed recursively as follows. In
generation n = 0, the record elements are constant, i. e. xi = 1 for all i = 1,...,N.
In the first step of the cascade (generation n = 1), the first half of the series
is multiplied by a factor a and the second half of the series is multiplied by a
factor b. This yields xi = a for i = 1,...,N/2 and xi = b for i = N/2 + 1,...,N. The
parameters a and b are between zero and one, 0 < a < b < 1. One need not
restrict the model to b = 1 − a as is often done in the literature [6]. In the
second step (generation n = 2), we apply the process of step
1 to the two subseries, yielding xi = a2 for i = 1,...,N/4, xi = ab for i =
N/4+1,...,N/2, xi = ba = ab for i = N/2+1,...,3N/4, and xi = b2 for i = 3N/4+1,...,N.
In general, in step n+1, each subseries of step n is divided into two subseries
of equal length, and the first half of the xi is multiplied by a while the second
half is multiplied by b. For example, in generation n = 3 the values in the
eight subseries are a3, a2b, a2b, ab2, a2b, ab2, ab2, b3. After nmax steps, the final
generation has been reached, where all subseries have length 1 and no more
splitting is possible. We note that the final record can be written as xi =
anmax−n(i−1)bn(i−1), where n(i) is the number of digits 1 in the binary
representation of the index i, e. g. n(13) = 3, since 13 corresponds to binary
1101.
For this multiplicative cascade model, the formula for τ(q) has been
derived earlier [6, 64, 32]. The result is τ(q) = [−ln(aq + bq) + q ln(a + b)]/ln2
or

. (42)

42
It is easy to see that h(1) = 1 for all values of a and b. Thus, in this form the
model is limited to cases where h(1), which is the exponent Hurst defined
originally in the R/S method, is equal to one.
In order to generalise this multifractal cascade process such that any
value of h(1) is possible, one can subtract the offset ∆h = ln(a+b)/ln(2) from
h(q) [99]. The constant offset ∆h corresponds to additional long-term
correlations incorporated in the multiplicative cascade model. For
generating records without this offset, we rescale the power spectrum.
First, we fast-Fourier transform (FFT) the simple multiplicative cascade
data into the frequency domain. Then, we multiply all Fourier coefficients
by f−∆h, where f is the frequency. This way, the slope β of the power spectra
S(f) ∼ f−β is decreased from β = 2h(2)−1 = [2ln(a+b)−ln(a2+b2)]/ln2 into β0 =
2[h(2)−∆h]−1 = −ln(a2+b2)/ln2. Finally, backward FFT is employed to
transform the signal back into the time domain.

8.4 The Bi-fractal Model

In some cases a simple bi-fractal model is already sufficient for modelling
apparently multifractal data [100]. For bi-fractal records the Renyi
exponents τ(q) are characterised by two distinct slopes α1 and α2,
τ(q) = ( qα1 − 1 q≤ (43)
q×
q > q×
or qα 2 + q×(α 1 − α 2) − 1
( qα1 + q×(α2 − α1) − 1 qα2 q ≤ q×
τ(q) = (44)
− 1 q > q× .
If this behaviour is translated into the h(q) picture using Eq. (29), we obtain
that h(q) exhibits a plateau from q = −∞ up to a certain q× and decays
hyperbolically for q > q×,

, (45)

or vice versa,

(46)

43
Both versions of this bi-fractal model require three parameters. The
multifractal spectrum is degenerated to two single points, thus its width can
be defined as ∆α = α1 − α2.

