Matlab Coding

Czech Technical University in Prague
Faculty of Electrical Engineering

Department of Cybernetics
Processing of neuronal single-cell recordings

from deep structures of human brain
Master thesis
Karel Schmidt
Supervisor: Ing. Daniel Novák Ph.D.

Study programme: Electrical Engineering and Information Technology
Study branch: Biomedical Engineering
leden 2008
Abstract
The micrororecording signal from deep structures of human brain was

analyzed. The signal was recorded from patients with Parkinson’s disease
receiving deep brain stimulation. First the algorithm was tested on artificially
generated data. State space attractors were reconstructed from the measured
time series using time delay embedding. The generalized dimensions of the
reconstructed attractor were estimated using algorithm based on generalized
correlation sum. The algorithm was found robust and sufficiently accurate
for q > 0. The results show that this analysis could not describe the chaotic
behaviour of the dynamical system of neurons in deep structures of human
brain, because of the insufficient length of data.
Abstrakt
Analyzovali jsme signál z mikroregistrace z hlubokých struktur lidského

mozku. Záznam byl pořı́zen během implantace hluboké mozkové stimulace
pacientům s Parkinsonovou chorobou. Nejdřı́ve jsme vyzkoušeli funkčnost
algoritmu na uměle vytvořených datech. Z naměřených dat jsme zrekon-
struovali atraktor ve stavovém prostoru. Pomocı́ algoritmu založeného na
zoběcněné korelačnı́ sumě jsme spočı́tali zobecněné dimenze zrekonstruo-
vaného atraktoru. Algoritmus byl shledán robustnı́ a dostatečně přesný pro
q > 0. Výsledky ukazujı́, že tato analýza vzhledem k nedostatečné délce
záznamu nemůže popsat chaotické chovánı́ dynamického systému neuronů v
hlubokých strukturách lidského mozku.
I would like to thank Dan Novák, my supervisor, for his many suggestions
and constant support in my work.
Prague, Czech Republic Karel Schmidt

January, 2008
Contents
1 Introduction 9
2 Medical Background 10
2.1 Parkinson’s disease . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1.1 Parkinson’s disease . . . . . . . . . . . . . . . . . . . . 10
2.1.2 Parkinson’s disease treatment . . . . . . . . . . . . . . 10
2.2 Deep brain stimulation . . . . . . . . . . . . . . . . . . . . . . 11
2.2.1 Neurosurgery . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.2 Navigation . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.3 Microrecording . . . . . . . . . . . . . . . . . . . . . . 13
3 Methods 14
3.1 State of Art . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2 Spike sorting . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3 Fractal analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3.1 Time delay embedding . . . . . . . . . . . . . . . . . . 17
3.3.2 Estimate the time delay . . . . . . . . . . . . . . . . . 17
3.3.3 Estimate embedding dimension . . . . . . . . . . . . . 18
3.3.4 Correlation dimension . . . . . . . . . . . . . . . . . . 21
3.3.5 Multifractal spectrum of generalized dimensions . . . . 22
3.3.6 Spectrum of scaling indices . . . . . . . . . . . . . . . . 24
4 Algorithms 27
4.1 Embedding . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.1.1 Estimate the time delay . . . . . . . . . . . . . . . . . 27
4.1.2 Estimate embedding dimension . . . . . . . . . . . . . 28
4.2 Fractal dimensions . . . . . . . . . . . . . . . . . . . . . . . . 29
4.2.1 Average Pointwise Mass Algorithms . . . . . . . . . . . 30
4.2.2 k-Nearest-Neighbor (Fixed-Mass) Algorithms . . . . . . 32
4.3 Specific implementations . . . . . . . . . . . . . . . . . . . . . 33
4.3.1 k-Nearest-Neighbor . . . . . . . . . . . . . . . . . . . . 33
1
4.3.2 Generalized correlation sum . . . . . . . . . . . . . . . 33
4.3.3 Generalized dimensions . . . . . . . . . . . . . . . . . . 34
4.3.4 Maximal and minimal generalized dimensions . . . . . 36
4.4 Spectrum of scaling indices . . . . . . . . . . . . . . . . . . . . 38
5 Results 39
5.1 Golden Standard data . . . . . . . . . . . . . . . . . . . . . . 39
5.1.1 Types of data used . . . . . . . . . . . . . . . . . . . . 39
5.1.2 Embedding the golden standard signals . . . . . . . . . 41
5.1.3 Generalized dimensions using k-nearest neighbor algo-
rithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.1.4 Calculating the generalized dimensions . . . . . . . . . 43
5.2 Measured data . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.2.1 Embedding the measured data . . . . . . . . . . . . . . 50
5.2.2 Generalized correlation sums from measured data . . . 53
5.2.3 Statistical analysis . . . . . . . . . . . . . . . . . . . . 57
5.2.4 Generalized dimensions from measured data . . . . . . 57
5.2.5 Analyzing the spike train . . . . . . . . . . . . . . . . . 61
6 Conclusion 63
6.1 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
Appendices 67
A State of Art 69
A.1 Fractal analysis of spinal dorsal horn neuron discharges by
means of sequential fractal dimension D . . . . . . . . . . . . 69
A.2 Multifractal statistics and underlying kinetics of neuron spike
time-series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
A.3 Fractal patterns in auditory nerve-spike trains . . . . . . . . . 70
A.4 Fractal character of the neural spike train in the visual system
of the cat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
A.5 Characterization of non-linear dynamics in rat cortical neu-
ronal networks . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
A.6 Bursting, spiking, chaos, fractals and universality in biological
rhythms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
A.7 Multiplicative multifractal modeling and discrimination of hu-
man neuronal activity . . . . . . . . . . . . . . . . . . . . . . 73
B Source codes 74
B.1 embed.m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
2
B.2 findembedparam.m . . . . . . . . . . . . . . . . . . . . . . . . 75
B.3 knnGolden.m . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
B.4 multifrac.m . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
C Generalized dimensions analysis 80
3
List of Figures
3.1 Auto mutual information of the Lorentz Attractor, τs = 15. . . 18

3.2 Cao’s method functions E1(d) and E2(d) for the Henon map. 20
3.3 Henon map with the data points partitioned in a grid of boxes.
Figure is from [4]. . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.4 Generalized dimension Dq as a function of q, f (q) and α(q)
for a typical multifractal. Figure is from [29]. . . . . . . . . . . 25
3.5 f (α) curve and significant values. Figure is from [29]. . . . . . 26
4.1 Complete process of multifractal analysis. . . . . . . . . . . . . 38
5.1 Attractor of the Henon map reconstructed from its x-component. 40

5.2 The Lorenz attractor reconstructed from its x-component. . . 40
5.3 Auto mutual information of the x-component of the Henon map. 41
5.4 Cao’s method functions E1(d) and E2(d) for Lorentz attractor. 42
5.5 Spectrum of generalized dimensions Dq calculated using the
k-nearest neighbor algorithm. . . . . . . . . . . . . . . . . . . 42
5.6 Scaling of the generalized correlation sum of the Henon map. . 43
5.7 Generalized dimensions Dq of the Henon map. . . . . . . . . . 45
5.8 Spectrum of scaling indices of the Henon map. . . . . . . . . . 46
5.9 Scaling of the generalized correlation sum of the Lorenz at-
tractor. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.10 Generalized dimensions Dq of the Lorenz attractor. . . . . . . 48
5.11 Spectrum of scaling indices of the Lorenz attractor. . . . . . . 49
5.12 Raw data of a signal from STN. . . . . . . . . . . . . . . . . . 50
5.13 Operation protocol from DBS implanting operation. . . . . . . 51
5.14 Auto mutual information of the measured data. . . . . . . . . 53
5.15 Cao’s method functions E1 and E2 for the measured data
signal 17, patient P. . . . . . . . . . . . . . . . . . . . . . . . . 54
5.16 Scaling of the generalized correlation sum of signal 17, patient
P. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.17 Scaling of the generalized correlation sum of signal 13, patient
P. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4
5.18 Spike train detection. . . . . . . . . . . . . . . . . . . . . . . . 57
5.19 Example, signal 18, patientP. Red line is the extraction thresh-
old. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.20 Example, the spike train extracted using the QSort algorithm
from the signal 18, patientP. . . . . . . . . . . . . . . . . . . . 58
5.21 Generalized correlation sums and Dq ’s for signals 6, 9, 13, 26
for patient P. . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.22 Plain signal, surrogate data and amplitude adjusted surrogate
data (top to bottom) for signal 13, patient P. . . . . . . . . . 60
5.23 Generalized correlation sums and Dq ’s for surrogate data from
signal 13, patient P. . . . . . . . . . . . . . . . . . . . . . . . . 60
5.24 Generalized correlation sums and Dq ’s for amplitude adjusted
surrogate data from signal 13, patient P. . . . . . . . . . . . . 61
5.25 Cao’s method functions E1 and E2 for a spike train extracted
from our signal. . . . . . . . . . . . . . . . . . . . . . . . . . . 62
C.1 Generalized correlation sums and Dq ’s for signals 21 and 27

from patient P. . . . . . . . . . . . . . . . . . . . . . . . . . . 80
C.2 Generalized correlation sums and Dq ’s for signals 11, 15, 17,
28 from patient P. . . . . . . . . . . . . . . . . . . . . . . . . . 81
C.3 Generalized correlation sums and Dq ’s for signals 7, 9, 13, 25
from patient J. . . . . . . . . . . . . . . . . . . . . . . . . . . 82
C.4 Generalized correlation sums and Dq ’s for signals 5, 10, 17, 27
from patient K. . . . . . . . . . . . . . . . . . . . . . . . . . . 83
C.5 Generalized correlation sums and Dq ’s for surrogate data from
different signals. . . . . . . . . . . . . . . . . . . . . . . . . . . 84
C.6 Generalized correlation sums and Dq ’s for amplitude adjusted
surrogate data from different signals. . . . . . . . . . . . . . . 85
5
Glossary
CT computed tomography.
DBS deep brain stimulation.
GPi globus pallidus interna.
IPG implantable pulse generator.
ISI inter-spike interval.
MRI magnetic resonance imaging.
PD Parkinson’s disease.
SNr substantia nigra.
STN subthalamic nucleus.
Th thalamus.
6
Nomenclature
a(i, d) ratio of d + 1 to d dimensional reconstructed space distance between

the reconstructed vector and it’s nearest neighbor
α scaling index of generalized dimensions
By pointwise mass function, numerically equal to Pi
C(r) correlation sum
Cq generalized correlation sum
d embedding dimension
D2 correlation dimension
Dq generalized dimension of chaotic attractor
E(d) Cao’s method first quantity
E ∗ (d) Cao’s method second quantity
E1(d) variation of Cao’s method first quantity
E2(d) variation of Cao’s method second quantity
f (α) spectrum of scaling indices
µ population mean
n() minimum number of open sets of diameter needed to cover a fractal

set
Pi probability for a random point to be ith box on the attractor, numer-

ically equal to By (r)
r search radius, the main parameter of the correlation sum
7
tj jump time, time interval between succesive embedded vectors
τ measure of generalized dimensions Dq
τs time delay between sampled values in time delay embedding
Θ(x) Heaviside step function
w Theiler window
xi time series sample
yi the embedded time delay vector
h·i denotes expectation value of a function, for uniformly distributed pa-

rameter, arithmetic mean
8
Chapter 1
Introduction
In this work we will analyze the micorecording signals from deep structures
of human brain. The signals were recorded from patients with Parkinson’s
disease. They were recorded during the stereotactic surgery performed for
implantation of the deep brain stimulation.
We will analyze the raw data and the detected spike trains. A multifractal
analysis will be done. Different numerical measures of chaos will be discussed.
This work will focus on the generalized dimensions and spectrum of scaling
indices.
Several algorithms for calculating these measures of chaos will be dis-
cussed. The more promising algorithms will be chosen and implemented. A
practical implementation will be thoroughly described.
The implementation will be tested on artificially generated data with
known properties. Later they will be used on real world signals. The results
and factors affecting the results will be discussed.
We will evaluate the suitability of the multifractal analysis for the mi-
crorecording signals.
9
Chapter 2
Medical Background
2.1 Parkinson’s disease

2.1.1 Parkinson’s disease
Parkinson’s disease (PD) belongs to a group of conditions called motor sys-
tem disorders, which are the result of the loss of dopamine-producing brain
cells. The four primary symptoms of PD are tremor, or trembling in hands,
arms, legs, jaw, and face; rigidity, or stiffness of the limbs and trunk; bradyki-
nesia, or slowness of movement; and postural instability, or impaired balance
and coordination. As these symptoms become more pronounced, patients
may have difficulty walking, talking, or completing other simple tasks. PD
usually affects people over the age of 50. Early symptoms of PD are subtle
and occur gradually. In some people the disease progresses more quickly than
in others. As the disease progresses, the shaking, or tremor, which affects the
majority of PD patients may begin to interfere with daily activities. Other
symptoms may include depression and other emotional changes; difficulty in
swallowing, chewing, and speaking; urinary problems or constipation; skin
problems; and sleep disruptions. There are currently no blood or laboratory
tests that have been proved to help in diagnosing sporadic PD. Therefore
the diagnosis is based on medical history and a neurological examination.
The disease can be difficult to diagnose accurately. Doctors may sometimes
request brain scans or laboratory tests in order to rule out other diseases.
[20]
2.1.2 Parkinson’s disease treatment

