Nonlinear Dynamics

Nonlinear Dynamics

Edited by
Todd Evans
Intech
IV















Published by Intech


Intech
Olajnica 19/2, 32000 Vukovar, Croatia

Abstracting and non-profit use of the material is permitted with credit to the source. Statements and
opinions expressed in the chapters are these of the individual contributors and not necessarily those of
the editors or publisher. No responsibility is accepted for the accuracy of information contained in the
published articles. Publisher assumes no responsibility liability for any damage or injury to persons or
property arising out of the use of any materials, instructions, methods or ideas contained inside. After
this work has been published by the Intech, authors have the right to republish it, in whole or part, in
any publication of which they are an author or editor, and the make other personal use of the work.

2010 Intech
Free online edition of this book you can find under www.sciyo.com
Additional copies can be obtained from:
[email protected]

First published January 2010
Printed in India

Technical Editor: Teodora Smiljanic

Nonlinear Dynamics, Edited by Todd Evans
p. cm.
ISBN 978-953-7619-61-9










Preface

This volume covers a diverse collection of topics dealing with some of the fundamental
concepts and applications embodied in the study of nonlinear dynamics. Each of the 15
chapters contained in this compendium generally fit into one of five topical areas: physics
applications, nonlinear oscillators, electrical and mechanical systems, biological and
behavioral applications or random processes. The authors of these chapters have
contributed a stimulating cross section of new results, which provide a fertile spectrum of
ideas that will inspire both seasoned researches and students.

Editor
Todd Evans
General Atomics
United States












Contents

Preface V

1. Nonlinear Absorption of Light in Materials with Long-lived Excited States 001

Francesca Serra and Eugene M. Terentjev


2. Exact Nonlinear Dynamics in Spinor Bose-Einstein Condensates 031

Junichi Ieda and Miki Wadati


3. A Conceptual Model for the Nonlinear Dynamics of Edge-localized
Modes in Tokamak Plasmas
059

Todd E. Evans, Andreas Wingen, Jon G. Watkins and Karl Heinz Spatschek


4. Nonlinear Dynamics of Cantilever Tip-Sample Surface
Interactions in Atomic Force Microscopy
079

John H. Cantrell and Sean A. Cantrell


5. Nonlinear Phenomena during
the Oxidation and Bromination of Pyrocatechol
109

Takashi Amemiya and Jichang Wang


6. Dynamics and Control of Nonlinear Variable Order Oscillators 129

Gerardo Diaz and Carlos F. M. Coimbra




7. Nonlinear Vibrations of Axially Moving Beams 145

Li-Qun Chen




8. The 3D Nonlinear Dynamics of Catenary Slender Structures
for Marine Applications
173

Ioannis K. Chatjigeorgiou and Spyros A. Mavrakos




9. Nonlinear Dynamics Traction Battery Modeling 199

Antoni Szumanowski

VIII
10. Entropic Geometry of Crowd Dynamics 221

Vladimir G. Ivancevic and Darryn J. Reid




11. Nonlinear Dynamics and Probabilistic Behavior in Medicine:
A Case Study
265

H. Nicolis




12. The Effect of Spatially Inhomogeneous Electromagnetic Field
and Local Inductive Hyperthermia on Nonlinear Dynamics
of the Growth for Transplanted Animal Tumors
285

Valerii Orel and Andriy Romanov




13. Advanced Computational Approaches for Predicting Tourist Arrivals:
the Case of Charter Air-Travel
309

Eleni I. Vlahogianni, Ph.D. and Matthew G. Karlaftis, Ph.D.




14. A Nonlinear Dynamics Approach
for Urban Water Resources Demand Forecasting and Planning
325

Xuehua Zhang, Hongwei Zhang and Baoan Zhang




15. A Detection-Estimation Method for Systems with Random Jumps
with Application to Target Tracking and Fault Diagnosis
343

Yury Grishin and Dariusz Janczak


1
Nonlinear Absorption of Light in Materials with
Long-lived Excited States
Francesca Serra and Eugene M. Terentjev
University of Cambridge
United Kingdom
1. Introduction
The absorption of light is an important phenomenon which has many applications in all the
natural sciences. One can say that all the chemical elements, molecules, complex substances,
and even galaxies, have their own fingerprint in the light absorption spectrum, as a
consequence of the allowed transitions between all electronic and vibronic levels.
The UV-Visible (UV-Vis) light (200-800 nm) has an energy comparable to that typical of the
transitions between the electrons in the outer shells or in molecular orbitals. Each atom has a
fixed number of atomic levels, and therefore those spectra are composed of narrow lines,
corresponding to the transitions between these levels. When molecules and macromolecules
are considered, the absorption spectrum is no longer characterised by thin lines but by wide
absorption bands. This is due to the fact that the electronic levels are split in many
vibrational and rotational sub-levels, which increase in number with the increasing
complexity of the molecules. IR spectroscopy is often used to investigate these lower energy
modes, but for very complex biological molecules not even this technique can resolve each
line precisely because the energy split between the various levels is too small. One possible
way to obtain higher resolution spectra is to lower the sample temperature, in order to
suppress many of the vibrational and rotational modes. For biological molecules, though,
lowering the temperature can be a problem if one wants to study, for example, the activity
of enzimes, which only work at physiological temperatures. One of the advantages of
absorption spectroscopy (IR and UV-Vis) is to be a non-disruptive technique, also for
delicate molecules like polymers and biomolecules.
In the process of light absorption by molecules, once a photon with the right energy is
absorbed, the molecule goes into an excited state at higher energy [Born and Wolf 1999,
Dunning & Hulet 1996]. Eventually, it spontaneously returns to the ground state, but it can
relax following several mechanisms. When excited, the molecule reaches, in general, one of
the sub-levels of a higher electronic state. The first process is then, generally, a relaxation to
the lower energy state of that electronic level (schematised in figure 1). This process is
usually very fast (in the femtosecond scale) and not radiative. From this level, there are
several pathways to dissipate the energy: a radiative transition from the lower level of the
excited state to the ground state (fluorescence), accompanied by the emission of a photon at
lower energy than the absorbed one; a flip of the electronic spin, which leads to a transition
between singlet and triplet state (intersystem crossing), often associated with another
Nonlinear Dynamics

2

Fig. 1. A scheme representing some possibility of excitation/disexcitation of a molecule.
Each electronic level is split into many vibrational and rotational sub-levels. The blue arrow
describes the absorption of a photon, the green arrow the emission of a photon from the
lower energy level of the excited state (fluorescence), while the black arrows indicate all the
nonradiative energy dissipation mechanisms, which can be alternative to fluorescence. The
intersystem crossing is another mechanism of disexcitation: the triplet state is represented
with the red curve, and the transition with the thick arrow. The molecule can relax over long
time to the ground state either with a nonradiative process or via phosphorescence (red
arrow).
radiative process (phosphorescence); a non radiative decay where the energy is released by
heat dissipation. In some molecules the relaxation pathway following the excitation is more
complex, and it can involve interaction with other molecules. In such cases the energy can
be transferred to other molecules via radiative or non radiative processes: azobenzene, for
example, is a photosensitive molecule which, after excitation, undergoes a conformational
change; a more common molecule, like chlorophyll in plant cell chloroplasts, transfers the
excitation to the neighbouring molecules until the energy reaches the photosynthetic
complex where the photosynthesis takes place.
The common characteristic shared by fluorescent molecules, molecules with a triplet state
and photosensitive molecules like azobenzene, is that the lifetime of the excited state is long
compared to the time it takes for the excitation to occur. This brings us to the subject of this
chapter, which deals with a phenomenon, closely associated with the lifetime of the excited
state, which we called dynamic photobleaching. In general usage, the term
photobleaching has been taken to refer to permanent damaging of a chemical, generally
due to prolongued exposure to light. Here, we will not consider this, but rather a reversible
phenomenon whereby the number of molecules in the ground state is depleted as a
consequence of the long lifetime of the excited state.
This effect has important consequences for UV-Visible spectroscopy measurements. In
practical use, UVVis light absorption experiments are simple and straightforward: a
collimated beam of light is sent onto a sample, the transmitted light is collected by a
Nonlinear Absorption of Light in Materials with Long-lived Excited States

3
spectrometer and the ratio between the incident and the transmitted light is measured. Its
simplicity means that this technique is widely used in many areas of science. The
information one can get from these measurements concerns the allowed electronic
transitions. On the other hand, once the electronic structure of a substance is known,
computer simulations are able to reproduce absorption spectra.
A very common use of UV-vis spectroscopy is to measure the concentration of substances,
and this requires the celebrated Lambert-Beer (LB) law. This semi-empirical law states that
the light propagating in a thick absorbing sample is attenuated at a constant rate, that is,
every layer absorbs the same proportion of light [Jaffe & Orchin 1962]. This can be expressed
simply as the remaining light intensity at a depth x into the sample is: I(x) = I
0
exp(x/D)
where I
0
is the incident intensity and D is a characteristic length which is called the
penetration depth of a given material. If an absorbing dye is dispersed in a solution (or in
an isotropic solid matrix) this penetration depth is inversely proportional to the dye
concentration. In this way it is possible to determine a dye concentration c by
experimentally measuring the absorbance, defined as the logarithm of intensity ratio

(1)
where x is the thickness of the sample (the light path length), D is the penetration depth, c
the concentration of the chromophore, and the universal length scale characteristic of a
specific molecule/solvent. One should note that in chemistry and biology one often uses
base-10 logarithm in defining the Absorbance, rather than the more intuitive natural
logarithm. If c is in molar units, the constant of proportionality is the molar absorption
coefficient and it is inversely proportional to the characteristic length defined above.


Fig. 2. Schematic diagram of a typical measurement of light absorption. The amount of
absorbed light dI across the layer dx is proportional to the number of chromophores in that
volume.
The derivation of this empirical law is straightforward. It assumes that the fraction of light
absorbed by a thin layer of sample (thickness dx) is proportional to the number of molecules
it contains (see figure 2), expressed as the volume fraction n times the volume of the thin
layer (Area dx)

Nonlinear Dynamics

4
where I is the intensity of the incident light. Introducing the cross section , which is a
measure of the probability of a photon being absorbed by a chromophore, the differential
equation becomes

Solving the equation from 0 to x (total thickness of the sample), with a light I
0
incident on
the front of the sample, one has

and we obtain equation 1 (rearranging the units opportunely).
Thanks to the Lambert-Beer law, UV-visible absorption spectroscopy is a useful and
practical tool in many areas of science [Serdyuk et al. 2007]. The technique is widely used in
organic chemistry and biology, as macromolecules often have a characteristic absorption in
the UV and, more rarely, in the visible region of the EM spectrum. For example, all proteins
have a characteristic absorption band around 190nm, due to the molecular orbital formed by
the peptide bond, and another band around 280nm due to the aromatic side chains of
aminoacids. Usually, this band is used to determine the concentration of proteins in a
compound. Nucleic acids also absorb in the UV region and have a strong absorption band at
260 nm. The ratio between the absorption peak at 260 and 280 nm can give information
about the relative quantity of DNA and protein in a biological complex, like ribosome. In
atmospheric sciences, absorption spectroscopy is used to identify the composition of the air
[Heard 2006 ]. Because the concentration of the species is very low, the light path must be
very big to yield a detectable signal. Because L is so large and the concentration can change
over the long range, a generalised Lambert-Beer law is preferred:

where i is the absorption cross section of each species i. Visible absorption can even be
applied as a diagnostic tool. In medicine, for example, it is used to measure microvascular
hemoglobin oxygen saturation (StO2) in small, thin tissue volumes (like small capillaries in
the mouth) to identify ischemia and hypoxemia [Benaron et al. 2005].
All these applications rely on the validity of the LB law. However, this empirical law has
limitations, and deviations are observed due to aggregation phenomena or electrostatic
interactions between particles. The simpler form of the LB law also fails to describe the two-
photon absorption and the excited state absorption process, and it must be substituted by a
generalised Lambert-Beer law [Nathan et al. 1985]. These phenomena are usually present
only at very high incident light intensity. Also, highly scattering media, very relevant for the
medical and geological applications, produce large deviations from LB law.
This chapter addresses the topic of deviations from the LB law occurring in photosensitive
media due to self-induced transparency, or photobleaching [McCall & Hahn 1967,
Armstrong 1965]. This effect has been reported in a number of different biological systems
such as rhodopsin [Merbs & Nathans 1992], green fluorescent protein [Henderson et al.
2007] and light harvesting complexes [Bopp et al. 1997] stimulated with strong laser
radiation.
Nonlinear Absorption of Light in Materials with Long-lived Excited States

5
In figure 1, we showed how the excitation/disexcitation of a molecule is essentially a 3-state
(or more!) process. Some of the energy loss, however, occurs very quickly and only involves
vibrational levels. Considering the different time scales, one can simplify this into a 2-state
model: an excitation process which promotes the molecule into a long-lived metastable state
and its relaxation to the ground state. The origin of the long life of the metastable state
depends on the particular system under study. In the case of spin flip of the excited electron,
the physical reason underlying the stability of the triplet state is to be found in the selection
rules, which practically forbid the transition between two different spin states (excited
triplet state- ground singlet state). This process has raised a vivid interest in the scientific
community in the last few decades, because triplet state is often a big problem in organic
semiconductor devices [Wohlgenannt & Vardeny 2003]. Alternatively, the molecule, excited
by light, gets trapped in a metastable state, separated from the ground state by an energy
barrier. This is the case for azobenzene, a small molecule which exists in two different forms
(isomers trans and cis). The transition between the two isomers requires breaking a double
bond. UV light with a certain energy induces this double-bond breakage and lets the
molecule rotate around its axis; with a certain probability, the bond will reform when the
molecule is in a metastable cis isomer. The relaxation to the ground (lower energy) state can
only happen if there is enough energy to break the double bond again. This can occur if the
molecule is excited with a light at a different wavelength, or if the thermal fluctuations
provide the molecule with enough energy to overcome the energy barrier and return to the
ground state. The thermal relaxation is very slow and the characteristic lifetime depends on
the nature of the chromophore and of the surrounding environment. This is a classical
Kramers problem of overcoming an energy barrier (the breakage of the double bond)
between the metastable and the ground state. In the case of this simple molecule, the
Lambert-Beer law is no longer accurate because of a phenomenon which we call here
dynamic photobleaching or saturable absorption. It means that the photons which shine
on a sample are absorbed by the chromophores in the first layers. If these molecules dont
return to their ground state immediately, when new photons fall on the sample they cant be
absorbed anymore in the initial layers and therefore propagate through the sample with a
sub-exponential law. So, the effective photo-bleaching of the first layers allows a further
propagation of light into the sample and this leads to nonlinear phenomena which are
interesting both from the theoretical [Andorn 1971, Berglund 2004, Statman & Janossi 2003,
Corbett & Warner 2007] and from the experimental point of view [Meitzner & Fischer 2002,
Barrett et al. 2007, Van Oosten et al. 2005, Van Oosten et al. 2007].
The aim of this chapter is to explore the effect that this phenomenon has on the typical
absorption measurements which are commonly performed on these kinds of molecules. We
will propose a new theory which can mathematically describe this effect and then we will
give experimental evidence of its validity both on azobenzene, a molecule with a very long-
lived excited state and whose kinetics of transition can be followed, and on more common
fluorescent molecules, like chlorophyll, focussing on the absorption of light at equilibrium.
2. Materials and methods
2.1 Azobenzene
The molecule 4-hexyloxy-4-((acryloyloxyhexyloxy)azobenzene (abbreviated as AC
6
AzoC
6
)
was synthesized in our lab by Dr. A.R. Tajbakhsh. Its molecular structure is shown in figure
3 and its synthesis is described in [Serra & Terentjev 2008a]. All azobenzene-based
Nonlinear Dynamics

6
molecules exist in two isomers, trans and cis: the transition between the trans isomer, more
stable, to the cis isomer, metastable, is stimulated by UV light, while the opposite reaction
can be spontaneous. The isomers of the described molecule are shown in fig. 3

Fig. 3. The monomer used AC
6
AzoC
6
has an acrylate head group followed by a carbon chain
where the azo-group is attached. It is schematised here in its two isomers, trans and cis.
The two isomers of this molecule absorb light at different wavelengths: the trans isomer has
a peak around 365 nm, while the cis isomer absorbs at 440 nm. It is thus possible to monitor
the kinetics of trans-cis transition.
Monitoring a conversion process in real-time presents difficulties for a traditional
spectrophotometer, because measurements over the whole spectrum of wavelengths take a
long time, and moreover it is often difficult to access the sample in order to provide the UV
illumination for isomerisation. For this reason, we chose a spectrometer equipped with a
CCD camera, which is able to collect signal across the whole visible spectrum
simultaneously. This technique works by illuminating the sample with white light; a system
of gratings then splits the transmitted light into its various spectral components, whose
intensity is measured by an array of photodiodes. This type of spectroscope does not require
a fixed or enclosed sample holder, therefore placing another source of illumination close to
the sample is easy.
For the measurements of light absorbtion a Thermo-Oriel MS260i (focal length 260mm)
spectrometer was used. The apparatus consists of a quartz probe lamp with an adjustable
slit, a quartz cuvette with 1cm optical path, an optical liquid lightguide to conduct the light
from the cuvette to the spectrometer, a 50 m slit at the entrance of the spectrometer, and
the Andor linear-array CCD camera connected to a computer. The simultaneous
measurement of all spectral frequencies allows for a response as fast as 0.021 s and the
possibility of reducing noise by averaging over many measurements. Before every
absorption spectrum, a background and a reference spectrum were collected: the
background is the spectrum collected without the illumination from the probe lamp, and the
reference was the spectrum collected with the probe lamp illuminating the cuvette filled
with solvent (without the chromophore dye). The absorbance was then calculated from the
counts of the detector as:

Nonlinear Absorption of Light in Materials with Long-lived Excited States

7
For all the experiments, it was important to verify that the linear relation between
absorbance and concentration held for the value of absorbances considered. It was shown
that the absorption-concentration relation was linear below A 1.2 in the base-10 defined
absorbance. At the intensity used for this experiment, for a concentration c expressed in
moles, the penetration depth at 365nm was = c 480nm.
At higher concentrations, the linearity fails because of various phenomena, including
aggregation of the molecules (especially with molecules like proteins or polymers), the
scattering of light from big particles and stray light in the spectroscope.
We provide monochromatic illumination to stimulate the isomerisation of azobenzene using a
Schott KL 1500 LCD lamp, placed at 90 degrees with respect to the incident probe light and the
optical fiber that collects the light from the sample. In this way the cuvette is irradiated with
UV light while the absorption spectrum is recorded along the perpendicular beam path, so the
absorption can be followed in real time without interference of the illumination light. The
intensity of the monochromatic light was in the order of a few tens of W cm
2
.
All isomerisation reactions were followed for 90 minutes, which was a sufficient time for
reaching the respective photostationary states. After this, the lamp was switched off and the
absorbance was measured during thermal isomerisation in the dark. An example of
spectrum measured with this technique is shown in fig. 4.


Fig. 4. Photo-induced isomerisation (a) and thermal relaxation (b) curves of AC6AzoC6
recorded as a time sequence. The arrows indicate the direction of the peak movement
during the reaction.
2.2 Other chromophores
Chlorophyll was extracted from Commelina Communis leaves
1
. The leaves were first boiled in
distilled water, in order to kill the enzymes which digest chlorophyll once the leaf is cut
from the plant. The leaves were then dried and ground up with a pestle, with a few drops of

1
Leaves were kindly provided by J. McGregor
Nonlinear Dynamics

8
acetone, and then left in a 50% hexane/water mixture (the hexane forms a layer on top of
the water), to separate the chlorophyll from the water-soluble compounds (vitamins, etc).
The extracted solution was filtered to avoid impurities, like dust particles or even intact
chloroplasts which could be responsible for light scattering effects. The whole extraction
process was carried out in the dark. This method of extraction does not allow the separation
of chlorophyll from the carotenoids which could be present in the leaves. However, the
collected spectrum shows that the stronger absorption bands are those of chlorophyll, and
this means that the other compounds are present only in low concentrations. Moreover, for
the purpose of our experiment, an highly purified chlorophyll is not needed, because the
analysis focusses on the absorption band around 660nm, far from the absorption band of
carotenoids (blue region). What is important to remark is that there are different kinds of
chlorophyll, whose relative content varies from species to species. The two main chlorophyll
components, called chlorophyll a and chlorophyll b, differ by a carboxylic group,
attached only to the porphyrine ring in chlorophyll b. The two molecules have slightly
different absorption spectra, but this is not relevant for the experiments, provided that the
plant species from which the chlorophyll is extracted is always the same (and has therefore
the same percentage of chlorophyll a and b).
From the discussion above, it is clear that chlorophyll is a very special molecule, and has
many peculiarities. In order to demonstrate that the theory is more general, a commercial
dye with a strong absorption in the visible region is also investigated. Nile Blue A, a dye
commonly used for staining DNA, but with spectral properties similar to chlorophyll, was
selected. The chemical structure of chlorophyll and Nile Blue is shown in figure 5.
The spectroscopy experiments were conducted with an Ocean Optics USB 4000
spectrometer, equipped with optical fibers. A 25W halogen lamp with spectral range 400-880
nm, whose intensity could be tuned, was used as illumination source. The light was
focussed with collimating lenses onto a 1 cm cuvette containing 3 ml of solution.
For comparison with more conventional spectroscopes (meaning, with a fixed sample
holder and a fixed intensity of incident light), a Cary UV-Vis Spectrophotometer was used
to measure spectra and absorbances of the two substances at various concentrations. The
intensity of the incident light from the spectrometer was also measured with a power meter.
In order to measure the intensity of the incident light, key quantity in our experiments, the
light from the source was shone onto the detector directly, in the absence of any sample or
cuvette in between. Knowing the characteristic response of the spectrometer detector, it is
possible to measure the intensity of light. Three different values, the number of counts at a
single wavelength or the integral of the intensity over the range of wavelength which
correspond to the absorption peak, and the integral over all the wavelengths could be used
to quantify the incident light intensity. In all cases the outcome of the experimental results
was the same. Using as a value of intensity the intensity at the single peak-wavelength made
it possible to compare it to the conventional spectroscope (which produce monochromatic
light). It was verified that the detector had a linear response in counts versus intensity over
the range of intensities we used.
For the absorption spectra, measured as A = log
10
(I
0
/I) reference spectra for the pure solvent
were taken before each measurement at every light intensity. The absorption and
fluorescence spectra of these materials are shown in figure 6. We chose to refer all the
absorption values to the wavelengths of the peaks in the yellow-red region.
Nonlinear Absorption of Light in Materials with Long-lived Excited States

9

(a) (b)
Fig. 5. Chemical structure of a) chlorophyll a and b) nile blue dye.


Fig. 6. Absorption spectra of chlorophyll (green curve) and nile blue (blue). The absorption
peaks which were considered in this work are those at 668nm and 628nm respectively.
For each solution, the linearity of the absorption/concentration curve was verified, in order
to avoid falling into a trivial nonlinear regime. The experiments were conducted in random
order of light intensity, and the reproducibility was verified. The absorption of each
chlorophyll solution at 660 nm was stable over a range of hours at constant incident light
intensity, indicating the absence of chemical irreversible bleaching. Fluorescence from the
dye was also ruled out as a possible source of disturbance, because at the light intensity we
used it is not detectable with our equipment.
3. Theory
Here we present a description of the dynamical photobleaching effect in the case of
azobenzene isomerisation, which was previously discussed. We will then generalise the
Nonlinear Dynamics

10
discussion to all the molecules with a long-lived excited state, and show how this affects the
measurements of light absorption.
The non-Lambertian propagation of light through a medium has important consequences
for the analysis of photo-isomerization kinetics: when the photo-bleaching becomes
important, the measured absorbance no longer follows a simple (traditionally used)
exponential law. In photosensitive molecules like azobenzene, irradiation with light of a
certain wavelength induces a conformational change (isomerization) from an equilibrium
trans state where the benzene rings are far apart to a bent cis state where they are closer. The
isomerization process follows first-order kinetics. Calling the fraction of molecules in the
two states trans and cis nt and nc we have

where I is the intensity of light, k is the trans-cis isomerisation rate, kb the stimulated back
transition rate (cis to trans), and the thermal relaxation rate. In the experiments on
azobenzene described below we use an illuminating light monochromated at the trans-cis
transition wavelength. In this case the stimulated cis-trans isomerization is negligible (that is,
Ikb 0) and, remembering that nc = 1 nt the kinetic equation reduces to

(2)
In this equation the intensity I = I(x) is the light intensity at a certain depth into the sample.
It is convenient to define a non-dimensional parameter = I
0
k/, which represent the
balance of photo- and thermal isomerization at a given incident intensity I
0
. In this notation,
the amount of molecules in the trans conformation in the photostationary state, when the
balance between nt and nc is stable, and therefore dnt/dt = 0 the equation reduces to simply

therefore

(3)
To express mathematically the reversible photobleaching phenomenon, it can be assumed
that the change in light intensity across a thin layer of sample (thickness dx) is proportional
to the number of molecules which are excited, i.e. the number of chromophores which
absorbed light in a small volume of sample of thickness dx; neglecting the stimulated cis-
trans isomerization (which is appropriate in our study), the model can be much simplified to
give, per unit time:

Then, combining all the parameters, such as the photon cross section and the transition
rates, we recover the penetration depth D = /c as the relevant parameter of the relation,
and the final expression is:
Nonlinear Absorption of Light in Materials with Long-lived Excited States

11

(4)
with D the penetration depth, inversely proportional to concentration. In order to study this
problem at the photostationary state, one needs to combine the equations (4) and (3).

(5)
Solving the differential equation, the stationary-state light intensity at a depth x is given by
the relationship [Corbett & Warner 2007]

(6)
Looking at the equation above some important insights can be gained. The most important
is that in the limit = 0 the equation reduces to the Lambert-Beer law, i.e. an exponential
decay in the transmittance through the medium. Therefore all the nonlinearity is included in
. The opposite limit, when is very big, leads to a linear relation between transmittance
and sample thickness. Figure 7 is a representation of equation 6 and it shows the intensity
variation I(x) for several values of the parameter : from the plot of transmittance as a
function of x, it is clear that, if the incident intensity, and therefore , is low enough, the
Lambert-Beer law is valid and the decrease is exponential, but if the incident intensity is
high the bleaching of the first layers becomes progressively more relevant such that they
become partially transparent to the radiation. The decay thus tends to become linear in the
bulk of the sample, I(x) I
0
(1 x/D).



Fig. 7. Transmitted intensity ratio I/I
0
in the photostationary state as a function of the
parameter x/D (proportional to sample thickness or inversely proportional to chromophore
concentration) for several values of . At small the decay is exponential; the light penetrates
deeper into the sample as increases, as the decay tends to be linear. [Serra & Terentjev 2008b]
In order to model the dynamics of photoisomerisation, which is evidently inhomogeneous
across the sample, it is not enough to model photobleaching with equation 6, but instead the
equations (2) and (4) should be coupled. Calculating a time derivative of equation 4 leads to
Nonlinear Dynamics

12

(7)
In the right hand side expression, equation 2 can be substituted, giving

(8)
This equation can be solved with the following method [Corbett et al. 2008]. Introducing the
variable y = ln(I/I
0
), which is a very sensible variable, being also the inverse of the
absorbance, the left-hand side is greatly simplified and one obtains

(9)
Also, from equation 4

(10)
Substituting nt in equation 9 and rearranging, one finds

(11)
In the next step, one has to keep in mind that
( ) exp( )
y
d dy dy
e y y
dx dx dx
+ = +
Rearranging equation 11 and exchanging the order of derivatives on the left-hand side, the
equation reduces to

(12)
It is now possible to integrate this expression. Integrating between 0 and x, the integral and
the derivative on the left-hand side cancel out and one finds:

(13)
The factor comes from the solution of the definite integral for x=0 (the lower integration
limit). In fact, if x=0, I = I
0
and therefore y=0 by definition.
The last step is a time integration. At time zero, the absorbance A = y must be equal to the
Lambert-Beer law value x/D. Including these considerations, the integral expression for the
intensity I(x, t) becomes:

(14)
Nonlinear Absorption of Light in Materials with Long-lived Excited States

13

Fig. 8. Transmitted intensity ratio I/I
0
as function of time through a fixed value of x/D = 2.7
and different incident light intensities. There are several things to observe in this figure. First,
as changes, the photostationary state reaches different levels, as expected: if is bigger, the
sample becomes more transparent. The second thing is that when increases (therefore
intensity of light) the sample reaches the stationary state more quickly. The last observation,
most important for our study, is that with increasing the deviation of the kinetics from a
simple exponential becomes more and more evident. [Serra & Terentjev 2008b]
The upper limit of this integral is the measurable absorbance A = ln[I
0
/I(x, t)] from a sample
of thickness x. Figure 8 shows a simulation predicting the time-evolution of intensity
transmitted along the path x/D. Note that at t = 0 all curves converge to the Lambert-Beer
I/I
0
= exp(x/D), while at long times a significant portion of chromophore is bleached and
the transmitted intensity increases.
We should note that the problem of non-linear photo-absorption dynamics is not only
restricted to azobenzene isomerisation. Even ordinary dye molecules that do not undergo
conformational changes stimulated by photon absorption, still follow the same dynamic
principles, but with electronic transitions in place of trans-cis isomerization. Therefore, the
results of this paper should be looked upon as widely applicable to other systems. In
particular, the two key conclusions, that the crossover intensity into the non-linear photo-
absorption regime is independent of dye concentration and that the rate of the transition is
independent on solvent viscosity, are probably completely general.
The model we propose only assumes a two-configuration system, and it does not imply
anything about the nature of the two states. Therefore, it is important to verify that this
model has a wider and broader validity, and, in detail, that it helps to understand the
behaviour of a large class of chromophores, like fluorescent molecules. Azobenzene
molecule exists in two physical states, trans or cis; for chlorophyll, one could make an
analogy and, considering the electronic transition, call the two states ground and
excited, we still find the same formula at the photostationary state

(15)
where x is the path length of light through the sample. It is important to see here that the
absorbance has a nonlinear dependence on the incident light intensity. The limits where the
Nonlinear Dynamics

14
LB law is recovered are either very low incident intensity (practically, it can never be
achieved) or a very fast recovery to the ground state compared to the excitation.
Equation 15 can has important implications for the interpretation of absorption data. Solving
the equation for the absorbance A = log(I
0
/I) leads to the expression

(16)
from which it is clear that the value of the absorption does not only explicitly depend on the
concentration and optical path length, but also on the intensity of the incident light I
0
.
4. Non-linear kinetics of photobleaching.
Azobenzene, as previously discussed, is a small molecule with two double-bonded nitrogen
atoms linked to two benzyl rings. It is photosensitive, because exposure to light induces an
isomerisation trans-cis (indicating two possible spatial arrangements of the benzyl rings)
around the double bond between the nitrogens, and this results in a molecule shape change.
The process is fully reversible either by stimulated backward photoconversion (with a light
at a different wavelength), or by spontaneous relaxation to the equilibrium trans
configuration.
The isomerization of azobenzene and its derivatives has been extensively studied for the last
fifty years [Sudesh Kumar & Neckers 1989, Renner & Moroder 2006, Rau 1990], because this
molecule has many interesting features and its applications range from electronics to
biomedicine. It has been used as a model molecule for all the biological processes that
involve similar reactions, like the isomerization of retinal in rhodopsin, or as a probe for
measuring the free volume in polymers [Victor & Torkelson 1987]. More recently, its
characteristic response to the polarization state of light made it a suitable molecule for
surface patterning [Nathanson et al. 1992, El Halabieh et al. 2004]. Finally, azobenzene-
containing elastomers can give rise to inhomogeneous photo-mechanical effects and their
applications as photoactuators and artificial muscles are under study [Hogan et al. 2002,
Finkelmann et al. 2001, Yu et al. 2004].
However, in spite of the large literature on the subject, many fundamental mechanisms and
effects have not been clarified yet. It is assumed that the isomerization reaction is very
sensitive to both electrical and mechanical characteristics of the environment which
surrounds the molecules, but identifying and separating these effects is a difficult and often
ambiguous task.
Of the two possible isomers of azobenzene, the trans form is the lowest energy form, since
the benzyl ring electron clouds are far apart (see fig. 3), but under UV light an isomerisation
occurs; once in the cis state, the molecules can return back to their equilibrium trans state
both by stimulated isomerisation with visible light or by thermal relaxation [Rau 1990]. The
rate constants of these two processes are usually different, thermal isomerisation being
slower. The microscopic mechanism that leads to the isomerisation is still not clear, but there
are suggestions for both rotational and inversional [Asano & Okada 1984] mechanisms may
be competing.
The isomerisation of azobenzene can be monitored by UV-Vis (ultraviolet and visible light)
spectroscopy, because the two trans and cis compounds have different absorption spectra in
this range: the trans isomer absorbs around 365 nm, while the cis isomer at around 440 nm
Nonlinear Absorption of Light in Materials with Long-lived Excited States

15
[Rau 1990]. Irradiation with light at the wavelength of the trans peak progressively depletes
the molecules in this conformation. This reaction can be followed by measuring the
intensity of the absorbance of the spectrum peak at 365 nm, which decreases as the trans-cis
photo-isomerisation reaction proceeds. An example of spectral evolution as isomerisation
proceeds can be found in figure 4.
The models that were proposed for reaction kinetic are basically first order models, with the
important exception of azobenzene in polymer matrices. The fraction of isomers in the cis
state, nc, varies as [Zimmerman et al. 1958, Mechau et al. 2005]:

(17)
where I is the irradiation intensity, kb and k are the cis-trans and trans-cis constant of
photoisomerisation respectively, is the thermal cis-trans isomerisation and nt represents
the fraction of molecules in the trans state, and it is equal to (1 nc). In the derivation of the
formula the sensible assumption was made that the trans-cis thermal isomerisation constant
is negligible.
A basic characteristic of the photoisomerisation problem is the rate of spontaneous thermal
cis-trans isomerisation . For a given azobenzene derivative, at fixed (room) temperature and
sufficiently low dye concentrations to avoid self-interaction, this rate is approximately the
same for all our solutions. We measured this rate after monitoring the relaxation of the
spectrum after the UV illumination is switched off (see [Serra & Terentjev 2008a] for detail)
and obtained 1.25 10
4
s
1
(or the corresponding relaxation time of ~ 8000s).
In order to test the predictions of the theory, dynamic absorption measurements were
performed for different dye concentrations and different light intensities. Considering
equation (14) this is equivalent to changing x/D (where D is inversely proportional to the
dye concentration) and , which is proportional to the incident intensity I
0
. With our
experimental setup it was possible to follow all the isomerisation kinetics and thus the time
dependence of I/I0

[Serra & Terentjev 2008b].
We prepared three different dye solutions with (non-dimensional) weight fractions c = 2.5
10
3
, 0.01 and 0.025, resulting in values of penetration depth ranging from D = 36 mm, to D
= 3.6 mm. We recall here the physical meaning of the penetration length, which is the
distance through the sample over which the light falls across a sample to 1/10 of its original
intensity. The cuvette containing the sample is 1 cm long; therefore a sample with D equal to
a few millimeters is almost completely opaque.
The measurements were performed using the Thermo-Oriel spectroscope described in the
materials and methods section. Illumination was provided by a Nichia chip-type UV-LED,
emitting at 365nm (bandwith about 10nm wide) whose output power was accurately
regulated by a power supply. The LED light was attenuated by passing through a black tube
of controlled length, placed in front of a quartz cuvette with 1 cm optical path. Several
values of intensity were used in reported measurements, ranging between I
0
= 4 and 60
W/cm
2
. It is important to point out that these intensity values are very low and that most
experiments on azobenzene isomerization are performed with intensities which are orders
of magnitude higher, making the photobleaching much more of an issue. The low values of
intensity allowed us to have a kinetics slow enough to detect the features which the theory
predicts at short times. Every point of the spectrum was collected as an average of 100
Nonlinear Dynamics

16
measurements. All isomerization reactions were followed for several hours, until a
photostationary state was reached. The measurements were repeated at different
illumination intensities (regulated with the power supply) and at different dye
concentrations.
In all cases it was important to verify that the dye concentration remained in a range where,
at time t = 0, the linear proportionality between absorbance and concentration (Lambert-
Beer law) held. This is important because the concentration of molecules in the trans state at
every instant was determined from the absorbance at 365 nm. Absorbance was measured at
several concentrations. The deviation from linearity started at A 1.2, which corresponds to
the dye weight fraction of c = 0.03 (3 wt%) in our 1 cm cuvette. After this point, aggregation
effects start playing a role and the basic Lambert-Beer law is no longer valid, undermining
the theoretical relationship given by the equation (14). We always kept the concentrations
below this value, so that the linearity at t = 0 was maintained, with A = x/D where the
penetration depth is inversely proportional to chromophore concentration, expressed as
weight fraction, D /c with = 91 m.
For our detailed dynamic experiments, a very important issue was the viscosity of the
solution. In fact, at high illumination intensity we have encountered an unexpected
problem. Figure 9 shows that the transmission of light through a low-viscosity dye solution
(in pure toluene) displays a characteristic oscillatory behaviour. Detailed analysis of this
phenomenon is under further investigation. Whether the oscillations are linked to the local
convection due to the heating of the sample spot [Nitzan & Ross 1973] or to the diffusion of
the less dense cis molecules or whether they are intrinsic to the non-linear photochemical
process [Borderie et al. 1992] is not clear at this stage.


Fig. 9. Kinetics of isomerisation monitored through the observation of I/I
0

over time for 3
different values of x/D ( - 0.2, - 0.7, - 1.1) and 2 different values of , corresponding to:
(a) I
0

= 4W cm
2
, and (b) I
0

= 20W cm
2
. The periodic instability was reproducible in all
low-viscosity experiments. [Serra & Terentjev 2008b]
In order to avoid this difficulty, the dye solutions were prepared in a mixture of toluene and
polystyrene of high molecular weight. Adding polystyrene increases the viscosity of the
solution by over 2 orders of magnitude, and in this way prevents fluid motion in the cuvette
on the time scales of our measurements. Polystyrenetoluene solutions were prepared at a
Nonlinear Absorption of Light in Materials with Long-lived Excited States

17
fixed weight ratio. Adding polystyrene to toluene increases the Rayleigh scattering of the
solution, but we felt that we could safely do that because on one hand the absorption
dynamics is not affected (we used the same concentration for all the measurements and for
the reference spectrum), and on the other we measured the transmittance of the toluene-
polystyrene solution, which is almost equal to the pure toluene solution at 365nm.
In figure 10 the representative experimental results are shown for the solution with the
highest chromophore concentration (D = 4.6 mm, leading to x/D = 2.2) and three values of
incident intensity. One finds that all curves converge to the same initial value corresponding
to the I/I
0
= exp[x/D], which for this concentration means quite a low transmission (I/I
0

0.11). If the isomerisation didnt take place, the sample absorption would be constant in time
according to the classical Lambert-Beer. A traditional description of the kinetics of
isomerisation would predict an exponential decay of the absorption over time, but from
figure 10 we see a strong deviation. We fit the data with the theoretical model given by
equation (14) where we input the values of (thermal relaxation) and x/D, leaving only =
I
0
k/ free. Two data sets at higher intensity show the transmitted I(x, t) reach saturated
values. In this case we are confident of the fit because we have to match both the slope and
the amplitude of the curve. We obtain 60 for I
0
= 60 W/cm
2
, 20 for I
0
= 20 W/cm
2
,
and 4 for I
0
= 4 W/cm
2

(the matching of values is pure coincidence). We found that one
particular output of experimental recording, the absorbance plateau value (photo-bleached)
at long times, was extremely sensitive to the reading of reference intensity I
0
. The latter
measurement could be affected by various stray factors and in a few cases we had to rescale
the raw absorbance readings with a proper reference value. This issue did not have any
effect on absorbance at t = 0, or the kinetics.



Fig. 10. The effect of photo-bleaching for samples with high dye concentration (x/D = 2.2).
Three values of irradiation intensity are labelled on the plot. Solid lines are fits to the data
with only one free parameter , giving = 60 for the highest intensity, = 20 for the middle
intensity, and = 4 for the lowest intensity. [Serra & Terentjev 2008b]
Nonlinear Dynamics

18

Fig. 11. The same experiment as in figure 10 but with an intermediate dye concentration
(x/D = 1.1), and the same values of irradiation intensity. Here the solid lines are not fits, but
theoretical plots of equation (14) for = 60, 20 and 4 for the decreasing I
0
, respectively.
[Serra & Terentjev 2008b]
At a lower concentration of chromophore, corresponding to D 9.2 mm and x/D = 1.1 (the
transmitted intensity is about 1/10 of the incident intensity), figure 11 shows the similar
features of the non-linearity, which are especially evident at very short times. Again all curves
start at the same I/I
0
0.33. At higher irradiation intensities we achieve the saturation and the
steady-state value I(x) corresponding to the solution of equation (6). The change of curvature,
notable in figures 8 and 10, is not so clear here even at the highest I
0
. However, in the
comparative analysis of data we now take a different approach. Assuming all the parameters
for the curves are now known ( and x/D from independent measurements, and from the
fitting in figure 10), we simply plot the theoretical equation (14) on top of the experimental
data. It is clear that the theory is in excellent agreement with the data.
Finally, we study the case of low dye concentration (D 91 mm, x/D = 0.14) in figure 12:
this is also the case which is more relevant for biological spectroscopy studies, where the
concentration of chromophore is usually small. In this exemplified case, the initial
transmittance is very high: almost 85% of the incident light goes through the sample. Here,
the complicated integral equation (14) simplifies dramatically, since at small x/D 1 the
difference between A = ln 10 ln(I
0
/I) and x/D (which is the range of integration in (14), is
also small. The integration can then be carried out analytically, giving

(19)
which gives in the stationary state the correct solution of equation (6) approximated at small
x/D:

Nonlinear Absorption of Light in Materials with Long-lived Excited States

19

Fig. 12. At low dye concentration (x/D = 0.14) the sample is relatively transparent. The data
are for the same three values of irradiation intensity as in the earlier plots (but note that the
I/I
0
axis starts from 0.8). The solid lines are theoretical fits for = 60, 20 and 5. The inset
shows the plot of exponential relaxation rate
1
against I
0
, with the linear fit. [Serra &
Terentjev 2008b]
The fits of the data for I(t) are again in good agreement with the full theory. More
importantly, we also see that that the rate of the process described by the approximation (19)
is given by the simple exponential,
1

= (1 + ) = + kI
0
. This is in fact the rate originally
seen in the kinetic equation (2). Therefore, if we instead fit the family of experimental curves
in figure 12 (and several other data sets we measured) by the simple exponential growth of
the absorbance, we can have an independent measure of the relaxation rates obtained by
this fit. The inset in figure 12 plots these rates for all the I
0
values we have studied. A clear
linear relation between the relaxation rate and I
0
allows us to independently determine the
molecular constant:

The measurement of k, with high accuracy, gives the ratio k/ 1 cm
2
s
1
W
1
, which
explains the fitted values of the non-dimensional parameter = I
0
k/.
The consequences of this nonlinear behaviour have in the last year raised an interested in
some research groups who studied the azobenzene-based actuators. The original work by
Corbett and Warner, in fact, focused only on the steady-state behaviour, could lead to
accurate prediction about the effect of dynamic photobleaching on the bending angle of
elastomers [Corbett & Warner 2007]. In fact, the dynamic photobleaching is the reason why
heavily doped cantilevers, where the penetration depth is very small, can still bend if
irradiated with sufficiently intense beams. Because the contraction of cantilevers is due to
the force generated by the differential contraction of the top and bottom layers, if the light
was propagating exponentially in the medium the bending would be impossible, because
the thin layer where the light propagates is too small to generate enough force. A non-
exponential propagation of light due to photobleaching, instead, can explain this effect.
Subsequent work by Van Oosten, Corbett et al. [Van Oosten et al. 2008, Corbett & Warner
Nonlinear Dynamics

20
2008] and White et al. [White et al. 2009] shown experimental evidence of this effect on the
bending of cantilevers. Lee et al. also shown the nonexponential kinetics on a different
azobenzene-based molecule [Lee et al. 2009].
5. Absorption of fluorescent molecules.
Because absorption spectroscopy is so widely used in biology, we want to show the effect of
dynamic photobleaching on a biological molecule, and we chose chlorophyll, an important
substance in biology (and in everyday life). Chlorophyll has a very recognisable absorption
spectrum, which shows two clear peaks, one in the blue and the other in the red region
(which procures its green colour) of the electromagnetic spectrum. It is also fluorescent in
the far red and the characteristic lifetime of its excited state is about 4 ns [Hipkins 1986, Jaffe
& Orchin 1962]. If it is irradiated by UV light or very strong visible light it undergoes a
photo-chemical bleaching which degrades the molecule irreversibly and leads it to
precipitate from solution, as many studies reported [Mirchin et al. 2003, Mirchin & Peled
2005, Carpentier et al. 1987]. We wish to observe a dynamical reversible bleaching due to the
absorption of light, rather than this chemical degradation process.
In the previous section, the theoretical model was verified in the case of azobenzene, a
molecule with a very long lived excited state. Because the kinetics of transition could be
followed by a spectrometer, it was also possible to model it with the kinetics law (equation
14). The model, as we said, does not make any hypothesis on the nature of the transition,
and can therefore be extended to all two-state (or more realistically, to the simplified 3-
state) systems. Fluorescent molecules have an excited state with a characteristic lifetime of a
few nanoseconds, which is still much slower than the typical time of excitation. These
characteristic times, though, are too short to be followed with conventional spectroscopes,
and the transition kinetics cannot be followed as in the previous case. The model, however,
also makes predictions also about the transmittance at the photostationary state, which
differs from the LB law transmittance. To clarify, in figure 10 the Beer limit would be the
transmittance at time zero, and the stationary state the transmittance at long times.
It was important, for our experiments, to rule out all possible mechanisms leading to failure
of LB law. As it was previously discussed, LB law has many limitations. It fails at high
concentration of dyes, when they start to interact with each other and form aggregates; it
fails if the stray light is high and the apparent absorption seems to reach a saturation level; it
can fail at high intensity of the incident light if nonlinear effects like multiple photon
absorption, or saturable absorption occur [Abitam et al. 2008, Correa et al. 2002]; it fails for
highly scattering samples because the light is sent out at a non-zero angle. In order to rule
out all these possible effects, we place ourselves in the most favourable experimental
conditions: low concentration of dye and low illumination intensity.
According to the model, the behaviour at the stationary state is described by equation 16.
The important thing to observe is that the absorbance (or, equivalently, the transmittance)
also depends on the intensity of the incident light I
0
. In order to experimentally verify this
dependence, five different solutions of chlorophyll at known concentrations were measured
at various light intensities. In this section all the absorbances will be reported in base-10
logarithmic form. Figure 13 shows the outcome of measurements of chlorophyll absorption
of the same solution using different incident intensities. The result was striking: the change
in the measured absorbance was very substantially affected by this parameter.
Nonlinear Absorption of Light in Materials with Long-lived Excited States

21

Fig. 13. Absorption of chlorophyll in ethanol at the same concentration (in fact, exactly the
same solutions) measured only changing the incident illumination intensity, I

1 = 6.5, I

2 =
13.1, I

3 = 27.5 Wm
2
s
1
.
Some interesting consequences of this effect are shown in figure 14 and 15. The values
correspond to the steady-state absorption at the peak wavelength. Indeed it is possible to see
a strong dependence on the incident light intensity which is enhanced at high solute
concentrations. A change in intensity of about 80% of the maximum value leads to a change
in absorbance of about 50%. Figure 14 shows the dependence of the absorbance on the
intensity at various concentrations. Equation 16 cannot be explicitly solved for A, but only
for I
0


which gives the fits in the plot. Figure 15 shows the same data in the classical absorbance-
concentration plot, for different intensities. It is important to remark that the experimental
points can be satisfactorily fitted with a straight line in all cases (as the LB law says) but the
line slopes are very different. Therefore the absorption coefficient may have different values
if it is measured with a different light source. The exchangeability of results between
different laboratories is thus in question.
We obtained analogous results with Nile Blue, a simpler chromophore. We decided to test
this dye, described in the Material and Methods section, because it has an absorption
spectrum similar to chlorophyll in the red region, but it is a simpler and well studied
molecule. This also proves that the results are general, and that aggregation phenomena
which may occur in chlorophyll solutions (giving rise to scattering phenomena from still
intact chloroplasts) are ruled out as a possible cause for the observed behaviour.
All of the experiments were repeated several times and the behaviour was reproducible.
Moreover, the intensity of light was increased and decreased alternatively to exclude the
hypothesis of a chemical permanent photobleaching as a reason for absorption decrease.
Due to the phenomenon of reversible (dynamic) photobleaching, a simple absorption
experiment like the one described in the introduction is in practice impossible. The values of
the absorption coefficients are meaningless if they dont carry the information about the
intensity of the incident light.
Nonlinear Dynamics

22

Fig. 14. Absorption of chlorophyll as a function of the intensity of incident light. One can see
an increase of absorbance at low intensities. The values are reported for five different
concentrations. In the figure, the black dotted line corresponds to the intensity of the
incident light of a commercial traditional spectrophotometer (Cary-UV-Vis). This
comparison is done in order to show that the range of intensity of our set-up is the same as a
more conventional one.

m

Fig. 15. Absorption as a function of concentration for the different values of incident light.
All the lines have a good Lambert-Beer linear form but different proportionality
coefficients. The LB limit was extrapolated from the ideal limit of zero intensity.
In light of this, can we use the theoretical model to find a new method to determine
concentrations using absorption spectroscopy, removing this dependence on the incident
light intensity? Looking at equation 15, knowing the ratio of the concentrations of two
solutions makes it possible to measure the combined ratio of parameters = (k/)I
0
. If one
solution has an unknown concentration c
1
and another solution is obtained by a dilution of
Nonlinear Absorption of Light in Materials with Long-lived Excited States

23
the first one, so that c
1
/c
2
= r, measuring the absorbance of the two solutions 1 and 2, at the
same incident light intensity and the same path length x, one obtains:

(20)
then

(21)
Knowing , one can simply determine the unknown concentration c
1
as

(22)
The relation yields the ratio c/, and therefore the knowledge of the absorption coefficient

1
is required. On the other hand, the same method allows determination of once the
concentration is known.
We took a series of concentrations of Nile Blue solutions and the corresponding absorbance
measured at different intensities of incident light. Taking pairs of measurements of
absorbance at known concentration, at the same value of light intensity, we extracted the
value of the parameter r from each pair, and from that we calculated and according to
equation 21 and 22. Averaging over all of them, we obtained the value = 120000
20000M
1
cm
1
. The literature reports = 77000M
1
cm
1
, but we attribute this to the fact that,
at intensities greater than zero, the absorption values are always systematically smaller (see
figures 14 and 15). Therefore the deviation from the literature value is still consistent with
our findings.
The limitation of this method is that it very sensitively depends on the value of the
concentration ratio, and therefore the errorbars are quite large. One should not, in fact, rely
on the value of measured only with one pair of measurements. Figure 16 shows the
dispersion of the estimate values of obtained using different pairs of values. The strong
dependence of the parameter on the value of r is evident from formula 21. It is possible to
see that the function is divergent when

but this happens
when r A
1
/A
2
, which is exactly the region of interest. For this reason, the values of are
very scattered (some of them are even negative, which is physically meaningless), and this
formula, although it is correct is principle, is hardly applicable to real experimental data.
6. How to use absorption spectroscopy
In order to overcome this disadvantage, a more robust method is suggested to determine the
value of the absorption coefficients. The strong dependence of on r obviously remains,
because it comes directly from formula 21. A method based on a linear regression can be
used instead to calculate from a series of absorbances and known concentrations. Also, a
series of measurements at known allow the evaluation of a substance concentration, just
like in traditional absorption experiments. Rearranging equation 15, one obtains:

(23)
Nonlinear Dynamics

24

Fig. 16. Values of obtained for three different values of light intensity, using various pairs
of measurements, according to the suggested method. It is important to notice that the
values of are systematically higher at higher intensity, as expected. As one would expect
from equation 21, the values of are very scattered and therefore the calculation of is not
precise: this is due to the strong dependence of on r around the point where r A
1
/A
2
.
If a set of concentrations and relative absorptions are known, one can plot the quantity
c(1 10

A
)
1
as a function of Aln(10)(1 10

A
)
1
. The result is a line whose slope is /x and
whose intercept is /x. It is therefore possible to determine all the important parameters.
Figure 17 shows this plot obtained for a set of Nile Blue dye. Promisingly, all the lines
obtained with this method are well fitted with parallel lines, which indicates that they all
converge to the same value of . Several lines indicate several values of incident light
intensity. From the plot one can find the parameter for all the intensities, and also , which


Fig. 17. Application of the suggested method to determine the relevant parameters and .
The slope depends on , which is the same for all the samples, but the intercept depends on
which changes with the intensity of light.
Nonlinear Absorption of Light in Materials with Long-lived Excited States

25
is simply the slope of the lines. Once is known, one can then determine, for x = 1, =
1
,
the molar absorption coefficient. Following classical error analysis we obtain
= 117700120M
1
cm
1
. The agreement with the previous method is good and this second
method has the advantage of greatly reducing the experimental errors.
Whilst this method appeared highly successful at a first glance, we discovered that plotting
the values of against the intensity of the incident light, measured from the number of
counts on the spectroscope, generated a relation which is not linear. According to the theory,
is simply the product between the incident intensity and some characteristic constant of
the material, therefore the nonlinearity shown in figure 18 is not acceptable.


Fig. 18. The parameter extracted from the intercepts of figure 17 as a function of the
intensity of the incident light. The theoretical model predicts a linear relationship, which is
not what is observed in the figure!
Considering the possible causes of this discrepancy, one can see in the model that
stimulated emission is completely neglected. Neglecting the light-stimulated back-transition
to the ground state was reasonable in the case of azobenzene, where the trans and cis peaks
were very far apart, but for fluorescent molecules the same light excites both transitions and
this factor should therefore be considered. The calculations become more complicated but
the procedure is the same as that described in the introductory section of this chapter.
One should now return to the kinetics equation, which we re-write here for simplicity. We call
n the fraction of molecules in the ground state and kb the rate of the stimulated back transition.

(24)
At the stationary state, the left hand side of the equation is zero and n becomes

(25)
This is the value which should be inserted in the expression for the photobleaching 4, giving

(26)
Nonlinear Dynamics

26
This can be simplified by dividing by k, introducing the parameter = kb/k and integrating
the equation

Note that earlier, neglecting the stimulated back-reaction, we essentially had = kb/k0.
While the integration on the right-hand side is trivial, the left-hand side splits into a sum

The integration gives

Final simplification leads to:

(27)
It is convenient here to reintroduce our usual non-dimensional parameter = I
0
k/

(28)
This expression is the full and general result. In many cases we expect to be small, so the
expansion at the first order correctly recovers the usual expression 6. Expansion to the
second order, instead, gives

(29)
Using this equation, the fit to the experimental data improved. The expression was readapted
to take into account the base-10 logarithm of the absorbance. The parameter space was
restricted because we expected and to be in the same range as previously determined. The
best fit to the curves was obtained using = 8.75 = 114000M
1
cm
1
and = 0.3. The value
of , quite substantially greater than zero, is consistent with the need to modify the original
equation. Figure 19 shows the concentration c on the y-axis and the absorbance A on the
abscissa (different curves for different light intensities): this is because formula 29 can be easily
inverted. The values of , obtained by fitting, increase linearly with the incident intensity, as
shown in figure 20. This is good evidence that the stimulated emission cannot be neglected: the
theory, thus modified, can well reproduce the experimental data.
7. Conclusions
The most important conclusion of our work is that one has to be cautious with the classical
concept of light absorption, represented by the Lambert-Beer law. Even without considering
Nonlinear Absorption of Light in Materials with Long-lived Excited States

27

Fig. 19. Fitting of the absorbance/concentration curves at different intensities, obtained with
the model which considers the stimulated back-transition.

Fig. 20. The parameter extracted from the fit in the figure above as a function of the
intensity of the incident light. In this case we observe the linear relationship with the correct
intercept in the origin.
multi-particle effects at high concentration or multiple photon absorption, even at very low
concentrations (corresponding in our case to the low x/D ratio) the illumination intensity
above a certain crossover level would always produce a non-linear dynamical effect
equivalent to the dynamic photo-bleaching, which increases the effective transmittance of
the sample. We emphasise that this is a totally reversible phenomenon, unrelated to the
chemical bleaching, which involves irreversible damage to the material. The crossover
between linear and strongly non-linear regimes is expressed by the non-dimensional
parameter = I
0
k/ and is, therefore, an intrinsic material parameter of every chromophore
molecule, but not dependent on the dye concentration. Note that the thermal cis-trans
isomerization rate is strongly temperature dependent, influencing the crossover intensity.
Azobenzene is an ideal molecule for this kind of study, because it allows investigation of the
transition kinetics using a simple spectrometer. The experimental data we obtained confirm
the predictions of the theoretical model, which provided a satisfactory fit to the data. At
high illumination intensity one finds a characteristic sigmoidal shape. Our experiments were
deliberately carried out in a highly viscous solvent to eliminate the additional complexities,
Nonlinear Dynamics

28
presented in figure 9, caused by the possible local convection flows of different isomers.
Certainly, a much more in-depth study will be required to take such effects into account.
The second result is related to the extension of the model to all fluorescent molecules, or
indeed any molecule which has a long-lived excited state. We showed that, even for those
molecules where the conversion between the two states is too fast to be followed by the
spectroscope, the nonlinearity has an important influence. The absorption values at the
stationary state were sensitive to the experimental conditions, and particularly the intensity
of the incident light.
It must be said that in the literature the light intensity of the spectrometer light source is
very rarely mentioned, therefore it is possible that many values of absorption coefficient
reported are wrong or meaningless. The reason why, to our knowledge, no one took this
phenomenon into consideration before is that many commercial spectroscopes always work
with the same light intensity, so the results are self-consistent. Also, weighing proteins or
other materials is often a difficult task, therefore a discrepancy between the expected value
and the literature values could be easily explained away. In fact, we showed that there is a
much deeper reason.
Moreover, this raises also problems of interpretation of data obtained comparing, for
example, the intensity of two different absorption peaks, because, even at the same incident
light intensity, the nonlinearity also depends on the factor (k/) which is different at each
wavelength. All these problems could be overcome using the method we suggested, which
is simple and straightforward and it allows reproducibility of results.
8. Acknowledgements
The authors wish to thank A. Tajbakhsh for the preparation of chemicals, A. Smith for the
advice in chlorophyll extraction, M. Warner and D. Corbett for discussing the theoretical
aspects of the work, J. McGregor for reviewing the manuscript, J. Huppert, P. Cicuta, M.
Kolle, K. Thomas and L. Payet for helpful discussion.
9. References
[Abitam et al. 2008] H. Abitam; H. Bohr; P. Buchhave. Correction to the beer-lambert-
bouguer law for optical absorption. Appl. Optics, 47:5354, 2008.
[Andorn 1971] M. Andorn; K. H. Bar-Eli. Optical bleaching and deviation from beers law of
solutions illuminated by a ruby laser. i. cryptocyanine solutions. J.Chem.Phys.,
55:50085016, 1971.
[Armstrong 1965] J.A. Armstrong. Saturable optical absorption in phthalocyanine dyes.
J.Appl.Phys., 36:471, 1965.
[Asano and Okada 1984] T. Asano; T. Okada. Thermal 2e isomerization of azobenzenes.
the pressure, solvent, and substituent effects. J.Org.Chem., 49, 43874391, 1984.
[Barrett et al. 2007] C. J. Barrett; J. Mamiya; K. G. Yager; T. Ikeda. Photo-mechanical effects in
azobenzenecontaining soft materials. Soft Matter, 3, 1249, 2007.
[Benaron et al. 2005] D. A. Benaron; I. H. Parachikov; W-F. Cheong; S. Friedland; B. E.
Rubinsky; D. M. Otten; F. W. Liu; C. J. Levinson; A.L. Murphy; J.W. Price; Y. Talmi;
J. P. Weersing; J. L. Duckworth; U. B. Horchner; E.L. Kermit. Design of a visible-
light spectroscopy clinical tissue oximeter. J. Biomed. Opt., 10, 044005, 2005.
[Berglund 2004] A. J. Berglund. Nonexponential statistics of fluorescence photobleaching.
J.Chem.Phys., 121: 28992903, 2004.
Nonlinear Absorption of Light in Materials with Long-lived Excited States

29
[Bopp et al. 1997] M. A. Bopp; Y. W. Jia; L. Q. Li; R. J. Cogdell; R. M. Hochstrasser.
Fluorescence and photobleaching dynamics of single light-harvesting complexes.
Proc.Natl.Am.Sci.USA, 94, 10630, 1997.
[Borderie et al. 1992] B. Borderie; D. Lavabre; J.C. Micheau; J.P. Laplante. Nonlinear
dynamics, multiple steady states, and oscillations in photochemlstry. J.Phys.Chem,
96, 2953, 1992.
[Born and Wolf 1999] M. Born; E. Wolf. Principles of Optics- 7th edition. Cambridge University
Press, Cambridge, 1999.
[Carpentier et al. 1987] R. Carpentier; R. M. Leblanc; M. Mimeault. Photoinhibition and
chlorophyll photobleaching in immobilized thylakoid membranes. Enzyme
Microb.Technol., 9, 489, 1987.
[Corbett and Warner 2007] D. Corbett; M. Warner. Linear and non-linear photo-induced
deformations of cantilevers. Phys.Rev.Lett., 99, 174302, 2007.
[Corbett and Warner 2008] D. Corbett; M. Warner. Polarization dependence of optically
driven polydomain elastomer mechanics. Phys. Rev. E, 78, 061701, 2008.
[Corbett et al. 2008] D. Corbett; C. L. van Oosten; M. Warner. Nonlinear dynamics of optical
absorption of intense beams. Phys. Rev. A, 78, 013823, 2008.
[Correa et al. 2002] D. S. Correa; L. de Boni; D.S. dos Santos jr.; N. M. Barbosa Neto; O. N.
Oliveira jr.; L. Misoguti; S. C. Zilio; C. R. Mendonca. Reverse saturable absorption
in chlorophyll a solutions. Appl. Phys. B, 74:559, 2002.
[Dunning and Hulet 1996] F. B. Dunning; R. G. Hulet. Atomic, molecular, and optical physics;
Atoms and Molecules. Academic Press Inc., San Diego, California, 1996.
[El Halabieh et al. 2004] H. El Halabieh; O. Mermut; C. J. Barrett. Using light to control
physical properties of polymers and surfaces with azobenzene chromophores. Pure
Appl.Chem., 76:14451465, 2004.
[Finkelmann et al. 2001] H. Finkelmann; E. Nishikawa; G. G. Pereira; M. Warner. A new
optomechanical effect in solids. Phys.Rev.Lett., 87, 015501, 2001.
[Heard 2006 ] D. E. Heard. Analytical techniques for atmospheric measurement. Blackwell
publishing Ltd., Oxford, 2006.
[Henderson et al. 2007] J. N. Henderson; H. W. Ai; R. E. Campbell; S. J. Remington.
Structural basis for reversible photobleaching of a green fluorescent protein
homologue. Proc.Natl.Am.Sci.USA, 104, 6672, 2007.
[Hipkins 1986] M. F. Hipkins; N. R. Baker. Photosynthesis energy transduction: a practical
approach. IRL Press, Oxford, Oxford, 1986.
[Hogan et al. 2002] P. M. Hogan; A. R. Tajbakhsh; E. M. Terentjev. Uv manipulation of order
and macroscopic shape in nematic elastomers. Phys.Rev.E, 65, 41720, 2002.
[Jaffe and Orchin 1962] H. H Jaffe; M. Orchin. Theory and applications of ultraviolet
spectroscopy. Wiley, New York, 1962.
[Lee et al. 2009] Y. J. Lee; S. I. Yanga; D. S. Kangb; S.-W. Joo. Solvent dependent photo-
isomerization of 4-dimethylaminoazobenzene carboxylic acid. Chem. Phys., 361,
176179, 2009.
[McCall and Hahn 1967 ] S. L. McCall; E. L. Hahn. Self-induced transparency by pulsed
coherent light. Phys.Rev.Lett., 18, 908, 1967.
[Mechau et al. 2005] N. Mechau; M. Saphiannikova; D. Neher. Dielectric and mechanical
properties of azobenzene polymer layers under visible and ultraviolet irradiation.
Macromolecules, 38, 38943902, 2005.
[Meitzner and Fischer 2002] G. D. Meitzner; D. A. Fischer. Distortions of fluorescence yield
x-ray absorption spectra due to sample thickness. Microchem.J., 71, 281, 2002.
Nonlinear Dynamics

30
[Merbs and Nathans 1992] S. L. Merbs; J. Nathans. Photobleaching difference absorption-
spectra of human cone pigments- quantitative analysis and comparison to other
methods. Photochem.Photobiol., 56, 869, 1992.
[Mirchin et al. 2003] N. Mirchin; A. Peled; Y. Dror. Modeling and analysis of bleaching
processes in photoexcited chlorophyll solutions. Synthetic Metals, 138, 323, 2003.
[Mirchin and Peled 2005] N. Mirchin; A. Peled. Photo-bleaching response in chlorophyll
solutions. Appl. Surface Sci., 248, 91, 2005.
[Nathan et al. 1985] V. Nathan; A. H. Guenther; S. S. Mitra. Review of multiphoton
absorption in crystalline solids. J.Opt.Soc.Am.B, 2, 294, 1985.
[Nathanson et al. 1992] A. Natansohn; P. Rochon; J. Gosselin; S. Xie. Azo polymers for
reversible optical storage. 1.poly[4- [ [2- (acr yloyloxy)ethyll ethylaminol-4-ni
troazo benzene]. Macromolecules, 25, 2268 2273, 1992.
[Nitzan and Ross 1973] A. Nitzan; J. Ross. Oscillations, multiple steady states, and
instabilities in illuminated systems. J.Chem.Phys., 59, 241, 1973.
[Rau 1990] H. Rau. Photochemistry of azobenzene. In J. F. Rabek, editor, Photochemistry and
Photophysics, pages 119142. CRC press; Boca Raton, 1990.
[Renner and Moroder 2006] C. Renner; L. Moroder. Azobenzene as conformational switch in
model peptides. ChemBioChem, 7, 868878, 2006.
[Serdyuk et al. 2007] I. N. Serdyuk; N. R. Zaccai; J. Zaccai. Methods in molecular biophysics.
Cambridge University Press, Cambridge, 2007.
[Serra and Terentjev 2008a] F. Serra; E. M. Terentjev. Effects of solvent viscosity and polarity
on the isomerization of azobenzene. Macromolecules, 123, 981986, 2008.
[Serra and Terentjev 2008b] F. Serra; E. M. Terentjev. Nonlinear dynamics of absorption and
photobleaching of dyes. J.Chem.Phys., 123, 224510, 2008.
[Statman and Janossi 2003] D. Statman ; I. Janossi. Study of photoisomerization of azo dyes
in liquid crystals. J.Chem.Phys., 118, 32223232, 2003.
[Sudesh Kumar and Neckers 1989] G. Sudesh Kumar ; D. C. Neckers. Photochemistry in
azobenzenecontaining polymers. Chem.Rev., 89, 19151925, 1989.
[Van Oosten et al. 2005] K. D. Harris; R. Cuypers; P. Scheibe; C. L. van Oosten; C.W.M.
Bastiaansen; J. Lub; J.D. Broer. Large amplitude light-induced motion in high elastic
modulus polymer actuators. J.Mater.Chem., 15, 5043, 2005.
[Van Oosten et al. 2007] C. L. Van Oosten; K. D. Harris; C. W. M. Bastiaansen; J.D. Broer.
Glassy photomechanical liquid-crystal network actuators for microscale devices.
Eur.Phys.J.E, 23:329, 2007.
[Van Oosten et al. 2008] C. L. Van Oosten; D. Corbett; D. Davies; M. Warner; C. W. M.
Bastiaansen; D. J. Broer. Bending dynamics and directionality reversal in liquid
crystal network photoactuators. Macromolecules, 41:85928596, 2008.
[Victor and Torkelson 1987] J. G. Victor; J. M. Torkelson. On measuring the distribution of
local free volume in glassy polymers by photochromic and fluorescence techniques.
Macromolecules, 20, 22412250, 1987.
[White et al. 2009] T. J. White; S. V. Serak; N. V. Tabiryan; R. A. Vaiaa; T. J. Bunning.
Polarization-controlled, photodriven bending in monodomain liquid crystal
elastomer cantilevers. J. Mater. Chem., 19, 10801085, 2009.
[Wohlgenannt and Vardeny 2003] M. Wohlgenannt; Z. V. Vardeny. Spin-dependent exciton
formation rates in -conjugated materials. J. Phys. Condens. Matter, 15, R83R107, 2003.
[Yu et al. 2004 ] Y. Yu; M. Nakano; T. Ikeda. Photoinduced bending and unbending behavior
of liquidcrystalline gels and elastomers. Pure Appl.Chem., 78, 14671477, 2004.
[Zimmerman et al. 1958] G. Zimmerman; L. Y. Chow; U. J. Paik. The photochemical
isomerization of azobenzene. J. Am. Chem. Soc., 80, 35283531, 1958.
2
Exact Nonlinear Dynamics
in Spinor Bose-Einstein Condensates
Junichi Ieda
1
and Miki Wadati
2

1
Institute for Materials Research, Tohoku University,
2
Department of Physics, Tokyo University of Science
Japan
1. Introduction
BoseEinstein Condensation (BEC) of atomic gases has attracted a renewed theoretical and
experimental interest in quantum many-body systems at extremely low temperatures
(Pethick & Smith; 2002). This excitement stems from two favorable features: (1) by applying
magnetic fields and lasers, most of the system parameters, such as the shape,
dimensionality, internal states of the condensates, and even the strength of the interatomic
interactions, are controllable; (2) due to the diluteness, the mean-field theory explains
experiments quite well. In particular, the GrossPitaevskii (GP) equation demonstrates its
validity as a basic equation for the condensate dynamics. The GP equation is the counterpart
of the nonlinear Schrdinger (NLS) equation in nonlinear optics. Thus, a study based on
nonlinear analysis is possible and important.
In nonlinear physics, a soliton is remarkable object not only for the fact that exact solutions
can be obtained but also for its usefulness as a communications tool due to its robustness. In
general, solitons are formed under the balance between nonlinearity and dispersion. For
atomic condensates, the former is attributed to the interatomic interactions, while the latter
comes from the kinetic energy. Either dark or bright solitons are allowable depending on the
positive or negative sign of the interatomic coupling constants g, respectively, and indeed
have been observed in a quasione-dimensional (q1D) optically constructed waveguide
(Strecker et al.; 2002) (Khaykovich et al.; 2002). Such matter-wave solitons are expected in
atom optics for applications in atom laser, atom interferometry, and coherent atom transport
(Meystre; 2002). In this chapter, we extend the analysis of the matter-wave solitons to a
multicomponent case by considering the so-called spinor condensate (Stenger et al.; 1998)
whose spin degrees of freedom are liberated under optical traps. Based on theoretical and
experimental results, we introduce a new integrable model which describes the dynamical
properties of the matter-wave soliton of spinor condensates (Ieda et al.; 2004a). We employ
the inverse scattering method to solve this model exactly. As a result, we predict the
occurrence of undiscovered physical phenomena such as macroscopic spin precession and
spin switching.
The chapter is organized as follows. In Sec. 2, the mean field theory of condensate is briefly
reviewed. Section 3 introduces an effective interatomic coupling in a q1D condensate. Using
these results, we consider a spinor condensate in q1D regime in Sec. 4. Then, in Sec. 5, we
Nonlinear Dynamics

32
show an integrable condition of the coupled nonlinear equations for spinor condensates in
which the exact soliton solutions are derived. In Sec. 6 and 7, we analyze the spin properties
of one-soliton and two-soliton, respectively. Finally we summarize our findings and remark
some current progresses on this topic in Sec. 8.
2. Mean field theory
The dynamics of BEC wave function can be described by an effective mean-field equation
known as the Gross-Pitaevskii (GP) equation. This is a classical nonlinear equation that takes
into account the effects of interatomic interactions through an effective mean field.
In this section, we derive the GP equation for a single component condensate and discuss
the theoretical background of it for later extension to a low dimensional case and a spinor
case.
2.1 Hamiltonian
In order to derive the mean-field equation for atomic BECs, we start with the second
quantized Hamiltonian. The Hamiltonian for the system of N interacting bosons with the
mass m in a trap potential U
trap
(r) can be written as

(1)

(2)

(3)
where v(r r) expresses the two-body interaction and the bosonic field operators satisfy the
equal-time commutation relations:

(4)
In most of the experiments, the trap is well approximated by a harmonic oscillator potential,

(5)
Condensates are pancake-shape for
z

x,y
whereas cigar-shape for
x,y

z
. For other
choice of trap potentials, say a linear or a 4-th order potential, the thermodynamic properties
can be changed (Ieda et al.; 2001). The discussion about non-harmonic potentials will be
given in a later section in connection with an implementation of quasi-one dimensional
system.
The atom-atom interaction v(r r) in a dilute and ultracold system can be approximated by

(6)

(7)
Exact Nonlinear Dynamics in Spinor Bose-Einstein Condensates

33
where a is the s-wave scattering length. The scattering length is the controllable parameter
which determines the properties of the low energy scattering between cold atoms. The
positive (negative) sign of a corresponds to the effectively repulsive (attractive) interaction.
2.2 Bogoliubov theory
The mean-field theory for weakly interacting dilute Bose gases (WIDBG) was proposed in
Bogoliubovs 1947 work (Pethick & Smith; 2002). The main idea of his approach consists in
separating out the condensate contribution from the bosonic field operator:

(8)
where n
0
= N
0
/ is a uniform condensate density (c-number) with N
0
the number of the
condensed atoms, the volume of the system, and the quantum part

is assumed to be a
small perturbation. Taking

and

terms up to quadratic, Bogoliubov built the first-oder
theory of uniform Bose gas.
This idea can be extended to non-uniform gases in trap potentials. If we introduce the r
dependence of the condensate part, the field operator is expressed as

(9)
The scalar function (r, t) is called the condensate wave function, which is normalized to be
the number of the condensed atoms,

(10)
In the case of BEC, the number of the condensed atoms becomes macroscopic, i.e.,

(11)
In this sense, the macroscopic wave function (r, t) is related to the first quantized N-
body wave function
N
(r
1
, . . . , r
N
; t) as

(12)
which obviously satisfies the symmetry under exchanges of two bosons.
Following the Bogoliubov prescription, we substitute (9) into (1) and retain and

terms
up to quadratic;

(13)

(14)

(15)
Nonlinear Dynamics

34


(16)
Equation (14) is called the Gross-Pitaevskii energy functional. The statistical and dynamical
properties of the condensate are determined through a variation of E

while the low-lying
excitations from the ground state can be analyzed by diagonalizing

. In the ground state,

part vanishes identically.
2.3 Gross-Pitaevskii equation
Even at the zero temperature, interactions may cause quantum correlation which gives rise
to occupation in the excited states. The assumption that the quantum fluctuation part (r, t)
gives a small contribution to the condensate is valid for a dilute system. In particular, if we
consider a dilute limit:

(17)
where na
3
is the gas parameter with n the particle number density, neglecting parts
provides an appropriate description of the condensate wave function at zero temperature.
By a variational principle,

(18)
we obtain the Gross-Pitaevskii (GP) equation:

(19)
This equation has been derived independently by Gross and Pitaevskii (Pethick & Smith;
2002) to deal with the superfluidity of
4
He-II. The GP equation is a classical field equation
for a scalar (complex) function but contains = explicitly. In this sense, the description of
the condensate in terms of is a manifestation of the macroscopic de Broglie wave, where
the corpuscular aspect of matter dose not play a role. Now the modulus and gradient of
phase of = ||exp(i) have a clear physical meaning,

(20)
where n and v denote number density and velocity of the condensate, respectively.
3. Confinement induced resonance
In this section, we derive an effective one-dimensional (1D) Hamiltonian for bosons
confined in an elongated trap. The interactions between atoms in the experiments are
always three-dimensional (3D) even when the kinetic motion of the atoms in such a tight
radial confinement is 1D like. Therefore, the trap-induced corrections to the strength of the
atomic interactions should be taken into account properly.
Exact Nonlinear Dynamics in Spinor Bose-Einstein Condensates

35
This problem was first solved by Olshanii (Olshanii; 1998) within the pseudopotential
approximation, yielding a new type of tuning mechanism for the scattering amplitude, now
called confinement induced resonance (CIR). In what follows, we show a detailed account of a
renormalization of the 3D interaction into an effective 1D interaction, which produces the
CIR. This technique plays a crucial role in Sec. 5 in order to realize an integrable condition
for spinor GP equations.
3.1 Model Hamiltonian
We start with the following model:
1. The trap potential is composed by an axially symmetric 2D harmonic potential of a
frequency

in the x-y plane.

2. Atomic motion for the z direction is free.
3. Interatomic interaction potential is represented by the Fermi-Huang pseudopotential:

(21)
where the coupling strength g is expressed by the 3D s-wave scattering length a as eq.
(7) (Meystre; 2002).
4. The energy of atoms for both transverse and longitudinal motions is well below the
transverse vibrational energy =

.
In the harmonic potential we can separate the center of mass and relative motion. Then we
consider the Schrdinger equation for the relative motion,

(22)
where the reduced mass m
r
= m/2, the relative coordinate r = r
1
r
2
, and the transverse
Hamiltonian:

(23)
From the above condition 4, we assume that the incident wave is factorized as
, where

is the transverse ground state

The
longitudinal kinetic energy is smaller than the energy separation between the ground state
and the first axially symmetric excited state:

(24)
where is the energy spectrum of 2D harmonic oscillator with n = 0, 1, 2, .
. . the principal quantum number, and m
z
the angular momentum around the z axis, which
takes on values m
z
= 0, 2, 4, . . . ,n (1, 3, 5, . . . ,n) for even (odd) n.
3.2 One-dimensional scattering amplitude
The asymptotic form of the scattering wave function is given by
Nonlinear Dynamics

36

(25)
where f
even
and f
odd
denote the one-dimensional scattering amplitudes for the even and odd
partial waves, respectively. While the transverse state (n = m
z
= 0) remains unchanged under
the assumption of low energy scattering considered above, the scattering amplitudes f
even,odd

are affected by a virtual excited state of the axially symmetric modes (n > 0,m
z
= 0) during
the collision.
To calculate the one-dimensional scattering amplitude we expand the solution,

(26)
where is the (axially symmetric) eigenstate of the transverse Hamiltonian (23), and
substitute this expansion into eq. (22) with the eigenvalue . Operating

(27)
to both side of the Schrdinger equation and taking the limit, in sequence, 0
+
, z ,
along with the asymptotic form (25), we can obtain and the following expression
for the scattering amplitudes:

(28)
Here we have used the normalization condition:

(29)
and the r0 limit of the regular (free of the 1/r divergence) part of the solution ,

(30)
We note that the regularization operator (r) that removes the 1/r divergence from the
scattered wave plays an important role in this derivation. All the expansion coefficients A
n

(n = 2, 4, . . . ) in eq. (26) can be obtained in the same procedure for each mode
n,0
(r

) with
the corresponding imaginary wave number:

(31)
the normalization condition of
n,0
(r

) and a simple relation: . Here a

is
the oscillator length of the (relative) transverse motion,

(32)
Exact Nonlinear Dynamics in Spinor Bose-Einstein Condensates

37
Recall that due to the condition (24) the value inside the parentheses in eq. (31) is positive
definite. Thus, the expression for the wave function along the z axis reads

(33)
where the function is defined as

(34)
the sum over s= n/2 originates from the sum appearing in eq. (26). We have chosen the
value
0,0
(0) to be real and positive. By subtracting and adding a sum,

(35)
to the function , and then, collecting term from the Taylor series of exp and
with respect to , one can show an expansion,

(36)
Here the zero-order term of the expansion has a form,

(37)
with

(38)
and

(39)
Substituting eq. (33) with eq. (36) into eq. (30), we get
reg
in an explicit form.
We then write the final expression of the one-dimensional scattering amplitudes (25) as

(40)
with the 1D scattering length:

(41)
Nonlinear Dynamics

38
3.3 Effective one-dimensional coupling strength
The expression (40) is an exact result for the potential (21) with arbitrary strength of the
transverse confinement a

. For atoms with the low kinetic energy, we can drop

term in the denominator of the scattering amplitudes (40), obtaining a one-dimensional
contact potential,

(42)
were the coupling strength:

(43)
Note that a simple average of the three-dimensional coupling g = 4=
2
a/m over the
transverse ground state only reproduces the coefficient of (43),

(44)
The resonance factor 1/[1 C(a/a

)] implies a possibility to control the strength of

atomatom scattering via tuning a confinement potential a

. The physical origin of the CIR is

attributed to a zero-energy Feshbach resonance in which the transverse modes of the
confining potential assume the roles of open and closed scattering channels.
4. Spinor BoseEinstein condensate
In this section, we extend the model of a single component condensate discussed in Sec. 2 to
that of a multicomponent condensate with the spin degrees of freedom, which we call a
spinor condensate for short (Pethick & Smith; 2002). In terms of spin, we mean the
hyperfine spin of atoms in this chapter.
4.1 Hamiltonian
The hyperfine spin f is defined by f = s + i, where s and i denote the electronic and nuclear
spins of the atoms. For simplicity, we consider bosons with the hyperfine spin f = 1. This
includes alkalis with nuclear spin i = 3/2 such as
7
Li,
87
Rb, and
23
Na. Alkali bosons with f > 1
such as
85
Rb (with i = 5/2), and
133
Cs (with i = 7/2) may have even richer structures.
Atoms in the f = 1 state are characterized by a vectorial field operator with the components
subject to the hyperfine spin manifold. The three-component field ,
where the superscript T denotes the transpose, satisfies the bosonic commutation relations:

(45)
In order to discuss the properties of spinor Bose gases, we start with the following second
quantized Hamiltonian,

(46)

(47)
Exact Nonlinear Dynamics in Spinor Bose-Einstein Condensates

39

(48)

(49)
where U
trap
(r) is the external trap potential, v(r r) expresses the two-body interaction and
subscripts {, , , = 1, 0,1} denote the components of the spin. The last term in eq. (46),
, is the response to an external magnetic field p (the linear Zeeman effect). This response
to the magnetic field necessarily selects one of several possible ground states, or the so-
called weak field seeking state, m
f
= 1 for f = 1 case where the spin degrees of freedom are
frozen. We set p = 0 throughout this chapter.
Due to the BoseEinstein statistics, the total spin F = f
1
+ f
2
of any two bosons whose relative
orbital angular momentum is zero should be restricted to even, F = 2 f , 2 f 2, . . . , 0. Thus,
the interatomic interaction

(r r) can be divided into several sectors labeled by F as

(50)
where

is the projection operator and g
F
characterizes the strength of the binary interaction
between bosonic atoms with the total spin F. This coupling constant g
F
is related to the
corresponding s-wave scattering length a
F
as

(51)
For f = 1 bosons, since F takes only on values 0 and 2, we can rewrite the potential (r r) in
a simple form using the following two properties of the projection operators the
completeness of the operators,

(52)
where is an identity operator, and the product of the angular momentum operators,

(53)
where a hat on f means an operator as projection. Solving these equations (52), (53) for
and , we obtain the form of the interaction in terms of the angular momentum
operators,

(54)
In this expression,

(55)
which are the magnitude of the density-density interaction and of the spin-spin interaction,
respectively. Thus, the interaction Hamiltonian is rewritten as
Nonlinear Dynamics

40

(56)
where we may use the following expressions of spin-1 matrices f = (f
x
, f
y
, f
z
) as

(57)
A construction of the interaction Hamiltonian for a general hyperfine spin f can be found in
(Ueda & Koashi; 2002).
4.2 f = 1 spinor condensate in quasi 1D regime
From now on, we assume that the system is quasi-one dimensional: the trapping potential is
suitably anisotropic such that the transverse spatial degrees of freedom (y-z plain) is
factorized from the longitudinal (x axis) and all the hyperfine states are in transverse ground
state.
As derived in Sec. 2, in the mean-field theory of the spinor BEC, the assembly of atoms in
the f = 1 state is characterized by a vectorial order parameter:

(58)
where the subscripts {1, 0,1} denote the magnetic quantum numbers with the components
subject to the hyperfine spin space. The normalization is imposed as

(59)
where N
T
is the total number of atoms.
According to the discussion in Sec. 3, the effective 1D couplings

and are represented by

(60)
where a
F
is the 3D s-wave scattering length of the total hyperfine spin F = 0, 2 channels,
respectively, and a

is the size of the ground state in the (relative) transverse motion.

Thus, the Gross-Pitaevskii energy functional of this system is given by

(61)
with the particle number and spin densities, respectively, defined by

(62)
Exact Nonlinear Dynamics in Spinor Bose-Einstein Condensates

41
The coupling constants and are connected to those in eqs. (60) (cf. eq. (43)) as

(63)
The time-evolution of spinor condensate wave function (x, t) can be derived from

(64)
Substituting eq. (61) into eq. (64), we get a set of equations for the longitudinal wave
functions of the spinor condensate:

(65a)

(65b)

(65c)
5. Integrable model
To analyze the dynamical properties of the coupled system (65), we propose an integrable
model as follows (Ieda et al.; 2004a,b). We consider the system with the coupling constants,

(66)
This situation corresponds to attractive mean-field interaction

< 0 and ferromagnetic spin-
exchange interaction

< 0. Note that in preceding investigations of spinor condensates
(Pethick & Smith; 2002), mean-field interaction is assumed to be repulsive c
0
> 0 and far
exceeding spin-exchange interaction in the magnitude c
0
|c
1
| in line with experimental
data. Thus, the parameter regime (66) was not been explored in detail.
The effective interactions between atoms in a BEC have been tuned with a Feshbach
resonance (Pethick & Smith; 2002). In spinor BECs, however, we should extend this to
alternative techniques such as an optically induced Feshbach resonance or a confinement
induced resonance (Olshanii; 1998), which do not affect the rotational symmetry of the
internal spin states. In the latter, the above condition is surely obtained by setting

(67)
in eq. (60) when
Nonlinear Dynamics

42

(68)
It is worth noting that the integrable property itself is independent of the sign of ( ) as
far as their magnitudes are equal to each other. The opposite sign case, i.e., = c > 0, can
be analyzed in the same manner (Uchiyama et al.; 2006).
In the dimensionless form:

(69)
where time and length are measured in units of

(70)
respectively, we rewrite eqs. (65) as follows, (we omit the arguments (x, t) hereafter.)

(71a)

(71b)

(71c)
Now we find that these coupled equations (71) are equivalent to a 22 matrix version of
nonlinear Schrdinger (NLS) equation:

(72)
with an identification,

(73)
Since the matrix NLS equation (72) is completely integrable (Tsuchida & Wadati; 1998), the
integrability of the reduced equations (71) are proved automatically (Ieda et al.; 2004a).
Remark that the general M L matrix NLS equation is also integrable. It is worthy to search
other integrable models for higher spin case (Uchiyama et al.; 2007).
5.1 Soliton solution
We summarize an explicit formula for the soliton solution of the 2 2 matrix version of NLS
equation (72) with eq. (73) by considering a reduction of a general formula obtained in
(Tsuchida & Wadati; 1998).
Under the vanishing boundary condition, one can apply the inverse scattering method (ISM)
to the nonlinear time evolution equation (72) associated with the generalized Zakharov-
Shabat eigenvalue problem:

(74)
Exact Nonlinear Dynamics in Spinor Bose-Einstein Condensates

43
Here
1
and
2
take their values in 2 2 matrices. The complex number k is the spectral
parameter. I is the 2 2 unit matrix. The 2 2 matrix Q plays a role as a potential function in
this linear system. According to (Tsuchida & Wadati; 1998), N-soliton solution of eq. (72)
with eq. (73) is expressed as

(75)
where the 2N 2N matrix S is given by

(76)
Here we have introduced the following parameterizations:

(77)

(78)
The 2 2 matrices
j
normalized to unity in a sense of the square norm,

(79)
must take the same form as Q from their definition. We call them polarization matrices,
which determine both the populations of three components {1, 0, 1} within each soliton and
the relative phases between them. The complex constants k
j
denote discrete eigenvalues,
each of which determines a bound state by the potential Q.
j
are real constants which can be
used to tune the initial displacements of solitons. It is worth noting that all x and t
dependence is only through the variables
j
(x, t). As we shall see in Sec. 6, the real part of

j
(x, t) represents the coordinate for observing soliton-js envelope while the imaginary part
of it represents the coordinate for observing soliton-js carrier waves.
The same procedure can be performed for nonvanishing boundary conditions (Ieda et al.;
2007) which is relevant to formation of spinor dark solitons (Uchiyama et al.; 2006).
Equation (72) is a completely integrable system whose initial value problems can be solved
via, for example, the ISM (Tsuchida & Wadati; 1998) (Ieda et al.; 2007). The existence of the
r-matrix for this system guarantees the existence of an infinite number of conservation laws
which restrict the dynamics of the system in an essential way. Here we show explicit forms
of some conserved quantities, i.e., total number, total spin (magnetization), total momentum
and total energy.

(80)

(81)
Nonlinear Dynamics

44

(82)

(83)

(84)

(85)

(86)

(87)
Here tr{} denotes the matrix trace and = (
x
,
y
,
z
)
T
are the Pauli matrices,

(88)
6. Spin property of one-soliton solution
In this section, we discuss one-soliton solutions and classify them by their spin states. If we
set N = 1 in the formula (75) we obtain the one-soliton solution:

(89)
where

(90)

(91)
We have omitted the subscripts of the soliton number. Here and hereafter, the subscripts R
and I denote real and imaginary parts, respectively. Throughout this section, we set k
R
>0
without loss of generality. We remark the significance of each parameter/coordinate as
follows,

We use the term amplitude to indicate the peak(s) height of solitons envelope. Actual
amplitude should be represented as k
R
multiplied by a factor from 1 to which is
Exact Nonlinear Dynamics in Spinor Bose-Einstein Condensates

45
determined by the type of polarization matrices. The explicit form will be shown later. As
mentioned before, solitons motion depends on both x and t via variables
R
and
I
, from
which we can see the meaning of velocity of soliton.
From a total spin conservation, one-soliton solution can be classified by the spin states. We
shall show that the only two spin states are allowable,

(92a)

(92b)
Substituting eqs. (89)(91) into eq. (83), we obtain the local spin density of the one-soliton
solution:

(93)
We also give the explicit form of the number density:

(94)
To clarify the physical meaning of det, we define here another important local density as

(95)
This quantity measures the formation of singlet pairs. Note that these pairs are
distinguished from Cooper pairs of electrons or those of
3
He owing to the different statistical
properties of ingredient particles. Since (x, t) does not contribute to the magnetization of
the soliton, it is invariant under any spin rotation. As far as ground state properties are
concerned, it is not necessary to introduce (x, t) for a system of spin-1 bosons, while a
counterpart to eq. (95) plays a crucial role for spin-2 case (Ueda & Koashi; 2002). As we shall
show later, however, it is useful to characterize solitons within energy degenerated states.
In the case of the one-soliton solution (89), the singlet pair density is proportional to the
determinant of the polarization matrix ,

(96)
This suggests that det represents the magnitude of the singlet pairs. For the general
N-soliton case, this singlet pair density can vary after each collision of solitons and is not the
conserved density. The detail will be discussed at the end of this section.
In what follows, we classify spin states of the one-soliton solution based on the values of
det.
6.1 Ferromagnetic state
Let det = 0, then eq. (89) becomes a simple form:

(97)
Nonlinear Dynamics

46
Now all of m
F
= 0, 1 components share the same wave function. Their distribution in the
internal state reflects directly the elements of the polarization matrix . One can see the
meaning of each parameter listed above. By definition, the singlet pair density (96) vanishes
everywhere. Thus, this type of soliton belongs to the ferromagnetic state and will be referred
to as a ferromagnetic soliton. The total number of atoms is given by integrating eq. (94) as

(98)
The total magnetization (82) becomes

(99)
with the modulus, . Equation (99) is connected to through a
gauge transformation and a spin rotation.
Next, we calculate the total momentum and the total energy of the ferromagnetic soliton.
Substituting eq. (97) into eqs. (84), (86) and using det = 0, we obtain

(100)
respectively. In infinite homogeneous 1D space as considered here, it can be shown that a
single component GP equation for BEC with attractive interactions, i.e., the self-focusing
NLS equation possesses the one-soliton solution that minimizes the total energy for fixed
number of particles and total momentum. This remains true for the spinor GP equations
(71). As we will see later, for given number of N
T
, the stationary (k
I
= 0) one-soliton solution
in the ferromagnetic state is the ground state of this system. On the other hand, in finite 1D
space case, the ground state is subject to a quantum phase transition between uniform and
soliton states (Kanamoto et al.; 2002).
6.2 Polar state
If det 0, the local spin density has one node, i.e., f(x
0
, t) = 0 at a point:

(101)
for each moment t. Setting x= x x
0
and A
1

2|det|, we get

(102)
Since each component of the local spin density is an odd function of x, its average value is
zero,

(103)
Exact Nonlinear Dynamics in Spinor Bose-Einstein Condensates

47
This implies that this type of soliton, on the average, belongs to the polar state (Pethick &
Smith; 2002). Let us also rewrite the number density (94) as

(104)
To elaborate on this type of soliton, we further divide into two cases.
(i) A
1

= 2|det| = 1 (

=0).
Under this constraint, we find the local spin (102) vanishes everywhere. Solitons in this state
possess the symmetry of polar state locally. We, therefore, refer to only those solitons as
polar solitons. Considering eq. (89) with the above condition, we recover a normal sech-type
soliton solution:

(105)
Note that the amplitude of soliton is different from that of the ferromagnetic soliton, which
leads to a relation between the total number and the spectral parameter as

(106)
The total momentum and the total energy are given by

(107)
respectively. The difference between ferromagnetic soliton energy and polar soliton energy
with the same number of atoms N
T
is

(108)
which is a natural consequence of the ferromagnetic interaction, i.e., = c < 0.
(ii) A
1

= 2|det|< 1.
In this case, the local spin retains nonzero value, although the average spin amounts to be
zero. The density profile (104) has the following structure. When A > 2, a peak of the density
splits into two (Fig. 1) due to different density profiles of m
F
= 0, 1 components.
For a large value of A, namely, when det gets close to zero, such twin peaks separate
away. In consequence, they behave as if a pair of two distinct ferromagnetic solitons with
antiparallel spins, traveling in parallel with the same velocity and the amplitudes half as
much as that of the polar soliton (A = 1) in the density profile [see the inset of Fig. 1(a) and
Fig. 1 (b)].
Hence, solitons of this type will be referred to as split solitons. The total number is the same
as the case (i),

(109)
The total momentum and the total energy are the same values as those in the case (i):

(110)
Nonlinear Dynamics

48

(a) (b)
Fig. 1. The density profiles of eq. (104). (a)We set k
R
= 0.5, and A = 1 (solid line), 2 (dashed
line), 5 (dash-dot line), 20 (dotted line). The inset shows a split soliton for A = 10
4
, consisting
of two ferromagnetic like solitons with the same velocity. (b) The density profiles of eq. (104)
(solid line) for k
R
= 0.5 and A = 10
4
, and the three components, m
F
= 0 (dashed line), m
F
= 1
(dotted line) and m
F
= 1 (dash-dot line) are shown simultaneously.
This degeneracy is ascribed to the integrable condition for the coupling constants, i.e.,
= . Comparing case (i) with case (ii), we find that a variety of dissimilar shaped solitons
are degenerated in the polar state. To characterize them, we can use, instead of A, a physical
quantity defined as

(111)
which is a monotone decreasing function of A [1, ); the maximum value, N
T
, at A = 1
(polar soliton) and limiting to 0 at A (ferromagnetic soliton). In this sense, S has the
meaning as the total singlet pairs of the whole system. As noted above, S is not the
conserved quantity in general (N 2); all the conserved densities should be expressed by the
matrix trace of products of Q

, Q and their derivatives (Tsuchida & Wadati; 1998) as eqs.

(81), (83), (85), and (87) while |(x, t)| is not. Nevertheless, S can be used to label solitons in
the polar state because it dose not change in the meanwhile prior to the subsequent collision.
7. Two-soliton collision and spin dynamics
In this section, we analyze two-soliton collisions in the spinor model. The two-soliton
solutions can be obtained by setting N = 2 in eq. (75). The derivation is straightforward but
rather lengthy. An explicit expression of the two-soliton solution is given in Appendix of
(Ieda et al.; 2004b) and, here, compute asymptotic forms of specific two-soliton solutions as
t , which define the collision properties of two-soliton in the spinor model.
For simplicity, we restrict the spectral parameters to regions:

(112a)
Exact Nonlinear Dynamics in Spinor Bose-Einstein Condensates

49
Under the conditions, we calculate the asymptotic forms in the final state (t) from those
in the initial state (t). Since each solitons envelope is located around x 2k
jI
t, soliton-1
and soliton-2 are initially isolated at x , and then, travel to the opposite directions at a
velocity of 2k
1I
and 2k
2I
, respectively. After a head-on collision, they pass through without
changing their velocities and arrive at x in the final state. Collisional effects appear not
only as usual phase shifts of solitons but also as a rotation of their polarization.
According to the classification of one-soliton solutions in the previous section, we choose the
following three cases: i) Polar-polar solitons collision, ii) Polar-ferromagnetic solitons
collision, iii) Ferromagnetic-ferromagnetic solitons collision. As we shall see later, the polar
soliton dose not affect the polarization of the other solitons apart from the total phase factor.
On the other hand, ferromagnetic solitons can rotate their partners polarization, which
allows for switching among the internal states.
7.1 Polar-polar solitons collision
We first deal with a collision between two polar solitons defined by k
j
and
j
(j = 1, 2) with
the conditions (112) and

, equivalently,

(113)
In the asymptotic regions, we can consider each soliton separately. Thus, the initial state is
given by the sum of two polar solitons as

(114)
where the asymptotic form of soliton-j (j = 1, 2) is

(115a)
These can be proved by taking the limit
2R
with keeping
1R
finite and, vice versa,

1R
with
2R
fixed. Phase factors which come from the values of |det
j
| are absorbed
by the arbitrary constants inside
jR
. In the final state, the opposite limit
2R
with
keeping
1R
finite and
1R
with |
2R
| < yields

(116)
where

(117)
with

(118)

(119)
Nonlinear Dynamics

50


(a) (b)
Fig. 2. Density plots of |
0
|
2

(a) and |
1
|
2

(b) for a polar-polar collision. Soliton 1 (left
mover) carries only 0 component and soliton 2 (right mover) consists of 1 components.
The parameters used here are k
1
= 0.25 0.25i, k
2
= 0.5 + 0.25i,
1
= 1/ ,
1
=
1
= 0,
2
= 0,

2
=
2
= 1/ .
Equations (115) and (117) are the same form as polar one-soliton solution (105). Collisional
effects appear only in the position shift (118) and the phase shifts (119). In Figs. 2, we show
the polar-polar collision with
1
= 1/ ,
1
=
1
= 0 and
2
= 0,
2
=
2
= 1/ . Thus, the
partial number N
j
, magnetization F
j
, momentum P
j
, and energy E
j
are defined for the
asymptotic form of soliton-j and calculated in the same manner as the previous section. The
integrals of motion are represented by the sum of those quantities for each soliton.
Moreover, we can prove that

(120)
which are by themselves conserved through the collision. In this sense, the polar-polar
collision is basically the same as that of the single-component NLS equation.
7.2 Polar-ferromagnetic solitons collision
Under the condition (112), we set soliton 1 to be polar soliton and soliton 2 to be
ferromagnetic soliton:

(121)
Then, the initial state is represented by eq. (114) with

(122a)

(122b)
The final state is given by eq. (116) with
Exact Nonlinear Dynamics in Spinor Bose-Einstein Condensates

51

(123a)

(123b)
Here we have defined

(124)

(125)
and also used eqs. (118), (119). Normalization of the new polarization matrix (125) turns out
to be unity,

(126)
The determinant of it becomes

(127)
We can see clearly that the initial polar soliton breaks into a split type,
after the collision with a ferromagnetic one. Only when where the spinor
part of wave function of two initial solitons is orthogonal to each other, we have .
Then, eqs. (123) are reduced to

(128a)

(128b)
which means that the polar soliton keeps its shape against the collision and shows no
mixing among the internal states except for the total phase shift. On the other hand, because
of the total spin conservation, the ferromagnetic soliton always retains its polarization
matrix and shows only the position and phase shifts similar to those of the polar-polar case.
In Fig. 3, we have density plots of a polar-ferromagnetic collision with the parameters
shown in the caption. These pictures correspond to each component of the exact two-soliton
solution for one collisional run. For simplicity, we choose the parameters to have
|
1
| = |
1
|. The polar soliton (soliton 1) initially prepared in m
F
= 1 are switched into a
soliton with a large population in m
F
= 0 and the remnant of m
F
= 1 after the collision.
Through the collision, the ferromagnetic soliton (soliton 2) plays only a switcher, showing
no mixing in the internal state of itself outside the collisional region, as clearly seen in eq.
(123b). In general, this kind of a drastic internal shift of polar soliton is likely observed for
large values of which appears in eqs. (124), (125). Although all the conserved
quantities such as the number of particles and the averaged spin of individual solitons are
Nonlinear Dynamics

52

(a) (b)
Fig. 3. Density plots of |
0
|
2

(a) and |
1
|
2

(b) for a polar-ferromagnetic collision. Soliton 1
(left mover) is a polar soliton and soliton 2 (right mover) is a ferromagnetic soliton.
The parameters used here are k
1
= 0.25 0.25i, k
2
= 0.5 + 0.25i,
1
= 0,
1
=
1
= 1/ ,

2
=
2
=
2
= 1/2.
invariant during this type of collision, the fraction of each component can vary not only in
each soliton level but also in the total after the collision. This contrasts to an intensity
coupled multicomponent NLS equation in which the total distribution among all
components is invariant throughout soliton collisions while a switching phenomenon
similar to Fig. 3 can be observed (Radhakrishnan et al.; 1997).
7.3 Ferromagnetic-ferromagnetic solitons collision
Finally, we discuss the collision between two ferromagnetic solitons,

(129)
The asymptotic forms are obtained for the initial state, where

(130)
and for the final state, where

(131a)
Here we have defined

(132)
and, for (j, l) = (1,2) or (2,1),

(133)
Exact Nonlinear Dynamics in Spinor Bose-Einstein Condensates

53
which are shown to be normalized in unity,

(134)
Each polarization matrix
j
of a ferromagnetic soliton can be expressed by three real
variables
j
,
j
,
j
as

(135)
In this expression, the polarization matrices in the initial state
j
and in the final state are
given by

(136)
where, with (j, l) = (1,2), (2,1),

(137)
This defines the collision property for the ferromagnetic-ferromagnetic soliton collision.
We can gain a better understanding of the collision between two ferromagnetic solitons by
recasting it in terms of the spin dynamics. The total spin conservation restricts the motion of
the spin of each soliton on a circumference around the total spin axis [Fig. 4(a)]. It will be
interpreted as a spin precession around the total magnetization.
We calculate the magnetization for each soliton to investigate their collision. In the initial
state, following eq. (99), we have the spin of soliton-j as

(138)
Thanks to the scattering property (137), the final state spins can be obtained through F
1,2
by
(139)
where

(140)
The conserved total spin, is given by

(141)
Nonlinear Dynamics

54
Considering spin rotation around the total spin F
T
, we can find rotated spin as

(142)
where

(143)
with

(144)
The rotation angle is determined by setting

through eqs. (139) and (142),

(145)
For the case that the magnitudes of the amplitude and velocity for each ferromagnetic
soliton are, respectively, identical with each other, |k
1R
| = |k
2R
| N
T
/4, |k
1I
| = |k
2I
| k
I
,
the final state magnetizations (139) are given by

(146)
where (j, l) = (1,2), (2,1). The rotation angle depends only on the ratio k
I
/k
R
and the
magnitude of the normalized total magnetization, F |F
T
|/N
T
, as

(147)
The principal value should be taken for the arccosine function: 0 arccos x .
Setting k
I
k
R
in eq. (147), one gets the small rotation angle, 0. In the opposite case, k
I
k
R
, each spin of two colliding solitons almost reverses its orientation, . Recall that k
I
is the speed of soliton. We can understand these phenomena since a slower soliton spends
the longer time inside the collisional region. Figure 4 shows the velocity dependence of the
rotation angle for various initial normalized spins. When F = 1, which corresponds to the
case of antiparallel spin collision, the spin precession can not occur as shown by the dotted
line in Fig. 4(b).
In Fig. 5Fig. 7, we give examples of this type of collisions for different k
I
, with the other
conditions fixed, to illustrate the velocity dependence. The initial normalized spin for the
parameter set given in the captions is F = 0.5. The rotation angles are 0.2, 0.5 and
0.9 for Fig. 5, Fig. 6 and Fig. 7, respectively. The internal shift
1

1
, and vice versa,
gradually increase by slowing down the velocity of the solitons.
Exact Nonlinear Dynamics in Spinor Bose-Einstein Condensates

55

(a) (b)
Fig. 4. (a) Schematic of spin precession of two colliding ferromagnetic solitons. (b) Velocity
dependence of the rotational angle in spin precession for the different initial relative angles,
F = 1 (solid line), 0.5 (dashed line), 0.0157 (dash-dot line) and 0 (dotted line).