9 Future Directions
The most straightforward future direction is to analyse more types of time
series from other complex systems than those listed in Chapter 2 to check
for the presence of fractal scaling and in particular long-term correlations.
Such applications may include (i) data that are not recorded as function of
time but as function of another parameter and (ii) higher dimensional data.
In particular, the inter-relationship between fractal time series and spatially
fractal structures can be studied. Studies of fields with fractal scaling in time
and space have already been performed in Geophysics. In some cases
studying new types of data will require dealing with more difficult types of
non-stationarities and transient behaviour, making further development of
the methods necessary. In many studies, detrending methods have not been
applied yet. However, discovering fractal scaling in more and more systems
cannot be an aim on its own.
Up to now, the reasons for observed fractal or multifractal scaling are
not clear in most applications. It is thus highly desirable to study causes for
fractal and multifractal correlations in time series, which is a difficult task,
of course. One approach might be based on modelling and comparing the
fractal aspects of real and modelled time series by applying the methods
described in this article. The fractal or multifractal characterisation can thus
be helpful in improving the models. For many applications, practically
usable models which display fractal or transient fractal scaling still have to
be developed. One example for a model explaining fractal scaling might be a
precipitation, storage, and runoff model, in which the fractal scaling of
runoff time series could be explained by fractional integration of rainfall in
soil, groundwater reservoirs, or river networks characterised by a fractal
structure. Also studies regarding the inter-relation between fractal scaling
and complex networks, representing the structure of a complex system, are
desirable. This way one could gain an interpretation of the causes for fractal
behaviour.
Another direction of future research is regarding the linear and
especially non-linear inter-relations between several time series. There is
great need for improved methods characterising cross-correlations and

44
similar statistical inter-relations between several non-stationary time
series. Most methods available so far are reserved to stationary data, which
is, however, hardly found in natural recordings. See [101] for a very recent
approach analysing fractal cross-correlations in non-stationary data. An
even more ambitious aim is the (time-dependent) characterisation of a
larger network of signals. In such a network, the signals themselves would
represent the nodes, while the (possibly directed) inter-relations between
each pair represent the links (or bonds) between the nodes. The properties
of both, nodes and links can vary with time or change abruptly, when the
represented complex system goes through a phase transition.
Finally, more work will have to be invested in studying the practical
consequences of fractal scaling in time series. Studies should particularly
focus on predictions of future values and behaviour of time series and
whole complex systems. This is very relevant, not only in hydrology and
climate research, where a clear distinguishing of trends and natural
fluctuations is crucial, but also for predicting dangerous medical events on-
line in patients based on the continuous recording of time series.

Acknowledgement: We thank Ronny Bartsch, Amir Bashan, Mikhail

Bogachev, Armin Bunde, Jan Eichner, Shlomo Havlin, Diego Rybski, Aicko
Schumann, and Stephan Zschiegner for helpful discussions. This work has
been supported by the Deutsche Forschungsgemeinschaft (grant KA
1676/3) and the European Union (STREP project DAPHNet, grant 018474-
2).

10 Bibliography

References
[1] B.B. Mandelbrot, J.W. van Ness, Fractional Brownian motions, fractional
noises and applications, SIAM Review 10, 422 (1968).

[2] B.B. Mandelbrot, J.R. Wallis, Some long-run properties of geophysical

records, Water Resour. Res. 5, 321-340 (1969).

[3] Benoit B. Mandelbrot, Multifractals and 1/f noise: wild self-affinity in

physics (Springer, Berlin 1999).

45
[4] H.E. Hurst, Long-term storage capacity of reservoirs, Transactions of the
American Society of Civil Engineering 116, 770 (1951).

[5] H.E. Hurst, R.P. Black, Y.M. Simaika, Long-term storage: an experimental
study (Constable and Co., Ltd, London, 1965).

[6] J. Feder, Fractals (Plenum Press, New York, 1988).

[7] M.F. Barnsley, Fractals everywhere (Academic Press, 1993).

[8] Fractals in science, edited by A. Bunde, S. Havlin (Springer, Berlin 1994).

[9] P.E.T. Jorgenssen, Analysis and probability: wavelets, signals, fractals
(Springer, Berlin 2000).

[10] The science of disasters – climate disruptions, heart attacks, and

market crashes, edited by A. Bunde, J. Kropp, H.-J. Schellnhuber (Springer,
Berlin 2002).

[11] H. Kantz, T. Schreiber, Nonlinear time series analysis (Cambridge

University Press, 2003).

[12] H.-O. Peitgen, H. J”urgens, D. Saupe, Chaos and fractals (Springer,

Berlin, 2004)

[13] D. Sornette, Critical phenomena in natural sciences (Springer, Berlin,

2004).

[14] C.-K. Peng, J. Mietus, J.M. Hausdorff, S. Havlin, H.E. Stanley, A.L.
Goldberger, Long-range anti-correlations and non-Gaussian behaviour of
the heartbeat, Phys. Rev. Lett. 70, 1343 (1993).