At present, there is no cure for PD, but a variety of medications provide
dramatic relief from the symptoms. Usually, patients are given levodopa
10
combined with carbidopa. Carbidopa delays the conversion of levodopa into
dopamine until it reaches the brain. Nerve cells can use levodopa to make
dopamine and replenish the brain’s dwindling supply. Although levodopa
helps at least three-quarters of parkinsonian cases, not all symptoms respond
equally to the drug. Bradykinesia and rigidity respond best, while tremor
may be only marginally reduced. Problems with balance and other symptoms
may not be alleviated at all. Anticholinergics may help control tremor and
rigidity. Other drugs, such as bromocriptine, pramipexole, and ropinirole,
mimic the role of dopamine in the brain, causing the neurons to react as
they would to dopamine. An antiviral drug, amantadine, also appears to
reduce symptoms.
Treating PD with surgery was once a common practice. But after the
discovery of levodopa, surgery was restricted to only a few cases. Studies in
the past few decades have led to great improvements in surgical techniques,
and surgery is again being used in people with advanced PD for whom drug
therapy is no longer sufficient. [20]
2.2 Deep brain stimulation

Deep brain stimulation (DBS) is a surgical procedure used to treat a variety
of disabling neurological symptoms—most commonly the debilitating symp-
toms of Parkinson’s disease (PD), such as tremor, rigidity, stiffness, slowed
movement, and walking problems. The procedure is also used to treat es-
sential tremor, a common neurological movement disorder. At present, the
procedure is used only for patients whose symptoms cannot be adequately
controlled with medications.[18]
The DBS system consists of three parts (1) DBS electrode, or lead, (2)
connecting wire and (3) implantable pulse generator (IPG). The electrical
pulses generated by the IPG are lead through the connecting wires to the
DBS electrodes that are placed in deep brain structures. Mostly subthalamic
nucleus (STN) or globus pallidus interna (GPi) is stimulated.
A small incision is made in the area of the collarbone, and the pulse
generator is placed there. The wire is placed under the scalp and skin. The
by far most time consuming and complicated part is the placement of the
DBS electrodes.
2.2.1 Neurosurgery
The stimulation electrodes are implanted by stereotactic surgery, a neuro-
surgeon who specializes in central nervous system function uses a stereotac-
11
tic head frame and imaging techniques such as magnetic resonance imaging
(MRI) or computed tomography (CT) scanning to map the brain and localize
the target within the brain.
The patient has the stereotactic head frame attached to the head prior to
the operation. Local anesthesia is used when attaching the frame. The frame
is providing three dimensional coordinate system for the brain mapping. Us-
ing MRI a three dimensional image of the brain is acquired. The calibrations
on the head frame are merged with the brain image to form a computerized
map of the brain. This map becomes the blueprint for planning and mea-
suring the trajectories of the electrodes into the deep brain structures. The
entire head frame structure is then bolted to the operating table to maintain
the head in a fixed position throughout the operation. Two burr hole are
made in the top of the skull. A strong topical anesthetic is used to numb the
skin while this hole is drilled. Since there are no pain receptors in the brain,
there is no need for deeper anesthetic. In addition, the patient must remain
awake in order to report any sensory changes during the surgery and later to
test the effect of the stimulation. [3]
2.2.2 Navigation
Targeting STN is problematic because of its variable location and relatively
small size (20–30 mm3). A combination of anatomic imaging with a stereo-
tactic frame, atlas coordinates, and intraoperative neurophysiology is cur-
rently considered the most reliable approach for STN targeting.
Anatomic targeting employing various imaging modalities, such as ven-
triculography, CT and MRI, and intraoperative electrophysiologic testing are
combined with information from time-tested autopsy-based atlases. These
atlases provide the coordinates of the STN and other structures relative to the
established anatomic landmarks, primarily the commissures and ventricles.
In case of thalamic localization, where borders of nuclei are variable and
essentially indeterminable with current anatomic imaging, this atlas-based
approach remains the gold standard.
The STN tends to change its borders, size, and shape significantly with
age, whereas the distance between the anterior and posterior commissures,
used as the main references for all atlas-based stereotactic procedures for
STN targeting, is almost constant during life. Therefor the combination of
MRI with intraoperative MER mapping of STN borders has proven to be
the most reliable technique for the best DBS targeting of STN. [24]
12
2.2.3 Microrecording
As an additional means to ensure the accuracy of the surgical probe electrical
brain mapping is performed during surgery. This technique uses micro elec-
trodes that can record electrical activity from individual brain cells within
deep brain regions. The micro-electrodes are much smaller and more delicate
than the electrodes that provide the deep brain stimulation. They are used
to identify cells within the thalamus, globus pallidus, subthalamic nucleus
and adjacent brain structures, and help steer the main probe towards the
desired surgical target.
These electrodes are inserted towards the target location estimated based
on the MRI. The signal that is being recorded is visually and acoustically
analyzed by the surgeon. The analysis is based on known characteristics of
the signals from different brain structures and on surgeons experience. If
the STN is not registered on the first pass of the microelectrode, the micro-
electrode is inserted parallel to the previous pass more in anterior,posterior,
medial and lateral covering a circle with approximately 4 mm diameter.
13
Chapter 3
Methods
3.1 State of Art

The scientific and technical works were searched in the INSPEC database,
because Inspec provides comprehensive coverage of traditional and cutting-
edge publications in the specific fields of physics, electrical and electronic
engineering, communications, computer science, control engineering, infor-
mation technology, manufacturing and mechanical engineering.
In addition there is significant coverage of areas such as operations re-
search, material science, oceanography, engineering mathematics, nuclear
engineering, environmental science, geophysics, nanotechnology, biomedical
technology and biophysics.
The database was searched through the Dialog Easy web interface.
Keywords
• chaos, fractal[s]
• spike train, time series
• single cell, neuron
Search queries
Dialog. Dtb: INSPEC. Query:
– All years (1994—2008)
– Entire text: fractal? AND spik? AND neuron?
17 records found. Relevant: A.1, A.2, A.3, A.4, A.6, A.7
Dialog. Dtb: INSPEC. Query:
– All years (1994—2008)
14
– Entire text: chao? AND neur? AND time-series AND cell
13 records found. Relevant: A.5.
INSPEC Results
Long-duration power-law correlation was found in a spike train in the visual
system of the cat A.4. Low-dimensional deterministic chaos in extra-cellular
recordings of mature rat cortical cells was reported by constructing the return
map and calculating the Lyapunov exponent A.5. Long-term correlations,
self-similarity of neuronal firing rates, power-law behavior of the Fano-factor
time curve and change in the firing pattern were analyzed A.3. Probabilistic
and multifractal properties of probability distributions of interspike intervals
were analyzed in a more recent study A.2. With the bifurcating diagrams, it
was shown how spiking can be transformed to bursting via a complex type
of dynamic structure when the key parameter in the model varied A.6. In
the above mentioned works, long sequences of interspike intervals of some
thousands spikes were analyzed, but in this study there are only 10 seconds
long records with some hundred spikes and therefor these methods cannot
be used in this case.
Neural firing pattern in two brain areas, the globus pallidus externa (GPe)
and the globus pallidus interna (GPi), was observed and chracterized by
mulitfractals. The generalized dimension spectrum Dq effectively differenti-
ates the two brain areas. A.7. This work also used microrecording data from
the surgery of deep brain stimulation. Our work is inspired by this work.
3.2 Spike sorting

We performed the spike sorting to analyze the firing pattern of the neu-
ron. We were looking for a difference in statistics of the firing pattern of a
oscillating neuron and a healthy neurons.
As seen in the state of art search, many studies have had promising results
when analyzing inter-spike interval (ISI) of spike trains from different neurons
using fractal analysis. The authors of these articles have usually measured
the firing pattern of neurons of experimental animals [26] [27]. We try ISI
to use the mulitfractal analysis on the detected spike trains from our signals
and compare them to the ISI statistics. We also compare the ISI statistics
of the spike trains to the multifractal analysis of the raw signals.
Online spike sorting is used. [21]
Spikes are detected using threshold crossings of a local energy measure-
ment p(t) of the bandpass filtered signal, which allows more reliable spike
15
detection than thresholding the raw signal. If p(t) is locally bigger than five
times the standard deviation of p(t) (or an other factor, referred to below as
the extraction threshold), a candidate spike is detected. For each threshold
crossing, a sample of 2.5 ms is extracted from the filtered signal.
The estimation of the number of neurons present, as well as the assign-
ment of each spike to a neuron, is based on a distance metric between two
spikes. Based on this distance, a threshold is used to decide: (i) how many
neurons are present and (ii) to assign each spike uniquely to one neuron or
to noise, if unsortable. A crucial element of this approach is the threshold,
which is calculated from the noise properties of the signal and is equal to
the squared average standard deviation of the signal, calculated with a slid-
ing window. The threshold is thus not a parameter as it is automatically
defined by the noise properties of the recording channel and is equal to (in
a theoretical sense) the minimal signal-to-noise ratio required to be able to
distinguish two neurons.
Each newly detected spike is sorted as soon as it is detected. The raw
waveform of a newly detected, as of yet unsorted spike, is used to calculate
the distance to all already known mean waveforms (clusters). The spike
is assigned to the existing cluster to which it has minimal distance if the
distance is smaller than a threshold value. If the minimal distance is larger
than the threshold, a new cluster is automatically created. Every time a
spike is assigned to a cluster, the mean waveform of that cluster is updated
by taking the mean of the last C spikes that were assigned to this cluster. This
causes the mean waveforms of each cluster to change as well, which might
result in two clusters which have mean waveforms whose distance is less than
the threshold. In this case, the two clusters become indistinguishable and
they are thus merged. The spikes assigned to both clusters will be assigned
to the newly created cluster. [21]
3.3 Fractal analysis

There are different ways of quantifying chaotical behavior of a system. The
more promising one for biological systems as follows from the state of art
search is the correlation dimension and especially the generalized dimensions
spectrum.
A time-series data is used as an input. It represents one variable of
the dynamical system. However the dynamical system usually has more
degrees of freedom than one. The dynamical system is than described with
trajectory in a state space of higher dimension. The different dimensions
of the dynamical system are quantifiers emphasizing the geometry of the
16
trajectory in state space. As the number of iterations goes to infinty the
trajectory approaches a set of points that is called an attractor. Dimension
of this attractor is measured. When the dimension is a non integer value it
is called a strange attractor and the dimension a fractal dimension. [29]
Short summary of our Fractal analyses...
3.3.1 Time delay embedding

Only a time series of one variable was measured, but this time-series data
can be used to reconstruct a multidimensional embedding space [25]. The
evolution of the trajectories in the embedding space mimics the behavior of
the actual trajectories in the full state space. The process of reconstuction
is called embedding. When a d-dimensional space should be reconstructed
d values from the time series are grouped together to form a vector. Each
number in the vector gives the value of the component of the vector along
one of hte axes in this d-dimensional space. Together, the d values give co-
ordinates of a single point in the d-dimnesional space. The time evolution of
the system is followed by seeing how successive vectors move in d-dimesional
space. The basic assumption is that the geometry and dynamics of the vec-
tors constructed in this way are the same as the geometry and dynamics of
the vectors in the full state space of the system. [16]
3.3.2 Estimate the time delay

Some of the questions that arise before the attractor reconstruction can be
done are:
How should be chosen the time delay betweem sampled values?
How do we choose the time interval between successive vectors?
We start from the second question. The time interval between succesive
vectors is called tj jump time. The jump time tj has to be small enough so
that most of the details of the attractor geometry can be maped. Therefore
it is usual to choose jump time tj to be equal to the sampling time. In this
case the number of vectors is equal to the number of samples in the time
series minus the embedding dimension. [16]
yi (d) = (xi , xi+τs , ..., xi+(d−1)τs ), i = 1, 2, ..., N − (d − 1)τs ,
where d is the embedding dimension and τs is the time delay between

sampled values.
17
Figure 3.1: Auto mutual information of the Lorentz Attractor, τs = 15.
The time delay between sampled values that form a vector τs can also
be estimated using the auto mutual information. The mutual information
I is the amount of information that is shared between two data sets. The
auto mutual information takes time delayed copies of one data set as the
second one. The first minimum of the auto mutual information is said to be
a prefered value for attractor reconstruction from time series, where one is
interested in independend coordinates, see Fig. 3.1.
X P (X, Y )
I(X, Y ) = P (X, Y )log , (3.1)
P (X)P (Y )
where P (X) is the probability of measuring a data value X, P (X, Y ) is
the joint probability of measuring X and Y at the same time. A low value
of the mutual information shows that there is little information common
between data sets. A normalized value of 1 shows the data sets are equal.
[8]
3.3.3 Estimate embedding dimension

Another parameter that needs to be estimated before we can reconstruct
the state space is the embedding dimesion d. To determine the optimal
embedding dimension from a the time series the reliable and commonly used
Cao’s method was used. [5]
We define the measure a(i, d)
18

yi (d + 1) − yn(i,d) (d + 1)
a(i, d) =
yi (d) − yn(i,d) (d) , i = 1, 2, ..., N − dτs , (3.2)
where k·k is some measurement of Euclidian distance and is given in this

paper by the maximum norm, i.e.,
kyk (m) − yl (m)k = max |xk+jτs − xl+jτs | ;

0≤j≤m−1
yi (d + 1) is the ith reconstructed vector with embedding dimension d + 1.