Fig. 5. Density plots of (a) |
0
|
2
, (b) |
1
|
2

and (c) |
1
|
2

for a fast ferromagnetic-
ferromagnetic collision. The parameters used here are k
1
= 0.5 0.75i, k
2
= 0.5 + 0.75i,
1
=
4/17,
1
= 16/17,
1
= 1/17,
2
= 4/17,
2
= 1/17,
2
= 16/17.

Fig. 6. Density plots of (a) |
0
|
2
, (b) |
1
|
2

and (c) |
1
|
2

for a medium speed
ferromagnetic-ferromagnetic collision. The parameters are the same as those of Fig. 5 except
for k
1I
= 0.25, k
2I
= 0.25.
Nonlinear Dynamics

56

Fig. 7. Density plots of (a) |
0
|
2
, (b) |
1
|
2

and (c) |
1
|
2

for a slow ferromagnetic-
ferromagnetic collision. The parameters are the same as those of Fig. 5 except for k
1I
= 0.05,
k
2I
= 0.05.
8. Concluding remarks
The soliton properties in spinor BoseEinstein condensates have been investigated.
Considering two experimental achievements in atomic condensates, the matter-wave soliton
and the spinor condensate, at the same time, we have predicted some new phenomena.
Based on the results provided in Sec. 24, in Sec. 5 we have introduced the new integrable
model which describes the dynamics of the multicomponent matter-wave soliton. The key
idea is finding the integrable condition of the original coupled nonlinear equations, i.e., the
spinor GP equations derived in Sec. 4. The integrable condition expressed by the coupling
constants, which is accessible via the confinement induced resonance explained in Sec. 3.
In Sec. 6, we classify the one-soliton solution. There exist two distinct spin states:
ferromagnetic, |F
T
| = N
T
and polar, |F
T
| = 0. In the ferromagnetic state, the spatial part and
the spinor part of the soliton are factorized (ferromagnetic soliton). In the polar state,
dissimilar shaped solitons which we call polar soliton for f(x) = 0 and split soliton otherwise
are energetically degenerate. The polar soliton has one peak and the space-spinor
factorization holds. On the other hand, a split soliton consists of twin peaks and the three
components show different profiles. Changing the polarization parameters one may control
the peak distance continuously.
In Sec. 7, we have analyzed two-soliton solutions which rule collisional phenomena of the
multiple solitons. Specifying the initial conditions, we have demonstrated two-soliton
collisions in three characteristic cases: polar-polar, polar-ferromagnetic, ferromagnetic-
ferromagnetic. In their collisions, the polar soliton is always passive which means that it
does not rotate its partners polarization while the ferromagnetic soliton does. Thus, in the
polar-ferromagnetic collision, one can use the polar soliton as a signal and ferromagnetic
soliton as a switch to realize a coherent matter-wave switching device. Collision of two
ferromagnetic solitons can be interpreted as the spin precession around the total spin. The
rotation angle depends on the total spin, amplitude and velocity of the solitons. Only
varying the velocity induces drastic change of the population shifts among the components.
Stability of spinor solitons has been investigated numerically and perturbatively (Li et al.;
2005) (Dabrowska-Wster et al.; 2007) (Doktorov et al.; 2008). It is also interesting to pursue
Exact Nonlinear Dynamics in Spinor Bose-Einstein Condensates

57
the soliton dynamics of spinor condensates under longitudinal harmonic trap (Zhang et al.;
2007). Recently, the integrability of the spinor GP equation has been studied in detail
(Gerdjikov et al.; 2009). The behavior of spinor solitons shows a variety of nonlinear
dynamics and it is worth exploring them experimentally.
9. Acknowledgment
This work was supported by Grant-in-Aid for Scientific Research No. 20740182 from MEXT.
10. References
Dabrowska-Wster, B. J.; Ostrovskaya, E. A.; Alexander, T. J.; Kivshar, Y. S. (2007).
Multicomponent gap solitons in spinor BoseEinstein condensates. Physical Review
A, 75, 2, (April 2008) 023617-1023617-11, ISSN 1094-1622.
Doktorov, E. V.; Wang, J. D.; Yang, J. K. (2008). Perturbation theory for bright spinor Bose-
Einstein condensate solitons. Physical Review A, 77, 4, (April 2008) 043617-1043617-
11, ISSN 1094-1622.
Gerdjikov, V.S.; Kostov, N.A.; Valchev, T. I. (2009). Solutions of multi-component NLS
models and Spinor BoseEinstein condensates. Physica D: Nonlinear Phenomena, 238,
15, (July 2009) 13061310, ISSN 0167-2789.
Ieda, J.; Tsurumi, T.; Wadati, M. (2001). BoseEinstein Condensation of Ideal Bose Gases.
Journal of the Physical Society of Japan, 70, 5, (May 2001) 12561259, ISSN 1347-4073.
Ieda, J.; Miyakawa, T.; Wadati, M. (2004). Exact Analysis of Soliton Dynamics in Spinor
BoseEinstein Condensates. Physical Review Letters, 93, 19, (November 2004) 194102-
1194102-4, ISSN 1079-7114.
Ieda, J.; Miyakawa, T.; Wadati, M. (2004). Matter-Wave Solitons in an F=1 Spinor Bose
Einstein Condensate. Journal of the Physical Society of Japan, 73, 11, (November 2004)
29963007, ISSN 1347-4073.
Ieda, J.; Uchiyama, M.; Wadati, M. (2007). Inverse scattering method for square matrix
nonlinear Schrodinger equation under nonvanishing boundary conditions. Journal
of Mathematical Physics, 48, 1, (January 2007) 013507-1013507-19, ISSN 1089-7658.
Khaykovich, L.; Schreck, F.; Ferrari, G.; Bourdel, T.; Cubizolles, J.; Carr, L. D.; Castin, Y.;
Salomon, C. (2002). Formation of a Matter-Wave Bright Soliton. Science, 96, 5571,
(May 2002) 12901293, ISSN 1095-9203.
Kanamoto, R; Saito, H; Ueda, M. (2003). Quantum phase transition in one-dimensional
Bose Einstein condensates with attractive interactions. Physical Review A, 67, 1,
(January 2003) 013608-1013608-7, ISSN 1094-1622.
Li, L.; Li, Z.; Malomed, B. A.; Mihalache, D.; Liu, W. M. (2005). Exact soliton solutions and
nonlinear modulation instability in spinor BoseEinstein condensates. Physical
Review A, 72, 3, (September 2005) 033611-1033611-11, ISSN 1094-1622.
Meystre, P. (2001). Atom Optics, Springer-Verlag New York, Inc., New York.
Olshanii, M. (1998). Atomic Scattering in the Presence of an External Confinement and a Gas
of Impenetrable Bosons. Physical Review Letters, 81, 5, (August 1998) 938941, ISSN
1079-7114.
Pethick, C. J. & Smith, H. (2002). BoseEinstein condensation in dilute Gases, Cambridge
University Press, Cambridge. Also 2nd ed. (2008).
Nonlinear Dynamics

58
Stenger, J.; Inoue, S.; Stamper-Kurn, D. M.; Miesner, M. R.; Chikkatur, A. P.; Ketterle, W.
(1998). Spin domains in ground-state BoseEinstein condenstate. Nature, 396, 6709,
(November 1998) 345348, ISSN 0028-0836.
Radhakrishnan, R.; Lakshmanan, M.; Hietarinta, J. (1997). Inelastic collision and switching of
coupled bright solitons in optical fibers. Physical Review E, 56, 2, (August 1997)
22132216, ISSN 1550-2376.
Strecker, K. E.; Partridge, G. B.; Truscott, A. G.; Hulet, R. G. (2002). Formation and
propagation of matter-wave soliton trains. Nature, 417, (May 2002) 150153, ISSN
0028-0836.
Tsuchida, T. & Wadati, M. (1998). The Coupled Modified Kortewegde Vries Equations.
Journal of the Physical Society of Japan, 67, 4, (April 1998) 11751187, ISSN 1347-4073.
Uchiyama, M.; Ieda, J.; Wadati, M. (2006). Dark Solitons in F=1 Spinor BoseEinstein
Condensate. Journal of the Physical Society of Japan, 75, 6, (June 2006), 064002-1
064002-9, ISSN 1347-4073.
Uchiyama, M.; Ieda, J.; Wadati, M. (2007). Multicomponent Bright Solitons in F=2 Spinor
Bose Einstein Condensates. Journal of the Physical Society of Japan, 76, 7, (July 2007),
074005- 1074005-6, ISSN 1347-4073.
Ueda, M. & Koashi, M. (2002). Theory of spin-2 BoseEinstein condensates: Spin
correlations, magnetic response, and excitation spectra. Physical Review A, 65, 6,
(May 2002), 063602-1063602-22, ISSN 1094-1622.
Zhang, W.; Mstecaplolu, . E.; You, L. (2007). Solitons in a trapped spin-1 atomic
condensate. Physical Review A, 75, 4, (April 2007), 043601-1043601-8, ISSN 1094-
1622.
3
A Conceptual Model for the Nonlinear Dynamics
of Edge-localized Modes in Tokamak Plasmas
Todd E. Evans
1
, Andreas Wingen
2
,
Jon G. Watkins
3
and Karl Heinz Spatschek
2

1
General Atomics,
2
Institute for Theoretical Physics, Heinrich-Heine-University,
3
Sandia National Laboratory,
1,3
United States
2
Germany
1. Introduction
High performance magnetically confined toroidal plasmas, such as those required for the
operation of a tokamak based fusion power plant, suffer from a troubling type of repetitive
edge instability known as edge-localized modes (ELMs). Magnetohydrodynamic (MHD),
peeling-ballooning, theory predicts that these instabilities are driven by a large current
density and pressure gradient that forms at the plasma edge as a consequence of the
enhanced confinement levels achieved in high performance H-mode plasmas. Although
ELMs are a common feature of high confinement tokamak plasmas, there are significant
gaps in our understanding of how these instabilities scale with the geometry of the plasma
and operating conditions expected in large tokamaks that are required for the generation of
fusion power. Thus, there is an urgent need for a model that can be validated with
experimental data from existing smaller tokamaks.
Here, we present a conceptual model describing the topological evolution of the magnetic
separatrix, in a tokamak plasma with a dominate lower hyperbolic point. Subsequently, the
nonlinear dynamics of the ELM instability, prescribed by the evolving separatrix topology,
is discussed. The model invokes a feedback amplification mechanism that causes the stable
and unstable invariant manifolds of the separatrix, comprising a homoclinic tangle
(Guckenheimer & Holmes, 1983), to grow explosively as the topology of the separatrix
manifolds unfold. The amplification process is driven by the rapid growth of helical, field-
aligned, thermoelectric currents that flow through relatively short edge plasma flux tubes
connecting high heat flux wall structures, known as divertor target plates, on both sides of
the plasma. These thermoelectric currents produce magnetic fields that couple to the
separatrix and modify its 3D (topological) structure. As the lobes of the separatrix tangle
grow, their area of intersection with the divertor target plates increases along with the size
of the flux tubes connecting target plates on both sides of the plasma. This increases the
thermoelectric current flow and completes the feedback loop. Numerical simulations have
shown that our model is consistent with measurements of the currents flowing between the
target plates and with camera images of the heat flux patterns on the divertor target plates
Nonlinear Dynamics

60
(Wingen, et al. 2009a). In addition, this model suggests that nonaxisymmetric external
magnetic coils can be used to force higher order separatrix bifurcations to prevent ELMs.
In the following sections, we discuss why ELMs are an important problem in tokamaks,
review what is known experimentally and theoretically about the characteristics of ELMs,
present a conceptual framework for the nonlinear evolution of an ELM and discuss results
from numerical simulations of the proposed model. We show a sequence of topological
bifurcations that are observed in the numerical simulations and discuss how these result in a
separatrix topology that produces heat flux patterns which are consistent with those
measured by infrared cameras in the DIII-D (Luxon, 2002) tokamak. A general description of
tokamaks and tokamak physics is given by Callen et al., (1992), Evans (2008) and Wesson
(2004).
2. Properties of ELMs in high performance tokamak plasmas
2.1 ELM dynamics
Type-I ELMs are naturally occurring MHD instabilities that release large bursts of particles
and energy from the boundary of the plasma (Suttrop 2000). These very fast growing
instabilities share properties that are somewhat similar to the eruption of prominences and
flares from the solar photosphere (Evans, et al., 1996). More specifically, expanding hot
plasma filaments carrying energy, particles and momentum away from the confined plasma
volume into the surrounding space are associated with these complex dynamical plasma
events that form on the surface of the sun and at the edge of a tokamak discharge.
Tokamaks operating in high confinement H-modes, with strong edge transport barriers, rely
exclusively on the formation of a large pressure gradient near the surface of the plasma to
obtain sufficiently high central temperatures and densities to carry out fusion energy
experiments in these devices. These large pressure gradients are believed to drive edge
MHD instabilities, referred to as peeling-ballooning modes, that are responsible for the
onset of ELMs (Snyder, et al., 2005). Since ELMs periodically release particles and energy
from the edge of the plasma, they limit the size of the pressure gradient that can be obtained
in tokamaks (Fenstermacher, et al. 2003). Scaling studies suggest that this limits the
maximum temperature of the core plasma and thus the ultimate performance of the
tokamak (Loarte, et al. 2003). In addition, the impulsive energy and particle flux released by
ELMs can cause a significant enhancement in the erosion of solid surfaces that make up the
internal walls and divertor components of the tokamak. The impulsive loading of these
structures due to ELMs releases non-hydrogenic impurities as a result of enhanced solid
surface erosion. These impurities change the properties of the divertor plasma and can be
transported out of the divertor chamber into the region of the scrape-off layer (SOL) plasma
located between the separatrix in the main chamber walls of the tokamak. While ELMs tend
to prevent the eroded divertor impurities from penetrating deeply into the high temperature
region of the core plasma, located inside the steep pressure gradient region referred to as
pedestal plasma, these impurities can accumulate in steady-state discharges and affect the
plasma performance unless the tokamak pumping system is capable of removing this
additional particle flux.
ELMs are typically classified by the amount of energy they eject from the pedestal plasma
and their dynamical properties. The largest of these instabilities, referred to as type-I ELMs,
are capable of reducing the energy stored in the pedestal plasma by as much as 20-25% in
tokamaks operating at the highest performance levels (Loarte, et al. 2003). In the largest
A Conceptual Model for the Nonlinear Dynamics of Edge-localized Modes in Tokamak Plasmas

61
tokamaks operating at the present time, this amounts to the ejection of up to 1 MJ of energy
within a period of about 200-300 s. In the next generation of tokamaks that are under
construction or being planned, the energy ejected by a single ELM is expected to increase by
about a factor of 20. These type-I ELMs are also characterized by an increase in frequency
f
ELM
as the injected power level of the neutral beam heating system P
NBI
increases and they
do not tend to have any clearly identifiable coherent magnetic fluctuations prior to their
onset (i.e., magnetic precursors) although an increase in the level of broadband plasma
turbulence is sometimes observed prior to their onset. Profiles of the edge electron density
(n
e
) and temperature (T
e
) just before an ELM are shown for a typical DIII-D type-I ELMing
discharge in Fig. 1(a) while Fig. 1(b) shows how the n
e
profile changes immediately
following a type-I ELM.




Fig. 1. (a) An example of the steep n
e
and T
e
profiles, as a function of the normalized
poloidal magnetic flux Psi (
N
), across the outer region of the plasma inside the separatrix
and outside the separatrix in a region referred to as the SOL and (b) the n
e
profile before and
after an ELM in DIII-D discharge 126006.
As seen in Fig. 1(b), plasma density from the top of the pedestal region inward to
approximately 1/2 the radius of the core plasma, at a normalized poloidal magnetic flux Psi
(
N
) equal 0.5, is ejected into the region outside the separatrix, referred to as the SOL, during
the explosive growth period of the ELM instability in a typical high performance, low
collisionality, DIII-D type-I ELMing discharge. Type-I ELMs typically have frequencies f
ELM

Nonlinear Dynamics

62
= 10-200 Hz and are triggered when the pedestal pressure gradient (
ped
p
ped
)
approaches the critical ideal MHD ballooning limit
ped
23
crit_ball
(Osborne et al., 2000).
They are often characterized by an isolated, very rapid, increase in the deuterium recycling
emissions when the particles ejected from inside the separatrix arrive at the divertor target
plates or the walls of the tokamak as shown in Fig. 2.


Fig. 2. A series of type-I ELM impulses seen in the lower (primary) divertor deuterium (D

)
recycling emissions during DIII-D discharge 126006 where f
ELM
= 5075 Hz is correlated to
an increase in the stored energy of the plasma.
A significant fraction of the energy released from the pedestal during type-I ELMs strikes
the divertor target plates [see Fig. 5(b) for a view of the DIII-D lower divertor and divertor
target plates] along with the particle flux responsible for the spikes in the recycling
emissions (Fig. 2). It is this combined, highly impulsive, heat and particle flux that can cause
enhanced erosion of the divertor targets and walls in large, high performance, tokamaks
leading to a substantial increase of non-hydrogenic impurities released into the divertor and
SOL plasmas.
Also associated with these impulsive heat and particle fluxes are large, rapidly growing,
electric currents that are an intrinsic part of type-I ELM dynamics. These currents are, in
fact, a basic element of the nonlinear model described below. In DIII-D Langmuir probes are
used to measure the parallel ion saturation current flowing through the divertor target
plates at several radial positions. These measurements show that these currents grow
explosively to a saturated amplitude exceeding 1 A/mm
2
in 50 s or less during the
nonlinear growth of a type-I ELM. Measurements of the toroidal distribution and dynamics
of these currents with a toroidal tile current array in DIII-D has shown that they are strongly
non-axisymmetric with dominate toroidal mode numbers consisting primarily of n=1 and 2
components (Evans et al., 1995) while the presence of higher n modes has been observed
during ELM precursors in some DIII-D discharges (Osborne et al., 2000). As an example of
the dynamics involved in the evolution of this current, the data shown in Fig. 3
demonstrates the explosive growth of the instability followed by a slow decay during a
single ELM. This data is obtained with two lower divertor Langmuir probes located just
outside the 15 mm SOL flux surface with major radii R = 1.500 m and 1.528 m in a double
A Conceptual Model for the Nonlinear Dynamics of Edge-localized Modes in Tokamak Plasmas

63
null (upper and lower hyperbolic point) DIII-D discharge biased upward by 11 mm. The
radial structure of this current indicates that the ELM produces a strong interaction with the
top of a pump duct more than 120 mm away from the point of intersection of the separatrix
with the lower divertor target (Fig. 5 shows a layout of the lower divertor geometry). This
data also suggests that current flowing in a type-I ELM may be associated with a relatively
ridged structure that forms during its initial explosive growth phase. Finally, it suggests that
as the current in this structure decays it appears to rotate past the two Langmuir probes
with a rather regular period of t ~ 480 s (i.e., a toroidal rotation velocity v
ELM
= 2R/t ~
19.6 km/s where R is the radial position of one of the Langmuir probes) as indicated by a
series of fairly regularly spaced peaks shown in Fig. 3. Signatures such as these have also
been seen in the DIII-D midplane reciprocating Langmuir probe data where v
ELM
=
13.5 km/s was observed in plasmas with edge toroidal carbon rotation velocities v
carbon
=
22 km/s (Boedo et al., 2005; Yu et al., 2008).


Fig. 3. Time evolution of the plasma current measured by a pair of lower divertor Langmuir
probes located 28 mm apart in major radius (R) during an ELM DIII-D discharge 138229.
Type-II ELMs are sometimes observed as the axisymmetric plasma shape becomes more
triangular and elongated. They often appear as small, irregular, fluctuations in the D

data
interspersed between the large type-I ELM D

spikes. Type-II ELMs do not appear to have a

distinct P
NBI
dependence or any signatures associated with coherent MHD precursors and
do not seem to be associated with a specific
crit
limit (Zohm, 1996). Type-III ELMs are small,
relatively high f
ELM
, instabilities that tend to have lower frequencies as P
NBI
increases
(Osborne et al., 2000). They typically have coherent magnetic precursor modes with
frequencies in the 50 kHz range and have low to intermediate toroidal mode numbers
(n=5-10). They are often found in relatively high-density, lower P
NBI
, plasmas with
ped
ranging from about 30% to 50% of
crit
(Suttrop, 2000). Other types of small ELMs (e.g., type-
V) have been identified in spherical tokamaks where they appear only in lower single null
plasmas and are often interspersed between large type-I ELMs (Maingi et al., 2005).
The conceptual model proposed here deals exclusively with the nonlinear dynamics and
associated topological evolution of the explosive growth seen during the initial growth of
Nonlinear Dynamics

64
type-I ELMs. The dynamics of the different types of ELMs outlined above, as well as those
observed during intermittent transport bursts in low confinement modes (L-modes) are, at
the most fundamental level, required to conform to the general framework of this model i.e.,
the fact that a divergence free vector field, such as the equilibrium magnetic field in a
tokamak or stellarator, must ultimately be consistent with a (conservative) Hamiltonian
representation such as that prescribed by dynamical systems theory (Guckenheimer &
Holmes, 1983; Lichtenberg & Lieberman, 1992).
2.2 ELM topology
Before elaborating the details of the nonlinear type-I ELM model below, it is instructive to
briefly describe the global topology of these instabilities. Fortunately, spherical tokamaks
such as MAST (Kirk et al., 2004; Kirk et al., 2007) and NSTX (Maingi et al., 2005) are
equipped with visible light fast framing cameras that can capture images of type-I ELMs.
Figure 4 provides a full view of the plasmas captured during a type-I ELM in MAST.


Fig. 4. A wide-angle view of the MAST plasma at one instant in time during the evolution of
a type-I ELM (Courtesy of A. Kirk, Culham Laboratory, UK).
Here, the bright emission bands, referred to as ELM filaments, wrap around the outer
surface of the plasma in helical patterns that connect the upper and lower divertors. The
pitch of these filaments is aligned with the local magnetic field, which typically has a rather
steep angle with respect to the equatorial plane of the plasma due to the relative strength of
the poloidal field compared to that of the toroidal field in spherical tokamaks such as MAST.
Note that the intensity of the emission in these filaments is not uniform along their helical
axis and that these structures are seen to protrude from the surface as they approach the
upper hyperbolic point where they become much more toroidally aligned. These
protrusions are consistent with the type of structure predicted by the topology of homoclinic
and heteroclinic tangles invoked in the ELM model presented below. Here, the protrusions
correspond to the lobes of the tangle, which become narrower in the poloidal direction and
more extended in the radial direction as they approach a hyperbolic point (Fig. 5 shows the
A Conceptual Model for the Nonlinear Dynamics of Edge-localized Modes in Tokamak Plasmas

65
lobes calculated when an n=1 homoclinic tangle is found in the DIII-D tokamak). This
poloidal compression, accompanied by a radial expansion, is a consequence of the
preservation of a constant value of the magnetic flux contained inside each lobe of the
structure as prescribed by the Hamiltonian nature of the tangle in the model as it
approaches a region of weak poloidal magnetic field near the hyperbolic points. These
protruding lobes form a spiraling magnetic footprint that converges to the unperturbed
intersection of the separatrix with the divertor target plate similar to the one shown in Fig. 4
of Roeder et al., (2003) for an n=1 homoclinic tangle in DIII-D. These magnetic footprints are
essential elements of the nonlinear ELM model presented below.
2.3 ELM theories
Linear ELM theories tend to fall into three general categories. The first and most well
developed of these includes ideal and resistive ballooning MHD models. These involve
pressure driven modes that couple to external kink modes (sometimes referred to as peeling
modes). The second involves dynamics described by a bifurcation of the confinement from
an H-mode to an L-mode forming a dynamical state described by a restricted type of limit
cycle. The third combines elements taken from the MHD and limit cycle models to construct
an appropriate set of dynamics. Each of these models is reviewed in a paper by Conner
(1998). Nonlinear ELM models are relatively sparse due to the complex nature of the
dynamics and topology involved in this phase of the instability. One example invokes the
explosive growth of a narrow finger of hot plasma that pushes its way through other field
lines (nonlinear ballooning) from a small region in the plasma interior and spreads across a
large section of the surface of the plasma (Cowley et al., 2003). These models are difficult to
validate in any practical way with tokamak data due to a lack of specific predictions on how
they relate to the various types of ELMs and operating regimes found in high performance
tokamak discharges. Clearly, a more quantitative model is needed. Thus, there is strong
motivation to develop a model that can be more easily tested with experimental data. The
model presented below provides a step in this direction since it can be used to numerically
calculate the global topology of ELMs including the distribution and size of magnetic
footprints that can be directly compared to divertor diagnostic data (Wingen et al., 2009a).
3. Description of the proposed nonlinear ELM model
3.1 Hamiltonian description of the separatrix topology in poloidally diverted tokamaks
Poloidally diverted tokamaks are formed by a set of external axisymmetric coils that result
in poloidal magnetic field nulls when superimposed on the magnetic field due to a toroidal
plasma current flow in the discharge. These poloidal field nulls, in combination with
magnetic fields from other shaping coils in the tokamak, form hyperbolic points (Zaslavsky,
2005) of the system along with their associated separatrices that divide field line trajectories
into trapped (inside the separatrix) versus passing (outside the separatrix) regions of space
(Evans, 2008). In an ideal axisymmetric poloidally diverted tokamak the trapped and
passing field line regions are referred to as closed and open field line regions
respectively. This is because field lines outside the separatrix intersect the walls of the
tokamak and thus are open with respect to the loss of heat and particles that flow parallel
to the field lines. Alternatively, field lines inside the ideal axisymmetric separatrix do not
intersect the walls of the tokamak and thus are considered closed in terms of heat and
Nonlinear Dynamics

66
particle transport parallel to these field lines. As discussed below this terminology is no
longer applicable when small non-axsymmetric magnetic pertrubations are present in the
system.
A fundamental element of the ELM model discussed here is the nonlinear evolution of the
separatrix topology in a poloidally diverted tokamak. Here, the growth of a small
topological defect known as a homoclinic tangle (Guckenheimer & Holmes, 1983), formed
by the separatrix due to intersections of stable and unstable invariant manifolds associated
with a hyperbolic point of the system, is the basic dynamical process invoked by the model.
In a poloidally diverted tokamak, following along the spatial trajectory of a stable invariant
manifold in the forward direction results in a series of converging steps that approaches the
hyperbolic point associated with the manifold. Similarly, following the unstable invariant
manifold in the opposite (backward) direction produces a series of converging steps toward
the hyperbolic point from the opposite side. Thus, the splitting of trajectories into stable and
unstable manifolds due to non-axisymmetric perturbations introduces a directional
dependence into the spatial trajectories of the field lines implying that following field lines
in opposite directions leads to very different spatial locations in the plasma (with the
exception of homoclinic points where the stable and unstable invariant manifolds intersect).
In general, a homoclinic (self-intersecting) tangle results in a 3D separatrix topology that is a
generic property of perturbed hyperbolic conservative systems which are composed of
divergence-free vector fields. The dynamics of such a system is described by integrating
Hamiltons equations of motion. Theoretically, it is well known that when sufficiently small
perturbations are introduced into such a system it remains Hamiltonian in nature and
preserves its well-behaved (deterministic) dynamics (Dankowicz, 1997; Lichtenberg &
Lieberman, 1992). Such systems are commonly referred to as near integrable and
generically have non-degenerate, transversely self-intersecting, separatrix manifold
topologies that form the lobes of the homoclinic tangle. Separatrix structures such as these
have been studied extensively in physics, mathematics, astrophysics, engineering and
neuroscience (Dankowicz, 1997; Guckenheimer & Holmes, 1983; Simiu, 2002). Additionally,
it is well known from conservative dynamical systems theory that the topology of a
homoclinic tangle is the fundamental element that dictates the behavior of the trajectories
which form the solutions to the differential equations describing the dynamics of the system
(Guckenheimer & Holmes, 1983). In toroidal plasma confinement devices such as
stellarators and tokamaks, the 3D topology of the field lines at any instant in time is found
by integrating a set of magnetic differential equations that are formulated in terms of the
toroidal () and poloidal () magnetic flux coordinates (Dhaeseler et al., 1991). Here, is
associated with the Hamiltonian H while serves as the canonical momentum of the system
and the equations describing the 3D spatial trajectories of the field lines are given in
HamiltonJacobi form as:

d dH
d d

= (1)

d dH
d d

= (2)
where and are the poloidal and toroidal angles respectively (Evans, 2008). The usual
Hamiltonian is recognized in terms of the familiar canonical coordinates p,q by substituting
A Conceptual Model for the Nonlinear Dynamics of Edge-localized Modes in Tokamak Plasmas

67
p and q and associating with time (t). In a tokamak 2 is the toroidal magnetic
flux enclosed by a surface of constant and 2 = 2H is the poloidal magnetic flux inside a
surface of constant H.
Equations (1) and (2) are generally integrable given an axisymmetric plasma equilibrium but
the addition of small non-axisymmetric magnetic fields transforms the Hamiltonian into an
arbitrary function of the toroidal and poloidal angles. In this system a symmetry breaking,
non-axisymmetric, magnetic perturbation can be expressed in terms of a perturbed
Hamiltonian H
1
(,,) where is a small dimensionless perturbation parameter. Then, the
total Hamiltonian H is given as:
H = H
0
() + H
1
(,,) (3)
or the sum of the axisymmetric part H
0
and the non-axisymmetric perturbed part H
1
. The
perturbed part of the Hamiltonian can be expressed in terms of a Fourier series as:

( ) ( ) ( )
1 , ,
,
, , cos
m n m n
m n
H H m n = +

(4)
where m and n are the poloidal and toroidal mode numbers respectively (Abdullaev, 2006).
In realistic tokamaks, the nominally degenerate invariant manifolds that form an ideally
axisymmetric separatrix are transformed into an infinite set of homoclinic intersections by
small field-errors associated with non-axisymmetric toroidal and poloidal magnetic field
coil positions and other random magnetic pertrubations that are an intrinsic part of the
tokamak environment (Evans et al., 2005). In addition, externally applied low toroidal mode
number (n=1) non-axisymmetric magnetic fields are commonly used to correct ambient
field-errors that amplify MHD modes in the core plasma.
Figure 5 shows a poloidal projection of the 3D separatrix structure at one toroidal angle in
the DIII-D tokamak during the application of an n=1 field-error correction perturbation. As
seen in the lower part of Fig. 5(a) just above the divertor region the lobes of the homoclinic
tangle intersect the high-field side (HFS) wall (R = 1.02 m, Z = -1.17 m) while on the low-
field side (LFS) the lobes intersect the horizontal divertor target plate (R = 1.35 m, Z = -
1.36 m). Here, R,Z are cylindrical coordinates representing the distance from the toroidal
axis of the tokamak and the displacement from equatorial plane respectively. A magnified
view of the lower divertor region is shown in Fig. 5(b) with a 45 divertor tile (dashed line)
connecting the HFS wall (R = 1.02 m) to the horizontal target plate tile (Z = -1.37 m). The
entrance to the pump duct is shown on the right-hand side of Fig. 5(b) (R 1.36 m) with the
top of the pump duct located at Z = -1.25 m. The connection length L
c
of magnetic flux tubes
between the LFS divertor target plate and the HFS wall is shown by the color bars in each
part of the figure.
This discharge is an example of a double null plasma equilibrium with the balance between
the upper and lower hyperbolic points displaced slightly downward. Here, the upper
hyperbolic point is located at R = 1.27 m, Z = 1.11 m while the lower hyperbolic point is
located at R = 1.28 m, Z = -1.13 m. The topology of the hetroclinic tangles formed in double
null equilibria has been shown to be a sensitive function of the relative positions of the
upper (secondary) and lower (primary) hyperbolic points (Evans et al., 2004). For a
downward biased equilibrium such as that shown in Fig. 5, the homoclinic tangle associated
with the lower hyperbolic point dominates the 3D topology of the separatrix and creates a
dramatic change in the magnetic topology inside the separatrix. Here, one of the lobes of the
Nonlinear Dynamics

68
tangle intersects the horizontal divertor target plate at R = 1.36 m, Z = -1.36 m while another
lobe intersects the vertical wall located at R = 1.02 m. The intersection of these homoclinic
lobes with the divertor target plate and wall opens some of the field lines that were
previously in the closed region inside the separatrix prior to the application of the n=1
magnetic perturbation field from the field-error correction coil (although some field lines are
always open due to intrinsic non-axisymmetric field errors without the correction coils).
This topological change creates a set of highly complex field line trajectories that traverse the
plasma volume inside the separatrix and connect the vertical high-field side (HFS) wall to
the low-field side (LFS) horizontal divertor target plate. This topology is similar to that of
line-tying found in the solar photosphere (Gibons & Spicer, 1981). Additionally, the field
line topology formed by this line-tying type of bifurcation is composed of a mixture of
stochastic fields, with a wide range of connection lengths (L
c
) that form fractal distributions
(Abdullaev, 2006), embedded inside a set of coherent flux tubes with short connection
lengths (Wingen et al., 2009b; Wingen et al., 2009c). It is the short L
c
flux tubes that play a
fundamental role in the nonlinear ELM model discussed below.


Fig. 5. (a) Full poloidal cross sectional view of a separatrix homoclinic tangle formed by an
applied external n=1 magnetic perturbation due to the DIII-D field-error correction coil with
a current of 8 kAt in discharge 133908 at t = 2000 ms. (b) An expanded cross sectional view
of the primary divertor. L
c
is the field line connection length between the horizontal target
plate (Z = -1.36 m) and the vertical wall (R = 1.02 m).
The intersection of the homoclinic and heteroclinic lobes with divertor targets and walls
forms objects referred to as magnetic footprints on the R, and Z, planes of the divertor
targets and walls respectively as shown in Fig. 6(a) for the HFS wall and Fig. 6(b) for the LFS
divertor target plate for the same conditions as in Fig. 5. Measurements of the heat (Evans
et al., 2005, Evans et al., 2007) and particle (Schmitz et al., 2008) flux distributions on the
A Conceptual Model for the Nonlinear Dynamics of Edge-localized Modes in Tokamak Plasmas

69
DIII-D divertors have been shown to be qualitatively consistent with numerical calculations
of magnetic footprints produced by homoclinic lobes during experiments with applied non-
axisymmetric magnetic fields from field-error correction and edge MHD (ELM) suppression
coils (Evans et al., 2006). Similar heat flux patterns have been observed in the ASDEX-U
tokamak (Eich et al., 2005) during ELMs. Quantitative comparisons of heat flux measure-
ments inside these footprints with numerical simulations indicate that the separation
between adjacent lobe intersections with the divertor targets can be a factor of 2-3 times
larger than that predicted suggesting that there is a significant amplification of the homo-
clinic tangle structure from the applied n=1 field due to the response of the plasma (Evans
et al., 2007). There are also indications that the topology of the lobes is affected by magnetic
perturbations from MHD modes deep in the core plasma (Evans et al., 2005).


Fig. 6. Lower divertor (a) magnetic footprint formed on HFS vertical wall by an externally
applied n=1 perturbation (no plasma response) from the DIII-D field-error correction coil
with a current of 8 kAt in discharge 133908 at t = 2000 ms and (b) the LFS magnetic footprint
formed on the horizontal divertor target plate. These footprints define the open field line hit
points due to the intersection of the lobes of the homoclinic tangle shown in Fig. 5 with the
target plate and wall. As in Fig. 5, L
c
is the field line connection length between the
horizontal target plate (Z = -1.36 m) and the vertical wall (R = 1.02 m).
3.2 Description of the temporal evolution prescribed by the model
A conceptual model describing the dynamics of the edge plasma and the evolution of the
pedestal magnetic topology following the linear growth phase of a type-I ELM is presented.
Understanding the physics, topology and dynamics of ELMs during their post-linear
growth phase is essential for predicting the characteristics of these instabilities as a funtion
of the pedestal plasma conditions. In particular, a model is needed that can be used to
predict the temporal evolution of the plasma heat and particle distributions on the vessel
wall and divertor components.
As discussed above, small quasi-static homoclinic and heteroclinic tangles result naturally
from a variety of non-axisymmetric magnetic field perturbations commonly found in high
Nonlinear Dynamics

70
performance poloidally diverted tokamaks. Examples of these non-axisymmetric field
perturbations include toroidal field ripple, field-errors, core and edge MHD modes, 3D
electromagnetic field control (trim) coils and small, spatially random, 3D field components
due to magnetic materials and tolerence build-ups in the electromagnetic coils used to
confine and shape the plasma (Evans et al., 2005; Evans et al., 2007). Thus, it is not
unreasonable to expect the formation of separatrix homoclinic and heteroclinic tangles to be
the norm rather than the exception, whether in a low confinement L-mode or in high
confinement H-mode plasma as well as between and during ELMs. It is the existence of the
separatrix topology associated with these tangles between ELMs that forms the basis of the
model (Evans et al., 2009) described here.
Given the basic topology shown in Fig. 5, the model assumes that small fluctuations in the
pedestal plasma pressure initiate a linearly growing MHD instability as the equilibrium
conditions in a narrow region just inside the separatrix approach a marginal stability point.
An example of this process is described by ideal MHD peeling-ballooning theory (Snyder
et al., 2005; Wilson et al., 2006) which presumes that linearly growing intermediate n modes
lead to the onset of the nonlinear growth phase. Peeling-ballooning theory predicts that the
onset of this edge MHD mode significantly increases the radial heat and particle transport.
Our model assumes that the energy associated with the linearly growing MHD mode flows
into the coherent, short connection length, homoclinic flux tubes connecting the HFS wall
and the LFS divertor target. At this point, fast parallel transport along these homoclinic flux
tubes causes a rapid increase in the electron temperature (T
e
) inside the magnetic footprints
near the wall and divertor surfaces. Experimental measurements taken during the early
growth of an ELM demonstrate that there is a rapid release of thermal energy from the area
located near the steep gradient region leading up to the top of the pedestal just inside the
separatrix (Kirk et al., 2007, Neuhauser et al., 2008). These observations are consistent with
our requirement of a rapid increase in the radial energy transport during this time. These
rapid bursts of energy flowing from the pedestal into the divertor appear to be correlated
with an increase in broadband magnetic fluctuations in the pedestal starting about 10 s
before the onset of the nonlinear growth phase (Neuhauser et al., 2008) suggesting that
currents in this region may play a key role in the onset of the nonlinear growth phase.
In our model, it is these inital heat pulses associated with the linearly growing MHD
instability that provide the mechanism needed to form a feedback amplification loop. It is
this feedback loop that causes the stable and unstable invariant manifolds of the initial
homoclinic tangle to grow explosively. Here, it is presumed that the amplification process is
triggered by the formation of field-aligned thermoelectric currents that flow through the
short, pedestal plasma, homoclinic flux tubes connecting the inner wall and outer divertor
target plate. These thermoelectric currents form when T
e
at one end of a flux tube increases
relative to T
e
at the other end (Staebler & Hinton, 1989). Since part of the heat pulse enters
the short flux tube near the equatorial plane on the LFS of the discharge it is expected to
arrive at the LFS target well before arriving at the HFS wall. Numerical simulations of these
two-poloidal-turn helical flux tubes (Wingen, et al. 2009a) show that the distance from the
LFS equatorial plane to the LFS target plate is ~25 m while the distance to the HFS wall is
~75 m. In DIII-D H-mode plasmas, T
e
just inside the separatrix, where these flux tubes
reside, is ~400-500 eV. Thus, given an electron thermal velocity v
Te
= (kT
e
/m
e
)
1/2
=
8.410
6
m/s where k is Boltzmanns constant and m
e
is the mass of an electron, these heat
pulses arrive at the LFS target plate approximately 6 s before reaching the HFS wall. This
A Conceptual Model for the Nonlinear Dynamics of Edge-localized Modes in Tokamak Plasmas

71
causes T
e
near the LFS target plate to increase relative to that near the HFS wall and initiates
the flow of a thermoelectric current from the LFS target to the HFS wall with a return
current connecting through the lower divertor vessel structure.
Thus, the model assumes that following the release of the initial heat pulse from the linearly
growing MHD mode a small field-aligned thermoelectric current begins to flow in a helical
flux tube formed by a small preexisting homoclinic tangle. Although the time evolution of
the current growth is not specifically predicted by the model, it is assumed that as the
current grows its magnetic field perturbs the upper and lower hyperbolic points causing the
lobes of the homoclinic tangle and their associated magnetic footprints to increase in size.
Simulations have been carried out assuming the current density in the flux tube is limited to
approximately 1/2 the initial ion saturation current density (~70-80 mA/mm
2
) during the
nonlinear phase of the instability (Wingen, et al., 2009a). These simulations have shown that
the magnetic footprint, associated with a single n=1 flux tube connecting the primary
(lower) LFS divertor target with the HFS wall, grows from an area of 1760 mm
2
, with a
current of 135.5 A, to an area of 3564 mm
2
with a current of 274.4 A. During this process, a
topological bifurcation takes place that creates a new set of n=2 flux tubes connecting the
primary divertor LFS target to the HFS wall (Wingen, et al., 2009a). It is then assumed, that
as the thermoelectric current grows with increasing footprint area there is a commensurately
increasing flow of energy from the pedestal into the flux tube that maintains the constant
current density. Here, the working hypothesis is that the growing helical thermoelectric
current filaments associated with the short connection length flux tubes also produce
resonant magnetic field components that open magnetic islands (i.e., Poincar islands) on
rational surfaces across the pedestal region in addtion to perturbing the nominally
axisymmetric hyperbolic points. As these islands grow and overlap they produce an
increase in the local magnetic field line stochasticity which enhances the effective radial heat
tranport into the homoclinic flux tube containing the thermoelectric current. This completes
the feedback amplification loop and results in the initial explosive growth phase of the
topological instability.
During the next step in the process, the initial helical current filament grows explosively and
acts to amplify the lobes of the homoclinic tangle while inducing a growing level of pedestal
stochasticity that penetrates deeper into the core plasma as it grows. This process results in a
self-amplification of the lobes due to a positive feedback loop between the size of the tangle
lobes, an increasing stochastic layer width and an increase in the heat flux to the target
plates that drives an increasing flow of current. This process takes on the appearance of
growing helical filaments that protrude beyond the edge of the plasma and seem to
propagate radially outward as they grow. A key feature of the processes involved up to this
point is that there is no need to invoke field line tearing and reconnection during the
evolution of an ELM. The entire process can be described using ideal MHD theory without
requiring resistive or dissipative effects that would cause the filaments to tear and separate
from the edge of the plasma. Such a process would rapidly shut down the thermal transport
responsible for the growth of the instability. This is seen by comparing the 1-2 ms decay
time following the current peak in Fig. 3 to a tearing mode growth time
-1
~
r
3/5

A
2/5
s
where is the growth rate of the tearing mode,
r
is the resistive time and
A
is the Alfvn
time in the pedestal. We find that
-1
0.1 ms where
r
= 1.2x10
-4
s and
A
= 5.8x10
-5
s or
approximately an order of magnitude shorter than the current decay time. Therefore, the
Nonlinear Dynamics

72
growth of a tearing mode following the peak in the current would cause a separation of the
filament from the pedestal and a rapid, sub-millisecond, termination of the of the current.
Instead, we see that the ELM is a radially extended structure, as indicated by the relatively
constant ratio of the signals from two adjacent Langmuir probes in Fig. 3, which is
reminiscent of the lobes of a homoclinic tangle and that this radial structure persists as the
current slowly decays. This relatively slow decay of the current is more consistent with a
slow shutdown of the heat flux from the pedestal as the energy reservoir in this region is
slowly depleted and T
e
in the short flux tubes drops. This reduction in T
e
causes an increase
in resistivity in the short flux tubes which, when coupled with a cooling of the plasma in
front of the divertor target plate due to an increase in particle recycling, as shown in Fig. 2,
slowly reduces the thermoelectric current flowing between the target plate and the wall.
Numerical simulations of the growth experienced by a pre-existing, field-error related,
homoclinic tangle have been carried out using current filaments that are proportional to the
area of the magnetic footprint on the divertor target plate. Results from these simulations
demonstrate that the calculated nonlinear dynamics of the tangles topology are consistent
with the heat flux patterns measured in the DIII-D divertor during a type-I ELM (Wingen
et al., 2009a). A key question studied during these simulations addresses how the
topological evolution prescribed by the model conforms to experimental measurements of
type-I ELM dynamics. In particular, data such as that shown in Fig. 4 suggest that the peak
in the toroidal mode spectrum of an ELM increases in mode number during the nonlinear
growth phase. As discussed below, a bifurcation in the separatrix topology has been
identified during the early growth phase of the instability. This bifurcation involves the
appearance of heteroclinic invariant manifolds associated with the upper (secondary)
hyberbolic point.
3.3 Dynamics of an ELM-induced homoclinic-to-heteroclinic separatrix bifurcation
Here, we describe the appearance of a homoclinic-to-heteroclinic bifurcation as the total
current flowing in a short flux tube, connecting the LFS divertor target plate to the HFS wall,
increases from 100 to 300 A. The simulation starts with an axisymmetric plasma equilibrium.
We then superimpose a spectrum of nonaxisymmetric magnetic perturbations due to field-
errors that have been systematically measured in the DIII-D tokamak (Luxon et al., 2003)
along with a 3D magnetic perturbation field produced by a field-error correction coil (refer-
red to as the I-coil) in DIII-D discharge 133908 at t = 2000 ms (Wingen et al., 2009a). Note
that this is the same plasma equilibrium shown in Fig. 5 but there an artificial n=1 nonaxi-
symmetric magnetic field is applied by a coil referred to as the C-coil with a relatively large
current in order to highlight the properties of the homoclinic tangle. In the simulation
discussed here, we use the actual coil currents that were employed during the experiment in
discharge 133908.
As a starting point for this simulation, the shortest flux tube produced by the field-errors
and an n=1 correction coil is selected. Initially, there is only one relatively small flux tube
connecting the LFS side lower (primary) divertor target plate with the HFS wall. We refer to
this as flux tube number 1. This flux tube makes two poloidal revolutions along its path
through the pedestal plasma just inside the separatrix and has a total length from the target
plate to the wall of ~100 m. The current flowing in a large divertor tile sensor is used to
establish a current density calibration for the simulation. This is done by calculating the area
A Conceptual Model for the Nonlinear Dynamics of Edge-localized Modes in Tokamak Plasmas

73
of overlap between the tile sensor and the magnetic footprint at the toroidal and radial
position of the tile when the maximum current during an ELM is reached. Using the calcu-
lated area of intersection with the tile sensor and an assumed current density of
77 mA/mm
2
we get 200 A which agrees with the measured current in this tile sensor at the
peak amplitude of the ELM. The assumed current density (about 1/2 the pre-ELM ion satu-
ration current) is held fixed throughout the remainder of simulation while the topology of
the separatrix unfolds. We start with a relatively small current in flux tube number 1 and
increase the current in steps. With each iteration of the code, the area of the magnetic
footprints increases as the size of the lobes produced by the homoclinic tangle associated
with the primary divertor hyperbolic point increases. The current is increased until the total
area of all the magnetic footprints overlapping the tile sensor equals ~3000 mm
2
. At this
point, the area of the three footprints associated with flux tubes 1, 2 and 3 is calculated and
using the assumed current density of 77 mA/mm
2
a total current of ~4.9 kA is obtained
(Wingen et al., 2009a).
During the sequence of iterations in the current flowing in flux tube number 1, a new pair of
flux tubes is formed followed by the formation of a fourth flux tube at a higher current. The
first pair of flux tubes, referred to as flux tube number 2 and 3, connect the primary LFS
divertor target plate to the HFS wall after one poloidal turn and have a length of ~50 m. Flux
tubes 2 and 3 are formed during a bifurcation of the separatrix topology that involves a
splitting of the invariant manifolds, caused by the presence of the secondary (upper) hyper-
bolic point, into a higher order set of stable and unstable branches of the original manifold
topology. We refer to this as a homoclinic-to-heteroclinic bifurcation although here we focus
only on the increased complexity of the homoclinic tangle associated with the primary
(lower) hyperbolic point.
Figure 7(b) shows the structure of the manifolds produced by the primary (lower divertor)
hyperbolic point in the secondary divertor region near the upper hyperbolic point with a
current of 100 A flowing in flux tube number 1. Flux tube number 1 is not large enough to
be clearly identified at this level of current. With this current, the initial formation of flux
tubes 2 and 3 has begun. Here, flux tubes are formed in the area between intersecting stable
and unstable manifolds. As seen in Fig. 8(a) flux tube number 3 is completed at 130 A when
the stable and unstable manifolds intersect while flux tube number 2 is not yet fully formed
at 150 A in Fig. 8(b).
As the current in flux tube number 1 is increased from 150 A to 200 A, flux tube number 2 is
completed and a new partially formed flux tube appears, flux tube number 4 as shown in
Fig. 9(a), on each side of flux tube number 2. Between 200 A and 300 A flux tube number 4 is
completed and manifold connections are made between the secondary (upper) divertor LFS
target plate and the primary (lower) HFS wall as well as between the primary LFS target
plate and the secondary HFS wall as shown in Fig. 9(b).
From this point on in the simulation a current proportional to the area of intersection of flux
tubes 2 and 3 with the primary LFS divertor target, having a current density of 77 mA/mm
2
,
is included at each subsequent step until the current limit discussed above is reached. As the
simulation proceeds flux tubes 2 and 3, which form a pair of single poloidal turn helical
structures that are displaced from each other toroidally by 180, produce an n=2
perturbation that dominates the growth of the lobes and the primary divertor LFS target
plate magnetic footprints.
Nonlinear Dynamics

74

Fig. 7. Poincar plots of (a) the calculated structure of the stable and unstable invariant
manifolds in the primary divertor with a current of 100 A in flux tube number 1 (not clearly
visible) and (b) the corresponding structure of the manifolds in the secondary divertor. The
numbers 2 and 3 indicate regions where flux tube number 2 and 3 will form as the current in
flux tube number 1 is increased in the simulation once the stable and unstable manifolds
intersect.

Fig. 8. Poincar plots of (a) the formation of flux tube number 3 in the secondary divertor as
the current in flux tube 1 is increased to 130 A, (b) flux tube number 2 is not completely
formed at 150 A.
4. Discussion and conclusion
A conceptual model describing the nonlinear gowth of type-I ELMs in high performance
tokamak plasmas has been presented along with a numerical simulation of the separatrix
evolution, described by the model, during an ELM in a typical DIII-D H-mode plasma. The
A Conceptual Model for the Nonlinear Dynamics of Edge-localized Modes in Tokamak Plasmas

75

Fig. 9. Poincar plots of (a) the formation of flux tube number 2 in the secondary divertor at
200 A in flux tube number 1 and the appearance of a new partially formed flux tube
(number 4) while (b) at 300 A in flux tube number 1 all of the new flux tubes (numbers 2, 3
and 4) are fully formed.
temporal evolution of the separatrix is driven by a plasma instability resulting from a
rapidly growing current that flows through the pedestal region of the plasma and changes
the global topology of the manifolds that make up the separatrix. This topological change
involves a homoclinic-to-heteroclinic bifurcation of the secondary (upper) hyperbolic point
in the equilibrium magnetic field. The bifurcation creates an n=2 helical structure, consisting
of two independent flux tubes separated by 180 toroidally, early in nonlinear growth phase
when a small, 150-200 A, field-aligned current flows in the original n=1 flux tube created by
field-errors and a field-error correction coil. Although reversing the direction of the current
in the n=1 flux tube does not have a significant effect on the structure of the invariant
manifolds associated with the tangle structures, distributing the current into multiple
filaments rather than allowing it in the single filament, as in the simulation shown in
Sec. 3.3, results in a much more complex topology that has significantly more lobes
intersecting the primary LFS divertor target plate. Thus, the model predicts the formation of
a new set of invariant manifolds associated with the secondary hyperbolic point. These new
invariant manifolds intersect the upper (secondary) divertor target plate and the HFS wall in
DIII-D during the nonlinear growth phase of an ELM. This has significant implications for
fusion reactor designs since it implies that complex heat and particle flux striations,
associated with the magnetic footprints and flux tubes due to these new separatrix
manifolds, should cause large impulsive energy bursts on secondary plasma facing surfaces
that are not typically designed to handle such interactions with high energy density
plasmas. Therfore, it is important that these predictions be tested using high time resolution
measurements of the transient heat and particle flux interactions with plasma facing
components near the secondary hyperbolic point during ELMs.
Another important question to ask of the model is whether it can be used to shed light on the
physics of ELM suppression when small (~50 G), stationary, n=3 magnetic pertrubations are
applied to ELMing H-modes in the DIII-D tokamak (Evans et al, 2006). Here, an interesting
Nonlinear Dynamics

76
hypothessis that can be tested is that the n=3 field interacts with the lobes of the n=1 field-error
tangle causing them to break up into relatively small scale structures that are more effective for
dissapating the steady-state heat and particle flux over a much larger area of the divertor. This
is expected to provide additional control over the pedestal transport that could be used to keep
the pressure gradient below the threshold required for the onset of the linearly growing MHD
installability. Constructive and destructive interference between nonaxisymmetric magnetic
pertrubations from various coils in DIII-D has been studied previously. This work
demonstrated that such interactions lead to much more complex lobe structures (Wingen et al.,
2009b) that tends to spread the footprints and open more flux tubes (Evans et al, 2007) which
can be used to fine tune the pedestal transport. Extending this hypothesis to higher n
homoclinic structures, such as n=4 up to n=6 or 7 or combinations of structures with toroidal
mode numbers ranging from 1 thorugh 7, suggests that the effect may provide much better
control over the height and width of the pedestal region thus allowing the possibility of fine
tuning of the pressure gradient profile. With advanced realtime pedestal profile diagnostics, it
should be possible to combine this multimode perturbation field approach with an edge
pressure and current gradient tracking algorithm to obtain a desired set of pedestal properties,
particularly if an edge-localized heating and current drive system such as electron cyclotron
system, were to be included as part of the feedback loop.
In general, the model presented here qualitatively fits some of the observed experimental
attributes of large type-I ELMs such as a slow decay rate of the current in the flux tube as
seen in Fig. 3. The model also has elements that may explain the variability seen in ELM
signatures such as the divertor recycling emissions when effects such as type-II ELMs
between the type-I ELMs are included. Other effects, such as an increase in the frequency of
the ELMs with increasing heating power and the apprent rotation of the ELM structure
during the nonlinear growth phase, have not yet been addressed by the model. These will be
the focus of future work along with more detailed experimental comparisons.
5. Acknowledgments
This work was supported by the US Department of Energy under DE-FC02-04ER54698, DE-
AC04-94AL85000, and DE-AC52-07NA27344.
6. References
Abdullaev, S.S. (2006). Construction of Mappings for Hamiltonian Systems and Their Applications,
Lecture Notes in Physics, Vol. 691, Springer, ISBN-10 3-540-30915-2, Berlin.
Boedo, J.A.; Rudakov, D.L.; Hollmann, E.M.; Gray, D.S.; Burrell, K.H.; et al. (2005). Edge-
localized mode dynamics and transport in the scrape-off layer of the DIII-D
tokamak. Physics of Plasmas 12, 072516:1-11.
Callen, J.D.; Carreras, B.A.; & Stambaugh, R.D. (1992) Stability and transport processes in
tokamak plasmas. Physics Today 45, 34-42.
Connor, J.W. (1998). A review of models for ELMs. Plasma Physics and Controlled Fusion 40,
191-213.
Cowley, S.C.; Wilson, H.; Hurricane, O & Fong, B. (2003). Explosive instabilities: from solar flares
to edge localized modes in tokamaks. Plasma Physics and Controlled Fusion 45, A31-A38.
Dankowicz, H. (1997). Chaotic Dynamics in Hamiltonian Systems with applications to celestial
mechanics, World Scientific Series on Nonlinear Science, Series A, Vol. 25, World
Scientific, ISBN 9810232217, Singapore.
A Conceptual Model for the Nonlinear Dynamics of Edge-localized Modes in Tokamak Plasmas

77
Dhaeseleer, W.D.; Hitchon, W.N.G.; Callen, J.D. & Shohet, J.L. (1991). Flux coordinates and
magnetic field structure, a guide to a fundamental tool for plasma theroy, Springer Series
in Computational Physics, Springer-Verlag ISBN 3-540-52419-3, Berlin.
Eich, T.; Herrmann, A.; Neuhauser, J.; Dux, R.; Fuchs, J.C.; et al. (2005). Type-I ELM structure
on divertor target plates in ASDEX Upgrade. Plasma Physics and Controlled Fusion
47, 815-842.
Evans, T.E.; Lasnier, C.J.; Hill, D.N.; Leonard, A.W.; Fenstermacher, M.E.; et al. (1995).
Measurements of non-axisymmetric effects in the DIII-D divertor. Journal of Nuclear
Materials 220-222, 235-239.
Evans, T.E.; Moyer, R.A.; Stephan, E.A.; Snider, R.T. & Coles, W.A. (1996). Causal spacio-
temporal correlations of short scale length solar wind accelaration and heating
mechanisma with a solar event correlation analyzer (SECA) instrument package.
Robotic exploration close to the sun: scientific basis, AIP Conference Proceedings 385,
pp. 145-152, Editor: S. R. Habbal, American Institute of Physics ISBN 1-56396-618-2,
New York.
Evans, T.E.; Roeder, R.K.W.; Carter, J.A. & Rapoport, B.I. (2004). Homoclinic tangles,
bifurcations and stochasticity in poloidally diverted tokamaks. Contributions to
Plasma Physics 44, 235-240.
Evans, T.E.; Roeder, R.K.W.; Carter, J.A.; Rapoport, B.I.; Fenstermacher, M.E.; & Lasnier, C.J.
(2005). Experimental signatures of homoclinic tangles in poloidally diverted
tokamaks. Journal of Physics: Confonference Proceedings Series 7, 174-190.
Evans, T.E.; Moyer, R.A.; Burrell, K.H.; Fenstermacher, M.E.; Joseph, I.; et al. (2006). Edge
stability and transport control with resonant magnetic pertrubations in collisionless
tokamak plasmas. Nature Physics 2, 419-423.
Evans, T.E.; Joseph, I.; Moyer, R.A.; Fenstermacher, M.E.; Lasnier, C.J.; Yan, L.W. (2007).
Experimental and numerical studies of seperatrix splitting and magnetic footfrints
in DIII-D. Journal of Nuclear Materials 363-365, 570-574.
Evans, T.E. (2008). Implications of topological complexity and Hamiltonian chaos in the
edge magnetic field of toroidal fusion plasmas. Chaos, Complexity and Transport:
Theory and Applications, pp. 147-176 Edited by: Chandre C.; Leoncini, X. &
Zaslavsky, G. World Scientific Press, ISBN-13 978-981-281-897-9, Singapore.
Evans, T.E.; Yu, J.H.; Jakubowski, M.W.; Schmitz, O.; Watkins, J.G. & Moyer, R.A. (2009). A
coneptual model of the magnetic topology and nonlinear dynamics of ELMs.
Journal of Nuclear Materials 390-391, 789-792.
Fenstermacher, M.E.; Leonard, A.W.; Snyder, P.B.; Boedo, J.A.; Brooks, N.H.; et al., (2003).
ELM particle and energy transport in the SOL and divertor of DIII-D. Plasma
Physics and Controlled Fusion 45, 1597-1626.
Gibons, M. & Spicer, D.S. (1981). On line tying. Solar Physics 69, 57-61.
Guckenheimer, J. & Holmes, P. (1983). Nonlinear Oscillations, Dynamical Systems, and
Bifurcations of Vector Fields, Applied Mathematical Science, Vol. 42 Springer-Verlag,
ISBN 0-387-90819-6, New York.
Kirk, A.; Wilson, H.R.; Counsell, G.F.; Akers, R.; Arends, E.; et al. (2004). Spatial and
temporal structure of edge-localized modes. Physical Review Letters 92, 245002:1-4.
Kirk, A.; Counsell, G.F.; Cunningham, G.; Dowling, J.; Dunstan, M.; et al. (2007). Evolution
of the pedestal on MAST and the implications fo ELM power loadings. Plasma
Physics and Controlled Fusion 49, 1259-1275.
Lichtenberg, A.J. & Lieberman, M.A. (1992). Regular and Chaotic Dynamics, Applied
Mathematical Sciences, Vol. 38 second edition Springer-Verlag, ISBN 0-387-97745-7,
New York.
Nonlinear Dynamics

78
Loarte, A.; Saibene, G.; Sartori, R.; Becoult, M.; Horton, L.; et al. (2003). ELM energy and
particle losses and their extrapolation to burning plasma experiments. Journal of
Nuclear Materials 313-316, 962-966.
Luxon, J.L. (2002). A design retrospective of the DIII-D tokamak. Nuclear Fusion 42, 614-633.
Luxon, J.L.; Schaffer, M.J.; Jackson, G.L.; Leuer, J.A.; Nagy, A.; et al. (2003). Anomalies in the
applied magnetic fields in DIII-D and their implications for the understanding of
stability experiments. Nuclear Fusion 43, 1813-1828.
Maingi, R.; Bush, C.E.; Fredrickson, E.D.; Gates, D.A.; Kaye, S.M.; (2005). H-mode pedestal,
ELM and power threshold studies in NSTX. Nuclear Fusion 45, 1066-1077.
Neuhauser, J.; Bobkov, V.; Conway, G.D.; et al. (2008). Structure and dynamics of
spontaneous and induced ELMs on ASDEX Upgrade. Nuclear Fusion 48, 045005:1-15.
Osborne, T. H.; Ferron, J.R.; Groebner, R.J.; Lao, L.L.; Leonard, A.W.; et al. (2000). The effect
of plasma shape on H-mode pedestal characteristics on DIII-D. Plasma Physics and
Controlled Fusion 42, A175-A184.
Roeder, R.K.W.; Rapoport, B.I. & Evans, T.E.; (2003). Elplicit calcualtions of homoclinic
tangles in tokamaks. Physics of Plasmas 10, 3796-3799.
Simiu, E. (2002). Chaotic Transitions in Deterministic and Stochastic Dynamical Systems,
Princeton Series in Applied Mathematics, Princeton University Press, ISBN 0-691-
05094-5, Princeton, New Jersey.
Schmitz, O.; Evans, T.E.; Fenstermacher, M.E.; Frerichs, H.; Jakubowski, M.W.; et al. (2008).
Aspects of three dimensional transport for ELM control experiments in ITER-
similar shape plasmas at low collisionality in DIII-D. Plasma Physics and Controlled
Fusion 50, 124029:1-19.
Snyder, P.B.; Wilson, H.R. & Xu, X.Q. (2005). Progress in the peeling-ballooning model of
edge localized modes: Numerical studies of the nonlinear dynamics. Physics of
Plasmas 12, 056115:1-7.
Staebler , G.M. & Hinton, F.L. (1989). Currents in the scrape-off layer of diverted tokamaks.
Nuclear Fusion 29, 1820-1824.
Suttrop, W. (2000). The physics of large and small edge localized modes. Plasma Physics and
Controlled Fusion 42 pp. A1-A14.
Wesson, J. (2004). Tokamaks, 3rd Ed. Oxford University Press, ISBN 0-19-8509227, New York.
Wilson, H.R.; Cowley, S.C.; Kirk, A.; & Snyder, P.B. (2006). Magnetohydrodynamic stability
of the H-mode transport barrier as a model for edge localized modes: an overview.
Plasma Physics and Controlled Fusion 48, A71-A84.
Wingen, A.; Evans, T.E.; Lasnier, C.J. & Spatschek, K.H. (2009a). Numerical modelling of the
nonlinear ELM cycle in tokamaks. Physical Review Letters Vol. -- pp. (submitted).
Wingen, A.; Evans, T.E. & Spatschek, K.H. (2009b). High resolution numerical studies of the
separatrix splitting due to non-axisymmetric perturbation in DIII-D. Nuclear Fusion
49, 055027:1-8.
Wingen, A.; Evans, T.E. & Spatschek, K.H. (2009c). Footprint structures due to resonant
magnetic perturbations in DIII-D. Physics of Plasmas 16, 042504:1-5.
Yu, J.H.; Boedo, J.A.; Hollmann, E.M.; Moyer, R.A. & Rudakov, D.L. (2008). Fast imaging of
edge localized mode structure and dynamics in DIII-D. Physics of Plasmas 15,
032504:1-7.
Zaslavsky, G.M. (2005). Hamiltonian chaos & fractional dynamics, Oxford University Press
ISBN 0-19-852604-0, Oxford.
Zohm, H. (1996). Edge localized modes (ELMs). Plasma Physics and Controlled Fusion 38,
105-128.
4
Nonlinear Dynamics of Cantilever
Tip-Sample Surface Interactions
in Atomic Force Microscopy
John H. Cantrell
1
and Sean A. Cantrell
2

1
NASA Langley Research Center

2
Johns Hopkins University
USA
1. Introduction
The atomic force microscope (AFM) (Bennig et al., 1986) has become an important nanoscale
characterization tool for the development of novel materials and devices. The rapid
development of new materials produced by the embedding of nanostructural constituents
into matrix materials has placed increasing demands on the development of new nanoscale
measurement methods and techniques to assess the microstructure-physical property
relationships of such materials. Dynamic implementations of the AFM (known variously as
acoustic-atomic force microscopies or A-AFM and scanning probe acoustic microscopies or
SPAM) utilize the interaction force between the cantilever tip and the sample surface to
extract information about sample material properties. Such properties include sample elastic
moduli, adhesion, surface viscoelasticity, embedded particle distributions, and topography.
The most commonly used A-AFM modalities include various implementations of amplitude
modulation-atomic force microscopy (AM-AFM) (including intermittent contact mode or
tapping mode) (Zhong et al., 1993), force modulation microscopy (FMM) (Maivald et al.,
1991), atomic force acoustic microscopy (AFAM) (Rabe & Arnold, 1994; Rabe et al., 2002),
ultrasonic force microscopy (UFM) (Kolosov & Yamanaka, 1993; Yamanaka et al., 1994),
heterodyne force microscopy (HFM) (Cuberes et al., 2000; Shekhawat & Dravid, 2005),
resonant difference-frequency atomic force ultrasonic microscopy (RDF-AFUM) (Cantrell et
al., 2007) and variations of these techniques

(Muthuswami & Geer, 2004; Hurley et al., 2003;
Geer et al., 2002; Kolosov et al., 1998; Yaralioglu et al., 2000; Zheng et al., 1006; Kopycinska-
Mller et al., 2006; Cuberes, 2009).
Central to all A-AFM modalities is the AFM. As illustrated in Fig. 1, the basic AFM consists of
a scan head, an AFM controller, and an image processor. The scan head consists of a cantilever
with a sharp tip, a piezoelement stack attached to the cantilever to control the distance
between the cantilever tip and sample surface (separation distance), and a light beam from a
laser source that reflects off the cantilever surface to a photo-diode detector used to monitor
the motion of the cantilever as the scan head moves over the sample surface. The output from
the photo-diode is used in the image processor to generate the micrograph.
The AFM output signal is derived from the interaction between the cantilever tip and the
sample surface. The interaction produces an interaction force that is highly dependent on the
Nonlinear Dynamics

80
tip-sample separation distance. A typical force-separation curve is shown in Fig. 2. Above the
separation distance z
A
the interaction force is negative, hence attractive, and below z
A
the
interaction force is positive, hence repulsive. The separation distance z
B
is the point on the
curve at which the maximum rate of change of the slope of the curve occurs and is thus the
point of maximum nonlinearity on the curve (the maximum nonlinearity regime).


Fig. 1. Schematic of the basic atomic force microscope.

Fig. 2. Interaction force plotted as a function of the separation distance z between cantilever
tip and sample surface.
Modalities, such as AFM and AM-AFM, are available for near-surface characterization,
while UFM, AFAM, FMM, HFM, and RDF-AFUM are generally used to assess deeper
(subsurface) features at the nanoscale. The nanoscale subsurface imaging modalities
combine the lateral resolution of the atomic force microscope with the nondestructive
capability of acoustic methodologies. The utilization of the AFM in principle provides the
necessary lateral resolution for obtaining subsurface images at the nanoscale, but the AFM
alone does not enable subsurface imaging. The propagation of acoustic waves through the
bulk of the specimen and the impinging of those waves on the specimen surface in contact
with the AFM cantilever enable such imaging. The use of acoustic waves in the ultrasonic
range of frequencies leads to a more optimal resolution, since both the intensity and the
phase variation of waves scattered from nanoscale, subsurface structures increase with
increasing frequency

(berall, 1997).
A schematic of the equipment arrangement for the various A-AFM modalities is shown in
Fig. 3. The arrangement used for AFAM and FMM is shown in Fig. 3 where the indicated
switches are in the open positions. AFAM and FMM utilize ultrasonic waves transmitted
into the material by a transducer attached to the bottom of the sample. After propagating
through the bulk of the sample, the wave impinges on the sample top surface where it
excites the engaged cantilever. For AFAM and FMM the cantilever tip is set to engage the
sample surface in hard contact corresponding to the roughly linear interaction region below
Nonlinear Dynamics of Cantilever Tip-Sample Surface Interactions in Atomic Force Microscopy

81
z
A
of the force-separation curve. The basic equipment arrangement used for UFM is the
same as that for AFAM and FMM, except that the cantilever tip for UFM is set to engage the
sample in the maximum nonlinearity regime of the force-separation curve. The UFM output
signal is a static or dc signal resulting from the interaction nonlinearity.


Fig. 3. Acoustic-atomic force microscope equipment configuration. Switches are open for
AFAM, FMM, and UFM. Switches are closed for HFM and RDF-AFUM.
The equipment arrangement for RDF-AFUM and HFM is shown in Fig. 3 where the
indicated switches are in the closed positions. Similar to the AFAM, FMM and UFM
modalities, RDF-AFUM and HFM employ ultrasonic waves launched from the bottom of the
sample. However, in contrast to the AFAM, FMM and UFM modalities, the cantilever in
RDF-AFUM and HFM is also driven into oscillation. RDF-AFUM and HFM operate in the
maximum nonlinearity regime of the force-separation curve, so the nonlinear interaction of
the surface and cantilever oscillations produces a strong difference-frequency output signal.
For the AM-AFM modality only the cantilever is driven into oscillation and the tip-sample
separation distance may be set to any position on the force-separation curve. In one mode
of AM-AFM operation the rest or quiescent separation distance z
0
lies well beyond the
region of strong tip-sample interaction, i.e. the quiescent separation z
0
>> z
B
.
Various approaches to assessing the nonlinear behavior of the cantilever probe dynamics
have been published

(Kolosov & Yamanaka, 1993; Yamanaka et al., 1994; Nony et al., 1999;
Yagasaki, 2004; Lee et al., 2006; Kokavecz et al., 2006; Wolf & Gottlieb, 2002; Turner, 2004;
Stark & Heckl, 2003; Stark et al., 2004; Hlscher et al., 1999; Garcia & Perez, 2002). We
present here a general, yet detailed, analytical treatment of the cantilever and the sample as
independent systems in which the nonlinear interaction force provides a coupling between
the cantilever tip and the small volume element of sample surface involved in the coupling.
The sample volume element is itself subject to a restoring force from the remainder of the
sample. We consider only the lowest-order terms in the cantilever tip-sample surface,
interaction force nonlinearity. Such terms are sufficient to account for the most important
operational characteristics and material properties obtained from each of the various
acoustic-atomic force microscopies cited above. A particular advantage of the coupled
independent systems model is that the equations are valid for all regions of the force-
separation curve and emphasize the local curvature properties (functional form) of the
curve. Another advantage is that the dynamics of the sample, hence energy transfer
characteristics, can be extracted straightforwardly from the solution set using the same
mathematical procedure as that for the cantilever.
Nonlinear Dynamics

82
We begin by developing a mathematical model of the interaction between the cantilever tip
and the sample surface that involves a coupling, via the nonlinear interaction force, of
separate dynamical equations for the cantilever and the sample surface. A general solution
is found that accounts for the positions of the excitation force (e.g., a piezo-transducer) and
the cantilever tip along the length of the cantilever as well as for the position of the laser
probe on the cantilever surface. The solution contains static terms (including static terms
generated by the nonlinearity), linear oscillatory terms, and nonlinear oscillatory terms.
Individual or various combinations of these terms are shown to apply as appropriate to a
quantitative description of signal generation for AM-AFM and RDF-AFUM as
representatives of the various A-AFM modalities. The two modalities represent opposite
extremes in measurement complexity, both in instrumentation and in the analytical
expressions used to calculate the output signal. This is followed by a quantitative analysis of
image contrast for the A-AFM techniques. As a test of the validity of the present model,
comparative measurements of the maximum fractional variation of the Young modulus in a
film of LaRC
TM
-CP2 polyimide polymer are presented using the RDF-AFUM and AM-AFM
modalities.
2. Analytical model of nonlinear cantilever dynamics
2.1 General dynamical equations
The cantilever of the AFM is able to vibrate in a number of different modes in free space
corresponding to various displacement types (flexural, longitudinal, shear, etc.), resonant
frequencies, and effective stiffness constants. Although any shape or oscillation mode of the
cantilever can in principle be used in the analysis to follow, for definiteness and expediency
we consider only the flexural modes of a cantilever modeled as a rectangular, elastic beam
of length L, width a, and height b. We assume the beam to be clamped at the position x = 0
and unclamped at the position x = L, as indicated in Fig. 4. We consider the flexural
displacement y(x,t) of the beam to be subjected to some general force per unit length H(x,t),
where x is the position along the beam and t is time. The dynamical equation for such a
beam is
) t , x ( H
t
) t , x ( y
A
x
) t , x ( y
I E
2
2
B B
4
4
B
=

(1)
where E
B
is the elastic modulus of the beam, I = ab
3
/12 is the bending moment of inertia,
B

is the beam mass density, and A
B
= ab is the cross-sectional area of the beam.
The solution to Eq. (1) may be obtained as a superposition of the natural vibrational modes
of the unforced cantilever as
=

=1 n
cn n
) t ( ) x ( Y ) t , x ( y (2)
where
cn
is the nth mode cantilever displacement (n = 1, 2, 3, ) and the spatial
eigenfunctions Y
n
(x) form an orthogonal basis set given by (Meirovitch, 1967)
( ) ( ) x q cosh x q cos x q sinh x q sin
x q cosh x q cos
x q sinh x q sin
) x ( Y
n n n n
n n
n n
n
+

= . (3)
Nonlinear Dynamics of Cantilever Tip-Sample Surface Interactions in Atomic Force Microscopy

83

Fig. 4. Schematic of cantilever tip-sample surface interaction: z
0
is the quiescent (rest) tip-
surface separation distance (setpoint), z the oscillating tip-surface separation distance,
c

the displacement (positive down) of the cantilever tip,
s
the displacement of the sample
surface (positive up),
cn
k is the nth mode cantilever stiffness constant (represented as an
nth mode spring),
c
m the cantilever mass,
s
k the sample stiffness constant (represented as
a single spring),
s
m the active sample mass, and ) z ( F
0
and ) z ( F
0
are the linear and first-
order nonlinear interaction force stiffness constants, respectively, at z
0
.
The flexural wave numbers q
n
in Eq.(3) are determined from the boundary conditions as
cos(q
n
L)cosh(q
n
L) = -1 and are related to the corresponding modal angular frequencies
n

via the dispersion relation I E / A q
B B B
2
n
4
n
= . The general force per unit length H(x,t) can
also be expanded in terms of the spatial eigenfunctions as (Sokolnikoff & Redheffer, 1958)
) x ( Y ) t ( B ) t , x ( H
n
1 n
n
=

=
. (4)

Applying the orthogonality condition

mn
L
0
n m
L dx Y ) x ( Y =

(5)
(
mn
are the Kronecker deltas) to Eq. (4), we obtain

= d ) ( Y ) t , ( H ) t ( B
n
L
0
n
. (6)
We now assume that the general force per unit length acting on the cantilever is composed
of (1) a cantilever driving force per unit length H
c
(x,t), (2) an interaction force per unit length
H
T
(x,t) between the cantilever tip and the sample surface, and (3) a dissipative force per unit
length H
d
(x,t). Thus, the general force per unit length H(x,t) = H
c
(x,t) + H
T
(x,t) + H
d
(x,t). We
now assume that the driving force per unit length is a purely sinusoidal oscillation of
angular frequency
c
and magnitude P
c
. We also assume the driving force to result from a
Nonlinear Dynamics

84
drive element (e.g., a piezo-transducer) applied at the point x
c
along the cantilever length.
We thus write ) x x ( e P ) t , x ( H
c
t i
c c
c
=

where (x x
c
) is the Dirac delta function. The
interaction force per unit length H
T
(x,t) of magnitude P
T
is applied at the cantilever tip at x =
x
T
and is not a direct function of time, since it serves as a passive coupling between the
independent cantilever and sample systems. We thus write the interaction force per unit
length as H
T
(x,t) = P
T
(x x
T
). We assume the modal dissipation force per unit length H
d
(x,t)
to be a product of the spatial eigenfunction and the cantilever displacement velocity given
as ) dt / d )( x ( Y P ) t , x ( H
cn n d d
= . The coefficient B
n
(t) is then obtained from Eq. (6) as
) dt / d ]( dx ) x ( Y P [ ) x ( Y P ) x ( Y e P ) t ( B
cn n d T n T c n
t i
c n
c
+ =

(7)
where the integration in the last term is taken over the range x = 0 to x = L. Substituting Eqs.
(2) and (4) into Eq.( 1) and collecting terms, we find that the dynamics for each mode n must
independently satisfy the relation
) x ( Y ) x ( Y e P
dx
) x ( Y d
I E
dt
) t ( d
) x ( Y A
n c n
t i
c cn
4
n
4
B
2
cn
2
n B B
c

= +

(8)
dt
d
] dx ) x ( Y P [ ) x ( Y ) x ( Y P
cn
L
0
n d n T n T

+ + .
From Eq. (3) we write
n
4
n
4
n
4
Y q dx / Y d = . Using this relation and the dispersion relation
between q
n
and
n
, we obtain that the coefficient of
cn
in Eq.( 8) is given by
B B
2
n
4
n
4
B
A ) dx / Y d ( I E = . Multiplying Eq. (8) by Y
m
(x) and integrating from x = 0 to x =
L, we obtain
F e F k m
t i
c cn cn cn c cn c
c
+ = + +


(9)
where the overdot denotes derivative with respect to time, m
c
=
B
A
B
L is the total mass of
the cantilever and F
c
= P
B
LY
n
(x
c
). The tip-sample interaction force F is defined by F =
P
T
LY
n
(x
T
) and the cantilever stiffness constant k
cn
is defined by
2
n c cn
m k = . The damping
coefficient
c
of the cantilever is defined as dx ) x ( Y L P
n d c
= . Note that, with regard to the
coupled system response, for a given mode n the effective magnitudes of the driving term F
c

and the interaction force F are dependent via Y
n
(x
c
) and Y
n
(x
T
), respectively, on the positions
x
c
and x
T
at which the forces are applied. The damping factor, in contrast, results from a
more general dependence on x via the integral of Y
n
(x) over the range zero to L. If the
excitation force per unit length is a distributed force over the cantilever surface rather than
at a point, then the resulting calculation for F
c
would involve an integral over Y
n
(x) as
obtained for the damping coefficient.
The interaction force F in Eq. (9) is derived without regard to the cantilever tip-sample
surface separation distance z. Realistically, the magnitude of F is quite dependent on the
separation distance. In particular, various parameters derived from the force-separation
curve play an essential role in the response of the cantilever to all driving forces. We further
consider that the interaction force not only involves the cantilever at the tip position x
T
but
Nonlinear Dynamics of Cantilever Tip-Sample Surface Interactions in Atomic Force Microscopy

85
also some elemental volume of material at the sample surface. To maintain equilibrium it is
appropriate to view the elemental volume of sample surface as a mass element m
s
(active
mass) that, in addition to the interaction force, is subjected to a linear restoring force from
material in the remainder of the sample. We assume that the restoring force per unit
displacement of m
s
in the direction z toward the cantilever tip is described by the sample
stiffness constant k
s
.
The interaction force F between the cantilever tip and the mass element m
s
is in general a
nonlinear function of the cantilever tip-sample surface separation distance z. A typical
nonlinear interaction force F(z) is shown schematically in Fig. 2 plotted as a function of the
cantilever tip-sample surface separation distance z. The interaction force results from a
number of possible fundamental mechanisms including electrostatic forces, van der Waals
forces, interatomic repulsive (e.g., Born-Mayer) potentials, and Casimir forces (Law &
Rieutord, 2002; Lantz et al., 2001; Polesel-Maris et al., 2003; Eguchi & Hasegawa, 2002; Saint
Jean et al.,; Chan et al., 2001). It is also influenced by chemical potentials as well as hydroxyl
groups formed from atmospheric moisture accumulation on the cantilever tip and sample
surface (Cantrell, 2004).
Since the force F(z) is common to the cantilever tip and the sample surface element, the
cantilever and the sample form a coupled dynamical system. We thus consider the
cantilever and the sample as independent dynamical systems coupled by their common
interaction force F(z). Fig. 4 shows a schematic representation of the various elements of the
coupled system. The dynamical equations expressing the responses of the cantilever and the
sample surface to all driving and damping forces may be written for each mode n of the
coupled system as
t cos F ) z ( F k m
c c cn cn cn c cn c
+ = + +

(10)
) t cos( F ) z ( F k m
s s sn s sn s sn s
+ + = + +

(11)
where
cn
(positive down) is the cantilever tip displacement for mode n,
sn
(positive up)
is the sample surface displacement for mode n,
c
is the angular frequency of the cantilever
oscillations,
s
is the angular frequency of the sample surface vibrations,
c
is the damping
coefficient for the cantilever,
s
is the damping coefficient for the sample surface,
c
F is the
magnitude of the cantilever driving force,
s
F is the magnitude of the sample driving
force that we assume here to result from an incident ultrasonic wave generated at the
opposite surface of the sample. The factor is a phase contribution resulting from the
propagation of the ultrasonic wave through the sample material and is considered in more
detail in Section 2.2.
Eqs. (10) and (11) are coupled equations representing the cantilever tip-sample surface
dynamics resulting from the nonlinear interaction forces. The equations govern the
cantilever and surface displacements
cn
and
sn
, respectively at x = x
T
. In a realistic AFM
measurement of the cantilever response to the driving forces, the measurement point is not
generally at x = x
T
, but at the point x = x
L
at which the laser beam of the AFM optical
detector system strikes the cantilever surface. The cantilever response at x = x
L
is found from
Eq. (2) to be
= =

=1 n
cn L n c L
) t ( ) x ( Y ) t ( ) t , x ( y (12)
Nonlinear Dynamics

86
We note from Fig. 4 that for a given mode n, ) ( z z
sn cn o
+ = , where z
0
is the quiescent
separation distance between the cantilever tip and the sample surface (setpoint distance).
We use this relationship in a power series expansion of ) z ( F about
o
z to obtain
+ + + =
2
0 0 0 0 0
) z z )( z ( F
2
1
) z z )( z ( F ) z ( F ) z ( F (13)
+ + + + =
2
sn cn 0 sn cn 0 0
) )( z ( F
2
1
) )( z ( F ) z ( F
where the prime denotes derivative with respect to z. Substitution of Eq. (13) into Eqs. (10)
and (11) gives
t cos F ) z ( F ) z ( F )] z ( F k [ m
c c 0 sn 0 cn 0 cn cn c cn c
+ = + + + +

(14)
+ + +
2
sn cn 0
) )( z ( F
2
1

) t cos( F ) z ( F ) z ( F )] z ( F k [ m
s s 0 cn 0 sn 0 s sn s sn s
+ + = + + + +

(15)
+ + +
2
sn cn 0
) )( z ( F
2
1
.

It is of interest to note that Eqs. (14) and (15) were obtained assuming that the cantilever is a
rectangular beam of constant cross-section. Such a restriction is not necessary, since the
mathematical procedure leading to Eqs. (14) and (15) is based on the assumption that the
general displacement of the cantilever can be expanded in terms of a set of eigenfunctions
that form an orthogonal basis set for the problem. For the beam cantilever the
eigenfunctions are Y
n
(x). For some other cantilever shape a different orthogonal basis set of
eigenfunctions would be appropriate. However, the mathematical procedure used here
would lead again to Eqs. (14) and (15) with values of the coefficients appropriate to the
different cantilever geometry.
2.2 Variations in signal amplitude and phase from subsurface features
We consider a traveling stress wave of unit amplitude of the form
( )
( )
] e e Re[ kx t cos e
kx t i x
s
x
s

= , where is the attenuation coefficient, x is the
propagation distance,
s
is the angular frequency, t is time, c / k
s
= , and c is the phase
velocity, propagating through a sample of thickness a/2. We assume that the wave is
generated at the bottom surface of the sample at the position x = 0 and that the wave is
reflected between the top and bottom surfaces of the sample. We assume that the effect of
the reflections is simply to change the direction of wave propagation.
For continuous waves the complex waveform at a point x in the material consists of the sum
of all contributions resulting from waves which had been generated at the point x = 0 and
have propagated to the point x after multiple reflections from the sample boundaries. We
thus write the complex wave

A (t) as
Nonlinear Dynamics of Cantilever Tip-Sample Surface Interactions in Atomic Force Microscopy

87

A (t) = e
x
e
i(
s
tkx)
[1 +e
(a+ika)
++e
n(a+ika)
+]


= e
x
e
i(
s
tkx)
e
(a+ika)






n=0

n
= e
x
e
i(
s
tkx)
1
1e
(a+ika)

(16)
where the last equality follows from the geometric series generated by the infinite sum. The
real waveform A(t) is obtained from Eq. (16) as


A(t) = Re[A (t)] = e
x
(A
1
2
+A
2
2
)
1/2
cos(
s
t kx ) = e
x
Bcos(
s
t kx ) (17)
where

) ka cos a (cosh 2
ka cos e
A
a
1

=

, (18)

) ka cos a (cosh 2
ka sin
A
2

= , (19)

ka cos e
ka sin
tan
a
1

, (20)
and

2 / 1 a a 2 2 / 1 2
2
2
1
) ka cos e 2 e 1 ( ) A A ( B

+ = + = . (21)
The evaluation (detection) of a continuous wave at the end of the sample opposite that of
the source is obtained by setting x = a/2 in the above equations. It is at x = a/2 that the AFM
cantilever engages the sample surface. In the following equations we set x = a/2.
The above results are derived for a homogeneous specimen. Consider now that the
specimen of thickness a/2 having phase velocity c contains embedded material of thickness
d/2 having phase velocity
d
c . The phase factor c / a ka
s
= in Eqs.(17)-(21) must then be
replaced by ka - where

d d
s
d
s
c
c
kd
c c
c
d
c
1
c
1
d

=

= (22)
and c = c
d
c. We thus set x = a/2 and re-write Eqs.(17), (20), and (21) as
]
2
) ka (
t cos[ B e ) t ( A
s
2 / a


=

(23)
where

) ka cos( e
) ka sin(
tan
a
1


=

, (24)
and
Nonlinear Dynamics

88

2 / 1 a a 2
)] ka cos( e 2 e 1 [ B

+ = . (25)
We have assumed in obtaining the above equations that the change in the attenuation
coefficient resulting from the embedded material is negligible.
For small we may expand Eq. (23) in a power series about = 0. Keeping only terms to
first order, we obtain
+ = (26)
where

+

=


ka sin ) ka cos e (
1 ka cos e
2 2 a
a
. (27)
Eq.(22) is thus approximated as
) t cos( B e )
2 2
ka
t cos( B e ) t ( A
s
2 / a
s
2 / a
+ =

+ =

(28)
where
( )

+ = + =
2 2
ka
, (29)
+ =
2
ka
(30)
and

+

+ = +

=


ka sin ) ka cos e (
1 ka cos e
2
1
2
2 2 a
a
. (31)
Equation (28) reveals that the total phase contribution at x = a/2 is and from Eqs.(29) and
(31) that the phase variation resulting from embedded material is .
The fractional change in the Young modulus E / E is related to the fractional change in the
ultrasonic longitudinal velocity c / c as
11 11
C / C E / E ) / ( ) c / c 2 ( + = where is
the mass density of the sample and
11
C is the Brugger longitudinal elastic constant.
Assuming that the fractional change in the mass density is small compared to the fractional
change in the wave velocity, we may estimate the relationship between E / E and c / c as
c / c 2 E / E . This relationship may be used to express , given in Eq.(22) in terms of
) c / c )( c / c ( c / c
d d
= , in terms of E / E .
2.3 Solution to the general dynamical equations
We solve the coupled nonlinear Eqs. (14) and (15) for the steady-state solution by writing the
coupled equations in matrix form and using an iteration procedure commonly employed in
the physics literature (Schiff, 1968) to solve the matrix expression. The first iteration involves
solving the equations for which the nonlinear terms are neglected. The second iteration is
obtained by substituting the first iterative solution into the nonlinear terms of Eqs.(14) and
Nonlinear Dynamics of Cantilever Tip-Sample Surface Interactions in Atomic Force Microscopy

89
(15) and solving the resulting equations. The procedure provides solutions both for the
cantilever tip and the sample surface displacements. Since the procedure is much too
lengthy to reproduce here in full detail, only the salient features of the procedure leading to
the steady state solution for the cantilever displacement =
cn n c
Y are given. We begin
by writing

cn cn cn cn
+ + = (32)
and

sn sn sn sn
+ + = (33)
where
cn
and
cn
represent the first iteration (i.e. linear) static and oscillatory solutions,
respectively, for the nth mode cantilever displacement,
cn
represents the second iteration
(i.e., nonlinear) solution for the nth mode cantilever displacement, and
sn
,
sn
, and
sn

are the corresponding first and second iteration nth mode displacements for the sample
surface.
We note that for the range of frequencies generally employed in A-AFM the contribution
from terms in the solution set involving the mass of the sample element m
s
is small
compared to the remaining terms and may to an excellent approximation be neglected. We
thus neglect the terms involving m
s
in the following equations.
2.3.1 First iterative solution
The first iterative solution is obtained by linearizing Eqs.(14) and (15), writing the resulting
expression in matrix form, and solving the matrix expression assuming sinusoidal driving
terms
t i
c
c
e F

and
t i
s
s
e F

for the cantilever and sample surface, respectively. The first
iteration yields a static solution
cn
and an oscillatory solution
cn
for the cantilever. The
static solution is given by

) k k )( z ( F k k
) z ( F k
s cn o s cn
o s
cn
+ +
= . (34)
The first iterative oscillatory solution is given by
) t cos( Q ) t cos( Q
ss s cs cc cc c cc cn
+ + + = (35)
where

cc
tan
1
(
c
k
s
+
s
k
cn
)
c

s
m
c

c
3
+ F (z
0
)(
c
+
s
)
c
k
cn
k
s
(m
c
k
s
+
c

s
)
c
2
+ F (z
0
)(k
cn
+ k
s
m
c

c
2
)
, (36)

ss
tan
1
(
c
k
s
+
s
k
cn
)
s

s
m
c

s
3
+ F (z
0
)(
c
+
s
)
s
k
cn
k
s
(m
c
k
s
+
c

s
)
s
2
+ F (z
0
)(k
cn
+ k
s
m
c

s
2
)
, (37)

Q
cc
F
c
{[k
s
+ F (z
o
)]
2
+
s
2

c
2
}
1/2
{[k
cn
k
s

c
2
(m
c
k
s
+
c

s
)


+ F (z
o
)(k
cn
+ k
s
m
c

c
2
)]
2
+[
c
(
s
k
cn
+
c
k
s
) (38)
Nonlinear Dynamics

90
2 / 1 2
c s c o c s
3
c
} )] ( ) z ( F m

+ + ,
and


Q
cs
F
s
F (z
o
){[k
cn
k
s

s
2
(m
c
k
s
+
c

s
) + F (z
o
)(k
cn
+ k
s
m
c

s
2
)]
2
(39)

+[
s
(
s
k
cn
+
c
k
s
)
s
3

s
m
c
+ F (z
o
)
s
(
s
+
c
)]
2
}
1/2
.
2.3.2 Second iterative solution
The second iterative solution
cn
for each mode n of the cantilever is considerably more
complicated, since it contains not only sum-frequency, difference-frequency, and generated
harmonic-frequency components, but linear and static components as well. The second
iterative solution
cn
is thus written as

harm , cn sum , cn diff , cn lin , cn stat , cn cn
+ + + + = (40)
where
stat , cn
is a static or dc contribution generated by the nonlinear tip-surface
interaction,
lin , cn
is a generated linear oscillatory contribution,
diff , cn
is a generated
difference-frequency contribution resulting from the nonlinear mixing of the cantilever and
sample oscillations,
sum , cn
is a generated sum-frequency contribution resulting from the
nonlinear mixing of the cantilever and sample oscillations, and
harm , cn
are generated
harmonic contributions.
Generally, the cantilever responds with decreasing displacement amplitudes as the drive
frequency is increased above the fundamental resonance (for some cantilevers the second
resonance mode has the largest amplitude), even when driven at higher modal frequencies.
Thus, acoustic-atomic force microscopy methods do not generally utilize harmonic or sum-
frequency signals. For expediency, such signals from the second iteration will not be
considered here. Only the static, linear, and difference-frequency terms from the second
iteration solution are relevant to the most commonly used A-AFM modalities.
The static contribution generated by the nonlinear interaction force is obtained to be

2
ss
2
sc
2
cs
2
cc
2
o
s cn o s cn
o s
stat , cn
Q Q Q Q 2 [
)] k k )( z ( F k k [
) z ( F k
4
1
+ + + +
+ +

= (41)
] cos Q Q 2 ) 2 cos( Q Q 2
ss ss cs cc cc sc cc
+ +
where

) k k )( z ( F k k
) z ( F ) k k (
s cn o s cn
o s cn
o
+ +
+
= , (42)


Q
sc
F
c
F (z
o
){[k
cn
k
s

c
2
(m
c
k
s
+
c

s
) + F (z
o
)(k
cn
+ k
s
m
c

c
2
)]
2
(43)

+[
c
(
s
k
cn
+
c
k
s
)
c
3

s
m
c
+ F (z
o
)
c
(
s
+
c
)]
2
}
1/2
,


Q
ss
F
s
{[k
s
+ F (z
o
)]
2
+
s
2

s
2
}
1/2
{[k
cn
k
s

s
2
(m
c
k
s
+
c

s
) (44)
Nonlinear Dynamics of Cantilever Tip-Sample Surface Interactions in Atomic Force Microscopy

91

+ F (z
o
)(k
cn
+ k
s
m
c

s
2
)]
2
+[
s
(
s
k
cn
+
c
k
s
)
2 / 1 2
c s c o c s
3
s
} )] ( ) z ( F m

+ + ,

) z ( F k
tan
o s
c s 1
cc
+

=

, (45)

2
s c o cn
s c 1
ss
m ) z ( F k
tan
+

=

(46)
and
cc
is given by Eq. (36), Q
cc
by Eq. (38) and Q
cs
by Eq.(39).
The linear oscillatory contribution
lin , cn
generated by the nonlinear interaction force in
the second iteration is obtained to be
) 2 t cos( ] cos Q Q 2 Q Q )[ z ( F
R
D
cc c cc c
2 / 1
cc sc cc
2
sc
2
cc o o
cc
c
lin , cn
+ + + + = (47)
) 2 t cos( ] cos Q Q 2 Q Q )[ z ( F
R
D
ss s ss s
2 / 1
ss cs ss
2
cs
2
ss o o
ss
s
+ + + + + +
where

sc cc cc
cc cc 1
cc
Q cos Q
sin Q
tan
+

=

, (48)

cs ss ss
ss ss 1
ss
Q cos Q
sin Q
tan
+

=

, (49)

s
c s 1
c
k
tan

=

, (50)

s
s s 1
s
k
tan

=

, (51)

2 / 1 2
c
2
s
2
s c
] k [ D + = , (52)

2 / 1 2
s
2
s
2
s s
] k [ D + = , (53)

2 2
s c s cn o s c s c
2
s s cn ss
)] m k k )( z ( F ) k m ( k k {[ R + + + = (54)
2 / 1 2
c s s o
3
s c s s c cn s s
} )] ( ) z ( F m ) k k ( [ + + + + ,
and

2 2
c c s cn o s c s c
2
c s cn cc
)] m k k )( z ( F ) k m ( k k {[ R + + + = (55)
Nonlinear Dynamics

92
2 / 1 2
c s c o
3
c c s s c cn s c
} )] ( ) z ( F m ) k k ( [ + + + + .
The difference-frequency contribution
diff , cn
generated by the nonlinear interaction force
in the second iteration is obtained to be
] t ) cos[( G
cs cs ss cc s c n diff , cn
+ + + = (56)
where

2
sc
2
cs
2
ss
2
cc
2
ss
2
sc
2
cs
2
cc o
cs
cs
n
Q Q Q Q Q Q Q Q ){ z ( F
R
D
2
1
G + + + = (57)

cc sc
2
cs cc ss ss cs
2
cc ss cc ss sc cs cc
cos Q Q Q 2 cos Q Q Q 2 ) cos( Q Q Q Q 2 + + + +
2 / 1
ss cc sc cs ss cc ss cs ss
2
sc
)} cos( Q Q Q Q 2 cos Q Q Q 2 + + ,

2
s c
2
s
2
s cs
) ( k D + = , (58)

2
2 cs
2
1 cs cs
R R R + = , (59)
] ) ( m k k )[ z ( F ) ( ) ( k m k k R
2
s c c s cn o
2
s c s c
2
s c s c s cn 1 cs
+ + = , (60)
) )( )( z ( F ) ( m ) k k )( ( R
c s s c o
3
s c c s s c c s s c 2 cs
+ + + = , (61)

1 cs
2 cs 1
cs
R
R
tan

= (62)


tan
1
(
c
k
s
+
s
k
cn
)(
c

s
)
s
m
c
(
c

s
)
3
+ F (z
0
)(
c
+
s
)(
c

s
)
k
cn
k
s
(m
c
k
s
+
c

s
)(
c

s
)
2
+ F (z
0
)[k
cn
+ k
s
m
c
(
c

s
)
2
]
,

s
s c s 1
cs
k
) (
tan

=

, (63)
and

sc cs ss cc ss cc ss ss sc cc cs cc
ss cc ss cc ss ss sc cc cs cc 1
Q Q ) cos( Q Q cos Q Q cos Q Q
) sin( Q Q sin Q Q sin Q Q
tan
+ + +
+
=

. (64)

The phase term given by Eq.(64) is quite complicated. However, advantage can be taken
of the fact that k
s
is generally large compared to other terms in the numerators of Q
cc
, Q
ss
,
Q
cs
, and Q
sc
; the denominators of these terms are very roughly all equal. Hence, the
magnitudes of Q
cc
and Q
ss
are usually large compared to those of Q
cs
and Q
sc
. The terms
involving the product

Q
cc
Q
ss
thus dominate in Eq. (64) and we may approximate as
Nonlinear Dynamics of Cantilever Tip-Sample Surface Interactions in Atomic Force Microscopy