[15] A. Bunde, S. Havlin, J. W. Kantelhardt, T. Penzel, J.-H. Peter, K. Voigt,

Correlated and uncorrelated regions in heart-rate fluctuations during
sleep, Phys. Rev. Lett. 85, 3736 (2000).

[16] D. Vyushin, I. Zhidkov, S. Havlin, A. Bunde, S. Brenner, Volcanic

forcing improves atmosphere-ocean coupled general circulation model
scaling performance, Geophys. Res. Lett. 31, L10206 (2004).

[17] E. Koscielny-Bunde, A. Bunde, S. Havlin, H.E. Roman, Y. Goldreich, H.J.

Schellnhuber, Indication of a universal persistence law governing
atmospheric variability, Phys. Rev. Lett. 81, 729 (1998).

46
[18] G.E.P. Box, G.M. Jenkins, G.C. Reinsel, Time-series Analysis (Prentice
Hall, New Jersey 1994).

[19] C. Chatfield, The analysis of time series. An introduction (Taylor &

Francis Ltd., 2003).

[20] D.T. Schmitt, M. Schulz, Analyzing memory effects of complex systems

from time series, Phys. Rev. E 73, 056204 (2006).

[21] M.S. Taqqu, V. Teverovsky, W. Willinger, Estimators for long-range

dependence: An empirical study, Fractals 3, 785 (1995).

[22] D. Delignieresa, S. Ramdania, L. Lemoinea, K. Torrea, M. Fortesb, G.

Ninot, Fractal analyses for ’short’ time series: A re-assessment of classical
methods, J. Math. Psychol. 50, 525 (2006).

[23] J. Mielniczuk, P. Wojdyllo, Estimation of Hurst exponent revisited,

Comp. Stat. Data. Anal. 51, 4510 (2007).

[24] G.A. Hunt, Random Fourier transforms, Trans. Amer. Math. Soc. 71, 38
(1951).

[25] G. Rangarajan, M. Ding, Integrated approach to the assessment of long

range correlation in time series data, Phys. Rev. E 61, 4991 (2000).

[26] C.-K. Peng, S.V. Buldyrev, A.L. Goldberger, S. Havlin, F. Sciortino, M.

Simons, H.E. Stanley, Long-range correlations in nucleotide sequences,
Nature 356, 168 (1992).

[27] P. Goupillaud, A. Grossmann, J. Morlet, Cycle-octave and related

transforms in seismic signal analysis, Geoexploration 23, 85 (1984).

[28] I. Daubechies, Orthogonal bases of compactly supported wavelets,

Commun. Pure Appl. Math. 41, 909 (1988).

[29] J. W. Kantelhardt, H. E. Roman, M. Greiner, Discrete wavelet approach

to multifractality, Physica A 220, 219 (1995).

[30] C.-K. Peng, S.V. Buldyrev, S. Havlin, M. Simons, H.E. Stanley, A.L.
Goldberger, Mosaic organization of DNA nucleotides, Phys. Rev. E 49,
1685 (1994).

47
[31] Y. Ashkenazy, P. Ch. Ivanov, S. Havlin, C.-K. Peng, A. L. Goldberger, H.
E. Stanley, Magnitude and sign correlations in heartbeat fluctuations, Phys.
Rev. Lett. 86, 1900 (2001).

[32] J.W. Kantelhardt, S.A. Zschiegner, A. Bunde, S. Havlin, E.

KoscielnyBunde, H.E. Stanley, Multifractal detrended fluctuation analysis
of nonstationary time series, Physica A 316, 87 (2002).

[33] G.-F. Gu, W.-X. Zhou, Detrended fluctuation analysis for fractals and
multifractals in higher dimensions, Phys. Rev. E 74, 061104 (2006)

[34] J.W. Kantelhardt, E. Koscielny-Bunde, H.H.A. Rego, S. Havlin, A.

Bunde, Detecting long-range correlations with detrended fluctuation
analysis, Physica A 295, 441 (2001).

[35] K. Hu, P.Ch. Ivanov, Z. Chen, P. Carpena, H.E. Stanley, Effect of trends
on detrended fluctuation analysis, Phys. Rev. E 64, 011114 (2001).