i.e., yi (d + 1) = (xi , xi+τs , ..., xi+dτs ); n(i, d)(1 ≤ n(i, d) ≤ N − dτs ) is an inte-
ger such that yn(i,d) (d) is the nearest neighbor of yi (d) in the d-dimensional
reconstructed phase space in the sense of distance k·k - we defined above.
n(i, d) depends on i and d.
Notes. (1) The n(i, d) in the numerator of Eq. 3.2 is the same as that
in the denominator. (2) If yn(i,d) (d) equals yi (d), we take the second nearest
neighbor instead of it.
The measure a(i, d) is a ratio of the Euclidian distance between the re-
constructed vector yi (d + 1) and it’s nearest neighbor yn(i,d) (d + 1) in a
d + 1-dimensional reconstructed space to the Euclidian distance between the
same reconstructed vector yi (d) and the same nearest neighbor yn(i,d) (d) but
this time in a d-dimensional reconstructed space
If d is qualified as an embedding dimension by the embedding theorems,
then any two points which stay close in the d-dimensional reconstructed space
will be still close in the (d + 1)-dimensional reconstructed space. Such a pair
of points are called true neighbors, otherwise, they are called false neighbors.
The following quantity is defined, i.e., the mean value of all a(i, d)’s,
−dτs
NX
1
E(d) = a(i, d) (3.3)
N − dτs i=1
E(d) is dependent only on the dimension d and the time delay τs . To

investigate its variation from d to d + 1. we define
E1(d) = E(d + 1)/E(d) (3.4)
We found that when increased by steps of 1 E1(d) stops changing when

d is greater than some value d0 if the time series comes from an attractor.
Then d0 + 1 is the minimum embedding dimension we look for.
19
Before we perform numerical tests, it is necessary to define another quan-
tity which is useful to distinguish deterministic signals from stochastic sig-
nals. Let
NX−dτs
∗ 1
xi+dτs − xn(i,d)+dτs
E (d) = (3.5)
N − dτs i=1
where the meaning of n(i, d) is the same as above, i.e., it is the integer
such that yn(i,d) (d) is the nearest neighbor of yi (d). We define
E2(d) = E ∗ (d + 1)/E ∗ (d) (3.6)
Figure 3.2: Cao’s method functions E1(d) and E2(d) for the Henon map.
For time series data from a random set of numbers, E1(d), in principle,
will never attain a saturation value as d increases. But in practical compu-
tations, it is difficult to resolve whether the E1(d) is slowly increasing or has
stopped changing if d is sufficiently large. In fact, since available observed
data samples are limited, it may happen that the E1(d) stops changing at
some d although the time series is random. To solve this problem, we can
consider the quantity E2(d). For random data, since the future values are
independent of the past values, E2(d) will be equal to 1 for any d in this case.
However, for deterministic data, E2(d) is certainly related to d, as a result,
it cannot be a constant for all d; in other words, there must exist some d’s
such that E2(d) 6= 1.
20
Cao recommends calculating both E1(d) and E2(d) for determining the
minimum embedding dimension of a scalar time series, and to distinguish
deterministic data from random data [5].
3.3.4 Correlation dimension

Generalized dimensions are one class of quantities to characterize the fractal-
ity. The Hausdorff or capacity dimension is, roughly speaking, the exponent
D in the expression n() = −D , where n() is the minimum number of open
sets of diameter needed to cover the set. It is, from the mathematical point
of view, the most natural concept to characterize fractal sets. Whereas the
information dimension takes into account the probability that a element is
visited and is therefore more attractive for physical systems. Finally, for the
characterization of measured data, other similar concepts, like the correla-
tion dimension, are more useful. One general remark is highly relevant in
order to understand the limitations of any numerical approach: dimensions
characterize a set or an invariant measure whose support is the set, whereas
any data set contains only a finite number of points representing the set or
the measure. By definition, the dimension of a finite set of points is zero.
When we determine the dimension of an attractor numerically, we extrap-
olate from finite length scales, where the statistics we apply is insensitive
to the finiteness of the number of data, to the infinitesimal scales, where
the concept of dimensions is defined. This extrapolation can fail for many
reasons which will be partly discussed below. Dimensions are invariant un-
der smooth transformations and thus computable in time delay embedding
spaces. [13]
The correlation dimension is computed most efficiently by the correlation
sum: [10]
N
1 X X
C(d, r) = Θ(r − |yi − yj |) (3.7)
Npairs i=d j<i−w
where yi are m-dimensional delay vectors. Npairs = (N − d − w)(N −

d − w + 1)/2 the number of pairs of points covered by the sums. Θ is the
Heaviside step function: Θ(x) is zero for x < 0 and one for x ≥ 0. w will
be discussed below. The summation counts the number of pairs (yi , yj ) for
which the distance |yi − yj | is less than r.On sufficiently small length scales
and when the embedding dimension d exceeds the correlation dimension of
the attractor,[23]
C(d, r) ∝ rD2
21
Since one does not know the correlation-dimension before doing this com-
putation, one checks for convergence of the estimated values of D2 in d. The
relevant caveats and misconceptions are reviewed, for example, in [29]. The
most prominent precaution is to exclude temporally correlated points from
the pair counting by the so called Theiler window w [29]. In order to become
a consistent estimator of the correlation integral from which the dimension
is derived the correlation sum should cover a random sample of points drawn
independently according to the invariant measure on the attractor. Succes-
sive elements of a time series are not usually independent. In particular, for
highly sampled flow data subsequent delay vectors are highly correlated.
log C(r)
D2 = lim (3.8)
r→0 log r
Theiler suggested to remove this spurious effect by simply ignoring all

pairs of points in Eq. (3.7) whose time indices differ by less than w, where w
should be chosen generously. With O(N 2 ) pairs available, the loss of O(wN )
pairs is not dramatic as long as w N . At the very least, pairs with
j = k have to be excluded, since otherwise the strong bias to D2 = 0, the
mathematically correct value for a finite set of points, reduces the scaling
range drastically. Choosing w, the first zero of the auto-correlation function,
sometimes even the decay time of the auto-correlation function, are not large
enough since they reflect only overall linear correlations.[29] The space–time-
separation plot provides a good means of determining a sufficient value for w.
In some cases, notably processes with inverse power law spectra, inspection
requires w to be of the order of the length of the time series. This indicates
that the data does not sample an invariant attractor sufficiently and the
estimation of invariants like D2 or Lyapunov exponents should be abandoned.
3.3.5 Multifractal spectrum of generalized dimensions

The generalized dimension that was introduced in [14], and independently
in [9], does take into account the number of points in a box. Let Bi denote
the ith box, and let Pi = µ(Bi )/µ(A) be the normalized measure of this box,
where A is the fractal whose dimension one wishes to calculate and µ is the
population mean of a set. Equivalently, it is the probability for a randomly
chosen point on the attractor to be in Bi , and it is usually estimated by
counting the number of points that are in the ith box and dividing by the
total number of points. If the attractor had been embedded in dimension d,
hypercubes of dimension d will be used.
22
Figure 3.3: Henon map with the data points partitioned in a grid of boxes.
Figure is from [4].
The generalized dimension is defined by
log i Piq
P
1
Dq = lim . (3.9)
q − 1 r→0 log r
the sum of Piq as a weighted average i Piq = i Pi (Piq−1 ) =

P P
Writing
q−1
Pi , one can associate bulk with the generalized average probability per

q−1 1/(q−1)
box ( Pi ) and identify Dq as a scaling of bulk with size. (h·i denotes
the expectation value of a function, for a function with uniformly distributed
parameter, like in this case the parameter i, it denotes the arithmetic mean.
For q = 2 the generalized average is the ordinary arithmetic average, and for
q = 3 it is a root mean square. It is not hard to show that the limit q → 1
leads to a geometric average. Finally, it is noted that q = 0 corresponds to
the Hausdorff or capacity dimension defined above.
For a uniform fractal, that is a fractal with all Pi equal, one obtains
a generalized dimension Dq that does not vary with q. For a nonuniform
fractal, however, the variation of Dq with q quantifies the nonuniformity. For
instance,
log (maxi Pi )
D∞ = lim . (3.10)
r→0 log r
23
log (mini Pi )
D−∞ = lim . (3.11)
r→0 log r
It is clear from Eq. (3.9) that Dq decreases with increasing q. From this
fact and the above equations, it is clear that the maximum dimension D−∞
is associated with the least-dense points on the fractal and the minimum
dimension D∞ corresponds to the most-dense points. This should not be
surprising: The densest set possible is a point, which has dimension zero.
The notion of generalized dimension first arose out of a need to under-
stand why various algorithms gave different answers for dimension. A fur-
ther motivation came from the need to characterize more fully fractals with
nonuniform measure. These sets are sometimes called multifractals and are
characterized by an a priori measure, that differs from the Hausdorff mea-
sure. The point is that, rather than measure just one dimension, one can
compute the full spectrum of generalized dimensions from D−∞ to D∞ .[29]
3.3.6 Spectrum of scaling indices

Let f (α) be the spectrum of scaling indices Halsey et al. [12] introduced
a change of variables that provides a new interpretation of generalized di-
mension. Let τ = (q − l)Dq and take a Legendre transformation from the
variables (q, τ ) into a new set of variables (a, f ):
∂τ
α= f = αq − τ (3.12)
∂q
and
∂f
q= τ = αq − f. (3.13)
∂α
An example of these functions can be seen in Fig. 3.4. In case either τ (q)
or f (α) is not differentiable, a more robust formulation is given by
f (α) = min{qα − τ (q)} (3.14)

q
τ (α) = min{qα − f (α)} (3.15)

α
The authors of [12] interpret f (α) as a ”spectrum of scaling indices”. 3.5

The variable α is the scaling index: µi = δiαi defines the local scaling at the
24
Figure 3.4: Generalized dimension Dq as a function of q, f (q) and α(q) for a
typical multifractal. Figure is from [29].
ith element of the cover. Then, the number n(α, r) of cover elements with
scaling index between α and α + ∆α scales as n(α, r) ∼ r−f (α) ∆α.
The curve f (α) in Fig. 3.5 is always convex upward, and the peak of the
curve occurs at q = 0. At this point f is equal to the fractal dimension D0 .
Also, the f (α) curve is tangent to the curve f = a, and the point of tangency
occurs at q = 1. In general, the left-hand branch corresponds to q > 0 and
the right-hand branch to q < 0.
To see that thisPinterpretation leads to Eqs. (3.12)-(3.15), consider the
fixed-box-size sum i Piq . The number of terms in this sum for which Pi = rα
is given by n(α, r). Thus
X q Z Z
Pi = n(α, r)r dα ∼ r−f (α) rqα dα ∼ rθ ,
qα
(3.16)
i
where θ = minα {qα − f (α)} since the integral will be dominated by

the smallest exponent of r when r → 0. Comparing P qthis to the definition of
(q−l)Dq
generalized dimension in Eq. (3.9), which gives i Pi ∼ r , one obtains
(q −1)Dq = mina {qα−f (α)}, which is the formula given in Eq. (3.15). From
this formula, Eqs. (3.12)-(3.14) follow.
This scaling n(α, r) ∼ r−f (α) suggests interpreting f (α) as the Hausdorff
dimension of the points with scaling index α. Indeed, the scaling index α is
associated with the pointwise dimension Dp , and the set
25
Figure 3.5: f (α) curve and significant values. Figure is from [29].
Sα = {X ∈ A; Dp (X) = α} (3.17)
is the set of all points in A for which the pointwise dimension is α. The
Hausdorff dimension of the set Sα is given by f (α).
The reader who is hoping for a dramatic picture of this fractal set Sα will
probably be disappointed. As Sakar [22] and others have pointed out, Sα is
not necessarily a closed set and may even be dense in the original fractal A.
Thus, although the set Sα may have a lower Hausdorff dimension than A, it
is possible for the box-counting dimension to be the same. And a picture of
Sα would look just like the picture of the original set A.
The f (α) formalism provides a tool for testing the notion of universality,
which states that a wide variety of dynamical systems should behave in a
similar way and should leave the same characteristic signatures. Indeed,
several researchers for example [11] have found physical systems whose f (α)
curves precisely matched the f (α) associated with a theoretical model of a
circle map undergoing a transition from quasi-periodicity to chaos. [29]
26
Chapter 4
Algorithms
In this chapter we will describe implementation of the methods which we

theoretically explained in the previous chapter 3. The first section describes
how to reconstruct the phase space and how to estimate the parameters
needed for the reconstruction. The second section gives the overview of the
algorithms used to calculate the fractal dimensions and explains the concrete
implementation. In the last section an algorithm that calculates the spectrum
of scaling indices is given.
4.1 Embedding
In this section a simple algorithm for the time delay embedding of a recon-
struction phase space is described. After the initial check if the input signal
is long enough for the required embedding, a simple loop runs over the em-
bedding dimensions and writes the values from the input signal with the time
delay τs and step of the jump time tj into the output, see source code 1.
for i=1:dim
start = (i-1)*delay;
d(:,i) = cin(start+1:shift:start+len);
end
Source code 1: for loop embedding a time series
4.1.1 Estimate the time delay

A tool that can help us determine the right value for time delay τs is the auto
mutual information function introduced in the previous chapter in section
27
3.3.2 and defined by Eq. (3.1). To calculate the auto mutual information
function itself we use the algorithm from TSTOOL package for Matlab [19].
To use this function the input signal has to be converted to rang values in
the interval < 0, 1), see source code 2. Then the amutual mex function from
the TSTOOL is used, syntax in source code 3.
[y,i] = sort(ts);
N=length(y);
y(i) = (0:N-1)/N;
Source code 2: time series conversion to rang values
a = amutual(y, maxtau, bins);
Source code 3: syntax of the amutual function from TSTOOL package
The preferred value for the time delay τs is said to be the first minimum
of the auto mutual information function a. As we know from observations
the auto mutual information function is always decreasing from τs = 0 to the
first minimum. Because of this assumption we can find by searching for the
first zero crossing of the first derivate of the function, see source code 4.
da=diff(a);
tau=find(da>=0,1,’first’)-1;
Source code 4: searching the first minimum of the auto mutual information
function
At the end the auto mutual information function is plotted so that the
user can check the results, an example is in Fig. 3.1.
4.1.2 Estimate embedding dimension