93



cc

ss
= tan
1

s

c
k
s
+ F (z
0
)
tan
1

c

s
k
cn
+ F (z
0
) m
c

s
2
(65)
where
cc
and
ss
are obtained from Eqs. (45) and (46), respectively. To the same extent that
may be approximated by Eq.( 65) we may also approximate G
n
, given by Eq. (57), as

ss cc
cs
cs 0
n
Q Q
R
D
2
) z ( F
G

. (66)
2.3.3 Important features of the solution set
The present derivation is based on the assumption that the cantilever tip-sample surface
interaction force is a multiply differentiable, nonlinear function of the tip-surface separation
distance as indicated in Fig. 2. Points on the curve below the separation distance z
A
in Fig. 2
correspond to a repulsive interaction force, while points above z
A
in Fig. 2 correspond to an
attractive interaction force. The force-separation curve has a minimum at a separation
distance z
B
corresponding to the maximum nonlinearity of the curve and that point lies in
the attractive force portion of the curve. Cantilever oscillations result in continuous
oscillatory changes in the tip-surface separation distance about the quiescent tip-surface
separation distance z
0
(see Fig. 4). Since the cantilever oscillations are constrained to follow
the force-separation curve, the fractions of the cantilever oscillation cycle in the repulsive
and attractive portions of the force-separation curve depend on the quiescent tip-surface
separation distance and the amplitude of the oscillations.
The cantilever oscillations are known to be bi-stable with the particular mode of oscillation
being determined by the initial conditions that includes the tip-surface separation distance
(Garcia & Perez, 2002). Unless some extraneous perturbation changes the mode of
oscillation, the cantilever continues to oscillate in a given bi-stable mode for a given set of
initial conditions. For large oscillation amplitudes the bi-stability coalesces to a single stable
mode. In the present model the bi-stable mode of cantilever oscillation is set by the value of
the effective sample stiffness constant k
s
that has one of two values one associated with
the dominantly repulsive portion of the force-separation curve and one associated with the
dominantly attractive portion (see Section 4.3). The value of the effective sample stiffness
constant, hence cantilever oscillation mode, must be determined experimentally in the
present model.
The total static solution to the coupled nonlinear equations (14) and (15) for the cantilever
stat , cn
is the sum of the contribution
cn
, given by Eq. (34), from the first iterative solution
and the contribution
stat , cn
, given by Eq. (41), from the second iteration as

stat , cn cn stat , cn
+ = . (67)
The total linear solution
lin , cn
to Eqs. (14) and (15) is the sum of the contribution
cn

given by Eq. (35) and the contribution
lin , cn
given by Eq. (47) as

lin , cn cn lin , cn
+ = . (68)
The total difference-frequency solution
diff , cn
to Eqs. (14) and (15) is simply the
contribution
diff , cn
given by Eq. (56).
Nonlinear Dynamics

94
It is interesting to note that
cn
and the component
o
in
stat , cn
do not explicitly involve
the cantilever drive amplitude
c
F and the sample surface drive amplitude
s
F , although
other terms involving the Q factors, given by Eqs. (38), (39), (43), and (44), in
stat , cn
do
involve these drive amplitudes. This means that only the contributions stemming from the
nonlinearity in the cantilever tip-sample surface interaction force respond directly to
variations in the drive amplitudes and in particular to the physical features of the material
giving rise to variations in
s
F . Further, the magnitudes of all second iteration (i.e. nonlinear)
contributions,
stat , cn
,
lin , cn
, and
diff , cn
are strongly dependent on the cantilever tip-
sample surface quiescent separation
o
z , since the value of the nonlinear stiffness constant
) z ( F
o
that dominates these contributions is highly sensitive to
o
z . Indeed, ) z ( F
o
attains
a maximum value near the bottom of the force-separation curve of Fig. 2.
For large deflections of the cantilever that may occur for sufficiently hard contact, large
bending moments may be introduced that produce frequency shifts in the cantilever
resonance frequencies quite apart from those introduced by the interaction force stiffness
constant ) z ( F
0
. For the assessment of ) z ( F
0
near the bottom of the force-separation curve
where the nonlinearity ) z ( F
0
is maximum (maximum nonlinearity regime) and ) z ( F
0
is
relatively small, the bending moments are generally negligible and a reasonable estimate of
) z ( F
0
can be obtained directly from differences in the engaged and non-engaged resonance
(free space) frequencies of the cantilever.
For large driving force amplitudes, nonlinear modes of oscillation may be generated in the
cantilever. Nonlinear tip-surface interactions are also known to excite nonlinear
(anharmonic) cantilever modes (Stark & Heckl, 2003; Garcia & Perez, 2002). It is assumed
that the nonlinear modes can be described in terms of a set of orthogonal eigenfunctions
Z
n
(x) describing the nonlinearities of the unforced cantilever that are generally different
from but orthogonal to Y
n
(x). In such case the nonlinear vibrational characteristics of the
cantilever may also be included in the general cantilever response in a manner similar to
that given above for the linear modes. The nonlinear modes are thus formally included in
the present model by extending the set of eigenvalues k
cn
, hence eigenvectors spanning the
function space, to allow for nonlinear eigenmodes. This requires no additional formal
analysis in the present model. All eigenvalues (including those from nonlinear modes) are
ascertained in the present model from experimental measurements.
3. Signal generation for representative A-AFM modalities
Generally, there are two working modes in A-AFM - the contact mode and the non-contact
mode. The contact mode is viewed as a modality for which the oscillating cantilever tip
makes periodic contact with the sample surface irrespective of the distance of separation
(setpont distance) between the non-oscillating (quiescent) cantilever tip and the sample
surface. When the setpoint distance z
0
lies close to the sample surface, the cantilever
operates near the dominantly repulsive portion of the cantilever tip-sample surface
interaction force-separation curve and experiences a dominantly repulsive force over some
appreciable fraction of an oscillation period (contact time). The oscillation amplitude is
usually small for this contact mode of operation and the tip-surface interaction force may be
approximated by a linear dependence of the tip-surface interaction force on the tip-surface
separation distance. A-AFM modalities that operate in the contact mode include force
Nonlinear Dynamics of Cantilever Tip-Sample Surface Interactions in Atomic Force Microscopy

95
modulation microscopy, atomic force acoustic microscopy, and a modality of amplitude
modulation-atomic force microscopy (AM-AFM) that may be descriptively called small
amplitude contact tapping mode.
Various other A-AFM modalities operate in the non-contact mode where the cantilever tip-
sample surface setpoint distance z
0
is sufficiently large that the cantilever tip, oscillating
with small amplitude, does not contact with the sample surface. In such cases the modalities
optimally operate in that portion of the force-separation curve that yields the maximum
force-separation nonlinearity, appropriately called the maximum nonlinearity regime of A-
AFM operation. Ultrasonic force microscopy, heterodyne force microscopy, and resonant
difference-frequency atomic force ultrasonic microscopy (RDF-AFUM) are examples of non-
contact A-AFM modalities. Non-contact amplitude modulation-atomic force microscopy
(noncontact tapping mode) also operates in this portion of the force-separation curve.
The equations derived in Section 2, describing the dynamical response of the cantilever
resulting from the cantilever tip-sample surface interaction forces, have been used to
quantify the signal generation and image contrast for all A-AFM modalities mentioned in
the introduction (Cantrell & Cantrell, 2008). We consider here, however, only resonant
difference-frequency atomic force ultrasonic microscopy (RDF-AFUM), and the commonly
used amplitude modulation-atomic force microscopy (AM-AFM), a modality that includes
the intermittent contact mode as well as contact and non-contact tapping modes. RDF-
AFUM and AM-AFM represent opposite extremes in complexity, both in instrumentation
and in the analytical expressions used to assess signal generation and image contrast.
RDF-AFUM uses input drive oscillations both to the cantilever and to the sample surface to
interrogate the sample. It is the most complex of the A-AFM modalities and the assessment
of signal generation and image contrast for RDF-AFUM requires application of the largest
number of equations from Section 2. The AM-AFM modality uses only an input drive
oscillation to the cantilever and is among the simplest of A-AFM modalities. The calculation
of the AM-AFM output signal thus requires relatively few equations from Section 2. The
AM-AFM modality may be viewed operationally and analytically as a subset of the RDF-
AFUM modality.
3.1 Resonant difference-frequency atomic force ultrasonic microscopy
Resonant difference-frequency atomic force ultrasonic microscopy

(RDF-AFUM) employs an
ultrasonic wave launched from the bottom of a sample, while the AFM cantilever tip
engages the sample top surface. The cantilever is driven at a frequency differing from the
ultrasonic frequency by one of the resonance frequencies of the engaged cantilever. It is
important to note that at high drive amplitudes of the ultrasonic wave or engaged cantilever
(or both) the resonance frequency generating the difference-frequency signal may
correspond to one of the nonlinear oscillation modes of the cantilever. The engaged
cantilever resonance frequency for the (linear or nonlinear) mode n, neglecting dissipation,
is given by
1
0 s cn 0 cn
2
cn c
)] z ( F k [ k ) z ( F k m

+ + = , where
cn
k is the cantilever stiffness
constant corresponding to the nth (linear or nonlinear) non-engaged (free space) resonance
mode. Since ) z ( F
0
may be positive or negative, depending on the shape of the force
separation curve, at the separation distance z
0
corresponding to maximum ) z ( F
0
, the
resonance frequency of the cantilever, when engaged at this value of z
0
, may be larger or
Nonlinear Dynamics

96
smaller, respectively, than the resonance frequency when not engaged. The nonlinear
mixing of the oscillating cantilever and the ultrasonic wave in the region defined by the
cantilever tip-sample surface interaction force generates difference-frequency oscillations at
the engaged cantilever resonance. Variations in the amplitude and phase of the bulk wave
due to the presence of subsurface nano/microstructures, as well as variations in near-
surface material parameters, affect the amplitude and phase of the difference-frequency
signal. These variations are used to create spatial mappings generated by subsurface and
near-surface structures.
In RDF-AFUM the cantilever difference-frequency response is obtained from the nonlinear
mixing in the region defined by the tip-surface interaction force. The interaction force varies
nonlinearly with the tip-surface separation distance. The deflection of the cantilever
obtained in calibration plots is related to this force. For small slopes of the deflection versus
separation distance, the interaction force and cantilever deflection curves are approximately
related via a constant of proportionality. The maximum difference-frequency signal
amplitude occurs when the quiescent deflection of the cantilever is near the bottom of the
force-separation curve (z
B
in Fig. 2). There the maximum change in the slope of the force
versus separation (hence maximum interaction force nonlinearity) occurs. We call this
region of operation the maximum nonlinearity regime.
The dominant term or terms for the cantilever difference-frequency displacement in Eqs.
(56) and (57) depend on the values of
cn
k for the free modes of cantilever oscillation, the
difference-frequency (
c

s
), and the value of ) z ( F
0
obtained at the quiescent separation
distance z
0
= (z
0
)B at which the maximum difference-frequency signal occurs. We designate
the non-engaged linear or nonlinear mode n for which the difference-frequency engaged
resonance occurs as n = p. The dominant difference-frequency component in Eqs.(56) and
(57) is thus
diff , cp diff , cp cp
= = and is given by Eq.(56) for n = p as
] t ) cos[( G
cs cs ss cc s c p diff , cp
+ + + = (69)
where Gp is given by Eq.(57) and in approximation by Eq.(66). The phase terms in Eq.(69)
are obtained from Eqs. (36), (37), (45), (46), and (62)-(64) where may be approximated by
Eq. (66).
It is important to point out in considering these equations that while the difference-
frequency resonance frequency ) (
s c
in RDF-AFUM is usually set to correspond to the
lowest resonance mode of the engaged cantilever (although a higher modal resonance could
be used), the cantilever driving frequency
c
and ultrasonic frequency
s
generally are set
near (but not necessary equal to) higher resonance modes n = q and n = r, respectively, of
the engaged cantilever. Thus, the cantilever stiffness constant k
cn
is appropriately given as
k
cp
when involving the difference-frequency terms in Eqs. (36)-(39), (42)-(46), and (58)-(64),
given as k
cq
when involving the cantilever drive frequency
c
at or near the frequency of
the qth cantilever resonance mode, and given as k
cr
when involving the ultrasonic frequency

s
at or near the frequency of the rth cantilever resonance mode. If
c
and
s
are not set at
or near a resonance modal frequency of the engaged cantilever, then it may be necessary to
include more than one term in Eqs. (12) and (32) corresponding to various values of q and r.
It is seen from Eq. (57) that for a given value of ) (
s c
the maximum value of
diff , cp

ideally occurs for a value of z
0
such that ) z ( F
0
is maximized. It is important to note,
Nonlinear Dynamics of Cantilever Tip-Sample Surface Interactions in Atomic Force Microscopy

97
however, that ) z ( F
0
, while relatively small in magnitude compared to that of the hard
contact regime, is generally not equal to zero at that point. Strictly, the values of ) z ( F
0
and
) z ( F
0
for a given z
0
are each dependent on the exact functional form of ) z ( F
0
. A functional
form for ) z ( F
0
sufficiently quantitative to quantify ) z ( F
0
and ) z ( F
0
is not typically
available. However, experimental curves for ) z ( F
0
can be obtained and compared to the
experimental curves of
diff , cp
plotted as a function of z
0
. An examination of Eq. (57)
suggests that a more exact approach to maximizing
diff , cp
would be not only to vary z
0

but also to vary slightly the difference-frequency from the free space resonance condition
until an optimal setting for both z
0
and the difference-frequency is achieved.
3.2 Amplitude modulation-atomic force microscopy
The amplitude modulation-atomic force microscopy (AM-AFM) mode (also called
intermittent contact mode or tapping mode) is a standard feature on many atomic force
microscopes for which the cantilever is driven in oscillation, but no surface oscillations
resulting from bulk ultrasonic waves are generated (i.e., Fs and
s
are zero). Thus, AM-AFM
cannot be used to image subsurface features, but interesting surface properties and features
can be imaged. Since AM-AFM can be used in both the hard contact and maximum
nonlinearity regimes (i.e. the linear and maximally nonlinear regimes, respectively, of the
force-separation curve), the cantilever displacement
lin , cn
for mode n is given most
generally as

lin , cn cn lin . cn
+ =
(70)
where
cn
is given by Eq.( 35) with the term involving
cs
Q set equal to zero and
lin , cn
is
given by Eq.(47) with all terms involving
cs
Q and
ss
Q set equal to zero.
3.2.1 Maximum nonlinearity regime
For the maximum nonlinearity regime the expression for
lin , cn
is
) t cos( H
cc c lin , cn
+ = (71)
where

) W / Q ( ) cos(
) sin(
tan
cc cc cc cc c
cc cc cc c 1
+ +
+
=

, (72)

2 / 1
cc sc cc
2
sc
2
cc 0 0
cc
c
) cos Q Q 2 Q Q )( z ( F
R
D
W + + = , (73)
and

2 / 1
cc cc cc c cc
2 2
cc
)] cos( W Q 2 W Q [ H + + + = (74)
where
cc
Q is given by Eq.(38),
sc
Q by Eq.(43),
cc
by Eq.(36),
cc
by Eq.(48),
0
by Eq.(42);
cc
a ,
c
,
c
D , and
cc
R , are given by Eqs.(45), (50), (52), and (53), respectively.
Nonlinear Dynamics

98
3.2.2 Hard contact regime
The complexity of the cantilever response
lin , cn
for AM-AFM is greatly reduced for the
hard contact regime, where ) z ( F
0
is negligibly small and ) z ( F
0
is very large and negative.
For sufficiently hard contact and
cc
are approximately zero and we obtain from Eq. (71)
that
) t cos( Q
cc c cc lin , cn
(75)
where

2 / 1 2
c
2
s c
2 2
c c s cn c cc
] ) ( ) m k k [( F Q

+ + + = (76)
and

2
c c s cn
c s c 1
cc
m k k
) (
tan
+
+
=

. (77)
The dependence of
lin , cn
on the material damping coefficient
s
and the sample stiffness
constant
s
k , both for the hard contact and the maximum nonlinearity regimes, means that
AM-AFM can be used to assess the viscoelastic properties of the material irrespective of the
regime of operation.
4. Image contrast for representative A-AFM modalities
All the above equations, except for Eqs.(26) - (31), were derived for constant values of the
cantilever and material parameters. If, in an area scan of the sample, the parameters remain
constant from point to point, the image generated from the scan would be flat and
featureless. We consider here that the sample stiffness constant
s
k may vary from point to
point on the sample surface. Since
s
k is dependent on the Young modulus E (see Section
4.3), this means that E also varies from point to point. We assume that the value of the
sample stiffness constant
s
k at a given point on the surface differs from the value
s
k at
another position as
s s s
k k k + = . For any function ) k ( f
s
having a functional dependence
on
s
k , a variation in
s
k generates a variation in ) k ( f
s
given by
s 0 s
k ) dk / df ( f = , where
the subscripted zero indicates evaluation at
s
k . A similar expression can be obtained for
the material damping parameter
s
, but we shall not consider such variations here.
A variation in
s
k produces a variation in both amplitude and phase of the signal generated
by the cantilever tip-sample surface interactions. The variations in amplitude and phase can
be used to generate amplitude and phase images, respectively, in a surface scan of the
sample. We consider here only images generated by the phase variations in the signal. The
equations for amplitude-generated images are given elsewhere (Cantrell & Cantrell, 2008).
The phase factors involved in RDF-AFUM are given from Eq.(69), (29), and (30) to be
cc
,
ss
,
cs
,
cs
, , and ; the phase factors involved in the AM-AFM mode are, from
Eq.(71),
cc
, and . Each of these phase factors is dependent on
s
k and the variations in
the phase factors resulting from variations in
s
k are responsible for image generation when
using phase detection of the A-AFM signal. The exact dependence of the phase on
s
k ,
however, is different for hard contact and maximum nonlinearity regimes.
Nonlinear Dynamics of Cantilever Tip-Sample Surface Interactions in Atomic Force Microscopy

99
4.1 Resonant difference-frequency atomic force ultrasonic microscopy
RDF-AFUM operates only in the maximum nonlinearity regime where the total variation in
phase is given as (
cs
+
cc
+
ss

cs
+ ). The phase factors relevant to RDF-
AFUM are given as

s
2 2
s
2
0 s
s
s
0
s
cs
cs
k
) ( )] z ( F k [
k
dk
d

+ +

=

= (78)
and

s
cc
cc
cc
k
B
A
= (79)
where

c s c
2
0 cq s 0
2
cq s cc
)] ( ) z ( F k ) z ( F 2 k [ A + + + = (80)
5
c s
2
c
3
c 0 cq c s s
2
c
m ))] z ( F k ( m 2 [ + + +
and

2 3
c c s c s c 0 cq s s c cc
} m )] )( z ( F k k {[ B + + + = (81)
2 2
c s c
2
c c cq 0 s 0
2
c c cq
} ) m k )( z ( F k )] z ( F m k {[ + + + ,

s
ss
ss
ss
k
B
A
= (82)
where

s s c
2
0 cr s 0
2
cr s ss
)] ( ) z ( F k ) z ( F 2 k [ A + + + = (83)
5
s s
2
c
3
s 0 cr c s s
2
c
m ))] z ( F k ( m 2 [ + + +
and

2 3
s c s s s c 0 cr s s c ss
} m )] )( z ( F k k {[ B + + + = (84)
2 2
s s c
2
s c cr 0 s 0
2
s c cr
} ) m k )( z ( F k )] z ( F m k {[ + + + ,
and

s
cs
cs
cs
k
B
A
= (85)
where
) )]( ( ) z ( F k ) z ( F 2 k [ A
s c
2
0 cp s 0
2
cp s cs
+ + + = (86)
Nonlinear Dynamics

100
5
s
2
c
3
0 cp c s s
2
c
) ( m ) ))]( z ( F k ( m 2 [ + + +
and

2 3
c s s c 0 cp s s c cs
} ) ( m ) )]( )( z ( F k k {[ B + + + = (87)
2 2
s c
2
c cp 0 s 0
2
c cp
} ) ( ] ) ( m k )[ z ( F k )] z ( F ) ( m k {[ + + + .

To the extent that
ss cc
= , as given by Eq.(65), we may write

s
2
c
2
s
2
0 s
c s
cc
k
)] z ( F k [

+ +

= = . (88)
The phase term is given by Eqs. (22) and (31).
4.2 Amplitude modulation-atomic force microscopy
The appropriate variations in the phase factors relevant to the AM-AFM or tapping mode
maximum nonlinearity regime are
cc
,
cc
, and . The factor is obtained from
Eq.(72) as

) ( sin )] W / Q ( ) [cos(
) cos( ) W / Q ( 1
cc cc cc c
2 2
cc cc cc cc c
cc cc cc c cc
+ + + +
+ +
= (89)
) (
cc cc cc c
+
where

s
2
c
2
s
2
s
c s
c
k
k

+

= , (90)
cc
is given by Eq. (79), and
cc
is obtained from Eq.( 48). To the extend that Q
sc
is much
smaller than Q
cc
, we get from Eq. (48) that
cc cc
= where
cc
is given by Eq. (88).
For the hard contact regime where ) z ( F
0
is very large and negative, the relevant phase
variation is obtained from Eq. (77) as

s
2
c
2
s c
2 2
c c cq s
c s c
cc
k
) ( ) m k k (
) (

+ + +
+
= . (91)
As a word of caution, the extent to which the hard contact equation applies depends on how
well the approximation ) z ( F
0
holds. In those cases where such an assumption is
suspect, the equations for the maximum nonlinearity regime should be used.
4.3 Dependence on the Young modulus
Hertzian contact theory provides that the sample stiffness constant k
s
is related to the Young
modulus E of the sample as (Yaralioglu et al., 2000)
Nonlinear Dynamics of Cantilever Tip-Sample Surface Interactions in Atomic Force Microscopy

101

1
2
T
2
T
c s
E
1
E
1
r 2 k

+

= (92)

where is the Poisson ratio of the sample material,
T
E and
T
are the Young modulus
and Poisson ratio, respectively, of the cantilever tip, and
c
r is the cantilever tip-sample
surface contact radius. Hence,

( )
E
E
E
1
E
1
) 1 (
E
k
E
E
1
E
1
E
1 r 2
k
1
2
T
2
T 2 s
2
2
T
2
T
2
2
c
s

+

=

+


=

. (93)

Strictly, Eq. (93) was derived for the case of repulsive interaction forces leading to a concave
elastic deformation of a flat sample surface from a contacting hard spherical object.
However, we consider here that to a crude approximation Eqs. (92) and (93) also hold for
attractive interactive forces providing that the elastic deformation of the sample surface is
viewed as a convex deformation (asperity) subtending an effective contact radius r
c
with the
cantilever tip that is appropriately different in magnitude from that of the repulsive force
case. As pointed out in Section 2.3.3, the cantilever oscillations are known to be bi-stable
with the particular mode of oscillation being determined by the initial conditions that
includes the tip-surface separation distance. In the present model the bi-stable mode of
cantilever oscillation is set by the value of the effective sample stiffness constant k
s

corresponding either to the dominantly repulsive region or dominantly attractive region of
the force-separation curve.
Eq. (93) can be used with Eqs. (78)-(91) to ascertain the fractional variation in the Young
modulus E / E from measurements of the phase variation in the signal from an appropriate
A-AFM modality. For the case where E
T
>> E, e.g. for polymeric or soft biological materials,
Eq. (92) reduces to k
s
= 2r
c
E and Eq. (93) reduces to k
s
= k
s
(E/E).
5. Assessment of model validity
We assess the validity of the above analytical model by comparing variations in the Young
modulus of a specimen as calculated from the model with independent experimental
measurements of the same specimen material. The choice of material is influenced by a
recent focus to develop high performance polymers having low density, high strength,
optical transparency, and high radiation resistance for a variety of applications in hostile
space environments. One such polymer is LaRC
TM
-CP2 polyimide. We consider here the
application of RDF-AFUM and AM-AFM to assess variations in the Young modulus of
nancomposites composed of nanoparticles embedded in a LaRC
TM
-CP2 polyimide matrix.
We consider two nanocomposites one embedded with gold nanoparticles and the other
embedded with single wall carbon (SWCNT) nanotube bundles.
We first consider a specimen of LaRC
TM
-CP2 polyimide polymer roughly 12.7 m thick
containing a monolayer of randomly distributed gold particles, roughly 10-15 nm in
diameter and embedded roughly 7 m beneath the specimen surface. Fig. 5a is an AM-AFM
Nonlinear Dynamics

102



Fig. 5. Micrographs of LaRC
TM
-CP2 polyimide polymer embedded with gold nanoparticles.
(a) Noncontact tapping mode (AM-AFM) phase-generated micrograph. (b) RDF-AFUM
phase-generated image over the same scan area as (a). (from Cantrell et al., 2007)
phase-generated image obtained in the maximum nonlinearity regime (noncontact tapping
mode). A commercial cantilever having a stiffness constant of 14 N m
-1
, a lowest-mode
resonance frequency of 302 kHz, and a cantilever damping coefficient of roughly 10
-8
kg s
-1

is driven at 2.1 MHz to obtain the micrograph of Fig.5a (Cantrell et al., 2007). The values of
the relevant model parameters for LaRC
TM
-CP2 polyimide polymer are 1.4 x 10
3
kg m
-3
for
the mass density , 2.4 GPa for the Young modulus E, 0.37 for the Poisson ratio , k
s
= 96.1
N m
-1
, and
s
= 4.8 x 10
-5
kg s
-1
(Park et al., 2002; Fay et al., 1999; Cantrell et al. 2007). Since
no bulk ultrasonic wave is involved, the image contrast results only from variations in the
specimen near-surface sample stiffness constant k
s
. The darker areas in the image
correspond to larger values of the sample stiffness constant, hence Young modulus, relative
to that of the brighter areas. The maximum phase difference between the bright and dark
areas in the image is approximately 1.5 degrees. Using the value 1.5 degrees, we obtain from
the model that the variation in the Young modulus E/E 18%. This value is consistent
with the value E/E 21%obtained from independent mechanical stretching experiments
on pure LaRC
TM
-CP2 polymer sheets (Fay et al., 1999).
An RDF-AFUM phase image of the same scan area as that of Fig. 5a is shown in Fig. 5b. The
RDF-AFUM image reveals bright and dark regions over the scan area that broadly
correspond to the bright and dark regions in the surface image of Fig. 5a, although the
image contrast and local detail appears to differ in the two images. F(z) is assessed to be
roughly 53 N m
-1
at the tip-surface separation corresponding to the maximum difference-
frequency signal. The acoustic wave has a frequency of 1.8 MHz. The maximum variation in
phase shown in Fig.5b is approximately 13.2 degrees. Using the value 13.2 degrees, we
obtain from the model that the variation in the Young modulus E/E 24%. This value is
also consistent with the value E/E 21%obtained from independent mechanical stretching
experiments on pure LaRC
TM
-CP2 polymer sheets.
The existence of contiguous material with differing elastic constants suggests that the
LaRC
TM
-CP2 material is not homogeneous. The broad coincidence of dark (bright) regions in
the images of Fig.5a and 5b suggests that the polymer structure giving rise to a larger
(smaller) elastic modulus in the bulk material occurs in varying amounts through the bulk
to the surface, the degree of darkness (brightness) in Fig. 5b being somewhat reflective of the
Nonlinear Dynamics of Cantilever Tip-Sample Surface Interactions in Atomic Force Microscopy

103
structural homogeneity of the material along the propagation path of the ultrasonic wave. It
is assumed that the appearance of contiguous material with different elastic coefficients may
result from the growth of a strain-nucleated harder material phase resulting from the
difference in the coefficients of thermal expansion between the polymer matrix and the
embedded gold particles.
To test the assumption of a strain-nucleated harder phase, micrographs were obtained of a
specimen formed from bundles of single-wall carbon nanotubes (SWCNTs) distributed
randomly through the bulk of a 50m-thick film of LaRC
TM
-CP2 polymer. Figure 6a shows a
conventional atomic force microscope (AFM) topographical image of the specimen showing
only surface features. A RDF-AFUM phase-image of the specimen, taken in the same scan
area as that of Fig 6a, is shown in Fig. 6b. Comparison of the two images reveals the
appearance of subsurface bundles of SWCNTs (dark contrast filamentary features) lying in
the plane of the RDF-AFUM image that do not appear in the AFM topographical scan.
Dramatic variations from dark to bright to slightly bright contrast occur in image plane
along portions of the boundary between the bundles of SWCNTs and the matrix material.
The variations follow the contour of the nanotube bundles and suggest the occurrence of an
interphase region (bright contrast feature) at the nanotube bundle-polymer interface. The
interphase consists of polymer material having dramatically different mechanical properties
from that of the matrix material. We note, however, that aside from the local interphase
regions in Fig. 6b there are no broad, contiguous regions of material with differing elastic
constants as observed in Fig. 5. Since the difference between the coefficients of thermal
expansion of LaRC
TM
-CP2 polymer and SWCNT bundles is considerable less than that for
LaRC
TM
-CP2 polymer and gold particles, we infer that the thermal strains in SWCNT
bundle-embedded polymer material are not sufficiently large to generate the larger
contiguous features observed in material embedded with gold particles.







Fig. 6. Micrographs of LaRC
TM
-CP2 polyimide polymer embedded with single wall carbon
nanotube bundles. (a) AFM topographical image. (b) RDF-AFUM phase-generated image
over the same scan area as (a).
Nonlinear Dynamics

104
6. Conclusion
The various dynamical implementations of the atomic force microscope have become
important nanoscale characterization tools for the development of novel materials and
devices. One of the most significant factors affecting all dynamical AFM modalities is the
cantilever tip-sample surface interaction force. We have developed a detailed mathematical
model of this interaction that includes a quantitative consideration of the nonlinearity of the
interaction force as a function of the cantilever tip-sample surface separation distance. The
model makes full use of cantilever beam dynamics and the multiply differentiability of the
continuous force-separation curve that results in a set of coupled differential equations,
Eqs.(14) and (15), for the displacement amplitudes of both the cantilever and the sample
surface. The coupled dynamical equations are recast in matrix form and solved by a
standard iteration procedure, but space limitations allow only a presentation of the salient
features of the procedure. Although the mathematical form of the coupled equations are
valid for any vibrational mode, only flexural vibrations of the cantilever and out-of-plane
oscillations of the sample surface are considered.
We emphasize that Eqs.(14) and (15) are obtained assuming that the cantilever is a
rectangular beam of constant cross-section, the dynamics of which are characterized by a set
of eigenfunctions that form an orthogonal basis for the solution set. For some other
cantilever shape a different orthogonal basis set of eigenfunctions would be appropriate.
However, the mathematical procedure used here would lead again to Eqs.(14) and (15) with
values of the coefficients appropriate to the different cantilever geometry. Practicably, this
means that the shape of the cantilever is not as important in the solution set as knowing the
cantilever modal resonant frequencies, obtained experimentally. The modal frequencies and
solution set are expanded to include nonlinear modes generated by nonlinear interaction
forces or large cantilever drive amplitudes.
A general steady state solution of the coupled dynamical equations is found that accounts
for the positions of the excitation force (e.g., a piezo-transducer) and the cantilever tip along
the length of the cantilever and for the position of the laser probe on the cantilever surface.
The solution is applied to two dynamical AFM modalities - resonant difference-frequency
atomic force ultrasonic microscopy, and the commonly used amplitude modulation-atomic
force microscopy. Image generation and contrast equations are obtained for each of the two
A-AFM modalities assuming for expediency that the contrast results only from variations in
the sample stiffness constant. Since the sample stiffness constant is related directly to the
Young modulus of the sample, the contrast can be expressed in terms of the variation in the
Young modulus from point to point as the sample is scanned. We note further the existence
of two values of the sample stiffness constant, corresponding to the dominantly attractive
and dominantly repulsive regimes of the force-separation curve. The two values allow for a
bi-stability in the cantilever oscillations that is experimentally observed.
Equations for both the maximum nonlinearity regime and the hard contact (linear) regime of
cantilever engagement with the sample surface are obtained. For dynamical AFM operation
outside these regimes, it is necessary to use all terms in the solution set given in Section 2 to
describe the signal output of a given A-AFM modality. The extent to which the hard contact
(linear regime) equations apply depends on how well the approximation ) z ( F
0
holds.
Nonlinear Dynamics of Cantilever Tip-Sample Surface Interactions in Atomic Force Microscopy

105
In those cases where such an assumption is suspect, all terms in the equations for a given
modality should be used.
In order to test the validity of the present model, comparative measurements of the
fractional variation of the Young modulus E/E in a film of LaRC
TM
-CP2 polyimide
polymer were obtained from phase-generated images obtained over the same scan area of
the specimen using the RDF-AFUM and AM-AFM maximum nonlinearity modalities. The
two modalities represent opposite extremes in measurement complexity, both in
instrumentation and in the analytical expressions used to calculate E/E. The values 24
percent calculated for RDF-AFUM and 18 percent calculated for the AM-AFM maximum
nonlinearity mode are in remarkably close agreement for such disparate techniques. The
agreement of both calculations with the value of 21 percent obtained from independent
mechanical stretching experiments of LaRC
TM
-CP2 polymer sheet material offers strong
evidence for the validity of the present model.
The present model can also be used to quantify the image contrast from variations in the
sample damping coefficient
s
in the material. Space limitations prohibit the inclusion of
such contrast mechanisms here, but the effects can be derived straightforwardly by the
reader from the equations derived in Section 2. Although the present model is developed for
flexural oscillations of the cantilever and out-of-plane vibrations of the sample surface, the
model can be extended to include other modes of cantilever oscillation and sample surface
response as well.
7. References
Binnig, G; Quate, C. F. & Gerber, Ch. (1986). Atomic force microscope. Physical Review
Letters, 56, 930-933.
Bolef, D. I. & and J. G. Miller, J. G. (1971). High-frequency continuous wave ultrasonics. In:
Physical Acoustics, Vol. VIII, W. P. Mason and R. N. Thurston, Ed., Academic, New
York, 95-201.
Cantrell, J. H. (2004). Determination of absolute bond strength from hydroxyl groups at
oxidized aluminum-epoxy interfaces by angle beam ultrasonic spectroscopy.
Journal of Applied Physics, 96, 3775-3781.
Cantrell, S. A.; Cantrell, J. H. & Lillehei, P. T. (2007). Nanoscale subsurface imaging via
resonant difference-frequency atomic force ultrasonic microscopy. Journal of
Applied Physics, 101, 114324.
Cantrell, J.H. & Cantrell, S. A. (2008). Analytical model of the nonlinear dynamics of
cantilever tip-sample surface interactions for various acoustic atomic force
microscopies. Physical Review B, 77, 165409.
Chan, H. B.; Aksyuk, V. A.; Kleiman, R. N.; Bishop, D. J. & Capasso, F. (2001). Nonlinear
micromechanical Casimir oscillator. Physical Review Letters, 97, 211801.
Cuberes, M. T.; Alexander, H. E.; Briggs, G. A. D. & and Kolosov, O. V. (2000). Heterodyne
force microscopy of PMMA/rubber nanocomposites: nanomapping of viscoelastic
response at ultrasonic frequencies. Journal of Physics D: Applied Physics, 33, 2347-
2355.
Nonlinear Dynamics

106
Cuberes, M. T. (2009). Intermittent-contact heterodyne force microscopy. Journal of
Nanomaterials, 2009, 762016.
Eguchi, T. & Hasegawa, Y. (2002). High resolution atomic force microscopic imaging of the
Si(111)-(7x7) surface: contribution of short-range force to the images. Physical
Review Letters, 89, 266105.
Fay, C. C.; Stoakley, D. M. & St. Clair, A. K. (1999). Molecularly oriented films for space
applications. High Performance Polymers, 11, 145-156.
Garcia, R & Perez, R. (2002). Dynamic atomic force microscopy methods. Surface Science
Reports, 47, 1-79.
Geer, R. E.; Kolosov, O. V.; Briggs, G. A. D. & Shekhawat, G. S. (2002). Nanometer-scale
mechanical imaging of aluminum damascene interconnect structures in a low-
dielectric-constant polymer. Journal of Applied Physics, 91, 9549-4555.
Hlscher, H.; Schwarz, U. D. & Wiesendanger, R. (1999). Calculation of the frequency shift
in dynamic force microscopy. Applied Surface Science, 140, 344-351.
Hurley, D. C.; Shen, K.; Jennett, N. M. & Turner, J. A. (2003). Atomic force acoustic
microscopy methods to determine thin-film elastic properties. Journal of Applied
Physics, 94, 2347-2354.
Kokavecz, J.; Marti, O.; Heszler, P. & and Mechler, A. (2006). Imaging bandwidth of the
tapping mode atomic force microscope probe. Physical Review B, 73, 155403.
Kolosov O. & Yamanaka, K. (1993). Nonlinear detection of ultrasonic vibrations in an atomic
force microscope. Japanese Journal of Applied Physics, 32, L1095-L1098.
Kolosov, O. V.; Castell, M. R.; Marsh, C. D.; Briggs, G. A. D.; Kamins, T. I. & Williams, R. S.
(1998). Imaging the elastic nanostructure of Ge islands by ultrasonic force
microscopy. Physical Review Letters, 81, 1046-1049.
Kopycinska-Mller, M.; Geiss, R. H. & Hurley, D. C. (2006). Contact mechanics and
tip shape in AFM-based nanomechanical measurements. Ultramicroscopy 106, 466-
474.
Lantz, M. A.; Hug, H. J.; Hoffmann, R.; van Schendel, P. J. A.; Kappenberger, P.; Martin, S.;
Baratoff, A. & Gntherodt, H.-J. (2001). Quantitative measurtement of short-range
chemical bonding forces. Science, 291, 2580-2583.
Law, B. M. & Rieutord, F. (2002). Electrostatic forces in atomic force microscopy. Physical
Review B, 66, 035402.
Lee, H.-L.; Yang, Y.-C.; Chang, W.-J. & Chu, S.-S. (2006). Effect of interactive damping on
vibration sensitivities of V-shaped atomic force microscope cantilevers. Japanese
Journal of Applied Physics, 45, 6017-6021.
Maivald, P.; Butt, H. J.; Gould, S. A.; Prater, C. B.; Drake, B.; Gurley, J. A.; Elings, V. B. &
Hansma, P. K. (1991). Using force modulation to image surface elasticities with the
atomic force microscope. Nanotechnology, 2, 103-106.
Meirovitch, L. (1967). Analytical Methods in Vibrations, Macmillan, New York.
Muthuswami, L. & Geer, R. E. (2004). Nanomechanical defect imaging in premetal
dielectrics for integrating circuits. Applied Physics Letters, 84, 5082-5084.
Nonlinear Dynamics of Cantilever Tip-Sample Surface Interactions in Atomic Force Microscopy

107
Nony, L.; Boisgard, R. & Aime, J. P. (1999). Nonlinear dynamical properties of an oscillating
tip-cantilever system in the tapping mode. Journal of Chemical Physics, 111, 1615-
1627.
Park, C.; Ounaies, Z.; Watson, K. A.; Crooks, R. E.; Smith, Jr., J.; Lowther, S. E.; J. Connell,
W.; Siochi, E. J.; Harrison, J. S. & St. Clair, T. L. (2002). Dispersion of single wall
carbon nanotubes by in situ polymerization under sonication. Chemical Physics
Letters, 364, 303-308.
Polesel-Maris, J; Piednoir, A.; Zambelli, T.; Bouju, X. & Gauthier, S. (2003). Experimental
investigation of resonance curves in dynamic force microscopy. Nanotechnology, 14,
1036-1042.
Rabe U. & Arnold, W. (1994). Acoustic microscopy by atomic force microscopy. Applied
Physics Letters, 64, 1493-1495.
Rabe, U.; Amelio, S.; Kopychinska, M.; Hirsekorn, S.; Kempf, M.; Goken, M. & Arnold, W.
(2002). Imaging and measurement of local mechanical properties by atomic force
micrscopy. Surface and Interface Analysis, 33, 65-70.
Saint Jean, M.; Hudlet, S.; Guthmann, C. & Berger, J. (1994). Van der Waals and capacitive
forces in atomic force microscopies. Journal of Applied Physics, 86, 5245-5248.
Schiff, L. I. (1968). Quantum Mechanics, McGraw-Hill, New York.
Shekhawat, G. S. & Dravid V. P. (2005). Nanoscale imaging of buried structures via scanning
near-field ultrasonic holography. Science, 310, 89-92.
Sokolnikoff, I. S. & Redheffer, R. M. (1958). Mathematics of Physics and Modern Engineering,
McGraw-Hill, New York.
Stark, R. W. & Heckl, W. M. (2003). Higher harmonics imaging in tapping-mode atomic-
force microscopy. Review of Scientific Instruments, 74, 5111-5114.
Stark, R. W.; Schitter, G.; Stark, M.; Guckenberger, R. & Stemmer, A. (2004). State-space
model of freely vibrating surface-coupled cantilever dynamics in atomic force
microscopy. Physical Review B, 69, 085412.
Turner, J. A. (2004). Nonlinear vibrations of a beam with cantilever-Hertzian contact
boundary conditions. Journal of Sound and Vibration, 275, 177-191.
berall, H. (1997). Interference and steady-state scattering of sound waves. In: Encyclopedia
of Acoustics, Vol. 1, Malcohm J. Crocker, (Ed.), 55-68, Wiley, ISBN 0-471-17767-9,
New York.
Wolf, K. & Gottlieb, O. (2002). Nonlinear dynamics of a noncontacting atomic force
microscope cantilever actuated by a piezoelectric layer. Journal of Applied Physics,
91, 4701-4709.
Yagasaki, K. (2004). Nonlinear dynamics of vibrating microcantilevers in tapping-mode
atomic force micrscopy. Physical Review B, 70, 245419.
Yamanaka, K.; Ogiso, H. & Kolosov, O. (1994). Ultrasonic force microscopy for nanometer
resolution subsurface imaging. Applied Physics Letters, 64, 178-180.
Yaralioglu, G. G.; Degertekin, F. L.; Crozier, K. B. & Quate, C. F. (2000). Contact stiffness of
layered materials for ultrasonic atomic force microscopy. Journal of Applied Physics,
87, 7491-7496.
Nonlinear Dynamics

108
Zheng, Y.; Geer, R. E.; Dovidenko, K.; Kopycinska-Mller, M. & Hurley, D. C. (2006).
Quantitative nanoscale modulus measurements and elastic imaging of SnO
2

nanobelts. Journal of Applied Physics, 100, 124308.
Zhong , Q.; Inniss, D.; Kjoller, K. & Elings, V. B. (1993). Fractured polymer/silica fiber
surface studied by tapping mode atomic force microscopy. Surface Science Letters,
290, L688 L692.
5
Nonlinear Phenomena during the
Oxidation and Bromination of Pyrocatechol
Takashi Amemiya
1
and Jichang Wang
1,2

1
Graduate School of Environment and Information Sciences,
Yokohama National University, Yokohama, 240-8501
2
Department of Chemistry and Biochemistry, University of Windsor,
Windsor, Ontario, N9B 3P4
1
Japan
2
Canada
1. Introduction
Many complex and interesting phenomena in nature are due to nonlinear interactions of the
constituents (Nicolis & Prigogine, 1977; Ball, 1999; Morowitz, 2002). The study of nonlinear
dynamical systems has achieved significant progress over the last four decades, which
allows scientists to understand various rather complicated behaviors such as self-
organization and pattern formation in the neuronal networks of brain (Scott Kelso, 1995).
Unusual properties of reagents in far-from-equilibrium conditions and the prevalence of
instability where small changes in initial conditions may lead to amplified effects have been
documented more than a century ago, but those nonlinear chemical phenomena did not get
much attention until late 1960s after the discovery of oscillatory behavior in a homogeneous
solution reaction between acidic bromate and malonic acid in the presence of metal catalyst
cerium (Field & Burgur, 1985; Scott, 1994; Epstein & Pojman 1998; Sagues & Epstein, 2003).
The system is now commonly known as the Belousov-Zhabotinsky (BZ) reaction (Zaikin &
Zhabotinsky, 1970; Field & Burger, 1985). Since then, the study of chemical oscillations and
wave formation has blossomed, which led to the observation of various nonlinear
spatiotemporal behaviours such as both simple and complex oscillations in a stirred system
(Smoes, 1979; Gyrgi & Field, 1992; Wang et al., 1995 & 1996; Zhao et al., 2005), Turing
pattern (Horvth et al., 2009), target and spiral waves in a two-dimensional reaction-
diffusion medium (Zaikin & Zhabotinsky, 1970; Winfree, 1972; Yamaguchi et al., 1991;
Steinbock et al., 1995; Kdr et al., 1998), and scroll waves in a 3-dimensional system (Welsh
et al., 1983; Winfree, 1987; Jahnke et al., 1988; Amemiya et al., 1996). Understanding the
onset of those exotic phenomena in chemical systems has provided important insight into
the formation of similar behaviour in nature (Goldbeter, 1996; Dutt & Menzinger, 1999;
Dhanarajan et al., 2002; Carlsson et al., 2006; Chiu et al., 2006).
As opposed to nonlinear systems in physical and biological areas, in which dynamic control
parameters are often inaccessible or difficult to adjust, chemical reactions can be
conveniently manipulated through adjusting the initial concentration of each reagent,
temperature, or flow rate in a continuously flow stirred tank reactor (CSTR) (Epstein, 1989;
Nonlinear Dynamics

110
Mori et al., 1993; Amemiya et al., 2002). As a result, chemical media have played a very
important role in gaining insights into various nonlinear behaviors encountered in nature
(Nicolis & Prigogine, 1989; Srensen et al., 1990; Kumli et al., 2003; Kurin-Csrgei et al.,
2004; McIIwaine et al., 2006).

Among existing chemical oscillators, the vast majority relies on
a few elements that possess multiple oxidation states, such as halogens, sulfur and some
transition metals. In 1978, Orbn and Krs carried out an extensive search to explore
chemical oscillations in the oxidations of aromatic compounds by acidic bromate (Krs &
Orbn, 1978; Orbn & Krs, 1978a; 1978b). Because of the absence of metal catalysts,
systems reported by Orbn and Krs in 1978 and discovered more recently by other groups
have been frequently referred as uncatalyzed bromate oscillators (UBO) (Farage & Janjic,
1982; Szalai & Krs, 1998; Adamcikova et al., 2001). In general, reactions of UBOs represent
the parallel running of oxidation and bromination of an organic substrate.
This chapter described nonlinear chemical kinetics in the bromate-pyrocatechol reaction
with or without the presence of metal catalysts (Harati & Wang, 2008a; 2008b). The bromate-
pyrocatechol reaction system was initially investigated by Orbn and Krs in 1978 (Orbn
& Krs, 1978a; 1978b). Unfortunately, no oscillatory behavior could be observed. The
absence of spontaneous oscillations in the earlier attempt has been attributed to two major
factors: First, the reaction between acidic bromate and pyrocatechol results in the
production of bromine, which inhibits autocatalytic reactions; secondly, the oxidation
product of pyrocatechol is a stable benzoquinone. As is shown in our recent reports, upon
extensive search in the concentration phase space the bromate-pyrocatechol reaction was
found to be capable of exhibiting spontaneous oscillations in a stirred batch system (Harati
& Wang, 2008b). A phase diagram established in the bromate and pyrocatechol
concentration space sheds light on why finding chemical oscillations in this chemical system
is such a challenging task. Same as reported in other UBOs (Wang et al., 2001; Zhao & Wang,
2006 & 2007), the bromate-pyrocatechol reaction exhibits subtle responses to illumination,
where, depending on the reaction conditions, either light-induced or light-quenched
oscillatory phenomena could be observed. The influence of metal catalysts on the nonlinear
dynamics of the bromate-pyrocatechol reaction was also discussed here.
2. Experimental observation of spontaneous oscillations
2.1 The uncatalyzed bromate-pyrocatechol reaction
Figure 1 presents three time series of the bromate-pyrocatechol (H
2
Q) reaction performed
under different initial concentrations of NaBrO
3
: (a) 0.085 M, (b) 0.093 M, and (c) 0.095 M.
Other reaction conditions are [H
2
Q] = 0.057 M and [H
2
SO
4
] = 1.4 M. Details of the
experimental procedure can be found in the original reports (Harati & Wang, 2007b). Shortly
after mixing all chemicals together, Pt potential as seen in Fig. 1a exhibited clock reaction
phenomenon, which was followed by gradual decrease for several hours.
Phenomenologically, the excursion of the Pt potential was accompanied by a dramatic color
change of the reaction solution from transparent to deep red. After the rapid color change,
which has been observed in all of the following experiments, the red color gradually turned
into yellow within the next two hours. Our experiments showed that for low bromate
concentrations (<0.09 M), Pt potential of the system decreased monotonically after the initial
excursion. Chemical oscillations were obtained when bromate concentration was increased
to 0.093 M. Further increase of bromate concentration led to slightly irregular oscillations in
Fig. 1c, where not only the amplitude but also the frequency of oscillation fluctuated. To
Nonlinear Phenomena during the Oxidation and Bromination of Pyrocatechol

111

Fig. 1. Time series of the pyrocatechol bromate reaction at different initial concentrations
of bromate: (a) 0.085 M, (b) 0.093 M and (c) 0.095M. Other reaction conditions are [H
2
Q] =
0.057 M, and [H
2
SO
4
] = 1.4 M.
show modulations in the oscillation frequency clearly, only the oscillation window is plotted
in Fig. 1c, in which the long induction time period, similar to the ones plotted in Figs. 1a and
1b, is omitted. As bromate concentration was increased continuously, the system underwent
reverse bifurcations leading the system back to non-oscillatory progress in time where the
evolution of Pt potential was the same as that in Fig. 1a. For conditions employed in Fig. 1,
spontaneous oscillations have been obtained when bromate concentration was between 0.09
and 0.11 M (Harati & Wang, 2007b).
Figure 2a plots the number of oscillation peak as a function of bromate concentration, where
it increases with bromate concentration and then drops sharply to 0 as the system moves out
of the oscillation window at the high bromate concentration. Fig. 2b illustrates that the
induction time (IP) of these spontaneous oscillations grows monotonically with the increase
Nonlinear Dynamics

112

Fig. 2. Dependence of the number of oscillations (N) and induction period (IP) on the initial
concentration of bromate. Other reaction conditions are [H
2
Q] = 0.057 M and [H
2
SO
4
] = 1.2
M.
of bromate concentration. The extremely long induction seen here is similar to that reported
in other uncatalyzed bromate oscillators (Farage & Janjic, 1982; Szalai & Krs, 1998;
Adamcikova et al., 2001)
Figure 3 presents temporal evolutions of the bromate-H
2
Q reaction under different initial
concentrations of H
2
Q: (a) 0.038 M, (b) 0.044 M, and (c) 0.047 M. In Fig. 3a Pt potential
exhibits a clock reaction phenomenon, followed by a gradual decrease. This behavior is the
same as seen at a low bromate concentration, where the clock variation of Pt potential is
accompanied by a dramatic color change of the reaction solution. When H
2
Q concentration
was increased to 0.044 M in Fig. 3b, spontaneous oscillations took place at about 2 hours
after the solution has turned into yellow. Further increase of H
2
Q concentration also resulted
in some irregularity in those transient oscillations such as the one shown in Fig. 3c. Again, to
show details of the chemical oscillations time scale in Fig. 3c is different from that used in
Figs. 3a and 3b. Within the oscillation window the induction time decreased monotonically
with the increase of H
2
Q concentration. On the other hand, the total number of oscillations
increased rapidly as H
2
Q concentration became larger than the lower bifurcation threshold
and then decreased gradually as H
2
Q concentration was increased further. The above results
indicate that bromate and H
2
Q have opposite effects on the oscillatory behavior.
Nonlinear Phenomena during the Oxidation and Bromination of Pyrocatechol

113

Fig. 3. Time series of the pyrocatechol bromate reaction at different initial concentrations
of pyrocatechol: (a) 0.038 M, (b) 0.044 M and (c) 0.047M. Other reaction conditions are
[NaBrO
3
] = 0.085 M, and [H
2
SO
4
] = 1.4 M
Figure 4 is a phase diagram in the pyrocatechol - bromate concentration plane, where filled
triangles denote the conditions at which the system exhibits spontaneous oscillations. Here,
the concentration of H
2
SO
4
is fixed at 1.4 M. First glance of this phase diagram indicates that
the oscillatory behavior exists over broad concentrations of pyrocatechol and bromate.
However, at each given concentration of bromate (or pyrocatechol) there is only a narrow
range of pyrocatechol (or bromate) concentration within that the system oscillates. This
diagonal narrow band window sheds light on the difficulty of landing the initial conditions
within such a window, when starting the experiments without existing information of this
system.
Dependence of the above chemical oscillations on H
2
SO
4
and bromate concentrations is
summarized in Fig. 5, in which H
2
Q concentration was fixed at 0.057 M. Filled triangles are
Nonlinear Dynamics

114

Fig. 4. Phase diagram of the bromate-pyrocatechol reaction in the pyrocatechol bromate
concentration plane. () denotes where the system exhibits transient oscillations. Sulfuric
acid concentration was fixed at 1.4 M.

Fig. 5. Phase diagram of the bromate-pyrocatechol reaction in the bromate H
2
SO
4

concentration plane. () denotes the conditions under which the system exhibits
oscillations. The concentration of pyrocatechol is 0.057 M.
the conditions under which the system exhibits spontaneous oscillations. This phase
diagram shows that when the concentration of H
2
SO
4
is larger than 2.5 M or smaller than 0.9
M, no oscillations can be obtained regardless bromate concentration. On the other hand, the
range of H
2
SO
4
concentration over which the system exhibits spontaneous oscillations is
broadened by lowering bromate concentration.
2.2 The ferroin-bromate-pyrocatechol reaction
Figure 6 presents time series of (a) the uncatalyzed and (b) ferroin-catalyzed bromate-
pyrocatechol reactions. In the uncatalyzed system, the Pt potential decreased gradually after
the initial excursion and then reached a plateau. In general, one might have considered that
Nonlinear Phenomena during the Oxidation and Bromination of Pyrocatechol

115
this closed reaction is over. However, the Pt potential suddenly started oscillating after
another two hours, and the oscillatory process lasted for longer than an hour with about 14
peaks. This result illustrates that under the conditions investigated here the uncatalyzed
bromate-pyrocatechol is capable of exhibiting spontaneous oscillations. There is no periodic
color change during the oscillation and thus the uncatalyzed system is deemed unsuitable
for studying chemical waves in spatially extended media.


Fig. 6. Time series of the (a) uncatalyzed, and (b) ferroin-catalyzed bromate-pyrocatechol
reaction. Other reaction conditions are: [H
2
SO
4
] = 1.30 M, [H
2
Q] = 0.043 M and [BrO
3
-
] =
0.078 M. The concentration of ferroin is equal to 1.0 x 10
-4
M in (b).
In Fig. 6b, when 1.0 x 10
-4
M ferroin was added to the bromate-pyracatechol reaction,
spontaneous oscillations commenced at about the same time as in the uncatalyzed system.
However, there are significant changes in the frequency of oscillation and the total number of
oscillations and both have been increased greatly. Notably, in this catalyzed system the
oscillation lasted for longer than 4 hours. Our experiments illustrate that this system exhibits
observable periodic color changes from yellowish to faint pink during the oscillatory window
when the concentration of ferroin is above 1.0 x 10
-4
M. Further increase of the concentration of
ferroin results in a better contrast, but reduces the lifetime of the oscillatory period.
Furthermore, when ferroin concentration is higher than 1.0 x 10
-3
M no obvious color change
could be seen in the stirred system. After oscillations in the ferroin-bromate-pyracatechol
system stopped, the solution has a blue color if the concentration of ferroin added is above 1.0
x 10
-3
M, or a pink color when the ferroin concentration is less than 5 10
-4
M.
Figure 7 summarizes the dependence of the number of oscillations (N) and the induction
time (IP) on the concentration of ferroin. There is a sharp increase in the number of
oscillations at a very low concentration of ferroin (10
-5
M), suggesting that the presence of
small amounts of metal catalyst favours the oscillatory behaviour. As the amount of ferroin
is increased, however, the number of oscillations decreases, which may be due to the
Nonlinear Dynamics

116
increased consumption of the reactants. Notably, ferroin shows a little effect on the
induction time (IP), where increasing ferroin concentration to 0.002 M only reduces the IP by
about 10 percent (Harati & Wang, 2008a).


Fig. 7. Dependence of the number of oscillations (N) and induction period (IP) on the
concentration of ferroin. Other reaction conditions are: [H
2
SO
4
] = 1.30 M, [BrO
3
-
] = 0.078 M,
and [pyrocatechol] = 0.043 M.


Fig. 8. Dependence of the number of oscillations (N) and induction time (IP) of the ferroin-
catalyzed system on the concentration of bromate and sulfuric acid. Other reaction
conditions are: [H
2
Q] = 0.044 M, [ferroin] = 1.0 x 10
-4
M, and (a &b) [H
2
SO
4
] = 1.40 M; (c&d)
[NaBrO
3
] = 0.085 M.
Nonlinear Phenomena during the Oxidation and Bromination of Pyrocatechol

117
Figure 8 plots, respectively, the number of oscillations (N) and induction time (IP) as a
function of concentrations of bromate and sulfuric acid in the ferroin-bromate-pyrocatechol
system, where the concentration of ferroin was fixed at 1.0 x 10
-4
M. Figs. 8a and 8b show
that increasing bromate concentration prolongs the induction period, which may arise from
the production of larger amounts of bromine in the reaction solution. The number of peaks
ascends first and then declines slightly with increasing bromate concentration. Under the
conditions studied here, the concentration of bromate must be between 0.070 and 0.085 M
for the system to show spontaneous oscillations. As shown in Figs. 8c and 8d, both N and IP
increase monotonically with the increase of H
2
SO
4
concentration. The system does not
oscillate when the concentration of H
2
SO
4
is higher than 1.4 M or lower than 1.0 M under
the conditions studied.
Figure 9 is a phase diagram of the ferroin-catalyzed system in the pyrocatechol and bromate
concentration plane, where () indicates the conditions under which the system exhibits
spontaneous oscillations. Similar to the situation of the uncatalyzed bromate-pyrocatechol
reaction, the first glance of this figure suggests that the system is able to exhibit oscillatory
dynamics over a broad range of bromate and pyrocatechol concentrations. However, at each
given concentration of pyrocatechol (or bromate), the proper concentration of bromate (or
pyrocatechol) is quite narrow. This narrow band shaped phase diagram suggests that
nonlinear behavior of this catalyzed system is more sensitive to the ratio of [H
2
Q]/[BrO
3
-
]
than their absolute concentrations. In comparison to the uncatalyzed bromate-pyrocatechol
system, the presence of ferroin does not change the shape of this phase diagram, but makes
the area of the parameter window slightly larger, implicating that ferroin favors the
oscillations.


Fig. 9. Phase diagram of the ferroin-catalyzed reaction in the bromatepyrocatechol
concentration plane. () denotes where the system exhibits simple periodic oscillations. The
concentration of ferroin is 1.0 x 10
-4
M.
Time series measured with a bromide selective electrode show that bromide concentration
increases slowly during the long induction time and then starts oscillating (Harati & Wang,
2008b). It is similar to the behavior reported in earlier studies of the uncatalyzed bromate-
1,4-cyclohexanedione and bromate-1,4-benzoquinone reactions (Szalai & Krs, 1998; Zhao
& Wang, 2006), in which the accumulation of bromide precursors has been suggested to be
responsible for the induction time. In this system, however, the initial addition of bromide,
which leads to the rapid production of bromine and then causes the bromination of
pyrocatechol, evidenced by mass spectrometry study (Harati & Wang, 2008b), does not
Nonlinear Dynamics

118
shorten the induction time. The slight decrease in the induction time observed at a very high
bromide concentration may result from decreases in H
2
Q and BrO
3
-
concentrations due to
reactions with bromine. The insensitivity of the induction time to the initial presence of
brominated substrates suggests that the governing mechanism of this oscillator may be
different from UBOs reported earlier.
2.3 The influence of Ce
4+
/Ce
3+
and Mn
3+
/Mn
2+

It is well known that metal catalysts such as ferroin participate the autocatalytic reactions
with bromine dioxide radicals (BrO
2
*) and therefore redox potential of the metal catalyst in
relative to the redox potential of HBrO
2
/BrO
2
* couple is an important parameter in
determining the rate of the autocatalytic cycle, which in turn has significant effects on the
overall reaction behavior. In the BZ reaction, four metal catalysts including ferroin,
ruthenium, cerium and manganese can be oxidized by bromine dioxide radicals, in which
the redox potential of HBrO
2
/BrO
2
* couple is larger than that of ferroin and ruthenium, but
smaller than that of Ce
4+
/Ce
3+
and Mn
3+
/Mn
2+
. Therefore, it is anticipated that when cerium
or manganese ions are introduced into the bromate-pyrocatechol reaction, behavior different
from that achieved in the ferroin-bromate-pyrocatechol system may emerge. Figure 10 plots
the number of oscilllations (N) and induction time (IP) of the catalyzed bromate-
pyrocatechol reaction as a function of catalyst (i.e. Ce
4+
and Mn
2+
) concentration.


Fig. 10. Dependence of the number of oscillations (N) and induction time (IP) on the initial
concentrations of cerium and menganese. Other reaction conditions are [H
2
SO
4
] = 1.3 M,
[NaBrO
3
] = 0.078 M, and [H
2
Q] = 0.043 M.
The sharp increase in the number of oscillations at the low concentration of cerium and
manganese illustrates that the presence of a small amount of metal catalyst favours the
oscillatory behaviour, similar to the case of ferroin. As the amount of catalyst (i.e. Ce
4+
or
Mn
2+
) increases, however, the number of oscillations decreases rapidly. It could be due to
the increased consumption of major reactants, in particular bromate. Overall, the effect of
Mn
2+
or Ce
4+
on the number of oscillations was not as significant as ferroin, although they
Nonlinear Phenomena during the Oxidation and Bromination of Pyrocatechol

119
doubled the number of peaks at an optimized condition. In contrast, the presence of a small
amount of cerium or manganese dramatically reduced the induction time, where the
induction time was shortened from about 3 hours in the uncatalyzed system to
approximately half an hour when the concentration of manganese and cerium reached,
respectively, 2.0 10
-4
and 5.0 10
-5
M. The IP became relative stable when the
concentration of manganese or cerium was increased further.
When comparing with the time series of the ferroin system presented in Fig. 6b, for the
cerium-catalyzed bromate-pyrocatechol reaction the Pt potential stayed flat after the initial
excursion. The amplitude of oscillation became significantly larger than that of the
uncatalyzed as well as the ferroin-catalyzed systems; but, there was no significant increase
in the total number of oscillations when compared with the uncatalyzed system. Unlike the
ferroin-catalyzed system, no periodic color change was achieved and thus is unfit for
studying waves. A short induction time and large oscillation amplitude (> 300 mV),
however, make the cerium-catalyzed system suitable for exploring temporal dynamics in a
stirred system. In particular, oscillations in the cerium system have a broad shoulder which
may potentially develop into complex oscillations. Times series of the Mn
2+
-catalyzed
bromate-pyrocatechol reaction is very similar to that of the cerium-catalyzed one, in which
the Pt potential stayed flat after the initial excursion and the oscillation commenced much
earlier than in the uncatalyzed system. The number of oscillations in the manganese system
is also slightly larger than that of the uncatalyzed system. Overall, cerium and manganese,
both have a redox potential above the redox potential of HBrO
2
/BrO
2
*, exhibit almost the
same influence on the reaction behavior.
2.4 Photochemical behavior
Ferroin-catalyzed BZ reaction is insensitive to the illumination of visible light. As a result,
the vast majority of existing studies on photosensitive chemical oscillators have been
performed with ruthenium as the metal catalyst, despite that ruthenium complex is
expensive and difficult to prepare. In Figure 11, the photosensitivity of the ferroin-catalyzed
bromate-pyrocatechol reaction was examined, in which the concentration of ferroin was
adjusted. As shown in Fig. 11a, when the system was exposed to light from the beginning of
the reaction, spontaneous oscillations emerged earlier, where the induction time was
shortened to about 6000 s, but the oscillatory process lasted for a shorter period of time. The
system then evolved into non-oscillatory evolution. Interestingly, after turning off the
illumination the Pt potential jumped to a higher value immediately and, more significantly,
another batch of oscillations developed after a long induction time. The above result
indicates that the ferroin-bromate-pyrocatechol reaction is photosensitive and influence of
light in this ferroin-catalyzed system is subtle. On one hand, illumination seems to favor the
oscillatory behavior by shortening the induction time, but it later quenches the oscillations.
In Fig. 11b the concentration of ferroin was doubled. When illuminated with the same light
as in Fig. 11a from the beginning, no oscillation was achieved, except there was a sharp drop
in the Pt potential at about the same time as that when oscillations occurred in Fig. 11a.
After turning off the light, the un-illuminated system exhibited oscillatory behaviour with a
long induction time. We have also applied illumination in the middle of the oscillatory
window, in which a strong illumination such as 100 mW/cm
2
immediately quenched the
oscillatory behaviour and oscillations revived shortly after reducing light intensity to a
lower level such as 30 mW/cm
2
. Interestingly, although ferroin itself is not a photosensitive
Nonlinear Dynamics

120

Fig. 11. Light effect on the bromate pyrocatechol ferroin reaction (a) and (b) light
illuminating from the beginning with intensity equal to 70 mW/cm
2
, under conditions
[NaBrO
3
] = 0.10 M, [H
2
SO
4
] = 1.40 M, [H
2
Q] = 0.057 M, (a) [Ferroin] = 5.010
-4
M, and (b)
[Ferroin] = 1.010
-3
M.
reagent, here its concentration nevertheless exhibits strong influence on the photoreaction
behaviour of the bromate-pyrocatechol system. Carrying out similar experiments with the
cerium- and manganese-catalyzed system under the otherwise the same reaction conditions
showed little photosensitivity, in which no quenching behaviour could be obtained,
although light did cause a visible decrease in the amplitude of oscillation.
3. Modelling
3.1 The model
To simulate the present experimental results, we employed the Orbn, Krs, and Noyes
(OKN) mechanism (Orbn et al., 1979) proposed for uncatalyzed reaction of aromatic
compounds with acidic bromate. The original OKN mechanism is composed of sixteen
reaction steps, i.e., ten steps K1 K10 in Scheme I and six steps K11 K16 in Scheme II as
listed in Table 1. We selected all ten reaction steps K1 K10 from Scheme I and the first four
reaction steps K11 K14 in Scheme II. The reason behind such a selection is that all reaction
steps in Scheme I as well as the first four reaction steps in Scheme II are suitable for an
aromatic compound containing at least two phenolic groups such as pyrocatechol used in
the present study.
Reaction steps K15 and K16 in Scheme II, on the other hand, suggest how phenol and its
derivatives could be involved in the oscillatory reactions. There is no experimental evidence
that pyrocatechol can be transformed into a substance of phenol type, we thus did not take
into account reactions involving phenol and its derivatives. The model used in our
Nonlinear Phenomena during the Oxidation and Bromination of Pyrocatechol

121
simulation consists of fourteen reaction steps K1 K14, and eleven variables, BrO
3
-
, Br
-
,
BrO
2
*, HBrO
2
, HOBr, Br*, Br
2
, HAr(OH)
2
, HAr(OH)O*, Q, and BrHQ, where HAr(OH)
2
is
pyrocatechol abbreviated as H
2
Q in the experimental section, HAr(OH)O* is pyrocatechol
radical, HArO
2
is 1,2-benzoquinone and BrAr(OH)
2
is brominted pyrocatechol.
The simulation was carried out by numerical integration of the set of differential equations
resulting from the application of the law of mass action to reactions K1 K14 with the rate
constants as listed in Table 1. The values of the rate constants for reactions K1 K3, K5, K8
have already been determined in the studies of the BZ reaction, and those of all other
reactions were either chosen from related work on the modified OKN mechanism by
Herbine and Field (Herbine & Field, 1980) or adjusted to give good agreement between
experimental results and simulations.


a
Herbine and Field 1980.
b
Adustable parameter chosen to give a good fit to data.
c
Not used
in the present model.
In this scheme, HAr(OH)
2
represents pyrocatechol compound containing two phenolic
groups, HAr(OH)O* is the radical obtained by hydrogen atom abstraction, HArO
2
is the
related quinone, BrAr(OH)
2
is the brominated derivative, and Ar
2
(OH)
4
is the coupling
product; HAr(OH) is phenol, HArO* is the hydrogen-atom abstracted radical, and Ar(OH)
2

is the product.
Table 1. OKN mechanism and rate constants used in the present simulation
Nonlinear Dynamics

122

Fig. 12. Numerical simulations of oscillations in (a) Br
-
(b) HBrO
2
, and (c) pyrocatechol
radical, HAr(OH)O*obtained from the present model K1 K14 by using the rate constants
listed in Table 1. The initilal concentraions were [BrO
3
-
]=0.08 M, [HAr(OH)
2
]=0.057 M,
[H
2
SO
4
]=1.4 M, and [Br
-
]=1.0 x 10
-10
M; the other initial concentrations were zero.
3.2 Numerical results
Figure 12 shows oscillations in three (Br
-
, HBrO
2
, and HAr(OH)O*) of the eleven variables
obtained in a simulation based on reactions K1 K14 and the rate constant values listed in
Table 1. The initial concentraions used in the simulation were [NaBrO
3
] = 0.08 M,
[HAr(OH)
2
] = 0.057 M, [H
2
SO
4
] = 1.4 M, and [Br
-
] = 1.0 x 10
-10
M with the other initial
concentrations to be zero with reference to those in the expreimental conditions as shown in
Fig. 1. Other four variables, BrO
2
*, Br*, HOBr, and Br
2
, exhibited oscillations, whereas the
rest variables, namely, BrO
3
-
, HAr(OH)
2
, HArO
2
, and BrAr(OH)
2
, did not exhibt oscillations
in the present simulation.
Figure 13 shows oscillations in [Br
-
] at different initial concentrations of bromate: (a) 0.08 M,
(b) 0.09 M, and (c) 0.1 M, with the same initial concentrations of [HAr(OH)
2
] = 0.057 M,
[H
2
SO
4
] = 1.4 M, and [Br
-
] = 1.0 x 10
-10
M with reference to the experimental conditions as
shown in Fig. 1. Although the concentration of bromate in the simulation is slightly smaller
than that in the experiments, the agreement between experimentally obtained redox
potential (Fig. 1) and simulated oscillations as shown in Figs. 12 and 13 is good. In
particular, the induction period and the period of oscillations are similar in magnitude, as
well as the degree of damping. The number of oscillations, and the prolonged period of
0 5000 10000 15000 20000
0
1x10
-9
2x10
-9
3x10
-9
4x10
-9
(a)
[
B
r
-
]

(
M
)
Tim e (s)
0 5000 10000 15000 20000
0
2x10
-9
4x10
-9
6x10
-9
8x10
-9
1x10
-8
(b)
[
H
B
r
O
2
]

(
M
)
Tim e (s)
0 5000 10000 15000 20000
0
5x10
-7
1x10
-6
2x10
-6
2x10
-6
(c)
[
H
A
r
(
O
H
)
O
*
]

(
M
)
Tim e (s)
Nonlinear Phenomena during the Oxidation and Bromination of Pyrocatechol

123


Fig. 13. Numerical simulations of the present model K1 K14 at different initial
concentrations of bromate: (a) 0.08 M, (b) 0.09 M, and (c) 0.1M. Other reaction conditions are
[HAr(OH)
2
] = 0.057 M, [H
2
SO
4
] = 1.4 M, and [Br
-
]=1.0 x 10
-10
M.
oscillations near the end of oscillations are also similar between experimental and simulated
results as shown in Fig. 1 (c), Fig.3 (c), Fig.12, and Fig.13. The above simulation not only
supports that the oscillatory phenomena seen in the batch system arises from intrinsic
dynamics, but also provides a tempelate for further understanding the mechanism of this
uncatalyzed bromate-pyrocatechol system.
While the above model is adequte in reproducing these spontaneous oscillations seen in
experiments, the concentration range over which oscillations could be achieved is somehow
different from what was determined in experiments. In the simulation, oscillatins were
obtained in the range of 0.02 M < [BrO
3
-
] < 0.1 M with [HAr(OH)
2
] = 0.057 M and [H
2
SO
4
] =
1.4 M in the present simuations, whereas no oscillation could be seen in experiments for the
condition of [BrO
3
-
] < 0.085 M. This discrepancy of range of the reactant concentrations for
exhibiting oscillations between experiments and simulations was also discerned for the
concentration of HAr(OH)
2
under the conditions [BrO
3
-
] = 0.085 M and [H
2
SO
4
] = 1.4 M:
Oscillations were exhibited in the range of 3 10
-4
M < [HAr(OH)
2
] < 0.3 M in the
simulation, whereas no oscillation could be observed in experiments under [HAr(OH)
2
] =
0.038 M as shown in Fig. 3 (a). The discrepancy in the suitable concentration range between
experiment and simulation may arise from two sources: (1) the currently employed model
may have skipped some of the unknown, but important reaction processes; (2) the rate
0 5000 10000 15000 20000
0
1x10
-9
2x10
-9
3x10
-9
4x10
-9
(a)
[
B
r
-
]

(
M
)
Tim e (s)
0 5000 10000 15000 20000
0
1x10
-9
2x10
-9
3x10
-9
4x10
-9
(b)
[
B
r
-
]

(
M
)
Tim e (s)
0 5000 10000 15000 20000
0
1x10
-9
2x10
-9
3x10
-9
4x10
-9
(c)
[
B
r
-
]

(
M
)
Tim e (s)
Nonlinear Dynamics

124
constants used in the calculation are too far away from their actual value. Note that those
values were original proposed for the phenol system (Herbine & Field, 1980). To shed light
on this issue, we have carefully adjusted the values of the adjustable rate constants in K4,
K6, K7, K9 K14, but so far no significant improvment was achhieved.
Two other sensitive properties that can help improve the modelling are the dependence of
the number of oscillations (N) and induction period (IP) on the reaction conditions. In
experiments, the N value increased monotonically from 4 to 15 as bromate concentration
was increased and then oscillatory behavior suddenly disappeared with the further increase
of bromate concentration. In contrast, in the simulation the number of oscillations decreased
gradually from 17 to 9 and then oscillatory behavior disappeared as the result of increasing
bromate concentration. On the positive side, IP values increased in both experiments and
simulations with respect to the increase of bromate concentration, i.e., from 9100 s to 11700 s
in the experiments, and from 8000 s to 9700 s in the simulations, respectively. We would like
to note that the simulated IP values firstly decreased from 12600 s to 7500 s with increase in
the initial concentration of bromate from 0.03 M to 0.06 M, then increased from 7600 s to
9700 s with increase in the bromate concentration from 0.07 M to 0.11 M.
3.3 Simplification of the model
In an attempt to catch the core of the above proposed model, we have examined the
influence of each individual step on the oscillatory behavior and found that reaction step
K12 in Scheme II is indispensable for oscillations under the present simulated conditions as
shown in Fig. 12. Such an observation is different from what has been suggested earlier
steps K1 to K10 would be sufficient to account for oscillations in the uncatalyzed bromate-
aromatic compounds oscillators (Orbn et al., 1979). For the Scheme II, our calculations
show that while setting one of the four rate constants k
11
to k
14
to zero; only when k
12
was set
to zero, no oscillation could be achieved. We further tested which reaction steps could be
eliminated by setting the rate constants to zero under the condition of k
12
0. The results are
as follows: (i) when three rate constants k
11
, k
13
, k
14
were simultaneously set to zero, no
oscillation was exhibited, (ii) when only one of the three rate constants was set to zero,
oscillation was observed in each case, and (iii) when two of the three rate constants were set
to zero, oscillations were exhibited under the condition of either k
13
0 (k
11
=k
14
=0) or k
14
0
(k
11
=k
13
=0) with the range of the rate constants as 3.0 10
3
< k
13
(M
-1
s
-1
) < 2.9 10
4
and 2.2
10
3
< k
14
(M
-1
s
-1
) < 6.0 10
4
, respectively. Thus our numerical investigation has concluded
that oscillations can be exhibited with minimal reaction steps as ten reaction steps in Scheme
I together with a combination of two reaction steps either K12 and K13 or K12 and K14 in
Scheme II.
Fig. 14 presents time series calculated under different combinations of reaction steps from
scheme II. This calculation result clearly illustrates that the oscillatory behavior is nearly
identical when the reaction step K11 was eliminated. Meanwhile, eliminating K13 or K14
seems to have the same influence on total number of oscillations (Fig.14 (c) ,(d)). However,
chemistry of the present reaction of aromatic compounds suggests that both reaction K13 and
K14 are equally important (Orbn et al., 1979). The equilibrium of step K13 is well
precedented, and equimolar mixtures of quinone and dihydroxybenzene are intensely colored,
and the radical HAr(OH)O* may be responsible for the color changes observed during
oscillations (Orbn et al., 1979). In addition, step K14 is said to explain the observed coupling
products and to prevent the buildup of quinone for further oscillations (Orbn et al., 1979).
Nonlinear Phenomena during the Oxidation and Bromination of Pyrocatechol

125

Fig. 14. Numerical simulations of the present model of K1 K10 with different reaction steps
in Scheme II: (a) K11 K14, (b) K12 K14, (c) K12 and K13, and (d) K12 and K14. The initilal
concentraions were [BrO
3
-
] = 0.08 M, [HAr(OH)
2
] = 0.057 M, [H
2
SO
4
] = 1.4 M, and [Br
-
] = 1.0
x 10
-10
M as shown in Fig. 10. Note that the scales of x and y axes are different from those in
Fig. 12.
In our numerical simulation, when we eliminated either step K13 or step K14, the simulated
numerical results such as (i) the time series of oscillations, (ii) the initial concentration range
of BrO
3
-
, H
2
SO
4
, and HAr(OH)
2
for oscillations, and (iii) the dependence of the number of
oscillations and induction period on the initial concentration of BrO
3
-
became significantly
different from those in experiments. In particular, the number of oscillations are too large
under the above conditions as shown in Figs.14 (c) and (d). Such observation suggests that
both K13 and K14 are important in the system studied here.
Consequently, we have concluded that the simplified model should include reaction steps
K1 to K10 in Scheme I, and K12, K13, and K14 in Scheme II to reproduce the experimental
results qualitatively.
3.4 Influence of reaction step K11 on the equilibrium of step K13
The numerical investigation presented in Fig. 14b suggests that reaction step K11 is not
necessary for qualitatively reproducing the experimental oscillations. Besides, more positive
reason for eliminatiing step K11 from the present model is that step K11 affects the range of
rate constant of the equilibrium step K13 significantly. The equilibrium must lie well to the
left (Orbn et al., 1979), i.e., the rate constant k
r13
to the left must be much larger than that k
13

0 10000 20000 30000
0
2x10
-9
4x10
-9
6x10
-9
8x10
-9
1x10
-8
(a)
[
B
r
-
]

(
M
)
Tim e (s)
0 10000 20000 30000
0
2x10
-9
4x10
-9
6x10
-9
8x10
-9
1x10
-8
(c)
[
B
r
-
]

(
M
)
Tim e (s)
0 10000 20000 30000
0
2x10
-9
4x10
-9
6x10
-9
8x10
-9
1x10
-8
(d)
[
B
r
-
]

(
M
)
Tim e (s)
0 10000 20000 30000
0
2x10
-9
4x10
-9
6x10
-9
8x10
-9
1x10
-8
(b)
[
B
r
-
]

(
M
)
Tim e (s)
Nonlinear Dynamics

126
to the right. However, when we included step K11 in the model, we found no upper limit of
the rate constant to the right; for instance, the rate constant can be more than 1.0 10
9
for the
system to exhibit scillations under the conditions as shown in Fig.10. This value is already
too large for the rate constant to the right, because we set the rate constant to the left to be
3.0 10
4
in the present simulations.
On the other hand, if we eliminated step K11 from the modelling, the range of the rate
constant to the right was 0.007 < k
13
(M
-1
s
-1
) < 0.03 for the system to exhibit oscillations,
which seems to be reasonable for the equilibrium reaction step K13 to lie well to the left.
Thus, this numerical analysis suggests that reaction step K11 should be eliminated from the
present model.
4. Conclusions
This chapter reviewed recent studies on the nonlinear dynamics in the bromate-
pyrocatechol reaction (Harati & Wang, 2008a and 2008b), which showed that spontaneous
oscillations could be obtained under broad range of reaction conditions. However, when the
concentration of bromate, the oxidant in this chemical oscillator, is fixed, the concentration
of pyrocatechol within which the system could exhibit spontaneous oscillations is quite
narrow. This accounts for the reason why earlier attempt of finding spontaneous oscillations
in the bromate-pyrocatechol system had failed. As illustrated by phase diagrams in the
concentration space, it is critical to keep the ratio of bromate/pyrocatechol within a proper
range. From the viewpoint of nonlinear dynamics, bromate is a parameter which has a
positive impact on the nonlinear feedback loop, where increasing bromate concentration
enhances the autocatalytic cycle (i.e. nonlinear feedback). On the other hand, pyrocatechol
involves in the production of bromide ions, a reagent which inhibits the autocatalytic
process, where an increase of pyrocatechol concentration accelerates the production of
bromide ions through reacting with such reagents as bromine molecules. The requirement
of having a proper ratio of bromate/pyrocatechol reflects the need of having a balanced
interaction between the activation cycle and inhibition process for the onset of oscillatory
behaviour in this chemical system. If the above conclusion is rational, one can expect that
the role that pyrocatechol reacts with bromine dioxide radicals to accomplish the
autocatalytic cycle is less important than its involvement in bromide production in this
uncatalyzed bromate oscillator, and therefore when a reagent such as metal catalyst is used
to replace pyrocatechol to react with bromine dioxide radicals for completing the
autocatalytic cycle, oscillations are still expected to be achievable. This is indeed the case.
Experiments have shown spontaneous oscillations when cerium, ferroin or manganese ions
were introduced into the bromate-pyrocatechol system.
Numerical simulations performed in this research show that the observed oscillatory
phenomena could be qualitatively reproduced with a generic model proposed for non-
catalyzed bromate oscillators. The simulation further indicates that while either two reaction
steps K12 and K13 or K12 and K14 together with ten steps K1 K10 in Scheme I in the OKN
mechanism are sufficient to qualitatively reproduce oscillations, three steps K12, K13, and
K14 with ten steps K1 K10 are more realistic for representing the chemistry involving the
oscillatory reactions, and also for reproducing oscillatory behaviors observed
experimentally. The ratio of the rate constants for the equilibrium reaction K13 was a key
reference to eliminate reaction step K11 from the original model. Although the present
model still needs to be improved to reproduce the experimental results quantitatively, it has
Nonlinear Phenomena during the Oxidation and Bromination of Pyrocatechol

127
given us a glimpse that the autocatalytic production of bromous acid could be modulated
periodically even in the absence of a bromide ion precursor such as bromomalonic acid in
the BZ reaction. Understanding the reproduction of bromide ion appears to be a key for
deciphering the oscillatory mechanism for the family of uncatalyzed oscillatory reactions of
substituted-aromatic compounds with bromate and should be given particularly attention in
the future research.
5. Acknowledgements
This material is based on work supported by Natural Science and Engineering Research
Council (NSERC), Canada, and Canada Foundation for Innovation (CFI). JW is grateful for
an invitation fellowship from Japan Society for the Promotion of Science (JSPS).
6. References
Adamkov, L.; Farbulov, Z. & evk, P. (2001) New J. Chem. Vol. 25, 487-490.
Amemiya, T.; Kdr, S.; Kettunen, P. & Showalter K. (1996). Phys. Rev. Lett. Vol. 77, 3244-
3247.
Amemiya, T.; Yamamoto, T.; Ohmori, T. & Yamaguchi, T. (2002) J. Phys. Chem. A Vol. 106,
612-620.
Ball P. (2001) The Self-Made Tapestry: Pattern Formation in Nature, Oxford University Press,
ISBN-10: 0198502435.
Carlsson, P.; Zhdanov, V. P. & Skoglundh, M. (2006) Phys. Chem. Chem. Phys. Vol. 8, 2703
2706.
Chiu, A. W. L.; Jahromi, S. S.; Khosravani, H.; Carlen, L. P. & Bardakjian, L. B. (2006) J.
Neural Eng. Vol. 3, 9-20.
Dhanarajan, A. P.; Misra, G. P. & Siegel, R. A. (2002) J. Phys. Chem. A Vol. 106, 8835-8838.
Dutt, A. K. & Menzinger, M. (1999) J. Chem. Phys. Vol. 110, 7591-7593.
Epstein, I. R. (1989). J. Chem. Edu. (1989) Vol. 66, 191-195.
Epstein, I. R. & Pojman, J. A. (1998) An Introduction to Nonlinear Chemical Dynamics, Oxford
University Press, ISBN10: 0-19-509670-3, Oxford.
Farage, V. J. & Janjic, D. (1982) Chem. Phys. Lett. Vol. 88. 301-304.
Field, R. J. & Burger, M. (1985) (Eds.), Oscillations and Traveling Waves in Chemical Systems,
Wiley-Interscience, ISBN-10: 0471893846, New York.
Goldbeter, A. (1996). Biochemical Oscillations and Cellular Rhythms, Cambridge University
Press, ISBN 0-521-59946-6, Cambridge.
Gyrgi, L. & Field, R. J. (1992) Nature Vol. 355, 808-810.
Harati, M. & Wang, J. (2008a) J. Phys. Chem. A Vol. 112, 4241-4245.
Harati, M. & Wang, J. (2008b) Z. Phys. Chem. A Vol. 222, 997-1011.
Herbine, P. & Field, R. J. (1980) J. Phys. Chem. Vol. 84, 1330-1333.
Horvth, J.; Szalai, I. & De Kepper, P. (2009) Science Vol. 324, 772-775.
Jahnke, W.; Henze C. & Winfree, A. T. (1988) Nature Vol. 336, 662-665.
Kdr, S.; Wang, J. & Showalter, K. (1998) Nature Vol. 391, 770-743.
Krs, E. & Orbn, M. (1978) Nature Vol. 273, 371-372.
Kumli, P. I.; Burger, M.; Hauser, M. J. B.; Mller, S. C. & Nagy-Ungvarai, Z. (2003) Phys.
Chem. Chem. Phys. Vol. 5, 5454-5458.
Kurin-Csrgei, K.; Epstein, I. R. & Orbn, M. (2004) J. Phys. Chem. B Vol. 108, 7352-7358.
Nonlinear Dynamics

128
McIIwaine, M.; Kovacs, K.; Scott, S. K. & Taylor, A. F. (2006) Chem. Phys. Lett. Vol. 417, 39-42.
Mori, Y.; Nakamichi Y.; Sekiguchi, T.; Okazaki, N.; Matsumura T. & Hanazaki, I. (1993)
Chem. Phys. Lett. Vol. 211, 421-424.
Morowitz, H. J. (2002), The Emergence of Everything: How the World Became Complex, Oxford
University Press, ISBN-13 978-0195135138, Oxford.
Nicolis, G. & Prigogine, I. (1977) Self-Organization in Non-Equilibrium Systems, Wiley, ISBN 10
- 0471024015.
Nicolis, G. & Prigogine, I. (1989) Exploring Complexity, FREEMAN, ISBN 0-7167-1859-6, New
York.
Orbn, M. & Krs, E. (1978a) J. Phys. Chem. Vol. 82, 1672-1674.
Orbn, M. & Krs, E. (1978b) React. Kinet. Catal. Lett. Vol. 8, 273-276.
Orbn, M.; Krs, E. & Noyes, R. M. (1979) J. Phys. Chem. Vol. 83, 3056-3057.
Sagues, F. & Epstein, I. R. (2003) Nonlinear Chemical Dynamics, Dalton Trans., 1201-1217.
Scott Kelso J. A. (1995), Dynamic Patterns: The self-organization of brain and behavior, The MIT
Press, ISBN-10: 0262611317, Cambridge, MA.
Scott, S. K. (1994) Chemical Chaos, Oxford University Press, ISBN 0-19-8556658-6, Oxford.
Smoes, M-L. J. Chem. Phys. (1979) Vol. 71, 4669-4679.
Srensen, P. G.; Hynne, F. & Nielsen, K. (1990) React. Kinet. Catal. Lett. Vol. 42, 309-315.
Steinbock, O.; Kettunen, P. & Showalter K. (1995) Science Vol. 269, 1857-1860.
Straube, R.; Flockerzi, D.; Mller, S. C. & Hauser, M. J. B. (2005) Phys. Rev. E. Vol. 72, 066205-
1 - 12.
Straube, R.; Mller, S. C. & Hauser, M. J. B. (2003) Z. Phys. Chem. Vol. 217, 1427-1442.
Szalai, I. & Krs, E. (1998) J. Phys. Chem. A Vol. 102, 6892-6897.
Yamaguchi, T.; Kuhnert, L.; Nagy-Ungvarai, Zs.; Mller, S. C. & Hess, B. (1991) J. Phys.
Chem. Vol. 95, 5831-5837.
Vanag, V. K.; Mguez, D. G. & Epstein, I. R. (2006) J. Chem. Phys. Vol. 125, 194515:1-12.
Wang, J.; Hynne, F.; Srensen, P. G. & Nielsen K. (1996) J. Phys. Chem. Vol. 100, 17593-17598.
Wang, J.; Srensen, P. G. & Hynne, F. (1995) Z. Phys. Chem. Vol. 192, 63-76.
Wang, J.; Yadav, Y.; Zhao, B.; Gao, Q. & Huh, D. (2004) J. Chem. Phys. Vol. 121, 10138-10144.
Welsh, B. J.; Gomatam, J. & Burgess, A. E. Nature Vol. 304, 611-614.
Winfree, A. T. (1972) Science Vol. 175, 634-636.
Winfree, A. T. (1987) When Time Breaks Down, Princeton University Press, ISBN 0-691-02402-
2, Princeton.
Witkowski, F. X.; Leon, L. J.; Penkoske, P. A.; Giles, W. R.; Spanol, M. L.; Ditto, W. L. &
Winfree, A. T. (1998) Nature Vol 392, 78-82.
Zaikin, A. N. & Zhabotinsky, A. M. (1970) Nature Vol. 225, 535-537.
Zhao, J.; Chen, Y. & Wang, J. (2005) J. Chem. Phys. Vol. 122, 114514:1-7.
Zhao, B. & Wang, J. (2006) Chem. Phys. Lett. Vol. 430, 41-44.
Zhao, B. & Wang, J. (2007) J. Photochem. Photobiol: Chemistry, Vol. 192, 204-210.
6
Dynamics and Control of Nonlinear Variable
Order Oscillators
Gerardo Diaz and Carlos F.M. Coimbra
University of California, Merced
U.S.A.
1. Introduction
The denomination Fractional Order Calculus has been widely used to describe the
mathematical analysis of differentiation and integration to an arbitrary non-integer order,
including irrational and complex orders. First proposed around three hundred years ago, it
has attracted much interest during the past three decades (Oldham & Spanier (1974), Miller
& Ross (1993), Podlubni (1999)). The increased interest in fractional systems in the past few
decades is due mainly to a large body of physical evidence describing fractional order
behavior in diverse areas such as fluid mechanics, mechanical systems, rheology,
electromagnetism, quantitative finances, electrochemistry, and biology. Fractional order
modeling provides exceptional capabilities for analysing memory-intense and delay systems
and it has been associated with the exact description of complex transport phenomena such
as fractional history effects in the unsteady viscous motion of small particles in suspension
(Coimbra et al. 2004, LEsperance et al. 2005). Although fractional order dynamical and
control systems were studied only marginally until a few decades ago, the recent
development of effective mathematical methods of integration of non-integer order
differential equations (Charef et al. (1992); Coimbra & Kobayashi (2002), Diethelm et al.
(2002); Momany (2006), Diethelm et al. (2005)) has resulted in a number of control schemes
and algorithms, many of which have shown better performance and disturbance rejection
compared to other traditional integer-order controllers (Podlubni (1999); Hartly & Lorenzo
(2002), Ladaci & Charef (2006), among others).
Variable order (VO) systems constitute a generalization of fractional order representations
to functional order. In VO systems the order of the derivative changes with respect to either
the dependent or the independent variables (or both), or parametrically with respect to an
external functional behavior (Samko & Ross, 1993). Compared to fractional order
applications, VO systems have not received much attention, although the potential to
characterize complex behavior by the functional order of differentiation or integration is
clear. Variable order formulations have been utilized, among other applications, to describe
the mechanics of an oscillating mass subjected to a variable viscoelasticity damper and a
linear spring (Coimbra, 2003), to analyze elastoplastic indentation problems (Ingman &
Suzdalnitsky (2004)), to interpolate the behavior of systems with multiple fractional terms
(Soon et al., 2005), and to develop a statistical mechanics model that yields a macroscopic
constitutive relation for a viscoelastic composite material undergoing compression at
varying strain rates (Ramirez & Coimbra, 2007). Concerning the dynamics and control of VO
Nonlinear Dynamics

130
systems, the authors of this chapter have previously analyzed the dynamics and linear
control of a variable viscoelasticity oscillator and have presented a generalization of the van
der Pol equation using the VO differential equation formulation (Diaz & Coimbra, 2009).
In the present work, we utilize the Coimbra Variable Order Differential Operator (VODOs)
to analyze the dynamics of the Duffing equation with a VO damping term. Coimbras
VODO returns the correct value of the p-th derivative for p < 2, as can be generalized to any
order, positive or negative.The behavior of the variable order differintegrals are shown in
variable phase space for different parameters that constitute a pictorial representation of the
dynamics of the variable order system, and help understand the transitional regimes
between the extreme values of the derivatives. Also, a tracking controller is developed and
applied to the oscillator for different expressions of the variable order q(x(t)). Finally, a
variable order controller is used to eliminate chaotic oscillations of Lorenz-type systems.
2. Fractional and variable order operators
Over the past few centuries, different definitions of a fractional operator have been
proposed. For instance the Riemann-Liouville integral is defined as


D
0, t

x(t) =
1
()
(t )
1
0
t

x( )d (1)
where R
+
is the order of integration of the function x(t) when the lower limit of
integration (initial condition) is chosen to be identically zero. The Riemann-Liouville
derivative of order is given as


D
0,t

x(t) =
1
(m)
d
m
dt
m
(t )
m1
0
t

x( )d , (2)
and the Grundwald-Letnikov differential operation is defined as

( )
0,
0,
0
( ) lim ( 1) ( )
n
p k
t k
h nh t
k
D x t h x t kh

=
=
=

. (3)
Finally, the Caputo derivative of fractional order of x(t) is defined as


D
0,t

x(t) =
1
(m)
(t )
m1
0
t

x
(m)
( )d , (4)
for which m-1 < <m Z
+
. More details about these operators can be found in Li & Deng
(2007), Diethelm (2002), and Hartley & Lorenzo (2002).
For variable order systems, Coimbra (2003) defined the canonical differential operator as:


D
q(x(t))
x(t) =
1
(1q(x(t)))
(t )
q (x(t))
D
1
0+
t

x()d +
(x(0
+
) x(0

))t
q (x(t))
(1q(x(t)))
(5)
where q(x(t)) < 1. The constraint on the upper limit of differentiation can be easily removed,
and is adopted here only for convenience. One of the important characteristics of Coimbras
Dynamics and Control of Nonlinear Variable Order Oscillators

131
operator is that it is dynamically consistent with causal behavior in the initial conditions, i.e.
the operator returns the appropriate Heaviside contribution to the integral value of
D
q(x(t))
x(t) when x(t) is not continuous between t=0
-
and t=0
+
(Coimbra, 2003; Ramirez &
Coimbra, 2007; Diaz & Coimbra (2009)). Also of relevance is that all integer and fractional
order differentials are returned correctly by the operator, including the upper limit. In this
work we used the extended version of this operator that covers the range of q(x(t))<2. The
generalized order differential operator can thus be calculated by the following numerical
algorithm:


D
q
x
n
=
1
(4 q)
a
i ,n
D
2
i=0
n

x
i
+
x(0
+
)(1q)(t
n
)
q
+D
1
x(0
+
)t
n
1q
(2q)
, (6)
with quadrature weights given by


a
i,n
= (3q)n
2q
n
3q
+ (n1)
3q
, if i=0

a
i,n
= (ni 1)n
3q
2(ni)
3q
+ (ni +1)
3q
, if 0<i<n.


a
i,n
= 1 , if i=n.
As stated earlier, one of the critical properties of this operator for generalized order
modeling is that it returns the p-th derivative of x(t) when q(x(t)) = p. This can be graphically
demonstrated by considering an arbitrary function with known derivatives such as


y = t
2
sin(t) (7)

Fig. 1. Comparison of values of function y=t
2
sin(t) and its derivatives with the results
obtained with operator described by Eq. (6) for several values of the order q.
Nonlinear Dynamics

132
Figure 1 shows the values of function y (Eq. 7) and its derivatives dy/dt, and d
2
y/dt
2

calculated analytically. The figure also shows that the operator described by Eq. (6) returns
values that match the functions y for q=0, dy/dt for q=1, and d
2
y/dt
2
for q2, respectively. The
values of q=0.5 and q=1.5 are also shown to indicate the matching of the rational order
derivatives with the values calculated using the VO operator.
3. Dynamics of the Duffing equation with variable order damping
Together with the van de Pol equation, the Duffing equation represents the behavior of one
of the most studied oscillators in the field of nonlinear dynamics (Guckenheimer & Holmes
(1983), Drazin (1994)). First introduced in 1918 by G. Duffing, different variations of the
equation have been used to analyze its dynamics for the automomous and forced cases.
Moon and Holmes (1979, 1980) considered a negative linear stiffness term to analyse the
forced vibrations of a cantilever beam near two magnets. Vincent & Kenfack (2008) recently
studied the bifurcation structure and synchronization of a double-well Duffing oscillator.
They were able to show regions of chaos and quasiperiodicity and they found threshold
parameters for which synchronization occured. With respect to fractional order systems,
Sheu et al. (2007) analyzed the Duffing equation with negative linear stiffness and a
fractional damping term. They reported a period doubling route to chaos in their study.
3.1 Forced oscillations
We generalize the concept of fractional damping to include a variable order term as:


D
2
x +D
q
x x +x
3
= sin(t) . (8)
The main difference with respect to the work by Sheu et al. (2007) is that they studied the
dynamics of Eq. (8) for a range of values of the fractional order q where this parameter was
kept constant for every case analyzed. Here, the oscillator is generalized to include a
damping term where the order of the derivative reacts to the effect of the forcing function
over time, thus q = q(t). In our analysis, we choose the value of parameters and to be 0.1
and 2, respectively.
Case = 1.5:
The first case considered in this work relates to the behavior of the oscillator given by Eq. (8)
for = 1.5 for two different conditions, i.e. q = 1 and q = (99/100) + sin( t). We note that the
operator described by Eq. 6 is valid for q(t) < 2, thus the expression used for the change in q
with respect to time ensures that this condition is met.
Figure 2 shows the dynamics of the oscillator given by Eq. (8) for q = 1 as the order of the
derivative in the damping term. The simulations cover the time range t [0, 700] where only
the results for t > 200 are plotted to exclude the initial transients. Chaotic behavior is observed
and a strange attractor is depicted in Fig. 2(a). The Poincar map is shown in Fig. 2(b).
The effect of the variable order derivative on the damping term of Eq. (8) significantly
changes the dynamics of the oscillator. This can be observed in Figs. 3(a) and 3(b) where it is
seen that after removing the intial transients, the dynamics of the oscillators are confined to
a narrower region in the phase space.
The dynamics of the VO oscillators can also be analyzed utilizing a modified version of the
phase diagram where the variable order derivative, D
q
x(t), is plotted on the ordinate axis
Dynamics and Control of Nonlinear Variable Order Oscillators

133
and the position, x(t), is plotted on the abcisa axis. Figure 4(a) shows the variable order
phase space (a plot of the value of the VO derivative, D
q
x(t), as a function position),
whereas Fig. 4(b) shows the behavior of D
q
x(t) as a function of the order of the derivative,
q(t). It is seen in Fig. 4(b) that q(t) < 2, thus meeting the upper limit of differentiation
mandated by the numerical algorithm used here (Eq. 6).


Fig. 2. Phase diagram and Poincare map for = 1.5 and q =1.

Fig. 3. Phase diagram and Poincare map for = 1.5 and q =(99/100)+ sin( t).
Nonlinear Dynamics

134
Figure 5(a) shows the change of x(t) and D
q
x(t) as a function of time. Figures 6(a) and 6(b)
show that q(t) also has an oscillatory behavior with D
q
x(t) having a minimum value when
x(t) and q(t) approach their maximum value. This is also depicted in the VO phase diagrams
shown in Figs. 4(a) and 4(b).

Fig. 4. Modified phase diagram and D
q
x(t) vs. q(t) plots for =1.5.

Fig. 5. Dynamics of VO Duffing equation with respect to time for =1.5. (a) - - - x(t), ____ =
D
q
x(t); (b) - - - q(t), ____ = D
q
x(t);
Dynamics and Control of Nonlinear Variable Order Oscillators

135

Fig. 6. Phase diagram and Poincare map for =0.5 and q=1.
Case = 0.5:
We now analyze the case where parameter = 0.5. After the initial transient, the standard
configuration (q = 1) shows an oscillatory behavior as depicted in Fig. 6(a) with a single
point appearing in the Poincare map, Fig. 6(b).

Fig. 7. Phase diagram and Poincare map for = 0.5 and q = (99/100)+ sin( t) for t > 200.
Nonlinear Dynamics

136
Figures 7(a) and 7(b) show the results of the simulations for = 0.5 and a variable order of
the derivative given by q(t) = (99/100) + sin( t). It is seen that the phase diagram and
Poincare maps differ significantly from the case q = 1. However, plotting x(t) as a function
of time, as depicted in Fig. 8, shows the transient effects seem to last longer than for the case
of q = 1. After t ~ 400, the system settles to an oscillatory behavior with a smaller amplitude.

Fig. 8. Phase diagram and Poincare map for = 0.5 and q = (99/100) + sin( t) for t > 200.

Fig. 9. Phase diagram and Poincare map for = 0.5 and q = (99/100)+ sin( t) for t > 400.
Dynamics and Control of Nonlinear Variable Order Oscillators

137
Plots of the phase diagram and the Poincare map for t > 400 are shown in Figs. 9(a) and 9(b),
respectively. Similar dynamics compared to q = 1 are displayed by the system.
3.2 Control of the VO Duffing equation
The dynamics of the variable order Duffing equations were analyzed in the previous section
for the cases = 0.1, = 2, with = 1.5 and = 0.5, respectively. In this section, we study
controls aspects of this equation subject to a VO damping term. An exact feedback
linearization is performed to obtain a tracking controller that drives the VO Duffing
oscillator to follow a periodic reference function, r (Khalil, 1996). The forcing function in Eq.
(8) can be replaced by a control action as shown by Eq. 9.


D
2
x = x x
3
D
q
x + u. (9)
Exact feedback linearization is obtained by choosing the control action


u = x
3
+D
q
x +v. (10)
Thus, Eq. 9 is converted to a linear equation of the form


D
2
x = x +v. (11)
This second order differential equation is transformed to a system of first order differential
equations

1 1
2 2
0 1 0
,
1 0 1
[1 0] .
x x
Ax Bv v
x x
y Cx x

= + = +


= =

(12)

A control action of form

u = K

x +Gr = k
1
x
1
k
2
x
2
+Gr is chosen where k
1
and k
2
are
constants that are used to select the location of the closed-loop eigenvalues, G is the
feedforward gain, and r is the reference. For the controllable system given by Eq. (12) we
arbitrarily select closed-loop egivenvalues
1,2
=-5 to obtain k
1
= 24 and k
2
= 10. The
feedforward gain is obained with Eq. (13) (Williams & Lawrence, 2007).