[36] Z. Chen, P.Ch. Ivanov, K. Hu, H.E. Stanley, Effect of non-stationarities

on detrended fluctuation analysis, Phys. Rev. E 65, 041107 (2002).

[37] Z. Chen, K. Hu, P. Carpena, P. Bernaola-Galvan, H.E. Stanley, P.Ch.

Ivanov, Effect of nonlinear filters on detrended fluctuation analysis, Phys.
Rev. E 71, 011104 (2005).

[38] P. Grau-Carles, Bootstrap testing for detrended fluctuation analysis,

Physics A 360, 89 (2006).

[39] R. Nagarajan, Effect of coarse-graining on detrended fluctuation

analysis, Physica A 363, 226 (2006).

[40] C. Heneghan, G. McDarby, Establishing the relation between

detrended fluctuation analysis and power spectral density analysis for
stochastic processes, Phys. Rev. E 62, 6103 (2000).

[41] R. Weron, Estimating long-range dependence: finite sample properties

and confidence intervals, Physica A 312, 285 (2002).

[42] A. Bashan, R. Bartsch, J.W. Kantelhardt, S. Havlin, Comparison of

detrending methods for fluctuation analysis, Physica A (in press, 2008).

48
[43] S. Bahar, J.W. Kantelhardt, A. Neiman, H.H.A. Rego, D.F. Russell, L.
Wilkens, A. Bunde, F. Moss, Long range temporal anti-correlations in
paddlefish electro-receptors, Europhys. Lett. 56, 454 (2001).

[44] R. Bartsch, T. Henning, A. Heinen, S. Heinrichs, P. Maass, Statistical

analysis of fluctuations in the ECG morphology, Physica A 354, 415
(2005).

[45] M.S. Santhanam, J.N. Bandyopadhyay, and D. Angom, Quantum

spectrum as a time series: fluctuation measures, Phys. Rev. E 73, 015201
(2006).

[46] Y. Ashkenazy, S. Havlin, P.Ch. Ivanov, C.-K. Peng, V. SchulteFrohlinde,

H.E. Stanley, Magnitude and sign scaling in power-law correlated time
series, Physica A 323, 19 (2003).

[47] T. Kalisky, Y. Ashkenazy, S. Havlin, Volatility of linear and nonlinear

time series, Phys. Rev. E 72, 011913 (2005).

[48] R.N. Mantegna, H.E. Stanley, An introduction to econophysics –

correlations and complexity in finance (Cambridge Univ. Press, 2000).

[49] J.-P. Bouchaud, M. Potters, Theory of financial risks: from statistical

physics to risk management (Cambridge Univ. Press, 2003).

[50] E. Alessio, A. Carbone, G. Castelli, V. Frappietro, Second-order moving

average and scaling of stochastic time series, Europ. Phys. J. B 27, 197
(2002).

[51] A. Carbone, G. Castelli, H.E. Stanley, Analysis of clusters formed by the

moving average of a long-range correlated time series, Phys. Rev. E 69,
026105 (2004).

[52] A. Carbone, G. Castelli, H.E. Stanley, Time-dependent Hurst exponent

in financial time series, Physica A 344, 267 (2004).

[53] J. Alvarez-Ramirez, E. Rodriguez, J.C. Echeverr´ıa, Detrending

fluctuation analysis based on moving average filtering, Physica A 354, 199
(2005).

49
[54] K. Kiyono, Z.R. Struzik, N. Aoyagi, F. Togo, Y. Yamamoto, Phase
transition in a healthy human heart rate, Phys. Rev. Lett. 95, 058101
(2005).

[55] M. Staudacher, S. Telser, A. Amann, H. Hinterhuber, M. Ritsch-Marte,

A new method for change-point detection developed for on-line analysis of
the heart beat variability during sleep, Physica A 349, 582 (2005).

[56] S. Telser, M. Staudacher, B. Hennig, Y. Ploner, A. Amann, H.

Hinterhuber, M. Ritsch-Marte, Temporally resolved fluctuation analysis of
sleepECG, J. Biol. Phys. (Preprint, 2007).