To determine the best embedding dimension for the reconstruction, we will
use the Cao’s method [5] explained in the previous chapter in section 3.3.3.
First we have to embed the signal with some high embedding dimension.
For embedding we use the algorithm mentioned above in section 4.1. As a
necessary embedding parameter we use the time delay τs obtained from the
algorithm in source codes 2, 3 and finally 4.
28
To calculate the functions E(d) from Eq. (3.3) and the function E ∗ (d)
from Eq. (3.5) we use a cao mex function from the TSTOOL package for
Matlab [19], syntax in source code 5.
[E, Estar] = cao(eData, ref, NNR);
Source code 5: syntax of the cao function from TSTOOL package
To speed up the computation we use only N random reference points for

this computation. As shown in [5], as few as 10000 reference points is enough
to estimate the embedding dimension. The computation can be enhanced by
calculating with more then one nearest neighbor.
Then we can compute the function E1(d) defined in Eq. (3.4) and func-
tion E2(d) defined in Eq. (3.6). This computation is in source code 6.
E1=E(2:end)./E(1:end-1);
E2=Estar(2:end)./Estar(1:end-1);
Source code 6: computing E1(d) and E2(d) from E and E ∗
The next part tries to estimate the embedding dimension as the point
where the function E1(d) stops changing as illustrated in the previous chapter
in Fig. 3.2. This point is found by looking for the second derivate of the
function E1(d), see source code 7.
ddE1=diff(diff(E1));
[dummy dim]=min(ddE1);
Source code 7: second derivate of the E1 function
Since there are more factors affecting the choice of the embedding dimen-
sion the functions E1(d) and E2(d) are plotted and the user can make his
own evaluation of the results as suggested by Cao [5] .
4.2 Fractal dimensions

It is emphasized that numerical techniques can only estimate the dimension
of a fractal. Practical estimation of dimension begins with a finite description
of the fractal object.
29
The mostly used way to compute dimension is to use the correlation
algorithm, which estimates dimension based on the statistics of pairwise dis-
tances. The correlation algorithm is in the class of fixed-size algorithms
because it is based on the scaling of mass with size for fixed-sized balls (or
grids). An alternative approach uses fixed-mass balls, usually by looking
at the statistics of distances to kth nearest neighbors. Both fixed size and
fixed-mass algorithms can be applied to estimation of generalized dimension
Dq .
4.2.1 Average Pointwise Mass Algorithms

The definition of the pointwise mass function is By (r) = µ[By (r)], where
By (r) is the ball of radius r about the point y. It is numerically equal to Pi ,
see also section 3.3.5 in previous chapter.
The scaling of By (r) with r is what defines the pointwise dimension at y.
If this scaling is the same for all y, then the fractal is uniform, and it follows
that Dq is a constant independent of q and has the value of the pointwise
dimension. For most fractals, however, pointwise dimension is not a global
quantity, and some averaging is necessary for the quantity to be made global.
From the point of view of practical computation, it is possible to compute
the pointwise dimension at a sample of points on the attractor and then to
average these values to obtain an estimate of the dimension of the attractor.
This average is one that in principle gives the information dimension D1 .
However, it is probably more efficient to estimate dimension by averaging
the mass function before the limit in Eq. (3.8) is taken. It is in this way that
the statistics of the interpoint distances can be most effectively exploited.
[29]
Correlation dimension
The most natural such averaging strategy was introduced by Grassberger
and Procaccia [10]. Here, a direct arithmetic average of the pointwise mass
function gives what Grassberger and Procaccia call the correlation integral:
C(r) = hBy (r)i. (4.1)

Here, By (r) can be approximated from a finite data set of size N by
N
1 X
Byj (r) ≈ Θ(R − kyi − yj k) (4.2)
N − 1 i=1,i6=j
30
where Θ is the Heaviside step function: Θ(x) is zero for x < 0 and one
for x ≥ 0. The importance of excluding i = j has been overlooked by some
authors, although Grassberger has stressed this point [9]. In fact, the case is
made in [29] for excluding all values of i for which |i − j| < W with W > 1.
From this, the correlation dimension D2 is defined:
log C(r)
D2 = lim as in Equation (3.8)
r→0 log r
It is now straightforward to approximate C(r) with a finite data set:
1 X
C(N, r) = Θ(R − kyi − yj k) similar to Eq. (3.7). (4.3)
N (N − 1) i6=j
In words,
number of distances less than r
C(N, r) =
number of distances altogether
Thus the correlation algorithm provides an estimate of dimension based
purely on the statistics of pairwise distances. Not only is this a particularly
elegant formulation but it has the substantial advantage that the function
C(r) is approximated even for r as small as the minimum interpoint distance.
For N points, C(N, r) has a dynamic range of O(N 2 ). Logarithmically
speaking, this range is twice that available to n(N,r) in the box-counting
method. It is also twice the range available in an estimate of the point-
wise mass function Bx (r) for a single point X. This greater range is the
one advantage that the correlation integral has over the average pointwise
dimension.
Generalized dimensions
A more general average than the direct arithmetic average used in the pre-
vious paragraphs of this section 4.2.1 is given by
Cq (r) = By (r)(q−l)

(4.4)
The scaling of this average with r gives the generalized dimension:

1 log Cq r
Dq (r) = lim (4.5)
q−1 r→0 log r
31
Note that q = 2 gives the direct arithmetic average that defines the
correlation dimension and that the q = 1 average (which is associated with
the information dimension).
From a finite set of points, generalized correlation sum Cq (N, r) can be
approximated by [16]
N
" N
#q−1
1 X 1 X
Cq (N, r) = Θ(R − kyi − yj k) (4.6)
N i=1 N − 1 i=1,i6=1
This formula is easiest to evaluate when q = 2, in which case it reduces

to the simple form given in Eq. (4.3). Algorithms based on this equation for
the general case q 6= 2 have been given in [16]. This approach was reported
not to work well for q < 1 since the term that is raised to the power of
q − 1 can be zero for r much larger than the smallest interpoint distance.
The formula ought to work well for large positive values of q, however. We
implementation this approach in this work.
4.2.2 k-Nearest-Neighbor (Fixed-Mass) Algorithms

In contrast to the average pointwise mass functions described in section 4.2.1,
in which the variation of mass inside fixed size balls is considered, the nearest-
neighbor algorithms consider the scaling of sizes in fixed-mass balls [29].
An early implementation of this notion in a chaotic-dynamics context
was suggested in [28]. Here one computes (rk ), which is the average distance
to the kth nearest neighbor, as a function of k. Let R(yi , k) denote the
distance between point yi and its kth nearest neighbor. [By convention, the
0th nearest neighbor is the point itself, so R(X, 0) = 0 for all y.] The average
is computed:
N
1 X
hrk i = R(yi , k) (4.7)
N i=1
The scaling (rk ) ∼ k l/D defines the dimension. (Actually, what was com-
puted in [28] was hrk2 i, since for Euclidean distances this equation is com-
putationally more efficient. And the scaling was taken to be hrk2i ∼ k 2/D .
The author of [1] considers moments of the average distance to the kth near-
est neighbor and recommends keeping k fixed and computing a dimension
function from the scaling of average moments of rk with the total number of
points N :
hrkγ i ∼ (k/N )γ/D(γ) (4.8)
32
This dimension function D(γ) is related to the generalized dimension by
the implicit formulas
γ = (q − 1)Dq , D(γ) = Dq (4.9)

and
q = γ/D(γ) + 1, Dq = D(γ) (4.10)
4.3 Specific implementations

4.3.1 k-Nearest-Neighbor
One implementation of the k-nearest neighbors algorithm for calculating the
generalized dimensions is implemented as gendimest mex function in the
TSTOOL package [19]. It implements the algorithm introduced in [30]. It
has the syntax shown in source code 8
[dims, moments]=gendimest(dists,gammas,kmin_low,kmin_high,kmax)
Source code 8: syntax of the gendimest function from TSTOOL package
The input argument dists is a matrix which contains distances from

reference points to their k nearest neighbors. This matrix should be obtained
by calling fnearneigh, which is nearest neighbor searcher from the TSTOOL
package, on the point set whose dimension spectrum is to be investigated.
Gammas is vector of the moment orders from Eq. (4.9) chosen by the user.
kmin_low, kmin_high and kmax are the parameters specifying the number of
neighbors used for the calculation.
This algorithm has some limitations it is unreliable for |q| > 4 as stated
in [30].
4.3.2 Generalized correlation sum

In the TSTOOL package there is also a mex function implementing the al-
gorithm for computation of correlation sum. We derived a new function
from the original TSTOOL mex function. This new function calculates the
generalized correlation sum introduced in the previous section in Eq. (4.6).
We modified the source code to calculate the generalized correlation sum.
The original implementation will not be discussed here. We left the original
33
computation of By (r) as is. The part of the code that is responsible to raising
the By (r)’s to the power of q − 1 is in source code 9.
for (long n=0; n < R; n++) {

...
sums[ni]++;
}
}
if (pairs > 1) {
double suml = 0;
double sumk = 0;
for (long bin=0; bin < bins; bin++) {
suml += sums[bin];
sums[bin] = 0;
if (suml > 0) {
sumk = suml / pairs;
corrsums[bin] += pow( sumk , (q-1) );
}
}
}
}
Source code 9: part of the process of computing the generalized correlation

sum
The highest for loop runs for all reference points. The original part
where the neighboring points within the distance r are searched is left out.
Each of the points found is added to the sum for the concrete distance, here
sums[ni]++. The number of found neighbors is divided by the number of
searched pairs and that gives the By (r). Each By (r) is raised to the power
of q − 1 using the C++ function pow. That is the implementation of the
inner member of the arithmetic average in Eq. (4.4).
To finish the computation of the Eq. (4.4) just the arithmetic average has
to be evaluated. The summation was done in the source code 9. The division
by the number of all reference points for which the distances to neighbors
were searched is done in the source code 10.
4.3.3 Generalized dimensions

The function for computing the generalized dimensions uses the generalized
correlation sum function in source codes 9 and 10. This function returns the
34
if (total_pairs > 0) {
double sum = 0;
sum = corrsums[bin];
corrsums[bin] = (sum / R);
}
}
Source code 10: finishing the computation of the code 9
correlations sums and the respective distances r for which the generalized
correlation sums were calculated. The distances are logarithmically spaced
in order to get a equal spacing in the log-log scaling plot. The limit in Eq.
(4.5) can be represented as scaling of the log Cq (r) to log r. In practical
realization there is need of user user-feedback since the log-log plot is not
linear everywhere. A linear scaling region has to be found manually [29] [15],
see source code 11.
The scale of log Cq (r) to log r is the Dq Eqs. (3.9) (4.5). It is fitted in the
scaling region using the least square method. Function lscov is a standard
Matlab function. The standard deviation of this process is also calculated to
know the accuracy of this method. The process is shown in source code 12.
figure; plot(logR,logC);
a=input(’scaling region from:’);
b=input(’scaling region to:’);
ScalingRegion=find((logR>a)&(logR<b));
Source code 11: requesting the user to select the scaling region
[limtau, limstdx, limmse] = ...

lscov([logR(ScalingRegion), ones(length(ScalingRegion), 1)],...
logC(ScalingRegion) );
D(q)=limtau(1); stdx(q)=limstdx(1); mse(q)=limmse(1);
stdD(q)=sqrt(mse(q));
Source code 12: Fitting the scale using least square method
35
4.3.4 Maximal and minimal generalized dimensions
To estimate the D∞ and D−∞ we implemented the Eq. (3.10) and (3.11).
The practical implementation is a small modification of the source code 9
and 10.
for (long n=0; n < R; n++) {
...
sums[ni]++;
}
}
if (pairs > 1) {
double suml = 0;
double sumk = 0;
suml += sums[bin];
sums[bin] = 0;
if (suml > 0) {
if (sumk > corrsums[bin]) {
corrsums[bin] = sumk;
}
}
}
}
}
Source code 13: Find the maximum By (r) to calculate D∞
The different part of the function is shown in source code 13. The highest
for loop runs for all reference points. The original part of the TSTOOL mex
function where the neighboring points within the distance r are searched
is left out. Each of the points found is added to the sum for the concrete
distance, here sums[ni]++. The number of found neighbors is divided by the
number of searched pairs and that gives the By (r). Each By (r) is compared
to the already stored value in corrsum[bin] if the value is bigger it is stored
in the corrsum[bin]. This way we end up with the maximal By (r) stored
in corrsum[bin] and that is the output of this function.
The D∞ is the scaling of the previous function with r. In practical real-
ization there is need of user user-feedback since the log-log plot is not linear
everywhere. A linear scaling region has to be found manually [29] [15], see
36
source code 11 in previous section 4.3.3.
The scale of log max By (r) to log r is the D∞ Eqs. (3.10). It is fitted in
the scaling region using the least square method. The standard deviation
of this process is also calculated to know the accuracy of this method. The
process is shown in source code 12 in previous section 4.3.3.
for (long n=0; n < R; n++) {