G = (C(ABK)
1
B)
1
. (13)

The tracking scheme is tested with the variable order derivative in the VO damping term
having the expression q = (99/100) + sin( t), where = 1.5 and = 2. Figures 10(a) to 10(d)
show the behavior of the tracking system for r(t) = 2 cos(/10) + sin(3/10). The ouput of
the system, y(t), follows the reference, r(t), consistently, as seen in Fig. 10(a). Figure 10(b)
shows the control action, u(t), and the sinusoidal behavior of the order of the VO derivative,
q(t), is shown in Fig. 10(c) where the value of the variable order derivative, D
q
y(t), is plotted
in Fig. 10(d).
Exact feedback linearization can be used for different functions of q(t). Figure 11(a) to 11(d)
show the tracking of reference r for q(t) = r(t)/3. Scaling of q(t) with respect to r is performed
so that the value of q(t) remains smaller than 2.
Nonlinear Dynamics

138




Fig. 10. Tracking control for the VO duffing equation for q(t)= (99/100)+ sin(t). (a) __ = r(t),
. . .=y(t); (b) u(t), (c) q(t), and (d) D
q
y(t).

We note that if the value of the order of the VO derivative, q(t), is known to remain within
the requirement of the operator (i.e. q(t)< 2) then an implicit form of the variation of q (i.e.
q=q(x)) can also be utilized (Diaz & Coimbra, 2009). It is also mentioned that if the closed-
loop eigenvalues are chosen to have positive real parts then the system becomes unstable.
4. VO control of the Lorenz system
So far, we have analyzed the dynamics and control of VO systems that have the term D
q
x(t)
as part of the expression describing their dynamics. We now apply the variable order
approach as the control action to stabilize a chaotic dynamical system. First proposed as a
way to discribe the dynamics of weather systems, the Lorenz system of equations (Lorenz,
1963) has been intensively studied as a dynamical system that displays chaotic behavior
where a strange attractor is encountered under certain values of its parameters. Control
techniques have been proposed in the past (Vincent & Yu, 1991) but to the best knowledge
of the authors, there is no study in the literature that has utilized a variable order controller
to stabilize the chaotic dynamics of the Lorenz system.
Dynamics and Control of Nonlinear Variable Order Oscillators

139

Fig. 11. Tracking control for the VO duffing equation for q(t)=(1/3) [2cos(/10)+sin(3/10)].
(a) __ = r(t), . . .= y(t); (b) u(t), (c) q(t), and (d) D
q
y(t).

The Lorenz system is described by the folowing equations

1
1 2
2
1 2 2 3
3
1 2 3
,
,
.
dx
x x
dt
dx
rx x x x
dt
dx
x x bx u
dt
= +
=
= +
(14)
For r > 1 there are two non-trivial equilibrium points, i.e.

x
1
= x
2
= (b (r 1))
1/2
, x
3
=r 1.
Linearizing the system with respect to the first non-trivial equilibrium point, we obtain


dz
1
dt
= z
1
+z
2
,
dz
2
dt
= z
1
z
2
b(r 1)z
3
,
dz
3
dt
= b(r 1)z
1
+ b(r 1)z
2
bz
3
+ u*,
(15)
Nonlinear Dynamics

140
which can be written as
dz
dt
= Az +Bu * , where


z
1
= x
1
b(r 1) ,
z
2
= x
2
b(r 1) ,
z
3
= x
3
(r 1).
(16)
Tavazoei et al. (2009) developed a control strategy using a fractional order controller with
three parameters that is used to suppress chaos. They showed that a chaotic system is
stabilized using the single control input u(t)=J
q
y(t), where J
q
is a fractional integral operator
and y(t) = -(T
1
+T
3
)(x(t)-x*), and where T
1
and T
3
are the first and third row of a
transformation matrix such that

1
0 1 0 0
0 0 1 , 0 .
1
A TAT B TB
a b c

= = = =




(17)
where the parameters a,b,c are the coefficients of the characteristic polynomial of the
Jacobian matrix A

3 2
0. s cs bs a + + + = (18)
Tavazoei et al. (2009) also showed that for the integral fractional operator with -1 < q < 0 the
controller stabilizes the system when

(1 /2) ( 1 /2)
0 ; .
cos( /2) cos( /2)
q q
cb ab
q q



< < >

(19)
We use the VO operator described by Eq. (5) with a negative value of q (i.e. integral variable
order operator) to suppress chaos of the Lorenz system. Choosing = 10, b = 8/3, r = 28 and
q =-0.2, we obtain 0 < < 2310.9 and > 23.7. Arbitrary values of = 23.1 and = 26.1 are
chosen that satisfy the constraints given by Eq. (19). Figure 12(a) depicts the chaotic
behaviour displayed by the Lorenz system for t < 25. At t = 25, the controller is turned on
and the system is stabilized around the selected equilibrium point. Figure 12(b) shows the
values of the control action, u(t). In this case q has been considered constant for the VO
operator.
The variable order capability of the controller can be verified by running a similar case
where the parameters and are kept constant and the order of the VO derivative is
changed. The controller works until the constraints given by Eq. (19) are no longer met.
Fixed values for and are used. However, for t > 25 the order of the VO derivative q(t) is
monotonically decreased starting from q = -0.2. Figure 13(a) shows the behaviour of the
system subject to the control action u(t) shown in Fig. 13(b). It is observed that once the
controller is turned on (t > 25) stabilization of the chaotic system is obtained for variable q
until parameters and fall outside of the constraints. Figure 13(c) shows the variation of q
over time. The controller reaches a point where it no longer stabilizes the chaotic behaviour
of the system. This situation is resolved by re-calculating the values of and for the VO
Dynamics and Control of Nonlinear Variable Order Oscillators

141

Fig. 12. Chaos suppression in the Lorenz system with = 10, b = 8/3, r = 28, q =-0.2, and
fixed values of and in VO operator in Eq. (5). (a) x, y, z vs t (b) u vs t.

Fig. 13. Performance of controllers for fixed values of and and decreasing value of q(t).
(a) x, y, z vs t (b) u vs t, (c) q vs t.
Nonlinear Dynamics

142
value of q to remain within the required constraints. Figure 14(a) shows that the controller
stabilizes the chaotic system under the variation of q with respect to time shown in Fig. 14(c)
that generates the control action displayed in Fig. 14(b). The variation in the values of and
is observed in Fig. 14(d) that shows that as q decreases the values of and also increase
rapidly.

Fig. 14. Performance of controllers for variable values of and and decreasing value of
q(t). (a) x, y, z vs t (b) u vs t, (c) q vs t, (d) , vs t.
Grigorenko and Grigorenko (2003) have shown that the generalized fractional order Lorenz
system also presents chaotic behaviour. Clearly, a VO controller technique as presented here
can also be utilized to suppress chaos in such a system.
5. Conclusion
Variable order systems, i.e. systems where the order of the derivative changes with respect
to either the dependent or the independent variables have not received as much attention as
fractional order systems, despite of the ability of variable order formulations to model
continuous spectral behavior in complex dynamics. We illustrate some of the characteristics
of variable order systems and controllers through the numerical simulation of nonlinear
dynamic oscillators and systems of equations. In this work, we analyze the dynamics of a
modified Duffing equation, which includes a variable order derivative as the damping term,
Dynamics and Control of Nonlinear Variable Order Oscillators

143
and illustrate its behavior as compared to the classical Duffing equation. Exact feedback
linearization is used to derive a linear controller of the Duffing equation with variable order
damping. Finally, a variable order controller is used to suppress chaos on the Lorenz system
of equations. To the best knowledge of the authors, this is the first time a variable order
controller is described.
6. References
Charef, A.; Sun, H.H.; Tsao, Y.Y. & Onaral, B. (1992) Fractal system as represented by
singularity function, IEEE Transactions on Automatic Control, 37(9) 14651470.
Coimbra, C.F.M & Kobayashi, M.H. (2002). On the Viscous Motion of a Small Particle in a
Rotating Cylinder. Journal of Fluid Mechanics (469) pp. 257-286.
Coimbra, C.F.M. (2003) Mechanics with variable-order differential operators, Annalen der
Physik, 12(11-12) 692703.
Coimbra, C.F.M.; LEsperance, D.; Lambert, A., Trolinger, J.D. & Rangel, R.H., (2004) An
experimental study on the history effects in high-frequency Stokes ows, Journal of
Fluid Mechanics (504) 353363.
Diaz, G. & Coimbra, C.F.M. (2009) Nonlinear dynamics and control of a variable order
oscillator with application to the van der Pol equation, Nonlinear Dynamics, 56:
145-157.
Diethelm, K.; N. J. Ford, N.J. & Freed, A.D. (2002) A predictor-corrector approach for the
numerical solution of fractional differential equations, Nonlinear Dynamics,
29(2002) 322.
Diethelm, K.; Ford, N.J.; Freed, A.D.& Luchko, Y. (2005) Algorithms for the fractional
calculus: A selection of numerical methods, Computational Methods in Applied
Mechanics and Engineering, 194 (6-8) 743-773.
Drazin, P.G. (1994) Nonlinear Systems, Cambridge Texts in Applied Mathematics,
Cambridge University Press, UK.
Grigorenko, I & Grigorenko, E. (2003) Chaotic dynamics of the fractional Lorenz system.
Physical Review Letters 91(3) 034101-1-0.4101-4.
Guckenheimer, J. & Holmes, P. (1983) Nonlinear Oscillators, Dynamical Systems, and
Bifurcations of Vector Fields, Applied Mathematical Sciences 42, Spriner-Verlag,
New York
Hartley, T.T. & Lorenzo, C.F. , Dynamics and control of initialized fractional-order systems,
Nonlinear Dynamics, 29(2002) 201233.
Ingman, D. & Suzdalnitsky, J., Control of damping oscillations by fractional differential
operator with time-dependent order, Computer Methods in Applied Mechanics
and Engineering, 193(2004), 55855595.
Khalil, H.K. (1996) Nonlinear Systems. Prentice Hall, 2
nd
Ed, USA. ISBN 0-13-228024-8.
Ladaci, S. and Charef, A. (2006), On fractional adaptive control, Nonlinear Dynamics, 43365
378.
LEsperance, D.; Coimbra, C.F.M.; Trolinger, J.D. & Rangel, R.H. (2005) Experimental
verication of fractional history effects on the viscous dynamics of small spherical
particles, Experiments in Fluids (38) 112-116.
Li, C. & Deng, W., Remarks on fractional derivatives, Applied Mathematics and
Computation, 187(2007) 777784.
Nonlinear Dynamics

144
Miller, K.S. & Ross,B.(1993) An Introduction to the Fractional Calculus and Fractional
Differential Equations, John Wiley and Sons, New York, NY.
Momani, S. (2006) A numerical scheme for the solution of multi-order fractional differential
equations, Applied Mathematics and Computation, 182 761770.
Moon, F.C. & Holmes, P.J. (1979) A magnetoelastic strange attractor. J Sound Vib, 65(2) 285-
296.
Moon, F.C. & Holmes, P.J. (1980) Addendum: A magnetoelastic strange attractor, J Sound
Vib, 69(2) 339.
Oldham, K.B. & Spanier, J. (1974) The Fractional Calculus, Academic Press, New York, NY.
Podlubni, I. (1999) Fractional Differential Equations, Academic Press, San Diego, CA.
Podlubni, I., Fractional-order systems and P I D -Controllers, IEEE Transactions on
Automatic Control, 44(1)(1999) 208214.
Ramirez, L.E.S. & Coimbra, C.F.M. (2007) A Variable Order Constitutive Relation for
Viscoelasticity Annalen der Physik (16), No. 7-8, pp. 543-552.
Samko, S.K. & Ross, B., Integration and differentiation to a variable fractional order, Integral
Transforms and Special Functions, 1 (4) (1993) 277300.
Sheu, L-J; Chen, H-K, Chen, J-H; Tam & L-M (2007) Chaotic dynamics of the fractionally
damped Duffing equation, Chaos Solitons & Fractals, 32 1459-1468.
Soon, C. M., Coimbra, C.F.M., & Kobayashi, M. H. (2005). "The Variable Viscoelasticity
Oscillator" Annalen der Physik, (14) N.6, pp. 378-389.
Vincent, U.E. & Kenfack, A. (2008) Synchronization and bifurcation structures in coupled
periodically forced non-identical Duffing oscillators, Physica Scripta, 77 045005
(7pp).
Williams, R.L. & Lawrence, D.A. (2007) Linear State-Space Control Systems. John Wiley and
Sons, USA. ISBN 978-0-471-73555-7.
7
Nonlinear Vibrations of Axially Moving Beams
Li-Qun Chen
Shanghai University
China
1. Introduction
Axially moving beams can represent many engineering devices, such as band saws, power
transmission belts, aerial cable tramways, crane hoist cables, flexible robotic manipulators,
and spacecraft deploying appendages. Despite usefulness and advantages of these devices,
vibrations associated with the devices have limited their applications. Therefore,
understanding transverse vibrations of axially moving beams is important for the design of
the devices. The investigations on vibrations of axially moving beams have theoretical
importance as well, because an axially moving beam is a typical representative of
distributed gyroscopic systems. The term gyroscopic arose in recognition of an early
problem in gyrodynamics. Actually, the Coriolis acceleration component experienced by
axially moving materials imparts a skew-symmetric or gyroscopic term to their governing
equations. Due to particular characteristics of the gyroscopic term, the approaches
developed in analyzing vibrations of an axially moving string can be applied to other more
complicated distributed gyroscopic systems. Because of the practical and theoretical
significance, the investigation on nonlinear vibrations of axially moving beams is a
challenging subject which has been studied for many years and is still of interest today.
The relevant researches on transverse vibrations of axially moving strings can be dated back
to (Aiken, 1878). There are several excellent and comprehensive survey papers, notably
(Mote, 1972), (Ulsoy and Mote, 1978), (Mote et al., 1982), (DAngelo et al., 1985), (Wickert
and Mote, 1988), (Wang and Liu, 1991), (Abrate, 1992), (Zhu, 2000), reviewing the state-of-
the-art in different time phases of investigations related to vibrations of axially moving
beams. The present chapter emphasizes on the recent achievements, although some early
results are mentioned for the sake of completeness. Besides, the chapter focuses the
nonlinear problem only. If the vibration amplitude is large, the nonlinearity should be taken
into account. Hence the chapter, unlike (Chen, 2005a) for axially moving strings, is not a
comprehensive survey with a complete and detailed representation of current researches.
Instead, the chapter is a counterpart of (Chen et al., 2008) for axially moving beams. The
author tries to put the some available results into a general framework, as well as to
highlight the work of the author and his collaborators. It is hoped that the chapter serves as
a collection of ideas, approaches, and main results in investigations on nonlinear vibration
of axially moving beams.
The chapter is organized as follows: Section 2 focuses on the mathematical models of
nonlinear transverse vibration. The special attentions are paid to the comparison of two
different nonlinear models and the introduction of the material time derivative into the
Nonlinear Dynamics

146
viscoelastic constitutive relations. Section 3 covers the developments and the applications of
approximately analytical methods, including the asymptotic method, the Lindstedt-Poincar
method, the method of normal forms, the method of nonlinear, complex modes, the method
of multiple scales, and the incremental harmonic balance method. Section 4 is devoted to the
numerical approaches, including the Galerkin discretization, the finite difference, and the
differential quadrature. Section 5 reveals the nonlinear behaviors such as bifurcation and
chaos based on the numerical solutions. Section 6 discusses energetics, conserved quantity
and the applications. The final section recommends future research directions.
2. Governing equations
2.1 Coupled vibration
The governing equation is the base of all analytical or numerical investigations. Generally,
an axially moving beam undergoes both the longitudinal vibration and the transverse
vibration, and they are coupled. (Thurman & Mote, 1969) obtained the governing equation
of coupled longitudinal and transverse vibrations of an axially moving beam. (Koivurova &
Salonen 1999) revisited the same modeling problem and clarified its kinematic aspects. Their
nonlinear formulation for the moving beam problem has two limitations: the material of the
beam is linear elastic constituted by Hookes law, and the axial speed of the beam is a
constant. As (Wickert & Mote 1988) pointed out, modeling of dissipative mechanisms is an
important vibrations analysis topic of axially moving materials. An effective approach is to
model the beam as a viscoelastic material. Therefore, it is necessary to deal with constitutive
laws other than Hookes law. Axial transport acceleration frequently appears in engineering
systems. For example, if an axially moving beam models a belt on a pair of rotating pulleys,
the rotation vibration of the pulleys will result in a small fluctuation in the axial speed of the
belt. The nonlinear model in (Thurman & Mote, 1969) for coupled vibration can be
generalized to an axially accelerating viscoelastic beam as follows.
Consider a uniform axially moving beam of density , cross-sectional area A, moment of
inertial I, and initial tension P
0
, as shown in Figure 1. The beam travels at the uniform
transport speed between two boundaries separated by distance l. Assume that the
deformation of the beam is confined to the vertical plane. A mixed Eulerian-Lagrangian
description is adopted. The distance from the left boundary is measured by fixed axial
coordinate x. The beam is subjected to an external excitation f
u
(x,t) and f
v
(x,t) in longitudinal
and transverse directions respectively, where t is the time. The in-plane motion of the beam
is specified by the longitudinal displacement u(x,t) related to coordinate translating at speed
and the transverse displacement v(x,t) related to the spatial frame.


Fig. 1. The physical model of an axially accelerating beam
Study the motion of the beam in a reference frame moving in the axial direction and at
speed . The reference system is not an inertial frame if is not a constant. The beam is a
Nonlinear Vibrations of Axially Moving Beams

147
one-dimensional continuum undergoing an in-plane motion in the moving reference frame,
the Eulerian equation of motion of a continuum gives

( )( )
( )
( )
( )
( )
( ) ( )
2
0
2
2
2
2
0
2
2
2
1 , , d 1
,
d
1 , ,
, , , , d 1
,
d
1 , ,
x u
x x
x xx v
x x
P A u f x t u
t x A A
u v
P A v M x t f x t v
t x A A A
u v

+ +

= +

+ +


+

= +

+ +

(1)
where a comma preceding x or t denotes partial differentiation with respect to x or t, (x,t) is
the axial disturbed stress, and M(x,t) is the bending moment. The viscoelastic material of the
beam obeys the Kelvin model, with the constitution relation
( ) ( )
d
, , ,
d
x t E x t
t


= +


N
(2)
where, E is the Young's modulus, is the dynamic viscosity, and the disturbed strain
N
(x, t)
of the beam is given by the nonlinear geometric relation
( )
2
2
N
1 , , 1
x x
u v = + + (3)
For a slender beam (for example, with I/(Al
2
)<0.001), the linear moment-curvature
relationship of Euler-Bernoulli beams is sufficiently accurate,
( )
d
, ,
d
xx
M x t E Iv
t

= +


(4)
In the moving reference frame, the beam itself is without any axial transportation, while the
boundaries are moving at speed . The axially moving beam is constrained by rotating
sleeves with rotational springs (Chen & Yang, 2006a). The stiffness constant of two springs is
the same, denoted as K. Nullifying the transverse displacements and balancing the bending
moment at both ends lead to the boundary conditions
( ) ( ) , 0, , 0, ; u s t u l s t = + = (5)
( ) ( ) ( ) ( ) ( ) ( ) , 0, , , , , 0; , 0, , , , , 0.
xx x xx x
v s t EIv s t Kv s t v l s t EIv l s t Kv l s t = = + = + + + = (6)
where s = . To avoid the moving boundary conditions (5), which are difficult to tackle, the
transformation of coordinates is introduced as follows

, . x x s t t +
(7)
Then, expressed in the new coordinates, the boundary conditions have a simpler form
( ) ( ) 0, 0, , 0; u t u l t = = (8)
( ) ( ) ( ) ( ) ( ) ( ) 0, 0, , 0, , 0, 0; , 0, , , , , 0.
xx x xx x
v t EIv t Kv t v l t EIv l t Kv l t = = = + = (9)
Nonlinear Dynamics

148
Under the new coordinates, the partial derivatives with respect to x and t remain invariant,
and the total time derivative changes as follows

d
dt x t

+

(10)
Substitution of equations (2), (4), and (10) into equation (1) yields

( ) ( )
( ) ( )
( )
( )
( )
( )
( )
( )
2
0 N N N
2
2
2
0 N N N
2
2
, 2 , 1 , ,
, , 1 ,
, ,
1 , ,
, 2 , , ,
, , ,
, , , ,
1 , ,
tt xt x xx
t x x
u
x x
tt xt x xx
t x x
xxxx xxxxt xxxxx v
x x
A u u u u
P A E u
f x t
x
u v
A v v v v
P A E v
EIv I v v f x
x
u v





+ + + + =

+ + + +

+ +

+ + + =

+ + +

+ + +

+ +

( ) , t
(11)
If other viscoelastic constitutive relations are used to describe the beam materials, they can
be incorporated into the governing equation in the similar way. However, a controversial
issue arises concerning the application of differential-type constitutive laws including the
Kelvin relation in axially moving materials. Some investigators used the partial time
derivative in the Kelvin model for axially moving strings (Zhang & Zu, 1998) (Zhang &
Song, 2007), (Chen et al., 2007) and (Ghayesh, 2008), or beams (Chen & Yang, 2005, 2006a,b),
(Ghayesh & Balar, 2008), (Ghayesh & Khadem, 2008), (Yang et al., 2009), and (zhan &
Pakdemirli, 2009). However, (Mochensturm & Guo, 2005) convincingly argued that the
Kelvin model generalized to axially moving materials should contain the material time
derivative to account for the added steady state dissipation of an axially moving
viscoelastic string. Actually the material time derivative was also used in other works on
axially moving viscoelastic beams (Marynowski, 2002, 2004, 2006), (Marynowski &
Kapitaniak, 2002, 2007), (Yang & Chen, 2005), (Ding & Chen, 2008), (Chen & Ding, 2008,
2009), (Chen & Wang, 2009) and (Chen, et al., 2008, 2009, 2010). Here a coordinate transform
will be proposed to develop the governing equations, which can introduce naturally the
material time derivative in the viscoelastic constitutive relations.
In small but finite stretching problems in literatures of nonlinear oscillations, only the lowest
order nonlinear terms need to be retained so that the governing equation of small-amplitude
motion will be obtained. Such simplified coupled governing equations were used in
analytical investigations on axially moving elastic beams (Thurman & Mote, 1969), (Riedel &
Tan, 2002), and (Sze et al., 2005).
It should be remarked that there are different types of governing equations for axially
moving beams (Tabarrok et al., 1974), (Wang & Mote, 1986, 1987), (Wang, 1991), (Hwang &
Perkins, 1992a,b, 1994), (Vu-Quoc & Li 1995), (Behdinan, et al, 1997), (Hochlenert et al.,
2007), (Pratiher & Dwivedy 2008), (Spelsbrg-Korspeter et al., 2008), and (Humer & Irschik,
2009). Actually, there are various beam theories such as Euler-Bernoulli theory, shear-
deformable theories, and three-dimensional theories, and geometric nonlinearities may take
different forms. Correspondingly, there are various governing equations of axially moving
beams. Even if an axially stationary slender structure is prescribed by more sophisticated
Nonlinear Vibrations of Axially Moving Beams

149
governing equations, the coordinate transform (7) is a still convenient approach to derive
the governing equations of the slender structure undergoing an axial motion.
2.2 Transverse vibration
Although the transverse vibration is generally coupled with the longitudinal vibration,
many researchers considered only the transverse vibration in order to derive a tractable
equation. Inserting u=0 into equation (3) and then omitting higher order nonlinear terms
yield a simplified strain-displacement relation termed as the Lagrange strain

2
L
, 2
x
v = (12)
Inserting u=0 into equation (11) and then retaining lower order nonlinear terms only yield a
nonlinear partial-differential equation

( ) ( )
( ) ( )
2
0
L L L
, 2 , , , , , , ,
, , , , .
tt xt x xx xx xxxx xxxxt xxxxx
t x x v
A v v v v P v EIv I v v
AE A A v f x t
x


+ + + + + +

= + + +

(13)
The quasi-static stretch assumption means that one can use the averaged value of the
disturbed tension ( )
1
L L L
0
, , d
t x
AE A x l + +

to replace the exact value
AE+A(,
t
+c,
x
). Thus equation (18) leads to nonlinear integro-partial-differential equation

( )
( )
( ) ( )
2
0
L L L
0
, 2 , , , , , , ,
,
, , d , .
tt xt x xx xx xxxx xxxxt xxxxx
l
xx
t x v
A v v v v P v EIv I v v
v
AE A A x f x t
l


+ + + + + +

= + + +

(14)
Both equation (13) and equation (14) are governing equations of transverse nonlinear
vibration.
Both the nonlinear partial-differential equation and the nonlinear integro-partial-differential
equation have been applied to some special cases such as free vibration without external
excitation (F
v
=0), elastic beams without viscoelasticity (=0), uniformly moving beams
without axially acceleration (

=0). The applications of the nonlinear partial-differential

equation include (Chen & Zu, 2004) for uniformly moving elastic beams without external
excitation, (Marynowski, 2002, 2004) and (Marynowski & Kapitaniak, 2007) for axially
moving viscoelastic beams without external excitation, (Yang & Chen, 2005) and (Chen &
Yang, 2006) for axially accelerating viscoelastic beams, and (Chen et al., 2007) for uniformly
moving elastic beams without external excitation. The applications of the nonlinear integro-
partial-differential equation include (Wickert, 1991), (Pellicano & Zirilli, F., 1997), (Pellicano
& Vestroni, 2000), (Chakraborty & Mallik, 2000a), (Pellicano, 2001), (Kong & Parker, 2004)
and (Chen & Zhao, 2005) for uniformly moving elastic beams without external excitation,
(Ghayesh, 2008) for uniformly moving viscoelastic beams without external excitation,
(Pellicano & Vestroni, 2000), (zhan & Pakdemirli, 2009) for uniformly moving elastic
beams, (Chakraborty & Mallik, 1999), (z et al, 2001) and (Ravindra & Zhu, 1998) for axially
accelerating elastic beams without external excitation, (Chakraborty & Mallik, 1998)
(Chakraborty et al., 1999), (Chakraborty & Mallik, 2000b) for axially moving elastic beams,
Nonlinear Dynamics

150
(Parker & Lin, 2001) for axially accelerating elastic beams, and (Yang et al., 2009), (Chen et
al., 2009) for axially accelerating viscoelastic beams, and (zhan & Pakdemirli, 2009) for
uniformly moving viscoelastic beams. Approximately analytical investigations on free
vibration of axially moving elastic (Chen & Yang, 2007), forced vibration of axially moving
viscoelastic beams (Yang & Chen, 2006), and parametric vibration of axially accelerating
viscoelastic beams (Chen & Yang, 2005) and (Chen & Ding, 2008) demonstrated that the
nonlinear partial-differential equation and the nonlinear integro-partial-differential equation
yield the qualitatively same results but there are quantitative differences.
The nonlinear integro-partial-differential equation can also be obtained through uncoupling
the governing equation for coupled longitudinal and transverse vibration under the quasi-
static stretch assumption in small but finite stretching problems, and a special case of free
vibration of axially moving elastic beam was treated in (Wickert, 1992). Under quasi-static
stretch assumption, the dynamic tension to be a function of time alone. In traditional
derivation in (Wickert, 1992), the nonlinear integro-partial-differential equation seems more
exact than the nonlinear partial-differential equation because it is the transverse equation of
motion in which the longitudinal displacement field is taken into account. However, the
derivation here indicates that the nonlinear partial-differential equation can be reduced to
the nonlinear integro-partial-differential equation based on the quasi-static stretch
assumption. Numerical investigations on free vibration of axially moving elastic beams
(Ding & Chen, 2008) and forced vibration of axially moving viscoelastic beams (Chen &
Ding, 2009) indicated that the nonlinear integro-partial-differential equation is superior to
the partial-differential equation, in the sense that approximates the coupled governing
equation of planar motion better (some details in Subsection 4.2). However, since there has
no decisive evidence to favor any models of transverse nonlinear vibration of axially
moving beams, it is still an open problem.
3. Approximate analytical methods
3.1 Direct-perturbation approaches
As exact solutions are usually unavailable, approximate analytical methods are widely
applied to investigate nonlinear vibration of axially moving beams. The approximate
analytical methods can be applied to the nonlinear (integro-)partial-differential equations
without discretization. Such a treatment is regarded as a direct-perturbation. The practice
can be dated back to (Thurman & Mote, 1969) in which a modified Lindstedt method was
used to calculate the fundamental frequency.
The method of multiple scales can be employed to analyze nonlinear vibration of axially
moving beams. Actually, a general framework of the multi-scale analysis has been proposed
for a linear gyroscopic continuous system under small nonlinear time-dependent
disturbances (Chen & Zu, 2008). Consider a gyroscopic continuous system with a weak
disturbance
( ) , , , ,
tt t
Mv Gv Kv N x t + + = (15)
where v(x,t) is the generalized displacement of the system at spatial coordinate x and time t,
linear, time-independent, spatial differential operators M, G and K represent mass,
gyroscopic and stiffness operators respectively, stands for a small dimensionless
parameter, and N(x,t) expresses a nonlinear function of x and t that may explicitly contain v
Nonlinear Vibrations of Axially Moving Beams

151
and its spatial and temporal partial derivatives as well as its integral over a spatial region or
a temporal interval. N(x,t) is periodic in time with the period 2/. Define an inner product
( ) ( ) , d ,
E
f g f x g x x =

(16)
for complex functions f and g defined in the gyroscopic continuum E, where the overbar
denotes the complex conjugate. M, K are symmetric and G is skew symmetric in the sense
, , , , , , , , Mf g f Mg Kf g f Kg Gf g f Gg = = = (17)
for all functions f and g satisfying appropriate boundary conditions. A uniform
approximation is sought in the form
( ) ( ) ( )
( )
2
0 0 1 1 0 1
, , , , , , v x t v x T T v x T T O = + + (18)
where T
0
=t, T
1
=t, and O(
2
) denotes the term with the same order as
2
or higher.
Substitution of equation (18) into equation (15) yields

0 0 0
0 0 0
, , 0,
T T T
Mv Gv Kv + + = (19)
( )
0 0 0
1 1 1 1 0 1
, , , , ,
T T T
Mv Gv Kv N x T T + + = (20)
where N
1
(x,T
0
,T
1
) stands for a nonlinear function of x, T
0
and T
1
, which usually depends
explicitly on v
0
and its derivatives and integrals. In addition, N
1
(x,T
0
,T
1
) is periodic in T
0

with the period 2/. Separation of variables leads to the solution of equation (19) as
( ) ( ) ( )
0
i
0 0 1 1
1
, , e ,
j
T
j j
j
v x T T A T x cc

=
= +

(21)
where A
j
denotes a complex function to be determined later,
j
and
j
represents
respectively the complex modal function and the natural frequency given by

2
i 0
j j j j j
M G K + + = (22)
and the boundary conditions, and cc stands for the complex conjugate of all preceding terms
on the right side of the equation. If approaches a linear combination of natural frequencies
of equation (19), the summation parametric resonance may occur. A detuning parameter is
introduced to quantify the deviation of from the combination, and is described by

1
,
j j
j
c

=
= +

(23)
where c
j
are real constants that are not all zero and only a finite of them are not zero. To
investigate the summation parametric resonance, substitution of equations (21) and (23) into
equation (20) leads to
Nonlinear Dynamics

152
( )
0
0 0 0
i
1 1 1 1
1
, , , e ,
j
T
T T T j
j
Mv Gv Kv F x T NST cc

=
+ + = + +

(24)
where F
j
(x, T
1
) (j=1,2,) are complex functions dependent explicitly on A
j
(T
1
) and their
temporal derivatives as well as
j
(x) and their spatial derivatives and integrals. (Chen & Zu,
2008) proved that the solvability condition is the orthogonality of the coefficient of the
resonant term in the first order equation and the corresponding modal function of the zero
order equation, e.g.
( )
1
, , 0.
j j
F x T = (25)
It should be noticed that the solvability condition (25) holds providing the boundary
conditions are appropriate. That is, M and K are symmetric and G is skew symmetric under
the boundary conditions. In a specific problem, these requirements can be checked for a
given the operators, boundary conditions and the modal functions. However, the
examination depends only on the unperturbed linear part of the problem. For example,
equation (25) holds for an axially moving beam under condition (9) (Chen & Zu, 2008).
Usually, it is assumed that only the modes involved in the resonance (23) need to be
considered in the linear solution (21), and the assumption is physically sound. Some case
studies demonstrated mathematically that the mode uninvolved in the resonance has no
effect on the steady-state response (Ding & Chen, 2008), (Chen & Wang, 2009), and (Chen et
al., 2009). (zhan & Pakdemirli, 2009) proposed multi-scale analysis on forced vibrations of
general continuous systems with cubic nonlinearities in the primary resonance case.
The method of multiple scales has been applied in various transverse nonlinear vibration
problems of axially moving beams. These problems include free (z et al, 2001) and (Chen &
Yang, 2007), forced (zhan & Pakdemirli, 2009), and parametric(z et al, 2001) and (zhan
& Pakdemirli, 2009) vibration of axially moving elastic beams, as well as forced (Yang &
Chen, 2006) and parametric (Chakraborty & Mallik, 1999), (Chen & Yang, 2005) and (Chen &
Ding, 2008) vibrations of axially moving viscoelastic beams. In addition to these works on
the base of the Euler-Bernoulli beam theory, the method of multiple scales has also be
applied to study free vibration of an axially moving beam with rotary inertia and
temperature variation effects (Ghayesh & Khadem, 2008), parametric vibration of axially
moving viscoelastic Rayleigh beams (Ghayesh & Balar, 2008), and forced (Tang et al. 2009)
and parametric (Tang et al., 2010) vibrations of axially moving elastic Timoshenko beams,
while the multi-scale analysis on axially moving viscoelastic Timoshenko beams has been
only limited to linear parametric vibration (Chen et al., 2010).
Addition to the method of multiple scales, the method of asymptotic analysis is also an
effective approach to treat parametric or nonlinear vibration. Based on the idea of Krylov,
Bogoliubov, and Mitropolsky, (Wickert, 1992) developed a asymptotic method for general
gyroscopic continuous systems with weak nonlinearities, and the method was specialized to
free nonlinear vibration of an axially moving elastic beam with supercritical transport speed.
(Maccari, 1999) proposed another asymptotic approach for analyzing transverse vibration of
axially stationary beams, which are disturbed conservative continuous systems, and
determined external force-response and frequency-response curves in the cases of primary
resonance and subharmonic resonance for a weakly periodically forced beam with quadratic
and cubic nonlinearities. The approach was extended to the gyroscopic continuous system
Nonlinear Vibrations of Axially Moving Beams

153
with a weak nonlinear and time-dependent disturbance in order to analyze transverse
vibration of an axially accelerating viscoelastic string constituted by the Kelvin model (Chen
et al., 2008) and the standard linear solid model (Chen & Chen, 2009). The method of
asymptotic analysis has been also presented for nonlinear parametric vibration of axially
accelerating viscoelastic beams constituted by the Kelvin model (Chen et al., 2009) as well as
linear parametric vibration of axially accelerating viscoelastic beams constituted by the
Kelvin model (Chen & Wang, 2009) and the standard linear solid model (Wang & Chen,
2010).
Nonlinear normal modes whose shapes depend on the amplitude provide a possible direct
treatment on nonlinear vibration of axially moving beams. (Chakraborty et al., 1999) used a
temporal harmonic balance and a spatial perturbation technique to determine the nonlinear
complex normal modes for free and forced vibrations of axially moving elastic beams. The
approach was adopted to study the response of a parametrically excited axially moving
beam both without and with an external harmonic force (Chakraborty & Mallik, 1998). The
results were justified by the wave propagation analysis (Chakraborty & Mallik, 2000a,b).
3.2 Discretization-perturbation approaches
Discretization of governing equations is a commonly used approach to obtain approximate
solutions of vibration problems of continuous systems. For the governing equations (11) of
coupled vibration of axially moving beams, one assumes an approximate solution in the
form
( ) ( ) ( )
1
, ,
m
i i
i
u x t p t x
=
=

(26)
( ) ( ) ( )
1
, ,
n
i i
i
v x t q t x
=
=

(27)
where p
i
(t) and q
i
(t) are generalized coordinates, and
i
(x) are
i
(x) base functions that are
usually chosen to be the linear vibration mode shapes of axially stationary beams or moving
beams (Wickert & Mote, 1990), (Chen & Yang, 2006), and (Tang et al. 2008). A weighted-
residual procedure such as the Galerkin procedure can be applied to truncate equation (11)
into m+n nonlinearly coupled second-order ordinary-differential equations. A general
description of the Galerkin procedure is as follows. Denote the differences between the left
and right sides of two equations in equation (11) as F
u
(x,u,v,t) and F
v
(x,u,v,t), which are
nonlinear functions of x and t that may explicitly contain v and its spatial and temporal
partial derivatives as well as its integral over a spatial region or a temporal interval. Then
approximate solution (31) and (32) satisfies

( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( ) ( )
1 1
1 1
, , , ,
0, , , , , 0, 1, , ; 1, ,
m n
u i i i i j
i i
m n
v i i i i k
i i
F x p t x v q t x t x
F x p t x v q t x t x j m k n


= =
= =





= = = =





(28)
where
j
(x) and
k
(x) are the weight functions.
Nonlinear Dynamics

154
After the discretization, various perturbation techniques such as the method of multiple
scales can be employed to analyze the resulting nonlinear ordinary-differential equations
approximately. Such a treatment is regarded as a discretization-perturbation
In practical problems, m and n in the discretization expressions (26) and (27) are rather
small, and they are usually 1 or 2. (Riedel & Tan, 2002) applied the method of multiple
scales to the discretized equations (m=n=2) to determine the forced response of an axially
moving elastic beam with internal resonance. The method of multiple scales was also
applied to the discretized problem of coupled vibration of axially moving beams (Feng &
Hu, 2002, 2003). (Sze et al., 2005) presented a general description of discretization of the
governing equation of an axially moving elastic beam, and used incremental harmonic
balance method to a concrete case of m=n=2 for forced response with internal resonance. In
both studies, the mode shapes of axially stationary beams were chosen as the base functions
and the weight functions.
Discretization-perturbation approaches have also been used in analyzing transverse
nonlinear vibration of axially moving beams. In this case, equation (27) will be substituted
into equation (13) or (14) and then the Galerkin procedure can be used to discretize equation
(13) or (14) into n nonlinearly coupled second-order ordinary-differential equations that can
be solved approximately via various perturbation techniques. The Lindstedt-Poincar
method was applied to discretized governing equations to evaluate transverse response of
axially moving beams (Pellicano & Zirilli, 1997) and to analyze parametric instability of
axially moving elastic beams subjected to multifrequency excitations (Parker & Lin, 2001).
The method of normal forms was used to evaluate free vibration of axially moving elastic
beams with internal resonances (Pellicano & Zirilli, 1997) as well as forced and parametric
vibration of axially moving elastic beams (Pellicano et al., 2001). In (Pellicano & Zirilli, 1997),
(Parker & Lin, 2001), and (Pellicano et al., 2001), the mode shapes of axially moving beams
were chosen as the base functions and the weight functions, and their orthogonality were
employed. The stationary mode shapes can also serve as the base functions and the weight
functions to discretize governing equations. Based on the discretization, the Lindstedt-
Poincar method was applied to determine the forced response of axially moving elastic
beams (Chen et al., 2007), and the method of multiple scales was applied to evaluate the
response of an axially moving viscoelastic beams subjected to multifrequency external
excitations (Yang et al., 2009). In their studies, n=2 (Chen et al., 2007) and n=1 (Yang et al.,
2009), respectively.
4. Numerical approaches
4.1 Galerkin procedure
Numerical calculation is an effective approach to studying nonlinear vibration of axially
moving beams. Based on the numerical solutions of the governing equations, some
changing tendencies of vibration characteristics, such as frequencies or amplitudes, with
related parameters can be predicted, the approximate analytical results can be verified, and
the nonlinear dynamical behaviours can be revealed.
Among numerical approaches, the Galerkin procedure can be applied to discretize the
governing equation of nonlinear vibration of axially moving beams. The Galerkin
discretization is not only the priority of discretization-perturbations reviewed in Subsection
3.2, but also feasible approach to numerical solutions. Using the 3 order Galerkin
discretization of governing equation (in the type of equation (18)) for transverse motion of
Nonlinear Vibrations of Axially Moving Beams

155
axially moving viscoelastic beams excited by the changing tension, (Marynowski, 2002) and
(Marynowski & Kapitaniak, 2002) numerically investigated the effects of different
viscoelastic models, such as the Kelvin model, the Maxwell model, and the standard linear
solid model, on the dynamic response and found that different viscoelastic models yield
very close numerical results for small damping. The Galerkin procedure has been mainly
use to calculate long time nonlinear dynamical behaviors, which will be addressed in
Subsection 5.1.
In the application of the Galerkin discretization, the main problem in the actual
computations is the complexity of the resulting discretized equations. If the number of terms
retained in the Galerkin discretization is rather large, the explicit expression of nonlinear
terms is very difficult to obtain. (Chen & Yang, 2006b) proposed a technique to simplify the
nonlinear terms in the equations derived from the Galerkin discretization. All nonlinear
terms are regrouped to combine the repeated terms and cancel the zero terms. Therefore, the
resulting equations can be easily coded for computers and then be effectively calculated. For
example, the Galerkin discretization of the governing equation (18) for transverse motion (in
the dimensionless form) is

( ) ( ) ( )
( ) ( ) ( ) ( ) ( )
( ) ( ) ( )
( )
2
1 1 1
2 2
f 1
1 1 1 1 1
2
1 1 1
, 2 , ( 1) ,
3
, , ,
2
2 , , 0
n n n
i i k i i k i i k
i i i
n n n n n
i i k i i k i j s i j s k
i i i j s
n n n
i j s i s j k i j k k
i j s
q t q t q t
k q t q t k q t q t q t
k q t q t q t k



= = =
= = = = =
= = =
+ +
+ +
+ = =

( ) 1, 2,..., n
(29)
If both the base and weight functions are chosen as sine functions, the stationary mode
shapes for the simply supported beams, equation (29) can be cast into a from convenient to
compute. Evaluating the corresponding inner products, regrouping the nonlinear terms to
combine the same terms and canceling all null terms in the resulting equation, one obtains

( ) ( ) ( )
( )
( )
( ) ( )
( )
{ }
{ }
( ) ( )
(
2 2 2 2 4 4 4 4
f
2 2
is odd
min 1,
4 1
2 2
1 2 1
1 max 1, 2 1
4 2 1
7
n
k j j k k k
j k
s n
n k n k s
l n j s j s j k s j s j
s k i s n l j
kj
q t q t q t k q t k k q k q
k j
k
s k q i s i q k q k q n l q j s j q k q

+
+
= + = = =

+ + + + +

= + + +



16


+

) }}
( ) ( )
( )
(
) ( )
{ }
{ }
( ) ( ) ( ) ( )
4 4 1 1
2 2
2 1 2
2 1 1
min 1,
1
max 1, 2 1
5
k s n k
s j k s j s j s j
s i s k
s n
n k s
s k j s j k s j s j n s j s
i s n s i
k k k
k q k s q j s j q k q k q s
k q j s j q q k s q j s j q q k s q j s j q q

+

= = = +

+
= = =



+




16 8

+ + +

1 1
2 1
k l
j
l i

= =

(30)
where the sum is defined to be zero if its lower limit is larger than its upper limit. Although
equation (30) seems rather complicated, it is very efficient when used for computer
implementing, because almost all repeated nonlinear terms are put together, and terms with
zero coefficients are eliminated. In fact, equation (29) contains 2n
3
nonlinear terms, while
equation (30) contains less than 2n
2
nonlinear terms. For large n, the difference is significant.
Nonlinear Dynamics

156
It should be remarked that, based on stationary mode shapes, the even order Galerkin
discretization can take the linear gyroscopic terms into full account, while the odd order
discretization will miss some effects of the gyroscopic terms.
4.2 Finite difference
The finite difference method is a numerical procedure to solve partial differential equations.
The method can be used to discretize both spatial coordinates and time or to discretize
spatial coordinates only. In the former case, the procedure consists of four steps: 1 Discretize
the continuous spatial domain and temporal interval, on which a partial differential
equation is defined, into a discrete finite difference grid; 2 Approximate the individual exact
partial derivatives in the partial differential equation by algebraic finite difference
approximations; 3 Substitute the finite difference approximations into the partial differential
equation in order to derive a set of algebraic finite difference equations; 4 Solve the resulting
algebraic finite difference equations.
The finite difference method can be applied to calculated nonlinear vibration of axially
moving beams. For example, the method will be employed to solve numerically equation
(11) (Chen & Ding, 2010). Introduce the LT equispaced mesh grid with time step and
space step h: x
j
=jh (j=0, 1,2,,L, h=l/L); t
n
=n (n=0,1,2,,N, =T/N), where T is the
calculation termination time. Denote the function values u(x,t) and v (x,t) at (x
j
,t
n
) as u
n
j
and
v
n
j
. Application of centered difference approximations to the spatial, temporal and mixed
partial derivatives leads to

1 1
1 1 1 1
2 2
1 1 1 1
1 1 1 1
2 2
, , , , , ,
2
, ,
4
n n n n n n n n
j j j j j j j j
x tt tt
n n n n
j j j j
xt
u u u u u u u u
u u u
h h
u u u u
u
h

+
+ +
+ +
+ +
+ +
= = =
+
=
(31)
and
1 1 1 1 2 1 1 2
2 3
2 1 1 2 3 2 1 1 2 3
4 5
1 1
2
2 2 2
, , , , , ,
2 2
4 6 4 4 5 5 4
, , , ,
2
2
, , ,
n n n n n n n n n
j j j j j j j j j
x xx xxx
n n n n n n n n n n n
j j j j j j j j j j j
xxxx xxxxx
n n n
j j j
tt xt
v v v v v v v v v
v v v
h h h
v v v v v v v v v v v
v v
h h
v v v v
v v

+ + + +
+ + + + +
+
+ + +
= = =
+ + + + +
= =
+
= =
1 1 1 1
1 1 1 1
1 1 1 1 1 1 1 1 1 1
2 1 1 2 2 1 1 2
4
,
4
4 6 4 4 6 4
, ,
2
n n n n
j j j j
n n n n n n n n n n
j j j j j j j j j j
xxxxt
v v v
h
v v v v v v v v v v
v
h

+ +
+ +
+ + + + +
+ + + +
+
+ + + +
=
(32)
Substitution of equations (31) and (32) into equation (11) leads to a set of algebraic equations
with respect to u
n
j
and v
n
j
that can be solved as under the boundary conditions (8) and (9) for
prescribed parameters and initial conditions. Then the resulting grid values u
n
j
and v
n
j
are
used in the finite difference schemes as an approximation to the continuous solutions u(x,t)
and v(x,t) to equation (11). When the external transverse load is a spatially uniformly
distributed periodic force, the amplitude of the beam center displacement changes with the
force frequency, which is shown in Fig. 2.
Nonlinear Vibrations of Axially Moving Beams

157
The finite difference method was applied to examine the validity of the two nonlinear
transverse models (equations (13) and (14)) and to determine the superiority in the sense of
approximating the coupled governing equation (11) of planar vibration. For forced vibration
of axially moving viscoelastic beams (Chen & Ding, 2010), the steady-state transverse
responses of the beam center calculated from the two transverse models are contrasted with
the results based on the coupled equations of planar vibration. Qualitatively, the three
models predict the same tendencies with the changing parameters. Quantitatively, there are
certain differences. In the view of both the center amplitude and the beam shape, the
nonlinear integro-partial-differential equation yields the results closer to those from the
governing equation of coupled vibration. The similar result was obtained by the finite
difference method for response in free vibrations of axially moving elastic beams (Ding &
Chen, 2009a).

0 20 40 60 80 100
0.00
0.04
0.08
A

16 18 20 22
0.00
0.04
0.08
A

(a) the amplitude-frequency relation (b) local magnification of (a) near
1
=18.107

49 50 51
0.0002
0.0003
0.0004
0.0005

A

90 95 100
0.000
0.001
0.002

A

(c) local magnification of (a) near
2
=49.706 (d) local magnification of (a) near
3
=97.140
Fig. 2. The response amplitude changing with the external excitation frequency
The finite difference method was used to confirm the analytical results of nonlinear
transverse vibration of axially moving beams. For free vibration of axially moving elastic
Nonlinear Dynamics

158
beams, (Pellicano & Zirilli, 1997) compared the beam center displacement changing with
time via the Lindstedt- Poincar method and the normal form method with the numerical
solutions via the finite difference method, and found that they are in good agreement. For
parametric vibration of axially accelerating viscoelastic beams, (Chen & Ding, 2008)
compared the stable steady-state response of the beam center via the method of multiple
scales with the numerical solutions via the finite difference method, and demonstrated that
they have the same qualitative tendencies changing with the related parameters and are
quantitatively with rather high precision.
4.3 Differential quadrature
The differential quadrature method, initiated from the idea of integral quadrature, is an
efficient discretization technique to seek accurate numerical solutions using a considerably
small number of grid points. The method can be used to discretize both spatial coordinates
and time or to discretize spatial coordinates only. In later case, the differential quadrature
discretization of a partial-differential equation yields a set of differential-algebraic equations
via the following four steps. 1 Discretize the continuous spatial domain, on which a partial
differential equation is defined, by grid points; 2 Approximate the individual exact partial
derivatives in the partial differential equation by a linear weighted sum of all the functional
values at all grid points; 3 Substitute the differential quadrature approximations into the
partial differential equation to obtain a set of ordinary-differential-algebraic equations; 4
Solve the resulting ordinary-differential-algebraic equations. The two extensively decisive
issues in the applications of the differential quadrature method are to choose grid points
and to determine the weighting coefficients for the discretization of a derivative of necessary
order.
The differential quadrature method can be applied to calculated nonlinear vibration of
axially moving beams. Equation (11) is treated as an example to show the application of the
differential quadrature method. Introduce N unequally spaced grid points as (Bert & Malik,
1996) and (Shu, 2001)

( )
( )
1 1
1 cos 1, 2, , .
2 1
i
i
x i N
N

= =

(33)

The quadrature rules for the derivatives of a function at the grid points yield

( )
( )
( )
( )
( )
( )
( )
( )
( )
( )
( )
( )
( )
( )
( )
( )
( )
( )
1 2
1 1
1 2
1 1
4 5
1 1
, , , , , , , ,
, , , , , , , ,
, , , , , , , ,
N N
x i j xx i j ij ij
j j
N N
x i j xx i j ij ij
j j
N N
xxxx i j xxxxx i j ij ij
j j
u x t A u x t u x t A u x t
v x t A v x t v x t A v x t
v x t A v x t v x t A v x t
= =
= =
= =
= =
= =
= =



(34)

where the weighting coefficients are the expression
Nonlinear Vibrations of Axially Moving Beams

159

( )
( ) ( )
( )
(1) 1,
1,
, 1, 2, , ;
N
i k
k k i
ij
N
i j j k
k k j
x x
A i j N j i
x x x x
=
=

= =

(35)
and the recurrence relationship

( )
( )
( 1)
( ) ( 1) (1)
( ) ( )
1,
2, 3, 4, 5; , 1, 2, , ; ,
1, 2, 3, 4, 5; 1, 2, , .
r
ij
r r
ij ii ij
i j
N
r r
ii ik
k k i
A
A r A A r i j N j i
x x
A A r i N

=


= = =

= = =

(36)
Consider the beam with simply supports at both ends (K=0 in equation (9)). Substitution of
equation (34) into equation (11) and modification of the weighting coefficient matrices to
implement the boundary conditions (Wang & Bert 1993) lead to the ordinary-differential-
algebraic equations

( ) ( ) ( )
( )
( )
( )
( )
( )
( )
( ) ( ) ( ) ( )
1 1 2 2
1 1
1
1
1 1
0
1 1
1 5 1 2 2
1
2
1
,
,
1
2
N N
j ij j ij ij j
j j
N
ij j
N N
u j j
ij j j ij j
j j
j
N
j ij ij j ij ij
j
u A u A A u
A u
f x t
A P A E A
A A
I EI
v A A v A A A
A A

= =
=
= =
=
+ + + + =

+





+ + + +


+




+ + + + +

( ) ( )
( )
( )
( )
( )
4 5
1
1
1 1 1
0
1 1
1 1
,
1
, ( 2, 3,..., 1) ,
1
0,
N
ij ij j
j
N
ij j
N N
v j j
ij j j ij j
j j
j
N N
A v
A
A v
f x t
A P A E A j N
A A
v v u u

=
=
= =

+






= + + + + =


+



= = = =

(37)
which can be numerically solved via the convenient integration routines.
The differential quadrature method was applied to check the validity and the superiority of
the two nonlinear transverse models. For free vibration of axially moving elastic beams
(Ding & Chen, 2009a), the transverse responses of the beam center calculated from the two
transverse models are contrasted with the results based on the coupled equations of planar
vibration. The computational investigation leads to the following conclusions: 1 The
differences between the two models are both relatively small for not very large vibration; 2
The model differences increase with the vibration amplitude and the axial speed; 3 The
integro-partial-differential equation yields better results.
The differential quadrature method was used to validate the analytical results of nonlinear
transverse vibration of axially moving beams. (Chen et al., 2009) developed a differential
Nonlinear Dynamics

160
quadrature scheme to verify the approximate analytical results of stable steady-state
response in parametric vibration of axially accelerating viscoelastic beams. Figure 3 shows
the comparison, in which the solid and dot lines represent the results of the asymptotic
analysis method and the differential quadrature method respectively. The amplitudes from
both methods are almost coincided, especially near the exact-resonance (=0) and in the first
resonance. The differential quadrature method was also applied to confirm the analytical
results of the stability regions in linear parametric vibration of axially accelerating beams
constituted by the Kelvin model (Chen & Wang, 2009) and the standard linear solid model
(Wang & Chen, 2009)

-1 0 1
0.00
0.02
0.04

A

-0.5 0.0 0.5
0.000
0.002
0.004

A

(a) the first principal parametric resonance (b) the second principal parametric resonance
Fig. 3. Comparison of analytical and numerical results
5. Nonlinear dynamical behaviours
5.1 Galerkin discretization
Axially moving beams undergo periodic vibrations in the aforementioned researches.
Nonlinear system may exhibit chaos, steady-state response sensitive to initial conditions
thus unpredictable after a certain time and recurrent but either periodic or quasiperiodic
hence like a random single with the continuous frequency spectrum. Besides, the dynamical
behaviors of nonlinear system may change qualitatively at the critical value of the
parametric variation, and the qualitative change is termed as bifurcation.
Many investigations on bifurcation and chaos are based on the Galerkin discretization of
various transverse models of axially moving beams. For transverse free vibration of
accelerating elastic beams in the supercritical regime, based on 1 order Galerkin
discretization, (Ravindra & Zhu, 1998) applied Melnikovs criterion to find out the
parameter condition of occurring chaos and performed numerical simulations to show both
period-doubling and intermittent routes to chaos. For transverse harmonically forced
vibration of axially moving elastic beams in the supercritical regime, based on 8 order
Galerkin discretization, (Pellicano & Vestroni, 2002) observed intricate scenario of chaos,
including cascade of bifurcations, blue-sky catastrophes and coexisting chaotic and periodic
orbits. Actually, they also considered 12 order Galerkin discretization and found that a few
number of degree-of-freedom is sufficient to furnish a good spatial representation and to
Nonlinear Vibrations of Axially Moving Beams

161
follow the actual dynamical behaviors. For transverse parametric vibration of axially
moving viscoelastic beams excited by the time-dependent tension, based on 4 order Galerkin
discretization, (Marynowski, 2004) and (Marynowski & Kapitaniak, 2007) observed the
inverse period doubling and inverse Hope bifurcation and occurrences of regular and
chaotic motions for beams constituted by the Kelvin model and the standard linear solid
model respectively. For transverse parametric vibration of axially accelerating viscoelastic
beams, based on 4 order Galerkin discretization, (Chen & Yang, 2006b) constructed
numerically the bifurcation diagrams in the case that the axial speed perturbation
amplitude, the mean axial speed, or the viscosity coefficient is respectively varied while
other parameters are fixed. They also calculated the largest Lyapunov exponent from the
discretized governing equation. Numerical results show that, with the increasing speed
perturbation amplitude, the increasing mean speed, and the decreasing viscosity coefficient,
the equilibrium loses its stability and bifurcates into a periodic motion, and the periodic
motion becomes chaotic motion via period doubling bifurcation. In addition, the chaotic
motion and the periodic motion exchange alternately for the sufficiently large speed


(a) From equilibrium to chaos (b) Local magnification
Fig. 4. Bifurcation versus the dimensionless speed fluctuation amplitude


(a) From chaos to equilibrium (b) Local magnification
Fig. 5. Bifurcation versus the dimensionless viscosity coefficient
Nonlinear Dynamics

162
perturbation amplitude and mean speed, and for the sufficiently small viscosity coefficient.
Figures 4 and 5 show respectively the bifurcation diagrams versus the dimensionless speed
fluctuation amplitude and the dimensionless viscosity coefficient.
5.2 Differential quadrature and time series
The differential quadrature method is an effective numerical technique for initial and
boundary problems, and it has much higher precision than the few term Galerkin
discretization. However, it has not been applied to calculate nonlinear behaviors of axially
moving materials until (Ding & Chen, 2009b). They used the differential quadrature method
to investigate bifurcation and chaos of an axially accelerating viscoelastic beam constituted
by the Kelvin model. Based on the numerical solutions, analysis of the time series yielded
the Lyapunov exponent to identify periodic and chaotic motions. Numerical results show
that, with the increasing mean axial speed, the equilibrium loses its stability and bifurcates
into a periodic motion, and the periodic motion becomes chaotic motion. The chaotic motion
and the periodic motion exchange alternately for the sufficiently large mean axial speed and
speed perturbation amplitude. Figures 6 and 7 show the Poincar map and the largest
Lyapunov exponent of periodic and chaotic motions respectively.

-0.001 0.000 0.001
-0.001
0.000
0.001
d
v
/
d
t
v( 0.5,t)

0 2500 5000
-0.2
0.0
0.2
0.4
0.6

t

(a) the Poincar map (b) the largest Lyapunov exponent
Fig. 6. Periodic motion of the beam centre
-0.001 0.000 0.001
-0.003
0.000
0.003
d
v
/
d
t
v( 0.5,t)

0 2500 5000
0.0
0.1
0.2
0.3

t

(a) the Poincar map (b) the largest Lyapunov exponent
Fig. 7. Chaotic motion of the beam centre
Nonlinear Vibrations of Axially Moving Beams

163
6. Energetics and conserved quantity
6.1 Energetics
It is well known that the total mechanical energy in free vibration of an undamped axially
stationary elastic beam with pinned or fixed ends is constant. However, many investigations
found that the total mechanical energy associated with free vibration of an axially moving
elastic beam is not constant even if the beam travels between two motionless supports.
(Barakat, 1968) considered the energetics of an axially moving beam and found that energy
flux through the supports can invalidate the linear theories of axially moving beams at
sufficiently high transporting speed. (Tabarrok, 1974) showed that the total energy of a
traveling beam without tension is periodic in time. (Wickert & Mote, 1989) presented the
temporal variation of the total energy related to the local rate of change and calculated the
temporal variation of energy associated with modes of moving beams. Considering the case
that there were nonconservative forces acting on two boundaries, (Lee & Mote 1997)
presented a generalized treatment of energetics of translating beams. (Renshaw et al., 1998)
examined the energy of axially moving beams from both Lagrangian and Eulerian views
and found that Lagrangian and Eulerian energy functionals are not conserved for axially
moving beams. (Zhu & Ni, 2000) investigated energetics of axially moving strings and
beams with arbitrarily varying lengths. (Chen & Zu, 2004) proposed energetics of axially
moving beams with geometric nonlinearity due to small but finite stretching of the beams.
Hence the variation of the total mechanical energy is a fundamental feature of free
transverse vibration of axially moving beams. However, all aforementioned investigations
on energetics and conserved quantities of axially moving beams have only been limited to
transverse vibration, in which longitudinal motion is assumed to be uncoupled and thus
neglectable. Actually, the energetics of coupled vibration of axially moving elastic strings
(Chen, 2006) can be extended to beams.
Assume that the axially moving beam described at the beginning of Subsection 2.1 is elastic
(=0), is without external excitations (f
u
=0, f
v
=0), and moves in a constant axially speed (=c).
Consider the total mechanical energy in a specified spatial domain, the span (0, L). The total
mechanical energy consists of the kinetic energy of all material particles and the potential
energy resulted from the initial tension, the disturbed tension, and the bending moment
caused by the beam deflection due to its motion
( ) ( )
2 2
2
0
0
1 1
, , , , , d
2 2 2
L
t x t x xx
A
c u cu v cv P EA EIv x

= + + + + + + +

E (38)

Then the time-rate of energy change is
( ) ( )
2 2
2
0
0
d 1 1
, , , , , d
d 2 2 2
L
t x t x xx
A
c u cu v cv P EA EIv x
t t




= + + + + + + +

E
(39)

Interchanging the order of differentiation and integration and inserting u,
tt
and v,
tt
solved
from equation (11), after some mathematical manipulations, one can express the time-rate of
energy change in the boundary values
Nonlinear Dynamics

164

( )
( )( )
( )
( )
( )
( )
( ) ( )
( ) ( )
0 0
2 2
2 2
0
2 2
2 2
0
0
1 , ,
, , , ,
d
1 , , 1 , ,
d
, , , , , ,
1 1 1
, , , , , .
2 2 2
l
x x
t x t x
x x x x
xx xt xx xxx t x
l
t x t x xx
P EA u P EA v
c u cu v cv
u v u v
t
EIv v cv EIv v cv
c A c u cu v cv P EA EIv


+ + +
+ + + +

=
+ + + +

+ + +



+ + + + + + +





E
(40)
Notice that

( )( )
( )
( )
( )
0 0
2 2
2 2
1 , ,
,
1 , , 1 , ,
x x
u v
x x x x
P EA u P EA v
P P
u v u v
+ + +
= =
+ + + +
(41)
are respectively the longitudinal and transverse components of the tension in the beam,
EIv,
xx
is the bending moment and EIv,
xxx
are the shear, while c+u,
t
+cu,
x
and v,
t
+cv,
x
are
respectively the absolute velocity in the longitudinal and transverse directions and v,
tx
+cv,
xx

is the absolute angle velocity. Hence the first term in equation (40) stands for the difference
of power of the beam tension, the beam bending moment, and the beam shear acting at two
ends. Meanwhile,
( ) ( )
2 2
2 2
0
1 1 1

, , , , ,
2 2 2
t x t x xx
A c u cu v cv P EA EIv


= + + + + + + +




E (42)
is the total mechanical energy per unit length. Hence the second term in equation (40) stands
for the energy change due to the axial motion of the beam. Physically, equation (40) means
the change rate of the energy consisting of two parts: the power of the beam tension,
moment and shear applying at two ends and the energy variation resulted from the axial
motion.
For a beam with the simple support (K=0 in equation (9)) or the fixed ends (K in
equation (9)), equation (40) leads to, respectively,

( )
2
0
d 1
1 , , , ,
d 2 2
L
xx xxx x
c EA Ac EI v v v
t



= + +


E
(43)

( )
2
0
, d
, 1 , .
d 2
L
x
x xx
u
c EA Ac u EIv
t


= + +


E
(44)
For an axially stationary beam, c=0. Equation (40) becomes

( )( )
( )
( )
( )
0 0
2 2
2 2
0
1 , ,
d
, , , , , , .
d
1 , , 1 , ,
L
x x
t t xx xt xxx t
x x x x
P EA u P EA v
u v EIv v EIv v
t
u v u v


+ + +

= + +

+ + + +


E
(45)
If the axially stationary beam is with pinned or fixed ends, equation (45) leads to the
conservation of the mechanical energy, which is a well known fact.
Nonlinear Vibrations of Axially Moving Beams

165
6.2 Conserved quantity
Although the total mechanical energy of axially moving beams is generally not constant,
there does exist an alternative conserved quantity. (Renshaw et al. 1998) presented both
Eulerian and Lagrangian conserved functionals for axially moving beams. (Chen & Zu,
2004) generalized their results to nonlinear free vibration of axially moving beams. They
adopted the partial-differential equation (a special case of equation (13)) for axially moving
beams undergoing nonlinear transverse vibration. (Chen & Zhao, 2005) also present a
conserved functional for a beam modeled by an integro-partial-differential equation derived
from the quasi-static assumption (a special case of equation (14)). They applied the
conserved functional to verify that the straight equilibrium configuration is stable for beams
at low axial speed.
Define the functional

( ) ( )
2 2 2 2 2 2 2
0
0
1 1
, , , , , d
2 2 2
L
t x t x xx
A
u c u v c v P EA EIv x

= + + + +

I (46)
Evaluation of the temporal differentiation by parts yield
( )
( )
( )
( )
( )
( )
( ) ( )
( )
2
2
0
3 2
2
2
0
2
2
0
3 2
2
2
1 , , , , , d
, , 2 , ,
d
1 , ,
1 , , 1 , , ,
, , 2 , , ,
1 , ,
L
x x xx x xx
t tt xt xx
x x
x xx x x xx
t tt xt xx xxxx
x x
u v v v u
u Au Acu Ac EA u EA P
t
u v
u v u v u
v Av Acv Ac EA u EIv EA P
u v




+

= + +

+ +



+ +
+ + + +

+ +

I
( ) ( )
( )( )
( )
( )
( )
0 0
0 2 2
2 2
0
d
1 , ,
, , , , , , , , , , .
1 , , 1 , ,
L
L
x x
t x t t x t t t xx xt
x x x x
x
P EA u P EA v
Ac u cu u v cv v u v EIv v
u v u v

+ + +

+ + + + + +


+ + + +


(47)
Substitution of equation (11) with =0, f
u
=0, f
v
=0, and =c into equation (47) leads to

( ) ( )
( )( )
( )
( )
( )
0
0 0
2 2
2 2
0
d
, , , , , ,
d
1 , ,
, , , , .
1 , , 1 , ,
L
t x t t x t
L
x x
t t xx xt
x x x x
Ac u cu u v cv v
t
P EA u P EA v
u v EIv v
u v u v

= + + +


+ + +

+ + +

+ + + +


I
(48)
At a pinned or fixed end, u,
t
=0, v,
t
=0, v,
xx
=0 or v,
x
=0 (hence v,
xt
=0). Therefore, equation (48)
results in dI/dt=0. There exists functional (46) that is conserved under pinned or fixed
boundary conditions for beams moving with a constant axial speed c.
The conserved quantity in a mechanical system is not only mathematically the first integral
leading to a reduction in the order of the system, but also reflects the physical essence of the
system closely related to the symmetries of the system. Therefore, it is theoretically
significant to investigate the conserved quantities. The conserved quantity in a mechanical
system can be used to check and develop numerical simulation algorithms. It is also useful
for stability analysis and controller design.
Nonlinear Dynamics

166
7. Concluding remarks
Because an axially moving beam is an effective mechanical model that can be used in
diverse engineering fields, many research activities in the area have been witnessed. The
chapter summarizes some resent works on modeling, analysis and simulations of nonlinear
vibrations of axially moving beams. It will remain to be an active research field. There are
many promising topics for future researches, including but surely not limited to the follows:
(1) modeling slender structures via sophisticated beam theories such as three-dimensional
beams or composite beams, (2) incorporating functionally graded, theromviscoelastic or
other advanced materials, (3) accounting for aerodynamic forces and heating and other
actions coupled with the vibration, (4) considering complex constraints and coupling such
as belts in drive systems, (5) developing analytical approaches especially for coupled
vibrations and strongly nonlinear vibrations, (6) investigating convergence, consistency, and
stability of numerical procedures, (7) exploring energetics of nonlinear and time-dependent
beams under general constraint conditions, (8) understanding complicated dynamical
behaviors such as global bifurcations, chaos, patterns, and spatio-temporal chaos.
8. Acknowledgments
This work was supported by the National Outstanding Young Scientists Foundation of
China (No. 10725209), the National Natural Science Foundation of China (No. 90816001), the
Specialized Research Fund for the Doctoral Program of Higher Education of China (Grant
No. 20093108110005), Shanghai Subject Chief Scientist Project (No. 09XD1401700), Shanghai
Leading Talent Program, Shanghai Leading Academic Discipline Project (No. S30106), and
the Program for Changjiang Scholars and Innovative Research Team in University (No.
IRT0844).
9. References
Abrate, A. S. (1992). Vibration of belts and belt drives. Mechanism and Machine Theory, 27, 6,
645-659, ISSN 0094-114X
Barakat, R. (1968). Transverse vibrations of a moving thin rod. The Journal of the Acoustical
Society of America, 43, 533-539, ISSN 0001-4966
Behdinan, K.; Stylianou M.C. & Tabarrok, B. (1997). Dynamics of flexible sliding beams
non-linear analysis part I: formulation. Journal of Sound and Vibration, 208, 4, 517-
539, ISSN 0022-460X
Bert, C. W. & Malik, M. (1996). The differential quadrature method in computational
mechanics: a review. Applied Mechanics Reviews, 49, 1-28, ISSN 0003-6900
Chakraborty, G. & Mallik, A.K. (1998). Parametrically excited nonlinear traveling beams
with and without external forcing. Nonlinear Dynamics, 17, 4, 301-324, ISSN 1573-
269X
Chakraborty, G. & Mallik, A.K. (1999). Stability of an accelerating beam. Journal of Sound and
Vibration, 227, 2, 309-320, ISSN 0022-460X
Nonlinear Vibrations of Axially Moving Beams

167
Chakraborty, G. & Mallik, A.K. (2000a). Wave propagation in and vibration of a travelling
beam with and without non-linear effects, part I: free vibration. Journal of Sound and
Vibration, 236, 2, 277-290, ISSN 0022-460X
Chakraborty, G. & Mallik, A.K. (2000b). Wave propagation in and vibration of a travelling
beam with and without non-linear effects, part II: forced vibration. Journal of Sound
and Vibration, 236, 2, 291-305, ISSN 0022-460X
Chakraborty, G.; Mallik, A.K. & Hatwal, H. (1999).Non-linear vibration of a travelling beam.
International Journal of Non-Linear Mechanics, 34, 655-670, ISSN 0020-7462
Chen, L.H.; Zhang, W. & Liu, Y.Q. (2007). Modeling of nonlinear oscillations for viscoelastic
moving belt using generalized Hamiltons Principle. ASME Journal of Vibration and
Acoustics, 129, 128-132, ISSN 1528-8927
Chen, L.Q. (2005). Analysis and control of transverse vibrations of axially moving strings.
Applied Mechanics Reviews, 58, 91-116, ISSN 0003-6900
Chen, L.Q. (2005). Principal parametric resonance of axially accelerating viscoelastic strings
constituted by the Boltzmann superposition principle, Proceedings of the Royal
Society of London A: Mathematical, Physical and Engineering Sciences, 461, 2061), 2701-
2720, ISSN 1471-2946
Chen, L.Q. (2006). The energetics and the stability of axially moving strings undergoing
planer motion. International Journal of Engineering Science, 44, 1346-1352, ISSN 0020-
7225
Chen, L.Q. & Chen, H. (2009). Asymptotic analysis on nonlinear vibration of axially
accelerating viscoelastic strings with the standard linear solid model. Journal of
Engineering Mathematics, accepted
Chen, L.Q.; Chen, H. & Lim, C.W. (2008). Asymptotic analysis of axially accelerating
viscoelastic strings. International Journal of Engineering Science, 46(10): 976-985, ISSN
0020-7225
Chen, L.Q. & Ding, H. (2008). Steady-state responses of axially accelerating viscoelastic
beams: approximate analysis and numerical confirmation. Science in China Series G:
Physics, Mechanics & Astronomy, 51(11): 1701-1721, ISSN 1862-2844
Chen, L.Q. & Ding H. (2009). Steady-state transverse response in planar vibration of axially
moving viscoelastic beams. ASME Journal of Vibration and Acoustics, in press, ISSN
1528-8927
Chen, L.Q.; Tang, Y.Q. & Lim, C.W. (2010). Dynamic stability in parametric resonance of
axially accelerating viscoelastic Timoshenko beams. Journal of Sound and Vibration,
329, 547-565, ISSN 0022-460X
Chen, L.Q. & Wang, B. (2009). Stability of axially accelerating viscoelastic beams: asymptotic
perturbation analysis and differential quadrature validation, European Journal of
Mechanics A/Solid, 28(4): 786-791, ISSN 0997-7538
Chen, L.Q.; Wang, B. & Ding, H. (2009). Nonlinear parametric vibration of axially moving
beams: asymptotic analysis and differential quadrature verification. Jorurnal of
Physics: Conference Series, 181, 012008, ISSN 1742-6596
Chen, L.Q. & Yang, X.D. (2005). Steady-state response of axially moving viscoelastic beams
with pulsating speed: comparison of two nonlinear models. International Journal of
Solids and Structures, 42, 37-50, ISSN 0020-7683
Nonlinear Dynamics

168
Chen, L.Q. & Yang, X.D. (2006a). Vibration and stability of an axially moving viscoelastic
beam with hybrid supports. European Journal of Mechanics A/Solids, 25, 996-1008,
ISSN 0997-7538
Chen, L.Q. & Yang, X.D. (2006b). Transverse nonlinear dynamics of axially accelerating
viscoelastic beams based on 4-term Galerkin truncation. Chaos, Solitons and Fractals,
2006, 27, 3, 748-757, ISSN 0960-0779
Chen, L.Q. & Yang, X.D. (2007). Nonlinear free vibration of an axially moving beam:
comparison of two models. Journal of Sound and Vibration, 299, 348-354, ISSN 0022-
460X
Chen, L.Q. & Zhao, W.J. (2005). A conserved quantity and the stability of axially moving
nonlinear beams. Journal of Sound and Vibration, 286, 663-668, ISSN 0022-460X
Chen, L.Q.; Zhang, W. & Zu, J.W. (2009). Nonlinear dynamics in transverse motion of axially
moving strings. Chaos, Solitons & Fractals, 40, 1, 78-90, ISSN 0960-0779
Chen, L.Q. & Zu, J.W. (2004). Energetics and Conserved Functional of Moving Materials
Undergoing Transverse Nonlinear Vibration. ASME Journal of Vibration and
Acoustics, 126, 452-455, ISSN 1528-8927
Chen, L.Q. & Zu, J.W. (2008). Solvability condition in multi-scale analysis of gyroscopic
continua. Journal of Sound and Vibration, 309, 338-342, ISSN 0022-460X
Chen, S.H.; Huang, J.L. & Sze, K.Y. (2007). Multidimensional Lindstedt-Poincar method for
nonlinear vibration of axially moving beams. Journal of Sound and Vibration, 306, 1-
11, ISSN, 0022-460X
DAngelo C. III; Alvarado, N.T.; Wang, K.W., and Mote, C.D.Jr. (1985). Current research on
circular saw and band saw vibration and stability. The Shock and Vibration Digest. 17,
5, 11-23, ISSN 1741-3184
Ding, H. & Chen, L.Q. (2008). Stability of axially accelerating viscoelastic beams: multi-scale
analysis with numerical confirmations. European Journal of Mechanics A/Solid, 27(6):
1108-1120, ISSN 0997-7538
Ding, H. & Chen, L.Q. (2009a). On two transverse nonlinear models of axially moving
beams. Science in China Series E: Technological Sciences, 52(3): 743-751, ISSN 1862-
281X
Ding, H. & Chen, L.Q. (2009b). Nonlinear dynamics of axially accelerating viscoelastic
beams based on differential quadrature. Acta Mechanica Sinica Solida, 22, 3, 267-275
ISSN 0894-9166
Feng, Z.H. & Hu, H.Y. (2002). Nonlinear dynamics modeling and periodic vibration of a
cantilever beam subjected to axial movement of basement. Acta Mechanica Solida
Sinica , 15, 2, 133-139, ISSN 0894-9166
Feng, Z.H. & Hu, H.Y. (2003). Principal parametric and three-to-one internal resonances of
flexible beams undergoing a large linear motion. Acta Mechanica Sinica, 19, 4, 355-
364, ISSN 0567-7718
Ghayesh, M.H. (2008). Nonlinear transversal vibration and stability of an axially moving
viscoelastic string supported by a partial viscoelastic guide. Journal of Sound and
Vibration, 314, 757-774, ISSN 0022-460X
Nonlinear Vibrations of Axially Moving Beams

169
Ghayesh, M.H. & Balar, S. (2008). Non-linear parametric vibration and stability of axially
moving visco-elastic Rayleigh beams. International Journal of Solids and Structures,
45, 6451-6467, ISSN 0020-7683
Ghayesh, M.H. & Khadem, S.E. (2008). Rotary inertia and temperature effects on non-
nonlinear vibration, steady-state response and stability of an axially moving beam
with time-dependent velocity. International Journal of Mechanical Sciences, 50, 389-
404, ISSN 0020-7403
Hochlenert, D.; Spelsberg-Korspeter, G. & Hagedorn, P. (2007). Friction induced vibrations
in moving continua and their application to brake squeal. ASME Journal of Applied
Mechanics, 74, 542-549, ISSN 1528-9036
Humer, A. & Irschik, H. 2009. Onset of transient vibrations of axially moving beams with
large displacements, finite deformations and an initially unknown length of the
reference configuration. Zeitschrift fur Angewandte Mathematik und Mechanik, 89, 4,
267-278, ISSN 0044-2267
Hwang, S.-J. & Perkins, N. C. (1992a). Supercritical stability of an axially moving beam part
1: model and equilibrium analysis. Journal of Sound and Vibration, 154, 381-396, ISSN
0022-460X
Hwang, S.-J. & Perkins, N. C. (1992b). Supercritical stability of an axially moving beam part
2: vibration and stability analysis. Journal of Sound and Vibration, 154, 397-409, ISSN
0022-460X
Hwang, S.-J. & Perkins, N. C. (1994). High speed stability of coupled band/wheel systems:
theory and experiment. Journal of Sound and Vibration, 169, 459-483, ISSN 0022-460X
Koivurova, H. & Salonen, E.M. (1999). Comments on nonlinear formulations for travelling
string and beam problems. Journal of Sound and Vibration, 225, 5, 845-856, ISSN 0022-
460X
Kong, L. & Parker, R. G. (2004). Coupled belt-pulley vibration in serpentine drives with belt
bending stiffness. ASME Journal of Applied Mechanics, 71, 109-119, ISSN 1528-9036
Lee, S. . & Mote, C.D.Jr. (1997). A generalized treatment of the energetics of translating
continua, part 2: beams and fluid conveying pipes, Journal of Sound and Vibration
204, 735-753, ISSN 0022-460X
Marynowski, K. (2002). Non-linear dynamic analysis of an axially moving viscoelastic beam.
Journal of Theoretical And Applied Mechanics, 40, ISSN 465-482, 0973-6085
Marynowski, K. (2004). Non-linear vibrations of an axially moving viscoelastic web with
time-dependent tension. Chaos, Solitons & Fractals, 21, 481-490, ISSN 0960-0779
Marynowski, K. (2006). Two-dimensional rheological element in modelling of axially
moving viscoelastic web. European Journal of Mechanics A/Solids, 25, 729-744, ISSN
0997-7538
Marynowski, K. & Kapitaniak, T. (2002). Kelvin-Voigt versus Bgers internal damping in
modeling of axially moving viscoelastic web. International Journal of Non-Linear
Mechanics, 37, 1147-1161, ISSN 0020-7462
Marynowski, K. & Kapitaniak, T. (2007). Zener internal damping in modelling of axially
moving viscoelastic beam with time-dependent tension. International Journal of Non-
Linear Mechanics, 42, 118-131, ISSN 0020-7462
Nonlinear Dynamics

170
Mockensturm, E. M. & Guo, J. (2005). Nonlinear vibration of parametrically excited,
viscoelastic, axially moving strings. ASME Journal of Applied Mechanics, 72, 374-380,
ISSN 1528-9036
Mote C.D.Jr. (1972). Dynamic stability of axially moving materials. The Shock and Vibration
Digest. 4, 4, 3-13, ISSN 1741-3184
Mote C.D.Jr.; Schajer, G.S. & Wu, W.Z. (1982). Band saw and circular saw vibration and
stability. The Shock and Vibration Digest. 14, 2, 19-25, ISSN 1741-3184
Maccari, A. (1999). The asymptotic perturbation method for nonlinear continuous systems.
Nonlinear Dynamics, 19, 1-18, ISSN 1573-269X
z, H.R.; Pakdemirli, M. & Boyaci, H. (2001). Non-linear vibrations and stability of an
axially moving beam with time-dependent velocity. International Journal of Non-
Linear Mechanics, 36, 107-115, ISSN 0020-7462
zhan, B. B. and Pakdemirli, M. (2009) A general solution procedure for the forced
vibrations of a continuous system with cubic nonlinearities: Primary resonance
case. Journal of Sound and Vibration, 325, 894-906, ISSN 0022-460X
Parker, R. G. & Lin, Y. (2001). Parametric instability of axially moving media subjected to
multifrequency tension and speed fluctuations. ASME Journal of Applied Mechanics,
68, 49-57, ISSN 1528-9036
Pellicano, F.; Fregolent, A.; Bertuzzi, A. & Vestroni F. (2001). Primary and parametric non-
linear resonances of a power transmission belt. Journal of Sound and Vibration, 244,
669-684, ISSN 0022-460X
Pellicano, F. & Vestroni, F. (2000). Nonlinear dynamics and bifurcations of an axially moving
beam. ASME Journal of Vibration and Acoustics, 122, 21-30, ISSN 1528-8927
Pellicano, F. & Vestroni, F. (2002). Complex dynamic of high-speed axially moving systems.
Journal of Sound and Vibration, 258, 31-44, ISSN 0022-460X
Pellicano, F. & Zirilli, F. (1997). Boundary layers and non-linear vibrations in an axially
moving beam. International Journal of Non-Linear Mechanics, 33, 691-711, ISSN 0020-
7462
Pratiher, B. & Dwivedy, S.K. (2008). Non-linear vibration of a single link viscoelastic
Cartesian manipulator. International Journal of Non-Linear Mechanics, 43, 683-696,
ISSN 0020-7462
Ravindra, B. & Zhu, W.D. (1998). Low dimensional chaotic response of axially accelerating
continuum in the supercritical regime. Archive of Applied Mechanics, 68, 195-205,
ISSN 1432-0681
Renshaw, A.A.; Rahn, C.D.; Wickert, J. & Mote, C.D.Jr. (1998). Energy and conserved
functionals for axially moving materials. ASME Journal of Vibration and Acoustics,
120, 634-636, ISSN 1528-8927
Riedel, C. H. & Tan, C. A. (2002). Coupled, forced response of an axially moving strip with
internal resonance. International Journal of Non-Linear Mechanics, 37, 101-116, ISSN
0020-7462
Spelsbrg-Korspeter, G; kirillov, & O.N. Hagedorn, P. (2008). Modeling and stability analysis
of an axially moving beam with frictional contact. ASME Journal of Applied
Mechanics, 75, 031001, ISSN 1528-9036
Nonlinear Vibrations of Axially Moving Beams

171
Shu, C. (2001). Differential Quadrature and Its Application in Engineering. Springer, ISBN 978-1-
85233-209-9, Berlin
Sze, K.Y.; Chen, S.H. & Huang, J.L. (2005). The incremental harmonic balance method for
nonlinear vibration of axially moving beams. Journal of Sound and Vibration, 281,
611-626, ISSN 0022-460X
Tang Y.Q.; Chen, L.Q., & Yang X.D. (2008). Natural frequencies, modes and critical speeds of
axially moving Timoshenko beams with different boundary conditions.
International Journal of Mechanical Sciences, 50 (10-11): 1448-1458, ISSN 0020-7403
Tang, Y.Q.; Chen, L.Q. & Yang, X.D. (2009). Non-linear vibrations of axially moving
Timoshenko beams under weak and strong external excitations. Journal of Sound and
Vibration, 320, 4/5, 1078-1099, ISSN 0022-460X
Tang, Y.Q.; Chen, L.Q. & Yang X.D. (2010). Parametric resonance of axially moving
Timoshenko beams with time-dependent speed. Nonlinear Dynamics, 58, 715-724,
ISSN 1573-269X
Thurman, A. L. & Mote Jr. C. D. (1969). Free, periodic, nonlinear oscillation of an axially
moving strip. ASME Journal of Applied Mechanics 36, 83-91, ISSN 1528-9036
Tabarrok, B., Leech, C. M. & Kim, Y. I. (1974). On the dynamics of an axially moving beam.
Journal of The Franklin Institute, 297, ISSN 201-220, 0016-0032
Ulsoy, A.G. & Mote, C.D.Jr. (1978). Band saw vibration and stability. The Shock and Vibration
Digest. 10, 1, 3-15, ISSN 1741-3184
Vu-Quoc, L. & Li, S. (1995). Dynamics of sliding geometrically-exact beams: large angle
maneuver and parametric resonance. Computer Methods in Applied Mechanics and
Engineering, 120, 1995, ISSN 0045-7825
Wang, B. & Chen, L.Q. (2009). Asymptotic stability analysis with numerical confirmation of
an axially accelerating beam constituted by the standard linear solid viscoelastic
model. Journal of Sound and Vibration, 328, 456-466, ISSN 0022-460X
Wang, K. W. (1991). Dynamic stability analysis of high speed axially moving bands with end
curvatures. ASME Journal of Vibration and Acoustics, 113, 62-68, ISSN 1528-8927
Wang, K. W. & Mote C. D. Jr. (1986). Vibration coupling analysis of Band/wheel mechanical
systems. Journal of Sound and Vibration. 109, 237-258, ISSN 0022-460X
Wang, K. W. & Mote C. D. Jr. (1987). Band/wheel system vibration under impulsive
boundary excitation. Journal of Sound and Vibration. 115, 203-216, ISSN 0022-460X
Wang, X. & Bert, C. W. (1993). A new approach in applying differential quadrature to static
and free vibration analyses of beams and plates. Journal of Sound and Vibration, 162,
566-572, ISSN 0022-460X
Wickert, J.A. (1992). Non-linear vibration of a traveling tensioned beam. International Journal
of Non-Linear Mechanics, 27, 503-517, ISSN 0020-7462
Wickert, J.A. & Mote, C.D.Jr. (1989). On the energetics of axially moving continua. The
Journal of the Acoustical Society of America, 85, 1365-1368, ISSN 0001-4966
Wickert, J.A. & Mote, C.D.Jr. (1990). Classical vibration analysis of axially moving continua.
ASME Journal of Applied Mechanics, 57, 738-744, ISSN 1528-9036
Wickert, J.A. & Mote, Jr. C. D. (1988). Current research on the vibration and stability of
axially-moving materials. The Shock and Vibration Digest. 20, 5, 3-13, ISSN 1741-3184
Nonlinear Dynamics

172
Yang, T.Z.; Fang, B.; Chen, Y. & Zhen, Y.X. (2009). Approximate solutions of axially moving
viscoelastic beams subject to multi-frequency excitations. International Journal of
Non-Linear Mechanics, 44, 240-248, ISSN 0020-7462
Yang, X. D. & Chen, L.Q. (2005). Bifurcation and chaos of an axially accelerating viscoelastic
beam. Chaos, Solitons & Fractals, 23(1): 249-258, ISSN 0960-0779
Yang, X. D. & Chen, L.Q. (2006). Non-linear forced vibration of axially moving viscoelastic
beams. Acta Mechanica Solida Sinica, 19, 4, 365-373, ISSN 0894-9166
Zhang, L. & Zu, J.W. (1999). Nonlinear vibration of parametrically excited moving belts.
ASME Journal of Applied Mechanics, 66, 396-402, ISSN 1528-9036
Zhang, W. & Song, C.Z. (2007). Higher-dimensional periodic and chaotic oscillations for
viscoelastic moving belt with multiple internal resonances. International Journal of
Bifurcation and Chaos, 17, 1637-1660, ISSN 1793-6551
Zhu, W.D. (2000). Vibration and stability of time-dependent translating media. The Shock and
Vibration Digest. 32, 5, 369-379, ISSN 1741-3184
Zhu, W.D. & Ni, J.(2000). Energetics and stability of translating media with an arbitrarily
varying length. ASME Journal of Vibration and Acoustics, 122, 295-304, ISSN 1528-
8927
8
The 3D Nonlinear Dynamics of Catenary Slender
Structures for Marine Applications
Ioannis K. Chatjigeorgiou and Spyros A. Mavrakos
National Technical University of Athens
Greece
1. Introduction
Riser systems are inextricable parts of integrated floating production and offloading systems
as they are used to convey oil from the seafloor to the offshore unit. Risers are installed
vertically or they are laid obtaining a catenary configuration. From the theoretical point of
view they can be formulated as slender structures obeying to the principles of the Euler-
Bernoulli beams. Riser-type catenary slender structures and especially Steel Catenary Risers
(SCRs) attract the attention of industry for many years as they are very promising for deep
water applications. According to the Committee V.5 of the International Ship and Offshore
Structures Congress (ISSC, 2003), flexible risers have been qualified to 1500m and are
expected to be installed in depths up to 3000m in the next few years. In such huge depths
where the suspended length of the catenary will unavoidably count several kilometers, the
equivalent elastic stiffness of the structure will be quite low enabling large displacements.
The later remark implies that even small excitations could cause significant excursions in
both in-plane and out-of-plane directions. Therefore a 2D formulation, although adequate in
predicting the associated dynamics in the reference plane of the static equilibrium, it would
be certainly a short approximation.
Furthermore, in deep water installations, for practical reasons mainly, the riser should be
configured nearly as a vertical structure in order to avoid suspending more material. The
nearly vertical configuration which ends in a sharp increase of the curvature close to the
bottom, results in extreme bending moments at the touch down region. The static bending
moment which is applied in the plane of reference of the catenary is further amplified due to
the imposed excitation set by the motions of the floating structure. It has been generally
acknowledged that the heave motion is the worst loading condition as it causes several
effects, which depending on the properties of the excitation, can be applied individually or
in combination between each other. Indicative examples are the seafloor interaction,
buckling-like effects, compression loading and heave induced out-of-plane motions.
For the formulation of the seafloor interaction, various approaches have been proposed and
it appears that the associated effects continue to attract the attention of the research
community (Leira et al., 2004; Aubeny et al., 2006; Pesce et al., 2006; Clukey et al., 2008).
Compression loading has been studied mainly in 2D (Passano & Larsen, 2006 & 2007;
Chatjigeorgiou et al., 2007; Chatjigeorgiou, 2008), while buckling-like effects and possible
Nonlinear Dynamics

174
destabilizations are mainly considered for completely vertical structures (Kuiper &
Mertikine, 2005; Gadagi & Benaroya, 2006; Chandrasekaran et al., 2006; Kuiper et al., 2008).
The content of the present work falls in the last category of the effects that were mentioned
previously. The main concern of the study is to identify the details of the out-of-plane
response which is induced due to motions imposed in the catenarys plane of reference and
in particular due to heave excitation. Relevant effects called as Heave Induced Motions
have been investigated experimentally in the past by Joint Industry Projects (JIP). According
to HILM (Heave Induced Lateral Motions of Steel Catenary Risers) JIP led by Institut
franais du ptrole (Ifp), the phenomenon was first recorded during the HCR (Highly
Compliant Riser Large Scale Model Tests) JIP led by PMB Engineering, in which a steel
catenary riser was excited by heave motion in a stillwater lake. The pipe was subjected to
out-of-plane cyclic motions. The same behaviour was observed during the HILM JIP
measurements (LeCunff et al., 2005).
Apparently, the associated phenomena can be captured numerically only by treating the
governing 3D dynamical system. To this end, the associated system is properly elaborated
and solved numerically using an efficient finite differences numerical scheme.
2. Definitions
A fully immersed catenary slender structure is considered. The catenary is modeled as an
Euler-Bernoulli slender beam, having the following geometrical and physical properties:
suspended length L, outer diameter d
o
, inner diameter d
i
, submerged weight w
o
, mass m,
hydrodynamic mass m
a
, cross sectional area A and moment of inertia I. The quantities d
o
, d
i
,
A and I, correspond to the unstretched condition, while w
o
, m and m
a
are defined per unit
unstretched length. The Young modulus of elasticity is denoted by E and accordingly EA
and EI define the elastic and bending stiffness respectively. Finally, it is assumed that the
catenary conforms to a linear stress-strain relation.
Next the generalized motion and loading vectors (Fig. 1) are defined. These are
[ ]
T
w v u t s V = ) ; (
K
(1)
[ ]
T
n b b n
M M S S T t s F = ) ; (
G
(2)
where u, v, w are the tangential (axial), normal and bi-normal velocities, respectively, is
the Eulerian angle which is formed between the tangent of the line and the horizontal in the
reference plane of the catenary, is the Eulerian angle in the out-of-plane direction, T is the
tension, S
n
and S
b
are the in-plane and the out-of-plane shear forces and finally M
b
and M
n

are the bending moments around the corresponding Lagrangian axes b
G
and n
G
, namely the
generalized loading that causes bending in the in-plane and the out-of-plane direction,
respectively. The moments M
b
and M
n
are associated with the corresponding curvatures
b

and
n
according to M
j
=EI
j
, for j=n,b.
In the general case where steady current is presented, the relative velocities should be
considered. These are written as
t tr
U u v = ,
n nr
U v v = and
b br
U w v = , where U
t
, U
n

and U
b
are the components of the steady current parallel to t
G
, n
G
and b
G
, respectively. The
elements of the vectors defined through Eqs. (1) and (2) are all functions of time t and the
unstretched Lagrangian coordinate s.
The 3D Nonlinear Dynamics of Catenary Slender Structures for Marine Applications

175
3. Dynamic system
The 3D dynamic equilibrium of the submerged catenary is governed by ten partial
differential equations. These equations are provided in the following without further details
on the derivation procedure. For more details the reader is referenced to the works of
Howell (1992), Burgess (1993), Triantafyllou (1994) and Tjavaras et al. (1998).

dt b n n b
R w S S
s
T
t
v
t
w
t
u
m + +

cos sin cos
0
(3)
( ) ( )
dn b b
n nr
a
R w S T
s
S
t
v
m w u
t t
v
m + + +

cos tan sin cos

0
(4)

db n b n
b br
a
R w T S
s
S
t
v
m
t
u
t
v
t
w
m +

sin sin tan sin

0
(5)
v w
s
u
t
T
EA
b n
+

1
(6)
( )

tan cos 1 w u
s
v
t EA
T
b
+ +

+ (7)
u v
s
w
t EA
T
n b

tan 1 (8)

3
2
1 tan

+ + =

EA
T
S EI
s
EI
b
b
n
(9)

3
1 tan

+ =

EA
T
S EI
s
EI
b b n
b
(10)

n
s
=

(11)

b
s
=

cos (12)
In Eqs. (3)-(5) R
dt
, R
dn
and R
db
denote the nonlinear drag forces which are expressed using
the Morisons formula. Thus,
( )
2 / 1
1
2
1
e v v C d R
tr tr dt o dt
+ = (13)
( )
2 / 1
2 / 1
2 2
1
2
1
e v v v C d R
br
nr nr dn o dn
+ + = (14)
Nonlinear Dynamics

176
( )
2 / 1
2 / 1
2 2
1
2
1
e v v v C d R
br
nr br db o db
+ + = (15)
where C
dt
, C
dn
and C
db
are the drag coefficients in tangential, normal and bi-normal
directions respectively. Normally, for a cylindrical structure, the in-plane and the out-of-
plane drag coefficients are equal while the tangential coefficient is very small and the
associated term can be ignored without loss of accuracy. Finally, e denotes the axial strain
deformation, which for a linear stress-strain relation is written as e=T/EA.
4. Numerical solution of the governing system using finite differences
The numerical method employed herein, is the finite differences box approximation
(Hoffman, 1993). Unlike the very popular finite element methods, the existing works which
are related to the application of numerical approximations that rely on finite differences,
concern mainly the dynamics of cables and mooring lines which have a negligible bending
stiffness (Burgess, 1993; Tjavaras et al., 1998; Ablow & Schechter, 1983; Howell, 1991;
Chatjigeorgiou & Mavrakos, 1999 & 2000; Gobat & Grosenbaugh, 2001 & 2006; Gobat et al.,
2002). The employment of the bending stiffness in mathematical formulations of cable
dynamics is done for special applications such as low tension cables, towing cables, highly
extensible cables and mooring lines in which the cycling loading leads to slacking
conditions, i.e. cancellation of the total tension.
With regard to the studies on pipes, for which the omission of the bending stiffness will
unavoidably lead to loss of important information, the finite differences approximation has
been used mainly for the solution of the static equilibrium problem (Zare & Datta, 1988; Jain
1994) or as a numerical scheme for the integration in the time domain, alternative to
Houbolt, Wilson- and Newmark- methods (Patel & Seyed, 1995). As far as the dynamic
equilibrium problem is concerned, box approximation has been employed recently by
Chatjigeorgiou (2008) for the development of a solution tool that treats the two dimensional
nonlinear dynamics of marine catenary risers.
For the governing system at hand (Eqs. (3)-(12)), the recommended procedure for employing
a finite differences approximation requires that the set of equations should be first cast in a
matrix-vector form. Thus, the concerned equations are written as
0 ) (
, ,
= +

t s
s t
Y F
Y
K
Y
M (16)
where [ ]
T
Y
b n n b
S S T w v u = . The mass and stiffness matrices, M
and K, and the forcing vector F are defined in Appendix A.
Next, Eq. (16) is discretized in both time and space using the finite differences box
approximation. This is the approach taken by several authors mentioned in the references
section of the present work. With this scheme, the discrete equations are written using what
look like traditional backward differences, but because the discetization is applied on the
half-grid points the method is second-order accurate. The result is a four point average,
centered around the half-grid point. Thus, Eq. (16) becomes
( ) ( )
1 1
1 1 1 1
1 1
Y Y Y Y
M M M M
i i i i
i i i i k k k k
k k k k
t t
+ +
+ +


+ + +




The 3D Nonlinear Dynamics of Catenary Slender Structures for Marine Applications

177

( ) ( )
( )
1 1
1 1 1 1
1 1
1 1
1 1
Y Y Y Y
K K K K
F F F F 0
i i i i
i i i i k k k k
k k k k
i i i i
k k k k
s s
+ +
+ +

+ +


+ + + +



+ + + + =
(17)
According to the matrix-vector Eq. (17) the governing partial differential equations are
defined in the center of [i,i+1] and [k-1,k], namely at [i+1/2, k-1/2]. The subscripts k define
the spatial grid points (the nodes) and the superscripts i define the temporal grid points (the
time steps). For n nodal points (k=1 corresponds to the touch down point at s=0 and k=n
corresponds to the top terminal point where the excitation is applied) Eq. (17) defines a
system of 10(n-1) equations to be solved for the 10n dependent variables at time step i+1.
The ten equations needed to complete the problem are provided by boundary conditions.
The algebraic equivalents of the governing Eqs. (3)-(12) are derived using the grid
transformation proposed by Eq. (17). The associated algebraic equations are given in
Appendix B of the present paper. The boundary conditions which are needed to complete
the final 10n algebraic system correspond to zero bending moments at both ends of the
catenary, zero motions at the bottom fixed point and specified time depended excitations at
the top in three directions. The final system is solved efficiently by the relaxation method.
5. Discussion on the contribution of the nonlinearities
The nonlinearities involved in the problem are either geometric or hydrodynamic
nonlinearities. Here the current is ignored and accordingly, the hydrodynamic action is
represented by the nonlinear drag forces induced due to the motions of the structure. It is
noted that the presence of current could stimulate possible vortex-induced-vibration
phenomena, the study of which exceeds the purposes of the present contribution. In
addition the structure is slender and therefore the diffraction phenomena are negligible.
This makes the drag forces the most determinative factor of hydrodynamic nature. Other
hydrodynamic effects involved in the problem are the added inertia forces which are
expressed through the added mass coefficients in the normal and the bi-normal directions.
Apart from the drag forces the dynamic equilibrium of the catenary involves also geometric
nonlinearities. Apparently, the most important are the internal loading-curvature terms. The
term internal loading refers to the tension and the shear forces. The question which easily
arises is how nonlinear contributions influence the motions of the structure, namely the
axial, the normal and the bi-normal displacements. It is evident that any excitation will
induce displacements in the same direction but the question herein concerns the details of
the motions which are induced in the other directions. The later remark is intimately
connected with the so called compression loading, i.e. the amplification of the bending
moments at the touch down region due to the dynamic components. The importance of the
subject regarding the in-plane bending moment has been extensively discussed by Passano
and Larsen (2006) and Chatjigeorgiou et al. (2007). Here the discussion is extended to the
out-of-plane bending moments as well.
In order to distinguish between the linear and the nonlinear effects it is indispensable to go
through the equivalent linearized dynamic problem. It is assumed that the generalized
loading terms and the Eulerian angles consist of a static and a dynamic component. These
will be denoted in the sequel by the indexes 0 and 1 respectively. In addition small motions
are considered. Thus the velocities are given by u=p/t, v=q/t and w=r/t, where p, q
Nonlinear Dynamics

178
and r are the motions in the axial, normal and bi-normal directions. Thus, the vector of the
unknowns of the linear problem [ ]
T
b n n b
S S T r q p t s Y = ) ; (
G
becomes
) ; ( ) ( ) ; (
1 0
t s Y s Y t s Y
G G
+ = (18)
where
[ ]
T
0 0 0 0 0 0 0 0
0 0 0 ) (
b n n b
S S T s Y = (19)
and
[ ]
T
1 1 1 1 1 1 1 1
) ; (
b n n b
S S T r q p t s Y =
G
(20)
The linearization procedure is outlined succinctly in the following. First, Eq. (18) is
introduced into the nonlinear system of Eqs. (3)-(12). After short mathematical
manipulations it can be seen that the resulting products will include the terms that define
the static equilibrium problem as well as nonlinear components. Static equilibrium terms
cancel each other while in the context of the linearized problem, the nonlinear terms are
ignored. The compatibility relations given by Eqs. (6)-(8), are integrated with respect to time
t. Finally, it is noted that the static terms
n0
,
0
and S
b0
are zero. This is due to the two-
dimensional static configuration of the catenary.
By employing the above procedure, the system of Eqs. (3)-(12) is reduced to the equivalent
linearized system.