[57] C.V. Chianca, A. Ticona, T.J.P. Penna, Fourier-detrended fluctuation

analysis, Physica A 357, 447 (2005).

[58] I.M. Ja´nosi, R. Mu¨ller, Empirical mode decomposition and correlation

properties of long daily ozone records, Phys. Rev. E 71, 056126 (2005).

[59] R. Nagarajan, R.G. Kavasseri, Minimizing the effect of trends on

detrended fluctuation analysis of long-range correlated noise, Physica A
354, 182 (2005).

[60] R. Nagarajan, Reliable scaling exponent estimation of long-range

correlated noise in the presence of random spikes, Physica A 366, 1 (2006).
[61] E. Rodriguez, J.C. Echeverria, J. Alvarez-Ramirez, Detrending
fluctuation analysis based on high-pass filtering, Physica A 375, 699
(2007).

[62] D. Grech and Z. Mazur, Statistical properties of old and new

techniques in detrended analysis of time series, Acta Phys. Pol. B 36, 2403
(2005).

[63] L. Xu, P.Ch. Ivanov, K.Hu, Z. Chen, A. Carbone, and H.E. Stanley,
Quantifying signals with power-law correlations: a comparative study of
detrended fluctuation analysis and detrended moving average techniques,
Phys. Rev. E 71, 051101 (2005).

[64] A.-L. Barab´asi, T. Vicsek, Multifractality of self-affine fractals, Phys.

Rev. A 44, 2730 (1991).

50
[65] E. Bacry, J. Delour, J.F. Muzy, Multifractal random walk , Phys. Rev. E
64, 026103 (2001).

[66] J.F. Muzy, E. Bacry, A. Arneodo, Wavelets and multifractal formalism

for singular signals: Application to turbulence data, Phys. Rev. Lett. 67,
3515 (1991).

[67] J.F. Muzy, E. Bacry, A. Arneodo, The multifractal formalism revisited

with wavelets, Int. J. Bifurcat. Chaos 4, 245 (1994).

[68] A. Arneodo, E. Bacry, P.V. Graves, J.F. Muzy, Characterizing longrange

correlations in DNA sequences from wavelet analysis, Phys. Rev. Lett. 74,
3293 (1995).

[69] A. Arneodo, S. Manneville, J.F. Muzy, Towards log-normal statistics in

high Reynolds number turbulence, Eur. Phys. J. B 1, 129 (1998).

[70] A. Arneodo, B. Audit, N. Decoster, J.F. Muzy, C. Vaillant, Wavelet based

multifractal formalism: applications to DNA sequences, satellite images of
the cloud structure, and stock market data, in: The science of disaster:
climate disruptions, market crashes, and heart attacks, edited by A. Bunde,
J. Kropp, H.-J. Schellnhuber (Springer, Berlin, 2002).

[71] J.W. Kantelhardt, D. Rybski, S.A. Zschiegner, P. Braun, E.

KoscielnyBunde, V. Livina, S. Havlin, A. Bunde, Multifractality of river
runoff and precipitation: comparison of fluctuation analysis and wavelet
methods, Physica A 330, 240 (2003).

[72] P. Oswiecimka, J. Kwapien, S. Drozdz, Wavelet versus detrended

fluctuation analysis of multifractal structures, Phys. Rev. E 74, 016103
(2006). [73] P.Ch. Ivanov, L.A.N. Amaral, A.L. Goldberger, S. Havlin, M.G.
Rosenblum, Z.R. Struzik, H.E. Stanley, Multifractality in human heartbeat
dynamics, Nature 399, 461 (1999).

[74] L.A.N. Amaral, P.Ch. Ivanov, N. Aoyagi, I. Hidaka, S. Tomono, A.L.

Goldberger, H.E. Stanley, Y. Yamamoto, Behavioral-independence
features of complex heartbeat dynamics, Phys. Rev. Lett. 86, 6026
(2001).

[75] M. Bogachev, A.Y. Schumann, J. W. Kantelhardt, in preparation

51
 (2008).

[76] A. Bunde, J.F. Eichner, J.W. Kantelhardt, S. Havlin, Long-term memory: A

natural mechanism for the clustering of extreme events and anomalous
residual times in climate records, Phys. Rev. Lett. 94, 048701 (2005).