...
sums[ni]++;
}
}
if (pairs > 1) {
double suml = 0;
double sumk = 0;
suml += sums[bin];
sums[bin] = 0;
if (suml > 0) {
if (sumk > corrsums[bin]) {
corrsums[bin] = sumk;
}
}
}
}
}
Source code 14: Find the minimal By (r) to calculate D−∞
The source code 14 shows the different part of the algorithm for finding
the minimal By (r). The only difference is the comparison of the new value
By (r) stored in sumk with the smallest value so far in corrsum[bin]. We are
looking for a smaller value, if it is found it will be stored corrsum[bin] and
sent to the output.
The process of the least square method slope fitting is same as in the case
for D∞ .
37
4.4 Spectrum of scaling indices
Our implementation also calculates the spectra of scaling indices f (α). The
implementation of the Legendre transformation of τ (q) (3.12) is shown in
source code 15.
alfa=diff(tau)’;
f=(-(qzero-1):(q-qzero-1)).*alfa-tau(1:numofqs-1)’;
Source code 15: Legendre transformation
The rest of the spectra is extrapolated using splines. Since for q → ∞

and q → −∞ the function f (α) converges to zero in a finite point α it can
be done and the viewer of the plot gets a better perspective. See source code
16.
alfa2=[max(alfa)+0.2:-0.001:max(alfa), ...
alfa, min(alfa):-0.001:min(alfa)-0.24];
f2=interp1(alfa, f, alfa2, ’spline’, ’extrap’);
Source code 16: interpolation of f (alpha)
Figure 4.1: Complete process of multifractal analysis.
38
Chapter 5
Results
5.1 Golden Standard data

First we evaluated our algorithms on artificially generated signals.
5.1.1 Types of data used

The Henon map is very often used fractal set used to evaluate different meth-
ods for quantifying chaos. The Henon map is defined as follows:
xn+1 = 1 − αx2n + yn (5.1a)

yn+1 = βxn (5.1b)
where α, β are parameters, we used the values α = 1.4, β = 0.5. These

values for parameters are mostly used for example in [5] [16]. For the calcu-
lations we took the x-component as time series.
We also used the Lorenz attractor, which is expressed by:
dx
= σ(y − x) (5.2a)
dt
dy
= ρx − y − xz (5.2b)
dt
dz
= xy − βbz (5.2c)
dt
σ, ρ, β are parameters and we used the values σ = 10, ρ = 28, β = 8/3.
For the calculations we again took the x-component.
39
Figure 5.1: Attractor of the Henon map reconstructed from its x-component.
Figure 5.2: The Lorenz attractor reconstructed from its x-component.
40
5.1.2 Embedding the golden standard signals
Figure 5.3: Auto mutual information of the x-component of the Henon map.
We calculated the auto mutual information using the amutual mex func-
tion from the TSTOOL package. The results for Lorentz attractor were
already shown in chapter 3 in Fig. 3.1 and τs = 15. The results are with
correspondence with the literature [5]. For the Henon map the results shown
in Fig. 5.3 do not show a distinct results and we used time delay τs = 1, this
is also with correspondence with the literature [5].
To find the embedding dimension we used the Cao’s method [5]. The
results are for the x-component of the Henon map d = 2 and the x-component
of the Lorenz attractor d = 3. The results for the Henon map are shown in
Fig. 3.2 and for Lorentz attractor in Fig. 5.4. These figures look exactly
same as in Cao’s article [5].
5.1.3 Generalized dimensions using k-nearest neighbor

algorithm
First we tried to use the implementation of the k-nearest neighbor algorithm,
described in sections 4.2.2 and 4.3.1. This function is part of the TSTOOL
package. We wrote a simple script, see appendix B.3, to test its capabilities
on the prepared artificial signals, the Henon map and the Lorenz attractor.
These results show, that this algorithm can be used only in a limited
range of q’s.
41
Figure 5.4: Cao’s method functions E1(d) and E2(d) for Lorentz attractor.
(a) the Henon map (b) the Lorenz attractor
Figure 5.5: Spectrum of generalized dimensions Dq calculated using the k-

nearest neighbor algorithm.
42
For the Henon map the Dq curve was decreasing only for q =< 1, 10 >,
outside of this interval the results contradicted the theory, that Dq is mono-
tonically decreasing function. Though the value D2 the correlation dimen-
sion was estimated to be D2 = 1.207 which is close to the theoretical value
D2 = 1.25 ± 0.02 [10].
For the Lorentz attractor the Dq curve was decreasing only for q =<
−3, 8 >, outside of this interval the results contradicted the theory, that Dq
is monotonically decreasing function. Though the value D2 the correlation
dimension was estimated to be D2 = 2.005 which is close to the theoretical
value D2 = 2.05 ± 0.01 [10].
Because of the limits of this algorithm we will further focus on the gen-
eralized correlation sum algorithm.
5.1.4 Calculating the generalized dimensions

the Henon map
We analyzed the x-component of the Henon map embedded with dimension
2 with time delay τs = 15.
Figure 5.6: Scaling of the generalized correlation sum of the Henon map.
43
To estimate the generalized dimension we calculated the generalized cor-
relation sums for various q’s. We have chosen the interval q ∈< −10, 10 >
because all the tendencies can be seen on this interval.
To estimate the D∞ we calculated the scaling of the maximal measure
Pi , see Eq. (3.10) and for the D−∞ the scaling of minimal measure Pi , see
Eq. (3.11).
On Fig. 5.6 the scaling for various q’s is plotted. Significant q’s are
labeled. The linear scaling region is highlighted with red line for q = −∞
and q = ∞. It is important to note that we have plotted log(q−1) Cq (r)
versus log r.
The difference between slopes for different q’s would be bigger if we plotted
just log Cq (r) versus log r. Furthermore in such a plot slopes for q ≤ 0 would
be negative since always the measure is 0 < Pi < 1 and when raised to power
of negative exponent the smaller the value the bigger the result. We prefer
to directly compare the graphs in one plot and therefor we have chosen to
plot log(q−1)
Cq (r)
versus log r.
The graph for q = 2 is linear in almost the whole interval in which it was
calculated. We expected that since for q = 2 our algorithm works almost ex-
actly the same way as the original correlation sum algorithm from TSTOOL
package [19].
With increasing q the slope of the graph is decreasing as expected from
theory 3.3.5. The shape of the graph is becoming more like the shape of the
graph of the maximal measure Pi . More precisely it is the graph of logarithm
of this measure log Pi , in this figure it is labeled q = ∞.
The stair step on the graph q = ∞ for small log r < −4.5 is the effect
of discretization made while generating the Henon map with finite number
of points. Similar behaviour of correlation sum can be found in [29] as an
example of the effect of discretization.
The small intrinsic oscillation on the graph q = ∞ can be caused by the
lacunarity of the attractor. Lacunarity is a measure of how the fractal fills
space, if the fractal is dense the lacunarity is small, the lacunarity increases
with coarseness. [29]
The graph q = 0 is first decreasing and later increasing. This is due to the
fact mention at the end of section 4.2.1. For log r < −3 there is a significant
number of points, that have the measure Pi equal to zero. Because of that
these points contribute to the sum with a zero but theoretically they should
contribute with something like ∞. Practically that cannot be realized. With
r ≈ −3 the biggest amount of points have Pi = N1 , where N is the total
−1
number of points. These points contribute to the sum with N1 = N and
the correlation sum C0 (r) is at its maximum for r ≈ −3. In our figure it is
the minimum of the graph q = 0, because we plotted log(q−1) Cq (r)
versus log r.
44
As q → −∞ the shape of the graphs is getting similar to the graph
q = −∞. The graph q = −∞ is equal to N1 for log r < −2.35 this means
that up to this scale the least dense region of the attractor has only one
relevant neighboring point within the distance r.
For q smaller than approximately −4 there are two regions with oscil-
lations. This can be again caused by the lacunarity of the attractor of the
Henon map.
Figure 5.7: Generalized dimensions Dq of the Henon map.
Generalized dimensions of the reconstructed attractor of the Henon map

are plotted in figure 5.7. Red lines show the standard deviation of the least
square method used to determine a slope from the graphs. For large q’s
the values are very inaccurate but the variation of the value with q is more
important than the exact value since the variation is the measure of non-
uniformity of the fractal.
For positive q’s the curve is convex, monotonically decreasing and with a
limit in ∞. That is exactly with accordance to the theory.
However for negative q’s the graph is oscillating. It seems that it is
oscillating around the expected function. That means the values that the
graph is oscillating around seem to converge to a finite value. Never the less
for q < −2 the estimation is unreliable.
The correlation dimension D2 was estimated to be D2 = 1.198 ± 0.008
that is close to the theoretical value D2 = 1.25 from [10]. In fact it is
45
approximately as accurate as the k-nearest neighbor algorithm evaluated in
the previous section 5.1.3.
Figure 5.8: Spectrum of scaling indices of the Henon map.
The spectrum of scaling indices α is a good representation of the scaling

behaviour of the correlation sum of the attractor. We constructed it from
τ = Dq (q − 1) using the Legendre transformation, see section 3.3.6.
For the Henon map we have plotted only f (α) calculated from Dq ’s for
q ∈< −2, 4 > in Fig. 5.8. It should be noted that the Legendre trans-
formation results in the desired concave curve f (α) only when the τ curve
is concave. Therefore we have limited the number of q’s from which the
Legendre transformation was calculated.
We can see that only a little more than the left part of the curve is
present which represents q > 0. The curve apparently leads to the point,
where α = D∞ and f (α) = 0 that corresponds with the theory, see Fig. 3.5.
On the peak of the curve f (α) = D0 which is in accordance with the
theory.
Lorentz attractor
We also analyzed the x-component of the Lorenz attractor embedded with
dimension 3 and with time delay τs = 15.
46
Figure 5.9: Scaling of the generalized correlation sum of the Lorenz attractor.
To estimate the generalized dimension we again calculated the generalized

correlation sums for various q’s. We have chosen the interval q ∈< −10, 10 >.
To estimate the D∞ we calculated the scaling of the maximal measure Pi ,
see Eq. (3.10) and for the D−∞ the scaling of minimal measure Pi , see Eq.
(3.11).
On Fig. 5.9 the scaling for q ∈< −10, 10 > is plotted. Significant q’s are
labeled. And the linear scaling region is highlighted with red line for each
individual q. It is important to note that we have plotted log(q−1)
Cq (r)
versus log r.
The graph q = 2 is linear in almost its whole length. The q = 0 has the
usual shape. It is decreasing for small r due to the points with the measure
Pi = 0. For log r > −0.5 it is approximately linear and increasing with a
measurable slope.
With increasing q’s the slope of the graph is decreasing corresponding to
the theory. It is also converging to the shape of the graph for q = ∞.
The graph for q = ∞ for log r < −10 has two the stair steps. These are
accounted to the discretization. Because the Lorenz attractor occupies the
phase space more densely there is no significant oscillation in the graph for
log r > −0.5.
47
For q < 0 we can see a heavy fluctuation for 0.5 < log r < 1. We
do not think this can be be accounted to the points with Pi = 0 since for
r this large there should be such point any more. Theiler [29] mentions
an anomalous shoulder that can be caused by auto-correlation in the time
series. But the shape does not correspond with the shoulder illustrated by
Theiler.The authors of [6] have made a thorough multifractal analysis of the
Lorenz attractor, but they haven’t accounted abnormalities like this. This
abnormality is therefor without a explanation. Though the rest of the graphs
can be used for the estimation of the slope and we used it.
Figure 5.10: Generalized dimensions Dq of the Lorenz attractor.
Generalized dimensions of the reconstructed attractor of the Lorenz at-

tractor are plotted in figure 5.10. Red lines show the standard deviation of
the least square method used to determine a slope from the graphs. The stan-
dard deviations show that the least square fitting was much more accurate
this time. That means that the scaling region was very close to linearity.
For q > 0 the graph has the expected shape. For q < −2 the graph is
oscillating again. Again the trend of the oscillating values is close to the
expected behaviour. Compare the graph for q < 0 to the one for k-nearest
neighbors estimator in Fig. 5.5b. There the graph starts to decrease for
decreasing q which is in contradiction with the theory.
The correlation dimension D2 was estimated to be D2 = 2.047 ± 0.006
that is in excellent accordance with the theoretical value of D2 = 2.05 ± 0.01
48
from [10].
Figure 5.11: Spectrum of scaling indices of the Lorenz attractor.
The spectrum of scaling indices f (α) for the Lorenz attractor constructed
from τ = Dq (q − 1) using the Legendre transformation is shown in Fig. 5.11.
We have plotted f (α) calculated only from Dq ’s for q > −2. We have
again limited the number of q’s from which the Legendre transformation was
calculated, because of the limited interval of the correct shape of the τ curve.
The graph shows only the left part of the curve which represents q > 0.
This part of the f (α) curve corresponds to the theory and has the direction
toward the point α = D∞ , f (α) = 0 as in Fig 3.5.
Also peak of the f (α) curve has the value of D0 as expected from theory.
5.2 Measured data

We received a database of data from MUDr. Robert Jech, Ph.D. from the
Department of Neurology of the 1st Faculty of Medicine at Charles University
in Prague and the General University Hospital in Prague. The data were
recorded during DBS implantation operation performed by MUDr. Jech
at the Department of Stereotactic and Radiation Neurosurgery of the Na
Homolce Hospital in Prague. For our analysis we have randomly chosen 3
patients. The data were recorded from 5 electrodes, one central and four
49
Figure 5.12: Raw data of a signal from STN.
lateral. An example of the operation protocol with the highlighted areas of

the brain is in Fig. 5.13. For reference into the database the codes of signals
are in table 5.1.
For the evaluation of the algorithms we have randomly chosen data differ-
ent patients and from different depths of the electrode. Included are signals
from different areas of the brain: the thalamus (Th), the STN, the substantia
nigra (SNr) and the brain tissue between them. The signals are 10 seconds
long. Each signal is from one or up to four spiking neurons with background
noise. The noise is produced by other neurons which are farther from the
tip of the microelectrode that is measuring the signal. In the brain tissue
between the nuclei there is usually no spiking. A 5 seconds long part of a
example signal from STN is for illustration shown in Fig. 5.12.
5.2.1 Embedding the measured data