1 0 0 0 1 1 0
1
2
2
cos w S S
s
T
t
p
m
b n b n

(21)
( )
t
q
q c w T T
s
S
t
q
m m
n b b
n
a

+ + +

+
1 0 0 0 1 0 1
1
2
2
sin (22)
( )
t
r
r c w T S
s
S
t
r
m m
b n b n
b
a

+
1 0 0 1 0 1 0 0
1
2
2
sin (23)

= q
s
p
EA T
b0 1
(24)
p
s
q
b0 1
+

= (25)

s
r

=
1
(26)

1 1 0
1
b b
n
S EI
s
EI =

(27)
The 3D Nonlinear Dynamics of Catenary Slender Structures for Marine Applications

179

1
1
n
b
S
s
EI =

(28)

1
1
n
s
=

(29)

1
1
b
s
=

(30)
In Eqs. (22) and (23) c
n
=4/(3)C
dn
d
o
and c
b
=4/(3)C
db
d
o
denote the linearized damping
coefficients which are determined through the linearization process of the nonlinear drag
forces R
dn
and R
db
. Also, the drag force in tangential direction was considered negligible,
whereas the elastic strain e was set equal to zero.
Eqs. (21)-(30) consists of two major groups, namely one set that governs the coupled axial
and normal motions (Eqs. (21), (22), (24), (25), (28) and (30)) and one set that governs the bi-
normal or out-of-plane motions (Eqs. (23), (26), (27) and (29)). Provided that the solution of
the static equilibrium problem is known, the two systems can be treated separately, which
implies that, at least in the context of the linear problem, the in-plane motions do not
influence the out-of-plane motions and vise versa. Thus, the axial and normal motions
induced out-of plane vibrations is only due to the nonlinear terms and especially due to the
geometric nonlinearities. This can be traced back to the fact that the out-of-plane static
components
n0
, S
b0
and
0
, were assumed equal to zero. In fact, this is the actual case when
the structure is perfect with no initial deformations, even marginal, and the excitations
coincide absolutely with the unit vectors t
G
and n
G
for the in-plane motions and b
G
for the
out of plane motions.
For the linear problem, which by default assumes that the motions are relatively small, the
in-plane and out-of-plane motions and their consequences, as regards the moments, the
shear forces and the tension, can be considered uncoupled without loss of accuracy.
Nevertheless, this is not a valid approach for the nonlinear problem. For a perfect structure
however and assuming only in-plane excitations it will be easy to confirm, through the
solution of the dynamic problem, that no out-of-plane motions are induced. This is a
shortcoming of the theoretical methods which is associated with the disability to represent
the marginal structural imperfections of the static configuration. However it is no difficult to
invent numerical tricks to override this practical problem. In the present contribution for
example, the numerical results which refer to the heave excitation induced out-of-plane
motions, were obtained by exciting the structure at the top with a combined motion that
consists of a vertical and a bi-normal component. The later is applied for a limited amount
of time, which is enough to produce non-zero out-of-plane angles, bending moments and
shear forces. Thus, at the cut-off time step the structure has obtained a 3D shape that
explicitly diverges from the perfect in-plane configuration and is accordingly used as the
initial condition for the subsequent time steps of the numerical simulation.
6. Numerical results and discussion
The numerical results which are presented in the following refer to the SCR that was used as
a model by Passano and Larsen (2006). The same model was employed also by
Chatjigeorgiou (2008). The physical and geometrical properties of the structure are: outer
Nonlinear Dynamics

180
diameter 0.429m, wall thickness 0.0022m, Young modulus of elasticity 207GPa, mass per
unit unstretched length 262.9kg/m, added mass per unit unstretched length 148.16kg/m,
submerged weight per unit unstretched length 915.6N/m, suspended length 2024m, elastic
stiffness 0.582310
10
N and bending stiffness 0.120910
9
Nm
2
. The drag coefficients in normal
and bi-normal directions were assumed equal to unity while the tangential drag coefficient
was set equal to zero. Finally, with regard to the installation characteristics, the catenary was
assumed suspended in water depth 1800m by applying a pretension at the top equal to
1860kN.
This work focuses mainly on the out-of-plane dynamics of the catenary, induced due to both
in-plane and out-of-plane motions. More interesting from the academic point of view is the
former type of excitation as in this case the out-of-plane motions are driven by nonlinearities.
6.1 Bi-normal (sway) excitation
Normally, nonlinear phenomena are stimulated at high frequencies and large amplitudes or
by combining both properties, at high excitation velocities. Therefore in order to expose and
study the associated impacts, the structure should be subjected to relatively severe loading.
The details of the sway excitation are examined having the structure excited with a
harmonic motion at the top with amplitude y
a
=1.0m and circular frequency =2.0rad/s.
The solution in the time domain and especially the one that accounts for the nonlinear terms
calculates the time histories of all time varying components at any point along the structure,
providing huge data records, which admittedly, are hard to be handled. In addition, in a
nonlinear formulation the records of the output signals will contain the contribution of sub-
and super-harmonics which are difficult to be identified by inspecting only the time
histories. Therefore, in order to present the results in a friendly and understandable format,
all records were processed using Fast Fourier Transformation (FFT) and adopted to 3D
spectrums. The spectrums reveal the prevailing frequencies at any point along the catenary.
For the test case mentioned before, the 3D spectrums for the dynamic tension T
1
, the normal
velocity v, the in-plane dynamic bending moment M
b1
, and the out-of-plane dynamic
bending moment M
n1
are depicted respectively in Figs. 2-5. It is noted that the out-of-plane
dynamic bending moment also represents the total out-of-plane bending moment as the
corresponding static counterpart is zero.
Fig. 5 shows that the out-of-plane bending moment responds at the excitation frequency.
This occurs for all points along the catenary. The maximum value occurs just before the top
terminal point where the excitation is applied. In addition, the variation of the out-of-plane
bending moment as a function of s exhibits a dentate configuration with a notable increase
at the touch down area. It is also important to note that no other harmonics are stimulated
and the response is restricted to the frequency of excitation only.
Figs. 2-4 demonstrate that the in-plane response due to the sway excitation is much more
complicated as various harmonics are detected. The most significant contribution comes
from the double of the excitation frequency (4.0rad/s) while it is visually evident that there
are peaks at 1/2, 3/2, 2, 5/2 and so on. The non-zero values of the spectral densities
for 0 or T, which exhibit a different pattern for the various dynamic components,
imply that the sway excitation causes a quasi-static application of the corresponding
component. In addition, the non-zero values for T, manifest that the response is in
general non periodic and it is composed by a fundamental frequency that tends to infinity
and practically a boundless number of harmonics.
The 3D Nonlinear Dynamics of Catenary Slender Structures for Marine Applications

181
6.2 In-plane heave excitation induced out-of-plane response
Here a single excitation case is examined that refers to excitation amplitude in heave
z
a
=1.0m with circular frequency =1.5rad/s. Again, a relatively high excitation velocity was
assumed, in order to investigate the effect of nonlinearities. In the specific static
configuration the heave motion acts nearly as an axial loading which, depending on the
conditions, may result in compression loading.
The details of the in-plane and the out-of-plane response due to the applied heave excitation
are examined with the aid of Figs. 6-19. Figs. 6-8 are given as a part of the discussion, started
in section 5, on the dependence of the out-of-plane motions, shear forces and bending
moments by the initial static configuration. Figs. 6-7 demonstrate a dependence of the
concerned variables on the amplitude of the sway excitation that is applied for practical
reasons and for a short time, just to provide an initial out-of-plane deformation to the
structure. Apparently, the records of the response, which in the specific case correspond to
the location where the maximum static bending moment M
b0
occurs, are different for
different amplitudes. Nevertheless, the output signals converge for large amplitudes. The
attainment of convergence is better shown in shear force S
b1
(Fig. 8), as in this case the
associated time history contains abnormal signals which however, do not dilute periodicity.
Nevertheless, it should be noted that the impotence to formulate accurately the marginal
static deformations in the out-of-plane direction, which it turn leads to the necessity to apply
artificially non-zero values of M
n0
(s) and
0
(s), constitutes in this connection, a numerical
uncertainty.
Next, we focus for a while in Figs. 6-8. Fig. 8 is a little bit confusing whereas a careful
inspection in Fig. 7 indicates the existence of a base harmonic and an additional harmonic.
The two harmonics are more evident in the time history of the out-of-plane velocity w (Fig.
6) and it can be shown that they correspond to 0.75rad/s and 2.25rad/s. In other words
none of the harmonics coincides with the excitation frequency. In particular, the concerned
harmonics correspond to 1/2 and 3/2 where is the frequency of the excitation.
Apparently the occurrence of these harmonics makes the motion of the structure quite
complicated. The latter remark is graphically shown in Figs. 9-11 which demonstrate the
path that is followed (in particular by node no 3 in a discretization grid of 100 nodes at
s=41m from touch down point) as seen from behind (v=f(w)), from above (u=f(w)) and from
the side (v=f(u)), respectively. It is noted that in Figs. 9-11 v and u respond following the
excitation frequency while w responds having contributions from both 1/2 and 3/2.
Fig. 9 shows that the general impression that the orbit of the structure follows a reclined
eight configuration is not absolutely true. In fact, the motion is more complicated, mainly
due to the contribution of 3/2. The reclined eight path or using a more symbolic term
the butterfly motion, is more appropriate to be used in order to describe the motion of the
structure from above, i.e. the function u=f(w). Finally, the fundamental frequency of the
response for v and u which are both in-plane components is equal to the excitation
frequency. This is shown with a more descriptive fashion in Fig. 11 where the function
v=f(u) is represented by two coinciding closed loops.
Figs. 9-11 have been plotted using the numerical predictions of two periods of the steady
state response. Another way to verify that the in-plane motions conform to the frequency of
excitation is to observe that the two loops of Fig. 11 practically coincide. However, this is not
the case when the out-of-plane motion is considered, which it is driven by a subharmonic
and a superharmonic of the excitation frequency. In this case, each of the loops in Figs. 9 and
10 (right or left) is covered during one period of the excitation. Nevertheless, the
Nonlinear Dynamics

182
fundamental frequency for the response of w, and in general for all out-of-plane
components, is the half of the excitation frequency and accordingly the steady state motion
at any point along the structure is completed after two excitation periods.
The contribution of the various harmonics, which are stimulated due to the heave excitation,
to both the in-plane and the out-of-plane dynamic components, is better shown in the 3D
spectral densities depicted in Figs. 12-17. Figs. 12-14 show in-plane components, namely the
dynamic tension T
1
(Fig. 12), the normal velocity v (Fig. 13) and the in-plane dynamic
bending moment M
b1
(Fig. 14). In the respective plots it is immediately apparent that the in-
plane components are primarily governed by the excitation frequency (=1.5rad/s in the
present case study), while it is evident that the in-plane response is affected by additional
harmonics that coincide with integer multipliers of the excitation frequency , i.e., 2, 3
etc. The 2 superharmonic is easily detectable in all three figures, whereas 3 is seen
(admittedly with relative difficulty), only in the dynamic tension spectral density (Fig. 12). It
should be stated however that it exists, together with the higher integer multipliers, in all in-
plane dynamic components.
Figs. 15-17 provide the 3D spectral densities of out-of-plane dynamic components, namely
the bi-normal velocity w (Fig. 15), the out-of-plane dynamic bending moment M
n1
(Fig. 16)
and the out-of-plane dynamic shear force S
b1
(Fig. 17). For enriching the discussion that
preceded with regard the dominant harmonics of the out-of-plane response due to the heave
excitation, it is again underlined that the motion herein is governed by frequencies that
correspond to 1/2, 3/2, 5/2 etc. The occurrence of all three of them can be detected only
in Fig. 15 (again, the latter is seen with relative difficulty), while for M
n1
and S
b1
the response
appears to be governed by 1/2. Moreover, we could positively claim that there is a slight
contribution from 3/2.
The question which easily arises is what exactly these findings mean. To provide an answer
we could generalize the visual observations on the 3D spectral densities of the out-of-plane
components and speculate that the contributing harmonics correspond to (n/2) for
n=1,2,. In addition, in order to be consistent with the above discussion we could claim that
the even terms of the sequence are negligible. As far as the in-plane response is concerned,
the logical sequence is to assume that the constituent harmonics could be approximated by
the same simple formula, but in this case, the components which could be omitted are the
odd terms of the sequence.
Correlating the above findings with the Mathieu equation, should not be considered as a
significant discovery as many authors did the same in the past. Nevertheless most of the
works in this subject discuss vertical slender structures (risers or tethers) (Gadagi &
Benaroya, 2006; Chandrasekaran et al., 2006; Kuiper et al., 2008; Park & Jung, 2002) for
marine applications where the heaving motions produce buckling and the associated
dynamic behaviour is directly connected to Mathieu equation. To extend the discussion in
the context of catenary structures, effort has been made to associate the numerical
predictions depicted graphically in 3D spectral densities to the solution(s) of Mathieu
equation. The issue for which we are mainly interested is that the global response consists of
harmonics (n/2) for n=1,2,, or equivalently n(2) for n=1,2,, provided that the
excitation frequency is 2. The Mathieu equation which is satisfied by periodic solutions is
given for reference in the following:
( ) 0 ) ( 2 cos 2
) (
2
2
= +

y q a
d
y d
(31)
The 3D Nonlinear Dynamics of Catenary Slender Structures for Marine Applications

183
where =t and q is referred as the Mathieu parameter. The solutions of Mathieu Eq. (31)
associated with the characteristic values a, are given by (Abramowitz & Stegun, 1970;
McLachlan, 1947; Meixner & Schfke, 1954)

=
=
0
2
2
2
] 2 cos[ ) ( ) ; ( ce
r
m
r
m
t r q A q t (32)
[ ]

=
+
+
+
+ =
0
1 2
1 2
1 2
) 1 2 ( cos ) ( ) ; ( ce
r
m
r
m
t r q A q t (33)
[ ]

=
+
+
+
+ =
0
1 2
1 2
1 2
) 1 2 ( sin ) ( ) ; ( se
r
m
r
m
t r q B q t (34)
[ ]

=
+
+
+
+ =
0
2 2
2 2
2 2
) 2 2 ( sin ) ( ) ; ( se
r
m
r
m
t r q B q t (35)
where ce
m
and se
m
are the even and odd periodic Mathieu functions and A and B are the
associated constants depending on the Mathieu parameter q. It is immediately apparent that
a stable solution of Mathieu Eq. (31) will include contributions originating from an infinite
number of harmonics. In any case the first harmonic will be equal to /2 provided that the
excitation frequency is . It is reminded that according to the numerical results that describe
the in-plane and the out-of-plane dynamic behaviour of the catenary structure due to heave
excitation, the response was assumed to include the same type and number of harmonics
regardless whether they are significant or not. The answer to the question why the in-plane
motions are governed by the harmonics , 2, 3,, and the out-of-plane motions by the
harmonics /2, 3/2, 5/2,is apparently a difficult task that requires deep and
comprehensive investigation and it could be the subject for a future work.
7. Conclusion
The 3D dynamic behaviour of catenary slender structures for marine applications was
considered. The investigation was based on the results obtained by solving the complete
nonlinear governing system that consists of ten partial differential equations. The solution
method employed was the finite differences box approximation. Particular attention was
given to the out-of-plane variables which are induced due to heave excitation.
The main finding in this context was the contribution of several harmonics that influence the
global response of the structure. In fact it was shown that under in-plane heave excitation at
the top terminal point the in-plane variables, motions and generalized loading components,
are governed by the harmonics , 2, 3,, whereas the out-of-plane variables by the
harmonics /2, 3/2, 5/2,
For the heave induced out-of-plane motions, the fundamental frequency is exactly the half
of the excitation frequency. This leads to cyclic motions which are completed during a time
interval that is equal to the double of the excitation period. It was shown graphically that the
Nonlinear Dynamics

184
orbit of the structure resembles a butterfly configuration. This interesting behaviour was
correlated to the even and odd periodic solutions of the canonical form of Mathieu equation.
Finally, the contribution of the nonlinearities was studied by deriving the equivalent
linearized system and it was commented that the out-of-plane motions induced due to in-
plane excitation are driven by the geometric nonlinear terms.
8. References
Ablow, C.M & Schechter, S. (1983). Numerical simulation of undersea cable dynamics. Ocean
Engineering; 10, 443-457
Abramowitz, M. & Stegun I.A. (1970). Handbook of mathematical functions, Dover Publications
Inc, New York
Aubeny, C.P., Biscotin, G. & Zhang, J (2006). Seafloor interaction with steel catenary risers, Final
Project Report, MMS Project No 510, Texas A&M University
Burgess, J.J. (1993). Bending stiffness in a simulation of undersea cable deployment.,
International Journal of Offshore and Polar Engineering, 3, 197-204
Chandrasekaran, S., Chandak, N.R. & Anupam, G. (2006). Stability analysis of TLP tethers,
Ocean Engineering, 33, 471-482.
Chatjigeorgiou, I.K. & Mavrakos, S.A. (1999). Comparison of numerical methods for predicting
the dynamic behavior of mooring lines. Proceedings of the 9th International Conference
on Offshore and Polar Engineering (ISOPE 1999), Brest, France, Vol. II, 332-339
Chatjigeorgiou, I.K. & Mavrakos, S.A. (2000). Comparative evaluation of numerical schemes
for 2D mooring dynamics, International Journal of Offshore and Polar Engineering, 10,
301-309
Chatjigeorgiou, I.K., Passano, E. & Larsen, C.M. (2007). Extreme bending moments on long
catenary risers due to heave excitation, Proceedings of the 26th International
Conference on Offshore Mechanics and Arctic Engineering (OMAE 2007), San Diego,
California, USA, Paper No 29384.
Chatjigeorgiou, I.K. (2008). A finite differences formulation for the linear and nonlinear
dynamics of 2D catenary risers, Ocean Engineering, 35, 616-636.
Clukey, E., Jacob, P. & Sharma, P. (2008). Investigation of riser seafloor interaction using
explicit finite element methods. Offshore Technology Conference, Houston, Texas,
OTC 19432
Gadagi, M.M. & Benaroya, H. (2006). Dynamic response of an axially loaded tendon of a
tension leg platform, Journal of Sound and Vibration, 293, 38-58.
Gobat, J.I. & Grosenbaugh, M.A. (2001). Application of the generalized- method to the time
integration of the cable dynamics equations, Computer Methods in Applied Mechanics
and Engineering, 190, 4817-4829.
Gobat, J.I., Grosenbaugh, M.A. & Triantafyllou, M.S. (2002). Generalized- time integration
solutions for hanging chain dynamics. Journal of Engineering Mechanics ASCE, 128,
677-687
Gobat, J.I. & Grosenbaugh, M.A. (2006). Time-domain numerical simulation of ocean cable
structures, Ocean Engineering, 33, 1373-1400.
Hoffman, J.D. (1993). Numerical methods for engineers and scientists, McGraw-Hill, New York
Howell, C.T. (1991). Numerical analysis of 2-D nonlinear cable equations with applications
to low tension problems, Proceedings of the 1st International Offshore and Polar
Engineering Conference (ISOPE 1991), Edinburgh, United Kingdom, Vol. II, 203-209.
The 3D Nonlinear Dynamics of Catenary Slender Structures for Marine Applications

185
Howell, C.T. Investigation of the dynamics of low tension cables, PhD Thesis, Massachusetts
Institute of Technology, Cambridge, Massachusetts
ISSC (2003). Report of the committee V.5: Floating Production Systems. Elsevier Science, Oxford,
Eds A.E. Mansour, R.C. Ertekin.
Jain, A.K. Review of flexible risers and articulated storage systems, Ocean Engineering, 21,
733-750.
Kuiper, G.L. & Metrikine, AV. (2005). Dynamic stability of a submerged, free-hanging riser
conveying fluid, Journal of Sound and Vibration, 280, 10511065.
Kuiper, G.L., Brugmans, J. & Metrikine, AV. (2008). Destabilization of deep-water risers by a
heaving platform, Journal of Sound and Vibration, 310, 541-557.
LeCunff, C., Biolley, F. & Damy, G. (2005). Experimental and numerical study of heave
induced lateral motion (HILM), Proceedings of the 24th International Conference on
Offshore Mechanics and Arctic Engineering (OMAE 2005), Halkidiki, Greece, Paper No
67019
Leira, B.J., Karunakaran, D., Giertsen, E., Passano, E. & Farnes, K-A. Analysis guidelines and
application of a riser-soil interaction model including trench effects, Proceedings of
the 23rd International Conference on Offshore Mechanics and Arctic Engineering (OMAE
2004), Vancouver, Canada, Paper No 51527
McLachlan N.W. (1947). Theory and applications of Mathieu functions, Dover Publications, New
York.
Meixner J. & Schfke F.W. (1954). Mathieusche funktionen und sphroidfunktionen, Springer,
Berlin
Milinazzo, F., Wilkie, M. & Latchman, S.A. (1987). An efficient algorithm for simulating the
dynamics of towed cable systems, Ocean Engineering, 14, 513-526.
Park, H-I., & Jung, D-H. (2002). A finite element method for dynamic analysis of long
slender marine structures under combined parametric and forcing excitations,
Ocean Engineering, 29, 1313-1325.
Passano, E. & Larsen, C.M. (2006). Efficient analysis of a catenary riser, Proceedings of the 25th
International Conference on Offshore Mechanics and Arctic Engineering (OMAE 2006),
Hamburg, Germany, Paper No 92308
Passano, E. & Larsen, C.M. (2007). Estimating distributions for extreme response of a
catenary riser. Proceedings of the 26th International Conference on Offshore Mechanics
and Arctic Engineering (OMAE 2007), San Diego, California, USA, Paper No 29547.
Patel, H.M. & Seyed, F.B. (1995). Review of flexible risers modelling and analysis techniques,
Engineering Structures, 17, 293-304.
Pesce, C.P., Martins, C.A. & Silveira, L.M.Y. (2006). Riser-soil interaction: Local dynamics at
TDP and a discussion on the eigenvalue and the VIV problems, Journal of Offshore
Mechanics and Arctic Engineering, 128, 39-55.
Tjavaras, A.A., Zhu, Q., Liu, Y., Triantafyllou, M.S. & Yue, D.K.P. (1998). The mechanics of
highly extensible cables, Journal of Sound and Vibration, 213, 709-737
Triantafyllou, M.S. (1994). Cable mechanics for moored floating structures. Proceedings of the
7th International Conference on the Behaviour of Offshore Structures (BOSS 1994),
Boston, Massachusetts, Vol. 2, 57-77.
Zare, K. & Datta, T.K. (1988). Vibration of Lazy-S risers due to vortex shedding under
lock-in, Proceedings of the 20th Offshore Technology Conference, OTC 5795.

Nonlinear Dynamics

186
Appendix A. Mass matrix M, stiffness matrix K and forcing vector F of Eq.
(16)

+

=
0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0
0 0 0 0 1 0 0 0 0 0
0 0 0 0 0 cos 1 0 0 0 0
0 0 0 0 0 0
1
0 0 0
0 0 0 0 sin 0 0 0
0 0 0 0 0 ) sin cos ( 0 0 0
0 0 0 0 cos 0 0 0
M
EA
T
EA
T
EA
mu mv m m
w u m m m
mw mv m
a
a

(A.1)

=
0 0 0 0 0 cos 0 0 0 0
0 0 0 0 1 0 0 0 0 0
0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 1 0 0
0 0 0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 0 0 1
0 0 0 1 0 0 0 0 0 0
0 0 1 0 0 0 0 0 0 0
0 0 0 0 0 0 1 0 0 0
K

EI
EI
(A.2)

( )
( )

+ +

+

+ +

+ + +
+
=
b
n
b b n
b
b
n b
b
b n
b
a db n b n
n
a dn b b
dt b n n b
EA
T
S EI
EA
T
S EI
u v
w u
v w
t
U
m R w T S
t
U
m R w S T
R w S S
3
3
2
0
0
0
1 tan
1 tan
tan
tan
sin sin tan
cos tan
cos sin
F

(A.3)
The 3D Nonlinear Dynamics of Catenary Slender Structures for Marine Applications

187
Appendix B. Algebraic expansions of the nonlinear system of dynamic
equilibrium Eqs. (3)-(12) using the finite differences box scheme
( )
( )
( )
( ) ( )
( ) ( )
( )
( ) 0
2
cos cos cos cos
4
1
2 4
1
2
1 1
1 1
4
1
2
1
cos sin cos sin cos sin cos sin
4
4
1
4
1
2
1
1
1
1
1 1
1
1
1
1
1 1
1
1
1
1
1
1
1
1 1
1
1
1
2 / 1
1 1 1
2 / 1
1
1
1
1
1
1
2 / 1 2 / 1
1 1 1
1 1
1
1
1
1
1 1 0
1 1
1
1
1
1
1 1
1 1
1
1
1
1
1 1 1
1
1
1
1
=

+
+ + +

+ + + +

+ + +

+ + +
+ + +
+ + +
+ + + +

+
=

+

+

+ +

+
+

+

+

+ + +

+

+ +

+

+ +

+

+ +
+

+
t
v v v v
t
w w w w
t
u u u u
m
e v v e v v
e v v e v v C d
w
S S S S
S S S S
s
T T T T
E
i
k
i
k
i
k
i
k i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
tr
i
k
tr
i
k
i
k
tr
i
k
tr
i
k
i
k
tr
i
k
tr
i
k
i
k
tr
i
k
tr dt o
i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
b
i
k
n
i
k
b
i
k
n
i
k
b
i
k
n
i
k
b
i
k
n
i
k
n
i
k
b
i
k
n
i
k
b
i
k
n
i
k
b
i
k
n
i
k
b
i
k
i
k
i
k
i
k

(B.1)
( )
( )
( )
( ) ( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( ) ( )
( )
( ) 0
2
sin sin sin sin
4
2
cos cos cos cos
4
2 2
1 1
1 1
4
1
2
1
cos cos cos cos
4
tan tan tan tan
4
1
4
1
2
1
1
1
1
1 1
1
1
1
1
1 1
1
1
1
1
1 1
1
1
1
1
1 1
1
1
1
1
1
1
1
1
2 / 1
1
2 / 1
2
1
2
1 1
2 / 1
1
1
2 / 1
2
1
1
2
1
1
1
1
2 / 1
2 / 1
2 2 2 / 1
1
2 / 1
2
1
2
1 1
1
1
1
1 0
1 1 1
1
1
1
1
1
1
1 1 1
1 1
1
1
1
1
1 1 1
1
1
1
2
=

+
+ + +

+
+ + +

+ + + + +

+ + + + +
+ + +
+ + + +
+ + + +

+
=

+

+

+ +

+

+

+ +

+

+

+ + + +

+

+

+ + +

+

+ +
+

+
t
w w w w
m
t
u u u u
m
t
v v v v
m
t
v v v v
m
e v v v e v v v
e v v v e v v v C d
w
S S S S
T T T T
s
S S S S
E
i
k
i
k
i
k
i
k i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
nr
i
k
nr
i
k
nr
i
k
nr
a
i
k
i
k
i
k
i
k
i
k
i
k
br
i
k
nr
i
k
nr
i
k
i
k
br
i
k
nr
i
k
nr
i
k
i
k
br
i
k
nr
i
k
nr
i
k
i
k
br
i
k
nr
i
k
nr dn o
i
k
i
k
i
k
i
k
i
k
i
k
b
i
k
b
i
k
i
k
b
i
k
b
i
k
i
k
b
i
k
b
i
k
i
k
b
i
k
b
i
k
b
i
k
i
k
b
i
k
i
k
b
i
k
i
k
b
i
k
i
k
n
i
k
n
i
k
n
i
k
n

(B.2)
Nonlinear Dynamics

188

( )
( )
( )
( ) ( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( ) ( )
( )
( ) 0
2 4
2
sin sin sin sin
4
2 2
1 1
1 1
4
1
2
1
sin sin sin sin sin sin sin sin
4
tan tan tan tan
4
1
4
1
2
1
1
1
1
1
1
1
1
1
1
1
1
1 1
1
1
1
1
1 1
1
1
1
1
1
1
1
1
2 / 1
1
2 / 1
2
1
2
1 1
2 / 1
1
1
2 / 1
2
1
1
2
1
1
1
1
2 / 1
2 / 1
2 2 2 / 1
1
2 / 1
2
1
2
1 1
1 1
1
1
1
1
1 1 0
1 1 1
1
1
1
1
1
1
1 1 1
1 1
1
1
1
1
1 1 1
1
1
1
3
=

+
+ + + +

+
+ + + +

+ + + + +

+ + + + +
+ + +
+ + +
+ + +

+
=

+

+

+ +

+

+

+ + + +

+

+ +

+

+ + +

+

+ +
+

+
t
u u u u
m
t
v v v v
m
t
v v v v
m
t
w w w w
m
e v v v e v v v
e v v v e v v v C d
w
S S S S
T T T T
s
S S S S
E
i
k
i
k
i
k
i
k i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
br
i
k
br
i
k
br
i
k
br
a
i
k
i
k
i
k
i
k
i
k
i
k
br
i
k
nr
i
k
br
i
k
i
k
br
i
k
nr
i
k
br
i
k
i
k
br
i
k
nr
i
k
br
i
k
i
k
br
i
k
nr
i
k
br db o
i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
n
i
k
b
i
k
i
k
n
i
k
b
i
k
i
k
n
i
k
b
i
k
i
k
n
i
k
b
i
k
n
i
k
i
k
n
i
k
i
k
n
i
k
i
k
n
i
k
i
k
b
i
k
b
i
k
b
i
k
b

(B.3)

( )
( ) 0
2
1
4
1
4
1
2
1
1
1
1
1 1
1
1
1
1
1 1
1 1
1
1
1
1
1 1 1
1
1
1
4
=

+
+ + +
+ + + +

+
=

+

+

+ +

+

+ +
+

+
t
T T T T
EA
v v v v
w w w w
s
u u u u
E
i
k
i
k
i
k
i
k i
k
b
i
k
i
k
b
i
k
i
k
b
i
k
i
k
b
i
k
i
k
n
i
k
i
k
n
i
k
i
k
n
i
k
i
k
n
i
k
i
k
i
k
i
k
i
k
(B.4)

( )
( )
( ) ( ) ( ) [
( ) ] 0
2
cos 1
cos 1 cos 1 cos 1
4
1
tan tan tan tan
4
1
4
1
2
1
1
1
1
1 1
1
1
1
1
1 1
1 1 1
1
1
1
1
1
1
1 1 1
1 1
1
1
1
1
1 1 1
1
1
1
5
=

+
+ +
+ + + + +
+ + + +
+ + + +

+
=

+

+

+ +

+

+ + +

+

+ +
+

+
t
e
e e e
w w w w
u u u u
s
v v v v
E
i
k
i
k
i
k
i
k i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
i
k
b
i
k
i
k
i
k
b
i
k
i
k
i
k
b
i
k
i
k
i
k
b
i
k
i
k
b
i
k
i
k
b
i
k
i
k
b
i
k
i
k
b
i
k
i
k
i
k
i
k
i
k

(B.5)

( )
( )
( ) ( ) ( ) [ ( )] 0
2
1 1 1 1
4
1
tan tan tan tan
4
1
4
1
2
1
1
1
1
1
1
1
1
1 1 1
1
1
1
1
1
1
1 1 1
1 1
1
1
1
1
1 1 1
1
1
1
6
=

+
+ + + + + + + +
+ + +
+ + +

+
=

+

+

+ + +

+

+ +
+

+
t
e e e e
v v v v
u u u u
s
w w w w
E
i
k
i
k
i
k
i
k i
k
i
k
i
k
i
k
i
k
i
k
b
i
k
i
k
i
k
b
i
k
i
k
i
k
b
i
k
i
k
i
k
b
i
k
i
k
n
i
k
i
k
n
i
k
i
k
n
i
k
i
k
n
i
k
i
k
i
k
i
k
i
k

(B.6)
The 3D Nonlinear Dynamics of Catenary Slender Structures for Marine Applications

189

(
)
( ) ( ) ( ) ( ) 0 1 1 1 1
4
1
tan tan
tan tan
4 2
3
1 1
3
1
1
1
1
3 3
1 1
1 1 1
1
1
1
1
1
1
1 1 1 1
1
1
1
7
=

+ + + + + + +
+ +
+

+
=

+

+ +

+

+ + +
+

+
i
k
i
k
b
i
k
i
k
b
i
k
i
k
b
i
k
i
k
b
i
k
i
k
b
i
k
b
i
k
i
k
b
i
k
b
i
k
i
k
b
i
k
b
i
k
i
k
b
i
k
b
i
k
n
i
k
n
i
k
n
i
k
n
e S e S e S e S
EI
s
EI E


(B.7)

(
)
( ) ( ) ( ) ( ) 0 1 1 1 1
4
1
tan tan
tan tan
4 2
3
1 1
3
1
1
1
1
3 3
1 1
1 1 1
1
1
1
1
1
1
1 1 1 1
1
1
1
8
=

+ + + + + + + +
+ +
+

+
=

+

+ +

+

+ + +
+

+
i
k
i
k
n
i
k
i
k
n
i
k
i
k
n
i
k
i
k
n
i
k
i
k
b
i
k
n
i
k
i
k
b
i
k
n
i
k
i
k
b
i
k
n
i
k
i
k
b
i
k
n
i
k
b
i
k
b
i
k
b
i
k
b
e S e S e S e S
EI
s
EI E


(B.8)
( ) 0
4
1
2
1
1
1
1 1
1
1
1
9
= + + +

+
=

+

+
+

+
i
k
n
i
k
n
i
k
n
i
k
n
i
k
i
k
i
k
i
k
s
E

(B.9)

( )
( ) 0
4
1
2
cos cos cos cos
4
1
1
1
1
1
1
1
1
1
1
1
1
1
10
= + + +

+
+ + + =

+
i
k
b
i
k
b
i
k
b
i
k
b
i
k
i
k
i
k
i
k i
k
i
k
i
k
i
k
s
E


(B.10)


Fig. 1. Stretched catenary segment and balance of internal loading.
Nonlinear Dynamics

190

Fig. 2. Spectral densities of the dynamic tension T
1
along the catenary under sway excitation
at the top, with amplitude y
a
=1.0m and circular frequency =2.0rad/s.



Fig. 3. Spectral densities of the normal velocity v along the catenary under sway excitation at
the top, with amplitude y
a
=1.0m and circular frequency =2.0rad/s.
The 3D Nonlinear Dynamics of Catenary Slender Structures for Marine Applications

191

Fig. 4. Spectral densities of the in-plane dynamic bending moment M
b1
along the catenary
under sway excitation at the top, with amplitude y
a
=1.0m and circular frequency
=2.0rad/s.

Fig. 5. Spectral densities of the out-of-plane dynamic bending moment M
n1
along the
catenary under sway excitation at the top, with amplitude y
a
=1.0m and circular frequency
=2.0rad/s.
Nonlinear Dynamics

192
100 102 104 106 108 110 112 114 116 118 120
2
1.5
1
0.5
0
0.5
1
1.5
2
time (s)
w

(
m
/
s
)


ya=2m
ya=3m
ya=4m
ya=5m

Fig. 6. Effect of the initial, short-time, sway displacement on the out-of-plane velocity w due
to heave excitation with amplitude z
a
=1.0m and circular frequency =1.5rad/s. The time
history depicts the variation of w at the location of the max static in-plane bending moment
M
b0
, namely at s41m from touch down (at node k=3 in a discretization grid of 100 nodes)

100 102 104 106 108 110 112 114 116 118 120
2.5
2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
x 10
5
time (s)
M
n

(
N
m
)


ya=2m
ya=3m
ya=4m
ya=5m

Fig. 7. Effect of the initial, short-time, sway displacement on the out-of-plane dynamic
bending moment M
n1
due to heave excitation with amplitude z
a
=1.0m and circular
frequency =1.5rad/s. The time history depicts the variation of M
n1
at the location of the
max static in-plane bending moment M
b0
, namely at s41m from touch down (at node k=3 in
a discretization grid of 100 nodes)
The 3D Nonlinear Dynamics of Catenary Slender Structures for Marine Applications

193
100 102 104 106 108 110 112 114 116 118 120
3000
2000
1000
0
1000
2000
3000
4000
time (s)
S
b

(
N
)


ya=2m
ya=3m
ya=4m
ya=5m

Fig. 8. Effect of the initial, short-time, sway displacement on the out-of-plane dynamic shear
force S
b1
due to heave excitation with amplitude z
a
=1.0m and circular frequency =1.5rad/s.
The time history depicts the variation of S
b1
at the location of the max static in-plane bending
moment M
b0
, namely at s41m from touch down (at node k=3 in a discretization grid of 100
nodes)

Fig. 9. Orbit of node no 3 (in a discretization grid of 100 nodes at s=41m from touch down)
as seen from behind (v=f(w)), under heave excitation at the top with amplitude z
a
=1.0m and
circular frequency =1.5rad/s.
Nonlinear Dynamics

194

Fig. 10. Orbit of node no 3 (in a discretization grid of 100 nodes at s=41m from touch down)
as seen from above (u=f(w)), under heave excitation at the top with amplitude z
a
=1.0m and
circular frequency =1.5rad/s.

Fig. 11. Orbit of node no 3 (in a discretization grid of 100 nodes at s=41m from touch down)
as seen from the side (v=f(u)), under heave excitation at the top with amplitude z
a
=1.0m and
circular frequency =1.5rad/s.
The 3D Nonlinear Dynamics of Catenary Slender Structures for Marine Applications

195

Fig. 12. Spectral densities of the dynamic tension T
1
along the catenary under heave
excitation at the top, with amplitude z
a
=1.0m and circular frequency =1.5rad/s.


Fig. 13. Spectral densities of the normal velocity v along the catenary under heave excitation
at the top, with amplitude z
a
=1.0m and circular frequency =1.5rad/s.
Nonlinear Dynamics

196

Fig. 14. Spectral densities of the in-plane dynamic bending moment M
b1
along the catenary
under heave excitation at the top, with amplitude z
a
=1.0m and circular frequency
=1.5rad/s.


Fig. 15. Spectral densities of the bi-normal velocity w along the catenary under heave
excitation at the top, with amplitude z
a
=1.0m and circular frequency =1.5rad/s.
The 3D Nonlinear Dynamics of Catenary Slender Structures for Marine Applications

197














Fig. 16. Spectral densities of the out-of-plane dynamic bending moment M
n1
along the
catenary under heave excitation at the top, with amplitude z
a
=1.0m and circular frequency
=1.5rad/s.
Nonlinear Dynamics

198















Fig. 17. Spectral densities of the out-of-plane dynamic shear force S
b1
along the catenary
under heave excitation at the top, with amplitude z
a
=1.0m and circular frequency
=1.5rad/s.
9
Nonlinear Dynamics Traction Battery Modeling
Antoni Szumanowski
Warsaw University of Technology,
Poland
1. Introduction
This chapter presents a method of determining electromotive force (EMF) and battery
internal resistance as time functions, which are depicted as functions of state of charge
(SOC). The model is based on battery discharge and charge characteristics under different
constant currents that are tested by a laboratory experiment.
Further the method of determining the battery SOC according to the battery modeling result
is considered. The influence of temperature on battery performance is analyzed according to
laboratory-tested data and the theoretical background for calculating the SOC is obtained.
The algorithm of battery SOC indication is depicted in detail. The algorithm of battery SOC
online indication considering the influence of temperature can be easily used in practice
by microprocessor. NiMH and Li-ion battery are taken under analyze. In fact, the method
also can be used for different types of contemporary batteries, if the required test data are
available.
Hybrid electric (HEVs) and electric (EVs) vehicles are remarkable solutions for the world
wide environmental and energy problem caused by automobiles. The research and
development of various technologies in HEVs is being actively conducted [1]-[8]. The role of
battery as power source in HEVs is significant. Dynamic nonlinear modeling and
simulations are the only tools for the optimal adjustment of battery parameters according to
analyzed driving cycles. The batterys capacity, voltage and mass should be minimized,
considering its over-load currents. This is the way to obtain the minimum cost of battery
according to the demands of its performance, robustness, and operating time.
The process of battery adjustment and its management is crucial during hybrid and electric
drives design. The generic model of electrochemical accumulator, which can be used in
every type of battery, is carried out. This model is based on physical and mathematical
modeling of the fundamental electrical impacts during energy conservation by a battery.
The model is oriented to the calculation of the parameters EMF and internal resistance. It is
easy to find direct relations between SOC and these two parameters. If the EMF is defined
and the function versus the SOC ( 0, 1 k < > ) is known, it is simple to depict the
discharge/charge state of a battery.
The model is really nonlinear because the correlative parameters of equations are functions
of time [or functions of SOC because ( ) SOC f t = ] during battery operation. The modeling
method presented in this chapter must use the laboratory data (for instance voltage for
different constant currents or internal resistance versus the battery SOC) that are expressed
Nonlinear Dynamics

200
in a static form. These data have to be obtained discharging and charging tests. The
considered generic model is easily adapted to different types of battery data and is
expressed in a dynamic way using approximation and iteration methods.
An HEV operation puts unique demands on battery when it operates as the auxiliary power
source. To optimize its operating life, the battery must spend minimal time in overcharge
and or overdischarge. The battery must be capable of furnishing or absorbing large currents
almost instantaneously while operating from a partial-state-of-charge baseline of roughly
50% [9]. For this reason, knowledge about battery internal loss (efficiency) is significant,
which influences the battery SOC.
There are many studies dedicated to determine the battery SOC [10]-[22]; however, these
solutions have some limitations for practical application [23]. Some solutions for practical
application are based on a loaded terminal voltage [17]-[20] or a simple calculation the flow
of charge to/from a battery [21]-[22], which is the integral that is based on current and time.
Both solutions are not considered the strong nonlinear behavior of a battery. It is possible to
determine transitory value of the SOC online in real drive conditions with proper
accuracy, considering the nonlinear characteristic of a battery by resolving the mathematical
model that is presented in this paper.
This is the background for optimal battery parameters as well as the proper battery
management system (BMS) design - particularly in the case of SOC indication [25]. The high
power (HP) NiMH and LiIon batteries so common used in HEV were considered.
Finally, for instance, the plots of battery voltage, current and SOC as alterations in time for
real experimental hybrid drive equipped with BMS especially design according to presented
original battery modeling method, are attached.
2. Battery dynamic modeling
2.1 Battery physical model
The basis enabling the formulation of the energy model of an electrochemical battery is
battery physical model shown in Fig.1.

i
a
R
el
R
e
R
p
E
U
a

Fig. 1. Substitute circuit for nonlinear battery modeling
2.2 Mathematical modeling
The internal resistance can be expressed in an analytical way [7], where:
( ) ( ) ( ) ( )
1
, , , , ,
w a el e a a
R i Q R Q R Q bE i Q I

= + + (1)
1
( , , )
a a
bE i Q I

is the resistance of polarization.
b is the coefficient that expresses the relative change of the polarizations EMF on the cells
terminals during the flow of the
a
I current in relation to the EMF E for nominal capacity.
Electrolyte resistance
el
R and electrode resistance
e
R are inversely proportional to
Nonlinear Dynamics Traction Battery Modeling

201
temporary capacity of the battery. During real operation, the capacity of the battery is
changeable with respect to current and temperature [7], i.e.,
( ) ( ) ( ) ( )
, , ,
u a w a
Q i t Q K i t t

= (2)
or
( ) ( ) ( )
0
, , , d
t
u a a a
Q i t Q i i t t

(3)
Where:
( ) ( )
,
w a
K i t t is the nonlinear function that is used to calculate the battery discharged capacity
( )
0
d
t
a
i t t

is the function that is used to calculate the used charge, which has been drawn from
the battery since the instant time t=0 till the time t
( ) ,
a
Q i

is the battery capacity as a function of temperature and load current, and

( ) n
w a
K i t

= (4)
where
w
K is the discharge capacity of the battery, n is the Peukerts constant, which varies
for different types of batteries.
Assuming temperature influence:
( ) ( )
( )
( )
0
, , d
t
a
u n a
n
i t
Q i t c Q i t t
I

=



(5)
where the ( ) c

coefficient can be defined as the temperature index of nominal capacity [7], i.e.,
( )
( )
1
1
n n
Q
c
Q

= =
+
(6)
where is the temperature capacity index (we can assume 0.01 deg
-1
).
According to the Peukert equation, we can get the following:

( ) ( )
( )
a a
n n
Q i U i t
I Q U

=


(7)
The left hand side of the equation (7) is the quotient of the electric power that is drawn
from the battery during the flow of
a n
i I current and the electric power that is drawn from
the battery during loading with the rated current. The quotient mentioned above defines the
usability index of the accumulated power, i.e.,
( )
( )
( )
,
a
A a
n
i t
i
I




=


(8)
Nonlinear Dynamics

202
When
a n
i I < , the value of the index can exceed 1.
During further solution of (5), it can be transformed by means of (8), i.e.,
( ) ( ) ( ) ( )
0
, , , d
t
u A a n a
Q i t c i Q i t t

=

(9)
Therefore, the real battery SOC can be expressed in the following way [7]:

( ) ( ) ( )
0
, d
t
A a n a
u
n n
c i Q i t t
Q
k
Q Q



= =

(10)
where 1 k = for a nominally charged battery, 0 1 k , and thus
( ) ( ) ( )
0
1
, d
t
A a a
n
k c i i t t
Q

=

(11)
For practical application, its necessary to transform aforementioned equations for
determining the internal resistance
w
R and EMF as functions of k (SOC) [7], i.e.,
( )
( ) ( )
( )
( )
1 2
, ,
, ,
, , , ,
a
w a
u a u a a
bE i Q l l
R i Q
Q i t Q i t i t

= + + (12)
( )
( )
( )
1
, ,
w a
a
E k
R i t lk b
i t

= + (13)
where
1
1 2
( )
n
l l l Q

= + , l const is a piecewise constant, assuming that the temporary change
of the battery capacity is significantly smaller than its nominal capacity; the coefficient l is
experimentally determined under static conditions. ( ) E k is the temporary value of
polarizations EMF, which is dependent on the SOC.
The EMF as a function of k is deduced from the well-know battery voltage equation,
including the momentary value of voltage and internal resistance, because the values
w
R and EMF are unknown. The solution can be obtained by a linearization and iterative
method, which is explained by following Fig.2 and following:

*
min
*
max
( )
( )
E k E
b k
E

= (14)
Take under consideration (12)-(14), its then possible to obtain the following:

*
min
*
max
*
1 min 1 1
1 *
max 1
( ) ( ) ( )
( )
( ) ( ) ( )
( )
n n n
w n
n n
n n n
w n
n n
E k E E k l k
R k
E I k
E k E E k l k
R k
E I k

= +

= +

(15)
Obviously, ( ) E k is the function that we need. To obtain it, its necessary to use the known
functions ( )
a
u k , which are obtained by laboratory tests.
Nonlinear Dynamics Traction Battery Modeling

203
0
SOC ( k )
EMF
E(k)
b(k
n
,k
n-1
)=
E(k) E*
min
E*
max
E
max

E
min

E***
max

E***
min

E*
max

E*
max
=E
max

E**
min

E*
min

k
n+1

k
n k
m
k
m+1 k
1=1
k
2


Fig. 2. Linearization method of EMF versus SOC (k)

SOC ( k )
u
u
a
(k)for I
a
=const.
k
m
k
m+1

k
1=
k
2

u
1

k
n-1
k
n

0.
I
a(n)
=const.
I
a(n-1)
=const.
u
2

u
m

u
m+1
u
n-1
u
n


Fig. 3. Linearization method of voltage versus SOC (k)
Similarly as in the case of Fig.3, the following equations are generated:

1 1 1
( ) ( ) ( )
( ) ( ) ( )
n n a w n
n n a w n
u k E k I R k
u k E k I R k

=

(16)
( )
n
u k and
1
( )
n
u k

are known from the family of voltage characteristics that are obtained by
laboratory tests.
( ) a n
I is also known because ( )
n
u k is determined for
( )
.
a n
I const =
Nonlinear Dynamics

204
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.8
0.9
1
1.1
1.2
1.3
1.4
1.5
State of charge
B
a
t
t
e
r
y

v
o
l
t
a
g
e

[
V
]
0.5C
1C
2C
3C
4C
5C

Fig. 4. Discharging data of a 14-Ah NiMH battery

Fig. 5. Charging data of a 14-Ah NiMH battery
Nonlinear Dynamics Traction Battery Modeling

205
+ is for discharge
- is for charge
0, 1 k < >
Using the above-presented approach, based on experimental data (shown in Figs.4 and 5),
its possible to construct a proper equation set as in the shape of (15) and (16) and resolve it
in an iterative way.
Last, the equations of
w
R and EMF take the shape of the following polynomial:

6 5 4 3 2
6 5 4 3 2
6 5 4 3 2
7 6 5 4 3 2
( )
( )
( )
( )
w r r r r r r r
e e e e e e e
b b b b b b b
l l l l l l l l
R k A k B k C k D k E k F k G
E k A k B k C k D k E k F k G
b k A k B k C k D k E k F k G
l k Ak Bk C k Dk E k Fk G k H
= + + + + + +
= + + + + + +
= + + + + + +
= + + + + + + +
(17)
3. Battery modeling results
The basic elements that are used to formulate the mathematical model of a NiMH battery
are the described iteration-approximation method and the approximations based on the
battery discharging and charging characteristics that are obtained by an experiment.
Experimental data are approximated to enable determination of the internal resistance in a
small-enough range k=0.001. The modeling results (Figs. 6-8) in the battery SOC operating
range of 0.1-0.95 show a small deviation (less than 1%) from the experimental data (Figs.9
and 10). The NiMH battery that is used in the experiment and the modeling is an HP battery
for HEV application. The nominal voltage of the battery is 1.2V, and the rated capacity
14Ah.
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
3
3.5
4
x 10
-3
State of charge
I
n
t
e
r
n
a
l

r
e
s
i
s
t
a
n
c
e

R
d
i
s
(
w
)

[
o
h
m
]

Fig. 6. Computed internal resistance characteristics of a 14-Ah NiMH battery for discharging
Nonlinear Dynamics

206
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
3.4
3.5
3.6
3.7
3.8
3.9
4
x 10
-3
State of Charge
I
n
t
e
r
n
a
l

R
e
s
i
s
t
a
n
c
e

R
c
(
w
)

[
o
h
m
]

Fig. 7. Computed internal resistance characteristics of a 14-Ah NiMH battery for charging
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
1.28
1.3
1.32
1.34
1.36
1.38
1.4
1.42
1.44
1.46
Electromotive Force of charging
State of Charge (k)
E
M
F

[
V
]

Fig. 8. Computed EMF of a 14-Ah NiMH battery
After approximation according to the computed results, approximated equations of (17) for 14-
Ah NiMH battery can be obtained. These factors of equations (17) are available in Table 1.
Nonlinear Dynamics Traction Battery Modeling

207
Coefficient b Coefficient l
Discharging Discharging
Factors of
Equation
(17)
Internal
resistance R(w)
during
discharging
Internal
resistance
Rd(w) during
charging
Electromotive
Force
Charging Charging
-0.015363 0.65917
A 0.65917 0.42073 13.504
0.015341 0.42073
0.10447 -2.0528
B -2.0397 -1.4434 -36.406
-0.10661 -1.4376
-0.18433 2.4978
C 2.4684 1.9362 36.881
0.22702 1.9195
0.13578 -1.495
D -1.4711 -1.2841 -17.198
-0.21788 -1.2661
-0.045129 0.45416
E 0.44578 0.43809 3.5264
0.10346 0.42896
0.0059814 -0.066422
F -0.065274 -0.071757 -0.10793
-0.023367 -0.06961
-9.416e-005 0.0099289
G 0.0099109 0.0078518 1.234
0.0020389 0.0076585
-1.2154e-015
H
1.9984e-008
Table 1. Factors of Eq. (17) for 14-Ah NiMH battery

Fig. 9. Error of experiment data and the computed voltage at different discharge currents
The basic element used to formulate the mathematical model of Li-ion battery module from
SAFT Company is the earlier described iteration-approximation method and the
approximated based on the battery discharging characteristics obtained by experiment. The
experimental data is approximated to enable determining the internal resistance in an
Nonlinear Dynamics

208
enough small range k = 0.001. The analyses, in the operating range SOC between 0.01~0.95,
gives us a small deviation (less than 2%) by using the iteration-approximation method from
the experimental data. The VL30P-12S module has 30Ah rated capacity and its special
designed for HEV application.

Fig. 10. Error of experiment data and computed voltage at different charge currents

Fig. 11. The discharging voltage characteristics of SAFT 30Ah Li-ion module
Nonlinear Dynamics Traction Battery Modeling

209
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.014
0.015
0.016
0.017
0.018
0.019
0.02
0.021
0.022
0.023
0.024
Internal Resistance R(w)
k
R
w

(
o
h
m
)

Fig. 12. The computed internal resistance of SAFT 30Ah Li-ion module


Fig. 13. The computed EMF of SAFT 30Ah Li-ion module
Nonlinear Dynamics

210
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
1
2
3
4
5
6
7
8
9
x 10
-4
Coefficient b
k
b

Fig. 14. The computed coefficient b of SAFT 30Ah Li-ion module
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.002
0.004
0.006
0.008
0.01
0.012
0.014
0.016
0.018
Coefficient l
k
l

(
o
h
m
)

Fig. 15. The computed coefficient l of SAFT 30Ah Li-ion module
After approximation according to the computed results, approximated equations of (17) for 30-
Ah Li-ion module can be obtained. These factors of equations (17) are available in Table 1.
Nonlinear Dynamics Traction Battery Modeling

211
Factors of
equations (6.56)
Internal resistance
R
w
Electromotive force
E
Coefficient b Coefficient l
A 0.71806 -28.091 0.0032193 0.71806
B -2.6569 157.05 -0.016116 -2.6545
C 3.7472 -296.92 0.036184 3.736
D -2.5575 265.34 -0.040738 -2.5406
E 0.8889 -119.29 0.023539 0.87755
F -0.14693 30.476 -0.0065159 -0.14352
G 0.023413 38.757 0.00078501 0.022978
H -1.7916e-015
Table 2. Factors of Eq. (17) for 30-Ah Li-ion module


Fig. 16. Errors between testing data and computed result of SAFT 30Ah Li-ion module
4. Temperature influence analysis on battery performance
The determination of the battery EMF and internal resistance gives unlimited possibilities of
calculating the batterys voltage versus SOC (k) relation for a different value of discharge-
charge current. For a real driving condition, the battery discharge or charge depends on the
drive architecture influencing the respective power distribution. In majority, battery
charging takes place during vehicle regenerative braking, which means that this situation
lasts for a relatively short time with a significant peak-current value. A discharging current
that is too high results in a rapid increase in the battery temperature.
Nonlinear Dynamics

212
The main role of this study is to find a theoretical background for calculating the
temperature influence on the battery SOC. The presented method is more accurate and
complicated compared with other methods, which doesnt mean that it is more difficult to
apply. First of all, it is necessary to make the following assumptions:
The considered battery is fully charged in nominal conditions: nominal current, nominal
temperature and nominal capacity (
b
i =1C,
b
=20C, the capacity is designed for nominal
parameters, respectively).
The EMF for the considered battery is defined as its nominal condition in the nominal SOC
alteration range 1, 0 k < > . The assumption is taken that the EMF value of k=0.15 is the
minimum EMF. For k=0, the EMF is defined as the minimum-minimorum, in practice
which should not be obtained. The same assumption is recommended for a value that is
different from the nominal temperature for the k

(SOC) definition. As shown in Fig.19, the

starting point value of the EMF for a different value from the nominal temperature can be
higher or lower, which means that the extension alteration of the SOC could be longer or
shorter. For instance (see Fig.17), in the case of the NiMH battery for a value that is higher
than the nominal temperature, the discharge capacity is smaller than the nominal, which
means that for a certain temperature, the battery capacity corresponding to this temperature
is also changed in file 1, 0 k

< > . However, the full k

doesnt mean the same discharge
capacity as in the case of nominal temperature but does mean the maximum discharge
capacity at this temperature. For this reason, in fact, k

for this temperature is

only ( ) ( ) k t k t

> , [in some case, ( ) ( ) k t k t

< , where ( ) k t is connected only with nominal

conditions].
-20 -10 0 10 20 30 40 50 60
12
12.5
13
13.5
14
14.5
Temperature dependence of discharge capacity
Temperature [C]
D
i
s
c
h
a
r
g
e

C
a
p
a
c
i
t
y

[
A
h
]
0.5C testing data
0.5C approximate curve
1C testing data
1C approximate curve

Fig. 17. Temperature dependence of the discharge capacity of the NiMH battery
Nonlinear Dynamics Traction Battery Modeling

213
-30 -20 -10 0 10 20 30
Temperature [C]
Q [Ah]
EMFnom
EMFmax
EMFmin
Qmin
Qmax
Qnom

Fig. 18. Temperature dependence of the battery usable discharge capacity and the EMF
starting point
From Fig.18, it is easy to note that the EMF (in the case of this battery type) value in the
nominal conditions is smaller than the EMF value for a temperature that is lower than 20C
(the nominal temperature), which means that for a maximum EMF value, the available
battery capacity is higher than in the case of the nominal terms. For nominal conditions, the
SOC can be defined by a k factor ( 1, 0 k < > ). If the EMF for the non-nominal conditions
reaches its highest value, the available charge (in ampere-hours) will be also greater. It is
easy to note the relation
max
Q Q

= and
nom
Q is defined as follows:
max
1
nom
Q Q
Q

=
> [If
1,
nom
nom
Q
Q Q
Q

< <
correspondingly,
max
1
nom
EMF
EMF EMF
EMF

< <
]
This corresponds to:
max
1
nom
EMF EMF
EMF

=
>
. On the other hand, for
nom
Q , 1, 0 k < > , but relating it to
nom
Q Q

> in
condition, the file 1, 0 < > means file
max
0,Q < > . Transforming k in nominal terms to k

is
necessary to use the general relation
nom
Q
Q

. Theoretically, the product
nom
nom
Q
k
Q

transfers the
SOC factor into other than nominal temperature conditions. The same transformation can be
obtained for
nom
nom
EMF
k
EMF

, where 1, 0
nom
k < > .
Nonlinear Dynamics

214

Fig. 19. Relation of
nom
EMF
EMF
and temperature
Using the transformation factor
nom
nom
EMF
k
EMF

or k s

(
,
nom
nom
EMF
k k s
EMF

= =
),it is possible to
relate the SOC of the battery that is determined for the nominal temperature to other
different temperatures.
5. Algorithm of battery SOC indication
The algorithm is given as follows.
1. By simulation, the family of ( )
b
u k for different constant currents 0.5 , 6
b
i C C < > and
nominal temperature (e.g. 20C) can be obtained according to battery modeling results
(EMF and internal resistance as functions of SOC).
2. From Fig.19,
nom
EMF
s
EMF

=
is defined for <-30C, +35C>
3. From Fig.20, for k=0.9, 0.2, the following lookup table can be obtained
11 11 1 81 81 8
12 12 1 82 82 8
1 1 1 8 8 8
, , , ,
, , , ,
0.9 0.2
, , , ,
n n n n
u i E u i E
u i E u i E
k k
u i E u i E



= =




""
" "

Because of the practical limitation of the SOC alteration of the battery that is applied in
hybrid drives, the range of k changes can be expressed as 0.9, 0.2 < > for the nominal
temperature.
Nonlinear Dynamics Traction Battery Modeling

215

Fig. 20. EMF and calculated discharging voltage characteristics at different discharging
current and nominal temperature
4. Considering the real temperatures, the SOC of the battery in relation to the nominal
temperature can be defined as s k k

= For instance, at a temperature of +5C,
1.06
nom
EMF
EMF

=
; hence, k
+5C
=1.06 k ,which means that at this moment and this
temperature, the available capacity is 1.06 times that of the nominal temperature. At a
temperature +30C,
0.89
nom
EMF
EMF

=
; hence, k
+30C
=0.89 k , which means that at this
moment and this temperature, the available capacity is 0.89 times that of the nominal
temperature.
A similar method and process can be used for the battery charging process (see Fig.21)
The above-depicted method can be used in design of battery management system ( BMS )
for the SOC determination and indication, especially in hybrid ( HEV ) and electric ( EV )
vehicle drives. Based on the aforementioned steps 1) - 4), the SOC indication algorithm can
be depicted as is shown in Fig.22.
In HEV the battery SOC changes faster ( because HP high power battery is used ) but not so
deep as in pure electric vehicles, equipped with high energy ( HE ) battery. It means that
the SOC indication - display process may not be realized as frequently. Its not necessary to
display the SOC of the battery every second. Certainly, the previous value of the SOC has to
be remembered by a microprocessor.
High accuracy of determination of battery SOC is at first of all necessary for entire drive
system control. In opposed to indication display, the feedback SOC signals from battery
must be available online.
Nonlinear Dynamics

216

Fig. 21. EMF and calculated charging voltage characteristics at different charging current
and nominal temperature

Battery

8 8 8
8 82 82
8 81 81
1 1 1
1 12 12
1 11 11
, ,
, ,
, ,
, ,
, ,
, ,
E i u
E i u
E i u
E i u
E i u
E i u
n n n n
"
"
"
al no min
=

k
b
i
b
u
Battery

8 8 8
8 82 82
8 81 81
1 1 1
1 12 12
1 11 11
, ,
, ,
, ,
, ,
, ,
, ,
E i u
E i u
E i u
E i u
E i u
E i u
n n n n
"
"
"
al no min
=

k
b
i
b
u

a) Discharging b) Charging
Fig. 22. SOC indication algorithm
Nonlinear Dynamics Traction Battery Modeling

217
The presented original method of EMF ( as function of k - SOC ) calculation is the
background for constructing BMS. This procedure is easily adopted for control application
in HEV and EV. Its high accuracy is very important for control drive systems ( master
controller ) based on feedback signals from BMS.
The following equation is the background to determine accurate value of SOC ( k ) for
dynamic conditions.

( ) ( ) ( ) ( )
w
nom
u t E k R k i t
k k s

=
=
(18)
+ is for discharge; - is for charge; where E(k) and R
w
(k) are taken from equation (17) for real
battery module.
Based on equation (18), the SOC calculation can be obtained in a direct way in online
dynamic battery voltage and current alterations. The solving (17) as high power factor
polynomial is really possible online by using two procedures: look-up table (dividing
polynomial function in shaped-line ranges) or bisection numerical iterative computation.
In some cases, when the accuracy of SOC indication can be lower (about 5%) , which is
accepted in HEV and EV drives, power factor of polynomial can be decreased by additional
approximation E(k) and R
w
(k). The accuracy of real time calculation is about 100 s.
The second method is bisection iterative calculation.
The exemplary plots of battery voltage, current and SOC is shown in following figures 23,
24, 25. Because The SOC of battery is much slower changeable than its voltage and current,
the SOC indication is computed and indicated by using moving average procedure.




Fig. 23. Exemplary test of battery load in hybrid drive ; blue battery current, green battery
voltage



Fig. 24. Exemplary test of battery SOC indication in real drive conditions corresponding to
battery load shown in Fig.23.
Nonlinear Dynamics

218

Fig. 25. Screen of control system based on dSpace programming for SOC indication.
6. Conclusions
The assumed method and effective model are very accurate according to error checking
results of the NiMH and Li-Ion batteries. The modeling method is valid for different types
of batteries. The model can be conveniently used for vehicle simulation because the battery
model is accurately approximated by mathematical equations. The model provides the
methodology for designing a battery management system and calculating the SOC. The
influence of temperature on battery performance is analyzed according to laboratory-tested
data and the theoretical background for the SOC calculation is obtained. The algorithm of
the battery SOC online indication considering the influence of temperature can be easily
used in practice by a microprocessor
7. References
[1] K. L. Butler, M. Ehsani, and P. Kamath, A matlab-based modeling and simulation
package for electric and hybrid electric vehicle design, IEEE Trans. Veh. Technol.,
vol. 48, no. 6, pp. 17701778, Nov. 1999.
[2] O. Caumont, P. L. Moigne, C. Rombaut, X. Muneret, and P. Lenain,Energy gauge for
lead acid batteries in electric vehicles, IEEE Trans. Energy Convers., vol. 15, no. 3,
pp. 354360, 2000.
[3] M. Ceraol and G. Pede, Techniques for estimating the residual range of an electric
vehicle, IEEE Trans. Veh. Technol., vol. 50, no. 1, pp. 109115,Jan. 2001.
[4] C. C. Chan, The state of the art of electric and hybrid vehicles, Proc.IEEE, vol. 90, no. 2,
pp. 247275, 2002.
[5] Valerie H. Johnson, Ahmad A. Pesaran, Temperature-dependent battery models for
high-power lithium-ion batteries, in Proc. International Electric Vehicle Symposium,
vol. 2, 2000, pp. 16.
Nonlinear Dynamics Traction Battery Modeling

219
[6] W. Gu and C. Wang, Thermal-electrochemical modeling of battery systems, Journal of
the Electrochemical Society vol.147, No.8, (2000), pp. 2910-22.
[7]
Szumanowski A. Fundamentals of hybrid vehicle drives Monograph Book, ISBN 83-
7204-114-8, Warsaw-Radom 2000.
[8]
Szumanowski A. Hybrid electric vehicle drives designedition based on urban buses
Monograph Book, ISBN 83-7204-456-2, Warsaw-Radom 2006.
[9] Robert F. Nelson, Power requirements for battery in HEVs, Journal of Power Sources,vol.
91, pp.2-26, 2000.
[10]
E. Karden, S. Buller, and R. W. De Doncker, A frequency-domain approach to
dynamical modeling of electrochemical power sources, Electrochimica Acta, vol.
47, no. 1314, pp. 23472356, 2002.D.
[11] J. Marcos, A. Lago, C. M. Penalver, J. Doval, A. Nogueira, C. Castro, and J. Chamadoira,
An approach to real behaviour modeling for traction lead-acid batteries, in Proc.
Power Electronics Specialists Conference, vol. 2, 2001, pp. 620624.
[12] A. Salkind, T. Atwater, P. Singh, S. Nelatury, S. Damodar, C. Fennie, and D. Reisner,
Dynamic characterization of small lead-acid cells, J. Power Sources, vol. 96, no. 1,
pp. 151159, 2001.
[13] G. Plett LiPB dynamic cell models for Kalman-Filter SOC estimation, Proc.
International Electric Vehicle Symposium, 2003, CD-ROM.
[14] S. Pang, J. Farrell, J. Du, and M. Barth, Battery state-of-charge estimation, in Proc.
American Control Conference, vol. 2, 2001, pp. 16441649.
[15] S. Malkhandi, S. K. Sinha, and K. Muthukumar, Estimation of state of charge of lead
acid battery using radial basis function, in Proc. Industrial Electronics Conference,
vol. 1, 2001, pp. 131136.
[16] S. Rodrigues, N. Munichandraiah, A. Shukla, A review of state-of-charge indication of
batteries by means of a.c. impedance measurements, Journal of Power Sources,
vol.87, No.1-2, 2000, pp.12-20.
[17] L. Jyunichi and T. Hiroya, Battery state-of-charge indicator for electric vehicle, in
Proc. International Electric Vehicle Symposium, vol. 2, 1996, pp. 229234.
[18] S. Sato and A. Kawamura, A new estimation method of state of charge using terminal
voltage and internal resistance for lead acid battery, in Proc. Power, vol. 2, 2002,
pp. 565570.
[19] W. X. Shen, C. C. Chan, E. W. C. Lo, and K. T. Chau, Estimation of battery available
capacity under variable discharge currents, J. Power Sources, vol. 103, no. 2, pp.
180187, 2002.
[20] W. X. Shen, K. T. Chau, C. C. Chan, Edward W. C. Lo, Neural network-based residual
capacity indicator for Nickel-Metal Hydride batteries in electric vehicles IEEE
Trans. Veh. Technol.,vol. 54, no. 5, pp. 17051712, Sep. 2005
[21] K. Morio, H. Kazuhiro, and P. Anil, Battery SOC and distance to empty meter of the
honda EV plus, in Proc. International Electric Vehicle Symposium, 1997, pp. 110.
[22] O. Caumont, P. L. Moigne, C. Rombaut, X. Muneret, and P. Lenain,Energy gauge for
lead-acid batteries in electric vehicles, IEEE Trans.Energy Convers., vol. 15, no. 3,
pp. 354360, Sep. 2000.
Nonlinear Dynamics

220
[23] Sabine Piller, Marion Perrin, Andreas Jossen Methods for stateofcharge
determination and their applications, Journal of Power Sources, vol. 96 , pp.113-120,
2001.
[24] Antoni Szumanowski, Jakub Dbicki, Arkadiusz Hajduga, Piotr Pirkowski, Chang
Yuhua, Li-ion battery modeling and monitoring approach for hybrid electric
vehicle applications, Proc. International Electric Vehicle Symposium, 2003, CD-ROM.
[25] Antoni Szumanowski, Yuhua Chang Battery Management System Based on Battery
Nonlinear Dynamics Modeling IEEE Transactions on Vehicular Technology, Vol.
57 no.3 May 2008

10
Entropic Geometry of Crowd Dynamics
Vladimir G. Ivancevic and Darryn J. Reid
Land Operations Division, Defence Science & Technology Organisation
Australia
1. Introduction
In this Chapter we propose a nonlinear entropic model of crowd generic psychophysical
1

dynamics. For this we use Feynmans actionamplitude formalism, operating on
microscopic, mesoscopic and macroscopic synergetic levels, which correspond to individual,
group (aggregate) and full crowd behavior dynamics, respectively. In all three levels, goal
directed behavior operates under entropy conservation,
t
S = 0, while naturally chaotic
behavior operates under (monotonically) increasing entropy,
t
S > 0. Between these two
distinct behavioral phases lies a topological phase transition with a chaotic inter-phase. We
formulate a geometrical representation of this behavioral transition in terms of the
Perelman-Ricci flow on the crowds Riemannian configuration manifold.
Recall that in psychology the term cognition
2
refers to an information processing view of an
individual psychological functions (see [3; 4; 68; 81; 88]). More generally, cognitive processes
can be natural and artificial, conscious and not conscious; therefore, they are analyzed from
different perspectives and in different contexts, e.g., anesthesia, neurology, psychology,
philosophy, logic (both Aristotelian and mathematical), systemics, computer science,
artificial intelligence (AI) and computational intelligence (CI). Both in psychology and in
AI/CI, cognition refers to the mental functions, mental processes and states of intelligent
entities (humans, human organizations, highly autonomous robots), with a particular focus
toward the study of comprehension, inferencing, decisionmaking, planning and learning (see,
e.g. [11]). The recently developed Scholarpedia, the free peer reviewed web encyclopedia of
computational neuroscience is largely based on cognitive neuroscience (see, e.g. [79]). The
concept of cognition is closely related to such abstract concepts as mind, reasoning, perception,
intelligence, learning, and many others that describe numerous capabilities of the human mind
and expected properties of AI/CI (see [51; 57] and references therein).
Yet disembodied cognition is a myth, albeit one that has had profound influence in Western
science since Rene Descartes and others gave it credence during the Scientific Revolution. In
fact, the mind-body separation had much more to do with explanation of method than with
explanation of the mind and cognition, yet it is with respect to the latter that its impact is most
widely felt. We find it to be an unsustainable assumption in the realm of crowd behavior.

1
The new term psychophysical should not be confused with the reserved psychological
term psychophysics. By psycho-physical we mean cognitivetophysical transition
behavior: from mental idea to physical manifestation.
2
Latin: cognoscere = to know
Nonlinear Dynamics

222
Mental intention is (almost immediately) followed by a physical action, that is, a human or
animal movement [82]. In animals, this physical action would be jumping, running, flying,
swimming, biting or grabbing. In humans, it can be talking, walking, driving, or shooting, etc.
Mathematical description of human/animal movement in terms of the corresponding neuro-
musculo-skeletal equations of motion, for the purpose of prediction and control, is formulated
within the realm of biodynamics (see [43; 44; 45; 46; 47; 48; 49; 55]).
The crowd (or, collective) behavior is clearly formed by some kind of superposition, contagion,
emergence, or convergence from the individual agents behavior. Le Bons 1895 contagion
theory, presented in The Crowd: A Study of the Popular Mind influenced many 20th
century figures. Sigmund Freud criticized Le Bons concept of collective soul, asserting that
crowds do not have a soul of their own. The main idea of Freudian crowd behavior theory was
that people who were in a crowd acted differently towards people than those who were
thinking individually: the minds of the group would merge together to form a collective way
of thinking. This idea was further developed in Jungian famous collective unconscious [63].
The term collective behavior [8] refers to social processes and events which do not reflect
existing social structure (laws, conventions, and institutions), but which emerge in a
spontaneous way. Collective behavior might also be defined as action which is neither
conforming (in which actors follow prevailing norms) nor deviant (in which actors violate
those norms). According to the emergence theory [86], crowds begin as collectivities composed
of people with mixed interests and motives; especially in the case of less stable crowds
(expressive, acting and protest crowds) norms may be vague and changing; people in crowds
make their own rules as they go along. According to currently popular convergence theory,
crowd behavior is not a product of the crowd itself, but is carried into the crowd by particular
individuals, thus crowds amount to a convergence of likeminded individuals.
We propose that the contagion and convergence theories may be unified by acknowledging
that both factors may coexist, even within a single scenario: we propose to refer to this third
approach as behavioral composition. It represents a substantial philosophical shift from
traditional analytical approaches, which have assumed either reduction of a whole into
parts or the emergence of the whole from the parts. In particular, both contagion and
convergence are related to social entropy, which is the natural decay of structure (such as
law, organization, and convention) in a social system [16]. Thus, social entropy provides an
entry point into realizing a behavioralcompositional theory of crowd dynamics.
Thus, while all mentioned psycho-social theories of crowd behavior are explanatory only, in
this paper we attempt to formulate a geometrically predictive modeltheory of crowd
psychophysical behavior.
In this chapter we attempt to formulate a geometrically predictive modeltheory of crowd
behavioral dynamics, based on the previously formulated individual Life Space Foam
concept [54].
3


3
General nonlinear stochastic dynamics, developed in a framework of Feynman path
integrals, have recently [54] been applied to Lewinian fieldtheoretic psychodynamics [67],
resulting in the development of a new concept of lifespace foam (LSF) as a natural medium
for motivational and cognitive psychodynamics. According to the LSFformalism, the
classic Lewinian life space can be macroscopically represented as a smooth manifold with
steady forcefields and behavioral paths, while at the microscopic level it is more
realistically represented as a collection of wildly fluctuating forcefields, (loco)motion paths
and local geometries (and topologies with holes).
Entropic Geometry of Crowd Dynamics

223
It is today well known that massive crowd movements can be precisely observed/moni-
tored from satellites and all that one can see is crowd physics. Therefore, all involved
psychology of individual crowd agents: cognitive, motivational and emotional is only a

A set of leastaction principles is used to model the smoothness of global, macrolevel LSF
paths, fields and geometry, according to the following prescription. The action S[], with
dimensions of Energy Time = Effort and depending on macroscopic paths, fields and
geometries (commonly denoted by an abstract field symbol
i
) is defined as a temporal
integral from the initial time instant t
ini
to the final time instant t
f in
,
[ ] = [ ] ,
t
fin
t
ini
S dt

L (1)
with Lagrangian density given by
[ ] = ( , ),
j
n i
i
x
d x

L L
where the integral is taken over all n coordinates x
j
= x
j
(t) of the LSF, and
j
i
x
are time and
space partial derivatives of the
i
-variables over coordinates. The standard least action
principle
[ ] = 0, S (2)
gives, in the form of the socalled EulerLagrangian equations, a shortest (loco)motion path,
an extreme forcefield, and a lifespace geometry of minimal curvature (and without holes).
In this way, we have obtained macroobjects in the global LSF: a single path described by
Newtonianlike equation of motion, a single forcefield described by Maxwellianlike field
equations, and a single obstaclefree Riemannian geometry (with global topology without
holes).
To model the corresponding local, microlevel LSF structures of rapidly fluctuating MD &
CD, an adaptive path integral is formulated, defining a multiphase and multipath (multi
field and multi geometry) transition amplitude from the motivational state of Intention to
the cognitive state of Action,

i [ ]
| := [ ]e ,
S
total
Action Intention w

D (3)
where the Lebesgue integration is performed over all continuous =
i
con
paths + fields +
geometries, while summation is performed over all discrete processes and regional topologies
j
dis
. The symbolic differential D[w] in the general path integral (24), represents an
adaptive path measure, defined as a weighted product

=1
[ ] = lim ,( = 1,..., = ).
N
i
s s
N
s
w w d i n con dis

D (4)
The adaptive path integral (3)(11) represents an dimensional neural network, with
weights w updating by the general rule [57]
new value(t + 1) = old value(t) + innovation(t).
Nonlinear Dynamics

224
non-transparent input (a hidden initial switch) for the fully observable crowd physics. In
this paper we will label this initial switch as mental preparation or loading, while the
manifested physical action is labeled hitting.
We propose the entropy formulation of crowd dynamics as a threestep process involving
individual behavioral dynamics and collective behavioral dynamics. The chaotic behavioral
phase transitions embedded in crowd dynamics may give a formal description for a
phenomenon called crowd turbulence by D. Helbing, depicting crowd disasters caused by the
panic stampede that can occur at high pedestrian densities and which is a serious concern
during mass events like soccer championship games or annual pilgrimage in Makkah (see
[37; 38; 39; 62]).
In this paper we propose the entropy formulation of crowd dynamics as a threestep
process involving individual dynamics and collective dynamics.
2. Generic threestep crowd psychophysical behavior
In this section we model a generic crowd dynamics (see e.g., [36; 69]) as a threestep process
based on a general partition function formalism. Note that the number of variables X
i
in the
standard partition function from statistical mechanics (see equation (59) in Appendix) need
not be countable, in which case the set of coordinates {x
i
} becomes a field
= (x), so the sum is to be replaced by the Euclidean path integral (that is a Wickrotated
Feynman transition amplitude in imaginary time, see subsection 3.4), as
[ ] ( ) = [ ]exp ( ) , Z H

D
More generally, in quantum field theory, instead of the field Hamiltonian H() we have the
action S() of the theory. Both Euclidean path integral,
[ ] ( ) = [ ]exp ( ) , real path integral in imaginary time Z S

D (5)
and Lorentzian one,
[ ] ( ) = [ ]exp ( ) , complex path integral in real time Z iS

D (6)
r epresent quantum field theory (QFT) partition functions. We will give formal definitions
of the above path integrals (i.e., general partition functions) in section 3. For the moment, we
only remark that the Lorentzian path integral (6) represents a QFT generalization of the
(nonlinear) Schrdinger equation, while the Euclidean path integral (5) in the (rectified) real
time represents a statistical field theory (SFT) generalization of the FokkerPlanck equation.
Now, following the framework of the Extended Second Law of Thermodynamics (see
Appendix),
t
S 0, for entropy S in any complex system described by its partition function,
we formulate a generic crowd dynamics, based on above partition functions, as the
following threestep process:
1. Individual dynamics (ID) is a transition process from an entropygrowing loading
phase of mental preparation, to the entropyconserving hitting/executing phase of
physical action. Formally, ID is given by the phasetransition map:

"LOADING": >0 "HITTING": =0
: MENTAL PREPARATION PHYSICAL ACTION
S S
t t

ID (7)
Entropic Geometry of Crowd Dynamics

225
defined by the individual (chaotic) phasetransition amplitude
=0 >0
[ ]
ID
ID
PHYS. ACTION MENTAL PREP. := [ ]e ,
S S
t t
iS
CHAOS

D
where the right-hand-side is the Lorentzian path-integral (or complex path-integral in
real time, see Appendix), with the individual action
ID ID
[ ] = [ ] ,
t
fin
t
ini
S L dt

where L
ID
[] is the behavioral Lagrangian, consisting of mental cognitive potential and
physical kinetic energy.
2. Aggregate dynamics (AD) represents the behavioral compositiontransition map:

"LOADING": >0 "HITTING": =0
AD AD
: MENTAL PREPARATION PHYSICAL ACTION
S S
t t
i
i i

AD (8)
where the (weighted) aggregate sum is taken over all individual agents, assuming
equipartition of the total energy. It is defined by the aggregate (chaotic) phase
transition amplitude
=0 >0
[ ]
AD
AD
PHYS. ACTION MENTAL PREP. := [ ]e ,
S S
t t
S
CHAOS

D
with the Euclidean path-integral in real time, that is the SFTpartition function, based
on the aggregate behavioral action
AD AD AD ID
AD
[ ] = [ ] , with [ ] = [ ].
t
fin i
t
ini
i
S L dt L L

3. Crowd dynamics (CD) represents the cumulative transition map:

"LOADING": >0 "HITTING": =0
CD CD
: MENTAL PREPARATION PHYSICAL ACTION
S S
t t
i
i i

CD (9)
where the (weighted) cumulative sum is taken over all individual agents, assuming
equipartition of the total behavioral energy. It is defined by the crowd (chaotic) phase
transition amplitude
=0 >0
[ ]
CD
CD
PHYS. ACTION MENTAL PREP. := [ ]e ,
S S
t t
iS
CHAOS

D
with the general Lorentzian path-integral, that is, the QFTpartition function), based on
the crowd behavioral action
CD CD CD ID AD
CD =#ofADsinCD
[ ] = [ ] , with [ ] = [ ] = [ ].
t
fin i k
t
ini
i k
S L dt L L L

Nonlinear Dynamics

226
All three entropic phasetransition maps, ID, AD and CD, are spatiotemporal biodynamic
cognition systems, evolving within their respective configuration manifolds (i.e., sets of their
respective degrees-of-freedom with equipartition of energy), according to biphasic action
functional formalisms with behavioral Lagrangian functions L
ID
, L
AD
and L
CD
, each
consisting of:
1. Cognitive mental potential (which is a mental preparation for the physical action), and
2. Physical kinetic energy (which describes the physical action itself).
To develop ID, AD and CD formalisms, we extend into a physical (or, more precisely,
biodynamic) crowd domain a purelymental individual LifeSpace Foam (LSF) framework
for motivational cognition [54], based on the quantumprobability concept.
4


4
The quantum probability concept is based on the following physical facts [58; 59]
1. The timedependent Schrdinger equation represents a complexvalued generalization
of the realvalued FokkerPlanck equation for describing the spatiotemporal
probability density function for the system exhibiting continuoustime Markov
stochastic process.
2. The Feynman path integral (including integration over continuous spectrum and
summation over discrete spectrum) is a generalization of the timedependent
Schrdinger equation, including both continuoustime and discretetime Markov
stochastic processes.
3. Both Schrdinger equation and path integral give physical description of any system
they are modelling in terms of its physical energy, instead of an abstract probabilistic
description of the FokkerPlanck equation.
Therefore, the Feynman path integral, as a generalization of the (nonlinear) timedependent
Schrdinger equation, gives a unique physical description for the general Markov stochastic
process, in terms of the physically based generalized probability density functions, valid
both for continuoustime and discretetime Markov systems. Its basic consequence is this: a
different way for calculating probabilities. The difference is rooted in the fact that sum of
squares is different from the square of sums, as is explained in the following text. Namely, in
DiracFeynman quantum formalism, each possible route from the initial system state A to
the final system state B is called a history. This history comprises any kind of a route,
ranging from continuous and smooth deterministic (mechanicallike) paths to completely
discontinues and random Markov chains (see, e.g., [23]). Each history (labelled by index i) is
quantitatively described by a complex number.
In this way, the overall probability of the systems transition from some initial state A to
some final state B is given not by adding up the probabilities for each historyroute, but by
headtotail adding up the sequence of amplitudes makingup each route first (i.e.,
performing the sumoverhistories) to get the total amplitude as a resultant vector, and
then squaring the total amplitude to get the overall transition probability.
Here we emphasize that the domain of validity of the quantum is not restricted to the
microscopic world [87]. There are macroscopic features of classically behaving systems,
which cannot be explained without recourse to the quantum dynamics. This field theoretic
model leads to the view of the phase transition as a condensation that is comparable to the
formation of fog and rain drops from water vapor, and that might serve to model both the
gamma and beta phase transitions. According to such a model, the production of activity
with longrange correlation in the brain takes place through the mechanism of spontaneous
Entropic Geometry of Crowd Dynamics

227
The behavioral dynamics approach to ID, AD and CD is based on entropic motor control [41;
42], which deals with neuro-physiological feedback information and environmental
uncertainty. The probabilistic nature of human motor action can be characterized by
entropies at the level of the organism, task, and environment. Systematic changes in motor
adaptation are characterized as taskorganism and environmentorganism tradeoffs in
entropy. Such compensatory adaptations lead to a view of goaldirected motor control as
the product of an underlying conservation of entropy across the taskorganism
environment system. In particular, an experiment conducted in [42] examined the changes
in entropy of the coordination of isometric force output under different levels of task
demands and feedback from the environment. The goal of the study was to examine the
hypothesis that human motor adaptation can be characterized as a process of entropy
conservation that is reflected in the compensation of entropy between the task, organism
motor output, and environment. Information entropy of the coordination dynamics relative
phase of the motor output was made conditional on the idealized situation of human
movement, for which the goal was always achieved. Conditional entropy of the motor
output decreased as the error tolerance and feedback frequency were decreased. Thus, as
the likelihood of meeting the task demands was decreased increased task entropy and/or
the amount of information from the environment is reduced increased environmental
entropy, the subjects of this experiment employed fewer coordination patterns in the force
output to achieve the goal. The conservation of entropy supports the view that context
dependent adaptations in human goaldirected action are guided fundamentally by natural
law and provides a novel means of examining human motor behavior. This is
fundamentally related to the Heisenberg uncertainty principle [59] and further supports the
argument for the primacy of a probabilistic approach toward the study of biodynamic
cognition systems.
5


breakdown of symmetry (SBS), which has for decades been shown to describe longrange
correlation in condensed matter physics. The adoption of such a field theoretic approach
enables modelling of the whole cerebral hemisphere and its hierarchy of components down to
the atomic level as a fully integrated macroscopic quantum system, namely as a macroscopic
system which is a quantum system not in the trivial sense that it is made, like all existing
matter, by quantum components such as atoms and molecules, but in the sense that some of its
macroscopic properties can best be described with recourse to quantum dynamics (see [22]
and references therein). Also, according to Freeman and Vitielo, manybody quantum field theory
appears to be the only existing theoretical tool capable to explain the dynamic origin of long
range correlations, their rapid and efficient formation and dissolution, their interim stability in
ground states, the multiplicity of coexisting and possibly noninterfering ground states, their
degree of ordering, and their rich textures relating to sensory and motor facets of behaviors. It
is historical fact that manybody quantum field theory has been devised and constructed in
past decades exactly to understand features like ordered pattern formation and phase
transitions in condensed matter physics that could not be understood in classical physics,
similar to those in the brain.
5
Our entropic actionamplitude formalism represents a kind of a generalization of the
Haken-Kelso- Bunz (HKB) model of self-organization in the individuals motor system [24;
65], including: multistability, phase transitions and hysteresis effects, presenting a contrary
view to the purely feedback driven systems. HKB uses the concepts of synergetics (order
Nonlinear Dynamics

228
On the other hand, it is well known that humans possess more degrees of freedom than are
needed to perform any defined motor task, but are required to co-ordinate them in order to
reliably accomplish high-level goals, while faced with intense motor variability. In an
attempt to explain how this takes place, Todorov and Jordan have formulated an alternative
theory of human motor co-ordination based on the concept of stochastic optimal feedback
control [84]. They were able to conciliate the requirement of goal achievement (e.g., grasping
an object) with that of motor variability (biomechanical degrees of freedom). Moreover, their
theory accommodates the idea that the human motor control mechanism uses internal
functional synergies to regulate taskirrelevant (redundant) movement.
Also, a developing field in coordination dynamics involves the theory of social coordination,
which attempts to relate the DC to normal human development of complex social cues
following certain patterns of interaction. This work is aimed at understanding how human
social interaction is mediated by meta-stability of neural networks. fMRI and EEG are
particularly useful in mapping thalamocortical response to social cues in experimental
studies. In particular, a new theory called the Phi complex has been developed by S. Kelso
and collaborators, to provide experimental results for the theory of social coordination
dynamics (see the recent nonlinear dynamics paper discussing social coordination and EEG
dynamics [85]). According to this theory, a pair of phi rhythms, likely generated in the
mirror neuron system, is the hallmark of human social coordination. Using a dualEEG
recording system, the authors monitored the interactions of eight pairs of subjects as they
moved their fingers with and without a view of the other individual in the pair.
Finally, the chaotic behavioral phase transitions embedded in CD may give a formal
description for a phenomenon called crowd turbulence by D. Helbing, depicting crowd
disasters caused by the panic stampede that can occur at high pedestrian densities and

parameters, control parameters, instability, etc) and the mathematical tools of nonlinearly
coupled (nonlinear) dynamical systems to account for self-organized behavior both at the
cooperative, coordinative level and at the level of the individual coordinating elements. The
HKB model stands as a building block upon which numerous extensions and elaborations
have been constructed. In particular, it has been possible to derive it from a realistic model
of the cortical sheet in which neural areas undergo a reorganization that is mediated by
intra- and inter-cortical connections. Also, the HKB model describes phase transitions
(switches) in coordinated human movement as follows: (i) when the agent begins in the
anti-phase mode and speed of movement is increased, a spontaneous switch to symmetrical,
in-phase movement occurs; (ii) this transition happens swiftly at a certain critical frequency;
(iii) after the switch has occurred and the movement rate is now decreased the subject
remains in the symmetrical mode, i.e. she does not switch back; and (iv) no such transitions
occur if the subject begins with symmetrical, in-phase movements. The HKB dynamics of
the order parameter relative phase as is given by a nonlinear first-order ODE:
2 2
=( 2 )sin sin2 , r r +
`

where is the phase relation (that characterizes the observed patterns of behavior, changes
abruptly at the transition and is only weakly dependent on parameters outside the phase
transition), r is the oscillator amplitude, while , are coupling parameters (from which the
critical frequency where the phase transition occurs can be calculated).

Entropic Geometry of Crowd Dynamics

229
which is a serious concern during mass events like soccer championship games or annual
pilgrimage in Makkah (see [37; 38; 39; 62]).
3. Formal crowd dynamics
In this section we formally develop a threestep crowd behavioral dynamics, conceptualized
by transition maps (7)(8)(9), in agreement with Hakens synergetics [25; 26]. We first
develop a macrolevel individual behavioral dynamics ID. Then we generalize ID into an
orchestrated behavioralcompositional crowd dynamics CD, using a quantumlike micro
level formalism with individual agents representing crowd quanta. Finally we develop a
mesolevel aggregate statisticalfield dynamics AD, such that composition of the aggregates
AD makesup the crowd.
3.1 Individual behavioral dynamics (ID)
ID transition map (7) is developed using the following actionamplitude formalism (see [53;
54]):
1. Macroscopically, as a smooth Riemannian nmanifold M
ID
(see Appendix) with steady
forcefields and behavioral paths, modelled by a realvalued classical action functional
S
ID
[], of the form
I I
[ ] = [ ] ,
t
fin
D D
t
ini
S L dt

(where macroscopic paths, fields and geometries are commonly denoted by an abstract
field symbol
i
) with the potentialenergy based Lagrangian L given by
I I
[ ] = ( , ),
n i
D D i j
x
L d x

L
where L is Lagrangian density, the integral is taken over all n local coordinates x
j
= x
j
(t)
of the ID, and
x
j
i
are time and space partial derivatives of the
i
variables over
coordinates. The standard least action principle
I
[ ] = 0,
D
S
gives, in the form of the EulerLagrangian equations, a shortest path, an extreme force
field, with a geometry of minimal curvature and topology without holes. We will see
below that high Riemannian curvature generates chaotic behavior, while holes in the
manifold produce topologically induced phase transitions.
2. Microscopically, as a collection of wildly fluctuating and jumping paths (histories),
forcefields and geometries/topologies, modelled by a complexvalued adaptive path
integral, formulated by defining a multiphase and multipath (multifield and multi
geometry) transition amplitude from the entropygrowing state of Mental Preparation
to the entropyconserving state of Physical Action,

[ ]
ID
ID
ID
Physical Action|Mental Preparation := [ ]e
iS

D (10)
where the functional IDmeasure D[w] is defined as a weighted product
Nonlinear Dynamics

230

=1
[ ] = lim , ( = 1,..., = ),
N
i
s s
N
s
w w d i n con dis

D (11)
representing an dimensional neural network [54], with weights w
s
updating by the
general rule
new value(t + 1) = old value(t) + innovation(t).
More precisely, the weights w
s
= w
s
(t) in (11) are updated according to one of the two
standard neural learning schemes, in which the microtime level is traversed in discrete
steps, i.e., if t = t
0
, t
1
, ..., t
s
then t + 1 = t
1
, t
2
, ..., t
s+1
:
6

a. A selforganized, unsupervised (e.g., Hebbianlike [35]) learning rule:
( 1) = ( ) ( ( ) ( )),
d a
s s s s
w t w t w t w t

+ + (12)
where = (t), = (t) denote signal and noise, respectively, while superscripts d
and a denote desired and achieved microstates, respectively; or
b. A certain form of a supervised gradient descent learning:
( 1)= ( ) ( ),
s s
w t w t J t + (13)
where is a small constant, called the step size, or the learning rate, and J(n)
denotes the gradient of the performance hypersurface at the tth iteration.
(Note that we could also use a rewardbased, reinforcement learning rule [83], in which
system learns its optimal policy: innovation(t) = |reward(t) penalty(t)|. )
In this way, we effectively derive a unique and globally smooth, causal and entropic phase
transition map (7), performed at a macroscopic (global) timelevel from some initial time t
ini

to the final time t
fin
. Thus, we have obtained macroobjects in the ID: a single path described
by Newtonianlike equation of motion, a single forcefield described by Maxwellianlike
field equations, and a single obstaclefree Riemannian geometry (with global topology
without holes).
In particular, on the macrolevel, we have the IDpaths, that is biodynamical trajectories
generated by the Hamilton action principle
I
[ ] = 0,
D
S x
with the Newtonian action S
ID
[x] given by (Einsteins summation convention over repeated
indices is always assumed)

I
1
[ ] = [ ] ,
2
t
fin j i
D ij
t
ini
S x g x x dt +

` ` (14)

6
The traditional neural networks approaches are known for their classes of functions they
can represent. Here we are talking about functions in an extensional rather than merely
intensional sense; that is, function can be read as input/output behavior [5; 6; 19; 34]. This
limitation has been attributed to their low-dimensionality (the largest neural networks are
limited to the order of 10
5

dimensions [61]). The proposed path integral approach represents
a new family of function-representation methods, which potentially offers a basis for a
fundamentally more expansive solution.
Entropic Geometry of Crowd Dynamics

231

Fig. 1. Riemannian configuration manifold M
ID
of human biodynamics is defined as a
topological product M =

i
SE(3)
i
of constrained Euclidean SE(3)groups of rigid body
motion in 3D Euclidean space (see [49; 52]), acting in all major (synovial) human joints. The
manifold M is a dynamical structure activated/controlled by potential covariant forces (16)
produced by a synergetic action of about 640 skeletal muscles [47].

where = (t, x
i
) denotes the mental LSFpotential field, while the second term,
1
= ,
2
j i
ij
T g x x ` `
represents the physical (biodynamic) kinetic energy generated by the Riemannian inertial
metric tensor g
ij
of the configuration biodynamic manifold M
ID
(see Figure 1). The
corresponding EulerLagrangian equations give the Newtonian equations of human
movement
= ,
i i i
x x
d
T T F
dt

`
(15)
where subscripts denote the partial derivatives and we have defined the covariant muscular
forces F
i

= F
i
(t, x
i
,
i
x` ) as negative gradients of the mental potential (x
i
),
= .
i i
x
F (16)
Equation (15) can be put into the standard Lagrangian form as
= , with = ( ),
i
i i
x x
d
L L L T x
dt

`
(17)
Nonlinear Dynamics

232
or (using the Legendre transform) into the forced, dissipative Hamiltonian form [44; 47]
= , = ,
i
p p i i i i
i i x x
x H R p F H R + + ` ` (18)
where p
i

are the generalized momenta (canonicallyconjugate to the coordinates x
i
),
H = H(p, x) is the Hamiltonian (total energy function) and R = R(p, x) is the general
dissipative function.
The human motor system possesses many independently controllable components that
often allow for more than a single movement pattern to be performed in order to achieve a
goal.
Hence, the motor system is endowed with a high level of adaptability to different tasks and
also environmental contexts [42]. The multiple SE(3)dynamics applied to human musculo
skeletal system gives the fundamental law of biodynamics, which is the covariant force law:
Force co vector field = Mass distribution Acceleration vector field, (19)
which is formally written:
= , ( , = 1,..., = dim( ))
j
i ij
F g a i j n M
where F
i
are the covariant force/torque components, g
ij
is the inertial metric tensor of the
configuration Riemannian manifold M =

i
SE(3)
i

(g
ij
defines the massdistribution of the
human body), while a
j

are the contravariant components of the linear and angular
acceleration vector-field. (This fundamental biodynamic law states that contrary to common
perception, acceleration and force are not quantities of the same nature: while acceleration is
a non-inertial vector-field, force is an inertial co-vector-field. This apparently insignificant
difference becomes crucial in injury prediction/prevention, especially in its derivative form
in which the massless jerk (= a` ) is relatively benign, while the massive jolt (= F
`
) is
deadly.) Both Lagrangian and (topologically equivalent) Hamiltonian development of the
covariant force law is fully elaborated in [47; 48; 49; 52]. This is consistent with the
postulation that human action is guided primarily by natural law [66].
On the microID level, instead of each single trajectory defined by the Newtonian equation
of motion (15), we have an ensemble of fluctuating and crossing paths on the configuration
manifold M with weighted probabilities (of the unit total sum). This ensemble of micro
paths is defined by the simplest instance of our adaptive path integral (10), similar to the
Feynmans original sum over histories,

i [ ]
I
P | = [ ]e ,
S x
M
D
hysical Action Mental Preparation wx

D (20)
where D[wx] is the functional IDmeasure on the space of all weighted paths, and the
exponential depends on the action S
ID
[x] given by (14).
3.2 Crowd behavioralcompositional dynamics (CD)
In this subsection we develop a generic crowd CD, as a unique and globally smooth, causal
and entropic phasetransition map (9), in which agents (or, crowds individual entities) can
Entropic Geometry of Crowd Dynamics

233
be both humans and robots. This crowd behavioral action takes place in a crowd smooth
Riemannian 3n-manifold M. Recall from Figure 1 that each individual segment of a human
body moves in the Euclidean 3space R
3

according to its own constrained SE(3)group.
Similarly, each individual agents trajectory, x
i
= x
i
(t), i = 1, ...n, is governed by the Euclidean
SE(2)group of rigid body motions in the plane. (Recall that a Lie group SE(2) SO(2) R is
a set of all 3 3 matrices of the form:

cos sin
sin cos ,
0 0 1
x
y

including both rigid translations (i.e., Cartesian x,ycoordinates) and rotation matrix
cos sin
sin cos

in Euclidean plane R
2
(see [49; 52]). The crowd configuration manifold M is
defined as a union of Euclidean SE(2)groups for all n individual agents in the crowd, that is
crowds configuration 3nmanifold is defined as a set


=1 =1
= (2) (2) ,
n n
k k k
k k
M SE SO

R (21)
coordinated by , , }, (for = 1, 2,..., ).
k k k k
x y k n x = {

In other words, the crowd configuration manifold M is a dynamical planar graph with
individual agents SE(2)groups of motion in the vertices and time-dependent inter-agent
distances = ( ) ( )
j i
ij i j
I x t x t

as edges.
Similarly to the individual case, the crowd action functional includes mental cognitive
potential and physical kinetic energy, formally given by (with i, j = 1, ..., 3n):


2
1 1
[ , ; , ] = ( ) ( ) ( ) ( ) ( ) ,
2 2
j j j i i i
i j ij i j i j ij
t t t
i j
A x x t t I x t x t dt dt g x t x t dt +

` ` ` ` (22)
2
2
with = ( ) ( ) , where , , .
j i
ij i j i j
I x t x t IN t t t OUT



The first term in (22) represents the mental potential for the interaction between any two
agents x
i
and x
i
within the total crowd matrix x
ij
. (Although, formally, this term contains
cognitive velocities, it still represents potential energy from the physical point of view.) It is
defined as a double integral over a delta function of the square of interval I
2
between two
points on the paths in their individual cognitive LSFs. Interaction occurs only when this
LSF distance between the two agents x
i
and x
j
vanishes. Note that the cognitive intentions
of any two agents generally occur at different times t
i

and t
j

unless t
i

= t
j
, when cognitive
synchronization occurs. This term effectively represents the crowd cognitive controller (see [53]).
Nonlinear Dynamics

234
The second term in (22) represents kinetic energy of the physical interaction of agents.
Namely, after the above cognitive synchronization is completed, the second term of physical
kinetic energy is activated in the common CD manifold, reducing it to just one of the agents
individual manifolds, which is equivalent to the center-of-mass segment in the human
musculo-skeletal system. Therefore, from (22) we can derive a generic EulerLagrangian
dynamics that is a composition of (17), which also means that we have in place a generic
Hamiltonian dynamics that is a amalgamate of (18), as well as the crowd covariant force law
(19), the governing law of crowd biodynamics:

Crowd force co vector field = Crowd mass distribution Crowd acceleration vector field,
formally: = , where is the inertial metric tensor of crowd manifold .
j
i ij ij
F g a g M (23)

The left-hand side of this equation defines forces acting on the crowd, while right-hand
defines its mass distribution coupled to the crowd kinematics (CK, described in the next
subsection).
At the slave level, the adaptive path integral, representing an dimensional neural
network, corresponding to the crowd behavioral action (22), reads


[ , ; , ]
CD
C
P | = [ , , ]e ,
iA x y t t
i j
D
hysical Action Mental Preparation w x y

D (24)

where the Lebesgue-type integration is performed over all continuous paths x
i
= x
i
(t
i
) and
y
j
= y
j
(t
j
), while summation is performed over all associated discrete Markov fluctuations
and jumps. The symbolic differential in the path integral (24) represents an adaptive path
measure, defined as the weighted product


=1
[ , , ] = , ( , = 1,..., ).
lim
N
j s i
ij
N
s
w x y w dx dy i j n

D (25)

The quantumfield path integral (24)(25) defines the microstate CDlevel, an ensemble of
fluctuating and crossing paths on the crowd 3nmanifold M.
The crowd manifold M itself has quite a sophisticated topological structure defined by its
macrostate EulerLagrangian dynamics. As a Riemannian smooth nmanifold, M gives rise
to its fundamental ngroupoid, or ncategory

n
(M) (see ([49; 52]). In

n
(M), 0cells are
points in M; 1cells are paths in M(i.e., parameterized smooth maps f : [0,1]M); 2cells are
smooth homotopies (denoted by ) of paths relative to endpoints (i.e., parameterized
smooth maps h : [0,1] [0,1] M); 3cells are smooth homotopies of homotopies of paths in
M (i.e., parameterized smooth maps j : [0,1] [0,1] [0,1] M). Categorical composition is
defined by pasting paths and homotopies. In this way, the following recursive homotopy
dynamics emerges on the crowd 3nmanifold M:

Entropic Geometry of Crowd Dynamics

235


3.3 Dissipative crowd kinematics (CD)
The crowd action (22) with its amalgamate Lagrangian dynamics (17) and amalgamate
Hamiltonian dynamics (18), as well as the crowd force law (23) define the macroscopic
crowd dynamics, CD. Suppose, for a moment, that CD is forcefree and dissipation free,
therefore conservative. Now, the basic characteristic of the conservative
Lagrangian/Hamiltonian systems evolving in the phase space spanned by the system
coordinates and their velocities/momenta, is that their flow
L
t
(explained below) preserves
the phasespace volume, as proposed by the Liouville theorem, which is the well known
fact in statistical mechanics. However, the preservation of the phase volume causes
structural instability of the conservative system, i.e., the phasespace spreading effect by
which small phase regions R
t
will tend to get distorted from the initial one R
o
during the
conservative system evolution. This problem, governed by entropy growth (
t
S > 0), is much
Nonlinear Dynamics

236
more serious in higher dimensions than in lower dimensions, since there are so many
directions in which the region can locally spread (see [49; 74]). This phenomenon is related
to conservative Hamiltonian chaos (see section 4 below).
However, this situation is not very frequent in case of organized human crowd. Its self-
organization mechanisms are clearly much stronger than the conservative statistical
mechanics effects, which we interpret in terms of Prigogines dissipative structures (see
Appendix). Formally, if dissipation of energy in a system is much stronger then its inertial
characteristics, then instead of the second-order NewtonLagrangian dynamic equations of
motion, we are actually dealing with the first-order driftless (non-acceleration, non-inertial)
kinematic equations of motion (see Appendix, eq. (64)), which is related to dissipative chaos
[71]. Briefly, the dissipative crowd flow can be depicted like this: from the set of initial
conditions for individual agents, the crowd evolves in time towards the set of the
corresponding entangled attractors,
7
which are mutually separated by fractal (non-integer
dimension) separatrices.
In this subsection we elaborate on the dissipative crowd kinematics (CK), which is self
controlled and dominates the CD if the crowds inertial forces are much weaker then the
crowds dissipation of energy, presented here in the form of nonlinear velocity controllers.