[77] A. Bunde, J.F. Eichner, J.W. Kantelhardt, S. Havlin, The effect of longterm
correlations on the return periods of rare events, Physica A 330, 1
(2003).

[78] E.G. Altmann, H. Kantz, Recurrence time analysis, long-term correlations,

and extreme events, Phys. Rev. E 71, 056106 (2005).

[79] J.F. Eichner, J.W. Kantelhardt, A. Bunde, S. Havlin, Statistics of return

intervals in long-term correlated records, Phys. Rev. E 75, 011128
(2007).

[80] J.F. Eichner, J.W. Kantelhardt, A. Bunde, S. Havlin, Extreme value

statistics in records with long-term persistence, Phys. Rev. E 73, 016130
(2006).

[81] M.I. Bogachev, J.F. Eichner, A. Bunde, Effect of nonlinear correlations on

the statistics of return intervals in multifractal data sets, Phys. Rev. Lett.
99, 240601 (2007).

[82] H. v. Storch, F.W. Zwiers, Statistical analysis in climate research

(Cambridge Univ. Press, 2001).

[83] G.F. Newell, M. Rosenblatt, Ann. Math. Statist. 33, 1306 (1962).

[84] D. Sornette, L. Knopoff, The paradox of the expected time until the next
earthquake, Bull. Seism. Soc. Am. 87, 789 (1997).

[85] R.A. Fisher, L.H.C. Tippett, Limiting forms of the frequency distribution
of the largest or smallest member of a sample, Proc. Camb. Phil. Soc. 24,
180 (1928).

[86] E.J. Gumbel, Statistics of extremes (Columbia University Press, New

York, 1958).

52
[87] J. Galambos, The asymptotic theory of extreme order statistics, (John
Wiley and Sons, New York, 1978).

[88] M.R. Leadbetter, G. Lindgren, H. Rootzen, Extremes and related

properties of random sequences and processes (Springer, New York,
1983).

[89] Extreme value theory and applications, edited by J. Galambos, J. Lechner,

E. Simin (Kluwer, Dordrecht, 1994).

[90] V. te Chow, Handbook of applied hydrology (McGraw-Hill Book

Company, New York, 1964).

[91] A.J. Raudkivi, Hydrology (Pergamon Press, Oxford, 1979).

[92] P.F. Rasmussen, N. Gautam, Alternative PWM-estimators of the Gumbel

distribution, J. Hydrol. 280, 265 (2003).

[93] B.B. Mandelbrot, A fast fractional Gaussian noise generator, Water

Resour. Res. 7, 543 (1971).

[94] R.F. Voss, in: Fundamental algorithms in computer graphics edited by

R.A. Earnshaw (Springer, Berlin, 1985).

[95] H.A. Makse, S. Havlin, M. Schwartz, H.E. Stanley, Method for generating
long-range correlations for large systems, Phys. Rev. E 53, 5445 (1996).

[96] I. Rodriguez-Iturbe, A. Rinaldo, Fractal river basins – change and

selforganization (Cambridge Univ. Press, 1997).

[97] T. Schreiber, A. Schmitz, Improved surrogate data for nonlinearity tests,

Phys. Rev. Lett. 77, 635 (1996).

[98] T. Schreiber, A. Schmitz, Surrogate time series, Physica D 142, 346

(2000).

[99] E. Koscielny-Bunde, J.W. Kantelhardt, P. Braun, A. Bunde und S. Havlin,

Long-term persistence and multifractality of river runoff records, J.
Hydrol. 322, 120 (2006).

53
[100] J.W. Kantelhardt, E. Koscielny-Bunde, D. Rybski, P. Braun, A. Bunde, S.
Havlin, Long-term persistence and multifractality of precipitation and
river runoff records, J. Geophys. Res. (Atmosph.) 111, D01106 (2006).

[101] B. Podobnik, H.E. Stanley, Detrended cross-correlation analysis: a new

method for analyzing two nonstationary time series, Phys. Rev. Lett.
100, 084102 (2008).