To reconstruct the attractor from the measured data we used the time delay
embedding 4.1. We calculated the auto mutual information function from
two different signals from the STN to try to estimate the time delay to be
used to form the time delay vectors τs . The results are shown in Fig. 5.14.
The auto mutual information was calculated using the matlab mex-function
from the TSTOOL toolbox for matlab [19].
The graph for signal 13, patient P in Fig. 5.14b is similar to the graph of
automutual information for the Henon map 5.3 and from this graph we can
not estimate the time delay τs . The graph for signal 17, patient P in Fig.
5.14a has the first minimum of auto mutual information function for τs = 4,
which has been used in the following experiments.
To estimate the dimension for the time delay reconstruction we used the
Cao’s method [5], see 3.3.3. We tested signal 17, patient P with the time
delay τs = 4. The results are shown in Fig. 5.15a. The E1 and E2 functions
are uncorrelated and E2 has approximately constant values. That means
50
Figure 5.13: Operation protocol from DBS implanting operation.
51
Table 5.1: Codes of the signals used for the analysis for reference to our
database.
code in database part of brain patient and signal code

Pra Z STNsin lat 6 Th P 6
Pra Z STNsin centr 9 n/a P 9
Pra Z STNsin centr 13 STN P 13
Pra Z STNsin centr 26 SNr P 26
Jan V STNdex post 9 n/a J 9
Jan V STNdex med 7 Th J 7
Jan V STNdex centr 13 STN J 13
Jan V STNdex lat 25 SNr J 25
Pra Z STNsin lat 11 Th P 11
Pra Z STNsin lat 15 n/a P 15
Pra Z STNsin lat 17 STN P 17
Pra Z STNsin centr 27 SNr P 27
Kub J STNdex post 5 n/a K 5
Kub J STNdex post 10 Th K 10
Kub J STNdex post 17 STN K 17
Kub J STNdex post 27 SNr K 27
52
(a) signal 17, patient P (b) signal 13, patient P
Figure 5.14: Auto mutual information of the measured data.
that the Cao’s method is not able to determine the embedding dimension in
this case. In fact the results shown in [5] for white noise look very similar.
That means that for the embedded signal 17, patient P with the time delay
τs = 4 there is probably no chaotic behaviour.
As did Cao in his work, we tried the time delay τs = 1. The results are
in Fig. 5.15b. The E1 and E2 functions are correlated and the E1 function
attains a saturation value for d = 4. We have applied the Cao’s method
on more randomly chosen signals and the results for all of them were very
similar to Fig. 5.15b.
After this analysis of embedding we have have chosen the following em-
bedding parameters τs = 1 and d = 4. We will embed all the signals in the
following analysis with these parameters.
5.2.2 Generalized correlation sums from measured data

To estimate the generalized dimensions of signal 17, patient P we calculated
the generalized correlation sums for the interval q ∈< −6, 6 >. To estimate
the D∞ we calculated the scaling of the maximal measure Pi , see Eq. (3.10)
and for the D−∞ the scaling of minimal measure Pi , see Eq. (3.11). We have
used N = 5000 random reference points from the reconstructed attractor to
speed up the computation.
On Fig. 5.16 the scaling for all q’s is plotted. Significant q’s are labeled.
The linear scaling regions are highlighted with red line for all q’s.
The graph q = 2 has properties that have to be noted. For Cq (r) <
1
N
the graph is braking down from the linear scaling behaviour. Since we
used N = 5000 reference points the graph in Fig. 5.16 is braking down for
53
(a) τs = 4 (b) τs = 1
Figure 5.15: Cao’s method functions E1 and E2 for the measured data signal
17, patient P.
f raclog Cq (r)q − 1 < 3.7. According to [29] this behaviour is typical for non
chaotic attractor. Also the graph is becoming saturated long before r reaches
the value equal to the size of the attractor. According to [15] this is typical
for a correlation sum of white noise. The graph q = 2 or the plain correlation
sum has aspects that do not support the chaotic behaviour of the measured
signal 17, patient P.
The graph q = ∞ has for small r a shape similar to the same graph for
Henon map in Fig. 5.6. For large r it is attaining saturation before the size
of the attractor is reached similarly to the graph q = 2. With increasing q
the slope of the graphs is decreasing and the shape of the graphs is becoming
more like the of graph q = ∞.
The graph q = 0 has two different slopes in its region and it was difficult
to find appropriate scaling region. The graph q = −∞ has a very long flat
part for r < 0.7. This means that the attractor is scattered much more than
the attractors the artificial signals. With decreasing q the shape is becoming
very irregular. We will discuss the properties of the generalized correlation
sum for q ≤ 0 after we look at the results for signal 13, patient P.
We calculated the generalized correlation sums of signal 13, patient P for
the interval q ∈< −6, 6 >. To estimate the D∞ we calculated the scaling of
the maximal measure Pi , see Eq. (3.10) and for the D−∞ the scaling of min-
imal measure Pi , see Eq. (3.11). We have used N = 5000 random reference
points from the reconstructed attractor to speed up the computation.
On Fig. 5.17 the scaling for all q’s is plotted. Significant q’s are labeled.
The linear scaling regions are again highlighted with red line for all q’s.
The graph q = 2 for signal 13, patient P has similar properties as the
54
Figure 5.16: Scaling of the generalized correlation sum of signal 17, patient
P.
same graph for signal 17, patient P. It seem to come from a non chaotic
attractor or a noisy signal. Also graph q = ∞ and and all graphs for q > 0
for signal 13, patient P look similar to these graphs for signal 17, patient P.
Graphs for q ≤ 0 are very irregular and finding a linear scaling region is
almost impossible. We will now focus on the graph q = −∞ and discuss the
possible source of the irregularity.
As we mentioned earlier in section 4.2.1 the main problem for q ≤ 0 are
the points where Pi = 0, this is the term in the generalized correlation sum
that is raised to the power of q − 1, see Eq. (4.5) and as By (r) in Eq. (4.4).
These zeros can figure in the sum for very large r’s. Our implementation
omits these terms, since the calculation could not be made at all with zeros
raised to power of q − 1 for q ≤ 0.
As long as the attractor is regular and dense in the phase space, eg. with
low lacunarity defined in section 5.1.4, there are no more points with the term
Pi equal to zero already for relatively small r. In such a case this problem
causes only small irregularities and only for r close to the minimal interpoint
distance.
55
Figure 5.17: Scaling of the generalized correlation sum of signal 13, patient
P.
However in this case the attractor is probably scattered a lot since the
original signal is noisy. Thus there are points with Pi = 0 for r much larger
than the minimal interpoint distance. The graph q = −∞ is calculated
according to Eq. (3.11) as the scaling of the minimal measure Pi with r.
When we follow this graph q = −∞ from the smallest r it has a constant
value of the minimal Pi = N1 . For log r ≈ 0.5 the smallest Pi found in the
attractor starts to increase. But there are some points in this attractor which
have the nearest neighbor point in a distance log r ≈ 1.2. It may be even a
single point. While computing the Cq (r) for q = −∞ only the smallest Pi
is counted to the Cq (r). For all log r < 1.2 these points with Pi = 0 where
omitted, but when one of these point gets the value Pi > 0 for log r ≈ 1.2 it
becomes the smallest Pi and the graph falls down to this value.
This graph is a perfect example of this problem and it is a proof that the
correlation algorithm is inaccurate for q ≤ 0. Therefor we will focus on the
analysis of the generalized dimension for q > 0.
56
5.2.3 Statistical analysis
In table 5.2 we can see features extracted from ISI spike trains by conventional
statistical methods. For description of the procedure of spike detection see
section 3.2. The extraction threshold used was 3.3 and the minimal number
of spikes to form a neuron during the clustering procedure was 100. The
procedure of spike detection and spike sorting can be see in Fig. 5.18. In
Fig. In Fig. 5.20 is an example of extracted spike train. From the table 5.2
can be seen that the parameters differ for different parts of the human brain.
Figure 5.18: Spike train detection.
Figure 5.19: Example, signal 18, patientP. Red line is the extraction thresh-
old.
5.2.4 Generalized dimensions from measured data

To see if there is a correlation between the area of the brain in which the
signal was recorded and the generalized dimension of this signal we have
analyzed many different signals from different patients. Here we will show
the analysis of four signals from patient P. The rest of results is in appendix
C in Figs. C.2, C.1, C.3, C.4.
In Fig. 5.21 we can see the correlation sums and generalized dimensions
for q = {2, 3, 4, 5}. However no difference can be seen for different signals.
The correlation dimension D2 that is approximately equal to the embed-
ding dimension d = 4. That is another sign that these results are same as for
white noise, thus they do not contain useful information about the signals.
57
Figure 5.20: Example, the spike train extracted using the QSort algorithm
from the signal 18, patientP.
Table 5.2: Features extracted from ISI spike trains by conventional statistical
methods. Features collected from 4 spike train data sets: number of spikes
detected, mean ISI interval and its standard deviation, coefficient of vari-
ance, percentage of ISI shorter than 3 ms, mean frequency and its standard
deviation, occurrence of a burst.
Dataset 6 9 13 26
part of brain Th n/a STN SNr
# of spikes 73 145 193 281
ISI (ms) 123.4 68.4 51.4 34.8
σ ISI (ms) 200.9 56.2 66.0 32.6
CV 1.6 0.8 1.3 1.0
< 3 ms (%) 0 0 3.1 20.0
f¯ (Hz) 8.1 14.6 19.5 28.7
σ f (Hz) 5.0 17.8 15.1 30.7
burst true 1 0 0 0
58
(a) scaling, signal 6 (b) Dq ’s, signal 6
(c) scaling, signal 9 (d) Dq ’s, signal 9
(e) scaling, signal 13 (f) Dq ’s, signal 13
(g) scaling, signal 26 (h) Dq ’s, signal 26
Figure 5.21: Generalized correlation sums and Dq ’s for signals 6, 9, 13, 26

for patient P.
59
To verify whether the results come from a non-linear dynamical system or
not we will perform a standard test using the same analysis on surrogate
data. Surrogate data is random data generated to have the same mean, vari-
ance, and autocorrelation function as the original data. Further we generated
”amplitude-adjusted” surrogate data, where the histogram of the surrogate
data is the same as that of the original data. To generate these data we used
Matlab scripts from [17]. The surrogate and amplitude adjusted surrogate
data for one signal are shown in Fig 5.22.
Figure 5.22: Plain signal, surrogate data and amplitude adjusted surrogate
data (top to bottom) for signal 13, patient P.
(a) scaling (b) Dq ’s
Figure 5.23: Generalized correlation sums and Dq ’s for surrogate data from
signal 13, patient P.
The results for surrogate data for signal 13, patient P are in Fig. 5.23
and the results for amplitude adjusted surrogate data are in Fig. 5.24. We
60
(a) scaling (b) Dq ’s
Figure 5.24: Generalized correlation sums and Dq ’s for amplitude adjusted

surrogate data from signal 13, patient P.
can see that the surrogate data and amplitude adjusted surrogate data have
same characteristics as the measured signal in figure 5.21f and 5.21e. We
used the test with surrogate and amplitude adjusted surrogate data also for
signals 6,9 and 26, for results see. Appendix C figures C.5 and C.6.
The fact that the results of the fractal analysis of surrogate data is not
differentiable from the results of the real signal is another sign that this
analysis is not able to describe the complexity of our microrecording signals
of 10 seconds in duration.
5.2.5 Analyzing the spike train

Because the analysis of raw signals did not supply any relevant results we
tried to analyze the detected spike trains. The spike train extracted from
our data by the Quick sort algorithm [21] was another data set that has been
analyzed. For description of the procedure see section 3.2. Example of a
detected spike train can be seen in Fig. 5.20. In the State of art search
3.1 we found several studies that analyzed the neural spike trains [7] [2] [27]
[26]. However these spike trains had some thousand samples. Because of the
limited length of our signals, the extracted spike train consisted of maximum
approximately 350 samples. This might not be enough to perform a fractal
analysis. Fig. 5.25 shows the results of the Cao’s method for the spike train.
This graph does not show correlation between E1 and E2 neither saturation
of E1. This means the spike train behaves as a noisy signal. However Cao
in his work [5] used at least 1000 samples to get somewhat useful result and
only for at least 5000 samples the results where clear. This means that these
results can be spurious, but we are not able to decide because the signal is
too short.
61
Figure 5.25: Cao’s method functions E1 and E2 for a spike train extracted
from our signal.
62
Chapter 6
Conclusion
We studied the theory of multifractal analysis of time series. First we de-

scribed a way of reconstructing the attractor in state space using time delay
embedding. Than we discussed the correlation dimension estimated from the
correlation sum. And we expanded the correlation sum to the generalized
correlation sum. We introduced the generalized dimensions as a measure of
multifractality estimated from the generalized correlation sum. And we also
explained the meaning of the spectrum of scaling indices and some significant
points on its graph, for example the peak as the capacity dimension q = 0
and the intersection of the graph with the α axis as the value of D−∞ and
D∞ .
We described the procedure of reconstructing the attractor in state space
and the algorithms used. We discussed algorithms for calculating the correla-
tion sum. And we expanded the idea on generalized correlation sum. We also
mentioned the k-nearest neighbor algorithm for estimating the generalized
dimension. And we described the algorithm of constructing the spectrum of
scaling indices from the generalized dimensions.
We implemented the algorithms partly in Matlab scripting language and
partly in C++. The C++ implementations were compiled as mex functions
for direct use in the Matlab environment. We evaluated the implementations
on artificial data. The implementations performed good, with limitations
when estimating generalized dimensions q for q < 0. But these limitations
were expected.
Then we used the programs and functions to analyze the microrecording
data. We reconstructed the attractor into the embedding state space. We
performed the multifractal analyzes on several time series from different pa-
tients. No relevant results have been found. We used the surrogate data to
test whether the multifractal analysis can describe the microrecording signal.
The test produced the same results for surrogate data and for the real signal.
63
From that we concluded that the low dimensional chaos cannot describe the
dynamical system of neurons in deep structures of human brain
We tried to analyze the detected spike trains for the used data of 10 sec-
onds duration. The spike train was rather too short for any analysis. We
were not able to determine the right parameters for the state space recon-
struction using time delay embedding. We can not make any conclusion from
that since the analyzed spike train was too short.
6.1 Future work