7
Recall that quantum entanglement is a quantum mechanical phenomenon in which the
quantum states of two or more objects are linked together so that one object can no longer be
adequately described without full mention of its counterpart even though the individual
objects may be spatially separated. This interconnection leads to correlations between
observable physical properties of remote systems. The related phenomenon of wave-function
collapse gives an impression that measurements performed on one system instantaneously
influence the other systems entangled with the measured system, even when far apart.
Entanglement has many applications in quantum information theory. Mixed state
entanglement can be viewed as a resource for quantum communication. A common
measure of entanglement is the entropy of a mixed quantum state (see, e.g. [59]). Since a
mixed quantum state is a probability distribution over a quantum ensemble, this leads
naturally to the definition of the von Neumann entropy, S() = Tr (log
2
) , which is
obviously similar to the classical Shannon entropy for probability distributions (p
1
, , p
n
),
defined as S(p
1
, , p
n
) =
i
p
i
log
2
p
i
. As in statistical mechanics, one can say that the more
uncertainty (number of microstates) the system should possess, the larger is its entropy.
Entropy gives a tool which can be used to quantify entanglement. If the overall system is
pure, the entropy of one subsystem can be used to measure its degree of entanglement with
the other subsystems.
The most popular issue in a research on dissipative quantum brain modelling has been
quantum entanglement between the brain and its environment [77; 78], where the brain
environment system has an entangled memory state, identified with the ground (vacuum)
state |0 >N, that cannot be factorized into two singlemode states. (In the VitielloPessa
dissipative quantum brain model [77; 78], the evolution of the Ncoded memory system was
represented as a trajectory of given initial condition running over timedependent states
|0(t) >N, each one minimizing the free energy functional.) Similar to this microscopic brain
environment entanglement, we propose a kind of macroscopic entanglement between the
operating modes of the crowd behavioral controller and its biodynamics, which can be
considered as a longrange correlation.
Applied externally to the dimension of the crowd 3nmanifold M, entanglement effectively
reduces the number of active degrees of freedom in (21).
Entropic Geometry of Crowd Dynamics

237
Recall that the essential concept in dynamical systems theory is the notion of a vectorfield
(that we will denote by a boldface symbol), which assigns a tangent vector to each point p in
the manifold in case. In particular, v is a gradient vectorfield if it equals the gradient of
some scalar function. A flowline of a vectorfield v is a path fl(t) satisfying the vector ODE,
fl
`
(t) = v(fl(t)), that is, v yields the velocity field of the path fl(t). The set of all flow lines of a
vectorfield v comprises its flow
t
that is (technically, see e.g., [49; 52]) a oneparameter Lie
group of diffeomorphisms (smooth bijective functions) generated by a vector-field v on M,
such that
0
= , = identity, which gives: ( ) = ( (0)).
t s t s t
t
+

Analytically, a vector-field v is defined as a set of autonomous ODEs. Its solution gives the
flow
t
, consisting of integral curves (or, flow lines) fl(t) of the vectorfield, such that all the
vectors from the vector-field are tangent to integral curves at different representative points
p M. In this way, through every representative point p M passes both a curve from the
flow and its tangent vector from the vector-field. Geometrically, vector-field is defined as a
cross-section of the tangent bundle TM of the manifold M.
In general, given an nD frame {
i
} {/x
i
} on a smooth nmanifold M (that is, a basis of
tangent vectors in a local coordinate chart x
i
= (x
1
, ..., x
n
) M), we can define any vector-field
v on M by its components v
i
= v
i
(t) as
1
1
= = = ... .
i i n
i i n
v v v v
x x x

+ +

v
Thus, a vector-field v X(M) (where X (M) is the set of all smooth vector-fields on M) is
actually a differential operator that can be used to differentiate any smooth scalar function
f = f (x
1
, ..., x
n
) on M, as a directional derivative of f in the direction of v. This is denoted simply
vf, such that
1
1
= = = ... .
i i n
i i n
f f f
f v f v v v
x x x

+ +

v
In particular, if v = ` (t) is a velocity vector-field of a space curve (t) = (x
1
(t), ..., x
n
(t)),
defined by its components v
i
=
i
x` (t), directional derivative of f (x
i
) in the direction of v
becomes
= = = = ,
i
i
i i
f df dx
f x f f
dt x dt

v
`
`
which is a rate-of-change of f along the curve (t) at a point x
i
(t).
Given two vector-fields, u = u
i

i
,v = v
i

i

X(M), their Lie bracket (or, commutator) is another
vector-field [u,v] X (M), defined by
[ , ] = = ,
j j i i
i j j i
u v v u u v uv vu
which, applied to any smooth function f on M, gives
( ) ( ) [ , ]( ) = ( ) ( ) . f f f u v u v v u
Nonlinear Dynamics

238
The Lie bracket measures the failure of mixed directional derivatives to commute. Clearly,
mixed partial derivatives do commute, [
i
,
j
] = 0, while in general it is not the case, [u,v] 0.
In addition, suppose that u generates the flow
t
and v generates the flow
s
. Then, for any
smooth function f on M, we have at any point p on M,
( )
2
[ , ]( )( ) = ( ( ( )) ( ( ( ))),
s t t s
f p f p f p
t s

u v
which means that in f (
s
(
t
(p))) we are starting at p, flowing along v a little bit, then along u
a little bit, and then evaluating f , while in f (
t
(
s
(p))) we are flowing first along u and then
v. Therefore, the Lie bracket infinitesimally measures how these flows fail to commute.
The Lie bracket satisfies the following three properties (for any three vector-fields u,v,w M
and two constants a, b thus forming a Lie algebra on the crowd manifold M):
i. [ , ] = [ , ] u v v u skew-symmetry;
ii. [ , ] = [ , ] [ , ] a b a b + + u v w u v u w bilinearity; and
iii. [ ,[ , ]] [ ,[ , ]] [ ,[ , ]] + + u v w v w u w u v Jacobi identity.
A new set of vector-fields on M can be generated by repeated Lie brackets of u, v, w M.
The Lie bracket is a standard tool in geometric nonlinear control theory (see, e.g. [49; 52]). Its
action on vector-fields can be best visualized using the popular car parking example, in
which the driver has two different vectorfield transformations at his disposal. They can
turn the steering wheel, or they can drive the car forward or backward. Here, we specify the
state of a car by four coordinates: the (x, y) coordinates of the center of the rear axle, the
direction of the car, and the angle between the front wheels and the direction of the car. l
is the constant length of the car. Therefore, the 4D configuration manifold of a car is a set
M SO(2) R
2
, coordinated by x {x, y, , }, which is slightly more complicated than the
individual crowd agents 3D configuration manifold SE(2) SO(2) R, coordinated by
x = {x, y, }. The driftless car kinematics can be defined as a vector ODE:

1 2
= ( ) ( ) , c c + x u x v x ` (26)
with two vectorfields, u,v X(M), and two scalar control inputs, c
1
and c
2
. The infinitesimal
carparking transformations will be the following vectorfields
cos
sin
tan
( ) DRIVE = cos sin ,
1
tan
0
x y l
l

+ +





u x

0
0
and ( ) STEER = .
0
1

v x
The car kinematics (26) therefore expands into a matrix ODE:
Entropic Geometry of Crowd Dynamics

239
1 2 1 2
cos
0
sin
0
= DRIVE STEER .
1
0 tan
1
0
x
y
c c c c
l

+ +






`
`
`
`

However, STEER and DRIVE do not commute (otherwise we could do all your steering at
home before driving of on a trip). Their combination is given by the Lie bracket
2
1
[ , ] [STEER, DRIVE] = WRIGGLE.
cos
l

v u
The operation [v,u] WRIGGLE [STEER,DRIVE] is the infinitesimal version of the
sequence of transformations: steer, drive, steer back, and drive back, i.e.,
1 1
{STEER, DRIVE, STEER , DRIVE }.


Now, WRIGGLE can get us out of some parking spaces, but not tight ones: we may not have
enough room to WRIGGLE out. The usual tight parking space restricts the DRIVE
transformation, but not STEER. A truly tight parking space restricts STEER as well by
putting your front wheels against the curb.
Fortunately, there is still another commutator available:
[ ,[ , ]] [DRIVE,[STEER, DRIVE]] =[[ , ], ] u v u u v u
2
1
[DRIVE, WRIGGLE] = sin cos SLIDE
cos
l x y

The operation [[u,v],u] SLIDE [DRIVE,WRIGGLE] is a displacement at right angles to
the car, and can get us out of any parking place. We just need to remember to steer, drive,
steer back, drive some more, steer, drive back, steer back, and drive back:
1 1 1 1
{STEER, DRIVE, STEER , DRIVE, STEER, DRIVE , STEER , DRIVE }.


We have to reverse steer in the middle of the parking place. This is not intuitive, and no
doubt is part of a common problem with parallel parking.
Thus, from only two controls, c
1
and c
2
, we can form the vectorfields DRIVE u,
STEER v, WRIGGLE [v,u], and SLIDE [[u,v],u], allowing us to move anywhere in the
car configuration manifold M SO(2) R
2
. All above computations are straightforward in
Mathematica
TM8
if we define the following three symbolic functions:

1. Jacobian matrix: JacMat[v_List, x_List] := Outer[D, v, x];
2. Lie bracket: LieBrc[u_List, v_List, x_List] := JacMat[v, x] . u - JacMat[u, x] . v;
3. Repeated Lie bracket: Adj[u_List, v_List, x_List, k_] :=
If[k == 0, v, LieBrc[u, Adj[u, v, x, k - 1], x]];


8
The above computations could instead be done in other available packages, such as Maple,
by suitably translating the provided example code.
Nonlinear Dynamics

240
In case of the human crowd, we have a slightly simpler, but multiplied problem, i.e.,
superposition of n individual agents motions. So, we can define the dissipative crowd
kinematics as a system of n vector ODEs:

1 2
= ( ) ( ) , where
k k k k k
c c + x u x v x ` (27)
cos
( ) DRIVE = , and
cos sin sin
0
k
k k k k k
k k
x y

+




u x
1 2
0
( ) STEER = 0 , while and are crowd controls.
1
k k k k
k
c c

v x
Thus, the crowd kinematics (27) expands into the matrix ODE:

1 2 1 2
0
cos
= DRIVE STEER 0 .
sin
0 1
k
k k k k k k k
x
y c c c c

+ +



`
`
`
(28)
A 3D simulation of random, dissipative crowd kinematics (27)(28) of 120 penguin-like
SE(2)robots, developed in C++/DirX is presented in Figure 2.


Fig. 2. Driving and steering random SE(2)dynamics of 120 penguin-like robots (with
embedded collision-detection). Compare with [2].
Entropic Geometry of Crowd Dynamics

241
The dissipative crowd kinematics (27)(28) obeys the set of n-tuple integral rules of motion
that are similar (though slightly simpler) to the above rules of the car kinematics, including
the following derived vector-fields:
WRIGGLE
k
[STEER
k
,DRIVE
k
] [v
k
,u
k
] and
SLIDE
k
[DRIVE
k
,WRIGGLE
k
] [[u
k
,v
k
],u
k
]
Thus, controlled by the two vector controls
1
k
c and
2
k
c , the crowd can form the vectorfields:
DRIVE u
k
, STEER v
k
, WRIGGLE [v
k
,u
k
], and SLIDE [[u
k
,v
k
],u
k
], allowing it to move
anywhere within its configuration manifold M given by (21). Solution of the dissipative
crowd kinematics (27)(28) defines the dissipative crowd flow,
K
t
.
Now, the general CDCK crowd behavior can be defined as a amalgamate flow (behavioral
Lagrangian flow,
L
t
, plus dissipative kinematic flow,
K
t
) on the crowd manifold M
defined by (21),
= : ( ( ), ( )),
L K
t t t
C t M t g t +
which is a one-parameter family of homeomorphic (topologically equivalent) Riemannian
manifolds
9
(M, g = g
ij
), parameterized by a time parameter t. That is, C
t
can be used for

9
Proper differentiation of vector and tensor fields on a smooth Riemannian manifold (like
the crowd 3nmanifold M) is performed using the LeviCivita covariant derivative (see, e.g.,
[49; 52]). Formally, let M be a Riemannian Nmanifold with the tangent bundle TM and a
local coordinate system
=1
{ }
i N
i
x defined in an open set U M. The covariant derivative
operator,
X
: C

(TM) C

(TM), is the unique linear map such that for any vector-fields
X,Y,Z, constant c, and scalar function f the following properties are valid:
= , ( ) = ( ) , =[ , ],
X cY X Y X X X X Y
c Y fZ Y Xf Z f Z Y X X Y
+
+ + + +
where [X,Y] is the Lie bracket of X and Y. In local coordinates, the metric g is defined for any
orthonormal basis (
i
= /x
i
) in U M by g
ij
= g(
i
,
j
) =
ij
,
k
g
ij
= 0. Then the affine Levi
Civita connection is defined on M by
( )
1
= , where = are the Christoffel symbols.
2
k k kl
j ij k ij i jl j il l ij
i
g g g g

+
Now, using the covariant derivative operator
X
we can define the Riemann curvature (3,1)
tensor Rm by
, ]
( , ) = ,
X Y Y X X Y
X Y Z Z Z Z Rm
which measures the curvature of the manifold by expressing how noncommutative
covariant differentiation is. The (3,1)components
l
ijk
R of Rm are defined in U M by
( )
, = , or = .
l l l l m l m l
i j k ijk l ijk i jk j ik jk im ik jm
R R + Rm
Also, the Riemann (4,0)tensor =
l m
ijk lm ijk
R g R is defined as the gbased inner product on M,
( )
= , , .
ijkl i j k l
R Rm
The first and second Bianchi identities for the Riemann (4,0)tensor R
ijkl
hold,
Nonlinear Dynamics

242
describing smooth deformations of the crowd manifold M over time. The manifold family
(M(t), g(t)) at time t determines the manifold family (M(t + dt), g(t + dt)) at an infinitesimal
time t + dt into the future, according to some presecribed geometric flow, like the celebrated
Ricci flow [30; 31; 32; 33] (that was an instrument for a proof of a 100year old Poincar
conjecture),
( ) = 2 ( ),
t ij ij
g t R t (29)
where R
ij
is the Ricci curvature tensor (see Appendix) of the crowd manifold M and
t
g(t) is
defined as

0
( ) ( )
( ) ( ) := .
lim
t
dt
g t dt g t d
g t g t
dt dt
+
(30)
3.4 Aggregate behavioralcompositional dynamics (AD)
To formally develop the meso-level aggregate behavioralcompositional dynamics (AD), we
start with the crowd path integral (24), which can be redefined if we Wickrotate the time
variable t to imaginary values, t = it, thereby transforming the Lorentzian path integral
in real time into the Euclidean path integral in imaginary time. Furthermore, if we rectify the
time axis back to the real line, we get the adaptive SFTpartition function as our proposed
AD model:

[ , ; , ]
AD
C
Physical Action|Mental Preparation = [ , , ]e .
A x y t t
i j
D
w x y



D (31)
The adaptive AD transition amplitude Physical Action|Mental Preparation
AD
as defined
by the SFTpartition function (31) is a general model of a Markov stochastic process. Recall
that Markov process is a random process characterized by a lack of memory, i.e., the statistical
properties of the immediate future are uniquely determined by the present, regardless of the
past (see, e.g. [23; 49]). The Ndimensional Markov process can be defined by the Ito
stochastic differential equation,
( ) = [ ( ), ] [ ( ), ] ( ),
j i i
i i ij
dx t A x t t dt B x t t dW t + (32)

= 0, = 0,
ijkl jkil kijl i jklm j kilm k ijlm
R R R R R R + + + +
while the twice contracted second Bianchi identity reads: 2
j
R
ij
=
i
R.
The (0,2) Ricci tensor Rc is the trace of the Riemann (3,1) tensor Rm,
( , ) tr( ( , ) ), so that ( , ) = ( ( , ) , ),
i i
Y Z X X Y Z X Y g X Y + Rc Rm Rc Rm
Its components R
jk
= Rc(
j
,
k
)are given in U M by the contraction
= , or = .
i i i i m i m
jk ijk jk i jk k ji mi jk mk ji
R R R +
Finally, the scalar curvature R is the trace of the Ricci tensor Rc, given in U M by: R = g
ij
R
ij
.

Entropic Geometry of Crowd Dynamics

243

0
(0) = , ( , = 1, , )
i
i
x x i j N (33)
or corresponding Ito stochastic integral equation

0 0
( ) = (0) [ ( ), ] ( ) [ ( ), ],
t t
j i i i i
i ij
x t x dsA x s s dW s B x s s + +

(34)
in which x
i
(t) is the variable of interest, the vector A
i
[x(t), t] denotes deterministic drift, the
matrix B
ij
[x(t), t] represents continuous stochastic diffusion fluctuations, and W

j
(t) is an N
variable Wiener process (i.e., generalized Brownian motion [23]) and
( ) = ( ) ( ).
j j j
dW t W t dt W t +
The two Ito equations (33)(34) are equivalent to the general ChapmanKolmogorov probability
equation (see equation (35) below). There are three well known special cases of the
Chapman Kolmogorov equation (see [23]):
1. When both B
ij
[x(t), t] and W(t) are zero, i.e., in the case of pure deterministic motion, it
reduces to the Liouville equation
{ } ( , | , ) = [ ( ), ] ( , | , ) .
t i i
i
P x t x t A x t t P x t x t
x

2. When only W(t) is zero, it reduces to the FokkerPlanck equation
{ } ( , | , ) = [ ( ), ] ( , | , )
t i i
i
P x t x t A x t t P x t x t
x

{ }
2
1
[ ( ), ] ( , | , ) .
2
ij j i
ij
B x t t P x t x t
x x

3. When both A
i
[x(t), t] and B
ij
[x(t), t) are zero, i.e., the statespace consists of integers
only, it reduces to the Master equation of discontinuous jumps
( , | , ) = ( | , ) ( , | , ) ( | , ) ( , | , ).
t
P x t x t dxW x x t P x t x t dxW x x t P x t x t



The Markov assumption can now be formulated in terms of the conditional probabilities P(x
i
,
t
i
): if the times t
i
increase from right to left, the conditional probability is determined entirely
by the knowledge of the most recent condition. Markov process is generated by a set of
conditional probabilities whose probabilitydensity P = P(x, t|x, t) evolution obeys the
general ChapmanKolmogorov integrodifferential equation
{ } { }
2
1
= [ ( ), ] [ ( ), ]
2
t i ij j i i
i ij
P A x t t P B x t t P
x x x

+



{ } ( | , ) ( | , ) dx W x x t P W x x t P +

including deterministic drift, diffusion fluctuations and discontinuous jumps (given respectively
in the first, second and third terms on the r.h.s.). This general ChapmanKolmogorov
integro-differential equation (35), with its conditional probability density evolution,
P = P(x, t|x, t), is represented by our SFTpartition function (31).
Nonlinear Dynamics

244
Furthermore, discretization of the adaptive SFTpartition function (31) gives the standard
partition function (see Appendix)

/
= e ,
j
w E T
j
j
Z

(35)
where E
j
is the motion energy eigenvalue (reflecting each possible motivational energetic
state), T is the temperaturelike environmental control parameter, and the sum runs over all
ID energy eigenstates (labelled by the index j). From (35), we can calculate the transition
entropy, as S = k
B

lnZ (see the next section).
4. Entropy, chaos and phase transitions in the crowd manifold
Recall that nonequilibrium phase transitions [25; 26; 27; 28; 29] are phenomena which bring
about qualitative physical changes at the macroscopic level in presence of the same
microscopic forces acting among the constituents of a system. In this section we extend the
CD formalism to incorporate both algorithmic and geometrical entropy as well as dynamical
chaos [50; 58; 60] between the entropygrowing phase of Mental Preparation and the
entropy conserving phase of Physical Action, together with the associated topological
phase transitions.
4.1 Algorithmic entropy
The Boltzmann and Shannon (hence also Gibbs entropy, which is Shannon entropy scaled
by k ln 2, where k is the Bolzmann constant) entropy definitions involve the notion of
ensembles. Membership of microscopic states in ensembles defines the probability density
function that underpins the entropy function; the result is that the entropy of a definite and
completely known microscopic state is precisely zero. Bolzmann entropy defines the
probabilistic model of the system by effectively discarding part of the information about the
system, while the Shannon entropy is concerned with measuring the ignorance of the
observer the amount of missing information about the system.
Zurek proposed a new physical entropy measure that can be applied to individual
microscopic system states and does not use the ensemble structure. This is based on the
notion of a fixed individually random object provided by Algorithmic Information Theory
and Kolmogorov Complexity: put simply, the randomness K(x) of a binary string x is the
length in terms of number of bits of the smallest program p on a universal computer that can
produce x.
While this is the basic idea, there are some important technical details involved with this
definition. The randomness definition uses the prefix complexity K(.) rather than the older
Kolmogorov complexity measure C(.): the prefix complexity K(x|y) of x given y is the
Kolmogorov complexity
u
C

(x|y)= min{p|x=
u
(y, p)} (with the convention that
u
C

(x|y)= if there is no such p) that is taken with respect to a reference universal partial
recursive function
u
that is a universal prefix function. Then the prefix complexity K(x) of x
is just K(x|) where is the empty string. A partial recursive prefix function : M N is a
partial recursive function such that if (p) < and (q) < then p is not a proper prefix of q:
that is, we restrict the complexity definition to a set of strings (which are descriptions of
effective procedures) such that none is a proper prefix of any other. In this way, all effective
procedure descriptions are self-delimiting: the total length of the description is given within
Entropic Geometry of Crowd Dynamics

245
the description itself. A universal prefix function
u
is a prefix function such that
n N
u
(y, n, p =
n
(y, p, where
n
is numbered n according to some Godel numbering
of the partial recursive functions; that is, a universal prefix function is a partial recursive
function that simulates any partial recursive function. Here, x,y stands for a total recusive
one-one mapping from NN into N, x
1
, x
2
, . . . , x
n
= x
1
, x
2
, . . . , x
n
,N is the set of natural
numbers, and M = {0,1}* is the set of all binary strings.
This notion of entropy circumvents the use of probability to give a concept of entropy that
can be applied to a fully specified macroscopic state: the algorithmic randomness of the state
is the length of the shortest possible effective description of it. To illustrate, suppose for the
moment that the set of microscopic states is countably infinite, with each state identified
with some natural number. It is known that the discrete version of the Gibbs entropy (and
hence of Shannons entropy) and the algorithmic entropy are asymptotically consistent
under mild assumptions. Consider a system with a countably infinite set of microscopic
states X supporting a probability density function P(.) so that P(x) is the probability that the
system is in microscopic state x X. Then the Gibbs entropy is ( ) = ( ln2) ( )log ( )
G
x X
S P k P x P x

(which is Shannons information-theoretic entropy H(P) scaled by k ln 2). Supposing that P(.)
is recursive, then ( ) = ( ln2) ( ) ( )
G
x X
S P k P x K x C

, where C

is a constant depending only on
the choice of the reference universal prefix function . Hence, as a measure of entropy, the
function K(.) manifests the same kind of behavior as Shannons and Gibbs entropy
measures.
Zureks proposal was of a new physical entropy measure that includes contributions from
both the randomness of a state and ignorance about it. Assume now that we have
determined the macroscopic parameters of the system, and encode this as a string - which
can always be converted into an equivalent binary string, which is just a natural number
under a standard encoding. It is standard to denote the binary string and its corresponding
natural number interchangeably; here let x be the encoded macroscopic parameters. Zureks
definition of algorithmic entropy of the macroscopic state is then K(x) + H
x
, where
H
x
= S
B
(x)/(k ln2), where S
B
(x) is the Bolzmann entropy of the system constrained by x and k
is Bolzmanns constant; the physical version of the algorithmic entropy is therefore defined
as S
A
(x) = (k ln2)(K(x) + H
x
). Here H
x
represents the level of ignorance about the microscopic
state, given the parameter set x; it can decrease towards zero as knowledge about the state of
the system increases, at which point the algorithmic entropy reduces to the Bolzmann entropy.
4.2 Ricci flow and Perelman entropyaction on the crowd manifold
Recall that the inertial metric crowd flow, C
t
: t (M(t), g(t)) on the crowd 3nmani-fold (21)
is a one-parameter family of homeomorphic Riemannian manifolds (M, g), evolving by the
Ricci flow (29)(30).
Now, given a smooth scalar function u : M R on the Riemannian crowd 3nmanifold M,
its Laplacian operator is locally defined as
= ,
ij
i j
u g u
where
i
is the covariant derivative (or, LeviCivita connection, see Appendix). We say that
a smooth function u : M [0,T)R, where T (0,], is a solution to the heat equation (see
Appendix, eq. (60)) on M if
Nonlinear Dynamics

246
= .
t
u u (36)
One of the most important properties satisfied by the heat equation is the maximum
principle, which says that for any smooth solution to the heat equation, whatever point-wise
bounds hold at t = 0 also hold for t > 0 [13]. This property exhibits the smoothing behavior
of the heat diffusion (36) on M.
Closely related to the heat diffusion (36) is the (the Fields medal winning) Perelman
entropyaction functional, which is on a 3nmanifold M with a Riemannian metric g
ij
and a
(temperature-like) scalar function f given by [75]

2
= ( | | )e
f
M
R f d

E (37)
where R is the scalar Riemann curvature on M, while d is the volume 3nform on M,
defined as

1 2 3
= det( ) ... .
n
ij
d g dx dx dx (38)
During the Ricci flow (29)(30) on the crowd manifold (21), that is, during the inertial metric
crowd flow, C
t
: t (M(t), g(t)), the Perelman entropy functional (37) evolves as

2
= 2 | | e .
f
t ij i j
R f d

E (39)
Now, the crowd breathers are solitonic crowd behaviors, which could be given by localized
periodic solutions of some nonlinear soliton PDEs, including the exactly solvable sine
Gordon equation and the focusing nonlinear Schrdinger equation. In particular, the time
dependent crowd inertial metric g
ij
(t), evolving by the Ricci flow g(t) given by (29)(30) on
the crowd 3nmanifold M is the Ricci crowd breather, if for some t
1
< t
2
and > 0 the metrics
g
ij
(t
1
) and g
ij
(t
2
) differ only by a diffeomorphism; the cases = 1, < 1, > 1 correspond to
steady, shrinking and expanding crowd breathers, respectively. Trivial crowd breathers, for
which the metrics g
ij
(t
1
) and g
ij
(t
2
) on M differ only by diffeomorphism and scaling for each
pair of t
1
and t
2
, are the crowd Ricci solitons. Thus, if we consider the Ricci flow (29)(30) as a
biodynamical system on the space of Riemannian metrics modulo diffeomorphism and
scaling, then crowd breathers and solitons correspond to periodic orbits and fixed points
respectively. At each time the Ricci soliton metric satisfies on M an equation of the form [75]
= 0,
ij ij i j j i
R cg b b + + +
where c is a number and b
i
is a 1form; in particular, when b
i
=
1
2

i
a for some function a on
M, we get a gradient Ricci soliton.
Define (g
ij
) = inf E (g
ij
, f ), where infimum is taken over all smooth f , satisfying
e = 1.
f
M
d

(40)
(g
ij
) is the lowest eigenvalue of the operator 4+ R. Then the entropy evolution formula
(39) implies that (g
ij
(t)) is non-decreasing in t, and moreover, if (t
1
) = (t
2
), then for t [t
1
,
t
2
] we have R
ij
+
i

j
f = 0 for f which minimizes E on M [75]. Therefore, a steady breather
on M is necessarily a steady soliton.
Entropic Geometry of Crowd Dynamics

247
If we define the conjugate heat operator on M as
= / t R

+
then we have the conjugate heat equation: = 0. u

The entropy functional (37) is nondecreasing under the coupled Riccidiffusion flow on M
[56]

2
| |
= 2 , = ,
2
t ij ij t
R u
g R u u u
u

+ (41)
where the second equation ensures
2
= 1,
M
u d

to be preserved by the Ricci flow g(t) on M.

If we define
2
= e
f
u

, then (41) is equivalent to fevolution equation on M (the nonlinear
backward heat equation),
2
= | | ,
t
f f f R +
which instead preserves (40). The coupled Riccidiffusion flow (41) is the most general
biodynamic model of the crowd reactiondiffusion processes on M. In a recent study [1] this
general model has been implemented for modelling a generic perceptionaction cycle with
applications to robot navigation in the form of a dynamical grid.
Perelmans functional E is analogous to negative thermodynamic entropy [75]. Recall (see
Appendix) that thermodynamic partition function for a generic canonical ensemble at
temperature
1
is given by
= e ( ),
E
Z d E

(42)
where (E) is a density measure, which does not depend on . From it, the average energy
is given by E=

lnZ, the entropy is S = E+lnZ, and the fluctuation is =(EE)
2

=
2

lnZ.
If we now fix a closed 3nmanifold M with a probability measure m and a metric g
ij
() that
depends on the temperature , then according to equation
= 2( ),
ij ij i j
g R f

+
the partition function (42) is given by
ln = ( ) .
2
n
Z f dm +

(43)
From (43) we get (see [75])
2 2 2
= ( | | ) , = ( ( | | ) ) ,
2
M M
n
E R f dm S R f f n dm

+ + +


4 2
2
1
= 2 | | , where = , = (4 ) e .
2
n
f
ij i j ij
M
R f g dm dm udV u

Nonlinear Dynamics

248
From the above formulas, we see that the fluctuation is nonnegative; it vanishes only on a
gradient shrinking soliton. E is nonnegative as well, whenever the flow exists for all
sufficiently small > 0. Furthermore, if the heat function u: (a) tends to a function as 0,
or (b) is a limit of a sequence of partial heat functions u
i
, such that each u
i
tends to a
function as
i
> 0, and
i
0, then the entropy S is also nonnegative. In case (a), all the
quantities E, S, tend to zero as 0, while in case (b), which may be interesting if g
ij
()
becomes singular at = 0, the entropy S may tend to a positive limit.
4.3 Chaotic inter-phase in crowd dynamics induced by its Riemannian geometry
change
Recall that CD transition map (9) is defined by the chaotic crowd phasetransition amplitude
=0 >0
[ ]
PHYS. ACTION MENTAL PREP. := [ ]e ,
S S
t t
iA x
M
CHAOS x

D
where we expect the inter-phase chaotic behavior (see [53]). To show that this chaotic
interphase is caused by the change in Riemannian geometry of the crowd 3nmanifold M,
we will first simplify the CD action functional (22) as

1
[ ] = [ ( , )] ,
2
t
fin j i
ij
t
ini
A x g x x V x x dt

` ` ` (44)
with the associated standard Hamiltonian, corresponding to the amalgamate version of (18),

2
=1
1
( , ) = ( , ),
2
N
i
i
H p x p V x x +

` (45)
where p
i
are the SE(2)momenta, canonically conjugate to the individual agents SE(2)
coordinates x
i
, (i = 1, ...,3n). Biodynamics of systems with action (44) and Hamiltonian (45)
are given by the set of geodesic equations [49; 52]

2
2
= 0,
j i k
i
jk
d x dx dx
ds ds ds
+ (46)
where
i
jk
are the Christoffel symbols of the affine LeviCivita connection of the
Riemannian CD manifold M (see Appendix). In this geometrical framework, the instability
of the trajectories is the instability of the geodesics, and it is completely determined by the
curvature properties of the CD manifold M according to the Jacobi equation of geodesic
deviation [49; 52]

2
2
= 0,
j i m
i k
jkm
D J dx dx
R J
ds ds ds
+ (47)
whose solution J, usually called Jacobi variation field, locally measures the distance between
nearby geodesics; D/ds stands for the covariant derivative along a geodesic and

i
jkm
R are
the components of the Riemann curvature tensor of the CD manifold M.
The relevant part of the Jacobi equation (47) is given by the tangent dynamics equation [12; 15]
Entropic Geometry of Crowd Dynamics

249

0 0
= 0, ( , = 1, , 3 ),
i i k
k
J R J i k n +
``
(48)
where the only non-vanishing components of the curvature tensor of the CD manifold M are

2
0 0
= / .
i i k
k
R V x x (49)
The tangent dynamics equation (48) can be used to define Lyapunov exponents in
dynamical systems given by the Riemannian action (44) and Hamiltonian (45), using the
formula [14]

2 2 2 2
1 =1 =1
= 1 /2 log( [ ( ) ( )] / [ (0) (0)]).
lim
N N
i i i i i i
t
t M J t J t M J J

+ + (50)
Lyapunov exponents measure the strength of dynamical chaos in the crowd behavior. The
sum of positive Lyapunov exponents defines the KolmogorovSinai entropy (see Appendix).
4.4 Crowd nonequilibrium phase transitions induced by manifold topology change
Now, to relate these results to topological phase transitions within the CD manifold M given
by (21), recall that any two highdimensional manifolds M
v
and M
v
have the same topology
if they can be continuously and differentiably deformed into one another, that is if they are
diffeomorphic. Thus by topology change the loss of diffeomorphicity is meant [80]. In this
respect, the socalled topological theorem [21] says that nonanalyticity is the shadow of a
more fundamental phenomenon occurring in the systems configuration manifold (in our
case the CD manifold): a topology change within the family of equipotential hypersurfaces
1 3 3 1 3
= {( , , ) | ( , , ) = },
n n n
v
M x x V x x v R
where V and x
i
are the microscopic interaction potential and coordinates respectively. This
topological approach to PTs stems from the numerical study of the dynamical counterpart of
phase transitions, and precisely from the observation of discontinuous or cuspy patterns
displayed by the largest Lyapunov exponent
1

at the transition energy [14]. Lyapunov
exponents cannot be measured in laboratory experiments, at variance with thermodynamic
observables, thus, being genuine dynamical observables they are only be estimated in
numerical simulations of the microscopic dynamics. If there are critical points of V in
configuration space, that is points
1 3
=[ , , ]
c n
x x x such that
=
( ) = 0
x x
c
V x , according to the
Morse Lemma [40], in the neighborhood of any critical point x
c

there always exists a
coordinate system x(t) = [x
1
(t), ...,x
3n
(t)] for which [14]

2 2 2 2
1 1 3
( ) = ( ) ,
c k k n
V x V x x x x x
+
+ + + (51)
where k is the index of the critical point, i.e., the number of negative eigenvalues of the
Hessian of the potential energy V. In the neighborhood of a critical point of the CDmanifold
M, equation (51) yields the simplified form of (49),
2
V/x
i
x
j
=
ij
, giving j unstable
directions that contribute to the exponential growth of the norm of the tangent vector J.
This means that the strength of dynamical chaos within the CDmanifold M, measured by
the largest Lyapunov exponent
1
given by (50), is affected by the existence of critical points
x
c
of the potential energy V(x). However, as V(x) is bounded below, it is a good Morse
Nonlinear Dynamics

250
function, with no vanishing eigenvalues of its Hessian matrix. According to Morse theory
[40], the existence of critical points of V is associated with topology changes of the
hypersurfaces {M
v
}
vR
. The topology change of the {M
v
}
vR
at some v
c
is a necessary
condition for a phase transition to take place at the corresponding energy value [21]. The
topology changes implied here are those described within the framework of Morse theory
through attachment of handles [40] to the CDmanifold M.
In our pathintegral language this means that suitable topology changes of equipotential
submanifolds of the CDmanifold M can entail thermodynamiclike phase transitions [25;
26; 27], according to the general formula:
[ ]
top ch
phase out|phase in := [ ]e .
iS
w

D
The statistical behavior of the crowd biodynamics system with the action functional (44) and
the Hamiltonian (45) is encompassed, in the canonical ensemble, by its partition function,
given by the Hamiltonian path integral [52]

3
top ch
= [ ] [ ]exp{i [ ( , )] },
'
t
i
n i
t
Z p x p x H p x d

` D D (52)
where we have used the shorthand notation
top ch
( ) ( )
[ ] [ ] .
2
dx dp
p x

D D
The path integral (52) can be calculated as the partition function [20],
3
3 3
2
( , ) ( )
3
=1 =1
( ) = e = e
n
n n
H p x V x i i
n i
i i
Z dp dx dx

3
2
0
= e ,
n
v
M
v
d
dv
V

(53)
where the last term is written using the socalled coarea formula [18], and v labels the
equipotential hypersurfaces M
v
of the CD manifold M,
1 3 3 1 3
= {( , , ) | ( , , ) = }.
n n n
v
M x x V x x v R
Equation (53) shows that the relevant statistical information is contained in the canonical
configurational partition function
( )
3
= ( )e .
V x C i
n
Z dx V x

Note that
3
C
n
Z is decomposed, in the last term of (53), into an infinite summation of
geometric integrals,
/ ,
M
v
d V

Entropic Geometry of Crowd Dynamics

251
defined on the {M
v
}
vR
. Once the microscopic interaction potential V(x) is given, the
configuration space of the system is automatically foliated into the family {M
v
}
vR
of these
equipotential hypersurfaces. Now, from standard statistical mechanical arguments we know
that, at any given value of the inverse temperature , the larger the number 3n, the closer to
v u
M M

are the microstates that significantly contribute to the averages, computed
through Z
3n
(), of thermodynamic observables. The hypersurface
u
M

is the one associated
with
( ) 1
3
= ( ) ( )e ,
V x C i
n
u Z dx V x

the average potential energy computed at a given . Thus, at any , if 3n is very large the
effective support of the canonical measure shrinks very close to a single .
v u
M M

= Hence,
the basic origin of a phase transition lies in a suitable topology change of the {M
v
}, occurring
at some v
c
[20]. This topology change induces the singular behavior of the thermodynamic
observables at a phase transition. It is conjectured that the counterpart of a phase transition
is a breaking of diffeomorphicity among the surfaces M
v
, it is appropriate to choose a
diffeomorphism invariant to probe if and how the topology of the M
v
changes as a function
of v. Fortunately, such a topological invariant exists, the Euler characteristic of the crowd
manifold M, defined by [49; 52]

3
=0
( ) = ( 1) ( ),
n
k
k
k
M b M

(54)
where the Betti numbers b
k
(M) are diffeomorphism invariants (b
k
are the dimensions of the
de Rhams cohomology groups H
k
(M;R); therefore the b
k
are integers). This homological
formula can be simplified by the use of the GaussBonnet theorem, that relates X(M) with
the total GaussKronecker curvature K
G
of the CDmanifold M given by [52; 58]
( ) = , where is given by (38).
G
M
M K d d

5. Conclusion
Our understanding of crowd dynamics is presently limited in important ways; in particular,
the lack of a geometrically predictive theory of crowd behavior restricts the ability for
authorities to intervene appropriately, or even to recognize when such intervention is
needed. This is not merely an idle theoretical investigation: given increasing population
sizes and thus increasing opportunity for the formation of large congregations of people,
death and injury due to trampling and crushing even within crowds that have not formed
under common malicious intent is a growing concern among police, military and
emergency services. This paper represents a contribution towards the understanding of
crowd behavior for the purpose of better informing decisionmakers about the dangers and
likely consequences of different intervention strategies in particular circumstances.
In this chapter, we have proposed an entropic geometrical model of crowd dynamics, with
dissipative kinematics, that operates across macro, micro and mesolevels. This
proposition is motivated by the need to explain the dynamics of crowds across these levels
simultaneously: we contend that only by doing this can we expect to adequately
Nonlinear Dynamics

252
characterize the geometrical properties of crowds with respect to regimes of behavior and
the changes of state that mark the boundaries between such regimes.
In pursuing this idea, we have set aside traditional assumptions with respect to the
separation of mind and body. Furthermore, we have attempted to transcend the long
running debate between contagion and convergence theories of crowd behavior with our
multi-layered approach: rather than representing a reduction of the whole into parts or the
emergence of the whole from the parts, our approach is build on the supposition that the
direction of logical implication can and does flow in both directions simultaneously. We
refer to this third alternative, which effectively unifies the other two, as behavioral
composition.
The most natural statistical descriptor is crowd entropy, which satisfies the extended second
thermodynamics law applicable to open systems comprised of many components.
Similarities between the configuration manifolds of individual (microlevel) and crowds
(macrolevel) motivate our claim that goaldirected movement operates under entropy
conservation, while natural crowd dynamics operates under monotonically increasing
entropy functions. Of particular interest is what happens between these distinct topological
phases: the phase transition is marked by chaotic movement.
We contend that backdrop gives us a basis on which we can build a geometrically predictive
modeltheory of crowd behavior dynamics. This contrasts with previous approaches, which
are explanatory only (explanation that is really narrative in nature). We propose an entropy
formulation of crowd dynamics as a three step process involving individual and collective
psycho-dynamics, and crucially non-equilibrium phase transitions whereby the forces
operating at the microscopic level result in geometrical change at the macroscopic level.
Here we have incorporated both geometrical and algorithmic notions of entropy as well as
chaos in studying the topological phase transition between the entropy conservation of
physical action and the entropy increase of mental preparation.
6. Appendix
6.1 Extended second law of thermodynamics
According to Boltzmanns interpretation of the second law of thermodynamics, there exists
a function of the state variables, usually chosen to be the physical entropy S of the system that
varies monotonically during the approach to the unique final state of thermodynamic
equilibrium:
0 (for any isolated system).
t
S (55)
It is usually interpreted as a tendency to increased disorder, i.e., an irreversible trend to
maximum disorder. The above interpretation of entropy and a second law is fairly obvious
for systems of weakly interacting particles, to which the arguments developed by Boltzmann
referred.
However, according to Prigogine [70], the above interpretation of entropy and a second law
is fairly obvious only for systems of weakly interacting particles, to which the arguments
developed by Boltzmann referred. On the other hand, for strongly interacting systems like
the crowd, the above interpretation does not apply in a straightforward manner since, we
know that for such systems there exists the possibility of evolving to more ordered states
through the mechanism of phase transitions.
Entropic Geometry of Crowd Dynamics

253
Let us now turn to nonisolated systems (like a human crowd), which exchange
energy/matter with the environment. The entropy variation will now be the sum of two
terms. One, entropy flux, d
e
S, is due to these exchanges; the other, entropy production, d
i
S, is
due to the phenomena going on within the system. Thus the entropy variation is
= .
i e
t
d S d S
S
dt dt
+ (56)
For an isolated system d
e
S = 0, and (56) together with (55) reduces to dS = d
i
S 0, the usual
statement of the second law. But even if the system is nonisolated, d
i
S will describe those
(irreversible) processes that would still go on even in the absence of the flux term d
e
S. We
thus require the following extended form of the second law:
0 (for any nonisolated system).
t
S (57)
As long as d
i
S is strictly positive, irreversible processes will go on continuously within the
system.
10
Thus, d
i
S > 0 is equivalent to the condition of dissipativity as time irreversibility. If,
on the other hand, d
i
S reduces to zero, the process will be reversible and will merely join
neighboring states of equilibrium through a slow variation of the flux term d
e
S.
From a computational perspective, we have a related algorithmic entropy. Suppose we have a
universal machine capable of simulating any effective procedure (i.e., a universal machine
that can compute any computable function). There are several models to choose from,
classically we would use a Universal Turing Machine but for technical reasons we are more
interested in Lambdatype Calculi or Combinatory Logics. Let us describe the system of
interest through some encoding as a combinatorial structure (classically this would be a

10
Among the most common irreversible processes contributing to d
i
S are chemical reactions,
heat conduction, diffusion, viscous dissipation, and relaxation phenomena in electrically or
magnetically polarized systems. For each of these phenomena two factors can be defined: an
appropriate internal flux, J
i
, denoting essentially its rate, and a driving force, X
i
, related to the
maintenance of the nonequilibrium constraint. A most remarkable feature is that d
i
S
becomes a bilinear form of J
i
and X
i
. The following table summarizes the fluxes and forces
associated with some commonly observed irreversible phenomena (see [48; 70])


In general, the fluxes J
k
are very complicated functions of the forces X
i
. A particularly simple
situation arises when their relation is linear, then we have the celebrated Onsager relations,
= , ( , = 1,..., )
i ik k
J L X i k n (58)
in which L
ik
denote the set of phenomenological coefficients. This is what happens near
equilibrium where they are also symmetric, L
ik
= L
ki
. Note, however, that certain states far
from equilibrium can still be characterized by a linear dependence of the form of (58) that
occurs either accidentally or because of the presence of special types of regulatory processes.
Nonlinear Dynamics

254
binary string, but again I prefer for technical reasons Normal Forms with respect to
alpha/beta/eta, weak, strong reduction, which are basically the Lambdatype Calculi and
Combinatory Logic notions roughly akin to a computational step). In other words, we
have states of our system now represented as sentences in some language. The entropy is
simply the minimum effective procedure against our computational model that generates
the description of the system state. This is a universal and absolute notion of compression of
our data the entropy is the strongest compression over all possible compression schemes,
in effect. Now here is the magic: this minimum is absolute in the sense that it does not vary
(except by a constant) with respect to our reference choice of machine.
6.2 Thermodynamic partition function
Recall that the partition function Z is a quantity that encodes the statistical properties of a
system in thermodynamic equilibrium. It is a function of temperature and other parameters,
such as the volume enclosing a gas. Other thermodynamic variables of the system, such as
the total energy, free energy, entropy, and pressure, can be expressed in terms of the
partition function or its derivatives.
A canonical ensemble is a statistical ensemble representing a probability distribution of
microscopic states of the system. Its probability distribution is characterized by the
proportion p
i
of members of the ensemble which exhibit a measurable macroscopic state i,
where the proportion of microscopic states for each macroscopic state i is given by the
Boltzmann distribution,
/( ) ( )/( )
1
= e = e ,
E kT E A kT
i i
i Z
p


where E
i
is the energy of state i. It can be shown that this is the distribution which is most
likely, if each system in the ensemble can exchange energy with a heat bath, or alternatively
with a large number of similar systems. In other words, it is the distribution which has
maximum entropy for a given average energy E
i
.
The partition function of a canonical ensemble is defined as a sum ( ) = e ,
E
j
j
Z

where = 1/(k
B
T) is the inverse temperature, where T is an ordinary temperature and k
B
is
the Boltzmanns constant. However, as the position x
i
and momentum p
i
variables of an ith
particle in a system can vary continuously, the set of microstates is actually uncountable. In
this case, some form of coarsegraining procedure must be carried out, which essentially
amounts to treating two mechanical states as the same microstate if the differences in their
position and momentum variables are small enough. The partition function then takes the
form of an integral. For instance, the partition function of a gas consisting of N molecules is
proportional to the 6Ndimensional phasespace integral,
3 3
6
( ) exp[ ( , )],
i i
i i N
Z d p d x H p x

R

where H = H(p
i
, x
i
), (i = 1, ...,N) is the classical Hamiltonian (total energy) function.
More generally, the socalled configuration integral, as used in probability theory,
information science and dynamical systems, is an abstraction of the above definition of a
partition function in statistical mechanics. It is a special case of a normalizing constant in
probability theory, for the Boltzmann distribution. The partition function occurs in many
problems of probability theory because, in situations where there is a natural symmetry, its
Entropic Geometry of Crowd Dynamics

255
associated probability measure, the Gibbs measure (see below), which generalizes the notion
of the canonical ensemble, has the Markov property.
Given a set of random variables X
i
taking on values x
i
, and purely potential Hamiltonian
function H(x
i
), (i = 1, ...,N), the partition function is defined as
( ) = exp ( ) .
i
i
x
Z H x

(59)
The function H is understood to be a real-valued function on the space of states {X
1
,X
2
}
while is a real-valued free parameter (conventionally, the inverse temperature). The sum
over the x
i

is understood to be a sum over all possible values that the random variable X
i
may take. Thus, the sum is to be replaced by an integral when the X
i
are continuous, rather
than discrete. Thus, one writes
( ) = exp ( ) ,
i i
Z dx H x


for the case of continuously-varying random variables X
i
.
The Gibbs measure of a random variable X
i
having the value x
i

is defined as the probability
density function
exp ( )
1
( = ) = exp ( ) = .
( ) exp ( )
i
i i
i
i
i
x
H x
P X x E x
Z H x

where E(x
i
) = H(x
i
) is the energy of the configuration x
i
. This probability, which is now
properly normalized so that 0 P(x
i
) 1, can be interpreted as a likelihood that a specific
configuration of values x
i
, (i = 1, 2, ...N) occurs in the system. P(x
i
) is also closely related to ,
the probability of a random partial recursive function halting.
As such, the partition function Z() can be understood to provide the Gibbs measure on the
space of states, which is the unique statistical distribution that maximizes the entropy for a
fixed expectation value of the energy,
log( ( ))
= .
Z
H

The associated entropy is given by
= ( )ln ( ) = log ( ),
i i
i
x
S P x P x H Z +

representing ignorance + randomness.
The principle of maximum entropy related to the expectation value of the energy H, is a
postulate about a universal feature of any probability assignment on a given set of
propositions (events, hypotheses, indices, etc.). Let some testable information about a
probability distribution function be given. Consider the set of all trial probability
distributions which encode this information. Then the probability distribution which
maximizes the information entropy is the true probability distribution, with respect to the
testable information prescribed.
Nonlinear Dynamics

256
Applied to the crowd dynamics, the Boltzmans theorem of equipartition of energy states that
the expectation value of the energy H is uniformly spread among all degrees-of-freedom of
the crowd (that is, across the whole crowd manifold M).
6.3 Free energy, Landaus phase transitions and Hakens synergetics
All thermodynamiclike properties of a multi-component system like a human (or robot)
crowd may be expressed in terms of its free energy potential, F = k
B
TlnZ(), and its partial
derivatives. In particular, the physical entropy S of the crowd is defined as the (negative) first
partial derivative of the free energy F with respect to the control parameter temperature T, i.e.,
S =
T
F, while the specific heat capacity C is the second derivative, C = T
T
S.
A phase of the crowd denotes a set of its states that have relatively uniform behavioral
properties. A crowd phase transition represents the its transformation from one phase to another
(see e.g., [48; 58]). In general, the crowd phase transitions are divided into two categories:
The firstorder phase transitions, or, discontinuous phase transitions, are those that involve a
latent heat C. During such a transition, a crowd either absorbs or releases a fixed (and
typically large) amount of energy. Because energy cannot be instantaneously
transferred between the system and its environment, firstorder crowd transitions are
associated with mixedphase regimes in which some parts of the crowd have completed
the transition and others have not. This forms a turbulent spatioi-temporal chaotic
interphase, difficult to study, because its dynamics can be violent and hard to control.
The secondorder phase transitions are the continuous phase transitions, in the entropy S is
continuous, without any latent heat C. They are purely entropic crowd transitions,
which are at the focus of the present study.
In Landaus theory od phase transitions (see [48; 58]), the probability density function P is
exponentially related to the free energy potential F, i.e., P e
F(T)
, if F is considered as a
function of some order parameter o. Thus, the most probable order parameter is determined
by the requirement F = min. Therefore, the most natural order parameter for the crowd
dynamics would be its entropy S.
The following table gives the analogy between various systems in thermal equilibrium and
the corresponding nonequilibrium systems analyzed in Hakens synergetics [25; 26; 27]:



In particular, in case of human biodynamics [48; 58], natural control inputs u
i
are muscular
forces and torques, F
i
, natural system outputs y
i
are joint coordinates q
i

and momenta p
i
,
while the system efficiencies e
i
represent the changes of coordinates and momenta with
changes of corresponding muscular torques for the ith active human joint, = , = .
i
q p i
i i
i i
q p
e e
F F



Entropic Geometry of Crowd Dynamics

257
6.4 Heat equation, Dirichlet action and gradient flow on a Riemannian manifold
The heat equation
= , u u ` (60)
on a compact Riemannian manifold M with static metric (
t
g = 0), where u : [0,T] M R is
a scalar field, can be interpreted as the gradient flow for the Dirichlet action

2
1
( ) := | | ,
2
g
M
E u u d

(61)
using the inner product,
1 2 1 2
, := ,
M
u u u u d

associated to the volume measure d. This
can be proved if we evolve u in time at some arbitrary rate u, an application of integration
by parts formula,
= ( )




M M
u X d u X d
(where div( ) := X X

is the divergence of the vector-field X

, which validates the Stokes

theorem, div( ) = 0),
M
X d

gives
( ) = ( ) = , ,
t
M
E u u u d u u

` ` (62)
from which we see that (60) is indeed the gradient flow for (62) with respect to the inner
product. In particular, if u solves the heat equation (60), we see that the Dirichlet energy is
decreasing in time,

2
( ) = | | .
t
M
E u u d

(63)
Thus we see that by representing the parabolic PDE (60) as a gradient flow, we
automatically gain a controlled quantity of the evolution, namely the energy functional that
is generating the gradient flow. This representation also strongly suggests that solutions of
(60) should eventually converge to stationary points of the Dirichlet energy (61), which by
(62) are harmonic functions (i.e., the functions u with u = 0). As an application of the
gradient flow interpretation, we can assert that the only periodic (or, breather) solutions
to the heat equation (60) are the harmonic functions (which must be constant if the manifold
M is compact). Indeed, if a solution u was periodic, then the monotone functional E must be
constant, which by (63) implies that u is harmonic as claimed.
6.5 Lyapunov exponents and KolmogorovSinai entropy
A branch of nonlinear dynamics has been developed with the aim of formalizing and
quantitatively characterizing the general sensitivity to initial conditions. The largest
Lyapunov exponent , together with the related KaplanYorke dimension d
KY
and the
KolmogorovSinai entropy h
KS
are the three indicators for measuring the rate of error growth
produced by a dynamical system [17; 50; 60].
The characteristic Lyapunov exponents are somehow an extension of the linear stability
analysis to the case of aperiodic motions. Roughly speaking, they measure the typical rate of
Nonlinear Dynamics

258
exponential divergence of nearby trajectories. In this sense they give information on the rate
of growth of a very small error on the initial state of a system [9; 10].
Consider an nD dynamical system given by the set of ODEs of the form
= ( ), x f x ` (64)
where x = (x
1
, . . . , x
n
) R
n
and f : R
n
R
n
. Recall that since the r.h.s of equation (64) does not
depend on t explicitly, the system is called autonomous. We assume that f is smooth enough
that the evolution is well defined for time intervals of arbitrary extension, and that the
motion occurs in a bounded region R of the system phase space M. We intend to study the
separation between two trajectories in M, x(t) and x(t), starting from two close initial
conditions, x(0) and x(0) = x(0) + x(0) in R
0
M, respectively.
As long as the difference between the trajectories, x(t) = x(t) x(t), remains infinitesimal, it
can be regarded as a vector, z(t), in the tangent space T
x
M of M. The time evolution of z(t) is
given by the linearized differential equations:
( )
( ) = ( ).
i
i j
j
x t
f
z t z t
x

`
Under rather general hypothesis, Oseledets [72] proved that for almost all initial conditions
x(0) R, there exists an orthonormal basis {e
i
} in the tangent space T
x
M such that, for large
times,
( ) = exp( ),
i i i
z t c e t (65)
where the coefficients {c
i
} depend on z(0). The exponents
1

2

d
are called
characteristic Lyapunov exponents. If the dynamical system has an ergodic invariant measure
on M, the spectrum of LEs {
i
} does not depend on the initial conditions, except for a set of
measure zero with respect to the natural invariant measure.
Equation (65) describes how an nD spherical region R = S
n

M, with radius centered in
x(0), deforms, with time, into an ellipsoid of semiaxes
i
(t) = exp(
i
t), directed along the e
i
vectors. Furthermore, for a generic small perturbation x(0), the distance between the
reference and the perturbed trajectory behaves as
( )
1 1 2
| ( )| | (0)|exp( ) 1 exp ( ) . x t x t O t +


If
1
> 0 we have a rapid (exponential) amplification of an error on the initial condition. In
such a case, the system is chaotic and, unpredictable on the long times. Indeed, if the initial
error amounts to
0
= |x(0)|, and we purpose to predict the states of the system with a
certain tolerance , then the prediction is reliable just up to a predictability time given by
1 0
1
ln .
p
T





This equation shows that T
p
is basically determined by the positive leading Lyapunov exponent,
since its dependence on
0
and is logarithmically weak. Because of its preeminent role,
1
is
often referred as the leading positive Lyapunov exponent, and denoted by .
Entropic Geometry of Crowd Dynamics

259
Therefore, Lyapunov exponents are average rates of expansion or contraction along the
principal axes. For the ith principal axis, the corresponding Lyapunov exponent is defined
as
= {(1 / )ln[ ( ) / (0)]},
lim
i i i
t
t L t L

(66)
where L
i
(t) is the radius of the ellipsoid along the ith principal axis at time t.
An initial volume V
0
of the phasespace region R
0
evolves on average as

( )
1 2 2
0
( ) = ,
t
n
V t V e
+ + +
(67)
and therefore the rate of change of V(t) is simply
2
=1
( ) = ( ).
n
i
i
V t V t

`

In the case of a 2D phase area A, evolving as
( )
1 2
0
( ) =
t
A t A e
+
, a Lyapunov dimension d
L
is
defined as
0
(ln( ( )))
= ,
lim
(ln(1 / ))
L
d N
d
d

where N() is the number of squares with sides of length required to cover A(t), and d
represents an ordinary capacity dimension,
0
ln
= .
lim
ln(1/ )



c
N
d


Lyapunov dimension can be extended to the case of nD phasespace by means of the
KaplanYorke dimension [64; 73; 89] as
1 2
1
= ,
| |
j
KY
j
d j

+
+ + +
+


where the
i
are ordered (
1
being the largest) and j is the index of the smallest nonnegative
Lyapunov exponent.
On the other hand, a state, initially determined with an error x(0), after a time enough
larger than 1/, may be found almost everywhere in the region of motion R M. In this
respect, the KolmogorovSinai (KS) entropy, h
KS
, supplies a more refined information. The
error on the initial state is due to the maximal resolution we use for observing the system.
For simplicity, let us assume the same resolution for each degree of freedom. We build a
partition of the phase space M with cells of volume
d
, so that the state of the system at t = t
0
is found in a region R
0
of volume V
0
=
d
around x(t
0
). Now we consider the trajectories
starting from V
0
at t
0
and sampled at discrete times t
j

= j (j =1, 2, 3, . . . , t). Since we are
considering motions that evolve in a bounded region R M, all the trajectories visit a finite
number of different cells, each one identified by a symbol. In this way a unique sequence of
symbols {s(0), s(1), s(2), . . . } is associated with a given trajectory x(t). In a chaotic system,
Nonlinear Dynamics

260
although each evolution x(t) is univocally determined by x(t
0
), a great number of different
symbolic sequences originates by the same initial cell, because of the divergence of nearby
trajectories. The total number of the admissible symbolic sequences, N

(, t), increases
exponentially with a rate given by the topological entropy
`
0
1
= ln ( , ).
limlim
T
t
h N t
t

However, if we consider only the number of sequences N
eff
(, t) N

(, t) which appear with

very high probability in the long time limit those that can be numerically or
experimentally detected and that are associated with the natural measure we arrive at a
more physical quantity called the KolmogorovSinai (or metric) entropy, which is the key
entropy notion in ergodic theory [17]:

0
1
= ln ( , ) .
limlim
KS eff T
t
h N t h
t

(68)
h
KS
quantifies the long time exponential rate of growth of the number of the effective coarse-
grained trajectories of a system. This suggests a link with information theory where the
Shannon entropy measures the mean asymptotic growth of the number of the typical
sequences the ensemble of which has probability almost one emitted by a source.
We may wonder what is the number of cells where, at a time t > t
0
, the points that evolved
from R
0
can be found, i.e., we wish to know how big is the coarsegrained volume V(, t),
occupied by the states evolved from the volume V
0
of the region R
0
, if the minimum volume
we can observe is V
min
=
d
. As stated above (67), we have
0
=1
( ) exp( ).
d
i
i
V t V t

However, this is true only in the limit 0. In this (unrealistic) limit, V(t) = V
0
for a
conservative system (where
=1
d
i

i
= 0) and V(t) < V
0
for a dissipative system (where
=1
d
i

i
< 0). As a consequence of limited resolution power, in the evolution of the volume
V
0
=
d

the effect of the contracting directions (associated with the negative Lyapunov
exponents) is completely lost. We can experience only the effect of the expanding directions,
associated with the positive Lyapunov exponents. As a consequence, in the typical case, the
coarse grained volume behaves as
0
0
( )
( , ) e ,
i
i
t
V t V


>

when V
0
is small enough. Since N
eff
(, t) V(, t)/V
0
, one has: h
KS
=
>0

i

i
. This argument
can be made more rigorous with a proper mathematical definition of the metric entropy. In
this case one derives the Pesin relation [17; 76]: h
KS

>0

i

i
. Because of its relation with the
Lyapunov exponents, or by the definition (68), it is clear that also h
KS
is a fine-grained and
global characterization of a dynamical system.
>0

i

The metric entropy is an invariant characteristic quantity of a dynamical system, i.e., given
two systems with invariant measures, their KSentropies exist and they are equal iff the
systems are isomorphic [7].
Entropic Geometry of Crowd Dynamics

261
Finally, the topological entropy on the manifold M equals the supremum of the
Kolmogorov-Sinai entropies,
( ) = sup{ ( ) = ( ) : ( )},
KS u
h u h u h u P M

where u : M M is a continuous map on M, and ranges over all uinvariant (Borel)
probability measures on M. Dynamical systems of positive topological entropy are often
considered topologically chaotic.
7. References
[1] Aidman, E., Ivancevic, V., Jennings, A. A Coupled ReactionDiffusion Field Model for
PerceptionAction Cycle with Applications to Robot Navigation. Int. J. Intel. Def.
Sup. Sys. 2008, 1(2), 93-115.
[2] Arizona State University. New Computer Model Predicts Crowd Behavior. ScienceDaily.
2007, May 22.
[3] Ashcraft M.H.Human Memory and Cognition (2nd ed.) Harper Collins: New York, 1994.
[4] Ashcraft, M.H. Cognition (4th ed.), Prentice Hall: New Jersey, 2005.
[5] Barendregt, H. The Lambda Calculus: Its syntax and semantics. Studies in Logic and the
Foundations of Mathematics. North Holland: Amsterdam, 1984.
[6] van Benthem, J. Reflections on epistemic logic. Logique & Analyse, 1991, 133134, 5 14.
[7] Billingsley, P. Ergodic theory and information. Wiley: New York, 1965.
[8] Blumer, H. Collective Behavior. In Principles of Sociology (A.M. Lee, ed.), Barnes &
Noble: New York, 1951, pp 67121.
[9] Boffetta, G., Lacorata, G., Vulpiani, A. (eds.) Introduction to chaos and diffusion. Chaos
in geophysical flows. Proc. ISSAOS, 2001.
[10] Boffetta, G., Cencini, M., Falcioni, M., Vulpiani, A. Predictability: a way to characterize
complexity. Phys. Rep. 2002, 356, 367474.
[11] Busemeyer, J.R., Diederich A. Survey of decision field theory. Math. Soc. Sci. 2002, 43,
345370.
[12] Caiani, L., Casetti, L., Clementi, C., Pettini, M. Geometry of Dynamics Lyapunov
Exponents and Phase Transitions. Phys. Rev. Lett. 1997, 79, 43614364.
[13] Cao, H.D., Chow, B. Recent developments on the Ricci flow. Bull. Amer. Math. Soc. 1999,
36, 5974.
[14] Casetti, L., Pettini, M., Cohen, E.G.D. Geometric Approach to Hamiltonian Dynamics
and Statistical Mechanics. Phys. Rep. 2000, 337, 237341.
[15] Casetti, L., Clementi, C., Pettini, M. Riemannian theory of Hamiltonian chaos and
Lyapunov exponents. Phys. Rev. E 1996, 54, 5969.
[16] Downarowicz, T. Entropy. Scholarpedia 2007, 2(11), 3901.
[17] Eckmann, J.P., Ruelle, D. Ergodic theory of chaos and strange attractors. Rev. Mod. Phys.
1985, 57, 617630.
[18] Federer, H. Geometric Measure Theory. Springer: New York, 1969.
[19] Forster, T., Logic, Induction and the Theory of Sets. London Math. Soc. Student Texts
56, Cambridge Univ. Press: Cambridge, 2003.
[20] Franzosi, R., Pettini, M., Spinelli, L. Topology and phase transitions: a paradigmatic
evidence. Phys. Rev. Lett. 2000, 84, 27742777.
Nonlinear Dynamics

262
[21] Franzosi, R., Pettini, M. Theorem on the origin of Phase Transitions. Phys. Rev. Lett.
2004, 92, 060601.
[22] Freeman, W.J., Vitiello, G. Nonlinear brain dynamics as macroscopic manifestation of
underlying manybody field dynamics. Phys. Life Rev. 2006, 3(2), 93118.
[23] Gardiner, C.W. Handbook of Stochastic Methods for Physics Chemistry and Natural
Sciences (2nd ed.). Springer, Berlin, 1985.
[24] Haken, H., Kelso, J.A.S., Bunz, H. A theoretical model of phase transitions in human
hand movements. Biol. Cybern. 1985, 51, 347356.
[25] Haken, H. Synergetics: An Introduction (3rd ed.). Springer: Berlin, 1983.
[26] Haken, H. Advanced Synergetics: Instability Hierarchies of SelfOrganizing Systems
and Devices (3rd ed.) Springer: Berlin, 1993.
[27] Haken, H. Principles of Brain Functioning: A Synergetic Approach to Brain Activity,
Behavior and Cognition, Springer: Berlin, 1996.
[28] Haken, H. Information and SelfOrganization: A Macroscopic Approach to Complex
Systems. Springer: Berlin, 2000.
[29] Haken, H. Brain Dynamics, Synchronization and Activity Patterns in PulseCodupled
Neural Nets with Delays and Noise, Springer: Berlin, 2002.
[30] Hamilton, R.S. Three-manifolds with positive Ricci curvature. J. Diff. Geom. 1982, 17,
255 306.
[31] Hamilton, R.S. Four-manifolds with positive curvature operator. J. Dif. Geom. 1986, 24,
153179.
[32] Hamilton, R.S. The Ricci flow on surfaces. Cont. Math. 1988, 71, 237261.
[33] Hamilton, R.S. The Harnack estimate for the Ricci flow. J. Dif. Geom. 1993, 37, 225 243.
[34] Hankin, C. An introduction to Lambda Calculi for Computer Scientists. Kings College
Pub. 2004.
[35] Hebb, D.O. The Organization of Behavior. Wiley: New York, 1949.
[36] Helbing, D., Molnar, P., Social force model for pedestrian dynamics. Phys. Rev. E 1995,
51(5), 42824286.
[37] Helbing, D., Farkas, I., Vicsek, T. Simulating dynamical features of escape panic. Nature
2000, 407, 487490.
[38] Helbing, D., Johansson, A., Mathiesen, J., Jensen, M.H., Hansen, A. Analytical approach
to continuous and intermittent bottleneck flows. Phys. Rev. Lett. 2006, 97, 168001.
[39] Helbing, D., Johansson, A., Zein Al-Abideen, H. The Dynamics of Crowd Disasters: An
Empirical Study. Phys. Rev. E 2007, 75, 046109.
[40] Hirsch, M.W. Differential Topology. Springer: New York, 1976.
[41] Hong, S.L., Newell, K.M. Entropy conservation in the control of human action. Nonl.
Dyn. Psych. Life. Sci. 2008, 12(2), 163190.
[42] Hong, S.L., Newell, K.M. Entropy compensation in human motor adaptation. Chaos
2008, 18(1), 013108.
[43] Ivancevic, V., Snoswell, M. Fuzzystochastic functor machine for general humanoid
robot dynamics. IEEE Trans. SMCB 2001, 31(3), 319330.
[44] Ivancevic, V. Symplectic Rotational Geometry in Human Biomechanics. SIAM Rev. 2004,
46(3), 455474.
[45] Ivancevic, V. Beagley, N. Brainlike functor control machine for general humanoid
biodynamics. Int. J. Math. Math. Sci. 2005, 11, 17591779.
Entropic Geometry of Crowd Dynamics

263
[46] Ivancevic, V. LieLagrangian model for realistic human bio-dynamics. Int. J. Hum. Rob.
2006, 3(2), 205218.
[47] Ivancevic, V., Ivancevic, T., HumanLike Biomechanics. Springer: Dordrecht, 2006.
[48] Ivancevic, V., Ivancevic, T. Natural Biodynamics.World Scientific: Singapore, 2006.
[49] Ivancevic, V., Ivancevic, T. Geometrical Dynamics of Complex Systems: A Unified
Modelling Approach to Physics Control Biomechanics Neurodynamics and
PsychoSocio Economical Dynamics. Springer: Dordrecht, 2006.
[50] Ivancevic, V., Ivancevic, T., HighDimensional Chaotic and Attractor Systems. Springer:
Berlin, 2007.
[51] Ivancevic, V., Ivancevic, T. Computational Mind: A Complex Dynamics Perspective.
Springer: Berlin, 2007.
[52] Ivancevic, V., Ivancevic, T., Applied Differential Geometry: A Modern Introduction.
World Scientific: Singapore, 2007.
[53] Ivancevic, V., Aidman, E., Yen, L. Extending Feynmans Formalisms for Modelling
Human Joint Action Coordination. Int. J. Biomath. 2008, (to appear).
[54] Ivancevic, V., Aidman, E. Life-space foam: A medium for motivational and cognitive
dynamics. Physica A 2007, 382, 616630.
[55] Ivancevic, V. Generalized Hamiltonian biodynamics and topology invariants of
humanoid robots. Int. J. Math. Math. Sci. 2002, 31(9), 555565.
[56] Ivancevic, V., Ivancevic, T. Ricci flow and bioreactiondiffusion systems. SIAM Rev.
2008 (submitted).
[57] Ivancevic, V., Ivancevic, T. NeuroFuzzy Associative Machinery for Comprehensive
Brain and Cognition Modelling. Springer: Berlin, 2007.
[58] Ivancevic, V., Ivancevic, T. Complex Nonlinearity: Chaos, Phase Transitions, Topology
Change and Path Integrals. Springer: 2008.
[59] Ivancevic, V., Ivancevic, T. Quantum Leap: From Dirac and Feynman Across the
Universe to Human Body and Mind.World Scientific: Singapore, 2008.
[60] Ivancevic, T., Jain, L., Pattison, J., Hariz, A. Nonlinear Dynamics and Chaos Methods in
Neurodynamics and Complex Data Analysis. Nonl. Dyn. 2008 (Springer Online
first).
[61] Izhikevich, E.M., Edelman, G.M. Large-Scale Model of Mammalian Thalamocortical
Systems. PNAS 2008, 105, 35933598.
[62] Johansson, A., Helbing, D., Z. Al-Abideen, H., Al-Bosta, S. From Crowd Dynamics to
Crowd Safety: A VideoBased Analysis. Adv. Com. Sys. 2008, 11(4), 497527.
[63] Jung, C.J. Collected Works of C.G. Jung. Princeton Univ. Press: New Jersey, 1970.
[64] Kaplan, J.L., Yorke, J.A. Numerical Solution of a Generalized Eigenvalue Problem for
Even Mapping. Peitgen, H.O.,Walther, H.O. (eds.). Functional Differential
Equations and Approximations of Fixed Points, Lecture Notes in Mathematics, 730,
Springer: Berlin, 1979, pp 228256.
[65] Kelso, JAS. Dynamic Patterns: The Self Organization of Brain and Behavior. MIT Press:
Cambridge, 1995.
[66] Kugler, P.N., Turvey, M.T. Information, Natural Law, and the SelfAssembly of
Rhythmic Movement: Theoretical and Experimental Investigations, Erlbaum:
Hillsdale, 1987.
[67] Lewin, K. Resolving Social Conflicts, and, Field Theory in Social Science. Am. Psych.
Assoc.,Washington, 1997.
Nonlinear Dynamics

264
[68] Matlin, M.W. Cognition. (7th ed.), Wiley: New York, 2008.
[69] Nara, A., Torrens, P.M. Spatial and temporal analysis of pedestrian egress behavior and
efficiency, In Association of Computing Machinery (ACM) Advances in
Geographic Information Systems, Samet, H.; Shahabi, C.; Schneider, M.(Eds.) 2007,
New York, ACM, 284-287.
[70] Nicolis, G., Prigogine, I. SelfOrganization in Nonequilibrium Systems: From
Dissipative Structures to Order through Fluctuations. Wiley: Europe, 1977.
[71] Nicolis, J.S. Dynamics of hierarchical systems: An evolutionary approach. Springer:
Berlin, 1986.
[72] Oseledets, V.I. A Multiplicative Ergodic Theorem: Characteristic Lyapunov Exponents
of Dynamical Systems. Trans. Moscow Math. Soc. 1968, 19, 197231.
[73] Ott, E., Grebogi, C., Yorke, J.A. Controlling chaos. Phys. Rev. Lett. 1990, 64, 1196 1199.
[74] Penrose, R. The Emperors New Mind. Oxford Univ. Press: Oxford, 1989.
[75] Perelman, G. The entropy formula for the Ricci flow and its geometric applications.
arXiv:math.DG/0211159, 2002.
[76] Pesin, Ya.B. Lyapunov Characteristic Exponents and Smooth Ergodic Theory. Russ.
Math. Surveys 1977, 32(4), 55114.
[77] Pessa, E., Vitiello, G. Quantum noise, entanglement and chaos in the quantum field
theory of mind/brain states. Mind and Matter 2003, 1, 5979.
[78] Pessa, E., Vitiello, G. Quantum noise induced entanglement and chaos in the dissipative
quantum model of brain. Int. J. Mod. Phys. 2004, 18B, 841858.
[79] Pessoa, L. On the relationship between emotion and cognition. Nat. Rev. Neurosci. 2008,
9, 148158.
[80] Pettini, M. Geometry and Topology in Hamiltonian Dynamics and Statistical
Mechanics. Springer, New York, 2007.
[81] Reed, S.K. Cognition: Theory and Applications. (7th ed.) Wadsworth Pub. 2006.
[82] Schner, G. Dynamical Systems Approaches to Cognition. In: Cambridge Handbook of
Computational Cognitive Modeling. Cambridge Univ. Press: Cambridge, 2007.
[83] Sutton, R.S., Barto, A.G. Reinforcement Learning: An Introduction. MIT Press:
Cambridge, MA, 1998.
[84] Todorov, E., Jordan, M.I. Optimal feedback control as a theory of motor coordination.
Nat. Neurosci. 2002, 5(11), 12261235.
[85] Tognoli, E., Lagarde, J., DeGuzman, G.C., Kelso, J.A.S. The phi complex as a
neuromarker of human social coordination. PNAS 2007, 104(19), 81908195.
[86] Turner, R.H., Killian, L.M. Collective Behavior (4th ed.) Englewood Cliffs: New Jersey,
1993.
[87] Umezawa, H. Advanced field theory: micro macro and thermal concepts. Am. Inst.
Phys.: New York, 1993.
[88] Willingham, D.T. Cognition: The Thinking Animal (3rd ed.) Prentice Hall: New York,
2006.
[89] Yorke, J.A., Alligood, K., Sauer, T. Chaos: An Introduction to Dynamical Systems.
Springer: New York, 1996.
11
Nonlinear Dynamics and Probabilistic Behavior
in Medicine: A Case Study
H. Nicolis
Unit RIMBAUD (adolescents), Service de Psychiatrie
CHU Brugman 4, place A. Van Gehuchten 1020 Bruxelles
Belgium
1. Introduction
Nonlinearity is ubiquitous in medicine and life sciences, from the molecular and cellular to
the organismic and population levels, owing to the presence of a variety of interactions,
feedbacks and other kinds of regulatory processes that ensure the harmonious coexistence of
the multitude of simultaneously ongoing activities (Mosekilde, 1996).
Nonlinearities arising from the cooperative interactions between the subunits constituting a
system in conjunction with appropriate environmental stimuli, give often rise to collective
behaviors transcending the individual subunits. A striking example of such collective
behavior is contagion, be it in the form of propagation of a disease, of a rumor or on a more
microscopic scale of a mutation, whereby a previously unaffected unit becomes affected in
its turn following an encounter with the information-carrying unit. In this chapter we will
be concerned with a particularly dramatic instance of contagion arising in the context of
adolescent psychiatry, namely, adolescent suicidal outbreaks.
Suicidal trends rank among the most serious disorders of adolescence. In most countries,
mortality from suicide is the second or the third leading cause (depending on the surveys) of
teenage deaths. The incidence of suicide attempts peaks during mid adolescence (Becker,
Schmidt, 2004). It is estimated that 20% of adolescents have suicidal thoughts and among
them as much as 5 to 8% have attempted to commit the act (Pommereau, 2001). Each of
these suicidal acts leaves behind surviving family members, friends and acquaintances who
must cope with the loss (Bridge et al, 2003).
A number of risk factors for adolescent suicidality have been identified. Among these the
most important are depression and exposure to suicide, suicide attempts or suicidal
thoughts by family and friends, suggesting that the adolescent can be considered at
potential risk of contagion with suicidality stimulations. Here, suicide contagion refers to
the link between adolescents exposure to a suicide stimulus and subsequent rise in the
frequency of suicide attempts or suicide rate and is considered most likely to occur in
already suicidal adolescent and to be a time-limited risk. In this respect, it appears
reasonable to view a suicidal trend as a behavioral attribute. If so, suicide contagion could
be regarded as a particular manifestation of behavioral contagion whereby, much like in an
infectious disease, an attitude or a mood passes from a person to the next. Jones and Jones
(1994) provided statistical support of behavioral contagion in a number of situations, and
Nonlinear Dynamics

266
the perspectives opened in their analysis constitute one of the principal motivations of the
present work.
Generally speaking, if a behavior is contagious, its prevalence increases with the number of
susceptible adolescents rather than the total number of individuals present. Wheeler (1970)
identifies behavioral contagion by 4 criteria:
1. An observer is motivated to behave in a certain way;
2. The observer knows how to perform the behavior in question but is not performing it;
3. The observer sees a model perform the behavior;
4. The observer after observing the model performs the behavior.
The theory of contagion rests on three central concepts apart from contagion itself:
susceptibility, mode of transmission and exposure. Susceptibility is necessary for contagious
transmission.
One aspect of youth suicide of particular concern is the repeated reports of suicide
outbreaks among young people. These outbreaks have been reported from as long ago as
ancient Greece and from around the world. They have been called suicide clusters, a term
that describes three or more suicides occurring within a defined space and time. The
incidence of cluster suicides is highest among teenagers and young adults (Gould, 2001) and
a growing concern has been that adolescents exposed to a peers suicide may be at increased
risk to engage in suicidal behavior (Brent et al, 1993 a, b). Many studies have also addressed
the question of whether indirect exposure to suicide through media or Internet accounts
contributes to subsequent suicide (Baume et al, 1997; Davidson et al, 1989).
The most common explanation for the above noted phenomena is that of imitation. This
mechanism is consistent with reported epidemics of suicide involving unusual methods
such as immolation etcImitation is also consistent with the short latency between publicity
and the increased rate of suicide within 1 to 2 weeks. According to McKenzie et al (2005)
there is indirect evidence that imitative suicide occurs among people with mental illnesses
and may account for about 10% of suicides by current and recent patients.
One could argue that individuals are influenced in their suicidal thoughts mainly through
their direct exposure to an actual suicidal attempt. If so, suicidal trend would be a
spontaneous process occurring at a rate equal to the size of the population of concerned
individuals multiplied by proportionality constant whose value depends on the exposure in
question. In this context, Joiner (1999) wonders if the pernicious agent of the hypothetical
contagion in suicide exists. He insists on the important role of exposure, external influence
rather than contagion and suggests that the concept of imitation may be not needed. He
emphasizes that the vulnerable people may become socially contagious via assortative
relating and thus simultaneously susceptible to the effects of life stress. Other studies report
that the predominant psychiatric sequelae observed in adolescents exposed to violent deaths
are anxiety, depression and post traumatic stress disorder. It has been suggested that the
degree to which the second person identifies with or feels similar to the deceased person
may influence the degree to which he is affected by this exposure.
While these mechanisms are undoubtedly operating in a number of circumstances of
interest, our main thesis here is that they cannot account properly for suicidal outbreaks, as
they lack the necessary ingredient of feedback. The alternative we thus propose is that of
cooperativity, when a population of susceptible individuals is mixed with a population of
suicidal ones. The nature of the suicidal attempt is in this perspective completely different,
as it now depends on the size of two subpopulations in close interaction. This double
Nonlinear Dynamics and Probabilistic Behavior in Medicine: A Case Study

267
dependence calls for a nonlinear approach to the problem and opens the way to self-
acceleration, abrupt transitions and other analogous behaviors concomitant to the well-
established syndrome of outbreaks.
In this chapter, the propagation of suicidal trends is viewed as the result of encounters in the
course of which a susceptible individual can change its mental state with some probability
following interaction with a suicidal one. The encounters can be short ranged like e.g. a
physical encounter in a hospital unit and a school class, or long ranged like e.g.
communication though the Internet. Different contagion scenarios are explored and the
main trends to be expected are identified. The results are confronted to the data available
and different strategies for improving current prevention practices are suggested. Two types
of methodologies are employed. In a first approach, the variability arising from the
individual decision making is ignored and a mean view is adopted. This maps the problem
to a problem of population growth in a medium of limited resources (here the total number
of susceptible and suicidal individuals). Various growth patterns are highlighted depending
on the contagion probability and the initial percentage of suicidal individuals. In a second
approach, variability is incorporated by means of the technique of Monte Carlo simulation
well suited to treat populations of limited size where randomness is expected to play an
important role. This approach has been used with success in several problems arising in
chemical kinetics, biochemistry and social insect behavior (Gillespie, 1992). A number of
different evolution scenarios are explored and some unexpected effects are brought out. The
novelty here is to give access to situations limited in space and time like e. g. those arising in
a given hospital unit over the usual hospitalization time, as opposed for those accounted for
in surveys where local and short scale trends are smeared out.
2. General setting
Let X
1
, X
2
be the populations of suicidal and of susceptible individuals respectively. In order
to bring out the role of nonlinearity and cooperativity in a clearcut manner, it is stipulated
that during the phenomenon of interest there is no major reshuffling of the organization,
entailing that the total population remains essentially constant:
X
1
+ X
2
= N = constant (1)
In addition to the above two types of individuals, a third type may also be present. In what
follows its role is viewed as that of a buffer, in the sense that while it does not participate
directly in the dynamics, it may play a role in determining the values of some of the
parameters present.
A first instance (hereafter referred as case I) explored in the sequel is that of contagion
arising though direct, physical encounters of type 1 and type 2 individuals, hereafter
denoted as S
1
and S
2
which are schematized as follows:

1
2
1 1 1 1
2 2 2 2
1 2 1
1 2 2
2
2
p
p
S S S S
S S S S
S S S
S S S
+ +
+ +
+
+
(2)
The first two steps correspond to the obvious idea that encounters of individuals of the same
kind do not give rise to a mental transition. On the contrary, the last two steps account for
Nonlinear Dynamics

268
imitation and thus cooperativity: upon encountering a susceptible individual, a suicidal one
can either switch to the susceptible state with a certain probability p
2
or induce a suicidal
trend to the susceptible partner with another probability p
1
. The corresponding probabilities
p
1
and p
2
are expected to fulfill the inequality p
1
>p
2
. Although there seems to be no direct
statistical evidence in support of this, we argue that in the absence of medical treatment
such a property reflects the well-established tendency of susceptible and suicidal
individuals to evolve uphill in the search of increasingly dramatic experiences rather than
to evacuate stress and evolve to the opposite way toward normality. We refer to
Pommereau (2001) for the definition of susceptible individuals. In fact adolescents with
mental dysfunction express affective immaturity, sensibility to frustrations, massive
dependence to genitors, depressivity of the mood without depressive episode and tendency
to acting out. These susceptible adolescents refer to the most deviant repairs including
suicidality. We emphasize that p
1
, p
2
are intrinsic parameters associated to individual 1-2
encounters, independent of the respective sizes of the populations X
1
and X
2
. Depending on
the latter, the overall process of contagion will of course become accentuated, as seen below.
It should also be noticed that in writing scheme (2), we tacitly assumed that individuals of
the type 1 and 2 can only exist in a single state. In a more refined analysis one could account
for further differentiation within a single subpopulation, like e.g. different degrees of
susceptibility in individuals of type 2. Other refinements would be to account for memory
effects and for changes in the parameters N, p
1
, p
2
arising for instance from medical care,
environmental stimuli or population renewal. Such extensions are likely to be important on
a long time scale. They are not carried out here, as our main purpose is to identify the role of
nonlinearity and cooperativity in the outbreak of suicidal attempts, a phenomenon expected
to be initiated in the short to intermediate time regime.
A second instance of interest (hereafter referred as case II) pertains to contagion through
long range interactions. To account for this possibility, we imagine that individuals
constitute the nodes of a network and the interactions between any two individuals give rise
to a connection between the corresponding nodes. In the previously presented case I, only
nearest neighbor nodes are connected (e.g. 1-2, 2-3 etc.). In the other extreme each node is
connected to any other node (e.g. 1-2, 1-3, 1-4, 2-3, 2-4,, etc). This corresponds to the
longest possible range that interactions can achieve. Intermediate cases may also be
envisaged. We emphasize that the model as defined above is in many respects generic. It
should thus apply suitably adapted to other types of behavioral contagion beyond the
suicidal one that constitutes the main focus of the present work.
We are now in the position to formulate the evolution of the subpopulations X
1
and X
2
in a
quantitative manner. Two complementary points of view are adopted for this purpose, as
specified below. The results to be reported depend crucially on the values of the contagion
probabilities p
1
and p
2
. These quantities or, more to the point, their difference p
1
-p
2

determine the time scale over which the suicidal trend will spread. In view of the scarcity of
relevant data, different values will be considered and the sensitivity of the results on the
choices will be assessed. Another important parameter, responsible for the sharpness of
contagion and for the importance of stochastic effects, is the total number N of the
individuals in the group and the initial numbers X
1
(0) of suicidal ones. In the following a
sensitivity analysis with respect to these parameters will be carried out and some robust
trends will be identified. The following possibilities will be considered.
Nonlinear Dynamics and Probabilistic Behavior in Medicine: A Case Study

269
1. All individuals N-X
1
(0) other than the suicidal ones are likely to be affected by the
contagion. This can be the case in a hospital unit or in an institution where non-suicidal
patients are already subjected to psychiatric disorders.
2. Among the X
2
=N-X
1
(0) individuals only a fraction X
2
(0) ( much smaller than 1) are
likely to be affected, the remaining ones being immune to any psychiatric disorders.
This can correspond to a school class or to hospital unit in which the adolescent patients
are treated for a completely different kind of disease.
3. Population dynamic approach: An averaged view
In this view, encompassing case I as well as case II above, it is assumed that individuals 1
and 2 are well mixed and interact at random. The strength of the interactions is proportional
to the corresponding fractions
1
=X
1
/N,
2
=X
2
/N, and only encounters between 1 and 2
lead to changes in the populations of either 1 or 2. This leads us to a rate law of the form

Rate of change of 1 over a time interval
=p
1
x (frequency of 1-2 encounters) - p
2
x (frequency of 1-2 encounters)

Taking the limit of the shortest time interval over which interactions become effective one
obtains the quantitative expression
d
1
/ dt = ( p
1
-p
2
)
1

2
or, with eq. (1)
d
1
/dt= p
1
(1-
1
) (3)
where we set
p=p
1
-p
2
(4)
This equation is formally identical to the logistic equation (Pielou, 1969). It can be integrated
exactly, yielding

1
(t) =

1
(0)
[1
1
(0)] e
pt
+
1
(0)
(5)
which is seen to depend solely on p and on the initial fraction
1
(0).
The two quantitatively different evolutions predicted by this equation are depicted in Fig. 1
and 2 corresponding respectively to
1
(0) being greater or smaller than 1/2. As can be seen,
in the first case one witnesses a smooth evolution toward a contagion of the entire
population, bound to occur on the time scale of
T
cont
~ 1/p (6)
In the second case one observes on the contrary a first period of quiescence during which
individuals 1 seem to have no contagion effect, followed by an explosive growth and
eventual saturation. The explosion time, corresponding to the inflexion point of the
1

versus t the curve of Fig. 2, can be evaluate explicitly and is given by


t
*
=
1
p
ln[
1
1
(0)

1
(0)
] (7)
Nonlinear Dynamics

270
For
1
(0) much smaller than unity it is therefore much longer than the contagion time
associated to the case of Fig. 1. In practice, saturation and explosion may never be achieved
if the corresponding times are longer than the hospitalization period. Nevertheless, the
above results may provide valuable indications on the trends that may be in elaboration
within the populations in interaction. They will also serve as reference for the Monte Carlo
approach presented below.

0.2
0.4
0.6
0.8
1
0 10 20 30 40
t

1
(t)

1
(0)

Fig. 1. Time evolution of the fraction of individuals of type 1 as deduced from eq. (5) under
the condition
1
(0)>1/2. Parameter values p=0.15,
1
(0)=0.55.

0.2
0.4
0.6
0.8
1
0 20 40 60 80

1
(t)
t
*
t

1
(0)

Fig. 2. As in Fig. 1 but with
1
(0)=0.01.
Nonlinear Dynamics and Probabilistic Behavior in Medicine: A Case Study

271
4. Monte Carlo simulation
When dealing with complex realities one is often led to recognize that a modeling approach
may be limited by the lack of detailed knowledge of the laws governing the system at hand
and of the values of the parameters involved in the description. A central point of the
present work is that to cope with this limitation it is important to set up a complementary
approach aiming at a direct simulation of the underlying process, rather that at the solution
of the evolution laws suggested by a certain model. The Monte Carlo simulation approach
described below provides an efficient way to achieve this goal. It also allows one to
incorporate in a natural way the role of individual variability expected to be of the utmost
importance, since the quantities featured are now fluctuating in both space and time rather
than being fully deterministic. Two types of studies have been conducted. In both cases, the
population sizes have deliberately been taken to be small to emulate real world situations as
they arise in a single hospital unit or in a school class. As it will turn out stochastic effects
will then play a very important role. Still, the averaged description serves as a useful
reference for apprehending the specific role of stochasticity in the overall process.
Case I
The physical space (school class, recreation area, hospital unit, space of common patient
activities, ...) is modeled as a regular square planar lattice. Each individual performs a
random walk between an initial position and its first neighbors. When two individuals are
led to occupy through this process the same lattice site processes (2) are locally switched on.
The various steps are weighted by the corresponding probabilities and the particular
transition to be performed at a given time is decided by a random number generator
(amounting essentially to throwing dice) compatible with these probabilities. After this
particular step is performed the populations X
1
, X
2
are updated and the process is restarted.
The simulation, which records the numbers of X
1
and X
2
at different parts of space, is
stopped at a number of steps beyond which the process becomes stationary in the sense of
reducing to fluctuations around a constant (time-independent) plateau. In addition to a
single realization of the simulation (referred as stochastic trajectory) averages over
realization giving access to mean values, variances etc are also performed.
The following instances are considered.
i. An institution or a big hospital unit with N=30, X
1
(0)=6 suicidal individuals and
X
2
(0)=24 individuals presenting other kinds of psychiatric disorders. The contagion
probabilities are set p
1
=0.25, p
2
=0.1 and the individuals are initially taken to be
distributed randomly.
ii. As before, but with N=20, X
1
(0)=4 in order to test the role of population size.
iii. A school class or a mixed hospital unit with N=30, X
1
(0)=2 suicidal individuals. It is
supposed that of the N-X
1
(0)=28 individuals 4 are susceptible of being affected and the
remaining 24 ones constitute the environment within which the process will take place.
Accordingly, the contagion probabilities are set to lower values p
1
=0.1, p
2
=0.05 since the
encounters are expected to be more scarce.
iv. N=8 individuals of which X
1
(0)=4 are suicidal and N-X
1
(0)=4 subject to other types of
disorder, functioning as a clan independent of its environment. This is accounted for
by resetting p
1
, p
2
to the values of 0.25 and 0.1 respectively.
v. As in iv. but now the two subpopulations are initially segregated (say in different
hospital rooms) and meet only in common activities.
Nonlinear Dynamics

272
Figures 3a,b depict the time dependence of the population density X
1
/N of X
1
averaged over
many realizations of the process and of the associated variance <X
1
2
>=<X
1
2
> - <X
1
>
2
.
Figure 4 provides a reformulation of the results of Fig. 3 when all cases (i) to (v) are
normalized to the same mean population. Figs 5 and 6a,b provide a more refined view of the
role of inherent variability by showing respectively a single stochastic trajectory under the
conditions of case (iii) and the probability histograms associated with (i) and (iii).
Case II
The physical space (e.g. Internet, a newsletter etc) is here lumped into a single cell within
which each individual may interact with any number of other ones with probabilities
determined as before. Again, stochastic trajectories recording the individual transitions as
well as averaged quantities over all trajectories are deduced. The context is now that of a
small number of heavily affected individuals communicating via Internet, newsletter or any
other kind of multimedia means with a small number of susceptible partners not attained so
far by the disease. Fig. 7 summarizes the results for N=6, X
1
(0)=3 using the same values for
parameters p
1
and p
2
as before.
5. Discussion
Building on evidence supporting the existence of suicidal contagion, we proposed and
developed a predictive model of how suicidal trends propagate in an adolescent population.
The principal feature underlying the model is the cooperative character of the contagion
process (last two steps in (2)). The model predictions depend entirely on two kinetic
parameters, the contagion probabilities p
1
and p
2
for susceptible and for suicidal individuals
to switch to the suicidal and susceptible state respectively; and on two population like
parameters, the total number N of individuals that may undergo a transition in their mental
state and the number X
1
(0) of suicidal individuals initially present.
A first result of interest has been that contagion is not always a smooth process but may
rather take an explosive form, depending on the values of X
1
(0)/N and p=p
1
-p
2
. In this latter
case there exists a well-defined time t

of switching toward a collective suicidal state (Figs 2,

3a and 4a). This provides a quantitative basis for the phenomenon of outbreak referred in
the Introduction as well as a strong support of the idea of contagion as a generic mechanism
of adolescent suicidal trends. Subsequently, the population attains a mean saturation level
on which is superimposed a random signal reflecting individual variability. This level may
actually never be attained since on a long time scale the refinements to the original model
discussed in section 2 will begin to play an increasingly crucial role.
A second series of results pertains to the role of stochasticity. The following comments are in
order on inspecting the key Figure 3.
- In all cases the mean value <X
1
> is increasing in time, in qualitative agreement with
Figs 1 and 2.
- The evolution is initially slower for segregated sub-populations (case (v)). What is
happening here is that few among 1 and 2 types first meet in a limited space which
constitutes a front of some sort, from which the trend can subsequently propagate.
- In cases (i), (ii), (iv) and (v), a saturation level in which the entire population of
susceptible individuals switches to the suicidal state is eventually reached. The time
scale for this to happen may be long with respect to the hospitalization or school period
times. Still, the explosive growth for short times should be emphasized, confirming the
prediction made in eq. (7) and Fig 2.
Nonlinear Dynamics and Probabilistic Behavior in Medicine: A Case Study

273
- The saturation level reached in case (iii) is significantly less than 100% in the same time
scale as (i), (ii), (iv) and (v). This at first sight unexpected emergence of a state of
undecidability is robust with respect to changes in the values of p
1
and p
2
. It arises
primarily from individual variability, here exacerbated by the smallness of the size of
the overall population compared to X
1
(0). There are long periods of hesitation and in
some realizations of the process the trend is inverted and the entire population reaches
the more favorable state.
0.2
0.4
0.6
0.8
1
0 200 400 600 800 1000 t
<X
1
>/N
a

0
0.05
0.1
0.15
0 200 400 600 800 1000 t
b
<X
1
2
>/N
2

Fig. 3. (a): Time dependence of the mean density of individuals of type 1 as deduced from
the Monte Carlo simulation; the full, dashed, heavy dotted, dashed-dotted and light dotted
lines refer to cases (i) to (v), respectively. (b) : Time dependence of the variance under the
conditions of Fig 3a. The physical space considered is a square planar lattice of size 10X10
space units, the number of statistical realizations is 10,000 and the initial positions of the
populations are random in space.
Nonlinear Dynamics

274
These trends are further illustrated in Fig. 3b where the variance<X
1
2
>=<X
1
2
> - <X
1
>
2
is
represented. In all case but (iii) <X
1
2
> is seen to reach a low final value, but prior to this it
goes though a well - marked maximum grossly at a time corresponding to the inflexion
point of the curves in Fig. 3a. As for case (iii), <X
1
2
> steadily increases and reaches a final
value orders of magnitude larger than for (i), (ii), (iv) and (v) which is comparable to the
mean value itself. This is in agreement with and provides an explanation of the statement in
Jones and Jones on the behavior of variance.

0.2
0.4
0.6
0.8
1
0 100 200 300 400 500
<X
1
>/N
t
a


0.05
0.1
0.15
0.2
0 100 200 300 400 500
<X
1
2
>/N
2
t
b

Fig. 4. (a): As in Fig. 3 but under conditions of identical overall population densities. Full,
dashed and dotted lines refer to cases (i), (ii) and (iii), respectively. Initial positions and
number of realizations as in Fig. 3.
Nonlinear Dynamics and Probabilistic Behavior in Medicine: A Case Study

275
Interestingly, when all cases above are normalized to the same mean population density,
cases (i), (ii), (iv), and (v) are essentially reduced to a universal" behavior both for the mean
and the variance while case (iii) still constitutes a different class (Fig. 4a, 4b). This suggests
that the model of eq. (3) is rather adequate for intermediate to long times as long as N is
sufficiently large (which in practice could be reached already for the rather modest value of
N=8), but even in these cases it may prove inadequate for short times and especially for
times around the maximum of the variance.

0
10
20
30
0 50 100 150 200 250 300 t
X
1
X
2
a


0
1
2
3
4
5
6
0 500 1000 1500 2000 t
X
1
X
2
b

Fig. 5. (a): Quasi-deterministic behavior modulated by small scale variability under the
conditions of case (i). (b): Situation of undecidability induced by the individual variability in
a small size population (case (iii)).
Nonlinear Dynamics

276
At the level of a single stochastic realization of the process (the analog of the type of
evolution observed in practice) variability and undecidability are reflected by the fact that
while in case (i) the switching of the population to state 1 occurs quite early in time (Fig.5a),
it needs a much longer induction time under the conditions of case (iii) (Fig. 5b). We next
comment on Figs 6a,b which depict the probability histograms associated with (i) and (iii)
respectively. In 6a, drawn after 80 time units (the time at which the variance reaches its
maximum in Fig. 3b) the histogram is clearly unimodal. It is peaked at a value corresponding

0.5
1.0
1.5
0 0.15 0.3 0.45 0.6 0.75
X
1
/N
P
t=80
a


2.0
4.0
0 0.15 0.3 0.45 0.6 0.75
P
X
1
/N
t=300
b

Fig. 6. Probability histograms associated with cases (i), Fig. 6a and (iii), Fig. 6b with an initial
population density 0.3. Initial positions as in Fig. 3 and number of realizations is 20,000.
Nonlinear Dynamics and Probabilistic Behavior in Medicine: A Case Study

277
to the instantaneous X
1
/N as deduced from Fig. 3a. For longer times the maximum slides to
the right and eventually tends to 1. The structure is radically different for Fig. 6b drawn
after 300 time units (the time for the value of the variance to exceed that of cases (i), (ii), (iv)
and (v)) which displays a bimodal structure. As can be seen, the two peaks are located at
low (close to 0) and high (close to 1) density of X
1
, reflecting the possibility of switching
from individuals of type 1 to type 2 with a non-negligible probability. Clearly, this type of
structure is quite different from the binomial distribution usually featured when
interpreting results of surveys (Jones & Jones, 1994). This reflects the cooperative character
of the contagion dynamics, an idea that has been central throughout this chapter.