Figure 1: (a) Examples of self-affine series xi characterized by different Hurst exponents H = 0.9, 0.7, 0.5, 0.3 (from top to bottom). The data has
been generated by Fourier filtering (see Section 8.1) using the same seed for the random number generator. (b) Differentiated series ∆xi of the data
from (a); the ∆xi are characterized by positive long-term correlations (persistence) with γ = 0.2 and 0.6 (top curve and second curve), uncorrelated
behaviour (third curve), and anti-correlations (bottom curve), respectively.

Figure 2: Comparison of the autocorrelation functions C(s) (decreasing functions) and fluctuation functions F2(s) (increasing functions) for
shortterm correlated data (top) and long-term correlated data (γ = 0.4, bottom). The asymptotic slope H ≈ α = 0.5 of F2(s) clearly indicates the
absence of long-term correlations, while H ≈ α = 1 – γ/2 > 0.5 indicates the presence long-term correlations. The difference is much harder to
observe in C(s), where there are more statistical fluctuations and even negative values (e.g., above s = 150 in (a) and between s = 300 and 400 in
(b), not shown). The data have been generated by an AR process (Eq. (4)) and Fourier filtering (Section 8.1) for (a) and (b), respectively. The
dashed lines indicate the theoretical curves.

54
 frequency

Figure 3: Spectral analysis of a fractal time series characterized by long-term correlations with γ = 0.4 (β = 0.6). The expected scaling behaviour
(dashed line indicating the slope –β) is observed only after binning of the spectrum (circles). The data has been generated by Fourier filtering (see
Section 8.1).

55
Figure 4: Illustration of the Fluctuation Analysis (FA) and the Detrended Fluctuation Analysis (DFA). For two segment durations (time scales) s
=
100 (a) and 200 (b), the profiles Y(j) (blue lines; defined in Eq. (10)), the values Y(νs) used for the FA in Eq. (11) (green circles), and least square
quadratic fits to the profiles used in DFA (red lines) are shown.

56
Figure 5: Application of discrete wavelet transform (WT) analysis on uncorrelated data (black circles), long-term correlated data (γ = 0.8, α = 0.6,
red squares), and short-term correlated data (summation of three AR processes, green diamonds). Averages of F2(s) averaged over 20 series with N
= 216 points and divided by s1/2 are shown, so that a horizontal line corresponds to uncorrelated behaviour. The blue open triangles show the result
for one selected extreme configuration, where it is hard to decide about the existence of long-term correlations (figure prepared by Mikhail
Bogachev).

57
Figure 6: Application of Detrended Fluctuation Analysis (DFA) on the data already studied in Fig. 5 (figure by Mikhail Bogachev).

58
Figure 7: Example of the Wavelet Transform Modulus Maxima (WTMM) method, showing the original data (top), its continuous wavelet
transform (grey scale coded amplitude of wavelet coefficients, middle), and the extracted maxima lines (bottom) (figure taken from [67]).

59
Figure 8: Multifractal Detrended Fluctuation Analysis (MF-DFA) of data from the binomial multifractal model (see Section 8.3) with a = 0.75.
Fq(s) is plotted versus s for the q values given in the legend; the slopes of the curves correspond to the values of h(q). The dashed lines have the
slopes of the theoretical slopes h(±∞) from Eq. (42). 100 configurations have been averaged.

60
Figure 9: Illustration for the definition of return intervals rq between extreme events above two quantiles (thresholds) q1 and q2 (figure by Jan
Eichner).

61
Figure 10: Normalized rescaled distribution density functions RqPq(r) of r values with Rq = 100 as a function of r/Rq for long-term correlated data
with γ = 0.4 (open symbols) and γ = 0.2 (filled symbols; we multiplied the data for the filled symbols by a factor 100 to avoid overlapping curves).
In (a) the original data was Gaussian distributed, in (b) exponentially distributed, in (c) power-law distributed with power −5.5, and in (d)
lognormally distributed. All four figures follow quite well stretched exponential curves (solid lines) over several decades. For small r/Rq values a
power-law regime seems to dominate, while on large scales deviations from the stretched exponential behaviour are due to finite-size effects (figure
by Jan Eichner).

62
 0 365 730 1095 1460 1825 2190 2555 2920
i [days]

Figure 11: Illustration for the definition of maxima mR within periods of R = 365 values (figure by Jan Eichner).

63