In future we will focus on the analysis of the spike trains. For this analysis
we will acquire longer data from experimental animals. The duration of the
signal would be probably rather hours than seconds.
For the study of microrecordings from human brain statistical analysis
is more suitable. We will use the statistical analysis on the whole database
consisting of hundreds of signals. The results could than have high signifi-
cance.
64
Bibliography
[1] H.D.I. Abarbanel, Analysis of observed chaotic data, Springer verlag,

1996.
[2] A. Bershadskii, E. Dremencov, D. Fukayama, and G. Yadid, Multifractal

statistics and underlying kinetics of neuron spike time-series, Physics
Letters A 289 (2001), no. 6, 337–42.
[3] F. Blair (ed.), Deep brain stimulation for parkinson’s disease, Parkin-
son’s Disease Foundation, 2007.
[4] A. Bueno-Orovio and V. M. Perez-Garcia, Enhanced box and prism as-

sisted algorithms for computing the correlation dimension, Chaos, Soli-
tons and Fractals 34 (2007), 509–518.
[5] L. Cao, Practical method for determining the minimum embedding di-
mension of a scalar time series, Physica D 110 (1997), no. 1-2, 43–50.
[6] R. Dominguez-Tenreiro, L.J. Roy, and V.J. Martinez, On the multifrac-

tal character of the lorenz attractor, Progress of Theoretical Physics 87
(1992), 1107–1118.
[7] A. Eblen-Zajjur, R. Salas, and H. Vanegas, Fractal analysis of spinal

dorsal horn neuron discharges by means of sequential fractal dimension
d, Computers in Biology and Medicine 26 (1996), no. 1, 87–95.
[8] A.M. Fraser and H.L. Swinney, Independent coordinates for strange at-
tractors from mutual information, Phys. Rev. A 33 (1986), no. 2, 1134–
1140.
[9] P. Grassberger, Generalized dimensions of strange attractors, Physics

Letters A 97 (1983), 227–230.
[10] P. Grassberger and I. Procaccia, Characterization of strange attractors,

Physical Review Letters 50 (1983), no. 5, 346–349.
65
[11] E.G. Gwinn and R.M. Westervelt, Scaling structure of attractors at the
transition from quasiperiodicity to chaos in electronic transport in ge,
Phys. Rev. Lett. 59 (1987), no. 2, 157–160.
[12] T.C. Halsey, M.H. Jensen, L.P. Kadanoff, I. Procaccia, and B.I.
Shraiman, Fractal measures and their singularities: The characteriza-
tion of strange sets, Physical Review A 33 (1986), no. 2, 1141–1151.
[13] R. Hegger, H. Kantz, and T. Schreiber, Practical implementation of

nonlinear time series methods: The tisean package, Chaos 9 (1999),
no. 2, 413–435.
[14] H.G.E. Hentschel and I. Procaccia, The infinite number of generalized

dimensions of fractals and strange attractors, Physica D 8D (1983),
no. 3, 435–44.
[15] R. C. Hilborn, Chaos and nonlinear dynamics, second ed., ch. 9 Quan-
tifying Chaos, pp. 319–374, Oxford University Press, 2000.
[16] R.C. Hilborn, Chaos and nonlinear dynamics, second ed., ch. 10 Many
Dimensions and Multifractals, pp. 375–430, Oxford University Press,
2000.
[17] D. Kaplan, Matlab software for nonlinear time-series analysis, 1997,

Macalester College , http://www.macalester.edu/ kaplan/software/.
[18] M.L. Kringelbach, N. Jenkinson, S.L. Owen, and T.Z. Aziz, Transla-
tional principles of deep brain stimulation, National Reviews Neuro-
science 8 (2007), no. 8, 623–635.
[19] Ch. Merkwirth, U. Parlitz, I. Wedekind, and W. Lauter-

born, Tstool version 1.11, December 2001, DPI Göttingen,
http://www.physik3.gwdg.de/tstool/.
[20] NINDS Office of Communications and Public Liaison, Parkinson’s dis-

ease: Hope through research, January 2006, NIH Publication No. 06-139.
[21] U. Rutishauser, E.M. Schuman, and A.N. Mamelak, Online detection

and sorting of extracellularly recorded action potentials in human medial
temporal lobe recordings, in vivo, Journal of Neuroscience Methods 154
(2006), no. 1-2, 204–224.
[22] S.K. Sarkar, Multifractal description of singular measures in dynamical

systems, Phys. Rev. A 36 (1987), no. 8, 4104–4106.
66
[23] T Sauer and J.A. Yorke, How many delay coordinates do you need?,
International Journal of Bifurcation and Chaos 3 (1993), no. 3, 737–
744.
[24] K.V. Slavin, K.R. Thulborn, C. Wess, and H. Nersesyan, Direct visual-
ization of the human subthalamic nucleus with 3t mr imaging, American
Journal of Neuroradiology 27 (2006), no. 1, 80–84.
[25] F. Takens, Dynamical systems and turbulence, Lecture Notes in Mathe-

matics, vol. 898, pp. 366–381, Springer Berlin, 1981.
[26] M.C. Teich, C. Heneghan, S.B. Lowen, T. Ozaki, and E. Kaplan, Fractal
character of the neural spike train in the visual system of the cat, Journal
of the Optical Society of America A (Optics, Image Science and Vision)
14 (1997), no. 3, 529–46.
[27] M.C. Teich and S.B. Lowen, Fractal patterns in auditory nerve-spike
trains, IEEE Engineering in Medicine and Biology Magazine 13 (1994),
no. 2, 197–202.
[28] Y. Termonia and Z. Alexandrowicz, Fractal dimension of strange attrac-

tors from radius versus size of arbitrary clusters, Phys. Rev. Lett. 51
(1983), no. 14, 1265–1268.
[29] J. Theiler, Estimating fractal dimension, Journal of the Optical Society

of America A 7 (1990), no. 6, 1055–1073.
[30] W. van de Water and P. Schram, Generalized dimensions from near-

neighbor information, Phys. Rev. A 37 (1988), no. 8, 3118–3125.
67
Appendices
68
Appendix A
State of Art
A.1 Fractal analysis of spinal dorsal horn neu-

ron discharges by means of sequential
fractal dimension D
Eblen-Zajjur, A. and Salas, R. and Vanegas, H., Fractal analysis of spinal
dorsal horn neuron discharges by means of sequential fractal dimension D,
Computers in Biology and Medicine 26, (1996), no. 1, 87-95.
Abstract
Describes a new methods for converting a typical point process, such as a
train of neuronal action potentials (spikes), into a planar curve which is
then processed by means of a fast algorithm to calculate and display the
fractal dimension D values of each of a sequence of blocks having an equal
and preselectable number of interspike intervals, hence the term sequential
fractal dimension D (SFD). This method is fast, does not require special
computing facilities, and provides a continuous, high temporal resolution
display of the neuronal discharge complexity along the course of spontaneous
activity or event relating changes. The method affords insight into short
duration changes in neuronal behaviour in a way independent of its discharge
rate. SFD analysis of spike trains from spinal dorsal horn neurons suggests
that the neuronal response to a given stimulus can be expressed as changes
in the discharge pattern complexity, thus revealing a novel sensory coding
strategy. (28 References)
69
A.2 Multifractal statistics and underlying ki-
netics of neuron spike time-series
Bershadskii, A. and Dremencov, E. and Fukayama, D. and Yadid, G., Multi-
fractal statistics and underlying kinetics of neuron spike time-series, Physics
Letters A 289, (2001), no. 6, 337-42.
Abstract
Probabilistic and multifractal properties of spiking time-series obtained in
vivo from singular neurons belonging to red nuclei from a rat’s brain are
analyzed. Lognormal and −1 power-law probability distributions of inter-
spike intervals are observed for healthy and for genetically depressive rats,
respectively. A simple thermodynamic model is elaborated to interpret the
obtained results. Investigations of long-range interspike correlations (both
probabilistic and multifractal) gives indications that the genetically defined
depression is related to individual neuron kinetic problems rather than to
brain system disorder. (33 References)
A.3 Fractal patterns in auditory nerve-spike

trains
Teich, M.C. and Lowen, S.B., Fractal patterns in auditory nerve-spike trains,
IEEE Engineering in Medicine and Biology Magazine 13, (1994), no. 2,
197-202.
Abstract
The authors discuss the following topics: traditional renewal point process
models; long-term correlations; self-similarity of neuronal firing rates; power-
law behavior of the Fano-factor time curve; change in the firing pattern
induced by the presence of a stimulus; neural information processing with
fractal nerve spikes; biophysical origin of the fractal behavior; fractal point-
process model; short-term correlations. (33 References)
70
A.4 Fractal character of the neural spike train
in the visual system of the cat
Teich, M.C. and Heneghan, C. and Lowen, S.B. and Ozaki, T. and Kaplan,
E., Fractal character of the neural spike train in the visual system of the
cat, Journal of the Optical Society of America A (Optics, Image Science and
Vision) 14, (1997), no. 3, 529-46.
Abstract
The authors used a variety of statistical measures to identify the point pro-
cess that describes the maintained discharge of retinal ganglion cells (RGC’s)
and neurons in the lateral geniculate nucleus (LGN) of the cat. These mea-
sures are based on both interevent intervals and event counts and include the
interevent-interval histogram, rescaled range analysis, the event-number his-
togram, the Fano factor, the Allan factor, and the periodogram. In addition,
the authors applied these measures to surrogate versions of the data, gener-
ated by random shuffling of the order of interevent intervals. The counting
statistics reveal 1/f-type fluctuations in the data (long-duration power-law
correlation), which are not present in the shuffled data. Estimates of the frac-
tal exponents measured for RGC- and their target LGN-spike trains are sim-
ilar in value, indicating that the fractal behavior either is transmitted from
one cell to the other or has a common origin. The gamma-r renewal process
model, often used in the analysis of visual-neuron interevent intervals, de-
scribes certain short-term features of the RGC and LGN data reasonably well
but fails to account for the long-duration correlation. The authors present a
new model for visual-system nerve-spike firings: a gamma-r renewal process
whose mean is modulated by fractal binomial noise. This fractal, doubly
stochastic point process characterizes the statistical behavior of both RGC
and LGN data sets remarkably well. (84 References)
A.5 Characterization of non-linear dynamics

in rat cortical neuronal networks
Boom, H. and Robinson, C. and Rutten, W. and Neuman, M. and Wijkstra,
H. (eds., Characterization of non-linear dynamics in rat cortical neuronal
networks, IEEE , New York, NY, USA, 1997, Proceedings of the 18th Annual
International Conference of the IEEE Engineering in Medicine and Biology
Society. ‘Bridging Disciplines for Biomedicine’.
71
Abstract
In this paper we report on the evidence for low-dimensional determinis-
tic chaos in extra-cellular recordings of mature rat cortical cells in culture
medium. In general, the available data sets are relatively short and are
heavily contaminated by noise making conclusive interpretations of chaotic
behaviour difficult to ascertain. In this study we describe two analysis tech-
niques used to detect chaotic behaviour, namely return map construction and
Lyapunov exponent calculation. Results show that both methods indicate
the presence of chaos which is consistent with other recent studies that have
also suggested chaotic behaviour in cultured cell networks. In addition, the
use of these two independent analysis techniques provides more rigorous ev-
idence for the existence of chaotic behaviour in cultured cell networks since
previous studios have relied solely on the method of return maps. (8 Refer-
ences)
A.6 Bursting, spiking, chaos, fractals and uni-

versality in biological rhythms
Chay, T.R. and Yin, S.F. and Young, S.L., Bursting, spiking, chaos, fractals,
and universality in biological rhythms, International Journal of Bifurcation
and Chaos in Applied Sciences and Engineering 5, (1995), no. 3, 595-635.
Abstract
Biological systems offer many interesting examples of oscillations, chaos, and
bifurcations. Oscillations in biology arise because most cellular processes con-
tain feedbacks that are appropriate for generating rhythms. These rhythms
are essential for regulating cellular function. In this tutorial review, we treat
two interesting nonlinear dynamic processes in biology that give rise to burst-
ing, spiking, chaos, and fractals: endogenous electrical activity of excitable
cells and Ca2+ releases from the Ca2+ stores in nonexcitable cells induced
by hormones and neurotransmitters. We will first show that each of these
complex processes can be described by a simple, yet elegant, mathematical
model. We then show how to utilize bifurcation analyses to gain a deeper
insight into the mechanisms involved in the neuronal and cellular oscillations.
With the bifurcating diagrams, we explain how spiking can be transformed
to bursting via a complex type of dynamic structure when the key param-
eter in the model varies. Understanding how this parameter would affect
the bifurcation structure is important in predicting and controlling abnor-
72
mal biological rhythms. Although we describe two very different dynamic
processes in biological rhythms, we will show that there is universality in
their bifurcation structures. (84 References)
A.7 Multiplicative multifractal modeling and