0.2
0.4
0.6
0.8
1
0 5 10 15 20
<X
2
>/N
<X
1
>/N
t
a

0.025
0.05
0 5 10 15 20
<X
1
2
>/N
2
t
b

Fig. 7. Time dependence of the mean density of individuals of type 1 and 2 (7a) and of the
variance of individuals of type 1 (7b) in the presence of long range interactions. Number of
realizations as in Fig. 3.
Nonlinear Dynamics

278
The results discussed so far pertain to Case I. Regarding now the new features concerning
Case II, summarized in Fig. 7, their most striking difference with Figs 3 and 4 is that the
process is now accelerated dramatically, such that saturation level is reached within an
observable time scale. Owing the small numbers involved this level is less than 100% in a
way analogous to case (iii) above. The variance remains substantial at saturation (Fig. 7b)
and goes through a maximum.
6. An augmented model
The results in the preceding sections depend crucially on the validity of the conservation
condition of the total population of suicidal and susceptible individuals (eq. (1)). Although
this may be a reasonable assumption over short to intermediate time scales it is bound to fail
in the long run, as the system becomes open to different kinds of interactions with its
environment. In this section we develop an augmented version of the model of eqs (2)
accounting for key processes expected to be present in real-world situations. Specifically, we
allow for the following additional steps.
- The influx of susceptible individuals S
2
from an external population A of size much
larger than S
2
:


A
a
S
2
(8a)
- The possibility that suicidal individuals may be removed from the population S
1

(recovery or on the contrary isolation):


S
1

k
1
S
1
*
(8b)
- The possibility that susceptible individuals may likewise be removed from the space of
coexistence with S
1
, spontaneously or deliberately:


S
2

k
2
S
2
*
(8c)
The rate equations associated to this augmented model read


d
1
dt
= p
1

2
k
1

1
d
2
dt
= a p
1

2
k
2

2
(9)
Choosing as before p>0, we notice that in the limit a=0, k
1
=k
2
=0 the total population
1
+
2

is conserved and one recovers for
1
the logistic equation (3). Here we are interested in the
new effects arising (a), from the opening of the susceptible population towards the influx a
of freshly arriving individuals; and (b), from the process by which both suicidal and
susceptible individuals tend to leave the system though the above mentioned mechanisms
of medical treatment, recovery or spatial constraints.
Nonlinear Dynamics and Probabilistic Behavior in Medicine: A Case Study

279
Contrary to eq. (3), eqs (9) do not admit an explicit analytic solution. We therefore proceed
by identifying first the stationary states in which the variables
1
and
2
no longer evolve in
time. Setting d
1
/dt= d
2
/dt=0 in eqs. (9), one finds:
- A semi trivial solution

1
= 0,
2
=
a
k
2
(10a)
- A fully non-trivial solution

1
=
1
k
1
(a
k
1
k
2
p
),
2
=
k
1
p
(10b)

To determine the conditions under which the system will eventually settle in (10a) or (10b)
we perturb slightly each of these states and seek for conditions on the parameters under
which the perturbations are amplified or on the contrary damped. In the first case the state -
which will be qualified as unstable- will not be sustainable under real-world conditions,
where perturbations of different origins are inevitable. In the second case the state which
will be qualified as stable- will represent the asymptotic regime towards which the system
will evolve after a transient period whose duration depends on the values of the parameters.
A standard linear stability analysis (Nicolis, 1995)) leads to the conclusion that there is a
well-defined transition separating these two situations, occurring at a value of the influx
parameter a given by


a
c
=
k
1
k
2
p
(11)

For a<a
c
state (10a) is the unique, stable steady state solution of eqs. (9) since state (10b) is
physically unacceptable (
1
<0). For a> a
c
state (10a) still exists but is unstable, and the
system evolves spontaneously towards state (10b) which becomes physically admissible as

1
is now positive. Notice that in the limit a=0, k
1
=k
2
=0, p>0 the semi-trivial state is always
unstable and the non-trivial one is always stable. This corresponds, in fact, to the situation
depicted in Figs 1 and 2 pertaining to the model of eq. (3).
Figures 8a,b summarize the time evolution of the fractions of
1
and
2
prior to the steady
state, under the condition a>a
c
(state (10b) is stable). We start with a sizable pool of
susceptible individuals in which a small fraction of suicidal ones has been introduced. The
evolution of
1
follows first a course quite similar to that of Fig 2, but once near the plateau
the situation changes radically: owing to the increasing effect of suicidal contagion, the pool
of susceptibles tends to be depleted and this in turn induces a sharp decrease of suicidal
incidents. The result is the appearance of a marked overshoot in the population of
1
and a
concomitant undershoot in
2
. Subsequently both
1
and
2
experience a slight undershoot
and overshoot respectively, before settling to their long terms values. We have here a second
manifestation of suicidal outbreak beyond the one identified for the model of eq. (3), where
outbreak was associated with the occurrence of an inflexion point of the function
1
(t) prior
to the attainment of the plateau (eq. (7)).
Nonlinear Dynamics

280

0.2
0.4
0.6
0.8
1
0 50 100

1
/
(

1
+

2
)
t
a



0.2
0.4
0.6
0.8
1
0 50 100
t

2
/
(

1
+

2
)
b


Fig. 8. Transient evolutions of the fractions of
1
(a) and
2
(b) obtained by solving
numerically eqs. (9). Parameter values a=1, k
1
=0.12, k
2
=0.01, k=0.02 and initial conditions
equal to 0.001 and 0.999, respectively.
Following the logic of the Monte Carlo analysis previously carried out for the scheme of eqs
(2), we now inquire on the effect of variability in the results derived so far in this section.
Rather than perform a full scale Monte Carlo study, we resort to a more phenomenological
approach in which variability is accounted for by adding to the right hand sides of both eqs
(9) uncorrelated random noises sampled from a Gaussian distribution. Fig 9 depicts the
Nonlinear Dynamics and Probabilistic Behavior in Medicine: A Case Study

281
response of
1
to a variability source of this kind. Keeping parameters values as in Fig. 8 we
see that variability tends to depress the extent of suicidal outbreak, presumably by
desynchronizing the action of the suicidal individuals that would otherwise have
manifested itself in a concerted fashion.



0.2
0.4
0.6
0.8
1
0 50 100 t

1
/
(

1
+

2
)



Fig. 9. Effect of variability in the form of uncorrelated Gaussian noise sources of variance
equal to 10
-2
added to eqs (9), on the evolution of the fraction of
1
. Parameter values and
initial conditions as in Fig. 8.
7. Conclusions and perspectives
We believe that the ideas put forward in this work have a methodological interest that may
be further enhanced by e.g. refining the model to account for several internal states or for
memory effects. In addition to this fundamental aspect we suggest that our results as they
stand can be the starting point for two kinds of applications. Firstly, the reassessment of
some of the results available from surveys. In particular the bimodal character of the
probability in Fig.6b, reflecting the cooperativity and the smallness of the population size,
suggests that the process does not always follow the trend of a purely random event as
reflected by a binomial probability distribution. Secondly, the elaboration of prevention
strategies. In particular, one may use the switching time t* (eq. (7)) and inflexion point in
Figs 2 and 8a) as alert level beyond which the process may get out of control. It may happen
as mentioned in Sec. 3 that under the conditions actually prevailing in a given environment
this time is much too long compared to the time scale imposed by the local conditions. If so
one should switch to a second indicator of an imminent catastrophic evolution, which in our
view is provided by the standard deviation (<X
1
2
>)
1/2
or more significantly the ratio
Nonlinear Dynamics

282
(<X
1
2
>)
1/2
/<X
1
>. As seen in Sec. 5 this quantity, easily monitored, tends to be enhanced in
the vicinity of a collective transition encompassing the populations of interest.
For all the situations analyzed in Sec. 4 with the exception of (iii), the propagation of suicide
is explosive and inevitable. The evolution of the propagation of suicide in case (v) is slower
because of the limited cooperativity between individuals who have few contacts between
them. It would be worthwhile to analyze in the future from this perspective contagion
trends in other behavioral disorders typical of adolescence such as running away and
addictions.
Another potential application pertains to prevention of situation (iii) in connection with the
nature of the class group. There is much discussion about the possibility to create classes
with mixed difficult adolescents, that is teenagers with conduct and affective disorders
inhibiting the faculty to learn and to succeed in school. In fact the adolescents suffering of
conduct disorder have often difficulties in mentalization of their essential depressive
symptoms. Even if they do not have the problem of suicidal symptoms in first place, they
commit repeatedly a lot of accidents, such as motor vehicle fatalities or even delinquent acts,
equivalent to suicidal act. Regrouping this kind of adolescents may be, in our view and
according to our results, an error as it will tend to induce further accidents. We see actually
that the mixing of susceptible individuals in a healthy class group limits the risk of
suicidal contagion.
Finally, there is according to our results an interactive Werther effect in the form of cyber
suicide. In 1774 Johann Wolfgang von Goethe published his by now famous novel Die
Leiden des jungen Werther, in which his hero a young artist, takes his own life after a
series of failed attempts to gain the love of beautiful Lotte. The novel had an immediate and
an immense impact: men of society used to dress like Werther and as many as 2000 readers
seem to have imitated the way he acted and died. As a result of this catastrophic situation,
Goethes novel was banned for a long time in many European countries. More than 200
years later, it appears that the availability of easy communication channels through the mass
media and in particular through the advent of the Internet, an increasingly important mode
of information and communication among adolescents and young adults is at the origin of a
comeback of an interactive Werther effect. Many studies have addressed the question of
an observer copying suicidal behavior that he has seen modeled in the media. Case reports
about cyber-suicide have been published, whereby indirect exposure to suicide through
media or Internet accounts contributes to subsequent suicide.
Suicide information is easily accessible over the web, as are special chat rooms for
discussions with like-minded people. Chat rooms are typical of adolescents and young
adults, a group at the highest risk for imitative suicidal behavior (Davidson et al, 1989). In
fact mass clusters are media related phenomena. They are regrouped more in time than in
space, and are purportedly in response of actual or fictional suicide.
Our results (case II, cf. Fig. 7) provide insights on the mechanisms underlying this collective
behavior. They also suggest certain ways of control of the phenomenon and of its follow
ups. Health group sites and qualified treatment for suicidal youths should be better
promoted. Psychiatrists, parents and teachers should take more interest in their
patients/childrens Internet consumption and discuss with them. Question on media and
Internet should be part of the anamnesis. The legal options to prevent cyber suicide should
be discussed from a national and international perspective because of the dramatic
Nonlinear Dynamics and Probabilistic Behavior in Medicine: A Case Study

283
contagion and the criminal abuse of the Internet communities (Becker, Schmidt, 2004). This
is crucial especially in view of our results on the dramatically fast pace of the process.
In summary, the major clinical insights afforded by our models are: the elaboration of
guidelines for slowing down the propagation of suicide; the identification of possible alert
indicators; and controlling Internet consumption. The main limitations of the models in their
present form are that memory effects are not incorporated and that an individual is taken to
be in either of only two mental states.
All in all we believe that in addition to and as complement of the all-important insights
afforded by the statistical analysis of surveys, a first principles approach of the kind
suggested in this chapter may contribute to the unveiling of some of the multiple facets of
the dramatic episodes surrounding adolescent suicidal trends.
8. References
Baume, P. Cantor, C.H., Rolfe, A. (1997). Cyber suicide: the role of interactive suicide , notes
on the Internet. Crisis, 18: 2, 73-79.
Becker, K., Schmidt, M.H. (2004). Internet chat rooms and suicide. Journal of the American
Academy of Child and Adolescent Psychiatry, 43: 3, 246.
Brent, D.A., Perper, J., Moritz, G., Allman, C., Liotus, L., Schweers, J., Roth, C., Balach, L.,
Cannobbio, R. (1993). Bereavement or depression? The impact of the loss of a friend
to suicide. Journal of the American Academy of Child and Adolescent Psychiatry, 32: 6,
1189-1197.
Brent, D.A., Perper, J., Moritz, G., Allman, C., Schweers, J., Roth, C., Balach, L., Cannobbio,
R., Liotus, L. (1993). Psychiatric sequelae to the loss of an adolescent peer to suicide.
Journal of the American Academy of Child and Adolescent Psychiatry, 32: 3, 509-517.
Bridge, J.A., Day, N.L., Day, R., Richardson, G. A., Birmaher, B., Brent, D.A. (2003). Major
depressive disorder in adolescents exposed to a friends suicide. Journal of the
American Academy of Child and Adolescent Psychiatry, 42: 11, 1294-1300.
Davidson, L.E., Rosenberg, M.L., Mercy, J.A., Franklin J., Simmons, J.T. (1989). An
epidemiologic study of risk factor in two teenage suicide clusters. JAMA, 17: 262,
2687-2692.
Gillespie, D. T. (1992). Markov Processes. New York: Academic Press.
Gould, M.S. (2001). Suicide and the media. Annals of the New York Academy of Sciences, 932,
200-224.
Joiner T.E., Jr. (1999). The clustering and contagion of suicide. Current directions in
psychological science, 8,3, 89-92.
Jones, M.B. and Jones, D.R. (1994). Testing for behavioural contagion in a case-control
design. Journal of psychiatric research, 28, 35-55.
McKenzie, N., Landau, S., Kapur, N., Meehan, J., Robinson, J., Bickley, H., Parsons, R.,
Appleby, L. (2005). Clustering of suicides among people with mental illness. The
British Journal of Psychiatry, 187: 476-480.
Mosekilde, E. (1996). Topics in Nonlinear Dynamics. Singapore: World Scientific.
Nicolis G. (1995). Introduction to Nonlinear Science. Cambridge: Cambridge University
Press.
Pielou, E. K. (1969). An Introduction to Mathematical Ecology. NewYork: Wiley-Interscience.
Nonlinear Dynamics

284
Pommereau, X. (2001). LAdolescent Suicidaire. Paris: Dunod.
Stolley, P.D., Tamar Lasky (1995). Investigating Disease Patterns : The Science of Epidemiology,
Freeman, New York.
Wheeler L. (1970). Interpersonal Influence. Boston : Allyn&Bacon.
12
The Effect of Spatially Inhomogeneous
Electromagnetic Field and Local Inductive
Hyperthermia on Nonlinear Dynamics of the
Growth for Transplanted Animal Tumors
Valerii Orel and Andriy Romanov
Medical Physics and Bioengineering Laboratory
National Cancer Institute
Ukraine
1. Introduction
Cancer is often characterized as a chaotic, poorly regulated growth. Cancer can be viewed as
a complex adaptive system. Complex adaptive systems can be described mathematically by
nonlinear (chaos) theory including asymmetry, fractal structure and autocorrelation factor
(Cramer, 1993). Atypical shape of tumor cells and chaotic structures of blood flow is one
from characteristic of cancer process. Atypical change of cell shape in conglomerates of
tumor cells and structure of blood flow is accompanied by increase of deterministic chaos
(Baish & Jain, 2000; Orel & Dzyatkovskaya, 2000). Complex natural phenomena such as
cancer are dynamical systems whose state changes by perturbation. The concept of
deterministic chaos is hierarchical for host in contemporary ideas about role of chaos for
potential application in oncology (Sedivy & Mader, 1997; Blazsek, 1992). The authors
introduced concepts related to chaos

theory, such as attractors, fractals and the Lotka-
Volterra equations,

as potentially useful approaches to allow for the analysis of

carcinogenic
biological processes as related to selection and

competition. In certain situations, these
equations give chaotic, non-linear,

and nonpredictable results. Given what is known about
the enormous

complexity of the carcinogenic process, use of models such as

these may be
perfectly justified, and might provide the theoretical

framework that is so desperately
needed in this age of data

overload to make real progress in the understanding of human

carcinogenesis (Garte, 2003).
Entropy is a measure of disorder. The thermodynamic entropy of a cancerous cell is
different from that of a normal cell due to the more disordered structure of the cancerous
cell. The reversal of entropy flow in tumour tissues may halt tumour development due to
reversed signal transmission in the tumour-host entity. This thermodynamic approach may
help in the design of cancer therapy (Molnar et al., 2009).
Transplanted animal tumors which can only be experimentally induced by transplanting
living tumor cells significant influence on complex adaptive systems include developing of
tumor formation for experimental animals. During recent years there has been increasing
public concern on potential cancer risks from radiofrequency radiation emissions (Hardell &
Nonlinear Dynamics

286
Sage, 2008). Inhomogeneous pulsing electromagnetic fields (EF) stimulation of biological
tissue was associated with the increase in the number of cells and/or with the enhancement
of the cellular differentiation (Diniz et al., 2002). Inhomogeneous (asymmetric) and
sinusoidal EF can cause different changes in protein synthesis of cells. It should be noted,
that pulsed asymmetric EF and heat shock produced different patterns of polypeptide
synthesis (Goodman & Henderson, 1988). Inhomogeneous pulsing EF caused significant
reductions in osteoclast formation of tumor necrosis factors, interleukins (Chang et al., 2004)
and in osteoblast-like cell of proliferation and gene expression (De Mattei et al., 2005). These
observations provide evidence that in vitro inhomogeneous EF affects the mechanisms
involved in cell proliferation and differentiation.
Magnetic resonance images demonstrate that malignant tumor can be inhomogeneous
media for spatially inhomogeneous EF (Fig. 1). Cancer patient exhibited higher values
within the spread parameter S range than healthy individual (Fig.2). Each wavefront will be
continued independently by an arbitrary inhomogeneous structure of tumor. Propagation of
inhomogeneous radio waves in tumor is accompanied by nonlinear effects with greater
changes in direction and energy of electromagnetic field than in normal tissues (Kattapuram
et al., 1999).


a b
Fig. 1. T1-weighted MR images of the stomach: a - healthy individual; b - cancer


a b
Fig. 2. Phase map of T1-weighted MR images of the stomach: a - healthy individual;
b - cancer
The Effect of Spatially Inhomogeneous Electromagnetic Field and Local Inductive Hyperthermia
on Nonlinear Dynamics of the Growth for Transplanted Animal Tumors

287
The complete wave field at a tumor will be then obtained as an integral superposition of all
wavefront arriving in some neighbourhood of the object. Inhomogeneous electromagnetic
wave can be written from Maxwell's equations in the form of an inhomogeneous
electromagnetic wave equation (or often "nonhomogeneous electromagnetic wave
equation") (Purcell, 1985). Relationships between transplanted animal tumors and external
inhomogeneous EF that initiated in them local hyperthermia are important for understand
of the principles nonlinear dynamics in cancer process and multimodal approach (and
typically nonlinearly) for him treatment (Furusawa & Kaneko, 2000).
Doxorubicin (DOXO) is an anthracycline quinone antineoplastic antibiotic that has been
shown to have a wide spectrum of clinical activity against a variety of solid tumors. The
mechanisms of DOXO-induced cytotoxicity have been extensively studied and have been
shown to include free radical formation and absorption of DOXO into the double helix of
DNA resulting in topoisomerase II-mediated DNA damage . DOXO also causes depolarization
of the membrane lipid bilayer in different cancer cell lines (Reszka et al., 2001).
Current forms of DOXO are higly toxic to the patient and can cause systematic comlications,
most notably cardiotoxicity. Systemic toxicity can seriously decrease the effectness of the
drug since a lower dose must be administrated to avoid toxicity. Another approach to avoid
toxicity include targeted delivery, however, it is often difficult to ensure that the
chemotherapy targets only the cancer tissue and the agent is localising in the target tissue.
Therefore in several studies DOXO was combined with electromagnetic hyperthermia with
an aim at enhancing antitumor efficacy of the drug (Shen et al., 2008). However, the
cytotoxicity of this antitumor agent is increased by elevated temperatures as shown in vitro
and in vivo (Marmor, 1979; Chen et al., 2004). Nonetheless, studies of DOXO and
electromagnetic hyperthermia are still controversial and often show no synergism or
synergism only at the doses that cannot be tolerated by subjects (Gaber, 2002). Positive
clinical results of combined treatment with DOXO and electromagnetic hyperthermia are
still unsatisfactory. Widespread clinical application of electromagnetic hyperthermia in the
patients is limited because temperatures in the range of 4150C produce heat shock
proteins and initiate drug resistance (thermoresistant) in tumor cells (Roemer, 1999).
Drug resistance is the single most important cause of cancer treatment failure and carries a
massive burden to patients, healthcare providers, drug developers and society. It is
estimated that multi drug resistance plays a major role in up to 50% of cancer cases. Today,
most drug therapies involve multiple agents, as it is almost universally the case that single
drugs (or single-target drugs) will encounter resistance. Drug resistance presents some of
the greatest challenges to the treatment and eradication of cancer. There are many studies
and reports on drug resistance in cancer cells. P-glycoprotein, the expression product of the
MDR-1 gene, is strongly associated with both de novo and acquired resistant. The protein
function as a transmembrane drug efflux pump, transporting cytostatic agents. Glutation
and it is dependent enzymes may be involved in resistance to drug by proving cellular
protection against free radicals damage. Resistance to drug occurs when the damaged DNA
undergoes excision repair. It is likely that many mechanisms of DOXO resistance exist and
that such mechanisms are cell specific. Thus, problems related to the development
multidrug resistance have led researchers to investigate alternative forms of administrating
DOXO for treatment of cancer.
One of complex approach may be in use of inhomogeneous pulsing EF for treatment of drug
resistance tumor (Miyagi et al., 2000). Pulsing EF used for stimulation of antiresistant
nonthermal effect in mouse osteosarcoma cell line (Hirata et al., 2001). It is known that
Nonlinear Dynamics

288
exposure to the pulsing EF causes depolarization of cell membranes and modifies drug
resistance of tumor cells (Pasquinelli et al., 1993; Ruiz-Gmez et al., 2002).
One of the branches in electromagnetic hyperthermia known as inductive hyperthermia (IH)
is based on the use of magnetic and electric components of EF in the radiofrequency

spectrum for the localization and the concentration of heat during anticancer neoadjuvant
therapy or activation

of susceptor material implanted in the tumor. The equivalent power
density (power density of the plane wave having the same field intensity) for magnetic field
is greater than that for the electric field by a factor of ten (Martino, 1962; Moseley, 1988).
During IH of tumor the process of irradiating realize by near-field. In the near-field the
maxima and minima of electric and magnetic fields do not occur at the same points along
the direction of propagation as they do in the case of the far-field. In this region, the
electromagnetic field structure may be highly spatially inhomogeneous and typically, there
may be substantial variations from the plane wave impedance i.e., in some regions, almost
pure electric fields may exist and, in other regions, almost pure magnetic fields (Jordan &
Balmain, 1968). The magnetic component of EF causes heating in

tumor

tissues through
induced eddy currents.

Incorporation of antitumor agents into the tumor cells is increased
by eddy current stimulation, which is induced by pulsing magnetic fields. Therefore, the cell
cycle shifts from the non-proliferative to proliferative phase that leads to increased
antitumor activity of the drug (Ivkov et al., 2005; Jin et al., 1998; Orel et al.,2005).
It is well known that EF can influence the chemical reactions to raise their activation
energies above threshold levels of thermal noise (Weaver et al.,1999). Nonthermal effects can
reduce existing disadvantages on all of the classical thermal treatment (Blank & Soo, 2001;
Longo & Ricci, 2007).
In paper (Boddie et al., 1987), it was suggested to produce an inhomogeneous EF pattern
with eddy current orthogonal to the magnetic force lines during regionally-focused
hyperthermia of a tumor. Really, it is possible to suppose that increased inhomogeneity of
EF will activate a non-equilibrium thermodynamical process in a tumor and increase
antitumor activity of DOXO. Separately nonthermal and hyperthermal effects (4150C) of
amplitude-frequency modulation for initiation EF inhomogeneity during treatment of
animal tumor is generally used. However, the influence of spatial inhomogeneity of EF and
local IH in the range physiological hyperthermia (3740C) on nonlinear dynamics of animal
tumor growth hasn't been well enough studied yet.
This paper examines the effects of spatially inhomogeneous EF, local IH in the range
physiological hyperthermia on nonlinear dynamics of the growth for transplanted animal
tumors and entropic action during treatment by DOXO of DOXO-resistant Guerin's
carcinoma.
2. Materials and methods
2.1 Experimental animals
In the study, 180 male rats weighing 170 20 g bred in the vivarium of National Cancer
Institute and 20 C57BL/6 male mice weighing 19 1 g bred in the vivarium of Bohomolets
Institute for Physiology Research, NAS of Ukraine (Kyiv, Ukraine) were used.
2.2 Tumor transplantation
The transplantation of Guerin carcinoma, Lewis lung carcinoma, sarcoma 45, Walker 256
carcinosarcoma and Pliss lymphosarcoma were performed according to the established
The Effect of Spatially Inhomogeneous Electromagnetic Field and Local Inductive Hyperthermia
on Nonlinear Dynamics of the Growth for Transplanted Animal Tumors

289
procedure. All animal procedures were carried out according to the rules of the regional
ethic committee. Animals were housed in 2 groups: group 1 control (no treatment); group
2 irradiation by elliptic applicator with straight profile (ASP) (40 MHz).
DOXO-resistant Guerin's carcinoma was acquired according to (Solyanik et al., 1999). Thirty
sequential subcutaneous transplantations of Guerin carcinoma cells (310
6
per animal)
received from DOXO-treated rats. The transplantation of DOXO-resistant Guerin's
carcinoma was performed subcutaneously by standard method into the right hind leg.
Animals were housed in four groups: 1 control (no treatment); 2 DOXO-administration;
3 DOXO-administration + electromagnetic irradiation (EI) by ASP; 4 DOXO-
administration + EI by elliptic applicator with the circular arc in profile (AAP). Each group
contained ten animals.
2.3 Electromagnetic irradiation
First prototype of the device for medical treatment called Magnetotherm (Radmir,
Ukraine) was used (Nikolov et al., 2008). The frequency of EI was 40 MHz with an initial
power of 100 W. The animal tumors irradiated locally (Fig. 3) by inductive coaxial
applicators that had differed by the geometry and spatial inhomogeneity of EF.


Fig. 3. Electromagnetic irradiation of animal tumors
ASP was an ellipse on a horizontal plane with the semi-axes 1.52.5 cm and straight profile
(Fig. 4a). AAP profile was an arc of the circle with the radius 2.3 cm (Fig. 4b) (Ares et al.,
1996).


a b
Fig. 4. Appearance of inductive applicator: a ASP; b AAP
EF distribution was computed according to (Mittra, 1973) (Fig. 5). Spatial inhomogeneity of
EF was estimated by asymmetry parameter of electric a
E
and magnetic a
H
field strength
distribution according to (Korn & Korn, 1968). Animal tumor was positioned in the center of
applicator loop at the distance 0.3 cm from tumor surface. Specific adsorption rates (SAR) of
EI were calculated according to (Mittra, 1973). Similar design was used in helical field
stellarator for the plasma to increase entropy of EF (Weller et al., 2001).
Nonlinear Dynamics

290

a b

c d
Fig. 5. The isolines of the electromagnetic field: a ASP, electrical component with
a
E
= 0.03 a.u.; b ASP, magnetic component with a
H
= 0.16 a.u., SAR = 8.8 W/kg;
c AAP, electrical component with a
E
= 0.89 a.u.; d AAP, magnetic component with a
H

= 0.48 a.u., SAR = 1.6 W/kg. Distance to the plane of applicator was 0.5 cm; the values on
isolines indicated the tension of the electrical field in V/m and the magnetic field in A/m;
the distance in cm is indicated on the axis of abscissas and ordinates
The change of thermal pattern on surface of phantom from fatty tissue of the pig after
irradiated by EF shown in Fig. 6. The structure of heat formation on the surface of phantoms
depends on the degree of electrmagnetic field nonuniformity and it is similar to computed.
2.4 Treatment of animals with doxorubicin-resistant Guerin's carcinoma
Experimental animals were treated by DOXO (Pharmacia & Upjohn) in the dose 1.5 mg/kg.
The treatment was performed five times by DOXO and EI from 10

to 18 days after tumor
transplantation every other two days. Tumor volume before treatment was 0.43 0.05 cm
3
.
The Effect of Spatially Inhomogeneous Electromagnetic Field and Local Inductive Hyperthermia
on Nonlinear Dynamics of the Growth for Transplanted Animal Tumors

291

a b
Fig. 6. The change of thermal pattern on surface of phantom from fatty tissue of the pig after
irradiated by: a ASP; b AAP
2.5 Temperature studies
The temperature was measured in the tumor centre of DOXO-resistant Guerins carcinoma
by the fiber-optic thermometer -4 (Radmir, Ukraine). The kinetics of typical temperature
changes for animal tumor under EI is represented in Fig. 7.


Fig. 7. The temperature changes in the centre of DOXO-resistant Guerins carcinoma during
EI by ASP (a) and AAP (b)
The temperature was reached up to 39.1 after 15 min and 40 after 30 min EI by ASP, as for
AAP that was 37.9 and 38.4, accordingly. The time between two measurements was 4 hours.
It is necessary to notice, that tumor temperature was slightly increased after EI by ASP in
comparison with AAP. The kinetics of temperature growth in the tumor was quasilinear. The
fluctuations of experimental values evaluated by standard error of temperature in linear
regression model. The standard error was 0.15 for ASP and 0.1 for AAP.
24
31
24
36
Nonlinear Dynamics

292
Preliminary research showed that 15 and 30 minutes of local EI on conventional Guerin
carcinoma initiated practically identical strengthening of DOXO antineoplastic activity.
Therefore, with aim of milder hyperthermic non-equilibrium effects at physiological
temperatures the irradiation was being performed during 15 minutes at once after treatment
by DOXO.
The animals were immobilized on the special panel to indicate the heat generation pattern of
EF. The thermography was conducted by remote thermograph (B.E. Loshkarev Institute of
semiconductors of NAS of Ukraine). The inhomogeneity structure of digital thermograms
was estimated by the Shannon entropy (S) equation meant for a statistical measure of the
disorder (non-equilibrium of thermodynamical process) of a system (Korn & Korn, 1968).
2.6 The analysis of nonlinear kinetics of tumor volume
Nonlinear kinetics of tumor volume was evaluated by growth factor according to
autocatalytic equation:
( ) ( ) = +
0
1
dx
x x x
dt
, (1)
where

=

0
0
x is relative tumor growth by time t;

=

0
0
0
x is relative tumor
volume at the moment of time t = 0;
0
and

is initial and limiting tumor volume
accordingly; is tumor volume at the moment of time t (Emanuel, 1977).
The solution of equation (1) is

= +

+

0
0
0 0
0
0
1
1
t
t
e
e
. (2)
The effect of EF and local IH on nonlinear dynamics of the growth of animal tumors was
evaluated with the braking ratio:

c
EI
, (3)
where
c
is growth factor for control group of animals,
EI
is growth factor for group
after EI.
2.7 The heterogeneity of tumor structure in ultrasound image
Ultrasonic studies were done before and right after EI by ultrasonic apparatus ATL HDI
3000 (Fillips, USA) with the use of 6 MHz transducer. During ultrasonic studies the
transducer was stationary fixed relative to animal tumor.
The heterogeneity of ultrasound image G in tumor tissues for studies of tumor vessels was
evaluated with spatial autocorrelation statistics r by Moran (Bailey & Gatrell, 1995; Orel et
al., 2007a):
G = 1 r, (4)
The Effect of Spatially Inhomogeneous Electromagnetic Field and Local Inductive Hyperthermia
on Nonlinear Dynamics of the Growth for Transplanted Animal Tumors

293

( ) ( )
( )
= =

= =
=

1 1
2
1 1
1
n n
ij i j
i j
i j
n n n
i ij
i j
i
i j
n w x x x x
r
x x w
(5)
where n is the number of pixels in selected region of interest in ultrasound image, x
i
is the
intensity of i
th
pixel, x is the mean intensity of whole region of interest, and w
i
is a distance-
based weight which is the inverse distance between pixels i and j (1/d
ij
).
2.8 Statistical and correlation analysis
Statistical processing of numerical results was carried out using Statistica 6.0 ( StatSoft, Inc.
19842001) computer program with parametric Students t-test. Correlation analysis was
performed with the MATLAB 7.0 (19842004 The MathWorks, Inc.) software.
3. Results
3.1 Changes in nonlinear dynamics of the growth for animal tumors under the
influence of spatially inhomogeneous electromagnetic field and local inductive
hyperthermia
As it is shown in table 1 the growth kinetics of animal tumors had very different nonlinear
responses under the influence of spatially inhomogeneous electromagnetic fields (a
E
=
0.03 a.u.; a
H
= 0.16 a.u.) and local IH initiated by ASP. The strongest inhibition effect under
the influence of EI was in Pliss lymphosarcoma and sarcoma 45. The growth stimulation of
animal tumors after EI was recorded in Walker 256 carcinosarcoma. Animal tumors for
Lewis lung carcinoma grew nonsignificantly but average number of metastases on a mouse
in the lungs was increased on 86%. Nonlinear dynamics of tumors growth was much
differed for each single animal in all investigated groups.
EI of Gueren carcinoma by AAP with inhomogeneous electromagnetic fields (a
E
= 0.89 a.u.;
a
H
= 0.48 a.u.) statistically not significant changed nonlinear dynamics of malignant growth
in comparison with control group of animal without treatment.

Parameters
Tumor

c
, day
-1

EI
, day
-1

Guerin carcinoma 0.45 0.01 0.46 0.05 0.99
Lewis lung carcinoma 0.39 0.02 0.36 0.01 1.07
Sarcoma 45 0.60 0.03 0.45 0.01
*
1.31
Walker 256 carcinosarcoma 0.60 0.01 0.66 0.01
*
0.91
Pliss lymphosarcoma 0.42 0.02 0.32 0.01
*
1.32
* Statistically significant difference from control group
Table 1. The growth kinetics of animal tumors
The ultrasonic studies were used for interpretation of peculiarities in tumor blood flow
during EI. Guerin carcinoma only was researched because there were problems in
Nonlinear Dynamics

294
visualization of ultrasound images on the monitor for other experimental tumors. Fig. 8
shows the sonogram of Guerin carcinoma on the 10
th
day after tumor transplantation before
and after EI. The sonograms show that tumor heterogeneity parameter G for Guerin
carcinoma was higher in 2.9 times after EI than without irradiation. This is in accordance
with well known medical observations that EI and mild hyperthermia in tumor is
characterized by intensive tumor blood flow (Song et al., 2005).


a b
Fig. 8. The sonogram of Guerin carcinoma and tumor heterogeneity parameter G:
a without EI (G = 0.24); b after 15 min EI (G = 0.69)
According to the presented data, one may suppose that recorded effects of inhibition or
stimulation growth for animal tumors after electromagnetic stimulation may be caused by
peculiarity of vascular damages in different experimental tumors.
3.2 The effect of spatially inhomogeneous electromagnetic field, local inductive
hyperthermia and doxorubicin on nonlinear dynamics of tumor growth for animals
with doxorubicin-resistant Guerin's carcinoma
As it is shown in Fig. 9, nonlinear dynamics of the growth for tumor volumes on 10 and 12
th

day after tumor transplantation was identical. Since 14
th
day after transplantation tumor
volumes for animals from 4 groups were statistically significant decreased in comparison with
the animals of 1, 2 and 3 groups on 88%, 79% and 82% ( < 0.05) accordingly in average. The
growth kinetics of animal tumors is shown in table 2. The growth kinetics for 3 group had
minimal response under the influence of DOXO and EI by ASP generated EF with a
E
=
0.03 a.u.; a
H
= 0.16 a.u. At the same time the complete resorption were observed on 20
th
day
after tumor transplantation for 40% animals from 4 group (DOXO + EI by AAP, a
E
= 0.89 a.u.
and a
H
= 0.48 a.u.). The recurrent tumor growth hadn't been detected for 4 months after the
treatment. Obtained results were testified by the study repeated in 4 months.
Our research showed that antitumor effect of DOXO was not depended on the rotation of
applicator on horizontal plane relative to tumor. Antitumor effect of DOXO didn't changed
significantly under EF after mechanochemical activation of drug before treatment.
The Effect of Spatially Inhomogeneous Electromagnetic Field and Local Inductive Hyperthermia
on Nonlinear Dynamics of the Growth for Transplanted Animal Tumors

295

Fig. 9. EI and DOXO-induced changes in nonlinear dynamics of the growth for DOXO-
resistant Guerin's carcinoma: 1 without DOXO and EI (control); 2 DOXO; 3 DOXO + EI
by ASP; 4 DOXO + EI by AAP
Parameters
N Treatment
, day
-1

1 Without DOXO and EI (control) 0.46 0.01
2 DOXO 0.42 0.01 1.08
3 DOXO + EI by ASP 0.47 0.02 0.97
4 DOXO + EI by AAP 0.32 0.02* 1.43
* Statistically significant difference from control group
Table 2. The growth kinetics of animal tumors
3.3 Thermography
Thermal patterns of tumors surface and the panel after EI are presented in Fig. 10. Maximal
inhomogeneity of tumor surface and indicative panel that estimated by entropy was

a b c
Fig. 10. Change of thermal pattern on tumor surface after transplantation on 15 day (1) and
indicative panel (2) after EI; without EI (control); b EI by ASP; c EI by AAP
2
1
1
2
1
2
Nonlinear Dynamics

296
obtained for AAP with increased spatial inhomogeneity of EF (Fig. 11). It testifies, that the
use of EF with increased spatial inhomogeneity influenced on nonuniform temperature

distribution on the surface of animal tumor.
a b
50
75
100
125
150
D
i
f
f
e
r
e
n
c
e

t
o

t
h
e

c
o
n
t
r
o
l
,

%




Fig. 11. The inhomogeneity (entropy) of thermal pattern on tumor surface after
transplantation on 15 day (a) and indicative panel (b) after EI:

by ASP;

by AAP. On
an axis there is a difference to the control (without EI)
3.4 Ultrasonic studies
Typical tumor sonograms on the 15
th
day after the tumor transplantation and 15 minutes of
EI are shown in Fig. 12. The computer nonlinear analysis of composite B-mode and steered
color Doppler acoustic image demonstrated that heterogeneity G was decreased by 30%
after EI with increased spatial inhomogeneity by AAP. It testifies, that the use of EF with
increased spatial inhomogeneity influenced on the vessel dilation in

malignant tissues. This
is in accordance with aforementioned observations that EI and moderate hyperthermia in a
tumor is characterized by the typical change of a tumors blood flow and increased
oxygenation of tumor cells (Song et al., 2005).
4. Discussion
4.1 The influence of spatially inhomogeneous electromagnetic field and inductive
hyperthermia on nonlinear aspects of malignant growth
Our study demonstrated that spatially inhomogeneous electromagnetic fields with
asymmetry parameters a
E
= 0.03 a.u. and a
H
= 0.16 a.u. and local IH in the range
physiological hyperthermia cause influence on nonlinear dynamic of the growth of
transplanted animal tumor (Orel et al., 2007b). The cancer processes are an example of non-
equilibrium, non-linear process. It is predictable locally in the very short-term, but not in the
medium- and long-term, as typical of systems exhibiting deterministic chaos (Rubin, 1984).
The effects of spatially inhomogeneous EF and local IH in the range physiological
hyperthermia warrant increased to create chaos for animal with cancer process. It effects of
inducing extremely large and very rapid surges of stochastic endogenous signals in tumor
The Effect of Spatially Inhomogeneous Electromagnetic Field and Local Inductive Hyperthermia
on Nonlinear Dynamics of the Growth for Transplanted Animal Tumors

297

a b

c d
Fig. 12. The change of heterogeneity (G) in composite B-mode and steered color Doppler
acoustic image of tumor: without EI (control), G = 0.55; b EI by ASP, G = 0.56;
c without EI (control), G = 0.60; d EI by applicator with AAP, G = 0.42
cells. They tend to be quasi (almost but not quite)-periodic, the periodicities are a complex of
many periods, and they can swing between different quasi-periodic states. But they are not
at all random (Waliszewski et al., 1998; Marino et al., 2000,2009).
Living systems are organized such that they manifest operational features ascribed to
hierarchical and heterarchical structures from quantum to organism levels (Dirks, 2008). In
mainstream biology that would enable us to understand how EF below the "thermal
threshold" could have any effects. That, despite the fact that consistent changes in gene
expression and DNA breakages considered to the most solid evidence have now been
obtained. Some biological effects are indeed associated with EF so weak that the energies in
those fields are below the energy of random thermal fluctuations. Molecular signaling in
Nonlinear Dynamics

298
eukaryotic cells is accomplished by complex and redundant pathways converging on key
molecules that are allosterically controlled by a limited number of signaling proteins. p53-
signaling pathway is an example of a complicated sequence of signals produced in response
to DNA damage. This pattern of signaling may arise from chance occurrences at the origin
of life and the necessities imposed on a nanomolar system (Yarosh, 2001; Schneider et al.,
2004). Signals from tumor cells look like stochastic processes although their latent
mechanism is deterministic. These are the butterfly effects: the molecule of DNA could
affect the metabolism in organism (in common with a proverbial butterfly flapping its wings
in the Amazon rainforest could affect the weather in London) (Carrubba et al., 2007;
Carrubba et al., 2008).
Thereby inhomogeneous EF influence on genetic instability gives rise to the diversity of
cancer process. Evidently above mentioned can incarnate of foundation for interpretation
different in nonlinear dynamics for transplanted animal tumors.
According to the presented data, one may suppose that recorded effects of inhibition or
stimulation growth for animal tumors after spatially inhomogeneous electromagnetic
stimulation may be caused by peculiarity of vascular damages in different experimental
tumors. These results are important for clinical application of medical technologies because
they testify against the use of electromagnetic hyperthermia as a basis for the monotherapy
of malignant human tumors and the necessity to facilitate local EI during anticancer
neoadjuvant therapy with the use of drugs or magnetic nanoparticles. In general, the
application of local electromagnetic hyperthermia in clinical oncology is effective when
combined with chemotherapy or radiochemotherapy as shown in (Falk & Issels, 2001).
4.2 An increase of doxorubicin antitumor effect by entopictic action of spatially
inhomogenous electromagnetic and heat fields
The spatially inhomogeneous field is definitely changed by the geometric and
mass/structure variance of the tumor itself. The effect of spatially inhomogeneous EF
during EI on transformation of radio waves and thermal descriptions in malignant tumors
was investigated. It is shown that structure of heat formation in the range physiological
hyperthermia on tumor surface depends on the degree of inhomogeneity of EF. In our next
experiments revealed entropic action in antitumor effect for DOXO of inhomogenous
electric (a
E
= 0.89 a.u.), magnetic fields (a
H
= 0.48 a.u.) and temperature in the range
physiological hyperthermia during EI.
This action we visualized for other antitumor drug too. The highest antitumor and
antimetastatic activity was caused by the combined action of cisplatin and irradiation by
spatially inhomogeneous EF and local I of animals with resistant to cisplatin substrain of
Lewis lung carcinoma too (Orel et al., 2009).
The heterogeneous structure of blood vessels in malignant tissue specified by greater
specific area of interaction with antitumor drug in comparison with normal tissue. Chaotic
signals of inhomogeneous EF can be applied to increase creativity of artificial intelligence, in
fluid dynamics of blood to induce turbulence to increase therapeutic effects for antitumor
drug, in biochemical processes to drive reactions toward otherwise improbable biochemical
compounds, or to raise bond energies above threshold levels without destructive heat. It can
be applied to the breaking up of separative attitudes among metastasized cancer cells and
aiding in the recovery from cancer (Orel et al., 2004).
What is physicochemical property of spatially inhomogeneous electric, magnetic and
temperature fields which influenced on nonlinear dynamics of biological process in the
tumor and initiated action as increased antitumor effect for DOXO?
The Effect of Spatially Inhomogeneous Electromagnetic Field and Local Inductive Hyperthermia
on Nonlinear Dynamics of the Growth for Transplanted Animal Tumors

299
The heterogeneity for tumor structure usually is more variable than for normal tissues.
Therefore, we studied influence of EF on transformation of electric, magnetic and thermal
fields in heterogeneous (rubber foam + 0.9% NaCl solution) and homogeneous (0.9% NaCl
solution) phantoms.
Preliminary research showed that transformation of EF and thermal patterns in phantoms
was investigated during EI by spatially inhomogeneous EF (Orel et al., 2008). The change of
electric (E) and magnetic (H) component under the influence of phantoms was calculated
as follows:
E =
0,
, (6)
H =
0
, (7)
where and H

is electric and magnetic field intensity under phantom,
0
and
0
is electric
and magnetic field intensity in the air, respectively.
It is shown in Fig. 13 that the structure of heat formation on the surface of phantoms
depends on the degree of EF nonuniformity and it is similar to computed in Fig. 5 EF
distribution. Relative increase of magnetic field strength H/H
0
in phantoms after EI by
AAP was in 3.5 times greater than by ASP on the average (Table 3). Relative increase of
temperature T/T
0
in phantoms was smaller in 5.4 times after EI by AAP compared to ASP
on the average. In rubber foam phantom the ratio T/T
0
increased in 8.6 times after EI by
AAP compared to 0.9% NaCl solution phantom. It testifies stronger transformation of
spatially inhomogeneous EF for heterogenous structure of rubber foam phantom than for
homogeneous structure of 0.9% NaCl solution phantom. The transformation of
inhomogeneous EF to thermal patterns for phantoms was similarly to an effect for animal
tumors (see chapter 3.3).


Fig. 13. The change of thermal pattern on phantom surface after electromagnetic irradiation
by ASP of foam rubber + 0.9% NaCl solution (a), AAP of foam rubber + 0.9% NaCl solution
(b), ASP of 0.9% NaCl solution (), AAP of 0.9% NaCl solution (d)
a
25C
29C
b
21C
30C
c
25C
35C
d
21C
29C
Nonlinear Dynamics

300
Phantom Applicator /
0
, % H/H
0
, % T/T
0
, %
NaCl 0.9% solution ASP 47 3 8.0 1.0 0.20 0.02
NaCl 0.9% solution AAP 19 3
*
20.0 3.1
*
0.10 0.01
Foam rubber ASP 49 6 7.0 0.5 6.2 1.0
Foam rubber AAP 28 4
*
31.0 3.5
*
0.7 0.2
*

*
p < 0.05 compared to similar parameter of ASP
Table 3. The ratios /
0
, H/H
0
and T/T
0
for phantoms
We studied the transformation of EF and thermal patterns in physiological phantoms
muscular, fatty, liver tissues and packed red blood cells too. The result was similarly to
physical phantoms.
Analyzing the above-mentioned phantom researchs, it is possible to mark the problem in
our discussion. Is an increase of antitumor effect for drug during treatment under the action
of spatially inhomogeneous EF and nonuniform temperature field with temperature peak
37.9C accompanied by the tendency of biological system to move toward randomness or
disorder that increased thermodynamical entropy in the tumor? As contrasted with our
experiments in classic electromagnetic hyperthermia the uniform heat with discrete peaks
temperature more 41C is basic for cancer therapy (Franckena et al., 2009) that is not enough
for essential change of the thermodynamic entropy in the tumor.
To answer on this question we studied the growth dynamics for Guerin carcinoma during
treatment by DOXO under influence of inhomogeneous EF and accessory uniform and
nonuniform heat in tumor activated by external water heating. Experimental animals were
treated by DOXO (Pharmacia & Upjohn) in the dose 1.5 mg/kg. The treatment was
performed four times by DOXO, EI and external uniform and nonuniform heating by the
rubber hot-water bottles from 9

to 15 days after tumor transplantation every other two days.
The growth kinetics of Guerin carcinoma was varied for different groups (Table 4). Spatially
inhomogeneous EF and nonuniform heat field in the range of physiological hyperthermia
was maximally increased antitumor effect of DOXO for transplanted Guerin carcinoma. But
temperature in the tumor for this case had a lesser value.
We can suppose that increase of antitumor effect by inhomogeneous EF for drug during
treatment of the tumor accompanied by the change of thermodynamical entropy.

Parameters
Treatment
Temperature in the
centre of tumor, C , day
-1

Control (without DOXO, EI and
accessory heat)
36.5 0.54 0.06 1.00
DOXO 36.5 0.42 0.02* 1.28
DOXO + accessory uniform heat +
EI by AAP
41.5 0.38 0.01* 1.43
DOXO + accessory uniform heat 40 0.37 0.01* 1.45
DOXO + accessory nonuniform
heat
38 0.36 0.01* 1.50
DOXO + EI by AAP 37.9 0.35 0.01* 1.53
* Statistically significant difference from control group
Table 4. The growth kinetics of Guerin carcinoma during 15 days after tumor transplantation
The Effect of Spatially Inhomogeneous Electromagnetic Field and Local Inductive Hyperthermia
on Nonlinear Dynamics of the Growth for Transplanted Animal Tumors

301
It is well known that EF can initiate electro- and magnetocaloric effects. The electro- and
magnetocaloric effects are electro- and magneto-thermodynamic phenomenons in which a
reversible change in temperature of a suitable material is caused by exposing the material to
a changing EF. It was accompanied by changes in transfers from electromagnetic to
thermodynamic entropy and enthalpy (Nikiforov, 2007; Crosignani & Tedeschi, 1976).
Therefore, we can symbolically included high-frequencies electromagnetic IH in separate
class of electro- and magnetocaloric effects.
Described above physicochemical interaction between spatially inhomogeneous electric,
magnetic and temperature fields in the phantoms was probably similar to physicochemical
interaction in the tumor. They could influence on nonlinear dynamics of biological process.
We suppose, that it was interconnection between nonlinear conversion effects of spatial
inhomogeneous electric, magnetic fields (a
E
= 0.89 a.u.; a
H
= 0.48 a.u.) and initiated spatial
inhomogeneous temperature field in the heterogeneity tumor structure during propagation
of radio waves through malignant tissues. Entropy action is expressed in increase of
antitumor effect for DOXO. Alongside located normal tissue toxicity effect was minimal
through low level their heterogeneity.
In future we will be able to develop of novel and effective strategies for prevention and
treating cancers on the basis of understanding of nonlinear dynamics of adaptive systems
associated with tumorigenesis aspects during signaling interaction between cancer cells and
the host for complex treatment of patients by whole-body irradiation with local varying
spatial inhomogeneous EF.
4.3 Nonlinear model of growth dynamics for transplanted animal tumor during
irradiation by spatially inhomogeneous electromagnetic field and inductive
hyperthermia
Spatially inhomogeneous EF and initiated it heat manage the formation and disintegration
of dissipative structures lying in the basis of self-organization processes in organism at
physiological hyperthermia. We applied Waddingtons epigenetic landscape model which is
a metaphor for how gene regulation modulates development to interpret the changes in
thermodynamical parameters (entropy, enthalpy etc.) during nonlinear tumor growth of
transplanted animal tumors (Goldberg et al., 2007). The traditional mechanist, pathway-
centered explanation assumes that a specific, instructive signal i.e., a messenger molecule
or external signal of that interacts with its cognate cell surface receptor, tells the cells which
particular genes to active in order to establish a new cell phenotype. Essentially, cell
distortion triggered the cell to select between different preexisting attractor states (Sole, R.
et al., 2006). A certain chemical reaction is performed at different temperatures and the
reaction rate is determined. The reaction rate (k) for a reactant or product in a particular
reaction is intuitively defined as how fast a reaction takes place according to the Eyring
Polanyi equation:

=
S H
B R RT
k T
k e e
h
, (8)
where: k
B
is Boltzmann's constant, h is Planck's constant, T is absolute temperature, S is
entropy of activation, H is enthalpy of activation, R is gas constant (Polanyi, 1987).
The interaction effect of spatially inhomogeneous EF with heterogenous structure of animal
tumors just as described above for the phantoms initiated spatially inhomogeneous thermal
Nonlinear Dynamics

302
field gradient in malignant tissues in the range physiological hyperthermia. It was
accompanied by stochastic changes in transfers from electromagnetic to thermodynamic
entropy S and enthalpy H of activation and, respectively, stochastic changes of the
reaction rate that influence on nonlinear (chaotic) aspects in malignant growth (random
effect of increase or decrease) for transplanted animal tumors (see chapter 3.1). Spatially
inhomogeneous EF with increased asymmetry parameters during treatment of animal
tumors by DOXO (Table. 4) accompanied by the change of entropy of activation (S), the
reaction rate k (eq.8) and initiate enzyme catalysis topoisomerase II-mediated DNA damage
and free radical formation, absorbing them into double helix of DNA and resulting damage
of tumor cells. In this case the number of free radicals increased, in our opinion, as a result
of the effect of spin conversion in radical electron pair.
Let us consider kinetic model of tumor growth under the action of DOXO and nonuniform
heat field in the range of physiological hyperthermia initiated by spatially heterogeneous
EF. Let tumor cells multiplied with the growth factor , and DNA of some part of cells loses
their ability for replication under the action of DOXO and nonuniform heat field. The
appropriate equation can be written as

dx
x v
dt
= . (9)
where x is the number of tumor cells in unit volume with capable of replication DNA, v is
the rate of appearing of tumor cells with damaged DNA, which is unable to replicate.
Doxorubicin is known to interact with DNA by intercalation and inhibits the progression of
the enzyme topoisomerase II, which unwinds DNA for transcription. Doxorubicin stabilizes
the topoisomerase II complex after it has broken the DNA chain for replication, preventing
the DNA double helix from being resealed and thereby stopping the process of replication.
Schematically this reaction can be written down as:
DOXO + [TOP+DNA] DNA*, (10)
where [TOP+DNA] is topoisomerase II complex, DNA* is damaged DNA.
Let y = C
DOXO
is the concentration of DOXO, y(0) = y
0
beginning maximal concentration of
DOXO, y0; u = C
TOP
is the concentration of topoisomerase II, u > 0. For the open system the
concentration of DOXO and TOP in the reaction (10) is described taking into account diffusion:
(11)
2
2
2
2
,
,
y
u
y y
r D
t l
u u
r D
t l

= +

= +

(12)
where r is reaction rate, D
y
and D
u
is effective diffusion rate, l is spatial coordinate.
In accordance with kinetic law of mass action during steady quasistationary regime in the
system the rate r of reaction (10) is expressed as
r = kyu, (13)
where k is the constant of reaction rate (Ederer & Gilles, 2007).
The concentration u of topoisomerase II is related with the number x of tumor cells in unit
volume:
The Effect of Spatially Inhomogeneous Electromagnetic Field and Local Inductive Hyperthermia
on Nonlinear Dynamics of the Growth for Transplanted Animal Tumors

303
u = ax, (14)
where a is a coefficient.
The rate v of appearing of tumor cells with damaged DNA determined by the cells with
topoisomerase II reacted in (10):

1 du
v
a dt
= . (15)

Putting in (15) the expression for
du
dt
from (12) and taking (14) into account, we will get

2
2 x
r x
v D
a l

. (16)
Thus, equations (9) and (11) it is possible to write down as a system:

2
2
2
2
,
.
x
y
dx r x
x D
dt a l
dy y
r D
dt l

= +

= +

(17)
The constant of reaction rate k depends on temperature T according to Arrhenius equation:

E
RT
k Ae

= . (18)
Taking (13) and (18) into account the system (17) will look like:

2
2
2
2
,
,
E
RT
x
E
RT
y
dx x
x Ae xy D
dt l
dy y
aAe xy D
dt l

= +

= +

(19)
with initial condition y(0) = y
0
and edge conditions x > 0 and y > 0.
The system of equations (11) describes the nonuniform thermal effect of the spatially
inhomogeneous EF on the growth kinetics of the number of tumor cells under the action of
DOXO.
According to the presented data, one may suppose that recorded effects of growth inhibition
for DOXO-resistant Guerin's carcinoma after treatment by DOXO and local EI by EF with
increased spatial inhomogeneity (a
E
= 0.89 a.u.; a
H
= 0.48 a.u.) may be connected with the
initiation of membrane depolarization due to two steps. Firstly ionic cyclotron resonance
and next paramagnetic resonance (Liboff AR, 1985; Blanchard & Blackman 1994; Bezrukov
& Vodyanoy, 1997), which initiated the antitumor activity of DOXO. Its biochemical
mechanisms may be the alteration of the tumor microenvironment via changes in the pH
gradient between the extracellular environment and the cell cytoplasm (De Milito & Fais,
2005) and probably EF influency on free radical metabolism of human body (Jin et al., 1998).
Thus, we can assert that spatially inhomogeneous EF and local IH initiated in tumor of the
reactions with multiple physicochemical properties.
Nonlinear Dynamics

304

a b

c
Fig. 14. Spatial distribution of entropy of activation in the tumor during treatment by
Doxorubicin hydrochloride C
27
H
29
NO
11
HCl and spatial inhomogeneity electromagnetic
field with increased asymmetry parameters: Doxorubicin hydrochloride; b Doxorubicin
hydrochloride under the action of spatially inhomogeneous EF and IH; - entropy of
activation and tumor growth
Our preclinical and early clinical data suggest that combining superficial and intracellular
agents can synergize and leverage single-agent activity. The aforementioned effect of
influence of spatially inhomogeneous EF and local IH at physiological temperatures on
increase of antitumor activity for drug used in clinical practice during chemotherapy of
cancer patients (Nikolov et al., 2008).
5. Conclusion
1. EI by spatially inhomogeneous EF and local IH in the range physiological hyperthermia
of transplanted animal tumors manifests many of nonlinear (chaotic) aspects in
malignant growth.
2. An increase of spatially inhomogeneous EF and local IH in the range physiological
hyperthermia increased antitumor effect of DOXO for transplanted DOXO-resistant
Guerin's carcinoma and accompanied by the change of thermodynamical entropy.
3. Understanding the chaotic theory

for cancer and its interplay may enable

similar
strategies to be employed in the treatment of cancer by spatially inhomogeneous EF and
local IH in the range physiological hyperthermia.
The Effect of Spatially Inhomogeneous Electromagnetic Field and Local Inductive Hyperthermia
on Nonlinear Dynamics of the Growth for Transplanted Animal Tumors

305
6. Acknowledgements
The authors would like to thank Ph.D. Dunaevsky V.I. for thermography measurements. We
wish to give special thanks to Ph.D. Dzyatkovskya N.N., Dr.-Ing. Nikolov N.A., Dr.-Ing.
Melnik Yu.I. and Dzyatkovskya I.I. for experimental studies and joint discussions.
7. References
Ares, F., Rengarajan, S, Lence, J., Trastoy, A. & Moreno, E. (1996). Synthesis of antenna
patterns of circular arc arrays. Electronics Letters, Vol. 32, No. 20, P. 18451846.
Bailey, T. & Gatrell, A. (1995). Interactive Spatial Data Analysism, Wiley, New York.
Baish, J. & Jain, R. (2000). Fractals and Cancer. Cancer Res., Vol. 60., P. 36833688.
Bezrukov, S. & Vodyanoy, I. (1997). Signal transduction across alamethicin ion channels in
the presence of noise. Biophys. J., Vol. 73, P. 24562464.
Blanchard, J. & Blackman, C. (1994). Clarification and application of an ion paramagnetic
resonance model for magnetic field interactions with biological resonance systems.
Bioelectromagnetics, Vol. 15, P. 217238.
Blank, M. & Soo, L. (2001). Electromagnetic acceleration of electron transfer reactions. J. Cell.
Biochem., Vol. 81, P. 278283.
Blazsek, I. Innate chaos: I. (1992). The origin and genesis of complex morphologies and
homeotic regulation. Biomed. Pharmacother., Vol. 46, No. 57, P. 219235.
Boddie, A.; Frazer, J. & Yamanashi, W. (1987). RF electromagnetic field generation apparatus
for regionally-focused hyperthermia, United States Patent N4674481 on 23. 06.1987.
Carrubba, S.; Frilot, C; Chesson, A. & Marino, A. (2007). Evidence of a nonlinear human
magnetic sense. Neuroscience, Vol. 144, No. 1, P. 356367.
Carrubba, S.; Frilot, C.; Chesson, A.; Webber, C.; Zbilut, J. & Marino, A. (2008).
Magnetosensory evoked potentials: consistent nonlinear phenomena. Neurosci. Res.,
Vol. 60, No. 1, P. 95105.
Chang, K., Chang, W., Yu, Y, Shih, C. (2004). Pulsed electromagnetic field stimulation of
bone marrow cells derived from ovariectomized rats affects osteoclast formation
and local factor production. Bioelectromagnetics, Vol. 25, No. 2, P. 134141.
Chen, Q.; Tong, M. & Dewhirst F. (2004). Targeting tumor microvessels using doxorubicin
encapsulated in a novel thermosensitive liposome. Mol. Cancer Ther., Vol. 3, P.
13111317.
Cramer, F. (1993). Chaos and order. The complex structure of living systems, VCH
Verlagsgesellschaft, Weinheim.
Crosignani,

B. & Tedeschi,

A. (1976). Variation of the entropy of an electromagnetic field
due to scattering. Lettere Al Nuovo Cimento, Vol. 17, No. 4, P. 141143.
De Mattei, M., Gagliano, N., Moscheni, C., Dellavia, C., Calastrini, C., Pellati, A., Gioia, M.,
Caruso, A. & Stabellini, G. (2005) Changes in polyamines, c-myc and c-fos gene
expression in osteoblast-like cells exposed to pulsed electromagnetic fields.
Bioelectromagnetics, Vol. 26, No. 3, P. 207214.
De Milito, A. & Fais, S. (2005). Proton pump inhibitors may reduce tumour resistance. Expert
Opinion on Pharmacotherapy, Vol. 6, P. 10491054.
Diniz, P., Shomura, K., Soejima, K. & Ito, G. (2002). Effects of pulsed electromagnetic field
(PEMF) stimulation on bone tissue like formation are dependent on the maturation
stages of the osteoblasts. Bioelectromagnetics.V. 23, Issue 5, Pages 398 405.
Nonlinear Dynamics

306
Dirks, P. (2008).Brain tumor stem cells: bringing order to the chaos of brain cancer. J Clin
Oncol, Vol. 26, No. 17, P. 29162924.
Ederer, M. & Gilles E. (2007). Thermodynamically feasible kinetic models of reaction
networks. Biophysical Journal, Vol. 92, No. 6, P. 18461857.
Emanuel, N. (1977). Kinetics of experimental tumor processes, Nauka, Moscow (in Russian).
Falk, M. & Issels, R. (2001). Hyperthermia in oncology. Int J Hyperthermia, Vol. 17, P. 118
Franckena, M.; Fatehi, D.; de Bruijne, M.; Canters, R.; van Norden, Y.; Mens, J.; van Rhoon,
G. & van der Zee, J. (2009). Hyperthermia dose-effect relationship in 420 patients
with cervical cancer treated with combined radiotherapy and hyperthermia. Eur J
Cancer, Vol. 45, No. 11, P. 19691978.
Furusawa, C. & Kaneko, K. (2000). Origin of complexity in multicellular organisms. Phys.
Rev. Lett., Vol. 84, No. 26, Pt. 1, P. 61306133.
Gaber, M. (2002). Modulation of doxorubicin resistance in multidrug-resistance cells by
targeted liposomes combined with hyperthermia. J Biochem Mol Biol Biophys, Vol. 6,
P. 309314.
Garte, S. (2003). Cancer epidemiology.Theory in carcinogenesis and epidemiology. Journal of
Epidemiology and Community Health, Vol. 57, P. 85.
Goldberg, A.; Allis, C. & Bernstein, E. (2007). Epigenetics: A landscape takes shape. Cell, Vol.
128, P. 635638.
Goodman, R. & Henderson, A. (1988). Exposure of salivary gland cells to low-frequency
electromagnetic fields alters polypeptide synthesis. Proc. Natl. Acad. Sci. USA, Vol.
11, P. 39283932.
Hardell, L. & Sage, C. (2008). Biological effects from electromagnetic field exposure and
public exposure standards. Biomed Pharmacother, Vol. 62, No. 2, P. 104109
Hirata, M.; Kusuzaki, K.; Takeshita, H.; Hashiguchi, S.; Hirasawa, Y. & Ashihara, T. (2001).
Drug resistance modification using pulsing electromagnetic field stimulation for
multidrug resistant mouse osteosarcoma cell line. Anticancer Res., Vol. 21, P. 317
320.
Ivkov, R., De Nardo, S., Daum, W., Foreman, A., Goldstein, R., Nemkov, V. & De Nardo G.
(2005). Application of high amplitude alternating magnetic fields for heat induction
of nanoparticles localized in cancer. Clinical Cancer Research, Vol. 11, P. 70937103.
Jin, Y., Wang, H., Cheng, Y. & Gu, H. (1998). Effects of static magnetic fields on free radical
metabolism of human body. Wei Sheng Yan Jiu, Vol. 27, No. 2, P. 9799.
Jordan, E. & Balmain, K. (1968). Electromagnetic waves and radiating system, Prentice Hall,
New Jersey.
Kattapuram, S., Rosol, M., Rosenthal, D., Palmer, W. & Mankin, H. (1999). Magnetic
resonance imaging features of allografts. Skeletal Radiol, Vol. 28, No. 7, P. 383389.
Korn, G. & Korn, T. (1968). Mathematical handbook for scientists and engineers, MacGraw-Hill
Book Company, New York.
Liboff, A. (1985). Cyclotron resonance in membrane transport. In: Interaction between
electromagnetic field and cells, Chiabrera, A.; Nicolini, C. & Schwan, H. (Eds.), P.
281296, Plenum, New York.
Longo, I. & Ricci A. (2007). Chemical activation using an open-end coaxial applicator. J.
Microw. Power Electromagn. Energy, Vol. 41, P. 419.
Marino, A.; Wolcott, M.; Chervenak, R.; Heuil, F.; Nilsen, E. & Frilot, C. (2000). Nonlinear
response of the immune system to power-frequency magnetic fields. Am. J. Physiol.
Regul. Integr. Comp. Physiol., Vol. 279, No. 3, P. 761768.
The Effect of Spatially Inhomogeneous Electromagnetic Field and Local Inductive Hyperthermia
on Nonlinear Dynamics of the Growth for Transplanted Animal Tumors

307
Marino, A.; Carrubba, S.; Frilot, C. & Chesson, A. (2009). Evidence that transduction of
electromagnetic field is mediated by a force receptor. Neurosci. Lett., Vol. 452, No. 2,
P. 119123.
Martino, F. Alternative inductothermia in cancer. (1962). A confirmation with therapeutic
applications of Warburg's theory. Cancro, Vol. 15, P. 358385.
Marmor, J. (1979). Interactions of hyperthermia and chemotherapy in animals. Cancer
Research, Vol. 39, P. 22692276.
Mittra, R. (1973). International Series of Monographs in Electrical Engineering. Computer
Techniques for Electromagnetics, Pergamon Press, Oxford & New York.
Miyagi, N., Sato, K., Rong, Y., Yamamura, S., Katagiri, H., Kobayashi, K. & Iwata, H. (2000).
Effects of PEMF on a murine osteosarcoma cell line: drug-resistant (P-glycoprotein-
positive) and non-resistant cells. Bioelectromagnetics, Vol. 21, No. 2, P. 112121.
Molnar, J.; Thornton, B.; Thornton-Benko, E.; Amaral, L.; Schelz, Z. & Novak, M. (2009).
Thermodynamics and Electro-Biologic Prospects for Therapies to Intervene in
Cancer Progression. Current Cancer Therapy Reviews, Vol. 5, No. 3, P. 158169.
Moseley, H. (1988). Non-ionizing radiation. Medical physics handbooks, Adam Hilger, Bristol &
Philadelphia.
Nikiforov, V. (2007). Magnetic induction hyperthermia. Russian Physics Journal, Vol. 50, No.
9, P.913924.
Nikolov, N.; Orel, V.; Smolanka, I.; Dzyatkovskaya, N.; Romanov, A.; Mel'nik, Y.; Klimanov
M. & Chernish, V. (2008). Apparatus for Short-Wave Inductothermy
Magnetotherm, Proceedings of NBC 2008, pp. 294298, Katushev, A.;
Dekhtyar, Yu. & Spigulis, J. (Eds), Springer-Verlag, Berlin, Heidelberg.
Orel, V. & Dzyatkovskaya, N. (2000). Mechanoemission of blood and oncogenesis. In:
Biophotonics and coherent systems, Beloussov, L.; Popp, F. & van Wijk, R. (Eds.), P.
347363, Moscow University Press.
Orel, V.; Grinevich, Y.; Dzyatkovskaya, N.; Danko, M.; Romanov, A.; Mel'nik, Y. &
Martynenko, S. (2004). Spatial & Mechanoemission Chaos of Mechanically
Deformed Tumor Cells. Journal of Mechanics in Medicine & Biology, Vol. 4, P. 3145.
Orel, V.; Kudryavets, Y.; Bezdenezhnih, N.; Danko, M.; Khronovskaya, N.; Romanov, A.;
Dzyatkovskaya, N. & Burlaka, A. (2005). Mechanochemically activated doxorubicin
nanoparticles in combination with 40MHz frequency irradiation on A-549 lung
carcinoma cells. Drug Delivery, Vol. 12, P. 171178.
Orel, V.; Kozarenko, T.; Galachin, K.; Romanov, A. & Morozoff, A. (2007a). Nonlinear
Analysis of Digital Images and Doppler Measurements for Trophoblastic Tumor.
Nonlinear Dynamics, Psyhology and Life Science, Vol. 11, P. 309331.
Orel, V.; Dzyatkovskaya, N.; Romanov, A. & Kozarenko, T. (2007b). The effect of
electromagnetic field and local inductive hyperthermia on nonlinear dynamics of
the growth of transplanted animal tumors. Experimental Oncology, Vol. 29, No. 2,
P. 156158.
Orel, V.; Nikolov, N.; Dzyatkovskaya, N.; Romanov, A.; Melnik, Y.; Dunaevsky V. &
Dzyatkovskaya, I. (2008). Influence of change of spatial nonuniformity of the
electromagnetic field on transformation of radio-waves and thermal characteristics
of phantoms and Lewis lung carcinoma. Physics of the Alive, Vol. 16, No. 2, P. 92
98 (In Ukrainian).
Orel, V.; Dzyatkovskaya, .; Nikolov, N.; Romanov, A.; Dzyatkovskaya, N.; Kulik, G.; Todor,
.; Chranovskaya, N. & Skachkova, . (2009). The influence of spatially nonuniform
electromagnetic field on antitumor activity of cisplatin during treatment of resistant
Nonlinear Dynamics

308
substrain of Lewis lung carcinoma. Ukrainian Radiology Journal, Vol. 17, P. 7277
(in Ukrainian).
Pasquinelli, P., Petrini, M., Mattii, L., Galimberti, S., Saviozzi, M. & Malvaldi G. (1993).
Biological effects of PEMF (pulsing electromagnetic field): an attempt to modify cell
resistance to anticancer agents. J Environ Pathol Toxicol Oncol, Vol. 12, P. 193197.
Polanyi, J. (1987). Some concepts in reaction dynamics. Science, Vol. 236, P. 680690.
Purcell, E. (1985). Electricity and Magnetism, McGraw-Hill, New York.
Reszka, K., Mc Cormick, M. & Britigan, B. (2001). Peroxidase- and nitrite-dependent
metabolism of the anthracycline anticancer agents daunorubicin and doxorubicin.
Biochemistry, Vol. 40, P. 1534915361.
Roemer, R. (1999). Engineering aspects of hyperthermia therapy. Annual Review of Biomedical
Engineering, Vol. 1, P. 347376.
Rubin, H. (1984). Cancer as a dynamic developmental. Cancer Res., Vol. 45, P. 29352942.
Ruiz-Gmez

M., de la Pea, L., Prieto-Barcia, M., Pastor, J., Gil, L. & Martnez-Morillo, M.
(2002). Influence of 1 and 25 Hz, 1.5 mT magnetic fields on antitumor drug potency
in a human adenocarcinoma cell line. Bioelectromagnetics, Vol. 23, No. 8, P. 578
525.
Sedivy, R. & Mader, M. (1997). Fractals, chaos, and cancer: do they coincide? Cancer Invest.,
Vol. 15, P. 601607.
Schneider, B. & Kulesz-Martin, M. (2004). Destructive cycles: the role of genomic instability
and adaptation in carcinogenesis. Carcinogenesis, Vol. 25, No. 11, P. 20332044.
Shen, J.; Zhang, W.; Wu, J. & Zhu, Y. (2008). The synergistic reversal effect of multidrug
resistance by quercetin and hyperthermia in doxorubicin-resistant human
myelogenous leukemia cells. Int J Hyperthermia, Vol. 24, No. 2, P. 151159.
Sole, R.; Garsia, I. & J.Costa. (2006). Spatial Dynamics in Cancer. In: Complex Systems Science
in Biomedicine Series: Topics in Biomedical Engineering, Deisboeck, T. & Kresh, J.(Ed.),
P. 557572, International Book Series, Springer US.
Solyanik, G.; Todor, I.; Kulik, G. & hekhun, V. (1999). Selective mechanism of the
emergence of Guerin`s carcinoma resistance to doxorubicin. Experimental oncology,
Vol. 21, P. 264268.
Song, C.; Park, H.; Lee, C. & Griffin, R. (2005). Implications of increased tumor blood flow
and oxygenation caused by mild temperature hyperthermia in tumor treatment.
Int. J. Hyperthermia, Vol. 21, P. 761767.
Waliszewski, P.; Molski, M. & Konarski, J. (1998). On the holistic approach in cellular and
cancer biology: nonlinearity, complexity, and quasi-determinism of the dynamic
cellular network. J. Surg. Oncol., Vol. 68, No. 2, P. 7078.
Weaver, J.; Vaughan, T. & Martin, G. (1999). Biological effects due to weak electric and
magnetic fields: the temperature variation threshold. Biophys J., Vol. 76, No. 6, P.
30263030.
Weller, A.; Anton, M.; Geiger, J.; Hirsch, M.; Jaenicke, R.; Werner, A. & Nhrenberg, C.
(2001). Survey of magnetohydrodynamic instabilities in the advanced stellarator.
Phys. Plasmas, Vol. 8, P. 931.
Yarosh, D. Why is DNA damage signaling so complicated? (2001). Chaos and molecular
signaling. Environmental and Molecular Mutagenesis, Vol. 38, No. 23, P. 132134.
13
Advanced Computational Approaches
for Predicting Tourist Arrivals:
the Case of Charter Air-Travel
Eleni I. Vlahogianni, Ph.D. and Matthew G. Karlaftis, Ph.D.
Department of Transportation Planning and Engineering, School of Civil Engineering,
National Technical University of Athens, 5, Iroon Polytechniou Str., Zografou Campus,
Athens 157 73,
Greece
1. Introduction
Tourism is one of the major industries profiting various sectors of the economy, such as the
transportation, accommodation, entertainment and so on. According to the World Tourism
Organization (2008), international tourism grew at around 5% during the first four months
of the year 2008. Fastest growth was observed in the Middle East, North-East and South
Asia, and Central and South America. Even though, uncertainty over the global economic
situation is affecting consumer confidence and could hurt tourism demand, for 2008 as a
whole, UNWTO maintains a cautiously positive forecast. Moreover, international trends
show that tourists are becoming more discerning in their choice of destinations and,
therefore, becoming less predictable and more spontaneous in terms of their consumption
patterns (Burger et al. 2001).
Air transportation is probably the most important mode for international travel and leisure.
A typical characteristic of air tourism in Europe is the extensive use of non-
scheduled/charter flights and the existence of low-cost carriers in the leisure travel market,
that account for 8% of passengers and 3% or revenues in the aviation industry (Dresner
2006). Non-scheduled demand is typical in Mediterranean countries where connections are
essentially touristic and characterized by non-scheduled services.
In this type of air travel, the ability to accurately predict tourist arrivals is of importance in
the successful management and operation of the airport facilities, as well as the adjacent
transportation network. Yet, the literature has little to offer in modeling demand stemming
from non-scheduled flights, as such series exhibit seasonality, intense variability and
inherent unpredictability.
This paper develops and tests advanced computational approaches in order to predict non-
scheduled/charter international tourist demand. The computational challenges that may
arise in such a problem are twofold: first, to treat seasonal and stochastic characteristics of
non-scheduled tourist demand, and, second, to develop models that consider past tourist
demand characterists. This paper focuses on developing ARFIMA models that consider both
non-stationarity and long-term memory effects in the auto-regressive process and temporal
neural networks with advance genetically optimized characteristics that treat both
nonlinearity and non-stationarity.
Nonlinear Dynamics

310
2. Motivators and prediction of non-scheduled air-travel demand
A major motivator for the emergence and growth of non-scheduled air travel has been the
low-cost carriers (LCC) and their prevalence in global aviation. From the period after 9/11
period that caused a decreasing trend in the airline travel demand, global aviation and
travel demand, particularly in Europe and the Mediterranean Region LCCs offered an
attractive alternative for price-sensitive clients during the tight economic times. Whereas
traditional airlines have concentrated on large cities and major airports, low-cost airlines
have turned to under-utilized airports at some distance from the main population centers
embracing a business model much different in its customer base, air network, and provision
of services by focusing on the more cost-sensitive leisure travel and working in a way that
traditional airlines cannot (Barrett 2000).
LCC market providing point-to-point (rather than hub-based) service owes its growth not
only to low-cost service, but also to the ability to focus on customer segments not
emphasized by larger carriers; European low-cost leaders Ryanair and EasyJet, for instance,
focus on providing air services for travelers seeking to visit friends and relatives. By
focusing on these groups, LCC have demonstrated an ability to grow the overall passenger
market, particularly on routes with strong tourist appeal (Dennis 2004).
Literature emphasizes the role of LCC in the development of multiple airport systems and
the emergence of secondary airports (Bonnefoy & Hansman 2004). LCC appeal to secondary
airport is in that they provide reduced congestion and lower cost, while still providing
access to key population centers.
The shift to secondary airports, along with the reduced gap between charter flights and no-
frills / budget flights have significant impact on the volatility of traffic for the entire airport
system; literature indicates that periods of high volatility and uncertainty in demand exist
during the developmental phases of secondary airports that can last up to 20 year after the
opening of such facilities (de Neufville, 1995).
Regarding leisure airline traffic, the ability to provide custom-made services to tourists has
been shown to be critical. Tourists increasingly expect to experience a personalized and
close to their life-style service (Graham 2006). A characteristic example of charter airports is
Greece where approximately 80% of the total tourist arrivals every year are accommodated
by air transportation. The importance of non-scheduled international arrivals is depicted in
Figure 1 that depicts annual evolution of total arrivals for 1989 and 2006 period, along with
the evolution of non-scheduled and scheduled international arrivals. As can be observed, for
the period after 2001, nearly 70% of air-travel arrivals concern international flights and 62%
of the international arrivals are accommodated by non-scheduled flights.
From a methodological standpoint, although the prediction of tourist demand has been
extensively treated (a review of approaches can be found in Law et al. 2007, Song & Li 2008)
little has been done towards the prediction of non-scheduled arrivals. Summarizing the
methodologies implemented to date for to tourist demand prediction, both econometrics
and other computational methods have been applied and compared. Law et al. (2007) state
that, comparing classical econometric prediction techniques that are highly exploited but
with marginal improvement to modeling touristic demand, the incorporation of data mining
techniques has led to some ground breaking outcomes.
Moreover, several papers on tourism forecasting problems report neural networks as having
better performance than classical statistical techniques, such as ARIMA models, exponential
Advanced Computational Approaches for Predicting Tourist Arrivals: the Case of Charter Air-Travel

311
smoothing and so on (Law and Au 1997, Law 2000, Burger et al. 2001, Kim et al. 2003, Cho
2003). These studies compare advanced computational approaches that have enhanced
capabilities in modeling nonlinear characteristics (for example neural networks) with simple
linear and stationary approaches such as the ARIMA models. Quite recently, hybrid ARIMA
and simple static neural networks, as well as mixtures of static neural network models have
also been found to perform better that classical time-series approaches (Aslanargun et al.
2007).
Regarding modeling of non-scheduled demand, previous work has applied regression
models to predict charter international arrivals to major Greek airports and has highlighted
that although there is uncertainty and variability in their evolution, historical data can be
used to provide good predictions (Karlaftis and Papastavrou 1998). However, no previous
work has been conducted in the direction of predicting non-scheduled international arrivals
in secondary airports with intense seasonal characteristics.
3. Computational approaches
3.1 Fractionally integrated autoregressive moving average processes
Commonly applied AR(I)MA models are able to describe processes that are covariance
stationary I(0) or non-stationary through differencing I(1). It has been observed that the
erroneous consideration of having a unit root leads to models with inflated estimates of the
moving average component (Box-Steffensmeier and Smith, 1998). In order to account for
long memory processes Fractional integration is introduced to autoregressive processes to
account for the processes that are neither I(0) or I(1) in the form of the differentiation
operator (Baillie 1999):
( ) ( ) ( )
2 3
1 1 1 1 ...
2! 3!
d L L
L dL d d d d

=


(1)
In the conditional mean, the fractionally integrated autoregressive moving average process
of orders p and q ARFIMA(p,d
m
,q) introduced by Granger and Joyeux (1980) and Hosking
(1981) is represented by the following equation:
( ) (1 ) ( ) ( )
m
d
t t
L L y L = (2)
~ (0, 1)
t t t t
z z N = (3)
where is the unconditional mean of y
t
,
2
1 2
( ) 1 ...
p
p
L L L L = and
2
1 2
( ) 1 ...
q
q
L L L L = + + + + are the AR and MA polynomials having all roots outside the
unit cycle, while innovations
t
are i.i.d distributed with
2

t
being the conditional variance
and a positive, time-varying, and measurable function with respect to the information set,
which is available at time t-1 (Baillie et al. 2002). The differentiation parameter (d
m
) is
associated with the following statistical properties of a (time) series (Hosking 1981, Odaki
1993):
For every region where
1
2 m
d < , then y
t
is stationary,
When

1
2
1
m
d < < , the series exhibits invertibility,
Nonlinear Dynamics

312
When
1
2
0
m
d < , the stationary process y
t
is antipersistent,
1

When

0
m
d = , the stationary process y
t

has short memory and is mean reverting,
When

1
2
0
m
d < , y
t
is fractionally integrated and exhibits long memory,
When

1
2
1
m
d < < , the process y
t
is mean-reverting, but the stationarity property cannot
be verified and,
When 1
m
d = , y
t
is a unit root process.
Fractionally integrated processes are significant in dealing with two issues: first, data is
being modeled more precisely, as the knife-edged restriction of an I(0) or I(1) process is
avoided and both long term persistence and, second, short-term correlation structure of a
series can be modeled (Hosking 1981).
3.2 Temporal genetically optimized neural networks
Temporal Neural Networks can be considered as an extension of the static Multi-layer
Perceptrons (MLP) that has been extensively applied to touristic demand prediction. They
differ from the commonly used MLPs in that they incorporate memory mechanisms in their
structure that can be limited to the input layer or extend to the entire network. The memory
acts as a time-series reconstruction module with the aim to embed the scalar series S(t) to a
vector { } ( ) ( ),..., ( ( 1) ) t S t S t m = S in an m-dimensional vector space known as Phase
Space, where is the time delay of and m is the dimension.
We implement a neural network called time-lagged neural networks (TLNN) with a
complex Gamma memory mechanism in the input layer and the hidden layer (de Vries and
Principe 1992). Moreover, in order to develop a fully non-stationary model we set the
network to predict under the iterative consideration: Given the time-series of a variable a
single step ahead model is constructed to produce a prediction

( ) S t at time t that is then fed

backwards to the network and is used as new input data in order to produce the next step

( 1) S t + prediction at t+1:

{ }

( 1) ( ), ( ), ( 1)... S t S t S t S t + = (4)
The training of TLNN under iterative consideration feeds back the prediction at time t+1
and utilizes it as an input for the generation of next prediction step t+2. The training in the
specific iterative neural network model is conducted via the temporal back-propagation
algorithm known as Back-propagation to time (BPTT) (Webros 1990); all weights are
duplicated spatially for an arbitrary number of time steps ; as such, each node that sends
activation to the next has number of copies as well. For a training cycle n, the weight
update is given by the following equation (Haykin 1999):
( 1) ( ) ( ) ( )
ji ji j i
n n n n + = + w w x (5)
where, w
ji
(n+1) and w
ji
(n) are the weights of the i-th synapse of the neuron j at training cycle
n+1 and n respectively, is the learning rate, xi(n) (i=1,2,n) is the input vector and
j
(n) is
given by:

1
Anti-persistence is a property of an ACF that exhibits slow decay, but the original series
are not characterized by the long memory property; rather, the autocorrelations (in the ACF)
alternate in signs.
Advanced Computational Approaches for Predicting Tourist Arrivals: the Case of Charter Air-Travel

313

( ) ( ( )), neuron in the output layer
( )
( ( )) ( ) , neuron in the hidden layer
j j
j
j r rj
r
e n n j
n
n n j

w
(6)
where, e
j
(n) is the networks error, is the nonlinear activation function. Moreover, if A is a
set of all neurons whose inputs are fed by the j neuron in the hidden layer is a forward
manner, then
1
( ) ( )
m
j ji i j
i
n n b
=
= +

w x is the induced local field of neuron r that belongs to the

A and [ ] ( ) ( ), ( 1),..., ( )
T
r r r r
n n n n m

= + + is the local gradient vector.

The iterative neural network approach introduced provides a fully non-stationary and
nonlinear environment for treating time series problems. However, regardless of being static
or dynamic, neural networks suffer from the lack of an automatic manner to self-
optimization mainly with respect to their structure (number of hidden units) and learning
parameters. Recently, genetic algorithms have gained significant interest as they can be
integrated to the neural network training to search for the optimal architecture without
outside interference, thus eliminating the tedious trial and error work of manually finding
an optimal network. Genetic algorithms are based on three distinct operations: selection,
cross-over and mutation (Mitchell 1998); these operations run sequentially in order for a
fitness criterion (in the specific case the minimization of the cross-validation error) to
converge. Details for the specific optimization approach can be found in Vlahogianni et al.
(2005).
4. Case study: greek island airports
We focus on the influence of Non-Scheduled International (NSI) arrivals to the secondary
airports and a prediction of their temporal evolution. Three case studies from Greek island
secondary airports are evaluated: Heraklion (Crete), Kefalonia and Rhodes. All three cases
exhibit significant demand during the peak summer period; however, these case studies
differ in the overall demand levels, as well as their seasonal arrival characteristics. As can be
observed in Figure 2, where the evolution of arrivals (passengers per year) and flights per
year and per airport for the period of 1999-2006 is depicted, Kerkyra is characterized by low
volumes, whereas Heraklion and Rhodes exhibit high demand during the year. The
difference is in the volume of the NSI arrivals; as can be seen in Figure 3, where monthly
arrival variation is depicted for all airports tested, Kerkyra and Rhodes have significantly
more acute monthly variation, reaching extremely low NSI demand during the off-peak
periods.
The analysis to follow will, first, focus on revealing long-term memory features in the
manner NSI arrivals evolve in time and, second, search for similarities or differences in the
dynamics of NSI arrivals across the airports selected with different demand distributions.
Third, advanced neural network predictors will be developed that will apply the iterative
approach in order to learn to approximate the dynamics of NSI arrivals; models will be
developed for all the three airports and compared to each other.
4.1 Fractional dynamics in NSI arrivals
Several ARFIMA models were fitted to the available time series in order to test whether
there exist fractional dynamics in the evolution of non-scheduled international arrivals. The
Nonlinear Dynamics

314
models are fitted to both three study airport, as well as to the pooled data, as well as data
from the peak (months from May to September). Moreover, in the same datasets I(1) ARIMA
processes are also fitted in order to compare the estimated autoregressive and moving
average parameters from ARFIMA and ARIMA models. The choice of the best fitting model
is done via Akaikes ( -2 2
LogL k
n n
+ , where logL is the log likelihood value, n is the number
of observations and k the number of estimated parameters) and Schwartzs
(
log
-2 2
LogL k
n n
+ ) criteria. Furthermore, the Jarque-Bera test (JB test) goodness-of-fit test
measuring the of departure from normality, Q
2
(i) statistics that indicate the possible
existence of serial correlation in the standardized residuals, as well as the LM ARCH
statistics that test the null hypothesis of no ARCH effect in the series tested are also
presented; the above test will provide information of the quality of the ARFIMA models
developed.
Results for the best fitted ARFIMA models are shown in Tables 1 to 3( parameter estimates
depicted in the tables are significantly different from zero at the 1% significance level).
Interestingly, for all case studies the fractional dynamics are similar. NSI arrivals in all
airports tested are found to be best described by a fractionally integrated ARMA process
with p=1 (autoregressive term) and q=1 (moving average term). Parameter d is found to
vary between 0.24 and 0.46 indicating that NSI arrivals regardless of study period (peak or
off-peak), as well as of the airport tested, exhibit long-term memory (for more details on the
memory properties see Washington et al. 2003). We observe that the ARFIMA modeling
results exhibit an apparent inability to approximate the monthly variability of NSI
arrivals, particularly at low demand levels (off-peak months) (Figure 4).
4.2 Iterative predictions of NSI arrivals using temporal neural networks
For iterative predictions, the specifications of the TLNN are shown in Table 4. As can be
observed, the depth of the Gamma memory of the TLNN (parameters and m) are
genetically optimized during the learning, along with the number of hidden units h in the
hidden layer and the learning rate and momentum of back-propagation. The available
data is separated into three subsets in order to test the training (cross-validation) and then
test the performance of the network (testing). Moreover, the genetic algorithm optimization
specifications are also depicted on Table 4; a roulette selection method is applied in order to
select the parents according to their fitness. Moreover, the probabilities of cross-over and
mutation are to be equal to 0.9 and 0.09 respectively, following literature that indicates that
crossover should usually be selected at high values and mutation should approximate the
inverse of the number of chromosomes (population) and be much lower than the crossover
probability to avoid permutation (Gen and Cheng, 2000).
Results concerning the optimization of the look-back time window, or else the depth of the
memory of the iterative temporal neural networks, are shown in Table 5. Interestingly, the
required data to produce accurate predictions as determined by the genetic optimization
of the parameters and m during learning differ between Heraklion airport and the rest of
the cases examined. The recurrence of the dynamics in the Heraklion case is every 6 months,
whereas NSI arrivals of Kerkyra and Rhodes are affected by data from up to 4 months in the
past.
Results of the predictions (test set) using TLNN are seen in Table 6; predictions for the same
period using ARFIMA (averaged for the three airports) are also illustrated. As can be
Advanced Computational Approaches for Predicting Tourist Arrivals: the Case of Charter Air-Travel

315
observed, the TLNN has, overall, better accuracy that is evident both in the high and low
demand periods in all three airport cases examined. The averaged behavior of the ARFIMA
and TLNN models developed with respect to the actual and predicted NSI arrivals is
graphically represented in Figure 5. Interestingly, the accuracy of predictions seems to
decrease significantly in the case of low demand time periods, such as months between
November and March, where touristic arrivals to Greek islands are, in general, significantly
lower than the ones during summer months. The decreased accuracy in the case of Kerkyra
indicates the existence of significant stochasticity in the manner in which arrivals evolve in
low demand and off-peak period cases.
5. Discussion and conclusions
A large portion of tourist demand is conducted by air. Several air links can have intrinsic
characteristics concerning the touristic demand evolution with strong non-stationary and
seasonal characteristics. In this paper we implemented recent data mining techniques to
model tourist demand and developed two advanced models of time-series prediction: a
fractionally integrated autoregressive moving average model (ARFIMA) and a temporal
genetically optimized iterative neural networks. These models originate from different
methodological backgrounds and aim to evaluate different statistical properties of tourist
demand (such as the existence of long-term memory, the parameters of memory depth for
predictions and so on). To evaluate the proposed methodologies, three cases studies were
examined that encompass three secondary airport located in the Greek Islands which exhibit
different yearly and monthly demand distributions.
In terms of prediction accuracy, the advanced form of temporal neural networks
implemented seems to outperform the ARFIMA model. This applies to both high and low
tourist demand periods. In terms of the knowledge acquired by the modeling, both
approaches revealed very interesting results; the fractional dynamics observed in both the
pooled data and the peak demand period, show that the tourist arrivals are not always
stationary or best described as most frequently - assumed - by ARIMA models. The
fractionally integrated processes fitted to the available data showed that all case studies
examined have similar fractional dynamics and exhibit long term memory. This finding has
significant implications to the process of modeling NSI arrivals, as it suggests the
persistence of the effect of several socio-economic issues to the evolution of NSI arrivals.
Moreover, the iterative approach to predicting NSI arrivals showed significant improvement
to the prediction accuracy. The advanced genetic optimization implemented with regards to
the look-back time window of the TLNN shows that the past could be utilized to predict the
evolution of tourist demand. Nevertheless, the differences in the memory depth of the three
TLNN models developed to approximate the dynamics of NSI arrivals in the three airports
indicates the stochasticity of the temporal evolution of NSI arrivals during periods of low
volume that significantly affect the accuracy of predictions.
Finally, lack of prediction accuracy during transitional conditions reveals that, as expected,
the demand evolution can have multiple causal dimensions that need to be considered in an
effective methodological framework that could integrate both the temporal and
causal/relational characteristics of other possible influential variables in the prediction
process. Our ongoing work focuses on extending the present methodological framework to
iterative neural network prediction that incorporates other socio-economic data to develop
Nonlinear Dynamics

316
influential relationships and evaluate whether they can improve predictability during
periods of stochasticity in tourist demand.
6. References
Aslanargun, A., Mammadov, M., Yazici, B. and Yolacan, S. (2007), Comparison of ARIMA,
neural networks and hybrid models in time series: tourist arrival forecasting,
Journal of Statistical Computation and Simulation, 77(1), 29-53.
Baillie, R.T., Han, Y.W. and Kwon, T.-G. (2002) Further long memory properties of
inflationary shocks, South. Econ. J. 68 496510.
Barrett, S. D. (2000). Airport competition in the deregulated European aviation market.
Journal of Air Transport Management, 6(1), 13 - 27.
Bonnefoy, P., and Hansman, R. (2004). Emergence and Impact of Secondary Airports in the
United States. Proceedings of the 4
th
AIAA ATIO Forum in Chicago, Illinois. Retrieved
August 7, 2007.
Burger, C.J.S.C., Dohnal M., Kathrada M., Law R. (2001), A practitioners guide to time-series
methods for tourism demand forecasting: a case study of Durban, South Africa,
Tourism Management, 22, 403-409.
Cho, V. (2003). A comparison of three different approaches to tourist arrival forecasting,
Tourism Management 24, 323330.
de Neufville, R. (1995). Management of multi-airport systems: A development strategy.
Journal of Air Transport Management, 2 (2), 99-110.
Dennis, N. (2004). Can the European low cost airline boom continue? Implications for
regional airports: The 44th European Congress of Regional Science Association. Porto,
Portugal. Retrieved online on January 4, 2007.
Dresner, M. (2006). Leisure versus business passengers: Similarities, differences, and
implications, Journal of Air Transport Management, 12, 2832.
Graham, A. (2006). Have the major forces driving leisure airline traffic changed?, Journal of
Air Transport Management, 12(1), 14-20.
Granger, C.W.J. and Joyeux, R., (1980) An introduction to long-memory time series models
and fractional differencing. Journal of Time Series Analysis 1, 15 29.
Haykin, S. (1999), Neural Networks: A comprehensive foundation, Prentice Hall Upper Saddle
River, NJ.
Hosking, J.R.M. (1981) Fractional differencing, Biometrika, 68, 165176.
Karlaftis, M. G. and Papastavrou, J. D. (1998), Demand characteristics for charter air-travel,
International Journal of Transportation Economics, XXV(1), 19-35.
Kim, J., Wei, S. and Ruys, H. (2003), Segmenting the market of West Australian senior
tourists using an artificial neural network, Tourism Management, 24, 2534.
Law, R. (2000), Back-propagation learning in improving the accuracy of neural network-
based tourism demand forecasting, Tourism Management, 21, 331340.
Law, R. and Au, N. (1999), A neural networks model to forecast Japanese demand for travel
to Hong Kong, Tourism Management, 20(1), 8997.
Law, R., Mok, H. and Goh, C. (2007), Data Mining in Tourism Demand Analysis: A
Retrospective Analysis, Advanced Data Mining and Applications, Book Series
Lecture Notes in Computer Science, Springer Berlin / Heidelberg, 508 515.
M. Gen and R.W. Cheng, Genetic Algorithms and Engineering Optimization, John Wiley &
Sons, New York (2000).
Advanced Computational Approaches for Predicting Tourist Arrivals: the Case of Charter Air-Travel

317
Mitchell, M. (1998). An introduction to genetic algorithms. MIT Press, ISBN: 0262631857.
Song, H. and Li, G. (2008), Tourism demand modeling and forecastingA review of recent
research, Tourism Management, 29, 203220.
Vlahogianni, E. I. Karlaftis M. G. and Golias, J. C. (2005), Optimized and meta-optimized
neural networks for short-term traffic flow prediction: A genetic approach,
Transportation Research C, 13(3), 211-234.
Washington, S.P. Karlaftis M.G. and Mannering, F.L. (2003), Statistical and Econometric
Methods for Transportation Data Analysis, Chapman & Hall/CRC Press, London.
Webros, P. J. (1990), Backpropagation Through time: What it does and How to do it, IEEE
Proceedings, 78(10), 1550-1567.
World Tourism Organization UNWTO (2008). Firm Tourism Demand - Advanced
Results, World Tourism Barometer, June, accessed at: http://www.unwto.org/
media/news/en/press_det.php?id=2532.
Nonlinear Dynamics

318


Pooled Peak Period
p=1,q=1 p=1,q=1
Degree of differentiation d
m
0.24 0.46

1
0.66 0.02
AR polynomial coefficients

2
- -

1
0.52 0.55
MA polynomial coefficients

2
- -

3
-

4
-

5
-
Log-likelihood -2622.36 -1079.26
JB Test
Null: normality
2.02 1.42
Q
2
(7)
Null: serial independence
136.25** 66.18**
LM ARCH (1)
Null: no ARCH effect
1.41 1.32
* rejection at 5% significance level
** rejection at 1% significance level
Table 1. Estimation Results for the ARFIMA(p,d
m
,q) models for the Heraklion airport.

Pooled Peak Period
p=1,q=1 p=1,q=1
Degree of differentiation d
m
0.15 0.31

1
0.66 0.05 AR polynomial coefficients

2
- -

1
0.35 0.48 MA polynomial coefficients

2
- -

3
-

4
-

5
-
Log-likelihood -2588.14 -1002.80
JB Test
Null: normality
4.43 1.24
Q
2
(7)
Null: serial independence
145.25** 64.54**
LM ARCH (1)
Null: no ARCH effect
1.65 0.03
* rejection at 5% significance level
** rejection at 1% significance level
Table 2. Estimation Results for the ARFIMA(p,d
m
,q) models for the Kerkyra airport.
Advanced Computational Approaches for Predicting Tourist Arrivals: the Case of Charter Air-Travel

319

Pooled Peak Period
p=1,q=1 p=1,q=1
Degree of differentiation d
m
0.34 0.37

1
0.67 0.05 AR polynomial coefficients

2
- -

1
0.43 0.58 MA polynomial coefficients

2
- -

3
-

4
-

5
-
Log-likelihood -2689.31 -1017.48
JB Test
Null: normality
3.48 2.48
Q
2
(7)
Null: serial independence
122.52** 75.67**
LM ARCH (2)
Null: no ARCH effect
0.82 0.10
* rejection at 5% significance level
** rejection at 1% significance level
Table 3. Estimation Results for the ARFIMA(p,d
m
,q) models for Rhodes airport.

Specifications

DATA TRCVTE *: 60%-20%-20%

Structure
Input layer: Gamma memory (genetically optimized memory
depth)
1 hidden layer (genetically optimized number of hidden units h)

Learning Back-propagation
Chromosome [5, 15] , [0.01 - 0.1], [0.5 - 0.9], [1, 5], m [1, 12] h **
Fitness function Mean square error (cross-validation set)
Selection Roulette
Cross-over Two point (p=0.9) G
e
n
e
t
i
c

a
l
g
o
r
i
t
h
m

o
p
t
i
m
i
z
a
t
i
o
n

Mutation Probability p=0.09
* Training - Cross-validation - Testing
** h: neurons in hidden layer, : learning rate, : momentum, : time delay, m:dimension
Table 4. Data and neural network specifications for iterative short-term prediction.

Pooled NSI Arrivals

m
Heraklion 1 6
Kerkyra 1 4
Rhodes 1 4
Table 5. Estimates of the depth of the Gamma memory (parameters and m) of the
genetically-optimized TLNNs for the three cases.
Nonlinear Dynamics

320
Pooled Data Peak Demand Period
GA-TLNN*
Heraklion
Kerkyra
Rhodes
17%
26%
18%
2.8
3.4
3.2
ARFIMA
(average over cases tested)
37% 8.2
*genetically optimized TLNN
Table 6. Mean Absolute Percent Error of predictions using ARFIMA and genetically
optimized TLNN.




0
5
10
15
20
25
1
9
8
9
1
9
9
0
1
9
9
1
1
9
9
2
1
9
9
3
1
9
9
4
1
9
9
5
1
9
9
6
1
9
9
7
1
9
9
8
1
9
9
9
2
0
0
0
2
0
0
1
2
0
0
2
2
0
0
3
2
0
0
4
2
0
0
5
2
0
0
6
P
a
s
e
n
g
e
r
s
M
i
l
l
i
o
n
s
Total Arrivals NSI Arrivals SI Arrivals




Fig. 1. Yearly evolution of the total arrivals, non-scheduled international arrivals (NSI
Arrivals) and scheduled international arrivals (SI Arrivals) for the Greek airports.
Advanced Computational Approaches for Predicting Tourist Arrivals: the Case of Charter Air-Travel

321

0
1
1
2
2
3
3
1
9
8
9
1
9
9
0
1
9
9
1
1
9
9
2
1
9
9
3
1
9
9
4
1
9
9
5
1
9
9
6
1
9
9
7
1
9
9
8
1
9
9
9
2
0
0
0
2
0
0
1
2
0
0
2
2
0
0
3
2
0
0
4
2
0
0
5
2
0
0
6
P
a
s
s
e
n
g
e
r
s
M
i
l
l
i
o
n
s
NSI Arrivals
Total International Arrivals
Domestic Arrivals
Total Arrivals

0
5
10
15
20
25
30
35
40
45
50
1
9
8
9
1
9
9
0
1
9
9
1
1
9
9
2
1
9
9
3
1
9
9
4
1
9
9
5
1
9
9
6
1
9
9
7
1
9
9
8
1
9
9
9
2
0
0
0
2
0
0
1
2
0
0
2
2
0
0
3
2
0
0
4
2
0
0
5
2
0
0
6
F
l
i
g
h
t
s
T
h
o
u
s
a
n
d
s
NSI Flights
International Flights
Domestic Flights
Total Flights

Heraklion (Crete)
0
0
0
1
1
1
1
1
9
8
9
1
9
9
0
1
9
9
1
1
9
9
2
1
9
9
3
1
9
9
4
1
9
9
5
1
9
9
6
1
9
9
7
1
9
9
8
1
9
9
9
2
0
0
0
2
0
0
1
2
0
0
2
2
0
0
3
2
0
0
4
2
0
0
5
2
0
0
6
P
a
s
s
e
n
g
e
r
s
M
i
l
l
i
o
n
s
NSI Arrivals
Total International Arrivals
Domestic Arrivals
Total Arrivals

0
2
4
6
8
10
12
14
16
18
1
9
8
9
1
9
9
0
1
9
9
1
1
9
9
2
1
9
9
3
1
9
9
4
1
9
9
5
1
9
9
6
1
9
9
7
1
9
9
8
1
9
9
9
2
0
0
0
2
0
0
1
2
0
0
2
2
0
0
3
2
0
0
4
2
0
0
5
2
0
0
6
F
l
i
g
h
t
s
T
h
o
u
s
a
n
d
s
NSI Flights
International Flights
Domestic Flights
Total Flights

Kerkyra
0
1
1
2
2
1
9
8
9
1
9
9
0
1
9
9
1
1
9
9
2
1
9
9
3
1
9
9
4
1
9
9
5
1
9
9
6
1
9
9
7
1
9
9
8
1
9
9
9
2
0
0
0
2
0
0
1
2
0
0
2
2
0
0
3
2
0
0
4
2
0
0
5
2
0
0
6
P
a
s
s
e
n
g
e
r
s
M
i
l
l
i
o
n
s
NSI Arrivals
Total International Arrivals
Domestic Arrivals
Total Arrivals

0
5
10
15
20
25
30
35
1
9
8
9
1
9
9
0
1
9
9
1
1
9
9
2
1
9
9
3
1
9
9
4
1
9
9
5
1
9
9
6
1
9
9
7
1
9
9
8
1
9
9
9
2
0
0
0
2
0
0
1
2
0
0
2
2
0
0
3
2
0
0
4
2
0
0
5
2
0
0
6
F
l
i
g
h
t
s
T
h
o
u
s
a
n
d
s
NSI Flights
International Flights
Domestic Flights
Total Flights

Rhodes
Fig. 2. Evolution of arrivals (passengers per year) and flights per year for the period of 1999-
2006.
Nonlinear Dynamics

322
39
8214
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
10000
Jan Feb Mar Apr May Jun Jul Aug Sept Oct Nov Dec
N
S
I

A
r
r
i
v
a
l
s
H
u
n
d
r
e
d
s
1999
2002
2004
2006
Average of 1999-2006 (%)

Heraklion (Crete)
5
1935
0
500
1,000
1,500
2,000
2,500
Jan Feb Mar Apr May Jun Jul Aug Sept Oct Nov Dec
N
S
I

A
r
r
i
v
a
l
s
H
u
n
d
r
e
d
s
1999
2002
2004
2006
Average of 1999-2006 (%)

Kerkyra
9
2646
0
500
1,000
1,500
2,000
2,500
3,000
3,500
Jan Feb Mar Apr May Jun Jul Aug Sept Oct Nov Dec
N
S
I

A
r
r
i
v
a
l
s
H
u
n
d
r
e
d
s
1999
2002
2004
2006
Average of 1999-2006 (%)

Rhodes
Fig. 3. Monthly variation of non-scheduled international arrivals in Rhodes for the period
between 1999 and 2006.
Advanced Computational Approaches for Predicting Tourist Arrivals: the Case of Charter Air-Travel

323
R = 0.7764
-100000
0
100000
200000
300000
400000
500000
0 100000 200000 300000 400000 500000
P
r
e
i
d
c
t
e
d

N
S
I

A
r
r
i
v
a
l
s
NSI Arrivals
Heraklion

R = 0.6784
0
50000
100000
150000
200000
250000
0 50000 100000 150000 200000 250000
P
r
e
i
d
c
t
e
d

N
S
I

A
r
r
i
v
a
l
s
NSI Arrivals
Kerkyra

R = 0.7894
-100000
-50000
0
50000
100000
150000
200000
250000
300000
350000
0 50000 100000 150000 200000 250000 300000 350000
P
r
e
i
d
c
t
e
d

N
S
I

A
r
r
i
v
a
l
s
NSI Arrivals
Rhodes

Fig. 4. Scatter plots of actual versus predicted values of NSI arrivals for the three airports.
Nonlinear Dynamics

324
0
100
200
300
400
500
600
700
J
u
l
-
0
5
A
u
g
-
0
5
S
e
p
-
0
5
O
c
t
-
0
5
N
o
v
-
0
5
D
e
c
-
0
5
J
a
n
-
0
6
F
e
b
-
0
6
M
a
r
-
0
6
A
p
r
-
0
6
M
a
y
-
0
6
J
u
n
-
0
6
J
u
l
-
0
6
A
u
g
-
0
6
S
e
p
-
0
6
O
c
t
-
0
6
N
o
v
-
0
6
T
h
o
u
s
a
n
d
s
NSI Arrivals
Predicted NSI Arrivals (TLNN)
Predicted NSI Arrivals (ARFIMA)


0
500
1000
1500
2000
2500
3000
Nov-05 Dec-05 Jan-06 Feb-06 Mar-06
NSI Arrivals
Predicted NSI Arrivals (TLNN)
Predicted NSI Arrivals (ARFIMA)

Fig. 5. Predictions using the ARFIMA and genetically optimized TLNN. Results from the
three case study airports are aggregated both for ARFIMA and TLNN.
14
A Nonlinear Dynamics Approach
for Urban Water Resources Demand
Forecasting and Planning
Xuehua Zhang, Hongwei Zhang and Baoan Zhang
1
Department of Environmental Economics, Tianjin Polytechnic University
2
School of Environmental Science and Technology, Tianjin University
China
1. Introduction
Over the past decades, controversial and conflict-laden water allocation issues among
competing domestic, industrial and agricultural water use as well as urban environmental
flows have raised increasing concerns (Huang & Chang, 2003); Particularly, Such
competition has been exacerbated by the growing population, rapidly economic growth,
deteriorating quality of water resources, and shrinking water availability due to a number of
natural and human-induced impacts. A sounding strategy for water resources allocation
and management can help to reduce or avoid the losses which are caused by water
resources scarcity. However, in the water management system, many components and their
interactions are uncertain. Such uncertainties could be multiplied not only by fasting
changes of socioeconomic boundary conditions but also by unpredictable extreme weather
events which caused by climate change. Thus, water resources management should be able
to deal with all challenges above. Therefore, an effective integrated approach is desired for
urban water adaptive management.
Many methods, such as stochastic, fuzzy, and interval-parameter programming techniques,
have been employed to counteract uncertainties in different fields of water management and
have made great progresses in managing uncertainties in model scale. Water resource is an
integral part of the socio-economic-environmental (SEE) system, which is a complex system
dominated by human. In order to reach a sounding decision, it is necessary for decision-
makers to obtain a better understanding of the significant factors that shape the urban and
the way the water resources system reacting to certain policy. Therefore, study of
sustainable water resource management should be based on general system theory that
addresses dynamic interactions amongst the related social-economic, environmental, and
institutional factors as well as non-linearity and multi-loop feedbacks.
System dynamics (SD) aims at solving of complex systems problems by simulating
development trends of the system and identifying the interrelations of each factor in the
system. This will help to explore the hidden mechanism and thus improve the performance
of the whole system. Hence, after proposed by W. Forrester (Forrester, 1968), SD model has
been widely used in global, national, and regional scales for sustainability assessment and
system development programme (Meadows 1973; Mashayekhi, 1990; Saeed, 1994). Due to
Nonlinear Dynamics

326
the complexity of problems in the water system, the use of dynamic simulation models in
water management has a long tradition (Biswas 1976; Roberts et al., 1983; Abbott and
Stanley, 1999; Ahmad & Simonovic, 2004). The development journaey of several sections of
applying system dynamics as a tool for integrated water management system analysis can
be traced as from focusing on water system itself, to having a strong economic examinations
on feedback relationships between industry and water availability, and then to having
interaction with population growth (Liu et al., 2007). The above development make SD
model has the flexibility and capability to support deliberative-analytical processes
effectively. Meanwhile, SD and Multi Objective Programme (MOP) integrated model as an
extension of the previous SD applications has been presented and used in urban water
management in recent years, which takes into account both optimization and simulation
(Guo et al, 1999; Zhang & Guo, 2002). This chapter will introduce a nonlinear dynamics
approach for urban water resources demand forecasting and planning based on SD-MOP
integrated model.
2. Uncertainties in Urban water system
2.1 Urban water system analysis
Generally, urban water system could be divided into four subsystems, i.e., social subsystem,
economic subsystem, environmental subsystem and water resources subsystem. The
relationships and interactions are complicate, as Fig. 1.


Fig. 1. Urban water management subsystems and relations
2.2 Uncertainties of urban water management system analysis
Urban water resources demand forecasting and planning are two important parts of urban
water integrated management. Commonly, integrated water management should provide a
framework for integrated decision-making and could be consists of system analysis, action
results forecast, planning formulate and implementation, and evaluation and monitoring the
goals and effects of implementation. At the system analysis stage, information collection and
investigation are the basic work. A system structure is built based on a careful consideration
of interactions among factors and subsystems. Long-term and short-term goals, problems,
and priority focused will then are identified with both experts and stakeholders take part in.
At the forecast stage, simulation model and evaluation model will be set up. Fixing on
parameters and variable values of models and listing alternative solutions are the key
process of the stage, based on field investigation, literature review and interviews with local
stakeholders. Then according to the simulation and evaluation results of the alternatives, the
selected solution can be identified and the corresponding desired actions can be determined.
Urban flow
Consumption
Labor
Environmental
capacity
Wastewater discharge
Production flow Wastewater discharge
Municipal flow
Water resources subsystem
Environmental subsystem
Economic subsystem
Social subsystem
A Nonlinear Dynamics Approach for Urban Water Resources Demand Forecasting and Planning

327
Implementation and re-evaluation cant be separated completely. Management and re-
evaluation is the mechanism that improves management goals and practices constantly.
Uncertainties limit the forecasting ability of and thus influence the quality of decision
making. They can be categorized into four types : (1) intransience uncertainties caused by
fasting changes of urban socioeconomic conditions; (2) external uncertainties caused by the
stress of factors beyond the urban boundary (Liu, 2007); (3) uncertainties associated with
raw data and model parameters driven from outdated or absent issues news, events, or
statistic data; and (4) uncertainties arising from multiple frames (e.g. peoples cognizing/
perceiving technique/ability advance, world and ethical view change) (Jamieson, 1996;
Pahl-Wostl, 2009). The above uncertainties are associated with all four stages, the details as
Fig. 2.


Fig. 2. The uncertainties in urban water management system
We can find that all above uncertainties are raised from the cognitive dimension (e.g. limited
understanding system behavior and interactions among composing factors, uncertainty
from fasting changes of socioeconomic conditions and change of natural conditions) and
technical dimension (e. g. outdated or absent issues news/events/data, absent specific to
techniques and countermeasures, limited of forecasting method) two aspects.
2.3 Overlook of counteracting measures to water system uncertainties
Whether we recognize it or not, socioeconomic laws and the natural laws are located in the
objective world. So we can say that uncertainty is raised from the limitations of human
cognition. Due to human cognitive abilities change, their understanding of the current
world and their forecast of the future world will change over time. Furthermore, SEE system
System analysis
Forecasting and
planning
Implement and Re-
evaluation
Outdated or absent
informations
Limited
understanding
system behavior and
interactions
Absent specific to
techniques and
countermeasures
Limited recognizing
priorities focused and
key problems
Uncertainty of system
structure
Uncertainty of and
parameter and
variable value
Uncertainty of
alternative solutions
Limitation of forecast
methods
U
n
c
e
r
t
a
i
n
t
y

o
f

f
o
r
e
c
a
s
t

r
e
s
u
l
t
s
Uncertainty from fasting changes of socioeconomic conditions and natural conditions
Uncertainty of
selected solution
Uncertainty of
desired action
Uncertainty of
available techniques
and countermeasures
U
n
c
e
r
t
a
i
n
t
y

o
f

I
m
p
l
e
m
e
n
t

a
n
d

r
e
-
e
v
a
l
u
a
t
i
o
n

r
e
s
u
l
t
s
Uncertainty of
external driving
force (e.g. global
change in general
and climate
change in
particular)
Nonlinear Dynamics

328
1
is a complexity system reflecting the mutual and complicated functions amongst the internal
elements, which can be characterized by the complicated system structure properties far
from balance status and with dissipation structures, as well as the behaviors of which the
input-output response shows uncertainty that beyond peoples experiential and qualitative
cognition. We can be in virtue of SD model as well as interactions between modelers and
stakeholders to interact the behavior uncertain from input-output response. The SD model
can be run by different scenarios, and thus the optimal scenario can be selected by the
analyses and discussions.
However, simulation model could be run in almost limitless scenarios according SEE
complex system parameters changed in different policies. Thus it is difficult to simulate all
possible scenarios constrained in time and fund. So it is difficult to ensure the optimal level
of selected scenarios and its corresponding programme design. Therefore, SD-MOP
integrated model (Zhang & Guo, 2002) is proposed to counteracts uncertainties with SD
model applying in different scenarios simulation and analysis, and MOP model applying in
optimization.
3. System dynamics model
3.1 The basic concepts of SD
The SD model takes certain steps along the time axis in the simulation process. At the end of
each step, the system variables denoting the state of the system are updated to represent the
consequences resulting from the previous simulation step. Initial conditions are needed for
the first time step. Variables representing flows of information and initials, arising as results
of system activities and producing the related consequences are named as level variables
described as in the flow diagram, and rate variables described as . Auxiliary
variable means the detailed steps by which information associated with current levels are
transformed into rates to bring about future changes. In addition, the symbol
represents the sinks or sources.
Fig. 3 is a sample flow diagram for the total population, in which the total population (TP) is
a level variable; birth population (BP), death population (DP), and net migrated population
(NP) are rate variables; and birth rate (BR), death rate (DR), and net migration rate (NR) are
auxiliary variables.
P
BP DP
BR
DR
NP
NMR

Fig. 3. SD flow chart of population subsystem
In SD level equation, three time points are denoted as Jpast, K (present), and L (future).
The step from J to K is referred to as JK and that from K to L as KL. The duration period
A Nonlinear Dynamics Approach for Urban Water Resources Demand Forecasting and Planning

329
between successive points is named DT. Therefore, a level variable could be referred to as
LEVEL.J, LEVEL.K, or LEVEL.L at a time pointRATE.JK and RATE.KL will function in the
duration period. We can express:
LEVEL.K=LEVEL.J+DT*RATE.JK
3.2 The procedures for applying SD model to simulate target system behavior
The proedures for applying SD model to simulate target system behavior can be
summarized into three steps.
(1) Construction SD model
The first step of the procedures is constructing SD model through analyses of the total
system, and identifying the model validity by historical examination, and sensitivity
analysis. Accordingly, parameters and relevance can be modified and confirmed.
(2) Validity examination
Validity examination examination includes direct observation, historical examination, and
sensitivity analyses. Direct observation is through SD model run, if there is no obviouse
unreasonable simulation results, we can to the historical examination.
Historical examination is checking the error between simulation and reality. The errors of
main forecasting level variables are accepted is one of the requirements of SD model being
used in reality system.
Another requirement is that the target system responds in lower degree sensitivity to most
of the parameters through a series of sensitivity analyses conducted to examine the systems
responses to variations of input parameters and/or their combinations. A concept of
sensitivity degree is defined as follows:

( ) ( )
( ) ( )
t t
Q
t t
Q X
S
Q X

(1)
where t is time; Q
(t)
denotes system state at time t; X
(t)
represents system parameter affecting
the system state at time t; S
Q
is sensitivity degree of state Q to parameter X; and Q
(t)
and
X
(t)
denote increments of state Q and parameter X at time t, respectively.
For the n state variables (Q
1
, Q
2
,, Q
n
), the general sensitivity degree of a parameter at time
t can be defined as follows:

1
1
.
i
n
Q
i
S S
n
=
=

(2)
Where n denotes a number of state variables; S
Q
is sensitivity degree of state Q
i
; and S is
general sensitivity degree of the n states to the parameter X.
If there are some departures from the model validity requirement standards, the SD model
should be adjusted until fix to the standards. Then, SD model could be used in target system
behavior simulation.
3.3 SD model validity in simulating nonlinear feedback mechanism
Although SD equations are linearity, they simulating in computer can describe nonlinear
characteristics produced by multi-feedback when consider temporal dynamic affection.
Nonlinear Dynamics

330
Figure 4 is a piece of water resources subsystem delay feedback circle- water supply
capacity building flow chart, which included two simple first-order delay feedbacks.
plan time, Pt
water demand, Wd
delay flow,
Df(t)
delay time, Dt
water supply
capacity, Ws(t)
+
plan for transfer water
from other area, Wr(t)
+
water transfer project
building, Wbr(t)
-

---

Fig. 4. Water supply capacity flow chart
Plan for transfer water from other area (Wr (t)) expressionin which had a first order delay,
was shown as the basic divided differences formula: ( ) ( ( )) / Wr t Wd Ws t Pt = .
Due to delay time to implement from confirming water transfer scheme to water supply
formation, water transfer project building (Wbr(t)) could be expressed as a simple first order
mater delay function: ( ) ( ) / Wbr t Df t Dt = .
As known, initialization of Df (t) is A m
3
, initialization of Ws (t) is B

m
3
, Wd = C m
3
, Pt = a,
Dt = b. According the above conditions can be established equations (3):

0
0
( ) ( ( )) /
( ) ( ) /
C
a
b
( )| A
( )| B
t
t
Wr t Wd Ws t Pt
Wbr t Df t Dt
Wd
Pt
Dt
Df t
Ws t
=
=
=

(3)
Confluence rate was the derivative of the flow to time t. Hereby, it could be obtained the
corresponding differential equations (4).

(4-1)
(4-2)
(4-3)
(4-4)
'
'
( ) 1
( ) (C ( ))
a b
(0) A
( )
( )
b
(0) B
Df t
Df t Ws t
Df
Df t
Ws t
Ws

(4)
A Nonlinear Dynamics Approach for Urban Water Resources Demand Forecasting and Planning

331
By equations (4), it could be derived the expression of flow, and the following equation
could be obtained by on both sides of equation (4-3) of equations (4) derivation.

1
"( ) '( ) ( ) C
2
Ws t Ws t Ws t + + = (5)
Solve equation (5), the curve of water supply capacity, the curve of the delay flow , the curve
of the plan rate, and the curve of project building could be derivate. Thus the results is
according follow three conditions.
1. Condition 1
when
a
b>
4
,
2
1 4
0
b ab
< , then
1, 2
1 1 4b
1
2b 2b a
i =
The solution of the equation (5) corresponding homogeneous equation is shown as:

1
2b
( ) 1 2
1 4b 1 4b
( cos sin )
2b a 2b a
t
s t
W e C t C t

= + (6)
Seeking the special solution of equation (5):

*
( ) s t
W C = (7)
According to equation (6) and (7), we can obtain the general solution of equation (5), which
is shown as equation (8).

1
2b
( ) 1 2
1 4b 1 4b
( cos sin
2b a 2b a
t
s t
W e C t C t C

= + + (8)
W
s(0)
=B will be into the equation (8). Then,
1
C B C = + ,
1
B C C =
From,

1
'
2b
( ) 2
1
2 2b
1 1 4b 1 4b
((B C)cos 1 sin 1 )
2b 2b a 2b a
C B 4b 1 4b 4b 1 4b
( 1 sin 1 1 cos 1 )
2b a 2b a 2b a 2b a
t
s t
t
W e t C t
C
e t t

= +

+ +
(9)
(0) (0)
A
'
b
s f
W D = = is into the equation (9). Then,
2
A 1 1 4b
(B C) 1
b 2b 2b a
C = + ,
2
2A B C
4b
1
a
C
+
=

According to the above, the special solution of equation (5) is shown as the follow:

1
2b
( )
1 4b 2A B C 1 4b
((B C)cos 1 sin 1 ) C
2b a 2b a 4b
1
a
t
s t
W e t t

+
= + +

(10)
Nonlinear Dynamics

332
The equation (10) is the curve of the Water supply capacity.
From:
( ) ( )
'
f t s t
D bW = , then the curve of the delay flow can be obtained as equation (11):

1
2b
( )
2b
A (B C)
1 4b 1 4b
a
(Acos 1 sin 1 )
2b a 2b a 4b
1
a
t
f t
D e t t

+
=

(11)
The curve of plan for transfer water from other area can also be obtained as equation (12):

1
2b
( )
1 4b 2A B C 1 4b
((B C)cos 1 sin 1 ) C
2b a 2b a 4b
1
a
t
r t
W e t t

+
= + +

(12)
The curve of project building can also be obtained as equation (13):

1
( )
2b
( )
2b
A (B C)
A 1 4b 1 4b
a
( cos 1 sin 1 )
b b 2b a 2b a 4b
b 1
a
t
f t
br t
D
W e t t

+
= =

(13)
2. Condition 2
When
a
b
4
=
2
1 4
0
b ab
= , Then:
1 2
1
2b
= =
The general solution of equation (5) is shown as the follow:

1
2b
( ) 1 2
( ) C
t
s t
W e C C t

= + + (14)
W
s(0)
=B will be into equation (14). Then,
1
B C C = + ,
1
B C C =
From,

1 1
2 2
( ) 2 2
1
' [(B-C) ]
2
t t
b b
s t
W e C t C e
b

= + +
(15)
(0)
(0)
A
'
f
s
D
W
b b
= = will be into the equation (15). Then,
2
A 1
(B C)
2
C
b b
= + ,
2
2A+B-C
2
C
b
=
According to the above, the special solution of equation (5) is shown as the follow:

1
2b
( )
2A B C
((B C) ) C
2b
t
s t
W e t

+
= + + (16)
The equation (16) is the curve of the water supply capacity.
A Nonlinear Dynamics Approach for Urban Water Resources Demand Forecasting and Planning

333
From
( ) ( )
b '
f t s t
D W = , then

1 1
2b 2b
( )
1 2A B C 2A B C
((B C) )
2 2b 2b
t t
f t
D e t e

+ +
= + + (17)

The equation (17) is the curve of the delay flow.
The curve of the water transfer rate can be obtained as (18) and the rate curve of Building
water supply facilities can be obtained as (19).

1
2b
( ) ( )
C-B 2A+B-C
(C ) / ( )
a 2ab
t
r t s t
W W a e t

= = (18)

1 1
( )
2b 2b
( ) 2
1 2A B C 2A B C
((B C) )
b 2b 2b 2b
t t
f t
br t
D
W e t e

+ +
= = + + (19)
3. Condition 3
when
a
b
4
< ,
2
1 4
0
b ab
> , Then,
1, 2
1 1 4b
1
2b 2b a
=

1 1 4b 1 1 4b
( 1 ) ( 1 )
2b 2b a 2b 2b a
( ) 1 2
C
t t
s t
W C e C e
+
= + + (20)
W
s(0
)=B will be into the equation (20). Then
1 2
C B C C + + = ,
1 2
B C C C =
From

1 1 4b 1 1 4b
( 1 ) ( 1 )
2b 2b a 2b 2b a
( ) 1 2
1 1 4b 1 1 4b
' ( 1 ) ( 1 )
2b 2b a 2b 2b a
t t
s t
W C e C e
+ +
= + + + (21)
( )
(0)
A
'
b b
f t
s
D
W = = will be into the equation (21). Then

1 2
A 1 1 4b 1 1 4b
( 1 ) ( 1 )
b 2b 2b a 2b 2b a
C C = + + (23)
Because
1 2
B C C C =

So,
1
4b
( 1 1)(B C) 2A
a
4b
2 1
a
C
+ +
=

,
2
4b
( 1 1)(B C) 2A
a
4b
2 1
a
C

=

According to the above, the special solution of equation (5) is shown as the follow:
Nonlinear Dynamics

334

1 1 4b
( 1 )
2b 2b a
( )
1 1 4b
( 1 )
2b 2b a
4b
( 1 1)(B-C)+2A
a
4b
2 1-
a
4b
( 1 1)(B-C)+2A
a
C
4b
2 1-
a
t
s t
t
W e
e
+
+
+
=

+ +
(22)
The equation (22) is the curve of the Water supply capacity.
From
( ) ( )
b '
f t s t
D W = , then

1 1 4b 1 1 4b
( 1- ) ( 1- )
2b 2b a 2b 2b a
( )
4b 4b 4b 4b
(C-B)+2A(-1+ 1- ) (C-B)+2A(1+ 1- )
a a a a
4b 4b
4 1- 4 1-
a a
t t
f t
D e e
+
= + (23)
The equation (23) is the curve of the delay flow.
And the curve of the water transfer rate can be obtained as (24).

( )
( )
1 1 4b 1 1 4b
( 1 ) ( 1 )
2b 2b a 2b 2b a
(C )
a
4b 4b
( 1 1)(C-B) 2A ( 1 1)(C-B) 2A
a a
4b 4b
2 1 2 1
a a
s t
r t
t t
W
W
e e
a a
+

=
+ +
= +

(24)
The rate curve of Building water supply facilities can be obtained as (25).

1 1 4b
( 1 )
( )
2b 2b a
( )
1 1 4b
( 1 )
2b 2b a
4b 4b
(C B)+2A( 1+ 1 )
a a
b 4b
4b 1
a
4b 4b
(B C)+2A(1+ 1 )
a a
4b
4b 1
a
t
f t
br t
t
D
W e
e
+


= =

(25)
From above deduction, we can know that although SD equations are linearity, they
simulating in computer can describe nonlinear characteristics produced by multi-feedback
when consider temporal dynamic affection.
4. Decision-making system based on SD-MOP integrated model for urban
water resources demand forecasting and planning
From above analysis, we can know that urban water resources demand forecasting is the
key procedure of urban water system management. In different scenarios, the forecasting
A Nonlinear Dynamics Approach for Urban Water Resources Demand Forecasting and Planning

335
outcomes may be different, which result in different corresponding planning. From above
deduction, we can also get the conclusion that SD model can be applying to simulate
nonlinear and complex system behavior though the basic equations are linear and simple.
Hence, we introduce a decision-making system which core in SD-MOP integrated model for
urban water resources demand forecasting and planning. The procedure of applying SD-
MOP integrated model as Fig.5.


Fig. 5. The procedure of SD-MOP integrated model applying
In SD-MOP integrated model, SD is used for water resources system dynamics nonlinear
behavior simulation, and MOP is used for optimal policy choice and optimal design
forming.
4.1 Setting up SD model
The first step of SD-MOP applying is constructing SD model based on information
collection system analysis. The procedures of constructing SD model are the follows:
1. identify the boundary of SD model;
2. classify sub-systems of urban water system;
3. determine the set of main level variables;
4. analysis the realtions of system parameters and variable;
5. design the flow diagram;
6. determine the basic value of parameters by mathmatic forecasting method both in
statistical method and experience according to current and historical imformation of the
target system;
7. set up basic mathmatic equations which consist of SD model;
8. test SD model validity and adjust it accoding testing results until it can be used in
realistic system simulation.
System analysis Identify the boundaryof the model
Information collect and analysis Set up SD model
IPV indentify
Set up and adjust assistant model
Solve MOP model and
Identify the value of IPV
Different scenarios
simulation and analysis
Exchange with stakholders
Planning design
Satisfied
Y
N
Interview with stakeholders
Set up MOP model
Nonlinear Dynamics

336
4.2 Analyzing IPV
Analyzing the sensitivity by sensitivity test and original run (run in the condition which the
system keep current behavior and tendency without any policy adjustment), the sensible
parameters and the closed relating variables can be identified, which are named as IPV
(Important Parameter and Variable).
IPV aggregation includes controllable factors and non-controllable two types. Non-control
lable factors can become system development neck, while adapting controllable factors in
suitable way could exploit urban development.
4.3 Setting up MOP model
Running the SD model based on the current situations (called original run). The gap
between the original run results and ideal level of the system can be identified. In order to
obtain optimal programme design and adjust the system function and behavior, MOP
model cored in IPV is set up. In the MOP model the controllable factors of IPV become
decison variables and non-controllable factors of IPV become constrains, while some level
variable which closely related to IPV become maximum or minimum aim.
General format of MOP model as follow:
max ( ) /
k k
f x (26)
s.t. ( ) ,
i i i
g x b (27)
0,
j j
x x x (28)
Where, x is decision variable (a set of real number in a closed boundary limit and is the
value of IPV or value of variable that are related to IPV), equation (26) is objective function,
(27) and (28) are the limiting conditions.
4.4 Setting up assistant model to solve MOP
Applying ODTL (Objective Deviation Tolerance Level) method (Zhou, 1998) to solve MOP
model. Here, there is some different from Zhou in interview process. First, we determin
each goal ODTL by interview with stakeholder based on giving them original run results
and the ideal goals. Second, the decision is not finished in one time, but in several times
based on showing them the former scenarios SD model simulation results which
corresponding to their choice of each goals ODTL, and the stake holders can adjust there
decision by comparing and discussing the former results. Finally, the optimal IPV can be
determined by several adjust assistant model, solve MOP, simulation corresponding system
tendency, and compare and selecte the desirable scenario.
4.5 Planning
Based on the optimum values of IPV, the proposals for running the model can be designed.
Accordingly the final plan proposal can be formulated.
5. Case study
Applying SD-MOP integrated model in a real urban system to test its validity [Zhang 2010].
A Nonlinear Dynamics Approach for Urban Water Resources Demand Forecasting and Planning

337
The boundary of the target system is the urban area of Qinhuangdao, which is a city of
Hebei province, located at latitude 3922-4037N and longitude 11833-11951E, and
covers an area of 7,812 km
2
. Qinhuangdao has jurisdiction over three districts (Shanhaiguan,
Beidahe, and Haibin) and four countries (Lulong, Qinglong, Funing, and Changli). The
annual rainfall in Qinhuangdao is about 670mm, with the water resource per capita in
Qinhuangdao is 600m
3
/a, which is 1/4 the average level in China. The system is composed
of population subsystem, industry subsystem, services subsystem, water supply subsystem
and water environmental protection subsystem. The planned period is 15 years (2006 -
2020). It is divided into two phases, i.e., 2006-2010 and 2010 - 2020. The base year is 2000.
5.1. Constructing SD model
Based on the analysis of the target system, SD model of Qinhuangdao (QHDWSD) can be
constructed, and thus the sensibility of the model can also be tested. There are more than
110 variables in SD model, in which there are more than 110 system dynamic equations. Fig.
6 is the flow chart of QHDWSD.


TP
BP
DP
NM
BR
DR
NMR
AP
UR
SIGDP
IVSIGDP
IRSIGDP
CA
CVP
RCAP
TWD
EWDDL
EWDDLC
EWDDLOP
SIWD
PSIGDPWC
AWD
PCPWC
TWRS
WRBF
APW
TADSC
TIWWD
DWWAT
RRSIWWD
IWWAT
DRW
WPI
EWD
TAWW
TWAT
AWWD
TIGDP
IVTIGDP
IRTIGDP
TIWD
PTIGDPWC
EFPCI
EFE
EFP
DWCC
QIWWD
AWWDC
MC
<Time>
NMC
TFNM
<Time>
TBB
<Time>
ACB
URF
<Time> MEUR
<EWDDLC>
PCADCODD
ADPCIGDP
TCODDA
DACODWP
<SIGDP>
WTRTF
PF
NAP
<NAP>
GDP
IVGDP
IRGDP
Per capita GDP
PCGA
GL
BUAGBCR
CTGSR
PGSP
<NAP>
<Time>
GWRSF
BUAC
CABUAS
CRBUAS
WDPMYGDP
<WRBF>
<WRBF>
<WRBF>
<WRBF>
DACOD MPDACOD
CODPI
<Per capita GDP>
<NAP>
<TIGDP>
<PCGA>
<EWD>
WRSDBI
<PF>
PCLA
ACPCHUA
EFPCHUA
RCPCHUA
FRCPCHUSA
PSIWDTF
DSCOD
RWRSTF
CTFEWD
<Time>
TCRBUA
<Time>
<SIGDP>
<WRBF>
TFIRTIGDP
<Time>
<Time>
TRRSIWWD
<Time>
TFADPCIGDP
TFDSCOD
<Time>
TFMPDACOD
IRGDP TF
TFDR
TOWUP
TEWDDLPCC
<Time>
PTIGDPWD
<Time>
TFIRSIGDP
WUPC
OWUP
TFWUPC
<Time>
OWD
EWDDLPCC
EWDDLPCOP


Fig. 6. QHDSD diagram
Nonlinear Dynamics

338
5.2 Identifying IPV
Based on original running and putting eight variables and fourteen parameters into
sensitivity analysis, IPV were identified. Those are: Increase rate of second industrial GDP,
per second industrial GDP water consumption, Per capita plow land water consumption.
5.3 Setting up and solving MOP model
In the original simulation, when GDP getting in the aim scale water resource supply and
demand balance index (water available supply to human social and economic activities
divided by water demand human social and economic activities) will be lower than 0.6 in
2020 (Fig. 2). The consequence will be that people actives water consumption invade and
occupy eco-environmental share and lead to water ecosystem quality degradation and water
resource sustainable supply capability decrease. According above analysis, the key issue is
the structure of the economic, thus MOP model is setting up as follow.

3
1
1
Z ( ) max
i
i
X X
=
=

(29)

3
2
1
Z ( ) min q
i i
i
X X
=
=

(30)

3
1
q Q
i i
i
X
=

(31)

3
1
Ymin /( ) Ymax
i i i i
i
X X
=

(32)
0
i
X (33)
where: X
i
=GDP of three industry (10
8
); q
i
=per GDP water consumption of three industry
(t/10
8
); Q=water resource amount could be supplied to human economic activities (t);
Ymin
i
= the lower bound of three industry proportion in total GDP; Ymax
i
= the higher
bound of three industry proportion in total GDP. Then set up assitant model and solved it
based on interaction with stakeholders who consists of the staff of water resources bureau,
the staff of the environmental protection agency, the staff of regional development and
reform Commission the staff of related bureaus, the staff of water supply and wastewater
treatment firms, delegates of the three industries, and representatives from the public.
5.4 Obtaining relative optimal programme
According IPV solution, the optimal design could be obtained and the corresponding water
resources plan of Qinhuangdao city was formulated. Table 1 shows the comparison of
different industry ratio in the total gross domestic production (GDP) respectively between
optimal solutions and original tendency. The comparison results of the water supply-
demand balance, GDP, population scale and water pollution index between the feasible
programme simulations with the original simulation as Fig. 7.
A Nonlinear Dynamics Approach for Urban Water Resources Demand Forecasting and Planning

339


year item
Primary industry
(%)
Secondary
industry
(%)
Tertiary industry
(%)
optimal designs 62 356 457.5
2010
original tendency 65 370 440.5
optimal designs 102 1220 1397.2
2020
original tendency 107 1256 1357.2

Table 1. Industrial structure (different industry ratio in GDP)


b
250
260
270
280
290
300
310
320
330
340
2005 2008 2011 2014 2017 2020
year
p
o
p
u
l
a
t
i
o
n
(
1
0
4
)
optimal
original
a
0
500
1000
1500
2000
2500
3000
3500
2005 2008 2011 2014 2017 2020
year
R
e
g
i
o
n
a
l

G
D
P
(
1
0
8

)
optimal
original
d
0
0.2
0.4
0.6
0.8
1
1.2
2005 2008 2011 2014 2017 2020
year
w
a
t
e
r

S
-
B

i
n
d
e
x
(
/
)
optimal
original
c
0
0.2
0.4
0.6
0.8
1
1.2
1.4
2005 2008 2011 2014 2017 2020
year
W
P

i
n
d
e
x
(
/
)
optimal
original


Fig. 7. Main level variable comparing between optimal design and original tendency
Nonlinear Dynamics

340
In Fig. 7, sub Fig-a is for gross domestic production, sub Fig-b is for total population scale,
sub Fig-c is for water pollution index (WPI-the ratio of simulating year water contamination
discharge to base year water contamination discharge ) contamination, and sub Fig-d is for
water resources supply-demand balance index (WRSDBI-the ratio of water supply quantity
to water demand quantity).
Fig. 7 and table 1 indicate that through adjusting system structure can realize water
sustainable utilization while not decreasing the speed of economic development. The water
resource strategy plan is based on nonlinear dynamics forecasting approach for water
resource demand.
5.5 Nonlinear dynamics approach validity test in practice
Follow is an example of Qinhuangdao water resource plan of 2000 to 2005. And it was
researched by our group during 1998 to 2000. In the plan, we used two methods, nonlinear
method and trend extending method, to forecast urban water resources demand. Fig. 8
shows the comparative errors for forecasting data and actual data between SD nonlinear
method and trend extending method. From Fig. 8, we can know that nonlinear forecasting
is more accurate with can give support to water resources plan.



0. 0
5. 0
10. 0
15. 0
20. 0
25. 0
30. 0
35. 0
2
0
0
1
2
0
0
2
2
0
0
3
2
0
0
4
2
0
0
5
a
p
r
e
d
i
c
t
i
o
n

e
r
r
o
r

(
%
)
SD nonl i ear met hod
t r end ext endi ng met hod



Fig. 8. The comparative analysis results
6. Conclusion
From above study, we can get the conclusion : (i) complex system analysis and nonlinear
dynamics simulation are very useful for urban water resource demand forecasting and
planning, (ii) the integrated model of SD-MOP can avoid the randomness of proposal
designed by experiences of planners and decision-makers, which results in the generated
planning proposal has high reliability.
A Nonlinear Dynamics Approach for Urban Water Resources Demand Forecasting and Planning

341
7. Acknowledgements
This work is supported by Hebei Education Department Natural Science Foundation
(No.2008324), and Tianjin philosophy and social sciences key project Foundation (No.
TJYY08-1-078).
8. References
Abbott MD, Stanley RS. (1999). Modeling groundwater recharge and flow in an upland
fractured bedrock aquifer. Syst Dyn Rev, 15, pp. 163-184
Ahmad S, Simonovic SP. (2004). Spatial system dynamics: new approach for simulation of
water resources systems. J Comput Civ Eng, 18, pp. 331-340
Biswas AK. (1976). Systems approach to water management. McGraw Hill, London. ISBN:
0070054800
Claudia Pahl-Wostl. (2009). A conceptual framework for analyzing adaptive capacity and
multi-level learning processes in resource governance regimes, Global Environmental
Change, 19, pp. 354-365.
Forrester JW, 1968. Principles of systems. Productivity, Portland
Guo HH, Xu YL, Zou R. (1999). Study on the environmental system planning method for
watershed under incomplete information. J Environ Sci, 19, pp. 421-6 (in Chinese)
Huang G H, Chang N B. (2003). The perspectives of environmental informatics and systems
analysis. J Environ Inform, 1 (1), pp. 1-6
Jamieson D. (1996). Scientific uncertainty: How do we know when to communicate research
findings to the public? Science of the Total Environment 184, pp. 103-107
Mashayekhi AN. (1990). Rangelands destruction under population growth: the case of Iran.
Syst Dyn Rev, 6, pp. 167-93
Meadows DL, Meadows DH. (1973). Toward global equilibrium: collected papers.
Cambridge, Mass: Wright-Allen Press
Roberts N, Andersen D, Deal R, Garet M, ShafferW. (1983). Introduction to computer
simulation: a system dynamics modeling approach. Productivity, Portland
Saeed K. (1994). Development planning and policy design: a system dynamics approach,
Brookfield: Avebury
Yong Liu, Huaicheng Guo, Zhenxing Zhang, Lijing Wang, Yongli Dai, Yingying Fan. (2007).
An Optimization Method Based on Scenario Analysis for Watershed Management
under Uncertainty. Environ Manage, 39, pp. 678-690
Zhou Rui, Guo Huaichen, Li lei. (1998). A new method based on the objective-
deviationtolerance-level from multi-objective-decision making, Journal of system
engineering, 13(3), pp. 41-47(in Chinese)
Zhang Xuehua, Guo Huaichen. (2002). Application of SD-MOP integrated model in urban
eco-environmental planning of Qinhuangdao. J Environ Sci, 22, pp. 92-97(in
Chinese)
Zhang Xuehua, Guo Huaichen, Zhang Baoan. (2002). Application of System Dynamics-Multi
Objedtive Programme integrated model in urban water resources planning of
Qinhuangdao.Advance in water seience, 13, pp. 351-357(in Chinese)
Nonlinear Dynamics

342
Zhang Xuehua, Zhang Hongwei, Zhang Baoan. (2010). SD-MOP integrate model and its
application in water resources plan: A case study of Qinhuangdao. Kybernetes, 39
(in press)
15
A Detection-Estimation Method for Systems
with Random Jumps with Application
to Target Tracking and Fault Diagnosis
Yury Grishin and Dariusz Janczak
Bialystok Technical University, Electrical Engineering Faculty
Poland
1. Introduction
Methods for detection and estimation of the structure or parameters of abrupt changes in
dynamic systems play an important role for solving a number of problems encountered in
practice. They have an important significance in different fields of telecommunications and
control applications, such as radar tracking of maneuvering targets, fault diagnosis and
identification (FDI), speech analysis, signal processing in geophysics and biomedical
systems. Most of these applications belong to the class of problems with nonlinear
dynamics. Among them an important role is played by a wide class of systems with abrupt
random jumps of parameters or structure.
A dynamic system with jumps of this kind can be defined as a system in which the structure
or parameters can change at any random time. Therefore, in order to describe such a system,
it is convenient to introduce an unknown random vector ( ) k that determines the current
system structure and parameters. Then the system state and observation equations are
dependent on this changing vector. The general case then is described as follows:
( 1) [ ( ), ( ), ( )] x k F x k k w k + = , (1)
( ) [ ( ), ( ), ( )] , ( ) y k h x k k v k k = , (2)
where F and h are known nonlinear functions, ) (k w and ) (k v are system and measurement
noises respectively and is the space of possible values of the vector ) (k .
The space can consist of finite or infinite sets of elements. The structure of the space
and evolution of the vector ) (k in time determine the main approaches to solving the
problem of detection-estimation in a dynamic system with jump structure. The classification
of the statistical characteristics of the parameter vector ) (k is presented in Fig. 1.
According to this classification, after the jump the parameter vector ) (k can take on finite
or infinite sets of values. In the case of the former the dynamic system can be in one of N
possible structures. It has been shown that a model of this kind (Willsky, 1976) is the most
comprehensive description of system jump changes. Such models demand a considerable
amount of prior information on probable jump changes in the system. At the same time,
they require a great deal of computation when used for state estimation or jump detection in
Nonlinear Dynamics

344
real-time systems. Modifications to these models are often used for solving problems related
to tracking maneuvering targets in radars (Gini & Rangaswamy, 2008) and in designing
reliable dynamic systems (Patton et al., 1989). Usually in these cases the multiple model
(MM) (Blackman & Popoli, 1999), multiple hypothesis test (MHT) (Bar-Shalom et al., 2001)
or interactive multiple model (IMM) approaches are used (Mazor et al., 1998).


Fig. 1. Classification of the parameter vector ( ) k
Evolution of the vector ) (k in time can be described in terms of a random process with
a known multidimensional probability density function (pdf), by the Markov sequence or by
single jumps. In practice it is difficult to obtain a priori information about the
multidimensional probability density function of the process. Therefore a model based on
these criteria is not readily applicable to solving the problem of detecting jumps in dynamic
systems.
Models in which the vector ) (k is defined by Markov properties can describe a broad
variety of jump changes and hence they are widely used in radar applications and FDI theory
(Grishin, 1994). Another class of system models with a jump structure is represented by
systems with single jumps that can occur at random time, the pdf of these moments being
unknown. This approach assumes that after the jump, the system parameters and structure
remain unchangeable. The latter assumption is often unjustified in practice because after the
jump the system may be non-stationary. More adequate models are required in order to
describe situations in which following the jump the parameter vector ) (k changes
according to the Markov sequence. A model of this kind will be considered below.
For a solution to the problem in a real-time system with a minimum computational burden
it is desirable to have simple but adequate models of the jumps. A method for modelling
jumps in dynamic systems by means of additive Gauss-Markov sequences with random
time rises in the state and observation equation is proposed in (Grishin, 1994). Nevertheless
such models also require a relatively large amount of prior information on the structure and
parameter of the jumps.
In order to resolve these difficulties a mixed multiple additive Gauss-Markov model
is proposed. For this model far less a priori information on probable system jumps
is required and it can be applied to a broad class of dynamic systems for which relatively
simple models can be used.
Two states
2 , 1 ), ( = i k
i

N states
N i k
i
, 1 ), ( =
Finite sets Infinite sets
Random vector
) (k
State equation Measurement equation
Dimension of
Markov sequence Single jumps
Evolution of ) (k in time
Random process
A Detection-Estimation Method for Systems with Random Jumps with Application
to Target Tracking and Fault Diagnosis

345
Using such models and a generalized likelihood ratio approach (GLR) (Katayama &
Sigimoto, 1997) it is easy to obtain suboptimal algorithms for state estimation and jump
detection. Such algorithms in comparison with the multiple model estimation algorithms
have relatively moderate computation requirements. They can be obtained in recursive form
and realized in real-time systems.
In the following section of this chapter we outline the applications of models of this kind
to the problem of radar maneuvering target tracking and failure detection.
2. The system model
The system and measurement equations are described by one of the following models:

( 1) ( 1, ) ( ) ( ) ( 1) ( 1, )1( 1, ),
( ) ( ) ( ) ( ), 1, , ,
S j i i
x k k k x k w k G k k t k t
y k H k x k v k j N
+ = + + + + + +
= + =
(3)
or:

0
( 1) ( 1, ) ( ) ( ),
( ) ( ) ( ) ( ) ( ) ( , )1( , ), 1, , ,
j i i
x k k k x k w k
y k H k x k v k H k k t k t j N
+ = + +
= + + =
(4)
where ( ) x k is the state vector, ( ), ( ) w k k are white Gaussian sequences with zero mean and
covariance matrices ( ) Q k and ( ) R k respectively, ( , )
j i
k t - an unknown Gauss-Markov
state vector modelling changes in the system after the jump at the time
i
t and 1( , )
i
k t is the
unit step function that is zero when
i
t k < .
The vector ( , )
j i
k t can be written in the general case as follows for a dynamic system
driven by the random signal ( )
j
k :
( 1, ) ( 1, ) ( , ) ( ), 1,..., ,
j i j j i j
k t k k k t k j N + = + + = (5)
where ( 1, )
j
k k + - a transition matrix, ( )
j
k is a white Gaussian sequences with zero mean
and covariance matrix ( )
oj
Q k , j - a number of possible jump models of which prior
probabilities ( )
j i
P t can be given or not. The other notations specified are commonly used
(Sorenson, 1985). The a priori distributions of a random value
i
t are assumed to be
unknown.
Thus the additional dynamic system can be described by a set of equations of the form (5)
with different transition matrices. The choice of a corresponding model can be carried out in
real time by an adaptive processing algorithm. The case of one of N possible models will be
considered below.
Depending on the nature of the parameter vector ( , )
j i
k t the model of changes may be
classified (Grishin & Janczak, 2006) as deterministic ( 0 ) ( = k
j
), stochastic ( ( 1, ) 0
j
k k + = )
or mixed ( 0 ) ( , 0 ) , 1 ( + k k k
j j
).
It is easy to demonstrate that the equations (3) - (5) describe a wide variety of system jumps
which take place in different parts of the system such as jump changes of the state vector
and its dimension, jumps of the system transition matrix elements, the covariance matrices
of observation and system noises. Let us consider a description of different jumps in the
system with the additive Gauss-Markov models.
Nonlinear Dynamics

346
Jump changes of the state vector dimension
For
i
k t > equation (3) can be rewritten as
( 1) ( 1, ) ( ) ( 1) ( 1, ) ( , ) ( 1) ( ) ( )
S i S
x k k k x k G k k k k t G k k w k + = + + + + + + + (6)
Defining the augmented state vector as
[ ] ( 1) ( 1) ( 1, ) ,
T
a i
x k x k k t + = + + from (5) and (6)
( 1) ( 1, ) ( ) ( 1) ( ),
a a a a
x k k k x k k w k + = + + + (7)
where

( 1, ) ( 1) ( 1, ) 1 ( 1)
( 1, ) , ( 1)
0 ( 1, ) 0 1
S S
a
k k G k k k G k
k k k
k k

+ + + +
+ = + =

+


are transition and input matrices, [ ] ( ) ( ) ( )
T
a
w k w k k = - the augmented input noise vector.
Thus equation (3) may be used for modelling the jumps in the system dimension.
As the dimension of the observation vector is the same, the observation matrix for
i
k t >
must be altered, such that [ ] ( ) ( ) 0
a
H k H k = .
Jump changes of the state vector variables
If in equation (3) the input matrix is:

, 1 ,
( 1)
0 , 1 ,
i
S
i
I k t
G k
k t
+ =
+ =

+

(8)
then the state equation of the system will be:
( 1) ( 1, ) ( ) ( ) ( 1, ) ( 1, ) .
i i
x k k k x k w k k t k t + = + + + + + (9)
Thus every variable of the state vector at time 1
i
k t + = changes abruptly. The values of
these changes are equal to the values of the corresponding variables of the random vector
( 1, )
i
k t + . If for ( 1)
i S
k t G k I > + = and the parameters of equation (5) are chosen
as ( )
0 0
( 1, ) , ( , ) , ( ) 0 0 ,
i i
k k I t t k Q + = = = = then one has:

0
( 1) ( 1, ) ( ) ( ) 1( 1, ) .
i
x k k k x k w k k t + = + + + + (10)
The preceding equation shows, that state variable bias appears at time
i
t .
Abrupt changes of the observation matrix
In considering jumps of the observation matrix elements it is necessary to restrict our
discussion to equation (4). If for
i
k t > the identity ( , ) ( )
i
k t x k = is valid, that is
( 1, ) ( 1, ) , ( ) ( ) , ( , ) ( )
i i i
k k k k k w k t t x t + = + = = , then the observation equation is:

0 0
( ) ( ) ( ) ( ) ( ) ( ) [ ( ) ( )] ( ) ( ) , .
i
y k H k x k v k H k x k H k H k x k v k k t = + + = + + > (11)
3. Detection-estimation algorithms in the systems with the additive Gauss-
Markov jumps
To design an appropriate detection-estimation algorithm for a system in which parameters
can be abruptly changed, it is necessary to detect the changes, to isolate them (that is to
A Detection-Estimation Method for Systems with Random Jumps with Application
to Target Tracking and Fault Diagnosis

347
determine the system element in which these changes take place) and then to estimate theirs
value. The main approaches to the design of such algorithms include the following:
- change-sensitive filters (Limit Memory Filters) (Willsky, 1976),
- an innovation-based approach that uses the generalized likelihood ratio (GLR) (Gertler,
1998),
- the multiple hypothesis test (Katayma & Sugimoto, 1997),
- an artificial neural network approach (Patton et al., 1989).
In this section we focus on the GLR approach. An approach of this kind involves the use of
the basic Kalman filter which is matched with the normal mode of the input process and the
GLR computation of the innovation process to detect the parameter or structure jumps
(Whang et al., 1994).
When the system changes have occurred, the innovation process is no longer zero mean and
it carries information about changes in the system.
3.1 Synthesis of the detection-estimation algorithm
Let us consider the system for which state and measurement equations are given by the
model (3). Then, calculating the propagation of all signals through the Kalman filter that
is matched with a system without jumps, we obtain that the innovation process ( / 1) z k k
of the filter in this case can be presented in the following form (Grishin, 1994):

1
( / 1) ( , ) ( , ) ( / 1).
S S i i
z k k T k t k t z k k = + (12)
where
1
( / 1) z k k is the innovation process of the matched Kalman filter

1

( / 1) ( ) ( ) ( / 1) z k k y k H k x k k = (13)
and

1 2
( , ) [ ( , ) ( , ) ( ) ( , 1)]
s i c i c i
T k t k t k t H k k k = , (14)

1 1
1
( ) ( ) , ,
( , )
( )[ ( , ) ( , 1) ( 1, ) ( , 1)], ;
i S i i
C i
C i C i i
H t G t k t
k t
H k k t k k F k t k k k t

=

>

(15)

( ) , ,
( , )
1
( ) ( , 1) ( 1, ) ( , 1) ,
S
S
G t k t
i i
k t
c i
G t k k k t k k k t
i c i i

=


+ >

(16)

1 1
1 1
( ) ( ) ( ) , ,
( , )
( ) ( , ) ( , 1) ( 1, ) ( , 1), ,
i i S i i
c i
c i c i i
K t H t G t k t
F k t
K k k t k k F k t k k k t

=

+ >

(17)

2 1
2
( ) , ,
( , )
( )[ ( , 1) ( 1, ) ( , 1)], ,
i i
C i
C i i
H t k t
k t
H k I k k F k t k k k t

=

>

(18)

2 1
2 2
( ) ( ) , ,
( , )
( ) ( , ) ( , 1) ( 1, ) ( , 1), .
i i i
C i
C i C i i
K t H t k t
F k t
K k k t k k F k t k k k t

=

=

+ >

(19)
Nonlinear Dynamics

348

(1) (2)
2 2
( , ) [ ( , ) ( , ) ( , )] ,
T T T T
i i i i
k t k t k t k t = (20)

( 1) ( 1)
2 2
( , ) ( , 1) ( 1, ) ( , ) ( 1) ,
i i i
k t k k k t L k t k = + (21)

( 2) ( 2)
2 2
( , ) ( , 1) ( 1, ) ( , ) ( 1)
i i i
k t C k k k t N k t k = + (22)

1
0, ,
( , )
( , 1)[ ( 1, ) ( 1)] ( , 1), ,
i
i
i S i
k t
L k t
k k L k t G k k k k t

=

>

(23)

1 2
( , ) ( , ) ( , ) ,
i i i
N k t N k t N k t = + (24)
1
1
1
0, ,
( , )
[ ( 1) ( 1) ( 1, ) ( , 1) ( 1, )] ( , 1) , ,
i
i
C i i i
k t
N k t
K k H k k t C k k N k t k k k t

=

+ >

2
1
2
0 , ,
( , )
[ ( 1) ( 1) ( , 1) ( 1, )] ( , 1) ( , ) , .
i
i
i i i
k t
N k t
K k H k C k k N k t k k L k t k t

=

+ >

It follows from equations (14) and (22) arising at time
i
t that the additive gauss-Markov
jump changes in the system dynamics result in the appearance of the random vector ( , )
i
k t
of which one of components is the vector ( , )
i
k t , in the innovation process of the matched
Kalman filter. When deducing expressions (14)-(22) we used the assumption that
the transition matrix ( 1, )
j
k k + from (5) is non-singular. This assumption is usually feasible
in engineering practice. The block diagram representation of the innovation process for
the system (3) is presented in Fig. 2.


Fig. 2. Block diagram representation of the innovation process for the system with structure
or parameters jumps in the system equation
Taking into consideration formulae (13) - (22) the system presented in Fig. 2 can be written
in the augmented form as follows:
( 1, ) ( 1, ) ( , ) ( 1, ) ( )
i i i
k t k k k t J k t k + = + + + (25)
where the state transition and input matrices of the augmented system are calculated as:
( ) ( 1, ) ( 1, ) ( 1, ) ( 1, ) k k diag k k k k C k k + = + + + and ( 1, ) [ ]
T T T
J k k I L N + = .
No abrupt changes
) 1 (
2

) 2 (
2

) 1 / ( k k z
) 1 / (
1
k k z
) (k ) (k
Delay
1 c
) , ( t k L

Delay H ) , ( t k N
C
Delay
2 c

Abrupt changes
A Detection-Estimation Method for Systems with Random Jumps with Application
to Target Tracking and Fault Diagnosis

349
When the system jumps take place in the observation channel described by equation (4) the
innovation process ( / 1) z k k has similar form to (12) :

1
( / 1) ( , ) ( , ) ( / 1)
o o i i
z k k T k t k t z k k = + , (26)
where all components of equation (26) can also be obtained in recursive form taking into
consideration propagation of the signals through the Kalman filter matched with the
undisturbed system :
[ ]
0 0
( , ) ( , ) ( ) ( , 1) ,
i i
T k t k t H k k k = (27)

1
( , ) [ ( , ) ( , )] ,
T T T
i i i
k t k t k t = (28)

1
0 0 0
( , ) ( ) ( ) ( , 1) ( 1, ) ( , 1),
i i
k t H k H k k k F k t k k

= (29)

1
0 0 0
( , ) ( ) ( , ) ( , 1) ( 1, ) ( , 1),
i i i
F k t K k k t k k F k t k k

= + (30)

1 1
( 1, ) ( 1, ) ( , ) ( 1, ) ( ),
i i i
k t C k k k t D k t k + = + + + (31)

1
0
( 1, ) [ ( ) ( ) ( 1, ) ( , )] ( 1, )
i i
D k t K k H k C k k D k t k k

+ = + + + , (32)
( 1, ) [ ( ) ( )] ( , 1) C k k I K k H k k k + = . (33)
Thus the problem under consideration can be formulated as a test of two hypotheses
the simple hypotheses
o
H with respect to the composite alternative
1
H :

, ) 1 / ( ) , ( ) , ( ) 1 / ( :
) 1 / ( ) 1 / ( :
1 1
1 0
+ =
=
k k z t k t k T k k z H
k k z k k z H
i i


(34)

where ) , ( ), , (
1 i i
t k t k T are described by (14) and (20) or (27) and (28).
Since the a priori distributions for
i
t and ( , )
i
k t are unknown we have to use the
generalized likelihood ratio (GLR) test. The GLR for the hypotheses (34) for
i
k t can be
written as follows (Grishin & Janczak, 2006):

1
1
0
[ ( / 1) / , ( , ( , ))]
( , ) ( 1, )
[ ( / 1) / ]
k
ti i i
i i
f z k k z H t k t
k t k t
f z k k H

(35)
Since the vector ( / 1) z k k in (34) is Gaussian the probability density functions [ ] f

in this
expression are also Gaussian. Thus the likelihood ratio can be written in the logarithmic form:
, 0 ) , 1 (
)], 1 / (
~
) 1 / ( [ ) ( )] 1 / (
~
) 1 / ( [ ) 1 / ( ) ( ) 1 / (
) ( det ln ) ( det ln ) , 1 ( ) , ( ln ) , (
1 1
1
1
=
+
+ + = =

i i
zo
T
z
T
zo z i i i
t t
k k z k k z k P k k z k k z k k z k P k k z
k P k P t k t k t k

(36)

where
1
( )
z
P k is the covariance matrix of the innovation process in the matched Kalman filter
(hypothesis
o
H ), the value
Nonlinear Dynamics

350

1
1

( / 1) [ ( / 1) / , ] ( , ) ( / 1, )
k
ti i i
z k k E z k k z H T k t k k t

= = (37)
is the prediction estimate of the innovation process for jumps which have occurred at
known time
i
t and


( / 1, ) ( , 1) ( 1 / 1, )
i i
k k t k k k k t =
(38)
is the prediction estimation of the Kalman filter for the system described by the expressions
(12) and (25).
The covariance matrix ( )
zo
P k from (36) is given by

1
( ) ( , ) ( / 1, ) ( , ) ( )
T
zo i o i i z
P k T k t P k k t T k t P k = + , (39)
where ( / 1, )
o i
P k k t is the covariance matrix of the estimate (38).
Therefore if the estimates

( / 1, )
i
k k t for each given
i
t are calculated the maximum
likelihood estimate is

. ) , ( max arg

i
t
i
t k t
i
=

(40)

Then the decision rule is

1
0
0

( , ) ( , ) , 1 ,
i i i
H
H
k t k t k M t k +
>
<
(41)
where )

, (
0 i
t k is the threshold value and

1
i
k M t k + is used to avoid a growing bank
of filters.
Thus the system of joint detection - estimation of jumps changes in a dynamic system
consists of the basic Kalman filter, which calculates values ) 1 / ( k k z , the bank of Kalman
filters, which compute the likelihood ratios ) , (
i
t k at different moments k M k t
i
,... 1 + = ,
the logic circuit, which selects the maximum value ) , (
i
t k and a threshold circuit for
detection of abrupt changes. Such a detection-estimation algorithm demonstrates
a moderate computational burden and can be carried out in real-time systems. Its structure
is presented in Fig. 3.


Fig. 3. Detection-estimation algorithm for the system with additive Gauss-Markov jumps
No
0
>
) (k y
) 1 (
0
+ M k
FK
1 + = M k t
1 + M k

) (
0
k
FK
k t =
k

) (k K
1
Iz
H



) , ( max arg
i
t
t k
i


)

, (

i
t k
Yes

A Detection-Estimation Method for Systems with Random Jumps with Application
to Target Tracking and Fault Diagnosis

351
The partial estimates ) , (

i
t k are obtained using M N = 1 samples of the innovation
process ) 1 / ( k k z and therefore they can be obtained using the finite memory filters of
which weights are calculated recursively.
3.2 Synthesis of the simplified detection-estimation algorithm
The method presented in section 3.1 is effective in supplying reasonably accurate estimates
of the state vector ) , (
i
t k . Moreover it does not require a priori knowledge of the additional
system state vector ) , 1 (
i i
t t initial value. However high order systems results in
a relatively high calculation burden. This is a consequence of the high order of the Kalman
filter for the system (12)-(33) and the necessity for filter parameter calculations at every time
step. To remediate these difficulties some simplifications may be introduced. As will be
shown in the following section, assuming an a priori knowledge of the vector initial value
) , 1 (
i i
t t , the decision filter equations (12) - (33) may be simplified. In this case the filter
parameters may be calculated prior to the estimation process (off line). Of course, a set of
adequately spaced initial values ) , 1 (
i i j
t t should be assumed and the corresponding
filters should be applied to the system structure (Fig. 3). Simulation investigations of the
detection method have shown it to be reasonably robust to inaccuracy of the vector
) , 1 (
i i j
t t value and the decision method chooses a filter initialised with ) , 1 (
i i j
t t that
is closest to the real one. The accuracy of the simplified method is not amenable to
the method described in the previous section but the calculation burden is smaller.
A detection-estimation algorithm can be obtained in a way similar to that described in
section 3.1 but with additional assumption that is known ) , 1 (
i i j
t t . A representation of
the residuals ) 1 / ( k k z for
i
t k can be divided into two components (one associated with
the undisturbed system and the other following a given failure) and has the following form
(in the case of system (4)):

1
0
( / 1, ) ( / 1) ( , ) ( , 1) ( 1, ) ( , ) ( 1) ,
i
k t
i z i i i i i z i i
n
z k k t z k k k t t t t t k t n t n

=
= + + + +

(42)
where ) 1 / (
1
k k z is the innovation process (zero mean white noise) related to the
unchanged system and the remaining elements represent the influence of specific system
change on the residuals of the filter matched to the undisturbed model.
All elements ) , (
i z
t k depend on the system matrices, onset time and filter gain and can be
calculated in a recursive way. In the case of failure described by the equation (4) these
elements can be calculated as follows:
( , ) ( ) ( , ) ( ) ( , 1) ( 1, ),
z i o z i z i
k t H k k t H k k k F k t = (43)
( , ) ( , 1) ( 1, ),
z i z i
k t k k k t = (44)
( , ) ( ) ( , ) ( , 1) ( 1, ) ,
z i z i z i
F k t K k k t k k F k t = + (45)
with initial conditions: 0 ) , 1 ( =
i i z
t t F , I ) , 1 ( =
i i z
t t where I is the identity matrix.
Considering equation (42) the detection problem can be formulated as a statistical test
of two hypotheses (
1 0
, H H ), the first of which ) (
0
H is intended to test the presence of
Nonlinear Dynamics

352
the white noise ) 1 / (
1
k k z and the second ) (
1
H , the presence (
1
H ) of the signal

0
) , (
i z
t k to ) 1 / (
1
k k z

noise background.

0 1
1 1 0
: ( / 1) ( / 1) ,
: ( / 1) ( / 1) ( , ) ,
z i
H z k k z k k
H z k k z k k k t

=
= +
(46)
where ) , 1 ( ) 1 , (
0 i i i i
t t t t =

and ) 1 / (
1
k k z

represents all noise components from
equation (42).
Since the distribution of the onset time
i
t is unknown a priori, the generalized likelihood
ratio (GLR) test is used:

1
0
max [ / ( )]

( , )
[ / ]
i
i
i
k
t i
k
i k
t
f Z H t
k t
f Z H
= , (47)
where ] [ f is the conditional probability density function and
)} 1 / ( , ... ), 1 / ( { = k k z t t z Z
i i
k
t
i
.
The decision procedure has the form (48) where the generalized likelihood logarithm
)

, (
i
t k is compared with the threshold )

, (
i p
t k . A variable threshold level is applied.
( )
1
0

( , ) ( , ) , arg max ( , ) , 1 ,
i
i p i i i i
t
H
H
k t k t t k t k M t k = +
>
<
(48)
where )

, (
i
t k is the logarithm of )

, (
i
t k , M is the width of the moving window used
to avoid an increasing number of additional filters matched to successive onset moments.
3.3 Threshold determination
The performance of the decision procedure is essential to the efficiency of detection and so
to the quality of estimation. The general principles of the applied GLR method are well
established (Willsky, 1976), (Sage & Melsa, 1971). Unfortunately, the use of the GLR
approach requires knowledge of the resulting probability distributions. For instance in the
detection - estimation structure based on the Kalman filter the usually resulting probability
distributions are unknown and the threshold value cannot be obtain in an analytical way.
The detailed solutions to the problem proposed in the literature are based on simplifications
such as the use of simplified statistics (not GLR) or experimental determination. Moreover
in numerical examples a constant threshold level is used. This approach is correct under
steady state conditions of the object and estimator when the corresponding probability
density functions are constant. It is not appropriate in a non-stationary state of the object or
filter and leads to permanent additional detection delay under such conditions. The solution
to the problem requires that changes in the probability distributions and application of
a variable threshold level be taken into consideration. This approach allows the constant
probability of false alarm (P
FA
) to be obtained, i.e. the probability of taking the decision that
a fault has occurred while the system is in a normal state. A method for obtaining a non-
A Detection-Estimation Method for Systems with Random Jumps with Application
to Target Tracking and Fault Diagnosis

353
constant threshold level variable for a simplified filter as described in the previous section
will be presented next.
The choice of a decision threshold )

, (
i p
t k can be obtained using the Neyman - Pearson
criterion, where a probability P
FA
of the false alarm level is assumed.

0
0 ( , )/
( , )
( ( , ) / ) 1 ( ( , )) ,
i
i
FA i o o k t H p i
p k t
P f k t H d F k t

= = =

(49)
where )) , ( (
0
/ ) , ( i p H t k
t k F
i

is the conditional probability distribution function of ) , (

i
t k .
As seen in (49), the decision threshold can be determined with the use of
0
( , )/
( ( , ))
i
k t H p i
F k t

.
It can be shown (Grishin, 1994) that the GLR logarithm can be computed in the following
way:

{
}
1
1
1
1
1
1
1
( , ) [ ( / 1)] ( / 1)[ ( / 1)] [ ( / 1) ( / 1, )]
2
( / 1)[ ( / 1) ( / 1, )] ln[det ( ( / 1)] ln[det ( ( / 1)] ,
i
k
T T
i z H i
l k
z H i z z
k t z l l P l l z l l z l l z l l t
P l l z l l z l l t P l l P l l

=
+

(50)

where
1
( / 1)
z
P l l , ( / 1, )
z i
P l l t , and
1
( / 1, )
H i
z l l t are covariance matrixes and
the expected value of the following conditional probability distributions for the Kalman
filter innovation process ) 1 / ( k k z :

1
1
0 0 1
1
1 1
[ ( / 1) / , ] [ ( / 1) / ; 0, ( / 1)] ,
[ ( / 1) / , ] [ ( / 1) / ; ( / 1, ), ( / 1)] .
i
i
l
t z
l
t H i z
f z l l Z H N z l l H P l l
f z l l Z H N z l l H z l l k P l l

=
=
(51)

Taking into consideration equation (42), the parameters of the distributions (51) can be
calculated as follows:

1
( / 1) ( ) ( / 1) ( ) ( ),
T
z
P l l H l P l l H l R l = + (52)

1
0
( / 1, ) ( / 1) ( , ) ( 1) ( , ) ,
i
k t
T
z i z z i i z i
n
P l l t P l l l t n Q t n l t n

=
= + + + +

(53)

[ ]
1
1 0
( / 1, ) ( / 1, ) / ( , ) ,
H i i z i
z l l t E z l l t H l t

= = (54)

where ( / 1) P l l is the covariance matrix of the state vector prediction

( / 1) x l l obtained
in the basic Kalman filter.
Unfortunately, as follows from (50) the GLR logarithm ( , )
i
k t is the difference between
a random variable with
2
distribution (first term) and a random variable with a non-
central
2
distribution (second term) in summation with the deterministic term (third part),
so an appropriate approximation of the distribution should be applied. The following
approximation of the sum (50) can be assumed:

0

( , ) ( , ) ( , ) ( , ) ( , ) ,
i i a i a i d i
k t k t k t k t c k t = + (55)
Nonlinear Dynamics

354
where ( , )
a i
k t ,
0
( , )
d i
c k t are coefficients, ( , )
a i
k t is a random variable with a known and
easy to compute distribution that would allow for approximation of the ( , )
i
k t
distribution.
The sum (50) can be written as:

( )
( )
2
0 0 0 0
1
1
( , ) ( , ) ( ) ( / 1) ( ) , ,
2
i
k s
i S i j j j d i
l t j
k t k t a l z l l b l c k t
= =


= + +





(56)
where:
2 2
1 0
0 2
1
( / 1) ( / 1)
( )
( / 1)
j j
j
j
l l l l
a l
l l

,
0
0 2 2
1 0
( / 1) ( / 1)
( )
( / 1) ( / 1)
j j
j
j j
l l z l l
b l
l l l l

=

,
( )
( )
1
0 0
1
det ( / 1) 1
( , ) ( ) ln
2 det ( / 1)
i
k s
z
d i j
l t j
z
P l l
c k t c l
P l l
= =


= +

2
0 2 2
0 1
( / 1)
( )
( / 1) ( / 1)
j
j
j j
z l l
c l
l l l l

=

,
2
0
( / 1)
j
l l ,
2
1
( / 1)
j
l l are j-th elements from the diagonals of matrices
1
( / 1)
z
P l l ,
( / 1)
z
P l l respectively, ( / 1)
j
z l l is j-th element of the vector ( / 1) z l l and
0
( / 1)
0
( / 1)
j
j
z l l
j
z l l


= , so
0
( / 1)
j
z l l is normally distributed [0, 1] N .
Defining a new variable
0
( , )
cd i
k t :

2
1
0 0 0 0 0 2
1
( , ) ( , ) ( , ) ( ) ( / 1) ( )
i
k s
cd i S i d i j j j
l t j
k t k t c k t a l z l l b l
= =
= = +

(57)
we can see that
0
( , )
cd i
k t is the weighted sum (with weights ) (
0
2
1
l a
j
) of squares of
( 1)
i
s k k + normally distributed (
0
[0, ]
j
N b ) variables. This leads to the idea of using the
non-central
2
distribution as an approximation distribution (the distribution of ( , )
a i
k t ).
In the case of the non-centrality parameter (
nc
), the number of degrees of freedom (
nc
N )
and the coefficient ( , )
a i
k t (
nc
) must be determined. Calculation of these parameters is
performed by matching three statistical moments (the first non-central, second and third
central) of the variable ( , ) ( , )
a i a i
k t k t (see (55)) and the sum
0
( , )
cd i
k t (see (57)).
As a result two sets of solutions ( { , , }
nc nc nc
N , { , , }
nc nc nc
N ) are obtained:
2 p
nc
m
S S
S

+
= ,
2 3
2
m p
nc
nc
S S
S S

=

,
( )
2 3
3 2
nc nc
nc
nc p
S S
N
S

,
2 p
nc
m
S S
S

= ,
2 3
2
m p
nc
nc
S S
S S

=

,
( )
2 3
3 2
nc nc
nc
nc p
S S
N
S

,
where
( )
2
0 0
1
( , ) ( ) 1 ( )
i
k s
m m i j j
l t j
S S k t a l b l
= =
= = +

,
( )
2 2
2 2 0 0
1
( , ) ( ) 1 2 ( )
i
k s
i j j
l t j
S S k t a l b l

= =
= = +

,
A Detection-Estimation Method for Systems with Random Jumps with Application
to Target Tracking and Fault Diagnosis

355
( )
3 2
3 3 0 0
1
( , ) ( ) 1 3 ( )
i
k s
i j j
l t j
S S k t a l b l

= =
= = +

,
2
2 3 p m
S S S S

= .
The set with 0
nc
and 0
nc
N > should be taken as the final solution. Moreover at the
beginning the following condition should be checked:
2
2 3
0
m
S S S


If the condition is not fulfilled the above approximation cannot be calculated. In this case an
approximation using the central
2
distribution was also derived and tested. However this
is less accurate in cases of low value of M (moving widow width) but has no numerical
constraint and needs less computation. Two of the required parameters (the number of
degrees of freedom and the coefficient ( , )
a i
k t ) can be determined by matching two
distribution parameters (mean value and variance) of the variable ( , ) ( , )
a i a i
k t k t and the
sum
0
( , )
cd i
k t .
In practice, the number of degrees of freedom obtained in both approximations is not
usually an integer number, so the distributions cannot be computed as typical central
2
or
noncentral
2
distributions. Instead of the central
2
distribution function the Gamma
distribution function (with parameters ( , ) /2
c i
N k t and 2) can be used. The other
distribution can be calculated in the following way (modification of the standard numerical
procedure):

( )
{ }
2
2
( , ) 2
0 0
/2
( ) ( ; /2) ( ; 2 , 2),
!
nc
i nc
i
nc
k t N i Po nc nc
i i
F x e P x f i F x N i
i

+
= =

= = +



(58)
where
Po
f - Poisson probability density function, F

- Gamma cumulative distribution

function.
The performance of the proposed method was tested by means of numerical simulations.
The results presented below were obtained for the first order process model and on the basis
of additive changes to the observation equation (see (4)) with the following parameters:
( , 1) 1 k k = , ( ) 1 H k = ,
2
( ) (0.2) Q k = , ( ) 1
o
H k = ,
2
( ) 10 R k = , ( , 1) 1 k k = ,
2
( ) (0.8) Q k

= ,
( 1, ) 1
i i
t t = ,
0 0
(0) : [ ; 12, 10] x x N x = . At the beginning the accuracy of the
approximations was tested using Monte Carlo simulation (number of simulations
100000
s
N = ). In Fig. 4 the distribution of ( , )
i
k t (determined by numerical experiment -
ex) and analytically calculated approximations (nc - noncentral, c - central
2
distribution) are compared for the case of 1 M = (the smallest width of the moving
window) and 5 M = (medium value of M).

-2 0 2 4
0
0.1
0.2
0.3
0.4
0.5
0.6
f[]
ex
nc
c
M = 1 ki = 10

-4 -2 0 2 4
0
0.05
0.1
0.15
0.2
f[]
ex
nc
c
M = 5 ki = 10

Fig. 4. Distribution of ( , )
i
k t (ex) and its approximations (nc, c) for 1 M = , 5 M =
Nonlinear Dynamics

356
As can be concluded from Fig. 4 the approximation nc is precise for all M. The accuracy of
approximation c is not so exact, especially for low value of M and low threshold level
(high P
FA
). These observations were confirmed by analytical measures. The Kullback
measure of distances between the distribution of ( , )
i
k t and its approximations were
calculated. The results are shown in table 1.

M=1 M=2 M=3 M=4 M=5
nc 0.0018 0.0023 0.0023 0.0024 0.0022
c 0.0161 0.0139 0.0110 0.0078 0.0058
Table 1. Kullback measure of distances between the distribution of ( , )
i
k t and its
approximations.
The numerical data presented in table 1 confirm that the approximation c is far less
accurate then nc for small M but is comparable for higher M values ( 5 M ).
Next, the threshold level was calculated. A constant probability P
FA
of false alarm was
assumed. This caused a change in the threshold value. The results are shown in Fig. 5.
It should be added that the character of the changes depends on system and failure
parameters and can vary from that presented.

0 10 20 30 40
2
3
4
5
6
M = 1
M = 4
M = 8
M = 12
M = 16
M = 21
k
p(k)

Fig. 5. Variation of threshold level in the case of constant P
FA

Finally a check of the validity of the threshold algorithms was performed by testing
the outcome probability P
FA
of false alarm. The results of
6
10
s
N = Monte Carlo simulations
are shown in Fig. 6. There were two P
FA
values assumed: 0.01
FA
P = and 0.001
FA
P = . The
parameter is verified for 1,..., 5 M = . The mean value of P
FA
was calculated and is shown as
FA
P .

0 5 10 15 k
0,0008
0,0009
0,001
0,0011
0,0012
M=1 P
FA
=0,0010
M=2 PFA=0,0011
M=3 P
FA
=0,0011
M=4 P
FA
=0,0010
M=5 P
FA
=0,0010
PFA = 0,001
PFA(k)

0 5 10 15 k
0.0096
0.0098
0.01
0.0102
0.0104
0.0106
0.0108
M=1 PFA=0.01
M=2 PFA=0.0101
M=3 PFA=0.0101
M=4 PFA=0.0101
M=5 PFA=0.0101
PFA = 0.01
PFA(k)

Fig. 6. P
FA
variation in time when thresholds were calculated for 0.001
FA
P = , 0.01
FA
P =
A Detection-Estimation Method for Systems with Random Jumps with Application
to Target Tracking and Fault Diagnosis

357
It can be seen from Fig. 6, that the proposed method demonstrates high accuracy.
The maximum difference between the obtained and assumed P
FA
was less than
4
8 10 P

= .
The difference diminishes as the number of simulations increases. Mean values
FA
P are very
close to the assumed P
FA
.
The simulation results demonstrate the effectiveness of the proposed probability
distribution approximations. The method allows a constant rate of the probability of false
alarm to be obtained in the non-stationary state of the object or filter.
4. Tracking of maneuvering targets
The demands of high precision tracking and guidance systems require accurate state
estimation of the targets. A variety of maneuvering target tracking methods have been
proposed in the literature. The main principles and techniques used to track target in real
situations and a comparative evaluation of some of the algorithms can be found in
(Blackman & Popoli, 1999). In recent years a great deal of new maneuvering target tracking
algorithms have been proposed. Among them, there are algorithms such as those which use
the input estimation (IE) technique, variable dimension (VD) filtering, multiple hypothesis
tracking (MHT) and the interacting multiple model (IMM) approach (Blackman & Popoli,
(1999), (Bar-Shalom & Fortmann, 1988), (Bar-Shalom et al., 2001), (Li & Bar-Shalom, 1993).
Although the structure of many optimal algorithms of maneuvering target tracking is
known, the computational complexity often limits theirs practical realization. Many
different tracking algorithms have been developed for the purposes of computational
feasibility. Some of them use combined techniques such as IMM/IE, IE/VD (Blackman &
Popoli, 1999). For a mathematical description of a maneuver the following models are
usually used: white noise models, a noisy jerk as a maneuver model, non-random maneuver
models and combined target maneuver models. The additive Gauss-Markov Models
(AGMM) presented earlier enable a realistic but simple description of quite complex
changes in a real process to be obtained. The maneuver of a moving object manifests as
a change in acceleration. Usually the change is modelled as a step or ramp function. In most
applications this approximation is sufficient but for precise or close distance tracking
the change model should be more representative. Reasonably accurate maneuver models
incorporate acceleration changes in the form of inertial system step response in the presence
of correlated noise. The acceleration dynamics (Blackman & Popoli, 1999) can be described
as:

) ( )] , ( 1 ) , ( 1 [ ) (
1
) ( t w t t t t t a t a
j i
+ + =

,

(59)


where ) (t a is acceleration, is acceleration level, is correlation time, w(t) is zero mean
white noise with covariance
w
Q and 1( , )
i
t t is unit step function with onset time
i
t and
j
t
is a time of maneuver termination.
An example of acceleration (
2
19.6
m
s
= for
2
4
1
m
w
s
Q = and
2
4
9
m
w
s
Q = ) used for simulation is
presented in Fig. 7.
Nonlinear Dynamics

358
0 20 40 60 80 100
0
5
10
15
[m/s
2
]
a
Qw
Qw
[s]
t

Fig. 7. Realization of an acceleration modelling maneuver
Defining the components of the state vector in terms of position, velocity and acceleration,
the target dynamics model on one axis can be written as:

)] , ( 1 ) , ( 1 )[ ( ) ( ) ( ) ( ) ( ) (
j i
t t t t t B t w t B t x t F t x + + =

,

(60)

where matrices ( ), ( ) F t B t are defined as:

1
0 0
1 0 0
0 1 0
) (t F ,

=
1
0
0
) (t B .
A discrete form of the model (60) is given by:

)], , 1 ( 1 ) , 1 ( 1 )[ 1 ( ) ( ) ( ) , 1 ( ) 1 (
j i d d d
t k t k k B k w k x k k k x + + + + + + = +

(61)

where the transition and system input matrices take the values:
2
1 ( 1 exp( ))
( 1, ) 0 1 (1 exp( ))
0 0 exp( )
T T
T
d
T
T
k k

+

+ =

,
2
2
2
2
(1 exp( ))
( ) ( 1 exp( )) ,
(1 exp( ))
T T T
T T
d
T
B k

+

= +






where T is the sampling time and ) (k w
d
is zero mean white noise with covariance matrix:

[ ]

= =
33 32 31
23 22 21
13 12 11
) ( * ) ( ) (
q q q
q q q
q q q
k w k w E k Q
T
d
d d
,

( ) ( ) ( ) ( ) ( ) ( )
2 2 3
2 4
2
3 11
1 2 2 2 exp ,
T T T T T
q



= + + + +




( ) ( ) ( ) ( ) , exp exp 2 2 1
2
3 2
21 12

+ + = =



T T T T
q q

( ) ( ) [ ]


T T T
q
2
2 2
22
exp exp 4 3

+ + = ,

( ) ( ) [ ]


T T
q q
2
2
32 23
exp exp 1

+ = = ,

A Detection-Estimation Method for Systems with Random Jumps with Application
to Target Tracking and Fault Diagnosis

359
( ) ( ) ( ) [ ]


T T T
q q
2
2 2
31 13
exp exp 2 1

= = ,

( )

T
q
2
33
exp 1

= .

This complex model can be described by means of AGMM additive to the state (63).
Maneuver is treated as a change in the order of target dynamics from the second (62) to
the third (61) and is modelled by means of vector ) , 1 (
i
t k + (64):

) ( ) (
1 0
1
) 1 (
1
k w k x
T
k x +

= +
,

(62)


), , 1 ( ) 1 ( ) ( ) ( ) , 1 ( ) 1 (
i
t k k G k w k x k k k x + + + + + = +
(63)

), ( ) , ( ) , 1 ( ) , 1 ( k t k k k t k
i i
+ + = +
(64)
where corresponding matrices take the following form:
1
0 1
T
=


,
1
( 1, )
0 1
T
e
k k

+ =



,
0
( 1, ) ,
1
T
i i
t t
e

( )
( )
( )
( ) ( )
( ) ( ) ( )
2
2
2 2
2
1 1 1 1 1
1 1 1
T T T
T T T
T T T T
T
e e e
G
e e e

+ + +

=

+ + + +

.
The performance characteristics of the proposed method were compared with the widely
used IMM technique (Bar-Shalom at al., 2001), (Blackman & Popoli, 1999), (Li & Bar-Shalom,
1993) using Monte Carlo simulations. Maneuver was modelled as acceleration change
described by the scenario shown in Fig. 7 ( 300
i
t T = , 600
j
t T = - ,
i j
t t - onset and
termination time). For a simulation of the IMM algorithm three models of the movement
have been used: the constant velocity model, Singers model with a correlation time 10s =
and
2
4
2
1
m
m
s
= , and model described by Singers model with constant acceleration of
2
19.6
m
s
= . The elements of transition matrix are equal to 0.9
ii
p = on the diagonal and
0.05
ij
p = elsewhere. Initially all models are assumed equiprobable.
In the Fig. 8 the root mean square errors (RMSE) of distance and velocity estimates are
shown. As follows from the schedules, the AGMM algorithm demonstrates a better
estimation performance in comparison with the IMM method everywhere apart from
transient parts of the manouver. Smaller estimation errors are achieved due to adaptation of
the AGMM filter dimension with respect to the real process model.
5. Failure detection in a multisensor integrated system
5.1 Fault tolerant airborne navigational system structure
As an example of the application of the methods developed to the problem of fault
detection-identification, let us consider reliable data processing in integrated GPS-based
Nonlinear Dynamics

360

k
RMSE
[m]
IMM
AMGM
0 20 40 60 80 100
0
2
4
6
8
10
12
14
16

RMSE
k
IMM
AMGM
0 20 40 60 80 100
0
2
4
6
8
10
12
14
16
[m/s]

Fig. 8. RMS error of position (left) and velocity (right)
airborne navigational equipment (Brown & Hwang, 1987), (Grishin, 2000). The possible
structure of a real airborne navigational aid is presented in Fig. 9. It may consist of a number
of radio-navigational and self-contained sensors such as the Microwave Landing System
(MLS) or the Instrument Landing System (ILS), the VOR/DME system, the Global
Positioning System (GPS), the Inertial Navigation System (INS) and the System of Air
Signals (SAS) supplying barometrical and altitude information (Fadden & Schwab, 1989).
Each sensor has independent diagnostic facilities (DF) which check the sensor serviceability
and control a state matrix circuit (SM). The latter determines the availability of the sensor
output data. When a sensor is out of order the integrated filter does not use the sensors data


Fig. 9. The structure of the fault-tolerant airborne navigation equipment ( DF - diagnostic
facilities, SM - state matrix circuit, CR - coordinate recalculation, FDIA - fault detection-
identification algorithm, GC - gate circuit, FAS - failure alarm signal, Tr - transmitter,
IFA - integrated filtering algorithm)
A Detection-Estimation Method for Systems with Random Jumps with Application
to Target Tracking and Fault Diagnosis

361
and the plane state vector estimate is computed with the aid of normally operating sensors
only. The corresponding failure alarm signal (FAS) has to be transmitted to the systems
users. It should be noted that the diagnostic facilities are able to detect only solid failures in
the airborne equipment and cannot determine faults in the ground-based or space-based
facilities.
In the absence of failures the integrated algorithm is usually based on non-linear
modifications of the Kalman filter (Sage & Melsa, 1971).
The main objective of this section is to present the algorithms for data processing in the
multisensor GPS-based airborne navigational equipment which, on the one hand would be
tolerant to possible failures of the information sources and on the other hand could enhance
the integrity of the whole navigational system. The main complicating factors accompanying
the solution to the problem are: rapid changes to the satellite geometry, the presence of
receiver clock error, increased dynamics of the aircraft and availability of additional
information from a number of the sensors mentioned above. In this case, fault-tolerant
signal processing can be based on analytical and/or physical redundancy (Grishin, 2000).
One of the main characteristics for a system of this kind is integrity (Brown, 1988) which can
be thought of as the ability of the system to provide a timely warning to users as to when the
system should not be used for navigation. The integrity performance characteristics such as
integrity warning time and accuracy threshold requirements vary with the phase of flight
(oceanic en route, domestic en route, terminal area and nonprecision approach). Higher
reliability and integrity of airborne equipment may be achieved as a result of the detection
of individual sensor failures and computation of the state estimates using data which have
their origin in the normal operated sensors only.
For modelling the failures of individual subsystems the additive Gauss-Markov models
considered in section 3 were used:
1. jump biases in observations (equation 4) with unknown onset time and value (antenna
beam distortion, time jumps in the GPS due to a gradual degradation of the satellite
clock, random bias in the INS due to drift of gyroscopes and so on);
2. random drifts (ramp-type incipient failures) which can be caused by multiple path
propagation effects in the ILS, frequency shifts in the GPS, soft failures in the INS and
a number of other failures that can be described by the equation (3).
Furthermore, it is necessary to take into consideration multiple malfunctions that can arise
in the sensors which result in outliers at the input of the integrated estimation filter. These
outliers can be caused by pulse interferences, by signal amplitude fluctuations or by clutter
or intentional jamming.
It is assumed here that outliers have a normal pdf (0, )
ki
N R

with a covariance matrix

2
( )
ki ki
R R k =

, where
2
1
ki
>> depending on the signal amplitude
i
A . This means that when
the outliers occur the pdf of measurements changes and their variances take on M different
values.
Thus the observation equation can be written as follows:
( ) ( ) ( ) ( ),
i
y k H k k k = + (65)
where H(k) is the observation matrix, the switching function ) (k
i
takes the value 1 when
the outliers and multiplicative interferences are absent (normal measurement process) and
2
( )
i ki
k = , under abnormal measurement conditions and v(k)

is the normal measurement
noise with the covariance matrix R(k) and zero mean vector.
Nonlinear Dynamics

362
In the general case, the switching function can be modeled by the finite state Markov chain
of which initial probabilities and the transition matrix are known or unknown depending
upon a priori information about the spectral characteristics of the outliers.
In the situation when not all sensors have failed, using the integrated filter estimates makes
it possible to detect failures of the individual sensors and to inform the user about them.
Our aim is to develop an integrated filter algorithm which would be fault-tolerant in the
presence of the failures and outliers mentioned above. Such an algorithm has been
developed for the aircraft state vector which contains nine components such as the x, y, z -
position, Vx, Vy, Vz - INS velocity errors, an altimeter bias and the GPS clocks shift and
velocity. But the above mentioned limitations concerning the state vector are not
fundamental and all the results can be applied to an arbitrary case.
The state and measurement equations in our case can be written in the following form:
1 1 ( , ) 1( , ),
s i i
x(k ) (k )x(k) U(k) w(k) k t k t + = + + + + (66)
[ ]
( , )1( , ),
o i i
y(k) h x(k) b(k) (k)v(k) k t k t = + + + (67)
where ( ) x k is the aircraft state vector, U(k) is the input control vector, ) , (
i s
t k is a failure
bias of the state vector arising at random time t
i
, ) , ( 1
i
t k is the unit step function, w(k) is
the system input noise vector, y(k) is the measurement vector, b(k) is the unknown constant
bias vector, ) , (
i o
t k is the Markov drift which models incipient failures of such sensors as
INS, SAS and errors due to the influence of multipath effects in the ILS, v(k) is zero mean
observation noise with covariance matrix R(k) , and { } 1 , 1 ) ( > = k is a multiplier which
describes the outliers in the observation channel.
The incipent failure model is described by (66). The a priori distributions of a random value
i
t are assumed to be unknown.
The time dependence of the sequence (k) can be described by a stationary Markov chain, for
which the initial probability vector ) 0 (

P and transition matrix

ij
P are

. ,
) 0 (
) 0 (
) 0 (
1
1 11 1
0

P P
P P
P
P
P
P
ij

(68)

Thus the system and failure model described by (66)-(67) differ from those proposed in
(Patton et al., 1989). Firstly, the failures here are treated as an additive Markov process in
the dynamic or observation equations with an unknown onset time and can describe both
deterministic and stochastic failure models. Secondly the outliers in the observation
channels are present at the system input simultaneously with possible failures. Thus, such
an approach makes it possible to describe both types of failures models - deterministic and
stochastic.
5.2 Algorithms for fault-tolerant data processing
As it follows from (66)-(67), the development of a reliable integrated filter can be advanced
by using non-linear filtering theory (Ristic et al., 2004). However, immediate application of
this theory yields too complicated an algorithm to use in real-time systems because of the
requirement for an infinite amount of memory. To overcome these difficulties it is necessary
A Detection-Estimation Method for Systems with Random Jumps with Application
to Target Tracking and Fault Diagnosis

363
to decompose the algorithm and to introduce the fault detection procedure as inherent part
of the process. Therefore it is necessary to modify the problem in the direction of
simplification. A simplification of this kind leads to a suboptimal algorithm which can be
applied to a real time system with limited memory requirements.
The first step in this direction is to separate the failure detection - estimation problem into an
independent task. A solution can be found if one knows the sensor error statistical models
and the integrated filter estimates. Using the approach presented in section 3 it is possible to
estimate failure onset time
i
t

and the value of the vector ) , (

i
t k . So in observation equation
(67) vector ) , (

i
t k can then be considered to be a known value.
The second step in solving the problem is synthesis of the integrated filtering algorithm so
that it will be sufficiently robust with respect to the presence of malfunctions (outliers) in the
observation channels.
In order to cope with this problem for the system described by equations (66) and (67), it is
necessary to use a general nonlinear filtering theory approach (Ristic et al., 2004). In this case
the estimates of the dynamic system state vector can be found as a conditional mean of the
following form (Janczak & Grishin, 2008):

1 1
2

( / ) [ ( ) / ] ( / ) ( / ),
k
i k i k
k
i
x k k E x k Y x k k P Y

= = (69)
where
1
{ (1), (2), , ( )}
k
Y y y y k = is the sequence of the input data, { (1), (2), , ( )}
k
i
k =
denotes the realization of the switching function and
], , / ) ( [ ) / (

1
k
i
k i
Y k x E k k x = (70)
are partial estimates that are calculated for each realization of the switching function. Thus
the optimal estimation algorithm requires infinitely increasing memory and cannot be
realized in practice. Practical realization can only be achieved by using different
approximations of the pdf of the estimates (69). One of the possible approaches to solving
this problem is using the Gaussian approximation method (Ristic et al., 2004). In such
an approach the state vector estimates

( / ) x k k can be expressed as the weighted sum of the

partial estimates

( / )
i
x k k corresponding to the presence and absence of the outliers in the
measurements:

= = =

, 1
1
2 2
). / ) ( ( ) ) ( , / (

) / (

i
k
ki
i
ki
i
Y k P k k k x k k x (71)
The posterior probability of the measurement channel state ) / ) ( (
1
2 k
ki
i
Y k P = depends on
the outlier stochastic characteristics. If the outliers are statistically independent,
the probability can be found from:

2 1
1 / 1
/ 2 1
1 / 1
1,
( ( ) / ( ) , )
,
( ( ) / ( ) , )
k
i ki i k
i k k
j ki j k
j
f y k k Y p
p
f y k k Y p

=
=
=
=
(72)
where
k i
p
/
is the a posteriori probability of the measurement noise covariance matrix
) (
~
2
k R R
ki
ki
= .
Nonlinear Dynamics

364
These probabilities can be calculated in real time using current data at the filter input based
on the pdf ) , ) ( / ) ( (
1
1
2
=
k
ki
i
Y k k y f of predicted estimates (Bar-Shalom et al., 2001). When
the fluctuations and outliers are independent in time, the probability
/ 1 i k ki
p q

= , where
ki
q
are the a priori probabilities.
It can be shown that for a system which contains N observation channels with outliers, this
method yields the following expression for the state vector estimate (Grishin, 2000):

}
1 2
1
1 2
1 2
1
1 2
, 1
1
2
, 1
1

( / ) ( / ) ( , / ) ( 1 / 1)
( / ) ( ) ( ) ( , / )

( ) ( ) ( 1 / 1) , 1, , 1, , ,
N
N
N
N
N
N
i i N
i i i
N
T
i i j j jj N
j i i i
j j j j j
x k k x k k p i i k x k k
P k k H k i R k p i i k
y k H k x k k i j N

=
= = +

= =

(73)

where ) / ( x

,..., ,
2 1
k k
N
i i i
is a partial estimate of the state vector for certain failure realisation
in the observation channels (sensors of navigational information),
1 2 1 2
( , , , / ) ( (1) , (2) , , ( ) / (1), , ( ))
N N
p i i i k p i i N i y y k = = = = are the a posteriori
probabilities of these realisations, (k/k) P
N
,...,i ,i i
2 1
is the update covariance matrix of the
partial estimate, i
j
=1, are values of the multiplier (k) in the j-th channel for a normal and
failure state of performance, and ) (k y
j
measurements at the output of the j-th navigational
information source.
It can be shown that a posteriori probabilities are calculated in real time as follows:

}
1 2 1 2
1 2
1 1 2
1 2
1 1
1 2 , ,..., 1 , ,..., 1
1
1 1
,..., 1 , ,..., 1
1 1 1
( , ,..., / ) ( ) / ( ), ( ) /
( ) / ( ), Y ( ) /Y ,
N N
N
N N
N
k k
N i i i i i i
k k
i i i i i
i i i
p i i i k f y k k Y p k Y
f y k k p k

= = =
=

(74)
where
1
1
, , 1
( ) / ( ),
N
k
i i
f y k k Y

is a value of the likelihood function at the point ) ( y k ,

[ ]
1
1
,..., ,
/ ) (
2 1
k
i i i
Y k p
N
- a priori probability of a certain combination of channel observation
serviceability, which can be calculated on the basis of a previous value of p and the Markov
chain characteristics:

1
( ) 1
, , 1 1
1
[ ( ) /Y ] P[ ,..., / 1], 1, ,
1
N
N
j k
i i n ij N j
j
p k P i i k i
n

=
= =
=

(75)
where
) ( j
ij n
P is the transition matrix elements of the Markov chain ) (
) (
k
j
in the j-th
observation channel. The algorithm described by (73) - (75) can be thought of as a soft
multichannel outlier screening procedure which is correct for arbitrary values of 1 > (not
necessarily for large ones).
Let us consider then, the part of the system structure (Fig. 9) which is responsible for
a decision of the failure detection-estimation problem in each information channel (sensor).
A Detection-Estimation Method for Systems with Random Jumps with Application
to Target Tracking and Fault Diagnosis

365
All of them contain a fault detection-identification algorithm (FDIA), which is used for
estimating the failures and for generating the failure alarm signal (FAS) to inform the user.
The failure detection-identification algorithm is designed on the basis of the GLR approach
for an additive Gauss-Markov model of the system failures. It can be constructed on the
assumption that no a priori information about failure onset time and the initial conditions
of vector ) , (
i
t k exists.
Since the failure vector ) , (
i
t k is part of ) , (
i
t k its estimate is also known. This estimate
can be used to cancel the input data biases, for example. The block diagram for
a cancellation of this kind is presented in Fig. 10.


Fig. 10. The fault bias cancellation method
After detecting abrupt changes to the sensor output, it is necessary to control the presence of
biases in the output estimates of the IFA to distinguish sensor failures from aircraft
manoeuvres.
It should be noted that the proposed structure also makes it possible to isolate failures, that
is, to determine if failures have occurred in the airborne navigation equipment or in the
space-based facilities. This can be realised by comparing the data of the FDIA and content of
the state matrix circuits. Following this, the failure alarm signal should be generated and
transmitted to the users.
6. Conclusion
We have presented a new recursive algorithm for joint detection and estimation of jump
changes in the dynamics and measurements of linear discrete-time systems in the presence
of outliers in observations. The algorithm has been developed on the basis of the GLR
method. The jumps were modelled as Gauss-Markov biases in state and observation
equations. The structure of the algorithm is sufficiently simple to enable it to be applied in
real-time systems with a relatively limited computational burden. The proposed models
describe a wide class of dynamic systems with jump parameters. The detection-estimation
algorithm developed, was successfully applied to the problem of radar maneuvering target
tracking and fault-tolerant signal processing for enhancing the integrity and reliability of
airborne navigation equipment. Simulation results revealed good estimation properties for
the algorithm.
7. References
Bar-Shalom, Y. & Fortmann, T. (1988). Tracking and data association, Academic Press, N.Y.
)

, (

i
t k
i
t

) (k y
) (
1
k y
+
Threshold
circuit
Failure
detection-identification
algorithm
Coordinate
recalculation
) (k H
o
+
from IFA

to IFA

- -
Nonlinear Dynamics

366
Bar-Shalom, Y.; Li, X. & Kirubarajan, R. (2001). Estimation with applications to tracking and
navigation, John Wiley & Sons, New York.
Blackman, S. & Popoli, R. (1999). Design and analysis of modern tracking systems. Artech
House, Boston.
Brown, G. & Hwang, P. (1987). GPS failure detection by autonomous means within cockpit.
Navigation, Vol. 33, No. 4, 1987, pp. 335-353.
Brown, A. (1988). Civil Aviation Integrity Requirements for the Global Positioning System.
Navigation, Vol. 35, No. 1, 1988, pp. 23-40.
Fadden, D. & Schwab, R. (1989) Aircraft Interface with Future ATC System, Proc. IEEE, Vol.
77, No 11, pp. 1745-1751.
Gertler, J. (1998). Fault Detection and Diagnosis in Engineering Systems, Marcel Dekker,
Inc,.N.Y.
Gini, F. & Rangaswamy, M./Ed. (2008). Knowledge-based radar detection, tracking and
classification, John Wiley&Sons, ISBN 978-0-470-14930-0, N. Y.
Grishin, Yu. (1994). An Application of the Additive Gauss-Markov Models of Discrete-Time
Dynamic Systems to the Problem of Abrupt Changes Detection, Proceedings of Int.
AMSE Conf. Systems: Analysis, Control & Design, pp. 211-220, v.1, July 1994, Lyon.
Grishin, Yu. (2000). Reliable data processing in an integrated GPS-based airborne
navigational equipment, Proceedings of the European Communication Conference
EUROCOMM 2000, pp 91-94, ISBN 0-7803-6323X, Munich (Germany), May 17,
2000, IEEE, NJ.
Grishin, Yu. & Janczak, D. (2006). Joint Adaptive Detection - Estimation Algorithm for
Maneuvering Target Tracking, Proceedings of the International Radar Symposium IRS
2006, pp. 375-378, ISBN 83-7207-621-9, Krakow (Poland), 24-26 May 2006, PIT,
Warszawa.
Janczak, D.

& Grishin, Yu. (2008). The GLR failure detection algorithm for a class of
nonlinear dynamic models with application to radar tracking problems, Proceedings
of the International Radar Symposium IRS 2008, pp. 233-236, ISBN 978-83-7207-757-8,
Wroclaw (Poland), 21-23 May 2008, PIT, Warszawa.
Katayma, T. & Sugimoto, S. (ed) (1997). Statistical methods in control and signal processing.
Marcel Dekker, Inc., N.Y.
Li, Rong. & Bar-Shalom, Y. (1993). Performance prediction of the interacting multiple model
algorithm, IEEE Trans., Vol. AES-29, No 3, 1993, pp. 755-771.
Mazor, E.; Dayan, J.; Averbuch, A. & Bar-Shalom, Y. (1998). Interacting multiple model
methods in target tracking: a survey, IEEE Trans., Vol. AES-34, No 1, 1998, pp. 103-
123.
Patton, R.; Frank, P. & Clark, R. (1989). Fault diagnosis in dynamic systems. Theory and
applications, Prentice Hall, N.Y.
Ristic, B.; Arulampalam, S. & Gordon, N. (2004). Beyond the Kalman filter-Partical filters for
tracking applications, Artech House, ISBN 1-58053-631-x, Boston.
Sage, A. & Melsa, I. (1971). Estimation Theory with Applications to Communication and Control,
Mc Graw-Hill, N. Y.
Sorenson, H. (ed) (1985). Kalman filtering: theory and application, IEEE Press, Piscataway, NJ.
Whang, I.; Lee, J. & Sung, T. (1994). Modified input estimation technique using
pseudoresiduals, IEEE Trans., Vol. AES-30, No 1, 1994, pp.220-227.
Willsky, A. (1976). A Survey of Design Methods for Failure Detection in Dynamic Systems,
Automatica, Vol. 12, 1976, pp. 601-611.