discrimination of human neuronal activ-
ity
Jianbo, G. and Yi, Z. and Sanchez, J.C. and Principe, J.C. and Okun, M.S.,
Multiplicative multifractal modeling and discrimination of human neuronal
activity, Physics Letters A 344, (2005), no. 2-4, 253-64.
Abstract
Understanding neuronal firing patterns is one of the most important problems
in theoretical neuroscience. It is also very important for clinical neurosurgery.
In this Letter, we introduce a computational procedure to examine whether
neuronal firing recordings could be characterized by cascade multiplicative
multifractals. By analyzing raw recording data as well as generated spike
train data from 3 patients collected in two brain areas, the globus pallidus
externa (GPe) and the globus pallidus interna (GPi), we show that the neural
firings are consistent with a multifractal process over certain time scale range
(t1 ,t2 ), where t1 is argued to be not smaller than the mean inter-spike-interval
of neuronal firings, while t2 may be related to the time that neuronal signals
propagate in the major neural branching structures pertinent to GPi and
GPe. The generalized dimension spectrum Dq effectively differentiates the
two brain areas, both intraand inter-patients. For distinguishing between
GPe and GPi, it is further shown that the cascade model is more effective
than the methods recently examined by Schiff et al. as well as the Fano factor
analysis. Therefore, the methodology may be useful in developing computer
aided tools to help clinicians perform precision neurosurgery in the operating
room. [All rights reserved Elsevier]. (37 References)
73
Appendix B
Source codes
B.1 embed.m
function cout=embed(cin, dim, delay, shift, windowtype)
N = length(cin);
M = floor((N-1-(dim-1)*delay)/shift)+1;
if M < 1
error(’time series to short for chosen embedding parameters’)
end
d = zeros(M, dim);
len = (M-1)*shift+1;
for i=1:dim
start = (i-1)*delay;
d(:,i) = cin(start+1:shift:start+len);
end
if ~strcmp(windowtype, ’Rect’)
d = d .* repmat(window(dim, windowtype)’, M, 1);
end
cout=d;
74
B.2 findembedparam.m
function [tau,dim,E1,E2]= findembedparam(ts,tau)
maxtau=40;
partitions=128;
[y,i] = sort(ts);
N=length(y);
y(i) = (0:N-1)/N;
a = amutual(y, maxtau, partitions);
da=diff(a);
if nargin<2 || isempty(tau)==1
tau=find(da>=0,1,’first’)-1;
end
figure; plot(1:maxtau,a(2:end),’k’,tau,a(tau+1),’kx’);
title(’Auto mutual information function’);
maxdim=9;
Nref=5000;
NNR=3;
eData = embed(ts, maxdim+2, tau,1,’Rect’);
L = length(eData);
ref = randref(2, L-2, Nref);
[E, Estar] = cao(eData, ref, NNR);
E1=E(2:end)./E(1:end-1);
E2=Estar(2:end)./Estar(1:end-1);
ddE1=diff(diff(E1));
[dummy dim]=min(ddE1);
dim=dim+1;
figure; plot(E1,’kx-’); hold on; plot(E2,’ko:’);
title(’Cao’’s method results’);
75
B.3 knnGolden.m
%% create simulated time series
N=10000;
Hen=henon(N);
Lor=lorentz(N,0,10,28,8/3);
%% Embedding
eHen = embed(Hen, 2, 1 ,1,’Rect’);
eLor = embed(Lor, 3, 15,1,’Rect’);
%% Random reference indices

n=3000;
ref = sort(randsample(N-100, n));
%% gedimest for Henon map

[nn, dist] = fnearneigh(eHen, ref, 128, 0);
gammas = -7.5:0.5:10;
gedimsHen = gendimest(dist, gammas, 10,10, 128);
qsHen=1-gammas./gedimsHen’;
figure; plot(qsHen, gedimsHen); hold on;

xlabel(’q’);ylabel(’D_q’);
[dummy mHen]=max(gedimsHen);
[dummy nHen]=min(gedimsHen);
plot(qsHen([mHen nHen]), gedimsHen([mHen nHen]),’x’)
%% gedimest for Lorenz attractor

[nn, dist] = fnearneigh(eLor, ref, 128, 0);
gammas = -18:0.5:25;
gedimsLor = gendimest(dist, gammas, 10,10, 128);
qsLor=1-gammas./gedimsLor’;
figure; plot(qsLor, gedimsLor); hold on;

xlabel(’q’);ylabel(’D_q’);
[dummy mLor]=max(gedimsLor);
[dummy nLor]=min(gedimsLor);
plot(qsLor([mLor nLor]), gedimsLor([mLor nLor]),’x’)
76
B.4 multifrac.m
function [D, alfa2, f2]=multifrac(data ,dim,delay)
%% Embedding
edata=embed(data, dim, delay,1,’Rect’);
atria=nn_prepare(edata);
%% Calculate Dq
numofqs=13; % number of q’s
qzero=ceil(numofqs/2);
tau=zeros(numofqs,1);
stdx=zeros(numofqs,1);
mse=zeros(numofqs,1);
stdD=zeros(numofqs,1);
D=ones(numofqs,1);
n=5000;
[N,dim] = size(edata);
ref = sort(randsample(N, n));
for q=8:numofqs
if q~=(qzero+1)
[logC1 logR1]=...
gencorrsum(atria, edata,ref,1,50,(q-qzero),32,2);
logC=log10(logC1(logC1~=0))/(q-qzero-1);
logR=log10(logR1(logC1~=0));
lR=min(logR); hR=max(logR); lC=min(logC)+1;
figure; plot(logR,logC); axis([lR hR lC 0]);
ScalingRegion=find((logR>a)&(logR<b));
[limtau, limstdx, limmse]= ...
lscov([logR(ScalingRegion),...
ones(length(ScalingRegion),1)], logC(ScalingRegion));
D(q)=limtau(1); stdx(q)=limstdx(1); mse(q)=limmse(1);
stdD(q)=sqrt(mse(q));
tau(q)=D(q)*(q-qzero-1);
hold on;
plot(logR(ScalingRegion),logR(ScalingRegion)*tau(q),’r’);
end
end
77
D(qzero+1)=mean([D(qzero) D(qzero+2)]);
%% Calculate Dinf
[logC1 logR1]=maxpcorr(atria, edata,ref,1,50,1,32,2);
logCx=log10(logC1(logC1~=0));
logRx=log10(logR1(logC1~=0));
figure; plot(logRx,logCx); axis([-6 1 -5 0]);
ScalingRegion=find((logRx>a)&(logRx<b));
[limtau, limstdx, limmse]=...
lscov([logRx(ScalingRegion),...
ones(length(ScalingRegion),1)],logCx(ScalingRegion));
Dinf=limtau(1); stdx=limstdx(1); mse=limmse(1);
hold on; plot((a:0.1:b),(a:0.1:b)*Dinf,’r’);
stdDinf=sqrt(mse);
%% Calculate D-inf
[logC1 logR1]=minpcorr(atria, edata,ref,1,50,1,32,2);
logCn=log10(logC1(logC1~=0));
logRn=log10(logR1(logC1~=0));
figure; plot(logRn,logCn); axis([-6 1 -5 0]);
ScalingRegion=find((logRn>a)&(logRn<b));
[limtau, limstdx, limmse]=...
lscov([logRn(ScalingRegion),...
ones(length(ScalingRegion),1)],logCn(ScalingRegion));
D_inf=limtau(1); stdx=limstdx(1); mse=limmse(1);
hold on; plot((a:0.1:b),(a:0.1:b)*D_inf+limtau(2)+0.2,’r’);
stdD_inf=sqrt(mse);
%% Spectrum of scaling indices f(alfa)

alfa=diff(tau)’;
f=(-(qzero-1):(q-qzero-1)).*alfa-tau(1:numofqs-1)’;
alfa=[D_inf alfa Dinf];
f=[0 f 0];
alfa2=[D_inf:-0.001:Dinf];
f2=interp1(alfa, f, alfa2, ’spline’);
%% Plot results
78
range=(-(qzero-1):(q-qzero));
figure(’Name’, ’tau’)
plot(range,tau);
ylabel(’tau’); xlabel(’q’)
figure(’Name’, ’Dq’)
plot(range,D, range,(D+stdD), range,(D-stdD));
ylabel(’Dq’); xlabel(’q’)
figure(’Name’, ’f(alfa)’)
plot(alfa(2:end-1),f(2:end-1),’x’);hold on;
plot(alfa2,f2,’r’)
plot(alfa2([1,end]),[0 0], ’or’);
xlabel(’alfa’); ylabel(’f’); hold on;
axis([min(alfa2)-0.1 max(alfa2)+0.1 -0.05 max(f2)+0.1]);
plot([min(alfa2)-0.1 max(alfa2)+0.1],[0 0],’k’);
text(Dinf+0.01,0.04,[’D _{\infty}=’ num2str(Dinf,4)])
text(D_inf-0.15,0.04,[’D _{-\infty}=’ num2str(D_inf,4)])
79
Appendix C
Generalized dimensions
analysis
(a) scaling, signal 21 patient P (b) Dq ’s, signal 21 patient P
(c) scaling, signal 27 patient P (d) Dq ’s, signal 27 patient P
Figure C.1: Generalized correlation sums and Dq ’s for signals 21 and 27 from
patient P.
80
(a) scaling, signal 5 patient P (b) Dq ’s, signal 5 patient P
(c) scaling, signal 15 patient P (d) Dq ’s, signal 15 patient P
(e) scaling, signal 17 patient P (f) Dq ’s, signal 17 patient P
(g) scaling, signal 18 patient P (h) Dq ’s, signal 18 patient P
Figure C.2: Generalized correlation sums and Dq ’s for signals 11, 15, 17, 28
from patient P.
81
(a) scaling, signal 7 patient J (b) Dq ’s, signal 7 patient J
(c) scaling, signal 9 patient J (d) Dq ’s, signal 9 patient J
(e) scaling, signal 13 patient J (f) Dq ’s, signal 13 patient J
(g) scaling, signal 25 patient J (h) Dq ’s, signal 25 patient J

from patient J.
82
(a) scaling, signal 5 patient K (b) Dq ’s, signal 5 patient K
(c) scaling, signal 10 patient K (d) Dq ’s, signal 10 patient K
(e) scaling, signal 17 patient K (f) Dq ’s, signal 17 patient K
(g) scaling, signal 27 patient K (h) Dq ’s, signal 27 patient K
from patient K.
83
Figure C.5: Generalized correlation sums and Dq ’s for surrogate data from
different signals.
84
Figure C.6: Generalized correlation sums and Dq ’s for amplitude adjusted

surrogate data from different signals.
85

Matlab Coding

Загружено:

Сведения о документе

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Matlab Coding

Загружено:

Авторское право:

Доступные форматы

Czech Technical University in Prague

Faculty of Electrical Engineering

Processing of neuronal single-cell recordings

Supervisor: Ing. Daniel Novák Ph.D.

The micrororecording signal from deep structures of human brain was

Analyzovali jsme signál z mikroregistrace z hlubokých struktur lidského

Prague, Czech Republic Karel Schmidt

C Generalized dimensions analysis 80

3.1 Auto mutual information of the Lorentz Attractor, τs = 15. . . 18

4.1 Complete process of multifractal analysis. . . . . . . . . . . . . 38

5.1 Attractor of the Henon map reconstructed from its x-component. 40

C.1 Generalized correlation sums and Dq ’s for signals 21 and 27

DBS deep brain stimulation.

GPi globus pallidus interna.

IPG implantable pulse generator.

ISI inter-spike interval.

MRI magnetic resonance imaging.

SNr substantia nigra.

STN subthalamic nucleus.

a(i, d) ratio of d + 1 to d dimensional reconstructed space distance between

α scaling index of generalized dimensions

By pointwise mass function, numerically equal to Pi

C(r) correlation sum

Cq generalized correlation sum

Dq generalized dimension of chaotic attractor

E(d) Cao’s method first quantity

E ∗ (d) Cao’s method second quantity

E1(d) variation of Cao’s method first quantity

E2(d) variation of Cao’s method second quantity

f (α) spectrum of scaling indices

n() minimum number of open sets of diameter  needed to cover a fractal

Pi probability for a random point to be ith box on the attractor, numer-

r search radius, the main parameter of the correlation sum

τ measure of generalized dimensions Dq

τs time delay between sampled values in time delay embedding

Θ(x) Heaviside step function

xi time series sample

yi the embedded time delay vector

h·i denotes expectation value of a function, for uniformly distributed pa-

2.1 Parkinson’s disease

2.1.2 Parkinson’s disease treatment

2.2 Deep brain stimulation

3.1 State of Art

• spike train, time series

• single cell, neuron

3.2 Spike sorting

3.3 Fractal analysis

3.3.1 Time delay embedding

3.3.2 Estimate the time delay

yi (d) = (xi , xi+τs , ..., xi+(d−1)τs ), i = 1, 2, ..., N − (d − 1)τs ,

where d is the embedding dimension and τs is the time delay between

3.3.3 Estimate embedding dimension

where k·k is some measurement of Euclidian distance and is given in this

kyk (m) − yl (m)k = max |xk+jτs − xl+jτs | ;

yi (d + 1) is the ith reconstructed vector with embedding dimension d + 1.

E(d) is dependent only on the dimension d and the time delay τs . To

E1(d) = E(d + 1)/E(d) (3.4)

We found that when increased by steps of 1 E1(d) stops changing when

E2(d) = E ∗ (d + 1)/E ∗ (d) (3.6)

3.3.4 Correlation dimension

where yi are m-dimensional delay vectors. Npairs = (N − d − w)(N −

Theiler suggested to remove this spurious effect by simply ignoring all

3.3.5 Multifractal spectrum of generalized dimensions

n() minimum number of open sets of diameter needed to cover a fractal